tcp_output.c revision e6c7d0857905f1d642cb8dbadae6794bfa1dff30
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#include <net/tcp.h> 38 39#include <linux/compiler.h> 40#include <linux/module.h> 41 42/* People can turn this off for buggy TCP's found in printers etc. */ 43int sysctl_tcp_retrans_collapse __read_mostly = 1; 44 45/* People can turn this on to work with those rare, broken TCPs that 46 * interpret the window field as a signed quantity. 47 */ 48int sysctl_tcp_workaround_signed_windows __read_mostly = 0; 49 50/* This limits the percentage of the congestion window which we 51 * will allow a single TSO frame to consume. Building TSO frames 52 * which are too large can cause TCP streams to be bursty. 53 */ 54int sysctl_tcp_tso_win_divisor __read_mostly = 3; 55 56int sysctl_tcp_mtu_probing __read_mostly = 0; 57int sysctl_tcp_base_mss __read_mostly = 512; 58 59/* By default, RFC2861 behavior. */ 60int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 61 62static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) 63{ 64 struct tcp_sock *tp = tcp_sk(sk); 65 unsigned int prior_packets = tp->packets_out; 66 67 tcp_advance_send_head(sk, skb); 68 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 69 70 /* Don't override Nagle indefinately with F-RTO */ 71 if (tp->frto_counter == 2) 72 tp->frto_counter = 3; 73 74 tp->packets_out += tcp_skb_pcount(skb); 75 if (!prior_packets) 76 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 77 inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 78} 79 80/* SND.NXT, if window was not shrunk. 81 * If window has been shrunk, what should we make? It is not clear at all. 82 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 83 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 84 * invalid. OK, let's make this for now: 85 */ 86static inline __u32 tcp_acceptable_seq(struct sock *sk) 87{ 88 struct tcp_sock *tp = tcp_sk(sk); 89 90 if (!before(tcp_wnd_end(tp), tp->snd_nxt)) 91 return tp->snd_nxt; 92 else 93 return tcp_wnd_end(tp); 94} 95 96/* Calculate mss to advertise in SYN segment. 97 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 98 * 99 * 1. It is independent of path mtu. 100 * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 101 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 102 * attached devices, because some buggy hosts are confused by 103 * large MSS. 104 * 4. We do not make 3, we advertise MSS, calculated from first 105 * hop device mtu, but allow to raise it to ip_rt_min_advmss. 106 * This may be overridden via information stored in routing table. 107 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 108 * probably even Jumbo". 109 */ 110static __u16 tcp_advertise_mss(struct sock *sk) 111{ 112 struct tcp_sock *tp = tcp_sk(sk); 113 struct dst_entry *dst = __sk_dst_get(sk); 114 int mss = tp->advmss; 115 116 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) { 117 mss = dst_metric(dst, RTAX_ADVMSS); 118 tp->advmss = mss; 119 } 120 121 return (__u16)mss; 122} 123 124/* RFC2861. Reset CWND after idle period longer RTO to "restart window". 125 * This is the first part of cwnd validation mechanism. */ 126static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst) 127{ 128 struct tcp_sock *tp = tcp_sk(sk); 129 s32 delta = tcp_time_stamp - tp->lsndtime; 130 u32 restart_cwnd = tcp_init_cwnd(tp, dst); 131 u32 cwnd = tp->snd_cwnd; 132 133 tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 134 135 tp->snd_ssthresh = tcp_current_ssthresh(sk); 136 restart_cwnd = min(restart_cwnd, cwnd); 137 138 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 139 cwnd >>= 1; 140 tp->snd_cwnd = max(cwnd, restart_cwnd); 141 tp->snd_cwnd_stamp = tcp_time_stamp; 142 tp->snd_cwnd_used = 0; 143} 144 145static void tcp_event_data_sent(struct tcp_sock *tp, 146 struct sk_buff *skb, struct sock *sk) 147{ 148 struct inet_connection_sock *icsk = inet_csk(sk); 149 const u32 now = tcp_time_stamp; 150 151 if (sysctl_tcp_slow_start_after_idle && 152 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) 153 tcp_cwnd_restart(sk, __sk_dst_get(sk)); 154 155 tp->lsndtime = now; 156 157 /* If it is a reply for ato after last received 158 * packet, enter pingpong mode. 159 */ 160 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 161 icsk->icsk_ack.pingpong = 1; 162} 163 164static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 165{ 166 tcp_dec_quickack_mode(sk, pkts); 167 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 168} 169 170/* Determine a window scaling and initial window to offer. 171 * Based on the assumption that the given amount of space 172 * will be offered. Store the results in the tp structure. 173 * NOTE: for smooth operation initial space offering should 174 * be a multiple of mss if possible. We assume here that mss >= 1. 175 * This MUST be enforced by all callers. 176 */ 177void tcp_select_initial_window(int __space, __u32 mss, 178 __u32 *rcv_wnd, __u32 *window_clamp, 179 int wscale_ok, __u8 *rcv_wscale) 180{ 181 unsigned int space = (__space < 0 ? 0 : __space); 182 183 /* If no clamp set the clamp to the max possible scaled window */ 184 if (*window_clamp == 0) 185 (*window_clamp) = (65535 << 14); 186 space = min(*window_clamp, space); 187 188 /* Quantize space offering to a multiple of mss if possible. */ 189 if (space > mss) 190 space = (space / mss) * mss; 191 192 /* NOTE: offering an initial window larger than 32767 193 * will break some buggy TCP stacks. If the admin tells us 194 * it is likely we could be speaking with such a buggy stack 195 * we will truncate our initial window offering to 32K-1 196 * unless the remote has sent us a window scaling option, 197 * which we interpret as a sign the remote TCP is not 198 * misinterpreting the window field as a signed quantity. 199 */ 200 if (sysctl_tcp_workaround_signed_windows) 201 (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 202 else 203 (*rcv_wnd) = space; 204 205 (*rcv_wscale) = 0; 206 if (wscale_ok) { 207 /* Set window scaling on max possible window 208 * See RFC1323 for an explanation of the limit to 14 209 */ 210 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 211 space = min_t(u32, space, *window_clamp); 212 while (space > 65535 && (*rcv_wscale) < 14) { 213 space >>= 1; 214 (*rcv_wscale)++; 215 } 216 } 217 218 /* Set initial window to value enough for senders, 219 * following RFC2414. Senders, not following this RFC, 220 * will be satisfied with 2. 221 */ 222 if (mss > (1 << *rcv_wscale)) { 223 int init_cwnd = 4; 224 if (mss > 1460 * 3) 225 init_cwnd = 2; 226 else if (mss > 1460) 227 init_cwnd = 3; 228 if (*rcv_wnd > init_cwnd * mss) 229 *rcv_wnd = init_cwnd * mss; 230 } 231 232 /* Set the clamp no higher than max representable value */ 233 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 234} 235 236/* Chose a new window to advertise, update state in tcp_sock for the 237 * socket, and return result with RFC1323 scaling applied. The return 238 * value can be stuffed directly into th->window for an outgoing 239 * frame. 240 */ 241static u16 tcp_select_window(struct sock *sk) 242{ 243 struct tcp_sock *tp = tcp_sk(sk); 244 u32 cur_win = tcp_receive_window(tp); 245 u32 new_win = __tcp_select_window(sk); 246 247 /* Never shrink the offered window */ 248 if (new_win < cur_win) { 249 /* Danger Will Robinson! 250 * Don't update rcv_wup/rcv_wnd here or else 251 * we will not be able to advertise a zero 252 * window in time. --DaveM 253 * 254 * Relax Will Robinson. 255 */ 256 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 257 } 258 tp->rcv_wnd = new_win; 259 tp->rcv_wup = tp->rcv_nxt; 260 261 /* Make sure we do not exceed the maximum possible 262 * scaled window. 263 */ 264 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 265 new_win = min(new_win, MAX_TCP_WINDOW); 266 else 267 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 268 269 /* RFC1323 scaling applied */ 270 new_win >>= tp->rx_opt.rcv_wscale; 271 272 /* If we advertise zero window, disable fast path. */ 273 if (new_win == 0) 274 tp->pred_flags = 0; 275 276 return new_win; 277} 278 279static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb) 280{ 281 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR; 282 if (!(tp->ecn_flags & TCP_ECN_OK)) 283 TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE; 284} 285 286static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) 287{ 288 struct tcp_sock *tp = tcp_sk(sk); 289 290 tp->ecn_flags = 0; 291 if (sysctl_tcp_ecn) { 292 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE | TCPCB_FLAG_CWR; 293 tp->ecn_flags = TCP_ECN_OK; 294 } 295} 296 297static __inline__ void 298TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th) 299{ 300 if (inet_rsk(req)->ecn_ok) 301 th->ece = 1; 302} 303 304static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, 305 int tcp_header_len) 306{ 307 struct tcp_sock *tp = tcp_sk(sk); 308 309 if (tp->ecn_flags & TCP_ECN_OK) { 310 /* Not-retransmitted data segment: set ECT and inject CWR. */ 311 if (skb->len != tcp_header_len && 312 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 313 INET_ECN_xmit(sk); 314 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 315 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 316 tcp_hdr(skb)->cwr = 1; 317 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 318 } 319 } else { 320 /* ACK or retransmitted segment: clear ECT|CE */ 321 INET_ECN_dontxmit(sk); 322 } 323 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 324 tcp_hdr(skb)->ece = 1; 325 } 326} 327 328/* Constructs common control bits of non-data skb. If SYN/FIN is present, 329 * auto increment end seqno. 330 */ 331static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 332{ 333 skb->csum = 0; 334 335 TCP_SKB_CB(skb)->flags = flags; 336 TCP_SKB_CB(skb)->sacked = 0; 337 338 skb_shinfo(skb)->gso_segs = 1; 339 skb_shinfo(skb)->gso_size = 0; 340 skb_shinfo(skb)->gso_type = 0; 341 342 TCP_SKB_CB(skb)->seq = seq; 343 if (flags & (TCPCB_FLAG_SYN | TCPCB_FLAG_FIN)) 344 seq++; 345 TCP_SKB_CB(skb)->end_seq = seq; 346} 347 348static inline int tcp_urg_mode(const struct tcp_sock *tp) 349{ 350 return tp->snd_una != tp->snd_up; 351} 352 353#define OPTION_SACK_ADVERTISE (1 << 0) 354#define OPTION_TS (1 << 1) 355#define OPTION_MD5 (1 << 2) 356 357struct tcp_out_options { 358 u8 options; /* bit field of OPTION_* */ 359 u8 ws; /* window scale, 0 to disable */ 360 u8 num_sack_blocks; /* number of SACK blocks to include */ 361 u16 mss; /* 0 to disable */ 362 __u32 tsval, tsecr; /* need to include OPTION_TS */ 363}; 364 365/* Beware: Something in the Internet is very sensitive to the ordering of 366 * TCP options, we learned this through the hard way, so be careful here. 367 * Luckily we can at least blame others for their non-compliance but from 368 * inter-operatibility perspective it seems that we're somewhat stuck with 369 * the ordering which we have been using if we want to keep working with 370 * those broken things (not that it currently hurts anybody as there isn't 371 * particular reason why the ordering would need to be changed). 372 * 373 * At least SACK_PERM as the first option is known to lead to a disaster 374 * (but it may well be that other scenarios fail similarly). 375 */ 376static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 377 const struct tcp_out_options *opts, 378 __u8 **md5_hash) { 379 if (unlikely(OPTION_MD5 & opts->options)) { 380 *ptr++ = htonl((TCPOPT_NOP << 24) | 381 (TCPOPT_NOP << 16) | 382 (TCPOPT_MD5SIG << 8) | 383 TCPOLEN_MD5SIG); 384 *md5_hash = (__u8 *)ptr; 385 ptr += 4; 386 } else { 387 *md5_hash = NULL; 388 } 389 390 if (unlikely(opts->mss)) { 391 *ptr++ = htonl((TCPOPT_MSS << 24) | 392 (TCPOLEN_MSS << 16) | 393 opts->mss); 394 } 395 396 if (likely(OPTION_TS & opts->options)) { 397 if (unlikely(OPTION_SACK_ADVERTISE & opts->options)) { 398 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 399 (TCPOLEN_SACK_PERM << 16) | 400 (TCPOPT_TIMESTAMP << 8) | 401 TCPOLEN_TIMESTAMP); 402 } else { 403 *ptr++ = htonl((TCPOPT_NOP << 24) | 404 (TCPOPT_NOP << 16) | 405 (TCPOPT_TIMESTAMP << 8) | 406 TCPOLEN_TIMESTAMP); 407 } 408 *ptr++ = htonl(opts->tsval); 409 *ptr++ = htonl(opts->tsecr); 410 } 411 412 if (unlikely(OPTION_SACK_ADVERTISE & opts->options && 413 !(OPTION_TS & opts->options))) { 414 *ptr++ = htonl((TCPOPT_NOP << 24) | 415 (TCPOPT_NOP << 16) | 416 (TCPOPT_SACK_PERM << 8) | 417 TCPOLEN_SACK_PERM); 418 } 419 420 if (unlikely(opts->ws)) { 421 *ptr++ = htonl((TCPOPT_NOP << 24) | 422 (TCPOPT_WINDOW << 16) | 423 (TCPOLEN_WINDOW << 8) | 424 opts->ws); 425 } 426 427 if (unlikely(opts->num_sack_blocks)) { 428 struct tcp_sack_block *sp = tp->rx_opt.dsack ? 429 tp->duplicate_sack : tp->selective_acks; 430 int this_sack; 431 432 *ptr++ = htonl((TCPOPT_NOP << 24) | 433 (TCPOPT_NOP << 16) | 434 (TCPOPT_SACK << 8) | 435 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 436 TCPOLEN_SACK_PERBLOCK))); 437 438 for (this_sack = 0; this_sack < opts->num_sack_blocks; 439 ++this_sack) { 440 *ptr++ = htonl(sp[this_sack].start_seq); 441 *ptr++ = htonl(sp[this_sack].end_seq); 442 } 443 444 if (tp->rx_opt.dsack) { 445 tp->rx_opt.dsack = 0; 446 tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks; 447 } 448 } 449} 450 451static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb, 452 struct tcp_out_options *opts, 453 struct tcp_md5sig_key **md5) { 454 struct tcp_sock *tp = tcp_sk(sk); 455 unsigned size = 0; 456 457#ifdef CONFIG_TCP_MD5SIG 458 *md5 = tp->af_specific->md5_lookup(sk, sk); 459 if (*md5) { 460 opts->options |= OPTION_MD5; 461 size += TCPOLEN_MD5SIG_ALIGNED; 462 } 463#else 464 *md5 = NULL; 465#endif 466 467 /* We always get an MSS option. The option bytes which will be seen in 468 * normal data packets should timestamps be used, must be in the MSS 469 * advertised. But we subtract them from tp->mss_cache so that 470 * calculations in tcp_sendmsg are simpler etc. So account for this 471 * fact here if necessary. If we don't do this correctly, as a 472 * receiver we won't recognize data packets as being full sized when we 473 * should, and thus we won't abide by the delayed ACK rules correctly. 474 * SACKs don't matter, we never delay an ACK when we have any of those 475 * going out. */ 476 opts->mss = tcp_advertise_mss(sk); 477 size += TCPOLEN_MSS_ALIGNED; 478 479 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { 480 opts->options |= OPTION_TS; 481 opts->tsval = TCP_SKB_CB(skb)->when; 482 opts->tsecr = tp->rx_opt.ts_recent; 483 size += TCPOLEN_TSTAMP_ALIGNED; 484 } 485 if (likely(sysctl_tcp_window_scaling)) { 486 opts->ws = tp->rx_opt.rcv_wscale; 487 if (likely(opts->ws)) 488 size += TCPOLEN_WSCALE_ALIGNED; 489 } 490 if (likely(sysctl_tcp_sack)) { 491 opts->options |= OPTION_SACK_ADVERTISE; 492 if (unlikely(!(OPTION_TS & opts->options))) 493 size += TCPOLEN_SACKPERM_ALIGNED; 494 } 495 496 return size; 497} 498 499static unsigned tcp_synack_options(struct sock *sk, 500 struct request_sock *req, 501 unsigned mss, struct sk_buff *skb, 502 struct tcp_out_options *opts, 503 struct tcp_md5sig_key **md5) { 504 unsigned size = 0; 505 struct inet_request_sock *ireq = inet_rsk(req); 506 char doing_ts; 507 508#ifdef CONFIG_TCP_MD5SIG 509 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 510 if (*md5) { 511 opts->options |= OPTION_MD5; 512 size += TCPOLEN_MD5SIG_ALIGNED; 513 } 514#else 515 *md5 = NULL; 516#endif 517 518 /* we can't fit any SACK blocks in a packet with MD5 + TS 519 options. There was discussion about disabling SACK rather than TS in 520 order to fit in better with old, buggy kernels, but that was deemed 521 to be unnecessary. */ 522 doing_ts = ireq->tstamp_ok && !(*md5 && ireq->sack_ok); 523 524 opts->mss = mss; 525 size += TCPOLEN_MSS_ALIGNED; 526 527 if (likely(ireq->wscale_ok)) { 528 opts->ws = ireq->rcv_wscale; 529 if (likely(opts->ws)) 530 size += TCPOLEN_WSCALE_ALIGNED; 531 } 532 if (likely(doing_ts)) { 533 opts->options |= OPTION_TS; 534 opts->tsval = TCP_SKB_CB(skb)->when; 535 opts->tsecr = req->ts_recent; 536 size += TCPOLEN_TSTAMP_ALIGNED; 537 } 538 if (likely(ireq->sack_ok)) { 539 opts->options |= OPTION_SACK_ADVERTISE; 540 if (unlikely(!doing_ts)) 541 size += TCPOLEN_SACKPERM_ALIGNED; 542 } 543 544 return size; 545} 546 547static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb, 548 struct tcp_out_options *opts, 549 struct tcp_md5sig_key **md5) { 550 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; 551 struct tcp_sock *tp = tcp_sk(sk); 552 unsigned size = 0; 553 554#ifdef CONFIG_TCP_MD5SIG 555 *md5 = tp->af_specific->md5_lookup(sk, sk); 556 if (unlikely(*md5)) { 557 opts->options |= OPTION_MD5; 558 size += TCPOLEN_MD5SIG_ALIGNED; 559 } 560#else 561 *md5 = NULL; 562#endif 563 564 if (likely(tp->rx_opt.tstamp_ok)) { 565 opts->options |= OPTION_TS; 566 opts->tsval = tcb ? tcb->when : 0; 567 opts->tsecr = tp->rx_opt.ts_recent; 568 size += TCPOLEN_TSTAMP_ALIGNED; 569 } 570 571 if (unlikely(tp->rx_opt.eff_sacks)) { 572 const unsigned remaining = MAX_TCP_OPTION_SPACE - size; 573 opts->num_sack_blocks = 574 min_t(unsigned, tp->rx_opt.eff_sacks, 575 (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 576 TCPOLEN_SACK_PERBLOCK); 577 size += TCPOLEN_SACK_BASE_ALIGNED + 578 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 579 } 580 581 return size; 582} 583 584/* This routine actually transmits TCP packets queued in by 585 * tcp_do_sendmsg(). This is used by both the initial 586 * transmission and possible later retransmissions. 587 * All SKB's seen here are completely headerless. It is our 588 * job to build the TCP header, and pass the packet down to 589 * IP so it can do the same plus pass the packet off to the 590 * device. 591 * 592 * We are working here with either a clone of the original 593 * SKB, or a fresh unique copy made by the retransmit engine. 594 */ 595static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 596 gfp_t gfp_mask) 597{ 598 const struct inet_connection_sock *icsk = inet_csk(sk); 599 struct inet_sock *inet; 600 struct tcp_sock *tp; 601 struct tcp_skb_cb *tcb; 602 struct tcp_out_options opts; 603 unsigned tcp_options_size, tcp_header_size; 604 struct tcp_md5sig_key *md5; 605 __u8 *md5_hash_location; 606 struct tcphdr *th; 607 int err; 608 609 BUG_ON(!skb || !tcp_skb_pcount(skb)); 610 611 /* If congestion control is doing timestamping, we must 612 * take such a timestamp before we potentially clone/copy. 613 */ 614 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) 615 __net_timestamp(skb); 616 617 if (likely(clone_it)) { 618 if (unlikely(skb_cloned(skb))) 619 skb = pskb_copy(skb, gfp_mask); 620 else 621 skb = skb_clone(skb, gfp_mask); 622 if (unlikely(!skb)) 623 return -ENOBUFS; 624 } 625 626 inet = inet_sk(sk); 627 tp = tcp_sk(sk); 628 tcb = TCP_SKB_CB(skb); 629 memset(&opts, 0, sizeof(opts)); 630 631 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) 632 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 633 else 634 tcp_options_size = tcp_established_options(sk, skb, &opts, 635 &md5); 636 tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 637 638 if (tcp_packets_in_flight(tp) == 0) 639 tcp_ca_event(sk, CA_EVENT_TX_START); 640 641 skb_push(skb, tcp_header_size); 642 skb_reset_transport_header(skb); 643 skb_set_owner_w(skb, sk); 644 645 /* Build TCP header and checksum it. */ 646 th = tcp_hdr(skb); 647 th->source = inet->sport; 648 th->dest = inet->dport; 649 th->seq = htonl(tcb->seq); 650 th->ack_seq = htonl(tp->rcv_nxt); 651 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 652 tcb->flags); 653 654 if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { 655 /* RFC1323: The window in SYN & SYN/ACK segments 656 * is never scaled. 657 */ 658 th->window = htons(min(tp->rcv_wnd, 65535U)); 659 } else { 660 th->window = htons(tcp_select_window(sk)); 661 } 662 th->check = 0; 663 th->urg_ptr = 0; 664 665 /* The urg_mode check is necessary during a below snd_una win probe */ 666 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 667 if (before(tp->snd_up, tcb->seq + 0x10000)) { 668 th->urg_ptr = htons(tp->snd_up - tcb->seq); 669 th->urg = 1; 670 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 671 th->urg_ptr = 0xFFFF; 672 th->urg = 1; 673 } 674 } 675 676 tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location); 677 if (likely((tcb->flags & TCPCB_FLAG_SYN) == 0)) 678 TCP_ECN_send(sk, skb, tcp_header_size); 679 680#ifdef CONFIG_TCP_MD5SIG 681 /* Calculate the MD5 hash, as we have all we need now */ 682 if (md5) { 683 sk->sk_route_caps &= ~NETIF_F_GSO_MASK; 684 tp->af_specific->calc_md5_hash(md5_hash_location, 685 md5, sk, NULL, skb); 686 } 687#endif 688 689 icsk->icsk_af_ops->send_check(sk, skb->len, skb); 690 691 if (likely(tcb->flags & TCPCB_FLAG_ACK)) 692 tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 693 694 if (skb->len != tcp_header_size) 695 tcp_event_data_sent(tp, skb, sk); 696 697 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 698 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 699 700 err = icsk->icsk_af_ops->queue_xmit(skb, 0); 701 if (likely(err <= 0)) 702 return err; 703 704 tcp_enter_cwr(sk, 1); 705 706 return net_xmit_eval(err); 707} 708 709/* This routine just queue's the buffer 710 * 711 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 712 * otherwise socket can stall. 713 */ 714static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 715{ 716 struct tcp_sock *tp = tcp_sk(sk); 717 718 /* Advance write_seq and place onto the write_queue. */ 719 tp->write_seq = TCP_SKB_CB(skb)->end_seq; 720 skb_header_release(skb); 721 tcp_add_write_queue_tail(sk, skb); 722 sk->sk_wmem_queued += skb->truesize; 723 sk_mem_charge(sk, skb->truesize); 724} 725 726static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, 727 unsigned int mss_now) 728{ 729 if (skb->len <= mss_now || !sk_can_gso(sk)) { 730 /* Avoid the costly divide in the normal 731 * non-TSO case. 732 */ 733 skb_shinfo(skb)->gso_segs = 1; 734 skb_shinfo(skb)->gso_size = 0; 735 skb_shinfo(skb)->gso_type = 0; 736 } else { 737 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 738 skb_shinfo(skb)->gso_size = mss_now; 739 skb_shinfo(skb)->gso_type = sk->sk_gso_type; 740 } 741} 742 743/* When a modification to fackets out becomes necessary, we need to check 744 * skb is counted to fackets_out or not. 745 */ 746static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb, 747 int decr) 748{ 749 struct tcp_sock *tp = tcp_sk(sk); 750 751 if (!tp->sacked_out || tcp_is_reno(tp)) 752 return; 753 754 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 755 tp->fackets_out -= decr; 756} 757 758/* Function to create two new TCP segments. Shrinks the given segment 759 * to the specified size and appends a new segment with the rest of the 760 * packet to the list. This won't be called frequently, I hope. 761 * Remember, these are still headerless SKBs at this point. 762 */ 763int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 764 unsigned int mss_now) 765{ 766 struct tcp_sock *tp = tcp_sk(sk); 767 struct sk_buff *buff; 768 int nsize, old_factor; 769 int nlen; 770 u16 flags; 771 772 BUG_ON(len > skb->len); 773 774 nsize = skb_headlen(skb) - len; 775 if (nsize < 0) 776 nsize = 0; 777 778 if (skb_cloned(skb) && 779 skb_is_nonlinear(skb) && 780 pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 781 return -ENOMEM; 782 783 /* Get a new skb... force flag on. */ 784 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 785 if (buff == NULL) 786 return -ENOMEM; /* We'll just try again later. */ 787 788 sk->sk_wmem_queued += buff->truesize; 789 sk_mem_charge(sk, buff->truesize); 790 nlen = skb->len - len - nsize; 791 buff->truesize += nlen; 792 skb->truesize -= nlen; 793 794 /* Correct the sequence numbers. */ 795 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 796 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 797 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 798 799 /* PSH and FIN should only be set in the second packet. */ 800 flags = TCP_SKB_CB(skb)->flags; 801 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH); 802 TCP_SKB_CB(buff)->flags = flags; 803 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 804 805 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 806 /* Copy and checksum data tail into the new buffer. */ 807 buff->csum = csum_partial_copy_nocheck(skb->data + len, 808 skb_put(buff, nsize), 809 nsize, 0); 810 811 skb_trim(skb, len); 812 813 skb->csum = csum_block_sub(skb->csum, buff->csum, len); 814 } else { 815 skb->ip_summed = CHECKSUM_PARTIAL; 816 skb_split(skb, buff, len); 817 } 818 819 buff->ip_summed = skb->ip_summed; 820 821 /* Looks stupid, but our code really uses when of 822 * skbs, which it never sent before. --ANK 823 */ 824 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 825 buff->tstamp = skb->tstamp; 826 827 old_factor = tcp_skb_pcount(skb); 828 829 /* Fix up tso_factor for both original and new SKB. */ 830 tcp_set_skb_tso_segs(sk, skb, mss_now); 831 tcp_set_skb_tso_segs(sk, buff, mss_now); 832 833 /* If this packet has been sent out already, we must 834 * adjust the various packet counters. 835 */ 836 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 837 int diff = old_factor - tcp_skb_pcount(skb) - 838 tcp_skb_pcount(buff); 839 840 tp->packets_out -= diff; 841 842 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 843 tp->sacked_out -= diff; 844 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 845 tp->retrans_out -= diff; 846 847 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 848 tp->lost_out -= diff; 849 850 /* Adjust Reno SACK estimate. */ 851 if (tcp_is_reno(tp) && diff > 0) { 852 tcp_dec_pcount_approx_int(&tp->sacked_out, diff); 853 tcp_verify_left_out(tp); 854 } 855 tcp_adjust_fackets_out(sk, skb, diff); 856 857 if (tp->lost_skb_hint && 858 before(TCP_SKB_CB(skb)->seq, 859 TCP_SKB_CB(tp->lost_skb_hint)->seq) && 860 (tcp_is_fack(tp) || TCP_SKB_CB(skb)->sacked)) 861 tp->lost_cnt_hint -= diff; 862 } 863 864 /* Link BUFF into the send queue. */ 865 skb_header_release(buff); 866 tcp_insert_write_queue_after(skb, buff, sk); 867 868 return 0; 869} 870 871/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 872 * eventually). The difference is that pulled data not copied, but 873 * immediately discarded. 874 */ 875static void __pskb_trim_head(struct sk_buff *skb, int len) 876{ 877 int i, k, eat; 878 879 eat = len; 880 k = 0; 881 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 882 if (skb_shinfo(skb)->frags[i].size <= eat) { 883 put_page(skb_shinfo(skb)->frags[i].page); 884 eat -= skb_shinfo(skb)->frags[i].size; 885 } else { 886 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 887 if (eat) { 888 skb_shinfo(skb)->frags[k].page_offset += eat; 889 skb_shinfo(skb)->frags[k].size -= eat; 890 eat = 0; 891 } 892 k++; 893 } 894 } 895 skb_shinfo(skb)->nr_frags = k; 896 897 skb_reset_tail_pointer(skb); 898 skb->data_len -= len; 899 skb->len = skb->data_len; 900} 901 902int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 903{ 904 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 905 return -ENOMEM; 906 907 /* If len == headlen, we avoid __skb_pull to preserve alignment. */ 908 if (unlikely(len < skb_headlen(skb))) 909 __skb_pull(skb, len); 910 else 911 __pskb_trim_head(skb, len - skb_headlen(skb)); 912 913 TCP_SKB_CB(skb)->seq += len; 914 skb->ip_summed = CHECKSUM_PARTIAL; 915 916 skb->truesize -= len; 917 sk->sk_wmem_queued -= len; 918 sk_mem_uncharge(sk, len); 919 sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 920 921 /* Any change of skb->len requires recalculation of tso 922 * factor and mss. 923 */ 924 if (tcp_skb_pcount(skb) > 1) 925 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1)); 926 927 return 0; 928} 929 930/* Not accounting for SACKs here. */ 931int tcp_mtu_to_mss(struct sock *sk, int pmtu) 932{ 933 struct tcp_sock *tp = tcp_sk(sk); 934 struct inet_connection_sock *icsk = inet_csk(sk); 935 int mss_now; 936 937 /* Calculate base mss without TCP options: 938 It is MMS_S - sizeof(tcphdr) of rfc1122 939 */ 940 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 941 942 /* Clamp it (mss_clamp does not include tcp options) */ 943 if (mss_now > tp->rx_opt.mss_clamp) 944 mss_now = tp->rx_opt.mss_clamp; 945 946 /* Now subtract optional transport overhead */ 947 mss_now -= icsk->icsk_ext_hdr_len; 948 949 /* Then reserve room for full set of TCP options and 8 bytes of data */ 950 if (mss_now < 48) 951 mss_now = 48; 952 953 /* Now subtract TCP options size, not including SACKs */ 954 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); 955 956 return mss_now; 957} 958 959/* Inverse of above */ 960int tcp_mss_to_mtu(struct sock *sk, int mss) 961{ 962 struct tcp_sock *tp = tcp_sk(sk); 963 struct inet_connection_sock *icsk = inet_csk(sk); 964 int mtu; 965 966 mtu = mss + 967 tp->tcp_header_len + 968 icsk->icsk_ext_hdr_len + 969 icsk->icsk_af_ops->net_header_len; 970 971 return mtu; 972} 973 974void tcp_mtup_init(struct sock *sk) 975{ 976 struct tcp_sock *tp = tcp_sk(sk); 977 struct inet_connection_sock *icsk = inet_csk(sk); 978 979 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 980 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 981 icsk->icsk_af_ops->net_header_len; 982 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 983 icsk->icsk_mtup.probe_size = 0; 984} 985 986/* Bound MSS / TSO packet size with the half of the window */ 987static int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize) 988{ 989 if (tp->max_window && pktsize > (tp->max_window >> 1)) 990 return max(tp->max_window >> 1, 68U - tp->tcp_header_len); 991 else 992 return pktsize; 993} 994 995/* This function synchronize snd mss to current pmtu/exthdr set. 996 997 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 998 for TCP options, but includes only bare TCP header. 999 1000 tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1001 It is minimum of user_mss and mss received with SYN. 1002 It also does not include TCP options. 1003 1004 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 1005 1006 tp->mss_cache is current effective sending mss, including 1007 all tcp options except for SACKs. It is evaluated, 1008 taking into account current pmtu, but never exceeds 1009 tp->rx_opt.mss_clamp. 1010 1011 NOTE1. rfc1122 clearly states that advertised MSS 1012 DOES NOT include either tcp or ip options. 1013 1014 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1015 are READ ONLY outside this function. --ANK (980731) 1016 */ 1017unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 1018{ 1019 struct tcp_sock *tp = tcp_sk(sk); 1020 struct inet_connection_sock *icsk = inet_csk(sk); 1021 int mss_now; 1022 1023 if (icsk->icsk_mtup.search_high > pmtu) 1024 icsk->icsk_mtup.search_high = pmtu; 1025 1026 mss_now = tcp_mtu_to_mss(sk, pmtu); 1027 mss_now = tcp_bound_to_half_wnd(tp, mss_now); 1028 1029 /* And store cached results */ 1030 icsk->icsk_pmtu_cookie = pmtu; 1031 if (icsk->icsk_mtup.enabled) 1032 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1033 tp->mss_cache = mss_now; 1034 1035 return mss_now; 1036} 1037 1038/* Compute the current effective MSS, taking SACKs and IP options, 1039 * and even PMTU discovery events into account. 1040 */ 1041unsigned int tcp_current_mss(struct sock *sk, int large_allowed) 1042{ 1043 struct tcp_sock *tp = tcp_sk(sk); 1044 struct dst_entry *dst = __sk_dst_get(sk); 1045 u32 mss_now; 1046 u16 xmit_size_goal; 1047 int doing_tso = 0; 1048 unsigned header_len; 1049 struct tcp_out_options opts; 1050 struct tcp_md5sig_key *md5; 1051 1052 mss_now = tp->mss_cache; 1053 1054 if (large_allowed && sk_can_gso(sk)) 1055 doing_tso = 1; 1056 1057 if (dst) { 1058 u32 mtu = dst_mtu(dst); 1059 if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 1060 mss_now = tcp_sync_mss(sk, mtu); 1061 } 1062 1063 header_len = tcp_established_options(sk, NULL, &opts, &md5) + 1064 sizeof(struct tcphdr); 1065 /* The mss_cache is sized based on tp->tcp_header_len, which assumes 1066 * some common options. If this is an odd packet (because we have SACK 1067 * blocks etc) then our calculated header_len will be different, and 1068 * we have to adjust mss_now correspondingly */ 1069 if (header_len != tp->tcp_header_len) { 1070 int delta = (int) header_len - tp->tcp_header_len; 1071 mss_now -= delta; 1072 } 1073 1074 xmit_size_goal = mss_now; 1075 1076 if (doing_tso) { 1077 xmit_size_goal = ((sk->sk_gso_max_size - 1) - 1078 inet_csk(sk)->icsk_af_ops->net_header_len - 1079 inet_csk(sk)->icsk_ext_hdr_len - 1080 tp->tcp_header_len); 1081 1082 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal); 1083 xmit_size_goal -= (xmit_size_goal % mss_now); 1084 } 1085 tp->xmit_size_goal = xmit_size_goal; 1086 1087 return mss_now; 1088} 1089 1090/* Congestion window validation. (RFC2861) */ 1091static void tcp_cwnd_validate(struct sock *sk) 1092{ 1093 struct tcp_sock *tp = tcp_sk(sk); 1094 1095 if (tp->packets_out >= tp->snd_cwnd) { 1096 /* Network is feed fully. */ 1097 tp->snd_cwnd_used = 0; 1098 tp->snd_cwnd_stamp = tcp_time_stamp; 1099 } else { 1100 /* Network starves. */ 1101 if (tp->packets_out > tp->snd_cwnd_used) 1102 tp->snd_cwnd_used = tp->packets_out; 1103 1104 if (sysctl_tcp_slow_start_after_idle && 1105 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1106 tcp_cwnd_application_limited(sk); 1107 } 1108} 1109 1110/* Returns the portion of skb which can be sent right away without 1111 * introducing MSS oddities to segment boundaries. In rare cases where 1112 * mss_now != mss_cache, we will request caller to create a small skb 1113 * per input skb which could be mostly avoided here (if desired). 1114 * 1115 * We explicitly want to create a request for splitting write queue tail 1116 * to a small skb for Nagle purposes while avoiding unnecessary modulos, 1117 * thus all the complexity (cwnd_len is always MSS multiple which we 1118 * return whenever allowed by the other factors). Basically we need the 1119 * modulo only when the receiver window alone is the limiting factor or 1120 * when we would be allowed to send the split-due-to-Nagle skb fully. 1121 */ 1122static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb, 1123 unsigned int mss_now, unsigned int cwnd) 1124{ 1125 struct tcp_sock *tp = tcp_sk(sk); 1126 u32 needed, window, cwnd_len; 1127 1128 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1129 cwnd_len = mss_now * cwnd; 1130 1131 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) 1132 return cwnd_len; 1133 1134 needed = min(skb->len, window); 1135 1136 if (cwnd_len <= needed) 1137 return cwnd_len; 1138 1139 return needed - needed % mss_now; 1140} 1141 1142/* Can at least one segment of SKB be sent right now, according to the 1143 * congestion window rules? If so, return how many segments are allowed. 1144 */ 1145static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, 1146 struct sk_buff *skb) 1147{ 1148 u32 in_flight, cwnd; 1149 1150 /* Don't be strict about the congestion window for the final FIN. */ 1151 if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && 1152 tcp_skb_pcount(skb) == 1) 1153 return 1; 1154 1155 in_flight = tcp_packets_in_flight(tp); 1156 cwnd = tp->snd_cwnd; 1157 if (in_flight < cwnd) 1158 return (cwnd - in_flight); 1159 1160 return 0; 1161} 1162 1163/* This must be invoked the first time we consider transmitting 1164 * SKB onto the wire. 1165 */ 1166static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, 1167 unsigned int mss_now) 1168{ 1169 int tso_segs = tcp_skb_pcount(skb); 1170 1171 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1172 tcp_set_skb_tso_segs(sk, skb, mss_now); 1173 tso_segs = tcp_skb_pcount(skb); 1174 } 1175 return tso_segs; 1176} 1177 1178static inline int tcp_minshall_check(const struct tcp_sock *tp) 1179{ 1180 return after(tp->snd_sml, tp->snd_una) && 1181 !after(tp->snd_sml, tp->snd_nxt); 1182} 1183 1184/* Return 0, if packet can be sent now without violation Nagle's rules: 1185 * 1. It is full sized. 1186 * 2. Or it contains FIN. (already checked by caller) 1187 * 3. Or TCP_NODELAY was set. 1188 * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1189 * With Minshall's modification: all sent small packets are ACKed. 1190 */ 1191static inline int tcp_nagle_check(const struct tcp_sock *tp, 1192 const struct sk_buff *skb, 1193 unsigned mss_now, int nonagle) 1194{ 1195 return (skb->len < mss_now && 1196 ((nonagle & TCP_NAGLE_CORK) || 1197 (!nonagle && tp->packets_out && tcp_minshall_check(tp)))); 1198} 1199 1200/* Return non-zero if the Nagle test allows this packet to be 1201 * sent now. 1202 */ 1203static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, 1204 unsigned int cur_mss, int nonagle) 1205{ 1206 /* Nagle rule does not apply to frames, which sit in the middle of the 1207 * write_queue (they have no chances to get new data). 1208 * 1209 * This is implemented in the callers, where they modify the 'nonagle' 1210 * argument based upon the location of SKB in the send queue. 1211 */ 1212 if (nonagle & TCP_NAGLE_PUSH) 1213 return 1; 1214 1215 /* Don't use the nagle rule for urgent data (or for the final FIN). 1216 * Nagle can be ignored during F-RTO too (see RFC4138). 1217 */ 1218 if (tcp_urg_mode(tp) || (tp->frto_counter == 2) || 1219 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)) 1220 return 1; 1221 1222 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 1223 return 1; 1224 1225 return 0; 1226} 1227 1228/* Does at least the first segment of SKB fit into the send window? */ 1229static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, 1230 unsigned int cur_mss) 1231{ 1232 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1233 1234 if (skb->len > cur_mss) 1235 end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1236 1237 return !after(end_seq, tcp_wnd_end(tp)); 1238} 1239 1240/* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1241 * should be put on the wire right now. If so, it returns the number of 1242 * packets allowed by the congestion window. 1243 */ 1244static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, 1245 unsigned int cur_mss, int nonagle) 1246{ 1247 struct tcp_sock *tp = tcp_sk(sk); 1248 unsigned int cwnd_quota; 1249 1250 tcp_init_tso_segs(sk, skb, cur_mss); 1251 1252 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1253 return 0; 1254 1255 cwnd_quota = tcp_cwnd_test(tp, skb); 1256 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1257 cwnd_quota = 0; 1258 1259 return cwnd_quota; 1260} 1261 1262int tcp_may_send_now(struct sock *sk) 1263{ 1264 struct tcp_sock *tp = tcp_sk(sk); 1265 struct sk_buff *skb = tcp_send_head(sk); 1266 1267 return (skb && 1268 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1), 1269 (tcp_skb_is_last(sk, skb) ? 1270 tp->nonagle : TCP_NAGLE_PUSH))); 1271} 1272 1273/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1274 * which is put after SKB on the list. It is very much like 1275 * tcp_fragment() except that it may make several kinds of assumptions 1276 * in order to speed up the splitting operation. In particular, we 1277 * know that all the data is in scatter-gather pages, and that the 1278 * packet has never been sent out before (and thus is not cloned). 1279 */ 1280static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1281 unsigned int mss_now) 1282{ 1283 struct sk_buff *buff; 1284 int nlen = skb->len - len; 1285 u16 flags; 1286 1287 /* All of a TSO frame must be composed of paged data. */ 1288 if (skb->len != skb->data_len) 1289 return tcp_fragment(sk, skb, len, mss_now); 1290 1291 buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC); 1292 if (unlikely(buff == NULL)) 1293 return -ENOMEM; 1294 1295 sk->sk_wmem_queued += buff->truesize; 1296 sk_mem_charge(sk, buff->truesize); 1297 buff->truesize += nlen; 1298 skb->truesize -= nlen; 1299 1300 /* Correct the sequence numbers. */ 1301 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1302 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1303 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1304 1305 /* PSH and FIN should only be set in the second packet. */ 1306 flags = TCP_SKB_CB(skb)->flags; 1307 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH); 1308 TCP_SKB_CB(buff)->flags = flags; 1309 1310 /* This packet was never sent out yet, so no SACK bits. */ 1311 TCP_SKB_CB(buff)->sacked = 0; 1312 1313 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1314 skb_split(skb, buff, len); 1315 1316 /* Fix up tso_factor for both original and new SKB. */ 1317 tcp_set_skb_tso_segs(sk, skb, mss_now); 1318 tcp_set_skb_tso_segs(sk, buff, mss_now); 1319 1320 /* Link BUFF into the send queue. */ 1321 skb_header_release(buff); 1322 tcp_insert_write_queue_after(skb, buff, sk); 1323 1324 return 0; 1325} 1326 1327/* Try to defer sending, if possible, in order to minimize the amount 1328 * of TSO splitting we do. View it as a kind of TSO Nagle test. 1329 * 1330 * This algorithm is from John Heffner. 1331 */ 1332static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1333{ 1334 struct tcp_sock *tp = tcp_sk(sk); 1335 const struct inet_connection_sock *icsk = inet_csk(sk); 1336 u32 send_win, cong_win, limit, in_flight; 1337 1338 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) 1339 goto send_now; 1340 1341 if (icsk->icsk_ca_state != TCP_CA_Open) 1342 goto send_now; 1343 1344 /* Defer for less than two clock ticks. */ 1345 if (tp->tso_deferred && 1346 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1347 goto send_now; 1348 1349 in_flight = tcp_packets_in_flight(tp); 1350 1351 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1352 1353 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1354 1355 /* From in_flight test above, we know that cwnd > in_flight. */ 1356 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1357 1358 limit = min(send_win, cong_win); 1359 1360 /* If a full-sized TSO skb can be sent, do it. */ 1361 if (limit >= sk->sk_gso_max_size) 1362 goto send_now; 1363 1364 /* Middle in queue won't get any more data, full sendable already? */ 1365 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 1366 goto send_now; 1367 1368 if (sysctl_tcp_tso_win_divisor) { 1369 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1370 1371 /* If at least some fraction of a window is available, 1372 * just use it. 1373 */ 1374 chunk /= sysctl_tcp_tso_win_divisor; 1375 if (limit >= chunk) 1376 goto send_now; 1377 } else { 1378 /* Different approach, try not to defer past a single 1379 * ACK. Receiver should ACK every other full sized 1380 * frame, so if we have space for more than 3 frames 1381 * then send now. 1382 */ 1383 if (limit > tcp_max_burst(tp) * tp->mss_cache) 1384 goto send_now; 1385 } 1386 1387 /* Ok, it looks like it is advisable to defer. */ 1388 tp->tso_deferred = 1 | (jiffies << 1); 1389 1390 return 1; 1391 1392send_now: 1393 tp->tso_deferred = 0; 1394 return 0; 1395} 1396 1397/* Create a new MTU probe if we are ready. 1398 * Returns 0 if we should wait to probe (no cwnd available), 1399 * 1 if a probe was sent, 1400 * -1 otherwise 1401 */ 1402static int tcp_mtu_probe(struct sock *sk) 1403{ 1404 struct tcp_sock *tp = tcp_sk(sk); 1405 struct inet_connection_sock *icsk = inet_csk(sk); 1406 struct sk_buff *skb, *nskb, *next; 1407 int len; 1408 int probe_size; 1409 int size_needed; 1410 int copy; 1411 int mss_now; 1412 1413 /* Not currently probing/verifying, 1414 * not in recovery, 1415 * have enough cwnd, and 1416 * not SACKing (the variable headers throw things off) */ 1417 if (!icsk->icsk_mtup.enabled || 1418 icsk->icsk_mtup.probe_size || 1419 inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 1420 tp->snd_cwnd < 11 || 1421 tp->rx_opt.eff_sacks) 1422 return -1; 1423 1424 /* Very simple search strategy: just double the MSS. */ 1425 mss_now = tcp_current_mss(sk, 0); 1426 probe_size = 2 * tp->mss_cache; 1427 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 1428 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 1429 /* TODO: set timer for probe_converge_event */ 1430 return -1; 1431 } 1432 1433 /* Have enough data in the send queue to probe? */ 1434 if (tp->write_seq - tp->snd_nxt < size_needed) 1435 return -1; 1436 1437 if (tp->snd_wnd < size_needed) 1438 return -1; 1439 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 1440 return 0; 1441 1442 /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1443 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1444 if (!tcp_packets_in_flight(tp)) 1445 return -1; 1446 else 1447 return 0; 1448 } 1449 1450 /* We're allowed to probe. Build it now. */ 1451 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 1452 return -1; 1453 sk->sk_wmem_queued += nskb->truesize; 1454 sk_mem_charge(sk, nskb->truesize); 1455 1456 skb = tcp_send_head(sk); 1457 1458 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 1459 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 1460 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK; 1461 TCP_SKB_CB(nskb)->sacked = 0; 1462 nskb->csum = 0; 1463 nskb->ip_summed = skb->ip_summed; 1464 1465 tcp_insert_write_queue_before(nskb, skb, sk); 1466 1467 len = 0; 1468 tcp_for_write_queue_from_safe(skb, next, sk) { 1469 copy = min_t(int, skb->len, probe_size - len); 1470 if (nskb->ip_summed) 1471 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 1472 else 1473 nskb->csum = skb_copy_and_csum_bits(skb, 0, 1474 skb_put(nskb, copy), 1475 copy, nskb->csum); 1476 1477 if (skb->len <= copy) { 1478 /* We've eaten all the data from this skb. 1479 * Throw it away. */ 1480 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags; 1481 tcp_unlink_write_queue(skb, sk); 1482 sk_wmem_free_skb(sk, skb); 1483 } else { 1484 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags & 1485 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); 1486 if (!skb_shinfo(skb)->nr_frags) { 1487 skb_pull(skb, copy); 1488 if (skb->ip_summed != CHECKSUM_PARTIAL) 1489 skb->csum = csum_partial(skb->data, 1490 skb->len, 0); 1491 } else { 1492 __pskb_trim_head(skb, copy); 1493 tcp_set_skb_tso_segs(sk, skb, mss_now); 1494 } 1495 TCP_SKB_CB(skb)->seq += copy; 1496 } 1497 1498 len += copy; 1499 1500 if (len >= probe_size) 1501 break; 1502 } 1503 tcp_init_tso_segs(sk, nskb, nskb->len); 1504 1505 /* We're ready to send. If this fails, the probe will 1506 * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 1507 TCP_SKB_CB(nskb)->when = tcp_time_stamp; 1508 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 1509 /* Decrement cwnd here because we are sending 1510 * effectively two packets. */ 1511 tp->snd_cwnd--; 1512 tcp_event_new_data_sent(sk, nskb); 1513 1514 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 1515 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 1516 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 1517 1518 return 1; 1519 } 1520 1521 return -1; 1522} 1523 1524/* This routine writes packets to the network. It advances the 1525 * send_head. This happens as incoming acks open up the remote 1526 * window for us. 1527 * 1528 * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1529 * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1530 * account rare use of URG, this is not a big flaw. 1531 * 1532 * Returns 1, if no segments are in flight and we have queued segments, but 1533 * cannot send anything now because of SWS or another problem. 1534 */ 1535static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1536 int push_one, gfp_t gfp) 1537{ 1538 struct tcp_sock *tp = tcp_sk(sk); 1539 struct sk_buff *skb; 1540 unsigned int tso_segs, sent_pkts; 1541 int cwnd_quota; 1542 int result; 1543 1544 sent_pkts = 0; 1545 1546 if (!push_one) { 1547 /* Do MTU probing. */ 1548 result = tcp_mtu_probe(sk); 1549 if (!result) { 1550 return 0; 1551 } else if (result > 0) { 1552 sent_pkts = 1; 1553 } 1554 } 1555 1556 while ((skb = tcp_send_head(sk))) { 1557 unsigned int limit; 1558 1559 tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1560 BUG_ON(!tso_segs); 1561 1562 cwnd_quota = tcp_cwnd_test(tp, skb); 1563 if (!cwnd_quota) 1564 break; 1565 1566 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1567 break; 1568 1569 if (tso_segs == 1) { 1570 if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1571 (tcp_skb_is_last(sk, skb) ? 1572 nonagle : TCP_NAGLE_PUSH)))) 1573 break; 1574 } else { 1575 if (!push_one && tcp_tso_should_defer(sk, skb)) 1576 break; 1577 } 1578 1579 limit = mss_now; 1580 if (tso_segs > 1 && !tcp_urg_mode(tp)) 1581 limit = tcp_mss_split_point(sk, skb, mss_now, 1582 cwnd_quota); 1583 1584 if (skb->len > limit && 1585 unlikely(tso_fragment(sk, skb, limit, mss_now))) 1586 break; 1587 1588 TCP_SKB_CB(skb)->when = tcp_time_stamp; 1589 1590 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 1591 break; 1592 1593 /* Advance the send_head. This one is sent out. 1594 * This call will increment packets_out. 1595 */ 1596 tcp_event_new_data_sent(sk, skb); 1597 1598 tcp_minshall_update(tp, mss_now, skb); 1599 sent_pkts++; 1600 1601 if (push_one) 1602 break; 1603 } 1604 1605 if (likely(sent_pkts)) { 1606 tcp_cwnd_validate(sk); 1607 return 0; 1608 } 1609 return !tp->packets_out && tcp_send_head(sk); 1610} 1611 1612/* Push out any pending frames which were held back due to 1613 * TCP_CORK or attempt at coalescing tiny packets. 1614 * The socket must be locked by the caller. 1615 */ 1616void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 1617 int nonagle) 1618{ 1619 struct sk_buff *skb = tcp_send_head(sk); 1620 1621 if (!skb) 1622 return; 1623 1624 /* If we are closed, the bytes will have to remain here. 1625 * In time closedown will finish, we empty the write queue and 1626 * all will be happy. 1627 */ 1628 if (unlikely(sk->sk_state == TCP_CLOSE)) 1629 return; 1630 1631 if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC)) 1632 tcp_check_probe_timer(sk); 1633} 1634 1635/* Send _single_ skb sitting at the send head. This function requires 1636 * true push pending frames to setup probe timer etc. 1637 */ 1638void tcp_push_one(struct sock *sk, unsigned int mss_now) 1639{ 1640 struct sk_buff *skb = tcp_send_head(sk); 1641 1642 BUG_ON(!skb || skb->len < mss_now); 1643 1644 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 1645} 1646 1647/* This function returns the amount that we can raise the 1648 * usable window based on the following constraints 1649 * 1650 * 1. The window can never be shrunk once it is offered (RFC 793) 1651 * 2. We limit memory per socket 1652 * 1653 * RFC 1122: 1654 * "the suggested [SWS] avoidance algorithm for the receiver is to keep 1655 * RECV.NEXT + RCV.WIN fixed until: 1656 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 1657 * 1658 * i.e. don't raise the right edge of the window until you can raise 1659 * it at least MSS bytes. 1660 * 1661 * Unfortunately, the recommended algorithm breaks header prediction, 1662 * since header prediction assumes th->window stays fixed. 1663 * 1664 * Strictly speaking, keeping th->window fixed violates the receiver 1665 * side SWS prevention criteria. The problem is that under this rule 1666 * a stream of single byte packets will cause the right side of the 1667 * window to always advance by a single byte. 1668 * 1669 * Of course, if the sender implements sender side SWS prevention 1670 * then this will not be a problem. 1671 * 1672 * BSD seems to make the following compromise: 1673 * 1674 * If the free space is less than the 1/4 of the maximum 1675 * space available and the free space is less than 1/2 mss, 1676 * then set the window to 0. 1677 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 1678 * Otherwise, just prevent the window from shrinking 1679 * and from being larger than the largest representable value. 1680 * 1681 * This prevents incremental opening of the window in the regime 1682 * where TCP is limited by the speed of the reader side taking 1683 * data out of the TCP receive queue. It does nothing about 1684 * those cases where the window is constrained on the sender side 1685 * because the pipeline is full. 1686 * 1687 * BSD also seems to "accidentally" limit itself to windows that are a 1688 * multiple of MSS, at least until the free space gets quite small. 1689 * This would appear to be a side effect of the mbuf implementation. 1690 * Combining these two algorithms results in the observed behavior 1691 * of having a fixed window size at almost all times. 1692 * 1693 * Below we obtain similar behavior by forcing the offered window to 1694 * a multiple of the mss when it is feasible to do so. 1695 * 1696 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 1697 * Regular options like TIMESTAMP are taken into account. 1698 */ 1699u32 __tcp_select_window(struct sock *sk) 1700{ 1701 struct inet_connection_sock *icsk = inet_csk(sk); 1702 struct tcp_sock *tp = tcp_sk(sk); 1703 /* MSS for the peer's data. Previous versions used mss_clamp 1704 * here. I don't know if the value based on our guesses 1705 * of peer's MSS is better for the performance. It's more correct 1706 * but may be worse for the performance because of rcv_mss 1707 * fluctuations. --SAW 1998/11/1 1708 */ 1709 int mss = icsk->icsk_ack.rcv_mss; 1710 int free_space = tcp_space(sk); 1711 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 1712 int window; 1713 1714 if (mss > full_space) 1715 mss = full_space; 1716 1717 if (free_space < (full_space >> 1)) { 1718 icsk->icsk_ack.quick = 0; 1719 1720 if (tcp_memory_pressure) 1721 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 1722 4U * tp->advmss); 1723 1724 if (free_space < mss) 1725 return 0; 1726 } 1727 1728 if (free_space > tp->rcv_ssthresh) 1729 free_space = tp->rcv_ssthresh; 1730 1731 /* Don't do rounding if we are using window scaling, since the 1732 * scaled window will not line up with the MSS boundary anyway. 1733 */ 1734 window = tp->rcv_wnd; 1735 if (tp->rx_opt.rcv_wscale) { 1736 window = free_space; 1737 1738 /* Advertise enough space so that it won't get scaled away. 1739 * Import case: prevent zero window announcement if 1740 * 1<<rcv_wscale > mss. 1741 */ 1742 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 1743 window = (((window >> tp->rx_opt.rcv_wscale) + 1) 1744 << tp->rx_opt.rcv_wscale); 1745 } else { 1746 /* Get the largest window that is a nice multiple of mss. 1747 * Window clamp already applied above. 1748 * If our current window offering is within 1 mss of the 1749 * free space we just keep it. This prevents the divide 1750 * and multiply from happening most of the time. 1751 * We also don't do any window rounding when the free space 1752 * is too small. 1753 */ 1754 if (window <= free_space - mss || window > free_space) 1755 window = (free_space / mss) * mss; 1756 else if (mss == full_space && 1757 free_space > window + (full_space >> 1)) 1758 window = free_space; 1759 } 1760 1761 return window; 1762} 1763 1764/* Collapses two adjacent SKB's during retransmission. */ 1765static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 1766{ 1767 struct tcp_sock *tp = tcp_sk(sk); 1768 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 1769 int skb_size, next_skb_size; 1770 1771 skb_size = skb->len; 1772 next_skb_size = next_skb->len; 1773 1774 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 1775 1776 tcp_highest_sack_combine(sk, next_skb, skb); 1777 1778 tcp_unlink_write_queue(next_skb, sk); 1779 1780 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 1781 next_skb_size); 1782 1783 if (next_skb->ip_summed == CHECKSUM_PARTIAL) 1784 skb->ip_summed = CHECKSUM_PARTIAL; 1785 1786 if (skb->ip_summed != CHECKSUM_PARTIAL) 1787 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 1788 1789 /* Update sequence range on original skb. */ 1790 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 1791 1792 /* Merge over control information. This moves PSH/FIN etc. over */ 1793 TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(next_skb)->flags; 1794 1795 /* All done, get rid of second SKB and account for it so 1796 * packet counting does not break. 1797 */ 1798 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 1799 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_RETRANS) 1800 tp->retrans_out -= tcp_skb_pcount(next_skb); 1801 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_LOST) 1802 tp->lost_out -= tcp_skb_pcount(next_skb); 1803 /* Reno case is special. Sigh... */ 1804 if (tcp_is_reno(tp) && tp->sacked_out) 1805 tcp_dec_pcount_approx(&tp->sacked_out, next_skb); 1806 1807 tcp_adjust_fackets_out(sk, next_skb, tcp_skb_pcount(next_skb)); 1808 tp->packets_out -= tcp_skb_pcount(next_skb); 1809 1810 /* changed transmit queue under us so clear hints */ 1811 tcp_clear_retrans_hints_partial(tp); 1812 if (next_skb == tp->retransmit_skb_hint) 1813 tp->retransmit_skb_hint = skb; 1814 1815 sk_wmem_free_skb(sk, next_skb); 1816} 1817 1818static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb) 1819{ 1820 if (tcp_skb_pcount(skb) > 1) 1821 return 0; 1822 /* TODO: SACK collapsing could be used to remove this condition */ 1823 if (skb_shinfo(skb)->nr_frags != 0) 1824 return 0; 1825 if (skb_cloned(skb)) 1826 return 0; 1827 if (skb == tcp_send_head(sk)) 1828 return 0; 1829 /* Some heurestics for collapsing over SACK'd could be invented */ 1830 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1831 return 0; 1832 1833 return 1; 1834} 1835 1836static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 1837 int space) 1838{ 1839 struct tcp_sock *tp = tcp_sk(sk); 1840 struct sk_buff *skb = to, *tmp; 1841 int first = 1; 1842 1843 if (!sysctl_tcp_retrans_collapse) 1844 return; 1845 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) 1846 return; 1847 1848 tcp_for_write_queue_from_safe(skb, tmp, sk) { 1849 if (!tcp_can_collapse(sk, skb)) 1850 break; 1851 1852 space -= skb->len; 1853 1854 if (first) { 1855 first = 0; 1856 continue; 1857 } 1858 1859 if (space < 0) 1860 break; 1861 /* Punt if not enough space exists in the first SKB for 1862 * the data in the second 1863 */ 1864 if (skb->len > skb_tailroom(to)) 1865 break; 1866 1867 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 1868 break; 1869 1870 tcp_collapse_retrans(sk, to); 1871 } 1872} 1873 1874/* This retransmits one SKB. Policy decisions and retransmit queue 1875 * state updates are done by the caller. Returns non-zero if an 1876 * error occurred which prevented the send. 1877 */ 1878int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 1879{ 1880 struct tcp_sock *tp = tcp_sk(sk); 1881 struct inet_connection_sock *icsk = inet_csk(sk); 1882 unsigned int cur_mss; 1883 int err; 1884 1885 /* Inconslusive MTU probe */ 1886 if (icsk->icsk_mtup.probe_size) { 1887 icsk->icsk_mtup.probe_size = 0; 1888 } 1889 1890 /* Do not sent more than we queued. 1/4 is reserved for possible 1891 * copying overhead: fragmentation, tunneling, mangling etc. 1892 */ 1893 if (atomic_read(&sk->sk_wmem_alloc) > 1894 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 1895 return -EAGAIN; 1896 1897 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 1898 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1899 BUG(); 1900 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 1901 return -ENOMEM; 1902 } 1903 1904 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 1905 return -EHOSTUNREACH; /* Routing failure or similar. */ 1906 1907 cur_mss = tcp_current_mss(sk, 0); 1908 1909 /* If receiver has shrunk his window, and skb is out of 1910 * new window, do not retransmit it. The exception is the 1911 * case, when window is shrunk to zero. In this case 1912 * our retransmit serves as a zero window probe. 1913 */ 1914 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) 1915 && TCP_SKB_CB(skb)->seq != tp->snd_una) 1916 return -EAGAIN; 1917 1918 if (skb->len > cur_mss) { 1919 if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 1920 return -ENOMEM; /* We'll try again later. */ 1921 } else { 1922 tcp_init_tso_segs(sk, skb, cur_mss); 1923 } 1924 1925 tcp_retrans_try_collapse(sk, skb, cur_mss); 1926 1927 /* Some Solaris stacks overoptimize and ignore the FIN on a 1928 * retransmit when old data is attached. So strip it off 1929 * since it is cheap to do so and saves bytes on the network. 1930 */ 1931 if (skb->len > 0 && 1932 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && 1933 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 1934 if (!pskb_trim(skb, 0)) { 1935 /* Reuse, even though it does some unnecessary work */ 1936 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, 1937 TCP_SKB_CB(skb)->flags); 1938 skb->ip_summed = CHECKSUM_NONE; 1939 } 1940 } 1941 1942 /* Make a copy, if the first transmission SKB clone we made 1943 * is still in somebody's hands, else make a clone. 1944 */ 1945 TCP_SKB_CB(skb)->when = tcp_time_stamp; 1946 1947 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 1948 1949 if (err == 0) { 1950 /* Update global TCP statistics. */ 1951 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 1952 1953 tp->total_retrans++; 1954 1955#if FASTRETRANS_DEBUG > 0 1956 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 1957 if (net_ratelimit()) 1958 printk(KERN_DEBUG "retrans_out leaked.\n"); 1959 } 1960#endif 1961 if (!tp->retrans_out) 1962 tp->lost_retrans_low = tp->snd_nxt; 1963 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 1964 tp->retrans_out += tcp_skb_pcount(skb); 1965 1966 /* Save stamp of the first retransmit. */ 1967 if (!tp->retrans_stamp) 1968 tp->retrans_stamp = TCP_SKB_CB(skb)->when; 1969 1970 tp->undo_retrans++; 1971 1972 /* snd_nxt is stored to detect loss of retransmitted segment, 1973 * see tcp_input.c tcp_sacktag_write_queue(). 1974 */ 1975 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 1976 } 1977 return err; 1978} 1979 1980static int tcp_can_forward_retransmit(struct sock *sk) 1981{ 1982 const struct inet_connection_sock *icsk = inet_csk(sk); 1983 struct tcp_sock *tp = tcp_sk(sk); 1984 1985 /* Forward retransmissions are possible only during Recovery. */ 1986 if (icsk->icsk_ca_state != TCP_CA_Recovery) 1987 return 0; 1988 1989 /* No forward retransmissions in Reno are possible. */ 1990 if (tcp_is_reno(tp)) 1991 return 0; 1992 1993 /* Yeah, we have to make difficult choice between forward transmission 1994 * and retransmission... Both ways have their merits... 1995 * 1996 * For now we do not retransmit anything, while we have some new 1997 * segments to send. In the other cases, follow rule 3 for 1998 * NextSeg() specified in RFC3517. 1999 */ 2000 2001 if (tcp_may_send_now(sk)) 2002 return 0; 2003 2004 return 1; 2005} 2006 2007/* This gets called after a retransmit timeout, and the initially 2008 * retransmitted data is acknowledged. It tries to continue 2009 * resending the rest of the retransmit queue, until either 2010 * we've sent it all or the congestion window limit is reached. 2011 * If doing SACK, the first ACK which comes back for a timeout 2012 * based retransmit packet might feed us FACK information again. 2013 * If so, we use it to avoid unnecessarily retransmissions. 2014 */ 2015void tcp_xmit_retransmit_queue(struct sock *sk) 2016{ 2017 const struct inet_connection_sock *icsk = inet_csk(sk); 2018 struct tcp_sock *tp = tcp_sk(sk); 2019 struct sk_buff *skb; 2020 struct sk_buff *hole = NULL; 2021 u32 last_lost; 2022 int mib_idx; 2023 int fwd_rexmitting = 0; 2024 2025 if (!tp->lost_out) 2026 tp->retransmit_high = tp->snd_una; 2027 2028 if (tp->retransmit_skb_hint) { 2029 skb = tp->retransmit_skb_hint; 2030 last_lost = TCP_SKB_CB(skb)->end_seq; 2031 if (after(last_lost, tp->retransmit_high)) 2032 last_lost = tp->retransmit_high; 2033 } else { 2034 skb = tcp_write_queue_head(sk); 2035 last_lost = tp->snd_una; 2036 } 2037 2038 tcp_for_write_queue_from(skb, sk) { 2039 __u8 sacked = TCP_SKB_CB(skb)->sacked; 2040 2041 if (skb == tcp_send_head(sk)) 2042 break; 2043 /* we could do better than to assign each time */ 2044 if (hole == NULL) 2045 tp->retransmit_skb_hint = skb; 2046 2047 /* Assume this retransmit will generate 2048 * only one packet for congestion window 2049 * calculation purposes. This works because 2050 * tcp_retransmit_skb() will chop up the 2051 * packet to be MSS sized and all the 2052 * packet counting works out. 2053 */ 2054 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 2055 return; 2056 2057 if (fwd_rexmitting) { 2058begin_fwd: 2059 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2060 break; 2061 mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 2062 2063 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2064 tp->retransmit_high = last_lost; 2065 if (!tcp_can_forward_retransmit(sk)) 2066 break; 2067 /* Backtrack if necessary to non-L'ed skb */ 2068 if (hole != NULL) { 2069 skb = hole; 2070 hole = NULL; 2071 } 2072 fwd_rexmitting = 1; 2073 goto begin_fwd; 2074 2075 } else if (!(sacked & TCPCB_LOST)) { 2076 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 2077 hole = skb; 2078 continue; 2079 2080 } else { 2081 last_lost = TCP_SKB_CB(skb)->end_seq; 2082 if (icsk->icsk_ca_state != TCP_CA_Loss) 2083 mib_idx = LINUX_MIB_TCPFASTRETRANS; 2084 else 2085 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 2086 } 2087 2088 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 2089 continue; 2090 2091 if (tcp_retransmit_skb(sk, skb)) 2092 return; 2093 NET_INC_STATS_BH(sock_net(sk), mib_idx); 2094 2095 if (skb == tcp_write_queue_head(sk)) 2096 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2097 inet_csk(sk)->icsk_rto, 2098 TCP_RTO_MAX); 2099 } 2100} 2101 2102/* Send a fin. The caller locks the socket for us. This cannot be 2103 * allowed to fail queueing a FIN frame under any circumstances. 2104 */ 2105void tcp_send_fin(struct sock *sk) 2106{ 2107 struct tcp_sock *tp = tcp_sk(sk); 2108 struct sk_buff *skb = tcp_write_queue_tail(sk); 2109 int mss_now; 2110 2111 /* Optimization, tack on the FIN if we have a queue of 2112 * unsent frames. But be careful about outgoing SACKS 2113 * and IP options. 2114 */ 2115 mss_now = tcp_current_mss(sk, 1); 2116 2117 if (tcp_send_head(sk) != NULL) { 2118 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN; 2119 TCP_SKB_CB(skb)->end_seq++; 2120 tp->write_seq++; 2121 } else { 2122 /* Socket is locked, keep trying until memory is available. */ 2123 for (;;) { 2124 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL); 2125 if (skb) 2126 break; 2127 yield(); 2128 } 2129 2130 /* Reserve space for headers and prepare control bits. */ 2131 skb_reserve(skb, MAX_TCP_HEADER); 2132 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2133 tcp_init_nondata_skb(skb, tp->write_seq, 2134 TCPCB_FLAG_ACK | TCPCB_FLAG_FIN); 2135 tcp_queue_skb(sk, skb); 2136 } 2137 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 2138} 2139 2140/* We get here when a process closes a file descriptor (either due to 2141 * an explicit close() or as a byproduct of exit()'ing) and there 2142 * was unread data in the receive queue. This behavior is recommended 2143 * by RFC 2525, section 2.17. -DaveM 2144 */ 2145void tcp_send_active_reset(struct sock *sk, gfp_t priority) 2146{ 2147 struct sk_buff *skb; 2148 2149 /* NOTE: No TCP options attached and we never retransmit this. */ 2150 skb = alloc_skb(MAX_TCP_HEADER, priority); 2151 if (!skb) { 2152 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 2153 return; 2154 } 2155 2156 /* Reserve space for headers and prepare control bits. */ 2157 skb_reserve(skb, MAX_TCP_HEADER); 2158 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2159 TCPCB_FLAG_ACK | TCPCB_FLAG_RST); 2160 /* Send it off. */ 2161 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2162 if (tcp_transmit_skb(sk, skb, 0, priority)) 2163 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 2164 2165 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 2166} 2167 2168/* WARNING: This routine must only be called when we have already sent 2169 * a SYN packet that crossed the incoming SYN that caused this routine 2170 * to get called. If this assumption fails then the initial rcv_wnd 2171 * and rcv_wscale values will not be correct. 2172 */ 2173int tcp_send_synack(struct sock *sk) 2174{ 2175 struct sk_buff *skb; 2176 2177 skb = tcp_write_queue_head(sk); 2178 if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) { 2179 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); 2180 return -EFAULT; 2181 } 2182 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) { 2183 if (skb_cloned(skb)) { 2184 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 2185 if (nskb == NULL) 2186 return -ENOMEM; 2187 tcp_unlink_write_queue(skb, sk); 2188 skb_header_release(nskb); 2189 __tcp_add_write_queue_head(sk, nskb); 2190 sk_wmem_free_skb(sk, skb); 2191 sk->sk_wmem_queued += nskb->truesize; 2192 sk_mem_charge(sk, nskb->truesize); 2193 skb = nskb; 2194 } 2195 2196 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK; 2197 TCP_ECN_send_synack(tcp_sk(sk), skb); 2198 } 2199 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2200 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2201} 2202 2203/* 2204 * Prepare a SYN-ACK. 2205 */ 2206struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2207 struct request_sock *req) 2208{ 2209 struct inet_request_sock *ireq = inet_rsk(req); 2210 struct tcp_sock *tp = tcp_sk(sk); 2211 struct tcphdr *th; 2212 int tcp_header_size; 2213 struct tcp_out_options opts; 2214 struct sk_buff *skb; 2215 struct tcp_md5sig_key *md5; 2216 __u8 *md5_hash_location; 2217 int mss; 2218 2219 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); 2220 if (skb == NULL) 2221 return NULL; 2222 2223 /* Reserve space for headers. */ 2224 skb_reserve(skb, MAX_TCP_HEADER); 2225 2226 skb->dst = dst_clone(dst); 2227 2228 mss = dst_metric(dst, RTAX_ADVMSS); 2229 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2230 mss = tp->rx_opt.user_mss; 2231 2232 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 2233 __u8 rcv_wscale; 2234 /* Set this up on the first call only */ 2235 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 2236 /* tcp_full_space because it is guaranteed to be the first packet */ 2237 tcp_select_initial_window(tcp_full_space(sk), 2238 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 2239 &req->rcv_wnd, 2240 &req->window_clamp, 2241 ireq->wscale_ok, 2242 &rcv_wscale); 2243 ireq->rcv_wscale = rcv_wscale; 2244 } 2245 2246 memset(&opts, 0, sizeof(opts)); 2247#ifdef CONFIG_SYN_COOKIES 2248 if (unlikely(req->cookie_ts)) 2249 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 2250 else 2251#endif 2252 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2253 tcp_header_size = tcp_synack_options(sk, req, mss, 2254 skb, &opts, &md5) + 2255 sizeof(struct tcphdr); 2256 2257 skb_push(skb, tcp_header_size); 2258 skb_reset_transport_header(skb); 2259 2260 th = tcp_hdr(skb); 2261 memset(th, 0, sizeof(struct tcphdr)); 2262 th->syn = 1; 2263 th->ack = 1; 2264 TCP_ECN_make_synack(req, th); 2265 th->source = ireq->loc_port; 2266 th->dest = ireq->rmt_port; 2267 /* Setting of flags are superfluous here for callers (and ECE is 2268 * not even correctly set) 2269 */ 2270 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2271 TCPCB_FLAG_SYN | TCPCB_FLAG_ACK); 2272 th->seq = htonl(TCP_SKB_CB(skb)->seq); 2273 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); 2274 2275 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2276 th->window = htons(min(req->rcv_wnd, 65535U)); 2277 tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location); 2278 th->doff = (tcp_header_size >> 2); 2279 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 2280 2281#ifdef CONFIG_TCP_MD5SIG 2282 /* Okay, we have all we need - do the md5 hash if needed */ 2283 if (md5) { 2284 tp->af_specific->calc_md5_hash(md5_hash_location, 2285 md5, NULL, req, skb); 2286 } 2287#endif 2288 2289 return skb; 2290} 2291 2292/* 2293 * Do all connect socket setups that can be done AF independent. 2294 */ 2295static void tcp_connect_init(struct sock *sk) 2296{ 2297 struct dst_entry *dst = __sk_dst_get(sk); 2298 struct tcp_sock *tp = tcp_sk(sk); 2299 __u8 rcv_wscale; 2300 2301 /* We'll fix this up when we get a response from the other end. 2302 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 2303 */ 2304 tp->tcp_header_len = sizeof(struct tcphdr) + 2305 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 2306 2307#ifdef CONFIG_TCP_MD5SIG 2308 if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2309 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2310#endif 2311 2312 /* If user gave his TCP_MAXSEG, record it to clamp */ 2313 if (tp->rx_opt.user_mss) 2314 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 2315 tp->max_window = 0; 2316 tcp_mtup_init(sk); 2317 tcp_sync_mss(sk, dst_mtu(dst)); 2318 2319 if (!tp->window_clamp) 2320 tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 2321 tp->advmss = dst_metric(dst, RTAX_ADVMSS); 2322 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2323 tp->advmss = tp->rx_opt.user_mss; 2324 2325 tcp_initialize_rcv_mss(sk); 2326 2327 tcp_select_initial_window(tcp_full_space(sk), 2328 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 2329 &tp->rcv_wnd, 2330 &tp->window_clamp, 2331 sysctl_tcp_window_scaling, 2332 &rcv_wscale); 2333 2334 tp->rx_opt.rcv_wscale = rcv_wscale; 2335 tp->rcv_ssthresh = tp->rcv_wnd; 2336 2337 sk->sk_err = 0; 2338 sock_reset_flag(sk, SOCK_DONE); 2339 tp->snd_wnd = 0; 2340 tcp_init_wl(tp, tp->write_seq, 0); 2341 tp->snd_una = tp->write_seq; 2342 tp->snd_sml = tp->write_seq; 2343 tp->snd_up = tp->write_seq; 2344 tp->rcv_nxt = 0; 2345 tp->rcv_wup = 0; 2346 tp->copied_seq = 0; 2347 2348 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2349 inet_csk(sk)->icsk_retransmits = 0; 2350 tcp_clear_retrans(tp); 2351} 2352 2353/* 2354 * Build a SYN and send it off. 2355 */ 2356int tcp_connect(struct sock *sk) 2357{ 2358 struct tcp_sock *tp = tcp_sk(sk); 2359 struct sk_buff *buff; 2360 2361 tcp_connect_init(sk); 2362 2363 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 2364 if (unlikely(buff == NULL)) 2365 return -ENOBUFS; 2366 2367 /* Reserve space for headers. */ 2368 skb_reserve(buff, MAX_TCP_HEADER); 2369 2370 tp->snd_nxt = tp->write_seq; 2371 tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN); 2372 TCP_ECN_send_syn(sk, buff); 2373 2374 /* Send it off. */ 2375 TCP_SKB_CB(buff)->when = tcp_time_stamp; 2376 tp->retrans_stamp = TCP_SKB_CB(buff)->when; 2377 skb_header_release(buff); 2378 __tcp_add_write_queue_tail(sk, buff); 2379 sk->sk_wmem_queued += buff->truesize; 2380 sk_mem_charge(sk, buff->truesize); 2381 tp->packets_out += tcp_skb_pcount(buff); 2382 tcp_transmit_skb(sk, buff, 1, GFP_KERNEL); 2383 2384 /* We change tp->snd_nxt after the tcp_transmit_skb() call 2385 * in order to make this packet get counted in tcpOutSegs. 2386 */ 2387 tp->snd_nxt = tp->write_seq; 2388 tp->pushed_seq = tp->write_seq; 2389 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 2390 2391 /* Timer for repeating the SYN until an answer. */ 2392 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2393 inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 2394 return 0; 2395} 2396 2397/* Send out a delayed ack, the caller does the policy checking 2398 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 2399 * for details. 2400 */ 2401void tcp_send_delayed_ack(struct sock *sk) 2402{ 2403 struct inet_connection_sock *icsk = inet_csk(sk); 2404 int ato = icsk->icsk_ack.ato; 2405 unsigned long timeout; 2406 2407 if (ato > TCP_DELACK_MIN) { 2408 const struct tcp_sock *tp = tcp_sk(sk); 2409 int max_ato = HZ / 2; 2410 2411 if (icsk->icsk_ack.pingpong || 2412 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 2413 max_ato = TCP_DELACK_MAX; 2414 2415 /* Slow path, intersegment interval is "high". */ 2416 2417 /* If some rtt estimate is known, use it to bound delayed ack. 2418 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 2419 * directly. 2420 */ 2421 if (tp->srtt) { 2422 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); 2423 2424 if (rtt < max_ato) 2425 max_ato = rtt; 2426 } 2427 2428 ato = min(ato, max_ato); 2429 } 2430 2431 /* Stay within the limit we were given */ 2432 timeout = jiffies + ato; 2433 2434 /* Use new timeout only if there wasn't a older one earlier. */ 2435 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 2436 /* If delack timer was blocked or is about to expire, 2437 * send ACK now. 2438 */ 2439 if (icsk->icsk_ack.blocked || 2440 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 2441 tcp_send_ack(sk); 2442 return; 2443 } 2444 2445 if (!time_before(timeout, icsk->icsk_ack.timeout)) 2446 timeout = icsk->icsk_ack.timeout; 2447 } 2448 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 2449 icsk->icsk_ack.timeout = timeout; 2450 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 2451} 2452 2453/* This routine sends an ack and also updates the window. */ 2454void tcp_send_ack(struct sock *sk) 2455{ 2456 struct sk_buff *buff; 2457 2458 /* If we have been reset, we may not send again. */ 2459 if (sk->sk_state == TCP_CLOSE) 2460 return; 2461 2462 /* We are not putting this on the write queue, so 2463 * tcp_transmit_skb() will set the ownership to this 2464 * sock. 2465 */ 2466 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 2467 if (buff == NULL) { 2468 inet_csk_schedule_ack(sk); 2469 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 2470 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 2471 TCP_DELACK_MAX, TCP_RTO_MAX); 2472 return; 2473 } 2474 2475 /* Reserve space for headers and prepare control bits. */ 2476 skb_reserve(buff, MAX_TCP_HEADER); 2477 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK); 2478 2479 /* Send it off, this clears delayed acks for us. */ 2480 TCP_SKB_CB(buff)->when = tcp_time_stamp; 2481 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); 2482} 2483 2484/* This routine sends a packet with an out of date sequence 2485 * number. It assumes the other end will try to ack it. 2486 * 2487 * Question: what should we make while urgent mode? 2488 * 4.4BSD forces sending single byte of data. We cannot send 2489 * out of window data, because we have SND.NXT==SND.MAX... 2490 * 2491 * Current solution: to send TWO zero-length segments in urgent mode: 2492 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 2493 * out-of-date with SND.UNA-1 to probe window. 2494 */ 2495static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 2496{ 2497 struct tcp_sock *tp = tcp_sk(sk); 2498 struct sk_buff *skb; 2499 2500 /* We don't queue it, tcp_transmit_skb() sets ownership. */ 2501 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 2502 if (skb == NULL) 2503 return -1; 2504 2505 /* Reserve space for headers and set control bits. */ 2506 skb_reserve(skb, MAX_TCP_HEADER); 2507 /* Use a previous sequence. This should cause the other 2508 * end to send an ack. Don't queue or clone SKB, just 2509 * send it. 2510 */ 2511 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK); 2512 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2513 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 2514} 2515 2516int tcp_write_wakeup(struct sock *sk) 2517{ 2518 struct tcp_sock *tp = tcp_sk(sk); 2519 struct sk_buff *skb; 2520 2521 if (sk->sk_state == TCP_CLOSE) 2522 return -1; 2523 2524 if ((skb = tcp_send_head(sk)) != NULL && 2525 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 2526 int err; 2527 unsigned int mss = tcp_current_mss(sk, 0); 2528 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 2529 2530 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 2531 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 2532 2533 /* We are probing the opening of a window 2534 * but the window size is != 0 2535 * must have been a result SWS avoidance ( sender ) 2536 */ 2537 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 2538 skb->len > mss) { 2539 seg_size = min(seg_size, mss); 2540 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 2541 if (tcp_fragment(sk, skb, seg_size, mss)) 2542 return -1; 2543 } else if (!tcp_skb_pcount(skb)) 2544 tcp_set_skb_tso_segs(sk, skb, mss); 2545 2546 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 2547 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2548 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2549 if (!err) 2550 tcp_event_new_data_sent(sk, skb); 2551 return err; 2552 } else { 2553 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 2554 tcp_xmit_probe_skb(sk, 1); 2555 return tcp_xmit_probe_skb(sk, 0); 2556 } 2557} 2558 2559/* A window probe timeout has occurred. If window is not closed send 2560 * a partial packet else a zero probe. 2561 */ 2562void tcp_send_probe0(struct sock *sk) 2563{ 2564 struct inet_connection_sock *icsk = inet_csk(sk); 2565 struct tcp_sock *tp = tcp_sk(sk); 2566 int err; 2567 2568 err = tcp_write_wakeup(sk); 2569 2570 if (tp->packets_out || !tcp_send_head(sk)) { 2571 /* Cancel probe timer, if it is not required. */ 2572 icsk->icsk_probes_out = 0; 2573 icsk->icsk_backoff = 0; 2574 return; 2575 } 2576 2577 if (err <= 0) { 2578 if (icsk->icsk_backoff < sysctl_tcp_retries2) 2579 icsk->icsk_backoff++; 2580 icsk->icsk_probes_out++; 2581 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2582 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 2583 TCP_RTO_MAX); 2584 } else { 2585 /* If packet was not sent due to local congestion, 2586 * do not backoff and do not remember icsk_probes_out. 2587 * Let local senders to fight for local resources. 2588 * 2589 * Use accumulated backoff yet. 2590 */ 2591 if (!icsk->icsk_probes_out) 2592 icsk->icsk_probes_out = 1; 2593 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2594 min(icsk->icsk_rto << icsk->icsk_backoff, 2595 TCP_RESOURCE_PROBE_INTERVAL), 2596 TCP_RTO_MAX); 2597 } 2598} 2599 2600EXPORT_SYMBOL(tcp_select_initial_window); 2601EXPORT_SYMBOL(tcp_connect); 2602EXPORT_SYMBOL(tcp_make_synack); 2603EXPORT_SYMBOL(tcp_simple_retransmit); 2604EXPORT_SYMBOL(tcp_sync_mss); 2605EXPORT_SYMBOL(tcp_mtup_init); 2606