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. */ 43 int 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 */ 48 int 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 */ 54 int sysctl_tcp_tso_win_divisor __read_mostly = 3; 55 56 int sysctl_tcp_mtu_probing __read_mostly = 0; 57 int sysctl_tcp_base_mss __read_mostly = 512; 58 59 /* By default, RFC2861 behavior. */ 60 int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 61 62 static 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 */ 86 static 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 */ 110 static __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. */ 126 static 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 } 144 145 static 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 } 163 164 static 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 */ 177 void 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 */ 241 static 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 } 278 279 static 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 } 285 286 static 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 } 296 297 static __inline__ void 298 TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th) 299 { 300 if (inet_rsk(req)->ecn_ok) 301 th->ece = 1; 302 } 303 304 static 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 */ 331 static 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 } 347 348 static 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) 356 357 struct 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 */ 376 static 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 } 450 451 static 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 } 498 499 static 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 } 546 547 static 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 */ 595 static 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) && 667 between(tp->snd_up, tcb->seq + 1, tcb->seq + 0xFFFF))) { 668 th->urg_ptr = htons(tp->snd_up - tcb->seq); 669 th->urg = 1; 670 } 671 672 tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location); 673 if (likely((tcb->flags & TCPCB_FLAG_SYN) == 0)) 674 TCP_ECN_send(sk, skb, tcp_header_size); 675 676 #ifdef CONFIG_TCP_MD5SIG 677 /* Calculate the MD5 hash, as we have all we need now */ 678 if (md5) { 679 sk->sk_route_caps &= ~NETIF_F_GSO_MASK; 680 tp->af_specific->calc_md5_hash(md5_hash_location, 681 md5, sk, NULL, skb); 682 } 683 #endif 684 685 icsk->icsk_af_ops->send_check(sk, skb->len, skb); 686 687 if (likely(tcb->flags & TCPCB_FLAG_ACK)) 688 tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 689 690 if (skb->len != tcp_header_size) 691 tcp_event_data_sent(tp, skb, sk); 692 693 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 694 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 695 696 err = icsk->icsk_af_ops->queue_xmit(skb, 0); 697 if (likely(err <= 0)) 698 return err; 699 700 tcp_enter_cwr(sk, 1); 701 702 return net_xmit_eval(err); 703 } 704 705 /* This routine just queue's the buffer 706 * 707 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 708 * otherwise socket can stall. 709 */ 710 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 711 { 712 struct tcp_sock *tp = tcp_sk(sk); 713 714 /* Advance write_seq and place onto the write_queue. */ 715 tp->write_seq = TCP_SKB_CB(skb)->end_seq; 716 skb_header_release(skb); 717 tcp_add_write_queue_tail(sk, skb); 718 sk->sk_wmem_queued += skb->truesize; 719 sk_mem_charge(sk, skb->truesize); 720 } 721 722 static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, 723 unsigned int mss_now) 724 { 725 if (skb->len <= mss_now || !sk_can_gso(sk)) { 726 /* Avoid the costly divide in the normal 727 * non-TSO case. 728 */ 729 skb_shinfo(skb)->gso_segs = 1; 730 skb_shinfo(skb)->gso_size = 0; 731 skb_shinfo(skb)->gso_type = 0; 732 } else { 733 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 734 skb_shinfo(skb)->gso_size = mss_now; 735 skb_shinfo(skb)->gso_type = sk->sk_gso_type; 736 } 737 } 738 739 /* When a modification to fackets out becomes necessary, we need to check 740 * skb is counted to fackets_out or not. 741 */ 742 static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb, 743 int decr) 744 { 745 struct tcp_sock *tp = tcp_sk(sk); 746 747 if (!tp->sacked_out || tcp_is_reno(tp)) 748 return; 749 750 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 751 tp->fackets_out -= decr; 752 } 753 754 /* Function to create two new TCP segments. Shrinks the given segment 755 * to the specified size and appends a new segment with the rest of the 756 * packet to the list. This won't be called frequently, I hope. 757 * Remember, these are still headerless SKBs at this point. 758 */ 759 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 760 unsigned int mss_now) 761 { 762 struct tcp_sock *tp = tcp_sk(sk); 763 struct sk_buff *buff; 764 int nsize, old_factor; 765 int nlen; 766 u16 flags; 767 768 BUG_ON(len > skb->len); 769 770 tcp_clear_retrans_hints_partial(tp); 771 nsize = skb_headlen(skb) - len; 772 if (nsize < 0) 773 nsize = 0; 774 775 if (skb_cloned(skb) && 776 skb_is_nonlinear(skb) && 777 pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 778 return -ENOMEM; 779 780 /* Get a new skb... force flag on. */ 781 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 782 if (buff == NULL) 783 return -ENOMEM; /* We'll just try again later. */ 784 785 sk->sk_wmem_queued += buff->truesize; 786 sk_mem_charge(sk, buff->truesize); 787 nlen = skb->len - len - nsize; 788 buff->truesize += nlen; 789 skb->truesize -= nlen; 790 791 /* Correct the sequence numbers. */ 792 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 793 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 794 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 795 796 /* PSH and FIN should only be set in the second packet. */ 797 flags = TCP_SKB_CB(skb)->flags; 798 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH); 799 TCP_SKB_CB(buff)->flags = flags; 800 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 801 802 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 803 /* Copy and checksum data tail into the new buffer. */ 804 buff->csum = csum_partial_copy_nocheck(skb->data + len, 805 skb_put(buff, nsize), 806 nsize, 0); 807 808 skb_trim(skb, len); 809 810 skb->csum = csum_block_sub(skb->csum, buff->csum, len); 811 } else { 812 skb->ip_summed = CHECKSUM_PARTIAL; 813 skb_split(skb, buff, len); 814 } 815 816 buff->ip_summed = skb->ip_summed; 817 818 /* Looks stupid, but our code really uses when of 819 * skbs, which it never sent before. --ANK 820 */ 821 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 822 buff->tstamp = skb->tstamp; 823 824 old_factor = tcp_skb_pcount(skb); 825 826 /* Fix up tso_factor for both original and new SKB. */ 827 tcp_set_skb_tso_segs(sk, skb, mss_now); 828 tcp_set_skb_tso_segs(sk, buff, mss_now); 829 830 /* If this packet has been sent out already, we must 831 * adjust the various packet counters. 832 */ 833 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 834 int diff = old_factor - tcp_skb_pcount(skb) - 835 tcp_skb_pcount(buff); 836 837 tp->packets_out -= diff; 838 839 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 840 tp->sacked_out -= diff; 841 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 842 tp->retrans_out -= diff; 843 844 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 845 tp->lost_out -= diff; 846 847 /* Adjust Reno SACK estimate. */ 848 if (tcp_is_reno(tp) && diff > 0) { 849 tcp_dec_pcount_approx_int(&tp->sacked_out, diff); 850 tcp_verify_left_out(tp); 851 } 852 tcp_adjust_fackets_out(sk, skb, diff); 853 } 854 855 /* Link BUFF into the send queue. */ 856 skb_header_release(buff); 857 tcp_insert_write_queue_after(skb, buff, sk); 858 859 return 0; 860 } 861 862 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 863 * eventually). The difference is that pulled data not copied, but 864 * immediately discarded. 865 */ 866 static void __pskb_trim_head(struct sk_buff *skb, int len) 867 { 868 int i, k, eat; 869 870 eat = len; 871 k = 0; 872 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 873 if (skb_shinfo(skb)->frags[i].size <= eat) { 874 put_page(skb_shinfo(skb)->frags[i].page); 875 eat -= skb_shinfo(skb)->frags[i].size; 876 } else { 877 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 878 if (eat) { 879 skb_shinfo(skb)->frags[k].page_offset += eat; 880 skb_shinfo(skb)->frags[k].size -= eat; 881 eat = 0; 882 } 883 k++; 884 } 885 } 886 skb_shinfo(skb)->nr_frags = k; 887 888 skb_reset_tail_pointer(skb); 889 skb->data_len -= len; 890 skb->len = skb->data_len; 891 } 892 893 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 894 { 895 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 896 return -ENOMEM; 897 898 /* If len == headlen, we avoid __skb_pull to preserve alignment. */ 899 if (unlikely(len < skb_headlen(skb))) 900 __skb_pull(skb, len); 901 else 902 __pskb_trim_head(skb, len - skb_headlen(skb)); 903 904 TCP_SKB_CB(skb)->seq += len; 905 skb->ip_summed = CHECKSUM_PARTIAL; 906 907 skb->truesize -= len; 908 sk->sk_wmem_queued -= len; 909 sk_mem_uncharge(sk, len); 910 sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 911 912 /* Any change of skb->len requires recalculation of tso 913 * factor and mss. 914 */ 915 if (tcp_skb_pcount(skb) > 1) 916 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1)); 917 918 return 0; 919 } 920 921 /* Not accounting for SACKs here. */ 922 int tcp_mtu_to_mss(struct sock *sk, int pmtu) 923 { 924 struct tcp_sock *tp = tcp_sk(sk); 925 struct inet_connection_sock *icsk = inet_csk(sk); 926 int mss_now; 927 928 /* Calculate base mss without TCP options: 929 It is MMS_S - sizeof(tcphdr) of rfc1122 930 */ 931 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 932 933 /* Clamp it (mss_clamp does not include tcp options) */ 934 if (mss_now > tp->rx_opt.mss_clamp) 935 mss_now = tp->rx_opt.mss_clamp; 936 937 /* Now subtract optional transport overhead */ 938 mss_now -= icsk->icsk_ext_hdr_len; 939 940 /* Then reserve room for full set of TCP options and 8 bytes of data */ 941 if (mss_now < 48) 942 mss_now = 48; 943 944 /* Now subtract TCP options size, not including SACKs */ 945 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); 946 947 return mss_now; 948 } 949 950 /* Inverse of above */ 951 int tcp_mss_to_mtu(struct sock *sk, int mss) 952 { 953 struct tcp_sock *tp = tcp_sk(sk); 954 struct inet_connection_sock *icsk = inet_csk(sk); 955 int mtu; 956 957 mtu = mss + 958 tp->tcp_header_len + 959 icsk->icsk_ext_hdr_len + 960 icsk->icsk_af_ops->net_header_len; 961 962 return mtu; 963 } 964 965 void tcp_mtup_init(struct sock *sk) 966 { 967 struct tcp_sock *tp = tcp_sk(sk); 968 struct inet_connection_sock *icsk = inet_csk(sk); 969 970 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 971 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 972 icsk->icsk_af_ops->net_header_len; 973 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 974 icsk->icsk_mtup.probe_size = 0; 975 } 976 977 /* Bound MSS / TSO packet size with the half of the window */ 978 static int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize) 979 { 980 if (tp->max_window && pktsize > (tp->max_window >> 1)) 981 return max(tp->max_window >> 1, 68U - tp->tcp_header_len); 982 else 983 return pktsize; 984 } 985 986 /* This function synchronize snd mss to current pmtu/exthdr set. 987 988 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 989 for TCP options, but includes only bare TCP header. 990 991 tp->rx_opt.mss_clamp is mss negotiated at connection setup. 992 It is minimum of user_mss and mss received with SYN. 993 It also does not include TCP options. 994 995 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 996 997 tp->mss_cache is current effective sending mss, including 998 all tcp options except for SACKs. It is evaluated, 999 taking into account current pmtu, but never exceeds 1000 tp->rx_opt.mss_clamp. 1001 1002 NOTE1. rfc1122 clearly states that advertised MSS 1003 DOES NOT include either tcp or ip options. 1004 1005 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1006 are READ ONLY outside this function. --ANK (980731) 1007 */ 1008 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 1009 { 1010 struct tcp_sock *tp = tcp_sk(sk); 1011 struct inet_connection_sock *icsk = inet_csk(sk); 1012 int mss_now; 1013 1014 if (icsk->icsk_mtup.search_high > pmtu) 1015 icsk->icsk_mtup.search_high = pmtu; 1016 1017 mss_now = tcp_mtu_to_mss(sk, pmtu); 1018 mss_now = tcp_bound_to_half_wnd(tp, mss_now); 1019 1020 /* And store cached results */ 1021 icsk->icsk_pmtu_cookie = pmtu; 1022 if (icsk->icsk_mtup.enabled) 1023 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1024 tp->mss_cache = mss_now; 1025 1026 return mss_now; 1027 } 1028 1029 /* Compute the current effective MSS, taking SACKs and IP options, 1030 * and even PMTU discovery events into account. 1031 * 1032 * LARGESEND note: !tcp_urg_mode is overkill, only frames up to snd_up 1033 * cannot be large. However, taking into account rare use of URG, this 1034 * is not a big flaw. 1035 */ 1036 unsigned int tcp_current_mss(struct sock *sk, int large_allowed) 1037 { 1038 struct tcp_sock *tp = tcp_sk(sk); 1039 struct dst_entry *dst = __sk_dst_get(sk); 1040 u32 mss_now; 1041 u16 xmit_size_goal; 1042 int doing_tso = 0; 1043 unsigned header_len; 1044 struct tcp_out_options opts; 1045 struct tcp_md5sig_key *md5; 1046 1047 mss_now = tp->mss_cache; 1048 1049 if (large_allowed && sk_can_gso(sk) && !tcp_urg_mode(tp)) 1050 doing_tso = 1; 1051 1052 if (dst) { 1053 u32 mtu = dst_mtu(dst); 1054 if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 1055 mss_now = tcp_sync_mss(sk, mtu); 1056 } 1057 1058 header_len = tcp_established_options(sk, NULL, &opts, &md5) + 1059 sizeof(struct tcphdr); 1060 /* The mss_cache is sized based on tp->tcp_header_len, which assumes 1061 * some common options. If this is an odd packet (because we have SACK 1062 * blocks etc) then our calculated header_len will be different, and 1063 * we have to adjust mss_now correspondingly */ 1064 if (header_len != tp->tcp_header_len) { 1065 int delta = (int) header_len - tp->tcp_header_len; 1066 mss_now -= delta; 1067 } 1068 1069 xmit_size_goal = mss_now; 1070 1071 if (doing_tso) { 1072 xmit_size_goal = ((sk->sk_gso_max_size - 1) - 1073 inet_csk(sk)->icsk_af_ops->net_header_len - 1074 inet_csk(sk)->icsk_ext_hdr_len - 1075 tp->tcp_header_len); 1076 1077 xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal); 1078 xmit_size_goal -= (xmit_size_goal % mss_now); 1079 } 1080 tp->xmit_size_goal = xmit_size_goal; 1081 1082 return mss_now; 1083 } 1084 1085 /* Congestion window validation. (RFC2861) */ 1086 static void tcp_cwnd_validate(struct sock *sk) 1087 { 1088 struct tcp_sock *tp = tcp_sk(sk); 1089 1090 if (tp->packets_out >= tp->snd_cwnd) { 1091 /* Network is feed fully. */ 1092 tp->snd_cwnd_used = 0; 1093 tp->snd_cwnd_stamp = tcp_time_stamp; 1094 } else { 1095 /* Network starves. */ 1096 if (tp->packets_out > tp->snd_cwnd_used) 1097 tp->snd_cwnd_used = tp->packets_out; 1098 1099 if (sysctl_tcp_slow_start_after_idle && 1100 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1101 tcp_cwnd_application_limited(sk); 1102 } 1103 } 1104 1105 /* Returns the portion of skb which can be sent right away without 1106 * introducing MSS oddities to segment boundaries. In rare cases where 1107 * mss_now != mss_cache, we will request caller to create a small skb 1108 * per input skb which could be mostly avoided here (if desired). 1109 * 1110 * We explicitly want to create a request for splitting write queue tail 1111 * to a small skb for Nagle purposes while avoiding unnecessary modulos, 1112 * thus all the complexity (cwnd_len is always MSS multiple which we 1113 * return whenever allowed by the other factors). Basically we need the 1114 * modulo only when the receiver window alone is the limiting factor or 1115 * when we would be allowed to send the split-due-to-Nagle skb fully. 1116 */ 1117 static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb, 1118 unsigned int mss_now, unsigned int cwnd) 1119 { 1120 struct tcp_sock *tp = tcp_sk(sk); 1121 u32 needed, window, cwnd_len; 1122 1123 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1124 cwnd_len = mss_now * cwnd; 1125 1126 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) 1127 return cwnd_len; 1128 1129 needed = min(skb->len, window); 1130 1131 if (cwnd_len <= needed) 1132 return cwnd_len; 1133 1134 return needed - needed % mss_now; 1135 } 1136 1137 /* Can at least one segment of SKB be sent right now, according to the 1138 * congestion window rules? If so, return how many segments are allowed. 1139 */ 1140 static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, 1141 struct sk_buff *skb) 1142 { 1143 u32 in_flight, cwnd; 1144 1145 /* Don't be strict about the congestion window for the final FIN. */ 1146 if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && 1147 tcp_skb_pcount(skb) == 1) 1148 return 1; 1149 1150 in_flight = tcp_packets_in_flight(tp); 1151 cwnd = tp->snd_cwnd; 1152 if (in_flight < cwnd) 1153 return (cwnd - in_flight); 1154 1155 return 0; 1156 } 1157 1158 /* This must be invoked the first time we consider transmitting 1159 * SKB onto the wire. 1160 */ 1161 static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, 1162 unsigned int mss_now) 1163 { 1164 int tso_segs = tcp_skb_pcount(skb); 1165 1166 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1167 tcp_set_skb_tso_segs(sk, skb, mss_now); 1168 tso_segs = tcp_skb_pcount(skb); 1169 } 1170 return tso_segs; 1171 } 1172 1173 static inline int tcp_minshall_check(const struct tcp_sock *tp) 1174 { 1175 return after(tp->snd_sml,tp->snd_una) && 1176 !after(tp->snd_sml, tp->snd_nxt); 1177 } 1178 1179 /* Return 0, if packet can be sent now without violation Nagle's rules: 1180 * 1. It is full sized. 1181 * 2. Or it contains FIN. (already checked by caller) 1182 * 3. Or TCP_NODELAY was set. 1183 * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1184 * With Minshall's modification: all sent small packets are ACKed. 1185 */ 1186 static inline int tcp_nagle_check(const struct tcp_sock *tp, 1187 const struct sk_buff *skb, 1188 unsigned mss_now, int nonagle) 1189 { 1190 return (skb->len < mss_now && 1191 ((nonagle & TCP_NAGLE_CORK) || 1192 (!nonagle && tp->packets_out && tcp_minshall_check(tp)))); 1193 } 1194 1195 /* Return non-zero if the Nagle test allows this packet to be 1196 * sent now. 1197 */ 1198 static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, 1199 unsigned int cur_mss, int nonagle) 1200 { 1201 /* Nagle rule does not apply to frames, which sit in the middle of the 1202 * write_queue (they have no chances to get new data). 1203 * 1204 * This is implemented in the callers, where they modify the 'nonagle' 1205 * argument based upon the location of SKB in the send queue. 1206 */ 1207 if (nonagle & TCP_NAGLE_PUSH) 1208 return 1; 1209 1210 /* Don't use the nagle rule for urgent data (or for the final FIN). 1211 * Nagle can be ignored during F-RTO too (see RFC4138). 1212 */ 1213 if (tcp_urg_mode(tp) || (tp->frto_counter == 2) || 1214 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)) 1215 return 1; 1216 1217 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 1218 return 1; 1219 1220 return 0; 1221 } 1222 1223 /* Does at least the first segment of SKB fit into the send window? */ 1224 static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, 1225 unsigned int cur_mss) 1226 { 1227 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1228 1229 if (skb->len > cur_mss) 1230 end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1231 1232 return !after(end_seq, tcp_wnd_end(tp)); 1233 } 1234 1235 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1236 * should be put on the wire right now. If so, it returns the number of 1237 * packets allowed by the congestion window. 1238 */ 1239 static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, 1240 unsigned int cur_mss, int nonagle) 1241 { 1242 struct tcp_sock *tp = tcp_sk(sk); 1243 unsigned int cwnd_quota; 1244 1245 tcp_init_tso_segs(sk, skb, cur_mss); 1246 1247 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1248 return 0; 1249 1250 cwnd_quota = tcp_cwnd_test(tp, skb); 1251 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1252 cwnd_quota = 0; 1253 1254 return cwnd_quota; 1255 } 1256 1257 int tcp_may_send_now(struct sock *sk) 1258 { 1259 struct tcp_sock *tp = tcp_sk(sk); 1260 struct sk_buff *skb = tcp_send_head(sk); 1261 1262 return (skb && 1263 tcp_snd_test(sk, skb, tcp_current_mss(sk, 1), 1264 (tcp_skb_is_last(sk, skb) ? 1265 tp->nonagle : TCP_NAGLE_PUSH))); 1266 } 1267 1268 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1269 * which is put after SKB on the list. It is very much like 1270 * tcp_fragment() except that it may make several kinds of assumptions 1271 * in order to speed up the splitting operation. In particular, we 1272 * know that all the data is in scatter-gather pages, and that the 1273 * packet has never been sent out before (and thus is not cloned). 1274 */ 1275 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1276 unsigned int mss_now) 1277 { 1278 struct sk_buff *buff; 1279 int nlen = skb->len - len; 1280 u16 flags; 1281 1282 /* All of a TSO frame must be composed of paged data. */ 1283 if (skb->len != skb->data_len) 1284 return tcp_fragment(sk, skb, len, mss_now); 1285 1286 buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC); 1287 if (unlikely(buff == NULL)) 1288 return -ENOMEM; 1289 1290 sk->sk_wmem_queued += buff->truesize; 1291 sk_mem_charge(sk, buff->truesize); 1292 buff->truesize += nlen; 1293 skb->truesize -= nlen; 1294 1295 /* Correct the sequence numbers. */ 1296 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1297 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1298 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1299 1300 /* PSH and FIN should only be set in the second packet. */ 1301 flags = TCP_SKB_CB(skb)->flags; 1302 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH); 1303 TCP_SKB_CB(buff)->flags = flags; 1304 1305 /* This packet was never sent out yet, so no SACK bits. */ 1306 TCP_SKB_CB(buff)->sacked = 0; 1307 1308 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1309 skb_split(skb, buff, len); 1310 1311 /* Fix up tso_factor for both original and new SKB. */ 1312 tcp_set_skb_tso_segs(sk, skb, mss_now); 1313 tcp_set_skb_tso_segs(sk, buff, mss_now); 1314 1315 /* Link BUFF into the send queue. */ 1316 skb_header_release(buff); 1317 tcp_insert_write_queue_after(skb, buff, sk); 1318 1319 return 0; 1320 } 1321 1322 /* Try to defer sending, if possible, in order to minimize the amount 1323 * of TSO splitting we do. View it as a kind of TSO Nagle test. 1324 * 1325 * This algorithm is from John Heffner. 1326 */ 1327 static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1328 { 1329 struct tcp_sock *tp = tcp_sk(sk); 1330 const struct inet_connection_sock *icsk = inet_csk(sk); 1331 u32 send_win, cong_win, limit, in_flight; 1332 1333 if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) 1334 goto send_now; 1335 1336 if (icsk->icsk_ca_state != TCP_CA_Open) 1337 goto send_now; 1338 1339 /* Defer for less than two clock ticks. */ 1340 if (tp->tso_deferred && 1341 ((jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1342 goto send_now; 1343 1344 in_flight = tcp_packets_in_flight(tp); 1345 1346 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1347 1348 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1349 1350 /* From in_flight test above, we know that cwnd > in_flight. */ 1351 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1352 1353 limit = min(send_win, cong_win); 1354 1355 /* If a full-sized TSO skb can be sent, do it. */ 1356 if (limit >= sk->sk_gso_max_size) 1357 goto send_now; 1358 1359 if (sysctl_tcp_tso_win_divisor) { 1360 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1361 1362 /* If at least some fraction of a window is available, 1363 * just use it. 1364 */ 1365 chunk /= sysctl_tcp_tso_win_divisor; 1366 if (limit >= chunk) 1367 goto send_now; 1368 } else { 1369 /* Different approach, try not to defer past a single 1370 * ACK. Receiver should ACK every other full sized 1371 * frame, so if we have space for more than 3 frames 1372 * then send now. 1373 */ 1374 if (limit > tcp_max_burst(tp) * tp->mss_cache) 1375 goto send_now; 1376 } 1377 1378 /* Ok, it looks like it is advisable to defer. */ 1379 tp->tso_deferred = 1 | (jiffies << 1); 1380 1381 return 1; 1382 1383 send_now: 1384 tp->tso_deferred = 0; 1385 return 0; 1386 } 1387 1388 /* Create a new MTU probe if we are ready. 1389 * Returns 0 if we should wait to probe (no cwnd available), 1390 * 1 if a probe was sent, 1391 * -1 otherwise 1392 */ 1393 static int tcp_mtu_probe(struct sock *sk) 1394 { 1395 struct tcp_sock *tp = tcp_sk(sk); 1396 struct inet_connection_sock *icsk = inet_csk(sk); 1397 struct sk_buff *skb, *nskb, *next; 1398 int len; 1399 int probe_size; 1400 int size_needed; 1401 int copy; 1402 int mss_now; 1403 1404 /* Not currently probing/verifying, 1405 * not in recovery, 1406 * have enough cwnd, and 1407 * not SACKing (the variable headers throw things off) */ 1408 if (!icsk->icsk_mtup.enabled || 1409 icsk->icsk_mtup.probe_size || 1410 inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 1411 tp->snd_cwnd < 11 || 1412 tp->rx_opt.eff_sacks) 1413 return -1; 1414 1415 /* Very simple search strategy: just double the MSS. */ 1416 mss_now = tcp_current_mss(sk, 0); 1417 probe_size = 2 * tp->mss_cache; 1418 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 1419 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 1420 /* TODO: set timer for probe_converge_event */ 1421 return -1; 1422 } 1423 1424 /* Have enough data in the send queue to probe? */ 1425 if (tp->write_seq - tp->snd_nxt < size_needed) 1426 return -1; 1427 1428 if (tp->snd_wnd < size_needed) 1429 return -1; 1430 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 1431 return 0; 1432 1433 /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1434 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1435 if (!tcp_packets_in_flight(tp)) 1436 return -1; 1437 else 1438 return 0; 1439 } 1440 1441 /* We're allowed to probe. Build it now. */ 1442 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 1443 return -1; 1444 sk->sk_wmem_queued += nskb->truesize; 1445 sk_mem_charge(sk, nskb->truesize); 1446 1447 skb = tcp_send_head(sk); 1448 1449 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 1450 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 1451 TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK; 1452 TCP_SKB_CB(nskb)->sacked = 0; 1453 nskb->csum = 0; 1454 nskb->ip_summed = skb->ip_summed; 1455 1456 tcp_insert_write_queue_before(nskb, skb, sk); 1457 1458 len = 0; 1459 tcp_for_write_queue_from_safe(skb, next, sk) { 1460 copy = min_t(int, skb->len, probe_size - len); 1461 if (nskb->ip_summed) 1462 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 1463 else 1464 nskb->csum = skb_copy_and_csum_bits(skb, 0, 1465 skb_put(nskb, copy), 1466 copy, nskb->csum); 1467 1468 if (skb->len <= copy) { 1469 /* We've eaten all the data from this skb. 1470 * Throw it away. */ 1471 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags; 1472 tcp_unlink_write_queue(skb, sk); 1473 sk_wmem_free_skb(sk, skb); 1474 } else { 1475 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags & 1476 ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); 1477 if (!skb_shinfo(skb)->nr_frags) { 1478 skb_pull(skb, copy); 1479 if (skb->ip_summed != CHECKSUM_PARTIAL) 1480 skb->csum = csum_partial(skb->data, 1481 skb->len, 0); 1482 } else { 1483 __pskb_trim_head(skb, copy); 1484 tcp_set_skb_tso_segs(sk, skb, mss_now); 1485 } 1486 TCP_SKB_CB(skb)->seq += copy; 1487 } 1488 1489 len += copy; 1490 1491 if (len >= probe_size) 1492 break; 1493 } 1494 tcp_init_tso_segs(sk, nskb, nskb->len); 1495 1496 /* We're ready to send. If this fails, the probe will 1497 * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 1498 TCP_SKB_CB(nskb)->when = tcp_time_stamp; 1499 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 1500 /* Decrement cwnd here because we are sending 1501 * effectively two packets. */ 1502 tp->snd_cwnd--; 1503 tcp_event_new_data_sent(sk, nskb); 1504 1505 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 1506 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 1507 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 1508 1509 return 1; 1510 } 1511 1512 return -1; 1513 } 1514 1515 /* This routine writes packets to the network. It advances the 1516 * send_head. This happens as incoming acks open up the remote 1517 * window for us. 1518 * 1519 * Returns 1, if no segments are in flight and we have queued segments, but 1520 * cannot send anything now because of SWS or another problem. 1521 */ 1522 static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle) 1523 { 1524 struct tcp_sock *tp = tcp_sk(sk); 1525 struct sk_buff *skb; 1526 unsigned int tso_segs, sent_pkts; 1527 int cwnd_quota; 1528 int result; 1529 1530 /* If we are closed, the bytes will have to remain here. 1531 * In time closedown will finish, we empty the write queue and all 1532 * will be happy. 1533 */ 1534 if (unlikely(sk->sk_state == TCP_CLOSE)) 1535 return 0; 1536 1537 sent_pkts = 0; 1538 1539 /* Do MTU probing. */ 1540 if ((result = tcp_mtu_probe(sk)) == 0) { 1541 return 0; 1542 } else if (result > 0) { 1543 sent_pkts = 1; 1544 } 1545 1546 while ((skb = tcp_send_head(sk))) { 1547 unsigned int limit; 1548 1549 tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1550 BUG_ON(!tso_segs); 1551 1552 cwnd_quota = tcp_cwnd_test(tp, skb); 1553 if (!cwnd_quota) 1554 break; 1555 1556 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1557 break; 1558 1559 if (tso_segs == 1) { 1560 if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1561 (tcp_skb_is_last(sk, skb) ? 1562 nonagle : TCP_NAGLE_PUSH)))) 1563 break; 1564 } else { 1565 if (tcp_tso_should_defer(sk, skb)) 1566 break; 1567 } 1568 1569 limit = mss_now; 1570 if (tso_segs > 1) 1571 limit = tcp_mss_split_point(sk, skb, mss_now, 1572 cwnd_quota); 1573 1574 if (skb->len > limit && 1575 unlikely(tso_fragment(sk, skb, limit, mss_now))) 1576 break; 1577 1578 TCP_SKB_CB(skb)->when = tcp_time_stamp; 1579 1580 if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC))) 1581 break; 1582 1583 /* Advance the send_head. This one is sent out. 1584 * This call will increment packets_out. 1585 */ 1586 tcp_event_new_data_sent(sk, skb); 1587 1588 tcp_minshall_update(tp, mss_now, skb); 1589 sent_pkts++; 1590 } 1591 1592 if (likely(sent_pkts)) { 1593 tcp_cwnd_validate(sk); 1594 return 0; 1595 } 1596 return !tp->packets_out && tcp_send_head(sk); 1597 } 1598 1599 /* Push out any pending frames which were held back due to 1600 * TCP_CORK or attempt at coalescing tiny packets. 1601 * The socket must be locked by the caller. 1602 */ 1603 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 1604 int nonagle) 1605 { 1606 struct sk_buff *skb = tcp_send_head(sk); 1607 1608 if (skb) { 1609 if (tcp_write_xmit(sk, cur_mss, nonagle)) 1610 tcp_check_probe_timer(sk); 1611 } 1612 } 1613 1614 /* Send _single_ skb sitting at the send head. This function requires 1615 * true push pending frames to setup probe timer etc. 1616 */ 1617 void tcp_push_one(struct sock *sk, unsigned int mss_now) 1618 { 1619 struct sk_buff *skb = tcp_send_head(sk); 1620 unsigned int tso_segs, cwnd_quota; 1621 1622 BUG_ON(!skb || skb->len < mss_now); 1623 1624 tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1625 cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH); 1626 1627 if (likely(cwnd_quota)) { 1628 unsigned int limit; 1629 1630 BUG_ON(!tso_segs); 1631 1632 limit = mss_now; 1633 if (tso_segs > 1) 1634 limit = tcp_mss_split_point(sk, skb, mss_now, 1635 cwnd_quota); 1636 1637 if (skb->len > limit && 1638 unlikely(tso_fragment(sk, skb, limit, mss_now))) 1639 return; 1640 1641 /* Send it out now. */ 1642 TCP_SKB_CB(skb)->when = tcp_time_stamp; 1643 1644 if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) { 1645 tcp_event_new_data_sent(sk, skb); 1646 tcp_cwnd_validate(sk); 1647 return; 1648 } 1649 } 1650 } 1651 1652 /* This function returns the amount that we can raise the 1653 * usable window based on the following constraints 1654 * 1655 * 1. The window can never be shrunk once it is offered (RFC 793) 1656 * 2. We limit memory per socket 1657 * 1658 * RFC 1122: 1659 * "the suggested [SWS] avoidance algorithm for the receiver is to keep 1660 * RECV.NEXT + RCV.WIN fixed until: 1661 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 1662 * 1663 * i.e. don't raise the right edge of the window until you can raise 1664 * it at least MSS bytes. 1665 * 1666 * Unfortunately, the recommended algorithm breaks header prediction, 1667 * since header prediction assumes th->window stays fixed. 1668 * 1669 * Strictly speaking, keeping th->window fixed violates the receiver 1670 * side SWS prevention criteria. The problem is that under this rule 1671 * a stream of single byte packets will cause the right side of the 1672 * window to always advance by a single byte. 1673 * 1674 * Of course, if the sender implements sender side SWS prevention 1675 * then this will not be a problem. 1676 * 1677 * BSD seems to make the following compromise: 1678 * 1679 * If the free space is less than the 1/4 of the maximum 1680 * space available and the free space is less than 1/2 mss, 1681 * then set the window to 0. 1682 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 1683 * Otherwise, just prevent the window from shrinking 1684 * and from being larger than the largest representable value. 1685 * 1686 * This prevents incremental opening of the window in the regime 1687 * where TCP is limited by the speed of the reader side taking 1688 * data out of the TCP receive queue. It does nothing about 1689 * those cases where the window is constrained on the sender side 1690 * because the pipeline is full. 1691 * 1692 * BSD also seems to "accidentally" limit itself to windows that are a 1693 * multiple of MSS, at least until the free space gets quite small. 1694 * This would appear to be a side effect of the mbuf implementation. 1695 * Combining these two algorithms results in the observed behavior 1696 * of having a fixed window size at almost all times. 1697 * 1698 * Below we obtain similar behavior by forcing the offered window to 1699 * a multiple of the mss when it is feasible to do so. 1700 * 1701 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 1702 * Regular options like TIMESTAMP are taken into account. 1703 */ 1704 u32 __tcp_select_window(struct sock *sk) 1705 { 1706 struct inet_connection_sock *icsk = inet_csk(sk); 1707 struct tcp_sock *tp = tcp_sk(sk); 1708 /* MSS for the peer's data. Previous versions used mss_clamp 1709 * here. I don't know if the value based on our guesses 1710 * of peer's MSS is better for the performance. It's more correct 1711 * but may be worse for the performance because of rcv_mss 1712 * fluctuations. --SAW 1998/11/1 1713 */ 1714 int mss = icsk->icsk_ack.rcv_mss; 1715 int free_space = tcp_space(sk); 1716 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 1717 int window; 1718 1719 if (mss > full_space) 1720 mss = full_space; 1721 1722 if (free_space < (full_space >> 1)) { 1723 icsk->icsk_ack.quick = 0; 1724 1725 if (tcp_memory_pressure) 1726 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 1727 4U * tp->advmss); 1728 1729 if (free_space < mss) 1730 return 0; 1731 } 1732 1733 if (free_space > tp->rcv_ssthresh) 1734 free_space = tp->rcv_ssthresh; 1735 1736 /* Don't do rounding if we are using window scaling, since the 1737 * scaled window will not line up with the MSS boundary anyway. 1738 */ 1739 window = tp->rcv_wnd; 1740 if (tp->rx_opt.rcv_wscale) { 1741 window = free_space; 1742 1743 /* Advertise enough space so that it won't get scaled away. 1744 * Import case: prevent zero window announcement if 1745 * 1<<rcv_wscale > mss. 1746 */ 1747 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 1748 window = (((window >> tp->rx_opt.rcv_wscale) + 1) 1749 << tp->rx_opt.rcv_wscale); 1750 } else { 1751 /* Get the largest window that is a nice multiple of mss. 1752 * Window clamp already applied above. 1753 * If our current window offering is within 1 mss of the 1754 * free space we just keep it. This prevents the divide 1755 * and multiply from happening most of the time. 1756 * We also don't do any window rounding when the free space 1757 * is too small. 1758 */ 1759 if (window <= free_space - mss || window > free_space) 1760 window = (free_space / mss) * mss; 1761 else if (mss == full_space && 1762 free_space > window + (full_space >> 1)) 1763 window = free_space; 1764 } 1765 1766 return window; 1767 } 1768 1769 /* Attempt to collapse two adjacent SKB's during retransmission. */ 1770 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, 1771 int mss_now) 1772 { 1773 struct tcp_sock *tp = tcp_sk(sk); 1774 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 1775 int skb_size, next_skb_size; 1776 u16 flags; 1777 1778 /* The first test we must make is that neither of these two 1779 * SKB's are still referenced by someone else. 1780 */ 1781 if (skb_cloned(skb) || skb_cloned(next_skb)) 1782 return; 1783 1784 skb_size = skb->len; 1785 next_skb_size = next_skb->len; 1786 flags = TCP_SKB_CB(skb)->flags; 1787 1788 /* Also punt if next skb has been SACK'd. */ 1789 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED) 1790 return; 1791 1792 /* Next skb is out of window. */ 1793 if (after(TCP_SKB_CB(next_skb)->end_seq, tcp_wnd_end(tp))) 1794 return; 1795 1796 /* Punt if not enough space exists in the first SKB for 1797 * the data in the second, or the total combined payload 1798 * would exceed the MSS. 1799 */ 1800 if ((next_skb_size > skb_tailroom(skb)) || 1801 ((skb_size + next_skb_size) > mss_now)) 1802 return; 1803 1804 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 1805 1806 tcp_highest_sack_combine(sk, next_skb, skb); 1807 1808 /* Ok. We will be able to collapse the packet. */ 1809 tcp_unlink_write_queue(next_skb, sk); 1810 1811 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 1812 next_skb_size); 1813 1814 if (next_skb->ip_summed == CHECKSUM_PARTIAL) 1815 skb->ip_summed = CHECKSUM_PARTIAL; 1816 1817 if (skb->ip_summed != CHECKSUM_PARTIAL) 1818 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 1819 1820 /* Update sequence range on original skb. */ 1821 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 1822 1823 /* Merge over control information. */ 1824 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */ 1825 TCP_SKB_CB(skb)->flags = flags; 1826 1827 /* All done, get rid of second SKB and account for it so 1828 * packet counting does not break. 1829 */ 1830 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 1831 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_RETRANS) 1832 tp->retrans_out -= tcp_skb_pcount(next_skb); 1833 if (TCP_SKB_CB(next_skb)->sacked & TCPCB_LOST) 1834 tp->lost_out -= tcp_skb_pcount(next_skb); 1835 /* Reno case is special. Sigh... */ 1836 if (tcp_is_reno(tp) && tp->sacked_out) 1837 tcp_dec_pcount_approx(&tp->sacked_out, next_skb); 1838 1839 tcp_adjust_fackets_out(sk, next_skb, tcp_skb_pcount(next_skb)); 1840 tp->packets_out -= tcp_skb_pcount(next_skb); 1841 1842 /* changed transmit queue under us so clear hints */ 1843 tcp_clear_retrans_hints_partial(tp); 1844 if (next_skb == tp->retransmit_skb_hint) 1845 tp->retransmit_skb_hint = skb; 1846 1847 sk_wmem_free_skb(sk, next_skb); 1848 } 1849 1850 /* Do a simple retransmit without using the backoff mechanisms in 1851 * tcp_timer. This is used for path mtu discovery. 1852 * The socket is already locked here. 1853 */ 1854 void tcp_simple_retransmit(struct sock *sk) 1855 { 1856 const struct inet_connection_sock *icsk = inet_csk(sk); 1857 struct tcp_sock *tp = tcp_sk(sk); 1858 struct sk_buff *skb; 1859 unsigned int mss = tcp_current_mss(sk, 0); 1860 u32 prior_lost = tp->lost_out; 1861 1862 tcp_for_write_queue(skb, sk) { 1863 if (skb == tcp_send_head(sk)) 1864 break; 1865 if (skb->len > mss && 1866 !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 1867 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 1868 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 1869 tp->retrans_out -= tcp_skb_pcount(skb); 1870 } 1871 tcp_skb_mark_lost_uncond_verify(tp, skb); 1872 } 1873 } 1874 1875 tcp_clear_retrans_hints_partial(tp); 1876 1877 if (prior_lost == tp->lost_out) 1878 return; 1879 1880 if (tcp_is_reno(tp)) 1881 tcp_limit_reno_sacked(tp); 1882 1883 tcp_verify_left_out(tp); 1884 1885 /* Don't muck with the congestion window here. 1886 * Reason is that we do not increase amount of _data_ 1887 * in network, but units changed and effective 1888 * cwnd/ssthresh really reduced now. 1889 */ 1890 if (icsk->icsk_ca_state != TCP_CA_Loss) { 1891 tp->high_seq = tp->snd_nxt; 1892 tp->snd_ssthresh = tcp_current_ssthresh(sk); 1893 tp->prior_ssthresh = 0; 1894 tp->undo_marker = 0; 1895 tcp_set_ca_state(sk, TCP_CA_Loss); 1896 } 1897 tcp_xmit_retransmit_queue(sk); 1898 } 1899 1900 /* This retransmits one SKB. Policy decisions and retransmit queue 1901 * state updates are done by the caller. Returns non-zero if an 1902 * error occurred which prevented the send. 1903 */ 1904 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 1905 { 1906 struct tcp_sock *tp = tcp_sk(sk); 1907 struct inet_connection_sock *icsk = inet_csk(sk); 1908 unsigned int cur_mss; 1909 int err; 1910 1911 /* Inconslusive MTU probe */ 1912 if (icsk->icsk_mtup.probe_size) { 1913 icsk->icsk_mtup.probe_size = 0; 1914 } 1915 1916 /* Do not sent more than we queued. 1/4 is reserved for possible 1917 * copying overhead: fragmentation, tunneling, mangling etc. 1918 */ 1919 if (atomic_read(&sk->sk_wmem_alloc) > 1920 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 1921 return -EAGAIN; 1922 1923 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 1924 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1925 BUG(); 1926 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 1927 return -ENOMEM; 1928 } 1929 1930 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 1931 return -EHOSTUNREACH; /* Routing failure or similar. */ 1932 1933 cur_mss = tcp_current_mss(sk, 0); 1934 1935 /* If receiver has shrunk his window, and skb is out of 1936 * new window, do not retransmit it. The exception is the 1937 * case, when window is shrunk to zero. In this case 1938 * our retransmit serves as a zero window probe. 1939 */ 1940 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) 1941 && TCP_SKB_CB(skb)->seq != tp->snd_una) 1942 return -EAGAIN; 1943 1944 if (skb->len > cur_mss) { 1945 if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 1946 return -ENOMEM; /* We'll try again later. */ 1947 } 1948 1949 /* Collapse two adjacent packets if worthwhile and we can. */ 1950 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) && 1951 (skb->len < (cur_mss >> 1)) && 1952 (!tcp_skb_is_last(sk, skb)) && 1953 (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) && 1954 (skb_shinfo(skb)->nr_frags == 0 && 1955 skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) && 1956 (tcp_skb_pcount(skb) == 1 && 1957 tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) && 1958 (sysctl_tcp_retrans_collapse != 0)) 1959 tcp_retrans_try_collapse(sk, skb, cur_mss); 1960 1961 /* Some Solaris stacks overoptimize and ignore the FIN on a 1962 * retransmit when old data is attached. So strip it off 1963 * since it is cheap to do so and saves bytes on the network. 1964 */ 1965 if (skb->len > 0 && 1966 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && 1967 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 1968 if (!pskb_trim(skb, 0)) { 1969 /* Reuse, even though it does some unnecessary work */ 1970 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, 1971 TCP_SKB_CB(skb)->flags); 1972 skb->ip_summed = CHECKSUM_NONE; 1973 } 1974 } 1975 1976 /* Make a copy, if the first transmission SKB clone we made 1977 * is still in somebody's hands, else make a clone. 1978 */ 1979 TCP_SKB_CB(skb)->when = tcp_time_stamp; 1980 1981 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 1982 1983 if (err == 0) { 1984 /* Update global TCP statistics. */ 1985 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 1986 1987 tp->total_retrans++; 1988 1989 #if FASTRETRANS_DEBUG > 0 1990 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 1991 if (net_ratelimit()) 1992 printk(KERN_DEBUG "retrans_out leaked.\n"); 1993 } 1994 #endif 1995 if (!tp->retrans_out) 1996 tp->lost_retrans_low = tp->snd_nxt; 1997 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 1998 tp->retrans_out += tcp_skb_pcount(skb); 1999 2000 /* Save stamp of the first retransmit. */ 2001 if (!tp->retrans_stamp) 2002 tp->retrans_stamp = TCP_SKB_CB(skb)->when; 2003 2004 tp->undo_retrans++; 2005 2006 /* snd_nxt is stored to detect loss of retransmitted segment, 2007 * see tcp_input.c tcp_sacktag_write_queue(). 2008 */ 2009 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 2010 } 2011 return err; 2012 } 2013 2014 static int tcp_can_forward_retransmit(struct sock *sk) 2015 { 2016 const struct inet_connection_sock *icsk = inet_csk(sk); 2017 struct tcp_sock *tp = tcp_sk(sk); 2018 2019 /* Forward retransmissions are possible only during Recovery. */ 2020 if (icsk->icsk_ca_state != TCP_CA_Recovery) 2021 return 0; 2022 2023 /* No forward retransmissions in Reno are possible. */ 2024 if (tcp_is_reno(tp)) 2025 return 0; 2026 2027 /* Yeah, we have to make difficult choice between forward transmission 2028 * and retransmission... Both ways have their merits... 2029 * 2030 * For now we do not retransmit anything, while we have some new 2031 * segments to send. In the other cases, follow rule 3 for 2032 * NextSeg() specified in RFC3517. 2033 */ 2034 2035 if (tcp_may_send_now(sk)) 2036 return 0; 2037 2038 return 1; 2039 } 2040 2041 /* This gets called after a retransmit timeout, and the initially 2042 * retransmitted data is acknowledged. It tries to continue 2043 * resending the rest of the retransmit queue, until either 2044 * we've sent it all or the congestion window limit is reached. 2045 * If doing SACK, the first ACK which comes back for a timeout 2046 * based retransmit packet might feed us FACK information again. 2047 * If so, we use it to avoid unnecessarily retransmissions. 2048 */ 2049 void tcp_xmit_retransmit_queue(struct sock *sk) 2050 { 2051 const struct inet_connection_sock *icsk = inet_csk(sk); 2052 struct tcp_sock *tp = tcp_sk(sk); 2053 struct sk_buff *skb; 2054 struct sk_buff *hole = NULL; 2055 u32 last_lost; 2056 int mib_idx; 2057 int fwd_rexmitting = 0; 2058 2059 if (!tp->lost_out) 2060 tp->retransmit_high = tp->snd_una; 2061 2062 if (tp->retransmit_skb_hint) { 2063 skb = tp->retransmit_skb_hint; 2064 last_lost = TCP_SKB_CB(skb)->end_seq; 2065 if (after(last_lost, tp->retransmit_high)) 2066 last_lost = tp->retransmit_high; 2067 } else { 2068 skb = tcp_write_queue_head(sk); 2069 last_lost = tp->snd_una; 2070 } 2071 2072 /* First pass: retransmit lost packets. */ 2073 tcp_for_write_queue_from(skb, sk) { 2074 __u8 sacked = TCP_SKB_CB(skb)->sacked; 2075 2076 if (skb == tcp_send_head(sk)) 2077 break; 2078 /* we could do better than to assign each time */ 2079 if (hole == NULL) 2080 tp->retransmit_skb_hint = skb; 2081 2082 /* Assume this retransmit will generate 2083 * only one packet for congestion window 2084 * calculation purposes. This works because 2085 * tcp_retransmit_skb() will chop up the 2086 * packet to be MSS sized and all the 2087 * packet counting works out. 2088 */ 2089 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 2090 return; 2091 2092 if (fwd_rexmitting) { 2093 begin_fwd: 2094 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2095 break; 2096 mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 2097 2098 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2099 tp->retransmit_high = last_lost; 2100 if (!tcp_can_forward_retransmit(sk)) 2101 break; 2102 /* Backtrack if necessary to non-L'ed skb */ 2103 if (hole != NULL) { 2104 skb = hole; 2105 hole = NULL; 2106 } 2107 fwd_rexmitting = 1; 2108 goto begin_fwd; 2109 2110 } else if (!(sacked & TCPCB_LOST)) { 2111 if (hole == NULL && !(sacked & TCPCB_SACKED_RETRANS)) 2112 hole = skb; 2113 continue; 2114 2115 } else { 2116 last_lost = TCP_SKB_CB(skb)->end_seq; 2117 if (icsk->icsk_ca_state != TCP_CA_Loss) 2118 mib_idx = LINUX_MIB_TCPFASTRETRANS; 2119 else 2120 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 2121 } 2122 2123 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 2124 continue; 2125 2126 if (tcp_retransmit_skb(sk, skb)) 2127 return; 2128 NET_INC_STATS_BH(sock_net(sk), mib_idx); 2129 2130 if (skb == tcp_write_queue_head(sk)) 2131 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2132 inet_csk(sk)->icsk_rto, 2133 TCP_RTO_MAX); 2134 } 2135 } 2136 2137 /* Send a fin. The caller locks the socket for us. This cannot be 2138 * allowed to fail queueing a FIN frame under any circumstances. 2139 */ 2140 void tcp_send_fin(struct sock *sk) 2141 { 2142 struct tcp_sock *tp = tcp_sk(sk); 2143 struct sk_buff *skb = tcp_write_queue_tail(sk); 2144 int mss_now; 2145 2146 /* Optimization, tack on the FIN if we have a queue of 2147 * unsent frames. But be careful about outgoing SACKS 2148 * and IP options. 2149 */ 2150 mss_now = tcp_current_mss(sk, 1); 2151 2152 if (tcp_send_head(sk) != NULL) { 2153 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN; 2154 TCP_SKB_CB(skb)->end_seq++; 2155 tp->write_seq++; 2156 } else { 2157 /* Socket is locked, keep trying until memory is available. */ 2158 for (;;) { 2159 skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL); 2160 if (skb) 2161 break; 2162 yield(); 2163 } 2164 2165 /* Reserve space for headers and prepare control bits. */ 2166 skb_reserve(skb, MAX_TCP_HEADER); 2167 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2168 tcp_init_nondata_skb(skb, tp->write_seq, 2169 TCPCB_FLAG_ACK | TCPCB_FLAG_FIN); 2170 tcp_queue_skb(sk, skb); 2171 } 2172 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 2173 } 2174 2175 /* We get here when a process closes a file descriptor (either due to 2176 * an explicit close() or as a byproduct of exit()'ing) and there 2177 * was unread data in the receive queue. This behavior is recommended 2178 * by RFC 2525, section 2.17. -DaveM 2179 */ 2180 void tcp_send_active_reset(struct sock *sk, gfp_t priority) 2181 { 2182 struct sk_buff *skb; 2183 2184 /* NOTE: No TCP options attached and we never retransmit this. */ 2185 skb = alloc_skb(MAX_TCP_HEADER, priority); 2186 if (!skb) { 2187 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 2188 return; 2189 } 2190 2191 /* Reserve space for headers and prepare control bits. */ 2192 skb_reserve(skb, MAX_TCP_HEADER); 2193 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2194 TCPCB_FLAG_ACK | TCPCB_FLAG_RST); 2195 /* Send it off. */ 2196 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2197 if (tcp_transmit_skb(sk, skb, 0, priority)) 2198 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 2199 2200 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 2201 } 2202 2203 /* WARNING: This routine must only be called when we have already sent 2204 * a SYN packet that crossed the incoming SYN that caused this routine 2205 * to get called. If this assumption fails then the initial rcv_wnd 2206 * and rcv_wscale values will not be correct. 2207 */ 2208 int tcp_send_synack(struct sock *sk) 2209 { 2210 struct sk_buff *skb; 2211 2212 skb = tcp_write_queue_head(sk); 2213 if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) { 2214 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); 2215 return -EFAULT; 2216 } 2217 if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) { 2218 if (skb_cloned(skb)) { 2219 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 2220 if (nskb == NULL) 2221 return -ENOMEM; 2222 tcp_unlink_write_queue(skb, sk); 2223 skb_header_release(nskb); 2224 __tcp_add_write_queue_head(sk, nskb); 2225 sk_wmem_free_skb(sk, skb); 2226 sk->sk_wmem_queued += nskb->truesize; 2227 sk_mem_charge(sk, nskb->truesize); 2228 skb = nskb; 2229 } 2230 2231 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK; 2232 TCP_ECN_send_synack(tcp_sk(sk), skb); 2233 } 2234 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2235 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2236 } 2237 2238 /* 2239 * Prepare a SYN-ACK. 2240 */ 2241 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2242 struct request_sock *req) 2243 { 2244 struct inet_request_sock *ireq = inet_rsk(req); 2245 struct tcp_sock *tp = tcp_sk(sk); 2246 struct tcphdr *th; 2247 int tcp_header_size; 2248 struct tcp_out_options opts; 2249 struct sk_buff *skb; 2250 struct tcp_md5sig_key *md5; 2251 __u8 *md5_hash_location; 2252 int mss; 2253 2254 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); 2255 if (skb == NULL) 2256 return NULL; 2257 2258 /* Reserve space for headers. */ 2259 skb_reserve(skb, MAX_TCP_HEADER); 2260 2261 skb->dst = dst_clone(dst); 2262 2263 mss = dst_metric(dst, RTAX_ADVMSS); 2264 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2265 mss = tp->rx_opt.user_mss; 2266 2267 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 2268 __u8 rcv_wscale; 2269 /* Set this up on the first call only */ 2270 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 2271 /* tcp_full_space because it is guaranteed to be the first packet */ 2272 tcp_select_initial_window(tcp_full_space(sk), 2273 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 2274 &req->rcv_wnd, 2275 &req->window_clamp, 2276 ireq->wscale_ok, 2277 &rcv_wscale); 2278 ireq->rcv_wscale = rcv_wscale; 2279 } 2280 2281 memset(&opts, 0, sizeof(opts)); 2282 #ifdef CONFIG_SYN_COOKIES 2283 if (unlikely(req->cookie_ts)) 2284 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 2285 else 2286 #endif 2287 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2288 tcp_header_size = tcp_synack_options(sk, req, mss, 2289 skb, &opts, &md5) + 2290 sizeof(struct tcphdr); 2291 2292 skb_push(skb, tcp_header_size); 2293 skb_reset_transport_header(skb); 2294 2295 th = tcp_hdr(skb); 2296 memset(th, 0, sizeof(struct tcphdr)); 2297 th->syn = 1; 2298 th->ack = 1; 2299 TCP_ECN_make_synack(req, th); 2300 th->source = ireq->loc_port; 2301 th->dest = ireq->rmt_port; 2302 /* Setting of flags are superfluous here for callers (and ECE is 2303 * not even correctly set) 2304 */ 2305 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2306 TCPCB_FLAG_SYN | TCPCB_FLAG_ACK); 2307 th->seq = htonl(TCP_SKB_CB(skb)->seq); 2308 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); 2309 2310 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2311 th->window = htons(min(req->rcv_wnd, 65535U)); 2312 tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location); 2313 th->doff = (tcp_header_size >> 2); 2314 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 2315 2316 #ifdef CONFIG_TCP_MD5SIG 2317 /* Okay, we have all we need - do the md5 hash if needed */ 2318 if (md5) { 2319 tp->af_specific->calc_md5_hash(md5_hash_location, 2320 md5, NULL, req, skb); 2321 } 2322 #endif 2323 2324 return skb; 2325 } 2326 2327 /* 2328 * Do all connect socket setups that can be done AF independent. 2329 */ 2330 static void tcp_connect_init(struct sock *sk) 2331 { 2332 struct dst_entry *dst = __sk_dst_get(sk); 2333 struct tcp_sock *tp = tcp_sk(sk); 2334 __u8 rcv_wscale; 2335 2336 /* We'll fix this up when we get a response from the other end. 2337 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 2338 */ 2339 tp->tcp_header_len = sizeof(struct tcphdr) + 2340 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 2341 2342 #ifdef CONFIG_TCP_MD5SIG 2343 if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2344 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2345 #endif 2346 2347 /* If user gave his TCP_MAXSEG, record it to clamp */ 2348 if (tp->rx_opt.user_mss) 2349 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 2350 tp->max_window = 0; 2351 tcp_mtup_init(sk); 2352 tcp_sync_mss(sk, dst_mtu(dst)); 2353 2354 if (!tp->window_clamp) 2355 tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 2356 tp->advmss = dst_metric(dst, RTAX_ADVMSS); 2357 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2358 tp->advmss = tp->rx_opt.user_mss; 2359 2360 tcp_initialize_rcv_mss(sk); 2361 2362 tcp_select_initial_window(tcp_full_space(sk), 2363 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 2364 &tp->rcv_wnd, 2365 &tp->window_clamp, 2366 sysctl_tcp_window_scaling, 2367 &rcv_wscale); 2368 2369 tp->rx_opt.rcv_wscale = rcv_wscale; 2370 tp->rcv_ssthresh = tp->rcv_wnd; 2371 2372 sk->sk_err = 0; 2373 sock_reset_flag(sk, SOCK_DONE); 2374 tp->snd_wnd = 0; 2375 tcp_init_wl(tp, tp->write_seq, 0); 2376 tp->snd_una = tp->write_seq; 2377 tp->snd_sml = tp->write_seq; 2378 tp->snd_up = tp->write_seq; 2379 tp->rcv_nxt = 0; 2380 tp->rcv_wup = 0; 2381 tp->copied_seq = 0; 2382 2383 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2384 inet_csk(sk)->icsk_retransmits = 0; 2385 tcp_clear_retrans(tp); 2386 } 2387 2388 /* 2389 * Build a SYN and send it off. 2390 */ 2391 int tcp_connect(struct sock *sk) 2392 { 2393 struct tcp_sock *tp = tcp_sk(sk); 2394 struct sk_buff *buff; 2395 2396 tcp_connect_init(sk); 2397 2398 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 2399 if (unlikely(buff == NULL)) 2400 return -ENOBUFS; 2401 2402 /* Reserve space for headers. */ 2403 skb_reserve(buff, MAX_TCP_HEADER); 2404 2405 tp->snd_nxt = tp->write_seq; 2406 tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN); 2407 TCP_ECN_send_syn(sk, buff); 2408 2409 /* Send it off. */ 2410 TCP_SKB_CB(buff)->when = tcp_time_stamp; 2411 tp->retrans_stamp = TCP_SKB_CB(buff)->when; 2412 skb_header_release(buff); 2413 __tcp_add_write_queue_tail(sk, buff); 2414 sk->sk_wmem_queued += buff->truesize; 2415 sk_mem_charge(sk, buff->truesize); 2416 tp->packets_out += tcp_skb_pcount(buff); 2417 tcp_transmit_skb(sk, buff, 1, GFP_KERNEL); 2418 2419 /* We change tp->snd_nxt after the tcp_transmit_skb() call 2420 * in order to make this packet get counted in tcpOutSegs. 2421 */ 2422 tp->snd_nxt = tp->write_seq; 2423 tp->pushed_seq = tp->write_seq; 2424 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 2425 2426 /* Timer for repeating the SYN until an answer. */ 2427 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2428 inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 2429 return 0; 2430 } 2431 2432 /* Send out a delayed ack, the caller does the policy checking 2433 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 2434 * for details. 2435 */ 2436 void tcp_send_delayed_ack(struct sock *sk) 2437 { 2438 struct inet_connection_sock *icsk = inet_csk(sk); 2439 int ato = icsk->icsk_ack.ato; 2440 unsigned long timeout; 2441 2442 if (ato > TCP_DELACK_MIN) { 2443 const struct tcp_sock *tp = tcp_sk(sk); 2444 int max_ato = HZ / 2; 2445 2446 if (icsk->icsk_ack.pingpong || 2447 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 2448 max_ato = TCP_DELACK_MAX; 2449 2450 /* Slow path, intersegment interval is "high". */ 2451 2452 /* If some rtt estimate is known, use it to bound delayed ack. 2453 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 2454 * directly. 2455 */ 2456 if (tp->srtt) { 2457 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); 2458 2459 if (rtt < max_ato) 2460 max_ato = rtt; 2461 } 2462 2463 ato = min(ato, max_ato); 2464 } 2465 2466 /* Stay within the limit we were given */ 2467 timeout = jiffies + ato; 2468 2469 /* Use new timeout only if there wasn't a older one earlier. */ 2470 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 2471 /* If delack timer was blocked or is about to expire, 2472 * send ACK now. 2473 */ 2474 if (icsk->icsk_ack.blocked || 2475 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 2476 tcp_send_ack(sk); 2477 return; 2478 } 2479 2480 if (!time_before(timeout, icsk->icsk_ack.timeout)) 2481 timeout = icsk->icsk_ack.timeout; 2482 } 2483 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 2484 icsk->icsk_ack.timeout = timeout; 2485 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 2486 } 2487 2488 /* This routine sends an ack and also updates the window. */ 2489 void tcp_send_ack(struct sock *sk) 2490 { 2491 struct sk_buff *buff; 2492 2493 /* If we have been reset, we may not send again. */ 2494 if (sk->sk_state == TCP_CLOSE) 2495 return; 2496 2497 /* We are not putting this on the write queue, so 2498 * tcp_transmit_skb() will set the ownership to this 2499 * sock. 2500 */ 2501 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 2502 if (buff == NULL) { 2503 inet_csk_schedule_ack(sk); 2504 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 2505 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 2506 TCP_DELACK_MAX, TCP_RTO_MAX); 2507 return; 2508 } 2509 2510 /* Reserve space for headers and prepare control bits. */ 2511 skb_reserve(buff, MAX_TCP_HEADER); 2512 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK); 2513 2514 /* Send it off, this clears delayed acks for us. */ 2515 TCP_SKB_CB(buff)->when = tcp_time_stamp; 2516 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); 2517 } 2518 2519 /* This routine sends a packet with an out of date sequence 2520 * number. It assumes the other end will try to ack it. 2521 * 2522 * Question: what should we make while urgent mode? 2523 * 4.4BSD forces sending single byte of data. We cannot send 2524 * out of window data, because we have SND.NXT==SND.MAX... 2525 * 2526 * Current solution: to send TWO zero-length segments in urgent mode: 2527 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 2528 * out-of-date with SND.UNA-1 to probe window. 2529 */ 2530 static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 2531 { 2532 struct tcp_sock *tp = tcp_sk(sk); 2533 struct sk_buff *skb; 2534 2535 /* We don't queue it, tcp_transmit_skb() sets ownership. */ 2536 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 2537 if (skb == NULL) 2538 return -1; 2539 2540 /* Reserve space for headers and set control bits. */ 2541 skb_reserve(skb, MAX_TCP_HEADER); 2542 /* Use a previous sequence. This should cause the other 2543 * end to send an ack. Don't queue or clone SKB, just 2544 * send it. 2545 */ 2546 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK); 2547 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2548 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 2549 } 2550 2551 int tcp_write_wakeup(struct sock *sk) 2552 { 2553 struct tcp_sock *tp = tcp_sk(sk); 2554 struct sk_buff *skb; 2555 2556 if (sk->sk_state == TCP_CLOSE) 2557 return -1; 2558 2559 if ((skb = tcp_send_head(sk)) != NULL && 2560 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 2561 int err; 2562 unsigned int mss = tcp_current_mss(sk, 0); 2563 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 2564 2565 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 2566 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 2567 2568 /* We are probing the opening of a window 2569 * but the window size is != 0 2570 * must have been a result SWS avoidance ( sender ) 2571 */ 2572 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 2573 skb->len > mss) { 2574 seg_size = min(seg_size, mss); 2575 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 2576 if (tcp_fragment(sk, skb, seg_size, mss)) 2577 return -1; 2578 } else if (!tcp_skb_pcount(skb)) 2579 tcp_set_skb_tso_segs(sk, skb, mss); 2580 2581 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 2582 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2583 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2584 if (!err) 2585 tcp_event_new_data_sent(sk, skb); 2586 return err; 2587 } else { 2588 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 2589 tcp_xmit_probe_skb(sk, 1); 2590 return tcp_xmit_probe_skb(sk, 0); 2591 } 2592 } 2593 2594 /* A window probe timeout has occurred. If window is not closed send 2595 * a partial packet else a zero probe. 2596 */ 2597 void tcp_send_probe0(struct sock *sk) 2598 { 2599 struct inet_connection_sock *icsk = inet_csk(sk); 2600 struct tcp_sock *tp = tcp_sk(sk); 2601 int err; 2602 2603 err = tcp_write_wakeup(sk); 2604 2605 if (tp->packets_out || !tcp_send_head(sk)) { 2606 /* Cancel probe timer, if it is not required. */ 2607 icsk->icsk_probes_out = 0; 2608 icsk->icsk_backoff = 0; 2609 return; 2610 } 2611 2612 if (err <= 0) { 2613 if (icsk->icsk_backoff < sysctl_tcp_retries2) 2614 icsk->icsk_backoff++; 2615 icsk->icsk_probes_out++; 2616 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2617 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 2618 TCP_RTO_MAX); 2619 } else { 2620 /* If packet was not sent due to local congestion, 2621 * do not backoff and do not remember icsk_probes_out. 2622 * Let local senders to fight for local resources. 2623 * 2624 * Use accumulated backoff yet. 2625 */ 2626 if (!icsk->icsk_probes_out) 2627 icsk->icsk_probes_out = 1; 2628 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2629 min(icsk->icsk_rto << icsk->icsk_backoff, 2630 TCP_RESOURCE_PROBE_INTERVAL), 2631 TCP_RTO_MAX); 2632 } 2633 } 2634 2635 EXPORT_SYMBOL(tcp_select_initial_window); 2636 EXPORT_SYMBOL(tcp_connect); 2637 EXPORT_SYMBOL(tcp_make_synack); 2638 EXPORT_SYMBOL(tcp_simple_retransmit); 2639 EXPORT_SYMBOL(tcp_sync_mss); 2640 EXPORT_SYMBOL(tcp_mtup_init); 2641