1 /* 2 * INET An implementation of the TCP/IP protocol suite for the LINUX 3 * operating system. INET is implemented using the BSD Socket 4 * interface as the means of communication with the user level. 5 * 6 * Implementation of the Transmission Control Protocol(TCP). 7 * 8 * Authors: Ross Biro 9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 10 * Mark Evans, <evansmp@uhura.aston.ac.uk> 11 * Corey Minyard <wf-rch!minyard@relay.EU.net> 12 * Florian La Roche, <flla@stud.uni-sb.de> 13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 14 * Linus Torvalds, <torvalds@cs.helsinki.fi> 15 * Alan Cox, <gw4pts@gw4pts.ampr.org> 16 * Matthew Dillon, <dillon@apollo.west.oic.com> 17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 18 * Jorge Cwik, <jorge@laser.satlink.net> 19 */ 20 21 /* 22 * Changes: Pedro Roque : Retransmit queue handled by TCP. 23 * : Fragmentation on mtu decrease 24 * : Segment collapse on retransmit 25 * : AF independence 26 * 27 * Linus Torvalds : send_delayed_ack 28 * David S. Miller : Charge memory using the right skb 29 * during syn/ack processing. 30 * David S. Miller : Output engine completely rewritten. 31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. 32 * Cacophonix Gaul : draft-minshall-nagle-01 33 * J Hadi Salim : ECN support 34 * 35 */ 36 37 #define pr_fmt(fmt) "TCP: " fmt 38 39 #include <net/tcp.h> 40 41 #include <linux/compiler.h> 42 #include <linux/gfp.h> 43 #include <linux/module.h> 44 45 /* People can turn this off for buggy TCP's found in printers etc. */ 46 int sysctl_tcp_retrans_collapse __read_mostly = 1; 47 48 /* People can turn this on to work with those rare, broken TCPs that 49 * interpret the window field as a signed quantity. 50 */ 51 int sysctl_tcp_workaround_signed_windows __read_mostly = 0; 52 53 /* Default TSQ limit of two TSO segments */ 54 int sysctl_tcp_limit_output_bytes __read_mostly = 131072; 55 56 /* This limits the percentage of the congestion window which we 57 * will allow a single TSO frame to consume. Building TSO frames 58 * which are too large can cause TCP streams to be bursty. 59 */ 60 int sysctl_tcp_tso_win_divisor __read_mostly = 3; 61 62 int sysctl_tcp_mtu_probing __read_mostly = 0; 63 int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS; 64 65 /* By default, RFC2861 behavior. */ 66 int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 67 68 unsigned int sysctl_tcp_notsent_lowat __read_mostly = UINT_MAX; 69 EXPORT_SYMBOL(sysctl_tcp_notsent_lowat); 70 71 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 72 int push_one, gfp_t gfp); 73 74 /* Account for new data that has been sent to the network. */ 75 static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) 76 { 77 struct inet_connection_sock *icsk = inet_csk(sk); 78 struct tcp_sock *tp = tcp_sk(sk); 79 unsigned int prior_packets = tp->packets_out; 80 81 tcp_advance_send_head(sk, skb); 82 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 83 84 tp->packets_out += tcp_skb_pcount(skb); 85 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || 86 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { 87 tcp_rearm_rto(sk); 88 } 89 90 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 91 tcp_skb_pcount(skb)); 92 } 93 94 /* SND.NXT, if window was not shrunk. 95 * If window has been shrunk, what should we make? It is not clear at all. 96 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 97 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 98 * invalid. OK, let's make this for now: 99 */ 100 static inline __u32 tcp_acceptable_seq(const struct sock *sk) 101 { 102 const struct tcp_sock *tp = tcp_sk(sk); 103 104 if (!before(tcp_wnd_end(tp), tp->snd_nxt)) 105 return tp->snd_nxt; 106 else 107 return tcp_wnd_end(tp); 108 } 109 110 /* Calculate mss to advertise in SYN segment. 111 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 112 * 113 * 1. It is independent of path mtu. 114 * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 115 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 116 * attached devices, because some buggy hosts are confused by 117 * large MSS. 118 * 4. We do not make 3, we advertise MSS, calculated from first 119 * hop device mtu, but allow to raise it to ip_rt_min_advmss. 120 * This may be overridden via information stored in routing table. 121 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 122 * probably even Jumbo". 123 */ 124 static __u16 tcp_advertise_mss(struct sock *sk) 125 { 126 struct tcp_sock *tp = tcp_sk(sk); 127 const struct dst_entry *dst = __sk_dst_get(sk); 128 int mss = tp->advmss; 129 130 if (dst) { 131 unsigned int metric = dst_metric_advmss(dst); 132 133 if (metric < mss) { 134 mss = metric; 135 tp->advmss = mss; 136 } 137 } 138 139 return (__u16)mss; 140 } 141 142 /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 143 * This is the first part of cwnd validation mechanism. */ 144 static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst) 145 { 146 struct tcp_sock *tp = tcp_sk(sk); 147 s32 delta = tcp_time_stamp - tp->lsndtime; 148 u32 restart_cwnd = tcp_init_cwnd(tp, dst); 149 u32 cwnd = tp->snd_cwnd; 150 151 tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 152 153 tp->snd_ssthresh = tcp_current_ssthresh(sk); 154 restart_cwnd = min(restart_cwnd, cwnd); 155 156 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 157 cwnd >>= 1; 158 tp->snd_cwnd = max(cwnd, restart_cwnd); 159 tp->snd_cwnd_stamp = tcp_time_stamp; 160 tp->snd_cwnd_used = 0; 161 } 162 163 /* Congestion state accounting after a packet has been sent. */ 164 static void tcp_event_data_sent(struct tcp_sock *tp, 165 struct sock *sk) 166 { 167 struct inet_connection_sock *icsk = inet_csk(sk); 168 const u32 now = tcp_time_stamp; 169 const struct dst_entry *dst = __sk_dst_get(sk); 170 171 if (sysctl_tcp_slow_start_after_idle && 172 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) 173 tcp_cwnd_restart(sk, __sk_dst_get(sk)); 174 175 tp->lsndtime = now; 176 177 /* If it is a reply for ato after last received 178 * packet, enter pingpong mode. 179 */ 180 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato && 181 (!dst || !dst_metric(dst, RTAX_QUICKACK))) 182 icsk->icsk_ack.pingpong = 1; 183 } 184 185 /* Account for an ACK we sent. */ 186 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 187 { 188 tcp_dec_quickack_mode(sk, pkts); 189 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 190 } 191 192 193 u32 tcp_default_init_rwnd(u32 mss) 194 { 195 /* Initial receive window should be twice of TCP_INIT_CWND to 196 * enable proper sending of new unsent data during fast recovery 197 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a 198 * limit when mss is larger than 1460. 199 */ 200 u32 init_rwnd = TCP_INIT_CWND * 2; 201 202 if (mss > 1460) 203 init_rwnd = max((1460 * init_rwnd) / mss, 2U); 204 return init_rwnd; 205 } 206 207 /* Determine a window scaling and initial window to offer. 208 * Based on the assumption that the given amount of space 209 * will be offered. Store the results in the tp structure. 210 * NOTE: for smooth operation initial space offering should 211 * be a multiple of mss if possible. We assume here that mss >= 1. 212 * This MUST be enforced by all callers. 213 */ 214 void tcp_select_initial_window(int __space, __u32 mss, 215 __u32 *rcv_wnd, __u32 *window_clamp, 216 int wscale_ok, __u8 *rcv_wscale, 217 __u32 init_rcv_wnd) 218 { 219 unsigned int space = (__space < 0 ? 0 : __space); 220 221 /* If no clamp set the clamp to the max possible scaled window */ 222 if (*window_clamp == 0) 223 (*window_clamp) = (65535 << 14); 224 space = min(*window_clamp, space); 225 226 /* Quantize space offering to a multiple of mss if possible. */ 227 if (space > mss) 228 space = (space / mss) * mss; 229 230 /* NOTE: offering an initial window larger than 32767 231 * will break some buggy TCP stacks. If the admin tells us 232 * it is likely we could be speaking with such a buggy stack 233 * we will truncate our initial window offering to 32K-1 234 * unless the remote has sent us a window scaling option, 235 * which we interpret as a sign the remote TCP is not 236 * misinterpreting the window field as a signed quantity. 237 */ 238 if (sysctl_tcp_workaround_signed_windows) 239 (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 240 else 241 (*rcv_wnd) = space; 242 243 (*rcv_wscale) = 0; 244 if (wscale_ok) { 245 /* Set window scaling on max possible window 246 * See RFC1323 for an explanation of the limit to 14 247 */ 248 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 249 space = min_t(u32, space, *window_clamp); 250 while (space > 65535 && (*rcv_wscale) < 14) { 251 space >>= 1; 252 (*rcv_wscale)++; 253 } 254 } 255 256 if (mss > (1 << *rcv_wscale)) { 257 if (!init_rcv_wnd) /* Use default unless specified otherwise */ 258 init_rcv_wnd = tcp_default_init_rwnd(mss); 259 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 260 } 261 262 /* Set the clamp no higher than max representable value */ 263 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 264 } 265 EXPORT_SYMBOL(tcp_select_initial_window); 266 267 /* Chose a new window to advertise, update state in tcp_sock for the 268 * socket, and return result with RFC1323 scaling applied. The return 269 * value can be stuffed directly into th->window for an outgoing 270 * frame. 271 */ 272 static u16 tcp_select_window(struct sock *sk) 273 { 274 struct tcp_sock *tp = tcp_sk(sk); 275 u32 old_win = tp->rcv_wnd; 276 u32 cur_win = tcp_receive_window(tp); 277 u32 new_win = __tcp_select_window(sk); 278 279 /* Never shrink the offered window */ 280 if (new_win < cur_win) { 281 /* Danger Will Robinson! 282 * Don't update rcv_wup/rcv_wnd here or else 283 * we will not be able to advertise a zero 284 * window in time. --DaveM 285 * 286 * Relax Will Robinson. 287 */ 288 if (new_win == 0) 289 NET_INC_STATS(sock_net(sk), 290 LINUX_MIB_TCPWANTZEROWINDOWADV); 291 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 292 } 293 tp->rcv_wnd = new_win; 294 tp->rcv_wup = tp->rcv_nxt; 295 296 /* Make sure we do not exceed the maximum possible 297 * scaled window. 298 */ 299 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 300 new_win = min(new_win, MAX_TCP_WINDOW); 301 else 302 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 303 304 /* RFC1323 scaling applied */ 305 new_win >>= tp->rx_opt.rcv_wscale; 306 307 /* If we advertise zero window, disable fast path. */ 308 if (new_win == 0) { 309 tp->pred_flags = 0; 310 if (old_win) 311 NET_INC_STATS(sock_net(sk), 312 LINUX_MIB_TCPTOZEROWINDOWADV); 313 } else if (old_win == 0) { 314 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 315 } 316 317 return new_win; 318 } 319 320 /* Packet ECN state for a SYN-ACK */ 321 static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb) 322 { 323 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 324 if (!(tp->ecn_flags & TCP_ECN_OK)) 325 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 326 } 327 328 /* Packet ECN state for a SYN. */ 329 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) 330 { 331 struct tcp_sock *tp = tcp_sk(sk); 332 333 tp->ecn_flags = 0; 334 if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) { 335 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 336 tp->ecn_flags = TCP_ECN_OK; 337 } 338 } 339 340 static __inline__ void 341 TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th) 342 { 343 if (inet_rsk(req)->ecn_ok) 344 th->ece = 1; 345 } 346 347 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 348 * be sent. 349 */ 350 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, 351 int tcp_header_len) 352 { 353 struct tcp_sock *tp = tcp_sk(sk); 354 355 if (tp->ecn_flags & TCP_ECN_OK) { 356 /* Not-retransmitted data segment: set ECT and inject CWR. */ 357 if (skb->len != tcp_header_len && 358 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 359 INET_ECN_xmit(sk); 360 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 361 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 362 tcp_hdr(skb)->cwr = 1; 363 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 364 } 365 } else { 366 /* ACK or retransmitted segment: clear ECT|CE */ 367 INET_ECN_dontxmit(sk); 368 } 369 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 370 tcp_hdr(skb)->ece = 1; 371 } 372 } 373 374 /* Constructs common control bits of non-data skb. If SYN/FIN is present, 375 * auto increment end seqno. 376 */ 377 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 378 { 379 struct skb_shared_info *shinfo = skb_shinfo(skb); 380 381 skb->ip_summed = CHECKSUM_PARTIAL; 382 skb->csum = 0; 383 384 TCP_SKB_CB(skb)->tcp_flags = flags; 385 TCP_SKB_CB(skb)->sacked = 0; 386 387 shinfo->gso_segs = 1; 388 shinfo->gso_size = 0; 389 shinfo->gso_type = 0; 390 391 TCP_SKB_CB(skb)->seq = seq; 392 if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 393 seq++; 394 TCP_SKB_CB(skb)->end_seq = seq; 395 } 396 397 static inline bool tcp_urg_mode(const struct tcp_sock *tp) 398 { 399 return tp->snd_una != tp->snd_up; 400 } 401 402 #define OPTION_SACK_ADVERTISE (1 << 0) 403 #define OPTION_TS (1 << 1) 404 #define OPTION_MD5 (1 << 2) 405 #define OPTION_WSCALE (1 << 3) 406 #define OPTION_FAST_OPEN_COOKIE (1 << 8) 407 408 struct tcp_out_options { 409 u16 options; /* bit field of OPTION_* */ 410 u16 mss; /* 0 to disable */ 411 u8 ws; /* window scale, 0 to disable */ 412 u8 num_sack_blocks; /* number of SACK blocks to include */ 413 u8 hash_size; /* bytes in hash_location */ 414 __u8 *hash_location; /* temporary pointer, overloaded */ 415 __u32 tsval, tsecr; /* need to include OPTION_TS */ 416 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 417 }; 418 419 /* Write previously computed TCP options to the packet. 420 * 421 * Beware: Something in the Internet is very sensitive to the ordering of 422 * TCP options, we learned this through the hard way, so be careful here. 423 * Luckily we can at least blame others for their non-compliance but from 424 * inter-operability perspective it seems that we're somewhat stuck with 425 * the ordering which we have been using if we want to keep working with 426 * those broken things (not that it currently hurts anybody as there isn't 427 * particular reason why the ordering would need to be changed). 428 * 429 * At least SACK_PERM as the first option is known to lead to a disaster 430 * (but it may well be that other scenarios fail similarly). 431 */ 432 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 433 struct tcp_out_options *opts) 434 { 435 u16 options = opts->options; /* mungable copy */ 436 437 if (unlikely(OPTION_MD5 & options)) { 438 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 439 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 440 /* overload cookie hash location */ 441 opts->hash_location = (__u8 *)ptr; 442 ptr += 4; 443 } 444 445 if (unlikely(opts->mss)) { 446 *ptr++ = htonl((TCPOPT_MSS << 24) | 447 (TCPOLEN_MSS << 16) | 448 opts->mss); 449 } 450 451 if (likely(OPTION_TS & options)) { 452 if (unlikely(OPTION_SACK_ADVERTISE & options)) { 453 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 454 (TCPOLEN_SACK_PERM << 16) | 455 (TCPOPT_TIMESTAMP << 8) | 456 TCPOLEN_TIMESTAMP); 457 options &= ~OPTION_SACK_ADVERTISE; 458 } else { 459 *ptr++ = htonl((TCPOPT_NOP << 24) | 460 (TCPOPT_NOP << 16) | 461 (TCPOPT_TIMESTAMP << 8) | 462 TCPOLEN_TIMESTAMP); 463 } 464 *ptr++ = htonl(opts->tsval); 465 *ptr++ = htonl(opts->tsecr); 466 } 467 468 if (unlikely(OPTION_SACK_ADVERTISE & options)) { 469 *ptr++ = htonl((TCPOPT_NOP << 24) | 470 (TCPOPT_NOP << 16) | 471 (TCPOPT_SACK_PERM << 8) | 472 TCPOLEN_SACK_PERM); 473 } 474 475 if (unlikely(OPTION_WSCALE & options)) { 476 *ptr++ = htonl((TCPOPT_NOP << 24) | 477 (TCPOPT_WINDOW << 16) | 478 (TCPOLEN_WINDOW << 8) | 479 opts->ws); 480 } 481 482 if (unlikely(opts->num_sack_blocks)) { 483 struct tcp_sack_block *sp = tp->rx_opt.dsack ? 484 tp->duplicate_sack : tp->selective_acks; 485 int this_sack; 486 487 *ptr++ = htonl((TCPOPT_NOP << 24) | 488 (TCPOPT_NOP << 16) | 489 (TCPOPT_SACK << 8) | 490 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 491 TCPOLEN_SACK_PERBLOCK))); 492 493 for (this_sack = 0; this_sack < opts->num_sack_blocks; 494 ++this_sack) { 495 *ptr++ = htonl(sp[this_sack].start_seq); 496 *ptr++ = htonl(sp[this_sack].end_seq); 497 } 498 499 tp->rx_opt.dsack = 0; 500 } 501 502 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 503 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 504 505 *ptr++ = htonl((TCPOPT_EXP << 24) | 506 ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) | 507 TCPOPT_FASTOPEN_MAGIC); 508 509 memcpy(ptr, foc->val, foc->len); 510 if ((foc->len & 3) == 2) { 511 u8 *align = ((u8 *)ptr) + foc->len; 512 align[0] = align[1] = TCPOPT_NOP; 513 } 514 ptr += (foc->len + 3) >> 2; 515 } 516 } 517 518 /* Compute TCP options for SYN packets. This is not the final 519 * network wire format yet. 520 */ 521 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 522 struct tcp_out_options *opts, 523 struct tcp_md5sig_key **md5) 524 { 525 struct tcp_sock *tp = tcp_sk(sk); 526 unsigned int remaining = MAX_TCP_OPTION_SPACE; 527 struct tcp_fastopen_request *fastopen = tp->fastopen_req; 528 529 #ifdef CONFIG_TCP_MD5SIG 530 *md5 = tp->af_specific->md5_lookup(sk, sk); 531 if (*md5) { 532 opts->options |= OPTION_MD5; 533 remaining -= TCPOLEN_MD5SIG_ALIGNED; 534 } 535 #else 536 *md5 = NULL; 537 #endif 538 539 /* We always get an MSS option. The option bytes which will be seen in 540 * normal data packets should timestamps be used, must be in the MSS 541 * advertised. But we subtract them from tp->mss_cache so that 542 * calculations in tcp_sendmsg are simpler etc. So account for this 543 * fact here if necessary. If we don't do this correctly, as a 544 * receiver we won't recognize data packets as being full sized when we 545 * should, and thus we won't abide by the delayed ACK rules correctly. 546 * SACKs don't matter, we never delay an ACK when we have any of those 547 * going out. */ 548 opts->mss = tcp_advertise_mss(sk); 549 remaining -= TCPOLEN_MSS_ALIGNED; 550 551 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { 552 opts->options |= OPTION_TS; 553 opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset; 554 opts->tsecr = tp->rx_opt.ts_recent; 555 remaining -= TCPOLEN_TSTAMP_ALIGNED; 556 } 557 if (likely(sysctl_tcp_window_scaling)) { 558 opts->ws = tp->rx_opt.rcv_wscale; 559 opts->options |= OPTION_WSCALE; 560 remaining -= TCPOLEN_WSCALE_ALIGNED; 561 } 562 if (likely(sysctl_tcp_sack)) { 563 opts->options |= OPTION_SACK_ADVERTISE; 564 if (unlikely(!(OPTION_TS & opts->options))) 565 remaining -= TCPOLEN_SACKPERM_ALIGNED; 566 } 567 568 if (fastopen && fastopen->cookie.len >= 0) { 569 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len; 570 need = (need + 3) & ~3U; /* Align to 32 bits */ 571 if (remaining >= need) { 572 opts->options |= OPTION_FAST_OPEN_COOKIE; 573 opts->fastopen_cookie = &fastopen->cookie; 574 remaining -= need; 575 tp->syn_fastopen = 1; 576 } 577 } 578 579 return MAX_TCP_OPTION_SPACE - remaining; 580 } 581 582 /* Set up TCP options for SYN-ACKs. */ 583 static unsigned int tcp_synack_options(struct sock *sk, 584 struct request_sock *req, 585 unsigned int mss, struct sk_buff *skb, 586 struct tcp_out_options *opts, 587 struct tcp_md5sig_key **md5, 588 struct tcp_fastopen_cookie *foc) 589 { 590 struct inet_request_sock *ireq = inet_rsk(req); 591 unsigned int remaining = MAX_TCP_OPTION_SPACE; 592 593 #ifdef CONFIG_TCP_MD5SIG 594 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 595 if (*md5) { 596 opts->options |= OPTION_MD5; 597 remaining -= TCPOLEN_MD5SIG_ALIGNED; 598 599 /* We can't fit any SACK blocks in a packet with MD5 + TS 600 * options. There was discussion about disabling SACK 601 * rather than TS in order to fit in better with old, 602 * buggy kernels, but that was deemed to be unnecessary. 603 */ 604 ireq->tstamp_ok &= !ireq->sack_ok; 605 } 606 #else 607 *md5 = NULL; 608 #endif 609 610 /* We always send an MSS option. */ 611 opts->mss = mss; 612 remaining -= TCPOLEN_MSS_ALIGNED; 613 614 if (likely(ireq->wscale_ok)) { 615 opts->ws = ireq->rcv_wscale; 616 opts->options |= OPTION_WSCALE; 617 remaining -= TCPOLEN_WSCALE_ALIGNED; 618 } 619 if (likely(ireq->tstamp_ok)) { 620 opts->options |= OPTION_TS; 621 opts->tsval = TCP_SKB_CB(skb)->when; 622 opts->tsecr = req->ts_recent; 623 remaining -= TCPOLEN_TSTAMP_ALIGNED; 624 } 625 if (likely(ireq->sack_ok)) { 626 opts->options |= OPTION_SACK_ADVERTISE; 627 if (unlikely(!ireq->tstamp_ok)) 628 remaining -= TCPOLEN_SACKPERM_ALIGNED; 629 } 630 if (foc != NULL && foc->len >= 0) { 631 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 632 need = (need + 3) & ~3U; /* Align to 32 bits */ 633 if (remaining >= need) { 634 opts->options |= OPTION_FAST_OPEN_COOKIE; 635 opts->fastopen_cookie = foc; 636 remaining -= need; 637 } 638 } 639 640 return MAX_TCP_OPTION_SPACE - remaining; 641 } 642 643 /* Compute TCP options for ESTABLISHED sockets. This is not the 644 * final wire format yet. 645 */ 646 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 647 struct tcp_out_options *opts, 648 struct tcp_md5sig_key **md5) 649 { 650 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; 651 struct tcp_sock *tp = tcp_sk(sk); 652 unsigned int size = 0; 653 unsigned int eff_sacks; 654 655 opts->options = 0; 656 657 #ifdef CONFIG_TCP_MD5SIG 658 *md5 = tp->af_specific->md5_lookup(sk, sk); 659 if (unlikely(*md5)) { 660 opts->options |= OPTION_MD5; 661 size += TCPOLEN_MD5SIG_ALIGNED; 662 } 663 #else 664 *md5 = NULL; 665 #endif 666 667 if (likely(tp->rx_opt.tstamp_ok)) { 668 opts->options |= OPTION_TS; 669 opts->tsval = tcb ? tcb->when + tp->tsoffset : 0; 670 opts->tsecr = tp->rx_opt.ts_recent; 671 size += TCPOLEN_TSTAMP_ALIGNED; 672 } 673 674 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 675 if (unlikely(eff_sacks)) { 676 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 677 opts->num_sack_blocks = 678 min_t(unsigned int, eff_sacks, 679 (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 680 TCPOLEN_SACK_PERBLOCK); 681 size += TCPOLEN_SACK_BASE_ALIGNED + 682 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 683 } 684 685 return size; 686 } 687 688 689 /* TCP SMALL QUEUES (TSQ) 690 * 691 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 692 * to reduce RTT and bufferbloat. 693 * We do this using a special skb destructor (tcp_wfree). 694 * 695 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 696 * needs to be reallocated in a driver. 697 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 698 * 699 * Since transmit from skb destructor is forbidden, we use a tasklet 700 * to process all sockets that eventually need to send more skbs. 701 * We use one tasklet per cpu, with its own queue of sockets. 702 */ 703 struct tsq_tasklet { 704 struct tasklet_struct tasklet; 705 struct list_head head; /* queue of tcp sockets */ 706 }; 707 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 708 709 static void tcp_tsq_handler(struct sock *sk) 710 { 711 if ((1 << sk->sk_state) & 712 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 713 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) 714 tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle, 715 0, GFP_ATOMIC); 716 } 717 /* 718 * One tasklet per cpu tries to send more skbs. 719 * We run in tasklet context but need to disable irqs when 720 * transferring tsq->head because tcp_wfree() might 721 * interrupt us (non NAPI drivers) 722 */ 723 static void tcp_tasklet_func(unsigned long data) 724 { 725 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 726 LIST_HEAD(list); 727 unsigned long flags; 728 struct list_head *q, *n; 729 struct tcp_sock *tp; 730 struct sock *sk; 731 732 local_irq_save(flags); 733 list_splice_init(&tsq->head, &list); 734 local_irq_restore(flags); 735 736 list_for_each_safe(q, n, &list) { 737 tp = list_entry(q, struct tcp_sock, tsq_node); 738 list_del(&tp->tsq_node); 739 740 sk = (struct sock *)tp; 741 bh_lock_sock(sk); 742 743 if (!sock_owned_by_user(sk)) { 744 tcp_tsq_handler(sk); 745 } else { 746 /* defer the work to tcp_release_cb() */ 747 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); 748 } 749 bh_unlock_sock(sk); 750 751 clear_bit(TSQ_QUEUED, &tp->tsq_flags); 752 sk_free(sk); 753 } 754 } 755 756 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ 757 (1UL << TCP_WRITE_TIMER_DEFERRED) | \ 758 (1UL << TCP_DELACK_TIMER_DEFERRED) | \ 759 (1UL << TCP_MTU_REDUCED_DEFERRED)) 760 /** 761 * tcp_release_cb - tcp release_sock() callback 762 * @sk: socket 763 * 764 * called from release_sock() to perform protocol dependent 765 * actions before socket release. 766 */ 767 void tcp_release_cb(struct sock *sk) 768 { 769 struct tcp_sock *tp = tcp_sk(sk); 770 unsigned long flags, nflags; 771 772 /* perform an atomic operation only if at least one flag is set */ 773 do { 774 flags = tp->tsq_flags; 775 if (!(flags & TCP_DEFERRED_ALL)) 776 return; 777 nflags = flags & ~TCP_DEFERRED_ALL; 778 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); 779 780 if (flags & (1UL << TCP_TSQ_DEFERRED)) 781 tcp_tsq_handler(sk); 782 783 /* Here begins the tricky part : 784 * We are called from release_sock() with : 785 * 1) BH disabled 786 * 2) sk_lock.slock spinlock held 787 * 3) socket owned by us (sk->sk_lock.owned == 1) 788 * 789 * But following code is meant to be called from BH handlers, 790 * so we should keep BH disabled, but early release socket ownership 791 */ 792 sock_release_ownership(sk); 793 794 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { 795 tcp_write_timer_handler(sk); 796 __sock_put(sk); 797 } 798 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { 799 tcp_delack_timer_handler(sk); 800 __sock_put(sk); 801 } 802 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { 803 sk->sk_prot->mtu_reduced(sk); 804 __sock_put(sk); 805 } 806 } 807 EXPORT_SYMBOL(tcp_release_cb); 808 809 void __init tcp_tasklet_init(void) 810 { 811 int i; 812 813 for_each_possible_cpu(i) { 814 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 815 816 INIT_LIST_HEAD(&tsq->head); 817 tasklet_init(&tsq->tasklet, 818 tcp_tasklet_func, 819 (unsigned long)tsq); 820 } 821 } 822 823 /* 824 * Write buffer destructor automatically called from kfree_skb. 825 * We can't xmit new skbs from this context, as we might already 826 * hold qdisc lock. 827 */ 828 void tcp_wfree(struct sk_buff *skb) 829 { 830 struct sock *sk = skb->sk; 831 struct tcp_sock *tp = tcp_sk(sk); 832 833 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && 834 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { 835 unsigned long flags; 836 struct tsq_tasklet *tsq; 837 838 /* Keep a ref on socket. 839 * This last ref will be released in tcp_tasklet_func() 840 */ 841 atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc); 842 843 /* queue this socket to tasklet queue */ 844 local_irq_save(flags); 845 tsq = &__get_cpu_var(tsq_tasklet); 846 list_add(&tp->tsq_node, &tsq->head); 847 tasklet_schedule(&tsq->tasklet); 848 local_irq_restore(flags); 849 } else { 850 sock_wfree(skb); 851 } 852 } 853 854 /* This routine actually transmits TCP packets queued in by 855 * tcp_do_sendmsg(). This is used by both the initial 856 * transmission and possible later retransmissions. 857 * All SKB's seen here are completely headerless. It is our 858 * job to build the TCP header, and pass the packet down to 859 * IP so it can do the same plus pass the packet off to the 860 * device. 861 * 862 * We are working here with either a clone of the original 863 * SKB, or a fresh unique copy made by the retransmit engine. 864 */ 865 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 866 gfp_t gfp_mask) 867 { 868 const struct inet_connection_sock *icsk = inet_csk(sk); 869 struct inet_sock *inet; 870 struct tcp_sock *tp; 871 struct tcp_skb_cb *tcb; 872 struct tcp_out_options opts; 873 unsigned int tcp_options_size, tcp_header_size; 874 struct tcp_md5sig_key *md5; 875 struct tcphdr *th; 876 int err; 877 878 BUG_ON(!skb || !tcp_skb_pcount(skb)); 879 880 if (clone_it) { 881 skb_mstamp_get(&skb->skb_mstamp); 882 883 if (unlikely(skb_cloned(skb))) 884 skb = pskb_copy(skb, gfp_mask); 885 else 886 skb = skb_clone(skb, gfp_mask); 887 if (unlikely(!skb)) 888 return -ENOBUFS; 889 /* Our usage of tstamp should remain private */ 890 skb->tstamp.tv64 = 0; 891 } 892 893 inet = inet_sk(sk); 894 tp = tcp_sk(sk); 895 tcb = TCP_SKB_CB(skb); 896 memset(&opts, 0, sizeof(opts)); 897 898 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 899 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 900 else 901 tcp_options_size = tcp_established_options(sk, skb, &opts, 902 &md5); 903 tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 904 905 if (tcp_packets_in_flight(tp) == 0) 906 tcp_ca_event(sk, CA_EVENT_TX_START); 907 908 /* if no packet is in qdisc/device queue, then allow XPS to select 909 * another queue. 910 */ 911 skb->ooo_okay = sk_wmem_alloc_get(sk) == 0; 912 913 skb_push(skb, tcp_header_size); 914 skb_reset_transport_header(skb); 915 916 skb_orphan(skb); 917 skb->sk = sk; 918 skb->destructor = tcp_wfree; 919 atomic_add(skb->truesize, &sk->sk_wmem_alloc); 920 921 /* Build TCP header and checksum it. */ 922 th = tcp_hdr(skb); 923 th->source = inet->inet_sport; 924 th->dest = inet->inet_dport; 925 th->seq = htonl(tcb->seq); 926 th->ack_seq = htonl(tp->rcv_nxt); 927 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 928 tcb->tcp_flags); 929 930 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 931 /* RFC1323: The window in SYN & SYN/ACK segments 932 * is never scaled. 933 */ 934 th->window = htons(min(tp->rcv_wnd, 65535U)); 935 } else { 936 th->window = htons(tcp_select_window(sk)); 937 } 938 th->check = 0; 939 th->urg_ptr = 0; 940 941 /* The urg_mode check is necessary during a below snd_una win probe */ 942 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 943 if (before(tp->snd_up, tcb->seq + 0x10000)) { 944 th->urg_ptr = htons(tp->snd_up - tcb->seq); 945 th->urg = 1; 946 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 947 th->urg_ptr = htons(0xFFFF); 948 th->urg = 1; 949 } 950 } 951 952 tcp_options_write((__be32 *)(th + 1), tp, &opts); 953 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) 954 TCP_ECN_send(sk, skb, tcp_header_size); 955 956 #ifdef CONFIG_TCP_MD5SIG 957 /* Calculate the MD5 hash, as we have all we need now */ 958 if (md5) { 959 sk_nocaps_add(sk, NETIF_F_GSO_MASK); 960 tp->af_specific->calc_md5_hash(opts.hash_location, 961 md5, sk, NULL, skb); 962 } 963 #endif 964 965 icsk->icsk_af_ops->send_check(sk, skb); 966 967 if (likely(tcb->tcp_flags & TCPHDR_ACK)) 968 tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 969 970 if (skb->len != tcp_header_size) 971 tcp_event_data_sent(tp, sk); 972 973 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 974 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 975 tcp_skb_pcount(skb)); 976 977 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 978 if (likely(err <= 0)) 979 return err; 980 981 tcp_enter_cwr(sk, 1); 982 983 return net_xmit_eval(err); 984 } 985 986 /* This routine just queues the buffer for sending. 987 * 988 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 989 * otherwise socket can stall. 990 */ 991 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 992 { 993 struct tcp_sock *tp = tcp_sk(sk); 994 995 /* Advance write_seq and place onto the write_queue. */ 996 tp->write_seq = TCP_SKB_CB(skb)->end_seq; 997 skb_header_release(skb); 998 tcp_add_write_queue_tail(sk, skb); 999 sk->sk_wmem_queued += skb->truesize; 1000 sk_mem_charge(sk, skb->truesize); 1001 } 1002 1003 /* Initialize TSO segments for a packet. */ 1004 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, 1005 unsigned int mss_now) 1006 { 1007 struct skb_shared_info *shinfo = skb_shinfo(skb); 1008 1009 /* Make sure we own this skb before messing gso_size/gso_segs */ 1010 WARN_ON_ONCE(skb_cloned(skb)); 1011 1012 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 1013 /* Avoid the costly divide in the normal 1014 * non-TSO case. 1015 */ 1016 shinfo->gso_segs = 1; 1017 shinfo->gso_size = 0; 1018 shinfo->gso_type = 0; 1019 } else { 1020 shinfo->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 1021 shinfo->gso_size = mss_now; 1022 shinfo->gso_type = sk->sk_gso_type; 1023 } 1024 } 1025 1026 /* When a modification to fackets out becomes necessary, we need to check 1027 * skb is counted to fackets_out or not. 1028 */ 1029 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 1030 int decr) 1031 { 1032 struct tcp_sock *tp = tcp_sk(sk); 1033 1034 if (!tp->sacked_out || tcp_is_reno(tp)) 1035 return; 1036 1037 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 1038 tp->fackets_out -= decr; 1039 } 1040 1041 /* Pcount in the middle of the write queue got changed, we need to do various 1042 * tweaks to fix counters 1043 */ 1044 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1045 { 1046 struct tcp_sock *tp = tcp_sk(sk); 1047 1048 tp->packets_out -= decr; 1049 1050 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1051 tp->sacked_out -= decr; 1052 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1053 tp->retrans_out -= decr; 1054 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1055 tp->lost_out -= decr; 1056 1057 /* Reno case is special. Sigh... */ 1058 if (tcp_is_reno(tp) && decr > 0) 1059 tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1060 1061 tcp_adjust_fackets_out(sk, skb, decr); 1062 1063 if (tp->lost_skb_hint && 1064 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1065 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1066 tp->lost_cnt_hint -= decr; 1067 1068 tcp_verify_left_out(tp); 1069 } 1070 1071 /* Function to create two new TCP segments. Shrinks the given segment 1072 * to the specified size and appends a new segment with the rest of the 1073 * packet to the list. This won't be called frequently, I hope. 1074 * Remember, these are still headerless SKBs at this point. 1075 */ 1076 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 1077 unsigned int mss_now, gfp_t gfp) 1078 { 1079 struct tcp_sock *tp = tcp_sk(sk); 1080 struct sk_buff *buff; 1081 int nsize, old_factor; 1082 int nlen; 1083 u8 flags; 1084 1085 if (WARN_ON(len > skb->len)) 1086 return -EINVAL; 1087 1088 nsize = skb_headlen(skb) - len; 1089 if (nsize < 0) 1090 nsize = 0; 1091 1092 if (skb_unclone(skb, gfp)) 1093 return -ENOMEM; 1094 1095 /* Get a new skb... force flag on. */ 1096 buff = sk_stream_alloc_skb(sk, nsize, gfp); 1097 if (buff == NULL) 1098 return -ENOMEM; /* We'll just try again later. */ 1099 1100 sk->sk_wmem_queued += buff->truesize; 1101 sk_mem_charge(sk, buff->truesize); 1102 nlen = skb->len - len - nsize; 1103 buff->truesize += nlen; 1104 skb->truesize -= nlen; 1105 1106 /* Correct the sequence numbers. */ 1107 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1108 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1109 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1110 1111 /* PSH and FIN should only be set in the second packet. */ 1112 flags = TCP_SKB_CB(skb)->tcp_flags; 1113 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 1114 TCP_SKB_CB(buff)->tcp_flags = flags; 1115 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1116 1117 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 1118 /* Copy and checksum data tail into the new buffer. */ 1119 buff->csum = csum_partial_copy_nocheck(skb->data + len, 1120 skb_put(buff, nsize), 1121 nsize, 0); 1122 1123 skb_trim(skb, len); 1124 1125 skb->csum = csum_block_sub(skb->csum, buff->csum, len); 1126 } else { 1127 skb->ip_summed = CHECKSUM_PARTIAL; 1128 skb_split(skb, buff, len); 1129 } 1130 1131 buff->ip_summed = skb->ip_summed; 1132 1133 /* Looks stupid, but our code really uses when of 1134 * skbs, which it never sent before. --ANK 1135 */ 1136 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 1137 buff->tstamp = skb->tstamp; 1138 1139 old_factor = tcp_skb_pcount(skb); 1140 1141 /* Fix up tso_factor for both original and new SKB. */ 1142 tcp_set_skb_tso_segs(sk, skb, mss_now); 1143 tcp_set_skb_tso_segs(sk, buff, mss_now); 1144 1145 /* If this packet has been sent out already, we must 1146 * adjust the various packet counters. 1147 */ 1148 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 1149 int diff = old_factor - tcp_skb_pcount(skb) - 1150 tcp_skb_pcount(buff); 1151 1152 if (diff) 1153 tcp_adjust_pcount(sk, skb, diff); 1154 } 1155 1156 /* Link BUFF into the send queue. */ 1157 skb_header_release(buff); 1158 tcp_insert_write_queue_after(skb, buff, sk); 1159 1160 return 0; 1161 } 1162 1163 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 1164 * eventually). The difference is that pulled data not copied, but 1165 * immediately discarded. 1166 */ 1167 static void __pskb_trim_head(struct sk_buff *skb, int len) 1168 { 1169 struct skb_shared_info *shinfo; 1170 int i, k, eat; 1171 1172 eat = min_t(int, len, skb_headlen(skb)); 1173 if (eat) { 1174 __skb_pull(skb, eat); 1175 len -= eat; 1176 if (!len) 1177 return; 1178 } 1179 eat = len; 1180 k = 0; 1181 shinfo = skb_shinfo(skb); 1182 for (i = 0; i < shinfo->nr_frags; i++) { 1183 int size = skb_frag_size(&shinfo->frags[i]); 1184 1185 if (size <= eat) { 1186 skb_frag_unref(skb, i); 1187 eat -= size; 1188 } else { 1189 shinfo->frags[k] = shinfo->frags[i]; 1190 if (eat) { 1191 shinfo->frags[k].page_offset += eat; 1192 skb_frag_size_sub(&shinfo->frags[k], eat); 1193 eat = 0; 1194 } 1195 k++; 1196 } 1197 } 1198 shinfo->nr_frags = k; 1199 1200 skb_reset_tail_pointer(skb); 1201 skb->data_len -= len; 1202 skb->len = skb->data_len; 1203 } 1204 1205 /* Remove acked data from a packet in the transmit queue. */ 1206 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 1207 { 1208 if (skb_unclone(skb, GFP_ATOMIC)) 1209 return -ENOMEM; 1210 1211 __pskb_trim_head(skb, len); 1212 1213 TCP_SKB_CB(skb)->seq += len; 1214 skb->ip_summed = CHECKSUM_PARTIAL; 1215 1216 skb->truesize -= len; 1217 sk->sk_wmem_queued -= len; 1218 sk_mem_uncharge(sk, len); 1219 sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 1220 1221 /* Any change of skb->len requires recalculation of tso factor. */ 1222 if (tcp_skb_pcount(skb) > 1) 1223 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); 1224 1225 return 0; 1226 } 1227 1228 /* Calculate MSS not accounting any TCP options. */ 1229 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 1230 { 1231 const struct tcp_sock *tp = tcp_sk(sk); 1232 const struct inet_connection_sock *icsk = inet_csk(sk); 1233 int mss_now; 1234 1235 /* Calculate base mss without TCP options: 1236 It is MMS_S - sizeof(tcphdr) of rfc1122 1237 */ 1238 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 1239 1240 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 1241 if (icsk->icsk_af_ops->net_frag_header_len) { 1242 const struct dst_entry *dst = __sk_dst_get(sk); 1243 1244 if (dst && dst_allfrag(dst)) 1245 mss_now -= icsk->icsk_af_ops->net_frag_header_len; 1246 } 1247 1248 /* Clamp it (mss_clamp does not include tcp options) */ 1249 if (mss_now > tp->rx_opt.mss_clamp) 1250 mss_now = tp->rx_opt.mss_clamp; 1251 1252 /* Now subtract optional transport overhead */ 1253 mss_now -= icsk->icsk_ext_hdr_len; 1254 1255 /* Then reserve room for full set of TCP options and 8 bytes of data */ 1256 if (mss_now < 48) 1257 mss_now = 48; 1258 return mss_now; 1259 } 1260 1261 /* Calculate MSS. Not accounting for SACKs here. */ 1262 int tcp_mtu_to_mss(struct sock *sk, int pmtu) 1263 { 1264 /* Subtract TCP options size, not including SACKs */ 1265 return __tcp_mtu_to_mss(sk, pmtu) - 1266 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 1267 } 1268 1269 /* Inverse of above */ 1270 int tcp_mss_to_mtu(struct sock *sk, int mss) 1271 { 1272 const struct tcp_sock *tp = tcp_sk(sk); 1273 const struct inet_connection_sock *icsk = inet_csk(sk); 1274 int mtu; 1275 1276 mtu = mss + 1277 tp->tcp_header_len + 1278 icsk->icsk_ext_hdr_len + 1279 icsk->icsk_af_ops->net_header_len; 1280 1281 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 1282 if (icsk->icsk_af_ops->net_frag_header_len) { 1283 const struct dst_entry *dst = __sk_dst_get(sk); 1284 1285 if (dst && dst_allfrag(dst)) 1286 mtu += icsk->icsk_af_ops->net_frag_header_len; 1287 } 1288 return mtu; 1289 } 1290 1291 /* MTU probing init per socket */ 1292 void tcp_mtup_init(struct sock *sk) 1293 { 1294 struct tcp_sock *tp = tcp_sk(sk); 1295 struct inet_connection_sock *icsk = inet_csk(sk); 1296 1297 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 1298 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 1299 icsk->icsk_af_ops->net_header_len; 1300 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 1301 icsk->icsk_mtup.probe_size = 0; 1302 } 1303 EXPORT_SYMBOL(tcp_mtup_init); 1304 1305 /* This function synchronize snd mss to current pmtu/exthdr set. 1306 1307 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 1308 for TCP options, but includes only bare TCP header. 1309 1310 tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1311 It is minimum of user_mss and mss received with SYN. 1312 It also does not include TCP options. 1313 1314 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 1315 1316 tp->mss_cache is current effective sending mss, including 1317 all tcp options except for SACKs. It is evaluated, 1318 taking into account current pmtu, but never exceeds 1319 tp->rx_opt.mss_clamp. 1320 1321 NOTE1. rfc1122 clearly states that advertised MSS 1322 DOES NOT include either tcp or ip options. 1323 1324 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1325 are READ ONLY outside this function. --ANK (980731) 1326 */ 1327 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 1328 { 1329 struct tcp_sock *tp = tcp_sk(sk); 1330 struct inet_connection_sock *icsk = inet_csk(sk); 1331 int mss_now; 1332 1333 if (icsk->icsk_mtup.search_high > pmtu) 1334 icsk->icsk_mtup.search_high = pmtu; 1335 1336 mss_now = tcp_mtu_to_mss(sk, pmtu); 1337 mss_now = tcp_bound_to_half_wnd(tp, mss_now); 1338 1339 /* And store cached results */ 1340 icsk->icsk_pmtu_cookie = pmtu; 1341 if (icsk->icsk_mtup.enabled) 1342 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1343 tp->mss_cache = mss_now; 1344 1345 return mss_now; 1346 } 1347 EXPORT_SYMBOL(tcp_sync_mss); 1348 1349 /* Compute the current effective MSS, taking SACKs and IP options, 1350 * and even PMTU discovery events into account. 1351 */ 1352 unsigned int tcp_current_mss(struct sock *sk) 1353 { 1354 const struct tcp_sock *tp = tcp_sk(sk); 1355 const struct dst_entry *dst = __sk_dst_get(sk); 1356 u32 mss_now; 1357 unsigned int header_len; 1358 struct tcp_out_options opts; 1359 struct tcp_md5sig_key *md5; 1360 1361 mss_now = tp->mss_cache; 1362 1363 if (dst) { 1364 u32 mtu = dst_mtu(dst); 1365 if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 1366 mss_now = tcp_sync_mss(sk, mtu); 1367 } 1368 1369 header_len = tcp_established_options(sk, NULL, &opts, &md5) + 1370 sizeof(struct tcphdr); 1371 /* The mss_cache is sized based on tp->tcp_header_len, which assumes 1372 * some common options. If this is an odd packet (because we have SACK 1373 * blocks etc) then our calculated header_len will be different, and 1374 * we have to adjust mss_now correspondingly */ 1375 if (header_len != tp->tcp_header_len) { 1376 int delta = (int) header_len - tp->tcp_header_len; 1377 mss_now -= delta; 1378 } 1379 1380 return mss_now; 1381 } 1382 1383 /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 1384 * As additional protections, we do not touch cwnd in retransmission phases, 1385 * and if application hit its sndbuf limit recently. 1386 */ 1387 static void tcp_cwnd_application_limited(struct sock *sk) 1388 { 1389 struct tcp_sock *tp = tcp_sk(sk); 1390 1391 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 1392 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 1393 /* Limited by application or receiver window. */ 1394 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 1395 u32 win_used = max(tp->snd_cwnd_used, init_win); 1396 if (win_used < tp->snd_cwnd) { 1397 tp->snd_ssthresh = tcp_current_ssthresh(sk); 1398 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 1399 } 1400 tp->snd_cwnd_used = 0; 1401 } 1402 tp->snd_cwnd_stamp = tcp_time_stamp; 1403 } 1404 1405 static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1406 { 1407 struct tcp_sock *tp = tcp_sk(sk); 1408 1409 /* Track the maximum number of outstanding packets in each 1410 * window, and remember whether we were cwnd-limited then. 1411 */ 1412 if (!before(tp->snd_una, tp->max_packets_seq) || 1413 tp->packets_out > tp->max_packets_out) { 1414 tp->max_packets_out = tp->packets_out; 1415 tp->max_packets_seq = tp->snd_nxt; 1416 tp->is_cwnd_limited = is_cwnd_limited; 1417 } 1418 1419 if (tcp_is_cwnd_limited(sk)) { 1420 /* Network is feed fully. */ 1421 tp->snd_cwnd_used = 0; 1422 tp->snd_cwnd_stamp = tcp_time_stamp; 1423 } else { 1424 /* Network starves. */ 1425 if (tp->packets_out > tp->snd_cwnd_used) 1426 tp->snd_cwnd_used = tp->packets_out; 1427 1428 if (sysctl_tcp_slow_start_after_idle && 1429 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1430 tcp_cwnd_application_limited(sk); 1431 } 1432 } 1433 1434 /* Minshall's variant of the Nagle send check. */ 1435 static bool tcp_minshall_check(const struct tcp_sock *tp) 1436 { 1437 return after(tp->snd_sml, tp->snd_una) && 1438 !after(tp->snd_sml, tp->snd_nxt); 1439 } 1440 1441 /* Update snd_sml if this skb is under mss 1442 * Note that a TSO packet might end with a sub-mss segment 1443 * The test is really : 1444 * if ((skb->len % mss) != 0) 1445 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1446 * But we can avoid doing the divide again given we already have 1447 * skb_pcount = skb->len / mss_now 1448 */ 1449 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1450 const struct sk_buff *skb) 1451 { 1452 if (skb->len < tcp_skb_pcount(skb) * mss_now) 1453 tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1454 } 1455 1456 /* Return false, if packet can be sent now without violation Nagle's rules: 1457 * 1. It is full sized. (provided by caller in %partial bool) 1458 * 2. Or it contains FIN. (already checked by caller) 1459 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1460 * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1461 * With Minshall's modification: all sent small packets are ACKed. 1462 */ 1463 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1464 int nonagle) 1465 { 1466 return partial && 1467 ((nonagle & TCP_NAGLE_CORK) || 1468 (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1469 } 1470 /* Returns the portion of skb which can be sent right away */ 1471 static unsigned int tcp_mss_split_point(const struct sock *sk, 1472 const struct sk_buff *skb, 1473 unsigned int mss_now, 1474 unsigned int max_segs, 1475 int nonagle) 1476 { 1477 const struct tcp_sock *tp = tcp_sk(sk); 1478 u32 partial, needed, window, max_len; 1479 1480 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1481 max_len = mss_now * max_segs; 1482 1483 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 1484 return max_len; 1485 1486 needed = min(skb->len, window); 1487 1488 if (max_len <= needed) 1489 return max_len; 1490 1491 partial = needed % mss_now; 1492 /* If last segment is not a full MSS, check if Nagle rules allow us 1493 * to include this last segment in this skb. 1494 * Otherwise, we'll split the skb at last MSS boundary 1495 */ 1496 if (tcp_nagle_check(partial != 0, tp, nonagle)) 1497 return needed - partial; 1498 1499 return needed; 1500 } 1501 1502 /* Can at least one segment of SKB be sent right now, according to the 1503 * congestion window rules? If so, return how many segments are allowed. 1504 */ 1505 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1506 const struct sk_buff *skb) 1507 { 1508 u32 in_flight, cwnd; 1509 1510 /* Don't be strict about the congestion window for the final FIN. */ 1511 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 1512 tcp_skb_pcount(skb) == 1) 1513 return 1; 1514 1515 in_flight = tcp_packets_in_flight(tp); 1516 cwnd = tp->snd_cwnd; 1517 if (in_flight < cwnd) 1518 return (cwnd - in_flight); 1519 1520 return 0; 1521 } 1522 1523 /* Initialize TSO state of a skb. 1524 * This must be invoked the first time we consider transmitting 1525 * SKB onto the wire. 1526 */ 1527 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, 1528 unsigned int mss_now) 1529 { 1530 int tso_segs = tcp_skb_pcount(skb); 1531 1532 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1533 tcp_set_skb_tso_segs(sk, skb, mss_now); 1534 tso_segs = tcp_skb_pcount(skb); 1535 } 1536 return tso_segs; 1537 } 1538 1539 1540 /* Return true if the Nagle test allows this packet to be 1541 * sent now. 1542 */ 1543 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1544 unsigned int cur_mss, int nonagle) 1545 { 1546 /* Nagle rule does not apply to frames, which sit in the middle of the 1547 * write_queue (they have no chances to get new data). 1548 * 1549 * This is implemented in the callers, where they modify the 'nonagle' 1550 * argument based upon the location of SKB in the send queue. 1551 */ 1552 if (nonagle & TCP_NAGLE_PUSH) 1553 return true; 1554 1555 /* Don't use the nagle rule for urgent data (or for the final FIN). */ 1556 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1557 return true; 1558 1559 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1560 return true; 1561 1562 return false; 1563 } 1564 1565 /* Does at least the first segment of SKB fit into the send window? */ 1566 static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1567 const struct sk_buff *skb, 1568 unsigned int cur_mss) 1569 { 1570 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1571 1572 if (skb->len > cur_mss) 1573 end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1574 1575 return !after(end_seq, tcp_wnd_end(tp)); 1576 } 1577 1578 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1579 * should be put on the wire right now. If so, it returns the number of 1580 * packets allowed by the congestion window. 1581 */ 1582 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1583 unsigned int cur_mss, int nonagle) 1584 { 1585 const struct tcp_sock *tp = tcp_sk(sk); 1586 unsigned int cwnd_quota; 1587 1588 tcp_init_tso_segs(sk, skb, cur_mss); 1589 1590 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1591 return 0; 1592 1593 cwnd_quota = tcp_cwnd_test(tp, skb); 1594 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1595 cwnd_quota = 0; 1596 1597 return cwnd_quota; 1598 } 1599 1600 /* Test if sending is allowed right now. */ 1601 bool tcp_may_send_now(struct sock *sk) 1602 { 1603 const struct tcp_sock *tp = tcp_sk(sk); 1604 struct sk_buff *skb = tcp_send_head(sk); 1605 1606 return skb && 1607 tcp_snd_test(sk, skb, tcp_current_mss(sk), 1608 (tcp_skb_is_last(sk, skb) ? 1609 tp->nonagle : TCP_NAGLE_PUSH)); 1610 } 1611 1612 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1613 * which is put after SKB on the list. It is very much like 1614 * tcp_fragment() except that it may make several kinds of assumptions 1615 * in order to speed up the splitting operation. In particular, we 1616 * know that all the data is in scatter-gather pages, and that the 1617 * packet has never been sent out before (and thus is not cloned). 1618 */ 1619 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1620 unsigned int mss_now, gfp_t gfp) 1621 { 1622 struct sk_buff *buff; 1623 int nlen = skb->len - len; 1624 u8 flags; 1625 1626 /* All of a TSO frame must be composed of paged data. */ 1627 if (skb->len != skb->data_len) 1628 return tcp_fragment(sk, skb, len, mss_now, gfp); 1629 1630 buff = sk_stream_alloc_skb(sk, 0, gfp); 1631 if (unlikely(buff == NULL)) 1632 return -ENOMEM; 1633 1634 sk->sk_wmem_queued += buff->truesize; 1635 sk_mem_charge(sk, buff->truesize); 1636 buff->truesize += nlen; 1637 skb->truesize -= nlen; 1638 1639 /* Correct the sequence numbers. */ 1640 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1641 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1642 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1643 1644 /* PSH and FIN should only be set in the second packet. */ 1645 flags = TCP_SKB_CB(skb)->tcp_flags; 1646 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 1647 TCP_SKB_CB(buff)->tcp_flags = flags; 1648 1649 /* This packet was never sent out yet, so no SACK bits. */ 1650 TCP_SKB_CB(buff)->sacked = 0; 1651 1652 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1653 skb_split(skb, buff, len); 1654 1655 /* Fix up tso_factor for both original and new SKB. */ 1656 tcp_set_skb_tso_segs(sk, skb, mss_now); 1657 tcp_set_skb_tso_segs(sk, buff, mss_now); 1658 1659 /* Link BUFF into the send queue. */ 1660 skb_header_release(buff); 1661 tcp_insert_write_queue_after(skb, buff, sk); 1662 1663 return 0; 1664 } 1665 1666 /* Try to defer sending, if possible, in order to minimize the amount 1667 * of TSO splitting we do. View it as a kind of TSO Nagle test. 1668 * 1669 * This algorithm is from John Heffner. 1670 */ 1671 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1672 bool *is_cwnd_limited) 1673 { 1674 struct tcp_sock *tp = tcp_sk(sk); 1675 const struct inet_connection_sock *icsk = inet_csk(sk); 1676 u32 send_win, cong_win, limit, in_flight; 1677 int win_divisor; 1678 1679 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1680 goto send_now; 1681 1682 if (icsk->icsk_ca_state != TCP_CA_Open) 1683 goto send_now; 1684 1685 /* Defer for less than two clock ticks. */ 1686 if (tp->tso_deferred && 1687 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1688 goto send_now; 1689 1690 in_flight = tcp_packets_in_flight(tp); 1691 1692 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1693 1694 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1695 1696 /* From in_flight test above, we know that cwnd > in_flight. */ 1697 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1698 1699 limit = min(send_win, cong_win); 1700 1701 /* If a full-sized TSO skb can be sent, do it. */ 1702 if (limit >= min_t(unsigned int, sk->sk_gso_max_size, 1703 tp->xmit_size_goal_segs * tp->mss_cache)) 1704 goto send_now; 1705 1706 /* Middle in queue won't get any more data, full sendable already? */ 1707 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 1708 goto send_now; 1709 1710 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1711 if (win_divisor) { 1712 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1713 1714 /* If at least some fraction of a window is available, 1715 * just use it. 1716 */ 1717 chunk /= win_divisor; 1718 if (limit >= chunk) 1719 goto send_now; 1720 } else { 1721 /* Different approach, try not to defer past a single 1722 * ACK. Receiver should ACK every other full sized 1723 * frame, so if we have space for more than 3 frames 1724 * then send now. 1725 */ 1726 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1727 goto send_now; 1728 } 1729 1730 /* Ok, it looks like it is advisable to defer. 1731 * Do not rearm the timer if already set to not break TCP ACK clocking. 1732 */ 1733 if (!tp->tso_deferred) 1734 tp->tso_deferred = 1 | (jiffies << 1); 1735 1736 if (cong_win < send_win && cong_win < skb->len) 1737 *is_cwnd_limited = true; 1738 1739 return true; 1740 1741 send_now: 1742 tp->tso_deferred = 0; 1743 return false; 1744 } 1745 1746 /* Create a new MTU probe if we are ready. 1747 * MTU probe is regularly attempting to increase the path MTU by 1748 * deliberately sending larger packets. This discovers routing 1749 * changes resulting in larger path MTUs. 1750 * 1751 * Returns 0 if we should wait to probe (no cwnd available), 1752 * 1 if a probe was sent, 1753 * -1 otherwise 1754 */ 1755 static int tcp_mtu_probe(struct sock *sk) 1756 { 1757 struct tcp_sock *tp = tcp_sk(sk); 1758 struct inet_connection_sock *icsk = inet_csk(sk); 1759 struct sk_buff *skb, *nskb, *next; 1760 int len; 1761 int probe_size; 1762 int size_needed; 1763 int copy; 1764 int mss_now; 1765 1766 /* Not currently probing/verifying, 1767 * not in recovery, 1768 * have enough cwnd, and 1769 * not SACKing (the variable headers throw things off) */ 1770 if (!icsk->icsk_mtup.enabled || 1771 icsk->icsk_mtup.probe_size || 1772 inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 1773 tp->snd_cwnd < 11 || 1774 tp->rx_opt.num_sacks || tp->rx_opt.dsack) 1775 return -1; 1776 1777 /* Very simple search strategy: just double the MSS. */ 1778 mss_now = tcp_current_mss(sk); 1779 probe_size = 2 * tp->mss_cache; 1780 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 1781 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 1782 /* TODO: set timer for probe_converge_event */ 1783 return -1; 1784 } 1785 1786 /* Have enough data in the send queue to probe? */ 1787 if (tp->write_seq - tp->snd_nxt < size_needed) 1788 return -1; 1789 1790 if (tp->snd_wnd < size_needed) 1791 return -1; 1792 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 1793 return 0; 1794 1795 /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1796 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1797 if (!tcp_packets_in_flight(tp)) 1798 return -1; 1799 else 1800 return 0; 1801 } 1802 1803 /* We're allowed to probe. Build it now. */ 1804 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 1805 return -1; 1806 sk->sk_wmem_queued += nskb->truesize; 1807 sk_mem_charge(sk, nskb->truesize); 1808 1809 skb = tcp_send_head(sk); 1810 1811 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 1812 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 1813 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 1814 TCP_SKB_CB(nskb)->sacked = 0; 1815 nskb->csum = 0; 1816 nskb->ip_summed = skb->ip_summed; 1817 1818 tcp_insert_write_queue_before(nskb, skb, sk); 1819 1820 len = 0; 1821 tcp_for_write_queue_from_safe(skb, next, sk) { 1822 copy = min_t(int, skb->len, probe_size - len); 1823 if (nskb->ip_summed) 1824 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 1825 else 1826 nskb->csum = skb_copy_and_csum_bits(skb, 0, 1827 skb_put(nskb, copy), 1828 copy, nskb->csum); 1829 1830 if (skb->len <= copy) { 1831 /* We've eaten all the data from this skb. 1832 * Throw it away. */ 1833 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1834 tcp_unlink_write_queue(skb, sk); 1835 sk_wmem_free_skb(sk, skb); 1836 } else { 1837 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1838 ~(TCPHDR_FIN|TCPHDR_PSH); 1839 if (!skb_shinfo(skb)->nr_frags) { 1840 skb_pull(skb, copy); 1841 if (skb->ip_summed != CHECKSUM_PARTIAL) 1842 skb->csum = csum_partial(skb->data, 1843 skb->len, 0); 1844 } else { 1845 __pskb_trim_head(skb, copy); 1846 tcp_set_skb_tso_segs(sk, skb, mss_now); 1847 } 1848 TCP_SKB_CB(skb)->seq += copy; 1849 } 1850 1851 len += copy; 1852 1853 if (len >= probe_size) 1854 break; 1855 } 1856 tcp_init_tso_segs(sk, nskb, nskb->len); 1857 1858 /* We're ready to send. If this fails, the probe will 1859 * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 1860 TCP_SKB_CB(nskb)->when = tcp_time_stamp; 1861 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 1862 /* Decrement cwnd here because we are sending 1863 * effectively two packets. */ 1864 tp->snd_cwnd--; 1865 tcp_event_new_data_sent(sk, nskb); 1866 1867 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 1868 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 1869 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 1870 1871 return 1; 1872 } 1873 1874 return -1; 1875 } 1876 1877 /* This routine writes packets to the network. It advances the 1878 * send_head. This happens as incoming acks open up the remote 1879 * window for us. 1880 * 1881 * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1882 * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1883 * account rare use of URG, this is not a big flaw. 1884 * 1885 * Send at most one packet when push_one > 0. Temporarily ignore 1886 * cwnd limit to force at most one packet out when push_one == 2. 1887 1888 * Returns true, if no segments are in flight and we have queued segments, 1889 * but cannot send anything now because of SWS or another problem. 1890 */ 1891 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1892 int push_one, gfp_t gfp) 1893 { 1894 struct tcp_sock *tp = tcp_sk(sk); 1895 struct sk_buff *skb; 1896 unsigned int tso_segs, sent_pkts; 1897 int cwnd_quota; 1898 int result; 1899 bool is_cwnd_limited = false; 1900 1901 sent_pkts = 0; 1902 1903 if (!push_one) { 1904 /* Do MTU probing. */ 1905 result = tcp_mtu_probe(sk); 1906 if (!result) { 1907 return false; 1908 } else if (result > 0) { 1909 sent_pkts = 1; 1910 } 1911 } 1912 1913 while ((skb = tcp_send_head(sk))) { 1914 unsigned int limit; 1915 1916 tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1917 BUG_ON(!tso_segs); 1918 1919 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) 1920 goto repair; /* Skip network transmission */ 1921 1922 cwnd_quota = tcp_cwnd_test(tp, skb); 1923 if (!cwnd_quota) { 1924 is_cwnd_limited = true; 1925 if (push_one == 2) 1926 /* Force out a loss probe pkt. */ 1927 cwnd_quota = 1; 1928 else 1929 break; 1930 } 1931 1932 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1933 break; 1934 1935 if (tso_segs == 1) { 1936 if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1937 (tcp_skb_is_last(sk, skb) ? 1938 nonagle : TCP_NAGLE_PUSH)))) 1939 break; 1940 } else { 1941 if (!push_one && 1942 tcp_tso_should_defer(sk, skb, &is_cwnd_limited)) 1943 break; 1944 } 1945 1946 /* TCP Small Queues : 1947 * Control number of packets in qdisc/devices to two packets / or ~1 ms. 1948 * This allows for : 1949 * - better RTT estimation and ACK scheduling 1950 * - faster recovery 1951 * - high rates 1952 * Alas, some drivers / subsystems require a fair amount 1953 * of queued bytes to ensure line rate. 1954 * One example is wifi aggregation (802.11 AMPDU) 1955 */ 1956 limit = max_t(unsigned int, sysctl_tcp_limit_output_bytes, 1957 sk->sk_pacing_rate >> 10); 1958 1959 if (atomic_read(&sk->sk_wmem_alloc) > limit) { 1960 set_bit(TSQ_THROTTLED, &tp->tsq_flags); 1961 /* It is possible TX completion already happened 1962 * before we set TSQ_THROTTLED, so we must 1963 * test again the condition. 1964 */ 1965 smp_mb__after_atomic(); 1966 if (atomic_read(&sk->sk_wmem_alloc) > limit) 1967 break; 1968 } 1969 1970 limit = mss_now; 1971 if (tso_segs > 1 && !tcp_urg_mode(tp)) 1972 limit = tcp_mss_split_point(sk, skb, mss_now, 1973 min_t(unsigned int, 1974 cwnd_quota, 1975 sk->sk_gso_max_segs), 1976 nonagle); 1977 1978 if (skb->len > limit && 1979 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 1980 break; 1981 1982 TCP_SKB_CB(skb)->when = tcp_time_stamp; 1983 1984 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 1985 break; 1986 1987 repair: 1988 /* Advance the send_head. This one is sent out. 1989 * This call will increment packets_out. 1990 */ 1991 tcp_event_new_data_sent(sk, skb); 1992 1993 tcp_minshall_update(tp, mss_now, skb); 1994 sent_pkts += tcp_skb_pcount(skb); 1995 1996 if (push_one) 1997 break; 1998 } 1999 2000 if (likely(sent_pkts)) { 2001 if (tcp_in_cwnd_reduction(sk)) 2002 tp->prr_out += sent_pkts; 2003 2004 /* Send one loss probe per tail loss episode. */ 2005 if (push_one != 2) 2006 tcp_schedule_loss_probe(sk); 2007 tcp_cwnd_validate(sk, is_cwnd_limited); 2008 return false; 2009 } 2010 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk)); 2011 } 2012 2013 bool tcp_schedule_loss_probe(struct sock *sk) 2014 { 2015 struct inet_connection_sock *icsk = inet_csk(sk); 2016 struct tcp_sock *tp = tcp_sk(sk); 2017 u32 timeout, tlp_time_stamp, rto_time_stamp; 2018 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3); 2019 2020 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) 2021 return false; 2022 /* No consecutive loss probes. */ 2023 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { 2024 tcp_rearm_rto(sk); 2025 return false; 2026 } 2027 /* Don't do any loss probe on a Fast Open connection before 3WHS 2028 * finishes. 2029 */ 2030 if (sk->sk_state == TCP_SYN_RECV) 2031 return false; 2032 2033 /* TLP is only scheduled when next timer event is RTO. */ 2034 if (icsk->icsk_pending != ICSK_TIME_RETRANS) 2035 return false; 2036 2037 /* Schedule a loss probe in 2*RTT for SACK capable connections 2038 * in Open state, that are either limited by cwnd or application. 2039 */ 2040 if (sysctl_tcp_early_retrans < 3 || !tp->srtt_us || !tp->packets_out || 2041 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 2042 return false; 2043 2044 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 2045 tcp_send_head(sk)) 2046 return false; 2047 2048 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account 2049 * for delayed ack when there's one outstanding packet. 2050 */ 2051 timeout = rtt << 1; 2052 if (tp->packets_out == 1) 2053 timeout = max_t(u32, timeout, 2054 (rtt + (rtt >> 1) + TCP_DELACK_MAX)); 2055 timeout = max_t(u32, timeout, msecs_to_jiffies(10)); 2056 2057 /* If RTO is shorter, just schedule TLP in its place. */ 2058 tlp_time_stamp = tcp_time_stamp + timeout; 2059 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; 2060 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { 2061 s32 delta = rto_time_stamp - tcp_time_stamp; 2062 if (delta > 0) 2063 timeout = delta; 2064 } 2065 2066 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 2067 TCP_RTO_MAX); 2068 return true; 2069 } 2070 2071 /* Thanks to skb fast clones, we can detect if a prior transmit of 2072 * a packet is still in a qdisc or driver queue. 2073 * In this case, there is very little point doing a retransmit ! 2074 * Note: This is called from BH context only. 2075 */ 2076 static bool skb_still_in_host_queue(const struct sock *sk, 2077 const struct sk_buff *skb) 2078 { 2079 const struct sk_buff *fclone = skb + 1; 2080 2081 if (unlikely(skb->fclone == SKB_FCLONE_ORIG && 2082 fclone->fclone == SKB_FCLONE_CLONE)) { 2083 NET_INC_STATS_BH(sock_net(sk), 2084 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 2085 return true; 2086 } 2087 return false; 2088 } 2089 2090 /* When probe timeout (PTO) fires, send a new segment if one exists, else 2091 * retransmit the last segment. 2092 */ 2093 void tcp_send_loss_probe(struct sock *sk) 2094 { 2095 struct tcp_sock *tp = tcp_sk(sk); 2096 struct sk_buff *skb; 2097 int pcount; 2098 int mss = tcp_current_mss(sk); 2099 int err = -1; 2100 2101 if (tcp_send_head(sk) != NULL) { 2102 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2103 goto rearm_timer; 2104 } 2105 2106 /* At most one outstanding TLP retransmission. */ 2107 if (tp->tlp_high_seq) 2108 goto rearm_timer; 2109 2110 /* Retransmit last segment. */ 2111 skb = tcp_write_queue_tail(sk); 2112 if (WARN_ON(!skb)) 2113 goto rearm_timer; 2114 2115 if (skb_still_in_host_queue(sk, skb)) 2116 goto rearm_timer; 2117 2118 pcount = tcp_skb_pcount(skb); 2119 if (WARN_ON(!pcount)) 2120 goto rearm_timer; 2121 2122 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 2123 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss, 2124 GFP_ATOMIC))) 2125 goto rearm_timer; 2126 skb = tcp_write_queue_tail(sk); 2127 } 2128 2129 if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 2130 goto rearm_timer; 2131 2132 err = __tcp_retransmit_skb(sk, skb); 2133 2134 /* Record snd_nxt for loss detection. */ 2135 if (likely(!err)) 2136 tp->tlp_high_seq = tp->snd_nxt; 2137 2138 rearm_timer: 2139 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2140 inet_csk(sk)->icsk_rto, 2141 TCP_RTO_MAX); 2142 2143 if (likely(!err)) 2144 NET_INC_STATS_BH(sock_net(sk), 2145 LINUX_MIB_TCPLOSSPROBES); 2146 } 2147 2148 /* Push out any pending frames which were held back due to 2149 * TCP_CORK or attempt at coalescing tiny packets. 2150 * The socket must be locked by the caller. 2151 */ 2152 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 2153 int nonagle) 2154 { 2155 /* If we are closed, the bytes will have to remain here. 2156 * In time closedown will finish, we empty the write queue and 2157 * all will be happy. 2158 */ 2159 if (unlikely(sk->sk_state == TCP_CLOSE)) 2160 return; 2161 2162 if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 2163 sk_gfp_atomic(sk, GFP_ATOMIC))) 2164 tcp_check_probe_timer(sk); 2165 } 2166 2167 /* Send _single_ skb sitting at the send head. This function requires 2168 * true push pending frames to setup probe timer etc. 2169 */ 2170 void tcp_push_one(struct sock *sk, unsigned int mss_now) 2171 { 2172 struct sk_buff *skb = tcp_send_head(sk); 2173 2174 BUG_ON(!skb || skb->len < mss_now); 2175 2176 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2177 } 2178 2179 /* This function returns the amount that we can raise the 2180 * usable window based on the following constraints 2181 * 2182 * 1. The window can never be shrunk once it is offered (RFC 793) 2183 * 2. We limit memory per socket 2184 * 2185 * RFC 1122: 2186 * "the suggested [SWS] avoidance algorithm for the receiver is to keep 2187 * RECV.NEXT + RCV.WIN fixed until: 2188 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 2189 * 2190 * i.e. don't raise the right edge of the window until you can raise 2191 * it at least MSS bytes. 2192 * 2193 * Unfortunately, the recommended algorithm breaks header prediction, 2194 * since header prediction assumes th->window stays fixed. 2195 * 2196 * Strictly speaking, keeping th->window fixed violates the receiver 2197 * side SWS prevention criteria. The problem is that under this rule 2198 * a stream of single byte packets will cause the right side of the 2199 * window to always advance by a single byte. 2200 * 2201 * Of course, if the sender implements sender side SWS prevention 2202 * then this will not be a problem. 2203 * 2204 * BSD seems to make the following compromise: 2205 * 2206 * If the free space is less than the 1/4 of the maximum 2207 * space available and the free space is less than 1/2 mss, 2208 * then set the window to 0. 2209 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 2210 * Otherwise, just prevent the window from shrinking 2211 * and from being larger than the largest representable value. 2212 * 2213 * This prevents incremental opening of the window in the regime 2214 * where TCP is limited by the speed of the reader side taking 2215 * data out of the TCP receive queue. It does nothing about 2216 * those cases where the window is constrained on the sender side 2217 * because the pipeline is full. 2218 * 2219 * BSD also seems to "accidentally" limit itself to windows that are a 2220 * multiple of MSS, at least until the free space gets quite small. 2221 * This would appear to be a side effect of the mbuf implementation. 2222 * Combining these two algorithms results in the observed behavior 2223 * of having a fixed window size at almost all times. 2224 * 2225 * Below we obtain similar behavior by forcing the offered window to 2226 * a multiple of the mss when it is feasible to do so. 2227 * 2228 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 2229 * Regular options like TIMESTAMP are taken into account. 2230 */ 2231 u32 __tcp_select_window(struct sock *sk) 2232 { 2233 struct inet_connection_sock *icsk = inet_csk(sk); 2234 struct tcp_sock *tp = tcp_sk(sk); 2235 /* MSS for the peer's data. Previous versions used mss_clamp 2236 * here. I don't know if the value based on our guesses 2237 * of peer's MSS is better for the performance. It's more correct 2238 * but may be worse for the performance because of rcv_mss 2239 * fluctuations. --SAW 1998/11/1 2240 */ 2241 int mss = icsk->icsk_ack.rcv_mss; 2242 int free_space = tcp_space(sk); 2243 int allowed_space = tcp_full_space(sk); 2244 int full_space = min_t(int, tp->window_clamp, allowed_space); 2245 int window; 2246 2247 if (mss > full_space) 2248 mss = full_space; 2249 2250 if (free_space < (full_space >> 1)) { 2251 icsk->icsk_ack.quick = 0; 2252 2253 if (sk_under_memory_pressure(sk)) 2254 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2255 4U * tp->advmss); 2256 2257 /* free_space might become our new window, make sure we don't 2258 * increase it due to wscale. 2259 */ 2260 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 2261 2262 /* if free space is less than mss estimate, or is below 1/16th 2263 * of the maximum allowed, try to move to zero-window, else 2264 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 2265 * new incoming data is dropped due to memory limits. 2266 * With large window, mss test triggers way too late in order 2267 * to announce zero window in time before rmem limit kicks in. 2268 */ 2269 if (free_space < (allowed_space >> 4) || free_space < mss) 2270 return 0; 2271 } 2272 2273 if (free_space > tp->rcv_ssthresh) 2274 free_space = tp->rcv_ssthresh; 2275 2276 /* Don't do rounding if we are using window scaling, since the 2277 * scaled window will not line up with the MSS boundary anyway. 2278 */ 2279 window = tp->rcv_wnd; 2280 if (tp->rx_opt.rcv_wscale) { 2281 window = free_space; 2282 2283 /* Advertise enough space so that it won't get scaled away. 2284 * Import case: prevent zero window announcement if 2285 * 1<<rcv_wscale > mss. 2286 */ 2287 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 2288 window = (((window >> tp->rx_opt.rcv_wscale) + 1) 2289 << tp->rx_opt.rcv_wscale); 2290 } else { 2291 /* Get the largest window that is a nice multiple of mss. 2292 * Window clamp already applied above. 2293 * If our current window offering is within 1 mss of the 2294 * free space we just keep it. This prevents the divide 2295 * and multiply from happening most of the time. 2296 * We also don't do any window rounding when the free space 2297 * is too small. 2298 */ 2299 if (window <= free_space - mss || window > free_space) 2300 window = (free_space / mss) * mss; 2301 else if (mss == full_space && 2302 free_space > window + (full_space >> 1)) 2303 window = free_space; 2304 } 2305 2306 return window; 2307 } 2308 2309 /* Collapses two adjacent SKB's during retransmission. */ 2310 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 2311 { 2312 struct tcp_sock *tp = tcp_sk(sk); 2313 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2314 int skb_size, next_skb_size; 2315 2316 skb_size = skb->len; 2317 next_skb_size = next_skb->len; 2318 2319 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 2320 2321 tcp_highest_sack_combine(sk, next_skb, skb); 2322 2323 tcp_unlink_write_queue(next_skb, sk); 2324 2325 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 2326 next_skb_size); 2327 2328 if (next_skb->ip_summed == CHECKSUM_PARTIAL) 2329 skb->ip_summed = CHECKSUM_PARTIAL; 2330 2331 if (skb->ip_summed != CHECKSUM_PARTIAL) 2332 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 2333 2334 /* Update sequence range on original skb. */ 2335 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 2336 2337 /* Merge over control information. This moves PSH/FIN etc. over */ 2338 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 2339 2340 /* All done, get rid of second SKB and account for it so 2341 * packet counting does not break. 2342 */ 2343 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2344 2345 /* changed transmit queue under us so clear hints */ 2346 tcp_clear_retrans_hints_partial(tp); 2347 if (next_skb == tp->retransmit_skb_hint) 2348 tp->retransmit_skb_hint = skb; 2349 2350 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2351 2352 sk_wmem_free_skb(sk, next_skb); 2353 } 2354 2355 /* Check if coalescing SKBs is legal. */ 2356 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 2357 { 2358 if (tcp_skb_pcount(skb) > 1) 2359 return false; 2360 /* TODO: SACK collapsing could be used to remove this condition */ 2361 if (skb_shinfo(skb)->nr_frags != 0) 2362 return false; 2363 if (skb_cloned(skb)) 2364 return false; 2365 if (skb == tcp_send_head(sk)) 2366 return false; 2367 /* Some heurestics for collapsing over SACK'd could be invented */ 2368 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2369 return false; 2370 2371 return true; 2372 } 2373 2374 /* Collapse packets in the retransmit queue to make to create 2375 * less packets on the wire. This is only done on retransmission. 2376 */ 2377 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 2378 int space) 2379 { 2380 struct tcp_sock *tp = tcp_sk(sk); 2381 struct sk_buff *skb = to, *tmp; 2382 bool first = true; 2383 2384 if (!sysctl_tcp_retrans_collapse) 2385 return; 2386 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2387 return; 2388 2389 tcp_for_write_queue_from_safe(skb, tmp, sk) { 2390 if (!tcp_can_collapse(sk, skb)) 2391 break; 2392 2393 space -= skb->len; 2394 2395 if (first) { 2396 first = false; 2397 continue; 2398 } 2399 2400 if (space < 0) 2401 break; 2402 /* Punt if not enough space exists in the first SKB for 2403 * the data in the second 2404 */ 2405 if (skb->len > skb_availroom(to)) 2406 break; 2407 2408 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 2409 break; 2410 2411 tcp_collapse_retrans(sk, to); 2412 } 2413 } 2414 2415 /* This retransmits one SKB. Policy decisions and retransmit queue 2416 * state updates are done by the caller. Returns non-zero if an 2417 * error occurred which prevented the send. 2418 */ 2419 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 2420 { 2421 struct tcp_sock *tp = tcp_sk(sk); 2422 struct inet_connection_sock *icsk = inet_csk(sk); 2423 unsigned int cur_mss; 2424 int err; 2425 2426 /* Inconslusive MTU probe */ 2427 if (icsk->icsk_mtup.probe_size) { 2428 icsk->icsk_mtup.probe_size = 0; 2429 } 2430 2431 /* Do not sent more than we queued. 1/4 is reserved for possible 2432 * copying overhead: fragmentation, tunneling, mangling etc. 2433 */ 2434 if (atomic_read(&sk->sk_wmem_alloc) > 2435 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 2436 return -EAGAIN; 2437 2438 if (skb_still_in_host_queue(sk, skb)) 2439 return -EBUSY; 2440 2441 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 2442 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 2443 BUG(); 2444 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 2445 return -ENOMEM; 2446 } 2447 2448 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 2449 return -EHOSTUNREACH; /* Routing failure or similar. */ 2450 2451 cur_mss = tcp_current_mss(sk); 2452 2453 /* If receiver has shrunk his window, and skb is out of 2454 * new window, do not retransmit it. The exception is the 2455 * case, when window is shrunk to zero. In this case 2456 * our retransmit serves as a zero window probe. 2457 */ 2458 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 2459 TCP_SKB_CB(skb)->seq != tp->snd_una) 2460 return -EAGAIN; 2461 2462 if (skb->len > cur_mss) { 2463 if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC)) 2464 return -ENOMEM; /* We'll try again later. */ 2465 } else { 2466 int oldpcount = tcp_skb_pcount(skb); 2467 2468 if (unlikely(oldpcount > 1)) { 2469 if (skb_unclone(skb, GFP_ATOMIC)) 2470 return -ENOMEM; 2471 tcp_init_tso_segs(sk, skb, cur_mss); 2472 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 2473 } 2474 } 2475 2476 tcp_retrans_try_collapse(sk, skb, cur_mss); 2477 2478 /* Make a copy, if the first transmission SKB clone we made 2479 * is still in somebody's hands, else make a clone. 2480 */ 2481 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2482 2483 /* make sure skb->data is aligned on arches that require it 2484 * and check if ack-trimming & collapsing extended the headroom 2485 * beyond what csum_start can cover. 2486 */ 2487 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 2488 skb_headroom(skb) >= 0xFFFF)) { 2489 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, 2490 GFP_ATOMIC); 2491 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2492 -ENOBUFS; 2493 } else { 2494 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2495 } 2496 2497 if (likely(!err)) { 2498 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2499 /* Update global TCP statistics. */ 2500 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 2501 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2502 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2503 tp->total_retrans++; 2504 } 2505 return err; 2506 } 2507 2508 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 2509 { 2510 struct tcp_sock *tp = tcp_sk(sk); 2511 int err = __tcp_retransmit_skb(sk, skb); 2512 2513 if (err == 0) { 2514 #if FASTRETRANS_DEBUG > 0 2515 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2516 net_dbg_ratelimited("retrans_out leaked\n"); 2517 } 2518 #endif 2519 if (!tp->retrans_out) 2520 tp->lost_retrans_low = tp->snd_nxt; 2521 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 2522 tp->retrans_out += tcp_skb_pcount(skb); 2523 2524 /* Save stamp of the first retransmit. */ 2525 if (!tp->retrans_stamp) 2526 tp->retrans_stamp = TCP_SKB_CB(skb)->when; 2527 2528 tp->undo_retrans += tcp_skb_pcount(skb); 2529 2530 /* snd_nxt is stored to detect loss of retransmitted segment, 2531 * see tcp_input.c tcp_sacktag_write_queue(). 2532 */ 2533 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 2534 } else if (err != -EBUSY) { 2535 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 2536 } 2537 return err; 2538 } 2539 2540 /* Check if we forward retransmits are possible in the current 2541 * window/congestion state. 2542 */ 2543 static bool tcp_can_forward_retransmit(struct sock *sk) 2544 { 2545 const struct inet_connection_sock *icsk = inet_csk(sk); 2546 const struct tcp_sock *tp = tcp_sk(sk); 2547 2548 /* Forward retransmissions are possible only during Recovery. */ 2549 if (icsk->icsk_ca_state != TCP_CA_Recovery) 2550 return false; 2551 2552 /* No forward retransmissions in Reno are possible. */ 2553 if (tcp_is_reno(tp)) 2554 return false; 2555 2556 /* Yeah, we have to make difficult choice between forward transmission 2557 * and retransmission... Both ways have their merits... 2558 * 2559 * For now we do not retransmit anything, while we have some new 2560 * segments to send. In the other cases, follow rule 3 for 2561 * NextSeg() specified in RFC3517. 2562 */ 2563 2564 if (tcp_may_send_now(sk)) 2565 return false; 2566 2567 return true; 2568 } 2569 2570 /* This gets called after a retransmit timeout, and the initially 2571 * retransmitted data is acknowledged. It tries to continue 2572 * resending the rest of the retransmit queue, until either 2573 * we've sent it all or the congestion window limit is reached. 2574 * If doing SACK, the first ACK which comes back for a timeout 2575 * based retransmit packet might feed us FACK information again. 2576 * If so, we use it to avoid unnecessarily retransmissions. 2577 */ 2578 void tcp_xmit_retransmit_queue(struct sock *sk) 2579 { 2580 const struct inet_connection_sock *icsk = inet_csk(sk); 2581 struct tcp_sock *tp = tcp_sk(sk); 2582 struct sk_buff *skb; 2583 struct sk_buff *hole = NULL; 2584 u32 last_lost; 2585 int mib_idx; 2586 int fwd_rexmitting = 0; 2587 2588 if (!tp->packets_out) 2589 return; 2590 2591 if (!tp->lost_out) 2592 tp->retransmit_high = tp->snd_una; 2593 2594 if (tp->retransmit_skb_hint) { 2595 skb = tp->retransmit_skb_hint; 2596 last_lost = TCP_SKB_CB(skb)->end_seq; 2597 if (after(last_lost, tp->retransmit_high)) 2598 last_lost = tp->retransmit_high; 2599 } else { 2600 skb = tcp_write_queue_head(sk); 2601 last_lost = tp->snd_una; 2602 } 2603 2604 tcp_for_write_queue_from(skb, sk) { 2605 __u8 sacked = TCP_SKB_CB(skb)->sacked; 2606 2607 if (skb == tcp_send_head(sk)) 2608 break; 2609 /* we could do better than to assign each time */ 2610 if (hole == NULL) 2611 tp->retransmit_skb_hint = skb; 2612 2613 /* Assume this retransmit will generate 2614 * only one packet for congestion window 2615 * calculation purposes. This works because 2616 * tcp_retransmit_skb() will chop up the 2617 * packet to be MSS sized and all the 2618 * packet counting works out. 2619 */ 2620 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 2621 return; 2622 2623 if (fwd_rexmitting) { 2624 begin_fwd: 2625 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2626 break; 2627 mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 2628 2629 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2630 tp->retransmit_high = last_lost; 2631 if (!tcp_can_forward_retransmit(sk)) 2632 break; 2633 /* Backtrack if necessary to non-L'ed skb */ 2634 if (hole != NULL) { 2635 skb = hole; 2636 hole = NULL; 2637 } 2638 fwd_rexmitting = 1; 2639 goto begin_fwd; 2640 2641 } else if (!(sacked & TCPCB_LOST)) { 2642 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 2643 hole = skb; 2644 continue; 2645 2646 } else { 2647 last_lost = TCP_SKB_CB(skb)->end_seq; 2648 if (icsk->icsk_ca_state != TCP_CA_Loss) 2649 mib_idx = LINUX_MIB_TCPFASTRETRANS; 2650 else 2651 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 2652 } 2653 2654 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 2655 continue; 2656 2657 if (tcp_retransmit_skb(sk, skb)) 2658 return; 2659 2660 NET_INC_STATS_BH(sock_net(sk), mib_idx); 2661 2662 if (tcp_in_cwnd_reduction(sk)) 2663 tp->prr_out += tcp_skb_pcount(skb); 2664 2665 if (skb == tcp_write_queue_head(sk)) 2666 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2667 inet_csk(sk)->icsk_rto, 2668 TCP_RTO_MAX); 2669 } 2670 } 2671 2672 /* Send a fin. The caller locks the socket for us. This cannot be 2673 * allowed to fail queueing a FIN frame under any circumstances. 2674 */ 2675 void tcp_send_fin(struct sock *sk) 2676 { 2677 struct tcp_sock *tp = tcp_sk(sk); 2678 struct sk_buff *skb = tcp_write_queue_tail(sk); 2679 int mss_now; 2680 2681 /* Optimization, tack on the FIN if we have a queue of 2682 * unsent frames. But be careful about outgoing SACKS 2683 * and IP options. 2684 */ 2685 mss_now = tcp_current_mss(sk); 2686 2687 if (tcp_send_head(sk) != NULL) { 2688 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; 2689 TCP_SKB_CB(skb)->end_seq++; 2690 tp->write_seq++; 2691 } else { 2692 /* Socket is locked, keep trying until memory is available. */ 2693 for (;;) { 2694 skb = alloc_skb_fclone(MAX_TCP_HEADER, 2695 sk->sk_allocation); 2696 if (skb) 2697 break; 2698 yield(); 2699 } 2700 2701 /* Reserve space for headers and prepare control bits. */ 2702 skb_reserve(skb, MAX_TCP_HEADER); 2703 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2704 tcp_init_nondata_skb(skb, tp->write_seq, 2705 TCPHDR_ACK | TCPHDR_FIN); 2706 tcp_queue_skb(sk, skb); 2707 } 2708 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 2709 } 2710 2711 /* We get here when a process closes a file descriptor (either due to 2712 * an explicit close() or as a byproduct of exit()'ing) and there 2713 * was unread data in the receive queue. This behavior is recommended 2714 * by RFC 2525, section 2.17. -DaveM 2715 */ 2716 void tcp_send_active_reset(struct sock *sk, gfp_t priority) 2717 { 2718 struct sk_buff *skb; 2719 2720 /* NOTE: No TCP options attached and we never retransmit this. */ 2721 skb = alloc_skb(MAX_TCP_HEADER, priority); 2722 if (!skb) { 2723 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 2724 return; 2725 } 2726 2727 /* Reserve space for headers and prepare control bits. */ 2728 skb_reserve(skb, MAX_TCP_HEADER); 2729 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2730 TCPHDR_ACK | TCPHDR_RST); 2731 /* Send it off. */ 2732 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2733 if (tcp_transmit_skb(sk, skb, 0, priority)) 2734 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 2735 2736 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 2737 } 2738 2739 /* Send a crossed SYN-ACK during socket establishment. 2740 * WARNING: This routine must only be called when we have already sent 2741 * a SYN packet that crossed the incoming SYN that caused this routine 2742 * to get called. If this assumption fails then the initial rcv_wnd 2743 * and rcv_wscale values will not be correct. 2744 */ 2745 int tcp_send_synack(struct sock *sk) 2746 { 2747 struct sk_buff *skb; 2748 2749 skb = tcp_write_queue_head(sk); 2750 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 2751 pr_debug("%s: wrong queue state\n", __func__); 2752 return -EFAULT; 2753 } 2754 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 2755 if (skb_cloned(skb)) { 2756 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 2757 if (nskb == NULL) 2758 return -ENOMEM; 2759 tcp_unlink_write_queue(skb, sk); 2760 skb_header_release(nskb); 2761 __tcp_add_write_queue_head(sk, nskb); 2762 sk_wmem_free_skb(sk, skb); 2763 sk->sk_wmem_queued += nskb->truesize; 2764 sk_mem_charge(sk, nskb->truesize); 2765 skb = nskb; 2766 } 2767 2768 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 2769 TCP_ECN_send_synack(tcp_sk(sk), skb); 2770 } 2771 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2772 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2773 } 2774 2775 /** 2776 * tcp_make_synack - Prepare a SYN-ACK. 2777 * sk: listener socket 2778 * dst: dst entry attached to the SYNACK 2779 * req: request_sock pointer 2780 * 2781 * Allocate one skb and build a SYNACK packet. 2782 * @dst is consumed : Caller should not use it again. 2783 */ 2784 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2785 struct request_sock *req, 2786 struct tcp_fastopen_cookie *foc) 2787 { 2788 struct tcp_out_options opts; 2789 struct inet_request_sock *ireq = inet_rsk(req); 2790 struct tcp_sock *tp = tcp_sk(sk); 2791 struct tcphdr *th; 2792 struct sk_buff *skb; 2793 struct tcp_md5sig_key *md5; 2794 int tcp_header_size; 2795 int mss; 2796 2797 skb = sock_wmalloc(sk, MAX_TCP_HEADER, 1, GFP_ATOMIC); 2798 if (unlikely(!skb)) { 2799 dst_release(dst); 2800 return NULL; 2801 } 2802 /* Reserve space for headers. */ 2803 skb_reserve(skb, MAX_TCP_HEADER); 2804 2805 skb_dst_set(skb, dst); 2806 security_skb_owned_by(skb, sk); 2807 2808 mss = dst_metric_advmss(dst); 2809 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2810 mss = tp->rx_opt.user_mss; 2811 2812 memset(&opts, 0, sizeof(opts)); 2813 #ifdef CONFIG_SYN_COOKIES 2814 if (unlikely(req->cookie_ts)) 2815 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 2816 else 2817 #endif 2818 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2819 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5, 2820 foc) + sizeof(*th); 2821 2822 skb_push(skb, tcp_header_size); 2823 skb_reset_transport_header(skb); 2824 2825 th = tcp_hdr(skb); 2826 memset(th, 0, sizeof(struct tcphdr)); 2827 th->syn = 1; 2828 th->ack = 1; 2829 TCP_ECN_make_synack(req, th); 2830 th->source = htons(ireq->ir_num); 2831 th->dest = ireq->ir_rmt_port; 2832 /* Setting of flags are superfluous here for callers (and ECE is 2833 * not even correctly set) 2834 */ 2835 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2836 TCPHDR_SYN | TCPHDR_ACK); 2837 2838 th->seq = htonl(TCP_SKB_CB(skb)->seq); 2839 /* XXX data is queued and acked as is. No buffer/window check */ 2840 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 2841 2842 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2843 th->window = htons(min(req->rcv_wnd, 65535U)); 2844 tcp_options_write((__be32 *)(th + 1), tp, &opts); 2845 th->doff = (tcp_header_size >> 2); 2846 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS); 2847 2848 #ifdef CONFIG_TCP_MD5SIG 2849 /* Okay, we have all we need - do the md5 hash if needed */ 2850 if (md5) { 2851 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 2852 md5, NULL, req, skb); 2853 } 2854 #endif 2855 2856 return skb; 2857 } 2858 EXPORT_SYMBOL(tcp_make_synack); 2859 2860 /* Do all connect socket setups that can be done AF independent. */ 2861 static void tcp_connect_init(struct sock *sk) 2862 { 2863 const struct dst_entry *dst = __sk_dst_get(sk); 2864 struct tcp_sock *tp = tcp_sk(sk); 2865 __u8 rcv_wscale; 2866 2867 /* We'll fix this up when we get a response from the other end. 2868 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 2869 */ 2870 tp->tcp_header_len = sizeof(struct tcphdr) + 2871 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 2872 2873 #ifdef CONFIG_TCP_MD5SIG 2874 if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2875 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2876 #endif 2877 2878 /* If user gave his TCP_MAXSEG, record it to clamp */ 2879 if (tp->rx_opt.user_mss) 2880 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 2881 tp->max_window = 0; 2882 tcp_mtup_init(sk); 2883 tcp_sync_mss(sk, dst_mtu(dst)); 2884 2885 if (!tp->window_clamp) 2886 tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 2887 tp->advmss = dst_metric_advmss(dst); 2888 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2889 tp->advmss = tp->rx_opt.user_mss; 2890 2891 tcp_initialize_rcv_mss(sk); 2892 2893 /* limit the window selection if the user enforce a smaller rx buffer */ 2894 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2895 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 2896 tp->window_clamp = tcp_full_space(sk); 2897 2898 tcp_select_initial_window(tcp_full_space(sk), 2899 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 2900 &tp->rcv_wnd, 2901 &tp->window_clamp, 2902 sysctl_tcp_window_scaling, 2903 &rcv_wscale, 2904 dst_metric(dst, RTAX_INITRWND)); 2905 2906 tp->rx_opt.rcv_wscale = rcv_wscale; 2907 tp->rcv_ssthresh = tp->rcv_wnd; 2908 2909 sk->sk_err = 0; 2910 sock_reset_flag(sk, SOCK_DONE); 2911 tp->snd_wnd = 0; 2912 tcp_init_wl(tp, 0); 2913 tp->snd_una = tp->write_seq; 2914 tp->snd_sml = tp->write_seq; 2915 tp->snd_up = tp->write_seq; 2916 tp->snd_nxt = tp->write_seq; 2917 2918 if (likely(!tp->repair)) 2919 tp->rcv_nxt = 0; 2920 else 2921 tp->rcv_tstamp = tcp_time_stamp; 2922 tp->rcv_wup = tp->rcv_nxt; 2923 tp->copied_seq = tp->rcv_nxt; 2924 2925 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2926 inet_csk(sk)->icsk_retransmits = 0; 2927 tcp_clear_retrans(tp); 2928 } 2929 2930 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 2931 { 2932 struct tcp_sock *tp = tcp_sk(sk); 2933 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 2934 2935 tcb->end_seq += skb->len; 2936 skb_header_release(skb); 2937 __tcp_add_write_queue_tail(sk, skb); 2938 sk->sk_wmem_queued += skb->truesize; 2939 sk_mem_charge(sk, skb->truesize); 2940 tp->write_seq = tcb->end_seq; 2941 tp->packets_out += tcp_skb_pcount(skb); 2942 } 2943 2944 /* Build and send a SYN with data and (cached) Fast Open cookie. However, 2945 * queue a data-only packet after the regular SYN, such that regular SYNs 2946 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 2947 * only the SYN sequence, the data are retransmitted in the first ACK. 2948 * If cookie is not cached or other error occurs, falls back to send a 2949 * regular SYN with Fast Open cookie request option. 2950 */ 2951 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 2952 { 2953 struct tcp_sock *tp = tcp_sk(sk); 2954 struct tcp_fastopen_request *fo = tp->fastopen_req; 2955 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen; 2956 struct sk_buff *syn_data = NULL, *data; 2957 unsigned long last_syn_loss = 0; 2958 2959 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 2960 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, 2961 &syn_loss, &last_syn_loss); 2962 /* Recurring FO SYN losses: revert to regular handshake temporarily */ 2963 if (syn_loss > 1 && 2964 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { 2965 fo->cookie.len = -1; 2966 goto fallback; 2967 } 2968 2969 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) 2970 fo->cookie.len = -1; 2971 else if (fo->cookie.len <= 0) 2972 goto fallback; 2973 2974 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 2975 * user-MSS. Reserve maximum option space for middleboxes that add 2976 * private TCP options. The cost is reduced data space in SYN :( 2977 */ 2978 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) 2979 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 2980 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 2981 MAX_TCP_OPTION_SPACE; 2982 2983 space = min_t(size_t, space, fo->size); 2984 2985 /* limit to order-0 allocations */ 2986 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 2987 2988 syn_data = skb_copy_expand(syn, MAX_TCP_HEADER, space, 2989 sk->sk_allocation); 2990 if (syn_data == NULL) 2991 goto fallback; 2992 2993 for (i = 0; i < iovlen && syn_data->len < space; ++i) { 2994 struct iovec *iov = &fo->data->msg_iov[i]; 2995 unsigned char __user *from = iov->iov_base; 2996 int len = iov->iov_len; 2997 2998 if (syn_data->len + len > space) 2999 len = space - syn_data->len; 3000 else if (i + 1 == iovlen) 3001 /* No more data pending in inet_wait_for_connect() */ 3002 fo->data = NULL; 3003 3004 if (skb_add_data(syn_data, from, len)) 3005 goto fallback; 3006 } 3007 3008 /* Queue a data-only packet after the regular SYN for retransmission */ 3009 data = pskb_copy(syn_data, sk->sk_allocation); 3010 if (data == NULL) 3011 goto fallback; 3012 TCP_SKB_CB(data)->seq++; 3013 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN; 3014 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH); 3015 tcp_connect_queue_skb(sk, data); 3016 fo->copied = data->len; 3017 3018 /* syn_data is about to be sent, we need to take current time stamps 3019 * for the packets that are in write queue : SYN packet and DATA 3020 */ 3021 skb_mstamp_get(&syn->skb_mstamp); 3022 data->skb_mstamp = syn->skb_mstamp; 3023 3024 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) { 3025 tp->syn_data = (fo->copied > 0); 3026 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3027 goto done; 3028 } 3029 syn_data = NULL; 3030 3031 fallback: 3032 /* Send a regular SYN with Fast Open cookie request option */ 3033 if (fo->cookie.len > 0) 3034 fo->cookie.len = 0; 3035 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3036 if (err) 3037 tp->syn_fastopen = 0; 3038 kfree_skb(syn_data); 3039 done: 3040 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3041 return err; 3042 } 3043 3044 /* Build a SYN and send it off. */ 3045 int tcp_connect(struct sock *sk) 3046 { 3047 struct tcp_sock *tp = tcp_sk(sk); 3048 struct sk_buff *buff; 3049 int err; 3050 3051 tcp_connect_init(sk); 3052 3053 if (unlikely(tp->repair)) { 3054 tcp_finish_connect(sk, NULL); 3055 return 0; 3056 } 3057 3058 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 3059 if (unlikely(buff == NULL)) 3060 return -ENOBUFS; 3061 3062 /* Reserve space for headers. */ 3063 skb_reserve(buff, MAX_TCP_HEADER); 3064 3065 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 3066 tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp; 3067 tcp_connect_queue_skb(sk, buff); 3068 TCP_ECN_send_syn(sk, buff); 3069 3070 /* Send off SYN; include data in Fast Open. */ 3071 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3072 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3073 if (err == -ECONNREFUSED) 3074 return err; 3075 3076 /* We change tp->snd_nxt after the tcp_transmit_skb() call 3077 * in order to make this packet get counted in tcpOutSegs. 3078 */ 3079 tp->snd_nxt = tp->write_seq; 3080 tp->pushed_seq = tp->write_seq; 3081 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 3082 3083 /* Timer for repeating the SYN until an answer. */ 3084 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 3085 inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 3086 return 0; 3087 } 3088 EXPORT_SYMBOL(tcp_connect); 3089 3090 /* Send out a delayed ack, the caller does the policy checking 3091 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 3092 * for details. 3093 */ 3094 void tcp_send_delayed_ack(struct sock *sk) 3095 { 3096 struct inet_connection_sock *icsk = inet_csk(sk); 3097 int ato = icsk->icsk_ack.ato; 3098 unsigned long timeout; 3099 3100 if (ato > TCP_DELACK_MIN) { 3101 const struct tcp_sock *tp = tcp_sk(sk); 3102 int max_ato = HZ / 2; 3103 3104 if (icsk->icsk_ack.pingpong || 3105 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 3106 max_ato = TCP_DELACK_MAX; 3107 3108 /* Slow path, intersegment interval is "high". */ 3109 3110 /* If some rtt estimate is known, use it to bound delayed ack. 3111 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 3112 * directly. 3113 */ 3114 if (tp->srtt_us) { 3115 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3116 TCP_DELACK_MIN); 3117 3118 if (rtt < max_ato) 3119 max_ato = rtt; 3120 } 3121 3122 ato = min(ato, max_ato); 3123 } 3124 3125 /* Stay within the limit we were given */ 3126 timeout = jiffies + ato; 3127 3128 /* Use new timeout only if there wasn't a older one earlier. */ 3129 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 3130 /* If delack timer was blocked or is about to expire, 3131 * send ACK now. 3132 */ 3133 if (icsk->icsk_ack.blocked || 3134 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 3135 tcp_send_ack(sk); 3136 return; 3137 } 3138 3139 if (!time_before(timeout, icsk->icsk_ack.timeout)) 3140 timeout = icsk->icsk_ack.timeout; 3141 } 3142 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3143 icsk->icsk_ack.timeout = timeout; 3144 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 3145 } 3146 3147 /* This routine sends an ack and also updates the window. */ 3148 void tcp_send_ack(struct sock *sk) 3149 { 3150 struct sk_buff *buff; 3151 3152 /* If we have been reset, we may not send again. */ 3153 if (sk->sk_state == TCP_CLOSE) 3154 return; 3155 3156 /* We are not putting this on the write queue, so 3157 * tcp_transmit_skb() will set the ownership to this 3158 * sock. 3159 */ 3160 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 3161 if (buff == NULL) { 3162 inet_csk_schedule_ack(sk); 3163 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 3164 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 3165 TCP_DELACK_MAX, TCP_RTO_MAX); 3166 return; 3167 } 3168 3169 /* Reserve space for headers and prepare control bits. */ 3170 skb_reserve(buff, MAX_TCP_HEADER); 3171 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 3172 3173 /* Send it off, this clears delayed acks for us. */ 3174 TCP_SKB_CB(buff)->when = tcp_time_stamp; 3175 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC)); 3176 } 3177 3178 /* This routine sends a packet with an out of date sequence 3179 * number. It assumes the other end will try to ack it. 3180 * 3181 * Question: what should we make while urgent mode? 3182 * 4.4BSD forces sending single byte of data. We cannot send 3183 * out of window data, because we have SND.NXT==SND.MAX... 3184 * 3185 * Current solution: to send TWO zero-length segments in urgent mode: 3186 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 3187 * out-of-date with SND.UNA-1 to probe window. 3188 */ 3189 static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 3190 { 3191 struct tcp_sock *tp = tcp_sk(sk); 3192 struct sk_buff *skb; 3193 3194 /* We don't queue it, tcp_transmit_skb() sets ownership. */ 3195 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 3196 if (skb == NULL) 3197 return -1; 3198 3199 /* Reserve space for headers and set control bits. */ 3200 skb_reserve(skb, MAX_TCP_HEADER); 3201 /* Use a previous sequence. This should cause the other 3202 * end to send an ack. Don't queue or clone SKB, just 3203 * send it. 3204 */ 3205 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 3206 TCP_SKB_CB(skb)->when = tcp_time_stamp; 3207 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 3208 } 3209 3210 void tcp_send_window_probe(struct sock *sk) 3211 { 3212 if (sk->sk_state == TCP_ESTABLISHED) { 3213 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3214 tcp_xmit_probe_skb(sk, 0); 3215 } 3216 } 3217 3218 /* Initiate keepalive or window probe from timer. */ 3219 int tcp_write_wakeup(struct sock *sk) 3220 { 3221 struct tcp_sock *tp = tcp_sk(sk); 3222 struct sk_buff *skb; 3223 3224 if (sk->sk_state == TCP_CLOSE) 3225 return -1; 3226 3227 if ((skb = tcp_send_head(sk)) != NULL && 3228 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 3229 int err; 3230 unsigned int mss = tcp_current_mss(sk); 3231 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 3232 3233 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 3234 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 3235 3236 /* We are probing the opening of a window 3237 * but the window size is != 0 3238 * must have been a result SWS avoidance ( sender ) 3239 */ 3240 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 3241 skb->len > mss) { 3242 seg_size = min(seg_size, mss); 3243 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3244 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC)) 3245 return -1; 3246 } else if (!tcp_skb_pcount(skb)) 3247 tcp_set_skb_tso_segs(sk, skb, mss); 3248 3249 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3250 TCP_SKB_CB(skb)->when = tcp_time_stamp; 3251 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 3252 if (!err) 3253 tcp_event_new_data_sent(sk, skb); 3254 return err; 3255 } else { 3256 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3257 tcp_xmit_probe_skb(sk, 1); 3258 return tcp_xmit_probe_skb(sk, 0); 3259 } 3260 } 3261 3262 /* A window probe timeout has occurred. If window is not closed send 3263 * a partial packet else a zero probe. 3264 */ 3265 void tcp_send_probe0(struct sock *sk) 3266 { 3267 struct inet_connection_sock *icsk = inet_csk(sk); 3268 struct tcp_sock *tp = tcp_sk(sk); 3269 int err; 3270 3271 err = tcp_write_wakeup(sk); 3272 3273 if (tp->packets_out || !tcp_send_head(sk)) { 3274 /* Cancel probe timer, if it is not required. */ 3275 icsk->icsk_probes_out = 0; 3276 icsk->icsk_backoff = 0; 3277 return; 3278 } 3279 3280 if (err <= 0) { 3281 if (icsk->icsk_backoff < sysctl_tcp_retries2) 3282 icsk->icsk_backoff++; 3283 icsk->icsk_probes_out++; 3284 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3285 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 3286 TCP_RTO_MAX); 3287 } else { 3288 /* If packet was not sent due to local congestion, 3289 * do not backoff and do not remember icsk_probes_out. 3290 * Let local senders to fight for local resources. 3291 * 3292 * Use accumulated backoff yet. 3293 */ 3294 if (!icsk->icsk_probes_out) 3295 icsk->icsk_probes_out = 1; 3296 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3297 min(icsk->icsk_rto << icsk->icsk_backoff, 3298 TCP_RESOURCE_PROBE_INTERVAL), 3299 TCP_RTO_MAX); 3300 } 3301 } 3302