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) { 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 const struct sk_buff *fclone = skb + 1; 882 883 skb_mstamp_get(&skb->skb_mstamp); 884 885 if (unlikely(skb->fclone == SKB_FCLONE_ORIG && 886 fclone->fclone == SKB_FCLONE_CLONE)) 887 NET_INC_STATS(sock_net(sk), 888 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 889 890 if (unlikely(skb_cloned(skb))) 891 skb = pskb_copy(skb, gfp_mask); 892 else 893 skb = skb_clone(skb, gfp_mask); 894 if (unlikely(!skb)) 895 return -ENOBUFS; 896 /* Our usage of tstamp should remain private */ 897 skb->tstamp.tv64 = 0; 898 } 899 900 inet = inet_sk(sk); 901 tp = tcp_sk(sk); 902 tcb = TCP_SKB_CB(skb); 903 memset(&opts, 0, sizeof(opts)); 904 905 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 906 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 907 else 908 tcp_options_size = tcp_established_options(sk, skb, &opts, 909 &md5); 910 tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 911 912 if (tcp_packets_in_flight(tp) == 0) 913 tcp_ca_event(sk, CA_EVENT_TX_START); 914 915 /* if no packet is in qdisc/device queue, then allow XPS to select 916 * another queue. 917 */ 918 skb->ooo_okay = sk_wmem_alloc_get(sk) == 0; 919 920 skb_push(skb, tcp_header_size); 921 skb_reset_transport_header(skb); 922 923 skb_orphan(skb); 924 skb->sk = sk; 925 skb->destructor = tcp_wfree; 926 atomic_add(skb->truesize, &sk->sk_wmem_alloc); 927 928 /* Build TCP header and checksum it. */ 929 th = tcp_hdr(skb); 930 th->source = inet->inet_sport; 931 th->dest = inet->inet_dport; 932 th->seq = htonl(tcb->seq); 933 th->ack_seq = htonl(tp->rcv_nxt); 934 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 935 tcb->tcp_flags); 936 937 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { 938 /* RFC1323: The window in SYN & SYN/ACK segments 939 * is never scaled. 940 */ 941 th->window = htons(min(tp->rcv_wnd, 65535U)); 942 } else { 943 th->window = htons(tcp_select_window(sk)); 944 } 945 th->check = 0; 946 th->urg_ptr = 0; 947 948 /* The urg_mode check is necessary during a below snd_una win probe */ 949 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 950 if (before(tp->snd_up, tcb->seq + 0x10000)) { 951 th->urg_ptr = htons(tp->snd_up - tcb->seq); 952 th->urg = 1; 953 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 954 th->urg_ptr = htons(0xFFFF); 955 th->urg = 1; 956 } 957 } 958 959 tcp_options_write((__be32 *)(th + 1), tp, &opts); 960 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) 961 TCP_ECN_send(sk, skb, tcp_header_size); 962 963 #ifdef CONFIG_TCP_MD5SIG 964 /* Calculate the MD5 hash, as we have all we need now */ 965 if (md5) { 966 sk_nocaps_add(sk, NETIF_F_GSO_MASK); 967 tp->af_specific->calc_md5_hash(opts.hash_location, 968 md5, sk, NULL, skb); 969 } 970 #endif 971 972 icsk->icsk_af_ops->send_check(sk, skb); 973 974 if (likely(tcb->tcp_flags & TCPHDR_ACK)) 975 tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 976 977 if (skb->len != tcp_header_size) 978 tcp_event_data_sent(tp, sk); 979 980 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 981 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 982 tcp_skb_pcount(skb)); 983 984 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 985 if (likely(err <= 0)) 986 return err; 987 988 tcp_enter_cwr(sk, 1); 989 990 return net_xmit_eval(err); 991 } 992 993 /* This routine just queues the buffer for sending. 994 * 995 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 996 * otherwise socket can stall. 997 */ 998 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 999 { 1000 struct tcp_sock *tp = tcp_sk(sk); 1001 1002 /* Advance write_seq and place onto the write_queue. */ 1003 tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1004 skb_header_release(skb); 1005 tcp_add_write_queue_tail(sk, skb); 1006 sk->sk_wmem_queued += skb->truesize; 1007 sk_mem_charge(sk, skb->truesize); 1008 } 1009 1010 /* Initialize TSO segments for a packet. */ 1011 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, 1012 unsigned int mss_now) 1013 { 1014 struct skb_shared_info *shinfo = skb_shinfo(skb); 1015 1016 /* Make sure we own this skb before messing gso_size/gso_segs */ 1017 WARN_ON_ONCE(skb_cloned(skb)); 1018 1019 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { 1020 /* Avoid the costly divide in the normal 1021 * non-TSO case. 1022 */ 1023 shinfo->gso_segs = 1; 1024 shinfo->gso_size = 0; 1025 shinfo->gso_type = 0; 1026 } else { 1027 shinfo->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 1028 shinfo->gso_size = mss_now; 1029 shinfo->gso_type = sk->sk_gso_type; 1030 } 1031 } 1032 1033 /* When a modification to fackets out becomes necessary, we need to check 1034 * skb is counted to fackets_out or not. 1035 */ 1036 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, 1037 int decr) 1038 { 1039 struct tcp_sock *tp = tcp_sk(sk); 1040 1041 if (!tp->sacked_out || tcp_is_reno(tp)) 1042 return; 1043 1044 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 1045 tp->fackets_out -= decr; 1046 } 1047 1048 /* Pcount in the middle of the write queue got changed, we need to do various 1049 * tweaks to fix counters 1050 */ 1051 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1052 { 1053 struct tcp_sock *tp = tcp_sk(sk); 1054 1055 tp->packets_out -= decr; 1056 1057 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1058 tp->sacked_out -= decr; 1059 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1060 tp->retrans_out -= decr; 1061 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1062 tp->lost_out -= decr; 1063 1064 /* Reno case is special. Sigh... */ 1065 if (tcp_is_reno(tp) && decr > 0) 1066 tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1067 1068 tcp_adjust_fackets_out(sk, skb, decr); 1069 1070 if (tp->lost_skb_hint && 1071 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1072 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 1073 tp->lost_cnt_hint -= decr; 1074 1075 tcp_verify_left_out(tp); 1076 } 1077 1078 /* Function to create two new TCP segments. Shrinks the given segment 1079 * to the specified size and appends a new segment with the rest of the 1080 * packet to the list. This won't be called frequently, I hope. 1081 * Remember, these are still headerless SKBs at this point. 1082 */ 1083 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 1084 unsigned int mss_now) 1085 { 1086 struct tcp_sock *tp = tcp_sk(sk); 1087 struct sk_buff *buff; 1088 int nsize, old_factor; 1089 int nlen; 1090 u8 flags; 1091 1092 if (WARN_ON(len > skb->len)) 1093 return -EINVAL; 1094 1095 nsize = skb_headlen(skb) - len; 1096 if (nsize < 0) 1097 nsize = 0; 1098 1099 if (skb_unclone(skb, GFP_ATOMIC)) 1100 return -ENOMEM; 1101 1102 /* Get a new skb... force flag on. */ 1103 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 1104 if (buff == NULL) 1105 return -ENOMEM; /* We'll just try again later. */ 1106 1107 sk->sk_wmem_queued += buff->truesize; 1108 sk_mem_charge(sk, buff->truesize); 1109 nlen = skb->len - len - nsize; 1110 buff->truesize += nlen; 1111 skb->truesize -= nlen; 1112 1113 /* Correct the sequence numbers. */ 1114 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1115 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1116 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1117 1118 /* PSH and FIN should only be set in the second packet. */ 1119 flags = TCP_SKB_CB(skb)->tcp_flags; 1120 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 1121 TCP_SKB_CB(buff)->tcp_flags = flags; 1122 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1123 1124 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 1125 /* Copy and checksum data tail into the new buffer. */ 1126 buff->csum = csum_partial_copy_nocheck(skb->data + len, 1127 skb_put(buff, nsize), 1128 nsize, 0); 1129 1130 skb_trim(skb, len); 1131 1132 skb->csum = csum_block_sub(skb->csum, buff->csum, len); 1133 } else { 1134 skb->ip_summed = CHECKSUM_PARTIAL; 1135 skb_split(skb, buff, len); 1136 } 1137 1138 buff->ip_summed = skb->ip_summed; 1139 1140 /* Looks stupid, but our code really uses when of 1141 * skbs, which it never sent before. --ANK 1142 */ 1143 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 1144 buff->tstamp = skb->tstamp; 1145 1146 old_factor = tcp_skb_pcount(skb); 1147 1148 /* Fix up tso_factor for both original and new SKB. */ 1149 tcp_set_skb_tso_segs(sk, skb, mss_now); 1150 tcp_set_skb_tso_segs(sk, buff, mss_now); 1151 1152 /* If this packet has been sent out already, we must 1153 * adjust the various packet counters. 1154 */ 1155 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 1156 int diff = old_factor - tcp_skb_pcount(skb) - 1157 tcp_skb_pcount(buff); 1158 1159 if (diff) 1160 tcp_adjust_pcount(sk, skb, diff); 1161 } 1162 1163 /* Link BUFF into the send queue. */ 1164 skb_header_release(buff); 1165 tcp_insert_write_queue_after(skb, buff, sk); 1166 1167 return 0; 1168 } 1169 1170 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 1171 * eventually). The difference is that pulled data not copied, but 1172 * immediately discarded. 1173 */ 1174 static void __pskb_trim_head(struct sk_buff *skb, int len) 1175 { 1176 struct skb_shared_info *shinfo; 1177 int i, k, eat; 1178 1179 eat = min_t(int, len, skb_headlen(skb)); 1180 if (eat) { 1181 __skb_pull(skb, eat); 1182 len -= eat; 1183 if (!len) 1184 return; 1185 } 1186 eat = len; 1187 k = 0; 1188 shinfo = skb_shinfo(skb); 1189 for (i = 0; i < shinfo->nr_frags; i++) { 1190 int size = skb_frag_size(&shinfo->frags[i]); 1191 1192 if (size <= eat) { 1193 skb_frag_unref(skb, i); 1194 eat -= size; 1195 } else { 1196 shinfo->frags[k] = shinfo->frags[i]; 1197 if (eat) { 1198 shinfo->frags[k].page_offset += eat; 1199 skb_frag_size_sub(&shinfo->frags[k], eat); 1200 eat = 0; 1201 } 1202 k++; 1203 } 1204 } 1205 shinfo->nr_frags = k; 1206 1207 skb_reset_tail_pointer(skb); 1208 skb->data_len -= len; 1209 skb->len = skb->data_len; 1210 } 1211 1212 /* Remove acked data from a packet in the transmit queue. */ 1213 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 1214 { 1215 if (skb_unclone(skb, GFP_ATOMIC)) 1216 return -ENOMEM; 1217 1218 __pskb_trim_head(skb, len); 1219 1220 TCP_SKB_CB(skb)->seq += len; 1221 skb->ip_summed = CHECKSUM_PARTIAL; 1222 1223 skb->truesize -= len; 1224 sk->sk_wmem_queued -= len; 1225 sk_mem_uncharge(sk, len); 1226 sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 1227 1228 /* Any change of skb->len requires recalculation of tso factor. */ 1229 if (tcp_skb_pcount(skb) > 1) 1230 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); 1231 1232 return 0; 1233 } 1234 1235 /* Calculate MSS not accounting any TCP options. */ 1236 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 1237 { 1238 const struct tcp_sock *tp = tcp_sk(sk); 1239 const struct inet_connection_sock *icsk = inet_csk(sk); 1240 int mss_now; 1241 1242 /* Calculate base mss without TCP options: 1243 It is MMS_S - sizeof(tcphdr) of rfc1122 1244 */ 1245 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 1246 1247 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 1248 if (icsk->icsk_af_ops->net_frag_header_len) { 1249 const struct dst_entry *dst = __sk_dst_get(sk); 1250 1251 if (dst && dst_allfrag(dst)) 1252 mss_now -= icsk->icsk_af_ops->net_frag_header_len; 1253 } 1254 1255 /* Clamp it (mss_clamp does not include tcp options) */ 1256 if (mss_now > tp->rx_opt.mss_clamp) 1257 mss_now = tp->rx_opt.mss_clamp; 1258 1259 /* Now subtract optional transport overhead */ 1260 mss_now -= icsk->icsk_ext_hdr_len; 1261 1262 /* Then reserve room for full set of TCP options and 8 bytes of data */ 1263 if (mss_now < 48) 1264 mss_now = 48; 1265 return mss_now; 1266 } 1267 1268 /* Calculate MSS. Not accounting for SACKs here. */ 1269 int tcp_mtu_to_mss(struct sock *sk, int pmtu) 1270 { 1271 /* Subtract TCP options size, not including SACKs */ 1272 return __tcp_mtu_to_mss(sk, pmtu) - 1273 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 1274 } 1275 1276 /* Inverse of above */ 1277 int tcp_mss_to_mtu(struct sock *sk, int mss) 1278 { 1279 const struct tcp_sock *tp = tcp_sk(sk); 1280 const struct inet_connection_sock *icsk = inet_csk(sk); 1281 int mtu; 1282 1283 mtu = mss + 1284 tp->tcp_header_len + 1285 icsk->icsk_ext_hdr_len + 1286 icsk->icsk_af_ops->net_header_len; 1287 1288 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 1289 if (icsk->icsk_af_ops->net_frag_header_len) { 1290 const struct dst_entry *dst = __sk_dst_get(sk); 1291 1292 if (dst && dst_allfrag(dst)) 1293 mtu += icsk->icsk_af_ops->net_frag_header_len; 1294 } 1295 return mtu; 1296 } 1297 1298 /* MTU probing init per socket */ 1299 void tcp_mtup_init(struct sock *sk) 1300 { 1301 struct tcp_sock *tp = tcp_sk(sk); 1302 struct inet_connection_sock *icsk = inet_csk(sk); 1303 1304 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 1305 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 1306 icsk->icsk_af_ops->net_header_len; 1307 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 1308 icsk->icsk_mtup.probe_size = 0; 1309 } 1310 EXPORT_SYMBOL(tcp_mtup_init); 1311 1312 /* This function synchronize snd mss to current pmtu/exthdr set. 1313 1314 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 1315 for TCP options, but includes only bare TCP header. 1316 1317 tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1318 It is minimum of user_mss and mss received with SYN. 1319 It also does not include TCP options. 1320 1321 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 1322 1323 tp->mss_cache is current effective sending mss, including 1324 all tcp options except for SACKs. It is evaluated, 1325 taking into account current pmtu, but never exceeds 1326 tp->rx_opt.mss_clamp. 1327 1328 NOTE1. rfc1122 clearly states that advertised MSS 1329 DOES NOT include either tcp or ip options. 1330 1331 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1332 are READ ONLY outside this function. --ANK (980731) 1333 */ 1334 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 1335 { 1336 struct tcp_sock *tp = tcp_sk(sk); 1337 struct inet_connection_sock *icsk = inet_csk(sk); 1338 int mss_now; 1339 1340 if (icsk->icsk_mtup.search_high > pmtu) 1341 icsk->icsk_mtup.search_high = pmtu; 1342 1343 mss_now = tcp_mtu_to_mss(sk, pmtu); 1344 mss_now = tcp_bound_to_half_wnd(tp, mss_now); 1345 1346 /* And store cached results */ 1347 icsk->icsk_pmtu_cookie = pmtu; 1348 if (icsk->icsk_mtup.enabled) 1349 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1350 tp->mss_cache = mss_now; 1351 1352 return mss_now; 1353 } 1354 EXPORT_SYMBOL(tcp_sync_mss); 1355 1356 /* Compute the current effective MSS, taking SACKs and IP options, 1357 * and even PMTU discovery events into account. 1358 */ 1359 unsigned int tcp_current_mss(struct sock *sk) 1360 { 1361 const struct tcp_sock *tp = tcp_sk(sk); 1362 const struct dst_entry *dst = __sk_dst_get(sk); 1363 u32 mss_now; 1364 unsigned int header_len; 1365 struct tcp_out_options opts; 1366 struct tcp_md5sig_key *md5; 1367 1368 mss_now = tp->mss_cache; 1369 1370 if (dst) { 1371 u32 mtu = dst_mtu(dst); 1372 if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 1373 mss_now = tcp_sync_mss(sk, mtu); 1374 } 1375 1376 header_len = tcp_established_options(sk, NULL, &opts, &md5) + 1377 sizeof(struct tcphdr); 1378 /* The mss_cache is sized based on tp->tcp_header_len, which assumes 1379 * some common options. If this is an odd packet (because we have SACK 1380 * blocks etc) then our calculated header_len will be different, and 1381 * we have to adjust mss_now correspondingly */ 1382 if (header_len != tp->tcp_header_len) { 1383 int delta = (int) header_len - tp->tcp_header_len; 1384 mss_now -= delta; 1385 } 1386 1387 return mss_now; 1388 } 1389 1390 /* Congestion window validation. (RFC2861) */ 1391 static void tcp_cwnd_validate(struct sock *sk) 1392 { 1393 struct tcp_sock *tp = tcp_sk(sk); 1394 1395 if (tp->packets_out >= tp->snd_cwnd) { 1396 /* Network is feed fully. */ 1397 tp->snd_cwnd_used = 0; 1398 tp->snd_cwnd_stamp = tcp_time_stamp; 1399 } else { 1400 /* Network starves. */ 1401 if (tp->packets_out > tp->snd_cwnd_used) 1402 tp->snd_cwnd_used = tp->packets_out; 1403 1404 if (sysctl_tcp_slow_start_after_idle && 1405 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1406 tcp_cwnd_application_limited(sk); 1407 } 1408 } 1409 1410 /* Minshall's variant of the Nagle send check. */ 1411 static bool tcp_minshall_check(const struct tcp_sock *tp) 1412 { 1413 return after(tp->snd_sml, tp->snd_una) && 1414 !after(tp->snd_sml, tp->snd_nxt); 1415 } 1416 1417 /* Update snd_sml if this skb is under mss 1418 * Note that a TSO packet might end with a sub-mss segment 1419 * The test is really : 1420 * if ((skb->len % mss) != 0) 1421 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1422 * But we can avoid doing the divide again given we already have 1423 * skb_pcount = skb->len / mss_now 1424 */ 1425 static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1426 const struct sk_buff *skb) 1427 { 1428 if (skb->len < tcp_skb_pcount(skb) * mss_now) 1429 tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1430 } 1431 1432 /* Return false, if packet can be sent now without violation Nagle's rules: 1433 * 1. It is full sized. (provided by caller in %partial bool) 1434 * 2. Or it contains FIN. (already checked by caller) 1435 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1436 * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1437 * With Minshall's modification: all sent small packets are ACKed. 1438 */ 1439 static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1440 int nonagle) 1441 { 1442 return partial && 1443 ((nonagle & TCP_NAGLE_CORK) || 1444 (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1445 } 1446 /* Returns the portion of skb which can be sent right away */ 1447 static unsigned int tcp_mss_split_point(const struct sock *sk, 1448 const struct sk_buff *skb, 1449 unsigned int mss_now, 1450 unsigned int max_segs, 1451 int nonagle) 1452 { 1453 const struct tcp_sock *tp = tcp_sk(sk); 1454 u32 partial, needed, window, max_len; 1455 1456 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1457 max_len = mss_now * max_segs; 1458 1459 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 1460 return max_len; 1461 1462 needed = min(skb->len, window); 1463 1464 if (max_len <= needed) 1465 return max_len; 1466 1467 partial = needed % mss_now; 1468 /* If last segment is not a full MSS, check if Nagle rules allow us 1469 * to include this last segment in this skb. 1470 * Otherwise, we'll split the skb at last MSS boundary 1471 */ 1472 if (tcp_nagle_check(partial != 0, tp, nonagle)) 1473 return needed - partial; 1474 1475 return needed; 1476 } 1477 1478 /* Can at least one segment of SKB be sent right now, according to the 1479 * congestion window rules? If so, return how many segments are allowed. 1480 */ 1481 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1482 const struct sk_buff *skb) 1483 { 1484 u32 in_flight, cwnd; 1485 1486 /* Don't be strict about the congestion window for the final FIN. */ 1487 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 1488 tcp_skb_pcount(skb) == 1) 1489 return 1; 1490 1491 in_flight = tcp_packets_in_flight(tp); 1492 cwnd = tp->snd_cwnd; 1493 if (in_flight < cwnd) 1494 return (cwnd - in_flight); 1495 1496 return 0; 1497 } 1498 1499 /* Initialize TSO state of a skb. 1500 * This must be invoked the first time we consider transmitting 1501 * SKB onto the wire. 1502 */ 1503 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, 1504 unsigned int mss_now) 1505 { 1506 int tso_segs = tcp_skb_pcount(skb); 1507 1508 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1509 tcp_set_skb_tso_segs(sk, skb, mss_now); 1510 tso_segs = tcp_skb_pcount(skb); 1511 } 1512 return tso_segs; 1513 } 1514 1515 1516 /* Return true if the Nagle test allows this packet to be 1517 * sent now. 1518 */ 1519 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1520 unsigned int cur_mss, int nonagle) 1521 { 1522 /* Nagle rule does not apply to frames, which sit in the middle of the 1523 * write_queue (they have no chances to get new data). 1524 * 1525 * This is implemented in the callers, where they modify the 'nonagle' 1526 * argument based upon the location of SKB in the send queue. 1527 */ 1528 if (nonagle & TCP_NAGLE_PUSH) 1529 return true; 1530 1531 /* Don't use the nagle rule for urgent data (or for the final FIN). */ 1532 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1533 return true; 1534 1535 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1536 return true; 1537 1538 return false; 1539 } 1540 1541 /* Does at least the first segment of SKB fit into the send window? */ 1542 static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1543 const struct sk_buff *skb, 1544 unsigned int cur_mss) 1545 { 1546 u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1547 1548 if (skb->len > cur_mss) 1549 end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1550 1551 return !after(end_seq, tcp_wnd_end(tp)); 1552 } 1553 1554 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1555 * should be put on the wire right now. If so, it returns the number of 1556 * packets allowed by the congestion window. 1557 */ 1558 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, 1559 unsigned int cur_mss, int nonagle) 1560 { 1561 const struct tcp_sock *tp = tcp_sk(sk); 1562 unsigned int cwnd_quota; 1563 1564 tcp_init_tso_segs(sk, skb, cur_mss); 1565 1566 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1567 return 0; 1568 1569 cwnd_quota = tcp_cwnd_test(tp, skb); 1570 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1571 cwnd_quota = 0; 1572 1573 return cwnd_quota; 1574 } 1575 1576 /* Test if sending is allowed right now. */ 1577 bool tcp_may_send_now(struct sock *sk) 1578 { 1579 const struct tcp_sock *tp = tcp_sk(sk); 1580 struct sk_buff *skb = tcp_send_head(sk); 1581 1582 return skb && 1583 tcp_snd_test(sk, skb, tcp_current_mss(sk), 1584 (tcp_skb_is_last(sk, skb) ? 1585 tp->nonagle : TCP_NAGLE_PUSH)); 1586 } 1587 1588 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1589 * which is put after SKB on the list. It is very much like 1590 * tcp_fragment() except that it may make several kinds of assumptions 1591 * in order to speed up the splitting operation. In particular, we 1592 * know that all the data is in scatter-gather pages, and that the 1593 * packet has never been sent out before (and thus is not cloned). 1594 */ 1595 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1596 unsigned int mss_now, gfp_t gfp) 1597 { 1598 struct sk_buff *buff; 1599 int nlen = skb->len - len; 1600 u8 flags; 1601 1602 /* All of a TSO frame must be composed of paged data. */ 1603 if (skb->len != skb->data_len) 1604 return tcp_fragment(sk, skb, len, mss_now); 1605 1606 buff = sk_stream_alloc_skb(sk, 0, gfp); 1607 if (unlikely(buff == NULL)) 1608 return -ENOMEM; 1609 1610 sk->sk_wmem_queued += buff->truesize; 1611 sk_mem_charge(sk, buff->truesize); 1612 buff->truesize += nlen; 1613 skb->truesize -= nlen; 1614 1615 /* Correct the sequence numbers. */ 1616 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1617 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1618 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1619 1620 /* PSH and FIN should only be set in the second packet. */ 1621 flags = TCP_SKB_CB(skb)->tcp_flags; 1622 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 1623 TCP_SKB_CB(buff)->tcp_flags = flags; 1624 1625 /* This packet was never sent out yet, so no SACK bits. */ 1626 TCP_SKB_CB(buff)->sacked = 0; 1627 1628 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1629 skb_split(skb, buff, len); 1630 1631 /* Fix up tso_factor for both original and new SKB. */ 1632 tcp_set_skb_tso_segs(sk, skb, mss_now); 1633 tcp_set_skb_tso_segs(sk, buff, mss_now); 1634 1635 /* Link BUFF into the send queue. */ 1636 skb_header_release(buff); 1637 tcp_insert_write_queue_after(skb, buff, sk); 1638 1639 return 0; 1640 } 1641 1642 /* Try to defer sending, if possible, in order to minimize the amount 1643 * of TSO splitting we do. View it as a kind of TSO Nagle test. 1644 * 1645 * This algorithm is from John Heffner. 1646 */ 1647 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1648 { 1649 struct tcp_sock *tp = tcp_sk(sk); 1650 const struct inet_connection_sock *icsk = inet_csk(sk); 1651 u32 send_win, cong_win, limit, in_flight; 1652 int win_divisor; 1653 1654 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) 1655 goto send_now; 1656 1657 if (icsk->icsk_ca_state != TCP_CA_Open) 1658 goto send_now; 1659 1660 /* Defer for less than two clock ticks. */ 1661 if (tp->tso_deferred && 1662 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1663 goto send_now; 1664 1665 in_flight = tcp_packets_in_flight(tp); 1666 1667 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1668 1669 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1670 1671 /* From in_flight test above, we know that cwnd > in_flight. */ 1672 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1673 1674 limit = min(send_win, cong_win); 1675 1676 /* If a full-sized TSO skb can be sent, do it. */ 1677 if (limit >= min_t(unsigned int, sk->sk_gso_max_size, 1678 tp->xmit_size_goal_segs * tp->mss_cache)) 1679 goto send_now; 1680 1681 /* Middle in queue won't get any more data, full sendable already? */ 1682 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 1683 goto send_now; 1684 1685 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); 1686 if (win_divisor) { 1687 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1688 1689 /* If at least some fraction of a window is available, 1690 * just use it. 1691 */ 1692 chunk /= win_divisor; 1693 if (limit >= chunk) 1694 goto send_now; 1695 } else { 1696 /* Different approach, try not to defer past a single 1697 * ACK. Receiver should ACK every other full sized 1698 * frame, so if we have space for more than 3 frames 1699 * then send now. 1700 */ 1701 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1702 goto send_now; 1703 } 1704 1705 /* Ok, it looks like it is advisable to defer. 1706 * Do not rearm the timer if already set to not break TCP ACK clocking. 1707 */ 1708 if (!tp->tso_deferred) 1709 tp->tso_deferred = 1 | (jiffies << 1); 1710 1711 return true; 1712 1713 send_now: 1714 tp->tso_deferred = 0; 1715 return false; 1716 } 1717 1718 /* Create a new MTU probe if we are ready. 1719 * MTU probe is regularly attempting to increase the path MTU by 1720 * deliberately sending larger packets. This discovers routing 1721 * changes resulting in larger path MTUs. 1722 * 1723 * Returns 0 if we should wait to probe (no cwnd available), 1724 * 1 if a probe was sent, 1725 * -1 otherwise 1726 */ 1727 static int tcp_mtu_probe(struct sock *sk) 1728 { 1729 struct tcp_sock *tp = tcp_sk(sk); 1730 struct inet_connection_sock *icsk = inet_csk(sk); 1731 struct sk_buff *skb, *nskb, *next; 1732 int len; 1733 int probe_size; 1734 int size_needed; 1735 int copy; 1736 int mss_now; 1737 1738 /* Not currently probing/verifying, 1739 * not in recovery, 1740 * have enough cwnd, and 1741 * not SACKing (the variable headers throw things off) */ 1742 if (!icsk->icsk_mtup.enabled || 1743 icsk->icsk_mtup.probe_size || 1744 inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 1745 tp->snd_cwnd < 11 || 1746 tp->rx_opt.num_sacks || tp->rx_opt.dsack) 1747 return -1; 1748 1749 /* Very simple search strategy: just double the MSS. */ 1750 mss_now = tcp_current_mss(sk); 1751 probe_size = 2 * tp->mss_cache; 1752 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 1753 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 1754 /* TODO: set timer for probe_converge_event */ 1755 return -1; 1756 } 1757 1758 /* Have enough data in the send queue to probe? */ 1759 if (tp->write_seq - tp->snd_nxt < size_needed) 1760 return -1; 1761 1762 if (tp->snd_wnd < size_needed) 1763 return -1; 1764 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 1765 return 0; 1766 1767 /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1768 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1769 if (!tcp_packets_in_flight(tp)) 1770 return -1; 1771 else 1772 return 0; 1773 } 1774 1775 /* We're allowed to probe. Build it now. */ 1776 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 1777 return -1; 1778 sk->sk_wmem_queued += nskb->truesize; 1779 sk_mem_charge(sk, nskb->truesize); 1780 1781 skb = tcp_send_head(sk); 1782 1783 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 1784 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 1785 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 1786 TCP_SKB_CB(nskb)->sacked = 0; 1787 nskb->csum = 0; 1788 nskb->ip_summed = skb->ip_summed; 1789 1790 tcp_insert_write_queue_before(nskb, skb, sk); 1791 1792 len = 0; 1793 tcp_for_write_queue_from_safe(skb, next, sk) { 1794 copy = min_t(int, skb->len, probe_size - len); 1795 if (nskb->ip_summed) 1796 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 1797 else 1798 nskb->csum = skb_copy_and_csum_bits(skb, 0, 1799 skb_put(nskb, copy), 1800 copy, nskb->csum); 1801 1802 if (skb->len <= copy) { 1803 /* We've eaten all the data from this skb. 1804 * Throw it away. */ 1805 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 1806 tcp_unlink_write_queue(skb, sk); 1807 sk_wmem_free_skb(sk, skb); 1808 } else { 1809 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 1810 ~(TCPHDR_FIN|TCPHDR_PSH); 1811 if (!skb_shinfo(skb)->nr_frags) { 1812 skb_pull(skb, copy); 1813 if (skb->ip_summed != CHECKSUM_PARTIAL) 1814 skb->csum = csum_partial(skb->data, 1815 skb->len, 0); 1816 } else { 1817 __pskb_trim_head(skb, copy); 1818 tcp_set_skb_tso_segs(sk, skb, mss_now); 1819 } 1820 TCP_SKB_CB(skb)->seq += copy; 1821 } 1822 1823 len += copy; 1824 1825 if (len >= probe_size) 1826 break; 1827 } 1828 tcp_init_tso_segs(sk, nskb, nskb->len); 1829 1830 /* We're ready to send. If this fails, the probe will 1831 * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 1832 TCP_SKB_CB(nskb)->when = tcp_time_stamp; 1833 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 1834 /* Decrement cwnd here because we are sending 1835 * effectively two packets. */ 1836 tp->snd_cwnd--; 1837 tcp_event_new_data_sent(sk, nskb); 1838 1839 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 1840 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 1841 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 1842 1843 return 1; 1844 } 1845 1846 return -1; 1847 } 1848 1849 /* This routine writes packets to the network. It advances the 1850 * send_head. This happens as incoming acks open up the remote 1851 * window for us. 1852 * 1853 * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1854 * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1855 * account rare use of URG, this is not a big flaw. 1856 * 1857 * Send at most one packet when push_one > 0. Temporarily ignore 1858 * cwnd limit to force at most one packet out when push_one == 2. 1859 1860 * Returns true, if no segments are in flight and we have queued segments, 1861 * but cannot send anything now because of SWS or another problem. 1862 */ 1863 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1864 int push_one, gfp_t gfp) 1865 { 1866 struct tcp_sock *tp = tcp_sk(sk); 1867 struct sk_buff *skb; 1868 unsigned int tso_segs, sent_pkts; 1869 int cwnd_quota; 1870 int result; 1871 1872 sent_pkts = 0; 1873 1874 if (!push_one) { 1875 /* Do MTU probing. */ 1876 result = tcp_mtu_probe(sk); 1877 if (!result) { 1878 return false; 1879 } else if (result > 0) { 1880 sent_pkts = 1; 1881 } 1882 } 1883 1884 while ((skb = tcp_send_head(sk))) { 1885 unsigned int limit; 1886 1887 tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1888 BUG_ON(!tso_segs); 1889 1890 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) 1891 goto repair; /* Skip network transmission */ 1892 1893 cwnd_quota = tcp_cwnd_test(tp, skb); 1894 if (!cwnd_quota) { 1895 if (push_one == 2) 1896 /* Force out a loss probe pkt. */ 1897 cwnd_quota = 1; 1898 else 1899 break; 1900 } 1901 1902 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1903 break; 1904 1905 if (tso_segs == 1) { 1906 if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1907 (tcp_skb_is_last(sk, skb) ? 1908 nonagle : TCP_NAGLE_PUSH)))) 1909 break; 1910 } else { 1911 if (!push_one && tcp_tso_should_defer(sk, skb)) 1912 break; 1913 } 1914 1915 /* TCP Small Queues : 1916 * Control number of packets in qdisc/devices to two packets / or ~1 ms. 1917 * This allows for : 1918 * - better RTT estimation and ACK scheduling 1919 * - faster recovery 1920 * - high rates 1921 * Alas, some drivers / subsystems require a fair amount 1922 * of queued bytes to ensure line rate. 1923 * One example is wifi aggregation (802.11 AMPDU) 1924 */ 1925 limit = max_t(unsigned int, sysctl_tcp_limit_output_bytes, 1926 sk->sk_pacing_rate >> 10); 1927 1928 if (atomic_read(&sk->sk_wmem_alloc) > limit) { 1929 set_bit(TSQ_THROTTLED, &tp->tsq_flags); 1930 /* It is possible TX completion already happened 1931 * before we set TSQ_THROTTLED, so we must 1932 * test again the condition. 1933 * We abuse smp_mb__after_clear_bit() because 1934 * there is no smp_mb__after_set_bit() yet 1935 */ 1936 smp_mb__after_clear_bit(); 1937 if (atomic_read(&sk->sk_wmem_alloc) > limit) 1938 break; 1939 } 1940 1941 limit = mss_now; 1942 if (tso_segs > 1 && !tcp_urg_mode(tp)) 1943 limit = tcp_mss_split_point(sk, skb, mss_now, 1944 min_t(unsigned int, 1945 cwnd_quota, 1946 sk->sk_gso_max_segs), 1947 nonagle); 1948 1949 if (skb->len > limit && 1950 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 1951 break; 1952 1953 TCP_SKB_CB(skb)->when = tcp_time_stamp; 1954 1955 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 1956 break; 1957 1958 repair: 1959 /* Advance the send_head. This one is sent out. 1960 * This call will increment packets_out. 1961 */ 1962 tcp_event_new_data_sent(sk, skb); 1963 1964 tcp_minshall_update(tp, mss_now, skb); 1965 sent_pkts += tcp_skb_pcount(skb); 1966 1967 if (push_one) 1968 break; 1969 } 1970 1971 if (likely(sent_pkts)) { 1972 if (tcp_in_cwnd_reduction(sk)) 1973 tp->prr_out += sent_pkts; 1974 1975 /* Send one loss probe per tail loss episode. */ 1976 if (push_one != 2) 1977 tcp_schedule_loss_probe(sk); 1978 tcp_cwnd_validate(sk); 1979 return false; 1980 } 1981 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk)); 1982 } 1983 1984 bool tcp_schedule_loss_probe(struct sock *sk) 1985 { 1986 struct inet_connection_sock *icsk = inet_csk(sk); 1987 struct tcp_sock *tp = tcp_sk(sk); 1988 u32 timeout, tlp_time_stamp, rto_time_stamp; 1989 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3); 1990 1991 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) 1992 return false; 1993 /* No consecutive loss probes. */ 1994 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { 1995 tcp_rearm_rto(sk); 1996 return false; 1997 } 1998 /* Don't do any loss probe on a Fast Open connection before 3WHS 1999 * finishes. 2000 */ 2001 if (sk->sk_state == TCP_SYN_RECV) 2002 return false; 2003 2004 /* TLP is only scheduled when next timer event is RTO. */ 2005 if (icsk->icsk_pending != ICSK_TIME_RETRANS) 2006 return false; 2007 2008 /* Schedule a loss probe in 2*RTT for SACK capable connections 2009 * in Open state, that are either limited by cwnd or application. 2010 */ 2011 if (sysctl_tcp_early_retrans < 3 || !tp->srtt_us || !tp->packets_out || 2012 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) 2013 return false; 2014 2015 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && 2016 tcp_send_head(sk)) 2017 return false; 2018 2019 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account 2020 * for delayed ack when there's one outstanding packet. 2021 */ 2022 timeout = rtt << 1; 2023 if (tp->packets_out == 1) 2024 timeout = max_t(u32, timeout, 2025 (rtt + (rtt >> 1) + TCP_DELACK_MAX)); 2026 timeout = max_t(u32, timeout, msecs_to_jiffies(10)); 2027 2028 /* If RTO is shorter, just schedule TLP in its place. */ 2029 tlp_time_stamp = tcp_time_stamp + timeout; 2030 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; 2031 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { 2032 s32 delta = rto_time_stamp - tcp_time_stamp; 2033 if (delta > 0) 2034 timeout = delta; 2035 } 2036 2037 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 2038 TCP_RTO_MAX); 2039 return true; 2040 } 2041 2042 /* When probe timeout (PTO) fires, send a new segment if one exists, else 2043 * retransmit the last segment. 2044 */ 2045 void tcp_send_loss_probe(struct sock *sk) 2046 { 2047 struct tcp_sock *tp = tcp_sk(sk); 2048 struct sk_buff *skb; 2049 int pcount; 2050 int mss = tcp_current_mss(sk); 2051 int err = -1; 2052 2053 if (tcp_send_head(sk) != NULL) { 2054 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2055 goto rearm_timer; 2056 } 2057 2058 /* At most one outstanding TLP retransmission. */ 2059 if (tp->tlp_high_seq) 2060 goto rearm_timer; 2061 2062 /* Retransmit last segment. */ 2063 skb = tcp_write_queue_tail(sk); 2064 if (WARN_ON(!skb)) 2065 goto rearm_timer; 2066 2067 pcount = tcp_skb_pcount(skb); 2068 if (WARN_ON(!pcount)) 2069 goto rearm_timer; 2070 2071 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 2072 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss))) 2073 goto rearm_timer; 2074 skb = tcp_write_queue_tail(sk); 2075 } 2076 2077 if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 2078 goto rearm_timer; 2079 2080 /* Probe with zero data doesn't trigger fast recovery. */ 2081 if (skb->len > 0) 2082 err = __tcp_retransmit_skb(sk, skb); 2083 2084 /* Record snd_nxt for loss detection. */ 2085 if (likely(!err)) 2086 tp->tlp_high_seq = tp->snd_nxt; 2087 2088 rearm_timer: 2089 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2090 inet_csk(sk)->icsk_rto, 2091 TCP_RTO_MAX); 2092 2093 if (likely(!err)) 2094 NET_INC_STATS_BH(sock_net(sk), 2095 LINUX_MIB_TCPLOSSPROBES); 2096 } 2097 2098 /* Push out any pending frames which were held back due to 2099 * TCP_CORK or attempt at coalescing tiny packets. 2100 * The socket must be locked by the caller. 2101 */ 2102 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 2103 int nonagle) 2104 { 2105 /* If we are closed, the bytes will have to remain here. 2106 * In time closedown will finish, we empty the write queue and 2107 * all will be happy. 2108 */ 2109 if (unlikely(sk->sk_state == TCP_CLOSE)) 2110 return; 2111 2112 if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 2113 sk_gfp_atomic(sk, GFP_ATOMIC))) 2114 tcp_check_probe_timer(sk); 2115 } 2116 2117 /* Send _single_ skb sitting at the send head. This function requires 2118 * true push pending frames to setup probe timer etc. 2119 */ 2120 void tcp_push_one(struct sock *sk, unsigned int mss_now) 2121 { 2122 struct sk_buff *skb = tcp_send_head(sk); 2123 2124 BUG_ON(!skb || skb->len < mss_now); 2125 2126 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2127 } 2128 2129 /* This function returns the amount that we can raise the 2130 * usable window based on the following constraints 2131 * 2132 * 1. The window can never be shrunk once it is offered (RFC 793) 2133 * 2. We limit memory per socket 2134 * 2135 * RFC 1122: 2136 * "the suggested [SWS] avoidance algorithm for the receiver is to keep 2137 * RECV.NEXT + RCV.WIN fixed until: 2138 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 2139 * 2140 * i.e. don't raise the right edge of the window until you can raise 2141 * it at least MSS bytes. 2142 * 2143 * Unfortunately, the recommended algorithm breaks header prediction, 2144 * since header prediction assumes th->window stays fixed. 2145 * 2146 * Strictly speaking, keeping th->window fixed violates the receiver 2147 * side SWS prevention criteria. The problem is that under this rule 2148 * a stream of single byte packets will cause the right side of the 2149 * window to always advance by a single byte. 2150 * 2151 * Of course, if the sender implements sender side SWS prevention 2152 * then this will not be a problem. 2153 * 2154 * BSD seems to make the following compromise: 2155 * 2156 * If the free space is less than the 1/4 of the maximum 2157 * space available and the free space is less than 1/2 mss, 2158 * then set the window to 0. 2159 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 2160 * Otherwise, just prevent the window from shrinking 2161 * and from being larger than the largest representable value. 2162 * 2163 * This prevents incremental opening of the window in the regime 2164 * where TCP is limited by the speed of the reader side taking 2165 * data out of the TCP receive queue. It does nothing about 2166 * those cases where the window is constrained on the sender side 2167 * because the pipeline is full. 2168 * 2169 * BSD also seems to "accidentally" limit itself to windows that are a 2170 * multiple of MSS, at least until the free space gets quite small. 2171 * This would appear to be a side effect of the mbuf implementation. 2172 * Combining these two algorithms results in the observed behavior 2173 * of having a fixed window size at almost all times. 2174 * 2175 * Below we obtain similar behavior by forcing the offered window to 2176 * a multiple of the mss when it is feasible to do so. 2177 * 2178 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 2179 * Regular options like TIMESTAMP are taken into account. 2180 */ 2181 u32 __tcp_select_window(struct sock *sk) 2182 { 2183 struct inet_connection_sock *icsk = inet_csk(sk); 2184 struct tcp_sock *tp = tcp_sk(sk); 2185 /* MSS for the peer's data. Previous versions used mss_clamp 2186 * here. I don't know if the value based on our guesses 2187 * of peer's MSS is better for the performance. It's more correct 2188 * but may be worse for the performance because of rcv_mss 2189 * fluctuations. --SAW 1998/11/1 2190 */ 2191 int mss = icsk->icsk_ack.rcv_mss; 2192 int free_space = tcp_space(sk); 2193 int allowed_space = tcp_full_space(sk); 2194 int full_space = min_t(int, tp->window_clamp, allowed_space); 2195 int window; 2196 2197 if (mss > full_space) 2198 mss = full_space; 2199 2200 if (free_space < (full_space >> 1)) { 2201 icsk->icsk_ack.quick = 0; 2202 2203 if (sk_under_memory_pressure(sk)) 2204 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2205 4U * tp->advmss); 2206 2207 /* free_space might become our new window, make sure we don't 2208 * increase it due to wscale. 2209 */ 2210 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 2211 2212 /* if free space is less than mss estimate, or is below 1/16th 2213 * of the maximum allowed, try to move to zero-window, else 2214 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 2215 * new incoming data is dropped due to memory limits. 2216 * With large window, mss test triggers way too late in order 2217 * to announce zero window in time before rmem limit kicks in. 2218 */ 2219 if (free_space < (allowed_space >> 4) || free_space < mss) 2220 return 0; 2221 } 2222 2223 if (free_space > tp->rcv_ssthresh) 2224 free_space = tp->rcv_ssthresh; 2225 2226 /* Don't do rounding if we are using window scaling, since the 2227 * scaled window will not line up with the MSS boundary anyway. 2228 */ 2229 window = tp->rcv_wnd; 2230 if (tp->rx_opt.rcv_wscale) { 2231 window = free_space; 2232 2233 /* Advertise enough space so that it won't get scaled away. 2234 * Import case: prevent zero window announcement if 2235 * 1<<rcv_wscale > mss. 2236 */ 2237 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 2238 window = (((window >> tp->rx_opt.rcv_wscale) + 1) 2239 << tp->rx_opt.rcv_wscale); 2240 } else { 2241 /* Get the largest window that is a nice multiple of mss. 2242 * Window clamp already applied above. 2243 * If our current window offering is within 1 mss of the 2244 * free space we just keep it. This prevents the divide 2245 * and multiply from happening most of the time. 2246 * We also don't do any window rounding when the free space 2247 * is too small. 2248 */ 2249 if (window <= free_space - mss || window > free_space) 2250 window = (free_space / mss) * mss; 2251 else if (mss == full_space && 2252 free_space > window + (full_space >> 1)) 2253 window = free_space; 2254 } 2255 2256 return window; 2257 } 2258 2259 /* Collapses two adjacent SKB's during retransmission. */ 2260 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 2261 { 2262 struct tcp_sock *tp = tcp_sk(sk); 2263 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 2264 int skb_size, next_skb_size; 2265 2266 skb_size = skb->len; 2267 next_skb_size = next_skb->len; 2268 2269 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 2270 2271 tcp_highest_sack_combine(sk, next_skb, skb); 2272 2273 tcp_unlink_write_queue(next_skb, sk); 2274 2275 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 2276 next_skb_size); 2277 2278 if (next_skb->ip_summed == CHECKSUM_PARTIAL) 2279 skb->ip_summed = CHECKSUM_PARTIAL; 2280 2281 if (skb->ip_summed != CHECKSUM_PARTIAL) 2282 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 2283 2284 /* Update sequence range on original skb. */ 2285 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 2286 2287 /* Merge over control information. This moves PSH/FIN etc. over */ 2288 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 2289 2290 /* All done, get rid of second SKB and account for it so 2291 * packet counting does not break. 2292 */ 2293 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2294 2295 /* changed transmit queue under us so clear hints */ 2296 tcp_clear_retrans_hints_partial(tp); 2297 if (next_skb == tp->retransmit_skb_hint) 2298 tp->retransmit_skb_hint = skb; 2299 2300 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2301 2302 sk_wmem_free_skb(sk, next_skb); 2303 } 2304 2305 /* Check if coalescing SKBs is legal. */ 2306 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 2307 { 2308 if (tcp_skb_pcount(skb) > 1) 2309 return false; 2310 /* TODO: SACK collapsing could be used to remove this condition */ 2311 if (skb_shinfo(skb)->nr_frags != 0) 2312 return false; 2313 if (skb_cloned(skb)) 2314 return false; 2315 if (skb == tcp_send_head(sk)) 2316 return false; 2317 /* Some heurestics for collapsing over SACK'd could be invented */ 2318 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2319 return false; 2320 2321 return true; 2322 } 2323 2324 /* Collapse packets in the retransmit queue to make to create 2325 * less packets on the wire. This is only done on retransmission. 2326 */ 2327 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 2328 int space) 2329 { 2330 struct tcp_sock *tp = tcp_sk(sk); 2331 struct sk_buff *skb = to, *tmp; 2332 bool first = true; 2333 2334 if (!sysctl_tcp_retrans_collapse) 2335 return; 2336 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2337 return; 2338 2339 tcp_for_write_queue_from_safe(skb, tmp, sk) { 2340 if (!tcp_can_collapse(sk, skb)) 2341 break; 2342 2343 space -= skb->len; 2344 2345 if (first) { 2346 first = false; 2347 continue; 2348 } 2349 2350 if (space < 0) 2351 break; 2352 /* Punt if not enough space exists in the first SKB for 2353 * the data in the second 2354 */ 2355 if (skb->len > skb_availroom(to)) 2356 break; 2357 2358 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 2359 break; 2360 2361 tcp_collapse_retrans(sk, to); 2362 } 2363 } 2364 2365 /* This retransmits one SKB. Policy decisions and retransmit queue 2366 * state updates are done by the caller. Returns non-zero if an 2367 * error occurred which prevented the send. 2368 */ 2369 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 2370 { 2371 struct tcp_sock *tp = tcp_sk(sk); 2372 struct inet_connection_sock *icsk = inet_csk(sk); 2373 unsigned int cur_mss; 2374 int err; 2375 2376 /* Inconslusive MTU probe */ 2377 if (icsk->icsk_mtup.probe_size) { 2378 icsk->icsk_mtup.probe_size = 0; 2379 } 2380 2381 /* Do not sent more than we queued. 1/4 is reserved for possible 2382 * copying overhead: fragmentation, tunneling, mangling etc. 2383 */ 2384 if (atomic_read(&sk->sk_wmem_alloc) > 2385 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 2386 return -EAGAIN; 2387 2388 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 2389 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 2390 BUG(); 2391 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 2392 return -ENOMEM; 2393 } 2394 2395 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 2396 return -EHOSTUNREACH; /* Routing failure or similar. */ 2397 2398 cur_mss = tcp_current_mss(sk); 2399 2400 /* If receiver has shrunk his window, and skb is out of 2401 * new window, do not retransmit it. The exception is the 2402 * case, when window is shrunk to zero. In this case 2403 * our retransmit serves as a zero window probe. 2404 */ 2405 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 2406 TCP_SKB_CB(skb)->seq != tp->snd_una) 2407 return -EAGAIN; 2408 2409 if (skb->len > cur_mss) { 2410 if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 2411 return -ENOMEM; /* We'll try again later. */ 2412 } else { 2413 int oldpcount = tcp_skb_pcount(skb); 2414 2415 if (unlikely(oldpcount > 1)) { 2416 if (skb_unclone(skb, GFP_ATOMIC)) 2417 return -ENOMEM; 2418 tcp_init_tso_segs(sk, skb, cur_mss); 2419 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 2420 } 2421 } 2422 2423 tcp_retrans_try_collapse(sk, skb, cur_mss); 2424 2425 /* Make a copy, if the first transmission SKB clone we made 2426 * is still in somebody's hands, else make a clone. 2427 */ 2428 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2429 2430 /* make sure skb->data is aligned on arches that require it 2431 * and check if ack-trimming & collapsing extended the headroom 2432 * beyond what csum_start can cover. 2433 */ 2434 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 2435 skb_headroom(skb) >= 0xFFFF)) { 2436 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, 2437 GFP_ATOMIC); 2438 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2439 -ENOBUFS; 2440 } else { 2441 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2442 } 2443 2444 if (likely(!err)) 2445 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 2446 return err; 2447 } 2448 2449 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 2450 { 2451 struct tcp_sock *tp = tcp_sk(sk); 2452 int err = __tcp_retransmit_skb(sk, skb); 2453 2454 if (err == 0) { 2455 /* Update global TCP statistics. */ 2456 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 2457 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2458 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2459 tp->total_retrans++; 2460 2461 #if FASTRETRANS_DEBUG > 0 2462 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2463 net_dbg_ratelimited("retrans_out leaked\n"); 2464 } 2465 #endif 2466 if (!tp->retrans_out) 2467 tp->lost_retrans_low = tp->snd_nxt; 2468 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 2469 tp->retrans_out += tcp_skb_pcount(skb); 2470 2471 /* Save stamp of the first retransmit. */ 2472 if (!tp->retrans_stamp) 2473 tp->retrans_stamp = TCP_SKB_CB(skb)->when; 2474 2475 tp->undo_retrans += tcp_skb_pcount(skb); 2476 2477 /* snd_nxt is stored to detect loss of retransmitted segment, 2478 * see tcp_input.c tcp_sacktag_write_queue(). 2479 */ 2480 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 2481 } else { 2482 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); 2483 } 2484 return err; 2485 } 2486 2487 /* Check if we forward retransmits are possible in the current 2488 * window/congestion state. 2489 */ 2490 static bool tcp_can_forward_retransmit(struct sock *sk) 2491 { 2492 const struct inet_connection_sock *icsk = inet_csk(sk); 2493 const struct tcp_sock *tp = tcp_sk(sk); 2494 2495 /* Forward retransmissions are possible only during Recovery. */ 2496 if (icsk->icsk_ca_state != TCP_CA_Recovery) 2497 return false; 2498 2499 /* No forward retransmissions in Reno are possible. */ 2500 if (tcp_is_reno(tp)) 2501 return false; 2502 2503 /* Yeah, we have to make difficult choice between forward transmission 2504 * and retransmission... Both ways have their merits... 2505 * 2506 * For now we do not retransmit anything, while we have some new 2507 * segments to send. In the other cases, follow rule 3 for 2508 * NextSeg() specified in RFC3517. 2509 */ 2510 2511 if (tcp_may_send_now(sk)) 2512 return false; 2513 2514 return true; 2515 } 2516 2517 /* This gets called after a retransmit timeout, and the initially 2518 * retransmitted data is acknowledged. It tries to continue 2519 * resending the rest of the retransmit queue, until either 2520 * we've sent it all or the congestion window limit is reached. 2521 * If doing SACK, the first ACK which comes back for a timeout 2522 * based retransmit packet might feed us FACK information again. 2523 * If so, we use it to avoid unnecessarily retransmissions. 2524 */ 2525 void tcp_xmit_retransmit_queue(struct sock *sk) 2526 { 2527 const struct inet_connection_sock *icsk = inet_csk(sk); 2528 struct tcp_sock *tp = tcp_sk(sk); 2529 struct sk_buff *skb; 2530 struct sk_buff *hole = NULL; 2531 u32 last_lost; 2532 int mib_idx; 2533 int fwd_rexmitting = 0; 2534 2535 if (!tp->packets_out) 2536 return; 2537 2538 if (!tp->lost_out) 2539 tp->retransmit_high = tp->snd_una; 2540 2541 if (tp->retransmit_skb_hint) { 2542 skb = tp->retransmit_skb_hint; 2543 last_lost = TCP_SKB_CB(skb)->end_seq; 2544 if (after(last_lost, tp->retransmit_high)) 2545 last_lost = tp->retransmit_high; 2546 } else { 2547 skb = tcp_write_queue_head(sk); 2548 last_lost = tp->snd_una; 2549 } 2550 2551 tcp_for_write_queue_from(skb, sk) { 2552 __u8 sacked = TCP_SKB_CB(skb)->sacked; 2553 2554 if (skb == tcp_send_head(sk)) 2555 break; 2556 /* we could do better than to assign each time */ 2557 if (hole == NULL) 2558 tp->retransmit_skb_hint = skb; 2559 2560 /* Assume this retransmit will generate 2561 * only one packet for congestion window 2562 * calculation purposes. This works because 2563 * tcp_retransmit_skb() will chop up the 2564 * packet to be MSS sized and all the 2565 * packet counting works out. 2566 */ 2567 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 2568 return; 2569 2570 if (fwd_rexmitting) { 2571 begin_fwd: 2572 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2573 break; 2574 mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 2575 2576 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2577 tp->retransmit_high = last_lost; 2578 if (!tcp_can_forward_retransmit(sk)) 2579 break; 2580 /* Backtrack if necessary to non-L'ed skb */ 2581 if (hole != NULL) { 2582 skb = hole; 2583 hole = NULL; 2584 } 2585 fwd_rexmitting = 1; 2586 goto begin_fwd; 2587 2588 } else if (!(sacked & TCPCB_LOST)) { 2589 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 2590 hole = skb; 2591 continue; 2592 2593 } else { 2594 last_lost = TCP_SKB_CB(skb)->end_seq; 2595 if (icsk->icsk_ca_state != TCP_CA_Loss) 2596 mib_idx = LINUX_MIB_TCPFASTRETRANS; 2597 else 2598 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 2599 } 2600 2601 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 2602 continue; 2603 2604 if (tcp_retransmit_skb(sk, skb)) 2605 return; 2606 2607 NET_INC_STATS_BH(sock_net(sk), mib_idx); 2608 2609 if (tcp_in_cwnd_reduction(sk)) 2610 tp->prr_out += tcp_skb_pcount(skb); 2611 2612 if (skb == tcp_write_queue_head(sk)) 2613 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 2614 inet_csk(sk)->icsk_rto, 2615 TCP_RTO_MAX); 2616 } 2617 } 2618 2619 /* Send a fin. The caller locks the socket for us. This cannot be 2620 * allowed to fail queueing a FIN frame under any circumstances. 2621 */ 2622 void tcp_send_fin(struct sock *sk) 2623 { 2624 struct tcp_sock *tp = tcp_sk(sk); 2625 struct sk_buff *skb = tcp_write_queue_tail(sk); 2626 int mss_now; 2627 2628 /* Optimization, tack on the FIN if we have a queue of 2629 * unsent frames. But be careful about outgoing SACKS 2630 * and IP options. 2631 */ 2632 mss_now = tcp_current_mss(sk); 2633 2634 if (tcp_send_head(sk) != NULL) { 2635 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; 2636 TCP_SKB_CB(skb)->end_seq++; 2637 tp->write_seq++; 2638 } else { 2639 /* Socket is locked, keep trying until memory is available. */ 2640 for (;;) { 2641 skb = alloc_skb_fclone(MAX_TCP_HEADER, 2642 sk->sk_allocation); 2643 if (skb) 2644 break; 2645 yield(); 2646 } 2647 2648 /* Reserve space for headers and prepare control bits. */ 2649 skb_reserve(skb, MAX_TCP_HEADER); 2650 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2651 tcp_init_nondata_skb(skb, tp->write_seq, 2652 TCPHDR_ACK | TCPHDR_FIN); 2653 tcp_queue_skb(sk, skb); 2654 } 2655 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 2656 } 2657 2658 /* We get here when a process closes a file descriptor (either due to 2659 * an explicit close() or as a byproduct of exit()'ing) and there 2660 * was unread data in the receive queue. This behavior is recommended 2661 * by RFC 2525, section 2.17. -DaveM 2662 */ 2663 void tcp_send_active_reset(struct sock *sk, gfp_t priority) 2664 { 2665 struct sk_buff *skb; 2666 2667 /* NOTE: No TCP options attached and we never retransmit this. */ 2668 skb = alloc_skb(MAX_TCP_HEADER, priority); 2669 if (!skb) { 2670 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 2671 return; 2672 } 2673 2674 /* Reserve space for headers and prepare control bits. */ 2675 skb_reserve(skb, MAX_TCP_HEADER); 2676 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2677 TCPHDR_ACK | TCPHDR_RST); 2678 /* Send it off. */ 2679 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2680 if (tcp_transmit_skb(sk, skb, 0, priority)) 2681 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 2682 2683 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 2684 } 2685 2686 /* Send a crossed SYN-ACK during socket establishment. 2687 * WARNING: This routine must only be called when we have already sent 2688 * a SYN packet that crossed the incoming SYN that caused this routine 2689 * to get called. If this assumption fails then the initial rcv_wnd 2690 * and rcv_wscale values will not be correct. 2691 */ 2692 int tcp_send_synack(struct sock *sk) 2693 { 2694 struct sk_buff *skb; 2695 2696 skb = tcp_write_queue_head(sk); 2697 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 2698 pr_debug("%s: wrong queue state\n", __func__); 2699 return -EFAULT; 2700 } 2701 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 2702 if (skb_cloned(skb)) { 2703 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 2704 if (nskb == NULL) 2705 return -ENOMEM; 2706 tcp_unlink_write_queue(skb, sk); 2707 skb_header_release(nskb); 2708 __tcp_add_write_queue_head(sk, nskb); 2709 sk_wmem_free_skb(sk, skb); 2710 sk->sk_wmem_queued += nskb->truesize; 2711 sk_mem_charge(sk, nskb->truesize); 2712 skb = nskb; 2713 } 2714 2715 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 2716 TCP_ECN_send_synack(tcp_sk(sk), skb); 2717 } 2718 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2719 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2720 } 2721 2722 /** 2723 * tcp_make_synack - Prepare a SYN-ACK. 2724 * sk: listener socket 2725 * dst: dst entry attached to the SYNACK 2726 * req: request_sock pointer 2727 * 2728 * Allocate one skb and build a SYNACK packet. 2729 * @dst is consumed : Caller should not use it again. 2730 */ 2731 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2732 struct request_sock *req, 2733 struct tcp_fastopen_cookie *foc) 2734 { 2735 struct tcp_out_options opts; 2736 struct inet_request_sock *ireq = inet_rsk(req); 2737 struct tcp_sock *tp = tcp_sk(sk); 2738 struct tcphdr *th; 2739 struct sk_buff *skb; 2740 struct tcp_md5sig_key *md5; 2741 int tcp_header_size; 2742 int mss; 2743 2744 skb = sock_wmalloc(sk, MAX_TCP_HEADER, 1, GFP_ATOMIC); 2745 if (unlikely(!skb)) { 2746 dst_release(dst); 2747 return NULL; 2748 } 2749 /* Reserve space for headers. */ 2750 skb_reserve(skb, MAX_TCP_HEADER); 2751 2752 skb_dst_set(skb, dst); 2753 security_skb_owned_by(skb, sk); 2754 2755 mss = dst_metric_advmss(dst); 2756 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2757 mss = tp->rx_opt.user_mss; 2758 2759 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 2760 __u8 rcv_wscale; 2761 /* Set this up on the first call only */ 2762 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 2763 2764 /* limit the window selection if the user enforce a smaller rx buffer */ 2765 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2766 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) 2767 req->window_clamp = tcp_full_space(sk); 2768 2769 /* tcp_full_space because it is guaranteed to be the first packet */ 2770 tcp_select_initial_window(tcp_full_space(sk), 2771 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 2772 &req->rcv_wnd, 2773 &req->window_clamp, 2774 ireq->wscale_ok, 2775 &rcv_wscale, 2776 dst_metric(dst, RTAX_INITRWND)); 2777 ireq->rcv_wscale = rcv_wscale; 2778 } 2779 2780 memset(&opts, 0, sizeof(opts)); 2781 #ifdef CONFIG_SYN_COOKIES 2782 if (unlikely(req->cookie_ts)) 2783 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 2784 else 2785 #endif 2786 TCP_SKB_CB(skb)->when = tcp_time_stamp; 2787 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5, 2788 foc) + sizeof(*th); 2789 2790 skb_push(skb, tcp_header_size); 2791 skb_reset_transport_header(skb); 2792 2793 th = tcp_hdr(skb); 2794 memset(th, 0, sizeof(struct tcphdr)); 2795 th->syn = 1; 2796 th->ack = 1; 2797 TCP_ECN_make_synack(req, th); 2798 th->source = htons(ireq->ir_num); 2799 th->dest = ireq->ir_rmt_port; 2800 /* Setting of flags are superfluous here for callers (and ECE is 2801 * not even correctly set) 2802 */ 2803 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2804 TCPHDR_SYN | TCPHDR_ACK); 2805 2806 th->seq = htonl(TCP_SKB_CB(skb)->seq); 2807 /* XXX data is queued and acked as is. No buffer/window check */ 2808 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 2809 2810 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2811 th->window = htons(min(req->rcv_wnd, 65535U)); 2812 tcp_options_write((__be32 *)(th + 1), tp, &opts); 2813 th->doff = (tcp_header_size >> 2); 2814 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS); 2815 2816 #ifdef CONFIG_TCP_MD5SIG 2817 /* Okay, we have all we need - do the md5 hash if needed */ 2818 if (md5) { 2819 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 2820 md5, NULL, req, skb); 2821 } 2822 #endif 2823 2824 return skb; 2825 } 2826 EXPORT_SYMBOL(tcp_make_synack); 2827 2828 /* Do all connect socket setups that can be done AF independent. */ 2829 static void tcp_connect_init(struct sock *sk) 2830 { 2831 const struct dst_entry *dst = __sk_dst_get(sk); 2832 struct tcp_sock *tp = tcp_sk(sk); 2833 __u8 rcv_wscale; 2834 2835 /* We'll fix this up when we get a response from the other end. 2836 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 2837 */ 2838 tp->tcp_header_len = sizeof(struct tcphdr) + 2839 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 2840 2841 #ifdef CONFIG_TCP_MD5SIG 2842 if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2843 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2844 #endif 2845 2846 /* If user gave his TCP_MAXSEG, record it to clamp */ 2847 if (tp->rx_opt.user_mss) 2848 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 2849 tp->max_window = 0; 2850 tcp_mtup_init(sk); 2851 tcp_sync_mss(sk, dst_mtu(dst)); 2852 2853 if (!tp->window_clamp) 2854 tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 2855 tp->advmss = dst_metric_advmss(dst); 2856 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2857 tp->advmss = tp->rx_opt.user_mss; 2858 2859 tcp_initialize_rcv_mss(sk); 2860 2861 /* limit the window selection if the user enforce a smaller rx buffer */ 2862 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2863 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 2864 tp->window_clamp = tcp_full_space(sk); 2865 2866 tcp_select_initial_window(tcp_full_space(sk), 2867 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 2868 &tp->rcv_wnd, 2869 &tp->window_clamp, 2870 sysctl_tcp_window_scaling, 2871 &rcv_wscale, 2872 dst_metric(dst, RTAX_INITRWND)); 2873 2874 tp->rx_opt.rcv_wscale = rcv_wscale; 2875 tp->rcv_ssthresh = tp->rcv_wnd; 2876 2877 sk->sk_err = 0; 2878 sock_reset_flag(sk, SOCK_DONE); 2879 tp->snd_wnd = 0; 2880 tcp_init_wl(tp, 0); 2881 tp->snd_una = tp->write_seq; 2882 tp->snd_sml = tp->write_seq; 2883 tp->snd_up = tp->write_seq; 2884 tp->snd_nxt = tp->write_seq; 2885 2886 if (likely(!tp->repair)) 2887 tp->rcv_nxt = 0; 2888 else 2889 tp->rcv_tstamp = tcp_time_stamp; 2890 tp->rcv_wup = tp->rcv_nxt; 2891 tp->copied_seq = tp->rcv_nxt; 2892 2893 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2894 inet_csk(sk)->icsk_retransmits = 0; 2895 tcp_clear_retrans(tp); 2896 } 2897 2898 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 2899 { 2900 struct tcp_sock *tp = tcp_sk(sk); 2901 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 2902 2903 tcb->end_seq += skb->len; 2904 skb_header_release(skb); 2905 __tcp_add_write_queue_tail(sk, skb); 2906 sk->sk_wmem_queued += skb->truesize; 2907 sk_mem_charge(sk, skb->truesize); 2908 tp->write_seq = tcb->end_seq; 2909 tp->packets_out += tcp_skb_pcount(skb); 2910 } 2911 2912 /* Build and send a SYN with data and (cached) Fast Open cookie. However, 2913 * queue a data-only packet after the regular SYN, such that regular SYNs 2914 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 2915 * only the SYN sequence, the data are retransmitted in the first ACK. 2916 * If cookie is not cached or other error occurs, falls back to send a 2917 * regular SYN with Fast Open cookie request option. 2918 */ 2919 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 2920 { 2921 struct tcp_sock *tp = tcp_sk(sk); 2922 struct tcp_fastopen_request *fo = tp->fastopen_req; 2923 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen; 2924 struct sk_buff *syn_data = NULL, *data; 2925 unsigned long last_syn_loss = 0; 2926 2927 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 2928 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, 2929 &syn_loss, &last_syn_loss); 2930 /* Recurring FO SYN losses: revert to regular handshake temporarily */ 2931 if (syn_loss > 1 && 2932 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { 2933 fo->cookie.len = -1; 2934 goto fallback; 2935 } 2936 2937 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) 2938 fo->cookie.len = -1; 2939 else if (fo->cookie.len <= 0) 2940 goto fallback; 2941 2942 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 2943 * user-MSS. Reserve maximum option space for middleboxes that add 2944 * private TCP options. The cost is reduced data space in SYN :( 2945 */ 2946 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) 2947 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 2948 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 2949 MAX_TCP_OPTION_SPACE; 2950 2951 space = min_t(size_t, space, fo->size); 2952 2953 /* limit to order-0 allocations */ 2954 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 2955 2956 syn_data = skb_copy_expand(syn, MAX_TCP_HEADER, space, 2957 sk->sk_allocation); 2958 if (syn_data == NULL) 2959 goto fallback; 2960 2961 for (i = 0; i < iovlen && syn_data->len < space; ++i) { 2962 struct iovec *iov = &fo->data->msg_iov[i]; 2963 unsigned char __user *from = iov->iov_base; 2964 int len = iov->iov_len; 2965 2966 if (syn_data->len + len > space) 2967 len = space - syn_data->len; 2968 else if (i + 1 == iovlen) 2969 /* No more data pending in inet_wait_for_connect() */ 2970 fo->data = NULL; 2971 2972 if (skb_add_data(syn_data, from, len)) 2973 goto fallback; 2974 } 2975 2976 /* Queue a data-only packet after the regular SYN for retransmission */ 2977 data = pskb_copy(syn_data, sk->sk_allocation); 2978 if (data == NULL) 2979 goto fallback; 2980 TCP_SKB_CB(data)->seq++; 2981 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN; 2982 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH); 2983 tcp_connect_queue_skb(sk, data); 2984 fo->copied = data->len; 2985 2986 /* syn_data is about to be sent, we need to take current time stamps 2987 * for the packets that are in write queue : SYN packet and DATA 2988 */ 2989 skb_mstamp_get(&syn->skb_mstamp); 2990 data->skb_mstamp = syn->skb_mstamp; 2991 2992 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) { 2993 tp->syn_data = (fo->copied > 0); 2994 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 2995 goto done; 2996 } 2997 syn_data = NULL; 2998 2999 fallback: 3000 /* Send a regular SYN with Fast Open cookie request option */ 3001 if (fo->cookie.len > 0) 3002 fo->cookie.len = 0; 3003 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3004 if (err) 3005 tp->syn_fastopen = 0; 3006 kfree_skb(syn_data); 3007 done: 3008 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3009 return err; 3010 } 3011 3012 /* Build a SYN and send it off. */ 3013 int tcp_connect(struct sock *sk) 3014 { 3015 struct tcp_sock *tp = tcp_sk(sk); 3016 struct sk_buff *buff; 3017 int err; 3018 3019 tcp_connect_init(sk); 3020 3021 if (unlikely(tp->repair)) { 3022 tcp_finish_connect(sk, NULL); 3023 return 0; 3024 } 3025 3026 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 3027 if (unlikely(buff == NULL)) 3028 return -ENOBUFS; 3029 3030 /* Reserve space for headers. */ 3031 skb_reserve(buff, MAX_TCP_HEADER); 3032 3033 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 3034 tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp; 3035 tcp_connect_queue_skb(sk, buff); 3036 TCP_ECN_send_syn(sk, buff); 3037 3038 /* Send off SYN; include data in Fast Open. */ 3039 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3040 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3041 if (err == -ECONNREFUSED) 3042 return err; 3043 3044 /* We change tp->snd_nxt after the tcp_transmit_skb() call 3045 * in order to make this packet get counted in tcpOutSegs. 3046 */ 3047 tp->snd_nxt = tp->write_seq; 3048 tp->pushed_seq = tp->write_seq; 3049 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 3050 3051 /* Timer for repeating the SYN until an answer. */ 3052 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 3053 inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 3054 return 0; 3055 } 3056 EXPORT_SYMBOL(tcp_connect); 3057 3058 /* Send out a delayed ack, the caller does the policy checking 3059 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 3060 * for details. 3061 */ 3062 void tcp_send_delayed_ack(struct sock *sk) 3063 { 3064 struct inet_connection_sock *icsk = inet_csk(sk); 3065 int ato = icsk->icsk_ack.ato; 3066 unsigned long timeout; 3067 3068 if (ato > TCP_DELACK_MIN) { 3069 const struct tcp_sock *tp = tcp_sk(sk); 3070 int max_ato = HZ / 2; 3071 3072 if (icsk->icsk_ack.pingpong || 3073 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 3074 max_ato = TCP_DELACK_MAX; 3075 3076 /* Slow path, intersegment interval is "high". */ 3077 3078 /* If some rtt estimate is known, use it to bound delayed ack. 3079 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 3080 * directly. 3081 */ 3082 if (tp->srtt_us) { 3083 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3084 TCP_DELACK_MIN); 3085 3086 if (rtt < max_ato) 3087 max_ato = rtt; 3088 } 3089 3090 ato = min(ato, max_ato); 3091 } 3092 3093 /* Stay within the limit we were given */ 3094 timeout = jiffies + ato; 3095 3096 /* Use new timeout only if there wasn't a older one earlier. */ 3097 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 3098 /* If delack timer was blocked or is about to expire, 3099 * send ACK now. 3100 */ 3101 if (icsk->icsk_ack.blocked || 3102 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 3103 tcp_send_ack(sk); 3104 return; 3105 } 3106 3107 if (!time_before(timeout, icsk->icsk_ack.timeout)) 3108 timeout = icsk->icsk_ack.timeout; 3109 } 3110 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3111 icsk->icsk_ack.timeout = timeout; 3112 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 3113 } 3114 3115 /* This routine sends an ack and also updates the window. */ 3116 void tcp_send_ack(struct sock *sk) 3117 { 3118 struct sk_buff *buff; 3119 3120 /* If we have been reset, we may not send again. */ 3121 if (sk->sk_state == TCP_CLOSE) 3122 return; 3123 3124 /* We are not putting this on the write queue, so 3125 * tcp_transmit_skb() will set the ownership to this 3126 * sock. 3127 */ 3128 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 3129 if (buff == NULL) { 3130 inet_csk_schedule_ack(sk); 3131 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 3132 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 3133 TCP_DELACK_MAX, TCP_RTO_MAX); 3134 return; 3135 } 3136 3137 /* Reserve space for headers and prepare control bits. */ 3138 skb_reserve(buff, MAX_TCP_HEADER); 3139 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 3140 3141 /* Send it off, this clears delayed acks for us. */ 3142 TCP_SKB_CB(buff)->when = tcp_time_stamp; 3143 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC)); 3144 } 3145 3146 /* This routine sends a packet with an out of date sequence 3147 * number. It assumes the other end will try to ack it. 3148 * 3149 * Question: what should we make while urgent mode? 3150 * 4.4BSD forces sending single byte of data. We cannot send 3151 * out of window data, because we have SND.NXT==SND.MAX... 3152 * 3153 * Current solution: to send TWO zero-length segments in urgent mode: 3154 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 3155 * out-of-date with SND.UNA-1 to probe window. 3156 */ 3157 static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 3158 { 3159 struct tcp_sock *tp = tcp_sk(sk); 3160 struct sk_buff *skb; 3161 3162 /* We don't queue it, tcp_transmit_skb() sets ownership. */ 3163 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); 3164 if (skb == NULL) 3165 return -1; 3166 3167 /* Reserve space for headers and set control bits. */ 3168 skb_reserve(skb, MAX_TCP_HEADER); 3169 /* Use a previous sequence. This should cause the other 3170 * end to send an ack. Don't queue or clone SKB, just 3171 * send it. 3172 */ 3173 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 3174 TCP_SKB_CB(skb)->when = tcp_time_stamp; 3175 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 3176 } 3177 3178 void tcp_send_window_probe(struct sock *sk) 3179 { 3180 if (sk->sk_state == TCP_ESTABLISHED) { 3181 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 3182 tcp_xmit_probe_skb(sk, 0); 3183 } 3184 } 3185 3186 /* Initiate keepalive or window probe from timer. */ 3187 int tcp_write_wakeup(struct sock *sk) 3188 { 3189 struct tcp_sock *tp = tcp_sk(sk); 3190 struct sk_buff *skb; 3191 3192 if (sk->sk_state == TCP_CLOSE) 3193 return -1; 3194 3195 if ((skb = tcp_send_head(sk)) != NULL && 3196 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 3197 int err; 3198 unsigned int mss = tcp_current_mss(sk); 3199 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 3200 3201 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 3202 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 3203 3204 /* We are probing the opening of a window 3205 * but the window size is != 0 3206 * must have been a result SWS avoidance ( sender ) 3207 */ 3208 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 3209 skb->len > mss) { 3210 seg_size = min(seg_size, mss); 3211 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3212 if (tcp_fragment(sk, skb, seg_size, mss)) 3213 return -1; 3214 } else if (!tcp_skb_pcount(skb)) 3215 tcp_set_skb_tso_segs(sk, skb, mss); 3216 3217 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3218 TCP_SKB_CB(skb)->when = tcp_time_stamp; 3219 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 3220 if (!err) 3221 tcp_event_new_data_sent(sk, skb); 3222 return err; 3223 } else { 3224 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3225 tcp_xmit_probe_skb(sk, 1); 3226 return tcp_xmit_probe_skb(sk, 0); 3227 } 3228 } 3229 3230 /* A window probe timeout has occurred. If window is not closed send 3231 * a partial packet else a zero probe. 3232 */ 3233 void tcp_send_probe0(struct sock *sk) 3234 { 3235 struct inet_connection_sock *icsk = inet_csk(sk); 3236 struct tcp_sock *tp = tcp_sk(sk); 3237 int err; 3238 3239 err = tcp_write_wakeup(sk); 3240 3241 if (tp->packets_out || !tcp_send_head(sk)) { 3242 /* Cancel probe timer, if it is not required. */ 3243 icsk->icsk_probes_out = 0; 3244 icsk->icsk_backoff = 0; 3245 return; 3246 } 3247 3248 if (err <= 0) { 3249 if (icsk->icsk_backoff < sysctl_tcp_retries2) 3250 icsk->icsk_backoff++; 3251 icsk->icsk_probes_out++; 3252 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3253 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 3254 TCP_RTO_MAX); 3255 } else { 3256 /* If packet was not sent due to local congestion, 3257 * do not backoff and do not remember icsk_probes_out. 3258 * Let local senders to fight for local resources. 3259 * 3260 * Use accumulated backoff yet. 3261 */ 3262 if (!icsk->icsk_probes_out) 3263 icsk->icsk_probes_out = 1; 3264 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 3265 min(icsk->icsk_rto << icsk->icsk_backoff, 3266 TCP_RESOURCE_PROBE_INTERVAL), 3267 TCP_RTO_MAX); 3268 } 3269 } 3270