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 * Version: $Id: tcp_output.c,v 1.146 2002/02/01 22:01:04 davem Exp $ 9 * 10 * Authors: Ross Biro, <bir7@leland.Stanford.Edu> 11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 12 * Mark Evans, <evansmp@uhura.aston.ac.uk> 13 * Corey Minyard <wf-rch!minyard@relay.EU.net> 14 * Florian La Roche, <flla@stud.uni-sb.de> 15 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 16 * Linus Torvalds, <torvalds@cs.helsinki.fi> 17 * Alan Cox, <gw4pts@gw4pts.ampr.org> 18 * Matthew Dillon, <dillon@apollo.west.oic.com> 19 * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 20 * Jorge Cwik, <jorge@laser.satlink.net> 21 */ 22 23 /* 24 * Changes: Pedro Roque : Retransmit queue handled by TCP. 25 * : Fragmentation on mtu decrease 26 * : Segment collapse on retransmit 27 * : AF independence 28 * 29 * Linus Torvalds : send_delayed_ack 30 * David S. Miller : Charge memory using the right skb 31 * during syn/ack processing. 32 * David S. Miller : Output engine completely rewritten. 33 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. 34 * Cacophonix Gaul : draft-minshall-nagle-01 35 * J Hadi Salim : ECN support 36 * 37 */ 38 39 #include <net/tcp.h> 40 41 #include <linux/compiler.h> 42 #include <linux/module.h> 43 #include <linux/smp_lock.h> 44 45 /* People can turn this off for buggy TCP's found in printers etc. */ 46 int sysctl_tcp_retrans_collapse = 1; 47 48 /* This limits the percentage of the congestion window which we 49 * will allow a single TSO frame to consume. Building TSO frames 50 * which are too large can cause TCP streams to be bursty. 51 */ 52 int sysctl_tcp_tso_win_divisor = 8; 53 54 static inline void update_send_head(struct sock *sk, struct tcp_sock *tp, 55 struct sk_buff *skb) 56 { 57 sk->sk_send_head = skb->next; 58 if (sk->sk_send_head == (struct sk_buff *)&sk->sk_write_queue) 59 sk->sk_send_head = NULL; 60 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 61 tcp_packets_out_inc(sk, tp, skb); 62 } 63 64 /* SND.NXT, if window was not shrunk. 65 * If window has been shrunk, what should we make? It is not clear at all. 66 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 67 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 68 * invalid. OK, let's make this for now: 69 */ 70 static inline __u32 tcp_acceptable_seq(struct sock *sk, struct tcp_sock *tp) 71 { 72 if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt)) 73 return tp->snd_nxt; 74 else 75 return tp->snd_una+tp->snd_wnd; 76 } 77 78 /* Calculate mss to advertise in SYN segment. 79 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 80 * 81 * 1. It is independent of path mtu. 82 * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 83 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 84 * attached devices, because some buggy hosts are confused by 85 * large MSS. 86 * 4. We do not make 3, we advertise MSS, calculated from first 87 * hop device mtu, but allow to raise it to ip_rt_min_advmss. 88 * This may be overridden via information stored in routing table. 89 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 90 * probably even Jumbo". 91 */ 92 static __u16 tcp_advertise_mss(struct sock *sk) 93 { 94 struct tcp_sock *tp = tcp_sk(sk); 95 struct dst_entry *dst = __sk_dst_get(sk); 96 int mss = tp->advmss; 97 98 if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) { 99 mss = dst_metric(dst, RTAX_ADVMSS); 100 tp->advmss = mss; 101 } 102 103 return (__u16)mss; 104 } 105 106 /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 107 * This is the first part of cwnd validation mechanism. */ 108 static void tcp_cwnd_restart(struct tcp_sock *tp, struct dst_entry *dst) 109 { 110 s32 delta = tcp_time_stamp - tp->lsndtime; 111 u32 restart_cwnd = tcp_init_cwnd(tp, dst); 112 u32 cwnd = tp->snd_cwnd; 113 114 if (tcp_is_vegas(tp)) 115 tcp_vegas_enable(tp); 116 117 tp->snd_ssthresh = tcp_current_ssthresh(tp); 118 restart_cwnd = min(restart_cwnd, cwnd); 119 120 while ((delta -= tp->rto) > 0 && cwnd > restart_cwnd) 121 cwnd >>= 1; 122 tp->snd_cwnd = max(cwnd, restart_cwnd); 123 tp->snd_cwnd_stamp = tcp_time_stamp; 124 tp->snd_cwnd_used = 0; 125 } 126 127 static inline void tcp_event_data_sent(struct tcp_sock *tp, 128 struct sk_buff *skb, struct sock *sk) 129 { 130 u32 now = tcp_time_stamp; 131 132 if (!tp->packets_out && (s32)(now - tp->lsndtime) > tp->rto) 133 tcp_cwnd_restart(tp, __sk_dst_get(sk)); 134 135 tp->lsndtime = now; 136 137 /* If it is a reply for ato after last received 138 * packet, enter pingpong mode. 139 */ 140 if ((u32)(now - tp->ack.lrcvtime) < tp->ack.ato) 141 tp->ack.pingpong = 1; 142 } 143 144 static __inline__ void tcp_event_ack_sent(struct sock *sk) 145 { 146 struct tcp_sock *tp = tcp_sk(sk); 147 148 tcp_dec_quickack_mode(tp); 149 tcp_clear_xmit_timer(sk, TCP_TIME_DACK); 150 } 151 152 /* Determine a window scaling and initial window to offer. 153 * Based on the assumption that the given amount of space 154 * will be offered. Store the results in the tp structure. 155 * NOTE: for smooth operation initial space offering should 156 * be a multiple of mss if possible. We assume here that mss >= 1. 157 * This MUST be enforced by all callers. 158 */ 159 void tcp_select_initial_window(int __space, __u32 mss, 160 __u32 *rcv_wnd, __u32 *window_clamp, 161 int wscale_ok, __u8 *rcv_wscale) 162 { 163 unsigned int space = (__space < 0 ? 0 : __space); 164 165 /* If no clamp set the clamp to the max possible scaled window */ 166 if (*window_clamp == 0) 167 (*window_clamp) = (65535 << 14); 168 space = min(*window_clamp, space); 169 170 /* Quantize space offering to a multiple of mss if possible. */ 171 if (space > mss) 172 space = (space / mss) * mss; 173 174 /* NOTE: offering an initial window larger than 32767 175 * will break some buggy TCP stacks. We try to be nice. 176 * If we are not window scaling, then this truncates 177 * our initial window offering to 32k. There should also 178 * be a sysctl option to stop being nice. 179 */ 180 (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 181 (*rcv_wscale) = 0; 182 if (wscale_ok) { 183 /* Set window scaling on max possible window 184 * See RFC1323 for an explanation of the limit to 14 185 */ 186 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 187 while (space > 65535 && (*rcv_wscale) < 14) { 188 space >>= 1; 189 (*rcv_wscale)++; 190 } 191 } 192 193 /* Set initial window to value enough for senders, 194 * following RFC1414. Senders, not following this RFC, 195 * will be satisfied with 2. 196 */ 197 if (mss > (1<<*rcv_wscale)) { 198 int init_cwnd = 4; 199 if (mss > 1460*3) 200 init_cwnd = 2; 201 else if (mss > 1460) 202 init_cwnd = 3; 203 if (*rcv_wnd > init_cwnd*mss) 204 *rcv_wnd = init_cwnd*mss; 205 } 206 207 /* Set the clamp no higher than max representable value */ 208 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 209 } 210 211 /* Chose a new window to advertise, update state in tcp_sock for the 212 * socket, and return result with RFC1323 scaling applied. The return 213 * value can be stuffed directly into th->window for an outgoing 214 * frame. 215 */ 216 static __inline__ u16 tcp_select_window(struct sock *sk) 217 { 218 struct tcp_sock *tp = tcp_sk(sk); 219 u32 cur_win = tcp_receive_window(tp); 220 u32 new_win = __tcp_select_window(sk); 221 222 /* Never shrink the offered window */ 223 if(new_win < cur_win) { 224 /* Danger Will Robinson! 225 * Don't update rcv_wup/rcv_wnd here or else 226 * we will not be able to advertise a zero 227 * window in time. --DaveM 228 * 229 * Relax Will Robinson. 230 */ 231 new_win = cur_win; 232 } 233 tp->rcv_wnd = new_win; 234 tp->rcv_wup = tp->rcv_nxt; 235 236 /* Make sure we do not exceed the maximum possible 237 * scaled window. 238 */ 239 if (!tp->rx_opt.rcv_wscale) 240 new_win = min(new_win, MAX_TCP_WINDOW); 241 else 242 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 243 244 /* RFC1323 scaling applied */ 245 new_win >>= tp->rx_opt.rcv_wscale; 246 247 /* If we advertise zero window, disable fast path. */ 248 if (new_win == 0) 249 tp->pred_flags = 0; 250 251 return new_win; 252 } 253 254 255 /* This routine actually transmits TCP packets queued in by 256 * tcp_do_sendmsg(). This is used by both the initial 257 * transmission and possible later retransmissions. 258 * All SKB's seen here are completely headerless. It is our 259 * job to build the TCP header, and pass the packet down to 260 * IP so it can do the same plus pass the packet off to the 261 * device. 262 * 263 * We are working here with either a clone of the original 264 * SKB, or a fresh unique copy made by the retransmit engine. 265 */ 266 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb) 267 { 268 if (skb != NULL) { 269 struct inet_sock *inet = inet_sk(sk); 270 struct tcp_sock *tp = tcp_sk(sk); 271 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 272 int tcp_header_size = tp->tcp_header_len; 273 struct tcphdr *th; 274 int sysctl_flags; 275 int err; 276 277 BUG_ON(!tcp_skb_pcount(skb)); 278 279 #define SYSCTL_FLAG_TSTAMPS 0x1 280 #define SYSCTL_FLAG_WSCALE 0x2 281 #define SYSCTL_FLAG_SACK 0x4 282 283 sysctl_flags = 0; 284 if (tcb->flags & TCPCB_FLAG_SYN) { 285 tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS; 286 if(sysctl_tcp_timestamps) { 287 tcp_header_size += TCPOLEN_TSTAMP_ALIGNED; 288 sysctl_flags |= SYSCTL_FLAG_TSTAMPS; 289 } 290 if(sysctl_tcp_window_scaling) { 291 tcp_header_size += TCPOLEN_WSCALE_ALIGNED; 292 sysctl_flags |= SYSCTL_FLAG_WSCALE; 293 } 294 if(sysctl_tcp_sack) { 295 sysctl_flags |= SYSCTL_FLAG_SACK; 296 if(!(sysctl_flags & SYSCTL_FLAG_TSTAMPS)) 297 tcp_header_size += TCPOLEN_SACKPERM_ALIGNED; 298 } 299 } else if (tp->rx_opt.eff_sacks) { 300 /* A SACK is 2 pad bytes, a 2 byte header, plus 301 * 2 32-bit sequence numbers for each SACK block. 302 */ 303 tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED + 304 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK)); 305 } 306 307 /* 308 * If the connection is idle and we are restarting, 309 * then we don't want to do any Vegas calculations 310 * until we get fresh RTT samples. So when we 311 * restart, we reset our Vegas state to a clean 312 * slate. After we get acks for this flight of 313 * packets, _then_ we can make Vegas calculations 314 * again. 315 */ 316 if (tcp_is_vegas(tp) && tcp_packets_in_flight(tp) == 0) 317 tcp_vegas_enable(tp); 318 319 th = (struct tcphdr *) skb_push(skb, tcp_header_size); 320 skb->h.th = th; 321 skb_set_owner_w(skb, sk); 322 323 /* Build TCP header and checksum it. */ 324 th->source = inet->sport; 325 th->dest = inet->dport; 326 th->seq = htonl(tcb->seq); 327 th->ack_seq = htonl(tp->rcv_nxt); 328 *(((__u16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | tcb->flags); 329 if (tcb->flags & TCPCB_FLAG_SYN) { 330 /* RFC1323: The window in SYN & SYN/ACK segments 331 * is never scaled. 332 */ 333 th->window = htons(tp->rcv_wnd); 334 } else { 335 th->window = htons(tcp_select_window(sk)); 336 } 337 th->check = 0; 338 th->urg_ptr = 0; 339 340 if (tp->urg_mode && 341 between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF)) { 342 th->urg_ptr = htons(tp->snd_up-tcb->seq); 343 th->urg = 1; 344 } 345 346 if (tcb->flags & TCPCB_FLAG_SYN) { 347 tcp_syn_build_options((__u32 *)(th + 1), 348 tcp_advertise_mss(sk), 349 (sysctl_flags & SYSCTL_FLAG_TSTAMPS), 350 (sysctl_flags & SYSCTL_FLAG_SACK), 351 (sysctl_flags & SYSCTL_FLAG_WSCALE), 352 tp->rx_opt.rcv_wscale, 353 tcb->when, 354 tp->rx_opt.ts_recent); 355 } else { 356 tcp_build_and_update_options((__u32 *)(th + 1), 357 tp, tcb->when); 358 359 TCP_ECN_send(sk, tp, skb, tcp_header_size); 360 } 361 tp->af_specific->send_check(sk, th, skb->len, skb); 362 363 if (tcb->flags & TCPCB_FLAG_ACK) 364 tcp_event_ack_sent(sk); 365 366 if (skb->len != tcp_header_size) 367 tcp_event_data_sent(tp, skb, sk); 368 369 TCP_INC_STATS(TCP_MIB_OUTSEGS); 370 371 err = tp->af_specific->queue_xmit(skb, 0); 372 if (err <= 0) 373 return err; 374 375 tcp_enter_cwr(tp); 376 377 /* NET_XMIT_CN is special. It does not guarantee, 378 * that this packet is lost. It tells that device 379 * is about to start to drop packets or already 380 * drops some packets of the same priority and 381 * invokes us to send less aggressively. 382 */ 383 return err == NET_XMIT_CN ? 0 : err; 384 } 385 return -ENOBUFS; 386 #undef SYSCTL_FLAG_TSTAMPS 387 #undef SYSCTL_FLAG_WSCALE 388 #undef SYSCTL_FLAG_SACK 389 } 390 391 392 /* This routine just queue's the buffer 393 * 394 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 395 * otherwise socket can stall. 396 */ 397 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 398 { 399 struct tcp_sock *tp = tcp_sk(sk); 400 401 /* Advance write_seq and place onto the write_queue. */ 402 tp->write_seq = TCP_SKB_CB(skb)->end_seq; 403 skb_header_release(skb); 404 __skb_queue_tail(&sk->sk_write_queue, skb); 405 sk_charge_skb(sk, skb); 406 407 /* Queue it, remembering where we must start sending. */ 408 if (sk->sk_send_head == NULL) 409 sk->sk_send_head = skb; 410 } 411 412 static inline void tcp_tso_set_push(struct sk_buff *skb) 413 { 414 /* Force push to be on for any TSO frames to workaround 415 * problems with busted implementations like Mac OS-X that 416 * hold off socket receive wakeups until push is seen. 417 */ 418 if (tcp_skb_pcount(skb) > 1) 419 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 420 } 421 422 /* Send _single_ skb sitting at the send head. This function requires 423 * true push pending frames to setup probe timer etc. 424 */ 425 void tcp_push_one(struct sock *sk, unsigned cur_mss) 426 { 427 struct tcp_sock *tp = tcp_sk(sk); 428 struct sk_buff *skb = sk->sk_send_head; 429 430 if (tcp_snd_test(tp, skb, cur_mss, TCP_NAGLE_PUSH)) { 431 /* Send it out now. */ 432 TCP_SKB_CB(skb)->when = tcp_time_stamp; 433 tcp_tso_set_push(skb); 434 if (!tcp_transmit_skb(sk, skb_clone(skb, sk->sk_allocation))) { 435 sk->sk_send_head = NULL; 436 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 437 tcp_packets_out_inc(sk, tp, skb); 438 return; 439 } 440 } 441 } 442 443 void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_std) 444 { 445 if (skb->len <= mss_std) { 446 /* Avoid the costly divide in the normal 447 * non-TSO case. 448 */ 449 skb_shinfo(skb)->tso_segs = 1; 450 skb_shinfo(skb)->tso_size = 0; 451 } else { 452 unsigned int factor; 453 454 factor = skb->len + (mss_std - 1); 455 factor /= mss_std; 456 skb_shinfo(skb)->tso_segs = factor; 457 skb_shinfo(skb)->tso_size = mss_std; 458 } 459 } 460 461 /* Function to create two new TCP segments. Shrinks the given segment 462 * to the specified size and appends a new segment with the rest of the 463 * packet to the list. This won't be called frequently, I hope. 464 * Remember, these are still headerless SKBs at this point. 465 */ 466 static int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len) 467 { 468 struct tcp_sock *tp = tcp_sk(sk); 469 struct sk_buff *buff; 470 int nsize; 471 u16 flags; 472 473 nsize = skb_headlen(skb) - len; 474 if (nsize < 0) 475 nsize = 0; 476 477 if (skb_cloned(skb) && 478 skb_is_nonlinear(skb) && 479 pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 480 return -ENOMEM; 481 482 /* Get a new skb... force flag on. */ 483 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 484 if (buff == NULL) 485 return -ENOMEM; /* We'll just try again later. */ 486 sk_charge_skb(sk, buff); 487 488 /* Correct the sequence numbers. */ 489 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 490 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 491 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 492 493 /* PSH and FIN should only be set in the second packet. */ 494 flags = TCP_SKB_CB(skb)->flags; 495 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); 496 TCP_SKB_CB(buff)->flags = flags; 497 TCP_SKB_CB(buff)->sacked = 498 (TCP_SKB_CB(skb)->sacked & 499 (TCPCB_LOST | TCPCB_EVER_RETRANS | TCPCB_AT_TAIL)); 500 TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL; 501 502 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_HW) { 503 /* Copy and checksum data tail into the new buffer. */ 504 buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize), 505 nsize, 0); 506 507 skb_trim(skb, len); 508 509 skb->csum = csum_block_sub(skb->csum, buff->csum, len); 510 } else { 511 skb->ip_summed = CHECKSUM_HW; 512 skb_split(skb, buff, len); 513 } 514 515 buff->ip_summed = skb->ip_summed; 516 517 /* Looks stupid, but our code really uses when of 518 * skbs, which it never sent before. --ANK 519 */ 520 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 521 522 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) { 523 tp->lost_out -= tcp_skb_pcount(skb); 524 tp->left_out -= tcp_skb_pcount(skb); 525 } 526 527 /* Fix up tso_factor for both original and new SKB. */ 528 tcp_set_skb_tso_segs(skb, tp->mss_cache_std); 529 tcp_set_skb_tso_segs(buff, tp->mss_cache_std); 530 531 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) { 532 tp->lost_out += tcp_skb_pcount(skb); 533 tp->left_out += tcp_skb_pcount(skb); 534 } 535 536 if (TCP_SKB_CB(buff)->sacked&TCPCB_LOST) { 537 tp->lost_out += tcp_skb_pcount(buff); 538 tp->left_out += tcp_skb_pcount(buff); 539 } 540 541 /* Link BUFF into the send queue. */ 542 __skb_append(skb, buff); 543 544 return 0; 545 } 546 547 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 548 * eventually). The difference is that pulled data not copied, but 549 * immediately discarded. 550 */ 551 static unsigned char *__pskb_trim_head(struct sk_buff *skb, int len) 552 { 553 int i, k, eat; 554 555 eat = len; 556 k = 0; 557 for (i=0; i<skb_shinfo(skb)->nr_frags; i++) { 558 if (skb_shinfo(skb)->frags[i].size <= eat) { 559 put_page(skb_shinfo(skb)->frags[i].page); 560 eat -= skb_shinfo(skb)->frags[i].size; 561 } else { 562 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 563 if (eat) { 564 skb_shinfo(skb)->frags[k].page_offset += eat; 565 skb_shinfo(skb)->frags[k].size -= eat; 566 eat = 0; 567 } 568 k++; 569 } 570 } 571 skb_shinfo(skb)->nr_frags = k; 572 573 skb->tail = skb->data; 574 skb->data_len -= len; 575 skb->len = skb->data_len; 576 return skb->tail; 577 } 578 579 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 580 { 581 if (skb_cloned(skb) && 582 pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 583 return -ENOMEM; 584 585 if (len <= skb_headlen(skb)) { 586 __skb_pull(skb, len); 587 } else { 588 if (__pskb_trim_head(skb, len-skb_headlen(skb)) == NULL) 589 return -ENOMEM; 590 } 591 592 TCP_SKB_CB(skb)->seq += len; 593 skb->ip_summed = CHECKSUM_HW; 594 595 skb->truesize -= len; 596 sk->sk_wmem_queued -= len; 597 sk->sk_forward_alloc += len; 598 sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 599 600 /* Any change of skb->len requires recalculation of tso 601 * factor and mss. 602 */ 603 if (tcp_skb_pcount(skb) > 1) 604 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 605 606 return 0; 607 } 608 609 /* This function synchronize snd mss to current pmtu/exthdr set. 610 611 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 612 for TCP options, but includes only bare TCP header. 613 614 tp->rx_opt.mss_clamp is mss negotiated at connection setup. 615 It is minumum of user_mss and mss received with SYN. 616 It also does not include TCP options. 617 618 tp->pmtu_cookie is last pmtu, seen by this function. 619 620 tp->mss_cache is current effective sending mss, including 621 all tcp options except for SACKs. It is evaluated, 622 taking into account current pmtu, but never exceeds 623 tp->rx_opt.mss_clamp. 624 625 NOTE1. rfc1122 clearly states that advertised MSS 626 DOES NOT include either tcp or ip options. 627 628 NOTE2. tp->pmtu_cookie and tp->mss_cache are READ ONLY outside 629 this function. --ANK (980731) 630 */ 631 632 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 633 { 634 struct tcp_sock *tp = tcp_sk(sk); 635 int mss_now; 636 637 /* Calculate base mss without TCP options: 638 It is MMS_S - sizeof(tcphdr) of rfc1122 639 */ 640 mss_now = pmtu - tp->af_specific->net_header_len - sizeof(struct tcphdr); 641 642 /* Clamp it (mss_clamp does not include tcp options) */ 643 if (mss_now > tp->rx_opt.mss_clamp) 644 mss_now = tp->rx_opt.mss_clamp; 645 646 /* Now subtract optional transport overhead */ 647 mss_now -= tp->ext_header_len; 648 649 /* Then reserve room for full set of TCP options and 8 bytes of data */ 650 if (mss_now < 48) 651 mss_now = 48; 652 653 /* Now subtract TCP options size, not including SACKs */ 654 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); 655 656 /* Bound mss with half of window */ 657 if (tp->max_window && mss_now > (tp->max_window>>1)) 658 mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len); 659 660 /* And store cached results */ 661 tp->pmtu_cookie = pmtu; 662 tp->mss_cache = tp->mss_cache_std = mss_now; 663 664 return mss_now; 665 } 666 667 /* Compute the current effective MSS, taking SACKs and IP options, 668 * and even PMTU discovery events into account. 669 * 670 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up 671 * cannot be large. However, taking into account rare use of URG, this 672 * is not a big flaw. 673 */ 674 675 unsigned int tcp_current_mss(struct sock *sk, int large) 676 { 677 struct tcp_sock *tp = tcp_sk(sk); 678 struct dst_entry *dst = __sk_dst_get(sk); 679 unsigned int do_large, mss_now; 680 681 mss_now = tp->mss_cache_std; 682 if (dst) { 683 u32 mtu = dst_mtu(dst); 684 if (mtu != tp->pmtu_cookie) 685 mss_now = tcp_sync_mss(sk, mtu); 686 } 687 688 do_large = (large && 689 (sk->sk_route_caps & NETIF_F_TSO) && 690 !tp->urg_mode); 691 692 if (do_large) { 693 unsigned int large_mss, factor, limit; 694 695 large_mss = 65535 - tp->af_specific->net_header_len - 696 tp->ext_header_len - tp->tcp_header_len; 697 698 if (tp->max_window && large_mss > (tp->max_window>>1)) 699 large_mss = max((tp->max_window>>1), 700 68U - tp->tcp_header_len); 701 702 factor = large_mss / mss_now; 703 704 /* Always keep large mss multiple of real mss, but 705 * do not exceed 1/tso_win_divisor of the congestion window 706 * so we can keep the ACK clock ticking and minimize 707 * bursting. 708 */ 709 limit = tp->snd_cwnd; 710 if (sysctl_tcp_tso_win_divisor) 711 limit /= sysctl_tcp_tso_win_divisor; 712 limit = max(1U, limit); 713 if (factor > limit) 714 factor = limit; 715 716 tp->mss_cache = mss_now * factor; 717 718 mss_now = tp->mss_cache; 719 } 720 721 if (tp->rx_opt.eff_sacks) 722 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED + 723 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK)); 724 return mss_now; 725 } 726 727 /* This routine writes packets to the network. It advances the 728 * send_head. This happens as incoming acks open up the remote 729 * window for us. 730 * 731 * Returns 1, if no segments are in flight and we have queued segments, but 732 * cannot send anything now because of SWS or another problem. 733 */ 734 int tcp_write_xmit(struct sock *sk, int nonagle) 735 { 736 struct tcp_sock *tp = tcp_sk(sk); 737 unsigned int mss_now; 738 739 /* If we are closed, the bytes will have to remain here. 740 * In time closedown will finish, we empty the write queue and all 741 * will be happy. 742 */ 743 if (sk->sk_state != TCP_CLOSE) { 744 struct sk_buff *skb; 745 int sent_pkts = 0; 746 747 /* Account for SACKS, we may need to fragment due to this. 748 * It is just like the real MSS changing on us midstream. 749 * We also handle things correctly when the user adds some 750 * IP options mid-stream. Silly to do, but cover it. 751 */ 752 mss_now = tcp_current_mss(sk, 1); 753 754 while ((skb = sk->sk_send_head) && 755 tcp_snd_test(tp, skb, mss_now, 756 tcp_skb_is_last(sk, skb) ? nonagle : 757 TCP_NAGLE_PUSH)) { 758 if (skb->len > mss_now) { 759 if (tcp_fragment(sk, skb, mss_now)) 760 break; 761 } 762 763 TCP_SKB_CB(skb)->when = tcp_time_stamp; 764 tcp_tso_set_push(skb); 765 if (tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC))) 766 break; 767 768 /* Advance the send_head. This one is sent out. 769 * This call will increment packets_out. 770 */ 771 update_send_head(sk, tp, skb); 772 773 tcp_minshall_update(tp, mss_now, skb); 774 sent_pkts = 1; 775 } 776 777 if (sent_pkts) { 778 tcp_cwnd_validate(sk, tp); 779 return 0; 780 } 781 782 return !tp->packets_out && sk->sk_send_head; 783 } 784 return 0; 785 } 786 787 /* This function returns the amount that we can raise the 788 * usable window based on the following constraints 789 * 790 * 1. The window can never be shrunk once it is offered (RFC 793) 791 * 2. We limit memory per socket 792 * 793 * RFC 1122: 794 * "the suggested [SWS] avoidance algorithm for the receiver is to keep 795 * RECV.NEXT + RCV.WIN fixed until: 796 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 797 * 798 * i.e. don't raise the right edge of the window until you can raise 799 * it at least MSS bytes. 800 * 801 * Unfortunately, the recommended algorithm breaks header prediction, 802 * since header prediction assumes th->window stays fixed. 803 * 804 * Strictly speaking, keeping th->window fixed violates the receiver 805 * side SWS prevention criteria. The problem is that under this rule 806 * a stream of single byte packets will cause the right side of the 807 * window to always advance by a single byte. 808 * 809 * Of course, if the sender implements sender side SWS prevention 810 * then this will not be a problem. 811 * 812 * BSD seems to make the following compromise: 813 * 814 * If the free space is less than the 1/4 of the maximum 815 * space available and the free space is less than 1/2 mss, 816 * then set the window to 0. 817 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 818 * Otherwise, just prevent the window from shrinking 819 * and from being larger than the largest representable value. 820 * 821 * This prevents incremental opening of the window in the regime 822 * where TCP is limited by the speed of the reader side taking 823 * data out of the TCP receive queue. It does nothing about 824 * those cases where the window is constrained on the sender side 825 * because the pipeline is full. 826 * 827 * BSD also seems to "accidentally" limit itself to windows that are a 828 * multiple of MSS, at least until the free space gets quite small. 829 * This would appear to be a side effect of the mbuf implementation. 830 * Combining these two algorithms results in the observed behavior 831 * of having a fixed window size at almost all times. 832 * 833 * Below we obtain similar behavior by forcing the offered window to 834 * a multiple of the mss when it is feasible to do so. 835 * 836 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 837 * Regular options like TIMESTAMP are taken into account. 838 */ 839 u32 __tcp_select_window(struct sock *sk) 840 { 841 struct tcp_sock *tp = tcp_sk(sk); 842 /* MSS for the peer's data. Previous verions used mss_clamp 843 * here. I don't know if the value based on our guesses 844 * of peer's MSS is better for the performance. It's more correct 845 * but may be worse for the performance because of rcv_mss 846 * fluctuations. --SAW 1998/11/1 847 */ 848 int mss = tp->ack.rcv_mss; 849 int free_space = tcp_space(sk); 850 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 851 int window; 852 853 if (mss > full_space) 854 mss = full_space; 855 856 if (free_space < full_space/2) { 857 tp->ack.quick = 0; 858 859 if (tcp_memory_pressure) 860 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss); 861 862 if (free_space < mss) 863 return 0; 864 } 865 866 if (free_space > tp->rcv_ssthresh) 867 free_space = tp->rcv_ssthresh; 868 869 /* Don't do rounding if we are using window scaling, since the 870 * scaled window will not line up with the MSS boundary anyway. 871 */ 872 window = tp->rcv_wnd; 873 if (tp->rx_opt.rcv_wscale) { 874 window = free_space; 875 876 /* Advertise enough space so that it won't get scaled away. 877 * Import case: prevent zero window announcement if 878 * 1<<rcv_wscale > mss. 879 */ 880 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 881 window = (((window >> tp->rx_opt.rcv_wscale) + 1) 882 << tp->rx_opt.rcv_wscale); 883 } else { 884 /* Get the largest window that is a nice multiple of mss. 885 * Window clamp already applied above. 886 * If our current window offering is within 1 mss of the 887 * free space we just keep it. This prevents the divide 888 * and multiply from happening most of the time. 889 * We also don't do any window rounding when the free space 890 * is too small. 891 */ 892 if (window <= free_space - mss || window > free_space) 893 window = (free_space/mss)*mss; 894 } 895 896 return window; 897 } 898 899 /* Attempt to collapse two adjacent SKB's during retransmission. */ 900 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now) 901 { 902 struct tcp_sock *tp = tcp_sk(sk); 903 struct sk_buff *next_skb = skb->next; 904 905 /* The first test we must make is that neither of these two 906 * SKB's are still referenced by someone else. 907 */ 908 if (!skb_cloned(skb) && !skb_cloned(next_skb)) { 909 int skb_size = skb->len, next_skb_size = next_skb->len; 910 u16 flags = TCP_SKB_CB(skb)->flags; 911 912 /* Also punt if next skb has been SACK'd. */ 913 if(TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED) 914 return; 915 916 /* Next skb is out of window. */ 917 if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd)) 918 return; 919 920 /* Punt if not enough space exists in the first SKB for 921 * the data in the second, or the total combined payload 922 * would exceed the MSS. 923 */ 924 if ((next_skb_size > skb_tailroom(skb)) || 925 ((skb_size + next_skb_size) > mss_now)) 926 return; 927 928 BUG_ON(tcp_skb_pcount(skb) != 1 || 929 tcp_skb_pcount(next_skb) != 1); 930 931 /* Ok. We will be able to collapse the packet. */ 932 __skb_unlink(next_skb, next_skb->list); 933 934 memcpy(skb_put(skb, next_skb_size), next_skb->data, next_skb_size); 935 936 if (next_skb->ip_summed == CHECKSUM_HW) 937 skb->ip_summed = CHECKSUM_HW; 938 939 if (skb->ip_summed != CHECKSUM_HW) 940 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 941 942 /* Update sequence range on original skb. */ 943 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 944 945 /* Merge over control information. */ 946 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */ 947 TCP_SKB_CB(skb)->flags = flags; 948 949 /* All done, get rid of second SKB and account for it so 950 * packet counting does not break. 951 */ 952 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL); 953 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS) 954 tp->retrans_out -= tcp_skb_pcount(next_skb); 955 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) { 956 tp->lost_out -= tcp_skb_pcount(next_skb); 957 tp->left_out -= tcp_skb_pcount(next_skb); 958 } 959 /* Reno case is special. Sigh... */ 960 if (!tp->rx_opt.sack_ok && tp->sacked_out) { 961 tcp_dec_pcount_approx(&tp->sacked_out, next_skb); 962 tp->left_out -= tcp_skb_pcount(next_skb); 963 } 964 965 /* Not quite right: it can be > snd.fack, but 966 * it is better to underestimate fackets. 967 */ 968 tcp_dec_pcount_approx(&tp->fackets_out, next_skb); 969 tcp_packets_out_dec(tp, next_skb); 970 sk_stream_free_skb(sk, next_skb); 971 } 972 } 973 974 /* Do a simple retransmit without using the backoff mechanisms in 975 * tcp_timer. This is used for path mtu discovery. 976 * The socket is already locked here. 977 */ 978 void tcp_simple_retransmit(struct sock *sk) 979 { 980 struct tcp_sock *tp = tcp_sk(sk); 981 struct sk_buff *skb; 982 unsigned int mss = tcp_current_mss(sk, 0); 983 int lost = 0; 984 985 sk_stream_for_retrans_queue(skb, sk) { 986 if (skb->len > mss && 987 !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) { 988 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) { 989 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 990 tp->retrans_out -= tcp_skb_pcount(skb); 991 } 992 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) { 993 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 994 tp->lost_out += tcp_skb_pcount(skb); 995 lost = 1; 996 } 997 } 998 } 999 1000 if (!lost) 1001 return; 1002 1003 tcp_sync_left_out(tp); 1004 1005 /* Don't muck with the congestion window here. 1006 * Reason is that we do not increase amount of _data_ 1007 * in network, but units changed and effective 1008 * cwnd/ssthresh really reduced now. 1009 */ 1010 if (tp->ca_state != TCP_CA_Loss) { 1011 tp->high_seq = tp->snd_nxt; 1012 tp->snd_ssthresh = tcp_current_ssthresh(tp); 1013 tp->prior_ssthresh = 0; 1014 tp->undo_marker = 0; 1015 tcp_set_ca_state(tp, TCP_CA_Loss); 1016 } 1017 tcp_xmit_retransmit_queue(sk); 1018 } 1019 1020 /* This retransmits one SKB. Policy decisions and retransmit queue 1021 * state updates are done by the caller. Returns non-zero if an 1022 * error occurred which prevented the send. 1023 */ 1024 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 1025 { 1026 struct tcp_sock *tp = tcp_sk(sk); 1027 unsigned int cur_mss = tcp_current_mss(sk, 0); 1028 int err; 1029 1030 /* Do not sent more than we queued. 1/4 is reserved for possible 1031 * copying overhead: frgagmentation, tunneling, mangling etc. 1032 */ 1033 if (atomic_read(&sk->sk_wmem_alloc) > 1034 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 1035 return -EAGAIN; 1036 1037 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 1038 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1039 BUG(); 1040 1041 if (sk->sk_route_caps & NETIF_F_TSO) { 1042 sk->sk_route_caps &= ~NETIF_F_TSO; 1043 sock_set_flag(sk, SOCK_NO_LARGESEND); 1044 tp->mss_cache = tp->mss_cache_std; 1045 } 1046 1047 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 1048 return -ENOMEM; 1049 } 1050 1051 /* If receiver has shrunk his window, and skb is out of 1052 * new window, do not retransmit it. The exception is the 1053 * case, when window is shrunk to zero. In this case 1054 * our retransmit serves as a zero window probe. 1055 */ 1056 if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd) 1057 && TCP_SKB_CB(skb)->seq != tp->snd_una) 1058 return -EAGAIN; 1059 1060 if (skb->len > cur_mss) { 1061 int old_factor = tcp_skb_pcount(skb); 1062 int new_factor; 1063 1064 if (tcp_fragment(sk, skb, cur_mss)) 1065 return -ENOMEM; /* We'll try again later. */ 1066 1067 /* New SKB created, account for it. */ 1068 new_factor = tcp_skb_pcount(skb); 1069 tp->packets_out -= old_factor - new_factor; 1070 tp->packets_out += tcp_skb_pcount(skb->next); 1071 } 1072 1073 /* Collapse two adjacent packets if worthwhile and we can. */ 1074 if(!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) && 1075 (skb->len < (cur_mss >> 1)) && 1076 (skb->next != sk->sk_send_head) && 1077 (skb->next != (struct sk_buff *)&sk->sk_write_queue) && 1078 (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(skb->next)->nr_frags == 0) && 1079 (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(skb->next) == 1) && 1080 (sysctl_tcp_retrans_collapse != 0)) 1081 tcp_retrans_try_collapse(sk, skb, cur_mss); 1082 1083 if(tp->af_specific->rebuild_header(sk)) 1084 return -EHOSTUNREACH; /* Routing failure or similar. */ 1085 1086 /* Some Solaris stacks overoptimize and ignore the FIN on a 1087 * retransmit when old data is attached. So strip it off 1088 * since it is cheap to do so and saves bytes on the network. 1089 */ 1090 if(skb->len > 0 && 1091 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && 1092 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 1093 if (!pskb_trim(skb, 0)) { 1094 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1; 1095 skb_shinfo(skb)->tso_segs = 1; 1096 skb_shinfo(skb)->tso_size = 0; 1097 skb->ip_summed = CHECKSUM_NONE; 1098 skb->csum = 0; 1099 } 1100 } 1101 1102 /* Make a copy, if the first transmission SKB clone we made 1103 * is still in somebody's hands, else make a clone. 1104 */ 1105 TCP_SKB_CB(skb)->when = tcp_time_stamp; 1106 tcp_tso_set_push(skb); 1107 1108 err = tcp_transmit_skb(sk, (skb_cloned(skb) ? 1109 pskb_copy(skb, GFP_ATOMIC): 1110 skb_clone(skb, GFP_ATOMIC))); 1111 1112 if (err == 0) { 1113 /* Update global TCP statistics. */ 1114 TCP_INC_STATS(TCP_MIB_RETRANSSEGS); 1115 1116 tp->total_retrans++; 1117 1118 #if FASTRETRANS_DEBUG > 0 1119 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) { 1120 if (net_ratelimit()) 1121 printk(KERN_DEBUG "retrans_out leaked.\n"); 1122 } 1123 #endif 1124 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 1125 tp->retrans_out += tcp_skb_pcount(skb); 1126 1127 /* Save stamp of the first retransmit. */ 1128 if (!tp->retrans_stamp) 1129 tp->retrans_stamp = TCP_SKB_CB(skb)->when; 1130 1131 tp->undo_retrans++; 1132 1133 /* snd_nxt is stored to detect loss of retransmitted segment, 1134 * see tcp_input.c tcp_sacktag_write_queue(). 1135 */ 1136 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 1137 } 1138 return err; 1139 } 1140 1141 /* This gets called after a retransmit timeout, and the initially 1142 * retransmitted data is acknowledged. It tries to continue 1143 * resending the rest of the retransmit queue, until either 1144 * we've sent it all or the congestion window limit is reached. 1145 * If doing SACK, the first ACK which comes back for a timeout 1146 * based retransmit packet might feed us FACK information again. 1147 * If so, we use it to avoid unnecessarily retransmissions. 1148 */ 1149 void tcp_xmit_retransmit_queue(struct sock *sk) 1150 { 1151 struct tcp_sock *tp = tcp_sk(sk); 1152 struct sk_buff *skb; 1153 int packet_cnt = tp->lost_out; 1154 1155 /* First pass: retransmit lost packets. */ 1156 if (packet_cnt) { 1157 sk_stream_for_retrans_queue(skb, sk) { 1158 __u8 sacked = TCP_SKB_CB(skb)->sacked; 1159 1160 /* Assume this retransmit will generate 1161 * only one packet for congestion window 1162 * calculation purposes. This works because 1163 * tcp_retransmit_skb() will chop up the 1164 * packet to be MSS sized and all the 1165 * packet counting works out. 1166 */ 1167 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 1168 return; 1169 1170 if (sacked&TCPCB_LOST) { 1171 if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) { 1172 if (tcp_retransmit_skb(sk, skb)) 1173 return; 1174 if (tp->ca_state != TCP_CA_Loss) 1175 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS); 1176 else 1177 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS); 1178 1179 if (skb == 1180 skb_peek(&sk->sk_write_queue)) 1181 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto); 1182 } 1183 1184 packet_cnt -= tcp_skb_pcount(skb); 1185 if (packet_cnt <= 0) 1186 break; 1187 } 1188 } 1189 } 1190 1191 /* OK, demanded retransmission is finished. */ 1192 1193 /* Forward retransmissions are possible only during Recovery. */ 1194 if (tp->ca_state != TCP_CA_Recovery) 1195 return; 1196 1197 /* No forward retransmissions in Reno are possible. */ 1198 if (!tp->rx_opt.sack_ok) 1199 return; 1200 1201 /* Yeah, we have to make difficult choice between forward transmission 1202 * and retransmission... Both ways have their merits... 1203 * 1204 * For now we do not retransmit anything, while we have some new 1205 * segments to send. 1206 */ 1207 1208 if (tcp_may_send_now(sk, tp)) 1209 return; 1210 1211 packet_cnt = 0; 1212 1213 sk_stream_for_retrans_queue(skb, sk) { 1214 /* Similar to the retransmit loop above we 1215 * can pretend that the retransmitted SKB 1216 * we send out here will be composed of one 1217 * real MSS sized packet because tcp_retransmit_skb() 1218 * will fragment it if necessary. 1219 */ 1220 if (++packet_cnt > tp->fackets_out) 1221 break; 1222 1223 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 1224 break; 1225 1226 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) 1227 continue; 1228 1229 /* Ok, retransmit it. */ 1230 if (tcp_retransmit_skb(sk, skb)) 1231 break; 1232 1233 if (skb == skb_peek(&sk->sk_write_queue)) 1234 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto); 1235 1236 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS); 1237 } 1238 } 1239 1240 1241 /* Send a fin. The caller locks the socket for us. This cannot be 1242 * allowed to fail queueing a FIN frame under any circumstances. 1243 */ 1244 void tcp_send_fin(struct sock *sk) 1245 { 1246 struct tcp_sock *tp = tcp_sk(sk); 1247 struct sk_buff *skb = skb_peek_tail(&sk->sk_write_queue); 1248 int mss_now; 1249 1250 /* Optimization, tack on the FIN if we have a queue of 1251 * unsent frames. But be careful about outgoing SACKS 1252 * and IP options. 1253 */ 1254 mss_now = tcp_current_mss(sk, 1); 1255 1256 if (sk->sk_send_head != NULL) { 1257 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN; 1258 TCP_SKB_CB(skb)->end_seq++; 1259 tp->write_seq++; 1260 } else { 1261 /* Socket is locked, keep trying until memory is available. */ 1262 for (;;) { 1263 skb = alloc_skb(MAX_TCP_HEADER, GFP_KERNEL); 1264 if (skb) 1265 break; 1266 yield(); 1267 } 1268 1269 /* Reserve space for headers and prepare control bits. */ 1270 skb_reserve(skb, MAX_TCP_HEADER); 1271 skb->csum = 0; 1272 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN); 1273 TCP_SKB_CB(skb)->sacked = 0; 1274 skb_shinfo(skb)->tso_segs = 1; 1275 skb_shinfo(skb)->tso_size = 0; 1276 1277 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 1278 TCP_SKB_CB(skb)->seq = tp->write_seq; 1279 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1; 1280 tcp_queue_skb(sk, skb); 1281 } 1282 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_OFF); 1283 } 1284 1285 /* We get here when a process closes a file descriptor (either due to 1286 * an explicit close() or as a byproduct of exit()'ing) and there 1287 * was unread data in the receive queue. This behavior is recommended 1288 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM 1289 */ 1290 void tcp_send_active_reset(struct sock *sk, int priority) 1291 { 1292 struct tcp_sock *tp = tcp_sk(sk); 1293 struct sk_buff *skb; 1294 1295 /* NOTE: No TCP options attached and we never retransmit this. */ 1296 skb = alloc_skb(MAX_TCP_HEADER, priority); 1297 if (!skb) { 1298 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); 1299 return; 1300 } 1301 1302 /* Reserve space for headers and prepare control bits. */ 1303 skb_reserve(skb, MAX_TCP_HEADER); 1304 skb->csum = 0; 1305 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST); 1306 TCP_SKB_CB(skb)->sacked = 0; 1307 skb_shinfo(skb)->tso_segs = 1; 1308 skb_shinfo(skb)->tso_size = 0; 1309 1310 /* Send it off. */ 1311 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk, tp); 1312 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq; 1313 TCP_SKB_CB(skb)->when = tcp_time_stamp; 1314 if (tcp_transmit_skb(sk, skb)) 1315 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); 1316 } 1317 1318 /* WARNING: This routine must only be called when we have already sent 1319 * a SYN packet that crossed the incoming SYN that caused this routine 1320 * to get called. If this assumption fails then the initial rcv_wnd 1321 * and rcv_wscale values will not be correct. 1322 */ 1323 int tcp_send_synack(struct sock *sk) 1324 { 1325 struct sk_buff* skb; 1326 1327 skb = skb_peek(&sk->sk_write_queue); 1328 if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) { 1329 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); 1330 return -EFAULT; 1331 } 1332 if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) { 1333 if (skb_cloned(skb)) { 1334 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 1335 if (nskb == NULL) 1336 return -ENOMEM; 1337 __skb_unlink(skb, &sk->sk_write_queue); 1338 skb_header_release(nskb); 1339 __skb_queue_head(&sk->sk_write_queue, nskb); 1340 sk_stream_free_skb(sk, skb); 1341 sk_charge_skb(sk, nskb); 1342 skb = nskb; 1343 } 1344 1345 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK; 1346 TCP_ECN_send_synack(tcp_sk(sk), skb); 1347 } 1348 TCP_SKB_CB(skb)->when = tcp_time_stamp; 1349 return tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC)); 1350 } 1351 1352 /* 1353 * Prepare a SYN-ACK. 1354 */ 1355 struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst, 1356 struct open_request *req) 1357 { 1358 struct tcp_sock *tp = tcp_sk(sk); 1359 struct tcphdr *th; 1360 int tcp_header_size; 1361 struct sk_buff *skb; 1362 1363 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); 1364 if (skb == NULL) 1365 return NULL; 1366 1367 /* Reserve space for headers. */ 1368 skb_reserve(skb, MAX_TCP_HEADER); 1369 1370 skb->dst = dst_clone(dst); 1371 1372 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS + 1373 (req->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) + 1374 (req->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) + 1375 /* SACK_PERM is in the place of NOP NOP of TS */ 1376 ((req->sack_ok && !req->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0)); 1377 skb->h.th = th = (struct tcphdr *) skb_push(skb, tcp_header_size); 1378 1379 memset(th, 0, sizeof(struct tcphdr)); 1380 th->syn = 1; 1381 th->ack = 1; 1382 if (dst->dev->features&NETIF_F_TSO) 1383 req->ecn_ok = 0; 1384 TCP_ECN_make_synack(req, th); 1385 th->source = inet_sk(sk)->sport; 1386 th->dest = req->rmt_port; 1387 TCP_SKB_CB(skb)->seq = req->snt_isn; 1388 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1; 1389 TCP_SKB_CB(skb)->sacked = 0; 1390 skb_shinfo(skb)->tso_segs = 1; 1391 skb_shinfo(skb)->tso_size = 0; 1392 th->seq = htonl(TCP_SKB_CB(skb)->seq); 1393 th->ack_seq = htonl(req->rcv_isn + 1); 1394 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 1395 __u8 rcv_wscale; 1396 /* Set this up on the first call only */ 1397 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 1398 /* tcp_full_space because it is guaranteed to be the first packet */ 1399 tcp_select_initial_window(tcp_full_space(sk), 1400 dst_metric(dst, RTAX_ADVMSS) - (req->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 1401 &req->rcv_wnd, 1402 &req->window_clamp, 1403 req->wscale_ok, 1404 &rcv_wscale); 1405 req->rcv_wscale = rcv_wscale; 1406 } 1407 1408 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 1409 th->window = htons(req->rcv_wnd); 1410 1411 TCP_SKB_CB(skb)->when = tcp_time_stamp; 1412 tcp_syn_build_options((__u32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), req->tstamp_ok, 1413 req->sack_ok, req->wscale_ok, req->rcv_wscale, 1414 TCP_SKB_CB(skb)->when, 1415 req->ts_recent); 1416 1417 skb->csum = 0; 1418 th->doff = (tcp_header_size >> 2); 1419 TCP_INC_STATS(TCP_MIB_OUTSEGS); 1420 return skb; 1421 } 1422 1423 /* 1424 * Do all connect socket setups that can be done AF independent. 1425 */ 1426 static inline void tcp_connect_init(struct sock *sk) 1427 { 1428 struct dst_entry *dst = __sk_dst_get(sk); 1429 struct tcp_sock *tp = tcp_sk(sk); 1430 __u8 rcv_wscale; 1431 1432 /* We'll fix this up when we get a response from the other end. 1433 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 1434 */ 1435 tp->tcp_header_len = sizeof(struct tcphdr) + 1436 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 1437 1438 /* If user gave his TCP_MAXSEG, record it to clamp */ 1439 if (tp->rx_opt.user_mss) 1440 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 1441 tp->max_window = 0; 1442 tcp_sync_mss(sk, dst_mtu(dst)); 1443 1444 if (!tp->window_clamp) 1445 tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 1446 tp->advmss = dst_metric(dst, RTAX_ADVMSS); 1447 tcp_initialize_rcv_mss(sk); 1448 tcp_ca_init(tp); 1449 1450 tcp_select_initial_window(tcp_full_space(sk), 1451 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 1452 &tp->rcv_wnd, 1453 &tp->window_clamp, 1454 sysctl_tcp_window_scaling, 1455 &rcv_wscale); 1456 1457 tp->rx_opt.rcv_wscale = rcv_wscale; 1458 tp->rcv_ssthresh = tp->rcv_wnd; 1459 1460 sk->sk_err = 0; 1461 sock_reset_flag(sk, SOCK_DONE); 1462 tp->snd_wnd = 0; 1463 tcp_init_wl(tp, tp->write_seq, 0); 1464 tp->snd_una = tp->write_seq; 1465 tp->snd_sml = tp->write_seq; 1466 tp->rcv_nxt = 0; 1467 tp->rcv_wup = 0; 1468 tp->copied_seq = 0; 1469 1470 tp->rto = TCP_TIMEOUT_INIT; 1471 tp->retransmits = 0; 1472 tcp_clear_retrans(tp); 1473 } 1474 1475 /* 1476 * Build a SYN and send it off. 1477 */ 1478 int tcp_connect(struct sock *sk) 1479 { 1480 struct tcp_sock *tp = tcp_sk(sk); 1481 struct sk_buff *buff; 1482 1483 tcp_connect_init(sk); 1484 1485 buff = alloc_skb(MAX_TCP_HEADER + 15, sk->sk_allocation); 1486 if (unlikely(buff == NULL)) 1487 return -ENOBUFS; 1488 1489 /* Reserve space for headers. */ 1490 skb_reserve(buff, MAX_TCP_HEADER); 1491 1492 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN; 1493 TCP_ECN_send_syn(sk, tp, buff); 1494 TCP_SKB_CB(buff)->sacked = 0; 1495 skb_shinfo(buff)->tso_segs = 1; 1496 skb_shinfo(buff)->tso_size = 0; 1497 buff->csum = 0; 1498 TCP_SKB_CB(buff)->seq = tp->write_seq++; 1499 TCP_SKB_CB(buff)->end_seq = tp->write_seq; 1500 tp->snd_nxt = tp->write_seq; 1501 tp->pushed_seq = tp->write_seq; 1502 tcp_ca_init(tp); 1503 1504 /* Send it off. */ 1505 TCP_SKB_CB(buff)->when = tcp_time_stamp; 1506 tp->retrans_stamp = TCP_SKB_CB(buff)->when; 1507 skb_header_release(buff); 1508 __skb_queue_tail(&sk->sk_write_queue, buff); 1509 sk_charge_skb(sk, buff); 1510 tp->packets_out += tcp_skb_pcount(buff); 1511 tcp_transmit_skb(sk, skb_clone(buff, GFP_KERNEL)); 1512 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS); 1513 1514 /* Timer for repeating the SYN until an answer. */ 1515 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto); 1516 return 0; 1517 } 1518 1519 /* Send out a delayed ack, the caller does the policy checking 1520 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 1521 * for details. 1522 */ 1523 void tcp_send_delayed_ack(struct sock *sk) 1524 { 1525 struct tcp_sock *tp = tcp_sk(sk); 1526 int ato = tp->ack.ato; 1527 unsigned long timeout; 1528 1529 if (ato > TCP_DELACK_MIN) { 1530 int max_ato = HZ/2; 1531 1532 if (tp->ack.pingpong || (tp->ack.pending&TCP_ACK_PUSHED)) 1533 max_ato = TCP_DELACK_MAX; 1534 1535 /* Slow path, intersegment interval is "high". */ 1536 1537 /* If some rtt estimate is known, use it to bound delayed ack. 1538 * Do not use tp->rto here, use results of rtt measurements 1539 * directly. 1540 */ 1541 if (tp->srtt) { 1542 int rtt = max(tp->srtt>>3, TCP_DELACK_MIN); 1543 1544 if (rtt < max_ato) 1545 max_ato = rtt; 1546 } 1547 1548 ato = min(ato, max_ato); 1549 } 1550 1551 /* Stay within the limit we were given */ 1552 timeout = jiffies + ato; 1553 1554 /* Use new timeout only if there wasn't a older one earlier. */ 1555 if (tp->ack.pending&TCP_ACK_TIMER) { 1556 /* If delack timer was blocked or is about to expire, 1557 * send ACK now. 1558 */ 1559 if (tp->ack.blocked || time_before_eq(tp->ack.timeout, jiffies+(ato>>2))) { 1560 tcp_send_ack(sk); 1561 return; 1562 } 1563 1564 if (!time_before(timeout, tp->ack.timeout)) 1565 timeout = tp->ack.timeout; 1566 } 1567 tp->ack.pending |= TCP_ACK_SCHED|TCP_ACK_TIMER; 1568 tp->ack.timeout = timeout; 1569 sk_reset_timer(sk, &tp->delack_timer, timeout); 1570 } 1571 1572 /* This routine sends an ack and also updates the window. */ 1573 void tcp_send_ack(struct sock *sk) 1574 { 1575 /* If we have been reset, we may not send again. */ 1576 if (sk->sk_state != TCP_CLOSE) { 1577 struct tcp_sock *tp = tcp_sk(sk); 1578 struct sk_buff *buff; 1579 1580 /* We are not putting this on the write queue, so 1581 * tcp_transmit_skb() will set the ownership to this 1582 * sock. 1583 */ 1584 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 1585 if (buff == NULL) { 1586 tcp_schedule_ack(tp); 1587 tp->ack.ato = TCP_ATO_MIN; 1588 tcp_reset_xmit_timer(sk, TCP_TIME_DACK, TCP_DELACK_MAX); 1589 return; 1590 } 1591 1592 /* Reserve space for headers and prepare control bits. */ 1593 skb_reserve(buff, MAX_TCP_HEADER); 1594 buff->csum = 0; 1595 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK; 1596 TCP_SKB_CB(buff)->sacked = 0; 1597 skb_shinfo(buff)->tso_segs = 1; 1598 skb_shinfo(buff)->tso_size = 0; 1599 1600 /* Send it off, this clears delayed acks for us. */ 1601 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk, tp); 1602 TCP_SKB_CB(buff)->when = tcp_time_stamp; 1603 tcp_transmit_skb(sk, buff); 1604 } 1605 } 1606 1607 /* This routine sends a packet with an out of date sequence 1608 * number. It assumes the other end will try to ack it. 1609 * 1610 * Question: what should we make while urgent mode? 1611 * 4.4BSD forces sending single byte of data. We cannot send 1612 * out of window data, because we have SND.NXT==SND.MAX... 1613 * 1614 * Current solution: to send TWO zero-length segments in urgent mode: 1615 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 1616 * out-of-date with SND.UNA-1 to probe window. 1617 */ 1618 static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 1619 { 1620 struct tcp_sock *tp = tcp_sk(sk); 1621 struct sk_buff *skb; 1622 1623 /* We don't queue it, tcp_transmit_skb() sets ownership. */ 1624 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 1625 if (skb == NULL) 1626 return -1; 1627 1628 /* Reserve space for headers and set control bits. */ 1629 skb_reserve(skb, MAX_TCP_HEADER); 1630 skb->csum = 0; 1631 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK; 1632 TCP_SKB_CB(skb)->sacked = urgent; 1633 skb_shinfo(skb)->tso_segs = 1; 1634 skb_shinfo(skb)->tso_size = 0; 1635 1636 /* Use a previous sequence. This should cause the other 1637 * end to send an ack. Don't queue or clone SKB, just 1638 * send it. 1639 */ 1640 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1; 1641 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq; 1642 TCP_SKB_CB(skb)->when = tcp_time_stamp; 1643 return tcp_transmit_skb(sk, skb); 1644 } 1645 1646 int tcp_write_wakeup(struct sock *sk) 1647 { 1648 if (sk->sk_state != TCP_CLOSE) { 1649 struct tcp_sock *tp = tcp_sk(sk); 1650 struct sk_buff *skb; 1651 1652 if ((skb = sk->sk_send_head) != NULL && 1653 before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) { 1654 int err; 1655 unsigned int mss = tcp_current_mss(sk, 0); 1656 unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq; 1657 1658 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 1659 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 1660 1661 /* We are probing the opening of a window 1662 * but the window size is != 0 1663 * must have been a result SWS avoidance ( sender ) 1664 */ 1665 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 1666 skb->len > mss) { 1667 seg_size = min(seg_size, mss); 1668 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 1669 if (tcp_fragment(sk, skb, seg_size)) 1670 return -1; 1671 /* SWS override triggered forced fragmentation. 1672 * Disable TSO, the connection is too sick. */ 1673 if (sk->sk_route_caps & NETIF_F_TSO) { 1674 sock_set_flag(sk, SOCK_NO_LARGESEND); 1675 sk->sk_route_caps &= ~NETIF_F_TSO; 1676 tp->mss_cache = tp->mss_cache_std; 1677 } 1678 } else if (!tcp_skb_pcount(skb)) 1679 tcp_set_skb_tso_segs(skb, tp->mss_cache_std); 1680 1681 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 1682 TCP_SKB_CB(skb)->when = tcp_time_stamp; 1683 tcp_tso_set_push(skb); 1684 err = tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC)); 1685 if (!err) { 1686 update_send_head(sk, tp, skb); 1687 } 1688 return err; 1689 } else { 1690 if (tp->urg_mode && 1691 between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF)) 1692 tcp_xmit_probe_skb(sk, TCPCB_URG); 1693 return tcp_xmit_probe_skb(sk, 0); 1694 } 1695 } 1696 return -1; 1697 } 1698 1699 /* A window probe timeout has occurred. If window is not closed send 1700 * a partial packet else a zero probe. 1701 */ 1702 void tcp_send_probe0(struct sock *sk) 1703 { 1704 struct tcp_sock *tp = tcp_sk(sk); 1705 int err; 1706 1707 err = tcp_write_wakeup(sk); 1708 1709 if (tp->packets_out || !sk->sk_send_head) { 1710 /* Cancel probe timer, if it is not required. */ 1711 tp->probes_out = 0; 1712 tp->backoff = 0; 1713 return; 1714 } 1715 1716 if (err <= 0) { 1717 if (tp->backoff < sysctl_tcp_retries2) 1718 tp->backoff++; 1719 tp->probes_out++; 1720 tcp_reset_xmit_timer (sk, TCP_TIME_PROBE0, 1721 min(tp->rto << tp->backoff, TCP_RTO_MAX)); 1722 } else { 1723 /* If packet was not sent due to local congestion, 1724 * do not backoff and do not remember probes_out. 1725 * Let local senders to fight for local resources. 1726 * 1727 * Use accumulated backoff yet. 1728 */ 1729 if (!tp->probes_out) 1730 tp->probes_out=1; 1731 tcp_reset_xmit_timer (sk, TCP_TIME_PROBE0, 1732 min(tp->rto << tp->backoff, TCP_RESOURCE_PROBE_INTERVAL)); 1733 } 1734 } 1735 1736 EXPORT_SYMBOL(tcp_connect); 1737 EXPORT_SYMBOL(tcp_make_synack); 1738 EXPORT_SYMBOL(tcp_simple_retransmit); 1739 EXPORT_SYMBOL(tcp_sync_mss); 1740