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 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(sk, 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 sock *sk, struct sk_buff *skb) 444 { 445 struct tcp_sock *tp = tcp_sk(sk); 446 447 if (skb->len <= tp->mss_cache_std || 448 !(sk->sk_route_caps & NETIF_F_TSO)) { 449 /* Avoid the costly divide in the normal 450 * non-TSO case. 451 */ 452 skb_shinfo(skb)->tso_segs = 1; 453 skb_shinfo(skb)->tso_size = 0; 454 } else { 455 unsigned int factor; 456 457 factor = skb->len + (tp->mss_cache_std - 1); 458 factor /= tp->mss_cache_std; 459 skb_shinfo(skb)->tso_segs = factor; 460 skb_shinfo(skb)->tso_size = tp->mss_cache_std; 461 } 462 } 463 464 /* Function to create two new TCP segments. Shrinks the given segment 465 * to the specified size and appends a new segment with the rest of the 466 * packet to the list. This won't be called frequently, I hope. 467 * Remember, these are still headerless SKBs at this point. 468 */ 469 static int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len) 470 { 471 struct tcp_sock *tp = tcp_sk(sk); 472 struct sk_buff *buff; 473 int nsize; 474 u16 flags; 475 476 nsize = skb_headlen(skb) - len; 477 if (nsize < 0) 478 nsize = 0; 479 480 if (skb_cloned(skb) && 481 skb_is_nonlinear(skb) && 482 pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 483 return -ENOMEM; 484 485 /* Get a new skb... force flag on. */ 486 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 487 if (buff == NULL) 488 return -ENOMEM; /* We'll just try again later. */ 489 sk_charge_skb(sk, buff); 490 491 /* Correct the sequence numbers. */ 492 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 493 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 494 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 495 496 /* PSH and FIN should only be set in the second packet. */ 497 flags = TCP_SKB_CB(skb)->flags; 498 TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); 499 TCP_SKB_CB(buff)->flags = flags; 500 TCP_SKB_CB(buff)->sacked = 501 (TCP_SKB_CB(skb)->sacked & 502 (TCPCB_LOST | TCPCB_EVER_RETRANS | TCPCB_AT_TAIL)); 503 TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL; 504 505 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_HW) { 506 /* Copy and checksum data tail into the new buffer. */ 507 buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize), 508 nsize, 0); 509 510 skb_trim(skb, len); 511 512 skb->csum = csum_block_sub(skb->csum, buff->csum, len); 513 } else { 514 skb->ip_summed = CHECKSUM_HW; 515 skb_split(skb, buff, len); 516 } 517 518 buff->ip_summed = skb->ip_summed; 519 520 /* Looks stupid, but our code really uses when of 521 * skbs, which it never sent before. --ANK 522 */ 523 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 524 525 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) { 526 tp->lost_out -= tcp_skb_pcount(skb); 527 tp->left_out -= tcp_skb_pcount(skb); 528 } 529 530 /* Fix up tso_factor for both original and new SKB. */ 531 tcp_set_skb_tso_segs(sk, skb); 532 tcp_set_skb_tso_segs(sk, buff); 533 534 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) { 535 tp->lost_out += tcp_skb_pcount(skb); 536 tp->left_out += tcp_skb_pcount(skb); 537 } 538 539 if (TCP_SKB_CB(buff)->sacked&TCPCB_LOST) { 540 tp->lost_out += tcp_skb_pcount(buff); 541 tp->left_out += tcp_skb_pcount(buff); 542 } 543 544 /* Link BUFF into the send queue. */ 545 __skb_append(skb, buff); 546 547 return 0; 548 } 549 550 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 551 * eventually). The difference is that pulled data not copied, but 552 * immediately discarded. 553 */ 554 static unsigned char *__pskb_trim_head(struct sk_buff *skb, int len) 555 { 556 int i, k, eat; 557 558 eat = len; 559 k = 0; 560 for (i=0; i<skb_shinfo(skb)->nr_frags; i++) { 561 if (skb_shinfo(skb)->frags[i].size <= eat) { 562 put_page(skb_shinfo(skb)->frags[i].page); 563 eat -= skb_shinfo(skb)->frags[i].size; 564 } else { 565 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 566 if (eat) { 567 skb_shinfo(skb)->frags[k].page_offset += eat; 568 skb_shinfo(skb)->frags[k].size -= eat; 569 eat = 0; 570 } 571 k++; 572 } 573 } 574 skb_shinfo(skb)->nr_frags = k; 575 576 skb->tail = skb->data; 577 skb->data_len -= len; 578 skb->len = skb->data_len; 579 return skb->tail; 580 } 581 582 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 583 { 584 if (skb_cloned(skb) && 585 pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 586 return -ENOMEM; 587 588 if (len <= skb_headlen(skb)) { 589 __skb_pull(skb, len); 590 } else { 591 if (__pskb_trim_head(skb, len-skb_headlen(skb)) == NULL) 592 return -ENOMEM; 593 } 594 595 TCP_SKB_CB(skb)->seq += len; 596 skb->ip_summed = CHECKSUM_HW; 597 598 skb->truesize -= len; 599 sk->sk_wmem_queued -= len; 600 sk->sk_forward_alloc += len; 601 sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 602 603 /* Any change of skb->len requires recalculation of tso 604 * factor and mss. 605 */ 606 if (tcp_skb_pcount(skb) > 1) 607 tcp_set_skb_tso_segs(sk, skb); 608 609 return 0; 610 } 611 612 /* This function synchronize snd mss to current pmtu/exthdr set. 613 614 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 615 for TCP options, but includes only bare TCP header. 616 617 tp->rx_opt.mss_clamp is mss negotiated at connection setup. 618 It is minumum of user_mss and mss received with SYN. 619 It also does not include TCP options. 620 621 tp->pmtu_cookie is last pmtu, seen by this function. 622 623 tp->mss_cache is current effective sending mss, including 624 all tcp options except for SACKs. It is evaluated, 625 taking into account current pmtu, but never exceeds 626 tp->rx_opt.mss_clamp. 627 628 NOTE1. rfc1122 clearly states that advertised MSS 629 DOES NOT include either tcp or ip options. 630 631 NOTE2. tp->pmtu_cookie and tp->mss_cache are READ ONLY outside 632 this function. --ANK (980731) 633 */ 634 635 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 636 { 637 struct tcp_sock *tp = tcp_sk(sk); 638 int mss_now; 639 640 /* Calculate base mss without TCP options: 641 It is MMS_S - sizeof(tcphdr) of rfc1122 642 */ 643 mss_now = pmtu - tp->af_specific->net_header_len - sizeof(struct tcphdr); 644 645 /* Clamp it (mss_clamp does not include tcp options) */ 646 if (mss_now > tp->rx_opt.mss_clamp) 647 mss_now = tp->rx_opt.mss_clamp; 648 649 /* Now subtract optional transport overhead */ 650 mss_now -= tp->ext_header_len; 651 652 /* Then reserve room for full set of TCP options and 8 bytes of data */ 653 if (mss_now < 48) 654 mss_now = 48; 655 656 /* Now subtract TCP options size, not including SACKs */ 657 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); 658 659 /* Bound mss with half of window */ 660 if (tp->max_window && mss_now > (tp->max_window>>1)) 661 mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len); 662 663 /* And store cached results */ 664 tp->pmtu_cookie = pmtu; 665 tp->mss_cache = tp->mss_cache_std = mss_now; 666 667 return mss_now; 668 } 669 670 /* Compute the current effective MSS, taking SACKs and IP options, 671 * and even PMTU discovery events into account. 672 * 673 * LARGESEND note: !urg_mode is overkill, only frames up to snd_up 674 * cannot be large. However, taking into account rare use of URG, this 675 * is not a big flaw. 676 */ 677 678 unsigned int tcp_current_mss(struct sock *sk, int large) 679 { 680 struct tcp_sock *tp = tcp_sk(sk); 681 struct dst_entry *dst = __sk_dst_get(sk); 682 unsigned int do_large, mss_now; 683 684 mss_now = tp->mss_cache_std; 685 if (dst) { 686 u32 mtu = dst_mtu(dst); 687 if (mtu != tp->pmtu_cookie) 688 mss_now = tcp_sync_mss(sk, mtu); 689 } 690 691 do_large = (large && 692 (sk->sk_route_caps & NETIF_F_TSO) && 693 !tp->urg_mode); 694 695 if (do_large) { 696 unsigned int large_mss, factor, limit; 697 698 large_mss = 65535 - tp->af_specific->net_header_len - 699 tp->ext_header_len - tp->tcp_header_len; 700 701 if (tp->max_window && large_mss > (tp->max_window>>1)) 702 large_mss = max((tp->max_window>>1), 703 68U - tp->tcp_header_len); 704 705 factor = large_mss / mss_now; 706 707 /* Always keep large mss multiple of real mss, but 708 * do not exceed 1/tso_win_divisor of the congestion window 709 * so we can keep the ACK clock ticking and minimize 710 * bursting. 711 */ 712 limit = tp->snd_cwnd; 713 if (sysctl_tcp_tso_win_divisor) 714 limit /= sysctl_tcp_tso_win_divisor; 715 limit = max(1U, limit); 716 if (factor > limit) 717 factor = limit; 718 719 tp->mss_cache = mss_now * factor; 720 721 mss_now = tp->mss_cache; 722 } 723 724 if (tp->rx_opt.eff_sacks) 725 mss_now -= (TCPOLEN_SACK_BASE_ALIGNED + 726 (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK)); 727 return mss_now; 728 } 729 730 /* This routine writes packets to the network. It advances the 731 * send_head. This happens as incoming acks open up the remote 732 * window for us. 733 * 734 * Returns 1, if no segments are in flight and we have queued segments, but 735 * cannot send anything now because of SWS or another problem. 736 */ 737 int tcp_write_xmit(struct sock *sk, int nonagle) 738 { 739 struct tcp_sock *tp = tcp_sk(sk); 740 unsigned int mss_now; 741 742 /* If we are closed, the bytes will have to remain here. 743 * In time closedown will finish, we empty the write queue and all 744 * will be happy. 745 */ 746 if (sk->sk_state != TCP_CLOSE) { 747 struct sk_buff *skb; 748 int sent_pkts = 0; 749 750 /* Account for SACKS, we may need to fragment due to this. 751 * It is just like the real MSS changing on us midstream. 752 * We also handle things correctly when the user adds some 753 * IP options mid-stream. Silly to do, but cover it. 754 */ 755 mss_now = tcp_current_mss(sk, 1); 756 757 while ((skb = sk->sk_send_head) && 758 tcp_snd_test(sk, skb, mss_now, 759 tcp_skb_is_last(sk, skb) ? nonagle : 760 TCP_NAGLE_PUSH)) { 761 if (skb->len > mss_now) { 762 if (tcp_fragment(sk, skb, mss_now)) 763 break; 764 } 765 766 TCP_SKB_CB(skb)->when = tcp_time_stamp; 767 tcp_tso_set_push(skb); 768 if (tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC))) 769 break; 770 771 /* Advance the send_head. This one is sent out. 772 * This call will increment packets_out. 773 */ 774 update_send_head(sk, tp, skb); 775 776 tcp_minshall_update(tp, mss_now, skb); 777 sent_pkts = 1; 778 } 779 780 if (sent_pkts) { 781 tcp_cwnd_validate(sk, tp); 782 return 0; 783 } 784 785 return !tp->packets_out && sk->sk_send_head; 786 } 787 return 0; 788 } 789 790 /* This function returns the amount that we can raise the 791 * usable window based on the following constraints 792 * 793 * 1. The window can never be shrunk once it is offered (RFC 793) 794 * 2. We limit memory per socket 795 * 796 * RFC 1122: 797 * "the suggested [SWS] avoidance algorithm for the receiver is to keep 798 * RECV.NEXT + RCV.WIN fixed until: 799 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 800 * 801 * i.e. don't raise the right edge of the window until you can raise 802 * it at least MSS bytes. 803 * 804 * Unfortunately, the recommended algorithm breaks header prediction, 805 * since header prediction assumes th->window stays fixed. 806 * 807 * Strictly speaking, keeping th->window fixed violates the receiver 808 * side SWS prevention criteria. The problem is that under this rule 809 * a stream of single byte packets will cause the right side of the 810 * window to always advance by a single byte. 811 * 812 * Of course, if the sender implements sender side SWS prevention 813 * then this will not be a problem. 814 * 815 * BSD seems to make the following compromise: 816 * 817 * If the free space is less than the 1/4 of the maximum 818 * space available and the free space is less than 1/2 mss, 819 * then set the window to 0. 820 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 821 * Otherwise, just prevent the window from shrinking 822 * and from being larger than the largest representable value. 823 * 824 * This prevents incremental opening of the window in the regime 825 * where TCP is limited by the speed of the reader side taking 826 * data out of the TCP receive queue. It does nothing about 827 * those cases where the window is constrained on the sender side 828 * because the pipeline is full. 829 * 830 * BSD also seems to "accidentally" limit itself to windows that are a 831 * multiple of MSS, at least until the free space gets quite small. 832 * This would appear to be a side effect of the mbuf implementation. 833 * Combining these two algorithms results in the observed behavior 834 * of having a fixed window size at almost all times. 835 * 836 * Below we obtain similar behavior by forcing the offered window to 837 * a multiple of the mss when it is feasible to do so. 838 * 839 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 840 * Regular options like TIMESTAMP are taken into account. 841 */ 842 u32 __tcp_select_window(struct sock *sk) 843 { 844 struct tcp_sock *tp = tcp_sk(sk); 845 /* MSS for the peer's data. Previous verions used mss_clamp 846 * here. I don't know if the value based on our guesses 847 * of peer's MSS is better for the performance. It's more correct 848 * but may be worse for the performance because of rcv_mss 849 * fluctuations. --SAW 1998/11/1 850 */ 851 int mss = tp->ack.rcv_mss; 852 int free_space = tcp_space(sk); 853 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 854 int window; 855 856 if (mss > full_space) 857 mss = full_space; 858 859 if (free_space < full_space/2) { 860 tp->ack.quick = 0; 861 862 if (tcp_memory_pressure) 863 tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss); 864 865 if (free_space < mss) 866 return 0; 867 } 868 869 if (free_space > tp->rcv_ssthresh) 870 free_space = tp->rcv_ssthresh; 871 872 /* Don't do rounding if we are using window scaling, since the 873 * scaled window will not line up with the MSS boundary anyway. 874 */ 875 window = tp->rcv_wnd; 876 if (tp->rx_opt.rcv_wscale) { 877 window = free_space; 878 879 /* Advertise enough space so that it won't get scaled away. 880 * Import case: prevent zero window announcement if 881 * 1<<rcv_wscale > mss. 882 */ 883 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 884 window = (((window >> tp->rx_opt.rcv_wscale) + 1) 885 << tp->rx_opt.rcv_wscale); 886 } else { 887 /* Get the largest window that is a nice multiple of mss. 888 * Window clamp already applied above. 889 * If our current window offering is within 1 mss of the 890 * free space we just keep it. This prevents the divide 891 * and multiply from happening most of the time. 892 * We also don't do any window rounding when the free space 893 * is too small. 894 */ 895 if (window <= free_space - mss || window > free_space) 896 window = (free_space/mss)*mss; 897 } 898 899 return window; 900 } 901 902 /* Attempt to collapse two adjacent SKB's during retransmission. */ 903 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now) 904 { 905 struct tcp_sock *tp = tcp_sk(sk); 906 struct sk_buff *next_skb = skb->next; 907 908 /* The first test we must make is that neither of these two 909 * SKB's are still referenced by someone else. 910 */ 911 if (!skb_cloned(skb) && !skb_cloned(next_skb)) { 912 int skb_size = skb->len, next_skb_size = next_skb->len; 913 u16 flags = TCP_SKB_CB(skb)->flags; 914 915 /* Also punt if next skb has been SACK'd. */ 916 if(TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED) 917 return; 918 919 /* Next skb is out of window. */ 920 if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd)) 921 return; 922 923 /* Punt if not enough space exists in the first SKB for 924 * the data in the second, or the total combined payload 925 * would exceed the MSS. 926 */ 927 if ((next_skb_size > skb_tailroom(skb)) || 928 ((skb_size + next_skb_size) > mss_now)) 929 return; 930 931 BUG_ON(tcp_skb_pcount(skb) != 1 || 932 tcp_skb_pcount(next_skb) != 1); 933 934 /* Ok. We will be able to collapse the packet. */ 935 __skb_unlink(next_skb, next_skb->list); 936 937 memcpy(skb_put(skb, next_skb_size), next_skb->data, next_skb_size); 938 939 if (next_skb->ip_summed == CHECKSUM_HW) 940 skb->ip_summed = CHECKSUM_HW; 941 942 if (skb->ip_summed != CHECKSUM_HW) 943 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 944 945 /* Update sequence range on original skb. */ 946 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 947 948 /* Merge over control information. */ 949 flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */ 950 TCP_SKB_CB(skb)->flags = flags; 951 952 /* All done, get rid of second SKB and account for it so 953 * packet counting does not break. 954 */ 955 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL); 956 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS) 957 tp->retrans_out -= tcp_skb_pcount(next_skb); 958 if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) { 959 tp->lost_out -= tcp_skb_pcount(next_skb); 960 tp->left_out -= tcp_skb_pcount(next_skb); 961 } 962 /* Reno case is special. Sigh... */ 963 if (!tp->rx_opt.sack_ok && tp->sacked_out) { 964 tcp_dec_pcount_approx(&tp->sacked_out, next_skb); 965 tp->left_out -= tcp_skb_pcount(next_skb); 966 } 967 968 /* Not quite right: it can be > snd.fack, but 969 * it is better to underestimate fackets. 970 */ 971 tcp_dec_pcount_approx(&tp->fackets_out, next_skb); 972 tcp_packets_out_dec(tp, next_skb); 973 sk_stream_free_skb(sk, next_skb); 974 } 975 } 976 977 /* Do a simple retransmit without using the backoff mechanisms in 978 * tcp_timer. This is used for path mtu discovery. 979 * The socket is already locked here. 980 */ 981 void tcp_simple_retransmit(struct sock *sk) 982 { 983 struct tcp_sock *tp = tcp_sk(sk); 984 struct sk_buff *skb; 985 unsigned int mss = tcp_current_mss(sk, 0); 986 int lost = 0; 987 988 sk_stream_for_retrans_queue(skb, sk) { 989 if (skb->len > mss && 990 !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) { 991 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) { 992 TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 993 tp->retrans_out -= tcp_skb_pcount(skb); 994 } 995 if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) { 996 TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 997 tp->lost_out += tcp_skb_pcount(skb); 998 lost = 1; 999 } 1000 } 1001 } 1002 1003 if (!lost) 1004 return; 1005 1006 tcp_sync_left_out(tp); 1007 1008 /* Don't muck with the congestion window here. 1009 * Reason is that we do not increase amount of _data_ 1010 * in network, but units changed and effective 1011 * cwnd/ssthresh really reduced now. 1012 */ 1013 if (tp->ca_state != TCP_CA_Loss) { 1014 tp->high_seq = tp->snd_nxt; 1015 tp->snd_ssthresh = tcp_current_ssthresh(tp); 1016 tp->prior_ssthresh = 0; 1017 tp->undo_marker = 0; 1018 tcp_set_ca_state(tp, TCP_CA_Loss); 1019 } 1020 tcp_xmit_retransmit_queue(sk); 1021 } 1022 1023 /* This retransmits one SKB. Policy decisions and retransmit queue 1024 * state updates are done by the caller. Returns non-zero if an 1025 * error occurred which prevented the send. 1026 */ 1027 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 1028 { 1029 struct tcp_sock *tp = tcp_sk(sk); 1030 unsigned int cur_mss = tcp_current_mss(sk, 0); 1031 int err; 1032 1033 /* Do not sent more than we queued. 1/4 is reserved for possible 1034 * copying overhead: frgagmentation, tunneling, mangling etc. 1035 */ 1036 if (atomic_read(&sk->sk_wmem_alloc) > 1037 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 1038 return -EAGAIN; 1039 1040 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 1041 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 1042 BUG(); 1043 1044 if (sk->sk_route_caps & NETIF_F_TSO) { 1045 sk->sk_route_caps &= ~NETIF_F_TSO; 1046 sock_set_flag(sk, SOCK_NO_LARGESEND); 1047 tp->mss_cache = tp->mss_cache_std; 1048 } 1049 1050 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 1051 return -ENOMEM; 1052 } 1053 1054 /* If receiver has shrunk his window, and skb is out of 1055 * new window, do not retransmit it. The exception is the 1056 * case, when window is shrunk to zero. In this case 1057 * our retransmit serves as a zero window probe. 1058 */ 1059 if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd) 1060 && TCP_SKB_CB(skb)->seq != tp->snd_una) 1061 return -EAGAIN; 1062 1063 if (skb->len > cur_mss) { 1064 int old_factor = tcp_skb_pcount(skb); 1065 int new_factor; 1066 1067 if (tcp_fragment(sk, skb, cur_mss)) 1068 return -ENOMEM; /* We'll try again later. */ 1069 1070 /* New SKB created, account for it. */ 1071 new_factor = tcp_skb_pcount(skb); 1072 tp->packets_out -= old_factor - new_factor; 1073 tp->packets_out += tcp_skb_pcount(skb->next); 1074 } 1075 1076 /* Collapse two adjacent packets if worthwhile and we can. */ 1077 if(!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) && 1078 (skb->len < (cur_mss >> 1)) && 1079 (skb->next != sk->sk_send_head) && 1080 (skb->next != (struct sk_buff *)&sk->sk_write_queue) && 1081 (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(skb->next)->nr_frags == 0) && 1082 (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(skb->next) == 1) && 1083 (sysctl_tcp_retrans_collapse != 0)) 1084 tcp_retrans_try_collapse(sk, skb, cur_mss); 1085 1086 if(tp->af_specific->rebuild_header(sk)) 1087 return -EHOSTUNREACH; /* Routing failure or similar. */ 1088 1089 /* Some Solaris stacks overoptimize and ignore the FIN on a 1090 * retransmit when old data is attached. So strip it off 1091 * since it is cheap to do so and saves bytes on the network. 1092 */ 1093 if(skb->len > 0 && 1094 (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && 1095 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 1096 if (!pskb_trim(skb, 0)) { 1097 TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1; 1098 skb_shinfo(skb)->tso_segs = 1; 1099 skb_shinfo(skb)->tso_size = 0; 1100 skb->ip_summed = CHECKSUM_NONE; 1101 skb->csum = 0; 1102 } 1103 } 1104 1105 /* Make a copy, if the first transmission SKB clone we made 1106 * is still in somebody's hands, else make a clone. 1107 */ 1108 TCP_SKB_CB(skb)->when = tcp_time_stamp; 1109 tcp_tso_set_push(skb); 1110 1111 err = tcp_transmit_skb(sk, (skb_cloned(skb) ? 1112 pskb_copy(skb, GFP_ATOMIC): 1113 skb_clone(skb, GFP_ATOMIC))); 1114 1115 if (err == 0) { 1116 /* Update global TCP statistics. */ 1117 TCP_INC_STATS(TCP_MIB_RETRANSSEGS); 1118 1119 tp->total_retrans++; 1120 1121 #if FASTRETRANS_DEBUG > 0 1122 if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) { 1123 if (net_ratelimit()) 1124 printk(KERN_DEBUG "retrans_out leaked.\n"); 1125 } 1126 #endif 1127 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 1128 tp->retrans_out += tcp_skb_pcount(skb); 1129 1130 /* Save stamp of the first retransmit. */ 1131 if (!tp->retrans_stamp) 1132 tp->retrans_stamp = TCP_SKB_CB(skb)->when; 1133 1134 tp->undo_retrans++; 1135 1136 /* snd_nxt is stored to detect loss of retransmitted segment, 1137 * see tcp_input.c tcp_sacktag_write_queue(). 1138 */ 1139 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 1140 } 1141 return err; 1142 } 1143 1144 /* This gets called after a retransmit timeout, and the initially 1145 * retransmitted data is acknowledged. It tries to continue 1146 * resending the rest of the retransmit queue, until either 1147 * we've sent it all or the congestion window limit is reached. 1148 * If doing SACK, the first ACK which comes back for a timeout 1149 * based retransmit packet might feed us FACK information again. 1150 * If so, we use it to avoid unnecessarily retransmissions. 1151 */ 1152 void tcp_xmit_retransmit_queue(struct sock *sk) 1153 { 1154 struct tcp_sock *tp = tcp_sk(sk); 1155 struct sk_buff *skb; 1156 int packet_cnt = tp->lost_out; 1157 1158 /* First pass: retransmit lost packets. */ 1159 if (packet_cnt) { 1160 sk_stream_for_retrans_queue(skb, sk) { 1161 __u8 sacked = TCP_SKB_CB(skb)->sacked; 1162 1163 /* Assume this retransmit will generate 1164 * only one packet for congestion window 1165 * calculation purposes. This works because 1166 * tcp_retransmit_skb() will chop up the 1167 * packet to be MSS sized and all the 1168 * packet counting works out. 1169 */ 1170 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 1171 return; 1172 1173 if (sacked&TCPCB_LOST) { 1174 if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) { 1175 if (tcp_retransmit_skb(sk, skb)) 1176 return; 1177 if (tp->ca_state != TCP_CA_Loss) 1178 NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS); 1179 else 1180 NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS); 1181 1182 if (skb == 1183 skb_peek(&sk->sk_write_queue)) 1184 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto); 1185 } 1186 1187 packet_cnt -= tcp_skb_pcount(skb); 1188 if (packet_cnt <= 0) 1189 break; 1190 } 1191 } 1192 } 1193 1194 /* OK, demanded retransmission is finished. */ 1195 1196 /* Forward retransmissions are possible only during Recovery. */ 1197 if (tp->ca_state != TCP_CA_Recovery) 1198 return; 1199 1200 /* No forward retransmissions in Reno are possible. */ 1201 if (!tp->rx_opt.sack_ok) 1202 return; 1203 1204 /* Yeah, we have to make difficult choice between forward transmission 1205 * and retransmission... Both ways have their merits... 1206 * 1207 * For now we do not retransmit anything, while we have some new 1208 * segments to send. 1209 */ 1210 1211 if (tcp_may_send_now(sk, tp)) 1212 return; 1213 1214 packet_cnt = 0; 1215 1216 sk_stream_for_retrans_queue(skb, sk) { 1217 /* Similar to the retransmit loop above we 1218 * can pretend that the retransmitted SKB 1219 * we send out here will be composed of one 1220 * real MSS sized packet because tcp_retransmit_skb() 1221 * will fragment it if necessary. 1222 */ 1223 if (++packet_cnt > tp->fackets_out) 1224 break; 1225 1226 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 1227 break; 1228 1229 if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) 1230 continue; 1231 1232 /* Ok, retransmit it. */ 1233 if (tcp_retransmit_skb(sk, skb)) 1234 break; 1235 1236 if (skb == skb_peek(&sk->sk_write_queue)) 1237 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto); 1238 1239 NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS); 1240 } 1241 } 1242 1243 1244 /* Send a fin. The caller locks the socket for us. This cannot be 1245 * allowed to fail queueing a FIN frame under any circumstances. 1246 */ 1247 void tcp_send_fin(struct sock *sk) 1248 { 1249 struct tcp_sock *tp = tcp_sk(sk); 1250 struct sk_buff *skb = skb_peek_tail(&sk->sk_write_queue); 1251 int mss_now; 1252 1253 /* Optimization, tack on the FIN if we have a queue of 1254 * unsent frames. But be careful about outgoing SACKS 1255 * and IP options. 1256 */ 1257 mss_now = tcp_current_mss(sk, 1); 1258 1259 if (sk->sk_send_head != NULL) { 1260 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN; 1261 TCP_SKB_CB(skb)->end_seq++; 1262 tp->write_seq++; 1263 } else { 1264 /* Socket is locked, keep trying until memory is available. */ 1265 for (;;) { 1266 skb = alloc_skb(MAX_TCP_HEADER, GFP_KERNEL); 1267 if (skb) 1268 break; 1269 yield(); 1270 } 1271 1272 /* Reserve space for headers and prepare control bits. */ 1273 skb_reserve(skb, MAX_TCP_HEADER); 1274 skb->csum = 0; 1275 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN); 1276 TCP_SKB_CB(skb)->sacked = 0; 1277 skb_shinfo(skb)->tso_segs = 1; 1278 skb_shinfo(skb)->tso_size = 0; 1279 1280 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 1281 TCP_SKB_CB(skb)->seq = tp->write_seq; 1282 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1; 1283 tcp_queue_skb(sk, skb); 1284 } 1285 __tcp_push_pending_frames(sk, tp, mss_now, TCP_NAGLE_OFF); 1286 } 1287 1288 /* We get here when a process closes a file descriptor (either due to 1289 * an explicit close() or as a byproduct of exit()'ing) and there 1290 * was unread data in the receive queue. This behavior is recommended 1291 * by draft-ietf-tcpimpl-prob-03.txt section 3.10. -DaveM 1292 */ 1293 void tcp_send_active_reset(struct sock *sk, int priority) 1294 { 1295 struct tcp_sock *tp = tcp_sk(sk); 1296 struct sk_buff *skb; 1297 1298 /* NOTE: No TCP options attached and we never retransmit this. */ 1299 skb = alloc_skb(MAX_TCP_HEADER, priority); 1300 if (!skb) { 1301 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); 1302 return; 1303 } 1304 1305 /* Reserve space for headers and prepare control bits. */ 1306 skb_reserve(skb, MAX_TCP_HEADER); 1307 skb->csum = 0; 1308 TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST); 1309 TCP_SKB_CB(skb)->sacked = 0; 1310 skb_shinfo(skb)->tso_segs = 1; 1311 skb_shinfo(skb)->tso_size = 0; 1312 1313 /* Send it off. */ 1314 TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk, tp); 1315 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq; 1316 TCP_SKB_CB(skb)->when = tcp_time_stamp; 1317 if (tcp_transmit_skb(sk, skb)) 1318 NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); 1319 } 1320 1321 /* WARNING: This routine must only be called when we have already sent 1322 * a SYN packet that crossed the incoming SYN that caused this routine 1323 * to get called. If this assumption fails then the initial rcv_wnd 1324 * and rcv_wscale values will not be correct. 1325 */ 1326 int tcp_send_synack(struct sock *sk) 1327 { 1328 struct sk_buff* skb; 1329 1330 skb = skb_peek(&sk->sk_write_queue); 1331 if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) { 1332 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); 1333 return -EFAULT; 1334 } 1335 if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) { 1336 if (skb_cloned(skb)) { 1337 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 1338 if (nskb == NULL) 1339 return -ENOMEM; 1340 __skb_unlink(skb, &sk->sk_write_queue); 1341 skb_header_release(nskb); 1342 __skb_queue_head(&sk->sk_write_queue, nskb); 1343 sk_stream_free_skb(sk, skb); 1344 sk_charge_skb(sk, nskb); 1345 skb = nskb; 1346 } 1347 1348 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK; 1349 TCP_ECN_send_synack(tcp_sk(sk), skb); 1350 } 1351 TCP_SKB_CB(skb)->when = tcp_time_stamp; 1352 return tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC)); 1353 } 1354 1355 /* 1356 * Prepare a SYN-ACK. 1357 */ 1358 struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst, 1359 struct request_sock *req) 1360 { 1361 struct inet_request_sock *ireq = inet_rsk(req); 1362 struct tcp_sock *tp = tcp_sk(sk); 1363 struct tcphdr *th; 1364 int tcp_header_size; 1365 struct sk_buff *skb; 1366 1367 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); 1368 if (skb == NULL) 1369 return NULL; 1370 1371 /* Reserve space for headers. */ 1372 skb_reserve(skb, MAX_TCP_HEADER); 1373 1374 skb->dst = dst_clone(dst); 1375 1376 tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS + 1377 (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) + 1378 (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) + 1379 /* SACK_PERM is in the place of NOP NOP of TS */ 1380 ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0)); 1381 skb->h.th = th = (struct tcphdr *) skb_push(skb, tcp_header_size); 1382 1383 memset(th, 0, sizeof(struct tcphdr)); 1384 th->syn = 1; 1385 th->ack = 1; 1386 if (dst->dev->features&NETIF_F_TSO) 1387 ireq->ecn_ok = 0; 1388 TCP_ECN_make_synack(req, th); 1389 th->source = inet_sk(sk)->sport; 1390 th->dest = ireq->rmt_port; 1391 TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn; 1392 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1; 1393 TCP_SKB_CB(skb)->sacked = 0; 1394 skb_shinfo(skb)->tso_segs = 1; 1395 skb_shinfo(skb)->tso_size = 0; 1396 th->seq = htonl(TCP_SKB_CB(skb)->seq); 1397 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); 1398 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 1399 __u8 rcv_wscale; 1400 /* Set this up on the first call only */ 1401 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 1402 /* tcp_full_space because it is guaranteed to be the first packet */ 1403 tcp_select_initial_window(tcp_full_space(sk), 1404 dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 1405 &req->rcv_wnd, 1406 &req->window_clamp, 1407 ireq->wscale_ok, 1408 &rcv_wscale); 1409 ireq->rcv_wscale = rcv_wscale; 1410 } 1411 1412 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 1413 th->window = htons(req->rcv_wnd); 1414 1415 TCP_SKB_CB(skb)->when = tcp_time_stamp; 1416 tcp_syn_build_options((__u32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok, 1417 ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale, 1418 TCP_SKB_CB(skb)->when, 1419 req->ts_recent); 1420 1421 skb->csum = 0; 1422 th->doff = (tcp_header_size >> 2); 1423 TCP_INC_STATS(TCP_MIB_OUTSEGS); 1424 return skb; 1425 } 1426 1427 /* 1428 * Do all connect socket setups that can be done AF independent. 1429 */ 1430 static inline void tcp_connect_init(struct sock *sk) 1431 { 1432 struct dst_entry *dst = __sk_dst_get(sk); 1433 struct tcp_sock *tp = tcp_sk(sk); 1434 __u8 rcv_wscale; 1435 1436 /* We'll fix this up when we get a response from the other end. 1437 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 1438 */ 1439 tp->tcp_header_len = sizeof(struct tcphdr) + 1440 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 1441 1442 /* If user gave his TCP_MAXSEG, record it to clamp */ 1443 if (tp->rx_opt.user_mss) 1444 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 1445 tp->max_window = 0; 1446 tcp_sync_mss(sk, dst_mtu(dst)); 1447 1448 if (!tp->window_clamp) 1449 tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 1450 tp->advmss = dst_metric(dst, RTAX_ADVMSS); 1451 tcp_initialize_rcv_mss(sk); 1452 tcp_ca_init(tp); 1453 1454 tcp_select_initial_window(tcp_full_space(sk), 1455 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 1456 &tp->rcv_wnd, 1457 &tp->window_clamp, 1458 sysctl_tcp_window_scaling, 1459 &rcv_wscale); 1460 1461 tp->rx_opt.rcv_wscale = rcv_wscale; 1462 tp->rcv_ssthresh = tp->rcv_wnd; 1463 1464 sk->sk_err = 0; 1465 sock_reset_flag(sk, SOCK_DONE); 1466 tp->snd_wnd = 0; 1467 tcp_init_wl(tp, tp->write_seq, 0); 1468 tp->snd_una = tp->write_seq; 1469 tp->snd_sml = tp->write_seq; 1470 tp->rcv_nxt = 0; 1471 tp->rcv_wup = 0; 1472 tp->copied_seq = 0; 1473 1474 tp->rto = TCP_TIMEOUT_INIT; 1475 tp->retransmits = 0; 1476 tcp_clear_retrans(tp); 1477 } 1478 1479 /* 1480 * Build a SYN and send it off. 1481 */ 1482 int tcp_connect(struct sock *sk) 1483 { 1484 struct tcp_sock *tp = tcp_sk(sk); 1485 struct sk_buff *buff; 1486 1487 tcp_connect_init(sk); 1488 1489 buff = alloc_skb(MAX_TCP_HEADER + 15, sk->sk_allocation); 1490 if (unlikely(buff == NULL)) 1491 return -ENOBUFS; 1492 1493 /* Reserve space for headers. */ 1494 skb_reserve(buff, MAX_TCP_HEADER); 1495 1496 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN; 1497 TCP_ECN_send_syn(sk, tp, buff); 1498 TCP_SKB_CB(buff)->sacked = 0; 1499 skb_shinfo(buff)->tso_segs = 1; 1500 skb_shinfo(buff)->tso_size = 0; 1501 buff->csum = 0; 1502 TCP_SKB_CB(buff)->seq = tp->write_seq++; 1503 TCP_SKB_CB(buff)->end_seq = tp->write_seq; 1504 tp->snd_nxt = tp->write_seq; 1505 tp->pushed_seq = tp->write_seq; 1506 tcp_ca_init(tp); 1507 1508 /* Send it off. */ 1509 TCP_SKB_CB(buff)->when = tcp_time_stamp; 1510 tp->retrans_stamp = TCP_SKB_CB(buff)->when; 1511 skb_header_release(buff); 1512 __skb_queue_tail(&sk->sk_write_queue, buff); 1513 sk_charge_skb(sk, buff); 1514 tp->packets_out += tcp_skb_pcount(buff); 1515 tcp_transmit_skb(sk, skb_clone(buff, GFP_KERNEL)); 1516 TCP_INC_STATS(TCP_MIB_ACTIVEOPENS); 1517 1518 /* Timer for repeating the SYN until an answer. */ 1519 tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto); 1520 return 0; 1521 } 1522 1523 /* Send out a delayed ack, the caller does the policy checking 1524 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 1525 * for details. 1526 */ 1527 void tcp_send_delayed_ack(struct sock *sk) 1528 { 1529 struct tcp_sock *tp = tcp_sk(sk); 1530 int ato = tp->ack.ato; 1531 unsigned long timeout; 1532 1533 if (ato > TCP_DELACK_MIN) { 1534 int max_ato = HZ/2; 1535 1536 if (tp->ack.pingpong || (tp->ack.pending&TCP_ACK_PUSHED)) 1537 max_ato = TCP_DELACK_MAX; 1538 1539 /* Slow path, intersegment interval is "high". */ 1540 1541 /* If some rtt estimate is known, use it to bound delayed ack. 1542 * Do not use tp->rto here, use results of rtt measurements 1543 * directly. 1544 */ 1545 if (tp->srtt) { 1546 int rtt = max(tp->srtt>>3, TCP_DELACK_MIN); 1547 1548 if (rtt < max_ato) 1549 max_ato = rtt; 1550 } 1551 1552 ato = min(ato, max_ato); 1553 } 1554 1555 /* Stay within the limit we were given */ 1556 timeout = jiffies + ato; 1557 1558 /* Use new timeout only if there wasn't a older one earlier. */ 1559 if (tp->ack.pending&TCP_ACK_TIMER) { 1560 /* If delack timer was blocked or is about to expire, 1561 * send ACK now. 1562 */ 1563 if (tp->ack.blocked || time_before_eq(tp->ack.timeout, jiffies+(ato>>2))) { 1564 tcp_send_ack(sk); 1565 return; 1566 } 1567 1568 if (!time_before(timeout, tp->ack.timeout)) 1569 timeout = tp->ack.timeout; 1570 } 1571 tp->ack.pending |= TCP_ACK_SCHED|TCP_ACK_TIMER; 1572 tp->ack.timeout = timeout; 1573 sk_reset_timer(sk, &tp->delack_timer, timeout); 1574 } 1575 1576 /* This routine sends an ack and also updates the window. */ 1577 void tcp_send_ack(struct sock *sk) 1578 { 1579 /* If we have been reset, we may not send again. */ 1580 if (sk->sk_state != TCP_CLOSE) { 1581 struct tcp_sock *tp = tcp_sk(sk); 1582 struct sk_buff *buff; 1583 1584 /* We are not putting this on the write queue, so 1585 * tcp_transmit_skb() will set the ownership to this 1586 * sock. 1587 */ 1588 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 1589 if (buff == NULL) { 1590 tcp_schedule_ack(tp); 1591 tp->ack.ato = TCP_ATO_MIN; 1592 tcp_reset_xmit_timer(sk, TCP_TIME_DACK, TCP_DELACK_MAX); 1593 return; 1594 } 1595 1596 /* Reserve space for headers and prepare control bits. */ 1597 skb_reserve(buff, MAX_TCP_HEADER); 1598 buff->csum = 0; 1599 TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK; 1600 TCP_SKB_CB(buff)->sacked = 0; 1601 skb_shinfo(buff)->tso_segs = 1; 1602 skb_shinfo(buff)->tso_size = 0; 1603 1604 /* Send it off, this clears delayed acks for us. */ 1605 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk, tp); 1606 TCP_SKB_CB(buff)->when = tcp_time_stamp; 1607 tcp_transmit_skb(sk, buff); 1608 } 1609 } 1610 1611 /* This routine sends a packet with an out of date sequence 1612 * number. It assumes the other end will try to ack it. 1613 * 1614 * Question: what should we make while urgent mode? 1615 * 4.4BSD forces sending single byte of data. We cannot send 1616 * out of window data, because we have SND.NXT==SND.MAX... 1617 * 1618 * Current solution: to send TWO zero-length segments in urgent mode: 1619 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 1620 * out-of-date with SND.UNA-1 to probe window. 1621 */ 1622 static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 1623 { 1624 struct tcp_sock *tp = tcp_sk(sk); 1625 struct sk_buff *skb; 1626 1627 /* We don't queue it, tcp_transmit_skb() sets ownership. */ 1628 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 1629 if (skb == NULL) 1630 return -1; 1631 1632 /* Reserve space for headers and set control bits. */ 1633 skb_reserve(skb, MAX_TCP_HEADER); 1634 skb->csum = 0; 1635 TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK; 1636 TCP_SKB_CB(skb)->sacked = urgent; 1637 skb_shinfo(skb)->tso_segs = 1; 1638 skb_shinfo(skb)->tso_size = 0; 1639 1640 /* Use a previous sequence. This should cause the other 1641 * end to send an ack. Don't queue or clone SKB, just 1642 * send it. 1643 */ 1644 TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1; 1645 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq; 1646 TCP_SKB_CB(skb)->when = tcp_time_stamp; 1647 return tcp_transmit_skb(sk, skb); 1648 } 1649 1650 int tcp_write_wakeup(struct sock *sk) 1651 { 1652 if (sk->sk_state != TCP_CLOSE) { 1653 struct tcp_sock *tp = tcp_sk(sk); 1654 struct sk_buff *skb; 1655 1656 if ((skb = sk->sk_send_head) != NULL && 1657 before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) { 1658 int err; 1659 unsigned int mss = tcp_current_mss(sk, 0); 1660 unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq; 1661 1662 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 1663 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 1664 1665 /* We are probing the opening of a window 1666 * but the window size is != 0 1667 * must have been a result SWS avoidance ( sender ) 1668 */ 1669 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 1670 skb->len > mss) { 1671 seg_size = min(seg_size, mss); 1672 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 1673 if (tcp_fragment(sk, skb, seg_size)) 1674 return -1; 1675 /* SWS override triggered forced fragmentation. 1676 * Disable TSO, the connection is too sick. */ 1677 if (sk->sk_route_caps & NETIF_F_TSO) { 1678 sock_set_flag(sk, SOCK_NO_LARGESEND); 1679 sk->sk_route_caps &= ~NETIF_F_TSO; 1680 tp->mss_cache = tp->mss_cache_std; 1681 } 1682 } else if (!tcp_skb_pcount(skb)) 1683 tcp_set_skb_tso_segs(sk, skb); 1684 1685 TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 1686 TCP_SKB_CB(skb)->when = tcp_time_stamp; 1687 tcp_tso_set_push(skb); 1688 err = tcp_transmit_skb(sk, skb_clone(skb, GFP_ATOMIC)); 1689 if (!err) { 1690 update_send_head(sk, tp, skb); 1691 } 1692 return err; 1693 } else { 1694 if (tp->urg_mode && 1695 between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF)) 1696 tcp_xmit_probe_skb(sk, TCPCB_URG); 1697 return tcp_xmit_probe_skb(sk, 0); 1698 } 1699 } 1700 return -1; 1701 } 1702 1703 /* A window probe timeout has occurred. If window is not closed send 1704 * a partial packet else a zero probe. 1705 */ 1706 void tcp_send_probe0(struct sock *sk) 1707 { 1708 struct tcp_sock *tp = tcp_sk(sk); 1709 int err; 1710 1711 err = tcp_write_wakeup(sk); 1712 1713 if (tp->packets_out || !sk->sk_send_head) { 1714 /* Cancel probe timer, if it is not required. */ 1715 tp->probes_out = 0; 1716 tp->backoff = 0; 1717 return; 1718 } 1719 1720 if (err <= 0) { 1721 if (tp->backoff < sysctl_tcp_retries2) 1722 tp->backoff++; 1723 tp->probes_out++; 1724 tcp_reset_xmit_timer (sk, TCP_TIME_PROBE0, 1725 min(tp->rto << tp->backoff, TCP_RTO_MAX)); 1726 } else { 1727 /* If packet was not sent due to local congestion, 1728 * do not backoff and do not remember probes_out. 1729 * Let local senders to fight for local resources. 1730 * 1731 * Use accumulated backoff yet. 1732 */ 1733 if (!tp->probes_out) 1734 tp->probes_out=1; 1735 tcp_reset_xmit_timer (sk, TCP_TIME_PROBE0, 1736 min(tp->rto << tp->backoff, TCP_RESOURCE_PROBE_INTERVAL)); 1737 } 1738 } 1739 1740 EXPORT_SYMBOL(tcp_connect); 1741 EXPORT_SYMBOL(tcp_make_synack); 1742 EXPORT_SYMBOL(tcp_simple_retransmit); 1743 EXPORT_SYMBOL(tcp_sync_mss); 1744