tcp_sack.c revision 1.16
1/* $NetBSD: tcp_sack.c,v 1.16 2006/10/07 19:26:07 yamt Exp $ */ 2 3/* 4 * Copyright (c) 2005 The NetBSD Foundation, Inc. 5 * All rights reserved. 6 * 7 * This code is derived from software contributed to The NetBSD Foundation 8 * by Kentaro A. Kurahone. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the NetBSD 21 * Foundation, Inc. and its contributors. 22 * 4. Neither the name of The NetBSD Foundation nor the names of its 23 * contributors may be used to endorse or promote products derived 24 * from this software without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 27 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 28 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 29 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 30 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 31 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 32 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 33 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 34 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 35 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 36 * POSSIBILITY OF SUCH DAMAGE. 37 */ 38 39/* 40 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995 41 * The Regents of the University of California. All rights reserved. 42 * 43 * Redistribution and use in source and binary forms, with or without 44 * modification, are permitted provided that the following conditions 45 * are met: 46 * 1. Redistributions of source code must retain the above copyright 47 * notice, this list of conditions and the following disclaimer. 48 * 2. Redistributions in binary form must reproduce the above copyright 49 * notice, this list of conditions and the following disclaimer in the 50 * documentation and/or other materials provided with the distribution. 51 * 4. Neither the name of the University nor the names of its contributors 52 * may be used to endorse or promote products derived from this software 53 * without specific prior written permission. 54 * 55 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 56 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 57 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 58 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 59 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 60 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 61 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 62 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 63 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 64 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 65 * SUCH DAMAGE. 66 * 67 * @(#)tcp_sack.c 8.12 (Berkeley) 5/24/95 68 * $FreeBSD: src/sys/netinet/tcp_sack.c,v 1.3.2.2 2004/12/25 23:02:57 rwatson Exp $ 69 */ 70 71/* 72 * @@(#)COPYRIGHT 1.1 (NRL) 17 January 1995 73 * 74 * NRL grants permission for redistribution and use in source and binary 75 * forms, with or without modification, of the software and documentation 76 * created at NRL provided that the following conditions are met: 77 * 78 * 1. Redistributions of source code must retain the above copyright 79 * notice, this list of conditions and the following disclaimer. 80 * 2. Redistributions in binary form must reproduce the above copyright 81 * notice, this list of conditions and the following disclaimer in the 82 * documentation and/or other materials provided with the distribution. 83 * 3. All advertising materials mentioning features or use of this software 84 * must display the following acknowledgements: 85 * This product includes software developed by the University of 86 * California, Berkeley and its contributors. 87 * This product includes software developed at the Information 88 * Technology Division, US Naval Research Laboratory. 89 * 4. Neither the name of the NRL nor the names of its contributors 90 * may be used to endorse or promote products derived from this software 91 * without specific prior written permission. 92 * 93 * THE SOFTWARE PROVIDED BY NRL IS PROVIDED BY NRL AND CONTRIBUTORS ``AS 94 * IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 95 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A 96 * PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NRL OR 97 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, 98 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, 99 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR 100 * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF 101 * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING 102 * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS 103 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 104 * 105 * The views and conclusions contained in the software and documentation 106 * are those of the authors and should not be interpreted as representing 107 * official policies, either expressed or implied, of the US Naval 108 * Research Laboratory (NRL). 109 */ 110 111#include <sys/cdefs.h> 112__KERNEL_RCSID(0, "$NetBSD: tcp_sack.c,v 1.16 2006/10/07 19:26:07 yamt Exp $"); 113 114#include "opt_inet.h" 115#include "opt_ipsec.h" 116#include "opt_inet_csum.h" 117#include "opt_tcp_debug.h" 118 119#include <sys/param.h> 120#include <sys/systm.h> 121#include <sys/malloc.h> 122#include <sys/mbuf.h> 123#include <sys/protosw.h> 124#include <sys/socket.h> 125#include <sys/socketvar.h> 126#include <sys/errno.h> 127#include <sys/syslog.h> 128#include <sys/pool.h> 129#include <sys/domain.h> 130#include <sys/kernel.h> 131 132#include <net/if.h> 133#include <net/route.h> 134#include <net/if_types.h> 135 136#include <netinet/in.h> 137#include <netinet/in_systm.h> 138#include <netinet/ip.h> 139#include <netinet/in_pcb.h> 140#include <netinet/in_var.h> 141#include <netinet/ip_var.h> 142 143#ifdef INET6 144#ifndef INET 145#include <netinet/in.h> 146#endif 147#include <netinet/ip6.h> 148#include <netinet6/ip6_var.h> 149#include <netinet6/in6_pcb.h> 150#include <netinet6/ip6_var.h> 151#include <netinet6/in6_var.h> 152#include <netinet/icmp6.h> 153#include <netinet6/nd6.h> 154#endif 155 156#ifndef INET6 157/* always need ip6.h for IP6_EXTHDR_GET */ 158#include <netinet/ip6.h> 159#endif 160 161#include <netinet/tcp.h> 162#include <netinet/tcp_fsm.h> 163#include <netinet/tcp_seq.h> 164#include <netinet/tcp_timer.h> 165#include <netinet/tcp_var.h> 166#include <netinet/tcpip.h> 167#include <netinet/tcp_debug.h> 168 169#include <machine/stdarg.h> 170 171/* SACK block pool. */ 172POOL_INIT(sackhole_pool, sizeof(struct sackhole), 0, 0, 0, "sackholepl", NULL); 173 174void 175tcp_new_dsack(struct tcpcb *tp, tcp_seq seq, u_int32_t len) 176{ 177 if (TCP_SACK_ENABLED(tp)) { 178 tp->rcv_dsack_block.left = seq; 179 tp->rcv_dsack_block.right = seq + len; 180 tp->rcv_sack_flags |= TCPSACK_HAVED; 181 } 182} 183 184void 185tcp_sack_option(struct tcpcb *tp, struct tcphdr *th, u_char *cp, int optlen) 186{ 187 struct sackblk 188 t_sack_block[(MAX_TCPOPTLEN - 2) / (sizeof(u_int32_t) * 2)]; 189 struct sackblk *sack = NULL; 190 struct sackhole *cur = NULL; 191 struct sackhole *tmp = NULL; 192 u_int32_t *lp = (u_int32_t *) (cp + 2); 193 int i, j, num_sack_blks, s; 194 tcp_seq left, right, acked; 195 196 /* 197 * If we aren't processing SACK responses, this is not an ACK 198 * or the peer sends us a sack option with invalid length, don't 199 * update the scoreboard. 200 */ 201 if (!TCP_SACK_ENABLED(tp) || ((th->th_flags & TH_ACK) == 0) || 202 (optlen % 8 != 2 || optlen < 10)) { 203 return; 204 } 205 206 /* 207 * If we don't want any SACK holes to be allocated, just return. 208 */ 209 if (tcp_sack_globalmaxholes == 0 || tcp_sack_tp_maxholes == 0) { 210 return; 211 } 212 213 /* If the ACK is outside [snd_una, snd_max], ignore the SACK options. */ 214 if (SEQ_LT(th->th_ack, tp->snd_una) || SEQ_GT(th->th_ack, tp->snd_max)) 215 return; 216 217 /* 218 * Extract SACK blocks. 219 * 220 * Note that t_sack_block is sorted so that we only need to do 221 * one pass over the sequence number space. (SACK "fast-path") 222 */ 223 num_sack_blks = optlen / 8; 224 acked = (SEQ_GT(th->th_ack, tp->snd_una)) ? th->th_ack : tp->snd_una; 225 for (i = 0; i < num_sack_blks; i++, lp += 2) { 226 memcpy(&left, lp, sizeof(*lp)); 227 memcpy(&right, lp + 1, sizeof(*lp)); 228 left = ntohl(left); 229 right = ntohl(right); 230 231 if (SEQ_LEQ(right, acked) || SEQ_GT(right, tp->snd_max) || 232 SEQ_GEQ(left, right)) { 233 /* SACK entry that's old, or invalid. */ 234 i--; 235 num_sack_blks--; 236 continue; 237 } 238 239 /* Insertion sort. */ 240 for (j = i; (j > 0) && SEQ_LT(left, t_sack_block[j - 1].left); 241 j--) { 242 t_sack_block[j].left = t_sack_block[j - 1].left; 243 t_sack_block[j].right = t_sack_block[j - 1].right; 244 } 245 t_sack_block[j].left = left; 246 t_sack_block[j].right = right; 247 } 248 249 /* XXX: Investigate making this a bit more fine-grained. */ 250 s = splsoftnet(); 251 252 /* Update the scoreboard. */ 253 cur = TAILQ_FIRST(&tp->snd_holes); 254 for (i = 0; i < num_sack_blks; i++) { 255 sack = &t_sack_block[i]; 256 /* 257 * FACK TCP. Update snd_fack so we can enter Fast 258 * Recovery early. 259 */ 260 if (SEQ_GEQ(sack->right, tp->snd_fack)) 261 tp->snd_fack = sack->right; 262 263 if (TAILQ_EMPTY(&tp->snd_holes)) { 264 /* First hole. */ 265 if (tcp_sack_globalholes >= tcp_sack_globalmaxholes) { 266 splx(s); 267 return; 268 } 269 cur = (struct sackhole *) 270 pool_get(&sackhole_pool, PR_NOWAIT); 271 if (cur == NULL) { 272 /* ENOBUFS, bail out*/ 273 splx(s); 274 return; 275 } 276 cur->start = th->th_ack; 277 cur->end = sack->left; 278 cur->rxmit = cur->start; 279 tp->rcv_lastsack = sack->right; 280 tp->snd_numholes++; 281 tcp_sack_globalholes++; 282 TAILQ_INSERT_HEAD(&tp->snd_holes, cur, sackhole_q); 283 continue; /* With next sack block */ 284 } 285 286 /* Go through the list of holes. */ 287 while (cur) { 288 if (SEQ_LEQ(sack->right, cur->start)) 289 /* SACKs data before the current hole */ 290 break; /* No use going through more holes */ 291 292 if (SEQ_GEQ(sack->left, cur->end)) { 293 /* SACKs data beyond the current hole */ 294 cur = TAILQ_NEXT(cur, sackhole_q); 295 continue; 296 } 297 298 if (SEQ_LEQ(sack->left, cur->start)) { 299 /* Data acks at least the beginning of hole */ 300 if (SEQ_GEQ(sack->right, cur->end)) { 301 /* Acks entire hole, so delete hole */ 302 tmp = cur; 303 cur = TAILQ_NEXT(cur, sackhole_q); 304 tp->snd_numholes--; 305 tcp_sack_globalholes--; 306 TAILQ_REMOVE(&tp->snd_holes, tmp, 307 sackhole_q); 308 pool_put(&sackhole_pool, tmp); 309 break; 310 } 311 312 /* Otherwise, move start of hole forward */ 313 cur->start = sack->right; 314 cur->rxmit = SEQ_MAX(cur->rxmit, cur->start); 315 break; 316 } 317 318 if (SEQ_GEQ(sack->right, cur->end)) { 319 /* Move end of hole backward. */ 320 cur->end = sack->left; 321 cur->rxmit = SEQ_MIN(cur->rxmit, cur->end); 322 cur = TAILQ_NEXT(cur, sackhole_q); 323 break; 324 } 325 326 if (SEQ_LT(cur->start, sack->left) && 327 SEQ_GT(cur->end, sack->right)) { 328 /* 329 * ACKs some data in middle of a hole; need to 330 * split current hole 331 */ 332 if (tcp_sack_globalholes >= 333 tcp_sack_globalmaxholes || 334 tp->snd_numholes >= 335 tcp_sack_tp_maxholes) { 336 splx(s); 337 return; 338 } 339 tmp = (struct sackhole *) 340 pool_get(&sackhole_pool, PR_NOWAIT); 341 if (tmp == NULL) { 342 /* ENOBUFS, bail out. */ 343 splx(s); 344 return; 345 } 346 tmp->start = sack->right; 347 tmp->end = cur->end; 348 tmp->rxmit = SEQ_MAX(cur->rxmit, tmp->start); 349 cur->end = sack->left; 350 cur->rxmit = SEQ_MIN(cur->rxmit, cur->end); 351 tp->snd_numholes++; 352 tcp_sack_globalholes++; 353 TAILQ_INSERT_AFTER(&tp->snd_holes, cur, tmp, 354 sackhole_q); 355 cur = tmp; 356 break; 357 } 358 } 359 360 /* At this point, we have reached the tail of the list. */ 361 if (SEQ_LT(tp->rcv_lastsack, sack->left)) { 362 /* 363 * Need to append new hole at end. 364 */ 365 if (tcp_sack_globalholes >= 366 tcp_sack_globalmaxholes || 367 tp->snd_numholes >= 368 tcp_sack_tp_maxholes) { 369 splx(s); 370 return; 371 } 372 tmp = (struct sackhole *) 373 pool_get(&sackhole_pool, PR_NOWAIT); 374 if (tmp == NULL) 375 continue; /* ENOBUFS */ 376 tmp->start = tp->rcv_lastsack; 377 tmp->end = sack->left; 378 tmp->rxmit = tmp->start; 379 tp->snd_numholes++; 380 tcp_sack_globalholes++; 381 TAILQ_INSERT_TAIL(&tp->snd_holes, tmp, sackhole_q); 382 cur = tmp; 383 } 384 if (SEQ_LT(tp->rcv_lastsack, sack->right)) { 385 tp->rcv_lastsack = sack->right; 386 } 387 } 388 389 splx(s); 390} 391 392void 393tcp_del_sackholes(struct tcpcb *tp, struct tcphdr *th) 394{ 395 /* Max because this could be an older ack that just arrived. */ 396 tcp_seq lastack = SEQ_GT(th->th_ack, tp->snd_una) ? 397 th->th_ack : tp->snd_una; 398 struct sackhole *cur = TAILQ_FIRST(&tp->snd_holes); 399 struct sackhole *tmp; 400 int s; 401 402 s = splsoftnet(); 403 404 while (cur) { 405 if (SEQ_LEQ(cur->end, lastack)) { 406 tmp = cur; 407 cur = TAILQ_NEXT(cur, sackhole_q); 408 tp->snd_numholes--; 409 tcp_sack_globalholes--; 410 TAILQ_REMOVE(&tp->snd_holes, tmp, sackhole_q); 411 pool_put(&sackhole_pool, tmp); 412 } else if (SEQ_LT(cur->start, lastack)) { 413 cur->start = lastack; 414 if (SEQ_LT(cur->rxmit, cur->start)) 415 cur->rxmit = cur->start; 416 break; 417 } else 418 break; 419 } 420 421 splx(s); 422} 423 424void 425tcp_free_sackholes(struct tcpcb *tp) 426{ 427 struct sackhole *sack; 428 int s; 429 430 s = splsoftnet(); 431 432 /* Free up the SACK hole list. */ 433 while (!TAILQ_EMPTY(&tp->snd_holes)) { 434 sack = TAILQ_FIRST(&tp->snd_holes); 435 tcp_sack_globalholes--; 436 TAILQ_REMOVE(&tp->snd_holes, sack, sackhole_q); 437 pool_put(&sackhole_pool, sack); 438 } 439 440 tp->snd_numholes = 0; 441 442 splx(s); 443} 444 445/* 446 * Implements the SACK response to a new ack, checking for partial acks 447 * in fast recovery. 448 */ 449void 450tcp_sack_newack(struct tcpcb *tp, struct tcphdr *th) 451{ 452 if (tp->t_partialacks < 0) { 453 /* 454 * Not in fast recovery. Reset the duplicate ack 455 * counter. 456 */ 457 tp->t_dupacks = 0; 458 } else if (SEQ_LT(th->th_ack, tp->snd_recover)) { 459 /* 460 * Partial ack handling within a sack recovery episode. 461 * Keeping this very simple for now. When a partial ack 462 * is received, force snd_cwnd to a value that will allow 463 * the sender to transmit no more than 2 segments. 464 * If necessary, a fancier scheme can be adopted at a 465 * later point, but for now, the goal is to prevent the 466 * sender from bursting a large amount of data in the midst 467 * of sack recovery. 468 */ 469 int num_segs = 1; 470 int sack_bytes_rxmt = 0; 471 472 tp->t_partialacks++; 473 TCP_TIMER_DISARM(tp, TCPT_REXMT); 474 tp->t_rtttime = 0; 475 476 /* 477 * send one or 2 segments based on how much new data was acked 478 */ 479 if (((th->th_ack - tp->snd_una) / tp->t_segsz) > 2) 480 num_segs = 2; 481 (void)tcp_sack_output(tp, &sack_bytes_rxmt); 482 tp->snd_cwnd = sack_bytes_rxmt + 483 (tp->snd_nxt - tp->sack_newdata) + num_segs * tp->t_segsz; 484 tp->t_flags |= TF_ACKNOW; 485 (void) tcp_output(tp); 486 } else { 487 /* 488 * Complete ack, inflate the congestion window to 489 * ssthresh and exit fast recovery. 490 * 491 * Window inflation should have left us with approx. 492 * snd_ssthresh outstanding data. But in case we 493 * would be inclined to send a burst, better to do 494 * it via the slow start mechanism. 495 */ 496 if (SEQ_SUB(tp->snd_max, th->th_ack) < tp->snd_ssthresh) 497 tp->snd_cwnd = SEQ_SUB(tp->snd_max, th->th_ack) 498 + tp->t_segsz; 499 else 500 tp->snd_cwnd = tp->snd_ssthresh; 501 tp->t_partialacks = -1; 502 tp->t_dupacks = 0; 503 if (SEQ_GT(th->th_ack, tp->snd_fack)) 504 tp->snd_fack = th->th_ack; 505 } 506} 507 508/* 509 * Returns pointer to a sackhole if there are any pending retransmissions; 510 * NULL otherwise. 511 */ 512struct sackhole * 513tcp_sack_output(struct tcpcb *tp, int *sack_bytes_rexmt) 514{ 515 struct sackhole *cur = NULL; 516 517 if(!TCP_SACK_ENABLED(tp)) 518 return (NULL); 519 520 *sack_bytes_rexmt = 0; 521 TAILQ_FOREACH(cur, &tp->snd_holes, sackhole_q) { 522 if (SEQ_LT(cur->rxmit, cur->end)) { 523 if (SEQ_LT(cur->rxmit, tp->snd_una)) { 524 /* old SACK hole */ 525 continue; 526 } 527 *sack_bytes_rexmt += (cur->rxmit - cur->start); 528 break; 529 } 530 *sack_bytes_rexmt += (cur->rxmit - cur->start); 531 } 532 533 return (cur); 534} 535 536/* 537 * After a timeout, the SACK list may be rebuilt. This SACK information 538 * should be used to avoid retransmitting SACKed data. This function 539 * traverses the SACK list to see if snd_nxt should be moved forward. 540 */ 541void 542tcp_sack_adjust(struct tcpcb *tp) 543{ 544 struct sackhole *cur = TAILQ_FIRST(&tp->snd_holes); 545 struct sackhole *n = NULL; 546 547 if (TAILQ_EMPTY(&tp->snd_holes)) 548 return; /* No holes */ 549 if (SEQ_GEQ(tp->snd_nxt, tp->rcv_lastsack)) 550 return; /* We're already beyond any SACKed blocks */ 551 552 /* 553 * Two cases for which we want to advance snd_nxt: 554 * i) snd_nxt lies between end of one hole and beginning of another 555 * ii) snd_nxt lies between end of last hole and rcv_lastsack 556 */ 557 while ((n = TAILQ_NEXT(cur, sackhole_q)) != NULL) { 558 if (SEQ_LT(tp->snd_nxt, cur->end)) 559 return; 560 if (SEQ_GEQ(tp->snd_nxt, n->start)) 561 cur = n; 562 else { 563 tp->snd_nxt = n->start; 564 return; 565 } 566 } 567 if (SEQ_LT(tp->snd_nxt, cur->end)) 568 return; 569 tp->snd_nxt = tp->rcv_lastsack; 570 571 return; 572} 573 574int 575tcp_sack_numblks(const struct tcpcb *tp) 576{ 577 int numblks; 578 579 if (!TCP_SACK_ENABLED(tp)) { 580 return 0; 581 } 582 583 numblks = (((tp->rcv_sack_flags & TCPSACK_HAVED) != 0) ? 1 : 0) + 584 tp->t_segqlen; 585 586 if (numblks == 0) { 587 return 0; 588 } 589 590 if (numblks > TCP_SACK_MAX) { 591 numblks = TCP_SACK_MAX; 592 } 593 594 return numblks; 595} 596