tcp_sack.c revision 185571
1247835Skib/*- 2247835Skib * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995 3247835Skib * The Regents of the University of California. 4247835Skib * All rights reserved. 5247835Skib * 6247835Skib * Redistribution and use in source and binary forms, with or without 7247835Skib * modification, are permitted provided that the following conditions 8247835Skib * are met: 9247835Skib * 1. Redistributions of source code must retain the above copyright 10247835Skib * notice, this list of conditions and the following disclaimer. 11247835Skib * 2. Redistributions in binary form must reproduce the above copyright 12247835Skib * notice, this list of conditions and the following disclaimer in the 13247835Skib * documentation and/or other materials provided with the distribution. 14247835Skib * 4. Neither the name of the University nor the names of its contributors 15247835Skib * may be used to endorse or promote products derived from this software 16247835Skib * without specific prior written permission. 17247835Skib * 18247835Skib * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 19247835Skib * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 20247835Skib * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 21247835Skib * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 22247835Skib * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 23247835Skib * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 24247835Skib * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 25247835Skib * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 26247835Skib * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 27247835Skib * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 28247835Skib * SUCH DAMAGE. 29247835Skib * 30247835Skib * @(#)tcp_sack.c 8.12 (Berkeley) 5/24/95 31247835Skib */ 32247835Skib 33247835Skib/*- 34247835Skib * @@(#)COPYRIGHT 1.1 (NRL) 17 January 1995 35247835Skib * 36247835Skib * NRL grants permission for redistribution and use in source and binary 37247835Skib * forms, with or without modification, of the software and documentation 38247835Skib * created at NRL provided that the following conditions are met: 39247835Skib * 40247835Skib * 1. Redistributions of source code must retain the above copyright 41247835Skib * notice, this list of conditions and the following disclaimer. 42247835Skib * 2. Redistributions in binary form must reproduce the above copyright 43247835Skib * notice, this list of conditions and the following disclaimer in the 44247835Skib * documentation and/or other materials provided with the distribution. 45247835Skib * 3. All advertising materials mentioning features or use of this software 46247835Skib * must display the following acknowledgements: 47247835Skib * This product includes software developed by the University of 48247835Skib * California, Berkeley and its contributors. 49247835Skib * This product includes software developed at the Information 50247835Skib * Technology Division, US Naval Research Laboratory. 51247835Skib * 4. Neither the name of the NRL nor the names of its contributors 52247835Skib * may be used to endorse or promote products derived from this software 53247835Skib * without specific prior written permission. 54247835Skib * 55247835Skib * THE SOFTWARE PROVIDED BY NRL IS PROVIDED BY NRL AND CONTRIBUTORS ``AS 56247835Skib * IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 57247835Skib * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A 58247835Skib * PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NRL OR 59247835Skib * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, 60247835Skib * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, 61247835Skib * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR 62247835Skib * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF 63247835Skib * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING 64247835Skib * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS 65247835Skib * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 66247835Skib * 67247835Skib * The views and conclusions contained in the software and documentation 68247835Skib * are those of the authors and should not be interpreted as representing 69247835Skib * official policies, either expressed or implied, of the US Naval 70247835Skib * Research Laboratory (NRL). 71247835Skib */ 72247835Skib 73247835Skib#include <sys/cdefs.h> 74247835Skib__FBSDID("$FreeBSD: head/sys/netinet/tcp_sack.c 185571 2008-12-02 21:37:28Z bz $"); 75247835Skib 76247835Skib#include "opt_inet.h" 77247835Skib#include "opt_inet6.h" 78247835Skib#include "opt_tcpdebug.h" 79247835Skib 80247835Skib#include <sys/param.h> 81247835Skib#include <sys/systm.h> 82247835Skib#include <sys/kernel.h> 83247835Skib#include <sys/sysctl.h> 84247835Skib#include <sys/malloc.h> 85247835Skib#include <sys/mbuf.h> 86247835Skib#include <sys/proc.h> /* for proc0 declaration */ 87247835Skib#include <sys/protosw.h> 88247835Skib#include <sys/socket.h> 89247835Skib#include <sys/socketvar.h> 90247835Skib#include <sys/syslog.h> 91247835Skib#include <sys/systm.h> 92247835Skib#include <sys/vimage.h> 93247835Skib 94247835Skib#include <machine/cpu.h> /* before tcp_seq.h, for tcp_random18() */ 95247835Skib 96247835Skib#include <vm/uma.h> 97247835Skib 98247835Skib#include <net/if.h> 99247835Skib#include <net/route.h> 100247835Skib 101247835Skib#include <netinet/in.h> 102247835Skib#include <netinet/in_systm.h> 103247835Skib#include <netinet/ip.h> 104247835Skib#include <netinet/in_var.h> 105247835Skib#include <netinet/in_pcb.h> 106247835Skib#include <netinet/ip_var.h> 107247835Skib#include <netinet/ip6.h> 108247835Skib#include <netinet/icmp6.h> 109247835Skib#include <netinet6/nd6.h> 110247835Skib#include <netinet6/ip6_var.h> 111247835Skib#include <netinet6/in6_pcb.h> 112247835Skib#include <netinet/tcp.h> 113247835Skib#include <netinet/tcp_fsm.h> 114247835Skib#include <netinet/tcp_seq.h> 115247835Skib#include <netinet/tcp_timer.h> 116247835Skib#include <netinet/tcp_var.h> 117247835Skib#include <netinet6/tcp6_var.h> 118247835Skib#include <netinet/tcpip.h> 119247835Skib#ifdef TCPDEBUG 120247835Skib#include <netinet/tcp_debug.h> 121247835Skib#endif /* TCPDEBUG */ 122247835Skib#include <netinet/vinet.h> 123247835Skib 124247835Skib#include <machine/in_cksum.h> 125247835Skib 126247835Skibextern struct uma_zone *sack_hole_zone; 127247835Skib 128247835Skib#ifdef VIMAGE_GLOBALS 129247835Skibint tcp_do_sack; 130247835Skibint tcp_sack_maxholes; 131247835Skibint tcp_sack_globalmaxholes; 132247835Skibint tcp_sack_globalholes; 133247835Skib#endif 134247835Skib 135247835SkibSYSCTL_NODE(_net_inet_tcp, OID_AUTO, sack, CTLFLAG_RW, 0, "TCP SACK"); 136247835SkibSYSCTL_V_INT(V_NET, vnet_inet, _net_inet_tcp_sack, OID_AUTO, enable, 137247835Skib CTLFLAG_RW, tcp_do_sack, 0, "Enable/Disable TCP SACK support"); 138247835SkibTUNABLE_INT("net.inet.tcp.sack.enable", &tcp_do_sack); 139247835Skib 140247835SkibSYSCTL_V_INT(V_NET, vnet_inet, _net_inet_tcp_sack, OID_AUTO, maxholes, 141247835Skib CTLFLAG_RW, tcp_sack_maxholes, 0, 142247835Skib "Maximum number of TCP SACK holes allowed per connection"); 143247835Skib 144247835SkibSYSCTL_V_INT(V_NET, vnet_inet, _net_inet_tcp_sack, OID_AUTO, globalmaxholes, 145247835Skib CTLFLAG_RW, tcp_sack_globalmaxholes, 0, 146247835Skib "Global maximum number of TCP SACK holes"); 147247835Skib 148247835SkibSYSCTL_V_INT(V_NET, vnet_inet, _net_inet_tcp_sack, OID_AUTO, globalholes, 149247835Skib CTLFLAG_RD, tcp_sack_globalholes, 0, 150 "Global number of TCP SACK holes currently allocated"); 151 152/* 153 * This function is called upon receipt of new valid data (while not in 154 * header prediction mode), and it updates the ordered list of sacks. 155 */ 156void 157tcp_update_sack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end) 158{ 159 /* 160 * First reported block MUST be the most recent one. Subsequent 161 * blocks SHOULD be in the order in which they arrived at the 162 * receiver. These two conditions make the implementation fully 163 * compliant with RFC 2018. 164 */ 165 struct sackblk head_blk, saved_blks[MAX_SACK_BLKS]; 166 int num_head, num_saved, i; 167 168 INP_WLOCK_ASSERT(tp->t_inpcb); 169 170 /* Check arguments. */ 171 KASSERT(SEQ_LT(rcv_start, rcv_end), ("rcv_start < rcv_end")); 172 173 /* SACK block for the received segment. */ 174 head_blk.start = rcv_start; 175 head_blk.end = rcv_end; 176 177 /* 178 * Merge updated SACK blocks into head_blk, and save unchanged SACK 179 * blocks into saved_blks[]. num_saved will have the number of the 180 * saved SACK blocks. 181 */ 182 num_saved = 0; 183 for (i = 0; i < tp->rcv_numsacks; i++) { 184 tcp_seq start = tp->sackblks[i].start; 185 tcp_seq end = tp->sackblks[i].end; 186 if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) { 187 /* 188 * Discard this SACK block. 189 */ 190 } else if (SEQ_LEQ(head_blk.start, end) && 191 SEQ_GEQ(head_blk.end, start)) { 192 /* 193 * Merge this SACK block into head_blk. This SACK 194 * block itself will be discarded. 195 */ 196 if (SEQ_GT(head_blk.start, start)) 197 head_blk.start = start; 198 if (SEQ_LT(head_blk.end, end)) 199 head_blk.end = end; 200 } else { 201 /* 202 * Save this SACK block. 203 */ 204 saved_blks[num_saved].start = start; 205 saved_blks[num_saved].end = end; 206 num_saved++; 207 } 208 } 209 210 /* 211 * Update SACK list in tp->sackblks[]. 212 */ 213 num_head = 0; 214 if (SEQ_GT(head_blk.start, tp->rcv_nxt)) { 215 /* 216 * The received data segment is an out-of-order segment. Put 217 * head_blk at the top of SACK list. 218 */ 219 tp->sackblks[0] = head_blk; 220 num_head = 1; 221 /* 222 * If the number of saved SACK blocks exceeds its limit, 223 * discard the last SACK block. 224 */ 225 if (num_saved >= MAX_SACK_BLKS) 226 num_saved--; 227 } 228 if (num_saved > 0) { 229 /* 230 * Copy the saved SACK blocks back. 231 */ 232 bcopy(saved_blks, &tp->sackblks[num_head], 233 sizeof(struct sackblk) * num_saved); 234 } 235 236 /* Save the number of SACK blocks. */ 237 tp->rcv_numsacks = num_head + num_saved; 238} 239 240/* 241 * Delete all receiver-side SACK information. 242 */ 243void 244tcp_clean_sackreport(struct tcpcb *tp) 245{ 246 int i; 247 248 INP_WLOCK_ASSERT(tp->t_inpcb); 249 tp->rcv_numsacks = 0; 250 for (i = 0; i < MAX_SACK_BLKS; i++) 251 tp->sackblks[i].start = tp->sackblks[i].end=0; 252} 253 254/* 255 * Allocate struct sackhole. 256 */ 257static struct sackhole * 258tcp_sackhole_alloc(struct tcpcb *tp, tcp_seq start, tcp_seq end) 259{ 260 INIT_VNET_INET(tp->t_inpcb->inp_vnet); 261 struct sackhole *hole; 262 263 if (tp->snd_numholes >= V_tcp_sack_maxholes || 264 V_tcp_sack_globalholes >= V_tcp_sack_globalmaxholes) { 265 V_tcpstat.tcps_sack_sboverflow++; 266 return NULL; 267 } 268 269 hole = (struct sackhole *)uma_zalloc(sack_hole_zone, M_NOWAIT); 270 if (hole == NULL) 271 return NULL; 272 273 hole->start = start; 274 hole->end = end; 275 hole->rxmit = start; 276 277 tp->snd_numholes++; 278 V_tcp_sack_globalholes++; 279 280 return hole; 281} 282 283/* 284 * Free struct sackhole. 285 */ 286static void 287tcp_sackhole_free(struct tcpcb *tp, struct sackhole *hole) 288{ 289 INIT_VNET_INET(tp->t_vnet); 290 291 uma_zfree(sack_hole_zone, hole); 292 293 tp->snd_numholes--; 294 V_tcp_sack_globalholes--; 295 296 KASSERT(tp->snd_numholes >= 0, ("tp->snd_numholes >= 0")); 297 KASSERT(V_tcp_sack_globalholes >= 0, ("tcp_sack_globalholes >= 0")); 298} 299 300/* 301 * Insert new SACK hole into scoreboard. 302 */ 303static struct sackhole * 304tcp_sackhole_insert(struct tcpcb *tp, tcp_seq start, tcp_seq end, 305 struct sackhole *after) 306{ 307 struct sackhole *hole; 308 309 /* Allocate a new SACK hole. */ 310 hole = tcp_sackhole_alloc(tp, start, end); 311 if (hole == NULL) 312 return NULL; 313 314 /* Insert the new SACK hole into scoreboard. */ 315 if (after != NULL) 316 TAILQ_INSERT_AFTER(&tp->snd_holes, after, hole, scblink); 317 else 318 TAILQ_INSERT_TAIL(&tp->snd_holes, hole, scblink); 319 320 /* Update SACK hint. */ 321 if (tp->sackhint.nexthole == NULL) 322 tp->sackhint.nexthole = hole; 323 324 return hole; 325} 326 327/* 328 * Remove SACK hole from scoreboard. 329 */ 330static void 331tcp_sackhole_remove(struct tcpcb *tp, struct sackhole *hole) 332{ 333 334 /* Update SACK hint. */ 335 if (tp->sackhint.nexthole == hole) 336 tp->sackhint.nexthole = TAILQ_NEXT(hole, scblink); 337 338 /* Remove this SACK hole. */ 339 TAILQ_REMOVE(&tp->snd_holes, hole, scblink); 340 341 /* Free this SACK hole. */ 342 tcp_sackhole_free(tp, hole); 343} 344 345/* 346 * Process cumulative ACK and the TCP SACK option to update the scoreboard. 347 * tp->snd_holes is an ordered list of holes (oldest to newest, in terms of 348 * the sequence space). 349 */ 350void 351tcp_sack_doack(struct tcpcb *tp, struct tcpopt *to, tcp_seq th_ack) 352{ 353 struct sackhole *cur, *temp; 354 struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1], *sblkp; 355 int i, j, num_sack_blks; 356 357 INP_WLOCK_ASSERT(tp->t_inpcb); 358 359 num_sack_blks = 0; 360 /* 361 * If SND.UNA will be advanced by SEG.ACK, and if SACK holes exist, 362 * treat [SND.UNA, SEG.ACK) as if it is a SACK block. 363 */ 364 if (SEQ_LT(tp->snd_una, th_ack) && !TAILQ_EMPTY(&tp->snd_holes)) { 365 sack_blocks[num_sack_blks].start = tp->snd_una; 366 sack_blocks[num_sack_blks++].end = th_ack; 367 } 368 /* 369 * Append received valid SACK blocks to sack_blocks[], but only if we 370 * received new blocks from the other side. 371 */ 372 if (to->to_flags & TOF_SACK) { 373 for (i = 0; i < to->to_nsacks; i++) { 374 bcopy((to->to_sacks + i * TCPOLEN_SACK), 375 &sack, sizeof(sack)); 376 sack.start = ntohl(sack.start); 377 sack.end = ntohl(sack.end); 378 if (SEQ_GT(sack.end, sack.start) && 379 SEQ_GT(sack.start, tp->snd_una) && 380 SEQ_GT(sack.start, th_ack) && 381 SEQ_LT(sack.start, tp->snd_max) && 382 SEQ_GT(sack.end, tp->snd_una) && 383 SEQ_LEQ(sack.end, tp->snd_max)) 384 sack_blocks[num_sack_blks++] = sack; 385 } 386 } 387 /* 388 * Return if SND.UNA is not advanced and no valid SACK block is 389 * received. 390 */ 391 if (num_sack_blks == 0) 392 return; 393 394 /* 395 * Sort the SACK blocks so we can update the scoreboard with just one 396 * pass. The overhead of sorting upto 4+1 elements is less than 397 * making upto 4+1 passes over the scoreboard. 398 */ 399 for (i = 0; i < num_sack_blks; i++) { 400 for (j = i + 1; j < num_sack_blks; j++) { 401 if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) { 402 sack = sack_blocks[i]; 403 sack_blocks[i] = sack_blocks[j]; 404 sack_blocks[j] = sack; 405 } 406 } 407 } 408 if (TAILQ_EMPTY(&tp->snd_holes)) 409 /* 410 * Empty scoreboard. Need to initialize snd_fack (it may be 411 * uninitialized or have a bogus value). Scoreboard holes 412 * (from the sack blocks received) are created later below 413 * (in the logic that adds holes to the tail of the 414 * scoreboard). 415 */ 416 tp->snd_fack = SEQ_MAX(tp->snd_una, th_ack); 417 /* 418 * In the while-loop below, incoming SACK blocks (sack_blocks[]) and 419 * SACK holes (snd_holes) are traversed from their tails with just 420 * one pass in order to reduce the number of compares especially when 421 * the bandwidth-delay product is large. 422 * 423 * Note: Typically, in the first RTT of SACK recovery, the highest 424 * three or four SACK blocks with the same ack number are received. 425 * In the second RTT, if retransmitted data segments are not lost, 426 * the highest three or four SACK blocks with ack number advancing 427 * are received. 428 */ 429 sblkp = &sack_blocks[num_sack_blks - 1]; /* Last SACK block */ 430 if (SEQ_LT(tp->snd_fack, sblkp->start)) { 431 /* 432 * The highest SACK block is beyond fack. Append new SACK 433 * hole at the tail. If the second or later highest SACK 434 * blocks are also beyond the current fack, they will be 435 * inserted by way of hole splitting in the while-loop below. 436 */ 437 temp = tcp_sackhole_insert(tp, tp->snd_fack,sblkp->start,NULL); 438 if (temp != NULL) { 439 tp->snd_fack = sblkp->end; 440 /* Go to the previous sack block. */ 441 sblkp--; 442 } else { 443 /* 444 * We failed to add a new hole based on the current 445 * sack block. Skip over all the sack blocks that 446 * fall completely to the right of snd_fack and 447 * proceed to trim the scoreboard based on the 448 * remaining sack blocks. This also trims the 449 * scoreboard for th_ack (which is sack_blocks[0]). 450 */ 451 while (sblkp >= sack_blocks && 452 SEQ_LT(tp->snd_fack, sblkp->start)) 453 sblkp--; 454 if (sblkp >= sack_blocks && 455 SEQ_LT(tp->snd_fack, sblkp->end)) 456 tp->snd_fack = sblkp->end; 457 } 458 } else if (SEQ_LT(tp->snd_fack, sblkp->end)) 459 /* fack is advanced. */ 460 tp->snd_fack = sblkp->end; 461 /* We must have at least one SACK hole in scoreboard. */ 462 KASSERT(!TAILQ_EMPTY(&tp->snd_holes), 463 ("SACK scoreboard must not be empty")); 464 cur = TAILQ_LAST(&tp->snd_holes, sackhole_head); /* Last SACK hole. */ 465 /* 466 * Since the incoming sack blocks are sorted, we can process them 467 * making one sweep of the scoreboard. 468 */ 469 while (sblkp >= sack_blocks && cur != NULL) { 470 if (SEQ_GEQ(sblkp->start, cur->end)) { 471 /* 472 * SACKs data beyond the current hole. Go to the 473 * previous sack block. 474 */ 475 sblkp--; 476 continue; 477 } 478 if (SEQ_LEQ(sblkp->end, cur->start)) { 479 /* 480 * SACKs data before the current hole. Go to the 481 * previous hole. 482 */ 483 cur = TAILQ_PREV(cur, sackhole_head, scblink); 484 continue; 485 } 486 tp->sackhint.sack_bytes_rexmit -= (cur->rxmit - cur->start); 487 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, 488 ("sackhint bytes rtx >= 0")); 489 if (SEQ_LEQ(sblkp->start, cur->start)) { 490 /* Data acks at least the beginning of hole. */ 491 if (SEQ_GEQ(sblkp->end, cur->end)) { 492 /* Acks entire hole, so delete hole. */ 493 temp = cur; 494 cur = TAILQ_PREV(cur, sackhole_head, scblink); 495 tcp_sackhole_remove(tp, temp); 496 /* 497 * The sack block may ack all or part of the 498 * next hole too, so continue onto the next 499 * hole. 500 */ 501 continue; 502 } else { 503 /* Move start of hole forward. */ 504 cur->start = sblkp->end; 505 cur->rxmit = SEQ_MAX(cur->rxmit, cur->start); 506 } 507 } else { 508 /* Data acks at least the end of hole. */ 509 if (SEQ_GEQ(sblkp->end, cur->end)) { 510 /* Move end of hole backward. */ 511 cur->end = sblkp->start; 512 cur->rxmit = SEQ_MIN(cur->rxmit, cur->end); 513 } else { 514 /* 515 * ACKs some data in middle of a hole; need 516 * to split current hole 517 */ 518 temp = tcp_sackhole_insert(tp, sblkp->end, 519 cur->end, cur); 520 if (temp != NULL) { 521 if (SEQ_GT(cur->rxmit, temp->rxmit)) { 522 temp->rxmit = cur->rxmit; 523 tp->sackhint.sack_bytes_rexmit 524 += (temp->rxmit 525 - temp->start); 526 } 527 cur->end = sblkp->start; 528 cur->rxmit = SEQ_MIN(cur->rxmit, 529 cur->end); 530 } 531 } 532 } 533 tp->sackhint.sack_bytes_rexmit += (cur->rxmit - cur->start); 534 /* 535 * Testing sblkp->start against cur->start tells us whether 536 * we're done with the sack block or the sack hole. 537 * Accordingly, we advance one or the other. 538 */ 539 if (SEQ_LEQ(sblkp->start, cur->start)) 540 cur = TAILQ_PREV(cur, sackhole_head, scblink); 541 else 542 sblkp--; 543 } 544} 545 546/* 547 * Free all SACK holes to clear the scoreboard. 548 */ 549void 550tcp_free_sackholes(struct tcpcb *tp) 551{ 552 struct sackhole *q; 553 554 INP_WLOCK_ASSERT(tp->t_inpcb); 555 while ((q = TAILQ_FIRST(&tp->snd_holes)) != NULL) 556 tcp_sackhole_remove(tp, q); 557 tp->sackhint.sack_bytes_rexmit = 0; 558 559 KASSERT(tp->snd_numholes == 0, ("tp->snd_numholes == 0")); 560 KASSERT(tp->sackhint.nexthole == NULL, 561 ("tp->sackhint.nexthole == NULL")); 562} 563 564/* 565 * Partial ack handling within a sack recovery episode. Keeping this very 566 * simple for now. When a partial ack is received, force snd_cwnd to a value 567 * that will allow the sender to transmit no more than 2 segments. If 568 * necessary, a better scheme can be adopted at a later point, but for now, 569 * the goal is to prevent the sender from bursting a large amount of data in 570 * the midst of sack recovery. 571 */ 572void 573tcp_sack_partialack(struct tcpcb *tp, struct tcphdr *th) 574{ 575 int num_segs = 1; 576 577 INP_WLOCK_ASSERT(tp->t_inpcb); 578 tcp_timer_activate(tp, TT_REXMT, 0); 579 tp->t_rtttime = 0; 580 /* Send one or 2 segments based on how much new data was acked. */ 581 if (((th->th_ack - tp->snd_una) / tp->t_maxseg) > 2) 582 num_segs = 2; 583 tp->snd_cwnd = (tp->sackhint.sack_bytes_rexmit + 584 (tp->snd_nxt - tp->sack_newdata) + num_segs * tp->t_maxseg); 585 if (tp->snd_cwnd > tp->snd_ssthresh) 586 tp->snd_cwnd = tp->snd_ssthresh; 587 tp->t_flags |= TF_ACKNOW; 588 (void) tcp_output(tp); 589} 590 591#if 0 592/* 593 * Debug version of tcp_sack_output() that walks the scoreboard. Used for 594 * now to sanity check the hint. 595 */ 596static struct sackhole * 597tcp_sack_output_debug(struct tcpcb *tp, int *sack_bytes_rexmt) 598{ 599 struct sackhole *p; 600 601 INP_WLOCK_ASSERT(tp->t_inpcb); 602 *sack_bytes_rexmt = 0; 603 TAILQ_FOREACH(p, &tp->snd_holes, scblink) { 604 if (SEQ_LT(p->rxmit, p->end)) { 605 if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */ 606 continue; 607 } 608 *sack_bytes_rexmt += (p->rxmit - p->start); 609 break; 610 } 611 *sack_bytes_rexmt += (p->rxmit - p->start); 612 } 613 return (p); 614} 615#endif 616 617/* 618 * Returns the next hole to retransmit and the number of retransmitted bytes 619 * from the scoreboard. We store both the next hole and the number of 620 * retransmitted bytes as hints (and recompute these on the fly upon SACK/ACK 621 * reception). This avoids scoreboard traversals completely. 622 * 623 * The loop here will traverse *at most* one link. Here's the argument. For 624 * the loop to traverse more than 1 link before finding the next hole to 625 * retransmit, we would need to have at least 1 node following the current 626 * hint with (rxmit == end). But, for all holes following the current hint, 627 * (start == rxmit), since we have not yet retransmitted from them. 628 * Therefore, in order to traverse more 1 link in the loop below, we need to 629 * have at least one node following the current hint with (start == rxmit == 630 * end). But that can't happen, (start == end) means that all the data in 631 * that hole has been sacked, in which case, the hole would have been removed 632 * from the scoreboard. 633 */ 634struct sackhole * 635tcp_sack_output(struct tcpcb *tp, int *sack_bytes_rexmt) 636{ 637 struct sackhole *hole = NULL; 638 639 INP_WLOCK_ASSERT(tp->t_inpcb); 640 *sack_bytes_rexmt = tp->sackhint.sack_bytes_rexmit; 641 hole = tp->sackhint.nexthole; 642 if (hole == NULL || SEQ_LT(hole->rxmit, hole->end)) 643 goto out; 644 while ((hole = TAILQ_NEXT(hole, scblink)) != NULL) { 645 if (SEQ_LT(hole->rxmit, hole->end)) { 646 tp->sackhint.nexthole = hole; 647 break; 648 } 649 } 650out: 651 return (hole); 652} 653 654/* 655 * After a timeout, the SACK list may be rebuilt. This SACK information 656 * should be used to avoid retransmitting SACKed data. This function 657 * traverses the SACK list to see if snd_nxt should be moved forward. 658 */ 659void 660tcp_sack_adjust(struct tcpcb *tp) 661{ 662 struct sackhole *p, *cur = TAILQ_FIRST(&tp->snd_holes); 663 664 INP_WLOCK_ASSERT(tp->t_inpcb); 665 if (cur == NULL) 666 return; /* No holes */ 667 if (SEQ_GEQ(tp->snd_nxt, tp->snd_fack)) 668 return; /* We're already beyond any SACKed blocks */ 669 /*- 670 * Two cases for which we want to advance snd_nxt: 671 * i) snd_nxt lies between end of one hole and beginning of another 672 * ii) snd_nxt lies between end of last hole and snd_fack 673 */ 674 while ((p = TAILQ_NEXT(cur, scblink)) != NULL) { 675 if (SEQ_LT(tp->snd_nxt, cur->end)) 676 return; 677 if (SEQ_GEQ(tp->snd_nxt, p->start)) 678 cur = p; 679 else { 680 tp->snd_nxt = p->start; 681 return; 682 } 683 } 684 if (SEQ_LT(tp->snd_nxt, cur->end)) 685 return; 686 tp->snd_nxt = tp->snd_fack; 687} 688