1 /*- 2 * SPDX-License-Identifier: BSD-3-Clause 3 * 4 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995 5 * The Regents of the University of California. 6 * All rights reserved. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 3. Neither the name of the University nor the names of its contributors 17 * may be used to endorse or promote products derived from this software 18 * without specific prior written permission. 19 * 20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 30 * SUCH DAMAGE. 31 * 32 * @(#)tcp_sack.c 8.12 (Berkeley) 5/24/95 33 */ 34 35 /*- 36 * @@(#)COPYRIGHT 1.1 (NRL) 17 January 1995 37 * 38 * NRL grants permission for redistribution and use in source and binary 39 * forms, with or without modification, of the software and documentation 40 * created at NRL provided that the following conditions are met: 41 * 42 * 1. Redistributions of source code must retain the above copyright 43 * notice, this list of conditions and the following disclaimer. 44 * 2. Redistributions in binary form must reproduce the above copyright 45 * notice, this list of conditions and the following disclaimer in the 46 * documentation and/or other materials provided with the distribution. 47 * 3. All advertising materials mentioning features or use of this software 48 * must display the following acknowledgements: 49 * This product includes software developed by the University of 50 * California, Berkeley and its contributors. 51 * This product includes software developed at the Information 52 * Technology Division, US Naval Research Laboratory. 53 * 4. Neither the name of the NRL nor the names of its contributors 54 * may be used to endorse or promote products derived from this software 55 * without specific prior written permission. 56 * 57 * THE SOFTWARE PROVIDED BY NRL IS PROVIDED BY NRL AND CONTRIBUTORS ``AS 58 * IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 59 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A 60 * PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NRL OR 61 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, 62 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, 63 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR 64 * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF 65 * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING 66 * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS 67 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 68 * 69 * The views and conclusions contained in the software and documentation 70 * are those of the authors and should not be interpreted as representing 71 * official policies, either expressed or implied, of the US Naval 72 * Research Laboratory (NRL). 73 */ 74 75 #include <sys/cdefs.h> 76 __FBSDID("$FreeBSD$"); 77 78 #include "opt_inet.h" 79 #include "opt_inet6.h" 80 #include "opt_tcpdebug.h" 81 82 #include <sys/param.h> 83 #include <sys/systm.h> 84 #include <sys/kernel.h> 85 #include <sys/sysctl.h> 86 #include <sys/malloc.h> 87 #include <sys/mbuf.h> 88 #include <sys/proc.h> /* for proc0 declaration */ 89 #include <sys/protosw.h> 90 #include <sys/socket.h> 91 #include <sys/socketvar.h> 92 #include <sys/syslog.h> 93 #include <sys/systm.h> 94 95 #include <machine/cpu.h> /* before tcp_seq.h, for tcp_random18() */ 96 97 #include <vm/uma.h> 98 99 #include <net/if.h> 100 #include <net/if_var.h> 101 #include <net/route.h> 102 #include <net/vnet.h> 103 104 #include <netinet/in.h> 105 #include <netinet/in_systm.h> 106 #include <netinet/ip.h> 107 #include <netinet/in_var.h> 108 #include <netinet/in_pcb.h> 109 #include <netinet/ip_var.h> 110 #include <netinet/ip6.h> 111 #include <netinet/icmp6.h> 112 #include <netinet6/nd6.h> 113 #include <netinet6/ip6_var.h> 114 #include <netinet6/in6_pcb.h> 115 #include <netinet/tcp.h> 116 #include <netinet/tcp_fsm.h> 117 #include <netinet/tcp_seq.h> 118 #include <netinet/tcp_timer.h> 119 #include <netinet/tcp_var.h> 120 #include <netinet6/tcp6_var.h> 121 #include <netinet/tcpip.h> 122 #ifdef TCPDEBUG 123 #include <netinet/tcp_debug.h> 124 #endif /* TCPDEBUG */ 125 126 #include <machine/in_cksum.h> 127 128 VNET_DECLARE(struct uma_zone *, sack_hole_zone); 129 #define V_sack_hole_zone VNET(sack_hole_zone) 130 131 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, sack, CTLFLAG_RW, 0, "TCP SACK"); 132 VNET_DEFINE(int, tcp_do_sack) = 1; 133 #define V_tcp_do_sack VNET(tcp_do_sack) 134 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, enable, CTLFLAG_VNET | CTLFLAG_RW, 135 &VNET_NAME(tcp_do_sack), 0, "Enable/Disable TCP SACK support"); 136 137 VNET_DEFINE(int, tcp_sack_maxholes) = 128; 138 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, maxholes, CTLFLAG_VNET | CTLFLAG_RW, 139 &VNET_NAME(tcp_sack_maxholes), 0, 140 "Maximum number of TCP SACK holes allowed per connection"); 141 142 VNET_DEFINE(int, tcp_sack_globalmaxholes) = 65536; 143 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, globalmaxholes, CTLFLAG_VNET | CTLFLAG_RW, 144 &VNET_NAME(tcp_sack_globalmaxholes), 0, 145 "Global maximum number of TCP SACK holes"); 146 147 VNET_DEFINE(int, tcp_sack_globalholes) = 0; 148 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, globalholes, CTLFLAG_VNET | CTLFLAG_RD, 149 &VNET_NAME(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 */ 156 void 157 tcp_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_LEQ(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 /* 197 * |-| 198 * |---| merge 199 * 200 * |-| 201 * |---| merge 202 * 203 * |-----| 204 * |-| DSACK smaller 205 * 206 * |-| 207 * |-----| DSACK smaller 208 */ 209 if (head_blk.start == end) 210 head_blk.start = start; 211 else if (head_blk.end == start) 212 head_blk.end = end; 213 else { 214 if (SEQ_LT(head_blk.start, start)) { 215 tcp_seq temp = start; 216 start = head_blk.start; 217 head_blk.start = temp; 218 } 219 if (SEQ_GT(head_blk.end, end)) { 220 tcp_seq temp = end; 221 end = head_blk.end; 222 head_blk.end = temp; 223 } 224 if ((head_blk.start != start) || 225 (head_blk.end != end)) { 226 if ((num_saved >= 1) && 227 SEQ_GEQ(saved_blks[num_saved-1].start, start) && 228 SEQ_LEQ(saved_blks[num_saved-1].end, end)) 229 num_saved--; 230 saved_blks[num_saved].start = start; 231 saved_blks[num_saved].end = end; 232 num_saved++; 233 } 234 } 235 } else { 236 /* 237 * This block supercedes the prior block 238 */ 239 if ((num_saved >= 1) && 240 SEQ_GEQ(saved_blks[num_saved-1].start, start) && 241 SEQ_LEQ(saved_blks[num_saved-1].end, end)) 242 num_saved--; 243 /* 244 * Save this SACK block. 245 */ 246 saved_blks[num_saved].start = start; 247 saved_blks[num_saved].end = end; 248 num_saved++; 249 } 250 } 251 252 /* 253 * Update SACK list in tp->sackblks[]. 254 */ 255 num_head = 0; 256 if (SEQ_LT(rcv_start, rcv_end)) { 257 /* 258 * The received data segment is an out-of-order segment. Put 259 * head_blk at the top of SACK list. 260 */ 261 tp->sackblks[0] = head_blk; 262 num_head = 1; 263 /* 264 * If the number of saved SACK blocks exceeds its limit, 265 * discard the last SACK block. 266 */ 267 if (num_saved >= MAX_SACK_BLKS) 268 num_saved--; 269 } 270 if (num_saved > 0) { 271 /* 272 * Copy the saved SACK blocks back. 273 */ 274 bcopy(saved_blks, &tp->sackblks[num_head], 275 sizeof(struct sackblk) * num_saved); 276 } 277 278 /* Save the number of SACK blocks. */ 279 tp->rcv_numsacks = num_head + num_saved; 280 } 281 282 void 283 tcp_clean_dsack_blocks(struct tcpcb *tp) 284 { 285 struct sackblk saved_blks[MAX_SACK_BLKS]; 286 int num_saved, i; 287 288 INP_WLOCK_ASSERT(tp->t_inpcb); 289 /* 290 * Clean up any DSACK blocks that 291 * are in our queue of sack blocks. 292 * 293 */ 294 num_saved = 0; 295 for (i = 0; i < tp->rcv_numsacks; i++) { 296 tcp_seq start = tp->sackblks[i].start; 297 tcp_seq end = tp->sackblks[i].end; 298 if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) { 299 /* 300 * Discard this D-SACK block. 301 */ 302 continue; 303 } 304 /* 305 * Save this SACK block. 306 */ 307 saved_blks[num_saved].start = start; 308 saved_blks[num_saved].end = end; 309 num_saved++; 310 } 311 if (num_saved > 0) { 312 /* 313 * Copy the saved SACK blocks back. 314 */ 315 bcopy(saved_blks, &tp->sackblks[0], 316 sizeof(struct sackblk) * num_saved); 317 } 318 tp->rcv_numsacks = num_saved; 319 } 320 321 /* 322 * Delete all receiver-side SACK information. 323 */ 324 void 325 tcp_clean_sackreport(struct tcpcb *tp) 326 { 327 int i; 328 329 INP_WLOCK_ASSERT(tp->t_inpcb); 330 tp->rcv_numsacks = 0; 331 for (i = 0; i < MAX_SACK_BLKS; i++) 332 tp->sackblks[i].start = tp->sackblks[i].end=0; 333 } 334 335 /* 336 * Allocate struct sackhole. 337 */ 338 static struct sackhole * 339 tcp_sackhole_alloc(struct tcpcb *tp, tcp_seq start, tcp_seq end) 340 { 341 struct sackhole *hole; 342 343 if (tp->snd_numholes >= V_tcp_sack_maxholes || 344 V_tcp_sack_globalholes >= V_tcp_sack_globalmaxholes) { 345 TCPSTAT_INC(tcps_sack_sboverflow); 346 return NULL; 347 } 348 349 hole = (struct sackhole *)uma_zalloc(V_sack_hole_zone, M_NOWAIT); 350 if (hole == NULL) 351 return NULL; 352 353 hole->start = start; 354 hole->end = end; 355 hole->rxmit = start; 356 357 tp->snd_numholes++; 358 atomic_add_int(&V_tcp_sack_globalholes, 1); 359 360 return hole; 361 } 362 363 /* 364 * Free struct sackhole. 365 */ 366 static void 367 tcp_sackhole_free(struct tcpcb *tp, struct sackhole *hole) 368 { 369 370 uma_zfree(V_sack_hole_zone, hole); 371 372 tp->snd_numholes--; 373 atomic_subtract_int(&V_tcp_sack_globalholes, 1); 374 375 KASSERT(tp->snd_numholes >= 0, ("tp->snd_numholes >= 0")); 376 KASSERT(V_tcp_sack_globalholes >= 0, ("tcp_sack_globalholes >= 0")); 377 } 378 379 /* 380 * Insert new SACK hole into scoreboard. 381 */ 382 static struct sackhole * 383 tcp_sackhole_insert(struct tcpcb *tp, tcp_seq start, tcp_seq end, 384 struct sackhole *after) 385 { 386 struct sackhole *hole; 387 388 /* Allocate a new SACK hole. */ 389 hole = tcp_sackhole_alloc(tp, start, end); 390 if (hole == NULL) 391 return NULL; 392 393 /* Insert the new SACK hole into scoreboard. */ 394 if (after != NULL) 395 TAILQ_INSERT_AFTER(&tp->snd_holes, after, hole, scblink); 396 else 397 TAILQ_INSERT_TAIL(&tp->snd_holes, hole, scblink); 398 399 /* Update SACK hint. */ 400 if (tp->sackhint.nexthole == NULL) 401 tp->sackhint.nexthole = hole; 402 403 return hole; 404 } 405 406 /* 407 * Remove SACK hole from scoreboard. 408 */ 409 static void 410 tcp_sackhole_remove(struct tcpcb *tp, struct sackhole *hole) 411 { 412 413 /* Update SACK hint. */ 414 if (tp->sackhint.nexthole == hole) 415 tp->sackhint.nexthole = TAILQ_NEXT(hole, scblink); 416 417 /* Remove this SACK hole. */ 418 TAILQ_REMOVE(&tp->snd_holes, hole, scblink); 419 420 /* Free this SACK hole. */ 421 tcp_sackhole_free(tp, hole); 422 } 423 424 /* 425 * Process cumulative ACK and the TCP SACK option to update the scoreboard. 426 * tp->snd_holes is an ordered list of holes (oldest to newest, in terms of 427 * the sequence space). 428 * Returns 1 if incoming ACK has previously unknown SACK information, 429 * 0 otherwise. Note: We treat (snd_una, th_ack) as a sack block so any changes 430 * to that (i.e. left edge moving) would also be considered a change in SACK 431 * information which is slightly different than rfc6675. 432 */ 433 int 434 tcp_sack_doack(struct tcpcb *tp, struct tcpopt *to, tcp_seq th_ack) 435 { 436 struct sackhole *cur, *temp; 437 struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1], *sblkp; 438 int i, j, num_sack_blks, sack_changed; 439 440 INP_WLOCK_ASSERT(tp->t_inpcb); 441 442 num_sack_blks = 0; 443 sack_changed = 0; 444 /* 445 * If SND.UNA will be advanced by SEG.ACK, and if SACK holes exist, 446 * treat [SND.UNA, SEG.ACK) as if it is a SACK block. 447 */ 448 if (SEQ_LT(tp->snd_una, th_ack) && !TAILQ_EMPTY(&tp->snd_holes)) { 449 sack_blocks[num_sack_blks].start = tp->snd_una; 450 sack_blocks[num_sack_blks++].end = th_ack; 451 } 452 /* 453 * Append received valid SACK blocks to sack_blocks[], but only if we 454 * received new blocks from the other side. 455 */ 456 if (to->to_flags & TOF_SACK) { 457 tp->sackhint.sacked_bytes = 0; /* reset */ 458 for (i = 0; i < to->to_nsacks; i++) { 459 bcopy((to->to_sacks + i * TCPOLEN_SACK), 460 &sack, sizeof(sack)); 461 sack.start = ntohl(sack.start); 462 sack.end = ntohl(sack.end); 463 if (SEQ_GT(sack.end, sack.start) && 464 SEQ_GT(sack.start, tp->snd_una) && 465 SEQ_GT(sack.start, th_ack) && 466 SEQ_LT(sack.start, tp->snd_max) && 467 SEQ_GT(sack.end, tp->snd_una) && 468 SEQ_LEQ(sack.end, tp->snd_max)) { 469 sack_blocks[num_sack_blks++] = sack; 470 tp->sackhint.sacked_bytes += 471 (sack.end-sack.start); 472 } 473 } 474 } 475 /* 476 * Return if SND.UNA is not advanced and no valid SACK block is 477 * received. 478 */ 479 if (num_sack_blks == 0) 480 return (sack_changed); 481 482 /* 483 * Sort the SACK blocks so we can update the scoreboard with just one 484 * pass. The overhead of sorting up to 4+1 elements is less than 485 * making up to 4+1 passes over the scoreboard. 486 */ 487 for (i = 0; i < num_sack_blks; i++) { 488 for (j = i + 1; j < num_sack_blks; j++) { 489 if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) { 490 sack = sack_blocks[i]; 491 sack_blocks[i] = sack_blocks[j]; 492 sack_blocks[j] = sack; 493 } 494 } 495 } 496 if (TAILQ_EMPTY(&tp->snd_holes)) 497 /* 498 * Empty scoreboard. Need to initialize snd_fack (it may be 499 * uninitialized or have a bogus value). Scoreboard holes 500 * (from the sack blocks received) are created later below 501 * (in the logic that adds holes to the tail of the 502 * scoreboard). 503 */ 504 tp->snd_fack = SEQ_MAX(tp->snd_una, th_ack); 505 /* 506 * In the while-loop below, incoming SACK blocks (sack_blocks[]) and 507 * SACK holes (snd_holes) are traversed from their tails with just 508 * one pass in order to reduce the number of compares especially when 509 * the bandwidth-delay product is large. 510 * 511 * Note: Typically, in the first RTT of SACK recovery, the highest 512 * three or four SACK blocks with the same ack number are received. 513 * In the second RTT, if retransmitted data segments are not lost, 514 * the highest three or four SACK blocks with ack number advancing 515 * are received. 516 */ 517 sblkp = &sack_blocks[num_sack_blks - 1]; /* Last SACK block */ 518 tp->sackhint.last_sack_ack = sblkp->end; 519 if (SEQ_LT(tp->snd_fack, sblkp->start)) { 520 /* 521 * The highest SACK block is beyond fack. Append new SACK 522 * hole at the tail. If the second or later highest SACK 523 * blocks are also beyond the current fack, they will be 524 * inserted by way of hole splitting in the while-loop below. 525 */ 526 temp = tcp_sackhole_insert(tp, tp->snd_fack,sblkp->start,NULL); 527 if (temp != NULL) { 528 tp->snd_fack = sblkp->end; 529 /* Go to the previous sack block. */ 530 sblkp--; 531 sack_changed = 1; 532 } else { 533 /* 534 * We failed to add a new hole based on the current 535 * sack block. Skip over all the sack blocks that 536 * fall completely to the right of snd_fack and 537 * proceed to trim the scoreboard based on the 538 * remaining sack blocks. This also trims the 539 * scoreboard for th_ack (which is sack_blocks[0]). 540 */ 541 while (sblkp >= sack_blocks && 542 SEQ_LT(tp->snd_fack, sblkp->start)) 543 sblkp--; 544 if (sblkp >= sack_blocks && 545 SEQ_LT(tp->snd_fack, sblkp->end)) 546 tp->snd_fack = sblkp->end; 547 } 548 } else if (SEQ_LT(tp->snd_fack, sblkp->end)) { 549 /* fack is advanced. */ 550 tp->snd_fack = sblkp->end; 551 sack_changed = 1; 552 } 553 cur = TAILQ_LAST(&tp->snd_holes, sackhole_head); /* Last SACK hole. */ 554 /* 555 * Since the incoming sack blocks are sorted, we can process them 556 * making one sweep of the scoreboard. 557 */ 558 while (sblkp >= sack_blocks && cur != NULL) { 559 if (SEQ_GEQ(sblkp->start, cur->end)) { 560 /* 561 * SACKs data beyond the current hole. Go to the 562 * previous sack block. 563 */ 564 sblkp--; 565 continue; 566 } 567 if (SEQ_LEQ(sblkp->end, cur->start)) { 568 /* 569 * SACKs data before the current hole. Go to the 570 * previous hole. 571 */ 572 cur = TAILQ_PREV(cur, sackhole_head, scblink); 573 continue; 574 } 575 tp->sackhint.sack_bytes_rexmit -= (cur->rxmit - cur->start); 576 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0, 577 ("sackhint bytes rtx >= 0")); 578 sack_changed = 1; 579 if (SEQ_LEQ(sblkp->start, cur->start)) { 580 /* Data acks at least the beginning of hole. */ 581 if (SEQ_GEQ(sblkp->end, cur->end)) { 582 /* Acks entire hole, so delete hole. */ 583 temp = cur; 584 cur = TAILQ_PREV(cur, sackhole_head, scblink); 585 tcp_sackhole_remove(tp, temp); 586 /* 587 * The sack block may ack all or part of the 588 * next hole too, so continue onto the next 589 * hole. 590 */ 591 continue; 592 } else { 593 /* Move start of hole forward. */ 594 cur->start = sblkp->end; 595 cur->rxmit = SEQ_MAX(cur->rxmit, cur->start); 596 } 597 } else { 598 /* Data acks at least the end of hole. */ 599 if (SEQ_GEQ(sblkp->end, cur->end)) { 600 /* Move end of hole backward. */ 601 cur->end = sblkp->start; 602 cur->rxmit = SEQ_MIN(cur->rxmit, cur->end); 603 } else { 604 /* 605 * ACKs some data in middle of a hole; need 606 * to split current hole 607 */ 608 temp = tcp_sackhole_insert(tp, sblkp->end, 609 cur->end, cur); 610 if (temp != NULL) { 611 if (SEQ_GT(cur->rxmit, temp->rxmit)) { 612 temp->rxmit = cur->rxmit; 613 tp->sackhint.sack_bytes_rexmit 614 += (temp->rxmit 615 - temp->start); 616 } 617 cur->end = sblkp->start; 618 cur->rxmit = SEQ_MIN(cur->rxmit, 619 cur->end); 620 } 621 } 622 } 623 tp->sackhint.sack_bytes_rexmit += (cur->rxmit - cur->start); 624 /* 625 * Testing sblkp->start against cur->start tells us whether 626 * we're done with the sack block or the sack hole. 627 * Accordingly, we advance one or the other. 628 */ 629 if (SEQ_LEQ(sblkp->start, cur->start)) 630 cur = TAILQ_PREV(cur, sackhole_head, scblink); 631 else 632 sblkp--; 633 } 634 return (sack_changed); 635 } 636 637 /* 638 * Free all SACK holes to clear the scoreboard. 639 */ 640 void 641 tcp_free_sackholes(struct tcpcb *tp) 642 { 643 struct sackhole *q; 644 645 INP_WLOCK_ASSERT(tp->t_inpcb); 646 while ((q = TAILQ_FIRST(&tp->snd_holes)) != NULL) 647 tcp_sackhole_remove(tp, q); 648 tp->sackhint.sack_bytes_rexmit = 0; 649 650 KASSERT(tp->snd_numholes == 0, ("tp->snd_numholes == 0")); 651 KASSERT(tp->sackhint.nexthole == NULL, 652 ("tp->sackhint.nexthole == NULL")); 653 } 654 655 /* 656 * Partial ack handling within a sack recovery episode. Keeping this very 657 * simple for now. When a partial ack is received, force snd_cwnd to a value 658 * that will allow the sender to transmit no more than 2 segments. If 659 * necessary, a better scheme can be adopted at a later point, but for now, 660 * the goal is to prevent the sender from bursting a large amount of data in 661 * the midst of sack recovery. 662 */ 663 void 664 tcp_sack_partialack(struct tcpcb *tp, struct tcphdr *th) 665 { 666 int num_segs = 1; 667 668 INP_WLOCK_ASSERT(tp->t_inpcb); 669 tcp_timer_activate(tp, TT_REXMT, 0); 670 tp->t_rtttime = 0; 671 /* Send one or 2 segments based on how much new data was acked. */ 672 if ((BYTES_THIS_ACK(tp, th) / tp->t_maxseg) >= 2) 673 num_segs = 2; 674 tp->snd_cwnd = (tp->sackhint.sack_bytes_rexmit + 675 (tp->snd_nxt - tp->sack_newdata) + num_segs * tp->t_maxseg); 676 if (tp->snd_cwnd > tp->snd_ssthresh) 677 tp->snd_cwnd = tp->snd_ssthresh; 678 tp->t_flags |= TF_ACKNOW; 679 (void) tp->t_fb->tfb_tcp_output(tp); 680 } 681 682 #if 0 683 /* 684 * Debug version of tcp_sack_output() that walks the scoreboard. Used for 685 * now to sanity check the hint. 686 */ 687 static struct sackhole * 688 tcp_sack_output_debug(struct tcpcb *tp, int *sack_bytes_rexmt) 689 { 690 struct sackhole *p; 691 692 INP_WLOCK_ASSERT(tp->t_inpcb); 693 *sack_bytes_rexmt = 0; 694 TAILQ_FOREACH(p, &tp->snd_holes, scblink) { 695 if (SEQ_LT(p->rxmit, p->end)) { 696 if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */ 697 continue; 698 } 699 *sack_bytes_rexmt += (p->rxmit - p->start); 700 break; 701 } 702 *sack_bytes_rexmt += (p->rxmit - p->start); 703 } 704 return (p); 705 } 706 #endif 707 708 /* 709 * Returns the next hole to retransmit and the number of retransmitted bytes 710 * from the scoreboard. We store both the next hole and the number of 711 * retransmitted bytes as hints (and recompute these on the fly upon SACK/ACK 712 * reception). This avoids scoreboard traversals completely. 713 * 714 * The loop here will traverse *at most* one link. Here's the argument. For 715 * the loop to traverse more than 1 link before finding the next hole to 716 * retransmit, we would need to have at least 1 node following the current 717 * hint with (rxmit == end). But, for all holes following the current hint, 718 * (start == rxmit), since we have not yet retransmitted from them. 719 * Therefore, in order to traverse more 1 link in the loop below, we need to 720 * have at least one node following the current hint with (start == rxmit == 721 * end). But that can't happen, (start == end) means that all the data in 722 * that hole has been sacked, in which case, the hole would have been removed 723 * from the scoreboard. 724 */ 725 struct sackhole * 726 tcp_sack_output(struct tcpcb *tp, int *sack_bytes_rexmt) 727 { 728 struct sackhole *hole = NULL; 729 730 INP_WLOCK_ASSERT(tp->t_inpcb); 731 *sack_bytes_rexmt = tp->sackhint.sack_bytes_rexmit; 732 hole = tp->sackhint.nexthole; 733 if (hole == NULL || SEQ_LT(hole->rxmit, hole->end)) 734 goto out; 735 while ((hole = TAILQ_NEXT(hole, scblink)) != NULL) { 736 if (SEQ_LT(hole->rxmit, hole->end)) { 737 tp->sackhint.nexthole = hole; 738 break; 739 } 740 } 741 out: 742 return (hole); 743 } 744 745 /* 746 * After a timeout, the SACK list may be rebuilt. This SACK information 747 * should be used to avoid retransmitting SACKed data. This function 748 * traverses the SACK list to see if snd_nxt should be moved forward. 749 */ 750 void 751 tcp_sack_adjust(struct tcpcb *tp) 752 { 753 struct sackhole *p, *cur = TAILQ_FIRST(&tp->snd_holes); 754 755 INP_WLOCK_ASSERT(tp->t_inpcb); 756 if (cur == NULL) 757 return; /* No holes */ 758 if (SEQ_GEQ(tp->snd_nxt, tp->snd_fack)) 759 return; /* We're already beyond any SACKed blocks */ 760 /*- 761 * Two cases for which we want to advance snd_nxt: 762 * i) snd_nxt lies between end of one hole and beginning of another 763 * ii) snd_nxt lies between end of last hole and snd_fack 764 */ 765 while ((p = TAILQ_NEXT(cur, scblink)) != NULL) { 766 if (SEQ_LT(tp->snd_nxt, cur->end)) 767 return; 768 if (SEQ_GEQ(tp->snd_nxt, p->start)) 769 cur = p; 770 else { 771 tp->snd_nxt = p->start; 772 return; 773 } 774 } 775 if (SEQ_LT(tp->snd_nxt, cur->end)) 776 return; 777 tp->snd_nxt = tp->snd_fack; 778 } 779