sys/netinet/tcp_sack.c

   1 /*
   2  * Copyright (c) 2003, 2004 Jeffrey M. Hsu.  All rights reserved.
   3  * Copyright (c) 2003, 2004 The DragonFly Project.  All rights reserved.
   4  *
   5  * This code is derived from software contributed to The DragonFly Project
   6  * by Jeffrey M. Hsu.
   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 DragonFly Project nor the names of its
  17  *    contributors may be used to endorse or promote products derived
  18  *    from this software without specific, prior written permission.
  19  *
  20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  23  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  24  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  25  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  27  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  28  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  29  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  30  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  31  * SUCH DAMAGE.
  32  *
  33  * $DragonFly: src/sys/netinet/tcp_sack.c,v 1.5 2007/03/04 18:51:59 swildner Exp $
  34  */
  35
  36 #include <sys/param.h>
  37 #include <sys/systm.h>
  38 #include <sys/kernel.h>
  39 #include <sys/malloc.h>
  40 #include <sys/queue.h>
  41 #include <sys/thread.h>
  42 #include <sys/types.h>
  43 #include <sys/socket.h>
  44 #include <sys/socketvar.h>
  45
  46 #include <net/if.h>
  47
  48 #include <netinet/in.h>
  49 #include <netinet/in_systm.h>
  50 #include <netinet/ip.h>
  51 #include <netinet/in_var.h>
  52 #include <netinet/in_pcb.h>
  53 #include <netinet/ip_var.h>
  54 #include <netinet/tcp.h>
  55 #include <netinet/tcp_seq.h>
  56 #include <netinet/tcp_var.h>
  57
  58 #include <vm/vm_zone.h>
  59
  60 struct sackblock {
  61         tcp_seq                 sblk_start;
  62         tcp_seq                 sblk_end;
  63         TAILQ_ENTRY(sackblock)  sblk_list;
  64 };
  65
  66 #define MAXSAVEDBLOCKS  8                       /* per connection limit */
  67
  68 static void insert_block(struct scoreboard *scb, struct sackblock *newblock);
  69 static void update_lostseq(struct scoreboard *scb, tcp_seq snd_una,
  70                            u_int maxseg);
  71
  72 static vm_zone_t sackblock_zone;
  73
  74 /*
  75  * Initialize SACK module.
  76  */
  77 void
  78 tcp_sack_init(void)
  79 {
  80         sackblock_zone = zinit("sack blocks", sizeof(struct sackblock),
  81             maxsockets, ZONE_INTERRUPT, 0);
  82 }
  83
  84 /*
  85  * Per-tcpcb initialization.
  86  */
  87 void
  88 tcp_sack_tcpcb_init(struct tcpcb *tp)
  89 {
  90         struct scoreboard *scb = &tp->scb;
  91
  92         scb->nblocks = 0;
  93         TAILQ_INIT(&scb->sackblocks);
  94         scb->lastfound = NULL;
  95 }
  96
  97 /*
  98  * Find the SACK block containing or immediately preceding "seq".
  99  * The boolean result indicates whether the sequence is actually
 100  * contained in the SACK block.
 101  */
 102 static boolean_t
 103 sack_block_lookup(struct scoreboard *scb, tcp_seq seq, struct sackblock **sb)
 104 {
 105         struct sackblock *hint = scb->lastfound;
 106         struct sackblock *cur, *last, *prev;
 107
 108         if (TAILQ_EMPTY(&scb->sackblocks)) {
 109                 *sb = NULL;
 110                 return FALSE;
 111         }
 112
 113         if (hint == NULL) {
 114                 /* No hint.  Search from start to end. */
 115                 cur = TAILQ_FIRST(&scb->sackblocks);
 116                 last = NULL;
 117                 prev = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 118         } else  {
 119                 if (SEQ_GEQ(seq, hint->sblk_start)) {
 120                         /* Search from hint to end of list. */
 121                         cur = hint;
 122                         last = NULL;
 123                         prev = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 124                 } else {
 125                         /* Search from front of list to hint. */
 126                         cur = TAILQ_FIRST(&scb->sackblocks);
 127                         last = hint;
 128                         prev = TAILQ_PREV(hint, sackblock_list, sblk_list);
 129                 }
 130         }
 131
 132         do {
 133                 if (SEQ_GT(cur->sblk_end, seq)) {
 134                         if (SEQ_GEQ(seq, cur->sblk_start)) {
 135                                 *sb = scb->lastfound = cur;
 136                                 return TRUE;
 137                         } else {
 138                                 *sb = scb->lastfound =
 139                                     TAILQ_PREV(cur, sackblock_list, sblk_list);
 140                                 return FALSE;
 141                         }
 142                 }
 143                 cur = TAILQ_NEXT(cur, sblk_list);
 144         } while (cur != last);
 145
 146         *sb = scb->lastfound = prev;
 147         return FALSE;
 148 }
 149
 150 /*
 151  * Allocate a SACK block.
 152  */
 153 static __inline struct sackblock *
 154 alloc_sackblock(void)
 155 {
 156         return zalloc(sackblock_zone);
 157 }
 158
 159 /*
 160  * Free a SACK block.
 161  */
 162 static __inline void
 163 free_sackblock(struct sackblock *s)
 164 {
 165         zfree(sackblock_zone, s);
 166 }
 167
 168 /*
 169  * Free up SACK blocks for data that's been acked.
 170  */
 171 static void
 172 tcp_sack_ack_blocks(struct scoreboard *scb, tcp_seq th_ack)
 173 {
 174         struct sackblock *sb, *nb;
 175
 176         sb = TAILQ_FIRST(&scb->sackblocks);
 177         while (sb && SEQ_LEQ(sb->sblk_end, th_ack)) {
 178                 nb = TAILQ_NEXT(sb, sblk_list);
 179                 if (sb == scb->lastfound)
 180                         scb->lastfound = NULL;
 181                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 182                 free_sackblock(sb);
 183                 --scb->nblocks;
 184                 KASSERT(scb->nblocks >= 0,
 185                     ("SACK block count underflow: %d < 0", scb->nblocks));
 186                 sb = nb;
 187         }
 188         if (sb && SEQ_GT(th_ack, sb->sblk_start))
 189                 sb->sblk_start = th_ack;        /* other side reneged? XXX */
 190 }
 191
 192 /*
 193  * Delete and free SACK blocks saved in scoreboard.
 194  */
 195 void
 196 tcp_sack_cleanup(struct scoreboard *scb)
 197 {
 198         struct sackblock *sb, *nb;
 199
 200         TAILQ_FOREACH_MUTABLE(sb, &scb->sackblocks, sblk_list, nb) {
 201                 free_sackblock(sb);
 202                 --scb->nblocks;
 203         }
 204         KASSERT(scb->nblocks == 0,
 205             ("SACK block %d count not zero", scb->nblocks));
 206         TAILQ_INIT(&scb->sackblocks);
 207         scb->lastfound = NULL;
 208 }
 209
 210 /*
 211  * Returns      0 if not D-SACK block,
 212  *              1 if D-SACK,
 213  *              2 if duplicate of out-of-order D-SACK block.
 214  */
 215 int
 216 tcp_sack_ndsack_blocks(struct raw_sackblock *blocks, const int numblocks,
 217                        tcp_seq snd_una)
 218 {
 219         if (numblocks == 0)
 220                 return 0;
 221
 222         if (SEQ_LT(blocks[0].rblk_start, snd_una))
 223                 return 1;
 224
 225         /* block 0 inside block 1 */
 226         if (numblocks > 1 &&
 227             SEQ_GEQ(blocks[0].rblk_start, blocks[1].rblk_start) &&
 228             SEQ_LEQ(blocks[0].rblk_end, blocks[1].rblk_end))
 229                 return 2;
 230
 231         return 0;
 232 }
 233
 234 /*
 235  * Update scoreboard on new incoming ACK.
 236  */
 237 static void
 238 tcp_sack_add_blocks(struct tcpcb *tp, struct tcpopt *to)
 239 {
 240         const int numblocks = to->to_nsackblocks;
 241         struct raw_sackblock *blocks = to->to_sackblocks;
 242         struct scoreboard *scb = &tp->scb;
 243         struct sackblock *sb;
 244         int startblock;
 245         int i;
 246
 247         if (tcp_sack_ndsack_blocks(blocks, numblocks, tp->snd_una) > 0)
 248                 startblock = 1;
 249         else
 250                 startblock = 0;
 251
 252         for (i = startblock; i < numblocks; i++) {
 253                 struct raw_sackblock *newsackblock = &blocks[i];
 254
 255                 /* don't accept bad SACK blocks */
 256                 if (SEQ_GT(newsackblock->rblk_end, tp->snd_max))
 257                         break;          /* skip all other blocks */
 258
 259                 sb = alloc_sackblock();
 260                 if (sb == NULL)         /* do some sort of cleanup? XXX */
 261                         break;          /* just skip rest of blocks */
 262                 sb->sblk_start = newsackblock->rblk_start;
 263                 sb->sblk_end = newsackblock->rblk_end;
 264                 if (TAILQ_EMPTY(&scb->sackblocks)) {
 265                         KASSERT(scb->nblocks == 0, ("emply scb w/ blocks"));
 266                         scb->nblocks = 1;
 267                         TAILQ_INSERT_HEAD(&scb->sackblocks, sb, sblk_list);
 268                 } else {
 269                         insert_block(scb, sb);
 270                 }
 271         }
 272 }
 273
 274 void
 275 tcp_sack_update_scoreboard(struct tcpcb *tp, struct tcpopt *to)
 276 {
 277         struct scoreboard *scb = &tp->scb;
 278
 279         tcp_sack_ack_blocks(scb, tp->snd_una);
 280         tcp_sack_add_blocks(tp, to);
 281         update_lostseq(scb, tp->snd_una, tp->t_maxseg);
 282         if (SEQ_LT(tp->rexmt_high, tp->snd_una))
 283                 tp->rexmt_high = tp->snd_una;
 284 }
 285
 286 /*
 287  * Insert SACK block into sender's scoreboard.
 288  */
 289 static void
 290 insert_block(struct scoreboard *scb, struct sackblock *newblock)
 291 {
 292         struct sackblock *sb, *workingblock;
 293         boolean_t overlap_front;
 294
 295         KASSERT(scb->nblocks > 0, ("insert_block() called w/ no blocks"));
 296
 297         if (scb->nblocks == MAXSAVEDBLOCKS) {
 298                 /*
 299                  * Should try to kick out older blocks XXX JH
 300                  * May be able to coalesce with existing block.
 301                  * Or, go other way and free all blocks if we hit this limit.
 302                  */
 303                 free_sackblock(newblock);
 304                 return;
 305         }
 306         KASSERT(scb->nblocks < MAXSAVEDBLOCKS,
 307             ("too many SACK blocks %d", scb->nblocks));
 308
 309         overlap_front = sack_block_lookup(scb, newblock->sblk_start,  &sb);
 310
 311         if (sb == NULL) {
 312                 workingblock = newblock;
 313                 TAILQ_INSERT_HEAD(&scb->sackblocks, newblock, sblk_list);
 314                 ++scb->nblocks;
 315         } else {
 316                 if (overlap_front || sb->sblk_end == newblock->sblk_start) {
 317                         /* extend old block and discard new one */
 318                         workingblock = sb;
 319                         if (SEQ_GT(newblock->sblk_end, sb->sblk_end))
 320                                 sb->sblk_end = newblock->sblk_end;
 321                         free_sackblock(newblock);
 322                 } else {
 323                         workingblock = newblock;
 324                         TAILQ_INSERT_AFTER(&scb->sackblocks, sb, newblock,
 325                                            sblk_list);
 326                         ++scb->nblocks;
 327                 }
 328         }
 329
 330         /* Consolidate right-hand side. */
 331         sb = TAILQ_NEXT(workingblock, sblk_list);
 332         while (sb != NULL &&
 333             SEQ_GEQ(workingblock->sblk_end, sb->sblk_end)) {
 334                 struct sackblock *nextblock;
 335
 336                 nextblock = TAILQ_NEXT(sb, sblk_list);
 337                 /* Remove completely overlapped block */
 338                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 339                 free_sackblock(sb);
 340                 --scb->nblocks;
 341                 KASSERT(scb->nblocks > 0,
 342                     ("removed overlapped block: %d blocks left", scb->nblocks));
 343                 sb = nextblock;
 344         }
 345         if (sb != NULL &&
 346             SEQ_GEQ(workingblock->sblk_end, sb->sblk_start)) {
 347                 /* Extend new block to cover partially overlapped old block. */
 348                 workingblock->sblk_end = sb->sblk_end;
 349                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 350                 free_sackblock(sb);
 351                 --scb->nblocks;
 352                 KASSERT(scb->nblocks > 0,
 353                     ("removed partial right: %d blocks left", scb->nblocks));
 354         }
 355 }
 356
 357 #ifdef DEBUG_SACK_BLOCKS
 358 static void
 359 tcp_sack_dump_blocks(struct scoreboard *scb)
 360 {
 361         struct sackblock *sb;
 362
 363         kprintf("%d blocks:", scb->nblocks);
 364         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
 365                 kprintf(" [%u, %u)", sb->sblk_start, sb->sblk_end);
 366         kprintf("\n");
 367 }
 368 #else
 369 static __inline void
 370 tcp_sack_dump_blocks(struct scoreboard *scb)
 371 {
 372 }
 373 #endif
 374
 375 /*
 376  * Optimization to quickly determine which packets are lost.
 377  */
 378 static void
 379 update_lostseq(struct scoreboard *scb, tcp_seq snd_una, u_int maxseg)
 380 {
 381         struct sackblock *sb;
 382         int nsackblocks = 0;
 383         int bytes_sacked = 0;
 384
 385         sb = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 386         while (sb != NULL) {
 387                 ++nsackblocks;
 388                 bytes_sacked += sb->sblk_end - sb->sblk_start;
 389                 if (nsackblocks == tcprexmtthresh ||
 390                     bytes_sacked >= tcprexmtthresh * maxseg) {
 391                         scb->lostseq = sb->sblk_start;
 392                         return;
 393                 }
 394                 sb = TAILQ_PREV(sb, sackblock_list, sblk_list);
 395         }
 396         scb->lostseq = snd_una;
 397 }
 398
 399 /*
 400  * Return whether the given sequence number is considered lost.
 401  */
 402 static boolean_t
 403 scb_islost(struct scoreboard *scb, tcp_seq seqnum)
 404 {
 405         return SEQ_LT(seqnum, scb->lostseq);
 406 }
 407
 408 /*
 409  * True if at least "amount" has been SACKed.  Used by Early Retransmit.
 410  */
 411 boolean_t
 412 tcp_sack_has_sacked(struct scoreboard *scb, u_int amount)
 413 {
 414         struct sackblock *sb;
 415         int bytes_sacked = 0;
 416
 417         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
 418                 bytes_sacked += sb->sblk_end - sb->sblk_start;
 419                 if (bytes_sacked >= amount)
 420                         return TRUE;
 421         }
 422         return FALSE;
 423 }
 424
 425 /*
 426  * Number of bytes SACKed below seq.
 427  */
 428 int
 429 tcp_sack_bytes_below(struct scoreboard *scb, tcp_seq seq)
 430 {
 431         struct sackblock *sb;
 432         int bytes_sacked = 0;
 433
 434         sb = TAILQ_FIRST(&scb->sackblocks);
 435         while (sb && SEQ_GT(seq, sb->sblk_start)) {
 436                 bytes_sacked += seq_min(seq, sb->sblk_end) - sb->sblk_start;
 437                 sb = TAILQ_NEXT(sb, sblk_list);
 438         }
 439         return bytes_sacked;
 440 }
 441
 442 /*
 443  * Return estimate of the number of bytes outstanding in the network.
 444  */
 445 uint32_t
 446 tcp_sack_compute_pipe(struct tcpcb *tp)
 447 {
 448         struct scoreboard *scb = &tp->scb;
 449         struct sackblock *sb;
 450         int nlost, nretransmitted;
 451         tcp_seq end;
 452
 453         nlost = tp->snd_max - scb->lostseq;
 454         nretransmitted = tp->rexmt_high - tp->snd_una;
 455
 456         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
 457                 if (SEQ_LT(sb->sblk_start, tp->rexmt_high)) {
 458                         end = seq_min(sb->sblk_end, tp->rexmt_high);
 459                         nretransmitted -= end - sb->sblk_start;
 460                 }
 461                 if (SEQ_GEQ(sb->sblk_start, scb->lostseq))
 462                         nlost -= sb->sblk_end - sb->sblk_start;
 463         }
 464
 465         return (nlost + nretransmitted);
 466 }
 467
 468 /*
 469  * Return the sequence number and length of the next segment to transmit
 470  * when in Fast Recovery.
 471  */
 472 boolean_t
 473 tcp_sack_nextseg(struct tcpcb *tp, tcp_seq *nextrexmt, uint32_t *plen,
 474                  boolean_t *lostdup)
 475 {
 476         struct scoreboard *scb = &tp->scb;
 477         struct socket *so = tp->t_inpcb->inp_socket;
 478         struct sackblock *sb;
 479         const struct sackblock *lastblock =
 480             TAILQ_LAST(&scb->sackblocks, sackblock_list);
 481         tcp_seq torexmt;
 482         long len, off;
 483
 484         /* skip SACKed data */
 485         tcp_sack_skip_sacked(scb, &tp->rexmt_high);
 486
 487         /* Look for lost data. */
 488         torexmt = tp->rexmt_high;
 489         *lostdup = FALSE;
 490         if (lastblock != NULL) {
 491                 if (SEQ_LT(torexmt, lastblock->sblk_end) &&
 492                     scb_islost(scb, torexmt)) {
 493 sendunsacked:
 494                         *nextrexmt = torexmt;
 495                         /* If the left-hand edge has been SACKed, pull it in. */
 496                         if (sack_block_lookup(scb, torexmt + tp->t_maxseg, &sb))
 497                                 *plen = sb->sblk_start - torexmt;
 498                         else
 499                                 *plen = tp->t_maxseg;
 500                         return TRUE;
 501                 }
 502         }
 503
 504         /* See if unsent data available within send window. */
 505         off = tp->snd_max - tp->snd_una;
 506         len = (long) ulmin(so->so_snd.sb_cc, tp->snd_wnd) - off;
 507         if (len > 0) {
 508                 *nextrexmt = tp->snd_max;       /* Send new data. */
 509                 *plen = tp->t_maxseg;
 510                 return TRUE;
 511         }
 512
 513         /* We're less certain this data has been lost. */
 514         if (lastblock == NULL || SEQ_LT(torexmt, lastblock->sblk_end))
 515                 goto sendunsacked;
 516
 517         return FALSE;
 518 }
 519
 520 /*
 521  * Return the next sequence number higher than "*prexmt" that has
 522  * not been SACKed.
 523  */
 524 void
 525 tcp_sack_skip_sacked(struct scoreboard *scb, tcp_seq *prexmt)
 526 {
 527         struct sackblock *sb;
 528
 529         /* skip SACKed data */
 530         if (sack_block_lookup(scb, *prexmt, &sb))
 531                 *prexmt = sb->sblk_end;
 532 }
 533
 534 #ifdef later
 535 void
 536 tcp_sack_save_scoreboard(struct scoreboard *scb)
 537 {
 538         struct scoreboard *scb = &tp->scb;
 539
 540         scb->sackblocks_prev = scb->sackblocks;
 541         TAILQ_INIT(&scb->sackblocks);
 542 }
 543
 544 void
 545 tcp_sack_revert_scoreboard(struct scoreboard *scb, tcp_seq snd_una,
 546                            u_int maxseg)
 547 {
 548         struct sackblock *sb;
 549
 550         scb->sackblocks = scb->sackblocks_prev;
 551         scb->nblocks = 0;
 552         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
 553                 ++scb->nblocks;
 554         tcp_sack_ack_blocks(scb, snd_una);
 555         scb->lastfound = NULL;
 556 }
 557 #endif
 558
 559 #ifdef DEBUG_SACK_HISTORY
 560 static void
 561 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
 562 {
 563         int i;
 564         static int ndumped;
 565
 566         /* only need a couple of these to debug most problems */
 567         if (++ndumped > 900)
 568                 return;
 569
 570         kprintf("%s:\tnsackhistory %d: ", msg, tp->nsackhistory);
 571         for (i = 0; i < tp->nsackhistory; ++i)
 572                 kprintf("[%u, %u) ", tp->sackhistory[i].rblk_start,
 573                     tp->sackhistory[i].rblk_end);
 574         kprintf("\n");
 575 }
 576 #else
 577 static __inline void
 578 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
 579 {
 580 }
 581 #endif
 582
 583 /*
 584  * Remove old SACK blocks from the SACK history that have already been ACKed.
 585  */
 586 static void
 587 tcp_sack_ack_history(struct tcpcb *tp)
 588 {
 589         int i, nblocks, openslot;
 590
 591         tcp_sack_dump_history("before tcp_sack_ack_history", tp);
 592         nblocks = tp->nsackhistory;
 593         for (i = openslot = 0; i < nblocks; ++i) {
 594                 if (SEQ_LEQ(tp->sackhistory[i].rblk_end, tp->rcv_nxt)) {
 595                         --tp->nsackhistory;
 596                         continue;
 597                 }
 598                 if (SEQ_LT(tp->sackhistory[i].rblk_start, tp->rcv_nxt))
 599                         tp->sackhistory[i].rblk_start = tp->rcv_nxt;
 600                 if (i == openslot)
 601                         ++openslot;
 602                 else
 603                         tp->sackhistory[openslot++] = tp->sackhistory[i];
 604         }
 605         tcp_sack_dump_history("after tcp_sack_ack_history", tp);
 606         KASSERT(openslot == tp->nsackhistory,
 607             ("tcp_sack_ack_history miscounted: %d != %d",
 608             openslot, tp->nsackhistory));
 609 }
 610
 611 /*
 612  * Add or merge newblock into reported history.
 613  * Also remove or update SACK blocks that will be acked.
 614  */
 615 static void
 616 tcp_sack_update_reported_history(struct tcpcb *tp, tcp_seq start, tcp_seq end)
 617 {
 618         struct raw_sackblock copy[MAX_SACK_REPORT_BLOCKS];
 619         int i, cindex;
 620
 621         tcp_sack_dump_history("before tcp_sack_update_reported_history", tp);
 622         /*
 623          * Six cases:
 624          *      0) no overlap
 625          *      1) newblock == oldblock
 626          *      2) oldblock contains newblock
 627          *      3) newblock contains oldblock
 628          *      4) tail of oldblock overlaps or abuts start of newblock
 629          *      5) tail of newblock overlaps or abuts head of oldblock
 630          */
 631         for (i = cindex = 0; i < tp->nsackhistory; ++i) {
 632                 struct raw_sackblock *oldblock = &tp->sackhistory[i];
 633                 tcp_seq old_start = oldblock->rblk_start;
 634                 tcp_seq old_end = oldblock->rblk_end;
 635
 636                 if (SEQ_LT(end, old_start) || SEQ_GT(start, old_end)) {
 637                         /* Case 0:  no overlap.  Copy old block. */
 638                         copy[cindex++] = *oldblock;
 639                         continue;
 640                 }
 641
 642                 if (SEQ_GEQ(start, old_start) && SEQ_LEQ(end, old_end)) {
 643                         /* Cases 1 & 2.  Move block to front of history. */
 644                         int j;
 645
 646                         start = old_start;
 647                         end = old_end;
 648                         /* no need to check rest of blocks */
 649                         for (j = i + 1; j < tp->nsackhistory; ++j)
 650                                 copy[cindex++] = tp->sackhistory[j];
 651                         break;
 652                 }
 653
 654                 if (SEQ_GEQ(old_end, start) && SEQ_LT(old_start, start)) {
 655                         /* Case 4:  extend start of new block. */
 656                         start = old_start;
 657                 } else if (SEQ_GEQ(end, old_start) && SEQ_GT(old_end, end)) {
 658                         /* Case 5: extend end of new block */
 659                         end = old_end;
 660                 } else {
 661                         /* Case 3.  Delete old block by not copying it. */
 662                         KASSERT(SEQ_LEQ(start, old_start) &&
 663                                 SEQ_GEQ(end, old_end),
 664                             ("bad logic: old [%u, %u), new [%u, %u)",
 665                              old_start, old_end, start, end));
 666                 }
 667         }
 668
 669         /* insert new block */
 670         tp->sackhistory[0].rblk_start = start;
 671         tp->sackhistory[0].rblk_end = end;
 672         cindex = min(cindex, MAX_SACK_REPORT_BLOCKS - 1);
 673         for (i = 0; i < cindex; ++i)
 674                 tp->sackhistory[i + 1] = copy[i];
 675         tp->nsackhistory = cindex + 1;
 676         tcp_sack_dump_history("after tcp_sack_update_reported_history", tp);
 677 }
 678
 679 /*
 680  * Fill in SACK report to return to data sender.
 681  */
 682 void
 683 tcp_sack_fill_report(struct tcpcb *tp, u_char *opt, u_int *plen)
 684 {
 685         u_int optlen = *plen;
 686         uint32_t *lp = (uint32_t *)(opt + optlen);
 687         uint32_t *olp;
 688         tcp_seq hstart = tp->rcv_nxt, hend;
 689         int nblocks;
 690
 691         KASSERT(TCP_MAXOLEN - optlen >=
 692             TCPOLEN_SACK_ALIGNED + TCPOLEN_SACK_BLOCK,
 693             ("no room for SACK header and one block: optlen %d", optlen));
 694
 695         olp = lp++;
 696         optlen += TCPOLEN_SACK_ALIGNED;
 697
 698         tcp_sack_ack_history(tp);
 699         if (tp->reportblk.rblk_start != tp->reportblk.rblk_end) {
 700                 *lp++ = htonl(tp->reportblk.rblk_start);
 701                 *lp++ = htonl(tp->reportblk.rblk_end);
 702                 optlen += TCPOLEN_SACK_BLOCK;
 703                 hstart = tp->reportblk.rblk_start;
 704                 hend = tp->reportblk.rblk_end;
 705                 if (tp->t_flags & TF_ENCLOSESEG) {
 706                         KASSERT(TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK,
 707                             ("no room for enclosing SACK block: oplen %d",
 708                             optlen));
 709                         *lp++ = htonl(tp->encloseblk.rblk_start);
 710                         *lp++ = htonl(tp->encloseblk.rblk_end);
 711                         optlen += TCPOLEN_SACK_BLOCK;
 712                         hstart = tp->encloseblk.rblk_start;
 713                         hend = tp->encloseblk.rblk_end;
 714                 }
 715                 if (SEQ_GT(hstart, tp->rcv_nxt))
 716                         tcp_sack_update_reported_history(tp, hstart, hend);
 717         }
 718         if (tcp_do_smartsack && (tp->t_flags & TF_SACKLEFT)) {
 719                 /* Fill in from left!  Walk re-assembly queue. */
 720                 struct tseg_qent *q;
 721
 722                 q = LIST_FIRST(&tp->t_segq);
 723                 while (q != NULL &&
 724                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
 725                         *lp++ = htonl(q->tqe_th->th_seq);
 726                         *lp++ = htonl(q->tqe_th->th_seq + q->tqe_len);
 727                         optlen += TCPOLEN_SACK_BLOCK;
 728                         q = LIST_NEXT(q, tqe_q);
 729                 }
 730         } else {
 731                 int n = 0;
 732
 733                 /* Fill in SACK blocks from right side. */
 734                 while (n < tp->nsackhistory &&
 735                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
 736                         if (tp->sackhistory[n].rblk_start != hstart) {
 737                                 *lp++ = htonl(tp->sackhistory[n].rblk_start);
 738                                 *lp++ = htonl(tp->sackhistory[n].rblk_end);
 739                                 optlen += TCPOLEN_SACK_BLOCK;
 740                         }
 741                         ++n;
 742                 }
 743         }
 744         tp->reportblk.rblk_start = tp->reportblk.rblk_end;
 745         tp->t_flags &= ~(TF_DUPSEG | TF_ENCLOSESEG | TF_SACKLEFT);
 746         nblocks = (lp - olp - 1) / 2;
 747         *olp = htonl(TCPOPT_SACK_ALIGNED |
 748                      (TCPOLEN_SACK + nblocks * TCPOLEN_SACK_BLOCK));
 749         *plen = optlen;
 750 }