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 1.7 2008/02/03 21:40:42 dillon 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 (scb->lastfound == sb)
 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                 if (scb->lastfound == sb)
 338                         scb->lastfound = NULL;
 339                 /* Remove completely overlapped block */
 340                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 341                 free_sackblock(sb);
 342                 --scb->nblocks;
 343                 KASSERT(scb->nblocks > 0,
 344                     ("removed overlapped block: %d blocks left", scb->nblocks));
 345                 sb = nextblock;
 346         }
 347         if (sb != NULL &&
 348             SEQ_GEQ(workingblock->sblk_end, sb->sblk_start)) {
 349                 /* Extend new block to cover partially overlapped old block. */
 350                 workingblock->sblk_end = sb->sblk_end;
 351                 if (scb->lastfound == sb)
 352                         scb->lastfound = NULL;
 353                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 354                 free_sackblock(sb);
 355                 --scb->nblocks;
 356                 KASSERT(scb->nblocks > 0,
 357                     ("removed partial right: %d blocks left", scb->nblocks));
 358         }
 359 }
 360
 361 #ifdef DEBUG_SACK_BLOCKS
 362 static void
 363 tcp_sack_dump_blocks(struct scoreboard *scb)
 364 {
 365         struct sackblock *sb;
 366
 367         kprintf("%d blocks:", scb->nblocks);
 368         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
 369                 kprintf(" [%u, %u)", sb->sblk_start, sb->sblk_end);
 370         kprintf("\n");
 371 }
 372 #else
 373 static __inline void
 374 tcp_sack_dump_blocks(struct scoreboard *scb)
 375 {
 376 }
 377 #endif
 378
 379 /*
 380  * Optimization to quickly determine which packets are lost.
 381  */
 382 static void
 383 update_lostseq(struct scoreboard *scb, tcp_seq snd_una, u_int maxseg)
 384 {
 385         struct sackblock *sb;
 386         int nsackblocks = 0;
 387         int bytes_sacked = 0;
 388
 389         sb = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 390         while (sb != NULL) {
 391                 ++nsackblocks;
 392                 bytes_sacked += sb->sblk_end - sb->sblk_start;
 393                 if (nsackblocks == tcprexmtthresh ||
 394                     bytes_sacked >= tcprexmtthresh * maxseg) {
 395                         scb->lostseq = sb->sblk_start;
 396                         return;
 397                 }
 398                 sb = TAILQ_PREV(sb, sackblock_list, sblk_list);
 399         }
 400         scb->lostseq = snd_una;
 401 }
 402
 403 /*
 404  * Return whether the given sequence number is considered lost.
 405  */
 406 static boolean_t
 407 scb_islost(struct scoreboard *scb, tcp_seq seqnum)
 408 {
 409         return SEQ_LT(seqnum, scb->lostseq);
 410 }
 411
 412 /*
 413  * True if at least "amount" has been SACKed.  Used by Early Retransmit.
 414  */
 415 boolean_t
 416 tcp_sack_has_sacked(struct scoreboard *scb, u_int amount)
 417 {
 418         struct sackblock *sb;
 419         int bytes_sacked = 0;
 420
 421         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
 422                 bytes_sacked += sb->sblk_end - sb->sblk_start;
 423                 if (bytes_sacked >= amount)
 424                         return TRUE;
 425         }
 426         return FALSE;
 427 }
 428
 429 /*
 430  * Number of bytes SACKed below seq.
 431  */
 432 int
 433 tcp_sack_bytes_below(struct scoreboard *scb, tcp_seq seq)
 434 {
 435         struct sackblock *sb;
 436         int bytes_sacked = 0;
 437
 438         sb = TAILQ_FIRST(&scb->sackblocks);
 439         while (sb && SEQ_GT(seq, sb->sblk_start)) {
 440                 bytes_sacked += seq_min(seq, sb->sblk_end) - sb->sblk_start;
 441                 sb = TAILQ_NEXT(sb, sblk_list);
 442         }
 443         return bytes_sacked;
 444 }
 445
 446 /*
 447  * Return estimate of the number of bytes outstanding in the network.
 448  */
 449 uint32_t
 450 tcp_sack_compute_pipe(struct tcpcb *tp)
 451 {
 452         struct scoreboard *scb = &tp->scb;
 453         struct sackblock *sb;
 454         int nlost, nretransmitted;
 455         tcp_seq end;
 456
 457         nlost = tp->snd_max - scb->lostseq;
 458         nretransmitted = tp->rexmt_high - tp->snd_una;
 459
 460         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
 461                 if (SEQ_LT(sb->sblk_start, tp->rexmt_high)) {
 462                         end = seq_min(sb->sblk_end, tp->rexmt_high);
 463                         nretransmitted -= end - sb->sblk_start;
 464                 }
 465                 if (SEQ_GEQ(sb->sblk_start, scb->lostseq))
 466                         nlost -= sb->sblk_end - sb->sblk_start;
 467         }
 468
 469         return (nlost + nretransmitted);
 470 }
 471
 472 /*
 473  * Return the sequence number and length of the next segment to transmit
 474  * when in Fast Recovery.
 475  */
 476 boolean_t
 477 tcp_sack_nextseg(struct tcpcb *tp, tcp_seq *nextrexmt, uint32_t *plen,
 478                  boolean_t *lostdup)
 479 {
 480         struct scoreboard *scb = &tp->scb;
 481         struct socket *so = tp->t_inpcb->inp_socket;
 482         struct sackblock *sb;
 483         const struct sackblock *lastblock =
 484             TAILQ_LAST(&scb->sackblocks, sackblock_list);
 485         tcp_seq torexmt;
 486         long len, off;
 487
 488         /* skip SACKed data */
 489         tcp_sack_skip_sacked(scb, &tp->rexmt_high);
 490
 491         /* Look for lost data. */
 492         torexmt = tp->rexmt_high;
 493         *lostdup = FALSE;
 494         if (lastblock != NULL) {
 495                 if (SEQ_LT(torexmt, lastblock->sblk_end) &&
 496                     scb_islost(scb, torexmt)) {
 497 sendunsacked:
 498                         *nextrexmt = torexmt;
 499                         /* If the left-hand edge has been SACKed, pull it in. */
 500                         if (sack_block_lookup(scb, torexmt + tp->t_maxseg, &sb))
 501                                 *plen = sb->sblk_start - torexmt;
 502                         else
 503                                 *plen = tp->t_maxseg;
 504                         return TRUE;
 505                 }
 506         }
 507
 508         /* See if unsent data available within send window. */
 509         off = tp->snd_max - tp->snd_una;
 510         len = (long) ulmin(so->so_snd.ssb_cc, tp->snd_wnd) - off;
 511         if (len > 0) {
 512                 *nextrexmt = tp->snd_max;       /* Send new data. */
 513                 *plen = tp->t_maxseg;
 514                 return TRUE;
 515         }
 516
 517         /* We're less certain this data has been lost. */
 518         if (lastblock == NULL || SEQ_LT(torexmt, lastblock->sblk_end))
 519                 goto sendunsacked;
 520
 521         return FALSE;
 522 }
 523
 524 /*
 525  * Return the next sequence number higher than "*prexmt" that has
 526  * not been SACKed.
 527  */
 528 void
 529 tcp_sack_skip_sacked(struct scoreboard *scb, tcp_seq *prexmt)
 530 {
 531         struct sackblock *sb;
 532
 533         /* skip SACKed data */
 534         if (sack_block_lookup(scb, *prexmt, &sb))
 535                 *prexmt = sb->sblk_end;
 536 }
 537
 538 #ifdef later
 539 void
 540 tcp_sack_save_scoreboard(struct scoreboard *scb)
 541 {
 542         struct scoreboard *scb = &tp->scb;
 543
 544         scb->sackblocks_prev = scb->sackblocks;
 545         TAILQ_INIT(&scb->sackblocks);
 546 }
 547
 548 void
 549 tcp_sack_revert_scoreboard(struct scoreboard *scb, tcp_seq snd_una,
 550                            u_int maxseg)
 551 {
 552         struct sackblock *sb;
 553
 554         scb->sackblocks = scb->sackblocks_prev;
 555         scb->nblocks = 0;
 556         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
 557                 ++scb->nblocks;
 558         tcp_sack_ack_blocks(scb, snd_una);
 559         scb->lastfound = NULL;
 560 }
 561 #endif
 562
 563 #ifdef DEBUG_SACK_HISTORY
 564 static void
 565 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
 566 {
 567         int i;
 568         static int ndumped;
 569
 570         /* only need a couple of these to debug most problems */
 571         if (++ndumped > 900)
 572                 return;
 573
 574         kprintf("%s:\tnsackhistory %d: ", msg, tp->nsackhistory);
 575         for (i = 0; i < tp->nsackhistory; ++i)
 576                 kprintf("[%u, %u) ", tp->sackhistory[i].rblk_start,
 577                     tp->sackhistory[i].rblk_end);
 578         kprintf("\n");
 579 }
 580 #else
 581 static __inline void
 582 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
 583 {
 584 }
 585 #endif
 586
 587 /*
 588  * Remove old SACK blocks from the SACK history that have already been ACKed.
 589  */
 590 static void
 591 tcp_sack_ack_history(struct tcpcb *tp)
 592 {
 593         int i, nblocks, openslot;
 594
 595         tcp_sack_dump_history("before tcp_sack_ack_history", tp);
 596         nblocks = tp->nsackhistory;
 597         for (i = openslot = 0; i < nblocks; ++i) {
 598                 if (SEQ_LEQ(tp->sackhistory[i].rblk_end, tp->rcv_nxt)) {
 599                         --tp->nsackhistory;
 600                         continue;
 601                 }
 602                 if (SEQ_LT(tp->sackhistory[i].rblk_start, tp->rcv_nxt))
 603                         tp->sackhistory[i].rblk_start = tp->rcv_nxt;
 604                 if (i == openslot)
 605                         ++openslot;
 606                 else
 607                         tp->sackhistory[openslot++] = tp->sackhistory[i];
 608         }
 609         tcp_sack_dump_history("after tcp_sack_ack_history", tp);
 610         KASSERT(openslot == tp->nsackhistory,
 611             ("tcp_sack_ack_history miscounted: %d != %d",
 612             openslot, tp->nsackhistory));
 613 }
 614
 615 /*
 616  * Add or merge newblock into reported history.
 617  * Also remove or update SACK blocks that will be acked.
 618  */
 619 static void
 620 tcp_sack_update_reported_history(struct tcpcb *tp, tcp_seq start, tcp_seq end)
 621 {
 622         struct raw_sackblock copy[MAX_SACK_REPORT_BLOCKS];
 623         int i, cindex;
 624
 625         tcp_sack_dump_history("before tcp_sack_update_reported_history", tp);
 626         /*
 627          * Six cases:
 628          *      0) no overlap
 629          *      1) newblock == oldblock
 630          *      2) oldblock contains newblock
 631          *      3) newblock contains oldblock
 632          *      4) tail of oldblock overlaps or abuts start of newblock
 633          *      5) tail of newblock overlaps or abuts head of oldblock
 634          */
 635         for (i = cindex = 0; i < tp->nsackhistory; ++i) {
 636                 struct raw_sackblock *oldblock = &tp->sackhistory[i];
 637                 tcp_seq old_start = oldblock->rblk_start;
 638                 tcp_seq old_end = oldblock->rblk_end;
 639
 640                 if (SEQ_LT(end, old_start) || SEQ_GT(start, old_end)) {
 641                         /* Case 0:  no overlap.  Copy old block. */
 642                         copy[cindex++] = *oldblock;
 643                         continue;
 644                 }
 645
 646                 if (SEQ_GEQ(start, old_start) && SEQ_LEQ(end, old_end)) {
 647                         /* Cases 1 & 2.  Move block to front of history. */
 648                         int j;
 649
 650                         start = old_start;
 651                         end = old_end;
 652                         /* no need to check rest of blocks */
 653                         for (j = i + 1; j < tp->nsackhistory; ++j)
 654                                 copy[cindex++] = tp->sackhistory[j];
 655                         break;
 656                 }
 657
 658                 if (SEQ_GEQ(old_end, start) && SEQ_LT(old_start, start)) {
 659                         /* Case 4:  extend start of new block. */
 660                         start = old_start;
 661                 } else if (SEQ_GEQ(end, old_start) && SEQ_GT(old_end, end)) {
 662                         /* Case 5: extend end of new block */
 663                         end = old_end;
 664                 } else {
 665                         /* Case 3.  Delete old block by not copying it. */
 666                         KASSERT(SEQ_LEQ(start, old_start) &&
 667                                 SEQ_GEQ(end, old_end),
 668                             ("bad logic: old [%u, %u), new [%u, %u)",
 669                              old_start, old_end, start, end));
 670                 }
 671         }
 672
 673         /* insert new block */
 674         tp->sackhistory[0].rblk_start = start;
 675         tp->sackhistory[0].rblk_end = end;
 676         cindex = min(cindex, MAX_SACK_REPORT_BLOCKS - 1);
 677         for (i = 0; i < cindex; ++i)
 678                 tp->sackhistory[i + 1] = copy[i];
 679         tp->nsackhistory = cindex + 1;
 680         tcp_sack_dump_history("after tcp_sack_update_reported_history", tp);
 681 }
 682
 683 /*
 684  * Fill in SACK report to return to data sender.
 685  */
 686 void
 687 tcp_sack_fill_report(struct tcpcb *tp, u_char *opt, u_int *plen)
 688 {
 689         u_int optlen = *plen;
 690         uint32_t *lp = (uint32_t *)(opt + optlen);
 691         uint32_t *olp;
 692         tcp_seq hstart = tp->rcv_nxt, hend;
 693         int nblocks;
 694
 695         KASSERT(TCP_MAXOLEN - optlen >=
 696             TCPOLEN_SACK_ALIGNED + TCPOLEN_SACK_BLOCK,
 697             ("no room for SACK header and one block: optlen %d", optlen));
 698
 699         olp = lp++;
 700         optlen += TCPOLEN_SACK_ALIGNED;
 701
 702         tcp_sack_ack_history(tp);
 703         if (tp->reportblk.rblk_start != tp->reportblk.rblk_end) {
 704                 *lp++ = htonl(tp->reportblk.rblk_start);
 705                 *lp++ = htonl(tp->reportblk.rblk_end);
 706                 optlen += TCPOLEN_SACK_BLOCK;
 707                 hstart = tp->reportblk.rblk_start;
 708                 hend = tp->reportblk.rblk_end;
 709                 if (tp->t_flags & TF_ENCLOSESEG) {
 710                         KASSERT(TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK,
 711                             ("no room for enclosing SACK block: oplen %d",
 712                             optlen));
 713                         *lp++ = htonl(tp->encloseblk.rblk_start);
 714                         *lp++ = htonl(tp->encloseblk.rblk_end);
 715                         optlen += TCPOLEN_SACK_BLOCK;
 716                         hstart = tp->encloseblk.rblk_start;
 717                         hend = tp->encloseblk.rblk_end;
 718                 }
 719                 if (SEQ_GT(hstart, tp->rcv_nxt))
 720                         tcp_sack_update_reported_history(tp, hstart, hend);
 721         }
 722         if (tcp_do_smartsack && (tp->t_flags & TF_SACKLEFT)) {
 723                 /* Fill in from left!  Walk re-assembly queue. */
 724                 struct tseg_qent *q;
 725
 726                 q = LIST_FIRST(&tp->t_segq);
 727                 while (q != NULL &&
 728                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
 729                         *lp++ = htonl(q->tqe_th->th_seq);
 730                         *lp++ = htonl(q->tqe_th->th_seq + q->tqe_len);
 731                         optlen += TCPOLEN_SACK_BLOCK;
 732                         q = LIST_NEXT(q, tqe_q);
 733                 }
 734         } else {
 735                 int n = 0;
 736
 737                 /* Fill in SACK blocks from right side. */
 738                 while (n < tp->nsackhistory &&
 739                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
 740                         if (tp->sackhistory[n].rblk_start != hstart) {
 741                                 *lp++ = htonl(tp->sackhistory[n].rblk_start);
 742                                 *lp++ = htonl(tp->sackhistory[n].rblk_end);
 743                                 optlen += TCPOLEN_SACK_BLOCK;
 744                         }
 745                         ++n;
 746                 }
 747         }
 748         tp->reportblk.rblk_start = tp->reportblk.rblk_end;
 749         tp->t_flags &= ~(TF_DUPSEG | TF_ENCLOSESEG | TF_SACKLEFT);
 750         nblocks = (lp - olp - 1) / 2;
 751         *olp = htonl(TCPOPT_SACK_ALIGNED |
 752                      (TCPOLEN_SACK + nblocks * TCPOLEN_SACK_BLOCK));
 753         *plen = optlen;
 754 }