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 1.8 2008/08/15 21:37:16 nth 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 struct sackblock {
  59         tcp_seq                 sblk_start;
  60         tcp_seq                 sblk_end;
  61         TAILQ_ENTRY(sackblock)  sblk_list;
  62 };
  63
  64 #define MAXSAVEDBLOCKS  8                       /* per connection limit */
  65
  66 static void insert_block(struct scoreboard *scb, struct sackblock *newblock);
  67 static void update_lostseq(struct scoreboard *scb, tcp_seq snd_una,
  68                            u_int maxseg);
  69
  70 static MALLOC_DEFINE(M_SACKBLOCK, "sblk", "sackblock struct");
  71
  72 /*
  73  * Per-tcpcb initialization.
  74  */
  75 void
  76 tcp_sack_tcpcb_init(struct tcpcb *tp)
  77 {
  78         struct scoreboard *scb = &tp->scb;
  79
  80         scb->nblocks = 0;
  81         TAILQ_INIT(&scb->sackblocks);
  82         scb->lastfound = NULL;
  83 }
  84
  85 /*
  86  * Find the SACK block containing or immediately preceding "seq".
  87  * The boolean result indicates whether the sequence is actually
  88  * contained in the SACK block.
  89  */
  90 static boolean_t
  91 sack_block_lookup(struct scoreboard *scb, tcp_seq seq, struct sackblock **sb)
  92 {
  93         struct sackblock *hint = scb->lastfound;
  94         struct sackblock *cur, *last, *prev;
  95
  96         if (TAILQ_EMPTY(&scb->sackblocks)) {
  97                 *sb = NULL;
  98                 return FALSE;
  99         }
 100
 101         if (hint == NULL) {
 102                 /* No hint.  Search from start to end. */
 103                 cur = TAILQ_FIRST(&scb->sackblocks);
 104                 last = NULL;
 105                 prev = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 106         } else  {
 107                 if (SEQ_GEQ(seq, hint->sblk_start)) {
 108                         /* Search from hint to end of list. */
 109                         cur = hint;
 110                         last = NULL;
 111                         prev = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 112                 } else {
 113                         /* Search from front of list to hint. */
 114                         cur = TAILQ_FIRST(&scb->sackblocks);
 115                         last = hint;
 116                         prev = TAILQ_PREV(hint, sackblock_list, sblk_list);
 117                 }
 118         }
 119
 120         do {
 121                 if (SEQ_GT(cur->sblk_end, seq)) {
 122                         if (SEQ_GEQ(seq, cur->sblk_start)) {
 123                                 *sb = scb->lastfound = cur;
 124                                 return TRUE;
 125                         } else {
 126                                 *sb = scb->lastfound =
 127                                     TAILQ_PREV(cur, sackblock_list, sblk_list);
 128                                 return FALSE;
 129                         }
 130                 }
 131                 cur = TAILQ_NEXT(cur, sblk_list);
 132         } while (cur != last);
 133
 134         *sb = scb->lastfound = prev;
 135         return FALSE;
 136 }
 137
 138 /*
 139  * Allocate a SACK block.
 140  */
 141 static __inline struct sackblock *
 142 alloc_sackblock(void)
 143 {
 144         return (kmalloc(sizeof(struct sackblock), M_SACKBLOCK, M_NOWAIT));
 145 }
 146
 147 /*
 148  * Free a SACK block.
 149  */
 150 static __inline void
 151 free_sackblock(struct sackblock *s)
 152 {
 153         kfree(s, M_SACKBLOCK);
 154 }
 155
 156 /*
 157  * Free up SACK blocks for data that's been acked.
 158  */
 159 static void
 160 tcp_sack_ack_blocks(struct scoreboard *scb, tcp_seq th_ack)
 161 {
 162         struct sackblock *sb, *nb;
 163
 164         sb = TAILQ_FIRST(&scb->sackblocks);
 165         while (sb && SEQ_LEQ(sb->sblk_end, th_ack)) {
 166                 nb = TAILQ_NEXT(sb, sblk_list);
 167                 if (scb->lastfound == sb)
 168                         scb->lastfound = NULL;
 169                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 170                 free_sackblock(sb);
 171                 --scb->nblocks;
 172                 KASSERT(scb->nblocks >= 0,
 173                     ("SACK block count underflow: %d < 0", scb->nblocks));
 174                 sb = nb;
 175         }
 176         if (sb && SEQ_GT(th_ack, sb->sblk_start))
 177                 sb->sblk_start = th_ack;        /* other side reneged? XXX */
 178 }
 179
 180 /*
 181  * Delete and free SACK blocks saved in scoreboard.
 182  */
 183 void
 184 tcp_sack_cleanup(struct scoreboard *scb)
 185 {
 186         struct sackblock *sb, *nb;
 187
 188         TAILQ_FOREACH_MUTABLE(sb, &scb->sackblocks, sblk_list, nb) {
 189                 free_sackblock(sb);
 190                 --scb->nblocks;
 191         }
 192         KASSERT(scb->nblocks == 0,
 193             ("SACK block %d count not zero", scb->nblocks));
 194         TAILQ_INIT(&scb->sackblocks);
 195         scb->lastfound = NULL;
 196 }
 197
 198 /*
 199  * Returns      0 if not D-SACK block,
 200  *              1 if D-SACK,
 201  *              2 if duplicate of out-of-order D-SACK block.
 202  */
 203 int
 204 tcp_sack_ndsack_blocks(struct raw_sackblock *blocks, const int numblocks,
 205                        tcp_seq snd_una)
 206 {
 207         if (numblocks == 0)
 208                 return 0;
 209
 210         if (SEQ_LT(blocks[0].rblk_start, snd_una))
 211                 return 1;
 212
 213         /* block 0 inside block 1 */
 214         if (numblocks > 1 &&
 215             SEQ_GEQ(blocks[0].rblk_start, blocks[1].rblk_start) &&
 216             SEQ_LEQ(blocks[0].rblk_end, blocks[1].rblk_end))
 217                 return 2;
 218
 219         return 0;
 220 }
 221
 222 /*
 223  * Update scoreboard on new incoming ACK.
 224  */
 225 static void
 226 tcp_sack_add_blocks(struct tcpcb *tp, struct tcpopt *to)
 227 {
 228         const int numblocks = to->to_nsackblocks;
 229         struct raw_sackblock *blocks = to->to_sackblocks;
 230         struct scoreboard *scb = &tp->scb;
 231         struct sackblock *sb;
 232         int startblock;
 233         int i;
 234
 235         if (tcp_sack_ndsack_blocks(blocks, numblocks, tp->snd_una) > 0)
 236                 startblock = 1;
 237         else
 238                 startblock = 0;
 239
 240         for (i = startblock; i < numblocks; i++) {
 241                 struct raw_sackblock *newsackblock = &blocks[i];
 242
 243                 /* don't accept bad SACK blocks */
 244                 if (SEQ_GT(newsackblock->rblk_end, tp->snd_max))
 245                         break;          /* skip all other blocks */
 246
 247                 sb = alloc_sackblock();
 248                 if (sb == NULL)         /* do some sort of cleanup? XXX */
 249                         break;          /* just skip rest of blocks */
 250                 sb->sblk_start = newsackblock->rblk_start;
 251                 sb->sblk_end = newsackblock->rblk_end;
 252                 if (TAILQ_EMPTY(&scb->sackblocks)) {
 253                         KASSERT(scb->nblocks == 0, ("emply scb w/ blocks"));
 254                         scb->nblocks = 1;
 255                         TAILQ_INSERT_HEAD(&scb->sackblocks, sb, sblk_list);
 256                 } else {
 257                         insert_block(scb, sb);
 258                 }
 259         }
 260 }
 261
 262 void
 263 tcp_sack_update_scoreboard(struct tcpcb *tp, struct tcpopt *to)
 264 {
 265         struct scoreboard *scb = &tp->scb;
 266
 267         tcp_sack_ack_blocks(scb, tp->snd_una);
 268         tcp_sack_add_blocks(tp, to);
 269         update_lostseq(scb, tp->snd_una, tp->t_maxseg);
 270         if (SEQ_LT(tp->rexmt_high, tp->snd_una))
 271                 tp->rexmt_high = tp->snd_una;
 272 }
 273
 274 /*
 275  * Insert SACK block into sender's scoreboard.
 276  */
 277 static void
 278 insert_block(struct scoreboard *scb, struct sackblock *newblock)
 279 {
 280         struct sackblock *sb, *workingblock;
 281         boolean_t overlap_front;
 282
 283         KASSERT(scb->nblocks > 0, ("insert_block() called w/ no blocks"));
 284
 285         if (scb->nblocks == MAXSAVEDBLOCKS) {
 286                 /*
 287                  * Should try to kick out older blocks XXX JH
 288                  * May be able to coalesce with existing block.
 289                  * Or, go other way and free all blocks if we hit this limit.
 290                  */
 291                 free_sackblock(newblock);
 292                 return;
 293         }
 294         KASSERT(scb->nblocks < MAXSAVEDBLOCKS,
 295             ("too many SACK blocks %d", scb->nblocks));
 296
 297         overlap_front = sack_block_lookup(scb, newblock->sblk_start,  &sb);
 298
 299         if (sb == NULL) {
 300                 workingblock = newblock;
 301                 TAILQ_INSERT_HEAD(&scb->sackblocks, newblock, sblk_list);
 302                 ++scb->nblocks;
 303         } else {
 304                 if (overlap_front || sb->sblk_end == newblock->sblk_start) {
 305                         /* extend old block and discard new one */
 306                         workingblock = sb;
 307                         if (SEQ_GT(newblock->sblk_end, sb->sblk_end))
 308                                 sb->sblk_end = newblock->sblk_end;
 309                         free_sackblock(newblock);
 310                 } else {
 311                         workingblock = newblock;
 312                         TAILQ_INSERT_AFTER(&scb->sackblocks, sb, newblock,
 313                                            sblk_list);
 314                         ++scb->nblocks;
 315                 }
 316         }
 317
 318         /* Consolidate right-hand side. */
 319         sb = TAILQ_NEXT(workingblock, sblk_list);
 320         while (sb != NULL &&
 321             SEQ_GEQ(workingblock->sblk_end, sb->sblk_end)) {
 322                 struct sackblock *nextblock;
 323
 324                 nextblock = TAILQ_NEXT(sb, sblk_list);
 325                 if (scb->lastfound == sb)
 326                         scb->lastfound = NULL;
 327                 /* Remove completely overlapped block */
 328                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 329                 free_sackblock(sb);
 330                 --scb->nblocks;
 331                 KASSERT(scb->nblocks > 0,
 332                     ("removed overlapped block: %d blocks left", scb->nblocks));
 333                 sb = nextblock;
 334         }
 335         if (sb != NULL &&
 336             SEQ_GEQ(workingblock->sblk_end, sb->sblk_start)) {
 337                 /* Extend new block to cover partially overlapped old block. */
 338                 workingblock->sblk_end = sb->sblk_end;
 339                 if (scb->lastfound == sb)
 340                         scb->lastfound = NULL;
 341                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 342                 free_sackblock(sb);
 343                 --scb->nblocks;
 344                 KASSERT(scb->nblocks > 0,
 345                     ("removed partial right: %d blocks left", scb->nblocks));
 346         }
 347 }
 348
 349 #ifdef DEBUG_SACK_BLOCKS
 350 static void
 351 tcp_sack_dump_blocks(struct scoreboard *scb)
 352 {
 353         struct sackblock *sb;
 354
 355         kprintf("%d blocks:", scb->nblocks);
 356         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
 357                 kprintf(" [%u, %u)", sb->sblk_start, sb->sblk_end);
 358         kprintf("\n");
 359 }
 360 #else
 361 static __inline void
 362 tcp_sack_dump_blocks(struct scoreboard *scb)
 363 {
 364 }
 365 #endif
 366
 367 /*
 368  * Optimization to quickly determine which packets are lost.
 369  */
 370 static void
 371 update_lostseq(struct scoreboard *scb, tcp_seq snd_una, u_int maxseg)
 372 {
 373         struct sackblock *sb;
 374         int nsackblocks = 0;
 375         int bytes_sacked = 0;
 376
 377         sb = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 378         while (sb != NULL) {
 379                 ++nsackblocks;
 380                 bytes_sacked += sb->sblk_end - sb->sblk_start;
 381                 if (nsackblocks == tcprexmtthresh ||
 382                     bytes_sacked >= tcprexmtthresh * maxseg) {
 383                         scb->lostseq = sb->sblk_start;
 384                         return;
 385                 }
 386                 sb = TAILQ_PREV(sb, sackblock_list, sblk_list);
 387         }
 388         scb->lostseq = snd_una;
 389 }
 390
 391 /*
 392  * Return whether the given sequence number is considered lost.
 393  */
 394 static boolean_t
 395 scb_islost(struct scoreboard *scb, tcp_seq seqnum)
 396 {
 397         return SEQ_LT(seqnum, scb->lostseq);
 398 }
 399
 400 /*
 401  * True if at least "amount" has been SACKed.  Used by Early Retransmit.
 402  */
 403 boolean_t
 404 tcp_sack_has_sacked(struct scoreboard *scb, u_int amount)
 405 {
 406         struct sackblock *sb;
 407         int bytes_sacked = 0;
 408
 409         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
 410                 bytes_sacked += sb->sblk_end - sb->sblk_start;
 411                 if (bytes_sacked >= amount)
 412                         return TRUE;
 413         }
 414         return FALSE;
 415 }
 416
 417 /*
 418  * Number of bytes SACKed below seq.
 419  */
 420 int
 421 tcp_sack_bytes_below(struct scoreboard *scb, tcp_seq seq)
 422 {
 423         struct sackblock *sb;
 424         int bytes_sacked = 0;
 425
 426         sb = TAILQ_FIRST(&scb->sackblocks);
 427         while (sb && SEQ_GT(seq, sb->sblk_start)) {
 428                 bytes_sacked += seq_min(seq, sb->sblk_end) - sb->sblk_start;
 429                 sb = TAILQ_NEXT(sb, sblk_list);
 430         }
 431         return bytes_sacked;
 432 }
 433
 434 /*
 435  * Return estimate of the number of bytes outstanding in the network.
 436  */
 437 uint32_t
 438 tcp_sack_compute_pipe(struct tcpcb *tp)
 439 {
 440         struct scoreboard *scb = &tp->scb;
 441         struct sackblock *sb;
 442         int nlost, nretransmitted;
 443         tcp_seq end;
 444
 445         nlost = tp->snd_max - scb->lostseq;
 446         nretransmitted = tp->rexmt_high - tp->snd_una;
 447
 448         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
 449                 if (SEQ_LT(sb->sblk_start, tp->rexmt_high)) {
 450                         end = seq_min(sb->sblk_end, tp->rexmt_high);
 451                         nretransmitted -= end - sb->sblk_start;
 452                 }
 453                 if (SEQ_GEQ(sb->sblk_start, scb->lostseq))
 454                         nlost -= sb->sblk_end - sb->sblk_start;
 455         }
 456
 457         return (nlost + nretransmitted);
 458 }
 459
 460 /*
 461  * Return the sequence number and length of the next segment to transmit
 462  * when in Fast Recovery.
 463  */
 464 boolean_t
 465 tcp_sack_nextseg(struct tcpcb *tp, tcp_seq *nextrexmt, uint32_t *plen,
 466                  boolean_t *lostdup)
 467 {
 468         struct scoreboard *scb = &tp->scb;
 469         struct socket *so = tp->t_inpcb->inp_socket;
 470         struct sackblock *sb;
 471         const struct sackblock *lastblock =
 472             TAILQ_LAST(&scb->sackblocks, sackblock_list);
 473         tcp_seq torexmt;
 474         long len, off;
 475
 476         /* skip SACKed data */
 477         tcp_sack_skip_sacked(scb, &tp->rexmt_high);
 478
 479         /* Look for lost data. */
 480         torexmt = tp->rexmt_high;
 481         *lostdup = FALSE;
 482         if (lastblock != NULL) {
 483                 if (SEQ_LT(torexmt, lastblock->sblk_end) &&
 484                     scb_islost(scb, torexmt)) {
 485 sendunsacked:
 486                         *nextrexmt = torexmt;
 487                         /* If the left-hand edge has been SACKed, pull it in. */
 488                         if (sack_block_lookup(scb, torexmt + tp->t_maxseg, &sb))
 489                                 *plen = sb->sblk_start - torexmt;
 490                         else
 491                                 *plen = tp->t_maxseg;
 492                         return TRUE;
 493                 }
 494         }
 495
 496         /* See if unsent data available within send window. */
 497         off = tp->snd_max - tp->snd_una;
 498         len = (long) ulmin(so->so_snd.ssb_cc, tp->snd_wnd) - off;
 499         if (len > 0) {
 500                 *nextrexmt = tp->snd_max;       /* Send new data. */
 501                 *plen = tp->t_maxseg;
 502                 return TRUE;
 503         }
 504
 505         /* We're less certain this data has been lost. */
 506         if (lastblock == NULL || SEQ_LT(torexmt, lastblock->sblk_end))
 507                 goto sendunsacked;
 508
 509         return FALSE;
 510 }
 511
 512 /*
 513  * Return the next sequence number higher than "*prexmt" that has
 514  * not been SACKed.
 515  */
 516 void
 517 tcp_sack_skip_sacked(struct scoreboard *scb, tcp_seq *prexmt)
 518 {
 519         struct sackblock *sb;
 520
 521         /* skip SACKed data */
 522         if (sack_block_lookup(scb, *prexmt, &sb))
 523                 *prexmt = sb->sblk_end;
 524 }
 525
 526 #ifdef later
 527 void
 528 tcp_sack_save_scoreboard(struct scoreboard *scb)
 529 {
 530         struct scoreboard *scb = &tp->scb;
 531
 532         scb->sackblocks_prev = scb->sackblocks;
 533         TAILQ_INIT(&scb->sackblocks);
 534 }
 535
 536 void
 537 tcp_sack_revert_scoreboard(struct scoreboard *scb, tcp_seq snd_una,
 538                            u_int maxseg)
 539 {
 540         struct sackblock *sb;
 541
 542         scb->sackblocks = scb->sackblocks_prev;
 543         scb->nblocks = 0;
 544         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
 545                 ++scb->nblocks;
 546         tcp_sack_ack_blocks(scb, snd_una);
 547         scb->lastfound = NULL;
 548 }
 549 #endif
 550
 551 #ifdef DEBUG_SACK_HISTORY
 552 static void
 553 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
 554 {
 555         int i;
 556         static int ndumped;
 557
 558         /* only need a couple of these to debug most problems */
 559         if (++ndumped > 900)
 560                 return;
 561
 562         kprintf("%s:\tnsackhistory %d: ", msg, tp->nsackhistory);
 563         for (i = 0; i < tp->nsackhistory; ++i)
 564                 kprintf("[%u, %u) ", tp->sackhistory[i].rblk_start,
 565                     tp->sackhistory[i].rblk_end);
 566         kprintf("\n");
 567 }
 568 #else
 569 static __inline void
 570 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
 571 {
 572 }
 573 #endif
 574
 575 /*
 576  * Remove old SACK blocks from the SACK history that have already been ACKed.
 577  */
 578 static void
 579 tcp_sack_ack_history(struct tcpcb *tp)
 580 {
 581         int i, nblocks, openslot;
 582
 583         tcp_sack_dump_history("before tcp_sack_ack_history", tp);
 584         nblocks = tp->nsackhistory;
 585         for (i = openslot = 0; i < nblocks; ++i) {
 586                 if (SEQ_LEQ(tp->sackhistory[i].rblk_end, tp->rcv_nxt)) {
 587                         --tp->nsackhistory;
 588                         continue;
 589                 }
 590                 if (SEQ_LT(tp->sackhistory[i].rblk_start, tp->rcv_nxt))
 591                         tp->sackhistory[i].rblk_start = tp->rcv_nxt;
 592                 if (i == openslot)
 593                         ++openslot;
 594                 else
 595                         tp->sackhistory[openslot++] = tp->sackhistory[i];
 596         }
 597         tcp_sack_dump_history("after tcp_sack_ack_history", tp);
 598         KASSERT(openslot == tp->nsackhistory,
 599             ("tcp_sack_ack_history miscounted: %d != %d",
 600             openslot, tp->nsackhistory));
 601 }
 602
 603 /*
 604  * Add or merge newblock into reported history.
 605  * Also remove or update SACK blocks that will be acked.
 606  */
 607 static void
 608 tcp_sack_update_reported_history(struct tcpcb *tp, tcp_seq start, tcp_seq end)
 609 {
 610         struct raw_sackblock copy[MAX_SACK_REPORT_BLOCKS];
 611         int i, cindex;
 612
 613         tcp_sack_dump_history("before tcp_sack_update_reported_history", tp);
 614         /*
 615          * Six cases:
 616          *      0) no overlap
 617          *      1) newblock == oldblock
 618          *      2) oldblock contains newblock
 619          *      3) newblock contains oldblock
 620          *      4) tail of oldblock overlaps or abuts start of newblock
 621          *      5) tail of newblock overlaps or abuts head of oldblock
 622          */
 623         for (i = cindex = 0; i < tp->nsackhistory; ++i) {
 624                 struct raw_sackblock *oldblock = &tp->sackhistory[i];
 625                 tcp_seq old_start = oldblock->rblk_start;
 626                 tcp_seq old_end = oldblock->rblk_end;
 627
 628                 if (SEQ_LT(end, old_start) || SEQ_GT(start, old_end)) {
 629                         /* Case 0:  no overlap.  Copy old block. */
 630                         copy[cindex++] = *oldblock;
 631                         continue;
 632                 }
 633
 634                 if (SEQ_GEQ(start, old_start) && SEQ_LEQ(end, old_end)) {
 635                         /* Cases 1 & 2.  Move block to front of history. */
 636                         int j;
 637
 638                         start = old_start;
 639                         end = old_end;
 640                         /* no need to check rest of blocks */
 641                         for (j = i + 1; j < tp->nsackhistory; ++j)
 642                                 copy[cindex++] = tp->sackhistory[j];
 643                         break;
 644                 }
 645
 646                 if (SEQ_GEQ(old_end, start) && SEQ_LT(old_start, start)) {
 647                         /* Case 4:  extend start of new block. */
 648                         start = old_start;
 649                 } else if (SEQ_GEQ(end, old_start) && SEQ_GT(old_end, end)) {
 650                         /* Case 5: extend end of new block */
 651                         end = old_end;
 652                 } else {
 653                         /* Case 3.  Delete old block by not copying it. */
 654                         KASSERT(SEQ_LEQ(start, old_start) &&
 655                                 SEQ_GEQ(end, old_end),
 656                             ("bad logic: old [%u, %u), new [%u, %u)",
 657                              old_start, old_end, start, end));
 658                 }
 659         }
 660
 661         /* insert new block */
 662         tp->sackhistory[0].rblk_start = start;
 663         tp->sackhistory[0].rblk_end = end;
 664         cindex = min(cindex, MAX_SACK_REPORT_BLOCKS - 1);
 665         for (i = 0; i < cindex; ++i)
 666                 tp->sackhistory[i + 1] = copy[i];
 667         tp->nsackhistory = cindex + 1;
 668         tcp_sack_dump_history("after tcp_sack_update_reported_history", tp);
 669 }
 670
 671 /*
 672  * Fill in SACK report to return to data sender.
 673  */
 674 void
 675 tcp_sack_fill_report(struct tcpcb *tp, u_char *opt, u_int *plen)
 676 {
 677         u_int optlen = *plen;
 678         uint32_t *lp = (uint32_t *)(opt + optlen);
 679         uint32_t *olp;
 680         tcp_seq hstart = tp->rcv_nxt, hend;
 681         int nblocks;
 682
 683         KASSERT(TCP_MAXOLEN - optlen >=
 684             TCPOLEN_SACK_ALIGNED + TCPOLEN_SACK_BLOCK,
 685             ("no room for SACK header and one block: optlen %d", optlen));
 686
 687         olp = lp++;
 688         optlen += TCPOLEN_SACK_ALIGNED;
 689
 690         tcp_sack_ack_history(tp);
 691         if (tp->reportblk.rblk_start != tp->reportblk.rblk_end) {
 692                 *lp++ = htonl(tp->reportblk.rblk_start);
 693                 *lp++ = htonl(tp->reportblk.rblk_end);
 694                 optlen += TCPOLEN_SACK_BLOCK;
 695                 hstart = tp->reportblk.rblk_start;
 696                 hend = tp->reportblk.rblk_end;
 697                 if (tp->t_flags & TF_ENCLOSESEG) {
 698                         KASSERT(TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK,
 699                             ("no room for enclosing SACK block: oplen %d",
 700                             optlen));
 701                         *lp++ = htonl(tp->encloseblk.rblk_start);
 702                         *lp++ = htonl(tp->encloseblk.rblk_end);
 703                         optlen += TCPOLEN_SACK_BLOCK;
 704                         hstart = tp->encloseblk.rblk_start;
 705                         hend = tp->encloseblk.rblk_end;
 706                 }
 707                 if (SEQ_GT(hstart, tp->rcv_nxt))
 708                         tcp_sack_update_reported_history(tp, hstart, hend);
 709         }
 710         if (tcp_do_smartsack && (tp->t_flags & TF_SACKLEFT)) {
 711                 /* Fill in from left!  Walk re-assembly queue. */
 712                 struct tseg_qent *q;
 713
 714                 q = LIST_FIRST(&tp->t_segq);
 715                 while (q != NULL &&
 716                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
 717                         *lp++ = htonl(q->tqe_th->th_seq);
 718                         *lp++ = htonl(q->tqe_th->th_seq + q->tqe_len);
 719                         optlen += TCPOLEN_SACK_BLOCK;
 720                         q = LIST_NEXT(q, tqe_q);
 721                 }
 722         } else {
 723                 int n = 0;
 724
 725                 /* Fill in SACK blocks from right side. */
 726                 while (n < tp->nsackhistory &&
 727                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
 728                         if (tp->sackhistory[n].rblk_start != hstart) {
 729                                 *lp++ = htonl(tp->sackhistory[n].rblk_start);
 730                                 *lp++ = htonl(tp->sackhistory[n].rblk_end);
 731                                 optlen += TCPOLEN_SACK_BLOCK;
 732                         }
 733                         ++n;
 734                 }
 735         }
 736         tp->reportblk.rblk_start = tp->reportblk.rblk_end;
 737         tp->t_flags &= ~(TF_DUPSEG | TF_ENCLOSESEG | TF_SACKLEFT);
 738         nblocks = (lp - olp - 1) / 2;
 739         *olp = htonl(TCPOPT_SACK_ALIGNED |
 740                      (TCPOLEN_SACK + nblocks * TCPOLEN_SACK_BLOCK));
 741         *plen = optlen;
 742 }