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 /*
  59  * Implemented:
  60  *
  61  * RFC 2018
  62  * RFC 2883
  63  * RFC 3517
  64  */
  65
  66 struct sackblock {
  67         tcp_seq                 sblk_start;
  68         tcp_seq                 sblk_end;
  69         TAILQ_ENTRY(sackblock)  sblk_list;
  70 };
  71
  72 #define MAXSAVEDBLOCKS  8                       /* per connection limit */
  73
  74 static void insert_block(struct scoreboard *scb, struct sackblock *newblock);
  75 static void update_lostseq(struct scoreboard *scb, tcp_seq snd_una,
  76                            u_int maxseg);
  77
  78 static MALLOC_DEFINE(M_SACKBLOCK, "sblk", "sackblock struct");
  79
  80 /*
  81  * Per-tcpcb initialization.
  82  */
  83 void
  84 tcp_sack_tcpcb_init(struct tcpcb *tp)
  85 {
  86         struct scoreboard *scb = &tp->scb;
  87
  88         scb->nblocks = 0;
  89         TAILQ_INIT(&scb->sackblocks);
  90         scb->lastfound = NULL;
  91 }
  92
  93 /*
  94  * Find the SACK block containing or immediately preceding "seq".
  95  * The boolean result indicates whether the sequence is actually
  96  * contained in the SACK block.
  97  */
  98 static boolean_t
  99 sack_block_lookup(struct scoreboard *scb, tcp_seq seq, struct sackblock **sb)
 100 {
 101         struct sackblock *hint = scb->lastfound;
 102         struct sackblock *cur, *last, *prev;
 103
 104         if (TAILQ_EMPTY(&scb->sackblocks)) {
 105                 *sb = NULL;
 106                 return FALSE;
 107         }
 108
 109         if (hint == NULL) {
 110                 /* No hint.  Search from start to end. */
 111                 cur = TAILQ_FIRST(&scb->sackblocks);
 112                 last = NULL;
 113                 prev = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 114         } else  {
 115                 if (SEQ_GEQ(seq, hint->sblk_start)) {
 116                         /* Search from hint to end of list. */
 117                         cur = hint;
 118                         last = NULL;
 119                         prev = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 120                 } else {
 121                         /* Search from front of list to hint. */
 122                         cur = TAILQ_FIRST(&scb->sackblocks);
 123                         last = hint;
 124                         prev = TAILQ_PREV(hint, sackblock_list, sblk_list);
 125                 }
 126         }
 127
 128         do {
 129                 if (SEQ_GT(cur->sblk_end, seq)) {
 130                         if (SEQ_GEQ(seq, cur->sblk_start)) {
 131                                 *sb = scb->lastfound = cur;
 132                                 return TRUE;
 133                         } else {
 134                                 *sb = scb->lastfound =
 135                                     TAILQ_PREV(cur, sackblock_list, sblk_list);
 136                                 return FALSE;
 137                         }
 138                 }
 139                 cur = TAILQ_NEXT(cur, sblk_list);
 140         } while (cur != last);
 141
 142         *sb = scb->lastfound = prev;
 143         return FALSE;
 144 }
 145
 146 /*
 147  * Allocate a SACK block.
 148  */
 149 static __inline struct sackblock *
 150 alloc_sackblock(void)
 151 {
 152         return (kmalloc(sizeof(struct sackblock), M_SACKBLOCK, M_NOWAIT));
 153 }
 154
 155 /*
 156  * Free a SACK block.
 157  */
 158 static __inline void
 159 free_sackblock(struct sackblock *s)
 160 {
 161         kfree(s, M_SACKBLOCK);
 162 }
 163
 164 /*
 165  * Free up SACK blocks for data that's been acked.
 166  */
 167 static void
 168 tcp_sack_ack_blocks(struct scoreboard *scb, tcp_seq th_ack)
 169 {
 170         struct sackblock *sb, *nb;
 171
 172         sb = TAILQ_FIRST(&scb->sackblocks);
 173         while (sb && SEQ_LEQ(sb->sblk_end, th_ack)) {
 174                 nb = TAILQ_NEXT(sb, sblk_list);
 175                 if (scb->lastfound == sb)
 176                         scb->lastfound = NULL;
 177                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 178                 free_sackblock(sb);
 179                 --scb->nblocks;
 180                 KASSERT(scb->nblocks >= 0,
 181                     ("SACK block count underflow: %d < 0", scb->nblocks));
 182                 sb = nb;
 183         }
 184         if (sb && SEQ_GT(th_ack, sb->sblk_start))
 185                 sb->sblk_start = th_ack;        /* other side reneged? XXX */
 186 }
 187
 188 /*
 189  * Delete and free SACK blocks saved in scoreboard.
 190  */
 191 void
 192 tcp_sack_cleanup(struct scoreboard *scb)
 193 {
 194         struct sackblock *sb, *nb;
 195
 196         TAILQ_FOREACH_MUTABLE(sb, &scb->sackblocks, sblk_list, nb) {
 197                 free_sackblock(sb);
 198                 --scb->nblocks;
 199         }
 200         KASSERT(scb->nblocks == 0,
 201             ("SACK block %d count not zero", scb->nblocks));
 202         TAILQ_INIT(&scb->sackblocks);
 203         scb->lastfound = NULL;
 204 }
 205
 206 /*
 207  * Cleanup the reported SACK block information
 208  */
 209 void
 210 tcp_sack_report_cleanup(struct tcpcb *tp)
 211 {
 212         tp->t_flags &= ~(TF_DUPSEG | TF_ENCLOSESEG | TF_SACKLEFT);
 213         tp->reportblk.rblk_start = tp->reportblk.rblk_end;
 214 }
 215
 216 /*
 217  * Returns      0 if not D-SACK block,
 218  *              1 if D-SACK,
 219  *              2 if duplicate of out-of-order D-SACK block.
 220  */
 221 int
 222 tcp_sack_ndsack_blocks(struct raw_sackblock *blocks, const int numblocks,
 223                        tcp_seq snd_una)
 224 {
 225         if (numblocks == 0)
 226                 return 0;
 227
 228         if (SEQ_LT(blocks[0].rblk_start, snd_una))
 229                 return 1;
 230
 231         /* block 0 inside block 1 */
 232         if (numblocks > 1 &&
 233             SEQ_GEQ(blocks[0].rblk_start, blocks[1].rblk_start) &&
 234             SEQ_LEQ(blocks[0].rblk_end, blocks[1].rblk_end))
 235                 return 2;
 236
 237         return 0;
 238 }
 239
 240 /*
 241  * Update scoreboard on new incoming ACK.
 242  */
 243 static void
 244 tcp_sack_add_blocks(struct tcpcb *tp, struct tcpopt *to)
 245 {
 246         const int numblocks = to->to_nsackblocks;
 247         struct raw_sackblock *blocks = to->to_sackblocks;
 248         struct scoreboard *scb = &tp->scb;
 249         struct sackblock *sb;
 250         int startblock;
 251         int i;
 252
 253         if (tcp_sack_ndsack_blocks(blocks, numblocks, tp->snd_una) > 0)
 254                 startblock = 1;
 255         else
 256                 startblock = 0;
 257
 258         for (i = startblock; i < numblocks; i++) {
 259                 struct raw_sackblock *newsackblock = &blocks[i];
 260
 261                 /* don't accept bad SACK blocks */
 262                 if (SEQ_GT(newsackblock->rblk_end, tp->snd_max)) {
 263                         tcpstat.tcps_rcvbadsackopt++;
 264                         break;          /* skip all other blocks */
 265                 }
 266                 tcpstat.tcps_sacksbupdate++;
 267
 268                 sb = alloc_sackblock();
 269                 if (sb == NULL) {       /* do some sort of cleanup? XXX */
 270                         tcpstat.tcps_sacksbfailed++;
 271                         break;          /* just skip rest of blocks */
 272                 }
 273                 sb->sblk_start = newsackblock->rblk_start;
 274                 sb->sblk_end = newsackblock->rblk_end;
 275                 insert_block(scb, sb);
 276         }
 277 }
 278
 279 void
 280 tcp_sack_update_scoreboard(struct tcpcb *tp, struct tcpopt *to)
 281 {
 282         struct scoreboard *scb = &tp->scb;
 283
 284         tcp_sack_ack_blocks(scb, tp->snd_una);
 285         tcp_sack_add_blocks(tp, to);
 286         update_lostseq(scb, tp->snd_una, tp->t_maxseg);
 287         if (SEQ_LT(tp->rexmt_high, tp->snd_una))
 288                 tp->rexmt_high = tp->snd_una;
 289 }
 290
 291 /*
 292  * Insert SACK block into sender's scoreboard.
 293  */
 294 static void
 295 insert_block(struct scoreboard *scb, struct sackblock *newblock)
 296 {
 297         struct sackblock *sb, *workingblock;
 298         boolean_t overlap_front;
 299
 300         if (TAILQ_EMPTY(&scb->sackblocks)) {
 301                 KASSERT(scb->nblocks == 0, ("emply scb w/ blocks"));
 302                 scb->nblocks = 1;
 303                 TAILQ_INSERT_HEAD(&scb->sackblocks, newblock, sblk_list);
 304                 return;
 305         }
 306         KASSERT(scb->nblocks > 0, ("insert_block() called w/ no blocks"));
 307
 308         if (scb->nblocks == MAXSAVEDBLOCKS) {
 309                 /*
 310                  * Should try to kick out older blocks XXX JH
 311                  * May be able to coalesce with existing block.
 312                  * Or, go other way and free all blocks if we hit this limit.
 313                  */
 314                 free_sackblock(newblock);
 315                 tcpstat.tcps_sacksboverflow++;
 316                 return;
 317         }
 318         KASSERT(scb->nblocks < MAXSAVEDBLOCKS,
 319             ("too many SACK blocks %d", scb->nblocks));
 320
 321         overlap_front = sack_block_lookup(scb, newblock->sblk_start,  &sb);
 322
 323         if (sb == NULL) {
 324                 workingblock = newblock;
 325                 TAILQ_INSERT_HEAD(&scb->sackblocks, newblock, sblk_list);
 326                 ++scb->nblocks;
 327         } else {
 328                 if (overlap_front || sb->sblk_end == newblock->sblk_start) {
 329                         /* extend old block and discard new one */
 330                         workingblock = sb;
 331                         if (SEQ_GT(newblock->sblk_end, sb->sblk_end))
 332                                 sb->sblk_end = newblock->sblk_end;
 333                         free_sackblock(newblock);
 334                         tcpstat.tcps_sacksbreused++;
 335                 } else {
 336                         workingblock = newblock;
 337                         TAILQ_INSERT_AFTER(&scb->sackblocks, sb, newblock,
 338                                            sblk_list);
 339                         ++scb->nblocks;
 340                 }
 341         }
 342
 343         /* Consolidate right-hand side. */
 344         sb = TAILQ_NEXT(workingblock, sblk_list);
 345         while (sb != NULL &&
 346             SEQ_GEQ(workingblock->sblk_end, sb->sblk_end)) {
 347                 struct sackblock *nextblock;
 348
 349                 nextblock = TAILQ_NEXT(sb, sblk_list);
 350                 if (scb->lastfound == sb)
 351                         scb->lastfound = NULL;
 352                 /* Remove completely overlapped block */
 353                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 354                 free_sackblock(sb);
 355                 --scb->nblocks;
 356                 KASSERT(scb->nblocks > 0,
 357                     ("removed overlapped block: %d blocks left", scb->nblocks));
 358                 sb = nextblock;
 359         }
 360         if (sb != NULL &&
 361             SEQ_GEQ(workingblock->sblk_end, sb->sblk_start)) {
 362                 /* Extend new block to cover partially overlapped old block. */
 363                 workingblock->sblk_end = sb->sblk_end;
 364                 if (scb->lastfound == sb)
 365                         scb->lastfound = NULL;
 366                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 367                 free_sackblock(sb);
 368                 --scb->nblocks;
 369                 KASSERT(scb->nblocks > 0,
 370                     ("removed partial right: %d blocks left", scb->nblocks));
 371         }
 372 }
 373
 374 #ifdef DEBUG_SACK_BLOCKS
 375 static void
 376 tcp_sack_dump_blocks(struct scoreboard *scb)
 377 {
 378         struct sackblock *sb;
 379
 380         kprintf("%d blocks:", scb->nblocks);
 381         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
 382                 kprintf(" [%u, %u)", sb->sblk_start, sb->sblk_end);
 383         kprintf("\n");
 384 }
 385 #else
 386 static __inline void
 387 tcp_sack_dump_blocks(struct scoreboard *scb)
 388 {
 389 }
 390 #endif
 391
 392 /*
 393  * Optimization to quickly determine which packets are lost.
 394  */
 395 static void
 396 update_lostseq(struct scoreboard *scb, tcp_seq snd_una, u_int maxseg)
 397 {
 398         struct sackblock *sb;
 399         int nsackblocks = 0;
 400         int bytes_sacked = 0;
 401
 402         sb = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 403         while (sb != NULL) {
 404                 ++nsackblocks;
 405                 bytes_sacked += sb->sblk_end - sb->sblk_start;
 406                 if (nsackblocks == tcprexmtthresh ||
 407                     bytes_sacked >= tcprexmtthresh * maxseg) {
 408                         scb->lostseq = sb->sblk_start;
 409                         return;
 410                 }
 411                 sb = TAILQ_PREV(sb, sackblock_list, sblk_list);
 412         }
 413         scb->lostseq = snd_una;
 414 }
 415
 416 /*
 417  * Return whether the given sequence number is considered lost.
 418  */
 419 static boolean_t
 420 scb_islost(struct scoreboard *scb, tcp_seq seqnum)
 421 {
 422         return SEQ_LT(seqnum, scb->lostseq);
 423 }
 424
 425 /*
 426  * True if at least "amount" has been SACKed.  Used by Early Retransmit.
 427  */
 428 boolean_t
 429 tcp_sack_has_sacked(struct scoreboard *scb, u_int amount)
 430 {
 431         struct sackblock *sb;
 432         int bytes_sacked = 0;
 433
 434         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
 435                 bytes_sacked += sb->sblk_end - sb->sblk_start;
 436                 if (bytes_sacked >= amount)
 437                         return TRUE;
 438         }
 439         return FALSE;
 440 }
 441
 442 /*
 443  * Number of bytes SACKed below seq.
 444  */
 445 int
 446 tcp_sack_bytes_below(struct scoreboard *scb, tcp_seq seq)
 447 {
 448         struct sackblock *sb;
 449         int bytes_sacked = 0;
 450
 451         sb = TAILQ_FIRST(&scb->sackblocks);
 452         while (sb && SEQ_GT(seq, sb->sblk_start)) {
 453                 bytes_sacked += seq_min(seq, sb->sblk_end) - sb->sblk_start;
 454                 sb = TAILQ_NEXT(sb, sblk_list);
 455         }
 456         return bytes_sacked;
 457 }
 458
 459 /*
 460  * Return estimate of the number of bytes outstanding in the network.
 461  */
 462 uint32_t
 463 tcp_sack_compute_pipe(struct tcpcb *tp)
 464 {
 465         struct scoreboard *scb = &tp->scb;
 466         struct sackblock *sb;
 467         int nlost, nretransmitted;
 468         tcp_seq end;
 469
 470         nlost = tp->snd_max - scb->lostseq;
 471         nretransmitted = tp->rexmt_high - tp->snd_una;
 472
 473         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
 474                 if (SEQ_LT(sb->sblk_start, tp->rexmt_high)) {
 475                         end = seq_min(sb->sblk_end, tp->rexmt_high);
 476                         nretransmitted -= end - sb->sblk_start;
 477                 }
 478                 if (SEQ_GEQ(sb->sblk_start, scb->lostseq))
 479                         nlost -= sb->sblk_end - sb->sblk_start;
 480         }
 481
 482         return (nlost + nretransmitted);
 483 }
 484
 485 /*
 486  * Return the sequence number and length of the next segment to transmit
 487  * when in Fast Recovery.
 488  */
 489 boolean_t
 490 tcp_sack_nextseg(struct tcpcb *tp, tcp_seq *nextrexmt, uint32_t *plen,
 491                  boolean_t *lostdup)
 492 {
 493         struct scoreboard *scb = &tp->scb;
 494         struct socket *so = tp->t_inpcb->inp_socket;
 495         struct sackblock *sb;
 496         const struct sackblock *lastblock =
 497             TAILQ_LAST(&scb->sackblocks, sackblock_list);
 498         tcp_seq torexmt;
 499         long len, off;
 500
 501         /* skip SACKed data */
 502         tcp_sack_skip_sacked(scb, &tp->rexmt_high);
 503
 504         /* Look for lost data. */
 505         torexmt = tp->rexmt_high;
 506         *lostdup = FALSE;
 507         if (lastblock != NULL) {
 508                 if (SEQ_LT(torexmt, lastblock->sblk_end) &&
 509                     scb_islost(scb, torexmt)) {
 510 sendunsacked:
 511                         *nextrexmt = torexmt;
 512                         /* If the left-hand edge has been SACKed, pull it in. */
 513                         if (sack_block_lookup(scb, torexmt + tp->t_maxseg, &sb))
 514                                 *plen = sb->sblk_start - torexmt;
 515                         else
 516                                 *plen = tp->t_maxseg;
 517                         return TRUE;
 518                 }
 519         }
 520
 521         /* See if unsent data available within send window. */
 522         off = tp->snd_max - tp->snd_una;
 523         len = (long) ulmin(so->so_snd.ssb_cc, tp->snd_wnd) - off;
 524         if (len > 0) {
 525                 *nextrexmt = tp->snd_max;       /* Send new data. */
 526                 *plen = tp->t_maxseg;
 527                 return TRUE;
 528         }
 529
 530         /* We're less certain this data has been lost. */
 531         if (lastblock == NULL || SEQ_LT(torexmt, lastblock->sblk_end))
 532                 goto sendunsacked;
 533
 534         return FALSE;
 535 }
 536
 537 /*
 538  * Return the next sequence number higher than "*prexmt" that has
 539  * not been SACKed.
 540  */
 541 void
 542 tcp_sack_skip_sacked(struct scoreboard *scb, tcp_seq *prexmt)
 543 {
 544         struct sackblock *sb;
 545
 546         /* skip SACKed data */
 547         if (sack_block_lookup(scb, *prexmt, &sb))
 548                 *prexmt = sb->sblk_end;
 549 }
 550
 551 #ifdef later
 552 void
 553 tcp_sack_save_scoreboard(struct scoreboard *scb)
 554 {
 555         struct scoreboard *scb = &tp->scb;
 556
 557         scb->sackblocks_prev = scb->sackblocks;
 558         TAILQ_INIT(&scb->sackblocks);
 559 }
 560
 561 void
 562 tcp_sack_revert_scoreboard(struct scoreboard *scb, tcp_seq snd_una,
 563                            u_int maxseg)
 564 {
 565         struct sackblock *sb;
 566
 567         scb->sackblocks = scb->sackblocks_prev;
 568         scb->nblocks = 0;
 569         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
 570                 ++scb->nblocks;
 571         tcp_sack_ack_blocks(scb, snd_una);
 572         scb->lastfound = NULL;
 573 }
 574 #endif
 575
 576 #ifdef DEBUG_SACK_HISTORY
 577 static void
 578 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
 579 {
 580         int i;
 581         static int ndumped;
 582
 583         /* only need a couple of these to debug most problems */
 584         if (++ndumped > 900)
 585                 return;
 586
 587         kprintf("%s:\tnsackhistory %d: ", msg, tp->nsackhistory);
 588         for (i = 0; i < tp->nsackhistory; ++i)
 589                 kprintf("[%u, %u) ", tp->sackhistory[i].rblk_start,
 590                     tp->sackhistory[i].rblk_end);
 591         kprintf("\n");
 592 }
 593 #else
 594 static __inline void
 595 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
 596 {
 597 }
 598 #endif
 599
 600 /*
 601  * Remove old SACK blocks from the SACK history that have already been ACKed.
 602  */
 603 static void
 604 tcp_sack_ack_history(struct tcpcb *tp)
 605 {
 606         int i, nblocks, openslot;
 607
 608         tcp_sack_dump_history("before tcp_sack_ack_history", tp);
 609         nblocks = tp->nsackhistory;
 610         for (i = openslot = 0; i < nblocks; ++i) {
 611                 if (SEQ_LEQ(tp->sackhistory[i].rblk_end, tp->rcv_nxt)) {
 612                         --tp->nsackhistory;
 613                         continue;
 614                 }
 615                 if (SEQ_LT(tp->sackhistory[i].rblk_start, tp->rcv_nxt))
 616                         tp->sackhistory[i].rblk_start = tp->rcv_nxt;
 617                 if (i == openslot)
 618                         ++openslot;
 619                 else
 620                         tp->sackhistory[openslot++] = tp->sackhistory[i];
 621         }
 622         tcp_sack_dump_history("after tcp_sack_ack_history", tp);
 623         KASSERT(openslot == tp->nsackhistory,
 624             ("tcp_sack_ack_history miscounted: %d != %d",
 625             openslot, tp->nsackhistory));
 626 }
 627
 628 /*
 629  * Add or merge newblock into reported history.
 630  * Also remove or update SACK blocks that will be acked.
 631  */
 632 static void
 633 tcp_sack_update_reported_history(struct tcpcb *tp, tcp_seq start, tcp_seq end)
 634 {
 635         struct raw_sackblock copy[MAX_SACK_REPORT_BLOCKS];
 636         int i, cindex;
 637
 638         tcp_sack_dump_history("before tcp_sack_update_reported_history", tp);
 639         /*
 640          * Six cases:
 641          *      0) no overlap
 642          *      1) newblock == oldblock
 643          *      2) oldblock contains newblock
 644          *      3) newblock contains oldblock
 645          *      4) tail of oldblock overlaps or abuts start of newblock
 646          *      5) tail of newblock overlaps or abuts head of oldblock
 647          */
 648         for (i = cindex = 0; i < tp->nsackhistory; ++i) {
 649                 struct raw_sackblock *oldblock = &tp->sackhistory[i];
 650                 tcp_seq old_start = oldblock->rblk_start;
 651                 tcp_seq old_end = oldblock->rblk_end;
 652
 653                 if (SEQ_LT(end, old_start) || SEQ_GT(start, old_end)) {
 654                         /* Case 0:  no overlap.  Copy old block. */
 655                         copy[cindex++] = *oldblock;
 656                         continue;
 657                 }
 658
 659                 if (SEQ_GEQ(start, old_start) && SEQ_LEQ(end, old_end)) {
 660                         /* Cases 1 & 2.  Move block to front of history. */
 661                         int j;
 662
 663                         start = old_start;
 664                         end = old_end;
 665                         /* no need to check rest of blocks */
 666                         for (j = i + 1; j < tp->nsackhistory; ++j)
 667                                 copy[cindex++] = tp->sackhistory[j];
 668                         break;
 669                 }
 670
 671                 if (SEQ_GEQ(old_end, start) && SEQ_LT(old_start, start)) {
 672                         /* Case 4:  extend start of new block. */
 673                         start = old_start;
 674                 } else if (SEQ_GEQ(end, old_start) && SEQ_GT(old_end, end)) {
 675                         /* Case 5: extend end of new block */
 676                         end = old_end;
 677                 } else {
 678                         /* Case 3.  Delete old block by not copying it. */
 679                         KASSERT(SEQ_LEQ(start, old_start) &&
 680                                 SEQ_GEQ(end, old_end),
 681                             ("bad logic: old [%u, %u), new [%u, %u)",
 682                              old_start, old_end, start, end));
 683                 }
 684         }
 685
 686         /* insert new block */
 687         tp->sackhistory[0].rblk_start = start;
 688         tp->sackhistory[0].rblk_end = end;
 689         cindex = min(cindex, MAX_SACK_REPORT_BLOCKS - 1);
 690         for (i = 0; i < cindex; ++i)
 691                 tp->sackhistory[i + 1] = copy[i];
 692         tp->nsackhistory = cindex + 1;
 693         tcp_sack_dump_history("after tcp_sack_update_reported_history", tp);
 694 }
 695
 696 /*
 697  * Fill in SACK report to return to data sender.
 698  */
 699 void
 700 tcp_sack_fill_report(struct tcpcb *tp, u_char *opt, u_int *plen)
 701 {
 702         u_int optlen = *plen;
 703         uint32_t *lp = (uint32_t *)(opt + optlen);
 704         uint32_t *olp;
 705         tcp_seq hstart = tp->rcv_nxt, hend;
 706         int nblocks;
 707
 708         KASSERT(TCP_MAXOLEN - optlen >=
 709             TCPOLEN_SACK_ALIGNED + TCPOLEN_SACK_BLOCK,
 710             ("no room for SACK header and one block: optlen %d", optlen));
 711
 712         if (tp->t_flags & TF_DUPSEG)
 713                 tcpstat.tcps_snddsackopt++;
 714         else
 715                 tcpstat.tcps_sndsackopt++;
 716
 717         olp = lp++;
 718         optlen += TCPOLEN_SACK_ALIGNED;
 719
 720         tcp_sack_ack_history(tp);
 721         if (tp->reportblk.rblk_start != tp->reportblk.rblk_end) {
 722                 *lp++ = htonl(tp->reportblk.rblk_start);
 723                 *lp++ = htonl(tp->reportblk.rblk_end);
 724                 optlen += TCPOLEN_SACK_BLOCK;
 725                 hstart = tp->reportblk.rblk_start;
 726                 hend = tp->reportblk.rblk_end;
 727                 if (tp->t_flags & TF_ENCLOSESEG) {
 728                         KASSERT(TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK,
 729                             ("no room for enclosing SACK block: oplen %d",
 730                             optlen));
 731                         *lp++ = htonl(tp->encloseblk.rblk_start);
 732                         *lp++ = htonl(tp->encloseblk.rblk_end);
 733                         optlen += TCPOLEN_SACK_BLOCK;
 734                         hstart = tp->encloseblk.rblk_start;
 735                         hend = tp->encloseblk.rblk_end;
 736                 }
 737                 if (SEQ_GT(hstart, tp->rcv_nxt))
 738                         tcp_sack_update_reported_history(tp, hstart, hend);
 739         }
 740         if (tcp_do_smartsack && (tp->t_flags & TF_SACKLEFT)) {
 741                 /* Fill in from left!  Walk re-assembly queue. */
 742                 struct tseg_qent *q;
 743
 744                 q = LIST_FIRST(&tp->t_segq);
 745                 while (q != NULL &&
 746                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
 747                         *lp++ = htonl(q->tqe_th->th_seq);
 748                         *lp++ = htonl(TCP_SACK_BLKEND(
 749                             q->tqe_th->th_seq + q->tqe_len,
 750                             q->tqe_th->th_flags));
 751                         optlen += TCPOLEN_SACK_BLOCK;
 752                         q = LIST_NEXT(q, tqe_q);
 753                 }
 754         } else {
 755                 int n = 0;
 756
 757                 /* Fill in SACK blocks from right side. */
 758                 while (n < tp->nsackhistory &&
 759                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
 760                         if (tp->sackhistory[n].rblk_start != hstart) {
 761                                 *lp++ = htonl(tp->sackhistory[n].rblk_start);
 762                                 *lp++ = htonl(tp->sackhistory[n].rblk_end);
 763                                 optlen += TCPOLEN_SACK_BLOCK;
 764                         }
 765                         ++n;
 766                 }
 767         }
 768         tp->reportblk.rblk_start = tp->reportblk.rblk_end;
 769         tp->t_flags &= ~(TF_DUPSEG | TF_ENCLOSESEG | TF_SACKLEFT);
 770         nblocks = (lp - olp - 1) / 2;
 771         *olp = htonl(TCPOPT_SACK_ALIGNED |
 772                      (TCPOLEN_SACK + nblocks * TCPOLEN_SACK_BLOCK));
 773         *plen = optlen;
 774 }