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(const struct raw_sackblock *raw_sb)
 151 {
 152         struct sackblock *sb;
 153
 154         sb = kmalloc(sizeof(struct sackblock), M_SACKBLOCK, M_NOWAIT);
 155         if (sb != NULL) {
 156                 sb->sblk_start = raw_sb->rblk_start;
 157                 sb->sblk_end = raw_sb->rblk_end;
 158         }
 159         return sb;
 160 }
 161
 162 /*
 163  * Free a SACK block.
 164  */
 165 static __inline void
 166 free_sackblock(struct sackblock *s)
 167 {
 168         kfree(s, M_SACKBLOCK);
 169 }
 170
 171 /*
 172  * Free up SACK blocks for data that's been acked.
 173  */
 174 static void
 175 tcp_sack_ack_blocks(struct scoreboard *scb, tcp_seq th_ack)
 176 {
 177         struct sackblock *sb, *nb;
 178
 179         sb = TAILQ_FIRST(&scb->sackblocks);
 180         while (sb && SEQ_LEQ(sb->sblk_end, th_ack)) {
 181                 nb = TAILQ_NEXT(sb, sblk_list);
 182                 if (scb->lastfound == sb)
 183                         scb->lastfound = NULL;
 184                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 185                 free_sackblock(sb);
 186                 --scb->nblocks;
 187                 KASSERT(scb->nblocks >= 0,
 188                     ("SACK block count underflow: %d < 0", scb->nblocks));
 189                 sb = nb;
 190         }
 191         if (sb && SEQ_GT(th_ack, sb->sblk_start))
 192                 sb->sblk_start = th_ack;        /* other side reneged? XXX */
 193 }
 194
 195 /*
 196  * Delete and free SACK blocks saved in scoreboard.
 197  */
 198 void
 199 tcp_sack_cleanup(struct scoreboard *scb)
 200 {
 201         struct sackblock *sb, *nb;
 202
 203         TAILQ_FOREACH_MUTABLE(sb, &scb->sackblocks, sblk_list, nb) {
 204                 free_sackblock(sb);
 205                 --scb->nblocks;
 206         }
 207         KASSERT(scb->nblocks == 0,
 208             ("SACK block %d count not zero", scb->nblocks));
 209         TAILQ_INIT(&scb->sackblocks);
 210         scb->lastfound = NULL;
 211 }
 212
 213 /*
 214  * Cleanup the reported SACK block information
 215  */
 216 void
 217 tcp_sack_report_cleanup(struct tcpcb *tp)
 218 {
 219         tp->t_flags &= ~(TF_DUPSEG | TF_ENCLOSESEG | TF_SACKLEFT);
 220         tp->reportblk.rblk_start = tp->reportblk.rblk_end;
 221 }
 222
 223 /*
 224  * Returns      0 if not D-SACK block,
 225  *              1 if D-SACK,
 226  *              2 if duplicate of out-of-order D-SACK block.
 227  */
 228 int
 229 tcp_sack_ndsack_blocks(struct raw_sackblock *blocks, const int numblocks,
 230                        tcp_seq snd_una)
 231 {
 232         if (numblocks == 0)
 233                 return 0;
 234
 235         if (SEQ_LT(blocks[0].rblk_start, snd_una))
 236                 return 1;
 237
 238         /* block 0 inside block 1 */
 239         if (numblocks > 1 &&
 240             SEQ_GEQ(blocks[0].rblk_start, blocks[1].rblk_start) &&
 241             SEQ_LEQ(blocks[0].rblk_end, blocks[1].rblk_end))
 242                 return 2;
 243
 244         return 0;
 245 }
 246
 247 /*
 248  * Update scoreboard on new incoming ACK.
 249  */
 250 static void
 251 tcp_sack_add_blocks(struct tcpcb *tp, struct tcpopt *to)
 252 {
 253         const int numblocks = to->to_nsackblocks;
 254         struct raw_sackblock *blocks = to->to_sackblocks;
 255         struct scoreboard *scb = &tp->scb;
 256         struct sackblock *sb;
 257         int startblock;
 258         int i;
 259
 260         if (tcp_sack_ndsack_blocks(blocks, numblocks, tp->snd_una) > 0)
 261                 startblock = 1;
 262         else
 263                 startblock = 0;
 264
 265         for (i = startblock; i < numblocks; i++) {
 266                 struct raw_sackblock *newsackblock = &blocks[i];
 267
 268                 /* don't accept bad SACK blocks */
 269                 if (SEQ_GT(newsackblock->rblk_end, tp->snd_max)) {
 270                         tcpstat.tcps_rcvbadsackopt++;
 271                         break;          /* skip all other blocks */
 272                 }
 273                 tcpstat.tcps_sacksbupdate++;
 274
 275                 sb = alloc_sackblock(newsackblock);
 276                 if (sb == NULL) {       /* do some sort of cleanup? XXX */
 277                         tcpstat.tcps_sacksbfailed++;
 278                         break;          /* just skip rest of blocks */
 279                 }
 280                 insert_block(scb, sb);
 281         }
 282 }
 283
 284 void
 285 tcp_sack_update_scoreboard(struct tcpcb *tp, struct tcpopt *to)
 286 {
 287         struct scoreboard *scb = &tp->scb;
 288
 289         tcp_sack_ack_blocks(scb, tp->snd_una);
 290         tcp_sack_add_blocks(tp, to);
 291         update_lostseq(scb, tp->snd_una, tp->t_maxseg);
 292         if (SEQ_LT(tp->rexmt_high, tp->snd_una))
 293                 tp->rexmt_high = tp->snd_una;
 294 }
 295
 296 /*
 297  * Insert SACK block into sender's scoreboard.
 298  */
 299 static void
 300 insert_block(struct scoreboard *scb, struct sackblock *newblock)
 301 {
 302         struct sackblock *sb, *workingblock;
 303         boolean_t overlap_front;
 304
 305         if (TAILQ_EMPTY(&scb->sackblocks)) {
 306                 KASSERT(scb->nblocks == 0, ("emply scb w/ blocks"));
 307                 scb->nblocks = 1;
 308                 TAILQ_INSERT_HEAD(&scb->sackblocks, newblock, sblk_list);
 309                 return;
 310         }
 311         KASSERT(scb->nblocks > 0, ("insert_block() called w/ no blocks"));
 312
 313         if (scb->nblocks == MAXSAVEDBLOCKS) {
 314                 /*
 315                  * Should try to kick out older blocks XXX JH
 316                  * May be able to coalesce with existing block.
 317                  * Or, go other way and free all blocks if we hit this limit.
 318                  */
 319                 free_sackblock(newblock);
 320                 tcpstat.tcps_sacksboverflow++;
 321                 return;
 322         }
 323         KASSERT(scb->nblocks < MAXSAVEDBLOCKS,
 324             ("too many SACK blocks %d", scb->nblocks));
 325
 326         overlap_front = sack_block_lookup(scb, newblock->sblk_start,  &sb);
 327
 328         if (sb == NULL) {
 329                 workingblock = newblock;
 330                 TAILQ_INSERT_HEAD(&scb->sackblocks, newblock, sblk_list);
 331                 ++scb->nblocks;
 332         } else {
 333                 if (overlap_front || sb->sblk_end == newblock->sblk_start) {
 334                         /* extend old block and discard new one */
 335                         workingblock = sb;
 336                         if (SEQ_GT(newblock->sblk_end, sb->sblk_end))
 337                                 sb->sblk_end = newblock->sblk_end;
 338                         free_sackblock(newblock);
 339                         tcpstat.tcps_sacksbreused++;
 340                 } else {
 341                         workingblock = newblock;
 342                         TAILQ_INSERT_AFTER(&scb->sackblocks, sb, newblock,
 343                                            sblk_list);
 344                         ++scb->nblocks;
 345                 }
 346         }
 347
 348         /* Consolidate right-hand side. */
 349         sb = TAILQ_NEXT(workingblock, sblk_list);
 350         while (sb != NULL &&
 351             SEQ_GEQ(workingblock->sblk_end, sb->sblk_end)) {
 352                 struct sackblock *nextblock;
 353
 354                 nextblock = TAILQ_NEXT(sb, sblk_list);
 355                 if (scb->lastfound == sb)
 356                         scb->lastfound = NULL;
 357                 /* Remove completely overlapped block */
 358                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 359                 free_sackblock(sb);
 360                 --scb->nblocks;
 361                 KASSERT(scb->nblocks > 0,
 362                     ("removed overlapped block: %d blocks left", scb->nblocks));
 363                 sb = nextblock;
 364         }
 365         if (sb != NULL &&
 366             SEQ_GEQ(workingblock->sblk_end, sb->sblk_start)) {
 367                 /* Extend new block to cover partially overlapped old block. */
 368                 workingblock->sblk_end = sb->sblk_end;
 369                 if (scb->lastfound == sb)
 370                         scb->lastfound = NULL;
 371                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 372                 free_sackblock(sb);
 373                 --scb->nblocks;
 374                 KASSERT(scb->nblocks > 0,
 375                     ("removed partial right: %d blocks left", scb->nblocks));
 376         }
 377 }
 378
 379 #ifdef DEBUG_SACK_BLOCKS
 380 static void
 381 tcp_sack_dump_blocks(struct scoreboard *scb)
 382 {
 383         struct sackblock *sb;
 384
 385         kprintf("%d blocks:", scb->nblocks);
 386         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
 387                 kprintf(" [%u, %u)", sb->sblk_start, sb->sblk_end);
 388         kprintf("\n");
 389 }
 390 #else
 391 static __inline void
 392 tcp_sack_dump_blocks(struct scoreboard *scb)
 393 {
 394 }
 395 #endif
 396
 397 /*
 398  * Optimization to quickly determine which packets are lost.
 399  */
 400 static void
 401 update_lostseq(struct scoreboard *scb, tcp_seq snd_una, u_int maxseg)
 402 {
 403         struct sackblock *sb;
 404         int nsackblocks = 0;
 405         int bytes_sacked = 0;
 406
 407         sb = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 408         while (sb != NULL) {
 409                 ++nsackblocks;
 410                 bytes_sacked += sb->sblk_end - sb->sblk_start;
 411                 if (nsackblocks == tcprexmtthresh ||
 412                     bytes_sacked >= tcprexmtthresh * maxseg) {
 413                         scb->lostseq = sb->sblk_start;
 414                         return;
 415                 }
 416                 sb = TAILQ_PREV(sb, sackblock_list, sblk_list);
 417         }
 418         scb->lostseq = snd_una;
 419 }
 420
 421 /*
 422  * Return whether the given sequence number is considered lost.
 423  */
 424 static boolean_t
 425 scb_islost(struct scoreboard *scb, tcp_seq seqnum)
 426 {
 427         return SEQ_LT(seqnum, scb->lostseq);
 428 }
 429
 430 /*
 431  * True if at least "amount" has been SACKed.  Used by Early Retransmit.
 432  */
 433 boolean_t
 434 tcp_sack_has_sacked(struct scoreboard *scb, u_int amount)
 435 {
 436         struct sackblock *sb;
 437         int bytes_sacked = 0;
 438
 439         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
 440                 bytes_sacked += sb->sblk_end - sb->sblk_start;
 441                 if (bytes_sacked >= amount)
 442                         return TRUE;
 443         }
 444         return FALSE;
 445 }
 446
 447 /*
 448  * Number of bytes SACKed below seq.
 449  */
 450 int
 451 tcp_sack_bytes_below(struct scoreboard *scb, tcp_seq seq)
 452 {
 453         struct sackblock *sb;
 454         int bytes_sacked = 0;
 455
 456         sb = TAILQ_FIRST(&scb->sackblocks);
 457         while (sb && SEQ_GT(seq, sb->sblk_start)) {
 458                 bytes_sacked += seq_min(seq, sb->sblk_end) - sb->sblk_start;
 459                 sb = TAILQ_NEXT(sb, sblk_list);
 460         }
 461         return bytes_sacked;
 462 }
 463
 464 /*
 465  * Return estimate of the number of bytes outstanding in the network.
 466  */
 467 uint32_t
 468 tcp_sack_compute_pipe(struct tcpcb *tp)
 469 {
 470         struct scoreboard *scb = &tp->scb;
 471         struct sackblock *sb;
 472         int nlost, nretransmitted;
 473         tcp_seq end;
 474
 475         nlost = tp->snd_max - scb->lostseq;
 476         nretransmitted = tp->rexmt_high - tp->snd_una;
 477
 478         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
 479                 if (SEQ_LT(sb->sblk_start, tp->rexmt_high)) {
 480                         end = seq_min(sb->sblk_end, tp->rexmt_high);
 481                         nretransmitted -= end - sb->sblk_start;
 482                 }
 483                 if (SEQ_GEQ(sb->sblk_start, scb->lostseq))
 484                         nlost -= sb->sblk_end - sb->sblk_start;
 485         }
 486
 487         return (nlost + nretransmitted);
 488 }
 489
 490 /*
 491  * Return the sequence number and length of the next segment to transmit
 492  * when in Fast Recovery.
 493  */
 494 boolean_t
 495 tcp_sack_nextseg(struct tcpcb *tp, tcp_seq *nextrexmt, uint32_t *plen,
 496                  boolean_t *lostdup)
 497 {
 498         struct scoreboard *scb = &tp->scb;
 499         struct socket *so = tp->t_inpcb->inp_socket;
 500         struct sackblock *sb;
 501         const struct sackblock *lastblock =
 502             TAILQ_LAST(&scb->sackblocks, sackblock_list);
 503         tcp_seq torexmt;
 504         long len, off;
 505
 506         /* skip SACKed data */
 507         tcp_sack_skip_sacked(scb, &tp->rexmt_high);
 508
 509         /* Look for lost data. */
 510         torexmt = tp->rexmt_high;
 511         *lostdup = FALSE;
 512         if (lastblock != NULL) {
 513                 if (SEQ_LT(torexmt, lastblock->sblk_end) &&
 514                     scb_islost(scb, torexmt)) {
 515 sendunsacked:
 516                         *nextrexmt = torexmt;
 517                         /* If the left-hand edge has been SACKed, pull it in. */
 518                         if (sack_block_lookup(scb, torexmt + tp->t_maxseg, &sb))
 519                                 *plen = sb->sblk_start - torexmt;
 520                         else
 521                                 *plen = tp->t_maxseg;
 522                         return TRUE;
 523                 }
 524         }
 525
 526         /* See if unsent data available within send window. */
 527         off = tp->snd_max - tp->snd_una;
 528         len = (long) ulmin(so->so_snd.ssb_cc, tp->snd_wnd) - off;
 529         if (len > 0) {
 530                 *nextrexmt = tp->snd_max;       /* Send new data. */
 531                 *plen = tp->t_maxseg;
 532                 return TRUE;
 533         }
 534
 535         /* We're less certain this data has been lost. */
 536         if (lastblock == NULL || SEQ_LT(torexmt, lastblock->sblk_end))
 537                 goto sendunsacked;
 538
 539         return FALSE;
 540 }
 541
 542 /*
 543  * Return the next sequence number higher than "*prexmt" that has
 544  * not been SACKed.
 545  */
 546 void
 547 tcp_sack_skip_sacked(struct scoreboard *scb, tcp_seq *prexmt)
 548 {
 549         struct sackblock *sb;
 550
 551         /* skip SACKed data */
 552         if (sack_block_lookup(scb, *prexmt, &sb))
 553                 *prexmt = sb->sblk_end;
 554 }
 555
 556 #ifdef later
 557 void
 558 tcp_sack_save_scoreboard(struct scoreboard *scb)
 559 {
 560         struct scoreboard *scb = &tp->scb;
 561
 562         scb->sackblocks_prev = scb->sackblocks;
 563         TAILQ_INIT(&scb->sackblocks);
 564 }
 565
 566 void
 567 tcp_sack_revert_scoreboard(struct scoreboard *scb, tcp_seq snd_una,
 568                            u_int maxseg)
 569 {
 570         struct sackblock *sb;
 571
 572         scb->sackblocks = scb->sackblocks_prev;
 573         scb->nblocks = 0;
 574         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
 575                 ++scb->nblocks;
 576         tcp_sack_ack_blocks(scb, snd_una);
 577         scb->lastfound = NULL;
 578 }
 579 #endif
 580
 581 #ifdef DEBUG_SACK_HISTORY
 582 static void
 583 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
 584 {
 585         int i;
 586         static int ndumped;
 587
 588         /* only need a couple of these to debug most problems */
 589         if (++ndumped > 900)
 590                 return;
 591
 592         kprintf("%s:\tnsackhistory %d: ", msg, tp->nsackhistory);
 593         for (i = 0; i < tp->nsackhistory; ++i)
 594                 kprintf("[%u, %u) ", tp->sackhistory[i].rblk_start,
 595                     tp->sackhistory[i].rblk_end);
 596         kprintf("\n");
 597 }
 598 #else
 599 static __inline void
 600 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
 601 {
 602 }
 603 #endif
 604
 605 /*
 606  * Remove old SACK blocks from the SACK history that have already been ACKed.
 607  */
 608 static void
 609 tcp_sack_ack_history(struct tcpcb *tp)
 610 {
 611         int i, nblocks, openslot;
 612
 613         tcp_sack_dump_history("before tcp_sack_ack_history", tp);
 614         nblocks = tp->nsackhistory;
 615         for (i = openslot = 0; i < nblocks; ++i) {
 616                 if (SEQ_LEQ(tp->sackhistory[i].rblk_end, tp->rcv_nxt)) {
 617                         --tp->nsackhistory;
 618                         continue;
 619                 }
 620                 if (SEQ_LT(tp->sackhistory[i].rblk_start, tp->rcv_nxt))
 621                         tp->sackhistory[i].rblk_start = tp->rcv_nxt;
 622                 if (i == openslot)
 623                         ++openslot;
 624                 else
 625                         tp->sackhistory[openslot++] = tp->sackhistory[i];
 626         }
 627         tcp_sack_dump_history("after tcp_sack_ack_history", tp);
 628         KASSERT(openslot == tp->nsackhistory,
 629             ("tcp_sack_ack_history miscounted: %d != %d",
 630             openslot, tp->nsackhistory));
 631 }
 632
 633 /*
 634  * Add or merge newblock into reported history.
 635  * Also remove or update SACK blocks that will be acked.
 636  */
 637 static void
 638 tcp_sack_update_reported_history(struct tcpcb *tp, tcp_seq start, tcp_seq end)
 639 {
 640         struct raw_sackblock copy[MAX_SACK_REPORT_BLOCKS];
 641         int i, cindex;
 642
 643         tcp_sack_dump_history("before tcp_sack_update_reported_history", tp);
 644         /*
 645          * Six cases:
 646          *      0) no overlap
 647          *      1) newblock == oldblock
 648          *      2) oldblock contains newblock
 649          *      3) newblock contains oldblock
 650          *      4) tail of oldblock overlaps or abuts start of newblock
 651          *      5) tail of newblock overlaps or abuts head of oldblock
 652          */
 653         for (i = cindex = 0; i < tp->nsackhistory; ++i) {
 654                 struct raw_sackblock *oldblock = &tp->sackhistory[i];
 655                 tcp_seq old_start = oldblock->rblk_start;
 656                 tcp_seq old_end = oldblock->rblk_end;
 657
 658                 if (SEQ_LT(end, old_start) || SEQ_GT(start, old_end)) {
 659                         /* Case 0:  no overlap.  Copy old block. */
 660                         copy[cindex++] = *oldblock;
 661                         continue;
 662                 }
 663
 664                 if (SEQ_GEQ(start, old_start) && SEQ_LEQ(end, old_end)) {
 665                         /* Cases 1 & 2.  Move block to front of history. */
 666                         int j;
 667
 668                         start = old_start;
 669                         end = old_end;
 670                         /* no need to check rest of blocks */
 671                         for (j = i + 1; j < tp->nsackhistory; ++j)
 672                                 copy[cindex++] = tp->sackhistory[j];
 673                         break;
 674                 }
 675
 676                 if (SEQ_GEQ(old_end, start) && SEQ_LT(old_start, start)) {
 677                         /* Case 4:  extend start of new block. */
 678                         start = old_start;
 679                 } else if (SEQ_GEQ(end, old_start) && SEQ_GT(old_end, end)) {
 680                         /* Case 5: extend end of new block */
 681                         end = old_end;
 682                 } else {
 683                         /* Case 3.  Delete old block by not copying it. */
 684                         KASSERT(SEQ_LEQ(start, old_start) &&
 685                                 SEQ_GEQ(end, old_end),
 686                             ("bad logic: old [%u, %u), new [%u, %u)",
 687                              old_start, old_end, start, end));
 688                 }
 689         }
 690
 691         /* insert new block */
 692         tp->sackhistory[0].rblk_start = start;
 693         tp->sackhistory[0].rblk_end = end;
 694         cindex = min(cindex, MAX_SACK_REPORT_BLOCKS - 1);
 695         for (i = 0; i < cindex; ++i)
 696                 tp->sackhistory[i + 1] = copy[i];
 697         tp->nsackhistory = cindex + 1;
 698         tcp_sack_dump_history("after tcp_sack_update_reported_history", tp);
 699 }
 700
 701 /*
 702  * Fill in SACK report to return to data sender.
 703  */
 704 void
 705 tcp_sack_fill_report(struct tcpcb *tp, u_char *opt, u_int *plen)
 706 {
 707         u_int optlen = *plen;
 708         uint32_t *lp = (uint32_t *)(opt + optlen);
 709         uint32_t *olp;
 710         tcp_seq hstart = tp->rcv_nxt, hend;
 711         int nblocks;
 712
 713         KASSERT(TCP_MAXOLEN - optlen >=
 714             TCPOLEN_SACK_ALIGNED + TCPOLEN_SACK_BLOCK,
 715             ("no room for SACK header and one block: optlen %d", optlen));
 716
 717         if (tp->t_flags & TF_DUPSEG)
 718                 tcpstat.tcps_snddsackopt++;
 719         else
 720                 tcpstat.tcps_sndsackopt++;
 721
 722         olp = lp++;
 723         optlen += TCPOLEN_SACK_ALIGNED;
 724
 725         tcp_sack_ack_history(tp);
 726         if (tp->reportblk.rblk_start != tp->reportblk.rblk_end) {
 727                 *lp++ = htonl(tp->reportblk.rblk_start);
 728                 *lp++ = htonl(tp->reportblk.rblk_end);
 729                 optlen += TCPOLEN_SACK_BLOCK;
 730                 hstart = tp->reportblk.rblk_start;
 731                 hend = tp->reportblk.rblk_end;
 732                 if (tp->t_flags & TF_ENCLOSESEG) {
 733                         KASSERT(TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK,
 734                             ("no room for enclosing SACK block: oplen %d",
 735                             optlen));
 736                         *lp++ = htonl(tp->encloseblk.rblk_start);
 737                         *lp++ = htonl(tp->encloseblk.rblk_end);
 738                         optlen += TCPOLEN_SACK_BLOCK;
 739                         hstart = tp->encloseblk.rblk_start;
 740                         hend = tp->encloseblk.rblk_end;
 741                 }
 742                 if (SEQ_GT(hstart, tp->rcv_nxt))
 743                         tcp_sack_update_reported_history(tp, hstart, hend);
 744         }
 745         if (tcp_do_smartsack && (tp->t_flags & TF_SACKLEFT)) {
 746                 /* Fill in from left!  Walk re-assembly queue. */
 747                 struct tseg_qent *q;
 748
 749                 q = LIST_FIRST(&tp->t_segq);
 750                 while (q != NULL &&
 751                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
 752                         *lp++ = htonl(q->tqe_th->th_seq);
 753                         *lp++ = htonl(TCP_SACK_BLKEND(
 754                             q->tqe_th->th_seq + q->tqe_len,
 755                             q->tqe_th->th_flags));
 756                         optlen += TCPOLEN_SACK_BLOCK;
 757                         q = LIST_NEXT(q, tqe_q);
 758                 }
 759         } else {
 760                 int n = 0;
 761
 762                 /* Fill in SACK blocks from right side. */
 763                 while (n < tp->nsackhistory &&
 764                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
 765                         if (tp->sackhistory[n].rblk_start != hstart) {
 766                                 *lp++ = htonl(tp->sackhistory[n].rblk_start);
 767                                 *lp++ = htonl(tp->sackhistory[n].rblk_end);
 768                                 optlen += TCPOLEN_SACK_BLOCK;
 769                         }
 770                         ++n;
 771                 }
 772         }
 773         tp->reportblk.rblk_start = tp->reportblk.rblk_end;
 774         tp->t_flags &= ~(TF_DUPSEG | TF_ENCLOSESEG | TF_SACKLEFT);
 775         nblocks = (lp - olp - 1) / 2;
 776         *olp = htonl(TCPOPT_SACK_ALIGNED |
 777                      (TCPOLEN_SACK + nblocks * TCPOLEN_SACK_BLOCK));
 778         *plen = optlen;
 779 }