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 int insert_block(struct scoreboard *scb,
  75                         const struct raw_sackblock *raw_sb);
  76 static void update_lostseq(struct scoreboard *scb, tcp_seq snd_una,
  77                            u_int maxseg);
  78
  79 static MALLOC_DEFINE(M_SACKBLOCK, "sblk", "sackblock struct");
  80
  81 /*
  82  * Per-tcpcb initialization.
  83  */
  84 void
  85 tcp_sack_tcpcb_init(struct tcpcb *tp)
  86 {
  87         struct scoreboard *scb = &tp->scb;
  88
  89         scb->nblocks = 0;
  90         TAILQ_INIT(&scb->sackblocks);
  91         scb->lastfound = NULL;
  92 }
  93
  94 /*
  95  * Find the SACK block containing or immediately preceding "seq".
  96  * The boolean result indicates whether the sequence is actually
  97  * contained in the SACK block.
  98  */
  99 static boolean_t
 100 sack_block_lookup(struct scoreboard *scb, tcp_seq seq, struct sackblock **sb)
 101 {
 102         struct sackblock *hint = scb->lastfound;
 103         struct sackblock *cur, *last, *prev;
 104
 105         if (TAILQ_EMPTY(&scb->sackblocks)) {
 106                 *sb = NULL;
 107                 return FALSE;
 108         }
 109
 110         if (hint == NULL) {
 111                 /* No hint.  Search from start to end. */
 112                 cur = TAILQ_FIRST(&scb->sackblocks);
 113                 last = NULL;
 114                 prev = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 115         } else  {
 116                 if (SEQ_GEQ(seq, hint->sblk_start)) {
 117                         /* Search from hint to end of list. */
 118                         cur = hint;
 119                         last = NULL;
 120                         prev = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 121                 } else {
 122                         /* Search from front of list to hint. */
 123                         cur = TAILQ_FIRST(&scb->sackblocks);
 124                         last = hint;
 125                         prev = TAILQ_PREV(hint, sackblock_list, sblk_list);
 126                 }
 127         }
 128
 129         do {
 130                 if (SEQ_GT(cur->sblk_end, seq)) {
 131                         if (SEQ_GEQ(seq, cur->sblk_start)) {
 132                                 *sb = scb->lastfound = cur;
 133                                 return TRUE;
 134                         } else {
 135                                 *sb = scb->lastfound =
 136                                     TAILQ_PREV(cur, sackblock_list, sblk_list);
 137                                 return FALSE;
 138                         }
 139                 }
 140                 cur = TAILQ_NEXT(cur, sblk_list);
 141         } while (cur != last);
 142
 143         *sb = scb->lastfound = prev;
 144         return FALSE;
 145 }
 146
 147 /*
 148  * Allocate a SACK block.
 149  */
 150 static __inline struct sackblock *
 151 alloc_sackblock(const struct raw_sackblock *raw_sb)
 152 {
 153         struct sackblock *sb;
 154
 155         sb = kmalloc(sizeof(struct sackblock), M_SACKBLOCK, M_NOWAIT);
 156         if (sb != NULL) {
 157                 sb->sblk_start = raw_sb->rblk_start;
 158                 sb->sblk_end = raw_sb->rblk_end;
 159         } else {
 160                 tcpstat.tcps_sacksbfailed++;
 161         }
 162         return sb;
 163 }
 164
 165 static __inline struct sackblock *
 166 alloc_sackblock_limit(struct scoreboard *scb,
 167     const struct raw_sackblock *raw_sb)
 168 {
 169         if (scb->nblocks == MAXSAVEDBLOCKS) {
 170                 /*
 171                  * Should try to kick out older blocks XXX JH
 172                  * May be able to coalesce with existing block.
 173                  * Or, go other way and free all blocks if we hit
 174                  * this limit.
 175                  */
 176                 tcpstat.tcps_sacksboverflow++;
 177                 return NULL;
 178         }
 179         return alloc_sackblock(raw_sb);
 180 }
 181
 182 /*
 183  * Free a SACK block.
 184  */
 185 static __inline void
 186 free_sackblock(struct sackblock *s)
 187 {
 188         kfree(s, M_SACKBLOCK);
 189 }
 190
 191 /*
 192  * Free up SACK blocks for data that's been acked.
 193  */
 194 static void
 195 tcp_sack_ack_blocks(struct scoreboard *scb, tcp_seq th_ack)
 196 {
 197         struct sackblock *sb, *nb;
 198
 199         sb = TAILQ_FIRST(&scb->sackblocks);
 200         while (sb && SEQ_LEQ(sb->sblk_end, th_ack)) {
 201                 nb = TAILQ_NEXT(sb, sblk_list);
 202                 if (scb->lastfound == sb)
 203                         scb->lastfound = NULL;
 204                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 205                 free_sackblock(sb);
 206                 --scb->nblocks;
 207                 KASSERT(scb->nblocks >= 0,
 208                     ("SACK block count underflow: %d < 0", scb->nblocks));
 209                 sb = nb;
 210         }
 211         if (sb && SEQ_GT(th_ack, sb->sblk_start))
 212                 sb->sblk_start = th_ack;        /* other side reneged? XXX */
 213 }
 214
 215 /*
 216  * Delete and free SACK blocks saved in scoreboard.
 217  */
 218 void
 219 tcp_sack_cleanup(struct scoreboard *scb)
 220 {
 221         struct sackblock *sb, *nb;
 222
 223         TAILQ_FOREACH_MUTABLE(sb, &scb->sackblocks, sblk_list, nb) {
 224                 free_sackblock(sb);
 225                 --scb->nblocks;
 226         }
 227         KASSERT(scb->nblocks == 0,
 228             ("SACK block %d count not zero", scb->nblocks));
 229         TAILQ_INIT(&scb->sackblocks);
 230         scb->lastfound = NULL;
 231 }
 232
 233 /*
 234  * Cleanup the reported SACK block information
 235  */
 236 void
 237 tcp_sack_report_cleanup(struct tcpcb *tp)
 238 {
 239         tp->t_flags &= ~(TF_DUPSEG | TF_ENCLOSESEG | TF_SACKLEFT);
 240         tp->reportblk.rblk_start = tp->reportblk.rblk_end;
 241 }
 242
 243 /*
 244  * Returns      0 if not D-SACK block,
 245  *              1 if D-SACK,
 246  *              2 if duplicate of out-of-order D-SACK block.
 247  */
 248 int
 249 tcp_sack_ndsack_blocks(struct raw_sackblock *blocks, const int numblocks,
 250                        tcp_seq snd_una)
 251 {
 252         if (numblocks == 0)
 253                 return 0;
 254
 255         if (SEQ_LT(blocks[0].rblk_start, snd_una))
 256                 return 1;
 257
 258         /* block 0 inside block 1 */
 259         if (numblocks > 1 &&
 260             SEQ_GEQ(blocks[0].rblk_start, blocks[1].rblk_start) &&
 261             SEQ_LEQ(blocks[0].rblk_end, blocks[1].rblk_end))
 262                 return 2;
 263
 264         return 0;
 265 }
 266
 267 /*
 268  * Update scoreboard on new incoming ACK.
 269  */
 270 static void
 271 tcp_sack_add_blocks(struct tcpcb *tp, struct tcpopt *to)
 272 {
 273         const int numblocks = to->to_nsackblocks;
 274         struct raw_sackblock *blocks = to->to_sackblocks;
 275         struct scoreboard *scb = &tp->scb;
 276         int startblock;
 277         int i;
 278
 279         if (tcp_sack_ndsack_blocks(blocks, numblocks, tp->snd_una) > 0)
 280                 startblock = 1;
 281         else
 282                 startblock = 0;
 283
 284         for (i = startblock; i < numblocks; i++) {
 285                 struct raw_sackblock *newsackblock = &blocks[i];
 286
 287                 /* don't accept bad SACK blocks */
 288                 if (SEQ_GT(newsackblock->rblk_end, tp->snd_max)) {
 289                         tcpstat.tcps_rcvbadsackopt++;
 290                         break;          /* skip all other blocks */
 291                 }
 292                 tcpstat.tcps_sacksbupdate++;
 293
 294                 if (insert_block(scb, newsackblock))
 295                         break;
 296         }
 297 }
 298
 299 void
 300 tcp_sack_update_scoreboard(struct tcpcb *tp, struct tcpopt *to)
 301 {
 302         struct scoreboard *scb = &tp->scb;
 303
 304         tcp_sack_ack_blocks(scb, tp->snd_una);
 305         tcp_sack_add_blocks(tp, to);
 306         update_lostseq(scb, tp->snd_una, tp->t_maxseg);
 307         if (SEQ_LT(tp->rexmt_high, tp->snd_una))
 308                 tp->rexmt_high = tp->snd_una;
 309 }
 310
 311 /*
 312  * Insert SACK block into sender's scoreboard.
 313  */
 314 static int
 315 insert_block(struct scoreboard *scb, const struct raw_sackblock *raw_sb)
 316 {
 317         struct sackblock *sb, *workingblock;
 318         boolean_t overlap_front;
 319
 320         if (TAILQ_EMPTY(&scb->sackblocks)) {
 321                 struct sackblock *newblock;
 322
 323                 KASSERT(scb->nblocks == 0, ("emply scb w/ blocks"));
 324
 325                 newblock = alloc_sackblock(raw_sb);
 326                 if (newblock == NULL)
 327                         return ENOMEM;
 328                 TAILQ_INSERT_HEAD(&scb->sackblocks, newblock, sblk_list);
 329                 scb->nblocks = 1;
 330                 return 0;
 331         }
 332
 333         KASSERT(scb->nblocks > 0, ("insert_block() called w/ no blocks"));
 334         KASSERT(scb->nblocks <= MAXSAVEDBLOCKS,
 335             ("too many SACK blocks %d", scb->nblocks));
 336
 337         overlap_front = sack_block_lookup(scb, raw_sb->rblk_start, &sb);
 338
 339         if (sb == NULL) {
 340                 workingblock = alloc_sackblock_limit(scb, raw_sb);
 341                 if (workingblock == NULL)
 342                         return ENOMEM;
 343                 TAILQ_INSERT_HEAD(&scb->sackblocks, workingblock, sblk_list);
 344                 ++scb->nblocks;
 345         } else {
 346                 if (overlap_front || sb->sblk_end == raw_sb->rblk_start) {
 347                         /* Extend old block */
 348                         workingblock = sb;
 349                         if (SEQ_GT(raw_sb->rblk_end, sb->sblk_end))
 350                                 sb->sblk_end = raw_sb->rblk_end;
 351                         tcpstat.tcps_sacksbreused++;
 352                 } else {
 353                         workingblock = alloc_sackblock_limit(scb, raw_sb);
 354                         if (workingblock == NULL)
 355                                 return ENOMEM;
 356                         TAILQ_INSERT_AFTER(&scb->sackblocks, sb, workingblock,
 357                             sblk_list);
 358                         ++scb->nblocks;
 359                 }
 360         }
 361
 362         /* Consolidate right-hand side. */
 363         sb = TAILQ_NEXT(workingblock, sblk_list);
 364         while (sb != NULL &&
 365             SEQ_GEQ(workingblock->sblk_end, sb->sblk_end)) {
 366                 struct sackblock *nextblock;
 367
 368                 nextblock = TAILQ_NEXT(sb, sblk_list);
 369                 if (scb->lastfound == sb)
 370                         scb->lastfound = NULL;
 371                 /* Remove completely overlapped block */
 372                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 373                 free_sackblock(sb);
 374                 --scb->nblocks;
 375                 KASSERT(scb->nblocks > 0,
 376                     ("removed overlapped block: %d blocks left", scb->nblocks));
 377                 sb = nextblock;
 378         }
 379         if (sb != NULL &&
 380             SEQ_GEQ(workingblock->sblk_end, sb->sblk_start)) {
 381                 /* Extend new block to cover partially overlapped old block. */
 382                 workingblock->sblk_end = sb->sblk_end;
 383                 if (scb->lastfound == sb)
 384                         scb->lastfound = NULL;
 385                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 386                 free_sackblock(sb);
 387                 --scb->nblocks;
 388                 KASSERT(scb->nblocks > 0,
 389                     ("removed partial right: %d blocks left", scb->nblocks));
 390         }
 391         return 0;
 392 }
 393
 394 #ifdef DEBUG_SACK_BLOCKS
 395 static void
 396 tcp_sack_dump_blocks(struct scoreboard *scb)
 397 {
 398         struct sackblock *sb;
 399
 400         kprintf("%d blocks:", scb->nblocks);
 401         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
 402                 kprintf(" [%u, %u)", sb->sblk_start, sb->sblk_end);
 403         kprintf("\n");
 404 }
 405 #else
 406 static __inline void
 407 tcp_sack_dump_blocks(struct scoreboard *scb)
 408 {
 409 }
 410 #endif
 411
 412 /*
 413  * Optimization to quickly determine which packets are lost.
 414  */
 415 static void
 416 update_lostseq(struct scoreboard *scb, tcp_seq snd_una, u_int maxseg)
 417 {
 418         struct sackblock *sb;
 419         int nsackblocks = 0;
 420         int bytes_sacked = 0;
 421
 422         sb = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 423         while (sb != NULL) {
 424                 ++nsackblocks;
 425                 bytes_sacked += sb->sblk_end - sb->sblk_start;
 426                 if (nsackblocks == tcprexmtthresh ||
 427                     bytes_sacked >= tcprexmtthresh * maxseg) {
 428                         scb->lostseq = sb->sblk_start;
 429                         return;
 430                 }
 431                 sb = TAILQ_PREV(sb, sackblock_list, sblk_list);
 432         }
 433         scb->lostseq = snd_una;
 434 }
 435
 436 /*
 437  * Return whether the given sequence number is considered lost.
 438  */
 439 static boolean_t
 440 scb_islost(struct scoreboard *scb, tcp_seq seqnum)
 441 {
 442         return SEQ_LT(seqnum, scb->lostseq);
 443 }
 444
 445 /*
 446  * True if at least "amount" has been SACKed.  Used by Early Retransmit.
 447  */
 448 boolean_t
 449 tcp_sack_has_sacked(struct scoreboard *scb, u_int amount)
 450 {
 451         struct sackblock *sb;
 452         int bytes_sacked = 0;
 453
 454         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
 455                 bytes_sacked += sb->sblk_end - sb->sblk_start;
 456                 if (bytes_sacked >= amount)
 457                         return TRUE;
 458         }
 459         return FALSE;
 460 }
 461
 462 /*
 463  * Number of bytes SACKed below seq.
 464  */
 465 int
 466 tcp_sack_bytes_below(struct scoreboard *scb, tcp_seq seq)
 467 {
 468         struct sackblock *sb;
 469         int bytes_sacked = 0;
 470
 471         sb = TAILQ_FIRST(&scb->sackblocks);
 472         while (sb && SEQ_GT(seq, sb->sblk_start)) {
 473                 bytes_sacked += seq_min(seq, sb->sblk_end) - sb->sblk_start;
 474                 sb = TAILQ_NEXT(sb, sblk_list);
 475         }
 476         return bytes_sacked;
 477 }
 478
 479 /*
 480  * Return estimate of the number of bytes outstanding in the network.
 481  */
 482 uint32_t
 483 tcp_sack_compute_pipe(struct tcpcb *tp)
 484 {
 485         struct scoreboard *scb = &tp->scb;
 486         struct sackblock *sb;
 487         int nlost, nretransmitted;
 488         tcp_seq end;
 489
 490         nlost = tp->snd_max - scb->lostseq;
 491         nretransmitted = tp->rexmt_high - tp->snd_una;
 492
 493         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
 494                 if (SEQ_LT(sb->sblk_start, tp->rexmt_high)) {
 495                         end = seq_min(sb->sblk_end, tp->rexmt_high);
 496                         nretransmitted -= end - sb->sblk_start;
 497                 }
 498                 if (SEQ_GEQ(sb->sblk_start, scb->lostseq))
 499                         nlost -= sb->sblk_end - sb->sblk_start;
 500         }
 501
 502         return (nlost + nretransmitted);
 503 }
 504
 505 /*
 506  * Return the sequence number and length of the next segment to transmit
 507  * when in Fast Recovery.
 508  */
 509 boolean_t
 510 tcp_sack_nextseg(struct tcpcb *tp, tcp_seq *nextrexmt, uint32_t *plen,
 511                  boolean_t *lostdup)
 512 {
 513         struct scoreboard *scb = &tp->scb;
 514         struct socket *so = tp->t_inpcb->inp_socket;
 515         struct sackblock *sb;
 516         const struct sackblock *lastblock =
 517             TAILQ_LAST(&scb->sackblocks, sackblock_list);
 518         tcp_seq torexmt;
 519         long len, off;
 520
 521         /* skip SACKed data */
 522         tcp_sack_skip_sacked(scb, &tp->rexmt_high);
 523
 524         /* Look for lost data. */
 525         torexmt = tp->rexmt_high;
 526         *lostdup = FALSE;
 527         if (lastblock != NULL) {
 528                 if (SEQ_LT(torexmt, lastblock->sblk_end) &&
 529                     scb_islost(scb, torexmt)) {
 530 sendunsacked:
 531                         *nextrexmt = torexmt;
 532                         /* If the left-hand edge has been SACKed, pull it in. */
 533                         if (sack_block_lookup(scb, torexmt + tp->t_maxseg, &sb))
 534                                 *plen = sb->sblk_start - torexmt;
 535                         else
 536                                 *plen = tp->t_maxseg;
 537                         return TRUE;
 538                 }
 539         }
 540
 541         /* See if unsent data available within send window. */
 542         off = tp->snd_max - tp->snd_una;
 543         len = (long) ulmin(so->so_snd.ssb_cc, tp->snd_wnd) - off;
 544         if (len > 0) {
 545                 *nextrexmt = tp->snd_max;       /* Send new data. */
 546                 *plen = tp->t_maxseg;
 547                 return TRUE;
 548         }
 549
 550         /* We're less certain this data has been lost. */
 551         if (lastblock == NULL || SEQ_LT(torexmt, lastblock->sblk_end))
 552                 goto sendunsacked;
 553
 554         return FALSE;
 555 }
 556
 557 /*
 558  * Return the next sequence number higher than "*prexmt" that has
 559  * not been SACKed.
 560  */
 561 void
 562 tcp_sack_skip_sacked(struct scoreboard *scb, tcp_seq *prexmt)
 563 {
 564         struct sackblock *sb;
 565
 566         /* skip SACKed data */
 567         if (sack_block_lookup(scb, *prexmt, &sb))
 568                 *prexmt = sb->sblk_end;
 569 }
 570
 571 #ifdef later
 572 void
 573 tcp_sack_save_scoreboard(struct scoreboard *scb)
 574 {
 575         struct scoreboard *scb = &tp->scb;
 576
 577         scb->sackblocks_prev = scb->sackblocks;
 578         TAILQ_INIT(&scb->sackblocks);
 579 }
 580
 581 void
 582 tcp_sack_revert_scoreboard(struct scoreboard *scb, tcp_seq snd_una,
 583                            u_int maxseg)
 584 {
 585         struct sackblock *sb;
 586
 587         scb->sackblocks = scb->sackblocks_prev;
 588         scb->nblocks = 0;
 589         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
 590                 ++scb->nblocks;
 591         tcp_sack_ack_blocks(scb, snd_una);
 592         scb->lastfound = NULL;
 593 }
 594 #endif
 595
 596 #ifdef DEBUG_SACK_HISTORY
 597 static void
 598 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
 599 {
 600         int i;
 601         static int ndumped;
 602
 603         /* only need a couple of these to debug most problems */
 604         if (++ndumped > 900)
 605                 return;
 606
 607         kprintf("%s:\tnsackhistory %d: ", msg, tp->nsackhistory);
 608         for (i = 0; i < tp->nsackhistory; ++i)
 609                 kprintf("[%u, %u) ", tp->sackhistory[i].rblk_start,
 610                     tp->sackhistory[i].rblk_end);
 611         kprintf("\n");
 612 }
 613 #else
 614 static __inline void
 615 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
 616 {
 617 }
 618 #endif
 619
 620 /*
 621  * Remove old SACK blocks from the SACK history that have already been ACKed.
 622  */
 623 static void
 624 tcp_sack_ack_history(struct tcpcb *tp)
 625 {
 626         int i, nblocks, openslot;
 627
 628         tcp_sack_dump_history("before tcp_sack_ack_history", tp);
 629         nblocks = tp->nsackhistory;
 630         for (i = openslot = 0; i < nblocks; ++i) {
 631                 if (SEQ_LEQ(tp->sackhistory[i].rblk_end, tp->rcv_nxt)) {
 632                         --tp->nsackhistory;
 633                         continue;
 634                 }
 635                 if (SEQ_LT(tp->sackhistory[i].rblk_start, tp->rcv_nxt))
 636                         tp->sackhistory[i].rblk_start = tp->rcv_nxt;
 637                 if (i == openslot)
 638                         ++openslot;
 639                 else
 640                         tp->sackhistory[openslot++] = tp->sackhistory[i];
 641         }
 642         tcp_sack_dump_history("after tcp_sack_ack_history", tp);
 643         KASSERT(openslot == tp->nsackhistory,
 644             ("tcp_sack_ack_history miscounted: %d != %d",
 645             openslot, tp->nsackhistory));
 646 }
 647
 648 /*
 649  * Add or merge newblock into reported history.
 650  * Also remove or update SACK blocks that will be acked.
 651  */
 652 static void
 653 tcp_sack_update_reported_history(struct tcpcb *tp, tcp_seq start, tcp_seq end)
 654 {
 655         struct raw_sackblock copy[MAX_SACK_REPORT_BLOCKS];
 656         int i, cindex;
 657
 658         tcp_sack_dump_history("before tcp_sack_update_reported_history", tp);
 659         /*
 660          * Six cases:
 661          *      0) no overlap
 662          *      1) newblock == oldblock
 663          *      2) oldblock contains newblock
 664          *      3) newblock contains oldblock
 665          *      4) tail of oldblock overlaps or abuts start of newblock
 666          *      5) tail of newblock overlaps or abuts head of oldblock
 667          */
 668         for (i = cindex = 0; i < tp->nsackhistory; ++i) {
 669                 struct raw_sackblock *oldblock = &tp->sackhistory[i];
 670                 tcp_seq old_start = oldblock->rblk_start;
 671                 tcp_seq old_end = oldblock->rblk_end;
 672
 673                 if (SEQ_LT(end, old_start) || SEQ_GT(start, old_end)) {
 674                         /* Case 0:  no overlap.  Copy old block. */
 675                         copy[cindex++] = *oldblock;
 676                         continue;
 677                 }
 678
 679                 if (SEQ_GEQ(start, old_start) && SEQ_LEQ(end, old_end)) {
 680                         /* Cases 1 & 2.  Move block to front of history. */
 681                         int j;
 682
 683                         start = old_start;
 684                         end = old_end;
 685                         /* no need to check rest of blocks */
 686                         for (j = i + 1; j < tp->nsackhistory; ++j)
 687                                 copy[cindex++] = tp->sackhistory[j];
 688                         break;
 689                 }
 690
 691                 if (SEQ_GEQ(old_end, start) && SEQ_LT(old_start, start)) {
 692                         /* Case 4:  extend start of new block. */
 693                         start = old_start;
 694                 } else if (SEQ_GEQ(end, old_start) && SEQ_GT(old_end, end)) {
 695                         /* Case 5: extend end of new block */
 696                         end = old_end;
 697                 } else {
 698                         /* Case 3.  Delete old block by not copying it. */
 699                         KASSERT(SEQ_LEQ(start, old_start) &&
 700                                 SEQ_GEQ(end, old_end),
 701                             ("bad logic: old [%u, %u), new [%u, %u)",
 702                              old_start, old_end, start, end));
 703                 }
 704         }
 705
 706         /* insert new block */
 707         tp->sackhistory[0].rblk_start = start;
 708         tp->sackhistory[0].rblk_end = end;
 709         cindex = min(cindex, MAX_SACK_REPORT_BLOCKS - 1);
 710         for (i = 0; i < cindex; ++i)
 711                 tp->sackhistory[i + 1] = copy[i];
 712         tp->nsackhistory = cindex + 1;
 713         tcp_sack_dump_history("after tcp_sack_update_reported_history", tp);
 714 }
 715
 716 /*
 717  * Fill in SACK report to return to data sender.
 718  */
 719 void
 720 tcp_sack_fill_report(struct tcpcb *tp, u_char *opt, u_int *plen)
 721 {
 722         u_int optlen = *plen;
 723         uint32_t *lp = (uint32_t *)(opt + optlen);
 724         uint32_t *olp;
 725         tcp_seq hstart = tp->rcv_nxt, hend;
 726         int nblocks;
 727
 728         KASSERT(TCP_MAXOLEN - optlen >=
 729             TCPOLEN_SACK_ALIGNED + TCPOLEN_SACK_BLOCK,
 730             ("no room for SACK header and one block: optlen %d", optlen));
 731
 732         if (tp->t_flags & TF_DUPSEG)
 733                 tcpstat.tcps_snddsackopt++;
 734         else
 735                 tcpstat.tcps_sndsackopt++;
 736
 737         olp = lp++;
 738         optlen += TCPOLEN_SACK_ALIGNED;
 739
 740         tcp_sack_ack_history(tp);
 741         if (tp->reportblk.rblk_start != tp->reportblk.rblk_end) {
 742                 *lp++ = htonl(tp->reportblk.rblk_start);
 743                 *lp++ = htonl(tp->reportblk.rblk_end);
 744                 optlen += TCPOLEN_SACK_BLOCK;
 745                 hstart = tp->reportblk.rblk_start;
 746                 hend = tp->reportblk.rblk_end;
 747                 if (tp->t_flags & TF_ENCLOSESEG) {
 748                         KASSERT(TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK,
 749                             ("no room for enclosing SACK block: oplen %d",
 750                             optlen));
 751                         *lp++ = htonl(tp->encloseblk.rblk_start);
 752                         *lp++ = htonl(tp->encloseblk.rblk_end);
 753                         optlen += TCPOLEN_SACK_BLOCK;
 754                         hstart = tp->encloseblk.rblk_start;
 755                         hend = tp->encloseblk.rblk_end;
 756                 }
 757                 if (SEQ_GT(hstart, tp->rcv_nxt))
 758                         tcp_sack_update_reported_history(tp, hstart, hend);
 759         }
 760         if (tcp_do_smartsack && (tp->t_flags & TF_SACKLEFT)) {
 761                 /* Fill in from left!  Walk re-assembly queue. */
 762                 struct tseg_qent *q;
 763
 764                 q = LIST_FIRST(&tp->t_segq);
 765                 while (q != NULL &&
 766                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
 767                         *lp++ = htonl(q->tqe_th->th_seq);
 768                         *lp++ = htonl(TCP_SACK_BLKEND(
 769                             q->tqe_th->th_seq + q->tqe_len,
 770                             q->tqe_th->th_flags));
 771                         optlen += TCPOLEN_SACK_BLOCK;
 772                         q = LIST_NEXT(q, tqe_q);
 773                 }
 774         } else {
 775                 int n = 0;
 776
 777                 /* Fill in SACK blocks from right side. */
 778                 while (n < tp->nsackhistory &&
 779                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
 780                         if (tp->sackhistory[n].rblk_start != hstart) {
 781                                 *lp++ = htonl(tp->sackhistory[n].rblk_start);
 782                                 *lp++ = htonl(tp->sackhistory[n].rblk_end);
 783                                 optlen += TCPOLEN_SACK_BLOCK;
 784                         }
 785                         ++n;
 786                 }
 787         }
 788         tp->reportblk.rblk_start = tp->reportblk.rblk_end;
 789         tp->t_flags &= ~(TF_DUPSEG | TF_ENCLOSESEG | TF_SACKLEFT);
 790         nblocks = (lp - olp - 1) / 2;
 791         *olp = htonl(TCPOPT_SACK_ALIGNED |
 792                      (TCPOLEN_SACK + nblocks * TCPOLEN_SACK_BLOCK));
 793         *plen = optlen;
 794 }