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