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 1.3 2006/01/14 11:33:50 swildner Exp $
  34  */
  35
  36 /*
  37  * Copyright (c) 2003, 2004 Jeffrey M. Hsu.  All rights reserved.
  38  *
  39  * License terms: all terms for the DragonFly license above plus the following:
  40  *
  41  * 4. All advertising materials mentioning features or use of this software
  42  *    must display the following acknowledgement:
  43  *
  44  *      This product includes software developed by Jeffrey M. Hsu
  45  *      for the DragonFly Project.
  46  *
  47  *    This requirement may be waived with permission from Jeffrey Hsu.
  48  *    This requirement will sunset and may be removed on November 13, 2005,
  49  *    after which the standard DragonFly license (as shown above) will
  50  *    apply.
  51  */
  52
  53 #include <sys/param.h>
  54 #include <sys/systm.h>
  55 #include <sys/kernel.h>
  56 #include <sys/malloc.h>
  57 #include <sys/queue.h>
  58 #include <sys/thread.h>
  59 #include <sys/types.h>
  60 #include <sys/socket.h>
  61 #include <sys/socketvar.h>
  62
  63 #include <net/if.h>
  64
  65 #include <netinet/in.h>
  66 #include <netinet/in_systm.h>
  67 #include <netinet/ip.h>
  68 #include <netinet/in_var.h>
  69 #include <netinet/in_pcb.h>
  70 #include <netinet/ip_var.h>
  71 #include <netinet/tcp.h>
  72 #include <netinet/tcp_seq.h>
  73 #include <netinet/tcp_var.h>
  74
  75 #include <vm/vm_zone.h>
  76
  77 struct sackblock {
  78         tcp_seq                 sblk_start;
  79         tcp_seq                 sblk_end;
  80         TAILQ_ENTRY(sackblock)  sblk_list;
  81 };
  82
  83 #define MAXSAVEDBLOCKS  8                       /* per connection limit */
  84
  85 static void insert_block(struct scoreboard *scb, struct sackblock *newblock);
  86 static void update_lostseq(struct scoreboard *scb, tcp_seq snd_una,
  87                            u_int maxseg);
  88
  89 static vm_zone_t sackblock_zone;
  90
  91 /*
  92  * Initialize SACK module.
  93  */
  94 void
  95 tcp_sack_init(void)
  96 {
  97         sackblock_zone = zinit("sack blocks", sizeof(struct sackblock),
  98             maxsockets, ZONE_INTERRUPT, 0);
  99 }
 100
 101 /*
 102  * Per-tcpcb initialization.
 103  */
 104 void
 105 tcp_sack_tcpcb_init(struct tcpcb *tp)
 106 {
 107         struct scoreboard *scb = &tp->scb;
 108
 109         scb->nblocks = 0;
 110         TAILQ_INIT(&scb->sackblocks);
 111         scb->lastfound = NULL;
 112 }
 113
 114 /*
 115  * Find the SACK block containing or immediately preceding "seq".
 116  * The boolean result indicates whether the sequence is actually
 117  * contained in the SACK block.
 118  */
 119 static boolean_t
 120 sack_block_lookup(struct scoreboard *scb, tcp_seq seq, struct sackblock **sb)
 121 {
 122         struct sackblock *hint = scb->lastfound;
 123         struct sackblock *cur, *last, *prev;
 124
 125         if (TAILQ_EMPTY(&scb->sackblocks)) {
 126                 *sb = NULL;
 127                 return FALSE;
 128         }
 129
 130         if (hint == NULL) {
 131                 /* No hint.  Search from start to end. */
 132                 cur = TAILQ_FIRST(&scb->sackblocks);
 133                 last = NULL;
 134                 prev = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 135         } else  {
 136                 if (SEQ_GEQ(seq, hint->sblk_start)) {
 137                         /* Search from hint to end of list. */
 138                         cur = hint;
 139                         last = NULL;
 140                         prev = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 141                 } else {
 142                         /* Search from front of list to hint. */
 143                         cur = TAILQ_FIRST(&scb->sackblocks);
 144                         last = hint;
 145                         prev = TAILQ_PREV(hint, sackblock_list, sblk_list);
 146                 }
 147         }
 148
 149         do {
 150                 if (SEQ_GT(cur->sblk_end, seq)) {
 151                         if (SEQ_GEQ(seq, cur->sblk_start)) {
 152                                 *sb = scb->lastfound = cur;
 153                                 return TRUE;
 154                         } else {
 155                                 *sb = scb->lastfound =
 156                                     TAILQ_PREV(cur, sackblock_list, sblk_list);
 157                                 return FALSE;
 158                         }
 159                 }
 160                 cur = TAILQ_NEXT(cur, sblk_list);
 161         } while (cur != last);
 162
 163         *sb = scb->lastfound = prev;
 164         return FALSE;
 165 }
 166
 167 /*
 168  * Allocate a SACK block.
 169  */
 170 static __inline struct sackblock *
 171 alloc_sackblock(void)
 172 {
 173         return zalloc(sackblock_zone);
 174 }
 175
 176 /*
 177  * Free a SACK block.
 178  */
 179 static __inline void
 180 free_sackblock(struct sackblock *s)
 181 {
 182         zfree(sackblock_zone, s);
 183 }
 184
 185 /*
 186  * Free up SACK blocks for data that's been acked.
 187  */
 188 static void
 189 tcp_sack_ack_blocks(struct scoreboard *scb, tcp_seq th_ack)
 190 {
 191         struct sackblock *sb, *nb;
 192
 193         sb = TAILQ_FIRST(&scb->sackblocks);
 194         while (sb && SEQ_LEQ(sb->sblk_end, th_ack)) {
 195                 nb = TAILQ_NEXT(sb, sblk_list);
 196                 if (sb == scb->lastfound)
 197                         scb->lastfound = NULL;
 198                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 199                 free_sackblock(sb);
 200                 --scb->nblocks;
 201                 KASSERT(scb->nblocks >= 0,
 202                     ("SACK block count underflow: %d < 0", scb->nblocks));
 203                 sb = nb;
 204         }
 205         if (sb && SEQ_GT(th_ack, sb->sblk_start))
 206                 sb->sblk_start = th_ack;        /* other side reneged? XXX */
 207 }
 208
 209 /*
 210  * Delete and free SACK blocks saved in scoreboard.
 211  */
 212 void
 213 tcp_sack_cleanup(struct scoreboard *scb)
 214 {
 215         struct sackblock *sb, *nb;
 216
 217         TAILQ_FOREACH_MUTABLE(sb, &scb->sackblocks, sblk_list, nb) {
 218                 free_sackblock(sb);
 219                 --scb->nblocks;
 220         }
 221         KASSERT(scb->nblocks == 0,
 222             ("SACK block %d count not zero", scb->nblocks));
 223         TAILQ_INIT(&scb->sackblocks);
 224         scb->lastfound = NULL;
 225 }
 226
 227 /*
 228  * Returns      0 if not D-SACK block,
 229  *              1 if D-SACK,
 230  *              2 if duplicate of out-of-order D-SACK block.
 231  */
 232 int
 233 tcp_sack_ndsack_blocks(struct raw_sackblock *blocks, const int numblocks,
 234                        tcp_seq snd_una)
 235 {
 236         if (numblocks == 0)
 237                 return 0;
 238
 239         if (SEQ_LT(blocks[0].rblk_start, snd_una))
 240                 return 1;
 241
 242         /* block 0 inside block 1 */
 243         if (numblocks > 1 &&
 244             SEQ_GEQ(blocks[0].rblk_start, blocks[1].rblk_start) &&
 245             SEQ_LEQ(blocks[0].rblk_end, blocks[1].rblk_end))
 246                 return 2;
 247
 248         return 0;
 249 }
 250
 251 /*
 252  * Update scoreboard on new incoming ACK.
 253  */
 254 static void
 255 tcp_sack_add_blocks(struct tcpcb *tp, struct tcpopt *to)
 256 {
 257         const int numblocks = to->to_nsackblocks;
 258         struct raw_sackblock *blocks = to->to_sackblocks;
 259         struct scoreboard *scb = &tp->scb;
 260         struct sackblock *sb;
 261         int startblock;
 262         int i;
 263
 264         if (tcp_sack_ndsack_blocks(blocks, numblocks, tp->snd_una) > 0)
 265                 startblock = 1;
 266         else
 267                 startblock = 0;
 268
 269         for (i = startblock; i < numblocks; i++) {
 270                 struct raw_sackblock *newsackblock = &blocks[i];
 271
 272                 /* don't accept bad SACK blocks */
 273                 if (SEQ_GT(newsackblock->rblk_end, tp->snd_max))
 274                         break;          /* skip all other blocks */
 275
 276                 sb = alloc_sackblock();
 277                 if (sb == NULL)         /* do some sort of cleanup? XXX */
 278                         break;          /* just skip rest of blocks */
 279                 sb->sblk_start = newsackblock->rblk_start;
 280                 sb->sblk_end = newsackblock->rblk_end;
 281                 if (TAILQ_EMPTY(&scb->sackblocks)) {
 282                         KASSERT(scb->nblocks == 0, ("emply scb w/ blocks"));
 283                         scb->nblocks = 1;
 284                         TAILQ_INSERT_HEAD(&scb->sackblocks, sb, sblk_list);
 285                 } else {
 286                         insert_block(scb, sb);
 287                 }
 288         }
 289 }
 290
 291 void
 292 tcp_sack_update_scoreboard(struct tcpcb *tp, struct tcpopt *to)
 293 {
 294         struct scoreboard *scb = &tp->scb;
 295
 296         tcp_sack_ack_blocks(scb, tp->snd_una);
 297         tcp_sack_add_blocks(tp, to);
 298         update_lostseq(scb, tp->snd_una, tp->t_maxseg);
 299         if (SEQ_LT(tp->rexmt_high, tp->snd_una))
 300                 tp->rexmt_high = tp->snd_una;
 301 }
 302
 303 /*
 304  * Insert SACK block into sender's scoreboard.
 305  */
 306 static void
 307 insert_block(struct scoreboard *scb, struct sackblock *newblock)
 308 {
 309         struct sackblock *sb, *workingblock;
 310         boolean_t overlap_front;
 311
 312         KASSERT(scb->nblocks > 0, ("insert_block() called w/ no blocks"));
 313
 314         if (scb->nblocks == MAXSAVEDBLOCKS) {
 315                 /*
 316                  * Should try to kick out older blocks XXX JH
 317                  * May be able to coalesce with existing block.
 318                  * Or, go other way and free all blocks if we hit this limit.
 319                  */
 320                 free_sackblock(newblock);
 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                 } else {
 340                         workingblock = newblock;
 341                         TAILQ_INSERT_AFTER(&scb->sackblocks, sb, newblock,
 342                                            sblk_list);
 343                         ++scb->nblocks;
 344                 }
 345         }
 346
 347         /* Consolidate right-hand side. */
 348         sb = TAILQ_NEXT(workingblock, sblk_list);
 349         while (sb != NULL &&
 350             SEQ_GEQ(workingblock->sblk_end, sb->sblk_end)) {
 351                 struct sackblock *nextblock;
 352
 353                 nextblock = TAILQ_NEXT(sb, sblk_list);
 354                 /* Remove completely overlapped block */
 355                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 356                 free_sackblock(sb);
 357                 --scb->nblocks;
 358                 KASSERT(scb->nblocks > 0,
 359                     ("removed overlapped block: %d blocks left", scb->nblocks));
 360                 sb = nextblock;
 361         }
 362         if (sb != NULL &&
 363             SEQ_GEQ(workingblock->sblk_end, sb->sblk_start)) {
 364                 /* Extend new block to cover partially overlapped old block. */
 365                 workingblock->sblk_end = sb->sblk_end;
 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         printf("%d blocks:", scb->nblocks);
 381         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
 382                 printf(" [%u, %u)", sb->sblk_start, sb->sblk_end);
 383         printf("\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.sb_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         printf("%s:\tnsackhistory %d: ", msg, tp->nsackhistory);
 588         for (i = 0; i < tp->nsackhistory; ++i)
 589                 printf("[%u, %u) ", tp->sackhistory[i].rblk_start,
 590                     tp->sackhistory[i].rblk_end);
 591         printf("\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         olp = lp++;
 713         optlen += TCPOLEN_SACK_ALIGNED;
 714
 715         tcp_sack_ack_history(tp);
 716         if (tp->reportblk.rblk_start != tp->reportblk.rblk_end) {
 717                 *lp++ = htonl(tp->reportblk.rblk_start);
 718                 *lp++ = htonl(tp->reportblk.rblk_end);
 719                 optlen += TCPOLEN_SACK_BLOCK;
 720                 hstart = tp->reportblk.rblk_start;
 721                 hend = tp->reportblk.rblk_end;
 722                 if (tp->t_flags & TF_ENCLOSESEG) {
 723                         KASSERT(TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK,
 724                             ("no room for enclosing SACK block: oplen %d",
 725                             optlen));
 726                         *lp++ = htonl(tp->encloseblk.rblk_start);
 727                         *lp++ = htonl(tp->encloseblk.rblk_end);
 728                         optlen += TCPOLEN_SACK_BLOCK;
 729                         hstart = tp->encloseblk.rblk_start;
 730                         hend = tp->encloseblk.rblk_end;
 731                 }
 732                 if (SEQ_GT(hstart, tp->rcv_nxt))
 733                         tcp_sack_update_reported_history(tp, hstart, hend);
 734         }
 735         if (tcp_do_smartsack && (tp->t_flags & TF_SACKLEFT)) {
 736                 /* Fill in from left!  Walk re-assembly queue. */
 737                 struct tseg_qent *q;
 738
 739                 q = LIST_FIRST(&tp->t_segq);
 740                 while (q != NULL &&
 741                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
 742                         *lp++ = htonl(q->tqe_th->th_seq);
 743                         *lp++ = htonl(q->tqe_th->th_seq + q->tqe_len);
 744                         optlen += TCPOLEN_SACK_BLOCK;
 745                         q = LIST_NEXT(q, tqe_q);
 746                 }
 747         } else {
 748                 int n = 0;
 749
 750                 /* Fill in SACK blocks from right side. */
 751                 while (n < tp->nsackhistory &&
 752                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
 753                         if (tp->sackhistory[n].rblk_start != hstart) {
 754                                 *lp++ = htonl(tp->sackhistory[n].rblk_start);
 755                                 *lp++ = htonl(tp->sackhistory[n].rblk_end);
 756                                 optlen += TCPOLEN_SACK_BLOCK;
 757                         }
 758                         ++n;
 759                 }
 760         }
 761         tp->reportblk.rblk_start = tp->reportblk.rblk_end;
 762         tp->t_flags &= ~(TF_DUPSEG | TF_ENCLOSESEG | TF_SACKLEFT);
 763         nblocks = (lp - olp - 1) / 2;
 764         *olp = htonl(TCPOPT_SACK_ALIGNED |
 765                      (TCPOLEN_SACK + nblocks * TCPOLEN_SACK_BLOCK));
 766         *plen = optlen;
 767 }