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, boolean_t *update);
  76 static void update_lostseq(struct scoreboard *scb, tcp_seq snd_una,
  77                            u_int maxseg, int rxtthresh);
  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(struct scoreboard *scb, const struct raw_sackblock *raw_sb)
 152 {
 153         struct sackblock *sb;
 154
 155         if (scb->freecache != NULL) {
 156                 sb = scb->freecache;
 157                 scb->freecache = NULL;
 158                 tcpstat.tcps_sacksbfast++;
 159         } else {
 160                 sb = kmalloc(sizeof(struct sackblock), M_SACKBLOCK, M_NOWAIT);
 161                 if (sb == NULL) {
 162                         tcpstat.tcps_sacksbfailed++;
 163                         return NULL;
 164                 }
 165         }
 166         sb->sblk_start = raw_sb->rblk_start;
 167         sb->sblk_end = raw_sb->rblk_end;
 168         return sb;
 169 }
 170
 171 static __inline struct sackblock *
 172 alloc_sackblock_limit(struct scoreboard *scb,
 173     const struct raw_sackblock *raw_sb)
 174 {
 175         if (scb->nblocks == MAXSAVEDBLOCKS) {
 176                 /*
 177                  * Should try to kick out older blocks XXX JH
 178                  * May be able to coalesce with existing block.
 179                  * Or, go other way and free all blocks if we hit
 180                  * this limit.
 181                  */
 182                 tcpstat.tcps_sacksboverflow++;
 183                 return NULL;
 184         }
 185         return alloc_sackblock(scb, raw_sb);
 186 }
 187
 188 /*
 189  * Free a SACK block.
 190  */
 191 static __inline void
 192 free_sackblock(struct scoreboard *scb, struct sackblock *s)
 193 {
 194         if (scb->freecache == NULL) {
 195                 /* YYY Maybe use the latest freed block? */
 196                 scb->freecache = s;
 197                 return;
 198         }
 199         kfree(s, M_SACKBLOCK);
 200 }
 201
 202 /*
 203  * Free up SACK blocks for data that's been acked.
 204  */
 205 static void
 206 tcp_sack_ack_blocks(struct scoreboard *scb, tcp_seq th_ack)
 207 {
 208         struct sackblock *sb, *nb;
 209
 210         sb = TAILQ_FIRST(&scb->sackblocks);
 211         while (sb && SEQ_LEQ(sb->sblk_end, th_ack)) {
 212                 nb = TAILQ_NEXT(sb, sblk_list);
 213                 if (scb->lastfound == sb)
 214                         scb->lastfound = NULL;
 215                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 216                 free_sackblock(scb, sb);
 217                 --scb->nblocks;
 218                 KASSERT(scb->nblocks >= 0,
 219                     ("SACK block count underflow: %d < 0", scb->nblocks));
 220                 sb = nb;
 221         }
 222         if (sb && SEQ_GT(th_ack, sb->sblk_start))
 223                 sb->sblk_start = th_ack;        /* other side reneged? XXX */
 224 }
 225
 226 /*
 227  * Delete and free SACK blocks saved in scoreboard.
 228  */
 229 void
 230 tcp_sack_cleanup(struct scoreboard *scb)
 231 {
 232         struct sackblock *sb, *nb;
 233
 234         TAILQ_FOREACH_MUTABLE(sb, &scb->sackblocks, sblk_list, nb) {
 235                 free_sackblock(scb, sb);
 236                 --scb->nblocks;
 237         }
 238         KASSERT(scb->nblocks == 0,
 239             ("SACK block %d count not zero", scb->nblocks));
 240         TAILQ_INIT(&scb->sackblocks);
 241         scb->lastfound = NULL;
 242 }
 243
 244 /*
 245  * Delete and free SACK blocks saved in scoreboard.
 246  * Delete the one slot block cache.
 247  */
 248 void
 249 tcp_sack_destroy(struct scoreboard *scb)
 250 {
 251         tcp_sack_cleanup(scb);
 252         if (scb->freecache != NULL) {
 253                 kfree(scb->freecache, M_SACKBLOCK);
 254                 scb->freecache = NULL;
 255         }
 256 }
 257
 258 /*
 259  * Cleanup the reported SACK block information
 260  */
 261 void
 262 tcp_sack_report_cleanup(struct tcpcb *tp)
 263 {
 264         tp->t_flags &= ~(TF_DUPSEG | TF_ENCLOSESEG | TF_SACKLEFT);
 265         tp->reportblk.rblk_start = tp->reportblk.rblk_end;
 266 }
 267
 268 /*
 269  * Returns      0 if not D-SACK block,
 270  *              1 if D-SACK,
 271  *              2 if duplicate of out-of-order D-SACK block.
 272  */
 273 int
 274 tcp_sack_ndsack_blocks(struct raw_sackblock *blocks, const int numblocks,
 275                        tcp_seq snd_una)
 276 {
 277         if (numblocks == 0)
 278                 return 0;
 279
 280         if (SEQ_LT(blocks[0].rblk_start, snd_una))
 281                 return 1;
 282
 283         /* block 0 inside block 1 */
 284         if (numblocks > 1 &&
 285             SEQ_GEQ(blocks[0].rblk_start, blocks[1].rblk_start) &&
 286             SEQ_LEQ(blocks[0].rblk_end, blocks[1].rblk_end))
 287                 return 2;
 288
 289         return 0;
 290 }
 291
 292 /*
 293  * Update scoreboard on new incoming ACK.
 294  */
 295 static void
 296 tcp_sack_add_blocks(struct tcpcb *tp, struct tcpopt *to)
 297 {
 298         const int numblocks = to->to_nsackblocks;
 299         struct raw_sackblock *blocks = to->to_sackblocks;
 300         struct scoreboard *scb = &tp->scb;
 301         int startblock, i;
 302
 303         if (tcp_sack_ndsack_blocks(blocks, numblocks, tp->snd_una) > 0)
 304                 startblock = 1;
 305         else
 306                 startblock = 0;
 307
 308         to->to_flags |= TOF_SACK_REDUNDANT;
 309         for (i = startblock; i < numblocks; i++) {
 310                 struct raw_sackblock *newsackblock = &blocks[i];
 311                 boolean_t update;
 312                 int error;
 313
 314                 /* don't accept bad SACK blocks */
 315                 if (SEQ_GT(newsackblock->rblk_end, tp->snd_max)) {
 316                         tcpstat.tcps_rcvbadsackopt++;
 317                         break;          /* skip all other blocks */
 318                 }
 319                 tcpstat.tcps_sacksbupdate++;
 320
 321                 error = insert_block(scb, newsackblock, &update);
 322                 if (update)
 323                         to->to_flags &= ~TOF_SACK_REDUNDANT;
 324                 if (error)
 325                         break;
 326         }
 327 }
 328
 329 void
 330 tcp_sack_update_scoreboard(struct tcpcb *tp, struct tcpopt *to)
 331 {
 332         struct scoreboard *scb = &tp->scb;
 333         int rexmt_high_update = 0;
 334
 335         tcp_sack_ack_blocks(scb, tp->snd_una);
 336         tcp_sack_add_blocks(tp, to);
 337         update_lostseq(scb, tp->snd_una, tp->t_maxseg, tp->t_rxtthresh);
 338         if (SEQ_LT(tp->rexmt_high, tp->snd_una)) {
 339                 tp->rexmt_high = tp->snd_una;
 340                 rexmt_high_update = 1;
 341         }
 342         if (tp->t_flags & TF_SACKRESCUED) {
 343                 if (SEQ_LT(tp->rexmt_rescue, tp->snd_una)) {
 344                         tp->t_flags &= ~TF_SACKRESCUED;
 345                 } else if (tcp_aggressive_rescuesack && rexmt_high_update &&
 346                     SEQ_LT(tp->rexmt_rescue, tp->rexmt_high)) {
 347                         /* Drag RescueRxt along with HighRxt */
 348                         tp->rexmt_rescue = tp->rexmt_high;
 349                 }
 350         }
 351 }
 352
 353 /*
 354  * Insert SACK block into sender's scoreboard.
 355  */
 356 static int
 357 insert_block(struct scoreboard *scb, const struct raw_sackblock *raw_sb,
 358     boolean_t *update)
 359 {
 360         struct sackblock *sb, *workingblock;
 361         boolean_t overlap_front;
 362
 363         *update = TRUE;
 364         if (TAILQ_EMPTY(&scb->sackblocks)) {
 365                 struct sackblock *newblock;
 366
 367                 KASSERT(scb->nblocks == 0, ("emply scb w/ blocks"));
 368
 369                 newblock = alloc_sackblock(scb, raw_sb);
 370                 if (newblock == NULL)
 371                         return ENOMEM;
 372                 TAILQ_INSERT_HEAD(&scb->sackblocks, newblock, sblk_list);
 373                 scb->nblocks = 1;
 374                 return 0;
 375         }
 376
 377         KASSERT(scb->nblocks > 0, ("insert_block() called w/ no blocks"));
 378         KASSERT(scb->nblocks <= MAXSAVEDBLOCKS,
 379             ("too many SACK blocks %d", scb->nblocks));
 380
 381         overlap_front = sack_block_lookup(scb, raw_sb->rblk_start, &sb);
 382
 383         if (sb == NULL) {
 384                 workingblock = alloc_sackblock_limit(scb, raw_sb);
 385                 if (workingblock == NULL)
 386                         return ENOMEM;
 387                 TAILQ_INSERT_HEAD(&scb->sackblocks, workingblock, sblk_list);
 388                 ++scb->nblocks;
 389         } else {
 390                 if (overlap_front || sb->sblk_end == raw_sb->rblk_start) {
 391                         /* Extend old block */
 392                         workingblock = sb;
 393                         if (SEQ_GT(raw_sb->rblk_end, sb->sblk_end))
 394                                 sb->sblk_end = raw_sb->rblk_end;
 395                         else
 396                                 *update = FALSE;
 397                         tcpstat.tcps_sacksbreused++;
 398                 } else {
 399                         workingblock = alloc_sackblock_limit(scb, raw_sb);
 400                         if (workingblock == NULL)
 401                                 return ENOMEM;
 402                         TAILQ_INSERT_AFTER(&scb->sackblocks, sb, workingblock,
 403                             sblk_list);
 404                         ++scb->nblocks;
 405                 }
 406         }
 407
 408         /* Consolidate right-hand side. */
 409         sb = TAILQ_NEXT(workingblock, sblk_list);
 410         while (sb != NULL &&
 411             SEQ_GEQ(workingblock->sblk_end, sb->sblk_end)) {
 412                 struct sackblock *nextblock;
 413
 414                 nextblock = TAILQ_NEXT(sb, sblk_list);
 415                 if (scb->lastfound == sb)
 416                         scb->lastfound = NULL;
 417                 /* Remove completely overlapped block */
 418                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 419                 free_sackblock(scb, sb);
 420                 --scb->nblocks;
 421                 KASSERT(scb->nblocks > 0,
 422                     ("removed overlapped block: %d blocks left", scb->nblocks));
 423                 sb = nextblock;
 424         }
 425         if (sb != NULL &&
 426             SEQ_GEQ(workingblock->sblk_end, sb->sblk_start)) {
 427                 /* Extend new block to cover partially overlapped old block. */
 428                 workingblock->sblk_end = sb->sblk_end;
 429                 if (scb->lastfound == sb)
 430                         scb->lastfound = NULL;
 431                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 432                 free_sackblock(scb, sb);
 433                 --scb->nblocks;
 434                 KASSERT(scb->nblocks > 0,
 435                     ("removed partial right: %d blocks left", scb->nblocks));
 436         }
 437         return 0;
 438 }
 439
 440 #ifdef DEBUG_SACK_BLOCKS
 441 static void
 442 tcp_sack_dump_blocks(struct scoreboard *scb)
 443 {
 444         struct sackblock *sb;
 445
 446         kprintf("%d blocks:", scb->nblocks);
 447         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
 448                 kprintf(" [%u, %u)", sb->sblk_start, sb->sblk_end);
 449         kprintf("\n");
 450 }
 451 #else
 452 static __inline void
 453 tcp_sack_dump_blocks(struct scoreboard *scb)
 454 {
 455 }
 456 #endif
 457
 458 /*
 459  * Optimization to quickly determine which packets are lost.
 460  */
 461 static void
 462 update_lostseq(struct scoreboard *scb, tcp_seq snd_una, u_int maxseg,
 463     int rxtthresh)
 464 {
 465         struct sackblock *sb;
 466         int nsackblocks = 0;
 467         int bytes_sacked = 0;
 468
 469         sb = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 470         while (sb != NULL) {
 471                 ++nsackblocks;
 472                 bytes_sacked += sb->sblk_end - sb->sblk_start;
 473                 if (nsackblocks == rxtthresh ||
 474                     bytes_sacked >= rxtthresh * maxseg) {
 475                         scb->lostseq = sb->sblk_start;
 476                         return;
 477                 }
 478                 sb = TAILQ_PREV(sb, sackblock_list, sblk_list);
 479         }
 480         scb->lostseq = snd_una;
 481 }
 482
 483 /*
 484  * Return whether the given sequence number is considered lost.
 485  */
 486 static boolean_t
 487 scb_islost(struct scoreboard *scb, tcp_seq seqnum)
 488 {
 489         return SEQ_LT(seqnum, scb->lostseq);
 490 }
 491
 492 /*
 493  * True if at least "amount" has been SACKed.  Used by Early Retransmit.
 494  */
 495 boolean_t
 496 tcp_sack_has_sacked(struct scoreboard *scb, u_int amount)
 497 {
 498         struct sackblock *sb;
 499         int bytes_sacked = 0;
 500
 501         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
 502                 bytes_sacked += sb->sblk_end - sb->sblk_start;
 503                 if (bytes_sacked >= amount)
 504                         return TRUE;
 505         }
 506         return FALSE;
 507 }
 508
 509 /*
 510  * Number of bytes SACKed below seq.
 511  */
 512 int
 513 tcp_sack_bytes_below(struct scoreboard *scb, tcp_seq seq)
 514 {
 515         struct sackblock *sb;
 516         int bytes_sacked = 0;
 517
 518         sb = TAILQ_FIRST(&scb->sackblocks);
 519         while (sb && SEQ_GT(seq, sb->sblk_start)) {
 520                 bytes_sacked += seq_min(seq, sb->sblk_end) - sb->sblk_start;
 521                 sb = TAILQ_NEXT(sb, sblk_list);
 522         }
 523         return bytes_sacked;
 524 }
 525
 526 /*
 527  * Return estimate of the number of bytes outstanding in the network.
 528  */
 529 uint32_t
 530 tcp_sack_compute_pipe(struct tcpcb *tp)
 531 {
 532         struct scoreboard *scb = &tp->scb;
 533         struct sackblock *sb;
 534         int nlost, nretransmitted;
 535         tcp_seq end;
 536
 537         nlost = tp->snd_max - scb->lostseq;
 538         nretransmitted = tp->rexmt_high - tp->snd_una;
 539
 540         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
 541                 if (SEQ_LT(sb->sblk_start, tp->rexmt_high)) {
 542                         end = seq_min(sb->sblk_end, tp->rexmt_high);
 543                         nretransmitted -= end - sb->sblk_start;
 544                 }
 545                 if (SEQ_GEQ(sb->sblk_start, scb->lostseq))
 546                         nlost -= sb->sblk_end - sb->sblk_start;
 547         }
 548
 549         return (nlost + nretransmitted);
 550 }
 551
 552 /*
 553  * Return the sequence number and length of the next segment to transmit
 554  * when in Fast Recovery.
 555  */
 556 boolean_t
 557 tcp_sack_nextseg(struct tcpcb *tp, tcp_seq *nextrexmt, uint32_t *plen,
 558     boolean_t *rescue)
 559 {
 560         struct scoreboard *scb = &tp->scb;
 561         struct socket *so = tp->t_inpcb->inp_socket;
 562         struct sackblock *sb;
 563         const struct sackblock *lastblock =
 564             TAILQ_LAST(&scb->sackblocks, sackblock_list);
 565         tcp_seq torexmt;
 566         long len, off;
 567
 568         /* skip SACKed data */
 569         tcp_sack_skip_sacked(scb, &tp->rexmt_high);
 570
 571         /* Look for lost data. */
 572         torexmt = tp->rexmt_high;
 573         *rescue = FALSE;
 574         if (lastblock != NULL) {
 575                 if (SEQ_LT(torexmt, lastblock->sblk_end) &&
 576                     scb_islost(scb, torexmt)) {
 577 sendunsacked:
 578                         *nextrexmt = torexmt;
 579                         /* If the left-hand edge has been SACKed, pull it in. */
 580                         if (sack_block_lookup(scb, torexmt + tp->t_maxseg, &sb))
 581                                 *plen = sb->sblk_start - torexmt;
 582                         else
 583                                 *plen = tp->t_maxseg;
 584                         return TRUE;
 585                 }
 586         }
 587
 588         /* See if unsent data available within send window. */
 589         off = tp->snd_max - tp->snd_una;
 590         len = (long) ulmin(so->so_snd.ssb_cc, tp->snd_wnd) - off;
 591         if (len > 0) {
 592                 *nextrexmt = tp->snd_max;       /* Send new data. */
 593                 *plen = tp->t_maxseg;
 594                 return TRUE;
 595         }
 596
 597         /* We're less certain this data has been lost. */
 598         if (lastblock != NULL && SEQ_LT(torexmt, lastblock->sblk_end))
 599                 goto sendunsacked;
 600
 601         /* Rescue retransmission */
 602         if (tcp_do_rescuesack) {
 603                 tcpstat.tcps_sackrescue_try++;
 604                 if (tp->t_flags & TF_SACKRESCUED) {
 605                         if (!tcp_aggressive_rescuesack)
 606                                 return FALSE;
 607
 608                         /*
 609                          * Aggressive variant of the rescue retransmission.
 610                          *
 611                          * The idea of the rescue retransmission is to sustain
 612                          * the ACK clock thus to avoid timeout retransmission.
 613                          *
 614                          * Under some situations, the conservative approach
 615                          * suggested in the draft
 616                          * http://tools.ietf.org/html/
 617                          * draft-nishida-tcpm-rescue-retransmission-00
 618                          * could not sustain ACK clock, since it only allows
 619                          * one rescue retransmission before a cumulative ACK
 620                          * covers the segement transmitted by rescue
 621                          * retransmission.
 622                          *
 623                          * We try to locate the next unSACKed segment which
 624                          * follows the previously sent rescue segment.  If
 625                          * there is no such segment, we loop back to the first
 626                          * unacknowledged segment.
 627                          */
 628
 629                         /*
 630                          * Skip SACKed data, but here we follow
 631                          * the last transmitted rescue segment.
 632                          */
 633                         torexmt = tp->rexmt_rescue;
 634                         tcp_sack_skip_sacked(scb, &torexmt);
 635                         if (torexmt == tp->snd_max) {
 636                                 /* Nothing left to retransmit; restart */
 637                                 torexmt = tp->snd_una;
 638                         }
 639                 }
 640                 *rescue = TRUE;
 641                 goto sendunsacked;
 642         } else if (tcp_do_smartsack && lastblock == NULL) {
 643                 tcpstat.tcps_sackrescue_try++;
 644                 *rescue = TRUE;
 645                 goto sendunsacked;
 646         }
 647
 648         return FALSE;
 649 }
 650
 651 /*
 652  * Return the next sequence number higher than "*prexmt" that has
 653  * not been SACKed.
 654  */
 655 void
 656 tcp_sack_skip_sacked(struct scoreboard *scb, tcp_seq *prexmt)
 657 {
 658         struct sackblock *sb;
 659
 660         /* skip SACKed data */
 661         if (sack_block_lookup(scb, *prexmt, &sb))
 662                 *prexmt = sb->sblk_end;
 663 }
 664
 665 #ifdef later
 666 void
 667 tcp_sack_save_scoreboard(struct scoreboard *scb)
 668 {
 669         struct scoreboard *scb = &tp->scb;
 670
 671         scb->sackblocks_prev = scb->sackblocks;
 672         TAILQ_INIT(&scb->sackblocks);
 673 }
 674
 675 void
 676 tcp_sack_revert_scoreboard(struct scoreboard *scb, tcp_seq snd_una,
 677                            u_int maxseg)
 678 {
 679         struct sackblock *sb;
 680
 681         scb->sackblocks = scb->sackblocks_prev;
 682         scb->nblocks = 0;
 683         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
 684                 ++scb->nblocks;
 685         tcp_sack_ack_blocks(scb, snd_una);
 686         scb->lastfound = NULL;
 687 }
 688 #endif
 689
 690 #ifdef DEBUG_SACK_HISTORY
 691 static void
 692 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
 693 {
 694         int i;
 695         static int ndumped;
 696
 697         /* only need a couple of these to debug most problems */
 698         if (++ndumped > 900)
 699                 return;
 700
 701         kprintf("%s:\tnsackhistory %d: ", msg, tp->nsackhistory);
 702         for (i = 0; i < tp->nsackhistory; ++i)
 703                 kprintf("[%u, %u) ", tp->sackhistory[i].rblk_start,
 704                     tp->sackhistory[i].rblk_end);
 705         kprintf("\n");
 706 }
 707 #else
 708 static __inline void
 709 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
 710 {
 711 }
 712 #endif
 713
 714 /*
 715  * Remove old SACK blocks from the SACK history that have already been ACKed.
 716  */
 717 static void
 718 tcp_sack_ack_history(struct tcpcb *tp)
 719 {
 720         int i, nblocks, openslot;
 721
 722         tcp_sack_dump_history("before tcp_sack_ack_history", tp);
 723         nblocks = tp->nsackhistory;
 724         for (i = openslot = 0; i < nblocks; ++i) {
 725                 if (SEQ_LEQ(tp->sackhistory[i].rblk_end, tp->rcv_nxt)) {
 726                         --tp->nsackhistory;
 727                         continue;
 728                 }
 729                 if (SEQ_LT(tp->sackhistory[i].rblk_start, tp->rcv_nxt))
 730                         tp->sackhistory[i].rblk_start = tp->rcv_nxt;
 731                 if (i == openslot)
 732                         ++openslot;
 733                 else
 734                         tp->sackhistory[openslot++] = tp->sackhistory[i];
 735         }
 736         tcp_sack_dump_history("after tcp_sack_ack_history", tp);
 737         KASSERT(openslot == tp->nsackhistory,
 738             ("tcp_sack_ack_history miscounted: %d != %d",
 739             openslot, tp->nsackhistory));
 740 }
 741
 742 /*
 743  * Add or merge newblock into reported history.
 744  * Also remove or update SACK blocks that will be acked.
 745  */
 746 static void
 747 tcp_sack_update_reported_history(struct tcpcb *tp, tcp_seq start, tcp_seq end)
 748 {
 749         struct raw_sackblock copy[MAX_SACK_REPORT_BLOCKS];
 750         int i, cindex;
 751
 752         tcp_sack_dump_history("before tcp_sack_update_reported_history", tp);
 753         /*
 754          * Six cases:
 755          *      0) no overlap
 756          *      1) newblock == oldblock
 757          *      2) oldblock contains newblock
 758          *      3) newblock contains oldblock
 759          *      4) tail of oldblock overlaps or abuts start of newblock
 760          *      5) tail of newblock overlaps or abuts head of oldblock
 761          */
 762         for (i = cindex = 0; i < tp->nsackhistory; ++i) {
 763                 struct raw_sackblock *oldblock = &tp->sackhistory[i];
 764                 tcp_seq old_start = oldblock->rblk_start;
 765                 tcp_seq old_end = oldblock->rblk_end;
 766
 767                 if (SEQ_LT(end, old_start) || SEQ_GT(start, old_end)) {
 768                         /* Case 0:  no overlap.  Copy old block. */
 769                         copy[cindex++] = *oldblock;
 770                         continue;
 771                 }
 772
 773                 if (SEQ_GEQ(start, old_start) && SEQ_LEQ(end, old_end)) {
 774                         /* Cases 1 & 2.  Move block to front of history. */
 775                         int j;
 776
 777                         start = old_start;
 778                         end = old_end;
 779                         /* no need to check rest of blocks */
 780                         for (j = i + 1; j < tp->nsackhistory; ++j)
 781                                 copy[cindex++] = tp->sackhistory[j];
 782                         break;
 783                 }
 784
 785                 if (SEQ_GEQ(old_end, start) && SEQ_LT(old_start, start)) {
 786                         /* Case 4:  extend start of new block. */
 787                         start = old_start;
 788                 } else if (SEQ_GEQ(end, old_start) && SEQ_GT(old_end, end)) {
 789                         /* Case 5: extend end of new block */
 790                         end = old_end;
 791                 } else {
 792                         /* Case 3.  Delete old block by not copying it. */
 793                         KASSERT(SEQ_LEQ(start, old_start) &&
 794                                 SEQ_GEQ(end, old_end),
 795                             ("bad logic: old [%u, %u), new [%u, %u)",
 796                              old_start, old_end, start, end));
 797                 }
 798         }
 799
 800         /* insert new block */
 801         tp->sackhistory[0].rblk_start = start;
 802         tp->sackhistory[0].rblk_end = end;
 803         cindex = min(cindex, MAX_SACK_REPORT_BLOCKS - 1);
 804         for (i = 0; i < cindex; ++i)
 805                 tp->sackhistory[i + 1] = copy[i];
 806         tp->nsackhistory = cindex + 1;
 807         tcp_sack_dump_history("after tcp_sack_update_reported_history", tp);
 808 }
 809
 810 /*
 811  * Fill in SACK report to return to data sender.
 812  */
 813 void
 814 tcp_sack_fill_report(struct tcpcb *tp, u_char *opt, u_int *plen)
 815 {
 816         u_int optlen = *plen;
 817         uint32_t *lp = (uint32_t *)(opt + optlen);
 818         uint32_t *olp;
 819         tcp_seq hstart = tp->rcv_nxt, hend;
 820         int nblocks;
 821
 822         KASSERT(TCP_MAXOLEN - optlen >=
 823             TCPOLEN_SACK_ALIGNED + TCPOLEN_SACK_BLOCK,
 824             ("no room for SACK header and one block: optlen %d", optlen));
 825
 826         if (tp->t_flags & TF_DUPSEG)
 827                 tcpstat.tcps_snddsackopt++;
 828         else
 829                 tcpstat.tcps_sndsackopt++;
 830
 831         olp = lp++;
 832         optlen += TCPOLEN_SACK_ALIGNED;
 833
 834         tcp_sack_ack_history(tp);
 835         if (tp->reportblk.rblk_start != tp->reportblk.rblk_end) {
 836                 *lp++ = htonl(tp->reportblk.rblk_start);
 837                 *lp++ = htonl(tp->reportblk.rblk_end);
 838                 optlen += TCPOLEN_SACK_BLOCK;
 839                 hstart = tp->reportblk.rblk_start;
 840                 hend = tp->reportblk.rblk_end;
 841                 if (tp->t_flags & TF_ENCLOSESEG) {
 842                         KASSERT(TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK,
 843                             ("no room for enclosing SACK block: oplen %d",
 844                             optlen));
 845                         *lp++ = htonl(tp->encloseblk.rblk_start);
 846                         *lp++ = htonl(tp->encloseblk.rblk_end);
 847                         optlen += TCPOLEN_SACK_BLOCK;
 848                         hstart = tp->encloseblk.rblk_start;
 849                         hend = tp->encloseblk.rblk_end;
 850                 }
 851                 if (SEQ_GT(hstart, tp->rcv_nxt))
 852                         tcp_sack_update_reported_history(tp, hstart, hend);
 853         }
 854         if (tcp_do_smartsack && (tp->t_flags & TF_SACKLEFT)) {
 855                 /* Fill in from left!  Walk re-assembly queue. */
 856                 struct tseg_qent *q;
 857
 858                 q = LIST_FIRST(&tp->t_segq);
 859                 while (q != NULL &&
 860                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
 861                         *lp++ = htonl(q->tqe_th->th_seq);
 862                         *lp++ = htonl(TCP_SACK_BLKEND(
 863                             q->tqe_th->th_seq + q->tqe_len,
 864                             q->tqe_th->th_flags));
 865                         optlen += TCPOLEN_SACK_BLOCK;
 866                         q = LIST_NEXT(q, tqe_q);
 867                 }
 868         } else {
 869                 int n = 0;
 870
 871                 /* Fill in SACK blocks from right side. */
 872                 while (n < tp->nsackhistory &&
 873                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
 874                         if (tp->sackhistory[n].rblk_start != hstart) {
 875                                 *lp++ = htonl(tp->sackhistory[n].rblk_start);
 876                                 *lp++ = htonl(tp->sackhistory[n].rblk_end);
 877                                 optlen += TCPOLEN_SACK_BLOCK;
 878                         }
 879                         ++n;
 880                 }
 881         }
 882         tp->reportblk.rblk_start = tp->reportblk.rblk_end;
 883         tp->t_flags &= ~(TF_DUPSEG | TF_ENCLOSESEG | TF_SACKLEFT);
 884         nblocks = (lp - olp - 1) / 2;
 885         *olp = htonl(TCPOPT_SACK_ALIGNED |
 886                      (TCPOLEN_SACK + nblocks * TCPOLEN_SACK_BLOCK));
 887         *plen = optlen;
 888 }