sys/netinet/tcp_sack.c

   1 /*
   2  * Copyright (c) 2003, 2004 Jeffrey M. Hsu.  All rights reserved.
   3  * Copyright (c) 2003, 2004 The DragonFly Project.  All rights reserved.
   4  *
   5  * This code is derived from software contributed to The DragonFly Project
   6  * by Jeffrey M. Hsu.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in the
  15  *    documentation and/or other materials provided with the distribution.
  16  * 3. Neither the name of The DragonFly Project nor the names of its
  17  *    contributors may be used to endorse or promote products derived
  18  *    from this software without specific, prior written permission.
  19  *
  20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  23  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  24  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  25  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  27  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  28  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  29  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  30  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  31  * SUCH DAMAGE.
  32  *
  33  * $DragonFly: src/sys/netinet/tcp_sack.c,v 1.8 2008/08/15 21:37:16 nth Exp $
  34  */
  35
  36 #include <sys/param.h>
  37 #include <sys/systm.h>
  38 #include <sys/kernel.h>
  39 #include <sys/malloc.h>
  40 #include <sys/queue.h>
  41 #include <sys/thread.h>
  42 #include <sys/types.h>
  43 #include <sys/socket.h>
  44 #include <sys/socketvar.h>
  45
  46 #include <net/if.h>
  47
  48 #include <netinet/in.h>
  49 #include <netinet/in_systm.h>
  50 #include <netinet/ip.h>
  51 #include <netinet/in_var.h>
  52 #include <netinet/in_pcb.h>
  53 #include <netinet/ip_var.h>
  54 #include <netinet/tcp.h>
  55 #include <netinet/tcp_seq.h>
  56 #include <netinet/tcp_var.h>
  57
  58 /*
  59  * Implemented:
  60  *
  61  * RFC 2018
  62  * RFC 2883
  63  * RFC 3517
  64  */
  65
  66 struct sackblock {
  67         tcp_seq                 sblk_start;
  68         tcp_seq                 sblk_end;
  69         TAILQ_ENTRY(sackblock)  sblk_list;
  70 };
  71
  72 #define MAXSAVEDBLOCKS  8                       /* per connection limit */
  73
  74 static int insert_block(struct scoreboard *scb,
  75                         const struct raw_sackblock *raw_sb);
  76 static void update_lostseq(struct scoreboard *scb, tcp_seq snd_una,
  77                            u_int maxseg);
  78
  79 static MALLOC_DEFINE(M_SACKBLOCK, "sblk", "sackblock struct");
  80
  81 /*
  82  * Per-tcpcb initialization.
  83  */
  84 void
  85 tcp_sack_tcpcb_init(struct tcpcb *tp)
  86 {
  87         struct scoreboard *scb = &tp->scb;
  88
  89         scb->nblocks = 0;
  90         TAILQ_INIT(&scb->sackblocks);
  91         scb->lastfound = NULL;
  92 }
  93
  94 /*
  95  * Find the SACK block containing or immediately preceding "seq".
  96  * The boolean result indicates whether the sequence is actually
  97  * contained in the SACK block.
  98  */
  99 static boolean_t
 100 sack_block_lookup(struct scoreboard *scb, tcp_seq seq, struct sackblock **sb)
 101 {
 102         struct sackblock *hint = scb->lastfound;
 103         struct sackblock *cur, *last, *prev;
 104
 105         if (TAILQ_EMPTY(&scb->sackblocks)) {
 106                 *sb = NULL;
 107                 return FALSE;
 108         }
 109
 110         if (hint == NULL) {
 111                 /* No hint.  Search from start to end. */
 112                 cur = TAILQ_FIRST(&scb->sackblocks);
 113                 last = NULL;
 114                 prev = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 115         } else  {
 116                 if (SEQ_GEQ(seq, hint->sblk_start)) {
 117                         /* Search from hint to end of list. */
 118                         cur = hint;
 119                         last = NULL;
 120                         prev = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 121                 } else {
 122                         /* Search from front of list to hint. */
 123                         cur = TAILQ_FIRST(&scb->sackblocks);
 124                         last = hint;
 125                         prev = TAILQ_PREV(hint, sackblock_list, sblk_list);
 126                 }
 127         }
 128
 129         do {
 130                 if (SEQ_GT(cur->sblk_end, seq)) {
 131                         if (SEQ_GEQ(seq, cur->sblk_start)) {
 132                                 *sb = scb->lastfound = cur;
 133                                 return TRUE;
 134                         } else {
 135                                 *sb = scb->lastfound =
 136                                     TAILQ_PREV(cur, sackblock_list, sblk_list);
 137                                 return FALSE;
 138                         }
 139                 }
 140                 cur = TAILQ_NEXT(cur, sblk_list);
 141         } while (cur != last);
 142
 143         *sb = scb->lastfound = prev;
 144         return FALSE;
 145 }
 146
 147 /*
 148  * Allocate a SACK block.
 149  */
 150 static __inline struct sackblock *
 151 alloc_sackblock(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;
 302         int i;
 303
 304         if (tcp_sack_ndsack_blocks(blocks, numblocks, tp->snd_una) > 0)
 305                 startblock = 1;
 306         else
 307                 startblock = 0;
 308
 309         for (i = startblock; i < numblocks; i++) {
 310                 struct raw_sackblock *newsackblock = &blocks[i];
 311
 312                 /* don't accept bad SACK blocks */
 313                 if (SEQ_GT(newsackblock->rblk_end, tp->snd_max)) {
 314                         tcpstat.tcps_rcvbadsackopt++;
 315                         break;          /* skip all other blocks */
 316                 }
 317                 tcpstat.tcps_sacksbupdate++;
 318
 319                 if (insert_block(scb, newsackblock))
 320                         break;
 321         }
 322 }
 323
 324 void
 325 tcp_sack_update_scoreboard(struct tcpcb *tp, struct tcpopt *to)
 326 {
 327         struct scoreboard *scb = &tp->scb;
 328         int rexmt_high_update = 0;
 329
 330         tcp_sack_ack_blocks(scb, tp->snd_una);
 331         tcp_sack_add_blocks(tp, to);
 332         update_lostseq(scb, tp->snd_una, tp->t_maxseg);
 333         if (SEQ_LT(tp->rexmt_high, tp->snd_una)) {
 334                 tp->rexmt_high = tp->snd_una;
 335                 rexmt_high_update = 1;
 336         }
 337         if (tp->t_flags & TF_SACKRESCUED) {
 338                 if (SEQ_LT(tp->rexmt_rescue, tp->snd_una)) {
 339                         tp->t_flags &= ~TF_SACKRESCUED;
 340                 } else if (rexmt_high_update &&
 341                     SEQ_LT(tp->rexmt_rescue, tp->rexmt_high)) {
 342                         /* Drag RescueRxt along with HighRxt */
 343                         tp->rexmt_rescue = tp->rexmt_high;
 344                 }
 345         }
 346 }
 347
 348 /*
 349  * Insert SACK block into sender's scoreboard.
 350  */
 351 static int
 352 insert_block(struct scoreboard *scb, const struct raw_sackblock *raw_sb)
 353 {
 354         struct sackblock *sb, *workingblock;
 355         boolean_t overlap_front;
 356
 357         if (TAILQ_EMPTY(&scb->sackblocks)) {
 358                 struct sackblock *newblock;
 359
 360                 KASSERT(scb->nblocks == 0, ("emply scb w/ blocks"));
 361
 362                 newblock = alloc_sackblock(scb, raw_sb);
 363                 if (newblock == NULL)
 364                         return ENOMEM;
 365                 TAILQ_INSERT_HEAD(&scb->sackblocks, newblock, sblk_list);
 366                 scb->nblocks = 1;
 367                 return 0;
 368         }
 369
 370         KASSERT(scb->nblocks > 0, ("insert_block() called w/ no blocks"));
 371         KASSERT(scb->nblocks <= MAXSAVEDBLOCKS,
 372             ("too many SACK blocks %d", scb->nblocks));
 373
 374         overlap_front = sack_block_lookup(scb, raw_sb->rblk_start, &sb);
 375
 376         if (sb == NULL) {
 377                 workingblock = alloc_sackblock_limit(scb, raw_sb);
 378                 if (workingblock == NULL)
 379                         return ENOMEM;
 380                 TAILQ_INSERT_HEAD(&scb->sackblocks, workingblock, sblk_list);
 381                 ++scb->nblocks;
 382         } else {
 383                 if (overlap_front || sb->sblk_end == raw_sb->rblk_start) {
 384                         /* Extend old block */
 385                         workingblock = sb;
 386                         if (SEQ_GT(raw_sb->rblk_end, sb->sblk_end))
 387                                 sb->sblk_end = raw_sb->rblk_end;
 388                         tcpstat.tcps_sacksbreused++;
 389                 } else {
 390                         workingblock = alloc_sackblock_limit(scb, raw_sb);
 391                         if (workingblock == NULL)
 392                                 return ENOMEM;
 393                         TAILQ_INSERT_AFTER(&scb->sackblocks, sb, workingblock,
 394                             sblk_list);
 395                         ++scb->nblocks;
 396                 }
 397         }
 398
 399         /* Consolidate right-hand side. */
 400         sb = TAILQ_NEXT(workingblock, sblk_list);
 401         while (sb != NULL &&
 402             SEQ_GEQ(workingblock->sblk_end, sb->sblk_end)) {
 403                 struct sackblock *nextblock;
 404
 405                 nextblock = TAILQ_NEXT(sb, sblk_list);
 406                 if (scb->lastfound == sb)
 407                         scb->lastfound = NULL;
 408                 /* Remove completely overlapped block */
 409                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 410                 free_sackblock(scb, sb);
 411                 --scb->nblocks;
 412                 KASSERT(scb->nblocks > 0,
 413                     ("removed overlapped block: %d blocks left", scb->nblocks));
 414                 sb = nextblock;
 415         }
 416         if (sb != NULL &&
 417             SEQ_GEQ(workingblock->sblk_end, sb->sblk_start)) {
 418                 /* Extend new block to cover partially overlapped old block. */
 419                 workingblock->sblk_end = sb->sblk_end;
 420                 if (scb->lastfound == sb)
 421                         scb->lastfound = NULL;
 422                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 423                 free_sackblock(scb, sb);
 424                 --scb->nblocks;
 425                 KASSERT(scb->nblocks > 0,
 426                     ("removed partial right: %d blocks left", scb->nblocks));
 427         }
 428         return 0;
 429 }
 430
 431 #ifdef DEBUG_SACK_BLOCKS
 432 static void
 433 tcp_sack_dump_blocks(struct scoreboard *scb)
 434 {
 435         struct sackblock *sb;
 436
 437         kprintf("%d blocks:", scb->nblocks);
 438         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
 439                 kprintf(" [%u, %u)", sb->sblk_start, sb->sblk_end);
 440         kprintf("\n");
 441 }
 442 #else
 443 static __inline void
 444 tcp_sack_dump_blocks(struct scoreboard *scb)
 445 {
 446 }
 447 #endif
 448
 449 /*
 450  * Optimization to quickly determine which packets are lost.
 451  */
 452 static void
 453 update_lostseq(struct scoreboard *scb, tcp_seq snd_una, u_int maxseg)
 454 {
 455         struct sackblock *sb;
 456         int nsackblocks = 0;
 457         int bytes_sacked = 0;
 458
 459         sb = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 460         while (sb != NULL) {
 461                 ++nsackblocks;
 462                 bytes_sacked += sb->sblk_end - sb->sblk_start;
 463                 if (nsackblocks == tcprexmtthresh ||
 464                     bytes_sacked >= tcprexmtthresh * maxseg) {
 465                         scb->lostseq = sb->sblk_start;
 466                         return;
 467                 }
 468                 sb = TAILQ_PREV(sb, sackblock_list, sblk_list);
 469         }
 470         scb->lostseq = snd_una;
 471 }
 472
 473 /*
 474  * Return whether the given sequence number is considered lost.
 475  */
 476 static boolean_t
 477 scb_islost(struct scoreboard *scb, tcp_seq seqnum)
 478 {
 479         return SEQ_LT(seqnum, scb->lostseq);
 480 }
 481
 482 /*
 483  * True if at least "amount" has been SACKed.  Used by Early Retransmit.
 484  */
 485 boolean_t
 486 tcp_sack_has_sacked(struct scoreboard *scb, u_int amount)
 487 {
 488         struct sackblock *sb;
 489         int bytes_sacked = 0;
 490
 491         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
 492                 bytes_sacked += sb->sblk_end - sb->sblk_start;
 493                 if (bytes_sacked >= amount)
 494                         return TRUE;
 495         }
 496         return FALSE;
 497 }
 498
 499 /*
 500  * Number of bytes SACKed below seq.
 501  */
 502 int
 503 tcp_sack_bytes_below(struct scoreboard *scb, tcp_seq seq)
 504 {
 505         struct sackblock *sb;
 506         int bytes_sacked = 0;
 507
 508         sb = TAILQ_FIRST(&scb->sackblocks);
 509         while (sb && SEQ_GT(seq, sb->sblk_start)) {
 510                 bytes_sacked += seq_min(seq, sb->sblk_end) - sb->sblk_start;
 511                 sb = TAILQ_NEXT(sb, sblk_list);
 512         }
 513         return bytes_sacked;
 514 }
 515
 516 /*
 517  * Return estimate of the number of bytes outstanding in the network.
 518  */
 519 uint32_t
 520 tcp_sack_compute_pipe(struct tcpcb *tp)
 521 {
 522         struct scoreboard *scb = &tp->scb;
 523         struct sackblock *sb;
 524         int nlost, nretransmitted;
 525         tcp_seq end;
 526
 527         nlost = tp->snd_max - scb->lostseq;
 528         nretransmitted = tp->rexmt_high - tp->snd_una;
 529
 530         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
 531                 if (SEQ_LT(sb->sblk_start, tp->rexmt_high)) {
 532                         end = seq_min(sb->sblk_end, tp->rexmt_high);
 533                         nretransmitted -= end - sb->sblk_start;
 534                 }
 535                 if (SEQ_GEQ(sb->sblk_start, scb->lostseq))
 536                         nlost -= sb->sblk_end - sb->sblk_start;
 537         }
 538
 539         return (nlost + nretransmitted);
 540 }
 541
 542 /*
 543  * Return the sequence number and length of the next segment to transmit
 544  * when in Fast Recovery.
 545  */
 546 boolean_t
 547 tcp_sack_nextseg(struct tcpcb *tp, tcp_seq *nextrexmt, uint32_t *plen,
 548     boolean_t *rescue)
 549 {
 550         struct scoreboard *scb = &tp->scb;
 551         struct socket *so = tp->t_inpcb->inp_socket;
 552         struct sackblock *sb;
 553         const struct sackblock *lastblock =
 554             TAILQ_LAST(&scb->sackblocks, sackblock_list);
 555         tcp_seq torexmt;
 556         long len, off;
 557
 558         /* skip SACKed data */
 559         tcp_sack_skip_sacked(scb, &tp->rexmt_high);
 560
 561         /* Look for lost data. */
 562         torexmt = tp->rexmt_high;
 563         *rescue = FALSE;
 564         if (lastblock != NULL) {
 565                 if (SEQ_LT(torexmt, lastblock->sblk_end) &&
 566                     scb_islost(scb, torexmt)) {
 567 sendunsacked:
 568                         *nextrexmt = torexmt;
 569                         /* If the left-hand edge has been SACKed, pull it in. */
 570                         if (sack_block_lookup(scb, torexmt + tp->t_maxseg, &sb))
 571                                 *plen = sb->sblk_start - torexmt;
 572                         else
 573                                 *plen = tp->t_maxseg;
 574                         return TRUE;
 575                 }
 576         }
 577
 578         /* See if unsent data available within send window. */
 579         off = tp->snd_max - tp->snd_una;
 580         len = (long) ulmin(so->so_snd.ssb_cc, tp->snd_wnd) - off;
 581         if (len > 0) {
 582                 *nextrexmt = tp->snd_max;       /* Send new data. */
 583                 *plen = tp->t_maxseg;
 584                 return TRUE;
 585         }
 586
 587         /* We're less certain this data has been lost. */
 588         if (lastblock != NULL && SEQ_LT(torexmt, lastblock->sblk_end))
 589                 goto sendunsacked;
 590
 591         /* Rescue retransmission */
 592         if (tcp_do_rescuesack) {
 593                 tcpstat.tcps_sackrescue_try++;
 594                 if (tp->t_flags & TF_SACKRESCUED) {
 595                         if (!tcp_aggressive_rescuesack)
 596                                 return FALSE;
 597
 598                         /*
 599                          * Aggressive variant of the rescue retransmission.
 600                          *
 601                          * The idea of the rescue retransmission is to sustain
 602                          * the ACK clock thus to avoid timeout retransmission.
 603                          *
 604                          * Under some situations, the conservative approach
 605                          * suggested in the draft
 606                          * http://tools.ietf.org/html/
 607                          * draft-nishida-tcpm-rescue-retransmission-00
 608                          * could not sustain ACK clock, since it only allows
 609                          * one rescue retransmission before a cumulative ACK
 610                          * covers the segement transmitted by rescue
 611                          * retransmission.
 612                          *
 613                          * We try to locate the next unSACKed segment which
 614                          * follows the previously sent rescue segment.  If
 615                          * there is no such segment, we loop back to the first
 616                          * unacknowledged segment.
 617                          */
 618
 619                         /*
 620                          * Skip SACKed data, but here we follow
 621                          * the last transmitted rescue segment.
 622                          */
 623                         torexmt = tp->rexmt_rescue;
 624                         tcp_sack_skip_sacked(scb, &torexmt);
 625                         if (torexmt == tp->snd_max) {
 626                                 /* Nothing left to retransmit; restart */
 627                                 torexmt = tp->snd_una;
 628                         }
 629                 }
 630                 *rescue = TRUE;
 631                 goto sendunsacked;
 632         } else if (tcp_do_smartsack && lastblock == NULL) {
 633                 tcpstat.tcps_sackrescue_try++;
 634                 *rescue = TRUE;
 635                 goto sendunsacked;
 636         }
 637
 638         return FALSE;
 639 }
 640
 641 /*
 642  * Return the next sequence number higher than "*prexmt" that has
 643  * not been SACKed.
 644  */
 645 void
 646 tcp_sack_skip_sacked(struct scoreboard *scb, tcp_seq *prexmt)
 647 {
 648         struct sackblock *sb;
 649
 650         /* skip SACKed data */
 651         if (sack_block_lookup(scb, *prexmt, &sb))
 652                 *prexmt = sb->sblk_end;
 653 }
 654
 655 #ifdef later
 656 void
 657 tcp_sack_save_scoreboard(struct scoreboard *scb)
 658 {
 659         struct scoreboard *scb = &tp->scb;
 660
 661         scb->sackblocks_prev = scb->sackblocks;
 662         TAILQ_INIT(&scb->sackblocks);
 663 }
 664
 665 void
 666 tcp_sack_revert_scoreboard(struct scoreboard *scb, tcp_seq snd_una,
 667                            u_int maxseg)
 668 {
 669         struct sackblock *sb;
 670
 671         scb->sackblocks = scb->sackblocks_prev;
 672         scb->nblocks = 0;
 673         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
 674                 ++scb->nblocks;
 675         tcp_sack_ack_blocks(scb, snd_una);
 676         scb->lastfound = NULL;
 677 }
 678 #endif
 679
 680 #ifdef DEBUG_SACK_HISTORY
 681 static void
 682 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
 683 {
 684         int i;
 685         static int ndumped;
 686
 687         /* only need a couple of these to debug most problems */
 688         if (++ndumped > 900)
 689                 return;
 690
 691         kprintf("%s:\tnsackhistory %d: ", msg, tp->nsackhistory);
 692         for (i = 0; i < tp->nsackhistory; ++i)
 693                 kprintf("[%u, %u) ", tp->sackhistory[i].rblk_start,
 694                     tp->sackhistory[i].rblk_end);
 695         kprintf("\n");
 696 }
 697 #else
 698 static __inline void
 699 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
 700 {
 701 }
 702 #endif
 703
 704 /*
 705  * Remove old SACK blocks from the SACK history that have already been ACKed.
 706  */
 707 static void
 708 tcp_sack_ack_history(struct tcpcb *tp)
 709 {
 710         int i, nblocks, openslot;
 711
 712         tcp_sack_dump_history("before tcp_sack_ack_history", tp);
 713         nblocks = tp->nsackhistory;
 714         for (i = openslot = 0; i < nblocks; ++i) {
 715                 if (SEQ_LEQ(tp->sackhistory[i].rblk_end, tp->rcv_nxt)) {
 716                         --tp->nsackhistory;
 717                         continue;
 718                 }
 719                 if (SEQ_LT(tp->sackhistory[i].rblk_start, tp->rcv_nxt))
 720                         tp->sackhistory[i].rblk_start = tp->rcv_nxt;
 721                 if (i == openslot)
 722                         ++openslot;
 723                 else
 724                         tp->sackhistory[openslot++] = tp->sackhistory[i];
 725         }
 726         tcp_sack_dump_history("after tcp_sack_ack_history", tp);
 727         KASSERT(openslot == tp->nsackhistory,
 728             ("tcp_sack_ack_history miscounted: %d != %d",
 729             openslot, tp->nsackhistory));
 730 }
 731
 732 /*
 733  * Add or merge newblock into reported history.
 734  * Also remove or update SACK blocks that will be acked.
 735  */
 736 static void
 737 tcp_sack_update_reported_history(struct tcpcb *tp, tcp_seq start, tcp_seq end)
 738 {
 739         struct raw_sackblock copy[MAX_SACK_REPORT_BLOCKS];
 740         int i, cindex;
 741
 742         tcp_sack_dump_history("before tcp_sack_update_reported_history", tp);
 743         /*
 744          * Six cases:
 745          *      0) no overlap
 746          *      1) newblock == oldblock
 747          *      2) oldblock contains newblock
 748          *      3) newblock contains oldblock
 749          *      4) tail of oldblock overlaps or abuts start of newblock
 750          *      5) tail of newblock overlaps or abuts head of oldblock
 751          */
 752         for (i = cindex = 0; i < tp->nsackhistory; ++i) {
 753                 struct raw_sackblock *oldblock = &tp->sackhistory[i];
 754                 tcp_seq old_start = oldblock->rblk_start;
 755                 tcp_seq old_end = oldblock->rblk_end;
 756
 757                 if (SEQ_LT(end, old_start) || SEQ_GT(start, old_end)) {
 758                         /* Case 0:  no overlap.  Copy old block. */
 759                         copy[cindex++] = *oldblock;
 760                         continue;
 761                 }
 762
 763                 if (SEQ_GEQ(start, old_start) && SEQ_LEQ(end, old_end)) {
 764                         /* Cases 1 & 2.  Move block to front of history. */
 765                         int j;
 766
 767                         start = old_start;
 768                         end = old_end;
 769                         /* no need to check rest of blocks */
 770                         for (j = i + 1; j < tp->nsackhistory; ++j)
 771                                 copy[cindex++] = tp->sackhistory[j];
 772                         break;
 773                 }
 774
 775                 if (SEQ_GEQ(old_end, start) && SEQ_LT(old_start, start)) {
 776                         /* Case 4:  extend start of new block. */
 777                         start = old_start;
 778                 } else if (SEQ_GEQ(end, old_start) && SEQ_GT(old_end, end)) {
 779                         /* Case 5: extend end of new block */
 780                         end = old_end;
 781                 } else {
 782                         /* Case 3.  Delete old block by not copying it. */
 783                         KASSERT(SEQ_LEQ(start, old_start) &&
 784                                 SEQ_GEQ(end, old_end),
 785                             ("bad logic: old [%u, %u), new [%u, %u)",
 786                              old_start, old_end, start, end));
 787                 }
 788         }
 789
 790         /* insert new block */
 791         tp->sackhistory[0].rblk_start = start;
 792         tp->sackhistory[0].rblk_end = end;
 793         cindex = min(cindex, MAX_SACK_REPORT_BLOCKS - 1);
 794         for (i = 0; i < cindex; ++i)
 795                 tp->sackhistory[i + 1] = copy[i];
 796         tp->nsackhistory = cindex + 1;
 797         tcp_sack_dump_history("after tcp_sack_update_reported_history", tp);
 798 }
 799
 800 /*
 801  * Fill in SACK report to return to data sender.
 802  */
 803 void
 804 tcp_sack_fill_report(struct tcpcb *tp, u_char *opt, u_int *plen)
 805 {
 806         u_int optlen = *plen;
 807         uint32_t *lp = (uint32_t *)(opt + optlen);
 808         uint32_t *olp;
 809         tcp_seq hstart = tp->rcv_nxt, hend;
 810         int nblocks;
 811
 812         KASSERT(TCP_MAXOLEN - optlen >=
 813             TCPOLEN_SACK_ALIGNED + TCPOLEN_SACK_BLOCK,
 814             ("no room for SACK header and one block: optlen %d", optlen));
 815
 816         if (tp->t_flags & TF_DUPSEG)
 817                 tcpstat.tcps_snddsackopt++;
 818         else
 819                 tcpstat.tcps_sndsackopt++;
 820
 821         olp = lp++;
 822         optlen += TCPOLEN_SACK_ALIGNED;
 823
 824         tcp_sack_ack_history(tp);
 825         if (tp->reportblk.rblk_start != tp->reportblk.rblk_end) {
 826                 *lp++ = htonl(tp->reportblk.rblk_start);
 827                 *lp++ = htonl(tp->reportblk.rblk_end);
 828                 optlen += TCPOLEN_SACK_BLOCK;
 829                 hstart = tp->reportblk.rblk_start;
 830                 hend = tp->reportblk.rblk_end;
 831                 if (tp->t_flags & TF_ENCLOSESEG) {
 832                         KASSERT(TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK,
 833                             ("no room for enclosing SACK block: oplen %d",
 834                             optlen));
 835                         *lp++ = htonl(tp->encloseblk.rblk_start);
 836                         *lp++ = htonl(tp->encloseblk.rblk_end);
 837                         optlen += TCPOLEN_SACK_BLOCK;
 838                         hstart = tp->encloseblk.rblk_start;
 839                         hend = tp->encloseblk.rblk_end;
 840                 }
 841                 if (SEQ_GT(hstart, tp->rcv_nxt))
 842                         tcp_sack_update_reported_history(tp, hstart, hend);
 843         }
 844         if (tcp_do_smartsack && (tp->t_flags & TF_SACKLEFT)) {
 845                 /* Fill in from left!  Walk re-assembly queue. */
 846                 struct tseg_qent *q;
 847
 848                 q = LIST_FIRST(&tp->t_segq);
 849                 while (q != NULL &&
 850                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
 851                         *lp++ = htonl(q->tqe_th->th_seq);
 852                         *lp++ = htonl(TCP_SACK_BLKEND(
 853                             q->tqe_th->th_seq + q->tqe_len,
 854                             q->tqe_th->th_flags));
 855                         optlen += TCPOLEN_SACK_BLOCK;
 856                         q = LIST_NEXT(q, tqe_q);
 857                 }
 858         } else {
 859                 int n = 0;
 860
 861                 /* Fill in SACK blocks from right side. */
 862                 while (n < tp->nsackhistory &&
 863                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
 864                         if (tp->sackhistory[n].rblk_start != hstart) {
 865                                 *lp++ = htonl(tp->sackhistory[n].rblk_start);
 866                                 *lp++ = htonl(tp->sackhistory[n].rblk_end);
 867                                 optlen += TCPOLEN_SACK_BLOCK;
 868                         }
 869                         ++n;
 870                 }
 871         }
 872         tp->reportblk.rblk_start = tp->reportblk.rblk_end;
 873         tp->t_flags &= ~(TF_DUPSEG | TF_ENCLOSESEG | TF_SACKLEFT);
 874         nblocks = (lp - olp - 1) / 2;
 875         *olp = htonl(TCPOPT_SACK_ALIGNED |
 876                      (TCPOLEN_SACK + nblocks * TCPOLEN_SACK_BLOCK));
 877         *plen = optlen;
 878 }