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  * RFC 6675
  65  */
  66
  67 struct sackblock {
  68         tcp_seq                 sblk_start;
  69         tcp_seq                 sblk_end;
  70         TAILQ_ENTRY(sackblock)  sblk_list;
  71 };
  72
  73 #define MAXSAVEDBLOCKS  8                       /* per connection limit */
  74
  75 static int insert_block(struct scoreboard *scb,
  76                         const struct raw_sackblock *raw_sb, boolean_t *update);
  77
  78 static MALLOC_DEFINE(M_SACKBLOCK, "sblk", "sackblock struct");
  79
  80 /*
  81  * Per-tcpcb initialization.
  82  */
  83 void
  84 tcp_sack_tcpcb_init(struct tcpcb *tp)
  85 {
  86         struct scoreboard *scb = &tp->scb;
  87
  88         scb->nblocks = 0;
  89         TAILQ_INIT(&scb->sackblocks);
  90         scb->lastfound = NULL;
  91 }
  92
  93 /*
  94  * Find the SACK block containing or immediately preceding "seq".
  95  * The boolean result indicates whether the sequence is actually
  96  * contained in the SACK block.
  97  */
  98 static boolean_t
  99 sack_block_lookup(struct scoreboard *scb, tcp_seq seq, struct sackblock **sb)
 100 {
 101         struct sackblock *hint = scb->lastfound;
 102         struct sackblock *cur, *last, *prev;
 103
 104         if (TAILQ_EMPTY(&scb->sackblocks)) {
 105                 *sb = NULL;
 106                 return FALSE;
 107         }
 108
 109         if (hint == NULL) {
 110                 /* No hint.  Search from start to end. */
 111                 cur = TAILQ_FIRST(&scb->sackblocks);
 112                 last = NULL;
 113                 prev = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 114         } else  {
 115                 if (SEQ_GEQ(seq, hint->sblk_start)) {
 116                         /* Search from hint to end of list. */
 117                         cur = hint;
 118                         last = NULL;
 119                         prev = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 120                 } else {
 121                         /* Search from front of list to hint. */
 122                         cur = TAILQ_FIRST(&scb->sackblocks);
 123                         last = hint;
 124                         prev = TAILQ_PREV(hint, sackblock_list, sblk_list);
 125                 }
 126         }
 127
 128         do {
 129                 if (SEQ_GT(cur->sblk_end, seq)) {
 130                         if (SEQ_GEQ(seq, cur->sblk_start)) {
 131                                 *sb = scb->lastfound = cur;
 132                                 return TRUE;
 133                         } else {
 134                                 *sb = scb->lastfound =
 135                                     TAILQ_PREV(cur, sackblock_list, sblk_list);
 136                                 return FALSE;
 137                         }
 138                 }
 139                 cur = TAILQ_NEXT(cur, sblk_list);
 140         } while (cur != last);
 141
 142         *sb = scb->lastfound = prev;
 143         return FALSE;
 144 }
 145
 146 /*
 147  * Allocate a SACK block.
 148  */
 149 static __inline struct sackblock *
 150 alloc_sackblock(struct scoreboard *scb, const struct raw_sackblock *raw_sb)
 151 {
 152         struct sackblock *sb;
 153
 154         if (scb->freecache != NULL) {
 155                 sb = scb->freecache;
 156                 scb->freecache = NULL;
 157                 tcpstat.tcps_sacksbfast++;
 158         } else {
 159                 sb = kmalloc(sizeof(struct sackblock), M_SACKBLOCK, M_NOWAIT);
 160                 if (sb == NULL) {
 161                         tcpstat.tcps_sacksbfailed++;
 162                         return NULL;
 163                 }
 164         }
 165         sb->sblk_start = raw_sb->rblk_start;
 166         sb->sblk_end = raw_sb->rblk_end;
 167         return sb;
 168 }
 169
 170 static __inline struct sackblock *
 171 alloc_sackblock_limit(struct scoreboard *scb,
 172     const struct raw_sackblock *raw_sb)
 173 {
 174         if (scb->nblocks == MAXSAVEDBLOCKS) {
 175                 /*
 176                  * Should try to kick out older blocks XXX JH
 177                  * May be able to coalesce with existing block.
 178                  * Or, go other way and free all blocks if we hit
 179                  * this limit.
 180                  */
 181                 tcpstat.tcps_sacksboverflow++;
 182                 return NULL;
 183         }
 184         return alloc_sackblock(scb, raw_sb);
 185 }
 186
 187 /*
 188  * Free a SACK block.
 189  */
 190 static __inline void
 191 free_sackblock(struct scoreboard *scb, struct sackblock *s)
 192 {
 193         if (scb->freecache == NULL) {
 194                 /* YYY Maybe use the latest freed block? */
 195                 scb->freecache = s;
 196                 return;
 197         }
 198         kfree(s, M_SACKBLOCK);
 199 }
 200
 201 /*
 202  * Free up SACK blocks for data that's been acked.
 203  */
 204 static void
 205 tcp_sack_ack_blocks(struct tcpcb *tp, tcp_seq th_ack)
 206 {
 207         struct scoreboard *scb = &tp->scb;
 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_GEQ(th_ack, sb->sblk_start)) {
 223                 /* Other side reneged? XXX */
 224                 tcpstat.tcps_sackrenege++;
 225                 tcp_sack_discard(tp);
 226         }
 227 }
 228
 229 /*
 230  * Delete and free SACK blocks saved in scoreboard.
 231  */
 232 static void
 233 tcp_sack_cleanup(struct scoreboard *scb)
 234 {
 235         struct sackblock *sb, *nb;
 236
 237         TAILQ_FOREACH_MUTABLE(sb, &scb->sackblocks, sblk_list, nb) {
 238                 free_sackblock(scb, sb);
 239                 --scb->nblocks;
 240         }
 241         KASSERT(scb->nblocks == 0,
 242             ("SACK block %d count not zero", scb->nblocks));
 243         TAILQ_INIT(&scb->sackblocks);
 244         scb->lastfound = NULL;
 245 }
 246
 247 /*
 248  * Discard SACK scoreboard, HighRxt, RescueRxt and LostSeq.
 249  */
 250 void
 251 tcp_sack_discard(struct tcpcb *tp)
 252 {
 253         tcp_sack_cleanup(&tp->scb);
 254         tp->rexmt_high = tp->snd_una;
 255         tp->sack_flags &= ~TSACK_F_SACKRESCUED;
 256         tp->scb.lostseq = tp->snd_una;
 257 }
 258
 259 /*
 260  * Delete and free SACK blocks saved in scoreboard.
 261  * Delete the one slot block cache.
 262  */
 263 void
 264 tcp_sack_destroy(struct scoreboard *scb)
 265 {
 266         tcp_sack_cleanup(scb);
 267         if (scb->freecache != NULL) {
 268                 kfree(scb->freecache, M_SACKBLOCK);
 269                 scb->freecache = NULL;
 270         }
 271 }
 272
 273 /*
 274  * Cleanup the reported SACK block information
 275  */
 276 void
 277 tcp_sack_report_cleanup(struct tcpcb *tp)
 278 {
 279         tp->sack_flags &=
 280             ~(TSACK_F_DUPSEG | TSACK_F_ENCLOSESEG | TSACK_F_SACKLEFT);
 281         tp->reportblk.rblk_start = tp->reportblk.rblk_end;
 282 }
 283
 284 /*
 285  * Returns      0 if not D-SACK block,
 286  *              1 if D-SACK,
 287  *              2 if duplicate of out-of-order D-SACK block.
 288  */
 289 int
 290 tcp_sack_ndsack_blocks(const struct raw_sackblock *blocks, const int numblocks,
 291     tcp_seq snd_una)
 292 {
 293         if (numblocks == 0)
 294                 return 0;
 295
 296         if (SEQ_LT(blocks[0].rblk_start, snd_una))
 297                 return 1;
 298
 299         /* block 0 inside block 1 */
 300         if (numblocks > 1 &&
 301             SEQ_GEQ(blocks[0].rblk_start, blocks[1].rblk_start) &&
 302             SEQ_LEQ(blocks[0].rblk_end, blocks[1].rblk_end))
 303                 return 2;
 304
 305         return 0;
 306 }
 307
 308 /*
 309  * Update scoreboard on new incoming ACK.
 310  */
 311 static void
 312 tcp_sack_add_blocks(struct tcpcb *tp, struct tcpopt *to)
 313 {
 314         const int numblocks = to->to_nsackblocks;
 315         struct raw_sackblock *blocks = to->to_sackblocks;
 316         struct scoreboard *scb = &tp->scb;
 317         int startblock, i;
 318
 319         if (tcp_sack_ndsack_blocks(blocks, numblocks, tp->snd_una) > 0)
 320                 startblock = 1;
 321         else
 322                 startblock = 0;
 323
 324         to->to_flags |= TOF_SACK_REDUNDANT;
 325         for (i = startblock; i < numblocks; i++) {
 326                 struct raw_sackblock *newsackblock = &blocks[i];
 327                 boolean_t update;
 328                 int error;
 329
 330                 /* Guard against ACK reordering */
 331                 if (SEQ_LEQ(newsackblock->rblk_start, tp->snd_una))
 332                         continue;
 333
 334                 /* Don't accept bad SACK blocks */
 335                 if (SEQ_GT(newsackblock->rblk_end, tp->snd_max)) {
 336                         tcpstat.tcps_rcvbadsackopt++;
 337                         break;          /* skip all other blocks */
 338                 }
 339                 tcpstat.tcps_sacksbupdate++;
 340
 341                 error = insert_block(scb, newsackblock, &update);
 342                 if (update)
 343                         to->to_flags &= ~TOF_SACK_REDUNDANT;
 344                 if (error)
 345                         break;
 346         }
 347 }
 348
 349 void
 350 tcp_sack_update_scoreboard(struct tcpcb *tp, struct tcpopt *to)
 351 {
 352         struct scoreboard *scb = &tp->scb;
 353         int rexmt_high_update = 0;
 354
 355         tcp_sack_ack_blocks(tp, tp->snd_una);
 356         tcp_sack_add_blocks(tp, to);
 357         tcp_sack_update_lostseq(scb, tp->snd_una, tp->t_maxseg,
 358             tp->t_rxtthresh);
 359         if (SEQ_LT(tp->rexmt_high, tp->snd_una)) {
 360                 tp->rexmt_high = tp->snd_una;
 361                 rexmt_high_update = 1;
 362         }
 363         if (tp->sack_flags & TSACK_F_SACKRESCUED) {
 364                 if (SEQ_LEQ(tp->rexmt_rescue, tp->snd_una)) {
 365                         tp->sack_flags &= ~TSACK_F_SACKRESCUED;
 366                 } else if (tcp_aggressive_rescuesack && rexmt_high_update &&
 367                     SEQ_LT(tp->rexmt_rescue, tp->rexmt_high)) {
 368                         /* Drag RescueRxt along with HighRxt */
 369                         tp->rexmt_rescue = tp->rexmt_high;
 370                 }
 371         }
 372 }
 373
 374 /*
 375  * Insert SACK block into sender's scoreboard.
 376  */
 377 static int
 378 insert_block(struct scoreboard *scb, const struct raw_sackblock *raw_sb,
 379     boolean_t *update)
 380 {
 381         struct sackblock *sb, *workingblock;
 382         boolean_t overlap_front;
 383
 384         *update = TRUE;
 385         if (TAILQ_EMPTY(&scb->sackblocks)) {
 386                 struct sackblock *newblock;
 387
 388                 KASSERT(scb->nblocks == 0, ("emply scb w/ blocks"));
 389
 390                 newblock = alloc_sackblock(scb, raw_sb);
 391                 if (newblock == NULL)
 392                         return ENOMEM;
 393                 TAILQ_INSERT_HEAD(&scb->sackblocks, newblock, sblk_list);
 394                 scb->nblocks = 1;
 395                 return 0;
 396         }
 397
 398         KASSERT(scb->nblocks > 0, ("insert_block() called w/ no blocks"));
 399         KASSERT(scb->nblocks <= MAXSAVEDBLOCKS,
 400             ("too many SACK blocks %d", scb->nblocks));
 401
 402         overlap_front = sack_block_lookup(scb, raw_sb->rblk_start, &sb);
 403
 404         if (sb == NULL) {
 405                 workingblock = alloc_sackblock_limit(scb, raw_sb);
 406                 if (workingblock == NULL)
 407                         return ENOMEM;
 408                 TAILQ_INSERT_HEAD(&scb->sackblocks, workingblock, sblk_list);
 409                 ++scb->nblocks;
 410         } else {
 411                 if (overlap_front || sb->sblk_end == raw_sb->rblk_start) {
 412                         tcpstat.tcps_sacksbreused++;
 413
 414                         /* Extend old block */
 415                         workingblock = sb;
 416                         if (SEQ_GT(raw_sb->rblk_end, sb->sblk_end)) {
 417                                 sb->sblk_end = raw_sb->rblk_end;
 418                         } else {
 419                                 /* Exact match, nothing to consolidate */
 420                                 *update = FALSE;
 421                                 return 0;
 422                         }
 423                 } else {
 424                         workingblock = alloc_sackblock_limit(scb, raw_sb);
 425                         if (workingblock == NULL)
 426                                 return ENOMEM;
 427                         TAILQ_INSERT_AFTER(&scb->sackblocks, sb, workingblock,
 428                             sblk_list);
 429                         ++scb->nblocks;
 430                 }
 431         }
 432
 433         /* Consolidate right-hand side. */
 434         sb = TAILQ_NEXT(workingblock, sblk_list);
 435         while (sb != NULL &&
 436             SEQ_GEQ(workingblock->sblk_end, sb->sblk_end)) {
 437                 struct sackblock *nextblock;
 438
 439                 nextblock = TAILQ_NEXT(sb, sblk_list);
 440                 if (scb->lastfound == sb)
 441                         scb->lastfound = NULL;
 442                 /* Remove completely overlapped block */
 443                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 444                 free_sackblock(scb, sb);
 445                 --scb->nblocks;
 446                 KASSERT(scb->nblocks > 0,
 447                     ("removed overlapped block: %d blocks left", scb->nblocks));
 448                 sb = nextblock;
 449         }
 450         if (sb != NULL &&
 451             SEQ_GEQ(workingblock->sblk_end, sb->sblk_start)) {
 452                 /* Extend new block to cover partially overlapped old block. */
 453                 workingblock->sblk_end = sb->sblk_end;
 454                 if (scb->lastfound == sb)
 455                         scb->lastfound = NULL;
 456                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
 457                 free_sackblock(scb, sb);
 458                 --scb->nblocks;
 459                 KASSERT(scb->nblocks > 0,
 460                     ("removed partial right: %d blocks left", scb->nblocks));
 461         }
 462         return 0;
 463 }
 464
 465 #ifdef DEBUG_SACK_BLOCKS
 466 static void
 467 tcp_sack_dump_blocks(const struct scoreboard *scb)
 468 {
 469         const struct sackblock *sb;
 470
 471         kprintf("%d blocks:", scb->nblocks);
 472         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
 473                 kprintf(" [%u, %u)", sb->sblk_start, sb->sblk_end);
 474         kprintf("\n");
 475 }
 476 #else
 477 static __inline void
 478 tcp_sack_dump_blocks(const struct scoreboard *scb)
 479 {
 480 }
 481 #endif
 482
 483 /*
 484  * Optimization to quickly determine which packets are lost.
 485  */
 486 void
 487 tcp_sack_update_lostseq(struct scoreboard *scb, tcp_seq snd_una, u_int maxseg,
 488     int rxtthresh)
 489 {
 490         struct sackblock *sb;
 491         int nsackblocks = 0;
 492         int bytes_sacked = 0;
 493         int rxtthresh_bytes;
 494
 495         if (tcp_do_rfc6675)
 496                 rxtthresh_bytes = (rxtthresh - 1) * maxseg;
 497         else
 498                 rxtthresh_bytes = rxtthresh * maxseg;
 499
 500         sb = TAILQ_LAST(&scb->sackblocks, sackblock_list);
 501         while (sb != NULL) {
 502                 ++nsackblocks;
 503                 bytes_sacked += sb->sblk_end - sb->sblk_start;
 504                 if (nsackblocks == rxtthresh ||
 505                     bytes_sacked >= rxtthresh_bytes) {
 506                         scb->lostseq = sb->sblk_start;
 507                         return;
 508                 }
 509                 sb = TAILQ_PREV(sb, sackblock_list, sblk_list);
 510         }
 511         scb->lostseq = snd_una;
 512 }
 513
 514 /*
 515  * Return whether the given sequence number is considered lost.
 516  */
 517 boolean_t
 518 tcp_sack_islost(const struct scoreboard *scb, tcp_seq seqnum)
 519 {
 520         return SEQ_LT(seqnum, scb->lostseq);
 521 }
 522
 523 /*
 524  * True if at least "amount" has been SACKed.  Used by Early Retransmit.
 525  */
 526 boolean_t
 527 tcp_sack_has_sacked(const struct scoreboard *scb, u_int amount)
 528 {
 529         const struct sackblock *sb;
 530         int bytes_sacked = 0;
 531
 532         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
 533                 bytes_sacked += sb->sblk_end - sb->sblk_start;
 534                 if (bytes_sacked >= amount)
 535                         return TRUE;
 536         }
 537         return FALSE;
 538 }
 539
 540 /*
 541  * Number of bytes SACKed below seq.
 542  */
 543 int
 544 tcp_sack_bytes_below(const struct scoreboard *scb, tcp_seq seq)
 545 {
 546         const struct sackblock *sb;
 547         int bytes_sacked = 0;
 548
 549         sb = TAILQ_FIRST(&scb->sackblocks);
 550         while (sb && SEQ_GT(seq, sb->sblk_start)) {
 551                 bytes_sacked += seq_min(seq, sb->sblk_end) - sb->sblk_start;
 552                 sb = TAILQ_NEXT(sb, sblk_list);
 553         }
 554         return bytes_sacked;
 555 }
 556
 557 /*
 558  * Return estimate of the number of bytes outstanding in the network.
 559  */
 560 uint32_t
 561 tcp_sack_compute_pipe(const struct tcpcb *tp)
 562 {
 563         const struct scoreboard *scb = &tp->scb;
 564         const struct sackblock *sb;
 565         int nlost, nretransmitted;
 566         tcp_seq end;
 567
 568         nlost = tp->snd_max - scb->lostseq;
 569         nretransmitted = tp->rexmt_high - tp->snd_una;
 570
 571         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
 572                 if (SEQ_LT(sb->sblk_start, tp->rexmt_high)) {
 573                         end = seq_min(sb->sblk_end, tp->rexmt_high);
 574                         nretransmitted -= end - sb->sblk_start;
 575                 }
 576                 if (SEQ_GEQ(sb->sblk_start, scb->lostseq))
 577                         nlost -= sb->sblk_end - sb->sblk_start;
 578         }
 579
 580         return (nlost + nretransmitted);
 581 }
 582
 583 /*
 584  * Return the sequence number and length of the next segment to transmit
 585  * when in Fast Recovery.
 586  */
 587 boolean_t
 588 tcp_sack_nextseg(struct tcpcb *tp, tcp_seq *nextrexmt, uint32_t *plen,
 589     boolean_t *rescue)
 590 {
 591         struct scoreboard *scb = &tp->scb;
 592         struct socket *so = tp->t_inpcb->inp_socket;
 593         struct sackblock *sb;
 594         const struct sackblock *lastblock =
 595             TAILQ_LAST(&scb->sackblocks, sackblock_list);
 596         tcp_seq torexmt;
 597         long len, off, sendwin;
 598
 599         /* skip SACKed data */
 600         tcp_sack_skip_sacked(scb, &tp->rexmt_high);
 601
 602         /* Look for lost data. */
 603         torexmt = tp->rexmt_high;
 604         *rescue = FALSE;
 605         if (lastblock != NULL) {
 606                 if (SEQ_LT(torexmt, lastblock->sblk_end) &&
 607                     tcp_sack_islost(scb, torexmt)) {
 608 sendunsacked:
 609                         *nextrexmt = torexmt;
 610                         /* If the left-hand edge has been SACKed, pull it in. */
 611                         if (sack_block_lookup(scb, torexmt + tp->t_maxseg, &sb))
 612                                 *plen = sb->sblk_start - torexmt;
 613                         else
 614                                 *plen = tp->t_maxseg;
 615                         return TRUE;
 616                 }
 617         }
 618
 619         /* See if unsent data available within send window. */
 620         off = tp->snd_max - tp->snd_una;
 621         sendwin = min(tp->snd_wnd, tp->snd_bwnd);
 622         len = (long) ulmin(so->so_snd.ssb_cc, sendwin) - off;
 623         if (len > 0) {
 624                 *nextrexmt = tp->snd_max;       /* Send new data. */
 625                 *plen = tp->t_maxseg;
 626                 return TRUE;
 627         }
 628
 629         /* We're less certain this data has been lost. */
 630         if (lastblock != NULL && SEQ_LT(torexmt, lastblock->sblk_end))
 631                 goto sendunsacked;
 632
 633         /* Rescue retransmission */
 634         if (tcp_do_rescuesack || tcp_do_rfc6675) {
 635                 tcpstat.tcps_sackrescue_try++;
 636                 if (tp->sack_flags & TSACK_F_SACKRESCUED) {
 637                         if (!tcp_aggressive_rescuesack)
 638                                 return FALSE;
 639
 640                         /*
 641                          * Aggressive variant of the rescue retransmission.
 642                          *
 643                          * The idea of the rescue retransmission is to sustain
 644                          * the ACK clock thus to avoid timeout retransmission.
 645                          *
 646                          * Under some situations, the conservative approach
 647                          * suggested in the draft
 648                          * http://tools.ietf.org/html/
 649                          * draft-nishida-tcpm-rescue-retransmission-00
 650                          * could not sustain ACK clock, since it only allows
 651                          * one rescue retransmission before a cumulative ACK
 652                          * covers the segement transmitted by rescue
 653                          * retransmission.
 654                          *
 655                          * We try to locate the next unSACKed segment which
 656                          * follows the previously sent rescue segment.  If
 657                          * there is no such segment, we loop back to the first
 658                          * unacknowledged segment.
 659                          */
 660
 661                         /*
 662                          * Skip SACKed data, but here we follow
 663                          * the last transmitted rescue segment.
 664                          */
 665                         torexmt = tp->rexmt_rescue;
 666                         tcp_sack_skip_sacked(scb, &torexmt);
 667                 }
 668                 if (torexmt == tp->snd_max) {
 669                         /* Nothing left to retransmit; restart */
 670                         torexmt = tp->snd_una;
 671                 }
 672                 *rescue = TRUE;
 673                 goto sendunsacked;
 674         } else if (tcp_do_smartsack && lastblock == NULL) {
 675                 tcpstat.tcps_sackrescue_try++;
 676                 *rescue = TRUE;
 677                 goto sendunsacked;
 678         }
 679
 680         return FALSE;
 681 }
 682
 683 /*
 684  * Return the next sequence number higher than "*prexmt" that has
 685  * not been SACKed.
 686  */
 687 void
 688 tcp_sack_skip_sacked(struct scoreboard *scb, tcp_seq *prexmt)
 689 {
 690         struct sackblock *sb;
 691
 692         /* skip SACKed data */
 693         if (sack_block_lookup(scb, *prexmt, &sb))
 694                 *prexmt = sb->sblk_end;
 695 }
 696
 697 /*
 698  * The length of the first amount of unSACKed data
 699  */
 700 uint32_t
 701 tcp_sack_first_unsacked_len(const struct tcpcb *tp)
 702 {
 703         const struct sackblock *sb;
 704
 705         sb = TAILQ_FIRST(&tp->scb.sackblocks);
 706         if (sb == NULL)
 707                 return tp->t_maxseg;
 708
 709         KASSERT(SEQ_LT(tp->snd_una, sb->sblk_start),
 710             ("invalid sb start %u, snd_una %u",
 711              sb->sblk_start, tp->snd_una));
 712         return (sb->sblk_start - tp->snd_una);
 713 }
 714
 715 #ifdef later
 716 void
 717 tcp_sack_save_scoreboard(struct scoreboard *scb)
 718 {
 719         struct scoreboard *scb = &tp->scb;
 720
 721         scb->sackblocks_prev = scb->sackblocks;
 722         TAILQ_INIT(&scb->sackblocks);
 723 }
 724
 725 void
 726 tcp_sack_revert_scoreboard(struct scoreboard *scb, tcp_seq snd_una,
 727                            u_int maxseg)
 728 {
 729         struct sackblock *sb;
 730
 731         scb->sackblocks = scb->sackblocks_prev;
 732         scb->nblocks = 0;
 733         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
 734                 ++scb->nblocks;
 735         tcp_sack_ack_blocks(scb, snd_una);
 736         scb->lastfound = NULL;
 737 }
 738 #endif
 739
 740 #ifdef DEBUG_SACK_HISTORY
 741 static void
 742 tcp_sack_dump_history(const char *msg, const struct tcpcb *tp)
 743 {
 744         int i;
 745         static int ndumped;
 746
 747         /* only need a couple of these to debug most problems */
 748         if (++ndumped > 900)
 749                 return;
 750
 751         kprintf("%s:\tnsackhistory %d: ", msg, tp->nsackhistory);
 752         for (i = 0; i < tp->nsackhistory; ++i)
 753                 kprintf("[%u, %u) ", tp->sackhistory[i].rblk_start,
 754                     tp->sackhistory[i].rblk_end);
 755         kprintf("\n");
 756 }
 757 #else
 758 static __inline void
 759 tcp_sack_dump_history(const char *msg, const struct tcpcb *tp)
 760 {
 761 }
 762 #endif
 763
 764 /*
 765  * Remove old SACK blocks from the SACK history that have already been ACKed.
 766  */
 767 static void
 768 tcp_sack_ack_history(struct tcpcb *tp)
 769 {
 770         int i, nblocks, openslot;
 771
 772         tcp_sack_dump_history("before tcp_sack_ack_history", tp);
 773         nblocks = tp->nsackhistory;
 774         for (i = openslot = 0; i < nblocks; ++i) {
 775                 if (SEQ_LEQ(tp->sackhistory[i].rblk_end, tp->rcv_nxt)) {
 776                         --tp->nsackhistory;
 777                         continue;
 778                 }
 779                 if (SEQ_LT(tp->sackhistory[i].rblk_start, tp->rcv_nxt))
 780                         tp->sackhistory[i].rblk_start = tp->rcv_nxt;
 781                 if (i == openslot)
 782                         ++openslot;
 783                 else
 784                         tp->sackhistory[openslot++] = tp->sackhistory[i];
 785         }
 786         tcp_sack_dump_history("after tcp_sack_ack_history", tp);
 787         KASSERT(openslot == tp->nsackhistory,
 788             ("tcp_sack_ack_history miscounted: %d != %d",
 789             openslot, tp->nsackhistory));
 790 }
 791
 792 /*
 793  * Add or merge newblock into reported history.
 794  * Also remove or update SACK blocks that will be acked.
 795  */
 796 static void
 797 tcp_sack_update_reported_history(struct tcpcb *tp, tcp_seq start, tcp_seq end)
 798 {
 799         struct raw_sackblock copy[MAX_SACK_REPORT_BLOCKS];
 800         int i, cindex;
 801
 802         tcp_sack_dump_history("before tcp_sack_update_reported_history", tp);
 803         /*
 804          * Six cases:
 805          *      0) no overlap
 806          *      1) newblock == oldblock
 807          *      2) oldblock contains newblock
 808          *      3) newblock contains oldblock
 809          *      4) tail of oldblock overlaps or abuts start of newblock
 810          *      5) tail of newblock overlaps or abuts head of oldblock
 811          */
 812         for (i = cindex = 0; i < tp->nsackhistory; ++i) {
 813                 struct raw_sackblock *oldblock = &tp->sackhistory[i];
 814                 tcp_seq old_start = oldblock->rblk_start;
 815                 tcp_seq old_end = oldblock->rblk_end;
 816
 817                 if (SEQ_LT(end, old_start) || SEQ_GT(start, old_end)) {
 818                         /* Case 0:  no overlap.  Copy old block. */
 819                         copy[cindex++] = *oldblock;
 820                         continue;
 821                 }
 822
 823                 if (SEQ_GEQ(start, old_start) && SEQ_LEQ(end, old_end)) {
 824                         /* Cases 1 & 2.  Move block to front of history. */
 825                         int j;
 826
 827                         start = old_start;
 828                         end = old_end;
 829                         /* no need to check rest of blocks */
 830                         for (j = i + 1; j < tp->nsackhistory; ++j)
 831                                 copy[cindex++] = tp->sackhistory[j];
 832                         break;
 833                 }
 834
 835                 if (SEQ_GEQ(old_end, start) && SEQ_LT(old_start, start)) {
 836                         /* Case 4:  extend start of new block. */
 837                         start = old_start;
 838                 } else if (SEQ_GEQ(end, old_start) && SEQ_GT(old_end, end)) {
 839                         /* Case 5: extend end of new block */
 840                         end = old_end;
 841                 } else {
 842                         /* Case 3.  Delete old block by not copying it. */
 843                         KASSERT(SEQ_LEQ(start, old_start) &&
 844                                 SEQ_GEQ(end, old_end),
 845                             ("bad logic: old [%u, %u), new [%u, %u)",
 846                              old_start, old_end, start, end));
 847                 }
 848         }
 849
 850         /* insert new block */
 851         tp->sackhistory[0].rblk_start = start;
 852         tp->sackhistory[0].rblk_end = end;
 853         cindex = min(cindex, MAX_SACK_REPORT_BLOCKS - 1);
 854         for (i = 0; i < cindex; ++i)
 855                 tp->sackhistory[i + 1] = copy[i];
 856         tp->nsackhistory = cindex + 1;
 857         tcp_sack_dump_history("after tcp_sack_update_reported_history", tp);
 858 }
 859
 860 /*
 861  * Fill in SACK report to return to data sender.
 862  */
 863 void
 864 tcp_sack_fill_report(struct tcpcb *tp, u_char *opt, u_int *plen)
 865 {
 866         u_int optlen = *plen;
 867         uint32_t *lp = (uint32_t *)(opt + optlen);
 868         uint32_t *olp;
 869         tcp_seq hstart = tp->rcv_nxt, hend;
 870         int nblocks;
 871
 872         KASSERT(TCP_MAXOLEN - optlen >=
 873             TCPOLEN_SACK_ALIGNED + TCPOLEN_SACK_BLOCK,
 874             ("no room for SACK header and one block: optlen %d", optlen));
 875
 876         if (tp->sack_flags & TSACK_F_DUPSEG)
 877                 tcpstat.tcps_snddsackopt++;
 878         else
 879                 tcpstat.tcps_sndsackopt++;
 880
 881         olp = lp++;
 882         optlen += TCPOLEN_SACK_ALIGNED;
 883
 884         tcp_sack_ack_history(tp);
 885         if (tp->reportblk.rblk_start != tp->reportblk.rblk_end) {
 886                 *lp++ = htonl(tp->reportblk.rblk_start);
 887                 *lp++ = htonl(tp->reportblk.rblk_end);
 888                 optlen += TCPOLEN_SACK_BLOCK;
 889                 hstart = tp->reportblk.rblk_start;
 890                 hend = tp->reportblk.rblk_end;
 891                 if (tp->sack_flags & TSACK_F_ENCLOSESEG) {
 892                         KASSERT(TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK,
 893                             ("no room for enclosing SACK block: oplen %d",
 894                             optlen));
 895                         *lp++ = htonl(tp->encloseblk.rblk_start);
 896                         *lp++ = htonl(tp->encloseblk.rblk_end);
 897                         optlen += TCPOLEN_SACK_BLOCK;
 898                         hstart = tp->encloseblk.rblk_start;
 899                         hend = tp->encloseblk.rblk_end;
 900                 }
 901                 if (SEQ_GT(hstart, tp->rcv_nxt))
 902                         tcp_sack_update_reported_history(tp, hstart, hend);
 903         }
 904         if (tcp_do_smartsack && (tp->sack_flags & TSACK_F_SACKLEFT)) {
 905                 /* Fill in from left!  Walk re-assembly queue. */
 906                 struct tseg_qent *q;
 907
 908                 q = TAILQ_FIRST(&tp->t_segq);
 909                 while (q != NULL &&
 910                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
 911                         *lp++ = htonl(q->tqe_th->th_seq);
 912                         *lp++ = htonl(TCP_SACK_BLKEND(
 913                             q->tqe_th->th_seq + q->tqe_len,
 914                             q->tqe_th->th_flags));
 915                         optlen += TCPOLEN_SACK_BLOCK;
 916                         q = TAILQ_NEXT(q, tqe_q);
 917                 }
 918         } else {
 919                 int n = 0;
 920
 921                 /* Fill in SACK blocks from right side. */
 922                 while (n < tp->nsackhistory &&
 923                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
 924                         if (tp->sackhistory[n].rblk_start != hstart) {
 925                                 *lp++ = htonl(tp->sackhistory[n].rblk_start);
 926                                 *lp++ = htonl(tp->sackhistory[n].rblk_end);
 927                                 optlen += TCPOLEN_SACK_BLOCK;
 928                         }
 929                         ++n;
 930                 }
 931         }
 932         tp->reportblk.rblk_start = tp->reportblk.rblk_end;
 933         tp->sack_flags &=
 934             ~(TSACK_F_DUPSEG | TSACK_F_ENCLOSESEG | TSACK_F_SACKLEFT);
 935         nblocks = (lp - olp - 1) / 2;
 936         *olp = htonl(TCPOPT_SACK_ALIGNED |
 937                      (TCPOLEN_SACK + nblocks * TCPOLEN_SACK_BLOCK));
 938         *plen = optlen;
 939 }