sys/netinet/tcp_sack.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-3-Clause
   3  *
   4  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995
   5  *      The Regents of the University of California.
   6  * All rights reserved.
   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 University nor the names of its contributors
  17  *    may be used to endorse or promote products derived from this software
  18  *    without specific prior written permission.
  19  *
  20  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  30  * SUCH DAMAGE.
  31  *
  32  *      @(#)tcp_sack.c  8.12 (Berkeley) 5/24/95
  33  */
  34
  35 /*-
  36  *      @@(#)COPYRIGHT  1.1 (NRL) 17 January 1995
  37  *
  38  * NRL grants permission for redistribution and use in source and binary
  39  * forms, with or without modification, of the software and documentation
  40  * created at NRL provided that the following conditions are met:
  41  *
  42  * 1. Redistributions of source code must retain the above copyright
  43  *    notice, this list of conditions and the following disclaimer.
  44  * 2. Redistributions in binary form must reproduce the above copyright
  45  *    notice, this list of conditions and the following disclaimer in the
  46  *    documentation and/or other materials provided with the distribution.
  47  * 3. All advertising materials mentioning features or use of this software
  48  *    must display the following acknowledgements:
  49  *      This product includes software developed by the University of
  50  *      California, Berkeley and its contributors.
  51  *      This product includes software developed at the Information
  52  *      Technology Division, US Naval Research Laboratory.
  53  * 4. Neither the name of the NRL nor the names of its contributors
  54  *    may be used to endorse or promote products derived from this software
  55  *    without specific prior written permission.
  56  *
  57  * THE SOFTWARE PROVIDED BY NRL IS PROVIDED BY NRL AND CONTRIBUTORS ``AS
  58  * IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
  59  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
  60  * PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL NRL OR
  61  * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
  62  * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
  63  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
  64  * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
  65  * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
  66  * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
  67  * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  68  *
  69  * The views and conclusions contained in the software and documentation
  70  * are those of the authors and should not be interpreted as representing
  71  * official policies, either expressed or implied, of the US Naval
  72  * Research Laboratory (NRL).
  73  */
  74
  75 #include <sys/cdefs.h>
  76 __FBSDID("$FreeBSD$");
  77
  78 #include "opt_inet.h"
  79 #include "opt_inet6.h"
  80 #include "opt_tcpdebug.h"
  81
  82 #include <sys/param.h>
  83 #include <sys/systm.h>
  84 #include <sys/kernel.h>
  85 #include <sys/sysctl.h>
  86 #include <sys/malloc.h>
  87 #include <sys/mbuf.h>
  88 #include <sys/proc.h>           /* for proc0 declaration */
  89 #include <sys/protosw.h>
  90 #include <sys/socket.h>
  91 #include <sys/socketvar.h>
  92 #include <sys/syslog.h>
  93 #include <sys/systm.h>
  94
  95 #include <machine/cpu.h>        /* before tcp_seq.h, for tcp_random18() */
  96
  97 #include <vm/uma.h>
  98
  99 #include <net/if.h>
 100 #include <net/if_var.h>
 101 #include <net/route.h>
 102 #include <net/vnet.h>
 103
 104 #include <netinet/in.h>
 105 #include <netinet/in_systm.h>
 106 #include <netinet/ip.h>
 107 #include <netinet/in_var.h>
 108 #include <netinet/in_pcb.h>
 109 #include <netinet/ip_var.h>
 110 #include <netinet/ip6.h>
 111 #include <netinet/icmp6.h>
 112 #include <netinet6/nd6.h>
 113 #include <netinet6/ip6_var.h>
 114 #include <netinet6/in6_pcb.h>
 115 #include <netinet/tcp.h>
 116 #include <netinet/tcp_fsm.h>
 117 #include <netinet/tcp_seq.h>
 118 #include <netinet/tcp_timer.h>
 119 #include <netinet/tcp_var.h>
 120 #include <netinet/tcpip.h>
 121 #include <netinet/cc/cc.h>
 122 #ifdef TCPDEBUG
 123 #include <netinet/tcp_debug.h>
 124 #endif /* TCPDEBUG */
 125
 126 #include <machine/in_cksum.h>
 127
 128 VNET_DECLARE(struct uma_zone *, sack_hole_zone);
 129 #define V_sack_hole_zone                VNET(sack_hole_zone)
 130
 131 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, sack, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 132     "TCP SACK");
 133
 134 VNET_DEFINE(int, tcp_do_sack) = 1;
 135 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, enable, CTLFLAG_VNET | CTLFLAG_RW,
 136     &VNET_NAME(tcp_do_sack), 0,
 137     "Enable/Disable TCP SACK support");
 138
 139 VNET_DEFINE(int, tcp_do_newsack) = 1;
 140 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, revised, CTLFLAG_VNET | CTLFLAG_RW,
 141     &VNET_NAME(tcp_do_newsack), 0,
 142     "Use revised SACK loss recovery per RFC 6675");
 143
 144 VNET_DEFINE(int, tcp_sack_maxholes) = 128;
 145 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, maxholes, CTLFLAG_VNET | CTLFLAG_RW,
 146     &VNET_NAME(tcp_sack_maxholes), 0,
 147     "Maximum number of TCP SACK holes allowed per connection");
 148
 149 VNET_DEFINE(int, tcp_sack_globalmaxholes) = 65536;
 150 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, globalmaxholes, CTLFLAG_VNET | CTLFLAG_RW,
 151     &VNET_NAME(tcp_sack_globalmaxholes), 0,
 152     "Global maximum number of TCP SACK holes");
 153
 154 VNET_DEFINE(int, tcp_sack_globalholes) = 0;
 155 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, globalholes, CTLFLAG_VNET | CTLFLAG_RD,
 156     &VNET_NAME(tcp_sack_globalholes), 0,
 157     "Global number of TCP SACK holes currently allocated");
 158
 159 int
 160 tcp_dsack_block_exists(struct tcpcb *tp)
 161 {
 162         /* Return true if a DSACK block exists */
 163         if (tp->rcv_numsacks == 0)
 164                 return (0);
 165         if (SEQ_LEQ(tp->sackblks[0].end, tp->rcv_nxt))
 166                 return(1);
 167         return (0);
 168 }
 169
 170 /*
 171  * This function will find overlaps with the currently stored sackblocks
 172  * and add any overlap as a dsack block upfront
 173  */
 174 void
 175 tcp_update_dsack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
 176 {
 177         struct sackblk head_blk,mid_blk,saved_blks[MAX_SACK_BLKS];
 178         int i, j, n, identical;
 179         tcp_seq start, end;
 180
 181         INP_WLOCK_ASSERT(tptoinpcb(tp));
 182
 183         KASSERT(SEQ_LT(rcv_start, rcv_end), ("rcv_start < rcv_end"));
 184
 185         if (SEQ_LT(rcv_end, tp->rcv_nxt) ||
 186             ((rcv_end == tp->rcv_nxt) &&
 187              (tp->rcv_numsacks > 0 ) &&
 188              (tp->sackblks[0].end == tp->rcv_nxt))) {
 189                 saved_blks[0].start = rcv_start;
 190                 saved_blks[0].end = rcv_end;
 191         } else {
 192                 saved_blks[0].start = saved_blks[0].end = 0;
 193         }
 194
 195         head_blk.start = head_blk.end = 0;
 196         mid_blk.start = rcv_start;
 197         mid_blk.end = rcv_end;
 198         identical = 0;
 199
 200         for (i = 0; i < tp->rcv_numsacks; i++) {
 201                 start = tp->sackblks[i].start;
 202                 end = tp->sackblks[i].end;
 203                 if (SEQ_LT(rcv_end, start)) {
 204                         /* pkt left to sack blk */
 205                         continue;
 206                 }
 207                 if (SEQ_GT(rcv_start, end)) {
 208                         /* pkt right to sack blk */
 209                         continue;
 210                 }
 211                 if (SEQ_GT(tp->rcv_nxt, end)) {
 212                         if ((SEQ_MAX(rcv_start, start) != SEQ_MIN(rcv_end, end)) &&
 213                             (SEQ_GT(head_blk.start, SEQ_MAX(rcv_start, start)) ||
 214                             (head_blk.start == head_blk.end))) {
 215                                 head_blk.start = SEQ_MAX(rcv_start, start);
 216                                 head_blk.end = SEQ_MIN(rcv_end, end);
 217                         }
 218                         continue;
 219                 }
 220                 if (((head_blk.start == head_blk.end) ||
 221                      SEQ_LT(start, head_blk.start)) &&
 222                      (SEQ_GT(end, rcv_start) &&
 223                       SEQ_LEQ(start, rcv_end))) {
 224                         head_blk.start = start;
 225                         head_blk.end = end;
 226                 }
 227                 mid_blk.start = SEQ_MIN(mid_blk.start, start);
 228                 mid_blk.end = SEQ_MAX(mid_blk.end, end);
 229                 if ((mid_blk.start == start) &&
 230                     (mid_blk.end == end))
 231                         identical = 1;
 232         }
 233         if (SEQ_LT(head_blk.start, head_blk.end)) {
 234                 /* store overlapping range */
 235                 saved_blks[0].start = SEQ_MAX(rcv_start, head_blk.start);
 236                 saved_blks[0].end   = SEQ_MIN(rcv_end, head_blk.end);
 237         }
 238         n = 1;
 239         /*
 240          * Second, if not ACKed, store the SACK block that
 241          * overlaps with the DSACK block unless it is identical
 242          */
 243         if ((SEQ_LT(tp->rcv_nxt, mid_blk.end) &&
 244             !((mid_blk.start == saved_blks[0].start) &&
 245             (mid_blk.end == saved_blks[0].end))) ||
 246             identical == 1) {
 247                 saved_blks[n].start = mid_blk.start;
 248                 saved_blks[n++].end = mid_blk.end;
 249         }
 250         for (j = 0; (j < tp->rcv_numsacks) && (n < MAX_SACK_BLKS); j++) {
 251                 if (((SEQ_LT(tp->sackblks[j].end, mid_blk.start) ||
 252                       SEQ_GT(tp->sackblks[j].start, mid_blk.end)) &&
 253                     (SEQ_GT(tp->sackblks[j].start, tp->rcv_nxt))))
 254                 saved_blks[n++] = tp->sackblks[j];
 255         }
 256         j = 0;
 257         for (i = 0; i < n; i++) {
 258                 /* we can end up with a stale initial entry */
 259                 if (SEQ_LT(saved_blks[i].start, saved_blks[i].end)) {
 260                         tp->sackblks[j++] = saved_blks[i];
 261                 }
 262         }
 263         tp->rcv_numsacks = j;
 264 }
 265
 266 /*
 267  * This function is called upon receipt of new valid data (while not in
 268  * header prediction mode), and it updates the ordered list of sacks.
 269  */
 270 void
 271 tcp_update_sack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
 272 {
 273         /*
 274          * First reported block MUST be the most recent one.  Subsequent
 275          * blocks SHOULD be in the order in which they arrived at the
 276          * receiver.  These two conditions make the implementation fully
 277          * compliant with RFC 2018.
 278          */
 279         struct sackblk head_blk, saved_blks[MAX_SACK_BLKS];
 280         int num_head, num_saved, i;
 281
 282         INP_WLOCK_ASSERT(tptoinpcb(tp));
 283
 284         /* Check arguments. */
 285         KASSERT(SEQ_LEQ(rcv_start, rcv_end), ("rcv_start <= rcv_end"));
 286
 287         if ((rcv_start == rcv_end) &&
 288             (tp->rcv_numsacks >= 1) &&
 289             (rcv_end == tp->sackblks[0].end)) {
 290                 /* retaining DSACK block below rcv_nxt (todrop) */
 291                 head_blk = tp->sackblks[0];
 292         } else {
 293                 /* SACK block for the received segment. */
 294                 head_blk.start = rcv_start;
 295                 head_blk.end = rcv_end;
 296         }
 297
 298         /*
 299          * Merge updated SACK blocks into head_blk, and save unchanged SACK
 300          * blocks into saved_blks[].  num_saved will have the number of the
 301          * saved SACK blocks.
 302          */
 303         num_saved = 0;
 304         for (i = 0; i < tp->rcv_numsacks; i++) {
 305                 tcp_seq start = tp->sackblks[i].start;
 306                 tcp_seq end = tp->sackblks[i].end;
 307                 if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
 308                         /*
 309                          * Discard this SACK block.
 310                          */
 311                 } else if (SEQ_LEQ(head_blk.start, end) &&
 312                            SEQ_GEQ(head_blk.end, start)) {
 313                         /*
 314                          * Merge this SACK block into head_blk.  This SACK
 315                          * block itself will be discarded.
 316                          */
 317                         /*
 318                          * |-|
 319                          *   |---|  merge
 320                          *
 321                          *     |-|
 322                          * |---|    merge
 323                          *
 324                          * |-----|
 325                          *   |-|    DSACK smaller
 326                          *
 327                          *   |-|
 328                          * |-----|  DSACK smaller
 329                          */
 330                         if (head_blk.start == end)
 331                                 head_blk.start = start;
 332                         else if (head_blk.end == start)
 333                                 head_blk.end = end;
 334                         else {
 335                                 if (SEQ_LT(head_blk.start, start)) {
 336                                         tcp_seq temp = start;
 337                                         start = head_blk.start;
 338                                         head_blk.start = temp;
 339                                 }
 340                                 if (SEQ_GT(head_blk.end, end)) {
 341                                         tcp_seq temp = end;
 342                                         end = head_blk.end;
 343                                         head_blk.end = temp;
 344                                 }
 345                                 if ((head_blk.start != start) ||
 346                                     (head_blk.end != end)) {
 347                                         if ((num_saved >= 1) &&
 348                                            SEQ_GEQ(saved_blks[num_saved-1].start, start) &&
 349                                            SEQ_LEQ(saved_blks[num_saved-1].end, end))
 350                                                 num_saved--;
 351                                         saved_blks[num_saved].start = start;
 352                                         saved_blks[num_saved].end = end;
 353                                         num_saved++;
 354                                 }
 355                         }
 356                 } else {
 357                         /*
 358                          * This block supercedes the prior block
 359                          */
 360                         if ((num_saved >= 1) &&
 361                            SEQ_GEQ(saved_blks[num_saved-1].start, start) &&
 362                            SEQ_LEQ(saved_blks[num_saved-1].end, end))
 363                                 num_saved--;
 364                         /*
 365                          * Save this SACK block.
 366                          */
 367                         saved_blks[num_saved].start = start;
 368                         saved_blks[num_saved].end = end;
 369                         num_saved++;
 370                 }
 371         }
 372
 373         /*
 374          * Update SACK list in tp->sackblks[].
 375          */
 376         num_head = 0;
 377         if (SEQ_LT(rcv_start, rcv_end)) {
 378                 /*
 379                  * The received data segment is an out-of-order segment.  Put
 380                  * head_blk at the top of SACK list.
 381                  */
 382                 tp->sackblks[0] = head_blk;
 383                 num_head = 1;
 384                 /*
 385                  * If the number of saved SACK blocks exceeds its limit,
 386                  * discard the last SACK block.
 387                  */
 388                 if (num_saved >= MAX_SACK_BLKS)
 389                         num_saved--;
 390         }
 391         if ((rcv_start == rcv_end) &&
 392             (rcv_start == tp->sackblks[0].end)) {
 393                 num_head = 1;
 394         }
 395         if (num_saved > 0) {
 396                 /*
 397                  * Copy the saved SACK blocks back.
 398                  */
 399                 bcopy(saved_blks, &tp->sackblks[num_head],
 400                       sizeof(struct sackblk) * num_saved);
 401         }
 402
 403         /* Save the number of SACK blocks. */
 404         tp->rcv_numsacks = num_head + num_saved;
 405 }
 406
 407 void
 408 tcp_clean_dsack_blocks(struct tcpcb *tp)
 409 {
 410         struct sackblk saved_blks[MAX_SACK_BLKS];
 411         int num_saved, i;
 412
 413         INP_WLOCK_ASSERT(tptoinpcb(tp));
 414         /*
 415          * Clean up any DSACK blocks that
 416          * are in our queue of sack blocks.
 417          *
 418          */
 419         num_saved = 0;
 420         for (i = 0; i < tp->rcv_numsacks; i++) {
 421                 tcp_seq start = tp->sackblks[i].start;
 422                 tcp_seq end = tp->sackblks[i].end;
 423                 if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
 424                         /*
 425                          * Discard this D-SACK block.
 426                          */
 427                         continue;
 428                 }
 429                 /*
 430                  * Save this SACK block.
 431                  */
 432                 saved_blks[num_saved].start = start;
 433                 saved_blks[num_saved].end = end;
 434                 num_saved++;
 435         }
 436         if (num_saved > 0) {
 437                 /*
 438                  * Copy the saved SACK blocks back.
 439                  */
 440                 bcopy(saved_blks, &tp->sackblks[0],
 441                       sizeof(struct sackblk) * num_saved);
 442         }
 443         tp->rcv_numsacks = num_saved;
 444 }
 445
 446 /*
 447  * Delete all receiver-side SACK information.
 448  */
 449 void
 450 tcp_clean_sackreport(struct tcpcb *tp)
 451 {
 452         int i;
 453
 454         INP_WLOCK_ASSERT(tptoinpcb(tp));
 455         tp->rcv_numsacks = 0;
 456         for (i = 0; i < MAX_SACK_BLKS; i++)
 457                 tp->sackblks[i].start = tp->sackblks[i].end=0;
 458 }
 459
 460 /*
 461  * Allocate struct sackhole.
 462  */
 463 static struct sackhole *
 464 tcp_sackhole_alloc(struct tcpcb *tp, tcp_seq start, tcp_seq end)
 465 {
 466         struct sackhole *hole;
 467
 468         if (tp->snd_numholes >= V_tcp_sack_maxholes ||
 469             V_tcp_sack_globalholes >= V_tcp_sack_globalmaxholes) {
 470                 TCPSTAT_INC(tcps_sack_sboverflow);
 471                 return NULL;
 472         }
 473
 474         hole = (struct sackhole *)uma_zalloc(V_sack_hole_zone, M_NOWAIT);
 475         if (hole == NULL)
 476                 return NULL;
 477
 478         hole->start = start;
 479         hole->end = end;
 480         hole->rxmit = start;
 481
 482         tp->snd_numholes++;
 483         atomic_add_int(&V_tcp_sack_globalholes, 1);
 484
 485         return hole;
 486 }
 487
 488 /*
 489  * Free struct sackhole.
 490  */
 491 static void
 492 tcp_sackhole_free(struct tcpcb *tp, struct sackhole *hole)
 493 {
 494
 495         uma_zfree(V_sack_hole_zone, hole);
 496
 497         tp->snd_numholes--;
 498         atomic_subtract_int(&V_tcp_sack_globalholes, 1);
 499
 500         KASSERT(tp->snd_numholes >= 0, ("tp->snd_numholes >= 0"));
 501         KASSERT(V_tcp_sack_globalholes >= 0, ("tcp_sack_globalholes >= 0"));
 502 }
 503
 504 /*
 505  * Insert new SACK hole into scoreboard.
 506  */
 507 static struct sackhole *
 508 tcp_sackhole_insert(struct tcpcb *tp, tcp_seq start, tcp_seq end,
 509     struct sackhole *after)
 510 {
 511         struct sackhole *hole;
 512
 513         /* Allocate a new SACK hole. */
 514         hole = tcp_sackhole_alloc(tp, start, end);
 515         if (hole == NULL)
 516                 return NULL;
 517
 518         /* Insert the new SACK hole into scoreboard. */
 519         if (after != NULL)
 520                 TAILQ_INSERT_AFTER(&tp->snd_holes, after, hole, scblink);
 521         else
 522                 TAILQ_INSERT_TAIL(&tp->snd_holes, hole, scblink);
 523
 524         /* Update SACK hint. */
 525         if (tp->sackhint.nexthole == NULL)
 526                 tp->sackhint.nexthole = hole;
 527
 528         return hole;
 529 }
 530
 531 /*
 532  * Remove SACK hole from scoreboard.
 533  */
 534 static void
 535 tcp_sackhole_remove(struct tcpcb *tp, struct sackhole *hole)
 536 {
 537
 538         /* Update SACK hint. */
 539         if (tp->sackhint.nexthole == hole)
 540                 tp->sackhint.nexthole = TAILQ_NEXT(hole, scblink);
 541
 542         /* Remove this SACK hole. */
 543         TAILQ_REMOVE(&tp->snd_holes, hole, scblink);
 544
 545         /* Free this SACK hole. */
 546         tcp_sackhole_free(tp, hole);
 547 }
 548
 549 /*
 550  * Process cumulative ACK and the TCP SACK option to update the scoreboard.
 551  * tp->snd_holes is an ordered list of holes (oldest to newest, in terms of
 552  * the sequence space).
 553  * Returns 1 if incoming ACK has previously unknown SACK information,
 554  * 0 otherwise.
 555  */
 556 int
 557 tcp_sack_doack(struct tcpcb *tp, struct tcpopt *to, tcp_seq th_ack)
 558 {
 559         struct sackhole *cur, *temp;
 560         struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1], *sblkp;
 561         int i, j, num_sack_blks, sack_changed;
 562         int delivered_data, left_edge_delta;
 563
 564         INP_WLOCK_ASSERT(tptoinpcb(tp));
 565
 566         num_sack_blks = 0;
 567         sack_changed = 0;
 568         delivered_data = 0;
 569         left_edge_delta = 0;
 570         /*
 571          * If SND.UNA will be advanced by SEG.ACK, and if SACK holes exist,
 572          * treat [SND.UNA, SEG.ACK) as if it is a SACK block.
 573          * Account changes to SND.UNA always in delivered data.
 574          */
 575         if (SEQ_LT(tp->snd_una, th_ack) && !TAILQ_EMPTY(&tp->snd_holes)) {
 576                 left_edge_delta = th_ack - tp->snd_una;
 577                 sack_blocks[num_sack_blks].start = tp->snd_una;
 578                 sack_blocks[num_sack_blks++].end = th_ack;
 579                 /*
 580                  * Pulling snd_fack forward if we got here
 581                  * due to DSACK blocks
 582                  */
 583                 if (SEQ_LT(tp->snd_fack, th_ack)) {
 584                         delivered_data += th_ack - tp->snd_una;
 585                         tp->snd_fack = th_ack;
 586                         sack_changed = 1;
 587                 }
 588         }
 589         /*
 590          * Append received valid SACK blocks to sack_blocks[], but only if we
 591          * received new blocks from the other side.
 592          */
 593         if (to->to_flags & TOF_SACK) {
 594                 for (i = 0; i < to->to_nsacks; i++) {
 595                         bcopy((to->to_sacks + i * TCPOLEN_SACK),
 596                             &sack, sizeof(sack));
 597                         sack.start = ntohl(sack.start);
 598                         sack.end = ntohl(sack.end);
 599                         if (SEQ_GT(sack.end, sack.start) &&
 600                             SEQ_GT(sack.start, tp->snd_una) &&
 601                             SEQ_GT(sack.start, th_ack) &&
 602                             SEQ_LT(sack.start, tp->snd_max) &&
 603                             SEQ_GT(sack.end, tp->snd_una) &&
 604                             SEQ_LEQ(sack.end, tp->snd_max)) {
 605                                 sack_blocks[num_sack_blks++] = sack;
 606                         } else if (SEQ_LEQ(sack.start, th_ack) &&
 607                             SEQ_LEQ(sack.end, th_ack)) {
 608                                 /*
 609                                  * Its a D-SACK block.
 610                                  */
 611                                 tcp_record_dsack(tp, sack.start, sack.end, 0);
 612                         }
 613                 }
 614         }
 615         /*
 616          * Return if SND.UNA is not advanced and no valid SACK block is
 617          * received.
 618          */
 619         if (num_sack_blks == 0)
 620                 return (sack_changed);
 621
 622         /*
 623          * Sort the SACK blocks so we can update the scoreboard with just one
 624          * pass. The overhead of sorting up to 4+1 elements is less than
 625          * making up to 4+1 passes over the scoreboard.
 626          */
 627         for (i = 0; i < num_sack_blks; i++) {
 628                 for (j = i + 1; j < num_sack_blks; j++) {
 629                         if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
 630                                 sack = sack_blocks[i];
 631                                 sack_blocks[i] = sack_blocks[j];
 632                                 sack_blocks[j] = sack;
 633                         }
 634                 }
 635         }
 636         if (TAILQ_EMPTY(&tp->snd_holes)) {
 637                 /*
 638                  * Empty scoreboard. Need to initialize snd_fack (it may be
 639                  * uninitialized or have a bogus value). Scoreboard holes
 640                  * (from the sack blocks received) are created later below
 641                  * (in the logic that adds holes to the tail of the
 642                  * scoreboard).
 643                  */
 644                 tp->snd_fack = SEQ_MAX(tp->snd_una, th_ack);
 645                 tp->sackhint.sacked_bytes = 0;  /* reset */
 646         }
 647         /*
 648          * In the while-loop below, incoming SACK blocks (sack_blocks[]) and
 649          * SACK holes (snd_holes) are traversed from their tails with just
 650          * one pass in order to reduce the number of compares especially when
 651          * the bandwidth-delay product is large.
 652          *
 653          * Note: Typically, in the first RTT of SACK recovery, the highest
 654          * three or four SACK blocks with the same ack number are received.
 655          * In the second RTT, if retransmitted data segments are not lost,
 656          * the highest three or four SACK blocks with ack number advancing
 657          * are received.
 658          */
 659         sblkp = &sack_blocks[num_sack_blks - 1];        /* Last SACK block */
 660         tp->sackhint.last_sack_ack = sblkp->end;
 661         if (SEQ_LT(tp->snd_fack, sblkp->start)) {
 662                 /*
 663                  * The highest SACK block is beyond fack.  First,
 664                  * check if there was a successful Rescue Retransmission,
 665                  * and move this hole left. With normal holes, snd_fack
 666                  * is always to the right of the end.
 667                  */
 668                 if (((temp = TAILQ_LAST(&tp->snd_holes, sackhole_head)) != NULL) &&
 669                     SEQ_LEQ(tp->snd_fack,temp->end)) {
 670                         temp->start = SEQ_MAX(tp->snd_fack, SEQ_MAX(tp->snd_una, th_ack));
 671                         temp->end = sblkp->start;
 672                         temp->rxmit = temp->start;
 673                         delivered_data += sblkp->end - sblkp->start;
 674                         tp->snd_fack = sblkp->end;
 675                         sblkp--;
 676                         sack_changed = 1;
 677                 } else {
 678                         /*
 679                          * Append a new SACK hole at the tail.  If the
 680                          * second or later highest SACK blocks are also
 681                          * beyond the current fack, they will be inserted
 682                          * by way of hole splitting in the while-loop below.
 683                          */
 684                         temp = tcp_sackhole_insert(tp, tp->snd_fack,sblkp->start,NULL);
 685                         if (temp != NULL) {
 686                                 delivered_data += sblkp->end - sblkp->start;
 687                                 tp->snd_fack = sblkp->end;
 688                                 /* Go to the previous sack block. */
 689                                 sblkp--;
 690                                 sack_changed = 1;
 691                         } else {
 692                                 /*
 693                                  * We failed to add a new hole based on the current
 694                                  * sack block.  Skip over all the sack blocks that
 695                                  * fall completely to the right of snd_fack and
 696                                  * proceed to trim the scoreboard based on the
 697                                  * remaining sack blocks.  This also trims the
 698                                  * scoreboard for th_ack (which is sack_blocks[0]).
 699                                  */
 700                                 while (sblkp >= sack_blocks &&
 701                                        SEQ_LT(tp->snd_fack, sblkp->start))
 702                                         sblkp--;
 703                                 if (sblkp >= sack_blocks &&
 704                                     SEQ_LT(tp->snd_fack, sblkp->end)) {
 705                                         delivered_data += sblkp->end - tp->snd_fack;
 706                                         tp->snd_fack = sblkp->end;
 707                                         sack_changed = 1;
 708                                 }
 709                         }
 710                 }
 711         } else if (SEQ_LT(tp->snd_fack, sblkp->end)) {
 712                 /* fack is advanced. */
 713                 delivered_data += sblkp->end - tp->snd_fack;
 714                 tp->snd_fack = sblkp->end;
 715                 sack_changed = 1;
 716         }
 717         cur = TAILQ_LAST(&tp->snd_holes, sackhole_head); /* Last SACK hole. */
 718         /*
 719          * Since the incoming sack blocks are sorted, we can process them
 720          * making one sweep of the scoreboard.
 721          */
 722         while (sblkp >= sack_blocks  && cur != NULL) {
 723                 if (SEQ_GEQ(sblkp->start, cur->end)) {
 724                         /*
 725                          * SACKs data beyond the current hole.  Go to the
 726                          * previous sack block.
 727                          */
 728                         sblkp--;
 729                         continue;
 730                 }
 731                 if (SEQ_LEQ(sblkp->end, cur->start)) {
 732                         /*
 733                          * SACKs data before the current hole.  Go to the
 734                          * previous hole.
 735                          */
 736                         cur = TAILQ_PREV(cur, sackhole_head, scblink);
 737                         continue;
 738                 }
 739                 tp->sackhint.sack_bytes_rexmit -=
 740                     (SEQ_MIN(cur->rxmit, cur->end) - cur->start);
 741                 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
 742                     ("sackhint bytes rtx >= 0"));
 743                 sack_changed = 1;
 744                 if (SEQ_LEQ(sblkp->start, cur->start)) {
 745                         /* Data acks at least the beginning of hole. */
 746                         if (SEQ_GEQ(sblkp->end, cur->end)) {
 747                                 /* Acks entire hole, so delete hole. */
 748                                 delivered_data += (cur->end - cur->start);
 749                                 temp = cur;
 750                                 cur = TAILQ_PREV(cur, sackhole_head, scblink);
 751                                 tcp_sackhole_remove(tp, temp);
 752                                 /*
 753                                  * The sack block may ack all or part of the
 754                                  * next hole too, so continue onto the next
 755                                  * hole.
 756                                  */
 757                                 continue;
 758                         } else {
 759                                 /* Move start of hole forward. */
 760                                 delivered_data += (sblkp->end - cur->start);
 761                                 cur->start = sblkp->end;
 762                                 cur->rxmit = SEQ_MAX(cur->rxmit, cur->start);
 763                         }
 764                 } else {
 765                         /* Data acks at least the end of hole. */
 766                         if (SEQ_GEQ(sblkp->end, cur->end)) {
 767                                 /* Move end of hole backward. */
 768                                 delivered_data += (cur->end - sblkp->start);
 769                                 cur->end = sblkp->start;
 770                                 cur->rxmit = SEQ_MIN(cur->rxmit, cur->end);
 771                                 if ((tp->t_flags & TF_LRD) && SEQ_GEQ(cur->rxmit, cur->end))
 772                                         cur->rxmit = tp->snd_recover;
 773                         } else {
 774                                 /*
 775                                  * ACKs some data in middle of a hole; need
 776                                  * to split current hole
 777                                  */
 778                                 temp = tcp_sackhole_insert(tp, sblkp->end,
 779                                     cur->end, cur);
 780                                 if (temp != NULL) {
 781                                         if (SEQ_GT(cur->rxmit, temp->rxmit)) {
 782                                                 temp->rxmit = cur->rxmit;
 783                                                 tp->sackhint.sack_bytes_rexmit +=
 784                                                     (SEQ_MIN(temp->rxmit,
 785                                                     temp->end) - temp->start);
 786                                         }
 787                                         cur->end = sblkp->start;
 788                                         cur->rxmit = SEQ_MIN(cur->rxmit,
 789                                             cur->end);
 790                                         if ((tp->t_flags & TF_LRD) && SEQ_GEQ(cur->rxmit, cur->end))
 791                                                 cur->rxmit = tp->snd_recover;
 792                                         delivered_data += (sblkp->end - sblkp->start);
 793                                 }
 794                         }
 795                 }
 796                 tp->sackhint.sack_bytes_rexmit +=
 797                     (SEQ_MIN(cur->rxmit, cur->end) - cur->start);
 798                 /*
 799                  * Testing sblkp->start against cur->start tells us whether
 800                  * we're done with the sack block or the sack hole.
 801                  * Accordingly, we advance one or the other.
 802                  */
 803                 if (SEQ_LEQ(sblkp->start, cur->start))
 804                         cur = TAILQ_PREV(cur, sackhole_head, scblink);
 805                 else
 806                         sblkp--;
 807         }
 808         if (!(to->to_flags & TOF_SACK))
 809                 /*
 810                  * If this ACK did not contain any
 811                  * SACK blocks, any only moved the
 812                  * left edge right, it is a pure
 813                  * cumulative ACK. Do not count
 814                  * DupAck for this. Also required
 815                  * for RFC6675 rescue retransmission.
 816                  */
 817                 sack_changed = 0;
 818         tp->sackhint.delivered_data = delivered_data;
 819         tp->sackhint.sacked_bytes += delivered_data - left_edge_delta;
 820         KASSERT((delivered_data >= 0), ("delivered_data < 0"));
 821         KASSERT((tp->sackhint.sacked_bytes >= 0), ("sacked_bytes < 0"));
 822         return (sack_changed);
 823 }
 824
 825 /*
 826  * Free all SACK holes to clear the scoreboard.
 827  */
 828 void
 829 tcp_free_sackholes(struct tcpcb *tp)
 830 {
 831         struct sackhole *q;
 832
 833         INP_WLOCK_ASSERT(tptoinpcb(tp));
 834         while ((q = TAILQ_FIRST(&tp->snd_holes)) != NULL)
 835                 tcp_sackhole_remove(tp, q);
 836         tp->sackhint.sack_bytes_rexmit = 0;
 837
 838         KASSERT(tp->snd_numholes == 0, ("tp->snd_numholes == 0"));
 839         KASSERT(tp->sackhint.nexthole == NULL,
 840                 ("tp->sackhint.nexthole == NULL"));
 841 }
 842
 843 /*
 844  * Partial ack handling within a sack recovery episode.  Keeping this very
 845  * simple for now.  When a partial ack is received, force snd_cwnd to a value
 846  * that will allow the sender to transmit no more than 2 segments.  If
 847  * necessary, a better scheme can be adopted at a later point, but for now,
 848  * the goal is to prevent the sender from bursting a large amount of data in
 849  * the midst of sack recovery.
 850  */
 851 void
 852 tcp_sack_partialack(struct tcpcb *tp, struct tcphdr *th)
 853 {
 854         int num_segs = 1;
 855         u_int maxseg = tcp_maxseg(tp);
 856
 857         INP_WLOCK_ASSERT(tptoinpcb(tp));
 858         tcp_timer_activate(tp, TT_REXMT, 0);
 859         tp->t_rtttime = 0;
 860         /* Send one or 2 segments based on how much new data was acked. */
 861         if ((BYTES_THIS_ACK(tp, th) / maxseg) >= 2)
 862                 num_segs = 2;
 863         tp->snd_cwnd = (tp->sackhint.sack_bytes_rexmit +
 864             (tp->snd_nxt - tp->snd_recover) + num_segs * maxseg);
 865         if (tp->snd_cwnd > tp->snd_ssthresh)
 866                 tp->snd_cwnd = tp->snd_ssthresh;
 867         tp->t_flags |= TF_ACKNOW;
 868         /*
 869          * RFC6675 rescue retransmission
 870          * Add a hole between th_ack (snd_una is not yet set) and snd_max,
 871          * if this was a pure cumulative ACK and no data was send beyond
 872          * recovery point. Since the data in the socket has not been freed
 873          * at this point, we check if the scoreboard is empty, and the ACK
 874          * delivered some new data, indicating a full ACK. Also, if the
 875          * recovery point is still at snd_max, we are probably application
 876          * limited. However, this inference might not always be true. The
 877          * rescue retransmission may rarely be slightly premature
 878          * compared to RFC6675.
 879          * The corresponding ACK+SACK will cause any further outstanding
 880          * segments to be retransmitted. This addresses a corner case, when
 881          * the trailing packets of a window are lost and no further data
 882          * is available for sending.
 883          */
 884         if ((V_tcp_do_newsack) &&
 885             SEQ_LT(th->th_ack, tp->snd_recover) &&
 886             (tp->snd_recover == tp->snd_max) &&
 887             TAILQ_EMPTY(&tp->snd_holes) &&
 888             (tp->sackhint.delivered_data > 0)) {
 889                 /*
 890                  * Exclude FIN sequence space in
 891                  * the hole for the rescue retransmission,
 892                  * and also don't create a hole, if only
 893                  * the ACK for a FIN is outstanding.
 894                  */
 895                 tcp_seq highdata = tp->snd_max;
 896                 if (tp->t_flags & TF_SENTFIN)
 897                         highdata--;
 898                 if (th->th_ack != highdata) {
 899                         tp->snd_fack = th->th_ack;
 900                         (void)tcp_sackhole_insert(tp, SEQ_MAX(th->th_ack,
 901                             highdata - maxseg), highdata, NULL);
 902                 }
 903         }
 904         (void) tcp_output(tp);
 905 }
 906
 907 #if 0
 908 /*
 909  * Debug version of tcp_sack_output() that walks the scoreboard.  Used for
 910  * now to sanity check the hint.
 911  */
 912 static struct sackhole *
 913 tcp_sack_output_debug(struct tcpcb *tp, int *sack_bytes_rexmt)
 914 {
 915         struct sackhole *p;
 916
 917         INP_WLOCK_ASSERT(tptoinpcb(tp));
 918         *sack_bytes_rexmt = 0;
 919         TAILQ_FOREACH(p, &tp->snd_holes, scblink) {
 920                 if (SEQ_LT(p->rxmit, p->end)) {
 921                         if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */
 922                                 continue;
 923                         }
 924                         *sack_bytes_rexmt += (p->rxmit - p->start);
 925                         break;
 926                 }
 927                 *sack_bytes_rexmt += (SEQ_MIN(p->rxmit, p->end) - p->start);
 928         }
 929         return (p);
 930 }
 931 #endif
 932
 933 /*
 934  * Returns the next hole to retransmit and the number of retransmitted bytes
 935  * from the scoreboard.  We store both the next hole and the number of
 936  * retransmitted bytes as hints (and recompute these on the fly upon SACK/ACK
 937  * reception).  This avoids scoreboard traversals completely.
 938  *
 939  * The loop here will traverse *at most* one link.  Here's the argument.  For
 940  * the loop to traverse more than 1 link before finding the next hole to
 941  * retransmit, we would need to have at least 1 node following the current
 942  * hint with (rxmit == end).  But, for all holes following the current hint,
 943  * (start == rxmit), since we have not yet retransmitted from them.
 944  * Therefore, in order to traverse more 1 link in the loop below, we need to
 945  * have at least one node following the current hint with (start == rxmit ==
 946  * end).  But that can't happen, (start == end) means that all the data in
 947  * that hole has been sacked, in which case, the hole would have been removed
 948  * from the scoreboard.
 949  */
 950 struct sackhole *
 951 tcp_sack_output(struct tcpcb *tp, int *sack_bytes_rexmt)
 952 {
 953         struct sackhole *hole = NULL;
 954
 955         INP_WLOCK_ASSERT(tptoinpcb(tp));
 956         *sack_bytes_rexmt = tp->sackhint.sack_bytes_rexmit;
 957         hole = tp->sackhint.nexthole;
 958         if (hole == NULL)
 959                 return (hole);
 960         if (SEQ_GEQ(hole->rxmit, hole->end)) {
 961                 for (;;) {
 962                         hole = TAILQ_NEXT(hole, scblink);
 963                         if (hole == NULL)
 964                                 return (hole);
 965                         if (SEQ_LT(hole->rxmit, hole->end)) {
 966                                 tp->sackhint.nexthole = hole;
 967                                 break;
 968                         }
 969                 }
 970         }
 971         KASSERT(SEQ_LT(hole->start, hole->end), ("%s: hole.start >= hole.end", __func__));
 972         if (!(V_tcp_do_newsack)) {
 973                 KASSERT(SEQ_LT(hole->start, tp->snd_fack), ("%s: hole.start >= snd.fack", __func__));
 974                 KASSERT(SEQ_LT(hole->end, tp->snd_fack), ("%s: hole.end >= snd.fack", __func__));
 975                 KASSERT(SEQ_LT(hole->rxmit, tp->snd_fack), ("%s: hole.rxmit >= snd.fack", __func__));
 976                 if (SEQ_GEQ(hole->start, hole->end) ||
 977                     SEQ_GEQ(hole->start, tp->snd_fack) ||
 978                     SEQ_GEQ(hole->end, tp->snd_fack) ||
 979                     SEQ_GEQ(hole->rxmit, tp->snd_fack)) {
 980                         log(LOG_CRIT,"tcp: invalid SACK hole (%u-%u,%u) vs fwd ack %u, ignoring.\n",
 981                                         hole->start, hole->end, hole->rxmit, tp->snd_fack);
 982                         return (NULL);
 983                 }
 984         }
 985         return (hole);
 986 }
 987
 988 /*
 989  * After a timeout, the SACK list may be rebuilt.  This SACK information
 990  * should be used to avoid retransmitting SACKed data.  This function
 991  * traverses the SACK list to see if snd_nxt should be moved forward.
 992  */
 993 void
 994 tcp_sack_adjust(struct tcpcb *tp)
 995 {
 996         struct sackhole *p, *cur = TAILQ_FIRST(&tp->snd_holes);
 997
 998         INP_WLOCK_ASSERT(tptoinpcb(tp));
 999         if (cur == NULL)
1000                 return; /* No holes */
1001         if (SEQ_GEQ(tp->snd_nxt, tp->snd_fack))
1002                 return; /* We're already beyond any SACKed blocks */
1003         /*-
1004          * Two cases for which we want to advance snd_nxt:
1005          * i) snd_nxt lies between end of one hole and beginning of another
1006          * ii) snd_nxt lies between end of last hole and snd_fack
1007          */
1008         while ((p = TAILQ_NEXT(cur, scblink)) != NULL) {
1009                 if (SEQ_LT(tp->snd_nxt, cur->end))
1010                         return;
1011                 if (SEQ_GEQ(tp->snd_nxt, p->start))
1012                         cur = p;
1013                 else {
1014                         tp->snd_nxt = p->start;
1015                         return;
1016                 }
1017         }
1018         if (SEQ_LT(tp->snd_nxt, cur->end))
1019                 return;
1020         tp->snd_nxt = tp->snd_fack;
1021 }
1022
1023 /*
1024  * Lost Retransmission Detection
1025  * Check is FACK is beyond the rexmit of the leftmost hole.
1026  * If yes, we restart sending from still existing holes,
1027  * and adjust cwnd via the congestion control module.
1028  */
1029 void
1030 tcp_sack_lost_retransmission(struct tcpcb *tp, struct tcphdr *th)
1031 {
1032         struct sackhole *temp;
1033
1034         if (IN_RECOVERY(tp->t_flags) &&
1035             SEQ_GT(tp->snd_fack, tp->snd_recover) &&
1036             ((temp = TAILQ_FIRST(&tp->snd_holes)) != NULL) &&
1037             SEQ_GEQ(temp->rxmit, temp->end) &&
1038             SEQ_GEQ(tp->snd_fack, temp->rxmit)) {
1039                 TCPSTAT_INC(tcps_sack_lostrexmt);
1040                 /*
1041                  * Start retransmissions from the first hole, and
1042                  * subsequently all other remaining holes, including
1043                  * those, which had been sent completely before.
1044                  */
1045                 tp->sackhint.nexthole = temp;
1046                 TAILQ_FOREACH(temp, &tp->snd_holes, scblink) {
1047                         if (SEQ_GEQ(tp->snd_fack, temp->rxmit) &&
1048                             SEQ_GEQ(temp->rxmit, temp->end))
1049                                 temp->rxmit = temp->start;
1050                 }
1051                 /*
1052                  * Remember the old ssthresh, to deduct the beta factor used
1053                  * by the CC module. Finally, set cwnd to ssthresh just
1054                  * prior to invoking another cwnd reduction by the CC
1055                  * module, to not shrink it excessively.
1056                  */
1057                 tp->snd_cwnd = tp->snd_ssthresh;
1058                 /*
1059                  * Formally exit recovery, and let the CC module adjust
1060                  * ssthresh as intended.
1061                  */
1062                 EXIT_RECOVERY(tp->t_flags);
1063                 cc_cong_signal(tp, th, CC_NDUPACK);
1064                 /*
1065                  * For PRR, adjust recover_fs as if this new reduction
1066                  * initialized this variable.
1067                  * cwnd will be adjusted by SACK or PRR processing
1068                  * subsequently, only set it to a safe value here.
1069                  */
1070                 tp->snd_cwnd = tcp_maxseg(tp);
1071                 tp->sackhint.recover_fs = (tp->snd_max - tp->snd_una) -
1072                                             tp->sackhint.recover_fs;
1073         }
1074 }