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 <netinet6/tcp6_var.h>
 121 #include <netinet/tcpip.h>
 122 #include <netinet/cc/cc.h>
 123 #ifdef TCPDEBUG
 124 #include <netinet/tcp_debug.h>
 125 #endif /* TCPDEBUG */
 126
 127 #include <machine/in_cksum.h>
 128
 129 VNET_DECLARE(struct uma_zone *, sack_hole_zone);
 130 #define V_sack_hole_zone                VNET(sack_hole_zone)
 131
 132 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, sack, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
 133     "TCP SACK");
 134
 135 VNET_DEFINE(int, tcp_do_sack) = 1;
 136 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, enable, CTLFLAG_VNET | CTLFLAG_RW,
 137     &VNET_NAME(tcp_do_sack), 0,
 138     "Enable/Disable TCP SACK support");
 139
 140 VNET_DEFINE(int, tcp_do_newsack) = 1;
 141 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, revised, CTLFLAG_VNET | CTLFLAG_RW,
 142     &VNET_NAME(tcp_do_newsack), 0,
 143     "Use revised SACK loss recovery per RFC 6675");
 144
 145 VNET_DEFINE(int, tcp_sack_maxholes) = 128;
 146 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, maxholes, CTLFLAG_VNET | CTLFLAG_RW,
 147     &VNET_NAME(tcp_sack_maxholes), 0,
 148     "Maximum number of TCP SACK holes allowed per connection");
 149
 150 VNET_DEFINE(int, tcp_sack_globalmaxholes) = 65536;
 151 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, globalmaxholes, CTLFLAG_VNET | CTLFLAG_RW,
 152     &VNET_NAME(tcp_sack_globalmaxholes), 0,
 153     "Global maximum number of TCP SACK holes");
 154
 155 VNET_DEFINE(int, tcp_sack_globalholes) = 0;
 156 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, globalholes, CTLFLAG_VNET | CTLFLAG_RD,
 157     &VNET_NAME(tcp_sack_globalholes), 0,
 158     "Global number of TCP SACK holes currently allocated");
 159
 160 int
 161 tcp_dsack_block_exists(struct tcpcb *tp)
 162 {
 163         /* Return true if a DSACK block exists */
 164         if (tp->rcv_numsacks == 0)
 165                 return (0);
 166         if (SEQ_LEQ(tp->sackblks[0].end, tp->rcv_nxt))
 167                 return(1);
 168         return (0);
 169 }
 170
 171 /*
 172  * This function will find overlaps with the currently stored sackblocks
 173  * and add any overlap as a dsack block upfront
 174  */
 175 void
 176 tcp_update_dsack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
 177 {
 178         struct sackblk head_blk,mid_blk,saved_blks[MAX_SACK_BLKS];
 179         int i, j, n, identical;
 180         tcp_seq start, end;
 181
 182         INP_WLOCK_ASSERT(tp->t_inpcb);
 183
 184         KASSERT(SEQ_LT(rcv_start, rcv_end), ("rcv_start < rcv_end"));
 185
 186         if (SEQ_LT(rcv_end, tp->rcv_nxt) ||
 187             ((rcv_end == tp->rcv_nxt) &&
 188              (tp->rcv_numsacks > 0 ) &&
 189              (tp->sackblks[0].end == tp->rcv_nxt))) {
 190                 saved_blks[0].start = rcv_start;
 191                 saved_blks[0].end = rcv_end;
 192         } else {
 193                 saved_blks[0].start = saved_blks[0].end = 0;
 194         }
 195
 196         head_blk.start = head_blk.end = 0;
 197         mid_blk.start = rcv_start;
 198         mid_blk.end = rcv_end;
 199         identical = 0;
 200
 201         for (i = 0; i < tp->rcv_numsacks; i++) {
 202                 start = tp->sackblks[i].start;
 203                 end = tp->sackblks[i].end;
 204                 if (SEQ_LT(rcv_end, start)) {
 205                         /* pkt left to sack blk */
 206                         continue;
 207                 }
 208                 if (SEQ_GT(rcv_start, end)) {
 209                         /* pkt right to sack blk */
 210                         continue;
 211                 }
 212                 if (SEQ_GT(tp->rcv_nxt, end)) {
 213                         if ((SEQ_MAX(rcv_start, start) != SEQ_MIN(rcv_end, end)) &&
 214                             (SEQ_GT(head_blk.start, SEQ_MAX(rcv_start, start)) ||
 215                             (head_blk.start == head_blk.end))) {
 216                                 head_blk.start = SEQ_MAX(rcv_start, start);
 217                                 head_blk.end = SEQ_MIN(rcv_end, end);
 218                         }
 219                         continue;
 220                 }
 221                 if (((head_blk.start == head_blk.end) ||
 222                      SEQ_LT(start, head_blk.start)) &&
 223                      (SEQ_GT(end, rcv_start) &&
 224                       SEQ_LEQ(start, rcv_end))) {
 225                         head_blk.start = start;
 226                         head_blk.end = end;
 227                 }
 228                 mid_blk.start = SEQ_MIN(mid_blk.start, start);
 229                 mid_blk.end = SEQ_MAX(mid_blk.end, end);
 230                 if ((mid_blk.start == start) &&
 231                     (mid_blk.end == end))
 232                         identical = 1;
 233         }
 234         if (SEQ_LT(head_blk.start, head_blk.end)) {
 235                 /* store overlapping range */
 236                 saved_blks[0].start = SEQ_MAX(rcv_start, head_blk.start);
 237                 saved_blks[0].end   = SEQ_MIN(rcv_end, head_blk.end);
 238         }
 239         n = 1;
 240         /*
 241          * Second, if not ACKed, store the SACK block that
 242          * overlaps with the DSACK block unless it is identical
 243          */
 244         if ((SEQ_LT(tp->rcv_nxt, mid_blk.end) &&
 245             !((mid_blk.start == saved_blks[0].start) &&
 246             (mid_blk.end == saved_blks[0].end))) ||
 247             identical == 1) {
 248                 saved_blks[n].start = mid_blk.start;
 249                 saved_blks[n++].end = mid_blk.end;
 250         }
 251         for (j = 0; (j < tp->rcv_numsacks) && (n < MAX_SACK_BLKS); j++) {
 252                 if (((SEQ_LT(tp->sackblks[j].end, mid_blk.start) ||
 253                       SEQ_GT(tp->sackblks[j].start, mid_blk.end)) &&
 254                     (SEQ_GT(tp->sackblks[j].start, tp->rcv_nxt))))
 255                 saved_blks[n++] = tp->sackblks[j];
 256         }
 257         j = 0;
 258         for (i = 0; i < n; i++) {
 259                 /* we can end up with a stale initial entry */
 260                 if (SEQ_LT(saved_blks[i].start, saved_blks[i].end)) {
 261                         tp->sackblks[j++] = saved_blks[i];
 262                 }
 263         }
 264         tp->rcv_numsacks = j;
 265 }
 266
 267 /*
 268  * This function is called upon receipt of new valid data (while not in
 269  * header prediction mode), and it updates the ordered list of sacks.
 270  */
 271 void
 272 tcp_update_sack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
 273 {
 274         /*
 275          * First reported block MUST be the most recent one.  Subsequent
 276          * blocks SHOULD be in the order in which they arrived at the
 277          * receiver.  These two conditions make the implementation fully
 278          * compliant with RFC 2018.
 279          */
 280         struct sackblk head_blk, saved_blks[MAX_SACK_BLKS];
 281         int num_head, num_saved, i;
 282
 283         INP_WLOCK_ASSERT(tp->t_inpcb);
 284
 285         /* Check arguments. */
 286         KASSERT(SEQ_LEQ(rcv_start, rcv_end), ("rcv_start <= rcv_end"));
 287
 288         if ((rcv_start == rcv_end) &&
 289             (tp->rcv_numsacks >= 1) &&
 290             (rcv_end == tp->sackblks[0].end)) {
 291                 /* retaining DSACK block below rcv_nxt (todrop) */
 292                 head_blk = tp->sackblks[0];
 293         } else {
 294                 /* SACK block for the received segment. */
 295                 head_blk.start = rcv_start;
 296                 head_blk.end = rcv_end;
 297         }
 298
 299         /*
 300          * Merge updated SACK blocks into head_blk, and save unchanged SACK
 301          * blocks into saved_blks[].  num_saved will have the number of the
 302          * saved SACK blocks.
 303          */
 304         num_saved = 0;
 305         for (i = 0; i < tp->rcv_numsacks; i++) {
 306                 tcp_seq start = tp->sackblks[i].start;
 307                 tcp_seq end = tp->sackblks[i].end;
 308                 if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
 309                         /*
 310                          * Discard this SACK block.
 311                          */
 312                 } else if (SEQ_LEQ(head_blk.start, end) &&
 313                            SEQ_GEQ(head_blk.end, start)) {
 314                         /*
 315                          * Merge this SACK block into head_blk.  This SACK
 316                          * block itself will be discarded.
 317                          */
 318                         /*
 319                          * |-|
 320                          *   |---|  merge
 321                          *
 322                          *     |-|
 323                          * |---|    merge
 324                          *
 325                          * |-----|
 326                          *   |-|    DSACK smaller
 327                          *
 328                          *   |-|
 329                          * |-----|  DSACK smaller
 330                          */
 331                         if (head_blk.start == end)
 332                                 head_blk.start = start;
 333                         else if (head_blk.end == start)
 334                                 head_blk.end = end;
 335                         else {
 336                                 if (SEQ_LT(head_blk.start, start)) {
 337                                         tcp_seq temp = start;
 338                                         start = head_blk.start;
 339                                         head_blk.start = temp;
 340                                 }
 341                                 if (SEQ_GT(head_blk.end, end)) {
 342                                         tcp_seq temp = end;
 343                                         end = head_blk.end;
 344                                         head_blk.end = temp;
 345                                 }
 346                                 if ((head_blk.start != start) ||
 347                                     (head_blk.end != end)) {
 348                                         if ((num_saved >= 1) &&
 349                                            SEQ_GEQ(saved_blks[num_saved-1].start, start) &&
 350                                            SEQ_LEQ(saved_blks[num_saved-1].end, end))
 351                                                 num_saved--;
 352                                         saved_blks[num_saved].start = start;
 353                                         saved_blks[num_saved].end = end;
 354                                         num_saved++;
 355                                 }
 356                         }
 357                 } else {
 358                         /*
 359                          * This block supercedes the prior block
 360                          */
 361                         if ((num_saved >= 1) &&
 362                            SEQ_GEQ(saved_blks[num_saved-1].start, start) &&
 363                            SEQ_LEQ(saved_blks[num_saved-1].end, end))
 364                                 num_saved--;
 365                         /*
 366                          * Save this SACK block.
 367                          */
 368                         saved_blks[num_saved].start = start;
 369                         saved_blks[num_saved].end = end;
 370                         num_saved++;
 371                 }
 372         }
 373
 374         /*
 375          * Update SACK list in tp->sackblks[].
 376          */
 377         num_head = 0;
 378         if (SEQ_LT(rcv_start, rcv_end)) {
 379                 /*
 380                  * The received data segment is an out-of-order segment.  Put
 381                  * head_blk at the top of SACK list.
 382                  */
 383                 tp->sackblks[0] = head_blk;
 384                 num_head = 1;
 385                 /*
 386                  * If the number of saved SACK blocks exceeds its limit,
 387                  * discard the last SACK block.
 388                  */
 389                 if (num_saved >= MAX_SACK_BLKS)
 390                         num_saved--;
 391         }
 392         if ((rcv_start == rcv_end) &&
 393             (rcv_start == tp->sackblks[0].end)) {
 394                 num_head = 1;
 395         }
 396         if (num_saved > 0) {
 397                 /*
 398                  * Copy the saved SACK blocks back.
 399                  */
 400                 bcopy(saved_blks, &tp->sackblks[num_head],
 401                       sizeof(struct sackblk) * num_saved);
 402         }
 403
 404         /* Save the number of SACK blocks. */
 405         tp->rcv_numsacks = num_head + num_saved;
 406 }
 407
 408 void
 409 tcp_clean_dsack_blocks(struct tcpcb *tp)
 410 {
 411         struct sackblk saved_blks[MAX_SACK_BLKS];
 412         int num_saved, i;
 413
 414         INP_WLOCK_ASSERT(tp->t_inpcb);
 415         /*
 416          * Clean up any DSACK blocks that
 417          * are in our queue of sack blocks.
 418          *
 419          */
 420         num_saved = 0;
 421         for (i = 0; i < tp->rcv_numsacks; i++) {
 422                 tcp_seq start = tp->sackblks[i].start;
 423                 tcp_seq end = tp->sackblks[i].end;
 424                 if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
 425                         /*
 426                          * Discard this D-SACK block.
 427                          */
 428                         continue;
 429                 }
 430                 /*
 431                  * Save this SACK block.
 432                  */
 433                 saved_blks[num_saved].start = start;
 434                 saved_blks[num_saved].end = end;
 435                 num_saved++;
 436         }
 437         if (num_saved > 0) {
 438                 /*
 439                  * Copy the saved SACK blocks back.
 440                  */
 441                 bcopy(saved_blks, &tp->sackblks[0],
 442                       sizeof(struct sackblk) * num_saved);
 443         }
 444         tp->rcv_numsacks = num_saved;
 445 }
 446
 447 /*
 448  * Delete all receiver-side SACK information.
 449  */
 450 void
 451 tcp_clean_sackreport(struct tcpcb *tp)
 452 {
 453         int i;
 454
 455         INP_WLOCK_ASSERT(tp->t_inpcb);
 456         tp->rcv_numsacks = 0;
 457         for (i = 0; i < MAX_SACK_BLKS; i++)
 458                 tp->sackblks[i].start = tp->sackblks[i].end=0;
 459 }
 460
 461 /*
 462  * Allocate struct sackhole.
 463  */
 464 static struct sackhole *
 465 tcp_sackhole_alloc(struct tcpcb *tp, tcp_seq start, tcp_seq end)
 466 {
 467         struct sackhole *hole;
 468
 469         if (tp->snd_numholes >= V_tcp_sack_maxholes ||
 470             V_tcp_sack_globalholes >= V_tcp_sack_globalmaxholes) {
 471                 TCPSTAT_INC(tcps_sack_sboverflow);
 472                 return NULL;
 473         }
 474
 475         hole = (struct sackhole *)uma_zalloc(V_sack_hole_zone, M_NOWAIT);
 476         if (hole == NULL)
 477                 return NULL;
 478
 479         hole->start = start;
 480         hole->end = end;
 481         hole->rxmit = start;
 482
 483         tp->snd_numholes++;
 484         atomic_add_int(&V_tcp_sack_globalholes, 1);
 485
 486         return hole;
 487 }
 488
 489 /*
 490  * Free struct sackhole.
 491  */
 492 static void
 493 tcp_sackhole_free(struct tcpcb *tp, struct sackhole *hole)
 494 {
 495
 496         uma_zfree(V_sack_hole_zone, hole);
 497
 498         tp->snd_numholes--;
 499         atomic_subtract_int(&V_tcp_sack_globalholes, 1);
 500
 501         KASSERT(tp->snd_numholes >= 0, ("tp->snd_numholes >= 0"));
 502         KASSERT(V_tcp_sack_globalholes >= 0, ("tcp_sack_globalholes >= 0"));
 503 }
 504
 505 /*
 506  * Insert new SACK hole into scoreboard.
 507  */
 508 static struct sackhole *
 509 tcp_sackhole_insert(struct tcpcb *tp, tcp_seq start, tcp_seq end,
 510     struct sackhole *after)
 511 {
 512         struct sackhole *hole;
 513
 514         /* Allocate a new SACK hole. */
 515         hole = tcp_sackhole_alloc(tp, start, end);
 516         if (hole == NULL)
 517                 return NULL;
 518
 519         /* Insert the new SACK hole into scoreboard. */
 520         if (after != NULL)
 521                 TAILQ_INSERT_AFTER(&tp->snd_holes, after, hole, scblink);
 522         else
 523                 TAILQ_INSERT_TAIL(&tp->snd_holes, hole, scblink);
 524
 525         /* Update SACK hint. */
 526         if (tp->sackhint.nexthole == NULL)
 527                 tp->sackhint.nexthole = hole;
 528
 529         return hole;
 530 }
 531
 532 /*
 533  * Remove SACK hole from scoreboard.
 534  */
 535 static void
 536 tcp_sackhole_remove(struct tcpcb *tp, struct sackhole *hole)
 537 {
 538
 539         /* Update SACK hint. */
 540         if (tp->sackhint.nexthole == hole)
 541                 tp->sackhint.nexthole = TAILQ_NEXT(hole, scblink);
 542
 543         /* Remove this SACK hole. */
 544         TAILQ_REMOVE(&tp->snd_holes, hole, scblink);
 545
 546         /* Free this SACK hole. */
 547         tcp_sackhole_free(tp, hole);
 548 }
 549
 550 /*
 551  * Process cumulative ACK and the TCP SACK option to update the scoreboard.
 552  * tp->snd_holes is an ordered list of holes (oldest to newest, in terms of
 553  * the sequence space).
 554  * Returns 1 if incoming ACK has previously unknown SACK information,
 555  * 0 otherwise.
 556  */
 557 int
 558 tcp_sack_doack(struct tcpcb *tp, struct tcpopt *to, tcp_seq th_ack)
 559 {
 560         struct sackhole *cur, *temp;
 561         struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1], *sblkp;
 562         int i, j, num_sack_blks, sack_changed;
 563         int delivered_data, left_edge_delta;
 564
 565         INP_WLOCK_ASSERT(tp->t_inpcb);
 566
 567         num_sack_blks = 0;
 568         sack_changed = 0;
 569         delivered_data = 0;
 570         left_edge_delta = 0;
 571         /*
 572          * If SND.UNA will be advanced by SEG.ACK, and if SACK holes exist,
 573          * treat [SND.UNA, SEG.ACK) as if it is a SACK block.
 574          * Account changes to SND.UNA always in delivered data.
 575          */
 576         if (SEQ_LT(tp->snd_una, th_ack) && !TAILQ_EMPTY(&tp->snd_holes)) {
 577                 left_edge_delta = th_ack - tp->snd_una;
 578                 sack_blocks[num_sack_blks].start = tp->snd_una;
 579                 sack_blocks[num_sack_blks++].end = th_ack;
 580                 /*
 581                  * Pulling snd_fack forward if we got here
 582                  * due to DSACK blocks
 583                  */
 584                 if (SEQ_LT(tp->snd_fack, th_ack)) {
 585                         delivered_data += th_ack - tp->snd_una;
 586                         tp->snd_fack = th_ack;
 587                         sack_changed = 1;
 588                 }
 589         }
 590         /*
 591          * Append received valid SACK blocks to sack_blocks[], but only if we
 592          * received new blocks from the other side.
 593          */
 594         if (to->to_flags & TOF_SACK) {
 595                 for (i = 0; i < to->to_nsacks; i++) {
 596                         bcopy((to->to_sacks + i * TCPOLEN_SACK),
 597                             &sack, sizeof(sack));
 598                         sack.start = ntohl(sack.start);
 599                         sack.end = ntohl(sack.end);
 600                         if (SEQ_GT(sack.end, sack.start) &&
 601                             SEQ_GT(sack.start, tp->snd_una) &&
 602                             SEQ_GT(sack.start, th_ack) &&
 603                             SEQ_LT(sack.start, tp->snd_max) &&
 604                             SEQ_GT(sack.end, tp->snd_una) &&
 605                             SEQ_LEQ(sack.end, tp->snd_max)) {
 606                                 sack_blocks[num_sack_blks++] = sack;
 607                         }
 608                 }
 609         }
 610         /*
 611          * Return if SND.UNA is not advanced and no valid SACK block is
 612          * received.
 613          */
 614         if (num_sack_blks == 0)
 615                 return (sack_changed);
 616
 617         /*
 618          * Sort the SACK blocks so we can update the scoreboard with just one
 619          * pass. The overhead of sorting up to 4+1 elements is less than
 620          * making up to 4+1 passes over the scoreboard.
 621          */
 622         for (i = 0; i < num_sack_blks; i++) {
 623                 for (j = i + 1; j < num_sack_blks; j++) {
 624                         if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
 625                                 sack = sack_blocks[i];
 626                                 sack_blocks[i] = sack_blocks[j];
 627                                 sack_blocks[j] = sack;
 628                         }
 629                 }
 630         }
 631         if (TAILQ_EMPTY(&tp->snd_holes)) {
 632                 /*
 633                  * Empty scoreboard. Need to initialize snd_fack (it may be
 634                  * uninitialized or have a bogus value). Scoreboard holes
 635                  * (from the sack blocks received) are created later below
 636                  * (in the logic that adds holes to the tail of the
 637                  * scoreboard).
 638                  */
 639                 tp->snd_fack = SEQ_MAX(tp->snd_una, th_ack);
 640                 tp->sackhint.sacked_bytes = 0;  /* reset */
 641         }
 642         /*
 643          * In the while-loop below, incoming SACK blocks (sack_blocks[]) and
 644          * SACK holes (snd_holes) are traversed from their tails with just
 645          * one pass in order to reduce the number of compares especially when
 646          * the bandwidth-delay product is large.
 647          *
 648          * Note: Typically, in the first RTT of SACK recovery, the highest
 649          * three or four SACK blocks with the same ack number are received.
 650          * In the second RTT, if retransmitted data segments are not lost,
 651          * the highest three or four SACK blocks with ack number advancing
 652          * are received.
 653          */
 654         sblkp = &sack_blocks[num_sack_blks - 1];        /* Last SACK block */
 655         tp->sackhint.last_sack_ack = sblkp->end;
 656         if (SEQ_LT(tp->snd_fack, sblkp->start)) {
 657                 /*
 658                  * The highest SACK block is beyond fack.  First,
 659                  * check if there was a successful Rescue Retransmission,
 660                  * and move this hole left. With normal holes, snd_fack
 661                  * is always to the right of the end.
 662                  */
 663                 if (((temp = TAILQ_LAST(&tp->snd_holes, sackhole_head)) != NULL) &&
 664                     SEQ_LEQ(tp->snd_fack,temp->end)) {
 665                         temp->start = SEQ_MAX(tp->snd_fack, SEQ_MAX(tp->snd_una, th_ack));
 666                         temp->end = sblkp->start;
 667                         temp->rxmit = temp->start;
 668                         delivered_data += sblkp->end - sblkp->start;
 669                         tp->snd_fack = sblkp->end;
 670                         sblkp--;
 671                         sack_changed = 1;
 672                 } else {
 673                         /*
 674                          * Append a new SACK hole at the tail.  If the
 675                          * second or later highest SACK blocks are also
 676                          * beyond the current fack, they will be inserted
 677                          * by way of hole splitting in the while-loop below.
 678                          */
 679                         temp = tcp_sackhole_insert(tp, tp->snd_fack,sblkp->start,NULL);
 680                         if (temp != NULL) {
 681                                 delivered_data += sblkp->end - sblkp->start;
 682                                 tp->snd_fack = sblkp->end;
 683                                 /* Go to the previous sack block. */
 684                                 sblkp--;
 685                                 sack_changed = 1;
 686                         } else {
 687                                 /*
 688                                  * We failed to add a new hole based on the current
 689                                  * sack block.  Skip over all the sack blocks that
 690                                  * fall completely to the right of snd_fack and
 691                                  * proceed to trim the scoreboard based on the
 692                                  * remaining sack blocks.  This also trims the
 693                                  * scoreboard for th_ack (which is sack_blocks[0]).
 694                                  */
 695                                 while (sblkp >= sack_blocks &&
 696                                        SEQ_LT(tp->snd_fack, sblkp->start))
 697                                         sblkp--;
 698                                 if (sblkp >= sack_blocks &&
 699                                     SEQ_LT(tp->snd_fack, sblkp->end)) {
 700                                         delivered_data += sblkp->end - tp->snd_fack;
 701                                         tp->snd_fack = sblkp->end;
 702                                         sack_changed = 1;
 703                                 }
 704                         }
 705                 }
 706         } else if (SEQ_LT(tp->snd_fack, sblkp->end)) {
 707                 /* fack is advanced. */
 708                 delivered_data += sblkp->end - tp->snd_fack;
 709                 tp->snd_fack = sblkp->end;
 710                 sack_changed = 1;
 711         }
 712         cur = TAILQ_LAST(&tp->snd_holes, sackhole_head); /* Last SACK hole. */
 713         /*
 714          * Since the incoming sack blocks are sorted, we can process them
 715          * making one sweep of the scoreboard.
 716          */
 717         while (sblkp >= sack_blocks  && cur != NULL) {
 718                 if (SEQ_GEQ(sblkp->start, cur->end)) {
 719                         /*
 720                          * SACKs data beyond the current hole.  Go to the
 721                          * previous sack block.
 722                          */
 723                         sblkp--;
 724                         continue;
 725                 }
 726                 if (SEQ_LEQ(sblkp->end, cur->start)) {
 727                         /*
 728                          * SACKs data before the current hole.  Go to the
 729                          * previous hole.
 730                          */
 731                         cur = TAILQ_PREV(cur, sackhole_head, scblink);
 732                         continue;
 733                 }
 734                 tp->sackhint.sack_bytes_rexmit -=
 735                     (SEQ_MIN(cur->rxmit, cur->end) - cur->start);
 736                 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
 737                     ("sackhint bytes rtx >= 0"));
 738                 sack_changed = 1;
 739                 if (SEQ_LEQ(sblkp->start, cur->start)) {
 740                         /* Data acks at least the beginning of hole. */
 741                         if (SEQ_GEQ(sblkp->end, cur->end)) {
 742                                 /* Acks entire hole, so delete hole. */
 743                                 delivered_data += (cur->end - cur->start);
 744                                 temp = cur;
 745                                 cur = TAILQ_PREV(cur, sackhole_head, scblink);
 746                                 tcp_sackhole_remove(tp, temp);
 747                                 /*
 748                                  * The sack block may ack all or part of the
 749                                  * next hole too, so continue onto the next
 750                                  * hole.
 751                                  */
 752                                 continue;
 753                         } else {
 754                                 /* Move start of hole forward. */
 755                                 delivered_data += (sblkp->end - cur->start);
 756                                 cur->start = sblkp->end;
 757                                 cur->rxmit = SEQ_MAX(cur->rxmit, cur->start);
 758                         }
 759                 } else {
 760                         /* Data acks at least the end of hole. */
 761                         if (SEQ_GEQ(sblkp->end, cur->end)) {
 762                                 /* Move end of hole backward. */
 763                                 delivered_data += (cur->end - sblkp->start);
 764                                 cur->end = sblkp->start;
 765                                 cur->rxmit = SEQ_MIN(cur->rxmit, cur->end);
 766                                 if ((tp->t_flags & TF_LRD) && SEQ_GEQ(cur->rxmit, cur->end))
 767                                         cur->rxmit = tp->snd_recover;
 768                         } else {
 769                                 /*
 770                                  * ACKs some data in middle of a hole; need
 771                                  * to split current hole
 772                                  */
 773                                 temp = tcp_sackhole_insert(tp, sblkp->end,
 774                                     cur->end, cur);
 775                                 if (temp != NULL) {
 776                                         if (SEQ_GT(cur->rxmit, temp->rxmit)) {
 777                                                 temp->rxmit = cur->rxmit;
 778                                                 tp->sackhint.sack_bytes_rexmit +=
 779                                                     (SEQ_MIN(temp->rxmit,
 780                                                     temp->end) - temp->start);
 781                                         }
 782                                         cur->end = sblkp->start;
 783                                         cur->rxmit = SEQ_MIN(cur->rxmit,
 784                                             cur->end);
 785                                         if ((tp->t_flags & TF_LRD) && SEQ_GEQ(cur->rxmit, cur->end))
 786                                                 cur->rxmit = tp->snd_recover;
 787                                         delivered_data += (sblkp->end - sblkp->start);
 788                                 }
 789                         }
 790                 }
 791                 tp->sackhint.sack_bytes_rexmit +=
 792                     (SEQ_MIN(cur->rxmit, cur->end) - cur->start);
 793                 /*
 794                  * Testing sblkp->start against cur->start tells us whether
 795                  * we're done with the sack block or the sack hole.
 796                  * Accordingly, we advance one or the other.
 797                  */
 798                 if (SEQ_LEQ(sblkp->start, cur->start))
 799                         cur = TAILQ_PREV(cur, sackhole_head, scblink);
 800                 else
 801                         sblkp--;
 802         }
 803         if (!(to->to_flags & TOF_SACK))
 804                 /*
 805                  * If this ACK did not contain any
 806                  * SACK blocks, any only moved the
 807                  * left edge right, it is a pure
 808                  * cumulative ACK. Do not count
 809                  * DupAck for this. Also required
 810                  * for RFC6675 rescue retransmission.
 811                  */
 812                 sack_changed = 0;
 813         tp->sackhint.delivered_data = delivered_data;
 814         tp->sackhint.sacked_bytes += delivered_data - left_edge_delta;
 815         KASSERT((delivered_data >= 0), ("delivered_data < 0"));
 816         KASSERT((tp->sackhint.sacked_bytes >= 0), ("sacked_bytes < 0"));
 817         return (sack_changed);
 818 }
 819
 820 /*
 821  * Free all SACK holes to clear the scoreboard.
 822  */
 823 void
 824 tcp_free_sackholes(struct tcpcb *tp)
 825 {
 826         struct sackhole *q;
 827
 828         INP_WLOCK_ASSERT(tp->t_inpcb);
 829         while ((q = TAILQ_FIRST(&tp->snd_holes)) != NULL)
 830                 tcp_sackhole_remove(tp, q);
 831         tp->sackhint.sack_bytes_rexmit = 0;
 832
 833         KASSERT(tp->snd_numholes == 0, ("tp->snd_numholes == 0"));
 834         KASSERT(tp->sackhint.nexthole == NULL,
 835                 ("tp->sackhint.nexthole == NULL"));
 836 }
 837
 838 /*
 839  * Partial ack handling within a sack recovery episode.  Keeping this very
 840  * simple for now.  When a partial ack is received, force snd_cwnd to a value
 841  * that will allow the sender to transmit no more than 2 segments.  If
 842  * necessary, a better scheme can be adopted at a later point, but for now,
 843  * the goal is to prevent the sender from bursting a large amount of data in
 844  * the midst of sack recovery.
 845  */
 846 void
 847 tcp_sack_partialack(struct tcpcb *tp, struct tcphdr *th)
 848 {
 849         int num_segs = 1;
 850         u_int maxseg = tcp_maxseg(tp);
 851
 852         INP_WLOCK_ASSERT(tp->t_inpcb);
 853         tcp_timer_activate(tp, TT_REXMT, 0);
 854         tp->t_rtttime = 0;
 855         /* Send one or 2 segments based on how much new data was acked. */
 856         if ((BYTES_THIS_ACK(tp, th) / maxseg) >= 2)
 857                 num_segs = 2;
 858         tp->snd_cwnd = (tp->sackhint.sack_bytes_rexmit +
 859             (tp->snd_nxt - tp->snd_recover) + num_segs * maxseg);
 860         if (tp->snd_cwnd > tp->snd_ssthresh)
 861                 tp->snd_cwnd = tp->snd_ssthresh;
 862         tp->t_flags |= TF_ACKNOW;
 863         /*
 864          * RFC6675 rescue retransmission
 865          * Add a hole between th_ack (snd_una is not yet set) and snd_max,
 866          * if this was a pure cumulative ACK and no data was send beyond
 867          * recovery point. Since the data in the socket has not been freed
 868          * at this point, we check if the scoreboard is empty, and the ACK
 869          * delivered some new data, indicating a full ACK. Also, if the
 870          * recovery point is still at snd_max, we are probably application
 871          * limited. However, this inference might not always be true. The
 872          * rescue retransmission may rarely be slightly premature
 873          * compared to RFC6675.
 874          * The corresponding ACK+SACK will cause any further outstanding
 875          * segments to be retransmitted. This addresses a corner case, when
 876          * the trailing packets of a window are lost and no further data
 877          * is available for sending.
 878          */
 879         if ((V_tcp_do_newsack) &&
 880             SEQ_LT(th->th_ack, tp->snd_recover) &&
 881             (tp->snd_recover == tp->snd_max) &&
 882             TAILQ_EMPTY(&tp->snd_holes) &&
 883             (tp->sackhint.delivered_data > 0)) {
 884                 /*
 885                  * Exclude FIN sequence space in
 886                  * the hole for the rescue retransmission,
 887                  * and also don't create a hole, if only
 888                  * the ACK for a FIN is outstanding.
 889                  */
 890                 tcp_seq highdata = tp->snd_max;
 891                 if (tp->t_flags & TF_SENTFIN)
 892                         highdata--;
 893                 if (th->th_ack != highdata) {
 894                         tp->snd_fack = th->th_ack;
 895                         (void)tcp_sackhole_insert(tp, SEQ_MAX(th->th_ack,
 896                             highdata - maxseg), highdata, NULL);
 897                 }
 898         }
 899         (void) tp->t_fb->tfb_tcp_output(tp);
 900 }
 901
 902 #if 0
 903 /*
 904  * Debug version of tcp_sack_output() that walks the scoreboard.  Used for
 905  * now to sanity check the hint.
 906  */
 907 static struct sackhole *
 908 tcp_sack_output_debug(struct tcpcb *tp, int *sack_bytes_rexmt)
 909 {
 910         struct sackhole *p;
 911
 912         INP_WLOCK_ASSERT(tp->t_inpcb);
 913         *sack_bytes_rexmt = 0;
 914         TAILQ_FOREACH(p, &tp->snd_holes, scblink) {
 915                 if (SEQ_LT(p->rxmit, p->end)) {
 916                         if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */
 917                                 continue;
 918                         }
 919                         *sack_bytes_rexmt += (p->rxmit - p->start);
 920                         break;
 921                 }
 922                 *sack_bytes_rexmt += (SEQ_MIN(p->rxmit, p->end) - p->start);
 923         }
 924         return (p);
 925 }
 926 #endif
 927
 928 /*
 929  * Returns the next hole to retransmit and the number of retransmitted bytes
 930  * from the scoreboard.  We store both the next hole and the number of
 931  * retransmitted bytes as hints (and recompute these on the fly upon SACK/ACK
 932  * reception).  This avoids scoreboard traversals completely.
 933  *
 934  * The loop here will traverse *at most* one link.  Here's the argument.  For
 935  * the loop to traverse more than 1 link before finding the next hole to
 936  * retransmit, we would need to have at least 1 node following the current
 937  * hint with (rxmit == end).  But, for all holes following the current hint,
 938  * (start == rxmit), since we have not yet retransmitted from them.
 939  * Therefore, in order to traverse more 1 link in the loop below, we need to
 940  * have at least one node following the current hint with (start == rxmit ==
 941  * end).  But that can't happen, (start == end) means that all the data in
 942  * that hole has been sacked, in which case, the hole would have been removed
 943  * from the scoreboard.
 944  */
 945 struct sackhole *
 946 tcp_sack_output(struct tcpcb *tp, int *sack_bytes_rexmt)
 947 {
 948         struct sackhole *hole = NULL;
 949
 950         INP_WLOCK_ASSERT(tp->t_inpcb);
 951         *sack_bytes_rexmt = tp->sackhint.sack_bytes_rexmit;
 952         hole = tp->sackhint.nexthole;
 953         if (hole == NULL || SEQ_LT(hole->rxmit, hole->end))
 954                 goto out;
 955         while ((hole = TAILQ_NEXT(hole, scblink)) != NULL) {
 956                 if (SEQ_LT(hole->rxmit, hole->end)) {
 957                         tp->sackhint.nexthole = hole;
 958                         break;
 959                 }
 960         }
 961 out:
 962         return (hole);
 963 }
 964
 965 /*
 966  * After a timeout, the SACK list may be rebuilt.  This SACK information
 967  * should be used to avoid retransmitting SACKed data.  This function
 968  * traverses the SACK list to see if snd_nxt should be moved forward.
 969  */
 970 void
 971 tcp_sack_adjust(struct tcpcb *tp)
 972 {
 973         struct sackhole *p, *cur = TAILQ_FIRST(&tp->snd_holes);
 974
 975         INP_WLOCK_ASSERT(tp->t_inpcb);
 976         if (cur == NULL)
 977                 return; /* No holes */
 978         if (SEQ_GEQ(tp->snd_nxt, tp->snd_fack))
 979                 return; /* We're already beyond any SACKed blocks */
 980         /*-
 981          * Two cases for which we want to advance snd_nxt:
 982          * i) snd_nxt lies between end of one hole and beginning of another
 983          * ii) snd_nxt lies between end of last hole and snd_fack
 984          */
 985         while ((p = TAILQ_NEXT(cur, scblink)) != NULL) {
 986                 if (SEQ_LT(tp->snd_nxt, cur->end))
 987                         return;
 988                 if (SEQ_GEQ(tp->snd_nxt, p->start))
 989                         cur = p;
 990                 else {
 991                         tp->snd_nxt = p->start;
 992                         return;
 993                 }
 994         }
 995         if (SEQ_LT(tp->snd_nxt, cur->end))
 996                 return;
 997         tp->snd_nxt = tp->snd_fack;
 998 }
 999
1000 /*
1001  * Lost Retransmission Detection
1002  * Check is FACK is beyond the rexmit of the leftmost hole.
1003  * If yes, we restart sending from still existing holes,
1004  * and adjust cwnd via the congestion control module.
1005  */
1006 void
1007 tcp_sack_lost_retransmission(struct tcpcb *tp, struct tcphdr *th)
1008 {
1009         struct sackhole *temp;
1010         uint32_t prev_cwnd;
1011         if (IN_RECOVERY(tp->t_flags) &&
1012             SEQ_GT(tp->snd_fack, tp->snd_recover) &&
1013             ((temp = TAILQ_FIRST(&tp->snd_holes)) != NULL) &&
1014             SEQ_GEQ(temp->rxmit, temp->end) &&
1015             SEQ_GEQ(tp->snd_fack, temp->rxmit)) {
1016                 TCPSTAT_INC(tcps_sack_lostrexmt);
1017                 /*
1018                  * Start retransmissions from the first hole, and
1019                  * subsequently all other remaining holes, including
1020                  * those, which had been sent completely before.
1021                  */
1022                 tp->sackhint.nexthole = temp;
1023                 TAILQ_FOREACH(temp, &tp->snd_holes, scblink) {
1024                         if (SEQ_GEQ(tp->snd_fack, temp->rxmit) &&
1025                             SEQ_GEQ(temp->rxmit, temp->end))
1026                                 temp->rxmit = temp->start;
1027                 }
1028                 /*
1029                  * Remember the old ssthresh, to deduct the beta factor used
1030                  * by the CC module. Finally, set cwnd to ssthresh just
1031                  * prior to invoking another cwnd reduction by the CC
1032                  * module, to not shrink it excessively.
1033                  */
1034                 prev_cwnd = tp->snd_cwnd;
1035                 tp->snd_cwnd = tp->snd_ssthresh;
1036                 /*
1037                  * Formally exit recovery, and let the CC module adjust
1038                  * ssthresh as intended.
1039                  */
1040                 EXIT_RECOVERY(tp->t_flags);
1041                 cc_cong_signal(tp, th, CC_NDUPACK);
1042                 /*
1043                  * For PRR, adjust recover_fs as if this new reduction
1044                  * initialized this variable.
1045                  * cwnd will be adjusted by SACK or PRR processing
1046                  * subsequently, only set it to a safe value here.
1047                  */
1048                 tp->snd_cwnd = tcp_maxseg(tp);
1049                 tp->sackhint.recover_fs = (tp->snd_max - tp->snd_una) -
1050                                             tp->sackhint.recover_fs;
1051         }
1052 }