2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995
5 * The Regents of the University of California.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
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.
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
32 * @(#)tcp_sack.c 8.12 (Berkeley) 5/24/95
36 * @@(#)COPYRIGHT 1.1 (NRL) 17 January 1995
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:
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.
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.
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).
75 #include <sys/cdefs.h>
76 __FBSDID("$FreeBSD$");
79 #include "opt_inet6.h"
80 #include "opt_tcpdebug.h"
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>
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>
95 #include <machine/cpu.h> /* before tcp_seq.h, for tcp_random18() */
100 #include <net/if_var.h>
101 #include <net/route.h>
102 #include <net/vnet.h>
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>
123 #include <netinet/tcp_debug.h>
124 #endif /* TCPDEBUG */
126 #include <machine/in_cksum.h>
128 VNET_DECLARE(struct uma_zone *, sack_hole_zone);
129 #define V_sack_hole_zone VNET(sack_hole_zone)
131 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, sack, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
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");
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");
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");
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");
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");
160 * This function will find overlaps with the currently stored sackblocks
161 * and add any overlap as a dsack block upfront
164 tcp_update_dsack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
166 struct sackblk head_blk,mid_blk,saved_blks[MAX_SACK_BLKS];
167 int i, j, n, identical;
170 INP_WLOCK_ASSERT(tp->t_inpcb);
172 KASSERT(SEQ_LT(rcv_start, rcv_end), ("rcv_start < rcv_end"));
174 if (SEQ_LT(rcv_end, tp->rcv_nxt) ||
175 ((rcv_end == tp->rcv_nxt) &&
176 (tp->rcv_numsacks > 0 ) &&
177 (tp->sackblks[0].end == tp->rcv_nxt))) {
178 saved_blks[0].start = rcv_start;
179 saved_blks[0].end = rcv_end;
181 saved_blks[0].start = saved_blks[0].end = 0;
184 head_blk.start = head_blk.end = 0;
185 mid_blk.start = rcv_start;
186 mid_blk.end = rcv_end;
189 for (i = 0; i < tp->rcv_numsacks; i++) {
190 start = tp->sackblks[i].start;
191 end = tp->sackblks[i].end;
192 if (SEQ_LT(rcv_end, start)) {
193 /* pkt left to sack blk */
196 if (SEQ_GT(rcv_start, end)) {
197 /* pkt right to sack blk */
200 if (SEQ_GT(tp->rcv_nxt, end)) {
201 if ((SEQ_MAX(rcv_start, start) != SEQ_MIN(rcv_end, end)) &&
202 (SEQ_GT(head_blk.start, SEQ_MAX(rcv_start, start)) ||
203 (head_blk.start == head_blk.end))) {
204 head_blk.start = SEQ_MAX(rcv_start, start);
205 head_blk.end = SEQ_MIN(rcv_end, end);
209 if (((head_blk.start == head_blk.end) ||
210 SEQ_LT(start, head_blk.start)) &&
211 (SEQ_GT(end, rcv_start) &&
212 SEQ_LEQ(start, rcv_end))) {
213 head_blk.start = start;
216 mid_blk.start = SEQ_MIN(mid_blk.start, start);
217 mid_blk.end = SEQ_MAX(mid_blk.end, end);
218 if ((mid_blk.start == start) &&
219 (mid_blk.end == end))
222 if (SEQ_LT(head_blk.start, head_blk.end)) {
223 /* store overlapping range */
224 saved_blks[0].start = SEQ_MAX(rcv_start, head_blk.start);
225 saved_blks[0].end = SEQ_MIN(rcv_end, head_blk.end);
229 * Second, if not ACKed, store the SACK block that
230 * overlaps with the DSACK block unless it is identical
232 if ((SEQ_LT(tp->rcv_nxt, mid_blk.end) &&
233 !((mid_blk.start == saved_blks[0].start) &&
234 (mid_blk.end == saved_blks[0].end))) ||
236 saved_blks[n].start = mid_blk.start;
237 saved_blks[n++].end = mid_blk.end;
239 for (j = 0; (j < tp->rcv_numsacks) && (n < MAX_SACK_BLKS); j++) {
240 if (((SEQ_LT(tp->sackblks[j].end, mid_blk.start) ||
241 SEQ_GT(tp->sackblks[j].start, mid_blk.end)) &&
242 (SEQ_GT(tp->sackblks[j].start, tp->rcv_nxt))))
243 saved_blks[n++] = tp->sackblks[j];
246 for (i = 0; i < n; i++) {
247 /* we can end up with a stale initial entry */
248 if (SEQ_LT(saved_blks[i].start, saved_blks[i].end)) {
249 tp->sackblks[j++] = saved_blks[i];
252 tp->rcv_numsacks = j;
256 * This function is called upon receipt of new valid data (while not in
257 * header prediction mode), and it updates the ordered list of sacks.
260 tcp_update_sack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
263 * First reported block MUST be the most recent one. Subsequent
264 * blocks SHOULD be in the order in which they arrived at the
265 * receiver. These two conditions make the implementation fully
266 * compliant with RFC 2018.
268 struct sackblk head_blk, saved_blks[MAX_SACK_BLKS];
269 int num_head, num_saved, i;
271 INP_WLOCK_ASSERT(tp->t_inpcb);
273 /* Check arguments. */
274 KASSERT(SEQ_LEQ(rcv_start, rcv_end), ("rcv_start <= rcv_end"));
276 if ((rcv_start == rcv_end) &&
277 (tp->rcv_numsacks >= 1) &&
278 (rcv_end == tp->sackblks[0].end)) {
279 /* retaining DSACK block below rcv_nxt (todrop) */
280 head_blk = tp->sackblks[0];
282 /* SACK block for the received segment. */
283 head_blk.start = rcv_start;
284 head_blk.end = rcv_end;
288 * Merge updated SACK blocks into head_blk, and save unchanged SACK
289 * blocks into saved_blks[]. num_saved will have the number of the
293 for (i = 0; i < tp->rcv_numsacks; i++) {
294 tcp_seq start = tp->sackblks[i].start;
295 tcp_seq end = tp->sackblks[i].end;
296 if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
298 * Discard this SACK block.
300 } else if (SEQ_LEQ(head_blk.start, end) &&
301 SEQ_GEQ(head_blk.end, start)) {
303 * Merge this SACK block into head_blk. This SACK
304 * block itself will be discarded.
317 * |-----| DSACK smaller
319 if (head_blk.start == end)
320 head_blk.start = start;
321 else if (head_blk.end == start)
324 if (SEQ_LT(head_blk.start, start)) {
325 tcp_seq temp = start;
326 start = head_blk.start;
327 head_blk.start = temp;
329 if (SEQ_GT(head_blk.end, end)) {
334 if ((head_blk.start != start) ||
335 (head_blk.end != end)) {
336 if ((num_saved >= 1) &&
337 SEQ_GEQ(saved_blks[num_saved-1].start, start) &&
338 SEQ_LEQ(saved_blks[num_saved-1].end, end))
340 saved_blks[num_saved].start = start;
341 saved_blks[num_saved].end = end;
347 * This block supercedes the prior block
349 if ((num_saved >= 1) &&
350 SEQ_GEQ(saved_blks[num_saved-1].start, start) &&
351 SEQ_LEQ(saved_blks[num_saved-1].end, end))
354 * Save this SACK block.
356 saved_blks[num_saved].start = start;
357 saved_blks[num_saved].end = end;
363 * Update SACK list in tp->sackblks[].
366 if (SEQ_LT(rcv_start, rcv_end)) {
368 * The received data segment is an out-of-order segment. Put
369 * head_blk at the top of SACK list.
371 tp->sackblks[0] = head_blk;
374 * If the number of saved SACK blocks exceeds its limit,
375 * discard the last SACK block.
377 if (num_saved >= MAX_SACK_BLKS)
380 if ((rcv_start == rcv_end) &&
381 (rcv_start == tp->sackblks[0].end)) {
386 * Copy the saved SACK blocks back.
388 bcopy(saved_blks, &tp->sackblks[num_head],
389 sizeof(struct sackblk) * num_saved);
392 /* Save the number of SACK blocks. */
393 tp->rcv_numsacks = num_head + num_saved;
397 tcp_clean_dsack_blocks(struct tcpcb *tp)
399 struct sackblk saved_blks[MAX_SACK_BLKS];
402 INP_WLOCK_ASSERT(tp->t_inpcb);
404 * Clean up any DSACK blocks that
405 * are in our queue of sack blocks.
409 for (i = 0; i < tp->rcv_numsacks; i++) {
410 tcp_seq start = tp->sackblks[i].start;
411 tcp_seq end = tp->sackblks[i].end;
412 if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
414 * Discard this D-SACK block.
419 * Save this SACK block.
421 saved_blks[num_saved].start = start;
422 saved_blks[num_saved].end = end;
427 * Copy the saved SACK blocks back.
429 bcopy(saved_blks, &tp->sackblks[0],
430 sizeof(struct sackblk) * num_saved);
432 tp->rcv_numsacks = num_saved;
436 * Delete all receiver-side SACK information.
439 tcp_clean_sackreport(struct tcpcb *tp)
443 INP_WLOCK_ASSERT(tp->t_inpcb);
444 tp->rcv_numsacks = 0;
445 for (i = 0; i < MAX_SACK_BLKS; i++)
446 tp->sackblks[i].start = tp->sackblks[i].end=0;
450 * Allocate struct sackhole.
452 static struct sackhole *
453 tcp_sackhole_alloc(struct tcpcb *tp, tcp_seq start, tcp_seq end)
455 struct sackhole *hole;
457 if (tp->snd_numholes >= V_tcp_sack_maxholes ||
458 V_tcp_sack_globalholes >= V_tcp_sack_globalmaxholes) {
459 TCPSTAT_INC(tcps_sack_sboverflow);
463 hole = (struct sackhole *)uma_zalloc(V_sack_hole_zone, M_NOWAIT);
472 atomic_add_int(&V_tcp_sack_globalholes, 1);
478 * Free struct sackhole.
481 tcp_sackhole_free(struct tcpcb *tp, struct sackhole *hole)
484 uma_zfree(V_sack_hole_zone, hole);
487 atomic_subtract_int(&V_tcp_sack_globalholes, 1);
489 KASSERT(tp->snd_numholes >= 0, ("tp->snd_numholes >= 0"));
490 KASSERT(V_tcp_sack_globalholes >= 0, ("tcp_sack_globalholes >= 0"));
494 * Insert new SACK hole into scoreboard.
496 static struct sackhole *
497 tcp_sackhole_insert(struct tcpcb *tp, tcp_seq start, tcp_seq end,
498 struct sackhole *after)
500 struct sackhole *hole;
502 /* Allocate a new SACK hole. */
503 hole = tcp_sackhole_alloc(tp, start, end);
507 /* Insert the new SACK hole into scoreboard. */
509 TAILQ_INSERT_AFTER(&tp->snd_holes, after, hole, scblink);
511 TAILQ_INSERT_TAIL(&tp->snd_holes, hole, scblink);
513 /* Update SACK hint. */
514 if (tp->sackhint.nexthole == NULL)
515 tp->sackhint.nexthole = hole;
521 * Remove SACK hole from scoreboard.
524 tcp_sackhole_remove(struct tcpcb *tp, struct sackhole *hole)
527 /* Update SACK hint. */
528 if (tp->sackhint.nexthole == hole)
529 tp->sackhint.nexthole = TAILQ_NEXT(hole, scblink);
531 /* Remove this SACK hole. */
532 TAILQ_REMOVE(&tp->snd_holes, hole, scblink);
534 /* Free this SACK hole. */
535 tcp_sackhole_free(tp, hole);
539 * Process cumulative ACK and the TCP SACK option to update the scoreboard.
540 * tp->snd_holes is an ordered list of holes (oldest to newest, in terms of
541 * the sequence space).
542 * Returns 1 if incoming ACK has previously unknown SACK information,
546 tcp_sack_doack(struct tcpcb *tp, struct tcpopt *to, tcp_seq th_ack)
548 struct sackhole *cur, *temp;
549 struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1], *sblkp;
550 int i, j, num_sack_blks, sack_changed;
551 int delivered_data, left_edge_delta;
553 INP_WLOCK_ASSERT(tp->t_inpcb);
560 * If SND.UNA will be advanced by SEG.ACK, and if SACK holes exist,
561 * treat [SND.UNA, SEG.ACK) as if it is a SACK block.
562 * Account changes to SND.UNA always in delivered data.
564 if (SEQ_LT(tp->snd_una, th_ack) && !TAILQ_EMPTY(&tp->snd_holes)) {
565 left_edge_delta = th_ack - tp->snd_una;
566 sack_blocks[num_sack_blks].start = tp->snd_una;
567 sack_blocks[num_sack_blks++].end = th_ack;
570 * Append received valid SACK blocks to sack_blocks[], but only if we
571 * received new blocks from the other side.
573 if (to->to_flags & TOF_SACK) {
574 for (i = 0; i < to->to_nsacks; i++) {
575 bcopy((to->to_sacks + i * TCPOLEN_SACK),
576 &sack, sizeof(sack));
577 sack.start = ntohl(sack.start);
578 sack.end = ntohl(sack.end);
579 if (SEQ_GT(sack.end, sack.start) &&
580 SEQ_GT(sack.start, tp->snd_una) &&
581 SEQ_GT(sack.start, th_ack) &&
582 SEQ_LT(sack.start, tp->snd_max) &&
583 SEQ_GT(sack.end, tp->snd_una) &&
584 SEQ_LEQ(sack.end, tp->snd_max)) {
585 sack_blocks[num_sack_blks++] = sack;
590 * Return if SND.UNA is not advanced and no valid SACK block is
593 if (num_sack_blks == 0)
594 return (sack_changed);
597 * Sort the SACK blocks so we can update the scoreboard with just one
598 * pass. The overhead of sorting up to 4+1 elements is less than
599 * making up to 4+1 passes over the scoreboard.
601 for (i = 0; i < num_sack_blks; i++) {
602 for (j = i + 1; j < num_sack_blks; j++) {
603 if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
604 sack = sack_blocks[i];
605 sack_blocks[i] = sack_blocks[j];
606 sack_blocks[j] = sack;
610 if (TAILQ_EMPTY(&tp->snd_holes)) {
612 * Empty scoreboard. Need to initialize snd_fack (it may be
613 * uninitialized or have a bogus value). Scoreboard holes
614 * (from the sack blocks received) are created later below
615 * (in the logic that adds holes to the tail of the
618 tp->snd_fack = SEQ_MAX(tp->snd_una, th_ack);
619 tp->sackhint.sacked_bytes = 0; /* reset */
622 * In the while-loop below, incoming SACK blocks (sack_blocks[]) and
623 * SACK holes (snd_holes) are traversed from their tails with just
624 * one pass in order to reduce the number of compares especially when
625 * the bandwidth-delay product is large.
627 * Note: Typically, in the first RTT of SACK recovery, the highest
628 * three or four SACK blocks with the same ack number are received.
629 * In the second RTT, if retransmitted data segments are not lost,
630 * the highest three or four SACK blocks with ack number advancing
633 sblkp = &sack_blocks[num_sack_blks - 1]; /* Last SACK block */
634 tp->sackhint.last_sack_ack = sblkp->end;
635 if (SEQ_LT(tp->snd_fack, sblkp->start)) {
637 * The highest SACK block is beyond fack. Append new SACK
638 * hole at the tail. If the second or later highest SACK
639 * blocks are also beyond the current fack, they will be
640 * inserted by way of hole splitting in the while-loop below.
642 temp = tcp_sackhole_insert(tp, tp->snd_fack,sblkp->start,NULL);
644 delivered_data += sblkp->end - sblkp->start;
645 tp->snd_fack = sblkp->end;
646 /* Go to the previous sack block. */
651 * We failed to add a new hole based on the current
652 * sack block. Skip over all the sack blocks that
653 * fall completely to the right of snd_fack and
654 * proceed to trim the scoreboard based on the
655 * remaining sack blocks. This also trims the
656 * scoreboard for th_ack (which is sack_blocks[0]).
658 while (sblkp >= sack_blocks &&
659 SEQ_LT(tp->snd_fack, sblkp->start))
661 if (sblkp >= sack_blocks &&
662 SEQ_LT(tp->snd_fack, sblkp->end)) {
663 delivered_data += sblkp->end - tp->snd_fack;
664 tp->snd_fack = sblkp->end;
668 } else if (SEQ_LT(tp->snd_fack, sblkp->end)) {
669 /* fack is advanced. */
670 delivered_data += sblkp->end - tp->snd_fack;
671 tp->snd_fack = sblkp->end;
674 cur = TAILQ_LAST(&tp->snd_holes, sackhole_head); /* Last SACK hole. */
676 * Since the incoming sack blocks are sorted, we can process them
677 * making one sweep of the scoreboard.
679 while (sblkp >= sack_blocks && cur != NULL) {
680 if (SEQ_GEQ(sblkp->start, cur->end)) {
682 * SACKs data beyond the current hole. Go to the
683 * previous sack block.
688 if (SEQ_LEQ(sblkp->end, cur->start)) {
690 * SACKs data before the current hole. Go to the
693 cur = TAILQ_PREV(cur, sackhole_head, scblink);
696 tp->sackhint.sack_bytes_rexmit -= (cur->rxmit - cur->start);
697 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
698 ("sackhint bytes rtx >= 0"));
700 if (SEQ_LEQ(sblkp->start, cur->start)) {
701 /* Data acks at least the beginning of hole. */
702 if (SEQ_GEQ(sblkp->end, cur->end)) {
703 /* Acks entire hole, so delete hole. */
704 delivered_data += (cur->end - cur->start);
706 cur = TAILQ_PREV(cur, sackhole_head, scblink);
707 tcp_sackhole_remove(tp, temp);
709 * The sack block may ack all or part of the
710 * next hole too, so continue onto the next
715 /* Move start of hole forward. */
716 delivered_data += (sblkp->end - cur->start);
717 cur->start = sblkp->end;
718 cur->rxmit = SEQ_MAX(cur->rxmit, cur->start);
721 /* Data acks at least the end of hole. */
722 if (SEQ_GEQ(sblkp->end, cur->end)) {
723 /* Move end of hole backward. */
724 delivered_data += (cur->end - sblkp->start);
725 cur->end = sblkp->start;
726 cur->rxmit = SEQ_MIN(cur->rxmit, cur->end);
729 * ACKs some data in middle of a hole; need
730 * to split current hole
732 temp = tcp_sackhole_insert(tp, sblkp->end,
735 if (SEQ_GT(cur->rxmit, temp->rxmit)) {
736 temp->rxmit = cur->rxmit;
737 tp->sackhint.sack_bytes_rexmit
741 cur->end = sblkp->start;
742 cur->rxmit = SEQ_MIN(cur->rxmit,
744 delivered_data += (sblkp->end - sblkp->start);
748 tp->sackhint.sack_bytes_rexmit += (cur->rxmit - cur->start);
750 * Testing sblkp->start against cur->start tells us whether
751 * we're done with the sack block or the sack hole.
752 * Accordingly, we advance one or the other.
754 if (SEQ_LEQ(sblkp->start, cur->start))
755 cur = TAILQ_PREV(cur, sackhole_head, scblink);
759 if (!(to->to_flags & TOF_SACK))
761 * If this ACK did not contain any
762 * SACK blocks, any only moved the
763 * left edge right, it is a pure
764 * cumulative ACK. Do not count
765 * DupAck for this. Also required
766 * for RFC6675 rescue retransmission.
769 tp->sackhint.delivered_data = delivered_data;
770 tp->sackhint.sacked_bytes += delivered_data - left_edge_delta;
771 KASSERT((delivered_data >= 0), ("delivered_data < 0"));
772 KASSERT((tp->sackhint.sacked_bytes >= 0), ("sacked_bytes < 0"));
773 return (sack_changed);
777 * Free all SACK holes to clear the scoreboard.
780 tcp_free_sackholes(struct tcpcb *tp)
784 INP_WLOCK_ASSERT(tp->t_inpcb);
785 while ((q = TAILQ_FIRST(&tp->snd_holes)) != NULL)
786 tcp_sackhole_remove(tp, q);
787 tp->sackhint.sack_bytes_rexmit = 0;
789 KASSERT(tp->snd_numholes == 0, ("tp->snd_numholes == 0"));
790 KASSERT(tp->sackhint.nexthole == NULL,
791 ("tp->sackhint.nexthole == NULL"));
795 * Partial ack handling within a sack recovery episode. Keeping this very
796 * simple for now. When a partial ack is received, force snd_cwnd to a value
797 * that will allow the sender to transmit no more than 2 segments. If
798 * necessary, a better scheme can be adopted at a later point, but for now,
799 * the goal is to prevent the sender from bursting a large amount of data in
800 * the midst of sack recovery.
803 tcp_sack_partialack(struct tcpcb *tp, struct tcphdr *th)
806 u_int maxseg = tcp_maxseg(tp);
808 INP_WLOCK_ASSERT(tp->t_inpcb);
809 tcp_timer_activate(tp, TT_REXMT, 0);
811 /* Send one or 2 segments based on how much new data was acked. */
812 if ((BYTES_THIS_ACK(tp, th) / maxseg) >= 2)
814 tp->snd_cwnd = (tp->sackhint.sack_bytes_rexmit +
815 (tp->snd_nxt - tp->snd_recover) + num_segs * maxseg);
816 if (tp->snd_cwnd > tp->snd_ssthresh)
817 tp->snd_cwnd = tp->snd_ssthresh;
818 tp->t_flags |= TF_ACKNOW;
820 * RFC6675 rescue retransmission
821 * Add a hole between th_ack (snd_una is not yet set) and snd_max,
822 * if this was a pure cumulative ACK and no data was send beyond
823 * recovery point. Since the data in the socket has not been freed
824 * at this point, we check if the scoreboard is empty, and the ACK
825 * delivered some new data, indicating a full ACK. Also, if the
826 * recovery point is still at snd_max, we are probably application
827 * limited. However, this inference might not always be true. The
828 * rescue retransmission may rarely be slightly premature
829 * compared to RFC6675.
830 * The corresponding ACK+SACK will cause any further outstanding
831 * segments to be retransmitted. This addresses a corner case, when
832 * the trailing packets of a window are lost and no further data
833 * is available for sending.
835 if ((V_tcp_do_newsack) &&
836 SEQ_LT(th->th_ack, tp->snd_recover) &&
837 (tp->snd_recover == tp->snd_max) &&
838 TAILQ_EMPTY(&tp->snd_holes) &&
839 (tp->sackhint.delivered_data > 0)) {
841 * Exclude FIN sequence space in
842 * the hole for the rescue retransmission,
843 * and also don't create a hole, if only
844 * the ACK for a FIN is outstanding.
846 tcp_seq highdata = tp->snd_max;
847 if (tp->t_flags & TF_SENTFIN)
849 if (th->th_ack != highdata)
850 (void)tcp_sackhole_insert(tp, SEQ_MAX(th->th_ack,
851 highdata - maxseg), highdata, NULL);
853 (void) tp->t_fb->tfb_tcp_output(tp);
858 * Debug version of tcp_sack_output() that walks the scoreboard. Used for
859 * now to sanity check the hint.
861 static struct sackhole *
862 tcp_sack_output_debug(struct tcpcb *tp, int *sack_bytes_rexmt)
866 INP_WLOCK_ASSERT(tp->t_inpcb);
867 *sack_bytes_rexmt = 0;
868 TAILQ_FOREACH(p, &tp->snd_holes, scblink) {
869 if (SEQ_LT(p->rxmit, p->end)) {
870 if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */
873 *sack_bytes_rexmt += (p->rxmit - p->start);
876 *sack_bytes_rexmt += (p->rxmit - p->start);
883 * Returns the next hole to retransmit and the number of retransmitted bytes
884 * from the scoreboard. We store both the next hole and the number of
885 * retransmitted bytes as hints (and recompute these on the fly upon SACK/ACK
886 * reception). This avoids scoreboard traversals completely.
888 * The loop here will traverse *at most* one link. Here's the argument. For
889 * the loop to traverse more than 1 link before finding the next hole to
890 * retransmit, we would need to have at least 1 node following the current
891 * hint with (rxmit == end). But, for all holes following the current hint,
892 * (start == rxmit), since we have not yet retransmitted from them.
893 * Therefore, in order to traverse more 1 link in the loop below, we need to
894 * have at least one node following the current hint with (start == rxmit ==
895 * end). But that can't happen, (start == end) means that all the data in
896 * that hole has been sacked, in which case, the hole would have been removed
897 * from the scoreboard.
900 tcp_sack_output(struct tcpcb *tp, int *sack_bytes_rexmt)
902 struct sackhole *hole = NULL;
904 INP_WLOCK_ASSERT(tp->t_inpcb);
905 *sack_bytes_rexmt = tp->sackhint.sack_bytes_rexmit;
906 hole = tp->sackhint.nexthole;
907 if (hole == NULL || SEQ_LT(hole->rxmit, hole->end))
909 while ((hole = TAILQ_NEXT(hole, scblink)) != NULL) {
910 if (SEQ_LT(hole->rxmit, hole->end)) {
911 tp->sackhint.nexthole = hole;
920 * After a timeout, the SACK list may be rebuilt. This SACK information
921 * should be used to avoid retransmitting SACKed data. This function
922 * traverses the SACK list to see if snd_nxt should be moved forward.
925 tcp_sack_adjust(struct tcpcb *tp)
927 struct sackhole *p, *cur = TAILQ_FIRST(&tp->snd_holes);
929 INP_WLOCK_ASSERT(tp->t_inpcb);
931 return; /* No holes */
932 if (SEQ_GEQ(tp->snd_nxt, tp->snd_fack))
933 return; /* We're already beyond any SACKed blocks */
935 * Two cases for which we want to advance snd_nxt:
936 * i) snd_nxt lies between end of one hole and beginning of another
937 * ii) snd_nxt lies between end of last hole and snd_fack
939 while ((p = TAILQ_NEXT(cur, scblink)) != NULL) {
940 if (SEQ_LT(tp->snd_nxt, cur->end))
942 if (SEQ_GEQ(tp->snd_nxt, p->start))
945 tp->snd_nxt = p->start;
949 if (SEQ_LT(tp->snd_nxt, cur->end))
951 tp->snd_nxt = tp->snd_fack;