2 * Copyright (c) 2003, 2004 Jeffrey M. Hsu. All rights reserved.
3 * Copyright (c) 2003, 2004 The DragonFly Project. All rights reserved.
5 * This code is derived from software contributed to The DragonFly Project
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 DragonFly Project nor the names of its
17 * contributors may be used to endorse or promote products derived
18 * from this software without specific, prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
23 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
24 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
25 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
26 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
27 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
28 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
29 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
30 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * $DragonFly: src/sys/netinet/tcp_sack.c,v 1.8 2008/08/15 21:37:16 nth Exp $
36 #include <sys/param.h>
37 #include <sys/systm.h>
38 #include <sys/kernel.h>
39 #include <sys/malloc.h>
40 #include <sys/queue.h>
41 #include <sys/thread.h>
42 #include <sys/types.h>
43 #include <sys/socket.h>
44 #include <sys/socketvar.h>
48 #include <netinet/in.h>
49 #include <netinet/in_systm.h>
50 #include <netinet/ip.h>
51 #include <netinet/in_var.h>
52 #include <netinet/in_pcb.h>
53 #include <netinet/ip_var.h>
54 #include <netinet/tcp.h>
55 #include <netinet/tcp_seq.h>
56 #include <netinet/tcp_var.h>
69 TAILQ_ENTRY(sackblock) sblk_list;
72 #define MAXSAVEDBLOCKS 8 /* per connection limit */
74 static void insert_block(struct scoreboard *scb, struct sackblock *newblock);
75 static void update_lostseq(struct scoreboard *scb, tcp_seq snd_una,
78 static MALLOC_DEFINE(M_SACKBLOCK, "sblk", "sackblock struct");
81 * Per-tcpcb initialization.
84 tcp_sack_tcpcb_init(struct tcpcb *tp)
86 struct scoreboard *scb = &tp->scb;
89 TAILQ_INIT(&scb->sackblocks);
90 scb->lastfound = NULL;
94 * Find the SACK block containing or immediately preceding "seq".
95 * The boolean result indicates whether the sequence is actually
96 * contained in the SACK block.
99 sack_block_lookup(struct scoreboard *scb, tcp_seq seq, struct sackblock **sb)
101 struct sackblock *hint = scb->lastfound;
102 struct sackblock *cur, *last, *prev;
104 if (TAILQ_EMPTY(&scb->sackblocks)) {
110 /* No hint. Search from start to end. */
111 cur = TAILQ_FIRST(&scb->sackblocks);
113 prev = TAILQ_LAST(&scb->sackblocks, sackblock_list);
115 if (SEQ_GEQ(seq, hint->sblk_start)) {
116 /* Search from hint to end of list. */
119 prev = TAILQ_LAST(&scb->sackblocks, sackblock_list);
121 /* Search from front of list to hint. */
122 cur = TAILQ_FIRST(&scb->sackblocks);
124 prev = TAILQ_PREV(hint, sackblock_list, sblk_list);
129 if (SEQ_GT(cur->sblk_end, seq)) {
130 if (SEQ_GEQ(seq, cur->sblk_start)) {
131 *sb = scb->lastfound = cur;
134 *sb = scb->lastfound =
135 TAILQ_PREV(cur, sackblock_list, sblk_list);
139 cur = TAILQ_NEXT(cur, sblk_list);
140 } while (cur != last);
142 *sb = scb->lastfound = prev;
147 * Allocate a SACK block.
149 static __inline struct sackblock *
150 alloc_sackblock(const struct raw_sackblock *raw_sb)
152 struct sackblock *sb;
154 sb = kmalloc(sizeof(struct sackblock), M_SACKBLOCK, M_NOWAIT);
156 sb->sblk_start = raw_sb->rblk_start;
157 sb->sblk_end = raw_sb->rblk_end;
166 free_sackblock(struct sackblock *s)
168 kfree(s, M_SACKBLOCK);
172 * Free up SACK blocks for data that's been acked.
175 tcp_sack_ack_blocks(struct scoreboard *scb, tcp_seq th_ack)
177 struct sackblock *sb, *nb;
179 sb = TAILQ_FIRST(&scb->sackblocks);
180 while (sb && SEQ_LEQ(sb->sblk_end, th_ack)) {
181 nb = TAILQ_NEXT(sb, sblk_list);
182 if (scb->lastfound == sb)
183 scb->lastfound = NULL;
184 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
187 KASSERT(scb->nblocks >= 0,
188 ("SACK block count underflow: %d < 0", scb->nblocks));
191 if (sb && SEQ_GT(th_ack, sb->sblk_start))
192 sb->sblk_start = th_ack; /* other side reneged? XXX */
196 * Delete and free SACK blocks saved in scoreboard.
199 tcp_sack_cleanup(struct scoreboard *scb)
201 struct sackblock *sb, *nb;
203 TAILQ_FOREACH_MUTABLE(sb, &scb->sackblocks, sblk_list, nb) {
207 KASSERT(scb->nblocks == 0,
208 ("SACK block %d count not zero", scb->nblocks));
209 TAILQ_INIT(&scb->sackblocks);
210 scb->lastfound = NULL;
214 * Cleanup the reported SACK block information
217 tcp_sack_report_cleanup(struct tcpcb *tp)
219 tp->t_flags &= ~(TF_DUPSEG | TF_ENCLOSESEG | TF_SACKLEFT);
220 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
224 * Returns 0 if not D-SACK block,
226 * 2 if duplicate of out-of-order D-SACK block.
229 tcp_sack_ndsack_blocks(struct raw_sackblock *blocks, const int numblocks,
235 if (SEQ_LT(blocks[0].rblk_start, snd_una))
238 /* block 0 inside block 1 */
240 SEQ_GEQ(blocks[0].rblk_start, blocks[1].rblk_start) &&
241 SEQ_LEQ(blocks[0].rblk_end, blocks[1].rblk_end))
248 * Update scoreboard on new incoming ACK.
251 tcp_sack_add_blocks(struct tcpcb *tp, struct tcpopt *to)
253 const int numblocks = to->to_nsackblocks;
254 struct raw_sackblock *blocks = to->to_sackblocks;
255 struct scoreboard *scb = &tp->scb;
256 struct sackblock *sb;
260 if (tcp_sack_ndsack_blocks(blocks, numblocks, tp->snd_una) > 0)
265 for (i = startblock; i < numblocks; i++) {
266 struct raw_sackblock *newsackblock = &blocks[i];
268 /* don't accept bad SACK blocks */
269 if (SEQ_GT(newsackblock->rblk_end, tp->snd_max)) {
270 tcpstat.tcps_rcvbadsackopt++;
271 break; /* skip all other blocks */
273 tcpstat.tcps_sacksbupdate++;
275 sb = alloc_sackblock(newsackblock);
276 if (sb == NULL) { /* do some sort of cleanup? XXX */
277 tcpstat.tcps_sacksbfailed++;
278 break; /* just skip rest of blocks */
280 insert_block(scb, sb);
285 tcp_sack_update_scoreboard(struct tcpcb *tp, struct tcpopt *to)
287 struct scoreboard *scb = &tp->scb;
289 tcp_sack_ack_blocks(scb, tp->snd_una);
290 tcp_sack_add_blocks(tp, to);
291 update_lostseq(scb, tp->snd_una, tp->t_maxseg);
292 if (SEQ_LT(tp->rexmt_high, tp->snd_una))
293 tp->rexmt_high = tp->snd_una;
297 * Insert SACK block into sender's scoreboard.
300 insert_block(struct scoreboard *scb, struct sackblock *newblock)
302 struct sackblock *sb, *workingblock;
303 boolean_t overlap_front;
305 if (TAILQ_EMPTY(&scb->sackblocks)) {
306 KASSERT(scb->nblocks == 0, ("emply scb w/ blocks"));
308 TAILQ_INSERT_HEAD(&scb->sackblocks, newblock, sblk_list);
311 KASSERT(scb->nblocks > 0, ("insert_block() called w/ no blocks"));
313 if (scb->nblocks == MAXSAVEDBLOCKS) {
315 * Should try to kick out older blocks XXX JH
316 * May be able to coalesce with existing block.
317 * Or, go other way and free all blocks if we hit this limit.
319 free_sackblock(newblock);
320 tcpstat.tcps_sacksboverflow++;
323 KASSERT(scb->nblocks < MAXSAVEDBLOCKS,
324 ("too many SACK blocks %d", scb->nblocks));
326 overlap_front = sack_block_lookup(scb, newblock->sblk_start, &sb);
329 workingblock = newblock;
330 TAILQ_INSERT_HEAD(&scb->sackblocks, newblock, sblk_list);
333 if (overlap_front || sb->sblk_end == newblock->sblk_start) {
334 /* extend old block and discard new one */
336 if (SEQ_GT(newblock->sblk_end, sb->sblk_end))
337 sb->sblk_end = newblock->sblk_end;
338 free_sackblock(newblock);
339 tcpstat.tcps_sacksbreused++;
341 workingblock = newblock;
342 TAILQ_INSERT_AFTER(&scb->sackblocks, sb, newblock,
348 /* Consolidate right-hand side. */
349 sb = TAILQ_NEXT(workingblock, sblk_list);
351 SEQ_GEQ(workingblock->sblk_end, sb->sblk_end)) {
352 struct sackblock *nextblock;
354 nextblock = TAILQ_NEXT(sb, sblk_list);
355 if (scb->lastfound == sb)
356 scb->lastfound = NULL;
357 /* Remove completely overlapped block */
358 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
361 KASSERT(scb->nblocks > 0,
362 ("removed overlapped block: %d blocks left", scb->nblocks));
366 SEQ_GEQ(workingblock->sblk_end, sb->sblk_start)) {
367 /* Extend new block to cover partially overlapped old block. */
368 workingblock->sblk_end = sb->sblk_end;
369 if (scb->lastfound == sb)
370 scb->lastfound = NULL;
371 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
374 KASSERT(scb->nblocks > 0,
375 ("removed partial right: %d blocks left", scb->nblocks));
379 #ifdef DEBUG_SACK_BLOCKS
381 tcp_sack_dump_blocks(struct scoreboard *scb)
383 struct sackblock *sb;
385 kprintf("%d blocks:", scb->nblocks);
386 TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
387 kprintf(" [%u, %u)", sb->sblk_start, sb->sblk_end);
392 tcp_sack_dump_blocks(struct scoreboard *scb)
398 * Optimization to quickly determine which packets are lost.
401 update_lostseq(struct scoreboard *scb, tcp_seq snd_una, u_int maxseg)
403 struct sackblock *sb;
405 int bytes_sacked = 0;
407 sb = TAILQ_LAST(&scb->sackblocks, sackblock_list);
410 bytes_sacked += sb->sblk_end - sb->sblk_start;
411 if (nsackblocks == tcprexmtthresh ||
412 bytes_sacked >= tcprexmtthresh * maxseg) {
413 scb->lostseq = sb->sblk_start;
416 sb = TAILQ_PREV(sb, sackblock_list, sblk_list);
418 scb->lostseq = snd_una;
422 * Return whether the given sequence number is considered lost.
425 scb_islost(struct scoreboard *scb, tcp_seq seqnum)
427 return SEQ_LT(seqnum, scb->lostseq);
431 * True if at least "amount" has been SACKed. Used by Early Retransmit.
434 tcp_sack_has_sacked(struct scoreboard *scb, u_int amount)
436 struct sackblock *sb;
437 int bytes_sacked = 0;
439 TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
440 bytes_sacked += sb->sblk_end - sb->sblk_start;
441 if (bytes_sacked >= amount)
448 * Number of bytes SACKed below seq.
451 tcp_sack_bytes_below(struct scoreboard *scb, tcp_seq seq)
453 struct sackblock *sb;
454 int bytes_sacked = 0;
456 sb = TAILQ_FIRST(&scb->sackblocks);
457 while (sb && SEQ_GT(seq, sb->sblk_start)) {
458 bytes_sacked += seq_min(seq, sb->sblk_end) - sb->sblk_start;
459 sb = TAILQ_NEXT(sb, sblk_list);
465 * Return estimate of the number of bytes outstanding in the network.
468 tcp_sack_compute_pipe(struct tcpcb *tp)
470 struct scoreboard *scb = &tp->scb;
471 struct sackblock *sb;
472 int nlost, nretransmitted;
475 nlost = tp->snd_max - scb->lostseq;
476 nretransmitted = tp->rexmt_high - tp->snd_una;
478 TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
479 if (SEQ_LT(sb->sblk_start, tp->rexmt_high)) {
480 end = seq_min(sb->sblk_end, tp->rexmt_high);
481 nretransmitted -= end - sb->sblk_start;
483 if (SEQ_GEQ(sb->sblk_start, scb->lostseq))
484 nlost -= sb->sblk_end - sb->sblk_start;
487 return (nlost + nretransmitted);
491 * Return the sequence number and length of the next segment to transmit
492 * when in Fast Recovery.
495 tcp_sack_nextseg(struct tcpcb *tp, tcp_seq *nextrexmt, uint32_t *plen,
498 struct scoreboard *scb = &tp->scb;
499 struct socket *so = tp->t_inpcb->inp_socket;
500 struct sackblock *sb;
501 const struct sackblock *lastblock =
502 TAILQ_LAST(&scb->sackblocks, sackblock_list);
506 /* skip SACKed data */
507 tcp_sack_skip_sacked(scb, &tp->rexmt_high);
509 /* Look for lost data. */
510 torexmt = tp->rexmt_high;
512 if (lastblock != NULL) {
513 if (SEQ_LT(torexmt, lastblock->sblk_end) &&
514 scb_islost(scb, torexmt)) {
516 *nextrexmt = torexmt;
517 /* If the left-hand edge has been SACKed, pull it in. */
518 if (sack_block_lookup(scb, torexmt + tp->t_maxseg, &sb))
519 *plen = sb->sblk_start - torexmt;
521 *plen = tp->t_maxseg;
526 /* See if unsent data available within send window. */
527 off = tp->snd_max - tp->snd_una;
528 len = (long) ulmin(so->so_snd.ssb_cc, tp->snd_wnd) - off;
530 *nextrexmt = tp->snd_max; /* Send new data. */
531 *plen = tp->t_maxseg;
535 /* We're less certain this data has been lost. */
536 if (lastblock == NULL || SEQ_LT(torexmt, lastblock->sblk_end))
543 * Return the next sequence number higher than "*prexmt" that has
547 tcp_sack_skip_sacked(struct scoreboard *scb, tcp_seq *prexmt)
549 struct sackblock *sb;
551 /* skip SACKed data */
552 if (sack_block_lookup(scb, *prexmt, &sb))
553 *prexmt = sb->sblk_end;
558 tcp_sack_save_scoreboard(struct scoreboard *scb)
560 struct scoreboard *scb = &tp->scb;
562 scb->sackblocks_prev = scb->sackblocks;
563 TAILQ_INIT(&scb->sackblocks);
567 tcp_sack_revert_scoreboard(struct scoreboard *scb, tcp_seq snd_una,
570 struct sackblock *sb;
572 scb->sackblocks = scb->sackblocks_prev;
574 TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
576 tcp_sack_ack_blocks(scb, snd_una);
577 scb->lastfound = NULL;
581 #ifdef DEBUG_SACK_HISTORY
583 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
588 /* only need a couple of these to debug most problems */
592 kprintf("%s:\tnsackhistory %d: ", msg, tp->nsackhistory);
593 for (i = 0; i < tp->nsackhistory; ++i)
594 kprintf("[%u, %u) ", tp->sackhistory[i].rblk_start,
595 tp->sackhistory[i].rblk_end);
600 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
606 * Remove old SACK blocks from the SACK history that have already been ACKed.
609 tcp_sack_ack_history(struct tcpcb *tp)
611 int i, nblocks, openslot;
613 tcp_sack_dump_history("before tcp_sack_ack_history", tp);
614 nblocks = tp->nsackhistory;
615 for (i = openslot = 0; i < nblocks; ++i) {
616 if (SEQ_LEQ(tp->sackhistory[i].rblk_end, tp->rcv_nxt)) {
620 if (SEQ_LT(tp->sackhistory[i].rblk_start, tp->rcv_nxt))
621 tp->sackhistory[i].rblk_start = tp->rcv_nxt;
625 tp->sackhistory[openslot++] = tp->sackhistory[i];
627 tcp_sack_dump_history("after tcp_sack_ack_history", tp);
628 KASSERT(openslot == tp->nsackhistory,
629 ("tcp_sack_ack_history miscounted: %d != %d",
630 openslot, tp->nsackhistory));
634 * Add or merge newblock into reported history.
635 * Also remove or update SACK blocks that will be acked.
638 tcp_sack_update_reported_history(struct tcpcb *tp, tcp_seq start, tcp_seq end)
640 struct raw_sackblock copy[MAX_SACK_REPORT_BLOCKS];
643 tcp_sack_dump_history("before tcp_sack_update_reported_history", tp);
647 * 1) newblock == oldblock
648 * 2) oldblock contains newblock
649 * 3) newblock contains oldblock
650 * 4) tail of oldblock overlaps or abuts start of newblock
651 * 5) tail of newblock overlaps or abuts head of oldblock
653 for (i = cindex = 0; i < tp->nsackhistory; ++i) {
654 struct raw_sackblock *oldblock = &tp->sackhistory[i];
655 tcp_seq old_start = oldblock->rblk_start;
656 tcp_seq old_end = oldblock->rblk_end;
658 if (SEQ_LT(end, old_start) || SEQ_GT(start, old_end)) {
659 /* Case 0: no overlap. Copy old block. */
660 copy[cindex++] = *oldblock;
664 if (SEQ_GEQ(start, old_start) && SEQ_LEQ(end, old_end)) {
665 /* Cases 1 & 2. Move block to front of history. */
670 /* no need to check rest of blocks */
671 for (j = i + 1; j < tp->nsackhistory; ++j)
672 copy[cindex++] = tp->sackhistory[j];
676 if (SEQ_GEQ(old_end, start) && SEQ_LT(old_start, start)) {
677 /* Case 4: extend start of new block. */
679 } else if (SEQ_GEQ(end, old_start) && SEQ_GT(old_end, end)) {
680 /* Case 5: extend end of new block */
683 /* Case 3. Delete old block by not copying it. */
684 KASSERT(SEQ_LEQ(start, old_start) &&
685 SEQ_GEQ(end, old_end),
686 ("bad logic: old [%u, %u), new [%u, %u)",
687 old_start, old_end, start, end));
691 /* insert new block */
692 tp->sackhistory[0].rblk_start = start;
693 tp->sackhistory[0].rblk_end = end;
694 cindex = min(cindex, MAX_SACK_REPORT_BLOCKS - 1);
695 for (i = 0; i < cindex; ++i)
696 tp->sackhistory[i + 1] = copy[i];
697 tp->nsackhistory = cindex + 1;
698 tcp_sack_dump_history("after tcp_sack_update_reported_history", tp);
702 * Fill in SACK report to return to data sender.
705 tcp_sack_fill_report(struct tcpcb *tp, u_char *opt, u_int *plen)
707 u_int optlen = *plen;
708 uint32_t *lp = (uint32_t *)(opt + optlen);
710 tcp_seq hstart = tp->rcv_nxt, hend;
713 KASSERT(TCP_MAXOLEN - optlen >=
714 TCPOLEN_SACK_ALIGNED + TCPOLEN_SACK_BLOCK,
715 ("no room for SACK header and one block: optlen %d", optlen));
717 if (tp->t_flags & TF_DUPSEG)
718 tcpstat.tcps_snddsackopt++;
720 tcpstat.tcps_sndsackopt++;
723 optlen += TCPOLEN_SACK_ALIGNED;
725 tcp_sack_ack_history(tp);
726 if (tp->reportblk.rblk_start != tp->reportblk.rblk_end) {
727 *lp++ = htonl(tp->reportblk.rblk_start);
728 *lp++ = htonl(tp->reportblk.rblk_end);
729 optlen += TCPOLEN_SACK_BLOCK;
730 hstart = tp->reportblk.rblk_start;
731 hend = tp->reportblk.rblk_end;
732 if (tp->t_flags & TF_ENCLOSESEG) {
733 KASSERT(TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK,
734 ("no room for enclosing SACK block: oplen %d",
736 *lp++ = htonl(tp->encloseblk.rblk_start);
737 *lp++ = htonl(tp->encloseblk.rblk_end);
738 optlen += TCPOLEN_SACK_BLOCK;
739 hstart = tp->encloseblk.rblk_start;
740 hend = tp->encloseblk.rblk_end;
742 if (SEQ_GT(hstart, tp->rcv_nxt))
743 tcp_sack_update_reported_history(tp, hstart, hend);
745 if (tcp_do_smartsack && (tp->t_flags & TF_SACKLEFT)) {
746 /* Fill in from left! Walk re-assembly queue. */
749 q = LIST_FIRST(&tp->t_segq);
751 TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
752 *lp++ = htonl(q->tqe_th->th_seq);
753 *lp++ = htonl(TCP_SACK_BLKEND(
754 q->tqe_th->th_seq + q->tqe_len,
755 q->tqe_th->th_flags));
756 optlen += TCPOLEN_SACK_BLOCK;
757 q = LIST_NEXT(q, tqe_q);
762 /* Fill in SACK blocks from right side. */
763 while (n < tp->nsackhistory &&
764 TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
765 if (tp->sackhistory[n].rblk_start != hstart) {
766 *lp++ = htonl(tp->sackhistory[n].rblk_start);
767 *lp++ = htonl(tp->sackhistory[n].rblk_end);
768 optlen += TCPOLEN_SACK_BLOCK;
773 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
774 tp->t_flags &= ~(TF_DUPSEG | TF_ENCLOSESEG | TF_SACKLEFT);
775 nblocks = (lp - olp - 1) / 2;
776 *olp = htonl(TCPOPT_SACK_ALIGNED |
777 (TCPOLEN_SACK + nblocks * TCPOLEN_SACK_BLOCK));