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(void)
152 return (kmalloc(sizeof(struct sackblock), M_SACKBLOCK, M_NOWAIT));
159 free_sackblock(struct sackblock *s)
161 kfree(s, M_SACKBLOCK);
165 * Free up SACK blocks for data that's been acked.
168 tcp_sack_ack_blocks(struct scoreboard *scb, tcp_seq th_ack)
170 struct sackblock *sb, *nb;
172 sb = TAILQ_FIRST(&scb->sackblocks);
173 while (sb && SEQ_LEQ(sb->sblk_end, th_ack)) {
174 nb = TAILQ_NEXT(sb, sblk_list);
175 if (scb->lastfound == sb)
176 scb->lastfound = NULL;
177 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
180 KASSERT(scb->nblocks >= 0,
181 ("SACK block count underflow: %d < 0", scb->nblocks));
184 if (sb && SEQ_GT(th_ack, sb->sblk_start))
185 sb->sblk_start = th_ack; /* other side reneged? XXX */
189 * Delete and free SACK blocks saved in scoreboard.
192 tcp_sack_cleanup(struct scoreboard *scb)
194 struct sackblock *sb, *nb;
196 TAILQ_FOREACH_MUTABLE(sb, &scb->sackblocks, sblk_list, nb) {
200 KASSERT(scb->nblocks == 0,
201 ("SACK block %d count not zero", scb->nblocks));
202 TAILQ_INIT(&scb->sackblocks);
203 scb->lastfound = NULL;
207 * Returns 0 if not D-SACK block,
209 * 2 if duplicate of out-of-order D-SACK block.
212 tcp_sack_ndsack_blocks(struct raw_sackblock *blocks, const int numblocks,
218 if (SEQ_LT(blocks[0].rblk_start, snd_una))
221 /* block 0 inside block 1 */
223 SEQ_GEQ(blocks[0].rblk_start, blocks[1].rblk_start) &&
224 SEQ_LEQ(blocks[0].rblk_end, blocks[1].rblk_end))
231 * Update scoreboard on new incoming ACK.
234 tcp_sack_add_blocks(struct tcpcb *tp, struct tcpopt *to)
236 const int numblocks = to->to_nsackblocks;
237 struct raw_sackblock *blocks = to->to_sackblocks;
238 struct scoreboard *scb = &tp->scb;
239 struct sackblock *sb;
243 if (tcp_sack_ndsack_blocks(blocks, numblocks, tp->snd_una) > 0)
248 for (i = startblock; i < numblocks; i++) {
249 struct raw_sackblock *newsackblock = &blocks[i];
251 /* don't accept bad SACK blocks */
252 if (SEQ_GT(newsackblock->rblk_end, tp->snd_max)) {
253 tcpstat.tcps_rcvbadsackopt++;
254 break; /* skip all other blocks */
256 tcpstat.tcps_sacksbupdate++;
258 sb = alloc_sackblock();
259 if (sb == NULL) { /* do some sort of cleanup? XXX */
260 tcpstat.tcps_sacksbfailed++;
261 break; /* just skip rest of blocks */
263 sb->sblk_start = newsackblock->rblk_start;
264 sb->sblk_end = newsackblock->rblk_end;
265 if (TAILQ_EMPTY(&scb->sackblocks)) {
266 KASSERT(scb->nblocks == 0, ("emply scb w/ blocks"));
268 TAILQ_INSERT_HEAD(&scb->sackblocks, sb, sblk_list);
270 insert_block(scb, sb);
276 tcp_sack_update_scoreboard(struct tcpcb *tp, struct tcpopt *to)
278 struct scoreboard *scb = &tp->scb;
280 tcp_sack_ack_blocks(scb, tp->snd_una);
281 tcp_sack_add_blocks(tp, to);
282 update_lostseq(scb, tp->snd_una, tp->t_maxseg);
283 if (SEQ_LT(tp->rexmt_high, tp->snd_una))
284 tp->rexmt_high = tp->snd_una;
288 * Insert SACK block into sender's scoreboard.
291 insert_block(struct scoreboard *scb, struct sackblock *newblock)
293 struct sackblock *sb, *workingblock;
294 boolean_t overlap_front;
296 KASSERT(scb->nblocks > 0, ("insert_block() called w/ no blocks"));
298 if (scb->nblocks == MAXSAVEDBLOCKS) {
300 * Should try to kick out older blocks XXX JH
301 * May be able to coalesce with existing block.
302 * Or, go other way and free all blocks if we hit this limit.
304 free_sackblock(newblock);
305 tcpstat.tcps_sacksboverflow++;
308 KASSERT(scb->nblocks < MAXSAVEDBLOCKS,
309 ("too many SACK blocks %d", scb->nblocks));
311 overlap_front = sack_block_lookup(scb, newblock->sblk_start, &sb);
314 workingblock = newblock;
315 TAILQ_INSERT_HEAD(&scb->sackblocks, newblock, sblk_list);
318 if (overlap_front || sb->sblk_end == newblock->sblk_start) {
319 /* extend old block and discard new one */
321 if (SEQ_GT(newblock->sblk_end, sb->sblk_end))
322 sb->sblk_end = newblock->sblk_end;
323 free_sackblock(newblock);
324 tcpstat.tcps_sacksbreused++;
326 workingblock = newblock;
327 TAILQ_INSERT_AFTER(&scb->sackblocks, sb, newblock,
333 /* Consolidate right-hand side. */
334 sb = TAILQ_NEXT(workingblock, sblk_list);
336 SEQ_GEQ(workingblock->sblk_end, sb->sblk_end)) {
337 struct sackblock *nextblock;
339 nextblock = TAILQ_NEXT(sb, sblk_list);
340 if (scb->lastfound == sb)
341 scb->lastfound = NULL;
342 /* Remove completely overlapped block */
343 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
346 KASSERT(scb->nblocks > 0,
347 ("removed overlapped block: %d blocks left", scb->nblocks));
351 SEQ_GEQ(workingblock->sblk_end, sb->sblk_start)) {
352 /* Extend new block to cover partially overlapped old block. */
353 workingblock->sblk_end = sb->sblk_end;
354 if (scb->lastfound == sb)
355 scb->lastfound = NULL;
356 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
359 KASSERT(scb->nblocks > 0,
360 ("removed partial right: %d blocks left", scb->nblocks));
364 #ifdef DEBUG_SACK_BLOCKS
366 tcp_sack_dump_blocks(struct scoreboard *scb)
368 struct sackblock *sb;
370 kprintf("%d blocks:", scb->nblocks);
371 TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
372 kprintf(" [%u, %u)", sb->sblk_start, sb->sblk_end);
377 tcp_sack_dump_blocks(struct scoreboard *scb)
383 * Optimization to quickly determine which packets are lost.
386 update_lostseq(struct scoreboard *scb, tcp_seq snd_una, u_int maxseg)
388 struct sackblock *sb;
390 int bytes_sacked = 0;
392 sb = TAILQ_LAST(&scb->sackblocks, sackblock_list);
395 bytes_sacked += sb->sblk_end - sb->sblk_start;
396 if (nsackblocks == tcprexmtthresh ||
397 bytes_sacked >= tcprexmtthresh * maxseg) {
398 scb->lostseq = sb->sblk_start;
401 sb = TAILQ_PREV(sb, sackblock_list, sblk_list);
403 scb->lostseq = snd_una;
407 * Return whether the given sequence number is considered lost.
410 scb_islost(struct scoreboard *scb, tcp_seq seqnum)
412 return SEQ_LT(seqnum, scb->lostseq);
416 * True if at least "amount" has been SACKed. Used by Early Retransmit.
419 tcp_sack_has_sacked(struct scoreboard *scb, u_int amount)
421 struct sackblock *sb;
422 int bytes_sacked = 0;
424 TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
425 bytes_sacked += sb->sblk_end - sb->sblk_start;
426 if (bytes_sacked >= amount)
433 * Number of bytes SACKed below seq.
436 tcp_sack_bytes_below(struct scoreboard *scb, tcp_seq seq)
438 struct sackblock *sb;
439 int bytes_sacked = 0;
441 sb = TAILQ_FIRST(&scb->sackblocks);
442 while (sb && SEQ_GT(seq, sb->sblk_start)) {
443 bytes_sacked += seq_min(seq, sb->sblk_end) - sb->sblk_start;
444 sb = TAILQ_NEXT(sb, sblk_list);
450 * Return estimate of the number of bytes outstanding in the network.
453 tcp_sack_compute_pipe(struct tcpcb *tp)
455 struct scoreboard *scb = &tp->scb;
456 struct sackblock *sb;
457 int nlost, nretransmitted;
460 nlost = tp->snd_max - scb->lostseq;
461 nretransmitted = tp->rexmt_high - tp->snd_una;
463 TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
464 if (SEQ_LT(sb->sblk_start, tp->rexmt_high)) {
465 end = seq_min(sb->sblk_end, tp->rexmt_high);
466 nretransmitted -= end - sb->sblk_start;
468 if (SEQ_GEQ(sb->sblk_start, scb->lostseq))
469 nlost -= sb->sblk_end - sb->sblk_start;
472 return (nlost + nretransmitted);
476 * Return the sequence number and length of the next segment to transmit
477 * when in Fast Recovery.
480 tcp_sack_nextseg(struct tcpcb *tp, tcp_seq *nextrexmt, uint32_t *plen,
483 struct scoreboard *scb = &tp->scb;
484 struct socket *so = tp->t_inpcb->inp_socket;
485 struct sackblock *sb;
486 const struct sackblock *lastblock =
487 TAILQ_LAST(&scb->sackblocks, sackblock_list);
491 /* skip SACKed data */
492 tcp_sack_skip_sacked(scb, &tp->rexmt_high);
494 /* Look for lost data. */
495 torexmt = tp->rexmt_high;
497 if (lastblock != NULL) {
498 if (SEQ_LT(torexmt, lastblock->sblk_end) &&
499 scb_islost(scb, torexmt)) {
501 *nextrexmt = torexmt;
502 /* If the left-hand edge has been SACKed, pull it in. */
503 if (sack_block_lookup(scb, torexmt + tp->t_maxseg, &sb))
504 *plen = sb->sblk_start - torexmt;
506 *plen = tp->t_maxseg;
511 /* See if unsent data available within send window. */
512 off = tp->snd_max - tp->snd_una;
513 len = (long) ulmin(so->so_snd.ssb_cc, tp->snd_wnd) - off;
515 *nextrexmt = tp->snd_max; /* Send new data. */
516 *plen = tp->t_maxseg;
520 /* We're less certain this data has been lost. */
521 if (lastblock == NULL || SEQ_LT(torexmt, lastblock->sblk_end))
528 * Return the next sequence number higher than "*prexmt" that has
532 tcp_sack_skip_sacked(struct scoreboard *scb, tcp_seq *prexmt)
534 struct sackblock *sb;
536 /* skip SACKed data */
537 if (sack_block_lookup(scb, *prexmt, &sb))
538 *prexmt = sb->sblk_end;
543 tcp_sack_save_scoreboard(struct scoreboard *scb)
545 struct scoreboard *scb = &tp->scb;
547 scb->sackblocks_prev = scb->sackblocks;
548 TAILQ_INIT(&scb->sackblocks);
552 tcp_sack_revert_scoreboard(struct scoreboard *scb, tcp_seq snd_una,
555 struct sackblock *sb;
557 scb->sackblocks = scb->sackblocks_prev;
559 TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
561 tcp_sack_ack_blocks(scb, snd_una);
562 scb->lastfound = NULL;
566 #ifdef DEBUG_SACK_HISTORY
568 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
573 /* only need a couple of these to debug most problems */
577 kprintf("%s:\tnsackhistory %d: ", msg, tp->nsackhistory);
578 for (i = 0; i < tp->nsackhistory; ++i)
579 kprintf("[%u, %u) ", tp->sackhistory[i].rblk_start,
580 tp->sackhistory[i].rblk_end);
585 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
591 * Remove old SACK blocks from the SACK history that have already been ACKed.
594 tcp_sack_ack_history(struct tcpcb *tp)
596 int i, nblocks, openslot;
598 tcp_sack_dump_history("before tcp_sack_ack_history", tp);
599 nblocks = tp->nsackhistory;
600 for (i = openslot = 0; i < nblocks; ++i) {
601 if (SEQ_LEQ(tp->sackhistory[i].rblk_end, tp->rcv_nxt)) {
605 if (SEQ_LT(tp->sackhistory[i].rblk_start, tp->rcv_nxt))
606 tp->sackhistory[i].rblk_start = tp->rcv_nxt;
610 tp->sackhistory[openslot++] = tp->sackhistory[i];
612 tcp_sack_dump_history("after tcp_sack_ack_history", tp);
613 KASSERT(openslot == tp->nsackhistory,
614 ("tcp_sack_ack_history miscounted: %d != %d",
615 openslot, tp->nsackhistory));
619 * Add or merge newblock into reported history.
620 * Also remove or update SACK blocks that will be acked.
623 tcp_sack_update_reported_history(struct tcpcb *tp, tcp_seq start, tcp_seq end)
625 struct raw_sackblock copy[MAX_SACK_REPORT_BLOCKS];
628 tcp_sack_dump_history("before tcp_sack_update_reported_history", tp);
632 * 1) newblock == oldblock
633 * 2) oldblock contains newblock
634 * 3) newblock contains oldblock
635 * 4) tail of oldblock overlaps or abuts start of newblock
636 * 5) tail of newblock overlaps or abuts head of oldblock
638 for (i = cindex = 0; i < tp->nsackhistory; ++i) {
639 struct raw_sackblock *oldblock = &tp->sackhistory[i];
640 tcp_seq old_start = oldblock->rblk_start;
641 tcp_seq old_end = oldblock->rblk_end;
643 if (SEQ_LT(end, old_start) || SEQ_GT(start, old_end)) {
644 /* Case 0: no overlap. Copy old block. */
645 copy[cindex++] = *oldblock;
649 if (SEQ_GEQ(start, old_start) && SEQ_LEQ(end, old_end)) {
650 /* Cases 1 & 2. Move block to front of history. */
655 /* no need to check rest of blocks */
656 for (j = i + 1; j < tp->nsackhistory; ++j)
657 copy[cindex++] = tp->sackhistory[j];
661 if (SEQ_GEQ(old_end, start) && SEQ_LT(old_start, start)) {
662 /* Case 4: extend start of new block. */
664 } else if (SEQ_GEQ(end, old_start) && SEQ_GT(old_end, end)) {
665 /* Case 5: extend end of new block */
668 /* Case 3. Delete old block by not copying it. */
669 KASSERT(SEQ_LEQ(start, old_start) &&
670 SEQ_GEQ(end, old_end),
671 ("bad logic: old [%u, %u), new [%u, %u)",
672 old_start, old_end, start, end));
676 /* insert new block */
677 tp->sackhistory[0].rblk_start = start;
678 tp->sackhistory[0].rblk_end = end;
679 cindex = min(cindex, MAX_SACK_REPORT_BLOCKS - 1);
680 for (i = 0; i < cindex; ++i)
681 tp->sackhistory[i + 1] = copy[i];
682 tp->nsackhistory = cindex + 1;
683 tcp_sack_dump_history("after tcp_sack_update_reported_history", tp);
687 * Fill in SACK report to return to data sender.
690 tcp_sack_fill_report(struct tcpcb *tp, u_char *opt, u_int *plen)
692 u_int optlen = *plen;
693 uint32_t *lp = (uint32_t *)(opt + optlen);
695 tcp_seq hstart = tp->rcv_nxt, hend;
698 KASSERT(TCP_MAXOLEN - optlen >=
699 TCPOLEN_SACK_ALIGNED + TCPOLEN_SACK_BLOCK,
700 ("no room for SACK header and one block: optlen %d", optlen));
702 if (tp->t_flags & TF_DUPSEG)
703 tcpstat.tcps_snddsackopt++;
705 tcpstat.tcps_sndsackopt++;
708 optlen += TCPOLEN_SACK_ALIGNED;
710 tcp_sack_ack_history(tp);
711 if (tp->reportblk.rblk_start != tp->reportblk.rblk_end) {
712 *lp++ = htonl(tp->reportblk.rblk_start);
713 *lp++ = htonl(tp->reportblk.rblk_end);
714 optlen += TCPOLEN_SACK_BLOCK;
715 hstart = tp->reportblk.rblk_start;
716 hend = tp->reportblk.rblk_end;
717 if (tp->t_flags & TF_ENCLOSESEG) {
718 KASSERT(TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK,
719 ("no room for enclosing SACK block: oplen %d",
721 *lp++ = htonl(tp->encloseblk.rblk_start);
722 *lp++ = htonl(tp->encloseblk.rblk_end);
723 optlen += TCPOLEN_SACK_BLOCK;
724 hstart = tp->encloseblk.rblk_start;
725 hend = tp->encloseblk.rblk_end;
727 if (SEQ_GT(hstart, tp->rcv_nxt))
728 tcp_sack_update_reported_history(tp, hstart, hend);
730 if (tcp_do_smartsack && (tp->t_flags & TF_SACKLEFT)) {
731 /* Fill in from left! Walk re-assembly queue. */
734 q = LIST_FIRST(&tp->t_segq);
736 TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
737 *lp++ = htonl(q->tqe_th->th_seq);
738 *lp++ = htonl(q->tqe_th->th_seq + q->tqe_len);
739 optlen += TCPOLEN_SACK_BLOCK;
740 q = LIST_NEXT(q, tqe_q);
745 /* Fill in SACK blocks from right side. */
746 while (n < tp->nsackhistory &&
747 TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
748 if (tp->sackhistory[n].rblk_start != hstart) {
749 *lp++ = htonl(tp->sackhistory[n].rblk_start);
750 *lp++ = htonl(tp->sackhistory[n].rblk_end);
751 optlen += TCPOLEN_SACK_BLOCK;
756 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
757 tp->t_flags &= ~(TF_DUPSEG | TF_ENCLOSESEG | TF_SACKLEFT);
758 nblocks = (lp - olp - 1) / 2;
759 *olp = htonl(TCPOPT_SACK_ALIGNED |
760 (TCPOLEN_SACK + nblocks * TCPOLEN_SACK_BLOCK));