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 1.1 2004/11/14 00:49:08 hsu Exp $
37 * Copyright (c) 2003, 2004 Jeffrey M. Hsu. All rights reserved.
39 * License terms: all terms for the DragonFly license above plus the following:
41 * 4. All advertising materials mentioning features or use of this software
42 * must display the following acknowledgement:
44 * This product includes software developed by Jeffrey M. Hsu
45 * for the DragonFly Project.
47 * This requirement may be waived with permission from Jeffrey Hsu.
48 * This requirement will sunset and may be removed on November 13, 2005,
49 * after which the standard DragonFly license (as shown above) will
53 #include <sys/param.h>
54 #include <sys/systm.h>
55 #include <sys/kernel.h>
56 #include <sys/malloc.h>
57 #include <sys/queue.h>
58 #include <sys/thread.h>
59 #include <sys/types.h>
60 #include <sys/socket.h>
61 #include <sys/socketvar.h>
65 #include <netinet/in.h>
66 #include <netinet/in_systm.h>
67 #include <netinet/ip.h>
68 #include <netinet/in_var.h>
69 #include <netinet/in_pcb.h>
70 #include <netinet/ip_var.h>
71 #include <netinet/tcp.h>
72 #include <netinet/tcp_seq.h>
73 #include <netinet/tcp_var.h>
75 #include <vm/vm_zone.h>
80 TAILQ_ENTRY(sackblock) sblk_list;
83 #define MAXSAVEDBLOCKS 8 /* per connection limit */
85 static void insert_block(struct scoreboard *scb, struct sackblock *newblock);
86 static void update_lostseq(struct scoreboard *scb, tcp_seq snd_una,
89 static vm_zone_t sackblock_zone;
92 * Initialize SACK module.
97 sackblock_zone = zinit("sack blocks", sizeof(struct sackblock),
98 maxsockets, ZONE_INTERRUPT, 0);
102 * Per-tcpcb initialization.
105 tcp_sack_tcpcb_init(struct tcpcb *tp)
107 struct scoreboard *scb = &tp->scb;
110 TAILQ_INIT(&scb->sackblocks);
111 scb->lastfound = NULL;
115 * Find the SACK block containing or immediately preceding "seq".
116 * The boolean result indicates whether the sequence is actually
117 * contained in the SACK block.
120 sack_block_lookup(struct scoreboard *scb, tcp_seq seq, struct sackblock **sb)
122 struct sackblock *hint = scb->lastfound;
123 struct sackblock *cur, *last, *prev;
125 if (TAILQ_EMPTY(&scb->sackblocks)) {
131 /* No hint. Search from start to end. */
132 cur = TAILQ_FIRST(&scb->sackblocks);
134 prev = TAILQ_LAST(&scb->sackblocks, sackblock_list);
136 if (SEQ_GEQ(seq, hint->sblk_start)) {
137 /* Search from hint to end of list. */
140 prev = TAILQ_LAST(&scb->sackblocks, sackblock_list);
142 /* Search from front of list to hint. */
143 cur = TAILQ_FIRST(&scb->sackblocks);
145 prev = TAILQ_PREV(hint, sackblock_list, sblk_list);
150 if (SEQ_GT(cur->sblk_end, seq)) {
151 if (SEQ_GEQ(seq, cur->sblk_start)) {
152 *sb = scb->lastfound = cur;
155 *sb = scb->lastfound =
156 TAILQ_PREV(cur, sackblock_list, sblk_list);
160 cur = TAILQ_NEXT(cur, sblk_list);
161 } while (cur != last);
163 *sb = scb->lastfound = prev;
168 * Allocate a SACK block.
170 static __inline struct sackblock *
171 alloc_sackblock(void)
173 return zalloc(sackblock_zone);
180 free_sackblock(struct sackblock *s)
182 zfree(sackblock_zone, s);
186 * Free up SACK blocks for data that's been acked.
189 tcp_sack_ack_blocks(struct scoreboard *scb, tcp_seq th_ack)
191 struct sackblock *sb, *nb;
193 sb = TAILQ_FIRST(&scb->sackblocks);
194 while (sb && SEQ_LEQ(sb->sblk_end, th_ack)) {
195 nb = TAILQ_NEXT(sb, sblk_list);
196 if (sb == scb->lastfound)
197 scb->lastfound = NULL;
198 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
201 KASSERT(scb->nblocks >= 0,
202 ("SACK block count underflow: %d < 0", scb->nblocks));
205 if (sb && SEQ_GT(th_ack, sb->sblk_start))
206 sb->sblk_start = th_ack; /* other side reneged? XXX */
210 * Delete and free SACK blocks saved in scoreboard.
213 tcp_sack_cleanup(struct scoreboard *scb)
215 struct sackblock *sb, *nb;
217 TAILQ_FOREACH_MUTABLE(sb, &scb->sackblocks, sblk_list, nb) {
221 KASSERT(scb->nblocks == 0,
222 ("SACK block %d count not zero", scb->nblocks));
223 TAILQ_INIT(&scb->sackblocks);
224 scb->lastfound = NULL;
228 * Returns 0 if not D-SACK block,
230 * 2 if duplicate of out-of-order D-SACK block.
233 tcp_sack_ndsack_blocks(struct raw_sackblock *blocks, const int numblocks,
239 if (SEQ_LT(blocks[0].rblk_start, snd_una))
242 /* block 0 inside block 1 */
244 SEQ_GEQ(blocks[0].rblk_start, blocks[1].rblk_start) &&
245 SEQ_LEQ(blocks[0].rblk_end, blocks[1].rblk_end))
252 * Update scoreboard on new incoming ACK.
255 tcp_sack_add_blocks(struct tcpcb *tp, struct tcpopt *to)
257 const int numblocks = to->to_nsackblocks;
258 struct raw_sackblock *blocks = to->to_sackblocks;
259 struct scoreboard *scb = &tp->scb;
260 struct sackblock *sb;
264 if (tcp_sack_ndsack_blocks(blocks, numblocks, tp->snd_una) > 0)
269 for (i = startblock; i < numblocks; i++) {
270 struct raw_sackblock *newsackblock = &blocks[i];
272 /* don't accept bad SACK blocks */
273 if (SEQ_GT(newsackblock->rblk_end, tp->snd_max))
274 break; /* skip all other blocks */
276 sb = alloc_sackblock();
277 if (sb == NULL) /* do some sort of cleanup? XXX */
278 break; /* just skip rest of blocks */
279 sb->sblk_start = newsackblock->rblk_start;
280 sb->sblk_end = newsackblock->rblk_end;
281 if (TAILQ_EMPTY(&scb->sackblocks)) {
282 KASSERT(scb->nblocks == 0, ("emply scb w/ blocks"));
284 TAILQ_INSERT_HEAD(&scb->sackblocks, sb, sblk_list);
286 insert_block(scb, sb);
292 tcp_sack_update_scoreboard(struct tcpcb *tp, struct tcpopt *to)
294 struct scoreboard *scb = &tp->scb;
296 tcp_sack_ack_blocks(scb, tp->snd_una);
297 tcp_sack_add_blocks(tp, to);
298 update_lostseq(scb, tp->snd_una, tp->t_maxseg);
299 if (SEQ_LT(tp->rexmt_high, tp->snd_una))
300 tp->rexmt_high = tp->snd_una;
304 * Insert SACK block into sender's scoreboard.
307 insert_block(struct scoreboard *scb, struct sackblock *newblock)
309 struct sackblock *sb, *workingblock;
310 boolean_t overlap_front;
312 KASSERT(scb->nblocks > 0, ("insert_block() called w/ no blocks"));
314 if (scb->nblocks == MAXSAVEDBLOCKS) {
316 * Should try to kick out older blocks XXX JH
317 * May be able to coalesce with existing block.
318 * Or, go other way and free all blocks if we hit this limit.
320 free_sackblock(newblock);
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);
340 workingblock = newblock;
341 TAILQ_INSERT_AFTER(&scb->sackblocks, sb, newblock,
347 /* Consolidate right-hand side. */
348 sb = TAILQ_NEXT(workingblock, sblk_list);
350 SEQ_GEQ(workingblock->sblk_end, sb->sblk_end)) {
351 struct sackblock *nextblock;
353 nextblock = TAILQ_NEXT(sb, sblk_list);
354 /* Remove completely overlapped block */
355 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
358 KASSERT(scb->nblocks > 0,
359 ("removed overlapped block: %d blocks left", scb->nblocks));
363 SEQ_GEQ(workingblock->sblk_end, sb->sblk_start)) {
364 /* Extend new block to cover partially overlapped old block. */
365 workingblock->sblk_end = sb->sblk_end;
366 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
369 KASSERT(scb->nblocks > 0,
370 ("removed partial right: %d blocks left", scb->nblocks));
374 #ifdef DEBUG_SACK_BLOCKS
376 tcp_sack_dump_blocks(struct scoreboard *scb)
378 struct sackblock *sb;
380 printf("%d blocks:", scb->nblocks);
381 TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
382 printf(" [%u, %u)", sb->sblk_start, sb->sblk_end);
387 tcp_sack_dump_blocks(struct scoreboard *scb)
393 * Optimization to quickly determine which packets are lost.
396 update_lostseq(struct scoreboard *scb, tcp_seq snd_una, u_int maxseg)
398 struct sackblock *sb;
400 int bytes_sacked = 0;
402 sb = TAILQ_LAST(&scb->sackblocks, sackblock_list);
405 bytes_sacked += sb->sblk_end - sb->sblk_start;
406 if (nsackblocks == tcprexmtthresh ||
407 bytes_sacked >= tcprexmtthresh * maxseg) {
408 scb->lostseq = sb->sblk_start;
411 sb = TAILQ_PREV(sb, sackblock_list, sblk_list);
413 scb->lostseq = snd_una;
417 * Return whether the given sequence number is considered lost.
420 scb_islost(struct scoreboard *scb, tcp_seq seqnum)
422 return SEQ_LT(seqnum, scb->lostseq);
426 * True if at least "amount" has been SACKed. Used by Early Retransmit.
429 tcp_sack_has_sacked(struct scoreboard *scb, u_int amount)
431 struct sackblock *sb;
432 int bytes_sacked = 0;
434 TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
435 bytes_sacked += sb->sblk_end - sb->sblk_start;
436 if (bytes_sacked >= amount)
443 * Return estimate of the number of bytes outstanding in the network.
446 tcp_sack_compute_pipe(struct tcpcb *tp)
448 struct scoreboard *scb = &tp->scb;
449 struct sackblock *sb;
450 int nlost, nretransmitted;
453 nlost = tp->snd_max - scb->lostseq;
454 nretransmitted = tp->rexmt_high - tp->snd_una;
456 TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
457 if (SEQ_LT(sb->sblk_start, tp->rexmt_high)) {
458 end = seq_min(sb->sblk_end, tp->rexmt_high);
459 nretransmitted -= end - sb->sblk_start;
461 if (SEQ_GEQ(sb->sblk_start, scb->lostseq))
462 nlost -= sb->sblk_end - sb->sblk_start;
465 return (nlost + nretransmitted);
469 * Return the sequence number and length of the next segment to transmit
470 * when in Fast Recovery.
473 tcp_sack_nextseg(struct tcpcb *tp, tcp_seq *nextrexmt, uint32_t *plen,
476 struct scoreboard *scb = &tp->scb;
477 struct socket *so = tp->t_inpcb->inp_socket;
478 struct sackblock *sb;
479 const struct sackblock *lastblock =
480 TAILQ_LAST(&scb->sackblocks, sackblock_list);
484 /* skip SACKed data */
485 tcp_sack_skip_sacked(scb, &tp->rexmt_high);
487 /* Look for lost data. */
488 torexmt = tp->rexmt_high;
490 if (lastblock != NULL) {
491 if (SEQ_LT(torexmt, lastblock->sblk_end) &&
492 scb_islost(scb, torexmt)) {
494 *nextrexmt = torexmt;
495 /* If the left-hand edge has been SACKed, pull it in. */
496 if (sack_block_lookup(scb, torexmt + tp->t_maxseg, &sb))
497 *plen = sb->sblk_start - torexmt;
499 *plen = tp->t_maxseg;
504 /* See if unsent data available within send window. */
505 off = tp->snd_max - tp->snd_una;
506 len = (long) ulmin(so->so_snd.sb_cc, tp->snd_wnd) - off;
508 *nextrexmt = tp->snd_max; /* Send new data. */
509 *plen = tp->t_maxseg;
513 /* We're less certain this data has been lost. */
514 if (lastblock == NULL || SEQ_LT(torexmt, lastblock->sblk_end))
521 * Return the next sequence number higher than "*prexmt" that has
525 tcp_sack_skip_sacked(struct scoreboard *scb, tcp_seq *prexmt)
527 struct sackblock *sb;
529 /* skip SACKed data */
530 if (sack_block_lookup(scb, *prexmt, &sb))
531 *prexmt = sb->sblk_end;
536 tcp_sack_save_scoreboard(struct scoreboard *scb)
538 struct scoreboard *scb = &tp->scb;
540 scb->sackblocks_prev = scb->sackblocks;
541 TAILQ_INIT(&scb->sackblocks);
545 tcp_sack_revert_scoreboard(struct scoreboard *scb, tcp_seq snd_una,
548 struct sackblock *sb;
550 scb->sackblocks = scb->sackblocks_prev;
552 TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
554 tcp_sack_ack_blocks(scb, snd_una);
555 scb->lastfound = NULL;
559 #ifdef DEBUG_SACK_HISTORY
561 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
566 /* only need a couple of these to debug most problems */
570 printf("%s:\tnsackhistory %d: ", msg, tp->nsackhistory);
571 for (i = 0; i < tp->nsackhistory; ++i)
572 printf("[%u, %u) ", tp->sackhistory[i].rblk_start,
573 tp->sackhistory[i].rblk_end);
578 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
584 * Remove old SACK blocks from the SACK history that have already been ACKed.
587 tcp_sack_ack_history(struct tcpcb *tp)
589 int i, nblocks, openslot;
591 tcp_sack_dump_history("before tcp_sack_ack_history", tp);
592 nblocks = tp->nsackhistory;
593 for (i = openslot = 0; i < nblocks; ++i) {
594 if (SEQ_LEQ(tp->sackhistory[i].rblk_end, tp->rcv_nxt)) {
598 if (SEQ_LT(tp->sackhistory[i].rblk_start, tp->rcv_nxt))
599 tp->sackhistory[i].rblk_start = tp->rcv_nxt;
603 tp->sackhistory[openslot++] = tp->sackhistory[i];
605 tcp_sack_dump_history("after tcp_sack_ack_history", tp);
606 KASSERT(openslot == tp->nsackhistory,
607 ("tcp_sack_ack_history miscounted: %d != %d",
608 openslot, tp->nsackhistory));
612 * Add or merge newblock into reported history.
613 * Also remove or update SACK blocks that will be acked.
616 tcp_sack_update_reported_history(struct tcpcb *tp, tcp_seq start, tcp_seq end)
618 struct raw_sackblock copy[MAX_SACK_REPORT_BLOCKS];
621 tcp_sack_dump_history("before tcp_sack_update_reported_history", tp);
625 * 1) newblock == oldblock
626 * 2) oldblock contains newblock
627 * 3) newblock contains oldblock
628 * 4) tail of oldblock overlaps or abuts start of newblock
629 * 5) tail of newblock overlaps or abuts head of oldblock
631 for (i = cindex = 0; i < tp->nsackhistory; ++i) {
632 struct raw_sackblock *oldblock = &tp->sackhistory[i];
633 tcp_seq old_start = oldblock->rblk_start;
634 tcp_seq old_end = oldblock->rblk_end;
636 if (SEQ_LT(end, old_start) || SEQ_GT(start, old_end)) {
637 /* Case 0: no overlap. Copy old block. */
638 copy[cindex++] = *oldblock;
642 if (SEQ_GEQ(start, old_start) && SEQ_LEQ(end, old_end)) {
643 /* Cases 1 & 2. Move block to front of history. */
648 /* no need to check rest of blocks */
649 for (j = i + 1; j < tp->nsackhistory; ++j)
650 copy[cindex++] = tp->sackhistory[j];
654 if (SEQ_GEQ(old_end, start) && SEQ_LT(old_start, start)) {
655 /* Case 4: extend start of new block. */
657 } else if (SEQ_GEQ(end, old_start) && SEQ_GT(old_end, end)) {
658 /* Case 5: extend end of new block */
661 /* Case 3. Delete old block by not copying it. */
662 KASSERT(SEQ_LEQ(start, old_start) &&
663 SEQ_GEQ(end, old_end),
664 ("bad logic: old [%u, %u), new [%u, %u)",
665 old_start, old_end, start, end));
669 /* insert new block */
670 tp->sackhistory[0].rblk_start = start;
671 tp->sackhistory[0].rblk_end = end;
672 cindex = min(cindex, MAX_SACK_REPORT_BLOCKS - 1);
673 for (i = 0; i < cindex; ++i)
674 tp->sackhistory[i + 1] = copy[i];
675 tp->nsackhistory = cindex + 1;
676 tcp_sack_dump_history("after tcp_sack_update_reported_history", tp);
680 * Fill in SACK report to return to data sender.
683 tcp_sack_fill_report(struct tcpcb *tp, u_char *opt, u_int *plen)
685 u_int optlen = *plen;
686 uint32_t *lp = (uint32_t *)(opt + optlen);
688 tcp_seq hstart = tp->rcv_nxt, hend;
691 KASSERT(TCP_MAXOLEN - optlen >=
692 TCPOLEN_SACK_ALIGNED + TCPOLEN_SACK_BLOCK,
693 ("no room for SACK header and one block: optlen %d", optlen));
696 optlen += TCPOLEN_SACK_ALIGNED;
698 tcp_sack_ack_history(tp);
699 if (tp->reportblk.rblk_start != tp->reportblk.rblk_end) {
700 *lp++ = htonl(tp->reportblk.rblk_start);
701 *lp++ = htonl(tp->reportblk.rblk_end);
702 optlen += TCPOLEN_SACK_BLOCK;
703 hstart = tp->reportblk.rblk_start;
704 hend = tp->reportblk.rblk_end;
705 if (tp->t_flags & TF_ENCLOSESEG) {
706 KASSERT(TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK,
707 ("no room for enclosing SACK block: oplen %d",
709 *lp++ = htonl(tp->encloseblk.rblk_start);
710 *lp++ = htonl(tp->encloseblk.rblk_end);
711 optlen += TCPOLEN_SACK_BLOCK;
712 hstart = tp->encloseblk.rblk_start;
713 hend = tp->encloseblk.rblk_end;
715 if (SEQ_GT(hstart, tp->rcv_nxt))
716 tcp_sack_update_reported_history(tp, hstart, hend);
718 if (tcp_do_smartsack && (tp->t_flags & TF_SACKLEFT)) {
719 /* Fill in from left! Walk re-assembly queue. */
722 q = LIST_FIRST(&tp->t_segq);
724 TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
725 *lp++ = htonl(q->tqe_th->th_seq);
726 *lp++ = htonl(q->tqe_th->th_seq + q->tqe_len);
727 optlen += TCPOLEN_SACK_BLOCK;
728 q = LIST_NEXT(q, tqe_q);
733 /* Fill in SACK blocks from right side. */
734 while (n < tp->nsackhistory &&
735 TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
736 if (tp->sackhistory[n].rblk_start != hstart) {
737 *lp++ = htonl(tp->sackhistory[n].rblk_start);
738 *lp++ = htonl(tp->sackhistory[n].rblk_end);
739 optlen += TCPOLEN_SACK_BLOCK;
744 tp->reportblk.rblk_start = tp->reportblk.rblk_end;
745 tp->t_flags &= ~(TF_DUPSEG | TF_ENCLOSESEG | TF_SACKLEFT);
746 nblocks = (lp - olp - 1) / 2;
747 *olp = htonl(TCPOPT_SACK_ALIGNED |
748 (TCPOLEN_SACK + nblocks * TCPOLEN_SACK_BLOCK));