tcp/sack: Implement RFC3517bis
[dragonfly.git] / sys / netinet / tcp_sack.c
1 /*
2  * Copyright (c) 2003, 2004 Jeffrey M. Hsu.  All rights reserved.
3  * Copyright (c) 2003, 2004 The DragonFly Project.  All rights reserved.
4  *
5  * This code is derived from software contributed to The DragonFly Project
6  * by Jeffrey M. Hsu.
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 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.
19  *
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
31  * SUCH DAMAGE.
32  *
33  * $DragonFly: src/sys/netinet/tcp_sack.c,v 1.8 2008/08/15 21:37:16 nth Exp $
34  */
35
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>
45
46 #include <net/if.h>
47
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>
57
58 /*
59  * Implemented:
60  *
61  * RFC 2018
62  * RFC 2883
63  * RFC 3517
64  */
65
66 struct sackblock {
67         tcp_seq                 sblk_start;
68         tcp_seq                 sblk_end;
69         TAILQ_ENTRY(sackblock)  sblk_list;
70 };
71
72 #define MAXSAVEDBLOCKS  8                       /* per connection limit */
73
74 static int insert_block(struct scoreboard *scb,
75                         const struct raw_sackblock *raw_sb, boolean_t *update);
76 static void update_lostseq(struct scoreboard *scb, tcp_seq snd_una,
77                            u_int maxseg, int rxtthresh);
78
79 static MALLOC_DEFINE(M_SACKBLOCK, "sblk", "sackblock struct");
80
81 /*
82  * Per-tcpcb initialization.
83  */
84 void
85 tcp_sack_tcpcb_init(struct tcpcb *tp)
86 {
87         struct scoreboard *scb = &tp->scb;
88
89         scb->nblocks = 0;
90         TAILQ_INIT(&scb->sackblocks);
91         scb->lastfound = NULL;
92 }
93
94 /*
95  * Find the SACK block containing or immediately preceding "seq".
96  * The boolean result indicates whether the sequence is actually
97  * contained in the SACK block.
98  */
99 static boolean_t
100 sack_block_lookup(struct scoreboard *scb, tcp_seq seq, struct sackblock **sb)
101 {
102         struct sackblock *hint = scb->lastfound;
103         struct sackblock *cur, *last, *prev;
104
105         if (TAILQ_EMPTY(&scb->sackblocks)) {
106                 *sb = NULL;
107                 return FALSE;
108         }
109
110         if (hint == NULL) {
111                 /* No hint.  Search from start to end. */
112                 cur = TAILQ_FIRST(&scb->sackblocks);
113                 last = NULL;
114                 prev = TAILQ_LAST(&scb->sackblocks, sackblock_list);
115         } else  {
116                 if (SEQ_GEQ(seq, hint->sblk_start)) {
117                         /* Search from hint to end of list. */
118                         cur = hint;
119                         last = NULL;
120                         prev = TAILQ_LAST(&scb->sackblocks, sackblock_list);
121                 } else {
122                         /* Search from front of list to hint. */
123                         cur = TAILQ_FIRST(&scb->sackblocks);
124                         last = hint;
125                         prev = TAILQ_PREV(hint, sackblock_list, sblk_list);
126                 }
127         }
128
129         do {
130                 if (SEQ_GT(cur->sblk_end, seq)) {
131                         if (SEQ_GEQ(seq, cur->sblk_start)) {
132                                 *sb = scb->lastfound = cur;
133                                 return TRUE;
134                         } else {
135                                 *sb = scb->lastfound =
136                                     TAILQ_PREV(cur, sackblock_list, sblk_list);
137                                 return FALSE;
138                         }
139                 }
140                 cur = TAILQ_NEXT(cur, sblk_list);
141         } while (cur != last);
142
143         *sb = scb->lastfound = prev;
144         return FALSE;
145 }
146
147 /*
148  * Allocate a SACK block.
149  */
150 static __inline struct sackblock *
151 alloc_sackblock(struct scoreboard *scb, const struct raw_sackblock *raw_sb)
152 {
153         struct sackblock *sb;
154
155         if (scb->freecache != NULL) {
156                 sb = scb->freecache;
157                 scb->freecache = NULL;
158                 tcpstat.tcps_sacksbfast++;
159         } else {
160                 sb = kmalloc(sizeof(struct sackblock), M_SACKBLOCK, M_NOWAIT);
161                 if (sb == NULL) {
162                         tcpstat.tcps_sacksbfailed++;
163                         return NULL;
164                 }
165         }
166         sb->sblk_start = raw_sb->rblk_start;
167         sb->sblk_end = raw_sb->rblk_end;
168         return sb;
169 }
170
171 static __inline struct sackblock *
172 alloc_sackblock_limit(struct scoreboard *scb,
173     const struct raw_sackblock *raw_sb)
174 {
175         if (scb->nblocks == MAXSAVEDBLOCKS) {
176                 /*
177                  * Should try to kick out older blocks XXX JH
178                  * May be able to coalesce with existing block.
179                  * Or, go other way and free all blocks if we hit
180                  * this limit.
181                  */
182                 tcpstat.tcps_sacksboverflow++;
183                 return NULL;
184         }
185         return alloc_sackblock(scb, raw_sb);
186 }
187
188 /*
189  * Free a SACK block.
190  */
191 static __inline void
192 free_sackblock(struct scoreboard *scb, struct sackblock *s)
193 {
194         if (scb->freecache == NULL) {
195                 /* YYY Maybe use the latest freed block? */
196                 scb->freecache = s;
197                 return;
198         }
199         kfree(s, M_SACKBLOCK);
200 }
201
202 /*
203  * Free up SACK blocks for data that's been acked.
204  */
205 static void
206 tcp_sack_ack_blocks(struct scoreboard *scb, tcp_seq th_ack)
207 {
208         struct sackblock *sb, *nb;
209
210         sb = TAILQ_FIRST(&scb->sackblocks);
211         while (sb && SEQ_LEQ(sb->sblk_end, th_ack)) {
212                 nb = TAILQ_NEXT(sb, sblk_list);
213                 if (scb->lastfound == sb)
214                         scb->lastfound = NULL;
215                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
216                 free_sackblock(scb, sb);
217                 --scb->nblocks;
218                 KASSERT(scb->nblocks >= 0,
219                     ("SACK block count underflow: %d < 0", scb->nblocks));
220                 sb = nb;
221         }
222         if (sb && SEQ_GT(th_ack, sb->sblk_start))
223                 sb->sblk_start = th_ack;        /* other side reneged? XXX */
224 }
225
226 /*
227  * Delete and free SACK blocks saved in scoreboard.
228  */
229 void
230 tcp_sack_cleanup(struct scoreboard *scb)
231 {
232         struct sackblock *sb, *nb;
233
234         TAILQ_FOREACH_MUTABLE(sb, &scb->sackblocks, sblk_list, nb) {
235                 free_sackblock(scb, sb);
236                 --scb->nblocks;
237         }
238         KASSERT(scb->nblocks == 0,
239             ("SACK block %d count not zero", scb->nblocks));
240         TAILQ_INIT(&scb->sackblocks);
241         scb->lastfound = NULL;
242 }
243
244 /*
245  * Delete and free SACK blocks saved in scoreboard.
246  * Delete the one slot block cache.
247  */
248 void
249 tcp_sack_destroy(struct scoreboard *scb)
250 {
251         tcp_sack_cleanup(scb);
252         if (scb->freecache != NULL) {
253                 kfree(scb->freecache, M_SACKBLOCK);
254                 scb->freecache = NULL;
255         }
256 }
257
258 /*
259  * Cleanup the reported SACK block information
260  */
261 void
262 tcp_sack_report_cleanup(struct tcpcb *tp)
263 {
264         tp->t_flags &= ~(TF_DUPSEG | TF_ENCLOSESEG | TF_SACKLEFT);
265         tp->reportblk.rblk_start = tp->reportblk.rblk_end;
266 }
267
268 /*
269  * Returns      0 if not D-SACK block,
270  *              1 if D-SACK,
271  *              2 if duplicate of out-of-order D-SACK block.
272  */
273 int
274 tcp_sack_ndsack_blocks(struct raw_sackblock *blocks, const int numblocks,
275                        tcp_seq snd_una)
276 {
277         if (numblocks == 0)
278                 return 0;
279
280         if (SEQ_LT(blocks[0].rblk_start, snd_una))
281                 return 1;
282
283         /* block 0 inside block 1 */
284         if (numblocks > 1 &&
285             SEQ_GEQ(blocks[0].rblk_start, blocks[1].rblk_start) &&
286             SEQ_LEQ(blocks[0].rblk_end, blocks[1].rblk_end))
287                 return 2;
288
289         return 0;
290 }
291
292 /*
293  * Update scoreboard on new incoming ACK.
294  */
295 static void
296 tcp_sack_add_blocks(struct tcpcb *tp, struct tcpopt *to)
297 {
298         const int numblocks = to->to_nsackblocks;
299         struct raw_sackblock *blocks = to->to_sackblocks;
300         struct scoreboard *scb = &tp->scb;
301         int startblock, i;
302
303         if (tcp_sack_ndsack_blocks(blocks, numblocks, tp->snd_una) > 0)
304                 startblock = 1;
305         else
306                 startblock = 0;
307
308         to->to_flags |= TOF_SACK_REDUNDANT;
309         for (i = startblock; i < numblocks; i++) {
310                 struct raw_sackblock *newsackblock = &blocks[i];
311                 boolean_t update;
312                 int error;
313
314                 /* Guard against ACK reordering */
315                 if (SEQ_LT(newsackblock->rblk_start, tp->snd_una))
316                         continue;
317
318                 /* Don't accept bad SACK blocks */
319                 if (SEQ_GT(newsackblock->rblk_end, tp->snd_max)) {
320                         tcpstat.tcps_rcvbadsackopt++;
321                         break;          /* skip all other blocks */
322                 }
323                 tcpstat.tcps_sacksbupdate++;
324
325                 error = insert_block(scb, newsackblock, &update);
326                 if (update)
327                         to->to_flags &= ~TOF_SACK_REDUNDANT;
328                 if (error)
329                         break;
330         }
331 }
332
333 void
334 tcp_sack_update_scoreboard(struct tcpcb *tp, struct tcpopt *to)
335 {
336         struct scoreboard *scb = &tp->scb;
337         int rexmt_high_update = 0;
338
339         tcp_sack_ack_blocks(scb, tp->snd_una);
340         tcp_sack_add_blocks(tp, to);
341         update_lostseq(scb, tp->snd_una, tp->t_maxseg, tp->t_rxtthresh);
342         if (SEQ_LT(tp->rexmt_high, tp->snd_una)) {
343                 tp->rexmt_high = tp->snd_una;
344                 rexmt_high_update = 1;
345         }
346         if (tp->t_flags & TF_SACKRESCUED) {
347                 if (SEQ_LT(tp->rexmt_rescue, tp->snd_una)) {
348                         tp->t_flags &= ~TF_SACKRESCUED;
349                 } else if (tcp_aggressive_rescuesack && rexmt_high_update &&
350                     SEQ_LT(tp->rexmt_rescue, tp->rexmt_high)) {
351                         /* Drag RescueRxt along with HighRxt */
352                         tp->rexmt_rescue = tp->rexmt_high;
353                 }
354         }
355 }
356
357 /*
358  * Insert SACK block into sender's scoreboard.
359  */
360 static int
361 insert_block(struct scoreboard *scb, const struct raw_sackblock *raw_sb,
362     boolean_t *update)
363 {
364         struct sackblock *sb, *workingblock;
365         boolean_t overlap_front;
366
367         *update = TRUE;
368         if (TAILQ_EMPTY(&scb->sackblocks)) {
369                 struct sackblock *newblock;
370
371                 KASSERT(scb->nblocks == 0, ("emply scb w/ blocks"));
372
373                 newblock = alloc_sackblock(scb, raw_sb);
374                 if (newblock == NULL)
375                         return ENOMEM;
376                 TAILQ_INSERT_HEAD(&scb->sackblocks, newblock, sblk_list);
377                 scb->nblocks = 1;
378                 return 0;
379         }
380
381         KASSERT(scb->nblocks > 0, ("insert_block() called w/ no blocks"));
382         KASSERT(scb->nblocks <= MAXSAVEDBLOCKS,
383             ("too many SACK blocks %d", scb->nblocks));
384
385         overlap_front = sack_block_lookup(scb, raw_sb->rblk_start, &sb);
386
387         if (sb == NULL) {
388                 workingblock = alloc_sackblock_limit(scb, raw_sb);
389                 if (workingblock == NULL)
390                         return ENOMEM;
391                 TAILQ_INSERT_HEAD(&scb->sackblocks, workingblock, sblk_list);
392                 ++scb->nblocks;
393         } else {
394                 if (overlap_front || sb->sblk_end == raw_sb->rblk_start) {
395                         /* Extend old block */
396                         workingblock = sb;
397                         if (SEQ_GT(raw_sb->rblk_end, sb->sblk_end))
398                                 sb->sblk_end = raw_sb->rblk_end;
399                         else
400                                 *update = FALSE;
401                         tcpstat.tcps_sacksbreused++;
402                 } else {
403                         workingblock = alloc_sackblock_limit(scb, raw_sb);
404                         if (workingblock == NULL)
405                                 return ENOMEM;
406                         TAILQ_INSERT_AFTER(&scb->sackblocks, sb, workingblock,
407                             sblk_list);
408                         ++scb->nblocks;
409                 }
410         }
411
412         /* Consolidate right-hand side. */
413         sb = TAILQ_NEXT(workingblock, sblk_list);
414         while (sb != NULL &&
415             SEQ_GEQ(workingblock->sblk_end, sb->sblk_end)) {
416                 struct sackblock *nextblock;
417
418                 nextblock = TAILQ_NEXT(sb, sblk_list);
419                 if (scb->lastfound == sb)
420                         scb->lastfound = NULL;
421                 /* Remove completely overlapped block */
422                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
423                 free_sackblock(scb, sb);
424                 --scb->nblocks;
425                 KASSERT(scb->nblocks > 0,
426                     ("removed overlapped block: %d blocks left", scb->nblocks));
427                 sb = nextblock;
428         }
429         if (sb != NULL &&
430             SEQ_GEQ(workingblock->sblk_end, sb->sblk_start)) {
431                 /* Extend new block to cover partially overlapped old block. */
432                 workingblock->sblk_end = sb->sblk_end;
433                 if (scb->lastfound == sb)
434                         scb->lastfound = NULL;
435                 TAILQ_REMOVE(&scb->sackblocks, sb, sblk_list);
436                 free_sackblock(scb, sb);
437                 --scb->nblocks;
438                 KASSERT(scb->nblocks > 0,
439                     ("removed partial right: %d blocks left", scb->nblocks));
440         }
441         return 0;
442 }
443
444 #ifdef DEBUG_SACK_BLOCKS
445 static void
446 tcp_sack_dump_blocks(struct scoreboard *scb)
447 {
448         struct sackblock *sb;
449
450         kprintf("%d blocks:", scb->nblocks);
451         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
452                 kprintf(" [%u, %u)", sb->sblk_start, sb->sblk_end);
453         kprintf("\n");
454 }
455 #else
456 static __inline void
457 tcp_sack_dump_blocks(struct scoreboard *scb)
458 {
459 }
460 #endif
461
462 /*
463  * Optimization to quickly determine which packets are lost.
464  */
465 static void
466 update_lostseq(struct scoreboard *scb, tcp_seq snd_una, u_int maxseg,
467     int rxtthresh)
468 {
469         struct sackblock *sb;
470         int nsackblocks = 0;
471         int bytes_sacked = 0;
472         int rxtthresh_bytes;
473
474         if (tcp_do_rfc3517bis)
475                 rxtthresh_bytes = (rxtthresh - 1) * maxseg;
476         else
477                 rxtthresh_bytes = rxtthresh * maxseg;
478
479         sb = TAILQ_LAST(&scb->sackblocks, sackblock_list);
480         while (sb != NULL) {
481                 ++nsackblocks;
482                 bytes_sacked += sb->sblk_end - sb->sblk_start;
483                 if (nsackblocks == rxtthresh ||
484                     bytes_sacked >= rxtthresh_bytes) {
485                         scb->lostseq = sb->sblk_start;
486                         return;
487                 }
488                 sb = TAILQ_PREV(sb, sackblock_list, sblk_list);
489         }
490         scb->lostseq = snd_una;
491 }
492
493 /*
494  * Return whether the given sequence number is considered lost.
495  */
496 boolean_t
497 tcp_sack_islost(struct scoreboard *scb, tcp_seq seqnum)
498 {
499         return SEQ_LT(seqnum, scb->lostseq);
500 }
501
502 /*
503  * True if at least "amount" has been SACKed.  Used by Early Retransmit.
504  */
505 boolean_t
506 tcp_sack_has_sacked(struct scoreboard *scb, u_int amount)
507 {
508         struct sackblock *sb;
509         int bytes_sacked = 0;
510
511         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
512                 bytes_sacked += sb->sblk_end - sb->sblk_start;
513                 if (bytes_sacked >= amount)
514                         return TRUE;
515         }
516         return FALSE;
517 }
518
519 /*
520  * Number of bytes SACKed below seq.
521  */
522 int
523 tcp_sack_bytes_below(struct scoreboard *scb, tcp_seq seq)
524 {
525         struct sackblock *sb;
526         int bytes_sacked = 0;
527
528         sb = TAILQ_FIRST(&scb->sackblocks);
529         while (sb && SEQ_GT(seq, sb->sblk_start)) {
530                 bytes_sacked += seq_min(seq, sb->sblk_end) - sb->sblk_start;
531                 sb = TAILQ_NEXT(sb, sblk_list);
532         }
533         return bytes_sacked;
534 }
535
536 /*
537  * Return estimate of the number of bytes outstanding in the network.
538  */
539 uint32_t
540 tcp_sack_compute_pipe(struct tcpcb *tp)
541 {
542         struct scoreboard *scb = &tp->scb;
543         struct sackblock *sb;
544         int nlost, nretransmitted;
545         tcp_seq end;
546
547         nlost = tp->snd_max - scb->lostseq;
548         nretransmitted = tp->rexmt_high - tp->snd_una;
549
550         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list) {
551                 if (SEQ_LT(sb->sblk_start, tp->rexmt_high)) {
552                         end = seq_min(sb->sblk_end, tp->rexmt_high);
553                         nretransmitted -= end - sb->sblk_start;
554                 }
555                 if (SEQ_GEQ(sb->sblk_start, scb->lostseq))
556                         nlost -= sb->sblk_end - sb->sblk_start;
557         }
558
559         return (nlost + nretransmitted);
560 }
561
562 /*
563  * Return the sequence number and length of the next segment to transmit
564  * when in Fast Recovery.
565  */
566 boolean_t
567 tcp_sack_nextseg(struct tcpcb *tp, tcp_seq *nextrexmt, uint32_t *plen,
568     boolean_t *rescue)
569 {
570         struct scoreboard *scb = &tp->scb;
571         struct socket *so = tp->t_inpcb->inp_socket;
572         struct sackblock *sb;
573         const struct sackblock *lastblock =
574             TAILQ_LAST(&scb->sackblocks, sackblock_list);
575         tcp_seq torexmt;
576         long len, off;
577
578         /* skip SACKed data */
579         tcp_sack_skip_sacked(scb, &tp->rexmt_high);
580
581         /* Look for lost data. */
582         torexmt = tp->rexmt_high;
583         *rescue = FALSE;
584         if (lastblock != NULL) {
585                 if (SEQ_LT(torexmt, lastblock->sblk_end) &&
586                     tcp_sack_islost(scb, torexmt)) {
587 sendunsacked:
588                         *nextrexmt = torexmt;
589                         /* If the left-hand edge has been SACKed, pull it in. */
590                         if (sack_block_lookup(scb, torexmt + tp->t_maxseg, &sb))
591                                 *plen = sb->sblk_start - torexmt;
592                         else
593                                 *plen = tp->t_maxseg;
594                         return TRUE;
595                 }
596         }
597
598         /* See if unsent data available within send window. */
599         off = tp->snd_max - tp->snd_una;
600         len = (long) ulmin(so->so_snd.ssb_cc, tp->snd_wnd) - off;
601         if (len > 0) {
602                 *nextrexmt = tp->snd_max;       /* Send new data. */
603                 *plen = tp->t_maxseg;
604                 return TRUE;
605         }
606
607         /* We're less certain this data has been lost. */
608         if (lastblock != NULL && SEQ_LT(torexmt, lastblock->sblk_end))
609                 goto sendunsacked;
610
611         /* Rescue retransmission */
612         if (tcp_do_rescuesack || tcp_do_rfc3517bis) {
613                 tcpstat.tcps_sackrescue_try++;
614                 if (tp->t_flags & TF_SACKRESCUED) {
615                         if (!tcp_aggressive_rescuesack)
616                                 return FALSE;
617
618                         /*
619                          * Aggressive variant of the rescue retransmission.
620                          *
621                          * The idea of the rescue retransmission is to sustain
622                          * the ACK clock thus to avoid timeout retransmission.
623                          *
624                          * Under some situations, the conservative approach
625                          * suggested in the draft
626                          * http://tools.ietf.org/html/
627                          * draft-nishida-tcpm-rescue-retransmission-00
628                          * could not sustain ACK clock, since it only allows
629                          * one rescue retransmission before a cumulative ACK
630                          * covers the segement transmitted by rescue
631                          * retransmission.
632                          *
633                          * We try to locate the next unSACKed segment which
634                          * follows the previously sent rescue segment.  If
635                          * there is no such segment, we loop back to the first
636                          * unacknowledged segment.
637                          */
638
639                         /*
640                          * Skip SACKed data, but here we follow
641                          * the last transmitted rescue segment.
642                          */
643                         torexmt = tp->rexmt_rescue;
644                         tcp_sack_skip_sacked(scb, &torexmt);
645                         if (torexmt == tp->snd_max) {
646                                 /* Nothing left to retransmit; restart */
647                                 torexmt = tp->snd_una;
648                         }
649                 }
650                 *rescue = TRUE;
651                 goto sendunsacked;
652         } else if (tcp_do_smartsack && lastblock == NULL) {
653                 tcpstat.tcps_sackrescue_try++;
654                 *rescue = TRUE;
655                 goto sendunsacked;
656         }
657
658         return FALSE;
659 }
660
661 /*
662  * Return the next sequence number higher than "*prexmt" that has
663  * not been SACKed.
664  */
665 void
666 tcp_sack_skip_sacked(struct scoreboard *scb, tcp_seq *prexmt)
667 {
668         struct sackblock *sb;
669
670         /* skip SACKed data */
671         if (sack_block_lookup(scb, *prexmt, &sb))
672                 *prexmt = sb->sblk_end;
673 }
674
675 #ifdef later
676 void
677 tcp_sack_save_scoreboard(struct scoreboard *scb)
678 {
679         struct scoreboard *scb = &tp->scb;
680
681         scb->sackblocks_prev = scb->sackblocks;
682         TAILQ_INIT(&scb->sackblocks);
683 }
684
685 void
686 tcp_sack_revert_scoreboard(struct scoreboard *scb, tcp_seq snd_una,
687                            u_int maxseg)
688 {
689         struct sackblock *sb;
690
691         scb->sackblocks = scb->sackblocks_prev;
692         scb->nblocks = 0;
693         TAILQ_FOREACH(sb, &scb->sackblocks, sblk_list)
694                 ++scb->nblocks;
695         tcp_sack_ack_blocks(scb, snd_una);
696         scb->lastfound = NULL;
697 }
698 #endif
699
700 #ifdef DEBUG_SACK_HISTORY
701 static void
702 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
703 {
704         int i;
705         static int ndumped;
706
707         /* only need a couple of these to debug most problems */
708         if (++ndumped > 900)
709                 return;
710
711         kprintf("%s:\tnsackhistory %d: ", msg, tp->nsackhistory);
712         for (i = 0; i < tp->nsackhistory; ++i)
713                 kprintf("[%u, %u) ", tp->sackhistory[i].rblk_start,
714                     tp->sackhistory[i].rblk_end);
715         kprintf("\n");
716 }
717 #else
718 static __inline void
719 tcp_sack_dump_history(char *msg, struct tcpcb *tp)
720 {
721 }
722 #endif
723
724 /*
725  * Remove old SACK blocks from the SACK history that have already been ACKed.
726  */
727 static void
728 tcp_sack_ack_history(struct tcpcb *tp)
729 {
730         int i, nblocks, openslot;
731
732         tcp_sack_dump_history("before tcp_sack_ack_history", tp);
733         nblocks = tp->nsackhistory;
734         for (i = openslot = 0; i < nblocks; ++i) {
735                 if (SEQ_LEQ(tp->sackhistory[i].rblk_end, tp->rcv_nxt)) {
736                         --tp->nsackhistory;
737                         continue;
738                 }
739                 if (SEQ_LT(tp->sackhistory[i].rblk_start, tp->rcv_nxt))
740                         tp->sackhistory[i].rblk_start = tp->rcv_nxt;
741                 if (i == openslot)
742                         ++openslot;
743                 else
744                         tp->sackhistory[openslot++] = tp->sackhistory[i];
745         }
746         tcp_sack_dump_history("after tcp_sack_ack_history", tp);
747         KASSERT(openslot == tp->nsackhistory,
748             ("tcp_sack_ack_history miscounted: %d != %d",
749             openslot, tp->nsackhistory));
750 }
751
752 /*
753  * Add or merge newblock into reported history.
754  * Also remove or update SACK blocks that will be acked.
755  */
756 static void
757 tcp_sack_update_reported_history(struct tcpcb *tp, tcp_seq start, tcp_seq end)
758 {
759         struct raw_sackblock copy[MAX_SACK_REPORT_BLOCKS];
760         int i, cindex;
761
762         tcp_sack_dump_history("before tcp_sack_update_reported_history", tp);
763         /*
764          * Six cases:
765          *      0) no overlap
766          *      1) newblock == oldblock
767          *      2) oldblock contains newblock
768          *      3) newblock contains oldblock
769          *      4) tail of oldblock overlaps or abuts start of newblock
770          *      5) tail of newblock overlaps or abuts head of oldblock
771          */
772         for (i = cindex = 0; i < tp->nsackhistory; ++i) {
773                 struct raw_sackblock *oldblock = &tp->sackhistory[i];
774                 tcp_seq old_start = oldblock->rblk_start;
775                 tcp_seq old_end = oldblock->rblk_end;
776
777                 if (SEQ_LT(end, old_start) || SEQ_GT(start, old_end)) {
778                         /* Case 0:  no overlap.  Copy old block. */
779                         copy[cindex++] = *oldblock;
780                         continue;
781                 }
782
783                 if (SEQ_GEQ(start, old_start) && SEQ_LEQ(end, old_end)) {
784                         /* Cases 1 & 2.  Move block to front of history. */
785                         int j;
786
787                         start = old_start;
788                         end = old_end;
789                         /* no need to check rest of blocks */
790                         for (j = i + 1; j < tp->nsackhistory; ++j)
791                                 copy[cindex++] = tp->sackhistory[j];
792                         break;
793                 }
794
795                 if (SEQ_GEQ(old_end, start) && SEQ_LT(old_start, start)) {
796                         /* Case 4:  extend start of new block. */
797                         start = old_start;
798                 } else if (SEQ_GEQ(end, old_start) && SEQ_GT(old_end, end)) {
799                         /* Case 5: extend end of new block */
800                         end = old_end;
801                 } else {
802                         /* Case 3.  Delete old block by not copying it. */
803                         KASSERT(SEQ_LEQ(start, old_start) &&
804                                 SEQ_GEQ(end, old_end),
805                             ("bad logic: old [%u, %u), new [%u, %u)",
806                              old_start, old_end, start, end));
807                 }
808         }
809
810         /* insert new block */
811         tp->sackhistory[0].rblk_start = start;
812         tp->sackhistory[0].rblk_end = end;
813         cindex = min(cindex, MAX_SACK_REPORT_BLOCKS - 1);
814         for (i = 0; i < cindex; ++i)
815                 tp->sackhistory[i + 1] = copy[i];
816         tp->nsackhistory = cindex + 1;
817         tcp_sack_dump_history("after tcp_sack_update_reported_history", tp);
818 }
819
820 /*
821  * Fill in SACK report to return to data sender.
822  */
823 void
824 tcp_sack_fill_report(struct tcpcb *tp, u_char *opt, u_int *plen)
825 {
826         u_int optlen = *plen;
827         uint32_t *lp = (uint32_t *)(opt + optlen);
828         uint32_t *olp;
829         tcp_seq hstart = tp->rcv_nxt, hend;
830         int nblocks;
831
832         KASSERT(TCP_MAXOLEN - optlen >=
833             TCPOLEN_SACK_ALIGNED + TCPOLEN_SACK_BLOCK,
834             ("no room for SACK header and one block: optlen %d", optlen));
835
836         if (tp->t_flags & TF_DUPSEG)
837                 tcpstat.tcps_snddsackopt++;
838         else
839                 tcpstat.tcps_sndsackopt++;
840
841         olp = lp++;
842         optlen += TCPOLEN_SACK_ALIGNED;
843
844         tcp_sack_ack_history(tp);
845         if (tp->reportblk.rblk_start != tp->reportblk.rblk_end) {
846                 *lp++ = htonl(tp->reportblk.rblk_start);
847                 *lp++ = htonl(tp->reportblk.rblk_end);
848                 optlen += TCPOLEN_SACK_BLOCK;
849                 hstart = tp->reportblk.rblk_start;
850                 hend = tp->reportblk.rblk_end;
851                 if (tp->t_flags & TF_ENCLOSESEG) {
852                         KASSERT(TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK,
853                             ("no room for enclosing SACK block: oplen %d",
854                             optlen));
855                         *lp++ = htonl(tp->encloseblk.rblk_start);
856                         *lp++ = htonl(tp->encloseblk.rblk_end);
857                         optlen += TCPOLEN_SACK_BLOCK;
858                         hstart = tp->encloseblk.rblk_start;
859                         hend = tp->encloseblk.rblk_end;
860                 }
861                 if (SEQ_GT(hstart, tp->rcv_nxt))
862                         tcp_sack_update_reported_history(tp, hstart, hend);
863         }
864         if (tcp_do_smartsack && (tp->t_flags & TF_SACKLEFT)) {
865                 /* Fill in from left!  Walk re-assembly queue. */
866                 struct tseg_qent *q;
867
868                 q = LIST_FIRST(&tp->t_segq);
869                 while (q != NULL &&
870                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
871                         *lp++ = htonl(q->tqe_th->th_seq);
872                         *lp++ = htonl(TCP_SACK_BLKEND(
873                             q->tqe_th->th_seq + q->tqe_len,
874                             q->tqe_th->th_flags));
875                         optlen += TCPOLEN_SACK_BLOCK;
876                         q = LIST_NEXT(q, tqe_q);
877                 }
878         } else {
879                 int n = 0;
880
881                 /* Fill in SACK blocks from right side. */
882                 while (n < tp->nsackhistory &&
883                     TCP_MAXOLEN - optlen >= TCPOLEN_SACK_BLOCK) {
884                         if (tp->sackhistory[n].rblk_start != hstart) {
885                                 *lp++ = htonl(tp->sackhistory[n].rblk_start);
886                                 *lp++ = htonl(tp->sackhistory[n].rblk_end);
887                                 optlen += TCPOLEN_SACK_BLOCK;
888                         }
889                         ++n;
890                 }
891         }
892         tp->reportblk.rblk_start = tp->reportblk.rblk_end;
893         tp->t_flags &= ~(TF_DUPSEG | TF_ENCLOSESEG | TF_SACKLEFT);
894         nblocks = (lp - olp - 1) / 2;
895         *olp = htonl(TCPOPT_SACK_ALIGNED |
896                      (TCPOLEN_SACK + nblocks * TCPOLEN_SACK_BLOCK));
897         *plen = optlen;
898 }