1/*
2 * Copyright (c) 2004-2016 Apple Inc. All rights reserved.
3 *
4 * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
5 *
6 * This file contains Original Code and/or Modifications of Original Code
7 * as defined in and that are subject to the Apple Public Source License
8 * Version 2.0 (the 'License'). You may not use this file except in
9 * compliance with the License. The rights granted to you under the License
10 * may not be used to create, or enable the creation or redistribution of,
11 * unlawful or unlicensed copies of an Apple operating system, or to
12 * circumvent, violate, or enable the circumvention or violation of, any
13 * terms of an Apple operating system software license agreement.
14 *
15 * Please obtain a copy of the License at
16 * http://www.opensource.apple.com/apsl/ and read it before using this file.
17 *
18 * The Original Code and all software distributed under the License are
19 * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
20 * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
21 * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
22 * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
23 * Please see the License for the specific language governing rights and
24 * limitations under the License.
25 *
26 * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
27 */
28/*
29 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
30 * The Regents of the University of California. All rights reserved.
31 *
32 * Redistribution and use in source and binary forms, with or without
33 * modification, are permitted provided that the following conditions
34 * are met:
35 * 1. Redistributions of source code must retain the above copyright
36 * notice, this list of conditions and the following disclaimer.
37 * 2. Redistributions in binary form must reproduce the above copyright
38 * notice, this list of conditions and the following disclaimer in the
39 * documentation and/or other materials provided with the distribution.
40 * 3. All advertising materials mentioning features or use of this software
41 * must display the following acknowledgement:
42 * This product includes software developed by the University of
43 * California, Berkeley and its contributors.
44 * 4. Neither the name of the University nor the names of its contributors
45 * may be used to endorse or promote products derived from this software
46 * without specific prior written permission.
47 *
48 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
49 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
50 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
51 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
52 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
53 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
54 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
55 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
56 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
57 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
58 * SUCH DAMAGE.
59 *
60 */
61
62#define _IP_VHL
63
64
65#include <sys/param.h>
66#include <sys/systm.h>
67#include <sys/kernel.h>
68#include <sys/sysctl.h>
69#include <sys/mbuf.h>
70#include <sys/domain.h>
71#include <sys/protosw.h>
72#include <sys/socket.h>
73#include <sys/socketvar.h>
74
75#include <kern/zalloc.h>
76
77#include <net/route.h>
78
79#include <netinet/in.h>
80#include <netinet/in_systm.h>
81#include <netinet/ip.h>
82#include <netinet/in_pcb.h>
83#include <netinet/ip_var.h>
84#if INET6
85#include <netinet6/in6_pcb.h>
86#include <netinet/ip6.h>
87#include <netinet6/ip6_var.h>
88#endif
89#include <netinet/tcp.h>
90//#define TCPOUTFLAGS
91#include <netinet/tcp_fsm.h>
92#include <netinet/tcp_seq.h>
93#include <netinet/tcp_timer.h>
94#include <netinet/tcp_var.h>
95#include <netinet/tcpip.h>
96#include <netinet/tcp_cache.h>
97#if TCPDEBUG
98#include <netinet/tcp_debug.h>
99#endif
100#include <sys/kdebug.h>
101
102#if IPSEC
103#include <netinet6/ipsec.h>
104#endif /*IPSEC*/
105
106#include <libkern/OSAtomic.h>
107
108SYSCTL_SKMEM_TCP_INT(OID_AUTO, sack, CTLFLAG_RW | CTLFLAG_LOCKED,
109 int, tcp_do_sack, 1, "Enable/Disable TCP SACK support");
110SYSCTL_SKMEM_TCP_INT(OID_AUTO, sack_maxholes, CTLFLAG_RW | CTLFLAG_LOCKED,
111 static int, tcp_sack_maxholes, 128,
112 "Maximum number of TCP SACK holes allowed per connection");
113
114SYSCTL_SKMEM_TCP_INT(OID_AUTO, sack_globalmaxholes,
115 CTLFLAG_RW | CTLFLAG_LOCKED, static int, tcp_sack_globalmaxholes, 65536,
116 "Global maximum number of TCP SACK holes");
117
118static SInt32 tcp_sack_globalholes = 0;
119SYSCTL_INT(_net_inet_tcp, OID_AUTO, sack_globalholes, CTLFLAG_RD | CTLFLAG_LOCKED,
120 &tcp_sack_globalholes, 0,
121 "Global number of TCP SACK holes currently allocated");
122
123static int tcp_detect_reordering = 1;
124static int tcp_dsack_ignore_hw_duplicates = 0;
125
126#if (DEVELOPMENT || DEBUG)
127SYSCTL_INT(_net_inet_tcp, OID_AUTO, detect_reordering,
128 CTLFLAG_RW | CTLFLAG_LOCKED,
129 &tcp_detect_reordering, 0, "");
130
131SYSCTL_INT(_net_inet_tcp, OID_AUTO, ignore_hw_duplicates,
132 CTLFLAG_RW | CTLFLAG_LOCKED,
133 &tcp_dsack_ignore_hw_duplicates, 0, "");
134#endif /* (DEVELOPMENT || DEBUG) */
135
136extern struct zone *sack_hole_zone;
137
138#define TCP_VALIDATE_SACK_SEQ_NUMBERS(_tp_, _sb_, _ack_) \
139 (SEQ_GT((_sb_)->end, (_sb_)->start) && \
140 SEQ_GT((_sb_)->start, (_tp_)->snd_una) && \
141 SEQ_GT((_sb_)->start, (_ack_)) && \
142 SEQ_LT((_sb_)->start, (_tp_)->snd_max) && \
143 SEQ_GT((_sb_)->end, (_tp_)->snd_una) && \
144 SEQ_LEQ((_sb_)->end, (_tp_)->snd_max))
145
146/*
147 * This function is called upon receipt of new valid data (while not in header
148 * prediction mode), and it updates the ordered list of sacks.
149 */
150void
151tcp_update_sack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
152{
153 /*
154 * First reported block MUST be the most recent one. Subsequent
155 * blocks SHOULD be in the order in which they arrived at the
156 * receiver. These two conditions make the implementation fully
157 * compliant with RFC 2018.
158 */
159 struct sackblk head_blk, saved_blks[MAX_SACK_BLKS];
160 int num_head, num_saved, i;
161
162 /* SACK block for the received segment. */
163 head_blk.start = rcv_start;
164 head_blk.end = rcv_end;
165
166 /*
167 * Merge updated SACK blocks into head_blk, and
168 * save unchanged SACK blocks into saved_blks[].
169 * num_saved will have the number of the saved SACK blocks.
170 */
171 num_saved = 0;
172 for (i = 0; i < tp->rcv_numsacks; i++) {
173 tcp_seq start = tp->sackblks[i].start;
174 tcp_seq end = tp->sackblks[i].end;
175 if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
176 /*
177 * Discard this SACK block.
178 */
179 } else if (SEQ_LEQ(head_blk.start, end) &&
180 SEQ_GEQ(head_blk.end, start)) {
181 /*
182 * Merge this SACK block into head_blk.
183 * This SACK block itself will be discarded.
184 */
185 if (SEQ_GT(head_blk.start, start))
186 head_blk.start = start;
187 if (SEQ_LT(head_blk.end, end))
188 head_blk.end = end;
189 } else {
190 /*
191 * Save this SACK block.
192 */
193 saved_blks[num_saved].start = start;
194 saved_blks[num_saved].end = end;
195 num_saved++;
196 }
197 }
198
199 /*
200 * Update SACK list in tp->sackblks[].
201 */
202 num_head = 0;
203 if (SEQ_GT(head_blk.start, tp->rcv_nxt)) {
204 /*
205 * The received data segment is an out-of-order segment.
206 * Put head_blk at the top of SACK list.
207 */
208 tp->sackblks[0] = head_blk;
209 num_head = 1;
210 /*
211 * If the number of saved SACK blocks exceeds its limit,
212 * discard the last SACK block.
213 */
214 if (num_saved >= MAX_SACK_BLKS)
215 num_saved--;
216 }
217 if (num_saved > 0) {
218 /*
219 * Copy the saved SACK blocks back.
220 */
221 bcopy(saved_blks, &tp->sackblks[num_head],
222 sizeof(struct sackblk) * num_saved);
223 }
224
225 /* Save the number of SACK blocks. */
226 tp->rcv_numsacks = num_head + num_saved;
227
228 /* If we are requesting SACK recovery, reset the stretch-ack state
229 * so that connection will generate more acks after recovery and
230 * sender's cwnd will open.
231 */
232 if ((tp->t_flags & TF_STRETCHACK) != 0 && tp->rcv_numsacks > 0)
233 tcp_reset_stretch_ack(tp);
234
235#if TRAFFIC_MGT
236 if (tp->acc_iaj > 0 && tp->rcv_numsacks > 0)
237 reset_acc_iaj(tp);
238#endif /* TRAFFIC_MGT */
239}
240
241/*
242 * Delete all receiver-side SACK information.
243 */
244void
245tcp_clean_sackreport( struct tcpcb *tp)
246{
247
248 tp->rcv_numsacks = 0;
249 bzero(&tp->sackblks[0], sizeof (struct sackblk) * MAX_SACK_BLKS);
250}
251
252/*
253 * Allocate struct sackhole.
254 */
255static struct sackhole *
256tcp_sackhole_alloc(struct tcpcb *tp, tcp_seq start, tcp_seq end)
257{
258 struct sackhole *hole;
259
260 if (tp->snd_numholes >= tcp_sack_maxholes ||
261 tcp_sack_globalholes >= tcp_sack_globalmaxholes) {
262 tcpstat.tcps_sack_sboverflow++;
263 return NULL;
264 }
265
266 hole = (struct sackhole *)zalloc(sack_hole_zone);
267 if (hole == NULL)
268 return NULL;
269
270 hole->start = start;
271 hole->end = end;
272 hole->rxmit = start;
273
274 tp->snd_numholes++;
275 OSIncrementAtomic(&tcp_sack_globalholes);
276
277 return hole;
278}
279
280/*
281 * Free struct sackhole.
282 */
283static void
284tcp_sackhole_free(struct tcpcb *tp, struct sackhole *hole)
285{
286 zfree(sack_hole_zone, hole);
287
288 tp->snd_numholes--;
289 OSDecrementAtomic(&tcp_sack_globalholes);
290}
291
292/*
293 * Insert new SACK hole into scoreboard.
294 */
295static struct sackhole *
296tcp_sackhole_insert(struct tcpcb *tp, tcp_seq start, tcp_seq end,
297 struct sackhole *after)
298{
299 struct sackhole *hole;
300
301 /* Allocate a new SACK hole. */
302 hole = tcp_sackhole_alloc(tp, start, end);
303 if (hole == NULL)
304 return NULL;
305 hole->rxmit_start = tcp_now;
306 /* Insert the new SACK hole into scoreboard */
307 if (after != NULL)
308 TAILQ_INSERT_AFTER(&tp->snd_holes, after, hole, scblink);
309 else
310 TAILQ_INSERT_TAIL(&tp->snd_holes, hole, scblink);
311
312 /* Update SACK hint. */
313 if (tp->sackhint.nexthole == NULL)
314 tp->sackhint.nexthole = hole;
315
316 return(hole);
317}
318
319/*
320 * Remove SACK hole from scoreboard.
321 */
322static void
323tcp_sackhole_remove(struct tcpcb *tp, struct sackhole *hole)
324{
325 /* Update SACK hint. */
326 if (tp->sackhint.nexthole == hole)
327 tp->sackhint.nexthole = TAILQ_NEXT(hole, scblink);
328
329 /* Remove this SACK hole. */
330 TAILQ_REMOVE(&tp->snd_holes, hole, scblink);
331
332 /* Free this SACK hole. */
333 tcp_sackhole_free(tp, hole);
334}
335/*
336 * When a new ack with SACK is received, check if it indicates packet
337 * reordering. If there is packet reordering, the socket is marked and
338 * the late time offset by which the packet was reordered with
339 * respect to its closest neighboring packets is computed.
340 */
341static void
342tcp_sack_detect_reordering(struct tcpcb *tp, struct sackhole *s,
343 tcp_seq sacked_seq, tcp_seq snd_fack)
344{
345 int32_t rext = 0, reordered = 0;
346
347 /*
348 * If the SACK hole is past snd_fack, this is from new SACK
349 * information, so we can ignore it.
350 */
351 if (SEQ_GT(s->end, snd_fack))
352 return;
353 /*
354 * If there has been a retransmit timeout, then the timestamp on
355 * the SACK segment will be newer. This might lead to a
356 * false-positive. Avoid re-ordering detection in this case.
357 */
358 if (tp->t_rxtshift > 0)
359 return;
360
361 /*
362 * Detect reordering from SACK information by checking
363 * if recently sacked data was never retransmitted from this hole.
364 */
365 if (SEQ_LT(s->rxmit, sacked_seq)) {
366 reordered = 1;
367 tcpstat.tcps_avoid_rxmt++;
368 }
369
370 if (reordered) {
371 if (tcp_detect_reordering == 1 &&
372 !(tp->t_flagsext & TF_PKTS_REORDERED)) {
373 tp->t_flagsext |= TF_PKTS_REORDERED;
374 tcpstat.tcps_detect_reordering++;
375 }
376
377 tcpstat.tcps_reordered_pkts++;
378 tp->t_reordered_pkts++;
379
380 /*
381 * If reordering is seen on a connection wth ECN enabled,
382 * increment the heuristic
383 */
384 if (TCP_ECN_ENABLED(tp)) {
385 INP_INC_IFNET_STAT(tp->t_inpcb, ecn_fallback_reorder);
386 tcpstat.tcps_ecn_fallback_reorder++;
387 tcp_heuristic_ecn_aggressive(tp);
388 }
389
390 VERIFY(SEQ_GEQ(snd_fack, s->rxmit));
391
392 if (s->rxmit_start > 0) {
393 rext = timer_diff(tcp_now, 0, s->rxmit_start, 0);
394 if (rext < 0)
395 return;
396
397 /*
398 * We take the maximum reorder window to schedule
399 * DELAYFR timer as that will take care of jitter
400 * on the network path.
401 *
402 * Computing average and standard deviation seems
403 * to cause unnecessary retransmissions when there
404 * is high jitter.
405 *
406 * We set a maximum of SRTT/2 and a minimum of
407 * 10 ms on the reorder window.
408 */
409 tp->t_reorderwin = max(tp->t_reorderwin, rext);
410 tp->t_reorderwin = min(tp->t_reorderwin,
411 (tp->t_srtt >> (TCP_RTT_SHIFT - 1)));
412 tp->t_reorderwin = max(tp->t_reorderwin, 10);
413 }
414 }
415}
416
417/*
418 * Process cumulative ACK and the TCP SACK option to update the scoreboard.
419 * tp->snd_holes is an ordered list of holes (oldest to newest, in terms of
420 * the sequence space).
421 */
422void
423tcp_sack_doack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th,
424 u_int32_t *newbytes_acked)
425{
426 struct sackhole *cur, *temp;
427 struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1], *sblkp;
428 int i, j, num_sack_blks;
429 tcp_seq old_snd_fack = 0, th_ack = th->th_ack;
430
431 num_sack_blks = 0;
432 /*
433 * If SND.UNA will be advanced by SEG.ACK, and if SACK holes exist,
434 * treat [SND.UNA, SEG.ACK) as if it is a SACK block.
435 */
436 if (SEQ_LT(tp->snd_una, th_ack) && !TAILQ_EMPTY(&tp->snd_holes)) {
437 sack_blocks[num_sack_blks].start = tp->snd_una;
438 sack_blocks[num_sack_blks++].end = th_ack;
439 }
440 /*
441 * Append received valid SACK blocks to sack_blocks[].
442 * Check that the SACK block range is valid.
443 */
444 for (i = 0; i < to->to_nsacks; i++) {
445 bcopy((to->to_sacks + i * TCPOLEN_SACK),
446 &sack, sizeof(sack));
447 sack.start = ntohl(sack.start);
448 sack.end = ntohl(sack.end);
449 if (TCP_VALIDATE_SACK_SEQ_NUMBERS(tp, &sack, th_ack))
450 sack_blocks[num_sack_blks++] = sack;
451 }
452
453 /*
454 * Return if SND.UNA is not advanced and no valid SACK block
455 * is received.
456 */
457 if (num_sack_blks == 0)
458 return;
459
460 VERIFY(num_sack_blks <= (TCP_MAX_SACK + 1));
461 /*
462 * Sort the SACK blocks so we can update the scoreboard
463 * with just one pass. The overhead of sorting upto 4+1 elements
464 * is less than making upto 4+1 passes over the scoreboard.
465 */
466 for (i = 0; i < num_sack_blks; i++) {
467 for (j = i + 1; j < num_sack_blks; j++) {
468 if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
469 sack = sack_blocks[i];
470 sack_blocks[i] = sack_blocks[j];
471 sack_blocks[j] = sack;
472 }
473 }
474 }
475 if (TAILQ_EMPTY(&tp->snd_holes)) {
476 /*
477 * Empty scoreboard. Need to initialize snd_fack (it may be
478 * uninitialized or have a bogus value). Scoreboard holes
479 * (from the sack blocks received) are created later below (in
480 * the logic that adds holes to the tail of the scoreboard).
481 */
482 tp->snd_fack = SEQ_MAX(tp->snd_una, th_ack);
483 *newbytes_acked += (tp->snd_fack - tp->snd_una);
484 }
485
486 old_snd_fack = tp->snd_fack;
487 /*
488 * In the while-loop below, incoming SACK blocks (sack_blocks[])
489 * and SACK holes (snd_holes) are traversed from their tails with
490 * just one pass in order to reduce the number of compares especially
491 * when the bandwidth-delay product is large.
492 * Note: Typically, in the first RTT of SACK recovery, the highest
493 * three or four SACK blocks with the same ack number are received.
494 * In the second RTT, if retransmitted data segments are not lost,
495 * the highest three or four SACK blocks with ack number advancing
496 * are received.
497 */
498 sblkp = &sack_blocks[num_sack_blks - 1]; /* Last SACK block */
499 if (SEQ_LT(tp->snd_fack, sblkp->start)) {
500 /*
501 * The highest SACK block is beyond fack.
502 * Append new SACK hole at the tail.
503 * If the second or later highest SACK blocks are also
504 * beyond the current fack, they will be inserted by
505 * way of hole splitting in the while-loop below.
506 */
507 temp = tcp_sackhole_insert(tp, tp->snd_fack,sblkp->start,NULL);
508 if (temp != NULL) {
509 tp->snd_fack = sblkp->end;
510 *newbytes_acked += (sblkp->end - sblkp->start);
511
512 /* Go to the previous sack block. */
513 sblkp--;
514 } else {
515 /*
516 * We failed to add a new hole based on the current
517 * sack block. Skip over all the sack blocks that
518 * fall completely to the right of snd_fack and proceed
519 * to trim the scoreboard based on the remaining sack
520 * blocks. This also trims the scoreboard for th_ack
521 * (which is sack_blocks[0]).
522 */
523 while (sblkp >= sack_blocks &&
524 SEQ_LT(tp->snd_fack, sblkp->start))
525 sblkp--;
526 if (sblkp >= sack_blocks &&
527 SEQ_LT(tp->snd_fack, sblkp->end)) {
528 *newbytes_acked += (sblkp->end - tp->snd_fack);
529 tp->snd_fack = sblkp->end;
530 }
531 }
532 } else if (SEQ_LT(tp->snd_fack, sblkp->end)) {
533 /* fack is advanced. */
534 *newbytes_acked += (sblkp->end - tp->snd_fack);
535 tp->snd_fack = sblkp->end;
536 }
537 /* We must have at least one SACK hole in scoreboard */
538 cur = TAILQ_LAST(&tp->snd_holes, sackhole_head); /* Last SACK hole */
539 /*
540 * Since the incoming sack blocks are sorted, we can process them
541 * making one sweep of the scoreboard.
542 */
543 while (sblkp >= sack_blocks && cur != NULL) {
544 if (SEQ_GEQ(sblkp->start, cur->end)) {
545 /*
546 * SACKs data beyond the current hole.
547 * Go to the previous sack block.
548 */
549 sblkp--;
550 continue;
551 }
552 if (SEQ_LEQ(sblkp->end, cur->start)) {
553 /*
554 * SACKs data before the current hole.
555 * Go to the previous hole.
556 */
557 cur = TAILQ_PREV(cur, sackhole_head, scblink);
558 continue;
559 }
560 tp->sackhint.sack_bytes_rexmit -= (cur->rxmit - cur->start);
561 if (SEQ_LEQ(sblkp->start, cur->start)) {
562 /* Data acks at least the beginning of hole */
563 if (SEQ_GEQ(sblkp->end, cur->end)) {
564 /* Acks entire hole, so delete hole */
565 *newbytes_acked += (cur->end - cur->start);
566
567 tcp_sack_detect_reordering(tp, cur,
568 cur->end, old_snd_fack);
569 temp = cur;
570 cur = TAILQ_PREV(cur, sackhole_head, scblink);
571 tcp_sackhole_remove(tp, temp);
572 /*
573 * The sack block may ack all or part of the next
574 * hole too, so continue onto the next hole.
575 */
576 continue;
577 } else {
578 /* Move start of hole forward */
579 *newbytes_acked += (sblkp->end - cur->start);
580 tcp_sack_detect_reordering(tp, cur,
581 sblkp->end, old_snd_fack);
582 cur->start = sblkp->end;
583 cur->rxmit = SEQ_MAX(cur->rxmit, cur->start);
584 }
585 } else {
586 /* Data acks at least the end of hole */
587 if (SEQ_GEQ(sblkp->end, cur->end)) {
588 /* Move end of hole backward */
589 *newbytes_acked += (cur->end - sblkp->start);
590 tcp_sack_detect_reordering(tp, cur,
591 cur->end, old_snd_fack);
592 cur->end = sblkp->start;
593 cur->rxmit = SEQ_MIN(cur->rxmit, cur->end);
594 } else {
595 /*
596 * ACKs some data in the middle of a hole;
597 * need to split current hole
598 */
599 *newbytes_acked += (sblkp->end - sblkp->start);
600 tcp_sack_detect_reordering(tp, cur,
601 sblkp->end, old_snd_fack);
602 temp = tcp_sackhole_insert(tp, sblkp->end,
603 cur->end, cur);
604 if (temp != NULL) {
605 if (SEQ_GT(cur->rxmit, temp->rxmit)) {
606 temp->rxmit = cur->rxmit;
607 tp->sackhint.sack_bytes_rexmit
608 += (temp->rxmit
609 - temp->start);
610 }
611 cur->end = sblkp->start;
612 cur->rxmit = SEQ_MIN(cur->rxmit,
613 cur->end);
614 /*
615 * Reset the rxmit_start to that of
616 * the current hole as that will
617 * help to compute the reorder
618 * window correctly
619 */
620 temp->rxmit_start = cur->rxmit_start;
621 }
622 }
623 }
624 tp->sackhint.sack_bytes_rexmit += (cur->rxmit - cur->start);
625 /*
626 * Testing sblkp->start against cur->start tells us whether
627 * we're done with the sack block or the sack hole.
628 * Accordingly, we advance one or the other.
629 */
630 if (SEQ_LEQ(sblkp->start, cur->start))
631 cur = TAILQ_PREV(cur, sackhole_head, scblink);
632 else
633 sblkp--;
634 }
635}
636
637/*
638 * Free all SACK holes to clear the scoreboard.
639 */
640void
641tcp_free_sackholes(struct tcpcb *tp)
642{
643 struct sackhole *q;
644
645 while ((q = TAILQ_FIRST(&tp->snd_holes)) != NULL)
646 tcp_sackhole_remove(tp, q);
647 tp->sackhint.sack_bytes_rexmit = 0;
648 tp->sackhint.nexthole = NULL;
649 tp->sack_newdata = 0;
650
651}
652
653/*
654 * Partial ack handling within a sack recovery episode.
655 * Keeping this very simple for now. When a partial ack
656 * is received, force snd_cwnd to a value that will allow
657 * the sender to transmit no more than 2 segments.
658 * If necessary, a better scheme can be adopted at a
659 * later point, but for now, the goal is to prevent the
660 * sender from bursting a large amount of data in the midst
661 * of sack recovery.
662 */
663void
664tcp_sack_partialack(struct tcpcb *tp, struct tcphdr *th)
665{
666 int num_segs = 1;
667
668 tp->t_timer[TCPT_REXMT] = 0;
669 tp->t_rtttime = 0;
670 /* send one or 2 segments based on how much new data was acked */
671 if (((BYTES_ACKED(th, tp)) / tp->t_maxseg) > 2)
672 num_segs = 2;
673 tp->snd_cwnd = (tp->sackhint.sack_bytes_rexmit +
674 (tp->snd_nxt - tp->sack_newdata) +
675 num_segs * tp->t_maxseg);
676 if (tp->snd_cwnd > tp->snd_ssthresh)
677 tp->snd_cwnd = tp->snd_ssthresh;
678 if (SEQ_LT(tp->snd_fack, tp->snd_recover) &&
679 tp->snd_fack == th->th_ack && TAILQ_EMPTY(&tp->snd_holes)) {
680 struct sackhole *temp;
681 /*
682 * we received a partial ack but there is no sack_hole
683 * that will cover the remaining seq space. In this case,
684 * create a hole from snd_fack to snd_recover so that
685 * the sack recovery will continue.
686 */
687 temp = tcp_sackhole_insert(tp, tp->snd_fack,
688 tp->snd_recover, NULL);
689 if (temp != NULL)
690 tp->snd_fack = tp->snd_recover;
691 }
692 (void) tcp_output(tp);
693}
694
695/*
696 * Debug version of tcp_sack_output() that walks the scoreboard. Used for
697 * now to sanity check the hint.
698 */
699static struct sackhole *
700tcp_sack_output_debug(struct tcpcb *tp, int *sack_bytes_rexmt)
701{
702 struct sackhole *p;
703
704 *sack_bytes_rexmt = 0;
705 TAILQ_FOREACH(p, &tp->snd_holes, scblink) {
706 if (SEQ_LT(p->rxmit, p->end)) {
707 if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */
708 continue;
709 }
710 *sack_bytes_rexmt += (p->rxmit - p->start);
711 break;
712 }
713 *sack_bytes_rexmt += (p->rxmit - p->start);
714 }
715 return (p);
716}
717
718/*
719 * Returns the next hole to retransmit and the number of retransmitted bytes
720 * from the scoreboard. We store both the next hole and the number of
721 * retransmitted bytes as hints (and recompute these on the fly upon SACK/ACK
722 * reception). This avoids scoreboard traversals completely.
723 *
724 * The loop here will traverse *at most* one link. Here's the argument.
725 * For the loop to traverse more than 1 link before finding the next hole to
726 * retransmit, we would need to have at least 1 node following the current hint
727 * with (rxmit == end). But, for all holes following the current hint,
728 * (start == rxmit), since we have not yet retransmitted from them. Therefore,
729 * in order to traverse more 1 link in the loop below, we need to have at least
730 * one node following the current hint with (start == rxmit == end).
731 * But that can't happen, (start == end) means that all the data in that hole
732 * has been sacked, in which case, the hole would have been removed from the
733 * scoreboard.
734 */
735struct sackhole *
736tcp_sack_output(struct tcpcb *tp, int *sack_bytes_rexmt)
737{
738 struct sackhole *hole = NULL, *dbg_hole = NULL;
739 int dbg_bytes_rexmt;
740
741 dbg_hole = tcp_sack_output_debug(tp, &dbg_bytes_rexmt);
742 *sack_bytes_rexmt = tp->sackhint.sack_bytes_rexmit;
743 hole = tp->sackhint.nexthole;
744 if (hole == NULL || SEQ_LT(hole->rxmit, hole->end))
745 goto out;
746 while ((hole = TAILQ_NEXT(hole, scblink)) != NULL) {
747 if (SEQ_LT(hole->rxmit, hole->end)) {
748 tp->sackhint.nexthole = hole;
749 break;
750 }
751 }
752out:
753 if (dbg_hole != hole) {
754 printf("%s: Computed sack hole not the same as cached value\n", __func__);
755 hole = dbg_hole;
756 }
757 if (*sack_bytes_rexmt != dbg_bytes_rexmt) {
758 printf("%s: Computed sack_bytes_retransmitted (%d) not "
759 "the same as cached value (%d)\n",
760 __func__, dbg_bytes_rexmt, *sack_bytes_rexmt);
761 *sack_bytes_rexmt = dbg_bytes_rexmt;
762 }
763 return (hole);
764}
765
766/*
767 * After a timeout, the SACK list may be rebuilt. This SACK information
768 * should be used to avoid retransmitting SACKed data. This function
769 * traverses the SACK list to see if snd_nxt should be moved forward.
770 */
771void
772tcp_sack_adjust(struct tcpcb *tp)
773{
774 struct sackhole *p, *cur = TAILQ_FIRST(&tp->snd_holes);
775
776 if (cur == NULL)
777 return; /* No holes */
778 if (SEQ_GEQ(tp->snd_nxt, tp->snd_fack))
779 return; /* We're already beyond any SACKed blocks */
780 /*
781 * Two cases for which we want to advance snd_nxt:
782 * i) snd_nxt lies between end of one hole and beginning of another
783 * ii) snd_nxt lies between end of last hole and snd_fack
784 */
785 while ((p = TAILQ_NEXT(cur, scblink)) != NULL) {
786 if (SEQ_LT(tp->snd_nxt, cur->end))
787 return;
788 if (SEQ_GEQ(tp->snd_nxt, p->start))
789 cur = p;
790 else {
791 tp->snd_nxt = p->start;
792 return;
793 }
794 }
795 if (SEQ_LT(tp->snd_nxt, cur->end))
796 return;
797 tp->snd_nxt = tp->snd_fack;
798 return;
799}
800
801/*
802 * This function returns TRUE if more than (tcprexmtthresh - 1) * SMSS
803 * bytes with sequence numbers greater than snd_una have been SACKed.
804 */
805boolean_t
806tcp_sack_byte_islost(struct tcpcb *tp)
807{
808 u_int32_t unacked_bytes, sndhole_bytes = 0;
809 struct sackhole *sndhole;
810 if (!SACK_ENABLED(tp) || IN_FASTRECOVERY(tp) ||
811 TAILQ_EMPTY(&tp->snd_holes) ||
812 (tp->t_flagsext & TF_PKTS_REORDERED))
813 return (FALSE);
814
815 unacked_bytes = tp->snd_max - tp->snd_una;
816
817 TAILQ_FOREACH(sndhole, &tp->snd_holes, scblink) {
818 sndhole_bytes += (sndhole->end - sndhole->start);
819 }
820
821 VERIFY(unacked_bytes >= sndhole_bytes);
822 return ((unacked_bytes - sndhole_bytes) >
823 ((tcprexmtthresh - 1) * tp->t_maxseg));
824}
825
826/*
827 * Process any DSACK options that might be present on an input packet
828 */
829
830boolean_t
831tcp_sack_process_dsack(struct tcpcb *tp, struct tcpopt *to,
832 struct tcphdr *th)
833{
834 struct sackblk first_sack, second_sack;
835 struct tcp_rxt_seg *rxseg;
836
837 bcopy(to->to_sacks, &first_sack, sizeof(first_sack));
838 first_sack.start = ntohl(first_sack.start);
839 first_sack.end = ntohl(first_sack.end);
840
841 if (to->to_nsacks > 1) {
842 bcopy((to->to_sacks + TCPOLEN_SACK), &second_sack,
843 sizeof(second_sack));
844 second_sack.start = ntohl(second_sack.start);
845 second_sack.end = ntohl(second_sack.end);
846 }
847
848 if (SEQ_LT(first_sack.start, th->th_ack) &&
849 SEQ_LEQ(first_sack.end, th->th_ack)) {
850 /*
851 * There is a dsack option reporting a duplicate segment
852 * also covered by cumulative acknowledgement.
853 *
854 * Validate the sequence numbers before looking at dsack
855 * option. The duplicate notification can come after
856 * snd_una moves forward. In order to set a window of valid
857 * sequence numbers to look for, we set a maximum send
858 * window within which the DSACK option will be processed.
859 */
860 if (!(TCP_DSACK_SEQ_IN_WINDOW(tp, first_sack.start, th->th_ack) &&
861 TCP_DSACK_SEQ_IN_WINDOW(tp, first_sack.end, th->th_ack))) {
862 to->to_nsacks--;
863 to->to_sacks += TCPOLEN_SACK;
864 tcpstat.tcps_dsack_recvd_old++;
865
866 /*
867 * returning true here so that the ack will not be
868 * treated as duplicate ack.
869 */
870 return (TRUE);
871 }
872 } else if (to->to_nsacks > 1 &&
873 SEQ_LEQ(second_sack.start, first_sack.start) &&
874 SEQ_GEQ(second_sack.end, first_sack.end)) {
875 /*
876 * there is a dsack option in the first block not
877 * covered by the cumulative acknowledgement but covered
878 * by the second sack block.
879 *
880 * verify the sequence numbes on the second sack block
881 * before processing the DSACK option. Returning false
882 * here will treat the ack as a duplicate ack.
883 */
884 if (!TCP_VALIDATE_SACK_SEQ_NUMBERS(tp, &second_sack,
885 th->th_ack)) {
886 to->to_nsacks--;
887 to->to_sacks += TCPOLEN_SACK;
888 tcpstat.tcps_dsack_recvd_old++;
889 return (TRUE);
890 }
891 } else {
892 /* no dsack options, proceed with processing the sack */
893 return (FALSE);
894 }
895
896 /* Update the tcpopt pointer to exclude dsack block */
897 to->to_nsacks--;
898 to->to_sacks += TCPOLEN_SACK;
899 tcpstat.tcps_dsack_recvd++;
900 tp->t_dsack_recvd++;
901
902 /* ignore DSACK option, if DSACK is disabled */
903 if (tp->t_flagsext & TF_DISABLE_DSACK)
904 return (TRUE);
905
906 /* If the DSACK is for TLP mark it as such */
907 if ((tp->t_flagsext & TF_SENT_TLPROBE) &&
908 first_sack.end == tp->t_tlphighrxt) {
909 if ((rxseg = tcp_rxtseg_find(tp, first_sack.start,
910 (first_sack.end - 1))) != NULL)
911 rxseg->rx_flags |= TCP_RXT_DSACK_FOR_TLP;
912 }
913 /* Update the sender's retransmit segment state */
914 if (((tp->t_rxtshift == 1 && first_sack.start == tp->snd_una) ||
915 ((tp->t_flagsext & TF_SENT_TLPROBE) &&
916 first_sack.end == tp->t_tlphighrxt)) &&
917 TAILQ_EMPTY(&tp->snd_holes) &&
918 SEQ_GT(th->th_ack, tp->snd_una)) {
919 /*
920 * If the dsack is for a retransmitted packet and one of
921 * the two cases is true, it indicates ack loss:
922 * - retransmit timeout and first_sack.start == snd_una
923 * - TLP probe and first_sack.end == tlphighrxt
924 *
925 * Ignore dsack and do not update state when there is
926 * ack loss
927 */
928 tcpstat.tcps_dsack_ackloss++;
929
930 return (TRUE);
931 } else if ((rxseg = tcp_rxtseg_find(tp, first_sack.start,
932 (first_sack.end - 1))) == NULL) {
933 /*
934 * Duplicate notification was not triggered by a
935 * retransmission. This might be due to network duplication,
936 * disable further DSACK processing.
937 */
938 if (!tcp_dsack_ignore_hw_duplicates) {
939 tp->t_flagsext |= TF_DISABLE_DSACK;
940 tcpstat.tcps_dsack_disable++;
941 }
942 } else {
943 /*
944 * If the segment was retransmitted only once, mark it as
945 * spurious. Otherwise ignore the duplicate notification.
946 */
947 if (rxseg->rx_count == 1)
948 rxseg->rx_flags |= TCP_RXT_SPURIOUS;
949 else
950 rxseg->rx_flags &= ~TCP_RXT_SPURIOUS;
951 }
952 return (TRUE);
953}
954