1 /*-
2 * SPDX-License-Identifier: BSD-3-Clause
3 *
4 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995
5 * The Regents of the University of California.
6 * All rights reserved.
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 University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
19 *
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * SUCH DAMAGE.
31 *
32 * @(#)tcp_sack.c 8.12 (Berkeley) 5/24/95
33 */
34
35 /*-
36 * @@(#)COPYRIGHT 1.1 (NRL) 17 January 1995
37 *
38 * NRL grants permission for redistribution and use in source and binary
39 * forms, with or without modification, of the software and documentation
40 * created at NRL provided that the following conditions are met:
41 *
42 * 1. Redistributions of source code must retain the above copyright
43 * notice, this list of conditions and the following disclaimer.
44 * 2. Redistributions in binary form must reproduce the above copyright
45 * notice, this list of conditions and the following disclaimer in the
46 * documentation and/or other materials provided with the distribution.
47 * 3. All advertising materials mentioning features or use of this software
48 * must display the following acknowledgements:
49 * This product includes software developed by the University of
50 * California, Berkeley and its contributors.
51 * This product includes software developed at the Information
52 * Technology Division, US Naval Research Laboratory.
53 * 4. Neither the name of the NRL nor the names of its contributors
54 * may be used to endorse or promote products derived from this software
55 * without specific prior written permission.
56 *
57 * THE SOFTWARE PROVIDED BY NRL IS PROVIDED BY NRL AND CONTRIBUTORS ``AS
58 * IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
59 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
60 * PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NRL OR
61 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
62 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
63 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
64 * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
65 * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
66 * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
67 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
68 *
69 * The views and conclusions contained in the software and documentation
70 * are those of the authors and should not be interpreted as representing
71 * official policies, either expressed or implied, of the US Naval
72 * Research Laboratory (NRL).
73 */
74
75 #include <sys/cdefs.h>
76 #include "opt_inet.h"
77 #include "opt_inet6.h"
78
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
82 #include <sys/sysctl.h>
83 #include <sys/malloc.h>
84 #include <sys/mbuf.h>
85 #include <sys/proc.h> /* for proc0 declaration */
86 #include <sys/protosw.h>
87 #include <sys/socket.h>
88 #include <sys/socketvar.h>
89 #include <sys/syslog.h>
90 #include <sys/systm.h>
91
92 #include <machine/cpu.h> /* before tcp_seq.h, for tcp_random18() */
93
94 #include <vm/uma.h>
95
96 #include <net/if.h>
97 #include <net/if_var.h>
98 #include <net/route.h>
99 #include <net/vnet.h>
100
101 #include <netinet/in.h>
102 #include <netinet/in_systm.h>
103 #include <netinet/ip.h>
104 #include <netinet/in_var.h>
105 #include <netinet/in_pcb.h>
106 #include <netinet/ip_var.h>
107 #include <netinet/ip6.h>
108 #include <netinet/icmp6.h>
109 #include <netinet6/nd6.h>
110 #include <netinet6/ip6_var.h>
111 #include <netinet6/in6_pcb.h>
112 #include <netinet/tcp.h>
113 #include <netinet/tcp_fsm.h>
114 #include <netinet/tcp_seq.h>
115 #include <netinet/tcp_timer.h>
116 #include <netinet/tcp_var.h>
117 #include <netinet/tcpip.h>
118 #include <netinet/cc/cc.h>
119
120 #include <machine/in_cksum.h>
121
122 VNET_DECLARE(struct uma_zone *, sack_hole_zone);
123 #define V_sack_hole_zone VNET(sack_hole_zone)
124
125 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, sack, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
126 "TCP SACK");
127
128 VNET_DEFINE(int, tcp_do_sack) = 1;
129 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, enable, CTLFLAG_VNET | CTLFLAG_RW,
130 &VNET_NAME(tcp_do_sack), 0,
131 "Enable/Disable TCP SACK support");
132
133 VNET_DEFINE(int, tcp_do_newsack) = 1;
134 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, revised, CTLFLAG_VNET | CTLFLAG_RW,
135 &VNET_NAME(tcp_do_newsack), 0,
136 "Use revised SACK loss recovery per RFC 6675");
137
138 VNET_DEFINE(int, tcp_sack_maxholes) = 128;
139 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, maxholes, CTLFLAG_VNET | CTLFLAG_RW,
140 &VNET_NAME(tcp_sack_maxholes), 0,
141 "Maximum number of TCP SACK holes allowed per connection");
142
143 VNET_DEFINE(int, tcp_sack_globalmaxholes) = 65536;
144 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, globalmaxholes, CTLFLAG_VNET | CTLFLAG_RW,
145 &VNET_NAME(tcp_sack_globalmaxholes), 0,
146 "Global maximum number of TCP SACK holes");
147
148 VNET_DEFINE(int, tcp_sack_globalholes) = 0;
149 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, globalholes, CTLFLAG_VNET | CTLFLAG_RD,
150 &VNET_NAME(tcp_sack_globalholes), 0,
151 "Global number of TCP SACK holes currently allocated");
152
153 int
tcp_dsack_block_exists(struct tcpcb * tp)154 tcp_dsack_block_exists(struct tcpcb *tp)
155 {
156 /* Return true if a DSACK block exists */
157 if (tp->rcv_numsacks == 0)
158 return (0);
159 if (SEQ_LEQ(tp->sackblks[0].end, tp->rcv_nxt))
160 return(1);
161 return (0);
162 }
163
164 /*
165 * This function will find overlaps with the currently stored sackblocks
166 * and add any overlap as a dsack block upfront
167 */
168 void
tcp_update_dsack_list(struct tcpcb * tp,tcp_seq rcv_start,tcp_seq rcv_end)169 tcp_update_dsack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
170 {
171 struct sackblk head_blk,mid_blk,saved_blks[MAX_SACK_BLKS];
172 int i, j, n, identical;
173 tcp_seq start, end;
174
175 INP_WLOCK_ASSERT(tptoinpcb(tp));
176
177 KASSERT(SEQ_LT(rcv_start, rcv_end), ("rcv_start < rcv_end"));
178
179 if (SEQ_LT(rcv_end, tp->rcv_nxt) ||
180 ((rcv_end == tp->rcv_nxt) &&
181 (tp->rcv_numsacks > 0 ) &&
182 (tp->sackblks[0].end == tp->rcv_nxt))) {
183 saved_blks[0].start = rcv_start;
184 saved_blks[0].end = rcv_end;
185 } else {
186 saved_blks[0].start = saved_blks[0].end = 0;
187 }
188
189 head_blk.start = head_blk.end = 0;
190 mid_blk.start = rcv_start;
191 mid_blk.end = rcv_end;
192 identical = 0;
193
194 for (i = 0; i < tp->rcv_numsacks; i++) {
195 start = tp->sackblks[i].start;
196 end = tp->sackblks[i].end;
197 if (SEQ_LT(rcv_end, start)) {
198 /* pkt left to sack blk */
199 continue;
200 }
201 if (SEQ_GT(rcv_start, end)) {
202 /* pkt right to sack blk */
203 continue;
204 }
205 if (SEQ_GT(tp->rcv_nxt, end)) {
206 if ((SEQ_MAX(rcv_start, start) != SEQ_MIN(rcv_end, end)) &&
207 (SEQ_GT(head_blk.start, SEQ_MAX(rcv_start, start)) ||
208 (head_blk.start == head_blk.end))) {
209 head_blk.start = SEQ_MAX(rcv_start, start);
210 head_blk.end = SEQ_MIN(rcv_end, end);
211 }
212 continue;
213 }
214 if (((head_blk.start == head_blk.end) ||
215 SEQ_LT(start, head_blk.start)) &&
216 (SEQ_GT(end, rcv_start) &&
217 SEQ_LEQ(start, rcv_end))) {
218 head_blk.start = start;
219 head_blk.end = end;
220 }
221 mid_blk.start = SEQ_MIN(mid_blk.start, start);
222 mid_blk.end = SEQ_MAX(mid_blk.end, end);
223 if ((mid_blk.start == start) &&
224 (mid_blk.end == end))
225 identical = 1;
226 }
227 if (SEQ_LT(head_blk.start, head_blk.end)) {
228 /* store overlapping range */
229 saved_blks[0].start = SEQ_MAX(rcv_start, head_blk.start);
230 saved_blks[0].end = SEQ_MIN(rcv_end, head_blk.end);
231 }
232 n = 1;
233 /*
234 * Second, if not ACKed, store the SACK block that
235 * overlaps with the DSACK block unless it is identical
236 */
237 if ((SEQ_LT(tp->rcv_nxt, mid_blk.end) &&
238 !((mid_blk.start == saved_blks[0].start) &&
239 (mid_blk.end == saved_blks[0].end))) ||
240 identical == 1) {
241 saved_blks[n].start = mid_blk.start;
242 saved_blks[n++].end = mid_blk.end;
243 }
244 for (j = 0; (j < tp->rcv_numsacks) && (n < MAX_SACK_BLKS); j++) {
245 if (((SEQ_LT(tp->sackblks[j].end, mid_blk.start) ||
246 SEQ_GT(tp->sackblks[j].start, mid_blk.end)) &&
247 (SEQ_GT(tp->sackblks[j].start, tp->rcv_nxt))))
248 saved_blks[n++] = tp->sackblks[j];
249 }
250 j = 0;
251 for (i = 0; i < n; i++) {
252 /* we can end up with a stale initial entry */
253 if (SEQ_LT(saved_blks[i].start, saved_blks[i].end)) {
254 tp->sackblks[j++] = saved_blks[i];
255 }
256 }
257 tp->rcv_numsacks = j;
258 }
259
260 /*
261 * This function is called upon receipt of new valid data (while not in
262 * header prediction mode), and it updates the ordered list of sacks.
263 */
264 void
tcp_update_sack_list(struct tcpcb * tp,tcp_seq rcv_start,tcp_seq rcv_end)265 tcp_update_sack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
266 {
267 /*
268 * First reported block MUST be the most recent one. Subsequent
269 * blocks SHOULD be in the order in which they arrived at the
270 * receiver. These two conditions make the implementation fully
271 * compliant with RFC 2018.
272 */
273 struct sackblk head_blk, saved_blks[MAX_SACK_BLKS];
274 int num_head, num_saved, i;
275
276 INP_WLOCK_ASSERT(tptoinpcb(tp));
277
278 /* Check arguments. */
279 KASSERT(SEQ_LEQ(rcv_start, rcv_end), ("rcv_start <= rcv_end"));
280
281 if ((rcv_start == rcv_end) &&
282 (tp->rcv_numsacks >= 1) &&
283 (rcv_end == tp->sackblks[0].end)) {
284 /* retaining DSACK block below rcv_nxt (todrop) */
285 head_blk = tp->sackblks[0];
286 } else {
287 /* SACK block for the received segment. */
288 head_blk.start = rcv_start;
289 head_blk.end = rcv_end;
290 }
291
292 /*
293 * Merge updated SACK blocks into head_blk, and save unchanged SACK
294 * blocks into saved_blks[]. num_saved will have the number of the
295 * saved SACK blocks.
296 */
297 num_saved = 0;
298 for (i = 0; i < tp->rcv_numsacks; i++) {
299 tcp_seq start = tp->sackblks[i].start;
300 tcp_seq end = tp->sackblks[i].end;
301 if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
302 /*
303 * Discard this SACK block.
304 */
305 } else if (SEQ_LEQ(head_blk.start, end) &&
306 SEQ_GEQ(head_blk.end, start)) {
307 /*
308 * Merge this SACK block into head_blk. This SACK
309 * block itself will be discarded.
310 */
311 /*
312 * |-|
313 * |---| merge
314 *
315 * |-|
316 * |---| merge
317 *
318 * |-----|
319 * |-| DSACK smaller
320 *
321 * |-|
322 * |-----| DSACK smaller
323 */
324 if (head_blk.start == end)
325 head_blk.start = start;
326 else if (head_blk.end == start)
327 head_blk.end = end;
328 else {
329 if (SEQ_LT(head_blk.start, start)) {
330 tcp_seq temp = start;
331 start = head_blk.start;
332 head_blk.start = temp;
333 }
334 if (SEQ_GT(head_blk.end, end)) {
335 tcp_seq temp = end;
336 end = head_blk.end;
337 head_blk.end = temp;
338 }
339 if ((head_blk.start != start) ||
340 (head_blk.end != end)) {
341 if ((num_saved >= 1) &&
342 SEQ_GEQ(saved_blks[num_saved-1].start, start) &&
343 SEQ_LEQ(saved_blks[num_saved-1].end, end))
344 num_saved--;
345 saved_blks[num_saved].start = start;
346 saved_blks[num_saved].end = end;
347 num_saved++;
348 }
349 }
350 } else {
351 /*
352 * This block supercedes the prior block
353 */
354 if ((num_saved >= 1) &&
355 SEQ_GEQ(saved_blks[num_saved-1].start, start) &&
356 SEQ_LEQ(saved_blks[num_saved-1].end, end))
357 num_saved--;
358 /*
359 * Save this SACK block.
360 */
361 saved_blks[num_saved].start = start;
362 saved_blks[num_saved].end = end;
363 num_saved++;
364 }
365 }
366
367 /*
368 * Update SACK list in tp->sackblks[].
369 */
370 num_head = 0;
371 if (SEQ_LT(rcv_start, rcv_end)) {
372 /*
373 * The received data segment is an out-of-order segment. Put
374 * head_blk at the top of SACK list.
375 */
376 tp->sackblks[0] = head_blk;
377 num_head = 1;
378 /*
379 * If the number of saved SACK blocks exceeds its limit,
380 * discard the last SACK block.
381 */
382 if (num_saved >= MAX_SACK_BLKS)
383 num_saved--;
384 }
385 if ((rcv_start == rcv_end) &&
386 (rcv_start == tp->sackblks[0].end)) {
387 num_head = 1;
388 }
389 if (num_saved > 0) {
390 /*
391 * Copy the saved SACK blocks back.
392 */
393 bcopy(saved_blks, &tp->sackblks[num_head],
394 sizeof(struct sackblk) * num_saved);
395 }
396
397 /* Save the number of SACK blocks. */
398 tp->rcv_numsacks = num_head + num_saved;
399 }
400
401 void
tcp_clean_dsack_blocks(struct tcpcb * tp)402 tcp_clean_dsack_blocks(struct tcpcb *tp)
403 {
404 struct sackblk saved_blks[MAX_SACK_BLKS];
405 int num_saved, i;
406
407 INP_WLOCK_ASSERT(tptoinpcb(tp));
408 /*
409 * Clean up any DSACK blocks that
410 * are in our queue of sack blocks.
411 *
412 */
413 num_saved = 0;
414 for (i = 0; i < tp->rcv_numsacks; i++) {
415 tcp_seq start = tp->sackblks[i].start;
416 tcp_seq end = tp->sackblks[i].end;
417 if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
418 /*
419 * Discard this D-SACK block.
420 */
421 continue;
422 }
423 /*
424 * Save this SACK block.
425 */
426 saved_blks[num_saved].start = start;
427 saved_blks[num_saved].end = end;
428 num_saved++;
429 }
430 if (num_saved > 0) {
431 /*
432 * Copy the saved SACK blocks back.
433 */
434 bcopy(saved_blks, &tp->sackblks[0],
435 sizeof(struct sackblk) * num_saved);
436 }
437 tp->rcv_numsacks = num_saved;
438 }
439
440 /*
441 * Delete all receiver-side SACK information.
442 */
443 void
tcp_clean_sackreport(struct tcpcb * tp)444 tcp_clean_sackreport(struct tcpcb *tp)
445 {
446 int i;
447
448 INP_WLOCK_ASSERT(tptoinpcb(tp));
449 tp->rcv_numsacks = 0;
450 for (i = 0; i < MAX_SACK_BLKS; i++)
451 tp->sackblks[i].start = tp->sackblks[i].end=0;
452 }
453
454 /*
455 * Allocate struct sackhole.
456 */
457 static struct sackhole *
tcp_sackhole_alloc(struct tcpcb * tp,tcp_seq start,tcp_seq end)458 tcp_sackhole_alloc(struct tcpcb *tp, tcp_seq start, tcp_seq end)
459 {
460 struct sackhole *hole;
461
462 if (tp->snd_numholes >= V_tcp_sack_maxholes ||
463 V_tcp_sack_globalholes >= V_tcp_sack_globalmaxholes) {
464 TCPSTAT_INC(tcps_sack_sboverflow);
465 return NULL;
466 }
467
468 hole = (struct sackhole *)uma_zalloc(V_sack_hole_zone, M_NOWAIT);
469 if (hole == NULL)
470 return NULL;
471
472 hole->start = start;
473 hole->end = end;
474 hole->rxmit = start;
475
476 tp->snd_numholes++;
477 atomic_add_int(&V_tcp_sack_globalholes, 1);
478
479 return hole;
480 }
481
482 /*
483 * Free struct sackhole.
484 */
485 static void
tcp_sackhole_free(struct tcpcb * tp,struct sackhole * hole)486 tcp_sackhole_free(struct tcpcb *tp, struct sackhole *hole)
487 {
488
489 uma_zfree(V_sack_hole_zone, hole);
490
491 tp->snd_numholes--;
492 atomic_subtract_int(&V_tcp_sack_globalholes, 1);
493
494 KASSERT(tp->snd_numholes >= 0, ("tp->snd_numholes >= 0"));
495 KASSERT(V_tcp_sack_globalholes >= 0, ("tcp_sack_globalholes >= 0"));
496 }
497
498 /*
499 * Insert new SACK hole into scoreboard.
500 */
501 static struct sackhole *
tcp_sackhole_insert(struct tcpcb * tp,tcp_seq start,tcp_seq end,struct sackhole * after)502 tcp_sackhole_insert(struct tcpcb *tp, tcp_seq start, tcp_seq end,
503 struct sackhole *after)
504 {
505 struct sackhole *hole;
506
507 /* Allocate a new SACK hole. */
508 hole = tcp_sackhole_alloc(tp, start, end);
509 if (hole == NULL)
510 return NULL;
511
512 /* Insert the new SACK hole into scoreboard. */
513 if (after != NULL)
514 TAILQ_INSERT_AFTER(&tp->snd_holes, after, hole, scblink);
515 else
516 TAILQ_INSERT_TAIL(&tp->snd_holes, hole, scblink);
517
518 /* Update SACK hint. */
519 if (tp->sackhint.nexthole == NULL)
520 tp->sackhint.nexthole = hole;
521
522 return hole;
523 }
524
525 /*
526 * Remove SACK hole from scoreboard.
527 */
528 static void
tcp_sackhole_remove(struct tcpcb * tp,struct sackhole * hole)529 tcp_sackhole_remove(struct tcpcb *tp, struct sackhole *hole)
530 {
531
532 /* Update SACK hint. */
533 if (tp->sackhint.nexthole == hole)
534 tp->sackhint.nexthole = TAILQ_NEXT(hole, scblink);
535
536 /* Remove this SACK hole. */
537 TAILQ_REMOVE(&tp->snd_holes, hole, scblink);
538
539 /* Free this SACK hole. */
540 tcp_sackhole_free(tp, hole);
541 }
542
543 /*
544 * Process cumulative ACK and the TCP SACK option to update the scoreboard.
545 * tp->snd_holes is an ordered list of holes (oldest to newest, in terms of
546 * the sequence space).
547 * Returns SACK_NEWLOSS if incoming ACK indicates ongoing loss (hole split, new hole),
548 * SACK_CHANGE if incoming ACK has previously unknown SACK information,
549 * SACK_NOCHANGE otherwise.
550 */
551 sackstatus_t
tcp_sack_doack(struct tcpcb * tp,struct tcpopt * to,tcp_seq th_ack)552 tcp_sack_doack(struct tcpcb *tp, struct tcpopt *to, tcp_seq th_ack)
553 {
554 struct sackhole *cur, *temp;
555 struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1], *sblkp;
556 int i, j, num_sack_blks;
557 sackstatus_t sack_changed;
558 int delivered_data, left_edge_delta;
559
560 INP_WLOCK_ASSERT(tptoinpcb(tp));
561
562 num_sack_blks = 0;
563 sack_changed = SACK_NOCHANGE;
564 delivered_data = 0;
565 left_edge_delta = 0;
566 /*
567 * If SND.UNA will be advanced by SEG.ACK, and if SACK holes exist,
568 * treat [SND.UNA, SEG.ACK) as if it is a SACK block.
569 * Account changes to SND.UNA always in delivered data.
570 */
571 if (SEQ_LT(tp->snd_una, th_ack) && !TAILQ_EMPTY(&tp->snd_holes)) {
572 left_edge_delta = th_ack - tp->snd_una;
573 sack_blocks[num_sack_blks].start = tp->snd_una;
574 sack_blocks[num_sack_blks++].end = th_ack;
575 /*
576 * Pulling snd_fack forward if we got here
577 * due to DSACK blocks
578 */
579 if (SEQ_LT(tp->snd_fack, th_ack)) {
580 delivered_data += th_ack - tp->snd_una;
581 tp->snd_fack = th_ack;
582 sack_changed = SACK_CHANGE;
583 }
584 }
585 /*
586 * Append received valid SACK blocks to sack_blocks[], but only if we
587 * received new blocks from the other side.
588 */
589 if (to->to_flags & TOF_SACK) {
590 for (i = 0; i < to->to_nsacks; i++) {
591 bcopy((to->to_sacks + i * TCPOLEN_SACK),
592 &sack, sizeof(sack));
593 sack.start = ntohl(sack.start);
594 sack.end = ntohl(sack.end);
595 if (SEQ_GT(sack.end, sack.start) &&
596 SEQ_GT(sack.start, tp->snd_una) &&
597 SEQ_GT(sack.start, th_ack) &&
598 SEQ_LT(sack.start, tp->snd_max) &&
599 SEQ_GT(sack.end, tp->snd_una) &&
600 SEQ_LEQ(sack.end, tp->snd_max)) {
601 sack_blocks[num_sack_blks++] = sack;
602 } else if (SEQ_LEQ(sack.start, th_ack) &&
603 SEQ_LEQ(sack.end, th_ack)) {
604 /*
605 * Its a D-SACK block.
606 */
607 tcp_record_dsack(tp, sack.start, sack.end, 0);
608 }
609 }
610 }
611 /*
612 * Return if SND.UNA is not advanced and no valid SACK block is
613 * received.
614 */
615 if (num_sack_blks == 0)
616 return (sack_changed);
617
618 /*
619 * Sort the SACK blocks so we can update the scoreboard with just one
620 * pass. The overhead of sorting up to 4+1 elements is less than
621 * making up to 4+1 passes over the scoreboard.
622 */
623 for (i = 0; i < num_sack_blks; i++) {
624 for (j = i + 1; j < num_sack_blks; j++) {
625 if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
626 sack = sack_blocks[i];
627 sack_blocks[i] = sack_blocks[j];
628 sack_blocks[j] = sack;
629 }
630 }
631 }
632 if (TAILQ_EMPTY(&tp->snd_holes)) {
633 /*
634 * Empty scoreboard. Need to initialize snd_fack (it may be
635 * uninitialized or have a bogus value). Scoreboard holes
636 * (from the sack blocks received) are created later below
637 * (in the logic that adds holes to the tail of the
638 * scoreboard).
639 */
640 tp->snd_fack = SEQ_MAX(tp->snd_una, th_ack);
641 tp->sackhint.sacked_bytes = 0; /* reset */
642 }
643 /*
644 * In the while-loop below, incoming SACK blocks (sack_blocks[]) and
645 * SACK holes (snd_holes) are traversed from their tails with just
646 * one pass in order to reduce the number of compares especially when
647 * the bandwidth-delay product is large.
648 *
649 * Note: Typically, in the first RTT of SACK recovery, the highest
650 * three or four SACK blocks with the same ack number are received.
651 * In the second RTT, if retransmitted data segments are not lost,
652 * the highest three or four SACK blocks with ack number advancing
653 * are received.
654 */
655 sblkp = &sack_blocks[num_sack_blks - 1]; /* Last SACK block */
656 tp->sackhint.last_sack_ack = sblkp->end;
657 if (SEQ_LT(tp->snd_fack, sblkp->start)) {
658 /*
659 * The highest SACK block is beyond fack. First,
660 * check if there was a successful Rescue Retransmission,
661 * and move this hole left. With normal holes, snd_fack
662 * is always to the right of the end.
663 */
664 if (((temp = TAILQ_LAST(&tp->snd_holes, sackhole_head)) != NULL) &&
665 SEQ_LEQ(tp->snd_fack,temp->end)) {
666 temp->start = SEQ_MAX(tp->snd_fack, SEQ_MAX(tp->snd_una, th_ack));
667 temp->end = sblkp->start;
668 temp->rxmit = temp->start;
669 delivered_data += sblkp->end - sblkp->start;
670 tp->snd_fack = sblkp->end;
671 sblkp--;
672 sack_changed = SACK_NEWLOSS;
673 } else {
674 /*
675 * Append a new SACK hole at the tail. If the
676 * second or later highest SACK blocks are also
677 * beyond the current fack, they will be inserted
678 * by way of hole splitting in the while-loop below.
679 */
680 temp = tcp_sackhole_insert(tp, tp->snd_fack,sblkp->start,NULL);
681 if (temp != NULL) {
682 delivered_data += sblkp->end - sblkp->start;
683 tp->snd_fack = sblkp->end;
684 /* Go to the previous sack block. */
685 sblkp--;
686 sack_changed = SACK_CHANGE;
687 } else {
688 /*
689 * We failed to add a new hole based on the current
690 * sack block. Skip over all the sack blocks that
691 * fall completely to the right of snd_fack and
692 * proceed to trim the scoreboard based on the
693 * remaining sack blocks. This also trims the
694 * scoreboard for th_ack (which is sack_blocks[0]).
695 */
696 while (sblkp >= sack_blocks &&
697 SEQ_LT(tp->snd_fack, sblkp->start))
698 sblkp--;
699 if (sblkp >= sack_blocks &&
700 SEQ_LT(tp->snd_fack, sblkp->end)) {
701 delivered_data += sblkp->end - tp->snd_fack;
702 tp->snd_fack = sblkp->end;
703 /*
704 * While the Scoreboard didn't change in
705 * size, we only ended up here because
706 * some SACK data had to be dismissed.
707 */
708 sack_changed = SACK_NEWLOSS;
709 }
710 }
711 }
712 } else if (SEQ_LT(tp->snd_fack, sblkp->end)) {
713 /* fack is advanced. */
714 delivered_data += sblkp->end - tp->snd_fack;
715 tp->snd_fack = sblkp->end;
716 sack_changed = SACK_CHANGE;
717 }
718 cur = TAILQ_LAST(&tp->snd_holes, sackhole_head); /* Last SACK hole. */
719 /*
720 * Since the incoming sack blocks are sorted, we can process them
721 * making one sweep of the scoreboard.
722 */
723 while (sblkp >= sack_blocks && cur != NULL) {
724 if (SEQ_GEQ(sblkp->start, cur->end)) {
725 /*
726 * SACKs data beyond the current hole. Go to the
727 * previous sack block.
728 */
729 sblkp--;
730 continue;
731 }
732 if (SEQ_LEQ(sblkp->end, cur->start)) {
733 /*
734 * SACKs data before the current hole. Go to the
735 * previous hole.
736 */
737 cur = TAILQ_PREV(cur, sackhole_head, scblink);
738 continue;
739 }
740 tp->sackhint.sack_bytes_rexmit -=
741 (SEQ_MIN(cur->rxmit, cur->end) - cur->start);
742 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
743 ("sackhint bytes rtx >= 0"));
744 sack_changed = SACK_CHANGE;
745 if (SEQ_LEQ(sblkp->start, cur->start)) {
746 /* Data acks at least the beginning of hole. */
747 if (SEQ_GEQ(sblkp->end, cur->end)) {
748 /* Acks entire hole, so delete hole. */
749 delivered_data += (cur->end - cur->start);
750 temp = cur;
751 cur = TAILQ_PREV(cur, sackhole_head, scblink);
752 tcp_sackhole_remove(tp, temp);
753 /*
754 * The sack block may ack all or part of the
755 * next hole too, so continue onto the next
756 * hole.
757 */
758 continue;
759 } else {
760 /* Move start of hole forward. */
761 delivered_data += (sblkp->end - cur->start);
762 cur->start = sblkp->end;
763 cur->rxmit = SEQ_MAX(cur->rxmit, cur->start);
764 }
765 } else {
766 /* Data acks at least the end of hole. */
767 if (SEQ_GEQ(sblkp->end, cur->end)) {
768 /* Move end of hole backward. */
769 delivered_data += (cur->end - sblkp->start);
770 cur->end = sblkp->start;
771 cur->rxmit = SEQ_MIN(cur->rxmit, cur->end);
772 if ((tp->t_flags & TF_LRD) && SEQ_GEQ(cur->rxmit, cur->end))
773 cur->rxmit = tp->snd_recover;
774 } else {
775 /*
776 * ACKs some data in middle of a hole; need
777 * to split current hole
778 */
779 temp = tcp_sackhole_insert(tp, sblkp->end,
780 cur->end, cur);
781 sack_changed = SACK_NEWLOSS;
782 if (temp != NULL) {
783 if (SEQ_GT(cur->rxmit, temp->rxmit)) {
784 temp->rxmit = cur->rxmit;
785 tp->sackhint.sack_bytes_rexmit +=
786 (SEQ_MIN(temp->rxmit,
787 temp->end) - temp->start);
788 }
789 cur->end = sblkp->start;
790 cur->rxmit = SEQ_MIN(cur->rxmit,
791 cur->end);
792 if ((tp->t_flags & TF_LRD) && SEQ_GEQ(cur->rxmit, cur->end))
793 cur->rxmit = tp->snd_recover;
794 delivered_data += (sblkp->end - sblkp->start);
795 }
796 }
797 }
798 tp->sackhint.sack_bytes_rexmit +=
799 (SEQ_MIN(cur->rxmit, cur->end) - cur->start);
800 /*
801 * Testing sblkp->start against cur->start tells us whether
802 * we're done with the sack block or the sack hole.
803 * Accordingly, we advance one or the other.
804 */
805 if (SEQ_LEQ(sblkp->start, cur->start))
806 cur = TAILQ_PREV(cur, sackhole_head, scblink);
807 else
808 sblkp--;
809 }
810 if (!(to->to_flags & TOF_SACK))
811 /*
812 * If this ACK did not contain any
813 * SACK blocks, any only moved the
814 * left edge right, it is a pure
815 * cumulative ACK. Do not count
816 * DupAck for this. Also required
817 * for RFC6675 rescue retransmission.
818 */
819 sack_changed = SACK_NOCHANGE;
820 tp->sackhint.delivered_data = delivered_data;
821 tp->sackhint.sacked_bytes += delivered_data - left_edge_delta;
822 KASSERT((delivered_data >= 0), ("delivered_data < 0"));
823 KASSERT((tp->sackhint.sacked_bytes >= 0), ("sacked_bytes < 0"));
824 return (sack_changed);
825 }
826
827 /*
828 * Free all SACK holes to clear the scoreboard.
829 */
830 void
tcp_free_sackholes(struct tcpcb * tp)831 tcp_free_sackholes(struct tcpcb *tp)
832 {
833 struct sackhole *q;
834
835 INP_WLOCK_ASSERT(tptoinpcb(tp));
836 while ((q = TAILQ_FIRST(&tp->snd_holes)) != NULL)
837 tcp_sackhole_remove(tp, q);
838 tp->sackhint.sack_bytes_rexmit = 0;
839
840 KASSERT(tp->snd_numholes == 0, ("tp->snd_numholes == 0"));
841 KASSERT(tp->sackhint.nexthole == NULL,
842 ("tp->sackhint.nexthole == NULL"));
843 }
844
845 /*
846 * Partial ack handling within a sack recovery episode. Keeping this very
847 * simple for now. When a partial ack is received, force snd_cwnd to a value
848 * that will allow the sender to transmit no more than 2 segments. If
849 * necessary, a better scheme can be adopted at a later point, but for now,
850 * the goal is to prevent the sender from bursting a large amount of data in
851 * the midst of sack recovery.
852 */
853 void
tcp_sack_partialack(struct tcpcb * tp,struct tcphdr * th)854 tcp_sack_partialack(struct tcpcb *tp, struct tcphdr *th)
855 {
856 int num_segs = 1;
857 u_int maxseg = tcp_maxseg(tp);
858
859 INP_WLOCK_ASSERT(tptoinpcb(tp));
860 tcp_timer_activate(tp, TT_REXMT, 0);
861 tp->t_rtttime = 0;
862 /* Send one or 2 segments based on how much new data was acked. */
863 if ((BYTES_THIS_ACK(tp, th) / maxseg) >= 2)
864 num_segs = 2;
865 tp->snd_cwnd = (tp->sackhint.sack_bytes_rexmit +
866 (tp->snd_nxt - tp->snd_recover) + num_segs * maxseg);
867 if (tp->snd_cwnd > tp->snd_ssthresh)
868 tp->snd_cwnd = tp->snd_ssthresh;
869 tp->t_flags |= TF_ACKNOW;
870 /*
871 * RFC6675 rescue retransmission
872 * Add a hole between th_ack (snd_una is not yet set) and snd_max,
873 * if this was a pure cumulative ACK and no data was send beyond
874 * recovery point. Since the data in the socket has not been freed
875 * at this point, we check if the scoreboard is empty, and the ACK
876 * delivered some new data, indicating a full ACK. Also, if the
877 * recovery point is still at snd_max, we are probably application
878 * limited. However, this inference might not always be true. The
879 * rescue retransmission may rarely be slightly premature
880 * compared to RFC6675.
881 * The corresponding ACK+SACK will cause any further outstanding
882 * segments to be retransmitted. This addresses a corner case, when
883 * the trailing packets of a window are lost and no further data
884 * is available for sending.
885 */
886 if ((V_tcp_do_newsack) &&
887 SEQ_LT(th->th_ack, tp->snd_recover) &&
888 TAILQ_EMPTY(&tp->snd_holes) &&
889 (tp->sackhint.delivered_data > 0)) {
890 /*
891 * Exclude FIN sequence space in
892 * the hole for the rescue retransmission,
893 * and also don't create a hole, if only
894 * the ACK for a FIN is outstanding.
895 */
896 tcp_seq highdata = tp->snd_max;
897 if (tp->t_flags & TF_SENTFIN)
898 highdata--;
899 highdata = SEQ_MIN(highdata, tp->snd_recover);
900 if (SEQ_LT(th->th_ack, highdata)) {
901 tp->snd_fack = SEQ_MAX(th->th_ack, tp->snd_fack);
902 (void)tcp_sackhole_insert(tp, SEQ_MAX(th->th_ack,
903 highdata - maxseg), highdata, NULL);
904 }
905 }
906 (void) tcp_output(tp);
907 }
908
909 #if 0
910 /*
911 * Debug version of tcp_sack_output() that walks the scoreboard. Used for
912 * now to sanity check the hint.
913 */
914 static struct sackhole *
915 tcp_sack_output_debug(struct tcpcb *tp, int *sack_bytes_rexmt)
916 {
917 struct sackhole *p;
918
919 INP_WLOCK_ASSERT(tptoinpcb(tp));
920 *sack_bytes_rexmt = 0;
921 TAILQ_FOREACH(p, &tp->snd_holes, scblink) {
922 if (SEQ_LT(p->rxmit, p->end)) {
923 if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */
924 continue;
925 }
926 *sack_bytes_rexmt += (p->rxmit - p->start);
927 break;
928 }
929 *sack_bytes_rexmt += (SEQ_MIN(p->rxmit, p->end) - p->start);
930 }
931 return (p);
932 }
933 #endif
934
935 /*
936 * Returns the next hole to retransmit and the number of retransmitted bytes
937 * from the scoreboard. We store both the next hole and the number of
938 * retransmitted bytes as hints (and recompute these on the fly upon SACK/ACK
939 * reception). This avoids scoreboard traversals completely.
940 *
941 * The loop here will traverse *at most* one link. Here's the argument. For
942 * the loop to traverse more than 1 link before finding the next hole to
943 * retransmit, we would need to have at least 1 node following the current
944 * hint with (rxmit == end). But, for all holes following the current hint,
945 * (start == rxmit), since we have not yet retransmitted from them.
946 * Therefore, in order to traverse more 1 link in the loop below, we need to
947 * have at least one node following the current hint with (start == rxmit ==
948 * end). But that can't happen, (start == end) means that all the data in
949 * that hole has been sacked, in which case, the hole would have been removed
950 * from the scoreboard.
951 */
952 struct sackhole *
tcp_sack_output(struct tcpcb * tp,int * sack_bytes_rexmt)953 tcp_sack_output(struct tcpcb *tp, int *sack_bytes_rexmt)
954 {
955 struct sackhole *hole = NULL;
956
957 INP_WLOCK_ASSERT(tptoinpcb(tp));
958 *sack_bytes_rexmt = tp->sackhint.sack_bytes_rexmit;
959 hole = tp->sackhint.nexthole;
960 if (hole == NULL)
961 return (hole);
962 if (SEQ_GEQ(hole->rxmit, hole->end)) {
963 for (;;) {
964 hole = TAILQ_NEXT(hole, scblink);
965 if (hole == NULL)
966 return (hole);
967 if (SEQ_LT(hole->rxmit, hole->end)) {
968 tp->sackhint.nexthole = hole;
969 break;
970 }
971 }
972 }
973 KASSERT(SEQ_LT(hole->start, hole->end), ("%s: hole.start >= hole.end", __func__));
974 if (!(V_tcp_do_newsack)) {
975 KASSERT(SEQ_LT(hole->start, tp->snd_fack), ("%s: hole.start >= snd.fack", __func__));
976 KASSERT(SEQ_LT(hole->end, tp->snd_fack), ("%s: hole.end >= snd.fack", __func__));
977 KASSERT(SEQ_LT(hole->rxmit, tp->snd_fack), ("%s: hole.rxmit >= snd.fack", __func__));
978 if (SEQ_GEQ(hole->start, hole->end) ||
979 SEQ_GEQ(hole->start, tp->snd_fack) ||
980 SEQ_GEQ(hole->end, tp->snd_fack) ||
981 SEQ_GEQ(hole->rxmit, tp->snd_fack)) {
982 log(LOG_CRIT,"tcp: invalid SACK hole (%u-%u,%u) vs fwd ack %u, ignoring.\n",
983 hole->start, hole->end, hole->rxmit, tp->snd_fack);
984 return (NULL);
985 }
986 }
987 return (hole);
988 }
989
990 /*
991 * After a timeout, the SACK list may be rebuilt. This SACK information
992 * should be used to avoid retransmitting SACKed data. This function
993 * traverses the SACK list to see if snd_nxt should be moved forward.
994 * In addition, cwnd will be inflated by the sacked bytes traversed when
995 * moving snd_nxt forward. This prevents a traffic burst after the final
996 * full ACK, and also keeps ACKs coming back.
997 */
998 int
tcp_sack_adjust(struct tcpcb * tp)999 tcp_sack_adjust(struct tcpcb *tp)
1000 {
1001 int sacked = 0;
1002 struct sackhole *p, *cur = TAILQ_FIRST(&tp->snd_holes);
1003
1004 INP_WLOCK_ASSERT(tptoinpcb(tp));
1005 if (cur == NULL) {
1006 /* No holes */
1007 return (0);
1008 }
1009 if (SEQ_GEQ(tp->snd_nxt, tp->snd_fack)) {
1010 /* We're already beyond any SACKed blocks */
1011 return (tp->sackhint.sacked_bytes);
1012 }
1013 /*
1014 * Two cases for which we want to advance snd_nxt:
1015 * i) snd_nxt lies between end of one hole and beginning of another
1016 * ii) snd_nxt lies between end of last hole and snd_fack
1017 */
1018 while ((p = TAILQ_NEXT(cur, scblink)) != NULL) {
1019 if (SEQ_LT(tp->snd_nxt, cur->end)) {
1020 return (sacked);
1021 }
1022 sacked += p->start - cur->end;
1023 if (SEQ_GEQ(tp->snd_nxt, p->start)) {
1024 cur = p;
1025 } else {
1026 tp->snd_nxt = p->start;
1027 return (sacked);
1028 }
1029 }
1030 if (SEQ_LT(tp->snd_nxt, cur->end)) {
1031 return (sacked);
1032 }
1033 tp->snd_nxt = tp->snd_fack;
1034 return (tp->sackhint.sacked_bytes);
1035 }
1036
1037 /*
1038 * Lost Retransmission Detection
1039 * Check is FACK is beyond the rexmit of the leftmost hole.
1040 * If yes, we restart sending from still existing holes,
1041 * and adjust cwnd via the congestion control module.
1042 */
1043 void
tcp_sack_lost_retransmission(struct tcpcb * tp,struct tcphdr * th)1044 tcp_sack_lost_retransmission(struct tcpcb *tp, struct tcphdr *th)
1045 {
1046 struct sackhole *temp;
1047
1048 if (IN_RECOVERY(tp->t_flags) &&
1049 SEQ_GT(tp->snd_fack, tp->snd_recover) &&
1050 ((temp = TAILQ_FIRST(&tp->snd_holes)) != NULL) &&
1051 SEQ_GEQ(temp->rxmit, temp->end) &&
1052 SEQ_GEQ(tp->snd_fack, temp->rxmit)) {
1053 TCPSTAT_INC(tcps_sack_lostrexmt);
1054 /*
1055 * Start retransmissions from the first hole, and
1056 * subsequently all other remaining holes, including
1057 * those, which had been sent completely before.
1058 */
1059 tp->sackhint.nexthole = temp;
1060 TAILQ_FOREACH(temp, &tp->snd_holes, scblink) {
1061 if (SEQ_GEQ(tp->snd_fack, temp->rxmit) &&
1062 SEQ_GEQ(temp->rxmit, temp->end))
1063 temp->rxmit = temp->start;
1064 }
1065 /*
1066 * Remember the old ssthresh, to deduct the beta factor used
1067 * by the CC module. Finally, set cwnd to ssthresh just
1068 * prior to invoking another cwnd reduction by the CC
1069 * module, to not shrink it excessively.
1070 */
1071 tp->snd_cwnd = tp->snd_ssthresh;
1072 /*
1073 * Formally exit recovery, and let the CC module adjust
1074 * ssthresh as intended.
1075 */
1076 EXIT_RECOVERY(tp->t_flags);
1077 cc_cong_signal(tp, th, CC_NDUPACK);
1078 /*
1079 * For PRR, adjust recover_fs as if this new reduction
1080 * initialized this variable.
1081 * cwnd will be adjusted by SACK or PRR processing
1082 * subsequently, only set it to a safe value here.
1083 */
1084 tp->snd_cwnd = tcp_maxseg(tp);
1085 tp->sackhint.recover_fs = (tp->snd_max - tp->snd_una) -
1086 tp->sackhint.recover_fs;
1087 }
1088 }
1089