xref: /freebsd/sys/netinet/tcp_sack.c (revision f6a3b357e9be4c6423c85eff9a847163a0d307c8)
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 __FBSDID("$FreeBSD$");
77 
78 #include "opt_inet.h"
79 #include "opt_inet6.h"
80 #include "opt_tcpdebug.h"
81 
82 #include <sys/param.h>
83 #include <sys/systm.h>
84 #include <sys/kernel.h>
85 #include <sys/sysctl.h>
86 #include <sys/malloc.h>
87 #include <sys/mbuf.h>
88 #include <sys/proc.h>		/* for proc0 declaration */
89 #include <sys/protosw.h>
90 #include <sys/socket.h>
91 #include <sys/socketvar.h>
92 #include <sys/syslog.h>
93 #include <sys/systm.h>
94 
95 #include <machine/cpu.h>	/* before tcp_seq.h, for tcp_random18() */
96 
97 #include <vm/uma.h>
98 
99 #include <net/if.h>
100 #include <net/if_var.h>
101 #include <net/route.h>
102 #include <net/vnet.h>
103 
104 #include <netinet/in.h>
105 #include <netinet/in_systm.h>
106 #include <netinet/ip.h>
107 #include <netinet/in_var.h>
108 #include <netinet/in_pcb.h>
109 #include <netinet/ip_var.h>
110 #include <netinet/ip6.h>
111 #include <netinet/icmp6.h>
112 #include <netinet6/nd6.h>
113 #include <netinet6/ip6_var.h>
114 #include <netinet6/in6_pcb.h>
115 #include <netinet/tcp.h>
116 #include <netinet/tcp_fsm.h>
117 #include <netinet/tcp_seq.h>
118 #include <netinet/tcp_timer.h>
119 #include <netinet/tcp_var.h>
120 #include <netinet6/tcp6_var.h>
121 #include <netinet/tcpip.h>
122 #ifdef TCPDEBUG
123 #include <netinet/tcp_debug.h>
124 #endif /* TCPDEBUG */
125 
126 #include <machine/in_cksum.h>
127 
128 VNET_DECLARE(struct uma_zone *, sack_hole_zone);
129 #define	V_sack_hole_zone		VNET(sack_hole_zone)
130 
131 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, sack, CTLFLAG_RW, 0, "TCP SACK");
132 VNET_DEFINE(int, tcp_do_sack) = 1;
133 #define	V_tcp_do_sack			VNET(tcp_do_sack)
134 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, enable, CTLFLAG_VNET | CTLFLAG_RW,
135     &VNET_NAME(tcp_do_sack), 0, "Enable/Disable TCP SACK support");
136 
137 VNET_DEFINE(int, tcp_sack_maxholes) = 128;
138 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, maxholes, CTLFLAG_VNET | CTLFLAG_RW,
139     &VNET_NAME(tcp_sack_maxholes), 0,
140     "Maximum number of TCP SACK holes allowed per connection");
141 
142 VNET_DEFINE(int, tcp_sack_globalmaxholes) = 65536;
143 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, globalmaxholes, CTLFLAG_VNET | CTLFLAG_RW,
144     &VNET_NAME(tcp_sack_globalmaxholes), 0,
145     "Global maximum number of TCP SACK holes");
146 
147 VNET_DEFINE(int, tcp_sack_globalholes) = 0;
148 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, globalholes, CTLFLAG_VNET | CTLFLAG_RD,
149     &VNET_NAME(tcp_sack_globalholes), 0,
150     "Global number of TCP SACK holes currently allocated");
151 
152 
153 /*
154  * This function will find overlaps with the currently stored sackblocks
155  * and add any overlap as a dsack block upfront
156  */
157 void
158 tcp_update_dsack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
159 {
160 	struct sackblk head_blk,mid_blk,saved_blks[MAX_SACK_BLKS];
161 	int i, j, n, identical;
162 	tcp_seq start, end;
163 
164 	INP_WLOCK_ASSERT(tp->t_inpcb);
165 
166 	KASSERT(SEQ_LT(rcv_start, rcv_end), ("rcv_start < rcv_end"));
167 
168 	if (tp->t_inpcb->inp_socket->so_options & SO_DEBUG) {
169 		log(LOG_DEBUG, "\nDSACK update: %d..%d, rcv_nxt: %u\n",
170 		rcv_start, rcv_end, tp->rcv_nxt);
171 	}
172 
173 	if (SEQ_LT(rcv_end, tp->rcv_nxt) ||
174 	    ((rcv_end == tp->rcv_nxt) &&
175 	     (tp->rcv_numsacks > 0 ) &&
176 	     (tp->sackblks[0].end == tp->rcv_nxt))) {
177 		saved_blks[0].start = rcv_start;
178 		saved_blks[0].end = rcv_end;
179 	} else {
180 		saved_blks[0].start = saved_blks[0].end = 0;
181 	}
182 
183 	head_blk.start = head_blk.end = 0;
184 	mid_blk.start = rcv_start;
185 	mid_blk.end = rcv_end;
186 	identical = 0;
187 
188 	for (i = 0; i < tp->rcv_numsacks; i++) {
189 		start = tp->sackblks[i].start;
190 		end = tp->sackblks[i].end;
191 		if (SEQ_LT(rcv_end, start)) {
192 			/* pkt left to sack blk */
193 			continue;
194 		}
195 		if (SEQ_GT(rcv_start, end)) {
196 			/* pkt right to sack blk */
197 			continue;
198 		}
199 		if (SEQ_GT(tp->rcv_nxt, end)) {
200 			if ((SEQ_MAX(rcv_start, start) != SEQ_MIN(rcv_end, end)) &&
201 			    (SEQ_GT(head_blk.start, SEQ_MAX(rcv_start, start)) ||
202 			    (head_blk.start == head_blk.end))) {
203 				head_blk.start = SEQ_MAX(rcv_start, start);
204 				head_blk.end = SEQ_MIN(rcv_end, end);
205 			}
206 			continue;
207 		}
208 		if (((head_blk.start == head_blk.end) ||
209 		     SEQ_LT(start, head_blk.start)) &&
210 		     (SEQ_GT(end, rcv_start) &&
211 		      SEQ_LEQ(start, rcv_end))) {
212 			head_blk.start = start;
213 			head_blk.end = end;
214 		}
215 		mid_blk.start = SEQ_MIN(mid_blk.start, start);
216 		mid_blk.end = SEQ_MAX(mid_blk.end, end);
217 		if ((mid_blk.start == start) &&
218 		    (mid_blk.end == end))
219 			identical = 1;
220 	}
221 	if (SEQ_LT(head_blk.start, head_blk.end)) {
222 		/* store overlapping range */
223 		saved_blks[0].start = SEQ_MAX(rcv_start, head_blk.start);
224 		saved_blks[0].end   = SEQ_MIN(rcv_end, head_blk.end);
225 	}
226 	n = 1;
227 	/*
228 	 * Second, if not ACKed, store the SACK block that
229 	 * overlaps with the DSACK block unless it is identical
230 	 */
231 	if ((SEQ_LT(tp->rcv_nxt, mid_blk.end) &&
232 	    !((mid_blk.start == saved_blks[0].start) &&
233 	    (mid_blk.end == saved_blks[0].end))) ||
234 	    identical == 1) {
235 		saved_blks[n].start = mid_blk.start;
236 		saved_blks[n++].end = mid_blk.end;
237 	}
238 	for (j = 0; (j < tp->rcv_numsacks) && (j < MAX_SACK_BLKS-1); j++) {
239 		if (((SEQ_LT(tp->sackblks[j].end, mid_blk.start) ||
240 		      SEQ_GT(tp->sackblks[j].start, mid_blk.end)) &&
241 		    (SEQ_GT(tp->sackblks[j].start, tp->rcv_nxt))))
242 		saved_blks[n++] = tp->sackblks[j];
243 	}
244 	j = 0;
245 	for (i = 0; i < n; i++) {
246 		/* we can end up with a stale inital entry */
247 		if (SEQ_LT(saved_blks[i].start, saved_blks[i].end)) {
248 			tp->sackblks[j++] = saved_blks[i];
249 		}
250 	}
251 	tp->rcv_numsacks = j;
252 }
253 
254 /*
255  * This function is called upon receipt of new valid data (while not in
256  * header prediction mode), and it updates the ordered list of sacks.
257  */
258 void
259 tcp_update_sack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
260 {
261 	/*
262 	 * First reported block MUST be the most recent one.  Subsequent
263 	 * blocks SHOULD be in the order in which they arrived at the
264 	 * receiver.  These two conditions make the implementation fully
265 	 * compliant with RFC 2018.
266 	 */
267 	struct sackblk head_blk, saved_blks[MAX_SACK_BLKS];
268 	int num_head, num_saved, i;
269 
270 	INP_WLOCK_ASSERT(tp->t_inpcb);
271 
272 	/* Check arguments. */
273 	KASSERT(SEQ_LEQ(rcv_start, rcv_end), ("rcv_start <= rcv_end"));
274 
275 	if ((rcv_start == rcv_end) &&
276 	    (tp->rcv_numsacks >= 1) &&
277 	    (rcv_end == tp->sackblks[0].end)) {
278 		/* retaining DSACK block below rcv_nxt (todrop) */
279 		head_blk = tp->sackblks[0];
280 	} else {
281 		/* SACK block for the received segment. */
282 		head_blk.start = rcv_start;
283 		head_blk.end = rcv_end;
284 	}
285 
286 	/*
287 	 * Merge updated SACK blocks into head_blk, and save unchanged SACK
288 	 * blocks into saved_blks[].  num_saved will have the number of the
289 	 * saved SACK blocks.
290 	 */
291 	num_saved = 0;
292 	for (i = 0; i < tp->rcv_numsacks; i++) {
293 		tcp_seq start = tp->sackblks[i].start;
294 		tcp_seq end = tp->sackblks[i].end;
295 		if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
296 			/*
297 			 * Discard this SACK block.
298 			 */
299 		} else if (SEQ_LEQ(head_blk.start, end) &&
300 			   SEQ_GEQ(head_blk.end, start)) {
301 			/*
302 			 * Merge this SACK block into head_blk.  This SACK
303 			 * block itself will be discarded.
304 			 */
305 			/*
306 			 * |-|
307 			 *   |---|  merge
308 			 *
309 			 *     |-|
310 			 * |---|    merge
311 			 *
312 			 * |-----|
313 			 *   |-|    DSACK smaller
314 			 *
315 			 *   |-|
316 			 * |-----|  DSACK smaller
317 			 */
318 			if (head_blk.start == end)
319 				head_blk.start = start;
320 			else if (head_blk.end == start)
321 				head_blk.end = end;
322 			else {
323 				if (SEQ_LT(head_blk.start, start)) {
324 					tcp_seq temp = start;
325 					start = head_blk.start;
326 					head_blk.start = temp;
327 				}
328 				if (SEQ_GT(head_blk.end, end)) {
329 					tcp_seq temp = end;
330 					end = head_blk.end;
331 					head_blk.end = temp;
332 				}
333 				if ((head_blk.start != start) ||
334 				    (head_blk.end != end)) {
335 					if ((num_saved >= 1) &&
336 					   SEQ_GEQ(saved_blks[num_saved-1].start, start) &&
337 					   SEQ_LEQ(saved_blks[num_saved-1].end, end))
338 						num_saved--;
339 					saved_blks[num_saved].start = start;
340 					saved_blks[num_saved].end = end;
341 					num_saved++;
342 				}
343 			}
344 		} else {
345 			/*
346 			 * This block supercedes the prior block
347 			 */
348 			if ((num_saved >= 1) &&
349 			   SEQ_GEQ(saved_blks[num_saved-1].start, start) &&
350 			   SEQ_LEQ(saved_blks[num_saved-1].end, end))
351 				num_saved--;
352 			/*
353 			 * Save this SACK block.
354 			 */
355 			saved_blks[num_saved].start = start;
356 			saved_blks[num_saved].end = end;
357 			num_saved++;
358 		}
359 	}
360 
361 	/*
362 	 * Update SACK list in tp->sackblks[].
363 	 */
364 	num_head = 0;
365 	if (SEQ_LT(rcv_start, rcv_end)) {
366 		/*
367 		 * The received data segment is an out-of-order segment.  Put
368 		 * head_blk at the top of SACK list.
369 		 */
370 		tp->sackblks[0] = head_blk;
371 		num_head = 1;
372 		/*
373 		 * If the number of saved SACK blocks exceeds its limit,
374 		 * discard the last SACK block.
375 		 */
376 		if (num_saved >= MAX_SACK_BLKS)
377 			num_saved--;
378 	}
379 	if ((rcv_start == rcv_end) &&
380 	    (rcv_start == tp->sackblks[0].end)) {
381 		num_head = 1;
382 	}
383 	if (num_saved > 0) {
384 		/*
385 		 * Copy the saved SACK blocks back.
386 		 */
387 		bcopy(saved_blks, &tp->sackblks[num_head],
388 		      sizeof(struct sackblk) * num_saved);
389 	}
390 
391 	/* Save the number of SACK blocks. */
392 	tp->rcv_numsacks = num_head + num_saved;
393 }
394 
395 void
396 tcp_clean_dsack_blocks(struct tcpcb *tp)
397 {
398 	struct sackblk saved_blks[MAX_SACK_BLKS];
399 	int num_saved, i;
400 
401 	INP_WLOCK_ASSERT(tp->t_inpcb);
402 	/*
403 	 * Clean up any DSACK blocks that
404 	 * are in our queue of sack blocks.
405 	 *
406 	 */
407 	num_saved = 0;
408 	for (i = 0; i < tp->rcv_numsacks; i++) {
409 		tcp_seq start = tp->sackblks[i].start;
410 		tcp_seq end = tp->sackblks[i].end;
411 		if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
412 			/*
413 			 * Discard this D-SACK block.
414 			 */
415 			continue;
416 		}
417 		/*
418 		 * Save this SACK block.
419 		 */
420 		saved_blks[num_saved].start = start;
421 		saved_blks[num_saved].end = end;
422 		num_saved++;
423 	}
424 	if (num_saved > 0) {
425 		/*
426 		 * Copy the saved SACK blocks back.
427 		 */
428 		bcopy(saved_blks, &tp->sackblks[0],
429 		      sizeof(struct sackblk) * num_saved);
430 	}
431 	tp->rcv_numsacks = num_saved;
432 }
433 
434 /*
435  * Delete all receiver-side SACK information.
436  */
437 void
438 tcp_clean_sackreport(struct tcpcb *tp)
439 {
440 	int i;
441 
442 	INP_WLOCK_ASSERT(tp->t_inpcb);
443 	tp->rcv_numsacks = 0;
444 	for (i = 0; i < MAX_SACK_BLKS; i++)
445 		tp->sackblks[i].start = tp->sackblks[i].end=0;
446 }
447 
448 /*
449  * Allocate struct sackhole.
450  */
451 static struct sackhole *
452 tcp_sackhole_alloc(struct tcpcb *tp, tcp_seq start, tcp_seq end)
453 {
454 	struct sackhole *hole;
455 
456 	if (tp->snd_numholes >= V_tcp_sack_maxholes ||
457 	    V_tcp_sack_globalholes >= V_tcp_sack_globalmaxholes) {
458 		TCPSTAT_INC(tcps_sack_sboverflow);
459 		return NULL;
460 	}
461 
462 	hole = (struct sackhole *)uma_zalloc(V_sack_hole_zone, M_NOWAIT);
463 	if (hole == NULL)
464 		return NULL;
465 
466 	hole->start = start;
467 	hole->end = end;
468 	hole->rxmit = start;
469 
470 	tp->snd_numholes++;
471 	atomic_add_int(&V_tcp_sack_globalholes, 1);
472 
473 	return hole;
474 }
475 
476 /*
477  * Free struct sackhole.
478  */
479 static void
480 tcp_sackhole_free(struct tcpcb *tp, struct sackhole *hole)
481 {
482 
483 	uma_zfree(V_sack_hole_zone, hole);
484 
485 	tp->snd_numholes--;
486 	atomic_subtract_int(&V_tcp_sack_globalholes, 1);
487 
488 	KASSERT(tp->snd_numholes >= 0, ("tp->snd_numholes >= 0"));
489 	KASSERT(V_tcp_sack_globalholes >= 0, ("tcp_sack_globalholes >= 0"));
490 }
491 
492 /*
493  * Insert new SACK hole into scoreboard.
494  */
495 static struct sackhole *
496 tcp_sackhole_insert(struct tcpcb *tp, tcp_seq start, tcp_seq end,
497     struct sackhole *after)
498 {
499 	struct sackhole *hole;
500 
501 	/* Allocate a new SACK hole. */
502 	hole = tcp_sackhole_alloc(tp, start, end);
503 	if (hole == NULL)
504 		return NULL;
505 
506 	/* Insert the new SACK hole into scoreboard. */
507 	if (after != NULL)
508 		TAILQ_INSERT_AFTER(&tp->snd_holes, after, hole, scblink);
509 	else
510 		TAILQ_INSERT_TAIL(&tp->snd_holes, hole, scblink);
511 
512 	/* Update SACK hint. */
513 	if (tp->sackhint.nexthole == NULL)
514 		tp->sackhint.nexthole = hole;
515 
516 	return hole;
517 }
518 
519 /*
520  * Remove SACK hole from scoreboard.
521  */
522 static void
523 tcp_sackhole_remove(struct tcpcb *tp, struct sackhole *hole)
524 {
525 
526 	/* Update SACK hint. */
527 	if (tp->sackhint.nexthole == hole)
528 		tp->sackhint.nexthole = TAILQ_NEXT(hole, scblink);
529 
530 	/* Remove this SACK hole. */
531 	TAILQ_REMOVE(&tp->snd_holes, hole, scblink);
532 
533 	/* Free this SACK hole. */
534 	tcp_sackhole_free(tp, hole);
535 }
536 
537 /*
538  * Process cumulative ACK and the TCP SACK option to update the scoreboard.
539  * tp->snd_holes is an ordered list of holes (oldest to newest, in terms of
540  * the sequence space).
541  * Returns 1 if incoming ACK has previously unknown SACK information,
542  * 0 otherwise. Note: We treat (snd_una, th_ack) as a sack block so any changes
543  * to that (i.e. left edge moving) would also be considered a change in SACK
544  * information which is slightly different than rfc6675.
545  */
546 int
547 tcp_sack_doack(struct tcpcb *tp, struct tcpopt *to, tcp_seq th_ack)
548 {
549 	struct sackhole *cur, *temp;
550 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1], *sblkp;
551 	int i, j, num_sack_blks, sack_changed;
552 
553 	INP_WLOCK_ASSERT(tp->t_inpcb);
554 
555 	num_sack_blks = 0;
556 	sack_changed = 0;
557 	/*
558 	 * If SND.UNA will be advanced by SEG.ACK, and if SACK holes exist,
559 	 * treat [SND.UNA, SEG.ACK) as if it is a SACK block.
560 	 */
561 	if (SEQ_LT(tp->snd_una, th_ack) && !TAILQ_EMPTY(&tp->snd_holes)) {
562 		sack_blocks[num_sack_blks].start = tp->snd_una;
563 		sack_blocks[num_sack_blks++].end = th_ack;
564 	}
565 	/*
566 	 * Append received valid SACK blocks to sack_blocks[], but only if we
567 	 * received new blocks from the other side.
568 	 */
569 	if (to->to_flags & TOF_SACK) {
570 		tp->sackhint.sacked_bytes = 0;	/* reset */
571 		for (i = 0; i < to->to_nsacks; i++) {
572 			bcopy((to->to_sacks + i * TCPOLEN_SACK),
573 			    &sack, sizeof(sack));
574 			sack.start = ntohl(sack.start);
575 			sack.end = ntohl(sack.end);
576 			if (SEQ_GT(sack.end, sack.start) &&
577 			    SEQ_GT(sack.start, tp->snd_una) &&
578 			    SEQ_GT(sack.start, th_ack) &&
579 			    SEQ_LT(sack.start, tp->snd_max) &&
580 			    SEQ_GT(sack.end, tp->snd_una) &&
581 			    SEQ_LEQ(sack.end, tp->snd_max)) {
582 				sack_blocks[num_sack_blks++] = sack;
583 				tp->sackhint.sacked_bytes +=
584 				    (sack.end-sack.start);
585 			}
586 		}
587 	}
588 	/*
589 	 * Return if SND.UNA is not advanced and no valid SACK block is
590 	 * received.
591 	 */
592 	if (num_sack_blks == 0)
593 		return (sack_changed);
594 
595 	/*
596 	 * Sort the SACK blocks so we can update the scoreboard with just one
597 	 * pass. The overhead of sorting up to 4+1 elements is less than
598 	 * making up to 4+1 passes over the scoreboard.
599 	 */
600 	for (i = 0; i < num_sack_blks; i++) {
601 		for (j = i + 1; j < num_sack_blks; j++) {
602 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
603 				sack = sack_blocks[i];
604 				sack_blocks[i] = sack_blocks[j];
605 				sack_blocks[j] = sack;
606 			}
607 		}
608 	}
609 	if (TAILQ_EMPTY(&tp->snd_holes))
610 		/*
611 		 * Empty scoreboard. Need to initialize snd_fack (it may be
612 		 * uninitialized or have a bogus value). Scoreboard holes
613 		 * (from the sack blocks received) are created later below
614 		 * (in the logic that adds holes to the tail of the
615 		 * scoreboard).
616 		 */
617 		tp->snd_fack = SEQ_MAX(tp->snd_una, th_ack);
618 	/*
619 	 * In the while-loop below, incoming SACK blocks (sack_blocks[]) and
620 	 * SACK holes (snd_holes) are traversed from their tails with just
621 	 * one pass in order to reduce the number of compares especially when
622 	 * the bandwidth-delay product is large.
623 	 *
624 	 * Note: Typically, in the first RTT of SACK recovery, the highest
625 	 * three or four SACK blocks with the same ack number are received.
626 	 * In the second RTT, if retransmitted data segments are not lost,
627 	 * the highest three or four SACK blocks with ack number advancing
628 	 * are received.
629 	 */
630 	sblkp = &sack_blocks[num_sack_blks - 1];	/* Last SACK block */
631 	tp->sackhint.last_sack_ack = sblkp->end;
632 	if (SEQ_LT(tp->snd_fack, sblkp->start)) {
633 		/*
634 		 * The highest SACK block is beyond fack.  Append new SACK
635 		 * hole at the tail.  If the second or later highest SACK
636 		 * blocks are also beyond the current fack, they will be
637 		 * inserted by way of hole splitting in the while-loop below.
638 		 */
639 		temp = tcp_sackhole_insert(tp, tp->snd_fack,sblkp->start,NULL);
640 		if (temp != NULL) {
641 			tp->snd_fack = sblkp->end;
642 			/* Go to the previous sack block. */
643 			sblkp--;
644 			sack_changed = 1;
645 		} else {
646 			/*
647 			 * We failed to add a new hole based on the current
648 			 * sack block.  Skip over all the sack blocks that
649 			 * fall completely to the right of snd_fack and
650 			 * proceed to trim the scoreboard based on the
651 			 * remaining sack blocks.  This also trims the
652 			 * scoreboard for th_ack (which is sack_blocks[0]).
653 			 */
654 			while (sblkp >= sack_blocks &&
655 			       SEQ_LT(tp->snd_fack, sblkp->start))
656 				sblkp--;
657 			if (sblkp >= sack_blocks &&
658 			    SEQ_LT(tp->snd_fack, sblkp->end))
659 				tp->snd_fack = sblkp->end;
660 		}
661 	} else if (SEQ_LT(tp->snd_fack, sblkp->end)) {
662 		/* fack is advanced. */
663 		tp->snd_fack = sblkp->end;
664 		sack_changed = 1;
665 	}
666 	cur = TAILQ_LAST(&tp->snd_holes, sackhole_head); /* Last SACK hole. */
667 	/*
668 	 * Since the incoming sack blocks are sorted, we can process them
669 	 * making one sweep of the scoreboard.
670 	 */
671 	while (sblkp >= sack_blocks  && cur != NULL) {
672 		if (SEQ_GEQ(sblkp->start, cur->end)) {
673 			/*
674 			 * SACKs data beyond the current hole.  Go to the
675 			 * previous sack block.
676 			 */
677 			sblkp--;
678 			continue;
679 		}
680 		if (SEQ_LEQ(sblkp->end, cur->start)) {
681 			/*
682 			 * SACKs data before the current hole.  Go to the
683 			 * previous hole.
684 			 */
685 			cur = TAILQ_PREV(cur, sackhole_head, scblink);
686 			continue;
687 		}
688 		tp->sackhint.sack_bytes_rexmit -= (cur->rxmit - cur->start);
689 		KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
690 		    ("sackhint bytes rtx >= 0"));
691 		sack_changed = 1;
692 		if (SEQ_LEQ(sblkp->start, cur->start)) {
693 			/* Data acks at least the beginning of hole. */
694 			if (SEQ_GEQ(sblkp->end, cur->end)) {
695 				/* Acks entire hole, so delete hole. */
696 				temp = cur;
697 				cur = TAILQ_PREV(cur, sackhole_head, scblink);
698 				tcp_sackhole_remove(tp, temp);
699 				/*
700 				 * The sack block may ack all or part of the
701 				 * next hole too, so continue onto the next
702 				 * hole.
703 				 */
704 				continue;
705 			} else {
706 				/* Move start of hole forward. */
707 				cur->start = sblkp->end;
708 				cur->rxmit = SEQ_MAX(cur->rxmit, cur->start);
709 			}
710 		} else {
711 			/* Data acks at least the end of hole. */
712 			if (SEQ_GEQ(sblkp->end, cur->end)) {
713 				/* Move end of hole backward. */
714 				cur->end = sblkp->start;
715 				cur->rxmit = SEQ_MIN(cur->rxmit, cur->end);
716 			} else {
717 				/*
718 				 * ACKs some data in middle of a hole; need
719 				 * to split current hole
720 				 */
721 				temp = tcp_sackhole_insert(tp, sblkp->end,
722 				    cur->end, cur);
723 				if (temp != NULL) {
724 					if (SEQ_GT(cur->rxmit, temp->rxmit)) {
725 						temp->rxmit = cur->rxmit;
726 						tp->sackhint.sack_bytes_rexmit
727 						    += (temp->rxmit
728 						    - temp->start);
729 					}
730 					cur->end = sblkp->start;
731 					cur->rxmit = SEQ_MIN(cur->rxmit,
732 					    cur->end);
733 				}
734 			}
735 		}
736 		tp->sackhint.sack_bytes_rexmit += (cur->rxmit - cur->start);
737 		/*
738 		 * Testing sblkp->start against cur->start tells us whether
739 		 * we're done with the sack block or the sack hole.
740 		 * Accordingly, we advance one or the other.
741 		 */
742 		if (SEQ_LEQ(sblkp->start, cur->start))
743 			cur = TAILQ_PREV(cur, sackhole_head, scblink);
744 		else
745 			sblkp--;
746 	}
747 	return (sack_changed);
748 }
749 
750 /*
751  * Free all SACK holes to clear the scoreboard.
752  */
753 void
754 tcp_free_sackholes(struct tcpcb *tp)
755 {
756 	struct sackhole *q;
757 
758 	INP_WLOCK_ASSERT(tp->t_inpcb);
759 	while ((q = TAILQ_FIRST(&tp->snd_holes)) != NULL)
760 		tcp_sackhole_remove(tp, q);
761 	tp->sackhint.sack_bytes_rexmit = 0;
762 
763 	KASSERT(tp->snd_numholes == 0, ("tp->snd_numholes == 0"));
764 	KASSERT(tp->sackhint.nexthole == NULL,
765 		("tp->sackhint.nexthole == NULL"));
766 }
767 
768 /*
769  * Partial ack handling within a sack recovery episode.  Keeping this very
770  * simple for now.  When a partial ack is received, force snd_cwnd to a value
771  * that will allow the sender to transmit no more than 2 segments.  If
772  * necessary, a better scheme can be adopted at a later point, but for now,
773  * the goal is to prevent the sender from bursting a large amount of data in
774  * the midst of sack recovery.
775  */
776 void
777 tcp_sack_partialack(struct tcpcb *tp, struct tcphdr *th)
778 {
779 	int num_segs = 1;
780 
781 	INP_WLOCK_ASSERT(tp->t_inpcb);
782 	tcp_timer_activate(tp, TT_REXMT, 0);
783 	tp->t_rtttime = 0;
784 	/* Send one or 2 segments based on how much new data was acked. */
785 	if ((BYTES_THIS_ACK(tp, th) / tp->t_maxseg) >= 2)
786 		num_segs = 2;
787 	tp->snd_cwnd = (tp->sackhint.sack_bytes_rexmit +
788 	    (tp->snd_nxt - tp->sack_newdata) + num_segs * tp->t_maxseg);
789 	if (tp->snd_cwnd > tp->snd_ssthresh)
790 		tp->snd_cwnd = tp->snd_ssthresh;
791 	tp->t_flags |= TF_ACKNOW;
792 	(void) tp->t_fb->tfb_tcp_output(tp);
793 }
794 
795 #if 0
796 /*
797  * Debug version of tcp_sack_output() that walks the scoreboard.  Used for
798  * now to sanity check the hint.
799  */
800 static struct sackhole *
801 tcp_sack_output_debug(struct tcpcb *tp, int *sack_bytes_rexmt)
802 {
803 	struct sackhole *p;
804 
805 	INP_WLOCK_ASSERT(tp->t_inpcb);
806 	*sack_bytes_rexmt = 0;
807 	TAILQ_FOREACH(p, &tp->snd_holes, scblink) {
808 		if (SEQ_LT(p->rxmit, p->end)) {
809 			if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */
810 				continue;
811 			}
812 			*sack_bytes_rexmt += (p->rxmit - p->start);
813 			break;
814 		}
815 		*sack_bytes_rexmt += (p->rxmit - p->start);
816 	}
817 	return (p);
818 }
819 #endif
820 
821 /*
822  * Returns the next hole to retransmit and the number of retransmitted bytes
823  * from the scoreboard.  We store both the next hole and the number of
824  * retransmitted bytes as hints (and recompute these on the fly upon SACK/ACK
825  * reception).  This avoids scoreboard traversals completely.
826  *
827  * The loop here will traverse *at most* one link.  Here's the argument.  For
828  * the loop to traverse more than 1 link before finding the next hole to
829  * retransmit, we would need to have at least 1 node following the current
830  * hint with (rxmit == end).  But, for all holes following the current hint,
831  * (start == rxmit), since we have not yet retransmitted from them.
832  * Therefore, in order to traverse more 1 link in the loop below, we need to
833  * have at least one node following the current hint with (start == rxmit ==
834  * end).  But that can't happen, (start == end) means that all the data in
835  * that hole has been sacked, in which case, the hole would have been removed
836  * from the scoreboard.
837  */
838 struct sackhole *
839 tcp_sack_output(struct tcpcb *tp, int *sack_bytes_rexmt)
840 {
841 	struct sackhole *hole = NULL;
842 
843 	INP_WLOCK_ASSERT(tp->t_inpcb);
844 	*sack_bytes_rexmt = tp->sackhint.sack_bytes_rexmit;
845 	hole = tp->sackhint.nexthole;
846 	if (hole == NULL || SEQ_LT(hole->rxmit, hole->end))
847 		goto out;
848 	while ((hole = TAILQ_NEXT(hole, scblink)) != NULL) {
849 		if (SEQ_LT(hole->rxmit, hole->end)) {
850 			tp->sackhint.nexthole = hole;
851 			break;
852 		}
853 	}
854 out:
855 	return (hole);
856 }
857 
858 /*
859  * After a timeout, the SACK list may be rebuilt.  This SACK information
860  * should be used to avoid retransmitting SACKed data.  This function
861  * traverses the SACK list to see if snd_nxt should be moved forward.
862  */
863 void
864 tcp_sack_adjust(struct tcpcb *tp)
865 {
866 	struct sackhole *p, *cur = TAILQ_FIRST(&tp->snd_holes);
867 
868 	INP_WLOCK_ASSERT(tp->t_inpcb);
869 	if (cur == NULL)
870 		return; /* No holes */
871 	if (SEQ_GEQ(tp->snd_nxt, tp->snd_fack))
872 		return; /* We're already beyond any SACKed blocks */
873 	/*-
874 	 * Two cases for which we want to advance snd_nxt:
875 	 * i) snd_nxt lies between end of one hole and beginning of another
876 	 * ii) snd_nxt lies between end of last hole and snd_fack
877 	 */
878 	while ((p = TAILQ_NEXT(cur, scblink)) != NULL) {
879 		if (SEQ_LT(tp->snd_nxt, cur->end))
880 			return;
881 		if (SEQ_GEQ(tp->snd_nxt, p->start))
882 			cur = p;
883 		else {
884 			tp->snd_nxt = p->start;
885 			return;
886 		}
887 	}
888 	if (SEQ_LT(tp->snd_nxt, cur->end))
889 		return;
890 	tp->snd_nxt = tp->snd_fack;
891 }
892