xref: /freebsd/sys/netinet/tcp_sack.c (revision 640235e2c2ba32947f7c59d168437ffa1280f1e6)
1 /*-
2  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995
3  *	The Regents of the University of California.
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  * 4. Neither the name of the University nor the names of its contributors
15  *    may be used to endorse or promote products derived from this software
16  *    without specific prior written permission.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  *
30  *	@(#)tcp_sack.c	8.12 (Berkeley) 5/24/95
31  */
32 
33 /*-
34  *	@@(#)COPYRIGHT	1.1 (NRL) 17 January 1995
35  *
36  * NRL grants permission for redistribution and use in source and binary
37  * forms, with or without modification, of the software and documentation
38  * created at NRL provided that the following conditions are met:
39  *
40  * 1. Redistributions of source code must retain the above copyright
41  *    notice, this list of conditions and the following disclaimer.
42  * 2. Redistributions in binary form must reproduce the above copyright
43  *    notice, this list of conditions and the following disclaimer in the
44  *    documentation and/or other materials provided with the distribution.
45  * 3. All advertising materials mentioning features or use of this software
46  *    must display the following acknowledgements:
47  *	This product includes software developed by the University of
48  *	California, Berkeley and its contributors.
49  *	This product includes software developed at the Information
50  *	Technology Division, US Naval Research Laboratory.
51  * 4. Neither the name of the NRL nor the names of its contributors
52  *    may be used to endorse or promote products derived from this software
53  *    without specific prior written permission.
54  *
55  * THE SOFTWARE PROVIDED BY NRL IS PROVIDED BY NRL AND CONTRIBUTORS ``AS
56  * IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
57  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
58  * PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL NRL OR
59  * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
60  * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
61  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
62  * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
63  * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
64  * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
65  * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
66  *
67  * The views and conclusions contained in the software and documentation
68  * are those of the authors and should not be interpreted as representing
69  * official policies, either expressed or implied, of the US Naval
70  * Research Laboratory (NRL).
71  */
72 
73 #include <sys/cdefs.h>
74 __FBSDID("$FreeBSD$");
75 
76 #include "opt_inet.h"
77 #include "opt_inet6.h"
78 #include "opt_tcpdebug.h"
79 
80 #include <sys/param.h>
81 #include <sys/systm.h>
82 #include <sys/kernel.h>
83 #include <sys/sysctl.h>
84 #include <sys/malloc.h>
85 #include <sys/mbuf.h>
86 #include <sys/proc.h>		/* for proc0 declaration */
87 #include <sys/protosw.h>
88 #include <sys/socket.h>
89 #include <sys/socketvar.h>
90 #include <sys/syslog.h>
91 #include <sys/systm.h>
92 
93 #include <machine/cpu.h>	/* before tcp_seq.h, for tcp_random18() */
94 
95 #include <vm/uma.h>
96 
97 #include <net/if.h>
98 #include <net/if_var.h>
99 #include <net/route.h>
100 #include <net/vnet.h>
101 
102 #include <netinet/in.h>
103 #include <netinet/in_systm.h>
104 #include <netinet/ip.h>
105 #include <netinet/in_var.h>
106 #include <netinet/in_pcb.h>
107 #include <netinet/ip_var.h>
108 #include <netinet/ip6.h>
109 #include <netinet/icmp6.h>
110 #include <netinet6/nd6.h>
111 #include <netinet6/ip6_var.h>
112 #include <netinet6/in6_pcb.h>
113 #include <netinet/tcp.h>
114 #include <netinet/tcp_fsm.h>
115 #include <netinet/tcp_seq.h>
116 #include <netinet/tcp_timer.h>
117 #include <netinet/tcp_var.h>
118 #include <netinet6/tcp6_var.h>
119 #include <netinet/tcpip.h>
120 #ifdef TCPDEBUG
121 #include <netinet/tcp_debug.h>
122 #endif /* TCPDEBUG */
123 
124 #include <machine/in_cksum.h>
125 
126 VNET_DECLARE(struct uma_zone *, sack_hole_zone);
127 #define	V_sack_hole_zone		VNET(sack_hole_zone)
128 
129 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, sack, CTLFLAG_RW, 0, "TCP SACK");
130 VNET_DEFINE(int, tcp_do_sack) = 1;
131 #define	V_tcp_do_sack			VNET(tcp_do_sack)
132 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, enable, CTLFLAG_VNET | CTLFLAG_RW,
133     &VNET_NAME(tcp_do_sack), 0, "Enable/Disable TCP SACK support");
134 
135 VNET_DEFINE(int, tcp_sack_maxholes) = 128;
136 #define	V_tcp_sack_maxholes		VNET(tcp_sack_maxholes)
137 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, maxholes, CTLFLAG_VNET | CTLFLAG_RW,
138     &VNET_NAME(tcp_sack_maxholes), 0,
139     "Maximum number of TCP SACK holes allowed per connection");
140 
141 VNET_DEFINE(int, tcp_sack_globalmaxholes) = 65536;
142 #define	V_tcp_sack_globalmaxholes	VNET(tcp_sack_globalmaxholes)
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 #define	V_tcp_sack_globalholes		VNET(tcp_sack_globalholes)
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 /*
154  * This function is called upon receipt of new valid data (while not in
155  * header prediction mode), and it updates the ordered list of sacks.
156  */
157 void
158 tcp_update_sack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
159 {
160 	/*
161 	 * First reported block MUST be the most recent one.  Subsequent
162 	 * blocks SHOULD be in the order in which they arrived at the
163 	 * receiver.  These two conditions make the implementation fully
164 	 * compliant with RFC 2018.
165 	 */
166 	struct sackblk head_blk, saved_blks[MAX_SACK_BLKS];
167 	int num_head, num_saved, i;
168 
169 	INP_WLOCK_ASSERT(tp->t_inpcb);
170 
171 	/* Check arguments. */
172 	KASSERT(SEQ_LT(rcv_start, rcv_end), ("rcv_start < rcv_end"));
173 
174 	/* SACK block for the received segment. */
175 	head_blk.start = rcv_start;
176 	head_blk.end = rcv_end;
177 
178 	/*
179 	 * Merge updated SACK blocks into head_blk, and save unchanged SACK
180 	 * blocks into saved_blks[].  num_saved will have the number of the
181 	 * saved SACK blocks.
182 	 */
183 	num_saved = 0;
184 	for (i = 0; i < tp->rcv_numsacks; i++) {
185 		tcp_seq start = tp->sackblks[i].start;
186 		tcp_seq end = tp->sackblks[i].end;
187 		if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
188 			/*
189 			 * Discard this SACK block.
190 			 */
191 		} else if (SEQ_LEQ(head_blk.start, end) &&
192 			   SEQ_GEQ(head_blk.end, start)) {
193 			/*
194 			 * Merge this SACK block into head_blk.  This SACK
195 			 * block itself will be discarded.
196 			 */
197 			if (SEQ_GT(head_blk.start, start))
198 				head_blk.start = start;
199 			if (SEQ_LT(head_blk.end, end))
200 				head_blk.end = end;
201 		} else {
202 			/*
203 			 * Save this SACK block.
204 			 */
205 			saved_blks[num_saved].start = start;
206 			saved_blks[num_saved].end = end;
207 			num_saved++;
208 		}
209 	}
210 
211 	/*
212 	 * Update SACK list in tp->sackblks[].
213 	 */
214 	num_head = 0;
215 	if (SEQ_GT(head_blk.start, tp->rcv_nxt)) {
216 		/*
217 		 * The received data segment is an out-of-order segment.  Put
218 		 * head_blk at the top of SACK list.
219 		 */
220 		tp->sackblks[0] = head_blk;
221 		num_head = 1;
222 		/*
223 		 * If the number of saved SACK blocks exceeds its limit,
224 		 * discard the last SACK block.
225 		 */
226 		if (num_saved >= MAX_SACK_BLKS)
227 			num_saved--;
228 	}
229 	if (num_saved > 0) {
230 		/*
231 		 * Copy the saved SACK blocks back.
232 		 */
233 		bcopy(saved_blks, &tp->sackblks[num_head],
234 		      sizeof(struct sackblk) * num_saved);
235 	}
236 
237 	/* Save the number of SACK blocks. */
238 	tp->rcv_numsacks = num_head + num_saved;
239 }
240 
241 /*
242  * Delete all receiver-side SACK information.
243  */
244 void
245 tcp_clean_sackreport(struct tcpcb *tp)
246 {
247 	int i;
248 
249 	INP_WLOCK_ASSERT(tp->t_inpcb);
250 	tp->rcv_numsacks = 0;
251 	for (i = 0; i < MAX_SACK_BLKS; i++)
252 		tp->sackblks[i].start = tp->sackblks[i].end=0;
253 }
254 
255 /*
256  * Allocate struct sackhole.
257  */
258 static struct sackhole *
259 tcp_sackhole_alloc(struct tcpcb *tp, tcp_seq start, tcp_seq end)
260 {
261 	struct sackhole *hole;
262 
263 	if (tp->snd_numholes >= V_tcp_sack_maxholes ||
264 	    V_tcp_sack_globalholes >= V_tcp_sack_globalmaxholes) {
265 		TCPSTAT_INC(tcps_sack_sboverflow);
266 		return NULL;
267 	}
268 
269 	hole = (struct sackhole *)uma_zalloc(V_sack_hole_zone, M_NOWAIT);
270 	if (hole == NULL)
271 		return NULL;
272 
273 	hole->start = start;
274 	hole->end = end;
275 	hole->rxmit = start;
276 
277 	tp->snd_numholes++;
278 	atomic_add_int(&V_tcp_sack_globalholes, 1);
279 
280 	return hole;
281 }
282 
283 /*
284  * Free struct sackhole.
285  */
286 static void
287 tcp_sackhole_free(struct tcpcb *tp, struct sackhole *hole)
288 {
289 
290 	uma_zfree(V_sack_hole_zone, hole);
291 
292 	tp->snd_numholes--;
293 	atomic_subtract_int(&V_tcp_sack_globalholes, 1);
294 
295 	KASSERT(tp->snd_numholes >= 0, ("tp->snd_numholes >= 0"));
296 	KASSERT(V_tcp_sack_globalholes >= 0, ("tcp_sack_globalholes >= 0"));
297 }
298 
299 /*
300  * Insert new SACK hole into scoreboard.
301  */
302 static struct sackhole *
303 tcp_sackhole_insert(struct tcpcb *tp, tcp_seq start, tcp_seq end,
304     struct sackhole *after)
305 {
306 	struct sackhole *hole;
307 
308 	/* Allocate a new SACK hole. */
309 	hole = tcp_sackhole_alloc(tp, start, end);
310 	if (hole == NULL)
311 		return NULL;
312 
313 	/* Insert the new SACK hole into scoreboard. */
314 	if (after != NULL)
315 		TAILQ_INSERT_AFTER(&tp->snd_holes, after, hole, scblink);
316 	else
317 		TAILQ_INSERT_TAIL(&tp->snd_holes, hole, scblink);
318 
319 	/* Update SACK hint. */
320 	if (tp->sackhint.nexthole == NULL)
321 		tp->sackhint.nexthole = hole;
322 
323 	return hole;
324 }
325 
326 /*
327  * Remove SACK hole from scoreboard.
328  */
329 static void
330 tcp_sackhole_remove(struct tcpcb *tp, struct sackhole *hole)
331 {
332 
333 	/* Update SACK hint. */
334 	if (tp->sackhint.nexthole == hole)
335 		tp->sackhint.nexthole = TAILQ_NEXT(hole, scblink);
336 
337 	/* Remove this SACK hole. */
338 	TAILQ_REMOVE(&tp->snd_holes, hole, scblink);
339 
340 	/* Free this SACK hole. */
341 	tcp_sackhole_free(tp, hole);
342 }
343 
344 /*
345  * Process cumulative ACK and the TCP SACK option to update the scoreboard.
346  * tp->snd_holes is an ordered list of holes (oldest to newest, in terms of
347  * the sequence space).
348  * Returns 1 if incoming ACK has previously unknown SACK information,
349  * 0 otherwise. Note: We treat (snd_una, th_ack) as a sack block so any changes
350  * to that (i.e. left edge moving) would also be considered a change in SACK
351  * information which is slightly different than rfc6675.
352  */
353 int
354 tcp_sack_doack(struct tcpcb *tp, struct tcpopt *to, tcp_seq th_ack)
355 {
356 	struct sackhole *cur, *temp;
357 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1], *sblkp;
358 	int i, j, num_sack_blks, sack_changed;
359 
360 	INP_WLOCK_ASSERT(tp->t_inpcb);
361 
362 	num_sack_blks = 0;
363 	sack_changed = 0;
364 	/*
365 	 * If SND.UNA will be advanced by SEG.ACK, and if SACK holes exist,
366 	 * treat [SND.UNA, SEG.ACK) as if it is a SACK block.
367 	 */
368 	if (SEQ_LT(tp->snd_una, th_ack) && !TAILQ_EMPTY(&tp->snd_holes)) {
369 		sack_blocks[num_sack_blks].start = tp->snd_una;
370 		sack_blocks[num_sack_blks++].end = th_ack;
371 	}
372 	/*
373 	 * Append received valid SACK blocks to sack_blocks[], but only if we
374 	 * received new blocks from the other side.
375 	 */
376 	if (to->to_flags & TOF_SACK) {
377 		tp->sackhint.sacked_bytes = 0;	/* reset */
378 		for (i = 0; i < to->to_nsacks; i++) {
379 			bcopy((to->to_sacks + i * TCPOLEN_SACK),
380 			    &sack, sizeof(sack));
381 			sack.start = ntohl(sack.start);
382 			sack.end = ntohl(sack.end);
383 			if (SEQ_GT(sack.end, sack.start) &&
384 			    SEQ_GT(sack.start, tp->snd_una) &&
385 			    SEQ_GT(sack.start, th_ack) &&
386 			    SEQ_LT(sack.start, tp->snd_max) &&
387 			    SEQ_GT(sack.end, tp->snd_una) &&
388 			    SEQ_LEQ(sack.end, tp->snd_max)) {
389 				sack_blocks[num_sack_blks++] = sack;
390 				tp->sackhint.sacked_bytes +=
391 				    (sack.end-sack.start);
392 			}
393 		}
394 	}
395 	/*
396 	 * Return if SND.UNA is not advanced and no valid SACK block is
397 	 * received.
398 	 */
399 	if (num_sack_blks == 0)
400 		return (sack_changed);
401 
402 	/*
403 	 * Sort the SACK blocks so we can update the scoreboard with just one
404 	 * pass. The overhead of sorting up to 4+1 elements is less than
405 	 * making up to 4+1 passes over the scoreboard.
406 	 */
407 	for (i = 0; i < num_sack_blks; i++) {
408 		for (j = i + 1; j < num_sack_blks; j++) {
409 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
410 				sack = sack_blocks[i];
411 				sack_blocks[i] = sack_blocks[j];
412 				sack_blocks[j] = sack;
413 			}
414 		}
415 	}
416 	if (TAILQ_EMPTY(&tp->snd_holes))
417 		/*
418 		 * Empty scoreboard. Need to initialize snd_fack (it may be
419 		 * uninitialized or have a bogus value). Scoreboard holes
420 		 * (from the sack blocks received) are created later below
421 		 * (in the logic that adds holes to the tail of the
422 		 * scoreboard).
423 		 */
424 		tp->snd_fack = SEQ_MAX(tp->snd_una, th_ack);
425 	/*
426 	 * In the while-loop below, incoming SACK blocks (sack_blocks[]) and
427 	 * SACK holes (snd_holes) are traversed from their tails with just
428 	 * one pass in order to reduce the number of compares especially when
429 	 * the bandwidth-delay product is large.
430 	 *
431 	 * Note: Typically, in the first RTT of SACK recovery, the highest
432 	 * three or four SACK blocks with the same ack number are received.
433 	 * In the second RTT, if retransmitted data segments are not lost,
434 	 * the highest three or four SACK blocks with ack number advancing
435 	 * are received.
436 	 */
437 	sblkp = &sack_blocks[num_sack_blks - 1];	/* Last SACK block */
438 	tp->sackhint.last_sack_ack = sblkp->end;
439 	if (SEQ_LT(tp->snd_fack, sblkp->start)) {
440 		/*
441 		 * The highest SACK block is beyond fack.  Append new SACK
442 		 * hole at the tail.  If the second or later highest SACK
443 		 * blocks are also beyond the current fack, they will be
444 		 * inserted by way of hole splitting in the while-loop below.
445 		 */
446 		temp = tcp_sackhole_insert(tp, tp->snd_fack,sblkp->start,NULL);
447 		if (temp != NULL) {
448 			tp->snd_fack = sblkp->end;
449 			/* Go to the previous sack block. */
450 			sblkp--;
451 			sack_changed = 1;
452 		} else {
453 			/*
454 			 * We failed to add a new hole based on the current
455 			 * sack block.  Skip over all the sack blocks that
456 			 * fall completely to the right of snd_fack and
457 			 * proceed to trim the scoreboard based on the
458 			 * remaining sack blocks.  This also trims the
459 			 * scoreboard for th_ack (which is sack_blocks[0]).
460 			 */
461 			while (sblkp >= sack_blocks &&
462 			       SEQ_LT(tp->snd_fack, sblkp->start))
463 				sblkp--;
464 			if (sblkp >= sack_blocks &&
465 			    SEQ_LT(tp->snd_fack, sblkp->end))
466 				tp->snd_fack = sblkp->end;
467 		}
468 	} else if (SEQ_LT(tp->snd_fack, sblkp->end)) {
469 		/* fack is advanced. */
470 		tp->snd_fack = sblkp->end;
471 		sack_changed = 1;
472 	}
473 	/* We must have at least one SACK hole in scoreboard. */
474 	KASSERT(!TAILQ_EMPTY(&tp->snd_holes),
475 	    ("SACK scoreboard must not be empty"));
476 	cur = TAILQ_LAST(&tp->snd_holes, sackhole_head); /* Last SACK hole. */
477 	/*
478 	 * Since the incoming sack blocks are sorted, we can process them
479 	 * making one sweep of the scoreboard.
480 	 */
481 	while (sblkp >= sack_blocks  && cur != NULL) {
482 		if (SEQ_GEQ(sblkp->start, cur->end)) {
483 			/*
484 			 * SACKs data beyond the current hole.  Go to the
485 			 * previous sack block.
486 			 */
487 			sblkp--;
488 			continue;
489 		}
490 		if (SEQ_LEQ(sblkp->end, cur->start)) {
491 			/*
492 			 * SACKs data before the current hole.  Go to the
493 			 * previous hole.
494 			 */
495 			cur = TAILQ_PREV(cur, sackhole_head, scblink);
496 			continue;
497 		}
498 		tp->sackhint.sack_bytes_rexmit -= (cur->rxmit - cur->start);
499 		KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
500 		    ("sackhint bytes rtx >= 0"));
501 		sack_changed = 1;
502 		if (SEQ_LEQ(sblkp->start, cur->start)) {
503 			/* Data acks at least the beginning of hole. */
504 			if (SEQ_GEQ(sblkp->end, cur->end)) {
505 				/* Acks entire hole, so delete hole. */
506 				temp = cur;
507 				cur = TAILQ_PREV(cur, sackhole_head, scblink);
508 				tcp_sackhole_remove(tp, temp);
509 				/*
510 				 * The sack block may ack all or part of the
511 				 * next hole too, so continue onto the next
512 				 * hole.
513 				 */
514 				continue;
515 			} else {
516 				/* Move start of hole forward. */
517 				cur->start = sblkp->end;
518 				cur->rxmit = SEQ_MAX(cur->rxmit, cur->start);
519 			}
520 		} else {
521 			/* Data acks at least the end of hole. */
522 			if (SEQ_GEQ(sblkp->end, cur->end)) {
523 				/* Move end of hole backward. */
524 				cur->end = sblkp->start;
525 				cur->rxmit = SEQ_MIN(cur->rxmit, cur->end);
526 			} else {
527 				/*
528 				 * ACKs some data in middle of a hole; need
529 				 * to split current hole
530 				 */
531 				temp = tcp_sackhole_insert(tp, sblkp->end,
532 				    cur->end, cur);
533 				if (temp != NULL) {
534 					if (SEQ_GT(cur->rxmit, temp->rxmit)) {
535 						temp->rxmit = cur->rxmit;
536 						tp->sackhint.sack_bytes_rexmit
537 						    += (temp->rxmit
538 						    - temp->start);
539 					}
540 					cur->end = sblkp->start;
541 					cur->rxmit = SEQ_MIN(cur->rxmit,
542 					    cur->end);
543 				}
544 			}
545 		}
546 		tp->sackhint.sack_bytes_rexmit += (cur->rxmit - cur->start);
547 		/*
548 		 * Testing sblkp->start against cur->start tells us whether
549 		 * we're done with the sack block or the sack hole.
550 		 * Accordingly, we advance one or the other.
551 		 */
552 		if (SEQ_LEQ(sblkp->start, cur->start))
553 			cur = TAILQ_PREV(cur, sackhole_head, scblink);
554 		else
555 			sblkp--;
556 	}
557 	return (sack_changed);
558 }
559 
560 /*
561  * Free all SACK holes to clear the scoreboard.
562  */
563 void
564 tcp_free_sackholes(struct tcpcb *tp)
565 {
566 	struct sackhole *q;
567 
568 	INP_WLOCK_ASSERT(tp->t_inpcb);
569 	while ((q = TAILQ_FIRST(&tp->snd_holes)) != NULL)
570 		tcp_sackhole_remove(tp, q);
571 	tp->sackhint.sack_bytes_rexmit = 0;
572 
573 	KASSERT(tp->snd_numholes == 0, ("tp->snd_numholes == 0"));
574 	KASSERT(tp->sackhint.nexthole == NULL,
575 		("tp->sackhint.nexthole == NULL"));
576 }
577 
578 /*
579  * Partial ack handling within a sack recovery episode.  Keeping this very
580  * simple for now.  When a partial ack is received, force snd_cwnd to a value
581  * that will allow the sender to transmit no more than 2 segments.  If
582  * necessary, a better scheme can be adopted at a later point, but for now,
583  * the goal is to prevent the sender from bursting a large amount of data in
584  * the midst of sack recovery.
585  */
586 void
587 tcp_sack_partialack(struct tcpcb *tp, struct tcphdr *th)
588 {
589 	int num_segs = 1;
590 
591 	INP_WLOCK_ASSERT(tp->t_inpcb);
592 	tcp_timer_activate(tp, TT_REXMT, 0);
593 	tp->t_rtttime = 0;
594 	/* Send one or 2 segments based on how much new data was acked. */
595 	if ((BYTES_THIS_ACK(tp, th) / tp->t_maxseg) >= 2)
596 		num_segs = 2;
597 	tp->snd_cwnd = (tp->sackhint.sack_bytes_rexmit +
598 	    (tp->snd_nxt - tp->sack_newdata) + num_segs * tp->t_maxseg);
599 	if (tp->snd_cwnd > tp->snd_ssthresh)
600 		tp->snd_cwnd = tp->snd_ssthresh;
601 	tp->t_flags |= TF_ACKNOW;
602 	(void) tp->t_fb->tfb_tcp_output(tp);
603 }
604 
605 #if 0
606 /*
607  * Debug version of tcp_sack_output() that walks the scoreboard.  Used for
608  * now to sanity check the hint.
609  */
610 static struct sackhole *
611 tcp_sack_output_debug(struct tcpcb *tp, int *sack_bytes_rexmt)
612 {
613 	struct sackhole *p;
614 
615 	INP_WLOCK_ASSERT(tp->t_inpcb);
616 	*sack_bytes_rexmt = 0;
617 	TAILQ_FOREACH(p, &tp->snd_holes, scblink) {
618 		if (SEQ_LT(p->rxmit, p->end)) {
619 			if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */
620 				continue;
621 			}
622 			*sack_bytes_rexmt += (p->rxmit - p->start);
623 			break;
624 		}
625 		*sack_bytes_rexmt += (p->rxmit - p->start);
626 	}
627 	return (p);
628 }
629 #endif
630 
631 /*
632  * Returns the next hole to retransmit and the number of retransmitted bytes
633  * from the scoreboard.  We store both the next hole and the number of
634  * retransmitted bytes as hints (and recompute these on the fly upon SACK/ACK
635  * reception).  This avoids scoreboard traversals completely.
636  *
637  * The loop here will traverse *at most* one link.  Here's the argument.  For
638  * the loop to traverse more than 1 link before finding the next hole to
639  * retransmit, we would need to have at least 1 node following the current
640  * hint with (rxmit == end).  But, for all holes following the current hint,
641  * (start == rxmit), since we have not yet retransmitted from them.
642  * Therefore, in order to traverse more 1 link in the loop below, we need to
643  * have at least one node following the current hint with (start == rxmit ==
644  * end).  But that can't happen, (start == end) means that all the data in
645  * that hole has been sacked, in which case, the hole would have been removed
646  * from the scoreboard.
647  */
648 struct sackhole *
649 tcp_sack_output(struct tcpcb *tp, int *sack_bytes_rexmt)
650 {
651 	struct sackhole *hole = NULL;
652 
653 	INP_WLOCK_ASSERT(tp->t_inpcb);
654 	*sack_bytes_rexmt = tp->sackhint.sack_bytes_rexmit;
655 	hole = tp->sackhint.nexthole;
656 	if (hole == NULL || SEQ_LT(hole->rxmit, hole->end))
657 		goto out;
658 	while ((hole = TAILQ_NEXT(hole, scblink)) != NULL) {
659 		if (SEQ_LT(hole->rxmit, hole->end)) {
660 			tp->sackhint.nexthole = hole;
661 			break;
662 		}
663 	}
664 out:
665 	return (hole);
666 }
667 
668 /*
669  * After a timeout, the SACK list may be rebuilt.  This SACK information
670  * should be used to avoid retransmitting SACKed data.  This function
671  * traverses the SACK list to see if snd_nxt should be moved forward.
672  */
673 void
674 tcp_sack_adjust(struct tcpcb *tp)
675 {
676 	struct sackhole *p, *cur = TAILQ_FIRST(&tp->snd_holes);
677 
678 	INP_WLOCK_ASSERT(tp->t_inpcb);
679 	if (cur == NULL)
680 		return; /* No holes */
681 	if (SEQ_GEQ(tp->snd_nxt, tp->snd_fack))
682 		return; /* We're already beyond any SACKed blocks */
683 	/*-
684 	 * Two cases for which we want to advance snd_nxt:
685 	 * i) snd_nxt lies between end of one hole and beginning of another
686 	 * ii) snd_nxt lies between end of last hole and snd_fack
687 	 */
688 	while ((p = TAILQ_NEXT(cur, scblink)) != NULL) {
689 		if (SEQ_LT(tp->snd_nxt, cur->end))
690 			return;
691 		if (SEQ_GEQ(tp->snd_nxt, p->start))
692 			cur = p;
693 		else {
694 			tp->snd_nxt = p->start;
695 			return;
696 		}
697 	}
698 	if (SEQ_LT(tp->snd_nxt, cur->end))
699 		return;
700 	tp->snd_nxt = tp->snd_fack;
701 }
702