xref: /freebsd/sys/netinet/tcp_sack.c (revision 81966bce06dac45f42bda62b14dba0756ef28505)
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/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 <netinet6/tcp6_var.h>
118 #include <netinet/tcpip.h>
119 #ifdef TCPDEBUG
120 #include <netinet/tcp_debug.h>
121 #endif /* TCPDEBUG */
122 
123 #include <machine/in_cksum.h>
124 
125 VNET_DECLARE(struct uma_zone *, sack_hole_zone);
126 #define	V_sack_hole_zone		VNET(sack_hole_zone)
127 
128 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, sack, CTLFLAG_RW, 0, "TCP SACK");
129 VNET_DEFINE(int, tcp_do_sack) = 1;
130 #define	V_tcp_do_sack			VNET(tcp_do_sack)
131 SYSCTL_VNET_INT(_net_inet_tcp_sack, OID_AUTO, enable, CTLFLAG_RW,
132     &VNET_NAME(tcp_do_sack), 0, "Enable/Disable TCP SACK support");
133 
134 VNET_DEFINE(int, tcp_sack_maxholes) = 128;
135 #define	V_tcp_sack_maxholes		VNET(tcp_sack_maxholes)
136 SYSCTL_VNET_INT(_net_inet_tcp_sack, OID_AUTO, maxholes, CTLFLAG_RW,
137     &VNET_NAME(tcp_sack_maxholes), 0,
138     "Maximum number of TCP SACK holes allowed per connection");
139 
140 VNET_DEFINE(int, tcp_sack_globalmaxholes) = 65536;
141 #define	V_tcp_sack_globalmaxholes	VNET(tcp_sack_globalmaxholes)
142 SYSCTL_VNET_INT(_net_inet_tcp_sack, OID_AUTO, globalmaxholes, CTLFLAG_RW,
143     &VNET_NAME(tcp_sack_globalmaxholes), 0,
144     "Global maximum number of TCP SACK holes");
145 
146 VNET_DEFINE(int, tcp_sack_globalholes) = 0;
147 #define	V_tcp_sack_globalholes		VNET(tcp_sack_globalholes)
148 SYSCTL_VNET_INT(_net_inet_tcp_sack, OID_AUTO, globalholes, CTLFLAG_RD,
149     &VNET_NAME(tcp_sack_globalholes), 0,
150     "Global number of TCP SACK holes currently allocated");
151 
152 /*
153  * This function is called upon receipt of new valid data (while not in
154  * header prediction mode), and it updates the ordered list of sacks.
155  */
156 void
157 tcp_update_sack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
158 {
159 	/*
160 	 * First reported block MUST be the most recent one.  Subsequent
161 	 * blocks SHOULD be in the order in which they arrived at the
162 	 * receiver.  These two conditions make the implementation fully
163 	 * compliant with RFC 2018.
164 	 */
165 	struct sackblk head_blk, saved_blks[MAX_SACK_BLKS];
166 	int num_head, num_saved, i;
167 
168 	INP_WLOCK_ASSERT(tp->t_inpcb);
169 
170 	/* Check arguments. */
171 	KASSERT(SEQ_LT(rcv_start, rcv_end), ("rcv_start < rcv_end"));
172 
173 	/* SACK block for the received segment. */
174 	head_blk.start = rcv_start;
175 	head_blk.end = rcv_end;
176 
177 	/*
178 	 * Merge updated SACK blocks into head_blk, and save unchanged SACK
179 	 * blocks into saved_blks[].  num_saved will have the number of the
180 	 * saved SACK blocks.
181 	 */
182 	num_saved = 0;
183 	for (i = 0; i < tp->rcv_numsacks; i++) {
184 		tcp_seq start = tp->sackblks[i].start;
185 		tcp_seq end = tp->sackblks[i].end;
186 		if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
187 			/*
188 			 * Discard this SACK block.
189 			 */
190 		} else if (SEQ_LEQ(head_blk.start, end) &&
191 			   SEQ_GEQ(head_blk.end, start)) {
192 			/*
193 			 * Merge this SACK block into head_blk.  This SACK
194 			 * block itself will be discarded.
195 			 */
196 			if (SEQ_GT(head_blk.start, start))
197 				head_blk.start = start;
198 			if (SEQ_LT(head_blk.end, end))
199 				head_blk.end = end;
200 		} else {
201 			/*
202 			 * Save this SACK block.
203 			 */
204 			saved_blks[num_saved].start = start;
205 			saved_blks[num_saved].end = end;
206 			num_saved++;
207 		}
208 	}
209 
210 	/*
211 	 * Update SACK list in tp->sackblks[].
212 	 */
213 	num_head = 0;
214 	if (SEQ_GT(head_blk.start, tp->rcv_nxt)) {
215 		/*
216 		 * The received data segment is an out-of-order segment.  Put
217 		 * head_blk at the top of SACK list.
218 		 */
219 		tp->sackblks[0] = head_blk;
220 		num_head = 1;
221 		/*
222 		 * If the number of saved SACK blocks exceeds its limit,
223 		 * discard the last SACK block.
224 		 */
225 		if (num_saved >= MAX_SACK_BLKS)
226 			num_saved--;
227 	}
228 	if (num_saved > 0) {
229 		/*
230 		 * Copy the saved SACK blocks back.
231 		 */
232 		bcopy(saved_blks, &tp->sackblks[num_head],
233 		      sizeof(struct sackblk) * num_saved);
234 	}
235 
236 	/* Save the number of SACK blocks. */
237 	tp->rcv_numsacks = num_head + num_saved;
238 }
239 
240 /*
241  * Delete all receiver-side SACK information.
242  */
243 void
244 tcp_clean_sackreport(struct tcpcb *tp)
245 {
246 	int i;
247 
248 	INP_WLOCK_ASSERT(tp->t_inpcb);
249 	tp->rcv_numsacks = 0;
250 	for (i = 0; i < MAX_SACK_BLKS; i++)
251 		tp->sackblks[i].start = tp->sackblks[i].end=0;
252 }
253 
254 /*
255  * Allocate struct sackhole.
256  */
257 static struct sackhole *
258 tcp_sackhole_alloc(struct tcpcb *tp, tcp_seq start, tcp_seq end)
259 {
260 	struct sackhole *hole;
261 
262 	if (tp->snd_numholes >= V_tcp_sack_maxholes ||
263 	    V_tcp_sack_globalholes >= V_tcp_sack_globalmaxholes) {
264 		TCPSTAT_INC(tcps_sack_sboverflow);
265 		return NULL;
266 	}
267 
268 	hole = (struct sackhole *)uma_zalloc(V_sack_hole_zone, M_NOWAIT);
269 	if (hole == NULL)
270 		return NULL;
271 
272 	hole->start = start;
273 	hole->end = end;
274 	hole->rxmit = start;
275 
276 	tp->snd_numholes++;
277 	atomic_add_int(&V_tcp_sack_globalholes, 1);
278 
279 	return hole;
280 }
281 
282 /*
283  * Free struct sackhole.
284  */
285 static void
286 tcp_sackhole_free(struct tcpcb *tp, struct sackhole *hole)
287 {
288 
289 	uma_zfree(V_sack_hole_zone, hole);
290 
291 	tp->snd_numholes--;
292 	atomic_subtract_int(&V_tcp_sack_globalholes, 1);
293 
294 	KASSERT(tp->snd_numholes >= 0, ("tp->snd_numholes >= 0"));
295 	KASSERT(V_tcp_sack_globalholes >= 0, ("tcp_sack_globalholes >= 0"));
296 }
297 
298 /*
299  * Insert new SACK hole into scoreboard.
300  */
301 static struct sackhole *
302 tcp_sackhole_insert(struct tcpcb *tp, tcp_seq start, tcp_seq end,
303     struct sackhole *after)
304 {
305 	struct sackhole *hole;
306 
307 	/* Allocate a new SACK hole. */
308 	hole = tcp_sackhole_alloc(tp, start, end);
309 	if (hole == NULL)
310 		return NULL;
311 
312 	/* Insert the new SACK hole into scoreboard. */
313 	if (after != NULL)
314 		TAILQ_INSERT_AFTER(&tp->snd_holes, after, hole, scblink);
315 	else
316 		TAILQ_INSERT_TAIL(&tp->snd_holes, hole, scblink);
317 
318 	/* Update SACK hint. */
319 	if (tp->sackhint.nexthole == NULL)
320 		tp->sackhint.nexthole = hole;
321 
322 	return hole;
323 }
324 
325 /*
326  * Remove SACK hole from scoreboard.
327  */
328 static void
329 tcp_sackhole_remove(struct tcpcb *tp, struct sackhole *hole)
330 {
331 
332 	/* Update SACK hint. */
333 	if (tp->sackhint.nexthole == hole)
334 		tp->sackhint.nexthole = TAILQ_NEXT(hole, scblink);
335 
336 	/* Remove this SACK hole. */
337 	TAILQ_REMOVE(&tp->snd_holes, hole, scblink);
338 
339 	/* Free this SACK hole. */
340 	tcp_sackhole_free(tp, hole);
341 }
342 
343 /*
344  * Process cumulative ACK and the TCP SACK option to update the scoreboard.
345  * tp->snd_holes is an ordered list of holes (oldest to newest, in terms of
346  * the sequence space).
347  */
348 void
349 tcp_sack_doack(struct tcpcb *tp, struct tcpopt *to, tcp_seq th_ack)
350 {
351 	struct sackhole *cur, *temp;
352 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1], *sblkp;
353 	int i, j, num_sack_blks;
354 
355 	INP_WLOCK_ASSERT(tp->t_inpcb);
356 
357 	num_sack_blks = 0;
358 	/*
359 	 * If SND.UNA will be advanced by SEG.ACK, and if SACK holes exist,
360 	 * treat [SND.UNA, SEG.ACK) as if it is a SACK block.
361 	 */
362 	if (SEQ_LT(tp->snd_una, th_ack) && !TAILQ_EMPTY(&tp->snd_holes)) {
363 		sack_blocks[num_sack_blks].start = tp->snd_una;
364 		sack_blocks[num_sack_blks++].end = th_ack;
365 	}
366 	/*
367 	 * Append received valid SACK blocks to sack_blocks[], but only if we
368 	 * received new blocks from the other side.
369 	 */
370 	if (to->to_flags & TOF_SACK) {
371 		for (i = 0; i < to->to_nsacks; i++) {
372 			bcopy((to->to_sacks + i * TCPOLEN_SACK),
373 			    &sack, sizeof(sack));
374 			sack.start = ntohl(sack.start);
375 			sack.end = ntohl(sack.end);
376 			if (SEQ_GT(sack.end, sack.start) &&
377 			    SEQ_GT(sack.start, tp->snd_una) &&
378 			    SEQ_GT(sack.start, th_ack) &&
379 			    SEQ_LT(sack.start, tp->snd_max) &&
380 			    SEQ_GT(sack.end, tp->snd_una) &&
381 			    SEQ_LEQ(sack.end, tp->snd_max))
382 				sack_blocks[num_sack_blks++] = sack;
383 		}
384 	}
385 	/*
386 	 * Return if SND.UNA is not advanced and no valid SACK block is
387 	 * received.
388 	 */
389 	if (num_sack_blks == 0)
390 		return;
391 
392 	/*
393 	 * Sort the SACK blocks so we can update the scoreboard with just one
394 	 * pass. The overhead of sorting upto 4+1 elements is less than
395 	 * making upto 4+1 passes over the scoreboard.
396 	 */
397 	for (i = 0; i < num_sack_blks; i++) {
398 		for (j = i + 1; j < num_sack_blks; j++) {
399 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
400 				sack = sack_blocks[i];
401 				sack_blocks[i] = sack_blocks[j];
402 				sack_blocks[j] = sack;
403 			}
404 		}
405 	}
406 	if (TAILQ_EMPTY(&tp->snd_holes))
407 		/*
408 		 * Empty scoreboard. Need to initialize snd_fack (it may be
409 		 * uninitialized or have a bogus value). Scoreboard holes
410 		 * (from the sack blocks received) are created later below
411 		 * (in the logic that adds holes to the tail of the
412 		 * scoreboard).
413 		 */
414 		tp->snd_fack = SEQ_MAX(tp->snd_una, th_ack);
415 	/*
416 	 * In the while-loop below, incoming SACK blocks (sack_blocks[]) and
417 	 * SACK holes (snd_holes) are traversed from their tails with just
418 	 * one pass in order to reduce the number of compares especially when
419 	 * the bandwidth-delay product is large.
420 	 *
421 	 * Note: Typically, in the first RTT of SACK recovery, the highest
422 	 * three or four SACK blocks with the same ack number are received.
423 	 * In the second RTT, if retransmitted data segments are not lost,
424 	 * the highest three or four SACK blocks with ack number advancing
425 	 * are received.
426 	 */
427 	sblkp = &sack_blocks[num_sack_blks - 1];	/* Last SACK block */
428 	tp->sackhint.last_sack_ack = sblkp->end;
429 	if (SEQ_LT(tp->snd_fack, sblkp->start)) {
430 		/*
431 		 * The highest SACK block is beyond fack.  Append new SACK
432 		 * hole at the tail.  If the second or later highest SACK
433 		 * blocks are also beyond the current fack, they will be
434 		 * inserted by way of hole splitting in the while-loop below.
435 		 */
436 		temp = tcp_sackhole_insert(tp, tp->snd_fack,sblkp->start,NULL);
437 		if (temp != NULL) {
438 			tp->snd_fack = sblkp->end;
439 			/* Go to the previous sack block. */
440 			sblkp--;
441 		} else {
442 			/*
443 			 * We failed to add a new hole based on the current
444 			 * sack block.  Skip over all the sack blocks that
445 			 * fall completely to the right of snd_fack and
446 			 * proceed to trim the scoreboard based on the
447 			 * remaining sack blocks.  This also trims the
448 			 * scoreboard for th_ack (which is sack_blocks[0]).
449 			 */
450 			while (sblkp >= sack_blocks &&
451 			       SEQ_LT(tp->snd_fack, sblkp->start))
452 				sblkp--;
453 			if (sblkp >= sack_blocks &&
454 			    SEQ_LT(tp->snd_fack, sblkp->end))
455 				tp->snd_fack = sblkp->end;
456 		}
457 	} else if (SEQ_LT(tp->snd_fack, sblkp->end))
458 		/* fack is advanced. */
459 		tp->snd_fack = sblkp->end;
460 	/* We must have at least one SACK hole in scoreboard. */
461 	KASSERT(!TAILQ_EMPTY(&tp->snd_holes),
462 	    ("SACK scoreboard must not be empty"));
463 	cur = TAILQ_LAST(&tp->snd_holes, sackhole_head); /* Last SACK hole. */
464 	/*
465 	 * Since the incoming sack blocks are sorted, we can process them
466 	 * making one sweep of the scoreboard.
467 	 */
468 	while (sblkp >= sack_blocks  && cur != NULL) {
469 		if (SEQ_GEQ(sblkp->start, cur->end)) {
470 			/*
471 			 * SACKs data beyond the current hole.  Go to the
472 			 * previous sack block.
473 			 */
474 			sblkp--;
475 			continue;
476 		}
477 		if (SEQ_LEQ(sblkp->end, cur->start)) {
478 			/*
479 			 * SACKs data before the current hole.  Go to the
480 			 * previous hole.
481 			 */
482 			cur = TAILQ_PREV(cur, sackhole_head, scblink);
483 			continue;
484 		}
485 		tp->sackhint.sack_bytes_rexmit -= (cur->rxmit - cur->start);
486 		KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
487 		    ("sackhint bytes rtx >= 0"));
488 		if (SEQ_LEQ(sblkp->start, cur->start)) {
489 			/* Data acks at least the beginning of hole. */
490 			if (SEQ_GEQ(sblkp->end, cur->end)) {
491 				/* Acks entire hole, so delete hole. */
492 				temp = cur;
493 				cur = TAILQ_PREV(cur, sackhole_head, scblink);
494 				tcp_sackhole_remove(tp, temp);
495 				/*
496 				 * The sack block may ack all or part of the
497 				 * next hole too, so continue onto the next
498 				 * hole.
499 				 */
500 				continue;
501 			} else {
502 				/* Move start of hole forward. */
503 				cur->start = sblkp->end;
504 				cur->rxmit = SEQ_MAX(cur->rxmit, cur->start);
505 			}
506 		} else {
507 			/* Data acks at least the end of hole. */
508 			if (SEQ_GEQ(sblkp->end, cur->end)) {
509 				/* Move end of hole backward. */
510 				cur->end = sblkp->start;
511 				cur->rxmit = SEQ_MIN(cur->rxmit, cur->end);
512 			} else {
513 				/*
514 				 * ACKs some data in middle of a hole; need
515 				 * to split current hole
516 				 */
517 				temp = tcp_sackhole_insert(tp, sblkp->end,
518 				    cur->end, cur);
519 				if (temp != NULL) {
520 					if (SEQ_GT(cur->rxmit, temp->rxmit)) {
521 						temp->rxmit = cur->rxmit;
522 						tp->sackhint.sack_bytes_rexmit
523 						    += (temp->rxmit
524 						    - temp->start);
525 					}
526 					cur->end = sblkp->start;
527 					cur->rxmit = SEQ_MIN(cur->rxmit,
528 					    cur->end);
529 				}
530 			}
531 		}
532 		tp->sackhint.sack_bytes_rexmit += (cur->rxmit - cur->start);
533 		/*
534 		 * Testing sblkp->start against cur->start tells us whether
535 		 * we're done with the sack block or the sack hole.
536 		 * Accordingly, we advance one or the other.
537 		 */
538 		if (SEQ_LEQ(sblkp->start, cur->start))
539 			cur = TAILQ_PREV(cur, sackhole_head, scblink);
540 		else
541 			sblkp--;
542 	}
543 }
544 
545 /*
546  * Free all SACK holes to clear the scoreboard.
547  */
548 void
549 tcp_free_sackholes(struct tcpcb *tp)
550 {
551 	struct sackhole *q;
552 
553 	INP_WLOCK_ASSERT(tp->t_inpcb);
554 	while ((q = TAILQ_FIRST(&tp->snd_holes)) != NULL)
555 		tcp_sackhole_remove(tp, q);
556 	tp->sackhint.sack_bytes_rexmit = 0;
557 
558 	KASSERT(tp->snd_numholes == 0, ("tp->snd_numholes == 0"));
559 	KASSERT(tp->sackhint.nexthole == NULL,
560 		("tp->sackhint.nexthole == NULL"));
561 }
562 
563 /*
564  * Partial ack handling within a sack recovery episode.  Keeping this very
565  * simple for now.  When a partial ack is received, force snd_cwnd to a value
566  * that will allow the sender to transmit no more than 2 segments.  If
567  * necessary, a better scheme can be adopted at a later point, but for now,
568  * the goal is to prevent the sender from bursting a large amount of data in
569  * the midst of sack recovery.
570  */
571 void
572 tcp_sack_partialack(struct tcpcb *tp, struct tcphdr *th)
573 {
574 	int num_segs = 1;
575 
576 	INP_WLOCK_ASSERT(tp->t_inpcb);
577 	tcp_timer_activate(tp, TT_REXMT, 0);
578 	tp->t_rtttime = 0;
579 	/* Send one or 2 segments based on how much new data was acked. */
580 	if ((BYTES_THIS_ACK(tp, th) / tp->t_maxseg) >= 2)
581 		num_segs = 2;
582 	tp->snd_cwnd = (tp->sackhint.sack_bytes_rexmit +
583 	    (tp->snd_nxt - tp->sack_newdata) + num_segs * tp->t_maxseg);
584 	if (tp->snd_cwnd > tp->snd_ssthresh)
585 		tp->snd_cwnd = tp->snd_ssthresh;
586 	tp->t_flags |= TF_ACKNOW;
587 	(void) tcp_output(tp);
588 }
589 
590 #if 0
591 /*
592  * Debug version of tcp_sack_output() that walks the scoreboard.  Used for
593  * now to sanity check the hint.
594  */
595 static struct sackhole *
596 tcp_sack_output_debug(struct tcpcb *tp, int *sack_bytes_rexmt)
597 {
598 	struct sackhole *p;
599 
600 	INP_WLOCK_ASSERT(tp->t_inpcb);
601 	*sack_bytes_rexmt = 0;
602 	TAILQ_FOREACH(p, &tp->snd_holes, scblink) {
603 		if (SEQ_LT(p->rxmit, p->end)) {
604 			if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */
605 				continue;
606 			}
607 			*sack_bytes_rexmt += (p->rxmit - p->start);
608 			break;
609 		}
610 		*sack_bytes_rexmt += (p->rxmit - p->start);
611 	}
612 	return (p);
613 }
614 #endif
615 
616 /*
617  * Returns the next hole to retransmit and the number of retransmitted bytes
618  * from the scoreboard.  We store both the next hole and the number of
619  * retransmitted bytes as hints (and recompute these on the fly upon SACK/ACK
620  * reception).  This avoids scoreboard traversals completely.
621  *
622  * The loop here will traverse *at most* one link.  Here's the argument.  For
623  * the loop to traverse more than 1 link before finding the next hole to
624  * retransmit, we would need to have at least 1 node following the current
625  * hint with (rxmit == end).  But, for all holes following the current hint,
626  * (start == rxmit), since we have not yet retransmitted from them.
627  * Therefore, in order to traverse more 1 link in the loop below, we need to
628  * have at least one node following the current hint with (start == rxmit ==
629  * end).  But that can't happen, (start == end) means that all the data in
630  * that hole has been sacked, in which case, the hole would have been removed
631  * from the scoreboard.
632  */
633 struct sackhole *
634 tcp_sack_output(struct tcpcb *tp, int *sack_bytes_rexmt)
635 {
636 	struct sackhole *hole = NULL;
637 
638 	INP_WLOCK_ASSERT(tp->t_inpcb);
639 	*sack_bytes_rexmt = tp->sackhint.sack_bytes_rexmit;
640 	hole = tp->sackhint.nexthole;
641 	if (hole == NULL || SEQ_LT(hole->rxmit, hole->end))
642 		goto out;
643 	while ((hole = TAILQ_NEXT(hole, scblink)) != NULL) {
644 		if (SEQ_LT(hole->rxmit, hole->end)) {
645 			tp->sackhint.nexthole = hole;
646 			break;
647 		}
648 	}
649 out:
650 	return (hole);
651 }
652 
653 /*
654  * After a timeout, the SACK list may be rebuilt.  This SACK information
655  * should be used to avoid retransmitting SACKed data.  This function
656  * traverses the SACK list to see if snd_nxt should be moved forward.
657  */
658 void
659 tcp_sack_adjust(struct tcpcb *tp)
660 {
661 	struct sackhole *p, *cur = TAILQ_FIRST(&tp->snd_holes);
662 
663 	INP_WLOCK_ASSERT(tp->t_inpcb);
664 	if (cur == NULL)
665 		return; /* No holes */
666 	if (SEQ_GEQ(tp->snd_nxt, tp->snd_fack))
667 		return; /* We're already beyond any SACKed blocks */
668 	/*-
669 	 * Two cases for which we want to advance snd_nxt:
670 	 * i) snd_nxt lies between end of one hole and beginning of another
671 	 * ii) snd_nxt lies between end of last hole and snd_fack
672 	 */
673 	while ((p = TAILQ_NEXT(cur, scblink)) != NULL) {
674 		if (SEQ_LT(tp->snd_nxt, cur->end))
675 			return;
676 		if (SEQ_GEQ(tp->snd_nxt, p->start))
677 			cur = p;
678 		else {
679 			tp->snd_nxt = p->start;
680 			return;
681 		}
682 	}
683 	if (SEQ_LT(tp->snd_nxt, cur->end))
684 		return;
685 	tp->snd_nxt = tp->snd_fack;
686 }
687