xref: /freebsd/sys/netinet/tcp_sack.c (revision 732a02b4e77866604a120a275c082bb6221bd2ff)
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 | CTLFLAG_MPSAFE, 0,
132     "TCP SACK");
133 VNET_DEFINE(int, tcp_do_sack) = 1;
134 #define	V_tcp_do_sack			VNET(tcp_do_sack)
135 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, enable, CTLFLAG_VNET | CTLFLAG_RW,
136     &VNET_NAME(tcp_do_sack), 0, "Enable/Disable TCP SACK support");
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 
154 /*
155  * This function will find overlaps with the currently stored sackblocks
156  * and add any overlap as a dsack block upfront
157  */
158 void
159 tcp_update_dsack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
160 {
161 	struct sackblk head_blk,mid_blk,saved_blks[MAX_SACK_BLKS];
162 	int i, j, n, identical;
163 	tcp_seq start, end;
164 
165 	INP_WLOCK_ASSERT(tp->t_inpcb);
166 
167 	KASSERT(SEQ_LT(rcv_start, rcv_end), ("rcv_start < rcv_end"));
168 
169 	if (SEQ_LT(rcv_end, tp->rcv_nxt) ||
170 	    ((rcv_end == tp->rcv_nxt) &&
171 	     (tp->rcv_numsacks > 0 ) &&
172 	     (tp->sackblks[0].end == tp->rcv_nxt))) {
173 		saved_blks[0].start = rcv_start;
174 		saved_blks[0].end = rcv_end;
175 	} else {
176 		saved_blks[0].start = saved_blks[0].end = 0;
177 	}
178 
179 	head_blk.start = head_blk.end = 0;
180 	mid_blk.start = rcv_start;
181 	mid_blk.end = rcv_end;
182 	identical = 0;
183 
184 	for (i = 0; i < tp->rcv_numsacks; i++) {
185 		start = tp->sackblks[i].start;
186 		end = tp->sackblks[i].end;
187 		if (SEQ_LT(rcv_end, start)) {
188 			/* pkt left to sack blk */
189 			continue;
190 		}
191 		if (SEQ_GT(rcv_start, end)) {
192 			/* pkt right to sack blk */
193 			continue;
194 		}
195 		if (SEQ_GT(tp->rcv_nxt, end)) {
196 			if ((SEQ_MAX(rcv_start, start) != SEQ_MIN(rcv_end, end)) &&
197 			    (SEQ_GT(head_blk.start, SEQ_MAX(rcv_start, start)) ||
198 			    (head_blk.start == head_blk.end))) {
199 				head_blk.start = SEQ_MAX(rcv_start, start);
200 				head_blk.end = SEQ_MIN(rcv_end, end);
201 			}
202 			continue;
203 		}
204 		if (((head_blk.start == head_blk.end) ||
205 		     SEQ_LT(start, head_blk.start)) &&
206 		     (SEQ_GT(end, rcv_start) &&
207 		      SEQ_LEQ(start, rcv_end))) {
208 			head_blk.start = start;
209 			head_blk.end = end;
210 		}
211 		mid_blk.start = SEQ_MIN(mid_blk.start, start);
212 		mid_blk.end = SEQ_MAX(mid_blk.end, end);
213 		if ((mid_blk.start == start) &&
214 		    (mid_blk.end == end))
215 			identical = 1;
216 	}
217 	if (SEQ_LT(head_blk.start, head_blk.end)) {
218 		/* store overlapping range */
219 		saved_blks[0].start = SEQ_MAX(rcv_start, head_blk.start);
220 		saved_blks[0].end   = SEQ_MIN(rcv_end, head_blk.end);
221 	}
222 	n = 1;
223 	/*
224 	 * Second, if not ACKed, store the SACK block that
225 	 * overlaps with the DSACK block unless it is identical
226 	 */
227 	if ((SEQ_LT(tp->rcv_nxt, mid_blk.end) &&
228 	    !((mid_blk.start == saved_blks[0].start) &&
229 	    (mid_blk.end == saved_blks[0].end))) ||
230 	    identical == 1) {
231 		saved_blks[n].start = mid_blk.start;
232 		saved_blks[n++].end = mid_blk.end;
233 	}
234 	for (j = 0; (j < tp->rcv_numsacks) && (n < MAX_SACK_BLKS); j++) {
235 		if (((SEQ_LT(tp->sackblks[j].end, mid_blk.start) ||
236 		      SEQ_GT(tp->sackblks[j].start, mid_blk.end)) &&
237 		    (SEQ_GT(tp->sackblks[j].start, tp->rcv_nxt))))
238 		saved_blks[n++] = tp->sackblks[j];
239 	}
240 	j = 0;
241 	for (i = 0; i < n; i++) {
242 		/* we can end up with a stale initial entry */
243 		if (SEQ_LT(saved_blks[i].start, saved_blks[i].end)) {
244 			tp->sackblks[j++] = saved_blks[i];
245 		}
246 	}
247 	tp->rcv_numsacks = j;
248 }
249 
250 /*
251  * This function is called upon receipt of new valid data (while not in
252  * header prediction mode), and it updates the ordered list of sacks.
253  */
254 void
255 tcp_update_sack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
256 {
257 	/*
258 	 * First reported block MUST be the most recent one.  Subsequent
259 	 * blocks SHOULD be in the order in which they arrived at the
260 	 * receiver.  These two conditions make the implementation fully
261 	 * compliant with RFC 2018.
262 	 */
263 	struct sackblk head_blk, saved_blks[MAX_SACK_BLKS];
264 	int num_head, num_saved, i;
265 
266 	INP_WLOCK_ASSERT(tp->t_inpcb);
267 
268 	/* Check arguments. */
269 	KASSERT(SEQ_LEQ(rcv_start, rcv_end), ("rcv_start <= rcv_end"));
270 
271 	if ((rcv_start == rcv_end) &&
272 	    (tp->rcv_numsacks >= 1) &&
273 	    (rcv_end == tp->sackblks[0].end)) {
274 		/* retaining DSACK block below rcv_nxt (todrop) */
275 		head_blk = tp->sackblks[0];
276 	} else {
277 		/* SACK block for the received segment. */
278 		head_blk.start = rcv_start;
279 		head_blk.end = rcv_end;
280 	}
281 
282 	/*
283 	 * Merge updated SACK blocks into head_blk, and save unchanged SACK
284 	 * blocks into saved_blks[].  num_saved will have the number of the
285 	 * saved SACK blocks.
286 	 */
287 	num_saved = 0;
288 	for (i = 0; i < tp->rcv_numsacks; i++) {
289 		tcp_seq start = tp->sackblks[i].start;
290 		tcp_seq end = tp->sackblks[i].end;
291 		if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
292 			/*
293 			 * Discard this SACK block.
294 			 */
295 		} else if (SEQ_LEQ(head_blk.start, end) &&
296 			   SEQ_GEQ(head_blk.end, start)) {
297 			/*
298 			 * Merge this SACK block into head_blk.  This SACK
299 			 * block itself will be discarded.
300 			 */
301 			/*
302 			 * |-|
303 			 *   |---|  merge
304 			 *
305 			 *     |-|
306 			 * |---|    merge
307 			 *
308 			 * |-----|
309 			 *   |-|    DSACK smaller
310 			 *
311 			 *   |-|
312 			 * |-----|  DSACK smaller
313 			 */
314 			if (head_blk.start == end)
315 				head_blk.start = start;
316 			else if (head_blk.end == start)
317 				head_blk.end = end;
318 			else {
319 				if (SEQ_LT(head_blk.start, start)) {
320 					tcp_seq temp = start;
321 					start = head_blk.start;
322 					head_blk.start = temp;
323 				}
324 				if (SEQ_GT(head_blk.end, end)) {
325 					tcp_seq temp = end;
326 					end = head_blk.end;
327 					head_blk.end = temp;
328 				}
329 				if ((head_blk.start != start) ||
330 				    (head_blk.end != end)) {
331 					if ((num_saved >= 1) &&
332 					   SEQ_GEQ(saved_blks[num_saved-1].start, start) &&
333 					   SEQ_LEQ(saved_blks[num_saved-1].end, end))
334 						num_saved--;
335 					saved_blks[num_saved].start = start;
336 					saved_blks[num_saved].end = end;
337 					num_saved++;
338 				}
339 			}
340 		} else {
341 			/*
342 			 * This block supercedes the prior block
343 			 */
344 			if ((num_saved >= 1) &&
345 			   SEQ_GEQ(saved_blks[num_saved-1].start, start) &&
346 			   SEQ_LEQ(saved_blks[num_saved-1].end, end))
347 				num_saved--;
348 			/*
349 			 * Save this SACK block.
350 			 */
351 			saved_blks[num_saved].start = start;
352 			saved_blks[num_saved].end = end;
353 			num_saved++;
354 		}
355 	}
356 
357 	/*
358 	 * Update SACK list in tp->sackblks[].
359 	 */
360 	num_head = 0;
361 	if (SEQ_LT(rcv_start, rcv_end)) {
362 		/*
363 		 * The received data segment is an out-of-order segment.  Put
364 		 * head_blk at the top of SACK list.
365 		 */
366 		tp->sackblks[0] = head_blk;
367 		num_head = 1;
368 		/*
369 		 * If the number of saved SACK blocks exceeds its limit,
370 		 * discard the last SACK block.
371 		 */
372 		if (num_saved >= MAX_SACK_BLKS)
373 			num_saved--;
374 	}
375 	if ((rcv_start == rcv_end) &&
376 	    (rcv_start == tp->sackblks[0].end)) {
377 		num_head = 1;
378 	}
379 	if (num_saved > 0) {
380 		/*
381 		 * Copy the saved SACK blocks back.
382 		 */
383 		bcopy(saved_blks, &tp->sackblks[num_head],
384 		      sizeof(struct sackblk) * num_saved);
385 	}
386 
387 	/* Save the number of SACK blocks. */
388 	tp->rcv_numsacks = num_head + num_saved;
389 }
390 
391 void
392 tcp_clean_dsack_blocks(struct tcpcb *tp)
393 {
394 	struct sackblk saved_blks[MAX_SACK_BLKS];
395 	int num_saved, i;
396 
397 	INP_WLOCK_ASSERT(tp->t_inpcb);
398 	/*
399 	 * Clean up any DSACK blocks that
400 	 * are in our queue of sack blocks.
401 	 *
402 	 */
403 	num_saved = 0;
404 	for (i = 0; i < tp->rcv_numsacks; i++) {
405 		tcp_seq start = tp->sackblks[i].start;
406 		tcp_seq end = tp->sackblks[i].end;
407 		if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
408 			/*
409 			 * Discard this D-SACK block.
410 			 */
411 			continue;
412 		}
413 		/*
414 		 * Save this SACK block.
415 		 */
416 		saved_blks[num_saved].start = start;
417 		saved_blks[num_saved].end = end;
418 		num_saved++;
419 	}
420 	if (num_saved > 0) {
421 		/*
422 		 * Copy the saved SACK blocks back.
423 		 */
424 		bcopy(saved_blks, &tp->sackblks[0],
425 		      sizeof(struct sackblk) * num_saved);
426 	}
427 	tp->rcv_numsacks = num_saved;
428 }
429 
430 /*
431  * Delete all receiver-side SACK information.
432  */
433 void
434 tcp_clean_sackreport(struct tcpcb *tp)
435 {
436 	int i;
437 
438 	INP_WLOCK_ASSERT(tp->t_inpcb);
439 	tp->rcv_numsacks = 0;
440 	for (i = 0; i < MAX_SACK_BLKS; i++)
441 		tp->sackblks[i].start = tp->sackblks[i].end=0;
442 }
443 
444 /*
445  * Allocate struct sackhole.
446  */
447 static struct sackhole *
448 tcp_sackhole_alloc(struct tcpcb *tp, tcp_seq start, tcp_seq end)
449 {
450 	struct sackhole *hole;
451 
452 	if (tp->snd_numholes >= V_tcp_sack_maxholes ||
453 	    V_tcp_sack_globalholes >= V_tcp_sack_globalmaxholes) {
454 		TCPSTAT_INC(tcps_sack_sboverflow);
455 		return NULL;
456 	}
457 
458 	hole = (struct sackhole *)uma_zalloc(V_sack_hole_zone, M_NOWAIT);
459 	if (hole == NULL)
460 		return NULL;
461 
462 	hole->start = start;
463 	hole->end = end;
464 	hole->rxmit = start;
465 
466 	tp->snd_numholes++;
467 	atomic_add_int(&V_tcp_sack_globalholes, 1);
468 
469 	return hole;
470 }
471 
472 /*
473  * Free struct sackhole.
474  */
475 static void
476 tcp_sackhole_free(struct tcpcb *tp, struct sackhole *hole)
477 {
478 
479 	uma_zfree(V_sack_hole_zone, hole);
480 
481 	tp->snd_numholes--;
482 	atomic_subtract_int(&V_tcp_sack_globalholes, 1);
483 
484 	KASSERT(tp->snd_numholes >= 0, ("tp->snd_numholes >= 0"));
485 	KASSERT(V_tcp_sack_globalholes >= 0, ("tcp_sack_globalholes >= 0"));
486 }
487 
488 /*
489  * Insert new SACK hole into scoreboard.
490  */
491 static struct sackhole *
492 tcp_sackhole_insert(struct tcpcb *tp, tcp_seq start, tcp_seq end,
493     struct sackhole *after)
494 {
495 	struct sackhole *hole;
496 
497 	/* Allocate a new SACK hole. */
498 	hole = tcp_sackhole_alloc(tp, start, end);
499 	if (hole == NULL)
500 		return NULL;
501 
502 	/* Insert the new SACK hole into scoreboard. */
503 	if (after != NULL)
504 		TAILQ_INSERT_AFTER(&tp->snd_holes, after, hole, scblink);
505 	else
506 		TAILQ_INSERT_TAIL(&tp->snd_holes, hole, scblink);
507 
508 	/* Update SACK hint. */
509 	if (tp->sackhint.nexthole == NULL)
510 		tp->sackhint.nexthole = hole;
511 
512 	return hole;
513 }
514 
515 /*
516  * Remove SACK hole from scoreboard.
517  */
518 static void
519 tcp_sackhole_remove(struct tcpcb *tp, struct sackhole *hole)
520 {
521 
522 	/* Update SACK hint. */
523 	if (tp->sackhint.nexthole == hole)
524 		tp->sackhint.nexthole = TAILQ_NEXT(hole, scblink);
525 
526 	/* Remove this SACK hole. */
527 	TAILQ_REMOVE(&tp->snd_holes, hole, scblink);
528 
529 	/* Free this SACK hole. */
530 	tcp_sackhole_free(tp, hole);
531 }
532 
533 /*
534  * Process cumulative ACK and the TCP SACK option to update the scoreboard.
535  * tp->snd_holes is an ordered list of holes (oldest to newest, in terms of
536  * the sequence space).
537  * Returns 1 if incoming ACK has previously unknown SACK information,
538  * 0 otherwise. Note: We treat (snd_una, th_ack) as a sack block so any changes
539  * to that (i.e. left edge moving) would also be considered a change in SACK
540  * information which is slightly different than rfc6675.
541  */
542 int
543 tcp_sack_doack(struct tcpcb *tp, struct tcpopt *to, tcp_seq th_ack)
544 {
545 	struct sackhole *cur, *temp;
546 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1], *sblkp;
547 	int i, j, num_sack_blks, sack_changed;
548 
549 	INP_WLOCK_ASSERT(tp->t_inpcb);
550 
551 	num_sack_blks = 0;
552 	sack_changed = 0;
553 	/*
554 	 * If SND.UNA will be advanced by SEG.ACK, and if SACK holes exist,
555 	 * treat [SND.UNA, SEG.ACK) as if it is a SACK block.
556 	 */
557 	if (SEQ_LT(tp->snd_una, th_ack) && !TAILQ_EMPTY(&tp->snd_holes)) {
558 		sack_blocks[num_sack_blks].start = tp->snd_una;
559 		sack_blocks[num_sack_blks++].end = th_ack;
560 	}
561 	/*
562 	 * Append received valid SACK blocks to sack_blocks[], but only if we
563 	 * received new blocks from the other side.
564 	 */
565 	if (to->to_flags & TOF_SACK) {
566 		tp->sackhint.sacked_bytes = 0;	/* reset */
567 		for (i = 0; i < to->to_nsacks; i++) {
568 			bcopy((to->to_sacks + i * TCPOLEN_SACK),
569 			    &sack, sizeof(sack));
570 			sack.start = ntohl(sack.start);
571 			sack.end = ntohl(sack.end);
572 			if (SEQ_GT(sack.end, sack.start) &&
573 			    SEQ_GT(sack.start, tp->snd_una) &&
574 			    SEQ_GT(sack.start, th_ack) &&
575 			    SEQ_LT(sack.start, tp->snd_max) &&
576 			    SEQ_GT(sack.end, tp->snd_una) &&
577 			    SEQ_LEQ(sack.end, tp->snd_max)) {
578 				sack_blocks[num_sack_blks++] = sack;
579 				tp->sackhint.sacked_bytes +=
580 				    (sack.end-sack.start);
581 			}
582 		}
583 	}
584 	/*
585 	 * Return if SND.UNA is not advanced and no valid SACK block is
586 	 * received.
587 	 */
588 	if (num_sack_blks == 0)
589 		return (sack_changed);
590 
591 	/*
592 	 * Sort the SACK blocks so we can update the scoreboard with just one
593 	 * pass. The overhead of sorting up to 4+1 elements is less than
594 	 * making up to 4+1 passes over the scoreboard.
595 	 */
596 	for (i = 0; i < num_sack_blks; i++) {
597 		for (j = i + 1; j < num_sack_blks; j++) {
598 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
599 				sack = sack_blocks[i];
600 				sack_blocks[i] = sack_blocks[j];
601 				sack_blocks[j] = sack;
602 			}
603 		}
604 	}
605 	if (TAILQ_EMPTY(&tp->snd_holes))
606 		/*
607 		 * Empty scoreboard. Need to initialize snd_fack (it may be
608 		 * uninitialized or have a bogus value). Scoreboard holes
609 		 * (from the sack blocks received) are created later below
610 		 * (in the logic that adds holes to the tail of the
611 		 * scoreboard).
612 		 */
613 		tp->snd_fack = SEQ_MAX(tp->snd_una, th_ack);
614 	/*
615 	 * In the while-loop below, incoming SACK blocks (sack_blocks[]) and
616 	 * SACK holes (snd_holes) are traversed from their tails with just
617 	 * one pass in order to reduce the number of compares especially when
618 	 * the bandwidth-delay product is large.
619 	 *
620 	 * Note: Typically, in the first RTT of SACK recovery, the highest
621 	 * three or four SACK blocks with the same ack number are received.
622 	 * In the second RTT, if retransmitted data segments are not lost,
623 	 * the highest three or four SACK blocks with ack number advancing
624 	 * are received.
625 	 */
626 	sblkp = &sack_blocks[num_sack_blks - 1];	/* Last SACK block */
627 	tp->sackhint.last_sack_ack = sblkp->end;
628 	if (SEQ_LT(tp->snd_fack, sblkp->start)) {
629 		/*
630 		 * The highest SACK block is beyond fack.  Append new SACK
631 		 * hole at the tail.  If the second or later highest SACK
632 		 * blocks are also beyond the current fack, they will be
633 		 * inserted by way of hole splitting in the while-loop below.
634 		 */
635 		temp = tcp_sackhole_insert(tp, tp->snd_fack,sblkp->start,NULL);
636 		if (temp != NULL) {
637 			tp->snd_fack = sblkp->end;
638 			/* Go to the previous sack block. */
639 			sblkp--;
640 			sack_changed = 1;
641 		} else {
642 			/*
643 			 * We failed to add a new hole based on the current
644 			 * sack block.  Skip over all the sack blocks that
645 			 * fall completely to the right of snd_fack and
646 			 * proceed to trim the scoreboard based on the
647 			 * remaining sack blocks.  This also trims the
648 			 * scoreboard for th_ack (which is sack_blocks[0]).
649 			 */
650 			while (sblkp >= sack_blocks &&
651 			       SEQ_LT(tp->snd_fack, sblkp->start))
652 				sblkp--;
653 			if (sblkp >= sack_blocks &&
654 			    SEQ_LT(tp->snd_fack, sblkp->end))
655 				tp->snd_fack = sblkp->end;
656 		}
657 	} else if (SEQ_LT(tp->snd_fack, sblkp->end)) {
658 		/* fack is advanced. */
659 		tp->snd_fack = sblkp->end;
660 		sack_changed = 1;
661 	}
662 	cur = TAILQ_LAST(&tp->snd_holes, sackhole_head); /* Last SACK hole. */
663 	/*
664 	 * Since the incoming sack blocks are sorted, we can process them
665 	 * making one sweep of the scoreboard.
666 	 */
667 	while (sblkp >= sack_blocks  && cur != NULL) {
668 		if (SEQ_GEQ(sblkp->start, cur->end)) {
669 			/*
670 			 * SACKs data beyond the current hole.  Go to the
671 			 * previous sack block.
672 			 */
673 			sblkp--;
674 			continue;
675 		}
676 		if (SEQ_LEQ(sblkp->end, cur->start)) {
677 			/*
678 			 * SACKs data before the current hole.  Go to the
679 			 * previous hole.
680 			 */
681 			cur = TAILQ_PREV(cur, sackhole_head, scblink);
682 			continue;
683 		}
684 		tp->sackhint.sack_bytes_rexmit -= (cur->rxmit - cur->start);
685 		KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
686 		    ("sackhint bytes rtx >= 0"));
687 		sack_changed = 1;
688 		if (SEQ_LEQ(sblkp->start, cur->start)) {
689 			/* Data acks at least the beginning of hole. */
690 			if (SEQ_GEQ(sblkp->end, cur->end)) {
691 				/* Acks entire hole, so delete hole. */
692 				temp = cur;
693 				cur = TAILQ_PREV(cur, sackhole_head, scblink);
694 				tcp_sackhole_remove(tp, temp);
695 				/*
696 				 * The sack block may ack all or part of the
697 				 * next hole too, so continue onto the next
698 				 * hole.
699 				 */
700 				continue;
701 			} else {
702 				/* Move start of hole forward. */
703 				cur->start = sblkp->end;
704 				cur->rxmit = SEQ_MAX(cur->rxmit, cur->start);
705 			}
706 		} else {
707 			/* Data acks at least the end of hole. */
708 			if (SEQ_GEQ(sblkp->end, cur->end)) {
709 				/* Move end of hole backward. */
710 				cur->end = sblkp->start;
711 				cur->rxmit = SEQ_MIN(cur->rxmit, cur->end);
712 			} else {
713 				/*
714 				 * ACKs some data in middle of a hole; need
715 				 * to split current hole
716 				 */
717 				temp = tcp_sackhole_insert(tp, sblkp->end,
718 				    cur->end, cur);
719 				if (temp != NULL) {
720 					if (SEQ_GT(cur->rxmit, temp->rxmit)) {
721 						temp->rxmit = cur->rxmit;
722 						tp->sackhint.sack_bytes_rexmit
723 						    += (temp->rxmit
724 						    - temp->start);
725 					}
726 					cur->end = sblkp->start;
727 					cur->rxmit = SEQ_MIN(cur->rxmit,
728 					    cur->end);
729 				}
730 			}
731 		}
732 		tp->sackhint.sack_bytes_rexmit += (cur->rxmit - cur->start);
733 		/*
734 		 * Testing sblkp->start against cur->start tells us whether
735 		 * we're done with the sack block or the sack hole.
736 		 * Accordingly, we advance one or the other.
737 		 */
738 		if (SEQ_LEQ(sblkp->start, cur->start))
739 			cur = TAILQ_PREV(cur, sackhole_head, scblink);
740 		else
741 			sblkp--;
742 	}
743 	return (sack_changed);
744 }
745 
746 /*
747  * Free all SACK holes to clear the scoreboard.
748  */
749 void
750 tcp_free_sackholes(struct tcpcb *tp)
751 {
752 	struct sackhole *q;
753 
754 	INP_WLOCK_ASSERT(tp->t_inpcb);
755 	while ((q = TAILQ_FIRST(&tp->snd_holes)) != NULL)
756 		tcp_sackhole_remove(tp, q);
757 	tp->sackhint.sack_bytes_rexmit = 0;
758 
759 	KASSERT(tp->snd_numholes == 0, ("tp->snd_numholes == 0"));
760 	KASSERT(tp->sackhint.nexthole == NULL,
761 		("tp->sackhint.nexthole == NULL"));
762 }
763 
764 /*
765  * Partial ack handling within a sack recovery episode.  Keeping this very
766  * simple for now.  When a partial ack is received, force snd_cwnd to a value
767  * that will allow the sender to transmit no more than 2 segments.  If
768  * necessary, a better scheme can be adopted at a later point, but for now,
769  * the goal is to prevent the sender from bursting a large amount of data in
770  * the midst of sack recovery.
771  */
772 void
773 tcp_sack_partialack(struct tcpcb *tp, struct tcphdr *th)
774 {
775 	int num_segs = 1;
776 
777 	INP_WLOCK_ASSERT(tp->t_inpcb);
778 	tcp_timer_activate(tp, TT_REXMT, 0);
779 	tp->t_rtttime = 0;
780 	/* Send one or 2 segments based on how much new data was acked. */
781 	if ((BYTES_THIS_ACK(tp, th) / tp->t_maxseg) >= 2)
782 		num_segs = 2;
783 	tp->snd_cwnd = (tp->sackhint.sack_bytes_rexmit +
784 	    (tp->snd_nxt - tp->snd_recover) + num_segs * tp->t_maxseg);
785 	if (tp->snd_cwnd > tp->snd_ssthresh)
786 		tp->snd_cwnd = tp->snd_ssthresh;
787 	tp->t_flags |= TF_ACKNOW;
788 	(void) tp->t_fb->tfb_tcp_output(tp);
789 }
790 
791 #if 0
792 /*
793  * Debug version of tcp_sack_output() that walks the scoreboard.  Used for
794  * now to sanity check the hint.
795  */
796 static struct sackhole *
797 tcp_sack_output_debug(struct tcpcb *tp, int *sack_bytes_rexmt)
798 {
799 	struct sackhole *p;
800 
801 	INP_WLOCK_ASSERT(tp->t_inpcb);
802 	*sack_bytes_rexmt = 0;
803 	TAILQ_FOREACH(p, &tp->snd_holes, scblink) {
804 		if (SEQ_LT(p->rxmit, p->end)) {
805 			if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */
806 				continue;
807 			}
808 			*sack_bytes_rexmt += (p->rxmit - p->start);
809 			break;
810 		}
811 		*sack_bytes_rexmt += (p->rxmit - p->start);
812 	}
813 	return (p);
814 }
815 #endif
816 
817 /*
818  * Returns the next hole to retransmit and the number of retransmitted bytes
819  * from the scoreboard.  We store both the next hole and the number of
820  * retransmitted bytes as hints (and recompute these on the fly upon SACK/ACK
821  * reception).  This avoids scoreboard traversals completely.
822  *
823  * The loop here will traverse *at most* one link.  Here's the argument.  For
824  * the loop to traverse more than 1 link before finding the next hole to
825  * retransmit, we would need to have at least 1 node following the current
826  * hint with (rxmit == end).  But, for all holes following the current hint,
827  * (start == rxmit), since we have not yet retransmitted from them.
828  * Therefore, in order to traverse more 1 link in the loop below, we need to
829  * have at least one node following the current hint with (start == rxmit ==
830  * end).  But that can't happen, (start == end) means that all the data in
831  * that hole has been sacked, in which case, the hole would have been removed
832  * from the scoreboard.
833  */
834 struct sackhole *
835 tcp_sack_output(struct tcpcb *tp, int *sack_bytes_rexmt)
836 {
837 	struct sackhole *hole = NULL;
838 
839 	INP_WLOCK_ASSERT(tp->t_inpcb);
840 	*sack_bytes_rexmt = tp->sackhint.sack_bytes_rexmit;
841 	hole = tp->sackhint.nexthole;
842 	if (hole == NULL || SEQ_LT(hole->rxmit, hole->end))
843 		goto out;
844 	while ((hole = TAILQ_NEXT(hole, scblink)) != NULL) {
845 		if (SEQ_LT(hole->rxmit, hole->end)) {
846 			tp->sackhint.nexthole = hole;
847 			break;
848 		}
849 	}
850 out:
851 	return (hole);
852 }
853 
854 /*
855  * After a timeout, the SACK list may be rebuilt.  This SACK information
856  * should be used to avoid retransmitting SACKed data.  This function
857  * traverses the SACK list to see if snd_nxt should be moved forward.
858  */
859 void
860 tcp_sack_adjust(struct tcpcb *tp)
861 {
862 	struct sackhole *p, *cur = TAILQ_FIRST(&tp->snd_holes);
863 
864 	INP_WLOCK_ASSERT(tp->t_inpcb);
865 	if (cur == NULL)
866 		return; /* No holes */
867 	if (SEQ_GEQ(tp->snd_nxt, tp->snd_fack))
868 		return; /* We're already beyond any SACKed blocks */
869 	/*-
870 	 * Two cases for which we want to advance snd_nxt:
871 	 * i) snd_nxt lies between end of one hole and beginning of another
872 	 * ii) snd_nxt lies between end of last hole and snd_fack
873 	 */
874 	while ((p = TAILQ_NEXT(cur, scblink)) != NULL) {
875 		if (SEQ_LT(tp->snd_nxt, cur->end))
876 			return;
877 		if (SEQ_GEQ(tp->snd_nxt, p->start))
878 			cur = p;
879 		else {
880 			tp->snd_nxt = p->start;
881 			return;
882 		}
883 	}
884 	if (SEQ_LT(tp->snd_nxt, cur->end))
885 		return;
886 	tp->snd_nxt = tp->snd_fack;
887 }
888