xref: /freebsd/sys/netinet/tcp_sack.c (revision 8a271827e7b5d5310e06df1f9f49ba0ef9efd263)
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 
81 #include <sys/param.h>
82 #include <sys/systm.h>
83 #include <sys/kernel.h>
84 #include <sys/sysctl.h>
85 #include <sys/malloc.h>
86 #include <sys/mbuf.h>
87 #include <sys/proc.h>		/* for proc0 declaration */
88 #include <sys/protosw.h>
89 #include <sys/socket.h>
90 #include <sys/socketvar.h>
91 #include <sys/syslog.h>
92 #include <sys/systm.h>
93 
94 #include <machine/cpu.h>	/* before tcp_seq.h, for tcp_random18() */
95 
96 #include <vm/uma.h>
97 
98 #include <net/if.h>
99 #include <net/if_var.h>
100 #include <net/route.h>
101 #include <net/vnet.h>
102 
103 #include <netinet/in.h>
104 #include <netinet/in_systm.h>
105 #include <netinet/ip.h>
106 #include <netinet/in_var.h>
107 #include <netinet/in_pcb.h>
108 #include <netinet/ip_var.h>
109 #include <netinet/ip6.h>
110 #include <netinet/icmp6.h>
111 #include <netinet6/nd6.h>
112 #include <netinet6/ip6_var.h>
113 #include <netinet6/in6_pcb.h>
114 #include <netinet/tcp.h>
115 #include <netinet/tcp_fsm.h>
116 #include <netinet/tcp_seq.h>
117 #include <netinet/tcp_timer.h>
118 #include <netinet/tcp_var.h>
119 #include <netinet/tcpip.h>
120 #include <netinet/cc/cc.h>
121 
122 #include <machine/in_cksum.h>
123 
124 VNET_DECLARE(struct uma_zone *, sack_hole_zone);
125 #define	V_sack_hole_zone		VNET(sack_hole_zone)
126 
127 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, sack, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
128     "TCP SACK");
129 
130 VNET_DEFINE(int, tcp_do_sack) = 1;
131 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, enable, CTLFLAG_VNET | CTLFLAG_RW,
132     &VNET_NAME(tcp_do_sack), 0,
133     "Enable/Disable TCP SACK support");
134 
135 VNET_DEFINE(int, tcp_do_newsack) = 1;
136 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, revised, CTLFLAG_VNET | CTLFLAG_RW,
137     &VNET_NAME(tcp_do_newsack), 0,
138     "Use revised SACK loss recovery per RFC 6675");
139 
140 VNET_DEFINE(int, tcp_sack_maxholes) = 128;
141 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, maxholes, CTLFLAG_VNET | CTLFLAG_RW,
142     &VNET_NAME(tcp_sack_maxholes), 0,
143     "Maximum number of TCP SACK holes allowed per connection");
144 
145 VNET_DEFINE(int, tcp_sack_globalmaxholes) = 65536;
146 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, globalmaxholes, CTLFLAG_VNET | CTLFLAG_RW,
147     &VNET_NAME(tcp_sack_globalmaxholes), 0,
148     "Global maximum number of TCP SACK holes");
149 
150 VNET_DEFINE(int, tcp_sack_globalholes) = 0;
151 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, globalholes, CTLFLAG_VNET | CTLFLAG_RD,
152     &VNET_NAME(tcp_sack_globalholes), 0,
153     "Global number of TCP SACK holes currently allocated");
154 
155 int
156 tcp_dsack_block_exists(struct tcpcb *tp)
157 {
158 	/* Return true if a DSACK block exists */
159 	if (tp->rcv_numsacks == 0)
160 		return (0);
161 	if (SEQ_LEQ(tp->sackblks[0].end, tp->rcv_nxt))
162 		return(1);
163 	return (0);
164 }
165 
166 /*
167  * This function will find overlaps with the currently stored sackblocks
168  * and add any overlap as a dsack block upfront
169  */
170 void
171 tcp_update_dsack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
172 {
173 	struct sackblk head_blk,mid_blk,saved_blks[MAX_SACK_BLKS];
174 	int i, j, n, identical;
175 	tcp_seq start, end;
176 
177 	INP_WLOCK_ASSERT(tptoinpcb(tp));
178 
179 	KASSERT(SEQ_LT(rcv_start, rcv_end), ("rcv_start < rcv_end"));
180 
181 	if (SEQ_LT(rcv_end, tp->rcv_nxt) ||
182 	    ((rcv_end == tp->rcv_nxt) &&
183 	     (tp->rcv_numsacks > 0 ) &&
184 	     (tp->sackblks[0].end == tp->rcv_nxt))) {
185 		saved_blks[0].start = rcv_start;
186 		saved_blks[0].end = rcv_end;
187 	} else {
188 		saved_blks[0].start = saved_blks[0].end = 0;
189 	}
190 
191 	head_blk.start = head_blk.end = 0;
192 	mid_blk.start = rcv_start;
193 	mid_blk.end = rcv_end;
194 	identical = 0;
195 
196 	for (i = 0; i < tp->rcv_numsacks; i++) {
197 		start = tp->sackblks[i].start;
198 		end = tp->sackblks[i].end;
199 		if (SEQ_LT(rcv_end, start)) {
200 			/* pkt left to sack blk */
201 			continue;
202 		}
203 		if (SEQ_GT(rcv_start, end)) {
204 			/* pkt right to sack blk */
205 			continue;
206 		}
207 		if (SEQ_GT(tp->rcv_nxt, end)) {
208 			if ((SEQ_MAX(rcv_start, start) != SEQ_MIN(rcv_end, end)) &&
209 			    (SEQ_GT(head_blk.start, SEQ_MAX(rcv_start, start)) ||
210 			    (head_blk.start == head_blk.end))) {
211 				head_blk.start = SEQ_MAX(rcv_start, start);
212 				head_blk.end = SEQ_MIN(rcv_end, end);
213 			}
214 			continue;
215 		}
216 		if (((head_blk.start == head_blk.end) ||
217 		     SEQ_LT(start, head_blk.start)) &&
218 		     (SEQ_GT(end, rcv_start) &&
219 		      SEQ_LEQ(start, rcv_end))) {
220 			head_blk.start = start;
221 			head_blk.end = end;
222 		}
223 		mid_blk.start = SEQ_MIN(mid_blk.start, start);
224 		mid_blk.end = SEQ_MAX(mid_blk.end, end);
225 		if ((mid_blk.start == start) &&
226 		    (mid_blk.end == end))
227 			identical = 1;
228 	}
229 	if (SEQ_LT(head_blk.start, head_blk.end)) {
230 		/* store overlapping range */
231 		saved_blks[0].start = SEQ_MAX(rcv_start, head_blk.start);
232 		saved_blks[0].end   = SEQ_MIN(rcv_end, head_blk.end);
233 	}
234 	n = 1;
235 	/*
236 	 * Second, if not ACKed, store the SACK block that
237 	 * overlaps with the DSACK block unless it is identical
238 	 */
239 	if ((SEQ_LT(tp->rcv_nxt, mid_blk.end) &&
240 	    !((mid_blk.start == saved_blks[0].start) &&
241 	    (mid_blk.end == saved_blks[0].end))) ||
242 	    identical == 1) {
243 		saved_blks[n].start = mid_blk.start;
244 		saved_blks[n++].end = mid_blk.end;
245 	}
246 	for (j = 0; (j < tp->rcv_numsacks) && (n < MAX_SACK_BLKS); j++) {
247 		if (((SEQ_LT(tp->sackblks[j].end, mid_blk.start) ||
248 		      SEQ_GT(tp->sackblks[j].start, mid_blk.end)) &&
249 		    (SEQ_GT(tp->sackblks[j].start, tp->rcv_nxt))))
250 		saved_blks[n++] = tp->sackblks[j];
251 	}
252 	j = 0;
253 	for (i = 0; i < n; i++) {
254 		/* we can end up with a stale initial entry */
255 		if (SEQ_LT(saved_blks[i].start, saved_blks[i].end)) {
256 			tp->sackblks[j++] = saved_blks[i];
257 		}
258 	}
259 	tp->rcv_numsacks = j;
260 }
261 
262 /*
263  * This function is called upon receipt of new valid data (while not in
264  * header prediction mode), and it updates the ordered list of sacks.
265  */
266 void
267 tcp_update_sack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
268 {
269 	/*
270 	 * First reported block MUST be the most recent one.  Subsequent
271 	 * blocks SHOULD be in the order in which they arrived at the
272 	 * receiver.  These two conditions make the implementation fully
273 	 * compliant with RFC 2018.
274 	 */
275 	struct sackblk head_blk, saved_blks[MAX_SACK_BLKS];
276 	int num_head, num_saved, i;
277 
278 	INP_WLOCK_ASSERT(tptoinpcb(tp));
279 
280 	/* Check arguments. */
281 	KASSERT(SEQ_LEQ(rcv_start, rcv_end), ("rcv_start <= rcv_end"));
282 
283 	if ((rcv_start == rcv_end) &&
284 	    (tp->rcv_numsacks >= 1) &&
285 	    (rcv_end == tp->sackblks[0].end)) {
286 		/* retaining DSACK block below rcv_nxt (todrop) */
287 		head_blk = tp->sackblks[0];
288 	} else {
289 		/* SACK block for the received segment. */
290 		head_blk.start = rcv_start;
291 		head_blk.end = rcv_end;
292 	}
293 
294 	/*
295 	 * Merge updated SACK blocks into head_blk, and save unchanged SACK
296 	 * blocks into saved_blks[].  num_saved will have the number of the
297 	 * saved SACK blocks.
298 	 */
299 	num_saved = 0;
300 	for (i = 0; i < tp->rcv_numsacks; i++) {
301 		tcp_seq start = tp->sackblks[i].start;
302 		tcp_seq end = tp->sackblks[i].end;
303 		if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
304 			/*
305 			 * Discard this SACK block.
306 			 */
307 		} else if (SEQ_LEQ(head_blk.start, end) &&
308 			   SEQ_GEQ(head_blk.end, start)) {
309 			/*
310 			 * Merge this SACK block into head_blk.  This SACK
311 			 * block itself will be discarded.
312 			 */
313 			/*
314 			 * |-|
315 			 *   |---|  merge
316 			 *
317 			 *     |-|
318 			 * |---|    merge
319 			 *
320 			 * |-----|
321 			 *   |-|    DSACK smaller
322 			 *
323 			 *   |-|
324 			 * |-----|  DSACK smaller
325 			 */
326 			if (head_blk.start == end)
327 				head_blk.start = start;
328 			else if (head_blk.end == start)
329 				head_blk.end = end;
330 			else {
331 				if (SEQ_LT(head_blk.start, start)) {
332 					tcp_seq temp = start;
333 					start = head_blk.start;
334 					head_blk.start = temp;
335 				}
336 				if (SEQ_GT(head_blk.end, end)) {
337 					tcp_seq temp = end;
338 					end = head_blk.end;
339 					head_blk.end = temp;
340 				}
341 				if ((head_blk.start != start) ||
342 				    (head_blk.end != end)) {
343 					if ((num_saved >= 1) &&
344 					   SEQ_GEQ(saved_blks[num_saved-1].start, start) &&
345 					   SEQ_LEQ(saved_blks[num_saved-1].end, end))
346 						num_saved--;
347 					saved_blks[num_saved].start = start;
348 					saved_blks[num_saved].end = end;
349 					num_saved++;
350 				}
351 			}
352 		} else {
353 			/*
354 			 * This block supercedes the prior block
355 			 */
356 			if ((num_saved >= 1) &&
357 			   SEQ_GEQ(saved_blks[num_saved-1].start, start) &&
358 			   SEQ_LEQ(saved_blks[num_saved-1].end, end))
359 				num_saved--;
360 			/*
361 			 * Save this SACK block.
362 			 */
363 			saved_blks[num_saved].start = start;
364 			saved_blks[num_saved].end = end;
365 			num_saved++;
366 		}
367 	}
368 
369 	/*
370 	 * Update SACK list in tp->sackblks[].
371 	 */
372 	num_head = 0;
373 	if (SEQ_LT(rcv_start, rcv_end)) {
374 		/*
375 		 * The received data segment is an out-of-order segment.  Put
376 		 * head_blk at the top of SACK list.
377 		 */
378 		tp->sackblks[0] = head_blk;
379 		num_head = 1;
380 		/*
381 		 * If the number of saved SACK blocks exceeds its limit,
382 		 * discard the last SACK block.
383 		 */
384 		if (num_saved >= MAX_SACK_BLKS)
385 			num_saved--;
386 	}
387 	if ((rcv_start == rcv_end) &&
388 	    (rcv_start == tp->sackblks[0].end)) {
389 		num_head = 1;
390 	}
391 	if (num_saved > 0) {
392 		/*
393 		 * Copy the saved SACK blocks back.
394 		 */
395 		bcopy(saved_blks, &tp->sackblks[num_head],
396 		      sizeof(struct sackblk) * num_saved);
397 	}
398 
399 	/* Save the number of SACK blocks. */
400 	tp->rcv_numsacks = num_head + num_saved;
401 }
402 
403 void
404 tcp_clean_dsack_blocks(struct tcpcb *tp)
405 {
406 	struct sackblk saved_blks[MAX_SACK_BLKS];
407 	int num_saved, i;
408 
409 	INP_WLOCK_ASSERT(tptoinpcb(tp));
410 	/*
411 	 * Clean up any DSACK blocks that
412 	 * are in our queue of sack blocks.
413 	 *
414 	 */
415 	num_saved = 0;
416 	for (i = 0; i < tp->rcv_numsacks; i++) {
417 		tcp_seq start = tp->sackblks[i].start;
418 		tcp_seq end = tp->sackblks[i].end;
419 		if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
420 			/*
421 			 * Discard this D-SACK block.
422 			 */
423 			continue;
424 		}
425 		/*
426 		 * Save this SACK block.
427 		 */
428 		saved_blks[num_saved].start = start;
429 		saved_blks[num_saved].end = end;
430 		num_saved++;
431 	}
432 	if (num_saved > 0) {
433 		/*
434 		 * Copy the saved SACK blocks back.
435 		 */
436 		bcopy(saved_blks, &tp->sackblks[0],
437 		      sizeof(struct sackblk) * num_saved);
438 	}
439 	tp->rcv_numsacks = num_saved;
440 }
441 
442 /*
443  * Delete all receiver-side SACK information.
444  */
445 void
446 tcp_clean_sackreport(struct tcpcb *tp)
447 {
448 	int i;
449 
450 	INP_WLOCK_ASSERT(tptoinpcb(tp));
451 	tp->rcv_numsacks = 0;
452 	for (i = 0; i < MAX_SACK_BLKS; i++)
453 		tp->sackblks[i].start = tp->sackblks[i].end=0;
454 }
455 
456 /*
457  * Allocate struct sackhole.
458  */
459 static struct sackhole *
460 tcp_sackhole_alloc(struct tcpcb *tp, tcp_seq start, tcp_seq end)
461 {
462 	struct sackhole *hole;
463 
464 	if (tp->snd_numholes >= V_tcp_sack_maxholes ||
465 	    V_tcp_sack_globalholes >= V_tcp_sack_globalmaxholes) {
466 		TCPSTAT_INC(tcps_sack_sboverflow);
467 		return NULL;
468 	}
469 
470 	hole = (struct sackhole *)uma_zalloc(V_sack_hole_zone, M_NOWAIT);
471 	if (hole == NULL)
472 		return NULL;
473 
474 	hole->start = start;
475 	hole->end = end;
476 	hole->rxmit = start;
477 
478 	tp->snd_numholes++;
479 	atomic_add_int(&V_tcp_sack_globalholes, 1);
480 
481 	return hole;
482 }
483 
484 /*
485  * Free struct sackhole.
486  */
487 static void
488 tcp_sackhole_free(struct tcpcb *tp, struct sackhole *hole)
489 {
490 
491 	uma_zfree(V_sack_hole_zone, hole);
492 
493 	tp->snd_numholes--;
494 	atomic_subtract_int(&V_tcp_sack_globalholes, 1);
495 
496 	KASSERT(tp->snd_numholes >= 0, ("tp->snd_numholes >= 0"));
497 	KASSERT(V_tcp_sack_globalholes >= 0, ("tcp_sack_globalholes >= 0"));
498 }
499 
500 /*
501  * Insert new SACK hole into scoreboard.
502  */
503 static struct sackhole *
504 tcp_sackhole_insert(struct tcpcb *tp, tcp_seq start, tcp_seq end,
505     struct sackhole *after)
506 {
507 	struct sackhole *hole;
508 
509 	/* Allocate a new SACK hole. */
510 	hole = tcp_sackhole_alloc(tp, start, end);
511 	if (hole == NULL)
512 		return NULL;
513 
514 	/* Insert the new SACK hole into scoreboard. */
515 	if (after != NULL)
516 		TAILQ_INSERT_AFTER(&tp->snd_holes, after, hole, scblink);
517 	else
518 		TAILQ_INSERT_TAIL(&tp->snd_holes, hole, scblink);
519 
520 	/* Update SACK hint. */
521 	if (tp->sackhint.nexthole == NULL)
522 		tp->sackhint.nexthole = hole;
523 
524 	return hole;
525 }
526 
527 /*
528  * Remove SACK hole from scoreboard.
529  */
530 static void
531 tcp_sackhole_remove(struct tcpcb *tp, struct sackhole *hole)
532 {
533 
534 	/* Update SACK hint. */
535 	if (tp->sackhint.nexthole == hole)
536 		tp->sackhint.nexthole = TAILQ_NEXT(hole, scblink);
537 
538 	/* Remove this SACK hole. */
539 	TAILQ_REMOVE(&tp->snd_holes, hole, scblink);
540 
541 	/* Free this SACK hole. */
542 	tcp_sackhole_free(tp, hole);
543 }
544 
545 /*
546  * Process cumulative ACK and the TCP SACK option to update the scoreboard.
547  * tp->snd_holes is an ordered list of holes (oldest to newest, in terms of
548  * the sequence space).
549  * Returns 1 if incoming ACK has previously unknown SACK information,
550  * 0 otherwise.
551  */
552 int
553 tcp_sack_doack(struct tcpcb *tp, struct tcpopt *to, tcp_seq th_ack)
554 {
555 	struct sackhole *cur, *temp;
556 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1], *sblkp;
557 	int i, j, num_sack_blks, sack_changed;
558 	int delivered_data, left_edge_delta;
559 
560 	INP_WLOCK_ASSERT(tptoinpcb(tp));
561 
562 	num_sack_blks = 0;
563 	sack_changed = 0;
564 	delivered_data = 0;
565 	left_edge_delta = 0;
566 	/*
567 	 * If SND.UNA will be advanced by SEG.ACK, and if SACK holes exist,
568 	 * treat [SND.UNA, SEG.ACK) as if it is a SACK block.
569 	 * Account changes to SND.UNA always in delivered data.
570 	 */
571 	if (SEQ_LT(tp->snd_una, th_ack) && !TAILQ_EMPTY(&tp->snd_holes)) {
572 		left_edge_delta = th_ack - tp->snd_una;
573 		sack_blocks[num_sack_blks].start = tp->snd_una;
574 		sack_blocks[num_sack_blks++].end = th_ack;
575 		/*
576 		 * Pulling snd_fack forward if we got here
577 		 * due to DSACK blocks
578 		 */
579 		if (SEQ_LT(tp->snd_fack, th_ack)) {
580 			delivered_data += th_ack - tp->snd_una;
581 			tp->snd_fack = th_ack;
582 			sack_changed = 1;
583 		}
584 	}
585 	/*
586 	 * Append received valid SACK blocks to sack_blocks[], but only if we
587 	 * received new blocks from the other side.
588 	 */
589 	if (to->to_flags & TOF_SACK) {
590 		for (i = 0; i < to->to_nsacks; i++) {
591 			bcopy((to->to_sacks + i * TCPOLEN_SACK),
592 			    &sack, sizeof(sack));
593 			sack.start = ntohl(sack.start);
594 			sack.end = ntohl(sack.end);
595 			if (SEQ_GT(sack.end, sack.start) &&
596 			    SEQ_GT(sack.start, tp->snd_una) &&
597 			    SEQ_GT(sack.start, th_ack) &&
598 			    SEQ_LT(sack.start, tp->snd_max) &&
599 			    SEQ_GT(sack.end, tp->snd_una) &&
600 			    SEQ_LEQ(sack.end, tp->snd_max)) {
601 				sack_blocks[num_sack_blks++] = sack;
602 			} else if (SEQ_LEQ(sack.start, th_ack) &&
603 			    SEQ_LEQ(sack.end, th_ack)) {
604 				/*
605 				 * Its a D-SACK block.
606 				 */
607 				tcp_record_dsack(tp, sack.start, sack.end, 0);
608 			}
609 		}
610 	}
611 	/*
612 	 * Return if SND.UNA is not advanced and no valid SACK block is
613 	 * received.
614 	 */
615 	if (num_sack_blks == 0)
616 		return (sack_changed);
617 
618 	/*
619 	 * Sort the SACK blocks so we can update the scoreboard with just one
620 	 * pass. The overhead of sorting up to 4+1 elements is less than
621 	 * making up to 4+1 passes over the scoreboard.
622 	 */
623 	for (i = 0; i < num_sack_blks; i++) {
624 		for (j = i + 1; j < num_sack_blks; j++) {
625 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
626 				sack = sack_blocks[i];
627 				sack_blocks[i] = sack_blocks[j];
628 				sack_blocks[j] = sack;
629 			}
630 		}
631 	}
632 	if (TAILQ_EMPTY(&tp->snd_holes)) {
633 		/*
634 		 * Empty scoreboard. Need to initialize snd_fack (it may be
635 		 * uninitialized or have a bogus value). Scoreboard holes
636 		 * (from the sack blocks received) are created later below
637 		 * (in the logic that adds holes to the tail of the
638 		 * scoreboard).
639 		 */
640 		tp->snd_fack = SEQ_MAX(tp->snd_una, th_ack);
641 		tp->sackhint.sacked_bytes = 0;	/* reset */
642 	}
643 	/*
644 	 * In the while-loop below, incoming SACK blocks (sack_blocks[]) and
645 	 * SACK holes (snd_holes) are traversed from their tails with just
646 	 * one pass in order to reduce the number of compares especially when
647 	 * the bandwidth-delay product is large.
648 	 *
649 	 * Note: Typically, in the first RTT of SACK recovery, the highest
650 	 * three or four SACK blocks with the same ack number are received.
651 	 * In the second RTT, if retransmitted data segments are not lost,
652 	 * the highest three or four SACK blocks with ack number advancing
653 	 * are received.
654 	 */
655 	sblkp = &sack_blocks[num_sack_blks - 1];	/* Last SACK block */
656 	tp->sackhint.last_sack_ack = sblkp->end;
657 	if (SEQ_LT(tp->snd_fack, sblkp->start)) {
658 		/*
659 		 * The highest SACK block is beyond fack.  First,
660 		 * check if there was a successful Rescue Retransmission,
661 		 * and move this hole left. With normal holes, snd_fack
662 		 * is always to the right of the end.
663 		 */
664 		if (((temp = TAILQ_LAST(&tp->snd_holes, sackhole_head)) != NULL) &&
665 		    SEQ_LEQ(tp->snd_fack,temp->end)) {
666 			temp->start = SEQ_MAX(tp->snd_fack, SEQ_MAX(tp->snd_una, th_ack));
667 			temp->end = sblkp->start;
668 			temp->rxmit = temp->start;
669 			delivered_data += sblkp->end - sblkp->start;
670 			tp->snd_fack = sblkp->end;
671 			sblkp--;
672 			sack_changed = 1;
673 		} else {
674 			/*
675 			 * Append a new SACK hole at the tail.  If the
676 			 * second or later highest SACK blocks are also
677 			 * beyond the current fack, they will be inserted
678 			 * by way of hole splitting in the while-loop below.
679 			 */
680 			temp = tcp_sackhole_insert(tp, tp->snd_fack,sblkp->start,NULL);
681 			if (temp != NULL) {
682 				delivered_data += sblkp->end - sblkp->start;
683 				tp->snd_fack = sblkp->end;
684 				/* Go to the previous sack block. */
685 				sblkp--;
686 				sack_changed = 1;
687 			} else {
688 				/*
689 				 * We failed to add a new hole based on the current
690 				 * sack block.  Skip over all the sack blocks that
691 				 * fall completely to the right of snd_fack and
692 				 * proceed to trim the scoreboard based on the
693 				 * remaining sack blocks.  This also trims the
694 				 * scoreboard for th_ack (which is sack_blocks[0]).
695 				 */
696 				while (sblkp >= sack_blocks &&
697 				       SEQ_LT(tp->snd_fack, sblkp->start))
698 					sblkp--;
699 				if (sblkp >= sack_blocks &&
700 				    SEQ_LT(tp->snd_fack, sblkp->end)) {
701 					delivered_data += sblkp->end - tp->snd_fack;
702 					tp->snd_fack = sblkp->end;
703 					sack_changed = 1;
704 				}
705 			}
706 		}
707 	} else if (SEQ_LT(tp->snd_fack, sblkp->end)) {
708 		/* fack is advanced. */
709 		delivered_data += sblkp->end - tp->snd_fack;
710 		tp->snd_fack = sblkp->end;
711 		sack_changed = 1;
712 	}
713 	cur = TAILQ_LAST(&tp->snd_holes, sackhole_head); /* Last SACK hole. */
714 	/*
715 	 * Since the incoming sack blocks are sorted, we can process them
716 	 * making one sweep of the scoreboard.
717 	 */
718 	while (sblkp >= sack_blocks  && cur != NULL) {
719 		if (SEQ_GEQ(sblkp->start, cur->end)) {
720 			/*
721 			 * SACKs data beyond the current hole.  Go to the
722 			 * previous sack block.
723 			 */
724 			sblkp--;
725 			continue;
726 		}
727 		if (SEQ_LEQ(sblkp->end, cur->start)) {
728 			/*
729 			 * SACKs data before the current hole.  Go to the
730 			 * previous hole.
731 			 */
732 			cur = TAILQ_PREV(cur, sackhole_head, scblink);
733 			continue;
734 		}
735 		tp->sackhint.sack_bytes_rexmit -=
736 		    (SEQ_MIN(cur->rxmit, cur->end) - cur->start);
737 		KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
738 		    ("sackhint bytes rtx >= 0"));
739 		sack_changed = 1;
740 		if (SEQ_LEQ(sblkp->start, cur->start)) {
741 			/* Data acks at least the beginning of hole. */
742 			if (SEQ_GEQ(sblkp->end, cur->end)) {
743 				/* Acks entire hole, so delete hole. */
744 				delivered_data += (cur->end - cur->start);
745 				temp = cur;
746 				cur = TAILQ_PREV(cur, sackhole_head, scblink);
747 				tcp_sackhole_remove(tp, temp);
748 				/*
749 				 * The sack block may ack all or part of the
750 				 * next hole too, so continue onto the next
751 				 * hole.
752 				 */
753 				continue;
754 			} else {
755 				/* Move start of hole forward. */
756 				delivered_data += (sblkp->end - cur->start);
757 				cur->start = sblkp->end;
758 				cur->rxmit = SEQ_MAX(cur->rxmit, cur->start);
759 			}
760 		} else {
761 			/* Data acks at least the end of hole. */
762 			if (SEQ_GEQ(sblkp->end, cur->end)) {
763 				/* Move end of hole backward. */
764 				delivered_data += (cur->end - sblkp->start);
765 				cur->end = sblkp->start;
766 				cur->rxmit = SEQ_MIN(cur->rxmit, cur->end);
767 				if ((tp->t_flags & TF_LRD) && SEQ_GEQ(cur->rxmit, cur->end))
768 					cur->rxmit = tp->snd_recover;
769 			} else {
770 				/*
771 				 * ACKs some data in middle of a hole; need
772 				 * to split current hole
773 				 */
774 				temp = tcp_sackhole_insert(tp, sblkp->end,
775 				    cur->end, cur);
776 				if (temp != NULL) {
777 					if (SEQ_GT(cur->rxmit, temp->rxmit)) {
778 						temp->rxmit = cur->rxmit;
779 						tp->sackhint.sack_bytes_rexmit +=
780 						    (SEQ_MIN(temp->rxmit,
781 						    temp->end) - temp->start);
782 					}
783 					cur->end = sblkp->start;
784 					cur->rxmit = SEQ_MIN(cur->rxmit,
785 					    cur->end);
786 					if ((tp->t_flags & TF_LRD) && SEQ_GEQ(cur->rxmit, cur->end))
787 						cur->rxmit = tp->snd_recover;
788 					delivered_data += (sblkp->end - sblkp->start);
789 				}
790 			}
791 		}
792 		tp->sackhint.sack_bytes_rexmit +=
793 		    (SEQ_MIN(cur->rxmit, cur->end) - cur->start);
794 		/*
795 		 * Testing sblkp->start against cur->start tells us whether
796 		 * we're done with the sack block or the sack hole.
797 		 * Accordingly, we advance one or the other.
798 		 */
799 		if (SEQ_LEQ(sblkp->start, cur->start))
800 			cur = TAILQ_PREV(cur, sackhole_head, scblink);
801 		else
802 			sblkp--;
803 	}
804 	if (!(to->to_flags & TOF_SACK))
805 		/*
806 		 * If this ACK did not contain any
807 		 * SACK blocks, any only moved the
808 		 * left edge right, it is a pure
809 		 * cumulative ACK. Do not count
810 		 * DupAck for this. Also required
811 		 * for RFC6675 rescue retransmission.
812 		 */
813 		sack_changed = 0;
814 	tp->sackhint.delivered_data = delivered_data;
815 	tp->sackhint.sacked_bytes += delivered_data - left_edge_delta;
816 	KASSERT((delivered_data >= 0), ("delivered_data < 0"));
817 	KASSERT((tp->sackhint.sacked_bytes >= 0), ("sacked_bytes < 0"));
818 	return (sack_changed);
819 }
820 
821 /*
822  * Free all SACK holes to clear the scoreboard.
823  */
824 void
825 tcp_free_sackholes(struct tcpcb *tp)
826 {
827 	struct sackhole *q;
828 
829 	INP_WLOCK_ASSERT(tptoinpcb(tp));
830 	while ((q = TAILQ_FIRST(&tp->snd_holes)) != NULL)
831 		tcp_sackhole_remove(tp, q);
832 	tp->sackhint.sack_bytes_rexmit = 0;
833 
834 	KASSERT(tp->snd_numholes == 0, ("tp->snd_numholes == 0"));
835 	KASSERT(tp->sackhint.nexthole == NULL,
836 		("tp->sackhint.nexthole == NULL"));
837 }
838 
839 /*
840  * Partial ack handling within a sack recovery episode.  Keeping this very
841  * simple for now.  When a partial ack is received, force snd_cwnd to a value
842  * that will allow the sender to transmit no more than 2 segments.  If
843  * necessary, a better scheme can be adopted at a later point, but for now,
844  * the goal is to prevent the sender from bursting a large amount of data in
845  * the midst of sack recovery.
846  */
847 void
848 tcp_sack_partialack(struct tcpcb *tp, struct tcphdr *th)
849 {
850 	int num_segs = 1;
851 	u_int maxseg = tcp_maxseg(tp);
852 
853 	INP_WLOCK_ASSERT(tptoinpcb(tp));
854 	tcp_timer_activate(tp, TT_REXMT, 0);
855 	tp->t_rtttime = 0;
856 	/* Send one or 2 segments based on how much new data was acked. */
857 	if ((BYTES_THIS_ACK(tp, th) / maxseg) >= 2)
858 		num_segs = 2;
859 	tp->snd_cwnd = (tp->sackhint.sack_bytes_rexmit +
860 	    (tp->snd_nxt - tp->snd_recover) + num_segs * maxseg);
861 	if (tp->snd_cwnd > tp->snd_ssthresh)
862 		tp->snd_cwnd = tp->snd_ssthresh;
863 	tp->t_flags |= TF_ACKNOW;
864 	/*
865 	 * RFC6675 rescue retransmission
866 	 * Add a hole between th_ack (snd_una is not yet set) and snd_max,
867 	 * if this was a pure cumulative ACK and no data was send beyond
868 	 * recovery point. Since the data in the socket has not been freed
869 	 * at this point, we check if the scoreboard is empty, and the ACK
870 	 * delivered some new data, indicating a full ACK. Also, if the
871 	 * recovery point is still at snd_max, we are probably application
872 	 * limited. However, this inference might not always be true. The
873 	 * rescue retransmission may rarely be slightly premature
874 	 * compared to RFC6675.
875 	 * The corresponding ACK+SACK will cause any further outstanding
876 	 * segments to be retransmitted. This addresses a corner case, when
877 	 * the trailing packets of a window are lost and no further data
878 	 * is available for sending.
879 	 */
880 	if ((V_tcp_do_newsack) &&
881 	    SEQ_LT(th->th_ack, tp->snd_recover) &&
882 	    TAILQ_EMPTY(&tp->snd_holes) &&
883 	    (tp->sackhint.delivered_data > 0)) {
884 		/*
885 		 * Exclude FIN sequence space in
886 		 * the hole for the rescue retransmission,
887 		 * and also don't create a hole, if only
888 		 * the ACK for a FIN is outstanding.
889 		 */
890 		tcp_seq highdata = tp->snd_max;
891 		if (tp->t_flags & TF_SENTFIN)
892 			highdata--;
893 		highdata = SEQ_MIN(highdata, tp->snd_recover);
894 		if (th->th_ack != highdata) {
895 			tp->snd_fack = th->th_ack;
896 			(void)tcp_sackhole_insert(tp, SEQ_MAX(th->th_ack,
897 			    highdata - maxseg), highdata, NULL);
898 		}
899 	}
900 	(void) tcp_output(tp);
901 }
902 
903 #if 0
904 /*
905  * Debug version of tcp_sack_output() that walks the scoreboard.  Used for
906  * now to sanity check the hint.
907  */
908 static struct sackhole *
909 tcp_sack_output_debug(struct tcpcb *tp, int *sack_bytes_rexmt)
910 {
911 	struct sackhole *p;
912 
913 	INP_WLOCK_ASSERT(tptoinpcb(tp));
914 	*sack_bytes_rexmt = 0;
915 	TAILQ_FOREACH(p, &tp->snd_holes, scblink) {
916 		if (SEQ_LT(p->rxmit, p->end)) {
917 			if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */
918 				continue;
919 			}
920 			*sack_bytes_rexmt += (p->rxmit - p->start);
921 			break;
922 		}
923 		*sack_bytes_rexmt += (SEQ_MIN(p->rxmit, p->end) - p->start);
924 	}
925 	return (p);
926 }
927 #endif
928 
929 /*
930  * Returns the next hole to retransmit and the number of retransmitted bytes
931  * from the scoreboard.  We store both the next hole and the number of
932  * retransmitted bytes as hints (and recompute these on the fly upon SACK/ACK
933  * reception).  This avoids scoreboard traversals completely.
934  *
935  * The loop here will traverse *at most* one link.  Here's the argument.  For
936  * the loop to traverse more than 1 link before finding the next hole to
937  * retransmit, we would need to have at least 1 node following the current
938  * hint with (rxmit == end).  But, for all holes following the current hint,
939  * (start == rxmit), since we have not yet retransmitted from them.
940  * Therefore, in order to traverse more 1 link in the loop below, we need to
941  * have at least one node following the current hint with (start == rxmit ==
942  * end).  But that can't happen, (start == end) means that all the data in
943  * that hole has been sacked, in which case, the hole would have been removed
944  * from the scoreboard.
945  */
946 struct sackhole *
947 tcp_sack_output(struct tcpcb *tp, int *sack_bytes_rexmt)
948 {
949 	struct sackhole *hole = NULL;
950 
951 	INP_WLOCK_ASSERT(tptoinpcb(tp));
952 	*sack_bytes_rexmt = tp->sackhint.sack_bytes_rexmit;
953 	hole = tp->sackhint.nexthole;
954 	if (hole == NULL)
955 		return (hole);
956 	if (SEQ_GEQ(hole->rxmit, hole->end)) {
957 		for (;;) {
958 			hole = TAILQ_NEXT(hole, scblink);
959 			if (hole == NULL)
960 				return (hole);
961 			if (SEQ_LT(hole->rxmit, hole->end)) {
962 				tp->sackhint.nexthole = hole;
963 				break;
964 			}
965 		}
966 	}
967 	KASSERT(SEQ_LT(hole->start, hole->end), ("%s: hole.start >= hole.end", __func__));
968 	if (!(V_tcp_do_newsack)) {
969 		KASSERT(SEQ_LT(hole->start, tp->snd_fack), ("%s: hole.start >= snd.fack", __func__));
970 		KASSERT(SEQ_LT(hole->end, tp->snd_fack), ("%s: hole.end >= snd.fack", __func__));
971 		KASSERT(SEQ_LT(hole->rxmit, tp->snd_fack), ("%s: hole.rxmit >= snd.fack", __func__));
972 		if (SEQ_GEQ(hole->start, hole->end) ||
973 		    SEQ_GEQ(hole->start, tp->snd_fack) ||
974 		    SEQ_GEQ(hole->end, tp->snd_fack) ||
975 		    SEQ_GEQ(hole->rxmit, tp->snd_fack)) {
976 			log(LOG_CRIT,"tcp: invalid SACK hole (%u-%u,%u) vs fwd ack %u, ignoring.\n",
977 					hole->start, hole->end, hole->rxmit, tp->snd_fack);
978 			return (NULL);
979 		}
980 	}
981 	return (hole);
982 }
983 
984 /*
985  * After a timeout, the SACK list may be rebuilt.  This SACK information
986  * should be used to avoid retransmitting SACKed data.  This function
987  * traverses the SACK list to see if snd_nxt should be moved forward.
988  */
989 void
990 tcp_sack_adjust(struct tcpcb *tp)
991 {
992 	struct sackhole *p, *cur = TAILQ_FIRST(&tp->snd_holes);
993 
994 	INP_WLOCK_ASSERT(tptoinpcb(tp));
995 	if (cur == NULL)
996 		return; /* No holes */
997 	if (SEQ_GEQ(tp->snd_nxt, tp->snd_fack))
998 		return; /* We're already beyond any SACKed blocks */
999 	/*-
1000 	 * Two cases for which we want to advance snd_nxt:
1001 	 * i) snd_nxt lies between end of one hole and beginning of another
1002 	 * ii) snd_nxt lies between end of last hole and snd_fack
1003 	 */
1004 	while ((p = TAILQ_NEXT(cur, scblink)) != NULL) {
1005 		if (SEQ_LT(tp->snd_nxt, cur->end))
1006 			return;
1007 		if (SEQ_GEQ(tp->snd_nxt, p->start))
1008 			cur = p;
1009 		else {
1010 			tp->snd_nxt = p->start;
1011 			return;
1012 		}
1013 	}
1014 	if (SEQ_LT(tp->snd_nxt, cur->end))
1015 		return;
1016 	tp->snd_nxt = tp->snd_fack;
1017 }
1018 
1019 /*
1020  * Lost Retransmission Detection
1021  * Check is FACK is beyond the rexmit of the leftmost hole.
1022  * If yes, we restart sending from still existing holes,
1023  * and adjust cwnd via the congestion control module.
1024  */
1025 void
1026 tcp_sack_lost_retransmission(struct tcpcb *tp, struct tcphdr *th)
1027 {
1028 	struct sackhole *temp;
1029 
1030 	if (IN_RECOVERY(tp->t_flags) &&
1031 	    SEQ_GT(tp->snd_fack, tp->snd_recover) &&
1032 	    ((temp = TAILQ_FIRST(&tp->snd_holes)) != NULL) &&
1033 	    SEQ_GEQ(temp->rxmit, temp->end) &&
1034 	    SEQ_GEQ(tp->snd_fack, temp->rxmit)) {
1035 		TCPSTAT_INC(tcps_sack_lostrexmt);
1036 		/*
1037 		 * Start retransmissions from the first hole, and
1038 		 * subsequently all other remaining holes, including
1039 		 * those, which had been sent completely before.
1040 		 */
1041 		tp->sackhint.nexthole = temp;
1042 		TAILQ_FOREACH(temp, &tp->snd_holes, scblink) {
1043 			if (SEQ_GEQ(tp->snd_fack, temp->rxmit) &&
1044 			    SEQ_GEQ(temp->rxmit, temp->end))
1045 				temp->rxmit = temp->start;
1046 		}
1047 		/*
1048 		 * Remember the old ssthresh, to deduct the beta factor used
1049 		 * by the CC module. Finally, set cwnd to ssthresh just
1050 		 * prior to invoking another cwnd reduction by the CC
1051 		 * module, to not shrink it excessively.
1052 		 */
1053 		tp->snd_cwnd = tp->snd_ssthresh;
1054 		/*
1055 		 * Formally exit recovery, and let the CC module adjust
1056 		 * ssthresh as intended.
1057 		 */
1058 		EXIT_RECOVERY(tp->t_flags);
1059 		cc_cong_signal(tp, th, CC_NDUPACK);
1060 		/*
1061 		 * For PRR, adjust recover_fs as if this new reduction
1062 		 * initialized this variable.
1063 		 * cwnd will be adjusted by SACK or PRR processing
1064 		 * subsequently, only set it to a safe value here.
1065 		 */
1066 		tp->snd_cwnd = tcp_maxseg(tp);
1067 		tp->sackhint.recover_fs = (tp->snd_max - tp->snd_una) -
1068 					    tp->sackhint.recover_fs;
1069 	}
1070 }
1071