xref: /freebsd/sys/netinet/tcp_sack.c (revision 8ddb146abcdf061be9f2c0db7e391697dafad85c)
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 #include <netinet/cc/cc.h>
123 #ifdef TCPDEBUG
124 #include <netinet/tcp_debug.h>
125 #endif /* TCPDEBUG */
126 
127 #include <machine/in_cksum.h>
128 
129 VNET_DECLARE(struct uma_zone *, sack_hole_zone);
130 #define	V_sack_hole_zone		VNET(sack_hole_zone)
131 
132 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, sack, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
133     "TCP SACK");
134 
135 VNET_DEFINE(int, tcp_do_sack) = 1;
136 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, enable, CTLFLAG_VNET | CTLFLAG_RW,
137     &VNET_NAME(tcp_do_sack), 0,
138     "Enable/Disable TCP SACK support");
139 
140 VNET_DEFINE(int, tcp_do_newsack) = 1;
141 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, revised, CTLFLAG_VNET | CTLFLAG_RW,
142     &VNET_NAME(tcp_do_newsack), 0,
143     "Use revised SACK loss recovery per RFC 6675");
144 
145 VNET_DEFINE(int, tcp_sack_maxholes) = 128;
146 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, maxholes, CTLFLAG_VNET | CTLFLAG_RW,
147     &VNET_NAME(tcp_sack_maxholes), 0,
148     "Maximum number of TCP SACK holes allowed per connection");
149 
150 VNET_DEFINE(int, tcp_sack_globalmaxholes) = 65536;
151 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, globalmaxholes, CTLFLAG_VNET | CTLFLAG_RW,
152     &VNET_NAME(tcp_sack_globalmaxholes), 0,
153     "Global maximum number of TCP SACK holes");
154 
155 VNET_DEFINE(int, tcp_sack_globalholes) = 0;
156 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, globalholes, CTLFLAG_VNET | CTLFLAG_RD,
157     &VNET_NAME(tcp_sack_globalholes), 0,
158     "Global number of TCP SACK holes currently allocated");
159 
160 int
161 tcp_dsack_block_exists(struct tcpcb *tp)
162 {
163 	/* Return true if a DSACK block exists */
164 	if (tp->rcv_numsacks == 0)
165 		return (0);
166 	if (SEQ_LEQ(tp->sackblks[0].end, tp->rcv_nxt))
167 		return(1);
168 	return (0);
169 }
170 
171 /*
172  * This function will find overlaps with the currently stored sackblocks
173  * and add any overlap as a dsack block upfront
174  */
175 void
176 tcp_update_dsack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
177 {
178 	struct sackblk head_blk,mid_blk,saved_blks[MAX_SACK_BLKS];
179 	int i, j, n, identical;
180 	tcp_seq start, end;
181 
182 	INP_WLOCK_ASSERT(tp->t_inpcb);
183 
184 	KASSERT(SEQ_LT(rcv_start, rcv_end), ("rcv_start < rcv_end"));
185 
186 	if (SEQ_LT(rcv_end, tp->rcv_nxt) ||
187 	    ((rcv_end == tp->rcv_nxt) &&
188 	     (tp->rcv_numsacks > 0 ) &&
189 	     (tp->sackblks[0].end == tp->rcv_nxt))) {
190 		saved_blks[0].start = rcv_start;
191 		saved_blks[0].end = rcv_end;
192 	} else {
193 		saved_blks[0].start = saved_blks[0].end = 0;
194 	}
195 
196 	head_blk.start = head_blk.end = 0;
197 	mid_blk.start = rcv_start;
198 	mid_blk.end = rcv_end;
199 	identical = 0;
200 
201 	for (i = 0; i < tp->rcv_numsacks; i++) {
202 		start = tp->sackblks[i].start;
203 		end = tp->sackblks[i].end;
204 		if (SEQ_LT(rcv_end, start)) {
205 			/* pkt left to sack blk */
206 			continue;
207 		}
208 		if (SEQ_GT(rcv_start, end)) {
209 			/* pkt right to sack blk */
210 			continue;
211 		}
212 		if (SEQ_GT(tp->rcv_nxt, end)) {
213 			if ((SEQ_MAX(rcv_start, start) != SEQ_MIN(rcv_end, end)) &&
214 			    (SEQ_GT(head_blk.start, SEQ_MAX(rcv_start, start)) ||
215 			    (head_blk.start == head_blk.end))) {
216 				head_blk.start = SEQ_MAX(rcv_start, start);
217 				head_blk.end = SEQ_MIN(rcv_end, end);
218 			}
219 			continue;
220 		}
221 		if (((head_blk.start == head_blk.end) ||
222 		     SEQ_LT(start, head_blk.start)) &&
223 		     (SEQ_GT(end, rcv_start) &&
224 		      SEQ_LEQ(start, rcv_end))) {
225 			head_blk.start = start;
226 			head_blk.end = end;
227 		}
228 		mid_blk.start = SEQ_MIN(mid_blk.start, start);
229 		mid_blk.end = SEQ_MAX(mid_blk.end, end);
230 		if ((mid_blk.start == start) &&
231 		    (mid_blk.end == end))
232 			identical = 1;
233 	}
234 	if (SEQ_LT(head_blk.start, head_blk.end)) {
235 		/* store overlapping range */
236 		saved_blks[0].start = SEQ_MAX(rcv_start, head_blk.start);
237 		saved_blks[0].end   = SEQ_MIN(rcv_end, head_blk.end);
238 	}
239 	n = 1;
240 	/*
241 	 * Second, if not ACKed, store the SACK block that
242 	 * overlaps with the DSACK block unless it is identical
243 	 */
244 	if ((SEQ_LT(tp->rcv_nxt, mid_blk.end) &&
245 	    !((mid_blk.start == saved_blks[0].start) &&
246 	    (mid_blk.end == saved_blks[0].end))) ||
247 	    identical == 1) {
248 		saved_blks[n].start = mid_blk.start;
249 		saved_blks[n++].end = mid_blk.end;
250 	}
251 	for (j = 0; (j < tp->rcv_numsacks) && (n < MAX_SACK_BLKS); j++) {
252 		if (((SEQ_LT(tp->sackblks[j].end, mid_blk.start) ||
253 		      SEQ_GT(tp->sackblks[j].start, mid_blk.end)) &&
254 		    (SEQ_GT(tp->sackblks[j].start, tp->rcv_nxt))))
255 		saved_blks[n++] = tp->sackblks[j];
256 	}
257 	j = 0;
258 	for (i = 0; i < n; i++) {
259 		/* we can end up with a stale initial entry */
260 		if (SEQ_LT(saved_blks[i].start, saved_blks[i].end)) {
261 			tp->sackblks[j++] = saved_blks[i];
262 		}
263 	}
264 	tp->rcv_numsacks = j;
265 }
266 
267 /*
268  * This function is called upon receipt of new valid data (while not in
269  * header prediction mode), and it updates the ordered list of sacks.
270  */
271 void
272 tcp_update_sack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
273 {
274 	/*
275 	 * First reported block MUST be the most recent one.  Subsequent
276 	 * blocks SHOULD be in the order in which they arrived at the
277 	 * receiver.  These two conditions make the implementation fully
278 	 * compliant with RFC 2018.
279 	 */
280 	struct sackblk head_blk, saved_blks[MAX_SACK_BLKS];
281 	int num_head, num_saved, i;
282 
283 	INP_WLOCK_ASSERT(tp->t_inpcb);
284 
285 	/* Check arguments. */
286 	KASSERT(SEQ_LEQ(rcv_start, rcv_end), ("rcv_start <= rcv_end"));
287 
288 	if ((rcv_start == rcv_end) &&
289 	    (tp->rcv_numsacks >= 1) &&
290 	    (rcv_end == tp->sackblks[0].end)) {
291 		/* retaining DSACK block below rcv_nxt (todrop) */
292 		head_blk = tp->sackblks[0];
293 	} else {
294 		/* SACK block for the received segment. */
295 		head_blk.start = rcv_start;
296 		head_blk.end = rcv_end;
297 	}
298 
299 	/*
300 	 * Merge updated SACK blocks into head_blk, and save unchanged SACK
301 	 * blocks into saved_blks[].  num_saved will have the number of the
302 	 * saved SACK blocks.
303 	 */
304 	num_saved = 0;
305 	for (i = 0; i < tp->rcv_numsacks; i++) {
306 		tcp_seq start = tp->sackblks[i].start;
307 		tcp_seq end = tp->sackblks[i].end;
308 		if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
309 			/*
310 			 * Discard this SACK block.
311 			 */
312 		} else if (SEQ_LEQ(head_blk.start, end) &&
313 			   SEQ_GEQ(head_blk.end, start)) {
314 			/*
315 			 * Merge this SACK block into head_blk.  This SACK
316 			 * block itself will be discarded.
317 			 */
318 			/*
319 			 * |-|
320 			 *   |---|  merge
321 			 *
322 			 *     |-|
323 			 * |---|    merge
324 			 *
325 			 * |-----|
326 			 *   |-|    DSACK smaller
327 			 *
328 			 *   |-|
329 			 * |-----|  DSACK smaller
330 			 */
331 			if (head_blk.start == end)
332 				head_blk.start = start;
333 			else if (head_blk.end == start)
334 				head_blk.end = end;
335 			else {
336 				if (SEQ_LT(head_blk.start, start)) {
337 					tcp_seq temp = start;
338 					start = head_blk.start;
339 					head_blk.start = temp;
340 				}
341 				if (SEQ_GT(head_blk.end, end)) {
342 					tcp_seq temp = end;
343 					end = head_blk.end;
344 					head_blk.end = temp;
345 				}
346 				if ((head_blk.start != start) ||
347 				    (head_blk.end != end)) {
348 					if ((num_saved >= 1) &&
349 					   SEQ_GEQ(saved_blks[num_saved-1].start, start) &&
350 					   SEQ_LEQ(saved_blks[num_saved-1].end, end))
351 						num_saved--;
352 					saved_blks[num_saved].start = start;
353 					saved_blks[num_saved].end = end;
354 					num_saved++;
355 				}
356 			}
357 		} else {
358 			/*
359 			 * This block supercedes the prior block
360 			 */
361 			if ((num_saved >= 1) &&
362 			   SEQ_GEQ(saved_blks[num_saved-1].start, start) &&
363 			   SEQ_LEQ(saved_blks[num_saved-1].end, end))
364 				num_saved--;
365 			/*
366 			 * Save this SACK block.
367 			 */
368 			saved_blks[num_saved].start = start;
369 			saved_blks[num_saved].end = end;
370 			num_saved++;
371 		}
372 	}
373 
374 	/*
375 	 * Update SACK list in tp->sackblks[].
376 	 */
377 	num_head = 0;
378 	if (SEQ_LT(rcv_start, rcv_end)) {
379 		/*
380 		 * The received data segment is an out-of-order segment.  Put
381 		 * head_blk at the top of SACK list.
382 		 */
383 		tp->sackblks[0] = head_blk;
384 		num_head = 1;
385 		/*
386 		 * If the number of saved SACK blocks exceeds its limit,
387 		 * discard the last SACK block.
388 		 */
389 		if (num_saved >= MAX_SACK_BLKS)
390 			num_saved--;
391 	}
392 	if ((rcv_start == rcv_end) &&
393 	    (rcv_start == tp->sackblks[0].end)) {
394 		num_head = 1;
395 	}
396 	if (num_saved > 0) {
397 		/*
398 		 * Copy the saved SACK blocks back.
399 		 */
400 		bcopy(saved_blks, &tp->sackblks[num_head],
401 		      sizeof(struct sackblk) * num_saved);
402 	}
403 
404 	/* Save the number of SACK blocks. */
405 	tp->rcv_numsacks = num_head + num_saved;
406 }
407 
408 void
409 tcp_clean_dsack_blocks(struct tcpcb *tp)
410 {
411 	struct sackblk saved_blks[MAX_SACK_BLKS];
412 	int num_saved, i;
413 
414 	INP_WLOCK_ASSERT(tp->t_inpcb);
415 	/*
416 	 * Clean up any DSACK blocks that
417 	 * are in our queue of sack blocks.
418 	 *
419 	 */
420 	num_saved = 0;
421 	for (i = 0; i < tp->rcv_numsacks; i++) {
422 		tcp_seq start = tp->sackblks[i].start;
423 		tcp_seq end = tp->sackblks[i].end;
424 		if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
425 			/*
426 			 * Discard this D-SACK block.
427 			 */
428 			continue;
429 		}
430 		/*
431 		 * Save this SACK block.
432 		 */
433 		saved_blks[num_saved].start = start;
434 		saved_blks[num_saved].end = end;
435 		num_saved++;
436 	}
437 	if (num_saved > 0) {
438 		/*
439 		 * Copy the saved SACK blocks back.
440 		 */
441 		bcopy(saved_blks, &tp->sackblks[0],
442 		      sizeof(struct sackblk) * num_saved);
443 	}
444 	tp->rcv_numsacks = num_saved;
445 }
446 
447 /*
448  * Delete all receiver-side SACK information.
449  */
450 void
451 tcp_clean_sackreport(struct tcpcb *tp)
452 {
453 	int i;
454 
455 	INP_WLOCK_ASSERT(tp->t_inpcb);
456 	tp->rcv_numsacks = 0;
457 	for (i = 0; i < MAX_SACK_BLKS; i++)
458 		tp->sackblks[i].start = tp->sackblks[i].end=0;
459 }
460 
461 /*
462  * Allocate struct sackhole.
463  */
464 static struct sackhole *
465 tcp_sackhole_alloc(struct tcpcb *tp, tcp_seq start, tcp_seq end)
466 {
467 	struct sackhole *hole;
468 
469 	if (tp->snd_numholes >= V_tcp_sack_maxholes ||
470 	    V_tcp_sack_globalholes >= V_tcp_sack_globalmaxholes) {
471 		TCPSTAT_INC(tcps_sack_sboverflow);
472 		return NULL;
473 	}
474 
475 	hole = (struct sackhole *)uma_zalloc(V_sack_hole_zone, M_NOWAIT);
476 	if (hole == NULL)
477 		return NULL;
478 
479 	hole->start = start;
480 	hole->end = end;
481 	hole->rxmit = start;
482 
483 	tp->snd_numholes++;
484 	atomic_add_int(&V_tcp_sack_globalholes, 1);
485 
486 	return hole;
487 }
488 
489 /*
490  * Free struct sackhole.
491  */
492 static void
493 tcp_sackhole_free(struct tcpcb *tp, struct sackhole *hole)
494 {
495 
496 	uma_zfree(V_sack_hole_zone, hole);
497 
498 	tp->snd_numholes--;
499 	atomic_subtract_int(&V_tcp_sack_globalholes, 1);
500 
501 	KASSERT(tp->snd_numholes >= 0, ("tp->snd_numholes >= 0"));
502 	KASSERT(V_tcp_sack_globalholes >= 0, ("tcp_sack_globalholes >= 0"));
503 }
504 
505 /*
506  * Insert new SACK hole into scoreboard.
507  */
508 static struct sackhole *
509 tcp_sackhole_insert(struct tcpcb *tp, tcp_seq start, tcp_seq end,
510     struct sackhole *after)
511 {
512 	struct sackhole *hole;
513 
514 	/* Allocate a new SACK hole. */
515 	hole = tcp_sackhole_alloc(tp, start, end);
516 	if (hole == NULL)
517 		return NULL;
518 
519 	/* Insert the new SACK hole into scoreboard. */
520 	if (after != NULL)
521 		TAILQ_INSERT_AFTER(&tp->snd_holes, after, hole, scblink);
522 	else
523 		TAILQ_INSERT_TAIL(&tp->snd_holes, hole, scblink);
524 
525 	/* Update SACK hint. */
526 	if (tp->sackhint.nexthole == NULL)
527 		tp->sackhint.nexthole = hole;
528 
529 	return hole;
530 }
531 
532 /*
533  * Remove SACK hole from scoreboard.
534  */
535 static void
536 tcp_sackhole_remove(struct tcpcb *tp, struct sackhole *hole)
537 {
538 
539 	/* Update SACK hint. */
540 	if (tp->sackhint.nexthole == hole)
541 		tp->sackhint.nexthole = TAILQ_NEXT(hole, scblink);
542 
543 	/* Remove this SACK hole. */
544 	TAILQ_REMOVE(&tp->snd_holes, hole, scblink);
545 
546 	/* Free this SACK hole. */
547 	tcp_sackhole_free(tp, hole);
548 }
549 
550 /*
551  * Process cumulative ACK and the TCP SACK option to update the scoreboard.
552  * tp->snd_holes is an ordered list of holes (oldest to newest, in terms of
553  * the sequence space).
554  * Returns 1 if incoming ACK has previously unknown SACK information,
555  * 0 otherwise.
556  */
557 int
558 tcp_sack_doack(struct tcpcb *tp, struct tcpopt *to, tcp_seq th_ack)
559 {
560 	struct sackhole *cur, *temp;
561 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1], *sblkp;
562 	int i, j, num_sack_blks, sack_changed;
563 	int delivered_data, left_edge_delta;
564 
565 	INP_WLOCK_ASSERT(tp->t_inpcb);
566 
567 	num_sack_blks = 0;
568 	sack_changed = 0;
569 	delivered_data = 0;
570 	left_edge_delta = 0;
571 	/*
572 	 * If SND.UNA will be advanced by SEG.ACK, and if SACK holes exist,
573 	 * treat [SND.UNA, SEG.ACK) as if it is a SACK block.
574 	 * Account changes to SND.UNA always in delivered data.
575 	 */
576 	if (SEQ_LT(tp->snd_una, th_ack) && !TAILQ_EMPTY(&tp->snd_holes)) {
577 		left_edge_delta = th_ack - tp->snd_una;
578 		sack_blocks[num_sack_blks].start = tp->snd_una;
579 		sack_blocks[num_sack_blks++].end = th_ack;
580 		/*
581 		 * Pulling snd_fack forward if we got here
582 		 * due to DSACK blocks
583 		 */
584 		if (SEQ_LT(tp->snd_fack, th_ack)) {
585 			delivered_data += th_ack - tp->snd_una;
586 			tp->snd_fack = th_ack;
587 			sack_changed = 1;
588 		}
589 	}
590 	/*
591 	 * Append received valid SACK blocks to sack_blocks[], but only if we
592 	 * received new blocks from the other side.
593 	 */
594 	if (to->to_flags & TOF_SACK) {
595 		for (i = 0; i < to->to_nsacks; i++) {
596 			bcopy((to->to_sacks + i * TCPOLEN_SACK),
597 			    &sack, sizeof(sack));
598 			sack.start = ntohl(sack.start);
599 			sack.end = ntohl(sack.end);
600 			if (SEQ_GT(sack.end, sack.start) &&
601 			    SEQ_GT(sack.start, tp->snd_una) &&
602 			    SEQ_GT(sack.start, th_ack) &&
603 			    SEQ_LT(sack.start, tp->snd_max) &&
604 			    SEQ_GT(sack.end, tp->snd_una) &&
605 			    SEQ_LEQ(sack.end, tp->snd_max)) {
606 				sack_blocks[num_sack_blks++] = sack;
607 			} else if (SEQ_LEQ(sack.start, th_ack) &&
608 			    SEQ_LEQ(sack.end, th_ack)) {
609 				/*
610 				 * Its a D-SACK block.
611 				 */
612 				tcp_record_dsack(tp, sack.start, sack.end, 0);
613 			}
614 		}
615 	}
616 	/*
617 	 * Return if SND.UNA is not advanced and no valid SACK block is
618 	 * received.
619 	 */
620 	if (num_sack_blks == 0)
621 		return (sack_changed);
622 
623 	/*
624 	 * Sort the SACK blocks so we can update the scoreboard with just one
625 	 * pass. The overhead of sorting up to 4+1 elements is less than
626 	 * making up to 4+1 passes over the scoreboard.
627 	 */
628 	for (i = 0; i < num_sack_blks; i++) {
629 		for (j = i + 1; j < num_sack_blks; j++) {
630 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
631 				sack = sack_blocks[i];
632 				sack_blocks[i] = sack_blocks[j];
633 				sack_blocks[j] = sack;
634 			}
635 		}
636 	}
637 	if (TAILQ_EMPTY(&tp->snd_holes)) {
638 		/*
639 		 * Empty scoreboard. Need to initialize snd_fack (it may be
640 		 * uninitialized or have a bogus value). Scoreboard holes
641 		 * (from the sack blocks received) are created later below
642 		 * (in the logic that adds holes to the tail of the
643 		 * scoreboard).
644 		 */
645 		tp->snd_fack = SEQ_MAX(tp->snd_una, th_ack);
646 		tp->sackhint.sacked_bytes = 0;	/* reset */
647 	}
648 	/*
649 	 * In the while-loop below, incoming SACK blocks (sack_blocks[]) and
650 	 * SACK holes (snd_holes) are traversed from their tails with just
651 	 * one pass in order to reduce the number of compares especially when
652 	 * the bandwidth-delay product is large.
653 	 *
654 	 * Note: Typically, in the first RTT of SACK recovery, the highest
655 	 * three or four SACK blocks with the same ack number are received.
656 	 * In the second RTT, if retransmitted data segments are not lost,
657 	 * the highest three or four SACK blocks with ack number advancing
658 	 * are received.
659 	 */
660 	sblkp = &sack_blocks[num_sack_blks - 1];	/* Last SACK block */
661 	tp->sackhint.last_sack_ack = sblkp->end;
662 	if (SEQ_LT(tp->snd_fack, sblkp->start)) {
663 		/*
664 		 * The highest SACK block is beyond fack.  First,
665 		 * check if there was a successful Rescue Retransmission,
666 		 * and move this hole left. With normal holes, snd_fack
667 		 * is always to the right of the end.
668 		 */
669 		if (((temp = TAILQ_LAST(&tp->snd_holes, sackhole_head)) != NULL) &&
670 		    SEQ_LEQ(tp->snd_fack,temp->end)) {
671 			temp->start = SEQ_MAX(tp->snd_fack, SEQ_MAX(tp->snd_una, th_ack));
672 			temp->end = sblkp->start;
673 			temp->rxmit = temp->start;
674 			delivered_data += sblkp->end - sblkp->start;
675 			tp->snd_fack = sblkp->end;
676 			sblkp--;
677 			sack_changed = 1;
678 		} else {
679 			/*
680 			 * Append a new SACK hole at the tail.  If the
681 			 * second or later highest SACK blocks are also
682 			 * beyond the current fack, they will be inserted
683 			 * by way of hole splitting in the while-loop below.
684 			 */
685 			temp = tcp_sackhole_insert(tp, tp->snd_fack,sblkp->start,NULL);
686 			if (temp != NULL) {
687 				delivered_data += sblkp->end - sblkp->start;
688 				tp->snd_fack = sblkp->end;
689 				/* Go to the previous sack block. */
690 				sblkp--;
691 				sack_changed = 1;
692 			} else {
693 				/*
694 				 * We failed to add a new hole based on the current
695 				 * sack block.  Skip over all the sack blocks that
696 				 * fall completely to the right of snd_fack and
697 				 * proceed to trim the scoreboard based on the
698 				 * remaining sack blocks.  This also trims the
699 				 * scoreboard for th_ack (which is sack_blocks[0]).
700 				 */
701 				while (sblkp >= sack_blocks &&
702 				       SEQ_LT(tp->snd_fack, sblkp->start))
703 					sblkp--;
704 				if (sblkp >= sack_blocks &&
705 				    SEQ_LT(tp->snd_fack, sblkp->end)) {
706 					delivered_data += sblkp->end - tp->snd_fack;
707 					tp->snd_fack = sblkp->end;
708 					sack_changed = 1;
709 				}
710 			}
711 		}
712 	} else if (SEQ_LT(tp->snd_fack, sblkp->end)) {
713 		/* fack is advanced. */
714 		delivered_data += sblkp->end - tp->snd_fack;
715 		tp->snd_fack = sblkp->end;
716 		sack_changed = 1;
717 	}
718 	cur = TAILQ_LAST(&tp->snd_holes, sackhole_head); /* Last SACK hole. */
719 	/*
720 	 * Since the incoming sack blocks are sorted, we can process them
721 	 * making one sweep of the scoreboard.
722 	 */
723 	while (sblkp >= sack_blocks  && cur != NULL) {
724 		if (SEQ_GEQ(sblkp->start, cur->end)) {
725 			/*
726 			 * SACKs data beyond the current hole.  Go to the
727 			 * previous sack block.
728 			 */
729 			sblkp--;
730 			continue;
731 		}
732 		if (SEQ_LEQ(sblkp->end, cur->start)) {
733 			/*
734 			 * SACKs data before the current hole.  Go to the
735 			 * previous hole.
736 			 */
737 			cur = TAILQ_PREV(cur, sackhole_head, scblink);
738 			continue;
739 		}
740 		tp->sackhint.sack_bytes_rexmit -=
741 		    (SEQ_MIN(cur->rxmit, cur->end) - cur->start);
742 		KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
743 		    ("sackhint bytes rtx >= 0"));
744 		sack_changed = 1;
745 		if (SEQ_LEQ(sblkp->start, cur->start)) {
746 			/* Data acks at least the beginning of hole. */
747 			if (SEQ_GEQ(sblkp->end, cur->end)) {
748 				/* Acks entire hole, so delete hole. */
749 				delivered_data += (cur->end - cur->start);
750 				temp = cur;
751 				cur = TAILQ_PREV(cur, sackhole_head, scblink);
752 				tcp_sackhole_remove(tp, temp);
753 				/*
754 				 * The sack block may ack all or part of the
755 				 * next hole too, so continue onto the next
756 				 * hole.
757 				 */
758 				continue;
759 			} else {
760 				/* Move start of hole forward. */
761 				delivered_data += (sblkp->end - cur->start);
762 				cur->start = sblkp->end;
763 				cur->rxmit = SEQ_MAX(cur->rxmit, cur->start);
764 			}
765 		} else {
766 			/* Data acks at least the end of hole. */
767 			if (SEQ_GEQ(sblkp->end, cur->end)) {
768 				/* Move end of hole backward. */
769 				delivered_data += (cur->end - sblkp->start);
770 				cur->end = sblkp->start;
771 				cur->rxmit = SEQ_MIN(cur->rxmit, cur->end);
772 				if ((tp->t_flags & TF_LRD) && SEQ_GEQ(cur->rxmit, cur->end))
773 					cur->rxmit = tp->snd_recover;
774 			} else {
775 				/*
776 				 * ACKs some data in middle of a hole; need
777 				 * to split current hole
778 				 */
779 				temp = tcp_sackhole_insert(tp, sblkp->end,
780 				    cur->end, cur);
781 				if (temp != NULL) {
782 					if (SEQ_GT(cur->rxmit, temp->rxmit)) {
783 						temp->rxmit = cur->rxmit;
784 						tp->sackhint.sack_bytes_rexmit +=
785 						    (SEQ_MIN(temp->rxmit,
786 						    temp->end) - temp->start);
787 					}
788 					cur->end = sblkp->start;
789 					cur->rxmit = SEQ_MIN(cur->rxmit,
790 					    cur->end);
791 					if ((tp->t_flags & TF_LRD) && SEQ_GEQ(cur->rxmit, cur->end))
792 						cur->rxmit = tp->snd_recover;
793 					delivered_data += (sblkp->end - sblkp->start);
794 				}
795 			}
796 		}
797 		tp->sackhint.sack_bytes_rexmit +=
798 		    (SEQ_MIN(cur->rxmit, cur->end) - cur->start);
799 		/*
800 		 * Testing sblkp->start against cur->start tells us whether
801 		 * we're done with the sack block or the sack hole.
802 		 * Accordingly, we advance one or the other.
803 		 */
804 		if (SEQ_LEQ(sblkp->start, cur->start))
805 			cur = TAILQ_PREV(cur, sackhole_head, scblink);
806 		else
807 			sblkp--;
808 	}
809 	if (!(to->to_flags & TOF_SACK))
810 		/*
811 		 * If this ACK did not contain any
812 		 * SACK blocks, any only moved the
813 		 * left edge right, it is a pure
814 		 * cumulative ACK. Do not count
815 		 * DupAck for this. Also required
816 		 * for RFC6675 rescue retransmission.
817 		 */
818 		sack_changed = 0;
819 	tp->sackhint.delivered_data = delivered_data;
820 	tp->sackhint.sacked_bytes += delivered_data - left_edge_delta;
821 	KASSERT((delivered_data >= 0), ("delivered_data < 0"));
822 	KASSERT((tp->sackhint.sacked_bytes >= 0), ("sacked_bytes < 0"));
823 	return (sack_changed);
824 }
825 
826 /*
827  * Free all SACK holes to clear the scoreboard.
828  */
829 void
830 tcp_free_sackholes(struct tcpcb *tp)
831 {
832 	struct sackhole *q;
833 
834 	INP_WLOCK_ASSERT(tp->t_inpcb);
835 	while ((q = TAILQ_FIRST(&tp->snd_holes)) != NULL)
836 		tcp_sackhole_remove(tp, q);
837 	tp->sackhint.sack_bytes_rexmit = 0;
838 
839 	KASSERT(tp->snd_numholes == 0, ("tp->snd_numholes == 0"));
840 	KASSERT(tp->sackhint.nexthole == NULL,
841 		("tp->sackhint.nexthole == NULL"));
842 }
843 
844 /*
845  * Partial ack handling within a sack recovery episode.  Keeping this very
846  * simple for now.  When a partial ack is received, force snd_cwnd to a value
847  * that will allow the sender to transmit no more than 2 segments.  If
848  * necessary, a better scheme can be adopted at a later point, but for now,
849  * the goal is to prevent the sender from bursting a large amount of data in
850  * the midst of sack recovery.
851  */
852 void
853 tcp_sack_partialack(struct tcpcb *tp, struct tcphdr *th)
854 {
855 	int num_segs = 1;
856 	u_int maxseg = tcp_maxseg(tp);
857 
858 	INP_WLOCK_ASSERT(tp->t_inpcb);
859 	tcp_timer_activate(tp, TT_REXMT, 0);
860 	tp->t_rtttime = 0;
861 	/* Send one or 2 segments based on how much new data was acked. */
862 	if ((BYTES_THIS_ACK(tp, th) / maxseg) >= 2)
863 		num_segs = 2;
864 	tp->snd_cwnd = (tp->sackhint.sack_bytes_rexmit +
865 	    (tp->snd_nxt - tp->snd_recover) + num_segs * maxseg);
866 	if (tp->snd_cwnd > tp->snd_ssthresh)
867 		tp->snd_cwnd = tp->snd_ssthresh;
868 	tp->t_flags |= TF_ACKNOW;
869 	/*
870 	 * RFC6675 rescue retransmission
871 	 * Add a hole between th_ack (snd_una is not yet set) and snd_max,
872 	 * if this was a pure cumulative ACK and no data was send beyond
873 	 * recovery point. Since the data in the socket has not been freed
874 	 * at this point, we check if the scoreboard is empty, and the ACK
875 	 * delivered some new data, indicating a full ACK. Also, if the
876 	 * recovery point is still at snd_max, we are probably application
877 	 * limited. However, this inference might not always be true. The
878 	 * rescue retransmission may rarely be slightly premature
879 	 * compared to RFC6675.
880 	 * The corresponding ACK+SACK will cause any further outstanding
881 	 * segments to be retransmitted. This addresses a corner case, when
882 	 * the trailing packets of a window are lost and no further data
883 	 * is available for sending.
884 	 */
885 	if ((V_tcp_do_newsack) &&
886 	    SEQ_LT(th->th_ack, tp->snd_recover) &&
887 	    (tp->snd_recover == tp->snd_max) &&
888 	    TAILQ_EMPTY(&tp->snd_holes) &&
889 	    (tp->sackhint.delivered_data > 0)) {
890 		/*
891 		 * Exclude FIN sequence space in
892 		 * the hole for the rescue retransmission,
893 		 * and also don't create a hole, if only
894 		 * the ACK for a FIN is outstanding.
895 		 */
896 		tcp_seq highdata = tp->snd_max;
897 		if (tp->t_flags & TF_SENTFIN)
898 			highdata--;
899 		if (th->th_ack != highdata) {
900 			tp->snd_fack = th->th_ack;
901 			(void)tcp_sackhole_insert(tp, SEQ_MAX(th->th_ack,
902 			    highdata - maxseg), highdata, NULL);
903 		}
904 	}
905 	(void) tcp_output(tp);
906 }
907 
908 #if 0
909 /*
910  * Debug version of tcp_sack_output() that walks the scoreboard.  Used for
911  * now to sanity check the hint.
912  */
913 static struct sackhole *
914 tcp_sack_output_debug(struct tcpcb *tp, int *sack_bytes_rexmt)
915 {
916 	struct sackhole *p;
917 
918 	INP_WLOCK_ASSERT(tp->t_inpcb);
919 	*sack_bytes_rexmt = 0;
920 	TAILQ_FOREACH(p, &tp->snd_holes, scblink) {
921 		if (SEQ_LT(p->rxmit, p->end)) {
922 			if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */
923 				continue;
924 			}
925 			*sack_bytes_rexmt += (p->rxmit - p->start);
926 			break;
927 		}
928 		*sack_bytes_rexmt += (SEQ_MIN(p->rxmit, p->end) - p->start);
929 	}
930 	return (p);
931 }
932 #endif
933 
934 /*
935  * Returns the next hole to retransmit and the number of retransmitted bytes
936  * from the scoreboard.  We store both the next hole and the number of
937  * retransmitted bytes as hints (and recompute these on the fly upon SACK/ACK
938  * reception).  This avoids scoreboard traversals completely.
939  *
940  * The loop here will traverse *at most* one link.  Here's the argument.  For
941  * the loop to traverse more than 1 link before finding the next hole to
942  * retransmit, we would need to have at least 1 node following the current
943  * hint with (rxmit == end).  But, for all holes following the current hint,
944  * (start == rxmit), since we have not yet retransmitted from them.
945  * Therefore, in order to traverse more 1 link in the loop below, we need to
946  * have at least one node following the current hint with (start == rxmit ==
947  * end).  But that can't happen, (start == end) means that all the data in
948  * that hole has been sacked, in which case, the hole would have been removed
949  * from the scoreboard.
950  */
951 struct sackhole *
952 tcp_sack_output(struct tcpcb *tp, int *sack_bytes_rexmt)
953 {
954 	struct sackhole *hole = NULL;
955 
956 	INP_WLOCK_ASSERT(tp->t_inpcb);
957 	*sack_bytes_rexmt = tp->sackhint.sack_bytes_rexmit;
958 	hole = tp->sackhint.nexthole;
959 	if (hole == NULL)
960 		return (hole);
961 	if (SEQ_GEQ(hole->rxmit, hole->end)) {
962 		for (;;) {
963 			hole = TAILQ_NEXT(hole, scblink);
964 			if (hole == NULL)
965 				return (hole);
966 			if (SEQ_LT(hole->rxmit, hole->end)) {
967 				tp->sackhint.nexthole = hole;
968 				break;
969 			}
970 		}
971 	}
972 	KASSERT(SEQ_LT(hole->start, hole->end), ("%s: hole.start >= hole.end", __func__));
973 	if (!(V_tcp_do_newsack)) {
974 		KASSERT(SEQ_LT(hole->start, tp->snd_fack), ("%s: hole.start >= snd.fack", __func__));
975 		KASSERT(SEQ_LT(hole->end, tp->snd_fack), ("%s: hole.end >= snd.fack", __func__));
976 		KASSERT(SEQ_LT(hole->rxmit, tp->snd_fack), ("%s: hole.rxmit >= snd.fack", __func__));
977 		if (SEQ_GEQ(hole->start, hole->end) ||
978 		    SEQ_GEQ(hole->start, tp->snd_fack) ||
979 		    SEQ_GEQ(hole->end, tp->snd_fack) ||
980 		    SEQ_GEQ(hole->rxmit, tp->snd_fack)) {
981 			log(LOG_CRIT,"tcp: invalid SACK hole (%u-%u,%u) vs fwd ack %u, ignoring.\n",
982 					hole->start, hole->end, hole->rxmit, tp->snd_fack);
983 			return (NULL);
984 		}
985 	}
986 	return (hole);
987 }
988 
989 /*
990  * After a timeout, the SACK list may be rebuilt.  This SACK information
991  * should be used to avoid retransmitting SACKed data.  This function
992  * traverses the SACK list to see if snd_nxt should be moved forward.
993  */
994 void
995 tcp_sack_adjust(struct tcpcb *tp)
996 {
997 	struct sackhole *p, *cur = TAILQ_FIRST(&tp->snd_holes);
998 
999 	INP_WLOCK_ASSERT(tp->t_inpcb);
1000 	if (cur == NULL)
1001 		return; /* No holes */
1002 	if (SEQ_GEQ(tp->snd_nxt, tp->snd_fack))
1003 		return; /* We're already beyond any SACKed blocks */
1004 	/*-
1005 	 * Two cases for which we want to advance snd_nxt:
1006 	 * i) snd_nxt lies between end of one hole and beginning of another
1007 	 * ii) snd_nxt lies between end of last hole and snd_fack
1008 	 */
1009 	while ((p = TAILQ_NEXT(cur, scblink)) != NULL) {
1010 		if (SEQ_LT(tp->snd_nxt, cur->end))
1011 			return;
1012 		if (SEQ_GEQ(tp->snd_nxt, p->start))
1013 			cur = p;
1014 		else {
1015 			tp->snd_nxt = p->start;
1016 			return;
1017 		}
1018 	}
1019 	if (SEQ_LT(tp->snd_nxt, cur->end))
1020 		return;
1021 	tp->snd_nxt = tp->snd_fack;
1022 }
1023 
1024 /*
1025  * Lost Retransmission Detection
1026  * Check is FACK is beyond the rexmit of the leftmost hole.
1027  * If yes, we restart sending from still existing holes,
1028  * and adjust cwnd via the congestion control module.
1029  */
1030 void
1031 tcp_sack_lost_retransmission(struct tcpcb *tp, struct tcphdr *th)
1032 {
1033 	struct sackhole *temp;
1034 
1035 	if (IN_RECOVERY(tp->t_flags) &&
1036 	    SEQ_GT(tp->snd_fack, tp->snd_recover) &&
1037 	    ((temp = TAILQ_FIRST(&tp->snd_holes)) != NULL) &&
1038 	    SEQ_GEQ(temp->rxmit, temp->end) &&
1039 	    SEQ_GEQ(tp->snd_fack, temp->rxmit)) {
1040 		TCPSTAT_INC(tcps_sack_lostrexmt);
1041 		/*
1042 		 * Start retransmissions from the first hole, and
1043 		 * subsequently all other remaining holes, including
1044 		 * those, which had been sent completely before.
1045 		 */
1046 		tp->sackhint.nexthole = temp;
1047 		TAILQ_FOREACH(temp, &tp->snd_holes, scblink) {
1048 			if (SEQ_GEQ(tp->snd_fack, temp->rxmit) &&
1049 			    SEQ_GEQ(temp->rxmit, temp->end))
1050 				temp->rxmit = temp->start;
1051 		}
1052 		/*
1053 		 * Remember the old ssthresh, to deduct the beta factor used
1054 		 * by the CC module. Finally, set cwnd to ssthresh just
1055 		 * prior to invoking another cwnd reduction by the CC
1056 		 * module, to not shrink it excessively.
1057 		 */
1058 		tp->snd_cwnd = tp->snd_ssthresh;
1059 		/*
1060 		 * Formally exit recovery, and let the CC module adjust
1061 		 * ssthresh as intended.
1062 		 */
1063 		EXIT_RECOVERY(tp->t_flags);
1064 		cc_cong_signal(tp, th, CC_NDUPACK);
1065 		/*
1066 		 * For PRR, adjust recover_fs as if this new reduction
1067 		 * initialized this variable.
1068 		 * cwnd will be adjusted by SACK or PRR processing
1069 		 * subsequently, only set it to a safe value here.
1070 		 */
1071 		tp->snd_cwnd = tcp_maxseg(tp);
1072 		tp->sackhint.recover_fs = (tp->snd_max - tp->snd_una) -
1073 					    tp->sackhint.recover_fs;
1074 	}
1075 }
1076