xref: /freebsd/sys/netinet/tcp_sack.c (revision 734e82fe33aa764367791a7d603b383996c6b40b)
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 #include "opt_inet.h"
77 #include "opt_inet6.h"
78 
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
82 #include <sys/sysctl.h>
83 #include <sys/malloc.h>
84 #include <sys/mbuf.h>
85 #include <sys/proc.h>		/* for proc0 declaration */
86 #include <sys/protosw.h>
87 #include <sys/socket.h>
88 #include <sys/socketvar.h>
89 #include <sys/syslog.h>
90 #include <sys/systm.h>
91 
92 #include <machine/cpu.h>	/* before tcp_seq.h, for tcp_random18() */
93 
94 #include <vm/uma.h>
95 
96 #include <net/if.h>
97 #include <net/if_var.h>
98 #include <net/route.h>
99 #include <net/vnet.h>
100 
101 #include <netinet/in.h>
102 #include <netinet/in_systm.h>
103 #include <netinet/ip.h>
104 #include <netinet/in_var.h>
105 #include <netinet/in_pcb.h>
106 #include <netinet/ip_var.h>
107 #include <netinet/ip6.h>
108 #include <netinet/icmp6.h>
109 #include <netinet6/nd6.h>
110 #include <netinet6/ip6_var.h>
111 #include <netinet6/in6_pcb.h>
112 #include <netinet/tcp.h>
113 #include <netinet/tcp_fsm.h>
114 #include <netinet/tcp_seq.h>
115 #include <netinet/tcp_timer.h>
116 #include <netinet/tcp_var.h>
117 #include <netinet/tcpip.h>
118 #include <netinet/cc/cc.h>
119 
120 #include <machine/in_cksum.h>
121 
122 VNET_DECLARE(struct uma_zone *, sack_hole_zone);
123 #define	V_sack_hole_zone		VNET(sack_hole_zone)
124 
125 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, sack, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
126     "TCP SACK");
127 
128 VNET_DEFINE(int, tcp_do_sack) = 1;
129 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, enable, CTLFLAG_VNET | CTLFLAG_RW,
130     &VNET_NAME(tcp_do_sack), 0,
131     "Enable/Disable TCP SACK support");
132 
133 VNET_DEFINE(int, tcp_do_newsack) = 1;
134 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, revised, CTLFLAG_VNET | CTLFLAG_RW,
135     &VNET_NAME(tcp_do_newsack), 0,
136     "Use revised SACK loss recovery per RFC 6675");
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 int
154 tcp_dsack_block_exists(struct tcpcb *tp)
155 {
156 	/* Return true if a DSACK block exists */
157 	if (tp->rcv_numsacks == 0)
158 		return (0);
159 	if (SEQ_LEQ(tp->sackblks[0].end, tp->rcv_nxt))
160 		return(1);
161 	return (0);
162 }
163 
164 /*
165  * This function will find overlaps with the currently stored sackblocks
166  * and add any overlap as a dsack block upfront
167  */
168 void
169 tcp_update_dsack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
170 {
171 	struct sackblk head_blk,mid_blk,saved_blks[MAX_SACK_BLKS];
172 	int i, j, n, identical;
173 	tcp_seq start, end;
174 
175 	INP_WLOCK_ASSERT(tptoinpcb(tp));
176 
177 	KASSERT(SEQ_LT(rcv_start, rcv_end), ("rcv_start < rcv_end"));
178 
179 	if (SEQ_LT(rcv_end, tp->rcv_nxt) ||
180 	    ((rcv_end == tp->rcv_nxt) &&
181 	     (tp->rcv_numsacks > 0 ) &&
182 	     (tp->sackblks[0].end == tp->rcv_nxt))) {
183 		saved_blks[0].start = rcv_start;
184 		saved_blks[0].end = rcv_end;
185 	} else {
186 		saved_blks[0].start = saved_blks[0].end = 0;
187 	}
188 
189 	head_blk.start = head_blk.end = 0;
190 	mid_blk.start = rcv_start;
191 	mid_blk.end = rcv_end;
192 	identical = 0;
193 
194 	for (i = 0; i < tp->rcv_numsacks; i++) {
195 		start = tp->sackblks[i].start;
196 		end = tp->sackblks[i].end;
197 		if (SEQ_LT(rcv_end, start)) {
198 			/* pkt left to sack blk */
199 			continue;
200 		}
201 		if (SEQ_GT(rcv_start, end)) {
202 			/* pkt right to sack blk */
203 			continue;
204 		}
205 		if (SEQ_GT(tp->rcv_nxt, end)) {
206 			if ((SEQ_MAX(rcv_start, start) != SEQ_MIN(rcv_end, end)) &&
207 			    (SEQ_GT(head_blk.start, SEQ_MAX(rcv_start, start)) ||
208 			    (head_blk.start == head_blk.end))) {
209 				head_blk.start = SEQ_MAX(rcv_start, start);
210 				head_blk.end = SEQ_MIN(rcv_end, end);
211 			}
212 			continue;
213 		}
214 		if (((head_blk.start == head_blk.end) ||
215 		     SEQ_LT(start, head_blk.start)) &&
216 		     (SEQ_GT(end, rcv_start) &&
217 		      SEQ_LEQ(start, rcv_end))) {
218 			head_blk.start = start;
219 			head_blk.end = end;
220 		}
221 		mid_blk.start = SEQ_MIN(mid_blk.start, start);
222 		mid_blk.end = SEQ_MAX(mid_blk.end, end);
223 		if ((mid_blk.start == start) &&
224 		    (mid_blk.end == end))
225 			identical = 1;
226 	}
227 	if (SEQ_LT(head_blk.start, head_blk.end)) {
228 		/* store overlapping range */
229 		saved_blks[0].start = SEQ_MAX(rcv_start, head_blk.start);
230 		saved_blks[0].end   = SEQ_MIN(rcv_end, head_blk.end);
231 	}
232 	n = 1;
233 	/*
234 	 * Second, if not ACKed, store the SACK block that
235 	 * overlaps with the DSACK block unless it is identical
236 	 */
237 	if ((SEQ_LT(tp->rcv_nxt, mid_blk.end) &&
238 	    !((mid_blk.start == saved_blks[0].start) &&
239 	    (mid_blk.end == saved_blks[0].end))) ||
240 	    identical == 1) {
241 		saved_blks[n].start = mid_blk.start;
242 		saved_blks[n++].end = mid_blk.end;
243 	}
244 	for (j = 0; (j < tp->rcv_numsacks) && (n < MAX_SACK_BLKS); j++) {
245 		if (((SEQ_LT(tp->sackblks[j].end, mid_blk.start) ||
246 		      SEQ_GT(tp->sackblks[j].start, mid_blk.end)) &&
247 		    (SEQ_GT(tp->sackblks[j].start, tp->rcv_nxt))))
248 		saved_blks[n++] = tp->sackblks[j];
249 	}
250 	j = 0;
251 	for (i = 0; i < n; i++) {
252 		/* we can end up with a stale initial entry */
253 		if (SEQ_LT(saved_blks[i].start, saved_blks[i].end)) {
254 			tp->sackblks[j++] = saved_blks[i];
255 		}
256 	}
257 	tp->rcv_numsacks = j;
258 }
259 
260 /*
261  * This function is called upon receipt of new valid data (while not in
262  * header prediction mode), and it updates the ordered list of sacks.
263  */
264 void
265 tcp_update_sack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
266 {
267 	/*
268 	 * First reported block MUST be the most recent one.  Subsequent
269 	 * blocks SHOULD be in the order in which they arrived at the
270 	 * receiver.  These two conditions make the implementation fully
271 	 * compliant with RFC 2018.
272 	 */
273 	struct sackblk head_blk, saved_blks[MAX_SACK_BLKS];
274 	int num_head, num_saved, i;
275 
276 	INP_WLOCK_ASSERT(tptoinpcb(tp));
277 
278 	/* Check arguments. */
279 	KASSERT(SEQ_LEQ(rcv_start, rcv_end), ("rcv_start <= rcv_end"));
280 
281 	if ((rcv_start == rcv_end) &&
282 	    (tp->rcv_numsacks >= 1) &&
283 	    (rcv_end == tp->sackblks[0].end)) {
284 		/* retaining DSACK block below rcv_nxt (todrop) */
285 		head_blk = tp->sackblks[0];
286 	} else {
287 		/* SACK block for the received segment. */
288 		head_blk.start = rcv_start;
289 		head_blk.end = rcv_end;
290 	}
291 
292 	/*
293 	 * Merge updated SACK blocks into head_blk, and save unchanged SACK
294 	 * blocks into saved_blks[].  num_saved will have the number of the
295 	 * saved SACK blocks.
296 	 */
297 	num_saved = 0;
298 	for (i = 0; i < tp->rcv_numsacks; i++) {
299 		tcp_seq start = tp->sackblks[i].start;
300 		tcp_seq end = tp->sackblks[i].end;
301 		if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
302 			/*
303 			 * Discard this SACK block.
304 			 */
305 		} else if (SEQ_LEQ(head_blk.start, end) &&
306 			   SEQ_GEQ(head_blk.end, start)) {
307 			/*
308 			 * Merge this SACK block into head_blk.  This SACK
309 			 * block itself will be discarded.
310 			 */
311 			/*
312 			 * |-|
313 			 *   |---|  merge
314 			 *
315 			 *     |-|
316 			 * |---|    merge
317 			 *
318 			 * |-----|
319 			 *   |-|    DSACK smaller
320 			 *
321 			 *   |-|
322 			 * |-----|  DSACK smaller
323 			 */
324 			if (head_blk.start == end)
325 				head_blk.start = start;
326 			else if (head_blk.end == start)
327 				head_blk.end = end;
328 			else {
329 				if (SEQ_LT(head_blk.start, start)) {
330 					tcp_seq temp = start;
331 					start = head_blk.start;
332 					head_blk.start = temp;
333 				}
334 				if (SEQ_GT(head_blk.end, end)) {
335 					tcp_seq temp = end;
336 					end = head_blk.end;
337 					head_blk.end = temp;
338 				}
339 				if ((head_blk.start != start) ||
340 				    (head_blk.end != end)) {
341 					if ((num_saved >= 1) &&
342 					   SEQ_GEQ(saved_blks[num_saved-1].start, start) &&
343 					   SEQ_LEQ(saved_blks[num_saved-1].end, end))
344 						num_saved--;
345 					saved_blks[num_saved].start = start;
346 					saved_blks[num_saved].end = end;
347 					num_saved++;
348 				}
349 			}
350 		} else {
351 			/*
352 			 * This block supercedes the prior block
353 			 */
354 			if ((num_saved >= 1) &&
355 			   SEQ_GEQ(saved_blks[num_saved-1].start, start) &&
356 			   SEQ_LEQ(saved_blks[num_saved-1].end, end))
357 				num_saved--;
358 			/*
359 			 * Save this SACK block.
360 			 */
361 			saved_blks[num_saved].start = start;
362 			saved_blks[num_saved].end = end;
363 			num_saved++;
364 		}
365 	}
366 
367 	/*
368 	 * Update SACK list in tp->sackblks[].
369 	 */
370 	num_head = 0;
371 	if (SEQ_LT(rcv_start, rcv_end)) {
372 		/*
373 		 * The received data segment is an out-of-order segment.  Put
374 		 * head_blk at the top of SACK list.
375 		 */
376 		tp->sackblks[0] = head_blk;
377 		num_head = 1;
378 		/*
379 		 * If the number of saved SACK blocks exceeds its limit,
380 		 * discard the last SACK block.
381 		 */
382 		if (num_saved >= MAX_SACK_BLKS)
383 			num_saved--;
384 	}
385 	if ((rcv_start == rcv_end) &&
386 	    (rcv_start == tp->sackblks[0].end)) {
387 		num_head = 1;
388 	}
389 	if (num_saved > 0) {
390 		/*
391 		 * Copy the saved SACK blocks back.
392 		 */
393 		bcopy(saved_blks, &tp->sackblks[num_head],
394 		      sizeof(struct sackblk) * num_saved);
395 	}
396 
397 	/* Save the number of SACK blocks. */
398 	tp->rcv_numsacks = num_head + num_saved;
399 }
400 
401 void
402 tcp_clean_dsack_blocks(struct tcpcb *tp)
403 {
404 	struct sackblk saved_blks[MAX_SACK_BLKS];
405 	int num_saved, i;
406 
407 	INP_WLOCK_ASSERT(tptoinpcb(tp));
408 	/*
409 	 * Clean up any DSACK blocks that
410 	 * are in our queue of sack blocks.
411 	 *
412 	 */
413 	num_saved = 0;
414 	for (i = 0; i < tp->rcv_numsacks; i++) {
415 		tcp_seq start = tp->sackblks[i].start;
416 		tcp_seq end = tp->sackblks[i].end;
417 		if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
418 			/*
419 			 * Discard this D-SACK block.
420 			 */
421 			continue;
422 		}
423 		/*
424 		 * Save this SACK block.
425 		 */
426 		saved_blks[num_saved].start = start;
427 		saved_blks[num_saved].end = end;
428 		num_saved++;
429 	}
430 	if (num_saved > 0) {
431 		/*
432 		 * Copy the saved SACK blocks back.
433 		 */
434 		bcopy(saved_blks, &tp->sackblks[0],
435 		      sizeof(struct sackblk) * num_saved);
436 	}
437 	tp->rcv_numsacks = num_saved;
438 }
439 
440 /*
441  * Delete all receiver-side SACK information.
442  */
443 void
444 tcp_clean_sackreport(struct tcpcb *tp)
445 {
446 	int i;
447 
448 	INP_WLOCK_ASSERT(tptoinpcb(tp));
449 	tp->rcv_numsacks = 0;
450 	for (i = 0; i < MAX_SACK_BLKS; i++)
451 		tp->sackblks[i].start = tp->sackblks[i].end=0;
452 }
453 
454 /*
455  * Allocate struct sackhole.
456  */
457 static struct sackhole *
458 tcp_sackhole_alloc(struct tcpcb *tp, tcp_seq start, tcp_seq end)
459 {
460 	struct sackhole *hole;
461 
462 	if (tp->snd_numholes >= V_tcp_sack_maxholes ||
463 	    V_tcp_sack_globalholes >= V_tcp_sack_globalmaxholes) {
464 		TCPSTAT_INC(tcps_sack_sboverflow);
465 		return NULL;
466 	}
467 
468 	hole = (struct sackhole *)uma_zalloc(V_sack_hole_zone, M_NOWAIT);
469 	if (hole == NULL)
470 		return NULL;
471 
472 	hole->start = start;
473 	hole->end = end;
474 	hole->rxmit = start;
475 
476 	tp->snd_numholes++;
477 	atomic_add_int(&V_tcp_sack_globalholes, 1);
478 
479 	return hole;
480 }
481 
482 /*
483  * Free struct sackhole.
484  */
485 static void
486 tcp_sackhole_free(struct tcpcb *tp, struct sackhole *hole)
487 {
488 
489 	uma_zfree(V_sack_hole_zone, hole);
490 
491 	tp->snd_numholes--;
492 	atomic_subtract_int(&V_tcp_sack_globalholes, 1);
493 
494 	KASSERT(tp->snd_numholes >= 0, ("tp->snd_numholes >= 0"));
495 	KASSERT(V_tcp_sack_globalholes >= 0, ("tcp_sack_globalholes >= 0"));
496 }
497 
498 /*
499  * Insert new SACK hole into scoreboard.
500  */
501 static struct sackhole *
502 tcp_sackhole_insert(struct tcpcb *tp, tcp_seq start, tcp_seq end,
503     struct sackhole *after)
504 {
505 	struct sackhole *hole;
506 
507 	/* Allocate a new SACK hole. */
508 	hole = tcp_sackhole_alloc(tp, start, end);
509 	if (hole == NULL)
510 		return NULL;
511 
512 	/* Insert the new SACK hole into scoreboard. */
513 	if (after != NULL)
514 		TAILQ_INSERT_AFTER(&tp->snd_holes, after, hole, scblink);
515 	else
516 		TAILQ_INSERT_TAIL(&tp->snd_holes, hole, scblink);
517 
518 	/* Update SACK hint. */
519 	if (tp->sackhint.nexthole == NULL)
520 		tp->sackhint.nexthole = hole;
521 
522 	return hole;
523 }
524 
525 /*
526  * Remove SACK hole from scoreboard.
527  */
528 static void
529 tcp_sackhole_remove(struct tcpcb *tp, struct sackhole *hole)
530 {
531 
532 	/* Update SACK hint. */
533 	if (tp->sackhint.nexthole == hole)
534 		tp->sackhint.nexthole = TAILQ_NEXT(hole, scblink);
535 
536 	/* Remove this SACK hole. */
537 	TAILQ_REMOVE(&tp->snd_holes, hole, scblink);
538 
539 	/* Free this SACK hole. */
540 	tcp_sackhole_free(tp, hole);
541 }
542 
543 /*
544  * Process cumulative ACK and the TCP SACK option to update the scoreboard.
545  * tp->snd_holes is an ordered list of holes (oldest to newest, in terms of
546  * the sequence space).
547  * Returns 1 if incoming ACK has previously unknown SACK information,
548  * 0 otherwise.
549  */
550 int
551 tcp_sack_doack(struct tcpcb *tp, struct tcpopt *to, tcp_seq th_ack)
552 {
553 	struct sackhole *cur, *temp;
554 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1], *sblkp;
555 	int i, j, num_sack_blks, sack_changed;
556 	int delivered_data, left_edge_delta;
557 
558 	INP_WLOCK_ASSERT(tptoinpcb(tp));
559 
560 	num_sack_blks = 0;
561 	sack_changed = 0;
562 	delivered_data = 0;
563 	left_edge_delta = 0;
564 	/*
565 	 * If SND.UNA will be advanced by SEG.ACK, and if SACK holes exist,
566 	 * treat [SND.UNA, SEG.ACK) as if it is a SACK block.
567 	 * Account changes to SND.UNA always in delivered data.
568 	 */
569 	if (SEQ_LT(tp->snd_una, th_ack) && !TAILQ_EMPTY(&tp->snd_holes)) {
570 		left_edge_delta = th_ack - tp->snd_una;
571 		sack_blocks[num_sack_blks].start = tp->snd_una;
572 		sack_blocks[num_sack_blks++].end = th_ack;
573 		/*
574 		 * Pulling snd_fack forward if we got here
575 		 * due to DSACK blocks
576 		 */
577 		if (SEQ_LT(tp->snd_fack, th_ack)) {
578 			delivered_data += th_ack - tp->snd_una;
579 			tp->snd_fack = th_ack;
580 			sack_changed = 1;
581 		}
582 	}
583 	/*
584 	 * Append received valid SACK blocks to sack_blocks[], but only if we
585 	 * received new blocks from the other side.
586 	 */
587 	if (to->to_flags & TOF_SACK) {
588 		for (i = 0; i < to->to_nsacks; i++) {
589 			bcopy((to->to_sacks + i * TCPOLEN_SACK),
590 			    &sack, sizeof(sack));
591 			sack.start = ntohl(sack.start);
592 			sack.end = ntohl(sack.end);
593 			if (SEQ_GT(sack.end, sack.start) &&
594 			    SEQ_GT(sack.start, tp->snd_una) &&
595 			    SEQ_GT(sack.start, th_ack) &&
596 			    SEQ_LT(sack.start, tp->snd_max) &&
597 			    SEQ_GT(sack.end, tp->snd_una) &&
598 			    SEQ_LEQ(sack.end, tp->snd_max)) {
599 				sack_blocks[num_sack_blks++] = sack;
600 			} else if (SEQ_LEQ(sack.start, th_ack) &&
601 			    SEQ_LEQ(sack.end, th_ack)) {
602 				/*
603 				 * Its a D-SACK block.
604 				 */
605 				tcp_record_dsack(tp, sack.start, sack.end, 0);
606 			}
607 		}
608 	}
609 	/*
610 	 * Return if SND.UNA is not advanced and no valid SACK block is
611 	 * received.
612 	 */
613 	if (num_sack_blks == 0)
614 		return (sack_changed);
615 
616 	/*
617 	 * Sort the SACK blocks so we can update the scoreboard with just one
618 	 * pass. The overhead of sorting up to 4+1 elements is less than
619 	 * making up to 4+1 passes over the scoreboard.
620 	 */
621 	for (i = 0; i < num_sack_blks; i++) {
622 		for (j = i + 1; j < num_sack_blks; j++) {
623 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
624 				sack = sack_blocks[i];
625 				sack_blocks[i] = sack_blocks[j];
626 				sack_blocks[j] = sack;
627 			}
628 		}
629 	}
630 	if (TAILQ_EMPTY(&tp->snd_holes)) {
631 		/*
632 		 * Empty scoreboard. Need to initialize snd_fack (it may be
633 		 * uninitialized or have a bogus value). Scoreboard holes
634 		 * (from the sack blocks received) are created later below
635 		 * (in the logic that adds holes to the tail of the
636 		 * scoreboard).
637 		 */
638 		tp->snd_fack = SEQ_MAX(tp->snd_una, th_ack);
639 		tp->sackhint.sacked_bytes = 0;	/* reset */
640 	}
641 	/*
642 	 * In the while-loop below, incoming SACK blocks (sack_blocks[]) and
643 	 * SACK holes (snd_holes) are traversed from their tails with just
644 	 * one pass in order to reduce the number of compares especially when
645 	 * the bandwidth-delay product is large.
646 	 *
647 	 * Note: Typically, in the first RTT of SACK recovery, the highest
648 	 * three or four SACK blocks with the same ack number are received.
649 	 * In the second RTT, if retransmitted data segments are not lost,
650 	 * the highest three or four SACK blocks with ack number advancing
651 	 * are received.
652 	 */
653 	sblkp = &sack_blocks[num_sack_blks - 1];	/* Last SACK block */
654 	tp->sackhint.last_sack_ack = sblkp->end;
655 	if (SEQ_LT(tp->snd_fack, sblkp->start)) {
656 		/*
657 		 * The highest SACK block is beyond fack.  First,
658 		 * check if there was a successful Rescue Retransmission,
659 		 * and move this hole left. With normal holes, snd_fack
660 		 * is always to the right of the end.
661 		 */
662 		if (((temp = TAILQ_LAST(&tp->snd_holes, sackhole_head)) != NULL) &&
663 		    SEQ_LEQ(tp->snd_fack,temp->end)) {
664 			temp->start = SEQ_MAX(tp->snd_fack, SEQ_MAX(tp->snd_una, th_ack));
665 			temp->end = sblkp->start;
666 			temp->rxmit = temp->start;
667 			delivered_data += sblkp->end - sblkp->start;
668 			tp->snd_fack = sblkp->end;
669 			sblkp--;
670 			sack_changed = 1;
671 		} else {
672 			/*
673 			 * Append a new SACK hole at the tail.  If the
674 			 * second or later highest SACK blocks are also
675 			 * beyond the current fack, they will be inserted
676 			 * by way of hole splitting in the while-loop below.
677 			 */
678 			temp = tcp_sackhole_insert(tp, tp->snd_fack,sblkp->start,NULL);
679 			if (temp != NULL) {
680 				delivered_data += sblkp->end - sblkp->start;
681 				tp->snd_fack = sblkp->end;
682 				/* Go to the previous sack block. */
683 				sblkp--;
684 				sack_changed = 1;
685 			} else {
686 				/*
687 				 * We failed to add a new hole based on the current
688 				 * sack block.  Skip over all the sack blocks that
689 				 * fall completely to the right of snd_fack and
690 				 * proceed to trim the scoreboard based on the
691 				 * remaining sack blocks.  This also trims the
692 				 * scoreboard for th_ack (which is sack_blocks[0]).
693 				 */
694 				while (sblkp >= sack_blocks &&
695 				       SEQ_LT(tp->snd_fack, sblkp->start))
696 					sblkp--;
697 				if (sblkp >= sack_blocks &&
698 				    SEQ_LT(tp->snd_fack, sblkp->end)) {
699 					delivered_data += sblkp->end - tp->snd_fack;
700 					tp->snd_fack = sblkp->end;
701 					sack_changed = 1;
702 				}
703 			}
704 		}
705 	} else if (SEQ_LT(tp->snd_fack, sblkp->end)) {
706 		/* fack is advanced. */
707 		delivered_data += sblkp->end - tp->snd_fack;
708 		tp->snd_fack = sblkp->end;
709 		sack_changed = 1;
710 	}
711 	cur = TAILQ_LAST(&tp->snd_holes, sackhole_head); /* Last SACK hole. */
712 	/*
713 	 * Since the incoming sack blocks are sorted, we can process them
714 	 * making one sweep of the scoreboard.
715 	 */
716 	while (sblkp >= sack_blocks  && cur != NULL) {
717 		if (SEQ_GEQ(sblkp->start, cur->end)) {
718 			/*
719 			 * SACKs data beyond the current hole.  Go to the
720 			 * previous sack block.
721 			 */
722 			sblkp--;
723 			continue;
724 		}
725 		if (SEQ_LEQ(sblkp->end, cur->start)) {
726 			/*
727 			 * SACKs data before the current hole.  Go to the
728 			 * previous hole.
729 			 */
730 			cur = TAILQ_PREV(cur, sackhole_head, scblink);
731 			continue;
732 		}
733 		tp->sackhint.sack_bytes_rexmit -=
734 		    (SEQ_MIN(cur->rxmit, cur->end) - cur->start);
735 		KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
736 		    ("sackhint bytes rtx >= 0"));
737 		sack_changed = 1;
738 		if (SEQ_LEQ(sblkp->start, cur->start)) {
739 			/* Data acks at least the beginning of hole. */
740 			if (SEQ_GEQ(sblkp->end, cur->end)) {
741 				/* Acks entire hole, so delete hole. */
742 				delivered_data += (cur->end - cur->start);
743 				temp = cur;
744 				cur = TAILQ_PREV(cur, sackhole_head, scblink);
745 				tcp_sackhole_remove(tp, temp);
746 				/*
747 				 * The sack block may ack all or part of the
748 				 * next hole too, so continue onto the next
749 				 * hole.
750 				 */
751 				continue;
752 			} else {
753 				/* Move start of hole forward. */
754 				delivered_data += (sblkp->end - cur->start);
755 				cur->start = sblkp->end;
756 				cur->rxmit = SEQ_MAX(cur->rxmit, cur->start);
757 			}
758 		} else {
759 			/* Data acks at least the end of hole. */
760 			if (SEQ_GEQ(sblkp->end, cur->end)) {
761 				/* Move end of hole backward. */
762 				delivered_data += (cur->end - sblkp->start);
763 				cur->end = sblkp->start;
764 				cur->rxmit = SEQ_MIN(cur->rxmit, cur->end);
765 				if ((tp->t_flags & TF_LRD) && SEQ_GEQ(cur->rxmit, cur->end))
766 					cur->rxmit = tp->snd_recover;
767 			} else {
768 				/*
769 				 * ACKs some data in middle of a hole; need
770 				 * to split current hole
771 				 */
772 				temp = tcp_sackhole_insert(tp, sblkp->end,
773 				    cur->end, cur);
774 				if (temp != NULL) {
775 					if (SEQ_GT(cur->rxmit, temp->rxmit)) {
776 						temp->rxmit = cur->rxmit;
777 						tp->sackhint.sack_bytes_rexmit +=
778 						    (SEQ_MIN(temp->rxmit,
779 						    temp->end) - temp->start);
780 					}
781 					cur->end = sblkp->start;
782 					cur->rxmit = SEQ_MIN(cur->rxmit,
783 					    cur->end);
784 					if ((tp->t_flags & TF_LRD) && SEQ_GEQ(cur->rxmit, cur->end))
785 						cur->rxmit = tp->snd_recover;
786 					delivered_data += (sblkp->end - sblkp->start);
787 				}
788 			}
789 		}
790 		tp->sackhint.sack_bytes_rexmit +=
791 		    (SEQ_MIN(cur->rxmit, cur->end) - cur->start);
792 		/*
793 		 * Testing sblkp->start against cur->start tells us whether
794 		 * we're done with the sack block or the sack hole.
795 		 * Accordingly, we advance one or the other.
796 		 */
797 		if (SEQ_LEQ(sblkp->start, cur->start))
798 			cur = TAILQ_PREV(cur, sackhole_head, scblink);
799 		else
800 			sblkp--;
801 	}
802 	if (!(to->to_flags & TOF_SACK))
803 		/*
804 		 * If this ACK did not contain any
805 		 * SACK blocks, any only moved the
806 		 * left edge right, it is a pure
807 		 * cumulative ACK. Do not count
808 		 * DupAck for this. Also required
809 		 * for RFC6675 rescue retransmission.
810 		 */
811 		sack_changed = 0;
812 	tp->sackhint.delivered_data = delivered_data;
813 	tp->sackhint.sacked_bytes += delivered_data - left_edge_delta;
814 	KASSERT((delivered_data >= 0), ("delivered_data < 0"));
815 	KASSERT((tp->sackhint.sacked_bytes >= 0), ("sacked_bytes < 0"));
816 	return (sack_changed);
817 }
818 
819 /*
820  * Free all SACK holes to clear the scoreboard.
821  */
822 void
823 tcp_free_sackholes(struct tcpcb *tp)
824 {
825 	struct sackhole *q;
826 
827 	INP_WLOCK_ASSERT(tptoinpcb(tp));
828 	while ((q = TAILQ_FIRST(&tp->snd_holes)) != NULL)
829 		tcp_sackhole_remove(tp, q);
830 	tp->sackhint.sack_bytes_rexmit = 0;
831 
832 	KASSERT(tp->snd_numholes == 0, ("tp->snd_numholes == 0"));
833 	KASSERT(tp->sackhint.nexthole == NULL,
834 		("tp->sackhint.nexthole == NULL"));
835 }
836 
837 /*
838  * Partial ack handling within a sack recovery episode.  Keeping this very
839  * simple for now.  When a partial ack is received, force snd_cwnd to a value
840  * that will allow the sender to transmit no more than 2 segments.  If
841  * necessary, a better scheme can be adopted at a later point, but for now,
842  * the goal is to prevent the sender from bursting a large amount of data in
843  * the midst of sack recovery.
844  */
845 void
846 tcp_sack_partialack(struct tcpcb *tp, struct tcphdr *th)
847 {
848 	int num_segs = 1;
849 	u_int maxseg = tcp_maxseg(tp);
850 
851 	INP_WLOCK_ASSERT(tptoinpcb(tp));
852 	tcp_timer_activate(tp, TT_REXMT, 0);
853 	tp->t_rtttime = 0;
854 	/* Send one or 2 segments based on how much new data was acked. */
855 	if ((BYTES_THIS_ACK(tp, th) / maxseg) >= 2)
856 		num_segs = 2;
857 	tp->snd_cwnd = (tp->sackhint.sack_bytes_rexmit +
858 	    (tp->snd_nxt - tp->snd_recover) + num_segs * maxseg);
859 	if (tp->snd_cwnd > tp->snd_ssthresh)
860 		tp->snd_cwnd = tp->snd_ssthresh;
861 	tp->t_flags |= TF_ACKNOW;
862 	/*
863 	 * RFC6675 rescue retransmission
864 	 * Add a hole between th_ack (snd_una is not yet set) and snd_max,
865 	 * if this was a pure cumulative ACK and no data was send beyond
866 	 * recovery point. Since the data in the socket has not been freed
867 	 * at this point, we check if the scoreboard is empty, and the ACK
868 	 * delivered some new data, indicating a full ACK. Also, if the
869 	 * recovery point is still at snd_max, we are probably application
870 	 * limited. However, this inference might not always be true. The
871 	 * rescue retransmission may rarely be slightly premature
872 	 * compared to RFC6675.
873 	 * The corresponding ACK+SACK will cause any further outstanding
874 	 * segments to be retransmitted. This addresses a corner case, when
875 	 * the trailing packets of a window are lost and no further data
876 	 * is available for sending.
877 	 */
878 	if ((V_tcp_do_newsack) &&
879 	    SEQ_LT(th->th_ack, tp->snd_recover) &&
880 	    TAILQ_EMPTY(&tp->snd_holes) &&
881 	    (tp->sackhint.delivered_data > 0)) {
882 		/*
883 		 * Exclude FIN sequence space in
884 		 * the hole for the rescue retransmission,
885 		 * and also don't create a hole, if only
886 		 * the ACK for a FIN is outstanding.
887 		 */
888 		tcp_seq highdata = tp->snd_max;
889 		if (tp->t_flags & TF_SENTFIN)
890 			highdata--;
891 		highdata = SEQ_MIN(highdata, tp->snd_recover);
892 		if (th->th_ack != highdata) {
893 			tp->snd_fack = th->th_ack;
894 			(void)tcp_sackhole_insert(tp, SEQ_MAX(th->th_ack,
895 			    highdata - maxseg), highdata, NULL);
896 		}
897 	}
898 	(void) tcp_output(tp);
899 }
900 
901 #if 0
902 /*
903  * Debug version of tcp_sack_output() that walks the scoreboard.  Used for
904  * now to sanity check the hint.
905  */
906 static struct sackhole *
907 tcp_sack_output_debug(struct tcpcb *tp, int *sack_bytes_rexmt)
908 {
909 	struct sackhole *p;
910 
911 	INP_WLOCK_ASSERT(tptoinpcb(tp));
912 	*sack_bytes_rexmt = 0;
913 	TAILQ_FOREACH(p, &tp->snd_holes, scblink) {
914 		if (SEQ_LT(p->rxmit, p->end)) {
915 			if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */
916 				continue;
917 			}
918 			*sack_bytes_rexmt += (p->rxmit - p->start);
919 			break;
920 		}
921 		*sack_bytes_rexmt += (SEQ_MIN(p->rxmit, p->end) - p->start);
922 	}
923 	return (p);
924 }
925 #endif
926 
927 /*
928  * Returns the next hole to retransmit and the number of retransmitted bytes
929  * from the scoreboard.  We store both the next hole and the number of
930  * retransmitted bytes as hints (and recompute these on the fly upon SACK/ACK
931  * reception).  This avoids scoreboard traversals completely.
932  *
933  * The loop here will traverse *at most* one link.  Here's the argument.  For
934  * the loop to traverse more than 1 link before finding the next hole to
935  * retransmit, we would need to have at least 1 node following the current
936  * hint with (rxmit == end).  But, for all holes following the current hint,
937  * (start == rxmit), since we have not yet retransmitted from them.
938  * Therefore, in order to traverse more 1 link in the loop below, we need to
939  * have at least one node following the current hint with (start == rxmit ==
940  * end).  But that can't happen, (start == end) means that all the data in
941  * that hole has been sacked, in which case, the hole would have been removed
942  * from the scoreboard.
943  */
944 struct sackhole *
945 tcp_sack_output(struct tcpcb *tp, int *sack_bytes_rexmt)
946 {
947 	struct sackhole *hole = NULL;
948 
949 	INP_WLOCK_ASSERT(tptoinpcb(tp));
950 	*sack_bytes_rexmt = tp->sackhint.sack_bytes_rexmit;
951 	hole = tp->sackhint.nexthole;
952 	if (hole == NULL)
953 		return (hole);
954 	if (SEQ_GEQ(hole->rxmit, hole->end)) {
955 		for (;;) {
956 			hole = TAILQ_NEXT(hole, scblink);
957 			if (hole == NULL)
958 				return (hole);
959 			if (SEQ_LT(hole->rxmit, hole->end)) {
960 				tp->sackhint.nexthole = hole;
961 				break;
962 			}
963 		}
964 	}
965 	KASSERT(SEQ_LT(hole->start, hole->end), ("%s: hole.start >= hole.end", __func__));
966 	if (!(V_tcp_do_newsack)) {
967 		KASSERT(SEQ_LT(hole->start, tp->snd_fack), ("%s: hole.start >= snd.fack", __func__));
968 		KASSERT(SEQ_LT(hole->end, tp->snd_fack), ("%s: hole.end >= snd.fack", __func__));
969 		KASSERT(SEQ_LT(hole->rxmit, tp->snd_fack), ("%s: hole.rxmit >= snd.fack", __func__));
970 		if (SEQ_GEQ(hole->start, hole->end) ||
971 		    SEQ_GEQ(hole->start, tp->snd_fack) ||
972 		    SEQ_GEQ(hole->end, tp->snd_fack) ||
973 		    SEQ_GEQ(hole->rxmit, tp->snd_fack)) {
974 			log(LOG_CRIT,"tcp: invalid SACK hole (%u-%u,%u) vs fwd ack %u, ignoring.\n",
975 					hole->start, hole->end, hole->rxmit, tp->snd_fack);
976 			return (NULL);
977 		}
978 	}
979 	return (hole);
980 }
981 
982 /*
983  * After a timeout, the SACK list may be rebuilt.  This SACK information
984  * should be used to avoid retransmitting SACKed data.  This function
985  * traverses the SACK list to see if snd_nxt should be moved forward.
986  */
987 void
988 tcp_sack_adjust(struct tcpcb *tp)
989 {
990 	struct sackhole *p, *cur = TAILQ_FIRST(&tp->snd_holes);
991 
992 	INP_WLOCK_ASSERT(tptoinpcb(tp));
993 	if (cur == NULL)
994 		return; /* No holes */
995 	if (SEQ_GEQ(tp->snd_nxt, tp->snd_fack))
996 		return; /* We're already beyond any SACKed blocks */
997 	/*-
998 	 * Two cases for which we want to advance snd_nxt:
999 	 * i) snd_nxt lies between end of one hole and beginning of another
1000 	 * ii) snd_nxt lies between end of last hole and snd_fack
1001 	 */
1002 	while ((p = TAILQ_NEXT(cur, scblink)) != NULL) {
1003 		if (SEQ_LT(tp->snd_nxt, cur->end))
1004 			return;
1005 		if (SEQ_GEQ(tp->snd_nxt, p->start))
1006 			cur = p;
1007 		else {
1008 			tp->snd_nxt = p->start;
1009 			return;
1010 		}
1011 	}
1012 	if (SEQ_LT(tp->snd_nxt, cur->end))
1013 		return;
1014 	tp->snd_nxt = tp->snd_fack;
1015 }
1016 
1017 /*
1018  * Lost Retransmission Detection
1019  * Check is FACK is beyond the rexmit of the leftmost hole.
1020  * If yes, we restart sending from still existing holes,
1021  * and adjust cwnd via the congestion control module.
1022  */
1023 void
1024 tcp_sack_lost_retransmission(struct tcpcb *tp, struct tcphdr *th)
1025 {
1026 	struct sackhole *temp;
1027 
1028 	if (IN_RECOVERY(tp->t_flags) &&
1029 	    SEQ_GT(tp->snd_fack, tp->snd_recover) &&
1030 	    ((temp = TAILQ_FIRST(&tp->snd_holes)) != NULL) &&
1031 	    SEQ_GEQ(temp->rxmit, temp->end) &&
1032 	    SEQ_GEQ(tp->snd_fack, temp->rxmit)) {
1033 		TCPSTAT_INC(tcps_sack_lostrexmt);
1034 		/*
1035 		 * Start retransmissions from the first hole, and
1036 		 * subsequently all other remaining holes, including
1037 		 * those, which had been sent completely before.
1038 		 */
1039 		tp->sackhint.nexthole = temp;
1040 		TAILQ_FOREACH(temp, &tp->snd_holes, scblink) {
1041 			if (SEQ_GEQ(tp->snd_fack, temp->rxmit) &&
1042 			    SEQ_GEQ(temp->rxmit, temp->end))
1043 				temp->rxmit = temp->start;
1044 		}
1045 		/*
1046 		 * Remember the old ssthresh, to deduct the beta factor used
1047 		 * by the CC module. Finally, set cwnd to ssthresh just
1048 		 * prior to invoking another cwnd reduction by the CC
1049 		 * module, to not shrink it excessively.
1050 		 */
1051 		tp->snd_cwnd = tp->snd_ssthresh;
1052 		/*
1053 		 * Formally exit recovery, and let the CC module adjust
1054 		 * ssthresh as intended.
1055 		 */
1056 		EXIT_RECOVERY(tp->t_flags);
1057 		cc_cong_signal(tp, th, CC_NDUPACK);
1058 		/*
1059 		 * For PRR, adjust recover_fs as if this new reduction
1060 		 * initialized this variable.
1061 		 * cwnd will be adjusted by SACK or PRR processing
1062 		 * subsequently, only set it to a safe value here.
1063 		 */
1064 		tp->snd_cwnd = tcp_maxseg(tp);
1065 		tp->sackhint.recover_fs = (tp->snd_max - tp->snd_una) -
1066 					    tp->sackhint.recover_fs;
1067 	}
1068 }
1069