xref: /freebsd/sys/netinet/tcp_sack.c (revision 7afc53b8dfcc7d5897920ce6cc7e842fbb4ab813)
1 /*-
2  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 4. Neither the name of the University nor the names of its contributors
14  *    may be used to endorse or promote products derived from this software
15  *    without specific prior written permission.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  *
29  *	@(#)tcp_sack.c	8.12 (Berkeley) 5/24/95
30  * $FreeBSD$
31  */
32 
33 /*-
34  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994
35  *	The Regents of the University of California.  All rights reserved.
36  *
37  * Redistribution and use in source and binary forms, with or without
38  * modification, are permitted provided that the following conditions
39  * are met:
40  * 1. Redistributions of source code must retain the above copyright
41  *    notice, this list of conditions and the following disclaimer.
42  * 2. Redistributions in binary form must reproduce the above copyright
43  *    notice, this list of conditions and the following disclaimer in the
44  *    documentation and/or other materials provided with the distribution.
45  * 3. Neither the name of the University nor the names of its contributors
46  *    may be used to endorse or promote products derived from this software
47  *    without specific prior written permission.
48  *
49  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
50  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
51  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
52  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
53  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
54  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
55  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
56  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
57  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
58  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
59  * SUCH DAMAGE.
60  *
61  *	@@(#)COPYRIGHT	1.1 (NRL) 17 January 1995
62  *
63  * NRL grants permission for redistribution and use in source and binary
64  * forms, with or without modification, of the software and documentation
65  * created at NRL provided that the following conditions are met:
66  *
67  * 1. Redistributions of source code must retain the above copyright
68  *    notice, this list of conditions and the following disclaimer.
69  * 2. Redistributions in binary form must reproduce the above copyright
70  *    notice, this list of conditions and the following disclaimer in the
71  *    documentation and/or other materials provided with the distribution.
72  * 3. All advertising materials mentioning features or use of this software
73  *    must display the following acknowledgements:
74  *	This product includes software developed by the University of
75  *	California, Berkeley and its contributors.
76  *	This product includes software developed at the Information
77  *	Technology Division, US Naval Research Laboratory.
78  * 4. Neither the name of the NRL nor the names of its contributors
79  *    may be used to endorse or promote products derived from this software
80  *    without specific prior written permission.
81  *
82  * THE SOFTWARE PROVIDED BY NRL IS PROVIDED BY NRL AND CONTRIBUTORS ``AS
83  * IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
84  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
85  * PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL NRL OR
86  * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
87  * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
88  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
89  * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
90  * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
91  * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
92  * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
93  *
94  * The views and conclusions contained in the software and documentation
95  * are those of the authors and should not be interpreted as representing
96  * official policies, either expressed or implied, of the US Naval
97  * Research Laboratory (NRL).
98  */
99 #include "opt_inet.h"
100 #include "opt_inet6.h"
101 #include "opt_ipsec.h"
102 #include "opt_tcpdebug.h"
103 #include "opt_tcp_input.h"
104 #include "opt_tcp_sack.h"
105 
106 #include <sys/param.h>
107 #include <sys/systm.h>
108 #include <sys/kernel.h>
109 #include <sys/sysctl.h>
110 #include <sys/malloc.h>
111 #include <sys/mbuf.h>
112 #include <sys/proc.h>		/* for proc0 declaration */
113 #include <sys/protosw.h>
114 #include <sys/socket.h>
115 #include <sys/socketvar.h>
116 #include <sys/syslog.h>
117 #include <sys/systm.h>
118 
119 #include <machine/cpu.h>	/* before tcp_seq.h, for tcp_random18() */
120 
121 #include <vm/uma.h>
122 
123 #include <net/if.h>
124 #include <net/route.h>
125 
126 #include <netinet/in.h>
127 #include <netinet/in_systm.h>
128 #include <netinet/ip.h>
129 #include <netinet/ip_icmp.h>	/* for ICMP_BANDLIM		*/
130 #include <netinet/in_var.h>
131 #include <netinet/icmp_var.h>	/* for ICMP_BANDLIM		*/
132 #include <netinet/in_pcb.h>
133 #include <netinet/ip_var.h>
134 #include <netinet/ip6.h>
135 #include <netinet/icmp6.h>
136 #include <netinet6/nd6.h>
137 #include <netinet6/ip6_var.h>
138 #include <netinet6/in6_pcb.h>
139 #include <netinet/tcp.h>
140 #include <netinet/tcp_fsm.h>
141 #include <netinet/tcp_seq.h>
142 #include <netinet/tcp_timer.h>
143 #include <netinet/tcp_var.h>
144 #include <netinet6/tcp6_var.h>
145 #include <netinet/tcpip.h>
146 #ifdef TCPDEBUG
147 #include <netinet/tcp_debug.h>
148 #endif /* TCPDEBUG */
149 
150 #ifdef FAST_IPSEC
151 #include <netipsec/ipsec.h>
152 #include <netipsec/ipsec6.h>
153 #endif
154 
155 #ifdef IPSEC
156 #include <netinet6/ipsec.h>
157 #include <netinet6/ipsec6.h>
158 #include <netkey/key.h>
159 #endif /*IPSEC*/
160 #include <machine/in_cksum.h>
161 
162 extern struct uma_zone *sack_hole_zone;
163 
164 SYSCTL_NODE(_net_inet_tcp, OID_AUTO, sack, CTLFLAG_RW, 0, "TCP SACK");
165 int tcp_do_sack = 1;
166 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, enable, CTLFLAG_RW,
167 	&tcp_do_sack, 0, "Enable/Disable TCP SACK support");
168 TUNABLE_INT("net.inet.tcp.sack.enable", &tcp_do_sack);
169 
170 static int tcp_sack_maxholes = 128;
171 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, maxholes, CTLFLAG_RW,
172 	&tcp_sack_maxholes, 0,
173     "Maximum number of TCP SACK holes allowed per connection");
174 
175 static int tcp_sack_globalmaxholes = 65536;
176 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, globalmaxholes, CTLFLAG_RW,
177 	&tcp_sack_globalmaxholes, 0,
178     "Global maximum number of TCP SACK holes");
179 
180 static int tcp_sack_globalholes = 0;
181 SYSCTL_INT(_net_inet_tcp_sack, OID_AUTO, globalholes, CTLFLAG_RD,
182     &tcp_sack_globalholes, 0,
183     "Global number of TCP SACK holes currently allocated");
184 
185 /*
186  * This function is called upon receipt of new valid data (while not in header
187  * prediction mode), and it updates the ordered list of sacks.
188  */
189 void
190 tcp_update_sack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
191 {
192 	/*
193 	 * First reported block MUST be the most recent one.  Subsequent
194 	 * blocks SHOULD be in the order in which they arrived at the
195 	 * receiver.  These two conditions make the implementation fully
196 	 * compliant with RFC 2018.
197 	 */
198 	struct sackblk head_blk, saved_blks[MAX_SACK_BLKS];
199 	int num_head, num_saved, i;
200 
201 	INP_LOCK_ASSERT(tp->t_inpcb);
202 
203 	/* Check arguments */
204 	KASSERT(SEQ_LT(rcv_start, rcv_end), ("rcv_start < rcv_end"));
205 
206 	/* SACK block for the received segment. */
207 	head_blk.start = rcv_start;
208 	head_blk.end = rcv_end;
209 
210 	/*
211 	 * Merge updated SACK blocks into head_blk, and
212 	 * save unchanged SACK blocks into saved_blks[].
213 	 * num_saved will have the number of the saved SACK blocks.
214 	 */
215 	num_saved = 0;
216 	for (i = 0; i < tp->rcv_numsacks; i++) {
217 		tcp_seq start = tp->sackblks[i].start;
218 		tcp_seq end = tp->sackblks[i].end;
219 		if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
220 			/*
221 			 * Discard this SACK block.
222 			 */
223 		} else if (SEQ_LEQ(head_blk.start, end) &&
224 			   SEQ_GEQ(head_blk.end, start)) {
225 			/*
226 			 * Merge this SACK block into head_blk.
227 			 * This SACK block itself will be discarded.
228 			 */
229 			if (SEQ_GT(head_blk.start, start))
230 				head_blk.start = start;
231 			if (SEQ_LT(head_blk.end, end))
232 				head_blk.end = end;
233 		} else {
234 			/*
235 			 * Save this SACK block.
236 			 */
237 			saved_blks[num_saved].start = start;
238 			saved_blks[num_saved].end = end;
239 			num_saved++;
240 		}
241 	}
242 
243 	/*
244 	 * Update SACK list in tp->sackblks[].
245 	 */
246 	num_head = 0;
247 	if (SEQ_GT(head_blk.start, tp->rcv_nxt)) {
248 		/*
249 		 * The received data segment is an out-of-order segment.
250 		 * Put head_blk at the top of SACK list.
251 		 */
252 		tp->sackblks[0] = head_blk;
253 		num_head = 1;
254 		/*
255 		 * If the number of saved SACK blocks exceeds its limit,
256 		 * discard the last SACK block.
257 		 */
258 		if (num_saved >= MAX_SACK_BLKS)
259 			num_saved--;
260 	}
261 	if (num_saved > 0) {
262 		/*
263 		 * Copy the saved SACK blocks back.
264 		 */
265 		bcopy(saved_blks, &tp->sackblks[num_head],
266 		      sizeof(struct sackblk) * num_saved);
267 	}
268 
269 	/* Save the number of SACK blocks. */
270 	tp->rcv_numsacks = num_head + num_saved;
271 }
272 
273 /*
274  * Delete all receiver-side SACK information.
275  */
276 void
277 tcp_clean_sackreport(tp)
278 	struct tcpcb *tp;
279 {
280 	int i;
281 
282 	INP_LOCK_ASSERT(tp->t_inpcb);
283 	tp->rcv_numsacks = 0;
284 	for (i = 0; i < MAX_SACK_BLKS; i++)
285 		tp->sackblks[i].start = tp->sackblks[i].end=0;
286 }
287 
288 /*
289  * Allocate struct sackhole.
290  */
291 static struct sackhole *
292 tcp_sackhole_alloc(struct tcpcb *tp, tcp_seq start, tcp_seq end)
293 {
294 	struct sackhole *hole;
295 
296 	if (tp->snd_numholes >= tcp_sack_maxholes ||
297 	    tcp_sack_globalholes >= tcp_sack_globalmaxholes) {
298 		tcpstat.tcps_sack_sboverflow++;
299 		return NULL;
300 	}
301 
302 	hole = (struct sackhole *)uma_zalloc(sack_hole_zone, M_NOWAIT);
303 	if (hole == NULL)
304 		return NULL;
305 
306 	hole->start = start;
307 	hole->end = end;
308 	hole->rxmit = start;
309 
310 	tp->snd_numholes++;
311 	tcp_sack_globalholes++;
312 
313 	return hole;
314 }
315 
316 /*
317  * Free struct sackhole.
318  */
319 static void
320 tcp_sackhole_free(struct tcpcb *tp, struct sackhole *hole)
321 {
322 	uma_zfree(sack_hole_zone, hole);
323 
324 	tp->snd_numholes--;
325 	tcp_sack_globalholes--;
326 
327 	KASSERT(tp->snd_numholes >= 0, ("tp->snd_numholes >= 0"));
328 	KASSERT(tcp_sack_globalholes >= 0, ("tcp_sack_globalholes >= 0"));
329 }
330 
331 /*
332  * Insert new SACK hole into scoreboard.
333  */
334 static struct sackhole *
335 tcp_sackhole_insert(struct tcpcb *tp, tcp_seq start, tcp_seq end,
336 		    struct sackhole *after)
337 {
338 	struct sackhole *hole;
339 
340 	/* Allocate a new SACK hole. */
341 	hole = tcp_sackhole_alloc(tp, start, end);
342 	if (hole == NULL)
343 		return NULL;
344 
345 	/* Insert the new SACK hole into scoreboard */
346 	if (after != NULL)
347 		TAILQ_INSERT_AFTER(&tp->snd_holes, after, hole, scblink);
348 	else
349 		TAILQ_INSERT_TAIL(&tp->snd_holes, hole, scblink);
350 
351 	/* Update SACK hint. */
352 	if (tp->sackhint.nexthole == NULL)
353 		tp->sackhint.nexthole = hole;
354 
355 	return hole;
356 }
357 
358 /*
359  * Remove SACK hole from scoreboard.
360  */
361 static void
362 tcp_sackhole_remove(struct tcpcb *tp, struct sackhole *hole)
363 {
364 	/* Update SACK hint. */
365 	if (tp->sackhint.nexthole == hole)
366 		tp->sackhint.nexthole = TAILQ_NEXT(hole, scblink);
367 
368 	/* Remove this SACK hole. */
369 	TAILQ_REMOVE(&tp->snd_holes, hole, scblink);
370 
371 	/* Free this SACK hole. */
372 	tcp_sackhole_free(tp, hole);
373 }
374 
375 /*
376  * Process the TCP SACK option.  Returns 1 if tcp_dooptions() should continue,
377  * and 0 otherwise, if the option was fine.  tp->snd_holes is an ordered list
378  * of holes (oldest to newest, in terms of the sequence space).
379  */
380 int
381 tcp_sack_option(struct tcpcb *tp, struct tcphdr *th, u_char *cp, int optlen)
382 {
383 	int tmp_olen;
384 	u_char *tmp_cp;
385 	struct sackhole *cur, *temp;
386 	struct sackblk sack, sack_blocks[TCP_MAX_SACK], *sblkp;
387 	int i, j, num_sack_blks;
388 
389 	INP_LOCK_ASSERT(tp->t_inpcb);
390 	if (!tp->sack_enable)
391 		return (1);
392 	if ((th->th_flags & TH_ACK) == 0)
393 		return (1);
394 	/* Note: TCPOLEN_SACK must be 2*sizeof(tcp_seq) */
395 	if (optlen <= 2 || (optlen - 2) % TCPOLEN_SACK != 0)
396 		return (1);
397 	/* If ack is outside [snd_una, snd_max], ignore the SACK options */
398 	if (SEQ_LT(th->th_ack, tp->snd_una) || SEQ_GT(th->th_ack, tp->snd_max))
399 		return (1);
400 	tmp_cp = cp + 2;
401 	tmp_olen = optlen - 2;
402 	tcpstat.tcps_sack_rcv_blocks++;
403 	/*
404 	 * Sort the SACK blocks so we can update the scoreboard
405 	 * with just one pass. The overhead of sorting upto 4 elements
406 	 * is less than making upto 4 passes over the scoreboard.
407 	 */
408 	num_sack_blks = 0;
409 	while (tmp_olen > 0) {
410 		bcopy(tmp_cp, &sack, sizeof(sack));
411 		sack.start = ntohl(sack.start);
412 		sack.end = ntohl(sack.end);
413 		if (SEQ_GT(sack.end, sack.start) &&
414 		    SEQ_GT(sack.start, tp->snd_una) &&
415 		    SEQ_GT(sack.start, th->th_ack) &&
416 		    SEQ_LEQ(sack.end, tp->snd_max))
417 			sack_blocks[num_sack_blks++] = sack;
418 		tmp_olen -= TCPOLEN_SACK;
419 		tmp_cp += TCPOLEN_SACK;
420 	}
421 	if (num_sack_blks == 0)
422 		return 0;
423 	/* Bubble sort */
424 	for (i = 0; i < num_sack_blks; i++) {
425 		for (j = i + 1; j < num_sack_blks; j++) {
426 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
427 				sack = sack_blocks[i];
428 				sack_blocks[i] = sack_blocks[j];
429 				sack_blocks[j] = sack;
430 			}
431 		}
432 	}
433 	if (TAILQ_EMPTY(&tp->snd_holes))
434 		/*
435 		 * Empty scoreboard. Need to initialize snd_fack (it may be
436 		 * uninitialized or have a bogus value). Scoreboard holes
437 		 * (from the sack blocks received) are created later below (in
438 		 * the logic that adds holes to the tail of the scoreboard).
439 		 */
440 		tp->snd_fack = tp->snd_una;
441 	/*
442 	 * In the while-loop below, incoming SACK blocks (sack_blocks[])
443 	 * and SACK holes (snd_holes) are traversed from their tails with
444 	 * just one pass in order to reduce the number of compares especially
445 	 * when the bandwidth-delay product is large.
446 	 * Note: Typically, in the first RTT of SACK recovery, the highest
447 	 * three or four SACK blocks with the same ack number are received.
448 	 * In the second RTT, if retransmitted data segments are not lost,
449 	 * the highest three or four SACK blocks with ack number advancing
450 	 * are received.
451 	 */
452 	sblkp = &sack_blocks[num_sack_blks - 1];	/* Last SACK block */
453 	if (SEQ_LT(tp->snd_fack, sblkp->start)) {
454 		/*
455 		 * The highest SACK block is beyond fack.
456 		 * Append new SACK hole at the tail.
457 		 * If the second or later highest SACK blocks are also
458 		 * beyond the current fack, they will be inserted by
459 		 * way of hole splitting in the while-loop below.
460 		 */
461 		tcp_sackhole_insert(tp, tp->snd_fack, sblkp->start, NULL);
462 		tp->snd_fack = sblkp->end;
463 		/* Go to the previous sack block. */
464 		sblkp--;
465 	} else if (SEQ_LT(tp->snd_fack, sblkp->end))
466 		/* fack is advanced. */
467 		tp->snd_fack = sblkp->end;
468 	cur = TAILQ_LAST(&tp->snd_holes, sackhole_head); /* Last SACK hole */
469 	/*
470 	 * Since the incoming sack blocks are sorted, we can process them
471 	 * making one sweep of the scoreboard.
472 	 */
473 	while (sblkp - sack_blocks >= 0) {
474 		KASSERT(cur != NULL, ("cur != NULL"));
475 		if (SEQ_GEQ(sblkp->start, cur->end)) {
476 			/*
477 			 * SACKs data beyond the current hole.
478 			 * Go to the previous sack block.
479 			 */
480 			sblkp--;
481 			continue;
482 		}
483 		if (SEQ_LEQ(sblkp->end, cur->start)) {
484 			/*
485 			 * SACKs data before the current hole.
486 			 * Go to the previous hole.
487 			 */
488 			cur = TAILQ_PREV(cur, sackhole_head, scblink);
489 			continue;
490 		}
491 		tp->sackhint.sack_bytes_rexmit -= (cur->rxmit - cur->start);
492 		KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
493 			("sackhint bytes rtx >= 0"));
494 		if (SEQ_LEQ(sblkp->start, cur->start)) {
495 			/* Data acks at least the beginning of hole */
496 			if (SEQ_GEQ(sblkp->end, cur->end)) {
497 				/* Acks entire hole, so delete hole */
498 				temp = cur;
499 				cur = TAILQ_PREV(cur, sackhole_head, scblink);
500 				tcp_sackhole_remove(tp, temp);
501 				/*
502 				 * The sack block may ack all or part of the next
503 				 * hole too, so continue onto the next hole.
504 				 */
505 				continue;
506 			} else {
507 				/* Move start of hole forward */
508 				cur->start = sblkp->end;
509 				cur->rxmit = SEQ_MAX(cur->rxmit, cur->start);
510 			}
511 			/* Go to the previous hole. */
512 			cur = TAILQ_PREV(cur, sackhole_head, scblink);
513 		} else {
514 			/* Data acks at least the end of hole */
515 			if (SEQ_GEQ(sblkp->end, cur->end)) {
516 				/* Move end of hole backward */
517 				cur->end = sblkp->start;
518 				cur->rxmit = SEQ_MIN(cur->rxmit, cur->end);
519 			} else {
520 				/*
521 				 * ACKs some data in middle of a hole; need to
522 				 * split current hole
523 				 */
524 				temp = tcp_sackhole_insert(tp, sblkp->end,
525 							   cur->end, cur);
526 				if (temp != NULL) {
527 					if (SEQ_GT(cur->rxmit, temp->rxmit)) {
528 						temp->rxmit = cur->rxmit;
529 						tp->sackhint.sack_bytes_rexmit
530 							+= (temp->rxmit
531 							    - temp->start);
532 					}
533 					cur->end = sblkp->start;
534 					cur->rxmit = SEQ_MIN(cur->rxmit,
535 							     cur->end);
536 				}
537 			}
538 			/* Go to the previous sack block. */
539 			sblkp--;
540 		}
541 		tp->sackhint.sack_bytes_rexmit += (cur->rxmit - cur->start);
542 	}
543 	return (0);
544 }
545 
546 /*
547  * Delete stale (i.e, cumulatively ack'd) holes.  Hole is deleted only if
548  * it is completely acked; otherwise, tcp_sack_option(), called from
549  * tcp_dooptions(), will fix up the hole.
550  */
551 void
552 tcp_del_sackholes(tp, th)
553 	struct tcpcb *tp;
554 	struct tcphdr *th;
555 {
556 	INP_LOCK_ASSERT(tp->t_inpcb);
557 	if (tp->sack_enable && tp->t_state != TCPS_LISTEN) {
558 		/* max because this could be an older ack just arrived */
559 		tcp_seq lastack = SEQ_GT(th->th_ack, tp->snd_una) ?
560 			th->th_ack : tp->snd_una;
561 		struct sackhole *cur = TAILQ_FIRST(&tp->snd_holes);
562 		struct sackhole *prev;
563 		while (cur)
564 			if (SEQ_LEQ(cur->end, lastack)) {
565 				prev = cur;
566 				cur = TAILQ_NEXT(cur, scblink);
567 				tp->sackhint.sack_bytes_rexmit -=
568 					(prev->rxmit - prev->start);
569 				tcp_sackhole_remove(tp, prev);
570 			} else if (SEQ_LT(cur->start, lastack)) {
571 				if (SEQ_LT(cur->rxmit, lastack)) {
572 					tp->sackhint.sack_bytes_rexmit -=
573 					    (cur->rxmit - cur->start);
574 					cur->rxmit = lastack;
575 				} else
576 					tp->sackhint.sack_bytes_rexmit -=
577 					    (lastack - cur->start);
578 				cur->start = lastack;
579 				break;
580 			} else
581 				break;
582 	}
583 }
584 
585 void
586 tcp_free_sackholes(struct tcpcb *tp)
587 {
588 	struct sackhole *q;
589 
590 	INP_LOCK_ASSERT(tp->t_inpcb);
591 	while ((q = TAILQ_FIRST(&tp->snd_holes)) != NULL)
592 		tcp_sackhole_remove(tp, q);
593 	tp->sackhint.sack_bytes_rexmit = 0;
594 
595 	KASSERT(tp->snd_numholes == 0, ("tp->snd_numholes == 0"));
596 	KASSERT(tp->sackhint.nexthole == NULL,
597 		("tp->sackhint.nexthole == NULL"));
598 }
599 
600 /*
601  * Partial ack handling within a sack recovery episode.
602  * Keeping this very simple for now. When a partial ack
603  * is received, force snd_cwnd to a value that will allow
604  * the sender to transmit no more than 2 segments.
605  * If necessary, a better scheme can be adopted at a
606  * later point, but for now, the goal is to prevent the
607  * sender from bursting a large amount of data in the midst
608  * of sack recovery.
609  */
610 void
611 tcp_sack_partialack(tp, th)
612 	struct tcpcb *tp;
613 	struct tcphdr *th;
614 {
615 	int num_segs = 1;
616 
617 	INP_LOCK_ASSERT(tp->t_inpcb);
618 	callout_stop(tp->tt_rexmt);
619 	tp->t_rtttime = 0;
620 	/* send one or 2 segments based on how much new data was acked */
621 	if (((th->th_ack - tp->snd_una) / tp->t_maxseg) > 2)
622 		num_segs = 2;
623 	tp->snd_cwnd = (tp->sackhint.sack_bytes_rexmit +
624 		(tp->snd_nxt - tp->sack_newdata) +
625 		num_segs * tp->t_maxseg);
626 	if (tp->snd_cwnd > tp->snd_ssthresh)
627 		tp->snd_cwnd = tp->snd_ssthresh;
628 	tp->t_flags |= TF_ACKNOW;
629 	(void) tcp_output(tp);
630 }
631 
632 /*
633  * Debug version of tcp_sack_output() that walks the scoreboard. Used for
634  * now to sanity check the hint.
635  */
636 static struct sackhole *
637 tcp_sack_output_debug(struct tcpcb *tp, int *sack_bytes_rexmt)
638 {
639 	struct sackhole *p;
640 
641 	INP_LOCK_ASSERT(tp->t_inpcb);
642 	*sack_bytes_rexmt = 0;
643 	TAILQ_FOREACH(p, &tp->snd_holes, scblink) {
644 		if (SEQ_LT(p->rxmit, p->end)) {
645 			if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */
646 				continue;
647 			}
648 			*sack_bytes_rexmt += (p->rxmit - p->start);
649 			break;
650 		}
651 		*sack_bytes_rexmt += (p->rxmit - p->start);
652 	}
653 	return (p);
654 }
655 
656 /*
657  * Returns the next hole to retransmit and the number of retransmitted bytes
658  * from the scoreboard. We store both the next hole and the number of
659  * retransmitted bytes as hints (and recompute these on the fly upon SACK/ACK
660  * reception). This avoids scoreboard traversals completely.
661  *
662  * The loop here will traverse *at most* one link. Here's the argument.
663  * For the loop to traverse more than 1 link before finding the next hole to
664  * retransmit, we would need to have at least 1 node following the current hint
665  * with (rxmit == end). But, for all holes following the current hint,
666  * (start == rxmit), since we have not yet retransmitted from them. Therefore,
667  * in order to traverse more 1 link in the loop below, we need to have at least
668  * one node following the current hint with (start == rxmit == end).
669  * But that can't happen, (start == end) means that all the data in that hole
670  * has been sacked, in which case, the hole would have been removed from the
671  * scoreboard.
672  */
673 struct sackhole *
674 tcp_sack_output(struct tcpcb *tp, int *sack_bytes_rexmt)
675 {
676 	struct sackhole *hole = NULL, *dbg_hole = NULL;
677 	int dbg_bytes_rexmt;
678 
679 	INP_LOCK_ASSERT(tp->t_inpcb);
680 	dbg_hole = tcp_sack_output_debug(tp, &dbg_bytes_rexmt);
681 	*sack_bytes_rexmt = tp->sackhint.sack_bytes_rexmit;
682 	hole = tp->sackhint.nexthole;
683 	if (hole == NULL || SEQ_LT(hole->rxmit, hole->end))
684 		goto out;
685 	while ((hole = TAILQ_NEXT(hole, scblink)) != NULL) {
686 		if (SEQ_LT(hole->rxmit, hole->end)) {
687 			tp->sackhint.nexthole = hole;
688 			break;
689 		}
690 	}
691 out:
692 	if (dbg_hole != hole) {
693 		printf("%s: Computed sack hole not the same as cached value\n", __func__);
694 		hole = dbg_hole;
695 	}
696 	if (*sack_bytes_rexmt != dbg_bytes_rexmt) {
697 		printf("%s: Computed sack_bytes_retransmitted (%d) not "
698 		       "the same as cached value (%d)\n",
699 		       __func__, dbg_bytes_rexmt, *sack_bytes_rexmt);
700 		*sack_bytes_rexmt = dbg_bytes_rexmt;
701 	}
702 	return (hole);
703 }
704 
705 /*
706  * After a timeout, the SACK list may be rebuilt.  This SACK information
707  * should be used to avoid retransmitting SACKed data.  This function
708  * traverses the SACK list to see if snd_nxt should be moved forward.
709  */
710 void
711 tcp_sack_adjust(struct tcpcb *tp)
712 {
713 	struct sackhole *p, *cur = TAILQ_FIRST(&tp->snd_holes);
714 
715 	INP_LOCK_ASSERT(tp->t_inpcb);
716 	if (cur == NULL)
717 		return; /* No holes */
718 	if (SEQ_GEQ(tp->snd_nxt, tp->snd_fack))
719 		return; /* We're already beyond any SACKed blocks */
720 	/*
721 	 * Two cases for which we want to advance snd_nxt:
722 	 * i) snd_nxt lies between end of one hole and beginning of another
723 	 * ii) snd_nxt lies between end of last hole and snd_fack
724 	 */
725 	while ((p = TAILQ_NEXT(cur, scblink)) != NULL) {
726 		if (SEQ_LT(tp->snd_nxt, cur->end))
727 			return;
728 		if (SEQ_GEQ(tp->snd_nxt, p->start))
729 			cur = p;
730 		else {
731 			tp->snd_nxt = p->start;
732 			return;
733 		}
734 	}
735 	if (SEQ_LT(tp->snd_nxt, cur->end))
736 		return;
737 	tp->snd_nxt = tp->snd_fack;
738 	return;
739 }
740