• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*-
2  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995
3  *	The Regents of the University of California.
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  * 4. Neither the name of the University nor the names of its contributors
15  *    may be used to endorse or promote products derived from this software
16  *    without specific prior written permission.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  *
30  *	@(#)tcp_sack.c	8.12 (Berkeley) 5/24/95
31  */
32 
33 /*-
34  *	@@(#)COPYRIGHT	1.1 (NRL) 17 January 1995
35  *
36  * NRL grants permission for redistribution and use in source and binary
37  * forms, with or without modification, of the software and documentation
38  * created at NRL provided that the following conditions are met:
39  *
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. All advertising materials mentioning features or use of this software
46  *    must display the following acknowledgements:
47  *	This product includes software developed by the University of
48  *	California, Berkeley and its contributors.
49  *	This product includes software developed at the Information
50  *	Technology Division, US Naval Research Laboratory.
51  * 4. Neither the name of the NRL nor the names of its contributors
52  *    may be used to endorse or promote products derived from this software
53  *    without specific prior written permission.
54  *
55  * THE SOFTWARE PROVIDED BY NRL IS PROVIDED BY NRL AND CONTRIBUTORS ``AS
56  * IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
57  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A
58  * PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL NRL OR
59  * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
60  * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
61  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
62  * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
63  * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
64  * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
65  * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
66  *
67  * The views and conclusions contained in the software and documentation
68  * are those of the authors and should not be interpreted as representing
69  * official policies, either expressed or implied, of the US Naval
70  * Research Laboratory (NRL).
71  */
72 
73 /* samkumar: Removed a bunch of #include's and VNET declarations. */
74 
75 #include <strings.h>
76 #include "tcp.h"
77 #include "tcp_fsm.h"
78 #include "tcp_seq.h"
79 #include "tcp_timer.h"
80 #include "tcp_var.h"
81 #include "sys/queue.h"
82 
83 enum tcp_sack_consts {
84 	V_tcp_sack_maxholes = MAX_SACKHOLES
85 };
86 
87 /*
88  * samkumar: Removed tcp_sack_globalmaxholes and tcp_sack_globalholes.
89  * There used to be a counter, V_tcp_sack_globalholes, that kept track of the
90  * total number of SACK holes allocated across all TCP connections.
91  */
92 
93 /*
94  * samkumar: I added these three functions. The first, tcp_sack_init,
95  * initializes a per-connection pool of SACK holes.
96  *
97  * The next two, sackhole_alloc and sackhole_free, allocate and deallocate SACK
98  * holes from the pool. Previously, the FreeBSD code would allocate SACK holes
99  * dynamically, for example, using the code
100  * "hole = (struct sackhole *)uma_zalloc(V_sack_hole_zone, M_NOWAIT);".
101  * TCPlp avoids dynamic memory allocation in the TCP implementation, so we
102  * replace it with this per-connection pool.
103  */
104 
105 void
tcp_sack_init(struct tcpcb * tp)106 tcp_sack_init(struct tcpcb* tp)
107 {
108 	bmp_init(tp->sackhole_bmp, SACKHOLE_BMP_SIZE);
109 }
110 
sackhole_alloc(struct tcpcb * tp)111 struct sackhole* sackhole_alloc(struct tcpcb* tp) {
112     size_t freeindex = bmp_countset(tp->sackhole_bmp, SACKHOLE_BMP_SIZE, 0, SACKHOLE_BMP_SIZE);
113     if (freeindex >= SACKHOLE_BMP_SIZE) {
114     	return NULL; // all sackholes are allocated already!
115     }
116     bmp_setrange(tp->sackhole_bmp, freeindex, 1);
117     return &tp->sackhole_pool[freeindex];
118 }
119 
sackhole_free(struct tcpcb * tp,struct sackhole * tofree)120 void sackhole_free(struct tcpcb* tp, struct sackhole* tofree) {
121 	size_t freeindex = (size_t) (tofree - &tp->sackhole_pool[0]);
122 	KASSERT(tofree == &tp->sackhole_pool[freeindex], ("sackhole pool unaligned"));
123 	bmp_clrrange(tp->sackhole_bmp, freeindex, 1);
124 }
125 
126 /*
127  * samkumar: Throughout the remaining functions, I have replaced allocation and
128  * deallocation of SACK holes, which previously used uma_zalloc and uma_zfree,
129  * with calls to sackhole_alloc and sackhole_free. I've also removed code for
130  * locking, global stats collection, global SACK hole limits, and debugging
131  * probes.
132  */
133 
134 
135 /*
136  * This function is called upon receipt of new valid data (while not in
137  * header prediction mode), and it updates the ordered list of sacks.
138  */
139 void
tcp_update_sack_list(struct tcpcb * tp,tcp_seq rcv_start,tcp_seq rcv_end)140 tcp_update_sack_list(struct tcpcb *tp, tcp_seq rcv_start, tcp_seq rcv_end)
141 {
142 	/*
143 	 * First reported block MUST be the most recent one.  Subsequent
144 	 * blocks SHOULD be in the order in which they arrived at the
145 	 * receiver.  These two conditions make the implementation fully
146 	 * compliant with RFC 2018.
147 	 */
148 	struct sackblk head_blk, saved_blks[MAX_SACK_BLKS];
149 	int num_head, num_saved, i;
150 
151 	/* Check arguments. */
152 	KASSERT(SEQ_LT(rcv_start, rcv_end), ("rcv_start < rcv_end"));
153 
154 	/* SACK block for the received segment. */
155 	head_blk.start = rcv_start;
156 	head_blk.end = rcv_end;
157 
158 	/*
159 	 * Merge updated SACK blocks into head_blk, and save unchanged SACK
160 	 * blocks into saved_blks[].  num_saved will have the number of the
161 	 * saved SACK blocks.
162 	 */
163 	num_saved = 0;
164 	for (i = 0; i < tp->rcv_numsacks; i++) {
165 		tcp_seq start = tp->sackblks[i].start;
166 		tcp_seq end = tp->sackblks[i].end;
167 		if (SEQ_GEQ(start, end) || SEQ_LEQ(start, tp->rcv_nxt)) {
168 			/*
169 			 * Discard this SACK block.
170 			 */
171 		} else if (SEQ_LEQ(head_blk.start, end) &&
172 			   SEQ_GEQ(head_blk.end, start)) {
173 			/*
174 			 * Merge this SACK block into head_blk.  This SACK
175 			 * block itself will be discarded.
176 			 */
177 			if (SEQ_GT(head_blk.start, start))
178 				head_blk.start = start;
179 			if (SEQ_LT(head_blk.end, end))
180 				head_blk.end = end;
181 		} else {
182 			/*
183 			 * Save this SACK block.
184 			 */
185 			saved_blks[num_saved].start = start;
186 			saved_blks[num_saved].end = end;
187 			num_saved++;
188 		}
189 	}
190 
191 	/*
192 	 * Update SACK list in tp->sackblks[].
193 	 */
194 	num_head = 0;
195 	if (SEQ_GT(head_blk.start, tp->rcv_nxt)) {
196 		/*
197 		 * The received data segment is an out-of-order segment.  Put
198 		 * head_blk at the top of SACK list.
199 		 */
200 		tp->sackblks[0] = head_blk;
201 		num_head = 1;
202 		/*
203 		 * If the number of saved SACK blocks exceeds its limit,
204 		 * discard the last SACK block.
205 		 */
206 		if (num_saved >= MAX_SACK_BLKS)
207 			num_saved--;
208 	}
209 	if (num_saved > 0) {
210 		/*
211 		 * Copy the saved SACK blocks back.
212 		 */
213 		bcopy(saved_blks, &tp->sackblks[num_head],
214 		      sizeof(struct sackblk) * num_saved);
215 	}
216 
217 	/* Save the number of SACK blocks. */
218 	tp->rcv_numsacks = num_head + num_saved;
219 }
220 
221 /*
222  * Delete all receiver-side SACK information.
223  */
224 void
tcp_clean_sackreport(struct tcpcb * tp)225 tcp_clean_sackreport(struct tcpcb *tp)
226 {
227 	int i;
228 
229 	tp->rcv_numsacks = 0;
230 	for (i = 0; i < MAX_SACK_BLKS; i++)
231 		tp->sackblks[i].start = tp->sackblks[i].end=0;
232 }
233 
234 /*
235  * Allocate struct sackhole.
236  */
237 static struct sackhole *
tcp_sackhole_alloc(struct tcpcb * tp,tcp_seq start,tcp_seq end)238 tcp_sackhole_alloc(struct tcpcb *tp, tcp_seq start, tcp_seq end)
239 {
240 	struct sackhole *hole;
241 
242 	/*
243 	 * samkumar: This if block also used to also return NULL if
244 	 * V_tcp_sack_globalholes >= V_tcp_sack_globalmaxholes
245 	 * but I removed that check since it doesn't make sense to enforce a global
246 	 * limit on SACK holes when we have a fixed-size pool (moreover, a separate
247 	 * pool per connection). The per-connection limit is sufficient.
248 	 */
249 	if (tp->snd_numholes >= V_tcp_sack_maxholes) {
250 		return NULL;
251 	}
252 
253 	hole = sackhole_alloc(tp);
254 	if (hole == NULL)
255 		return NULL;
256 
257 	hole->start = start;
258 	hole->end = end;
259 	hole->rxmit = start;
260 
261 	tp->snd_numholes++;
262 
263 	return hole;
264 }
265 
266 /*
267  * Free struct sackhole.
268  */
269 static void
tcp_sackhole_free(struct tcpcb * tp,struct sackhole * hole)270 tcp_sackhole_free(struct tcpcb *tp, struct sackhole *hole)
271 {
272 	sackhole_free(tp, hole);
273 
274 	tp->snd_numholes--;
275 
276 	KASSERT(tp->snd_numholes >= 0, ("tp->snd_numholes >= 0"));
277 }
278 
279 /*
280  * Insert new SACK hole into scoreboard.
281  */
282 static struct sackhole *
tcp_sackhole_insert(struct tcpcb * tp,tcp_seq start,tcp_seq end,struct sackhole * after)283 tcp_sackhole_insert(struct tcpcb *tp, tcp_seq start, tcp_seq end,
284     struct sackhole *after)
285 {
286 	struct sackhole *hole;
287 
288 	/* Allocate a new SACK hole. */
289 	hole = tcp_sackhole_alloc(tp, start, end);
290 	if (hole == NULL)
291 		return NULL;
292 
293 	/* Insert the new SACK hole into scoreboard. */
294 	if (after != NULL)
295 		TAILQ_INSERT_AFTER(&tp->snd_holes, after, hole, scblink);
296 	else
297 		TAILQ_INSERT_TAIL(&tp->snd_holes, hole, scblink);
298 
299 	/* Update SACK hint. */
300 	if (tp->sackhint.nexthole == NULL)
301 		tp->sackhint.nexthole = hole;
302 
303 	return hole;
304 }
305 
306 /*
307  * Remove SACK hole from scoreboard.
308  */
309 static void
tcp_sackhole_remove(struct tcpcb * tp,struct sackhole * hole)310 tcp_sackhole_remove(struct tcpcb *tp, struct sackhole *hole)
311 {
312 
313 	/* Update SACK hint. */
314 	if (tp->sackhint.nexthole == hole)
315 		tp->sackhint.nexthole = TAILQ_NEXT(hole, scblink);
316 
317 	/* Remove this SACK hole. */
318 	TAILQ_REMOVE(&tp->snd_holes, hole, scblink);
319 
320 	/* Free this SACK hole. */
321 	tcp_sackhole_free(tp, hole);
322 }
323 
324 /*
325  * Process cumulative ACK and the TCP SACK option to update the scoreboard.
326  * tp->snd_holes is an ordered list of holes (oldest to newest, in terms of
327  * the sequence space).
328  */
329 void
tcp_sack_doack(struct tcpcb * tp,struct tcpopt * to,tcp_seq th_ack)330 tcp_sack_doack(struct tcpcb *tp, struct tcpopt *to, tcp_seq th_ack)
331 {
332 	struct sackhole *cur, *temp;
333 	struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1], *sblkp;
334 	int i, j, num_sack_blks;
335 
336 	num_sack_blks = 0;
337 	/*
338 	 * If SND.UNA will be advanced by SEG.ACK, and if SACK holes exist,
339 	 * treat [SND.UNA, SEG.ACK) as if it is a SACK block.
340 	 */
341 	if (SEQ_LT(tp->snd_una, th_ack) && !TAILQ_EMPTY(&tp->snd_holes)) {
342 		sack_blocks[num_sack_blks].start = tp->snd_una;
343 		sack_blocks[num_sack_blks++].end = th_ack;
344 	}
345 	/*
346 	 * Append received valid SACK blocks to sack_blocks[], but only if we
347 	 * received new blocks from the other side.
348 	 */
349 	if (to->to_flags & TOF_SACK) {
350 		for (i = 0; i < to->to_nsacks; i++) {
351 			bcopy((to->to_sacks + i * TCPOLEN_SACK),
352 			    &sack, sizeof(sack));
353 			sack.start = ntohl(sack.start);
354 			sack.end = ntohl(sack.end);
355 			if (SEQ_GT(sack.end, sack.start) &&
356 			    SEQ_GT(sack.start, tp->snd_una) &&
357 			    SEQ_GT(sack.start, th_ack) &&
358 			    SEQ_LT(sack.start, tp->snd_max) &&
359 			    SEQ_GT(sack.end, tp->snd_una) &&
360 			    SEQ_LEQ(sack.end, tp->snd_max))
361 				sack_blocks[num_sack_blks++] = sack;
362 		}
363 	}
364 	/*
365 	 * Return if SND.UNA is not advanced and no valid SACK block is
366 	 * received.
367 	 */
368 	if (num_sack_blks == 0)
369 		return;
370 
371 	/*
372 	 * Sort the SACK blocks so we can update the scoreboard with just one
373 	 * pass. The overhead of sorting upto 4+1 elements is less than
374 	 * making upto 4+1 passes over the scoreboard.
375 	 */
376 	for (i = 0; i < num_sack_blks; i++) {
377 		for (j = i + 1; j < num_sack_blks; j++) {
378 			if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) {
379 				sack = sack_blocks[i];
380 				sack_blocks[i] = sack_blocks[j];
381 				sack_blocks[j] = sack;
382 			}
383 		}
384 	}
385 	if (TAILQ_EMPTY(&tp->snd_holes))
386 		/*
387 		 * Empty scoreboard. Need to initialize snd_fack (it may be
388 		 * uninitialized or have a bogus value). Scoreboard holes
389 		 * (from the sack blocks received) are created later below
390 		 * (in the logic that adds holes to the tail of the
391 		 * scoreboard).
392 		 */
393 		tp->snd_fack = SEQ_MAX(tp->snd_una, th_ack);
394 	/*
395 	 * In the while-loop below, incoming SACK blocks (sack_blocks[]) and
396 	 * SACK holes (snd_holes) are traversed from their tails with just
397 	 * one pass in order to reduce the number of compares especially when
398 	 * the bandwidth-delay product is large.
399 	 *
400 	 * Note: Typically, in the first RTT of SACK recovery, the highest
401 	 * three or four SACK blocks with the same ack number are received.
402 	 * In the second RTT, if retransmitted data segments are not lost,
403 	 * the highest three or four SACK blocks with ack number advancing
404 	 * are received.
405 	 */
406 	sblkp = &sack_blocks[num_sack_blks - 1];	/* Last SACK block */
407 	tp->sackhint.last_sack_ack = sblkp->end;
408 	if (SEQ_LT(tp->snd_fack, sblkp->start)) {
409 		/*
410 		 * The highest SACK block is beyond fack.  Append new SACK
411 		 * hole at the tail.  If the second or later highest SACK
412 		 * blocks are also beyond the current fack, they will be
413 		 * inserted by way of hole splitting in the while-loop below.
414 		 */
415 		temp = tcp_sackhole_insert(tp, tp->snd_fack,sblkp->start,NULL);
416 		if (temp != NULL) {
417 			tp->snd_fack = sblkp->end;
418 			/* Go to the previous sack block. */
419 			sblkp--;
420 		} else {
421 			/*
422 			 * We failed to add a new hole based on the current
423 			 * sack block.  Skip over all the sack blocks that
424 			 * fall completely to the right of snd_fack and
425 			 * proceed to trim the scoreboard based on the
426 			 * remaining sack blocks.  This also trims the
427 			 * scoreboard for th_ack (which is sack_blocks[0]).
428 			 */
429 			while (sblkp >= sack_blocks &&
430 			       SEQ_LT(tp->snd_fack, sblkp->start))
431 				sblkp--;
432 			if (sblkp >= sack_blocks &&
433 			    SEQ_LT(tp->snd_fack, sblkp->end))
434 				tp->snd_fack = sblkp->end;
435 		}
436 	} else if (SEQ_LT(tp->snd_fack, sblkp->end))
437 		/* fack is advanced. */
438 		tp->snd_fack = sblkp->end;
439 	/* We must have at least one SACK hole in scoreboard. */
440 	KASSERT(!TAILQ_EMPTY(&tp->snd_holes),
441 	    ("SACK scoreboard must not be empty"));
442 	cur = TAILQ_LAST(&tp->snd_holes, sackhole_head); /* Last SACK hole. */
443 	/*
444 	 * Since the incoming sack blocks are sorted, we can process them
445 	 * making one sweep of the scoreboard.
446 	 */
447 	while (sblkp >= sack_blocks  && cur != NULL) {
448 		if (SEQ_GEQ(sblkp->start, cur->end)) {
449 			/*
450 			 * SACKs data beyond the current hole.  Go to the
451 			 * previous sack block.
452 			 */
453 			sblkp--;
454 			continue;
455 		}
456 		if (SEQ_LEQ(sblkp->end, cur->start)) {
457 			/*
458 			 * SACKs data before the current hole.  Go to the
459 			 * previous hole.
460 			 */
461 			cur = TAILQ_PREV(cur, sackhole_head, scblink);
462 			continue;
463 		}
464 		tp->sackhint.sack_bytes_rexmit -= (cur->rxmit - cur->start);
465 		KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
466 		    ("sackhint bytes rtx >= 0"));
467 		if (SEQ_LEQ(sblkp->start, cur->start)) {
468 			/* Data acks at least the beginning of hole. */
469 			if (SEQ_GEQ(sblkp->end, cur->end)) {
470 				/* Acks entire hole, so delete hole. */
471 				temp = cur;
472 				cur = TAILQ_PREV(cur, sackhole_head, scblink);
473 				tcp_sackhole_remove(tp, temp);
474 				/*
475 				 * The sack block may ack all or part of the
476 				 * next hole too, so continue onto the next
477 				 * hole.
478 				 */
479 				continue;
480 			} else {
481 				/* Move start of hole forward. */
482 				cur->start = sblkp->end;
483 				cur->rxmit = SEQ_MAX(cur->rxmit, cur->start);
484 			}
485 		} else {
486 			/* Data acks at least the end of hole. */
487 			if (SEQ_GEQ(sblkp->end, cur->end)) {
488 				/* Move end of hole backward. */
489 				cur->end = sblkp->start;
490 				cur->rxmit = SEQ_MIN(cur->rxmit, cur->end);
491 			} else {
492 				/*
493 				 * ACKs some data in middle of a hole; need
494 				 * to split current hole
495 				 */
496 				temp = tcp_sackhole_insert(tp, sblkp->end,
497 				    cur->end, cur);
498 				if (temp != NULL) {
499 					if (SEQ_GT(cur->rxmit, temp->rxmit)) {
500 						temp->rxmit = cur->rxmit;
501 						tp->sackhint.sack_bytes_rexmit
502 						    += (temp->rxmit
503 						    - temp->start);
504 					}
505 					cur->end = sblkp->start;
506 					cur->rxmit = SEQ_MIN(cur->rxmit,
507 					    cur->end);
508 				}
509 			}
510 		}
511 		tp->sackhint.sack_bytes_rexmit += (cur->rxmit - cur->start);
512 		/*
513 		 * Testing sblkp->start against cur->start tells us whether
514 		 * we're done with the sack block or the sack hole.
515 		 * Accordingly, we advance one or the other.
516 		 */
517 		if (SEQ_LEQ(sblkp->start, cur->start))
518 			cur = TAILQ_PREV(cur, sackhole_head, scblink);
519 		else
520 			sblkp--;
521 	}
522 }
523 
524 /*
525  * Free all SACK holes to clear the scoreboard.
526  */
527 void
tcp_free_sackholes(struct tcpcb * tp)528 tcp_free_sackholes(struct tcpcb *tp)
529 {
530 	struct sackhole *q;
531 
532 	while ((q = TAILQ_FIRST(&tp->snd_holes)) != NULL)
533 		tcp_sackhole_remove(tp, q);
534 	tp->sackhint.sack_bytes_rexmit = 0;
535 
536 	KASSERT(tp->snd_numholes == 0, ("tp->snd_numholes == 0"));
537 	KASSERT(tp->sackhint.nexthole == NULL,
538 		("tp->sackhint.nexthole == NULL"));
539 }
540 
541 /*
542  * Partial ack handling within a sack recovery episode.  Keeping this very
543  * simple for now.  When a partial ack is received, force snd_cwnd to a value
544  * that will allow the sender to transmit no more than 2 segments.  If
545  * necessary, a better scheme can be adopted at a later point, but for now,
546  * the goal is to prevent the sender from bursting a large amount of data in
547  * the midst of sack recovery.
548  */
549 void
tcp_sack_partialack(struct tcpcb * tp,struct tcphdr * th)550 tcp_sack_partialack(struct tcpcb *tp, struct tcphdr *th)
551 {
552 	int num_segs = 1;
553 
554 	tcp_timer_activate(tp, TT_REXMT, 0);
555 	tp->t_rtttime = 0;
556 	/* Send one or 2 segments based on how much new data was acked. */
557 	if ((BYTES_THIS_ACK(tp, th) / tp->t_maxseg) >= 2)
558 		num_segs = 2;
559 	tp->snd_cwnd = (tp->sackhint.sack_bytes_rexmit +
560 	    (tp->snd_nxt - tp->sack_newdata) + num_segs * tp->t_maxseg);
561 	if (tp->snd_cwnd > tp->snd_ssthresh)
562 		tp->snd_cwnd = tp->snd_ssthresh;
563 	tp->t_flags |= TF_ACKNOW;
564 	(void) tcp_output(tp);
565 }
566 
567 /*
568  * samkumar: Removed this function for now, but I left it in as a comment
569  * (using #if 0) in case it is useful later for debugging.
570  */
571 #if 0
572 /*
573  * Debug version of tcp_sack_output() that walks the scoreboard.  Used for
574  * now to sanity check the hint.
575  */
576 static struct sackhole *
577 tcp_sack_output_debug(struct tcpcb *tp, int *sack_bytes_rexmt)
578 {
579 	struct sackhole *p;
580 
581 	INP_WLOCK_ASSERT(tp->t_inpcb);
582 	*sack_bytes_rexmt = 0;
583 	TAILQ_FOREACH(p, &tp->snd_holes, scblink) {
584 		if (SEQ_LT(p->rxmit, p->end)) {
585 			if (SEQ_LT(p->rxmit, tp->snd_una)) {/* old SACK hole */
586 				continue;
587 			}
588 			*sack_bytes_rexmt += (p->rxmit - p->start);
589 			break;
590 		}
591 		*sack_bytes_rexmt += (p->rxmit - p->start);
592 	}
593 	return (p);
594 }
595 #endif
596 
597 /*
598  * Returns the next hole to retransmit and the number of retransmitted bytes
599  * from the scoreboard.  We store both the next hole and the number of
600  * retransmitted bytes as hints (and recompute these on the fly upon SACK/ACK
601  * reception).  This avoids scoreboard traversals completely.
602  *
603  * The loop here will traverse *at most* one link.  Here's the argument.  For
604  * the loop to traverse more than 1 link before finding the next hole to
605  * retransmit, we would need to have at least 1 node following the current
606  * hint with (rxmit == end).  But, for all holes following the current hint,
607  * (start == rxmit), since we have not yet retransmitted from them.
608  * Therefore, in order to traverse more 1 link in the loop below, we need to
609  * have at least one node following the current hint with (start == rxmit ==
610  * end).  But that can't happen, (start == end) means that all the data in
611  * that hole has been sacked, in which case, the hole would have been removed
612  * from the scoreboard.
613  */
614 struct sackhole *
tcp_sack_output(struct tcpcb * tp,int * sack_bytes_rexmt)615 tcp_sack_output(struct tcpcb *tp, int *sack_bytes_rexmt)
616 {
617 	struct sackhole *hole = NULL;
618 
619 	*sack_bytes_rexmt = tp->sackhint.sack_bytes_rexmit;
620 	hole = tp->sackhint.nexthole;
621 	if (hole == NULL || SEQ_LT(hole->rxmit, hole->end))
622 		goto out;
623 	while ((hole = TAILQ_NEXT(hole, scblink)) != NULL) {
624 		if (SEQ_LT(hole->rxmit, hole->end)) {
625 			tp->sackhint.nexthole = hole;
626 			break;
627 		}
628 	}
629 out:
630 	return (hole);
631 }
632 
633 /*
634  * After a timeout, the SACK list may be rebuilt.  This SACK information
635  * should be used to avoid retransmitting SACKed data.  This function
636  * traverses the SACK list to see if snd_nxt should be moved forward.
637  */
638 void
tcp_sack_adjust(struct tcpcb * tp)639 tcp_sack_adjust(struct tcpcb *tp)
640 {
641 	struct sackhole *p, *cur = TAILQ_FIRST(&tp->snd_holes);
642 
643 	if (cur == NULL)
644 		return; /* No holes */
645 	if (SEQ_GEQ(tp->snd_nxt, tp->snd_fack))
646 		return; /* We're already beyond any SACKed blocks */
647 	/*-
648 	 * Two cases for which we want to advance snd_nxt:
649 	 * i) snd_nxt lies between end of one hole and beginning of another
650 	 * ii) snd_nxt lies between end of last hole and snd_fack
651 	 */
652 	while ((p = TAILQ_NEXT(cur, scblink)) != NULL) {
653 		if (SEQ_LT(tp->snd_nxt, cur->end))
654 			return;
655 		if (SEQ_GEQ(tp->snd_nxt, p->start))
656 			cur = p;
657 		else {
658 			tp->snd_nxt = p->start;
659 			return;
660 		}
661 	}
662 	if (SEQ_LT(tp->snd_nxt, cur->end))
663 		return;
664 	tp->snd_nxt = tp->snd_fack;
665 }
666