xref: /original-bsd/sys/kern/vfs_cluster.c (revision 6066c21e)
1 /*-
2  * Copyright (c) 1993
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * %sccs.include.redist.c%
6  *
7  *	@(#)vfs_cluster.c	8.7 (Berkeley) 02/13/94
8  */
9 
10 #include <sys/param.h>
11 #include <sys/proc.h>
12 #include <sys/buf.h>
13 #include <sys/vnode.h>
14 #include <sys/mount.h>
15 #include <sys/trace.h>
16 #include <sys/malloc.h>
17 #include <sys/resourcevar.h>
18 #include <libkern/libkern.h>
19 
20 #ifdef DEBUG
21 #include <vm/vm.h>
22 #include <sys/sysctl.h>
23 int doreallocblks = 1;
24 struct ctldebug debug13 = { "doreallocblks", &doreallocblks };
25 #else
26 /* XXX for cluster_write */
27 #define doreallocblks 1
28 #endif
29 
30 /*
31  * Local declarations
32  */
33 struct buf *cluster_newbuf __P((struct vnode *, struct buf *, long, daddr_t,
34 	    daddr_t, long, int));
35 struct buf *cluster_rbuild __P((struct vnode *, u_quad_t, struct buf *,
36 	    daddr_t, daddr_t, long, int, long));
37 void	    cluster_wbuild __P((struct vnode *, struct buf *, long,
38 	    daddr_t, int, daddr_t));
39 struct cluster_save *cluster_collectbufs __P((struct vnode *, struct buf *));
40 
41 #ifdef DIAGNOSTIC
42 /*
43  * Set to 1 if reads of block zero should cause readahead to be done.
44  * Set to 0 treats a read of block zero as a non-sequential read.
45  *
46  * Setting to one assumes that most reads of block zero of files are due to
47  * sequential passes over the files (e.g. cat, sum) where additional blocks
48  * will soon be needed.  Setting to zero assumes that the majority are
49  * surgical strikes to get particular info (e.g. size, file) where readahead
50  * blocks will not be used and, in fact, push out other potentially useful
51  * blocks from the cache.  The former seems intuitive, but some quick tests
52  * showed that the latter performed better from a system-wide point of view.
53  */
54 int	doclusterraz = 0;
55 #define ISSEQREAD(vp, blk) \
56 	(((blk) != 0 || doclusterraz) && \
57 	 ((blk) == (vp)->v_lastr + 1 || (blk) == (vp)->v_lastr))
58 #else
59 #define ISSEQREAD(vp, blk) \
60 	((blk) != 0 && ((blk) == (vp)->v_lastr + 1 || (blk) == (vp)->v_lastr))
61 #endif
62 
63 /*
64  * This replaces bread.  If this is a bread at the beginning of a file and
65  * lastr is 0, we assume this is the first read and we'll read up to two
66  * blocks if they are sequential.  After that, we'll do regular read ahead
67  * in clustered chunks.
68  *
69  * There are 4 or 5 cases depending on how you count:
70  *	Desired block is in the cache:
71  *	    1 Not sequential access (0 I/Os).
72  *	    2 Access is sequential, do read-ahead (1 ASYNC).
73  *	Desired block is not in cache:
74  *	    3 Not sequential access (1 SYNC).
75  *	    4 Sequential access, next block is contiguous (1 SYNC).
76  *	    5 Sequential access, next block is not contiguous (1 SYNC, 1 ASYNC)
77  *
78  * There are potentially two buffers that require I/O.
79  * 	bp is the block requested.
80  *	rbp is the read-ahead block.
81  *	If either is NULL, then you don't have to do the I/O.
82  */
83 cluster_read(vp, filesize, lblkno, size, cred, bpp)
84 	struct vnode *vp;
85 	u_quad_t filesize;
86 	daddr_t lblkno;
87 	long size;
88 	struct ucred *cred;
89 	struct buf **bpp;
90 {
91 	struct buf *bp, *rbp;
92 	daddr_t blkno, ioblkno;
93 	long flags;
94 	int error, num_ra, alreadyincore;
95 
96 #ifdef DIAGNOSTIC
97 	if (size == 0)
98 		panic("cluster_read: size = 0");
99 #endif
100 
101 	error = 0;
102 	flags = B_READ;
103 	*bpp = bp = getblk(vp, lblkno, size, 0, 0);
104 	if (bp->b_flags & B_CACHE) {
105 		/*
106 		 * Desired block is in cache; do any readahead ASYNC.
107 		 * Case 1, 2.
108 		 */
109 		trace(TR_BREADHIT, pack(vp, size), lblkno);
110 		flags |= B_ASYNC;
111 		ioblkno = lblkno + (vp->v_ralen ? vp->v_ralen : 1);
112 		alreadyincore = (int)incore(vp, ioblkno);
113 		bp = NULL;
114 	} else {
115 		/* Block wasn't in cache, case 3, 4, 5. */
116 		trace(TR_BREADMISS, pack(vp, size), lblkno);
117 		bp->b_flags |= B_READ;
118 		ioblkno = lblkno;
119 		alreadyincore = 0;
120 		curproc->p_stats->p_ru.ru_inblock++;		/* XXX */
121 	}
122 	/*
123 	 * XXX
124 	 * Replace 1 with a window size based on some permutation of
125 	 * maxcontig and rot_delay.  This will let you figure out how
126 	 * many blocks you should read-ahead (case 2, 4, 5).
127 	 *
128 	 * If the access isn't sequential, reset the window to 1.
129 	 * Note that a read to the same block is considered sequential.
130 	 * This catches the case where the file is being read sequentially,
131 	 * but at smaller than the filesystem block size.
132 	 */
133 	rbp = NULL;
134 	if (!ISSEQREAD(vp, lblkno)) {
135 		vp->v_ralen = 0;
136 		vp->v_maxra = lblkno;
137 	} else if ((ioblkno + 1) * size <= filesize && !alreadyincore &&
138 	    !(error = VOP_BMAP(vp, ioblkno, NULL, &blkno, &num_ra)) &&
139 	    blkno != -1) {
140 		/*
141 		 * Reading sequentially, and the next block is not in the
142 		 * cache.  We are going to try reading ahead.
143 		 */
144 		if (num_ra) {
145 			/*
146 			 * If our desired readahead block had been read
147 			 * in a previous readahead but is no longer in
148 			 * core, then we may be reading ahead too far
149 			 * or are not using our readahead very rapidly.
150 			 * In this case we scale back the window.
151 			 */
152 			if (!alreadyincore && ioblkno <= vp->v_maxra)
153 				vp->v_ralen = max(vp->v_ralen >> 1, 1);
154 			/*
155 			 * There are more sequential blocks than our current
156 			 * window allows, scale up.  Ideally we want to get
157 			 * in sync with the filesystem maxcontig value.
158 			 */
159 			else if (num_ra > vp->v_ralen && lblkno != vp->v_lastr)
160 				vp->v_ralen = vp->v_ralen ?
161 					min(num_ra, vp->v_ralen << 1) : 1;
162 
163 			if (num_ra > vp->v_ralen)
164 				num_ra = vp->v_ralen;
165 		}
166 
167 		if (num_ra)				/* case 2, 4 */
168 			rbp = cluster_rbuild(vp, filesize,
169 			    bp, ioblkno, blkno, size, num_ra, flags);
170 		else if (ioblkno == lblkno) {
171 			bp->b_blkno = blkno;
172 			/* Case 5: check how many blocks to read ahead */
173 			++ioblkno;
174 			if ((ioblkno + 1) * size > filesize ||
175 			    incore(vp, ioblkno) || (error = VOP_BMAP(vp,
176 			     ioblkno, NULL, &blkno, &num_ra)) || blkno == -1)
177 				goto skip_readahead;
178 			/*
179 			 * Adjust readahead as above
180 			 */
181 			if (num_ra) {
182 				if (!alreadyincore && ioblkno <= vp->v_maxra)
183 					vp->v_ralen = max(vp->v_ralen >> 1, 1);
184 				else if (num_ra > vp->v_ralen &&
185 					 lblkno != vp->v_lastr)
186 					vp->v_ralen = vp->v_ralen ?
187 						min(num_ra,vp->v_ralen<<1) : 1;
188 				if (num_ra > vp->v_ralen)
189 					num_ra = vp->v_ralen;
190 			}
191 			flags |= B_ASYNC;
192 			if (num_ra)
193 				rbp = cluster_rbuild(vp, filesize,
194 				    NULL, ioblkno, blkno, size, num_ra, flags);
195 			else {
196 				rbp = getblk(vp, ioblkno, size, 0, 0);
197 				rbp->b_flags |= flags;
198 				rbp->b_blkno = blkno;
199 			}
200 		} else {
201 			/* case 2; read ahead single block */
202 			rbp = getblk(vp, ioblkno, size, 0, 0);
203 			rbp->b_flags |= flags;
204 			rbp->b_blkno = blkno;
205 		}
206 
207 		if (rbp == bp)			/* case 4 */
208 			rbp = NULL;
209 		else if (rbp) {			/* case 2, 5 */
210 			trace(TR_BREADMISSRA,
211 			    pack(vp, (num_ra + 1) * size), ioblkno);
212 			curproc->p_stats->p_ru.ru_inblock++;	/* XXX */
213 		}
214 	}
215 
216 	/* XXX Kirk, do we need to make sure the bp has creds? */
217 skip_readahead:
218 	if (bp)
219 		if (bp->b_flags & (B_DONE | B_DELWRI))
220 			panic("cluster_read: DONE bp");
221 		else
222 			error = VOP_STRATEGY(bp);
223 
224 	if (rbp)
225 		if (error || rbp->b_flags & (B_DONE | B_DELWRI)) {
226 			rbp->b_flags &= ~(B_ASYNC | B_READ);
227 			brelse(rbp);
228 		} else
229 			(void) VOP_STRATEGY(rbp);
230 
231 	/*
232 	 * Recalculate our maximum readahead
233 	 */
234 	if (rbp == NULL)
235 		rbp = bp;
236 	if (rbp)
237 		vp->v_maxra = rbp->b_lblkno + (rbp->b_bufsize / size) - 1;
238 
239 	if (bp)
240 		return(biowait(bp));
241 	return(error);
242 }
243 
244 /*
245  * If blocks are contiguous on disk, use this to provide clustered
246  * read ahead.  We will read as many blocks as possible sequentially
247  * and then parcel them up into logical blocks in the buffer hash table.
248  */
249 struct buf *
250 cluster_rbuild(vp, filesize, bp, lbn, blkno, size, run, flags)
251 	struct vnode *vp;
252 	u_quad_t filesize;
253 	struct buf *bp;
254 	daddr_t lbn;
255 	daddr_t blkno;
256 	long size;
257 	int run;
258 	long flags;
259 {
260 	struct cluster_save *b_save;
261 	struct buf *tbp;
262 	daddr_t bn;
263 	int i, inc;
264 
265 #ifdef DIAGNOSTIC
266 	if (size != vp->v_mount->mnt_stat.f_iosize)
267 		panic("cluster_rbuild: size %d != filesize %d\n",
268 			size, vp->v_mount->mnt_stat.f_iosize);
269 #endif
270 	if (size * (lbn + run + 1) > filesize)
271 		--run;
272 	if (run == 0) {
273 		if (!bp) {
274 			bp = getblk(vp, lbn, size, 0, 0);
275 			bp->b_blkno = blkno;
276 			bp->b_flags |= flags;
277 		}
278 		return(bp);
279 	}
280 
281 	bp = cluster_newbuf(vp, bp, flags, blkno, lbn, size, run + 1);
282 	if (bp->b_flags & (B_DONE | B_DELWRI))
283 		return (bp);
284 
285 	b_save = malloc(sizeof(struct buf *) * run + sizeof(struct cluster_save),
286 	    M_SEGMENT, M_WAITOK);
287 	b_save->bs_bufsize = b_save->bs_bcount = size;
288 	b_save->bs_nchildren = 0;
289 	b_save->bs_children = (struct buf **)(b_save + 1);
290 	b_save->bs_saveaddr = bp->b_saveaddr;
291 	bp->b_saveaddr = (caddr_t) b_save;
292 
293 	inc = btodb(size);
294 	for (bn = blkno + inc, i = 1; i <= run; ++i, bn += inc) {
295 		if (incore(vp, lbn + i)) {
296 			if (i == 1) {
297 				bp->b_saveaddr = b_save->bs_saveaddr;
298 				bp->b_flags &= ~B_CALL;
299 				bp->b_iodone = NULL;
300 				allocbuf(bp, size);
301 				free(b_save, M_SEGMENT);
302 			} else
303 				allocbuf(bp, size * i);
304 			break;
305 		}
306 		tbp = getblk(vp, lbn + i, 0, 0, 0);
307 		/*
308 		 * getblk may return some memory in the buffer if there were
309 		 * no empty buffers to shed it to.  If there is currently
310 		 * memory in the buffer, we move it down size bytes to make
311 		 * room for the valid pages that cluster_callback will insert.
312 		 * We do this now so we don't have to do it at interrupt time
313 		 * in the callback routine.
314 		 */
315 		if (tbp->b_bufsize != 0) {
316 			caddr_t bdata = (char *)tbp->b_data;
317 
318 			if (tbp->b_bufsize + size > MAXBSIZE)
319 				panic("cluster_rbuild: too much memory");
320 			if (tbp->b_bufsize > size) {
321 				/*
322 				 * XXX if the source and destination regions
323 				 * overlap we have to copy backward to avoid
324 				 * clobbering any valid pages (i.e. pagemove
325 				 * implementations typically can't handle
326 				 * overlap).
327 				 */
328 				bdata += tbp->b_bufsize;
329 				while (bdata > (char *)tbp->b_data) {
330 					bdata -= CLBYTES;
331 					pagemove(bdata, bdata + size, CLBYTES);
332 				}
333 			} else
334 				pagemove(bdata, bdata + size, tbp->b_bufsize);
335 		}
336 		tbp->b_blkno = bn;
337 		tbp->b_flags |= flags | B_READ | B_ASYNC;
338 		++b_save->bs_nchildren;
339 		b_save->bs_children[i - 1] = tbp;
340 	}
341 	return(bp);
342 }
343 
344 /*
345  * Either get a new buffer or grow the existing one.
346  */
347 struct buf *
348 cluster_newbuf(vp, bp, flags, blkno, lblkno, size, run)
349 	struct vnode *vp;
350 	struct buf *bp;
351 	long flags;
352 	daddr_t blkno;
353 	daddr_t lblkno;
354 	long size;
355 	int run;
356 {
357 	if (!bp) {
358 		bp = getblk(vp, lblkno, size, 0, 0);
359 		if (bp->b_flags & (B_DONE | B_DELWRI)) {
360 			bp->b_blkno = blkno;
361 			return(bp);
362 		}
363 	}
364 	allocbuf(bp, run * size);
365 	bp->b_blkno = blkno;
366 	bp->b_iodone = cluster_callback;
367 	bp->b_flags |= flags | B_CALL;
368 	return(bp);
369 }
370 
371 /*
372  * Cleanup after a clustered read or write.
373  * This is complicated by the fact that any of the buffers might have
374  * extra memory (if there were no empty buffer headers at allocbuf time)
375  * that we will need to shift around.
376  */
377 void
378 cluster_callback(bp)
379 	struct buf *bp;
380 {
381 	struct cluster_save *b_save;
382 	struct buf **bpp, *tbp;
383 	long bsize;
384 	caddr_t cp;
385 	int error = 0;
386 
387 	/*
388 	 * Must propogate errors to all the components.
389 	 */
390 	if (bp->b_flags & B_ERROR)
391 		error = bp->b_error;
392 
393 	b_save = (struct cluster_save *)(bp->b_saveaddr);
394 	bp->b_saveaddr = b_save->bs_saveaddr;
395 
396 	bsize = b_save->bs_bufsize;
397 	cp = (char *)bp->b_data + bsize;
398 	/*
399 	 * Move memory from the large cluster buffer into the component
400 	 * buffers and mark IO as done on these.
401 	 */
402 	for (bpp = b_save->bs_children; b_save->bs_nchildren--; ++bpp) {
403 		tbp = *bpp;
404 		pagemove(cp, tbp->b_data, bsize);
405 		tbp->b_bufsize += bsize;
406 		tbp->b_bcount = bsize;
407 		if (error) {
408 			tbp->b_flags |= B_ERROR;
409 			tbp->b_error = error;
410 		}
411 		biodone(tbp);
412 		bp->b_bufsize -= bsize;
413 		cp += bsize;
414 	}
415 	/*
416 	 * If there was excess memory in the cluster buffer,
417 	 * slide it up adjacent to the remaining valid data.
418 	 */
419 	if (bp->b_bufsize != bsize) {
420 		if (bp->b_bufsize < bsize)
421 			panic("cluster_callback: too little memory");
422 		pagemove(cp, (char *)bp->b_data + bsize, bp->b_bufsize - bsize);
423 	}
424 	bp->b_bcount = bsize;
425 	bp->b_iodone = NULL;
426 	free(b_save, M_SEGMENT);
427 	if (bp->b_flags & B_ASYNC)
428 		brelse(bp);
429 	else {
430 		bp->b_flags &= ~B_WANTED;
431 		wakeup((caddr_t)bp);
432 	}
433 }
434 
435 /*
436  * Do clustered write for FFS.
437  *
438  * Three cases:
439  *	1. Write is not sequential (write asynchronously)
440  *	Write is sequential:
441  *	2.	beginning of cluster - begin cluster
442  *	3.	middle of a cluster - add to cluster
443  *	4.	end of a cluster - asynchronously write cluster
444  */
445 void
446 cluster_write(bp, filesize)
447         struct buf *bp;
448 	u_quad_t filesize;
449 {
450         struct vnode *vp;
451         daddr_t lbn;
452         int maxclen, cursize;
453 
454         vp = bp->b_vp;
455         lbn = bp->b_lblkno;
456 
457 	/* Initialize vnode to beginning of file. */
458 	if (lbn == 0)
459 		vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
460 
461         if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
462 	    (bp->b_blkno != vp->v_lasta + btodb(bp->b_bcount))) {
463 		maxclen = MAXBSIZE / vp->v_mount->mnt_stat.f_iosize - 1;
464 		if (vp->v_clen != 0) {
465 			/*
466 			 * Next block is not sequential.
467 			 *
468 			 * If we are not writing at end of file, the process
469 			 * seeked to another point in the file since its
470 			 * last write, or we have reached our maximum
471 			 * cluster size, then push the previous cluster.
472 			 * Otherwise try reallocating to make it sequential.
473 			 */
474 			cursize = vp->v_lastw - vp->v_cstart + 1;
475 			if (!doreallocblks ||
476 			    (lbn + 1) * bp->b_bcount != filesize ||
477 			    lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
478 				cluster_wbuild(vp, NULL, bp->b_bcount,
479 				    vp->v_cstart, cursize, lbn);
480 			} else {
481 				struct buf **bpp, **endbp;
482 				struct cluster_save *buflist;
483 
484 				buflist = cluster_collectbufs(vp, bp);
485 				endbp = &buflist->bs_children
486 				    [buflist->bs_nchildren - 1];
487 				if (VOP_REALLOCBLKS(vp, buflist)) {
488 					/*
489 					 * Failed, push the previous cluster.
490 					 */
491 					for (bpp = buflist->bs_children;
492 					     bpp < endbp; bpp++)
493 						brelse(*bpp);
494 					free(buflist, M_SEGMENT);
495 					cluster_wbuild(vp, NULL, bp->b_bcount,
496 					    vp->v_cstart, cursize, lbn);
497 				} else {
498 					/*
499 					 * Succeeded, keep building cluster.
500 					 */
501 					for (bpp = buflist->bs_children;
502 					     bpp <= endbp; bpp++)
503 						bdwrite(*bpp);
504 					free(buflist, M_SEGMENT);
505 					vp->v_lastw = lbn;
506 					vp->v_lasta = bp->b_blkno;
507 					return;
508 				}
509 			}
510 		}
511 		/*
512 		 * Consider beginning a cluster.
513 		 * If at end of file, make cluster as large as possible,
514 		 * otherwise find size of existing cluster.
515 		 */
516 		if ((lbn + 1) * bp->b_bcount != filesize &&
517 		    (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen) ||
518 		     bp->b_blkno == -1)) {
519 			bawrite(bp);
520 			vp->v_clen = 0;
521 			vp->v_lasta = bp->b_blkno;
522 			vp->v_cstart = lbn + 1;
523 			vp->v_lastw = lbn;
524 			return;
525 		}
526                 vp->v_clen = maxclen;
527                 if (maxclen == 0) {		/* I/O not contiguous */
528 			vp->v_cstart = lbn + 1;
529                         bawrite(bp);
530                 } else {			/* Wait for rest of cluster */
531 			vp->v_cstart = lbn;
532                         bdwrite(bp);
533 		}
534 	} else if (lbn == vp->v_cstart + vp->v_clen) {
535 		/*
536 		 * At end of cluster, write it out.
537 		 */
538 		cluster_wbuild(vp, bp, bp->b_bcount, vp->v_cstart,
539 		    vp->v_clen + 1, lbn);
540 		vp->v_clen = 0;
541 		vp->v_cstart = lbn + 1;
542 	} else
543 		/*
544 		 * In the middle of a cluster, so just delay the
545 		 * I/O for now.
546 		 */
547 		bdwrite(bp);
548 	vp->v_lastw = lbn;
549 	vp->v_lasta = bp->b_blkno;
550 }
551 
552 
553 /*
554  * This is an awful lot like cluster_rbuild...wish they could be combined.
555  * The last lbn argument is the current block on which I/O is being
556  * performed.  Check to see that it doesn't fall in the middle of
557  * the current block (if last_bp == NULL).
558  */
559 void
560 cluster_wbuild(vp, last_bp, size, start_lbn, len, lbn)
561 	struct vnode *vp;
562 	struct buf *last_bp;
563 	long size;
564 	daddr_t start_lbn;
565 	int len;
566 	daddr_t	lbn;
567 {
568 	struct cluster_save *b_save;
569 	struct buf *bp, *tbp;
570 	caddr_t	cp;
571 	int i, s;
572 
573 #ifdef DIAGNOSTIC
574 	if (size != vp->v_mount->mnt_stat.f_iosize)
575 		panic("cluster_wbuild: size %d != filesize %d\n",
576 			size, vp->v_mount->mnt_stat.f_iosize);
577 #endif
578 redo:
579 	while ((!incore(vp, start_lbn) || start_lbn == lbn) && len) {
580 		++start_lbn;
581 		--len;
582 	}
583 
584 	/* Get more memory for current buffer */
585 	if (len <= 1) {
586 		if (last_bp) {
587 			bawrite(last_bp);
588 		} else if (len) {
589 			bp = getblk(vp, start_lbn, size, 0, 0);
590 			bawrite(bp);
591 		}
592 		return;
593 	}
594 
595 	bp = getblk(vp, start_lbn, size, 0, 0);
596 	if (!(bp->b_flags & B_DELWRI)) {
597 		++start_lbn;
598 		--len;
599 		brelse(bp);
600 		goto redo;
601 	}
602 
603 	/*
604 	 * Extra memory in the buffer, punt on this buffer.
605 	 * XXX we could handle this in most cases, but we would have to
606 	 * push the extra memory down to after our max possible cluster
607 	 * size and then potentially pull it back up if the cluster was
608 	 * terminated prematurely--too much hassle.
609 	 */
610 	if (bp->b_bcount != bp->b_bufsize) {
611 		++start_lbn;
612 		--len;
613 		bawrite(bp);
614 		goto redo;
615 	}
616 
617 	--len;
618 	b_save = malloc(sizeof(struct buf *) * len + sizeof(struct cluster_save),
619 	    M_SEGMENT, M_WAITOK);
620 	b_save->bs_bcount = bp->b_bcount;
621 	b_save->bs_bufsize = bp->b_bufsize;
622 	b_save->bs_nchildren = 0;
623 	b_save->bs_children = (struct buf **)(b_save + 1);
624 	b_save->bs_saveaddr = bp->b_saveaddr;
625 	bp->b_saveaddr = (caddr_t) b_save;
626 
627 	bp->b_flags |= B_CALL;
628 	bp->b_iodone = cluster_callback;
629 	cp = (char *)bp->b_data + size;
630 	for (++start_lbn, i = 0; i < len; ++i, ++start_lbn) {
631 		/*
632 		 * Block is not in core or the non-sequential block
633 		 * ending our cluster was part of the cluster (in which
634 		 * case we don't want to write it twice).
635 		 */
636 		if (!incore(vp, start_lbn) ||
637 		    last_bp == NULL && start_lbn == lbn)
638 			break;
639 
640 		/*
641 		 * Get the desired block buffer (unless it is the final
642 		 * sequential block whose buffer was passed in explictly
643 		 * as last_bp).
644 		 */
645 		if (last_bp == NULL || start_lbn != lbn) {
646 			tbp = getblk(vp, start_lbn, size, 0, 0);
647 			if (!(tbp->b_flags & B_DELWRI)) {
648 				brelse(tbp);
649 				break;
650 			}
651 		} else
652 			tbp = last_bp;
653 
654 		++b_save->bs_nchildren;
655 
656 		/* Move memory from children to parent */
657 		if (tbp->b_blkno != (bp->b_blkno + btodb(bp->b_bufsize))) {
658 			printf("Clustered Block: %d addr %x bufsize: %d\n",
659 			    bp->b_lblkno, bp->b_blkno, bp->b_bufsize);
660 			printf("Child Block: %d addr: %x\n", tbp->b_lblkno,
661 			    tbp->b_blkno);
662 			panic("Clustered write to wrong blocks");
663 		}
664 
665 		pagemove(tbp->b_data, cp, size);
666 		bp->b_bcount += size;
667 		bp->b_bufsize += size;
668 
669 		tbp->b_bufsize -= size;
670 		tbp->b_flags &= ~(B_READ | B_DONE | B_ERROR | B_DELWRI);
671 		tbp->b_flags |= (B_ASYNC | B_AGE);
672 		s = splbio();
673 		reassignbuf(tbp, tbp->b_vp);		/* put on clean list */
674 		++tbp->b_vp->v_numoutput;
675 		splx(s);
676 		b_save->bs_children[i] = tbp;
677 
678 		cp += size;
679 	}
680 
681 	if (i == 0) {
682 		/* None to cluster */
683 		bp->b_saveaddr = b_save->bs_saveaddr;
684 		bp->b_flags &= ~B_CALL;
685 		bp->b_iodone = NULL;
686 		free(b_save, M_SEGMENT);
687 	}
688 	bawrite(bp);
689 	if (i < len) {
690 		len -= i + 1;
691 		start_lbn += 1;
692 		goto redo;
693 	}
694 }
695 
696 /*
697  * Collect together all the buffers in a cluster.
698  * Plus add one additional buffer.
699  */
700 struct cluster_save *
701 cluster_collectbufs(vp, last_bp)
702 	struct vnode *vp;
703 	struct buf *last_bp;
704 {
705 	struct cluster_save *buflist;
706 	daddr_t	lbn;
707 	int i, len;
708 
709 	len = vp->v_lastw - vp->v_cstart + 1;
710 	buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
711 	    M_SEGMENT, M_WAITOK);
712 	buflist->bs_nchildren = 0;
713 	buflist->bs_children = (struct buf **)(buflist + 1);
714 	for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++)
715 		    (void)bread(vp, lbn, last_bp->b_bcount, NOCRED,
716 			&buflist->bs_children[i]);
717 	buflist->bs_children[i] = last_bp;
718 	buflist->bs_nchildren = i + 1;
719 	return (buflist);
720 }
721