xref: /dflybsd-src/sys/kern/vfs_cluster.c (revision 41871674d0079dec70d55eb824f39d07dc7b3310)
1 /*-
2  * Copyright (c) 1993
3  *	The Regents of the University of California.  All rights reserved.
4  * Modifications/enhancements:
5  * 	Copyright (c) 1995 John S. Dyson.  All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. All advertising materials mentioning features or use of this software
16  *    must display the following acknowledgement:
17  *	This product includes software developed by the University of
18  *	California, Berkeley and its contributors.
19  * 4. Neither the name of the University nor the names of its contributors
20  *    may be used to endorse or promote products derived from this software
21  *    without specific prior written permission.
22  *
23  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33  * SUCH DAMAGE.
34  *
35  *	@(#)vfs_cluster.c	8.7 (Berkeley) 2/13/94
36  * $FreeBSD: src/sys/kern/vfs_cluster.c,v 1.92.2.9 2001/11/18 07:10:59 dillon Exp $
37  * $DragonFly: src/sys/kern/vfs_cluster.c,v 1.20 2006/03/26 07:56:54 swildner Exp $
38  */
39 
40 #include "opt_debug_cluster.h"
41 
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
45 #include <sys/proc.h>
46 #include <sys/buf.h>
47 #include <sys/vnode.h>
48 #include <sys/malloc.h>
49 #include <sys/mount.h>
50 #include <sys/resourcevar.h>
51 #include <sys/vmmeter.h>
52 #include <vm/vm.h>
53 #include <vm/vm_object.h>
54 #include <vm/vm_page.h>
55 #include <sys/sysctl.h>
56 #include <sys/buf2.h>
57 #include <vm/vm_page2.h>
58 
59 #if defined(CLUSTERDEBUG)
60 #include <sys/sysctl.h>
61 static int	rcluster= 0;
62 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, "");
63 #endif
64 
65 static MALLOC_DEFINE(M_SEGMENT, "cluster_save", "cluster_save buffer");
66 
67 static struct cluster_save *
68 	cluster_collectbufs (struct vnode *vp, struct buf *last_bp,
69 			    int lblocksize);
70 static struct buf *
71 	cluster_rbuild (struct vnode *vp, off_t filesize, off_t loffset,
72 			    off_t doffset, int size, int run, struct buf *fbp);
73 static void cluster_callback (struct bio *);
74 
75 
76 static int write_behind = 1;
77 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, "");
78 
79 extern vm_page_t	bogus_page;
80 
81 extern int cluster_pbuf_freecnt;
82 
83 /*
84  * Maximum number of blocks for read-ahead.
85  */
86 #define MAXRA 32
87 
88 /*
89  * This replaces bread.
90  */
91 int
92 cluster_read(struct vnode *vp, off_t filesize, off_t loffset,
93 	     int size, int totread, int seqcount, struct buf **bpp)
94 {
95 	struct buf *bp, *rbp, *reqbp;
96 	off_t origoffset;
97 	off_t doffset;
98 	int error;
99 	int i;
100 	int maxra, racluster;
101 
102 	error = 0;
103 
104 	/*
105 	 * Try to limit the amount of read-ahead by a few
106 	 * ad-hoc parameters.  This needs work!!!
107 	 */
108 	racluster = vp->v_mount->mnt_iosize_max / size;
109 	maxra = 2 * racluster + (totread / size);
110 	if (maxra > MAXRA)
111 		maxra = MAXRA;
112 	if (maxra > nbuf/8)
113 		maxra = nbuf/8;
114 
115 	/*
116 	 * get the requested block
117 	 */
118 	*bpp = reqbp = bp = getblk(vp, loffset, size, 0, 0);
119 	origoffset = loffset;
120 
121 	/*
122 	 * if it is in the cache, then check to see if the reads have been
123 	 * sequential.  If they have, then try some read-ahead, otherwise
124 	 * back-off on prospective read-aheads.
125 	 */
126 	if (bp->b_flags & B_CACHE) {
127 		if (!seqcount) {
128 			return 0;
129 		} else if ((bp->b_flags & B_RAM) == 0) {
130 			return 0;
131 		} else {
132 			struct buf *tbp;
133 			bp->b_flags &= ~B_RAM;
134 			/*
135 			 * We do the crit here so that there is no window
136 			 * between the findblk and the b_usecount increment
137 			 * below.  We opt to keep the crit out of the loop
138 			 * for efficiency.
139 			 */
140 			crit_enter();
141 			for (i = 1; i < maxra; i++) {
142 				if (!(tbp = findblk(vp, loffset + i * size))) {
143 					break;
144 				}
145 
146 				/*
147 				 * Set another read-ahead mark so we know
148 				 * to check again.
149 				 */
150 				if (((i % racluster) == (racluster - 1)) ||
151 					(i == (maxra - 1)))
152 					tbp->b_flags |= B_RAM;
153 			}
154 			crit_exit();
155 			if (i >= maxra) {
156 				return 0;
157 			}
158 			loffset += i * size;
159 		}
160 		reqbp = bp = NULL;
161 	} else {
162 		off_t firstread = bp->b_loffset;
163 		int nblks;
164 
165 		KASSERT(firstread != NOOFFSET,
166 			("cluster_read: no buffer offset"));
167 		if (firstread + totread > filesize)
168 			totread = (int)(filesize - firstread);
169 		nblks = totread / size;
170 		if (nblks) {
171 			int burstbytes;
172 
173 			if (nblks > racluster)
174 				nblks = racluster;
175 
176 	    		error = VOP_BMAP(vp, loffset, NULL,
177 					 &doffset, &burstbytes, NULL);
178 			if (error)
179 				goto single_block_read;
180 			if (doffset == NOOFFSET)
181 				goto single_block_read;
182 			if (burstbytes < size * 2)
183 				goto single_block_read;
184 			if (nblks > burstbytes / size)
185 				nblks = burstbytes / size;
186 
187 			bp = cluster_rbuild(vp, filesize, loffset,
188 					    doffset, size, nblks, bp);
189 			loffset += bp->b_bufsize;
190 		} else {
191 single_block_read:
192 			/*
193 			 * if it isn't in the cache, then get a chunk from
194 			 * disk if sequential, otherwise just get the block.
195 			 */
196 			bp->b_flags |= B_READ | B_RAM;
197 			loffset += size;
198 		}
199 	}
200 
201 	/*
202 	 * If we have been doing sequential I/O, then do some read-ahead.
203 	 */
204 	rbp = NULL;
205 	if (seqcount &&
206 	    loffset < origoffset + seqcount * size &&
207 	    loffset + size <= filesize
208 	) {
209 		rbp = getblk(vp, loffset, size, 0, 0);
210 		if ((rbp->b_flags & B_CACHE) == 0) {
211 			int nblksread;
212 			int ntoread;
213 			int burstbytes;
214 
215 			error = VOP_BMAP(vp, loffset, NULL,
216 					 &doffset, &burstbytes, NULL);
217 			if (error || doffset == NOOFFSET) {
218 				rbp->b_flags &= ~(B_ASYNC | B_READ);
219 				brelse(rbp);
220 				rbp = NULL;
221 				goto no_read_ahead;
222 			}
223 			ntoread = burstbytes / size;
224 			nblksread = (totread + size - 1) / size;
225 			if (seqcount < nblksread)
226 				seqcount = nblksread;
227 			if (seqcount < ntoread)
228 				ntoread = seqcount;
229 
230 			rbp->b_flags |= B_READ | B_ASYNC | B_RAM;
231 			if (burstbytes) {
232 				rbp = cluster_rbuild(vp, filesize, loffset,
233 						     doffset, size,
234 						     ntoread, rbp);
235 			} else {
236 				rbp->b_bio2.bio_offset = doffset;
237 			}
238 		}
239 	}
240 no_read_ahead:
241 
242 	/*
243 	 * Handle the synchronous read
244 	 */
245 	if (bp) {
246 #if defined(CLUSTERDEBUG)
247 		if (rcluster)
248 			printf("S(%lld,%d,%d) ",
249 			    bp->b_loffset, bp->b_bcount, seqcount);
250 #endif
251 		if ((bp->b_flags & B_CLUSTER) == 0) {
252 			vfs_busy_pages(bp, 0);
253 		}
254 		bp->b_flags &= ~(B_ERROR|B_INVAL);
255 		if ((bp->b_flags & B_ASYNC) || bp->b_bio1.bio_done != NULL)
256 			BUF_KERNPROC(bp);
257 		vn_strategy(vp, &bp->b_bio1);
258 		error = bp->b_error;
259 	}
260 
261 	/*
262 	 * And if we have read-aheads, do them too
263 	 */
264 	if (rbp) {
265 		if (error) {
266 			rbp->b_flags &= ~(B_ASYNC | B_READ);
267 			brelse(rbp);
268 		} else if (rbp->b_flags & B_CACHE) {
269 			rbp->b_flags &= ~(B_ASYNC | B_READ);
270 			bqrelse(rbp);
271 		} else {
272 #if defined(CLUSTERDEBUG)
273 			if (rcluster) {
274 				if (bp)
275 					printf("A+(%lld,%d,%lld,%d) ",
276 					    rbp->b_loffset, rbp->b_bcount,
277 					    rbp->b_loffset - origoffset,
278 					    seqcount);
279 				else
280 					printf("A(%lld,%d,%lld,%d) ",
281 					    rbp->b_loffset, rbp->b_bcount,
282 					    rbp->b_loffset - origoffset,
283 					    seqcount);
284 			}
285 #endif
286 
287 			if ((rbp->b_flags & B_CLUSTER) == 0) {
288 				vfs_busy_pages(rbp, 0);
289 			}
290 			rbp->b_flags &= ~(B_ERROR|B_INVAL);
291 			if ((rbp->b_flags & B_ASYNC) || rbp->b_bio1.bio_done != NULL)
292 				BUF_KERNPROC(rbp);
293 			vn_strategy(vp, &rbp->b_bio1);
294 		}
295 	}
296 	if (reqbp)
297 		return (biowait(reqbp));
298 	else
299 		return (error);
300 }
301 
302 /*
303  * If blocks are contiguous on disk, use this to provide clustered
304  * read ahead.  We will read as many blocks as possible sequentially
305  * and then parcel them up into logical blocks in the buffer hash table.
306  */
307 static struct buf *
308 cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset,
309 	off_t doffset, int size, int run, struct buf *fbp)
310 {
311 	struct buf *bp, *tbp;
312 	off_t boffset;
313 	int i, j;
314 
315 	KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
316 	    ("cluster_rbuild: size %d != filesize %ld\n",
317 	    size, vp->v_mount->mnt_stat.f_iosize));
318 
319 	/*
320 	 * avoid a division
321 	 */
322 	while (loffset + run * size > filesize) {
323 		--run;
324 	}
325 
326 	tbp = fbp;
327 	tbp->b_flags |= B_READ;
328 	tbp->b_bio2.bio_offset = doffset;
329 	if( (tbp->b_flags & B_MALLOC) ||
330 		((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
331 		return tbp;
332 
333 	bp = trypbuf(&cluster_pbuf_freecnt);
334 	if (bp == 0)
335 		return tbp;
336 
337 	/*
338 	 * We are synthesizing a buffer out of vm_page_t's, but
339 	 * if the block size is not page aligned then the starting
340 	 * address may not be either.  Inherit the b_data offset
341 	 * from the original buffer.
342 	 */
343 	bp->b_data = (char *)((vm_offset_t)bp->b_data |
344 	    ((vm_offset_t)tbp->b_data & PAGE_MASK));
345 	bp->b_flags = B_ASYNC | B_READ | B_CLUSTER | B_VMIO;
346 	bp->b_bio1.bio_done = cluster_callback;
347 	bp->b_bio1.bio_caller_info1.cluster_head = NULL;
348 	bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
349 	bp->b_loffset = loffset;
350 	bp->b_bio2.bio_offset = NOOFFSET;
351 	KASSERT(bp->b_loffset != NOOFFSET,
352 		("cluster_rbuild: no buffer offset"));
353 	pbgetvp(vp, bp);
354 
355 	bp->b_bcount = 0;
356 	bp->b_bufsize = 0;
357 	bp->b_xio.xio_npages = 0;
358 
359 	for (boffset = doffset, i = 0; i < run; ++i, boffset += size) {
360 		if (i != 0) {
361 			if ((bp->b_xio.xio_npages * PAGE_SIZE) +
362 			    round_page(size) > vp->v_mount->mnt_iosize_max) {
363 				break;
364 			}
365 
366 			/*
367 			 * Shortcut some checks and try to avoid buffers that
368 			 * would block in the lock.  The same checks have to
369 			 * be made again after we officially get the buffer.
370 			 */
371 			if ((tbp = findblk(vp, loffset + i * size)) != NULL) {
372 				if (BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT))
373 					break;
374 				BUF_UNLOCK(tbp);
375 
376 				for (j = 0; j < tbp->b_xio.xio_npages; j++) {
377 					if (tbp->b_xio.xio_pages[j]->valid)
378 						break;
379 				}
380 
381 				if (j != tbp->b_xio.xio_npages)
382 					break;
383 
384 				if (tbp->b_bcount != size)
385 					break;
386 			}
387 
388 			tbp = getblk(vp, loffset + i * size, size, 0, 0);
389 
390 			/*
391 			 * Stop scanning if the buffer is fuly valid
392 			 * (marked B_CACHE), or locked (may be doing a
393 			 * background write), or if the buffer is not
394 			 * VMIO backed.  The clustering code can only deal
395 			 * with VMIO-backed buffers.
396 			 */
397 			if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
398 			    (tbp->b_flags & B_VMIO) == 0) {
399 				bqrelse(tbp);
400 				break;
401 			}
402 
403 			/*
404 			 * The buffer must be completely invalid in order to
405 			 * take part in the cluster.  If it is partially valid
406 			 * then we stop.
407 			 */
408 			for (j = 0;j < tbp->b_xio.xio_npages; j++) {
409 				if (tbp->b_xio.xio_pages[j]->valid)
410 					break;
411 			}
412 			if (j != tbp->b_xio.xio_npages) {
413 				bqrelse(tbp);
414 				break;
415 			}
416 
417 			/*
418 			 * Set a read-ahead mark as appropriate
419 			 */
420 			if (i == 1 || i == (run - 1))
421 				tbp->b_flags |= B_RAM;
422 
423 			/*
424 			 * Set the buffer up for an async read (XXX should
425 			 * we do this only if we do not wind up brelse()ing?).
426 			 * Set the block number if it isn't set, otherwise
427 			 * if it is make sure it matches the block number we
428 			 * expect.
429 			 */
430 			tbp->b_flags |= B_READ | B_ASYNC;
431 			if (tbp->b_bio2.bio_offset == NOOFFSET) {
432 				tbp->b_bio2.bio_offset = boffset;
433 			} else if (tbp->b_bio2.bio_offset != boffset) {
434 				brelse(tbp);
435 				break;
436 			}
437 		}
438 		/*
439 		 * XXX fbp from caller may not be B_ASYNC, but we are going
440 		 * to biodone() it in cluster_callback() anyway
441 		 */
442 		BUF_KERNPROC(tbp);
443 		cluster_append(&bp->b_bio1, tbp);
444 		for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
445 			vm_page_t m;
446 			m = tbp->b_xio.xio_pages[j];
447 			vm_page_io_start(m);
448 			vm_object_pip_add(m->object, 1);
449 			if ((bp->b_xio.xio_npages == 0) ||
450 				(bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
451 				bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
452 				bp->b_xio.xio_npages++;
453 			}
454 			if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
455 				tbp->b_xio.xio_pages[j] = bogus_page;
456 		}
457 		/*
458 		 * XXX shouldn't this be += size for both, like in
459 		 * cluster_wbuild()?
460 		 *
461 		 * Don't inherit tbp->b_bufsize as it may be larger due to
462 		 * a non-page-aligned size.  Instead just aggregate using
463 		 * 'size'.
464 		 */
465 		if (tbp->b_bcount != size)
466 		    printf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, size);
467 		if (tbp->b_bufsize != size)
468 		    printf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, size);
469 		bp->b_bcount += size;
470 		bp->b_bufsize += size;
471 	}
472 
473 	/*
474 	 * Fully valid pages in the cluster are already good and do not need
475 	 * to be re-read from disk.  Replace the page with bogus_page
476 	 */
477 	for (j = 0; j < bp->b_xio.xio_npages; j++) {
478 		if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
479 		    VM_PAGE_BITS_ALL) {
480 			bp->b_xio.xio_pages[j] = bogus_page;
481 		}
482 	}
483 	if (bp->b_bufsize > bp->b_kvasize)
484 		panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)",
485 		    bp->b_bufsize, bp->b_kvasize);
486 	bp->b_kvasize = bp->b_bufsize;
487 
488 	pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
489 		(vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages);
490 	return (bp);
491 }
492 
493 /*
494  * Cleanup after a clustered read or write.
495  * This is complicated by the fact that any of the buffers might have
496  * extra memory (if there were no empty buffer headers at allocbuf time)
497  * that we will need to shift around.
498  *
499  * The returned bio is &bp->b_bio1
500  */
501 void
502 cluster_callback(struct bio *bio)
503 {
504 	struct buf *bp = bio->bio_buf;
505 	struct buf *tbp;
506 	int error = 0;
507 
508 	/*
509 	 * Must propogate errors to all the components.
510 	 */
511 	if (bp->b_flags & B_ERROR)
512 		error = bp->b_error;
513 
514 	pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
515 	/*
516 	 * Move memory from the large cluster buffer into the component
517 	 * buffers and mark IO as done on these.  Since the memory map
518 	 * is the same, no actual copying is required.
519 	 */
520 	while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) {
521 		bio->bio_caller_info1.cluster_head = tbp->b_cluster_next;
522 		if (error) {
523 			tbp->b_flags |= B_ERROR;
524 			tbp->b_error = error;
525 		} else {
526 			tbp->b_dirtyoff = tbp->b_dirtyend = 0;
527 			tbp->b_flags &= ~(B_ERROR|B_INVAL);
528 			/*
529 			 * XXX the bdwrite()/bqrelse() issued during
530 			 * cluster building clears B_RELBUF (see bqrelse()
531 			 * comment).  If direct I/O was specified, we have
532 			 * to restore it here to allow the buffer and VM
533 			 * to be freed.
534 			 */
535 			if (tbp->b_flags & B_DIRECT)
536 				tbp->b_flags |= B_RELBUF;
537 		}
538 		biodone(&tbp->b_bio1);
539 	}
540 	relpbuf(bp, &cluster_pbuf_freecnt);
541 }
542 
543 /*
544  *	cluster_wbuild_wb:
545  *
546  *	Implement modified write build for cluster.
547  *
548  *		write_behind = 0	write behind disabled
549  *		write_behind = 1	write behind normal (default)
550  *		write_behind = 2	write behind backed-off
551  */
552 
553 static __inline int
554 cluster_wbuild_wb(struct vnode *vp, int size, off_t start_loffset, int len)
555 {
556 	int r = 0;
557 
558 	switch(write_behind) {
559 	case 2:
560 		if (start_loffset < len)
561 			break;
562 		start_loffset -= len;
563 		/* fall through */
564 	case 1:
565 		r = cluster_wbuild(vp, size, start_loffset, len);
566 		/* fall through */
567 	default:
568 		/* fall through */
569 		break;
570 	}
571 	return(r);
572 }
573 
574 /*
575  * Do clustered write for FFS.
576  *
577  * Three cases:
578  *	1. Write is not sequential (write asynchronously)
579  *	Write is sequential:
580  *	2.	beginning of cluster - begin cluster
581  *	3.	middle of a cluster - add to cluster
582  *	4.	end of a cluster - asynchronously write cluster
583  */
584 void
585 cluster_write(struct buf *bp, off_t filesize, int seqcount)
586 {
587 	struct vnode *vp;
588 	off_t loffset;
589 	int maxclen, cursize;
590 	int lblocksize;
591 	int async;
592 
593 	vp = bp->b_vp;
594 	if (vp->v_type == VREG) {
595 		async = vp->v_mount->mnt_flag & MNT_ASYNC;
596 		lblocksize = vp->v_mount->mnt_stat.f_iosize;
597 	} else {
598 		async = 0;
599 		lblocksize = bp->b_bufsize;
600 	}
601 	loffset = bp->b_loffset;
602 	KASSERT(bp->b_loffset != NOOFFSET,
603 		("cluster_write: no buffer offset"));
604 
605 	/* Initialize vnode to beginning of file. */
606 	if (loffset == 0)
607 		vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
608 
609 	if (vp->v_clen == 0 || loffset != vp->v_lastw + lblocksize ||
610 	    bp->b_bio2.bio_offset == NOOFFSET ||
611 	    (bp->b_bio2.bio_offset != vp->v_lasta + lblocksize)) {
612 		maxclen = vp->v_mount->mnt_iosize_max;
613 		if (vp->v_clen != 0) {
614 			/*
615 			 * Next block is not sequential.
616 			 *
617 			 * If we are not writing at end of file, the process
618 			 * seeked to another point in the file since its last
619 			 * write, or we have reached our maximum cluster size,
620 			 * then push the previous cluster. Otherwise try
621 			 * reallocating to make it sequential.
622 			 *
623 			 * Change to algorithm: only push previous cluster if
624 			 * it was sequential from the point of view of the
625 			 * seqcount heuristic, otherwise leave the buffer
626 			 * intact so we can potentially optimize the I/O
627 			 * later on in the buf_daemon or update daemon
628 			 * flush.
629 			 */
630 			cursize = vp->v_lastw - vp->v_cstart + lblocksize;
631 			if (bp->b_loffset + lblocksize != filesize ||
632 			    loffset != vp->v_lastw + lblocksize || vp->v_clen <= cursize) {
633 				if (!async && seqcount > 0) {
634 					cluster_wbuild_wb(vp, lblocksize,
635 						vp->v_cstart, cursize);
636 				}
637 			} else {
638 				struct buf **bpp, **endbp;
639 				struct cluster_save *buflist;
640 
641 				buflist = cluster_collectbufs(vp, bp,
642 							      lblocksize);
643 				endbp = &buflist->bs_children
644 				    [buflist->bs_nchildren - 1];
645 				if (VOP_REALLOCBLKS(vp, buflist)) {
646 					/*
647 					 * Failed, push the previous cluster
648 					 * if *really* writing sequentially
649 					 * in the logical file (seqcount > 1),
650 					 * otherwise delay it in the hopes that
651 					 * the low level disk driver can
652 					 * optimize the write ordering.
653 					 */
654 					for (bpp = buflist->bs_children;
655 					     bpp < endbp; bpp++)
656 						brelse(*bpp);
657 					free(buflist, M_SEGMENT);
658 					if (seqcount > 1) {
659 						cluster_wbuild_wb(vp,
660 						    lblocksize, vp->v_cstart,
661 						    cursize);
662 					}
663 				} else {
664 					/*
665 					 * Succeeded, keep building cluster.
666 					 */
667 					for (bpp = buflist->bs_children;
668 					     bpp <= endbp; bpp++)
669 						bdwrite(*bpp);
670 					free(buflist, M_SEGMENT);
671 					vp->v_lastw = loffset;
672 					vp->v_lasta = bp->b_bio2.bio_offset;
673 					return;
674 				}
675 			}
676 		}
677 		/*
678 		 * Consider beginning a cluster. If at end of file, make
679 		 * cluster as large as possible, otherwise find size of
680 		 * existing cluster.
681 		 */
682 		if ((vp->v_type == VREG) &&
683 		    bp->b_loffset + lblocksize != filesize &&
684 		    (bp->b_bio2.bio_offset == NOOFFSET) &&
685 		    (VOP_BMAP(vp, loffset, NULL, &bp->b_bio2.bio_offset, &maxclen, NULL) ||
686 		     bp->b_bio2.bio_offset == NOOFFSET)) {
687 			bawrite(bp);
688 			vp->v_clen = 0;
689 			vp->v_lasta = bp->b_bio2.bio_offset;
690 			vp->v_cstart = loffset + lblocksize;
691 			vp->v_lastw = loffset;
692 			return;
693 		}
694 		if (maxclen > lblocksize)
695 			vp->v_clen = maxclen - lblocksize;
696 		else
697 			vp->v_clen = 0;
698 		if (!async && vp->v_clen == 0) { /* I/O not contiguous */
699 			vp->v_cstart = loffset + lblocksize;
700 			bawrite(bp);
701 		} else {	/* Wait for rest of cluster */
702 			vp->v_cstart = loffset;
703 			bdwrite(bp);
704 		}
705 	} else if (loffset == vp->v_cstart + vp->v_clen) {
706 		/*
707 		 * At end of cluster, write it out if seqcount tells us we
708 		 * are operating sequentially, otherwise let the buf or
709 		 * update daemon handle it.
710 		 */
711 		bdwrite(bp);
712 		if (seqcount > 1)
713 			cluster_wbuild_wb(vp, lblocksize, vp->v_cstart,
714 					  vp->v_clen + lblocksize);
715 		vp->v_clen = 0;
716 		vp->v_cstart = loffset + lblocksize;
717 	} else if (vm_page_count_severe()) {
718 		/*
719 		 * We are low on memory, get it going NOW
720 		 */
721 		bawrite(bp);
722 	} else {
723 		/*
724 		 * In the middle of a cluster, so just delay the I/O for now.
725 		 */
726 		bdwrite(bp);
727 	}
728 	vp->v_lastw = loffset;
729 	vp->v_lasta = bp->b_bio2.bio_offset;
730 }
731 
732 
733 /*
734  * This is an awful lot like cluster_rbuild...wish they could be combined.
735  * The last lbn argument is the current block on which I/O is being
736  * performed.  Check to see that it doesn't fall in the middle of
737  * the current block (if last_bp == NULL).
738  */
739 int
740 cluster_wbuild(struct vnode *vp, int size, off_t start_loffset, int bytes)
741 {
742 	struct buf *bp, *tbp;
743 	int i, j;
744 	int totalwritten = 0;
745 
746 	while (bytes > 0) {
747 		crit_enter();
748 		/*
749 		 * If the buffer is not delayed-write (i.e. dirty), or it
750 		 * is delayed-write but either locked or inval, it cannot
751 		 * partake in the clustered write.
752 		 */
753 		if (((tbp = findblk(vp, start_loffset)) == NULL) ||
754 		  ((tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI) ||
755 		  BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
756 			start_loffset += size;
757 			bytes -= size;
758 			crit_exit();
759 			continue;
760 		}
761 		bremfree(tbp);
762 		tbp->b_flags &= ~B_DONE;
763 		crit_exit();
764 
765 		/*
766 		 * Extra memory in the buffer, punt on this buffer.
767 		 * XXX we could handle this in most cases, but we would
768 		 * have to push the extra memory down to after our max
769 		 * possible cluster size and then potentially pull it back
770 		 * up if the cluster was terminated prematurely--too much
771 		 * hassle.
772 		 */
773 		if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
774 		  (tbp->b_bcount != tbp->b_bufsize) ||
775 		  (tbp->b_bcount != size) ||
776 		  (bytes == size) ||
777 		  ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
778 			totalwritten += tbp->b_bufsize;
779 			bawrite(tbp);
780 			start_loffset += size;
781 			bytes -= size;
782 			continue;
783 		}
784 
785 		/*
786 		 * We got a pbuf to make the cluster in.
787 		 * so initialise it.
788 		 */
789 		bp->b_bcount = 0;
790 		bp->b_bufsize = 0;
791 		bp->b_xio.xio_npages = 0;
792 		bp->b_loffset = tbp->b_loffset;
793 		bp->b_bio2.bio_offset = tbp->b_bio2.bio_offset;
794 
795 		/*
796 		 * We are synthesizing a buffer out of vm_page_t's, but
797 		 * if the block size is not page aligned then the starting
798 		 * address may not be either.  Inherit the b_data offset
799 		 * from the original buffer.
800 		 */
801 		bp->b_data = (char *)((vm_offset_t)bp->b_data |
802 		    ((vm_offset_t)tbp->b_data & PAGE_MASK));
803 		bp->b_flags |= B_CLUSTER |
804 			(tbp->b_flags & (B_VMIO | B_NEEDCOMMIT | B_NOWDRAIN));
805 		bp->b_bio1.bio_done = cluster_callback;
806 		bp->b_bio1.bio_caller_info1.cluster_head = NULL;
807 		bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
808 		pbgetvp(vp, bp);
809 		/*
810 		 * From this location in the file, scan forward to see
811 		 * if there are buffers with adjacent data that need to
812 		 * be written as well.
813 		 */
814 		for (i = 0; i < bytes; (i += size), (start_loffset += size)) {
815 			if (i != 0) { /* If not the first buffer */
816 				crit_enter();
817 				/*
818 				 * If the adjacent data is not even in core it
819 				 * can't need to be written.
820 				 */
821 				if ((tbp = findblk(vp, start_loffset)) == NULL) {
822 					crit_exit();
823 					break;
824 				}
825 
826 				/*
827 				 * If it IS in core, but has different
828 				 * characteristics, or is locked (which
829 				 * means it could be undergoing a background
830 				 * I/O or be in a weird state), then don't
831 				 * cluster with it.
832 				 */
833 				if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
834 				    B_INVAL | B_DELWRI | B_NEEDCOMMIT))
835 				  != (B_DELWRI | B_CLUSTEROK |
836 				    (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
837 				    (tbp->b_flags & B_LOCKED) ||
838 				    BUF_LOCK(tbp, LK_EXCLUSIVE | LK_NOWAIT)) {
839 					crit_exit();
840 					break;
841 				}
842 
843 				/*
844 				 * Check that the combined cluster
845 				 * would make sense with regard to pages
846 				 * and would not be too large
847 				 */
848 				if ((tbp->b_bcount != size) ||
849 				  ((bp->b_bio2.bio_offset + i) !=
850 				    tbp->b_bio2.bio_offset) ||
851 				  ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
852 				    (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
853 					BUF_UNLOCK(tbp);
854 					crit_exit();
855 					break;
856 				}
857 				/*
858 				 * Ok, it's passed all the tests,
859 				 * so remove it from the free list
860 				 * and mark it busy. We will use it.
861 				 */
862 				bremfree(tbp);
863 				tbp->b_flags &= ~B_DONE;
864 				crit_exit();
865 			} /* end of code for non-first buffers only */
866 
867 			/*
868 			 * If the IO is via the VM then we do some
869 			 * special VM hackery (yuck).  Since the buffer's
870 			 * block size may not be page-aligned it is possible
871 			 * for a page to be shared between two buffers.  We
872 			 * have to get rid of the duplication when building
873 			 * the cluster.
874 			 */
875 			if (tbp->b_flags & B_VMIO) {
876 				vm_page_t m;
877 
878 				if (i != 0) { /* if not first buffer */
879 					for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
880 						m = tbp->b_xio.xio_pages[j];
881 						if (m->flags & PG_BUSY) {
882 							bqrelse(tbp);
883 							goto finishcluster;
884 						}
885 					}
886 				}
887 
888 				for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
889 					m = tbp->b_xio.xio_pages[j];
890 					vm_page_io_start(m);
891 					vm_object_pip_add(m->object, 1);
892 					if ((bp->b_xio.xio_npages == 0) ||
893 					  (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
894 						bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
895 						bp->b_xio.xio_npages++;
896 					}
897 				}
898 			}
899 			bp->b_bcount += size;
900 			bp->b_bufsize += size;
901 
902 			crit_enter();
903 			bundirty(tbp);
904 			tbp->b_flags &= ~(B_READ | B_DONE | B_ERROR);
905 			tbp->b_flags |= B_ASYNC;
906 			crit_exit();
907 			BUF_KERNPROC(tbp);
908 			cluster_append(&bp->b_bio1, tbp);
909 
910 			/*
911 			 * check for latent dependencies to be handled
912 			 */
913 			if (LIST_FIRST(&tbp->b_dep) != NULL && bioops.io_start)
914 				(*bioops.io_start)(tbp);
915 
916 		}
917 	finishcluster:
918 		pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
919 			(vm_page_t *) bp->b_xio.xio_pages, bp->b_xio.xio_npages);
920 		if (bp->b_bufsize > bp->b_kvasize)
921 			panic(
922 			    "cluster_wbuild: b_bufsize(%d) > b_kvasize(%d)\n",
923 			    bp->b_bufsize, bp->b_kvasize);
924 		bp->b_kvasize = bp->b_bufsize;
925 		totalwritten += bp->b_bufsize;
926 		bp->b_dirtyoff = 0;
927 		bp->b_dirtyend = bp->b_bufsize;
928 		bawrite(bp);
929 
930 		bytes -= i;
931 	}
932 	return totalwritten;
933 }
934 
935 /*
936  * Collect together all the buffers in a cluster.
937  * Plus add one additional buffer.
938  */
939 static struct cluster_save *
940 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int lblocksize)
941 {
942 	struct cluster_save *buflist;
943 	struct buf *bp;
944 	off_t loffset;
945 	int i, len;
946 
947 	len = (int)(vp->v_lastw - vp->v_cstart + lblocksize) / lblocksize;
948 	buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
949 			 M_SEGMENT, M_WAITOK);
950 	buflist->bs_nchildren = 0;
951 	buflist->bs_children = (struct buf **) (buflist + 1);
952 	for (loffset = vp->v_cstart, i = 0; i < len; (loffset += lblocksize), i++) {
953 		(void) bread(vp, loffset, last_bp->b_bcount, &bp);
954 		buflist->bs_children[i] = bp;
955 		if (bp->b_bio2.bio_offset == NOOFFSET) {
956 			VOP_BMAP(bp->b_vp, bp->b_loffset, NULL,
957 				&bp->b_bio2.bio_offset, NULL, NULL);
958 		}
959 	}
960 	buflist->bs_children[i] = bp = last_bp;
961 	if (bp->b_bio2.bio_offset == NOOFFSET) {
962 		VOP_BMAP(bp->b_vp, bp->b_loffset, NULL,
963 			 &bp->b_bio2.bio_offset, NULL, NULL);
964 	}
965 	buflist->bs_nchildren = i + 1;
966 	return (buflist);
967 }
968 
969 void
970 cluster_append(struct bio *bio, struct buf *tbp)
971 {
972 	tbp->b_cluster_next = NULL;
973 	if (bio->bio_caller_info1.cluster_head == NULL) {
974 		bio->bio_caller_info1.cluster_head = tbp;
975 		bio->bio_caller_info2.cluster_tail = tbp;
976 	} else {
977 		bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp;
978 		bio->bio_caller_info2.cluster_tail = tbp;
979 	}
980 }
981 
982