xref: /freebsd/sys/kern/vfs_cluster.c (revision 6e0da4f753ed6b5d26395001a6194b4fdea70177)
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  * 4. Neither the name of the University nor the names of its contributors
16  *    may be used to endorse or promote products derived from this software
17  *    without specific prior written permission.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  *
31  *	@(#)vfs_cluster.c	8.7 (Berkeley) 2/13/94
32  */
33 
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36 
37 #include "opt_debug_cluster.h"
38 
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/kernel.h>
42 #include <sys/proc.h>
43 #include <sys/bio.h>
44 #include <sys/buf.h>
45 #include <sys/vnode.h>
46 #include <sys/malloc.h>
47 #include <sys/mount.h>
48 #include <sys/resourcevar.h>
49 #include <sys/vmmeter.h>
50 #include <vm/vm.h>
51 #include <vm/vm_object.h>
52 #include <vm/vm_page.h>
53 #include <sys/sysctl.h>
54 
55 #if defined(CLUSTERDEBUG)
56 static int	rcluster= 0;
57 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0,
58     "Debug VFS clustering code");
59 #endif
60 
61 static MALLOC_DEFINE(M_SEGMENT, "cluster_save buffer", "cluster_save buffer");
62 
63 static struct cluster_save *
64 	cluster_collectbufs(struct vnode *vp, struct buf *last_bp);
65 static struct buf *
66 	cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn,
67 			 daddr_t blkno, long size, int run, struct buf *fbp);
68 
69 static int write_behind = 1;
70 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
71     "Cluster write-behind; 0: disable, 1: enable, 2: backed off");
72 
73 static int read_max = 8;
74 SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0,
75     "Cluster read-ahead max block count");
76 
77 /* Page expended to mark partially backed buffers */
78 extern vm_page_t	bogus_page;
79 
80 /*
81  * Number of physical bufs (pbufs) this subsystem is allowed.
82  * Manipulated by vm_pager.c
83  */
84 extern int cluster_pbuf_freecnt;
85 
86 /*
87  * Read data to a buf, including read-ahead if we find this to be beneficial.
88  * cluster_read replaces bread.
89  */
90 int
91 cluster_read(vp, filesize, lblkno, size, cred, totread, seqcount, bpp)
92 	struct vnode *vp;
93 	u_quad_t filesize;
94 	daddr_t lblkno;
95 	long size;
96 	struct ucred *cred;
97 	long totread;
98 	int seqcount;
99 	struct buf **bpp;
100 {
101 	struct buf *bp, *rbp, *reqbp;
102 	daddr_t blkno, origblkno;
103 	int maxra, racluster;
104 	int error, ncontig;
105 	int i;
106 
107 	error = 0;
108 
109 	/*
110 	 * Try to limit the amount of read-ahead by a few
111 	 * ad-hoc parameters.  This needs work!!!
112 	 */
113 	racluster = vp->v_mount->mnt_iosize_max / size;
114 	maxra = seqcount;
115 	maxra = min(read_max, maxra);
116 	maxra = min(nbuf/8, maxra);
117 	if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize)
118 		maxra = (filesize / size) - lblkno;
119 
120 	/*
121 	 * get the requested block
122 	 */
123 	*bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0, 0);
124 	origblkno = lblkno;
125 
126 	/*
127 	 * if it is in the cache, then check to see if the reads have been
128 	 * sequential.  If they have, then try some read-ahead, otherwise
129 	 * back-off on prospective read-aheads.
130 	 */
131 	if (bp->b_flags & B_CACHE) {
132 		if (!seqcount) {
133 			return 0;
134 		} else if ((bp->b_flags & B_RAM) == 0) {
135 			return 0;
136 		} else {
137 			int s;
138 			bp->b_flags &= ~B_RAM;
139 			/*
140 			 * We do the spl here so that there is no window
141 			 * between the incore and the b_usecount increment
142 			 * below.  We opt to keep the spl out of the loop
143 			 * for efficiency.
144 			 */
145 			s = splbio();
146 			VI_LOCK(vp);
147 			for (i = 1; i < maxra; i++) {
148 				/*
149 				 * Stop if the buffer does not exist or it
150 				 * is invalid (about to go away?)
151 				 */
152 				rbp = gbincore(&vp->v_bufobj, lblkno+i);
153 				if (rbp == NULL || (rbp->b_flags & B_INVAL))
154 					break;
155 
156 				/*
157 				 * Set another read-ahead mark so we know
158 				 * to check again.
159 				 */
160 				if (((i % racluster) == (racluster - 1)) ||
161 					(i == (maxra - 1)))
162 					rbp->b_flags |= B_RAM;
163 			}
164 			VI_UNLOCK(vp);
165 			splx(s);
166 			if (i >= maxra) {
167 				return 0;
168 			}
169 			lblkno += i;
170 		}
171 		reqbp = bp = NULL;
172 	/*
173 	 * If it isn't in the cache, then get a chunk from
174 	 * disk if sequential, otherwise just get the block.
175 	 */
176 	} else {
177 		off_t firstread = bp->b_offset;
178 		int nblks;
179 
180 		KASSERT(bp->b_offset != NOOFFSET,
181 		    ("cluster_read: no buffer offset"));
182 
183 		ncontig = 0;
184 
185 		/*
186 		 * Compute the total number of blocks that we should read
187 		 * synchronously.
188 		 */
189 		if (firstread + totread > filesize)
190 			totread = filesize - firstread;
191 		nblks = howmany(totread, size);
192 		if (nblks > racluster)
193 			nblks = racluster;
194 
195 		/*
196 		 * Now compute the number of contiguous blocks.
197 		 */
198 		if (nblks > 1) {
199 	    		error = VOP_BMAP(vp, lblkno, NULL,
200 				&blkno, &ncontig, NULL);
201 			/*
202 			 * If this failed to map just do the original block.
203 			 */
204 			if (error || blkno == -1)
205 				ncontig = 0;
206 		}
207 
208 		/*
209 		 * If we have contiguous data available do a cluster
210 		 * otherwise just read the requested block.
211 		 */
212 		if (ncontig) {
213 			/* Account for our first block. */
214 			ncontig = min(ncontig + 1, nblks);
215 			if (ncontig < nblks)
216 				nblks = ncontig;
217 			bp = cluster_rbuild(vp, filesize, lblkno,
218 				blkno, size, nblks, bp);
219 			lblkno += (bp->b_bufsize / size);
220 		} else {
221 			bp->b_flags |= B_RAM;
222 			bp->b_iocmd = BIO_READ;
223 			lblkno += 1;
224 		}
225 	}
226 
227 	/*
228 	 * handle the synchronous read so that it is available ASAP.
229 	 */
230 	if (bp) {
231 		if ((bp->b_flags & B_CLUSTER) == 0) {
232 			vfs_busy_pages(bp, 0);
233 		}
234 		bp->b_flags &= ~B_INVAL;
235 		bp->b_ioflags &= ~BIO_ERROR;
236 		if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL)
237 			BUF_KERNPROC(bp);
238 		bp->b_iooffset = dbtob(bp->b_blkno);
239 		bstrategy(bp);
240 		curproc->p_stats->p_ru.ru_inblock++;
241 	}
242 
243 	/*
244 	 * If we have been doing sequential I/O, then do some read-ahead.
245 	 */
246 	while (lblkno < (origblkno + maxra)) {
247 		error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL);
248 		if (error)
249 			break;
250 
251 		if (blkno == -1)
252 			break;
253 
254 		/*
255 		 * We could throttle ncontig here by maxra but we might as
256 		 * well read the data if it is contiguous.  We're throttled
257 		 * by racluster anyway.
258 		 */
259 		if (ncontig) {
260 			ncontig = min(ncontig + 1, racluster);
261 			rbp = cluster_rbuild(vp, filesize, lblkno, blkno,
262 				size, ncontig, NULL);
263 			lblkno += (rbp->b_bufsize / size);
264 			if (rbp->b_flags & B_DELWRI) {
265 				bqrelse(rbp);
266 				continue;
267 			}
268 		} else {
269 			rbp = getblk(vp, lblkno, size, 0, 0, 0);
270 			lblkno += 1;
271 			if (rbp->b_flags & B_DELWRI) {
272 				bqrelse(rbp);
273 				continue;
274 			}
275 			rbp->b_flags |= B_ASYNC | B_RAM;
276 			rbp->b_iocmd = BIO_READ;
277 			rbp->b_blkno = blkno;
278 		}
279 		if (rbp->b_flags & B_CACHE) {
280 			rbp->b_flags &= ~B_ASYNC;
281 			bqrelse(rbp);
282 			continue;
283 		}
284 		if ((rbp->b_flags & B_CLUSTER) == 0) {
285 			vfs_busy_pages(rbp, 0);
286 		}
287 		rbp->b_flags &= ~B_INVAL;
288 		rbp->b_ioflags &= ~BIO_ERROR;
289 		if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL)
290 			BUF_KERNPROC(rbp);
291 		rbp->b_iooffset = dbtob(rbp->b_blkno);
292 		bstrategy(rbp);
293 		curproc->p_stats->p_ru.ru_inblock++;
294 	}
295 
296 	if (reqbp)
297 		return (bufwait(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(vp, filesize, lbn, blkno, size, run, fbp)
309 	struct vnode *vp;
310 	u_quad_t filesize;
311 	daddr_t lbn;
312 	daddr_t blkno;
313 	long size;
314 	int run;
315 	struct buf *fbp;
316 {
317 	struct buf *bp, *tbp;
318 	daddr_t bn;
319 	int i, inc, j;
320 
321 	GIANT_REQUIRED;
322 
323 	KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
324 	    ("cluster_rbuild: size %ld != filesize %jd\n",
325 	    size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
326 
327 	/*
328 	 * avoid a division
329 	 */
330 	while ((u_quad_t) size * (lbn + run) > filesize) {
331 		--run;
332 	}
333 
334 	if (fbp) {
335 		tbp = fbp;
336 		tbp->b_iocmd = BIO_READ;
337 	} else {
338 		tbp = getblk(vp, lbn, size, 0, 0, 0);
339 		if (tbp->b_flags & B_CACHE)
340 			return tbp;
341 		tbp->b_flags |= B_ASYNC | B_RAM;
342 		tbp->b_iocmd = BIO_READ;
343 	}
344 
345 	tbp->b_blkno = blkno;
346 	if( (tbp->b_flags & B_MALLOC) ||
347 		((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
348 		return tbp;
349 
350 	bp = trypbuf(&cluster_pbuf_freecnt);
351 	if (bp == 0)
352 		return tbp;
353 
354 	/*
355 	 * We are synthesizing a buffer out of vm_page_t's, but
356 	 * if the block size is not page aligned then the starting
357 	 * address may not be either.  Inherit the b_data offset
358 	 * from the original buffer.
359 	 */
360 	bp->b_data = (char *)((vm_offset_t)bp->b_data |
361 	    ((vm_offset_t)tbp->b_data & PAGE_MASK));
362 	bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO;
363 	bp->b_iocmd = BIO_READ;
364 	bp->b_iodone = cluster_callback;
365 	bp->b_blkno = blkno;
366 	bp->b_lblkno = lbn;
367 	bp->b_offset = tbp->b_offset;
368 	KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
369 	pbgetvp(vp, bp);
370 
371 	TAILQ_INIT(&bp->b_cluster.cluster_head);
372 
373 	bp->b_bcount = 0;
374 	bp->b_bufsize = 0;
375 	bp->b_npages = 0;
376 
377 	inc = btodb(size);
378 	for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
379 		if (i != 0) {
380 			if ((bp->b_npages * PAGE_SIZE) +
381 			    round_page(size) > vp->v_mount->mnt_iosize_max) {
382 				break;
383 			}
384 
385 			tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT);
386 
387 			/* Don't wait around for locked bufs. */
388 			if (tbp == NULL)
389 				break;
390 
391 			/*
392 			 * Stop scanning if the buffer is fully valid
393 			 * (marked B_CACHE), or locked (may be doing a
394 			 * background write), or if the buffer is not
395 			 * VMIO backed.  The clustering code can only deal
396 			 * with VMIO-backed buffers.
397 			 */
398 			VI_LOCK(vp);
399 			if ((tbp->b_vflags & BV_BKGRDINPROG) ||
400 			    (tbp->b_flags & B_CACHE) ||
401 			    (tbp->b_flags & B_VMIO) == 0) {
402 				VI_UNLOCK(vp);
403 				bqrelse(tbp);
404 				break;
405 			}
406 			VI_UNLOCK(vp);
407 
408 			/*
409 			 * The buffer must be completely invalid in order to
410 			 * take part in the cluster.  If it is partially valid
411 			 * then we stop.
412 			 */
413 			VM_OBJECT_LOCK(tbp->b_bufobj->bo_object);
414 			for (j = 0;j < tbp->b_npages; j++) {
415 				VM_OBJECT_LOCK_ASSERT(tbp->b_pages[j]->object,
416 				    MA_OWNED);
417 				if (tbp->b_pages[j]->valid)
418 					break;
419 			}
420 			VM_OBJECT_UNLOCK(tbp->b_bufobj->bo_object);
421 			if (j != tbp->b_npages) {
422 				bqrelse(tbp);
423 				break;
424 			}
425 
426 			/*
427 			 * Set a read-ahead mark as appropriate
428 			 */
429 			if ((fbp && (i == 1)) || (i == (run - 1)))
430 				tbp->b_flags |= B_RAM;
431 
432 			/*
433 			 * Set the buffer up for an async read (XXX should
434 			 * we do this only if we do not wind up brelse()ing?).
435 			 * Set the block number if it isn't set, otherwise
436 			 * if it is make sure it matches the block number we
437 			 * expect.
438 			 */
439 			tbp->b_flags |= B_ASYNC;
440 			tbp->b_iocmd = BIO_READ;
441 			if (tbp->b_blkno == tbp->b_lblkno) {
442 				tbp->b_blkno = bn;
443 			} else if (tbp->b_blkno != bn) {
444 				brelse(tbp);
445 				break;
446 			}
447 		}
448 		/*
449 		 * XXX fbp from caller may not be B_ASYNC, but we are going
450 		 * to biodone() it in cluster_callback() anyway
451 		 */
452 		BUF_KERNPROC(tbp);
453 		TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
454 			tbp, b_cluster.cluster_entry);
455 		VM_OBJECT_LOCK(tbp->b_bufobj->bo_object);
456 		for (j = 0; j < tbp->b_npages; j += 1) {
457 			vm_page_t m;
458 			m = tbp->b_pages[j];
459 			vm_page_io_start(m);
460 			vm_object_pip_add(m->object, 1);
461 			if ((bp->b_npages == 0) ||
462 				(bp->b_pages[bp->b_npages-1] != m)) {
463 				bp->b_pages[bp->b_npages] = m;
464 				bp->b_npages++;
465 			}
466 			if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
467 				tbp->b_pages[j] = bogus_page;
468 		}
469 		VM_OBJECT_UNLOCK(tbp->b_bufobj->bo_object);
470 		/*
471 		 * XXX shouldn't this be += size for both, like in
472 		 * cluster_wbuild()?
473 		 *
474 		 * Don't inherit tbp->b_bufsize as it may be larger due to
475 		 * a non-page-aligned size.  Instead just aggregate using
476 		 * 'size'.
477 		 */
478 		if (tbp->b_bcount != size)
479 			printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
480 		if (tbp->b_bufsize != size)
481 			printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
482 		bp->b_bcount += size;
483 		bp->b_bufsize += size;
484 	}
485 
486 	/*
487 	 * Fully valid pages in the cluster are already good and do not need
488 	 * to be re-read from disk.  Replace the page with bogus_page
489 	 */
490 	VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
491 	for (j = 0; j < bp->b_npages; j++) {
492 		VM_OBJECT_LOCK_ASSERT(bp->b_pages[j]->object, MA_OWNED);
493 		if ((bp->b_pages[j]->valid & VM_PAGE_BITS_ALL) ==
494 		    VM_PAGE_BITS_ALL) {
495 			bp->b_pages[j] = bogus_page;
496 		}
497 	}
498 	VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
499 	if (bp->b_bufsize > bp->b_kvasize)
500 		panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
501 		    bp->b_bufsize, bp->b_kvasize);
502 	bp->b_kvasize = bp->b_bufsize;
503 
504 	pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
505 		(vm_page_t *)bp->b_pages, bp->b_npages);
506 	return (bp);
507 }
508 
509 /*
510  * Cleanup after a clustered read or write.
511  * This is complicated by the fact that any of the buffers might have
512  * extra memory (if there were no empty buffer headers at allocbuf time)
513  * that we will need to shift around.
514  */
515 void
516 cluster_callback(bp)
517 	struct buf *bp;
518 {
519 	struct buf *nbp, *tbp;
520 	int error = 0;
521 
522 	GIANT_REQUIRED;
523 
524 	/*
525 	 * Must propogate errors to all the components.
526 	 */
527 	if (bp->b_ioflags & BIO_ERROR)
528 		error = bp->b_error;
529 
530 	pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
531 	/*
532 	 * Move memory from the large cluster buffer into the component
533 	 * buffers and mark IO as done on these.
534 	 */
535 	for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
536 		tbp; tbp = nbp) {
537 		nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
538 		if (error) {
539 			tbp->b_ioflags |= BIO_ERROR;
540 			tbp->b_error = error;
541 		} else {
542 			tbp->b_dirtyoff = tbp->b_dirtyend = 0;
543 			tbp->b_flags &= ~B_INVAL;
544 			tbp->b_ioflags &= ~BIO_ERROR;
545 			/*
546 			 * XXX the bdwrite()/bqrelse() issued during
547 			 * cluster building clears B_RELBUF (see bqrelse()
548 			 * comment).  If direct I/O was specified, we have
549 			 * to restore it here to allow the buffer and VM
550 			 * to be freed.
551 			 */
552 			if (tbp->b_flags & B_DIRECT)
553 				tbp->b_flags |= B_RELBUF;
554 		}
555 		bufdone(tbp);
556 	}
557 	pbrelvp(bp);
558 	relpbuf(bp, &cluster_pbuf_freecnt);
559 }
560 
561 /*
562  *	cluster_wbuild_wb:
563  *
564  *	Implement modified write build for cluster.
565  *
566  *		write_behind = 0	write behind disabled
567  *		write_behind = 1	write behind normal (default)
568  *		write_behind = 2	write behind backed-off
569  */
570 
571 static __inline int
572 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len)
573 {
574 	int r = 0;
575 
576 	switch(write_behind) {
577 	case 2:
578 		if (start_lbn < len)
579 			break;
580 		start_lbn -= len;
581 		/* FALLTHROUGH */
582 	case 1:
583 		r = cluster_wbuild(vp, size, start_lbn, len);
584 		/* FALLTHROUGH */
585 	default:
586 		/* FALLTHROUGH */
587 		break;
588 	}
589 	return(r);
590 }
591 
592 /*
593  * Do clustered write for FFS.
594  *
595  * Three cases:
596  *	1. Write is not sequential (write asynchronously)
597  *	Write is sequential:
598  *	2.	beginning of cluster - begin cluster
599  *	3.	middle of a cluster - add to cluster
600  *	4.	end of a cluster - asynchronously write cluster
601  */
602 void
603 cluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount)
604 {
605 	daddr_t lbn;
606 	int maxclen, cursize;
607 	int lblocksize;
608 	int async;
609 
610 	if (vp->v_type == VREG) {
611 		async = vp->v_mount->mnt_flag & MNT_ASYNC;
612 		lblocksize = vp->v_mount->mnt_stat.f_iosize;
613 	} else {
614 		async = 0;
615 		lblocksize = bp->b_bufsize;
616 	}
617 	lbn = bp->b_lblkno;
618 	KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
619 
620 	/* Initialize vnode to beginning of file. */
621 	if (lbn == 0)
622 		vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
623 
624 	if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
625 	    (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
626 		maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
627 		if (vp->v_clen != 0) {
628 			/*
629 			 * Next block is not sequential.
630 			 *
631 			 * If we are not writing at end of file, the process
632 			 * seeked to another point in the file since its last
633 			 * write, or we have reached our maximum cluster size,
634 			 * then push the previous cluster. Otherwise try
635 			 * reallocating to make it sequential.
636 			 *
637 			 * Change to algorithm: only push previous cluster if
638 			 * it was sequential from the point of view of the
639 			 * seqcount heuristic, otherwise leave the buffer
640 			 * intact so we can potentially optimize the I/O
641 			 * later on in the buf_daemon or update daemon
642 			 * flush.
643 			 */
644 			cursize = vp->v_lastw - vp->v_cstart + 1;
645 			if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
646 			    lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
647 				if (!async && seqcount > 0) {
648 					cluster_wbuild_wb(vp, lblocksize,
649 						vp->v_cstart, cursize);
650 				}
651 			} else {
652 				struct buf **bpp, **endbp;
653 				struct cluster_save *buflist;
654 
655 				buflist = cluster_collectbufs(vp, bp);
656 				endbp = &buflist->bs_children
657 				    [buflist->bs_nchildren - 1];
658 				if (VOP_REALLOCBLKS(vp, buflist)) {
659 					/*
660 					 * Failed, push the previous cluster
661 					 * if *really* writing sequentially
662 					 * in the logical file (seqcount > 1),
663 					 * otherwise delay it in the hopes that
664 					 * the low level disk driver can
665 					 * optimize the write ordering.
666 					 */
667 					for (bpp = buflist->bs_children;
668 					     bpp < endbp; bpp++)
669 						brelse(*bpp);
670 					free(buflist, M_SEGMENT);
671 					if (seqcount > 1) {
672 						cluster_wbuild_wb(vp,
673 						    lblocksize, vp->v_cstart,
674 						    cursize);
675 					}
676 				} else {
677 					/*
678 					 * Succeeded, keep building cluster.
679 					 */
680 					for (bpp = buflist->bs_children;
681 					     bpp <= endbp; bpp++)
682 						bdwrite(*bpp);
683 					free(buflist, M_SEGMENT);
684 					vp->v_lastw = lbn;
685 					vp->v_lasta = bp->b_blkno;
686 					return;
687 				}
688 			}
689 		}
690 		/*
691 		 * Consider beginning a cluster. If at end of file, make
692 		 * cluster as large as possible, otherwise find size of
693 		 * existing cluster.
694 		 */
695 		if ((vp->v_type == VREG) &&
696 			((u_quad_t) bp->b_offset + lblocksize) != filesize &&
697 		    (bp->b_blkno == bp->b_lblkno) &&
698 		    (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
699 		     bp->b_blkno == -1)) {
700 			bawrite(bp);
701 			vp->v_clen = 0;
702 			vp->v_lasta = bp->b_blkno;
703 			vp->v_cstart = lbn + 1;
704 			vp->v_lastw = lbn;
705 			return;
706 		}
707 		vp->v_clen = maxclen;
708 		if (!async && maxclen == 0) {	/* I/O not contiguous */
709 			vp->v_cstart = lbn + 1;
710 			bawrite(bp);
711 		} else {	/* Wait for rest of cluster */
712 			vp->v_cstart = lbn;
713 			bdwrite(bp);
714 		}
715 	} else if (lbn == vp->v_cstart + vp->v_clen) {
716 		/*
717 		 * At end of cluster, write it out if seqcount tells us we
718 		 * are operating sequentially, otherwise let the buf or
719 		 * update daemon handle it.
720 		 */
721 		bdwrite(bp);
722 		if (seqcount > 1)
723 			cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1);
724 		vp->v_clen = 0;
725 		vp->v_cstart = lbn + 1;
726 	} else if (vm_page_count_severe()) {
727 		/*
728 		 * We are low on memory, get it going NOW
729 		 */
730 		bawrite(bp);
731 	} else {
732 		/*
733 		 * In the middle of a cluster, so just delay the I/O for now.
734 		 */
735 		bdwrite(bp);
736 	}
737 	vp->v_lastw = lbn;
738 	vp->v_lasta = bp->b_blkno;
739 }
740 
741 
742 /*
743  * This is an awful lot like cluster_rbuild...wish they could be combined.
744  * The last lbn argument is the current block on which I/O is being
745  * performed.  Check to see that it doesn't fall in the middle of
746  * the current block (if last_bp == NULL).
747  */
748 int
749 cluster_wbuild(vp, size, start_lbn, len)
750 	struct vnode *vp;
751 	long size;
752 	daddr_t start_lbn;
753 	int len;
754 {
755 	struct buf *bp, *tbp;
756 	int i, j, s;
757 	int totalwritten = 0;
758 	int dbsize = btodb(size);
759 
760 	GIANT_REQUIRED;
761 
762 	while (len > 0) {
763 		s = splbio();
764 		/*
765 		 * If the buffer is not delayed-write (i.e. dirty), or it
766 		 * is delayed-write but either locked or inval, it cannot
767 		 * partake in the clustered write.
768 		 */
769 		VI_LOCK(vp);
770 		if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
771 		    (tbp->b_vflags & BV_BKGRDINPROG)) {
772 			VI_UNLOCK(vp);
773 			++start_lbn;
774 			--len;
775 			splx(s);
776 			continue;
777 		}
778 		if (BUF_LOCK(tbp,
779 		    LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, VI_MTX(vp))) {
780 			++start_lbn;
781 			--len;
782 			splx(s);
783 			continue;
784 		}
785 		if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
786 			BUF_UNLOCK(tbp);
787 			++start_lbn;
788 			--len;
789 			splx(s);
790 			continue;
791 		}
792 		bremfree(tbp);
793 		tbp->b_flags &= ~B_DONE;
794 		splx(s);
795 
796 		/*
797 		 * Extra memory in the buffer, punt on this buffer.
798 		 * XXX we could handle this in most cases, but we would
799 		 * have to push the extra memory down to after our max
800 		 * possible cluster size and then potentially pull it back
801 		 * up if the cluster was terminated prematurely--too much
802 		 * hassle.
803 		 */
804 		if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
805 		     (B_CLUSTEROK | B_VMIO)) ||
806 		  (tbp->b_bcount != tbp->b_bufsize) ||
807 		  (tbp->b_bcount != size) ||
808 		  (len == 1) ||
809 		  ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
810 			totalwritten += tbp->b_bufsize;
811 			bawrite(tbp);
812 			++start_lbn;
813 			--len;
814 			continue;
815 		}
816 
817 		/*
818 		 * We got a pbuf to make the cluster in.
819 		 * so initialise it.
820 		 */
821 		TAILQ_INIT(&bp->b_cluster.cluster_head);
822 		bp->b_bcount = 0;
823 		bp->b_bufsize = 0;
824 		bp->b_npages = 0;
825 		if (tbp->b_wcred != NOCRED)
826 			bp->b_wcred = crhold(tbp->b_wcred);
827 
828 		bp->b_blkno = tbp->b_blkno;
829 		bp->b_lblkno = tbp->b_lblkno;
830 		bp->b_offset = tbp->b_offset;
831 
832 		/*
833 		 * We are synthesizing a buffer out of vm_page_t's, but
834 		 * if the block size is not page aligned then the starting
835 		 * address may not be either.  Inherit the b_data offset
836 		 * from the original buffer.
837 		 */
838 		bp->b_data = (char *)((vm_offset_t)bp->b_data |
839 		    ((vm_offset_t)tbp->b_data & PAGE_MASK));
840 		bp->b_flags |= B_CLUSTER |
841 				(tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
842 		bp->b_iodone = cluster_callback;
843 		pbgetvp(vp, bp);
844 		/*
845 		 * From this location in the file, scan forward to see
846 		 * if there are buffers with adjacent data that need to
847 		 * be written as well.
848 		 */
849 		for (i = 0; i < len; ++i, ++start_lbn) {
850 			if (i != 0) { /* If not the first buffer */
851 				s = splbio();
852 				/*
853 				 * If the adjacent data is not even in core it
854 				 * can't need to be written.
855 				 */
856 				VI_LOCK(vp);
857 				if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
858 				    (tbp->b_vflags & BV_BKGRDINPROG)) {
859 					VI_UNLOCK(vp);
860 					splx(s);
861 					break;
862 				}
863 
864 				/*
865 				 * If it IS in core, but has different
866 				 * characteristics, or is locked (which
867 				 * means it could be undergoing a background
868 				 * I/O or be in a weird state), then don't
869 				 * cluster with it.
870 				 */
871 				if (BUF_LOCK(tbp,
872 				    LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
873 				    VI_MTX(vp))) {
874 					splx(s);
875 					break;
876 				}
877 
878 				if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
879 				    B_INVAL | B_DELWRI | B_NEEDCOMMIT))
880 				    != (B_DELWRI | B_CLUSTEROK |
881 				    (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
882 				    tbp->b_wcred != bp->b_wcred) {
883 					BUF_UNLOCK(tbp);
884 					splx(s);
885 					break;
886 				}
887 
888 				/*
889 				 * Check that the combined cluster
890 				 * would make sense with regard to pages
891 				 * and would not be too large
892 				 */
893 				if ((tbp->b_bcount != size) ||
894 				  ((bp->b_blkno + (dbsize * i)) !=
895 				    tbp->b_blkno) ||
896 				  ((tbp->b_npages + bp->b_npages) >
897 				    (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
898 					BUF_UNLOCK(tbp);
899 					splx(s);
900 					break;
901 				}
902 				/*
903 				 * Ok, it's passed all the tests,
904 				 * so remove it from the free list
905 				 * and mark it busy. We will use it.
906 				 */
907 				bremfree(tbp);
908 				tbp->b_flags &= ~B_DONE;
909 				splx(s);
910 			} /* end of code for non-first buffers only */
911 			/* check for latent dependencies to be handled */
912 			if ((LIST_FIRST(&tbp->b_dep)) != NULL) {
913 				tbp->b_iocmd = BIO_WRITE;
914 				buf_start(tbp);
915 			}
916 			/*
917 			 * If the IO is via the VM then we do some
918 			 * special VM hackery (yuck).  Since the buffer's
919 			 * block size may not be page-aligned it is possible
920 			 * for a page to be shared between two buffers.  We
921 			 * have to get rid of the duplication when building
922 			 * the cluster.
923 			 */
924 			if (tbp->b_flags & B_VMIO) {
925 				vm_page_t m;
926 
927 				VM_OBJECT_LOCK(tbp->b_bufobj->bo_object);
928 				if (i != 0) { /* if not first buffer */
929 					for (j = 0; j < tbp->b_npages; j += 1) {
930 						m = tbp->b_pages[j];
931 						if (m->flags & PG_BUSY) {
932 							VM_OBJECT_UNLOCK(
933 							    tbp->b_object);
934 							bqrelse(tbp);
935 							goto finishcluster;
936 						}
937 					}
938 				}
939 				for (j = 0; j < tbp->b_npages; j += 1) {
940 					m = tbp->b_pages[j];
941 					vm_page_io_start(m);
942 					vm_object_pip_add(m->object, 1);
943 					if ((bp->b_npages == 0) ||
944 					  (bp->b_pages[bp->b_npages - 1] != m)) {
945 						bp->b_pages[bp->b_npages] = m;
946 						bp->b_npages++;
947 					}
948 				}
949 				VM_OBJECT_UNLOCK(tbp->b_bufobj->bo_object);
950 			}
951 			bp->b_bcount += size;
952 			bp->b_bufsize += size;
953 
954 			s = splbio();
955 			bundirty(tbp);
956 			tbp->b_flags &= ~B_DONE;
957 			tbp->b_ioflags &= ~BIO_ERROR;
958 			tbp->b_flags |= B_ASYNC;
959 			tbp->b_iocmd = BIO_WRITE;
960 			reassignbuf(tbp);		/* put on clean list */
961 			bufobj_wref(tbp->b_bufobj);
962 			splx(s);
963 			BUF_KERNPROC(tbp);
964 			TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
965 				tbp, b_cluster.cluster_entry);
966 		}
967 	finishcluster:
968 		pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
969 			(vm_page_t *) bp->b_pages, bp->b_npages);
970 		if (bp->b_bufsize > bp->b_kvasize)
971 			panic(
972 			    "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
973 			    bp->b_bufsize, bp->b_kvasize);
974 		bp->b_kvasize = bp->b_bufsize;
975 		totalwritten += bp->b_bufsize;
976 		bp->b_dirtyoff = 0;
977 		bp->b_dirtyend = bp->b_bufsize;
978 		bawrite(bp);
979 
980 		len -= i;
981 	}
982 	return totalwritten;
983 }
984 
985 /*
986  * Collect together all the buffers in a cluster.
987  * Plus add one additional buffer.
988  */
989 static struct cluster_save *
990 cluster_collectbufs(vp, last_bp)
991 	struct vnode *vp;
992 	struct buf *last_bp;
993 {
994 	struct cluster_save *buflist;
995 	struct buf *bp;
996 	daddr_t lbn;
997 	int i, len;
998 
999 	len = vp->v_lastw - vp->v_cstart + 1;
1000 	buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1001 	    M_SEGMENT, M_WAITOK);
1002 	buflist->bs_nchildren = 0;
1003 	buflist->bs_children = (struct buf **) (buflist + 1);
1004 	for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
1005 		(void) bread(vp, lbn, last_bp->b_bcount, NOCRED, &bp);
1006 		buflist->bs_children[i] = bp;
1007 		if (bp->b_blkno == bp->b_lblkno)
1008 			VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno,
1009 				NULL, NULL);
1010 	}
1011 	buflist->bs_children[i] = bp = last_bp;
1012 	if (bp->b_blkno == bp->b_lblkno)
1013 		VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1014 	buflist->bs_nchildren = i + 1;
1015 	return (buflist);
1016 }
1017