3 * The Regents of the University of California. All rights reserved.
4 * Modifications/enhancements:
5 * Copyright (c) 1995 John S. Dyson. All rights reserved.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
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.
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
31 * @(#)vfs_cluster.c 8.7 (Berkeley) 2/13/94
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
37 #include "opt_debug_cluster.h"
39 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/kernel.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>
51 #include <vm/vm_object.h>
52 #include <vm/vm_page.h>
53 #include <sys/sysctl.h>
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");
61 static MALLOC_DEFINE(M_SEGMENT, "cluster_save buffer", "cluster_save buffer");
63 static struct cluster_save *
64 cluster_collectbufs(struct vnode *vp, struct buf *last_bp);
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 static void cluster_callback(struct buf *);
70 static int write_behind = 1;
71 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
72 "Cluster write-behind; 0: disable, 1: enable, 2: backed off");
74 static int read_max = 8;
75 SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0,
76 "Cluster read-ahead max block count");
78 /* Page expended to mark partially backed buffers */
79 extern vm_page_t bogus_page;
82 * Read data to a buf, including read-ahead if we find this to be beneficial.
83 * cluster_read replaces bread.
86 cluster_read(vp, filesize, lblkno, size, cred, totread, seqcount, bpp)
96 struct buf *bp, *rbp, *reqbp;
97 daddr_t blkno, origblkno;
105 * Try to limit the amount of read-ahead by a few
106 * ad-hoc parameters. This needs work!!!
108 racluster = vp->v_mount->mnt_iosize_max / size;
110 maxra = min(read_max, maxra);
111 maxra = min(nbuf/8, maxra);
112 if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize)
113 maxra = (filesize / size) - lblkno;
116 * get the requested block
118 *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0, 0);
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.
126 if (bp->b_flags & B_CACHE) {
129 } else if ((bp->b_flags & B_RAM) == 0) {
132 bp->b_flags &= ~B_RAM;
134 for (i = 1; i < maxra; i++) {
136 * Stop if the buffer does not exist or it
137 * is invalid (about to go away?)
139 rbp = gbincore(&vp->v_bufobj, lblkno+i);
140 if (rbp == NULL || (rbp->b_flags & B_INVAL))
144 * Set another read-ahead mark so we know
147 if (((i % racluster) == (racluster - 1)) ||
149 rbp->b_flags |= B_RAM;
159 * If it isn't in the cache, then get a chunk from
160 * disk if sequential, otherwise just get the block.
163 off_t firstread = bp->b_offset;
166 KASSERT(bp->b_offset != NOOFFSET,
167 ("cluster_read: no buffer offset"));
172 * Compute the total number of blocks that we should read
175 if (firstread + totread > filesize)
176 totread = filesize - firstread;
177 nblks = howmany(totread, size);
178 if (nblks > racluster)
182 * Now compute the number of contiguous blocks.
185 error = VOP_BMAP(vp, lblkno, NULL,
186 &blkno, &ncontig, NULL);
188 * If this failed to map just do the original block.
190 if (error || blkno == -1)
195 * If we have contiguous data available do a cluster
196 * otherwise just read the requested block.
199 /* Account for our first block. */
200 ncontig = min(ncontig + 1, nblks);
203 bp = cluster_rbuild(vp, filesize, lblkno,
204 blkno, size, nblks, bp);
205 lblkno += (bp->b_bufsize / size);
207 bp->b_flags |= B_RAM;
208 bp->b_iocmd = BIO_READ;
214 * handle the synchronous read so that it is available ASAP.
217 if ((bp->b_flags & B_CLUSTER) == 0) {
218 vfs_busy_pages(bp, 0);
220 bp->b_flags &= ~B_INVAL;
221 bp->b_ioflags &= ~BIO_ERROR;
222 if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL)
224 bp->b_iooffset = dbtob(bp->b_blkno);
226 curproc->p_stats->p_ru.ru_inblock++;
230 * If we have been doing sequential I/O, then do some read-ahead.
232 while (lblkno < (origblkno + maxra)) {
233 error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL);
241 * We could throttle ncontig here by maxra but we might as
242 * well read the data if it is contiguous. We're throttled
243 * by racluster anyway.
246 ncontig = min(ncontig + 1, racluster);
247 rbp = cluster_rbuild(vp, filesize, lblkno, blkno,
248 size, ncontig, NULL);
249 lblkno += (rbp->b_bufsize / size);
250 if (rbp->b_flags & B_DELWRI) {
255 rbp = getblk(vp, lblkno, size, 0, 0, 0);
257 if (rbp->b_flags & B_DELWRI) {
261 rbp->b_flags |= B_ASYNC | B_RAM;
262 rbp->b_iocmd = BIO_READ;
263 rbp->b_blkno = blkno;
265 if (rbp->b_flags & B_CACHE) {
266 rbp->b_flags &= ~B_ASYNC;
270 if ((rbp->b_flags & B_CLUSTER) == 0) {
271 vfs_busy_pages(rbp, 0);
273 rbp->b_flags &= ~B_INVAL;
274 rbp->b_ioflags &= ~BIO_ERROR;
275 if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL)
277 rbp->b_iooffset = dbtob(rbp->b_blkno);
279 curproc->p_stats->p_ru.ru_inblock++;
283 return (bufwait(reqbp));
289 * If blocks are contiguous on disk, use this to provide clustered
290 * read ahead. We will read as many blocks as possible sequentially
291 * and then parcel them up into logical blocks in the buffer hash table.
294 cluster_rbuild(vp, filesize, lbn, blkno, size, run, fbp)
303 struct buf *bp, *tbp;
307 KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
308 ("cluster_rbuild: size %ld != filesize %jd\n",
309 size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
314 while ((u_quad_t) size * (lbn + run) > filesize) {
320 tbp->b_iocmd = BIO_READ;
322 tbp = getblk(vp, lbn, size, 0, 0, 0);
323 if (tbp->b_flags & B_CACHE)
325 tbp->b_flags |= B_ASYNC | B_RAM;
326 tbp->b_iocmd = BIO_READ;
329 tbp->b_blkno = blkno;
330 if( (tbp->b_flags & B_MALLOC) ||
331 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
334 bp = trypbuf(&cluster_pbuf_freecnt);
339 * We are synthesizing a buffer out of vm_page_t's, but
340 * if the block size is not page aligned then the starting
341 * address may not be either. Inherit the b_data offset
342 * from the original buffer.
344 bp->b_data = (char *)((vm_offset_t)bp->b_data |
345 ((vm_offset_t)tbp->b_data & PAGE_MASK));
346 bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO;
347 bp->b_iocmd = BIO_READ;
348 bp->b_iodone = cluster_callback;
351 bp->b_offset = tbp->b_offset;
352 KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
355 TAILQ_INIT(&bp->b_cluster.cluster_head);
362 for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
364 if ((bp->b_npages * PAGE_SIZE) +
365 round_page(size) > vp->v_mount->mnt_iosize_max) {
369 tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT);
371 /* Don't wait around for locked bufs. */
376 * Stop scanning if the buffer is fully valid
377 * (marked B_CACHE), or locked (may be doing a
378 * background write), or if the buffer is not
379 * VMIO backed. The clustering code can only deal
380 * with VMIO-backed buffers.
383 if ((tbp->b_vflags & BV_BKGRDINPROG) ||
384 (tbp->b_flags & B_CACHE) ||
385 (tbp->b_flags & B_VMIO) == 0) {
393 * The buffer must be completely invalid in order to
394 * take part in the cluster. If it is partially valid
397 VM_OBJECT_LOCK(tbp->b_bufobj->bo_object);
398 for (j = 0;j < tbp->b_npages; j++) {
399 VM_OBJECT_LOCK_ASSERT(tbp->b_pages[j]->object,
401 if (tbp->b_pages[j]->valid)
404 VM_OBJECT_UNLOCK(tbp->b_bufobj->bo_object);
405 if (j != tbp->b_npages) {
411 * Set a read-ahead mark as appropriate
413 if ((fbp && (i == 1)) || (i == (run - 1)))
414 tbp->b_flags |= B_RAM;
417 * Set the buffer up for an async read (XXX should
418 * we do this only if we do not wind up brelse()ing?).
419 * Set the block number if it isn't set, otherwise
420 * if it is make sure it matches the block number we
423 tbp->b_flags |= B_ASYNC;
424 tbp->b_iocmd = BIO_READ;
425 if (tbp->b_blkno == tbp->b_lblkno) {
427 } else if (tbp->b_blkno != bn) {
433 * XXX fbp from caller may not be B_ASYNC, but we are going
434 * to biodone() it in cluster_callback() anyway
437 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
438 tbp, b_cluster.cluster_entry);
439 VM_OBJECT_LOCK(tbp->b_bufobj->bo_object);
440 for (j = 0; j < tbp->b_npages; j += 1) {
444 vm_object_pip_add(m->object, 1);
445 if ((bp->b_npages == 0) ||
446 (bp->b_pages[bp->b_npages-1] != m)) {
447 bp->b_pages[bp->b_npages] = m;
450 if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
451 tbp->b_pages[j] = bogus_page;
453 VM_OBJECT_UNLOCK(tbp->b_bufobj->bo_object);
455 * XXX shouldn't this be += size for both, like in
458 * Don't inherit tbp->b_bufsize as it may be larger due to
459 * a non-page-aligned size. Instead just aggregate using
462 if (tbp->b_bcount != size)
463 printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
464 if (tbp->b_bufsize != size)
465 printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
466 bp->b_bcount += size;
467 bp->b_bufsize += size;
471 * Fully valid pages in the cluster are already good and do not need
472 * to be re-read from disk. Replace the page with bogus_page
474 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
475 for (j = 0; j < bp->b_npages; j++) {
476 VM_OBJECT_LOCK_ASSERT(bp->b_pages[j]->object, MA_OWNED);
477 if ((bp->b_pages[j]->valid & VM_PAGE_BITS_ALL) ==
479 bp->b_pages[j] = bogus_page;
482 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
483 if (bp->b_bufsize > bp->b_kvasize)
484 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
485 bp->b_bufsize, bp->b_kvasize);
486 bp->b_kvasize = bp->b_bufsize;
488 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
489 (vm_page_t *)bp->b_pages, bp->b_npages);
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.
503 struct buf *nbp, *tbp;
507 * Must propogate errors to all the components.
509 if (bp->b_ioflags & BIO_ERROR)
512 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
514 * Move memory from the large cluster buffer into the component
515 * buffers and mark IO as done on these.
517 for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
519 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
521 tbp->b_ioflags |= BIO_ERROR;
522 tbp->b_error = error;
524 tbp->b_dirtyoff = tbp->b_dirtyend = 0;
525 tbp->b_flags &= ~B_INVAL;
526 tbp->b_ioflags &= ~BIO_ERROR;
528 * XXX the bdwrite()/bqrelse() issued during
529 * cluster building clears B_RELBUF (see bqrelse()
530 * comment). If direct I/O was specified, we have
531 * to restore it here to allow the buffer and VM
534 if (tbp->b_flags & B_DIRECT)
535 tbp->b_flags |= B_RELBUF;
540 relpbuf(bp, &cluster_pbuf_freecnt);
546 * Implement modified write build for cluster.
548 * write_behind = 0 write behind disabled
549 * write_behind = 1 write behind normal (default)
550 * write_behind = 2 write behind backed-off
554 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len)
558 switch(write_behind) {
565 r = cluster_wbuild(vp, size, start_lbn, len);
575 * Do clustered write for FFS.
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
585 cluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount)
588 int maxclen, cursize;
592 if (vp->v_type == VREG) {
593 async = vp->v_mount->mnt_flag & MNT_ASYNC;
594 lblocksize = vp->v_mount->mnt_stat.f_iosize;
597 lblocksize = bp->b_bufsize;
600 KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
602 /* Initialize vnode to beginning of file. */
604 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
606 if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
607 (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
608 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
609 if (vp->v_clen != 0) {
611 * Next block is not sequential.
613 * If we are not writing at end of file, the process
614 * seeked to another point in the file since its last
615 * write, or we have reached our maximum cluster size,
616 * then push the previous cluster. Otherwise try
617 * reallocating to make it sequential.
619 * Change to algorithm: only push previous cluster if
620 * it was sequential from the point of view of the
621 * seqcount heuristic, otherwise leave the buffer
622 * intact so we can potentially optimize the I/O
623 * later on in the buf_daemon or update daemon
626 cursize = vp->v_lastw - vp->v_cstart + 1;
627 if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
628 lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
629 if (!async && seqcount > 0) {
630 cluster_wbuild_wb(vp, lblocksize,
631 vp->v_cstart, cursize);
634 struct buf **bpp, **endbp;
635 struct cluster_save *buflist;
637 buflist = cluster_collectbufs(vp, bp);
638 endbp = &buflist->bs_children
639 [buflist->bs_nchildren - 1];
640 if (VOP_REALLOCBLKS(vp, buflist)) {
642 * Failed, push the previous cluster
643 * if *really* writing sequentially
644 * in the logical file (seqcount > 1),
645 * otherwise delay it in the hopes that
646 * the low level disk driver can
647 * optimize the write ordering.
649 for (bpp = buflist->bs_children;
652 free(buflist, M_SEGMENT);
654 cluster_wbuild_wb(vp,
655 lblocksize, vp->v_cstart,
660 * Succeeded, keep building cluster.
662 for (bpp = buflist->bs_children;
665 free(buflist, M_SEGMENT);
667 vp->v_lasta = bp->b_blkno;
673 * Consider beginning a cluster. If at end of file, make
674 * cluster as large as possible, otherwise find size of
677 if ((vp->v_type == VREG) &&
678 ((u_quad_t) bp->b_offset + lblocksize) != filesize &&
679 (bp->b_blkno == bp->b_lblkno) &&
680 (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
681 bp->b_blkno == -1)) {
684 vp->v_lasta = bp->b_blkno;
685 vp->v_cstart = lbn + 1;
689 vp->v_clen = maxclen;
690 if (!async && maxclen == 0) { /* I/O not contiguous */
691 vp->v_cstart = lbn + 1;
693 } else { /* Wait for rest of cluster */
697 } else if (lbn == vp->v_cstart + vp->v_clen) {
699 * At end of cluster, write it out if seqcount tells us we
700 * are operating sequentially, otherwise let the buf or
701 * update daemon handle it.
705 cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1);
707 vp->v_cstart = lbn + 1;
708 } else if (vm_page_count_severe()) {
710 * We are low on memory, get it going NOW
715 * In the middle of a cluster, so just delay the I/O for now.
720 vp->v_lasta = bp->b_blkno;
725 * This is an awful lot like cluster_rbuild...wish they could be combined.
726 * The last lbn argument is the current block on which I/O is being
727 * performed. Check to see that it doesn't fall in the middle of
728 * the current block (if last_bp == NULL).
731 cluster_wbuild(vp, size, start_lbn, len)
737 struct buf *bp, *tbp;
739 int totalwritten = 0;
740 int dbsize = btodb(size);
744 * If the buffer is not delayed-write (i.e. dirty), or it
745 * is delayed-write but either locked or inval, it cannot
746 * partake in the clustered write.
749 if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
750 (tbp->b_vflags & BV_BKGRDINPROG)) {
757 LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, VI_MTX(vp))) {
762 if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
769 tbp->b_flags &= ~B_DONE;
772 * Extra memory in the buffer, punt on this buffer.
773 * XXX we could handle this in most cases, but we would
774 * have to push the extra memory down to after our max
775 * possible cluster size and then potentially pull it back
776 * up if the cluster was terminated prematurely--too much
779 if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
780 (B_CLUSTEROK | B_VMIO)) ||
781 (tbp->b_bcount != tbp->b_bufsize) ||
782 (tbp->b_bcount != size) ||
784 ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
785 totalwritten += tbp->b_bufsize;
793 * We got a pbuf to make the cluster in.
796 TAILQ_INIT(&bp->b_cluster.cluster_head);
800 if (tbp->b_wcred != NOCRED)
801 bp->b_wcred = crhold(tbp->b_wcred);
803 bp->b_blkno = tbp->b_blkno;
804 bp->b_lblkno = tbp->b_lblkno;
805 bp->b_offset = tbp->b_offset;
808 * We are synthesizing a buffer out of vm_page_t's, but
809 * if the block size is not page aligned then the starting
810 * address may not be either. Inherit the b_data offset
811 * from the original buffer.
813 bp->b_data = (char *)((vm_offset_t)bp->b_data |
814 ((vm_offset_t)tbp->b_data & PAGE_MASK));
815 bp->b_flags |= B_CLUSTER |
816 (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
817 bp->b_iodone = cluster_callback;
820 * From this location in the file, scan forward to see
821 * if there are buffers with adjacent data that need to
822 * be written as well.
824 for (i = 0; i < len; ++i, ++start_lbn) {
825 if (i != 0) { /* If not the first buffer */
827 * If the adjacent data is not even in core it
828 * can't need to be written.
831 if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
832 (tbp->b_vflags & BV_BKGRDINPROG)) {
838 * If it IS in core, but has different
839 * characteristics, or is locked (which
840 * means it could be undergoing a background
841 * I/O or be in a weird state), then don't
845 LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
849 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
850 B_INVAL | B_DELWRI | B_NEEDCOMMIT))
851 != (B_DELWRI | B_CLUSTEROK |
852 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
853 tbp->b_wcred != bp->b_wcred) {
859 * Check that the combined cluster
860 * would make sense with regard to pages
861 * and would not be too large
863 if ((tbp->b_bcount != size) ||
864 ((bp->b_blkno + (dbsize * i)) !=
866 ((tbp->b_npages + bp->b_npages) >
867 (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
872 * Ok, it's passed all the tests,
873 * so remove it from the free list
874 * and mark it busy. We will use it.
877 tbp->b_flags &= ~B_DONE;
878 } /* end of code for non-first buffers only */
879 /* check for latent dependencies to be handled */
880 if ((LIST_FIRST(&tbp->b_dep)) != NULL) {
881 tbp->b_iocmd = BIO_WRITE;
885 * If the IO is via the VM then we do some
886 * special VM hackery (yuck). Since the buffer's
887 * block size may not be page-aligned it is possible
888 * for a page to be shared between two buffers. We
889 * have to get rid of the duplication when building
892 if (tbp->b_flags & B_VMIO) {
895 VM_OBJECT_LOCK(tbp->b_bufobj->bo_object);
896 if (i != 0) { /* if not first buffer */
897 for (j = 0; j < tbp->b_npages; j += 1) {
899 if (m->flags & PG_BUSY) {
907 for (j = 0; j < tbp->b_npages; j += 1) {
910 vm_object_pip_add(m->object, 1);
911 if ((bp->b_npages == 0) ||
912 (bp->b_pages[bp->b_npages - 1] != m)) {
913 bp->b_pages[bp->b_npages] = m;
917 VM_OBJECT_UNLOCK(tbp->b_bufobj->bo_object);
919 bp->b_bcount += size;
920 bp->b_bufsize += size;
922 tbp->b_flags &= ~B_DONE;
923 tbp->b_ioflags &= ~BIO_ERROR;
924 tbp->b_flags |= B_ASYNC;
925 tbp->b_iocmd = BIO_WRITE;
926 reassignbuf(tbp); /* put on clean list */
927 bufobj_wref(tbp->b_bufobj);
929 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
930 tbp, b_cluster.cluster_entry);
933 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
934 (vm_page_t *) bp->b_pages, bp->b_npages);
935 if (bp->b_bufsize > bp->b_kvasize)
937 "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
938 bp->b_bufsize, bp->b_kvasize);
939 bp->b_kvasize = bp->b_bufsize;
940 totalwritten += bp->b_bufsize;
942 bp->b_dirtyend = bp->b_bufsize;
951 * Collect together all the buffers in a cluster.
952 * Plus add one additional buffer.
954 static struct cluster_save *
955 cluster_collectbufs(vp, last_bp)
959 struct cluster_save *buflist;
964 len = vp->v_lastw - vp->v_cstart + 1;
965 buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
966 M_SEGMENT, M_WAITOK);
967 buflist->bs_nchildren = 0;
968 buflist->bs_children = (struct buf **) (buflist + 1);
969 for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
970 (void) bread(vp, lbn, last_bp->b_bcount, NOCRED, &bp);
971 buflist->bs_children[i] = bp;
972 if (bp->b_blkno == bp->b_lblkno)
973 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno,
976 buflist->bs_children[i] = bp = last_bp;
977 if (bp->b_blkno == bp->b_lblkno)
978 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
979 buflist->bs_nchildren = i + 1;