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 * Number of physical bufs (pbufs) this subsystem is allowed.
83 * Manipulated by vm_pager.c
85 extern int cluster_pbuf_freecnt;
88 * Read data to a buf, including read-ahead if we find this to be beneficial.
89 * cluster_read replaces bread.
92 cluster_read(vp, filesize, lblkno, size, cred, totread, seqcount, bpp)
102 struct buf *bp, *rbp, *reqbp;
103 daddr_t blkno, origblkno;
104 int maxra, racluster;
111 * Try to limit the amount of read-ahead by a few
112 * ad-hoc parameters. This needs work!!!
114 racluster = vp->v_mount->mnt_iosize_max / size;
116 maxra = min(read_max, maxra);
117 maxra = min(nbuf/8, maxra);
118 if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize)
119 maxra = (filesize / size) - lblkno;
122 * get the requested block
124 *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0, 0);
128 * if it is in the cache, then check to see if the reads have been
129 * sequential. If they have, then try some read-ahead, otherwise
130 * back-off on prospective read-aheads.
132 if (bp->b_flags & B_CACHE) {
135 } else if ((bp->b_flags & B_RAM) == 0) {
138 bp->b_flags &= ~B_RAM;
140 for (i = 1; i < maxra; i++) {
142 * Stop if the buffer does not exist or it
143 * is invalid (about to go away?)
145 rbp = gbincore(&vp->v_bufobj, lblkno+i);
146 if (rbp == NULL || (rbp->b_flags & B_INVAL))
150 * Set another read-ahead mark so we know
153 if (((i % racluster) == (racluster - 1)) ||
155 rbp->b_flags |= B_RAM;
165 * If it isn't in the cache, then get a chunk from
166 * disk if sequential, otherwise just get the block.
169 off_t firstread = bp->b_offset;
172 KASSERT(bp->b_offset != NOOFFSET,
173 ("cluster_read: no buffer offset"));
178 * Compute the total number of blocks that we should read
181 if (firstread + totread > filesize)
182 totread = filesize - firstread;
183 nblks = howmany(totread, size);
184 if (nblks > racluster)
188 * Now compute the number of contiguous blocks.
191 error = VOP_BMAP(vp, lblkno, NULL,
192 &blkno, &ncontig, NULL);
194 * If this failed to map just do the original block.
196 if (error || blkno == -1)
201 * If we have contiguous data available do a cluster
202 * otherwise just read the requested block.
205 /* Account for our first block. */
206 ncontig = min(ncontig + 1, nblks);
209 bp = cluster_rbuild(vp, filesize, lblkno,
210 blkno, size, nblks, bp);
211 lblkno += (bp->b_bufsize / size);
213 bp->b_flags |= B_RAM;
214 bp->b_iocmd = BIO_READ;
220 * handle the synchronous read so that it is available ASAP.
223 if ((bp->b_flags & B_CLUSTER) == 0) {
224 vfs_busy_pages(bp, 0);
226 bp->b_flags &= ~B_INVAL;
227 bp->b_ioflags &= ~BIO_ERROR;
228 if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL)
230 bp->b_iooffset = dbtob(bp->b_blkno);
232 curproc->p_stats->p_ru.ru_inblock++;
236 * If we have been doing sequential I/O, then do some read-ahead.
238 while (lblkno < (origblkno + maxra)) {
239 error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL);
247 * We could throttle ncontig here by maxra but we might as
248 * well read the data if it is contiguous. We're throttled
249 * by racluster anyway.
252 ncontig = min(ncontig + 1, racluster);
253 rbp = cluster_rbuild(vp, filesize, lblkno, blkno,
254 size, ncontig, NULL);
255 lblkno += (rbp->b_bufsize / size);
256 if (rbp->b_flags & B_DELWRI) {
261 rbp = getblk(vp, lblkno, size, 0, 0, 0);
263 if (rbp->b_flags & B_DELWRI) {
267 rbp->b_flags |= B_ASYNC | B_RAM;
268 rbp->b_iocmd = BIO_READ;
269 rbp->b_blkno = blkno;
271 if (rbp->b_flags & B_CACHE) {
272 rbp->b_flags &= ~B_ASYNC;
276 if ((rbp->b_flags & B_CLUSTER) == 0) {
277 vfs_busy_pages(rbp, 0);
279 rbp->b_flags &= ~B_INVAL;
280 rbp->b_ioflags &= ~BIO_ERROR;
281 if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL)
283 rbp->b_iooffset = dbtob(rbp->b_blkno);
285 curproc->p_stats->p_ru.ru_inblock++;
289 return (bufwait(reqbp));
295 * If blocks are contiguous on disk, use this to provide clustered
296 * read ahead. We will read as many blocks as possible sequentially
297 * and then parcel them up into logical blocks in the buffer hash table.
300 cluster_rbuild(vp, filesize, lbn, blkno, size, run, fbp)
309 struct buf *bp, *tbp;
313 KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
314 ("cluster_rbuild: size %ld != filesize %jd\n",
315 size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
320 while ((u_quad_t) size * (lbn + run) > filesize) {
326 tbp->b_iocmd = BIO_READ;
328 tbp = getblk(vp, lbn, size, 0, 0, 0);
329 if (tbp->b_flags & B_CACHE)
331 tbp->b_flags |= B_ASYNC | B_RAM;
332 tbp->b_iocmd = BIO_READ;
335 tbp->b_blkno = blkno;
336 if( (tbp->b_flags & B_MALLOC) ||
337 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
340 bp = trypbuf(&cluster_pbuf_freecnt);
345 * We are synthesizing a buffer out of vm_page_t's, but
346 * if the block size is not page aligned then the starting
347 * address may not be either. Inherit the b_data offset
348 * from the original buffer.
350 bp->b_data = (char *)((vm_offset_t)bp->b_data |
351 ((vm_offset_t)tbp->b_data & PAGE_MASK));
352 bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO;
353 bp->b_iocmd = BIO_READ;
354 bp->b_iodone = cluster_callback;
357 bp->b_offset = tbp->b_offset;
358 KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
361 TAILQ_INIT(&bp->b_cluster.cluster_head);
368 for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
370 if ((bp->b_npages * PAGE_SIZE) +
371 round_page(size) > vp->v_mount->mnt_iosize_max) {
375 tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT);
377 /* Don't wait around for locked bufs. */
382 * Stop scanning if the buffer is fully valid
383 * (marked B_CACHE), or locked (may be doing a
384 * background write), or if the buffer is not
385 * VMIO backed. The clustering code can only deal
386 * with VMIO-backed buffers.
389 if ((tbp->b_vflags & BV_BKGRDINPROG) ||
390 (tbp->b_flags & B_CACHE) ||
391 (tbp->b_flags & B_VMIO) == 0) {
399 * The buffer must be completely invalid in order to
400 * take part in the cluster. If it is partially valid
403 VM_OBJECT_LOCK(tbp->b_bufobj->bo_object);
404 for (j = 0;j < tbp->b_npages; j++) {
405 VM_OBJECT_LOCK_ASSERT(tbp->b_pages[j]->object,
407 if (tbp->b_pages[j]->valid)
410 VM_OBJECT_UNLOCK(tbp->b_bufobj->bo_object);
411 if (j != tbp->b_npages) {
417 * Set a read-ahead mark as appropriate
419 if ((fbp && (i == 1)) || (i == (run - 1)))
420 tbp->b_flags |= B_RAM;
423 * Set the buffer up for an async read (XXX should
424 * we do this only if we do not wind up brelse()ing?).
425 * Set the block number if it isn't set, otherwise
426 * if it is make sure it matches the block number we
429 tbp->b_flags |= B_ASYNC;
430 tbp->b_iocmd = BIO_READ;
431 if (tbp->b_blkno == tbp->b_lblkno) {
433 } else if (tbp->b_blkno != bn) {
439 * XXX fbp from caller may not be B_ASYNC, but we are going
440 * to biodone() it in cluster_callback() anyway
443 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
444 tbp, b_cluster.cluster_entry);
445 VM_OBJECT_LOCK(tbp->b_bufobj->bo_object);
446 for (j = 0; j < tbp->b_npages; j += 1) {
450 vm_object_pip_add(m->object, 1);
451 if ((bp->b_npages == 0) ||
452 (bp->b_pages[bp->b_npages-1] != m)) {
453 bp->b_pages[bp->b_npages] = m;
456 if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
457 tbp->b_pages[j] = bogus_page;
459 VM_OBJECT_UNLOCK(tbp->b_bufobj->bo_object);
461 * XXX shouldn't this be += size for both, like in
464 * Don't inherit tbp->b_bufsize as it may be larger due to
465 * a non-page-aligned size. Instead just aggregate using
468 if (tbp->b_bcount != size)
469 printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
470 if (tbp->b_bufsize != size)
471 printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
472 bp->b_bcount += size;
473 bp->b_bufsize += size;
477 * Fully valid pages in the cluster are already good and do not need
478 * to be re-read from disk. Replace the page with bogus_page
480 VM_OBJECT_LOCK(bp->b_bufobj->bo_object);
481 for (j = 0; j < bp->b_npages; j++) {
482 VM_OBJECT_LOCK_ASSERT(bp->b_pages[j]->object, MA_OWNED);
483 if ((bp->b_pages[j]->valid & VM_PAGE_BITS_ALL) ==
485 bp->b_pages[j] = bogus_page;
488 VM_OBJECT_UNLOCK(bp->b_bufobj->bo_object);
489 if (bp->b_bufsize > bp->b_kvasize)
490 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
491 bp->b_bufsize, bp->b_kvasize);
492 bp->b_kvasize = bp->b_bufsize;
494 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
495 (vm_page_t *)bp->b_pages, bp->b_npages);
500 * Cleanup after a clustered read or write.
501 * This is complicated by the fact that any of the buffers might have
502 * extra memory (if there were no empty buffer headers at allocbuf time)
503 * that we will need to shift around.
509 struct buf *nbp, *tbp;
513 * Must propogate errors to all the components.
515 if (bp->b_ioflags & BIO_ERROR)
519 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_npages);
522 * Move memory from the large cluster buffer into the component
523 * buffers and mark IO as done on these.
525 for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
527 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
529 tbp->b_ioflags |= BIO_ERROR;
530 tbp->b_error = error;
532 tbp->b_dirtyoff = tbp->b_dirtyend = 0;
533 tbp->b_flags &= ~B_INVAL;
534 tbp->b_ioflags &= ~BIO_ERROR;
536 * XXX the bdwrite()/bqrelse() issued during
537 * cluster building clears B_RELBUF (see bqrelse()
538 * comment). If direct I/O was specified, we have
539 * to restore it here to allow the buffer and VM
542 if (tbp->b_flags & B_DIRECT)
543 tbp->b_flags |= B_RELBUF;
548 relpbuf(bp, &cluster_pbuf_freecnt);
554 * Implement modified write build for cluster.
556 * write_behind = 0 write behind disabled
557 * write_behind = 1 write behind normal (default)
558 * write_behind = 2 write behind backed-off
562 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len)
566 switch(write_behind) {
573 r = cluster_wbuild(vp, size, start_lbn, len);
583 * Do clustered write for FFS.
586 * 1. Write is not sequential (write asynchronously)
587 * Write is sequential:
588 * 2. beginning of cluster - begin cluster
589 * 3. middle of a cluster - add to cluster
590 * 4. end of a cluster - asynchronously write cluster
593 cluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount)
596 int maxclen, cursize;
600 if (vp->v_type == VREG) {
601 async = vp->v_mount->mnt_flag & MNT_ASYNC;
602 lblocksize = vp->v_mount->mnt_stat.f_iosize;
605 lblocksize = bp->b_bufsize;
608 KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
610 /* Initialize vnode to beginning of file. */
612 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
614 if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
615 (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
616 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
617 if (vp->v_clen != 0) {
619 * Next block is not sequential.
621 * If we are not writing at end of file, the process
622 * seeked to another point in the file since its last
623 * write, or we have reached our maximum cluster size,
624 * then push the previous cluster. Otherwise try
625 * reallocating to make it sequential.
627 * Change to algorithm: only push previous cluster if
628 * it was sequential from the point of view of the
629 * seqcount heuristic, otherwise leave the buffer
630 * intact so we can potentially optimize the I/O
631 * later on in the buf_daemon or update daemon
634 cursize = vp->v_lastw - vp->v_cstart + 1;
635 if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
636 lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
637 if (!async && seqcount > 0) {
638 cluster_wbuild_wb(vp, lblocksize,
639 vp->v_cstart, cursize);
642 struct buf **bpp, **endbp;
643 struct cluster_save *buflist;
645 buflist = cluster_collectbufs(vp, bp);
646 endbp = &buflist->bs_children
647 [buflist->bs_nchildren - 1];
648 if (VOP_REALLOCBLKS(vp, buflist)) {
650 * Failed, push the previous cluster
651 * if *really* writing sequentially
652 * in the logical file (seqcount > 1),
653 * otherwise delay it in the hopes that
654 * the low level disk driver can
655 * optimize the write ordering.
657 for (bpp = buflist->bs_children;
660 free(buflist, M_SEGMENT);
662 cluster_wbuild_wb(vp,
663 lblocksize, vp->v_cstart,
668 * Succeeded, keep building cluster.
670 for (bpp = buflist->bs_children;
673 free(buflist, M_SEGMENT);
675 vp->v_lasta = bp->b_blkno;
681 * Consider beginning a cluster. If at end of file, make
682 * cluster as large as possible, otherwise find size of
685 if ((vp->v_type == VREG) &&
686 ((u_quad_t) bp->b_offset + lblocksize) != filesize &&
687 (bp->b_blkno == bp->b_lblkno) &&
688 (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
689 bp->b_blkno == -1)) {
692 vp->v_lasta = bp->b_blkno;
693 vp->v_cstart = lbn + 1;
697 vp->v_clen = maxclen;
698 if (!async && maxclen == 0) { /* I/O not contiguous */
699 vp->v_cstart = lbn + 1;
701 } else { /* Wait for rest of cluster */
705 } else if (lbn == vp->v_cstart + vp->v_clen) {
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.
713 cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1);
715 vp->v_cstart = lbn + 1;
716 } else if (vm_page_count_severe()) {
718 * We are low on memory, get it going NOW
723 * In the middle of a cluster, so just delay the I/O for now.
728 vp->v_lasta = bp->b_blkno;
733 * This is an awful lot like cluster_rbuild...wish they could be combined.
734 * The last lbn argument is the current block on which I/O is being
735 * performed. Check to see that it doesn't fall in the middle of
736 * the current block (if last_bp == NULL).
739 cluster_wbuild(vp, size, start_lbn, len)
745 struct buf *bp, *tbp;
747 int totalwritten = 0;
748 int dbsize = btodb(size);
752 * If the buffer is not delayed-write (i.e. dirty), or it
753 * is delayed-write but either locked or inval, it cannot
754 * partake in the clustered write.
757 if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
758 (tbp->b_vflags & BV_BKGRDINPROG)) {
765 LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, VI_MTX(vp))) {
770 if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
777 tbp->b_flags &= ~B_DONE;
780 * Extra memory in the buffer, punt on this buffer.
781 * XXX we could handle this in most cases, but we would
782 * have to push the extra memory down to after our max
783 * possible cluster size and then potentially pull it back
784 * up if the cluster was terminated prematurely--too much
787 if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
788 (B_CLUSTEROK | B_VMIO)) ||
789 (tbp->b_bcount != tbp->b_bufsize) ||
790 (tbp->b_bcount != size) ||
792 ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
793 totalwritten += tbp->b_bufsize;
801 * We got a pbuf to make the cluster in.
804 TAILQ_INIT(&bp->b_cluster.cluster_head);
808 if (tbp->b_wcred != NOCRED)
809 bp->b_wcred = crhold(tbp->b_wcred);
811 bp->b_blkno = tbp->b_blkno;
812 bp->b_lblkno = tbp->b_lblkno;
813 bp->b_offset = tbp->b_offset;
816 * We are synthesizing a buffer out of vm_page_t's, but
817 * if the block size is not page aligned then the starting
818 * address may not be either. Inherit the b_data offset
819 * from the original buffer.
821 bp->b_data = (char *)((vm_offset_t)bp->b_data |
822 ((vm_offset_t)tbp->b_data & PAGE_MASK));
823 bp->b_flags |= B_CLUSTER |
824 (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
825 bp->b_iodone = cluster_callback;
828 * From this location in the file, scan forward to see
829 * if there are buffers with adjacent data that need to
830 * be written as well.
832 for (i = 0; i < len; ++i, ++start_lbn) {
833 if (i != 0) { /* If not the first buffer */
835 * If the adjacent data is not even in core it
836 * can't need to be written.
839 if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
840 (tbp->b_vflags & BV_BKGRDINPROG)) {
846 * If it IS in core, but has different
847 * characteristics, or is locked (which
848 * means it could be undergoing a background
849 * I/O or be in a weird state), then don't
853 LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
857 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
858 B_INVAL | B_DELWRI | B_NEEDCOMMIT))
859 != (B_DELWRI | B_CLUSTEROK |
860 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
861 tbp->b_wcred != bp->b_wcred) {
867 * Check that the combined cluster
868 * would make sense with regard to pages
869 * and would not be too large
871 if ((tbp->b_bcount != size) ||
872 ((bp->b_blkno + (dbsize * i)) !=
874 ((tbp->b_npages + bp->b_npages) >
875 (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
880 * Ok, it's passed all the tests,
881 * so remove it from the free list
882 * and mark it busy. We will use it.
885 tbp->b_flags &= ~B_DONE;
886 } /* end of code for non-first buffers only */
887 /* check for latent dependencies to be handled */
888 if ((LIST_FIRST(&tbp->b_dep)) != NULL) {
889 tbp->b_iocmd = BIO_WRITE;
893 * If the IO is via the VM then we do some
894 * special VM hackery (yuck). Since the buffer's
895 * block size may not be page-aligned it is possible
896 * for a page to be shared between two buffers. We
897 * have to get rid of the duplication when building
900 if (tbp->b_flags & B_VMIO) {
903 VM_OBJECT_LOCK(tbp->b_bufobj->bo_object);
904 if (i != 0) { /* if not first buffer */
905 for (j = 0; j < tbp->b_npages; j += 1) {
907 if (m->flags & PG_BUSY) {
915 for (j = 0; j < tbp->b_npages; j += 1) {
918 vm_object_pip_add(m->object, 1);
919 if ((bp->b_npages == 0) ||
920 (bp->b_pages[bp->b_npages - 1] != m)) {
921 bp->b_pages[bp->b_npages] = m;
925 VM_OBJECT_UNLOCK(tbp->b_bufobj->bo_object);
927 bp->b_bcount += size;
928 bp->b_bufsize += size;
930 tbp->b_flags &= ~B_DONE;
931 tbp->b_ioflags &= ~BIO_ERROR;
932 tbp->b_flags |= B_ASYNC;
933 tbp->b_iocmd = BIO_WRITE;
934 reassignbuf(tbp); /* put on clean list */
935 bufobj_wref(tbp->b_bufobj);
937 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
938 tbp, b_cluster.cluster_entry);
941 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
942 (vm_page_t *) bp->b_pages, bp->b_npages);
943 if (bp->b_bufsize > bp->b_kvasize)
945 "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
946 bp->b_bufsize, bp->b_kvasize);
947 bp->b_kvasize = bp->b_bufsize;
948 totalwritten += bp->b_bufsize;
950 bp->b_dirtyend = bp->b_bufsize;
959 * Collect together all the buffers in a cluster.
960 * Plus add one additional buffer.
962 static struct cluster_save *
963 cluster_collectbufs(vp, last_bp)
967 struct cluster_save *buflist;
972 len = vp->v_lastw - vp->v_cstart + 1;
973 buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
974 M_SEGMENT, M_WAITOK);
975 buflist->bs_nchildren = 0;
976 buflist->bs_children = (struct buf **) (buflist + 1);
977 for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
978 (void) bread(vp, lbn, last_bp->b_bcount, NOCRED, &bp);
979 buflist->bs_children[i] = bp;
980 if (bp->b_blkno == bp->b_lblkno)
981 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno,
984 buflist->bs_children[i] = bp = last_bp;
985 if (bp->b_blkno == bp->b_lblkno)
986 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
987 buflist->bs_nchildren = i + 1;