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/rwlock.h>
50 #include <sys/vmmeter.h>
52 #include <vm/vm_object.h>
53 #include <vm/vm_page.h>
54 #include <sys/sysctl.h>
56 #if defined(CLUSTERDEBUG)
57 static int rcluster= 0;
58 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0,
59 "Debug VFS clustering code");
62 static MALLOC_DEFINE(M_SEGMENT, "cl_savebuf", "cluster_save buffer");
64 static struct cluster_save *cluster_collectbufs(struct vnode *vp,
65 struct buf *last_bp, int gbflags);
66 static struct buf *cluster_rbuild(struct vnode *vp, u_quad_t filesize,
67 daddr_t lbn, daddr_t blkno, long size, int run, int gbflags,
69 static void cluster_callback(struct buf *);
71 static int write_behind = 1;
72 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
73 "Cluster write-behind; 0: disable, 1: enable, 2: backed off");
75 static int read_max = 64;
76 SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0,
77 "Cluster read-ahead max block count");
79 /* Page expended to mark partially backed buffers */
80 extern vm_page_t bogus_page;
83 * Read data to a buf, including read-ahead if we find this to be beneficial.
84 * cluster_read replaces bread.
87 cluster_read(struct vnode *vp, u_quad_t filesize, daddr_t lblkno, long size,
88 struct ucred *cred, long totread, int seqcount, int gbflags,
91 struct buf *bp, *rbp, *reqbp;
93 daddr_t blkno, origblkno;
100 if (!unmapped_buf_allowed)
101 gbflags &= ~GB_UNMAPPED;
104 * Try to limit the amount of read-ahead by a few
105 * ad-hoc parameters. This needs work!!!
107 racluster = vp->v_mount->mnt_iosize_max / size;
109 maxra = min(read_max, maxra);
110 maxra = min(nbuf/8, maxra);
111 if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize)
112 maxra = (filesize / size) - lblkno;
115 * get the requested block
117 *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0, gbflags);
121 * if it is in the cache, then check to see if the reads have been
122 * sequential. If they have, then try some read-ahead, otherwise
123 * back-off on prospective read-aheads.
125 if (bp->b_flags & B_CACHE) {
128 } else if ((bp->b_flags & B_RAM) == 0) {
131 bp->b_flags &= ~B_RAM;
133 for (i = 1; i < maxra; i++) {
135 * Stop if the buffer does not exist or it
136 * is invalid (about to go away?)
138 rbp = gbincore(&vp->v_bufobj, lblkno+i);
139 if (rbp == NULL || (rbp->b_flags & B_INVAL))
143 * Set another read-ahead mark so we know
144 * to check again. (If we can lock the
145 * buffer without waiting)
147 if ((((i % racluster) == (racluster - 1)) ||
149 && (0 == BUF_LOCK(rbp,
150 LK_EXCLUSIVE | LK_NOWAIT, NULL))) {
151 rbp->b_flags |= B_RAM;
163 * If it isn't in the cache, then get a chunk from
164 * disk if sequential, otherwise just get the block.
167 off_t firstread = bp->b_offset;
170 KASSERT(bp->b_offset != NOOFFSET,
171 ("cluster_read: no buffer offset"));
176 * Compute the total number of blocks that we should read
179 if (firstread + totread > filesize)
180 totread = filesize - firstread;
181 nblks = howmany(totread, size);
182 if (nblks > racluster)
186 * Now compute the number of contiguous blocks.
189 error = VOP_BMAP(vp, lblkno, NULL,
190 &blkno, &ncontig, NULL);
192 * If this failed to map just do the original block.
194 if (error || blkno == -1)
199 * If we have contiguous data available do a cluster
200 * otherwise just read the requested block.
203 /* Account for our first block. */
204 ncontig = min(ncontig + 1, nblks);
207 bp = cluster_rbuild(vp, filesize, lblkno,
208 blkno, size, nblks, gbflags, bp);
209 lblkno += (bp->b_bufsize / size);
211 bp->b_flags |= B_RAM;
212 bp->b_iocmd = BIO_READ;
218 * handle the synchronous read so that it is available ASAP.
221 if ((bp->b_flags & B_CLUSTER) == 0) {
222 vfs_busy_pages(bp, 0);
224 bp->b_flags &= ~B_INVAL;
225 bp->b_ioflags &= ~BIO_ERROR;
226 if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL)
228 bp->b_iooffset = dbtob(bp->b_blkno);
230 curthread->td_ru.ru_inblock++;
234 * If we have been doing sequential I/O, then do some read-ahead.
236 while (lblkno < (origblkno + maxra)) {
237 error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL);
245 * We could throttle ncontig here by maxra but we might as
246 * well read the data if it is contiguous. We're throttled
247 * by racluster anyway.
250 ncontig = min(ncontig + 1, racluster);
251 rbp = cluster_rbuild(vp, filesize, lblkno, blkno,
252 size, ncontig, gbflags, NULL);
253 lblkno += (rbp->b_bufsize / size);
254 if (rbp->b_flags & B_DELWRI) {
259 rbp = getblk(vp, lblkno, size, 0, 0, gbflags);
261 if (rbp->b_flags & B_DELWRI) {
265 rbp->b_flags |= B_ASYNC | B_RAM;
266 rbp->b_iocmd = BIO_READ;
267 rbp->b_blkno = blkno;
269 if (rbp->b_flags & B_CACHE) {
270 rbp->b_flags &= ~B_ASYNC;
274 if ((rbp->b_flags & B_CLUSTER) == 0) {
275 vfs_busy_pages(rbp, 0);
277 rbp->b_flags &= ~B_INVAL;
278 rbp->b_ioflags &= ~BIO_ERROR;
279 if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL)
281 rbp->b_iooffset = dbtob(rbp->b_blkno);
283 curthread->td_ru.ru_inblock++;
287 return (bufwait(reqbp));
293 * If blocks are contiguous on disk, use this to provide clustered
294 * read ahead. We will read as many blocks as possible sequentially
295 * and then parcel them up into logical blocks in the buffer hash table.
298 cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn,
299 daddr_t blkno, long size, int run, int gbflags, struct buf *fbp)
302 struct buf *bp, *tbp;
308 KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
309 ("cluster_rbuild: size %ld != f_iosize %jd\n",
310 size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
315 while ((u_quad_t) size * (lbn + run) > filesize) {
321 tbp->b_iocmd = BIO_READ;
323 tbp = getblk(vp, lbn, size, 0, 0, gbflags);
324 if (tbp->b_flags & B_CACHE)
326 tbp->b_flags |= B_ASYNC | B_RAM;
327 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_flags = B_ASYNC | B_CLUSTER | B_VMIO;
345 if ((gbflags & GB_UNMAPPED) != 0) {
346 bp->b_flags |= B_UNMAPPED;
347 bp->b_data = unmapped_buf;
349 bp->b_data = (char *)((vm_offset_t)bp->b_data |
350 ((vm_offset_t)tbp->b_data & PAGE_MASK));
352 bp->b_iocmd = BIO_READ;
353 bp->b_iodone = cluster_callback;
356 bp->b_offset = tbp->b_offset;
357 KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
360 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 |
376 (gbflags & GB_UNMAPPED));
378 /* Don't wait around for locked bufs. */
383 * Stop scanning if the buffer is fully valid
384 * (marked B_CACHE), or locked (may be doing a
385 * background write), or if the buffer is not
386 * VMIO backed. The clustering code can only deal
387 * with VMIO-backed buffers.
390 if ((tbp->b_vflags & BV_BKGRDINPROG) ||
391 (tbp->b_flags & B_CACHE) ||
392 (tbp->b_flags & B_VMIO) == 0) {
400 * The buffer must be completely invalid in order to
401 * take part in the cluster. If it is partially valid
406 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
407 for (j = 0; tsize > 0; j++) {
408 toff = off & PAGE_MASK;
410 if (toff + tinc > PAGE_SIZE)
411 tinc = PAGE_SIZE - toff;
412 VM_OBJECT_ASSERT_WLOCKED(tbp->b_pages[j]->object);
413 if ((tbp->b_pages[j]->valid &
414 vm_page_bits(toff, tinc)) != 0)
419 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
426 * Set a read-ahead mark as appropriate
428 if ((fbp && (i == 1)) || (i == (run - 1)))
429 tbp->b_flags |= B_RAM;
432 * Set the buffer up for an async read (XXX should
433 * we do this only if we do not wind up brelse()ing?).
434 * Set the block number if it isn't set, otherwise
435 * if it is make sure it matches the block number we
438 tbp->b_flags |= B_ASYNC;
439 tbp->b_iocmd = BIO_READ;
440 if (tbp->b_blkno == tbp->b_lblkno) {
442 } else if (tbp->b_blkno != bn) {
448 * XXX fbp from caller may not be B_ASYNC, but we are going
449 * to biodone() it in cluster_callback() anyway
452 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
453 tbp, b_cluster.cluster_entry);
454 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
455 for (j = 0; j < tbp->b_npages; j += 1) {
459 vm_object_pip_add(m->object, 1);
460 if ((bp->b_npages == 0) ||
461 (bp->b_pages[bp->b_npages-1] != m)) {
462 bp->b_pages[bp->b_npages] = m;
465 if (m->valid == VM_PAGE_BITS_ALL)
466 tbp->b_pages[j] = bogus_page;
468 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
470 * Don't inherit tbp->b_bufsize as it may be larger due to
471 * a non-page-aligned size. Instead just aggregate using
474 if (tbp->b_bcount != size)
475 printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
476 if (tbp->b_bufsize != size)
477 printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
478 bp->b_bcount += size;
479 bp->b_bufsize += size;
483 * Fully valid pages in the cluster are already good and do not need
484 * to be re-read from disk. Replace the page with bogus_page
486 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
487 for (j = 0; j < bp->b_npages; j++) {
488 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[j]->object);
489 if (bp->b_pages[j]->valid == VM_PAGE_BITS_ALL)
490 bp->b_pages[j] = bogus_page;
492 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
493 if (bp->b_bufsize > bp->b_kvasize)
494 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
495 bp->b_bufsize, bp->b_kvasize);
496 bp->b_kvasize = bp->b_bufsize;
498 if ((bp->b_flags & B_UNMAPPED) == 0) {
499 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
500 (vm_page_t *)bp->b_pages, bp->b_npages);
506 * Cleanup after a clustered read or write.
507 * This is complicated by the fact that any of the buffers might have
508 * extra memory (if there were no empty buffer headers at allocbuf time)
509 * that we will need to shift around.
515 struct buf *nbp, *tbp;
519 * Must propogate errors to all the components.
521 if (bp->b_ioflags & BIO_ERROR)
524 if ((bp->b_flags & B_UNMAPPED) == 0) {
525 pmap_qremove(trunc_page((vm_offset_t) bp->b_data),
529 * Move memory from the large cluster buffer into the component
530 * buffers and mark IO as done on these.
532 for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
534 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
536 tbp->b_ioflags |= BIO_ERROR;
537 tbp->b_error = error;
539 tbp->b_dirtyoff = tbp->b_dirtyend = 0;
540 tbp->b_flags &= ~B_INVAL;
541 tbp->b_ioflags &= ~BIO_ERROR;
543 * XXX the bdwrite()/bqrelse() issued during
544 * cluster building clears B_RELBUF (see bqrelse()
545 * comment). If direct I/O was specified, we have
546 * to restore it here to allow the buffer and VM
549 if (tbp->b_flags & B_DIRECT)
550 tbp->b_flags |= B_RELBUF;
555 relpbuf(bp, &cluster_pbuf_freecnt);
561 * Implement modified write build for cluster.
563 * write_behind = 0 write behind disabled
564 * write_behind = 1 write behind normal (default)
565 * write_behind = 2 write behind backed-off
569 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len,
574 switch (write_behind) {
581 r = cluster_wbuild(vp, size, start_lbn, len, gbflags);
591 * Do clustered write for FFS.
594 * 1. Write is not sequential (write asynchronously)
595 * Write is sequential:
596 * 2. beginning of cluster - begin cluster
597 * 3. middle of a cluster - add to cluster
598 * 4. end of a cluster - asynchronously write cluster
601 cluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount,
605 int maxclen, cursize;
609 if (!unmapped_buf_allowed)
610 gbflags &= ~GB_UNMAPPED;
612 if (vp->v_type == VREG) {
613 async = DOINGASYNC(vp);
614 lblocksize = vp->v_mount->mnt_stat.f_iosize;
617 lblocksize = bp->b_bufsize;
620 KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
622 /* Initialize vnode to beginning of file. */
624 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
626 if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
627 (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
628 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
629 if (vp->v_clen != 0) {
631 * Next block is not sequential.
633 * If we are not writing at end of file, the process
634 * seeked to another point in the file since its last
635 * write, or we have reached our maximum cluster size,
636 * then push the previous cluster. Otherwise try
637 * reallocating to make it sequential.
639 * Change to algorithm: only push previous cluster if
640 * it was sequential from the point of view of the
641 * seqcount heuristic, otherwise leave the buffer
642 * intact so we can potentially optimize the I/O
643 * later on in the buf_daemon or update daemon
646 cursize = vp->v_lastw - vp->v_cstart + 1;
647 if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
648 lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
649 if (!async && seqcount > 0) {
650 cluster_wbuild_wb(vp, lblocksize,
651 vp->v_cstart, cursize, gbflags);
654 struct buf **bpp, **endbp;
655 struct cluster_save *buflist;
657 buflist = cluster_collectbufs(vp, bp, gbflags);
658 endbp = &buflist->bs_children
659 [buflist->bs_nchildren - 1];
660 if (VOP_REALLOCBLKS(vp, buflist)) {
662 * Failed, push the previous cluster
663 * if *really* writing sequentially
664 * in the logical file (seqcount > 1),
665 * otherwise delay it in the hopes that
666 * the low level disk driver can
667 * optimize the write ordering.
669 for (bpp = buflist->bs_children;
672 free(buflist, M_SEGMENT);
674 cluster_wbuild_wb(vp,
675 lblocksize, vp->v_cstart,
680 * Succeeded, keep building cluster.
682 for (bpp = buflist->bs_children;
685 free(buflist, M_SEGMENT);
687 vp->v_lasta = bp->b_blkno;
693 * Consider beginning a cluster. If at end of file, make
694 * cluster as large as possible, otherwise find size of
697 if ((vp->v_type == VREG) &&
698 ((u_quad_t) bp->b_offset + lblocksize) != filesize &&
699 (bp->b_blkno == bp->b_lblkno) &&
700 (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
701 bp->b_blkno == -1)) {
704 vp->v_lasta = bp->b_blkno;
705 vp->v_cstart = lbn + 1;
709 vp->v_clen = maxclen;
710 if (!async && maxclen == 0) { /* I/O not contiguous */
711 vp->v_cstart = lbn + 1;
713 } else { /* Wait for rest of cluster */
717 } else if (lbn == vp->v_cstart + vp->v_clen) {
719 * At end of cluster, write it out if seqcount tells us we
720 * are operating sequentially, otherwise let the buf or
721 * update daemon handle it.
725 cluster_wbuild_wb(vp, lblocksize, vp->v_cstart,
726 vp->v_clen + 1, gbflags);
729 vp->v_cstart = lbn + 1;
730 } else if (vm_page_count_severe()) {
732 * We are low on memory, get it going NOW
737 * In the middle of a cluster, so just delay the I/O for now.
742 vp->v_lasta = bp->b_blkno;
747 * This is an awful lot like cluster_rbuild...wish they could be combined.
748 * The last lbn argument is the current block on which I/O is being
749 * performed. Check to see that it doesn't fall in the middle of
750 * the current block (if last_bp == NULL).
753 cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len,
756 struct buf *bp, *tbp;
759 int totalwritten = 0;
760 int dbsize = btodb(size);
762 if (!unmapped_buf_allowed)
763 gbflags &= ~GB_UNMAPPED;
768 * If the buffer is not delayed-write (i.e. dirty), or it
769 * is delayed-write but either locked or inval, it cannot
770 * partake in the clustered write.
773 if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
774 (tbp->b_vflags & BV_BKGRDINPROG)) {
781 LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_MTX(bo))) {
786 if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
792 if (tbp->b_pin_count > 0) {
799 tbp->b_flags &= ~B_DONE;
802 * Extra memory in the buffer, punt on this buffer.
803 * XXX we could handle this in most cases, but we would
804 * have to push the extra memory down to after our max
805 * possible cluster size and then potentially pull it back
806 * up if the cluster was terminated prematurely--too much
809 if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
810 (B_CLUSTEROK | B_VMIO)) ||
811 (tbp->b_bcount != tbp->b_bufsize) ||
812 (tbp->b_bcount != size) ||
814 ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
815 totalwritten += tbp->b_bufsize;
823 * We got a pbuf to make the cluster in.
826 TAILQ_INIT(&bp->b_cluster.cluster_head);
830 if (tbp->b_wcred != NOCRED)
831 bp->b_wcred = crhold(tbp->b_wcred);
833 bp->b_blkno = tbp->b_blkno;
834 bp->b_lblkno = tbp->b_lblkno;
835 bp->b_offset = tbp->b_offset;
838 * We are synthesizing a buffer out of vm_page_t's, but
839 * if the block size is not page aligned then the starting
840 * address may not be either. Inherit the b_data offset
841 * from the original buffer.
843 if ((gbflags & GB_UNMAPPED) == 0 ||
844 (tbp->b_flags & B_VMIO) == 0) {
845 bp->b_data = (char *)((vm_offset_t)bp->b_data |
846 ((vm_offset_t)tbp->b_data & PAGE_MASK));
848 bp->b_flags |= B_UNMAPPED;
849 bp->b_data = unmapped_buf;
851 bp->b_flags |= B_CLUSTER | (tbp->b_flags & (B_VMIO |
853 bp->b_iodone = cluster_callback;
856 * From this location in the file, scan forward to see
857 * if there are buffers with adjacent data that need to
858 * be written as well.
860 for (i = 0; i < len; ++i, ++start_lbn) {
861 if (i != 0) { /* If not the first buffer */
863 * If the adjacent data is not even in core it
864 * can't need to be written.
867 if ((tbp = gbincore(bo, start_lbn)) == NULL ||
868 (tbp->b_vflags & BV_BKGRDINPROG)) {
874 * If it IS in core, but has different
875 * characteristics, or is locked (which
876 * means it could be undergoing a background
877 * I/O or be in a weird state), then don't
881 LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
885 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
886 B_INVAL | B_DELWRI | B_NEEDCOMMIT))
887 != (B_DELWRI | B_CLUSTEROK |
888 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
889 tbp->b_wcred != bp->b_wcred) {
895 * Check that the combined cluster
896 * would make sense with regard to pages
897 * and would not be too large
899 if ((tbp->b_bcount != size) ||
900 ((bp->b_blkno + (dbsize * i)) !=
902 ((tbp->b_npages + bp->b_npages) >
903 (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
909 * Do not pull in pinned buffers.
911 if (tbp->b_pin_count > 0) {
917 * Ok, it's passed all the tests,
918 * so remove it from the free list
919 * and mark it busy. We will use it.
922 tbp->b_flags &= ~B_DONE;
923 } /* end of code for non-first buffers only */
925 * If the IO is via the VM then we do some
926 * special VM hackery (yuck). Since the buffer's
927 * block size may not be page-aligned it is possible
928 * for a page to be shared between two buffers. We
929 * have to get rid of the duplication when building
932 if (tbp->b_flags & B_VMIO) {
935 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
936 if (i != 0) { /* if not first buffer */
937 for (j = 0; j < tbp->b_npages; j += 1) {
939 if (m->oflags & VPO_BUSY) {
947 for (j = 0; j < tbp->b_npages; j += 1) {
950 vm_object_pip_add(m->object, 1);
951 if ((bp->b_npages == 0) ||
952 (bp->b_pages[bp->b_npages - 1] != m)) {
953 bp->b_pages[bp->b_npages] = m;
957 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
959 bp->b_bcount += size;
960 bp->b_bufsize += size;
962 * If any of the clustered buffers have their
963 * B_BARRIER flag set, transfer that request to
966 bp->b_flags |= (tbp->b_flags & B_BARRIER);
967 tbp->b_flags &= ~(B_DONE | B_BARRIER);
968 tbp->b_flags |= B_ASYNC;
969 tbp->b_ioflags &= ~BIO_ERROR;
970 tbp->b_iocmd = BIO_WRITE;
972 reassignbuf(tbp); /* put on clean list */
973 bufobj_wref(tbp->b_bufobj);
975 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
976 tbp, b_cluster.cluster_entry);
979 if ((bp->b_flags & B_UNMAPPED) == 0) {
980 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
981 (vm_page_t *)bp->b_pages, bp->b_npages);
983 if (bp->b_bufsize > bp->b_kvasize)
985 "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
986 bp->b_bufsize, bp->b_kvasize);
987 bp->b_kvasize = bp->b_bufsize;
988 totalwritten += bp->b_bufsize;
990 bp->b_dirtyend = bp->b_bufsize;
999 * Collect together all the buffers in a cluster.
1000 * Plus add one additional buffer.
1002 static struct cluster_save *
1003 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int gbflags)
1005 struct cluster_save *buflist;
1010 len = vp->v_lastw - vp->v_cstart + 1;
1011 buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1012 M_SEGMENT, M_WAITOK);
1013 buflist->bs_nchildren = 0;
1014 buflist->bs_children = (struct buf **) (buflist + 1);
1015 for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
1016 (void)bread_gb(vp, lbn, last_bp->b_bcount, NOCRED,
1018 buflist->bs_children[i] = bp;
1019 if (bp->b_blkno == bp->b_lblkno)
1020 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno,
1023 buflist->bs_children[i] = bp = last_bp;
1024 if (bp->b_blkno == bp->b_lblkno)
1025 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1026 buflist->bs_nchildren = i + 1;