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 *
65 cluster_collectbufs(struct vnode *vp, struct buf *last_bp);
67 cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn,
68 daddr_t blkno, long size, int run, struct buf *fbp);
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;
102 * Try to limit the amount of read-ahead by a few
103 * ad-hoc parameters. This needs work!!!
105 racluster = vp->v_mount->mnt_iosize_max / size;
107 maxra = min(read_max, maxra);
108 maxra = min(nbuf/8, maxra);
109 if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize)
110 maxra = (filesize / size) - lblkno;
113 * get the requested block
115 *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0, 0);
119 * if it is in the cache, then check to see if the reads have been
120 * sequential. If they have, then try some read-ahead, otherwise
121 * back-off on prospective read-aheads.
123 if (bp->b_flags & B_CACHE) {
126 } else if ((bp->b_flags & B_RAM) == 0) {
129 bp->b_flags &= ~B_RAM;
131 for (i = 1; i < maxra; i++) {
133 * Stop if the buffer does not exist or it
134 * is invalid (about to go away?)
136 rbp = gbincore(&vp->v_bufobj, lblkno+i);
137 if (rbp == NULL || (rbp->b_flags & B_INVAL))
141 * Set another read-ahead mark so we know
142 * to check again. (If we can lock the
143 * buffer without waiting)
145 if ((((i % racluster) == (racluster - 1)) ||
147 && (0 == BUF_LOCK(rbp,
148 LK_EXCLUSIVE | LK_NOWAIT, NULL))) {
149 rbp->b_flags |= B_RAM;
161 * If it isn't in the cache, then get a chunk from
162 * disk if sequential, otherwise just get the block.
165 off_t firstread = bp->b_offset;
168 KASSERT(bp->b_offset != NOOFFSET,
169 ("cluster_read: no buffer offset"));
174 * Compute the total number of blocks that we should read
177 if (firstread + totread > filesize)
178 totread = filesize - firstread;
179 nblks = howmany(totread, size);
180 if (nblks > racluster)
184 * Now compute the number of contiguous blocks.
187 error = VOP_BMAP(vp, lblkno, NULL,
188 &blkno, &ncontig, NULL);
190 * If this failed to map just do the original block.
192 if (error || blkno == -1)
197 * If we have contiguous data available do a cluster
198 * otherwise just read the requested block.
201 /* Account for our first block. */
202 ncontig = min(ncontig + 1, nblks);
205 bp = cluster_rbuild(vp, filesize, lblkno,
206 blkno, size, nblks, bp);
207 lblkno += (bp->b_bufsize / size);
209 bp->b_flags |= B_RAM;
210 bp->b_iocmd = BIO_READ;
216 * handle the synchronous read so that it is available ASAP.
219 if ((bp->b_flags & B_CLUSTER) == 0) {
220 vfs_busy_pages(bp, 0);
222 bp->b_flags &= ~B_INVAL;
223 bp->b_ioflags &= ~BIO_ERROR;
224 if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL)
226 bp->b_iooffset = dbtob(bp->b_blkno);
228 curthread->td_ru.ru_inblock++;
232 * If we have been doing sequential I/O, then do some read-ahead.
234 while (lblkno < (origblkno + maxra)) {
235 error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL);
243 * We could throttle ncontig here by maxra but we might as
244 * well read the data if it is contiguous. We're throttled
245 * by racluster anyway.
248 ncontig = min(ncontig + 1, racluster);
249 rbp = cluster_rbuild(vp, filesize, lblkno, blkno,
250 size, ncontig, NULL);
251 lblkno += (rbp->b_bufsize / size);
252 if (rbp->b_flags & B_DELWRI) {
257 rbp = getblk(vp, lblkno, size, 0, 0, 0);
259 if (rbp->b_flags & B_DELWRI) {
263 rbp->b_flags |= B_ASYNC | B_RAM;
264 rbp->b_iocmd = BIO_READ;
265 rbp->b_blkno = blkno;
267 if (rbp->b_flags & B_CACHE) {
268 rbp->b_flags &= ~B_ASYNC;
272 if ((rbp->b_flags & B_CLUSTER) == 0) {
273 vfs_busy_pages(rbp, 0);
275 rbp->b_flags &= ~B_INVAL;
276 rbp->b_ioflags &= ~BIO_ERROR;
277 if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL)
279 rbp->b_iooffset = dbtob(rbp->b_blkno);
281 curthread->td_ru.ru_inblock++;
285 return (bufwait(reqbp));
291 * If blocks are contiguous on disk, use this to provide clustered
292 * read ahead. We will read as many blocks as possible sequentially
293 * and then parcel them up into logical blocks in the buffer hash table.
296 cluster_rbuild(vp, filesize, lbn, blkno, size, run, fbp)
306 struct buf *bp, *tbp;
312 KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
313 ("cluster_rbuild: size %ld != f_iosize %jd\n",
314 size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
319 while ((u_quad_t) size * (lbn + run) > filesize) {
325 tbp->b_iocmd = BIO_READ;
327 tbp = getblk(vp, lbn, size, 0, 0, 0);
328 if (tbp->b_flags & B_CACHE)
330 tbp->b_flags |= B_ASYNC | B_RAM;
331 tbp->b_iocmd = BIO_READ;
333 tbp->b_blkno = blkno;
334 if( (tbp->b_flags & B_MALLOC) ||
335 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
338 bp = trypbuf(&cluster_pbuf_freecnt);
343 * We are synthesizing a buffer out of vm_page_t's, but
344 * if the block size is not page aligned then the starting
345 * address may not be either. Inherit the b_data offset
346 * from the original buffer.
348 bp->b_data = (char *)((vm_offset_t)bp->b_data |
349 ((vm_offset_t)tbp->b_data & PAGE_MASK));
350 bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO;
351 bp->b_iocmd = BIO_READ;
352 bp->b_iodone = cluster_callback;
355 bp->b_offset = tbp->b_offset;
356 KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
359 TAILQ_INIT(&bp->b_cluster.cluster_head);
367 for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
369 if ((bp->b_npages * PAGE_SIZE) +
370 round_page(size) > vp->v_mount->mnt_iosize_max) {
374 tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT);
376 /* Don't wait around for locked bufs. */
381 * Stop scanning if the buffer is fully valid
382 * (marked B_CACHE), or locked (may be doing a
383 * background write), or if the buffer is not
384 * VMIO backed. The clustering code can only deal
385 * with VMIO-backed buffers.
388 if ((tbp->b_vflags & BV_BKGRDINPROG) ||
389 (tbp->b_flags & B_CACHE) ||
390 (tbp->b_flags & B_VMIO) == 0) {
398 * The buffer must be completely invalid in order to
399 * take part in the cluster. If it is partially valid
404 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
405 for (j = 0; tsize > 0; j++) {
406 toff = off & PAGE_MASK;
408 if (toff + tinc > PAGE_SIZE)
409 tinc = PAGE_SIZE - toff;
410 VM_OBJECT_ASSERT_WLOCKED(tbp->b_pages[j]->object);
411 if ((tbp->b_pages[j]->valid &
412 vm_page_bits(toff, tinc)) != 0)
417 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
424 * Set a read-ahead mark as appropriate
426 if ((fbp && (i == 1)) || (i == (run - 1)))
427 tbp->b_flags |= B_RAM;
430 * Set the buffer up for an async read (XXX should
431 * we do this only if we do not wind up brelse()ing?).
432 * Set the block number if it isn't set, otherwise
433 * if it is make sure it matches the block number we
436 tbp->b_flags |= B_ASYNC;
437 tbp->b_iocmd = BIO_READ;
438 if (tbp->b_blkno == tbp->b_lblkno) {
440 } else if (tbp->b_blkno != bn) {
446 * XXX fbp from caller may not be B_ASYNC, but we are going
447 * to biodone() it in cluster_callback() anyway
450 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
451 tbp, b_cluster.cluster_entry);
452 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
453 for (j = 0; j < tbp->b_npages; j += 1) {
457 vm_object_pip_add(m->object, 1);
458 if ((bp->b_npages == 0) ||
459 (bp->b_pages[bp->b_npages-1] != m)) {
460 bp->b_pages[bp->b_npages] = m;
463 if (m->valid == VM_PAGE_BITS_ALL)
464 tbp->b_pages[j] = bogus_page;
466 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
468 * Don't inherit tbp->b_bufsize as it may be larger due to
469 * a non-page-aligned size. Instead just aggregate using
472 if (tbp->b_bcount != size)
473 printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
474 if (tbp->b_bufsize != size)
475 printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
476 bp->b_bcount += size;
477 bp->b_bufsize += size;
481 * Fully valid pages in the cluster are already good and do not need
482 * to be re-read from disk. Replace the page with bogus_page
484 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
485 for (j = 0; j < bp->b_npages; j++) {
486 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[j]->object);
487 if (bp->b_pages[j]->valid == VM_PAGE_BITS_ALL)
488 bp->b_pages[j] = bogus_page;
490 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
491 if (bp->b_bufsize > bp->b_kvasize)
492 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
493 bp->b_bufsize, bp->b_kvasize);
494 bp->b_kvasize = bp->b_bufsize;
496 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
497 (vm_page_t *)bp->b_pages, bp->b_npages);
502 * Cleanup after a clustered read or write.
503 * This is complicated by the fact that any of the buffers might have
504 * extra memory (if there were no empty buffer headers at allocbuf time)
505 * that we will need to shift around.
511 struct buf *nbp, *tbp;
515 * Must propogate errors to all the components.
517 if (bp->b_ioflags & BIO_ERROR)
520 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, 0);
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,
597 int maxclen, cursize;
601 if (vp->v_type == VREG) {
602 async = DOINGASYNC(vp);
603 lblocksize = vp->v_mount->mnt_stat.f_iosize;
606 lblocksize = bp->b_bufsize;
609 KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
611 /* Initialize vnode to beginning of file. */
613 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
615 if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
616 (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
617 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
618 if (vp->v_clen != 0) {
620 * Next block is not sequential.
622 * If we are not writing at end of file, the process
623 * seeked to another point in the file since its last
624 * write, or we have reached our maximum cluster size,
625 * then push the previous cluster. Otherwise try
626 * reallocating to make it sequential.
628 * Change to algorithm: only push previous cluster if
629 * it was sequential from the point of view of the
630 * seqcount heuristic, otherwise leave the buffer
631 * intact so we can potentially optimize the I/O
632 * later on in the buf_daemon or update daemon
635 cursize = vp->v_lastw - vp->v_cstart + 1;
636 if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
637 lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
638 if (!async && seqcount > 0) {
639 cluster_wbuild_wb(vp, lblocksize,
640 vp->v_cstart, cursize);
643 struct buf **bpp, **endbp;
644 struct cluster_save *buflist;
646 buflist = cluster_collectbufs(vp, bp);
647 endbp = &buflist->bs_children
648 [buflist->bs_nchildren - 1];
649 if (VOP_REALLOCBLKS(vp, buflist)) {
651 * Failed, push the previous cluster
652 * if *really* writing sequentially
653 * in the logical file (seqcount > 1),
654 * otherwise delay it in the hopes that
655 * the low level disk driver can
656 * optimize the write ordering.
658 for (bpp = buflist->bs_children;
661 free(buflist, M_SEGMENT);
663 cluster_wbuild_wb(vp,
664 lblocksize, vp->v_cstart,
669 * Succeeded, keep building cluster.
671 for (bpp = buflist->bs_children;
674 free(buflist, M_SEGMENT);
676 vp->v_lasta = bp->b_blkno;
682 * Consider beginning a cluster. If at end of file, make
683 * cluster as large as possible, otherwise find size of
686 if ((vp->v_type == VREG) &&
687 ((u_quad_t) bp->b_offset + lblocksize) != filesize &&
688 (bp->b_blkno == bp->b_lblkno) &&
689 (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
690 bp->b_blkno == -1)) {
693 vp->v_lasta = bp->b_blkno;
694 vp->v_cstart = lbn + 1;
698 vp->v_clen = maxclen;
699 if (!async && maxclen == 0) { /* I/O not contiguous */
700 vp->v_cstart = lbn + 1;
702 } else { /* Wait for rest of cluster */
706 } else if (lbn == vp->v_cstart + vp->v_clen) {
708 * At end of cluster, write it out if seqcount tells us we
709 * are operating sequentially, otherwise let the buf or
710 * update daemon handle it.
714 cluster_wbuild_wb(vp, lblocksize, vp->v_cstart, vp->v_clen + 1);
716 vp->v_cstart = lbn + 1;
717 } else if (vm_page_count_severe()) {
719 * We are low on memory, get it going NOW
724 * In the middle of a cluster, so just delay the I/O for now.
729 vp->v_lasta = bp->b_blkno;
734 * This is an awful lot like cluster_rbuild...wish they could be combined.
735 * The last lbn argument is the current block on which I/O is being
736 * performed. Check to see that it doesn't fall in the middle of
737 * the current block (if last_bp == NULL).
740 cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len,
743 struct buf *bp, *tbp;
746 int totalwritten = 0;
747 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, BO_MTX(bo))) {
770 if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
776 if (tbp->b_pin_count > 0) {
783 tbp->b_flags &= ~B_DONE;
786 * Extra memory in the buffer, punt on this buffer.
787 * XXX we could handle this in most cases, but we would
788 * have to push the extra memory down to after our max
789 * possible cluster size and then potentially pull it back
790 * up if the cluster was terminated prematurely--too much
793 if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
794 (B_CLUSTEROK | B_VMIO)) ||
795 (tbp->b_bcount != tbp->b_bufsize) ||
796 (tbp->b_bcount != size) ||
798 ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
799 totalwritten += tbp->b_bufsize;
807 * We got a pbuf to make the cluster in.
810 TAILQ_INIT(&bp->b_cluster.cluster_head);
814 if (tbp->b_wcred != NOCRED)
815 bp->b_wcred = crhold(tbp->b_wcred);
817 bp->b_blkno = tbp->b_blkno;
818 bp->b_lblkno = tbp->b_lblkno;
819 bp->b_offset = tbp->b_offset;
822 * We are synthesizing a buffer out of vm_page_t's, but
823 * if the block size is not page aligned then the starting
824 * address may not be either. Inherit the b_data offset
825 * from the original buffer.
827 bp->b_data = (char *)((vm_offset_t)bp->b_data |
828 ((vm_offset_t)tbp->b_data & PAGE_MASK));
829 bp->b_flags |= B_CLUSTER |
830 (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
831 bp->b_iodone = cluster_callback;
834 * From this location in the file, scan forward to see
835 * if there are buffers with adjacent data that need to
836 * be written as well.
838 for (i = 0; i < len; ++i, ++start_lbn) {
839 if (i != 0) { /* If not the first buffer */
841 * If the adjacent data is not even in core it
842 * can't need to be written.
845 if ((tbp = gbincore(bo, start_lbn)) == NULL ||
846 (tbp->b_vflags & BV_BKGRDINPROG)) {
852 * If it IS in core, but has different
853 * characteristics, or is locked (which
854 * means it could be undergoing a background
855 * I/O or be in a weird state), then don't
859 LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
863 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
864 B_INVAL | B_DELWRI | B_NEEDCOMMIT))
865 != (B_DELWRI | B_CLUSTEROK |
866 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
867 tbp->b_wcred != bp->b_wcred) {
873 * Check that the combined cluster
874 * would make sense with regard to pages
875 * and would not be too large
877 if ((tbp->b_bcount != size) ||
878 ((bp->b_blkno + (dbsize * i)) !=
880 ((tbp->b_npages + bp->b_npages) >
881 (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
887 * Do not pull in pinned buffers.
889 if (tbp->b_pin_count > 0) {
895 * Ok, it's passed all the tests,
896 * so remove it from the free list
897 * and mark it busy. We will use it.
900 tbp->b_flags &= ~B_DONE;
901 } /* end of code for non-first buffers only */
903 * If the IO is via the VM then we do some
904 * special VM hackery (yuck). Since the buffer's
905 * block size may not be page-aligned it is possible
906 * for a page to be shared between two buffers. We
907 * have to get rid of the duplication when building
910 if (tbp->b_flags & B_VMIO) {
913 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
914 if (i != 0) { /* if not first buffer */
915 for (j = 0; j < tbp->b_npages; j += 1) {
917 if (m->oflags & VPO_BUSY) {
925 for (j = 0; j < tbp->b_npages; j += 1) {
928 vm_object_pip_add(m->object, 1);
929 if ((bp->b_npages == 0) ||
930 (bp->b_pages[bp->b_npages - 1] != m)) {
931 bp->b_pages[bp->b_npages] = m;
935 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
937 bp->b_bcount += size;
938 bp->b_bufsize += size;
940 * If any of the clustered buffers have their
941 * B_BARRIER flag set, transfer that request to
944 bp->b_flags |= (tbp->b_flags & B_BARRIER);
945 tbp->b_flags &= ~(B_DONE | B_BARRIER);
946 tbp->b_flags |= B_ASYNC;
947 tbp->b_ioflags &= ~BIO_ERROR;
948 tbp->b_iocmd = BIO_WRITE;
950 reassignbuf(tbp); /* put on clean list */
951 bufobj_wref(tbp->b_bufobj);
953 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
954 tbp, b_cluster.cluster_entry);
957 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
958 (vm_page_t *) bp->b_pages, bp->b_npages);
959 if (bp->b_bufsize > bp->b_kvasize)
961 "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
962 bp->b_bufsize, bp->b_kvasize);
963 bp->b_kvasize = bp->b_bufsize;
964 totalwritten += bp->b_bufsize;
966 bp->b_dirtyend = bp->b_bufsize;
975 * Collect together all the buffers in a cluster.
976 * Plus add one additional buffer.
978 static struct cluster_save *
979 cluster_collectbufs(vp, last_bp)
983 struct cluster_save *buflist;
988 len = vp->v_lastw - vp->v_cstart + 1;
989 buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
990 M_SEGMENT, M_WAITOK);
991 buflist->bs_nchildren = 0;
992 buflist->bs_children = (struct buf **) (buflist + 1);
993 for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
994 (void) bread(vp, lbn, last_bp->b_bcount, NOCRED, &bp);
995 buflist->bs_children[i] = bp;
996 if (bp->b_blkno == bp->b_lblkno)
997 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno,
1000 buflist->bs_children[i] = bp = last_bp;
1001 if (bp->b_blkno == bp->b_lblkno)
1002 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1003 buflist->bs_nchildren = i + 1;