2 * SPDX-License-Identifier: BSD-3-Clause
5 * The Regents of the University of California. All rights reserved.
6 * Modifications/enhancements:
7 * Copyright (c) 1995 John S. Dyson. All rights reserved.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 * 3. Neither the name of the University nor the names of its contributors
18 * may be used to endorse or promote products derived from this software
19 * without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * @(#)vfs_cluster.c 8.7 (Berkeley) 2/13/94
36 #include <sys/cdefs.h>
37 __FBSDID("$FreeBSD$");
39 #include "opt_debug_cluster.h"
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/kernel.h>
47 #include <sys/vnode.h>
48 #include <sys/malloc.h>
49 #include <sys/mount.h>
50 #include <sys/racct.h>
51 #include <sys/resourcevar.h>
52 #include <sys/rwlock.h>
53 #include <sys/vmmeter.h>
55 #include <vm/vm_object.h>
56 #include <vm/vm_page.h>
57 #include <sys/sysctl.h>
59 #if defined(CLUSTERDEBUG)
60 static int rcluster= 0;
61 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0,
62 "Debug VFS clustering code");
65 static MALLOC_DEFINE(M_SEGMENT, "cl_savebuf", "cluster_save buffer");
67 static struct cluster_save *cluster_collectbufs(struct vnode *vp,
68 struct buf *last_bp, int gbflags);
69 static struct buf *cluster_rbuild(struct vnode *vp, u_quad_t filesize,
70 daddr_t lbn, daddr_t blkno, long size, int run, int gbflags,
72 static void cluster_callback(struct buf *);
74 static int write_behind = 1;
75 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
76 "Cluster write-behind; 0: disable, 1: enable, 2: backed off");
78 static int read_max = 64;
79 SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0,
80 "Cluster read-ahead max block count");
82 static int read_min = 1;
83 SYSCTL_INT(_vfs, OID_AUTO, read_min, CTLFLAG_RW, &read_min, 0,
84 "Cluster read min block count");
87 * Read data to a buf, including read-ahead if we find this to be beneficial.
88 * cluster_read replaces bread.
91 cluster_read(struct vnode *vp, u_quad_t filesize, daddr_t lblkno, long size,
92 struct ucred *cred, long totread, int seqcount, int gbflags,
95 struct buf *bp, *rbp, *reqbp;
97 daddr_t blkno, origblkno;
104 if (!unmapped_buf_allowed)
105 gbflags &= ~GB_UNMAPPED;
108 * Try to limit the amount of read-ahead by a few
109 * ad-hoc parameters. This needs work!!!
111 racluster = vp->v_mount->mnt_iosize_max / size;
113 maxra = min(read_max, maxra);
114 maxra = min(nbuf/8, maxra);
115 if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize)
116 maxra = (filesize / size) - lblkno;
119 * get the requested block
121 *bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0, gbflags);
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.
131 if (bp->b_flags & B_CACHE) {
134 } else if ((bp->b_flags & B_RAM) == 0) {
137 bp->b_flags &= ~B_RAM;
139 for (i = 1; i < maxra; i++) {
141 * Stop if the buffer does not exist or it
142 * is invalid (about to go away?)
144 rbp = gbincore(&vp->v_bufobj, lblkno+i);
145 if (rbp == NULL || (rbp->b_flags & B_INVAL))
149 * Set another read-ahead mark so we know
150 * to check again. (If we can lock the
151 * buffer without waiting)
153 if ((((i % racluster) == (racluster - 1)) ||
155 && (0 == BUF_LOCK(rbp,
156 LK_EXCLUSIVE | LK_NOWAIT, NULL))) {
157 rbp->b_flags |= B_RAM;
169 * If it isn't in the cache, then get a chunk from
170 * disk if sequential, otherwise just get the block.
173 off_t firstread = bp->b_offset;
177 KASSERT(bp->b_offset != NOOFFSET,
178 ("cluster_read: no buffer offset"));
183 * Adjust totread if needed
185 minread = read_min * size;
186 if (minread > totread)
190 * Compute the total number of blocks that we should read
193 if (firstread + totread > filesize)
194 totread = filesize - firstread;
195 nblks = howmany(totread, size);
196 if (nblks > racluster)
200 * Now compute the number of contiguous blocks.
203 error = VOP_BMAP(vp, lblkno, NULL,
204 &blkno, &ncontig, NULL);
206 * If this failed to map just do the original block.
208 if (error || blkno == -1)
213 * If we have contiguous data available do a cluster
214 * otherwise just read the requested block.
217 /* Account for our first block. */
218 ncontig = min(ncontig + 1, nblks);
221 bp = cluster_rbuild(vp, filesize, lblkno,
222 blkno, size, nblks, gbflags, bp);
223 lblkno += (bp->b_bufsize / size);
225 bp->b_flags |= B_RAM;
226 bp->b_iocmd = BIO_READ;
232 * handle the synchronous read so that it is available ASAP.
235 if ((bp->b_flags & B_CLUSTER) == 0) {
236 vfs_busy_pages(bp, 0);
238 bp->b_flags &= ~B_INVAL;
239 bp->b_ioflags &= ~BIO_ERROR;
240 if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL)
242 bp->b_iooffset = dbtob(bp->b_blkno);
247 racct_add_buf(curproc, bp, 0);
248 PROC_UNLOCK(curproc);
251 curthread->td_ru.ru_inblock++;
255 * If we have been doing sequential I/O, then do some read-ahead.
257 while (lblkno < (origblkno + maxra)) {
258 error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL);
266 * We could throttle ncontig here by maxra but we might as
267 * well read the data if it is contiguous. We're throttled
268 * by racluster anyway.
271 ncontig = min(ncontig + 1, racluster);
272 rbp = cluster_rbuild(vp, filesize, lblkno, blkno,
273 size, ncontig, gbflags, NULL);
274 lblkno += (rbp->b_bufsize / size);
275 if (rbp->b_flags & B_DELWRI) {
280 rbp = getblk(vp, lblkno, size, 0, 0, gbflags);
282 if (rbp->b_flags & B_DELWRI) {
286 rbp->b_flags |= B_ASYNC | B_RAM;
287 rbp->b_iocmd = BIO_READ;
288 rbp->b_blkno = blkno;
290 if (rbp->b_flags & B_CACHE) {
291 rbp->b_flags &= ~B_ASYNC;
295 if ((rbp->b_flags & B_CLUSTER) == 0) {
296 vfs_busy_pages(rbp, 0);
298 rbp->b_flags &= ~B_INVAL;
299 rbp->b_ioflags &= ~BIO_ERROR;
300 if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL)
302 rbp->b_iooffset = dbtob(rbp->b_blkno);
307 racct_add_buf(curproc, rbp, 0);
308 PROC_UNLOCK(curproc);
311 curthread->td_ru.ru_inblock++;
316 * Like bread, always brelse() the buffer when
317 * returning an error.
319 error = bufwait(reqbp);
329 * If blocks are contiguous on disk, use this to provide clustered
330 * read ahead. We will read as many blocks as possible sequentially
331 * and then parcel them up into logical blocks in the buffer hash table.
334 cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn,
335 daddr_t blkno, long size, int run, int gbflags, struct buf *fbp)
337 struct buf *bp, *tbp;
341 int i, inc, j, k, toff;
343 KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
344 ("cluster_rbuild: size %ld != f_iosize %jd\n",
345 size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
350 while ((u_quad_t) size * (lbn + run) > filesize) {
356 tbp->b_iocmd = BIO_READ;
358 tbp = getblk(vp, lbn, size, 0, 0, gbflags);
359 if (tbp->b_flags & B_CACHE)
361 tbp->b_flags |= B_ASYNC | B_RAM;
362 tbp->b_iocmd = BIO_READ;
364 tbp->b_blkno = blkno;
365 if( (tbp->b_flags & B_MALLOC) ||
366 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
369 bp = trypbuf(&cluster_pbuf_freecnt);
374 * We are synthesizing a buffer out of vm_page_t's, but
375 * if the block size is not page aligned then the starting
376 * address may not be either. Inherit the b_data offset
377 * from the original buffer.
379 bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO;
380 if ((gbflags & GB_UNMAPPED) != 0) {
381 bp->b_data = unmapped_buf;
383 bp->b_data = (char *)((vm_offset_t)bp->b_data |
384 ((vm_offset_t)tbp->b_data & PAGE_MASK));
386 bp->b_iocmd = BIO_READ;
387 bp->b_iodone = cluster_callback;
390 bp->b_offset = tbp->b_offset;
391 KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
394 TAILQ_INIT(&bp->b_cluster.cluster_head);
401 for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
403 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
404 vfs_drain_busy_pages(tbp);
405 vm_object_pip_add(tbp->b_bufobj->bo_object,
407 for (k = 0; k < tbp->b_npages; k++)
408 vm_page_sbusy(tbp->b_pages[k]);
409 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
411 if ((bp->b_npages * PAGE_SIZE) +
412 round_page(size) > vp->v_mount->mnt_iosize_max) {
416 tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT |
417 (gbflags & GB_UNMAPPED));
419 /* Don't wait around for locked bufs. */
424 * Stop scanning if the buffer is fully valid
425 * (marked B_CACHE), or locked (may be doing a
426 * background write), or if the buffer is not
427 * VMIO backed. The clustering code can only deal
428 * with VMIO-backed buffers. The bo lock is not
429 * required for the BKGRDINPROG check since it
430 * can not be set without the buf lock.
432 if ((tbp->b_vflags & BV_BKGRDINPROG) ||
433 (tbp->b_flags & B_CACHE) ||
434 (tbp->b_flags & B_VMIO) == 0) {
440 * The buffer must be completely invalid in order to
441 * take part in the cluster. If it is partially valid
446 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
447 for (j = 0; tsize > 0; j++) {
448 toff = off & PAGE_MASK;
450 if (toff + tinc > PAGE_SIZE)
451 tinc = PAGE_SIZE - toff;
452 VM_OBJECT_ASSERT_WLOCKED(tbp->b_pages[j]->object);
453 if ((tbp->b_pages[j]->valid &
454 vm_page_bits(toff, tinc)) != 0)
456 if (vm_page_xbusied(tbp->b_pages[j]))
458 vm_object_pip_add(tbp->b_bufobj->bo_object, 1);
459 vm_page_sbusy(tbp->b_pages[j]);
465 vm_object_pip_add(tbp->b_bufobj->bo_object, -j);
466 for (k = 0; k < j; k++)
467 vm_page_sunbusy(tbp->b_pages[k]);
468 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
472 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
475 * Set a read-ahead mark as appropriate
477 if ((fbp && (i == 1)) || (i == (run - 1)))
478 tbp->b_flags |= B_RAM;
481 * Set the buffer up for an async read (XXX should
482 * we do this only if we do not wind up brelse()ing?).
483 * Set the block number if it isn't set, otherwise
484 * if it is make sure it matches the block number we
487 tbp->b_flags |= B_ASYNC;
488 tbp->b_iocmd = BIO_READ;
489 if (tbp->b_blkno == tbp->b_lblkno) {
491 } else if (tbp->b_blkno != bn) {
492 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
497 * XXX fbp from caller may not be B_ASYNC, but we are going
498 * to biodone() it in cluster_callback() anyway
501 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
502 tbp, b_cluster.cluster_entry);
503 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
504 for (j = 0; j < tbp->b_npages; j += 1) {
507 if ((bp->b_npages == 0) ||
508 (bp->b_pages[bp->b_npages-1] != m)) {
509 bp->b_pages[bp->b_npages] = m;
512 if (m->valid == VM_PAGE_BITS_ALL)
513 tbp->b_pages[j] = bogus_page;
515 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
517 * Don't inherit tbp->b_bufsize as it may be larger due to
518 * a non-page-aligned size. Instead just aggregate using
521 if (tbp->b_bcount != size)
522 printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
523 if (tbp->b_bufsize != size)
524 printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
525 bp->b_bcount += size;
526 bp->b_bufsize += size;
530 * Fully valid pages in the cluster are already good and do not need
531 * to be re-read from disk. Replace the page with bogus_page
533 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
534 for (j = 0; j < bp->b_npages; j++) {
535 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[j]->object);
536 if (bp->b_pages[j]->valid == VM_PAGE_BITS_ALL)
537 bp->b_pages[j] = bogus_page;
539 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
540 if (bp->b_bufsize > bp->b_kvasize)
541 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
542 bp->b_bufsize, bp->b_kvasize);
544 if (buf_mapped(bp)) {
545 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
546 (vm_page_t *)bp->b_pages, bp->b_npages);
552 * Cleanup after a clustered read or write.
553 * This is complicated by the fact that any of the buffers might have
554 * extra memory (if there were no empty buffer headers at allocbuf time)
555 * that we will need to shift around.
561 struct buf *nbp, *tbp;
565 * Must propagate errors to all the components.
567 if (bp->b_ioflags & BIO_ERROR)
570 if (buf_mapped(bp)) {
571 pmap_qremove(trunc_page((vm_offset_t) bp->b_data),
575 * Move memory from the large cluster buffer into the component
576 * buffers and mark IO as done on these.
578 for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
580 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
582 tbp->b_ioflags |= BIO_ERROR;
583 tbp->b_error = error;
585 tbp->b_dirtyoff = tbp->b_dirtyend = 0;
586 tbp->b_flags &= ~B_INVAL;
587 tbp->b_ioflags &= ~BIO_ERROR;
589 * XXX the bdwrite()/bqrelse() issued during
590 * cluster building clears B_RELBUF (see bqrelse()
591 * comment). If direct I/O was specified, we have
592 * to restore it here to allow the buffer and VM
595 if (tbp->b_flags & B_DIRECT)
596 tbp->b_flags |= B_RELBUF;
601 relpbuf(bp, &cluster_pbuf_freecnt);
607 * Implement modified write build for cluster.
609 * write_behind = 0 write behind disabled
610 * write_behind = 1 write behind normal (default)
611 * write_behind = 2 write behind backed-off
615 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len,
620 switch (write_behind) {
627 r = cluster_wbuild(vp, size, start_lbn, len, gbflags);
637 * Do clustered write for FFS.
640 * 1. Write is not sequential (write asynchronously)
641 * Write is sequential:
642 * 2. beginning of cluster - begin cluster
643 * 3. middle of a cluster - add to cluster
644 * 4. end of a cluster - asynchronously write cluster
647 cluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount,
651 int maxclen, cursize;
655 if (!unmapped_buf_allowed)
656 gbflags &= ~GB_UNMAPPED;
658 if (vp->v_type == VREG) {
659 async = DOINGASYNC(vp);
660 lblocksize = vp->v_mount->mnt_stat.f_iosize;
663 lblocksize = bp->b_bufsize;
666 KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
668 /* Initialize vnode to beginning of file. */
670 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
672 if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
673 (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
674 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
675 if (vp->v_clen != 0) {
677 * Next block is not sequential.
679 * If we are not writing at end of file, the process
680 * seeked to another point in the file since its last
681 * write, or we have reached our maximum cluster size,
682 * then push the previous cluster. Otherwise try
683 * reallocating to make it sequential.
685 * Change to algorithm: only push previous cluster if
686 * it was sequential from the point of view of the
687 * seqcount heuristic, otherwise leave the buffer
688 * intact so we can potentially optimize the I/O
689 * later on in the buf_daemon or update daemon
692 cursize = vp->v_lastw - vp->v_cstart + 1;
693 if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
694 lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
695 if (!async && seqcount > 0) {
696 cluster_wbuild_wb(vp, lblocksize,
697 vp->v_cstart, cursize, gbflags);
700 struct buf **bpp, **endbp;
701 struct cluster_save *buflist;
703 buflist = cluster_collectbufs(vp, bp, gbflags);
704 endbp = &buflist->bs_children
705 [buflist->bs_nchildren - 1];
706 if (VOP_REALLOCBLKS(vp, buflist)) {
708 * Failed, push the previous cluster
709 * if *really* writing sequentially
710 * in the logical file (seqcount > 1),
711 * otherwise delay it in the hopes that
712 * the low level disk driver can
713 * optimize the write ordering.
715 for (bpp = buflist->bs_children;
718 free(buflist, M_SEGMENT);
720 cluster_wbuild_wb(vp,
721 lblocksize, vp->v_cstart,
726 * Succeeded, keep building cluster.
728 for (bpp = buflist->bs_children;
731 free(buflist, M_SEGMENT);
733 vp->v_lasta = bp->b_blkno;
739 * Consider beginning a cluster. If at end of file, make
740 * cluster as large as possible, otherwise find size of
743 if ((vp->v_type == VREG) &&
744 ((u_quad_t) bp->b_offset + lblocksize) != filesize &&
745 (bp->b_blkno == bp->b_lblkno) &&
746 (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
747 bp->b_blkno == -1)) {
750 vp->v_lasta = bp->b_blkno;
751 vp->v_cstart = lbn + 1;
755 vp->v_clen = maxclen;
756 if (!async && maxclen == 0) { /* I/O not contiguous */
757 vp->v_cstart = lbn + 1;
759 } else { /* Wait for rest of cluster */
763 } else if (lbn == vp->v_cstart + vp->v_clen) {
765 * At end of cluster, write it out if seqcount tells us we
766 * are operating sequentially, otherwise let the buf or
767 * update daemon handle it.
771 cluster_wbuild_wb(vp, lblocksize, vp->v_cstart,
772 vp->v_clen + 1, gbflags);
775 vp->v_cstart = lbn + 1;
776 } else if (vm_page_count_severe()) {
778 * We are low on memory, get it going NOW
783 * In the middle of a cluster, so just delay the I/O for now.
788 vp->v_lasta = bp->b_blkno;
793 * This is an awful lot like cluster_rbuild...wish they could be combined.
794 * The last lbn argument is the current block on which I/O is being
795 * performed. Check to see that it doesn't fall in the middle of
796 * the current block (if last_bp == NULL).
799 cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len,
802 struct buf *bp, *tbp;
805 int totalwritten = 0;
806 int dbsize = btodb(size);
808 if (!unmapped_buf_allowed)
809 gbflags &= ~GB_UNMAPPED;
814 * If the buffer is not delayed-write (i.e. dirty), or it
815 * is delayed-write but either locked or inval, it cannot
816 * partake in the clustered write.
819 if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
820 (tbp->b_vflags & BV_BKGRDINPROG)) {
827 LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_LOCKPTR(bo))) {
832 if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
839 tbp->b_flags &= ~B_DONE;
842 * Extra memory in the buffer, punt on this buffer.
843 * XXX we could handle this in most cases, but we would
844 * have to push the extra memory down to after our max
845 * possible cluster size and then potentially pull it back
846 * up if the cluster was terminated prematurely--too much
849 if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
850 (B_CLUSTEROK | B_VMIO)) ||
851 (tbp->b_bcount != tbp->b_bufsize) ||
852 (tbp->b_bcount != size) ||
854 ((bp = (vp->v_vflag & VV_MD) != 0 ?
855 trypbuf(&cluster_pbuf_freecnt) :
856 getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
857 totalwritten += tbp->b_bufsize;
865 * We got a pbuf to make the cluster in.
868 TAILQ_INIT(&bp->b_cluster.cluster_head);
872 if (tbp->b_wcred != NOCRED)
873 bp->b_wcred = crhold(tbp->b_wcred);
875 bp->b_blkno = tbp->b_blkno;
876 bp->b_lblkno = tbp->b_lblkno;
877 bp->b_offset = tbp->b_offset;
880 * We are synthesizing a buffer out of vm_page_t's, but
881 * if the block size is not page aligned then the starting
882 * address may not be either. Inherit the b_data offset
883 * from the original buffer.
885 if ((gbflags & GB_UNMAPPED) == 0 ||
886 (tbp->b_flags & B_VMIO) == 0) {
887 bp->b_data = (char *)((vm_offset_t)bp->b_data |
888 ((vm_offset_t)tbp->b_data & PAGE_MASK));
890 bp->b_data = unmapped_buf;
892 bp->b_flags |= B_CLUSTER | (tbp->b_flags & (B_VMIO |
894 bp->b_iodone = cluster_callback;
897 * From this location in the file, scan forward to see
898 * if there are buffers with adjacent data that need to
899 * be written as well.
901 for (i = 0; i < len; ++i, ++start_lbn) {
902 if (i != 0) { /* If not the first buffer */
904 * If the adjacent data is not even in core it
905 * can't need to be written.
908 if ((tbp = gbincore(bo, start_lbn)) == NULL ||
909 (tbp->b_vflags & BV_BKGRDINPROG)) {
915 * If it IS in core, but has different
916 * characteristics, or is locked (which
917 * means it could be undergoing a background
918 * I/O or be in a weird state), then don't
922 LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
926 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
927 B_INVAL | B_DELWRI | B_NEEDCOMMIT))
928 != (B_DELWRI | B_CLUSTEROK |
929 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
930 tbp->b_wcred != bp->b_wcred) {
936 * Check that the combined cluster
937 * would make sense with regard to pages
938 * and would not be too large
940 if ((tbp->b_bcount != size) ||
941 ((bp->b_blkno + (dbsize * i)) !=
943 ((tbp->b_npages + bp->b_npages) >
944 (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
950 * Ok, it's passed all the tests,
951 * so remove it from the free list
952 * and mark it busy. We will use it.
955 tbp->b_flags &= ~B_DONE;
956 } /* end of code for non-first buffers only */
958 * If the IO is via the VM then we do some
959 * special VM hackery (yuck). Since the buffer's
960 * block size may not be page-aligned it is possible
961 * for a page to be shared between two buffers. We
962 * have to get rid of the duplication when building
965 if (tbp->b_flags & B_VMIO) {
968 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
970 vfs_drain_busy_pages(tbp);
971 } else { /* if not first buffer */
972 for (j = 0; j < tbp->b_npages; j += 1) {
974 if (vm_page_xbusied(m)) {
982 for (j = 0; j < tbp->b_npages; j += 1) {
985 vm_object_pip_add(m->object, 1);
986 if ((bp->b_npages == 0) ||
987 (bp->b_pages[bp->b_npages - 1] != m)) {
988 bp->b_pages[bp->b_npages] = m;
992 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
994 bp->b_bcount += size;
995 bp->b_bufsize += size;
997 * If any of the clustered buffers have their
998 * B_BARRIER flag set, transfer that request to
1001 bp->b_flags |= (tbp->b_flags & B_BARRIER);
1002 tbp->b_flags &= ~(B_DONE | B_BARRIER);
1003 tbp->b_flags |= B_ASYNC;
1004 tbp->b_ioflags &= ~BIO_ERROR;
1005 tbp->b_iocmd = BIO_WRITE;
1007 reassignbuf(tbp); /* put on clean list */
1008 bufobj_wref(tbp->b_bufobj);
1010 buf_track(tbp, __func__);
1011 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
1012 tbp, b_cluster.cluster_entry);
1015 if (buf_mapped(bp)) {
1016 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
1017 (vm_page_t *)bp->b_pages, bp->b_npages);
1019 if (bp->b_bufsize > bp->b_kvasize)
1021 "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
1022 bp->b_bufsize, bp->b_kvasize);
1023 totalwritten += bp->b_bufsize;
1025 bp->b_dirtyend = bp->b_bufsize;
1030 return totalwritten;
1034 * Collect together all the buffers in a cluster.
1035 * Plus add one additional buffer.
1037 static struct cluster_save *
1038 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int gbflags)
1040 struct cluster_save *buflist;
1045 len = vp->v_lastw - vp->v_cstart + 1;
1046 buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1047 M_SEGMENT, M_WAITOK);
1048 buflist->bs_nchildren = 0;
1049 buflist->bs_children = (struct buf **) (buflist + 1);
1050 for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
1051 (void)bread_gb(vp, lbn, last_bp->b_bcount, NOCRED,
1053 buflist->bs_children[i] = bp;
1054 if (bp->b_blkno == bp->b_lblkno)
1055 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno,
1058 buflist->bs_children[i] = bp = last_bp;
1059 if (bp->b_blkno == bp->b_lblkno)
1060 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1061 buflist->bs_nchildren = i + 1;