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");
66 static uma_zone_t cluster_pbuf_zone;
68 static void cluster_init(void *);
69 static struct cluster_save *cluster_collectbufs(struct vnode *vp,
70 struct buf *last_bp, int gbflags);
71 static struct buf *cluster_rbuild(struct vnode *vp, u_quad_t filesize,
72 daddr_t lbn, daddr_t blkno, long size, int run, int gbflags,
74 static void cluster_callback(struct buf *);
76 static int write_behind = 1;
77 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
78 "Cluster write-behind; 0: disable, 1: enable, 2: backed off");
80 static int read_max = 64;
81 SYSCTL_INT(_vfs, OID_AUTO, read_max, CTLFLAG_RW, &read_max, 0,
82 "Cluster read-ahead max block count");
84 static int read_min = 1;
85 SYSCTL_INT(_vfs, OID_AUTO, read_min, CTLFLAG_RW, &read_min, 0,
86 "Cluster read min block count");
88 SYSINIT(cluster, SI_SUB_CPU, SI_ORDER_ANY, cluster_init, NULL);
91 cluster_init(void *dummy)
94 cluster_pbuf_zone = pbuf_zsecond_create("clpbuf", nswbuf / 2);
98 * Read data to a buf, including read-ahead if we find this to be beneficial.
99 * cluster_read replaces bread.
102 cluster_read(struct vnode *vp, u_quad_t filesize, daddr_t lblkno, long size,
103 struct ucred *cred, long totread, int seqcount, int gbflags,
106 struct buf *bp, *rbp, *reqbp;
109 daddr_t blkno, origblkno;
110 int maxra, racluster;
117 if (!unmapped_buf_allowed)
118 gbflags &= ~GB_UNMAPPED;
121 * Try to limit the amount of read-ahead by a few
122 * ad-hoc parameters. This needs work!!!
124 racluster = vp->v_mount->mnt_iosize_max / size;
126 maxra = min(read_max, maxra);
127 maxra = min(nbuf/8, maxra);
128 if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize)
129 maxra = (filesize / size) - lblkno;
132 * get the requested block
134 error = getblkx(vp, lblkno, size, 0, 0, gbflags, &bp);
139 gbflags &= ~GB_NOSPARSE;
144 * if it is in the cache, then check to see if the reads have been
145 * sequential. If they have, then try some read-ahead, otherwise
146 * back-off on prospective read-aheads.
148 if (bp->b_flags & B_CACHE) {
151 } else if ((bp->b_flags & B_RAM) == 0) {
154 bp->b_flags &= ~B_RAM;
156 for (i = 1; i < maxra; i++) {
158 * Stop if the buffer does not exist or it
159 * is invalid (about to go away?)
161 rbp = gbincore(&vp->v_bufobj, lblkno+i);
162 if (rbp == NULL || (rbp->b_flags & B_INVAL))
166 * Set another read-ahead mark so we know
167 * to check again. (If we can lock the
168 * buffer without waiting)
170 if ((((i % racluster) == (racluster - 1)) ||
172 && (0 == BUF_LOCK(rbp,
173 LK_EXCLUSIVE | LK_NOWAIT, NULL))) {
174 rbp->b_flags |= B_RAM;
186 * If it isn't in the cache, then get a chunk from
187 * disk if sequential, otherwise just get the block.
190 off_t firstread = bp->b_offset;
194 KASSERT(bp->b_offset != NOOFFSET,
195 ("cluster_read: no buffer offset"));
200 * Adjust totread if needed
202 minread = read_min * size;
203 if (minread > totread)
207 * Compute the total number of blocks that we should read
210 if (firstread + totread > filesize)
211 totread = filesize - firstread;
212 nblks = howmany(totread, size);
213 if (nblks > racluster)
217 * Now compute the number of contiguous blocks.
220 error = VOP_BMAP(vp, lblkno, NULL,
221 &blkno, &ncontig, NULL);
223 * If this failed to map just do the original block.
225 if (error || blkno == -1)
230 * If we have contiguous data available do a cluster
231 * otherwise just read the requested block.
234 /* Account for our first block. */
235 ncontig = min(ncontig + 1, nblks);
238 bp = cluster_rbuild(vp, filesize, lblkno,
239 blkno, size, nblks, gbflags, bp);
240 lblkno += (bp->b_bufsize / size);
242 bp->b_flags |= B_RAM;
243 bp->b_iocmd = BIO_READ;
249 * handle the synchronous read so that it is available ASAP.
252 if ((bp->b_flags & B_CLUSTER) == 0) {
253 vfs_busy_pages(bp, 0);
255 bp->b_flags &= ~B_INVAL;
256 bp->b_ioflags &= ~BIO_ERROR;
257 if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL)
259 bp->b_iooffset = dbtob(bp->b_blkno);
263 PROC_LOCK(td->td_proc);
264 racct_add_buf(td->td_proc, bp, 0);
265 PROC_UNLOCK(td->td_proc);
268 td->td_ru.ru_inblock++;
272 * If we have been doing sequential I/O, then do some read-ahead.
274 while (lblkno < (origblkno + maxra)) {
275 error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL);
283 * We could throttle ncontig here by maxra but we might as
284 * well read the data if it is contiguous. We're throttled
285 * by racluster anyway.
288 ncontig = min(ncontig + 1, racluster);
289 rbp = cluster_rbuild(vp, filesize, lblkno, blkno,
290 size, ncontig, gbflags, NULL);
291 lblkno += (rbp->b_bufsize / size);
292 if (rbp->b_flags & B_DELWRI) {
297 rbp = getblk(vp, lblkno, size, 0, 0, gbflags);
299 if (rbp->b_flags & B_DELWRI) {
303 rbp->b_flags |= B_ASYNC | B_RAM;
304 rbp->b_iocmd = BIO_READ;
305 rbp->b_blkno = blkno;
307 if (rbp->b_flags & B_CACHE) {
308 rbp->b_flags &= ~B_ASYNC;
312 if ((rbp->b_flags & B_CLUSTER) == 0) {
313 vfs_busy_pages(rbp, 0);
315 rbp->b_flags &= ~B_INVAL;
316 rbp->b_ioflags &= ~BIO_ERROR;
317 if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL)
319 rbp->b_iooffset = dbtob(rbp->b_blkno);
323 PROC_LOCK(td->td_proc);
324 racct_add_buf(td->td_proc, rbp, 0);
325 PROC_UNLOCK(td->td_proc);
328 td->td_ru.ru_inblock++;
333 * Like bread, always brelse() the buffer when
334 * returning an error.
336 error = bufwait(reqbp);
346 * If blocks are contiguous on disk, use this to provide clustered
347 * read ahead. We will read as many blocks as possible sequentially
348 * and then parcel them up into logical blocks in the buffer hash table.
351 cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn,
352 daddr_t blkno, long size, int run, int gbflags, struct buf *fbp)
354 struct buf *bp, *tbp;
358 int i, inc, j, k, toff;
360 KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
361 ("cluster_rbuild: size %ld != f_iosize %jd\n",
362 size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
367 while ((u_quad_t) size * (lbn + run) > filesize) {
373 tbp->b_iocmd = BIO_READ;
375 tbp = getblk(vp, lbn, size, 0, 0, gbflags);
376 if (tbp->b_flags & B_CACHE)
378 tbp->b_flags |= B_ASYNC | B_RAM;
379 tbp->b_iocmd = BIO_READ;
381 tbp->b_blkno = blkno;
382 if( (tbp->b_flags & B_MALLOC) ||
383 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1) )
386 bp = uma_zalloc(cluster_pbuf_zone, M_NOWAIT);
391 * We are synthesizing a buffer out of vm_page_t's, but
392 * if the block size is not page aligned then the starting
393 * address may not be either. Inherit the b_data offset
394 * from the original buffer.
396 bp->b_flags = B_ASYNC | B_CLUSTER | B_VMIO;
397 if ((gbflags & GB_UNMAPPED) != 0) {
398 bp->b_data = unmapped_buf;
400 bp->b_data = (char *)((vm_offset_t)bp->b_data |
401 ((vm_offset_t)tbp->b_data & PAGE_MASK));
403 bp->b_iocmd = BIO_READ;
404 bp->b_iodone = cluster_callback;
407 bp->b_offset = tbp->b_offset;
408 KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
411 TAILQ_INIT(&bp->b_cluster.cluster_head);
418 for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
420 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
421 vfs_drain_busy_pages(tbp);
422 vm_object_pip_add(tbp->b_bufobj->bo_object,
424 for (k = 0; k < tbp->b_npages; k++)
425 vm_page_sbusy(tbp->b_pages[k]);
426 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
428 if ((bp->b_npages * PAGE_SIZE) +
429 round_page(size) > vp->v_mount->mnt_iosize_max) {
433 tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT |
434 (gbflags & GB_UNMAPPED));
436 /* Don't wait around for locked bufs. */
441 * Stop scanning if the buffer is fully valid
442 * (marked B_CACHE), or locked (may be doing a
443 * background write), or if the buffer is not
444 * VMIO backed. The clustering code can only deal
445 * with VMIO-backed buffers. The bo lock is not
446 * required for the BKGRDINPROG check since it
447 * can not be set without the buf lock.
449 if ((tbp->b_vflags & BV_BKGRDINPROG) ||
450 (tbp->b_flags & B_CACHE) ||
451 (tbp->b_flags & B_VMIO) == 0) {
457 * The buffer must be completely invalid in order to
458 * take part in the cluster. If it is partially valid
463 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
464 for (j = 0; tsize > 0; j++) {
465 toff = off & PAGE_MASK;
467 if (toff + tinc > PAGE_SIZE)
468 tinc = PAGE_SIZE - toff;
469 VM_OBJECT_ASSERT_WLOCKED(tbp->b_pages[j]->object);
470 if ((tbp->b_pages[j]->valid &
471 vm_page_bits(toff, tinc)) != 0)
473 if (vm_page_xbusied(tbp->b_pages[j]))
475 vm_object_pip_add(tbp->b_bufobj->bo_object, 1);
476 vm_page_sbusy(tbp->b_pages[j]);
482 vm_object_pip_add(tbp->b_bufobj->bo_object, -j);
483 for (k = 0; k < j; k++)
484 vm_page_sunbusy(tbp->b_pages[k]);
485 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
489 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
492 * Set a read-ahead mark as appropriate
494 if ((fbp && (i == 1)) || (i == (run - 1)))
495 tbp->b_flags |= B_RAM;
498 * Set the buffer up for an async read (XXX should
499 * we do this only if we do not wind up brelse()ing?).
500 * Set the block number if it isn't set, otherwise
501 * if it is make sure it matches the block number we
504 tbp->b_flags |= B_ASYNC;
505 tbp->b_iocmd = BIO_READ;
506 if (tbp->b_blkno == tbp->b_lblkno) {
508 } else if (tbp->b_blkno != bn) {
509 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
514 * XXX fbp from caller may not be B_ASYNC, but we are going
515 * to biodone() it in cluster_callback() anyway
518 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
519 tbp, b_cluster.cluster_entry);
520 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
521 for (j = 0; j < tbp->b_npages; j += 1) {
524 if ((bp->b_npages == 0) ||
525 (bp->b_pages[bp->b_npages-1] != m)) {
526 bp->b_pages[bp->b_npages] = m;
529 if (m->valid == VM_PAGE_BITS_ALL)
530 tbp->b_pages[j] = bogus_page;
532 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
534 * Don't inherit tbp->b_bufsize as it may be larger due to
535 * a non-page-aligned size. Instead just aggregate using
538 if (tbp->b_bcount != size)
539 printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
540 if (tbp->b_bufsize != size)
541 printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
542 bp->b_bcount += size;
543 bp->b_bufsize += size;
547 * Fully valid pages in the cluster are already good and do not need
548 * to be re-read from disk. Replace the page with bogus_page
550 VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
551 for (j = 0; j < bp->b_npages; j++) {
552 VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[j]->object);
553 if (bp->b_pages[j]->valid == VM_PAGE_BITS_ALL)
554 bp->b_pages[j] = bogus_page;
556 VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
557 if (bp->b_bufsize > bp->b_kvasize)
558 panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
559 bp->b_bufsize, bp->b_kvasize);
561 if (buf_mapped(bp)) {
562 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
563 (vm_page_t *)bp->b_pages, bp->b_npages);
569 * Cleanup after a clustered read or write.
570 * This is complicated by the fact that any of the buffers might have
571 * extra memory (if there were no empty buffer headers at allocbuf time)
572 * that we will need to shift around.
575 cluster_callback(struct buf *bp)
577 struct buf *nbp, *tbp;
581 * Must propagate errors to all the components.
583 if (bp->b_ioflags & BIO_ERROR)
586 if (buf_mapped(bp)) {
587 pmap_qremove(trunc_page((vm_offset_t) bp->b_data),
591 * Move memory from the large cluster buffer into the component
592 * buffers and mark IO as done on these.
594 for (tbp = TAILQ_FIRST(&bp->b_cluster.cluster_head);
596 nbp = TAILQ_NEXT(&tbp->b_cluster, cluster_entry);
598 tbp->b_ioflags |= BIO_ERROR;
599 tbp->b_error = error;
601 tbp->b_dirtyoff = tbp->b_dirtyend = 0;
602 tbp->b_flags &= ~B_INVAL;
603 tbp->b_ioflags &= ~BIO_ERROR;
605 * XXX the bdwrite()/bqrelse() issued during
606 * cluster building clears B_RELBUF (see bqrelse()
607 * comment). If direct I/O was specified, we have
608 * to restore it here to allow the buffer and VM
611 if (tbp->b_flags & B_DIRECT)
612 tbp->b_flags |= B_RELBUF;
617 uma_zfree(cluster_pbuf_zone, bp);
623 * Implement modified write build for cluster.
625 * write_behind = 0 write behind disabled
626 * write_behind = 1 write behind normal (default)
627 * write_behind = 2 write behind backed-off
631 cluster_wbuild_wb(struct vnode *vp, long size, daddr_t start_lbn, int len,
636 switch (write_behind) {
643 r = cluster_wbuild(vp, size, start_lbn, len, gbflags);
653 * Do clustered write for FFS.
656 * 1. Write is not sequential (write asynchronously)
657 * Write is sequential:
658 * 2. beginning of cluster - begin cluster
659 * 3. middle of a cluster - add to cluster
660 * 4. end of a cluster - asynchronously write cluster
663 cluster_write(struct vnode *vp, struct buf *bp, u_quad_t filesize, int seqcount,
667 int maxclen, cursize;
671 if (!unmapped_buf_allowed)
672 gbflags &= ~GB_UNMAPPED;
674 if (vp->v_type == VREG) {
675 async = DOINGASYNC(vp);
676 lblocksize = vp->v_mount->mnt_stat.f_iosize;
679 lblocksize = bp->b_bufsize;
682 KASSERT(bp->b_offset != NOOFFSET, ("cluster_write: no buffer offset"));
684 /* Initialize vnode to beginning of file. */
686 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
688 if (vp->v_clen == 0 || lbn != vp->v_lastw + 1 ||
689 (bp->b_blkno != vp->v_lasta + btodb(lblocksize))) {
690 maxclen = vp->v_mount->mnt_iosize_max / lblocksize - 1;
691 if (vp->v_clen != 0) {
693 * Next block is not sequential.
695 * If we are not writing at end of file, the process
696 * seeked to another point in the file since its last
697 * write, or we have reached our maximum cluster size,
698 * then push the previous cluster. Otherwise try
699 * reallocating to make it sequential.
701 * Change to algorithm: only push previous cluster if
702 * it was sequential from the point of view of the
703 * seqcount heuristic, otherwise leave the buffer
704 * intact so we can potentially optimize the I/O
705 * later on in the buf_daemon or update daemon
708 cursize = vp->v_lastw - vp->v_cstart + 1;
709 if (((u_quad_t) bp->b_offset + lblocksize) != filesize ||
710 lbn != vp->v_lastw + 1 || vp->v_clen <= cursize) {
711 if (!async && seqcount > 0) {
712 cluster_wbuild_wb(vp, lblocksize,
713 vp->v_cstart, cursize, gbflags);
716 struct buf **bpp, **endbp;
717 struct cluster_save *buflist;
719 buflist = cluster_collectbufs(vp, bp, gbflags);
720 endbp = &buflist->bs_children
721 [buflist->bs_nchildren - 1];
722 if (VOP_REALLOCBLKS(vp, buflist)) {
724 * Failed, push the previous cluster
725 * if *really* writing sequentially
726 * in the logical file (seqcount > 1),
727 * otherwise delay it in the hopes that
728 * the low level disk driver can
729 * optimize the write ordering.
731 for (bpp = buflist->bs_children;
734 free(buflist, M_SEGMENT);
736 cluster_wbuild_wb(vp,
737 lblocksize, vp->v_cstart,
742 * Succeeded, keep building cluster.
744 for (bpp = buflist->bs_children;
747 free(buflist, M_SEGMENT);
749 vp->v_lasta = bp->b_blkno;
755 * Consider beginning a cluster. If at end of file, make
756 * cluster as large as possible, otherwise find size of
759 if ((vp->v_type == VREG) &&
760 ((u_quad_t) bp->b_offset + lblocksize) != filesize &&
761 (bp->b_blkno == bp->b_lblkno) &&
762 (VOP_BMAP(vp, lbn, NULL, &bp->b_blkno, &maxclen, NULL) ||
763 bp->b_blkno == -1)) {
766 vp->v_lasta = bp->b_blkno;
767 vp->v_cstart = lbn + 1;
771 vp->v_clen = maxclen;
772 if (!async && maxclen == 0) { /* I/O not contiguous */
773 vp->v_cstart = lbn + 1;
775 } else { /* Wait for rest of cluster */
779 } else if (lbn == vp->v_cstart + vp->v_clen) {
781 * At end of cluster, write it out if seqcount tells us we
782 * are operating sequentially, otherwise let the buf or
783 * update daemon handle it.
787 cluster_wbuild_wb(vp, lblocksize, vp->v_cstart,
788 vp->v_clen + 1, gbflags);
791 vp->v_cstart = lbn + 1;
792 } else if (vm_page_count_severe()) {
794 * We are low on memory, get it going NOW
799 * In the middle of a cluster, so just delay the I/O for now.
804 vp->v_lasta = bp->b_blkno;
809 * This is an awful lot like cluster_rbuild...wish they could be combined.
810 * The last lbn argument is the current block on which I/O is being
811 * performed. Check to see that it doesn't fall in the middle of
812 * the current block (if last_bp == NULL).
815 cluster_wbuild(struct vnode *vp, long size, daddr_t start_lbn, int len,
818 struct buf *bp, *tbp;
821 int totalwritten = 0;
822 int dbsize = btodb(size);
824 if (!unmapped_buf_allowed)
825 gbflags &= ~GB_UNMAPPED;
830 * If the buffer is not delayed-write (i.e. dirty), or it
831 * is delayed-write but either locked or inval, it cannot
832 * partake in the clustered write.
835 if ((tbp = gbincore(&vp->v_bufobj, start_lbn)) == NULL ||
836 (tbp->b_vflags & BV_BKGRDINPROG)) {
843 LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK, BO_LOCKPTR(bo))) {
848 if ((tbp->b_flags & (B_INVAL | B_DELWRI)) != B_DELWRI) {
855 tbp->b_flags &= ~B_DONE;
858 * Extra memory in the buffer, punt on this buffer.
859 * XXX we could handle this in most cases, but we would
860 * have to push the extra memory down to after our max
861 * possible cluster size and then potentially pull it back
862 * up if the cluster was terminated prematurely--too much
865 if (((tbp->b_flags & (B_CLUSTEROK | B_MALLOC | B_VMIO)) !=
866 (B_CLUSTEROK | B_VMIO)) ||
867 (tbp->b_bcount != tbp->b_bufsize) ||
868 (tbp->b_bcount != size) ||
870 ((bp = uma_zalloc(cluster_pbuf_zone,
871 (vp->v_vflag & VV_MD) != 0 ? M_NOWAIT : M_WAITOK)) == NULL)) {
872 totalwritten += tbp->b_bufsize;
880 * We got a pbuf to make the cluster in.
883 TAILQ_INIT(&bp->b_cluster.cluster_head);
887 if (tbp->b_wcred != NOCRED)
888 bp->b_wcred = crhold(tbp->b_wcred);
890 bp->b_blkno = tbp->b_blkno;
891 bp->b_lblkno = tbp->b_lblkno;
892 bp->b_offset = tbp->b_offset;
895 * We are synthesizing a buffer out of vm_page_t's, but
896 * if the block size is not page aligned then the starting
897 * address may not be either. Inherit the b_data offset
898 * from the original buffer.
900 if ((gbflags & GB_UNMAPPED) == 0 ||
901 (tbp->b_flags & B_VMIO) == 0) {
902 bp->b_data = (char *)((vm_offset_t)bp->b_data |
903 ((vm_offset_t)tbp->b_data & PAGE_MASK));
905 bp->b_data = unmapped_buf;
907 bp->b_flags |= B_CLUSTER | (tbp->b_flags & (B_VMIO |
909 bp->b_iodone = cluster_callback;
912 * From this location in the file, scan forward to see
913 * if there are buffers with adjacent data that need to
914 * be written as well.
916 for (i = 0; i < len; ++i, ++start_lbn) {
917 if (i != 0) { /* If not the first buffer */
919 * If the adjacent data is not even in core it
920 * can't need to be written.
923 if ((tbp = gbincore(bo, start_lbn)) == NULL ||
924 (tbp->b_vflags & BV_BKGRDINPROG)) {
930 * If it IS in core, but has different
931 * characteristics, or is locked (which
932 * means it could be undergoing a background
933 * I/O or be in a weird state), then don't
937 LK_EXCLUSIVE | LK_NOWAIT | LK_INTERLOCK,
941 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
942 B_INVAL | B_DELWRI | B_NEEDCOMMIT))
943 != (B_DELWRI | B_CLUSTEROK |
944 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
945 tbp->b_wcred != bp->b_wcred) {
951 * Check that the combined cluster
952 * would make sense with regard to pages
953 * and would not be too large
955 if ((tbp->b_bcount != size) ||
956 ((bp->b_blkno + (dbsize * i)) !=
958 ((tbp->b_npages + bp->b_npages) >
959 (vp->v_mount->mnt_iosize_max / PAGE_SIZE))) {
965 * Ok, it's passed all the tests,
966 * so remove it from the free list
967 * and mark it busy. We will use it.
970 tbp->b_flags &= ~B_DONE;
971 } /* end of code for non-first buffers only */
973 * If the IO is via the VM then we do some
974 * special VM hackery (yuck). Since the buffer's
975 * block size may not be page-aligned it is possible
976 * for a page to be shared between two buffers. We
977 * have to get rid of the duplication when building
980 if (tbp->b_flags & B_VMIO) {
983 VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
985 vfs_drain_busy_pages(tbp);
986 } else { /* if not first buffer */
987 for (j = 0; j < tbp->b_npages; j += 1) {
989 if (vm_page_xbusied(m)) {
997 for (j = 0; j < tbp->b_npages; j += 1) {
1000 vm_object_pip_add(m->object, 1);
1001 if ((bp->b_npages == 0) ||
1002 (bp->b_pages[bp->b_npages - 1] != m)) {
1003 bp->b_pages[bp->b_npages] = m;
1007 VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
1009 bp->b_bcount += size;
1010 bp->b_bufsize += size;
1012 * If any of the clustered buffers have their
1013 * B_BARRIER flag set, transfer that request to
1016 bp->b_flags |= (tbp->b_flags & B_BARRIER);
1017 tbp->b_flags &= ~(B_DONE | B_BARRIER);
1018 tbp->b_flags |= B_ASYNC;
1019 tbp->b_ioflags &= ~BIO_ERROR;
1020 tbp->b_iocmd = BIO_WRITE;
1022 reassignbuf(tbp); /* put on clean list */
1023 bufobj_wref(tbp->b_bufobj);
1025 buf_track(tbp, __func__);
1026 TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
1027 tbp, b_cluster.cluster_entry);
1030 if (buf_mapped(bp)) {
1031 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
1032 (vm_page_t *)bp->b_pages, bp->b_npages);
1034 if (bp->b_bufsize > bp->b_kvasize)
1036 "cluster_wbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
1037 bp->b_bufsize, bp->b_kvasize);
1038 totalwritten += bp->b_bufsize;
1040 bp->b_dirtyend = bp->b_bufsize;
1045 return totalwritten;
1049 * Collect together all the buffers in a cluster.
1050 * Plus add one additional buffer.
1052 static struct cluster_save *
1053 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int gbflags)
1055 struct cluster_save *buflist;
1060 len = vp->v_lastw - vp->v_cstart + 1;
1061 buflist = malloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1062 M_SEGMENT, M_WAITOK);
1063 buflist->bs_nchildren = 0;
1064 buflist->bs_children = (struct buf **) (buflist + 1);
1065 for (lbn = vp->v_cstart, i = 0; i < len; lbn++, i++) {
1066 (void)bread_gb(vp, lbn, last_bp->b_bcount, NOCRED,
1068 buflist->bs_children[i] = bp;
1069 if (bp->b_blkno == bp->b_lblkno)
1070 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno,
1073 buflist->bs_children[i] = bp = last_bp;
1074 if (bp->b_blkno == bp->b_lblkno)
1075 VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL, NULL);
1076 buflist->bs_nchildren = i + 1;