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 * 3. All advertising materials mentioning features or use of this software
16 * must display the following acknowledgement:
17 * This product includes software developed by the University of
18 * California, Berkeley and its contributors.
19 * 4. Neither the name of the University nor the names of its contributors
20 * may be used to endorse or promote products derived from this software
21 * without specific prior written permission.
23 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 * @(#)vfs_cluster.c 8.7 (Berkeley) 2/13/94
36 * $FreeBSD: src/sys/kern/vfs_cluster.c,v 1.92.2.9 2001/11/18 07:10:59 dillon Exp $
37 * $DragonFly: src/sys/kern/vfs_cluster.c,v 1.40 2008/07/14 03:09:00 dillon Exp $
40 #include "opt_debug_cluster.h"
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/kernel.h>
47 #include <sys/vnode.h>
48 #include <sys/malloc.h>
49 #include <sys/mount.h>
50 #include <sys/resourcevar.h>
51 #include <sys/vmmeter.h>
53 #include <vm/vm_object.h>
54 #include <vm/vm_page.h>
55 #include <sys/sysctl.h>
58 #include <vm/vm_page2.h>
60 #include <machine/limits.h>
62 #if defined(CLUSTERDEBUG)
63 #include <sys/sysctl.h>
64 static int rcluster= 0;
65 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, "");
68 static MALLOC_DEFINE(M_SEGMENT, "cluster_save", "cluster_save buffer");
70 static struct cluster_save *
71 cluster_collectbufs (struct vnode *vp, struct buf *last_bp,
74 cluster_rbuild (struct vnode *vp, off_t filesize, off_t loffset,
75 off_t doffset, int blksize, int run,
77 static void cluster_callback (struct bio *);
78 static void cluster_setram (struct buf *);
80 static int write_behind = 1;
81 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
82 "Cluster write-behind setting");
83 static int max_readahead = 2 * 1024 * 1024;
84 SYSCTL_INT(_vfs, OID_AUTO, max_readahead, CTLFLAG_RW, &max_readahead, 0,
85 "Limit in bytes for desired cluster read-ahead");
87 extern vm_page_t bogus_page;
89 extern int cluster_pbuf_freecnt;
92 * This replaces bread.
94 * filesize - read-ahead @ blksize will not cross this boundary
95 * loffset - loffset for returned *bpp
96 * blksize - blocksize for returned *bpp and read-ahead bps
97 * minreq - minimum (not a hard minimum) in bytes, typically reflects
98 * a higher level uio resid.
99 * maxreq - maximum (sequential heuristic) in bytes (highet typ ~2MB)
100 * bpp - return buffer (*bpp) for (loffset,blksize)
103 cluster_readx(struct vnode *vp, off_t filesize, off_t loffset,
104 int blksize, size_t minreq, size_t maxreq, struct buf **bpp)
106 struct buf *bp, *rbp, *reqbp;
117 * Calculate the desired read-ahead in blksize'd blocks (maxra).
118 * To do this we calculate maxreq.
120 * maxreq typically starts out as a sequential heuristic. If the
121 * high level uio/resid is bigger (minreq), we pop maxreq up to
122 * minreq. This represents the case where random I/O is being
123 * performed by the userland is issuing big read()'s.
125 * Then we limit maxreq to max_readahead to ensure it is a reasonable
128 * Finally we must ensure that (loffset + maxreq) does not cross the
129 * boundary (filesize) for the current blocksize. If we allowed it
130 * to cross we could end up with buffers past the boundary with the
131 * wrong block size (HAMMER large-data areas use mixed block sizes).
132 * minreq is also absolutely limited to filesize.
136 /* minreq not used beyond this point */
138 if (maxreq > max_readahead) {
139 maxreq = max_readahead;
140 if (maxreq > 16 * 1024 * 1024)
141 maxreq = 16 * 1024 * 1024;
143 if (maxreq < blksize)
145 if (loffset + maxreq > filesize) {
146 if (loffset > filesize)
149 maxreq = filesize - loffset;
152 maxra = (int)(maxreq / blksize);
155 * Get the requested block.
160 *bpp = reqbp = bp = getblk(vp, loffset, blksize, 0, 0);
161 origoffset = loffset;
164 * Calculate the maximum cluster size for a single I/O, used
165 * by cluster_rbuild().
167 maxrbuild = vmaxiosize(vp) / blksize;
170 * if it is in the cache, then check to see if the reads have been
171 * sequential. If they have, then try some read-ahead, otherwise
172 * back-off on prospective read-aheads.
174 if (bp->b_flags & B_CACHE) {
176 * Not sequential, do not do any read-ahead
182 * No read-ahead mark, do not do any read-ahead
185 if ((bp->b_flags & B_RAM) == 0)
189 * We hit a read-ahead-mark, figure out how much read-ahead
190 * to do (maxra) and where to start (loffset).
192 * Shortcut the scan. Typically the way this works is that
193 * we've built up all the blocks inbetween except for the
194 * last in previous iterations, so if the second-to-last
195 * block is present we just skip ahead to it.
197 * This algorithm has O(1) cpu in the steady state no
198 * matter how large maxra is.
200 bp->b_flags &= ~B_RAM;
202 if (findblk(vp, loffset + (maxra - 2) * blksize, FINDBLK_TEST))
207 if (findblk(vp, loffset + i * blksize,
208 FINDBLK_TEST) == NULL) {
215 * We got everything or everything is in the cache, no
222 * Calculate where to start the read-ahead and how much
223 * to do. Generally speaking we want to read-ahead by
224 * (maxra) when we've found a read-ahead mark. We do
225 * not want to reduce maxra here as it will cause
226 * successive read-ahead I/O's to be smaller and smaller.
228 * However, we have to make sure we don't break the
229 * filesize limitation for the clustered operation.
231 loffset += i * blksize;
234 if (loffset >= filesize)
236 if (loffset + maxra * blksize > filesize) {
237 maxreq = filesize - loffset;
238 maxra = (int)(maxreq / blksize);
241 __debugvar off_t firstread = bp->b_loffset;
245 * Set-up synchronous read for bp.
247 bp->b_cmd = BUF_CMD_READ;
248 bp->b_bio1.bio_done = biodone_sync;
249 bp->b_bio1.bio_flags |= BIO_SYNC;
251 KASSERT(firstread != NOOFFSET,
252 ("cluster_read: no buffer offset"));
255 * nblks is our cluster_rbuild request size, limited
256 * primarily by the device.
258 if ((nblks = maxra) > maxrbuild)
264 error = VOP_BMAP(vp, loffset, &doffset,
265 &burstbytes, NULL, BUF_CMD_READ);
267 goto single_block_read;
268 if (nblks > burstbytes / blksize)
269 nblks = burstbytes / blksize;
270 if (doffset == NOOFFSET)
271 goto single_block_read;
273 goto single_block_read;
275 bp = cluster_rbuild(vp, filesize, loffset,
276 doffset, blksize, nblks, bp);
277 loffset += bp->b_bufsize;
278 maxra -= bp->b_bufsize / blksize;
282 * If it isn't in the cache, then get a chunk from
283 * disk if sequential, otherwise just get the block.
292 * If B_CACHE was not set issue bp. bp will either be an
293 * asynchronous cluster buf or a synchronous single-buf.
294 * If it is a single buf it will be the same as reqbp.
296 * NOTE: Once an async cluster buf is issued bp becomes invalid.
299 #if defined(CLUSTERDEBUG)
301 kprintf("S(%012jx,%d,%d)\n",
302 (intmax_t)bp->b_loffset, bp->b_bcount, maxra);
304 if ((bp->b_flags & B_CLUSTER) == 0)
305 vfs_busy_pages(vp, bp);
306 bp->b_flags &= ~(B_ERROR|B_INVAL);
307 vn_strategy(vp, &bp->b_bio1);
313 * If we have been doing sequential I/O, then do some read-ahead.
314 * The code above us should have positioned us at the next likely
317 * Only mess with buffers which we can immediately lock. HAMMER
318 * will do device-readahead irrespective of what the blocks
321 while (error == 0 && maxra > 0) {
326 rbp = getblk(vp, loffset, blksize,
327 GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
330 if ((rbp->b_flags & B_CACHE)) {
336 * An error from the read-ahead bmap has nothing to do
337 * with the caller's original request.
339 tmp_error = VOP_BMAP(vp, loffset, &doffset,
340 &burstbytes, NULL, BUF_CMD_READ);
341 if (tmp_error || doffset == NOOFFSET) {
342 rbp->b_flags |= B_INVAL;
347 if ((nblks = maxra) > maxrbuild)
349 if (nblks > burstbytes / blksize)
350 nblks = burstbytes / blksize;
355 rbp->b_cmd = BUF_CMD_READ;
356 /*rbp->b_flags |= B_AGE*/;
360 rbp = cluster_rbuild(vp, filesize, loffset,
364 rbp->b_bio2.bio_offset = doffset;
367 #if defined(CLUSTERDEBUG)
370 kprintf("A+(%012jx,%d,%jd) "
371 "doff=%012jx minr=%zd ra=%d\n",
372 (intmax_t)loffset, rbp->b_bcount,
373 (intmax_t)(loffset - origoffset),
374 (intmax_t)doffset, minreq, maxra);
376 kprintf("A-(%012jx,%d,%jd) "
377 "doff=%012jx minr=%zd ra=%d\n",
378 (intmax_t)rbp->b_loffset, rbp->b_bcount,
379 (intmax_t)(loffset - origoffset),
380 (intmax_t)doffset, minreq, maxra);
384 rbp->b_flags &= ~(B_ERROR|B_INVAL);
386 if ((rbp->b_flags & B_CLUSTER) == 0)
387 vfs_busy_pages(vp, rbp);
389 loffset += rbp->b_bufsize;
390 maxra -= rbp->b_bufsize / blksize;
391 vn_strategy(vp, &rbp->b_bio1);
392 /* rbp invalid now */
396 * Wait for our original buffer to complete its I/O. reqbp will
397 * be NULL if the original buffer was B_CACHE. We are returning
398 * (*bpp) which is the same as reqbp when reqbp != NULL.
402 KKASSERT(reqbp->b_bio1.bio_flags & BIO_SYNC);
403 error = biowait(&reqbp->b_bio1, "clurd");
409 * If blocks are contiguous on disk, use this to provide clustered
410 * read ahead. We will read as many blocks as possible sequentially
411 * and then parcel them up into logical blocks in the buffer hash table.
413 * This function either returns a cluster buf or it returns fbp. fbp is
414 * already expected to be set up as a synchronous or asynchronous request.
416 * If a cluster buf is returned it will always be async.
419 cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset, off_t doffset,
420 int blksize, int run, struct buf *fbp)
422 struct buf *bp, *tbp;
425 int maxiosize = vmaxiosize(vp);
430 while (loffset + run * blksize > filesize) {
435 tbp->b_bio2.bio_offset = doffset;
436 if((tbp->b_flags & B_MALLOC) ||
437 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1)) {
441 bp = trypbuf_kva(&cluster_pbuf_freecnt);
447 * We are synthesizing a buffer out of vm_page_t's, but
448 * if the block size is not page aligned then the starting
449 * address may not be either. Inherit the b_data offset
450 * from the original buffer.
452 bp->b_data = (char *)((vm_offset_t)bp->b_data |
453 ((vm_offset_t)tbp->b_data & PAGE_MASK));
454 bp->b_flags |= B_CLUSTER | B_VMIO;
455 bp->b_cmd = BUF_CMD_READ;
456 bp->b_bio1.bio_done = cluster_callback; /* default to async */
457 bp->b_bio1.bio_caller_info1.cluster_head = NULL;
458 bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
459 bp->b_loffset = loffset;
460 bp->b_bio2.bio_offset = doffset;
461 KASSERT(bp->b_loffset != NOOFFSET,
462 ("cluster_rbuild: no buffer offset"));
466 bp->b_xio.xio_npages = 0;
468 for (boffset = doffset, i = 0; i < run; ++i, boffset += blksize) {
470 if ((bp->b_xio.xio_npages * PAGE_SIZE) +
471 round_page(blksize) > maxiosize) {
476 * Shortcut some checks and try to avoid buffers that
477 * would block in the lock. The same checks have to
478 * be made again after we officially get the buffer.
480 tbp = getblk(vp, loffset + i * blksize, blksize,
481 GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
484 for (j = 0; j < tbp->b_xio.xio_npages; j++) {
485 if (tbp->b_xio.xio_pages[j]->valid)
488 if (j != tbp->b_xio.xio_npages) {
494 * Stop scanning if the buffer is fuly valid
495 * (marked B_CACHE), or locked (may be doing a
496 * background write), or if the buffer is not
497 * VMIO backed. The clustering code can only deal
498 * with VMIO-backed buffers.
500 if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
501 (tbp->b_flags & B_VMIO) == 0 ||
502 (LIST_FIRST(&tbp->b_dep) != NULL &&
510 * The buffer must be completely invalid in order to
511 * take part in the cluster. If it is partially valid
514 for (j = 0;j < tbp->b_xio.xio_npages; j++) {
515 if (tbp->b_xio.xio_pages[j]->valid)
518 if (j != tbp->b_xio.xio_npages) {
524 * Set a read-ahead mark as appropriate. Always
525 * set the read-ahead mark at (run - 1). It is
526 * unclear why we were also setting it at i == 1.
528 if (/*i == 1 ||*/ i == (run - 1))
532 * Depress the priority of buffers not explicitly
535 /* tbp->b_flags |= B_AGE; */
538 * Set the block number if it isn't set, otherwise
539 * if it is make sure it matches the block number we
542 if (tbp->b_bio2.bio_offset == NOOFFSET) {
543 tbp->b_bio2.bio_offset = boffset;
544 } else if (tbp->b_bio2.bio_offset != boffset) {
551 * The passed-in tbp (i == 0) will already be set up for
552 * async or sync operation. All other tbp's acquire in
553 * our loop are set up for async operation.
555 tbp->b_cmd = BUF_CMD_READ;
557 cluster_append(&bp->b_bio1, tbp);
558 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
561 m = tbp->b_xio.xio_pages[j];
562 vm_page_busy_wait(m, FALSE, "clurpg");
565 vm_object_pip_add(m->object, 1);
566 if ((bp->b_xio.xio_npages == 0) ||
567 (bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
568 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
569 bp->b_xio.xio_npages++;
571 if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
572 tbp->b_xio.xio_pages[j] = bogus_page;
575 * XXX shouldn't this be += size for both, like in
578 * Don't inherit tbp->b_bufsize as it may be larger due to
579 * a non-page-aligned size. Instead just aggregate using
582 if (tbp->b_bcount != blksize)
583 kprintf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, blksize);
584 if (tbp->b_bufsize != blksize)
585 kprintf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, blksize);
586 bp->b_bcount += blksize;
587 bp->b_bufsize += blksize;
591 * Fully valid pages in the cluster are already good and do not need
592 * to be re-read from disk. Replace the page with bogus_page
594 for (j = 0; j < bp->b_xio.xio_npages; j++) {
595 if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
597 bp->b_xio.xio_pages[j] = bogus_page;
600 if (bp->b_bufsize > bp->b_kvasize) {
601 panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)",
602 bp->b_bufsize, bp->b_kvasize);
604 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
605 (vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages);
611 * Cleanup after a clustered read or write.
612 * This is complicated by the fact that any of the buffers might have
613 * extra memory (if there were no empty buffer headers at allocbuf time)
614 * that we will need to shift around.
616 * The returned bio is &bp->b_bio1
619 cluster_callback(struct bio *bio)
621 struct buf *bp = bio->bio_buf;
626 * Must propogate errors to all the components. A short read (EOF)
627 * is a critical error.
629 if (bp->b_flags & B_ERROR) {
631 } else if (bp->b_bcount != bp->b_bufsize) {
632 panic("cluster_callback: unexpected EOF on cluster %p!", bio);
635 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
637 * Move memory from the large cluster buffer into the component
638 * buffers and mark IO as done on these. Since the memory map
639 * is the same, no actual copying is required.
641 while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) {
642 bio->bio_caller_info1.cluster_head = tbp->b_cluster_next;
644 tbp->b_flags |= B_ERROR | B_IODEBUG;
645 tbp->b_error = error;
647 tbp->b_dirtyoff = tbp->b_dirtyend = 0;
648 tbp->b_flags &= ~(B_ERROR|B_INVAL);
649 tbp->b_flags |= B_IODEBUG;
651 * XXX the bdwrite()/bqrelse() issued during
652 * cluster building clears B_RELBUF (see bqrelse()
653 * comment). If direct I/O was specified, we have
654 * to restore it here to allow the buffer and VM
657 if (tbp->b_flags & B_DIRECT)
658 tbp->b_flags |= B_RELBUF;
660 biodone(&tbp->b_bio1);
662 relpbuf(bp, &cluster_pbuf_freecnt);
668 * Implement modified write build for cluster.
670 * write_behind = 0 write behind disabled
671 * write_behind = 1 write behind normal (default)
672 * write_behind = 2 write behind backed-off
676 cluster_wbuild_wb(struct vnode *vp, int blksize, off_t start_loffset, int len)
680 switch(write_behind) {
682 if (start_loffset < len)
684 start_loffset -= len;
687 r = cluster_wbuild(vp, blksize, start_loffset, len);
697 * Do clustered write for FFS.
700 * 1. Write is not sequential (write asynchronously)
701 * Write is sequential:
702 * 2. beginning of cluster - begin cluster
703 * 3. middle of a cluster - add to cluster
704 * 4. end of a cluster - asynchronously write cluster
707 cluster_write(struct buf *bp, off_t filesize, int blksize, int seqcount)
711 int maxclen, cursize;
715 if (vp->v_type == VREG)
716 async = vp->v_mount->mnt_flag & MNT_ASYNC;
719 loffset = bp->b_loffset;
720 KASSERT(bp->b_loffset != NOOFFSET,
721 ("cluster_write: no buffer offset"));
723 /* Initialize vnode to beginning of file. */
725 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
727 if (vp->v_clen == 0 || loffset != vp->v_lastw + blksize ||
728 bp->b_bio2.bio_offset == NOOFFSET ||
729 (bp->b_bio2.bio_offset != vp->v_lasta + blksize)) {
730 maxclen = vmaxiosize(vp);
731 if (vp->v_clen != 0) {
733 * Next block is not sequential.
735 * If we are not writing at end of file, the process
736 * seeked to another point in the file since its last
737 * write, or we have reached our maximum cluster size,
738 * then push the previous cluster. Otherwise try
739 * reallocating to make it sequential.
741 * Change to algorithm: only push previous cluster if
742 * it was sequential from the point of view of the
743 * seqcount heuristic, otherwise leave the buffer
744 * intact so we can potentially optimize the I/O
745 * later on in the buf_daemon or update daemon
748 cursize = vp->v_lastw - vp->v_cstart + blksize;
749 if (bp->b_loffset + blksize != filesize ||
750 loffset != vp->v_lastw + blksize || vp->v_clen <= cursize) {
751 if (!async && seqcount > 0) {
752 cluster_wbuild_wb(vp, blksize,
753 vp->v_cstart, cursize);
756 struct buf **bpp, **endbp;
757 struct cluster_save *buflist;
759 buflist = cluster_collectbufs(vp, bp, blksize);
760 endbp = &buflist->bs_children
761 [buflist->bs_nchildren - 1];
762 if (VOP_REALLOCBLKS(vp, buflist)) {
764 * Failed, push the previous cluster
765 * if *really* writing sequentially
766 * in the logical file (seqcount > 1),
767 * otherwise delay it in the hopes that
768 * the low level disk driver can
769 * optimize the write ordering.
771 for (bpp = buflist->bs_children;
774 kfree(buflist, M_SEGMENT);
776 cluster_wbuild_wb(vp,
777 blksize, vp->v_cstart,
782 * Succeeded, keep building cluster.
784 for (bpp = buflist->bs_children;
787 kfree(buflist, M_SEGMENT);
788 vp->v_lastw = loffset;
789 vp->v_lasta = bp->b_bio2.bio_offset;
795 * Consider beginning a cluster. If at end of file, make
796 * cluster as large as possible, otherwise find size of
799 if ((vp->v_type == VREG) &&
800 bp->b_loffset + blksize != filesize &&
801 (bp->b_bio2.bio_offset == NOOFFSET) &&
802 (VOP_BMAP(vp, loffset, &bp->b_bio2.bio_offset, &maxclen, NULL, BUF_CMD_WRITE) ||
803 bp->b_bio2.bio_offset == NOOFFSET)) {
806 vp->v_lasta = bp->b_bio2.bio_offset;
807 vp->v_cstart = loffset + blksize;
808 vp->v_lastw = loffset;
811 if (maxclen > blksize)
812 vp->v_clen = maxclen - blksize;
815 if (!async && vp->v_clen == 0) { /* I/O not contiguous */
816 vp->v_cstart = loffset + blksize;
818 } else { /* Wait for rest of cluster */
819 vp->v_cstart = loffset;
822 } else if (loffset == vp->v_cstart + vp->v_clen) {
824 * At end of cluster, write it out if seqcount tells us we
825 * are operating sequentially, otherwise let the buf or
826 * update daemon handle it.
830 cluster_wbuild_wb(vp, blksize, vp->v_cstart,
831 vp->v_clen + blksize);
833 vp->v_cstart = loffset + blksize;
834 } else if (vm_page_count_severe()) {
836 * We are low on memory, get it going NOW
841 * In the middle of a cluster, so just delay the I/O for now.
845 vp->v_lastw = loffset;
846 vp->v_lasta = bp->b_bio2.bio_offset;
851 * This is an awful lot like cluster_rbuild...wish they could be combined.
852 * The last lbn argument is the current block on which I/O is being
853 * performed. Check to see that it doesn't fall in the middle of
854 * the current block (if last_bp == NULL).
857 cluster_wbuild(struct vnode *vp, int blksize, off_t start_loffset, int bytes)
859 struct buf *bp, *tbp;
861 int totalwritten = 0;
862 int maxiosize = vmaxiosize(vp);
866 * If the buffer is not delayed-write (i.e. dirty), or it
867 * is delayed-write but either locked or inval, it cannot
868 * partake in the clustered write.
870 tbp = findblk(vp, start_loffset, FINDBLK_NBLOCK);
872 (tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI ||
873 (LIST_FIRST(&tbp->b_dep) && buf_checkwrite(tbp))) {
876 start_loffset += blksize;
881 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
884 * Extra memory in the buffer, punt on this buffer.
885 * XXX we could handle this in most cases, but we would
886 * have to push the extra memory down to after our max
887 * possible cluster size and then potentially pull it back
888 * up if the cluster was terminated prematurely--too much
891 if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
892 (tbp->b_bcount != tbp->b_bufsize) ||
893 (tbp->b_bcount != blksize) ||
894 (bytes == blksize) ||
895 ((bp = getpbuf_kva(&cluster_pbuf_freecnt)) == NULL)) {
896 totalwritten += tbp->b_bufsize;
898 start_loffset += blksize;
904 * Set up the pbuf. Track our append point with b_bcount
905 * and b_bufsize. b_bufsize is not used by the device but
906 * our caller uses it to loop clusters and we use it to
907 * detect a premature EOF on the block device.
911 bp->b_xio.xio_npages = 0;
912 bp->b_loffset = tbp->b_loffset;
913 bp->b_bio2.bio_offset = tbp->b_bio2.bio_offset;
916 * We are synthesizing a buffer out of vm_page_t's, but
917 * if the block size is not page aligned then the starting
918 * address may not be either. Inherit the b_data offset
919 * from the original buffer.
921 bp->b_data = (char *)((vm_offset_t)bp->b_data |
922 ((vm_offset_t)tbp->b_data & PAGE_MASK));
923 bp->b_flags &= ~B_ERROR;
924 bp->b_flags |= B_CLUSTER | B_BNOCLIP |
925 (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
926 bp->b_bio1.bio_caller_info1.cluster_head = NULL;
927 bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
930 * From this location in the file, scan forward to see
931 * if there are buffers with adjacent data that need to
932 * be written as well.
934 for (i = 0; i < bytes; (i += blksize), (start_loffset += blksize)) {
935 if (i != 0) { /* If not the first buffer */
936 tbp = findblk(vp, start_loffset,
939 * Buffer not found or could not be locked
946 * If it IS in core, but has different
947 * characteristics, then don't cluster
950 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
951 B_INVAL | B_DELWRI | B_NEEDCOMMIT))
952 != (B_DELWRI | B_CLUSTEROK |
953 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
954 (tbp->b_flags & B_LOCKED) ||
955 (LIST_FIRST(&tbp->b_dep) &&
963 * Check that the combined cluster
964 * would make sense with regard to pages
965 * and would not be too large
967 if ((tbp->b_bcount != blksize) ||
968 ((bp->b_bio2.bio_offset + i) !=
969 tbp->b_bio2.bio_offset) ||
970 ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
971 (maxiosize / PAGE_SIZE))) {
976 * Ok, it's passed all the tests,
977 * so remove it from the free list
978 * and mark it busy. We will use it.
981 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
982 } /* end of code for non-first buffers only */
985 * If the IO is via the VM then we do some
986 * special VM hackery (yuck). Since the buffer's
987 * block size may not be page-aligned it is possible
988 * for a page to be shared between two buffers. We
989 * have to get rid of the duplication when building
992 if (tbp->b_flags & B_VMIO) {
995 if (i != 0) { /* if not first buffer */
996 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
997 m = tbp->b_xio.xio_pages[j];
998 if (m->flags & PG_BUSY) {
1005 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
1006 m = tbp->b_xio.xio_pages[j];
1007 vm_page_busy_wait(m, FALSE, "clurpg");
1008 vm_page_io_start(m);
1010 vm_object_pip_add(m->object, 1);
1011 if ((bp->b_xio.xio_npages == 0) ||
1012 (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
1013 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
1014 bp->b_xio.xio_npages++;
1018 bp->b_bcount += blksize;
1019 bp->b_bufsize += blksize;
1022 tbp->b_flags &= ~B_ERROR;
1023 tbp->b_cmd = BUF_CMD_WRITE;
1025 cluster_append(&bp->b_bio1, tbp);
1028 * check for latent dependencies to be handled
1030 if (LIST_FIRST(&tbp->b_dep) != NULL)
1034 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
1035 (vm_page_t *) bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1036 if (bp->b_bufsize > bp->b_kvasize) {
1038 "cluster_wbuild: b_bufsize(%d) > b_kvasize(%d)\n",
1039 bp->b_bufsize, bp->b_kvasize);
1041 totalwritten += bp->b_bufsize;
1043 bp->b_dirtyend = bp->b_bufsize;
1044 bp->b_bio1.bio_done = cluster_callback;
1045 bp->b_cmd = BUF_CMD_WRITE;
1047 vfs_busy_pages(vp, bp);
1048 bsetrunningbufspace(bp, bp->b_bufsize);
1050 vn_strategy(vp, &bp->b_bio1);
1054 return totalwritten;
1058 * Collect together all the buffers in a cluster.
1059 * Plus add one additional buffer.
1061 static struct cluster_save *
1062 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int blksize)
1064 struct cluster_save *buflist;
1069 len = (int)(vp->v_lastw - vp->v_cstart + blksize) / blksize;
1070 buflist = kmalloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1071 M_SEGMENT, M_WAITOK);
1072 buflist->bs_nchildren = 0;
1073 buflist->bs_children = (struct buf **) (buflist + 1);
1074 for (loffset = vp->v_cstart, i = 0; i < len; (loffset += blksize), i++) {
1075 (void) bread(vp, loffset, last_bp->b_bcount, &bp);
1076 buflist->bs_children[i] = bp;
1077 if (bp->b_bio2.bio_offset == NOOFFSET) {
1078 VOP_BMAP(bp->b_vp, bp->b_loffset,
1079 &bp->b_bio2.bio_offset,
1080 NULL, NULL, BUF_CMD_WRITE);
1083 buflist->bs_children[i] = bp = last_bp;
1084 if (bp->b_bio2.bio_offset == NOOFFSET) {
1085 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1086 NULL, NULL, BUF_CMD_WRITE);
1088 buflist->bs_nchildren = i + 1;
1093 cluster_append(struct bio *bio, struct buf *tbp)
1095 tbp->b_cluster_next = NULL;
1096 if (bio->bio_caller_info1.cluster_head == NULL) {
1097 bio->bio_caller_info1.cluster_head = tbp;
1098 bio->bio_caller_info2.cluster_tail = tbp;
1100 bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp;
1101 bio->bio_caller_info2.cluster_tail = tbp;
1107 cluster_setram (struct buf *bp)
1109 bp->b_flags |= B_RAM;
1110 if (bp->b_xio.xio_npages)
1111 vm_page_flag_set(bp->b_xio.xio_pages[0], PG_RAM);