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>
57 #include <vm/vm_page2.h>
59 #include <machine/limits.h>
61 #if defined(CLUSTERDEBUG)
62 #include <sys/sysctl.h>
63 static int rcluster= 0;
64 SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, "");
67 static MALLOC_DEFINE(M_SEGMENT, "cluster_save", "cluster_save buffer");
69 static struct cluster_save *
70 cluster_collectbufs (struct vnode *vp, struct buf *last_bp,
73 cluster_rbuild (struct vnode *vp, off_t filesize, off_t loffset,
74 off_t doffset, int blksize, int run,
76 static void cluster_callback (struct bio *);
77 static void cluster_setram (struct buf *);
79 static int write_behind = 1;
80 SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0,
81 "Cluster write-behind setting");
82 static int max_readahead = 2 * 1024 * 1024;
83 SYSCTL_INT(_vfs, OID_AUTO, max_readahead, CTLFLAG_RW, &max_readahead, 0,
84 "Limit in bytes for desired cluster read-ahead");
86 extern vm_page_t bogus_page;
88 extern int cluster_pbuf_freecnt;
91 * This replaces bread.
93 * filesize - read-ahead @ blksize will not cross this boundary
94 * loffset - loffset for returned *bpp
95 * blksize - blocksize for returned *bpp and read-ahead bps
96 * minreq - minimum (not a hard minimum) in bytes, typically reflects
97 * a higher level uio resid.
98 * maxreq - maximum (sequential heuristic) in bytes (highet typ ~2MB)
99 * bpp - return buffer (*bpp) for (loffset,blksize)
102 cluster_read(struct vnode *vp, off_t filesize, off_t loffset,
103 int blksize, size_t minreq, size_t maxreq, struct buf **bpp)
105 struct buf *bp, *rbp, *reqbp;
116 * Calculate the desired read-ahead in blksize'd blocks (maxra).
117 * To do this we calculate maxreq.
119 * maxreq typically starts out as a sequential heuristic. If the
120 * high level uio/resid is bigger (minreq), we pop maxreq up to
121 * minreq. This represents the case where random I/O is being
122 * performed by the userland is issuing big read()'s.
124 * Then we limit maxreq to max_readahead to ensure it is a reasonable
127 * Finally we must ensure that (loffset + maxreq) does not cross the
128 * boundary (filesize) for the current blocksize. If we allowed it
129 * to cross we could end up with buffers past the boundary with the
130 * wrong block size (HAMMER large-data areas use mixed block sizes).
131 * minreq is also absolutely limited to filesize.
135 /* minreq not used beyond this point */
137 if (maxreq > max_readahead) {
138 maxreq = max_readahead;
139 if (maxreq > 16 * 1024 * 1024)
140 maxreq = 16 * 1024 * 1024;
142 if (maxreq < blksize)
144 if (loffset + maxreq > filesize) {
145 if (loffset > filesize)
148 maxreq = filesize - loffset;
151 maxra = (int)(maxreq / blksize);
154 * Get the requested block.
156 *bpp = reqbp = bp = getblk(vp, loffset, blksize, 0, 0);
157 origoffset = loffset;
160 * Calculate the maximum cluster size for a single I/O, used
161 * by cluster_rbuild().
163 maxrbuild = vmaxiosize(vp) / blksize;
166 * if it is in the cache, then check to see if the reads have been
167 * sequential. If they have, then try some read-ahead, otherwise
168 * back-off on prospective read-aheads.
170 if (bp->b_flags & B_CACHE) {
172 * Not sequential, do not do any read-ahead
178 * No read-ahead mark, do not do any read-ahead
181 if ((bp->b_flags & B_RAM) == 0)
185 * We hit a read-ahead-mark, figure out how much read-ahead
186 * to do (maxra) and where to start (loffset).
188 * Shortcut the scan. Typically the way this works is that
189 * we've built up all the blocks inbetween except for the
190 * last in previous iterations, so if the second-to-last
191 * block is present we just skip ahead to it.
193 * This algorithm has O(1) cpu in the steady state no
194 * matter how large maxra is.
196 bp->b_flags &= ~B_RAM;
198 if (findblk(vp, loffset + (maxra - 2) * blksize, FINDBLK_TEST))
203 if (findblk(vp, loffset + i * blksize,
204 FINDBLK_TEST) == NULL) {
211 * We got everything or everything is in the cache, no
217 loffset += i * blksize;
220 __debugvar off_t firstread = bp->b_loffset;
224 * Set-up synchronous read for bp.
226 bp->b_cmd = BUF_CMD_READ;
227 bp->b_bio1.bio_done = biodone_sync;
228 bp->b_bio1.bio_flags |= BIO_SYNC;
230 KASSERT(firstread != NOOFFSET,
231 ("cluster_read: no buffer offset"));
234 * nblks is our cluster_rbuild request size, limited
235 * primarily by the device.
237 if ((nblks = maxra) > maxrbuild)
243 error = VOP_BMAP(vp, loffset, &doffset,
244 &burstbytes, NULL, BUF_CMD_READ);
246 goto single_block_read;
247 if (nblks > burstbytes / blksize)
248 nblks = burstbytes / blksize;
249 if (doffset == NOOFFSET)
250 goto single_block_read;
252 goto single_block_read;
254 bp = cluster_rbuild(vp, filesize, loffset,
255 doffset, blksize, nblks, bp);
256 loffset += bp->b_bufsize;
257 maxra -= bp->b_bufsize / blksize;
261 * If it isn't in the cache, then get a chunk from
262 * disk if sequential, otherwise just get the block.
271 * If B_CACHE was not set issue bp. bp will either be an
272 * asynchronous cluster buf or a synchronous single-buf.
273 * If it is a single buf it will be the same as reqbp.
275 * NOTE: Once an async cluster buf is issued bp becomes invalid.
278 #if defined(CLUSTERDEBUG)
280 kprintf("S(%012jx,%d,%d)\n",
281 (intmax_t)bp->b_loffset, bp->b_bcount, maxra);
283 if ((bp->b_flags & B_CLUSTER) == 0)
284 vfs_busy_pages(vp, bp);
285 bp->b_flags &= ~(B_ERROR|B_INVAL);
286 vn_strategy(vp, &bp->b_bio1);
292 * If we have been doing sequential I/O, then do some read-ahead.
293 * The code above us should have positioned us at the next likely
296 * Only mess with buffers which we can immediately lock. HAMMER
297 * will do device-readahead irrespective of what the blocks
300 while (error == 0 && maxra > 0) {
305 rbp = getblk(vp, loffset, blksize,
306 GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
309 if ((rbp->b_flags & B_CACHE)) {
315 * An error from the read-ahead bmap has nothing to do
316 * with the caller's original request.
318 tmp_error = VOP_BMAP(vp, loffset, &doffset,
319 &burstbytes, NULL, BUF_CMD_READ);
320 if (tmp_error || doffset == NOOFFSET) {
321 rbp->b_flags |= B_INVAL;
326 if ((nblks = maxra) > maxrbuild)
328 if (nblks > burstbytes / blksize)
329 nblks = burstbytes / blksize;
334 rbp->b_cmd = BUF_CMD_READ;
335 /*rbp->b_flags |= B_AGE*/;
339 rbp = cluster_rbuild(vp, filesize, loffset,
343 rbp->b_bio2.bio_offset = doffset;
346 #if defined(CLUSTERDEBUG)
349 kprintf("A+(%012jx,%d,%jd) "
350 "doff=%012jx minr=%zd ra=%d\n",
351 (intmax_t)loffset, rbp->b_bcount,
352 (intmax_t)(loffset - origoffset),
353 (intmax_t)doffset, minreq, maxra);
355 kprintf("A-(%012jx,%d,%jd) "
356 "doff=%012jx minr=%zd ra=%d\n",
357 (intmax_t)rbp->b_loffset, rbp->b_bcount,
358 (intmax_t)(loffset - origoffset),
359 (intmax_t)doffset, minreq, maxra);
363 rbp->b_flags &= ~(B_ERROR|B_INVAL);
365 if ((rbp->b_flags & B_CLUSTER) == 0)
366 vfs_busy_pages(vp, rbp);
368 loffset += rbp->b_bufsize;
369 maxra -= rbp->b_bufsize / blksize;
370 vn_strategy(vp, &rbp->b_bio1);
371 /* rbp invalid now */
375 * Wait for our original buffer to complete its I/O. reqbp will
376 * be NULL if the original buffer was B_CACHE. We are returning
377 * (*bpp) which is the same as reqbp when reqbp != NULL.
381 KKASSERT(reqbp->b_bio1.bio_flags & BIO_SYNC);
382 error = biowait(&reqbp->b_bio1, "clurd");
388 * If blocks are contiguous on disk, use this to provide clustered
389 * read ahead. We will read as many blocks as possible sequentially
390 * and then parcel them up into logical blocks in the buffer hash table.
392 * This function either returns a cluster buf or it returns fbp. fbp is
393 * already expected to be set up as a synchronous or asynchronous request.
395 * If a cluster buf is returned it will always be async.
398 cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset, off_t doffset,
399 int blksize, int run, struct buf *fbp)
401 struct buf *bp, *tbp;
404 int maxiosize = vmaxiosize(vp);
409 while (loffset + run * blksize > filesize) {
414 tbp->b_bio2.bio_offset = doffset;
415 if((tbp->b_flags & B_MALLOC) ||
416 ((tbp->b_flags & B_VMIO) == 0) || (run <= 1)) {
420 bp = trypbuf_kva(&cluster_pbuf_freecnt);
426 * We are synthesizing a buffer out of vm_page_t's, but
427 * if the block size is not page aligned then the starting
428 * address may not be either. Inherit the b_data offset
429 * from the original buffer.
431 bp->b_data = (char *)((vm_offset_t)bp->b_data |
432 ((vm_offset_t)tbp->b_data & PAGE_MASK));
433 bp->b_flags |= B_CLUSTER | B_VMIO;
434 bp->b_cmd = BUF_CMD_READ;
435 bp->b_bio1.bio_done = cluster_callback; /* default to async */
436 bp->b_bio1.bio_caller_info1.cluster_head = NULL;
437 bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
438 bp->b_loffset = loffset;
439 bp->b_bio2.bio_offset = doffset;
440 KASSERT(bp->b_loffset != NOOFFSET,
441 ("cluster_rbuild: no buffer offset"));
445 bp->b_xio.xio_npages = 0;
447 for (boffset = doffset, i = 0; i < run; ++i, boffset += blksize) {
449 if ((bp->b_xio.xio_npages * PAGE_SIZE) +
450 round_page(blksize) > maxiosize) {
455 * Shortcut some checks and try to avoid buffers that
456 * would block in the lock. The same checks have to
457 * be made again after we officially get the buffer.
459 tbp = getblk(vp, loffset + i * blksize, blksize,
460 GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
463 for (j = 0; j < tbp->b_xio.xio_npages; j++) {
464 if (tbp->b_xio.xio_pages[j]->valid)
467 if (j != tbp->b_xio.xio_npages) {
473 * Stop scanning if the buffer is fuly valid
474 * (marked B_CACHE), or locked (may be doing a
475 * background write), or if the buffer is not
476 * VMIO backed. The clustering code can only deal
477 * with VMIO-backed buffers.
479 if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
480 (tbp->b_flags & B_VMIO) == 0 ||
481 (LIST_FIRST(&tbp->b_dep) != NULL &&
489 * The buffer must be completely invalid in order to
490 * take part in the cluster. If it is partially valid
493 for (j = 0;j < tbp->b_xio.xio_npages; j++) {
494 if (tbp->b_xio.xio_pages[j]->valid)
497 if (j != tbp->b_xio.xio_npages) {
503 * Set a read-ahead mark as appropriate
505 if (i == 1 || i == (run - 1))
509 * Depress the priority of buffers not explicitly
512 /* tbp->b_flags |= B_AGE; */
515 * Set the block number if it isn't set, otherwise
516 * if it is make sure it matches the block number we
519 if (tbp->b_bio2.bio_offset == NOOFFSET) {
520 tbp->b_bio2.bio_offset = boffset;
521 } else if (tbp->b_bio2.bio_offset != boffset) {
528 * The passed-in tbp (i == 0) will already be set up for
529 * async or sync operation. All other tbp's acquire in
530 * our loop are set up for async operation.
532 tbp->b_cmd = BUF_CMD_READ;
534 cluster_append(&bp->b_bio1, tbp);
535 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
537 m = tbp->b_xio.xio_pages[j];
539 vm_object_pip_add(m->object, 1);
540 if ((bp->b_xio.xio_npages == 0) ||
541 (bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
542 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
543 bp->b_xio.xio_npages++;
545 if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
546 tbp->b_xio.xio_pages[j] = bogus_page;
549 * XXX shouldn't this be += size for both, like in
552 * Don't inherit tbp->b_bufsize as it may be larger due to
553 * a non-page-aligned size. Instead just aggregate using
556 if (tbp->b_bcount != blksize)
557 kprintf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, blksize);
558 if (tbp->b_bufsize != blksize)
559 kprintf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, blksize);
560 bp->b_bcount += blksize;
561 bp->b_bufsize += blksize;
565 * Fully valid pages in the cluster are already good and do not need
566 * to be re-read from disk. Replace the page with bogus_page
568 for (j = 0; j < bp->b_xio.xio_npages; j++) {
569 if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
571 bp->b_xio.xio_pages[j] = bogus_page;
574 if (bp->b_bufsize > bp->b_kvasize) {
575 panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)",
576 bp->b_bufsize, bp->b_kvasize);
578 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
579 (vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages);
585 * Cleanup after a clustered read or write.
586 * This is complicated by the fact that any of the buffers might have
587 * extra memory (if there were no empty buffer headers at allocbuf time)
588 * that we will need to shift around.
590 * The returned bio is &bp->b_bio1
593 cluster_callback(struct bio *bio)
595 struct buf *bp = bio->bio_buf;
600 * Must propogate errors to all the components. A short read (EOF)
601 * is a critical error.
603 if (bp->b_flags & B_ERROR) {
605 } else if (bp->b_bcount != bp->b_bufsize) {
606 panic("cluster_callback: unexpected EOF on cluster %p!", bio);
609 pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
611 * Move memory from the large cluster buffer into the component
612 * buffers and mark IO as done on these. Since the memory map
613 * is the same, no actual copying is required.
615 while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) {
616 bio->bio_caller_info1.cluster_head = tbp->b_cluster_next;
618 tbp->b_flags |= B_ERROR | B_IODEBUG;
619 tbp->b_error = error;
621 tbp->b_dirtyoff = tbp->b_dirtyend = 0;
622 tbp->b_flags &= ~(B_ERROR|B_INVAL);
623 tbp->b_flags |= B_IODEBUG;
625 * XXX the bdwrite()/bqrelse() issued during
626 * cluster building clears B_RELBUF (see bqrelse()
627 * comment). If direct I/O was specified, we have
628 * to restore it here to allow the buffer and VM
631 if (tbp->b_flags & B_DIRECT)
632 tbp->b_flags |= B_RELBUF;
634 biodone(&tbp->b_bio1);
636 relpbuf(bp, &cluster_pbuf_freecnt);
642 * Implement modified write build for cluster.
644 * write_behind = 0 write behind disabled
645 * write_behind = 1 write behind normal (default)
646 * write_behind = 2 write behind backed-off
650 cluster_wbuild_wb(struct vnode *vp, int blksize, off_t start_loffset, int len)
654 switch(write_behind) {
656 if (start_loffset < len)
658 start_loffset -= len;
661 r = cluster_wbuild(vp, blksize, start_loffset, len);
671 * Do clustered write for FFS.
674 * 1. Write is not sequential (write asynchronously)
675 * Write is sequential:
676 * 2. beginning of cluster - begin cluster
677 * 3. middle of a cluster - add to cluster
678 * 4. end of a cluster - asynchronously write cluster
681 cluster_write(struct buf *bp, off_t filesize, int blksize, int seqcount)
685 int maxclen, cursize;
689 if (vp->v_type == VREG)
690 async = vp->v_mount->mnt_flag & MNT_ASYNC;
693 loffset = bp->b_loffset;
694 KASSERT(bp->b_loffset != NOOFFSET,
695 ("cluster_write: no buffer offset"));
697 /* Initialize vnode to beginning of file. */
699 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
701 if (vp->v_clen == 0 || loffset != vp->v_lastw + blksize ||
702 bp->b_bio2.bio_offset == NOOFFSET ||
703 (bp->b_bio2.bio_offset != vp->v_lasta + blksize)) {
704 maxclen = vmaxiosize(vp);
705 if (vp->v_clen != 0) {
707 * Next block is not sequential.
709 * If we are not writing at end of file, the process
710 * seeked to another point in the file since its last
711 * write, or we have reached our maximum cluster size,
712 * then push the previous cluster. Otherwise try
713 * reallocating to make it sequential.
715 * Change to algorithm: only push previous cluster if
716 * it was sequential from the point of view of the
717 * seqcount heuristic, otherwise leave the buffer
718 * intact so we can potentially optimize the I/O
719 * later on in the buf_daemon or update daemon
722 cursize = vp->v_lastw - vp->v_cstart + blksize;
723 if (bp->b_loffset + blksize != filesize ||
724 loffset != vp->v_lastw + blksize || vp->v_clen <= cursize) {
725 if (!async && seqcount > 0) {
726 cluster_wbuild_wb(vp, blksize,
727 vp->v_cstart, cursize);
730 struct buf **bpp, **endbp;
731 struct cluster_save *buflist;
733 buflist = cluster_collectbufs(vp, bp, blksize);
734 endbp = &buflist->bs_children
735 [buflist->bs_nchildren - 1];
736 if (VOP_REALLOCBLKS(vp, buflist)) {
738 * Failed, push the previous cluster
739 * if *really* writing sequentially
740 * in the logical file (seqcount > 1),
741 * otherwise delay it in the hopes that
742 * the low level disk driver can
743 * optimize the write ordering.
745 for (bpp = buflist->bs_children;
748 kfree(buflist, M_SEGMENT);
750 cluster_wbuild_wb(vp,
751 blksize, vp->v_cstart,
756 * Succeeded, keep building cluster.
758 for (bpp = buflist->bs_children;
761 kfree(buflist, M_SEGMENT);
762 vp->v_lastw = loffset;
763 vp->v_lasta = bp->b_bio2.bio_offset;
769 * Consider beginning a cluster. If at end of file, make
770 * cluster as large as possible, otherwise find size of
773 if ((vp->v_type == VREG) &&
774 bp->b_loffset + blksize != filesize &&
775 (bp->b_bio2.bio_offset == NOOFFSET) &&
776 (VOP_BMAP(vp, loffset, &bp->b_bio2.bio_offset, &maxclen, NULL, BUF_CMD_WRITE) ||
777 bp->b_bio2.bio_offset == NOOFFSET)) {
780 vp->v_lasta = bp->b_bio2.bio_offset;
781 vp->v_cstart = loffset + blksize;
782 vp->v_lastw = loffset;
785 if (maxclen > blksize)
786 vp->v_clen = maxclen - blksize;
789 if (!async && vp->v_clen == 0) { /* I/O not contiguous */
790 vp->v_cstart = loffset + blksize;
792 } else { /* Wait for rest of cluster */
793 vp->v_cstart = loffset;
796 } else if (loffset == vp->v_cstart + vp->v_clen) {
798 * At end of cluster, write it out if seqcount tells us we
799 * are operating sequentially, otherwise let the buf or
800 * update daemon handle it.
804 cluster_wbuild_wb(vp, blksize, vp->v_cstart,
805 vp->v_clen + blksize);
807 vp->v_cstart = loffset + blksize;
808 } else if (vm_page_count_severe()) {
810 * We are low on memory, get it going NOW
815 * In the middle of a cluster, so just delay the I/O for now.
819 vp->v_lastw = loffset;
820 vp->v_lasta = bp->b_bio2.bio_offset;
825 * This is an awful lot like cluster_rbuild...wish they could be combined.
826 * The last lbn argument is the current block on which I/O is being
827 * performed. Check to see that it doesn't fall in the middle of
828 * the current block (if last_bp == NULL).
831 cluster_wbuild(struct vnode *vp, int blksize, off_t start_loffset, int bytes)
833 struct buf *bp, *tbp;
835 int totalwritten = 0;
836 int maxiosize = vmaxiosize(vp);
840 * If the buffer is not delayed-write (i.e. dirty), or it
841 * is delayed-write but either locked or inval, it cannot
842 * partake in the clustered write.
844 tbp = findblk(vp, start_loffset, FINDBLK_NBLOCK);
846 (tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI ||
847 (LIST_FIRST(&tbp->b_dep) && buf_checkwrite(tbp))) {
850 start_loffset += blksize;
855 KKASSERT(tbp->b_cmd == BUF_CMD_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_CLUSTEROK) ||
866 (tbp->b_bcount != tbp->b_bufsize) ||
867 (tbp->b_bcount != blksize) ||
868 (bytes == blksize) ||
869 ((bp = getpbuf_kva(&cluster_pbuf_freecnt)) == NULL)) {
870 totalwritten += tbp->b_bufsize;
872 start_loffset += blksize;
878 * Set up the pbuf. Track our append point with b_bcount
879 * and b_bufsize. b_bufsize is not used by the device but
880 * our caller uses it to loop clusters and we use it to
881 * detect a premature EOF on the block device.
885 bp->b_xio.xio_npages = 0;
886 bp->b_loffset = tbp->b_loffset;
887 bp->b_bio2.bio_offset = tbp->b_bio2.bio_offset;
890 * We are synthesizing a buffer out of vm_page_t's, but
891 * if the block size is not page aligned then the starting
892 * address may not be either. Inherit the b_data offset
893 * from the original buffer.
895 bp->b_data = (char *)((vm_offset_t)bp->b_data |
896 ((vm_offset_t)tbp->b_data & PAGE_MASK));
897 bp->b_flags &= ~B_ERROR;
898 bp->b_flags |= B_CLUSTER | B_BNOCLIP |
899 (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
900 bp->b_bio1.bio_caller_info1.cluster_head = NULL;
901 bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
904 * From this location in the file, scan forward to see
905 * if there are buffers with adjacent data that need to
906 * be written as well.
908 for (i = 0; i < bytes; (i += blksize), (start_loffset += blksize)) {
909 if (i != 0) { /* If not the first buffer */
910 tbp = findblk(vp, start_loffset,
913 * Buffer not found or could not be locked
920 * If it IS in core, but has different
921 * characteristics, then don't cluster
924 if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
925 B_INVAL | B_DELWRI | B_NEEDCOMMIT))
926 != (B_DELWRI | B_CLUSTEROK |
927 (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
928 (tbp->b_flags & B_LOCKED) ||
929 (LIST_FIRST(&tbp->b_dep) &&
937 * Check that the combined cluster
938 * would make sense with regard to pages
939 * and would not be too large
941 if ((tbp->b_bcount != blksize) ||
942 ((bp->b_bio2.bio_offset + i) !=
943 tbp->b_bio2.bio_offset) ||
944 ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
945 (maxiosize / 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 KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
956 } /* end of code for non-first buffers only */
959 * If the IO is via the VM then we do some
960 * special VM hackery (yuck). Since the buffer's
961 * block size may not be page-aligned it is possible
962 * for a page to be shared between two buffers. We
963 * have to get rid of the duplication when building
966 if (tbp->b_flags & B_VMIO) {
969 if (i != 0) { /* if not first buffer */
970 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
971 m = tbp->b_xio.xio_pages[j];
972 if (m->flags & PG_BUSY) {
979 for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
980 m = tbp->b_xio.xio_pages[j];
982 vm_object_pip_add(m->object, 1);
983 if ((bp->b_xio.xio_npages == 0) ||
984 (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
985 bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
986 bp->b_xio.xio_npages++;
990 bp->b_bcount += blksize;
991 bp->b_bufsize += blksize;
994 tbp->b_flags &= ~B_ERROR;
995 tbp->b_cmd = BUF_CMD_WRITE;
997 cluster_append(&bp->b_bio1, tbp);
1000 * check for latent dependencies to be handled
1002 if (LIST_FIRST(&tbp->b_dep) != NULL)
1006 pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
1007 (vm_page_t *) bp->b_xio.xio_pages, bp->b_xio.xio_npages);
1008 if (bp->b_bufsize > bp->b_kvasize) {
1010 "cluster_wbuild: b_bufsize(%d) > b_kvasize(%d)\n",
1011 bp->b_bufsize, bp->b_kvasize);
1013 totalwritten += bp->b_bufsize;
1015 bp->b_dirtyend = bp->b_bufsize;
1016 bp->b_bio1.bio_done = cluster_callback;
1017 bp->b_cmd = BUF_CMD_WRITE;
1019 vfs_busy_pages(vp, bp);
1020 bsetrunningbufspace(bp, bp->b_bufsize);
1022 vn_strategy(vp, &bp->b_bio1);
1026 return totalwritten;
1030 * Collect together all the buffers in a cluster.
1031 * Plus add one additional buffer.
1033 static struct cluster_save *
1034 cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int blksize)
1036 struct cluster_save *buflist;
1041 len = (int)(vp->v_lastw - vp->v_cstart + blksize) / blksize;
1042 buflist = kmalloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
1043 M_SEGMENT, M_WAITOK);
1044 buflist->bs_nchildren = 0;
1045 buflist->bs_children = (struct buf **) (buflist + 1);
1046 for (loffset = vp->v_cstart, i = 0; i < len; (loffset += blksize), i++) {
1047 (void) bread(vp, loffset, last_bp->b_bcount, &bp);
1048 buflist->bs_children[i] = bp;
1049 if (bp->b_bio2.bio_offset == NOOFFSET) {
1050 VOP_BMAP(bp->b_vp, bp->b_loffset,
1051 &bp->b_bio2.bio_offset,
1052 NULL, NULL, BUF_CMD_WRITE);
1055 buflist->bs_children[i] = bp = last_bp;
1056 if (bp->b_bio2.bio_offset == NOOFFSET) {
1057 VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
1058 NULL, NULL, BUF_CMD_WRITE);
1060 buflist->bs_nchildren = i + 1;
1065 cluster_append(struct bio *bio, struct buf *tbp)
1067 tbp->b_cluster_next = NULL;
1068 if (bio->bio_caller_info1.cluster_head == NULL) {
1069 bio->bio_caller_info1.cluster_head = tbp;
1070 bio->bio_caller_info2.cluster_tail = tbp;
1072 bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp;
1073 bio->bio_caller_info2.cluster_tail = tbp;
1079 cluster_setram (struct buf *bp)
1081 bp->b_flags |= B_RAM;
1082 if (bp->b_xio.xio_npages)
1083 vm_page_flag_set(bp->b_xio.xio_pages[0], PG_RAM);