linprocfs - Introduce /proc/mounts

[dragonfly.git] / sys / kern / vfs_cluster.c

/*-
 * Copyright (c) 1993
 *      The Regents of the University of California.  All rights reserved.
 * Modifications/enhancements:
 *      Copyright (c) 1995 John S. Dyson.  All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 * 1. Redistributions of source code must retain the above copyright
 *    notice, this list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in the
 *    documentation and/or other materials provided with the distribution.
 * 3. All advertising materials mentioning features or use of this software
 *    must display the following acknowledgement:
 *      This product includes software developed by the University of
 *      California, Berkeley and its contributors.
 * 4. Neither the name of the University nor the names of its contributors
 *    may be used to endorse or promote products derived from this software
 *    without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
 * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
 * SUCH DAMAGE.
 *
 *      @(#)vfs_cluster.c       8.7 (Berkeley) 2/13/94
 * $FreeBSD: src/sys/kern/vfs_cluster.c,v 1.92.2.9 2001/11/18 07:10:59 dillon Exp $
 * $DragonFly: src/sys/kern/vfs_cluster.c,v 1.40 2008/07/14 03:09:00 dillon Exp $
 */

#include "opt_debug_cluster.h"

#include <sys/param.h>
#include <sys/systm.h>
#include <sys/kernel.h>
#include <sys/proc.h>
#include <sys/buf.h>
#include <sys/vnode.h>
#include <sys/malloc.h>
#include <sys/mount.h>
#include <sys/resourcevar.h>
#include <sys/vmmeter.h>
#include <vm/vm.h>
#include <vm/vm_object.h>
#include <vm/vm_page.h>
#include <sys/sysctl.h>
#include <sys/buf2.h>
#include <vm/vm_page2.h>

#include <machine/limits.h>

#if defined(CLUSTERDEBUG)
#include <sys/sysctl.h>
static int      rcluster= 0;
SYSCTL_INT(_debug, OID_AUTO, rcluster, CTLFLAG_RW, &rcluster, 0, "");
#endif

static MALLOC_DEFINE(M_SEGMENT, "cluster_save", "cluster_save buffer");

static struct cluster_save *
        cluster_collectbufs (struct vnode *vp, struct buf *last_bp,
                            int blksize);
static struct buf *
        cluster_rbuild (struct vnode *vp, off_t filesize, off_t loffset,
                            off_t doffset, int blksize, int run, 
                            struct buf *fbp);
static void cluster_callback (struct bio *);
static void cluster_setram (struct buf *);

static int write_behind = 1;
SYSCTL_INT(_vfs, OID_AUTO, write_behind, CTLFLAG_RW, &write_behind, 0, "");

extern vm_page_t        bogus_page;

extern int cluster_pbuf_freecnt;

/*
 * Maximum number of blocks for read-ahead.
 */
#define MAXRA 32

/*
 * This replaces bread.
 */
int
cluster_read(struct vnode *vp, off_t filesize, off_t loffset, 
             int blksize, size_t resid, int seqcount, struct buf **bpp)
{
        struct buf *bp, *rbp, *reqbp;
        off_t origoffset;
        off_t doffset;
        int error;
        int i;
        int maxra, racluster;
        int totread;

        error = 0;
        totread = (resid > INT_MAX) ? INT_MAX : (int)resid;

        /*
         * racluster - calculate maximum cluster IO size (limited by
         *             backing block device).
         *
         * Try to limit the amount of read-ahead by a few ad-hoc parameters.
         * This needs work!!!
         *
         * NOTE!  The BMAP operations may involve synchronous I/O so we
         *        really want several cluster IOs in progress to absorb
         *        the time lag.
         */
        racluster = vmaxiosize(vp) / blksize;
        maxra = 2 * racluster + (totread / blksize);
        if (maxra > MAXRA)
                maxra = MAXRA;
        if (maxra > nbuf / 8)
                maxra = nbuf / 8;

        /*
         * Get the requested block.
         */
        *bpp = reqbp = bp = getblk(vp, loffset, blksize, 0, 0);
        origoffset = loffset;

        /*
         * if it is in the cache, then check to see if the reads have been
         * sequential.  If they have, then try some read-ahead, otherwise
         * back-off on prospective read-aheads.
         */
        if (bp->b_flags & B_CACHE) {
                /*
                 * Not sequential, do not do any read-ahead
                 */
                if (seqcount == 0 || maxra == 0)
                        return 0;

                /*
                 * No read-ahead mark, do not do any read-ahead
                 * yet.
                 */
                if ((bp->b_flags & B_RAM) == 0)
                        return 0;

                /*
                 * We hit a read-ahead-mark, figure out how much read-ahead
                 * to do (maxra) and where to start (loffset).
                 *
                 * Shortcut the scan.  Typically the way this works is that
                 * we've built up all the blocks inbetween except for the
                 * last in previous iterations, so if the second-to-last
                 * block is present we just skip ahead to it.
                 *
                 * This algorithm has O(1) cpu in the steady state no
                 * matter how large maxra is.
                 */
                bp->b_flags &= ~B_RAM;

                if (findblk(vp, loffset + (maxra - 2) * blksize, FINDBLK_TEST))
                        i = maxra - 1;
                else
                        i = 1;
                while (i < maxra) {
                        if (findblk(vp, loffset + i * blksize,
                                    FINDBLK_TEST) == NULL) {
                                break;
                        }
                        ++i;
                }
                if (i >= maxra)
                        return 0;
                maxra -= i;
                loffset += i * blksize;
                reqbp = bp = NULL;
        } else {
                off_t firstread = bp->b_loffset;
                int nblks;

                /*
                 * Set-up synchronous read for bp.
                 */
                bp->b_cmd = BUF_CMD_READ;
                bp->b_bio1.bio_done = biodone_sync;
                bp->b_bio1.bio_flags |= BIO_SYNC;

                KASSERT(firstread != NOOFFSET, 
                        ("cluster_read: no buffer offset"));
                if (firstread + totread > filesize)
                        totread = (int)(filesize - firstread);
                nblks = totread / blksize;
                if (nblks) {
                        int burstbytes;

                        if (nblks > racluster)
                                nblks = racluster;

                        error = VOP_BMAP(vp, loffset, &doffset,
                                         &burstbytes, NULL, BUF_CMD_READ);
                        if (error)
                                goto single_block_read;
                        if (doffset == NOOFFSET)
                                goto single_block_read;
                        if (burstbytes < blksize * 2)
                                goto single_block_read;
                        if (nblks > burstbytes / blksize)
                                nblks = burstbytes / blksize;

                        bp = cluster_rbuild(vp, filesize, loffset,
                                            doffset, blksize, nblks, bp);
                        loffset += bp->b_bufsize;
                        maxra -= (bp->b_bufsize - blksize) / blksize;
                } else {
single_block_read:
                        /*
                         * if it isn't in the cache, then get a chunk from
                         * disk if sequential, otherwise just get the block.
                         */
                        cluster_setram(bp);
                        loffset += blksize;
                }
        }

        /*
         * If B_CACHE was not set issue bp.  bp will either be an
         * asynchronous cluster buf or a synchronous single-buf.
         * If it is a single buf it will be the same as reqbp.
         *
         * NOTE: Once an async cluster buf is issued bp becomes invalid.
         */
        if (bp) {
#if defined(CLUSTERDEBUG)
                if (rcluster)
                        kprintf("S(%lld,%d,%d) ",
                            bp->b_loffset, bp->b_bcount, seqcount);
#endif
                if ((bp->b_flags & B_CLUSTER) == 0)
                        vfs_busy_pages(vp, bp);
                bp->b_flags &= ~(B_ERROR|B_INVAL);
                vn_strategy(vp, &bp->b_bio1);
                error = 0;
                /* bp invalid now */
        }

        /*
         * If we have been doing sequential I/O, then do some read-ahead.
         * The code above us should have positioned us at the next likely
         * offset.
         *
         * Only mess with buffers which we can immediately lock.  HAMMER
         * will do device-readahead irrespective of what the blocks
         * represent.
         */
        while (!error && seqcount && maxra > 0 &&
               loffset + blksize <= filesize) {
                int nblksread;
                int ntoread;
                int burstbytes;
                int tmp_error;

                rbp = getblk(vp, loffset, blksize,
                             GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
                if (rbp == NULL)
                        goto no_read_ahead;
                if ((rbp->b_flags & B_CACHE)) {
                        bqrelse(rbp);
                        goto no_read_ahead;
                }

                /*
                 * An error from the read-ahead bmap has nothing to do
                 * with the caller's original request.
                 */
                tmp_error = VOP_BMAP(vp, loffset, &doffset,
                                     &burstbytes, NULL, BUF_CMD_READ);
                if (tmp_error || doffset == NOOFFSET) {
                        rbp->b_flags |= B_INVAL;
                        brelse(rbp);
                        rbp = NULL;
                        goto no_read_ahead;
                }
                ntoread = burstbytes / blksize;
                nblksread = (totread + blksize - 1) / blksize;
                if (seqcount < nblksread)
                        seqcount = nblksread;
                if (ntoread > seqcount)
                        ntoread = seqcount;

                /*
                 * rbp: async read
                 */
                rbp->b_cmd = BUF_CMD_READ;
                /*rbp->b_flags |= B_AGE*/;
                cluster_setram(rbp);

                if (burstbytes) {
                        rbp = cluster_rbuild(vp, filesize, loffset,
                                             doffset, blksize, 
                                             ntoread, rbp);
                } else {
                        rbp->b_bio2.bio_offset = doffset;
                }
#if defined(CLUSTERDEBUG)
                if (rcluster) {
                        if (bp)
                                kprintf("A+(%lld,%d,%lld,%d) ",
                                    rbp->b_loffset, rbp->b_bcount,
                                    rbp->b_loffset - origoffset,
                                    seqcount);
                        else
                                kprintf("A(%lld,%d,%lld,%d) ",
                                    rbp->b_loffset, rbp->b_bcount,
                                    rbp->b_loffset - origoffset,
                                    seqcount);
                }
#endif
                rbp->b_flags &= ~(B_ERROR|B_INVAL);

                if ((rbp->b_flags & B_CLUSTER) == 0)
                        vfs_busy_pages(vp, rbp);
                BUF_KERNPROC(rbp);
                loffset += rbp->b_bufsize;
                maxra -= rbp->b_bufsize / blksize;
                vn_strategy(vp, &rbp->b_bio1);
                /* rbp invalid now */
        }

        /*
         * Wait for our original buffer to complete its I/O.  reqbp will
         * be NULL if the original buffer was B_CACHE.  We are returning
         * (*bpp) which is the same as reqbp when reqbp != NULL.
         */
no_read_ahead:
        if (reqbp) {
                KKASSERT(reqbp->b_bio1.bio_flags & BIO_SYNC);
                error = biowait(&reqbp->b_bio1, "clurd");
        }
        return (error);
}

/*
 * If blocks are contiguous on disk, use this to provide clustered
 * read ahead.  We will read as many blocks as possible sequentially
 * and then parcel them up into logical blocks in the buffer hash table.
 *
 * This function either returns a cluster buf or it returns fbp.  fbp is
 * already expected to be set up as a synchronous or asynchronous request.
 *
 * If a cluster buf is returned it will always be async.
 */
static struct buf *
cluster_rbuild(struct vnode *vp, off_t filesize, off_t loffset, off_t doffset,
               int blksize, int run, struct buf *fbp)
{
        struct buf *bp, *tbp;
        off_t boffset;
        int i, j;
        int maxiosize = vmaxiosize(vp);

        /*
         * avoid a division
         */
        while (loffset + run * blksize > filesize) {
                --run;
        }

        tbp = fbp;
        tbp->b_bio2.bio_offset = doffset;
        if((tbp->b_flags & B_MALLOC) ||
            ((tbp->b_flags & B_VMIO) == 0) || (run <= 1)) {
                return tbp;
        }

        bp = trypbuf(&cluster_pbuf_freecnt);
        if (bp == NULL) {
                return tbp;
        }

        /*
         * We are synthesizing a buffer out of vm_page_t's, but
         * if the block size is not page aligned then the starting
         * address may not be either.  Inherit the b_data offset
         * from the original buffer.
         */
        bp->b_data = (char *)((vm_offset_t)bp->b_data |
            ((vm_offset_t)tbp->b_data & PAGE_MASK));
        bp->b_flags |= B_CLUSTER | B_VMIO;
        bp->b_cmd = BUF_CMD_READ;
        bp->b_bio1.bio_done = cluster_callback;         /* default to async */
        bp->b_bio1.bio_caller_info1.cluster_head = NULL;
        bp->b_bio1.bio_caller_info2.cluster_tail = NULL;
        bp->b_loffset = loffset;
        bp->b_bio2.bio_offset = doffset;
        KASSERT(bp->b_loffset != NOOFFSET,
                ("cluster_rbuild: no buffer offset"));

        bp->b_bcount = 0;
        bp->b_bufsize = 0;
        bp->b_xio.xio_npages = 0;

        for (boffset = doffset, i = 0; i < run; ++i, boffset += blksize) {
                if (i) {
                        if ((bp->b_xio.xio_npages * PAGE_SIZE) +
                            round_page(blksize) > maxiosize) {
                                break;
                        }

                        /*
                         * Shortcut some checks and try to avoid buffers that
                         * would block in the lock.  The same checks have to
                         * be made again after we officially get the buffer.
                         */
                        tbp = getblk(vp, loffset + i * blksize, blksize,
                                     GETBLK_SZMATCH|GETBLK_NOWAIT, 0);
                        if (tbp == NULL)
                                break;
                        for (j = 0; j < tbp->b_xio.xio_npages; j++) {
                                if (tbp->b_xio.xio_pages[j]->valid)
                                        break;
                        }
                        if (j != tbp->b_xio.xio_npages) {
                                bqrelse(tbp);
                                break;
                        }

                        /*
                         * Stop scanning if the buffer is fuly valid 
                         * (marked B_CACHE), or locked (may be doing a
                         * background write), or if the buffer is not
                         * VMIO backed.  The clustering code can only deal
                         * with VMIO-backed buffers.
                         */
                        if ((tbp->b_flags & (B_CACHE|B_LOCKED)) ||
                            (tbp->b_flags & B_VMIO) == 0 ||
                            (LIST_FIRST(&tbp->b_dep) != NULL &&
                             buf_checkread(tbp))
                        ) {
                                bqrelse(tbp);
                                break;
                        }

                        /*
                         * The buffer must be completely invalid in order to
                         * take part in the cluster.  If it is partially valid
                         * then we stop.
                         */
                        for (j = 0;j < tbp->b_xio.xio_npages; j++) {
                                if (tbp->b_xio.xio_pages[j]->valid)
                                        break;
                        }
                        if (j != tbp->b_xio.xio_npages) {
                                bqrelse(tbp);
                                break;
                        }

                        /*
                         * Set a read-ahead mark as appropriate
                         */
                        if (i == 1 || i == (run - 1))
                                cluster_setram(tbp);

                        /*
                         * Depress the priority of buffers not explicitly
                         * requested.
                         */
                        /* tbp->b_flags |= B_AGE; */

                        /*
                         * Set the block number if it isn't set, otherwise
                         * if it is make sure it matches the block number we
                         * expect.
                         */
                        if (tbp->b_bio2.bio_offset == NOOFFSET) {
                                tbp->b_bio2.bio_offset = boffset;
                        } else if (tbp->b_bio2.bio_offset != boffset) {
                                brelse(tbp);
                                break;
                        }
                }

                /*
                 * The passed-in tbp (i == 0) will already be set up for
                 * async or sync operation.  All other tbp's acquire in
                 * our loop are set up for async operation.
                 */
                tbp->b_cmd = BUF_CMD_READ;
                BUF_KERNPROC(tbp);
                cluster_append(&bp->b_bio1, tbp);
                for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
                        vm_page_t m;
                        m = tbp->b_xio.xio_pages[j];
                        vm_page_io_start(m);
                        vm_object_pip_add(m->object, 1);
                        if ((bp->b_xio.xio_npages == 0) ||
                                (bp->b_xio.xio_pages[bp->b_xio.xio_npages-1] != m)) {
                                bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
                                bp->b_xio.xio_npages++;
                        }
                        if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
                                tbp->b_xio.xio_pages[j] = bogus_page;
                }
                /*
                 * XXX shouldn't this be += size for both, like in 
                 * cluster_wbuild()?
                 *
                 * Don't inherit tbp->b_bufsize as it may be larger due to
                 * a non-page-aligned size.  Instead just aggregate using
                 * 'size'.
                 */
                if (tbp->b_bcount != blksize)
                    kprintf("warning: tbp->b_bcount wrong %d vs %d\n", tbp->b_bcount, blksize);
                if (tbp->b_bufsize != blksize)
                    kprintf("warning: tbp->b_bufsize wrong %d vs %d\n", tbp->b_bufsize, blksize);
                bp->b_bcount += blksize;
                bp->b_bufsize += blksize;
        }

        /*
         * Fully valid pages in the cluster are already good and do not need
         * to be re-read from disk.  Replace the page with bogus_page
         */
        for (j = 0; j < bp->b_xio.xio_npages; j++) {
                if ((bp->b_xio.xio_pages[j]->valid & VM_PAGE_BITS_ALL) ==
                    VM_PAGE_BITS_ALL) {
                        bp->b_xio.xio_pages[j] = bogus_page;
                }
        }
        if (bp->b_bufsize > bp->b_kvasize) {
                panic("cluster_rbuild: b_bufsize(%d) > b_kvasize(%d)",
                    bp->b_bufsize, bp->b_kvasize);
        }
        pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
                (vm_page_t *)bp->b_xio.xio_pages, bp->b_xio.xio_npages);
        BUF_KERNPROC(bp);
        return (bp);
}

/*
 * Cleanup after a clustered read or write.
 * This is complicated by the fact that any of the buffers might have
 * extra memory (if there were no empty buffer headers at allocbuf time)
 * that we will need to shift around.
 *
 * The returned bio is &bp->b_bio1
 */
void
cluster_callback(struct bio *bio)
{
        struct buf *bp = bio->bio_buf;
        struct buf *tbp;
        int error = 0;

        /*
         * Must propogate errors to all the components.  A short read (EOF)
         * is a critical error.
         */
        if (bp->b_flags & B_ERROR) {
                error = bp->b_error;
        } else if (bp->b_bcount != bp->b_bufsize) {
                panic("cluster_callback: unexpected EOF on cluster %p!", bio);
        }

        pmap_qremove(trunc_page((vm_offset_t) bp->b_data), bp->b_xio.xio_npages);
        /*
         * Move memory from the large cluster buffer into the component
         * buffers and mark IO as done on these.  Since the memory map
         * is the same, no actual copying is required.
         */
        while ((tbp = bio->bio_caller_info1.cluster_head) != NULL) {
                bio->bio_caller_info1.cluster_head = tbp->b_cluster_next;
                if (error) {
                        tbp->b_flags |= B_ERROR;
                        tbp->b_error = error;
                } else {
                        tbp->b_dirtyoff = tbp->b_dirtyend = 0;
                        tbp->b_flags &= ~(B_ERROR|B_INVAL);
                        /*
                         * XXX the bdwrite()/bqrelse() issued during
                         * cluster building clears B_RELBUF (see bqrelse()
                         * comment).  If direct I/O was specified, we have
                         * to restore it here to allow the buffer and VM
                         * to be freed.
                         */
                        if (tbp->b_flags & B_DIRECT)
                                tbp->b_flags |= B_RELBUF;
                }
                biodone(&tbp->b_bio1);
        }
        relpbuf(bp, &cluster_pbuf_freecnt);
}

/*
 *      cluster_wbuild_wb:
 *
 *      Implement modified write build for cluster.
 *
 *              write_behind = 0        write behind disabled
 *              write_behind = 1        write behind normal (default)
 *              write_behind = 2        write behind backed-off
 */

static __inline int
cluster_wbuild_wb(struct vnode *vp, int blksize, off_t start_loffset, int len)
{
        int r = 0;

        switch(write_behind) {
        case 2:
                if (start_loffset < len)
                        break;
                start_loffset -= len;
                /* fall through */
        case 1:
                r = cluster_wbuild(vp, blksize, start_loffset, len);
                /* fall through */
        default:
                /* fall through */
                break;
        }
        return(r);
}

/*
 * Do clustered write for FFS.
 *
 * Three cases:
 *      1. Write is not sequential (write asynchronously)
 *      Write is sequential:
 *      2.      beginning of cluster - begin cluster
 *      3.      middle of a cluster - add to cluster
 *      4.      end of a cluster - asynchronously write cluster
 */
void
cluster_write(struct buf *bp, off_t filesize, int blksize, int seqcount)
{
        struct vnode *vp;
        off_t loffset;
        int maxclen, cursize;
        int async;

        vp = bp->b_vp;
        if (vp->v_type == VREG)
                async = vp->v_mount->mnt_flag & MNT_ASYNC;
        else
                async = 0;
        loffset = bp->b_loffset;
        KASSERT(bp->b_loffset != NOOFFSET, 
                ("cluster_write: no buffer offset"));

        /* Initialize vnode to beginning of file. */
        if (loffset == 0)
                vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;

        if (vp->v_clen == 0 || loffset != vp->v_lastw + blksize ||
            bp->b_bio2.bio_offset == NOOFFSET ||
            (bp->b_bio2.bio_offset != vp->v_lasta + blksize)) {
                maxclen = vmaxiosize(vp);
                if (vp->v_clen != 0) {
                        /*
                         * Next block is not sequential.
                         *
                         * If we are not writing at end of file, the process
                         * seeked to another point in the file since its last
                         * write, or we have reached our maximum cluster size,
                         * then push the previous cluster. Otherwise try
                         * reallocating to make it sequential.
                         *
                         * Change to algorithm: only push previous cluster if
                         * it was sequential from the point of view of the
                         * seqcount heuristic, otherwise leave the buffer 
                         * intact so we can potentially optimize the I/O
                         * later on in the buf_daemon or update daemon
                         * flush.
                         */
                        cursize = vp->v_lastw - vp->v_cstart + blksize;
                        if (bp->b_loffset + blksize != filesize ||
                            loffset != vp->v_lastw + blksize || vp->v_clen <= cursize) {
                                if (!async && seqcount > 0) {
                                        cluster_wbuild_wb(vp, blksize,
                                                vp->v_cstart, cursize);
                                }
                        } else {
                                struct buf **bpp, **endbp;
                                struct cluster_save *buflist;

                                buflist = cluster_collectbufs(vp, bp, blksize);
                                endbp = &buflist->bs_children
                                    [buflist->bs_nchildren - 1];
                                if (VOP_REALLOCBLKS(vp, buflist)) {
                                        /*
                                         * Failed, push the previous cluster
                                         * if *really* writing sequentially
                                         * in the logical file (seqcount > 1),
                                         * otherwise delay it in the hopes that
                                         * the low level disk driver can
                                         * optimize the write ordering.
                                         */
                                        for (bpp = buflist->bs_children;
                                             bpp < endbp; bpp++)
                                                brelse(*bpp);
                                        kfree(buflist, M_SEGMENT);
                                        if (seqcount > 1) {
                                                cluster_wbuild_wb(vp, 
                                                    blksize, vp->v_cstart, 
                                                    cursize);
                                        }
                                } else {
                                        /*
                                         * Succeeded, keep building cluster.
                                         */
                                        for (bpp = buflist->bs_children;
                                             bpp <= endbp; bpp++)
                                                bdwrite(*bpp);
                                        kfree(buflist, M_SEGMENT);
                                        vp->v_lastw = loffset;
                                        vp->v_lasta = bp->b_bio2.bio_offset;
                                        return;
                                }
                        }
                }
                /*
                 * Consider beginning a cluster. If at end of file, make
                 * cluster as large as possible, otherwise find size of
                 * existing cluster.
                 */
                if ((vp->v_type == VREG) &&
                    bp->b_loffset + blksize != filesize &&
                    (bp->b_bio2.bio_offset == NOOFFSET) &&
                    (VOP_BMAP(vp, loffset, &bp->b_bio2.bio_offset, &maxclen, NULL, BUF_CMD_WRITE) ||
                     bp->b_bio2.bio_offset == NOOFFSET)) {
                        bawrite(bp);
                        vp->v_clen = 0;
                        vp->v_lasta = bp->b_bio2.bio_offset;
                        vp->v_cstart = loffset + blksize;
                        vp->v_lastw = loffset;
                        return;
                }
                if (maxclen > blksize)
                        vp->v_clen = maxclen - blksize;
                else
                        vp->v_clen = 0;
                if (!async && vp->v_clen == 0) { /* I/O not contiguous */
                        vp->v_cstart = loffset + blksize;
                        bawrite(bp);
                } else {        /* Wait for rest of cluster */
                        vp->v_cstart = loffset;
                        bdwrite(bp);
                }
        } else if (loffset == vp->v_cstart + vp->v_clen) {
                /*
                 * At end of cluster, write it out if seqcount tells us we
                 * are operating sequentially, otherwise let the buf or
                 * update daemon handle it.
                 */
                bdwrite(bp);
                if (seqcount > 1)
                        cluster_wbuild_wb(vp, blksize, vp->v_cstart,
                                          vp->v_clen + blksize);
                vp->v_clen = 0;
                vp->v_cstart = loffset + blksize;
        } else if (vm_page_count_severe()) {
                /*
                 * We are low on memory, get it going NOW
                 */
                bawrite(bp);
        } else {
                /*
                 * In the middle of a cluster, so just delay the I/O for now.
                 */
                bdwrite(bp);
        }
        vp->v_lastw = loffset;
        vp->v_lasta = bp->b_bio2.bio_offset;
}


/*
 * This is an awful lot like cluster_rbuild...wish they could be combined.
 * The last lbn argument is the current block on which I/O is being
 * performed.  Check to see that it doesn't fall in the middle of
 * the current block (if last_bp == NULL).
 */
int
cluster_wbuild(struct vnode *vp, int blksize, off_t start_loffset, int bytes)
{
        struct buf *bp, *tbp;
        int i, j;
        int totalwritten = 0;
        int maxiosize = vmaxiosize(vp);

        while (bytes > 0) {
                /*
                 * If the buffer is not delayed-write (i.e. dirty), or it 
                 * is delayed-write but either locked or inval, it cannot 
                 * partake in the clustered write.
                 */
                tbp = findblk(vp, start_loffset, FINDBLK_NBLOCK);
                if (tbp == NULL ||
                    (tbp->b_flags & (B_LOCKED | B_INVAL | B_DELWRI)) != B_DELWRI ||
                    (LIST_FIRST(&tbp->b_dep) && buf_checkwrite(tbp))) {
                        if (tbp)
                                BUF_UNLOCK(tbp);
                        start_loffset += blksize;
                        bytes -= blksize;
                        continue;
                }
                bremfree(tbp);
                KKASSERT(tbp->b_cmd == BUF_CMD_DONE);

                /*
                 * Extra memory in the buffer, punt on this buffer.
                 * XXX we could handle this in most cases, but we would
                 * have to push the extra memory down to after our max
                 * possible cluster size and then potentially pull it back
                 * up if the cluster was terminated prematurely--too much
                 * hassle.
                 */
                if (((tbp->b_flags & (B_CLUSTEROK|B_MALLOC)) != B_CLUSTEROK) ||
                    (tbp->b_bcount != tbp->b_bufsize) ||
                    (tbp->b_bcount != blksize) ||
                    (bytes == blksize) ||
                    ((bp = getpbuf(&cluster_pbuf_freecnt)) == NULL)) {
                        totalwritten += tbp->b_bufsize;
                        bawrite(tbp);
                        start_loffset += blksize;
                        bytes -= blksize;
                        continue;
                }

                /*
                 * Set up the pbuf.  Track our append point with b_bcount
                 * and b_bufsize.  b_bufsize is not used by the device but
                 * our caller uses it to loop clusters and we use it to
                 * detect a premature EOF on the block device.
                 */
                bp->b_bcount = 0;
                bp->b_bufsize = 0;
                bp->b_xio.xio_npages = 0;
                bp->b_loffset = tbp->b_loffset;
                bp->b_bio2.bio_offset = tbp->b_bio2.bio_offset;

                /*
                 * We are synthesizing a buffer out of vm_page_t's, but
                 * if the block size is not page aligned then the starting
                 * address may not be either.  Inherit the b_data offset
                 * from the original buffer.
                 */
                bp->b_data = (char *)((vm_offset_t)bp->b_data |
                    ((vm_offset_t)tbp->b_data & PAGE_MASK));
                bp->b_flags &= ~B_ERROR;
                bp->b_flags |= B_CLUSTER | B_BNOCLIP |
                        (tbp->b_flags & (B_VMIO | B_NEEDCOMMIT));
                bp->b_bio1.bio_caller_info1.cluster_head = NULL;
                bp->b_bio1.bio_caller_info2.cluster_tail = NULL;

                /*
                 * From this location in the file, scan forward to see
                 * if there are buffers with adjacent data that need to
                 * be written as well.
                 */
                for (i = 0; i < bytes; (i += blksize), (start_loffset += blksize)) {
                        if (i != 0) { /* If not the first buffer */
                                tbp = findblk(vp, start_loffset,
                                              FINDBLK_NBLOCK);
                                /*
                                 * Buffer not found or could not be locked
                                 * non-blocking.
                                 */
                                if (tbp == NULL)
                                        break;

                                /*
                                 * If it IS in core, but has different
                                 * characteristics, then don't cluster
                                 * with it.
                                 */
                                if ((tbp->b_flags & (B_VMIO | B_CLUSTEROK |
                                     B_INVAL | B_DELWRI | B_NEEDCOMMIT))
                                    != (B_DELWRI | B_CLUSTEROK |
                                     (bp->b_flags & (B_VMIO | B_NEEDCOMMIT))) ||
                                    (tbp->b_flags & B_LOCKED) ||
                                    (LIST_FIRST(&tbp->b_dep) &&
                                     buf_checkwrite(tbp))
                                ) {
                                        BUF_UNLOCK(tbp);
                                        break;
                                }

                                /*
                                 * Check that the combined cluster
                                 * would make sense with regard to pages
                                 * and would not be too large
                                 */
                                if ((tbp->b_bcount != blksize) ||
                                  ((bp->b_bio2.bio_offset + i) !=
                                    tbp->b_bio2.bio_offset) ||
                                  ((tbp->b_xio.xio_npages + bp->b_xio.xio_npages) >
                                    (maxiosize / PAGE_SIZE))) {
                                        BUF_UNLOCK(tbp);
                                        break;
                                }
                                /*
                                 * Ok, it's passed all the tests,
                                 * so remove it from the free list
                                 * and mark it busy. We will use it.
                                 */
                                bremfree(tbp);
                                KKASSERT(tbp->b_cmd == BUF_CMD_DONE);
                        } /* end of code for non-first buffers only */

                        /*
                         * If the IO is via the VM then we do some
                         * special VM hackery (yuck).  Since the buffer's
                         * block size may not be page-aligned it is possible
                         * for a page to be shared between two buffers.  We
                         * have to get rid of the duplication when building
                         * the cluster.
                         */
                        if (tbp->b_flags & B_VMIO) {
                                vm_page_t m;

                                if (i != 0) { /* if not first buffer */
                                        for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
                                                m = tbp->b_xio.xio_pages[j];
                                                if (m->flags & PG_BUSY) {
                                                        bqrelse(tbp);
                                                        goto finishcluster;
                                                }
                                        }
                                }
                                        
                                for (j = 0; j < tbp->b_xio.xio_npages; ++j) {
                                        m = tbp->b_xio.xio_pages[j];
                                        vm_page_io_start(m);
                                        vm_object_pip_add(m->object, 1);
                                        if ((bp->b_xio.xio_npages == 0) ||
                                          (bp->b_xio.xio_pages[bp->b_xio.xio_npages - 1] != m)) {
                                                bp->b_xio.xio_pages[bp->b_xio.xio_npages] = m;
                                                bp->b_xio.xio_npages++;
                                        }
                                }
                        }
                        bp->b_bcount += blksize;
                        bp->b_bufsize += blksize;

                        bundirty(tbp);
                        tbp->b_flags &= ~B_ERROR;
                        tbp->b_cmd = BUF_CMD_WRITE;
                        BUF_KERNPROC(tbp);
                        cluster_append(&bp->b_bio1, tbp);

                        /*
                         * check for latent dependencies to be handled 
                         */
                        if (LIST_FIRST(&tbp->b_dep) != NULL)
                                buf_start(tbp);
                }
        finishcluster:
                pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
                        (vm_page_t *) bp->b_xio.xio_pages, bp->b_xio.xio_npages);
                if (bp->b_bufsize > bp->b_kvasize) {
                        panic(
                            "cluster_wbuild: b_bufsize(%d) > b_kvasize(%d)\n",
                            bp->b_bufsize, bp->b_kvasize);
                }
                totalwritten += bp->b_bufsize;
                bp->b_dirtyoff = 0;
                bp->b_dirtyend = bp->b_bufsize;
                bp->b_bio1.bio_done = cluster_callback;
                bp->b_cmd = BUF_CMD_WRITE;

                vfs_busy_pages(vp, bp);
                bp->b_runningbufspace = bp->b_bufsize;
                if (bp->b_runningbufspace) {
                        runningbufspace += bp->b_runningbufspace;
                        ++runningbufcount;
                }
                BUF_KERNPROC(bp);
                vn_strategy(vp, &bp->b_bio1);

                bytes -= i;
        }
        return totalwritten;
}

/*
 * Collect together all the buffers in a cluster.
 * Plus add one additional buffer.
 */
static struct cluster_save *
cluster_collectbufs(struct vnode *vp, struct buf *last_bp, int blksize)
{
        struct cluster_save *buflist;
        struct buf *bp;
        off_t loffset;
        int i, len;

        len = (int)(vp->v_lastw - vp->v_cstart + blksize) / blksize;
        buflist = kmalloc(sizeof(struct buf *) * (len + 1) + sizeof(*buflist),
                         M_SEGMENT, M_WAITOK);
        buflist->bs_nchildren = 0;
        buflist->bs_children = (struct buf **) (buflist + 1);
        for (loffset = vp->v_cstart, i = 0; i < len; (loffset += blksize), i++) {
                (void) bread(vp, loffset, last_bp->b_bcount, &bp);
                buflist->bs_children[i] = bp;
                if (bp->b_bio2.bio_offset == NOOFFSET) {
                        VOP_BMAP(bp->b_vp, bp->b_loffset,
                                 &bp->b_bio2.bio_offset,
                                 NULL, NULL, BUF_CMD_WRITE);
                }
        }
        buflist->bs_children[i] = bp = last_bp;
        if (bp->b_bio2.bio_offset == NOOFFSET) {
                VOP_BMAP(bp->b_vp, bp->b_loffset, &bp->b_bio2.bio_offset,
                         NULL, NULL, BUF_CMD_WRITE);
        }
        buflist->bs_nchildren = i + 1;
        return (buflist);
}

void
cluster_append(struct bio *bio, struct buf *tbp)
{
        tbp->b_cluster_next = NULL;
        if (bio->bio_caller_info1.cluster_head == NULL) {
                bio->bio_caller_info1.cluster_head = tbp;
                bio->bio_caller_info2.cluster_tail = tbp;
        } else {
                bio->bio_caller_info2.cluster_tail->b_cluster_next = tbp;
                bio->bio_caller_info2.cluster_tail = tbp;
        }
}

static
void
cluster_setram (struct buf *bp)
{
        bp->b_flags |= B_RAM;
        if (bp->b_xio.xio_npages)
                vm_page_flag_set(bp->b_xio.xio_pages[0], PG_RAM);
}