/*
* Same as above, but forces the page to be detached from the object
* and go into free pool.
*/
void
sf_ext_free_nocache(void *arg1, void *arg2)
{
struct sf_buf *sf = arg1;
struct sendfile_sync *sfs = arg2;
vm_page_t pg = sf_buf_page(sf);
sf_buf_free(sf);
vm_page_lock(pg);
if (vm_page_unwire(pg, PQ_NONE)) {
vm_object_t obj;
/* Try to free the page, but only if it is cheap to. */
if ((obj = pg->object) == NULL)
vm_page_free(pg);
else if (!vm_page_xbusied(pg) && VM_OBJECT_TRYWLOCK(obj)) {
vm_page_free(pg);
VM_OBJECT_WUNLOCK(obj);
} else
vm_page_deactivate(pg);
}
vm_page_unlock(pg);
if (sfs != NULL) {
mtx_lock(&sfs->mtx);
KASSERT(sfs->count > 0, ("Sendfile sync botchup count == 0"));
if (--sfs->count == 0)
cv_signal(&sfs->cv);
mtx_unlock(&sfs->mtx);
}
}
/*
* 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.
*/
static struct buf *
cluster_rbuild(struct vnode *vp, u_quad_t filesize, daddr_t lbn,
daddr_t blkno, long size, int run, int gbflags, struct buf *fbp)
{
struct buf *bp, *tbp;
daddr_t bn;
off_t off;
long tinc, tsize;
int i, inc, j, k, toff;
KASSERT(size == vp->v_mount->mnt_stat.f_iosize,
("cluster_rbuild: size %ld != f_iosize %jd\n",
size, (intmax_t)vp->v_mount->mnt_stat.f_iosize));
/*
* avoid a division
*/
while ((u_quad_t) size * (lbn + run) > filesize) {
--run;
}
if (fbp) {
tbp = fbp;
tbp->b_iocmd = BIO_READ;
} else {
tbp = getblk(vp, lbn, size, 0, 0, gbflags);
if (tbp->b_flags & B_CACHE)
return tbp;
tbp->b_flags |= B_ASYNC | B_RAM;
tbp->b_iocmd = BIO_READ;
}
tbp->b_blkno = blkno;
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_flags = B_ASYNC | B_CLUSTER | B_VMIO;
if ((gbflags & GB_UNMAPPED) != 0) {
bp->b_data = unmapped_buf;
} else {
bp->b_data = (char *)((vm_offset_t)bp->b_data |
((vm_offset_t)tbp->b_data & PAGE_MASK));
}
bp->b_iocmd = BIO_READ;
bp->b_iodone = cluster_callback;
bp->b_blkno = blkno;
bp->b_lblkno = lbn;
bp->b_offset = tbp->b_offset;
KASSERT(bp->b_offset != NOOFFSET, ("cluster_rbuild: no buffer offset"));
pbgetvp(vp, bp);
TAILQ_INIT(&bp->b_cluster.cluster_head);
bp->b_bcount = 0;
bp->b_bufsize = 0;
bp->b_npages = 0;
inc = btodb(size);
for (bn = blkno, i = 0; i < run; ++i, bn += inc) {
if (i == 0) {
VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
vfs_drain_busy_pages(tbp);
vm_object_pip_add(tbp->b_bufobj->bo_object,
tbp->b_npages);
for (k = 0; k < tbp->b_npages; k++)
vm_page_sbusy(tbp->b_pages[k]);
VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
} else {
if ((bp->b_npages * PAGE_SIZE) +
round_page(size) > vp->v_mount->mnt_iosize_max) {
break;
}
tbp = getblk(vp, lbn + i, size, 0, 0, GB_LOCK_NOWAIT |
(gbflags & GB_UNMAPPED));
/* Don't wait around for locked bufs. */
if (tbp == NULL)
break;
/*
* Stop scanning if the buffer is fully 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. The bo lock is not
* required for the BKGRDINPROG check since it
* can not be set without the buf lock.
*/
if ((tbp->b_vflags & BV_BKGRDINPROG) ||
(tbp->b_flags & B_CACHE) ||
(tbp->b_flags & B_VMIO) == 0) {
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.
*/
off = tbp->b_offset;
tsize = size;
VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
for (j = 0; tsize > 0; j++) {
toff = off & PAGE_MASK;
tinc = tsize;
if (toff + tinc > PAGE_SIZE)
tinc = PAGE_SIZE - toff;
VM_OBJECT_ASSERT_WLOCKED(tbp->b_pages[j]->object);
if ((tbp->b_pages[j]->valid &
vm_page_bits(toff, tinc)) != 0)
break;
if (vm_page_xbusied(tbp->b_pages[j]))
break;
vm_object_pip_add(tbp->b_bufobj->bo_object, 1);
vm_page_sbusy(tbp->b_pages[j]);
off += tinc;
tsize -= tinc;
}
if (tsize > 0) {
clean_sbusy:
vm_object_pip_add(tbp->b_bufobj->bo_object, -j);
for (k = 0; k < j; k++)
vm_page_sunbusy(tbp->b_pages[k]);
VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
bqrelse(tbp);
break;
}
VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
/*
* Set a read-ahead mark as appropriate
*/
if ((fbp && (i == 1)) || (i == (run - 1)))
tbp->b_flags |= B_RAM;
/*
* Set the buffer up for an async read (XXX should
* we do this only if we do not wind up brelse()ing?).
* Set the block number if it isn't set, otherwise
* if it is make sure it matches the block number we
* expect.
*/
tbp->b_flags |= B_ASYNC;
tbp->b_iocmd = BIO_READ;
if (tbp->b_blkno == tbp->b_lblkno) {
tbp->b_blkno = bn;
} else if (tbp->b_blkno != bn) {
VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
goto clean_sbusy;
}
}
/*
* XXX fbp from caller may not be B_ASYNC, but we are going
* to biodone() it in cluster_callback() anyway
*/
BUF_KERNPROC(tbp);
TAILQ_INSERT_TAIL(&bp->b_cluster.cluster_head,
tbp, b_cluster.cluster_entry);
VM_OBJECT_WLOCK(tbp->b_bufobj->bo_object);
for (j = 0; j < tbp->b_npages; j += 1) {
vm_page_t m;
m = tbp->b_pages[j];
if ((bp->b_npages == 0) ||
(bp->b_pages[bp->b_npages-1] != m)) {
bp->b_pages[bp->b_npages] = m;
bp->b_npages++;
}
if (m->valid == VM_PAGE_BITS_ALL)
tbp->b_pages[j] = bogus_page;
}
VM_OBJECT_WUNLOCK(tbp->b_bufobj->bo_object);
/*
* 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 != size)
printf("warning: tbp->b_bcount wrong %ld vs %ld\n", tbp->b_bcount, size);
if (tbp->b_bufsize != size)
printf("warning: tbp->b_bufsize wrong %ld vs %ld\n", tbp->b_bufsize, size);
bp->b_bcount += size;
bp->b_bufsize += size;
}
/*
* 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
*/
VM_OBJECT_WLOCK(bp->b_bufobj->bo_object);
for (j = 0; j < bp->b_npages; j++) {
VM_OBJECT_ASSERT_WLOCKED(bp->b_pages[j]->object);
if (bp->b_pages[j]->valid == VM_PAGE_BITS_ALL)
bp->b_pages[j] = bogus_page;
}
VM_OBJECT_WUNLOCK(bp->b_bufobj->bo_object);
if (bp->b_bufsize > bp->b_kvasize)
panic("cluster_rbuild: b_bufsize(%ld) > b_kvasize(%d)\n",
bp->b_bufsize, bp->b_kvasize);
if (buf_mapped(bp)) {
pmap_qenter(trunc_page((vm_offset_t) bp->b_data),
(vm_page_t *)bp->b_pages, bp->b_npages);
}
return (bp);
}