/*
* Fill a previously allocated cache entry with data.
*/
static void
vdev_cache_fill(zio_t *fio)
{
vdev_t *vd = fio->io_vd;
vdev_cache_t *vc = &vd->vdev_cache;
vdev_cache_entry_t *ve = fio->io_private;
zio_t *pio;
ASSERT(fio->io_size == VCBS);
/*
* Add data to the cache.
*/
mutex_enter(&vc->vc_lock);
ASSERT(ve->ve_fill_io == fio);
ASSERT(ve->ve_offset == fio->io_offset);
ASSERT(ve->ve_data == fio->io_data);
ve->ve_fill_io = NULL;
/*
* Even if this cache line was invalidated by a missed write update,
* any reads that were queued up before the missed update are still
* valid, so we can satisfy them from this line before we evict it.
*/
while ((pio = zio_walk_parents(fio)) != NULL)
vdev_cache_hit(vc, ve, pio);
if (fio->io_error || ve->ve_missed_update)
vdev_cache_evict(vc, ve);
mutex_exit(&vc->vc_lock);
}
/*
* Allocate an entry in the cache. At the point we don't have the data,
* we're just creating a placeholder so that multiple threads don't all
* go off and read the same blocks.
*/
static vdev_cache_entry_t *
vdev_cache_allocate(zio_t *zio)
{
vdev_cache_t *vc = &zio->io_vd->vdev_cache;
uint64_t offset = P2ALIGN(zio->io_offset, VCBS);
vdev_cache_entry_t *ve;
ASSERT(MUTEX_HELD(&vc->vc_lock));
if (zfs_vdev_cache_size == 0)
return (NULL);
/*
* If adding a new entry would exceed the cache size,
* evict the oldest entry (LRU).
*/
if ((avl_numnodes(&vc->vc_lastused_tree) << zfs_vdev_cache_bshift) >
zfs_vdev_cache_size) {
ve = avl_first(&vc->vc_lastused_tree);
if (ve->ve_fill_io != NULL)
return (NULL);
ASSERT(ve->ve_hits != 0);
vdev_cache_evict(vc, ve);
}
ve = kmem_zalloc(sizeof (vdev_cache_entry_t), KM_SLEEP);
ve->ve_offset = offset;
ve->ve_lastused = lbolt;
ve->ve_data = zio_buf_alloc(VCBS);
avl_add(&vc->vc_offset_tree, ve);
avl_add(&vc->vc_lastused_tree, ve);
return (ve);
}
void
vdev_cache_purge(vdev_t *vd)
{
vdev_cache_t *vc = &vd->vdev_cache;
vdev_cache_entry_t *ve;
mutex_enter(&vc->vc_lock);
while ((ve = avl_first(&vc->vc_offset_tree)) != NULL)
vdev_cache_evict(vc, ve);
mutex_exit(&vc->vc_lock);
}
/*
* Fill a previously allocated cache entry with data.
*/
static void
vdev_cache_fill(zio_t *zio)
{
vdev_t *vd = zio->io_vd;
vdev_cache_t *vc = &vd->vdev_cache;
vdev_cache_entry_t *ve = zio->io_private;
zio_t *dio;
ASSERT(zio->io_size == VCBS);
/*
* Add data to the cache.
*/
mutex_enter(&vc->vc_lock);
ASSERT(ve->ve_fill_io == zio);
ASSERT(ve->ve_offset == zio->io_offset);
ASSERT(ve->ve_data == zio->io_data);
ve->ve_fill_io = NULL;
/*
* Even if this cache line was invalidated by a missed write update,
* any reads that were queued up before the missed update are still
* valid, so we can satisfy them from this line before we evict it.
*/
for (dio = zio->io_delegate_list; dio; dio = dio->io_delegate_next)
vdev_cache_hit(vc, ve, dio);
if (zio->io_error || ve->ve_missed_update)
vdev_cache_evict(vc, ve);
mutex_exit(&vc->vc_lock);
while ((dio = zio->io_delegate_list) != NULL) {
zio->io_delegate_list = dio->io_delegate_next;
dio->io_delegate_next = NULL;
dio->io_error = zio->io_error;
zio_next_stage(dio);
}
}
/*
* Allocate an entry in the cache. At the point we don't have the data,
* we're just creating a placeholder so that multiple threads don't all
* go off and read the same blocks.
*/
static vdev_cache_entry_t *
vdev_cache_allocate(zio_t *zio)
{
vdev_cache_t *vc = &zio->io_vd->vdev_cache;
uint64_t offset = P2ALIGN(zio->io_offset, VCBS);
vdev_cache_entry_t *ve;
ASSERT(MUTEX_HELD(&vc->vc_lock));
if (zfs_vdev_cache_size == 0)
return (NULL);
/*
* If adding a new entry would exceed the cache size,
* evict the oldest entry (LRU).
*/
if ((avl_numnodes(&vc->vc_lastused_tree) << zfs_vdev_cache_bshift) >
zfs_vdev_cache_size) {
ve = avl_first(&vc->vc_lastused_tree);
if (ve->ve_fill_io != NULL)
return (NULL);
ASSERT3U(ve->ve_hits, !=, 0);
vdev_cache_evict(vc, ve);
}