Re: [Qemu-block] [PATCH v12 05/10] qcow2: Optimize zero_single_l2() to minimize L2 churn

[Max Reitz]

On 04.05.2017 05:07, Eric Blake wrote:
> Similar to discard_single_l2(), we should try to avoid dirtying
> the L2 cache when the cluster we are changing already has the
> right characteristics.
> 
> Note that by the time we get to zero_single_l2(), BDRV_REQ_MAY_UNMAP
> is a requirement to unallocate a cluster (this is because the block
> layer clears that flag if discard.* flags during open requested that
> we never punch holes - see the conversation around commit 170f4b2e,
> https://lists.gnu.org/archive/html/qemu-devel/2016-09/msg07306.html).
> Therefore, this patch can only reuse a zero cluster as-is if either
> unmapping is not requested, or if the zero cluster was not associated
> with an allocation.
> 
> Technically, there are some cases where an unallocated cluster
> already reads as all zeroes (namely, when there is no backing file
> [easy: check bs->backing], or when the backing file also reads as
> zeroes [harder: we can't check bdrv_get_block_status since we are
> already holding the lock]), where the guest would not immediately see
> a difference if we left that cluster unallocated.  But if the user
> did not request unmapping, leaving an unallocated cluster is wrong;
> and even if the user DID request unmapping, keeping a cluster
> unallocated risks a subtle semantic change of guest-visible contents
> if a backing file is later added, and it is not worth auditing
> whether all internal uses such as mirror properly avoid an unmap
> request.  Thus, this patch is intentionally limited to just clusters
> that are already marked as zero.
> 
> Signed-off-by: Eric Blake <address@hidden>
> 
> ---
> v12: store cluster type in temporary
> v11: reserved for blkdebug half of v10
> v10: new patch, replacing earlier attempt to use unallocated clusters,
> and ditching any optimization of v2 files
> ---
>  block/qcow2-cluster.c | 15 +++++++++++++--
>  1 file changed, 13 insertions(+), 2 deletions(-)

Reviewed-by: Max Reitz <address@hidden>

> 
> diff --git a/block/qcow2-cluster.c b/block/qcow2-cluster.c
> index 14e2086..78fbe34 100644
> --- a/block/qcow2-cluster.c
> +++ b/block/qcow2-cluster.c
> @@ -1599,6 +1599,7 @@ static int zero_single_l2(BlockDriverState *bs, 
> uint64_t offset,
>      int l2_index;
>      int ret;
>      int i;
> +    bool unmap = !!(flags & BDRV_REQ_MAY_UNMAP);
> 
>      ret = get_cluster_table(bs, offset, &l2_table, &l2_index);
>      if (ret < 0) {
> @@ -1611,12 +1612,22 @@ static int zero_single_l2(BlockDriverState *bs, 
> uint64_t offset,
> 
>      for (i = 0; i < nb_clusters; i++) {
>          uint64_t old_offset;
> +        int cluster_type;

Hm, why doesn't this enum have a name yet... Well, if only *someone*
could address that. */me ducks*

Max

> 
>          old_offset = be64_to_cpu(l2_table[l2_index + i]);
> 
> -        /* Update L2 entries */
> +        /*
> +         * Minimize L2 changes if the cluster already reads back as
> +         * zeroes with correct allocation.
> +         */
> +        cluster_type = qcow2_get_cluster_type(old_offset);
> +        if (cluster_type == QCOW2_CLUSTER_ZERO_PLAIN ||
> +            (cluster_type == QCOW2_CLUSTER_ZERO_ALLOC && !unmap)) {
> +            continue;
> +        }
> +
>          qcow2_cache_entry_mark_dirty(bs, s->l2_table_cache, l2_table);
> -        if (old_offset & QCOW_OFLAG_COMPRESSED || flags & 
> BDRV_REQ_MAY_UNMAP) {
> +        if (cluster_type == QCOW2_CLUSTER_COMPRESSED || unmap) {
>              l2_table[l2_index + i] = cpu_to_be64(QCOW_OFLAG_ZERO);
>              qcow2_free_any_clusters(bs, old_offset, 1, 
> QCOW2_DISCARD_REQUEST);
>          } else {
>

signature.asc
Description: OpenPGP digital signature