Re: [Qemu-devel] [PATCH 1/5] qcow2: Work with bytes in qcow2_get_cluster_offset()

[Eric Blake <eblake@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 4334 bytes --]

On 06/03/2016 11:21 AM, Kevin Wolf wrote:
> This patch changes the units that qcow2_get_cluster_offset() uses
> internally, without touching the interface just yet. This will be done
> in another patch.
> 
> Signed-off-by: Kevin Wolf <kwolf@redhat.com>
> ---
>  block/qcow2-cluster.c | 43 ++++++++++++++++++++++---------------------
>  1 file changed, 22 insertions(+), 21 deletions(-)
> 
> diff --git a/block/qcow2-cluster.c b/block/qcow2-cluster.c
> index d901d89..b2405b1 100644
> --- a/block/qcow2-cluster.c
> +++ b/block/qcow2-cluster.c
> @@ -482,29 +482,27 @@ int qcow2_get_cluster_offset(BlockDriverState *bs, uint64_t offset,
>      unsigned int l2_index;
>      uint64_t l1_index, l2_offset, *l2_table;
>      int l1_bits, c;
> -    unsigned int index_in_cluster, nb_clusters;
> -    uint64_t nb_available, nb_needed;
> +    unsigned int offset_in_cluster, nb_clusters;
> +    uint64_t bytes_available, bytes_needed;
>      int ret;
> +    unsigned int bytes;
>  
> -    index_in_cluster = (offset >> 9) & (s->cluster_sectors - 1);
> -    nb_needed = *num + index_in_cluster;
> +    bytes = *num * BDRV_SECTOR_SIZE;

bytes can overflow if num > BDRV_REQUEST_MAX_SECTORS.  Is that ever a
problem? Would an assert() help, or else clamping it (since it is really
just a hint of how far we are allowed to look for similar clusters, but
we can always quit looking early):

bytes = MIN(*num * BDRV_SECTOR_SIZE, UINT32_MAX);

Then again, the interface is about to change, so this may just be a
transient problem.

>  
> -    l1_bits = s->l2_bits + s->cluster_bits;
> -
> -    /* compute how many bytes there are between the offset and
> -     * the end of the l1 entry
> -     */
> +    offset_in_cluster = offset_into_cluster(s, offset);
> +    bytes_needed = bytes + offset_in_cluster;

Hmm, my idea for clamped bytes above may need tweaking (that is,
UINT32_MAX is inappropriate as the clamp value, if you are going to
round it up here).  Or, since bytes_needed is 64-bits, make either bytes
or offset_in_cluster also be 64 bits, or start with '0ULL +', to avoid
overflow.

>  
> -    nb_available = (1ULL << l1_bits) - (offset & ((1ULL << l1_bits) - 1));
> -
> -    /* compute the number of available sectors */
> +    l1_bits = s->l2_bits + s->cluster_bits;
>  
> -    nb_available = (nb_available >> 9) + index_in_cluster;
> +    /* compute how many bytes there are between the start of the cluster
> +     * containing offset and the end of the l1 entry */
> +    bytes_available = (1ULL << l1_bits) - (offset & ((1ULL << l1_bits) - 1))
> +                    + offset_in_cluster;
>  
> -    if (nb_needed > nb_available) {
> -        nb_needed = nb_available;
> +    if (bytes_needed > bytes_available) {
> +        bytes_needed = bytes_available;
>      }
> -    assert(nb_needed <= INT_MAX);
> +    assert(bytes_needed <= INT_MAX);

Is this assertion too late given the spots I pointed out above as
possible overflow points?

>  
>      *cluster_offset = 0;
>  
> @@ -542,7 +540,7 @@ int qcow2_get_cluster_offset(BlockDriverState *bs, uint64_t offset,
>      *cluster_offset = be64_to_cpu(l2_table[l2_index]);
>  
>      /* nb_needed <= INT_MAX, thus nb_clusters <= INT_MAX, too */
> -    nb_clusters = size_to_clusters(s, nb_needed << 9);
> +    nb_clusters = size_to_clusters(s, bytes_needed);
>  
>      ret = qcow2_get_cluster_type(*cluster_offset);
>      switch (ret) {
> @@ -589,13 +587,16 @@ int qcow2_get_cluster_offset(BlockDriverState *bs, uint64_t offset,
>  
>      qcow2_cache_put(bs, s->l2_table_cache, (void**) &l2_table);
>  
> -    nb_available = (c * s->cluster_sectors);
> +    bytes_available = (c * s->cluster_size);
>  
>  out:
> -    if (nb_available > nb_needed)
> -        nb_available = nb_needed;
> +    if (bytes_available > bytes_needed) {
> +        bytes_available = bytes_needed;
> +    }
>  
> -    *num = nb_available - index_in_cluster;
> +    bytes = bytes_available - offset_in_cluster;
> +    assert((bytes & (BDRV_SECTOR_SIZE - 1)) == 0);
> +    *num = bytes >> BDRV_SECTOR_BITS;
>  
>      return ret;

Looks like it is headed in the right direction, though.

-- 
Eric Blake   eblake redhat com    +1-919-301-3266
Libvirt virtualization library http://libvirt.org


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 604 bytes --]