qcow2_discard_clusters() is set up to silently ignore sub-cluster
head or tail on unaligned requests.  However, it is easy to audit
the various callers: qcow2_snapshot_create() has always passed
aligned data since the call was introduced in 1ebf561;
qcow2_co_pdiscard() has passed aligned clusters since commit
ecdbead taught the block layer the preferred discard alignment (the
block layer can still pass sub-cluster values, but those are
handled directly in qcow2_co_pdiscard()); and qcow2_make_empty()
was fixed to pass aligned clusters in commit a3e1505.  Replace
rounding with assertions to hold us to the tighter contract,
eliminating the now-impossible case of an early exit for a
sub-cluster request.

qcow2_zero_clusters() has always been called with cluster-aligned
arguments from its lone caller qcow2_co_pwrite_zeroes() (like
qcow2_co_pdiscard(), the caller takes care of sub-cluster requests
from the block layer; and qcow2_zero_clusters() would have
misbehaved on unaligned requests), but it deserves the same
assertion for symmetry.

Signed-off-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>

---
v5: no change
v4: new patch
---
 block/qcow2-cluster.c | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

diff --git a/block/qcow2-cluster.c b/block/qcow2-cluster.c
index 928c1e2..3304a15 100644
--- a/block/qcow2-cluster.c
+++ b/block/qcow2-cluster.c
@@ -1521,13 +1521,9 @@ int qcow2_discard_clusters(BlockDriverState *bs, uint64_t offset,

     end_offset = offset + (nb_sectors << BDRV_SECTOR_BITS);

-    /* Round start up and end down */
-    offset = align_offset(offset, s->cluster_size);
-    end_offset = start_of_cluster(s, end_offset);
-
-    if (offset > end_offset) {
-        return 0;
-    }
+    /* Caller must pass aligned values */
+    assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
+    assert(QEMU_IS_ALIGNED(end_offset, s->cluster_size));

     nb_clusters = size_to_clusters(s, end_offset - offset);

@@ -1602,6 +1598,10 @@ int qcow2_zero_clusters(BlockDriverState *bs, uint64_t offset, int nb_sectors,
     uint64_t nb_clusters;
     int ret;

+    /* Caller must pass aligned values */
+    assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
+    assert(QEMU_IS_ALIGNED(nb_sectors, s->cluster_size >> BDRV_SECTOR_BITS));
+
     /* The zero flag is only supported by version 3 and newer */
     if (s->qcow_version < 3) {
         return -ENOTSUP;
-- 
2.9.3

Am 14.02.2017 um 20:25 hat Eric Blake geschrieben:
> qcow2_discard_clusters() is set up to silently ignore sub-cluster
> head or tail on unaligned requests.  However, it is easy to audit
> the various callers: qcow2_snapshot_create() has always passed
> aligned data since the call was introduced in 1ebf561;
> qcow2_co_pdiscard() has passed aligned clusters since commit
> ecdbead taught the block layer the preferred discard alignment (the
> block layer can still pass sub-cluster values, but those are
> handled directly in qcow2_co_pdiscard()); and qcow2_make_empty()
> was fixed to pass aligned clusters in commit a3e1505.

I don't think this is true for the very part in the image if the image
size isn't cluster aligned:

    ret = qcow2_discard_clusters(bs, start_sector * BDRV_SECTOR_SIZE,
                                 MIN(sector_step,
                                     bs->total_sectors - start_sector),
                                 QCOW2_DISCARD_SNAPSHOT, true);

sector_step is alright after commit a3e1505, but bs->total_sectors can
be unaligned.

The other callers look okay.

> Replace
> rounding with assertions to hold us to the tighter contract,
> eliminating the now-impossible case of an early exit for a
> sub-cluster request.
> 
> qcow2_zero_clusters() has always been called with cluster-aligned
> arguments from its lone caller qcow2_co_pwrite_zeroes() (like
> qcow2_co_pdiscard(), the caller takes care of sub-cluster requests
> from the block layer; and qcow2_zero_clusters() would have
> misbehaved on unaligned requests), but it deserves the same
> assertion for symmetry.
> 
> Signed-off-by: Eric Blake <eblake@redhat.com>
> Reviewed-by: Max Reitz <mreitz@redhat.com>
> 
> ---
> v5: no change
> v4: new patch
> ---
>  block/qcow2-cluster.c | 14 +++++++-------
>  1 file changed, 7 insertions(+), 7 deletions(-)
> 
> diff --git a/block/qcow2-cluster.c b/block/qcow2-cluster.c
> index 928c1e2..3304a15 100644
> --- a/block/qcow2-cluster.c
> +++ b/block/qcow2-cluster.c
> @@ -1521,13 +1521,9 @@ int qcow2_discard_clusters(BlockDriverState *bs, uint64_t offset,
> 
>      end_offset = offset + (nb_sectors << BDRV_SECTOR_BITS);
> 
> -    /* Round start up and end down */
> -    offset = align_offset(offset, s->cluster_size);
> -    end_offset = start_of_cluster(s, end_offset);
> -
> -    if (offset > end_offset) {
> -        return 0;
> -    }
> +    /* Caller must pass aligned values */
> +    assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
> +    assert(QEMU_IS_ALIGNED(end_offset, s->cluster_size));
> 
>      nb_clusters = size_to_clusters(s, end_offset - offset);
> 
> @@ -1602,6 +1598,10 @@ int qcow2_zero_clusters(BlockDriverState *bs, uint64_t offset, int nb_sectors,
>      uint64_t nb_clusters;
>      int ret;
> 
> +    /* Caller must pass aligned values */
> +    assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
> +    assert(QEMU_IS_ALIGNED(nb_sectors, s->cluster_size >> BDRV_SECTOR_BITS));

Maybe using s->cluster_sectors is nicer than doing the shift here. Not
too important, the code is replaced in the next patch anyway.

>      /* The zero flag is only supported by version 3 and newer */
>      if (s->qcow_version < 3) {
>          return -ENOTSUP;

Kevin

On 02/15/2017 06:25 AM, Kevin Wolf wrote:
> Am 14.02.2017 um 20:25 hat Eric Blake geschrieben:
>> qcow2_discard_clusters() is set up to silently ignore sub-cluster
>> head or tail on unaligned requests.  However, it is easy to audit
>> the various callers: qcow2_snapshot_create() has always passed
>> aligned data since the call was introduced in 1ebf561;
>> qcow2_co_pdiscard() has passed aligned clusters since commit
>> ecdbead taught the block layer the preferred discard alignment (the
>> block layer can still pass sub-cluster values, but those are
>> handled directly in qcow2_co_pdiscard()); and qcow2_make_empty()
>> was fixed to pass aligned clusters in commit a3e1505.
> 
> I don't think this is true for the very part in the image if the image
> size isn't cluster aligned:
> 
>     ret = qcow2_discard_clusters(bs, start_sector * BDRV_SECTOR_SIZE,
>                                  MIN(sector_step,
>                                      bs->total_sectors - start_sector),
>                                  QCOW2_DISCARD_SNAPSHOT, true);
> 
> sector_step is alright after commit a3e1505, but bs->total_sectors can
> be unaligned.

Indeed; a slight tweak to qemu-iotests/097 exposes the problem.  I'll
include that in my next spin (although I note that Dan has patches on
the list to split 097 into two tests, in order to give qcow coverage
rather than just qcow2).

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