[PATCH] qemu: fix excluding disk from internal inactive snapshot

[Peter Krempa]

On Wed, Feb 16, 2022 at 15:33:57 +0300, Nikolay Shirokovskiy wrote:
> ср, 16 февр. 2022 г. в 11:58, Peter Krempa <pkrempa at redhat.com>:
> 
> > On Wed, Feb 16, 2022 at 11:38:01 +0300, Nikolay Shirokovskiy wrote:
> > > Currently taking an internal inactive snapshot with a disk excluded from
> > > snapshot using snapshot="no" will fail with error "Disk device '%s' does
> > > not support snapshotting".
> >
> > The idea for internal snapshots since we wanted them to behave the same
> > as they do with 'savevm' is that you can't actually exclude a disk from
> > the snapshot.
> >
> > Could you elaborate how you expect this to co-operate with live internal
> > snapshots?
> >
> >
> I have an issue with readonly usb raw disk. Making internal snapshot of an
> active domain
> is possible in this case if I exclude the disk from snapshot explicitly or
> implicitly.

Well, yeah, you can set it as explicitly excluded since it would be
excluded implicitly. The other way should not be possible.

> However in case of inactive domain snapshot fails with the above error. So
> I guess
> we should allow this case.
> 
> So maybe we should check for readonly property of the disk instead? On
> snapshot creation
> there will be no difference as check in virDomainSnapshotAlignDisks will
> fail for non
> readonly disks.

[...]

> Yet it makes more sense to me to check snapdef disk
> because qemuDomainSnapshotForEachQcow2Raw
> is only to call qemu-img and the decision about the disk set is not its
> concern. Also
> the function is used on snapshot reverting/deletion and we want to base
> on snapdef again and not the current state of readonly attribute on disk.

Okay, so the problem isn't actually with the code in this patch, because
virDomainSnapshotAlignDisks already ensures that the snapshot wouldn't
be done under the same conditions as with the live snapshots.

It's the commit message that is misleading because you didn't at all
mention that the disk was readonly and thus wouldn not take part in the
snapshot anyway.

Thus the code you are changing is correctly addressing the bug, where
the offline snapshotting function is attempting to take the snapshot of
the readonly disk.

I suggest the following commit message:

'qemuDomainSnapshotForEachQcow2Raw' doesn't properly handle the
'VIR_DOMAIN_SNAPSHOT_LOCATION_NONE' setting and thus doesn't skip disks
which were excluded from the snapshot due to being read-only.


With that commit message you can use:

Reviewed-by: Peter Krempa <pkrempa at redhat.com>

> However it is not clear to me why there should be parity between active and
> inactive
> snapshot in terms of possible disk sets. We have constraint on set in case
> of active
> snapshot just due to API limitations. What are the reasons to mimic this
> set for inactive
> snapshot?

It minimizes the state matrix and possible corner cases especially since
the old qemu impl is not flexible in handling partial snapshots.