Re: qcow2: Zero-initialization of external data files

[Eric Blake]

On 4/9/20 9:10 AM, Max Reitz wrote:

> > What happens when an operation attempts to unmap things?  Do we reject
> > all unmap operations when data-file-raw is set (thus leaving a cluster
> > marked as allocated at all times, if we can first guarantee that
> > preallocation set things up that way)?
> No, unmap operations currently work.  qcow2_free_any_clusters() passes
> them through to the external data file.
>
> The problem is that the unmap also zeroes the L2 entry, so if you then
> write data to the raw file, it won’t be visible from the qcow2 side of
> things.  However, I’m not sure whether we support modifications of a raw
> file when it is already “in use” by a qcow2 image, so maybe that’s fine.

We don't support concurrent modification. But if the guest is running 
and unmaps things, then shuts off, then we edit the raw file offline, 
then we restart the guest, the guest should see the results of those 
offline edits.

We have to special-case the qcow2 code to either treat unallocated 
cluster + raw-data-file as not really unallocated (so that we can see 
those edits), or we have to special case it to handle unmap + 
raw-data-file specially (pass unmap to the raw file, but do NOT mark the 
cluster unallocated in the qcow2 wrapper).  Which special case we choose 
for unmap may in turn affect whether it is easier to require 
preallocation=metadata (because we can now guarantee that no cluster 
will ever be marked unallocated in qcow2) or whether the default 
preallocation of leaving clusters "unallocated" in qcow2 still sees the 
proper guest data.

--
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3226
Virtualization:  qemu.org | libvirt.org