Re: qcow2 preallocation and backing files

[Max Reitz]

On 20.11.19 16:58, Vladimir Sementsov-Ogievskiy wrote:
> 20.11.2019 18:18, Alberto Garcia wrote:
>> On Wed 20 Nov 2019 01:27:53 PM CET, Vladimir Semeeausntsov-Ogievskiy wrote:
>>
>>> 3. Also, the latter way is inconsistent with discard. Discarded
>>> regions returns zeroes, not clusters from backing. I think discard and
>>> truncate should behave in the same safe zero way.
>>
>> But then PREALLOC_MODE_OFF implies that the L2 metadata should be
>> preallocated (all clusters should be QCOW2_CLUSTER_ZERO_PLAIN), at least
>> when there is a backing file.
>>
>> Or maybe we just forbid PREALLOC_MODE_OFF during resize if there is a
>> backing file ?
>>
> 
> Kevin proposed a fix that alters PREALLOC_MODE_OFF behavior if there is
> a backing file, to allocate L2 metadata with ZERO clusters..
> 
> I don't think that it's the best thing to do, but it's already done, it works
> and seems appropriate for rc3..
> 
> I see now, that change PREALLOC_MODE_OFF behavior may break things, first of
> all qemu-img create, which creating UNALLOCATED qcow2 by default for years.
> 
> Still, I think that it would be safer to always ZERO expanded part of qcow2,
> regardless of backing file..
> 
> We may add PREALLOC_MODE_ZERO, and use it in mirror, commit, and some other 
> calls
> to bdrv_truncate, except for qcow2 image creation of course.

Well, the good news is that block_resize currently has no such parameter
and thus we could make a non-OFF value the default.

The bad news is that the reason it has no such parameter is that it
would need to be a job in order to support any form of preallocation.

So actually I don’t think we can make block_resize write zeroes to the
new region by default, because then it needs to be a job, and it just
isn’t.  (Of course, we can get to it through a deprecation cycle, but we
can’t “fix” block_resize directly.)

Max

signature.asc
Description: OpenPGP digital signature