Re: [PATCH v6 04/10] qcow2: Support BDRV_REQ_ZERO_WRITE for truncate

[Eric Blake <eblake@redhat.com>]

On 4/24/20 7:17 AM, Kevin Wolf wrote:

> 
>>> +            ret = qcow2_co_pwritev_part(bs, old_length, qiov.size, &qiov, 0, 0);
>>
>> Better not do it if this cluster is already ZERO.. But it may be a
>> later patch to improve that.
> 
> I doubt it's worth writing code to optimise a corner case inside a
> corner case.

I spotted the same issue, and my suggestion was to use 
qcow2_co_pwrite_zero instead of qcow2_co_pwritev_part, but Kevin pointed 
out that my idea would probably not work the way I thought.  Then again, 
checking if the page is already zero is all the more benefit that 
qcow2_co_pwrite_zero would have given for me to have raised the 
question.  The following example illustrates why it might be worthwhile:

Suppose we have an image with 1M clusters, which is an unaligned 100.5M 
in length but started life with all clusters sparse.  We then resize it 
to another unaligned 200.5M.  The call to qcow2_co_pwritev_part will 
cause the image to be non-sparse for one cluster out of 201, whereas 
adding a check to skip the pwritev because the original unaligned tail 
cluster already read as zero will let us keep all clusters sparse.

At any rate, I argued that any such optimization would be a followup 
patch, and Kevin is right that it is a corner case optimization unlikely 
to affect many real-life images.

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