Re: [PATCH v12 2/3] qcow2: Allow writing compressed data of multiple clusters

[Vladimir Sementsov-Ogievskiy]

10.04.2020 14:12, Alberto Garcia wrote:
> On Thu 09 Apr 2020 08:39:12 PM CEST, Vladimir Sementsov-Ogievskiy wrote:
> > > Because of this a test cannot expect that running the same commands on
> > > an empty image produces always the same results.
> > >
> > > Is this something that we should be concerned about?
> >
> > Parallel writing compressed clusters is significant improvement, as it
> > allow compressing in really parallel threads.
>
> I see, I just wasn't sure if you were aware of this side effect.

No, we didn't thought about it, so good to know, thanks.

> > So, I don't think we should make specific workaround for
> > testing... What exactly is the case?
>
> I noticed this while writing some tests for the subcluster allocation
> feature, but this is not a problem for me. Many of our iotests make
> assumptions about the location of L2 and refcount tables so changing
> those would break a lot of them. This thing only changes the offset of
> the compressed data clusters (and their L2 entries), but as far as I'm
> aware no one relies on them being predictable. I just need to make sure
> that I don't do it either.

OK. I had similar problems (because of asynchronicity) with existing iotests
in may series for backup. As I remember, I had to add options to just disable
asynchronicity for some tests. So, if needed, we can add some options for
qcow2 (which can be used to justify number of parallel requests, not only to
disable them at all). Still, of course, it's better to avoid testing only
sequential IO when it is async without options.



--
Best regards,
Vladimir