Libguestfs - Mar 2019 - Re: [PATCH nbdkit v4 05/15] cow: Disable extents information in this filter.

On 3/26/19 4:17 PM, Richard W.M. Jones wrote:
> The cow filter doesn't support trimming and we should assume the
> underlying plugin is fully allocated too.  Note that both of these
> limitations might be lifted with a more advanced filter
> implementation.
> 
> However we ought to support this in future because xz files contain
> sparseness information, so add a note.
Stale copy-and-paste comment?

Also, since the series is mostly alphabetical, I suspect you're missing
4.5/15 for filters/cache/cache.c (blindly passing through .extents to
the underlying plugin); easiest is probably forcing .can_extents false,
but nicer would be remembering status of portions that we have cached
locally to avoid repeated callouts to the plugin.

ACK for the cow filter.

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

On Wed, Mar 27, 2019 at 09:45:58PM -0500, Eric Blake
wrote:
> On 3/26/19 4:17 PM, Richard W.M. Jones wrote:
> > The cow filter doesn't support trimming and we should assume the
> > underlying plugin is fully allocated too.  Note that both of these
> > limitations might be lifted with a more advanced filter
> > implementation.
> > 
> > However we ought to support this in future because xz files contain
> > sparseness information, so add a note.
> 
> Stale copy-and-paste comment?
Yes that was a copy-paste bug (or rather commit -c).  I've fixed it.
> Also, since the series is mostly alphabetical, I suspect you're missing
> 4.5/15 for filters/cache/cache.c (blindly passing through .extents to
> the underlying plugin); easiest is probably forcing .can_extents false,
> but nicer would be remembering status of portions that we have cached
> locally to avoid repeated callouts to the plugin.
I examined the cache filter and I don't think it needs any changes for
correctness.  It could - as you note - be improved so it caches
extents information, but I think if it does nothing then it's still
correct.

Note that nbdkit will automatically pass the extents call through - we
don't need to add an explicit extents callback in filters which don't
want to modify the result.  Also .can_extents is properly taken into
account.  (I think the man page should be clearer about all this.)

Rich.

-- 
Richard Jones, Virtualization Group, Red Hat http://people.redhat.com/~rjones
Read my programming and virtualization blog: http://rwmj.wordpress.com
virt-top is 'top' for virtual machines.  Tiny program with many
powerful monitoring features, net stats, disk stats, logging, etc.
http://people.redhat.com/~rjones/virt-top

On 3/28/19 5:32 AM, Richard W.M. Jones wrote:
> On Wed, Mar 27, 2019 at 09:45:58PM -0500, Eric Blake wrote:
>> On 3/26/19 4:17 PM, Richard W.M. Jones wrote:
>>> The cow filter doesn't support trimming and we should assume
the
>>> underlying plugin is fully allocated too.  Note that both of these
>>> limitations might be lifted with a more advanced filter
>>> implementation.
>>>
>>> However we ought to support this in future because xz files contain
>>> sparseness information, so add a note.
>>
>> Stale copy-and-paste comment?
> 
> Yes that was a copy-paste bug (or rather commit -c).  I've fixed it.
> 
>> Also, since the series is mostly alphabetical, I suspect you're
missing
>> 4.5/15 for filters/cache/cache.c (blindly passing through .extents to
>> the underlying plugin); easiest is probably forcing .can_extents false,
>> but nicer would be remembering status of portions that we have cached
>> locally to avoid repeated callouts to the plugin.
> 
> I examined the cache filter and I don't think it needs any changes for
> correctness.  It could - as you note - be improved so it caches
> extents information, but I think if it does nothing then it's still
> correct.
Still, it may be worth a comment in the file, to specifically document
that we've thought about it, and that if performance proves to be slow
(vddk, anyone? and qemu-img's current known inefficiencies with REQ_ONE)
that it may be helpful as a future addition.
> 
> Note that nbdkit will automatically pass the extents call through - we
> don't need to add an explicit extents callback in filters which
don't
> want to modify the result.  Also .can_extents is properly taken into
> account.  (I think the man page should be clearer about all this.)
A doc update might not hurt, but yeah, I already realized that the
default behavior (if you don't do anything else) is for both
.can_extents and .extents to blindly pass through.

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