Re: [Qemu-devel] [PATCH for-4.0 v4 0/2] virtio: Provide version-specific variants of virtio PCI devices

[Andrea Bolognani]

On Tue, 2019-03-05 at 15:38 +0100, Gerd Hoffmann wrote:
>   Hi,
> 
> >   -device virtio-blk-pci-non-transitional \
> >   -device virtio-net-pci-non-transitional \
> >   -device virtio-gpu-pci-non-transitional \
> > 
> > and you wouldn't have to question why you can use the
> > non-transitional variant for pretty much everything, except for the
> > few cases where you can't - for no apparent reason...
> 
> Well, there are no variants, only a single virtio-$foo-pci* device.  So
> you don't have to worry about picking one of the available variants,
> there is no choice in the first place.
> 
> When adding an virtio-gpu-pci-non-transitional variant we'll create
> confusion too, because it wouldn't be a real variant.  We would have two
> 100% identical devices then, and people will probably wonder why they
> exist and what the difference is ...

When looking at a single device, I mostly agree with your assessment;
however, when looking at the overall situation with VirtIO devices,
one might quite reasonably infer the following rules:

  * devices marked as (non-)transitional are going to show up as
    (non-)transitional;

  * unmarked devices might show up as either one, depending on some
    factor which is not immediately obvious.

So if you knew you wanted non-transitional devices you would expect
to just use the non-transitional variant for *all* VirtIO devices,
including virtio-gpu, without necessarily caring whether the unmarked
devices behaves any differently; if you tried to use the transitional
device, you'd get an error message telling you that device doesn't
exist, which is pretty reasonable and easy to research / understand.

With the current situation, once you've tried using non-transitional
virtio-gpu and gotten back an error message, there's quite a bit more
digging required to figure out *why* the device is not there in the
first place.

So I agree neither scenario is exactly perfect, but I still think
adding non-transitional alias devices would overall be more
user-friendly.

> So I can't see how this would be so much better.  We have to document
> the mess no matter what.

We have some documentation in libvirt:

  https://libvirt.org/formatdomain.html#elementsVirtioTransitional

Not that more / improved documentation is ever a bad idea :)

-- 
Andrea Bolognani / Red Hat / Virtualization