Re: [PATCH 0/3] VFIO-based PCI device assignment for QEMU 1.2

[Jan Kiszka <jan.kiszka@web.de>]

[-- Attachment #1: Type: text/plain, Size: 1673 bytes --]

On 2012-08-13 21:31, Anthony Liguori wrote:
> Jan Kiszka <jan.kiszka@siemens.com> writes:
> 
>> On 2012-08-13 15:58, Avi Kivity wrote:
>>> On 08/13/2012 04:27 PM, Anthony Liguori wrote:
>>>
>>>> Thanks for pushing this forward!  Hopefully this will finally kill off
>>>> qemu-kvm.git for good.
>>>
>>> No, it won't.  vfio requires a 3.6 kernel, which we cannot assume anyone
>>> has.  We'll need the original device assignment code side-by-side.
>>
>> ...which is on my to-do list for 1.3.
> 
> Is there a deprecation plan for the old device assignment code?
> 
> I'm not really against the idea of requiring a new kernel for new
> features.
> 
> From a Fedora/OpenSUSE point of view, would supporting old kernels be a
> requirement to stop shipping qemu-kvm.git over qemu.git?
> 
> Since distros ship new kernels and new userspaces, I don't think distros
> would care so I'm not sure who we're trying to support old kernels for.

We are supporting KVM down to 2.6.3x, if not 2.6.2x. Also, device
assignment is a new feature for upstream, but not for the masses of KVM
users of QEMU (due to qemu-kvm and corresponding libvirt support). I
think it will take some more kernel releases to have all feature there
that allows performance-wise equivalent device assignment via VFIO. And
it can even be helpful to cross-check issues of VFIO in the field.

Except for some self-contained helper functions in the KVM layer,
classic device assignment will be as isolated as VFIO. So I don't think
we would take any noteworthy burden to maintain it as long as the kernel
supports this interface.

Can't comment on the other questions.

Jan



[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 262 bytes --]