[Qemu-devel] Re: [RFC PATCH 0/5] QEMU VFIO device assignment

[Alex Williamson <alex.williamson@redhat.com>]

On Sun, 2010-07-11 at 21:43 +0300, Avi Kivity wrote:
> On 07/11/2010 09:37 PM, Alex Williamson wrote:
> > On Sun, 2010-07-11 at 21:17 +0300, Avi Kivity wrote:
> >    
> >> On 07/11/2010 09:09 PM, Alex Williamson wrote:
> >>      
> >>> The following series implements QEMU device assignment using the
> >>> proposed VFIO/UIOMMU kernel interfaces.  See the last patch for
> >>> further vfio description.  I've tested this on the v2 VFIO patch,
> >>> with a number of fixes hacked in along the way.  I'll update when
> >>> Tom releases a new version of VFIO.  Hopefully this will provide
> >>> some support for the usefulness of such an interfaces.  Thanks,
> >>>
> >>>
> >>>        
> >> What's the plan for supporting this alongside the existing kvm device
> >> assignment code?
> >>
> >> vfio will only exist in very new kernels, so we have to support the old
> >> code for a while to give people chance to adjust (say 12-18 months).
> >>      
> > I was thinking that vfio device assignment might be the qemu acceptable
> > version of device assignment, and we can let kvm style device assignment
> > live out it's remaining time in the qemu-kvm tree, before it gets
> > deprecated.
> >    
> 
> Definitely, the effort to make qemu-kvm device assignment mergable 
> probably isn't worth it.
> 
> >> Ideally we'd have compatible command line syntax with qemu choosing vfio
> >> if available and falling back to kvm device assignment if not.
> >>      
> > Ideally, yes, but I'm not sure how how feasible that is.  For the
> > command line root user, the syntax is nearly same (s/pci-assign/vfio/),
> > but once we start trying to do libvirt based assignment, passing vfiofd
> > &  uiommufd, it will need to know the difference anyway.  Thanks,
> >    
> 
> I see.  Well, we can probably live with two separate invocations, users 
> who assign devices should be savvy or use libvirt.

One other thing to be aware of is that vfio requires devices to be
PCI-2.3 compliant in order to support DisINTx.  This allows vfio to
support devices making use of shared INTx interrupts, but excludes older
devices that users maybe managed to get assigned to an exclusive
interrupt for kvm style assignment.

I suppose we might be able to make vfio work with either pci 2.3 devices
or older devices with exclusive interrupts if that ends up affecting
many users.

Alex