Re: AMD KVM Pci Passthrough reports device busy

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

On Tue, 2012-06-05 at 18:55 +0200, Andreas Hartmann wrote:
> Alex Williamson wrote:
> [...]
> > Yep, I think the previous suggestion about reloading vfio_iommu_type1
> > with allow_unsafe_interrupts=1 will solve it.
> 
> Yes! Works now. Success!!!!!
> 
> Works means: Device is seen in VM. I couldn't test it up to now, because
> I don't have any driver in the VM for this device.

I tested it as far as doing 'cat /dev/video0 > /tmp/test'.  The file
grows, the device gets interrupts and `file` thinks it's a video (no
input connected to the card).

> > It has nothing to do with
> > file permissions, you're getting EPERM at the point where we set the
> > iommu type, which enables access to devices.  By default we want an
> > iommu which protects against malicious MSI attacks, which requires
> > interrupt remapping on x86.  Joerg has been working to add this for AMD
> > since the hardware supports it, but for now, the above options lets us
> > bypass the check.  Thanks,
> 
> Now, I have to check my problem with the filesystem first, which isn't
> cleanly unmounted. I suspect a few missing options for kernel compile.
> Next will be fglrx :-) for 3.4
> 
> Hmmm, is there a chance to get your extension isolated to get it
> compiled for 3.1, too?

The patches are all layered on top of that branch, so you should be able
to easily 'git show' to get patches or cherry-pick them across trees.  I
don't think there should be too much trouble backporting them.  Thanks,

Alex