[vfio-users] VGA Passthrough limitations?

[Alex Williamson]

On Wed, May 18, 2016 at 1:33 PM, Abdulla Bubshait <darkstego at gmail.com>
wrote:

> I don't think this is an identical MSRS spam, but it might be caused by
> the same underlying principal. I think the bottleneck here is the CPU,
> since Virtual CPUs are not exactly like CPUs.
>
> Take the Heroes MSRS spam. Each event is a register read in hardware, but
> since these are unique registers they will be emulated (if function works)
> as software in KVM. What I am guessing is happening here is the 3D code is
> running something that requires KVM to do things, which ends up being much
> slower than hardware.
>
> I would suggest trying to do something like CPU pinning to see if you can
> bump up performance a bit. But if this thing turns out to be a code path
> that isn't using hardware virtualization then you might never get the
> performance you are looking for.
>
> Maybe someone with more experience with the inner workings of
> virtualization could be more useful here.
>

This is where you probably want to post your analysis to the KVM mailing
list <kvm at vger.kernel.org>.  There are folks there with quite a bit more
knowledge of the low level processor details than me.  Device assignment
itself can almost entirely get out of the way, the limitations are
interrupt latency (which has hardware solutions in newer processors for
some cases) and device regions that are partially emulated or virtualized,
like device quirks or msi-x vector pages.  Actual processor virtualization
is material for the KVM list.  Thanks,

Alex
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listman.redhat.com/archives/vfio-users/attachments/20160518/85b3b2d8/attachment.htm>