[vfio-users] What makes or breaks proper unbinding of vfio-pci?

[Mario Goebbels]

Hi,

I'm currently trying to figure weird behaviour with vfio-pci. When I fire
up my Windows VM with VGA passthrough, I seem to be able to boot and shut
it down just fine, binding and unbinding vfio-pci without issues, so long I
just do simple things with the GPU inside the VM.

Whenever I run a game inside the VM, after shutting it down, unbinding
vfio-pci usually ends up in a kernel panic, either saying...

[  311.157926] BUG: unable to handle kernel paging request at
00007f23c8b47000

-or-

[ 2782.316934] kernel tried to execute NX-protected page - exploit attempt?
(uid: 0)

It's pretty easy to reproduce, pretty much as described above. I start my
VM with Windows 10, run say GTA V for just a minute, shut it down
immediately, and when the helper script unbinds vfio-pci, it panics.
However, if I start the VM, do some simpler things, like browsing the web,
watching some video with hardware accelerated decoding, and even run some
WebGL demos (that appear hardware accelerated, I suppose similar to a
game), and shut it down again, everything unbinds cleanly. And I can
repeatedly start the VM without needing to reboot in-between or other
tricks, as long I don't run a fullblown game.

Any idea what's going on and how I can fix this?

The system is an Intel Core i7-5820K, an Asrock X99 Extreme4 mainboard and
the card being passed through is an NVidia GTX780 from EVGA. This happens
on both Linux 4.4 and 4.5-rc1.

My current version of the helper script for reference:
http://pastebin.com/kz79Cppv

Thanks for any help and ideas,
-mg
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listman.redhat.com/archives/vfio-users/attachments/20160125/5328d471/attachment.htm>