[vfio-users] Win 10 VM often freezes Linux when shutting down

[Alex Williamson]

On Wed, 17 Feb 2016 22:41:53 +0100
Jens Zimmermann <zimmermannjens888 at gmail.com> wrote:

> Hello guys!
> 
> I have a Windows 10 VM with GPU passsthrough working for around 2 months
> now and while there were no problems in the very beginning as far as I can
> tell, soon after the VM started to freeze my Linux sometimes after I shut
> it down.
> The strange thing is that it does not happen every time. It seems like it
> usually happens when the VM was running for quite some time like an hour or
> two.
> I thought it may be the Windows shut down which is by default some kind of
> hibernate, so I shut down Windows by using the command "shutdown /s /t0",
> but that did not help. I also tried force stopping the VM, that led to a
> freeze as well.
> 
> I presume the error is caused by Linux not being able to integrate the
> virtual/passed through hardware into the system again somehow. Perhaps it's
> the memory since the freezes usually happen when the VM was running for
> some time, but this is just a wild guess, I am no linux expert at all.
> 
> This is journalctl of the most recent freeze, VM was started at 20:13 and
> shut down at 21:09, followed by a freeze: http://pastebin.com/mwQsFsKB
> In comparison this time it did not freeze, start at 21:33, finishes at
> 21:55: http://pastebin.com/YGPnVPBD
> 
> My setup: i7-5820k, ASRock x99 Extreme 4, 16GB RAM, nvidia 210 & 750 Ti,
> two monitors, mice and keyboards, Manjaro 15.12 with kernel 4.1, qemu
> 2.4.1-2, libvirt 1.3.1-2, virt-manager 1.3.2-2
> qemu xml: http://pastebin.com/WPupqDEv
> 
> I would be glad to provide you more information, just tell me what you want
> and how I can get it.
> 
> I really appreciate any suggestion, I am out of ideas now.

I'd be suspicious of any devices that you're assigning that gets
re-bound to host drivers when the VM is shutdown.  I see from the XML
that you're not only assigning the GPU audio, but 00:1b.0 on the system
as well, which I'd guess is the onboard audio.  You can bind all the
assigned devices with virsh nodedev-detach and then set managed='no' in
the xml to make sure they don't get re-attached to host drivers.