[vfio-users] cpu usage in guest != cpu usage in host, even with exclusive pinning

[Jon Panozzo]

Can confirm 4.2.8 kernel doesn’t exhibit this behavior with the same QEMU package and VM guest.  So taking the exact same software configuration to the 4.2.8 kernel results in normal CPU usage, but with any 4.4.x kernel variant I’ve tried thus far, I’ve seen the same 100% usage.

It seems to apply when any media application is in use (3D, audio, or video).  Just playing an MP3 for me caused the host to go to 100%.  And if I leave shadowplay running, it stays at 100% as well.  Any application that pulls on a VFIO passed-through device seems to cause this to occur for me.

> On Mar 17, 2016, at 3:55 PM, Jon Panozzo <jonp at lime-technology.com> wrote:
> 
> Also experiencing this with Linux kernel 4.4.5 and QEMU 2.5.0.  Odd behavior indeed.
> 
>> On Feb 14, 2016, at 9:34 PM, globalgorrilla at fastmail.fm wrote:
>> 
>> Still the same with 4.5-rc4 and AUR qemu-git.
>> 
>> On 23 Jan 2016, at 18:55, Dan Ziemba wrote:
>> 
>>> On Tue, 2016-01-19 at 16:52 +0100, Friedrich Oslage wrote:
>>>> I did some more testing and it turns out to be a kernel error,
>>>> introduced somewhere around linux 4.3.
>>>> 
>>>> Short description of the error: 100% cpu usage while using 3d in a vm
>>>> with vfio
>>>> 
>>>> My last known good kernel is 4.2.8, which gives me about 20% cpu
>>>> usage
>>>> while playing Diablo III. My first known bad kernel is 4.3.3, which
>>>> gives me 100% cpu usage with every game, even while idle-ling in the
>>>> menu.
>>>> 
>>>> On 01/17/2016 10:29 AM, Friedrich Oslage wrote:
>>>>> my host system is linux-4.4.0 with qemu-2.5.0 and a 4-core i7.
>>>>> Linux is
>>>>> booted with isolcpus=1-3,5-7 to reserve 3 cores + threads for the
>>>>> Windows 10 VM.
>>>>> The VM's 3 cores(2 threads each) are pinned to the respective
>>>>> physical
>>>>> core/thread. The iothread is pinned to 1-3,5-7.
>>> 
>>> Interesting, I've also noticed similar behavior since upgrading to
>>> kernel 4.4 and qemu 2.5.0.  I am running 4 cores / 8 threads for the VM
>>> and pinned to individual host threads by libvirt, but I didn't boot
>>> with the isolcpus arg. Emulator threads are pinned to the remaining 4
>>> host threads.
>>> 
>>> Host is a i7-5930K, so 6 cores.  Guest is running Win 8.1.
>>> 
>>> Just sitting at the pause screen in Fallout: New Vegas, I have a single
>>> CPU running at 100% according to both the windows task manager and the
>>> host.  The host cpu thread that is at 100% is one of the ones with a
>>> vcpu pinned to it.  In the guest, it does appear that the game is the
>>> thing using all the CPU cycles, and the GPU is basically idle at this
>>> time.
>>> 
>>> When I was previously running kernel 4.1.15 and qemu 2.4, New Vegas was
>>> never demanding enough to max out any one cpu no matter what I was
>>> doing with the game.  Same goes for more demanding games such as GTA V.
>>> 
>>> 
>>> I upgraded both kernel and qemu at the same time, so I can't say which
>>> caused it.  I haven't really noticed much difference in performance. If
>>> anything, it might be slightly better.  I seemed to notice less frame
>>> rate drops and stuttering while playing GTA V after the upgrade.
>>> 
>>> Dan
>>> 
>>> _______________________________________________
>>> vfio-users mailing list
>>> vfio-users at redhat.com
>>> https://www.redhat.com/mailman/listinfo/vfio-users
>> 
>> _______________________________________________
>> vfio-users mailing list
>> vfio-users at redhat.com
>> https://www.redhat.com/mailman/listinfo/vfio-users
>