Possible APIC conflict with 2.6.38.8 Kernel

Possible APIC conflict with 2.6.38.8 Kernel

[Simon]

We're building a virtual windows 7 x64 model on top of a custom Linux OS.
We're currently using 2.6.38.8 kernel.
I was rather attracted to the new kvm-apic/qemu-kvm merge you've currently
made available, as this may hold the answer to a stubborn problem; 
every time we load or attempt to create a windows 7 image while 
our RTAI(real time) base is running, at an early stage of loading 
windows the real time breaks and stops dead.

We suspect this is due to APIC conflicts between RTAI and Qemu-kvm.

I've read with interest your internal debate on making kvm-apic options
available, and I wonder if these now exist and if so, what they are
(the online help may not have kept pace).
Note that these are 64-bit OS's. With older kernels and 0.13.0 
with 32 bit XP, we didn't have this problem.

Re: Possible APIC conflict with 2.6.38.8 Kernel

[Jan Kiszka]

On 2012-02-07 16:25, Simon wrote:
> We're building a virtual windows 7 x64 model on top of a custom Linux OS.
> We're currently using 2.6.38.8 kernel.
> I was rather attracted to the new kvm-apic/qemu-kvm merge you've currently
> made available, as this may hold the answer to a stubborn problem; 
> every time we load or attempt to create a windows 7 image while 
> our RTAI(real time) base is running, at an early stage of loading 
> windows the real time breaks and stops dead.
> 
> We suspect this is due to APIC conflicts between RTAI and Qemu-kvm.
> 
> I've read with interest your internal debate on making kvm-apic options
> available, and I wonder if these now exist and if so, what they are
> (the online help may not have kept pace).
> Note that these are 64-bit OS's. With older kernels and 0.13.0 
> with 32 bit XP, we didn't have this problem.

That's Russian roulette. You can't simply use KVM (or any other
virtualization subsystem) on an I-pipe-patched host kernel. See [1] for
the requirements on I-pipe (will be contained a bit differently in
upcoming 3.x patches for x86) and [2] for the real-time extension on
top. We implemented it for Xenomai, but RTAI should be patchable as well
- in principle.

Note that there is potentially a FPU-related guest corruption with these
patches remaining, but I wasn't able to reproduce it yet. May depend on
host CPU features.

Just to clarify: this is not really a topic of KVM or upstream Linux -
or even QEMU. The I-pipe patch changes the semantics of low-level
interrupt masking in Linux and opens new context switching points. Thus,
it is responsible for making the KVM subsystem aware of this and for
coordination with the real-time extension.

Jan

[1] http://thread.gmane.org/gmane.linux.kernel.adeos.general/1898
[2]
http://git.xenomai.org/?p=xenomai-2.6.git;a=commitdiff;h=35505fe10c0dae3d1eab8a409b978d338b84b89e

-- 
Siemens AG, Corporate Technology, CT T DE IT 1
Corporate Competence Center Embedded Linux