[Qemu-devel] xsave instruction not passed through correctly on AMD hardware

[Qemu-devel] xsave instruction not passed through correctly on AMD hardware

[Owen Tuz]

[-- Attachment #1: Type: text/plain, Size: 1192 bytes --]

Hi all,

We've been seeing a problem lately running FreeBSD 9.1 and 9.2 (latest
stable) which causes the guest to crash during boot when QEMU is run on an
AMD processor with the 'xsave' flag set.

To reproduce this behaviour:

- Boot a FreeBSD 9.1 or 9.2 guest or even installation CD on an AMD
processor with xsave enabled. Use '-cpu host'.

- After the bootloader, the guest will crash almost immediately with the
message 'kernel trap 12 with interrupts disabled'. This occurs before any
disks are loaded, so it's not possible to get a memory dump from the guest
OS for backtrace.

- Boot again with '-cpu host,-xsave'. The guest should boot successfully.

This was seen on AMD Opteron 6238 processor family, and does not affect our
Opteron 6128s (due to lack of the xsave flag).

We've also tested on an Intel Xeon E5-2640 processor which has the xsave
flag set and verified that we do not see this behaviour.

Based on this, I believe that the xsave instruction is not being correctly
emulated on some hardware. Is this a known issue?

Thanks in advance for looking, and please let me know if we can provide any
more useful information to help diagnose/fix this.

Best regards,

Owen Tuz

[-- Attachment #2: Type: text/html, Size: 1526 bytes --]

Re: [Qemu-devel] xsave instruction not passed through correctly on AMD hardware

[Paolo Bonzini]

Il 11/11/2013 15:30, Owen Tuz ha scritto:
> Hi all,
> 
> We've been seeing a problem lately running FreeBSD 9.1 and 9.2 (latest
> stable) which causes the guest to crash during boot when QEMU is run on
> an AMD processor with the 'xsave' flag set. 
> 
> To reproduce this behaviour:
> 
> - Boot a FreeBSD 9.1 or 9.2 guest or even installation CD on an AMD
> processor with xsave enabled. Use '-cpu host'.
> 
> - After the bootloader, the guest will crash almost immediately with the
> message 'kernel trap 12 with interrupts disabled'. This occurs before
> any disks are loaded, so it's not possible to get a memory dump from the
> guest OS for backtrace.
> 
> - Boot again with '-cpu host,-xsave'. The guest should boot successfully.

This is probably cause by the "lightweight profiling" XSAVE extension.
KVM does not support it.  Using "-cpu host" may show problems when the
CPU has features that are not supported by KVM.

It's possible though that this has been recently fixed.  Please try
branch "next" of git://git.kernel.org/pub/scm/virt/kvm/kvm.git and
report back.

Paolo

Re: [Qemu-devel] xsave instruction not passed through correctly on AMD hardware

[Owen Tuz]

[-- Attachment #1: Type: text/plain, Size: 1445 bytes --]

Thanks, Paolo. We will test and let you know.
I'm not familiar with LWP (some reading to do there) - are there any plans
to support this in future, or is this just something that we're not
interested in emulating?

Best regards,

Owen


On Mon, Nov 11, 2013 at 3:25 PM, Paolo Bonzini <pbonzini@redhat.com> wrote:

> Il 11/11/2013 15:30, Owen Tuz ha scritto:
> > Hi all,
> >
> > We've been seeing a problem lately running FreeBSD 9.1 and 9.2 (latest
> > stable) which causes the guest to crash during boot when QEMU is run on
> > an AMD processor with the 'xsave' flag set.
> >
> > To reproduce this behaviour:
> >
> > - Boot a FreeBSD 9.1 or 9.2 guest or even installation CD on an AMD
> > processor with xsave enabled. Use '-cpu host'.
> >
> > - After the bootloader, the guest will crash almost immediately with the
> > message 'kernel trap 12 with interrupts disabled'. This occurs before
> > any disks are loaded, so it's not possible to get a memory dump from the
> > guest OS for backtrace.
> >
> > - Boot again with '-cpu host,-xsave'. The guest should boot successfully.
>
> This is probably cause by the "lightweight profiling" XSAVE extension.
> KVM does not support it.  Using "-cpu host" may show problems when the
> CPU has features that are not supported by KVM.
>
> It's possible though that this has been recently fixed.  Please try
> branch "next" of git://git.kernel.org/pub/scm/virt/kvm/kvm.git and
> report back.
>
> Paolo
>

[-- Attachment #2: Type: text/html, Size: 2189 bytes --]

Re: [Qemu-devel] xsave instruction not passed through correctly on AMD hardware

[Paolo Bonzini]

Il 11/11/2013 16:43, Owen Tuz ha scritto:
> Thanks, Paolo. We will test and let you know.
> I'm not familiar with LWP (some reading to do there) - are there any
> plans to support this in future, or is this just something that we're
> not interested in emulating?

Given Linux does not support LWP, and AMD is not really trying to
compete with Intel anymore, it is quite unlikely that KVM will support
LWP in the future.

Paolo

Re: [Qemu-devel] xsave instruction not passed through correctly on AMD hardware

[Owen Tuz]

[-- Attachment #1: Type: text/plain, Size: 660 bytes --]

Great, thanks for the clarification. Testing with kvm-next shows that this
is still true in the latest build.

Best regards,

Owen


On Mon, Nov 11, 2013 at 4:56 PM, Paolo Bonzini <pbonzini@redhat.com> wrote:

> Il 11/11/2013 16:43, Owen Tuz ha scritto:
> > Thanks, Paolo. We will test and let you know.
> > I'm not familiar with LWP (some reading to do there) - are there any
> > plans to support this in future, or is this just something that we're
> > not interested in emulating?
>
> Given Linux does not support LWP, and AMD is not really trying to
> compete with Intel anymore, it is quite unlikely that KVM will support
> LWP in the future.
>
> Paolo
>

[-- Attachment #2: Type: text/html, Size: 1130 bytes --]