Xen guest floating point problem

Xen guest floating point problem

[Mick Jordan]

We have a simple C test program that runs on our evolved version of 
Mini-OS (GUK) that exhibits sporadic non-deterministic failure regarding 
floating point computations (mostly calling the __ieee754_remainder 
method of fdlibm). Typically about 5 computations in 10000 fail, usually 
with Nan or Infinite values showing up. We are as convinced as we can be 
at this point that the GUK code is not corrupting the floating point 
state, so are down to two ideas. One is a bug in Xen on guest context 
switching or event delivery regarding the FP state. The other is that 
GUK is failing to obey some Xen API contract regarding FP state. For 
example, timer and xenstore events are being delivered during the run 
although we can't correlate them directly to the fault.

The program fails on both Solaris xVM 3.1.4, xVM 3.3, various Linux 
3.1.4 systems, but does not fail when run on Solaris or Linux domUs. 
This points the finger at GUK but we are wondering whether they have a 
workaround for some problem we are not aware of.

Thanks for any help.

Mick

Re: Xen guest floating point problem

[Keir Fraser]

On 04/08/2009 00:51, "Mick Jordan" <Mick.Jordan@sun.com> wrote:

> We have a simple C test program that runs on our evolved version of
> Mini-OS (GUK) that exhibits sporadic non-deterministic failure regarding
> floating point computations (mostly calling the __ieee754_remainder
> method of fdlibm). Typically about 5 computations in 10000 fail, usually
> with Nan or Infinite values showing up. We are as convinced as we can be
> at this point that the GUK code is not corrupting the floating point
> state, so are down to two ideas. One is a bug in Xen on guest context
> switching or event delivery regarding the FP state. The other is that
> GUK is failing to obey some Xen API contract regarding FP state. For
> example, timer and xenstore events are being delivered during the run
> although we can't correlate them directly to the fault.
> 
> The program fails on both Solaris xVM 3.1.4, xVM 3.3, various Linux
> 3.1.4 systems, but does not fail when run on Solaris or Linux domUs.
> This points the finger at GUK but we are wondering whether they have a
> workaround for some problem we are not aware of.

I don't think the guest should have to do anything special regarding
handling FP state. Xen should automatically save/restore the state as
necessary. We had bugs in this area long ago, but they were fixed and I'd be
surprised if problems had crept back in.

 -- Keir