On Mon, Dec 19, 2005 at 09:34:01AM +1100, Paul Ripke wrote:
> OK, I've had a closer look at this, and come up with the included patch,
> which appears to work fine for me, with threaded and non-threaded 
> programs.
> I can't repeat the behaviour of the FPU state being not being restored,
> restored to the wrong thread, or wiped.

This is great, many thanks ! I commited the attached patch, which
calls HYPERVISOR_fpu_taskswitch() at the end of npxsave_lwp() instead of the
various places in which call npxsave_lwp() in the code.

> 
> Thoughts behind the patch:
> - since clts() is a no-op under xen 2 (in xen 3, 
> HYPERVISOR_fpu_taskswitch
>   takes an int parameter allowing set/clear), it makes no sense to 
> stts()
>   inside the DNA handler. We can't clear it again.
> - I don't get the npxsave_lwp() calls in the DNA handler. On SMP, it's
>   used to dump an LWP's FPU state from a different CPU. On a 
> uniprocessor,
>   the FPU state has already been dumped by npxsave_cpu(). SMP support 
> isn't
>   relevant for xen 2.

Sure, but as it's #ifdef MULTIPROCESSOR, there's no problem keeping
it here. The goal is to eventually be able to share more code with i386.

-- 
Manuel Bouyer <bouyer@antioche.eu.org>
     NetBSD: 26 ans d'experience feront toujours la difference
--