Re: [PATCH/RFC] Math-emu kills the kernel on Athlon64 X2

[Randy Dunlap <rdunlap@xenotime.net>]

On Tue, 3 Oct 2006 09:37:45 -0700 (PDT) Linus Torvalds wrote:

> 
> 
> On Tue, 3 Oct 2006, Randy Dunlap wrote:
> > 
> > Yes, that works.
> 
> Ok. I'll commit that simple thing, and add a comment on why we're 
> apparently doing the same thing twice (you do need _both_ of those things: 
> the "disable_x86_fxsr" will make sure all other CPU's also get cleared, 
> while the "clear_bit()" will clear it immediately on the boot CPU)
> 
> I'll leave the no387/nofxsr linking alone for now. The main reason to use 
> no387 would seem to be just testing that emulation works at all, and I 
> guess we can just tell people to use the "no387 nofxsr" combination.
> 
> So Randy, with this you can boot all the way into user space, and some FP 
> apps still work too?

My few trivial float and double apps work.
Is there any particular test/workload that you want me to run?

> (Of course, user-space may be buggy and use SSE etc without testing for 
> whether the CPU actually supports it - if the install process has 
> installed some special SSE-version of a library depending on what the CPU 
> claimed at that point, or if somebody uses "cpuid" directly rather than 
> asking the kernel. So there's no way we're going to _guarantee_ that this 
> works in user space, but at least a well-behaved user-space that works on 
> a i486 should hopefully be ok).


---
~Randy