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

[Randy Dunlap <rdunlap@xenotime.net>]

On Mon, 2 Oct 2006 19:34:27 -0700 (PDT) Linus Torvalds wrote:

> 
> 
> On Mon, 2 Oct 2006, Randy Dunlap wrote:
> > 
> > I had no trouble reproducing the boot failure (on Pentium-M), then
> > I tried TRACE_RESUME().  Nifty, but not really needed here since
> > earlyprintk worked and contained the fault messages:
> > 
> > [   16.841784] math_emulate: 0060:c01062dd
> > [   16.845579] Kernel panic - not syncing: Math emulation needed in kernel
> > 
> > But CONFIG_MATH_EMULATION=y, so what now?
> 
> The "Math emulation needed in kernel" message means that it was asked to 
> emulate a kernel instruction, and it refuses to do so. The emulation is 
> _not_ meant to be a real FPU, it simply looks like one to user space. A 
> lot of things aren't really emulated (there's no global x87 context, for 
> example: the context is all strictly per-process).
> 
> > Linus mentioned CPU feature bits.  The message log above didn't
> > make me feel good about them.  Sure enough, we are playing with
> > features before reading the feature bits.
> 
> Please look up address c01062dd in the system map (or just using gdb), 
> that will tell you what code _tried_ to use the math coprocessor in kernel 
> space.

Sure, it's in arch/i386/kernel/i387.c::mxcsr_feature_mask_init(),
on the fxsave instruction:

void mxcsr_feature_mask_init(void)
{
	unsigned long mask = 0;
	clts();
	if (cpu_has_fxsr) {
		memset(&current->thread.i387.fxsave, 0, sizeof(struct i387_fxsave_struct));
		asm volatile("fxsave %0" : : "m" (current->thread.i387.fxsave)); 
		mask = current->thread.i387.fxsave.mxcsr_mask;
		if (mask == 0) mask = 0x0000ffbf;
	}

---
~Randy