Re: [x86] do_arch_prctl - bug?

[Eric Lacombe <goretux@gmail.com>]

Le mercredi 19 novembre 2008 02:07:23 Jeremy Fitzhardinge, vous avez écrit :
> Eric Lacombe wrote:
> > I look at the Intel docs (vol. 3A) again, and see that in 64 bits mode
> > the hidden field gs.base are physically mapped to the MSR, so it seems
> > that in order to load gs.base we don't need to load gs (like in 32 bits
> > mode), but rather we only need to load the MSR.
> >
> > So I don't understand the purpose of load_gs_index in that context :
> >
> > if (doit) {
> > 	load_gs_index(0);
> > 	ret = checking_wrmsrl(MSR_KERNEL_GS_BASE, addr);
> > }
> >
> > Why don't we only load the MSR ?
> > What is the purpose of calling load_gs_index with 0 as parameter ?
>
> Because %gs of 0 means "base too large, go to MSR".  If you have a
> 32-bit base, then loading it into the gdt and loading %gs with the right
> selector is faster.  wrmsr/rdmsr are slow instructions.

Ok, thanks, so I suppose now that only doing :
asm volatile("movl %0,%%gs" :: "r" (0));
could corrupt the address of the PDA that resides actually in the MSR_GS_BASE. 
And that's why load_gs_index is used as it contains "swapgs" before and after 
the "mov to gs".

Is that correct?

Regards,

	Eric