Re: [patch 1/2] x86_64 page fault NMI-safe

[Jeremy Fitzhardinge <jeremy@goop.org>]

On 07/14/2010 03:37 PM, Linus Torvalds wrote:
> I think the %rip check should be pretty simple - exactly because there
> is only a single point where the race is open between that 'mov' and
> the 'iret'. So it's simpler than the (similar) thing we do for
> debug/nmi stack fixup for sysenter that has to check a range.
>
> The only worry is if that crazy paravirt code wants to paravirtualize
> the iretq. Afaik, paravirt does that exactly because they screw up
> iret handling themselves. Maybe we could stop doing that stupid iretq
> paravirtualization, and just tell the paravirt people to do the same
> thing I propose, and just allow nesting.
>   

We screw around with iret because there's a separate interrupt mask flag
which can't be set atomically with respect to a stack/ring change (well,
there's more to it, but I won't confuse matters).

But it only really matters if the PV guest can also get NMI-like
interrupts.  While Xen supports NMI for PV guests, we don't use it much
and I haven't looked into implementing support for it yet.

    J