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

[Steven Rostedt <rostedt@goodmis.org>]

On Sun, 2010-07-18 at 11:17 -0700, Linus Torvalds wrote:

> Oh, and the non-nested case obviously needs all the regular "make the
> kernel state look right" code. Like the swapgs stuff etc if required.
> My example code was really meant to just document the nesting
> handling, not the existing stuff we already need to do with
> save_paranoid etc.
> 
> And I really think it should work, but I'd again like to stress that
> it's just a RFD code sequence with no testing what-so-ever etc.
> 

Are you sure you don't want to use Mathieu's 2/2 patch? We are fixing
the x86 problem that iret re-enables NMIs, and you don't want to touch
anything else but the NMI code. But it may be saner to just fix the
places that call iret. We can perhaps encapsulate those into a single
macro that we can get right and will be correct everywhere it is used.

The ugliest part of Mathieu's code is dealing with paravirt, but
paravirt is ugly to begin with.

Doing this prevents nested NMIs as well as all the unknowns that will
come with dealing with nested NMIs. Where as, handling all iret's should
be straight forward, although a bit more intrusive than what we would
like.

Just saying,

-- Steve