lkml.org 
[lkml]   [2010]   [Jul]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [patch 1/2] x86_64 page fault NMI-safe
* Linus Torvalds (torvalds@linux-foundation.org) wrote:
> On Thu, Jul 15, 2010 at 3:01 PM, Mathieu Desnoyers
> <mathieu.desnoyers@efficios.com> wrote:
> >
> > . NMI exit code
> > and fake NMI entry are made reentrant with respect to NMI handler interruption
> > by testing, at the very beginning of the NMI handler, if a NMI is nested over
> > the whole nmi_atomic .. nmi_atomic_end code region.
>
> That is totally bogus. The NMI can be nested by exceptions and
> function calls - the whole _point_ of this thing. So testing "rip" for
> anything else than the specific final "iret" is meaningless. You will
> be in an NMI region regardless of what rip is.

There are 2 tests done on NMI handler entry:

1) test if nested over nmi_atomic region (which is a very restrained region
around nmi_exit, which does not do any function call nor take traps).
2) test if the per-cpu nmi_nesting flag is set.

Test #2 takes care of NMIs nested over functions called and traps.

>
> > This code assumes NMIs have a separate stack.
>
> It also needs to be made per-cpu (and the flags be per-cpu).

Sure, that was implied ;)

>
> Then you could in fact possibly test the stack pointer for whether it
> is in the NMI stack area, and use the value of %rsp itself as the
> flag. So you could avoid the flag entirely. Because testing %rsp is
> valid - testing %rip is not.

That could be used as a way to detect "nesting over NMI", but I'm not entirely
sure it would deal with the "we need a fake NMI" flag set/clear (more or less
equivalent to setting CS to 0 in your implementation and then back to some other
value). The "set" is done with NMIs disabled, but the "clear" is done at fake
NMI entry, where NMIs are active.

>
> That would also avoid the race, because %rsp (as a flag) now gets
> cleared atomically by the "iret". So that might actually solve things.

Well, I'm still unconvinced there is anything to solve, as I built my NMI entry
with 2 tests: one for "nmi_atomic" code range and the other for per-cpu nesting
flag. Given that I set/clear the per-cpu nesting flag either with NMIs off or
within the nmi_atomic code range, this should all work fine.

Unless I am missing something else ?

Thanks,

Mathieu

--
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com


\
 
 \ /
  Last update: 2010-07-16 00:33    [from the cache]
©2003-2011 Jasper Spaans