Re: [PATCH 2/3] x86: Remove cmpxchg from i386 NMI nesting code

[Steven Rostedt <rostedt@goodmis.org>]

On Fri, 2012-06-08 at 10:28 -0700, H. Peter Anvin wrote:
> On 06/08/2012 09:41 AM, Steven Rostedt wrote:
>  
> >>  
> >>  #define nmi_nesting_postprocess()                                      \
> >>         do {                                                            \
> >> -               if (cmpxchg(&__get_cpu_var(nmi_state),                  \
> >> -                   NMI_EXECUTING, NMI_NOT_RUNNING) != NMI_EXECUTING)   \
> >> +               if (!local_dec_and_test(&__get_cpu_var(nmi_state)))     \
> > 
> > Now this is where I think the patch helps. I'm almost certain that
> > local_dec_and_test is faster than a cmpxchg by many cycles. Especially
> > on i386.
> > 
> 
> On i386 it's infinite, but again, I don't think the code will ever be
> exercised on i386.  I'm much more concerned about performance on current
> processors.

I'm not worried about non i386 here. Some context missing from the patch
is that this code is surrounded by:

#ifdef CONFIG_X86_32

[...]

#else

The '#else' part does something completely different, as this work is
done in the assembly handler on x86_64.

And yes, this code is executed on i386 (well when I boot my x86_64 into
an i386 kernel it does).


> 
> But yes, local_dec_and_test should at least not be more expensive.  Even
> better, use this_cpu_dec_return().

Oh! That would make this much better! I'll update the patch set.

Thanks!

-- Steve