Re: Q: smp.c && barriers (Was: [PATCH 1/4] generic-smp: removesingle ipi fallback for smp_call_function_many())

[Suresh Siddha]

On Wed, 2009-02-18 at 09:14 -0800, Linus Torvalds wrote:
> 
> On Wed, 18 Feb 2009, Ingo Molnar wrote:
> >
> > an off-list comment pointed out this piece of information as 
> > well:
> > 
> >   http://www.sandpile.org/ia32/coherent.htm
> > 
> >   A WRMSR to one of the x2APIC MSRs (0000_0800h...0000_0BFFh) is 
> >   not guaranteed to be serializing.
> > 
> > So i suspect we should just enclose it in smp_mb() pairs to make 
> > sure it's a full barrier in both directions?
> 
> Why would we care about "both directions"?

on x86 we don't need in both directions.

> 
> I think putting an sfence _before_ the wrmsr (and not even all of them - 
> just put it in front of the "send IPI" sequence) should be fine. Any other 
> ordering sounds like just unnecessary overhead to me.

For x2apic ipi's, we should use a serializing instruction or a "mfence"
instruction. "sfence" will not help in this scenario.

thanks,
suresh

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/