Re: [patch] x86: Rendezvous all the cpu's for MTRR/PAT init

[Ingo Molnar <mingo@elte.hu>]

* Nick Piggin <npiggin@suse.de> wrote:

> On Tue, Aug 18, 2009 at 11:20:57PM -0700, Suresh B wrote:
> > To make it clean I can move the smp_store_cpu_info() call before
> > local_irq_disable() in smp_callin(). But that needs more changes (for
> > xen etc). So thinking more, I think it is safe to do smp_call_function()
> > with interrupts disabled as the caller is currently not in the
> > cpu_online_mask.
> > 
> > i.e., no one else sends smp_call_function interrupt to this AP who is
> > doing smp_call_function() with interrupts disabled and as such there
> > won't be any deadlocks typically associated with calling
> > smp_call_function() with interrupts disabled. Copied Nick to confirm or
> > correct my understanding.
> > 
> > New patch appended removes this irq enable/disable sequence around
> > mtrr_ap_init() and add's a cpu_online() check in smp_call_function
> > warn-on's.
> 
> Yes this seems like a fine idea to me. Maybe also add a
> WARN_ON(cpu_online) in the interrupt-side as well just to
> make it clear.
> 
> If you split the patch out with its own changelog and give
> a comment for the special case, then you can add an
> Acked-by: Nick Piggin <npiggin@suse.de>
> 
> Although until you get acks from all arch maintainers, the 
> functionality would have to only be used on a per-arch basis but 
> that's probably OK as it's a pretty tricky thing for generic code 
> to be doing :)

Also, Suresh, please generate patches with diffstat included so 
that the arch impact can be deducted at a glance.

	Ingo