[PATCH v2] ARM: implement optimized percpu variable access

[Nicolas Pitre]

On Thu, 29 Nov 2012, Russell King - ARM Linux wrote:

> On Thu, Nov 29, 2012 at 10:45:50AM -0500, Nicolas Pitre wrote:
> > On Thu, 29 Nov 2012, Rob Herring wrote:
> > > Signed-off-by: Rob Herring <rob.herring at calxeda.com>
> > > Acked-by: Will Deacon <will.deacon at arm.com>
> > 
> > With the above, and moving the call to cpu_init() after the call to 
> > cpu_switch_mm(mm->pgd, mm) to fix Will's concerns (personally I'd put it 
> > right after local_flush_tlb_all())... 
> 
> You're confused.  We were suggesting before the printk().
> 
> The reasoning is: printk() is not guaranteed not to access per-cpu
> variables, so it needs to be before the first printk.  It can't be
> before cpu_switch_mm(), and putting it before the TLB flush does
> _not_ guarantee that TLB/MMU isn't going to still be seeing the
> strongly-ordered attribute - so it _must_ be after the TLB flush.

But isn't that what I wrote above?  I said "personally I'd put it right 
after local_flush_tlb_all()".

> As for the setup of the active MM, that's something that I still
> think should come as early as possible (even before the per-cpu
> stuff) because that is getting everything properly initialized and
> setup for this new thread; there's a risk that a fault occuring
> before that point may cause issues, specially if active_mm were NULL.

Good point.

> So, between cpumask_set_cpu() and printk() please.

Agreed.


Nicolas