Re: [PATCH 6/6] percpu: Add preemption checks to __this_cpu ops

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Oct 16, 2013 at 12:52:38PM -0400, Steven Rostedt wrote:
> On Wed, 16 Oct 2013 18:25:37 +0200
> Peter Zijlstra <peterz@infradead.org> wrote:
>  
> > But yes, the way its set-up an arch could indeed provide __this_cup_$op
> > itself -- without providing the _$n variants; in which case the
> > raw_cpu_$op provided by you is broken.
> > 
> > Can't we have a 'simple' coccinelle script rename the entire __this_cpu*
> > implementation over to raw_cpu* and then provide generic __this_cpu* ->
> > raw_cpu maps?
> > 
> 
> Perhaps we should match the way spinlocks are.
> 
> 
> this_cpu*() be the normal use.
> 
> raw_this_cpu() could perhaps not do the checks?
> 
> arch_this_cpu() be the architecture specific version of this_cpu*

In that case we'd need to do something like:

 this_cpu_$op   -> this_cpu_$op_irq (disables irqs itself)
 __this_cpu_$op -> this_cpu_$op     (with check)
 		-> raw_cpu_$op      (without the check)

I don't think the arch bits feature heavily for percpu; normally an arch
provides __this_cpu_$op_$n; raw_cpu_$op_$n in my latest proposal.


Anyway; I don't think the spinlock pattern matches too good and I don't
mind the proposed:

  this_cpu_$op    (disables IRQs itself)
  __this_cpu_$op  (with preemption check)
  raw_cpu_$op	  (without preemption check)