Re: + ignore-stolen-time-in-the-softlockup-watchdog-fix.patch added to -mm tree

[Andrew Morton <akpm@linux-foundation.org>]

On Tue, 24 Apr 2007 21:37:47 -0700 Jeremy Fitzhardinge <jeremy@goop.org> wrote:

> Andrew Morton wrote:
> > Any which enable CONFIG_DEBUG_PREEMPT...
> >   
> 
> Hm.  Normally smp_processor_id() CSEs nicely, so it isn't generally
> necessary to do it manually.

For all architectures?

>  Not sure what CONFIG_DEBUG_PREEMPT will do
> to it.

#ifdef CONFIG_DEBUG_PREEMPT
  extern unsigned int debug_smp_processor_id(void);
# define smp_processor_id() debug_smp_processor_id()
#else
# define smp_processor_id() raw_smp_processor_id()
#endif

We could perhaps mark debug_smp_processor_id() as pure or const or whatever
the trick is to say it doesn't need to be reevaluated.  That would usually
be OK, because as long as preempt is disabled that _is_ how it works,
however code might do:

	preempt_disable();
	smp_processor_id();
	preempt_enable();

	<preempt or schedule()>

	preempt_disable();
	smp_processor_id();
	preempt_enable();

and caching the value would be wrong.

hm.  If "smp_processor_id() CSEs nicely", what prevents the compiler from
making the same mistake?  IOW, what are the barriers for CSE?  Function
calls?