Re: [RFC][PATCH] PM: disable nonboot cpus before suspending devices

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

On Wed, 3 Feb 2010 23:34:37 +0100
"Rafael J. Wysocki" <rjw@sisk.pl> wrote:

> On Wednesday 03 February 2010, Andrew Morton wrote:
> > On Wed, 3 Feb 2010 02:44:23 +0100 "Rafael J. Wysocki" <rjw@sisk.pl> wrote:
> > 
> > > +static inline gfp_t clear_gfp_allowed_mask(gfp_t mask)
> > > +{
> > > +	gfp_t ret = gfp_allowed_mask;
> > > +	gfp_allowed_mask &= ~mask;
> > > +	return ret;
> > > +}
> > 
> > Fair enuf.
> > 
> > Of course, this is all horridly racy/buggy without locking.  Would I be
> > correct in hoping that all the callers happen when the system is in
> > everyone-is-frozen mode?
> 
> As far as I can tell, gfp_allowed_mask is only touched during init apart from
> this.

Well yes - the new interfaces are the problem - they're racy!

> > Perhaps we should add some documentation (or even an assertion) to
> > prevent someone from using these interfaces from within normal code.
> 
> I thought about that, but didn't invent anything smart enough.
> 
> Well, maybe except for a comment like "this must be called with pm_mutex held",
> because that's the only case when it would be really safe.

Is that the locking rule?  My above guess was incorrect?

Maybe slip a

	BUG_ON(!mutex_is_locked(&pm_mutex));

in there?