Re: [RFC][PATCH 0/2] preempt: Track unmatched preempt_enable() to preempt_disable()

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Jan 16, 2014 at 01:39:06PM -0500, Steven Rostedt wrote:
> On Thu, 16 Jan 2014 18:45:36 +0100
> Peter Zijlstra <peterz@infradead.org> wrote:
> 
> > On Thu, Jan 16, 2014 at 12:18:24PM -0500, Steven Rostedt wrote:
> > 
> > Didn't see anything horrible in there, so fine with me.
> > 
> > Another thing I recently ran into was:
> > 
> >   local_irq_disable();
> >   preempt_disable();
> >   local_irq_enable();
> > 
> >   local_irq_disable();
> >   preempt_enable();
> >   local_irq_enable();
> > 
> > I'm not sure any of our debugging catches that, I meant to extend
> > lockdep, but maybe it could also be hooked into something else.
> 
> Hmm, what's wrong with the above?

Suppose a reschedule interrupt happens in the middle; we set
NEED_RESCHED but won't act because preempt is disabled.

When we hit preempt_enable() we'll call into preempt_schedule() but fail
because preemptible() finds irqs_disabled().

And in the end we do not schedule, but have NEED_RESCHED set.

FAIL.