Re: [RFC][PATCH 2/3] perf: Add a bit of paranoia

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Feb 04, 2015 at 03:51:36PM +0100, Jiri Olsa wrote:
> On Mon, Feb 02, 2015 at 06:32:32PM +0100, Peter Zijlstra wrote:
> > > That looks like tail recursive fun! An irq work that raises and irq work
> > > ad infinitum. Lemme see if I can squash that.. didn't we have something
> > > like this before... /me goes look.
> > 
> > 
> > Does this make it go away?
> > 
> > --- a/kernel/events/core.c
> > +++ b/kernel/events/core.c
> > @@ -4413,6 +4413,8 @@ static void perf_pending_event(struct ir
> >  	struct perf_event *event = container_of(entry,
> >  			struct perf_event, pending);
> >  
> > +	int rctx = perf_swevent_get_recursion_context();
> > +
> 
> hum, you should check the rctx
> 
> 	if (rctx == -1)
> 		return;

D'uh, yes.

> also this recursion is bound to swevent_htable, should we rather add
> separate ctx data for irq_work to limit the clashing with SW events?

No, we explicitly want to disable software events while handling the
irq_work. The problem as reported looks like irq_work triggering a
swevent (tp actually, but that's classed the same) generates a new
irq_work, and we get stuck in an endless cycle of that.

So by effectively disabling swevents while processing the irq_work we
should break the cycle.