Re: [PATCH 2/2] kernel/events: Add option to enable counting sideband events in wakeup_events

[Naveen N. Rao]

On 2017/07/12 12:48PM, Jiri Olsa wrote:
> On Mon, Jun 19, 2017 at 08:01:08PM +0530, Naveen N. Rao wrote:
> 
> SNIP
> 
> > -	if (!event->attr.watermark) {
> > +	if (!event->attr.count_sb_events && !event->attr.watermark) {
> >  		int wakeup_events = event->attr.wakeup_events;
> >  
> >  		if (wakeup_events) {
> > diff --git a/kernel/events/ring_buffer.c b/kernel/events/ring_buffer.c
> > index 4e7c728569a8..f43a6081141f 100644
> > --- a/kernel/events/ring_buffer.c
> > +++ b/kernel/events/ring_buffer.c
> > @@ -197,6 +197,19 @@ __perf_output_begin(struct perf_output_handle *handle,
> >  	 * none of the data stores below can be lifted up by the compiler.
> >  	 */
> >  
> > +	if (event->attr.count_sb_events && !event->attr.watermark) {
> > +		int wakeup_events = event->attr.wakeup_events;
> > +
> > +		if (wakeup_events) {
> > +			int events = local_inc_return(&rb->events);
> > +
> > +			if (events >= wakeup_events) {
> > +				local_sub(wakeup_events, &rb->events);
> > +				local_inc(&rb->wakeup);
> > +			}
> > +		}
> > +	}
> 
> hum, so there's the same wakeup code in perf_output_sample,
> but it's not called for non-sample (sideband) events

Good point.

> 
> it'd be nice to have this wakeup only once in __perf_output_begin,
> but it'd change the behaviour for sideband events, which would start to
> follow the wakeup_events logic.. and possibly disturb Vince's tests?

I suppose you meant 'change the behaviour for _sample_ events'. Yes, the 
problem with having the check here is that it will now start firing for 
all events, rather than just the sampling events. I am not sure if we 
can differentiate sample events from sideband events in 
__perf_output_begin(). The other aspect is also Peter's concerns around 
performance.

Thanks,
Naveen