Re: [PATCH v2 1/1] perf: Sharing PMU counters across compatible events

[Jiri Olsa]

On Tue, Sep 11, 2018 at 01:29:32PM +0000, Song Liu wrote:

SNIP

> >>> 
> >>> jirka
> >> 
> >> I am not sure I am following. The pmu is disabled when we call
> >> event_pmu_add(). Why do we need to read before calling pmu->add()? 
> >> And this is the first added dup event for this master, so we don't
> >> need to worry about others. 
> >> 
> >> Does this make sense? 
> > 
> > I was just thinking since the pmu is disable we could
> > we don't need to read the event on 2 places.. it's almost
> > identic code
> 
> How about something like:
> 
> 
> +/* PMU sharing aware version of event->pmu->add() */
> +static int event_pmu_add(struct perf_event *event,
> +			 struct perf_event_context *ctx)
> +{
> +	struct perf_event_dup *dup;
> +	int ret;
> +
> +	/* no sharing, just do event->pmu->add() */
> +	if (event->dup_id == -1)
> +		return event->pmu->add(event, PERF_EF_START);
> +
> +	dup = &ctx->dup_events[event->dup_id];
> +
> +	if (dup->active_event_count = 0) {
> +		/* try add master */
> +		ret = event->pmu->add(dup->master, PERF_EF_START);
> +		if (ret)
> +			return ret;
> +	}
> +
> +	dup->active_event_count++;
> +	event->pmu->read(dup->master);
> +	event->dup_base_count = dup_read_count(dup);
> +	event->dup_base_child_count = dup_read_child_count(dup);
> +	
> +	return 0;
> +}

yep, seems ok

jirka