Re: [PATCH V6 5/8] perf/x86/intel/uncore: add infrastructure for free running counter

[Peter Zijlstra]

On Thu, Jan 25, 2018 at 12:13:45PM -0800, kan.liang@xxxxxxxxx wrote:
> @@ -512,6 +538,17 @@ static int uncore_pmu_event_add(struct perf_event *event, int flags)
>  	if (!box)
>  		return -ENODEV;
>  
> +	/*
> +	 * The free funning counter is assigned in event_init().
> +	 * The free running counter event and free running counter
> +	 * are 1:1 mapped. It doesn't need to be tracked in event_list.
> +	 */
> +	if (uncore_pmc_freerunning(hwc->idx)) {
> +		if (flags & PERF_EF_START)
> +			uncore_pmu_event_start(event, 0);
> +		return 0;
> +	}
> +
>  	ret = n = uncore_collect_events(box, event, false);
>  	if (ret < 0)
>  		return ret;

This will result in some (probably) unexpected scheduling artifacts.
Probably the only way to really cure that is to have the free running
counters in their own PMU and not share with the GP counters of this
box.