Re: [PATCH V7 08/14] perf/x86/intel: Generic support for hardware TopDown metrics

["Liang, Kan" <kan.liang@linux.intel.com>]

On 7/24/2020 11:27 AM, peterz@infradead.org wrote:
> On Fri, Jul 24, 2020 at 03:19:06PM +0200, peterz@infradead.org wrote:
>> On Thu, Jul 23, 2020 at 10:11:11AM -0700, kan.liang@linux.intel.com wrote:
>>> @@ -3375,6 +3428,72 @@ static int intel_pmu_hw_config(struct perf_event *event)
>>>   	if (event->attr.type != PERF_TYPE_RAW)
>>>   		return 0;
>>>   
>>> +	/*
>>> +	 * Config Topdown slots and metric events
>>> +	 *
>>> +	 * The slots event on Fixed Counter 3 can support sampling,
>>> +	 * which will be handled normally in x86_perf_event_update().
>>> +	 *
>>> +	 * The metric events don't support sampling.
>>> +	 *
>>> +	 * For counting, topdown slots and metric events will be
>>> +	 * handled specially for event update.
>>> +	 * A flag PERF_X86_EVENT_TOPDOWN is applied for the case.
>>> +	 */
>>> +	if (x86_pmu.intel_cap.perf_metrics && is_topdown_event(event)) {
>>> +		if (is_metric_event(event)) {
>>> +			struct perf_event *leader = event->group_leader;
>>> +			struct perf_event *sibling;
>>> +
>>> +			/* The metric events don't support sampling. */
>>> +			if (is_sampling_event(event))
>>> +				return -EINVAL;
>>> +
>>> +			/* The metric events cannot be a group leader. */
>>> +			if (leader == event)
>>> +				return -EINVAL;
>>> +
>>> +			/*
>>> +			 * The slots event cannot be the leader of a topdown
>>> +			 * sample-read group, e.g., {slots, topdown-retiring}:S
>>> +			 */
>>> +			if (is_slots_event(leader) && is_sampling_event(leader))
>>> +				return -EINVAL;
>>
>> This has nothing to do with sample-read; SLOTS cannot be sampling when
>> coupled with the METRIC stuff because hardware is daft.
>>
>> And you can have SAMPLE_READ on non-leader events just fine.
>>
>>> +
>>> +			/*
>>> +			 * The slots event must be before the metric events,
>>> +			 * because we only update the values of a topdown
>>> +			 * group once with the slots event.
>>> +			 */
>>> +			if (!is_slots_event(leader)) {
>>> +				for_each_sibling_event(sibling, leader) {
>>> +					if (is_slots_event(sibling))
>>> +						break;
>>> +					if (is_metric_event(sibling))
>>> +						return -EINVAL;
>>> +				}
>>> +			}
>>
>> Per the SIBLING patch this then wants to be:
>>
>> 			if (!is_slots_event(leader))
>> 				return -EINVAL;
>>
>> 			event->event_caps |= PERF_EV_CAP_SIBLING.
>> 			/*
>> 			 * Only once we have a METRICs sibling to we
>> 			 * need TopDown magic.
>> 			 */
>> 			leader->hw.flags |= PERF_X86_EVENT_TOPDOWN;
>>> +		}
>>> +
>>> +		if (!is_sampling_event(event)) {
>>> +			if (event->attr.config1 != 0)
>>> +				return -EINVAL;
>>
>> How does this depend on sampling?
>>
>>> +			/*
>>> +			 * The TopDown metrics events and slots event don't
>>> +			 * support any filters.
>>> +			 */
>>> +			if (event->attr.config & X86_ALL_EVENT_FLAGS)
>>> +				return -EINVAL;
>>
>> That seems independent of sampling too. Even a sampling SLOTS shouldn't
>> be having any of those afaict.
>>
>>> +
>>> +			event->hw.flags |= PERF_X86_EVENT_TOPDOWN;
>>
>> This is confusing too, a !sampling SLOTS event without METRIC siblings
>> shouldn't have this set, right? So arguably, this should be like above.
>>
>>> +
>>> +			event->event_caps |= PERF_EV_CAP_COEXIST;
>>> +
>>> +			if (is_metric_event(event))
>>> +				event->hw.flags &= ~PERF_X86_EVENT_RDPMC_ALLOWED;
>>
>> This too seems like something that should be in the is_metric_event()
>> branch above.
>>
>>> +		}
>>> +	}
>>> +
>>>   	if (!(event->attr.config & ARCH_PERFMON_EVENTSEL_ANY))
>>>   		return 0;
>>>   
> 
> FWIW, I pushed out a branch with all these changes in:
> 
>    git://git.kernel.org/pub/scm/linux/kernel/git/peterz/queue.git perf/metric
> 
> Just to get it some build love, if you want it differently, I'm happy to
> throw it all out again.

Thanks Peter.

I will pull the branch and do more tests.

Thanks,
Kan