Re: [PATCH 2/4] drivers/nvdimm: Add perf interface to expose nvdimm performance stats

[kajoljain]

On 6/8/21 11:06 PM, Peter Zijlstra wrote:
> On Tue, Jun 08, 2021 at 05:26:58PM +0530, Kajol Jain wrote:
>> +static int nvdimm_pmu_cpu_offline(unsigned int cpu, struct hlist_node *node)
>> +{
>> +	struct nvdimm_pmu *nd_pmu;
>> +	u32 target;
>> +	int nodeid;
>> +	const struct cpumask *cpumask;
>> +
>> +	nd_pmu = hlist_entry_safe(node, struct nvdimm_pmu, node);
>> +
>> +	/* Clear it, incase given cpu is set in nd_pmu->arch_cpumask */
>> +	cpumask_test_and_clear_cpu(cpu, &nd_pmu->arch_cpumask);
>> +
>> +	/*
>> +	 * If given cpu is not same as current designated cpu for
>> +	 * counter access, just return.
>> +	 */
>> +	if (cpu != nd_pmu->cpu)
>> +		return 0;
>> +
>> +	/* Check for any active cpu in nd_pmu->arch_cpumask */
>> +	target = cpumask_any(&nd_pmu->arch_cpumask);
>> +	nd_pmu->cpu = target;
>> +
>> +	/*
>> +	 * Incase we don't have any active cpu in nd_pmu->arch_cpumask,
>> +	 * check in given cpu's numa node list.
>> +	 */
>> +	if (target >= nr_cpu_ids) {
>> +		nodeid = cpu_to_node(cpu);
>> +		cpumask = cpumask_of_node(nodeid);
>> +		target = cpumask_any_but(cpumask, cpu);
>> +		nd_pmu->cpu = target;
>> +
>> +		if (target >= nr_cpu_ids)
>> +			return -1;
>> +	}
>> +
>> +	return 0;
>> +}
>> +
>> +static int nvdimm_pmu_cpu_online(unsigned int cpu, struct hlist_node *node)
>> +{
>> +	struct nvdimm_pmu *nd_pmu;
>> +
>> +	nd_pmu = hlist_entry_safe(node, struct nvdimm_pmu, node);
>> +
>> +	if (nd_pmu->cpu >= nr_cpu_ids)
>> +		nd_pmu->cpu = cpu;
>> +
>> +	return 0;
>> +}
> 
>> +static int nvdimm_pmu_cpu_hotplug_init(struct nvdimm_pmu *nd_pmu)
>> +{
>> +	int nodeid, rc;
>> +	const struct cpumask *cpumask;
>> +
>> +	/*
>> +	 * Incase cpu hotplug is not handled by arch specific code
>> +	 * they can still provide required cpumask which can be used
>> +	 * to get designatd cpu for counter access.
>> +	 * Check for any active cpu in nd_pmu->arch_cpumask.
>> +	 */
>> +	if (!cpumask_empty(&nd_pmu->arch_cpumask)) {
>> +		nd_pmu->cpu = cpumask_any(&nd_pmu->arch_cpumask);
>> +	} else {
>> +		/* pick active cpu from the cpumask of device numa node. */
>> +		nodeid = dev_to_node(nd_pmu->dev);
>> +		cpumask = cpumask_of_node(nodeid);
>> +		nd_pmu->cpu = cpumask_any(cpumask);
>> +	}
>> +
>> +	rc = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "perf/nvdimm:online",
>> +				     nvdimm_pmu_cpu_online, nvdimm_pmu_cpu_offline);
>> +
> 
> Did you actually test this hotplug stuff?
> 
> That is, create a counter, unplug the CPU the counter was on, and
> continue counting? "perf stat -I" is a good option for this, concurrent
> with a hotplug.
>
> Because I don't think it's actually correct. The thing is perf core is
> strictly per-cpu, and it will place the event on a specific CPU context.
> If you then unplug that CPU, nothing will touch the events on that CPU
> anymore.
> 
> What drivers that span CPUs need to do is call
> perf_pmu_migrate_context() whenever the CPU they were assigned to goes
> away. Please have a look at arch/x86/events/rapl.c or
> arch/x86/events/amd/power.c for relatively simple drivers that have this
> property.
> 


Hi Peter,
    Primarily I tested off-lining multiple cpus and checking if cpumask file is updating as expected,
followed with perf stat commands.
But I missed the scenario where we are off-lining CPU while running perf stat. My bad, thanks
for pointing it out.
I will fix this issue and send new version of the patchset.

Thanks,
Kajol Jain
>