Re: [V14 3/8] drivers: perf: arm_pmuv3: Enable branch stack sampling framework

[James Clark <james.clark@arm.com>]

On 15/11/2023 07:22, Anshuman Khandual wrote:
> On 11/14/23 17:44, James Clark wrote:
>>
>>
>> On 14/11/2023 05:13, Anshuman Khandual wrote:
>> [...]
>>
>>> +/*
>>> + * This is a read only constant and safe during multi threaded access
>>> + */
>>> +static struct perf_branch_stack zero_branch_stack = { .nr = 0, .hw_idx = -1ULL};
>>> +
>>> +static void read_branch_records(struct pmu_hw_events *cpuc,
>>> +				struct perf_event *event,
>>> +				struct perf_sample_data *data,
>>> +				bool *branch_captured)
>>> +{
>>> +	/*
>>> +	 * CPU specific branch records buffer must have been allocated already
>>> +	 * for the hardware records to be captured and processed further.
>>> +	 */
>>> +	if (WARN_ON(!cpuc->branches))
>>> +		return;
>>> +
>>> +	/*
>>> +	 * Overflowed event's branch_sample_type does not match the configured
>>> +	 * branch filters in the BRBE HW. So the captured branch records here
>>> +	 * cannot be co-related to the overflowed event. Report to the user as
>>> +	 * if no branch records have been captured, and flush branch records.
>>> +	 * The same scenario is applicable when the current task context does
>>> +	 * not match with overflown event.
>>> +	 */
>>> +	if ((cpuc->brbe_sample_type != event->attr.branch_sample_type) ||
>>> +	    (event->ctx->task && cpuc->brbe_context != event->ctx)) {
>>> +		perf_sample_save_brstack(data, event, &zero_branch_stack);
>>
>> Is there any benefit to outputting a zero size stack vs not outputting
>> anything at all?
> 
> The event has got PERF_SAMPLE_BRANCH_STACK marked and hence perf_sample_data
> must have PERF_SAMPLE_BRANCH_STACK with it's br_stack pointing to the branch
> records. Hence without assigning a zeroed struct perf_branch_stack, there is
> a chance, that perf_sample_data will pass on some garbage branch records to
> the ring buffer.
> 

I don't think that's an issue, the perf core code handles the case where
no branch stack exists on a sample. It even outputs the zero length for
you, but there is other stuff that can be skipped if you just never call
perf_sample_save_brstack():

from kernel/events/core.c:

 if (sample_type & PERF_SAMPLE_BRANCH_STACK) {
  if (data->br_stack) {
    size_t size;

    size = data->br_stack->nr
      * sizeof(struct perf_branch_entry);

    perf_output_put(handle, data->br_stack->nr);
    if (branch_sample_hw_index(event))
      perf_output_put(handle, data->br_stack->hw_idx);
    perf_output_copy(handle, data->br_stack->entries, size);
  } else {
    /*
     * we always store at least the value of nr
     */
    u64 nr = 0;
    perf_output_put(handle, nr);
  }
}


>>
>>> +		return;
>>> +	}
>>> +
>>> +	/*
>>> +	 * Read the branch records from the hardware once after the PMU IRQ
>>> +	 * has been triggered but subsequently same records can be used for
>>> +	 * other events that might have been overflowed simultaneously thus
>>> +	 * saving much CPU cycles.
>>> +	 */
>>> +	if (!*branch_captured) {
>>> +		armv8pmu_branch_read(cpuc, event);
>>> +		*branch_captured = true;
>>> +	}
>>> +	perf_sample_save_brstack(data, event, &cpuc->branches->branch_stack);
>>> +}
>>> +
>>>  static irqreturn_t armv8pmu_handle_irq(struct arm_pmu *cpu_pmu)
>>>  {
>>>  	u32 pmovsr;
>>> @@ -766,6 +815,7 @@ static irqreturn_t armv8pmu_handle_irq(struct arm_pmu *cpu_pmu)
>>>  	struct pmu_hw_events *cpuc = this_cpu_ptr(cpu_pmu->hw_events);
>>>  	struct pt_regs *regs;
>>>  	int idx;
>>> +	bool branch_captured = false;
>>>  
>>>  	/*
>>>  	 * Get and reset the IRQ flags
>>> @@ -809,6 +859,13 @@ static irqreturn_t armv8pmu_handle_irq(struct arm_pmu *cpu_pmu)
>>>  		if (!armpmu_event_set_period(event))
>>>  			continue;
>>>  
>>> +		/*
>>> +		 * PMU IRQ should remain asserted until all branch records
>>> +		 * are captured and processed into struct perf_sample_data.
>>> +		 */
>>> +		if (has_branch_stack(event) && cpu_pmu->has_branch_stack)
>>> +			read_branch_records(cpuc, event, &data, &branch_captured);
>>
>> You could return instead of using the out param, not really any
>> different, but maybe a bit more normal:
>>
>>   branch_captured |= read_branch_records(cpuc, event, &data,
>> branch_captured);
> 
> I am just wondering - how that would be any better ?
> 

Maybe it wouldn't, but I suppose it's just the same way you don't write
returns like:

  armv8pmu_task_ctx_cache_alloc(cpu_pmu, &ret);

instead of:

  ret = armv8pmu_task_ctx_cache_alloc(cpu_pmu);

Out params can be hard to reason about sometimes. Maybe not in this case
though.

>>
>>> +
>>>  		/*
>>>  		 * Perf event overflow will queue the processing of the event as
>>>  		 * an irq_work which will be taken care of in the handling of
>>> @@ -818,6 +875,8 @@ static irqreturn_t armv8pmu_handle_irq(struct arm_pmu *cpu_pmu)
>>>  			cpu_pmu->disable(event);
>>>  	}
>>>  	armv8pmu_start(cpu_pmu);
>>> +	if (cpu_pmu->has_branch_stack)
>>> +		armv8pmu_branch_reset();
>>>  
>>>  	return IRQ_HANDLED;
>>>  }
>>> @@ -907,6 +966,24 @@ static int armv8pmu_user_event_idx(struct perf_event *event)
>>>  	return event->hw.idx;
>>>  }
>>>  
>>> +static void armv8pmu_sched_task(struct perf_event_pmu_context *pmu_ctx, bool sched_in)
>>> +{
>>> +	struct arm_pmu *armpmu = to_arm_pmu(pmu_ctx->pmu);
>>> +	void *task_ctx = pmu_ctx->task_ctx_data;
>>> +
>>> +	if (armpmu->has_branch_stack) {
>>> +		/* Save branch records in task_ctx on sched out */
>>> +		if (task_ctx && !sched_in) {
>>> +			armv8pmu_branch_save(armpmu, task_ctx);
>>> +			return;
>>> +		}
>>> +
>>> +		/* Reset branch records on sched in */
>>> +		if (sched_in)
>>> +			armv8pmu_branch_reset();
>>> +	}
>>> +}
>>> +
>>>  /*
>>>   * Add an event filter to a given event.
>>>   */
>>> @@ -977,6 +1054,9 @@ static void armv8pmu_reset(void *info)
>>>  		pmcr |= ARMV8_PMU_PMCR_LP;
>>>  
>>>  	armv8pmu_pmcr_write(pmcr);
>>> +
>>> +	if (cpu_pmu->has_branch_stack)
>>> +		armv8pmu_branch_reset();
>>>  }
>>>  
>>>  static int __armv8_pmuv3_map_event_id(struct arm_pmu *armpmu,
>>> @@ -1014,6 +1094,20 @@ static int __armv8_pmuv3_map_event(struct perf_event *event,
>>>  
>>>  	hw_event_id = __armv8_pmuv3_map_event_id(armpmu, event);
>>>  
>>> +	if (has_branch_stack(event)) {
>>> +		if (!armv8pmu_branch_attr_valid(event))
>>> +			return -EOPNOTSUPP;
>>> +
>>> +		/*
>>> +		 * If a task gets scheduled out, the current branch records
>>> +		 * get saved in the task's context data, which can be later
>>> +		 * used to fill in the records upon an event overflow. Let's
>>> +		 * enable PERF_ATTACH_TASK_DATA in 'event->attach_state' for
>>> +		 * all branch stack sampling perf events.
>>> +		 */
>>> +		event->attach_state |= PERF_ATTACH_TASK_DATA;
>>> +	}
>>> +
>>>  	/*
>>>  	 * CHAIN events only work when paired with an adjacent counter, and it
>>>  	 * never makes sense for a user to open one in isolation, as they'll be
>>> @@ -1130,6 +1224,35 @@ static void __armv8pmu_probe_pmu(void *info)
>>>  		cpu_pmu->reg_pmmir = read_pmmir();
>>>  	else
>>>  		cpu_pmu->reg_pmmir = 0;
>>> +	armv8pmu_branch_probe(cpu_pmu);
>>
>> I'm not sure if this is splitting hairs or not, but
>> __armv8pmu_probe_pmu() is run on only one of 'any' of the supported CPUs
>> for this PMU.
> 
> Right.
> 
>>
>> Is it not possible to have some of those CPUs support and some not
>> support BRBE, even though they are all the same PMU type? Maybe we could
> 
> I am not sure, but not something I have come across.
> 
>> wait for it to explode with some weird system, or change it so that the
>> BRBE probe is run on every CPU, with a second 'supported_brbe_mask' field.
> 
> Right, but for now, the current solutions looks sufficient.
> 

I suppose it means people will have to split their PMUs to ones that do
and don't support BRBE. I'm not sure if that's worth adding a comment in
the docs or it's too obscure.

>>
>>> +}
>>> +
>>> +static int branch_records_alloc(struct arm_pmu *armpmu)
>>> +{
>>> +	struct branch_records __percpu *records;
>>> +	int cpu;
>>> +
>>> +	records = alloc_percpu_gfp(struct branch_records, GFP_KERNEL);
>>> +	if (!records)
>>> +		return -ENOMEM;
>>> +
>>
>> Doesn't this technically need to take the CPU mask where BRBE is
>> supported into account? Otherwise you are allocating for cores that
>> never use it.
>>
>> Also it's done per-CPU _and_ per-PMU type, multiplying the number of
>> BRBE buffers allocated, even if they can only ever be used per-CPU.
> 
> Agreed, but I believe we have already been though this discussion, and
> settled for this method - for being a simpler approach.
> 
>>
>>> +	/*
>>> +	 * percpu memory allocated for 'records' gets completely consumed
>>> +	 * here, and never required to be freed up later. So permanently
>>> +	 * losing access to this anchor i.e 'records' is acceptable.
>>> +	 *
>>> +	 * Otherwise this allocation handle would have to be saved up for
>>> +	 * free_percpu() release later if required.
>>> +	 */
>>> +	for_each_possible_cpu(cpu) {
>>> +		struct pmu_hw_events *events_cpu;
>>> +		struct branch_records *records_cpu;
>>> +
>>> +		events_cpu = per_cpu_ptr(armpmu->hw_events, cpu);
>>> +		records_cpu = per_cpu_ptr(records, cpu);
>>> +		events_cpu->branches = records_cpu;
>>> +	}
>>> +	return 0;
>>>  }
>>>  
>>>  static int armv8pmu_probe_pmu(struct arm_pmu *cpu_pmu)
>>> @@ -1146,7 +1269,21 @@ static int armv8pmu_probe_pmu(struct arm_pmu *cpu_pmu)
>>>  	if (ret)
>>>  		return ret;
>>>  
>>> -	return probe.present ? 0 : -ENODEV;
>>> +	if (!probe.present)
>>> +		return -ENODEV;
>>> +
>>> +	if (cpu_pmu->has_branch_stack) {
>>> +		ret = armv8pmu_task_ctx_cache_alloc(cpu_pmu);
>>> +		if (ret)
>>> +			return ret;
>>> +
>>> +		ret = branch_records_alloc(cpu_pmu);
>>> +		if (ret) {
>>> +			armv8pmu_task_ctx_cache_free(cpu_pmu);
>>> +			return ret;
>>> +		}
>>> +	}
>>> +	return 0;
>>>  }
>>>  
>>
>> [...]
>>> diff --git a/include/linux/perf/arm_pmuv3.h b/include/linux/perf/arm_pmuv3.h
>>> index 9c226adf938a..72da4522397c 100644
>>> --- a/include/linux/perf/arm_pmuv3.h
>>> +++ b/include/linux/perf/arm_pmuv3.h
>>> @@ -303,4 +303,50 @@
>>>  		}						\
>>>  	} while (0)
>>>  
>>> +struct pmu_hw_events;
>>> +struct arm_pmu;
>>> +struct perf_event;
>>> +
>>> +#ifdef CONFIG_PERF_EVENTS
>>
>> Very minor nit, but if you end up moving the stubs to the brbe header
>> you probably don't need the #ifdef CONFIG_PERF_EVENTS because it just
>> won't be included in that case.
> 
> Right, will drop CONFIG_PERF_EVENTS wrapper.
> 
>>
>>> +static inline void armv8pmu_branch_reset(void)
>>> +{
>>> +}
>>> +
>>> +static inline void armv8pmu_branch_probe(struct arm_pmu *arm_pmu)
>>> +{
>>> +}
>>> +
>>> +static inline bool armv8pmu_branch_attr_valid(struct perf_event *event)
>>> +{
>>> +	WARN_ON_ONCE(!has_branch_stack(event));
>>> +	return false;
>>> +}
>>> +
>>> +static inline void armv8pmu_branch_enable(struct arm_pmu *arm_pmu)
>>> +{
>>> +}
>>> +
>>> +static inline void armv8pmu_branch_disable(void)
>>> +{
>>> +}
>>> +
>>> +static inline void armv8pmu_branch_read(struct pmu_hw_events *cpuc,
>>> +					struct perf_event *event)
>>> +{
>>> +	WARN_ON_ONCE(!has_branch_stack(event));
>>> +}
>>> +
>>> +static inline void armv8pmu_branch_save(struct arm_pmu *arm_pmu, void *ctx)
>>> +{
>>> +}
>>> +
>>> +static inline int armv8pmu_task_ctx_cache_alloc(struct arm_pmu *arm_pmu)
>>> +{
>>> +	return 0;
>>> +}
>>> +
>>> +static inline void armv8pmu_task_ctx_cache_free(struct arm_pmu *arm_pmu)
>>> +{
>>> +}
>>> +#endif /* CONFIG_PERF_EVENTS */
>>>  #endif