diff mbox series

[v1,2/2] arm64: perf: Report error if PMU fails to support current CPU

Message ID 20210608145228.36595-2-leo.yan@linaro.org (mailing list archive)
State New, archived
Headers show
Series [v1,1/2] arm64: perf: Correct per-thread mode if the event is not supported | expand

Commit Message

Leo Yan June 8, 2021, 2:52 p.m. UTC
When run perf command on the Arm big.LITTLE system (Juno-r2 board):

  perf record -e cycles --per-thread program

The command executes normally without any errors; but when report the
result with "perf report", it fails to parse the symbols.  This is
because the perf data file doesn't contain MMAP2 events, thus it cannot
find the correct object file for parsing symbols.

On the big.LITTLE system, if the initialized PMU doesn't match with the
CPU the profiled task is scheduled on; for example, the initialized PMU
is on CPU0 in the LITTLE cluster, when invoke the function
perf_event_mmap_event() on CPU2 in the big cluster, the event is always
filtered out due to the CPU2 is not supported by the initialized PMU.
Finally, there have no any MMAP2 samples are generated for the
profiling.

This patch doesn't fix for this issue, alternatively, it simply reports
an error when detect the current CPU is not supported by PMU, so can
remind the user for the abnormal situation.

Signed-off-by: Leo Yan <leo.yan@linaro.org>
---
 drivers/perf/arm_pmu.c | 3 +++
 1 file changed, 3 insertions(+)

Comments

Mark Rutland June 8, 2021, 3:21 p.m. UTC | #1
Hi Leo,

On Tue, Jun 08, 2021 at 10:52:28PM +0800, Leo Yan wrote:
> When run perf command on the Arm big.LITTLE system (Juno-r2 board):
> 
>   perf record -e cycles --per-thread program
> 
> The command executes normally without any errors; but when report the
> result with "perf report", it fails to parse the symbols.  This is
> because the perf data file doesn't contain MMAP2 events, thus it cannot
> find the correct object file for parsing symbols.
> 
> On the big.LITTLE system, if the initialized PMU doesn't match with the
> CPU the profiled task is scheduled on; for example, the initialized PMU
> is on CPU0 in the LITTLE cluster, when invoke the function
> perf_event_mmap_event() on CPU2 in the big cluster, the event is always
> filtered out due to the CPU2 is not supported by the initialized PMU.
> Finally, there have no any MMAP2 samples are generated for the
> profiling.
> 
> This patch doesn't fix for this issue, alternatively, it simply reports
> an error when detect the current CPU is not supported by PMU, so can
> remind the user for the abnormal situation.

This behaviour is by design, and is not a kernel error, so it isn't
appropriate to use pr_err() here. NAK to adding the pr_err().

I agree that the behaviour for PERF_TYPE_HARDWARE is confusing and not
all that great, but as it's ABI it's not something that we can change.

What really needs to happen here is for userspace to gain some awareness
of big.LITTLE. For example, for the command:

$ perf record -e cycles --per-thread program

... what we should do is have userspace identify the set of CPU PMUs
(based on the `cpus` attribute in sysfs), and open a per-thread event on
each of those, which we can then summarize together.

Either that, or print a warning that the system is big.LITTLE and an
event will not count on all CPUs.

Thanks,
Mark.

 
> Signed-off-by: Leo Yan <leo.yan@linaro.org>
> ---
>  drivers/perf/arm_pmu.c | 3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/drivers/perf/arm_pmu.c b/drivers/perf/arm_pmu.c
> index aedea060ca8b..99ddc8bf6466 100644
> --- a/drivers/perf/arm_pmu.c
> +++ b/drivers/perf/arm_pmu.c
> @@ -565,6 +565,9 @@ static int armpmu_filter_match(struct perf_event *event)
>  	int ret;
>  
>  	ret = cpumask_test_cpu(cpu, &armpmu->supported_cpus);
> +	if (!ret)
> +		pr_err("PMU doesn't support current CPU %d\n", cpu);
> +
>  	if (ret && armpmu->filter_match)
>  		return armpmu->filter_match(event);
>  
> -- 
> 2.25.1
>
diff mbox series

Patch

diff --git a/drivers/perf/arm_pmu.c b/drivers/perf/arm_pmu.c
index aedea060ca8b..99ddc8bf6466 100644
--- a/drivers/perf/arm_pmu.c
+++ b/drivers/perf/arm_pmu.c
@@ -565,6 +565,9 @@  static int armpmu_filter_match(struct perf_event *event)
 	int ret;
 
 	ret = cpumask_test_cpu(cpu, &armpmu->supported_cpus);
+	if (!ret)
+		pr_err("PMU doesn't support current CPU %d\n", cpu);
+
 	if (ret && armpmu->filter_match)
 		return armpmu->filter_match(event);