Re: [PATCH V6] arm64: perf: Make exporting of pmu events configurable

[Srinivasarao Pathipati <quic_c_spathi@quicinc.com>]

On 6/24/2022 5:48 PM, Will Deacon wrote:
> On Thu, Jun 09, 2022 at 06:29:49PM +0530, Srinivasarao Pathipati wrote:
>> The PMU export bit (PMCR_EL0.X) is getting reset during pmu reset,
>> Make it configurable using sysctls to enable/disable at runtime.
>>
>> Signed-off-by: Srinivasarao Pathipati <quic_c_spathi@quicinc.com>
>> ---
>> Changes since V5:
>> 	- removed configuring with kernel parameters.
>> Changes since V4:
>> 	- Registering sysctls dynamically for only arm64 as suggested by Will
>> 	- Not removed the code to configure with kernel parameters
>> 	  as the sysctl's kernel parameter(sysctl.kernel.export_pmu_events)
>> 	  is not working at early bootup. pmu_reset() getting called before
>> 	  sysctl's kernel parameter is set.
>> Changes since V3:
>> 	- export bit is now configurable with sysctl
>> 	- enabling export bit on reset instead of retaining
>>
>> Changes since V2:
>> 	Done below changes as per Will's comments
>> 	- enabling pmcr_x now configurable with kernel parameters and
>> 	  by default it is disabled.
>> 	
>> Changes since V1:
>> 	- Preserving only PMCR_X bit as per Robin Murphy's comment.
>>
>> ---
>>   Documentation/admin-guide/sysctl/kernel.rst | 11 +++++++++++
>>   arch/arm64/kernel/perf_event.c              | 13 +++++++++++++
>>   2 files changed, 24 insertions(+)
>>
>> diff --git a/Documentation/admin-guide/sysctl/kernel.rst b/Documentation/admin-guide/sysctl/kernel.rst
>> index ddccd10..c2ecd84 100644
>> --- a/Documentation/admin-guide/sysctl/kernel.rst
>> +++ b/Documentation/admin-guide/sysctl/kernel.rst
>> @@ -267,6 +267,17 @@ domain names are in general different. For a detailed discussion
>>   see the ``hostname(1)`` man page.
>>   
>>   
>> +export_pmu_events (arm64 only)
>> +==============================
>> +
>> +Controls the PMU export bit (PMCR_EL0.X), which enables the exporting of
>> +events over an IMPLEMENTATION DEFINED PMU event export bus to another device.
>> +
>> +0: disables exporting of events (default).
>> +
>> +1: enables exporting of events.
>> +
>> +
>>   firmware_config
>>   ===============
>>   
>> diff --git a/arch/arm64/kernel/perf_event.c b/arch/arm64/kernel/perf_event.c
>> index cb69ff1..a8c32a0 100644
>> --- a/arch/arm64/kernel/perf_event.c
>> +++ b/arch/arm64/kernel/perf_event.c
>> @@ -298,6 +298,7 @@ PMU_FORMAT_ATTR(long, "config1:0");
>>   PMU_FORMAT_ATTR(rdpmc, "config1:1");
>>   
>>   static int sysctl_perf_user_access __read_mostly;
>> +static int sysctl_export_pmu_events __read_mostly;
>>   
>>   static inline bool armv8pmu_event_is_64bit(struct perf_event *event)
>>   {
>> @@ -1047,6 +1048,9 @@ static void armv8pmu_reset(void *info)
>>   	if (armv8pmu_has_long_event(cpu_pmu))
>>   		pmcr |= ARMV8_PMU_PMCR_LP;
>>   
>> +	if (sysctl_export_pmu_events)
>> +		pmcr |= ARMV8_PMU_PMCR_X;
> I think we need to do this in armv8pmu_start() rather than armv8pmu_reset(),
> otherwise any changes to the sysctl at runtime won't take effect unless you
> do something like re-online the CPU.
>
> Will

Hi Will ,

We think it may not work.   Say for example pmu->reset is called which 
disables the exporting of events,
and pmu->start not called. This would lead to missing of events.

How about enabling it in both reset and start functions? do you see any 
issue?

--Srinivas