Re: [PATCH 00/20] perf vendors events arm64: Multiple Arm CPUs

[Nick Forrington]

On 15/05/2022 23:03, Ian Rogers wrote:
> On Thu, May 12, 2022 at 8:53 AM John Garry <john.garry@xxxxxxxxxx> wrote:
>
> Generally this looks ok:
>
> Reviewed-by: John Garry <john.garry@xxxxxxxxxx>
>
> If you are feeling particularly helpful then you can add support for any
> events missing to pre-existing core support, like a57-a72.
>
> I'll raise John's "ok" and say this looks great! :-D Some thoughts:

Thanks Ian!

> The mapfile.csv cpuid values don't directly align with:
> https://github.com/ARM-software/data/blob/master/cpus.json
> but this definitely looks deliberate.

Correct - they use different formats.

mapfile.csv uses the MIDR format

https://developer.arm.com/documentation/100442/0100/register-descriptions/aarch64-system-registers/midr-el1--main-id-register--el1

The cpus.json "cpuid" is the implementer and part number from the MIDR 
(the other fields are always fixed in mapfile.csv)
> The new events lack the PMU "Unit" value. The current perf json is
> pretty free form and leads to problems if two PMUs are present.
> Context is here:
> https://lore.kernel.org/lkml/CAP-5=fWRRZsyJZ-gky-FOFz79zW_3r78d_0APpj5sf66HqTpLw@xxxxxxxxxxxxxx/
>
> My idea to rationalize this is to mirror what is already done in
> sysfs, that is the event data is specific to a PMU. As a lot of "Unit"
> values are missing from events on x86 a reasonable guess if the "Unit"
> is missing is to use "cpu". Poking a Google Pixel 4a, I see that all
> PMU data is in "armv8_pmuv3". So for ARM I could guess this is always
> the case, ie all events should belong to armv8_pmuv3. This may not be
> right and could lead to confusion like an event BR_COND_MIS_PRED
> having an alias of "armv8_pmuv3/BR_COND_MIS_PRED/" but it really
> should have some other PMU name in there. I just raise this in case
> there is a fix for this we could incorporate into this patch series,
> maybe "armv8_pmuv3" is always the PMU and my life is easy.

My understanding is that all JSON events under arm64/arm apply to the 
CPU PMU, although there could be 2 (or more) armv8_pmuv3 devices in a 
herterogeneous system (armv8_pmuv3_0, armv8_pmuv3_1, ...) - each with 
different events.

So I don't think static "Unit" data would be helpful, but it should be 
possible to map JSON events to appropriate CPUs with existing data.

e.g. /sys/bus/event_source/devices/armv8_pmuv3_0/cpus shows the CPUs 
associated with a PMU device, and each CPU can be mapped to JSON events 
via the MIDR (as is done already)

Thanks,
Nick