Re: [PATCH V5 0/9] perf: Driver specific configuration for PMU

[Mathieu Poirier]

On 23 August 2016 at 11:02, Alexander Shishkin
<alexander.shishkin@xxxxxxxxxxxxxxx> wrote:
> Mathieu Poirier <mathieu.poirier@xxxxxxxxxx> writes:
>
>> On 22 August 2016 at 09:15, Alexander Shishkin
>> <alexander.shishkin@xxxxxxxxxxxxxxx> wrote:
>>> Mathieu Poirier <mathieu.poirier@xxxxxxxxxx> writes:
>>>
>>>> As such something that used to be a two-step process:
>>>>
>>>> # echo 1 > /sys/bus/coresight/devices/20070000.etr/enable_sink
>>>> # perf record -e cs_etm//u --per-thread  uname
>>>>
>>>> is integrated in a single command:
>>>>
>>>> # perf record -e cs_etm/@sink=20070000.etr/u --per-thread  uname
>>>
>>> Can't we simply teach perf record to write 1 to that sysfs attribute and
>>> avoid parsing more ascii strings in the kernel? I suspect that would also
>>> take way less code.
>>
>> That, in my opinion, would be a big hack.  Peter and Jiri, any thoughts on this?
>
> Why would you say it's a hack? The whole tracer->sink configuration is
> outside of scope of perf framework, there is absolutely no reason why
> perf core should do this, especially, add this to perf ABI.

That is why the solution was made generic - sink configuration is
simply using it.

>
> It would have somewhat made sense if you could configure different
> events on the same pmu to send data to different sinks, but even that
> won't work, because you simply cannot guarantee sink's availability if
> you have to release it when your event gets scheduled out.
>
>>> Are there any other use cases for this besides specifying @sink for a
>>> ETM?
>>
>> Not at this time but there is so many configuration option for the
>> ETM/PTM tracers (that aren't filters) that I wanted the right
>> infrastructure to be there should/when we need to expand.
>
> As I've asked before, what exactly is the need for expansion? That is,
> something more than purely hypothetical that needs to be configured in
> the tracer, on per event basis, *and* does not fit into the event
> attribute. Note that the sink configuration also doesn't fit the bill.

The first thing that comes to mind is counters and state machine,
something that I simply don't see fitting into the event attributes.
But those could also be set via sysFS - the end result is exactly the
same.

>
> Regards,
> --
> Alex