[PATCH V8 23/23] perf tools: adding coresight etm PMU record capabilities

[Mathieu Poirier]

On 2 February 2016 at 09:41, Arnaldo Carvalho de Melo <acme at kernel.org> wrote:
> Em Tue, Feb 02, 2016 at 09:20:51AM -0700, Mathieu Poirier escreveu:
>> [...]
>>
>> >>> >
>> >>> > Looks OK, apart from adding linux/coresight-pmu.h to the manifest, but I
>> >>> > mentioned that on another patch.
>> >>> >
>> >>> > However there is no decoder, which begs the question, is there anything you
>> >>> > can actually do with the perf.data file?  Might be a bit confusing for users
>> >>> > if they can capture traces but not use perf tools on the resulting perf.data
>> >>> > file?
>> >>>
>> >>> We are working on a decoding library in parallel to this work.
>> >>
>> >> Would be nice to be able to get both in the same patch kit, no? So that
>> >> one can both record and process the traces, verifying it all works.
>> >
>> > We are still a few weeks away from being in a position where the
>> > community can start playing with the decoding library.  I can hold off
>> > on the "perf tools" patches when I queue the kernel side of the work
>> > for 4.6 but since you and Adrian have already reviewed the work it
>> > would be nice to have that part included as well.
>> >
>> > We've been playing with the perf.data files for a couple of months now
>> > and things look at the right place.  This isn't surprising since we
>> > are using the same framework as X86.
>> >
>> > I think the generation of the perf.data file should be coupled with
>> > the submission of the kernel driver but would also respect a diverging
>> > point of view.  Simply let me know what you prefer and I will adjust
>> > V9 accordingly.
>>
>> Arnaldo,
>>
>> I'm preparing V9 at this time - what's your view on the above?
>
> I'd say go with something we can test, i.e. if we generate a perf.data
> file we can't then process to figure out if what was inserted is right,
> how can we decide if it is ok?

Very well - I'll hold off on the perf tools patches for now.  We can
make a submission of all the user space components when things are in
a working order.

Mathieu

>
> Otherwise please describe how you test it, preferably by having this in
> the commit log, i.e. if you decide that using plain 'perf report -D' is
> enough, state that and show the output, etc.
>
> - Arnaldo