Re: perf_counters issue with PERF_SAMPLE_GROUP

[stephane eranian <eranian@googlemail.com>]

On Tue, Aug 11, 2009 at 10:55 PM, Peter Zijlstra<a.p.zijlstra@chello.nl> wrote:
> On Tue, 2009-08-11 at 21:40 +0200, stephane eranian wrote:
>
>> > You seem to have forgotten to append your test.c though :-)
>> >
>> Can't send you the program because it uses extra bits and pieces
>> which are hard to remove. Otherwise I would have send it already.
>
> Those other bits aren't open source? tskk :-)
>

You don't know me well!
They will be but they are not ready yet.

>> But I think it boils down to the following piece of code in
>> perf_counter_output():
>>                 leader = counter->group_leader;
>>                 list_for_each_entry(sub, &leader->sibling_list, list_entry) {
>>                         if (sub != counter)
>>                                 sub->pmu->read(sub);
>>
>>                         group_entry.id = primary_counter_id(sub);
>>                         group_entry.counter = atomic64_read(&sub->count);
>>
>>                         perf_output_put(&handle, group_entry);
>>                 }
>
> Well, likely, but nothing obviously wrong stands out there, so now I get
> to write a reproduces to see what's going wrong.
>

>> >> Related to PERF_SAMPLE_GROUP, I believe there is some information missing.
>> >> You need to provide the TIMING information because in the case of SAMPLE_GROUP
>> >> you'd like to be able to scale the values of the counters you are
>> >> collecting. And you
>> >> need the timing at the moment, the sample was recorded not later.
>> >
>> > Right, so something like the below, possibly complemented with having
>> > PERF_COUNTER_IOC_RESET also reset the run-times?
>> >
>> Yes, but don't you have a namespace issue between PERF_FORMAT_* and
>> PERF_SAMPLE_* in the patch below? I would think you want to keep them separate.
>
> Maybe, otoh we've consistently used it whenever exposing the timing
> data.
>
How do you make sure the bits used by the PERF_FORMAT_* stuff does not collide
with PERF_SAMPLE_*?

>> I am also wondering about why one would want one timing value and not the other.
>> In other words, why not group them under a single name. But maybe it is harder
>> to return more than one u64 per PERF_FORMAT?
>
> Not really, Paul did it like that initially and we've been consistently
> doing it like that -- changing it now is a bit late.
>

But aren't you going to change the cpu, pid target stuff we discussed a couple
of weeks ago anyway?