Re: [PATCH 2/5] libperf: Propagate maps only if necessary

[Adrian Hunter <adrian.hunter@intel.com>]

On 29/09/22 23:42, Namhyung Kim wrote:
> On Wed, Sep 28, 2022 at 10:19 PM Adrian Hunter <adrian.hunter@intel.com> wrote:
>>
>> On 29/09/22 08:09, Namhyung Kim wrote:
>>> On Wed, Sep 28, 2022 at 7:08 PM Ian Rogers <irogers@google.com> wrote:
>>>>
>>>> On Wed, Sep 28, 2022 at 4:46 PM Namhyung Kim <namhyung@kernel.org> wrote:
>>>>>
>>>>> On Wed, Sep 28, 2022 at 12:54 AM Adrian Hunter <adrian.hunter@intel.com> wrote:
>>>>>>
>>>>>> On 27/09/22 20:28, Namhyung Kim wrote:
>>>>>>> Hi Adrian,
>>>>>>>
>>>>>>> On Tue, Sep 27, 2022 at 12:06 AM Adrian Hunter <adrian.hunter@intel.com> wrote:
>>>>>>>>
>>>>>>>> On 24/09/22 19:57, Namhyung Kim wrote:
>>>>>>>>> The current code propagate evsel's cpu map settings to evlist when it's
>>>>>>>>> added to an evlist.  But the evlist->all_cpus and each evsel's cpus will
>>>>>>>>> be updated in perf_evlist__set_maps() later.  No need to do it before
>>>>>>>>> evlist's cpus are set actually.
>>>>>>>>>
>>>>>>>>> Actually we discarded this intermediate all_cpus maps at the beginning
>>>>>>>>> of perf_evlist__set_maps().  Let's not do this.  It's only needed when
>>>>>>>>> an evsel is added after the evlist cpu maps are set.
>>>>>>>>
>>>>>>>> That might not be true.  Consider evlist__fix_hybrid_cpus() which fiddles
>>>>>>>> with evsel->core.cpus and evsel->core.own_cpus after the evsel has been
>>>>>>>> added to the evlist.  It can also remove an evsel from the evlist.
>>>>>>>
>>>>>>> Thanks for your review.  I think it's fine to change evsel cpus or to remove
>>>>>>> an evsel from evlist before calling evlist__create_maps().  The function
>>>>>>> will take care of setting evlist's all_cpus from the evsels in the evlist.
>>>>>>> So previous changes in evsel/cpus wouldn't be any special.
>>>>>>>
>>>>>>> After this point, adding a new evsel needs to update evlist all cpus by
>>>>>>> propagating cpu maps.  So I think hybrid cpus should be fine.
>>>>>>> Did I miss something?
>>>>>>
>>>>>> I wondered how it might play out if evlist__fix_hybrid_cpus() reduced the
>>>>>> cpus from the target->cpu_list (using perf record -C) , since after this
>>>>>> patch all_cpus always starts with the target->cpu_list instead of an empty
>>>>>> list.  But then, in the hybrid case, it puts a dummy event that uses the
>>>>>> target cpu list anyway, so the result is the same.
>>>>>>
>>>>>> I don't know if there are any cases where all_cpus would actually need to
>>>>>> exclude some of the cpus from target->cpu_list.
>>>>>
>>>>> I'm not aware of other cases to reduce cpu list.  I think it'd be fine
>>>>> if it has a cpu in the evlist->all_cpus even if it's not used.  The evsel
>>>>> should have a correct list anyway and we mostly use the evsel cpus
>>>>> to do the real work.
>>>>>
>>>>> Thanks,
>>>>> Namhyung
>>>>
>>>> The affinity changes made it so that we use all_cpus probably more
>>>> often than the evsel CPU maps for real work. The reason being we want
>>>> to avoid IPIs so we do all the work on 1 CPU and then move to the next
>>>> CPU in evlist all_cpus. evsel CPU maps are used to make sure the
>>>> indices are kept accurate - for example, if an uncore event is
>>>> measured with a CPU event:
>>>> https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/perf/util/evlist.h?h=perf/core#n366
>>>> https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/perf/util/evlist.c?h=perf/core#n404
>>>
>>> Right, I meant it'd check the evsel cpus eventually even if it iterates
>>> on the evlist all_cpus.  The evlist_cpu_iterator__next() will skip a
>>> CPU if it's not in the evsel cpus.
>>>
>>> Thanks,
>>> Namhyung
>>
>> Perhaps an alternative is to be explicit about deferring map
>> propagation e.g.
> 
> Thanks for your patch.  Yeah, we can use this.
> 
> But I still think it'd be better doing it unconditionally
> since any propagation before perf_evlist__set_maps
> will be discarded anyway.  With this change, other
> than perf record will collect all cpus before _set_maps
> and then discard it.  It seems like a waste, no?
> 
> Or else, we can have allow_map_propagation initialized
> to false and set it to true in perf_evlist__set_maps().
> 

That sounds fine.