Re: [PATCH v2 2/7] perf evlist: Add evlist__findnew_tracking_event() helper

[Yang Jihong]

Hello,

On 2023/7/29 0:40, Ian Rogers wrote:
> On Thu, Jul 20, 2023 at 12:24 AM Yang Jihong <yangjihong1@xxxxxxxxxx> wrote:
> > Hello,
> >
> > On 2023/7/20 0:44, Ian Rogers wrote:
> > > On Fri, Jul 14, 2023 at 8:31 PM Yang Jihong <yangjihong1@xxxxxxxxxx> wrote:
> > > > Currently, intel-bts, intel-pt, and arm-spe may add a dummy event for
> > > > tracking to the evlist. We may need to search for the dummy event for
> > > > some settings. Therefore, add evlist__findnew_tracking_event() helper.
> > > >
> > > > evlist__findnew_tracking_event() also deal with system_wide maps if
> > > > system_wide is true.
> > >
> > > I'm wondering if we can simplify the naming in the API, we have "dummy
> > > event" which makes sense as we literally call the event "dummy",
> > > "sideband" which refers to the kind of samples/events the dummy event
> > > will record but "tracking" I think tends to get used as a verb rather
> > > than a noun. So I think evlist__findnew_tracking_event should be
> > > evlist__findnew_dummy_event.
> > Uh, from the discussion that followed, it seems that there is no
> > consensus yet...
> > If there is a clear consensus on whether to use "dummy event" or
> > "tracking event", I will change the name of the API.
> >
> > I think sideband event is equivalent to tracking event (refer
> > evsel__config(), tracking events include task, mmap, mmap2, and comm
> > sideband events, which are all sideband).
> >
> > tracking event are instances of dummy event. For example, we create
> > another dummy event to record the text poke event of ksymbol (refer perf
> > record --kcore).
> >
> > An evlist contains only one tracking event, but can contain multiple
> > dummy events.
>
> Thanks for the feedback. So the tracking event is by definition the
> first dummy event in the evlist? What is the purpose of the other
Uh... It may not be the first dummy event, but evsel->track must be 
true. Only one evsel in an evlist meets this condition.

> dummy events in this case? Perhaps we can get to an intention
> revealing implementation something like:
>
> /** The "tracking event" gathering sideband data is the first dummy
> event in the list. */
> struct evsel *evlist__findnew_tracking_event(struct evlist *evlist)
> {
>     struct evsel *dummy = evlist__find_first_dummy_event(evlist);
>
>     if (!dummy) {
>        dummy = evlist__add_dummy(evlist);
>     }
>     return dummy;
> }
>
> But I think the key thing for me is I'm still not sure what is going
> on when there are multiple dummy events for you, what are the other
> dummy events for other than tracking sideband data?
For other dummy events, perf record will open a dummy event to track 
ksymbol text_poke when "--kcore" option is used.
I thinks tracking ksymbol text_poke separately it needs to be processed 
independently, go system_wide and enable immediately.

All of the above is my understanding, may need Adrian to confirm whether 
it is accurate.

Thanks,
Yang