Re: [PATCH v4 0/4] perf: Support PERF_SAMPLE_READ with inherit_stat

From: Namhyung Kim
Date: Wed Mar 27 2024 - 16:34:41 EST


Hello,

On Mon, Mar 25, 2024 at 3:12 AM Ben Gainey <Ben.Gainey@xxxxxxx> wrote:
>
> On Fri, 2024-03-22 at 18:22 -0700, Namhyung Kim wrote:
> > On Fri, Mar 22, 2024 at 9:42 AM Ben Gainey <ben.gainey@xxxxxxx>
> > wrote:
> > >
> > > This change allows events to use PERF_SAMPLE READ with inherit so
> > > long
> > > as both inherit_stat and PERF_SAMPLE_TID are set.
> > >
> > > Currently it is not possible to use PERF_SAMPLE_READ with inherit.
> > > This
> > > restriction assumes the user is interested in collecting aggregate
> > > statistics as per `perf stat`. It prevents a user from collecting
> > > per-thread samples using counter groups from a multi-threaded or
> > > multi-process application, as with `perf record -e '{....}:S'`.
> > > Instead
> > > users must use system-wide mode, or forgo the ability to sample
> > > counter
> > > groups. System-wide mode is often problematic as it requires
> > > specific
> > > permissions (no CAP_PERFMON / root access), or may lead to capture
> > > of
> > > significant amounts of extra data from other processes running on
> > > the
> > > system.
> > >
> > > Perf already supports the ability to collect per-thread counts with
> > > `inherit` via the `inherit_stat` flag. This patch changes
> >
> > I'm not sure about this part. IIUC inherit and inherit_stat is not
> > for
> > per-thread counts, it only supports per-process (including children)
> > events.
>
> Hi Namhyung
>
> Thanks for the comments...
>
> I don't think this is correct, if you compare the behaviour of
>
> perf record --no-inherit ... <some-forking-processes>
> perf script -F pid,tid | sort -u
> and
> perf record --no-inherit ... <some-multithreaded-processes>
> perf script -F pid,tid | sort -u
>
> vs
>
> perf record ... <some-forking-processes>
> perf script -F pid,tid | sort -u
> and
> perf record .. <some-multithreaded-processes>
> perf script -F pid,tid | sort -u
>
> The behaviour is consistent with the fact that no-inherit only records
> the primary thread of the primary process, whereas in the inherit case
> any child tasks (either threads or forked processes) is recorded.

Right, I was talking about the counting behavior not sampling
as inherit_stat is only for the counting. I think it'd return an error
if event attr has both sample_freq and inherit_stat.


> >
> >
> > > `perf_event_alloc` relaxing the restriction to combine `inherit`
> > > with
> > > `PERF_SAMPLE_READ` so that the combination will be allowed so long
> > > as
> > > `inherit_stat` and `PERF_SAMPLE_TID` are enabled.
> >
> > Anyway, does it really need 'inherit_stat'? I think it's only for
> > counting use cases (e.g. 'perf stat') not for sampling.
>
>
> I would be very happy to remove the inherit_stat requirement. When I
> first came to this it seemed like the logic was all there in
> inherit_stat already, but now that I have to take a different path in
> `perf_event_context_sched_out` I suspect it should be trivial to remove
> the inherit_stat requirement.

ok.

> >
> > Also technically, it can have PERF_SAMPLE_STREAM_ID instead
> > of PERF_SAMPLE_TID to distinguish the counter values.
>
>
> It looks like you are correct, but the ID given in the read_format part
> of PERF_SAMPLE_RECORD is the ID rather than STREAM_ID. (I had
> incorrectly thought/stated it was the latter). Hence when processing
> the read_format values in the sample record, we either need to use the
> TID to uniquely identify the source, or we would need to modify the
> read_format to (additionally) include the STREAM_ID.
>
> * The current approach in tools uses the ID+TID, which puts more
> complexity in the tools but means there isn't an extra field in the
> read_format data (for each value).
> * Alternatively I could introduce a PERF_FORMAT_STREAM_ID; I would
> expect that the user/tool would need to specify
> PERF_FORMAT_ID|PERF_FORMAT_STREAM_ID as they would need to use the ID
> to lookup the correct perf_event_attr, but could use the STREAM_ID to
> uniquely identify the child event. This approach would add an extra u64
> per value in the read_format data but is possibly simpler/safer for
> tools?
>
> Any preferences?

I think it's better to use TID + ID. IIUC there's no way to track
STREAM_ID for new children other than getting it from sample.
As sample has TID already it'd be meaningless using STREAM_ID
to distinguish events.

Thanks,
Namhyung