On Mittwoch, 11. Dezember 2024 17:09:19 Mitteleuropäische Normalzeit Christian Hergert wrote: > On vacation until 2025 so my replies will be sparse here :) > > On 12/11/24 3:24 AM, Milian Wolff wrote: > > For symmetry reasons, we also need an `..._end` method. And what about > > explicitly removing processes once they have finished (PERF_RECORD_EXIT)? > > One area you can run into issues with this is that there is no > synchronization between Perf streams from different CPUs. So you could > end up processing an EXIT from PID N on CPU 1 before a SAMPLE from PID N > on CPU 2.
I thought that the perf events are never ordered when you obtain them, you always first have to sort them based on their timestamps. For optimization purposes thankfully we have PERF_RECORD_FINISHED_ROUND so we don't need to buffer all events and sort them. That said, we do sometimes see violations of this, with some events arriving after a PERF_RECORD_FINISHED_ROUND with a timestamp earlier than what was previously handled in the last round. Is that what you have in mind, or are you thinking of something else that I am not aware of yet? Cheers -- Milian Wolff m...@milianw.de http://milianw.de
signature.asc
Description: This is a digitally signed message part.
