Messages in this thread | | | Date | Mon, 20 Dec 2021 10:30:34 +0100 | From | Peter Zijlstra <> | Subject | Re: [PATCH v3] perf/core: Set event shadow time for inactive events too |
| |
On Sat, Dec 18, 2021 at 09:09:05AM +0000, Song Liu wrote: > > > > On Dec 17, 2021, at 8:35 AM, Peter Zijlstra <peterz@infradead.org> wrote: > > > > On Fri, Dec 10, 2021 at 11:33:41AM +0100, Peter Zijlstra wrote: > > > >> I'm thinking this is a cgroup specific thing. Normally the shadow_time > >> thing is simply a relative displacement between event-time and the > >> global clock. That displacement never changes, except when you do > >> IOC_DISABLE/IOC_ENABLE. > >> > >> However, for cgroup things are different, since the cgroup events aren't > >> unconditionally runnable, that is, the enabled time should only count > >> when the cgroup is active, right? > >> > >> So perhaps perf_event_read_local() should use a cgroup clock instead of > >> perf_clock() for cgroup events. > >> > >> Let me think about that some more... > > > > How's this then? Song, could you also please test and or better explain > > the problem f79256532682 pretends to cure? Because the below is > > reverting that, I *really* hate having to touch the events we're not > > scheduling. > > Unfortunately, this change bring the bug back. For time_enabled in rdpmc > case to work properly, we have to touch all the enabled but not running > events, right?
Ohh.. argh. I think I see why, it looses the context time enable edge, and because this is all strictly per-event in the uapi (there is no ctx representation) it can't be cured by improving ctx time handling :/
Bah, I so hate this.
| |