Messages in this thread | | | Date | Tue, 5 Jul 2011 15:31:05 +0200 | From | Ingo Molnar <> | Subject | Re: [PATCH -tip, final] perf, x86: Add hw_watchdog_set_attr() in a sake of nmi-watchdog on P4 |
| |
* Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:
> > So the question is, why does the NMI watchdog prevent 'perf top' > > from working on a P4? > > because the NMI watchdog is a pinned event, you don't want to share > the counter, that would be very bad, suppose you lock up when the > NMI watchdog was scheduled out. Unreliably debug tools are worse > than no tools.
Yeah, indeed that explains the symptom.
Firstly, we should fix/enhance perf top to print out an error message in this case, not just hang there doing nothing.
Secondly, the proper solution would be to allow the multiplexing of like-minded hw events. Here if we have two events:
- pinned NMI watchdog, set to a period of 2 billion cycles - perf top with a default of 1 khz auto-freq cycles
We should first change the NMI watchdog to use auto-freq samples - the hw_nmi_get_sample_period() looks unnecessary - if we set the NMI watchdog to 1 Hz by default it should be more than enough.
Thus we'd have two events:
- pinned NMI watchdog, set to 1 Hz - perf top, set to 1000 Hz
The 1000 Hz event could serve the 1 Hz event just fine.
Thirdly, even if we mucked the NMI event to be somewhat different from the real cycles event (so the P4 PMU can co-schedule it with perf top), the real solution there would *still* be to express this as a variation of a cycles event: use the same trick to allow up to two cycles events to be present in the PMU - but hide this from the generic interfaces, just allow up to two cycles events to be scheduled at once.
This will have the advantage of not only fixing the NMI watchdog, but other users as well.
Now, if there's some real behavioral difference between the two events (one is halted cycles the other is unhalted cycles), then i'd suggest to use PERF_COUNT_HW_BUS_CYCLES in the NMI watchdog - that is the generic 'constant frequency' cycles event.
So there's lots of options to fix/improve this more intelligently.
Thanks,
Ingo
| |