Re: [perfmon2] [PATCH] perf_events: fix time tracking in samples

[Stephane E. <er...@go...>]

On Tue, Oct 19, 2010 at 7:09 PM, Peter Zijlstra <pe...@in...> wrote:
> On Tue, 2010-10-19 at 19:01 +0200, Stephane Eranian wrote:
>> On Tue, Oct 19, 2010 at 6:52 PM, Peter Zijlstra <pe...@in...> wrote:
>> > On Tue, 2010-10-19 at 18:47 +0200, Stephane Eranian wrote:
>> >> This patch corrects time tracking in samples. Without this patch
>> >> both time_enabled and time_running may be reported as zero when
>> >> user asks for PERF_SAMPLE_READ.
>> >>
>> >> You use PERF_SAMPLE_READ when you want to sample the values of
>> >> other counters in each sample. Because of multiplexing, it is
>> >> necessary to know both time_enable, time_running to be able
>> >> to scale counts correctly.
>> >>
>> >> We defer updating timing until we know it is really needed, i.e.,
>> >> only when we have PERF_SAMPLE_READ.
>> >>
>> >> With this patch, the libpfm4 example task_smpl now reports
>> >> correct counts (shown on 2.4GHz Core 2):
>> >>
>> >> $ task_smpl -p 2400000000 -e unhalted_core_cycles:u,instructions_retired:u,baclears  noploop 5
>> >> noploop for 5 seconds
>> >> IIP:0x000000004006d6 PID:5596 TID:5596 TIME:466,210,211,430 STREAM_ID:33 PERIOD:2,400,000,000 ENA=1,010,157,814 RUN=1,010,157,814 NR=3
>> >>       2,400,000,254 unhalted_core_cycles:u (33)
>> >>       2,399,273,744 instructions_retired:u (34)
>> >>       53,340 baclears (35)
>> >>
>> >> Signed-off-by: Stephane Eranian <er...@go...>
>> >>
>> >> ---
>> >>
>> >> diff --git a/kernel/perf_event.c b/kernel/perf_event.c
>> >> index f309e80..04611dd 100644
>> >> --- a/kernel/perf_event.c
>> >> +++ b/kernel/perf_event.c
>> >> @@ -3494,6 +3494,9 @@ static void perf_output_read_group(struct perf_output_handle *handle,
>> >>  static void perf_output_read(struct perf_output_handle *handle,
>> >>                            struct perf_event *event)
>> >>  {
>> >> +     update_context_time(event->ctx);
>> >> +     update_event_times(event);
>> >> +
>> >>       if (event->attr.read_format & PERF_FORMAT_GROUP)
>> >>               perf_output_read_group(handle, event);
>> >>       else
>> >
>> >
>> > Right, except that this can actually corrupt the time measurements... :/
>> >
>> > Usually context times are updated under ctx->lock, and this is called
>> > from NMI context, which can interrupt ctx->lock..
>> >
>> Ok, I missed that. But I don't understand why you need the lock to
>> udpate the time. The lower-level clock is lockless if I recall. Can't you
>> use an atomic ops in update_context_time()?
>
> atomic ops would slow down those code paths, also, I don't think you can
> fully get the ordering between ->tstamp_$foo and ->total_time_$foo just
> right.
>

I don't get that. Could you give an example?

>> > I was thinking about updating a local copy of the times, in that case
>> > you can only get funny times from samples, but it won't corrupt the
>> > actual running data.
>> >
>> You want time to be correct in every sample How would you detect
>> bogus timing?
>
> Not sure, but barring 64bit atomics for all these, 32bit archs and NMI
> are going to be 'interesting'
>

Every sample needs to be correct, otherwise you run the risk of introducing
bias.

I think if the tradeoffs is correctness vs. speed, I'd choose correctness.