Re: [PATCH v5] perf: Use monotonic clock as a source for timestamps

[Adrian Hunter]

On 12/02/15 12:28, Peter Zijlstra wrote:
> On Thu, Feb 12, 2015 at 12:04:54PM +0200, Adrian Hunter wrote:
>> On 11/02/15 18:12, Peter Zijlstra wrote:
>>>
>>> How about something like the below? I _think_ it should mostly work for
>>> x86, where the tsc is a 64bit wide cycle counter.
>>
>> It would have to be based on CLOCK_MONOTONIC_RAW not CLOCK_MONOTONIC 
> 
> Why?

In the CLOCK_MONOTONIC case, the components of the calculation (mult and
shift etc) are subject to change, so the calculation would be increasingly
inaccurate the greater the time between reading those values the reading TSC
or capturing perf events.

Accuracy is important for matching sideband events against Intel PT. e.g.
did a mmap event happen before or after a given TSC timestamp.

Adding another sample value (Pawel's patch 3) is more accurate and simpler
to understand. It just needs to be extended to allow TSC.

> 
>> and you would have to check the clocksource is TSC.
> 
> It implicitly does that; it has that sched_clock_stable() thing, but
> yeah I suppose someone could change the clocksource even though the tsc
> is stable.
> 
> Not using TSC when its available is quite crazy though.. but sure.
> 
>> Why is CLOCK_MONOTONIC preferred anyway - I would have thought any
>> adjustment would skew performance timings?
> 
> Because you can do inter-machine stuff with MONOTONIC and that's
> entirely impossible with MONO_RAW.

Ok, the man page does not make it sound as bad as that.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/