Message-ID: <54E607AE.8080707@intel.com>
Date:	Thu, 19 Feb 2015 17:56:30 +0200
From:	Adrian Hunter <adrian.hunter@...el.com>
To:	Peter Zijlstra <peterz@...radead.org>
CC:	Ingo Molnar <mingo@...hat.com>,
	Arnaldo Carvalho de Melo <acme@...nel.org>,
	linux-kernel@...r.kernel.org, David Ahern <dsahern@...il.com>,
	Frederic Weisbecker <fweisbec@...il.com>,
	Jiri Olsa <jolsa@...hat.com>,
	Namhyung Kim <namhyung@...il.com>,
	Paul Mackerras <paulus@...ba.org>,
	Stephane Eranian <eranian@...gle.com>,
	John Stultz <john.stultz@...aro.org>,
	Thomas Gleixner <tglx@...utronix.de>,
	Pawel Moll <pawel.moll@....com>,
	Steven Rostedt <rostedt@...dmis.org>,
	Andi Kleen <ak@...ux.intel.com>
Subject: Re: [PATCH 0/2] perf/x86: Add ability to sample TSC

On 19/02/2015 5:05 p.m., Peter Zijlstra wrote:
> On Thu, Feb 19, 2015 at 04:38:57PM +0200, Adrian Hunter wrote:
>> On 19/02/15 15:50, Peter Zijlstra wrote:
>>> On Thu, Feb 19, 2015 at 02:11:08PM +0200, Adrian Hunter wrote:
>>>> Hi
>>>>
>>>> With the advent of switching perf_clock to CLOCK_MONOTONIC,
>>>> it will not be possible to convert perf_clock directly to/from
>>>> TSC. So add the ability to sample TSC instead.
>>>
>>> Well, you can, mostly. MONOTONIC is only affected by NTP slew rate
>>> changes, not offset changes.
>>
>> man page says is also subject to adjtime(3)
>
> which is slew adjustment; read the adjtime manpage :-)
>
>>> And NTP limits the slew rate to 500 PPM, so even if you would get a
>>
>> Assuming it is not broken.
>
> NTP people are a cautious crowd, sure they get it wrong just like the
> rest of us, but mostly it needs to work.
>
>>> slew change and then not update the userpage data for a second you'd be
>>> maximally off by 0.0005 seconds.
>>
>> That could still be enough to break the decoder. It will certainly
>> misrepresent the order of events, which is a big loss of information.
>
> What decoder? perf report is already subject to much larger shifts in
> time if you run it on say a core2 machine.

Any decoder of Intel PT data. Side-band events like sched_switch or mmap
have to be sync'ed with Intel PT TSC timestamps to decode the trace. But
synchronizing any kind of event could be useful for analysis.

>
>>> And that is way below what the current perf clock guarantees on funny
>>> hardware.
>>>
>>> If you're really worried about this; we could maybe get John and Thomas
>>> to allow us a callback on every slew change so we can update the
>>> userpage data ASAP, much reducing the max error.
>>>
>>> Say it takes a 10e5 cycles to update your userpage, then you're never
>>> further off than 50 cycles, which is below your ART multiplier.
>>
>> You still need to wake up user space to read the userpage.
>
> Uhm what? Userspace is already awake.

For Intel PT recording, perf record will be sleeping on poll().

>
>>> Does that really matter? Also, if you have a stable crystal, the slew
>>> rate change should be minimal and infrequent, never getting you close to
>>> these numbers.
>>>
>>> So no, I'm not convinced we need this.
>>
>> Adding TSC to the sample is a lot simpler and more accurate.
>
> Finding multiple samples and interpolating between them is much simpler
> than reading tsc and doing the mult, shift and offset addition?
>
> I suspect you're talking about something else entirely; your changelogs
> are inadequate for they tell ntohing of your usecase and have me
> guessing. Don't do that.

Sorry. I did mention Intel PT in patch 2, but I basically assumed the
need to synchronize events with other time sources was understood.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists