Re: [PATCH 06/14] perf, x86: PEBS infrastructure

[Stephane Eranian <eranian@google.com>]

On Fri, Mar 5, 2010 at 1:20 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> On Fri, 2010-03-05 at 17:19 +1100, Paul Mackerras wrote:
>> On Thu, Mar 04, 2010 at 03:00:52PM +0100, Peter Zijlstra wrote:
>>
>> > Implement a simple PEBS model that always takes a single PEBS event at
>> > a time. This is done so that the interaction with the rest of the
>> > system is as expected (freq adjust, period randomization, lbr).
>> >
>> > Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
>> > LKML-Reference: <new-submission>
>> > ---
>>
>> ...
>>
>> > @@ -203,8 +203,9 @@ struct perf_event_attr {
>> >                             enable_on_exec :  1, /* next exec enables     */
>> >                             task           :  1, /* trace fork/exit       */
>> >                             watermark      :  1, /* wakeup_watermark      */
>> > +                           precise        :  1, /* OoO invariant counter */
>>
>> Could you explain in a bit more detail what this means?
>>
>> Also, it would be good to mention the ABI addition in the patch
>> description, and explain it briefly there.
>
> Quite so, my bad.
>
> So on Intel regular PMIs can happen several instructions later than the
> actual event due to out-of-order processing of the instruction stream,
> that is, it doesn't keep the IP of the actual instruction that triggered
> the event, so all we have is the IP of where the interrupt happened (the
> difference between these IPs is called skid).
>
I don't think the skid is strictly related to OoO. You had skid on Itanium
which was in-order.

> Now Intel came up with something called Precise Event Based Sampling
> (PEBS) which stores a (partial) register set in some memory buffer at
> event time (trap like for some daft reason).
>
> So from that we can obtain the IP of the instruction _after_ the
> instruction that caused the event. This is reliably so (mostly [*]) and
> does not contain out-of-order artifacts (0-skid).
>
The sampled instruction is guaranteed to have caused the event but
it is unrelated to your sampling period, i.e., it can be N events after
your sampling period due to the way PEBS is implemented.

> So the ->precise flag tells us to use a more precise sampling method if
> available on the hardware (AMD could be using IBS to implement this for
> their instruction counter).
>
> If you look at patch 9/14 you'll see we use the Last Branch Recording
> (LBR) facility of the Intel cpus (patch 8/14) to find the last basic
> block in the instruction stream and use that to rewind the instruction
> stream to get the actual instruction that triggered the event. In case
> that works I also set PERF_RECORD_MISC_EXACT to indicate we got the IP
> dead on (mostly [*]).
>
By rewinding you get the IP of the instruction that caused the event. But the
register state is STILL reflecting the situation at retirement of that
instruction.

> I suspect CPUs that are strictly in-order, like Atom, might always have
> it right, but I need to validate that.
>
Not sure about that.