Re: [RFC][PATCH] perf_events, x86: PEBS support

[Ingo Molnar <mingo@elte.hu>]

* Peter Zijlstra <peterz@infradead.org> wrote:

> Totally uncompiled and untested, but it looks to be about mostly there so I 
> thought I'd post it.
> 
> One of the things that is missing is keeping the count value sane while 
> using PEBS -- another is dealing with auto frequency things, I thought 
> about single shot PEBS assist for that.
> 
> After this we can do something like PERF_SAMPLE_REGS, but for that we need 
> to think about how to expose pt_regs to userspace or something (or maybe it 
> already is, I haven't checked).
> 
> Also, initially I'll go through all the other hw perf implementations 
> (powerpc, sparc, arm, sh) and make then refuse to create attr.precise 
> counters -- precise meaning the reported IP is not influenced by OoO 
> artefacts.
> 
> Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
> ---
>  arch/x86/kernel/cpu/perf_event.c |  354 ++++++++++++++++++++++++++++++++++-----
>  include/linux/perf_event.h       |    4 
>  2 files changed, 314 insertions(+), 44 deletions(-)

Very nice!

> @@ -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,

I think we want to default to precise events even if not specifically 
requested by user-space, in the cases where that's possible on the CPU 
without additional limitations.

That way people will default to better (and possibly cheaper) PEBS profiling 
on modern Intel CPUs.

	Ingo