Re: [PATCH v3 1/3] perf: Allow using AUX data in perf samples

From: Alexander Shishkin
Date: Mon Nov 04 2019 - 05:40:12 EST


Peter Zijlstra <peterz@xxxxxxxxxxxxx> writes:

> On Mon, Oct 28, 2019 at 07:08:18PM +0200, Alexander Shishkin wrote:
>
>> > @@ -6318,11 +6318,12 @@ static void perf_aux_sample_output(struc
>> >
>> > /*
>> > * Guard against NMI hits inside the critical section;
>> > - * see also perf_aux_sample_size().
>> > + * see also perf_prepare_sample_aux().
>> > */
>> > WRITE_ONCE(rb->aux_in_sampling, 1);
>> > + barrier();
>>
>> Isn't WRITE_ONCE() barrier enough on its own? My thinking was that we
>> only need a compiler barrier here, hence the WRITE_ONCE.
>
> WRITE_ONCE() is a volatile store and (IIRC) the compiler ensures order
> against other volatile things, but not in general.
>
> barrier() OTOH clobbers all of memory and thereby ensures nothing can
> get hoised over it.
>
> Now, the only thing we do inside this region is an indirect call, which
> on its own already implies a sync point for as long as the compiler
> cannot inline it, so it might be a bit paranoid on my end (I don't think
> even LTO can reduce this indirection and cause inlining).

I see what you mean. I was only thinking about not having to order the
AUX STOREs vs the rb->aux_in_sampling. Ordering the call itself makes
sense.

Thanks,
--
Alex