Re: [PATCH v2 2/5] perf: Free aux pages in unmap path

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Mar 14, 2016 at 04:04:44PM +0200, Alexander Shishkin wrote:
> Peter Zijlstra <peterz@infradead.org> writes:

> >> +static void perf_pmu_output_stop(struct perf_event *event)
> >> +{
> >> +	int cpu, err;
> >> +
> >> +	/* better be thorough */
> >> +	get_online_cpus();
> >> +restart:
> >> +	for_each_online_cpu(cpu) {
> >> +		err = cpu_function_call(cpu, __perf_pmu_output_stop, event);
> >> +		if (err)
> >> +			goto restart;
> >> +	}
> >> +	put_online_cpus();
> >> +}
> >
> > This seems wildly overkill, could we not iterate rb->event_list like we
> > do for the normal buffer?
> 
> Actually we can. One problem though is that iterating rb::event_list
> requires rcu read section or irqsafe rb::event_lock and we need to send
> IPIs. 

We should be able to send IPIs with rcu_read_lock() held; doing so with
IRQs disabled is a bit harder.

> The normal buffer case tears down the rb::event_list as it goes,
> so it can close the rcu read section right after it fetches one event
> from it. In this case however, we must keep the list intact.

Yep..

> > Sure, we need to IPI for each event found, but that seems better than
> > unconditionally sending IPIs to all CPUs.
> 
> Actually, won't it "often" be the case that the number of events will be
> a multiple of the number of cpus? The usual use case being one event per
> task per cpu and inheritance enabled. In this case we'll zap multiple
> events per IPI.

Right, but then each event (or set thereof) will be for one particular
CPU. So for the one munmap() you typically only end up sending IPIs to
that one CPU.

If OTOH you send IPIs to all CPUs for all events, you end up with n^2
IPIs, because for each CPUs munmap() you send IPIs to all other CPUs.