Re: [PATCH 4/7] perf: Free aux pages in unmap path

[Peter Zijlstra <peterz@infradead.org>]

On Fri, Dec 04, 2015 at 11:17:23PM +0100, Peter Zijlstra wrote:
> On Fri, Dec 04, 2015 at 06:02:06PM +0100, Peter Zijlstra wrote:
> > The only solution I've come up with is:
> > 
> > 	struct rb_aux *aux = rb->aux;
> > 
> > 	if (aux && vma->vm_pgoff == aux->pgoff) {
> > 		ctx = perf_event_ctx_lock(event);
> 
> Can't do this at all, see the comment in put_event(). perf_read_group()
> accesses user memory (and hence causes faults, which in turn take
> mmap_sem) while holding ctx::mutex.
> 
> So neither this, not what you proposed can work.
> 
> Will need moar thinking.

So we could try and see if we can get this working:

static int __perf_event_stop(void *info)
{
	struct perf_event *event = info;

	/* IRQs disabled, cannot get scheduled away */
	if (event->oncpu == smp_processor_id()) {
		event->pmu->stop(event);
		return 0;
	}

	return -EAGAIN;
}

perf_event_stop(struct perf_event *event)
{
	for (;;) {
		if (READ_ONCE(event->state) != PERF_EVENT_STATE_ACTIVE)
			break;

		smp_rmb(); /* if we see ACTIVE, ->oncpu must be set */

		if (!cpu_function_call(READ_ONCE(event->oncpu), __perf_event_stop, event))
			break;
	}
}


That probably wants some WRITE_ONCE() and maybe some memory barriers in
event_sched_in() as well, like:

	WRITE_ONCE(event->oncpu, smp_processor_id());
	smp_wmb();
	WRITE_ONCE(event->state, PERF_EVENT_STATE_ACTIVE);