Re: use-after-free in __perf_install_in_context

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Dec 08, 2015 at 07:57:38PM +0100, Ingo Molnar wrote:
> Btw., could we add more redundancy / debug code to the refcounting code? It seems 
> to be a frequent source of very hard to find/fix races/bugs - so it should be ripe 
> for some extra debug infrastructure ...

I'll try, but its not easy. The biggest problem so far is the
ctx::parent relation. We cannot easily track the reverse of that because
of locking (we'd need to hold two ctx->lock, and we cannot because
perf_event_context_sched_out()).

The other two relations we have reverse maps for:

  task->perf_event_ctxp[] <-> ctx->task
  event->ctx              <-> ctx->event_list

Also, all 3 relations are (more or less) protected under ctx->lock:

  - unclone_ctx():
	removes the ctx->parent link, with ctx->lock held

  - perf_remove_from_context():
	removes the event from ctx::event_list, with ctx->lock held;
	leaves event->ctx set, because nasty games with ctx migration,
	it will get set to the new context, but can never be NULL.

  - perf_event_exit_task_context():
	clears task->perf_event_ctxp[], with ctx->lock held;
	leaves ctx->task set, still looking into this.