Re: hit a KASan bug related to Perf during stress test

[Peter Zijlstra]

On Mon, Oct 24, 2016 at 01:15:27PM +0200, Oleg Nesterov wrote:
> On 10/24, Peter Zijlstra wrote:
> >
> > > [32738.867020]  [<ffffffff810d9975>] task_tgid_nr_ns+0x35/0xb0
> >
> > So here we did: perf_event_[pt]id(event, current);
> >
> > How can _current_ not be valid anymore?
> 
> ...
> 
> > > [32739.040207]  [<ffffffff81135a4c>] __call_rcu+0x12c/0x450
> >
> > And while we just called release_task(), that call_rcu() should still be
> > pending at this point,
> 
> Yes, current is still valid.
> 
> But nothing protects current->group_leader or parent/real_parent, they
> can point to the exited/freed task. We really need to nullify them in
> __unhash_process() to catch the problems like this, I wanted to do this
> many times...
> 
> So you simply can't know your tgid or even tid after release_task() calls
> __unhash_process(). Actually after exit_notify() unless the exiting task
> autoreaps itself.
> 
> How about the trivial fix below?
> 
> Oleg.
> 
> --- x/kernel/events/core.c
> +++ x/kernel/events/core.c
> @@ -1257,7 +1257,7 @@ static u32 perf_event_pid(struct perf_ev
>  	if (event->parent)
>  		event = event->parent;
>  
> -	return task_tgid_nr_ns(p, event->ns);
> +	return pid_alive(p) ? task_tgid_nr_ns(p, event->ns) : 0;
>  }

Hurm.. should we not push this into task_tgid_nr_ns() ? I mean, now the
user needs to be aware of this dinky detail.