Re: [RFC][PATCH 3/4] tracing/kprobes: Fail to unregister if probe event files are open

[Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>]

(2013/07/05 3:48), Oleg Nesterov wrote:
> On 07/04, Masami Hiramatsu wrote:
>>
>> (2013/07/04 12:33), Steven Rostedt wrote:
>>> +	/* Will fail if probe is being used by ftrace or perf */
>>> +	if (unregister_probe_event(tp))
>>> +		return -EBUSY;
>>> +
>>>  	__unregister_trace_probe(tp);
>>>  	list_del(&tp->list);
>>> -	unregister_probe_event(tp);
>>>
>>>  	return 0;
>>>  }
>>
>> This may cause an unexpected access violation at kprobe handler because
>> unregister_probe_event frees event_call/event_files and it will be
>> accessed until kprobe is *completely* disabled.
> 
> I don't think so... Please correct me.
> 
> (but yes I think the patch needs a small update, see below).
> 
>> Actually disable_kprobe() doesn't ensure to finish the current running
>> kprobe handlers.
> 
> Yes. in fact disable_trace_probe(file != NULL) does, but perf doesn't.

Ah, right. we did that.

> 
>> Thus, even if trace_probe_is_enabled() returns false,
>> we must do synchronize_sched() for waiting, before unregister_probe_event().
> 
> No, I think we should simply kill trace_probe_is_enabled() here.
> And synchronize_sched() _before_ unregister_probe_event() can't
> help, exactly because trace_probe_is_enabled() is racy.

Right, it should be useless.


>> OTOH, unregister_kprobe() waits for that.
> 
> Yes.
> 
> So I think we only need to move kfree(tp->call.print_fmt). In fact I
> already wrote the patch assuming that trace_remove_event_call() will
> be changed as we discussed.
> 
> So the sequence should be:
> 
> 	if (trace_remove_event_call(...))
> 		return;
> 
> 	/* does synchronize_sched */
> 	unregister_kprobe();
> 
> 	kfree(everything);
> 
> Agreed?

If we can free everything after all, I'd like to do so.
Hmm, but AFAICS, trace_remove_event_call() supposes that
all event is disabled completely.

A safe way is to wait rcu always right after disable_*probe
in disable_trace_probe. If we have an unused link, we can
free it after that.

Or, do more aggressively, introducing a dying-bit for each
trace-probe could be another way. If the bit is set, all
enable operations are failed. It works like as a per-event lock.

Thank you,

-- 
Masami HIRAMATSU
IT Management Research Dept. Linux Technology Center
Hitachi, Ltd., Yokohama Research Laboratory
E-mail: masami.hiramatsu.pt@hitachi.com