Re: PATCH? trace_remove_event_call() should fail if call is active

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

(2013/07/03 7:23), Oleg Nesterov wrote:
> On 07/02, Oleg Nesterov wrote:
>>
>> So please ignore modules ;)
> 
> Or lets discuss the change above.

No, I think this still doesn't ensure that we can remove dynamic
event safely. Since the event is related to several files under
events/ dir and buffer instances, someone can just stay open the
files while the event is removed and read/write it.
Perhaps, we need per-event_call refcounter not per-trace_array
one, and do as below.

Open file:
 -> lock event_mutex
 -> find event_call and event_file
 -> get event_call
 -> unlock event_mutex

Close:
 -> lock event_mutex
 -> put event_call
 -> unlock event_mutex

Remove event (via kprobe_events):
 -> lock event_mutex
 -> find event_call
 -> -EBUSY if event is enabled or refcount != 0
    (here, no one accessing the event and not enabled)
 -> unregister_kprobe
 -> remove event
 -> unlock event_mutex

The key is holding event_mutex *and* getting refcount
under any operation.
And of course, we can unregister the kprobe outside of
the event_mutex, but it still need a synchronize_sched
for safety.

 -> lock event_mutex
 -> wait_for_rcu (to ensure no disabled kprobe accesses the event)
 -> find event_call
 -> -EBUSY if event is enabled or refcount != 0
    (here, no one accessing the event and not enabled)
 -> remove event
 -> unlock event_mutex
 -> unregister_kprobe

Thank you,


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