Re: [PATCH 2/3] tracing/kprobes: Kill probe_enable_lock

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

(2013/06/18 0:18), Oleg Nesterov wrote:
> On 06/17, Masami Hiramatsu wrote:
>>
>> (2013/06/17 2:21), Oleg Nesterov wrote:
>>> enable_trace_probe() and disable_trace_probe() should not worry about
>>> serialization, the caller (perf_trace_init or __ftrace_set_clr_event)
>>> holds event_mutex.
>>>
>>> They are also called by kprobe_trace_self_tests_init(), but this __init
>>> function can't race with itself or trace_events.c
>>
>> Right,
>> For safety, we should comment this at the caller side,
> 
> Which caller do you mean?

I meant the caller was kprobe_test_self_tests_init().
Since that function calls enable/disable_trace_probe()
without holding event_mutex, we need to notice that
(this is safe because there is no race) at the calling
places :)

Thank you,

> 
> The patch adds
> 
> 	/*
> 	 * This and enable_trace_probe/disable_trace_probe rely on event_mutex
> 	 * held by the caller, __ftrace_set_clr_event().
> 	 */
> 
> above trace_probe_nr_files() but the next patch removes this function
> with the comment...
> 
> Will you agree with this patch if I add something like
> 
> 	/*
> 	 * called by perf_trace_init() or __ftrace_set_clr_event() under event_mutex
> 	 */
> 
> above kprobe_register() ? Perhaps it makes sense to add
> lockdep_assert_held(&event_mutex) into the body?
> 
> And:
> 
>> because
>> those calls are the reason why I have introduced this lock.
> 
> Please do not hesitate to nack this patch if you think that we should
> keep probe_enable_lock for safety even if it is not currently needed.
> In this case I'd suggest to move lock/unlock into kprobe_register()
> but this is minor.
> 
> Oleg.
> 
> 


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