Re: [PATCH -tip -v10 7/7] tracing: add kprobe-based event tracer

[Masami Hiramatsu <mhiramat at redhat dot com>]

Masami Hiramatsu wrote:
> Frederic Weisbecker wrote:
>> On Tue, Jul 07, 2009 at 05:31:25PM -0400, Masami Hiramatsu wrote:
>>> Frederic Weisbecker wrote:
>>>> On Tue, Jul 07, 2009 at 04:42:32PM -0400, Masami Hiramatsu wrote:
>>>>> Frederic Weisbecker wrote:
>>>>>> On Tue, Jul 07, 2009 at 03:55:28PM -0400, Masami Hiramatsu wrote:
>>>>>>> Frederic Weisbecker wrote:
>>>>>>>>> diff --git a/kernel/trace/trace.h b/kernel/trace/trace.h
>>>>>>>>> index 206cb7d..65945eb 100644
>>>>>>>>> --- a/kernel/trace/trace.h
>>>>>>>>> +++ b/kernel/trace/trace.h
>>>>>>>>> @@ -45,6 +45,8 @@ enum trace_type {
>>>>>>>>>  	TRACE_POWER,
>>>>>>>>>  	TRACE_BLK,
>>>>>>>>>  	TRACE_KSYM,
>>>>>>>>> +	TRACE_KPROBE,
>>>>>>>>> +	TRACE_KRETPROBE,
>>>>>>>>>  
>>>>>>>>>  	__TRACE_LAST_TYPE,
>>>>>>>>>  };
>>>>>>>>> @@ -227,6 +229,22 @@ struct trace_ksym {
>>>>>>>>>  	char			ksym_name[KSYM_NAME_LEN];
>>>>>>>>>  	char			p_name[TASK_COMM_LEN];
>>>>>>>>>  };
>>>>>>>>> +#define TRACE_KPROBE_ARGS 6
>>>>>>>>> +
>>>>>>>>> +struct kprobe_trace_entry {
>>>>>>>>> +	struct trace_entry	ent;
>>>>>>>>> +	unsigned long		ip;
>>>>>>>>> +	int			nargs;
>>>>>>>>> +	unsigned long		args[TRACE_KPROBE_ARGS];
>>>>>>>> I see that you actually make use of arg as a dynamic sizeable
>>>>>>>> array.
>>>>>>>> For clarity, args[TRACE_KPROBE_ARGS] could be args[0].
>>>>>>>>
>>>>>>>> It's just a neat and wouldn't affect the code nor the data
>>>>>>>> but would be clearer for readers of that code.
>>>>>>> Hmm. In that case, I think we'll need a new macro for field
>>>>>>> definition, like TRACE_FIELD_ZERO(type, item).
>>>>>> You mean that for trace_define_field() to describe fields of events?
>>>>>> Actually the fields should be defined dynamically depending on how
>>>>>> is built the kprobe event (which arguments are requested, how many,
>>>>>> etc..).
>>>>> Yeah, if you specified a probe point with its event name, the tracer
>>>>> will make a corresponding event dynamically. There are also anonymous
>>>>> probes which don't have corresponding events. For those anonymous
>>>>> probes, I need to define two generic event types(kprobe and kretprobe).
>>>>>
>>>>> Thank you,
>>>> Ok. Btw, why do you need to define those two anonymous events?
>>>> Actually your event types are always dynamically created.
>>>> Those you defined through TRACE_FORMAT_EVENT are only "ghost events",
>>>> they only stand there as a abstract pattern, right?
>>>>
>>> Not always created.
>>>
>>> Below command will create an event "event1";
>>> p probe_point:event1 a1 a2 a3 ... > /debug/tracing/kprobe_events
>>>
>>> But next command doesn't create.
>>> p probe_point a1 a2 a3 ... > /debug/tracing/kprobe_events
>>
>> Aah, ok.
>>
>>
>>> This just inserts a kprobe to probe_point. the advantage of this
>>> "simple" command is that you never be annoyed by making different
>>> name for new events :-)
>>
>> Indeed.
>> But speaking about that, may be you could dynamically create a name
>> following this simple model: func+offset

hmm, and we have two probe types, p(robe) and r(et probe).
so, event name should be type@func+offset or type@address.

>> Unless we can set several kprobes on the exact same address?
> 
> Actually, we can...
> I thought that someone might want to insert events in the same
> address for retrieving more than 6 arguments.

Anyway, I can improve the interface according to user's voice.
If you have good idea, I'm happy to hear that:-)

Thank you,

-- 
Masami Hiramatsu

Software Engineer
Hitachi Computer Products (America), Inc.
Software Solutions Division

e-mail: mhiramat@redhat.com