Re: [PATCH 0/5][RFC] kprobes/ftrace: Have kprobes use ftrace on ftrace nops

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

(2011/08/11 22:01), Steven Rostedt wrote:
> On Thu, 2011-08-11 at 15:28 +0900, Masami Hiramatsu wrote:
>> (2011/08/11 9:34), Steven Rostedt wrote:
>>> On Thu, 2011-08-11 at 09:21 +0900, Masami Hiramatsu wrote:
>>>> Hi Steven,
>>>>
>>>> Thanks for this nice feature!
>>>>
>>>> (2011/08/11 1:22), Steven Rostedt wrote:
>>>>> Hi All,
>>>>>
>>>>> I started working on adding the -mfentry switch to ftrace, which
>>>>> allows us to remove the frame pointers requirement from function tracing
>>>>> as well as makes mcount (fentry) work just better.
>>>>>
>>>>> But when I did this in another branch, I noticed that I broke kprobes
>>>>> in its most common usage. The attaching a probe at the beginning of
>>>>> a function to use get to its parameters.
>>>>>
>>>>> So I started this branch. This branch is to have kprobes use ftrace
>>>>> directly when a probe is attached to a ftrace nop. Currently, kprobes
>>>>> will just error when that happens. With this patch set, it will hook
>>>>> into the ftrace infrastructure and use ftrace instead. This is more
>>>>> like an optimized probe as no breakpoints need to be set. A call to
>>>>> the function is done directly via the mcount trampoline. If ftrace
>>>>> pt_regs is implemented for an arch, kprobes gets this feature for free.
>>>>
>>>> I agreed this idea, this looks good to me too :)
>>>> With -fentry, this can improve dynamic trace events very much.
>>>>
>>>> BTW (OT), it seems that current kprobe data structure becomes a bit
>>>> fat. Maybe what we need is just a "holder of hooking handler" as
>>>> what ftrace provides, not a full storage data structure of copied
>>>> instrucutions. Perhaps, we'd better diet the kprobe structure for
>>>> transparency of hooking infrastructure.
>>>
>>> Sure, I can make the ftrace_ops field in kprobes dynamically allocated
>>> instead. That shouldn't be an issue.
>>
>> By the way (again), perhaps, much simpler solution is using ftrace
>> not in kprobe, but in the trace_kprobe. Of course, there are several
>> pros and cons...
>>
>> The pros:
>> - Arch independent solution (anyway, ftrace still needs passing pt_regs
>>  to their handler)
>> - Don't need to introduce more complexity into kprobes itself.
>> - Maybe systemtap also can catch up with this as using same method.
>
> Note that systemtap and others will be hooking into kprobes version, not
> the trace_kprobe one. If we do it in trace_kprobe, then everyone else
> needs to reimplement it too.  I have bigger ideas for the future of
> this, and I really want to get this working. If it doesn't work for
> kprobes, then it won't work for anything else.

I don't think it won't work. It can work but on a long way.
Could you tell me your "bigger ideas"? Perhaps, we are on the different
way but aim to same goal.

>> The cons:
>> - Native kprobes users will be disappointed... anyway, they just need to
>>   move their probes to the next instruction (usually addr+=5 is OK).
>
> I've been told that doing the addr+=5 (which is also arch specific) can
> break things for other tools.

As I told in previous mail, I think kprobes can do that transparently.

Thank you,

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