Re: how to get a kernel callgraph starting from a syscall

[Francis Moreau <francis.moro@gmail.com>]

On Sat, Apr 23, 2011 at 10:40 AM, Stefan Hajnoczi <stefanha@gmail.com> wrote:
> On Fri, Apr 22, 2011 at 12:46 PM, Francis Moreau <francis.moro@gmail.com> wrote:
>> On Fri, Apr 22, 2011 at 7:29 AM, Francis Moreau <francis.moro@gmail.com> wrote:
>>> On Fri, Apr 22, 2011 at 2:47 AM, David Ahern <dsahern@gmail.com> wrote:
>>>> On 04/21/11 14:07, Francis Moreau wrote:
>>>>> On Thu, Apr 21, 2011 at 9:49 PM, David Ahern <dsahern@gmail.com> wrote:
>>>>>>
>>>>>> You want the individual samples? if so, perf script dumps them and you
>>>>>> can see the callchain for each sys_read.
>>>>>>
>>>>>
>>>>> hmm, I don't see what you mean by individual samples.
>>>>>
>>>>> I don't think sampling can help in my case (tracing the callchain of
>>>>> one syscall).
>>>>>
>>>>> Could you give me an example of what you have in mind ?
>>>>>
>>>>> Thanks
>>>>
>>>> I think I get it now. 'perf record' for syscall event stops at sys_read
>>>> as that is the point the event is generated. You want where the kernel
>>>> goes starting with sys_read. In that case you are limited to the
>>>> pre-existing tracepoints (see 'perf list -e tracepoints') to trigger an
>>>> event and back trace, or if the read causes the processes to block you
>>>> can you use the context-switch event or sched_switch trace point.
>>>>
>>>> You could also insert probe points using perf probe. Lin Ming posted an
>>>> example of this a couple of weeks ago:
>>>> # ./perf probe -k ~/vmlinux -s ~/linux-2.6/  find_get_page
>>>> # ./perf record -e probe:find_get_page -f -g -a
>>>>
>>>
>>> Thanks for you answer.
>>>
>>> I think I'll use probe.
>>
>>
>> Thinking more about it, I think that both solution (probe or
>> tracepoint) are not adapted.
>>
>> Because I'm interested in discovering the path taken by the kernel
>> from a given syscall, therefore I don't know this path. However both
>> solution imply that the path is known since you need to insert some
>> 'probes' in this path to see it...
>
> You can use "Dynamic ftrace with the function graph tracer"
> (set_graph_function):
> http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=blob;f=Documentation/trace/ftrace.txt;hb=HEAD#l1664
>
> I don't know of way to do this using perf(1).

I don't know either but since perf interface is more convenient than
sysfs one, I was wondering if that was possible with perf.

Thanks
-- 
Francis