On Thu, 2013-04-11 at 13:59 +0200, Oleg Nesterov wrote:
> On 04/11, Masami Hiramatsu wrote:
> >
> > (2013/04/10 23:58), Oleg Nesterov wrote:
> >
> > > And... Cough, another question ;) To simplify, lets discuss
> > > kprobe_perf_func()
> > > only. Suppose that a task hits the kprobe but this task/cpu doesn't have
> > > a counter. Can't we avoid perf_trace_buf_prepare/submit in this case?
> > > IOW, what do you think about the change below?
> >
> > Hmm, I'm not so sure how frequently this happens.
>
> Suppose that you do, say, "perf record -e probe:some_func workload". Only
> "workload" will have the active counter, any other task which hits the
> probed some_func() will do perf_trace_buf_prepare/perf_trace_buf_submit
> just to realize that nobody wants perf_swevent_event().
Wow, you're right. Seems that perf goes through a lot of work for every
time a tracepoint is hit for *all tasks*.
>
> Simple test-case:
>
> #include <unistd.h>
>
> int main(void)
> {
> int n;
>
> for (n = 0; n < 1000 * 1000; ++n)
> getppid();
>
> return 0;
> }
>
> Without kprobe:
>
> # time ./ppid
>
> real 0m0.663s
> user 0m0.163s
> sys 0m0.500s
>
> Activate the probe:
>
> # perf probe sys_getppid
>
> # perf record -e probe:sys_getppid sleep 1000 &
> [1] 546
>
> Test it again 3 times:
>
> # time ./ppid
>
> Before the patch:
>
> real 0m9.727s
> user 0m0.177s
> sys 0m9.547s
>
> real 0m9.752s
> user 0m0.180s
> sys 0m9.573s
>
> real 0m9.761s
> user 0m0.187s
> sys 0m9.573s
>
> After the patch:
>
> real 0m9.605s
> user 0m0.163s
> sys 0m9.437s
>
> real 0m9.592s
> user 0m0.167s
> sys 0m9.423s
>
> real 0m9.613s
> user 0m0.183s
> sys 0m9.427s
>
> So the difference looks measurable but small, and I did the testing
> under qemu so I do not really know if we can trust the numbers.
>
> > And, is this right way to
> > handle that case?
>
> If only I was sure ;) I am asking.
>
> And, to clarify, it is not that I think this change can really
> improve the perfomance. Just I am trying to understand what I have
> missed.
>
> > If so, we can do same thing also on trace_events.
> > (perf_trace_##call in include/trace/ftrace.h)
>
> Yes, yes, this is not kprobe-specific. It seems that more users of
> perf_trace_buf_submit() could be changed the same way.
Yeah, looks like include/trace/ftrace.h needs an update.
Frederic?
-- Steve
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
