Re: Re: Re: Regarding systemtap support for AArch64

[Sandeepa Prabhu <sandeepa dot prabhu at linaro dot org>]

On 15 October 2013 15:09, Masami Hiramatsu
<masami.hiramatsu.pt@hitachi.com> wrote:
> (2013/10/07 20:12), Sandeepa Prabhu wrote:
>>>>>>>>>  - Is it really need to use spinlock to protect break_hook?
>>>>>>>> Any cpu can remove breakpoint hooks right, and traversal happen in
>>>>>>>> debug exception context so mutex are not safe (can sleep/schedule out)
>>>>>>>> in debug exception.
>>>>>>>
>>>>>>> I don't think we need to remove the breakpoint hooks after starting
>>>>>>> up the kernel. If we use the spinlock there, we'll pay a big cost
>>>>>>> because of the lock contention.
>>>>>> Not in kprobes. But kgdb can remove breakpoint handler and use same
>>>>>> API. or atleast while providing an api we should not assume race
>>>>>> cannot happen right?
>>>>>
>>>>> In that case, we'd better add a wrapper handler for kgdb so that
>>>>> the list isn't updated even if the kgdb removes its handler.
>>>>>
>>>>>> And there wont be much lock contention, i'ts only if the debug
>>>>>> framework (like kgdb) is wrapping-up, not is normal use-case.
>>>>>
>>>>> Hmm, it seems that the spinlock is locked while handling a breakpoint.
>>>>> This will cause a bad performance issue when we put many kprobes
>>>>> on SMP system.
>>>> arm maintainers prefer a reader/writer spin-locks, so there wont be
>>>> lock contention in debug path, each instance of kprobe hook trap (on
>>>> any CPU) would be a reader, not blocking.
>>>
>>> OK for the first step, and it eventually should be fixed to lockless.
>>> (depends on the performance improvement)
>> Hmm, is there a performance requirement for systemtap or perf? -like
>> how much time each test suite should consume etc?
>
> Basically, for the enterprise use, we aims to get less than 5% loss
> of runtime performance. Of course it depends on the configuration.
> This requirement comes from the usage of tracing, it's usually used
> as a "flight-recorder" in such system. For analyzing the root cause
> of the trouble, some fundamental events are always recorded into a
> memory buffer. When encountering a trouble, the buffer will be dumped,
> and trouble shooting team analyzes it.
>
> Thus, I'd like to make the performance overhead of tracing as
> small as possible.
Hmm, my worry is whether we can really measure and improve performance
or not -running on foundation model, do not have real hardware access
right now :(

>
> However, for debugging use, the performance degradation is not
> so important.
>
>> Want to know the acceptance criteria for systemtap or perf to say
>> 'kprobes/uprobes on an architecture' is complaint and good enough for
>> tracing?
>
> I think there is no such criteria. The overhead problem depends on the
> use-cases as I said above. If it is functional, it's enough to use by
> perf/ftrace ;) Performance optimization can be done afterwords.
>
> Thank you,
>
> --
> Masami HIRAMATSU
> IT Management Research Dept. Linux Technology Center
> Hitachi, Ltd., Yokohama Research Laboratory
> E-mail: masami.hiramatsu.pt@hitachi.com
>
>