RE: Re: [PATCH v2 1/5] perf probe: Split add_perf_probe_events()

[åæéå / HIRAMATUïMASAMI]

>From: Namhyung Kim [mailto:namhyung@xxxxxxxxx] On Behalf Of Namhyung Kim
>
>On Sun, Sep 06, 2015 at 03:47:37PM +0800, Wangnan (F) wrote:
>> Hi Namhyung,
>
>Hi,
>
>>
>> Thanks for this patchset.
>>
>> Could you plase have a look at patch 5/27 and 6/27 in my newest pull
>> request?
>> These 2 patches utilize new probing API to create probe point and collect
>> probe_trace_events. I'm not very sure I fully understand your design
>> principle,
>> especially the cleanup part, because I can see different functions dealing
>> with
>> cleanup:
>>
>> cleanup_perf_probe_events

This is for clearing an array of probe events.

>> del_perf_probe_events

This is not for cleanup, but for removing probes in the kernel.

>> clear_perf_probe_event
>> clear_probe_trace_event

These are the cleanup each event. Ah, right, since now perf_probe_event has probe_trace_events,
clear_perf_probe_event has to call clear_probe_trace_event.

>>
>> But non of them works perfectly for me.
>
>The cleanup_perf_probe_events() is just to keep the existing logic as
>long as possible.  But I think it needs to call
>clear_perf_probe_event().
>
>The del_perf_probe_events() uses strfilter, but I think it can be
>problematic if other instances or users are using similar events at
>the same time.

Yeah, since perf probe doesn't lock the ftrace, there should be a
timing bug, but it can be fixed easily by ignoring -ENOENT. :) 

>So for your case, IMHO it'd better keeping the perf/trace events after
>probing and reusing the events for unprobing.  I'll take a look at it.
>
>
>>
>> In bpf_prog_priv__clear() function of 6/27, I copied some code from
>> cleanup_perf_probe_events(), because I think when destroying bpf programs,
>> the probe_trace_events should also be cleanuped, but we don't need call
>> exit_symbol_maps() many times, because we are in 'perf record', and not
>> sure whether other parts of perf need symbol maps. Otherwise I think
>> directly
>> calling cleanup_perf_probe_events() sould be better.
>
>Yeah, I also think exit_symbol_maps() should not be a part of the
>cleanup.  I'll send a patch soon.

OK.


Thanks!