Re: [PATCH v4 0/4] perf timechart io mode

[Jiri Olsa <jolsa@redhat.com>]

On Wed, Jun 25, 2014 at 08:01:17AM -0600, David Ahern wrote:
> On 6/25/14, 5:20 AM, Stanislav Fomichev wrote:
> >>>maybe we dont need to fail in this case.. seems like it should
> >>>not be hard to detect, wanr and recover? ;-)
> >The easiest way is just to bail out and don't try to be smart, that's
> >what I do now.
> >
> >Another approach would be to just convert all my consistency
> >checks to warnings and skip invalid events (and hope it somehow works). I
> >think it makes sense when for some reason we have perf.data which
> >is not really consistent but we still want to get something out of it.
> 
> If the event type does not make sense how can you trust any other field --
> like event size?
> 
> >
> >>seems to me it should not be happening at all. It seems like the head caught
> >>the tail and was not properly detected.
> >So you mean it perf record's fault?
> >
> 
> In my cases I am running perf-record. At the end of the session it walks the
> events and in too many cases I see the message:
> 
>  0x113f80 [0x8]: failed to process type: 68
> 

hum.. 68 is PERF_RECORD_FINISHED_ROUND, so there was something wrong
during the queue flush.. thus processing all the queue events and
failing only if either the parsing or delivery fails

jirka