Re: [PATCH] perf timechart: dynamically determine event data offset

[Arnaldo Carvalho de Melo <acme@ghostprotocols.net>]

Em Wed, Nov 27, 2013 at 11:17:43PM +0900, Namhyung Kim escreveu:
> 2013-11-27 (수), 10:44 -0300, Arnaldo Carvalho de Melo:
> > Same signature, 32-bit, 64-bit userland, so whoever wrote timechart, Arjan, I
> > think, made no mistakes at using the kernel exported interface, choosing the
> > most efficient way to extract the data, casting to a struct.

> > Yeah, using the interface that searches for a field name to get the offset and
> > then add it to a pointer to then cast to the type will allow maximum flexibility,
> > but is not really efficient.

> > Doing that for something that is not performance critical (probably) as
> > timechart probably is not a problem, but so is not a problem using the patch
> > that does the cast after finding the offset of the first non-common field.
 
> I'm not sure how it affects the performance really.

Hey, everytime we need to get the value of a field when processing each
tracepoint sample we need to do it via:

struct format_field *pevent_find_field(struct event_format *event, const char *name)
{
        struct format_field *format;

        for (format = event->format.fields; format; format = format->next) {
                if (strcmp(format->name, name) == 0)
                        break;
        }

        return format;
}

I don't think this is optimal, no.

But yeah, for things that don't have performance needs, don't process that many
samples and hey, machines are fast and cheap these days ;-)

> But in most cases, using the libtraceevent would be a better solution IMHO.

It is the one that gives the most flexibily to change the kernel exported
tracepoint interface at will, yes.
 
> > Having said that, I'll take the latest patch, using perf_evsel__intval et all,
> > if Namhyung sees no further problems with it.
 
> Okay, after a quick glance I don't see anymore issue.  You can add my
> Ack there.

Ok, thanks for checking!

- Arnaldo