Re: [PATCH 2/3 -tip] perf_counter tools: Add support to set ofmultiple events in one shot

From: Ingo Molnar
Date: Tue Jun 30 2009 - 18:47:39 EST



* Jaswinder Singh Rajput <jaswinder@xxxxxxxxxx> wrote:

> On Tue, 2009-06-30 at 11:57 +0200, Ingo Molnar wrote:
> > * Jaswinder Singh Rajput <jaswinder@xxxxxxxxxx> wrote:
> >
> > > On Mon, 2009-06-29 at 05:57 +0200, Ingo Molnar wrote:
> > >
> > > > The above patterns i suggested _already cover_ 'multiple events'.
> > > >
> > > > We might define further aliases like:
> > > >
> > > > all := "*"
> > > > all-sw := "sw-*"
> > > >
> > > > but it should all be in terms of patterns and regular
> > > > expressions, not via some hardcoded special-case thing as your
> > > > posted patches did.
> > > >
> > >
> > > It seems to me very confusing and needs lot of book-keeping and
> > > need to rewrite whole tools/perf/util/parse-events.c because :
> > >
> > > * means all perf_event_types :
> > > PERF_TYPE_HARDWARE,
> > > PERF_TYPE_SOFTWARE,
> > > PERF_TYPE_TRACEPOINT,
> > > PERF_TYPE_HW_CACHE,
> > > PERF_TYPE_RAW
> > >
> > > hw-* means all hardware events :
> > > PERF_TYPE_HARDWARE,
> > > PERF_TYPE_HW_CACHE,
> > > PERF_TYPE_RAW
> > >
> > > sw-* means all software events :
> > > PERF_TYPE_SOFTWARE,
> > > PERF_TYPE_TRACEPOINT
> > >
> > > *cache* means all cache based events :
> > > PERF_COUNT_CACHE_REFERENCES, /* Generalized H/W */
> > > PERF_COUNT_CACHE_MISSES, /* Generalized H/W */
> > > PERF_TYPE_HW_CACHE, /* Generalized Cache */
> > >
> > > *write* means all write based events :
> > > (L1D, WRITE, ACCESS),
> > > (L1D, WRITE, MISS),
> > > (LL, WRITE, ACCESS),
> > > (LL, WRITE, MISS),
> > > (DTLB, WRITE, ACCESS),
> > > (DTLB, WRITE, MISS)
> > >
> > > Please let me know why it looks complex to me, is it really
> > > complex or I am going in wrong direction.
> >
> > It would certainly need some reorganization of the code but the end
> > result would be more flexible and other places could use it too, for
> > example:
> >
> > perf test -e hw-*
> >
> > would test all (known) hardware counters.
> >
>
> Its true.
>
> Can you please verify that the assumptions I made above are
> correct.

Well, the right way to approach this is to assign each event a "full
name" and a list of aliases/shortcuts (like we have now), and then
do pattern matching on the full name.

So we'd have full/long event names like:

hw-cycles
hw-instructions
hw-l1-cache-load-misses
sw-minor-page-faults
...

to implement regex patterns over these, no event specific knowledge
should be put into the pattern matching engine itself - it just
blindly goes over the full names as strings.

As long as the full names are unique and structured well, this will
work fine. The only non-trivial piece of restructuring is to make it
easy for the pattern matching engine to iterate over all events.
Right now they are in separate tables - perhaps they should be
collected into a single table or so.

Ingo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/