Re: [PATCH] perf mem: add priv level filtering support

[Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com>]

Arnaldo Carvalho de Melo [acme@redhat.com] wrote:
| Em Wed, Aug 28, 2013 at 03:38:28PM +0200, Stephane Eranian escreveu:
| > On Wed, Aug 28, 2013 at 3:27 PM, Arnaldo Carvalho de Melo <acme@redhat.com> wrote:
| > > So perhaps we should change both to (and add this to 'report' as well):
| > >
| > >     -U, --hide_kernel_symbols   hide kernel symbols
| > >     -K, --hide_user_symbols     hide user symbols
| > >
| > 
| > Well, I don't know what perf top does here but I don't want to hide
| > the samples. I simply don't want to collect them (do not appear
| > in the perf.data file). If that's what is happening in perf top, then
| > I'll be glad to use the same options.
| 
| Indeed, its for different purposes, 'perf top' when used with one of
| those options will still collect samples for all priv levels and will
| just toggle a flag to not zap the ones asked not to show when decaying
| the samples.
| 
| When the user presses 'U' or 'K' on the UI, the flags gets toggled and
| samples start being considered/zapped.
| 
| But my worry here is about consistency accross tools for the single
| letter options, so perhaps if you could use:
| 
|      -U		collect only user level samples
|      -K		collect only kernel level samples
| 
| I think it would stay consistent and clear, what do you think?

But, we use lower case qualifiers :u, :k to select user or kernel mode
monitoring. 

	perf record -e cycles		# both kernel and user
	perf record -e cycles:u ...	# just user

(tools/perf/util/parse-events.c:

        struct event_modifier {
                int eu;
                int ek;
                int eh;
                int eH;
                int eG;
                int precise;
                int exclude_GH;
        };

Will we ever need hypervisor and host monitoring for 'perf mem' ?

Or can we add a '-e' option to 'perf mem' so user can specify the events
and qualfiers same as they do for 'perf record' ?

	perf mem -e mem-loads:u record .....

(this would of course expose the mem-loads and mem-stores events to
the user)

Sukadev