Re: [PATCH 1/2] perf tools: Add reference timestamp to perf header

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

Em Mon, Dec 13, 2010 at 06:09:26PM +0100, Frederic Weisbecker escreveu:
> On Mon, Dec 13, 2010 at 02:48:54PM -0200, Arnaldo Carvalho de Melo wrote:
> > Em Mon, Dec 13, 2010 at 04:54:56PM +0100, Frederic Weisbecker escreveu:
> > > On Mon, Dec 13, 2010 at 07:39:24AM -0700, David S. Ahern wrote:
> > > > On 12/12/10 13:16, Frederic Weisbecker wrote:
> > > > >>  int perf_session__create_kernel_maps(struct perf_session *self)
> > > > >>  {
> > > > >>  	int ret = machine__create_kernel_maps(&self->host_machine);
> > > > >> @@ -167,6 +190,9 @@ struct perf_session *perf_session__new(const char *filename, int mode, bool forc
> > > > >>  		 */
> > > > >>  		if (perf_session__create_kernel_maps(self) < 0)
> > > > >>  			goto out_delete;
> > > > >> +
> > > > >> +		if (perf_session__create_ref_time(self) < 0)
> > > > >> +			goto out_delete;
> > 
> > > > > So, it does record it anytime?
> > 
> > > > Simplest to always add it the header as it does no harm if not used.
> > > > To make it conditional means adding an input parameter to
> > > > perf_session__new() and then updating all the callers.
> > 
> > > > Is the preference to make it conditional?
> > 
> > > No, I suspect the overhead is plain unnoticeable. And if it becomes
> > > any problem one day, we can still make that conditional later.
> > 
> > Right, but it just ocurred to me, can't we encode this in the file stat?
> > Also, how do we deal with:
> > 
> > $ perf record --help
> > <SNIP>
> >        -A, --append
> >            Append to the output file to do incremental profiling.
> > <SNIP>
> > 
> > ?
> > 
> > I now think this should be implemented as a PERF_RECORD_WALL_CLOCK
> > synthesized event, to be inserted just before we start collecting the
> > other events.
> > 
> > That way, when --appending, we just insert another and use that from
> > that point on.
> > 
> > Look at tools/perf/util/event.h, enum perf_user_event_type and look how
> > those are synthesized.
> 
> (Adding peterz)

Adding tglx
 
> Right, but Peter wanted us to get rid of these user events types.
> I guess we can't really do this as new perf tools must be able
> to support old perf files.
> 
> So this should be the last one to add. Something like PERF_RECORD_GEN_EVT
> that has a field in its headers containing a sub-type which can be this
> wall clock but can also host about everything in the future.
> 
> This way we don't propagate more the possible overlap with the kernel.

That works for me. I would call it:

PERF_RECORD_SYNTH

our ioctl ;-)

But whatever, just a name.

- Arnaldo