Re: [PATCH v2 1/6] perf: Add interface to add general events to sysfs

[Lin Ming <ming.m.lin@intel.com>]

On Mon, 2011-07-25 at 15:57 +0800, Peter Zijlstra wrote:
> On Tue, 2011-07-19 at 15:52 +0800, Lin Ming wrote:
> 
> > > I'd rather not have a function pointer here, and all the kobj thingies
> > > can be hooking into the existing struct device, right?
> > 
> > I forgot to mention one important reason why I added a function pointer.
> > 
> > The events can only be added to sysfs after PMU sysfs is initialized in
> > perf_event_sysfs_init ->
> >     pmu_dev_alloc
> 
> Right.
> 
> > > @@ -5571,6 +5571,8 @@ static int pmu_dev_alloc(struct pmu *pmu)
> > >        if (ret)
> > >                goto free_dev;
> > >
> > > +       if (pmu->add_events)
> > > +               pmu->add_events();
> > 
> > So we need a pmu callback which is called in pmu_dev_alloc to add events
> > to sysfs.
> > 
> > You suggested a new interface,
> > int perf_pmu_add_event(struct pmu *pmu, const char *name, u64 config)
> > 
> > But where should it be called?
> > I guess you mean to call it in pmu init function, for example,
> > uncore_pmu_init.
> > 
> > But pmu init function maybe called before perf_event_sysfs_init, which
> > means the pmu sysfs has not been initialized yet.
> 
> Right, so there is no reason to call perf_pmu_register() really early,
> except for the normal pmu and the software pmu (since they're used by
> the watchdog muck).
> 
> The uncore for example can use late_initcall() just fine and use both
> perf_pmu_register() and the proposed perf_pmu_add_event() from the same 
> init call.
> 
> Only for the primary pmu and software thingies do we need to add an
> extra init call.

This seems a solution.

I'll use late_initcall for uncore.
Then later we can add an extra late_initcall for normal pmu and software
pmu.