Re: On trace_*_rcuidle functions in modules

[Paul E. McKenney]

On Tue, Apr 14, 2020 at 08:47:18PM -0700, John Stultz wrote:
> On Tue, Apr 14, 2020 at 7:57 PM Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
> >
> > On Tue, Apr 14, 2020 at 07:20:01PM -0700, John Stultz wrote:
> > > Hey folks,
> > >   So recently I was looking at converting some drivers to be loadable
> > > modules instead of built-in only, and one of my patches just landed in
> > > -next and started getting build error reports.
> > >
> > > It ends up, recently in the merge window, the driver I was converting
> > > to module switched a trace_*() function to trace_*_rcuidle() to fix a
> > > bug.  Now when building as a module, if tracing is configured on, it
> > > can't seem to find the trace_*_rcuidle() symbol.
> > >
> > > This is because, as you are aware, we don't declare trace_*_rcuidle
> > > functions in modules - and haven't for quite some time:
> > >   https://lore.kernel.org/lkml/20120905062306.GA14756@leaf/
> > >
> > > I wanted to better understand the background rationale for that patch,
> > > to understand if not exporting the rcu_idle_exit and rcu_idle_enter,
> > > calls was because they weren't used or if it was a more intentional
> > > decision to avoid allowing modules to use them.
> > >
> > > Would it be reasonable to revisit that patch? Or is there some
> > > recommended alternative solution?
> >
> > I will defer to Steven and Josh on the rationale.  (Cowardly of me,
> > I know!)
> >
> > What I do is to maintain a wrapper for tracepoints within a built-in
> > portion of RCU, export the wrapper, and invoke the wrapper from the
> > rcutorture module.  Maybe you can do something similar?
> 
> That feels a little hackish, but I guess if there isn't a better option...

It is just rcutorture, so I am not going to claim that I put a huge
amount of energy into researching better options.  ;-)

> > But why would a module be invoked from the idle loop?  Is the module
> > supplying an idle driver or some such?
> 
> The driver (qcom rpmh driver) registers a cpu_pm notifier callback,
> which gets called when entering idle.

That would do it!  And yes, the idle-loop power-management code is in
an interesting situation in that RCU is not watching it by default, but
it does need to be debugged.  One straightforward approach would be for
the PM code to tell RCU to watch on entry and exit, but I don't have a
feel for how this would affect energy efficiency.

							Thanx, Paul