Re: On trace_*_rcuidle functions in modules

[Paul E. McKenney]

On Wed, Apr 15, 2020 at 06:51:21PM -0400, Steven Rostedt wrote:
> On Wed, 15 Apr 2020 15:04:59 -0700
> "Paul E. McKenney" <paulmck@xxxxxxxxxx> wrote:
> 
> > On Wed, Apr 15, 2020 at 05:49:18PM -0400, Steven Rostedt wrote:
> > > On Wed, 15 Apr 2020 14:02:04 -0700
> > > John Stultz <john.stultz@xxxxxxxxxx> wrote:
> > >   
> > > > 
> > > > So in my case your concerns may not be a problem, but I guess
> > > > generally it might. Though I'd hope the callback would be unregistered
> > > > (and whatever waiting for the grace period to complete be done) before
> > > > the module removal is complete. But maybe I'm still missing your
> > > > point?  
> > > 
> > > Hmm, you may have just brought up a problem here...
> > > 
> > > You're saying that cpu_pm_register_notifier() callers are called from non
> > > RCU watching context? If that's the case, we have this:
> > > 
> > > int cpu_pm_unregister_notifier(struct notifier_block *nb)
> > > {
> > > 	return atomic_notifier_chain_unregister(&cpu_pm_notifier_chain, nb);
> > > }
> > > 
> > > And this:
> > > 
> > > int atomic_notifier_chain_unregister(struct atomic_notifier_head *nh,
> > > 		struct notifier_block *n)
> > > {
> > > 	unsigned long flags;
> > > 	int ret;
> > > 
> > > 	spin_lock_irqsave(&nh->lock, flags);
> > > 	ret = notifier_chain_unregister(&nh->head, n);
> > > 	spin_unlock_irqrestore(&nh->lock, flags);
> > > 	synchronize_rcu();
> > > 	return ret;
> > > }
> > > 
> > > Which means that if something registered a cpu_pm notifier, then
> > > unregistered it, and freed whatever the notifier accesses, then there's a
> > > chance that the synchronize_rcu() can return before the called notifier
> > > finishes, and anything that notifier accesses could have been freed.
> > > 
> > > I believe that module code should not be able to be run in RCU non watching
> > > context, and neither should notifiers. I think we just stumbled on a bug.
> > > 
> > > Paul?  
> > 
> > Or we say that such modules cannot be unloaded.  Or that such modules'
> > exit handlers, after disentangling themselves from the idle loop, must
> > invoke synchronize_rcu_rude() or similar, just as modules that use
> > call_rcu() are currently required to invoke rcu_barrier().
> > 
> > Or is it possible to upgrade the protection that modules use?
> > 
> > My guess is that invoking rcu_irq_enter() and rcu_irq_exit() around every
> > potential call into module code out of the PM code is a non-starter,
> > but I cannot prove that either way.
> >
> 
> No this has nothing to do with modules. This is a bug right now with the
> cpu_pm notifier (after looking at the code, it's not a bug right now, see
> below).
> 
> Say you have something that allocates some data and registers a
> callback to the cpu_pm notifier that access that data. Then for some
> reason, you want to remove that notifier and free the data. Usually you
> would do:
> 
> 	cpu_pm_unregister_notifier(my_notifier);
> 	kfree(my_data);
> 
> But the problem is that the callback of that my_notifier could be executing
> in a RCU non-watching space, and the cpu_pm_unregister_notifier() can
> return before the my_notifier is done, and the my_data is freed. Then the
> callback for the my_notifier could still be accessing the my_data.
> 
> 
> /me goes and reads the code and sees this is not an issue, and you can
> ignore the above concern.
> 
> I was about to suggest a patch, but that has already been written...
> 
> 313c8c16ee62b ("PM / CPU: replace raw_notifier with atomic_notifier")
> 
> Which surrounds the notifier callbacks with rcu_irq_enter_irqson()
> 
> Which means that if John moves the code to use the notifier, then he could
> also remove the _rcuidle(), because RCU will be watching.

Whew!!!  ;-)

							Thanx, Paul