Re: [PATCH v1 1/6] rcu: Add support for consolidated-RCU reader checking

[Paul E. McKenney]

On Fri, Jul 12, 2019 at 01:06:31PM -0400, Joel Fernandes wrote:
> On Fri, Jul 12, 2019 at 09:45:31AM -0700, Paul E. McKenney wrote:
> > On Fri, Jul 12, 2019 at 11:10:51AM -0400, Joel Fernandes wrote:
> > > On Fri, Jul 12, 2019 at 01:11:25PM +0200, Peter Zijlstra wrote:
> > > > On Thu, Jul 11, 2019 at 07:43:56PM -0400, Joel Fernandes (Google) wrote:
> > > > > +int rcu_read_lock_any_held(void)
> > > > > +{
> > > > > +	int lockdep_opinion = 0;
> > > > > +
> > > > > +	if (!debug_lockdep_rcu_enabled())
> > > > > +		return 1;
> > > > > +	if (!rcu_is_watching())
> > > > > +		return 0;
> > > > > +	if (!rcu_lockdep_current_cpu_online())
> > > > > +		return 0;
> > > > > +
> > > > > +	/* Preemptible RCU flavor */
> > > > > +	if (lock_is_held(&rcu_lock_map))
> > > > 
> > > > you forgot debug_locks here.
> > > 
> > > Actually, it turns out debug_locks checking is not even needed. If
> > > debug_locks == 0, then debug_lockdep_rcu_enabled() returns 0 and we would not
> > > get to this point.
> > > 
> > > > > +		return 1;
> > > > > +
> > > > > +	/* BH flavor */
> > > > > +	if (in_softirq() || irqs_disabled())
> > > > 
> > > > I'm not sure I'd put irqs_disabled() under BH, also this entire
> > > > condition is superfluous, see below.
> > > > 
> > > > > +		return 1;
> > > > > +
> > > > > +	/* Sched flavor */
> > > > > +	if (debug_locks)
> > > > > +		lockdep_opinion = lock_is_held(&rcu_sched_lock_map);
> > > > > +	return lockdep_opinion || !preemptible();
> > > > 
> > > > that !preemptible() turns into:
> > > > 
> > > >   !(preempt_count()==0 && !irqs_disabled())
> > > > 
> > > > which is:
> > > > 
> > > >   preempt_count() != 0 || irqs_disabled()
> > > > 
> > > > and already includes irqs_disabled() and in_softirq().
> > > > 
> > > > > +}
> > > > 
> > > > So maybe something lke:
> > > > 
> > > > 	if (debug_locks && (lock_is_held(&rcu_lock_map) ||
> > > > 			    lock_is_held(&rcu_sched_lock_map)))
> > > > 		return true;
> > > 
> > > Agreed, I will do it this way (without the debug_locks) like:
> > > 
> > > ---8<-----------------------
> > > 
> > > diff --git a/kernel/rcu/update.c b/kernel/rcu/update.c
> > > index ba861d1716d3..339aebc330db 100644
> > > --- a/kernel/rcu/update.c
> > > +++ b/kernel/rcu/update.c
> > > @@ -296,27 +296,15 @@ EXPORT_SYMBOL_GPL(rcu_read_lock_bh_held);
> > >  
> > >  int rcu_read_lock_any_held(void)
> > >  {
> > > -	int lockdep_opinion = 0;
> > > -
> > >  	if (!debug_lockdep_rcu_enabled())
> > >  		return 1;
> > >  	if (!rcu_is_watching())
> > >  		return 0;
> > >  	if (!rcu_lockdep_current_cpu_online())
> > >  		return 0;
> > > -
> > > -	/* Preemptible RCU flavor */
> > > -	if (lock_is_held(&rcu_lock_map))
> > > -		return 1;
> > > -
> > > -	/* BH flavor */
> > > -	if (in_softirq() || irqs_disabled())
> > > -		return 1;
> > > -
> > > -	/* Sched flavor */
> > > -	if (debug_locks)
> > > -		lockdep_opinion = lock_is_held(&rcu_sched_lock_map);
> > > -	return lockdep_opinion || !preemptible();
> > > +	if (lock_is_held(&rcu_lock_map) || lock_is_held(&rcu_sched_lock_map))
> > 
> > OK, I will bite...  Why not also lock_is_held(&rcu_bh_lock_map)?
> 
> Hmm, I was borrowing the strategy from rcu_read_lock_bh_held() which does not
> check for a lock held in this map.
> 
> Honestly, even  lock_is_held(&rcu_sched_lock_map) seems unnecessary per-se
> since !preemptible() will catch that? rcu_read_lock_sched() disables
> preemption already, so lockdep's opinion of the matter seems redundant there.

Good point!  At least as long as the lockdep splats list RCU-bh among
the locks held, which they did last I checked.

Of course, you could make the same argument for getting rid of
rcu_sched_lock_map.  Does it make sense to have the one without
the other?

> Sorry I already sent out patches again before seeing your comment but I can
> rework and resend them based on any other suggestions.

Not a problem!

							Thax, Paul