Re: [PATCH tip/core/rcu 1/6] rcu: Abstract the dynticks momentary-idle operation

[Paul E. McKenney]

On Sun, Jan 15, 2017 at 11:39:51PM -0800, Josh Triplett wrote:
> On Sat, Jan 14, 2017 at 12:54:40AM -0800, Paul E. McKenney wrote:
> > This commit is the first step towards full abstraction of all accesses to
> > the ->dynticks counter, implementing the previously open-coded atomic add
> > of two in a new rcu_dynticks_momentary_idle() function.  This abstraction
> > will ease changes to the ->dynticks counter operation.
> > 
> > Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxxxxxxxxxx>
> 
> This change has an additional effect not documented in the commit
> message: it eliminates the smp_mb__before_atomic and
> smp_mb__after_atomic calls.  Can you please document that in the commit
> message, and explain why that doesn't cause a problem?

The trick is that the old code used the non-value-returning atomic_add(),
which does not imply ordering, hence the smp_mb__before_atomic() and
smp_mb__after_atomic() calls.  The new code uses atomic_add_return(),
which does return a value, and therefore implies full ordering in and
of itself.

How would you like me to proceed?

							Thanx, Paul

> > --- a/kernel/rcu/tree.c
> > +++ b/kernel/rcu/tree.c
> > @@ -281,6 +281,19 @@ static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
> >  #endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
> >  };
> >  
> > +/*
> > + * Do a double-increment of the ->dynticks counter to emulate a
> > + * momentary idle-CPU quiescent state.
> > + */
> > +static void rcu_dynticks_momentary_idle(void)
> > +{
> > +	struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
> > +	int special = atomic_add_return(2, &rdtp->dynticks);
> > +
> > +	/* It is illegal to call this from idle state. */
> > +	WARN_ON_ONCE(!(special & 0x1));
> > +}
> > +
> >  DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr);
> >  EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr);
> >  
> > @@ -300,7 +313,6 @@ EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr);
> >  static void rcu_momentary_dyntick_idle(void)
> >  {
> >  	struct rcu_data *rdp;
> > -	struct rcu_dynticks *rdtp;
> >  	int resched_mask;
> >  	struct rcu_state *rsp;
> >  
> > @@ -327,10 +339,7 @@ static void rcu_momentary_dyntick_idle(void)
> >  		 * quiescent state, with no need for this CPU to do anything
> >  		 * further.
> >  		 */
> > -		rdtp = this_cpu_ptr(&rcu_dynticks);
> > -		smp_mb__before_atomic(); /* Earlier stuff before QS. */
> > -		atomic_add(2, &rdtp->dynticks);  /* QS. */
> > -		smp_mb__after_atomic(); /* Later stuff after QS. */
> > +		rcu_dynticks_momentary_idle();
> >  		break;
> >  	}
> >  }
> > -- 
> > 2.5.2
> > 
>