Re: [PATCH 2/4] rcu: Remove useless WRITE_ONCE() on rcu_data.exp_deferred_qs

[Frederic Weisbecker]

On Fri, Sep 17, 2021 at 11:10:24AM -0700, Paul E. McKenney wrote:
> On Thu, Sep 16, 2021 at 11:05:14PM +0200, Frederic Weisbecker wrote:
> > On Thu, Sep 16, 2021 at 09:43:40AM -0700, Paul E. McKenney wrote:
> > > On Thu, Sep 16, 2021 at 02:10:46PM +0200, Frederic Weisbecker wrote:
> > > > This variable is never written nor read remotely. Remove this confusion.
> > > > 
> > > > Signed-off-by: Frederic Weisbecker <frederic@xxxxxxxxxx>
> > > > ---
> > > >  kernel/rcu/tree_exp.h | 2 +-
> > > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > > > 
> > > > diff --git a/kernel/rcu/tree_exp.h b/kernel/rcu/tree_exp.h
> > > > index f3947c49eee7..4266610b4587 100644
> > > > --- a/kernel/rcu/tree_exp.h
> > > > +++ b/kernel/rcu/tree_exp.h
> > > > @@ -255,7 +255,7 @@ static void rcu_report_exp_cpu_mult(struct rcu_node *rnp,
> > > >   */
> > > >  static void rcu_report_exp_rdp(struct rcu_data *rdp)
> > > >  {
> > > > -	WRITE_ONCE(rdp->exp_deferred_qs, false);
> > > > +	rdp->exp_deferred_qs = false;
> > > 
> > > Are you sure that this can never be invoked from an interrupt handler?
> > > And that rdp->exp_deferred_qs is never read from an interrupt handler?
> > > If either can happen, then the WRITE_ONCE() does play a role, right?
> > 
> > Well, the only effect I can imagine is that it can partly prevent from an
> > interrupt to report concurrently the quiescent state during the few
> > instructions before we mask interrupts and lock the node.
> > 
> > That's a micro performance benefit that avoid a second call to
> > rcu_report_exp_cpu_mult() with the extra locking and early exit.
> 
> I am not claiming that current compilers would mess this up, though I
> have learned to have great respect for what future compilers might do...

:)

> 
> > But then that racy interrupt can still happen before we clear exp_deferred_qs.
> > In this case __this_cpu_cmpxchg() would have been more efficient.
> 
> Except that __this_cpu_cmpxchg() would have a possibility of failure,
> and thus an extra branch not needed by WRITE_ONCE().  Or am I missing
> your point here?

Right, but an extra branch that could spare a call to rcu_report_exp_cpu_mult().

Anyway I don't mind the WRITE_ONCE(), but you know how ordering (whether
compiler or CPU) makes me anxious when undocumented or not self-explanatory,
(although arguably the latter can vary depending on the reviewer :)

Thanks.