Re: [PATCH v2 18/27] rcu: Rename rcu_dynticks_in_eqs_since() into rcu_watching_changed_since()

[Frederic Weisbecker <frederic@kernel.org>]

Le Tue, May 07, 2024 at 10:14:08AM -0700, Paul E. McKenney a écrit :
> On Tue, May 07, 2024 at 03:48:18PM +0200, Frederic Weisbecker wrote:
> > Le Tue, Apr 30, 2024 at 11:17:22AM +0200, Valentin Schneider a écrit :
> > > The context_tracking.state RCU_DYNTICKS subvariable has been renamed to
> > > RCU_WATCHING, the dynticks prefix can go.
> > > 
> > > Furthermore, the "in_eqs_since" part confuses me, as IIUC this only checks
> > > for a change in watching/eqs state, not that RCU transitionned *into* a
> > > EQS after the snapshot was taken.
> > > 
> > > e.g. if
> > >   snap = 0b1000 (EQS)
> > > and the following rcu_watching_snap(CPU) is:
> > > 	 0b1100 (watching)
> > > then
> > >   rcu_watching_in_eqs_since(rdp, snap) -> true
> > > 
> > > but because RCU was already in EQS at the time of the
> > > snap - it hasn't entered EQS "since" the snap was taken.
> > > 
> > > Update the name to reflect that we're only looking at watching/EQS
> > > transitions, not specifically transitions into EQS.
> > 
> > Indeed in practice the function only checks a change. But semantically it really
> > checks a trip to eqs because this function is only ever called after a failing
> > call to rcu_dynticks_in_eqs().
> > 
> > So not sure about that one rename. Paul?
> 
> As you say, Valentin is technically correct.  Me, I am having a hard
> time getting too excited one way or the other.  ;-)
> 
> I suggest thinking in terms of rate-bounding the change.  If you do
> change it, don't change it again for a few years.

Makes sense!

> 
> Either way, should comments be changed or added?
> 
> Of course, the scientific way to evaluate this is to whose a couple
> dozen people the old code and a couple dozen other people the new code,
> and see if one group or the other has statistically significantly lower
> levels of confusion.  I don't see how this is feasible, but it is the
> (painfully) correct way.  On the other hand, it would have the beneficial
> side effect of getting more people exposed to Linux-kernel-RCU internals.
> Unfortunately, it might also have the additional side effect of making
> them (more) annoyed at RCU.  ;-)

Sounds good!

I divided myself in two blank RCU subjects for a double blind study
and locked those people up overnight with a paper containing both proposals.

I opened the door five minutes ago and they both elected by mutual agreement
rcu_watching_changed_since()! Also they are thirsty.

Congratulations Valentin! :-)

> 							Thanx, Paul