Re: [PATCH v3 1/1] rcu: Reduce synchronize_rcu() waiting time

From: Uladzislau Rezki
Date: Wed Oct 18 2023 - 13:37:52 EST


On Wed, Oct 18, 2023 at 10:32:22AM -0400, Joel Fernandes wrote:
> On Tue, Oct 17, 2023 at 10:06 AM Uladzislau Rezki <urezki@xxxxxxxxx> wrote:
> [...]
> > > > +
> > > > + /* Finally. */
> > > > + complete(&rs->completion);
> > > > +}
> > > > +
> > > > +static void rcu_sr_normal_gp_cleanup_work(struct work_struct *work)
> > > > +{
> > > > + struct llist_node *done, *rcu, *next;
> > > > +
> > > > + done = llist_del_all(&sr.done);
> > > > + if (!done)
> > > > + return;
> > > > +
> > > > + llist_for_each_safe(rcu, next, done)
> > > > + rcu_sr_normal_complete(rcu);
> > > > +}
> > > [...]
> > > > +static void rcu_sr_normal_add_req(struct rcu_synchronize *rs)
> > > > +{
> > > > + atomic_inc(&sr.active);
> > > > + if (llist_add((struct llist_node *) &rs->head, &sr.curr))
> > > > + /* Set the tail. Only first and one user can do that. */
> > > > + WRITE_ONCE(sr.curr_tail, (struct llist_node *) &rs->head);
> > > > + atomic_dec(&sr.active);
> > >
> > > Here there is no memory ordering provided by the atomic ops. Is that really Ok?
> > >
> > This needs to be reworked since there is no ordering guaranteed. I think
> > there is a version of "atomic_inc_something" that guarantees it?
>
> Yeah there is atomic_fetch_{inc,dec}{_acquire,_release}()
>
> Or:
> atomic_inc(&sr.active);
> smp_mb__after_atomic();
>
> smp_mb__before_atomic();
> atomic_dec(&sr.active);
>
> ?
>
> That's probably better because we don't need ordering before the inc
> or after the dec, AFAICS.
>
There are two variants, atomic_inc_return() and atomic_dec_return()
which are fully ordered. Any thoughts about them? One line instead of
two as in your case.

Your concern is about, that atomic_inc() can be reodered? There is a
llist_add() that has inside the try_cmpxchg() that should have barrier.

Any thoughts?

Thank you for the review and help, Joel!

--
Uladzislau Rezki