Le Fri, Mar 24, 2023 at 03:51:54PM -0700, Paul E. McKenney a écrit :
> On Fri, Mar 24, 2023 at 11:09:08PM +0100, Frederic Weisbecker wrote:
> > Le Wed, Mar 22, 2023 at 04:18:24PM -0700, Paul E. McKenney a écrit :
> > > > @@ -1336,13 +1336,25 @@ lazy_rcu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
> > > > unsigned long flags;
> > > > unsigned long count = 0;
> > > >
> > > > + /*
> > > > + * Protect against concurrent (de-)offloading. Otherwise nocb locking
> > > > + * may be ignored or imbalanced.
> > > > + */
> > > > + mutex_lock(&rcu_state.barrier_mutex);
> > >
> > > I was worried about this possibly leading to out-of-memory deadlock,
> > > but if I recall correctly, the (de-)offloading process never allocates
> > > memory, so this should be OK?
> >
> > Good point. It _should_ be fine but like you, Joel and Hillf pointed out
> > it's asking for trouble.
> >
> > We could try Joel's idea to use mutex_trylock() as a best effort, which
> > should be fine as it's mostly uncontended.
> >
> > The alternative is to force nocb locking and check the offloading state
> > right after. So instead of:
> >
> > rcu_nocb_lock_irqsave(rdp, flags);
> > //flush stuff
> > rcu_nocb_unlock_irqrestore(rdp, flags);
> >
> > Have:
> >
> > raw_spin_lock_irqsave(rdp->nocb_lock, flags);
> > if (!rcu_rdp_is_offloaded(rdp))
> > raw_spin_unlock_irqrestore(rdp->nocb_lock, flags);
> > continue;
> > }
> > //flush stuff
> > rcu_nocb_unlock_irqrestore(rdp, flags);
> >
> > But it's not pretty and also disqualifies the last two patches as
> > rcu_nocb_mask can't be iterated safely anymore.
> >
> > What do you think?
>
> The mutex_trylock() approach does have the advantage of simplicity,
> and as you say should do well given low contention.
>
> Which reminds me, what sort of test strategy did you have in mind?
> Memory exhaustion can have surprising effects.
The best I can do is to trigger the count and scan callbacks through
the shrinker debugfs and see if it crashes or not :-)
>
> > > > /* Snapshot count of all CPUs */
> > > > for_each_possible_cpu(cpu) {
> > > > struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
> > > > - int _count = READ_ONCE(rdp->lazy_len);
> > > > + int _count;
> > > > +
> > > > + if (!rcu_rdp_is_offloaded(rdp))
> > > > + continue;
> > >
> > > If the CPU is offloaded, isn't ->lazy_len guaranteed to be zero?
> > >
> > > Or can it contain garbage after a de-offloading operation?
> >
> > If it's deoffloaded, ->lazy_len is indeed (supposed to be) guaranteed to be zero.
> > Bypass is flushed and disabled atomically early on de-offloading and the
> > flush resets ->lazy_len.
>
> Whew! At the moment, I don't feel strongly about whether or not
> the following code should (1) read the value, (2) warn on non-zero,
> (3) assume zero without reading, or (4) some other option that is not
> occurring to me. Your choice!
(2) looks like a good idea!
Thanks.
