On Wed, Mar 29, 2023 at 06:02:02PM +0200, Frederic Weisbecker wrote:
> The ->lazy_len is only checked locklessly. Recheck again under the
> ->nocb_lock to avoid spending more time on flushing/waking if not
> necessary. The ->lazy_len can still increment concurrently (from 1 to
> infinity) but under the ->nocb_lock we at least know for sure if there
> are lazy callbacks at all (->lazy_len > 0).
>
> Signed-off-by: Frederic Weisbecker <frederic@xxxxxxxxxx>
> ---
> kernel/rcu/tree_nocb.h | 16 ++++++++++++----
> 1 file changed, 12 insertions(+), 4 deletions(-)
>
> diff --git a/kernel/rcu/tree_nocb.h b/kernel/rcu/tree_nocb.h
> index c321fce2af8e..dfa9c10d6727 100644
> --- a/kernel/rcu/tree_nocb.h
> +++ b/kernel/rcu/tree_nocb.h
> @@ -1358,12 +1358,20 @@ lazy_rcu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
> if (!rcu_rdp_is_offloaded(rdp))
> continue;
>
> + if (!READ_ONCE(rdp->lazy_len))
> + continue;
Do you depend on the ordering of the above read of ->lazy_len against
anything in the following, aside from the re-read of ->lazy_len? (Same
variable, both READ_ONCE() or stronger, so you do get that ordering.)
If you do need that ordering, the above READ_ONCE() needs to instead
be smp_load_acquire() or similar. If you don't need that ordering,
what you have is good.
> + rcu_nocb_lock_irqsave(rdp, flags);
> + /*
> + * Recheck under the nocb lock. Since we are not holding the bypass
> + * lock we may still race with increments from the enqueuer but still
> + * we know for sure if there is at least one lazy callback.
> + */
> _count = READ_ONCE(rdp->lazy_len);
> -
> - if (_count == 0)
> + if (!_count) {
> + rcu_nocb_unlock_irqrestore(rdp, flags);
> continue;
> -
> - rcu_nocb_lock_irqsave(rdp, flags);
> + }
> WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies, false));
> rcu_nocb_unlock_irqrestore(rdp, flags);
> wake_nocb_gp(rdp, false);
> --
> 2.34.1
>
