On Wed, Sep 19, 2018 at 03:55:07PM +0800, Ming Lei wrote: > On Wed, Sep 19, 2018 at 01:19:10PM +0800, jianchao.wang wrote: > > Hi Ming > > > > On 09/18/2018 06:19 PM, Ming Lei wrote: > > > + unsigned long __percpu *percpu_count; > > > + > > > + WARN_ON_ONCE(__ref_is_percpu(ref, &percpu_count)); > > > + > > > + /* get one extra ref for avoiding race with .release */ > > > + rcu_read_lock_sched(); > > > + atomic_long_add(1, &ref->count); > > > + rcu_read_unlock_sched(); > > > + } > > > > The rcu_read_lock_sched here is redundant. We have been in the critical section > > of a spin_lock_irqsave. > > Right. > > > > > The atomic_long_add(1, &ref->count) may have two result. > > 1. ref->count > 1 > > it will not drop to zero any more. > > 2. ref->count == 1 > > it has dropped to zero and .release may be running. > > IMO, both the two cases are fine and supported, or do you have other > concern about this way? It is too quick, :-) Yeah, the .release() may be running. For blk-mq/NVMe's use case, it won't be an issue. We may comment on this race and let user handle it if it is a problem. thanks, Ming