On 2019-10-02 08:08:52 [-0700], Paul E. McKenney wrote:
> On Wed, Oct 02, 2019 at 01:22:53PM +0200, Sebastian Andrzej Siewior wrote:
> > This is a revert of commit
> > a4244454df129 ("percpu-refcount: use RCU-sched insted of normal RCU")
> >
> > which claims the only reason for using RCU-sched is
> > "rcu_read_[un]lock() … are slightly more expensive than
> > preempt_disable/enable()"
> >
> > and
> > "As the RCU critical sections are extremely short, using sched-RCU
> > shouldn't have any latency implications."
> >
> > The problem with using RCU-sched here is that it disables preemption and
> > the callback must not acquire any sleeping locks like spinlock_t on
> > PREEMPT_RT which is the case with some of the users.
>
> Looks good in general, but changing to RCU-preempt does not change the
> fact that the callbacks execute with bh disabled. There is a newish
> queue_rcu_work() that invokes a workqueue handler after a grace period.
>
> Or am I missing your point here?
That is fine, no the RCU callback. The problem is that
percpu_ref_put_many() as of now does:
rcu_read_lock_sched(): /* aka preempt_disable(); */
if (__ref_is_percpu(ref, &percpu_count))
this_cpu_sub(*percpu_count, nr);
else if (unlikely(atomic_long_sub_and_test(nr, &ref->count)))
ref->release(ref);
and then the callback invoked via ref->release() acquires a spinlock_t
with disabled preemption.
> Thanx, Paul
Sebastian
