On Mon, Aug 05, 2019 at 10:58:13PM +0800, Boqun Feng wrote:
> On Mon, Aug 05, 2019 at 10:43:18PM +0800, Boqun Feng wrote:
> > On Mon, Aug 05, 2019 at 04:02:41PM +0200, Peter Zijlstra wrote:
> > [...]
> > >
> > > static inline void percpu_up_read(struct percpu_rw_semaphore *sem)
> > > {
> > > + rwsem_release(&sem->dep_map, 1, _RET_IP_);
> > > +
> > > preempt_disable();
> > > /*
> > > * Same as in percpu_down_read().
> > > */
> > > - if (likely(rcu_sync_is_idle(&sem->rss)))
> > > + if (likely(rcu_sync_is_idle(&sem->rss))) {
> > > __this_cpu_dec(*sem->read_count);
> > > - else
> > > - __percpu_up_read(sem); /* Unconditional memory barrier */
> > > - preempt_enable();
> > > + preempt_enable();
> > > + return;
> > > + }
> > >
> > > - rwsem_release(&sem->rw_sem.dep_map, 1, _RET_IP_);
> >
> > Missing a preempt_enable() here?
> >
>
> Ah.. you modified the semantics of __percpu_up_read() to imply a
> preempt_enable(), sorry for the noise...
Yes indeed; I suppose I should've noted that in the Changlog. The reason
is that waitqueues use spin_lock() which change into a sleepable lock on
RT and thus cannot be used with preeption disabled. We also cannot
(easily) switch to swait because we use both exclusive and !exclusive
waits.
