On Wed 11-05-16 11:17:33, Peter Zijlstra wrote:
> On Wed, May 11, 2016 at 11:04:42AM +0200, Michal Hocko wrote:
> > On Wed 11-05-16 10:44:01, Peter Zijlstra wrote:
> > [...]
> > > @@ -504,6 +502,18 @@ __rwsem_down_write_failed_common(struct rw_semaphore
> > > *sem, int state)
> > > raw_spin_unlock_irq(&sem->wait_lock);
> > >
> > > return ret;
> > > +
> > > +out_nolock:
> > > + __set_current_state(TASK_RUNNING);
> > > + raw_spin_lock_irq(&sem->wait_lock);
> > > + list_del(&waiter.list);
> > > + if (list_empty(&sem->wait_list))
> > > + rwsem_atomic_update(-RWSEM_WAITING_BIAS, sem);
> > > + else
> > > + __rwsem_do_wake(sem, RWSEM_WAKE_READERS);
> > > + raw_spin_unlock_irq(&sem->wait_lock);
> > > +
> > > + return ERR_PTR(-EINTR);
> > > }
> >
> > Looks much better but don't we have to check the count for potentially
> > pending writers?
>
> Ah, so I was thinking that if we get here, there must still be an owner,
> otherwise we'd have acquired the lock. And if there is an owner, we
> cannot go wake writers. Hence the WAKE_READERS thing.
I was worried about the case where the owner is writer and we would wake
readers but I have missed that this wouldn't happen because of
if (wake_type != RWSEM_WAKE_READ_OWNED) {
adjustment = RWSEM_ACTIVE_READ_BIAS;
try_reader_grant:
oldcount = rwsem_atomic_update(adjustment, sem) - adjustment;
if (unlikely(oldcount < RWSEM_WAITING_BIAS)) {
/* A writer stole the lock. Undo our reader grant. */
if (rwsem_atomic_update(-adjustment, sem) &
RWSEM_ACTIVE_MASK)
goto out;
/* Last active locker left. Retry waking readers. */
goto try_reader_grant;
}
}
> Then again, WAKE_ANY would not harm, in that if we do wake a pending
> writer it will not proceed if it cannot and it'll go back to sleep
> again.
true
> So yeah, maybe WAKE_ANY is the prudent thing to do.
I guess so.
Care to cook up a full patch?
Thanks!
--
Michal Hocko
SUSE Labs
