Re: [PATCH] rwsem: fix missed wakeup due to reordering of load

[Peter Zijlstra]

On Thu, Aug 24, 2017 at 02:33:04PM +0200, Peter Zijlstra wrote:
> On Thu, Aug 24, 2017 at 01:29:27PM +0200, Peter Zijlstra wrote:
> > 
> > WTH did you not Cc the people that commented on your patch last time?
> > 
> > On Wed, Aug 23, 2017 at 04:58:55PM +0530, Prateek Sood wrote:
> > > If a spinner is present, there is a chance that the load of
> > > rwsem_has_spinner() in rwsem_wake() can be reordered with
> > > respect to decrement of rwsem count in __up_write() leading
> > > to wakeup being missed.
> > 
> > >  spinning writer                  up_write caller
> > >  ---------------                  -----------------------
> > >  [S] osq_unlock()                 [L] osq
> > >   spin_lock(wait_lock)
> > >   sem->count=0xFFFFFFFF00000001
> > >             +0xFFFFFFFF00000000
> > >   count=sem->count
> > >   MB
> > >                                    sem->count=0xFFFFFFFE00000001
> > >                                              -0xFFFFFFFF00000001
> > >                                    RMB
> > 
> > This doesn't make sense, it appears to order a STORE against something
> > else.
> > 
> > >                                    spin_trylock(wait_lock)
> > >                                    return
> > >  rwsem_try_write_lock(count)
> > >  spin_unlock(wait_lock)
> > >  schedule()
> 
> Is this what you wanted to write?

And ideally there should be a comment near the atomic_long_add_return()
in __rwsem_down_write_failed_common() to indicate we rely on the implied
smp_mb() before it -- just in case someone goes and makes it
atomic_long_add_return_relaxed().

And I suppose someone should look at the waiting branch of that thing
too.. because I'm not sure what happens if waiting is true but count
isn't big enough.

I bloody hate the rwsem code, that BIAS stuff forever confuses me. I
have a start at rewriting the thing to put the owner in the lock word
just like we now do for mutex, but never seem to get around to finishing
it.

> ---
>  kernel/locking/rwsem-xadd.c | 27 +++++++++++++++++++++++++++
>  1 file changed, 27 insertions(+)
> 
> diff --git a/kernel/locking/rwsem-xadd.c b/kernel/locking/rwsem-xadd.c
> index 02f660666ab8..813b5d3654ce 100644
> --- a/kernel/locking/rwsem-xadd.c
> +++ b/kernel/locking/rwsem-xadd.c
> @@ -613,6 +613,33 @@ struct rw_semaphore *rwsem_wake(struct rw_semaphore *sem)
>  	DEFINE_WAKE_Q(wake_q);
>  
>  	/*
> +	 * __rwsem_down_write_failed_common(sem)
> +	 *   rwsem_optimistic_spin(sem)
> +	 *     osq_unlock(sem->osq)
> +	 *   ...
> +	 *   atomic_long_add_return(&sem->count)
> +	 *
> +	 *		- VS -
> +	 *
> +	 *			__up_write()
> +	 *			  if (atomic_long_sub_return_release(&sem->count) < 0)
> +	 *			    rwsem_wake(sem)
> +	 *			      osq_is_locked(&sem->osq)
> +	 *
> +	 * And __up_write() must observe !osq_is_locked() when it observes the
> +	 * atomic_long_add_return() in order to not miss a wakeup.
> +	 *
> +	 * This boils down to:
> +	 *
> +	 * [S.rel] X = 1		[RmW] r0 = (Y += 0)
> +	 *	   MB			      RMB
> +	 * [RmW]   Y += 1		[L]   r1 = X
> +	 *
> +	 * exists (r0=1 /\ r1=0)
> +	 */
> +	smp_rmb();
> +
> +	/*
>  	 * If a spinner is present, it is not necessary to do the wakeup.
>  	 * Try to do wakeup only if the trylock succeeds to minimize
>  	 * spinlock contention which may introduce too much delay in the