Re: [PATCH v2] locking/rwsem: add acquire barrier to read_slowpath exit when queue is empty

[Will Deacon]

On Thu, Jul 18, 2019 at 12:58:12PM +0200, Peter Zijlstra wrote:
> On Thu, Jul 18, 2019 at 10:26:41AM +0100, Will Deacon wrote:
> 
> > /*
> >  * We need to ensure ACQUIRE semantics when reading sem->count so that
> >  * we pair with the RELEASE store performed by an unlocking/downgrading
> >  * writer.
> >  *
> >  * P0 (writer)			P1 (reader)
> >  *
> >  * down_write(sem);
> >  * <write shared data>
> >  * downgrade_write(sem);
> >  * -> fetch_add_release(&sem->count)
> >  *
> >  *				down_read_slowpath(sem);
> >  *				-> atomic_read(&sem->count)
> >  *				   <ctrl dep>
> >  *				   smp_acquire__after_ctrl_dep()
> >  *				<read shared data>
> >  */
> 
> So I'm thinking all this is excessive; the simple rule is: lock acquire
> should imply ACQUIRE, we all know why.

Fair enough, I just thought this was worth highlighting because you can't
reply on the wait_lock to give you ACQUIRE ordering.

> > In writing this, I also noticed that we don't have any explicit ordering
> > at the end of the reader slowpath when we wait on the queue but get woken
> > immediately:
> > 
> > 	if (!waiter.task)
> > 		break;
> > 
> > Am I missing something?
> 
> Ha!, I ran into the very same one. I keep confusing myself, but I think
> you're right and that needs to be smp_load_acquire() to match the
> smp_store_release() in rwsem_mark_wake().
> 
> (the actual race there is _tiny_ due to the smp_mb() right before it,
> but I cannot convince myself that is indeed sufficient)
> 
> The signal_pending_state() case is also fun, but I think wait_lock there
> is sufficient (even under RCpc).
> 
> I've ended up with this..
> 
> ---
> diff --git a/kernel/locking/rwsem.c b/kernel/locking/rwsem.c
> index 37524a47f002..9eb630904a17 100644
> --- a/kernel/locking/rwsem.c
> +++ b/kernel/locking/rwsem.c
> @@ -1000,6 +1000,7 @@ rwsem_down_read_slowpath(struct rw_semaphore *sem, int state)
>  	atomic_long_add(-RWSEM_READER_BIAS, &sem->count);
>  	adjustment = 0;
>  	if (rwsem_optimistic_spin(sem, false)) {
> +		/* rwsem_optimistic_spin() implies ACQUIRE through rwsem_*trylock() */

I couldn't figure out if this was dependent on the return value or not,
and looking at osq_lock() I also couldn't see the ACQUIRE barrier when we're
spinning on node->locked. Hmm.

>  		/*
>  		 * Wake up other readers in the wait list if the front
>  		 * waiter is a reader.
> @@ -1014,6 +1015,7 @@ rwsem_down_read_slowpath(struct rw_semaphore *sem, int state)
>  		}
>  		return sem;
>  	} else if (rwsem_reader_phase_trylock(sem, waiter.last_rowner)) {
> +		/* rwsem_reader_phase_trylock() implies ACQUIRE */

Can we add "on success" to the end of this, please?

>  		return sem;
>  	}
>  
> @@ -1032,6 +1034,8 @@ rwsem_down_read_slowpath(struct rw_semaphore *sem, int state)
>  		 */
>  		if (adjustment && !(atomic_long_read(&sem->count) &
>  		     (RWSEM_WRITER_MASK | RWSEM_FLAG_HANDOFF))) {
> +			/* Provide lock ACQUIRE */
> +			smp_acquire__after_ctrl_dep();
>  			raw_spin_unlock_irq(&sem->wait_lock);
>  			rwsem_set_reader_owned(sem);
>  			lockevent_inc(rwsem_rlock_fast);
> @@ -1065,15 +1069,25 @@ rwsem_down_read_slowpath(struct rw_semaphore *sem, int state)
>  	wake_up_q(&wake_q);
>  
>  	/* wait to be given the lock */
> -	while (true) {
> +	for (;;) {
>  		set_current_state(state);
> -		if (!waiter.task)
> +		if (!smp_load_acquire(&waiter.task)) {
> +			/*
> +			 * Matches rwsem_mark_wake()'s smp_store_release() and ensures
> +			 * we're ordered against its sem->count operations.
> +			 */
>  			break;
> +		}

Ack. Also, grepping for 'waiter.task' reveals a similar usage in
drivers/tty/tty_ldsem.c if you're feeling brave enough.

Will