Re: [RFC PATCH] ipc/sem.c: Add one more memory barrier to sem_lock().

[Oleg Nesterov <oleg@redhat.com>]

Sorry Manfred, I initiated this discussion and then disappeared. Currently
I am buried in the ancient 2.16.18 bugs ;)

On 02/25, Manfred Spraul wrote:
> Hi,
>
> What do you think about the following patch for sem_lock()?
>
> Other options:
>
> 1) I don't like
>
> 	#define smp_mb__after_unlock_wait()	smp_rmb()
>
> 	I think it is too specific: the last block in sem_lock uses
>
> 		if (sma->complex_count == 0) {
> 			smp_rmb();
> 			return;
> 		}

See below.

>
> 2) What about
>
> 	#define smp_aquire__after_control_barrier()	smp_rmb()


I agree with any naming. The only point of the new helper is that we can
factor out the comment, otherwise we would need to repeat it again and again.


> @@ -341,7 +359,13 @@ static inline int sem_lock(struct sem_array *sma, struct sembuf *sops,
>  			 * Thus: if is now 0, then it will stay 0.
>  			 */
>  			if (sma->complex_count == 0) {
> -				/* fast path successful! */
> +				/*
> +				 * Fast path successful!
> +				 * We only need a final memory barrier.
> +				 * (see sem_wait_array() for details).
> +				 */
> +				smp_rmb();
> +

I'll try to read this again tomorrow, but so far I am confused.

Most probably I missed something, but this looks unneeded at first glance.

We already have another smp_rmb() above this check. And it should act as
a "final" barrier, or we can not trust this ->complex_count check ?

And (if I am right) this means that the comment above that rmb() should
be updated. And that is why I think the helper makes sense, the comment
should be almost the same as in sem_wait_array().

If not, could you please spell to explain why do we need another rmb() ?

Oleg.