Re: [PATCH 2/6] locking/rwsem: Document barrier need when waking tasks

[Peter Zijlstra <peterz@infradead.org>]

On Sun, Jan 25, 2015 at 11:36:05PM -0800, Davidlohr Bueso wrote:
> The need for the smp_mb in __rwsem_do_wake should be
> properly documented. Applies to both xadd and spinlock
> variants.
> 
> Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
> ---
>  kernel/locking/rwsem-spinlock.c | 5 +++++
>  kernel/locking/rwsem-xadd.c     | 5 +++++
>  2 files changed, 10 insertions(+)
> 
> diff --git a/kernel/locking/rwsem-spinlock.c b/kernel/locking/rwsem-spinlock.c
> index 2555ae1..54f7a17 100644
> --- a/kernel/locking/rwsem-spinlock.c
> +++ b/kernel/locking/rwsem-spinlock.c
> @@ -85,6 +85,11 @@ __rwsem_do_wake(struct rw_semaphore *sem, int wakewrite)
>  
>  		list_del(&waiter->list);
>  		tsk = waiter->task;
> +		/*
> +		 * Ensure that all cores see the read before
> +		 * setting it to the waiter task to nil, as that
> +		 * grants the read lock to the next task.
> +		 */
>  		smp_mb();
>  		waiter->task = NULL;
>  		wake_up_process(tsk);

Could you enhance that comment by pointing at the pairing code? Is that
the wait loop in rwsem_down_read_failed()?

Also, the comment confuses, how can all cores observe a read into a
local variable?