Re: [PATCH 4/4] qspinlock for x86: smp_mb__after_spin_lock() is free

[Peter Zijlstra]

On Sun, Aug 28, 2016 at 01:56:16PM +0200, Manfred Spraul wrote:
> For x86 qspinlocks, no additional memory barrier is required in
> smp_mb__after_spin_lock:
> 
> Theoretically, for qspinlock we could define two barriers:
> - smp_mb__after_spin_lock: Free for x86, not free for powerpc
> - smp_mb__between_spin_lock_and_spin_unlock_wait():
> 	Free for all archs, see queued_spin_unlock_wait for details.
> 
> As smp_mb__between_spin_lock_and_spin_unlock_wait() is not used
> in any hotpaths, the patch does not create that define yet.
> 
> Signed-off-by: Manfred Spraul <manfred@xxxxxxxxxxxxxxxx>
> ---
>  arch/x86/include/asm/qspinlock.h | 11 +++++++++++
>  1 file changed, 11 insertions(+)
> 
> diff --git a/arch/x86/include/asm/qspinlock.h b/arch/x86/include/asm/qspinlock.h
> index eaba080..da06433 100644
> --- a/arch/x86/include/asm/qspinlock.h
> +++ b/arch/x86/include/asm/qspinlock.h
> @@ -61,6 +61,17 @@ static inline bool virt_spin_lock(struct qspinlock *lock)
>  }
>  #endif /* CONFIG_PARAVIRT */
>  
> +#ifndef smp_mb__after_spin_lock
> +/**
> + * smp_mb__after_spin_lock() - Provide smp_mb() after spin_lock
> + *
> + * queued_spin_lock() provides full memory barriers semantics,
> + * thus no further memory barrier is required. See
> + * queued_spin_unlock_wait() for further details.
> + */
> +#define smp_mb__after_spin_lock()	barrier()
> +#endif


I don't get this barrier, and I from my understanding this isn't
correct.

Please explain more.