Re: [PATCH 1.5/5] alpha: atomic: provide asm for the fastpath for _atomic_dec_and_lock_irqsave

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Jun 04, 2018 at 12:27:57PM +0200, Sebastian Andrzej Siewior wrote:
> I just looked at Alpha's atomic_dec_and_lock assembly and did something
> that should work for atomic_dec_and_lock_irqsave. I think it works but I
> would prefer for someone from the Alpha-Camp to ack this before it goes
> in. It is not critical because the non-optimized version should work.

I would vote to simply delete this entire file and get alpha on the
generic code.

Afaict, this asm gets the ordering wrong, and I doubt it is much faster
than using atomic_add_unless() in any case (+- the ordering of course).

>  arch/alpha/lib/dec_and_lock.c | 33 ++++++++++++++++++++++++++-------
>  1 file changed, 26 insertions(+), 7 deletions(-)
> 
> diff --git a/arch/alpha/lib/dec_and_lock.c b/arch/alpha/lib/dec_and_lock.c
> index 069fef7372dc..d29ba1de4f68 100644
> --- a/arch/alpha/lib/dec_and_lock.c
> +++ b/arch/alpha/lib/dec_and_lock.c
> @@ -32,6 +32,28 @@ _atomic_dec_and_lock:				\n\
>  	.previous				\n\
>  	.end _atomic_dec_and_lock");
>  
> +  asm (".text					\n\
> +	.global _atomic_dec_and_lock_irqsave	\n\
> +	.ent _atomic_dec_and_lock_irqsave	\n\
> +	.align	4				\n\
> +_atomic_dec_and_lock_irqsave:			\n\
> +	.prologue 0				\n\
> +1:	ldl_l	$1, 0($16)			\n\
> +	subl	$1, 1, $1			\n\
> +	beq	$1, 2f				\n\
> +	stl_c	$1, 0($16)			\n\
> +	beq	$1, 4f				\n\
> +	mb					\n\
> +	clr	$0				\n\
> +	ret					\n\
> +2:	br	$29, 3f				\n\
> +3:	ldgp	$29, 0($29)			\n\
> +	br	$atomic_dec_and_lock_irqsave1..ng	\n\
> +	.subsection 2				\n\
> +4:	br	1b				\n\
> +	.previous				\n\
> +	.end _atomic_dec_and_lock_irqsave");
> +
>  static int __used atomic_dec_and_lock_1(atomic_t *atomic, spinlock_t *lock)
>  {
>  	/* Slow path */
> @@ -43,14 +65,11 @@ static int __used atomic_dec_and_lock_1(atomic_t *atomic, spinlock_t *lock)
>  }
>  EXPORT_SYMBOL(_atomic_dec_and_lock);
>  
> -int _atomic_dec_and_lock_irqsave(atomic_t *atomic, spinlock_t *lock,
> -				 unsigned long *flags)
> +static int __used atomic_dec_and_lock_irqsave1(atomic_t *atomic,
> +					       spinlock_t *lock,
> +					       unsigned long *flags)
>  {
> -	/* Subtract 1 from counter unless that drops it to 0 (ie. it was 1) */
> -	if (atomic_add_unless(atomic, -1, 1))
> -		return 0;
> -
> -	/* Otherwise do it the slow way */
> +	/* Slow way */
>  	spin_lock_irqsave(lock, *flags);
>  	if (atomic_dec_and_test(atomic))
>  		return 1;
> -- 
> 2.17.1
>