Re: [PATCH 4/6] ARC: LLOCK/SCOND based rwlock

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Aug 03, 2015 at 03:33:06PM +0530, Vineet Gupta wrote:
> With LLOCK/SCOND, the rwlock counter can be atomically updated w/o need
> for a guarding spin lock.

Maybe re-iterate the exclusive vs shared spin story again.

And aside from the far too many full barriers (again), I was just
wondering about:

> +static inline void arch_write_unlock(arch_rwlock_t *rw)
> +{
> +	unsigned int val;
> +
> +	smp_mb();
> +
> +	/*
> +	 * rw->counter = __ARCH_RW_LOCK_UNLOCKED__;
> +	 */
> +	__asm__ __volatile__(
> +	"1:	llock	%[val], [%[rwlock]]	\n"
> +	"	scond	%[UNLOCKED], [%[rwlock]]\n"
> +	"	bnz	1b			\n"
> +	"					\n"
> +	: [val]		"=&r"	(val)
> +	: [rwlock]	"r"	(&(rw->counter)),
> +	  [UNLOCKED]	"r"	(__ARCH_RW_LOCK_UNLOCKED__)
> +	: "memory", "cc");
> +
> +	smp_mb();
> +}

Why can't that be a straight store?