[concurrency-interest] Semantics of compareAndSwapX

[Andrew Haley]

On 02/14/2014 02:46 PM, Doug Lea wrote:
> On 02/14/2014 08:12 AM, Andrew Haley wrote:
>> On 02/14/2014 11:52 AM, David Holmes wrote:
>>> Andrew Haley writes:
>>>>
>>>> What are the semantics of Unsafe.compareAndSwapX?  The javadoc is
>>>> rather thin.
>>>
>>> They are the implementation for the associated AtomicX.compareAndSet methods
>>> and so have to adhere to the specification of those methods. But in terms of
>>> memory barriers the key factor is that the variables act as volatiles so the
>>> semantics are of a volatile read combined with a volatile write.
>>
>> Yes, but is that all?
> 
> That's all wrt ordering effects, but CAS must also be atomic.
> For ARMv8/Aarch64 (which I suspect you have in mind), the best mapping
> is likely the one now used in linux.
> (http://lists.infradead.org/pipermail/linux-arm-kernel/2014-February/229588.html)
> 
> 	<Access [A]>
> 
> 	// atomic_op (B)
> 1:	ldxr	x0, [B]		// Exclusive load
> 	<op(B)>
> 	stlxr	w1, x0, [B]	// Exclusive store with release
> 	cbnz	w1, 1b
> 	dmb	ish		// Full barrier
> 
> 	<Access [C]>

Okay, but if you have a look at the patch in that email you will see
that the sequence for CAS is different:

static inline long atomic64_cmpxchg(atomic64_t *ptr, long old, long new)
 	long oldval;
 	unsigned long res;

+	smp_mb();
+
 	asm volatile("// atomic64_cmpxchg\n"
-"1:	ldaxr	%1, %2\n"
+"1:	ldxr	%1, %2\n"
 "	cmp	%1, %3\n"
 "	b.ne	2f\n"
-"	stlxr	%w0, %4, %2\n"
+"	stxr	%w0, %4, %2\n"
 "	cbnz	%w0, 1b\n"
 "2:"
 	: "=&r" (res), "=&r" (oldval), "+Q" (ptr->counter)
 	: "Ir" (old), "r" (new)
 	: "cc", "memory");

+	smp_mb();
 	return oldval;

Note that this has a full barrier at each end.

> Aside: The JMM update is likely to finally pin down
> multiple modes of CAS/WeakCAS (full, acquire, release),
> which are not all that common, but are specializable into
> more efficient mappings on some processors.

Ah, good.

Andrew.