[lttng-dev] [PATCH 5/7] Use __atomic builtins to implement CMM_{LOAD, STORE}_SHARED

[Mathieu Desnoyers]

On 2023-03-17 17:37, Ondřej Surý via lttng-dev wrote:
> Instead of using CMM_ACCESS_ONCE() with memory barriers, use __atomic
> builtins with relaxed memory ordering to implement CMM_LOAD_SHARED() and
> CMM_STORE_SHARED().
> 
> Signed-off-by: Ondřej Surý <ondrej at sury.org>
> ---
>   include/urcu/system.h | 7 +++----
>   1 file changed, 3 insertions(+), 4 deletions(-)
> 
> diff --git a/include/urcu/system.h b/include/urcu/system.h
> index faae390..99e7443 100644
> --- a/include/urcu/system.h
> +++ b/include/urcu/system.h
> @@ -26,7 +26,7 @@
>    * Identify a shared load. A cmm_smp_rmc() or cmm_smp_mc() should come
>    * before the load.
>    */
> -#define _CMM_LOAD_SHARED(p)	       CMM_ACCESS_ONCE(p)
> +#define _CMM_LOAD_SHARED(p)	       __atomic_load_n(&(p), __ATOMIC_RELAXED)
>   
>   /*
>    * Load a data from shared memory, doing a cache flush if required.
> @@ -42,7 +42,7 @@
>    * Identify a shared store. A cmm_smp_wmc() or cmm_smp_mc() should
>    * follow the store.
>    */
> -#define _CMM_STORE_SHARED(x, v)	__extension__ ({ CMM_ACCESS_ONCE(x) = (v); })
> +#define _CMM_STORE_SHARED(x, v)	__atomic_store_n(&(x), (v), __ATOMIC_RELAXED)

__atomic_store_n() is void. _CMM_STORE_SHARED() should evaluate to (v) 
(unless we decide to change the semantic, which I would rather avoid).

Thanks,

Mathieu

>   
>   /*
>    * Store v into x, where x is located in shared memory. Performs the
> @@ -51,9 +51,8 @@
>   #define CMM_STORE_SHARED(x, v)						\
>   	__extension__							\
>   	({								\
> -		__typeof__(x) _v = _CMM_STORE_SHARED(x, v);		\
> +		_CMM_STORE_SHARED(x, v);				\
>   		cmm_smp_wmc();						\
> -		_v = _v;	/* Work around clang "unused result" */	\
>   	})
>   
>   #endif /* _URCU_SYSTEM_H */

-- 
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com