On Wed, Nov 08, 2023 at 06:04:47PM +0100, Thomas Monjalon wrote:
> 02/11/2023 04:04, Tyler Retzlaff:
> > Replace use of __atomic_thread_fence with __rte_atomic_thread_fence.
> >
> > It may be appropriate to use rte_atomic_thread_fence instead but it
> > will be up to maintainers to evaluate and make the change if appropriate.
>
> I don't understand the use of __rte_atomic_thread_fence
> which is supposed to be EAL-internal only, isn't it?
>
> On x86, we have this:
> static __rte_always_inline void
> rte_atomic_thread_fence(rte_memory_order memorder)
> {
> if (memorder == rte_memory_order_seq_cst)
> rte_smp_mb();
> else
> __rte_atomic_thread_fence(memorder);
> }
>
> This is skipped if you use __rte_atomic_thread_fence() directly.
correct. that is on purpose because the original code was skipping
condition by using __atomic_thread_fence directly.
this series intends no functional change which is why the replacements
are __rte_atomic_thread_fence instead of to rte_atomic_thread_fence.
i would encourage the maintainers to evaluate whether the code can use
rte_atomic_thread_fence directly without functional regression.
ty
