On Fri, Feb 08, 2019 at 05:17:39PM +0100, Bernard Metzler wrote: > I agree. In siw READ_ONCE/WRITE_ONCE is used, or the appropriate > macros, which add memory barriers before or after, as needed. > I currently clean up a few occurrences of explicit mb's. > Why do we need an extra macro like proposed RDMA_WRITE_UAPI_ATOMIC? > Is it to catch cases where one wants to introduce non-atomic values > to be shared? I think this was to get some kind of type safety so the compiler would complain if people touch the atomics wrongly. Maybe we don't need it. Jason
