Re: [PATCH for 4.6 03/13] xen: Introduce ACCESS_ONCE macro

[Andrew Cooper <andrew.cooper3@citrix.com>]

On 17/12/14 17:10, Jan Beulich wrote:
>>>> Julien Grall <julien.grall@linaro.org> 12/17/14 1:55 PM >>>
>> On 17/12/14 10:05, Jan Beulich wrote:
>>>>> On 16.12.14 at 21:08, <julien.grall@linaro.org> wrote:
>>>> +#define ACCESS_ONCE(x) (*(volatile typeof(x) *)&(x))
>>> Any reason not to simply use {read,write}_atomic() instead, which we
>>> already have?
>> To avoid modifying Linux drivers when it's not necessary and doesn't harm.
> I realize that's the motivation, but I also view it as problematic to have two
> different constructs doing the same thing. Defining the new one in terms of
> the existing ones doesn't seem possible (or else I would suggest that in
> order for the connection to be obvious). We'll see what other maintainers
> think...

Personally, I find the semantics of ACCESS_ONCE() more intuitive than
read/write_atomic(), and it is certainly more familiar to Linux developers.

Furthermore, ACCESS_ONCE() doesn't force an mov instruction if the
compiler can identify a better instruction to use.

There are only a handful of user users of read/write_atomic().  It would
not be hard to make a blanket switch, if we chose to go in that direction.

~Andrew