[RFC PATCH] lib: Introduce generic __cmpxchg_u64() and use it where needed

[Peter Zijlstra]

On Thu, Nov 01, 2018 at 06:46:50PM +0100, Dmitry Vyukov wrote:
> If there is a warning that we don't want to see at all, then we can
> disable it. It supposed to be a useful tool, rather than a thing in
> itself that lives own life. We already I think removed 1 particularly
> noisy warning and made another optional via a config.

> But the thing with overflows is that, even if it's defined, it's not
> necessary the intended behavior. For example, take allocation size
> calculation done via unsigned size_t. If it overflows it does not help
> if C defines result or not, it still gives a user controlled write
> primitive. We've seen similar cases with timeout/deadline calculation
> in kernel, we really don't want it to just wrap modulo-2, right. Some
> user-space projects even test with unsigned overflow warnings or
> implicit truncation warnings, which are formally legal, but frequently
> bugs.

Sure; but then don't call it UB.

If we want to have an additional integer over/underflow checker (ideally
with a gcc plugin that has explicit annotations like __wrap to make it
go away) that is fine; and it can be done on unsigned and signed.