Re: [PATCH] x86/mmuext: Unify okay/rc error handling in do_mmuext_op()

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

On 22/12/2015 08:57, Jan Beulich wrote:
>>>> On 21.12.15 at 18:16, <andrew.cooper3@citrix.com> wrote:
>> c/s 506db90 "x86/HVM: merge HVM and PVH hypercall tables" introduced a path
>> whereby 'okay' was used uninitialised, with broke compilation on CentOS 7.
> It appeared to be used uninitialized, but wasn't in fact (i.e. the
> outcome - the value rc gets set to - didn't depend on the value
> due to
>
>         if ( unlikely(!okay) && !rc )
>             rc = -EINVAL;
>
> being equivalent to
>
>         if ( !rc && unlikely(!okay) )
>             rc = -EINVAL;
>
> (no side effects for the expressions on either side of the &&).
> I'll re-word accordingly upon committing, to not give the false
> impression of there having been other than a cosmetic problem.

There is a real problem.  Because the compiler is able to prove that
okay is genuinely read uninitialised in one case, the rules concerning
undefined behaviour permit it to do anything it wishes, including
omitting this if statement.

As far as practical problems go however, it is the build breakage which
is relevant, and it breaks because of a -Werror=maybe-uninitialised.

~Andrew