Re: [PATCH] x86/cmpxchg: Add safety for bad sizes

[Jan Beulich <jbeulich@xxxxxxxx>]

On 27.01.2026 11:23, Andrew Cooper wrote:
> @@ -66,6 +68,8 @@ static always_inline unsigned long __xchg(
>                         : [x] "+r" (x), [ptr] "+m" (*(volatile uint64_t *)ptr)
>                         :: "memory" );
>          break;
> +    default:
> +        __bad_xchg_size();

What has come of the plans to emit an assembly error directive in such
situations?

Also for Misra's sake "break" will be wanted.

> @@ -106,6 +110,8 @@ static always_inline unsigned long __cmpxchg(
>                         : [new] "r" (new), "a" (old)
>                         : "memory" );
>          return prev;
> +    default:
> +        BUG();
>      }
>      return old;
>  }
> @@ -137,6 +143,8 @@ static always_inline unsigned long cmpxchg_local_(
>                         : "=a" (prev), [ptr] "+m" (*(uint64_t *)ptr)
>                         : [new] "r" (new), "a" (old) );
>          break;
> +    default:
> +        BUG();
>      }
>  
>      return prev;

Hmm. If for some reason hvmemul_cmpxchg() ended up hitting either of these,
we'd immediately have an XSA. Imo these want to be ASSERT_UNREACHABLE()
with plausible recovery for release builds.

Jan