Hi Thomas,

On Thu, 02 Apr 2020 00:39:55 +0200 Thomas Gleixner <tglx@linutronix.de> wrote:
>
> Stephen Rothwell <sfr@canb.auug.org.au> writes:
> > On Wed, 01 Apr 2020 12:25:25 +0200 Thomas Gleixner <tglx@linutronix.de> wrote:  
> >> Me neither. Which compiler version?  
> >
> > arm-linux-gnueabi-gcc (Debian 9.2.1-21) 9.2.1 20191130
> >  
> >> I'm using arm-linux-gnueabi-gcc (Debian 8.3.0-2) 8.3.0 which does not
> >> show that oddity.  
> >
> > I assume it is because of the change to arch_futex_atomic_op_inuser()
> > for arm and the compiler is not clever enough to work out that the early
> > return from arch_futex_atomic_op_inuser() means that oldval is not
> > referenced in its caller.  
> 
> Actually no. It's the ASM part which causes this. With the following
> hack applied it compiles:
> 
> diff --git a/arch/arm/include/asm/futex.h b/arch/arm/include/asm/futex.h
> index e133da303a98..2c6b40f71009 100644
> --- a/arch/arm/include/asm/futex.h
> +++ b/arch/arm/include/asm/futex.h
> @@ -132,7 +132,7 @@ futex_atomic_cmpxchg_inatomic(u32 *uval, u32 __user *uaddr,
>  static inline int
>  arch_futex_atomic_op_inuser(int op, int oparg, int *oval, u32 __user *uaddr)
>  {
> -	int oldval = 0, ret, tmp;
> +	int oldval = 0, ret;
>  
>  	if (!access_ok(uaddr, sizeof(u32)))
>  		return -EFAULT;
> @@ -142,6 +142,7 @@ arch_futex_atomic_op_inuser(int op, int oparg, int *oval, u32 __user *uaddr)
>  #endif
>  
>  	switch (op) {
> +#if 0
>  	case FUTEX_OP_SET:
>  		__futex_atomic_op("mov	%0, %4", ret, oldval, tmp, uaddr, oparg);
>  		break;
> @@ -157,6 +158,7 @@ arch_futex_atomic_op_inuser(int op, int oparg, int *oval, u32 __user *uaddr)
>  	case FUTEX_OP_XOR:
>  		__futex_atomic_op("eor	%0, %1, %4", ret, oldval, tmp, uaddr, oparg);
>  		break;
> +#endif
>  	default:
>  		ret = -ENOSYS;
>  	}
> 
> but with this is emits the warning:
> 
> diff --git a/arch/arm/include/asm/futex.h b/arch/arm/include/asm/futex.h
> index e133da303a98..5191d7b61b83 100644
> --- a/arch/arm/include/asm/futex.h
> +++ b/arch/arm/include/asm/futex.h
> @@ -145,6 +145,7 @@ arch_futex_atomic_op_inuser(int op, int oparg, int *oval, u32 __user *uaddr)
>  	case FUTEX_OP_SET:
>  		__futex_atomic_op("mov	%0, %4", ret, oldval, tmp, uaddr, oparg);
>  		break;
> +#if 0
>  	case FUTEX_OP_ADD:
>  		__futex_atomic_op("add	%0, %1, %4", ret, oldval, tmp, uaddr, oparg);
>  		break;
> @@ -157,6 +158,7 @@ arch_futex_atomic_op_inuser(int op, int oparg, int *oval, u32 __user *uaddr)
>  	case FUTEX_OP_XOR:
>  		__futex_atomic_op("eor	%0, %1, %4", ret, oldval, tmp, uaddr, oparg);
>  		break;
> +#endif
>  	default:
>  		ret = -ENOSYS;
>  	}
> 
> and the below proves it:
> 
> diff --git a/arch/arm/include/asm/futex.h b/arch/arm/include/asm/futex.h
> index e133da303a98..a9151884bc85 100644
> --- a/arch/arm/include/asm/futex.h
> +++ b/arch/arm/include/asm/futex.h
> @@ -165,8 +165,13 @@ arch_futex_atomic_op_inuser(int op, int oparg, int *oval, u32 __user *uaddr)
>  	preempt_enable();
>  #endif
>  
> -	if (!ret)
> -		*oval = oldval;
> +	/*
> +	 * Store unconditionally. If ret != 0 the extra store is the least
> +	 * of the worries but GCC cannot figure out that __futex_atomic_op()
> +	 * is either setting ret to -EFAULT or storing the old value in
> +	 * oldval which results in a uninitialized warning at the call site.
> +	 */
> +	*oval = oldval;
>  
>  	return ret;
>  }
> 
> I think that's the right thing to do anyway. The conditional is pointless.

Thanks for the analysis.

I am still getting this warning, now from Linus' tree builds.

-- 
Cheers,
Stephen Rothwell