Re: [RFC PATCH 7/9] um: nommu: add SMP futex operations

On Mon, 2026-07-20 at 20:26 +0200, Johannes Berg wrote:
>=20
> +static inline int
> +arch_futex_atomic_op_inuser(int op, u32 oparg, int *oval, u32 __user *ua=
ddr)
> +{
> +	u32 oldval, newval;
> +	u32 *p =3D (u32 __force *)uaddr;
> +
> +	do {
> +		oldval =3D READ_ONCE(*p);
> +
> +		switch (op) {
> +		case FUTEX_OP_SET:
> +			newval =3D oparg;
> +			break;
> +		case FUTEX_OP_ADD:
> +			newval =3D oldval + oparg;
> +			break;
> +		case FUTEX_OP_OR:
> +			newval =3D oldval | oparg;
> +			break;
> +		case FUTEX_OP_ANDN:
> +			newval =3D oldval & ~oparg;
> +			break;
> +		case FUTEX_OP_XOR:
> +			newval =3D oldval ^ oparg;
> +			break;
> +		default:
> +			return -ENOSYS;
> +		}
> +	} while (!try_cmpxchg(p, &oldval, newval));
> +
> +	*oval =3D oldval;

Wait ... this is also garbage, it should just be the switch() and then

	*oval =3D cmpxchg(p, oldval, newval);

not the loop...

> +	/* FIXME - shouldn't it be *uval =3D cmpxchg()? */
> +	*uval =3D *p;
> +	cmpxchg(p, oldval, newval);

And yes, I think the FIXME is right and our uaccess.c implementation is
wrong.

johannes