Re: [PATCH v2] uaccess: rust: use newtype for user pointers

[Alice Ryhl <aliceryhl@google.com>]

On Tue, May 27, 2025 at 04:13:03PM -0700, Boqun Feng wrote:
> On Tue, May 27, 2025 at 11:12:11PM +0100, Al Viro wrote:
> > On Tue, May 27, 2025 at 01:53:12PM +0000, Alice Ryhl wrote:
> > > In C code we use sparse with the __user annotation to detect cases where
> > > a user pointer is mixed up with other things. To replicate that, we
> > > introduce a new struct UserPtr that serves the same purpose using the
> > > newtype pattern.
> > > 
> > > The UserPtr type is not marked with #[derive(Debug)], which means that
> > > it's not possible to print values of this type. This avoids ASLR
> > > leakage.
> > > 
> > > The type is added to the prelude as it is a fairly fundamental type
> > > similar to c_int. The wrapping_add() method is renamed to
> > > wrapping_byte_add() for consistency with the method name found on raw
> > > pointers.
> > 
> > That's considerably weaker than __user, though - with
> > 	struct foo {struct bar x; struct baz y[2]; };
> 
> Translate to Rust this is:
> 
>     struct Foo {
>         x: Bar,
> 	y: Baz[2],
>     }
> 
> > 	struct foo __user *p;
> 
> UserPtr should probably be generic over pointee, so:
> 
>     pub struct UserPtr<T>(*mut c_void, PhantomData<*mut T>);
> 
> and
> 
>     let p: UserPtr<Foo> = ...;
> 
> > 	void f(struct bar __user *);
> 
> and this is:
> 
>     pub fn f(bar: UserPtr<Bar>)
> 
> and the checking should work, a (maybe unrelated) tricky part though..
> 
> > sparse does figure out that f(&p->y[1]) is a type error - &p->y[1] is
> 
> In Rust, you will need to play a little unsafe game to get &p->y[1]:
> 
>     let foo_ptr: *mut Foo = p.as_mut_ptr();
>     let y_ptr: *mut Baz = unsafe { addr_of_mut!((*foo_ptr).y[1]) };
>     let y: UserPtr<Baz> = unsafe { UserPtr::from_ptr(y_ptr) };
> 
> passing y to f() will get a type mismatch, so the detection/checking
> works. To avoid the unsafe game we need field projection [1].

That looks pretty unergonomic. We can do better than that.

Alice

> > struct baz __user * and f() expects struct bar __user *.
> > 
> > It's not just mixing userland pointers with other things - it's not mixing
> > userland pointers to different types, etc.
> > 
> 
> In short, with UserPtr generic over pointee, we can have the similar
> detection as sparse.
> 
> [1]: https://github.com/rust-lang/rfcs/pull/3735
> 
> Regards,
> Boqun
> 
> > In practice I've seen quite a few brainos caught by that...