Re: [PATCH] rust: drm: Drop the use of Opaque for ioctl arguments

[Boqun Feng <boqun.feng@gmail.com>]

On Thu, Jun 19, 2025 at 03:17:31PM +0200, Benno Lossin wrote:
> On Thu Jun 19, 2025 at 2:26 PM CEST, Daniel Almeida wrote:
> > Hi Benno,
> >
> >> On 19 Jun 2025, at 08:01, Benno Lossin <lossin@kernel.org> wrote:
> >> 
> >> On Thu Jun 19, 2025 at 12:55 PM CEST, Danilo Krummrich wrote:
> >>> On Thu, Jun 19, 2025 at 12:21:02PM +0200, Beata Michalska wrote:
> >>>> diff --git a/rust/kernel/drm/ioctl.rs b/rust/kernel/drm/ioctl.rs
> >>>> index 445639404fb7..12b296131672 100644
> >>>> --- a/rust/kernel/drm/ioctl.rs
> >>>> +++ b/rust/kernel/drm/ioctl.rs
> >>>> @@ -139,7 +139,7 @@ pub mod internal {
> >>>>                             // asserted above matches the size of this type, and all bit patterns of
> >>>>                             // UAPI structs must be valid.
> >>>>                             let data = unsafe {
> >>>> -                                &*(raw_data as *const $crate::types::Opaque<$crate::uapi::$struct>)
> >>>> +                                &mut *(raw_data as *mut $crate::uapi::$struct)
> >>> 
> >>> I think we have to document the guarantees we rely on to create this mutable
> >>> reference.
> >> 
> >> If the C side is using pointers to read/write the value concurrently,
> >> this is wrong, it needs to be wrapped in Opaque.
> >> 
> >> ---
> >> Cheers,
> >> Benno
> >
> > How can this happen, exactly? Can you provide an example that corroborates it?
> 
> I don't have the context on this, I only saw a raw pointer being turned
> into a mutable reference and that's only possible if there are no shared
> or other exclusive references for the duration of its existence and no
> raw pointers are being used to access the value.
> 

I think in this case, as Daniel described below, `data` points to a
buffer either on the stack of drm_ioctl() function or allocated from
kmalloc() in drm_ioctl(), and drm_ioctl() copies userspace data into
that buffer, so at the this point, the data should be owned solely by
this function. But of course the safety comments need to be adjusted.

Regards,
Boqun

> ---
> Cheers,
> Benno
> 
> > The general pattern for drivers is to fill an uapi type and then wait on an
> > ioctl. The kernel then copies that using copy_from_user, so we're safe from
> > that perspective (i.e.: from the perspective of concurrent access from
> > userspace).
> >
> > In kernelspace, we usually extract arguments from the uapi types to then
> > dictate further processing inside drivers. In what way are these shared with
> > "the C side" ?
> >
> > If the result of this discussion is that we agree that this Opaque is not
> > needed, then we definitely need this patch, because using Opaque<T> complicates
> > all ioctls implementations by making it harder to get to the inner T in the
> > first place. We would have to needlessly add a lot of unsafe blocks for drivers
> > that wouldn't otherwise be there.
> >
> >
> > -- Daniel
>