Re: [net-next PATCH v10 7/7] rust: net::phy sync with match_phy_device C changes

["Benno Lossin" <lossin@kernel.org>]

On Sat May 17, 2025 at 8:27 AM CEST, FUJITA Tomonori wrote:
> On Fri, 16 May 2025 22:16:23 +0200
> "Benno Lossin" <lossin@kernel.org> wrote:
>> On Fri May 16, 2025 at 5:12 PM CEST, Christian Marangi wrote:
>>> On Fri, May 16, 2025 at 04:48:53PM +0200, Benno Lossin wrote:
>>>> On Fri May 16, 2025 at 2:30 PM CEST, FUJITA Tomonori wrote:
>>>> > On Thu, 15 May 2025 13:27:12 +0200
>>>> > Christian Marangi <ansuelsmth@gmail.com> wrote:
>>>> >> @@ -574,6 +577,23 @@ pub const fn create_phy_driver<T: Driver>() -> DriverVTable {
>>>> >>  /// This trait is used to create a [`DriverVTable`].
>>>> >>  #[vtable]
>>>> >>  pub trait Driver {
>>>> >> +    /// # Safety
>>>> >> +    ///
>>>> >> +    /// For the duration of `'a`,
>>>> >> +    /// - the pointer must point at a valid `phy_driver`, and the caller
>>>> >> +    ///   must be in a context where all methods defined on this struct
>>>> >> +    ///   are safe to call.
>>>> >> +    unsafe fn from_raw<'a>(ptr: *const bindings::phy_driver) -> &'a Self
>>>> >> +    where
>>>> >> +        Self: Sized,
>>>> >> +    {
>>>> >> +        // CAST: `Self` is a `repr(transparent)` wrapper around `bindings::phy_driver`.
>>>> >> +        let ptr = ptr.cast::<Self>();
>>>> >> +        // SAFETY: by the function requirements the pointer is valid and we have unique access for
>>>> >> +        // the duration of `'a`.
>>>> >> +        unsafe { &*ptr }
>>>> >> +    }
>>>> >
>>>> > We might need to update the comment. phy_driver is const so I think
>>>> > that we can access to it any time.
>>>> 
>>>> Why is any type implementing `Driver` a transparent wrapper around
>>>> `bindings::phy_driver`?
>>>> 
>>>
>>> Is this referred to a problem with using from_raw or more of a general
>>> question on how the rust wrapper are done for phy code?
>> 
>> I looked at the `phy.rs` file again and now I'm pretty sure the above
>> code is wrong. `Self` can be implemented on any type (even types like
>> `Infallible` that do not have any valid bit patterns, since it's an
>> empty enum). The abstraction for `bindings::phy_driver` is
>> `DriverVTable` not an object of type `Self`, so you should cast to that
>> pointer instead.
>
> Yeah.
>
> I don't want to delay this patchset due to Rust side changes so
> casting a pointer to bindings::phy_driver to DriverVTable is ok but
> the following signature doesn't look useful for Rust phy drivers:
>
> fn match_phy_device(_dev: &mut Device, _drv: &DriverVTable) -> bool
>
> struct DriverVTable is only used to create an array of
> bindings::phy_driver for C side, and it doesn't provide any
> information to the Rust driver.

Yeah, but we could add accessor functions that provide that information.
Although that doesn't really make sense at the moment, see below.

> In match_phy_device(), for example, a device driver accesses to
> PHY_DEVICE_ID, which the Driver trait provides. I think we need to
> create an instance of the device driver's own type that implements the
> Driver trait and make it accessible.

I think that's wrong, nothing stops me from implementing `Driver` for an
empty enum and that can't be instantiated. The reason that one wants to
have this in C is because the same `match` function is used for
different drivers (or maybe devices? I'm not too familiar with the
terminology). In Rust, you must implement the match function for a
single PHY_DEVICE_ID only, so maybe we don't need to change the
signature at all?

---
Cheers,
Benno