Re: [PATCH net-next v7 1/5] rust: core abstractions for network PHY drivers

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

On Fri, Oct 27, 2023 at 09:19:38PM +0000, Benno Lossin wrote:
[...]
> 
> Good catch. `phy_device` is rather large (did not look at the exact
> size) and this will not be optimized on debug builds, so it could lead
> to stackoverflows.
> 

IIRC, kernel only supports O2 build, but yes, if we don't want the copy
here, we should avoid the copy semantics.

> > 	struct phy_device phydev = *ptr;
> > 
> > Sure, both compilers can figure this out, therefore no extra copy is
> > done, but still it's better to avoid this copy semantics by doing:
> > 
> > 	let phydev = unsafe { &*self.0.get() };
> 
> We need to be careful here, since doing this creates a reference
> `&bindings::phy_device` which asserts that it is immutable. That is not
> the case, since the C side might change it at any point (this is the
> reason we wrap things in `Opaque`, since that allows mutatation even
> through sharde references).
> 
> I did not notice this before, but this means we cannot use the `link`
> function from bindgen, since that takes `&self`. We would need a
> function that takes `*const Self` instead.
> 

Hmm... but does it mean even `set_speed()` has the similar issue?

	let phydev: *mut phy_device = self.0.get();
	unsafe { (*phydev).speed = ...; }

The `(*phydev)` creates a `&mut` IIUC. So we need the following maybe?

	let phydev: *mut phy_device = self.0.get();
	unsafe { *addr_mut_of!((*phydev).speed) = ...; }

because at least from phylib core guarantee, we know no one accessing
`speed` in the same time. However, yes, bit fields are tricky...

Regards,
Boqun

> -- 
> Cheers,
> Benno
>