Re: [PATCH 0/4] Re-export Zeroable and zeroed() from ffi module

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

On Mon Jan 26, 2026 at 5:14 PM CET, Alice Ryhl wrote:
> On Mon, Jan 26, 2026 at 5:11 PM Gary Guo <gary@garyguo.net> wrote:
>>
>> On Mon Jan 26, 2026 at 4:08 PM GMT, Benno Lossin wrote:
>> > On Mon Jan 26, 2026 at 2:25 PM CET, Alice Ryhl wrote:
>> >> On Mon, Jan 26, 2026 at 01:15:53PM +0000, Gary Guo wrote:
>> >>> On Mon Jan 26, 2026 at 1:05 PM GMT, Alice Ryhl wrote:
>> >>> > Currently, the Zeroable trait is defined by pin-init because pin-init
>> >>> > happens to use the trait. However, zeroed types are useful for many
>> >>> > purposes other than pin-init. Also, we wish to implement Zeroable for
>> >>> > types generated by bindgen. For both of these reasons, re-export
>> >>> > Zeroable from the ffi crate, which is a already dependency of the crates
>> >>> > with bindgen output.
>> >>>
>> >>> I don't see a benefit of re-exporting these from the `ffi` crate? Especially
>> >>> that we re-export `ffi` crate from kernel crate anyway, and `Zeroable` is
>> >>> already in the kernel prelude.
>> >>>
>> >>> We already derive `Zeroable` for bindgen via `MaybeZeroable` derive in
>> >>> rust/bindgen_parameters.
>> >>
>> >> I can't find the convo now, but this change is on my list from when we
>> >> discussed also implementing FromBytes / IntoBytes for the bindings
>> >> types. To do that, we need to move our FromBytes / IntoBytes traits
>> >> somewhere that bindings/uapi can access, and we agreed that the ffi
>> >> crate was a good place for it.
>> >>
>> >> And then for consistency, also reexport Zeroable from the same location.
>> >
>> > I think you also mentioned at some point that using `pin_init` from
>> > `bindings` seemed strange and also using the `pin_init::zeroed()`
>> > function also doesn't fit, since it doesn't have to do with pinned
>> > initialization.
>>
>> Shouldn't it be that a crate that implements Zeroable / FromBytes / IntoBytes
>> and then pin_init becoming an user of that crate, then?

I think that's a little bit too much effort just for "nice naming".

> The Zeroable trait has to be in pin-init because it's also outside the
> kernel. You *could* add yet another crate just for this and let
> pin-init depend on it, but just putting it in the existing ffi seems
> reasonable to me, and ffi is not a bad name for the owner of those
> traits anyway.
>
> Though I guess if we add zerocopy, that concern goes away.

Indeed, I would just move pin-init to that.

Cheers,
Benno