Re: [PATCH 4/8] rust: dma: introduce dma::CoherentInit for memory initialization

["Alexandre Courbot" <acourbot@nvidia.com>]

On Wed Mar 4, 2026 at 1:22 AM JST, Danilo Krummrich wrote:
<snip>
> @@ -352,6 +358,151 @@ fn from(direction: DataDirection) -> Self {
>      }
>  }
>  
> +/// Initializer type for [`Coherent`].
> +///
> +/// A [`Coherent`] object can't provide access to its memory as (mutable) slice safely, since it
> +/// can't fulfill the requirements for creating a slice. For instance, it is not valid to have a
> +/// (mutable) slice to of the memory of a [`Coherent`] while the memory might be accessed by a
> +/// device.
> +///
> +/// In contrast, this initializer type is able to fulfill the requirements to safely obtain a
> +/// (mutable) slice, as it neither provides access to the DMA address of the embedded [`Coherent`],
> +/// nor can it be used with the DMA projection accessors.
> +///
> +/// Once initialized, this type can be converted to a regular [`Coherent`] object.
> +///
> +/// # Examples
> +///
> +/// `CoherentInit<T>`:
> +///
> +/// ```
> +/// # use kernel::device::{
> +/// #     Bound,
> +/// #     Device,
> +/// # };
> +/// use kernel::dma::{attrs::*,
> +///     Coherent,
> +///     CoherentInit,
> +/// };
> +///
> +/// # fn test(dev: &Device<Bound>) -> Result {
> +/// let mut dmem: CoherentInit<u64> =
> +///     CoherentInit::zeroed_with_attrs(dev, GFP_KERNEL, DMA_ATTR_NO_WARN)?;

Since this is an example, shall we use the simpler
`CoherentInit::zeroed`?

> +/// *dmem = 42;
> +/// let dmem: Coherent<u64> = dmem.into();
> +/// # Ok::<(), Error>(()) }
> +/// ```
> +///
> +/// `CoherentInit<[T]>`:
> +///
> +///
> +/// ```
> +/// # use kernel::device::{
> +/// #     Bound,
> +/// #     Device,
> +/// # };
> +/// use kernel::dma::{attrs::*,
> +///     Coherent,
> +///     CoherentInit,
> +/// };
> +///
> +/// # fn test(dev: &Device<Bound>) -> Result {
> +/// let mut dmem: CoherentInit<[u64]> =
> +///     CoherentInit::zeroed_slice_with_attrs(dev, 4, GFP_KERNEL, DMA_ATTR_NO_WARN)?;

Same here.

> +/// dmem.fill(42);
> +/// let dmem: Coherent<[u64]> = dmem.into();
> +/// # Ok::<(), Error>(()) }
> +/// ```
> +pub struct CoherentInit<T: AsBytes + FromBytes + KnownSize + ?Sized>(Coherent<T>);
> +
> +impl<T: AsBytes + FromBytes> CoherentInit<[T]> {
> +    /// Initializer variant of [`Coherent::zeroed_slice_with_attrs`].
> +    pub fn zeroed_slice_with_attrs(
> +        dev: &device::Device<Bound>,
> +        count: usize,
> +        gfp_flags: kernel::alloc::Flags,
> +        dma_attrs: Attrs,
> +    ) -> Result<Self> {
> +        Coherent::zeroed_slice_with_attrs(dev, count, gfp_flags, dma_attrs).map(Self)
> +    }
> +
> +    /// Same as [CoherentInit::zeroed_slice_with_attrs], but with `dma::Attrs(0)`.
> +    pub fn zeroed_slice(
> +        dev: &device::Device<Bound>,
> +        count: usize,
> +        gfp_flags: kernel::alloc::Flags,
> +    ) -> Result<Self> {
> +        Self::zeroed_slice_with_attrs(dev, count, gfp_flags, Attrs(0))
> +    }
> +
> +    /// Initializes the element at `i` using the given initializer.
> +    ///
> +    /// Returns `EINVAL` if `i` is out of bounds.
> +    pub fn init_at<E>(&mut self, i: usize, init: impl Init<T, E>) -> Result

Should this method be introduced in the next patch, or even in its own
patch? It feels a bit out of place at this stage since the non-array
`CoherentInit` doesn't have an equivalent.

I was also wondering whether we could have an `init` method that
initializes all the elements without having to zero the whole array
first, but I guess it might be a bit difficult to implement in a
flexible enough way.