Re: [PATCH v8 2/2] rust: add dma coherent allocator abstraction.

[Thomas Hampton]

On 1/8/2025 6:27 AM, Abdiel Janulgue wrote:

> +    /// Returns the base address, dma handle, attributes and the size of the allocated region.
> +    /// The caller takes ownership of the returned resources, i.e., will have the responsibility
> +    /// in calling `bindings::dma_free_attrs`.
> +    pub fn into_parts(self) -> (*mut T, bindings::dma_addr_t, usize, usize) {
> +        let size = self.count * core::mem::size_of::<T>();
> +        let ret = (
> +            self.cpu_addr,
> +            self.dma_handle,
> +            self.dma_attrs.as_raw(),
> +            size,
> +        );
> +        // Drop the device's reference count associated with this object. This is needed as no
> +        // destructor will be called on this object once this function returns.
> +        // SAFETY: the device pointer is still valid as of this point due to the type invariants
> +        // on `CoherentAllocation`.
> +        unsafe { bindings::put_device(self.dev.as_raw()) }
> +        core::mem::forget(self);
> +        ret
> +    }

Would it be feasible and prudent to avoid exposing `pub fn
into_parts`, `pub fn first_ptr`, `pub fn first_ptr_mut`?

It seems like some sort of functional access method might be better
for keeping future Rust code maintainable and Rusty.

For example, making client code from:

> {
>    let ctrl_id = unsafe { &*(id.first_ptr() as *const NvmeIdCtrl) };
>    number_of_namespaces = ctrl_id.nn.into();
>    mdts = ctrl_id.mdts;
> }

Into something like,
pub fn expose(self, op: F)
  where F: Fn(*const T) {
    op(self.cpu_addr);
}

let number_of_namespaces;
let mdts;
id.expose(|data| {
let ctrl_id = &*(data as *const NvmeIdCtrl);
    number_of_namespaces = ctrl_id.nn.into();
    mdts = ctrl_id.mdts;
});

And to avoid code docs like:

> /// The caller <...> will have the responsibility in calling `bindings::dma_free_attrs`.