Re: [PATCH] rust: dma: add `Range` type

[Robin Murphy <robin.murphy@arm.com>]

On 2026-08-06 8:46 am, Vasileios Almpanis wrote:
> The base dma address of `Coherent` and `CoherentHandle` is a bare
> `DmaAddress` integer so any arithmetic, a driver does on it is unchecked
> and can potentially go past the end of the allocation or overflow.
> 
> Add a `dma::Range` type that couples a base `DmaAddress` with a length
> and only hands out addresses and sub-ranges within `[start, start + len)`
> with all arithmetics checked against both the length and the overflow of
> of the underlying `dma_addr_t`. Let `Coherent` and `CoherentHandle`
> provide the `Range` covering their allocation.
> 
> Suggested-by: Danilo Krummrich <dakr@kernel.org>
> Link: https://github.com/Rust-for-Linux/linux/issues/1248
> Signed-off-by: Vasileios Almpanis <vasilisalmpanis@gmail.com>
> ---
> This is one of my first rust-for-linux patches, so any suggestions are
> extremely welcome. Some words about the decisions taken:
> 
> - I chose lengths offsets to be `DmaAddress` instead of usize since the
>    bus space can be 64-bit on 32-bit CPUS.

The base address may be 64-bit, but the underlying dma_alloc_*/dma_map_* 
APIs all express lengths in size_t, so we shouldn't need to accommodate 
anything larger in Rust either. Sizes-as-address-types always seem a bit 
weird and clunky (lookin' at you, resource_size_t...), so probably 
better avoided if not absolutely necessary, IMO.

Thanks,
Robin.

> - The constructor returns EOVERFLOW for `dma_addr_t` overflow, while
>    out-of-bounds requests return EINVAL; happy to use a single error
>    code if preferred.
> 
> - `dma_range()` is added to both `Coherent` and `CoherentHandle` so the
>    type has users from the start. I could split it into a follow-up if
>    that is preferred.
> 
> Tested with `make rustdoc`, `make rustfmtcheck`, `CLIPPY=1`, and the
> KUnit doctests `rust_doctests_kernel`.
> ---
>   rust/kernel/dma.rs | 174 +++++++++++++++++++++++++++++++++++++++++++++++++++++
>   1 file changed, 174 insertions(+)
> 
> diff --git a/rust/kernel/dma.rs b/rust/kernel/dma.rs
> index 200def84fb69e006bca0b1c578bac9f1dc8da708..e65e99ae4d915ac2e385df8b72fc518c07e4c82d 100644
> --- a/rust/kernel/dma.rs
> +++ b/rust/kernel/dma.rs
> @@ -41,6 +41,142 @@
>   /// Note that this may be `u64` even on 32-bit architectures.
>   pub type DmaAddress = bindings::dma_addr_t;
>   
> +/// A range of DMA addresses.
> +///
> +/// Couples a base [`DmaAddress`] with the length in bytes of the region it belongs to,
> +/// representing the half-open range `[start, start + len)` of DMA addresses.
> +///
> +/// Unlike a bare [`DmaAddress`], a [`Range`] only hands out addresses and sub-ranges that are
> +/// guaranteed to lie within `[start, start + len)`; all arithmetic is checked against both the
> +/// length of the range and overflow of the underlying [`DmaAddress`].
> +///
> +/// # Invariants
> +///
> +/// `start + len` does not overflow [`DmaAddress`].
> +#[derive(Clone, Copy, Debug, PartialEq, Eq)]
> +pub struct Range {
> +    start: DmaAddress,
> +    len: DmaAddress,
> +}
> +
> +impl Range {
> +    /// Creates a new [`Range`] of `len` bytes, starting at `start`.
> +    ///
> +    /// Returns [`EOVERFLOW`] if `start + len` overflows [`DmaAddress`].
> +    ///
> +    /// # Examples
> +    ///
> +    /// ```
> +    /// use kernel::dma::{DmaAddress, Range};
> +    ///
> +    /// let range = Range::new(0x1000, 0x200)?;
> +    /// assert_eq!(range.start(), 0x1000);
> +    /// assert_eq!(range.end(), 0x1200);
> +    /// assert_eq!(range.len(), 0x200);
> +    ///
> +    /// assert!(Range::new(DmaAddress::MAX, 1).is_err());
> +    /// # Ok::<(), Error>(())
> +    /// ```
> +    #[inline]
> +    pub const fn new(start: DmaAddress, len: DmaAddress) -> Result<Self> {
> +        if start.checked_add(len).is_none() {
> +            return Err(EOVERFLOW);
> +        }
> +
> +        // INVARIANT: We just checked that `start + len` does not overflow `DmaAddress`.
> +        Ok(Self { start, len })
> +    }
> +
> +    /// Returns the first address of the range.
> +    #[inline]
> +    pub const fn start(&self) -> DmaAddress {
> +        self.start
> +    }
> +
> +    /// Returns the first address after the end of the range.
> +    #[inline]
> +    pub const fn end(&self) -> DmaAddress {
> +        // By the type invariant, `start + len` does not overflow `DmaAddress`.
> +        self.start + self.len
> +    }
> +
> +    /// Returns the length of the range in bytes.
> +    #[inline]
> +    pub const fn len(&self) -> DmaAddress {
> +        self.len
> +    }
> +
> +    /// Returns `true` if the range is empty.
> +    #[inline]
> +    pub const fn is_empty(&self) -> bool {
> +        self.len == 0
> +    }
> +
> +    /// Returns the address at `offset` bytes into the range.
> +    ///
> +    /// The returned address is guaranteed to lie within the range; returns [`EINVAL`] if `offset`
> +    /// is not smaller than the length of the range.
> +    ///
> +    /// # Examples
> +    ///
> +    /// ```
> +    /// use kernel::dma::Range;
> +    ///
> +    /// let range = Range::new(0x1000, 0x200)?;
> +    ///
> +    /// assert_eq!(range.address(0)?, 0x1000);
> +    /// assert_eq!(range.address(0x1ff)?, 0x11ff);
> +    /// assert!(range.address(0x200).is_err());
> +    /// # Ok::<(), Error>(())
> +    /// ```
> +    #[inline]
> +    pub const fn address(&self, offset: DmaAddress) -> Result<DmaAddress> {
> +        if offset >= self.len {
> +            return Err(EINVAL);
> +        }
> +
> +        // By the type invariant, `start + offset < start + len` does not overflow `DmaAddress`.
> +        Ok(self.start + offset)
> +    }
> +
> +    /// Returns the sub-range of `len` bytes, starting `offset` bytes into the range.
> +    ///
> +    /// The returned range is guaranteed to lie within the range; returns [`EINVAL`] if
> +    /// `offset + len` overflows [`DmaAddress`] or exceeds the length of the range.
> +    ///
> +    /// # Examples
> +    ///
> +    /// ```
> +    /// use kernel::dma::Range;
> +    ///
> +    /// let range = Range::new(0x1000, 0x200)?;
> +    ///
> +    /// let sub = range.subrange(0x100, 0x80)?;
> +    /// assert_eq!(sub.start(), 0x1100);
> +    /// assert_eq!(sub.end(), 0x1180);
> +    ///
> +    /// assert!(range.subrange(0x100, 0x101).is_err());
> +    /// # Ok::<(), Error>(())
> +    /// ```
> +    #[inline]
> +    pub const fn subrange(&self, offset: DmaAddress, len: DmaAddress) -> Result<Self> {
> +        let Some(end) = offset.checked_add(len) else {
> +            return Err(EINVAL);
> +        };
> +
> +        if end > self.len {
> +            return Err(EINVAL);
> +        }
> +
> +        // INVARIANT: `start + offset + len <= start + self.len`, which by the type invariant of
> +        // `self` does not overflow `DmaAddress`.
> +        Ok(Self {
> +            start: self.start + offset,
> +            len,
> +        })
> +    }
> +}
> +
>   /// Trait to be implemented by DMA capable bus devices.
>   ///
>   /// The [`dma::Device`](Device) trait should be implemented by bus specific device representations,
> @@ -626,6 +762,25 @@ pub fn dma_handle(&self) -> DmaAddress {
>           self.dma_handle
>       }
>   
> +    /// Returns the [`Range`] of DMA addresses covering this allocation.
> +    ///
> +    /// Unlike [`Self::dma_handle`], which hands out the base address as a bare integer, the
> +    /// returned [`Range`] couples the base address with the size of the allocation, such that
> +    /// any offset arithmetic performed on it is checked.
> +    #[inline]
> +    pub fn dma_range(&self) -> Range {
> +        // INVARIANT: By the type invariants of `Self`, `dma_handle` is the DMA address base of an
> +        // allocated region of `self.size()` bytes; the DMA API guarantees that a mapped region
> +        // never wraps the DMA address space, hence `dma_handle + size` does not overflow
> +        // `DmaAddress`.
> +        Range {
> +            start: self.dma_handle,
> +            // CAST: `usize` always fits in `DmaAddress`, which is at least 32 bits wide and
> +            // always 64 bits wide on 64-bit architectures.
> +            len: self.size() as DmaAddress,
> +        }
> +    }
> +
>       /// Returns a reference to the data in the region.
>       ///
>       /// # Safety
> @@ -1101,6 +1256,25 @@ pub fn dma_handle(&self) -> DmaAddress {
>           self.dma_handle
>       }
>   
> +    /// Returns the [`Range`] of DMA addresses covering this allocation.
> +    ///
> +    /// Unlike [`Self::dma_handle`], which hands out the base address as a bare integer, the
> +    /// returned [`Range`] couples the base address with the size of the allocation, such that
> +    /// any offset arithmetic performed on it is checked.
> +    #[inline]
> +    pub fn dma_range(&self) -> Range {
> +        // INVARIANT: By the type invariants of `Self`, `dma_handle` is the DMA address base of an
> +        // allocated region of `self.size` bytes; the DMA API guarantees that a mapped region
> +        // never wraps the DMA address space, hence `dma_handle + size` does not overflow
> +        // `DmaAddress`.
> +        Range {
> +            start: self.dma_handle,
> +            // CAST: `usize` always fits in `DmaAddress`, which is at least 32 bits wide and
> +            // always 64 bits wide on 64-bit architectures.
> +            len: self.size as DmaAddress,
> +        }
> +    }
> +
>       /// Returns the size in bytes of this allocation.
>       #[inline]
>       pub fn size(&self) -> usize {
> 
> ---
> base-commit: dc01dfb37b34beeefcfe1c3055364d41a4070c7e
> change-id: 20260805-dma-7c5330baadb6
> 
> Best regards,