On Tue Sep 15, 2026 at 11:57 AM BST, Laura Nao wrote:
> SysMemBackend's IoCapable<u64> impl is currently gated on CONFIG_64BIT,
> copying the MMIO backend's restriction. MMIO needs that gate because
> readq() is not available on 32bit. System memory has no such dependency:
> a u64 volatile load/store compiles on any architecture, it's just not
> single-copy atomic on 32bit.
>
> Drop the gate so u64-backed types, such as bitfields, work on 32bit too.
> Document the non-atomicity at the impl instead of enforcing it at build
> time.
>
> Co-developed-by: Daniel Almeida <[email protected]>
> Signed-off-by: Daniel Almeida <[email protected]>
> Signed-off-by: Laura Nao <[email protected]>
No. If you don't need atomicity, please use `copy_read` to read the value. It'll
compile down to the same volatile read, just without atomicity guarantee.
Best,
Gary
> ---
> rust/kernel/io.rs | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/rust/kernel/io.rs b/rust/kernel/io.rs
> index de8ef8e2aec4..a85ddd07cb7f 100644
> --- a/rust/kernel/io.rs
> +++ b/rust/kernel/io.rs
> @@ -1428,7 +1428,9 @@ fn io_write(view: SysMem<'_, $ty>, value: $ty) {
> impl_sysmem_io_capable!(u8);
> impl_sysmem_io_capable!(u16);
> impl_sysmem_io_capable!(u32);
> -#[cfg(CONFIG_64BIT)]
> +// Unlike MMIO, that needs `readq` which is not available on 32-bit, a
> +// system-memory `u64` access compiles on any target. It is just not
> +// single-copy atomic on 32-bit.
> impl_sysmem_io_capable!(u64);
>
> impl IoCopyable for SysMemBackend {
