Re: [PATCH v2 1/2] rust: introduce abstractions for fwctl

[Zhi Wang <zhiw@nvidia.com>]

On Mon, 26 Jan 2026 14:19:12 -0400
Jason Gunthorpe <jgg@nvidia.com> wrote:

> On Thu, Jan 22, 2026 at 10:42:30PM +0200, Zhi Wang wrote:
> > --- a/drivers/fwctl/Kconfig
> > +++ b/drivers/fwctl/Kconfig
> > @@ -8,6 +8,18 @@ menuconfig FWCTL
> >  	  manipulating device FLASH, debugging, and other activities
> > that don't fit neatly into an existing subsystem.
> >  

snip

> > +/// Represents a fwctl device type.
> > +///
> > +/// This enum corresponds to the C `enum fwctl_device_type` and is
> > used to identify +/// the specific firmware control interface
> > implemented by a device. +#[repr(u32)]
> > +#[derive(Copy, Clone, Debug, Eq, PartialEq)]
> > +pub enum DeviceType {
> > +    /// Error/invalid device type.
> > +    Error = bindings::fwctl_device_type_FWCTL_DEVICE_TYPE_ERROR,
> > +    /// MLX5 device type.
> > +    Mlx5 = bindings::fwctl_device_type_FWCTL_DEVICE_TYPE_MLX5,
> > +    /// CXL device type.
> > +    Cxl = bindings::fwctl_device_type_FWCTL_DEVICE_TYPE_CXL,
> > +    /// PDS device type.
> > +    Pds = bindings::fwctl_device_type_FWCTL_DEVICE_TYPE_PDS,
> > +    /// Rust fwctl test device type.
> > +    RustFwctlTest =
> > bindings::fwctl_device_type_FWCTL_DEVICE_TYPE_RUST_FWCTL_TEST, +}
> 
> Do we really need these contentless comments?
> 

I will try to remove them if the compiler allows me to do that in the next
re-spin.

> > +impl Device {
> > +    /// # Safety
> > +    ///
> > +    /// `ptr` must be a valid pointer to a `struct fwctl_device`.
> > +    unsafe fn from_raw<'a>(ptr: *mut bindings::fwctl_device) -> &'a
> > Self {
> > +        // CAST: `Self` is a transparent wrapper around
> > `bindings::fwctl_device`.
> > +        // SAFETY: By the safety requirement, `ptr` is valid.
> > +        unsafe { &*ptr.cast() }
> > +    }

snip

> > +            // ensuring it remains valid during allocation.
> > +            let dev = unsafe {
> > +                bindings::_fwctl_alloc_device(
> > +                    parent.as_raw(),
> > +                    ops,
> > +                    core::mem::size_of::<bindings::fwctl_device>(),
> > +                )
> > +            };
> > +
> > +            if dev.is_null() {
> > +                return Err(ENOMEM);
> > +            }
> > +
> > +            // SAFETY: dev is guaranteed to be a valid pointer from
> > `_fwctl_alloc_device()`.
> > +            let ret = unsafe { bindings::fwctl_register(dev) };
> > +            if ret != 0 {
> > +                // SAFETY: dev is guaranteed to be a valid pointer
> > from `_fwctl_alloc_device()`.
> > +                unsafe {
> > +                    bindings::fwctl_put(dev);
> > +                }
> > +                return Err(Error::from_errno(ret));
> > +            }
> 
> This looks weirdly sequenced, the driver's object has to be fully
> initialized before you can call register, so it is quite strange to
> see a wrapper that does both alloc and register in one function.
> 

The fwctl_alloc_device() helper allocates a raw struct fwctl_device
without private driver data here. The Rust driver object should be
already allocated and initialized separately before reaching this
point.

We rely on the standard dev->parent chain to access the rust driver
object from the fwctl callbacks.

> > +                bindings::_fwctl_alloc_device(
> > +                    parent.as_raw(),
> > +                    ops,
> > +                    core::mem::size_of::<bindings::fwctl_device>(),
> > +// SAFETY: `Registration` can be sent to other threads because:
> > +// - It only contains a `NonNull<fwctl_device>` pointer and a
> > PhantomData marker +// - The underlying C fwctl_device is thread-safe
> > with internal locking +// - `Drop` calls
> > `fwctl_unregister()/fwctl_put()` which are safe from any sleepable
> > context
> 
> fwctl_unregister is not safe from any context, it must be called
> while the Device is still bound.
> 

The registration is wrapped in Devres<> in the sample driver, which
guarantees that drop is called while the Device is still bound.

I agree that the current abstraction itself does not strictly enforce this
(e.g., if the object is moved out of Devres). I will investigate an
approach to enforce this requirement in the next re-spin.

Z.

> Jason
>