Re: [RFC PATCH v5 4/6] rust: DmaFence: [RFC] Add Signaler

[sashiko-bot@kernel.org]

Thank you for your contribution! Sashiko AI review found 2 potential issue(s) to consider:
- [High] dma_fence: counter leak in FenceSignaler::signal causes panic on drop
- [Low] rust: missing inline annotations on small forwarding functions in abstraction crate
--

commit 31b644dae0c3cdaaa05c7d23cd49d1e5209e51c7
Author: Philipp Stanner <phasta@kernel.org>

rust: DmaFence: [RFC] Add Signaler

This commit introduces a FenceSignaler type that allows signaling an
associated DriverFence independently of the locked data structure it
resides in.

> diff --git a/rust/kernel/dma_buf/dma_fence.rs b/rust/kernel/dma_buf/dma_fence.rs
> index b67121c7abd41..4e9b4a6a9471b 100644
> --- a/rust/kernel/dma_buf/dma_fence.rs
> +++ b/rust/kernel/dma_buf/dma_fence.rs
> @@ -590,6 +590,45 @@ unsafe fn dec_ref(ptr: NonNull<Self>) {
>      }
>  }
>  
> +/// An object that can signal an associated [`DriverFence`], independently of
> +/// that one's lifetime.
> +///
> +/// Use with great caucion! Only the producer of a fence should ever be able to
> +/// signal it.
> +///
> +/// This helper is necessary to signal DriverFences that are hidden behind locks
> +/// which cannot be taken in certain situations.
> +pub struct FenceSignaler {
> +    fence: ARef<Fence>,
> +}
> +
> +impl FenceSignaler {
> +    /// Signal the associated [`DriverFence`].
> +    pub fn signal(&self, res: Result) {
> +        let fence = self.fence.lock();
> +
> +        // SAFETY: `fence` is valid because `self` is valid. The lock must be
> +        // held, which we acquired directly above.
> +        if !unsafe { bindings::dma_fence_test_signaled_flag(fence.as_raw()) } {
> +            if let Err(err) = res {
> +                // SAFETY: `fence` is valid because `self` is valid. The fence
> +                // must not have been signaled yet, which we check directly above.
> +                unsafe { bindings::dma_fence_set_error(fence.as_raw(), err.to_errno()) };
> +            }
> +            // SAFETY: `fence` is valid because `self` is valid. The lock must
> +            // be held, which we acquired above.
> +            unsafe { bindings::dma_fence_signal_locked(fence.as_raw()) };
> +        }
> +
> +        // TODO: This should decrement the nr-of-unsignaled-fences tracker in fctx
> +    }

[Severity: High]
Does skipping the counter decrement here cause a resource leak leading to a
kernel panic?

When FenceSignaler::signal() signals the fence without decrementing
nr_of_unsignaled_fences, it creates an issue during teardown. If the
associated DriverFence is subsequently dropped, DriverFence::drop() will
observe that the fence is already signaled and skip its own decrement:

DriverFence::drop() {
    ...
    let signaled: bool = unsafe { bindings::dma_fence_test_signaled_flag(guard.as_raw()) };
    if !signaled {
    ...
}

As a result, the unsignaled counter permanently leaks 1 count per use of
FenceSignaler. When the FenceContext is eventually dropped, its non-zero
counter triggers panic!("Forgotten fences in FenceContext.").

Can this TODO be resolved before merging to prevent drivers from triggering a
denial of service on normal cleanup paths?

> +
> +    /// Get the associated fence's sequence number.
> +    pub fn seqno(&self) -> u64 {
> +        self.fence.seqno()
> +    }

[Severity: Low]
Are small forwarding functions like this missing the #[inline] attribute?

According to the Rust subsystem coding guidelines, small functions or functions
forwarding to a C binding call in abstractions should be annotated with
#[inline] to avoid minor performance overhead across crate boundaries.

[ ... ]

> @@ -805,6 +844,16 @@ pub fn as_fence(&self) -> &Fence {
>          unsafe { Fence::from_raw(self.as_raw()) }
>      }
>  
> +    /// Create a [`FenceSignaler`] associated with this `DriverFence`.
> +    ///
> +    /// Use with caution! Only the producer of a [`DriverFence`] should be able
> +    /// to signal it.
> +    pub fn new_signaler(&self) -> FenceSignaler {
> +        FenceSignaler {
> +            fence: ARef::from(self.as_fence()),
> +        }
> +    }

[Severity: Low]
Should this small constructor also have an #[inline] annotation?

Additionally, FenceContext::deref() is a trivial deref implementation in
this file missing the #[inline] attribute. Could these annotations be added?

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20261009191124.1022902-2-phasta@kernel.org?part=4