Re: [PATCH 2/2] drm/tyr: unplug the GPU when reset fails

[Alice Ryhl <aliceryhl@google.com>]

On Sun, Aug 23, 2026 at 01:50:08PM +0300, Onur Özkan wrote:
> @@ -56,7 +57,7 @@ pub(crate) struct TyrPlatformDriverData<'bound> {
>      // ensure the correct drop order.
>      #[pin]
>      _reset: reset::ResetHandle<'bound>,
> -    _reg: drm::Registration<'bound, TyrDrmDriver>,
> +    _reg: Arc<Revocable<drm::Registration<'bound, TyrDrmDriver>>>,
>      _device: ARef<TyrDrmDevice>,
>  }
>  
> @@ -161,17 +162,18 @@ fn probe<'bound>(
>                      gpu_info,
>              });
>  
> -            // SAFETY: `reg` is stored in the platform driver data and is not leaked or
> -            // forgotten, so it is dropped before the `'bound` registration data can become
> -            // invalid.
> +            // SAFETY: `reg` is wrapped in a `Revocable` owned by the platform driver
> +            // data and its reset handle. Neither is leaked or forgotten, so `reg` is
> +            // dropped before the `'bound` registration data can become invalid.
>              let reg =
>                  unsafe { drm::Registration::new_with_lt(pdev.as_ref(), unreg_dev, reg_data, 0)? };
>              let device = reg.device().into();
> +            let reg = Arc::pin_init(Revocable::new(reg), GFP_KERNEL)?;

`drm::Registration::new` requires that the `Registration` is dropped
before `'bound` expires so that `drm_dev_unplug()` runs and
`TyrDrmRegistrationData<'bound>` is dropped while `'bound` is still
valid. Wrapping it in an `Arc` makes this invariant fragile: if any
`Arc` clone outlives `TyrPlatformDriverData` or is leaked,
`drm::Registration` will not be dropped on platform unbind and
`TyrDrmRegistrationData<'bound>` will outlive `'bound`.

With `pin-init`'s new safe self-referential struct support in `rust-next`
(see e.g. [1]), you can avoid both `Arc` and raw pointers altogether by
storing `_reg: Revocable<drm::Registration<'bound, TyrDrmDriver>>`
directly as a `#[pin]` field of `TyrPlatformDriverData<'bound>` and
having `_reset` borrow it via field + existential lifetimes:

    #[pin_data]
    pub(crate) struct TyrPlatformDriverData<'bound> {
        #[pin]
        #[uses('_reg: invariant)]
        _reset: reset::ResetHandle<'_reg>,
        #[pin]
        _reg: Revocable<drm::Registration<'bound, TyrDrmDriver>>,
    }

and in `reset.rs`:

    #[pin_data]
    struct Controller<'ctrl>
    where
        exists<'bound>: 'ctrl,
    {
        pdev: &'ctrl platform::Device<Bound>,
        state: Atomic<ResetState>,
        hw: Arc<HwGate<'ctrl>>,
        #[uses('bound: invariant)]
        registration:
            Pin<&'ctrl Revocable<drm::Registration<'bound, TyrDrmDriver>>>,
    }

Then `probe()` simply initializes `_reg` first and passes `_reg` to
`ResetHandle::new(pdev, hw, _reg)?`, `reset_work()` calls
`self.with_registration(|reg| reg.revoke())`, and `#[pin_data]`
statically checks that `_reset` drops before `_reg`.

Also, as a heads-up for rebasing onto `drm-rust-next`:
`_device: ARef<TyrDrmDevice>` has been removed from
`TyrPlatformDriverData` (along with the empty `PinnedDrop` impl), and
`drm::Registration::new()` now returns `impl PinInit<Self, Error>`
directly without the `flags` argument, so `_reg` can be initialized in
place via `Revocable::new(unsafe { drm::Registration::new(...) })`.

[1] https://lore.kernel.org/r/20261019133647.1279779-1-gary@kernel.org

Alice