Re: [PATCH 1/3] rust: drm: gem: shmem: implement gem_create_object for prime imports
From: Sami Tolvanen
Date: Wed Nov 18 2026 - 13:58:42 EST
On Wed, Nov 18, 2026 at 3:41 AM Alice Ryhl <aliceryhl@xxxxxxxxxx> wrote:
>
> On Mon, Oct 26, 2026 at 12:14:08PM +0100, Janne Grunau wrote:
> > From: Sami Tolvanen <samitolvanen@xxxxxxxxxx>
> >
> > For a prime-imported BO with no `gem_create_object` callback,
> > `__drm_gem_shmem_create()` allocates `kzalloc(sizeof(struct
> > drm_gem_shmem_object))`, which is smaller than the Rust `Object<T>`
> > wrapper. A later `container_of!()` from a `drm_gem_object` back to
> > `Object<T>` then runs past the allocation and corrupts the heap.
> >
> > Provide a `gem_create_object` callback that allocates the full
> > `Object<T>` via `KBox::pin_init`. Move `obj` to the first field so the
> > returned pointer is at offset zero, satisfying the `kfree()` on the
> > init-failure path. This is now compile-checked with `const_assert!`.
> >
> > Imported BOs have no userspace creation arguments, so extend
> > `DriverObject` with a `create_imported(dev, size)` constructor and
> > implement it for Tyr's `BoData`, Nova, and the in-tree KUnit driver.
> ...
> > + /// Create a new driver data object for a GEM object that is being imported
> > + /// from another driver (e.g. via dma-buf).
> > + ///
> > + /// Imported objects are allocated by the importing helper, which only
> > + /// provides the device and size. There is no driver-supplied
> > + /// [`DriverObject::Args`], so drivers must initialize the per-object
> > + /// state without those arguments.
> > + fn create_imported(dev: &drm::Device<Self::Driver>, size: usize) -> impl PinInit<Self, Error>;
> ...
> > + // SAFETY: We never move out of `self`. Ownership of the allocation is transferred to the
> > + // C side, which will run `__drm_gem_shmem_init()` next and later free us via
> > + // `free_callback` (or `kfree()` on the `__drm_gem_shmem_init()` error path, matching the
> > + // C kzalloc fallback contract).
> > + let new = KBox::into_raw(unsafe { Pin::into_inner_unchecked(new) });
> > +
> > + // SAFETY: `new` was just produced by `KBox::into_raw` and so is a valid pointer to an
> > + // initialized `Self`. `obj.base` is the embedded `drm_gem_object` the caller expects.
> > + unsafe { &raw mut (*(*new).obj.get()).base }
> > + }
>
> I ran this through a review agent, and it found these two things. As far as I
> can tell, these are valid concerns.
>
> 1. On the `__drm_gem_shmem_init()` failure path, `__drm_gem_shmem_create()`
> calls `kfree(obj)` directly rather than reconstructing `KBox<Self>`.
> Having `obj` at offset 0 ensures `kfree(obj)` frees the start of the slab
> allocation, but it skips running `Drop` on `Object<T>` and `inner: T`. If a
> driver's `T` owns any fields with non-trivial `Drop` implementations
> initialized in `T::create_imported()`, they will be leaked if
> `__drm_gem_shmem_init()` fails (e.g. if `drm_gem_create_mmap_offset()`
> fails).
I agree, this looks like a valid issue.
> 2. `Object<T>::ALLOC_OPS` also sets:
>
> dumb_create: Some(bindings::drm_gem_shmem_dumb_create),
>
> `drm_gem_shmem_dumb_create()` calls `drm_gem_shmem_create()` ->
> `__drm_gem_shmem_create(dev, size, false)` -> `gem_create_object_callback()`,
> and `DRM_IOCTL_MODE_CREATE_DUMB` only checks `dev->driver->dumb_create`
> (not `DRIVER_MODESET`). So `T::create_imported()` can also be called to
> create a non-imported dumb buffer (`obj.is_imported() == false`).
>
> Instead of implementing `gem_create_object`, what do you think about
> exporting a `private = true` counterpart to `drm_gem_shmem_init()` (similar to
> commit e3f4bdaf2c5b ("drm/gem/shmem: Extract drm_gem_shmem_init() from
> drm_gem_shmem_create()")), implementing a Rust `gem_prime_import_sg_table`
> callback on `shmem::Object<T>` directly, and setting `dumb_create: None` in
> `Object<T>::ALLOC_OPS`? That way `new: Pin<KBox<Self>>` stays owned by Rust
> across `__drm_gem_shmem_init()` just like in `Object::new()`, so `Drop` runs
> normally on error and `obj` does not need to sit at offset 0 for `kfree(obj)`.
A later commit in tyr-dev lets drivers opt out of dumb_create, which
addresses this part. However, your proposal sounds cleaner and fixes
both issues, so let's go with that.
Sami