Re: [PATCH RFC 18/18] drm/asahi: Add the Asahi driver for Apple AGX GPUs

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 05/04/2023 23.44, Daniel Vetter wrote:
On Tue, Mar 07, 2023 at 11:25:43PM +0900, Asahi Lina wrote:
+/// Look up a GEM object handle for a `File` and return an `ObjectRef` for it.
+pub(crate) fn lookup_handle(file: &DrmFile, handle: u32) -> Result<ObjectRef> {
+    Ok(ObjectRef::new(shmem::Object::lookup_handle(file, handle)?))
+}

So maybe my expectations for rust typing is a bit too much, but I kinda
expected this to be fully generic:

- trait Driver (drm_driver) knows the driver's object type
- a generic create_handle function could ensure that for drm_file (which
   is always for a specific drm_device and hence Driver) can ensure at the
   type level that you only put the right objects into the drm_file
- a generic lookup_handle function on the drm_file knows the Driver trait
   and so can give you back the right type right away.

Why the wrapping, what do I miss?

Sigh, so this is one of the many ways I'm trying to work around the "Rust doesn't do subclasses" problem (so we can figure out what the best one is ^^).

The generic shmem::Object::lookup_handle() call *is* fully generic and will get you back a driver-specific object. But since Rust doesn't do subclassing, what you get back isn't a driver-specific type T, but rather a (reference to a) shmem::Object<T>. T represents the inner driver-specific data/functionality (only), and the outer shmem::Object<T> includes the actual drm_gem_shmem_object plus a T. This is backwards from C, where you expect the opposite situation where T contains a shmem object, but that just doesn't work with Rust because there's no way to build a safe API around that model as far as I know.

Now the problem is from the higher layers I want object operations that interact with the shmem::Object<T> (that is, they call generic GEM functions on the object). Options so far:

1. Add an outer wrapper and put that functionality there.
2. Just have the functions on T as helpers, so you need to call T::foo(obj) instead of obj.foo(). 3. Use the undocumented method receiver trait thing to make shmem::Object<T> a valid `self` type, plus add auto-Deref to shmem::Object. Then obj.foo() works.

#1 is what I use here. #2 is how the driver-specific File ioctl callbacks are implemented, and also sched::Job<T>. #3 is used for fence callbacks (FenceObject<T>). None of them are great, and I'd love to hear what people think of the various options...

There are other unexplored options, like in this GEM case it could be covered with a driver-internal auxiliary trait impl'd on shmem::Object<T> buuut that doesn't work when you actually need callbacks on T itself to circle back to shmem::Object<T>, as is the case with File/Job/FenceObject.

~~ Lina




[Index of Archives]     [Linux DRI Users]     [Linux Intel Graphics]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]     [XFree86]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Linux Kernel]     [Linux SCSI]     [XFree86]
  Powered by Linux