Re: [PATCH RFC 01/18] rust: drm: ioctl: Add DRM ioctl abstraction

From: Asahi Lina
Date: Fri Mar 10 2023 - 01:21:38 EST


On 10/03/2023 05.39, Karol Herbst wrote:
> On Thu, Mar 9, 2023 at 9:24 PM Faith Ekstrand
> <faith.ekstrand@xxxxxxxxxxxxx> wrote:
>>
>> On Thu, 2023-03-09 at 15:04 +0900, Asahi Lina wrote:
>>> On 08/03/2023 02.34, Björn Roy Baron wrote:
>>>>> + // SAFETY: This is just the ioctl
>>>>> argument, which hopefully has the right type
>>>>> + // (we've done our best checking the
>>>>> size).
>>>>
>>>> In the rust tree there is the ReadableFromBytes [1] trait which
>>>> indicates that it is safe to read arbitrary bytes into the type.
>>>> Maybe you could add it as bound on the argument type when it lands
>>>> in rust-next? This way you can't end up with for example a struct
>>>> containing a bool with the byte value 2, which is UB.
>>>
>>> There's actually a much bigger story here, because that trait isn't
>>> really very useful without a way to auto-derive it. I need the same
>>> kind
>>> of guarantee for all the GPU firmware structs...
>>>
>>> There's one using only declarative macros [1] and one using proc
>>> macros
>>> [2]. And then, since ioctl arguments are declared in C UAPI header
>>> files, we need a way to be able to derive those traits for them...
>>> which
>>> I guess means bindgen changes?
>>
>> It'd be cool to be able to auto-verify that uAPI structs are all
>> tightly packed and use the right subset of types. Maybe not possible
>> this iteration but it'd be cool to see in future. I'd like to see it
>> for C as well, ideally.
>>
>> ~Faith
>>
>
> I'm sure that with a macro you could verify that a struct definition
> doesn't contain any gaps, just not sure on how one would enforce that.
> Could add a trait which can only be implemented through a proc_macro?
> Maybe we can have a proc_macro ensuring no gaps? Would be cool tech to
> have indeed.

You just make the trait unsafe, as usual, then implement it via that
macro. It's how the things I linked work ^^

The tricky thing with C UAPI definitions is just that we need to get
bindgen to emit those macro instantiations around struct definitions
somehow. Or maybe it could be done with a brute force text-based
postprocessing pass? If we put all UAPI defs into their own crate, you
could probably just do it with sed or a python script or something on
the bindgen output to add it for all struct types...

@Rust folks: Should I try creating a uapi crate for this? I think we can
just mirror the bindings crate logic, and we don't need helpers or
anything like that here, so it shouldn't be very difficult. Then I could
(eventually) eliminate all usage of the full bindings crate in the
driver, and also try experimenting with stuff like this to validate all
UAPI types and implement special traits for them...

~~ Lina