Re: [PATCH v6 3/6] rust: irq: add support for non-threaded IRQs and handlers

["Danilo Krummrich" <dakr@kernel.org>]

On Sun Jul 13, 2025 at 1:19 PM CEST, Benno Lossin wrote:
> On Sun Jul 13, 2025 at 12:24 PM CEST, Danilo Krummrich wrote:
>> On Sun Jul 13, 2025 at 1:32 AM CEST, Daniel Almeida wrote:
>>>
>>>
>>>> On 12 Jul 2025, at 18:24, Danilo Krummrich <dakr@kernel.org> wrote:
>>>> 
>>>> On Thu Jul 3, 2025 at 9:30 PM CEST, Daniel Almeida wrote:
>>>>> +/// Callbacks for an IRQ handler.
>>>>> +pub trait Handler: Sync {
>>>>> +    /// The hard IRQ handler.
>>>>> +    ///
>>>>> +    /// This is executed in interrupt context, hence all corresponding
>>>>> +    /// limitations do apply.
>>>>> +    ///
>>>>> +    /// All work that does not necessarily need to be executed from
>>>>> +    /// interrupt context, should be deferred to a threaded handler.
>>>>> +    /// See also [`ThreadedRegistration`].
>>>>> +    fn handle(&self) -> IrqReturn;
>>>>> +}
>>>> 
>>>> One thing I forgot, the IRQ handlers should have a &Device<Bound> argument,
>>>> i.e.:
>>>> 
>>>> fn handle(&self, dev: &Device<Bound>) -> IrqReturn
>>>> 
>>>> IRQ registrations naturally give us this guarantee, so we should take advantage
>>>> of that.
>>>> 
>>>> - Danilo
>>>
>>> Hi Danilo,
>>>
>>> I do not immediately see a way to get a Device<Bound> from here:
>>>
>>> unsafe extern "C" fn handle_irq_callback<T: Handler>(_irq: i32, ptr: *mut c_void) -> c_uint {
>>>
>>> Refall that we've established `ptr` to be the address of the handler. This
>>> came after some back and forth and after the extensive discussion that Benno
>>> and Boqun had w.r.t to pinning in request_irq().
>>
>> You can just wrap the Handler in a new type and store the pointer there:
>>
>> 	#[pin_data]
>> 	struct Wrapper {
>> 	   #[pin]
>> 	   handler: T,
>> 	   dev: NonNull<Device<Bound>>,
>> 	}
>>
>> And then pass a pointer to the Wrapper field to request_irq();
>> handle_irq_callback() can construct a &T and a &Device<Bound> from this.
>>
>> Note that storing a device pointer, without its own reference count, is
>> perfectly fine, since inner (Devres<RegistrationInner>) already holds a
>> reference to the device and guarantees the bound scope for the handler
>> callbacks.
>
> Can't we just add an accessor function to `Devres`?

	#[pin_data]
	pub struct Registration<T: Handler + 'static> {
	    #[pin]
	    inner: Devres<RegistrationInner>,
	
	    #[pin]
	    handler: T,
	
	    /// Pinned because we need address stability so that we can pass a pointer
	    /// to the callback.
	    #[pin]
	    _pin: PhantomPinned,
	}

Currently we pass the address of handler to request_irq(), so this doesn't help,
hence my proposal to replace the above T with Wrapper (actually Wrapper<T>).

> Also `Devres` only stores `Device<Normal>`, not `Device<Bound>`...

The Devres instance itself may out-live device unbind, but it ensures that the
encapsulated data does not, hence it holds a reference count, i.e. ARef<Device>.

Device<Bound> or ARef<Device<Bound>> *never* exists, only &'a Device<Bound>
within a corresponding scope for which we can guarantee the device is bound.

In the proposed wrapper we can store a NonNull<Device<Bound>> though, because we
can safely give out a &Device<Bound> in the IRQ's handle() callback. This is
because:

  (1) RegistrationInner is guarded by Devres and guarantees that free_irq() is
      completed *before* the device is unbound.

  (2) It is guaranteed that the device pointer is valid because (1) guarantees
      it's even bound and because Devres<RegistrationInner> itself has a
      reference count.