Message-Id: <DBB18YX7MBDW.3C5Q5Y1O28NFL@kernel.org>
Date: Sun, 13 Jul 2025 17:29:36 +0200
From: "Benno Lossin" <lossin@...nel.org>
To: "Danilo Krummrich" <dakr@...nel.org>
Cc: "Daniel Almeida" <daniel.almeida@...labora.com>, "Miguel Ojeda"
 <ojeda@...nel.org>, "Alex Gaynor" <alex.gaynor@...il.com>, "Boqun Feng"
 <boqun.feng@...il.com>, "Gary Guo" <gary@...yguo.net>,
 Björn Roy Baron <bjorn3_gh@...tonmail.com>, "Andreas
 Hindborg" <a.hindborg@...nel.org>, "Alice Ryhl" <aliceryhl@...gle.com>,
 "Trevor Gross" <tmgross@...ch.edu>, "Greg Kroah-Hartman"
 <gregkh@...uxfoundation.org>, "Rafael J. Wysocki" <rafael@...nel.org>,
 "Thomas Gleixner" <tglx@...utronix.de>, "Bjorn Helgaas"
 <bhelgaas@...gle.com>, Krzysztof Wilczyński
 <kwilczynski@...nel.org>, <linux-kernel@...r.kernel.org>,
 <rust-for-linux@...r.kernel.org>, <linux-pci@...r.kernel.org>
Subject: Re: [PATCH v6 3/6] rust: irq: add support for non-threaded IRQs and
 handlers

On Sun Jul 13, 2025 at 2:42 PM CEST, Danilo Krummrich wrote:
> On Sun Jul 13, 2025 at 2:16 PM CEST, Benno Lossin wrote:
>> On Sun Jul 13, 2025 at 1:57 PM CEST, Danilo Krummrich wrote:
>>> 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@...nel.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>).
>>
>> You can just use `container_of!`?
>
> Sure, that's possible too.
>
>>>> 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.
>>
>> How does it ensure that?
>
> RegistrationInner calls free_irq() in it's drop() method; Devres revokes it on
> device unbind.

Makes sense, so we probably do need the unsafe typestate change
function in this case.

>>>
>>>   (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.
>>
>> Yeah but I would find it much more natural (and also useful in other
>> circumstances) if `Devres<T>` would give you access to `Device` (at
>> least the `Normal` type state).
>
> If we use container_of!() instead or just pass the address of Self (i.e.
> Registration) to request_irq() instead,
>
> 	pub fn device(&self) -> &Device
>
> is absolutely possible to add to Devres, of course.
>
>> Depending on how (1) is ensured, we might just need an unsafe function
>> that turns `Device<Normal>` into `Device<Bound>`.
>
> `&Device<Normal>` in `&Device<Bound>`, yes. I have such a method locally
> already (but haven't sent it yet), because that's going to be a use-case for
> other abstractions as well. One specific example is the PWM Chip abstraction
> [1].
>
> [1] https://lore.kernel.org/lkml/20250710-rust-next-pwm-working-fan-for-sending-v11-3-93824a16f9ec@samsung.com/

I think this solution sounds much better than storing an additional
`NonNull<Device<Bound>>`.

---
Cheers,
Benno

Powered by blists - more mailing lists