> On 23 Jul 2025, at 11:26, Boqun Feng <boqun.feng@xxxxxxxxx> wrote:
>
> On Wed, Jul 23, 2025 at 10:55:20AM -0300, Daniel Almeida wrote:
>> Hi Boqun,
>>
>> [...]
>>
>>>> + IrqRequest { dev, irq }
>>>> + }
>>>> +
>>>> + /// Returns the IRQ number of an [`IrqRequest`].
>>>> + pub fn irq(&self) -> u32 {
>>>> + self.irq
>>>> + }
>>>> +}
>>>> +
>>>> +/// A registration of an IRQ handler for a given IRQ line.
>>>> +///
>>>> +/// # Examples
>>>> +///
>>>> +/// The following is an example of using `Registration`. It uses a
>>>> +/// [`AtomicU32`](core::sync::AtomicU32) to provide the interior mutability.
>>>
>>> We are going to remove all usage of core::sync::Atomic* when the LKMM
>>> atomics [1] land. You can probably use `Completion` here (handler does
>>> complete_all(), and registration uses wait_for_completion()) because
>>> `Completion` is irq-safe. And this brings my next comment..
>>
>> How are completions equivalent to atomics? I am trying to highlight interior
>> mutability in this example.
>>
>
> Well, `Completion` also has interior mutability.
>
>> Is the LKMM atomic series getting merged during the upcoming merge window? Because my
>> understanding was that the IRQ series was ready to go in 6.17, pending a reply
>
> Nope, it's likely to be in 6.18.
>
>> from Thomas and some minor comments that have been mentioned in v7.
>>
>> If the LKMM series is not ready yet, my proposal is to leave the
>> Atomics->Completion change for a future patch (or really, to just use the new
>> Atomic types introduced by your series, because again, I don't think Completion
>> is the right thing to have there).
>>
>
> Why? I can find a few examples that an irq handler does a
> complete_all(), e.g. gpi_process_ch_ctrl_irq() in
> drivers/dma/qcom/gpi.c. I think it's very normal for a driver thread to
> use completions to wait for an irq to happen.
>
> But sure, this and the handler pinned initializer thing is not a blocker
> issue. However, I would like to see them resolved as soon as possible
> once merged.
>
> Regards,
> Boqun
>
>>
>> - Daniel
Because it is not as explicit. The main thing we should be conveying to users
here is how to get a &mut or otherwise mutate the data when running the
handler. When people see AtomicU32, it's a quick jump to "I can make this work
by using other locks, like SpinLockIrq". Completions hide this, IMHO.
It's totally possible for someone to see this and say "ok, I can call
complete() on this, but how can I mutate the data in some random T struct?",
even though these are essentially the same thing from an interior mutability
point of view.
-- Daniel
