Re: [PATCH v3 1/2] rust: irq: add support for request_irq()

[Danilo Krummrich <dakr@kernel.org>]

On Tue, Jun 03, 2025 at 11:57:22AM +0200, Alice Ryhl wrote:
> On Tue, Jun 3, 2025 at 11:43 AM Danilo Krummrich <dakr@kernel.org> wrote:
> >
> > On Tue, Jun 03, 2025 at 11:18:40AM +0200, Alice Ryhl wrote:
> > > I don't think that helps. If Devres::drop gets to swap is_available
> > > before the devm callback performs the swap, then the devm callback is
> > > just a no-op and the device still doesn't wait for free_irq() to
> > > finish running.
> >
> > True, this will indeed always be racy. The rule from the C API has always been
> > that devm_{remove,release}_action() must not be called if a concurrent unbind
> > can't be ruled out. Consequently, the same is true for Revocable::revoke() in
> > this case.
> >
> > I think Devres::drop() shouldn't do anything then and instead we should provide
> > Devres::release() and Devres::remove(), which require the &Device<Bound>
> > reference the Devres object was created with, in order to prove that there
> > can't be a concurrent unbind, just like Devres::access().
> 
> What I suggested with the mutex would work if you remove the devm
> callback *after* calling free_irq.
> 
>     // drop Registration
>     mutex_lock();
>     free_irq();
>     mutex_unlock();
>     devm_remove_callback();

I think it would need to be

	if (!devm_remove_callback()) {
		mutex_lock();
		free_irq();
		mutex_unlock();
	}

>     // devm callback
>     mutex_lock();
>     free_irq();
>     mutex_unlock();

Yes, we could solve this with a lock as well, but it would be an additional
lock, just to maintain the current drop() semantics, which I don't see much
value in.

The common case is that the object wrapped in a Devres is meant to live for the
entire duration the device is bound to the driver.

> Another simpler option is to just not support unregistering the irq
> callback except through devm. Then you don't have a registration at
> all. Creating the callback can take an irq number and a ForeignOwnable
> to put in the void pointer. The devm callback calls free_irq and drops
> the ForeignOwnable.

That's basically what Devres::new_foreign_owned() already does.