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

[Alice Ryhl <aliceryhl@google.com>]

On Tue, Jun 03, 2025 at 10:46:28AM +0200, Danilo Krummrich wrote:
> On Tue, Jun 03, 2025 at 08:28:42AM +0000, Alice Ryhl wrote:
> > That optimization sounds like something we definitely want, but I have
> > one question: is free_irq() safe to use in atomic context / inside
> > rcu_read_lock()? What about the threaded-irq variant?
> 
> No, free_irq() must not be called from atomic context. Hence, it's not valid to
> call it from within an RCU read-side critical section.
> 
> I assume you're confusing something, free_irq() is called from the destructor of
> the irq::Registration object, hence it is either called when the object itself
> is dropped or from the devres callback, which is called after the
> synchronize_rcu(), but not from an RCU read-side critical section.

Ok hold on ... I guess the issue I thought was there manifests itself in
another way. What about this situation?

Thread 1                 Thread 2
device removal starts
                         Drop for Devres starts running
                         devm_remove_action() = 0
device is fully unbound
                         free_irq()

Now the call to free_irq() happens too late, because there's nothing in
the devm callback stack to wait for it.

Alice