On Mon, Jul 06, 2026 at 04:22:23PM -0700, Rosen Penev wrote:
> During driver removal, devres cleans up in LIFO order. The IRQ domain
> was created and its devm cleanup action registered after
> devm_gpiochip_add_data(), so the domain was destroyed before the GPIO
> chip was deregistered. If gpiod_to_irq() is called on a pin during
> this window, mvebu_gpio_to_irq() passes the freed mvchip->domain to
> irq_create_mapping().

Without having looked at your patch, I would expect that there is a
similar window during probe where calling gpiod_to_irq() between
devm_gpiochip_add_data() and IRQ domain creation results in an invalid
mvchip->domain being passed to irq_create_mapping(), right?

> Fix by moving the IRQ domain creation, devm cleanup action registration,
> generic chip allocation, and chip type setup before
> devm_gpiochip_add_data(). This ensures the GPIO chip is torn down
> first (preventing new IRQ mappings), then the IRQ domain is removed,
> and finally mvchip is freed.
> 
> Fixes: 644ee70267a9 ("gpio: mvebu: fix irq domain leak")

If the above is right, the referenced commit only creates the issue at
driver detach. The issue at probe time exists since 
fefe7b092345 ("gpio: introduce gpio-mvebu driver for Marvell SoCs")
which has `gpiochip_add(&mvchip->chip);` before coping for irqs.

Best regards
Uwe