On Fri, May 30, 2014 at 4:12 AM, Zhu, Lejun [off-list ref] wrote:

        retval = gpiochip_add(&cg->chip);
        if (retval) {
                dev_warn(&pdev->dev, "add gpio chip error: %d\n", retval);
                return ret;
        }

        gpiochip_irqchip_add(&cg->chip, &crystalcove_irqchip, 0,
                             handle_simple_irq, IRQ_TYPE_NONE);

        retval = request_threaded_irq(irq, NULL, crystalcove_gpio_irq_handler,
                                      IRQF_ONESHOT, KBUILD_MODNAME, cg);

You should request the interrupt before you add the irqchip
I think. But it shouldn't really matter, mainly to avoid tearing
down the irqchip if getting the irq should fail.

But this code will trigger a crash in gpiolib-acpi. Currently at the end
of gpiochip_add(), it calls:

gpiochip_add() -> acpi_gpiochip_add() -> acpi_gpiochip_request_interrupts()

acpi_gpiochip_request_interrupts() needs ->to_irq to work. Without having
called gpiochip_irqchip_add() already, this will be NULL:

        if (!chip->to_irq)
                return;    <-- It will return here.

        INIT_LIST_HEAD(&acpi_gpio->events);

In the tear down path, acpi_gpiochip_free_interrupts() will find to_irq is
no longer NULL, then it will walk an uninitialized list.

So, should this be fixed in gpiolib-acpi?

Maybe, maybe in the drivers. I think Mika has a proposed solution...

Yours,
Linus Walleij