Re: [PATCH] pinctrl: cherryview: fix issues caused by dynamic gpio irqs mapping

From: Linus Walleij
Date: Sat Oct 07 2017 - 20:42:33 EST


On Tue, Oct 3, 2017 at 7:00 PM, Grygorii Strashko
<grygorii.strashko@xxxxxx> wrote:

> New GPIO IRQs are allocated and mapped dynamically by default when
> GPIO IRQ infrastructure is used by cherryview-pinctrl driver.
> This causes issues on some Intel platforms [1][2] with broken BIOS which
> hardcodes Linux IRQ numbers in their ACPI tables.
>
> On such platforms cherryview-pinctrl driver should allocate and map all
> GPIO IRQs at probe time.
> Side effect - "Cannot allocate irq_descs @ IRQ%d, assuming pre-allocated\n"
> can be seen at boot log.
>
> NOTE. It still may fail if boot sequence will changed and some interrupt
> controller will be probed before cherryview-pinctrl which will shift Linux IRQ
> numbering (expected with CONFIG_SPARCE_IRQ enabled).
>
> [1] https://bugzilla.kernel.org/show_bug.cgi?id=194945
> [2] https://lkml.org/lkml/2017/9/28/153
> Cc: Andy Shevchenko <andy.shevchenko@xxxxxxxxx>
> Cc: Chris Gorman <chrisjohgorman@xxxxxxxxx>
> Cc: Mika Westerberg <mika.westerberg@xxxxxxxxxxxxxxx>
> Cc: Heikki Krogerus <heikki.krogerus@xxxxxxxxxxxxxxx>
> Signed-off-by: Grygorii Strashko <grygorii.strashko@xxxxxx>
> Reported-by: Chris Gorman <chrisjohgorman@xxxxxxxxx>
> Reported-by: Mika Westerberg <mika.westerberg@xxxxxxxxxxxxxxx>

OK patch applied for fixes.

But I'mm still very sceptical about this.

Look at the following (just from grep irq_base drivers/gpio/):

drivers/gpio/gpio-ml-ioh.c:

static int ioh_gpio_to_irq(struct gpio_chip *gpio, unsigned offset)
{
struct ioh_gpio *chip = gpiochip_get_data(gpio);
return chip->irq_base + offset;
}

(...)
ch = irq - chip->irq_base;
if (irq <= chip->irq_base + 7) {
im_reg = &chip->reg->regs[chip->ch].im_0;
im_pos = ch;
(...)

drivers/gpio/gpio-sta2x11.c:

static int gsta_gpio_to_irq(struct gpio_chip *gpio, unsigned offset)
{
struct gsta_gpio *chip = gpiochip_get_data(gpio);
return chip->irq_base + offset;
}

(etc)

The thing is that the lines you deleted from gpiolib were the only
thing ever assigning chip->irq_base. This patch, if performed
properly should have removed the .irq_base field from struct
gpio_chip altogether without regressions.

As it is not, anything using .irq_base is regressing.

Either you need to convince me you can quickfix all of these
users, or we need to simply revert this change.

Yours,
Linus Walleij