On Mon, Nov 03, 2025 at 10:57:44AM +0100, Hans de Goede wrote:
On 3-Nov-25 10:35 AM, Andy Shevchenko wrote:
On Sun, Nov 02, 2025 at 08:09:21PM +0100, Hans de Goede wrote:
Since commit d24cfee7f63d ("spi: Fix acpi deferred irq probe"), the acpi_dev_gpio_irq_get() call gets delayed till spi_probe() is called on the SPI device.
If there is no driver for the SPI device then the move to spi_probe() results in acpi_dev_gpio_irq_get() never getting called. This may cause problems by leaving the GPIO pin floating because this call is responsible for setting up the GPIO pin direction and/or bias according to the values from the ACPI tables.
Re-add the removed acpi_dev_gpio_irq_get() in acpi_register_spi_device() to ensure the GPIO pin is always correctly setup, while keeping the acpi_dev_gpio_irq_get() call added to spi_probe() to deal with -EPROBE_DEFER returns caused by the GPIO controller not having a driver yet.
Even before following the link to some papering over module via the link below I wondered, if the I²C case should be covered as well. The https://github.com/alexpevzner/hotfix-kvadra-touchpad refers to I²C enabled touchpads.
...
I'm not against the SPI fix, but is it confirmed that it really fixes the issue?
Yes Mark and I got an offlist email bisecting this to the:
Fixes: d24cfee7f63d ("spi: Fix acpi deferred irq probe")
commit (on a stable kernel series) and a later email confirming that this patch fixes it.
Shouldn't we use Closes in this case instead of Link?
It seems that leaving the fingerprint reader enable pin (the first GPIO listed in _CRS which is an output only pin, is likely the enable pin) floating is causing these issues. So in a way the acpi_dev_gpio_irq_get() fixing this by forcing the enable pin to no longer float is a bit of luck. But things did work before d24cfee7f63d ("spi: Fix acpi deferred irq probe") so we need this to fix a regression
Yeah, fixing a regression is good thing, but not papering over the issue.
and as you indicate it seems like a good idea in general and maybe we should also do this for i2c.
As for doing something similar for I2C devices, that is an interesting question. Even though it is possible I'm not aware of any i2c-devices which have a userspace driver like SPI/USB fingerprint readers do, so on i2c I would expect probe() to always get called. So I'm not sure it is necessary there.
Reading the problem statement (the second paragraph) I lean towards a generic solution residing somewhere in drivers/acpi/scan.c (like acpi_init_device_object() / acpi_bus_attach() calls), although I don't see a quick way how to achieve this. It seems would require a bit of refactoring to allow ACPI glue code to call specific subsystem calls or making a GPIOLIB to provide some "early" initialisation flow(s).