Re: [PATCH] gpiolib: acpi: Fix failed in acpi_gpiochip_find() by adding parent node match

[Benjamin Tissoires <bentiss@xxxxxxxxxx>]

On May 13 2024, Andy Shevchenko wrote:
> On Mon, May 13, 2024 at 03:59:01PM +0800, Devyn Liu wrote:
> > Previous patch modified the standard used by acpi_gpiochip_find()
> > to match device nodes. Using the device node set in gc->gpiodev->d-
> > ev instead of gc->parent.
> > 
> > However, there is a situation in gpio-dwapb where the GPIO device
> > driver will set gc->fwnode for each port corresponding to a child
> > node under a GPIO device, so gc->gpiodev->dev will be assigned the
> > value of each child node in gpiochip_add_data().
> > 
> > gpio-dwapb.c:
> > 128,31 static int dwapb_gpio_add_port(struct dwapb_gpio *gpio,
> > 			       struct dwapb_port_property *pp,
> > 			       unsigned int offs);
> > port->gc.fwnode = pp->fwnode;
> > 
> > 693,39 static int dwapb_gpio_probe;
> > err = dwapb_gpio_add_port(gpio, &pdata->properties[i], i);
> > 
> > When other drivers request GPIO pin resources through the GPIO device
> > node provided by ACPI (corresponding to the parent node), the change
> > of the matching object to gc->gpiodev->dev in acpi_gpiochip_find()
> > only allows finding the value of each port (child node), resulting
> > in a failed request.
> > 
> > Reapply the condition of using gc->parent for match in acpi_gpio-
> > chip_find() in the code can compatible with the problem of gpio-dwapb,
> > and will not affect the two cases mentioned in the patch:
> > 1. There is no setting for gc->fwnode.
> > 2. The case that depends on using gc->fwnode for match.
> 
> Thanks for the report, analysis, and patch.
> 
> ...
> 
> >  static int acpi_gpiochip_find(struct gpio_chip *gc, const void *data)
> >  {
> > -	return device_match_acpi_handle(&gc->gpiodev->dev, data);
> > +	return device_match_acpi_handle(&gc->gpiodev->dev, data) ||
> > +		(gc->parent && device_match_acpi_handle(gc->parent, data));
> >  }

The original patch (from Devyn) is:
Tested-by: Benjamin Tissoires <bentiss@xxxxxxxxxx>

For reference, a successful run can be seen at:
https://gitlab.freedesktop.org/bentiss/hid/-/jobs/58661693

> 
> I'm wondering if the below approach will work for all:
> 
> static int acpi_gpiochip_find(struct gpio_chip *gc, const void *data)
> {
> 	struct device *dev = acpi_get_first_physical_node(ACPI_COMPANION(&gc->gpiodev->dev));
> 
> 	return device_match_acpi_handle(dev, data);
> }

Looks like I get a bad dev pointer in this situation:
https://gitlab.freedesktop.org/bentiss/hid/-/jobs/58662689#L704

Not sure if this is because the not-yet-upstream patches I have in
hid-cp2112 are doing something wrong or if there is a good reason for
it...

The patch that adds fwnode to hid-cp2112 are:
https://gitlab.freedesktop.org/bentiss/gitlab-kernel-ci/-/blob/master/VM/0002-HID-usbhid-Share-USB-device-firmware-node-with-child.patch
https://gitlab.freedesktop.org/bentiss/gitlab-kernel-ci/-/blob/master/VM/0003-HID-cp2112-Fwnode-Support.patch

both of those patches are applied before compilation in the CI run from
above.

> 
> Cc'ing to Benjamin for testing and commenting.

TL;DR: initial patch is fine, yours will probably need a check on the
dev return value, so not sure if we gain a lot...

Cheers,
Benjamin