Re: pxa_defconfig runtime failures due to 'ARM: pxa: activate pinctrl for device-tree machines'

From: Guenter Roeck
Date: Sat May 28 2016 - 17:17:23 EST


On 05/28/2016 01:24 AM, Robert Jarzmik wrote:
Guenter Roeck <linux@xxxxxxxxxxxx> writes:

Hi,

your mainline commit f806dac5938b ("ARM: pxa: activate pinctrl for device-tree
machines") causes various non-devicetree systems to fail with the following
error messages when running a pxa_defconfig image.
Ah yes, you're right.


During boot:

Can't request reset_gpio

At reboot:

reboot: Restarting system
------------[ cut here ]------------
kernel BUG at arch/arm/mach-pxa/reset.c:59!
...

Added logging shows that the error seen when trying to request the reset gpio
is -EPROBE_DEFER, and that the pxa gpio driver is not instantiated.

This is seen when attempting to run akita, borzoi, spitz, terrier, or tosa
in qemu with pxa_defconfig. Reverting your patch fixes the problem.

Is this on purpose ?
Well no :)
The real reason behind is that gpio handling for pxa in its current state cannot
be built for _both_ a devicetree machine (ie. pxa-dt.c) and a non devicetree
machine (ie. corgi, tosa, ...).

This is turn is because for devicetree a pinctrl is enforced for the machine,
and a pinctrl driver is required. If it's not available, pxa_gpio_request()
fails on pinctrl_request_gpio() and returns -EPROBE_DEFER.

Now the true chicken and egg problem is than machine files,
ie. arch/arm/mach-pxa/xxx.c are using gpio before the drivers are probed, in the
init_machine() function, and that's why pinctrl/gpio for legacy machine files is
a bit difficult.

Unless I am missing something, it effectively means that
pxa_defconfig no longer works for pxa3xx systems, since those do not support
devicetree (or at least there is no devicetree file which includes
pxa3xx.dtsi).
I'd rather say that pxa_defconfig doesn't work anymore on any legacy system.
Its first purpose was to ensure compilation coverage of all legacy pxa systems.
Its second one was to have a single kernel bootable on all legacy pxa systems.

Therefore, would you tell me if the patch in [1] fixes your issue ?


Yes, it does. It is actually the workaround I deployed for my test builds,
since I thought that the change was on purpose.

Thanks,
Guenter