On Mon, Jul 16, 2018 at 7:58 AM Linus Walleij [off-list ref] wrote:

On Fri, Jul 13, 2018 at 5:27 PM Rob Herring [off-list ref] wrote:

quoted

On Fri, Jul 13, 2018 at 3:47 AM Alexandre Belloni
[off-list ref] wrote:

quoted

Hi Rob,

On 12/07/2018 13:22:22-0600, Rob Herring wrote:

quoted

AT91 pinctrl deferred probing support is broken if the GPIO devices
haven't probed first and set gpio_banks to non-zero. The later condition
that only some GPIO devices haven't probed can't actually happen as
either all the GPIO devices have probed first or none of them have. Plus
the pinctrl driver has already returned -EINVAL, so it's not on the
deferred list to retry probing.

Fix this by consolidating the checking that all GPIO devices are probed.

Cc: Jean-Christophe Plagniol-Villard <redacted>
Cc: Linus Walleij <redacted>
Cc: Nicolas Ferre <nicolas.ferre@microchip.com>
Cc: Alexandre Belloni <alexandre.belloni@bootlin.com>
Cc: linux-arm-kernel at lists.infradead.org
Cc: linux-gpio at vger.kernel.org
Signed-off-by: Rob Herring <robh@kernel.org>
---
This is a result of trying to remove of_platform_default_populate from
at91 code and relying on the DT core handling populating devices. That
changed the probe order and broke booting.

This solves part of the issue but when tested with the
of_platform_default_populate removal, many drivers will fail with
gpiod_set_value: invalid GPIO (errorpointer)

or

gpiod_get_value: invalid GPIO (errorpointer)

This happens both before and after the pinctrl driver probed.

I didn't investigate much yet.

Looks to me like it may be a circular dependency. The GPIO request
functions depend on the pinctrl driver which depends on the gpio
driver. Maybe returning EPROBE_DEFER in at91_gpio_request_enable and
removing the requirement that the GPIO driver probe first would fix
it...

I think you're spot in.

This kind of problem has cropped up in these subsystems since day
one and we probably just have different lame attempts to paper
over it all over.

I suspect that what we need to do is more akin to the DRM model
which I recently learned about (drivers/gpu/drm/vc4/vc4_drv.c is a very
clean and good example) using the core kernel component
infrastructure:

1. Drivers probe() independently, doing intiialization of state
  container struct and retreiveing resources like I/O memory,
  regulators, IRQ, DMA... but no registering interfaces to
  the subsystems yet.

2. The probe of the main component (master) matches the
  subcomponents on the platform bus. In this case, whatever
  needs to come first is the master. I guess pin control.
  return component_master_add_with_match(dev, &vc4_drm_ops, match);

3. Subcomponents:
   return component_add(&pdev->dev, &vc4_dsi_ops);

4. Later on we take control over the binding order.
   the .bind callback of the main component issues
   component_bind_all(dev, drm) and there the subdevices
   are bound.

The process is perfectly reversible. The idea here is to break the
neck of the problem by doing enough set-up in probe() so that
bind() can commence with all its resources. My Nomadik driver
already makes a half-assed attempt at this (not using the component
framework) but I could refactor it to provide an example if
there is interest.

I'm not really a fan of the component f/w. It feels bolted on and
should be more integrated into the driver model IMHO (though I've not
given any thought to what that would look like).

The problem in this case could be easily avoided by removing
"simple-bus" from the pinctrl node (having a driver for a node should
be a clue that "simple-bus" is not appropriate). Then the pinctrl
driver can properly sequence things by calling of_platform_populate to
instantiate the GPIO devices first and then registering with pinctrl
subsys. However, we can't just remove "simple-bus" to fix this without
having a compatibility problem.

Rob