[PATCH] dm: pinctrl: Revert "pinctrl: probe pinctrl drivers during post-bind"

[Simon Glass]

Hi Pali,

On Fri, 30 Dec 2022 at 12:02, Pali Rohár <pali at kernel.org> wrote:
>
> On Friday 30 December 2022 11:47:29 Simon Glass wrote:
> > Hi Pali,
> >
> > On Fri, 30 Dec 2022 at 10:13, Pali Rohár <pali at kernel.org> wrote:
> > >
> > > On Friday 30 December 2022 10:00:11 Simon Glass wrote:
> > > > Hi Pali,
> > > >
> > > > On Fri, 30 Dec 2022 at 09:53, Pali Rohár <pali at kernel.org> wrote:
> > > > >
> > > > > On Friday 30 December 2022 09:30:27 Simon Glass wrote:
> > > > > > Hi Pali,
> > > > > >
> > > > > > On Wed, 21 Dec 2022 at 17:02, Pali Rohár <pali at kernel.org> wrote:
> > > > > > >
> > > > > > > On Wednesday 21 December 2022 07:27:39 Simon Glass wrote:
> > > > > > > > This breaks chromebook_coral and it is also not how things should work. If
> > > > > > > > a board needs to bind GPIOs as part of a pinctrl driver this can be done
> > > > > > > > during the bind step, if needed.
> > > > > > > >
> > > > > > > > We cannot probe pinctrl devices when binding as a rule, since it cannot be
> > > > > > > > supported on some platforms.
> > > > > > > >
> > > > > > > > The bind and probe steps are separate in U-Boot and they should remain
> > > > > > > > separate.
> > > > > > > >
> > > > > > > > This reverts commit f9ec791b5e24378b71590877499f8683d5f54dac.
> > > > > > >
> > > > > > > Unfortunately reverting this patch would break other devices, mostly
> > > > > > > A3720 based where pinctrl driver acts also as gpio driver. Because no
> > > > > > > other caller then register gpio driver and so other drivers which parses
> > > > > > > gpios from DT (which belongs to that gpio driver) will fail during
> > > > > > > probe.
> > > > > >
> > > > > > That is something to be sorted out for that platform. You can bind
> > > > > > GPIO devices in the pinctrl driver's bind() method as other SoCs do.
> > > > > > Even better, the device tree typically has that info in it, i.e. GPIO
> > > > > > subnodes within the pinctrl node.
> > > > > >
> > > > > > Probing pinctrl in a bind function is unfortunately just wrong. It
> > > > > > will cause all sorts of problems, and perhaps already has.
> > > > >
> > > > > Ok, so it means that drivers/pinctrl/mvebu/pinctrl-armada-37xx.c needs
> > > > > to be refactored and fixed to handle these restrictions.
> > > >
> > > > It is not a restriction. It is simply that binding and probing are
> > > > different things and we should not tie them together. It will just
> > > > become a nightmare for board bringup and other drivers.
> > > >
> > > > >
> > > > > > > Also I think that pinctrl command would not work in this case if pinctrl
> > > > > > > would not be probed.
> > > > > >
> > > > > > Devices are probed before use, including by commands.
> > > > > >
> > > > > > This is quite important to fix before the release.
> > > > >
> > > > > Unfortunately in this time I do not have any a3720 board for testing.
> > > > > Robert was able to easily trigger this issue and also introduced that
> > > > > commit f9ec791b5e24 ("pinctrl: probe pinctrl drivers during post-bind").
> > > > >
> > > > > So may I ask Robert to look at it? In past days I looked at powerpc
> > > > > release issues and I do not have time for other things.
> > > >
> > > > That driver has a few FIXMEs in it and could use a look anyway. I see
> > > > that the gpio controller is a subnode of pinctrl anyway. Adding a
> > > > compatible string for it would fix the problem  just like that, and
> > > > remove a lot of ugly code in the driver. This Linux-centric nature of
> > > > device tree bindings really is infuriating:
> > >
> > > All a3720 DTS files are 1:1 copied from Linux. So if problem is in DTS
> > > file then it should be discussed with Linux dt/mvebu maintainers. This
> > > year I fixed U-Boot code to handle Linux a3720 DTS files and then copied
> > > DTS files from Linux. So it is not a good idea to have again different
> > > DTS file for u-boot and different for kernel.

That was not my suggestion. I would simply like the bindings in Linux
to be more explicit, rather than having driver code manually written
to do what the device tree is supposed to do.

> > >
> > > > /* FIXME: Use livtree and check the result of device_bind() below */
> > > > fdt_for_each_subnode(subnode, blob, node) {
> > > >     if (fdtdec_get_bool(blob, subnode, "gpio-controller")) {
> > > >         ret = 0;
> > > >         break;
> > > >     }
> > > > };
> > > > if (ret)
> > > >    return ret;
> > > >
> > > >
> > > > Failing that it just needs a bind() method that calls
> > > > armada_37xx_gpiochip_register()
> > > >
> > > > Regards,
> > > > Simon
> > >
> > > Seems that it is not such simple. armada_37xx_gpiochip_register()
> > > depends on initialized and bound pinctrl part of driver. So before
> > > binding gpio you need to bind pinctrl as they share internal structures.
> >
> > Where are you seeing that? From what I can tell it just binds the GPIO
> > driver. It doesn't probe it. So long as you bind the GPIO driver in
> > armada_37xx_pinctrl_bind() it should be equivalent.
>
> armada_37xx_gpiochip_register() calls device_bind() for
> &armada_37xx_gpio_driver which probe method armada_37xx_gpio_probe() and
> ops &armada_37xx_gpio_ops callbacks access a37xx pinctrl internal
> structures.

Yes but probing is different from binding, please see [1]

>
> So I'm not sure if there is an issue or not. But for sure a37xx gpio
> must be probed after a37xx pinctrl is probed because a37xx pinctrl probe
> function fills internal a37xx pinctrl strucutre used by a37xx gpio probe
> function.

Yes, children are probed after parents, as in docs:

"All parent devices are probed. It is not possible to activate a
device unless its predecessors (all the way up to the root device) are
activated. This means (for example) that an I2C driver will require
that its bus be activated."

>
> > And don't forget the root cause of this whole problem is Linux-centrix
> > DT bindings.

Regards,
Simon

[1] https://u-boot.readthedocs.io/en/latest/develop/driver-model/design.html#driver-lifecycle