Re: [PATCH v11 2/6] gpiolib: allow to specify the firmware node in struct gpio_chip

[Andy Shevchenko <andriy.shevchenko@xxxxxxxxxxxxxxx>]

On Thu, Dec 02, 2021 at 02:52:55PM +0100, Bartosz Golaszewski wrote:
> On Thu, Dec 2, 2021 at 2:45 PM Andy Shevchenko
> <andriy.shevchenko@xxxxxxxxxxxxxxx> wrote:
> >
> > On Thu, Dec 02, 2021 at 02:06:57PM +0100, Bartosz Golaszewski wrote:
> > > On Thu, Dec 2, 2021 at 12:38 PM Andy Shevchenko
> > > <andriy.shevchenko@xxxxxxxxxxxxxxx> wrote:
> > > >
> > > > On Thu, Dec 02, 2021 at 01:35:01PM +0200, Andy Shevchenko wrote:
> > > > > On Thu, Dec 02, 2021 at 12:24:06PM +0100, Bartosz Golaszewski wrote:
> > > > > > On Thu, Dec 2, 2021 at 11:58 AM Andy Shevchenko
> > > > > > <andriy.shevchenko@xxxxxxxxxxxxxxx> wrote:
> > > > > > >
> > > > > > > On Wed, Dec 01, 2021 at 02:11:28PM +0100, Bartosz Golaszewski wrote:
> > > > > > > > On Tue, Nov 30, 2021 at 10:04 PM Bartosz Golaszewski <brgl@xxxxxxxx> wrote:
> > > > > > >
> > > > > > > ...
> > > > > > >
> > > > > > > > Let me maybe rephrase the problem: currently, for GPIO devices
> > > > > > > > instantiating multiple banks created outside of the OF or ACPI
> > > > > > > > frameworks (e.g. instantiated manually and configured using a
> > > > > > > > hierarchy of software nodes with a single parent swnode and a number
> > > > > > > > of child swnodes representing the children), it is impossible to
> > > > > > > > assign firmware nodes other than the one representing the top GPIO
> > > > > > > > device to the gpiochip child devices.
> > > > > > > >
> > > > > > > > In fact if we want to drop the OF APIs entirely from gpiolib - this
> > > > > > > > would be the right first step as for gpio-sim it actually replaces the
> > > > > > > > gc->of_node = some_of_node; assignment that OF-based drivers do for
> > > > > > > > sub-nodes defining banks and it does work with device-tree (I verified
> > > > > > > > that too) thanks to the fwnode abstraction layer.
> > > > > > >
> > > > > > > In exchange of acknowledgements I confirm that I understood the issue
> > > > > > > you are describing. What I still don't like is this band-aid:ish approach.
> > > > > > > What we really need is to replace of_node by fwnode in GPIO library once
> > > > > > > for all. But it can be done later after your simulation series (or before,
> > > > > > > i.o.w. independently), hence I propose to update TODO and do it separately.
> > > > > > >
> > > > > >
> > > > > > But this is what we already do for OF. How would the core gpiolib know
> > > > > > how the firmware nodes represent the banks? It's the driver's job to
> > > > > > tell the framework which node corresponds with what. If anything, we
> > > > > > should start replacing of_nodes with fwnodes in drivers and eventually
> > > > > > we'd drop the of_node pointer from gpio_chip entirely, but we'd keep
> > > > > > the fwnode pointer I added as the driver still needs to assign it
> > > > > > itself.
> > > > > >
> > > > > > Again: I may be missing something here but I've been going through
> > > > > > this on and on and can't figure out any other way. Looking at
> > > > > > gpiolib-acpi.c I don't see it correctly assigning fwnodes to
> > > > > > sub-devices either but I don't have any HW to test it.
> > > > > >
> > > > > > As for this series: I can't really drop this patch as gpio-sim relies
> > > > > > on swnodes being correctly associated with gpio_chips to identify the
> > > > > > gpiodevs from configfs callbacks.
> > > > >
> > > > > Then we need to replace of_node by fwnode as a first step. I have looked
> > > > > briefly into the list of drivers that may have been cleaned up and it doesn't
> > > > > look too long.
> > > >
> > > > Let me kick this off by sending couple of patches.
> > >
> > > Are you fine with merging this in the meantime to get gpio-sim into mainline?
> >
> > gpio-sim, yes, (though I may bikeshed about naming of the configfs attributes,
> > etc) but not this patch.
> >
> 
> There's no way around it though AFAIK. First - the 'gpio-line-names'
> property will not work for banks. 'ngpios' will only work because we
> read it manually in probe() to figure out the number of sysfs groups.
> And also configfs callbacks will not be able to associate bank devices
> with configfs groups. I would really like to hear an alternative -
> even if it's just an idea and not actual implementation.
> 
> I'm really curious to see how you'll remove the of_node pointer and
> not introduce the corresponding fwnode pointer actually.

Seems I was unclear, fwnode pointer will be needed, but what I'm against of is
having of_node and fwnode at the same time in the struct gpio_chip.

Yes, we may modify this patch to work without that ugly ifdeffery and with both
in the structure, but I don't think it's a good solution.

Now clearly we have to clean up of_node first.

-- 
With Best Regards,
Andy Shevchenko