Re: [PATCH] ARM: shmobile: r8a73a4: Move pfc node to work around probe ordering bug

[Tony Lindgren <tony@atomide.com>]

* Geert Uytterhoeven <geert@linux-m68k.org> [150209 09:17]:
> On Mon, Feb 9, 2015 at 5:24 PM, Tony Lindgren <tony@atomide.com> wrote:
> > * Geert Uytterhoeven <geert+renesas@glider.be> [150206 12:26]:
> >> Notes:
> >>   - It seems several people tried to solve this in the core OF probing
> >>     code before, but the final solution never went in?
> >>   - This can be reproduced on other SoCs (e.g. sh73a0 and r8a7740) by
> >>     moving their pfc nodes before their interrupt controller nodes.
> >>   - This patch is against my working tree, so it doesn't apply to
> >>     Simon's repository, but you get the idea....
> >
> > No issues with the patch, but here are few comments on the core
> > reasons (without looking at the code in this case) that might help
> > fix similar issues.
> >
> > In all the cases I've seen these errors are caused by non-standard
> > custom initcall levels for drivers like i2c bus. The real solution
> > is to initialize drivers later with standard module_init, and stop
> > the race to the bottom with custom initcall levels.
> >
> > If there is legacy board specific platofrm init code that needs
> > i2c gpios early, that code can probably be moved to initialize
> > later on.
> 
> In this case no i2c is involved. The drivers for both pinctrl
> (renesas,pfc-r8a73a4) and irqchip (renesas,irqc) are registered
> at the same level:
>   - postcore_initcall(sh_pfc_init);
>   - postcore_initcall(irqc_init);
> Hence the system uses the "natural" order from within the DTS,
> and decided to instantiate the pfc before the irqchip.

OK
 
> > Also, there should not be any need for custom driver initcall
> > levels from Linux generic framework point of view as for example
> > irqchip implementing drivers work just fine as a loadable module.
> 
> As long as no other device that's instantiated earlier references that
> irqchip?

Right :) And deferred probe won't still remove the warning in
that case. From what I remember, that's a valid warning from irq
framework point of view.

Regards,

Tony