[PATCH] mmc: mmci.c: Defer probe() in case of missing GPIOs

[linux@arm.linux.org.uk (Russell King - ARM Linux)]

On Sat, Jun 16, 2012 at 04:57:48PM +0200, Roland Stigge wrote:
> On 16/06/12 16:26, Russell King - ARM Linux wrote:
> > On Sat, Jun 16, 2012 at 04:14:59PM +0200, Roland Stigge wrote:
> >> If the GPIOs used by the MMCI driver are not registered yet when the
> >> driver is probe()d, they can't be used. This happens if the mmci driver
> >> is probed before the respective GPIO controller (e.g. on the LPC32xx
> >> EA3250 board, the PCA9532 GPIO controller would be initialized via DT
> >> after mmci). Therefore, we defer mmci in this case.
> > 
> > This code is wrong.  There are platforms where plat->gpio_cd is negative
> > (because there isn't an associated GPIO) and we still expect the driver
> > to successfully bind.  In that case, the driver gets the CD and WP
> > information via the status callback.
> > 
> > So this is an incompatible change with existing (and required) driver
> > behaviour.
> 
> As someone just told me, in the case of no GPIO, we would have gpio_cd
> == -ENODEV. Would it be sufficient to check for -ENODEV (in which case
> we would do without GPIO), and otherwise return -EPROBE_DEFER?

Sigh.  New rules which jump up when no one's read the bloody code.
That's intensely frustrating.

Why can't the *new* DT parsing code set the GPIOs to -EPROBE_DEFER when
*it* wants them to mean "we should defer" ?  Why should the DT code
have sole use over "-1" ?  That would be *far* more logical and wouldn't
clash with anything in existing use.