[PATCH v4 1/5] drivers: memory: Introduce Marvell EBU Device Bus driver

[ezequiel.garcia@free-electrons.com (Ezequiel Garcia)]

Hi Arnd,

On Tue, Apr 09, 2013 at 11:40:51AM +0200, Arnd Bergmann wrote:
> On Tuesday 09 April 2013, Ezequiel Garcia wrote:
> > > This can probably just be of_platform_populate or similar to do all
> > > children? For instance, I often use many DT nodes to represent a FPGA,
> > > since the my FPGA's tend to have many functionally orthogonal units
> > > inside.
> > > 
> > 
> > In the past I've found it's not possible to use of_platform_populate,
> > altough I can be wrong.
> > 
> > The problem seems to be that through of_platform_populate() it's
> > possible that the child device (cfi-flash, for instance) gets probed
> > before the devbus controller. In other words, the actual parent-child
> > relationship gets lost and it's necesarry to use some child probe deferal
> > mechanism.
> > 
> > This has been recently discussed [1] when implementing GPMC where for simplicity
> > the of_platform_device_create() solution was chosen.
> > Anyway, I'll try of_platform_populate early tomorrow.
> 
> I think the problem mentioned in that thread was that the device would
> automatically get probed by the top-level of_platform_populate() when
> you mark the devbus as compatible="simple-bus", which is not what Jason
> was referring to. Instead you should call of_platform_populate in the
> probe() function of the devbus device to probe its children after
> the bus is set up.
> 

Ah! yes, you're right...

Well, in that case the only issue I can foresee is that if we decide
to use of_platform_populate we won't be able to unregister child
devices from the remove() callback.

Indeed, the benefits of using of_platform_populate are interesting,
but I don't know how much of an issue this represents.

If we can't unregister child devices, we can't remove address windows.
Now, this is not a big deal, since we plan to define them statically in
the DT anyway.
In that case, it seems we shouldn't allow this driver to be a module, uh?

(actually we currently can't have mvebu-devbus as a module, because
mbus API is not exported, but we can fix that if we want).

What do you think?
-- 
Ezequiel Garc?a, Free Electrons
Embedded Linux, Kernel and Android Engineering
http://free-electrons.com