[RFC/PATCH 7/7] WIP: HACK/RFC: omap_device: begin to decouple platform_device from omap_device

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

On Thu, Jul 21, 2011 at 04:52:18PM -0700, Kevin Hilman wrote:
> Rather than embedding a struct platform_device inside a struct
> omap_device, decouple them, leaving only a pointer to the
> platform_device inside the omap_device.
> 
> This patch uses devres to allocate and attach the omap_device to the
> struct device, so finding an omap_device from a struct device means
> using devres_find(), and the to_omap_device() helper function was
> modified accordingly.
> 
> RFC/Hack alert:
> 
> Currently the driver core (drivers/base/dd.c) doesn't expect any
> devres resources to exist before the driver's ->probe() is called.  In
> this patch, I just comment out the warning, but we'll need to
> understand why that limitation exists, and if it's a real limitation.
> A first glance suggests that it's not really needed.  If this is a
> true limitation, we'll need to find some way other than devres to
> attach an omap_device to a struct device.
> 
> On OMAP, we will need an omap_device attached to a struct device
> before probe because device HW may be disabled in probe and drivers
> are expected to use runtime PM in ->probe() to activate the hardware
> before access.  Because the runtime PM API calls use omap_device (via
> our PM domain layer), we need omap_device attached to a
> platform_device before probe.

This feels really wrong to overload devres with this.  devres purpose is
to provide the device's _drivers_ with a way to allocate and free resources
in such a way to avoid leaks on .remove or probe failure.  So I think that
overloading it with something that has a different lifetime is completely
wrong.

We could add a new member to struct dev_archdata or pdev_archdata to carry
a pointer to this data, which I think would be a far cleaner (and saner)
way to deal with this.  In much the same way as we already have an of_node
member in struct device.