On 26/07/12 15:43, Mark Brown wrote:
On Thu, Jul 26, 2012 at 03:15:01PM +0100, Lee Jones wrote:
Sorry missed this:
Why are we doing this? The MFD cells are a totally Linux specific thing, there's no reason to represent them in the device tree unless they're in some way reusable and the "ab8500-codec" name suggests that's unlikely. Just put the properties on the parent node and instantiate the MFD cell as normal.
We have all of the AB8500 devices into the Device Tree to accurately represent the hardware. We will also be passing configuration information into the AB8500 Codec from Device Tree. The only reason we're still registering them using the MFD API is to overcome addressing issues encountered earlier. Each 'device' still belongs in the 'device' tree.
The device here is the AB8500. The fact that Linux chooses to represent it as an MFD with a particular set of subdrivers is a Linux specific decision and may well change over time. For example it's likely that we'll want to migrate the clocks out of the audio driver and into the clock API when that becomes useful. Similarly currently a lot of these devices use ASoC level jack detection but that's going to move over to extcon over time.
There's no way you're going to be able to reuse this for anything that isn't an AB8500, there's no abstraction of the SoC integration here. If you had clearly identifiable, repeatable IPs which you could reasonably bind to a different bit of silicon then that'd be different but there's nothing like that here. We already know that the functionality covered by the driver is going to be present simply by virtue of knowing that there's an AB8500 and similarly there's no real way this driver could ever be used without the core driver. All the "device" in the device tree is doing is serving as a container to place some of the DT properties into, this needs to be separated out from the instantiation of the Linux device driver. There's nothing stopping the driver from looking at the OF node of its parent here.
The goal here isn't just to copy the Linux device model and platform data into device tree bindings, the device tree bindings need to think about what the chip actually is so they can be reused by other OSs and by future versions of Linux.
If we were to take this Device Tree and use it on something non-Linux, that OS will still need to know about each of the AB8500 devices and every associated configuration option. Only in Linux do we continue to register them though a different API, which doesn't affect any other OS.
Another OS might have a different idea about how it's going to split up the chip which better fits with the models which that OS has for the functions present on the device. The reason this is a distinct device in Linux is all to do with how Linux models the hardware.
Okay, so your suggestion is to strip out all of the sub-devices under the AB8500. It's doable, but will take some restructuring and thinking about. This is a job for another day. I think it's okay to continue with the current semantics for the time-being. The line we're discussing does need to be split out though. I didn't mean to merge it in with the ASoC stuff.