Re: [RFC, PATCH 0/4] SoC base drivers

[Ben Dooks <ben-linux-arm@fluff.org>]

On Tue, May 01, 2007 at 08:08:06AM +0300, Paul Sokolovsky wrote:
> Hello linux-kernel,
> 
> 
> In contemporary systems, lots of functionality oftentimes handled by various
> kinds of SoCs (system-on-chip), representing a number of deversified
> controllers packaged in one chip. Handling them is arguably an ongoing 
> problem, as diversity and number of devices included make it hard to
> come with maintainable and reusable solution for writing drivers. Common
> examples are developing one monolithic driver for all internal devices, or
> making one of device drivers export functions required by others, leading
> to not very clean dependencies like touchscreen driver depending on
> a sound one.
> 
> Handhelds.org kernel project (dealing with Linux porting to consumer,
> real-world embedded devices, and majority of our devices have different kinds
> of SoCs) for few years developed a more clean approach to handling
> SoCs - using the concept of "SoC base drivers". This is not the first time
> we're submitting these patches, with reworking and elaborating them based on
> the feedback received. We also extend supported machine base for these 
> drivers (for example, asic3_base driver in the following patches is used by 
> a dozen of machines).
> 
> SoC base drivers - concept
> --------------------------
> 
> A SoC base driver is a special kind of driver which manages a SoC chip's 
> common and shared resources and functionality, and provides interfaces
> for subdevice drivers to use. This in particukar solves "tangled 
> dependencies" issues mentioned above - different subdevices now depend
> not on each other, but on a common foundation, forming a hierarchial 
> structure.
> 
> Term "interfaces" is used intentionally, as besides offering an API adhoc
> to specific SoC functionality, the intent is to reuse existing kernel
> susbystems/interfaces as much as possible, essentially breaking tight
> coupledness of subdevice drivers and base driver. For example, Generic
> IRQ Subsystem is used to handle SoC interrupts, so subdevice drivers
> just use existing generic API to request and free IRQs.
> 
> More importantly, existing driver model is used to handle subdevices 
> of a SoC. For each subdevice, a standard struct device is allocated,
> and LDM is used to handle driver binding and further lifecycle. So,
> this is a special trair of SoC base drivers: these are the drivers
> which register other devices (note that it is not a case of "legacy" 
> driver where single module registers both a device and a driver
> serving it - SoC base driver registers other devices to be handled by
> other driver, namely SoC individual subdevices and drivers for them).
> Initial implementation from few years ago registered per-SoC bus
> for the purpose of attaching subdevices, but during LKML reviews it
> was pointed out that there're no good reasons for that, as such bus
> does not have any special functionality attached, so now platform_bus
> is used instead, for good.
> 
> For the most part, subdevices are allocated dynamically, and SoC
> base driver calculates/fixes up resources and parameters for them,
> to be suitable for specific configuration (for example, different
> base address of SoC).
> 
> What exact functionality and API a SoC base driver provides depends
> largely on specific chip, there's no specific API a SoC driver should
> implement. Here is a list of common tasks the driver usually would do:
> 
> 1. Access handling to the chip (serialization, locking, etc.)
> 2. Managing common chip resources:
> 2.1. Interrupts control, demultiplexing, etc. (using Generic IRQ subsystem)
> 2.2. GPIO handling (adhoc, while eagerly waiting for an extensible GPIO API, 
> we posted our implementation at http://lkml.org/lkml/2007/4/10/299 ).
> 2.3. Clocks (Using Platform Clock API)
> 2.4. Other kinds of "enable" and "power" switches (in adhoc manner or 
> (ab)using the Clocks API, and waiting for generalization of it).
> 3. Calculating properties and registerting subdevices.

Wow, platform devices with a new name. I don't see how any of this is
not handled by platfrom device.

GPIO devices could be handled by a new resource type of GPIO

The only other item in the list which we do not yet have is a
form of the clock API which can be extended past the base CPU
clock implementation.

Anything registering new IRQs could create sub platform devices
with the correct resources. 
 
> There's a helper soc-core API to help SoC base drivers with common tasks,
> like registering subdevices.
> 
> The implementation and patches following is structured as follows:
> 
> 1. include/linux/soc/ and drivers/soc/ directories are created to
> keep namespaces clean (we have around 10 SoC base drivers now). Note
> that these dirs are for base drivers only, subdevice drivers go to
> one of standard dirs based on their functionality (for a example,
> video driver goes to driver/video/).
> 
> 2. soc-core.c, soc-core.h: helper API to calculate subdevice resources
> and register them based on template definitions provided by SoC drivers.
> 
> 3. asic3_base.c and associated headers: SoC base driver for HTC ASIC3,
> a popular SoC for ARM-based handheld devices, currently known to be used
> in 12 machines, including one machine which is already in mainline.
> 
> 4. mach-3715.c: A patch for HP iPaq rx3715 machine, already in mainline,
> to add ASIC3 support using asic3_base driver.
> 
> 
> I would like to end this RFC with the fact that mainline of course 
> already contains drivers for non-CPU SoC chips. arch/arm/common/sa1111.c 
> and arch/arm/common/locomo.c are two examples from the ARM realm. This RFC
> grows from this fact, and aims to provide consistent definition and
> best prctices for handling such SoCs. 
> 
> Finally, in the beginning, term "reusability" is used, and this is not empty
> word. Besides clean structural approach and improved source-level reusability
> due to this, the approach proposed proved to offer immediate subdevice 
> driver reusability across different SoCs. One vivid example is ds1wm driver,
> for a W1 bus controller. It was found that this controller is found in few
> SoC chips used in consumer handhelds (at least ASIC3, SAMCOP, HAMCOP chips),
> and immediately usable, as long as base drivers prepares IRQ resources and
> clocks for it.

-- 
Ben

Q:      What's a light-year?
A:      One-third less calories than a regular year.