[RFC PATCH] CLK: Allow parent clock and rate to be configured in DT.

[Tomasz Figa]

Hi,

[RANT]

I tend to disagree about this whole hype about device tree usage for other 
things than pure hardware description. I don't think device tree should be  
a way to force some kind of new world order, but rather a more convenient 
and more maintainable (than board files) way of support hardware platforms 
in Linux kernel.

On Friday 05 of April 2013 20:07:09 Matt Sealey wrote:
> On Thu, Apr 4, 2013 at 6:08 PM, Fabio Estevam <festevam at gmail.com> 
wrote:
> > This could be useful for removing the imx6q_sabrelite_cko1_setup()
> > function from arch/arm/mach-imx/mach-imx6q.c.
> > 
> > I would like to add audio support for another board and would like to
> > avoid to do the same as imx6q_sabrelite_cko1_setup()  for setting up
> > the CLKO, if possible.
> 
> You know, we have the same problem on one of our designs here (CLKO is
> used on MX53 for audio codec and camera device, shared) - the current
> solution is hack it into platform support in a BSP kernel.
> 
> If we move to device tree, we know and you know the solution already;
> all this clock setup HAS to be done in the bootloader.

Assuming that you can change the bootloader... Isn't bootloader this piece 
of code that you shouldn't touch if it's working (especially in production 
environments)?

> The device tree really, really is only a way to describe the
> configuration as it exists or to describe alternatives - for instance,
> a clock with 10 parents may be described as having 10 parents, and the
> binding (in complicated cases) or driver (if it is simple as a value
> from 0 to 7 and the field width is 3 bits for example) will determine
> how that alternative can be selected (and by consequence, what the
> current setting is).
> 
> The device tree concept is NOT a place to dump configuration items for
> your board as hardcoded values to try and force something you could
> have done elsewhere. On i.MX53 you cannot boot a kernel verbatim - you
> at least have to run through the Boot ROM and supply a DCD table or
> plugins first. This is where you figure things like this out.

Why not? For Linux and most of ARM-based platforms device tree is just a 
replacement for board files. They are not going to be used with any other 
imaginary kernel supporting device tree. On many of them even the device 
tree blob will be distributed along with kernel image, if not simply 
appended to the zImage.

I think it should be up to the board/platform maintainer whether they want 
to limit device tree on particular boards/platforms just to hardware 
description or extend it to handle everything that was originally handled 
by board files.

> In a case where you have, say, an audio codec and it has a dynamic
> input clock that you want to change on the fly, first of all you would
> not hardcode a frequency into a device tree, second of all there is
> nothing stopping you from doing that right now anyway. If the clock is
> static and fixed frequency and can only be turned on and off, there
> are clock bindings for this already. If it is static and can never be
> turned off, then there are clock bindings for this too, and in this
> case the proviso is that the clock is ALREADY turned on and ALREADY
> stable before you enter the kernel otherwise the description is by
> it's very definition invalid.
> 
> If the clock frequency in hardware is not what you wanted when the
> driver comes up, and you only have an on/off switch, it is not up to
> the device tree binding to describe how to go about configuring the
> system to make sure. You cannot in good conscience put a clock
> frequency "to be set later" in the device tree along with a clock
> phandle, simply because that means the device tree no longer describes
> the hardware accurately - the clock phandle is a valid clock, the
> hardware will tell you a frequency from registers in the clock driver,
> the device tree will disagree. How do you know which one is the
> correct value, or if the frequency in the tree is a suggestion rather
> than a description?

Sure. This is (and has always been) something to account for when bringing 
up new platforms from the ground up.

But you can't always have control over the bootloader. What's wrong in 
letting the kernel configure the board itself? It must configure most of 
the hardware anyway, based on platform data (either located in board files 
or parsed from device tree).

Why not to make the kernel independent from the bootloader at all? Then 
the bootloader could just do some minimal initialization needed to load a 
kernel image from flash memory and launch it (+ some code to allow 
flashing of new images).

> I am totally against this (sorry Sascha..). Let's put some effort into
> fixing the bootloaders rather than trying to use the device tree to
> enforce the ridiculous assumption that "Linux kernel does not trust
> the bootloader". Once the bindings and trees are out of the kernel
> source, you're going to HAVE to trust what the bootloader passes, be
> it pregenerated compiled blob (which needs to be written to match the
> hardware state the bootloader finishes up at) or dynamically generated
> at runtime during the boot process (which can describe to the bit what
> the hardware is doing). If you are testing this on Beagle XM you can
> fix your U-Boot easily. New boards will have to be designed *with the
> idea of device trees and working configurations in mind* - that is a
> fact of life, in fact this is how it should be in reality these days
> anyway (most HW designers will do initial bringup of the board - at
> least a minimal working configuration, in a restricted environment
> where they can use test pads to measure clocks, voltage outputs and
> levels of devices to ensure both production was successful and that
> the design is in itself electrically sound. This code 99% of the time
> ends up in the bootloader... sometimes not, when the board was
> designed by one group and the software written by the community, but
> in this case the community needs to learn board bringup and proper
> behavior...)

Well, so you are basically suggesting to throw away Linux support for 
boards/platforms designed without device tree in mind and on which 
replacing the bootloader is simply infeasible or sometimes even 
impossible.

I don't think that ARM Linux is adopting device tree just to have device 
tree. There are many problems that can be solved using device tree, like 
multiplatform support, separation of board support from kernel code, 
binding consumers with providers (PHYs, GPIOs, clocks, etc.). Why do we 
have to cover all these advantages behind a curtain of new problems?

[/RANT]

Best regards,
Tomasz