Re: [RFC 6/6] ARM: dts: exynos4210: Add platform-specific descriptions for pin controllers

Hi Stephen,

On Monday 24 of September 2012 17:14:38 Stephen Warren wrote:

On 09/24/2012 03:31 PM, Tomasz Figa wrote:

quoted

On Monday 24 of September 2012 11:42:15 Stephen Warren wrote:

quoted

On 09/21/2012 01:54 PM, Tomasz Figa wrote:

quoted

On Friday 21 of September 2012 12:56:41 Stephen Warren wrote:

quoted

On 09/20/2012 02:53 AM, Tomasz Figa wrote:

quoted

The patch "pinctrl: samsung: Parse pin banks from DT" introduced
platform-specific data parsing from DT.

This patch adds all necessary nodes and properties to exynos4210
device
tree sources.

+++ b/arch/arm/boot/dts/exynos4210-pinctrl-banks.dtsi

+			samsung,pctl-offset = <0x000>;
+			samsung,pin-bank = "gpa0";
+			samsung,pin-count = <8>;
+			samsung,func-width = <4>;
+			samsung,pud-width = <2>;
+			samsung,drv-width = <2>;
+			samsung,conpdn-width = <2>;
+			samsung,pudpdn-width = <2>;

The properties above represent the width of the fields. Must all
fields
always be packed together into say the LSB of the registers? What if
there are gaps between the fields in a future SoC variant, or the
order
of the fields in the register changes? I think you want to add
either
a
samsung,func-bit/samsung,func-position property for each of the
fields,
or change from samsung,func-width=<4> to samsung,field-mask=<0xf>.
IIRC,
the generic pinctrl binding used a mask for this purpose.

What if a future SoC variant adds more fields to the register? At
that
point, you'd need to edit the driver anyway in order to define a new
DT
property to represent the new field. Perhaps instead of having an
explicit named property per field in the register, you want a simple
list of fields:

samsung,pin-property-names = "func", "pud", "drv", "conpdn",
"pudpdn";
samsung,pin-propert-masks = <0xf 0x30 0xc0 0x300 0xc00>;

That would allow a completely arbitrary number of fields and layouts
to
be described.

I wonder if for absolute generality you want a samsung,pin-stride
property to represent the difference in register address per pin. I
assume that's hard-coded as 4 right now.

Hmm, considering so many different possible changes, maybe a more
conservative solution would be better, like reducing the amount of
information held in DT to bank type, e.g.

	samsung,bank-type = "exynos4";

or maybe

	compatible = "samsung,exynos4-pin-bank*;

and then define supported bank types in the driver. SoC-specific data
would remain in DT, i.e. pctl-offset, pin-bank, pin-count,
eint-offset, etc.

Yes, removing much of the data from DT and putting it into a tiny
table
in the driver makes sense to me in this case.

A hybrid solution came to my mind, define bank types in device tree
once

and then reference them in banks. Something like:
	pinctrl-bank-types {
	
		bank_off: bank-off {
		
			samsung,func-width = <4>;
			samsung,pud-width = <2>;
			samsung,drv-width = <2>;
			samsung,conpdn-width = <2>;
			samsung,pudpdn-width = <2>;
		
		};
		
		bank_alive: bank-alive {
		
			samsung,func-width = <4>;
			samsung,pud-width = <2>;
			samsung,drv-width = <2>;
		
		};
	
	};
	
	/* ... */
	
	pinctrl@12345678 {
	
		/* ... */
		gpa0: gpa0 {
		
			/* ... */
			samsung,bank-type = <&bank_off>;
			/* ... */
		
		};
		/* ... */
	
	};

This would add the possibility to define new banks types quickly, but
would not add too much overhead.

What do you think?

That does solve the issue of parsing those same values over and over,
but at the cost of making the binding a lot more conceptually complex.

However, it doesn't solve any of the extensibility issues I raised such
as whether pos+size or mask would be a better representation, what about
if the fields end up being in different (separate) registers on newer
SoCs, etc.

Hmm, could you elaborate on the idea of using mask instead of field widths? 
I don't see how this could be better and there is an additional drawback of 
having to calculate width and pos from every mask.

Anyway, back to your concern, the values that are written to the bit fields 
specified by those bindings are arbitrary SoC-specific values anyway, so 
if, for example, we get a SoC with following register layout:

bits:    7 | 6 - 4  | 3 | 2 - 0
meaning: 0 | func 1 | 0 | func 0

or

bits:    7 - 5  | 4 | 3 - 1  | 0
meaning: func 1 | 0 | func 0 | 0

we can easily define the width as 4 and use appropriate 4-bit function 
values with zeroes on reserved positions.

I forget, do you actually have multiple different SoCs right now (or in
the near future where the HW design is known now for certain even if the
chip isn't available) that have different values for all these *-width
properties and hence can be represented just using this binding and
without altering the driver at all? If so, I suppose the original
binding is at least useful (although I would certainly still request to
use *-mask instead of *-width properties).

The binding I proposed covers all Samsung SoCs currently available, 
starting from s3c24xx, through s3c64xx (except 4bit2 bank type, with two 
function registers), to the whole Exynos series, including latest Exynos5.

Best regards,
--
Tomasz Figa
Samsung Poland R&D Center