[RFC] pinmux: group and function definitions in the device tree

[Ludovic Desroches]

On Mon, Mar 23, 2015 at 07:44:24AM +0100, Sascha Hauer wrote:
> On Fri, Mar 20, 2015 at 04:06:09PM +0100, Ludovic Desroches wrote:
> > On Thu, Mar 19, 2015 at 07:56:37PM +0100, Sascha Hauer wrote:
> > > On Thu, Mar 19, 2015 at 04:39:50PM +0100, Ludovic Desroches wrote:
> > > > 	[...]
> > > > 
> > > > 	pinctrl_defs {
> > > > 		mci0 {
> > > > 			mci0_ioset0_1bit_grp {
> > > > 				at91,pins = <68 69 70>;
> > > > 				at91,mux = <2>;
> > > > 			};
> > > > 
> > > > 			mci0_ioset0_4bit_grp {
> > > > 				at91,pins = <68 69 70 71 72 73>;
> > > > 				at91,mux = <2>;
> > > > 			};
> > > > 
> > > > 			mci0_ioset0_8bit_grp {
> > > > 				at91,pins = <68 69 70 71 72 73 74 75 76 77>;
> > > > 				at91,mux = <2>;
> > > > 			};
> > > > 		};
> > > > 	};
> > > 
> > > Why are different groups here? Do you want to put them into the dtsi?
> > 
> > We used to have a configuration per pin in our products. On next ones we
> > will have some constraints ie. on the controller side we still have a
> > configuration per pin but we will introduced the notion of iosets. This
> > notion involves that timings are guaranteed only in one ioset. That's
> > why we can't mix signals from several iosets because. On the controller side
> > we can do all we want so I would like to use groups as a software protection.
> 
> What does happen when you mix signals of different iosets? It won't work
> so the developer will change it. What do you need the software
> protection for?
> 

I can't say it won't work, it could work in some cases. My fear is to
have some support cases because of this. It seems easy to spot this kind
of issue but experience tell us that we can loose time for this kind of
"stupid" error.

> > i2c0 {
> > 	i2c0_ioset0_grp {
> > 		pins  = <PINMUX_PIN(71, 4),
> > 			 PINMUX_PIN(72, 4)>;
> > 	};
> > 	i2c0_ioset1_grp {
> > 		pins  = <PINMUX_PIN(91, 6),
> > 			 PINMUX_PIN(92, 6)>;
> > 	};
> > 	i2c0_ioset2_grp {
> > 		pins  = <PINMUX_PIN(108, 1),
> > 			 PINMUX_PIN(109, 1)>;
> > 	};
> > };
> > 
> > 
> > > This would mean you have to carry a lot of groups in each dtsi from
> > > which only a small fraction is used. We did that on i.MX but no longer
> > > do this since the dtbs get very big.
> > 
> > Yes I would like to put it in the dtsi file. For sure we will have big
> > dtbs. If we put that in the driver we will increase the size of our
> > kernel. If every one does this, we will have a lot of unused code for a
> > multiplatform kernel. Where is the best place to put this?
> 
> If you put only the needed ones in each dts then you won't have to carry
> all unused groups in all dtbs.
> 
> I think the idea you follow here is that you want to put all possible
> combinations (mci0_ioset0_4bit_grp, mci0_ioset1_1bit_grp) into the dtsi
> and let the board dts just pick one. For i.MX we realized that this does
> not scale because the number of possibilities is too high. Very often
> new boards had to introduce new pingroups in the dtsi because they did
> not yet exist or even did exist in a slightly different configuration.
> Out of tree dts files were painful because they often caused merge
> conflicts in the dtsi files. Also there there was the risk that two new
> groups were introduced with the same name but different meanings because
> someone else was faster (and chose the next free number in the 'ioset'
> counter).
> We removed the groups in the dtsi files altogether and put them into
> each board dts individually. This also made it easy to add the
> additional pins (Write protect, card detect) to the same groups.
>

It validates the decision to not put group declaration in the
driver. If we put all the configurations, it will add a big amount of
code. If we put only the configurations we use, the user will have to
compile a new kernel to add the ones he needs.

Well, maybe you're right about putting the groups in the dts files. I've
to think more about this solution.

> > > > - A subnode for these definitions in order to not parse the whole
> > > >   pinctrl node to retrieve groups and functions.
> > > > - Using node names as function and group names.
> > > > - Can we get generic properties to define the groups? Of course a 'pins'
> > > >   property is mandatory. In my case I will need an extra one to tell the
> > > >   controller how to mux the pins (a same pin can have up to 7 muxing
> > > >   possibilities).
> > > 
> > > Did you have a look at the RFC I sent for these kind of controllers [1] and
> > > the final result for the Mediatek driver currently in Linux-next [2]?.
> > > 
> > > The binding has both the config and the pins in a single node and thus
> > > is very compact.
> > > 
> > 
> > Thanks for the links. Well I had a look to them and now I am a bit
> > lost...
> > 
> > I agree with this binding but it involves to get rid of
> > pinconf_generic_dt_node_to_map_all, isn't it? What do group and function
> > become? It seems these concepts have disappeared.
> 
> The binding I suggested changes nothing with pinconf, only the pinmux
> information is added to the same node. You can still call
> pinconf_generic_dt_node_to_map_all() on the nodes, it will simply ignore
> the pinmux information. You would have to handle them separately (or
> write some generic helper if you like)

Yes, I can still use it. What I mean is there is no generic helper at the
moment to get both pinmux (excepting using the function property) and
pinconf information. Is it planned to have something generic for the
pinmux property? I see MTK_GET_PIN_NO and MTK_GET_PIN_FUNC macros, on my side,
I think it will feet my needs. Maybe we only need to remove the MTK
prefix and put these macros in another header.

In the mediatek driver, I have also noticed that we have a group for
each pin. I have the feeling that the concept of groups disappear, isn't
it?


Ludovic