Re: [PATCH v10 3/4] dtc: Plugin and fixup support

Hi Phil,

On Nov 29, 2016, at 14:11 , Phil Elwell [off-list ref] wrote:

Pantelis,

On 29/11/2016 10:55, Pantelis Antoniou wrote:

quoted

Manually adding symbols by targeting __symbols__ is just bad. There is absolutely
no guarantee that the symbol/fixup node(s) will still be there in following iterations
of the patches.

Remember that this is now part of the Linux kernel - it isn't something
you can just change at will.

Since I’m the one that put it there, I’m sure I have a little bit of leverage.

quoted

I am thinking of parsing them, recording the information in kernel structures and then
deleting them altogether.

quoted

How does your patch handle duplicate symbols?

It doesn’t. Having duplicate global symbols is bad. 

It appears you want scoping rules instead. Care to paste a concrete example?

Concrete non-trivial examples are hard to come by. There are some simple
cases where we've attached labels to __overlay__ nodes so that the
contents can be patched by our overlay parameter mechanism - they could
just be given unique names instead of just "frag0", "frag1" etc. I'm
more concerned about parameterised macro-expanded overlays.

Consider an overlay that defines a CAN controller on an SPI bus. We
currently have two such overlays in the RPi tree, one for SPI 0.0 and
one for SPI 0.1. Here's an extract from one of them:

   /* the interrupt pin of the can-controller */
   fragment@2 {
       target = <&gpio>;
       __overlay__ {
           can0_pins: can0_pins {
               brcm,pins = <25>;
               brcm,function = <0>; /* input */
           };
       };
   };
...
   fragment@4 {
       target = <&spi0>;
       __overlay__ {
           /* needed to avoid dtc warning */
           #address-cells = <1>;
           #size-cells = <0>;
           can0: mcp2515@0 {
               reg = <0>;
               compatible = "microchip,mcp2515";
               pinctrl-names = "default";
               pinctrl-0 = <&can0_pins>;
               spi-max-frequency = <10000000>;
               interrupt-parent = <&gpio>;
               interrupts = <25 0x2>;
               clocks = <&can0_osc>;
           };
       };
   };

One day I'd like to merge these into a single parameterised version that
could target any CS line on any SPI controller. This requires that any
created node names are unique with the scope of the parent ("mcp2515@0",
"can0_pins"), and that the name of the target label (spi0) is patched to
select the correct SPI bus. Our existing, limited overlay parameter
mechanism uses labels to identify properties to patch:

       spimaxfrequency = <&can0>,"spi-max-frequency:0";

(I would have attached labels to the properties themselves, but that
doesn't seem to work, contrary to the ePAPR spec.)

If the labels that locate the node, property and label names to change
also themselves have to be made unique then that adds an extra level of
complexity.

The parameter application is a pre-processing step before the overlay is
merged, so there is nothing preventing me from filtering the symbols
node before passing it on based on rules of my own choosing, but I
wanted to make more people aware of this change.

First note is that this is exactly what the portable connector is supposed to
do; abstract away the SoC differences.

Second note is that it’s not the overlay application that’s having problems, it’s
your parameter patching method.

The 'spimaxfrequency = <&can0>,"spi-max-frequency:0”’ form could more easily be
done by targeting aliases instead of node labels.

I.e. You can apply the overlay, set an alias to the node and instead of referencing
the label, reference the alias.

Again, this is a stop-gap until the portable connector is done, but what I take out
of this is the need for a parameterization step in which an overlay is modified before
it is applied according to an external parameter.

Phil

Regards

— Pantelis