On Wed, Sep 07, 2016 at 04:44:58PM -0700, Stephen Boyd wrote:

Quoting David Gibson (2016-08-30 16:55:23)

quoted

On Mon, Aug 29, 2016 at 07:07:49PM -0700, Stephen Boyd wrote:

quoted

Quoting David Gibson (2016-08-29 06:45:11)

quoted

So, combining those thoughts together, I'm thinking dtc format for
something connecting to two different widget sockets (pretty much the
worst case) would look something like:

/plugin/ foo,widget-socket {
};

/plugin/ foo,widget-socket {
        realias {
                i2c-b = "i2c";
                intc-b = "intc";
                mmio-b = "mmio";
        };
};

&i2c {
        .. devices on the i2c from the first plug ..
};

&i2c-b {
        .. devices on the i2c from the second plug ..
};

Obviously we'd also need to devise an encoding for this to compile
into, since the one I proposed previously won't work in this case.

I suppose we can distribute the realias nodes when we compile the plugin
into overlay fragments. The socket matching is a little vague though.
How would we know which socket to apply to when we have two identical
looking sockets? I'm thinking we could put some of that information into
the fragment itself.

So, my assumption in this example was that the plugin could plug into
*any* two widget sockets.  If it needs to connect to specific ones,
then pretty much by definition, the sockets aren't really of
indistinguishable type.

Ah ok. It wasn't clear on how we target where the extension board
connects into the full tree. It sounds like the decision of where to
plug in the extension board falls to the overlay manager then?

Yes.  The whole point of connectors is that they're interchangeable,
so neither the "plug" side nor the "socket" side should specify which
instance of the other part they're connected to.

So to resolve the whole tree in this case, extra information would be
needed.  Ideally it would come from probing (even something ad-hoc
like an ID register), failing that it would need to be admin
configured.

quoted

quoted

/{
      compatible = "foo,whirlgig-widget";

      fragment@0 { /* corresponds to i2c in example above */
              target-socket = "foo,widget-socket-a";
              target-alias = "i2c";
              __overlay__ {
                      ....
              };
      };

      fragment@1 { /* corresponds to i2c-b in example above */
              target-socket = "foo,widget-socket-b";
              target-alias = "i2c";
              __overlay__ {
                      ...
              };
      };
};

We don't need any new construct here.  In this case the sockets aren't
100% compatible, which we can notate with
        compatible = "widget-socket-a", "widget-socket";
In the base board.

Devices which can plug into any widget socket will use target-socket =
"widget-socket", those which require a specific one (including
requiring both) can specifically list "widget-socket-a" and/or
"widget-socket-b".

Agreed. How do we express the realias node in the overlay format though?
I was trying to come up with some way to make that work by
redistributing the alias to the fragments but that doesn't work so
well.

So, I'm thinking of the overlay dt info in 3 pieces

1) Global metadata

  Not much here yet, but it seems like it could be useful.

2) Per-connector metadata

  This specifies what sort of "socket" each "plug" on the overlay
  needs to connect to, and any other information we need that's
  specific to a particular "plug"

3) DT fragments

  The actual device info applied once we've resolved all the aliases
  and whatnot.

The realias information would need to be in part (2)

How we encode those pieces into the final overlay dt is pretty close
to an arbitrary choice.

quoted

quoted

If we have two identical connectors maybe we'll have to enforce that the
connectors have some more specific unique compatible string so that we
can match up the right socket. But I don't see how we can require that
the overlays know this detail if they only care about one socket and
could go into either one of them. In that case we should have the loader
ask the user which socket they connected this extension board to?

I was also thinking it would be better to leave the gpio-map and
interrupt-map properties at the connector level. For example:

      widget1 {
              compatible = "foo,widget-socket";
              interrupt-map-mask = <0xffffffff>;
              interrupt-map = <0 &intc 7 0>,
                              <1 &intc 8 0>;
      };

That could work - but we should (and implicitly, do) support either
way.  Using subnodes might be useful for particularly complex irq or
gpio mappings.

Sounds good.

quoted

quoted

and then we could put a label on the plugin/expansion syntax so we can
reference the connector as a whole:

      /plugin/ connector: foo,widget-socket {
              compatible = "foo,whirlgig-widget";
      };

      &i2c {
              device@40 {
                      interrupt-parent = <&connector>;
                      interrupts = <1>;
              };
      };

I also thought about making another alias inside the connector node to
point to itself, but that fails when you get into the situation of two
connectors and collisions, unless you rename them of course. It felt
better to leave that choice to the overlay though.

In conclusion, I see a few topics/patterns emerging:

1) Expose phandles through the connectors in some way that allows us to
   limit what the plugin/expansion boards can modify or use

Yes, I definitely think we want that.

quoted

2) Have some flexible syntax to remap cell sizes from the baseboard
   through the connector to provide a consistent connector size (i.e.
   remap interrupts and gpios from multiple sources, etc. into a fixed
   number of cells)

I don't think we need any new constructs here.  If there are
mismatches we can put dummy bridges with appropriate ranges properties
on one side or the other.

The only thing I see that might want some help is that the connector
type should certainly imply a specific set of cell widths for all the
included buses.  So possibly we should supply some stuff to help
enforce that.

I'm specifically thinking that anything that has a #<name>-cells
associated with it (#gpio-cells, #clock-cells, #power-domain-cells,
etc.) would need to have an associated <name>-map property and
associated parsing code so that we can make a consistent cell width for
the connector. If we have existing ways to make this work, e.g.
interrupt-map or ranges, then we don't need.

Yes, I agree.  All the "interrupt tree like" things need interrupt
tree like nexus nodes, with a map and map-mask.

However, we also need some way of ensuring that the #address-cells and
#size-cells for any buses we expose match properly.

Or maybe we need to invent something like a cross between 'ranges' and
'interrupt-map' that can remap ranges on one address space to the
address space attached to a different node.

quoted

quoted

3) Allow plugin/expansion boards to use multiple connectors from the
   baseboard in a consistent way

Seems reasonable.

quoted

4) Attempt to maintain almost all of the current overlay syntax with
   syntactic sugar

I'm not really sure what you mean by that.

I mean that we're not really trying to change the general structure of
the DT overlay syntax. At least from what I can tell we're planning to
convert this /plugin/ format into a DT overlay that gets compiled into
binary all inside DTC.

Roughly speaking, yes.  However I think we should take the opportunity
to address design flaws in the current overlay encoding as they come
up (e.g. the fact that it uses phandles which *always* have to be
resolved - better to directly use an alias / symbol / label).

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson