On 10/03/2022 18:32, Ioana Ciornei wrote:

On Thu, Mar 10, 2022 at 05:47:31PM +0100, Krzysztof Kozlowski wrote:

quoted

On 10/03/2022 15:51, Ioana Ciornei wrote:

quoted

Describe the "fsl,lynx-28g" compatible used by the Lynx 28G SerDes PHY
driver on Layerscape based SoCs.

The message is a bit misleading, because it suggests you add only
compatible to existing bindings. Instead please look at the git log how
people usually describe it in subject and message.

Sure, I can change the title and commit message.

quoted

quoted

+patternProperties:
+  '^phy@[0-9a-f]$':
+    type: object
+    properties:
+      reg:
+        description:
+          Number of the SerDes lane.
+        minimum: 0
+        maximum: 7
+
+      "#phy-cells":
+        const: 0

Why do you need all these children? You just enumerated them, without
statuses, resources or any properties. This should be rather just index
of lynx-28g phy.

I am just describing each lane of the SerDes block so that each ethernet
dts node references it directly.

Instead, phy user should reference phy device node and phy ID. Just like
we do for other providers (everything with #xxxxx-cells).

Since I am new to the generic PHY infrastructure I was using the COMPHY
for the Marvell MVEBU SoCs (phy-mvebu-comphy.txt) as a loose example.

I don't know it but it might not be the best example... Just because we
have already some solution it does not mean it is good. :)

Each lane there is described as a different child node as well. The only
difference from the COMPHY is that Lynx 28G does not need #phy-cells =
<1> to reference the input port, we just use '#phy-cells = <0>' on each
lane.

What is wrong with this approach? Or better, is there an easier way to
do this?

Because the nodes look artificial. It looks like you have nodes only
differentiate by index. As I said before - there are no other properties
in these nodes.

Imagine now a clock provider with 500 clocks like this...

The easier approach, especially since you have a shared registers, is to
use phy-cells = 1, without artificial nodes just to pass the index.

It would be entirely different if you actually had any properties in the
children. IOW, if these were actually some blocks with their own
characteristics and programming model.

Best regards,
Krzysztof