[PATCH RFC net-next v2 11/12] net: dsa: Separate C22 and C45 MDIO bus transaction methods

[Vladimir Oltean]

On Tue, Jan 03, 2023 at 04:48:59PM +0100, Andrew Lunn wrote:
> > Since clause 45 PHYs are identified by the "ethernet-phy-ieee802.3-c45"
> > compatible string (otherwise they are C22), then a PHY which is not
> > described in the device tree can only be C22. So this is why
> > ds->slave_mii_bus only deals with clause 22 methods, and the true reason
> > behind the comment above.
> > 
> > But actually this premise is no longer true since Luiz' commit
> > fe7324b93222 ("net: dsa: OF-ware slave_mii_bus"), which introduced the
> > strange concept of an "OF-aware helper for internal PHYs which are not
> > described in the device tree". After his patch, it is possible to have
> > something like this:
> > 
> > 	ethernet-switch {
> > 		ethernet-ports {
> > 			port at 1 {
> > 				reg = <1>;
> > 			};
> > 		};
> > 
> > 		mdio {
> > 			ethernet-phy at 1 {
> > 				compatible = "ethernet-phy-ieee802.3-c45"
> > 				reg = <1>;
> > 			};
> > 		};
> > 	};
> > 
> > As you can see, this is a clause 45 internal PHY which lacks a
> > phy-handle, so its bus must be put in ds->slave_mii_bus in order for
> > dsa_slave_phy_connect() to see it without that phy-handle (based on the
> > port number matching with the PHY number). After Luiz' patch, this kind
> > of device tree is possible, and it invalidates the assumption about
> > ds->slave_mii_bus only driving C22 PHYs.
> 
> My memory is hazy, but i think at the time i wrote these patches,
> there was no DSA driver which made use of ds->slave_mii_bus with
> C45. So i took the short cut of only supporting C22.

Actually I believe that in v1 you did extend ds->ops with C45 methods,
but it's me who told you to remove them:
https://patchwork.kernel.org/project/netdevbpf/patch/20220508153049.427227-10-andrew@lunn.ch/#24852813

> 
> Those DSA drivers which do support C45 all register their bus directly
> with the MDIO core.

And rightfully so. IMHO, letting DSA allocate ds->slave_mii_bus out of
driver writer sheer convenience (a secondary purpose) should be deprecated,
unless the reason for using ds->slave_mii_bus is the lack of a phy-handle
(the primary purpose). It becomes that more confusing to have to extend
dsa_switch_ops with 2 more methods which serve the secondary purpose but
not the primary one.

> So Luiz patches may allow a C45 bus, but are there any drivers today
> actually using it?

I sent a private email to Luiz a few minutes ago asking him to confirm.