Re: [PATCH phy 13/14] dt-bindings: phy: lynx-28g: add compatible strings per SerDes and instantiation

[Vladimir Oltean <vladimir.oltean@nxp.com>]

On Thu, Sep 04, 2025 at 08:22:16PM +0100, Conor Dooley wrote:
> On Thu, Sep 04, 2025 at 06:44:01PM +0300, Vladimir Oltean wrote:
> > Going by the generic "fsl,lynx-28g" compatible string and expecting all
> > SerDes instantiations on all SoCs to use it was a mistake.
> > 
> > They all share the same register map, sure, but the number of protocol
> > converters and lanes which are instantiated differs in a way that isn't
> > detectable by software. So distinguish them by compatible strings.
> > At the same time, keep "fsl,lynx-28g" as backup.
> 
> Why keep the backup? Doesn't sound like you can use it for anything,
> unless there's some minimum set of capabilities that all devices
> support. If that's not the case, should it not just be marked deprecated
> or removed entirely?

To be honest, I could use some guidance on the best way to handle this.

When I had written this patch downstream, lx2160a.dtsi only had serdes_1
defined, as "fsl,lynx-28g", and this patch made more sense. Keep
"fsl,lynx-28g" as a synonym for "fsl,lx2160a-serdes1", so that new
device trees still work with old kernels (as is sometimes needed during
'git bisect', etc), for some definition of the word "work" (more often
than not, unsatisfactory - for example, fw_devlink blocks probing the PHY
consumer driver if the PHY driver doesn't exist, but the 'phys' property
exists in the device tree).

Unbeknownst to me, commit 2f2900176b44 ("arm64: dts: lx2160a: describe
the SerDes block #2") came and defined the second SerDes also with
"fsl,lynx-28g".

The second SerDes is less capable than the first one, so the same
developer then started battling with the fact that the driver doesn't
know that serdes_2 doesn't support some protocols, and wrote some
patches like 9bef84d30f1f ("phy: lynx-28g: check return value when
calling lynx_28g_pll_get"), which in all likelihood could have been
avoided using a specific compatible string. The lynx_info ::
lane_supports_mode() method from patch 14/14 is supposed to say what is
supported per SerDes and what not.

In terms of implementation, what does "deprecating" the "fsl,lynx-28g"
compatible string mean, compared to removing it entirely? Would there be
any remaining driver support for it? Should I compute the common set of
capabilities between SerDes #1 and #2, and only support that? What
impact would this have upon old device trees? Is it acceptable to just
remove support for them?