On Mon, Nov 14, 2022 at 01:08:03PM -0500, Sean Anderson wrote:
> On 11/14/22 12:23, Vladimir Oltean wrote:
> > On Thu, Nov 10, 2022 at 11:56:15AM -0500, Sean Anderson wrote:
> >> these will probably be in device trees for a year before the kernel
> >> starts using them. But once that is done, we are free to require them.
> >
> > Sorry, you need to propose something that is not "we can break compatibility
> > with today's device trees one year from now".
>
> But only if the kernel gets updated and not the device tree. When can
> such a situation occur? Are we stuck with this for the next 10 years all
> because someone may have a device tree which they compiled in 2017, and
> *insist* on using the latest kernel with? Is this how you run your
> systems?
I'm a developer (and I work on other platforms than the ones you're
planning to break), so the answer to this question doesn't mean a thing.
> We don't get the device tree from firmware on this platform; usually it
> is bundled with the kernel in a FIT or loaded from the same disk
> partition as the kernel. I can imagine that they might not always be
> updated at exactly the same time, but this is nuts.
What does "this" platform mean exactly? There are many platforms to
which you've added compatible strings to keep things working assuming a
dtb update, many of them very old. And those to which you did are not by
far all that exist. There is no requirement that all platform device
trees are upstreamed to the Linux kernel.
> The original device tree is broken because it doesn't include compatible
> strings for devices on a generic bus. There's no way to fix that other
> than hard-coding the driver. This can be done for some buses, but this
> is an MDIO bus and we already assume devices without compatibles are
> PHYs.
Let's be clear about this. It's "broken" in the sense that you don't like
the way in which it works, not in the sense that it results in a system
that doesn't work. And not having a compatible string is just as broken
as it is for other devices with detectable device IDs, like Ethernet
PHYs in general, PCI devices, etc.
The way in which that works here, specifically, is that a generic PHY driver
is bound to the Lynx PCS devices, driver which does nothing since nobody
calls phy_attach_direct() to it. Then, using fwnode_mdio_find_device(),
you follow the pcsphy-handle and you get a reference to the mdio_device
(parent class of phy_device) object that resulted from the generic PHY
driver probing on the PCS, and you program the PCS to do what you want.
The PHY core does assume that mdio_devices without compatible strings
are phy_devices, but also makes exceptions (and warns about it) - see
commit ae461131960b ("of: of_mdio: Add a whitelist of PHY compatibilities.").
Maybe the reverse exception could also be made, and a warning for that
be added as well.
> In the next version of this series, I will include a compatibility
> function which can bind a driver automatically if one is missing when
> looking up a phy. But I would really like to have an exit strategy.
You'll have to get agreement from higher level maintainers than me that
the strategy "wait one year, break old device trees" is okay. Generally
we wouldn't have answers to this kind of questions that depend on whom
you ask. Otherwise.. we would all know whom to ask and whom not to ;)
Sadly I haven't found anything "official" in either Documentation/devicetree/usage-model.rst
or Documentation/process/submitting-patches.rst. Maybe I missed it?
I've added Arnd Bergmann for an ack, and also Marc Zyngier, not because
of any particular connection to what's being changed here, but because
I happen to know that he might have strong opinions on the topic :)
Full context here:
https://patchwork.kernel.org/project/netdevbpf/cover/20221103210650.2325784-1-sean.anderson@xxxxxxxx/
If I'm the only one opposing this, I guess I'll look elsewhere.
