It works at a device level, so it doesn't know about resources.  The
only information it has is of the "this device may depend on that
other device" type and it uses that information to figure out a usable
probe ordering for drivers.

And that simplification is the problem. A phandle does not point to a
device, it points to a resource of a device. It should really be doing
what the driver would do, follow the phandle to the resource and see
if it exists yet. If it does not exist then yes it can defer the
probe. If the resource does exist, allow the driver to probe.

Also if the probe has already started, it may still return
-EPROBE_DEFER at any time in theory

Sure it can, and does. And any driver which is not broken will
unregister its resources on the error path. And that causes users of
the resources to release them. It all nicely unravels, and then tries
again later. This all works, it is what these drivers do.

However, making children wait for their parents to complete probing is
generally artificial, especially in the cases when the children are
registered by the parent's driver.  So waiting should be an exception
in these cases, not a rule.

So are you suggesting that fw_devlink core needs to change, recognise
the dependency on a parent, and allow the probe? Works for me. Gets us
back to before fw_devlink.

    Andrew