On Thu, Aug 2, 2018 at 1:31 PM Stefan Agner [off-list ref] wrote:

A while back at least using those init lists were not well received even
for GPIO/pinctrl drivers:

https://lore.kernel.org/lkml/CACRpkdYk0zW12qNXgOstTLmdVDYacu0Un+8quTN+J_azOic7AA@mail.gmail.com/T/#mf0596982324a6489b5537b0531ac5aed60a316ba (local)

You shouldn't listen too much to that guy he's not trustworthy.

I still think we should make an exception for GPIO/pinctrl and use
earlier initcalls. Platform GPIO/pinctrl drivers provide basic
infrastructure often used by many other drivers, we want to have them
loaded early. It avoids unnecessary EPROBE_DEFER and hence probably even
boots faster.

When we have the pin control and GPIO at different initlevels it makes me
uneasy because I feel we have implicit init dependencies that seem more
than a little fragile.

My recent thinking has involved the component method used in DRM drivers
such as drivers/gpu/drm/vc4/vc4_drv.c where a few different component
subdrivers are linked together at bind time (not probe time!) into a master
component.

Rob was no big fan of this but the DRM people like it and I was thinking to
make a try at it.

This way we could at least probe and bind the pin control and GPIO drivers
at the *same* initlevel and express the dependencies between them
somewhat.

This should definitely go in, at least as a stop gap solution.

Agreed. (And patch applied.)

Yours,
Linus Walleij