Re: [PATCH 4/9] PCI: create platform devices for child OF nodes of the port node

[Bartosz Golaszewski <brgl@bgdev.pl>]

On Wed, Jan 17, 2024 at 5:45 PM Greg Kroah-Hartman
<gregkh@linuxfoundation.org> wrote:
>
> On Wed, Jan 17, 2024 at 05:07:43PM +0100, Bartosz Golaszewski wrote:
> > From: Bartosz Golaszewski <bartosz.golaszewski@linaro.org>
> >
> > In order to introduce PCI power-sequencing, we need to create platform
> > devices for child nodes of the port node.
>
> Ick, why a platform device?  What is the parent of this device, a PCI
> device?  If so, then this can't be a platform device, as that's not what
> it is, it's something else so make it a device of that type,.
>

Greg,

This is literally what we agreed on at LPC. In fact: during one of the
hall track discussions I said that you typically NAK any attempts at
using the platform bus for "fake" devices but you responded that this
is what the USB on-board HUB does and while it's not pretty, this is
what we need to do.

Now as for the implementation, the way I see it we have two solutions:
either we introduce a fake, top-level PCI slot platform device device
that will reference the PCI host controller by phandle or we will live
with a secondary, "virtual" platform device for power sequencing that
is tied to the actual PCI device. The former requires us to add DT
bindings, add a totally fake DT node representing the "slot" which
doesn't really exist (and Krzysztof already expressed his negative
opinion of that) and then have code that will be more complex than it
needs to be. The latter allows us to not change DT at all (other than
adding regulators, clocks and GPIOs to already existing WLAN nodes),
reuse the existing parent-child relationship between the port node and
the instantiated platform device as well as result in simpler code.

Given that DT needs to be stable while the underlying C code can
freely change if we find a better solution, I think that the second
option is a no-brainer here.

> > They will get matched against
> > the pwrseq drivers (if one exists) and then the actual PCI device will
> > reuse the node once it's detected on the bus.
>
> Reuse it how?
>

By consuming the same DT node using device_set_of_node_from_dev() when
the PCI device is registered. This ensures we don't try to bind
pinctrl twice etc.

> > Signed-off-by: Bartosz Golaszewski <bartosz.golaszewski@linaro.org>
> > ---
> >  drivers/pci/bus.c    | 9 ++++++++-
> >  drivers/pci/remove.c | 3 ++-
> >  2 files changed, 10 insertions(+), 2 deletions(-)
> >
> > diff --git a/drivers/pci/bus.c b/drivers/pci/bus.c
> > index 9c2137dae429..8ab07f711834 100644
> > --- a/drivers/pci/bus.c
> > +++ b/drivers/pci/bus.c
> > @@ -12,6 +12,7 @@
> >  #include <linux/errno.h>
> >  #include <linux/ioport.h>
> >  #include <linux/of.h>
> > +#include <linux/of_platform.h>
> >  #include <linux/proc_fs.h>
> >  #include <linux/slab.h>
> >
> > @@ -342,8 +343,14 @@ void pci_bus_add_device(struct pci_dev *dev)
> >        */
> >       pcibios_bus_add_device(dev);
> >       pci_fixup_device(pci_fixup_final, dev);
> > -     if (pci_is_bridge(dev))
> > +     if (pci_is_bridge(dev)) {
> >               of_pci_make_dev_node(dev);
> > +             retval = of_platform_populate(dev->dev.of_node, NULL, NULL,
> > +                                           &dev->dev);
>
> So this is a pci bridge device, not a platform device, please don't do
> this, make it a real device of a new type.
>

Not sure what you mean. Are you suggesting adding a new bus? Or do we
already have a concept of PCI bridge devices in the kernel?

Bartosz

> thanks,
>
> greg k-h