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

[Greg Kroah-Hartman <gregkh@linuxfoundation.org>]

On Wed, Feb 07, 2024 at 05:32:38PM +0100, Bartosz Golaszewski wrote:
> On Fri, Feb 2, 2024 at 11:02 AM Bartosz Golaszewski <brgl@bgdev.pl> wrote:
> >
> > On Fri, Feb 2, 2024 at 1:03 AM Bjorn Andersson <andersson@kernel.org> wrote:
> > >
> >
> > [snip]
> >
> > > > >
> > > > > I believe I missed this part of the discussion, why does this need to be
> > > > > a platform_device? What does the platform_bus bring that can't be
> > > > > provided by some other bus?
> > > > >
> > > >
> > > > Does it need to be a platform_device? No, of course not. Does it make
> > > > sense for it to be one? Yes, for two reasons:
> > > >
> > > > 1. The ATH11K WLAN module is represented on the device tree like a
> > > > platform device, we know it's always there and it consumes regulators
> > > > from another platform device. The fact it uses PCIe doesn't change the
> > > > fact that it is logically a platform device.
> > >
> > > Are you referring to the ath11k SNOC (firmware running on co-processor
> > > in the SoC) variant?
> > >
> > > Afaict the PCIe-attached ath11k is not represented as a platform_device
> > > in DeviceTree.
> > >
> >
> > My bad. In RB5 it isn't (yet - I want to add it in the power
> > sequencing series). It is in X13s though[1].
> >
> > > Said platform_device is also not a child under the PCIe bus, so this
> > > would be a different platform_device...
> > >
> >
> > It's the child of the PCIe port node but there's a reason for it to
> > have the `compatible` property. It's because it's an entity of whose
> > existence we are aware before the system boots.
> >
> > > > 2. The platform bus already provides us with the entire infrastructure
> > > > that we'd now need to duplicate (possibly adding bugs) in order to
> > > > introduce a "power sequencing bus".
> > > >
> > >
> > > This is a perfectly reasonable desire. Look at our PMICs, they are full
> > > of platform_devices. But through the years it's been said many times,
> > > that this is not a valid or good reason for using platform_devices, and
> > > as a result we have e.g. auxiliary bus.
> > >
> >
> > Ok, so I cannot find this information anywhere (nor any example). Do
> > you happen to know if the auxiliary bus offers any software node
> > integration so that the `compatible` property from DT can get
> > seamlessly mapped to auxiliary device IDs?
> >
> 
> So I was just trying to port this to using the auxiliary bus, only to
> find myself literally reimplementing functions from
> drivers/of/device.c. I have a feeling that this is simply wrong. If
> we're instantiating devices well defined on the device-tree then IMO
> we *should* make them platform devices. Anything else and we'll be
> reimplementing drivers/of/ because we will need to parse the device
> nodes, check the compatible, match it against drivers etc. Things that
> are already implemented for the platform bus and of_* APIs.
> 
> Greg: Could you chime in and confirm that it's alright to use the
> platform bus here? Or maybe there is some infrastructure to create
> auxiliary devices from software nodes?

Note, I HATE the use of the platform bus here, but I don't have a better
suggestion.

I'd love for the auxbus to work, and if you can create that from
software nodes, all the better!  But I don't think that's possible just
yet, and you would end up implementing all the same stuff that the
platform bus has today for this functionality, so I doubt it would be
worth it.

thanks,

greg k-h