On Thu, 2020-04-02 at 14:38 -0500, Bjorn Helgaas wrote:
> [+cc Rob for DT platform device dependency question]
> 
> On Thu, Apr 02, 2020 at 04:27:23PM +0200, Nicolas Saenz Julienne wrote:

[...]

> > Sorry it wasn't clear enough, I'll redo this comment. Also note that
> > the PCIe bus and the XHCI chip are hardwired, so that's the only
> > device that'll ever be available on the bus.
> > 
> > VIA805's XHCI firmware has to be loaded trough RPi's firmware
> > mailbox in between the PCIe bus probe and the subsequent USB probe.
> > Note that a PCI reset clears the firmware. The only mechanism
> > available in between the two operations are PCI Fixups. These are
> > limited in their own way, as I can't return -EPROBE_DEFER if the
> > firmware interface isn't available yet. Hence the need for an
> > explicit dependency between pcie-brcmstb and raspberrypi's firmware
> > mailbox device.
> > 
> > Your concern here showcases this series' limitations. From a high
> > level perspective it's not clear to me who should be responsible for
> > downloading the firmware. 
> 
> I think it's fairly common for drivers to download firmware to their
> devices.  I guess there's not really any need to download the firmware
> until a driver wants to use the device, right?
> 
> > And I get the feeling I'm abusing PCI fixups. I haven't found any
> > smart way to deal with this three way dependency of
> > platform/non-platform devices.
> 
> So IIUC, the three-way dependency involves:
> 
>   1) brcm_pcie_probe(), which initialize the PCI host controller
>   platform device, enumerates PCI devices, and makes them available
>   for driver binding,

Yes, and also resets the PCI bus, which will clear VL805's firmware (the XHCI
chip).

>   2) the firmware mailbox initialization (maybe
>   rpi_firmware_probe()?),
>
>   3) quirk_usb_early_handoff(), which downloads firmware to the VL805
>   PCI USB adapter via rpi_firmware_init_vl805(), which uses the
>   firmware mailbox?

And yes, that's the general idea.

> Is there some way to express a dependency between
> "raspberrypi,bcm2835-firmware" (the platform device claimed by
> rpi_firmware_probe() and "brcm,bcm2711-pcie"?  If we could ensure that
> rpi_firmware_probe() runs before brcm_pcie_probe(), would that solve
> part of this?

That's ultimately what this patch tries to achieve. If there was a way to
offload it to DT it would be way nicer.

Regards,
Nicolas