Re: [RFC 3/9] PCI/portdrv: create platform devices for child OF nodes

[Dan Williams]

[ add Terry ]


Lukas Wunner wrote:
> On Wed, Jan 10, 2024 at 01:55:18PM +0100, Bartosz Golaszewski wrote:
> > On Tue, Jan 9, 2024 at 3:43???PM Lukas Wunner <lukas@xxxxxxxxx> wrote:
> > > On Thu, Jan 04, 2024 at 02:01:17PM +0100, Bartosz Golaszewski wrote:
> > > > In order to introduce PCIe power-sequencing, we need to create platform
> > > > devices for child nodes of the port driver node. They will get matched
> > > > against the pwrseq drivers (if one exists) and then the actuak PCIe
> > > > device will reuse the node once it's detected on the bus.
> > > [...]
> > > > --- a/drivers/pci/pcie/portdrv.c
> > > > +++ b/drivers/pci/pcie/portdrv.c
> > > > @@ -715,7 +716,7 @@ static int pcie_portdrv_probe(struct pci_dev *dev,
> > > >               pm_runtime_allow(&dev->dev);
> > > >       }
> > > >
> > > > -     return 0;
> > > > +     return devm_of_platform_populate(&dev->dev);
> > > >  }
> > >
> > > I think this belongs in of_pci_make_dev_node(), portdrv seems totally
> > > the wrong place.  Note that you're currently calling this for RCECs
> > > (Root Complex Event Collectors) as well, which is likely not what
> > > you want.
> > >
> > 
> > of_pci_make_dev_node() is only called when the relevant PCI device is
> > instantiated which doesn't happen until it's powered-up and scanned -
> > precisely the problem I'm trying to address.
> 
> No, of_pci_make_dev_node() is called *before* device_attach(),
> i.e. before portdrv has even probed.  So it seems this should
> work perfectly well for your use case.
> 
> 
> > > devm functions can't be used in the PCI core, so symmetrically call
> > > of_platform_unpopulate() from of_pci_remove_node().
> > 
> > I don't doubt what you're saying is true (I've seen worse things) but
> > this is the probe() callback of a driver using the driver model. Why
> > wouldn't devres work?
> 
> The long term plan is to move the functionality in portdrv to
> the PCI core.  Because devm functions can't be used in the PCI
> core, adding new ones to portdrv will *add* a new roadblock to
> migrating portdrv to the PCI core.  In other words, it makes
> future maintenance more difficult.
> 
> Generally, only PCIe port services which share the same interrupt
> (hotplug, PME, bandwith notification, flit error counter, ...)
> need to live in portdrv.  Arbitrary other stuff should not be
> shoehorned into portdrv.

I came here to say the same thing. It is already the case that portdrv
is not a good model to build new functionality upon [1], and PCI core
enlightenment should be considered first.

The portdrv model is impeding Terry's CXL Port error handling effort, so
I am on the lookout for portdrv growing new entanglements to unwind
later.

[1]: http://lore.kernel.org/r/20221025232535.GA579167@bhelgaas