On Fri, Jul 4, 2025 at 3:57 AM Herve Codina <herve.codina@xxxxxxxxxxx> wrote:
>
> Hi Rob,
>
> On Thu, 3 Jul 2025 09:33:02 +0200
> Herve Codina <herve.codina@xxxxxxxxxxx> wrote:
>
> > Hi Rob,
> >
> > On Fri, 27 Jun 2025 10:52:00 -0500
> > Rob Herring <robh@xxxxxxxxxx> wrote:
> >
> > > On Fri, Jun 13, 2025 at 03:47:45PM +0200, Herve Codina wrote:
> > > > The simple-pm-bus driver handles several simple busses. When it is used
> > > > with busses other than a compatible "simple-pm-bus", it doesn't populate
> > > > its child devices during its probe.
> > > >
> > > > This confuses fw_devlink and results in wrong or missing devlinks.
> > > >
> > > > Once a driver is bound to a device and the probe() has been called,
> > > > device_links_driver_bound() is called.
> > > >
> > > > This function performs operation based on the following assumption:
> > > > If a child firmware node of the bound device is not added as a
> > > > device, it will never be added.
> > > >
> > > > Among operations done on fw_devlinks of those "never be added" devices,
> > > > device_links_driver_bound() changes their supplier.
> > > >
> > > > With devices attached to a simple-bus compatible device, this change
> > > > leads to wrong devlinks where supplier of devices points to the device
> > > > parent (i.e. simple-bus compatible device) instead of the device itself
> > > > (i.e. simple-bus child).
> > > >
> > > > When the device attached to the simple-bus is removed, because devlinks
> > > > are not correct, its consumers are not removed first.
> > > >
> > > > In order to have correct devlinks created, make the simple-pm-bus driver
> > > > compliant with the devlink assumption and create its child devices
> > > > during its probe.
> > >
> > > IIRC, skipping child nodes was because there were problems with
> > > letting the driver handle 'simple-bus'. How does this avoid that now?
> >
> > I don't know about the specific issues related to those problems. Do you
> > have some pointers about them?
> >
> > >
> > > The root of_platform_populate() that created the simple-bus device that
> > > gets us to the probe here will continue descending into child nodes.
> > > Meanwhile, the probe here is also descending into those same child
> > > nodes. Best case, that's just redundant. Worst case, won't you still
> > > have the same problem if the first of_platform_populate() creates the
> > > devices first?
> > >
> >
> > Maybe we could simply avoid of_platform_populate() to be recursive when a
> > device populate by of_platform_populate() is one of devices handled by
> > the simple-bus driver and let the simple-bus driver do the job.
> >
> > of_platform_populate will handle the first level. It will populate children
> > of the node given to of_platform_populate() and the children of those
> > children will be populate by the simple-bus driver.
> >
> > I could try a modification in that way. Do you think it could be a correct
> > solution?
> >
>
> I have started to look at this solution and it's going to be more complex
> than than I thought.
>
> Many MFD drivers uses a compatible of this kind (the same exist for bus
> driver with "simple-bus"):
> compatible = "foo,bar", "simple-mfd";
>
> Usually the last compatible string ("simple-mfd" here) is a last fallback
> and the first string is the more specific one.
>
> In the problematic case, "foo,bar" has a specific driver and the driver
> performs some operations at probe() but doesn't call of_platform_populate()
> and relies on the core to do the device creations (recursively) based on
> the "simple,mfd" string present in the compatible property.
>
> Some other calls of_platform_populate() in they probe (which I think is
> correct) and in that case, the child device creation can be done at two
> location: specific driver probe() and core.
>
> You pointed out that the core could create devices before the specific
> driver is probed. In that case, some of existing drivers calling
> of_platform_populate() are going to have issues.
>
> I can try to modify existing MFD and bus drivers (compatible fallback to
> simple-mfd, simple-bus, ...) in order to have them call of_platform_populate()
> in they probe() and after all problematic drivers are converted, the
> recursive creation of devices done in the core could be removed.
The problem is how does a bus driver know if there is a specific MFD
driver or not? It doesn't. The MFD driver could be a module and loaded
any time later. We'd really need some sort of unbind the generic
driver and re-bind to a more specific driver when and if that driver
appears. We could perhaps have a list of devices with a driver because
in theory that should be a short list as the (broken) promise of
simple-mfd is the child nodes have no dependency on the parent node
which implies the parent doesn't have a driver. The specific
compatible is there in case that assumption turns out wrong.
Rob
