[RFC PATCH 1/4] dt: early platform devices support

[Grant Likely]

On Sat, Jun 25, 2011 at 12:11:42PM +0100, Marc Zyngier wrote:
> On Fri, 24 Jun 2011 22:49:56 -0600
> Grant Likely <grant.likely at secretlab.ca> wrote:
> 
> > [cc'ing linux kernel since I discuss driver core issues]
> > [cc'ing greg and kay, 'cause I've included a patch I'd like you to
> > look at (see end of email)]
> > 
> > On Fri, Jun 24, 2011 at 8:10 AM, Marc Zyngier <marc.zyngier at arm.com> wrote:
> > > Add support for populating early platform devices from the device
> > > tree, by walking the tree and adding nodes whose 'compatible'
> > > property matches the 'class' string passed as a parameter.
> > >
> > > This allows devices to be probed long before the whole device
> > > infrastructure is available.
> > >
> > > Signed-off-by: Marc Zyngier <marc.zyngier at arm.com>
> > > ---
> > >  drivers/of/platform.c       |   26 ++++++++++++++++++++++++++
> > >  include/linux/of_platform.h |    1 +
> > >  2 files changed, 27 insertions(+), 0 deletions(-)
> > >
> > > diff --git a/drivers/of/platform.c b/drivers/of/platform.c
> > > index e75af39..2a323ee 100644
> > > --- a/drivers/of/platform.c
> > > +++ b/drivers/of/platform.c
> > > @@ -458,4 +458,30 @@ int of_platform_populate(struct device_node *root,
> > >        of_node_put(root);
> > >        return rc;
> > >  }
> > > +
> > > +/**
> > > + * of_early_platform_populate() - Populate early platform devices from DT
> > > + * @class: string to compare to the 'compatible' attributes
> > > + *
> > > + * This function walks the device tree and register devices whose
> > > + * 'compatible' property matches the 'class' parameter.
> > > + *
> > > + * Returns 0 on success, < 0 on failure.
> > > + */
> > > +int of_early_platform_populate(const char *class)
> > > +{
> > > +       struct platform_device *pdev;
> > > +       struct device_node *np = NULL;
> > > +
> > > +       while ((np = of_find_compatible_node(np, NULL, class))) {
> > 
> > for_each_compatible_node()
> > 
> > > +               pdev = of_device_alloc(np, NULL, NULL);
> > > +               if (!pdev)
> > > +                       return -ENOMEM;
> > > +               list_del(&pdev->dev.devres_head);
> > > +               memset(&pdev->dev.devres_head, 0, sizeof(pdev->dev.devres_head));
> > > +               early_platform_add_devices(&pdev, 1);
> > 
> > I'm not sure this will work reliably.  The of_platform semantics are
> > still slightly different from 'regular' platform devices (which I do
> > intend to fix though), so it may cause problems with how the device
> > name gets assigned.  I'd need to dig into it though.
> > 
> > > +       }
> > > +
> > > +       return 0;
> > > +}
> > 
> > Hmmm, I'd really rather not go down the path of having different
> > 'classes' of devices that need to be registered at different times.
> > There just isn't a really good generic way to know which devices
> > should be the 'early' ones, it gets hairy with regard to making sure
> > devices don't get registered twice, and it doesn't actually solve the
> > problem that there is no way to handle dependencies between devices in
> > the Linux device model.
> > 
> > What I /want/ to do is allow drivers to defer initialization at .probe
> > time if some of the resources it needs aren't yet available.
> > Unfortunately, that doesn't work so well for interrupt controllers
> > since irq numbers are expected to be correctly populated in the
> > platform device resource table.  What complicates things further is
> > that most gpio controllers are doubling as irq controllers nowdays,
> > and those are typically modelled with platform devices themselves.
> 
> While I totally agree with all the above, this patch tries to address a
> slightly different problem. On top of device/device dependencies, there
> is also a number of implicit dependencies.
> 
> In the case I'm currently trying to solve, the kernel expects a timer
> to be up and running when hitting the delay loop calibration. At that
> stage, it is impossible to register a platform device, hence the use
> (abuse?) of early platform devices.

:-)  Unfortunately I never did get a chance to read through your
entire patch set, so I missed this.

> The current ARM code relies on the timers *not* being a standard
> device, and being directly setup by an board specific method. The SMP
> timers are even worse, as they are directly called by the core code,
> meaning that we can only have *one* implementation at a time in the
> kernel.
> 
> So the early platform stuff is a potential solution for that, though
> there is no way it would handle any kind of dependency. What I dream of
> is to have the full device/driver infrastructure available much
> earlier, before the rest of the kernel starts relying on some hardware
> being up and running...

My suggestion: don't use platform_device for this.  I think it is
entirely the wrong model.  Keep the timers *not* standard devices.

Instead, use a timer setup function that accepts a list of compatible
properties and an initialization hook for each timer type.  Since
timers have exist, have to be called by core code, and cannot ever be
built as modules, the Linux device model really doesn't offer much
advantage.

It is completely valid and often done to access device tree data from
early setup code without any context of a struct device.

g.