On Thu, 2019-10-03 at 20:53 -0500, Rob Herring wrote:
> > > > But I think that with this series, given the fact that we now treat the
> > > > lack
> > > > of
> > > > dma-ranges as a 1:1 mapping instead of an error, we could rewrite the
> > > > function
> > > > like this:
> > > 
> > > Now, I'm reconsidering allowing this abuse... It's better if the code
> > > which understands the bus structure in DT for a specific bus passes in
> > > the right thing. Maybe I should go back to Robin's version (below).
> > > OTOH, the existing assumption that 'dma-ranges' was in the immediate
> > > parent was an assumption on the bus structure which maybe doesn't
> > > always apply.
> > > 
> > > diff --git a/drivers/of/device.c b/drivers/of/device.c
> > > index a45261e21144..6951450bb8f3 100644
> > > --- a/drivers/of/device.c
> > > +++ b/drivers/of/device.c
> > > @@ -98,12 +98,15 @@ int of_dma_configure(struct device *dev, struct
> > > device_node *parent, bool force_
> > >         u64 mask;
> > > 
> > >         np = dev->of_node;
> > > -       if (!np)
> > > -               np = parent;
> > > +       if (np)
> > > +               parent = of_get_dma_parent(np);
> > > +       else
> > > +               np = of_node_get(parent);
> > >         if (!np)
> > >                 return -ENODEV;
> > > 
> > > -       ret = of_dma_get_range(np, &dma_addr, &paddr, &size);
> > > +       ret = of_dma_get_range(parent, &dma_addr, &paddr, &size);
> > > +       of_node_put(parent);
> > >         if (ret < 0) {
> > >                 /*
> > >                  * For legacy reasons, we have to assume some devices need
> > 
> > I spent some time thinking about your comments and researching. I came to
> > the
> > realization that both these solutions break the usage in
> > drivers/gpu/drm/sun4i/sun4i_backend.c:805. In that specific case both
> > 'dev->of_node' and 'parent' exist yet the device receiving the configuration
> > and 'parent' aren't related in any way.
> 
> I knew there was some reason I didn't like those virtual DT nodes...
> 
> That does seem to be the oddest case. Several of the others are just
> non-DT child platform devices. Perhaps we need a "copy the DMA config
> from another struct device (or parent struct device)" function to
> avoid using a DT function on a non-DT device.
> 
> > IOW we can't just use 'dev->of_node' as a starting point to walk upwards the
> > tree. We always have to respect whatever DT node the bus provided, and start
> > there. This clashes with the current solutions, as they are based on the
> > fact
> > that we can use dev->of_node when present.
> 
> Yes, you are right.
> 
> > My guess at this point, if we're forced to honor that behaviour, is that we
> > have to create a new API for the PCI use case. Something the likes of
> > of_dma_configure_parent().
> 
> I think of_dma_configure just has to work with the device_node of
> either the device or the device parent and dev->of_node is never used
> unless the caller sets it.

Fine, so given the following two distinct uses of
of_dma_configure(struct device *dev, struct device_node *np, bool ...):

- dev->of_node == np: Platform bus' typical use, we imperatively have to start
  parsing dma-ranges from np's DMA parent, as the device we're configuring
  might be a bus containing dma-ranges himself. For example a platform PCIe bus.

- dev->of_node != np: Here the bus is pulling some trick. The device might or
  might not be represented in DT and np might be a bus or a device. But one
  thing I realised is that device being configured never represents a memory
  mapped bus. Assuming this assumption is acceptable, we can traverse the DT
  tree starting from np and get a correct configuration as long as dma-ranges
  not being present is interpreted as a 1:1 mapping.

The resulting code, which I tested on an RPi4, Freescale Layerscape and passes
OF's unit tests, looks like this:

int of_dma_configure(struct device *dev, struct device_node *np, bool force_dma)
{
	u64 dma_addr, paddr, size = 0;
	struct device_node *parent;
	u64 mask;
	int ret;

	if (!np)
		return -ENODEV;

	parent = of_node_get(np);
	if (dev->of_node == parent)
		parent = of_get_next_dma_parent(np);

	ret = of_dma_get_range(parent, &dma_addr, &paddr, &size);
	of_node_put(parent);

	[...]
}

Would that be acceptable?

Regards,
Nicolas