Hi Russell and Tomi,

On Wednesday 12 March 2014 12:47:09 Tomi Valkeinen wrote:

On 12/03/14 12:25, Russell King - ARM Linux wrote:

quoted

On Mon, Mar 10, 2014 at 02:52:53PM +0100, Laurent Pinchart wrote:

quoted

In theory unidirectional links in DT are indeed enough. However, let's
not forget the following.

- There's no such thing as single start points for graphs. Sure, in some
simple cases the graph will have a single start point, but that's not a
generic rule. For instance the camera graphs
http://ideasonboard.org/media/omap3isp.ps and
http://ideasonboard.org/media/eyecam.ps have two camera sensors, and thus
two starting points from a data flow point of view.

I think we need to stop thinking of a graph linked in terms of data
flow - that's really not useful.

Consider a display subsystem.  The CRTC is the primary interface for
the CPU - this is the "most interesting" interface, it's the interface
which provides access to the picture to be displayed for the CPU.  Other
interfaces are secondary to that purpose - reading the I2C DDC bus for
the display information is all secondary to the primary purpose of
displaying a picture.

For a capture subsystem, the primary interface for the CPU is the frame
grabber (whether it be an already encoded frame or not.)  The sensor
devices are all secondary to that.

So, the primary software interface in each case is where the data for
the primary purpose is transferred.  This is the point at which these
graphs should commence since this is where we would normally start
enumeration of the secondary interfaces.

V4L2 even provides interfaces for this: you open the capture device,
which then allows you to enumerate the capture device's inputs, and
this in turn allows you to enumerate their properties.  You don't open
a particular sensor and work back up the tree.

Please note that this has partly changed a couple of years ago with the 
introduction of the media controller framework. Userspace now opens a logical 
media device that describes the topology of the hardware, and then accesses 
individual components directly, from sensor to DMA engine.

We do it the other way around in OMAP DSS. It's the displays the user is
interested in, so we enumerate the displays, and if the user wants to
enable a display, we then follow the links from the display towards the
SoC, configuring and enabling the components on the way.

The logical view of a device from a CPU perspective evolves over time, as APIs 
are refactored or created to support new hardware that comes with new 
paradigms and additional complexity. The hardware data flow direction, 
however, doesn't change. Only modeling the data flow direction in DT might be 
tempting but is probably too hasty of a conclusion though : if DT should model 
the hardware, it ends up modeling a logical view of the hardware, and is thus 
not as closed as one might believe.

In the particular case of display devices I believe that using the data flow 
direction for links (assuming we can't use bidirectional links in DT) is a 
good model. It would allow parsing the whole graph at a reasonable cost (still 
higher than with bidirectional links) while making clear how to represent 
links. Let's not forgot that with more complex devices not all components can 
be referenced directly from the CPU-side display controller.

I don't have a strong opinion on the direction, I think both have their
pros. In any case, that's more of a driver model thing, and I'm fine
with linking in the DT outwards from the SoC (presuming the
double-linking is not ok, which I still like best).

quoted

I believe trying to do this according to the flow of data is just wrong.
You should always describe things from the primary device for the CPU
towards the peripheral devices and never the opposite direction.

In that case there's possibly the issue I mentioned in other email in
this thread: an encoder can be used in both a display and a capture
pipeline. Describing the links outwards from CPU means that sometimes
the encoder's input port is pointed at, and sometimes the encoder's
output port is pointed at.

That's possibly ok, but I think Grant was of the opinion that things
should be explicitly described in the binding documentation: either a
device's port must contain a 'remote-endpoint' property, or it must not,
but no "sometimes". But maybe I took his words too literally.

Then there's also the audio example Philipp mentioned, where there is no
clear "outward from Soc" direction for the link, as the link was
bi-directional and between two non-SoC devices.

Even if the link was unidirectional the "outward from SoC" direction isn't 
always defined for links between non-SoC devices.

-- 
Regards,

Laurent Pinchart