Re: [PATCH v3 13/24] platform: add video-multiplexer subdevice driver

[Laurent Pinchart]

Hi Benoit,

On Tuesday 07 Feb 2017 07:36:48 Benoit Parrot wrote:
> Laurent Pinchart wrote on Tue [2017-Feb-07 12:26:32 +0200]:
> > On Monday 06 Feb 2017 15:10:46 Steve Longerbeam wrote:
> >> On 02/06/2017 02:33 PM, Laurent Pinchart wrote:
> >>> On Monday 06 Feb 2017 10:50:22 Hans Verkuil wrote:
> >>>> On 02/05/2017 04:48 PM, Laurent Pinchart wrote:
> >>>>> On Tuesday 24 Jan 2017 18:07:55 Steve Longerbeam wrote:
> >>>>>> On 01/24/2017 04:02 AM, Philipp Zabel wrote:
> >>>>>>> On Fri, 2017-01-20 at 15:03 +0100, Hans Verkuil wrote:
> >>>>>>>>> +
> >>>>>>>>> +int vidsw_g_mbus_config(struct v4l2_subdev *sd, struct
> >>>>>>>>> v4l2_mbus_config *cfg)
> > 
> > [snip]
> > 
> >>>>>>>> I am not certain this op is needed at all. In the current kernel
> >>>>>>>> this op is only used by soc_camera, pxa_camera and omap3isp
> >>>>>>>> (somewhat dubious). Normally this information should come from the
> >>>>>>>> device tree and there should be no need for this op.
> >>>>>>>> 
> >>>>>>>> My (tentative) long-term plan was to get rid of this op.
> >>>>>>>> 
> >>>>>>>> If you don't need it, then I recommend it is removed.
> >>>>>> 
> >>>>>> Hi Hans, the imx-media driver was only calling g_mbus_config to the
> >>>>>> camera sensor, and it was doing that to determine the sensor's bus
> >>>>>> type. This info was already available from parsing a
> >>>>>> v4l2_of_endpoint from the sensor node. So it was simple to remove the
> >>>>>> g_mbus_config calls, and instead rely on the parsed sensor
> >>>>>> v4l2_of_endpoint.
> >>>>> 
> >>>>> That's not a good point.
> >>> 
> >>> (mea culpa, s/point/idea/)
> >>> 
> >>>>> The imx-media driver must not parse the sensor DT node as it is not
> >>>>> aware of what bindings the sensor is compatible with.
> >> 
> >> Hi Laurent,
> >> 
> >> I don't really understand this argument. The sensor node has been found
> >> by parsing the OF graph, so it is known to be a camera sensor node at
> >> that point.
> > 
> > All you know in the i.MX6 driver is that the remote node is a video
> > source. You can rely on the fact that it implements the OF graph bindings
> > to locate other ports in that DT node, but that's more or less it.
> > 
> > DT properties are defined by DT bindings and thus qualified by a
> > compatible string. Unless you match on sensor compat strings in the i.MX6
> > driver (which you shouldn't do, to keep the driver generic) you can't know
> > for certain how to parse the sensor node DT properties. For all you know,
> > the video source could be a bridge such as an HDMI to CSI-2 converter for
> > instance, so you can't even rely on the fact that it's a sensor.
> > 
> >>>>> Information must instead be queried from the sensor subdev at
> >>>>> runtime, through the g_mbus_config() operation.
> >>>>> 
> >>>>> Of course, if you can get the information from the imx-media DT
> >>>>> node, that's certainly an option. It's only information provided by
> >>>>> the sensor driver that you have no choice but query using a subdev
> >>>>> operation.
> >>>> 
> >>>> Shouldn't this come from the imx-media DT node? BTW, why is omap3isp
> >>>> using this?
> >>> 
> >>> It all depends on what type of information needs to be retrieved, and
> >>> whether it can change at runtime or is fixed. Adding properties to the
> >>> imx-media DT node is certainly fine as long as those properties
> >>> describe the i.MX side.
> >> 
> >> In this case the info needed is the media bus type. That info is most
> >> easily available by calling v4l2_of_parse_endpoint() on the sensor's
> >> endpoint node.
> > 
> > I haven't had time to check the code in details yet, so I can't really
> > comment on what you need and how it should be implemented exactly.
> > 
> >> The media bus type is not something that can be added to the
> >> imx-media node since it contains no endpoint nodes.
> > 
> > Agreed. You have endpoints in the CSI nodes though.
> > 
> >>> In the omap3isp case, we use the operation to query whether parallel
> >>> data contains embedded sync (BT.656) or uses separate h/v sync signals.
> >>> 
> >>>> The reason I am suspicious about this op is that it came from
> >>>> soc-camera and predates the DT. The contents of v4l2_mbus_config seems
> >>>> very much like a HW description to me, i.e. something that belongs in
> >>>> the DT.
> >>> 
> >>> Part of it is possibly outdated, but for buses that support multiple
> >>> modes of operation (such as the parallel bus case described above) we
> >>> need to make that information discoverable at runtime. Maybe this should
> >>> be considered as related to Sakari's efforts to support VC/DT for CSI-2,
> >>> and supported through the API he is working on.
> >> 
> >> That sounds interesting, can you point me to some info on this effort?
> > 
> > Sure.
> > 
> > http://git.retiisi.org.uk/?p=~sailus/linux.git;a=shortlog;h=refs/heads/vc
> > 
> >> I've been thinking the DT should contain virtual channel info for CSI-2
> >> buses.
> > 
> > I don't think it should. CSI-2 virtual channels and data types should be
> > handled as a software concept, and thus supported through driver code
> > without involving DT.
> 
> Laurent,
> 
> So when you have a CSI2 port aggregator for instance where traffic from up
> to 4 CSI2 sources where each source is now assigned its own VC by the
> aggregator and interleaved into a single CSI2 Receiver. I was hoping that
> in this case the VC would be DT discoverable as a specicic source
> identifier. So the CSI-RX side could associate a specific source and create
> its own video device.

In this specific example, I believe the aggregator should be modelled with 4 
input ports and one output port in DT, and with 4 sink pads and one source pad 
in MC. Information about the VCs multiplexed over the aggregator's source link 
should not be part of the device tree, but should be discoverable at runtime 
through V4L2 subdev operations. This would include information about how the 4 
input streams are routed to VCs inside the aggregator.

> I am guessing that no such thing exist today?

There's very little (to not say nothing) in terms of VC support in V4L2 today.

-- 
Regards,

Laurent Pinchart