Re: [RFC 0/2] drm/dsi: DSI for devices with different control bus

From: Archit Taneja
Date: Thu Sep 10 2015 - 02:15:47 EST




On 09/08/2015 03:57 PM, Andrzej Hajda wrote:
On 09/07/2015 01:46 PM, Archit Taneja wrote:
Thierry,

On 08/21/2015 11:39 AM, Archit Taneja wrote:


On 08/20/2015 05:18 PM, Thierry Reding wrote:
On Thu, Aug 20, 2015 at 09:46:14AM +0530, Archit Taneja wrote:
Hi Thierry, Lucas,


On 08/19/2015 08:32 PM, Thierry Reding wrote:
On Wed, Aug 19, 2015 at 04:52:24PM +0200, Lucas Stach wrote:
Am Mittwoch, den 19.08.2015, 16:34 +0200 schrieb Thierry Reding:
On Wed, Aug 19, 2015 at 04:17:08PM +0200, Lucas Stach wrote:
Hi Thierry, Archit,

[...]
Perhaps a better way would be to invert this relationship.
According to
your proposal we'd have to have DT like this:

i2c@... {
...

dsi-device@... {
...
dsi-bus = <&dsi>;
...
};

...
};

dsi@... {
...
};

Inversing the relationship would become something like this:

i2c@... {
...
};

dsi@... {
...

peripheral@... {
...
i2c-bus = <&i2c>;
...
};

...
};

Both of those aren't fundamentally different, and they both have
the
disavantage of lacking ways to transport configuration data that
the
other bus needs to instantiate the dummy device (such as the reg
property for example, denoting the I2C slave address or the DSI
VC).

So how about we create two devices in the device tree and fuse
them at
the driver level:

i2c@... {
...

i2cdsi: dsi-device@... {
...
};

...
};

dsi@... {
...

peripheral@... {
...
control = <&i2cdsi>;
...
};

...
};

This way we'll get both an I2C device and a DSI device that we
can fully
describe using the standard device tree bindings. At driver time
we can
get the I2C device from the phandle in the control property of
the DSI
device and use it to execute I2C transactions.

I don't really like to see that you are inventing yet-another-way to
handle devices connected to multiple buses.

Devicetree is structured along the control buses, even if the
devices
are connected to multiple buses, in the DT they are always
children of
the bus that is used to control their registers from the CPUs
perspective. So a DSI encoder that is controlled through i2c is
clearly
a child of the i2c master controller and only of that one.

I think that's a flawed interpretation of what's going on here. The
device in fact has two interfaces: one is I2C, the other is DSI.
In my
opinion that's reason enough to represent it as two logical devices.

Does it really have 2 control interfaces that are used at the same
time?
Or is the DSI connection a passive data bus if the register control
happens through i2c?

The interfaces may not be used at the same time, and the DSI interface
may even be crippled, but the device is still addressable from the DSI
host controller, if for nothing else than for routing the video stream.

If you need to model connections between devices that are not
reflected
through the control bus hierarchy you should really consider
using the
standardized of-graph bindings.
(Documentation/devicetree/bindings/graph.txt)

The problem is that the original proposal would instantiate a dummy
device, so it wouldn't be represented in DT at all. So unless you
do add
two logical devices to DT (one for each bus interface), you don't
have
anything to glue together with an OF graph.

I see that the having dummy device is the least desirable solution.
But
if there is only one control bus to the device I think it should be
one
device sitting beneath the control bus.

You can then use of-graph to model the data path between the DSI
encoder
and device.

But you will be needing a device below the DSI host controller to
represent that endpoint of the connection. The DSI host controller
itself is in no way connected to the I2C adapter. You would have to
add some sort of quirk to the DSI controller binding to allow it to

Thanks for the review.

I implemented this to support ADV7533 DSI to HDMI encoder chip, which
has a DSI video bus and an i2c control bus.

There weren't any quirks as such in the device tree when I tried to
implement this. The DT seems to manage fine without a node
corresponding to a mipi_dsi_device:

i2c_adap@.. {
adv7533@.. {

port {
adv_in: endpoint {
remote-endpoint = <&dsi_out>;
};
};
};
};

dsi_host@.. {
...
...

port {
dsi_out: endpoint {
remote-endpoint = <&adv_in>;
}
};
};

It's the i2c driver's job to parse the graph and retrieve the
phandle to the dsi host. Using this, it can proceed with
registering itself to this host using the new dsi funcs. This
patch does the same for the adv7533 i2c driver:

http://www.spinics.net/lists/dri-devel/msg86840.html

hook up with a control endpoint. And then you'll need more quirks
to describe what kind of DSI device this is.

Could you explain what you meant by this? I.e. describing the kind
of DSI device?

Describing the number of lanes, specifying the virtual channel, mode
flags, etc. You could probably set the number of lanes and mode flags
via the I2C driver, but especially the virtual channel cannot be set
because it isn't known to the I2C DT branch of the device.

I agree with the VC part. It could be a DT entry within the i2c client
node, but that does make it seem like a quirk. The 'reg' way under the
DSI host is definitely better to populate the virtual channel.


The dsi device created isn't really a dummy device as such. It's
dummy in the sense that there isn't a real dsi driver associated
with it. The dsi device is still used to attach to a mipi dsi host,
the way normal dsi devices do.

I understand, but I don't see why it has to be instantiated by the I2C
driver, that's what I find backwards. There is already a standard way
for instantiating DSI devices, why not use it?

I assumed we could either represent the device using an i2c driver, or
dsi, but not both. Hence, I came up with this approach.


On the other hand if you properly instantiate the DSI device you can
easily write a driver for it, and the driver will set up the correct
properties as implied by the compatible string. Once you have that you
can easily hook it up to the I2C control interface in whatever way you
like, be that an OF graph or just a simple unidirectional link by
phandle.


With the fused approach you suggested, we would have 2 drivers: one i2c
and the other dsi. The i2c client driver would be more or less minimal,
preparing an i2c_client device for the dsi driver to use. Is my
understanding correct?

Correct. That's kind of similar to the way an HDMI encoder driver would
use an I2C adapter to query EDID. The i2c_client device would be a means
for the DSI driver to access the control interface.

Okay.

Although, I'm not sure about the HDMI encoder example. An HDMI
encoder would read off edid directly from the adapter(with an address
specified), it wouldn't need to create an i2c client device. Therefore,
an HDMI encoder wouldn't need to have a separate node for i2c in DT.


We can do without dummy dsi devices with this method. But, representing
a device with 2 DT nodes seems a bit off. We'd also need to compatible
strings for the same device, one for the i2c part, and the other for
the dsi part.

I agree that this somewhat stretches the capabilities of device tree.
Another alternative I guess would be to not have a compatible string for
the I2C device at all (that's technically not valid, I guess) because we
really don't need an I2C driver for the device. What we really need is a
DSI driver with a means to talk over some I2C bus with some other part
of its device.

I think what the driver should 'really' be is a bit subjective, and can
vary based on what the buses are used for in the device. For the Toshiba
chip that Jani mentioned, it tends more towards a DSI driver. Whereas,
for an ADV75xx chip, it's closer to an I2C driver since only I2C can be
used to configure the chip registers.

Although, I agree with the point you made about the DSI bus here:

"and the DSI interface may even be crippled, but the device is still
addressable from the DSI host controller, if for nothing else than for
routing the video stream."

The fact that the data on the DSI bus contains routing information (i.e,
virtual channel number) always gives it some 'control' aspect.


From an adv75xx driver perspective, it should also support the ADV7511
chip, which is a RGB/DPI to HDMI encoder. For adv7511, we don't need a
DSI DT node. It would be a bit inconsistent to have the bindings
require both DSI and I2C nodes for one chip, and only I2C node for the
other, with both chips being supported by the same driver.

Why would that be inconsistent? That sounds like the most accurate
representation of the hardware to me.

Inconsistent wasn't the right term. I should have used 'uncommon' :)
It's common for two chips of the same family to have a different set
optional properties in DT, but it's not common for two chips of the
same family to be represented by a different number of devices in
DT.

I don't have an issue with the fused approach you suggested, as long
as people are okay with the DT parts. Especially the part of needing 2
compatible strings in the DT.

I implemented the ADV7533 driver with the approach you suggested above
(2 drivers for 2 different components of the chip). I posted it out
just a while back (with you in loop).

The DT node with this apporach would look like this:

https://github.com/boddob/linux/blob/c24cbf63a6998d00095c10122ce5e37b764c7dba/arch/arm64/boot/dts/qcom/apq8016-sbc.dtsi#L162

The main irritant with the '2 driver' approach is that we need to
share the per-device driver data with them. For ADV7533, I've made
the i2c driver allocate driver data (struct adv7511).

The dsi driver gets the driver data in the following way:

- The i2c driver sets the driver data as its client data using
i2c_set_clientdata()
- Parse the i2c-control phandle to get the corresponding i2c client.
- Extract the adv7511 struct by getting i2c_get_clientdata()

This way of getting the same driver data is a bit strange, but it
works. For this, we do need to ensure that the dsi driver defers
as long as the i2c driver isn't probed.

I've now implemented both approaches for the driver. The first using
a dummy dsi device, and this one using 2 drivers (with both being
represented in DT). The advantage of the latter is that we don't need
to create any dummy device stuff, the disadvantage is that DT is a bit
uncommon.

Can we now come to a conclusion on what approach is better?

DSI by design is data bus which can be used additionally as a control bus, but
in this particular case it is purely data bus. So of-graph bindings seem to be
better choice. As already Lucas Stach said DT hierarchy should describe control
buses and of-graph bindings should describe data bus. Argument that hw has two
interfaces does not seem to be valid here - it has only one control interface.
The other one is only for data, representing every data interface using DT
hierarchy would lead to inflation of pseudo devices.

On the other side I do not see dummy device approach ideal solution, I guess
lightweight framework providing DSI hosts detached from Linux device model could
work better here.
The only problem here is that it should coexist somehow with dsi bus used to
control devices. Anyway implementing it shouldn't be hard, question is if it
would be eventually accepted :) I guess we can live for now with dummy devs.

Summarizing I would prefer version with dummy devices, as it seems more
compatible with DT design.

Thanks for the feedback. I'll spin a newer version of the dummy dsi dev patches after waiting for some more comments.

Archit

--
To unsubscribe from this list: send the line "unsubscribe linux-arm-msm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html


--
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/