[v4,0/8] i2c-atr and FPDLink

  Hi,

Intro
-----

This is, kind of, v4 of Luca's i2c-atr and FPDLink series, v3 of which
you can find from:

https://lore.kernel.org/all/20220206115939.3091265-1-luca@lucaceresoli.net/

I say "kind of", as the FPDLink drivers have diverged from Luca's
version quite a bit and the drivers support different HW versions. A
Big thanks for Luca for working on the drivers!

I'd really like to send and review i2c-atr and FPDLink drivers
separately, but as the concepts are new and those drivers are linked
together, in the end I decided to keep them in one series. Even so, I
think these patches divide quite clearly into to areas:

1) i2c-atr and the related code in the FPDLink drivers
2) Everything else about FPDLink

The FPDLink drivers support multiple streams and for that reason the
series is based on V4L2 streams series v15:

https://lore.kernel.org/all/20221003121852.616745-1-tomi.valkeinen@ideasonboard.com/

HW overview
-----------

TI's DS90UB9xx IC (UB9xx for short) family is a set of deserializer and
serializer ICs for video, both display and capture. These ICs support
FPDLink 3 and some also support FPDLink 4. From the user's point of view
there's really not much difference between FPDLink 3 and 4.

An example HW setup with two cameras could be like this:

                          +-- [Ser 1] <-- [Camera 1]
[CSI-2 RX] <-- [Deser] <--|
                          +-- [Ser 2] <-- [Camera 2]

The cameras send video streams over CSI-2 to the serializers. The
serializers encode the received data to FPDLink and send it to the
deserializer. The deserializer decodes the data and sends it forward
over CSI-2 bus.

The FPDLink bus is a bi-directional bus, with a fast forward-channel and
a slow back-channel. In addition to the video, the devices support
forwarding GPIOs (both ways) and forwarding i2c transactions with
address translation.

This series is only about the capture ICs. The HW versions supported
by these drivers are:

- UB960, FPDLink 3 deserializer
- UB9702, FPDLink 3/4 deserializer
- UB953, FPDLink 3 CSI-2 serializer
- UB913, FPDLink 3 parallel video serializer
- UB971, FPDLink 3/4 CSI-2 serializer

Note that this series does not support UB954 deserializer, which Luca
has. I don't have it and decided early to drop it for simpler
development. As UB954 is a subset of UB960, adding support for it should
be trivial.

FPDLink Deser and Ser Devices & Drivers
---------------------------------------

The serializer devices are, in a way, child devices of the deserializer.
It is the deserializer driver which creates and removes the serializer
devices, although the serializer device itself is not a child of the
deserializer, but rather the serializer is created as a device on the
i2c bus.

There are perhaps other methods to add the serializer devices. I think a
real bus is an overkill, as FPDLink is a point-to-point link. But if
there's no FPDLink bus, what are the options? Platform device? That's
not good either. So, for the time being, the serializers are i2c clients
on the main i2c bus, even if they're really behind the FPDLink.

Also, the deserializer driver needs to share some information with the
serializers, and it is done with platform data set to the serializer
device. If we had a bus, the bus could probably be used to convey this
information.

I2C & I2C ATR
-------------

We have three different i2c client types:

1) The deserializer is a normal i2c slave, and there's nothing special
about accessing it. The deserializer is connected to the "main" i2c bus.

2) The serializers are also accessible with i2c via FPDLink: when a
transaction to a specific (programmable) address happens on the main
i2c-bus the deserializer takes the transaction and forwards it via
FPDLink to the serializer. This is implemented in the deserializer
driver by just creating the serializer i2c device on the main i2c-bus.
This is not 100% correct, as the serializers are not directly connected
to an i2c bus as i2c slaves.

3) The serializer is an i2c master, and the camera and possibly other
i2c peripherals can be connected to that "remote" i2c bus. The
deserializer hardware has an i2c-alias table which describes which
i2c-alias address the deserializer will react to, what is the real
i2c-address and to which serializer the transactions will be forwarded
to. This is called address translation (ATR).

I2C ATR Thoughts
----------------

I have addressed most of the comments Luca received for the i2c-atr
driver, but other than that the i2c-atr driver is mostly the same.
However, there's one big difference: the i2c bus can be given by the
drivers.

So, looking at the DT, in Luca's version the deserializer node looked
like this for serializers and i2c-atr:

deser {
	remote-chips {
		remote-chip@0 {
			// The serializer
		};
	};

	i2c-atr {
		i2c@0 {
			// This is the bus behind the serializer
		};
	};
};

I have:

deser {
	links {
		link@0 {
			serializer {
				// The serializer
				i2c {
					// This is the bus
				};
			}
		};
	};
};

I think this reflects the hardware much better. But this means that the
i2c-atr "device", which in Luca's version is only the deserializer, is
in my version split between the deser and serializer. However, the only
thing I changed in the i2c-atr driver is that the i2c_atr_add_adapter()
function takes an fwnode handle which tells where the i2c bus is found
(but if it's NULL, it looks for i2c-atr just like in Luca's version).

Perhaps the biggest difference is that in Luca's version the i2c-atr was
private to the deserializer, and the deser driver called
i2c_atr_add_adapter(). In my version the deser shares the i2c-atr with
the serializers, and the serializers call i2c_atr_add_adapter(). I think
this is much better, as the i2c-bus is behind the serializer, and the
serializer's registers affect the bus, and thus the bus should really be
create/destroyed by the serializer driver as it controls the bus'
i2c-master hardware. 

Now, i2c-mux and i2c-atr are quite similar as has been discussed in the
earlier reviews. And while the FPDLink ICs support ATR, we can easily
imagine a simpler deserializer which only supports mux-style forwarding.
For these reasons I believe we have two topics: 1) i2c-atr without
FPDLink, and 2) i2c-atr (and i2c-mux) with FPDlink.

1)
I am not aware of a stand-alone IC that performs address translation.
If there is, I think i2c-atr as it is in this series is a good solution
(but the bus fwnode feature mentioned above can be dropped).

2)
If I suggested adding an i2c-bus fwnode parameter to
i2c_mux_add_adapter(), and the i2c-bus might be under some other device,
I think the reception could be quite negative (and I would agree). For
this reason I'm not very happy with the i2c-atr and using it with
FPDLink.

In fact, I'm thinking that it might be better to just drop the i2c-atr
driver, and add the support directly to the FPDlink drivers. But that
could mean possibly duplicating the same code for other deser/ser
architectures, so I have kept the i2c-atr driver for now.

In any case, I think when figuring out 2), we can forget about ATR and
the added complexity the translation brings, and just think how i2c-mux
could be used in a deser-ser setup. And maybe the answer is "don't use
it, just write the support directly to the deser-ser drivers".

 Tomi

Luca Ceresoli (2):
  i2c: core: let adapters be notified of client attach/detach
  i2c: add I2C Address Translator (ATR) support

Tomi Valkeinen (6):
  dt-bindings: media: add bindings for TI DS90UB960
  dt-bindings: media: add bindings for TI DS90UB913
  dt-bindings: media: add bindings for TI DS90UB953
  media: i2c: add DS90UB960 driver
  media: i2c: add DS90UB913 driver
  media: i2c: add DS90UB953 driver

 .../bindings/media/i2c/ti,ds90ub913.yaml      |  127 +
 .../bindings/media/i2c/ti,ds90ub953.yaml      |  120 +
 .../bindings/media/i2c/ti,ds90ub960.yaml      |  392 ++
 Documentation/i2c/index.rst                   |    1 +
 Documentation/i2c/muxes/i2c-atr.rst           |   78 +
 MAINTAINERS                                   |    7 +
 drivers/i2c/Kconfig                           |    9 +
 drivers/i2c/Makefile                          |    1 +
 drivers/i2c/i2c-atr.c                         |  497 ++
 drivers/i2c/i2c-core-base.c                   |   18 +-
 drivers/media/i2c/Kconfig                     |   28 +
 drivers/media/i2c/Makefile                    |    3 +
 drivers/media/i2c/ds90ub913.c                 |  892 ++++
 drivers/media/i2c/ds90ub953.c                 | 1607 +++++++
 drivers/media/i2c/ds90ub960.c                 | 4198 +++++++++++++++++
 include/linux/i2c-atr.h                       |   80 +
 include/linux/i2c.h                           |   16 +
 include/media/i2c/ds90ub9xx.h                 |   16 +
 18 files changed, 8089 insertions(+), 1 deletion(-)
 create mode 100644 Documentation/devicetree/bindings/media/i2c/ti,ds90ub913.yaml
 create mode 100644 Documentation/devicetree/bindings/media/i2c/ti,ds90ub953.yaml
 create mode 100644 Documentation/devicetree/bindings/media/i2c/ti,ds90ub960.yaml
 create mode 100644 Documentation/i2c/muxes/i2c-atr.rst
 create mode 100644 drivers/i2c/i2c-atr.c
 create mode 100644 drivers/media/i2c/ds90ub913.c
 create mode 100644 drivers/media/i2c/ds90ub953.c
 create mode 100644 drivers/media/i2c/ds90ub960.c
 create mode 100644 include/linux/i2c-atr.h
 create mode 100644 include/media/i2c/ds90ub9xx.h
  

Hi Tomi, All,

On Tue,  1 Nov 2022 15:20:24 +0200
Tomi Valkeinen <tomi.valkeinen@ideasonboard.com> wrote:

> Hi,
> 
> Intro
> -----
> 
> This is, kind of, v4 of Luca's i2c-atr and FPDLink series, v3 of which
> you can find from:
> 
> https://lore.kernel.org/all/20220206115939.3091265-1-luca@lucaceresoli.net/
> 
> I say "kind of", as the FPDLink drivers have diverged from Luca's
> version quite a bit and the drivers support different HW versions. A
> Big thanks for Luca for working on the drivers!
> 
> I'd really like to send and review i2c-atr and FPDLink drivers
> separately, but as the concepts are new and those drivers are linked
> together, in the end I decided to keep them in one series. Even so, I
> think these patches divide quite clearly into to areas:
> 
> 1) i2c-atr and the related code in the FPDLink drivers
> 2) Everything else about FPDLink
> 
> The FPDLink drivers support multiple streams and for that reason the
> series is based on V4L2 streams series v15:
> 
> https://lore.kernel.org/all/20221003121852.616745-1-tomi.valkeinen@ideasonboard.com/
> 
> HW overview
> -----------
> 
> TI's DS90UB9xx IC (UB9xx for short) family is a set of deserializer and
> serializer ICs for video, both display and capture. These ICs support
> FPDLink 3 and some also support FPDLink 4. From the user's point of view
> there's really not much difference between FPDLink 3 and 4.
> 
> An example HW setup with two cameras could be like this:
> 
>                           +-- [Ser 1] <-- [Camera 1]
> [CSI-2 RX] <-- [Deser] <--|
>                           +-- [Ser 2] <-- [Camera 2]
> 
> The cameras send video streams over CSI-2 to the serializers. The
> serializers encode the received data to FPDLink and send it to the
> deserializer. The deserializer decodes the data and sends it forward
> over CSI-2 bus.
> 
> The FPDLink bus is a bi-directional bus, with a fast forward-channel and
> a slow back-channel. In addition to the video, the devices support
> forwarding GPIOs (both ways) and forwarding i2c transactions with
> address translation.
> 
> This series is only about the capture ICs. The HW versions supported
> by these drivers are:
> 
> - UB960, FPDLink 3 deserializer
> - UB9702, FPDLink 3/4 deserializer
> - UB953, FPDLink 3 CSI-2 serializer
> - UB913, FPDLink 3 parallel video serializer
> - UB971, FPDLink 3/4 CSI-2 serializer
> 
> Note that this series does not support UB954 deserializer, which Luca
> has. I don't have it and decided early to drop it for simpler
> development. As UB954 is a subset of UB960, adding support for it should
> be trivial.
> 
> FPDLink Deser and Ser Devices & Drivers
> ---------------------------------------
> 
> The serializer devices are, in a way, child devices of the deserializer.
> It is the deserializer driver which creates and removes the serializer
> devices, although the serializer device itself is not a child of the
> deserializer, but rather the serializer is created as a device on the
> i2c bus.
> 
> There are perhaps other methods to add the serializer devices. I think a
> real bus is an overkill, as FPDLink is a point-to-point link. But if
> there's no FPDLink bus, what are the options? Platform device? That's
> not good either. So, for the time being, the serializers are i2c clients
> on the main i2c bus, even if they're really behind the FPDLink.
> 
> Also, the deserializer driver needs to share some information with the
> serializers, and it is done with platform data set to the serializer
> device. If we had a bus, the bus could probably be used to convey this
> information.
> 
> I2C & I2C ATR
> -------------
> 
> We have three different i2c client types:
> 
> 1) The deserializer is a normal i2c slave, and there's nothing special
> about accessing it. The deserializer is connected to the "main" i2c bus.
> 
> 2) The serializers are also accessible with i2c via FPDLink: when a
> transaction to a specific (programmable) address happens on the main
> i2c-bus the deserializer takes the transaction and forwards it via
> FPDLink to the serializer. This is implemented in the deserializer
> driver by just creating the serializer i2c device on the main i2c-bus.
> This is not 100% correct, as the serializers are not directly connected
> to an i2c bus as i2c slaves.
> 
> 3) The serializer is an i2c master, and the camera and possibly other
> i2c peripherals can be connected to that "remote" i2c bus. The
> deserializer hardware has an i2c-alias table which describes which
> i2c-alias address the deserializer will react to, what is the real
> i2c-address and to which serializer the transactions will be forwarded
> to. This is called address translation (ATR).

An excellent description so far!

> I2C ATR Thoughts
> ----------------
> 
> I have addressed most of the comments Luca received for the i2c-atr
> driver, but other than that the i2c-atr driver is mostly the same.
> However, there's one big difference: the i2c bus can be given by the
> drivers.
> 
> So, looking at the DT, in Luca's version the deserializer node looked
> like this for serializers and i2c-atr:
> 
> deser {
> 	remote-chips {
> 		remote-chip@0 {
> 			// The serializer
> 		};
> 	};
> 
> 	i2c-atr {
> 		i2c@0 {
> 			// This is the bus behind the serializer
> 		};
> 	};
> };
> 
> I have:
> 
> deser {
> 	links {
> 		link@0 {
> 			serializer {
> 				// The serializer
> 				i2c {
> 					// This is the bus
> 				};
> 			}
> 		};
> 	};
> };
> 
> I think this reflects the hardware much better. But this means that the
> i2c-atr "device", which in Luca's version is only the deserializer, is
> in my version split between the deser and serializer. However, the only
> thing I changed in the i2c-atr driver is that the i2c_atr_add_adapter()
> function takes an fwnode handle which tells where the i2c bus is found
> (but if it's NULL, it looks for i2c-atr just like in Luca's version).
> 
> Perhaps the biggest difference is that in Luca's version the i2c-atr was
> private to the deserializer, and the deser driver called
> i2c_atr_add_adapter(). In my version the deser shares the i2c-atr with
> the serializers, and the serializers call i2c_atr_add_adapter(). I think
> this is much better, as the i2c-bus is behind the serializer, and the
> serializer's registers affect the bus, and thus the bus should really be
> create/destroyed by the serializer driver as it controls the bus'
> i2c-master hardware. 
> 
> Now, i2c-mux and i2c-atr are quite similar as has been discussed in the
> earlier reviews. And while the FPDLink ICs support ATR, we can easily
> imagine a simpler deserializer which only supports mux-style forwarding.
> For these reasons I believe we have two topics: 1) i2c-atr without
> FPDLink, and 2) i2c-atr (and i2c-mux) with FPDlink.
> 
> 1)
> I am not aware of a stand-alone IC that performs address translation.
> If there is, I think i2c-atr as it is in this series is a good solution
> (but the bus fwnode feature mentioned above can be dropped).
> 
> 2)
> If I suggested adding an i2c-bus fwnode parameter to
> i2c_mux_add_adapter(), and the i2c-bus might be under some other device,
> I think the reception could be quite negative (and I would agree). For
> this reason I'm not very happy with the i2c-atr and using it with
> FPDLink.
> 
> In fact, I'm thinking that it might be better to just drop the i2c-atr
> driver, and add the support directly to the FPDlink drivers. But that
> could mean possibly duplicating the same code for other deser/ser
> architectures, so I have kept the i2c-atr driver for now.

Indeed I think the ROHM serdes chips do have an address translation
feature that works pretty much like the TI ones, and the ATR should be
cleanly reusable across the two brands. The ATR code might be
simplified to just provide helpers for common code maybe, but I'd
rather avoid code duplication.
  

On 11/7/22 13:48, Luca Ceresoli wrote:
> Hi Tomi, All,
> 
> On Tue,  1 Nov 2022 15:20:24 +0200
> Tomi Valkeinen <tomi.valkeinen@ideasonboard.com> wrote:
>>
>> 2)
>> If I suggested adding an i2c-bus fwnode parameter to
>> i2c_mux_add_adapter(), and the i2c-bus might be under some other device,
>> I think the reception could be quite negative (and I would agree). For
>> this reason I'm not very happy with the i2c-atr and using it with
>> FPDLink.
>>
>> In fact, I'm thinking that it might be better to just drop the i2c-atr
>> driver, and add the support directly to the FPDlink drivers. But that
>> could mean possibly duplicating the same code for other deser/ser
>> architectures, so I have kept the i2c-atr driver for now.
> 
> Indeed I think the ROHM serdes chips do have an address translation
> feature that works pretty much like the TI ones, and the ATR should be
> cleanly reusable across the two brands. The ATR code might be
> simplified to just provide helpers for common code maybe, but I'd
> rather avoid code duplication.

I tend to agree with Luca on keeping the I2C-atr a re-usable generic 
code. The reason why I did not (yet) bring this (or other concerns) up 
is that currently I don't have a permission to upstream my work for ROHM 
SerDes Linux drivers :( I still do discuss the decision but to be 
honest, I believe the chances to get my drivers upstream are getting 
slim. Hence I did not see it reasonable to try pushing Tomi's / Luca's 
work to any direction.

Yet, after that being said - these SerDes devices which can combine I2C 
buses from multiple remote devices to one local I2C controller do bring 
a real need for a solution like the address translation. Thus, I would 
not be surprized if we saw new devices with this kind of feature. 
Potentially also from other vendors. So, regardless what happens with 
ROHM SERDESes - I (too) think we should keep the ATR as independently 
reusable component.

Yours,
	-- Matti Vaittinen

-- 
Matti Vaittinen
Linux kernel developer at ROHM Semiconductors
Oulu Finland

~~ When things go utterly wrong vim users can always type :help! ~~
  

On 07/11/2022 13:48, Luca Ceresoli wrote:

>> In fact, I'm thinking that it might be better to just drop the i2c-atr
>> driver, and add the support directly to the FPDlink drivers. But that
>> could mean possibly duplicating the same code for other deser/ser
>> architectures, so I have kept the i2c-atr driver for now.
> 
> Indeed I think the ROHM serdes chips do have an address translation
> feature that works pretty much like the TI ones, and the ATR should be
> cleanly reusable across the two brands. The ATR code might be
> simplified to just provide helpers for common code maybe, but I'd
> rather avoid code duplication.

I agree. The reason I'm wondering about this is the fact that the i2c 
slave side (deser) is in a different IC and driver than the i2c master 
side and driver (ser). That makes it quite different from the i2c-mux 
(at least how the i2c-mux exists now).

  Tomi
  

On 11/1/22 15:20, Tomi Valkeinen wrote:
> Hi,
> 
> Intro
> -----
> 
> This is, kind of, v4 of Luca's i2c-atr and FPDLink series, v3 of which
> you can find from:
> 
> https://lore.kernel.org/all/20220206115939.3091265-1-luca@lucaceresoli.net/
> 

//snip

First of all, I like this series.

There is just one thing that slightly bothers me here. It is putting 
everything the deserializer has inside one driver. I think I mentioned 
this already earlier - but to me these devices seem like MFD ICs.

Hence, my attempt to support one of SerDes devices resembling the TI 
SerDes devices had following split. MFD I2C drivers for SER and DES. Own 
platform drivers for DES ATR, V4L2, pinctrl/GPIO. Also own platform 
drivers for SER V4L2, pinctrl/GPIO.

Below you can see the relevant pieces of simplified (pseudo-)code 
explaining the (optional?) design which attempts creating somewhat 
smaller and re-usable individual drivers for DES/SER blocks. Please, do 
not treat this as a requirement - please treat it as a food for thoughts ;)


DES MFD-core: as I2C - device. Registers regmap and IRQ-chip for 
sub-devices. Creates the sub-devices for DES. Parses the required SerDes 
topology from device-tree and makes the minimum required configuration 
to get the actual link (FPDLink in TI's case) run in a level that I2C 
connection is possible when ATR is configured. (Actually, I noticed this 
could be put in ATR driver, but in my opinion it's better to parse the 
overall DT in MFD and put it in the MFD driver-data so all sub-devices 
can get the parsed bits of DT data from parent w/o duplicating the 
parsing code).

static struct mfd_cell bu18rm84_mfd_cells[] = {
	{ .name = "bu18rm84-i2c", },
	{ .name = "bu18rm84-video", },
	{ .name = "bu18rm84-pinctrl", },
};

...

static int link_init(struct device *dev, struct regmap *regmap,
		    struct bu18rm84_core_data *data,
		    struct bu18rm84_ser_conf *ser, int ser_id)
{
	...
}

static int prepare_hw_for_subdevs(struct device *dev, struct regmap *regmap,
				  struct bu18rm84_core_data *data)
{
	int ret, i;

	ret = bu18rm84_common_init_pre(dev, regmap, data);
	if (ret)
		return ret;

	for (i = 0; i < BU18RM84_MAX_NUM_SER_NODES; i++) {
		if (data->ser[i].ser_ep) {
			/*
			 * We initialize link between DES => SERsin MFD so
			 * the sub-devices have working link to SER devices.
			 */
			ret = link_init(dev, regmap, data, &data->ser[i], i);
			if (ret)
				return ret;
		}
	}

	...

	return ret;
}

static int connected_devices(struct device *dev, struct 
bu18rm84_core_data *data)
{
	int i;

	for (i = 0; i < BU18RM84_MAX_NUM_SER_NODES; i++) {
		/*
		 * TODO: See where/when/if we should drop the reference to
		 * the node
		 */
		data->ser[i].ser_ep = fwnode_graph_get_endpoint_by_id(data->fw, i,
								  0, 0);
		if (data->ser[i].ser_ep) {
			dev_dbg(dev, "Found node for SER %d\n", i);
			data->num_ser++;
		}
	}


	for (; i < BU18RM84_MAX_NUM_SER_NODES + BU18RM84_MAX_NUM_CSI_NODES; i++) {
		int csi = i - BU18RM84_MAX_NUM_SER_NODES;

		data->csi_ep[csi] = fwnode_graph_get_endpoint_by_id(data->fw, i,
								    0, 0);
		if (data->csi_ep[csi])
			data->num_csi++;
	}
	dev_dbg(dev, "found %d ser, %d csi\n", data->num_ser, data->num_csi);

	return (!data->num_ser) ? -ENODEV : 0;
}

static int bu18rm84_i2c_probe(struct i2c_client *i2c)
{
	data = devm_kzalloc(&i2c->dev, sizeof(*data), GFP_KERNEL);
	regmap = devm_regmap_init_i2c(i2c, &bu18rm84_regmap);

	ret = connected_devices(&i2c->dev, data);

	/*
	 * We need to set-up the link between DES and SER(s) before
	 * the I2C to SER(s) is operational. Sub-devices like the SER GPIO, pinmux
	 * or media do need this I2C so we'd better to prepare things up-to-the
	 * point where I2C works prior registering the subdev cells.
	 *
	 * Unfortunately, this means the MFD needs to go and read & understand
	 * the DT for SER topology - and do some magic settings to get things
	 * up & running.
	 */
	ret = prepare_hw_for_subdevs(&i2c->dev, regmap, data);

	ret = devm_regmap_add_irq_chip(&i2c->dev, regmap, i2c->irq,
				       IRQF_SHARED, 0, &bu18rm84_irq_chip,
				       &irq_data);

	ret = devm_mfd_add_devices(&i2c->dev, PLATFORM_DEVID_AUTO,
				   bu18rm84_mfd_cells,
				   ARRAY_SIZE(bu18rm84_mfd_cells),
				   NULL, 0, regmap_irq_get_domain(irq_data));

}

===============

DES ATR platform driver: Kicked by DES MFD. Uses I2C-atr to configure 
address translation. Creates SER devices and creates the I2C adapter for 
the remote SER buses. This in-turn launches SER MFD driver + remote I2C 
slave devices. And yes. I think the remote I2C slave devices must not be 
created before SER side link configurations are done by SER MFD driver. 
Thanks Tomi - I guess I've been lucky this far (or there is something I 
didn't spot quite yet ;] )

static int bu18rm84_atr_attach_client(struct i2c_atr *atr, u32 chan_id,
				   const struct i2c_board_info *info,
				   const struct i2c_client *client,
				   u16 *alias_id)
{
	/* Configure the ATR for slave to DES registers */
}

static void bu18rm84_atr_detach_client(struct i2c_atr *atr, u32 chan_id,
				    const struct i2c_client *client)
{
	/* Clear the configuration for the ATR for slave at DES registers */
}

static const struct i2c_atr_ops bu18rm84_atr_ops = {
	.attach_client = bu18rm84_atr_attach_client,
	.detach_client = bu18rm84_atr_detach_client,
};

static int bu18rm84_populate_ser(struct bu18rm84_i2c_atr *data, u32 *addrs,
				 int num_addrs, struct fwnode_handle *des_node,
				 int ser_id)
{
	/*
	 * Configure the I2C alias for SER devices in DES registers based on
	 * reg-names in the DT
	 */
	idx = fwnode_property_match_string(des_node, "reg-names",
					   ser_names[ser_id]);
	...
}

static int setup_ser_i2c(struct bu18rm84_i2c_atr *data, int ser_id,
			 struct fwnode_handle *ser_ep)
{
	int ret;
	const char *name;
	struct i2c_client **new = &data->ser_client[ser_id];
	struct i2c_board_info ser_info = { 0 };

	ser_info.fwnode = fwnode_get_named_child_node(ser_ep, "remote-chip");
	name = fwnode_get_name(ser_info.fwnode);

	ser_info.addr = data->ser_alias[ser_id];
	*new = i2c_new_client_device(data->parent_i2c->adapter, &ser_info);

	/* Add adapter to correctly enable i2c to devices behind SER */
	/* We need to have ATR done before this */
	ret = i2c_atr_add_adapter(data->atr, ser_id);
	...

	return ret;
}

static int bu18rm84_setup_ser_i2cs(const struct bu18rm84_core_data *core,
				   struct bu18rm84_i2c_atr *data)
{
	for (i = 0; i < BU18RM84_MAX_NUM_SER_NODES; i++) {
		if (core->ser[i].ser_ep) {
			ret = setup_ser_i2c(data, i, core->ser[i].ser_ep);
			...
		}
	}
}

static int bu18rm84_atr_probe(struct platform_device *pdev)
{
	parent = dev->parent;
	core = dev_get_drvdata(parent);

	data = devm_kzalloc(dev, sizeof(*data), GFP_KERNEL);

	data->parent_i2c = to_i2c_client(parent);

	data->regmap = dev_get_regmap(parent, NULL);

	dev_set_drvdata(&pdev->dev, data);

	/* Go read the I2C addresses from DT */
	ret = bu18rm84_i2c_parse_dt(data, core);

	/*
	 * Create device for SER. We won't add SER to ATR in order to avoid
	 * messing with the SER specific aliases in ATR.
	 *
	 * Also, add adapter.
	 */
	for (i = 0; i < BU18RM84_MAX_NUM_SER_NODES; i++) {
		if (core->ser[i].ser_ep) {
			ret = bu18rm84_populate_ser(data, &addrs[0], num_addrs,
						    core->fw, i);
		}
	}

	data->atr = i2c_atr_new(data->parent_i2c->adapter, parent,
				&bu18rm84_atr_ops, 4);

	i2c_atr_set_clientdata(data->atr, data);

	return bu18rm84_setup_ser_i2cs(core, data);

=============================

Ser MFD (i2c)driver. Kicked by DES ATR platform device when creating the 
I2C devices with the SER MFD fwnode. Finalizes the SER <=> DES I2C 
connection and kicks up the SER slave devices for DES pinctrl / V4L2 or 
others.


static struct mfd_cell bu18tm41_mfd_cells[] = {
	{ .name = "bu18tm41-video", },	/* Video forwarding */
	{ .name = "bu18tm41-pinctrl", }, /* pinmux, pinconf and gpio */
};

static int bu18tm41_conf_ser(struct device *dev, struct regmap *regmap,
			     struct bu18tm41_core *core)
{
	fw = dev_fwnode(dev);

	ret = fwnode_property_read_u64(fw, "compulsory stuff to get link up",
				       &foo);

	ret = bu18tm41_enable_link( ... );

	...

	return 0;
}

static int bu18tm41_i2c_probe(struct i2c_client *i2c)
{
	regmap = devm_regmap_init_i2c(i2c, &bu18tm41_regmap);
	core = devm_kzalloc(&i2c->dev, sizeof(*core), GFP_KERNEL);
	dev_set_drvdata(&i2c->dev, core);

	/*
  	 * Do the minimum required link config on SER config to get I2C
  	 * between DES <=> SER up
	 *
	 * Unfortunately, this means the MFD needs to do some magic settings
	 * to get things up & running.
	 */
	ret = bu18tm41_conf_ser(&i2c->dev, regmap, core);

	ret = devm_mfd_add_devices(&i2c->dev, PLATFORM_DEVID_AUTO,
				   bu18tm41_mfd_cells,
				   ARRAY_SIZE(bu18tm41_mfd_cells),
				   NULL, 0, /*regmap_irq_get_domain(irq_data) */ NULL);

	return ret;
}

===========================

Des media platform driver. V4L2/media stuff.

...

static int bu18rm84_media_probe(struct platform_device *pdev)
{

	core = dev_get_drvdata(dev->parent);

	data = devm_kzalloc(dev, sizeof(*data), GFP_KERNEL);

	data->dev = dev;
	data->parent_i2c = to_i2c_client(dev->parent);
	data->regmap = dev_get_regmap(dev->parent, NULL);

	platform_set_drvdata(pdev, data);


	ret = bu18rm84_parse_csi_dt(core, data);

	ret = bu18rm84_config_csi2(data);

	ret = bu18rm84_add_ser_data(core, data);

	...

	/* Create the V4L2 subdevices */
	ret = bu18rm84_create_subdev(core, data);
	if (ret)
		goto err_subdev;

	...
}

SER video, Des/Ser pinctrl/GPIO platform drivers:
...
  (Like usual)


I think it was Tomi who asked me the benefit of using MFD. In some cases 
the digital interface towards pinctrl/GPIO or other functional blocks in 
SER/DES is re-used from other products - or the blocks are re-used on 
other products. Separating them in own platform-drivers is a nice way to 
re-use drivers and avoid code duplication.

Whether it is worth or correct can be evaluated by brighter minds :) As 
I said, I am in a no way demanding a change if you see the current 
proposal being better! I am just presenting an optional way to feed your 
thoughts ;)

Ps. Sorry for _very_ stripped down code in this explanation. I am 
currently not allowed to publish any details of the ICs these drivers 
are written for.

Yours
	-- Matti

-- 
Matti Vaittinen
Linux kernel developer at ROHM Semiconductors
Oulu Finland

~~ When things go utterly wrong vim users can always type :help! ~~
  

Hi Matti,

On 07/11/2022 16:37, Vaittinen, Matti wrote:

I only had time to have a brief look at your code, but I have a few 
quick questions.

> I think it was Tomi who asked me the benefit of using MFD. In some cases
> the digital interface towards pinctrl/GPIO or other functional blocks in
> SER/DES is re-used from other products - or the blocks are re-used on
> other products. Separating them in own platform-drivers is a nice way to
> re-use drivers and avoid code duplication.

Is there anything that prevents us (or makes it difficult) from 
refactoring a "monolithic" driver into an MFD later? If we see such IP 
re-use, can we then move to an MFD?

I admit I have never written an MFD driver (but I have hacked with a few 
years back). As I see it, the "subcomponents" in FPDLink ICs are more or 
less tied to the FPDLink. It's not like they're independent. Compare to, 
for example, a PMIC which provides regulators and GPIOs, and possibly 
the only shared part between those two features are the pins.

So, I think I'm still wondering about the benefit...

In the current version I have the deser driver supporting UB960 and 
UB9702. I guess I could split those into separate drivers, and using MFD 
I could share lot of code between them. But I still can't see why that's 
better than having both UB960 and UB9702 in the same driver (and, if the 
amount of supported devices increases, perhaps dividing some parts to 
separate files and using function points for a few things).

The benefit would be more obvious if there was some other type of IC 
that uses the same IP subcomponents. Maybe the display side FPDLink 
devices are such, but I have never done a deep dive in their 
documentation. And, even then, I think I still have the question: can we 
just move to an MFD later when the need comes?

Also, isn't the use or non-use of MFD strictly a driver private thing, 
it should not affect any shared parts or shared designs? In other words, 
if you have your ROHM hat on, why do you care how the UB9xx driver is 
structured internally? ;)

  Tomi
  

Evening Tomi,

On 11/7/22 17:37, Tomi Valkeinen wrote:
> Hi Matti,
> 
> On 07/11/2022 16:37, Vaittinen, Matti wrote:
> 
> I only had time to have a brief look at your code, but I have a few 
> quick questions.
> 
>> I think it was Tomi who asked me the benefit of using MFD. In some cases
>> the digital interface towards pinctrl/GPIO or other functional blocks in
>> SER/DES is re-used from other products - or the blocks are re-used on
>> other products. Separating them in own platform-drivers is a nice way to
>> re-use drivers and avoid code duplication.
> 
> Is there anything that prevents us (or makes it difficult) from 
> refactoring a "monolithic" driver into an MFD later? If we see such IP 
> re-use, can we then move to an MFD?

I haven't done such conversion. I think the work for doing the 
conversion at later phase is roughly same it would be now. However, 
synchronizing such change with related subsystem trees might be some 
extra work.

> I admit I have never written an MFD driver (but I have hacked with a few 
> years back). As I see it, the "subcomponents" in FPDLink ICs are more or 
> less tied to the FPDLink. It's not like they're independent. Compare to, 
> for example, a PMIC which provides regulators and GPIOs, and possibly 
> the only shared part between those two features are the pins.

I think that in SerDes driver case the benefit may come from re-use and 
clarity. Smaller drivers tend to be easier to comprehend, although I 
liked the way you had divided the drivers in sections.

> So, I think I'm still wondering about the benefit...
> 
> In the current version I have the deser driver supporting UB960 and 
> UB9702. I guess I could split those into separate drivers,

I wouldn't break the driver per IC. If the ICs are similar enough to be 
nicely handled by same driver, then they probably should.

> 
> The benefit would be more obvious if there was some other type of IC 
> that uses the same IP subcomponents.

Yes. Same or similar subcomponents. This indeed is the benefit I see. I 
don't know if TI could use same - say GPIO - control logic on another 
type of device? Or, maybe separating the logic could guide one to use 
some generic stuff like regmap_gpio driver? And finally, submitting 
small platform drivers via respective subsystems can yield better review ;)

> 
> Also, isn't the use or non-use of MFD strictly a driver private thing, 

I am tempted to say "yes", but when giving it a thought - it's really 
not fully that. Splitting a driver to subdrivers can allow re-use of 
subcomponents by unrelated ICs. OTOH, always stuffing everything in a 
driver "because it is driver internal decision" can lead to code 
duplication and bloat.

> it should not affect any shared parts or shared designs? In other words, 
> if you have your ROHM hat on, why do you care how the UB9xx driver is 
> structured internally? ;)

As I wrote:
 > > Please, do not treat this as a requirement - please treat it as a 
food for thoughts ;)

Eg, I am not trying to tell you how to do the UB9xx drivers. I just 
think that considering another design _might_ result more optimal design 
- but I do leave the decision to you who know this area better than I do.

Yours
	-- Matti

-- 
Matti Vaittinen
Linux kernel developer at ROHM Semiconductors
Oulu Finland

~~ When things go utterly wrong vim users can always type :help! ~~