Re: [PATCH 2/7] dt-bindings: mfd: ds90ux9xx: add description of TI DS90Ux9xx I2C bridge

[Vladimir Zapolskiy <vz@xxxxxxxxx>]

Hi Luca,

On 11/03/2018 11:00 PM, Luca Ceresoli wrote:
> Hi Vladimir,
> 
> On 31/10/18 21:12, Vladimir Zapolskiy wrote:
>> Hi Luca,
>>
>> thank you for review.
>>
>> On 10/30/2018 06:43 PM, Luca Ceresoli wrote:
>>> Hi Vladimir,
>>>
>>> On 08/10/18 23:12, Vladimir Zapolskiy wrote:
>>>> From: Vladimir Zapolskiy <vladimir_zapolskiy@xxxxxxxxxx>
>>>>
>>>> TI DS90Ux9xx de-/serializers are capable to route I2C messages to
>>>> I2C slave devices connected to a remote de-/serializer in a pair,
>>>> the change adds description of device tree bindings of the subcontroller
>>>> to configure and enable this functionality.
>>>>
>>>> Signed-off-by: Vladimir Zapolskiy <vladimir_zapolskiy@xxxxxxxxxx>
>>>> ---
>>>>  .../bindings/mfd/ti,ds90ux9xx-i2c-bridge.txt  | 61 +++++++++++++++++++
>>>>  1 file changed, 61 insertions(+)
>>>>  create mode 100644 Documentation/devicetree/bindings/mfd/ti,ds90ux9xx-i2c-bridge.txt
>>>>
>>>> diff --git a/Documentation/devicetree/bindings/mfd/ti,ds90ux9xx-i2c-bridge.txt b/Documentation/devicetree/bindings/mfd/ti,ds90ux9xx-i2c-bridge.txt
>>>> new file mode 100644
>>>> index 000000000000..4169e382073a
>>>> --- /dev/null
>>>> +++ b/Documentation/devicetree/bindings/mfd/ti,ds90ux9xx-i2c-bridge.txt
>>>> @@ -0,0 +1,61 @@
>>>> +TI DS90Ux9xx de-/serializer I2C bridge subcontroller
>>>> +
>>>> +Required properties:
>>>> +- compatible: Must contain a generic "ti,ds90ux9xx-i2c-bridge" value and
>>>> +	may contain one more specific value from the list:
>>>> +	"ti,ds90ux925-i2c-bridge",
>>>> +	"ti,ds90ux926-i2c-bridge",
>>>> +	"ti,ds90ux927-i2c-bridge",
>>>> +	"ti,ds90ux928-i2c-bridge",
>>>> +	"ti,ds90ux940-i2c-bridge".
>>>> +
>>>> +Required properties of a de-/serializer device connected to a local I2C bus:
>>>> +- ti,i2c-bridges: List of phandles to remote de-/serializer devices with
>>>> +	two arguments: id of a local de-/serializer FPD link and an assigned
>>>> +	I2C address of a remote de-/serializer to be accessed on a local
>>>> +	I2C bus.
>>>> +
>>>> +Optional properties of a de-/serializer device connected to a local I2C bus:
>>>> +- ti,i2c-bridge-maps: List of 3-cell values:
>>>> +	- the first argument is id of a local de-/serializer FPD link,
>>>> +	- the second argument is an I2C address of a device connected to
>>>> +	  a remote de-/serializer IC,
>>>> +	- the third argument is an I2C address of the remote I2C device
>>>> +	  for access on a local I2C bus.
>>>
>>> BTW I usually use names "remove slave" address and "alias" for bullets 2
>>> and 3. These are the names from the datasheets, and are clearer IMO.
>>>
>>
>> Definitely you are correct, I find that verbose descriptions might be
>> more appropriate and self-explanatory for anyone, who is not closely familiar
>> with the IC series. I'll consider to add the names from the datasheets
>> as well.
>>
>>> Now to the big stuff.
>>>
>>> I find a static map in the "local" chip DT node is a limit. You might
>>> have to support multiple models of remote device, where you'll know the
>>> model only when after it gets connected. Think Beaglebone capes, but
>>> over FPD-Link 3. This scenario opens several issues, but specifically
>>> for I2C address mapping I addressed it by adding in the "local" chip's
>>> DT node a pool of I2C aliases it can use. The DT author is responsible
>>> to pick addresses that are not used on the same I2C bus, which cannot be
>>> done at runtime reliably.
>>
>> Here I see several important topics raised.
>>
>> 1) A static map in the "local" chip DT node is not a limit in sense that
>>    it is optional, so it would be a working model just to omit the property,
>>    however it may (or may not) require another handlers to bridge remote
>>    I2C devices, for instance 'ti,i2c-bridge-pass-all' property, or new
>>    UAPI.
> 
> Do you mean when the "pass-all" method and the "static map in the local
> node" method are insufficient, then the solution is to implement a third
> method that is powerful enough? If that's what you mean, then I think we
> should rather [try to] implement from the beginning a method that is
> powerful enough to handle all the cases we can foresee.

To my best knowledge "static map in the local node" method from my design
is completely sufficient alone. On top of it, if hardware design of a series
of remote PCBs is thought-out enough, then adding DT overlays allows to
manage connections and instantiatings of unknown in advance remote PCBs
in runtime, but this is out of scope of classic IC device tree design,
because device trees are for description of non-discoverable hardware, thus
clearly a complex PCB behind an FPD-Link III shall find its description
in DTB in more or less precise way.

PASS ALL setting may be helpful for testing or adding initial support
of remote PCBs, by the way I've recollected that ICs support simultaneous
slave/alias mappings and PASS ALL function, the practical sense is not
clear though.

>> 2) About supporting multiple models of remote PCBs in the same dts file,
>>    it might be an excessive complication to predict a proper description
>>    of an unknown in advance complex device, so, a better solution should
>>    be to apply DT overlays in runtime, but at any time the hardware
>>    description and the mapping shall be precisely defined.
> 
> I agree runtime DT overlays is the correct way of handling multiple
> remote boards, but not if they contain a ti,i2c-bridge-maps
> (physical-to-alias I2C address map).
> 
> Consider the case where we have N different main board (with SoC and
> "local" [de]serializer) and M peripheral boards (display+deserializer or
> sensor+serializer). Then we'd need up to N*M DT overlays, each with the
> alias map for a specific <main_board, peripheral_board> pair. This is
> because the map maps a physical address that exists on the remote side
> (peripheral board) to an alias that exists on the local side (main board).

If you take into account my statement above about thought-out PCB design,
then multiplication is not needed, but in general you are correct.

And again, device trees are used for description of any non-discoverable
hardware, both main boards and peripheral boards fall into this category,
so the argument about complexity is quite vain, think of an option to
reduce a number of, let say, arch/arm/boot/dts/imx6*.dts files in 3 times
and still get all boards properly supported.

So, DT overlays is the only right method to deal with complexity of
M peripheral boards.

> To realistically model the hardware, the DT overlay should contain the
> removable components, including the physical address on the remote side
> (e.g. touchscreen slave address), but not including the alias on the
> local side (a slave address that is not used on the SoC i2c bus).
> 

The alias on the local side is also a hardware description in sense
that 

>> 3) About a pool of vacant I2C addresses, I dislike the idea that there
>>    will be no definite or constant I2C address in runtime for a particular
>>    remote slave device.
> 
> Indeed that's a little annoyance. But you'll face it only during kernel
> development and debugging, not at the userspace interface. See below the
> discussion about i2c_adapter.
> 

It depends, most likely the good written userspace applications can
deal with the remote side devices correctly, but, for instance, I can't
instantly imagine how to distinguish two identical eeproms on a peripheral
boards with different I2C addresses. My method allows to overcome this
kind of problems.

What I think, is that 

> [...]
> 
>>> Here's my current draft on a dual/quad port deserializer:
>>>
>>> &i2c0 {
>>>     serializer@3d {
>>>         reg = <0x3d>;
>>>         ...
>>>
>>>         /* Guaranteed not physically present on i2c0 */
>>>         i2c-alias-pool = /bits/ 8 <0x20 0x21 0x22 0x23 0x24 0x25>;
>>>
>>>         rxports {
>>>             #address-cells = <1>;
>>>             #size-cells = <0>;
>>>
>>>             rxport@0 {
>>>                 reg = <0>;
>>>                 remote-i2c-bus { /* The proxied I2C bus on rxport 0 */
>>>                     #address-cells = <1>;
>>>                     #size-cells = <0>;
>>>
>>>                     eeprom@51 {
>>>                         reg = <0x51>;
>>>                         compatible = "at,24c02";
>>>                     };
>>>                 };
>>>
>>>             rxport@1 {
>>>                 reg = <1>;
>>>                 remote-i2c-bus { /* The proxied I2C bus on rxport 1 */
>>>                     #address-cells = <1>;
>>>                     #size-cells = <0>;
>>>
>>>                     eeprom@51 {
>>>                         reg = <0x51>;
>>>                         compatible = "at,24c02";
>>>                     };
>>>                 };
>>>             };
>>>         };
>>>     };
>>> };
>>>
>>> At probe time the serializer driver instantiates one new i2c_adapter for
>>> each rxport. Any remote device is added (removed) to that adapter, then
>>> the driver finds an available alias and maps (unmaps) it. The
>>
>> I avoid using i2c_adapter object, because then you get a confusing access
>> to right the same device on two logical I2C buses. This is not the way
>> how I2C muxes operate or are expected to operate, commonly I2C muxes contain
>> a protocol to access muxed devices, which are "invisible" on a host bus,
>> and here a local IC behaves like an I2C device with multiple addresses.
> 
> Sorry, this whole paragraph is not quite clear to me.
> 
> Well, except "I avoid using i2c_adapter object", which is very clear and
> opposite to my approach. :) I'm instantiating an i2c_adapter for each rx
> port because it's the best model of the real world that I could devise.

If you use an i2c_adapter in the discussed case, it allows to discover and
access right the same I2C device on two logical buses at the same time.

> See the picture:
> 
>                 .----------------.
>                 |                |   i2c-5    ,-- touch@50
> ----.   i2c-0   |        adap0----------------+-- eeprom@51
> SoC |-----------| atr@3d         |
> ----'           |        adap1----------------+-- eeprom@51
>                 |                |   i2c-6    `-- touch@50
>                 `----------------'
> 
> I called the central object ATR (address translator), which is the i2c
> remotizer feature in the DS90Ux chipset. Topologically it is somewhat
> similar to a mux, but:
> 
>  * it also translates addresses (hence the name)
>  * electrically it never attaches the upstream and downstream busses,
>    they are driven independently, with message buffering etc.

To my point above, "eeprom@51" device will be accessible both on
i2c-0 and i2c-5, this is something that I want to avoid by any means.

> The touch and eeprom in the example are instantiated on the physical
> busses where they are wired (i2c-5 or i2c-6) with their physical
> addresses on those busses.

At the cost of disadvantage that touch and eeprom are discoverable on
i2c-0 also. The same device mapped into two logical I2C busses at the
same time does not sound correct or even thoroughly tested, sorry.

>> Note, that following an advice from Wolfram I'm going to send the i2c-bridge
>> cell driver into inclusion under drivers/i2c/muxes/ , even if the device
>> driver does not register a mux.
>>
>>> transactions are handled in a way similar to i2c-mux, i.e. the ds90*
>>> i2c_adapter has a master_xfer callback that changes the remote slave
>>> address to the corresponding alias, then calls parent->algo->master_xfer().
>>>
>>> Note how both eeproms in the example have the same physical address.
>>> They will be given two different aliases.
>>
>> The question is how to determine which runtime assigned address represents
>> which eeprom of two. The remote/alias scheme I propose makes it transparent.
> 
> You generally don't care about the assigned alias. To access e.g. each
> of the eeproms the path is always reliable and represents the physical
> world:
> 
>  - i2c-0 -> 00-003d -> rxport0 -> ??-0051
>  - i2c-0 -> 00-003d -> rxport1 -> ??-0051
> 
> The alias matters only when debugging the driver or inspecting the
> physical bus with a protocol analyzer or oscilloscope.
> 

Please see my point about userspace interfaces. udev won't tell if mapped
eeprom@20 is actually an eeprom@50 or eeprom@51 device on a peripheral
board. This uncertainty shall be avoided, my DT design removes the riddle.


--
Best wishes,
Vladimir