Re: [PATCH RFC v3 1/4] dt-bindings: mux: Increase the number of arguments in mux-controls

From: Aswath Govindraju
Date: Mon Nov 29 2021 - 07:58:12 EST


Hi Peter,

On 29/11/21 5:58 pm, Peter Rosin wrote:
> Hi Aswath,
>
> On 2021-11-29 10:31, Aswath Govindraju wrote:
>> Hi Peter,
>>
>> On 29/11/21 1:45 pm, Peter Rosin wrote:
>>>
>>>
>>> On 2021-11-29 05:36, Aswath Govindraju wrote:
>>>> Hi Peter,
>>>>
>>>> On 25/11/21 7:05 pm, Peter Rosin wrote:
>>>>> Hi!
>>>>>
>>>>> You need to have some description on how #mux-control-cells now work.
>>>>> The previous description is in mux-consumer.yaml and an update there
>>>>> is needed.
>>>>>
>>>>> However, I have realized that the adg792a binding uses #mux-control-cells
>>>>> to indicate if it should expose its three muxes with one mux-control
>>>>> and operate the muxes in parallel, or if it should be expose three
>>>>> independent mux-controls. So, the approach in this series to always
>>>>> have the #mux-control-cells property fixed at <2> when indicating a
>>>>> state will not work for that binding. And I see no fix for that binding
>>>>> without adding a new property.
>>>>>
>>>>> So, I would like a different approach. Since I dislike how mux-controls
>>>>> -after this series- is not (always) specifying a mux-control like the name
>>>>> says, but instead optionally a specific state, the new property I would
>>>>> like to add is #mux-state-cells such that it would always be one more
>>>>> than #mux-control-cells.
>>>>>
>>>>> mux: mux-controller {
>>>>> compatible = "gpio-mux";
>>>>> #mux-control-cells = <0>;
>>>>> #mux-state-cells = <1>;
>>>>>
>>>>> mux-gpios = <...>;
>>>>> };
>>>>>
>>>>> can-phy {
>>>>> compatible = "ti,tcan1043";
>>>>> ...
>>>>> mux-states = <&mux 1>;
>>>>> };
>>>>>
>>>>> That solves the naming issue, the unused argument for mux-conrtrollers
>>>>> that previously had #mux-control-cells = <0>, and the binding for adg792a
>>>>> need no longer be inconsistent.
>>>>>
>>>>> Or, how should this be solved? I'm sure there are other options...
>>>>>
>>>>
>>>>
>>>> I feel that the new approach using mux-state-cells seems to be
>>>> overpopulating the device tree nodes, when the state can be represented
>>>> using the control cells. I understand that the definition for
>>>> mux-controls is to only specify the control line to be used in a given
>>>> mux. Can't it now be upgraded to also represent the state at which the
>>>> control line has to be set to?
>>>>
>>>> With respect to adg792a, it is inline with the current implementation
>>>> and the only change I think would be required in the driver is,
>>>
>>> No, that does not work. See below.
>>>
>>>>
>>>> diff --git a/drivers/mux/adg792a.c b/drivers/mux/adg792a.c
>>>> index e8fc2fc1ab09..2cd3bb8a40d4 100644
>>>> --- a/drivers/mux/adg792a.c
>>>> +++ b/drivers/mux/adg792a.c
>>>> @@ -73,8 +73,6 @@ static int adg792a_probe(struct i2c_client *i2c)
>>>> ret = device_property_read_u32(dev, "#mux-control-cells", &cells);
>>>> if (ret < 0)
>>>> return ret;
>>>> - if (cells >= 2)
>>>> - return -EINVAL;
>>>>
>>>> mux_chip = devm_mux_chip_alloc(dev, cells ? 3 : 1, 0);
>>>
>>> When you add cell #2 with the state, the cells variable will end up
>>> as 2 always. Which means that there is no way to alloc one mux
>>> control since "cells ? 3 : 1" will always end up as "3", with no
>>> easy fix.
>>>
>>> So, your approach does not work for this driver.
>>>
>>
>> Sorry, but how is this different from the case of
>>
>> #mux-control-cells = 1
>>
>> If #mux-control-cells is equal to 1 it means the consumer will use a
>> given control line from the mux chip. The same would be the case when we
>> will be using, #mux-control-cells is equal to 2, except we can also
>> provide the state.
>>
>> If the consumer will use all the lines then #mux-control-cells will be
>> set to 0. In this condition the state can't be provided from the DT and
>> the consumer will be controlling the entire mux chip. If
>> #mux-control-cells is greater than 0 then we will not be able to provide
>> multiple lines of control using a single mux-controls entry(mux-controls
>> = <...>;) right? We would have the using multiple mux-controls
>> entries(mux-controls = <...>, <...>;).
>
> I think you misunderstand. The adg792a driver operates the chip in
> different modes depending on if you specify 0 or 1 cells. With 0,
> it's not just that the consumer operates three muxes. It is also, and
> more importantly, that the three muxes are operated in parallel without
> the consumer doing anything different with the single mux control it
> sees (even if there are three muxes operated by that single mux
> control).
>
> That said, yes, you can make it limp along like you describe above.
> But why should it not be possible to specify a specific state when
> the adg792a driver operates the muxes in parallel? And yes, you could
> add some other flag to indicate this mode, but my point is that it
> is silly to add special cases like this if you don't need to. Since
> adding a specific state is the new thing, that is what should be
> added in a way that fits with the old stuff without imposing new
> flags on that old stuff.
>
> An example: the three muxes in an adg792a chip could be used as
> two muxes for some I2C bus (SCL and SDA signals) and the third mux
> for something unrelated. Suppose that you want to operate the adg792a
> as three parallel muxes so that you mux SCL and SDA simultaneously
> (as is expected by the i2c-mux-gpmux binding, it only expects one
> mux control), and that you want to use the third mux as the enable-
> state for your phy. With your suggested binding you cannot, unless
> you add a mechanism to make the adg792a driver operate its muxes in
> parallel even if there are two cells instead of zero. I.e. without
> that new flag the i2c-mux-gpmux binding needs to see
>
> #mux-control-cells = <0>;
>
> while your new phy binding instead needs to see
>
> #mux-control-cells = <2>;
>
> And you obviously can't have it both ways.
>
> (Sure, it would not be possible to mux the I2C bus while the phy
> is enabled in the above example, but there could be some other
> limitation in place that makes that invalid anyway. And it's just
> an example anyway...)
>
> A mux-control is potentially a shared resource, and bindings have
> to take this into account.
>

Understood. Thank you for the explanation. Will correct the
implementation and post a respin.

Regards,
Aswath

> Cheers,
> Peter
>