Re: [PATCH v2 3/3] hwmon: ltc2990: support all measurement modes

[Tom Levens]

On 17 Nov 2016, at 22:54, Guenter Roeck <linux@xxxxxxxxxxxx> wrote:
> On Thu, Nov 17, 2016 at 08:52:12PM +0100, Mike Looijmans wrote:
>> ïOn 17-11-2016 19:56, Guenter Roeck wrote:
>>> On Thu, Nov 17, 2016 at 06:40:17PM +0100, Mike Looijmans wrote:
>>>> ïOn 17-11-16 17:56, Guenter Roeck wrote:
>>>>> On 11/17/2016 04:10 AM, Tom Levens wrote:
>>>>>> Updated version of the ltc2990 driver which supports all measurement
>>>>>> modes available in the chip. The mode can be set through a devicetree
>>>>>> attribute.
>>>>> 
>>> [ ... ]
>>> 
>>>>>> 
>>>>>> static int ltc2990_i2c_probe(struct i2c_client *i2c,
>>>>>>                 const struct i2c_device_id *id)
>>>>>> {
>>>>>>    int ret;
>>>>>>    struct device *hwmon_dev;
>>>>>> +    struct ltc2990_data *data;
>>>>>> +    struct device_node *of_node = i2c->dev.of_node;
>>>>>> 
>>>>>>    if (!i2c_check_functionality(i2c->adapter,
>>>>>> I2C_FUNC_SMBUS_BYTE_DATA |
>>>>>>                     I2C_FUNC_SMBUS_WORD_DATA))
>>>>>>        return -ENODEV;
>>>>>> 
>>>>>> -    /* Setup continuous mode, current monitor */
>>>>>> +    data = devm_kzalloc(&i2c->dev, sizeof(struct ltc2990_data),
>>>>>> GFP_KERNEL);
>>>>>> +    if (unlikely(!data))
>>>>>> +        return -ENOMEM;
>>>>>> +    data->i2c = i2c;
>>>>>> +
>>>>>> +    if (!of_node || of_property_read_u32(of_node, "lltc,mode",
>>>>>> &data->mode))
>>>>>> +        data->mode = LTC2990_CONTROL_MODE_DEFAULT;
>>>>> 
>>>>> Iam arguing with myself if we should still do this or if we should read
>>>>> the mode
>>>>> from the chip instead if it isn't provided (after all, it may have been
>>>>> initialized
>>>>> by the BIOS/ROMMON).
>>>> 
>>>> I think the mode should be explicitly set, without default. There's no way
>>>> to tell whether the BIOS or bootloader has really set it up or whether the
>>>> chip is just reporting whatever it happened to default to. And given the
>>>> chip's function, it's unlikely a bootloader would want to initialize it.
>>>> 
>>> Unlikely but possible. Even if we all agree that the chip should be configured
>>> by the driver, I don't like imposing that view on everyone else.
>>> 
>>>> My advice would be to make it a required property. If not set, display an
>>>> error and bail out.
>>>> 
>>> It is not that easy, unfortunately. It also has to work on a non-devicetree
>>> system. I would not object to making the property mandatory, but we would
>>> still need to provide non-DT support.
>>> 
>>> My "use case" for taking the current mode from the chip if not specified
>>> is that it would enable me to run a module test with all modes. I consider
>>> this extremely valuable.
>> 
>> Good point.
>> 
>> The chip defaults to measuring internal temperature only, and the mode
>> defaults to "0".
>> 
>> Choosing a mode that doesn't match the actual circuitry could be bad for the
>> chip or the board (though unlikely, it'll probably just be useless) since it
>> will actively drive some of the inputs in the temperature modes (which is
>> default for V3/V4 pins).
>> 
>> Instead of failing, one could choose to set the default mode to "7", which
>> just measures the 4 voltages, which would be a harmless mode in all cases.
>> 
>> As a way to let a bootloader set things up, I think it would be a good check
>> to see if CONTROL register bits 4:3 are set. If "00", the chip is not
>> acquiring data at all, and probably needs configuration still. In that case,
>> the mode must be provided by the devicetree (or the default "7").
>> If bits 4:3 are "11", it has already been set up to measure its inputs, and
>> it's okay to continue doing just that and use the current value of 2:0
>> register as default mode (if the devicetree didn't specify any mode at all).
>> 
> 
> At first glance, agreed, though by default b[3:4] are 00, and only the
> internal temperature is measured. Actually, the 5 mode bits are all
> relevant to determine what is measured. Maybe it would be better to take
> all 5 bits into account instead of blindly setting b[34]:=11 and a specific
> setting of b[0:2]. Sure, that would make the mode table a bit larger,
> but then ltc2990_attrs_ena[] could be made an u16 array, and a table size
> of 64 bytes would not be that bad.

I would tend to agree that it should be possible to configure all 5 mode
bits through the devicetree. What I would propose is as follows.

If a devicetree node exists, the mode parameter(s?) are required and the 
chip is initialised.

If a devicetree node doesn't exist, it is assumed that the chip has 
already been configured (by the BIOS, etc). The mode is read from the 
chip to set the visibility of the sysfs attributes. In the worst case, where the 
chip has not been configured by another source, it would only be possible
to measure the internal temperature -- but I think this is an acceptable
limitation.

The only case that this does not cover is if the device tree node 
exists but the chip is expected to be configured by some other source. 
Maybe I am wrong, but I would not expect this to be a terribly common
situation.

What do you think?

Cheers,
Tom