Re: [PATCH v3 2/2] hwmon: acpi_power_meter: convert to hwmon_device_register_with_info

[Guenter Roeck <linux@roeck-us.net>]

On 5/15/22 23:21, LABBE Corentin wrote:
> Le Sun, May 15, 2022 at 05:29:54PM -0700, Guenter Roeck a écrit :
>> On 5/15/22 12:36, LABBE Corentin wrote:
>>> Le Wed, May 11, 2022 at 07:10:29PM -0700, Guenter Roeck a écrit :
>>>> Corentin,
>>>>
>>>> On 5/8/22 23:30, Corentin Labbe wrote:
>>>>> Booting lead to a hwmon_device_register() is deprecated. Please convert the driver to use hwmon_device_register_with_info().
>>>>> So let's convert the driver to use hwmon_device_register_with_info().
>>>>>
>>>>> Signed-off-by: Corentin Labbe <clabbe@baylibre.com>
>>>>> ---
>>>> [ ... ]
>>>>
>>>>> @@ -836,20 +740,20 @@ static void acpi_power_meter_notify(struct acpi_device *device, u32 event)
>>>>>     		if (res)
>>>>>     			break;
>>>>>     
>>>>> -		remove_attrs(resource);
>>>>> +		remove_domain_devices(resource);
>>>>>     		setup_attrs(resource);
>>>>
>>>> Zhang Rui found an interesting problem with this code:
>>>> It needs a call to sysfs_update_groups(hwmon_dev->groups)
>>>> to update sysfs attribute visibility, probably between
>>>> remove_domain_devices() and setup_attrs().
>>>>
>>>>>     		break;
>>>>>     	case METER_NOTIFY_TRIP:
>>>>> -		sysfs_notify(&device->dev.kobj, NULL, POWER_AVERAGE_NAME);
>>>>> +		hwmon_notify_event(&device->dev, hwmon_power, hwmon_power_average, 0);
>>>>
>>>> ... which makes realize: The notification device should be the hwmon device.
>>>> That would be resource->hwmon_dev, not the acpi device.
>>>>
>>>
>>> Hello
>>>
>>> Since my hardware lacks capabilities testing this, I have emulated it on qemu:
>>> https://github.com/montjoie/qemu/commit/320f2ddacb954ab308ef699f66fca6313f75bc2b
>>>
>>> I have added a custom ACPI _DBX method for triggering some ACPI state change. (like config change, like enabling CAP).
>>>
>>> For testing config change I have tried lot of way:
>>>                   res = read_capabilities(resource);
>>> @@ -742,18 +758,22 @@ static void acpi_power_meter_notify(struct acpi_device *device, u32 event)
>>>    
>>>                   remove_domain_devices(resource);
>>>                   setup_attrs(resource);
>>> +               res = sysfs_update_groups(&resource->hwmon_dev->kobj, acpi_power_groups);
>>> +               res = sysfs_update_groups(&resource->acpi_dev->dev.kobj, acpi_power_groups);
>>> +               res = hwmon_notify_event(resource->hwmon_dev, hwmon_power, hwmon_power_cap, 0);
>>> +               res = hwmon_notify_event(resource->hwmon_dev, hwmon_power, hwmon_power_average, 0);
>>
>> Did you add a debug log here ?
> 
> Yes I added debug log to check what is called.
> 
>>
>> acpi_power_groups would be the wrong parameter for sysfs_update_groups().
>> It would have to be resource->hwmon_dev->groups.
>>
> 
> Even with that, no call to is_visible:
> @@ -742,18 +758,22 @@ static void acpi_power_meter_notify(struct acpi_device *device, u32 event)
>   
>                  remove_domain_devices(resource);
>                  setup_attrs(resource);
> +               res = sysfs_update_groups(&resource->hwmon_dev->kobj, resource->hwmon_dev->groups);
> +               res = sysfs_update_groups(&resource->acpi_dev->dev.kobj, resource->hwmon_dev->groups);
> +               res = hwmon_notify_event(resource->hwmon_dev, hwmon_power, hwmon_power_cap, 0);
> +               res = hwmon_notify_event(resource->hwmon_dev, hwmon_power, hwmon_power_average, 0);
>                  break;
> 
> I checked drivers/hwmon/hwmon.c is seems that is_visible is only called by gen_attr/gen_attrs which is only called by __hwmon_create_attrs and then by registers functions.
> So perhaps it explain why it is never called.

Ah yes, you are correct. Sorry, it has been too long ago that I wrote that code.
Effectively that means we'll have to rework the hwmon core to generate attributes
anyway and leave it up to the driver core to call the is_visible function.

Guenter