Re: [PATCH RFC 0/6] power: supply: extension API

[Armin Wolf <W_Armin@gmx.de>]

Am 07.06.24 um 12:26 schrieb Thomas Weißschuh:

> On 2024-06-07 01:10:02+0000, Armin Wolf wrote:
>> Am 06.06.24 um 16:50 schrieb Thomas Weißschuh:
>>
>>> Introduce a mechanism for drivers to extend the properties implemented
>>> by a power supply.
>>>
>>> Motivation
>>> ----------
>>>
>>> Various drivers, mostly in platform/x86 extend the ACPI battery driver
>>> with additional sysfs attributes to implement more UAPIs than are
>>> exposed through ACPI by using various side-channels, like WMI,
>>> nonstandard ACPI or EC communication.
>>>
>>> While the created sysfs attributes look similar to the attributes
>>> provided by the powersupply core, there are various deficiencies:
>>>
>>> * They don't show up in uevent payload.
>>> * They can't be queried with the standard in-kernel APIs.
>>> * They don't work with triggers.
>>> * The extending driver has to reimplement all of the parsing,
>>>     formatting and sysfs display logic.
>>> * Writing a extension driver is completely different from writing a
>>>     normal power supply driver.
>>> * Properties can not be properly overriden.
>>>
>>> The proposed extension API avoids all of these issues.
>>> An extension is just a "struct power_supply_ext" with the same kind of
>>> callbacks as in a normal "struct power_supply_desc".
>>>
>>> The API is meant to be used via battery_hook_register(), the same way as
>>> the current extensions.
>>>
>>> For example my upcoming cros_ec charge control driver[0] saves 80 lines
>>> of code with this patchset.
>>>
>>> Contents
>>> --------
>>>
>>> * Patch 1 and 2 are generic preparation patches, that probably make
>>>     sense without this series.
>>> * Patch 3 implements the extension API itself.
>>> * Patch 4 implements a PoC locking scheme for the extension API.
>>> * Patch 5 adds extension support to test_power.c
>>> * Patch 6 converts the in-tree platform/x86/system76 driver to the
>>>     extension API.
>>>
>>> Open issues
>>> -----------
>>>
>>> * Newly registered properties will not show up in hwmon.
>>>     To do that properly would require some changes in the hwmon core.
>>>     As far as I know, no current driver would extend the hwmon properties anyways.
>>> * As this is only useful with the hooks of CONFIG_ACPI_BATTERY, should
>>>     it also be gated behind this or another config?
>>> * Only one extension can be used at a time.
>>>     So far this should be enough, having more would complicate the
>>>     implementation.
>>> * Is an rw_semaphore acceptable?
>>>
>>> [0] https://lore.kernel.org/lkml/20240528-cros_ec-charge-control-v2-0-81fb27e1cff4@weissschuh.net/
>> Nice, i love this proposal!
> Good to hear!
>
>> I agree that the hwmon update functionality will need some changes in the hwmon core to work,
>> but there would be at least one driver benefiting from this (dell-wmi-ddv). Maybe we can add
>> support for this at a later point in time.
> Surely. Alternatively we could re-register the hwmon device after an
> extension was added.
>
>> The possibility of registering multiple power supply extensions on a single power supply will
>> be necessary to support battery charge control on Dell notebooks in the future. This is because
>> there will be two drivers on Dell notebooks which register battery extensions: dell-wmi-ddv and
>> dell-laptop (when support for battery charge control is supported someday).
>>
>> How difficult would it be to support such scenarios? If its very difficult, then maybe we can implement
>> this later when the need arises.
> It's not really difficult. The problem is in the callback functions
> going from a 'struct power_supply' back to the correct extension struct
> for use with container_of() to access the drivers private data.
>
> But we can add a marker member to 'struct power_supply_ext' with which
> the callback can figure out which of the registered extensions is its
> own. Something like "led_hw_trigger_type" in the LED subsystem.

Maybe we can do the same thing as the battery hook API and just pass a pointer to
the power_supply_ext instance to the callbacks. They then can use container_of()
to access the drivers private data if the struct power_supply_ext is embedded
inside the private data struct.

>
> And some documentation about how conflicts are to be resolved.
>
> Thomas

Sound like a plan, i suggest that extensions be prevented from registering with
a power supply containing conflicting properties or containing extensions with
conflicting properties.

As a side note, maybe there is a way to make power_supply_update_groups() available
for other power supply drivers? Afaik the ACPI battery driver would benefit from this too.

Thanks,
Armin Wolf

>>> Signed-off-by: Thomas Weißschuh <linux@weissschuh.net>
>>> ---
>>> Thomas Weißschuh (6):
>>>         power: supply: sysfs: use power_supply_property_is_writeable()
>>>         power: supply: core: avoid iterating properties directly
>>>         power: supply: core: implement extension API
>>>         power: supply: core: add locking around extension access
>>>         power: supply: test-power: implement a power supply extension
>>>         platform/x86: system76: Use power_supply extension API
>>>
>>>    drivers/platform/x86/system76_acpi.c      |  83 +++++++++---------
>>>    drivers/power/supply/power_supply.h       |   9 ++
>>>    drivers/power/supply/power_supply_core.c  | 136 ++++++++++++++++++++++++++++--
>>>    drivers/power/supply/power_supply_hwmon.c |  48 +++++------
>>>    drivers/power/supply/power_supply_sysfs.c |  39 ++++++---
>>>    drivers/power/supply/test_power.c         | 102 ++++++++++++++++++++++
>>>    include/linux/power_supply.h              |  25 ++++++
>>>    7 files changed, 357 insertions(+), 85 deletions(-)
>>> ---
>>> base-commit: 2df0193e62cf887f373995fb8a91068562784adc
>>> change-id: 20240602-power-supply-extensions-07d949f509d9