Re: [RFC 0/1] hwmon: (hp-wmi-sensors) Support autoloading

[James Seo]

On Wed, Mar 20, 2024 at 04:13:59PM +0100, Armin Wolf wrote:
> Am 20.03.24 um 01:40 schrieb James Seo:
> 
>> On Tue, Mar 19, 2024 at 02:00:06PM +0100, Armin Wolf wrote:
>>> Am 19.03.24 um 06:47 schrieb James Seo:
>>> 
>>>> On Mon, Mar 18, 2024 at 10:57:31PM +0100, Armin Wolf wrote:
>>>>> Currently, the hp-wmi-sensors driver needs to be loaded manually
>>>>> on supported machines. This however is unnecessary since the WMI
>>>>> id table can be used to support autoloading.
>>>>> 
>>>>> However the driver might conflict with the hp-wmi driver since both
>>>>> seem to use the same WMI GUID for registering notify handler.
>>>>> 
>>>>> I am thus submitting this patch as an RFC for now.
>>>>> 
>>>>> Armin Wolf (1):
>>>>>     hwmon: (hp-wmi-sensors) Add missing MODULE_DEVICE_TABLE()
>>>>> 
>>>>>    drivers/hwmon/hp-wmi-sensors.c | 2 ++
>>>>>    1 file changed, 2 insertions(+)
>>>>> 
>>>>> --
>>>>> 2.39.2
>>>>> 
>>>> Autoloading was deliberately left out for now because of the GUID
>>>> conflict with hp-wmi's WMI notify handler.
>>>> 
>>>> HP's GUID reuse across product lines for different types of WMI
>>>> objects with different names and shapes means that with a patch like
>>>> this, many systems that should only load hp-wmi-sensors but not
>>>> hp-wmi will try to autoload both. (Perhaps all of them; I want to say
>>>> that the GUID 5FB7F034-2C63-45e9-BE91-3D44E2C707E4, which is the
>>>> second of the two GUIDs that hp-wmi uses to autoload, exists on every
>>>> HP system I've examined.)
>>>> 
>>>> Meanwhile, hp-wmi does various other platform things, and there's so
>>>> much hardware out there that who knows, maybe there are some systems
>>>> that really should load both. I don't think so but I can't rule it
>>>> out.
>>>> 
>>>> Unlike hp-wmi-sensors, hp-wmi doesn't survive failure to install its
>>>> notify handler, which sets up a potential race condition depending on
>>>> when hp-wmi and hp-wmi-sensors loads on a given system.
>>>> 
>>>> Therefore, I intended to add autoloading at the same time as
>>>> converting hp-wmi-sensors to use the bus-based WMI interface once
>>>> aggregate WMI devices are better supported.
>>>> 
>>>> As you mentioned [1], I ran into issues when I tried to do the
>>>> conversion by simply adding the GUID to struct wmi_driver.id_table.
>>>> That resulted in two separate independent instances of hp_wmi_sensors
>>>> being loaded, which isn't what I wanted.
>>> After taking a look at a ACPI table dump of a HP machine, i noticed that
>>> HPBIOS_BIOSEvent has the GUID 2B814318-4BE8-4707-9D84-A190A859B5D0, which is
>>> different than the event GUID used by hp-wmi.
>>> 
>>> According your comment in hp_wmi_notify(), i assume that some machines have
>>> mixed-up event GUIDs.
>> I investigated further. Every HP machine in the Linux Hardware Database that
>> has \\.\root\WMI\hpqBEvnt at 95F24279-4D7B-4334-9387-ACCDC67EF61C also has
>> \\.\root\WMI\HPBIOS_BIOSEvent at 2B814318-4BE8-4707-9D84-A190A859B5D0.
> 
> Could it be that using 95F24279-4D7B-4334-9387-ACCDC67EF61C is a mistake?
> Or do you know of a machine which indeed uses this GUID to deliver sensor events?
> Because it not, then we can just avoid this GUID conflict entirely by using the
> other GUID.
>

No, it's not a mistake, it's HP reusing GUIDs. Both my test machines use
95F24279-4D7B-4334-9387-ACCDC67EF61C for \\.\root\WMI\HPBIOS_BIOSEvent.

Previously I examined a sample of ACPI dumps from machines with both
hpqBEvnt and HPBIOS_BIOSEvent, and concluded:

  - hpqBEvnt is for various events on both business and non-business
    machines that are of no interest to hp-wmi-sensors (e.g. hotkeys)

  - some machines with hpqBEvnt also have HPBIOS_BIOSEvent at GUID
    2B814318-4BE8-4707-9D84-A190A859B5D0

  - no machines with both hpqBEvnt and HPBIOS_BIOSEvent actually surface
    relevant sensor events (e.g. fan speed too high) via HPBIOS_BIOSEvent;
    they only surface non-sensor events (e.g. BIOS setting was changed)
    that are of no interest to hp-wmi-sensors

  - therefore, 2B814318-4BE8-4707-9D84-A190A859B5D0 does not need to be
    handled in hp-wmi-sensors

But this time I have done an exhaustive examination and concluded that a
few machines with both events do surface sensor events via HPBIOS_BIOSEvent.

>>> I thing it would be best to create a separate WMI driver for the event and
>>> use a notifier chain (see include/linux/notifier.h) to distribute the event data.
>>> 
>>> In case of event GUID 95F24279-4D7B-4334-9387-ACCDC67EF61C, both hp-wmi and
>>> hp-wmi-sensors can subscribe on this notifier and receive event data without
>>> stepping on each other's toes.
>>> 
>>> The same can be done for the event GUID 2B814318-4BE8-4707-9D84-A190A859B5D0,
>>> with a separate notifier chain.
>>> 
>>> This would decouple the event handling from the event data consumers, allowing
>>> both hp-wmi and hp-wmi-sensors to coexist.
>> No objections from me for this specific use case to work around the GUID conflict.
>> hp-wmi-sensors should indeed subscribe on 2B814318-4BE8-4707-9D84-A190A859B5D0
>> for some of those machines.
>> 
>> Any ideas for getting rid of wmi_query_block() for fetching
>> \\.\root\HP\InstrumentedBIOS\HPBIOS_PlatformEvents? I know other drivers are
>> also using it for getting blocks other than their "main" GUID.
> 
> Good question, it seems that HPBIOS_PlatformEvents is optional, so using the component
> framework will not work.
> 

Yes, HPBIOS_PlatformEvents is optional, but it's pretty much necessary for
alarm and intrusion events. Without it, it's not possible to know whether a
machine even reports such events until after they occur (rare). We'd have
to assume that all machines always support such events.

> If those WMI data blocks are always associated with the same ACPI device used by the
> sensors GUID, then maybe i could create some sort of API for checking if a given GUID
> exists the ACPI device associated with a WMI device.

For all HP machines in the Linux Hardware Database, all machines with
HPBIOS_PlatformEvents also have HPBIOS_BIOSNumericSensor. The reverse is
not true. Neither WMI object appears under multiple GUIDs.

> However i thing the event GUID issue is more important right now.

Sure. I also wonder if your idea could be expanded into a generic driver
for publishing simple WMI events. This would be usable in other drivers
that are currently using legacy handlers for receiving event data.

More broadly, if hp-wmi-drivers is any indication, aggregate WMI devices
could be a pain. Primary WMI object, associated WMI objects (optional or
mandatory), multiple aggregate devices allowed to bind to the same
objects. And if using GUIDs for identification, multiple allowable GUIDs.

Thanks,

James

> Thanks,
> Armin Wolf
> 
>>> I can provide a prototype implementation, but unfortunately i have no HP machine
>>> myself for testing. But i might be able to find one to test my changes.
>> Happy to test. (Also happy to try it myself, but I'd need some help.)
>> 
>>> Thanks,
>>> Armin Wolf
>>> 
>>>> [1] https://lore.kernel.org/linux-hwmon/cd81a7d6-4b81-f074-1f28-6d1b5300b937@xxxxxx/
>>>>