Am 13.11.24 um 12:05 schrieb Hans de Goede:
Hi All,
On 12-Nov-24 5:31 PM, Armin Wolf wrote:
Am 12.11.24 um 16:10 schrieb Werner Sembach:
Am 12.11.24 um 13:51 schrieb Armin Wolf:
Am 12.11.24 um 13:42 schrieb Werner Sembach:
Hi,
Am 12.11.24 um 13:01 schrieb Armin Wolf:
Am 12.11.24 um 12:52 schrieb Werner Sembach:
Hi,
quick learning question: Why is wmi_bus_type not exported unlike, for
example, acpi_bus_type, and platform_bus_type?
Wanted to use bus_find_device_by_name in an acpi driver that might
need additional infos from a wmi interface that might or might not be
present.
Kind regards,
Werner Sembach
What kind of information do you have in mind? wmi_bus_type is not
being exported for historic reasons, i can change that if necessary.
It's for the tuxedo-drivers part for the Sirius 16 Gen 1 & 2 which has
a slow wmi and a quick acpi interface, however the quick acpi
interface can not get the max rpm of the cooling fans, but the wmi
interface can.
Thing is for the acpi driver we might plan an earlier upstream date
and it might get multi-odm support, while the wmi interface is and
stays odm specific. So my idea was to only couple both drivers in a
dynamic way using bus_find_device_by_name.
Interesting, how is the ACPI interface not ODM specific? Can you
elaborate a bit on how the ACPI and the WMI interfaces work?
We have an ODM that was willing to include ACPI code by us in their
BIOS blob and we hope that in the future we can carry that API over to
other ODMs for future TUXEDO devices.
In pseudocode that API looks like this:
v1:
void SMOD(bool mode): Toggle firmware controlled fans vs manually (aka
via the commands below) controlled fans
bool GMOD(): Get current SMOD setting
int GCNT(): Get number of fans
enum GTYP(int index): Returns "CPU-fan" or "GPU-fan"
void SSPD(int index, int value): Set fan speed target as a fraction of
max speed
int GSPD(int index): Get current fan speed target as a fraction of max
speed
v2 same as v1 but with added:
int GRPM(int index): Get current actual fan speed in revolutions per
minute
int GTMP(int index): Get temperature of thing fan with respective
index is pointed at (CPU or GPU die, see GTYP)
Like I said, what is missing is a "Get Max RPM" function even in v2,
which we might add a future iteration, but, well this bios is now out
in the wild. However these released devices have a "get info" function
in the wmi code which returns the v2 infos and the max rpm.
I want to write the code in a way that it probes the acpi interface
for function existence and wherever something is missing tries to fall
back to infos gathered from the wmi interface, but that one is
implemented in a stand alone module (the tuxedo_nb04_* stuff in
tuxedo-drivers) and I would like to keep it that way in honor of KISS.
My plan is that the first time max rpm is pulled the acpi driver uses
bus_find_device_* to get the wmi device, if present, and pulls max rpm
from the driver data there and copies it over to it's own driver data.
If not possible it returns a dummy value or falls back to another
method. Maybe a hard coded list of max rpm values, currently only 2
devices have the new interface, so it wouldn't be a long list.
Directly going to the hard coded list is our current fallback plan,
but it is not an elegant solution as the info is actually there, if
you know what i mean?
Kind regards,
Werner
I see, we once had a similar case with the dell-wmi driver, see commit f97e058cfe80 ("platform/x86: wmi: Don't allow drivers to get each other's GUIDs"):
The only driver using this was dell-wmi, and it really was a hack.
The driver was getting a data attribute from another driver and this
type of action should not be encouraged.
Rather drivers that need to interact with one another should pass
data back and forth via exported functions.
I would be quite unhappy with drivers interacting with WMI devices without a proper WMI driver, but i can see your point here.
Agreed on that 1 driver should not be poking the [wmi_]dev of another
driver. This usually works until it doesn't for some reason so it
should just be avoided.
Maybe we can keep the retrieval of the fanX_max values out of the kernel? I propose the following:
- have a driver for your generic ACPI interface
- have a driver for the WMI interface (with fanX_max hwmon attributes)
The driver for the generic ACPI interface exposes the fan speed controls as pwmX attributes if the interface does not support
the "Get Max RPM" function. The userspace application in this case searches for the hwmon chip exposed by the WMI driver and
reads the fanX_max attributes there. Then the application can convert the target fan speed into values for the pwmX attributes
itself.
If the ACPI interface however supports the "Get Max RPM" function, then it exposes fanX_max and fanX_target hwmon attributes
themself and the userspace application uses them directly.
This would keep the kernel drivers simple.
That would indeed keep the kernel drivers simple, but at the cost of
providing a non standard hwmon interface.
Whatever implementation is written it really MUST follow the standard
hwmon API so that any hwmon tools like the lm_sensors fancontrol script
will work properly.
So NACK from me for exposing fanX_max on a separate hwmon device.
What I think works best in cases like this is to have the wmi-driver
expose a function to retrieve the fan max value.
This function can use a static global array of fan max values +
a global fan_max_values_initialized bool and it can return -EPROBE_DEFER
when the bool is not set yet.
This will also require the ACPI driver to have a Kconfig "depends on"
the WMI driver but that should be fine.
And then the ACPI driver can simply call the exported helper function
to get the max fan values.
This sort of cross driver function calling is not ideal, but it is
better then poking at a struct device owned by another driver.
Problem is that when we really bring over the ACPI interface to other ODMs the
WMI driver will no longer be there, that's why I wanted to avoid a static
dependency on the WMI module to be able to reuse the code.
Or maybe just hardcode the max fan values in the ACPI driver if
it is only 2 models and the next version of the ACPI interface
is supposed to fix this shortcoming ?
Well, at least I hope so.
Regards,
Hans
