Am 01.10.24 um 15:41 schrieb Benjamin Tissoires:
On Oct 01 2024, Werner Sembach wrote:
(sorry resend because thunderbird made it a html mail)
Hi,
Am 30.09.24 um 19:06 schrieb Benjamin Tissoires:
On Sep 30 2024, Werner Sembach wrote:
[...]
Thinking about it, maybe it's not to bad that it only changes once udev is
ready, like this udev could decide if leds should be used or if it should
directly be passed to OpenRGB for example, giving at least some consistency
only changing once: i.e. firmware -> OpenRGB setting and not firmware->leds
setting->OpenRGB setting.
That would work if OpenRGB gets to ship the LampArray bpf object (not
saying that it should). Because if OpenRGB is not installed, you'll get
a led class device, and if/when OpenRGB is installed, full LampArray
would be presented.
The idea in my head is still that there is some kind of sysfs switch to
enable/disable leds.
FWIW, I'm never a big fan of sysfs. They become UAPI and we are screwed
without possibility to change them...
Why not having a simple led driver for HID LampArray devices which exposes the
whole LampArray as a single LED?
If userspace wants to have direct control over the underlying LampArray device,
it just needs to unbind the default driver (maybe udev can be useful here?).
My idea is then that a udev rule shipped with OpenRGB sets this switch to
disable before loading the BPF driver so leds never get initialized for the
final LampArray device.
FWIW, udev-hid-bpf can inject a udev property into a HID-BPF. So
basically we can have a udev property set (or not) by openrgb which
makes the BPF program decide whether to present the keyboard as
LampArray or not.
I do not think that using HID-BPF makes sense here, since the underlying HID device
is already purely virtual.
Using HID-BPF on top of the already virtual HID device would be a bit strange.
But anyway, BPF allows to dynamically change the behaviour of the
device, so that's IMO one bonus point of it.
FWIW, the use of BPF only allows you to not corner yourself. If you
failed at your LampArray implementation, you'll have to deal with it
forever-ish. So it's perfectly sensible to use BPF as an intermediate step
where you develop both userspace and kernel space and then convert back
the BPF into a proper HID driver.
I don't really see this point: The LampArray API is defined by the HID Usage
Table and the report descriptor, so there is not API to mess up and
everything else has to be parsed dynamically by userspace anyway, so it can
easily be changed and userspace just adopts automatically.
And for this case the proper HID driver is already ready.
Yeah, except we don't have the fallback LED class. If you are confident
enough with your implementation, then maybe yes we can include it as a
driver from day one, but that looks like looking for troubles from my
point of view.
To be on the safe side that we don't talk about different things: My current
plan is that the leds subsystem builds on top of the LampArray
implementation.
I would say that the HID subsystem knows how to translate LampArray into
a subset of LEDs. But I think that's what you are saying.
Like this the leds part has to be only implemented once for all LampArray
devices be it emulated via a driver or native via firmware in the device
itself.
yep, that's the plan. However, not sure how to fit LampArray into LED.
Maybe the LED driver can present the whole LampArry as a single RGB LED. This
might be enough for basic LED control (on/off, changing color, ...).
For advanced LED control (effects, large number of LEDs, ...) userspace can just
unbind the default LED driver and use hidraw to drive the LampArray themself.
And I feel confident that the UAPI should be that the userspace gets a
hidraw device with a LampArray HID descriptor, and every thing else is, by
the HID spec, dynamic anyway so I can still change my mind in implementation
specifics there, can't I?
Yeah... I think?
From my point of view we are just bikeshedding on to where put that
"firmware" extension, in WMI, in HID (kernel with a subdriver), or
externally in BPF.
Just a insane idea: can Tuxedo change the ACPI code supplied by the BIOS?
After a second look at the LampArray code here... Aren't you forgetting
the to/from CPU conversions in case you are on a little endian system?
Since this driver is for built in keyboards of x86 notebooks it isn't
required or is it?
Good point. Let's just hope you don't start shipping a LE laptop with
the same keyboard hardware :)
I believe we should do those CPU conversions regardless, so future driver developers
have good examples to copy from.
So the only point for me currently is: Is it ok to have key position/usage
description tables in the kernel driver or not?
good question :)
I would say, probably not in the WMI driver itself. I would rather have
a hid-tuxedo.c HID driver that does that. But even there, we already had
Linus complaining once regarding the report descriptors we sometimes
insert in drivers, which are looking like opaque blobs. So it might not be
the best either.
Isn't tuxedo_nb04_wmi_ab_virtual_lamp_array.c not something like
hid-tuxedo.c? or should it be a separate file with just the arrays?
It is, in a way. I think it's more a question for Hans and the other
platform maintainers, whether they would accept this.
Is there a possibility to query the physical keyboard layout using the WMI interface?
Sorry I don't have a clear yes/no answer.
Hm... Well if it's no problem I would keep the current implementation with
minor adjustments because, like i described above, I don't see a benefit now
that this already works to rewrite it in BPF again.
If it is a problem then i don't see another way then to rewrite it in BPF.
Note: For future devices there might be more keyboard layouts added,
basically every time the chassis form factor changes.
If the WMI part doesn't change, then maybe having BPF would be easier
for you in the future. Adding a HID-BPF file would cost basically
nothing, and it'll be out of band with the kernel, meaning you can ship
it in already running kernels (assuming the same WMI driver doesn't need
any updates).
Cheers,
Benjamin
To sum up the architechture (not mutally exclusive technically)
/- leds
WMI <- WMI to LampArray Kernel driver <-switch-|
\- OpenRGB
/- leds
WMI <- WMI to Custom HID Kernel driver <- Custom HID to LampArray BPF
driver<-switch-|
\- OpenRGB
With the "switch" and "leds" implemented in hid core, automatically
initialized every time a LampArray device pops up (regardless if it is from
native firmware, a bpf driver or a kernel driver)
Writing this down I think it was never decided how the switch should look like:
It should not be a sysfs attribute of the leds device as the leds device
should disappear when the switch is set away from it, but should it be a
sysfs variable of the hid device? This would mean that hid core needs to add
that switch variable to every hid device having a LampArray section in the
descriptor.
Again, not a big fan of the sysfs, because it's UAPI and need root to
trigger it (though the udev rule sort this one out).
BPF allows already to re-enumerate the device with a different look and
feel, so it seems more appropriate to me.
Also, having a sysfs that depends on the report descriptor basically
means that we will lose it whenever we re-enumerate it (kind of what the
LED problem you mentioned above). So we would need to have a sysfs on
*every* HID devices???
Actually, what would work is (ignoring the BPF bikeshedding for Tuxedos
HW):
- a device presents a report descriptor with LampArray (wherever it
comes from)
- hid-led.c takes over it (assuming we extend it for LampArray), and
creates a few LEDs based on the Input usage (one global rgb color for
regular keys, another one for the few other LEDs known to userspace)
- when openRGB is present (special udev property), a BPF program is
inserted that only contains a report descriptor fixup that prevent the
use of hid-led.c
Can we just manually unbind the hid-led driver?
- the device gets re-enumerated, cleaning the in-kernel leds, and only
present the LampArray through hidraw, waiting for OpenRGB to take
over.
- at any time we can remove the BPF and restore the LEDs functionality
of hid-led.c
Being able to develop a kernel driver without having to reboot and
being sure you won't crash your kernel is a game changer ;)
Cheers,
Benjamin
Best regards and sorry for the many questions,
Werner Sembach
PS: on a side node: How does hid core handle HID devices with a broken HID
implementation fixed by bpf, if bpf is loaded after hid-core? Does the hid
device get reinitialized by hid core once the bpf driver got loaded? If yes,
is there a way to avoid side effects by this double initialization or is
there a way to avoid this double initialization, like marking the device id
as broken so that hid core- does not initialize it unless it's fixed by bpf?
- broken HID device:
it depends on what you call "broken" HID device. If the report
descriptor is boggus, hid-core will reject the device and will not
present it to user space (by returning -EINVAL).
If the device is sensible in terms of HID protocol, it will present it
to userspace, but the fact that it creates an input node or LED or
whatever just depends on what is inside the report descriptor.
- HID-BPF:
HID-BPF is inserted between the device itself and the rest of the
in-kernel HID stack.
Whenever you load and attach (or detach) a BPF program which has a
report descriptor fixup, HID-core will reconnect the device,
re-enumerate it (calling ->probe()), and will re-present it to
userspace as if it were a new device (you get all uevents).
- double initialization:
nowadays hid-generic doesn't do anything on the device itself except
calling the powerup/powerdown, by calling ->start and ->stop on the
HID transport driver. It's not a problem on 99% of the devices AFAICT.
technically, if the report descriptor is bogus, start/stop won't be
called, but you'll get an error in the dmesg. So if you really want to
rely on that "broken" scenario we can always add a specific quirk in
HID to not spew that error.
Cheers,
Benjamin
PPS: sorry for pushing that hard on HID-BPF, but I can see that it fits
all of the requirements here:
- need to be dynamic
- still unsure of the userspace implementation, meaning that userspace
might do something wrong, which might require kernel changes
- possibility to extend later the kernel API
- lots of fun :)
