Re: hidraw with exclusive access ?

[Timo Teras <timo.teras@iki.fi>]

On Mon, 16 Mar 2015 15:15:33 -0400
Benjamin Tissoires <benjamin.tissoires@gmail.com> wrote:

> On Mon, Mar 16, 2015 at 2:39 PM, Timo Teras <timo.teras@iki.fi> wrote:
> > On Mon, 16 Mar 2015 11:29:10 -0400
> > Benjamin Tissoires <benjamin.tissoires@gmail.com> wrote:
> >
> >> On Fri, Mar 13, 2015 at 2:37 AM, Timo Teras <timo.teras@iki.fi>
> >> wrote:
> >> > As an alternative I think we could have hid-rawonly driver, that
> >> > would by default bind to nothing. And when bound, would connect
> >> > to the hid device with HID_CONNECT_HIDRAW flag only. I could
> >> > then in userland rebind the device from hid-generic to
> >> > hid-rawonly.
> >>
> >> That could work (make a special driver and use HID_CONNECT_HIDRAW),
> >> but that's a terrible idea and will be refused I guess.
> >> Users expects their keyboard to be working everywhere, even at boot
> >> and in the console. By doing so, you will just break existing
> >> keyboards.
> >
> > No. The whole point was to have driver that is not automatically
> > bound to anything. So we would not break anything existing. You'd
> > get that driver only by manual rebinding via sysfs.
> 
> Hmm, still not convinced by the manual re-binding through sysfs. If
> the keyboard does not work directly with the generic hid protocol or
> send garbage that users can not deal with, then why not simply bind it
> to the hidraw driver only?

They are complicated and configurable devices. They can work in the
macro mode, and I believe the device can send a sequence of keypresses
- at least some of them, but not all. Additionally some of the devices
appear as multiple HID devices. E.g. the joystick models show up as
three HID devices: their "SPLAT" interface, one keyboard and one
joystick (or mouse) device.

It might make sense to allow the Joystick to show up as regular
joystick by default. But to make the SPLAT interface work properly,
also the other HID events need to be processed. So currently I have
code that reads all three related hidraw devices, but acts only based
on the SPLAT one.

It also depends on the device model whether or not it generates
"conflicting" input events that I'd like to suppress.

In any case, the simplest way for me to was to use hidraw devices.. As
I'm using dedicated machine, there's no real harm by the conflicting
input events (as no X is running currently). The only annoyance is the
dmesg errors about bad input events. But if I ever build boxes running
also X, it'll be a problem. That's why'd appreciate a way to grab the
hidraw device.

> >> > Or finally, to implement a full kernel side driver it. But I'd
> >> > rather not go there. Especially since the device has multiple
> >> > leds, and the input system allows only limited leds. (The leds
> >> > could be exported as led devices, but then I'd need more
> >> > userland logic to figure out which led devices map to which
> >> > input device.)
> >>
> >> No, that's not true. If you have LEDs, then use the LED class and
> >> then user space deal with the LEDs. If the LEDs are not standard,
> >> then you will need userspace tools to access them so I don't think
> >> it is a problem. Standard LEDs will be handled properly through
> >> the input node (CAPS Lock, Ver Num, etc...).
> >
> > There's individial LEDs per each key. Namely I'm looking at the
> > devices at: http://xkeys.com/XkeysKeyboards/index.php
> >
> > The protocol they run is published at:
> > http://xkeys.com/PISupport/DeveloperHIDDataReports.php
> >
> > And as mentioned - it'd be lot of work to combine which LED is for
> > which key.
> 
> Then in that case, why not exporting your own sysfs interface to
> illuminate the LEDs and set the macro?
> 
> We do that for the OLED of the wacom device, and that's fine (I
> guess).

I found it simpler to implement the protocol in my app, instead of
writing a kernel interface for it. Doing the kernel driver would make
more sense if there was a way to map LEDs to keys already. Otherwise
I'm just inventing an abstraction that just makes me write more code...
and will probably eventually be replaced by something else.

> >> The good side of having your driver in the kernel is also less
> >> maintenance, and less pain for the users. Once it hits mainline,
> >> users will have a functional keyboard without having to rely on
> >> anything else. For you, if some interface change, you will have
> >> less head aches.
> >>
> >> If you *really* don't want to work in the kernel, you should simply
> >> ignore the generic keys in your driver and work only on the special
> >> keys macros that you want to support.
> >
> > It's not possible. Few of the generic keys default to mouse buttons.
> > Those I'd like to ignore.
> >
> > In fact, there's already some code out there that use hidraw and
> > feed the application from it. The code also opens the corresponding
> > input device with grab to make sure the events do not leak out to
> > x11. So I'm not the only one wanting to do this.
> 
> So, IMO you can approach the problem through 2 ways:
> - make a short out-of-the-tree driver for the X-keys that bind only to
> hidraw, when you launch the user-space driver, unbind hid-generic and
> bind the device to your small driver and then deal with the keyboard
> in userspace.
> - keep the kernel input node, reimplement the .raw_event() in the
> kernel driver, and provide a control API through sysfs.
>
> The second way allows you to have a UI/config tool which is not a
> daemon and which just controls what the kernel does.

I'd like to have a generic in-tree solution. For my immediate needs,
I'd be happy using hidraw as long as I have some mechanism to exclude
feeding that HID device to regular input system.

> BTW, I guessed that the macros are not stored in the keyboard but
> rather in your user-space program. If the macros are actually stored
> and generated by the device itself, then I think fixing the kernel to
> handle them would be a better path, and you need hidraw only to
> configure the device.

As said, depends likely on the device, and the way it is configured.
But to make the backlighting control work usage of the SPLAT interface
is required.

Btw. one example of code trying to accomplish something similar is:
https://github.com/chrippa/ds4drv/blob/master/ds4drv/backends/hidraw.py
it uses and parses the hidraw stuff, but also opens the corresponding
input device just to grab it for exclusive access.

Thanks,
Timo