6.3 KiB
Why does input-remapper not use xkb configs?
Initial target You write a symbols file based on your specified mapping, and that's pretty much it. There were two mappings: The first one is in the keycodes file and contains "<10> = 10", which is super redundant but needed for xkb. The second one mapped "<10>" to characters, modifiers, etc. using symbol files in xkb. However, if you had one keyboard layout for your mouse that writes SHIFT keys on keycode 10, and one for your keyboard that is normal and writes 1/! on keycode 10, then you would not be able to write ! by pressing that mouse button and that keyboard button at the same time.
This was quite mature, pretty much finished and tested.
It still exists in the first branch
The second idea was to write special keycodes known only to input-remapper (256 - 511) into the input device of your mouse in /dev/input, and map those to SHIFT and such, whenever a button is clicked. A mapping would have existed to prevent the original keycode 10 from writing a 1. But this device doesn't have the capabilities set for those keycodes, so it won't use them. At that time I didn't know about capabilities though.
The third idea is to create a new input device that uses 8 - 255, just like other layouts, and input-remapper always tries to use the same keycodes for SHIFT as already used in the system default. The pipeline is like this:
- A human thumb presses an extra-button of the device "mouse"
- input-remapper uses evdev to get the event from "mouse", sees "ahh, it's a 10, I know that one and will now write 50 into my own device". 50 is the keycode for SHIFT on my regular keyboard, so it won't clash anymore with alphanumeric keys and such.
- X has input-remappers configs for the input-remapper device loaded and checks in it's keycodes config file "50, that would be <50>", then looks into it's symbols config "<50> is mapped to SHIFT", and then it actually presses the SHIFT down to modify all other future buttons.
- X has another config for "mouse" loaded, which prevents any system default mapping to print the overwritten key "1" into the session.
But this is a rather complicated approach. The mapping of 10 -> 50 would have to be stored somewhere as well. It would make the mess of configuration files already needed for xkb even worse. This idea was not considered for long
Fourth idea: Based on the second idea, instead of using keycodes greater than 255, use unused keycodes starting from 255, going down. For example pressing key 10 triggers input-remapper to write key 253 into the /dev device while mapping key 10 to nothing. This has the same problem, the device capabilities ignore many of those keycodes. 140 works, 145 won't, 150 works.
Fifth idea: Instead of writing xkb symbol files, just disable all mouse buttons with a single symbol file. Input-remapper listens for key events in /dev and then writes the mapped keycode into a new device in /dev. For example, if 10 should be mapped to Shift_L, xkb configs would disable key 10 and input-remapper would write 50 into /dev, which is Shift_L in the system mapping. This sounds incredibly simple and makes me throw away tons of code.
But somehow writing into the new /dev file makes the original keycode not mapped by xbk symbol files, and therefore leak through. In the previous example, it would still write '1', and then after that the other key. By adding a timeout single keys work, but holding down a button that is mapped to shift will (would usually have a keycode of 10, now triggers writing 50) write "!!!!!!!!!". Even though no symbols are loaded for that button.
The Sixth idea The described problem is because the second device that starts writing an event.value of 2 will take control of what is happening. Following example: (KB = keyboard, example devices)
- hold a on KB1:
a-1,a-2,a-2,a-2, ... - hold shift on KB2:
shift-2,shift-2,shift-2, ... No a-2 on KB1 happening anymore. The xkb symbols of KB2 will be used! So if KB2 maps shift+a to b, it will write b, even though KB1 maps shift+a to c! And if you reverse this, hold shift on KB2 first and then a on KB1, the xkb mapping of KB1 will take effect and write c!
In the context of the fifth idea, KB1 would be the mouse, KB2 would be the new /dev device. The KB1 keycode comes first and is then realized as '!' when KB2 comes in and applies a different mapping.
Which means in order to prevent "!!!!!!" being written while holding down keycode 10 on the mouse, which is supposed to be shift, the 10 of the input-remapper /dev node has to be mapped to none as well. But that would prevent a key that is mapped to "1", which translates to 10, from working. So instead of using the output from xmodmap to determine the correct keycode, use a custom mapping that starts at 255 and just offsets xmodmap by 255. The correct capabilities need to exist this time. Everything below 255 is disabled. This mapping is applied to input-remappers custom /dev node.
However, if you try to map Shift to button 10 of your mouse, and use mouse-shift + keyboard-1, you need to press keyboard-1 again to do anything. I assume this is because:
- mouse-10 down
- inputremapper says: 50 down
- xkb mapping: 10 is none. 50 is shift.
- keyboard-10 down (down again? X/Linux ignores that)
- keyboard-10 up
- keyboard-10 down, "!" written
Seventh, final solution By grabbing the mouse device (EVIOCGRAB) this won't happen. Since this prevents all the keycodes from doing stuff, no empty xkb symbols file is needed anymore. If 10 is mapped to 'a', it will figure out the keycode for 'a' in the system configuration (via setxkbmap -pke) and write it into a new device that has proper capabilities. So no xkb configurations are needed at all anymore.
How I would have liked it to be
This solution would have made the macro thing impossible though
setxkbmap -layout ~/.config/input-remapper/mouse -Foo Device3
config looks like:
10 = a A
11 = Shift_L
282 = b # middle mouse
done. Without crashing X. Without printing generic useless errors. Without colliding with other devices using the same keycodes. Xkb also can't map 282 afaik. If it was that easy, an app to map keys would have already existed.
The current solution supports a config like that in json format.