UCW keyboard layout on XWayland
This is a story/writeup of how I debugged UCW layout on XWayland. You may learn
here how keyboard layouts work on Linux.
Background: The UCW layout
The UCW keyboard layout is one of the interesting methods of typing Czech
letters on a rather American keyboard. The main idea is having a classic US
(QWERTY) layout, but without the CapsLock, which serves as a key to add
diacritics, so that e.g. Caps+s creates "š". It is naturally possible to
combine Caps with Shift to create uppercase letters, and because not all
letters can have diacritics in Czech (e.g. there is no "ǩ"), it also manages to
cover all the German and Slovak letters. (Sometimes, there are multiple
diacritics for a single letter, then the additional diacritics are nearby:
Caps+e is "é", Caps+w is "ě".) The CapsLock can still be pressed by using
Alt+Caps.
The layout has some nice features like avoiding deficiencies with the Czech
layout (diacritics on number row are too far, both round parentheses are on the
same key, other parentheses only with AltGr at random places, no asterisk, for
some reason we have "§" though, …). Another nice feature is that it is rather
interoperable: I am able to type on any computer with the "most standard"
layout and when writing in Czech I can just not press the CapsLock key and only
lose my diacritics, computers can have this layout system-wide even when
foreign people use it,…
I would not say many people use the layout but at least several my friends do
and I have come across several random people on the Internet who do also. More
on this below :-)
The important part is how the layout is set up. Fortunately, it is contained in
the xkeyboard-config database, so the following command just enables it in Xorg:
setxkbmap us,cz -variant ,ucw -option grp:caps_switch
Technically, this sets two layouts (or groups), which are switched by
pressing the CapsLock key. This has some disadvantages (I have managed a few
times to have the order accidentally swapped and layout switchers (or their
users) are sometimes confused), but actual issues are rare. Under X11, that is…
Corollary 1: I want UCW to work differently.
My custom keyboard layout?
I didn't like fact that UCW layout is in fact just an overlay and not a
"proper" level 3, so I started studying how to create a custom keyboard layout.
Few notable resources: XKB page on ArchWiki, manpages for
xkeyboard-config(7), setxkbmap(1) and xkbcomp(1) and looking into
/usr/share/X11/xkb/symbols/cz (found by grepping /usr/share/X11 for
"ucw").
Another reason for creating my own layouts is various tweaks esp. on laptop
keyboards. For example, the laptop I am typing this sentence on has broken the
up arrow key, so I would like to map it somewhere. I use xmodmap for that,
but if this could be contained in a single layout, it would make stuff simpler
for me. (And as I will later learn, xmodmap does not work well with xkb
layouts…)
Also, I have another tweak on my keyboards: Compose key. This really calls for
the custom layout! And when at it, I should create one for the ttys as well…
But this had quite a low priority (the kbd.sh script in all my homes does
the job well enough), so I didn't get to it in time. Luckily, maybe – if I were
quicker I would not end up in this rabbit hole. At this point, I knew the basic
stuff and was reasonably sure that I could hack it together, at least somehow;
what was stopping me was lack of time and not being sure how I want to manage
machine-specific tweaks in two different layout syntaxes (xkb for X11 and kbd
for ttys).
Aside: A few notes about xkb layouts for completeness
This all has been described elsewhere in more detail, but it is useful to
know when debugging layouts, so I will mention it here again.
There are two forms xkb layouts may be specified. The low-level one is called
KcCGST (keycodes, compat, geometry, symbols, types) and describes very much
everything that should happen on any keypress. These actions are described in
subdirectories of /usr/share/X11/xkb/ of the same name.
The high-level descriptions are called RMLVO (rules, model, layout, variant,
option) and are what you usually configure, either with setxkbmap or using
GUI tools. Files in /usr/share/X11/xkb/rules/ describe the translation of
RMLVO to KcCGST (the XML files add human-readable names for the GUIs and the
like). Maybe confusingly, the RMLVO names of options and layouts are very
similar to the names of KcCGST compat and symbols, but are generally a
different thing AFAIK.
KcCGST also come in two forms: the readable one and the complete one. The
readable one does not determine everything in one file, instead including
others. The complete description is, well, complete. It is kind of similar to
the C preprocessing.
The readable KcCGST are what is stored in the filesystem as well as what you
get from setxkbmap -print and may serve as a good starting point for
tweaks. To get the complete ones one can use xkbcomp $DISPLAY -.
The processing of these description is different in Xorg and in Wayland. If
I were to bet where the bug is before digging into it, this sounds like a very
likely culprit.
A wild friend has appeared
… and he had a problem and asked me if I would have a look into it. Apparently,
group switching didn't work in XWayland. At first, I wanted to work it around
by finally creating the Unified UCW keymap™, but I wanted to learn more about
XKB (from TODO GUIDES), which took me quite long, again. (Studying, working,
helping with camps and fixing layouts takes some time…)
Also, since the issue is apparently bigger, solving all group switches would be
a better and more useful solution than just hacking up the Unified UCW keymap
(even though I want my tweaks to eventually be xkb-based anyway).
I had very little experience with Wayland until then, mostly because X11 worked
for me rather well. I tried running Sway few times before, but usually quickly
reverted to using Xorg for various reasons (Nvidia GPU in laptop, very laggy
mouse, maybe this exact issue with UCW layout) which I didn't feel like
solving at that moment.
I wanted to help my friend, though, so I started looking more into using Sway
on another laptop and reading about internals of input system on Peter
Hutterer's blog TODO NAME SPELLING? Occasionally I would randomly google
(with DuckDuckGo :-D) for the issue with group switching, just in case…
The nerd snipe issue
TODO SOMEONE on GitHub found out an interesting workaround: when they
disabled the group switch, the UCW layout started working. Complete "huh?"
moment, I knew I wanted to know more about why that worked. The
xkeyboard-common(7) manpage says about more group switching options, so I
randomly tried switching by Menu key in XWayland (using setxkbmap even
though it complained and was supposed to do nothing), and ended up with both
CapsLock and Menu adding diacritics.
After a bit of mathematic thinking (more in my comment of the main issue) I
concluded that it was actually working fine, just that the SetGroup action
was evaluated twice, which would overflow back to the first group when two
layouts are set (probably the most common case). Mathematic term for this issue
is "arithmetic in the Z_2 group" :-)
Corollary 2: A simple workaround is adding the UCW overlay twice, e.g.
setxkbmap us,cz,cz -variant ,ucw,ucw (and thus changing the algebraic group
to be Z_3).
I think that another workaround would be to change the compat rules for
XWayland, but it feels nasty to have such a quirk in xkeyboard-config database.
Changing them for everyone (in compat/evdev directly) might break other
systems, so that also does not seem to be a good way.
Corollary 3: We are fixing XWayland (or the way it processes events from Wayland).
Learning what happens
I continued reading, this time mostly who-t's post series about custom
layouts, core Wayland protocol as well as the source code of various
tools (xev, wev, xkbcli interactive-wayland &c.) Given that I knew
very little about the inner workings of the stack, I wanted to find some code
that would be run and "enhance" it with so many debug prints that I would
understand what was the state of various parts of the Xwayland's "stacked xkb".
I kind of knew that one part of this is somewhere in the XWayland server,
whose code felt intimidating, so I wanted to determine where to start nudging
it from what APIs the clients use.
Tip
Downloading various packages in Linux is rather simple, as well as their
rebuilding. In Arch, I can get the package's PKGBUILD just with yay -G
package, build it with makepkg and install it using yay -U
package-….pkg.tar. Makepkg also has some options which allow me to tweak
the source code after downloading it and before building it.
Other distros are probably similar, e.g. for Debian-based distros one can
use apt source, debuild and apt install to do the same.
One interesting observation is that contrary to my expectation, XWayland does
not seem to use libxkbcommon (according to the /proc/PID/maps file). This
can have several reasons, but its source code also contains a slightly tweaked
version of (Xorg-style) xkb, which might mean I will be dealing with the ugly
code :-/ (Actually, XWayland might not process key events itself, instead
just passing them to clients, but this seems inconsistent with the issue – what
else would be introducing the second group switch?)
So now I need to understand a (hopefully small) parts of two protocols: X11 to
understand what the X11 clients receive from XWayland, and Wayland to
understand what XWayland gets from my compositor.
Understanding X11
This part is maybe the easier one, because xev pretty much dumps the data.
I should check in the server code as well, but for the Czech chords xev
seems to receive KeyEvents with state 0x2000 under Xorg (i3wm) and with
0x4000 under XWayland when three groups are set up (for two groups the
state is 0x0, i.e. the same as for English keypresses). State bits this
high encode the active group and are described in the relevant part
of the XKB extension documentation.
We can see that if we set the GroupsWrap control of XKB (I don't know
where/how yet) to ClampIntoRange it should also help work around the
problem. But our goal is to align behaviour of XWayland and JustWayland, i.e.
prevent XWayland seeing two group shifts, not to hack our way in either
protocol, so we do not stop here.
At this point, I had an approximate idea of how XKB is supposed to do,
hopefully. I have not yet read the implementation in XWayland though…
Understanding Wayland
This part is supposed to also be simple: the core Wayland protocol will send us
wl_keyboard::keymap, wl_keyboard::key and
wl_keyboard::modifiers events containing the required data. To prove
our hypotheses we can use wev which does very little post-processing as
with xev.
The client does not seem to have a say in what kind of keymap it will get, it
is up to the compositor. This means that even XWayland cannot say it wants raw
data, but xkb_v1-style keycodes are probably close (they are off by 8,
which is kind of expected).
While exploring what happens I found out there is a bug in the stable version
1.0.0 of wev, where it forgets to include the serial number for
wl_keyboard::modifiers. It has already been patched
15 months ago, but the patch is not included in a release yet.
But I am also interested in what is exchanged between the compositor and
XWayland server. And there does not seem to be any kind of sniffer yet, so I
started writing one
. Only then I learned that I can set the environment variable
WAYLAND_DEBUG=1 to get dumps of all the calls that are happening. (To my
defense, this does not seem to be very documented. I only found a small note in
the building guide. Even the debugging extras page does
not mention this.)
By running WAYLAND_DEBUG=1 Xwayland :15 and DISPLAY=:15 xev, we do
learn what we thought was happening:
- XWayland gets XKB v1 keymap (I trust sway to supply the same map as to
wev earlier.)
- XWayland gets the correct (Wayland core) events from the compositor
- XWayland sends a wrong (X11 with XKB) events to xev.
This confirms that XWayland is really the culprit here.
Digging into XWayland
At this point, we have all the tools, references to documentation and knowledge
we could possibly have, so now we get our hands dirty. A simple grep -r
wl_keyboard src/ tells us that the keyboard handling occurs in the file
hw/xwayland/xwayland-input.c in functions keyboard_handle_*, which look
like the libwayland-client handlers we have seen in wev.
However, the fun ends here. I found the generic code of Xorg (which these
handlers quickly call) rather hard to make sense of (though ctags,
printf(3), debug build and gdb might help) and the implementation of XKB is
rather massive: about 20 000 loc just in the xkb/ directory. And touching
this code in a bad way might break something somewhere for somebody.
We can get creative though. We know that this issue is only present in
XWayland, which (by a quick grep -l) seems to live in hw/xwayland. Yay,
only about 12 000 loc to sift through :-) This code is safe(r) to destroy,
though, since other stuff will supposedly not be influenced.
This is about a month since I begun exploring the issue (I attend to this
mostly on my commute or during free time in some evenings, so it goes slowly).
Because of that, I became too fixated on a task of "getting xev not to show
group 2 instead of group 1" (in the state bits). What I needed to remember was
that my task could also be "getting XWayland's implementation of XKB group not
interfere with the compositor's one".
Can I have some fun, please? Imagine a James Veitch's style of lines appearing here…
static void
keyboard_handle_modifiers(void *data, struct wl_keyboard *keyboard,
uint32_t serial, uint32_t mods_depressed,
uint32_t mods_latched, uint32_t mods_locked,
uint32_t group)
{
/*lol*/group = 0;
// the rest of function…
Okay, xev tells me this is actually the way. Can I have more fun?
/*rofl*/return;
Oh yes I can! Even with this piece of heresy xev resolves state correctly,
even with combination with other modifiers like control and shift, three
groups, …
This seems to be a sensible way, because the Wayland protocol seems to promise
to send wl_keyboard::key before wl_keyboard::modifiers:
If this event [wl_keyboard::key] produces a change in modifiers, then the
resulting wl_keyboard.modifiers event must be sent after this event.
My hypothesis is that XWayland's XKB updates its state when receiving the ::key
events and the following ::modifiers event confuses it. However, I do not yet
know what that code is supposed to do.