14 fuzz tests (pytest -m fuzz) that launch foot, alacritty,
gnome-calculator, evince, xmessage, and combinations thereof inside
kick, then exercise basic interactions (cycle, overlay, close, rapid
cycling, tab-groups stacking, unresponsive close).
Excluded from the default pytest run via the fuzz marker, like flaky.
1. Four identical set_title/set_app_id handlers (toplevel, sub, xdg)
collapsed into handle_overlay_refresh().
2. Map activation decision extracted to map_should_activate(), shared
by both the xdg and X11 map handlers.
3. Timer cleanup removed from finish_app_destroy() -- remove_app()
already handles it unconditionally.
meson setup --reconfigure touches subprojects/.wraplock every run,
making it newer than build/kick and triggering an infinite rebuild
loop. The file is meson housekeeping, not a real source dependency.
Three places detached signal listeners from tab-group members:
handle_sub_destroy, handle_sub_x_destroy, and remove_app. The
remove_app copy had already drifted (missing set_title/set_app_id).
Extract two helpers so the logic lives in one place.
When remove_app() force-cleans a tab's group members (e.g. the main
toplevel's destruction cascades before a sub-member's own destroy fires),
it removed destroy/map/unmap/commit but not set_title/set_app_id. The
freed member's listener stayed on the sub-toplevel's signal list; when
wlroots later destroyed that toplevel the non-empty listener assertion
fired.
Crash repro: foot+foot tab, run gnome-calculator in one terminal, cycle
to the main surface, Alt+Delete. Both processes die from SIGTERM;
remove_app() frees the calculator member without detaching set_title,
and wlroots asserts on the calculator toplevel teardown.
enabling direct scanout everywhere reintroduced the exact failure the
global disable once guarded against: in nested mode a "scanned-out"
buffer on the headless output is displayed nowhere by itself -- kick
must blit every presented buffer into the host window, importing the
client's dmabuf into the host EGL context per frame. vlc's
scanout-feedback-tuned buffers reject that import per frame
(EGL_BAD_PARAMETER "modifier attributes not equal"), leaving black video
behind playing audio.
policy is now per mode:
- standalone (DRM/VT console): scanout enabled -- true zero-copy
presentation with no composite or blit. an attach-render lock is held
while the tab overlay is visible so the overlay's second pass lands on
the composited frame instead of being painted onto the client's own
buffer, and recreated outputs inherit the lock.
- nested: always composites. KICK_NESTED_SCANOUT=1 re-enables scanout
there for testing.
also scope the software-cursor lock to headless standalone outputs:
their set_cursor hook pretends a hardware plane exists while drawing
nothing, so the lock was needed there -- but on real DRM it made direct
scanout permanently ineligible (a composited cursor cannot survive
scanout), defeating zero-copy fullscreen entirely.
direct scanout was globally disabled after nested-mode black frames
(Firefox): the scene handed the client's own buffer to the output, and
the host-window blit could not keep it in sync. wlroots 0.19.1 forwards
explicit-sync timelines in the scanned-out state and re-evaluates
eligibility every frame, so the global kill is no longer warranted:
- remove the WLR_SCENE_DISABLE_DIRECT_SCANOUT env: scanout now engages
whenever a single opaque client covers the output -- fullscreen video,
full-bleed tabs -- and skips the scene composite entirely
- hold an attach-render lock while the tab overlay is visible so the
overlay's second pass always lands on kick's composite (the same
lock/unlock pattern labwc uses for its own panels), releasing it the
moment the prefix is released
- recreated outputs inherit the lock if the overlay is up
verified under labwc: scan-out enable/disable logged around overlay
transitions during playback, and fullscreen placement probes fill the
monitor correctly with the resizing test client.
destroying the previous wl_buffer immediately after attaching its
replacement raced the compositor's use of the last committed frame and
pulled the rug from under data-pointer access to buffers still on
screen -- black frames exactly when buffers turned over. a few KB of
shm per resize is an acceptable price for a short-lived test client;
resizes are rare.
libdecor sat on the critical path of every host-window transition even
where it drew nothing: on compositors honoring server-side decorations
(labwc/sway/KDE) it negotiated SSD, then still injected permanent
set_window_geometry hints and batched acks inside its own commit call.
labwc 0.9.8 expires a 100 ms configure deadline under load and answers
"move but keep old size" -- anchoring fullscreen views to stale geometry
forever.
the host window now selects its backend at startup: compositors
advertising zxdg_decoration_manager_v1 get kick-owned xdg_shell objects
with immediate configure acks, an explicit server-side decoration request,
and no geometry hints at all (KICK_LIBDECOR=1 forces the libdecor path);
hosts without the manager -- Mutter/GNOME forces CSD -- keep libdecor
unchanged. this also restores mirrored fullscreen semantics: a client's
fullscreen request again fullscreens the host window in one step.
create_buffer() early-returned once a buffer existed, so the client kept
attaching its original 1280x800 pixels no matter what size the compositor
configured -- every resize/fullscreen test was silently verifying against
a client that cannot resize, painting old-extent content inside whatever
window state the compositor applied.
_ensure_built() moves from Session.__init__/_spawn_standalone to module
import: a broken build now aborts pytest at collection with one clear
error, rather than erroring every test at setup (which made failing-tree
runs slow and noisy). direct-import scripts keep their coverage since
they import this module too.
the feature itself works interactively, but the test depends on which
member holds focus when the stack maps -- a startup race between the
main window activating and the extra member entering -- so it failed
deterministically in automated runs. removed rather than papered over;
the chord is still exercised manually.
tests could silently run against a stale binary when the tree failed to
compile (the suite stayed green against the previous build). Session
construction and _spawn_standalone now run ./build.sh -- incremental and
a no-op when up to date -- and raise immediately on failure, so a broken
tree errors every spawning test at setup with the compiler output above
it instead. test.sh drops its build step; pytest triggers rebuilds
itself, which also covers scripts that import this module directly.
a client's fullscreen request used to be mirrored onto the host window
via libdecor, but labwc 0.9.8 expires its 100 ms configure deadline under
load and then applies "move but keep old size" permanently -- the
enlarged image anchored at the old origin with black margins. the mirror
is gone: a client's fullscreen now fills kick's output only, leaving the
host window alone. fullscreening the host remains an explicit act
(prefix+F11 or the compositor), and the host-to-clients propagation is
unchanged.
three changes that make our side of the fullscreen round-trip as fast
and deterministic as the protocol allows:
- queue libdecor fullscreen requests until the present pipeline drains
(frame callback received), so a throttled eglSwapBuffers started
before the resize can no longer complete after the ack and marshal a
stale-size frame past it; rapid requests collapse to the latest target
- flush the ack immediately instead of waiting for the end-of-loop
flush, which can land after labwc's 100 ms configure-response deadline
- allocate the resized swapchain during host_resize rather than lazily
at first render, moving the expensive GBM allocation ahead of the ack
labwc 0.9.8 can still expire its deadline under load and then applies
"move but keep old size" permanently; these shrink that window without
curing it.
wlr_renderer_init_wl_display() is the legacy all-in-one helper: it creates
its own wl_shm (v2) and linux-dmabuf (v4) globals, which duplicated the
explicit wl_shm v1 and linux-dmabuf v5 created further down. clients that
keep per-instance state -- Qt's Wayland platform re-enters its dmabuf
feedback setup for every advertised instance, and VLC's wl_dmabuf video
output likewise -- misbehave on the duplicates: vlc's Qt UI segfaulted at
startup and, once running, its video stayed black behind playing audio.
drop the legacy call and keep the explicitly versioned pair.
x11_pending_teardown only removed the pending's associate listener when
the surface had been hooked (associated). the listener is attached
unconditionally at new_surface, so an X11 window destroyed before ever
associating -- VLC's short-lived probe windows do exactly that -- left
it linked on events.associate, and wlroots' xwayland_surface_destroy
asserted on the non-empty list, killing kick.
- Alt+Shift+Left/Right moves the focused member left/right in its group's
window list, analogous to Alt+Shift+Up/Down tab reordering
- Wraps at ends (moving right from newest lands at front)
- Overlay summoned like other navigation chords
- Terminals never see Shift+arrow (chord is consumed)
- Wrap-ternary type-error eliminated; unused 'wrapped' variable removed
Kick advertised only pointer+keyboard and ignored every touch device, so
clients never saw a touchscreen and taps were silently dropped.
The seat now advertises WL_SEAT_CAPABILITY_TOUCH. Standalone mode wires
libinput touch devices through wlr_cursor with normalized-absolute warping
and delivers via the seat touch API with slot ids preserved (multi-touch).
Nested mode binds wl_touch when the host seat offers the capability,
forwards events through the same coordinate space as the pointer path, and
ends live sequences cleanly on capability loss or host cancel.
The ctl test feed gains CTL_TOUCH so headless tests can inject taps, and
wltest learns --touch-log to record received events; an end-to-end
regression test proves down/motion/up reach the client with correct slot
and coordinates.
host_frame_configure followed every libdecor_frame_commit (the ack) with
a bare wl_surface_commit, which re-committed the still-attached previous
frame. labwc applies the acked state at that first commit, so entering
fullscreen anchored the new layout to the stale window size: the enlarged
image painted from the old origin with black margins top/left and cropped
bottom/right. pre-libdecor, the ack went out on its own listener long
before any commit, so the state-applying commit was always the next
correctly sized frame -- restore that ordering by only flushing the empty
commit for size-less initial configures, where it is still needed to get
libdecor mapped.
clients_set_fullscreen only told each tab's main surface, so members
stacked before a toggle never learned they were fullscreen, and members
that mapped during fullscreen were never released from it. walk each
tab's member list with a member_set_fullscreen helper (mirroring the app
one) and give xdg members the same re-maximize/resize mains get on the
way out; X11 members keep their own geometry. logged as
"members fullscreen=" next to the existing line.
a grabbed prefix+F11 chord flips the host window's fullscreen state via
libdecor; the resulting configure fans out to every tab through the
usual host->clients propagation, exactly like a host-compositor
fullscreen. latched against key repeats and a no-op standalone, where
there is no host window. plain unprefixed F11 still passes through to
clients.
e1b137b added the late-attach damage nudge only to the X11 map path, so
Wayland members -- GTK4 dialogs and menus -- raced their first frame:
black until an unrelated repaint (a tab switch) forced one. The nudge
now fires for every member arrival on the shown tab regardless of
backend. Found via git bisect with a deterministic same-client
second-toplevel pixel probe.
X11 windows now pair their Wayland surface after membership decisions run
(membership starts at first map), so instantly-mapped transients -- GTK
menu dropdowns -- attached without ever wiring their scene surface: emacs
menus did nothing, silently. The associate handlers now replay the map
when the surface is already mapped, members join the owner's list before
that replay (show_current walks the list to pick visibility), X11 members
keep their own geometry instead of being stretched maximized, and late
attaches force whole-scene damage so the first frame is never black.
Members render through wlr_scene_subsurface_tree_create; the ctl test feed
gains CTL_CMD_PTR_BUTTON so tests can click for real; tests/xpopup.c
probes the exact menu shape (instantly-mapped managed transient), and a
regression test locks in the attach/replay/focus lifecycle.
GDK_BACKEND/QT_QPA_PLATFORM forced to wayland stripped toolkits' fallback,
killing X11-only builds like emacs-gtk ('cannot open display' against a
healthy Xwayland). Children now self-select from WAYLAND_DISPLAY/DISPLAY,
and XAUTHORITY is cleared.
Separately, X11 membership decisions moved from window-creation time to
first map: GTK leader windows never map, yet claimed the tab slot at
creation, leaving a permanent black tab while the real frame was demoted
to a stacked member -- and Alt-cycling onto that phantom main member
crashed on wlr_seat_keyboard_notify_enter(NULL). Never-mapped windows now
free silently; keyboard-entry paths guard on live surfaces. New
tests/xleader.c probe (unmapped leader + mapped frame) locks the behaviour
in.
title and app_id arrive as protocol events with no pixel damage, so the
overlay (which reads them live) never repainted while held -- a tab going
from black to content kept showing the command until the menu was closed
and reopened. subscribe to the xdg/x11 set_title and set_app_id signals
(and refresh on map) and repaint in place when the overlay is visible.
a tab whose window has not opened yet has no mapped members, so its
overlay row drew nothing and showed a blank title. draw the launch
command instead, sized into the panel (which now measures and counts
pending rows) and truncated to the row width if it would overflow.
each window title now sits in its own grey box matching the blue
highlight size, lightened so it separates from the panel, and rows are
spaced so the boxes no longer touch.
the tab overlay now lists every window of a tab (main plus stack members)
with a per-window icon and title instead of one icon and title per tab, and
highlights only the active window so the overlay cursor can move across all
of a tab's windows. titles are also centered within each row.
set_active_app used to discard the tab's focus_member when leaving it,
so coming back always re-focused the newest mapped member instead of the
window the user was last on. keep the pointer on leave and restore it on
enter (falling back to the newest member only if it is no longer mapped).
