komorebi/docs/dev/ARCHITECTURE.md

Architecture

komorebi is a tiling window manager. It decides where every window goes and asks a compositor to put it there. Everything it needs from the compositor is behind one trait, Backend, so the window manager is written, and tested, without one.

This page describes that trait, who implements and calls it, where each of its rules is enforced, the model of the desktop the window manager tiles, the backend that reaches KWin, and how an event about a window becomes a tile. It grows with the code: each part of the daemon that is added has its place here.

Vocabulary

• Window: a top-level surface a client has asked the compositor for, named by the compositor's u32 id.
• Monitor: a connected output, with its bounds and its work area (its bounds minus the panels).
• Workspace: the containers a monitor shows, in order, and the layout that tiles them. A monitor has a ring of workspaces and shows one of them.
• Container: what a layout places. It gets one tile, and every window in it is put on that tile.
• Layout: the arrangement (BSP, columns, rows, ...) that turns a work area and a number of containers into one rectangle, a tile, for each container.
• Event: what the compositor reports about a window: it was shown, destroyed, given the focus, minimized or restored (WindowManagerEvent).
• Compositor: the program that owns the screen and the windows. On KDE Plasma it is KWin.
• Backend: the implementation of the Backend trait that talks to a compositor.
• Fake: FakeBackend, the in-memory backend the tests use.
• Bridge: the script komorebi loads into KWin (bridge.js). It reports what KWin tells it and carries out what the daemon asks.
• Table: the KWin backend's model of the desktop (Table), which the bridge's events keep up to date and the queries are answered from.
• Held call: the Poll call the bridge always has open, which the daemon answers when it has commands for the bridge, or after 10 seconds with none.

Design

The window manager reads the desktop and changes it through Backend and through nothing else. A real backend answers from a table that compositor events keep up to date, and sends commands to the compositor. The fake answers from a table the test fills in, and records the commands. The window manager cannot tell them apart, which is the point: the same logic runs against both. The reasons are in DESIGN.md.

The model

The desktop, as the window manager sees it, is two lists and a pointer:

• Windows, each a WindowInfo: its id, process, title, app_id, executable, frame, whether it is minimized, whether it is a normal application window (as opposed to a dialog, a splash screen or a dock), and whether it is alive. A window that has been destroyed stays in the table for a short time with alive false, so that the event that mentions it can still be logged.
• Monitors, each a MonitorInfo: its id, name, an identity that survives a hot-plug, its bounds and its work area.
• The pointer, a position if the compositor says where it is, and the focus, the id of the window that has it.

A command is one of: move and resize a window, focus it, raise it, minimize it or bring it back, close it, and put the pointer in the middle of a rectangle.

The window manager tiles a tree of four levels, each level a Ring: a list with one element focused, so that "the focused workspace" or "the focused container" means the same thing at every level.

• A monitor (Monitor) is one of the backend's monitors: its bounds, its work area, its padding, an optional work area offset, and its workspaces. It shows the focused one.
• A workspace (Workspace) has a layout, its padding, and its containers, in order.
• A container (Container) has its windows. A new window gets a container of its own. A container can be locked, which keeps it at its index while others are inserted, removed or swapped around it.
• A window (Window) is a Copy handle that carries only the compositor's id. Everything else about the window, and every change to it, goes through the backend.

Tiling a workspace (Workspace::update) is a function of this tree and a work area. The monitor gives its focused workspace the work area, the padding to fall back to and the offset (WorkspaceGlobals). The workspace takes the offset and its padding off the work area, the layout gives one tile per container, in order, each shrunk by the container padding, and every window of a container is asked to move onto its tile with set_rect. A tile is clamped to the work area, so it cannot reach onto a neighboring monitor.

The WindowManager keeps the monitors, the windows the user minimized (they are in no workspace until they are restored), and an index from each window it manages to the monitor and workspace that have it. It is loaded from the backend at startup: every monitor, and every open window that is normal, alive and not minimized, on the focused workspace of the monitor its middle is on. After that it changes only when an event comes in, and each event ends with the workspace in front tiled again.

The KWin backend

KWinBackend is the backend of a real Plasma session. KWin has no protocol that lets another process read and set window geometry, so the backend is in two processes:

• The bridge, bridge.js, a script KWin runs inside itself. It subscribes to KWin's signals and sends each as a JSON Event call over the session bus, and it carries out the commands it gets as the answer to its held Poll call. When it starts it sends a snapshot of the desktop. It holds no window manager logic.
• The daemon's side, kwin, owns org.komorebi.Bridge on the session bus, with one object for Event and one for Poll. Event calls are handled one at a time, in the order they arrive, and never wait. A Poll call is held until there are commands, so that the bridge gets them as soon as they are queued, without a timer and without a call from the daemon into KWin.

Every decision about what an event means is made in the daemon, by the table (kwin::state): an event goes in, the table is updated, and what follows comes out, a window event for the window manager or the answer to a query. The table is pure (no D-Bus, no threads), so the unit tests drive it with the JSON the bridge sends.

The bridge is JavaScript, which no Rust build compiles, and KWin loads the file as it is, so it is checked without a build step. tsc checks it as strict TypeScript through its JSDoc comments, which type every message, against kwin-api.d.ts, the part of KWin's scripting API the bridge uses. bridge.test.js runs it with node --test against a fake of that API. The two sides share the messages of a session in fixtures: the bridge's tests check that it sends exactly those events, and answer it with those commands, and the Rust tests feed the events to the table and check that the backend sends those commands. A change to a message on one side fails a test on the other.

The path of an event: a KWin signal, the bridge's handler, Event(json) on the session bus, Table::handle, a WindowManagerEvent on the channel of window_manager_event_listener, the thread that listen_for_events starts, and WindowManager::process_event. The queries (windows, window, foreground_window, monitors) are answered from the table. cursor_pos is the exception: only the bridge can read the pointer, so it is asked through the held call, and given 250 ms to answer before the table's last known position is returned.

A command goes the other way. The backend looks up the UUID KWin knows the window by, queues a JSON command for the bridge, and returns. The held Poll call is answered with the queue, as one batch: the daemon waits a millisecond after the first command for the rest, so that a retile of every window of a workspace reaches KWin at once. The bridge carries the commands out in order and reports the ones it could not carry out (a window that closed in the meantime) as a failed event, which is logged. set_rect also records the frame it asked for in the table, so that the window manager sees it straight away; the geometry event that follows says what the client made of it. The pointer cannot be moved from a script: when it follows the focus, the backend invokes KWin's "Move Mouse to Focus" shortcut, which centers it on the window KWin has focused. When komorebi stops, the backend asks KWin to unload the bridge.

connect makes the backend: it connects to the session bus, loads the bridge into KWin, and returns once the snapshot has arrived, so that the window manager reads a full table when it starts. PLATFORM.md has the D-Bus names, the files and the environment variables; DESIGN.md has the reasons.

The parts

• backend: the trait, the two data types, and the process-wide handle (set_backend, backend, with_backend).
• backend::fake: the fake.
• ring: the Ring, and the macro that gives every type holding one the same accessors (containers(), focused_container_mut(), ...).
• monitor, workspace, container, window: the four levels of the tree.
• lockable_sequence: inserting, removing and swapping in a list where some elements are locked in place, which is how a workspace changes its containers.
• window_manager: the WindowManager and the operations on it (tile the workspace in front, tile every monitor, take a minimized window out).
• desktop: loads the monitors and the open windows from the backend.
• window_manager_event and process_event: the events, and what the window manager does with each.
• window_manager_event_listener: the channel that carries events from the backend to the window manager, and listen_for_events, the thread that takes them off it.
• kwin: the KWin backend, its D-Bus service and the held call; kwin::state, the table; and bridge.js, the bridge.
• main.rs: the daemon. It connects to KWin, or installs the fake with --fake-backend, loads and tiles the desktop, starts the thread that handles events, and waits to be stopped.

The handle is a process-wide singleton so that a window can be a small Copy value that carries only its id, and ask the backend for what it needs. A test replaces it for its own thread with with_backend, which puts the previous one back when the test is done, also if it panics.

Traits and roles

Backend is the compositor. It is implemented once for each compositor and once for the fake. The window manager is its only caller, through backend. The contract:

• Queries (windows, window, foreground_window, cursor_pos) answer from the backend's own table and never wait for the compositor, except the KWin backend's cursor_pos, which waits up to 250 ms for the bridge. monitors may fail.
• A command returns once the compositor has accepted it, not once the client has acted on it. The frame a window ends up with is what window_rect says afterwards, because a client may refuse a size.
• A command about a window that does not exist, or is no longer alive, is an error. It never panics. The KWin backend knows a closed window for a minute after it closed, and sends its commands to the bridge, which reports that it could not carry them out.
• shutdown, is_window, window_rect and monitor_from_point have defaults, written in terms of the queries. An implementor should not need to override them, except shutdown when it created something in the compositor that has to be given up.

What is guaranteed, and by what

Rule	Enforced by	Proof	Escape hatch
A backend implements every query and command	type system: the methods have no defaults	compile_fail,E0046 on Backend	none
A backend can be shared between threads	type system: Backend: Send + Sync + 'static	compile_fail,E0277 on Backend	none
The window manager asks for a backend that was never installed	runtime error: backend panics, naming set_backend and with_backend	# Panics on backend	none
A test's backend does not outlive the test	framework path: with_backend restores the previous one on drop	with_backend_puts_the_previous_backend_back_when_the_closure_panics	none
A command about a missing window is an error, not a panic	test	a_command_about_a_window_nobody_opened_is_an_error and a_command_about_a_closed_window_is_an_error against the fake, a_command_about_a_window_the_table_does_not_know_is_an_error_and_sends_nothing against the KWin backend	none
A locked container keeps its index	framework path: a Workspace inserts, removes and swaps its containers through LockableSequence	test_locked_containers_with_new_window and test_locked_containers_remove_window	containers_mut() is the VecDeque itself
A window komorebi does not manage is never moved	framework path: handle_event asks Window::should_manage before it acts on any event but a destroy	dialogs_and_other_non_normal_windows_are_left_alone	none
A destroyed window leaves whichever workspace has it	framework path: the destroy is looked up in the index of managed windows, which is rebuilt at startup and before and after every event	a_window_that_was_open_before_komorebi_started_is_removed_when_it_closes	none
The daemon does not start without KWin, unless it is told to run against the fake	startup check: connect fails, and the daemon exits with the error, code 1	without_a_session_bus_komorebi_exits_with_an_error_naming_it	--fake-backend
One komorebi per session	startup check: the session bus gives org.komorebi.Bridge to one connection	none	none
Events reach the window manager in the order KWin raised them	framework path: the Event interface is spawn = false, so the calls are handled inline, and one channel and one thread carry the events to the window manager	vm-kwin-events ("the Show came first")	none
No event names a window before the event that announces it	convention in the bridge, which announces a window it has not seen when it is activated; the table shows a window once	a_new_window_raises_show_and_then_focus_in_that_order, announcing_the_same_window_twice_shows_it_once, vm-kwin-events	none
A window's id names one window while the daemon runs	framework path: the table numbers each KWin window the first time it sees it, and never gives a number out twice	ids_are_assigned_once_per_window_and_survive_a_new_snapshot, tombstones_are_dropped_after_a_while_and_their_ids_are_not_reused	none
A backend never waits for the window manager	framework path: an event that does not fit in the channel (20 waiting) is dropped and logged	none	none
Every window is where komorebi asked for it, within the size its client keeps to	VM test	vm-core-e2e, vm-client-geometry	a client may refuse a size: the frame is what window_rect says
A retile reaches KWin as one batch	framework path: the answer to Poll waits a millisecond after the first command for the others	vm-latency	none
The window manager sees a frame it asked for at once	framework path: the KWin backend's set_rect records it in the table before KWin answers	a_requested_rect_is_visible_at_once_and_replaced_by_the_real_one	none
The bridge holds no window manager logic	convention, with every decision about an event in the table, which the kwin::state unit tests pin	the unit tests in komorebi/src/kwin/state.rs	none
What the bridge sends is what the table reads	type checker: the JSDoc types of the messages in bridge.js; tests: the events of fixtures are what the bridge's tests check it sends and what the table's tests feed it	what the bridge sends for a session is what the daemon's tests read, the_bridges_events_for_a_session_are_the_window_events_komorebi_expects	none
What the backend sends is what the bridge carries out	tests: the commands of fixtures are what the backend's tests check it sends and what the bridge's tests answer it with, checking what KWin is asked to do for each	the_commands_of_a_session_go_to_the_bridge_as_the_bridges_tests_answer_it, what the bridge sends for a session is what the daemon's tests read, a command for a window KWin no longer has is given back, and the rest are carried out	none
Everything the bridge sends over D-Bus is one string	type checker: kwin-api.d.ts gives callDBus one string argument	tsc, every call goes to the daemon's names, with one JSON string	none
The bridge uses nothing a KWin script does not have	type checker: the library is ECMAScript 2016 alone, and KWin's API is what kwin-api.d.ts declares	tsc	a declaration KWin does not match: the VM tests run the bridge in KWin
Queries never block on the compositor	convention only	none	the KWin backend's cursor_pos waits up to 250 ms
Only the backend module reaches the compositor	convention only	none	none

The rows enforced by convention are rules nothing checks. A real backend that waits inside a query breaks one of them, a module that talks to D-Bus directly breaks another, and a bridge that decides something in JavaScript breaks a third: the bridge's tests check what it sends, not that it decides nothing.

One event, end to end

A window is shown, and ends up on a tile:

1. KWin raises windowAdded. The bridge sends an added event with what KWin says about the window, and Table::handle gives the window an id, records it, and returns a WindowManagerEvent::Show carrying the window and the compositor signal it came from. The event goes on the channel, and the event thread calls WindowManager::process_event, which hands it to handle_event and logs how long that took. The end-to-end tests build the event themselves and call process_event.
2. handle_event asks Window::should_manage, which asks the backend's window(id) whether the window is alive and normal. If it is not, the event changes nothing.
3. The index of managed windows is rebuilt from the containers. If the workspace in front already has the window, or the index has it on another monitor or workspace, the event creates nothing.
4. Workspace::new_container_for_window puts the window in a new container after the focused one, inserted through LockableSequence so that a locked container keeps its index, and focuses the new container.
5. WindowManager::update_focused_workspace asks the focused Monitor to give its focused workspace the work area, the padding and the offset (WorkspaceGlobals), and calls Workspace::update.
6. Workspace::update asks the layout for one tile per container, clamps each to the work area, and calls Window::set_position for each window, which is the backend's set_rect. The fake puts the window there at once; the frame a window ends up with is what window_rect says afterwards. The KWin backend queues a geo command for the bridge, the bridge sets the window's frame, the client draws at the size it accepts, and KWin reports the frame it ended up with as a geometry event, which updates the table.
7. If the new container is the only one, the window is given the focus with the backend's focus, and the pointer follows it (on KWin, through its "Move Mouse to Focus" shortcut).
8. The index is rebuilt again, so the next event finds the window.

Where to look in the code

• komorebi/src/backend/mod.rs: the trait.
• komorebi/src/backend/fake.rs: the fake and its tests.
• komorebi/src/core: the vocabulary the rest is written in.
• komorebi/src/workspace.rs: tiling, in Workspace::update.
• komorebi/src/monitor.rs: the work area, padding and offset a monitor gives the workspace it shows.
• komorebi/src/process_event.rs: what each event does, and the thread that handles them.
• komorebi/src/kwin/mod.rs: the KWin backend, its D-Bus service and the held call.
• komorebi/src/kwin/state.rs: the table, with the shapes of the bridge's events, and its tests.
• komorebi/src/kwin/bridge.js: the bridge, with the types of its messages; kwin-api.d.ts and tsconfig.json, what tsc checks it against; bridge.test.js, its tests; fixtures, the messages both sides' tests read.
• nix/tests/vm/suites/kwin-events: the events of a real KWin, through the daemon.
• nix/tests/vm/suites/core-e2e: the daemon tiling a real KWin, checked against what KWin has.
• komorebi/tests/wm_events.rs: the window manager driven end to end against the fake.