Skip to main content
Ladybird’s multi-process architecture depends on two tightly coupled libraries: LibCore, which abstracts the operating system and drives asynchronous I/O through a cooperative event loop, and LibIPC, which layers type-safe message passing on top of UNIX sockets so that the browser’s isolated processes can communicate without sharing memory. Together they underpin nearly every long-running component in the browser — from the UI process to the per-tab WebContent renderer, the ImageDecoder, and the RequestServer.
LibCore is the POSIX abstraction layer used throughout Ladybird. It wraps file descriptors, processes, sockets, timers, and signals behind C++ classes that integrate with the event loop. It targets Linux, macOS, and (via separate implementation files) Windows and other POSIX systems.

​
EventLoop

Core::EventLoop is the heart of any LibCore application. It runs a cooperative, single-threaded event dispatcher: exec() enters the loop, repeatedly calling pump(), which blocks in select(2) until an event is ready, then dispatches all pending callbacks before sleeping again.Key operations on EventLoop:
  • exec() — enters the event loop and runs until quit() is called. Returns an exit code, so the pattern return app.exec() is conventional in GUI entrypoints.
  • pump(WaitMode) — processes one batch of pending events. Called by exec() in a loop. The WaitMode controls whether pump() uses select(2) to sleep until an event is ready (WaitForEvents) or polls without blocking (PollForEvents). Under the hood, pump() calls into the platform-specific EventLoopImplementation which invokes select(2) on registered notifier file descriptors and the wake pipe, using the nearest timer deadline as the select(2) timeout.
  • deferred_invoke(callback) — schedules callback to run on the next event loop iteration without blocking the caller.
  • wake() — writes to the event loop’s wake pipe so that a sleeping select(2) call returns immediately, causing the loop to process any newly posted events. This is the safe way to signal an event loop from another thread.
  • EventLoop::register_signal(signo, handler) — registers a POSIX signal handler that fires as a normal event on the calling thread’s event loop, avoiding the re-entrancy pitfalls of raw POSIX signal handlers.
The event loop maintains a per-thread event loop stack. Nested event loops (e.g. for modal windows in GUI applications) push onto this stack; when a nested loop exits, its pending events are returned to the lower loop’s queue. EventLoop::current() returns the topmost loop on the calling thread.Thread safety and wake pipeEvent loops are per-thread: all global state (notifiers, timers, event loop stack) is stored in thread-local variables. To safely signal an event loop running on another thread, call wake(), which writes to an internal pipe that select(2) is watching. This causes the sleeping loop to wake and process the new event.

​
Core::Timer

Core::Timer wraps the event loop’s timer registration behind a convenient class. Set an interval in milliseconds, attach a callback with on_timeout, and start the timer. Timers can fire once or repeatedly.

​
Core::Notifier

Core::Notifier fires a callback whenever a file descriptor becomes readable or writable. It registers the fd with the event loop’s select(2) watch set so that the loop wakes up and dispatches the on_activation callback as soon as the fd is ready.

​
Other LibCore Utilities

​
Dos and Don’ts with EventLoop

DO NOT store an EventLoop in a global variable. The event loop relies on its own global variable initialisation, so storing it globally can trigger an initialisation-order fiasco that UBSAN will catch. DO create your main event loop inside main() (or ladybird_main()) and pass it to objects that need it.DO NOT access EventLoop::current() unless you know which thread you are on. If no event loop exists on the current thread the program will crash. DO receive a reference to the specific event loop you need at construction time.DO NOT call pump() or exec() on an event loop belonging to a different thread. Sleeping and waking rely on thread-local variables. DO use wake() or deferred_invoke() to post work to another thread’s event loop.
LibIPC enables Ladybird’s separate processes to communicate by exchanging typed messages over UNIX-domain socket pairs. Each service defines its own IPC protocol; the framework handles serialisation, deserialisation, and connection lifecycle.

​
Architecture

Ladybird uses a multi-process architecture:
  • Main UI process — coordinates the browser window and user interaction.
  • WebContent (one per tab) — renders pages and runs JavaScript in a sandboxed process.
  • ImageDecoder — decodes images out-of-process to limit the blast radius of malformed image data.
  • RequestServer — performs network connections on behalf of the renderer, keeping raw socket access out of the sandboxed tab process.
Each service boundary is represented by a pair of classes generated from an IPC protocol definition:
  • IPC::ConnectionFromClient — the server side of a connection, representing one connected client.
  • IPC::ConnectionToServer — the client side, used by the process that initiated the connection.
Both sides share an IPC::Connection base that owns the underlying Core::LocalSocket and integrates with the event loop via a Core::Notifier on the socket’s file descriptor.

​
Message Passing

IPC::Message is the base type for all IPC messages. Each message carries a message ID and a serialised payload. The framework provides IPC::Encoder and IPC::Decoder for converting C++ types to and from the binary wire format.Messages can be:
  • One-way (fire-and-forget) — the caller sends the message and continues immediately.
  • Synchronous (blocking call) — the caller sends a request message and pumps the event loop in a nested fashion until the reply arrives.

​
Transport and File Descriptor Passing

IPC::Transport abstracts the underlying channel. On POSIX systems this is a socketpair(2) UNIX-domain socket, which also supports sending open file descriptors via SCM_RIGHTS — used to pass AnonymousBuffer handles (shared bitmaps) and other OS resources between processes without data copying.IPC::Attachment encapsulates file descriptors that travel alongside a message.

​
Service Roles

Because each tab’s WebContent process is fully isolated, a crash or hang in one tab’s renderer does not bring down the rest of the browser — the UI process simply tears down the dead IPC connection and spawns a replacement.