ladybird/Documentation/ProcessArchitecture.md
Sam Atkins 86744449e3 Documentation: Flatten the Browser/ directory
Separating out the browser documentation doesn't make much sense now
that we are only a browser. :^)
2024-10-23 09:02:46 +02:00

3.3 KiB

SerenityOS Browser process architecture

NOTE: This document is partly aspirational, in that the state of the code does not yet fully reflect what's described here. Implementation is underway.

The SerenityOS web browser ("Browser") uses a multi-process architecture to improve stability and security in the face of arbitrary (and possibly hostile) web content.

Process overview

Every instance of the Browser application can have one or more tabs open. Each tab has a unique WebContent service process spawned on its behalf.

Two important aspects of web browsing are further separated from the WebContent process: network requests and image decoding, segregated to the RequestServer and ImageDecoder processes respectively.

All processes are aggressively sandboxed using the pledge() and unveil() mechanisms. Furthermore, all processes except Browser run as an unprivileged user, separate from the primary logged-in desktop user.

Process: WebContent

This process hosts the main HTML/CSS engine (LibWeb.) It also runs JavaScript (LibJS.) It gets input events from Browser and paints the web content into shared bitmaps. It can only communicate with the outside world via RequestServer.

Process: RequestServer

This process can use networking protocols (like HTTP or HTTPS) to request files from the outside world. Each WebContent process gets its own RequestServer to do uploading or downloading on its behalf.

For DNS lookups, RequestServer asks for help from the system's global LookupServer service, which handles all outgoing DNS requests.

Process: ImageDecoder

This process can decode images (PNG, JPEG, BMP, ICO, PBM, etc.) into bitmaps. Each image is decoded in a fresh ImageDecoder process. These are strongly sandboxed and can't do much except receive encoded bitmap data and return a bitmap to WebContent if decoding is successful.

How processes are spawned

To get a fresh WebContent process, anyone with the suitable file system permissions can spawn one by connecting to the socket at /tmp/session/%sid/portal/webcontent, with %sid being the current login session id. This socket is managed by SystemServer and will spawn a new instance of WebContent for every connection.

The same basic concept applies to RequestServer and ImageDecoder as well, except that those services are spawned by WebContent as needed, not by Browser.

Class overview

In the GUI application process, a OutOfProcessWebView widget is placed somewhere in a window, and it takes care of spawning all of the helper processes, etc.

Internally, the OutOfProcessWebView has a WebContentClient object that implements the client side of the WebContent IPC protocol.

The WebContentClient speaks to a WebContent::ConnectionFromClient in the WebContent process. Internally, the WebContent::ConnectionFromClient has a WebContent::PageHost which hosts the LibWeb engine's main Web::Page object.

Inside LibWeb, a Web::Page has a main Web::Frame, which may have subframes corresponding to <frame> or <iframe> HTML elements. Each Web::Frame has a Web::Document, which is the root node of the DOM tree.

Describing the LibWeb object model in detail is outside the scope of this document.