Multi-process Architecture
Last updated
Last updated
This document describes Chromium's high-level architecture.
It's nearly impossible to build a rendering engine that never crashes or hangs. It's also nearly impossible to build a rendering engine that is perfectly secure.
In some ways, the current state of web browsers is like that of the single-user, co-operatively multi-tasked operating systems of the past. As a misbehaving application in such an operating system could take down the entire system, so can a misbehaving web page in a modern web browser. All it takes is one browser or plug-in bug to bring down the entire browser and all of the currently running tabs.
Modern operating systems are more robust because they put applications into separate processes that are walled off from one another. A crash in one application generally does not impair other applications or the integrity of the operating system, and each user's access to other users' data is restricted.
We use separate processes for browser tabs to protect the overall application from bugs and glitches in the rendering engine. We also restrict access from each rendering engine process to others and to the rest of the system. In some ways, this brings to web browsing the benefits that memory protection and access control brought to operating systems.
We refer to the main process that runs the UI and manages tab and plugin processes as the "browser process" or "browser." Likewise, the tab-specific processes are called "render processes" or "renderers." The renderers use the WebKit open-source layout engine for interpreting and laying out HTML.
Each render process has a global RenderProcess object that manages communication with the parent browser process and maintains global state. The browser maintains a corresponding RenderProcessHost for each render process, which manages browser state and communication for the renderer. The browser and the renderers communicate using Chromium's IPC system.
Each render process has one or more RenderView objects, managed by the RenderProcess, which correspond to tabs of content. The corresponding RenderProcessHost maintains a RenderViewHost corresponding to each view in the renderer. Each view is given a view ID that is used to differentiate multiple views in the same renderer. These IDs are unique inside one renderer but not within the browser, so identifying a view requires a RenderProcessHost and a view ID. Communication from the browser to a specific tab of content is done through these RenderViewHost objects, which know how to send messages through their RenderProcessHost to the RenderProcess and on to the RenderView.
In the render process:
The RenderProcess handles IPC with the corresponding RenderProcessHost in the browser. There is exactly one RenderProcess object per render process. This is how all browser ↔ renderer communication happens.
The RenderView object communicates with its corresponding RenderViewHost in the browser process (via the RenderProcess), and our WebKit embedding layer. This object represents the contents of one web page in a tab or popup window
In the browser process:
The Browser object represents a top-level browser window.
The RenderProcessHost object represents the browser side of a single browser ↔ renderer IPC connection. There is one RenderProcessHost in the browser process for each render process.
The RenderViewHost object encapsulates communication with the remote RenderView, and RenderWidgetHost handles the input and painting for RenderWidget in the browser.
For more detailed information on how this embedding works, see the How Chromium displays web pages design document.
In general, each new window or tab opens in a new process. The browser will spawn a new process and instruct it to create a single RenderView.
Sometimes it is necessary or desirable to share the render process between tabs or windows. A web application opens a new window that it expects to communicate with synchronously, for example, using window.open in JavaScript. In this case, when we create a new window or tab, we need to reuse the process that the window was opened with. We also have strategies to assign new tabs to existing processes if the total number of processes is too large, or if the user already has a process open navigated to that domain. These strategies are described in Process Models.
Each IPC connection to a browser process watches the process handles. If these handles are signaled, the render process has crashed and the tabs are notified of the crash. For now, we show a "sad tab" screen that notifies the user that the renderer has crashed. The page can be reloaded by pressing the reload button or by starting a new navigation. When this happens, we notice that there is no process and create a new one.
Given Webkit is running in a separate process, we have the opportunity to restrict its access to system resources. For example, we can ensure that the renderer's only access to the network is via its parent browser process. Likewise, we can restrict its access to the filesystem using the host operating system's built-in permissions.
In addition to restricting the renderer's access to the filesystem and network, we can also place limitations on its access to the user's display and related objects. We run each render process on a separate Windows "Desktop" which is not visible to the user. This prevents a compromised renderer from opening new windows or capturing keystrokes.
Given renderers running in separate processes, it becomes straightforward to treat hidden tabs as lower priority. Normally, minimized processes on Windows have their memory automatically put into a pool of "available memory." In low-memory situations, Windows will swap this memory to disk before it swaps out higher-priority memory, helping to keep the user-visible programs more responsive. We can apply this same principle to hidden tabs. When a render process has no top-level tabs, we can release that process's "working set" size as a hint to the system to swap that memory out to disk first if necessary. Because we found that reducing the working set size also reduces tab switching performance when the user is switching between two tabs, we release this memory gradually. This means that if the user switches back to a recently used tab, that tab's memory is more likely to be paged in than less recently used tabs. Users with enough memory to run all their programs will not notice this process at all: Windows will only actually reclaim such data if it needs it, so there is no performance hit when there is ample memory.
This helps us get a more optimal memory footprint in low-memory situations. The memory associated with seldom-used background tabs can get entirely swapped out while foreground tabs' data can be entirely loaded into memory. In contrast, a single-process browser will have all tabs' data randomly distributed in its memory, and it is impossible to separate the used and unused data so cleanly, wasting both memory and performance.
Firefox-style NPAPI plug-ins run in their own process, separate from renderers. This is described in detail in Plugin Architecture.
Historically, new features (i.e. autofill to pick an example) have been added by bolting on their code to RenderView in the renderer process, and RenderViewHost in the browser process. If a feature was handled on the IO thread in the browser process, then its IPC messages were dispatched in BrowserMessageFilter. RenderViewHost would often dispatch the IPC message only to call WebContents (through the RenderViewHostDelegate interface), which would then call to another piece of code. All the IPC messages between the browser and renderer were declared in a massive render_messages_internal.h. Touching each of these files for every feature meant that the classes became bloated.
We have added helper classes and mechanisms for filtering IPC messages on each of the above threads. This makes it easier to write self contained features.
If you want to filter and send IPC messages, implement the RenderViewObserver interface (content/renderer/render_view_observer.h). The RenderViewObserver base class takes a RenderView and manages the object's lifetime so that it's tied to RenderView (this is overridable). The class can then filter and send IPC messages, and additionally gets notification about frame loading and closing, which many features need. As an example, see ChromeExtensionHelper (chrome/renderer/extensions/chrome_extension_helper.h).
If your feature has part of the code in WebKit, avoid having callbacks go through WebViewClient interface so that we don't bloat it. Consider creating a new WebKit interface that the WebKit code calls, and have the renderer side class implement it. As an example, see WebAutoFillClient (WebKit/chromium/public/WebAutoFillClient.h).
The WebContentsObserver (content/public/browser/web_contents_observer.h) interface allows objects on the UI thread to filter IPC messages and also gives notifications about frame navigation. As an example, see TabHelper (chrome/browser/extensions/tab_helper.h).
To filter and send IPC messages on other browser threads, such as IO/FILE/WEBKIT etc, implement BrowserMessageFilter interface (content/browser/browser_message_filter.h). The BrowserRenderProcessHost object creates and adds the filters in its CreateMessageFilters function.
In general, if a feature has more than a few IPC messages, they should be moved into a separate file (i.e. not be added to render_messages_internal.h). This also helps with filtering messages on a thread other than the IO thread. As an example, see content/common/pepper_file_messages.h. This allows their filter, PepperFileMessageFilter, to easily send the messages to the file thread without having to specify their IDs in multiple places.
void PepperFileMessageFilter::OverrideThreadForMessage(
const IPC::Message& message,
BrowserThread::ID* thread) {
if (IPC_MESSAGE_CLASS(message) == PepperFileMsgStart)
*thread = BrowserThread::FILE;
}