Multi-process Architecture

This document describes Chromium's high-level architecture.

Problem

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 state of web browsers around 2006 was 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 could a misbehaving web page in a web browser. All it took 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.

Architectural overview

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 Blink open-source layout engine for interpreting and laying out HTML.

Managing render processes

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.

Managing views

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.

Components and interfaces

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.

Sharing the render process

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.

Detecting crashed or misbehaving renderers

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.

Sandboxing the renderer

Given the renderer is running in a separate process, we have the opportunity to restrict its access to system resources via sandboxing. 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.

Giving back memory

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.

Plug-ins and Extensions

Firefox-style NPAPI plug-ins ran in their own process, separate from renderers. This is described in detail in Plugin Architecture

The Site Isolation project aims to provide more isolation between renderers, an early deliverable for this project includes running Chrome's HTML/JavaScript content extensions in isolated processes.

Multi-process Architecture的更多相关文章

  1. python multi process multi thread

    muti thread: python threading: https://docs.python.org/2/library/threading.html#thread-objects https ...

  2. 6 Multi-Cloud Architecture Designs for an Effective Cloud

    https://www.simform.com/multi-cloud-architecture/ Enterprises increasingly want to take advantage of ...

  3. Node.js进程管理之Process模块

    在前面Node.js事件运行机制也有提到,Node.js应用在单个线程运行,但是现在大部分服务器都是多处理器,为了方便使用多个进程,Node.js提供了3个模块.Process模块提供了访问正在运行的 ...

  4. Flash: An Efficient and Portable Web Server

    Introduction This paper presents the design of a new Web server architecture called the asymmetric m ...

  5. LIST OF NOSQL DATABASES [currently 150]

    http://nosql-database.org Core NoSQL Systems: [Mostly originated out of a Web 2.0 need] Wide Column ...

  6. 谷歌浏览器使用IE内核

    经常使用谷歌浏览器去到处溜达,可是,居然还不知道谷歌浏览的还有很多新鲜的设置,昨天遇到了调试一个脚本没有出现效果,而在IE中就可以显示所谓的效果: 查询了下,看到谷歌浏览器确实对脚本的部分样式要求还是 ...

  7. MySQL、MongoDB、Redis数据库Docker镜像制作

    MySQL.MongoDB.Redis数据库Docker镜像制作 在多台主机上进行数据库部署时,如果使用传统的MySQL的交互式的安装方式将会重复很多遍.如果做成镜像,那么我们只需要make once ...

  8. python 多线程和多进程基本写法

    #coding=utf-8 def aJob(arg): """ 提供给多线程调用 """ import threading t = thr ...

  9. Windows Server 2003 Sp2 下无法安装SQL Server 2008 Management Studio Express问题

    Windows Server 2003 Sp2 下无法安装SQL Server 2008 Management Studio Express问题钉子 发表于 2010-5-22 1:42:51问题描述 ...

  10. Python标准库10 多进程初步 (multiprocessing包)

    作者:Vamei 出处:http://www.cnblogs.com/vamei 欢迎转载,也请保留这段声明.谢谢! 我们已经见过了使用subprocess包来创建子进程,但这个包有两个很大的局限性: ...

随机推荐

  1. UI布局【转】

    转载自: https://www.cnblogs.com/wangdaijun/p/5519459.html https://www.jianshu.com/p/f781c40df57c Good U ...

  2. Python学习第二天-编写三级菜单

    编写三级菜单:1. 运行程序输出第一级菜单2. 选择一级菜单某项,输出二级菜单,同理输出三级菜单3. 菜单数据保存在文件中4. 让用户选择是否要退出5. 有返回上一级菜单的功能 # Author: z ...

  3. 不安装Oracle客户端,用plsql连接远程Oracle数据库(绝对解决你的问题)

    1,首先准备下载两个软件,一个是instantclient.zip,另一个是plsql安装包.但是得确定您的电脑是32位还是64位,我这边提供了32位和64位的供您下载: 百度网盘:https://p ...

  4. hihocoder 1124 : 好矩阵 dp

    好矩阵 时间限制:3000ms 单点时限:1000ms 内存限制:256MB 描写叙述 给定n, m.一个n × m矩阵是好矩阵当且仅当它的每一个位置都是非负整数,且每行每列的和 ≤ 2.求好矩阵的个 ...

  5. so near yet so far

    Dear little yang So beautiful boy as you, the most beautiful boy is you who i ever saw, like a sun , ...

  6. Configuration file schema for the .NET Framework

    https://docs.microsoft.com/en-us/dotnet/framework/configure-apps/file-schema/index Runtime Settings ...

  7. legend---八、php对象如何转换成js对象

    legend---八.php对象如何转换成js对象 一.总结 一句话总结:a.直接转换:b.通过json对象做中间桥梁 1.为什么传递给父亲构造函数的参数不能写默认值? 这里的第三行的比如$type不 ...

  8. tomcat和nginx相互结合的优化调整

    在工作中遇到这样的情况 Tomcat为后台 nginx为反向代理 需要往后台导入数据,由于处理时间过长,导致访问时出现504和500  通过修改tomcat中maxParameterCount=&qu ...

  9. List methods

    Python provides methods that operate on lists. For example, append adds a new element to the end of ...

  10. Spring Boot: Tuning your Undertow application for throughput--转

    原文地址:https://jmnarloch.wordpress.com/2016/04/26/spring-boot-tuning-your-undertow-application-for-thr ...