Android Interprocess Communication(一)

Abstract

This paper is an analysis of the interprocess communication in Android mobile operating system, provided by a custom software called Binder. It gives an overview about the capabilities and different layers of the Binder framework.

1. Introduction

A Gartner forecast [8] stated in late 2010, that Android will ”… challenge symbian for no. 1 position by 2014”.That means, that Android is an increasing factor in smartphone computing. Android introduces new extensions and features to the Linux kernel. The Binder framework for interprocess communication represents such a feature. For this framework a complete documentation does not exist. The Java classes are well documented, but it gets more fragmented and is completely missing for the kernel module.

This work tries to give an overview of the Binder framework, its features and functionalities.

Chapter 2 provides all needed background knowledge to understand the Binder framework. It gives an overview about Linux and outlines its basic concepts.

Chapter 3 introduces the Android operating system and its basic concepts.

Chapter 4 discusses the Binder and its capabilities. The explanation covers an abstract level, without going into implementation details.

The Chapter 5 discusses the different layers of the Binder framework and where its the features are implemented.

Chapter 6 gives an example of an applied Binder interprocess communication, where an Android client application accesses an Android remote service application and performs an remote procedure call.

2. Background

In this chapter we discuss the underlying theory. We give an overview on the topic of multitasking and processes and we point out why the concept of interprocess communication is needed.

2.1. Multitasking, Processes and Threads

Multitasking is the ability to execute multiple instances of programs or processes at the same time. An operating system therefore creates for every binary executable file a certain memory frame with its own stack, heap, data and shared mapped libraries. It also assigns special internal management structures. This is called a process.

The operating system must provide fair proportioning, because only one process can use the CPU at the same time. All processes must be interruptible. The operating system sends them to sleep or wakes them on their time slot. This work is done by a scheduler, supplying each process with an optimal time slot.

A thread is a process without own address space in memory, it shares the address space with the parent process. Processes can have child threads, and a thread must be assigned to a process.

2.2. Process Isolation

Due to security and safety reasons, one process must not manipulate the data of another process. For this purpose an operating system must integrate a concept for process isolation. In Linux, the virtual memory mechanism achieves that by assigning each process accesses to one linear and contiguous memory space. This virtual memory space is mapped to physical memory by the operating system. Each process has its own virtual memory space, so that a process cannot manipulate the memory space of another process. The memory access of a process is limited to its virtual memory. Only the operating system has access to physical and therefore all memory.

The process isolation ensures for each process memory security, but in many cases the communication between process is wanted and needed. The operating system must provide mechanisms to approve interprocess communication.

2.3. User Space and Kernel Space

Processes run normally in an unprivileged operation mode, that means they have no access to physical memory or devices. This operation mode is called in Linux user space. More abstractly, the concept of security boundaries of an operating system introduces the term ring. Note, that this must be a hardware supported feature of the platform. A certain group of rights is assigned to a ring. Intel hardware [21] supports four rings, but only two rings are used by Linux. These are ring 0 with full rights and ring 3 with least rights. Ring 1 and 2 are unused. System processes run in ring 0 and user processes in ring 3. If a process needs higher privileges, it must perform a transition from ring 3 to ring 0. The transition passes a gateway, that performs security checks on arguments. This transition is called system call and produces a certain amount of calculating overhead.

2.4. Interprocess Communication in Linux

If one process exchanges data with another process, it is called interprocess communication (IPC). Linux offers a variety of mechanisms for IPC. These are the following listed: [23]

Signals Oldest IPC method. A process can send signals to processes with the same uid and gid or in the same process group.

Pipes Pipes are unidirectional bytestreams that connect the standard output from one process with the standard input of another process.

Sockets A socket is an endpoint of bidirectional communication. Two processes can communicate with bytestreams by opening the same socket.

Message queues Processes can write a message to a message queue that is readable for other Processes.
Semaphores A semaphore is a shared variable that can be read and written by many processes.

Shared Memory A location in system memory mapped into virtual address spaces of two processes, that each process can fully access.

3. Android

In this chapter, the basic mechanisms and concepts of the mobile operating system Android are presented. Android was developed by the Open Handset Alliance and Google and is available since 2008. 1

3.1. Kernel

Android is based on a Linux 2.6 standard kernel but enhanced with new extensions for mobile needs. These are kernel modules Alarm, Ashmem, Binder, power management, Low Memory Killer, a kernel debugger and a logger. We will analyze the Binder driver in this work, that offers a new IPC mechanism to Linux. [17]

3.2. Programming Languages

Four programming languages are used for system development: Assembler, C, C++ and Java. The kernel has a small amount of Assembler but is mainly written in C. Some native applications and libraries are written in C++. All other applications, especially custom apps, are written in Java. [10]

3.3. Java Native Interface

A distinction is made between programs compiled for the virtual machine and programs compiled to run on a specific computation platform, like Intel x86 or ARM. Programs compiled for a specific platform are called native. Because Java is executed in a virtual machine with its own byte-code, no native code can be executed directly. Due to the need to access low-level os mechanism like kernel calls, Java has to overcome this obstacle. This is done by the Java native interface (JNI) [22], which allows Java to execute compiled code from libraries written in other languages, e.g. C++. This is a trade-off between gaining capabilities of accessing the system and decreasing the level of security in Java.

### 3.4. Dalvik Virtual Machine

The Dalvik virtual machine (DVM) [5] runs the Java programmed apps. The DVM does not claim to be a Java virtual machine (JVM) due to license reasons, but fulfills the same purpose. Java 5 programs can run in that environment.

The Sun JVM is stack based, because a stack machine can be run on every hardware. Hardware and platform independence were major design principles of Java. The DVM is register based for performance reasons and well adapted to ARM hardware. This is a different design principle, taking the advantage of hardware independence for high performance and less power consumption, which is essential for mobile purposes with limited battery capability. The possibility to use the Java native interface weakens the security guarantying property of Java to implicit checking the bounds of variables and to encapsulate system calls and the force to use JVM defined interfaces to the system. The use of native libraries can allow bypassing the type and border checking of the virtual machine and opens the door to stack-overflow attacks. [4]

Even it is a security issue, the JNI is essential for the interprocess communication mechanism because the middleware of Binder are C++ libraries and must be accessed with JNI.

3.5. Zygote

Due to performance reasons, the DVM is started only once. Each new instance of it is cloned. This is done by a system service called Zygote. 2

First, it preinitializes and preloads common Android classes in its heap. [25] Then, it listens on a socket for commands to start a new Android application. On receiving a start command, it forks a new process with the loaded application. This process becomes the started application and shares the heap with the original Zygote process by copy-on-write mapping and so the memory pages of Zygote’s heap are linked to this new process. While the application reads only from the heap, it stays shared. But when the application performs write operations on its heap, the corresponding memory page is copied and the link is changed to the new page. Now the heap can be manipulated, without manipulating the original data from the parent Zygote process.

When an Android application forks, it uses Zygote’s memory layout and therefore the layout is the same for each application.

3.6. Application Concept

Each Android application is composed from up to 4 different components. [12] Each component has a special subject. Figure 3.1 presents the components as a ierarchically class diagram since they are actually Java classes.

Android Interprocess Communication(一)_第1张图片

The activity represents the user interface of an application. It is responsible for performing the screen and receiving interaction created by the user. It is not intended to hold persistent data because it can be sent to sleep by the operating system if another activity is brought to the front.

For long duration purposes Android offers the service component. All tasks running in the background of an application must be implemented here, because a foreground service is only stopped if the system runs out of memory and apps must be terminated to free memory.

Even if the service is persistent in task performing, it is depreciated to hold persistent data. This is the subject of the content provider, which gives an interface for accessing persistent data like file or network streams as SQLlike databases.

The broadcast receiver is for receiving system wide messages, i.e. the message that a new SMS has come in is provided to all subscribers. A low battery level warning is also sent on this channel. Broadcast reveivers handle these messages and marshal certain action, e.g. saving the state of an app in prospect to a soon shutdown of the mobile device.

The application manifest [13] keeps the information for Android about the component. In this file, the basic application configuration is set. E.g., if an service starts in its own process or if it is attached to local process. Listing 3.1 gives an example of an Android application manifest XML file.

<?xml version="1.0" encoding="utf-8"?><manifest xmlns:android="http://schemas.android.com/apk/res/android"    package="com.msi.manning.binder">    <application android:icon="@drawable/icon">        <activity android:name=".ActivityExample" android:label="@string/app_name">            <intent-filter>                <action android:name="android.intent.action.MAIN" />                <category android:name="android.intent.category.LAUNCHER" />            </intent-filter>        </activity>        <service android:name=".SimpleMathService" />    </application></manifest>

3.7. Component Communication Concepts

As different components have to exchange data, this is realized through intercom- ponent communication, or interprocess communication, if the specific components belong to different processes (apps).

The communication works with so called intents. These are representations for operations to be performed. An intent is basically a datastructure which contains a URI and an action. The URI uniquely identifies an application component and the action identifies the operation to be executed.

This intent is submitted by the interprocess communication system. Figure 3.2 gives an overview about the different forms of component interaction. An activity is started by an intent and can bring another activity to the front.

A service can be started, stopped and bound by IPC. Also the call and return methods are implemented by IPC.

A content provider can be queried by an activity via IPC and returns the result accordingly. The Android source code files show an extensive use of IPC to exchange the abstract data.

A broadcast receiver gets all the intents (eg. messages) it has subscripted to via IPC. [12]

Android Interprocess Communication(一)_第2张图片

At this point, the importance of the IPC mechanism becomes apparent. The Android OS with its framework is a distributed system and the major and key technology to achieve that design is the IPC Binder mechanism.

3.8. Security Concept

The security mechanism in Android consists of three layers. The basic layer consists of a division of the persistent memory in two partitions, called system and data. The system partition is mounted as read only to prevent system data manipulation. The data partition is the place where application states and persistent data can be stored. Note, that the system partition can be remounted in write mode by the App Store application to install new apps.

To separate the apps from each other, the discretionary access control (DAC) model of Linux is used. Each app has its unique user ID and group ID and can only access its own directory. Only apps from the same author run under the same user ID and can access their data. All apps must be signed by the author to prevent data manipulation and to identify the author. So the system can determine if apps are from the same author and can run under one UID.

Since the Linux DAC model allows only a rudimentary level of access control, for fine granulated rights Android’s middleware offers a permission label system which implements a mandatory access control (MAC) model. This system is based on a set of permissions and a set of exported services including access restrictions.

For each action performed on the operating system a permission label exists. At installation time the application asks the user for a set of permissions, the user has the choice between granting all asked permissions or aborting the installation. Once granted at installation time, a permission can never be removed except by uninstalling the app.

Every app can specify an intent filter, a white list mechanism, that defines the types of intents the components of the application should receive. Intents that are not listed will be filtered out by the reference monitor.

A problem is that the apps of an author can communicate freely because of same UID and GID. That means, if multiple apps from the same author are installed on the phone the different rights of each app accumulate through transitive property. App A asks app B of the same author to perform an action, for which app A has no rights but app B. So app B gets in possession of rights that are not granted by user for this dedicated app. [6] [18] [16]

(continue…)