Android Security Internals: An In-Depth Guide to Android's Security Architecture

About the Author

Nikolay Elenkov has been working on enterprise security projects for the past 10 years. He has developed security software on various platforms, ranging from smart cards and HSMs to Windows and Linux servers. He became interested in Android shortly after the initial public release and has been developing applications for it since version 1.5. Nikolay’s interest in Android internals intensified after the release of Android 4.0 (Ice Cream Sandwich), and for the past three years he’s been documenting his findings and writing about Android security on his blog, http://nelenkov.blogspot.com/.

About the Technical Reviewer

Kenny Root has been a core contributor to the Android platform at Google since 2009, where his focus has been primarily on security and cryptography. He is the author of ConnectBot, the first SSH app for Android, and is an avid open source contributor. When he’s not hacking on software, he’s spending time with his wife and two boys. He is an alumnus of Stanford University, Columbia University, Chinese University of Hong Kong, and Baker College, but he’s originally from Kansas City, which has the best barbecue.

Foreword

I first became aware of the quality of Nikolay’s work in Android security with the release of Android 4.0, Ice Cream Sandwich. I needed a better explanation of the new Android backup format; I was struggling to exploit a vulnerability I had found, because I didn’t have a full grasp of the new feature and format. His clear, in-depth explanation helped me understand the issue, exploit the vulnerability, and get a patch into production devices quickly. I have since been a frequent visitor to his blog, often referring to it when I need a refresher.

While I was honored to be asked to write this foreword, I honestly didn’t believe I’d learn much from the book because I’ve been working on Android security for many years. This belief could not have been more wrong. As I read and digested new information regarding subjects I thought I knew thoroughly, my mind whirled with thoughts of what I had missed and what I could have done better. Why wasn’t a reference like this available when I first engrossed myself in Android?

This book exposes the reader to a wide range of security topics, from Android permissions and sandboxing to the Android SELinux implementation, SEAndroid. It provides excellent explanations of minute details and rarely seen features such as dm-verify. Like me, you’ll walk away from this book with a better understanding of Android security features.

Android Security Internals has earned a permanent spot on my office bookshelf.

Jon jcase Sawyer

CTO, Applied Cybersecurity LLC

Port Angeles, WA

Acknowledgments

I would like to thank everyone at No Starch Press who worked on this book. Special thanks to Bill Pollock for making my ramblings readable and to Alison Law for her patience in turning them into an actual book.

A big thanks to Kenny Root for reviewing all chapters and sharing the backstories behind some of Android’s security features.

Thanks to Jorrit Chainfire Jongma for maintaining SuperSU, which has been an invaluable tool for poking at Android’s internals, and for reviewing my coverage of it in Chapter 13.

Thanks to Jon jcase Sawyer for continuing to challenge our assumptions about Android security and for contributing a foreword to my book.

Introduction

In a relatively short period of time, Android has become the world’s most popular mobile platform. Although originally designed for smartphones, it now powers tablets, TVs, and wearable devices, and will soon even be found in cars. Android is being developed at a breathtaking pace, with an average of two major releases per year. Each new release brings a better UI, performance improvements, and a host of new user-facing features which are typically blogged about and dissected in excruciating detail by Android enthusiasts.

One aspect of the Android platform that has seen major improvements over the last few years, but which has received little public attention, is security. Over the years, Android has become more resistant to common exploit techniques (such as buffer overflows), its application isolation (sandboxing) has been reinforced, and its attack surface has been considerably reduced by aggressively decreasing the number of system processes that run as root. In addition to these exploit mitigations, recent versions of Android have introduced major new security features such as restricted user support, full-disk encryption, hardware-backed credential storage, and support for centralized device management and provisioning. Even more enterprise-oriented features and security improvements such as managed profile support, improved full-disk encryption, and support for biometric authentication have been announced for the next Android release (referred to as Android L as I write this).

As with any new platform feature, discussing cutting-edge security improvements is exciting, but it’s arguably more important to understand Android’s security architecture from the bottom up because each new security feature builds upon and integrates with the platform’s core security model. Android’s sandboxing model (in which each application runs as a separate Linux user and has a dedicated data directory) and permission system (which requires each application to explicitly declare the platform features it requires) are fairly well understood and documented. However, the internals of other fundamental platform features that have an impact on device security, such as package management and code signing, are largely treated as a black box beyond the security research community.

One of the reasons for Android’s popularity is the relative ease with which a device can be flashed with a custom build of Android, rooted by applying a third-party update package, or otherwise customized. Android enthusiast forums and blogs feature many practical How to guides that take users through the steps necessary to unlock a device and apply various customization packages, but they offer very little structured information about how such system updates operate under the hood and what risks they carry.

This books aims to fill these gaps by providing an exploration of how Android works by describing its security architecture from the bottom up and delving deep into the implementation of major Android subsystems and components that relate to device and data security. The coverage includes broad topics that affect all applications, such as package and user management, permissions and device policy, as well as more specific ones such as cryptographic providers, credential storage, and support for secure elements.

It’s not uncommon for entire Android subsystems to be replaced or rewritten between releases, but security-related development is conservative by nature, and while the described behavior might be changed or augmented across releases, Android’s core security architecture should remain fairly stable in future releases.

Who This Book Is For

This book should be useful to anyone interested in learning more about Android’s security architecture. Both security researchers looking to evaluate the security level of Android as a whole or of a specific subsystem and platform developers working on customizing and extending Android will find the high-level description of each security feature and the provided implementation details to be a useful starting point for understanding the underlying platform source code. Application developers can gain a deeper understanding of how the platform works, which will enable them to write more secure applications and take better advantage of the security-related APIs that the platform provides. While some parts of the book are accessible to a non-technical audience, the bulk of the discussion is closely tied to Android source code or system files, so familiarity with the core concepts of software development in a Unix environment is useful.

Prerequisites

The book assumes basic familiarity with Unix-style operating systems, preferably Linux, and does not explain common concepts such as processes, user groups, file permissions, and so on. Linux-specific or recently added OS features (such as capability and mount namespaces) are generally introduced briefly before discussing Android subsystems that use them. Most of the presented platform code comes from core Android daemons (usually implemented in C or C++) and system services (usually implemented in Java), so basic familiarity with at least one of these languages is also required. Some code examples feature sequences of Linux system calls, so familiarity with Linux system programming can be helpful in understanding the code, but is not absolutely required. Finally, while the basic structure and core components (such as activities and services) of Android apps are briefly described in the initial chapters, basic understanding of Android development is assumed.

Android Versions

The description of Android’s architecture and implementation in this book (except for several proprietary Google features) is based on source code publicly released as part of the Android Open Source Project (AOSP). Most of the discussion and code excerpts reference Android 4.4, which is the latest publicly available version released with source code at the time of this writing. The master branch of AOSP is also referenced a few times, because commits to master are generally a good indicator of the direction future Android releases will take. However, not all changes to the master branch are incorporated in public releases as is, so it’s quite possible that future releases will change and even remove some of the presented functionality.

A developer preview version of the next Android release (Android L, mentioned earlier) was announced shortly after the draft of this book was completed. However, as of this writing, the full source code of Android L is not available and its exact public release date is unknown. While the preview release does include some new security features, such as improvements to device encryption, managed profiles, and device management, none of these features are final and so are subject to change. That is why this book does not discuss any of these new features. Although we could introduce some of Android L’s security improvements based on their observed behavior, without the underlying source code, any discussion about their implementation would be incomplete and speculative.

How Is This Book Organized?

This book consists of 13 chapters that are designed to be read in sequence. Each chapter discusses a different aspect or feature of Android security, and subsequent chapters build on the concepts introduced by their predecessors. Even if you’re already familiar with Android’s architecture and security model and are looking for details about a specific topic, you should at least skim Chapter 1 through Chapter 3 because the topics they cover form the foundation for the rest of the book.

Chapter 1 gives a high-level overview of Android’s architecture and security model.

Chapter 2 describes how Android permissions are declared, used, and enforced by the system.

Chapter 3 discusses code signing and details how Android’s application installation and management process works.

Chapter 4 explores Android’s multi-user support and describes how data isolation is implemented on multi-user devices.

Chapter 5 gives an overview of the Java Cryptography Architecture (JCA) framework and describes Android’s JCA cryptographic providers.

Chapter 6 introduces the architecture of the Java Secure Socket Extension (JSSE) framework and delves into its Android implementation.

Chapter 7 explores Android’s credential store and introduces the APIs it provides to applications that need to store cryptographic keys securely.

Chapter 8 discusses Android’s online account management framework and shows how support for Google accounts is integrated into Android.

Chapter 9 presents Android’s device management framework, details how VPN support is implemented, and delves into Android’s support for the Extensible Authentication Protocol (EAP).

Chapter 10 introduces verified boot, disk encryption, and Android’s lockscreen implementation, and shows how secure USB debugging and encrypted device backups are implemented.

Chapter 11 gives an overview of Android’s NFC stack, delves into secure element (SE) integration and APIs, and introduces host-based card emulation (HCE).

Chapter 12 starts with a brief introduction to SELinux’s architecture and policy language, details the changes made to SELinux in order to integrate it in Android, and gives an overview of Android’s base SELinux policy.

Chapter 13 discusses how Android’s bootloader and recovery OS are used to perform full system updates, and details how root access can be obtained on both engineering and production Android builds.

Conventions

Because the main topic of this book is Android’s architecture and implementation, it contains multiple code excerpts and file listings, which are extensively referenced in the sections that follow each listing or code example. A few format conventions are used to set those references (which typically include multiple OS or programming language constructs) apart from the rest of the text.

Commands; function and variable names; XML attributes; and SQL object names are set in monospace (for example: the id command, the getCallingUid() method, the name attribute, and so on). The names of files and directories, Linux users and groups, processes, and other OS objects are set in italic (for example: "the packages.xml file, the system user, the vold daemon," and so on). String literals are also set in italic (for example: "the AndroidOpenSSL provider). If you use such string literals in a program, you typically need to enclose them in double or single quotes (for example: Signature.getInstance(SHA1withRSA, AndroidOpenSSL")).

Java class names are typically in their unqualified format without the package name (for example: the Binder class); fully qualified names are only used when multiple classes with the same name exist in the discussed API or package, or when specifying the containing package is otherwise important (for example: the javax.net.ssl.SSLSocketFactory class). When referenced in the text, function and method names are shown with parentheses, but their parameters are typically omitted for brevity (for example: the getInstance() factory method). See the relevant reference documentation for the full function or method signature.

Most chapters include diagrams that illustrate the architecture or structure of the discussed security subsystem or component. All diagrams follow an informal boxes and arrows style and do not conform strictly to a particular format. That said, most diagrams borrow ideas from UML class and deployment diagrams, and boxes typically represent classes or objects, while arrows represent dependency or communication paths.

Chapter 1. Android’s Security Model

This chapter will first briefly introduce Android’s architecture, inter-process communication (IPC) mechanism, and main components. We then describe Android’s security model and how it relates to the underlying Linux security infrastructure and code signing. We conclude with a brief overview of some newer additions to Android’s security model, namely multi-user support, mandatory access control (MAC) based on SELinux, and verified boot. Android’s architecture and security model are built on top of the traditional Unix process, user, and file paradigm, but this paradigm is not described from scratch here. We assume a basic familiarity with Unix-like systems, particularly Linux.

Android’s Architecture

Let’s briefly examine Android’s architecture from the bottom up. Figure 1-1 shows a simplified representation of the Android stack.

Figure 1-1. The Android architecture

Linux Kernel

As you can see in Figure 1-1, Android is built on top of the Linux kernel. As in any Unix system, the kernel provides drivers for hardware, networking, file-system access, and process management. Thanks to the Android Mainlining Project,[1] you can now run Android with a recent vanilla kernel (with some effort), but an Android kernel is slightly different from a regular Linux kernel that you might find on a desktop machine or a non-Android embedded device. The differences are due to a set of new features (sometimes called Androidisms[2]) that were originally added to support Android. Some of the main Androidisms are the low memory killer, wakelocks (integrated as part of wakeup sources support in the mainline Linux kernel), anonymous shared memory (ashmem), alarms, paranoid networking, and Binder.

The most important Androidisms for our discussion are Binder and paranoid networking. Binder implements IPC and an associated security mechanism, which we discuss in more detail in Binder. Paranoid networking restricts access to network sockets to applications that hold specific permissions. We delve deeper into this topic in Chapter 2.

Native Userspace

On top of the kernel is the native userspace layer, consisting of the init binary (the first process started, which starts all other processes), several native daemons, and a few hundred native libraries that are used throughout the system. While the presence of an init binary and daemons is reminiscent of a traditional Linux system, note that both init and the associated startup scripts have been developed from scratch and are quite different from their mainline Linux counterparts.

Dalvik VM

The bulk of Android is implemented in Java and as such is executed by a Java Virtual Machine (JVM). Android’s current Java VM implementation is called Dalvik and it is the next layer in our stack. Dalvik was designed with mobile devices in mind and cannot run Java bytecode (.class files) directly: its native input format is called Dalvik Executable (DEX) and is packaged in .dex files. In turn, .dex files are packaged either inside system Java libraries (JAR files), or inside Android applications (APK files, discussed in Chapter 3).

Dalvik and Oracle’s JVM have different architectures—register-based in Dalvik versus stack-based in the JVM—and different instruction sets. Let’s look at a simple example to illustrate the differences between the two VMs (see Example 1-1).

Example 1-1. Static Java method that adds two integers

public static int add(int i, int j) {

    return i + j;

}

When compiled for each VM, the add() static method, which simply adds two integers and returns the result, would generate the bytecode shown in Example 1-2.

Example 1-2. JVM and Dalvik bytecode

JVM Bytecode

public static int add(int, int);

  Code:

    0: iload_0➊

    1: iload_1➋

    2: iadd➌

    3: ireturn➍

Dalvik Bytecode

.method public static add(II)I

    add-int v0, p0, p1➎

    return v0➏

.end method

Here, the JVM uses two instructions to load the parameters onto the stack (➊ and ➋), then executes the addition ➌, and finally returns the result ➍. In contrast, Dalvik uses a single instruction to add parameters (in registers p0 and p1) and puts the result in the v0 register ➎. Finally, it returns the contents of the v0 register ➏. As you can see, Dalvik uses fewer instructions to achieve the same result. Generally speaking, register-based VMs use fewer instructions, but the resulting code is larger than the corresponding code in a stack-based VM. However, on most architectures, loading code is less expensive than instruction dispatch, so register-based VMs can be interpreted more efficiently.[3]

In most production devices, system libraries and preinstalled applications do not contain device-independent DEX code directly. As a performance optimization, DEX code is converted to a device-dependent format and stored in an Optimized DEX (.odex) file, which typically resides in the same directory as its parent JAR or APK file. A similar optimization process is performed for user-installed applications at install time.

Java Runtime Libraries

A Java language implementation requires a set of runtime libraries, defined mostly in the java.* and javax.* packages. Android’s core Java libraries are originally derived from the Apache Harmony project[4] and are the next layer on our stack. As Android has evolved, the original Harmony code has changed significantly. In the process, some features have been replaced entirely (such as internationalization support, the cryptographic provider, and some related classes), while others have been extended and improved. The core libraries are developed mostly in Java, but they have some native code dependencies as well. Native code is linked into Android’s Java libraries using the standard Java Native Interface (JNI),[5] which allows Java code to call native code and vice versa. The Java runtime libraries layer is directly accessed both from system services and applications.

System Services

The layers introduced up until now make up the plumbing necessary to implement the core of Android —system services. System services (79 as of version 4.4) implement most of the fundamental Android features, including display and touch screen support, telephony, and network connectivity. Most system services are implemented in Java; some fundamental ones are written in native code.

With a few exceptions, each system service defines a remote interface that can be called from other services and applications. Coupled with the service discovery, mediation, and IPC provided by Binder, system services effectively implement an object-oriented OS on top of Linux.

Let’s look at how Binder enables IPC on Android in detail, as this is one of the cornerstones of Android’s security model.

Inter-Process Communication

As mentioned previously, Binder is an inter-process communication (IPC) mechanism. Before getting into detail about how Binder works, let’s briefly review IPC.

As in any Unix-like system, processes in Android have separate address spaces and a process cannot directly access another process’s memory (this is called process isolation). This is usually a good thing, both for stability and security reasons: multiple processes modifying the same memory can be catastrophic, and you don’t want a potentially rogue process that was started by another user to dump your email by accessing your mail client’s memory. However, if a process wants to offer some useful service(s) to other processes, it needs to provide some mechanism that allows other processes to discover and interact with those services. That mechanism is referred to as IPC.

The need for a standard IPC mechanism is not new, so several options predate Android. These include files, signals, sockets, pipes, semaphores, shared memory, message queues, and so on. While Android uses some of these (such as local sockets), it does not support others (namely System V IPCs like semaphores, shared memory segments, and message queues).

Binder

Because the standard IPC mechanisms weren’t flexible or reliable enough, a new IPC mechanism called Binder was developed for Android. While Android’s Binder is a new implementation, it’s based on the architecture and ideas of OpenBinder.[6]

Binder implements a distributed component architecture based on abstract interfaces. It is similar to Windows Common Object Model (COM) and Common Object Broker Request Architectures (CORBA) on Unix, but unlike those frameworks, it runs on a single device and does not support remote procedure calls (RPC) across the network (although RPC support could be implemented on top of Binder). A full description of the Binder framework is outside the scope of this book, but we introduce its main components briefly in the following sections.

Binder Implementation

As mentioned earlier, on a Unix-like system, a process cannot access another process’s memory. However, the kernel has control over all processes and therefore can expose an interface that enables IPC. In Binder, this interface is the /dev/binder device, which is implemented by the Binder kernel driver. The Binder driver is the central object of the framework, and all IPC calls go through it. Inter-process communication is implemented with a single ioctl() call that both sends and receives data through the binder_write_read structure, which consists of a write_buffer containing commands for the driver, and a read_buffer containing commands that the userspace needs to perform.

But how is data actually passed between processes? The Binder driver manages part of the address space of each process. The Binder driver-managed chunk of memory is read-only to the process, and all writing is performed by the kernel module. When a process sends a message to another process, the kernel allocates some space in the destination process’s memory, and copies the message data directly from the sending process. It then queues a short message to the receiving process telling it where the received message is. The recipient can then access that message directly (because it is in its own memory space). When a process is finished with the message, it notifies the Binder driver to mark the memory as free. Figure 1-2 shows a simplified illustration of the Binder IPC architecture.

Figure 1-2. Binder IPC

Higher-level IPC abstractions in Android such as Intents (commands with associated data that are delivered to components across processes), Messengers (objects that enable message-based communication across processes), and ContentProviders (components that expose a cross-process data management interface) are built on top of Binder. Additionally, service interfaces that need to be exposed to other processes can be defined using the Android Interface Definition Language (AIDL), which enables clients to call remote services as if they were local Java objects. The associated aidl tool automatically generates stubs (client-side representations of the remote object) and proxies that map interface methods to the lower-level transact() Binder method and take care of converting parameters to a format that Binder can transmit (this is called parameter marshalling/unmarshalling). Because Binder is inherently typeless, AIDL-generated stubs and proxies also provide type safety by including the target interface name in each Binder transaction (in the proxy) and validating it in the stub.

Binder Security

On a higher level, each object that can be accessed through the Binder framework implements the IBinder interface and is called a Binder object. Calls to a Binder object are performed inside a Binder transaction, which contains a reference to the target object, the ID of the method to execute, and a data buffer. The Binder driver automatically adds the process ID (PID) and effective user ID (EUID) of the calling process to the transaction data. The called process (callee) can inspect the PID and EUID and decide whether it should execute the requested method based on its internal logic or system-wide metadata about the calling application.

Since the PID and EUID are filled in by the kernel, caller processes cannot fake their identity to get more privileges than allowed by the system (that is, Binder prevents privilege escalation). This is one of the central pieces of Android’s security model, and all higher-level abstractions, such as permissions, build upon it. The EUID and PID of the caller are accessible via the getCallingPid() and getCallingUid() methods of the android.os.Binder class, which is part of Android’s public API.

Note

The calling process’s EUID may not map to a single application if more than one application is executing under the same UID (see Chapter 2 for details). However, this does not affect security decisions, as processes running under the same UID are typically granted the same set of permissions and privileges (unless process-specific SELinux rules have been defined).

Binder Identity

One of the most important properties of Binder objects is that they maintain a unique identity across processes. Thus if process A creates a Binder object and passes it to process B, which in turn passes it to process C, calls from all three processes will be processed by the same Binder object. In practice, process A will reference the Binder object directly by its memory address (because it is in process A’s memory space), while process B and C will receive only a handle to the Binder object.

The kernel maintains the mapping between live Binder objects and their handles in other processes. Because a Binder object’s identity is unique and maintained by the kernel, it is impossible for userspace processes to create a copy of a Binder object or obtain a reference to one unless they have been handed one through IPC. Thus a Binder object is a unique, unforgeable, and communicable object that can act as a security token. This enables the use of capability-based security in Android.

Capability-Based Security

In a capability-based security model, programs are granted access to a particular resource by giving them an unforgeable capability that both references the target object and encapsulates a set of access rights to it. Because capabilities are unforgeable, the mere fact that a program possesses a capability is sufficient to give it access to the target resource; there is no need to maintain access control lists (ACLs) or similar structures associated with actual resources.

Binder Tokens

In Android, Binder objects can act as capabilities and are called Binder tokens when used in this fashion. A Binder token can be both a capability and a target resource. The possession of a Binder token grants the owning process full access to a Binder object, enabling it to perform Binder transactions on the target object. If the Binder object implements multiple actions (by selecting the action to perform based on the code parameter of the Binder transaction), the caller can perform any action when it has a reference to that Binder object. If more granular access control is required, the implementation of each action needs to implement the necessary permission checks, typically by utilizing the PID and EUID of the caller process.

A common pattern in Android is to allow all actions to callers running as system (UID 1000) or root (UID 0), but perform additional permission checks for all other processes. Thus access to important Binder objects such as system services is controlled in two ways: by limiting who can get a reference to that Binder object and by checking the caller identity before performing an action on the Binder object. (This check is optional and implemented by the Binder object itself, if required.)

Alternatively, a Binder object can be used only as a capability without implementing any other functionality. In this usage pattern, the same Binder object is held by two (or more) cooperating processes, and the one acting as a server (processing some kind of client requests) uses the Binder token to authenticate its clients, much like web servers use session cookies.

This usage pattern is used internally by the Android framework and is mostly invisible to applications. One notable use case of Binder tokens that is visible in the public API is window tokens. The top-level window of each activity is associated with a Binder token (called a window token), which Android’s window manager (the system service responsible for managing application windows) keeps track of. Applications can obtain their own window token but cannot get access to the window tokens of other applications. Typically you don’t want other applications adding or removing windows on top of your own; each request to do so must provide the window token associated with the application, thus guaranteeing that window requests are coming from your own application or from the system.

Accessing Binder Objects

Although Android controls access to Binder objects for security purposes, and the only way to communicate with a Binder object is to be given a reference to it, some Binder objects (most notably system services) need to be universally accessible. It is, however, impractical to hand out references to all system services to each and every process, so we need some mechanism that allows processes to discover and obtain references to system services as needed.

In order to enable service discovery, the Binder framework has a single context manager, which maintains references to Binder objects. Android’s context manager implementation is the servicemanager native daemon. It is started very early in the boot process so that system services can register with it as they start up. Services are registered by passing a service name and a Binder reference to the service manager. Once a service is registered, any client can obtain its Binder reference by using its name. However, most system services implement additional permission checks, so obtaining a reference does not automatically guarantee access to all of its functionality. Because anyone can access a Binder reference when it is registered with the service manager, only a small set of whitelisted system processes can register system services. For example, only a process executing as UID 1002 (AID_BLUETOOTH) can register the bluetooth system service.

You can view a list of registered services by using the service list command, which returns the name of each registered service and the implemented IBinder interface. Sample output from running the command on an Android 4.4 device is shown in Example 1-3.

Example 1-3. Obtaining a list of registered system services with the service list command

$ service list

service list

Found 79 services:

0      sip: [android.net.sip.ISipService]

1      phone: [com.android.internal.telephony.ITelephony]

2      iphonesubinfo: [com.android.internal.telephony.IPhoneSubInfo]

3      simphonebook: [com.android.internal.telephony.IIccPhoneBook]

4      isms: [com.android.internal.telephony.ISms]

5      nfc: [android.nfc.INfcAdapter]

6      media_router: [android.media.IMediaRouterService]

7      print: [android.print.IPrintManager]

8      assetatlas: [android.view.IAssetAtlas]

9      dreams: [android.service.dreams.IdreamManager]

--

snip--

Other Binder Features

While not directly related to Android’s security model, two other notable Binder features are reference counting and death notification (also known as link to death). Reference counting guarantees that Binder objects are automatically freed when no one references them and is implemented in the kernel driver with the BC_INCREFS, BC_ACQUIRE, BC_RELEASE, and BC_DECREFS commands. Reference counting is integrated at various levels of the Android framework but is not directly visible to applications.

Death notification allows applications that use Binder objects that are hosted by other processes to be notified when those processes are killed by the kernel and to perform any necessary cleanup. Death notification is implemented with the BC_REQUEST_DEATH_NOTIFICATION and BC_CLEAR_DEATH_NOTIFICATION commands in the kernel driver and the linkToDeath() and unlinkToDeath() methods of the IBinder interface[7] in the framework. (Death notifications for local binders are not sent, because local binders cannot die without the hosting process dying as well.)

Android Framework Libraries

Next on the stack are the Android framework libraries, sometimes called just the framework. The framework includes all Java libraries that are not part of the standard Java runtime (java.*, javax.*, and so on) and is for the most part hosted under the android top-level package. The framework includes the basic blocks for building Android applications, such as the base classes for activities, services, and content providers (in the android.app.* packages); GUI widgets (in the android.view.* and android.widget packages); and classes for file and database access (mostly in the android.database.* and android.content.* packages). It also includes classes that let you interact with device hardware, as well as classes that take advantage of higher-level services offered by the system.

Even though almost all Android OS functionality above the kernel level is implemented as system services, it is not exposed directly in the framework but is accessed via facade classes called managers. Typically, each manager is backed by a corresponding system service; for example, the BluetoothManager is a facade for the BluetoothManagerService.

Applications

On the highest level of the stack are applications (or apps), which are the programs that users directly interact with. While all apps have the same structure and are built on top of the Android framework, we distinguish between system apps and user-installed apps.

System Apps

System apps are included in the OS image, which is read-only on production devices (typically mounted as /system), and cannot be uninstalled or changed by users. Therefore, these apps are considered secure and are given many more privileges than user-installed apps. System apps can be part of the core Android OS or can simply be preinstalled user applications, such as email clients or browsers. While all apps installed under /system were treated equally in earlier versions of Android (except by OS features that check the app signing certificate), Android 4.4 and higher treat apps installed in /system/priv-app/ as privileged applications and will only grant permissions with protection level signatureOrSystem to privileged apps, not to all apps installed under /system. Apps that are signed with the platform signing key can be granted system permissions with the signature protection level, and thus can get OS-level privileges even if they are not preinstalled under /system. (See Chapter 2 for details on permissions and code signing.)

While system apps cannot be uninstalled or changed, they can be updated by users as long as the updates are signed with the same private key, and some can be overridden by user-installed apps. For example, a user can choose to replace the preinstalled application launcher or input method with a third-party application.

User-Installed Apps

User-installed apps are installed on a dedicated read-write partition (typically mounted as /data) that hosts user data and can be uninstalled at will. Each application lives in a dedicated security sandbox and typically cannot affect other applications or access their data. Additionally, apps can only access resources that they have explicitly been granted a permission to use. Privilege separation and the principle of least privilege are central to Android’s security model, and we will explore how they are implemented in the next section.

Android App Components

Android applications are a combination of loosely coupled components and, unlike traditional applications, can have more than one entry point. Each component can offer multiple entry points that can be reached based on user actions in the same or another application, or triggered by a system event that the application has registered to be notified about.

Components and their entry points, as well as additional metadata, are defined in the application’s manifest file, called AndroidManifest.xml. Like most Android resource files, this file is compiled into a binary XML format (similar to ASN.1) before bundling it in the application package (APK) file in order to decrease size and speed up parsing. The most important application property defined in the manifest file is the application package name, which uniquely identifies each application in the system. The package name is in the same format as Java package names (reverse domain name notation; for example, com.google.email).

The AndroidManifest.xml file is parsed at application install time, and the package and components it defines are registered with the system. Android requires each application to be signed using a key controlled by its developer. This guarantees that an installed application cannot be replaced by another application that claims to have the same package name (unless it is signed with the same key, in which case the existing application is updated). We’ll discuss code signing and application packages in Chapter 3.

The main components of Android apps are listed below.

Activities

An activity is a single screen with a user interface. Activities are the main building blocks of Android GUI applications. An application can have multiple activities and while they are usually designed to be displayed in a particular order, each activity can be started independently, potentially by a different app (if allowed).

Services

A service is a component that runs in the background and has no user interface. Services are typically used to perform some long-running operation, such as downloading a file or playing music, without blocking the user interface. Services can also define a remote interface using AIDL and provide some functionality to other apps. However, unlike system services, which are part of the OS and are always running, application services are started and stopped on demand.

Content providers

Content providers provide an interface to app data, which is typically stored in a database or files. Content providers can be accessed via IPC and are mainly used to share an app’s data with other apps. Content providers offer fine-grained control over what parts of data are accessible, allowing an application to share only a subset of its data.

Broadcast receivers

A broadcast receiver is a component that responds to systemwide events, called broadcasts. Broadcasts can originate from the system (for example, announcing changes in network connectivity), or from a user application (for example, announcing that background data update has completed).

Android’s Security Model

Like the rest of the system, Android’s security model also takes advantage of the security features offered by the Linux kernel. Linux is a multiuser operating system and the kernel can isolate user resources from one another, just as it isolates processes. In a Linux system, one user cannot access another user’s files (unless explicitly granted permission) and each process runs with the identity (user and group ID, usually referred to as UID and GID) of the user that started it, unless the set-user-ID or set-group-ID (SUID and SGID) bits are set on the corresponding executable file.

Android takes advantage of this user isolation, but treats users differently than a traditional Linux system (desktop or server) does. In a traditional system, a UID is given either to a physical user that can log into the system and execute commands via the shell, or to a system service (daemon) that executes in the background (because system daemons are often accessible over the network, running each daemon with a dedicated UID can limit the damage if one is compromised). Android was originally designed for smartphones, and because mobile phones are personal devices, there was no need to register different physical users with the system. The physical user is implicit, and UIDs are used to distinguish applications instead. This forms the basis of Android’s application sandboxing.

Application Sandboxing

Android automatically assigns a unique UID, often called an app ID, to each application at installation and executes that application in a dedicated process running as that UID. Additionally, each application is given a dedicated data directory which only it has permission to read and write to. Thus, applications are isolated, or sandboxed, both at the process level (by having each run in a dedicated process) and at the file level (by having a private data directory). This creates a kernel-level application sandbox, which applies to all applications, regardless of whether they are executed in a native or virtual machine process.

System daemons and applications run under well-defined and constant UIDs, and very few daemons run as the root user (UID 0). Android does not have the traditional /etc/password file and its system UIDs are statically defined in the android_filesystem_config.h header file. UIDs for system services start from 1000, with 1000 being the system (AID_SYSTEM) user, which has special (but still limited) privileges. Automatically generated UIDs for applications start at 10000 (AID_APP), and the corresponding usernames are in the form app_XXX or uY_aXXX (on Android versions that support multiple physical users), where XXX is the offset from AID_APP and Y is the Android user ID (not the same as UID). For example, the 10037 UID corresponds to the u0_a37 username and may be assigned to the Google email client application (com.google.android.email package). Example 1-4 shows that the email application process executes as the u0_a37 user ➊, while other application processes execute as different users.

Example 1-4. Each application process executes as a dedicated user on Android

$ ps

--

snip

--

u0_a37    16973 182  941052  60800 ffffffff 400d073c S com.google.android.email➊

u0_a8    18788 182  925864  50236 ffffffff 400d073c S com.google.android.dialer

u0_a29    23128 182  875972  35120 ffffffff 400d073c S com.google.android.calendar

u0_a34    23264 182  868424  31980 ffffffff 400d073c S com.google.android.deskclock

--

snip--

The data directory of the email application is named after its package name and is created under /data/data/ on single-user devices. (Multi-user devices use a different naming scheme as discussed in Chapter 4.) All files inside the data directory are owned by the dedicated Linux user, u0_a37, as shown in Example 1-5 (with timestamps omitted). Applications can optionally create files using the MODE_WORLD_READABLE and MODE_WORLD_WRITEABLE flags to allow direct access to files by other applications, which effectively sets the S_IROTH and S_IWOTH access bits on the file, respectively. However, the direct sharing of files is discouraged, and those flags are deprecated in Android versions 4.2 and higher.

Example 1-5. Application directories are owned by the dedicated Linux user

# ls -l /data/data/com.google.android.email

drwxrwx--x u0_a37  u0_a37            app_webview

drwxrwx--x u0_a37  u0_a37            cache

drwxrwx--x u0_a37  u0_a37            databases

drwxrwx--x u0_a37  u0_a37            files

--

snip--

Application UIDs are managed alongside other package metadata in the /data/system/packages.xml file (the canonical source) and also written to the /data/system/packages.list file. (We discuss package management and the packages.xml file in Chapter 3.) Example 1-6 shows the UID assigned to the com.google.android.email package as it appears in packages.list.

Example 1-6. The UID corresponding to each application is stored in /data/system/packages.list

# grep 'com.google.android.email' /data/system/packages.list

com.google.android.email 10037 0 /data/data/com.google.android.email default 3003,1028,1015

Here, the first field is the package name, the second is the UID assigned to the application, the third is the debuggable flag (1 if debuggable), the fourth is the application’s data directory path, and the fifth is the seinfo label (used by SELinux). The last field is a list of the supplementary GIDs that the app launches with. Each GID is typically associated with an Android permission (discussed next) and the GID list is generated based on the permissions granted to the application.

Applications can be installed using the same UID, called a shared user ID, in which case they can share files and even run in the same process. Shared user IDs are used extensively by system applications, which often need to use the same resources across different packages for modularity. For example, in Android 4.4 the system