Systematic setup of compartmentalised systems on the seL4

Systematic setup of compartmentalised systems on the seL4
Diplomarbeit
Systematic setup of compartmentalised
systems on the seL4 microkernel
Lukas Hänel
February 26, 2010
Technische Universität Dresden
Fakultät Informatik
Institut für Systemarchitektur
Professur Betriebssysteme
Betreuender Hochschullehrer: Prof. Dr. rer. nat. Hermann Härtig
Betreuender Mitarbeiter:
Dr. Kevin Elphinstone
Erklärung
Hiermit erkläre ich, dass ich diese Arbeit selbstständig erstellt und keine anderen als die
angegebenen Hilfsmittel benutzt habe.
Dresden, den February 26, 2010
Lukas Hänel
Acknowledgement
I would like to thank everybody that made my overseas trip such a rich experience. My
special thanks goes to Hermann Härtig for affording me the opportunity to study abroad
and for his inspiring lectures that first awakened my interest in operating systems. I
want to thank Gernot Heiser for the ERTOS group at NICTA. My thanks goes to Kevin
Elphinstone and Peter Chubb for welcoming me in Australia and guiding my work. I
would like to thank the many people who worked with me in the project, namely Ben
Kalman for developing the build tools, Michael von Tessin for support on the seL4
microkernel, and Ihor Kuz for guidance on component architectures and proof-reading
my thesis. Thanks to Robert Sison for asking for more features and Adam Walker for
maturing my code. I would like to thank the people responsible for my decision to go
overseas, namely Christian Helmuth for his supervision of my study project, the guys
at ST Microelectronics for the opportunity of my internship, and my dear parents for
supporting me in many ways.
Contents
1 Introduction
1.1 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Background and Related Work
2.1 Information flow architectures . . . . . . . . . . . . . . . .
2.1.1 Definition . . . . . . . . . . . . . . . . . . . . . . .
2.1.2 Problems . . . . . . . . . . . . . . . . . . . . . . .
2.2 Security Architectures . . . . . . . . . . . . . . . . . . . .
2.3 Component Architecture for microkernel-based Embedded
2.4 Hardware compartmentalisation support . . . . . . . . . .
2.4.1 CPU compartmentalisation . . . . . . . . . . . . .
2.4.2 I/O Channels . . . . . . . . . . . . . . . . . . . . .
2.4.3 IOMMU . . . . . . . . . . . . . . . . . . . . . . . .
2.4.4 PCI device discovery . . . . . . . . . . . . . . . . .
2.5 Virtualisation . . . . . . . . . . . . . . . . . . . . . . . . .
2.5.1 Virtualisation of PCI device discovery . . . . . . .
2.6 Related Work . . . . . . . . . . . . . . . . . . . . . . . . .
1
2
.
.
.
.
.
.
.
.
.
.
.
.
.
5
5
5
6
8
9
12
12
13
14
14
17
18
18
.
.
.
.
.
.
.
.
.
.
.
.
.
.
21
21
21
23
23
25
26
26
28
28
30
30
30
31
31
4 Design
4.1 Handling I/O devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.1 PCI device discovery by the system . . . . . . . . . . . . . . . .
4.1.2 PCI device discovery by compartments . . . . . . . . . . . . . . .
33
33
34
34
3 seL4
3.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 What it is . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3 How to . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.1 Capabilities and CSpaces . . . . . . . . . . . . .
3.3.2 Kernel memory management - Untyped_Retype
3.3.3 VSpaces . . . . . . . . . . . . . . . . . . . . . . .
3.3.4 Threads, Endpoints and IRQs . . . . . . . . . . .
3.3.5 Start up . . . . . . . . . . . . . . . . . . . . . . .
3.3.6 x86 and IO protection . . . . . . . . . . . . . . .
3.3.7 Devices . . . . . . . . . . . . . . . . . . . . . . .
3.4 Current implementation status . . . . . . . . . . . . . .
3.4.1 Iwana - the kernel object allocator . . . . . . . .
3.4.2 Paravirtualisation - seL4::Wombat . . . . . . . .
3.5 Application to this thesis . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. . . . .
. . . . .
. . . . .
. . . . .
Systems
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
IX
Contents
.
.
.
.
.
.
35
36
37
38
41
46
.
.
.
.
.
.
.
.
.
49
49
49
50
51
53
53
54
55
55
6 Evaluation and Future Work
6.1 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
59
62
63
7 Conclusions
65
Glossary
67
Bibliography
69
4.2
4.1.3 PCI configuration . . . . .
Architecture specification . . . . .
4.2.1 Device specification . . . .
4.2.2 Compartment specification
4.2.3 Connection specification . .
4.2.4 Compartment bootinfo . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5 Implementation
5.1 System initialisation . . . . . . . . . . . .
5.1.1 The loading process . . . . . . . .
5.1.2 Loading compartments . . . . . . .
5.1.3 Mapping devices to compartments
5.1.4 Establishing connections . . . . . .
5.1.5 Starting compartments . . . . . . .
5.2 Device access in seL4::Wombat . . . . . .
5.3 Virtual PCI configuration space . . . . . .
5.4 Timer server . . . . . . . . . . . . . . . .
X
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
List of Figures
2.1
2.2
2.3
2.4
2.5
2.6
2.7
Example information flow architecture . . . . . . . .
MILS layer . . . . . . . . . . . . . . . . . . . . . . .
Computer architecture . . . . . . . . . . . . . . . . .
PCI bus structure . . . . . . . . . . . . . . . . . . .
PCI configuration address format . . . . . . . . . . .
PCI configuration header . . . . . . . . . . . . . . .
Capturing memory accesses using a memory-mapped
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
I/O region
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
6
8
13
15
15
16
17
3.1
3.2
3.3
3.4
3.5
3.6
Capability references . . . . . . .
Guarded page table example . .
Capability moving . . . . . . . .
Retyping untyped memory . . . .
Example with seL4 address space
seL4 initial CSpace on QEMU . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
23
24
25
26
27
29
4.1
4.2
4.3
4.4
4.5
Offlining PCI device discovery . . . . . . . .
PCI virtualisation . . . . . . . . . . . . . .
Connection options . . . . . . . . . . . . . .
Conflict of select and compartmentalisation
Example of endpoint connection model . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
34
35
43
44
45
5.1
Information flow of the timer server . . . . . . . . . . . . . . . . . . . .
56
6.1
6.2
Architecture of secure firewall . . . . . . . . . . . . . . . . . . . . . . . .
Architecture of SOSP demo . . . . . . . . . . . . . . . . . . . . . . . . .
61
61
. . . . .
. . . . .
. . . . .
. . . . .
objects
. . . . .
XI
Listings
2.1
4.1
4.2
4.3
4.4
4.5
4.6
4.7
4.8
4.9
4.10
5.1
Example of a CAmkES architecture specification
PCI device identification data structure . . . . .
I/O region data structure . . . . . . . . . . . . .
Device description data structure . . . . . . . . .
Example minimal compartment specification . .
Example device server specification . . . . . . . .
Example virtual machine specification . . . . . .
Compartment specification data structure . . . .
Connection specification data structure . . . . .
Outgoing endpoint reporting data structure . . .
Compartment bootinfo data structure . . . . . .
Specification of timer compartment . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
11
38
38
39
40
40
41
42
43
45
47
57
XIII
Chapter 1
Introduction
Developing high assurance systems is a hard problem. It requires intimate knowledge of
system hardware and careful design and verification of the controlling software. When
following a security architecture like MILS [AFOTH06] a microkernel controls the computer’s central processing unit. The microkernel abstracts hardware protection mechanisms and provides compartments that are perfectly isolated except for explicitly-allowed
compartment communications. The microkernel does not control buses and I/O devices
of the system. They are controlled by bus managers and user level device drivers. The
system architecture grants those compartments access to their hardware and connects
them to client applications that require services of these devices. This thesis is about
instantiation of such architectures on modern x86 computers using the seL4 microkernel [EDE08].
Several developments in computer technology call for more complexity in trustworthy
systems. First, ever capable mobile devices create the desire to have a do-it-all device
that handles entertainment, communication, personal data and financial transactions.
Banks will only accept such devices when they are convinced that they are secure. Second, increasing performance capabilities of servers make it intolerable to have only one
customer per machine. While untrusted server virtualisation is not a problem for many
industries, military and governments demand solutions that provide isolation equivalent to physical separation. By default, embedded systems in aircrafts and automobiles
require high assurance of their safety. With the advent of smaller sizes and lower costs,
these systems control more and more parts of their environments and hence increase in
complexity.
While microkernels improve the system architecture, they do not necessarily make a
system secure. The next operating system issue that a secure system has to address is
device and resource management. Common solutions to these tasks fail to be secure. I
therefore propose an approach that leverages security architectures to do secure resource
and device management.
Common operating systems are designed with extensibility and universality in mind.
Device and resource management is geared to support new devices and components
with unknown resource requirements. In turn, the kernel does not preallocate resources
but follows a demand-allocation scheme where modules and applications continuously
request resources. The kernel serves these requests with a best effort to achieve a fair
resource distribution. In reality, however, faulty or malicious applications can request
all resources of the OS and thereby deny service to other applications. Additionally,
1
Chapter 1 Introduction
availability of resources can be used as a covert channel. Two applications that can each
exhaust and determine exhaustion of a resource can communicate.
In contrast, applications in seL4 cannot exhaust kernel resources. Isolated compartments only access seL4 to manage their kernel objects or to communicate with other
compartments. Standard resource management in microkernel systems is traditionally
performed by user level resource managers [GJP+ 00, APJ+ 01]. However, these resource
managers also implemented a best-effort strategy and were therefore also prone to provide covert channels. More generally, global services that connect untrusted applications
can compromise application isolation. Because of this threat, the developer must either
analyse them rigorously or avoid them entirely.
A high assurance system cannot give control to untrusted applications. Hence this thesis proposes an architecture where applications do not request resources at runtime. By
not allowing applications to control availability of resources, denial-of-service attacks and
resource-exhaustion channels are prevented. To replace on-demand requests, resource
allocations are determined at development time. An architecture specification is used
to describe the resource requirements of components and the assignment of devices to
components.
1.1 Outline
The remaining part of this thesis is structured as follows:
Chapter 2 Chapter 2 introduces the background and basis for this thesis. I will define
information flow architectures and compartments and show their relation to computer security. I will then survey software architectures for secure and embedded
systems. Those are based on hardware protection mechanisms that I will explain
afterwards. Specialisation in computer hardware emerged the requirement for
an I/O device discovery mechanism that plays an important role in computer
compartmentalisation as well. I will therefore explain it. Then I will give an
introduction to virtualisation as background for the compartmentalisation of PCI
device discovery. Also I will review related work.
Chapter 3 Following this, I will introduce the seL4 project because the component
developed in this thesis forms a basic component of every new seL4 system. I will
summarise the goals and security guarantees, explain the microkernel API, and
show current applications.
Chapter 4 In chapter 4, I will design a secure approach to device discovery for the
system and for compartments. Then I will design the models for I/O devices,
compartments, and connections and show the seL4 objects they consist of.
Chapter 5 In this chapter, I will describe my implementation and substantiate implementation decisions. I will handle the loading process, device access in Linux,
device discovery mechanisms and compartmentalisation of the timer.
Chapter 6 In chapter 6, I will evaluate the achievement of my goals, show limitations
of the approach and outline possibilities for future work.
2
1.1 Outline
Chapter 7 Finally, in chapter 7, I will draw conclusions about my work.
3
Chapter 2
Background and Related Work
This chapter will give the necessary background for this thesis. I will define the term
information flow architecture that is the basis for my work. I then survey existing
system security research and highlight its relation to information flow architectures. I
introduce hardware protection mechanisms of modern computers because I employ them
to implement an information flow architecture. I will then explain PCI device discovery
and the issues and opportunities of using it. The chapter continues with a survey of
virtualisation solutions and their approach to virtualisation of device discovery. I will
conclude with related work.
2.1 Information flow architectures
2.1.1 Definition
Access control policies [San93] define how information can flow between subjects and
objects. Subjects and objects get assigned security classifications, or labels and the
policy defines how information can or may not flow from one label to another. For
example, information may not flow from high to low in the Bell-La Padula model.
However, special secure applications may be authorised to allow such an information
flow when they ensure the secrecy goal by other means. For example, a declassifier
could remove the high label from outdated information, an encryption program could
maintain secrecy by outputting scrambled data only, and a filtering component could
remove classified bits from documents.
In a simple example (see Figure 2.1), a filtering system would accept data with secret
classification on one I/O device, remove compromising parts in a filter component and
then transmit the data as unclassified on another I/O device.
Subjects and objects are implemented by a secure operating system. This system is
secure, when it always enforces the restrictions set by the access control policy. That is,
the OS has to make sure that information can flow from one subject to another, only if
this flow is allowed in the policy. To do so, the OS splits the computer system into partitions or compartments that are isolated from each other. Between these compartments,
the OS may enable monitored or unmonitored communication. A compartment hosts a
subject and the OS ensures that two compartments are only connected if information
can flow between respective subjects.
A compartment as part of a computer system consists of several parts. Because a compartment is an active software component it consists of one or more threads of activity
5
Chapter 2 Background and Related Work
Compartments in a computer system
Computer system
low
I/O
device
high
Filter
compartment
I/O
device
Figure 2.1: Example information flow architecture
running on one or more processors. Consequently, a compartment also comprises a partition of the computer’s memory to store the program code and data. A compartment
can also contain input–output (I/O) channels and hence include I/O devices. If a compartment is connected to another compartment or to the OS, one side of the connection
is also part of the compartment.
Compartments and inter-compartment connections form a graph. This graph specifies how information can flow in a computer system. I call this graph an information
flow architecture to highlight that it represents both a software architecture and an
information-flow restricted system.
2.1.2 Problems
Systems implementing access control policies face two main problems: Privilege escalation attacks and the presence of covert channels. I will briefly outline the two.
Privilege escalation
In a privilege escalation attack a program breaks out of its compartment. It can then
control a bigger part of the computer and eventually access classified data.
Privilege escalation attacks require that a malicious program can communicate with
programs with higher privileges or with the OS. During this communication, the malicious program tries to drive the attacked component into a state the developer did not
foresee. In a successful attack, the attacker then communicates prepared data that
overwrites the program in the other compartment with his own program.
Common operating systems and applications suffer from these attacks because their
specification does not disallow this behaviour or because they do not implement their
specifications correctly. High-assurance secure systems have to show convincingly that
they are correct and that their specification does not allow privilege escalation attacks.
However, they may still have covert channels that allow unauthorised access to classified
data.
6
2.1 Information flow architectures
Covert channels
A system contains covert channels [Lam73] if unconnected compartments can communicate. Because this connection conflicts with the access control policy it allows for two
types of attacks. First, a malicious program can observe classified data from another
program. Second, a confined but untrusted program handling classified data can communicate this data to a less classified program.
A system contains covert channels when compartments can induce and assess change
in global resources. Depending on the type of resource, channels are classified as storage
channels or timing channels. Furthermore, resources can be provided by the OS or by
the hardware.
Typical OS provided resources for covert channels are namespaces and file contents.
Also, the exhaustion of disk or main memory can be used as a covert channel. An
example for a covert timing channel in hardware is the utilisation of a cache. Two
compartments can communicate, if one compartment can deliberately cause cache misses
in the other compartment and the other compartment can monitor the resulting changes
in access time.
To mitigate covert channels, the OS has to be designed in a way that does not allow a
compartment to induce or assess changes in global resources. This requirement in turn
demands a thorough examination of every shared hardware and software component.
In this section I defined the requirements for a high assurance secure system. I showed
how privilege escalation attacks and covert channels can overcome the system compartmentalisation set up by the OS. The next section will show how to mitigate these
problems when building high-assurance secure systems.
7
Chapter 2 Background and Related Work
Application layer
Middleware services
Distribution layer
User level device drivers
Separation Kernel
Hardware
Figure 2.2: MILS layer
2.2 Security Architectures
Developing complex high-assurance systems seems to be impossible. Formal verification as developed in the L4.verified project [KEH+ 09] shows the feasibility of writing
correct software. However, the technology does not scale to complex operating systems.
Complexity is often addressed with modularisation and componentisation techniques
because complex problems often consist of various sub problems of lower complexity.
Individual components can then be developed and examined independently.
Usually however, systems are decomposed into functional units. While this approach
aids in developing and managing the system, it does not necessarily make the system
secure. When component boundaries are not enforced at runtime, such systems cannot use untrusted components and rely on the correctness of all components. Systems
based on secure languages (e.g., Singularity/Sing# [FAH+ 06]), enforce boundaries but
do not support legacy components and require expensive reengineering of all parts.
Microkernel-based security architectures [HHF+ 05, FH06] enforce boundaries and support legacy components.
To increase trustworthiness, a security architecture splits the system into trusted parts
and untrusted parts. The trusted parts are called the trusted computing base (TCB)
and it is a goal of security architectures to minimise the trusted computing base. The
TCB includes the OS kernel and core components that all need verification. Classical
systems based on a trusted computing base became too complex to verify. The MILS
architecture tries to remedy this situation by a new approach that I will introduce next.
Multiple Independent Levels of Security and Safety
The ERTOS group tries to develop and verify a high-assurance secure system following
the Multiple Independent Levels of Security and Safety (MILS) approach [AFOTH06].
I will introduce this approach to give an overview of the architectures my work shall
support.
The main idea of the MILS approach is the system decomposition rule. MILS decomposes the security goal of a complex system into several security properties that are
established by components at different security layers (see Figure 2.2). Each component
is build upon the guarantees of its lower layers and is only responsible for its own security property. Security analysis focuses on individual security components to achieve
the overall security goal.
8
2.3 Component Architecture for microkernel-based Embedded Systems
At its lowest layer the MILS architecture requires modern processors that support
efficient context switches to allow isolated OS components. Also, a privileged mode
to control the security features and I/O protection is required. (see Section 2.4 for an
explanation of the x86 architecture.)
The second layer is a separation kernel that controls the processor and abstracts the
processor’s features. It has to provide isolated partitions (compartments) of the computer that can only communicate via confined inter-partition communication primitives.
It must be impossible to circumvent the kernel separation, that is, to communicate to
other partitions by other means than the communication primitives.
The next layer in the MILS architecture is formed by userlevel device drivers that
control I/O devices. Because these devices are not used to partition the computer’s
computing resources they need not be controlled from inside the separation kernel.
Instead they are controlled by device driver partitions. Access control to devices is
implemented using the separation kernel’s communication primitives.
The MILS architecture also contains a layer for distribution and parallelisation that
I disregard in this thesis.
The next layer is the middleware services layer. This layer is built on top of the
separation kernel’s communication primitives and provides more traditional OS services
and information flow control. Example services include resource allocation, higher-level
communication services and real-time data distribution. Information flow control in the
middleware layer ensures that partitions and messages are labelled and inappropriate
messages are filtered. Also it should enable one-way communication between partitions.
To do so, the middleware service layer sets up an information flow architecture and
controls key nodes in the information flow graph.
The preceding layers sum up to what a secure operating system would usually provide.
Based on these guarantees, the final application layer is responsible only for security
functionality. Typical applications are cryptographic services that take data at a high
classification level, encrypt it, and sent it out at a low classification level.
The MILS approach provides guidelines and requirements for the design of secure
systems. However, it does not address software development issues like software reuse
and hardware specifics like bus mastering by I/O devices and device discovery that
are popular on modern computers. These issues are addressed by other means that
I will present and use during this thesis. Also the MILS approach does not present a
separation kernel. Research into microkernels promises to achieve the separation kernel’s
requirements (see Chapter 3). The next section will introduce a software development
technique for microkernel-based systems.
2.3 Component Architecture for microkernel-based Embedded
Systems
CAmkES, the Component Architecture for microkernel-based Embedded Systems [KLGH07, NIC] is a framework for the specification and instantiation of software
architectures. Because its architecture is based on a microkernel, CAmkES was a
natural basis for this thesis. However, it was not used.
9
Chapter 2 Background and Related Work
CAmkES consists of three parts: The architecture description language specifies the
components and how they are connected. A code generator transforms this description
into stub code and an initialisation routine. A startup program and template code
bind the architecture to the underlying microkernel system. In this system, compartments are connected via synchronous and asynchronous inter process communication
(IPC) interfaces, and via shared memory. The component model takes this into account
by providing components and connectors that map to these concepts. Additionally,
CAmkES is based on typed interfaces between components. These interfaces are specified in an interface description language (IDL) and components can provide or use
(require) such an interface.
The architecture description specifies components with their component interfaces and
connections that connect components (see Figure 2.1 for an example). The specification
starts with a list of import statements that include specifications for IDL interfaces and
connectors. It then contains a list of component descriptions. Each component can
provide or use an IDL interface. Components can emit and consume typed events. To
specify usage of shared memory, components can have a typed dataport. Components
with the control keyword, have a thread of control and thus will run from the beginning.
Components can be composed of other, internally connected components. After the
list of component descriptions, an assembly and composition section specifies, which
components are instantiated and how the instances are connected. It starts with a list
of instantiated components. Each entry specifies the type of the component and the
name in this composition. The composition then contains a list of connections. Each
connection specifies the connector, a name for the connection, and the partners. For
each partner the connections also specifies the used stub.
From an architecture description, the build system will generate stub code and infer
a directory layout to find and build the components. The stub code consists of headers
and library functions that translate the invocation of an interface into the appropriate microkernel communication primitive. Also a startup program is generated that
instantiates the architecture.
To port CAmkES to seL4, we have to integrate the CAmkES and seL4 build systems.
This would ease development of seL4 component systems. Instead of this step, this
thesis will provide a basis for seL4 stub code and the startup program.
CAmkES is a powerful tool for the specification and instantiation of software architectures, however, it lacks in two main areas: While being designed to specify operating
systems, it does not take into account the specification and integration of I/O devices.
Also, the current connectors do not restrict information flow. While CAmkES can specify information flow architectures, two connected components can communicate in both
directions. To overcome the first problem area in my thesis, I will present the x86 I/O
architecture and hardware protection mechanisms in the next section.
10
2.3 Component Architecture for microkernel-based Embedded Systems
import " H e l l o . i d l 4 " ;
import " s t d _ c o n n e c t o r . camkes " ;
component HelloComponent {
provides Hello h ;
d a t a p o r t Data buf ;
e m i t s EHello hevent ;
}
component H e l l o C l i e n t {
control ;
uses Hello h ;
d a t a p o r t Data buf ;
consumes EHello hevent ;
}
assembly {
composition {
component HelloComponent hcomp ;
component H e l l o C l i e n t h c l i e n t ;
c o n n e c t i o n IguanaRPC h e l l o ( from h c l i e n t . h , t o hcomp . h ) ;
c o n n e c t i o n IguanaAsynchEvent h e l l o _ e ( from h c l i e n t . hevent ,
t o hcomp . hevent ) ;
c o n n e c t i o n IguanaSharedData h e l l o _ d ( from h c l i e n t . buf ,
t o hcomp . buf ) ;
}
}
Listing 2.1: Example of a CAmkES architecture specification
11
Chapter 2 Background and Related Work
2.4 Hardware compartmentalisation support
A modern Intel x86 computer contains processors [Int08a] and PCI devices [SA99] as
active components, the DRAM as passive memory and the North Bridge and buses as
connection devices (see Figure 2.3). Whereas PCI devices provide specialised operations for specific I/O protocols, the CPU performs general purpose processing. The
CPU is also in charge of controlling the system compartmentalisation. As such, the
CPU executes instructions that modify CPU and machine state. CPU compartmentalisation is achieved via the controlling supervisor mode and the restricted user mode.
I/O compartmentalisation is achieved via the CPU, the PCI bus, and the IOMMU. In
this section I will first present CPU compartmentalisation concepts and then explain
how I/O channels are controlled. Afterwards I introduce the IOMMU and PCI device
discovery.
2.4.1 CPU compartmentalisation
The CPU performs general purpose processing and system control operations. Complete
system control is only possible in supervisor mode. In user mode, only a subset of
the CPU state and machine state can be accessed. This subset is defined by data
structures that are ultimately set up in supervisor mode. To support independent,
isolated execution environments the CPU supports controlled switching between the
modes. The program that runs in the supervisor mode is called the operating system
(OS) or the kernel.
Exceptions
To isolate a single program the CPU has exceptions. When the processor is in user mode
and the user program issues a privileged instruction or addresses unmapped memory,
the processor switches to supervisor mode. The operating system can then decide how
to handle that exception. That is, to stop the program, to handle the exception and
restart the program, or to ignore the exception and continue the user program.
System Calls
To make functionality that requires the processor’s privileged mode to be available to
user mode programs, operating systems use the system call functionality of the processor. A system call instruction is used by a user mode program to voluntarily switch to
the OS and to pass parameters to it. The OS then decodes the parameters, checks the
validity of the request, and performs the requested operation.
Timer
To limit the time a user mode program can use the processor, the operating system
uses a timer device. When the time of the currently running program has run out, the
timer sends an I/O interrupt request to the processor. The processor will then perform
a switch of the execution environment to supervisor mode. The OS can then decide
12
2.4 Hardware compartmentalisation support
Processor
MMU
System Bus
North Bridge
DRAM
IOMMU
PCI Express
Config
PCI Device
Config
PCI Device
Figure 2.3: Computer architecture
whether to schedule another task, continue the current task, or to terminate the current
task.
2.4.2 I/O Channels
The CPU accesses memory and I/O devices via memory requests on the system bus.
The North Bridge decodes these requests and forwards them to either DRAM or to I/O
devices (see Figure 2.3). The so addressed memory is called the physical memory, the
set of addresses the physical address space. The CPU supports memory partitioning
and protection via the memory management unit (MMU).
MMU
The MMU divides the physical address space into 4 kByte pages. It confines isolated
programs into virtual address spaces that hide the physical addresses. Virtual address
spaces are implemented by a translation from virtual addresses to physical addresses that
traverses page directories and page tables. For 32-bit processors, the 10 most significant
bits of a virtual address are used to index into the page directory. A page directory
entry contains the physical address of a page table. The next 10 bit are used to index
into the page table and to get the physical address of the memory frame. The resulting
memory access uses the remaining 12 bit as offset into the frame. The currently used
page directory can only be changed in supervisor mode. When the supervisor program
refrains from mapping page directories and page tables into the virtual address space,
the corresponding user program cannot access any other parts of the memory directly.
13
Chapter 2 Background and Related Work
As devices are memory-mapped, they can, giving a mapping into a virtual address space,
be accessed in user mode.
I/O Ports
Prior to the introduction of 32-bit memory addresses, x86 CPUs accessed I/O devices
via I/O ports in an I/O space. The I/O space is a separate, 16-bit address space that
is still used in modern computers to control legacy hardware, compatibility interfaces
of modern devices, and the PCI bus. The I/O space can only be accessed by special
I/O load/store instructions. It has a separate protection concept but is also routed via
the system bus. Using the I/O permission bitmap, a supervisor mode program can give
user mode tasks access to individual I/O ports.
I/O Interrupts
While I/O devices are active components, they usually perform tasks for the processor.
When they have finished a task, they notify the processor via an I/O interrupt. The
interrupt causes the processor to switch to supervisor mode. The OS then identifies
and handles the interrupt. To continue normal operation, the processor must also
acknowledge the interrupt.
2.4.3 IOMMU
Modern I/O devices are active components and can access memory independently of the
CPU. Using bus mastering [SA99] they can perform direct memory access (DMA) and
bypass the CPU’s MMU protection. To set up a device for DMA, a user mode program
requires the physical addresses of its virtual address space. More importantly, DMA is
not subject to MMU protection. A user program can instruct a device to use DMA to
access any part of memory. The introduction of an input–output memory-management
unit (IOMMU) [Int08b] between PCI devices and the DRAM solves that problem. It
introduces device virtual address spaces for I/O devices. Several devices can share one
address space, and IOMMU address spaces can share pages with MMU address spaces.
IOMMU address spaces are implemented using a similar structure, but with several
9-bit page tables to support 64-bit addresses. The CPU programs the IOMMU via
memory-mapped registers. Those link to context entry tables that contain entries for
each PCI device. Context entries link PCI devices to IOMMU address spaces. Using
proper setup of these data structures, a PCI device can use the same virtual addresses
as its controlling program and can only access the same memory.
2.4.4 PCI device discovery
The operating system has to perform device discovery to identify the I/O devices present
in the system. To do so, it accesses common device addresses and evaluates the reaction
(probing). For legacy x86 devices, the OS assumes their presence and uses them. To
find ACPI [Hew06] reported devices, the OS scans the BIOS part of the memory to find,
14
2.4 Hardware compartmentalisation support
CPU
PCI Bus 0
Device 0
PCI-to-PCI
Bridge
Device 1
PCI Bus 1
Device 0
Device 1
Figure 2.4: PCI bus structure
31 30
28 27
1 Reserved
24 23
0
16 15
Bus
11 10
Device number
8 7
2 1
Function Register number
0
x x
Figure 2.5: PCI configuration address format
by means of a magic number, the pointer to the ACPI tables that describe devices. To
find PCI devices, the OS scans the PCI bus [SA99].
A system can have 1 to 256 PCI buses, each bus can have 1 to 32 devices and each
device can have 1 to 8 functions. Functions correspond to different functional units a
chip might have. For means of identification and configuration every PCI function must
implement a PCI configuration space. A PCI configuration space consists of 64, 32-bit
words of which the first 16 are defined by the standard (see Figure 2.6). However, Intel
x86 processors cannot access the PCI configuration space directly. Instead, they use the
north bridge as proxy. They access the north bridge via the following two 32-bit I/O
ports:
• The Configuration Address Port, at 0CF8h.
• The Configuration Data Port, at 0CFCh.
To access configuration data, the OS writes the address to the address port and then
reads or writes data from the data port. The address is decoded as follows: Bits 23–16
denote the Bus number, Bits 15–11 denote the Device number, and Bits 10–8 denote
the Function number. Bits 7–2 address the word in the function’s configuration space.
The remaining bits are either reserved or have fixed values (see Figure 2.5).
To scan for PCI devices, the OS starts reading identity information of all devices of
bus 0. If a device slot does not contain a device, the PCI bridge reports values of all
1’s. If one of the devices was a PCI bridge, the OS starts probing the bus behind that
bridge (see Figure 2.4). To identify devices, the OS reads the Vendor ID and Device ID
from the first word of the header. Via the Status register, the OS can find out about the
15
Chapter 2 Background and Related Work
31
16 15
0
Device ID
Vendor ID
00h
Status
Command
04h
Class Code
BIST
Revision ID
Cache Line S. 0Ch
Base Address Register 0
10h
Base Address Register 1
14h
Base Address Register 2
18h
Base Address Register 3
1Ch
Base Address Register 4
20h
Base Address Register 5
24h
Cardbus CIS Pointer
28h
Subsystem ID
Subsystem Vendor ID
Reserved
Cap. Pointer
Interrupt Pin
34h
38h
Reserved
Min Gnt.
2Ch
30h
Expansion ROM Base Address
Max Lat.
08h
Lat. Timer
Header Type
Interrupt Line
3Ch
Figure 2.6: PCI configuration header
capabilities of the device. Base Address Register (BAR) 0 to 5 configure I/O regions
of the device. Usually the BIOS configures these regions already such that they do not
overlap. The OS still has to find out where each device’s memory-mapped I/O regions
are mapped in the physical address space. The same applies for I/O regions accessed
by I/O ports.
While access to individual devices can be compartmentalised via memory-mapped I/O
regions and I/O port protection, PCI device configuration cannot be compartmentalised
in hardware. Access to the configuration I/O ports allows a program to change the
hardware configuration. This leads to two problems: a) Two programs can use writeable
configuration space memory as covert channel. b) Specific configurations of the hardware
may break isolation. For example, the Base Address Registers of a PCI device use
physical addresses. When a program changes those to overlay a portion of the physical
memory, another program will read and write to the device’s memory instead of to
DRAM. When the original program changes the configuration back, it can read the
data of the other program (see Figure 2.7).
In summary, PCI configuration space is an important means to communicate device
information to the OS. The device discovery protocol is mainly based on reading information. However, writing data back is necessary as well. There is no compartmentalisation
support for PCI configuration in hardware. When compartments have to use the PCI
configuration space, an OS component must dispatch accesses securely.
16
2.5 Virtualisation
Physical address space
Compartment 1
1. Access PCI configuration
Config
Compartment 2
3.
I/O Region
2. Change I/O region address
I/O Region
PCI Device
I/O Region
3. Capture
memory
accesses
Figure 2.7: Capturing memory accesses using a memory-mapped I/O region
2.5 Virtualisation
Part of this thesis is about reusing existing commodity software using virtualisation. I
will introduce techniques and terminology and outline existing solutions for virtualisation of PCI device discovery.
Various flavours of virtualisation exist for different use cases. In all kinds, a virtual
machine is supposed to be an efficient, isolated duplicate of a real machine [PG74].
However, often, virtual machines do not exist as real machines. The computer the
virtualisation system runs on is called the host system, the software running inside the
virtual machine is called the guest system, the OS the guest OS.
A virtual machine consists of virtual CPU’s, virtual memory and virtual devices. They
are implemented by a combination of hardware and software mechanisms. A virtual
machine monitor (VMM) software controls the physical machine’s hardware that might
have more or less support for virtualisation.
In modern computers, virtual memory can be used to virtualise the memory.
Processor virtualisation requires virtual instructions that modify virtual processor
state. Depending on the physical processor and the virtual processor, virtual instructions must be implemented in a software interpreter (Emulation) or can run directly on
the physical processor (Virtualisation) [NSL+ 06].
To maintain proper isolation, virtualised programs cannot be allowed to execute privileged instructions. Virtualisation systems handle privileged instructions in several ways
[NSL+ 06]:
• they emulate them in hardware (Hardware Virtualisation),
• they emulate them with the help of the OS (Trap and Emulate), or
• they avoid them all together (Binary Rewriting, Paravirtualisation).
For microkernel-based systems, paravirtualisation is used to leverage the code base of
legacy operating systems. To paravirtualise an operating system is to modify its source
17
Chapter 2 Background and Related Work
code to run on the microkernel instead of physical hardware. Code portions that would
use privileged instructions are replaced by calls to the microkernel system call API.
A virtual machine can use the following virtual devices [AJM+ 06]:
• devices simulated by software (Emulation),
• abstract software interfaces (Paravirtualisation),
• physical devices passed-through (Direct Assignment).
To virtualise an I/O device, we have to consider four types of interactions of the
virtual machine and the physical device: device discovery and configuration, device
control (MMIO), data transfers (DMA), and I/O interrupts. Modern systems support
secure efficient virtualisation of the last three operations in hardware.
2.5.1 Virtualisation of PCI device discovery
PCI device discovery on x86 processors uses I/O port access. Respective instructions are
privileged and have to be handled by the virtualisation solution. The virtual machine
monitor then presents the guest OS the devices the user has configured for this virtual
machine. Respective devices can be physical devices passed through to the virtual
machine or virtual devices emulated by the VMM.
Hardware support for virtualisation of device discovery is currently not available.
Virtualisation solutions therefore have to use the software-based mechanisms they use
for privileged instructions. Two such examples are the QEMU open source processor
emulator [Bel05] that uses emulation and the Xen hypervisor [FHN+ 04] that uses paravirtualisation. Respective uses of I/O port instructions or PCI configuration space
operations are translated into accesses to a PCI device model. The VMM implements
this model and emulates a reaction as specified by the PCI discovery protocol and the
virtual devices the VMM wants to present. For a more detailed example see also [Zeh06].
Because modern computers are modular regarding their device configuration, device discovery mechanisms are a requirement. When compartmentalising the computer, device
discovery must be compartmentalised as well. Virtualisation is a well-known compartmentalisation tool. I have surveyed existing techniques to extend an existing virtualisation solution with support for I/O device discovery (see Section 4.1). The next section
will show related work.
2.6 Related Work
Device management
In my study thesis [Hän07], I analysed solutions for device management and I/O compartmentalisation. In summary, current server and desktop OSes do not encapsulate
drivers. Drivers run in the supervisor mode of the processor and can access any memory
18
2.6 Related Work
and I/O channel. Nevertheless, these OSes have mechanisms for controlled access to
I/O resources [CRKH05]. However the OSes do not enforce usage of these mechanisms
and trust all drivers.
Microkernel-based systems mitigate this problem and encapsulate drivers in address
spaces [HLM+ 03, LCFD+ 05]. To manage I/O access these system reused the aforementioned mechanisms [Hel01, Hei05]. While they can guarantee that a driver compartment
only accesses the I/O resources it requested, a driver can request any unclaimed resource.
Memory management
Similarly to I/O resources, common OSes handle memory and kernel object requests
without restrictions. Hence a process can request arbitrary amounts of memory and can
request creation of all kinds of kernel objects, that is, file handles, network connections,
new processes, page tables, and so on. Again, microkernel-based systems faced similar
problems [Elp04].
Virtual machine architectures
The recent move to system architectures based on virtual machines changes the resource
management problem. Applications are no longer handled as processes that request
resources at runtime, but as virtual machines with a predetermined resource allocation.
This changes the approach to resource management, because a (guest) operating system
can usually not send requests to the (virtualised) hardware to provide more resources.
Therefore, I/O device allocations and physical memory allocations are defined by the
user and not by the application.
However, virtualisation solutions often introduce load-balancing techniques that result
in another form of resource sharing. Common memory load-balancing mechanisms like
swapping and the more recent ballooning allow a virtual machine monitor to distribute
physical memory between virtual machines depending on the load. This can lead to a
covert timing channel because the physical memory becomes a cache of disk memory
(see Section 2.1.2). While this attack against resource management is viable, it is rather
theoretical. Because physical memory is big in comparison to disk transfer rates this
attack would only provide a low-capacity covert channel. Additionally, virtualisation
solutions are usually based on common operating systems and hence have a huge trusted
computing base that should have bigger holes in other areas.
In conclusion, high assurance secure systems require isolated components. This can
be achieved with microkernel-based system architectures. However, microkernels alone
provide no security. The next step is secure I/O and resource management. Traditional
solutions are not secure, however, virtual machine architectures have more secure I/O
and resource management.
In this chapter I laid the basis for this thesis. I started with the theoretical background, access control policies and computer compartmentalisation. Then I introduced
a security architecture that promises to enable high assurance secure systems. For such
19
Chapter 2 Background and Related Work
architectures, we require appropriate development methodologies. For this purpose I
outlined CAmkES, the Component Architecture for microkernel-based Embedded Systems. However, my work is not based on embedded systems. Instead I use the x86 and
PCI hardware architecture. I introduced the relevant hardware compartmentalisation
mechanisms and detailed the I/O memory management unit and PCI device discovery.
Also I showed how virtualisation is used to compartmentalise systems and how related
work addresses I/O and resource management. The next chapter will introduce the seL4
microkernel and show how it abstracts the hardware compartmentalisation mechanisms.
20
Chapter 3
seL4
Because the software developed in this thesis is a major component of a seL4 system,
I reserve a complete chapter for seL4, the secure embedded L4 microkernel [EDE08].
I first restate the reasons to use seL4. Then I define the kernel, highlight associated
security guarantees and explain the programming interface. I will conclude with the
state of current user-level applications.
3.1 Motivation
Common operating systems were designed with performance and functionality in mind.
They run most of the OS in the processor’s supervisor mode. Hence a flaw in one part of
the OS kernel affects the security and safety of the entire system. Software verification is
used to prove software bug-free. However, software required to operate today’s systems
has reached a complexity that is beyond conventional analysis and verification methods.
Microkernel-based systems remedy that situation by splitting the system into parts that
can be verified independently. The seL4 microkernel has in turn undergone functional
specification and formal verification [KEH+ 09].
To give guarantees about subsystem isolation, the microkernel may not enable information flow except through defined interfaces. However, OS kernels and initial L4
microkernels enabled communication over global name spaces and resource availability [Elp04]. Restricting access to resource availability implies an explicit resource policy.
The main resources the microkernel cares about are microkernel objects and memory to
back these objects. The seL4 microkernel therefore exports kernel memory management
to user level. To address kernel objects, seL4 defines capabilities. Capabilities form local
name spaces and hence are kernel objects as well. In turn processes also need to manage
their capability spaces.
3.2 What it is
The seL4 microkernel operates in the privileged mode of the processor. As a microkernel it only contains functionality required to export hardware features securely and
efficiently. The basic security guarantee states that a properly isolated compartment
cannot extend its authority.
21
Chapter 3 seL4
The following abstractions are defined by seL4:
Threads as units of activity that time-share the processor and use system calls. Threads
are handled via thread control blocks (TCB) Thread control blocks have a thread
state, a register set, a scheduling priority, an address space, a capability space, an
IPC buffer, and a fault endpoint.
Address spaces called VSpaces, as memory abstractions that partition the main memory. VSpaces are build up from page directory objects, page table objects and
frame objects.
Endpoints as means to communicate between threads. Endpoints can transfer data and
capabilities. IRQs and exceptions are received on Endpoints.
Capability spaces called CSpaces, as containers for references to kernel objects.
CSpaces are implemented as guarded page tables and built up from page-table-like
CNodes.
Physical memory called untyped memory, to control kernel and application memory
usage. Untyped objects have different power-of-2 sizes and can be retyped into
kernel objects.
Hardware control objects like IRQ control, address space ID control, IOMMU address
space ID control, I/O port access. Hardware control objects have specific interfaces.
Via the untyped memory abstraction seL4 externalises kernel memory management
and prevents kernel memory exhaustion. Any memory the kernel consumes is abstracted
in a kernel object. Applications can only create kernel objects to the extent of their
access to untyped memory. Furthermore, once retyped, untyped memory cannot be
typed into something else before it is revoked. Hence, the seL4 kernel guarantees type
safety of its objects. Together with the local name spaces provided by the CSpaces, this
property enables the seL4 security guarantee that “a capability cannot ever cross an
isolation domain”.
For the isolation guarantee to hold, isolated domains must meet the following restrictions [EDE08]:
• No sharing of writeable page tables or writeable CNodes.
• No shared endpoint that can transfer capabilities.
• No access to thread control blocks across domains.
In addition to the seL4 isolation guarantee, a secure system often requires spatial and
temporal isolation between components. While temporal isolation is beyond the scope
of this thesis, spatial isolation is not.
For spatial isolation, additional restrictions must be met:
• No sharing of frame objects.
22
3.3 How to
TCB
Root CNode
TCB
(a)
2-level CSpace
1-level CSpace
0x0
0x3
Endpoint
(b)
Root CNode
CNode
Endpoint
0x3
Endpoint
Endpoint
Figure 3.1: Capability references: (a) CSpace with one level: capability reference 0x3 is used
as an index into the one CNode. (b) CSpace with two levels: capability reference
0x03 is translated, first 4 bits as index into the level-1 CNode, bits 4-7 as index
into the level-2 CNode.
That is, no shared memory. A frame may also not be shared via IOMMU address
spaces.
• No sharing of endpoints, untyped objects, or hardware control objects.
That is, no communication via seL4 objects.
• No sharing of I/O devices.
That is, one domain per device. seL4 represents a device by several capabilities.
All those should be allocated to one compartment.
3.3 How to
Because the software developed in this thesis closely interacts with the seL4 kernel, I will
give a more detailed explanation of the kernel API [NIC06]. I will define the different
seL4 abstractions more precisely and explain their interactions.
3.3.1 Capabilities and CSpaces
A capability gives a user the right to use a kernel object. Using the capability is called
invoking the capability. To invoke a capability, a user specifies the capability reference
during the system call. The kernel uses the capability reference to index into the user’s
32-bit wide CSpace that contains the capability (see Figure 3.1). A capability also states
the seL4 permissions, like read, write and grant, a user has to the kernel object.
In an analogy, capabilities are like pointers to kernel objects. However, these pointers
cannot be on the stack, but come from an array on the heap. To access an object, an
index into the pointer array must be used. It is now the job of the user program to
manage empty slots in this array and to maintain the semantics and relationships of
23
Chapter 3 seL4
31
24 23
16 15
8 7
Guarded page table
20
Translates: 20 bit
Guard-size: 10 bit
Guard-value: 0x000
0x00000
Entries
.
.
.
GPT entry
0x003FF
0x00400
Address space
GPT root
0xFFFFF
.
.
.
0
12
4k page
.
.
.
Data word
.
.
.
No entries
.
.
.
Figure 3.2: Guarded page table example. One GPT can cover a complete 32-bit address space.
filled slots. In another analogy, capability references resemble Unix file handles, they
are pure numbers to the application but array indexes to the kernel. However, CSpaces
are not simple arrays.
CSpaces are implemented like page tables because they require fast, bounded lookup
times, and big maximum sizes, but low average sizes. To allow for variable-sized
CSpaces, seL4 uses guarded page tables (GPT) [Lie94].
In a GPT, every table can have a different, power-of-2 size. Additionally, a table can
cover a bigger address range than it actually translates (see Figure 3.2). Usually, when a
page table translates x bits, it has n = 2x entries. A GPT however can have less entries.
Therefore, every GPT has a number t of bits it covers, a number n of entries, and a
guard with size s and value v such that n = 2t−s . The value v determines the value
that the first s address bits must have to be translated by the GPT. Addresses that fall
into the address range of the GPT, but do not pass the guard, cause page faults.
The guarded page table equivalent in seL4 is the CNode that has a guard and a
number of entries. Capabilities are entries (slots) in CNodes. A CSpace consists of an
hierarchy of CNodes (see Figure 3.1).
Similarly to a virtual address space, not every capability reference in a CSpace translates to a final object. Empty slots can be used to accommodate new capabilities.
Generic capability invocations allow a user to move or copy capabilities from one place
to another. Also this can be used to reduce permissions and to set further attributes.
24
3.3 How to
Server
Client
Root CNode
Root CNode
root sub-CSpace (8 bit)
index CNode
thread's CSpace (4 bit)
index =
SelfCSpace =
0x2
0x1
CNode
CNode
offset1 =
root =
0x4
CNode
0x3
offset2 =
0x5
Endpoint
Endpoint
Endpoint
Figure 3.3: Moving an endpoint capability. To specify the endpoint capability-slot as the
server, set root to 4, index to 1, offset to 3 and depth to 8.
When doing so, the user specifies for source and destination respectively, four parameters
as addresses into the guarded CSpace.
Moving capabilities in a CSpace:
root is a capability reference in the thread’s CSpace and has to point to a CNode
that starts the root sub-CSpace. To modify its own CSpace, a thread needs a
capability in its CSpace, to the top-level CNode of its CSpace. This CNode capability is usually placed at constant capability reference seL4_SelfCSpace := 2.
Most capability invocations will set root to seL4_SelfCSpace. However, a server
with capabilities to client CSpaces, that is, CNode capabilities, can modify client
CSpaces.
index is a capability reference in the root sub-CSpace and points to the CNode that
the user wants to modify.
offset notifies the slot or entry in the index-CNode that the user wants to modify.
depth allows to stop the translation before the lowest level of the CSpace and enables
a user to place capabilities in earlier CNodes. It denotes the number of bits
translated during the lookup.
For an example, see Figure 3.3.
3.3.2 Kernel memory management - Untyped_Retype
Untyped objects can be cast into other kernel objects. To do so, the user invokes
retype on an untyped object, specifying the type and size for the new objects, the
number of new objects to create, and a range in one CNode to place these objects.
For an example, see Figure 3.4. Space not used by the retype operation is lost to
internal fragmentation. However, the untyped object can be revoked, resulting in the
deletion of all derived objects. Deletion of individual derived objects increases internal
25
Chapter 3 seL4
TCB
Root CNode
thread's CSpace
CNode
0x3 Untyped
0x4 4k Frame
0x5 4k Frame
8k untyped
memory
0x2
4k Frame
4k Frame
Figure 3.4: Retyping untyped memory. Transform 8k untyped-memory at capability reference
3 into 2 4k frames, create capabilities in the CNode with capability reference 2, in
the range 4 to 5.
fragmentation. Untyped objects can be retyped into other, smaller untyped objects to
prevent fragmentation.
3.3.3 VSpaces
seL4 does not have memory mappings via flexpages as L4 [HWL96] has. Instead, system
applications have to build and manage their address spaces with hardware objects.
However, they do not need to take care to not overwrite them accidentally.
To create a new VSpace, an application associates a page directory object with an
address space id (ASID). It then maps page tables and pages into the VSpace at an
address and with associated rights. Existing mappings will be overwritten, and one
page capability can only be mapped into one VSpace. To share a frame, the frame
capability must be copied, and the copy mapped into the other VSpace.
For an example, see Figure 3.5.
3.3.4 Threads, Endpoints and IRQs
seL4 user-space threads have similar semantics as L4 threads [Lie96]. A thread is identified by a capability to its Thread Control Block (TCB). A TCB allows setting of the
address space, the capability space, the IPC Buffer and the scheduling priority (ThreadControl). Furthermore, the register set can be read and written (ExchangeRegisters),
and the thread can be stopped (Suspend) and be started (Resume). The IPCBuffer
is a shared memory page between the kernel and the user-mode thread to exchange
parameters during system calls. A thread has to map the IPCBuffer into its VSpace to
access it.
Endpoints are connector objects that, when copied to several CSpaces, can be used
by respective threads to exchange messages and capabilities via IPC.
Capabilities to endpoints, like all other capabilities, have permissions associated with
them. Endpoint capabilities without seL4_Grant cannot be used to transfer capabilities,
26
3.3 How to
ASID pool
TCB
Root VSpace
Root CNode
VSpace
Page table
0x001 Page table
thread's CSpace
.
.
.
ASID
Page directory
.
.
.
.
.
.
0x3fe
0x3ff
4k Frame
4k Frame
0x3 Page directory
Page table
0x5 4k Frame
0x6 4k Frame
4k Frame
4k Frame
0x4
Figure 3.5: Example with seL4 address space objects
endpoint capabilities without seL4_Read cannot be used to receive from the endpoint,
and endpoint capabilities without seL4_Write permission cannot be used to send to the
endpoint.
In addition to permissions, endpoint capabilities have 32-bit badges that allow a program to distinguish between different senders on one endpoint. To use a badge, a server
creates a copy with badge of an endpoint, and then moves the copy to the client. When
the client sends to this endpoint capability with badge, the server receives the badge
back.
seL4 supports synchronous and asynchronous endpoints. Synchronous endpoints are
used for conventional, blocking or non-blocking L4 IPC with message transfer. Asynchronous endpoints are used for non-blocking send and only modify one buffer word
of the endpoint by XOR-ing the bits of the sender’s badge and the buffer word. This
allows a receiver to distinguish between 32 senders. When reading from an asynchronous
endpoint, the buffer word contains all bits set since the previous read and is reset.
Each thread also has a synchronous fault endpoint. When the thread pagefaults, traps
or executes an undefined instruction, seL4 sends an IPC to the fault endpoint containing
the registers associated with the exception. To continue the faulting thread, the fault
handler thread has to reply with an IPC with new registers or restart the thread.
To use a service, clients call an endpoint associated with a server. Calling translates
to synchronous sending and immediate waiting.
With the IRQ control object, individual IRQ objects can be created. Individual
IRQ objects can link to an endpoint and send interrupts as asynchronous IPCs to this
endpoint. To make it possible to receive an interrupt again, the user has to invoke the
individual IRQ’s acknowledge operation.
27
Chapter 3 seL4
3.3.5 Start up
At startup, seL4 loads the initial program’s ELF-file and starts an initial thread. This
thread has an initial CSpace that contains capabilities to all the thread’s objects, to all
hardware control objects, and to all untyped memory objects (see Figure 3.6).
The initial CSpace is a 2-level CSpace with a big level-1 CNode and several level-2
CNodes with 256 entries, aka 8-bit translation, each. However, for historic reasons,
level-2 CNodes translate 12 bit, to look like pages in a 2-level page table. Hence, level-2
CNodes have a 4-bit guard with value 0h and are 4096 capability slots distant from each
other. (That leaves gaps of 3840 (=4096-256) slots).
To inform the thread about the semantics of its capabilities, seL4 provides a static
bootinfo data structure at a fixed address. Many drawbacks of the bootinfo’s coding
make using it a nightmare. Entries in this array generally describe ranges of capabilities
with the same semantics.
The following semantics are defined by the seL4 bootinfo for regions of the initial
CSpace:
seL4_Region_Empty a region not covered by L1 or L2 CNodes
seL4_Region_RootTask frame objects for bootinfo, IPCBuffer and code
seL4_Region_L1Node number of entries and guard of the L1 CNode
seL4_Region_L2Node number of entries and guards of the L2 CNodes
seL4_Region_FreeSlots free slots in L2 CNodes
seL4_Region_InitCaps the initial thread’s TCB, CSpace, VSpace and hardware control objects
seL4_Region_SmallBlocks 4-kB untyped memory objects
seL4_Region_LargeBlocks sets of different power-of-2 sized untyped memory objects
seL4_Region_DeviceCaps frame objects to I/O device memory
3.3.6 x86 and IO protection
seL4 does not use hardware protection for I/O ports (see Section 2.4.2). Instead, I/O
ports are mapped to special system calls. To invoke these system calls, a thread needs
the seL4_IA32_InitIOPort := 7 capability.
seL4 uses IOMMUs (see Section 2.4.3) to give devices secure, direct memory
access [Pal09]. seL4 only allows one PCI device per IOMMU address space. The seL4
kernel already sets up root tables for all IOMMUs and context tables for all PCI buses.
To create an IOMMU address space object, the user needs the IOMMU control object
and invokes the create operation with the address (bus, device, function) of the PCI
device. The user can then map IOMMU page tables and frames into the address space.
28
3.3 How to
TCB
Root VSpace
VSpace
Page directory
Root CNode
Level-2 CNode
Level-1 CNode
Guard
0x00 (6 bits)
0x00
Guard
0x0000
0x0001
0x0002
CNode
Roottask
Empty
0x09fe
0x15b5
0x02
0x03
0x04
0x0003
0x09ff
0x01
0x05
0x06
0x07
Roottask
(4k frames)
0x15b6
Empty
0x3fff
(16k slots)
0x08
0x09
0x0E
0x1D
0x00 (4 bits)
Nil cap
TCB
CNode
Page directory
IRQ control
ASID control
ASID pool
IO port
IOMMU control
Page table
4k untyped
memory
0x1E
Device caps
(frames)
0x46
0x47
PDE
Large blocks Untyped
memory
0x6c
0x6d
Free slots
0xff
Page table
PTE
PTE
4k Frame
4k Frame
4k Untyped
4k Untyped
Network card
4k Frame
4k Frame
VGA card
4M Frame
4k Frame
1k Untyped
1M Untyped
16M Untyped
32M Untyped
(256 slots)
Figure 3.6: seL4 initial CSpace on QEMU
29
Chapter 3 seL4
3.3.7 Devices
In the current IA32 version of seL4, the kernel does PCI device discovery (see Section 2.4.4) at boot time. However, it only reads out the BARs and creates frame
capabilities for every I/O page. These capabilities end up in the CSpace of the initial
thread and are reported to it as DeviceCaps in the bootinfo (see for example Figure 3.6)
Regions with contiguous physical addresses are merged to one capability region in the
bootinfo. To identify them, bootinfo entries contain the first physical address of the
I/O memory region. As device regions often span several, for example 64, pages, several devices fill up the 256-entry L2 CNodes. As it happened that caused regions of
one device to span two L2 CNodes. Finally, in the initial CSpace (see Section 3.3.5),
L2 CNodes have gaps between them, so contiguous physical ranges are represented by
non-contiguous capability ranges.
A user program is supposed to do PCI device discovery as well and parse the bootinfo
to match the BAR-reported I/O region with frame capabilities. To do PCI device
discovery, a user program uses the seL4_IA32_InitIOPort capability.
3.4 Current implementation status
At the time this thesis was started, the major challenges introduced by the new seL4 concepts already had detached solutions. There was a paravirtualised version of Linux that
used a SLAB-like memory allocator [Bon94] (Iwana). Also small examples demonstrated
the use of PCI device discovery and direct memory access via the IOMMU [Pal09]. In
this section, I will explain the object allocator and paravirtualisation in more detail.
3.4.1 Iwana - the kernel object allocator
The Iwana memory library [Elk09] is closely interwoven with the startup program. This
program handles the initial CSpace and transforms it into a much easier CSpace. The
Iwana allocator is initialised on that CSpace and used to load Linux. Linux in turn uses
Iwana to create processes and to manage their address spaces.
In contrast to the initial CSpace (see Figure 3.6), Iwana requires a CSpace where
the level-1 CNode and the level-2 CNodes both translate 10 bits. It only support 4 kB
and 1 MB untyped memory objects. Capabilities for each memory type must be in one
contiguous capability region. The startup program creates that Iwana-CSpace by using
ad hoc object allocation. Using seL4 compile time options, the developer ensures that
the initial program has enough objects for the ad hoc allocation.
After setting up the Iwana-CSpace, the initial program stores information about that
CSpace into a static bootinfo data structure at a fixed address. It then starts a new
thread in the new CSpace. The thread then initialises the Iwana allocator with the
range of 4 kB untyped memory objects, the range of 1 MB untyped memory objects,
and the location of free slots in the L1 CNode and the current L2 CNode.
After initialisation, programs can use Iwana to allocate any seL4 objects: thread
control blocks, endpoints, page directories, page tables, frames, CNodes, and slots.
Iwana internally allocates CNodes to grow the CSpace as necessary and retypes untyped
30
3.5 Application to this thesis
memory in larger chunks of 256 objects at a time. It stores the affected capability
references in pools per object type and hands them out when requested. Although
objects can be freed into these pools, they are never revoked into untyped memory.
3.4.2 Paravirtualisation - seL4::Wombat
The ERTOS group has a paravirtualised version of Linux [Elk09] that follows the
L4Linux approach [HHL+ 97] and is based on Wombat [LvSH05]. This version uses the
Iwana allocator and was the only application loadable by the startup program. After
initialising Iwana, the startup program loads the Linux ELF-image into a VSpace that
it creates on-the-fly. For simplicity reasons, Linux uses the same CSpace as the startup
program. After starting Linux, the startup program assumes the role of the timer-server.
The timer-server, too, is in the same CSpace. Linux will initialise the Iwana library in
a free area of this CSpace.
When Linux is started, it allocates a heap with a known mapping from physical
address to capability reference, allocates and starts the system call thread that initialises
the architecture independent part.
Processes are implemented as individual seL4 threads when they are in user mode and
execute on the L4Linux system call thread when they are in the Linux kernel. Kernel
entry and exit are mapped to fault IPC and reply. As such, Linux maintains endpoints
for all its processes and waits on them whenever it has scheduled the corresponding
process.
When this thesis was started, paravirtualisation of processor and memory was working, however Linux did not have access to I/O devices. The only device access was an
interface to an I/O port based timer server and the seL4 kernel’s debug putchar system
call.
3.5 Application to this thesis
In this chapter I presented seL4. The seL4 microkernel employs new design approaches
that require proper handling at user level. Object capabilities allow fine-grain control
over resources and can prevent covert channels. However, they require handling of local
names and local name spaces. Also, I/O devices must be identified and capabilities
to access device resources must be created. To create kernel objects, the user must
manage physical memory. To create compartments, an entity has to create all required
kernel objects and assemble them. When instantiating an information flow architecture,
compartments must get access to their program, to device resources, and to other compartments via defined interfaces. The next chapter will present my solutions to these
tasks.
31
Chapter 4
Design
The task for this thesis is to compartmentalise an x86 system using the seL4 microkernel. The goal is to achieve the strongest compartmentalisation possible by design.
Therefore, standard OS services and features are cut back and replaced by a build time
specification mechanism. The resulting system is static and cannot change at run-time.
However, a specification mechanism shall allow to describe different hardware-software
compositions.
I extended an existing seL4 startup program with support for I/O devices and multiple
compartments. This program can now be parameterised by data structures that describe
I/O devices and compartments.
In this chapter I will develop a secure approach to device discovery and show how I
built seL4 compartments and how these compartments are parameterised. I will first
provide my solution for PCI device discovery by the OS. Then I show how a similar
mechanism can be used in untrusted, deprivileged, paravirtualised guest operating systems. I will then give an approach to secure PCI configuration. In the second part
of this chapter I will derive the specifications that are required to instantiate secure
architectures.
4.1 Handling I/O devices
In an information flow architecture a computer consists of I/O devices that require
special consideration. Four kinds of interaction between devices and compartments
are important: device discovery, device control, data transfers, and I/O interrupts.
seL4 already provides the last three operations, the device communication channels, to
compartments. However, it lacks understanding of what a device is (as appropriate for
a microkernel). To give a compartment access to a device, the loading process has to
discover devices on the machine and collect their associated seL4 resources. It then has
to give these resources to the compartments.
Device discovery is in conflict with a completely static system. I therefore propose
to treat x86 systems as embedded systems for security purposes. Instead of dynamic
discovery and usage of devices, the system developer has to specify all I/O devices he
wants to use. The loading process then assumes these devices are present and fails
otherwise.
The OS learns about PCI devices by scanning the PCI bus (see Section 2.4.4). To
dispatch PCI devices to compartments, the loading process has to know the communication channels each device uses. For each device, the PCI configuration space contains
33
Chapter 4 Design
Loading process
Device explorer
PCI discovery code
PCI discovery code
Hardware
Hardware
PCI
a)
PCI
Machine
specification
Loading process
Technical
Documentation
Hardware
PCI
b)
Figure 4.1: Offlining PCI device discovery: a) Device discovery code is in the trusted computing base, b) only the machine specification has to be trusted.
this information. The allocation of I/O resources to PCI devices is performed by the
BIOS on system startup.
4.1.1 PCI device discovery by the system
The loading process is responsible for splitting the resources between compartments. As
such it is in the trusted computing base and must be correct. To show correct device
discovery, one must have a formal model of PCI and show that the loading process
uses it correctly (see Figure 4.1, a). Under the assumption that PCI discovery always
returns the same result for the same machine, this result can be saved and reused. That
way, device discovery can be removed from the loading process. By offlining access to
the actual hardware, verification of the loading process is simplified. The offline device
discovery yields a copy of the PCI configuration space. This representation is then
analysed, confirmed, and linked into the binary of the loading process (see Figure 4.1, b).
The prototype implementation showed that this design is feasible and that PCI devices
can be operated without prior access to their configuration spaces. To ease machine
specification, I developed the deviceexplorer tool to read out the PCI configuration space
of a machine. It runs on Linux and accesses the /proc/bus/pci virtual directory to read
device information. It then formats this information according to a device specification
and writes it into a C-header file. I will present this specification in Section 4.2.1 of
this chapter. The developer then assigns identifiers to the devices he wants to use. In
the architecture description the developer uses these identifiers to grant a compartment
access to a device.
4.1.2 PCI device discovery by compartments
Compartments, too, must have a way to discover the I/O devices they consist of.
Because the PCI configuration space can also be used to modify the I/O resources
and has no protection concept, hardware access to it cannot be given to isolated compartments. Instead, we require another way to transfer device descriptions into a com-
34
4.1 Handling I/O devices
Guest OS
Guest OS
PCI discovery code
PCI discovery code
virtual PCI
VMM
virtual PCI
a)
loading process,
seL4
b)
Figure 4.2: PCI virtualisation. a) classical: emulation by VMM, b) emulation inside the VM
compartment
partment. For custom made components, device information can be encoded in a data
structure or the initialisation code.
To make a paravirtualised OS use a set of devices, the developer can configure the
OS to only have drivers for these devices. For PCI devices, one additionally needs to
make the OS activate the drivers. In a non virtualised OS, this activation is done by the
PCI discovery code that starts drivers for every device it finds. This same functionality
is required in the virtual environment as well. Instead of scanning the hardware, the
OS has to scan a software generated list of devices. It then has to activate the PCI
drivers. However, the interface between the PCI code and PCI drivers is quite complex,
is a moving target and generally belongs to the OS internals. A less intrusive approach
is the paravirtualisation of the PCI device discovery code itself. Instead of replacing
that code, you can intercept its access to hardware. By emulating reads and writes to
the PCI configuration space, the existing code can be reused without changes. To do
so, the paravirtualised OS must be provided with a copy of the configuration space of
all devices it is supposed to drive (see Figure 4.2). Reads can then just return values
of the copy. However, the discovery protocol also contains writes. On closer analysis,
these writes are only required to access another static value, the I/O region sizes. In
this case, the guest OS writes all 1’s to each base address registers (BAR) of every PCI
device. This action has to make the next read to the BAR report the encoded size of the
I/O region. Implementing these two mechanisms, read from copy, and emulate write to
BARs, proved to be enough to paravirtualise PCI device discovery.
4.1.3 PCI configuration
PCI is not only used for device discovery, but also to configure individual devices. For
devices that require configuration via PCI, a secure way to access the PCI configuration
space must be devised. I will present three designs.
First, the system developer can try to avoid using devices that require PCI configuration access. Under the assumptions that the configuration set up by the BIOS is secure,
this approach reduces the complexity of the necessary trusted software. However, it
might be impossible to avoid these devices because certain features cannot be config-
35
Chapter 4 Design
ured by the BIOS. Examples include message signalled interrupts [SA99] and DMA
modes of popular disk controllers [Int07].
Second, the developer can offline PCI device configuration when his system does not
change device configurations at run time. Many device drivers only configure devices
when they are started. A developer could therefore extract the configuration part from
the driver and move that code to a trusted instance like the loading process. The
loading process then configures the device before it starts the untrusted OS. The OS
then only accesses a local copy of the PCI configuration space. To make the OS think
it is controlling the device, the PCI virtualisation layer might be required to emulate
effects of configuration space writes.
Third, if configuration access is necessary during the runtime of the system, the
developer has to include a secure server that controls writes to the PCI configuration space. That server needs access to the SEL4_IA32_InitIOPort capability (see
Section 3.3.6) and needs to have an interface that allows confining access to individual
devices (see also [Hän07]). Similar to the seL4 kernel interface, it could provide a control
capability that allows creating capabilities for single device objects. A secure instance
then creates these capabilities and hands them over to untrusted components. Such
capabilities can be implemented with endpoints (see also [LW09]) and it would be the
servers responsibility to ensure that requests on one endpoint only access the associated
device. Additionally, the developer has to decide, whether to export byte access to the
configuration space or just a set of functions like enable_PCI_COMMAND_MEMORY() and
set_PCI_PM_CTRL(int value). However, most settings determine the consumption of
PCI resources by the device. As these resources are also global, they cannot be controlled
by an untrusted compartment. For example, access to the PCI BARs determines the
mapping of device memory into the physical address space. If an untrusted application
can change this mapping it can capture accesses to memory (see Section 2.4.4).
For the prototype I chose the first design, because the devices I used did not require
PCI configuration.
In this section I handled device discovery and configuration. I showed that it is
useful to separate PCI device discovery from PCI device configuration. I also moved
virtualisation of device discovery from the virtual machine monitor into the virtual
machine compartment. Thereby I reduced the complexity of the VMM and made a
global PCI discovery service unnecessary. The next section will handle further device
operations and the modelling of compartments.
4.2 Architecture specification
This section describes the manifestation of an information flow architecture in an x86seL4 system. I define the relevant partitions of a computer (compartments) and derive a
method to specify them. This specification is then processed by a loading process. The
main tasks of the loading process is to split resources and to instantiate an architecture
on top of the seL4 microkernel. This thesis is about the second part. To instantiate
all the compartments I needed to find out their minimal resource requirements. The
remainder of this section will show what seL4 resources are required to control a device,
36
4.2 Architecture specification
to create different types of compartments, and to connect compartments. Furthermore,
I will give the developed specifications as well as the interface by which compartments
get to know about their resources.
4.2.1 Device specification
To assemble the seL4 resources of a device, the loading process must know the identity
and communication channels of the device. It can acquire them via PCI device discovery
or via a static configuration.
To assemble the seL4 resources of a device, the following information is required:
• the physical addresses of the I/O memory regions to control the device,
• whether the device is controlled via I/O ports,
• the IRQ number to receive I/O interrupts from the device,
• the PCI device address to create an IOMMU address space.
When this information is acquired, the loading process has to create or find the
corresponding seL4 capabilities.
The following seL4 capabilities are required to use an I/O device:
• a range of frame capabilities to I/O memory of the device, mapped into a VSpace
• the seL4_IA32_InitIOPOrt capability for I/O port access
• an IRQ capability and an endpoint capability associated with the IRQ
• a device-specific IOMMU address space id, IOMMU page tables and frame capabilities mapped into this IOMMU space and a VSpace.
To reduce the trusted computing base I chose a static configuration. The developer
uses another OS and my deviceexplorer tool to dump the PCI configuration space into a
list of I/O devices of the target machine. Using this list, the loading process acquires the
relevant capabilities and copies or maps them to the compartments driving the devices.
To maintain the meaning of the newly created capabilities I derived a data structure
that stores device capability references. This data structure contains sections for each
device operation. Each section describes both the physical property of the device as
well as the corresponding seL4 capabilities.
The first section is about device identification. Currently only PCI devices are handled. A PCI identifier contains the device address (source id) and a copy of the PCI
configuration space (see Listing 4.1).
Memory mapped I/O regions are characterised by a range of physical addresses and a
range of capabilities. For later purposes, flags and seL4 permissions were also required
in this specification (see Listing 4.2).
The complete device specification (see Listing 4.3) contains up to 6 MMIO regions,
the PCI structure and capability references for the IRQ and the IOMMU space. The
37
Chapter 4 Design
typedef struct {
unsigned int p c i _ s o u r c e _ i d ;
unsigned int c o n f i g _ s p a c e [ 6 4 ] ;
unsigned int b a r _ s i z e [ 6 ] ;
} wombat_pci_device_t ;
Listing 4.1: PCI device identification data structure
struct s e L 4 _ r e s o u r c e {
uint32_t
uint32_t
unsigned long
seL4_CapRights
seL4_CPtr
seL4_CPtr
};
start ;
end ;
flags ;
rights ;
cap_start ;
cap_end ;
Listing 4.2: I/O region data structure
IO port capability is not mentioned because it is always at the same capability reference. However, the device specification states whether a device requires IO port access,
whether it uses DMA, and whether it can send IRQs. The IRQ number is also included.
Using the device specification, the loading process assembles the seL4 resources and
assigns them to a compartment. To report the device and its seL4 resources to the
compartment, the loading process copies the device specification to the compartment’s
bootinfo.
4.2.2 Compartment specification
To create a compartment in seL4 is to assemble all the necessary kernel objects. Different
compartment types require different seL4 resources. Resource requirements are based on
four classes: minimal isolated domains, device driver domains, virtual machine domains,
and compartment connections.
Minimal isolated domains
For every isolated seL4 application, the following capabilities must be assembled:
• a Thread Control Block (TCB), to execute the application,
• a VSpace, filled with enough page tables and page frames, to
– hold the program,
38
4.2 Architecture specification
t y p e d e f s t r u c t se L4_ device {
/∗ A numeric i d e n t i f y i n g name f o r t h i s d e v i c e . ∗/
enum device_name name ;
#d e f i n e SEL4_DEVICE_USES_DMA
#d e f i n e SEL4_DEVICE_HAS_IOPORTS
#d e f i n e SEL4_DEVICE_HAS_IRQ
#d e f i n e SEL4_DEVICE_VIRTUAL_IRQ
unsigned i n t
flags ;
0 x00000002
0 x00000004
0 x00000008
0 x00000010
unsigned i n t
unsigned i n t
irq ;
irq_cap ;
unsigned i n t
iommu_space_cap ;
#d e f i n e SEL4_DEVICE_IS_PCI
0 x00000000
unsigned i n t
type ;
union {
wombat_pci_device_t p c i ;
} id ;
s t r u c t seL4_resource resource [ 6 ] ;
} seL4_device_t ;
Listing 4.3: Device description data structure
– hold the IPCBuffer, and to
– hold the stack, and
• an empty CSpace to confine authority of the application.
Most of these properties are determined by the application binary. The existing loader
has a function to build a VSpace from an application binary. Binaries are identified in
a combined ELF-file by a string. To parameterise loading of a program, I refer to the
binary’s file name. For a sample specification of a simple compartment, see Listing 4.4.
Device driver domains
Device driver domains have all the requirements of a simple application. In addition,
they need access to devices. As discussed previously, the following is required to drive
a device:
• a description of the device and its seL4 capabilities,
39
Chapter 4 Design
component_t app = {
. name = " Simple App"
. binary_name = " app . e l f "
. CSpace_depth = 4 ,
};
Listing 4.4: Example minimal compartment specification
component_t n e t w o r k s e r v e r = {
. name = " Network s e r v e r "
. binary_name = " network−s e r v e r . e l f "
. CSpace_depth = 4 ,
. d e v i c e s = {HARDWARE_NIC_0, } ,
};
Listing 4.5: Example device server specification
• the I/O memory of the device mapped into the VSpace, possibly involving page
tables,
• the I/O port access capability,
• the IRQ capability and the associated endpoint capability,
• an IOMMU address space that shares memory with the driver’s VSpace.
The device-specific capabilities to create all these resources are already referenced by
the device specification. Device specifications in turn are referenced by an identifier. To
notify assignment of devices to a compartment, I gave the compartment specification a
list of device identifiers. For a sample specification of a device server, see Listing 4.5.
Virtual machine domains
Virtual machine domains also have the requirements of seL4 applications and of driver
domains, when they drive devices. In addition, they must be able to manage and control
compartments on their own. Therefore, they need access to many seL4 objects. To keep
the complexity of the loading process small, the virtual machine should create all these
objects itself. Therefore, a virtual machine domain requires the following:
• write-access to its CSpace,
• untyped memory to create seL4 objects,
40
4.2 Architecture specification
component_t n e t w o r k r o u t e r = {
. name = "Wombat 1 "
. binary_name = " vmlinux . e l f "
. CSpace_depth = 2 0 ,
. d e v i c e s = {HARDWARE_NIC_0, HARDWARE_NIC_1} ,
. smallUT = 9 , / / ( i . e 9 ∗ 4k )
. largeUT = 2 1 , / / ( i . e . 21 ∗ 1MB)
. c m d l i n e = " i n i t =/ e t c / i n i t . d/ d r o p b e a r 0 " ,
};
Listing 4.6: Example virtual machine specification
• a big, or a grow-able CSpace,
• the ASID control capability,
• a capability to its own VSpace, and
• a description of its initial capabilities and initial CSpace and VSpace.
To mark a compartment as virtual machine domain, I use the size of the CSpace in
the compartment specification. A CSpace_depth of 20 signifies to the loading process to create an Iwana-compatible 2-level CSpace. The loading process will also give
the compartment the required, above mentioned, capabilities. Of these requirements
I parameterised the amount of untyped memory. I also added a command-line string
to the compartment description. This parameter can be used to start virtual machines
with different kernel command-lines. For a sample specification of a virtual machine,
see Listing 4.6.
For the entire compartment description data structure, see Listing 4.7.
4.2.3 Connection specification
To connect seL4 domains, the following options exist:
• connect via shared seL4 objects, like endpoints, TCBs, CNodes, page tables, and
so on;
• connect via shared hardware objects, like memory frames, devices.
To maintain isolation, only shared memory and non-grant endpoints are acceptable (see
Section 3.1). Because connections do not belong to one of the connected parties, I treat
them as separate objects in the specification.
A connection object can be shared between multiple compartments. To model this,
the object can be modelled as a passive compartment and links between the object and
41
Chapter 4 Design
typedef struct {
c h a r ∗name ;
int id ;
loader_DeviceIDs_t d e v i c e s [ 5 ] ;
char ∗ cmdline ;
c h a r ∗ binary_name ;
c h a r ∗ramdisk_name ;
u n s i g n e d i n t CSpace_depth ;
u n s i g n e d i n t smallUT ;
u n s i g n e d i n t largeUT ;
unsigned i n t d e v i c e O f f s e t ;
// i n t e r n a l
s t r u c t app_info ∗app ;
} component_t ;
Listing 4.7: Compartment specification data structure
a compartment denote mapping of the object into the compartment (see Figure 4.3, a).
To ensure information flow control, mappings can be restricted to read-only or writeonly access to the shared object. For simple two-way connections, this model adds
overhead, that is, specifying the object and two mappings, instead of one specification
object. Therefore I used a simpler version. In my model, connections are specified for
two compartments only (see Figure 4.3, b).
In the specification (see Listing 4.8), a connection item specifies the two compartments
it connects. Furthermore, I developed two connection types, one based on endpoints,
the other based on shared memory. A type parameter specifies this type. The next
parameter encompasses type-specific information and the final two parameters specify
the seL4 permissions each compartment has to the objects of this connection.
In addition to setting up the connection, the loading process also has to inform both
domains about their communication capabilities. Later argument will show that I model
connections close to I/O devices. I therefore overloaded the list of I/O devices in a
compartment’s bootinfo to also describe compartment connections.
The problem with select
When a compartment has several endpoint connections, it might want to wait on all
connections at the same time (select()). Especially in a client-server scenario, the
server wants to wait for all clients. This can be achieved with a multithreaded server
that has one thread per endpoint, or with an endpoint shared between the server and
all clients.
A shared endpoint poses a problem to the compartmentalisation (see Figure 4.4, a).
When an endpoint is shared between a server and its clients, the clients can communicate
via this endpoint. To prevent this unintended information flow, the shared endpoint
42
4.2 Architecture specification
Compartment
Compartment
Compartment
Endpoint
Endpoint
Compartment
Memory
Compartment
Endpoint
Memory
Compartment
Compartment
Compartment
a)
b)
Figure 4.3: Connection options. a) Objects shared between multiple compartments, b) Connection between two compartments
typedef struct {
component_id_t
component_id_t
connection_types_t
union {
unsigned i n t
unsigned i n t
void ∗
} data ;
seL4_CapRights
seL4_CapRights
} connection_t ;
from ;
to ;
type ;
protocol ;
irq ;
dev ;
from_rights ;
to_rights ;
Listing 4.8: Connection specification data structure
43
Chapter 4 Design
Server
Server
shared
memory
Endpoint
Client
Client
a)
Endpoint
Client
shared
memory
Client
b)
Figure 4.4: a) Conflict of select and compartmentalisation. b) Solution
must be read-only or write-only for all clients. This restriction prevents the classical
call-reply server protocol. My conclusion from this observation is that the current seL4
API does not allow call-reply with a shared endpoint without enabling communication
between all partners.
However, the unidirectional notification mechanism, that is, write-only clients, is still
useful, if a client and a server are connected via multiple links (see Figure 4.4, b).
Compare the situation to an I/O device that has I/O regions shared with a compartment
and can send an interrupt to the compartment. The compartment can receive interrupts
from this or other devices, however devices cannot send interrupts to each other.
I disregarded the multithreaded server as being too complex a solution for this problem. Instead, I use shared notification endpoints for client-server scenarios.
Asynchronous endpoint connection
Following the above discussion, I decided for a model where each compartment can be a
server waiting for several clients at once. Also, a compartment can send notifications to
several servers. Therefore, each compartment has one endpoint to receive from several
clients and multiple endpoints to send to different servers. For one compartment I call
these endpoints the incoming endpoint and the outgoing endpoints (for an example, see
Figure 4.5).
Note also that a write-only endpoint has a covert back-channel, if it is a synchronous
endpoint. Although a sender may only write to an endpoint, it can determine whether
the receiver is waiting or not. To enforce unidirectional connections, I have to use
asynchronous endpoints.
An asynchronous endpoint does not allow data transfer. Instead, non-blocking IPC
only modifies one buffer word of the endpoint by XOR-ing it with the bits of the sender’s
badge. When reading from this endpoint, the buffer word contains all bits set since the
previous read. The server uses these bits to distinguish between different clients. In
addition, the incoming endpoint is used to receive I/O interrupts. In my model, clients
44
4.2 Architecture specification
Compartment
Compartment
Incoming endpoint
Compartment
Outgoing endpoint
Figure 4.5: Example of endpoint connection model
s t r u c t seL4_outgoingEP {
seL4_CPtr
cap ;
unsigned i n t
protocol ;
};
Listing 4.9: Outgoing endpoint reporting data structure
are represented to the server as connections. A connection specifies the “IRQ number”
as the bit the server will see when somebody uses this connection.
When specifying an endpoint connection, the ’from’ compartment is the client id,
and the ’to’ compartment is the server id. The server receives client requests as virtual
interrupts, whereas the client has an endpoint to send notifications. The data parameter
of an endpoint connection specifies the IRQ number the server associates with this client.
An extension to the current endpoint connection could allow several incoming endpoints to allow for the call-reply scheme between two compartments.
When a compartment is the target (to) of an endpoint connection, the loader will
create a new device entry in the compartment’s bootinfo. The loader sets the device’s
IRQ number and sets the flags field to HAS_IRQ and VIRTUAL_IRQ. This tells the
interrupt code to correctly dispatch this notification but to not acknowledge an interrupt.
For the from compartment, the loader creates a new entry (see Listing 4.9) in the
outgoing endpoints list in the compartment’s bootinfo.
Shared memory connections
To connect compartments via shared memory, I have to devise virtual addresses in the
VSpaces to which to map the memory. To allocate virtual addresses I have two options:
Reserve a portion of the virtual address space and allocate ranges from that area. Or,
ask the developer to provide virtual addresses. I decided for the second option to refrain
from putting restrictions on the usage of virtual addresses.
45
Chapter 4 Design
The second task is to report the shared memory region to the compartment. When
the developer devises the virtual addresses, he can also tell its program where each
shared memory region is. That is, use the virtual addresses as identifiers for the regions.
However, when he reuses his program in another context he has to adapt his program to
the new memory layout. To increase modularity I propose to use identifiers for shared
memory connections and to report shared region descriptors via the compartment’s
bootinfo. Also I want to allow a developer to specify multiple shared memory regions
between two compartments in one connection object. I decided to represent shared
memory as a virtual device, both in the connection specification and when reporting to
compartments.
To allow a developer to specify virtual addresses for a shared memory object, I reused
the device specification data structure (see Figure 4.3). The data-part of the connection
data structure is therefore considered a pointer to a device data structure that contains
an identifier and memory regions. The regions’ memory is then allocated and mapped
into the compartments at the specified addresses. Consequently, these device specifications are also used to report the shared memory connection to the compartment.
Because of this specification method, I call shared memory connections virtual device
connections.
In this section I showed how to model connection of compartments using endpoints
and shared memory. I analysed the problem with shared endpoints and came up with a
solution. I have devised a way to specify shared memory and a way to report connections
to compartments. The next section will summarise the reporting facility.
4.2.4 Compartment bootinfo
When giving a compartment capabilities, the compartment also has to know their meanings. For that purpose, the loading process generates a static bootinfo that contains
the necessary information. This bootinfo is a data structure (see Listing 4.10) that the
loading process fills when creating the compartment. The loading process then copies
the data to a fixed address in the compartment’s VSpace.
The bootinfo first describe the capability reference ranges for untyped memory of 4kB
and 1MB sizes. Then, it describes the CSpace layout, assuming a two-level CSpace, that
is filled from 0 onwards, the first entry states the first free slot. The second entry states
the first free slot in the level 1 CNode. Of the next entries, the command line reports
the string the developer wants this virtual machine to use as kernel command line. The
next entries are self-describing, the incoming endpoint and array and count of outgoing
endpoints describe corresponding capabilities. The device count and array describe I/O
devices and connections of the compartment. Note that the device descriptions contain
copies of the devices’ PCI configuration spaces and form part of the PCI virtualisation
support.
This data structure concludes the design chapter. I have developed a secure solution
for PCI device discovery that was not handled by seL4. I have designed models of seL4
compartments including I/O devices and connections. I explained what seL4 resources
are required to set up these models and how to parameterise them. The next chapter
will show how these compartments are instantiated.
46
4.2 Architecture specification
struct bootinfo {
/∗ S t a r t / end o f s m a l l untyped o b j e c t s . ∗/
unsigned long small_ut_start ;
u n s i g n e d l o n g small_ut_end ;
/∗ S t a r t / end o f l a r g e untyped o b j e c t s . ∗/
unsigned long large_ut_start ;
u n s i g n e d l o n g large_ut_end ;
/∗ CNode l a y o u t . ∗/
unsigned long cspace_free_slot ;
unsigned long cspace_free_L1_slot ;
/∗ K e r n e l command l i n e . ∗/
c h a r command_line [ 1 2 8 ] ;
/∗ Endpoint t o r e c e i v e IRQ e v e n t s . ∗/
seL4_CPtr incomingEP ;
/∗ Endpoints t o e . g . timer −c l i e n t s , network s e r v e r ∗/
u n s i g n e d l o n g outgoing_count ;
s t r u c t seL4_outgoingEP outgoingEPs [ 8 ] ;
/∗ D ev ic e i n f o r m a t i o n . ∗/
unsigned long device_count ;
s t r u c t s eL4 _de vice d e v i c e s [ 8 ] ;
};
Listing 4.10: Compartment bootinfo data structure
47
Chapter 5
Implementation
This chapter illustrates the implementation part of this thesis. I will show individual
steps of loading an information flow architecture and substantiate implementation decisions. I start with an in-depth discussion of the loading process. Then I discuss the
other half, device access in seL4::Wombat. Afterwards, I resolve implementation issues
of PCI virtualisation and conclude with the details of the timer server.
5.1 System initialisation
In the existing system, a bootloader loads the microkernel and the initial program
into memory. The microkernel then starts the initial program. An ELF-combining tool
concatenates multiple ELF files to the initial program. The first program then reads the
combined ELF-file’s header and loads additional ELF files. I extended this mechanism
to make it instantiate a specified software architecture.
To transform the architecture specification into a running system, I first allocate
and load the compartments. Then I give each compartment the device capabilities it
needs. Afterwards, I allocate and connect communication channels. Finally, I start
the compartments. Because the loading process has the highest scheduling priority, the
compartments start when the loading process finishes.
5.1.1 The loading process
I based the loading process on an existing startup program (see Section 3.4.1). This program was used to load one instance of the paravirtualised Linux. The startup program
uses two stages. It starts off with the initial CSpace in the setup stage. There it creates a new CSpace that is usable by the Iwana object allocator. The startup program
then starts the loader-thread in this CSpace, initiating the second stage, the loading
stage. The loader first initialises Iwana and then loads the Linux ELF image. Note that
the first and the second stage use the same VSpace and therefore share code and data
structures.
Considering device support, the existing startup program already transferred the
generic hardware control capabilities into the second CSpace. However, it did not transfer the page frames of the I/O regions. When adding this functionality, the mapping
of frame-object to physical address must be preserved. First however, the loader has to
calculate this mapping from the initial CSpace layout (see Section 3.3.7).
49
Chapter 5 Implementation
In the design section I defined the list of devices of the target machine (see Section 4.2.1). The loading process uses this list to find all the I/O regions whose frame
capabilities need transferring to the second CSpace. For each device in the list, each
I/O region is processed. I use the existing implementation that, given a physical address
range, finds the list of non-contiguous capabilities ranges. I then move the frame capabilities into a contiguous region in the second CSpace. Then I update the device’s
I/O region description with the new capability range. After this procedure, the second
CSpace has all capabilities required to control the requested I/O devices.
5.1.2 Loading compartments
The loading process loads compartments by parsing the compartment-specification data
structures and allocating and assembling corresponding seL4 objects. Another data
structure is used to track the allocated objects during different phases of the loading
process.
Creating the minimal compartment
First, the Thread Control Block (TCB), IPCBuffer and incoming endpoint are allocated.
Then, the compartment’s ELF file is located in the combined ELF file and loaded into
a VSpace. The loader therefore allocates page tables and frames as necessary. Frames
are first mapped into the loader’s VSpace, filled with content and then mapped into the
compartment’s VSpace.
Building CSpaces
Afterwards, the loader creates a CSpace for the compartment. Two types of CSpaces are
currently supported. The simple CSpace consists of one CNode with 16 entries and can
be used for compartments that only have a maximum of 16 capabilities, that is, drivers
that only have endpoints to clients and to hardware control objects. This CSpace does
not contain an entry to its own CNode and hence cannot be modified. The VM CSpace
is a two-level CSpace with the same layout as the Iwana CSpace to support the Iwana
allocator. Initially, this CSpace is constructed with two 1024-entry CNodes, one for the
first level, one for the second level. The second CNode’s capability is copied into the first
CNode at slot 0 to establish valid capabilities from 0 to 1024. To give the compartment
authority over its CSpace, the first CNode’s capability is copied into the second CNode
at seL4_SelfCSpace := 2 with all permissions and the 20-bit depth information. Also,
the VSpace and TCB capability are copied into the second CNode at the standardised
slots (seL4_SelfTCB := 1, seL4_SelfVSpace := 3).
Allocating capabilities
When the CSpace is initialised, the loader can fill the CSpace with the requested capabilities. To do so, the loader has to devise capability references in the new CSpace.
Many of them were already defined by the existing program, for example slots 20 to 99
for 4k untyped memory, slots from 100 for 1MB untyped memory. While these ranges
50
5.1 System initialisation
work for standardised anonymous objects, endpoints and hardware control objects have
individual semantics. I already defined data structures to track these semantics, however, I also had to allocate capability references for them. Because these references can
be chosen freely in a free range, I implemented a simple allocator with a moving freepointer. For each CSpace type, the allocator is initialised with the range of free slots.
On every allocation, the free-pointer is moved by the number of requested capabilities.
This allocator is also used in other places, for example to track the loader’s untyped
memory objects that are allocated to compartments.
Adding optional objects
In the next step the loader copies untyped memory objects to the compartment. The
required amounts for 4k and 1MB-sized objects are specified in the compartment description. Because a compartment has to know the number of objects, it is here also that
the loader updates the compartment’s bootinfo with the ranges and amounts of copied
objects.
Then, the loader connects devices to the compartment. Therefore it loops through
the compartment’s array of assigned device identifiers and uses them to index the global
device array. It then calls a function to connect this device to this compartment. I will
describe this function in the next subsection.
Finally, the loader copies the incoming endpoint into a free address in the compartment’s CSpace. It also updates the compartment’s bootinfo. In the next step connections are handled, after device access in the next subsection.
5.1.3 Mapping devices to compartments
To allow a compartment access to a device, the four device operations must be accessible
to the compartment.
Device discovery and configuration
For identification, I copy the device specification into the compartment’s bootinfo. I
use the compartment’s device counter to index into the compartment’s device array.
Also I keep a reference to this object to update it with compartment-local capability
references.
The prototype implementation does not contain support for secure access to PCI
configuration. For uncontrolled access, the compartment can use I/O ports to configure
PCI devices.
I/O Interrupts
If the device has an interrupt, I have to give the compartment the IRQ’s capability and
an endpoint associated with this IRQ. I allocate a free slot in the compartment’s CSpace,
update the bootinfo with this slot, and copy the IRQ capability to the compartment’s
CSpace. Note that the IRQ capability is derived from the IRQ control capability using
the IRQ number. As IRQ endpoint I use the compartment’s incoming endpoint. To
51
Chapter 5 Implementation
allow the compartment to distinguish different IRQs, I associate this IRQ with a badge.
Therefore I have to create a copy with badge of the endpoint capability. The badge has
only the IRQ number’s bit set. Then I associate this badged endpoint capability with
the IRQ capability. Thereby, the IRQ is linked to the incoming endpoint.
I/O Ports
For devices that are controlled using I/O ports, the seL4_IA32_InitIOPort capability is
required in the compartment. When copying it into the CSpace at the standardised slot,
I have to make sure it is copied only once. Note also that the current seL4 implementation does not allow to compartmentalise access to individual I/O ports. Therefore, a
compartment with this capability can control many devices and the PCI bus. However,
for testing and development, it is vital to have I/O port access in compartments.
Memory-mapped I/O
When giving a compartment access to memory-mapped I/O regions, there are two
options. Either I can copy the frame capabilities into the compartment’s CSpace, or
I can map the frame capabilities into the compartment’s VSpace. The first option
removes code from the loader and gives the compartment more control over these frames.
However, a compartment now needs to access its CSpace and VSpace to map the frames
into its VSpace. Also, the compartment might need page tables to establish the mapping
and untyped memory to create the page tables. While all these requirements are fulfilled
by seL4::Wombat, this option adds requirements and complexity to simple driver servers.
The second option allows to use driver servers that contain only minimal seL4 code.
However, with this option, a VM cannot map I/O memory regions to compartments it
creates on its own. Despite this drawback I chose the second option.
To map I/O regions into the VSpace, I now have to devise virtual addresses for the I/O
regions. Because I virtualise both PCI discovery and device addresses, I can almost freely
choose an address for the I/O region. However, I decided to not change the PCI configuration space information and to keep the original physical addresses as virtual addresses.
Note that on the testing machine, those were in the range 0xf0000000 - 0xffffffff.
This same range is also occupied by the seL4 kernel that reserves this part of each
compartment’s VSpace as kernel address space (256MB). I therefore devised the range
0xe0000000 - 0xefffffff for memory mapped I/O device regions and implemented a
simple offset of -0x10000000 from physical to virtual addresses. Compartments have to
respect this offset to access I/O regions. For Linux, a central place to install this offset
exists.
For mapping I/O regions in the VSpace, an existing helper function tests whether
mapping of page tables is necessary, and if so, allocates and maps them, prior to mapping
the page. The bootinfo is not updated, because the addresses are kept. However, it is
used to determine the addresses where to map the pages to.
52
5.1 System initialisation
Data transfer - DMA
To enable a compartment to have DMA data transfers with an I/O device, the compartment’s VSpace must share frames with the device’s IOMMU space. To create the
IOMMU space, a program needs the device’s IOMMU-address-space capability, IOMMU
page tables and useful frame capabilities, that is, frames that are or can be in the compartment’s VSpace. Similar to the argument for memory-mapped I/O regions, either the
loader can set up this IOMMU address space or it can be made a task of the compartment. In this case I decided for the compartment option, because this reduces complexity
of the loader more than is the case for MMIO regions and because of the assumption that
compartments using DMA-capabilities of devices are implemented using VMs. Hence,
giving a compartment DMA control over a device, consists in creating the IOMMU
ASID from the device’s PCI address and copying this capability into the compartment’s
CSpace. Also, the compartment’s bootinfo is updated with the newly allocated IOMMU
ASID capability. For more information on IOMMU address spaces, see the next section.
After these operations, compartments have all capabilities to drive I/O devices. However, they still need glue code to use them. Also, the loader has to connect compartments
to establish an information flow architecture.
5.1.4 Establishing connections
After the loader has assembled compartments, it connects them. Therefore it loops
through the list of connections and establishes each connection. It uses the connection’s
from and to field to find the compartment’s data structures and seL4 objects. Each
connection type is handled differently.
For the endpoint connection, the loader finds the already allocated incoming endpoint
of the to compartment. Then it copies the incoming endpoint to a newly allocated slot
in the from compartment’s CSpace. Thereby it assigns a badge to the endpoint using
the IRQ number of the connection. It reports this connection as a new device to the
from compartment and as new outgoing endpoint to the from compartment.
For the shared-memory virtual-device connection, the loader parses the device description and allocates the required amount of pages for each I/O region of the device. Simultaneously, it updates the capability ranges in the I/O region. Then it connects this
device to one compartment, using the method described above. Afterwards it creates
copies of the frame capabilities and connects the device to the other compartment.
5.1.5 Starting compartments
When all generic seL4 objects are allocated and assembled, the loader will start the
applications. Prior to this, for virtual machines, it copies the command line into the
compartment’s bootinfo and copies the seL4_InitASIDPool capability. Note that sharing of the initial ASID pool also provides a covert channel. Future work should allocate
separate ASID pools for all compartment.
To start the application, the loader maps into the compartment’s VSpace, the stack,
the IPCBuffer, and the bootinfo. It assigns the VSpace and CSpace, IPCBuffer and
53
Chapter 5 Implementation
priority to the TCB. Then it sets the TCBs registers to the ELF-file’s entry point and
the stack position, and starts the compartment with this operation.
5.2 Device access in seL4::Wombat
The existing seL4::Wombat version already had paravirtualisation code that interacted
with Iguana [Hei05]. My work was to find these places and replace them with seL4
equivalents. Note that the Wombat/Iguana system was based on a request system,
where I/O-region requests from Linux drivers were forwarded and served by Iguana. In
my thesis, I/O regions are premapped. The I/O region request functions [CRKH05]
therefore only calculate the offset. Furthermore, I mapped I/O port functions to the
seL4 capability invocations.
I implemented an initial version for I/O interrupt dispatching and acknowledging.
The interrupt thread waits on the incoming endpoint and reads the badge on receipt
of an IPC. It then calls the Linux interrupt handler. Also it has to acknowledge the
interrupt by invoking the acknowledge operation on the IRQ capability. Initially, it finds
this capability in the bootinfo, later it uses an array that maps IRQ numbers to IRQ
capabilities.
Linux initiates DMA for objects on its heap. During an I/O operation, Linux will ask
the paravirtualisation code to give an I/O device access to a part of the heap. While
a lazy approach of mapping only the requested portions of the heap into the IOMMU
address space has its merits, I decided to premap the Linux heap into each IOMMU
address space Linux controls. One of the first actions of the paravirtualisation code
is the allocation of the heap, with a fixed mapping from virtual address to capability
reference. Later in the initialisation code, I added functionality to create IOMMU
spaces. Therefore I loop through all devices in the bootinfo, get the IOMMU ASID
capability, calculate from the heap size the required number of I/O page tables, and
allocate and map them into the IOMMU space. Then I create copies of the heap’s
frame capabilities and map them into the IOMMU space at the same virtual addresses.
Therefore I also had to extend the Iwana object allocator with support for allocation of
I/O page tables.
When first activating I/O device drivers in the existing seL4::Wombat version, a
strange Linux kernel page fault had to be solved. Further analysis revealed that arbitrarily data structures were overwritten and indicated a memory corruption. Digging
through the driver code while trying to determine the first occurrence of this fault
revealed that unrelated data structures had the same addresses. Eventually I could
pinpoint this to the allocation of arrays in the driver’s code that were allocated using
vmalloc() [CRKH05]. In the seL4::Wombat code, the vmalloc() area was set to the
heap region. I resolved that problem by reducing the heap region and reserving a part
as vmalloc() area. Note that the seL4::Wombat kernel cannot resolve its own page
faults.
With these provisions in place, seL4::Wombat can access I/O devices. To initialise
PCI drivers, I gave Linux access to the PCI bus at first and prevented it from using
54
5.3 Virtual PCI configuration space
other devices by a hack. Later I removed that requirement by implementing virtual PCI
discovery.
5.3 Virtual PCI configuration space
When paravirtualising PCI device discovery, a few implementation problems have to
be solved. Following the design (see Section 4.1), the paravirtualised OS is intercepted
at the pci_write_config and pci_read_config interface. Accesses specify the bus,
the device, the function, and the address in the function’s configuration space; also
the size of the access (8-bit, 16-bit, 32-bit) and a value to set or to return is specified.
On the other hand, a list of device specifications contains the copies of the devices’
configuration spaces. A translation has to map the guest OS’s access to the device list.
This translation should be efficient, because it has to be invoked on every configuration
access during device discovery.
A simplistic translation uses only the device specification list as data structure. On
every access the glue code walks through this list and checks whether the current device
entry has the requested PCI address. If so, it performs the requested operation on
the entry in the list. If no device is present, the emulation returns a value of all 1’s,
corresponding to the discovery protocol. This approach does not scale with bigger
numbers of devices but has a low memory overhead.
A simple fast translation uses lookup tables to translate accesses. It would need one
table with 256 entries for each bus, one table for each bus with 32 entries for each device
and one entry for each device with 8 entries for each function. The function entries then
point to the device specifications in the device list. To reduce the space consumption of
this device tree, paths that do not lead to devices can be pruned. The lookup process
then has to check whether an entry points to a table before it dereferences the pointer.
I used a translation with lookup tables and a pruned device tree. This tree is built in
the startup code of the paravirtualised OS. Later the paravirtualisation code redirects
the configuration space accesses through the lookup tables to the device list entries.
This way, I made the performance overhead of PCI-device-discovery paravirtualisation constant. While this feature is not performance critical, it reduces the loading
time of seL4::Wombat. I estimate that the loading time is already reduced because of
using memory instead of I/O devices as store for device information While the constant number of accesses is similar to the simplistic translation, it should scale better
with more devices. However, I did not evaluate this prediction because my focus was
implementability of the high-assurance design and not performance.
5.4 Timer server
seL4::Wombat relies on a hardware time source for scheduling and keeping time. The
existing system already had a timer server similar to the one in Iguana [Hei05]. However,
the server was not compartmentalised and relied on residing in the same CSpace as
seL4::Wombat.
55
Chapter 5 Implementation
Timer server
endpoint
Timer server
endpoint
Client
Client
a)
Client
Client
b)
Figure 5.1: Information flow of the timer server. a) The timer server must be trusted to not
transfer information, b) Isolation is independent from the timer server
To run multiple instances of seL4::Wombat, each instance requires access to a time
source. To maintain isolation between these instances, they cannot share direct access to
the timer hardware. Instead, access to the timer hardware must be compartmentalised.
That is, arbitrary numbers of individual instances must be able to program timeouts and
receive notifications independently and isolated from each other. Such a functionality
is not provided by current x86 hardware. Instead, the original timer server uses the
programmable interval timer of the chipset [Int07] that receives timeouts on IRQ 0.
The timer server then sends IPCs to endpoints that it had previously created for clients
registering a timer.
Looking at the information flow of the existing timer server, clients that program a
timeout send data to the server and clients that receive return-values or timeout-IPC’s
receive data from the server. Hence, the timer server must be trusted to not transfer
information from one client to another (see Figure 5.1, a). To alleviate this situation, I
propose a simpler timer server that does not allow programming. Instead, it is configured
to send timeout IPC’s at a preconfigured rate to a programmed set of clients. This way,
information only flows from the timer server to its clients. No information can flow via
the timer server from one client to another (see Figure 5.1, b).
To drive the timer hardware, the timer server requires several seL4 capabilities. It
needs the SEL4_IA32_InitIOPort capability to access the timer’s registers, the IRQ
capability of IRQ 0 to acknowledge the interrupt, and an endpoint capability associated
with the IRQ capability to receive interrupt IPCs. To notify its clients, the timer server
requires one endpoint for each client. These endpoints have to be asynchronous and
write-only for the timer server. Otherwise, clients can send information to the server by
selectively waiting or not waiting on the synchronous endpoint.
The required seL4 capabilities are available if the timer server is specified as seen in
Listing 5.1 in the architecture specification.
When started, the timer server reads out the capability references and the number of
clients. It then programs the timer hardware and waits for timer interrupts. On every
interrupt it sends IPC’s to each of its client endpoints.
56
5.4 Timer server
SYSTEM_COMPONENTS[COMPONENT_TIMER] = ( component_t ) {
. name = " timer −s e r v e r " ,
. d e v i c e s = {HARDWARE_TIMER, } ,
. binary_name = " timer −s e r v e r . e l f " ,
. CSpace_depth = CLIENT_CSPACE_BITS,
};
SYSTEM_CONNECTIONS[ 0 ] = ( c o n n e c t i o n _ t ) {
. from = COMPONENT_TIMER,
. t o = COMPONENT_WOMBAT_1,
. type = INCOMING_EP,
{ . i r q =0} ,
. f r o m _ r i g h t s = seL4_CanWrite ,
. to_rights = seL4_AllRights } ;
Listing 5.1: Specification of timer compartment
With the timer server I showed how to compartmentalise access to a simple platform
device. This approach does not require verification of the timer server and relies solely
on the correctness of the compartmentalisation mechanisms. However, the functionality
of the timer server is reduced. Future work has to determine whether this type of time
source is sufficient.
This chapter showed individual steps of loading an information flow architecture. I
covered details of the loading process, device access in seL4::Wombat, implementation
issues of PCI virtualisation, and the timer server. The following chapter will evaluate
the achievement of my goals.
57
Chapter 6
Evaluation and Future Work
In this chapter I will revisit my goals and evaluate how I achieved them. Then I show
limitations of my approach and give an outlook to future work.
6.1 Evaluation
Security
The task for this thesis was to compartmentalise an x86 system using the seL4 microkernel. The goal was to achieve the strongest compartmentalisation possible by design.
To evaluate this goal I will summarise my approach here. I started by defining how a
compartmentalised system should look like. Then I showed how the x86 architecture
supports compartmentalisation and how an operating system can use these mechanisms.
Afterwards I introduced the seL4 microkernel and showed how it uses the x86 mechanisms. I also highlighted the parts that still needed compartmentalisation. For one of
them, device discovery, I developed an approach that has a low runtime-code complexity.
Then I showed how to achieve a compartmentalised system by assembling compartments
only of the required seL4 resources and I/O resources. Compartments as developed in
the design chapter are isolated except for explicitly specified communication channels.
In contrast to other systems, compartments cannot use resource requests, I/O devices,
the timer server, or the PCI bus to break the compartmentalisation. While I am confident about these claims, I cannot prove them. Future work to develop a high assurance
system has to analyse the correctness of my work.
In addition to the assumption that my code is correct, I make assumption about the
BIOS. For the correctness of device discovery, I assume the BIOS creates a compartmentalised PCI device configuration. I also assume this configuration does not change
across reboots.
Complexity
The trusted computing base of my compartmentalisation solution encompasses the seL4
microkernel and the loading program. To give an impression of the complexity of these
components I will use the source lines of code (SLOC) count.
The seL4 microkernel has about 8700 SLOC. The original loading program had a
1916 SLOC count. It is dependent on the Iwana library that adds about 1500 SLOC.
From the original loading program I removed the timer server that has now 364 SLOC.
59
Chapter 6 Evaluation and Future Work
This leaves 1552 SLOC for loading VSpaces and building the Iwana CSpace. My loading program now has 2260 SLOC. So I added 708 SLOC to support I/O devices and
compartmentalisation.
Because of my decision to use static device discovery I did not require PCI device
discovery code in the trusted computing base. This decision removed otherwise necessary
700 SLOC of similar projects. However, I only worked on a small variety of devices.
Other devices and features will make this code necessary.
When adding virtualisation support for PCI discovery I chose to emulate PCI configuration space inside the virtual machine. The resulting code only required 65 new
SLOC in the paravirtualised Linux. Hence this extension is small and easy to maintain. Because I am intercepting hardware accesses, I do not even touch Linux software
infrastructure.
Performance
I did not develop a new component for a well evaluated system. Instead, my thesis
adds a fundamental component for the new field of compartmentalised systems on seL4.
Performance of these systems is expected to be in the range of previous microkernelbased systems. I did not conduct measurements to validate this assumption, because I
expect the performance to be dominated by the seL4 microkernel’s design decisions.
During testing however, the paravirtualised Linux on a Core 2 Duo 2.33 GHz with
an Intel PCIe Gigabit Ethernet card was reactive via a serial console and via an ssh
connection.
Functionality
A goal of this thesis was to leverage the properties of the seL4 microkernel that employs
new approaches to kernel resource management. This feature removes the necessity for
a memory management server. To leverage this property, the initial seL4 program has to
split and distribute kernel memory between compartments. Different approaches to solve
this memory allocation problem in seL4 have not resulted in a usable system. Only the
Iwana memory allocation library was sufficiently sophisticated and developed to provide
a basis for this thesis. Choosing Iwana allowed me to concentrate on instantiating
compartmentalised systems.
Extensibility
One main goal of this thesis was to systematically support compartmentalised applications on the seL4 kernel, including compartmentalising access to hardware devices. To
achieve this goal, I developed the loading process that can be parameterised with the
specification of a static compartmentalised application. It also includes compartmentalisation support for hardware devices that are connected via I/O ports, memory-mapped
I/O regions, interrupts, and DMA. That is, legacy x86 devices, PCI devices, and ACPIreported devices.
To demonstrate the generality of this approach I will present two subsequent works
that were based on my work. In his honours project, Robert Sison [Sis09] developed a
60
6.1 Evaluation
NIC a
seL4::Wombat
Firewall
server
seL4::Wombat
NIC b
Timer server
Figure 6.1: Architecture of secure firewall
VGA
Keyboard
Timer
seL4::Wombat
with a shell
VGA
server
Timer server
Figure 6.2: Architecture of SOSP demo
secure firewall that contains two Linux instances each connected to one network card and
to each other via a trusted firewall compartment (see Figure 6.1). His work made use of
the hardware compartmentalisation functionality and the endpoint and shared memory
connections. Also the timer server was used to implement the timeout of network
connections in the firewall. To instantiate his architecture he only had to specify it
using the data structures developed in my thesis.
The ERTOS group presented the seL4 work at SOSP and included a demo. This demo
ran a Linux instance next to a VGA server on a notebook (see Figure 6.2). Therefore
it was required to specify the x86 keyboard and VGA devices. This was possible with
minor additions and compartments were then able to use these devices. Note that
Linux and the VGA server were sharing the VGA device, however, the VGA server was
unaffected by restarts of Linux. Additionally, the VGA server also used the timer server
to show time elapsed since bootup. Instead of producing an implementation only valid
for this specific demo, the demo resulted in a reusable toy VGA server and evaluated
the requirements for a secure terminal (VGA and keyboard access).
61
Chapter 6 Evaluation and Future Work
6.2 Limitations
Connections
In the design chapter I derived ways to connect seL4 compartments and a method to
specify these connections. When analysing the design spectrum I only found three useful
and secure connection types. However, I only implemented two of them. Synchronous
endpoints cannot be used in the current version.
Considering the abstraction of connections my solution has two drawbacks. seL4 can
connect compartments in a finite number of ways, using a shared object and the capabilities that connect the object to the CSpaces of the compartments. I modelled connections
as links between two compartments. A more generic model would model connections
as connections, thus include as separate objects, shared objects and mappings of shared
objects to compartments. This model allows to specify connections between multiple
compartments and multiple endpoints per compartment. In the design chapter I showed
that the resulting connections conflict with compartmentalisation or do not lend themselves to a single-threaded server scheme. However, for mutually trusting compartments
a call-reply connection or globally shared memory reduces complexity. Also, a compartment does not have to wait on all its endpoints at the same time or can use multiple
threads to do so.
To specify shared memory connections I use the virtual device abstraction and require
the developer to provide virtual addresses. In hindsight I think these two decisions were
wrong. Component developers often need not care about virtual memory layout but have
to with my solution. Also, virtual addresses in the architecture specification introduce a
dependency between the architecture and the component that reduces modularity. I also
think that developers do not need multiple shared memory regions between two compartments. A simpler specification could therefore only state the size and an identifier
for a shared memory region.
Limitations of the approach
Static system
The current design trades adaptability and online reconfiguration for simplicity in an
attempt to increase trustworthiness. The analysed system is entirely static in what it
expects from the hardware and in what it creates in terms of software components.
Once started, changing the architecture is not possible.
Static device discovery
The current system uses offline device discovery. When that information is assembled
into the loader binary, the loader can only run on machines closely resembling the
specified machine. For faster development, it is useful to have only one binary execute
on different machines. That can be achieved by adding device discovery code to the
loading process. That code adds to the trusted computing base and the overall size of
62
6.3 Future Work
the trusted code. However, its functionality is fairly isolated and does not interact with
other components.
Information flow architectures
This thesis is limited to static information flow architectures. Only information can
flow between components, not resources. Therefore it is only useful for single purpose
machines that do not change their functionality at runtime.
6.3 Future Work
In this section I show problem areas that involve my work and have to be solved to
deliver a secure and versatile high-assurance system.
Verification
Several steps are necessary to analyse the correctness of my work. The goal of this
analysis would be, to show that my code indeed transforms a specified information flow
architecture into a compartmentalised x86 system. Therefore, a formal model has to
describe the I/O channels and I/O devices of the x86 architecture and the PCI bus.
This model also has to contain the definition of compartment and it has to be decidable
whether a given hardware configuration is compartmentalised, that is, divided into
isolated partitions. A next model has to show how seL4 interacts with the I/O hardware
model and how the seL4 API controls this interaction. My program instantiates an
architecture by performing seL4 system calls. Given an architecture specification and
a model of my program, these system calls can be calculated. Applying the sequence
of systems calls to the seL4 API model will create a sequence of transitions in the seL4
kernel model, which will in turn create a sequence of transitions in the hardware model.
Starting from an initial state in the hardware model, this sequence will create a new
hardware configuration. Using the compartmentalisation test, we could now show that
we created a compartmentalised hardware state.
Dynamic information flow architectures
This thesis is based on static information flow architectures. This allows to build systems
that only contain the kernel and untrusted run-time components. Future work should
also investigate using trusted run-time components to enable dynamic information flow
architectures.
In such systems, untrusted compartments would still only have explicit data-transfer
connections to other compartments. Thus, the same class of secure information flow
graphs can be instantiated. However, that loading process does not have to run at
startup time only. Instead, the system specification should be modifiable at runtime
and incur a change of the current compartment configuration. That way an authorised
user of the system can set up a new use case.
63
Chapter 6 Evaluation and Future Work
Because a dynamic system will deviate from its initial state, a full specification at
development time is impossible. Instead, a new component, the compartment manager,
has to maintain the system configuration. It also has to accept configuration changes
and organise transitions between configurations. The compartment manager has to
bookkeep resource information for compartments to destroy them and to reuse the
resources. To set up compartments, it also requires compartment descriptions similar
to the ones defined in this thesis. Otherwise it could not create isolated compartments
that do not request resources.
I/O compartmentalisation
A further challenge is the compartmentalisation of more I/O devices. In general, this
will incur device servers that compartmentalise I/O devices for compartments.
For the simple timer device, I showed a compartmentalisation solution that does not
require verification. This solution was based on unidirectional information flow. Future
work should investigate whether a similar mechanism is possible for other I/O devices
that, conceptually, are input-only or output-only. That is, keyboard, mouse, sensors for
input-only devices, sound, graphics for output-only devices.
Network and storage devices are conceptually bidirectional and require a more
advanced compartmentalisation solution. One compartment would directly control the
device and provide virtual devices or a similar service to client-compartments. The
device server has to ensure that no information can flow between its clients. Therefore
I propose to not allow untrusted compartments to create, modify, delete, or list virtual
devices. Instead, these operations should be reserved for trusted components.
My thesis analysed I/O compartmentalisation for the PCI bus. The current design
only supports I/O devices that are statically connected via I/O ports, MMIO, DMA,
and interrupts. Modern device classes exist that use more abstract, software-like interfaces. Examples are USB devices, Intel network cards implementing VT-c technology
and graphic cards supporting Microsoft DirectX 10. These devices provide compartmentalisation for I/O channels similar to the device servers I just described and can
result in faster and less complex I/O compartmentalisation.
In this chapter I evaluated the achievement of my goals, showed limitations of my
approach and presented my ideas for future work. The next chapter will conclude.
64
Chapter 7
Conclusions
In my thesis I developed a key component of future seL4 systems. This component is
responsible to instantiate a compartmentalised software architecture on the seL4 microkernel. I showed what architectures secure high assurance systems require and how
microkernel-based architectures can be described in component systems.
One requirement of security architectures is the compartmentalisation of the underlying hardware. I showed how both the x86 architecture and the seL4 microkernel support
compartmentalisation of the system. I further showed how to handle the resources both
mechanisms provide. I therefore developed a description of seL4 compartments including
I/O devices and explicit compartment connections. I implemented a loader component
that interprets these descriptions and instantiates the specified compartments.
Of the four operations of I/O devices, x86 and seL4 already compartmentalise device
control, data transfer and I/O interrupts. In this thesis I proposed to make device
discovery static to reduce the interdependability of system components. I demonstrated
that this idea can be implemented for the operating system and for compartments. For
compartmentalised operating systems I showed a mapping from the OS’s native device
discovery to the static device list. For static device discovery I developed specifications
for I/O devices that comprise the four device operations and describe the x86 I/O
channels as well as the corresponding seL4 capabilities.
I showed that seL4 compartments have five classes of requirements. I showed how
minimal compartments look like and argued that they are dominated by their address
space. Then I showed that for a compartment to access a device the compartment must
have the capabilities referenced in the device specification. Virtual machine compartments manage subcompartments. Therefore they need access to their capability space,
address space, and to seL4 kernel memory. I analysed compartment connections in seL4
and devised secure connections based on message passing and shared memory. Finally,
I developed a reporting facility to tell a compartment its capabilities.
I evaluated my goals and concluded that my work can be used as a central component
in a seL4 system. I also argued that my design decisions removed many attack vectors
of common operating systems. However, I can only hope that I did not miss other vulnerabilities. Therefore I outlined how future work could formalise the x86 architecture
and proof correctness of my results.
65
Glossary
ACPI advanced configuration and power management interface
API application programming interface
ASID address space identifier
BAR base address register (PCI)
BIOS Basic Input Output System
CNode capability table (seL4)
CPU central processing unit
CSpace capability space (seL4)
DMA direct memory access
DRAM dynamic random access memory
ELF executable and linking format
IDL interface description language
IO input–output
IOMMU input–output memory management unit
IPC inter process communication
IRQ interrupt request
MILS multiple independent levels of security and safety
MMIO memory-mapped I/O
MMU memory management unit
OS operating system
PCI peripheral component interconnect
seL4 secure embedded L4 microkernel
TCB thread control block
VSpace (virtual) address space (seL4)
67
Bibliography
[AFOTH06] Jim Alves-Foss, Paul W. Oman, Carol Taylor, and Scott Harrison. The
MILS architecture for high-assurance embedded systems. International
Journal on Embedded Systems, 2:239–247, 2006.
[AJM+ 06]
D. Abramson, J. Jackson, S. Muthrasanallur, G. Neiger, G. Regnier,
R. Sankaran, I. Schoinas, R. Uhlig, B. Vembu, and J. Wiegert. Intel virtualization technology for directed I/O. Intel Technology Journal, 10(3):179–
192, 2006.
[APJ+ 01]
Mohit Aron, Yoonho Park, Trent Jaeger, Jochen Liedtke, Kevin Elphinstone, and Luke Deller. The SawMill framework for VM diversity. In
Proceedings of the 6th Australasian Computer Systems Architecture Conference, Gold Coast, Australia, January 2001. IEEE CS Press.
[Bel05]
Fabrice Bellard. QEMU, a fast and portable dynamic translator. In Proceedings of the 2005 USENIX Annual Technical Conference, FREENIX
Track, pages 41–46, Anaheim, CA, April 2005.
[Bon94]
Jeff Bonwick. The Slab Allocator: An Object-Caching Kernel Memory
Allocator. In USENIX Technical Conference, Boston, MA, USA, Winter
1994.
[CRKH05]
Jonathan Corbet, Alessandro Rubini, and Greg Kroah-Hartman. Linux
Device Drivers. O’Reilly & Associates, Inc, 3rd edition, 2005.
[EDE08]
Dhammika Elkaduwe, Philip Derrin, and Kevin Elphinstone. Kernel design
for isolation and assurance of physical memory. In 1st Workshop on Isolation and Integration in Embedded Systems, pages 35–40, Glasgow, UK,
April 2008. ACM SIGOPS.
[Elk09]
Dhammika Elkaduwe. A Principled Approach To Kernel Memory Management. PhD thesis, School of Computer Science and Engineering, University
of NSW, Sydney 2052, Australia, February 2009.
[Elp04]
Kevin Elphinstone. Future directions in the evolution of the L4 microkernel. In Gerwin Klein, editor, Proceedings of the NICTA workshop on OS
verification 2004, Technical Report 0401005T-1, Sydney, Australia, October 2004. National ICT Australia.
[FAH+ 06]
Manuel Fähndrich, Mark Aiken, Chris Hawblitzel, Orion Hodson, Galen C.
Hunt, James R. Larus, and Steven Levi. Language support for fast and
69
Bibliography
reliable message-based communication in Singularity OS. In Proceedings of
the 1st EuroSys Conference, pages 177–190, Leuven, Belgium, April 2006.
[FH06]
Norman Feske and Christian Helmuth. Design of the Bastei OS architecture, subsequently called Genode OS Framework. Technical report, TUDFI06-07-Dezember-2006, TU Dresden, 2006. 6, 74, 2006.
[FHN+ 04]
Keir Fraser, Steven Hand, Rolf Neugebauer, Ian Pratt, Andrew Warfield,
and Mark Williamson. Safe hardware access with the Xen virtual machine
monitor. In Proceedings of the 1st Workshop on Operating System and
Architectural Support for the On-Demand IT Infrastructure, 2004.
[GJP+ 00]
Alain Gefflaut, Trent Jaeger, Yoonho Park, Jochen Liedtke, Kevin J.
Elphinstone, Volkmar Uhlig, Jonathon E. Tidswell, Luke Deller, and Lars
Reuther. The Sawmill multiserver approach. In Proceedings of the 9th
SIGOPS European Workshop, pages 109–114, Kolding, Denmark, 2000.
ACM Press.
[Hei05]
G. Heiser. Iguana User Manual. National ICT Australia, 2005.
[Hel01]
Christian Helmuth. Generische Portierung von Linux-Gerätetreibern
auf die DROPS Architektur.
Master’s thesis, TU Dresden, Germany, July 2001. URL http://os.inf.tu-dresden.de/papers_ps/
helmuth-diplom.pdf.
[Hew06]
Hewlett-Packard Corporation, Intel Corporation, Microsoft Corporation,
Phoenix Technologies Ltd., Toshiba Corporation. Advanced Configuration
and Power Interface Specification, 3.0b edition, October 2006. URL http:
//acpi.info.
[HHF+ 05]
Hermann Härtig, Michael Hohmuth, Norman Feske, Christian Helmuth,
Adam Lackorzynski, Frank Mehnert, and Michael Peter. The Nizza securesystem architecture. In colcom05, San Jose, CA, USA, December 2005.
[HHL+ 97]
Hermann Härtig, Michael Hohmuth, Jochen Liedtke, Sebastian Schönberg,
and Jean Wolter. The performance of µ-kernel-based systems. In Proceedings of the 16th ACM Symposium on Operating Systems Principles, pages
66–77, St. Malo, France, October 1997.
[HLM+ 03]
Hermann Härtig, Jork Löser, Frank Mehnert, Lars Reuther, Martin
Pohlack, and Alexander Warg. An I/O architecture for microkernel-based
operating systems. Technical Report TUD-FI03-08, TU Dresden, July
2003.
[HWL96]
Hermann Härtig, Jean Wolter, and Jochen Liedtke. Flexible-sized pageobjects. In Proceedings of the 5th IEEE International Workshop on Object
Orientation in Operating Systems, page 102, Seattle, WA, USA, October
1996.
70
Bibliography
[Hän07]
Lukas Hänel. ACPI for L4Env. Student thesis, TU Dresden, Germany, June 2007. URL http://os.inf.tu-dresden.de/papers_ps/
haenel-beleg.pdf.
[Int07]
Intel Corp. Intel I/O Controller Hub 8 (ICH8) Family, 2007. URL http:
//www.intel.com/Assets/PDF/datasheet/313056.pdf.
[Int08a]
Intel Corp. Intel 64 and IA-32 Architectures Software Developer’s Manuals, 2008. URL http://www.intel.com/products/processor/manuals/
index.htm.
[Int08b]
Intel Corp.
Intel Virtualization Technology for Directed I/O,
2008.
URL http://download.intel.com/technology/computing/
vptech/Intel(r)_VT_for_Direct_IO.pdf.
[KEH+ 09]
Gerwin Klein, Kevin Elphinstone, Gernot Heiser, June Andronick, David
Cock, Philip Derrin, Dhammika Elkaduwe, Kai Engelhardt, Rafal Kolanski, Michael Norrish, Thomas Sewell, Harvey Tuch, and Simon Winwood.
seL4: Formal verification of an OS kernel. In Proceedings of the 22nd ACM
Symposium on Operating Systems Principles, pages 207–220, Big Sky, MT,
USA, October 2009. ACM.
[KLGH07]
Ihor Kuz, Yan Liu, Ian Gorton, and Gernot Heiser. CAmkES: A component
model for secure microkernel-based embedded systems. Journal of Systems
and Software Special Edition on Component-Based Software Engineering
of Trustworthy Embedded Systems, 80(5):687–699, May 2007.
[Lam73]
Butler W. Lampson. A note on the confinement problem. Communications
of the ACM, 16:613–615, 1973.
[LCFD+ 05] Ben Leslie, Peter Chubb, Nicholas Fitzroy-Dale, Stefan Götz, Charles Gray,
Luke Macpherson, Daniel Potts, Yueting (Rita) Shen, Kevin Elphinstone,
and Gernot Heiser. User-level device drivers: Achieved performance. Journal of Computer Science and Technology, 20(5):654–664, September 2005.
[Lie94]
Jochen Liedtke. Page table structures for fine-grain virtual memory. IEEE
Technical Committee on Computer Architecture Newsletter, 1994.
[Lie96]
Jochen Liedtke. L4 Reference Manual — 486/Pentium/PentiumPro, Version 2.0. GMD, Schloß Birlighofen, Germany, September 1996. Working
Paper 1021.
[LvSH05]
Ben Leslie, Carl van Schaik, and Gernot Heiser. Wombat: A portable
user-mode Linux for embedded systems. In Proceedings of the 6th
Linux.Conf.Au, Canberra, April 2005.
[LW09]
Adam Lackorzynski and Alexander Warg. Taming subsystems: capabilities
as universal resource access control in L4. In 2nd Workshop on Isolation
and Integration in Embedded Systems, Nuremburg, Germany, 2009. ACM.
71
Bibliography
[NIC]
NICTA.
camkes.
[NIC06]
National ICT Australia. seL4 Reference Manual, 2006. URL http://www.
ertos.nicta.com.au/research/sel4/sel4-refman.pdf.
[NSL+ 06]
Gil Neiger, Amy Santoni, Felix Leung, Dion Rodgers, and Rich Uhlig.
Intel virtualization technology: Hardware support for efficient processor
virtualization. Intel Technology Journal, 10(3), August 2006.
[Pal09]
Ameya Palande. Capability-based secure DMA in seL4. Masters thesis,
Vrije Universiteit, Amsterdam, January 2009.
[PG74]
Gerald J. Popek and Robert P. Goldberg. Formal requirements for virtualizable third generation architectures. Communications of the ACM,
17(7):413–421, 1974.
[SA99]
Tom Shanley and Don Anderson. PCI System Architecture. Mindshare,
Inc., 1999.
[San93]
Ravi S. Sandhu. Lattice-based access control models. IEEE Computer,
26(11):9–19, 1993.
[Sis09]
Robert Sison. Secure Web Server on seL4. BE thesis, School of Computer Science and Engineering, University of NSW, Sydney 2052, Australia,
October 2009.
[Zeh06]
Myrto Zehnder. Virtualisation of PCI devices in LinuxOnLinux. Masters thesis, Department of Computer Science, Swiss Federal Institute of
Technology, Zurich, Switzerland, August 2006.
72
CAmkES website.
http://ertos.nicta.com.au/software/
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement