genode/doc/challenges.txt
2015-09-09 15:14:29 +02:00

493 lines
26 KiB
Plaintext

=======================================
Future Challenges of the Genode project
=======================================
Abstract
########
This document compiles various ideas to pursue in the context of Genode. It is
meant as source of inspiration for individuals who are interested in getting
involved with the project and for students who want to base their student
research projects on Genode.
Applications and library infrastructure
#######################################
:Chrome web browser:
The Chrome web browser promises to address the most pressing security
concerns of web application by isolation measures, in particular the
sandboxing of plugins and the confinement of individual web applications. As
we demonstrated with the Genode Live CD 10.11, Genode facilitates a more
natural way to pursue such techniques compared with current commodity
operating systems. Furthermore, the use of Genode as base platform for Chrome
would strengthen the web-browser security by dwarfing its trusted computing
base by two orders of magnitude compared to the use of Linux as base
platform. This would allow Chrome to be considered as a secure interface to
the web for use cases in the high-assurance domain.
:Qemu or Virtual Box on Genode:
Using Genode as hosting platform for virtual machines running in Qemu or
Virtual Box would enable the execution of security-sensitive functions (for
example cryptography) with a minimally-complex trusted computing base
beside running Windows on the same machine.
:VNC server implementing Genode's framebuffer session interface:
With 'Input' and 'Framebuffer', Genode provides two low-level interfaces
used by interactive applications. For example, the Nitpicker GUI server uses
these interfaces as a client and, in turn, exports multiple virtual
'Framebuffer' and 'Input' interfaces to its clients. This enables a
highly modular use of applications such as the nesting of GUIs. By
implementing the 'Framebuffer' and 'Input' interfaces with a VNC server
implementation, all graphical workloads of Genode would become available over
the network. One immediate application of this implementation is the remote
testing of graphical Genode applications running on a headless server.
:Tiled window manager:
At Genode Labs, we pursue the goal to shape Genode into an general-purpose
operating system suitable for productive work. The feature set needed to
achieve this goal largely depends on the tools and applications daily used by
the Genode engineers. As one particularly important tool for being highly
productive, we identified a tiled user interface. Currently, all developers
at Genode Labs embrace either the Ion3 window manager or the tiled Terminator
terminal emulator. Hence, we desire to have a similar mode of user
interaction on Genode as well. The goal of this challenge is to identify the
most important usage patters and the implementation of a tiled GUI that
multiplexes the framebuffer into a set of tiled and tabbed virtual
framebuffers.
Related to this work, the low-level 'Framebuffer' and 'Input' interfaces
should be subject to a revision, for example for enabling the flexible change
of framebuffer sizes as needed by a tiled user interface.
:Interactive sound switchbox based on Genode's Audio_out session interface:
Since version 10.05, Genode features a highly flexible configuration concept
that allows the arbitrary routing of session requests throughout the
hierarchic process structure. Even though primarily designed for expressing
mandatory-access control rules, the concept scales far beyond this use case.
For example, it can be used to run an arbitrary number of processes
implementing the same interface and connecting the different interface
implementations. One special case of this scenario is a chain of audio
filters with each using the 'Audio_out' session interface for both roles
client and server. Combined with the Nitpicker GUI server and Genode's
support for real-time priorities, this base techniques enable the creation of
flexible audio mixer / switchboard applications, which require dedicated
frameworks (e.g., Jack audio) on traditional operating systems. The goal of
this project is to create a show case implementation demonstrating the
feasibility for creating high-quality audio applications on Genode.
Furthermore, we wish for feedback regarding the current design of our bulk
streaming interface when used for low-latency applications.
:PDF reader for E-Government use:
A facility for reading PDF and E-Book documents is one of the indispensable
features Genode has to provide to be considered for general-purpose
computing. The goal of this work is to identify a suitable open-source PDF
engine and port it as native application to Genode. The challenging part is
to keep the complexity of this application as low as possible in order to
enable the use of this application as a trusted document reader. Further
ideas envision the use of PDF files as medium for sensitive documents
combined with measures for protecting the integrity of the displayed
information. For example, when processing contracts or similar sensitive
documents in E-Government scenarios, the consumer of such documents expects
the correct display of all the information as expressed by the creator of the
document. In the event of a compromised PDF engine or a man-in-the middle
attacker manipulating the PDF file, the consumer of the document requires a
way to identify such security breaches. In this context, running the PDF
engine in a sandboxed Genode subsystem has two incentives. First, the attack
surface for manipulating the PDF engine gets dramatically reduced, and
second, the integrity of the result of the PDF engine can be measured by an
independent trusted component facilitating Genode secure GUI server
(Nitpicker).
:Graphical on-target IPC tracing tool using Qt:
Analysing the interaction of components of a multi-server operating system
such as Genode is important to discover bottlenecks of the system and for
debugging highly complex usage scenarios involving many processes. Currently,
Genode handles this problem with two approaches. First, Genode's
recursive structure enables the integration of a subsystem in a basic
OS setup featuring only those drivers and components used for the particular
subsystem. After the successful integration of such a subsystem, it can
be embedded into a far more complex application scenario without any changes.
With this approach, the subject to analyse can be kept at a reasonable level
at integration time. For debugging purposes, the current approach is using
the debugging facilities of the respective base platforms (e.g., using
GDB on Linux, the Fiasco kernel debugger, the OKL4 kernel debugger).
However, in many cases, bottlenecks do not occur when integrating individual
sub systems but after integrating multiple of such subsystems into a large
application scenario. For such scenarios, existing debugging methodologies do
not scale. A tool is desired that is able to capture the relationships
between processes of a potentially large process hierarchy, to display
communication and control flows between those processes, and to visualize the
interaction of threads with the kernel's scheduler.
Since Qt is available natively on Genode, the creation of both offline and
on-target analysis tools has become feasible. The first step of this project
is creating an interactive on-target tool, that displays the interaction
of communicating threads as captured on the running system. The tool should
work on a selected kernel that provides a facility for tracing IPC messages.
Application frameworks
######################
:Running the Meego application stack on Genode using Qt:
With Genode 11.02, Qt has become available. The most prominent feature
of this version is the new QML language to design GUIs using a declarative
language. This technique is targeted specifically to mobile applications and
other touch-based devices. The goal of this project is to run the Meego
application stack natively on Genode. First, the software components and
Meego-specific Linux customizations must be identified. For each such
component, it must be decided whether to port its code or to reimplement its
interface. The immediate goal of the first step is running one Meego example
application natively on Genode.
:Python Qt bindings:
With the Python interpreter and the port of the Qt framework, the principle
components for Python-based GUIs on Genode are available. However, the glue
between both components is missing. The incentive of this work is supplementing
our Python port with the modules needed for real applications and porting the
Qt bindings to Genode. This would bring Genode one step closer to executing
modern Python-based GUI applications (in particular KDE4 applications).
:Evaluation of porting GTK+ to Genode:
With Qt, we have demonstrated the feasibility to run a highly-complex
application framework via Genode on a wide range of microkernels. That leaves
the question of looking into the other major toolkit in town, namely GTK+ as
used by Firefox and the Gnome desktop.
:Cairographics:
Cairo is a high-quality 2D vector graphics engine used by a large number of
open-source projects, in particular GTK+. Hence the port of Cairo is a
prerequisite for the GTK+ challenge. In addition, it would enable the
use of further libraries such as Poppler.
Device drivers
##############
:Enhancing Gallium3D support:
Genode 10.08 introduced Gallium3D including the GPU driver for Intel GMA
CPUs. With this initial version, we demonstrated that the powerful software
stack for running hardware-accelerated 3D applications can be deployed on
Genode. At the same time, it motivates us to reach out for even more
ambitious goals:
First, the current approach executes the GPU driver alongside the complete
Gallium3D software stack and the application code in one address space. To
enable the use of multiple hardware-accelerated applications running at the
same time, the GPU driver must run separated from the Gallium3D code as done
on Linux. The preliminary interfaces for this decomposition are already in
place but there are several open questions. Most importantly, the page-fault
handling of buffer objects mapped in the application's address space.
Second, we'd like to complement our current Intel GMA GPU driver with
interrupt-based synchronization, namely vblank handling. This requires an
understanding of the Intel GMA interrupt code and the enhancement of our
driver environment.
Third, we desire the use of further Gallium3D drivers, in particular the
Nouveau and r300 drivers. The basic approach to bring these drivers to Genode
is the same as for Intel GMA but the respective driver environments are yet
to be developed.
If you are interested in low-level graphics hacking, GPUs, and
high-performance graphics, this project is ideal to get you on track.
:Split USB core from USB device drivers:
Genode's current USB support is based on the Linux USB stack running as a
single process on Genode. This process includes the USB core logic, USB host
controller driver as well as the USB device drivers such as HID or USB
storage. This monolithic USB process is rather inflexible. Hence, we desire a
decomposition of this solution such that the USB host driver and each USB
device driver runs in a separate process.
:IOMMU support on the NOVA Hypervisor:
The NOVA hypervisor is the first open-source microkernel with thorough
support for IOMMUs, which principally enables the use of untrusted device
drivers alongside sensitive software on one machine. Without an IOMMU, each
device driver for a device that operates with DMA, is able to indirectly
access the whole physical memory through programming the device. With IOMMU,
the physical memory addressable by DMA operations can be restrained per
device. The goal of this challenge is to enhance Genode with I/O protection
when running on the NOVA kernel. This would clear the way towards reusing
complex untrusted device drivers running in dedicated device-driver OS
instances.
:I/O Kit:
I/O Kit is the device-driver framework as used by the Darwin operating
system, which forms the basis for Mac OS X. The port of I/O Kit would enable
the easy re-use of the library of I/O-Kit-based device drivers on Genode. As
foundation of this project, we recommend to use the DDE Kit API featured by
Genode.
:Support for multi-touch input devices:
The efforts towards enabling mobile application stacks such as Meego and
Android on Genode must be accompanied by a revision of Genode's 'Input'
session interface to accommodate multi-touch input devices. First, existing
APIs such as multi-touch support in X11, Qt, and Android should be analysed.
Based on these findings, we expect a proposal for changing Genode's input
interface. The interface extension should be validated by a example driver
implementing the interface as well as an example applications.
System services
###############
:Copy-on-write memory manager:
Genode's managed dataspaces provide a generalized page-table concept,
enabling servers to provide on-demand paged memory objects (dataspaces) to
clients. This concept is showcased by the ISO9660 driver, which provides
on-demand paged ROM dataspaces to its clients. Depending on the access
pattern of the client, the ISO9660 server loads the used parts of the ROM
file from CDROM. Managed dataspaces principally allow for a wide variety of
interesting applications such as the transparent migration of non-local and
local memory in a NUMA system, sparse dataspaces, swapping, and copy-on-write
dataspaces. The goal of this project is a dataspace manager that implements
copy-on-write semantics combined with a merging technique optimizing the
memory footprint at runtime. Pages of two managed dataspaces that share the
same content should be provided via read-only page sharing. If one client
attempts to change the content of a shared page, a new physical copy of the
page get created. Vice versa, if the content of different pages converge, the
sharing should be re-established. This work is a follow-up of the diploma
thesis of Sebastian Sumpf
[http://os.inf.tu-dresden.de/papers_ps/sumpf-diplom.pdf - Cloning L4Linux].
On the course of this project, the managed dataspace concept of Genode
will be refined, in particular regarding the creation of read-only
dataspaces from read-write dataspaces.
:Using Haskell as systems-development language:
The goal of this project is the application of functional programming
i.e., Haskell, for the implementation of low-level Genode components.
Implementing critical functionalities in such a high-level language instead
of a classical systems language such as C or C++ would pave the way towards
analyzing such components with formal methods.
The use of Haskell for systems development was pioneered by the
[http://programatica.cs.pdx.edu/House/ - House Project]. A more recent
development is [http://halvm.org - HalVM] - a light-weight OS runtime for
Xen that is based on Haskell.
:Dbus emulation:
Dbus is a popular inter-process communication mechanism on Linux, which
enables user applications to respond to global system events and announce
state changes to other applications. It is extensively used by modern desktop
environments. To enable such applications to integrate well with Genode, a
Dbus emulation solution has to be developed.
:Wayland:
With the availability of Gallium3D on Genode, the prospect for incorporating
further projects of the Linux graphics ecosystem into Genode arises.
[http://wayland.freedesktop.org - Wayland] is a window server especially
designed to be used with Gallium3D. Its design has many similarities with
Genode's Nitpicker GUI server, in particular the decision to move window
handling policies to the client and thereby minimize the complexity of the
GUI server. Whereas Nitpicker was designed for high security, Wayland is
targeted to creating GUIs with fluid and tearless animations using
hardware-accelerated graphics. We believe that because of the many conceptual
parallels with Nitpicker, Wayland would fit very well into the Genode system.
However, as a prerequisite for this project, Genode's Gallium3D support must
be decomposed first. See the challenges regarding our Gallium3D support for
further information.
Runtime environments
####################
:Android's Dalvik VM natively on Genode:
Dalvik is a Java virtual machine that is used for executing applications on
Android. By running Dalvik directly on Genode, the Linux kernel could be
removed from the trusted computing base of Android, facilitating the use of
this mobile OS in high-assurance settings.
:Vancouver VMM for Genode on the NOVA hypervisor:
Vancouver is the user-level virtual-machine monitor accompanying the NOVA
hypervisor. It combines a VT-based CPU virtualization with a rich set of
device models to run unmodified guest operating systems at near-native
performance. Since NOVA is a supported base platform of Genode, running
Vancouver in the dynamic Genode environment has become feasible. By running
Vancouver on Genode instead of NOVA's original static userland would open up
new use cases where the combination of faithful virtualization with dynamic
applications is desired.
Genode 11.11 introduced the initial integration of Vancouver into Genode.
This version of Vancouver is able to bootstrap another kernel (e.g.,
Fiasco.OC) within the virtual machine. However, several pieces are missing
for reaching the goal of running a fully-fledged Linux OS as guest.
:Runtime for the D programming language:
The D systems programming language was designed to overcome many gripes that
exists with C++. In particular, it introduces a sane syntax for meta
programming, supports unit tests, and contract-based programming. These
features make D a compelling language to explore when implementing OS
components. Even though D is a compiled language, it comes with a runtime
providing support for exception handling and garbage collection. The goal of
the project is to explore the use of D for Genode programs, porting the
runtime to Genode, adapting the Genode build system to accommodate D
programs, and interfacing D programs with other Genode components written in
C++.
Platforms
#########
:Evaluation of MP scheduling models on different Genode base platforms:
Several of Genode's supported base platforms come with multi-processor
support, i.e., Linux, NOVA, L4ka::Pistachio, and Fiasco.OC. Each of
these kernels follows a different approach for utilizing multiple CPUs. For
example, Linux manages the association of threads with CPUs
largely transparent for user-level programs. In contrast, NOVA makes the use
of multiple CPUs explicit and constraints the modes of IPC interaction of
threads running on different CPUs. Furthermore, kernels differ with regard to
thread migration and scheduling. The goal of this project is to identify ways
to support the SMP features of the respective kernels at Genode's API level
such that SMP can be easily utilized by Genode programs in a largely kernel
agnostic way.
:Microkernelizing Linux:
Thanks to Genode's generic interfaces for I/O access as provided by core, all
Genode device drivers including drivers ported from Linux and gPXE can be
executed as user-level components on all supported microkernels. However, so
far, we have not enabled the use of these device drivers on Linux as base
platform. The goal of this project is the systematic replacement of in-kernel
Linux device drivers by Genode processes running in user space, effectively
reducing the Linux kernel to a runtime for Genode's core process. But moving
drivers to Genode processes is just the beginning. By employing further
Genode functionality such as its native GUI, lwIP, and Noux, many protocol
stacks can effectively be removed from the Linux kernel.
The goal of this project is to evaluate how small the Linux kernel can get
when used as a microkernel.
:Support for the HelenOS/SPARTAN kernel:
[http://www.helenos.org - HelenOS] is a microkernel-based multi-server OS
developed at the university of Prague. It is based on the SPARTAN microkernel,
which runs on a wide variety of CPU architectures including Sparc, MIPS, and
PowerPC. This broad platform support makes SPARTAN an interesting kernel to
look at alone. But a further motivation is the fact that SPARTAN does not
follow the classical L4 road, providing a kernel API that comes with an own
terminology and different kernel primitives. This makes the mapping of
SPARTAN's kernel API to Genode a challenging endeavour and would provide us
with feedback regarding the universality of Genode's internal interfaces.
Finally, this project has the potential to ignite a further collaboration
between the HelenOS and Genode communities.
:Support for the seL4 kernel:
The seL4 kernel developed by NICTA and OK-Labs is the first formally verified
microkernel. It runs on the x86 and ARM architectures and supports the
execution of a paravirtualized version of Linux on top. Even though seL4 is
proprietary technology, a free binary release and the specification of the
kernel API has been published early 2011. Being a capability-based kernel,
seL4 is in the line of the current-generation L4 kernels alongside NOVA and
Fiasco.OC. Genode already supports the latter two kernel, which hints at the
feasibility to support seL4 as well. Currently, the seL4 kernel comes with a
rather static user land, which is far from utilizing the full potential of
the kernel with regard to dynamic resource management. By adapting Genode to
seL4, a rich dynamic application workload would become available to this
kernel, which could potentially spawn interest in extending the formal
verification efforts at NICTA to the Genode system executing dynamic
real-world applications.
:Support for the Barrelfish kernel:
[http://barrelfish.org - `Barrelfish] is a so-called multi-kernel OS designed
for heterogeneous multi-processor systems. At its heart, it is a
microkernel-based multi-server OS. Its kernel provides different mechanisms
than L4-based kernels. Instead of managing threads in the kernel, there is a
mechanism for implementing preemptive multi-threading at user level.
Consequently, inter-process communication does not address threads but
protection domains. This makes the Barrelfish kernel a very interesting and
challenging target for running Genode.
:Support for the XNU kernel (Darwin):
XNU is the kernel used by Darwin and Mac OS X. It is derived from the
MACH microkernel and extended with a UNIX-like syscall API. Because the
kernel is used for Mac OS X, it could represent an industry-strength
base platform for Genode supporting all CPU features as used by Mac OS X.
:Xen as kernel for Genode:
Using Xen as kernel for Genode would clear the way to remove the
overly complex Linux OS from the trusted computing base of Xen
guests OSes.
Xen is a hypervisor that can host multiple virtual machines on one physical
machine. For driving physical devices and for virtual-machine management, Xen
relies on a privileged guest OS called Dom0. Currently, Linux is the
predominant choice to be used as Dom0, which implicates a trusted computing
base of millions of lines of code for the other guest OSes.
Even though Xen was designed as hypervisor, a thorough analysis done by Julian
Stecklina concludes that Xen qualifies well as a kernel for Genode. For
example, Julian implemented a version of Genode's IPC framework that utilizes
Xen's communication mechanisms (event channels and shared memory).
:Linux process containers for supporting Genode`s resource trading:
Even though the Linux version of Genode is primarily meant as a development
platform, there exist interesting opportunities to explore when combining
Genode with Linux, in particular Linux' process containers.
Linux process containers provide a mechanism to partition physical resources,
foremost CPU time, between Linux processes. This raises the interesting
question of whether this mechanism could be used for a proper implementation
of Genode's resource trading on Linux.
[http://lwn.net/Articles/236038/ - Process containers introduction...]
Optimizations
#############
:Low-latency audio streaming:
Genode comes with an audio streaming interface called 'Audio_out' session.
It is based on a shared-memory packet stream accompanied with asynchronous
data-flow signals. For real-time audio processing involving chains of Genode
components, streams of audio data must be carried at low latency, imposing
constraints to buffer sizes and the modes of operation of the audio mixer and
audio drivers. The goal of this project is to create a holistic design of the
whole chain of audio processing, taking thread-scheduling into account. A
particular challenge is the mixed output of real-time (small buffer, low
latency) and non-real-time (larger buffer to compensate jitter, higher
latency) audio sources.
:De-privileging the VESA graphics driver:
The VESA graphics driver executes the graphics initialization code provided
by the graphics card via an x86 emulator. To initialize a graphics mode, this
code needs to access device hardware. Currently, we permit access to all
device registers requested by the graphics-card's code. These devices include
the system timer, the PCI configuration registers, and the interrupt
controller, which are critical for the proper operating of the kernel. The
goal of this work is to restrict the permissions of the VESA driver to a
minimum by virtualizing all devices but the actual graphics card.