Introduction to Genode
Genode is a construction kit for building special-purpose operating systems out of a number of components such as device drivers, protocol stacks, and applications. Those components are organized using only a few yet powerful architectual prinicples, and thereby, allow for the composition of a wide range of different systems. The live CD is meant to showcase how far the concept scales as of today.
The following introduction will provide you with hands-on experience with the basics of Genode:
The creation and destruction of single processes as well as arbitrarily complex sub systems
The trusted-path facility of the Nitpicker secure GUI
The assignment of resource quotas to sub systems
The multiple instantiation of services
The usage of run-time adaptable policy for routing client requests to different services
The launchpad application starter
Figure 2 shows the main window of the launchpad application. It consists of three areas. The upper area contains status information about launchpad itself. The available memory quota is presented by a grey-colored bar. The middle area of the window contains the list of available applications that can be started by clicking on the application's name. Before starting an application, the user can define the amount of memory quota to donate to the new application by adjusting the red bar using the mouse. exec-once:launchpad(22M) - Start the launchpad by clicking on this link...
For a first test, you may set the memory quota of the program named scout to 10MB and then click its name. Thereupon, another instance of the scout text browser will be started and the lower area of launchpad becomes populated with status information about launchpad's children. Currently, launchpad has scout as its only child. For each child, its name, its memory quota, and a kill button are presented. After having started scout, you will further notice a change of launchpad's own status information as the memory quota spent for scout is not directly available to launchpad anymore.
In Figure 3, you see an illustration of the current setup (slightly simplified, leaving out the main menu and the other parts of the live CD). At the very bottom, there are the kernel, core, and init. Init has started the framebuffer driver, the timer driver, the nitpicker GUI server, and launchpad as it children. Launchpad, in turn, has started the second instance of scout as its only child. You can get a further idea about the relationship between the applications by pressing the ScrLock key, which gets especially handled by the nitpicker GUI server. We call this key the X-ray key because it makes the identity of each window on screen visible to the user. Each screen region gets labeled by its chain of parents and their grandparents respectively. During the walk through the demo scenario, you may press the X-ray key at any time to make the parent-child relationships visible on screen.
By pressing the kill button (labeled with x) of the scout child in launchpad's window, scout will disappear and launchpad regains its original memory quota. Although killing a process may sound like a simple thing to do, it is worthwhile to mention that scout was using a number of services, for example core's LOG service, the nitpicker GUI service, and the timer service. While using these services, scout made portions of its own memory quota available to them. When scout was killed by launchpad, all those relationships were gracefully reverted such that there is no resource leakage.
Recursive system structure
Thanks to the recursive structure of Genode, the mechanisms that function for a single application are also applicable to whole sub systems. As a test, you may configure the launchpad application entry within the launchpad window to 15MB and start another instance of launchpad. A new launchpad window will appear. Apart from the status information at the upper part of its window, it looks completely identical to the first instance. You may notice that the displayed available quota of the second launchpad instance is lower then the 15MB. The difference corresponds to the application's static memory usage including the BSS segment and the double-buffer backing store. With the new instance, you may start further applications, for example by clicking on testnit. To distinguish the different instances of the applications on screen, the X-ray key becomes handy again. Figure 4 shows a screenshot of the described setup in X-ray mode. Now, after creating a whole hierarchy of applications, you can try killing the whole tree at once by clicking the kill button of the launchpad entry in the original launchpad window. You will notice that whole sub system gets properly destructed and the original system state is regained.
The flexibility of nested policies
Beside providing the ability to construct and destruct hierarchically structured sub systems, the recursive system structure allows for an extremely flexible definition and management of system policies that can be implanted into each parent. As an example, launchpad has a simple built-in policy of how children are connected to services.
If a child requests a service, launchpad looks if such a service is provided by any of the other children and, if so, a connection gets established. If the service is not offered by any child, launchpad delegates the request to its parent. For example, a request for the LOG service will always end up at core, which implements the service by the means of terminal (or kernel debug) output. By starting a child that offers the same service interface, however, we can shadow core's LOG service by an alternative implementation. You can try this out by first starting testnit and observing its log output at the terminal window. When started, testnit tells us some status information. By further starting the program called nitlog, we create a new LOG service as a child of launchpad. On screen, this application appears just as a black window that can be dragged to any screen position with the mouse. When now starting a new instance of testnit, launchpad will resolve the request for the LOG service by establishing a connection to nitlog instead of propagating the request to its parent. Consequently, we can now observe the status output of the second testnit instance inside the nitlog window.
The same methodology can be applied to arbitrarily complex services. For example, you can create a new instance of the framebuffer service by starting the liquid_fb application. This application provides the framebuffer service and, in turn, uses the nitpicker GUI server to get displayed on screen. Because any new requests for a framebuffer will now be served by the liquid_fb application, we can start another instance of nitpicker. This instance uses liquid_fb as its graphics back end and, in turn, provides the GUI service. Now, when starting another instance of scout, the new scout window will appear within liquid_fb too (Figure 5).
The extremely simple example policy implemented in launchpad in combination with the recursive system structure of Genode already provides a wealth of flexibility without the need to recompile or reconfigure any application. The policy implemented and enforced by a parent may also deny services for its children or impose other restrictions. For example, the window labels presented in X-ray mode are successively defined by all parents and grandparents that mediate the request of an application to the GUI service. The scout window as the parent of launchpad imposes its policy of labeling the GUI session with the label "launchpad". Init as the parent of scout again overrides this label with the name of its immediate child from which the GUI request comes from. Hence the label becomes "scout -> launchpad".
Where to go from here?
Although this little demonstration scratches only the surface of Genode, we hope that the power of its underlying design becomes apparent. The most distinctive property of Genode, however, is its extremely low complexity. The functionality of the complete demo scenario is implemented in less than 20,000 lines of source code (LOC), including the GUI and the demo applications. As a point of reference, when relying on libpng for decompressing the images as seen in the text browser, this number doubles. In fact, the complete base OS framework accounts for less source-code complexity than the code needed for decoding the PNG images. To these numbers, the complexity of the used underlying kernel must be added, for example 10-20 KLOC for an L4 microkernel (or far more than 500 KLOC when relying on the Linux kernel). In combination with a microkernel, Genode enables the implementation of security-sensitive applications with a trusted computing base (TCB) of some thousands rather than millions of lines of code. If using a hypervisor as kernel for Genode, this advantage can further be combined with compatibility to existing applications executed on virtual machines.
More details, architectural and technical documents, our road map, and the complete source code are available at http://genode.org.
The development of the Genode OS Framework is conducted as an open-source community project, coordinated by Genode Labs, a company founded by the original authors of Genode. If you are interested in supporting our project through participation or funding, please consider joining our community (http://genode.org) or contact Genode Labs (http://www.genode-labs.com).