OSGi – bundles / fragments / dependencies

In the last weeks I needed to look at several issues regarding OSGi dependencies in different products. A lot of these issues were IMHO related to wrong usage of OSGi bundle fragments. As I needed to search for various solutions, I will publish my results and my opinion on the usage of fragments in this post. Partly also for myself to remind me about it in the future.

What is a fragment?

As explained in the OSGi Wiki, a fragment is a bundle that makes its contents available to another bundle. And most importantly, a fragment and its host bundle share the same classloader.

Looking at this from a more abstract point of view, a fragment is an extension to an existing bundle. This might be a simplified statement. But considering this statement helped me solving several issues.

What are fragments used for?

I have seen a lot of different usage scenarios for fragments. Considering the above statement, some of them where wrong by design. But before explaining when not to use fragments, let’s look when they are the agent of choice. Basically fragments need to be used whenever a resource needs to be accessible by the classloader of the host bundle. There are several use cases for that, most of them rely on technologies and patterns that are based on standard Java. For example:

  • Add configuration files to a third-party-plugin
    e.g. provide the logging configuration (log4j.xml for the org.apache.log4j bundle)
  • Add new language files for a resource bundle
    e.g. a properties file for locale fr_FR that needs to be located next to the other properties files by specification
  • Add classes that need to be dynamically loaded by a framework
    e.g. provide a custom logging appender
  • Provide native code
    This can be done in several ways, but more on that shortly.

In short: fragments are used to customize a bundle

When are fragments the wrong agent of choice?

To explain this we will look at the different ways to provide native code as an example.

One way is to use the Bundle-NativeCode manifest header. This way the native code for all environments are packaged in the same bundle. So no fragments here, but sometimes not easy to setup. At least I struggled with this approach some years ago.

A more common approach is to use fragments. For every supported platform there is a corresponding fragment that contains the platform specific native library. The host bundle on the other side typically contains the Java code that loads the native library and provides the interface to access it (e.g. via JNI). This scenario is IMHO a good example for using fragments to provide native code. The fragment only extend the host bundle without exposing something public.

Another approach is the SWT approach. The difference to the above scenario is, that the host bundle org.eclipse.swt is an almost empty bundle that only contains the OSGi meta-information in the MANIFEST.MF. The native libraries aswell as the corresponding Java code is supplied via platform dependent fragments. Although SWT is often referred as reference for dealing with native libraries in OSGi, I think that approach is wrong.

To elaborate why I think the approach org.eclipse.swt is using is wrong, we will have a look at a small example.

  1. Create a host bundle in Eclipse via File -> New -> Plug-in Project and name it org.fipro.host. Ensure to not creating an Activator or anything else.
  2. Create a fragment for that host bundle via File -> New -> Other -> Plug-in Development -> Fragment Project and name it org.fipro.host.fragment. Specify the host bundle org.fipro.host on the second wizard page.
  3. Create the package org.fipro.host in the fragment project.
  4. Create the following simple class (yes, it has nothing to do with native code in fragments, but it also shows the issues).
    package org.fipro.host;
    public class MyHelper {
    	public static void doSomething() {
    		System.out.println("do something");

So far, so good. Now let’s consume the helper class.

  1. Create a new bundle via File -> New -> Plug-in Project and name it org.fipro.consumer. This time let the wizard create an Activator.
  2. In Activator#start(BundleContext) try to call MyHelper#doSomething()

Now the fun begins. Of course MyHelper can not be resolved at this time. We first need to make the package consumable in OSGi. This can be done in the fragment or the host bundle. I personally tend to configure Export-Package in the bundle/fragment where the package is located. We therefore add the Export-Package manifest header to the fragment. To do this open the file org.fipro.host.fragment/META-INF/MANIFEST.MF. Switch to the Runtime tab and click Add… to add the package org.fipro.host.

Note: As a fragment is an extension to a bundle, you can also specify the Export-Package header for org.fipro.host in the host bundle org.fipro.host. org.eclipse.swt is configured this way. But notice that the fragment packages are not automatically resolved using the PDE Manifest Editor and you need to add the manifest header manually.

After that the package org.fipro.host can be consumed by other bundles. Open the file org.fipro.consumer/META-INF/MANIFEST.MF and switch to the Dependencies tab. At this time it doesn’t matter if you use Required Plug-ins or Imported Packages. Although Import-Package should be always the preferred way, as we will see shortly.

Althought the manifest headers are configured correctly, the MyHelper class can not be resolved. The reason for this is PDE tooling. It needs additional information to construct proper class paths for building. This can be done by adding the following line to the manifest file of org.fipro.host

Eclipse-ExtensibleAPI: true

After this additional header is added, the compilation errors are gone.

Note: This additional manifest header is not necessary and not used at runtime. At runtime a fragment is always allowed to add additional packages, classes and resources to the API of the host.

After the compilation errors are gone in our workspace and the application runs fine, let’s try to build it using Maven Tycho. I don’t want to walk through the whole process of setting up a Tycho build. So let’s simply assume you have a running Tycho build and include the three projects to that build. Using POM-less Tycho this simply means to add the three projects to the modules section of the build.

You can find further information on Tycho here:
Eclipse Tycho for building Eclipse Plug-ins and RCP applications
POM-less Tycho builds for structured environments

Running the build will fail because of a Compilation failure. The Activator class does not compile because the import org.fipro.host cannot be resolved. Similar to PDE, Tycho is not aware of the build dependency to the fragment. This can be solved by adding an extra. entry to the build.properties of the org.fipro.consumer project.

extra.. = platform:/fragment/org.fipro.host.fragment

See the Plug-in Development Environment Guide for further information about build configuration.

After that entry was added to the build.properties of the consumer bundle, also the Tycho build succeeds.

What is wrong with the above?

At first sight it is quite obvious what is wrong with the above solution. You need to configure the tooling at several places to make the compilation and the build work. These workarounds even introduce dependencies where there shouldn’t be any. In the above example this might be not a big issue, but think about platform dependent fragments. Do you really want to configure a build dependency to a win32.win32.x86 fragment on the consumer side?

The above scenario even introduces issues for installations with p2. Using the empty host with implementations in the fragments forces you to ensure that at least (or exactly) one fragment is installed together with the host. Which is another workaround in my opinion (see Bug 361901 for further information).

OSGi purists will say that the main issue is located in PDE tooling and Tycho, because the build dependencies are kept as close as possible to the runtime dependencies (see for example here). And using tools like Bndtools you don’t need these workarounds. And in first place I agree with that. But unfortunately it is not possible (or only hard to achieve) to use Bndtools for Eclipse application development. Mainly because in plain OSGi, Eclipse features, applications and products are not known. Therefore also the feature based update mechanism of p2 is not usable. But I don’t want to start the discussion PDE vs. Bndtools. That is worth another (series) of posts.

In my opinion the real issue in the above scenario, and therefore also in org.eclipse.swt, is the wrong usage of fragments. Why is there a host bundle that only contains the OSGi meta information? After thinking a while about this, I realized that the only reason can be laziness! Users want to use Require-Bundle instead of configuring the several needed Import-Package entries. IMHO this is the only reason that the org.eclipse.swt bundle with the multiple platform dependent fragments exists.

Let’s try to think about possible changes. Make every platform dependent fragment a bundle and configure the Export-Package manifest header for every bundle. That’s it on the provider side. If you wonder about the Eclipse-PlatformFilter manifest header, that works for bundles aswell as for fragments. So we don’t loose anything here. On the consumer side we need to ensure that Import-Package is used instead of Require-Bundle. This way we declare dependencies on the functionality, not the bundle where the functionality originated. That’s all! Using this approach, the workarounds mentioned above can be removed. PDE and Tycho are working as intended, as they can simply resolve bundle dependencies. I have to admit that I’m not sure about p2 regarding the platform dependent bundles. Would need to check this separately.


Having a look at the two initial statements about fragments

  • a fragment is an extension to an existing bundle
  • fragments are used to customize a bundle

it is IMHO wrong to make API public available from a fragment. These statements could even be modified to become the following:

  • a fragment is an optional extension to an existing bundle

Having that statement in mind, things are getting even clearer when thinking about fragments. Here is another example to strengthen my statement. Guess you have a host bundle that already exports a package org.fipro.host. Now you have a fragment that adds an additional public class via that package, and in a consumer bundle that class is used. Using Bndtools or the workarounds for PDE and Tycho showed above, this should compile and build fine. But what if the fragment is not deployed or started at runtime? Since there is no constraint for the consumer bundle that would identify the missing fragment, the consumer bundle would start. And you will get a ClassNotFoundException at runtime.

Personally I think that everytime a direct dependency to a fragment is introduced, there is something wrong.

There might be exceptions to that rule. One could be to create a custom logging appender that needs to be accessible in other places, e.g. for programmatically configurations. As the logging appender needs to be in the same classloader as the logging framework (e.g. org.apache.log4j), it needs to be provided via fragment. And to access it programmatically, a direct dependency to the fragment is needed. But honestly, even in such a case a direct dependency to the fragment can be avoided with a good module design. Such a design could be for example to make the appender an OSGi service. The service interface would be defined in a separate API bundle and the programmatic access would be implemented against the service interface. Therefore no direct dependency to the fragment would be necessary.

As I struggled several days with searching for solutions on fragment dependency issues, I hope this post can help others, solving such issues. Basically my solution is to get rid of all fragments that export API and make them either separate bundles or let them provide their API via services.

If someone with a deeper knowledge in OSGi ever comes by this post and has some comments or remarks about my statements, please let me know. I’m always happy to learn something new or getting new insights.

About Dirk Fauth

Dirk Fauth is a Software Architect for Rich Client Systems working for the Robert Bosch GmbH in Stuttgart and a lecturer in Java basics for the Baden-Wuerttemberg Cooperative State University (DHBW). He is active in developing, teaching and talking about OSGi, Eclipse RCP applications and Eclipse related technologies. He is project lead of the Nebula NatTable project, Eclipse Platform committer and also a committer and contributor to several other Eclipse projects. (Twitter: @fipro78)
This entry was posted in Dirk Fauth, Eclipse, OSGi. Bookmark the permalink.