Copyright © 2006 -, 2011 VMware Inc., Oracle Inc.
Documentation made available under the terms of the Eclipse Public License v1.0 and Apache License v2.0 which accompanies this distribution. The Eclipse Public License is available at http://www.eclipse.org/legal/epl-v10.html and the Apache License v2.0 is available at http://www.opensource.org/licenses/apache2.0.php. You may elect to redistribute this code under either of these licenses.
java.lang.LinkageError: loader constraint violationif I use Gemini Blueprint. Things work fine with vanilla OSGi. What's wrong?
Eclipse Gemini Blueprint is based on Spring Dynamic Modules 2.0 which has been migrated to the Eclipse Foundation. While the name and the packages have changed the internals are still the same; one will find both names are used interchangeably (especially by those that have used Spring DM for a while). Further more, all Spring DM applications are supported by Gemini Blueprint (do pay attention to the migration guide available here).
The OSGi term is a trademark belonging to The OSGi Alliance. In order to comply with their guidelines, it was decided that the project name be changed to "Spring Dynamic Modules for OSGi Service Platforms" (aka Spring DM). The new name is still pending final approval by the OSGi Alliance. The name change was the result of an amicable discussion between the OSGi Alliance and Interface21. Interface21 is a member of the OSGi Alliance, and the OSGi Alliance remain very supportive of the project. Note that Spring Dynamic Modules has moved to Eclipse Foundation to form Gemini Blueprint project (see above)
org.eclipse.gemini.blueprint.*.internal packages are
meant (as the name implies) to be private and non-public package. Thus,
there is no documentation, support or compatibility guarantee for
them. In fact, the Gemini Blueprint bundle does not even export
them to prevent accidental usage.
If you find classes under this package, which you really, really depend on, then consider raising an issue on JIRA to have access opened up.
Gemini Blueprint requires at least Java 1.7, OSGi 4.1+ and Spring 4.2 It might be possible to use Gemini Blueprint on a lower execution environment (such CDC) but it is not guaranteed to work. Both Spring and Gemini Blueprint rely on JavaBeans (java.beans package) which, unfortunately, is missing in most restricted environments. See this PDF for information on CDC profiles. Note that, Spring 4.2 also requires Java 1.7.
Nevertheless, experiences and feedback on running Gemini Blueprint in restricted environments is welcomed - please use our mailing list.
There are no current plans to support other dynamic module frameworks.
See the requirements entry. The OSGi Service Platform is designed to run in very constrained environments however, Gemini Blueprint depends on the Spring Framework v4.2.x which in turn depends on JDK 1.7. Thus Gemini Blueprint cannot run on more constrained environments (such as the OSGi Minimum Execution Environment) unless Spring itself also runs in those environments. There are no current plans to make such a version of Spring. However as existing OSGi developers adopt Gemini Blueprint to simplify creation of OSGi applications and the user base expands, the target audience can cover domains much broader than enterprise Java applications. In time this could create a large enough demand to justify the investment needed to allow Spring and Gemini Blueprint to run in restricted environments.
Gemini Blueprint requires an OSGi 4.2 (though it might work on OSGi 4.0 and 4.1) platform. The framework has been tested on Equinox, Felix and Knopflerfish - in fact, the test suite is ran against all of them as part of our continuous integration process.
The best place to start is The Osgi Alliance home and developer pages which provide the OSGi specifications, introductions and many links and blogs on the topic. Please see the reference documentation appendix for more information.
In addition, all OSGi implementation websites host detailed, step-by-step tutorials and introduction.
If you discover any additional materials useful for OSGi newbies, please let us know to update the list. Thank you.
Please see the file called
readme-building.txt found in the source tree.
OSGi platforms do not change the way libraries work, it just enforces tighter classloading. Thus, you can, in most of the cases, use the same logging strategy used in non-OSGi environments.
Spring (and Gemini Blueprint) use internally the commons-logging API which acts as an "ultra-thin bridge between different logging implementations". In OSGi, just like in a non-OSGi environment, Spring and Gemini Blueprint delegate all logging (including initialisation) to the actual commons-logging API implementation.
Out of the box, SLF4J library is provided, which
shares the same purpose as commons-logging but without the
class loading discovery mechanism (that causes loading issues), using
static wiring (see the SLF4J site for more info). To use slf4j, make sure
slf4j-log4j-XXX.jar (where XXX stands for the slf4j version).
The last jar provides the static wiring between slf4j and log4j - if another implementation
is desired (such as jdk14), then a different jar is required (for the jdk14 that would be
slf4j-jdk14-XXX.jar) - see the official SLF4J site for more information.
Please see this
question for more details on why commons-logging jar is not
Gemini Blueprint uses SLF4J on top of Log4J but this can be
easily changed. As part of log4j initialisation, a
configuration fille needs to be present in the bundle classpath. This
means that the configuration file has to be part of your bundle or one
of its attached fragments. Besides SLF4J, for another OSGi-aware
solution, one can try Pax
To learn more about log4j setup process, follow this link.
The commons-logging project provides the commons-logging API
commons-logging-api-nn.jar) along with an
that provides a wrapper between the API and the actual logging libraries
used underneath (such as log4j, java.util.logging, etc). However, in
order to determine what implementation should be used, commons-logging
library tries to do some classloading-based discovery that is fragile
and can fail unexpectedly. In an strict classloading environment such
as OSGi, this mechanism adds unnecessary complexity - that's why we
decided to use SFL4J which is not just simpler and actively maintained
but is also OSGi-friendly out of the box.
If you have to use commons-logging (for example the jar is required by certain bundles) then try using the most recent version commons-logging version (1.1+) as it provides more options on the discovery process. Below are some settings that can be used to make commons-logging work inside an OSGi environment:
Unfortunately, commons-logging 1.0.x uses the thread context class loader (TCCL) always for loading
loggers implementations. Inside an OSGi environment, the TCCL is undefined and cannot be relied upon.
Since managing the TCCL is almost impossible as most loggers are defined as static fields that need to
resolved on class loading, using a different
LogFactory is advised. One can use
the org.apache.commons.logging.LogFactory system property to specify a different
log factory however, the commons-logging bundle should be able to load this class.
If using commons logging 1.1.x, one can turn off the tccl usage through use_tccl property, part of the commons-logging.properties file. http://commons.apache.org/logging/commons-logging-1.1/troubleshooting.html. Additionally, 1.1.x provides several system properties (such as org.apache.commons.logging.Log.allowFlawedContext, org.apache.commons.logging.Log.allowFlawedDiscovery and org.apache.commons.logging.Log.allowFlawedHierarchy) that can change the behavious of the discovery process. See the LogFactoryImpl javadoc for more details.
In our tests, commons logging 1.1.x can be used with reasonable success inside OSGi. We haven't been able to find a generic configuration for commons logging 1.0.x that works and that does not rely on fragile hacks dependent on the running environment.
It is completely up to you what logging implementation you want Gemini Blueprint to use. To route log messages to the OSGi logging service, just use a commons-logging API implementation that delegates to the OSGi logging service, such as Pax Logging.
OSGi requires JARs to contain certain
MANIFEST.MF entries which indicate what classes are
required and shared by each archive. This means that tradition jars cannot be used inside an OSGi environment.
To solve the problem one can:
Use a repository of pre-wrapped libraries such as Orbit, Felix Commons or Knopflerfish repository. Gemini Blueprint uses the SpringSource Enterprise Bundle Repository for its dependencies, which you might find useful. Additionally, for artifacts that have not yet made it into SpringSource Repository, Gemini Blueprint provides a small, temporary (Amazon S3) Maven repository (link | browser-friendly link) for its internal usage.
Wrap the necessary jars with proper OSGi manifest. While this can be done by hand, we strongly recommend Peter Kriens excellent bnd tool which can do this for you automatically. For Maven, see Felix maven-bundle-plugin.
Include the jar inside your OSGi bundle and include it in the bundle classpath through Bundle-ClassPath directive. See the OSGi specification for more information.
This problem is likely to be caused by bad class wiring. All 1.3+ JDKs include incomplete javax.transaction and javax.transaction.xa packages for usage inside ORB environments. To address the problem, use a proper JTA library which contains all the classes from the forementioned packages and exports them with a specific version to prevent confusion.
Gemini Blueprint wraps JTA 1.1 library for OSGI environments which can be found at Spring snapshot repository. One can deploy this library and specify version 1.1 for javax.transaction* packages inside Import-Package header. By specifying the version, one can be sure that the proper package is used.
Note that JTA 1.1 is compatible with version 1.0.1.
Remove the official OSGi jars (osgi.jar or osgi-r4-core.jar) from the classpath and use only the actual OSGi platform (Equinox/Knopflerfish/Felix) jars. The former provides only the public classes without an actual implementation and thus cannot be used during runtime, only during the compilation stage.
autoExport flag, part of the service exporter, will discover and
include for exporting only the visible interfaces/classes implemented
by the service object. Consider class
GenericApplicationContext which implements among
org.springframework.beans.factory package) and
Depending on your OSGi imports, the exporting bundle can see
only one of the packages, none or both. Based on these visibility
settings, the exporter will only export the classes that are 'known'
to the exporting bundle. For example, if the exporting bundle sees
org.springframework.core.io but not
org.springframework.beans.factory, the service will
be exported as a
ResourceLoader but not as a
BeanFactory. In fact, exporting the object as a
BeanFactory will fail since the bundle doesn't see
this interface and thus doesn't know how to handle its contract.
This problem occurs on Spring DM versions up to 1.0 - consider upgrading to 1.0.1 or better. If you are stuck with 1.0 see below. When running an integration test, Gemini Blueprint will duplicate the test instance and execute it inside OSGi. To avoid problems like this one, make sure you are using the same libraries (with the same version) as Spring DM when running your test. This particular problem for example is caused by a JUnit 3.8.2 vs 3.8.x serialization compatibility. Make sure that you are using at least JUnit 3.8.2 for the execution of your tests.
Eclipse PDE uses Equinox OSGi platform underneath which (like other OSGi platforms) caches the bundles between re-runs. When
the cache is not properly updated, one can encounter strange behaviour (such as the new services/code being picked up)
or errors ranging from class versioning to linkage. Consider doing a complete clean build or, in case of Eclipse,
creating a new workspace or deleting the bundle folder (depends on each project settings but most users should find it at:
In Spring DM 1.1 M2, the proxy infrastructure has been refined to avoid type leaks, the usage of dynamic imports or exposure of class loader chain delegation. If you encounter class visibility problems during the upgrade then it's likely you have missing imports which were previously resolved as a side effect of Spring DM proxy weaving process.
To deal with dynamics, Gemini Blueprint creates proxies around the imported services. The proxies are classes (generated at runtime), different
from the target but able to intercept the calls made to it. Since a proxy is different then its target, comparing objects against it can
yield different results then when the comparison is done against the target. In most scenarios this is not a problem but there might be
corner cases where this contract matters. Since 1.1, Gemini Blueprint importer proxies implement
interface (from Spring framework) which allow access to the raw target.
Make sure the Gemini Blueprint collections are injected into object of compatible types (for example
java.util.Collection). If the types are not compatible,
the container will have to perform
transforming the Gemini Blueprint managed collection into a 'normal' one, unaware of the OSGi dynamics.
Since Spring 2.5.6, the symbolic names of the artifacts have changed slightly. Spring DM/Gemini Blueprint aligned its symbolic names as well with the
new patter since Spring DM 1.2.0 M2. Thus the prefix
org.springframework.bundle.osgi has been changed to
org.eclipse.gemini.blueprint; for example Gemini BLueprint extender symbolic name was changed from
(notice the missing
To fix this problem, change the reference to the old symbolic name (usually inside the fragments manifests or LDAP filters) to the new one.
Most likely the bundle imports do not identify all the transitive packages dependencies (through the uses directive). Packages
imported from a bundle, can in turn, depend on other packages from other bundles - these are sometime called transitive or indirect dependencies. If
these are not properly identified, the OSGi platform cannot validate the wiring and allows multiple versions of the same class to be available inside the
same class space. This problem appears whether Gemini Blueprint is used or not. However, due to the usage of proxies inside Gemini Blueprint (which forces eager class loading
for the classes proxies), the class graph is evaluated at runtime causing the problem to occur early. With vanilla OSGi, the loading occurs lazy which means
the problem is going to occur at runtime rather then at startup.
To fix the problem, add the relevant transitive packages to the list of exported packages, either manually or automatically through tools such as
Bundlor or Bnd.
For more information on the
uses directive please see the OSGi spec,
section 3.6.4 or this SpringSource blog entry.
Gemini Blueprint (Spring DM 2.0) is the Reference Implementation for OSGi 4.2 Blueprint Container specification. Simply deploy your Blueprint bundles in a platform where Gemini Blueprint is already activated and you should be done. For more information about Blueprint and Gemini Blueprint, please see the Blueprint Container chapter in the reference documentation.
Since 2.3.0, Knopflerfish changed the way it does bootpath delegation which causes classes to be loaded from inside and outside OSGi. Set
the system property
true before starting up the Knopflerfish
platform to prevent this from happening. See Knopflerfish 2.3.x release notes for more
Some of the Spring jars (optionally) import each other packages for bootstrapping reasons. While cycles between jars are allowed and supported by the OSGi platform, by default, Eclipse PDE complains about them. Since Eclipse Galileo, it is possible to enable binary cycles in a target platform. allowing the build to compile. For more information, see this entry from the Eclipse Reference Guide.