Re: [jboss-as7-dev] [OSGi WAB support] webbundle: URL handler

I was looking into it a little more too and came to the conclusion that the whole approach via the DeploymentUnitProcessor can't work. The problem is that things passed via a webbundle: URL don't need to be OSGi bundles. One of the things that the webbundle: URL handler can do is turn non-OSGi bundles (i.e. WAR files) into OSGi bundles on the fly, so you could do something like: bundle.install("webbundle:file://myplainold.war?Bundle-SymbolicName=foo&&Web-ContextPath=/test") The problem is that from very early on in the process, even before it hits the the DUPs, the bundle.install code expects the thing that is being deployed to be a valid OSGi bundle. And I think it's right in doing that. So - the only clean solution that I can see is to have the URL handler to the conversion to OSGi bundle on the fly. With that you can use these webbundle: URLs from anywhere I think as they will work from anywhere in the vm where you do URL().openStream(). Cheers, David On 30/11/2012 16:22, Thomas Diesler wrote: > I was thinking about this a little more … > > We need to distinguish between two cases > > #1 Deployment through CLI, Console, Management API > #2 Deployment through BundleContext.install(…) > > Re #1: In this case the webbundle URI needs to be passed through the mgmnt layer. What happens when you try this in CLI/Console? In both UIs there must be a way to pass in a URI that is then accessible from a DUP. > Given that we can see the URI in a DUP, it can be parsed after OSGiBundleInfoParseProcessor but before BundleDeploymentProcessor. The BundleDeploymentProcessor would see valid OSGiMetaData and everything after would proceed as normal. > > Re #2: In this case the Framework already provides the Deployment object that is then passed to the BundleLifecycleIntegration. By that time the metadata is already parsed and part of the Deployment. This happens in AbstractBundleContext > > DeploymentFactoryPlugin deploymentPlugin = frameworkState.getDeploymentFactoryPlugin(); > dep = deploymentPlugin.createDeployment(location, rootFile); > > Fortunately the DeploymentFactoryPlugin is an integration point that exits in anticipation of a requirement like this. You would need to provide a "DeploymentFactoryIntegration" that parses the webbundle URI in createDeployment. The returned Deployment is already a valid OSGi deployment. No additional work is needed in the DUPs. > > cheers > --thomas > > > On Nov 30, 2012, at 3:35 PM, Thomas Diesler <thomas.diesler(a)jboss.com <mailto:thomas.diesler@jboss.com>> wrote: > >> Good progress >> >> > I can get the Deployment object back, given its DU.name >> >> You don't need to do that. Your DUP should be placed after the BundleDeploymentProcessor, in which case the Deployment is readily available from the DU. >> >> Re the Bundle.location to DU.name mapping … >> The runtime name must be unique and at the same time simple enough to be usable as bundle key on the CLI. How about using the something like this >> >> String toRuntimeName(String location) { >> - if location is valid URI: return URI path (i.e. strip protocol & params) >> - else if location has a suffix .???: return location without suffix >> - else return location >> } >> >> >> On Nov 30, 2012, at 3:24 PM, David Bosschaert <david(a)redhat.com <mailto:david@redhat.com>> wrote: >> >>> Ah, yes. Via a static method on BundleLifecycleIntegration I can get the Deployment object back, given its DU.name. Once I have the Deployment object I can indeed find out the original URL and read the parameters off that. >>> >>> The getRuntimeName() simply returns the thing after the last "/" in the installation location. This may be alright for things that come from the filesystem, but if you install something from a URL the results can be a bit random, as people can have '/'-es in their parameters, e.g. >>> http://myhost/mybundle?param1=/whatever >>> I will have a look at making that a bit better in that context. >>> >>> I noticed that the DeploymentPlanBuilder.add() has an overload where you can provide a name and a common name. This might be useful for OSGi in that for the common name we could pass in the name we currently do, but for the name we could pass in the original Bundle.location argument. I haven't fully tested this but it seems to me that that could ensure that even in the case that getRuntimeName() returns the same thing for two different bundles things will still continue to work. >>> >>> Cheers, >>> >>> David >>> >>> On 30/11/2012 10:55, Thomas Diesler wrote: >>>> The DU runtime name is computed here <https://github.com/jbosgi/jboss-as/blob/master/osgi/service/src/main/java... >>>> >>>> LOGGER.debugf("Install deployment: %s", dep); >>>> String runtimeName = getRuntimeName(dep); >>>> putDeployment(runtimeName, dep); >>>> try { >>>> InputStream input = dep.getRoot().openStream(); >>>> try { >>>> ServerDeploymentHelper server = new ServerDeploymentHelper(deploymentManager); >>>> server.deploy(runtimeName, input); >>>> } finally { >>>> VFSUtils.safeClose(input); >>>> } >>>> } catch (RuntimeException rte) { >>>> throw rte; >>>> } catch (Exception ex) { >>>> throw MESSAGES.cannotDeployBundle(ex, dep); >>>> } >>>> >>>> That mapping from the Bundle.location to the DU.name is historical and I'm not sure what the limitations currently are. Perhaps you could find out while you're doing this and document why getRuntimeName(dep) is doing what its doing. That name also has to feed through the management layer as the deployment's runtimeName. Perhaps there are some limitations at that level. Have a look at the variations of DeploymentPlanBuilder.add(…) >>>> >>>> Ideally, we would like to pass the Bundle.location directly to the DeploymentPlanBuilder API so that it becomes DU.name unchanged. >>>> >>>> xxxxxxxxxxxxxxxxxxxxxxxxxxxx >>>> Thomas Diesler >>>> JBoss OSGi Lead >>>> JBoss, a division of Red Hat >>>> xxxxxxxxxxxxxxxxxxxxxxxxxxxx >>>> >>>> >>>> >>>> On Nov 30, 2012, at 11:29 AM, Thomas Diesler <thomas.diesler(a)jboss.com <mailto:thomas.diesler@jboss.com>> wrote: >>>> >>>>> > DeploymentUnit.getName() only returns the bare name of the item being deployed >>>>> >>>>> Is this really true? What is the DU.name when you do BundleContext.install("webbundle://foo?key=value", input) ? >>>>> >>>>> > I could not find the Deployment.location >>>>> >>>>> Have a look at the BundleDeploymentProcessor <https://github.com/jbosgi/jboss-as/blob/master/osgi/service/src/main/java... >>>>> >>>>> xxxxxxxxxxxxxxxxxxxxxxxxxxxx >>>>> Thomas Diesler >>>>> JBoss OSGi Lead >>>>> JBoss, a division of Red Hat >>>>> xxxxxxxxxxxxxxxxxxxxxxxxxxxx >>>>> >>>>> >>>>> >>>>> On Nov 29, 2012, at 10:11 PM, David Bosschaert <david(a)redhat.com <mailto:david@redhat.com>> wrote: >>>>> >>>>>> Hi Thomas, >>>>>> >>>>>> If I can get the original location/URI back in the DeploymentUnitProcessor.deploy() so I can associate them with each other that would work for me too. DeploymentUnit.getName() only returns the bare name of the item being deployed (e.g. just "mywebapp"), which isn't really precise enough. I could not find the Deployment.location that you're referring to from the deploy() method, is it available as an attachment of some sort? >>>>>> >>>>>> Cheers, >>>>>> >>>>>> David >>>>>> >>>>>> On 29/11/2012 19:22, Thomas Diesler wrote: >>>>>>> Hi David, >>>>>>> >>>>>>> The webbundle://foo?key=value URL is mainly a transport vehicle for meta data. I don't think it is intended to give access to the bytes of the war (however, we could do this too - see below). That URL spec (as a string) is the Bundle.location that is given in BundleContext(location, input). That location identifier was originally meant to be a URL that could give access to the Bundle's bytes. This is no longer the case and any string (in most cases an URI) can be given as the location. Internally, I think the location becomes the DU name. If not, it is definitely the Deployment.location. >>>>>>> >>>>>>> So a DUP does have access to that web bundle location. The URL handler is mainly a URI parser that is supposed to give access to the OSGi metadata that need to be put in the manifest (in our case the OSGiMetaData not the Manifest). AFAIK, the Framework tries to construct a URL from the location only if no input bytes are given. When we talk about an URLHandler we are mainly talking about a simple URI parser. A URLHandler would need to be implemented for BundleContext(location) to work. >>>>>>> >>>>>>> Given that the URI parsing works and that we can generate OSGiMetaData from it, the bytes that make up the WAR are maintained by the DeploymentRepository and available through the DU roots. >>>>>>> >>>>>>> In the unlikely case that the TCK does something like this: >>>>>>> >>>>>>> Manifest manifest = new Manifest(new URL("webbundle://foo?key=value").openStream()); >>>>>>> validateGeneratedManifest(manifest) >>>>>>> >>>>>>> we would need to feed back the generated OSGiMetaData to a byte buffer. In any case that would have to access the DU root content and amend it by a generated Manifest. >>>>>>> I'd have to check if the above is really required. >>>>>>> >>>>>>> If this does not help either, give me a shout and I put together a quick prototype. >>>>>>> >>>>>>> cheers >>>>>>> --thomas >>>>>>> >>>>>>> >>>>>>> On Nov 29, 2012, at 12:29 PM, David Bosschaert <david(a)redhat.com <mailto:david@redhat.com>> wrote: >>>>>>> >>>>>>>> Hi Thomas, >>>>>>>> >>>>>>>> I have the following issues with your suggestion. >>>>>>>> >>>>>>>> 1. I don't fully see how the information available to the URL handler can be associated with the information available to the DeploymentUnitProcessor. The URL handler has the URL, that's all, while AFAICS the original URL (or whatever was inside the webbundle: url) is no longer available when the deploy() method is called. >>>>>>>> 2. If we find a way to fix 1. this will only work if people use BundleContext.install(String location). It will fail when people call url.openStream() on the webbundle: url and does not work with BundleContext.install(String, InputStream). >>>>>>>> >>>>>>>> Another approach would be to simply let the URL handler do all the work, i.e. modify the stream being passed through. Then those URLs will work in any context. >>>>>>>> >>>>>>>> Cheers, >>>>>>>> >>>>>>>> David >>>>>>>> >>>>>>>> On 26/11/2012 17:25, Thomas Diesler wrote: >>>>>>>>> Hi David, >>>>>>>>> >>>>>>>>> here a quick summary of what I suggested today: >>>>>>>>> >>>>>>>>> The first thing that sees the URL coming from BundleContext.install(...) is the Framework, which has a notion of pluggable URL handlers. >>>>>>>>> In AS7 the URL handler should be an integration plugin and a DUP at the same time. The DUP would do nothing as long as the plugin >>>>>>>>> is not activated (i.e. the framework is down). When the Framework activates the URL handler gets registered with the framework and the DUP becomes active. >>>>>>>>> >>>>>>>>> The DUP would then need to provide OSGiMetaData with a Bunde-SymbolicName and Bundle-Classpath. The Bundle-Classpath should point to WEB-INF/classes and >>>>>>>>> the collection of stuff in WEB-INF/lib. For completeness it could generate Package-Import requirements on the javax.servlet.* APIs. The DUP should be placed after >>>>>>>>> the DUP that normally provides OSGiMetaData and should do nothing if the OSGiMetaData is already there. >>>>>>>>> >>>>>>>>> Hope that helps, cheers >>>>>>>>> --thomas >>>>>>>>> >>>>>>>>> >>>>>>>>> On 11/21/2012 05:58 PM, David Bosschaert wrote: >>>>>>>>>> Hi all, >>>>>>>>>> >>>>>>>>>> As part of making the JBoss OSGi Web Application Support compliant with >>>>>>>>>> the spec I have started running it through the OSGi TCK. >>>>>>>>>> I noticed that the TCK depends heavily on the webbundle: URL protocol >>>>>>>>>> which is specified in section 128.4 of the specification - it is not an >>>>>>>>>> optional piece. So in order to support this we need to provide such a >>>>>>>>>> URL handler. >>>>>>>>>> >>>>>>>>>> As the webbundle: handler is never part of runtime operation (it only >>>>>>>>>> converts a WAR file into a WAB file on the fly) I was looking into >>>>>>>>>> possibly using existing implementations of the URL handler instead. >>>>>>>>>> However the ones that I found are quite heavy on the dependencies. The >>>>>>>>>> implementation in Aries depends on Blueprint being present and the one >>>>>>>>>> in Pax has about 10 other dependencies (including junit) - they drag in >>>>>>>>>> too much baggage IMHO. >>>>>>>>>> >>>>>>>>>> So I'm starting to come to the conclusion that we need to provide such >>>>>>>>>> an implementation as part of the OSGi webbundle support in AS7. The JIRA >>>>>>>>>> is https://issues.jboss.org/browse/AS7-6006 >>>>>>>>>> >>>>>>>>>> Any other ideas? >>>>>>>>>> >>>>>>>>>> Cheers, >>>>>>>>>> >>>>>>>>>> David >>>>>>>>>> _______________________________________________ >>>>>>>>>> jboss-as7-dev mailing list >>>>>>>>>> jboss-as7-dev(a)lists.jboss.org <mailto:jboss-as7-dev@lists.jboss.org> >>>>>>>>>> https://lists.jboss.org/mailman/listinfo/jboss-as7-dev >>>>>> >>>>> >>>>> _______________________________________________ >>>>> jboss-as7-dev mailing list >>>>> jboss-as7-dev(a)lists.jboss.org <mailto:jboss-as7-dev@lists.jboss.org> >>>>> https://lists.jboss.org/mailman/listinfo/jboss-as7-dev >>>> >>> >> >> xxxxxxxxxxxxxxxxxxxxxxxxxxxx >> Thomas Diesler >> JBoss OSGi Lead >> JBoss, a division of Red Hat >> xxxxxxxxxxxxxxxxxxxxxxxxxxxx >> >> >> >> _______________________________________________ >> jboss-as7-dev mailing list >> jboss-as7-dev(a)lists.jboss.org <mailto:jboss-as7-dev@lists.jboss.org> >> https://lists.jboss.org/mailman/listinfo/jboss-as7-dev > > xxxxxxxxxxxxxxxxxxxxxxxxxxxx > Thomas Diesler > JBoss OSGi Lead > JBoss, a division of Red Hat > xxxxxxxxxxxxxxxxxxxxxxxxxxxx > > >