[jsword-devel] OSGi update
DM Smith
dmsmith555 at yahoo.com
Mon Dec 18 17:45:54 MST 2006
Phiillip,
I really appreciate you looking into this.
On Dec 17, 2006, at 2:50 PM, P. R. B. wrote:
> Update and questions from the OSGi front.
>
> Additional general information about OSGi services.
> A service can be associated with arbitrary key/value properties,
> and a service can be found by querying against those properties.
> For example, a Filter service that has a FilterName property of
> "gbf" can be searched for specifically by adding another query
> clause to the service filter string. Also, multiple services that
> implement the same service interface(s) can be retrieved at once. A
> ServiceFactory object can be used in place of service instances to
> allow lazy service instantiation.
>
> Links to OSGi coding tutorials:
> http://oscar-osgi.sourceforge.net/tutorial/
> http://www.knopflerfish.org/osgi_service_tutorial.html
>
> Link to OSGi framework javadocs:
> http://www2.osgi.org/javadoc/r4/index.html
>
> Proposed changes for the OSGi version of Common:
>
> Name the OSGi project org.crosswire.common according to Eclipse
> naming conventions and to reduce the chance of name collisions with
> the bundle in other projects and systems. This would also help us
> distinguish between the OSGi project and any RCP wrapper project we
> create for it.
So if I understand you correctly, Under incubator, you'd create a
folder to hold all the rcp projects, say rcp or jsword-rcp. Under
that folder you would have a project named,
org.crosswire.common
And under there you would hold the common source.
I don't see a problem with this. There is nothing important in the
current project naming convention. With little problem we could
rename the existing projects (only the build.xml and Eclipse project
files know of the project names)
Currently, we are following the project structure defined by Maven as
we were thinking of migrating to using that as a build management
tool. The current layout provides the following (and if we can still
satisfy it via some other mechanism, I'll be happy)
a) Automated build that can be run from the command line on a Linux
platform (currently Fedora Core 6) and also invoked within Eclipse.
The current build system understands the dependency relationships
between projects and ensures that dependencies are processed in the
proper order.
b) Test is separate from the source that it is testing, but the tests
are in the same package as the classes they test. We use JUnit for
testing and this probably should be preserved. The tests can be run
from within the ant build, but currently are turned off. I run them
before I check in code.
c) Resources that drive the application or a part of it but don't
have a particular location are homed together in each project in a
separate resource directory.
>
> Convert the Job class to an interface. Methods that use Job (mostly
> in JSword) will take a Job as a parameter and run in the caller's
> thread. An RCP wrapper would have an adapter to convert an
> IProgressMonitor into a Job. This would allow generic OSGi callers
> and RCP callers to track the job themselves.
I was suprised to find that Job was not an interface. It should be. I
still think that it is fairly important (i.e. I could be pursuaded
otherwise) that the interface should not presume that progress is
being monitored.
>
> Logger - This class will wrap a collection of logger services,
> similar to how Log4J allows multiple appenders. Callers will use
> the class as usual.
Good. That's what this Logger is meant to do.
>
> Converter - This interface will be moved to a new package to
> indicate that it is used for services (appears to be an OSGi
> convention). TransformingSAXEventProviderConverter will not be a
> service because it's constructor takes a parameter (we can modify
> this class or the interface later to make it service-friendly).
> Implementors will register instances along with a name property to
> simplify retrieval.
OK.
>
> Proposed changes for the OSGi version of JSword
>
> Filters, Installers, other plugin/service classes and interfaces:
> These will follow the same principal as Converter above, i.e. move
> the interfaces to packages to indicate they're service interfaces,
> register instances with a name property for retrieval.
I'm not sure I understand this. For example, Filter.java is an
interface and it already is in a package o.c.j.book.Filter is a
package that contains Filter.java and a few support classes. If I
understand correctly OSGi obviates the need for FilterFactory.java.
Msg.java or something that is equivalent for internationalization is
needed.
We have an effort underway to internationalize into German and Farsi.
It would be best to retain as much of that effort as possible. So far
they have not gotten to the JSword properties.
>
> Begin removing calls to ClassUtil since the framework's class
> loaders and service API will reduce the need for and usability of
> the class.
I think this is mostly true. One of the primary purposes of this
class is to produce useful "debug" messages for developers but
presented to users, so they can tell us what the application says the
problem is. As long as the purpose is satisfied, I don't think that
it matters how.
>
> Eclipse / RCP wrapper plugins:
> Do we want to write our own Eclipse / RCP wrapper plugins around
> the Common and JSword bundles? This would allow us to convert
> plugin extensions into service objects ourselves, rather than have
> the RCP application juggle OSGi service metaphors with extension
> ones. I don't know if we're at the point where it matters, but I
> thought it would be worth checking.
I don't have a clue (because of lack of experience/knowledge). My
general principle is to: Consider everything, plan for it, but
postpone implementation as long as possible and then only implement
what is needed.
>
> Vendor bundles:
> We'll need a handful of bundles based on 3rd party code. Where's
> the best place in the incubator for them?
Yes!
>
> Versioning:
> Do we want the bundle versions to adhere to the jar versions?
> Keeping them the same makes synchronizing the code much easier, but
> bundle/plugin updates adhere strongly to the major/minor/service
> numbering changes, which is important since other bundles can
> specify which versions they're compatible with.
So far the points at which we do releases have affected most of our
jars. This is especially true with releases that are further apart. I
think that it is important for it to be understandable that what can
be used together. Having each release revision all jars has been the
easiest way to manage this. Also in a jnlp world the signing
requirements are that the collection of jars need to be signed by the
same signer. I don't think there is a problem with declaring a new
version of a jar or a plugin, even if it has not changed one line of
code.
But as long as it can be managed well, I don't really care.
>
> Thanks,
> -Phillip
>
> "P. R. B." <dysbiote at yahoo.com> wrote:
> DM,
>
> Eclipse 3.1 requires a few more steps to get to OSGi compatibility
> than Eclipse 3.2, but nothing too painful. There are some OSGi
> settings that the 3.2 plugin editor can modify that the 3.1 editor
> can't, but these could always be added in the manifest.mf file
> directly. Equinox is also apparently built into Eclipse 3.1. I have
> no clue what the differences are between versions, but it's
> probably documented on the Equinox site.
>
> Here's a link to the steps to I used to build and test the bundles.
> Nothing pretty. I glossed over some of the details in the first
> section regarding bundle dependencies for brevity. Feel free to
> point out any difficulties you run into following the steps, should
> you decide to try them, of if you think of something you'd like to
> see added to the list.
>
> http://paristano.org/public/osgi_instructions.html
>
> Thanks. =)
> -Phillip
>
> DM Smith <dmsmith555 at yahoo.com> wrote:
> Phillip,
> This is super. So, if I understand correctly, prior to Eclipse 3.2
> working with OSGI bundles is very unfriendly. But the Equinox
> implementation in 3.2 is reasonably supported. If we do go this
> route of adding OSGi as standalone then I think it would be good to
> have the gory details.
>
> I think Logger and Converter are two good examples. In the case of
> Logger, Common provides the behavior and it would be reasonable for
> it to be replace in a particular application.
>
> WRT Converter, the implementations are entirely outside of Common,
> so it would be an optional extension to provide. Optional because
> it is conceived that Common might be used by something other than
> JSword and that something might not do xml transformations.
>
> Keep us posted.
>
> Many Thanks!
> DM
>
> On Dec 12, 2006, at 11:19 PM, P. R. B. wrote:
>
>> Hi all,
>>
>> Okay, here's how far I've gotten so far. First, some quick
>> definitions for this message, to avoid confusion:
>>
>> Bundle: an OSGi-compatible bundle that has no dependency on
>> anything related to a specific implementation of OSGi (e.g.
>> Equinox / Eclipse, Knopflerfish). By the nature of OSGi bundles,
>> there are no extension points for plugins to connect to.
>>
>> Plugin: an Eclipse / RCP plugin. Plugins are bundles, but bundles
>> aren't plugins. Plugins may provide extension points for other
>> plugins, may provide extensions to other plugins' extension
>> points, or both.
>>
>>
>> The executive summary:
>>
>> A new plugin created in Eclipse 3.x can easily be reduced to a
>> bundle. I was able to convert Commons to a simple bundle without
>> coding changes, and I was able to make calls to said bundle's
>> classes from another bundle. The call to Commons required hitting
>> Apache code (which depends on other Apache code, and so on), so
>> bundle dependency chains work as expected. It all ran successfully
>> from within two different OSGi framework implementations.
>>
>> The longer version:
>>
>> My goal for getting a feel for OSGi was to turn the Common jar
>> into a simple bundle, and make a call to its code from a separate
>> bundle. The OSGi spec for a bundle's manifest file is difficult to
>> manage manually, so I thought I'd find an Eclipse plugin that
>> could simplify this for me, moreso than Eclipse's built-in plugin
>> editor. My first go was with Knopflerfish (http://
>> www.knopflerfish.org). They provide an Eclipse plugin to create
>> custom bundle projects and a UI-based OSGi framework. I was
>> discouraged by a couple things in the plugin. First, a bundle
>> isn't exported correctly from a bundle project using the standard
>> "Export..." action; it has to be exported through a button in a
>> custom editor tied to a bundle.manifest file. Second, I couldn't
>> find a way to make one bundle completely dependent on another like
>> I can when working with plugins (the functionality may be there,
>> but I couldn't find it). The OSGi framework itself is slick, but
>> the plugin didn't help me much.
>>
>> Next I looked at Equinox, which is what Eclipse (now) uses for its
>> OSGi framework implementation. Eclipse 3.2 allows a plugin to be
>> created that targets Equinox. In other words, it creates a bundle
>> and not a plugin, and all plugin-related editors also handle
>> bundle properties. Using Eclipse 3.1 to create a bundle required a
>> little more work. I had to modify the new plugin's manifest and
>> code to remove its dependency on the org.eclipse.osgi bundle. The
>> changes were minor, but worth noting.
>>
>> I was able to create the Common bundle in Eclipse this way, along
>> with a test bundle that would call some Common code when the test
>> bundle was activated. I had already created plugins for the
>> various jars that Common depends on. These plugins were simple
>> enough to qualify as bundles (no extensions, no plugin.xml, etc),
>> so I used them to fulfill the Common bundle's dependencies. I took
>> this batch of bundles and loaded them into the Equinox command
>> line OSGi console, and the test bundle made its call successfully
>> when it started up. I did the same with the bundles in the
>> Knopflerfish console.
>>
>> Going forward, I'll look at the Common code to determine what
>> classes need to be changed to make them bundle- and plugin-
>> friendly. Logger and Converter are the first two that come to
>> mind. I imagine we'd want a plugin wrapper to be able to use its
>> own log for output and provide Converter extension points to allow
>> other plugins to add converters. I don't know whether these can be
>> done in a practical manner or not yet. I'll also consider the
>> classes that DM listed. I'll post what I come up with.
>>
>> Let me know if anyone is interested in knowing the specific steps
>> to create the bundles and test them. I can write them up and post
>> a link to them here.
>>
>> Thanks,
>> -Phillip
>>
>>
>> Check out the all-new Yahoo! Mail beta - Fire up a more powerful
>> email and get things done faster.
>> _______________________________________________
>> jsword-devel mailing list
>> jsword-devel at crosswire.org
>> http://www.crosswire.org/mailman/listinfo/jsword-devel
>
> _______________________________________________
> jsword-devel mailing list
> jsword-devel at crosswire.org
> http://www.crosswire.org/mailman/listinfo/jsword-devel
>
> Access over 1 million songs - Yahoo! Music
> Unlimited._______________________________________________
> jsword-devel mailing list
> jsword-devel at crosswire.org
> http://www.crosswire.org/mailman/listinfo/jsword-devel
>
>
> Everyone is raving about the all-new Yahoo! Mail beta.
> _______________________________________________
> jsword-devel mailing list
> jsword-devel at crosswire.org
> http://www.crosswire.org/mailman/listinfo/jsword-devel
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.crosswire.org/pipermail/jsword-devel/attachments/20061218/fe7fb431/attachment-0001.html
More information about the jsword-devel
mailing list