Hi,
So, I've had hellish troubles the past couple of days with memory leaks
in Glassfish/OSGI/Felix. Basically, when I uninstall my Qi4j app not
everything was let go of, and the consequence is that the classloaders
stay in memory, eventually causing a nice OutOfMemory error.
I've found some really bizarre things through debugging this, such as
the MacOSX Preferences API implementation having a Timer thread
referencing the ContextClassLoader that was set when the timer happened
to be started (lazy start->random). This caused the OSGi classloader to
be non-GC'able, and the only way I could get around it was to force this
timer to be started while knowing that there was no context classloader
for the thread. This is just one example, and they are in the same category.
So, having chased these issues one at a time, using VisualVM, heap
dumps, and "Find GC Root", I'm now down to basically no objects being
reachable from any GC root, and yet they're not GC'ed. Very annoying.
This has forced me to start looking into ditching CGLIB entirely. I'm
now looking at ASM, since it can generate the classes we want. The main
problem was figuring out how to use it. But it turns out there's a nice
helper for it: basically write the Java code you want generated, and
then the helper will create Java code to use ASM to create it.
So I've done that and am now trying to generify it so that it works for
any class, apart from the test one I did. The general idea is this: if a
mixin is marked as abstract we need to subclass it and implement all
missing methods. When those methods are called the proxy shall be
invoked. The basic template I'm using looks something like this:
public class SomeMixin_Stub
extends SomeMixin
{
private Object proxy;
public SomeMixin_Stub( Object proxy )
{
super();
this.proxy = proxy;
}
public String other()
{
return ((Other)proxy).other();
}
public String foo( String bar, int x )
{
return ((Other)proxy).foo( bar, x );
}
public void bar( double doub, boolean bool, float fl, char ch )
{
((Other)proxy).bar( doub, bool, fl, ch );
}
}
--
In the above I have a SomeMixin that implements an Other-interface with
methods it's not actually implementing, so they have to be filled in. If
I can generate the above, and do classloading that is OSGi-friendly, all
should be fine. It might even be faster than what we do today. We'll see.
Any thoughts on this? Any gotchas in OSGi I should be aware of while
doing this?
/Rickard
_______________________________________________
qi4j-dev mailing list
[email protected]
http://lists.ops4j.org/mailman/listinfo/qi4j-dev
