Re: PCRegistry ClassLoader memory leak

[Craig L Russell]

On Jul 18, 2007, at 3:51 PM, Marc Prud'hommeaux wrote:

On Jul 18, 2007, at 2:52 PM, Kevan Miller wrote:

On Jul 17, 2007, at 5:32 PM, Marc Prud'hommeaux wrote:

Correct me if I am wrong, but I think the problem is essentially that we have a Map of SomeClass:Object->Object->...->SomeClass, and therefore SomeClass can never be garbage collected, regardless of whether the keys in the Map are soft or weak, and regardless of whether the class' ClassLoader is referenced elsewhere or not. Another possible solution might be to find every place where we have a Class field that could possibly be set to a PersistenceCapable, and turn them all into a WeakReference<Class>. I haven't checked to see how big a change that would be, but it might be one of the simpler solutions (and it might be faster than storing the class name as a String and doing a Class.forName() every time we need access to the Class object).

There are two categories of references that are keeping ClassLoaders alive. (1) References to classes (i.e. Meta fields pcClass, fieldTypes[], and pcSuper). These can all be wrapped fairly easily using WeakReferences. I've fixed these, locally.

Sounds good.

(2) References to objects (i.e Meta.pc). This is a problem. Meta.pc is the only reference to the PersistenceCapable object. Since Meta.pc is the only reference to the PersistenceCapable object, it can't be a WeakReference.

Craig

If the only time the WeakReference will be nullified is when the class is no longer referenced, isn't it ok to use WeakReference? What are the circumstances under which the class is not referenced? If its class loader is garbage collected, or no class references the pc class? The Registry need only keep track of currently used classes. If the class loader decides to unload the class, it's ok to remove it from the registry, as long as when the class loader needs it again it initializes it. It's been a long time since this architecture was put into place but I don't recall any real issues like these...

Any thoughts on fixing the problem with Meta.pc? Does it need to be an object? Can we generate an instance of PersistenceCapable as needed?

That is unfortunate. To my knowledge, the only reason we need to keep around an instance of the object is that it is easy to construct new instances using PersistenceCapable.pcNewInstance() without having to use reflection each time or worry about security. I'm not sure of any way around this.

Or, on a different tack, can PCRegistry be made non-static? Associated with the PU, for instance?

If PCRegistry weren't static, then we wouldn't be able to do this.

I personally think it would be nice if we could get rid of PCRegistry being static. However, this part of the problem with the Meta.pc issue in the previous paragraph: the static initializer of any enhanced persistent class will do: PCRegistry.register(MyClass.class, pcFieldNames, pcFieldTypes, pcFieldFlags, pcPCSuperclass, "MyClass", new MyClass()); Also, I think there are a number of places where we rely on PCRegistry being static for other reasons. I think it would be nice to get rid of these, but that project might be too large a scope just to fix this bug.

As Patrick mentions, another approach is to load OpenJPA in every application classloader (or a unique OpenJPA ClassLoader per application ClassLoader). I really don't want to do this... I'd prefer a private interface to notify you of when ClassLoaders are no longer valid...

--kevan

On Jul 17, 2007, at 10:14 AM, Patrick Linskey wrote:

Of course, another approach is for Geronimo to trash the classloader

that OpenJPA was loaded in when the app is undeployed.

-Patrick

On 7/17/07, robert burrell donkin <[EMAIL PROTECTED]> wrote:

On 7/17/07, Kevan Miller <[EMAIL PROTECTED]> wrote:
>
> On Jul 17, 2007, at 12:15 AM, Craig L Russell wrote:
>
> >
> > On Jul 16, 2007, at 7:48 PM, Kevan Miller wrote:
> >
> >>
> >> On Jul 16, 2007, at 10:26 PM, Pinaki Poddar wrote:
> >>
> >>> Just to clarify:

> >>> Person is
> >>> the persistence capable class, *not*

> >>> ().
> >>>
> >>> But testing in multi-classloader environment is required to
> >>> validate any
> >>> changes of this nature.
> >>

> >> SSN is a field of Person, the ClassLoader of SSN is not

> >> wouldn't work...
> >
> > Just to clarify, if SSN is the type of a field of Person, the

> > loading Person.
>
> Craig,

> of an SSNImpl type which need not be. However, that's not really
> relevant to the problem at hand...
>
> Pinaki,

> keep Classes/ClassLoaders alive...

> >>> I meant it should be the loader that loaded Person.class where > >>> org.apache.openjpa.enhance.PersistenceCapable.class.getClassLoader > >> Yes, that's how I interpreted your statement. My point is that if > >> necessarily the ClassLoader for Person. I'd be concerned that the > >> Person ClassLoader can't load SSN. In that case, your technique > > ClassLoader of Person.class must be able to load SSN (either itself > > or via a parent ClassLoader) or you will have a linkage error while > You are correct. The declared type must be loadable. I was thinking > I'm afraid that I may have improperly narrowed your objectives by > singling out fieldTypes. PCRegistry$Meta.pc and pcSuper could also if there are several places where this may happen then perhaps try a variant on http://svn.apache.org/repos/asf/jakarta/commons/ proper/logging/trunk/src/java/org/apache/commons/logging/impl/ WeakHashtable.java

- robert

--
Patrick Linskey
202 669 5907

Craig Russell
Architect, Sun Java Enterprise System http://java.sun.com/products/jdo
408 276-5638 mailto:[EMAIL PROTECTED]
P.S. A good JDO? O, Gasp!

smime.p7s
Description: S/MIME cryptographic signature