You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@cayenne.apache.org by Andrus Adamchik <an...@objectstyle.org> on 2007/02/20 13:31:23 UTC
JPA enhancer behavior - fat agent?
Dain Sundstrom (who is a member of Geronimo project) did some
preliminary testing of Cayenne JPA provider in the OpenEJB container
last week - you've seen some jiras that he opened. I am taking
further discussion to the list (not sure if Dain is subscribed, so I
am cc'ying to him as well).
So the enhancer... I tried to run the OpenEJB JpaTest [1] with
Cayenne provider. It fails to enhance the Employee object [2], even
if it is explicitly mentioned in the PersistenceUnit, because it was
loaded in the JVM *before* Cayenne provider was started and
Provider.createContainerEntityManagerFactory(..) was called.
I was trying to think of ways to "fix" it, but considering that
java.lang.instrument.Instrumentation doesn't allow to enhance already
loaded classes ("redefineClasses" method places serious limitations
on what can be done, namely adding new fields is not allowed), looks
like JPA users will have to deal with specific startup order
requirements. In-container and Java SE operation modes will look
different:
1. In-container: provider shouldn't directly access its own agent,
and simply register a needed ClassTransformer with container-provided
PersistenceUnitInfo. Container on the other hand has the
responsibility to hook up PersistenceUnitInfo implementor to the
agent, and ensure that provider is started *before* and persistence
classes are loaded in the app class loader. So I wonder whether the
fact that JpaTest worked with OpenJPA provider by some coincidence?
2. Standalone: for the enhancer to work, an application must ensure a
call to Persistence.createEntityManagerFactory() prior to the entity
classes loaded in runtime. This sucks, and the only way around is to
run enhancer straight from the agent. I think that's what OpenJPA
does. Some time ago we decided [3] to avoid putting any Cayenne
runtime classes in the agent, but it looks like we don't have a
choice. Unless somebody can suggest a better way of providing
reliable runtime enhancement, I will change the behavior to provide a
"fat" agent.
[1] https://svn.apache.org/repos/asf/incubator/openejb/trunk/openejb3/
container/openejb-core/src/test/java/org/apache/openejb/core/cmp/jpa/
JpaTest.java
[2] https://svn.apache.org/repos/asf/incubator/openejb/trunk/openejb3/
container/openejb-core/src/test/java/org/apache/openejb/core/cmp/jpa/
Employee.java
[3] http://objectstyle.org/cayenne/lists/cayenne-devel/2006/08/0088.html
Andrus
P.S. Dain - are there any other show stopper issues in your testing
(except for no EJBQL)?
Re: JPA enhancer behavior - fat agent?
Posted by Craig L Russell <Cr...@Sun.COM>.
+1 to what Dain says here. All the persistence providers that I'm
familiar with use the -javaagent on the java command line in Java SE
mode to do the instrumentation (bytecode enhancement) regardless of
the load order of the persistent classes and the persistence provider.
Craig
On Feb 22, 2007, at 10:45 AM, Dain Sundstrom wrote:
> You can't really push a class back into the app. Instead you must
> wait for the app to load (pull) a class and hook the load via the
> javaagent.
>
> In a managed (container) environment this works quite well because
> the container can guarantee that the JPA provider is full
> initialized before any app code is called. In an unmanaged
> environment life is much more difficult since it is by definition
> the application code that initializes the system. To get around
> this, you will need the javaagent to search for persistence units
> and determine which classes need to be enhanced. One nice thing,
> is most people will use the @Entity annotation instead of XML to
> declare their persistent beans, and this entity is easy to detect
> directly using asm. You should ask David Blevins about his xbean-
> finder code which is designed to detect such annotations.
>
> -dain
>
> On Feb 22, 2007, at 10:21 AM, Andrus Adamchik wrote:
>
>> I see how TemporaryClassLoader operates. We haven't been using
>> 'getNewTempClassLoader' in the provider code as of yet. Let me
>> poke around and see if I am overcomplicating the issue... I still
>> don't understand how to push an enhanced class back into the app
>> classloader space, but maybe I am wrong (still thinking in Cayenne
>> classic terms), and this is not needed.
>>
>> Andrus
>>
>>
>> On Feb 22, 2007, at 7:46 PM, Dain Sundstrom wrote:
>>> Sorry, got busy and missed this email.
>>>
>>> This is such a common problem, it was addressed directly by the
>>> JPA specs. When loading classes for inspection use the class
>>> loader returned from
>>> javax.persistence.spi.PersistenceUnitInfo.getNewTempClassLoader().
>>>
>>> For Cayenne's own PersistenceUnitInfo implementation, I suggest
>>> you copy org.apache.openejb.core.TemporaryClassLoader which I
>>> copied from OpenJPA and cleaned up.
>>>
>>> -dain
>>
>
Craig Russell
Architect, Sun Java Enterprise System http://java.sun.com/products/jdo
408 276-5638 mailto:Craig.Russell@sun.com
P.S. A good JDO? O, Gasp!
Re: JPA enhancer behavior - fat agent?
Posted by Dain Sundstrom <da...@iq80.com>.
On Feb 23, 2007, at 3:11 AM, Andrus Adamchik wrote:
> BTW, this discussion happened to be very helpful in straightening
> Cayenne PersistentUnitInfo ClassLoaders. We had our own version of
> TemporaryClassLoader from the beginning, but is was used as the
> main PersistentUnitInfo class loader, not the temp one (and
> provider incorrectly called PersistentUnitInfo.getClassLoader() as
> a result). I fixed it and got further down with OpenEJB JpaTest.
> Enhancer now works. I'll see if I can address other errors.
I suggest you compare you temp class loader to the one I pointed
out. It is very easy to get the code wrong (I did my first two
attempts at it).
-dain
Re: JPA enhancer behavior - fat agent?
Posted by Andrus Adamchik <an...@objectstyle.org>.
Ok, that matches my assessment - we need a JPA-aware agent (what I
called fat agent).
BTW, this discussion happened to be very helpful in straightening
Cayenne PersistentUnitInfo ClassLoaders. We had our own version of
TemporaryClassLoader from the beginning, but is was used as the main
PersistentUnitInfo class loader, not the temp one (and provider
incorrectly called PersistentUnitInfo.getClassLoader() as a result).
I fixed it and got further down with OpenEJB JpaTest. Enhancer now
works. I'll see if I can address other errors.
> You should ask David Blevins about his xbean-finder code which is
> designed to detect such annotations.
I'll check it out.
Andrus
On Feb 22, 2007, at 8:45 PM, Dain Sundstrom wrote:
> You can't really push a class back into the app. Instead you must
> wait for the app to load (pull) a class and hook the load via the
> javaagent.
>
> In a managed (container) environment this works quite well because
> the container can guarantee that the JPA provider is full
> initialized before any app code is called. In an unmanaged
> environment life is much more difficult since it is by definition
> the application code that initializes the system. To get around
> this, you will need the javaagent to search for persistence units
> and determine which classes need to be enhanced. One nice thing,
> is most people will use the @Entity annotation instead of XML to
> declare their persistent beans, and this entity is easy to detect
> directly using asm. You should ask David Blevins about his xbean-
> finder code which is designed to detect such annotations.
>
> -dain
>
> On Feb 22, 2007, at 10:21 AM, Andrus Adamchik wrote:
>
>> I see how TemporaryClassLoader operates. We haven't been using
>> 'getNewTempClassLoader' in the provider code as of yet. Let me
>> poke around and see if I am overcomplicating the issue... I still
>> don't understand how to push an enhanced class back into the app
>> classloader space, but maybe I am wrong (still thinking in Cayenne
>> classic terms), and this is not needed.
>>
>> Andrus
>>
>>
>> On Feb 22, 2007, at 7:46 PM, Dain Sundstrom wrote:
>>> Sorry, got busy and missed this email.
>>>
>>> This is such a common problem, it was addressed directly by the
>>> JPA specs. When loading classes for inspection use the class
>>> loader returned from
>>> javax.persistence.spi.PersistenceUnitInfo.getNewTempClassLoader().
>>>
>>> For Cayenne's own PersistenceUnitInfo implementation, I suggest
>>> you copy org.apache.openejb.core.TemporaryClassLoader which I
>>> copied from OpenJPA and cleaned up.
>>>
>>> -dain
>>
>
>
Re: JPA enhancer behavior - fat agent?
Posted by Dain Sundstrom <da...@iq80.com>.
You can't really push a class back into the app. Instead you must
wait for the app to load (pull) a class and hook the load via the
javaagent.
In a managed (container) environment this works quite well because
the container can guarantee that the JPA provider is full initialized
before any app code is called. In an unmanaged environment life is
much more difficult since it is by definition the application code
that initializes the system. To get around this, you will need the
javaagent to search for persistence units and determine which classes
need to be enhanced. One nice thing, is most people will use the
@Entity annotation instead of XML to declare their persistent beans,
and this entity is easy to detect directly using asm. You should ask
David Blevins about his xbean-finder code which is designed to detect
such annotations.
-dain
On Feb 22, 2007, at 10:21 AM, Andrus Adamchik wrote:
> I see how TemporaryClassLoader operates. We haven't been using
> 'getNewTempClassLoader' in the provider code as of yet. Let me poke
> around and see if I am overcomplicating the issue... I still don't
> understand how to push an enhanced class back into the app
> classloader space, but maybe I am wrong (still thinking in Cayenne
> classic terms), and this is not needed.
>
> Andrus
>
>
> On Feb 22, 2007, at 7:46 PM, Dain Sundstrom wrote:
>> Sorry, got busy and missed this email.
>>
>> This is such a common problem, it was addressed directly by the
>> JPA specs. When loading classes for inspection use the class
>> loader returned from
>> javax.persistence.spi.PersistenceUnitInfo.getNewTempClassLoader().
>>
>> For Cayenne's own PersistenceUnitInfo implementation, I suggest
>> you copy org.apache.openejb.core.TemporaryClassLoader which I
>> copied from OpenJPA and cleaned up.
>>
>> -dain
>
Re: JPA enhancer behavior - fat agent?
Posted by Andrus Adamchik <an...@objectstyle.org>.
I see how TemporaryClassLoader operates. We haven't been using
'getNewTempClassLoader' in the provider code as of yet. Let me poke
around and see if I am overcomplicating the issue... I still don't
understand how to push an enhanced class back into the app
classloader space, but maybe I am wrong (still thinking in Cayenne
classic terms), and this is not needed.
Andrus
On Feb 22, 2007, at 7:46 PM, Dain Sundstrom wrote:
> Sorry, got busy and missed this email.
>
> This is such a common problem, it was addressed directly by the JPA
> specs. When loading classes for inspection use the class loader
> returned from
> javax.persistence.spi.PersistenceUnitInfo.getNewTempClassLoader().
>
> For Cayenne's own PersistenceUnitInfo implementation, I suggest you
> copy org.apache.openejb.core.TemporaryClassLoader which I copied
> from OpenJPA and cleaned up.
>
> -dain
Re: JPA enhancer behavior - fat agent?
Posted by Dain Sundstrom <da...@iq80.com>.
Sorry, got busy and missed this email.
This is such a common problem, it was addressed directly by the JPA
specs. When loading classes for inspection use the class loader
returned from
javax.persistence.spi.PersistenceUnitInfo.getNewTempClassLoader().
For Cayenne's own PersistenceUnitInfo implementation, I suggest you
copy org.apache.openejb.core.TemporaryClassLoader which I copied from
OpenJPA and cleaned up.
-dain
On Feb 20, 2007, at 4:31 AM, Andrus Adamchik wrote:
> Dain Sundstrom (who is a member of Geronimo project) did some
> preliminary testing of Cayenne JPA provider in the OpenEJB
> container last week - you've seen some jiras that he opened. I am
> taking further discussion to the list (not sure if Dain is
> subscribed, so I am cc'ying to him as well).
>
> So the enhancer... I tried to run the OpenEJB JpaTest [1] with
> Cayenne provider. It fails to enhance the Employee object [2], even
> if it is explicitly mentioned in the PersistenceUnit, because it
> was loaded in the JVM *before* Cayenne provider was started and
> Provider.createContainerEntityManagerFactory(..) was called.
>
> I was trying to think of ways to "fix" it, but considering that
> java.lang.instrument.Instrumentation doesn't allow to enhance
> already loaded classes ("redefineClasses" method places serious
> limitations on what can be done, namely adding new fields is not
> allowed), looks like JPA users will have to deal with specific
> startup order requirements. In-container and Java SE operation
> modes will look different:
>
> 1. In-container: provider shouldn't directly access its own agent,
> and simply register a needed ClassTransformer with container-
> provided PersistenceUnitInfo. Container on the other hand has the
> responsibility to hook up PersistenceUnitInfo implementor to the
> agent, and ensure that provider is started *before* and persistence
> classes are loaded in the app class loader. So I wonder whether the
> fact that JpaTest worked with OpenJPA provider by some coincidence?
>
> 2. Standalone: for the enhancer to work, an application must ensure
> a call to Persistence.createEntityManagerFactory() prior to the
> entity classes loaded in runtime. This sucks, and the only way
> around is to run enhancer straight from the agent. I think that's
> what OpenJPA does. Some time ago we decided [3] to avoid putting
> any Cayenne runtime classes in the agent, but it looks like we
> don't have a choice. Unless somebody can suggest a better way of
> providing reliable runtime enhancement, I will change the behavior
> to provide a "fat" agent.
>
> [1] https://svn.apache.org/repos/asf/incubator/openejb/trunk/
> openejb3/container/openejb-core/src/test/java/org/apache/openejb/
> core/cmp/jpa/JpaTest.java
> [2] https://svn.apache.org/repos/asf/incubator/openejb/trunk/
> openejb3/container/openejb-core/src/test/java/org/apache/openejb/
> core/cmp/jpa/Employee.java
> [3] http://objectstyle.org/cayenne/lists/cayenne-devel/
> 2006/08/0088.html
>
> Andrus
>
> P.S. Dain - are there any other show stopper issues in your testing
> (except for no EJBQL)?