Under heavy load and sharing an EJB proxy between different threads, there's a
chance of NullPointerException like the one below being reported, specially when
using proxies configured with RoundRobin as load balance policy, such as SLSBs
or EJB2 home proxies:

java.lang.NullPointerException
at org.jboss.invocation.unified.interfaces.UnifiedInvokerHAProxy.invoke(UnifiedInvokerHAProxy.java:186)
at org.jboss.invocation.InvokerInterceptor.invokeInvoker(InvokerInterceptor.java:365)
at org.jboss.invocation.MarshallingInvokerInterceptor.invoke(MarshallingInvokerInterceptor.java:63)
at org.jboss.proxy.ejb.RetryInterceptor.invoke(RetryInterceptor.java:176)
at org.jboss.proxy.TransactionInterceptor.invoke(TransactionInterceptor.java:61)
at org.jboss.proxy.SecurityInterceptor.invoke(SecurityInterceptor.java:70)
at org.jboss.proxy.ejb.HomeInterceptor.invoke(HomeInterceptor.java:184)
at org.jboss.proxy.ClientContainer.invoke(ClientContainer.java:100)
at $Proxy1.create(Unknown Source)

Whenever a new target for an invocation is chosen, init(InvokerLocator locator)
method in UnifiedInvokerProxy is executed:

protected void init(InvokerLocator locator)
{
this.locator = locator;

try

{ client = createClient(locator, getSubSystem()); client.connect(); }

catch(Exception e)

{ log.fatal("Could not initialize UnifiedInvokerProxy.", e); }

}

This method could be executed in paralell to the following in
UnifiedInvokerHAProxy.invoke(Invocation invocation):

Client clientInstance = getClient(invocation);
log.debug("Making invocation on " + clientInstance.getInvoker().getLocator());

When the new client instance is created, the invoker is null until client.connect() is
called, so something like this can happen:

1.- T1 (Thread 1) executes getClient(invocation); and returns client#1
2.- T2 (Thread 2) selects a new target for the invocation and calls init(InvokerLocator locator)
3.- T2 calls createClient() and the instance variable client becomes client#2
4.- T1 now calls clientInstance.getInvoker().getLocator() which is now client#2 but
connect() has not been called on this instance, so results on NPE.

The simplest of all solutions would be to make assignment of client instance thread safe,
something I already did for JBAS-4815 to get around another thread safety issue.

I'm gonna take the tests I created for JBAS-4815 and use them to replicate something like
this and check whether the fix for that fixes this as well.