|
Description:
|
Whenever getInstanceFromFactory() is called and a factory is used, a (application wide) lock is used: factoryLock.
In my environments the Factory uses expensive calls to the backend to create an object, and during this time no other Factory method is allowed to be called. This is a severe bottleneck in a multiuser environment.
This lock is important for APPLICATION scoped components, but other components that are i.e. CONVERSATION scoped, Seam already ensures that only one Thread has access to the current conversation. The same should be true for PAGE and EVENT. In this cases no lock should be acquired. UPDATE: Factories that use injection should not be called multithreaded, therefore an updated condition.
See the follwing code snippet. I also attach a patch.
I tested the patch for Seam 2.2.2.Final - I would be most happy to have this included in 2.2 and 2.3 branch.
Br Alexander.
{code:title=Component.java}
ScopeType scopeResult = getOutScope(factoryMethod.getScope(),
factoryMethod.getComponent());
ScopeType scopeFactory = factoryMethod.getComponent().getScope();
boolean lockingNeeded = false;
/*
* we need this lock in the following cases: (1) the target scope is
* accessed by more than one thread (as we don't want to create the same
* object by two threads at the same time for the same scope) OR (2) the
* factory is accessed by more than one thread and is using interceptors
* (as accessing a factory multiple times might mess up interceptors).
* This assumes that (1) the scopes CONVERSATION, EVENT and PAGE can't
* be accessed by more than one thread anyway due to CONVERSATION being
* locked for the current thread anyway, and EVENT and PAGE being only
* short-lived anyway. (2) a factory that doesn't use injection can be
* accessed multi threaded. See JBSEAM-4669/JBSEAM-2419 for the original
* discussion.
*/
if ((scopeResult != ScopeType.CONVERSATION
&& scopeResult != ScopeType.EVENT && scopeResult != ScopeType.PAGE)
|| (scopeFactory != ScopeType.CONVERSATION
&& scopeFactory != ScopeType.EVENT
&& scopeFactory != ScopeType.PAGE && factoryMethod
.getComponent().interceptionEnabled)) {
lockingNeeded = true;
}
if (lockingNeeded) {
/*
* synchronize
only on
all instances of this factory (they might outject to
the
* same scope, i.e.
factory
instance
in EVENT scope
,
outjecting to APPLICATION scope),
*
let the rest of the world
*
continue.
*/
synchronized (factory
.getClass(
)
)
{
return createInstanceFromFactory(name, scope, factoryMethod,
factory);
}
} else {
return createInstanceFromFactory(name, scope, factoryMethod, factory);
}
[...]
private static Object createInstanceFromFactory(String name, ScopeType scope,
Init.FactoryMethod factoryMethod, Object factory) {
// check whether there has been created an instance by another thread
// while waiting for this function's lock
if (scope != STATELESS) {
Object value = (scope == null) ? Contexts.lookupInStatefulContexts(name)
: scope.getContext().get(name);
if (value != null) {
return value;
}
}
if (factory == null) {
return null;
} else {
Object result = factoryMethod.getComponent().callComponentMethod(factory,
factoryMethod.getMethod());
return handleFactoryMethodResult(name, factoryMethod.getComponent(),
result, factoryMethod.getScope());
}
}
{code}
|
Hello Marek Schmidt,
thank you for describing this scenario, and also for the adding the test case.
By now locking on the factory's class, this now ensures that the different instances of the same factory in different contexts can't be called at the same time. This now fixes the test cases you described.
$ mvn clean integration-test -Dtest=ConcurrentFactoryTest,FactoryLockTest
I don't know how much the performance can be improved by determining when to lock the instance or the class - for now I want to keep it simple and don't add another if. Please let me know if I should update the patch.
Best regards,
Alexander.