[JAMES-3724] Lifecycle API should embed a leak detector

• Type: Improvement
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 3.7.0
• Fix Version/s: 3.8.0
• Component/s: James Core
• Assignee: Unassigned
• Reporter: Benoit Tellier
• Created: 06/Mar/22 07:43
• Updated: 04/May/22 20:52
• Resolved: 16/Mar/22 01:35

Description

Why ?

JAMES-3720 showed how easy to miss calling Lifecycle::dispose on email objects, which causes potentially temporary file leaks. If core developer / core components are subject to leaks while core contributors are aware of the importance of life-cycle management of mail objects, I cannot but be worry of users writing extensions and messing up.

Goals

Thus I would like a defense mechanism that:

• Log errors if an email is not disposed, and ideally contains a stacktrace to where that email was created to ease debug.
• GC should eventually reclaim those temporary files, this is better than leaking them.

Solution overview

I have in mind something much similar to Netty 4 pooled buffer leak detection:

• Rely on phantom references to track GCed items
• Fill a stack trace upon object creation
• Upon GC verify that resources were released, if not do it and warn the user about it.

References:

• About phantom references: https://www.baeldung.com/java-phantom-reference
• Netty resource leak detector: https://github.com/netty/netty/blob/4.1/common/src/main/java/io/netty/util/ResourceLeakDetector.java

API overview

Use system properties to configure leak detection:

• `-Djames.ligecycle.leak.detection.mode=none` would behave as today.
• `-Djames.ligecycle.leak.detection.mode=simple` would output a simple ERROR log if a leak is encountered and would free the resources.
• `-Djames.ligecycle.leak.detection.mode=advanced` would output an extented ERROR log implying the place of allocation of the underlying object and would free resources.
• `-Djames.ligecycle.leak.detection.mode=testing` would output an extented ERROR log implying the place of allocation of the underlying object and rethrow an error, unsuring test fails and that action is being taken by the development team.

`jvm.properties` files and documentation would document this setting.

Also we IMO should default to `simple`.

Implementation details

Provide an abstract class, Disposable.LeakAware:

public interface Disposable {

    /**
     * Dispose the object
     */
    void dispose();

    static abstract class LeakAware implements Disposable {
        private static final ReferenceQueue<Object> REFERENCE_QUEUE = new ReferenceQueue<>();

        private Atomic<Boolean> isDisposed:

        protected LeakAware() {
            isDisposed = new AtomicBoolean(false);
        }

        public void track() {
             // Create a phantom reference, of course if detection mode is none
             new LeakAwareFinalizer(this, REFERENCE_QUEUE)

             // POLL the reference queue here to check for leaks!
             // of course don't do it if the leak detection mode is none
             while ((referenceFromQueue = REFERENCE_QUEUE.poll()) != null) {
                 ((LeakAwareFinalizer)referenceFromQueue).detectLeak();
                 referenceFromQueue.clear();
             }
        }
    }

    public class LeakAwareFinalizer extends PhantomReference<LeakAware> {

        public LeakAwareFinalizer(LeakAware referent, ReferenceQueue<? super LeakAware> q) {
            super(referent, q);
        }

        public void detectLeak() {
            if (!referent.isDisposed()) {
                // add starttrace of creation if needed
                LOGGER.error("Leak detected!!!");
                // clean if needed
                referent.dispose();
            }
        }
    }
}

Then sensitive classes, likely only `MimeMessageInputStreamSource` would extend it, and thus automatically gain leak detection.

We can turn `MimeMessageSource` into an interface with a default size method if needed.

Acceptance criteria

• Run tests in testing leak detection mode. This can be done by passing system properties to the surefire plugin.
• Unit tests including logging checks for each mode.
• See this for logging assertions: PeriodicalHealthCheckTest
• Call System.gc() to force a GC and get the leak detection running
• Manage dynamically system variables in unit tests to test each mode.
• Verify release was called (via atomic boolean in a custom Disposable.LeakAware object?)
• Document this feature.

Issue Links

• is related to: JAMES-3759 Leak Detector is not running in some tests
• links to: GitHub Pull Request #911
• links to: GitHub Pull Request #913