[timfox]

Ok, I wrote a simple test case that sends and consumes messages (10K) in size. Can't see any leak.

I'm going to wait until someone creates a real JUnit test case that replicates the problem since I'm just wasting my time here.

[clebert.suconic]

Procedures to replicate the issue:

- Using Trunk

- build a distribution

(You will need maven on the PATH, otherwise it's failing ATM)

- Unzip tx-leaks under ./build/hornetq-2.2.0.CR1/examples/soak

- Change the run.sh, make MaxMemory to 2G, and add verbose:gc

(I have attached a version of run.sh just for your convenience here)

- Start the server with this option:

./run.sh /path-to-build-output/hornetq-2.2.0.CR1/examples/soak/tx-leaks/server0

- Run tx-leaks with

./build.sh runRemote

It should fail in 10 minutes

The same will never fail with Branch_2_1. (I have tried it with 512M maxMemory)

It seems something gets into a redelivery loop until all the memory is consumed.

Notice the test is using diverts. And I have disabled paging for this test.

[clebert.suconic]

@Tim: BTW: isn't this task going to break JDK 1.5 support? It seems you replaced a class that's only available on jdk 1.6+. Maybe it would be better to keep it as is on the client classes?

[clebert.suconic]

Actually, I just committed org.hornetq.tests.soak.client.ClientSoakTest

It would be interesting to add -verbose:gc to your VM options when running the test on Eclipse (run->RunConfiguration->ClientsoakTest).. (You will need to run it at least once before adding the option).

The same test will OME on trunk, and it will run forever on Branch_2_1.

I'm not sure if the divert used here will have any play on the issue. I'm just reproducing the exact same case I had on /soak/tx-leaks

[clebert.suconic]

Nothing to do with Diverts.

I've added org.hornetq.tests.soak.client.ClientNonDivertedSoakTest which also fails on trunk.

This test is doing inVM, without diverts.

[clebert.suconic]

@Tim: We talked earlier that a possible fix on LinkedListImpl is to change removeAfter like this:

   private void removeAfter(Node<E> after)

   {

      Node<E> toRemove = after.next;

      after.next = toRemove.next;

      if (toRemove.next != null)

      {

         toRemove.next.prev = after;

      }

      if (toRemove == tail)

      {

         tail = after;

      }

      size--;

      if (toRemove.iterCount != 0)

      {

         LinkedListImpl.this.nudgeIterators(toRemove);

      }

     

     toRemove.next = null;

     toRemove.val = null;

     toRemove.prev = null;

   }

   private void removeAfter(Node<E> after)

   {

      Node<E> toRemove = after.next;

      after.next = toRemove.next;

      if (toRemove.next != null)

      {

         toRemove.next.prev = after;

      }

      if (toRemove == tail)

      {

         tail = after;

      }

      size--;

      if (toRemove.iterCount != 0)

      {

         LinkedListImpl.this.nudgeIterators(toRemove);

      }

     toRemove.next = null;

     toRemove.val = null;

     toRemove.prev = null;

   }

i.e., setting those fields to null at the end of the method.

I probably misunderstood the change you asked me to do, as all the tests on org.hornetq.tests.integration.client,PagingTest are failing after I do that. Something related to receve(TIMEOUT) returning null (missing messages) (not being related to paging itself of course).

So, I just wanted to point you that. If you run the testsuite and it's passing, then you're safe (and I probably misunderstood the fix).

[clebert.suconic]

I also tried to set .val to null what made the LInkedListImplTest to pass, however the SoakTest was failing.

Besides, I also added a new test that would fail if toRemove.val was set to true

(Look at SimpleSendReceiveSoakTest, which is a simple test, very similar to the LinkedListTest one, always keeping 1000 elements on the queue and sending and consuming elements)

[timfox]

Clebert Suconic wrote:

(Look at SimpleSendReceiveSoakTest, which is a simple test, very similar to the LinkedListTest one, always keeping 1000 elements on the queue and sending and consuming elements)

I guess it was a typo, but this test is calling commit(), but the session is auto commit.

[timfox]

The tests seem to pass if I do:

private void removeAfter(Node<E> after)

   {

      Node<E> toRemove = after.next;

      after.next = toRemove.next;

      if (toRemove.next != null)

      {

         toRemove.next.prev = after;

      }

      if (toRemove == tail)

      {

         tail = after;

      }

      size--;

      if (toRemove.iterCount != 0)

      {

         LinkedListImpl.this.nudgeIterators(toRemove);

      }

     toRemove.next = null;

     toRemove.prev = null;

   }

[timfox]

This fix seems to work, tests pass etc and is committed.

I am currently fixing a related intermittent ordering issue, should be done by end of day.

[timfox]

This should be fixed in TRUNK now.

I'd like to see this merged into the 2_1 branch so it's in the next release for EAP too.