Github user zhijiangW commented on a diff in the pull request: https://github.com/apache/flink/pull/4509#discussion_r152512614 --- Diff: flink-runtime/src/test/java/org/apache/flink/runtime/io/network/partition/consumer/RemoteInputChannelTest.java --- @@ -301,81 +306,388 @@ public void testProducerFailedException() throws Exception { } /** - * Tests {@link RemoteInputChannel#recycle(MemorySegment)}, verifying the exclusive segment is - * recycled to available buffers directly and it triggers notify of announced credit. + * Tests to verify that the input channel requests floating buffers from buffer pool + * in order to maintain backlog + initialCredit buffers available once receiving the + * sender's backlog, and registers as listener if no floating buffers available. */ @Test - public void testRecycleExclusiveBufferBeforeReleased() throws Exception { - final SingleInputGate inputGate = mock(SingleInputGate.class); - final RemoteInputChannel inputChannel = spy(createRemoteInputChannel(inputGate)); + public void testRequestFloatingBufferOnSenderBacklog() throws Exception { + // Setup + final NetworkBufferPool networkBufferPool = new NetworkBufferPool(12, 32, MemoryType.HEAP); + final SingleInputGate inputGate = createSingleInputGate(); + final RemoteInputChannel inputChannel = createRemoteInputChannel(inputGate); + try { + final int numFloatingBuffers = 10; + final BufferPool bufferPool = spy(networkBufferPool.createBufferPool(numFloatingBuffers, numFloatingBuffers)); + inputGate.setBufferPool(bufferPool); + + // Assign exclusive segments to the channel + final int numExclusiveBuffers = 2; + inputGate.setInputChannel(inputChannel.partitionId.getPartitionId(), inputChannel); + inputGate.assignExclusiveSegments(networkBufferPool, numExclusiveBuffers); + + assertEquals("There should be " + numExclusiveBuffers + " buffers available in the channel", + numExclusiveBuffers, inputChannel.getNumberOfAvailableBuffers()); - // Recycle exclusive segment - inputChannel.recycle(MemorySegmentFactory.allocateUnpooledSegment(1024, inputChannel)); + // Receive the producer's backlog + inputChannel.onSenderBacklog(8); - assertEquals("There should be one buffer available after recycle.", - 1, inputChannel.getNumberOfAvailableBuffers()); - verify(inputChannel, times(1)).notifyCreditAvailable(); + // Request the number of floating buffers by the formula of backlog + initialCredit - availableBuffers + verify(bufferPool, times(8)).requestBuffer(); + verify(bufferPool, times(0)).addBufferListener(inputChannel); + assertEquals("There should be 10 buffers available in the channel", + 10, inputChannel.getNumberOfAvailableBuffers()); - inputChannel.recycle(MemorySegmentFactory.allocateUnpooledSegment(1024, inputChannel)); + inputChannel.onSenderBacklog(11); - assertEquals("There should be two buffers available after recycle.", - 2, inputChannel.getNumberOfAvailableBuffers()); - // It should be called only once when increased from zero. - verify(inputChannel, times(1)).notifyCreditAvailable(); + // Need extra three floating buffers, but only two buffers available in buffer pool, register as listener as a result + verify(bufferPool, times(11)).requestBuffer(); + verify(bufferPool, times(1)).addBufferListener(inputChannel); + assertEquals("There should be 12 buffers available in the channel", + 12, inputChannel.getNumberOfAvailableBuffers()); + + inputChannel.onSenderBacklog(12); + + // Already in the status of waiting for buffers and will not request any more + verify(bufferPool, times(11)).requestBuffer(); + verify(bufferPool, times(1)).addBufferListener(inputChannel); + --- End diff -- Actually I tried to test the two logics in two separate tests `testRequestFloatingBufferOnSenderBacklog` and `testFairDistributionFloatingBuffers`. For `testRequestFloatingBufferOnSenderBacklog`, it only wants to verify the request logic on input channel side. The key point is that the input channel will not request repeated if it is already as listener in pool. For 'testFairDistributionFloatingBuffers`, it only wants to verify that the input channel listener is getting buffer fairly during buffer `recycle()` on `bufferPool` side. I think it can cover the comment you mentioned "Can you also verify the behaviour when the buffers become available?" I will further check that later. I agree with that we missed the tests to verify that we stick to `senderBacklog + initialCredit` in different processes. And I will add them later.
---