Slice front-end request messages
[controller.git] / opendaylight / md-sal / cds-access-client / src / main / java / org / opendaylight / controller / cluster / access / client / TransmitQueue.java
index 4a1b3a2f293a499a64e805250a5a3776222caa67..0313a72a8319fc107a967a2bfbb7c188b4a08ac4 100644 (file)
@@ -8,11 +8,13 @@
 package org.opendaylight.controller.cluster.access.client;
 
 import akka.actor.ActorRef;
+import com.google.common.annotations.VisibleForTesting;
 import com.google.common.base.Preconditions;
 import com.google.common.base.Verify;
-import com.google.common.collect.Iterables;
 import edu.umd.cs.findbugs.annotations.SuppressFBWarnings;
 import java.util.ArrayDeque;
+import java.util.Collection;
+import java.util.Deque;
 import java.util.Iterator;
 import java.util.Optional;
 import java.util.Queue;
@@ -22,6 +24,10 @@ import org.opendaylight.controller.cluster.access.concepts.RequestEnvelope;
 import org.opendaylight.controller.cluster.access.concepts.RequestException;
 import org.opendaylight.controller.cluster.access.concepts.Response;
 import org.opendaylight.controller.cluster.access.concepts.ResponseEnvelope;
+import org.opendaylight.controller.cluster.access.concepts.RuntimeRequestException;
+import org.opendaylight.controller.cluster.access.concepts.SliceableMessage;
+import org.opendaylight.controller.cluster.messaging.MessageSlicer;
+import org.opendaylight.controller.cluster.messaging.SliceOptions;
 import org.slf4j.Logger;
 import org.slf4j.LoggerFactory;
 
@@ -52,28 +58,45 @@ import org.slf4j.LoggerFactory;
 @NotThreadSafe
 abstract class TransmitQueue {
     static final class Halted extends TransmitQueue {
+        // For ConnectingClientConnection.
         Halted(final int targetDepth) {
             super(targetDepth);
         }
 
+        // For ReconnectingClientConnection.
+        Halted(final TransmitQueue oldQueue, final long now) {
+            super(oldQueue, now);
+        }
+
         @Override
         int canTransmitCount(final int inflightSize) {
             return 0;
         }
 
         @Override
-        TransmittedConnectionEntry transmit(final ConnectionEntry entry, final long now) {
+        Optional<TransmittedConnectionEntry> transmit(final ConnectionEntry entry, final long now) {
             throw new UnsupportedOperationException("Attempted to transmit on a halted queue");
         }
+
+        @Override
+        void preComplete(ResponseEnvelope<?> envelope) {
+        }
     }
 
     static final class Transmitting extends TransmitQueue {
+        private static final long NOT_SLICING = -1;
+
         private final BackendInfo backend;
+        private final MessageSlicer messageSlicer;
         private long nextTxSequence;
+        private long currentSlicedEnvSequenceId = NOT_SLICING;
 
-        Transmitting(final int targetDepth, final BackendInfo backend) {
-            super(targetDepth);
+        // For ConnectedClientConnection.
+        Transmitting(final TransmitQueue oldQueue, final int targetDepth, final BackendInfo backend, final long now,
+                final MessageSlicer messageSlicer) {
+            super(oldQueue, targetDepth, now);
             this.backend = Preconditions.checkNotNull(backend);
+            this.messageSlicer = Preconditions.checkNotNull(messageSlicer);
         }
 
         @Override
@@ -82,38 +105,108 @@ abstract class TransmitQueue {
         }
 
         @Override
-        TransmittedConnectionEntry transmit(final ConnectionEntry entry, final long now) {
-            final RequestEnvelope env = new RequestEnvelope(entry.getRequest().toVersion(backend.getVersion()),
+        Optional<TransmittedConnectionEntry> transmit(final ConnectionEntry entry, final long now) {
+            // If we're currently slicing a message we can't send any subsequent requests until slicing completes to
+            // avoid an out-of-sequence request envelope failure on the backend. In this case we return an empty
+            // Optional to indicate the request was not transmitted.
+            if (currentSlicedEnvSequenceId >= 0) {
+                return Optional.empty();
+            }
+
+            final Request<?, ?> request = entry.getRequest();
+            final RequestEnvelope env = new RequestEnvelope(request.toVersion(backend.getVersion()),
                 backend.getSessionId(), nextTxSequence++);
 
-            final TransmittedConnectionEntry ret = new TransmittedConnectionEntry(entry, env.getSessionId(),
-                env.getTxSequence(), now);
-            backend.getActor().tell(env, ActorRef.noSender());
-            return ret;
+            if (request instanceof SliceableMessage) {
+                if (messageSlicer.slice(SliceOptions.builder().identifier(request.getTarget())
+                        .message(env).replyTo(request.getReplyTo()).sendTo(backend.getActor())
+                        .onFailureCallback(t -> env.sendFailure(new RuntimeRequestException(
+                                "Failed to slice request " + request, t), 0L)).build())) {
+                    // The request was sliced so record the envelope sequence id to prevent transmitting
+                    // subsequent requests until slicing completes.
+                    currentSlicedEnvSequenceId = env.getTxSequence();
+                }
+            } else {
+                backend.getActor().tell(env, ActorRef.noSender());
+            }
+
+            return Optional.of(new TransmittedConnectionEntry(entry, env.getSessionId(),
+                    env.getTxSequence(), now));
+        }
+
+        @Override
+        void preComplete(ResponseEnvelope<?> envelope) {
+            if (envelope.getTxSequence() == currentSlicedEnvSequenceId) {
+                // Slicing completed for the prior request - clear the cached sequence id field to enable subsequent
+                // requests to be transmitted.
+                currentSlicedEnvSequenceId = NOT_SLICING;
+            }
         }
     }
 
     private static final Logger LOG = LoggerFactory.getLogger(TransmitQueue.class);
 
-    private final ArrayDeque<TransmittedConnectionEntry> inflight = new ArrayDeque<>();
-    private final ArrayDeque<ConnectionEntry> pending = new ArrayDeque<>();
-    private final ProgressTracker tracker;
+    private final Deque<TransmittedConnectionEntry> inflight = new ArrayDeque<>();
+    private final Deque<ConnectionEntry> pending = new ArrayDeque<>();
+    private final AveragingProgressTracker tracker;  // Cannot be just ProgressTracker as we are inheriting limits.
     private ReconnectForwarder successor;
 
+    /**
+     * Construct initial transmitting queue.
+     */
     TransmitQueue(final int targetDepth) {
         tracker = new AveragingProgressTracker(targetDepth);
     }
 
-    final Iterable<ConnectionEntry> asIterable() {
-        return Iterables.concat(inflight, pending);
+    /**
+     * Construct new transmitting queue while inheriting timing data from the previous transmit queue instance.
+     */
+    TransmitQueue(final TransmitQueue oldQueue, final int targetDepth, final long now) {
+        tracker = new AveragingProgressTracker(oldQueue.tracker, targetDepth, now);
+    }
+
+    /**
+     * Construct new transmitting queue while inheriting timing and size data from the previous transmit queue instance.
+     */
+    TransmitQueue(final TransmitQueue oldQueue, final long now) {
+        tracker = new AveragingProgressTracker(oldQueue.tracker, now);
+    }
+
+    /**
+     * Cancel the accumulated sum of delays as we expect the new backend to work now.
+     */
+    void cancelDebt(final long now) {
+        tracker.cancelDebt(now);
+    }
+
+    /**
+     * Drain the contents of the connection into a list. This will leave the queue empty and allow further entries
+     * to be added to it during replay. When we set the successor all entries enqueued between when this methods
+     * returns and the successor is set will be replayed to the successor.
+     *
+     * @return Collection of entries present in the queue.
+     */
+    final Collection<ConnectionEntry> drain() {
+        final Collection<ConnectionEntry> ret = new ArrayDeque<>(inflight.size() + pending.size());
+        ret.addAll(inflight);
+        ret.addAll(pending);
+        inflight.clear();
+        pending.clear();
+        return ret;
     }
 
     final long ticksStalling(final long now) {
         return tracker.ticksStalling(now);
     }
 
+    final boolean hasSuccessor() {
+        return successor != null;
+    }
+
     // If a matching request was found, this will track a task was closed.
     final Optional<TransmittedConnectionEntry> complete(final ResponseEnvelope<?> envelope, final long now) {
+        preComplete(envelope);
+
         Optional<TransmittedConnectionEntry> maybeEntry = findMatchingEntry(inflight, envelope);
         if (maybeEntry == null) {
             LOG.debug("Request for {} not found in inflight queue, checking pending queue", envelope);
@@ -129,19 +222,60 @@ abstract class TransmitQueue {
         tracker.closeTask(now, entry.getEnqueuedTicks(), entry.getTxTicks(), envelope.getExecutionTimeNanos());
 
         // We have freed up a slot, try to transmit something
-        int toSend = canTransmitCount(inflight.size());
-        while (toSend > 0) {
+        tryTransmit(now);
+
+        return Optional.of(entry);
+    }
+
+    final void tryTransmit(final long now) {
+        final int toSend = canTransmitCount(inflight.size());
+        if (toSend > 0 && !pending.isEmpty()) {
+            transmitEntries(toSend, now);
+        }
+    }
+
+    private void transmitEntries(final int maxTransmit, final long now) {
+        for (int i = 0; i < maxTransmit; ++i) {
             final ConnectionEntry e = pending.poll();
-            if (e == null) {
-                break;
+            if (e == null || !transmitEntry(e, now)) {
+                LOG.debug("Queue {} transmitted {} requests", this, i);
+                return;
             }
+        }
+
+        LOG.debug("Queue {} transmitted {} requests", this, maxTransmit);
+    }
 
-            LOG.debug("Transmitting entry {}", e);
-            transmit(e, now);
-            toSend--;
+    private boolean transmitEntry(final ConnectionEntry entry, final long now) {
+        LOG.debug("Queue {} transmitting entry {}", this, entry);
+        // We are not thread-safe and are supposed to be externally-guarded,
+        // hence send-before-record should be fine.
+        // This needs to be revisited if the external guards are lowered.
+        final Optional<TransmittedConnectionEntry> maybeTransmitted = transmit(entry, now);
+        if (!maybeTransmitted.isPresent()) {
+            return false;
         }
 
-        return Optional.of(entry);
+        inflight.addLast(maybeTransmitted.get());
+        return true;
+    }
+
+    final long enqueueOrForward(final ConnectionEntry entry, final long now) {
+        if (successor != null) {
+            // This call will pay the enqueuing price, hence the caller does not have to
+            successor.forwardEntry(entry, now);
+            return 0;
+        }
+
+        return enqueue(entry, now);
+    }
+
+    final void enqueueOrReplay(final ConnectionEntry entry, final long now) {
+        if (successor != null) {
+            successor.replayEntry(entry, now);
+        } else {
+            enqueue(entry, now);
+        }
     }
 
     /**
@@ -149,24 +283,36 @@ abstract class TransmitQueue {
      *
      * @return Delay to be forced on the calling thread, in nanoseconds.
      */
-    final long enqueue(final ConnectionEntry entry, final long now) {
-        if (successor != null) {
-            successor.forwardEntry(entry, now);
-            return 0;
-        }
+    private long enqueue(final ConnectionEntry entry, final long now) {
+
+        // XXX: we should place a guard against incorrect entry sequences:
+        // entry.getEnqueueTicks() should have non-negative difference from the last entry present in the queues
 
         // Reserve an entry before we do anything that can fail
         final long delay = tracker.openTask(now);
-        if (canTransmitCount(inflight.size()) <= 0) {
+
+        /*
+         * This is defensive to make sure we do not do the wrong thing here and reorder messages if we ever happen
+         * to have available send slots and non-empty pending queue.
+         */
+        final int toSend = canTransmitCount(inflight.size());
+        if (toSend <= 0) {
             LOG.trace("Queue is at capacity, delayed sending of request {}", entry.getRequest());
-            pending.add(entry);
-        } else {
-            // We are not thread-safe and are supposed to be externally-guarded,
-            // hence send-before-record should be fine.
-            // This needs to be revisited if the external guards are lowered.
-            inflight.offer(transmit(entry, now));
-            LOG.debug("Sent request {} on queue {}", entry.getRequest(), this);
+            pending.addLast(entry);
+            return delay;
         }
+
+        if (pending.isEmpty()) {
+            if (!transmitEntry(entry, now)) {
+                LOG.debug("Queue {} cannot transmit request {} - delaying it", this, entry.getRequest());
+                pending.addLast(entry);
+            }
+
+            return delay;
+        }
+
+        pending.addLast(entry);
+        transmitEntries(toSend, now);
         return delay;
     }
 
@@ -175,7 +321,9 @@ abstract class TransmitQueue {
      */
     abstract int canTransmitCount(int inflightSize);
 
-    abstract TransmittedConnectionEntry transmit(ConnectionEntry entry, long now);
+    abstract Optional<TransmittedConnectionEntry> transmit(ConnectionEntry entry, long now);
+
+    abstract void preComplete(ResponseEnvelope<?> envelope);
 
     final boolean isEmpty() {
         return inflight.isEmpty() && pending.isEmpty();
@@ -196,23 +344,53 @@ abstract class TransmitQueue {
     }
 
     final void setForwarder(final ReconnectForwarder forwarder, final long now) {
-        Verify.verify(successor == null, "Successor {} already set on connection {}", successor, this);
+        Verify.verify(successor == null, "Successor %s already set on connection %s", successor, this);
         successor = Preconditions.checkNotNull(forwarder);
         LOG.debug("Connection {} superseded by {}, splicing queue", this, successor);
 
+        /*
+         * We need to account for entries which have been added between the time drain() was called and this method
+         * is invoked. Since the old connection is visible during replay and some entries may have completed on the
+         * replay thread, there was an avenue for this to happen.
+         */
+        int count = 0;
         ConnectionEntry entry = inflight.poll();
         while (entry != null) {
-            successor.forwardEntry(entry, now);
+            successor.replayEntry(entry, now);
             entry = inflight.poll();
+            count++;
         }
 
         entry = pending.poll();
         while (entry != null) {
-            successor.forwardEntry(entry, now);
+            successor.replayEntry(entry, now);
             entry = pending.poll();
+            count++;
+        }
+
+        LOG.debug("Connection {} queue spliced {} messages", this, count);
+    }
+
+    final void remove(final long now) {
+        final TransmittedConnectionEntry txe = inflight.poll();
+        if (txe == null) {
+            final ConnectionEntry entry = pending.pop();
+            tracker.closeTask(now, entry.getEnqueuedTicks(), 0, 0);
+        } else {
+            tracker.closeTask(now, txe.getEnqueuedTicks(), txe.getTxTicks(), 0);
         }
     }
 
+    @VisibleForTesting
+    Deque<TransmittedConnectionEntry> getInflight() {
+        return inflight;
+    }
+
+    @VisibleForTesting
+    Deque<ConnectionEntry> getPending() {
+        return pending;
+    }
+
     /*
      * We are using tri-state return here to indicate one of three conditions:
      * - if a matching entry is found, return an Optional containing it
@@ -275,5 +453,4 @@ abstract class TransmitQueue {
         }
         queue.clear();
     }
-
 }