package org.opendaylight.controller.cluster.access.client;
import com.google.common.annotations.Beta;
-import java.util.Optional;
-import java.util.concurrent.CompletionStage;
+import com.google.common.base.Preconditions;
+import java.util.Map;
+import java.util.concurrent.ConcurrentHashMap;
import javax.annotation.Nonnull;
import javax.annotation.Nullable;
-import org.opendaylight.controller.cluster.access.commands.TransactionRequest;
+import javax.annotation.concurrent.GuardedBy;
import org.opendaylight.controller.cluster.access.concepts.ClientIdentifier;
import org.opendaylight.controller.cluster.access.concepts.FailureEnvelope;
import org.opendaylight.controller.cluster.access.concepts.RequestException;
import org.opendaylight.controller.cluster.access.concepts.RequestFailure;
+import org.opendaylight.controller.cluster.access.concepts.ResponseEnvelope;
import org.opendaylight.controller.cluster.access.concepts.RetiredGenerationException;
import org.opendaylight.controller.cluster.access.concepts.SuccessEnvelope;
+import org.opendaylight.controller.cluster.access.concepts.TransactionIdentifier;
import org.opendaylight.yangtools.concepts.Identifiable;
+import org.opendaylight.yangtools.concepts.WritableIdentifier;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;
-import scala.concurrent.duration.FiniteDuration;
/**
* A behavior, which handles messages sent to a {@link AbstractClientActor}.
* @author Robert Varga
*/
@Beta
-public abstract class ClientActorBehavior extends RecoveredClientActorBehavior<ClientActorContext>
- implements Identifiable<ClientIdentifier> {
+public abstract class ClientActorBehavior<T extends BackendInfo> extends
+ RecoveredClientActorBehavior<ClientActorContext> implements Identifiable<ClientIdentifier> {
private static final Logger LOG = LoggerFactory.getLogger(ClientActorBehavior.class);
- protected ClientActorBehavior(@Nonnull final ClientActorContext context) {
+ /**
+ * Map of connections to the backend. This map is concurrent to allow lookups, but given complex operations
+ * involved in connection transitions it is protected by a {@link InversibleLock}. Write-side of the lock is taken
+ * during connection transitions. Optimistic read-side of the lock is taken when new connections are introduced
+ * into the map.
+ *
+ * <p>
+ * The lock detects potential AB/BA deadlock scenarios and will force the reader side out by throwing
+ * a {@link InversibleLockException} -- which must be propagated up, releasing locks as it propagates. The initial
+ * entry point causing the the conflicting lookup must then call {@link InversibleLockException#awaitResolution()}
+ * before retrying the operation.
+ */
+ // TODO: it should be possible to move these two into ClientActorContext
+ private final Map<Long, AbstractClientConnection<T>> connections = new ConcurrentHashMap<>();
+ private final InversibleLock connectionsLock = new InversibleLock();
+ private final BackendInfoResolver<T> resolver;
+
+ protected ClientActorBehavior(@Nonnull final ClientActorContext context,
+ @Nonnull final BackendInfoResolver<T> resolver) {
super(context);
+ this.resolver = Preconditions.checkNotNull(resolver);
}
@Override
return context().getIdentifier();
}
+ /**
+ * Get a connection to a shard.
+ *
+ * @param shard Shard cookie
+ * @return Connection to a shard
+ * @throws InversibleLockException if the shard is being reconnected
+ */
+ public final AbstractClientConnection<T> getConnection(final Long shard) {
+ while (true) {
+ final long stamp = connectionsLock.optimisticRead();
+ final AbstractClientConnection<T> conn = connections.computeIfAbsent(shard, this::createConnection);
+ if (connectionsLock.validate(stamp)) {
+ // No write-lock in-between, return success
+ return conn;
+ }
+ }
+ }
+
+ @SuppressWarnings("unchecked")
@Override
- final ClientActorBehavior onReceiveCommand(final Object command) {
+ final ClientActorBehavior<T> onReceiveCommand(final Object command) {
if (command instanceof InternalCommand) {
- return ((InternalCommand) command).execute(this);
+ return ((InternalCommand<T>) command).execute(this);
}
if (command instanceof SuccessEnvelope) {
return onRequestSuccess((SuccessEnvelope) command);
}
if (command instanceof FailureEnvelope) {
- return onRequestFailure((FailureEnvelope) command);
+ return internalOnRequestFailure((FailureEnvelope) command);
}
return onCommand(command);
}
- private ClientActorBehavior onRequestSuccess(final SuccessEnvelope command) {
- return context().completeRequest(this, command);
- }
+ private void onResponse(final ResponseEnvelope<?> response) {
+ final WritableIdentifier id = response.getMessage().getTarget();
- private ClientActorBehavior onRequestFailure(final FailureEnvelope command) {
- final RequestFailure<?, ?> failure = command.getMessage();
- final RequestException cause = failure.getCause();
- if (cause instanceof RetiredGenerationException) {
- LOG.error("{}: current generation {} has been superseded", persistenceId(), getIdentifier(), cause);
- haltClient(cause);
- context().poison(cause);
- return null;
- }
+ // FIXME: this will need to be updated for other Request/Response types to extract cookie
+ Preconditions.checkArgument(id instanceof TransactionIdentifier);
+ final TransactionIdentifier txId = (TransactionIdentifier) id;
- if (failure.isHardFailure()) {
- return context().completeRequest(this, command);
+ final AbstractClientConnection<T> connection = connections.get(txId.getHistoryId().getCookie());
+ if (connection != null) {
+ connection.receiveResponse(response);
+ } else {
+ LOG.info("{}: Ignoring unknown response {}", persistenceId(), response);
}
-
- // TODO: add instanceof checks on cause to detect more problems
-
- LOG.warn("{}: Unhandled retriable failure {}, promoting to hard failure", persistenceId(), command);
- return context().completeRequest(this, command);
}
- // This method is executing in the actor context, hence we can safely interact with the queue
- private ClientActorBehavior doSendRequest(final TransactionRequest<?> request, final RequestCallback callback) {
- // Get or allocate queue for the request
- final SequencedQueue queue = context().queueFor(request.getTarget().getHistoryId().getCookie());
-
- // Note this is a tri-state return and can be null
- final Optional<FiniteDuration> result = queue.enqueueRequest(request, callback);
- if (result == null) {
- // Happy path: we are done here
- return this;
- }
-
- if (result.isPresent()) {
- // Less happy path: we need to schedule a timer
- scheduleQueueTimeout(queue, result.get());
- return this;
- }
-
- startResolve(queue, request.getTarget().getHistoryId().getCookie());
+ private ClientActorBehavior<T> onRequestSuccess(final SuccessEnvelope success) {
+ onResponse(success);
return this;
}
- // This method is executing in the actor context, hence we can safely interact with the queue
- private void startResolve(final SequencedQueue queue, final long cookie) {
- // Queue does not have backend information. Initiate resolution, which may actually be piggy-backing on to a
- // previous request to resolve.
- final CompletionStage<? extends BackendInfo> f = resolver().getBackendInfo(cookie);
-
- // This is the tricky part: depending on timing, the queue may have a stale request for resolution, which has
- // been invalidated or it may already have a reference to this resolution request. Let us give it a chance to
- // update and it will indicate if this resolution request is an update. If it is, we'll piggy-back on it and
- // run backend information update in the actor thread. If it is not, we do not need to do anything, as we will
- // bulk-process all requests.
- if (queue.expectProof(f)) {
- f.thenAccept(backend -> context().executeInActor(cb -> cb.finishResolve(queue, f, backend)));
- }
+ private ClientActorBehavior<T> onRequestFailure(final FailureEnvelope failure) {
+ onResponse(failure);
+ return this;
}
- // This method is executing in the actor context, hence we can safely interact with the queue
- private ClientActorBehavior finishResolve(final SequencedQueue queue,
- final CompletionStage<? extends BackendInfo> futureBackend, final BackendInfo backend) {
-
- final Optional<FiniteDuration> maybeTimeout = queue.setBackendInfo(futureBackend, backend);
- if (maybeTimeout.isPresent()) {
- scheduleQueueTimeout(queue, maybeTimeout.get());
+ private ClientActorBehavior<T> internalOnRequestFailure(final FailureEnvelope command) {
+ final RequestFailure<?, ?> failure = command.getMessage();
+ final RequestException cause = failure.getCause();
+ if (cause instanceof RetiredGenerationException) {
+ LOG.error("{}: current generation {} has been superseded", persistenceId(), getIdentifier(), cause);
+ haltClient(cause);
+ poison(cause);
+ return null;
}
- return this;
- }
- // This method is executing in the actor context, hence we can safely interact with the queue
- private void scheduleQueueTimeout(final SequencedQueue queue, final FiniteDuration timeout) {
- LOG.debug("{}: scheduling timeout in {}", persistenceId(), timeout);
- context().executeInActor(cb -> cb.queueTimeout(queue), timeout);
+ return onRequestFailure(command);
}
- // This method is executing in the actor context, hence we can safely interact with the queue
- private ClientActorBehavior queueTimeout(final SequencedQueue queue) {
- final boolean needBackend;
-
+ private void poison(final RequestException cause) {
+ final long stamp = connectionsLock.writeLock();
try {
- needBackend = queue.runTimeout();
- } catch (NoProgressException e) {
- // Uh-oh, no progress. The queue has already killed itself, now we need to remove it
- LOG.debug("{}: No progress made - removing queue", persistenceId(), e);
- context().removeQueue(queue);
- return this;
- }
+ for (AbstractClientConnection<T> q : connections.values()) {
+ q.poison(cause);
+ }
- if (needBackend) {
- startResolve(queue, queue.getCookie());
+ connections.clear();
+ } finally {
+ connectionsLock.unlockWrite(stamp);
}
-
- return this;
}
/**
* @return Next behavior to use, null if this actor should shut down.
*/
@Nullable
- protected abstract ClientActorBehavior onCommand(@Nonnull Object command);
+ protected abstract ClientActorBehavior<T> onCommand(@Nonnull Object command);
/**
* Override this method to provide a backend resolver instance.
*
* @return a backend resolver instance
*/
- @Nonnull
- protected abstract BackendInfoResolver<?> resolver();
+ protected final @Nonnull BackendInfoResolver<T> resolver() {
+ return resolver;
+ }
/**
- * Send a request to the backend and invoke a specified callback when it finishes. This method is safe to invoke
- * from any thread.
+ * Callback invoked when a new connection has been established.
*
- * @param request Request to send
- * @param callback Callback to invoke
+ * @param conn Old connection
+ * @param backend New backend
+ * @return Newly-connected connection.
*/
- public final void sendRequest(final TransactionRequest<?> request, final RequestCallback callback) {
- context().executeInActor(cb -> cb.doSendRequest(request, callback));
+ @GuardedBy("connectionsLock")
+ protected abstract @Nonnull ConnectedClientConnection<T> connectionUp(
+ final @Nonnull AbstractClientConnection<T> conn, final @Nonnull T backend);
+
+ private void backendConnectFinished(final Long shard, final AbstractClientConnection<T> conn,
+ final T backend, final Throwable failure) {
+ if (failure != null) {
+ LOG.error("{}: failed to resolve shard {}", persistenceId(), shard, failure);
+ return;
+ }
+
+ final long stamp = connectionsLock.writeLock();
+ try {
+ // Bring the connection up
+ final ConnectedClientConnection<T> newConn = connectionUp(conn, backend);
+
+ // Make sure new lookups pick up the new connection
+ connections.replace(shard, conn, newConn);
+ LOG.debug("{}: replaced connection {} with {}", persistenceId(), conn, newConn);
+ } finally {
+ connectionsLock.unlockWrite(stamp);
+ }
+ }
+
+ void removeConnection(final AbstractClientConnection<?> conn) {
+ connections.remove(conn.cookie(), conn);
+ LOG.debug("{}: removed connection {}", persistenceId(), conn);
+ }
+
+ @SuppressWarnings("unchecked")
+ void reconnectConnection(final ConnectedClientConnection<?> oldConn,
+ final ReconnectingClientConnection<?> newConn) {
+ final ReconnectingClientConnection<T> conn = (ReconnectingClientConnection<T>)newConn;
+ connections.replace(oldConn.cookie(), (AbstractClientConnection<T>)oldConn, conn);
+ LOG.debug("{}: connection {} reconnecting as {}", persistenceId(), oldConn, newConn);
+
+ final Long shard = oldConn.cookie();
+ resolver().refreshBackendInfo(shard, conn.getBackendInfo().get()).whenComplete(
+ (backend, failure) -> context().executeInActor(behavior -> {
+ backendConnectFinished(shard, conn, backend, failure);
+ return behavior;
+ }));
+ }
+
+ private ConnectingClientConnection<T> createConnection(final Long shard) {
+ final ConnectingClientConnection<T> conn = new ConnectingClientConnection<>(context(), shard);
+
+ resolver().getBackendInfo(shard).whenComplete((backend, failure) -> context().executeInActor(behavior -> {
+ backendConnectFinished(shard, conn, backend, failure);
+ return behavior;
+ }));
+
+ return conn;
}
}
Code Review / controller.git / blobdiff