opendaylight/md-sal/cds-access-client/src/main/java/org/opendaylight/controller/cluster/access/client/ClientActorBehavior.java

   1 /*
   2  * Copyright (c) 2016 Cisco Systems, Inc. and others.  All rights reserved.
   3  *
   4  * This program and the accompanying materials are made available under the
   5  * terms of the Eclipse Public License v1.0 which accompanies this distribution,
   6  * and is available at http://www.eclipse.org/legal/epl-v10.html
   7  */
   8 package org.opendaylight.controller.cluster.access.client;
   9
  10 import com.google.common.annotations.Beta;
  11 import com.google.common.base.Preconditions;
  12 import com.google.common.base.Verify;
  13 import java.util.Map;
  14 import java.util.concurrent.ConcurrentHashMap;
  15 import java.util.concurrent.TimeUnit;
  16 import java.util.concurrent.TimeoutException;
  17 import javax.annotation.Nonnull;
  18 import javax.annotation.Nullable;
  19 import javax.annotation.concurrent.GuardedBy;
  20 import org.opendaylight.controller.cluster.access.commands.NotLeaderException;
  21 import org.opendaylight.controller.cluster.access.commands.OutOfSequenceEnvelopeException;
  22 import org.opendaylight.controller.cluster.access.concepts.ClientIdentifier;
  23 import org.opendaylight.controller.cluster.access.concepts.FailureEnvelope;
  24 import org.opendaylight.controller.cluster.access.concepts.LocalHistoryIdentifier;
  25 import org.opendaylight.controller.cluster.access.concepts.RequestException;
  26 import org.opendaylight.controller.cluster.access.concepts.RequestFailure;
  27 import org.opendaylight.controller.cluster.access.concepts.ResponseEnvelope;
  28 import org.opendaylight.controller.cluster.access.concepts.RetiredGenerationException;
  29 import org.opendaylight.controller.cluster.access.concepts.RuntimeRequestException;
  30 import org.opendaylight.controller.cluster.access.concepts.SuccessEnvelope;
  31 import org.opendaylight.controller.cluster.access.concepts.TransactionIdentifier;
  32 import org.opendaylight.yangtools.concepts.Identifiable;
  33 import org.opendaylight.yangtools.concepts.WritableIdentifier;
  34 import org.slf4j.Logger;
  35 import org.slf4j.LoggerFactory;
  36 import scala.concurrent.duration.FiniteDuration;
  37
  38 /**
  39  * A behavior, which handles messages sent to a {@link AbstractClientActor}.
  40  *
  41  * @author Robert Varga
  42  */
  43 @Beta
  44 public abstract class ClientActorBehavior<T extends BackendInfo> extends
  45         RecoveredClientActorBehavior<ClientActorContext> implements Identifiable<ClientIdentifier> {
  46     /**
  47      * Connection reconnect cohort, driven by this class.
  48      */
  49     @FunctionalInterface
  50     protected interface ConnectionConnectCohort {
  51         /**
  52          * Finish the connection by replaying previous messages onto the new connection.
  53          *
  54          * @param enqueuedEntries Previously-enqueued entries
  55          * @return A {@link ReconnectForwarder} to handle any straggler messages which arrive after this method returns.
  56          */
  57         @Nonnull ReconnectForwarder finishReconnect(@Nonnull Iterable<ConnectionEntry> enqueuedEntries);
  58     }
  59
  60     private static final Logger LOG = LoggerFactory.getLogger(ClientActorBehavior.class);
  61     private static final FiniteDuration RESOLVE_RETRY_DURATION = FiniteDuration.apply(5, TimeUnit.SECONDS);
  62
  63     /**
  64      * Map of connections to the backend. This map is concurrent to allow lookups, but given complex operations
  65      * involved in connection transitions it is protected by a {@link InversibleLock}. Write-side of the lock is taken
  66      * during connection transitions. Optimistic read-side of the lock is taken when new connections are introduced
  67      * into the map.
  68      *
  69      * <p>
  70      * The lock detects potential AB/BA deadlock scenarios and will force the reader side out by throwing
  71      * a {@link InversibleLockException} -- which must be propagated up, releasing locks as it propagates. The initial
  72      * entry point causing the the conflicting lookup must then call {@link InversibleLockException#awaitResolution()}
  73      * before retrying the operation.
  74      */
  75     // TODO: it should be possible to move these two into ClientActorContext
  76     private final Map<Long, AbstractClientConnection<T>> connections = new ConcurrentHashMap<>();
  77     private final InversibleLock connectionsLock = new InversibleLock();
  78     private final BackendInfoResolver<T> resolver;
  79
  80     protected ClientActorBehavior(@Nonnull final ClientActorContext context,
  81             @Nonnull final BackendInfoResolver<T> resolver) {
  82         super(context);
  83         this.resolver = Preconditions.checkNotNull(resolver);
  84     }
  85
  86     @Override
  87     @Nonnull
  88     public final ClientIdentifier getIdentifier() {
  89         return context().getIdentifier();
  90     }
  91
  92     /**
  93      * Get a connection to a shard.
  94      *
  95      * @param shard Shard cookie
  96      * @return Connection to a shard
  97      * @throws InversibleLockException if the shard is being reconnected
  98      */
  99     public final AbstractClientConnection<T> getConnection(final Long shard) {
 100         while (true) {
 101             final long stamp = connectionsLock.optimisticRead();
 102             final AbstractClientConnection<T> conn = connections.computeIfAbsent(shard, this::createConnection);
 103             if (connectionsLock.validate(stamp)) {
 104                 // No write-lock in-between, return success
 105                 return conn;
 106             }
 107         }
 108     }
 109
 110     private AbstractClientConnection<T> getConnection(final ResponseEnvelope<?> response) {
 111         // Always called from actor context: no locking required
 112         return connections.get(extractCookie(response.getMessage().getTarget()));
 113     }
 114
 115     @SuppressWarnings("unchecked")
 116     @Override
 117     final ClientActorBehavior<T> onReceiveCommand(final Object command) {
 118         if (command instanceof InternalCommand) {
 119             return ((InternalCommand<T>) command).execute(this);
 120         }
 121         if (command instanceof SuccessEnvelope) {
 122             return onRequestSuccess((SuccessEnvelope) command);
 123         }
 124         if (command instanceof FailureEnvelope) {
 125             return internalOnRequestFailure((FailureEnvelope) command);
 126         }
 127
 128         return onCommand(command);
 129     }
 130
 131     private static long extractCookie(final WritableIdentifier id) {
 132         if (id instanceof TransactionIdentifier) {
 133             return ((TransactionIdentifier) id).getHistoryId().getCookie();
 134         } else if (id instanceof LocalHistoryIdentifier) {
 135             return ((LocalHistoryIdentifier) id).getCookie();
 136         } else {
 137             throw new IllegalArgumentException("Unhandled identifier " + id);
 138         }
 139     }
 140
 141     private void onResponse(final ResponseEnvelope<?> response) {
 142         final AbstractClientConnection<T> connection = getConnection(response);
 143         if (connection != null) {
 144             connection.receiveResponse(response);
 145         } else {
 146             LOG.info("{}: Ignoring unknown response {}", persistenceId(), response);
 147         }
 148     }
 149
 150     private ClientActorBehavior<T> onRequestSuccess(final SuccessEnvelope success) {
 151         onResponse(success);
 152         return this;
 153     }
 154
 155     private ClientActorBehavior<T> onRequestFailure(final FailureEnvelope failure) {
 156         onResponse(failure);
 157         return this;
 158     }
 159
 160     private ClientActorBehavior<T> internalOnRequestFailure(final FailureEnvelope command) {
 161         final RequestFailure<?, ?> failure = command.getMessage();
 162         final RequestException cause = failure.getCause();
 163         if (cause instanceof RetiredGenerationException) {
 164             LOG.error("{}: current generation {} has been superseded", persistenceId(), getIdentifier(), cause);
 165             haltClient(cause);
 166             poison(cause);
 167             return null;
 168         }
 169         if (cause instanceof NotLeaderException) {
 170             final AbstractClientConnection<T> conn = getConnection(command);
 171             if (conn instanceof ReconnectingClientConnection) {
 172                 // Already reconnecting, do not churn the logs
 173                 return this;
 174             } else if (conn != null) {
 175                 LOG.info("{}: connection {} indicated no leadership, reconnecting it", persistenceId(), conn, cause);
 176                 return conn.reconnect(this);
 177             }
 178         }
 179         if (cause instanceof OutOfSequenceEnvelopeException) {
 180             final AbstractClientConnection<T> conn = getConnection(command);
 181             if (conn instanceof ReconnectingClientConnection) {
 182                 // Already reconnecting, do not churn the logs
 183                 return this;
 184             } else if (conn != null) {
 185                 LOG.info("{}: connection {} indicated no sequencing mismatch on {} sequence {}, reconnecting it",
 186                     persistenceId(), conn, failure.getTarget(), failure.getSequence(), cause);
 187                 return conn.reconnect(this);
 188             }
 189         }
 190
 191         return onRequestFailure(command);
 192     }
 193
 194     private void poison(final RequestException cause) {
 195         final long stamp = connectionsLock.writeLock();
 196         try {
 197             for (AbstractClientConnection<T> q : connections.values()) {
 198                 q.poison(cause);
 199             }
 200
 201             connections.clear();
 202         } finally {
 203             connectionsLock.unlockWrite(stamp);
 204         }
 205     }
 206
 207     /**
 208      * Halt And Catch Fire. Halt processing on this client. Implementations need to ensure they initiate state flush
 209      * procedures. No attempt to use this instance should be made after this method returns. Any such use may result
 210      * in undefined behavior.
 211      *
 212      * @param cause Failure cause
 213      */
 214     protected abstract void haltClient(@Nonnull Throwable cause);
 215
 216     /**
 217      * Override this method to handle any command which is not handled by the base behavior.
 218      *
 219      * @param command the command to process
 220      * @return Next behavior to use, null if this actor should shut down.
 221      */
 222     @Nullable
 223     protected abstract ClientActorBehavior<T> onCommand(@Nonnull Object command);
 224
 225     /**
 226      * Override this method to provide a backend resolver instance.
 227      *
 228      * @return a backend resolver instance
 229      */
 230     protected final @Nonnull BackendInfoResolver<T> resolver() {
 231         return resolver;
 232     }
 233
 234     /**
 235      * Callback invoked when a new connection has been established. Implementations are expected perform preparatory
 236      * tasks before the previous connection is frozen.
 237      *
 238      * @param newConn New connection
 239      * @return ConnectionConnectCohort which will be used to complete the process of bringing the connection up.
 240      */
 241     @GuardedBy("connectionsLock")
 242     @Nonnull protected abstract ConnectionConnectCohort connectionUp(@Nonnull ConnectedClientConnection<T> newConn);
 243
 244     private void backendConnectFinished(final Long shard, final AbstractClientConnection<T> conn,
 245             final T backend, final Throwable failure) {
 246         if (failure != null) {
 247             if (failure instanceof TimeoutException) {
 248                 if (!conn.equals(connections.get(shard))) {
 249                     // AbstractClientConnection will remove itself when it decides there is no point in continuing,
 250                     // at which point we want to stop retrying
 251                     LOG.info("{}: stopping resolution of shard {} on stale connection {}", persistenceId(), shard, conn,
 252                         failure);
 253                     return;
 254                 }
 255
 256                 LOG.debug("{}: timed out resolving shard {}, scheduling retry in {}", persistenceId(), shard,
 257                     RESOLVE_RETRY_DURATION, failure);
 258                 context().executeInActor(b -> {
 259                     resolveConnection(shard, conn);
 260                     return b;
 261                 }, RESOLVE_RETRY_DURATION);
 262                 return;
 263             }
 264
 265             LOG.error("{}: failed to resolve shard {}", persistenceId(), shard, failure);
 266             final RequestException cause;
 267             if (failure instanceof RequestException) {
 268                 cause = (RequestException) failure;
 269             } else {
 270                 cause = new RuntimeRequestException("Failed to resolve shard " + shard, failure);
 271             }
 272
 273             conn.poison(cause);
 274             return;
 275         }
 276
 277         LOG.info("{}: resolved shard {} to {}", persistenceId(), shard, backend);
 278         final long stamp = connectionsLock.writeLock();
 279         try {
 280             // Create a new connected connection
 281             final ConnectedClientConnection<T> newConn = new ConnectedClientConnection<>(conn.context(),
 282                     conn.cookie(), backend);
 283             LOG.info("{}: resolving connection {} to {}", persistenceId(), conn, newConn);
 284
 285             // Start reconnecting without the old connection lock held
 286             final ConnectionConnectCohort cohort = Verify.verifyNotNull(connectionUp(newConn));
 287
 288             // Lock the old connection and get a reference to its entries
 289             final Iterable<ConnectionEntry> replayIterable = conn.startReplay();
 290
 291             // Finish the connection attempt
 292             final ReconnectForwarder forwarder = Verify.verifyNotNull(cohort.finishReconnect(replayIterable));
 293
 294             // Install the forwarder, unlocking the old connection
 295             conn.finishReplay(forwarder);
 296
 297             // Make sure new lookups pick up the new connection
 298             connections.replace(shard, conn, newConn);
 299             LOG.info("{}: replaced connection {} with {}", persistenceId(), conn, newConn);
 300         } finally {
 301             connectionsLock.unlockWrite(stamp);
 302         }
 303     }
 304
 305     void removeConnection(final AbstractClientConnection<?> conn) {
 306         connections.remove(conn.cookie(), conn);
 307         LOG.debug("{}: removed connection {}", persistenceId(), conn);
 308     }
 309
 310     @SuppressWarnings("unchecked")
 311     void reconnectConnection(final ConnectedClientConnection<?> oldConn,
 312             final ReconnectingClientConnection<?> newConn) {
 313         final ReconnectingClientConnection<T> conn = (ReconnectingClientConnection<T>)newConn;
 314         LOG.info("{}: connection {} reconnecting as {}", persistenceId(), oldConn, newConn);
 315
 316         final boolean replaced = connections.replace(oldConn.cookie(), (AbstractClientConnection<T>)oldConn, conn);
 317         if (!replaced) {
 318             final AbstractClientConnection<T> existing = connections.get(oldConn.cookie());
 319             LOG.warn("{}: old connection {} does not match existing {}, new connection {} in limbo", persistenceId(),
 320                 oldConn, existing, newConn);
 321         }
 322
 323         final Long shard = oldConn.cookie();
 324         LOG.info("{}: refreshing backend for shard {}", persistenceId(), shard);
 325         resolver().refreshBackendInfo(shard, conn.getBackendInfo().get()).whenComplete(
 326             (backend, failure) -> context().executeInActor(behavior -> {
 327                 backendConnectFinished(shard, conn, backend, failure);
 328                 return behavior;
 329             }));
 330     }
 331
 332     private ConnectingClientConnection<T> createConnection(final Long shard) {
 333         final ConnectingClientConnection<T> conn = new ConnectingClientConnection<>(context(), shard);
 334         resolveConnection(shard, conn);
 335         return conn;
 336     }
 337
 338     private void resolveConnection(final Long shard, final AbstractClientConnection<T> conn) {
 339         LOG.debug("{}: resolving shard {} connection {}", persistenceId(), shard, conn);
 340         resolver().getBackendInfo(shard).whenComplete((backend, failure) -> context().executeInActor(behavior -> {
 341             backendConnectFinished(shard, conn, backend, failure);
 342             return behavior;
 343         }));
 344     }
 345 }