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