opendaylight/md-sal/sal-distributed-datastore/src/main/java/org/opendaylight/controller/cluster/datastore/ShardDataTree.java

   1 /*
   2  * Copyright (c) 2015 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.datastore;
   9
  10 import akka.actor.ActorRef;
  11 import akka.util.Timeout;
  12 import com.google.common.annotations.VisibleForTesting;
  13 import com.google.common.base.Optional;
  14 import com.google.common.base.Preconditions;
  15 import com.google.common.base.Stopwatch;
  16 import com.google.common.base.Ticker;
  17 import com.google.common.base.Verify;
  18 import com.google.common.collect.ImmutableList;
  19 import com.google.common.collect.ImmutableMap;
  20 import com.google.common.collect.ImmutableMap.Builder;
  21 import com.google.common.primitives.UnsignedLong;
  22 import edu.umd.cs.findbugs.annotations.SuppressFBWarnings;
  23 import java.io.File;
  24 import java.io.IOException;
  25 import java.util.AbstractMap.SimpleEntry;
  26 import java.util.ArrayDeque;
  27 import java.util.ArrayList;
  28 import java.util.Collection;
  29 import java.util.HashMap;
  30 import java.util.Iterator;
  31 import java.util.Map;
  32 import java.util.Map.Entry;
  33 import java.util.Queue;
  34 import java.util.concurrent.ExecutionException;
  35 import java.util.concurrent.TimeUnit;
  36 import java.util.concurrent.TimeoutException;
  37 import java.util.function.UnaryOperator;
  38 import javax.annotation.Nonnull;
  39 import javax.annotation.concurrent.NotThreadSafe;
  40 import org.opendaylight.controller.cluster.access.concepts.LocalHistoryIdentifier;
  41 import org.opendaylight.controller.cluster.access.concepts.TransactionIdentifier;
  42 import org.opendaylight.controller.cluster.datastore.DataTreeCohortActorRegistry.CohortRegistryCommand;
  43 import org.opendaylight.controller.cluster.datastore.ShardDataTreeCohort.State;
  44 import org.opendaylight.controller.cluster.datastore.persisted.CommitTransactionPayload;
  45 import org.opendaylight.controller.cluster.datastore.persisted.MetadataShardDataTreeSnapshot;
  46 import org.opendaylight.controller.cluster.datastore.persisted.ShardDataTreeSnapshot;
  47 import org.opendaylight.controller.cluster.datastore.persisted.ShardDataTreeSnapshotMetadata;
  48 import org.opendaylight.controller.cluster.datastore.utils.DataTreeModificationOutput;
  49 import org.opendaylight.controller.cluster.datastore.utils.PruningDataTreeModification;
  50 import org.opendaylight.controller.cluster.raft.protobuff.client.messages.Payload;
  51 import org.opendaylight.controller.md.sal.common.api.data.AsyncDataBroker.DataChangeScope;
  52 import org.opendaylight.controller.md.sal.common.api.data.AsyncDataChangeListener;
  53 import org.opendaylight.controller.md.sal.common.api.data.OptimisticLockFailedException;
  54 import org.opendaylight.controller.md.sal.common.api.data.TransactionCommitFailedException;
  55 import org.opendaylight.controller.md.sal.dom.api.DOMDataTreeChangeListener;
  56 import org.opendaylight.controller.md.sal.dom.store.impl.DataChangeListenerRegistration;
  57 import org.opendaylight.yangtools.concepts.Identifier;
  58 import org.opendaylight.yangtools.concepts.ListenerRegistration;
  59 import org.opendaylight.yangtools.yang.data.api.YangInstanceIdentifier;
  60 import org.opendaylight.yangtools.yang.data.api.schema.NormalizedNode;
  61 import org.opendaylight.yangtools.yang.data.api.schema.tree.ConflictingModificationAppliedException;
  62 import org.opendaylight.yangtools.yang.data.api.schema.tree.DataTreeCandidate;
  63 import org.opendaylight.yangtools.yang.data.api.schema.tree.DataTreeCandidateTip;
  64 import org.opendaylight.yangtools.yang.data.api.schema.tree.DataTreeCandidates;
  65 import org.opendaylight.yangtools.yang.data.api.schema.tree.DataTreeModification;
  66 import org.opendaylight.yangtools.yang.data.api.schema.tree.DataTreeSnapshot;
  67 import org.opendaylight.yangtools.yang.data.api.schema.tree.DataTreeTip;
  68 import org.opendaylight.yangtools.yang.data.api.schema.tree.DataValidationFailedException;
  69 import org.opendaylight.yangtools.yang.data.api.schema.tree.ModificationType;
  70 import org.opendaylight.yangtools.yang.data.api.schema.tree.TipProducingDataTree;
  71 import org.opendaylight.yangtools.yang.data.api.schema.tree.TipProducingDataTreeTip;
  72 import org.opendaylight.yangtools.yang.data.api.schema.tree.TreeType;
  73 import org.opendaylight.yangtools.yang.data.impl.schema.tree.InMemoryDataTreeFactory;
  74 import org.opendaylight.yangtools.yang.model.api.SchemaContext;
  75 import org.slf4j.Logger;
  76 import org.slf4j.LoggerFactory;
  77 import scala.concurrent.duration.Duration;
  78
  79 /**
  80  * Internal shard state, similar to a DOMStore, but optimized for use in the actor system,
  81  * e.g. it does not expose public interfaces and assumes it is only ever called from a
  82  * single thread.
  83  *
  84  * <p>
  85  * This class is not part of the API contract and is subject to change at any time.
  86  */
  87 @NotThreadSafe
  88 public class ShardDataTree extends ShardDataTreeTransactionParent {
  89     private static final class CommitEntry {
  90         final SimpleShardDataTreeCohort cohort;
  91         long lastAccess;
  92
  93         CommitEntry(final SimpleShardDataTreeCohort cohort, final long now) {
  94             this.cohort = Preconditions.checkNotNull(cohort);
  95             lastAccess = now;
  96         }
  97     }
  98
  99     private static final Timeout COMMIT_STEP_TIMEOUT = new Timeout(Duration.create(5, TimeUnit.SECONDS));
 100     private static final Logger LOG = LoggerFactory.getLogger(ShardDataTree.class);
 101
 102     private final Map<LocalHistoryIdentifier, ShardDataTreeTransactionChain> transactionChains = new HashMap<>();
 103     private final DataTreeCohortActorRegistry cohortRegistry = new DataTreeCohortActorRegistry();
 104     private final Queue<CommitEntry> pendingTransactions = new ArrayDeque<>();
 105     private final ShardDataTreeChangeListenerPublisher treeChangeListenerPublisher;
 106     private final ShardDataChangeListenerPublisher dataChangeListenerPublisher;
 107     private final Collection<ShardDataTreeMetadata<?>> metadata;
 108     private final TipProducingDataTree dataTree;
 109     private final String logContext;
 110     private final Shard shard;
 111     private Runnable runOnPendingTransactionsComplete;
 112
 113     /**
 114      * Optimistic {@link DataTreeCandidate} preparation. Since our DataTree implementation is a
 115      * {@link TipProducingDataTree}, each {@link DataTreeCandidate} is also a {@link DataTreeTip}, e.g. another
 116      * candidate can be prepared on top of it. They still need to be committed in sequence. Here we track the current
 117      * tip of the data tree, which is the last DataTreeCandidate we have in flight, or the DataTree itself.
 118      */
 119     private TipProducingDataTreeTip tip;
 120
 121     private SchemaContext schemaContext;
 122
 123     public ShardDataTree(final Shard shard, final SchemaContext schemaContext, final TipProducingDataTree dataTree,
 124             final ShardDataTreeChangeListenerPublisher treeChangeListenerPublisher,
 125             final ShardDataChangeListenerPublisher dataChangeListenerPublisher, final String logContext,
 126             final ShardDataTreeMetadata<?>... metadata) {
 127         this.dataTree = Preconditions.checkNotNull(dataTree);
 128         updateSchemaContext(schemaContext);
 129
 130         this.shard = Preconditions.checkNotNull(shard);
 131         this.treeChangeListenerPublisher = Preconditions.checkNotNull(treeChangeListenerPublisher);
 132         this.dataChangeListenerPublisher = Preconditions.checkNotNull(dataChangeListenerPublisher);
 133         this.logContext = Preconditions.checkNotNull(logContext);
 134         this.metadata = ImmutableList.copyOf(metadata);
 135         tip = dataTree;
 136     }
 137
 138     public ShardDataTree(final Shard shard, final SchemaContext schemaContext, final TreeType treeType,
 139             final ShardDataTreeChangeListenerPublisher treeChangeListenerPublisher,
 140             final ShardDataChangeListenerPublisher dataChangeListenerPublisher, final String logContext) {
 141         this(shard, schemaContext, InMemoryDataTreeFactory.getInstance().create(treeType),
 142                 treeChangeListenerPublisher, dataChangeListenerPublisher, logContext);
 143     }
 144
 145     @VisibleForTesting
 146     public ShardDataTree(final Shard shard, final SchemaContext schemaContext, final TreeType treeType) {
 147         this(shard, schemaContext, treeType, new DefaultShardDataTreeChangeListenerPublisher(),
 148                 new DefaultShardDataChangeListenerPublisher(), "");
 149     }
 150
 151     final String logContext() {
 152         return logContext;
 153     }
 154
 155     final Ticker ticker() {
 156         return shard.ticker();
 157     }
 158
 159     public TipProducingDataTree getDataTree() {
 160         return dataTree;
 161     }
 162
 163     SchemaContext getSchemaContext() {
 164         return schemaContext;
 165     }
 166
 167     void updateSchemaContext(final SchemaContext newSchemaContext) {
 168         dataTree.setSchemaContext(newSchemaContext);
 169         this.schemaContext = Preconditions.checkNotNull(newSchemaContext);
 170     }
 171
 172     /**
 173      * Take a snapshot of current state for later recovery.
 174      *
 175      * @return A state snapshot
 176      */
 177     @Nonnull ShardDataTreeSnapshot takeStateSnapshot() {
 178         final NormalizedNode<?, ?> rootNode = dataTree.takeSnapshot().readNode(YangInstanceIdentifier.EMPTY).get();
 179         final Builder<Class<? extends ShardDataTreeSnapshotMetadata<?>>, ShardDataTreeSnapshotMetadata<?>> metaBuilder =
 180                 ImmutableMap.builder();
 181
 182         for (ShardDataTreeMetadata<?> m : metadata) {
 183             final ShardDataTreeSnapshotMetadata<?> meta = m.toSnapshot();
 184             if (meta != null) {
 185                 metaBuilder.put(meta.getType(), meta);
 186             }
 187         }
 188
 189         return new MetadataShardDataTreeSnapshot(rootNode, metaBuilder.build());
 190     }
 191
 192     private void applySnapshot(@Nonnull final ShardDataTreeSnapshot snapshot,
 193             final UnaryOperator<DataTreeModification> wrapper) throws DataValidationFailedException {
 194         final Stopwatch elapsed = Stopwatch.createStarted();
 195
 196         if (!pendingTransactions.isEmpty()) {
 197             LOG.warn("{}: applying state snapshot with pending transactions", logContext);
 198         }
 199
 200         final Map<Class<? extends ShardDataTreeSnapshotMetadata<?>>, ShardDataTreeSnapshotMetadata<?>> snapshotMeta;
 201         if (snapshot instanceof MetadataShardDataTreeSnapshot) {
 202             snapshotMeta = ((MetadataShardDataTreeSnapshot) snapshot).getMetadata();
 203         } else {
 204             snapshotMeta = ImmutableMap.of();
 205         }
 206
 207         for (ShardDataTreeMetadata<?> m : metadata) {
 208             final ShardDataTreeSnapshotMetadata<?> s = snapshotMeta.get(m.getSupportedType());
 209             if (s != null) {
 210                 m.applySnapshot(s);
 211             } else {
 212                 m.reset();
 213             }
 214         }
 215
 216         final DataTreeModification mod = wrapper.apply(dataTree.takeSnapshot().newModification());
 217         // delete everything first
 218         mod.delete(YangInstanceIdentifier.EMPTY);
 219
 220         final java.util.Optional<NormalizedNode<?, ?>> maybeNode = snapshot.getRootNode();
 221         if (maybeNode.isPresent()) {
 222             // Add everything from the remote node back
 223             mod.write(YangInstanceIdentifier.EMPTY, maybeNode.get());
 224         }
 225         mod.ready();
 226
 227         final DataTreeModification unwrapped = unwrap(mod);
 228         dataTree.validate(unwrapped);
 229         DataTreeCandidateTip candidate = dataTree.prepare(unwrapped);
 230         dataTree.commit(candidate);
 231         notifyListeners(candidate);
 232
 233         LOG.debug("{}: state snapshot applied in %s", logContext, elapsed);
 234     }
 235
 236     /**
 237      * Apply a snapshot coming from the leader. This method assumes the leader and follower SchemaContexts match and
 238      * does not perform any pruning.
 239      *
 240      * @param snapshot Snapshot that needs to be applied
 241      * @throws DataValidationFailedException when the snapshot fails to apply
 242      */
 243     void applySnapshot(@Nonnull final ShardDataTreeSnapshot snapshot) throws DataValidationFailedException {
 244         applySnapshot(snapshot, UnaryOperator.identity());
 245     }
 246
 247     private PruningDataTreeModification wrapWithPruning(final DataTreeModification delegate) {
 248         return new PruningDataTreeModification(delegate, dataTree, schemaContext);
 249     }
 250
 251     private static DataTreeModification unwrap(final DataTreeModification modification) {
 252         if (modification instanceof PruningDataTreeModification) {
 253             return ((PruningDataTreeModification)modification).delegate();
 254         }
 255         return modification;
 256     }
 257
 258     /**
 259      * Apply a snapshot coming from recovery. This method does not assume the SchemaContexts match and performs data
 260      * pruning in an attempt to adjust the state to our current SchemaContext.
 261      *
 262      * @param snapshot Snapshot that needs to be applied
 263      * @throws DataValidationFailedException when the snapshot fails to apply
 264      */
 265     void applyRecoverySnapshot(final @Nonnull ShardDataTreeSnapshot snapshot) throws DataValidationFailedException {
 266         applySnapshot(snapshot, this::wrapWithPruning);
 267     }
 268
 269     @SuppressWarnings("checkstyle:IllegalCatch")
 270     private void applyRecoveryCandidate(final DataTreeCandidate candidate) throws DataValidationFailedException {
 271         final PruningDataTreeModification mod = wrapWithPruning(dataTree.takeSnapshot().newModification());
 272         DataTreeCandidates.applyToModification(mod, candidate);
 273         mod.ready();
 274
 275         final DataTreeModification unwrapped = mod.delegate();
 276         LOG.trace("{}: Applying recovery modification {}", logContext, unwrapped);
 277
 278         try {
 279             dataTree.validate(unwrapped);
 280             dataTree.commit(dataTree.prepare(unwrapped));
 281         } catch (Exception e) {
 282             File file = new File(System.getProperty("karaf.data", "."),
 283                     "failed-recovery-payload-" + logContext + ".out");
 284             DataTreeModificationOutput.toFile(file, unwrapped);
 285             throw new IllegalStateException(String.format(
 286                     "%s: Failed to apply recovery payload. Modification data was written to file %s",
 287                     logContext, file), e);
 288         }
 289     }
 290
 291     /**
 292      * Apply a payload coming from recovery. This method does not assume the SchemaContexts match and performs data
 293      * pruning in an attempt to adjust the state to our current SchemaContext.
 294      *
 295      * @param payload Payload
 296      * @throws IOException when the snapshot fails to deserialize
 297      * @throws DataValidationFailedException when the snapshot fails to apply
 298      */
 299     void applyRecoveryPayload(final @Nonnull Payload payload) throws IOException, DataValidationFailedException {
 300         if (payload instanceof CommitTransactionPayload) {
 301             final Entry<TransactionIdentifier, DataTreeCandidate> e =
 302                     ((CommitTransactionPayload) payload).getCandidate();
 303             applyRecoveryCandidate(e.getValue());
 304             allMetadataCommittedTransaction(e.getKey());
 305         } else if (payload instanceof DataTreeCandidatePayload) {
 306             applyRecoveryCandidate(((DataTreeCandidatePayload) payload).getCandidate());
 307         } else {
 308             LOG.debug("{}: ignoring unhandled payload {}", logContext, payload);
 309         }
 310     }
 311
 312     private void applyReplicatedCandidate(final Identifier identifier, final DataTreeCandidate foreign)
 313             throws DataValidationFailedException {
 314         LOG.debug("{}: Applying foreign transaction {}", logContext, identifier);
 315
 316         final DataTreeModification mod = dataTree.takeSnapshot().newModification();
 317         DataTreeCandidates.applyToModification(mod, foreign);
 318         mod.ready();
 319
 320         LOG.trace("{}: Applying foreign modification {}", logContext, mod);
 321         dataTree.validate(mod);
 322         final DataTreeCandidate candidate = dataTree.prepare(mod);
 323         dataTree.commit(candidate);
 324
 325         notifyListeners(candidate);
 326     }
 327
 328     /**
 329      * Apply a payload coming from the leader, which could actually be us. This method assumes the leader and follower
 330      * SchemaContexts match and does not perform any pruning.
 331      *
 332      * @param identifier Payload identifier as returned from RaftActor
 333      * @param payload Payload
 334      * @throws IOException when the snapshot fails to deserialize
 335      * @throws DataValidationFailedException when the snapshot fails to apply
 336      */
 337     void applyReplicatedPayload(final Identifier identifier, final Payload payload) throws IOException,
 338             DataValidationFailedException {
 339         /*
 340          * This is a bit more involved than it needs to be due to to the fact we do not want to be touching the payload
 341          * if we are the leader and it has originated with us.
 342          *
 343          * The identifier will only ever be non-null when we were the leader which achieved consensus. Unfortunately,
 344          * though, this may not be the case anymore, as we are being called some time afterwards and we may not be
 345          * acting in that capacity anymore.
 346          *
 347          * In any case, we know that this is an entry coming from replication, hence we can be sure we will not observe
 348          * pre-Boron state -- which limits the number of options here.
 349          */
 350         if (payload instanceof CommitTransactionPayload) {
 351             if (identifier == null) {
 352                 final Entry<TransactionIdentifier, DataTreeCandidate> e =
 353                         ((CommitTransactionPayload) payload).getCandidate();
 354                 applyReplicatedCandidate(e.getKey(), e.getValue());
 355                 allMetadataCommittedTransaction(e.getKey());
 356             } else {
 357                 Verify.verify(identifier instanceof TransactionIdentifier);
 358                 payloadReplicationComplete((TransactionIdentifier) identifier);
 359             }
 360         } else {
 361             LOG.warn("{}: ignoring unhandled identifier {} payload {}", logContext, identifier, payload);
 362         }
 363     }
 364
 365     private void payloadReplicationComplete(final TransactionIdentifier txId) {
 366         final CommitEntry current = pendingTransactions.peek();
 367         if (current == null) {
 368             LOG.warn("{}: No outstanding transactions, ignoring consensus on transaction {}", logContext, txId);
 369             return;
 370         }
 371
 372         if (!current.cohort.getIdentifier().equals(txId)) {
 373             LOG.warn("{}: Head of queue is {}, ignoring consensus on transaction {}", logContext,
 374                 current.cohort.getIdentifier(), txId);
 375             return;
 376         }
 377
 378         finishCommit(current.cohort);
 379     }
 380
 381     private void allMetadataCommittedTransaction(final TransactionIdentifier txId) {
 382         for (ShardDataTreeMetadata<?> m : metadata) {
 383             m.onTransactionCommitted(txId);
 384         }
 385     }
 386
 387     ShardDataTreeTransactionChain ensureTransactionChain(final LocalHistoryIdentifier localHistoryIdentifier) {
 388         ShardDataTreeTransactionChain chain = transactionChains.get(localHistoryIdentifier);
 389         if (chain == null) {
 390             chain = new ShardDataTreeTransactionChain(localHistoryIdentifier, this);
 391             transactionChains.put(localHistoryIdentifier, chain);
 392         }
 393
 394         return chain;
 395     }
 396
 397     ReadOnlyShardDataTreeTransaction newReadOnlyTransaction(final TransactionIdentifier txId) {
 398         if (txId.getHistoryId().getHistoryId() == 0) {
 399             return new ReadOnlyShardDataTreeTransaction(txId, dataTree.takeSnapshot());
 400         }
 401
 402         return ensureTransactionChain(txId.getHistoryId()).newReadOnlyTransaction(txId);
 403     }
 404
 405     ReadWriteShardDataTreeTransaction newReadWriteTransaction(final TransactionIdentifier txId) {
 406         if (txId.getHistoryId().getHistoryId() == 0) {
 407             return new ReadWriteShardDataTreeTransaction(ShardDataTree.this, txId, dataTree.takeSnapshot()
 408                     .newModification());
 409         }
 410
 411         return ensureTransactionChain(txId.getHistoryId()).newReadWriteTransaction(txId);
 412     }
 413
 414     @VisibleForTesting
 415     public void notifyListeners(final DataTreeCandidate candidate) {
 416         treeChangeListenerPublisher.publishChanges(candidate, logContext);
 417         dataChangeListenerPublisher.publishChanges(candidate, logContext);
 418     }
 419
 420     void notifyOfInitialData(final DataChangeListenerRegistration<AsyncDataChangeListener<YangInstanceIdentifier,
 421             NormalizedNode<?, ?>>> listenerReg, final Optional<DataTreeCandidate> currentState) {
 422         if (currentState.isPresent()) {
 423             ShardDataChangeListenerPublisher localPublisher = dataChangeListenerPublisher.newInstance();
 424             localPublisher.registerDataChangeListener(listenerReg.getPath(), listenerReg.getInstance(),
 425                     listenerReg.getScope());
 426             localPublisher.publishChanges(currentState.get(), logContext);
 427         }
 428     }
 429
 430     void notifyOfInitialData(final YangInstanceIdentifier path, final DOMDataTreeChangeListener listener,
 431             final Optional<DataTreeCandidate> currentState) {
 432         if (currentState.isPresent()) {
 433             ShardDataTreeChangeListenerPublisher localPublisher = treeChangeListenerPublisher.newInstance();
 434             localPublisher.registerTreeChangeListener(path, listener);
 435             localPublisher.publishChanges(currentState.get(), logContext);
 436         }
 437     }
 438
 439     void closeAllTransactionChains() {
 440         for (ShardDataTreeTransactionChain chain : transactionChains.values()) {
 441             chain.close();
 442         }
 443
 444         transactionChains.clear();
 445     }
 446
 447     void closeTransactionChain(final LocalHistoryIdentifier transactionChainId) {
 448         final ShardDataTreeTransactionChain chain = transactionChains.remove(transactionChainId);
 449         if (chain != null) {
 450             chain.close();
 451         } else {
 452             LOG.debug("{}: Closing non-existent transaction chain {}", logContext, transactionChainId);
 453         }
 454     }
 455
 456     Entry<DataChangeListenerRegistration<AsyncDataChangeListener<YangInstanceIdentifier, NormalizedNode<?, ?>>>,
 457             Optional<DataTreeCandidate>> registerChangeListener(final YangInstanceIdentifier path,
 458                     final AsyncDataChangeListener<YangInstanceIdentifier, NormalizedNode<?, ?>> listener,
 459                     final DataChangeScope scope) {
 460         DataChangeListenerRegistration<AsyncDataChangeListener<YangInstanceIdentifier, NormalizedNode<?, ?>>> reg =
 461                 dataChangeListenerPublisher.registerDataChangeListener(path, listener, scope);
 462
 463         return new SimpleEntry<>(reg, readCurrentData());
 464     }
 465
 466     private Optional<DataTreeCandidate> readCurrentData() {
 467         final Optional<NormalizedNode<?, ?>> currentState =
 468                 dataTree.takeSnapshot().readNode(YangInstanceIdentifier.EMPTY);
 469         return currentState.isPresent() ? Optional.of(DataTreeCandidates.fromNormalizedNode(
 470             YangInstanceIdentifier.EMPTY, currentState.get())) : Optional.<DataTreeCandidate>absent();
 471     }
 472
 473     public Entry<ListenerRegistration<DOMDataTreeChangeListener>, Optional<DataTreeCandidate>>
 474             registerTreeChangeListener(final YangInstanceIdentifier path, final DOMDataTreeChangeListener listener) {
 475         final ListenerRegistration<DOMDataTreeChangeListener> reg =
 476                 treeChangeListenerPublisher.registerTreeChangeListener(path, listener);
 477
 478         return new SimpleEntry<>(reg, readCurrentData());
 479     }
 480
 481     int getQueueSize() {
 482         return pendingTransactions.size();
 483     }
 484
 485     @Override
 486     void abortTransaction(final AbstractShardDataTreeTransaction<?> transaction) {
 487         // Intentional no-op
 488     }
 489
 490     @Override
 491     ShardDataTreeCohort finishTransaction(final ReadWriteShardDataTreeTransaction transaction) {
 492         final DataTreeModification snapshot = transaction.getSnapshot();
 493         snapshot.ready();
 494
 495         return createReadyCohort(transaction.getIdentifier(), snapshot);
 496     }
 497
 498     public Optional<NormalizedNode<?, ?>> readNode(final YangInstanceIdentifier path) {
 499         return dataTree.takeSnapshot().readNode(path);
 500     }
 501
 502     DataTreeSnapshot takeSnapshot() {
 503         return dataTree.takeSnapshot();
 504     }
 505
 506     @VisibleForTesting
 507     public DataTreeModification newModification() {
 508         return dataTree.takeSnapshot().newModification();
 509     }
 510
 511     /**
 512      * Commits a modification.
 513      *
 514      * @deprecated This method violates DataTree containment and will be removed.
 515      */
 516     @VisibleForTesting
 517     @Deprecated
 518     public DataTreeCandidate commit(final DataTreeModification modification) throws DataValidationFailedException {
 519         // Direct modification commit is a utility, which cannot be used while we have transactions in-flight
 520         Preconditions.checkState(tip == dataTree, "Cannot modify data tree while transacgitons are pending");
 521
 522         modification.ready();
 523         dataTree.validate(modification);
 524         DataTreeCandidate candidate = dataTree.prepare(modification);
 525         dataTree.commit(candidate);
 526         return candidate;
 527     }
 528
 529     public Collection<ShardDataTreeCohort> getAndClearPendingTransactions() {
 530         Collection<ShardDataTreeCohort> ret = new ArrayList<>(pendingTransactions.size());
 531         for (CommitEntry entry: pendingTransactions) {
 532             ret.add(entry.cohort);
 533         }
 534
 535         pendingTransactions.clear();
 536         tip = dataTree;
 537         return ret;
 538     }
 539
 540     @SuppressWarnings("checkstyle:IllegalCatch")
 541     private void processNextTransaction() {
 542         while (!pendingTransactions.isEmpty()) {
 543             final CommitEntry entry = pendingTransactions.peek();
 544             final SimpleShardDataTreeCohort cohort = entry.cohort;
 545             final DataTreeModification modification = cohort.getDataTreeModification();
 546
 547             if (cohort.getState() != State.CAN_COMMIT_PENDING) {
 548                 break;
 549             }
 550
 551             LOG.debug("{}: Validating transaction {}", logContext, cohort.getIdentifier());
 552             Exception cause;
 553             try {
 554                 tip.validate(modification);
 555                 LOG.debug("{}: Transaction {} validated", logContext, cohort.getIdentifier());
 556                 cohort.successfulCanCommit();
 557                 entry.lastAccess = shard.ticker().read();
 558                 return;
 559             } catch (ConflictingModificationAppliedException e) {
 560                 LOG.warn("{}: Store Tx {}: Conflicting modification for path {}.", logContext, cohort.getIdentifier(),
 561                     e.getPath());
 562                 cause = new OptimisticLockFailedException("Optimistic lock failed.", e);
 563             } catch (DataValidationFailedException e) {
 564                 LOG.warn("{}: Store Tx {}: Data validation failed for path {}.", logContext, cohort.getIdentifier(),
 565                     e.getPath(), e);
 566
 567                 // For debugging purposes, allow dumping of the modification. Coupled with the above
 568                 // precondition log, it should allow us to understand what went on.
 569                 LOG.debug("{}: Store Tx {}: modifications: {} tree: {}", cohort.getIdentifier(), modification,
 570                         dataTree);
 571                 cause = new TransactionCommitFailedException("Data did not pass validation.", e);
 572             } catch (Exception e) {
 573                 LOG.warn("{}: Unexpected failure in validation phase", logContext, e);
 574                 cause = e;
 575             }
 576
 577             // Failure path: propagate the failure, remove the transaction from the queue and loop to the next one
 578             pendingTransactions.poll().cohort.failedCanCommit(cause);
 579         }
 580
 581         maybeRunOperationOnPendingTransactionsComplete();
 582     }
 583
 584     void startCanCommit(final SimpleShardDataTreeCohort cohort) {
 585         final SimpleShardDataTreeCohort current = pendingTransactions.peek().cohort;
 586         if (!cohort.equals(current)) {
 587             LOG.debug("{}: Transaction {} scheduled for canCommit step", logContext, cohort.getIdentifier());
 588             return;
 589         }
 590
 591         processNextTransaction();
 592     }
 593
 594     private void failPreCommit(final Exception cause) {
 595         shard.getShardMBean().incrementFailedTransactionsCount();
 596         pendingTransactions.poll().cohort.failedPreCommit(cause);
 597         processNextTransaction();
 598     }
 599
 600     @SuppressWarnings("checkstyle:IllegalCatch")
 601     void startPreCommit(final SimpleShardDataTreeCohort cohort) {
 602         final CommitEntry entry = pendingTransactions.peek();
 603         Preconditions.checkState(entry != null, "Attempted to pre-commit of %s when no transactions pending", cohort);
 604
 605         final SimpleShardDataTreeCohort current = entry.cohort;
 606         Verify.verify(cohort.equals(current), "Attempted to pre-commit %s while %s is pending", cohort, current);
 607         final DataTreeCandidateTip candidate;
 608         try {
 609             candidate = tip.prepare(cohort.getDataTreeModification());
 610         } catch (Exception e) {
 611             failPreCommit(e);
 612             return;
 613         }
 614
 615         try {
 616             cohort.userPreCommit(candidate);
 617         } catch (ExecutionException | TimeoutException e) {
 618             failPreCommit(e);
 619             return;
 620         }
 621
 622         // Set the tip of the data tree.
 623         tip = Verify.verifyNotNull(candidate);
 624
 625         entry.lastAccess = shard.ticker().read();
 626         cohort.successfulPreCommit(candidate);
 627     }
 628
 629     private void failCommit(final Exception cause) {
 630         shard.getShardMBean().incrementFailedTransactionsCount();
 631         pendingTransactions.poll().cohort.failedCommit(cause);
 632         processNextTransaction();
 633     }
 634
 635     @SuppressWarnings("checkstyle:IllegalCatch")
 636     private void finishCommit(final SimpleShardDataTreeCohort cohort) {
 637         final TransactionIdentifier txId = cohort.getIdentifier();
 638         final DataTreeCandidate candidate = cohort.getCandidate();
 639
 640         LOG.debug("{}: Resuming commit of transaction {}", logContext, txId);
 641
 642         try {
 643             dataTree.commit(candidate);
 644         } catch (Exception e) {
 645             LOG.error("{}: Failed to commit transaction {}", logContext, txId, e);
 646             failCommit(e);
 647             return;
 648         }
 649
 650         // All pending candidates have been committed, reset the tip to the data tree
 651         if (tip == candidate) {
 652             tip = dataTree;
 653         }
 654
 655         shard.getShardMBean().incrementCommittedTransactionCount();
 656         shard.getShardMBean().setLastCommittedTransactionTime(System.currentTimeMillis());
 657
 658         // FIXME: propagate journal index
 659         pendingTransactions.poll().cohort.successfulCommit(UnsignedLong.ZERO);
 660
 661         LOG.trace("{}: Transaction {} committed, proceeding to notify", logContext, txId);
 662         notifyListeners(candidate);
 663
 664         processNextTransaction();
 665     }
 666
 667     void startCommit(final SimpleShardDataTreeCohort cohort, final DataTreeCandidate candidate) {
 668         final CommitEntry entry = pendingTransactions.peek();
 669         Preconditions.checkState(entry != null, "Attempted to start commit of %s when no transactions pending", cohort);
 670
 671         final SimpleShardDataTreeCohort current = entry.cohort;
 672         Verify.verify(cohort.equals(current), "Attempted to commit %s while %s is pending", cohort, current);
 673
 674         if (shard.canSkipPayload() || candidate.getRootNode().getModificationType() == ModificationType.UNMODIFIED) {
 675             LOG.debug("{}: No replication required, proceeding to finish commit", logContext);
 676             finishCommit(cohort);
 677             return;
 678         }
 679
 680         final TransactionIdentifier txId = cohort.getIdentifier();
 681         final Payload payload;
 682         try {
 683             payload = CommitTransactionPayload.create(txId, candidate);
 684         } catch (IOException e) {
 685             LOG.error("{}: Failed to encode transaction {} candidate {}", logContext, txId, candidate, e);
 686             pendingTransactions.poll().cohort.failedCommit(e);
 687             return;
 688         }
 689
 690         // Once completed, we will continue via payloadReplicationComplete
 691         entry.lastAccess = shard.ticker().read();
 692         shard.persistPayload(txId, payload);
 693         LOG.debug("{}: Transaction {} submitted to persistence", logContext, txId);
 694     }
 695
 696     void processCohortRegistryCommand(final ActorRef sender, final CohortRegistryCommand message) {
 697         cohortRegistry.process(sender, message);
 698     }
 699
 700     @Override
 701     ShardDataTreeCohort createReadyCohort(final TransactionIdentifier txId,
 702             final DataTreeModification modification) {
 703         SimpleShardDataTreeCohort cohort = new SimpleShardDataTreeCohort(this, modification, txId,
 704                 cohortRegistry.createCohort(schemaContext, txId, COMMIT_STEP_TIMEOUT));
 705         pendingTransactions.add(new CommitEntry(cohort, shard.ticker().read()));
 706         return cohort;
 707     }
 708
 709     @SuppressFBWarnings(value = "DB_DUPLICATE_SWITCH_CLAUSES", justification = "See inline comments below.")
 710     void checkForExpiredTransactions(final long transactionCommitTimeoutMillis) {
 711         final long timeout = TimeUnit.MILLISECONDS.toNanos(transactionCommitTimeoutMillis);
 712         final long now = shard.ticker().read();
 713         final CommitEntry currentTx = pendingTransactions.peek();
 714         if (currentTx != null && currentTx.lastAccess + timeout < now) {
 715             LOG.warn("{}: Current transaction {} has timed out after {} ms in state {}", logContext,
 716                     currentTx.cohort.getIdentifier(), transactionCommitTimeoutMillis, currentTx.cohort.getState());
 717             boolean processNext = true;
 718             switch (currentTx.cohort.getState()) {
 719                 case CAN_COMMIT_PENDING:
 720                     pendingTransactions.remove().cohort.failedCanCommit(new TimeoutException());
 721                     break;
 722                 case CAN_COMMIT_COMPLETE:
 723                     // The suppression of the FindBugs "DB_DUPLICATE_SWITCH_CLAUSES" warning pertains to this clause
 724                     // whose code is duplicated with PRE_COMMIT_COMPLETE. The clauses aren't combined in case the code
 725                     // in PRE_COMMIT_COMPLETE is changed.
 726                     pendingTransactions.remove().cohort.reportFailure(new TimeoutException());
 727                     break;
 728                 case PRE_COMMIT_PENDING:
 729                     pendingTransactions.remove().cohort.failedPreCommit(new TimeoutException());
 730                     break;
 731                 case PRE_COMMIT_COMPLETE:
 732                     // FIXME: this is a legacy behavior problem. Three-phase commit protocol specifies that after we
 733                     //        are ready we should commit the transaction, not abort it. Our current software stack does
 734                     //        not allow us to do that consistently, because we persist at the time of commit, hence
 735                     //        we can end up in a state where we have pre-committed a transaction, then a leader failover
 736                     //        occurred ... the new leader does not see the pre-committed transaction and does not have
 737                     //        a running timer. To fix this we really need two persistence events.
 738                     //
 739                     //        The first one, done at pre-commit time will hold the transaction payload. When consensus
 740                     //        is reached, we exit the pre-commit phase and start the pre-commit timer. Followers do not
 741                     //        apply the state in this event.
 742                     //
 743                     //        The second one, done at commit (or abort) time holds only the transaction identifier and
 744                     //        signals to followers that the state should (or should not) be applied.
 745                     //
 746                     //        In order to make the pre-commit timer working across failovers, though, we need
 747                     //        a per-shard cluster-wide monotonic time, so a follower becoming the leader can accurately
 748                     //        restart the timer.
 749                     pendingTransactions.remove().cohort.reportFailure(new TimeoutException());
 750                     break;
 751                 case COMMIT_PENDING:
 752                     LOG.warn("{}: Transaction {} is still committing, cannot abort", logContext,
 753                         currentTx.cohort.getIdentifier());
 754                     currentTx.lastAccess = now;
 755                     processNext = false;
 756                     return;
 757                 case ABORTED:
 758                 case COMMITTED:
 759                 case FAILED:
 760                 case READY:
 761                 default:
 762                     pendingTransactions.remove();
 763             }
 764
 765             if (processNext) {
 766                 processNextTransaction();
 767             }
 768         }
 769     }
 770
 771     boolean startAbort(final SimpleShardDataTreeCohort cohort) {
 772         final Iterator<CommitEntry> it = pendingTransactions.iterator();
 773         if (!it.hasNext()) {
 774             LOG.debug("{}: no open transaction while attempting to abort {}", logContext, cohort.getIdentifier());
 775             return true;
 776         }
 777
 778         // First entry is special, as it may already be committing
 779         final CommitEntry first = it.next();
 780         if (cohort.equals(first.cohort)) {
 781             if (cohort.getState() != State.COMMIT_PENDING) {
 782                 LOG.debug("{}: aborting head of queue {} in state {}", logContext, cohort.getIdentifier(),
 783                     cohort.getIdentifier());
 784
 785                 it.remove();
 786                 rebasePreCommittedTransactions(it, dataTree);
 787                 processNextTransaction();
 788                 return true;
 789             }
 790
 791             LOG.warn("{}: transaction {} is committing, skipping abort", logContext, cohort.getIdentifier());
 792             return false;
 793         }
 794
 795         TipProducingDataTreeTip newTip = dataTree;
 796         while (it.hasNext()) {
 797             final CommitEntry e = it.next();
 798             if (cohort.equals(e.cohort)) {
 799                 LOG.debug("{}: aborting queued transaction {}", logContext, cohort.getIdentifier());
 800                 it.remove();
 801                 rebasePreCommittedTransactions(it, newTip);
 802                 return true;
 803             } else {
 804                 newTip = cohort.getCandidate();
 805             }
 806         }
 807
 808         LOG.debug("{}: aborted transaction {} not found in the queue", logContext, cohort.getIdentifier());
 809         return true;
 810     }
 811
 812     @SuppressWarnings("checkstyle:IllegalCatch")
 813     private void rebasePreCommittedTransactions(Iterator<CommitEntry> iter, TipProducingDataTreeTip newTip) {
 814         tip = newTip;
 815         while (iter.hasNext()) {
 816             final SimpleShardDataTreeCohort cohort = iter.next().cohort;
 817             if (cohort.getState() == State.CAN_COMMIT_COMPLETE) {
 818                 LOG.debug("{}: Revalidating queued transaction {}", logContext, cohort.getIdentifier());
 819
 820                 try {
 821                     tip.validate(cohort.getDataTreeModification());
 822                 } catch (DataValidationFailedException | RuntimeException e) {
 823                     LOG.debug("{}: Failed to revalidate queued transaction {}", logContext, cohort.getIdentifier(), e);
 824                     cohort.reportFailure(e);
 825                 }
 826             } else if (cohort.getState() == State.PRE_COMMIT_COMPLETE) {
 827                 LOG.debug("{}: Repreparing queued transaction {}", logContext, cohort.getIdentifier());
 828
 829                 try {
 830                     tip.validate(cohort.getDataTreeModification());
 831                     DataTreeCandidateTip candidate = tip.prepare(cohort.getDataTreeModification());
 832                     cohort.userPreCommit(candidate);
 833
 834                     cohort.setNewCandidate(candidate);
 835                     tip = candidate;
 836                 } catch (ExecutionException | TimeoutException | RuntimeException | DataValidationFailedException e) {
 837                     LOG.debug("{}: Failed to reprepare queued transaction {}", logContext, cohort.getIdentifier(), e);
 838                     cohort.reportFailure(e);
 839                 }
 840             }
 841         }
 842     }
 843
 844     void setRunOnPendingTransactionsComplete(final Runnable operation) {
 845         runOnPendingTransactionsComplete = operation;
 846         maybeRunOperationOnPendingTransactionsComplete();
 847     }
 848
 849     private void maybeRunOperationOnPendingTransactionsComplete() {
 850         if (runOnPendingTransactionsComplete != null && pendingTransactions.isEmpty()) {
 851             LOG.debug("{}: Pending transactions complete - running operation {}", logContext,
 852                     runOnPendingTransactionsComplete);
 853
 854             runOnPendingTransactionsComplete.run();
 855             runOnPendingTransactionsComplete = null;
 856         }
 857     }
 858 }