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 YangInstanceIdentifier root,
 140             final ShardDataTreeChangeListenerPublisher treeChangeListenerPublisher,
 141             final ShardDataChangeListenerPublisher dataChangeListenerPublisher, final String logContext) {
 142         this(shard, schemaContext, InMemoryDataTreeFactory.getInstance().create(treeType, root),
 143                 treeChangeListenerPublisher, dataChangeListenerPublisher, logContext);
 144     }
 145
 146     @VisibleForTesting
 147     public ShardDataTree(final Shard shard, final SchemaContext schemaContext, final TreeType treeType) {
 148         this(shard, schemaContext, treeType, YangInstanceIdentifier.EMPTY,
 149                 new DefaultShardDataTreeChangeListenerPublisher(),
 150                 new DefaultShardDataChangeListenerPublisher(), "");
 151     }
 152
 153     final String logContext() {
 154         return logContext;
 155     }
 156
 157     final Ticker ticker() {
 158         return shard.ticker();
 159     }
 160
 161     public TipProducingDataTree getDataTree() {
 162         return dataTree;
 163     }
 164
 165     SchemaContext getSchemaContext() {
 166         return schemaContext;
 167     }
 168
 169     void updateSchemaContext(final SchemaContext newSchemaContext) {
 170         dataTree.setSchemaContext(newSchemaContext);
 171         this.schemaContext = Preconditions.checkNotNull(newSchemaContext);
 172     }
 173
 174     /**
 175      * Take a snapshot of current state for later recovery.
 176      *
 177      * @return A state snapshot
 178      */
 179     @Nonnull ShardDataTreeSnapshot takeStateSnapshot() {
 180         final NormalizedNode<?, ?> rootNode = dataTree.takeSnapshot().readNode(YangInstanceIdentifier.EMPTY).get();
 181         final Builder<Class<? extends ShardDataTreeSnapshotMetadata<?>>, ShardDataTreeSnapshotMetadata<?>> metaBuilder =
 182                 ImmutableMap.builder();
 183
 184         for (ShardDataTreeMetadata<?> m : metadata) {
 185             final ShardDataTreeSnapshotMetadata<?> meta = m.toSnapshot();
 186             if (meta != null) {
 187                 metaBuilder.put(meta.getType(), meta);
 188             }
 189         }
 190
 191         return new MetadataShardDataTreeSnapshot(rootNode, metaBuilder.build());
 192     }
 193
 194     private void applySnapshot(@Nonnull final ShardDataTreeSnapshot snapshot,
 195             final UnaryOperator<DataTreeModification> wrapper) throws DataValidationFailedException {
 196         final Stopwatch elapsed = Stopwatch.createStarted();
 197
 198         if (!pendingTransactions.isEmpty()) {
 199             LOG.warn("{}: applying state snapshot with pending transactions", logContext);
 200         }
 201
 202         final Map<Class<? extends ShardDataTreeSnapshotMetadata<?>>, ShardDataTreeSnapshotMetadata<?>> snapshotMeta;
 203         if (snapshot instanceof MetadataShardDataTreeSnapshot) {
 204             snapshotMeta = ((MetadataShardDataTreeSnapshot) snapshot).getMetadata();
 205         } else {
 206             snapshotMeta = ImmutableMap.of();
 207         }
 208
 209         for (ShardDataTreeMetadata<?> m : metadata) {
 210             final ShardDataTreeSnapshotMetadata<?> s = snapshotMeta.get(m.getSupportedType());
 211             if (s != null) {
 212                 m.applySnapshot(s);
 213             } else {
 214                 m.reset();
 215             }
 216         }
 217
 218         final DataTreeModification mod = wrapper.apply(dataTree.takeSnapshot().newModification());
 219         // delete everything first
 220         mod.delete(YangInstanceIdentifier.EMPTY);
 221
 222         final java.util.Optional<NormalizedNode<?, ?>> maybeNode = snapshot.getRootNode();
 223         if (maybeNode.isPresent()) {
 224             // Add everything from the remote node back
 225             mod.write(YangInstanceIdentifier.EMPTY, maybeNode.get());
 226         }
 227         mod.ready();
 228
 229         final DataTreeModification unwrapped = unwrap(mod);
 230         dataTree.validate(unwrapped);
 231         DataTreeCandidateTip candidate = dataTree.prepare(unwrapped);
 232         dataTree.commit(candidate);
 233         notifyListeners(candidate);
 234
 235         LOG.debug("{}: state snapshot applied in %s", logContext, elapsed);
 236     }
 237
 238     /**
 239      * Apply a snapshot coming from the leader. This method assumes the leader and follower SchemaContexts match and
 240      * does not perform any pruning.
 241      *
 242      * @param snapshot Snapshot that needs to be applied
 243      * @throws DataValidationFailedException when the snapshot fails to apply
 244      */
 245     void applySnapshot(@Nonnull final ShardDataTreeSnapshot snapshot) throws DataValidationFailedException {
 246         applySnapshot(snapshot, UnaryOperator.identity());
 247     }
 248
 249     private PruningDataTreeModification wrapWithPruning(final DataTreeModification delegate) {
 250         return new PruningDataTreeModification(delegate, dataTree, schemaContext);
 251     }
 252
 253     private static DataTreeModification unwrap(final DataTreeModification modification) {
 254         if (modification instanceof PruningDataTreeModification) {
 255             return ((PruningDataTreeModification)modification).delegate();
 256         }
 257         return modification;
 258     }
 259
 260     /**
 261      * Apply a snapshot coming from recovery. This method does not assume the SchemaContexts match and performs data
 262      * pruning in an attempt to adjust the state to our current SchemaContext.
 263      *
 264      * @param snapshot Snapshot that needs to be applied
 265      * @throws DataValidationFailedException when the snapshot fails to apply
 266      */
 267     void applyRecoverySnapshot(final @Nonnull ShardDataTreeSnapshot snapshot) throws DataValidationFailedException {
 268         applySnapshot(snapshot, this::wrapWithPruning);
 269     }
 270
 271     @SuppressWarnings("checkstyle:IllegalCatch")
 272     private void applyRecoveryCandidate(final DataTreeCandidate candidate) throws DataValidationFailedException {
 273         final PruningDataTreeModification mod = wrapWithPruning(dataTree.takeSnapshot().newModification());
 274         DataTreeCandidates.applyToModification(mod, candidate);
 275         mod.ready();
 276
 277         final DataTreeModification unwrapped = mod.delegate();
 278         LOG.trace("{}: Applying recovery modification {}", logContext, unwrapped);
 279
 280         try {
 281             dataTree.validate(unwrapped);
 282             dataTree.commit(dataTree.prepare(unwrapped));
 283         } catch (Exception e) {
 284             File file = new File(System.getProperty("karaf.data", "."),
 285                     "failed-recovery-payload-" + logContext + ".out");
 286             DataTreeModificationOutput.toFile(file, unwrapped);
 287             throw new IllegalStateException(String.format(
 288                     "%s: Failed to apply recovery payload. Modification data was written to file %s",
 289                     logContext, file), e);
 290         }
 291     }
 292
 293     /**
 294      * Apply a payload coming from recovery. This method does not assume the SchemaContexts match and performs data
 295      * pruning in an attempt to adjust the state to our current SchemaContext.
 296      *
 297      * @param payload Payload
 298      * @throws IOException when the snapshot fails to deserialize
 299      * @throws DataValidationFailedException when the snapshot fails to apply
 300      */
 301     void applyRecoveryPayload(final @Nonnull Payload payload) throws IOException, DataValidationFailedException {
 302         if (payload instanceof CommitTransactionPayload) {
 303             final Entry<TransactionIdentifier, DataTreeCandidate> e =
 304                     ((CommitTransactionPayload) payload).getCandidate();
 305             applyRecoveryCandidate(e.getValue());
 306             allMetadataCommittedTransaction(e.getKey());
 307         } else if (payload instanceof DataTreeCandidatePayload) {
 308             applyRecoveryCandidate(((DataTreeCandidatePayload) payload).getCandidate());
 309         } else {
 310             LOG.debug("{}: ignoring unhandled payload {}", logContext, payload);
 311         }
 312     }
 313
 314     private void applyReplicatedCandidate(final Identifier identifier, final DataTreeCandidate foreign)
 315             throws DataValidationFailedException {
 316         LOG.debug("{}: Applying foreign transaction {}", logContext, identifier);
 317
 318         final DataTreeModification mod = dataTree.takeSnapshot().newModification();
 319         DataTreeCandidates.applyToModification(mod, foreign);
 320         mod.ready();
 321
 322         LOG.trace("{}: Applying foreign modification {}", logContext, mod);
 323         dataTree.validate(mod);
 324         final DataTreeCandidate candidate = dataTree.prepare(mod);
 325         dataTree.commit(candidate);
 326
 327         notifyListeners(candidate);
 328     }
 329
 330     /**
 331      * Apply a payload coming from the leader, which could actually be us. This method assumes the leader and follower
 332      * SchemaContexts match and does not perform any pruning.
 333      *
 334      * @param identifier Payload identifier as returned from RaftActor
 335      * @param payload Payload
 336      * @throws IOException when the snapshot fails to deserialize
 337      * @throws DataValidationFailedException when the snapshot fails to apply
 338      */
 339     void applyReplicatedPayload(final Identifier identifier, final Payload payload) throws IOException,
 340             DataValidationFailedException {
 341         /*
 342          * 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
 343          * if we are the leader and it has originated with us.
 344          *
 345          * The identifier will only ever be non-null when we were the leader which achieved consensus. Unfortunately,
 346          * though, this may not be the case anymore, as we are being called some time afterwards and we may not be
 347          * acting in that capacity anymore.
 348          *
 349          * In any case, we know that this is an entry coming from replication, hence we can be sure we will not observe
 350          * pre-Boron state -- which limits the number of options here.
 351          */
 352         if (payload instanceof CommitTransactionPayload) {
 353             if (identifier == null) {
 354                 final Entry<TransactionIdentifier, DataTreeCandidate> e =
 355                         ((CommitTransactionPayload) payload).getCandidate();
 356                 applyReplicatedCandidate(e.getKey(), e.getValue());
 357                 allMetadataCommittedTransaction(e.getKey());
 358             } else {
 359                 Verify.verify(identifier instanceof TransactionIdentifier);
 360                 payloadReplicationComplete((TransactionIdentifier) identifier);
 361             }
 362         } else {
 363             LOG.warn("{}: ignoring unhandled identifier {} payload {}", logContext, identifier, payload);
 364         }
 365     }
 366
 367     private void payloadReplicationComplete(final TransactionIdentifier txId) {
 368         final CommitEntry current = pendingTransactions.peek();
 369         if (current == null) {
 370             LOG.warn("{}: No outstanding transactions, ignoring consensus on transaction {}", logContext, txId);
 371             return;
 372         }
 373
 374         if (!current.cohort.getIdentifier().equals(txId)) {
 375             LOG.warn("{}: Head of queue is {}, ignoring consensus on transaction {}", logContext,
 376                 current.cohort.getIdentifier(), txId);
 377             return;
 378         }
 379
 380         finishCommit(current.cohort);
 381     }
 382
 383     private void allMetadataCommittedTransaction(final TransactionIdentifier txId) {
 384         for (ShardDataTreeMetadata<?> m : metadata) {
 385             m.onTransactionCommitted(txId);
 386         }
 387     }
 388
 389     ShardDataTreeTransactionChain ensureTransactionChain(final LocalHistoryIdentifier localHistoryIdentifier) {
 390         ShardDataTreeTransactionChain chain = transactionChains.get(localHistoryIdentifier);
 391         if (chain == null) {
 392             chain = new ShardDataTreeTransactionChain(localHistoryIdentifier, this);
 393             transactionChains.put(localHistoryIdentifier, chain);
 394         }
 395
 396         return chain;
 397     }
 398
 399     ReadOnlyShardDataTreeTransaction newReadOnlyTransaction(final TransactionIdentifier txId) {
 400         if (txId.getHistoryId().getHistoryId() == 0) {
 401             return new ReadOnlyShardDataTreeTransaction(txId, dataTree.takeSnapshot());
 402         }
 403
 404         return ensureTransactionChain(txId.getHistoryId()).newReadOnlyTransaction(txId);
 405     }
 406
 407     ReadWriteShardDataTreeTransaction newReadWriteTransaction(final TransactionIdentifier txId) {
 408         if (txId.getHistoryId().getHistoryId() == 0) {
 409             return new ReadWriteShardDataTreeTransaction(ShardDataTree.this, txId, dataTree.takeSnapshot()
 410                     .newModification());
 411         }
 412
 413         return ensureTransactionChain(txId.getHistoryId()).newReadWriteTransaction(txId);
 414     }
 415
 416     @VisibleForTesting
 417     public void notifyListeners(final DataTreeCandidate candidate) {
 418         treeChangeListenerPublisher.publishChanges(candidate, logContext);
 419         dataChangeListenerPublisher.publishChanges(candidate, logContext);
 420     }
 421
 422     void notifyOfInitialData(final DataChangeListenerRegistration<AsyncDataChangeListener<YangInstanceIdentifier,
 423             NormalizedNode<?, ?>>> listenerReg, final Optional<DataTreeCandidate> currentState) {
 424         if (currentState.isPresent()) {
 425             ShardDataChangeListenerPublisher localPublisher = dataChangeListenerPublisher.newInstance();
 426             localPublisher.registerDataChangeListener(listenerReg.getPath(), listenerReg.getInstance(),
 427                     listenerReg.getScope());
 428             localPublisher.publishChanges(currentState.get(), logContext);
 429         }
 430     }
 431
 432     void notifyOfInitialData(final YangInstanceIdentifier path, final DOMDataTreeChangeListener listener,
 433             final Optional<DataTreeCandidate> currentState) {
 434         if (currentState.isPresent()) {
 435             ShardDataTreeChangeListenerPublisher localPublisher = treeChangeListenerPublisher.newInstance();
 436             localPublisher.registerTreeChangeListener(path, listener);
 437             localPublisher.publishChanges(currentState.get(), logContext);
 438         }
 439     }
 440
 441     void closeAllTransactionChains() {
 442         for (ShardDataTreeTransactionChain chain : transactionChains.values()) {
 443             chain.close();
 444         }
 445
 446         transactionChains.clear();
 447     }
 448
 449     void closeTransactionChain(final LocalHistoryIdentifier transactionChainId) {
 450         final ShardDataTreeTransactionChain chain = transactionChains.remove(transactionChainId);
 451         if (chain != null) {
 452             chain.close();
 453         } else {
 454             LOG.debug("{}: Closing non-existent transaction chain {}", logContext, transactionChainId);
 455         }
 456     }
 457
 458     Entry<DataChangeListenerRegistration<AsyncDataChangeListener<YangInstanceIdentifier, NormalizedNode<?, ?>>>,
 459             Optional<DataTreeCandidate>> registerChangeListener(final YangInstanceIdentifier path,
 460                     final AsyncDataChangeListener<YangInstanceIdentifier, NormalizedNode<?, ?>> listener,
 461                     final DataChangeScope scope) {
 462         DataChangeListenerRegistration<AsyncDataChangeListener<YangInstanceIdentifier, NormalizedNode<?, ?>>> reg =
 463                 dataChangeListenerPublisher.registerDataChangeListener(path, listener, scope);
 464
 465         return new SimpleEntry<>(reg, readCurrentData());
 466     }
 467
 468     private Optional<DataTreeCandidate> readCurrentData() {
 469         final Optional<NormalizedNode<?, ?>> currentState =
 470                 dataTree.takeSnapshot().readNode(YangInstanceIdentifier.EMPTY);
 471         return currentState.isPresent() ? Optional.of(DataTreeCandidates.fromNormalizedNode(
 472             YangInstanceIdentifier.EMPTY, currentState.get())) : Optional.<DataTreeCandidate>absent();
 473     }
 474
 475     public Entry<ListenerRegistration<DOMDataTreeChangeListener>, Optional<DataTreeCandidate>>
 476             registerTreeChangeListener(final YangInstanceIdentifier path, final DOMDataTreeChangeListener listener) {
 477         final ListenerRegistration<DOMDataTreeChangeListener> reg =
 478                 treeChangeListenerPublisher.registerTreeChangeListener(path, listener);
 479
 480         return new SimpleEntry<>(reg, readCurrentData());
 481     }
 482
 483     int getQueueSize() {
 484         return pendingTransactions.size();
 485     }
 486
 487     @Override
 488     void abortTransaction(final AbstractShardDataTreeTransaction<?> transaction) {
 489         // Intentional no-op
 490     }
 491
 492     @Override
 493     ShardDataTreeCohort finishTransaction(final ReadWriteShardDataTreeTransaction transaction) {
 494         final DataTreeModification snapshot = transaction.getSnapshot();
 495         snapshot.ready();
 496
 497         return createReadyCohort(transaction.getIdentifier(), snapshot);
 498     }
 499
 500     public Optional<NormalizedNode<?, ?>> readNode(final YangInstanceIdentifier path) {
 501         return dataTree.takeSnapshot().readNode(path);
 502     }
 503
 504     DataTreeSnapshot takeSnapshot() {
 505         return dataTree.takeSnapshot();
 506     }
 507
 508     @VisibleForTesting
 509     public DataTreeModification newModification() {
 510         return dataTree.takeSnapshot().newModification();
 511     }
 512
 513     /**
 514      * Commits a modification.
 515      *
 516      * @deprecated This method violates DataTree containment and will be removed.
 517      */
 518     @VisibleForTesting
 519     @Deprecated
 520     public DataTreeCandidate commit(final DataTreeModification modification) throws DataValidationFailedException {
 521         // Direct modification commit is a utility, which cannot be used while we have transactions in-flight
 522         Preconditions.checkState(tip == dataTree, "Cannot modify data tree while transacgitons are pending");
 523
 524         modification.ready();
 525         dataTree.validate(modification);
 526         DataTreeCandidate candidate = dataTree.prepare(modification);
 527         dataTree.commit(candidate);
 528         return candidate;
 529     }
 530
 531     public Collection<ShardDataTreeCohort> getAndClearPendingTransactions() {
 532         Collection<ShardDataTreeCohort> ret = new ArrayList<>(pendingTransactions.size());
 533         for (CommitEntry entry: pendingTransactions) {
 534             ret.add(entry.cohort);
 535         }
 536
 537         pendingTransactions.clear();
 538         tip = dataTree;
 539         return ret;
 540     }
 541
 542     @SuppressWarnings("checkstyle:IllegalCatch")
 543     private void processNextTransaction() {
 544         while (!pendingTransactions.isEmpty()) {
 545             final CommitEntry entry = pendingTransactions.peek();
 546             final SimpleShardDataTreeCohort cohort = entry.cohort;
 547             final DataTreeModification modification = cohort.getDataTreeModification();
 548
 549             if (cohort.getState() != State.CAN_COMMIT_PENDING) {
 550                 break;
 551             }
 552
 553             LOG.debug("{}: Validating transaction {}", logContext, cohort.getIdentifier());
 554             Exception cause;
 555             try {
 556                 tip.validate(modification);
 557                 LOG.debug("{}: Transaction {} validated", logContext, cohort.getIdentifier());
 558                 cohort.successfulCanCommit();
 559                 entry.lastAccess = shard.ticker().read();
 560                 return;
 561             } catch (ConflictingModificationAppliedException e) {
 562                 LOG.warn("{}: Store Tx {}: Conflicting modification for path {}.", logContext, cohort.getIdentifier(),
 563                     e.getPath());
 564                 cause = new OptimisticLockFailedException("Optimistic lock failed.", e);
 565             } catch (DataValidationFailedException e) {
 566                 LOG.warn("{}: Store Tx {}: Data validation failed for path {}.", logContext, cohort.getIdentifier(),
 567                     e.getPath(), e);
 568
 569                 // For debugging purposes, allow dumping of the modification. Coupled with the above
 570                 // precondition log, it should allow us to understand what went on.
 571                 LOG.debug("{}: Store Tx {}: modifications: {} tree: {}", cohort.getIdentifier(), modification,
 572                         dataTree);
 573                 cause = new TransactionCommitFailedException("Data did not pass validation.", e);
 574             } catch (Exception e) {
 575                 LOG.warn("{}: Unexpected failure in validation phase", logContext, e);
 576                 cause = e;
 577             }
 578
 579             // Failure path: propagate the failure, remove the transaction from the queue and loop to the next one
 580             pendingTransactions.poll().cohort.failedCanCommit(cause);
 581         }
 582
 583         maybeRunOperationOnPendingTransactionsComplete();
 584     }
 585
 586     void startCanCommit(final SimpleShardDataTreeCohort cohort) {
 587         final SimpleShardDataTreeCohort current = pendingTransactions.peek().cohort;
 588         if (!cohort.equals(current)) {
 589             LOG.debug("{}: Transaction {} scheduled for canCommit step", logContext, cohort.getIdentifier());
 590             return;
 591         }
 592
 593         processNextTransaction();
 594     }
 595
 596     private void failPreCommit(final Exception cause) {
 597         shard.getShardMBean().incrementFailedTransactionsCount();
 598         pendingTransactions.poll().cohort.failedPreCommit(cause);
 599         processNextTransaction();
 600     }
 601
 602     @SuppressWarnings("checkstyle:IllegalCatch")
 603     void startPreCommit(final SimpleShardDataTreeCohort cohort) {
 604         final CommitEntry entry = pendingTransactions.peek();
 605         Preconditions.checkState(entry != null, "Attempted to pre-commit of %s when no transactions pending", cohort);
 606
 607         final SimpleShardDataTreeCohort current = entry.cohort;
 608         Verify.verify(cohort.equals(current), "Attempted to pre-commit %s while %s is pending", cohort, current);
 609         final DataTreeCandidateTip candidate;
 610         try {
 611             candidate = tip.prepare(cohort.getDataTreeModification());
 612         } catch (Exception e) {
 613             failPreCommit(e);
 614             return;
 615         }
 616
 617         try {
 618             cohort.userPreCommit(candidate);
 619         } catch (ExecutionException | TimeoutException e) {
 620             failPreCommit(e);
 621             return;
 622         }
 623
 624         // Set the tip of the data tree.
 625         tip = Verify.verifyNotNull(candidate);
 626
 627         entry.lastAccess = shard.ticker().read();
 628         cohort.successfulPreCommit(candidate);
 629     }
 630
 631     private void failCommit(final Exception cause) {
 632         shard.getShardMBean().incrementFailedTransactionsCount();
 633         pendingTransactions.poll().cohort.failedCommit(cause);
 634         processNextTransaction();
 635     }
 636
 637     @SuppressWarnings("checkstyle:IllegalCatch")
 638     private void finishCommit(final SimpleShardDataTreeCohort cohort) {
 639         final TransactionIdentifier txId = cohort.getIdentifier();
 640         final DataTreeCandidate candidate = cohort.getCandidate();
 641
 642         LOG.debug("{}: Resuming commit of transaction {}", logContext, txId);
 643
 644         try {
 645             dataTree.commit(candidate);
 646         } catch (Exception e) {
 647             LOG.error("{}: Failed to commit transaction {}", logContext, txId, e);
 648             failCommit(e);
 649             return;
 650         }
 651
 652         // All pending candidates have been committed, reset the tip to the data tree
 653         if (tip == candidate) {
 654             tip = dataTree;
 655         }
 656
 657         shard.getShardMBean().incrementCommittedTransactionCount();
 658         shard.getShardMBean().setLastCommittedTransactionTime(System.currentTimeMillis());
 659
 660         // FIXME: propagate journal index
 661         pendingTransactions.poll().cohort.successfulCommit(UnsignedLong.ZERO);
 662
 663         LOG.trace("{}: Transaction {} committed, proceeding to notify", logContext, txId);
 664         notifyListeners(candidate);
 665
 666         processNextTransaction();
 667     }
 668
 669     void startCommit(final SimpleShardDataTreeCohort cohort, final DataTreeCandidate candidate) {
 670         final CommitEntry entry = pendingTransactions.peek();
 671         Preconditions.checkState(entry != null, "Attempted to start commit of %s when no transactions pending", cohort);
 672
 673         final SimpleShardDataTreeCohort current = entry.cohort;
 674         Verify.verify(cohort.equals(current), "Attempted to commit %s while %s is pending", cohort, current);
 675
 676         if (shard.canSkipPayload() || candidate.getRootNode().getModificationType() == ModificationType.UNMODIFIED) {
 677             LOG.debug("{}: No replication required, proceeding to finish commit", logContext);
 678             finishCommit(cohort);
 679             return;
 680         }
 681
 682         final TransactionIdentifier txId = cohort.getIdentifier();
 683         final Payload payload;
 684         try {
 685             payload = CommitTransactionPayload.create(txId, candidate);
 686         } catch (IOException e) {
 687             LOG.error("{}: Failed to encode transaction {} candidate {}", logContext, txId, candidate, e);
 688             pendingTransactions.poll().cohort.failedCommit(e);
 689             return;
 690         }
 691
 692         // Once completed, we will continue via payloadReplicationComplete
 693         entry.lastAccess = shard.ticker().read();
 694         shard.persistPayload(txId, payload);
 695         LOG.debug("{}: Transaction {} submitted to persistence", logContext, txId);
 696     }
 697
 698     void processCohortRegistryCommand(final ActorRef sender, final CohortRegistryCommand message) {
 699         cohortRegistry.process(sender, message);
 700     }
 701
 702     @Override
 703     ShardDataTreeCohort createReadyCohort(final TransactionIdentifier txId,
 704             final DataTreeModification modification) {
 705         SimpleShardDataTreeCohort cohort = new SimpleShardDataTreeCohort(this, modification, txId,
 706                 cohortRegistry.createCohort(schemaContext, txId, COMMIT_STEP_TIMEOUT));
 707         pendingTransactions.add(new CommitEntry(cohort, shard.ticker().read()));
 708         return cohort;
 709     }
 710
 711     @SuppressFBWarnings(value = "DB_DUPLICATE_SWITCH_CLAUSES", justification = "See inline comments below.")
 712     void checkForExpiredTransactions(final long transactionCommitTimeoutMillis) {
 713         final long timeout = TimeUnit.MILLISECONDS.toNanos(transactionCommitTimeoutMillis);
 714         final long now = shard.ticker().read();
 715         final CommitEntry currentTx = pendingTransactions.peek();
 716         if (currentTx != null && currentTx.lastAccess + timeout < now) {
 717             LOG.warn("{}: Current transaction {} has timed out after {} ms in state {}", logContext,
 718                     currentTx.cohort.getIdentifier(), transactionCommitTimeoutMillis, currentTx.cohort.getState());
 719             boolean processNext = true;
 720             switch (currentTx.cohort.getState()) {
 721                 case CAN_COMMIT_PENDING:
 722                     pendingTransactions.remove().cohort.failedCanCommit(new TimeoutException());
 723                     break;
 724                 case CAN_COMMIT_COMPLETE:
 725                     // The suppression of the FindBugs "DB_DUPLICATE_SWITCH_CLAUSES" warning pertains to this clause
 726                     // whose code is duplicated with PRE_COMMIT_COMPLETE. The clauses aren't combined in case the code
 727                     // in PRE_COMMIT_COMPLETE is changed.
 728                     pendingTransactions.remove().cohort.reportFailure(new TimeoutException());
 729                     break;
 730                 case PRE_COMMIT_PENDING:
 731                     pendingTransactions.remove().cohort.failedPreCommit(new TimeoutException());
 732                     break;
 733                 case PRE_COMMIT_COMPLETE:
 734                     // FIXME: this is a legacy behavior problem. Three-phase commit protocol specifies that after we
 735                     //        are ready we should commit the transaction, not abort it. Our current software stack does
 736                     //        not allow us to do that consistently, because we persist at the time of commit, hence
 737                     //        we can end up in a state where we have pre-committed a transaction, then a leader failover
 738                     //        occurred ... the new leader does not see the pre-committed transaction and does not have
 739                     //        a running timer. To fix this we really need two persistence events.
 740                     //
 741                     //        The first one, done at pre-commit time will hold the transaction payload. When consensus
 742                     //        is reached, we exit the pre-commit phase and start the pre-commit timer. Followers do not
 743                     //        apply the state in this event.
 744                     //
 745                     //        The second one, done at commit (or abort) time holds only the transaction identifier and
 746                     //        signals to followers that the state should (or should not) be applied.
 747                     //
 748                     //        In order to make the pre-commit timer working across failovers, though, we need
 749                     //        a per-shard cluster-wide monotonic time, so a follower becoming the leader can accurately
 750                     //        restart the timer.
 751                     pendingTransactions.remove().cohort.reportFailure(new TimeoutException());
 752                     break;
 753                 case COMMIT_PENDING:
 754                     LOG.warn("{}: Transaction {} is still committing, cannot abort", logContext,
 755                         currentTx.cohort.getIdentifier());
 756                     currentTx.lastAccess = now;
 757                     processNext = false;
 758                     return;
 759                 case ABORTED:
 760                 case COMMITTED:
 761                 case FAILED:
 762                 case READY:
 763                 default:
 764                     pendingTransactions.remove();
 765             }
 766
 767             if (processNext) {
 768                 processNextTransaction();
 769             }
 770         }
 771     }
 772
 773     boolean startAbort(final SimpleShardDataTreeCohort cohort) {
 774         final Iterator<CommitEntry> it = pendingTransactions.iterator();
 775         if (!it.hasNext()) {
 776             LOG.debug("{}: no open transaction while attempting to abort {}", logContext, cohort.getIdentifier());
 777             return true;
 778         }
 779
 780         // First entry is special, as it may already be committing
 781         final CommitEntry first = it.next();
 782         if (cohort.equals(first.cohort)) {
 783             if (cohort.getState() != State.COMMIT_PENDING) {
 784                 LOG.debug("{}: aborting head of queue {} in state {}", logContext, cohort.getIdentifier(),
 785                     cohort.getIdentifier());
 786
 787                 it.remove();
 788                 rebasePreCommittedTransactions(it, dataTree);
 789                 processNextTransaction();
 790                 return true;
 791             }
 792
 793             LOG.warn("{}: transaction {} is committing, skipping abort", logContext, cohort.getIdentifier());
 794             return false;
 795         }
 796
 797         TipProducingDataTreeTip newTip = dataTree;
 798         while (it.hasNext()) {
 799             final CommitEntry e = it.next();
 800             if (cohort.equals(e.cohort)) {
 801                 LOG.debug("{}: aborting queued transaction {}", logContext, cohort.getIdentifier());
 802                 it.remove();
 803                 rebasePreCommittedTransactions(it, newTip);
 804                 return true;
 805             } else {
 806                 newTip = cohort.getCandidate();
 807             }
 808         }
 809
 810         LOG.debug("{}: aborted transaction {} not found in the queue", logContext, cohort.getIdentifier());
 811         return true;
 812     }
 813
 814     @SuppressWarnings("checkstyle:IllegalCatch")
 815     private void rebasePreCommittedTransactions(Iterator<CommitEntry> iter, TipProducingDataTreeTip newTip) {
 816         tip = newTip;
 817         while (iter.hasNext()) {
 818             final SimpleShardDataTreeCohort cohort = iter.next().cohort;
 819             if (cohort.getState() == State.CAN_COMMIT_COMPLETE) {
 820                 LOG.debug("{}: Revalidating queued transaction {}", logContext, cohort.getIdentifier());
 821
 822                 try {
 823                     tip.validate(cohort.getDataTreeModification());
 824                 } catch (DataValidationFailedException | RuntimeException e) {
 825                     LOG.debug("{}: Failed to revalidate queued transaction {}", logContext, cohort.getIdentifier(), e);
 826                     cohort.reportFailure(e);
 827                 }
 828             } else if (cohort.getState() == State.PRE_COMMIT_COMPLETE) {
 829                 LOG.debug("{}: Repreparing queued transaction {}", logContext, cohort.getIdentifier());
 830
 831                 try {
 832                     tip.validate(cohort.getDataTreeModification());
 833                     DataTreeCandidateTip candidate = tip.prepare(cohort.getDataTreeModification());
 834                     cohort.userPreCommit(candidate);
 835
 836                     cohort.setNewCandidate(candidate);
 837                     tip = candidate;
 838                 } catch (ExecutionException | TimeoutException | RuntimeException | DataValidationFailedException e) {
 839                     LOG.debug("{}: Failed to reprepare queued transaction {}", logContext, cohort.getIdentifier(), e);
 840                     cohort.reportFailure(e);
 841                 }
 842             }
 843         }
 844     }
 845
 846     void setRunOnPendingTransactionsComplete(final Runnable operation) {
 847         runOnPendingTransactionsComplete = operation;
 848         maybeRunOperationOnPendingTransactionsComplete();
 849     }
 850
 851     private void maybeRunOperationOnPendingTransactionsComplete() {
 852         if (runOnPendingTransactionsComplete != null && pendingTransactions.isEmpty()) {
 853             LOG.debug("{}: Pending transactions complete - running operation {}", logContext,
 854                     runOnPendingTransactionsComplete);
 855
 856             runOnPendingTransactionsComplete.run();
 857             runOnPendingTransactionsComplete = null;
 858         }
 859     }
 860 }