2 * Copyright (c) 2014 Brocade Communications Systems, Inc. and others. All rights reserved.
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
8 package org.opendaylight.controller.cluster.datastore;
10 import akka.actor.ActorRef;
11 import akka.actor.Status.Failure;
12 import akka.serialization.Serialization;
13 import com.google.common.annotations.VisibleForTesting;
14 import com.google.common.base.Preconditions;
15 import java.util.ArrayList;
16 import java.util.Collection;
17 import java.util.Collections;
18 import java.util.HashMap;
19 import java.util.Iterator;
20 import java.util.LinkedList;
21 import java.util.List;
23 import java.util.Queue;
24 import java.util.concurrent.ExecutionException;
25 import org.opendaylight.controller.cluster.datastore.DataTreeCohortActorRegistry.CohortRegistryCommand;
26 import org.opendaylight.controller.cluster.datastore.messages.AbortTransactionReply;
27 import org.opendaylight.controller.cluster.datastore.messages.BatchedModifications;
28 import org.opendaylight.controller.cluster.datastore.messages.BatchedModificationsReply;
29 import org.opendaylight.controller.cluster.datastore.messages.CanCommitTransaction;
30 import org.opendaylight.controller.cluster.datastore.messages.CanCommitTransactionReply;
31 import org.opendaylight.controller.cluster.datastore.messages.CommitTransaction;
32 import org.opendaylight.controller.cluster.datastore.messages.ForwardedReadyTransaction;
33 import org.opendaylight.controller.cluster.datastore.messages.ReadyLocalTransaction;
34 import org.opendaylight.controller.cluster.datastore.messages.ReadyTransactionReply;
35 import org.opendaylight.controller.cluster.datastore.utils.AbstractBatchedModificationsCursor;
36 import org.opendaylight.controller.md.sal.common.api.data.TransactionCommitFailedException;
37 import org.opendaylight.yangtools.concepts.Identifier;
38 import org.opendaylight.yangtools.yang.model.api.SchemaContext;
39 import org.slf4j.Logger;
42 * Coordinates commits for a shard ensuring only one concurrent 3-phase commit.
44 * @author Thomas Pantelis
46 final class ShardCommitCoordinator {
48 // Interface hook for unit tests to replace or decorate the DOMStoreThreePhaseCommitCohorts.
49 public interface CohortDecorator {
50 ShardDataTreeCohort decorate(Identifier transactionID, ShardDataTreeCohort actual);
53 private final Map<Identifier, CohortEntry> cohortCache = new HashMap<>();
55 private CohortEntry currentCohortEntry;
57 private final ShardDataTree dataTree;
59 private final DataTreeCohortActorRegistry cohortRegistry = new DataTreeCohortActorRegistry();
61 // We use a LinkedList here to avoid synchronization overhead with concurrent queue impls
62 // since this should only be accessed on the shard's dispatcher.
63 private final Queue<CohortEntry> queuedCohortEntries = new LinkedList<>();
65 private int queueCapacity;
67 private final Logger log;
69 private final String name;
71 private final long cacheExpiryTimeoutInMillis;
73 // This is a hook for unit tests to replace or decorate the DOMStoreThreePhaseCommitCohorts.
74 private CohortDecorator cohortDecorator;
76 private ReadyTransactionReply readyTransactionReply;
78 private Runnable runOnPendingTransactionsComplete;
80 ShardCommitCoordinator(ShardDataTree dataTree, long cacheExpiryTimeoutInMillis, int queueCapacity, Logger log,
83 this.queueCapacity = queueCapacity;
86 this.dataTree = Preconditions.checkNotNull(dataTree);
87 this.cacheExpiryTimeoutInMillis = cacheExpiryTimeoutInMillis;
91 return queuedCohortEntries.size();
94 int getCohortCacheSize() {
95 return cohortCache.size();
98 void setQueueCapacity(int queueCapacity) {
99 this.queueCapacity = queueCapacity;
102 private ReadyTransactionReply readyTransactionReply(Shard shard) {
103 if(readyTransactionReply == null) {
104 readyTransactionReply = new ReadyTransactionReply(Serialization.serializedActorPath(shard.self()));
107 return readyTransactionReply;
110 private boolean queueCohortEntry(CohortEntry cohortEntry, ActorRef sender, Shard shard) {
111 if(queuedCohortEntries.size() < queueCapacity) {
112 queuedCohortEntries.offer(cohortEntry);
114 log.debug("{}: Enqueued transaction {}, queue size {}", name, cohortEntry.getTransactionID(),
115 queuedCohortEntries.size());
119 cohortCache.remove(cohortEntry.getTransactionID());
121 final RuntimeException ex = new RuntimeException(
122 String.format("%s: Could not enqueue transaction %s - the maximum commit queue"+
123 " capacity %d has been reached.",
124 name, cohortEntry.getTransactionID(), queueCapacity));
125 log.error(ex.getMessage());
126 sender.tell(new Failure(ex), shard.self());
132 * This method is called to ready a transaction that was prepared by ShardTransaction actor. It caches
133 * the prepared cohort entry for the given transactions ID in preparation for the subsequent 3-phase commit.
135 * @param ready the ForwardedReadyTransaction message to process
136 * @param sender the sender of the message
137 * @param shard the transaction's shard actor
140 void handleForwardedReadyTransaction(ForwardedReadyTransaction ready, ActorRef sender, Shard shard,
141 SchemaContext schema) {
142 log.debug("{}: Readying transaction {}, client version {}", name,
143 ready.getTransactionID(), ready.getTxnClientVersion());
145 final ShardDataTreeCohort cohort = ready.getTransaction().ready();
146 final CohortEntry cohortEntry = new CohortEntry(ready.getTransactionID(), cohort, cohortRegistry, schema, ready.getTxnClientVersion());
147 cohortCache.put(cohortEntry.getTransactionID(), cohortEntry);
149 if(!queueCohortEntry(cohortEntry, sender, shard)) {
153 if(ready.isDoImmediateCommit()) {
154 cohortEntry.setDoImmediateCommit(true);
155 cohortEntry.setReplySender(sender);
156 cohortEntry.setShard(shard);
157 handleCanCommit(cohortEntry);
159 // The caller does not want immediate commit - the 3-phase commit will be coordinated by the
160 // front-end so send back a ReadyTransactionReply with our actor path.
161 sender.tell(readyTransactionReply(shard), shard.self());
166 * This method handles a BatchedModifications message for a transaction being prepared directly on the
167 * Shard actor instead of via a ShardTransaction actor. If there's no currently cached
168 * DOMStoreWriteTransaction, one is created. The batched modifications are applied to the write Tx. If
169 * the BatchedModifications is ready to commit then a DOMStoreThreePhaseCommitCohort is created.
171 * @param batched the BatchedModifications message to process
172 * @param sender the sender of the message
173 * @param shard the transaction's shard actor
175 void handleBatchedModifications(BatchedModifications batched, ActorRef sender, Shard shard, SchemaContext schema) {
176 CohortEntry cohortEntry = cohortCache.get(batched.getTransactionID());
177 if(cohortEntry == null) {
178 cohortEntry = new CohortEntry(batched.getTransactionID(),
179 dataTree.newReadWriteTransaction(batched.getTransactionID()),
180 cohortRegistry, schema, batched.getVersion());
181 cohortCache.put(cohortEntry.getTransactionID(), cohortEntry);
184 if(log.isDebugEnabled()) {
185 log.debug("{}: Applying {} batched modifications for Tx {}", name,
186 batched.getModifications().size(), batched.getTransactionID());
189 cohortEntry.applyModifications(batched.getModifications());
191 if(batched.isReady()) {
192 if(cohortEntry.getLastBatchedModificationsException() != null) {
193 cohortCache.remove(cohortEntry.getTransactionID());
194 throw cohortEntry.getLastBatchedModificationsException();
197 if(cohortEntry.getTotalBatchedModificationsReceived() != batched.getTotalMessagesSent()) {
198 cohortCache.remove(cohortEntry.getTransactionID());
199 throw new IllegalStateException(String.format(
200 "The total number of batched messages received %d does not match the number sent %d",
201 cohortEntry.getTotalBatchedModificationsReceived(), batched.getTotalMessagesSent()));
204 if(!queueCohortEntry(cohortEntry, sender, shard)) {
208 if(log.isDebugEnabled()) {
209 log.debug("{}: Readying Tx {}, client version {}", name,
210 batched.getTransactionID(), batched.getVersion());
213 cohortEntry.ready(cohortDecorator, batched.isDoCommitOnReady());
215 if(batched.isDoCommitOnReady()) {
216 cohortEntry.setReplySender(sender);
217 cohortEntry.setShard(shard);
218 handleCanCommit(cohortEntry);
220 sender.tell(readyTransactionReply(shard), shard.self());
223 sender.tell(new BatchedModificationsReply(batched.getModifications().size()), shard.self());
228 * This method handles {@link ReadyLocalTransaction} message. All transaction modifications have
229 * been prepared beforehand by the sender and we just need to drive them through into the
232 * @param message the ReadyLocalTransaction message to process
233 * @param sender the sender of the message
234 * @param shard the transaction's shard actor
236 void handleReadyLocalTransaction(ReadyLocalTransaction message, ActorRef sender, Shard shard,
237 SchemaContext schema) {
238 final ShardDataTreeCohort cohort = new SimpleShardDataTreeCohort(dataTree, message.getModification(),
239 message.getTransactionID());
240 final CohortEntry cohortEntry = new CohortEntry(message.getTransactionID(), cohort, cohortRegistry, schema,
241 DataStoreVersions.CURRENT_VERSION);
242 cohortCache.put(cohortEntry.getTransactionID(), cohortEntry);
243 cohortEntry.setDoImmediateCommit(message.isDoCommitOnReady());
245 if(!queueCohortEntry(cohortEntry, sender, shard)) {
249 log.debug("{}: Applying local modifications for Tx {}", name, message.getTransactionID());
251 if (message.isDoCommitOnReady()) {
252 cohortEntry.setReplySender(sender);
253 cohortEntry.setShard(shard);
254 handleCanCommit(cohortEntry);
256 sender.tell(readyTransactionReply(shard), shard.self());
260 Collection<BatchedModifications> createForwardedBatchedModifications(final BatchedModifications from,
261 final int maxModificationsPerBatch) {
262 CohortEntry cohortEntry = getAndRemoveCohortEntry(from.getTransactionID());
263 if(cohortEntry == null || cohortEntry.getTransaction() == null) {
264 return Collections.singletonList(from);
267 cohortEntry.applyModifications(from.getModifications());
269 final LinkedList<BatchedModifications> newModifications = new LinkedList<>();
270 cohortEntry.getTransaction().getSnapshot().applyToCursor(new AbstractBatchedModificationsCursor() {
272 protected BatchedModifications getModifications() {
273 if(newModifications.isEmpty() ||
274 newModifications.getLast().getModifications().size() >= maxModificationsPerBatch) {
275 newModifications.add(new BatchedModifications(from.getTransactionID(), from.getVersion()));
278 return newModifications.getLast();
282 BatchedModifications last = newModifications.getLast();
283 last.setDoCommitOnReady(from.isDoCommitOnReady());
284 last.setReady(from.isReady());
285 last.setTotalMessagesSent(newModifications.size());
286 return newModifications;
289 private void handleCanCommit(CohortEntry cohortEntry) {
290 cohortEntry.updateLastAccessTime();
292 if(currentCohortEntry != null) {
293 // There's already a Tx commit in progress so we can't process this entry yet - but it's in the
294 // queue and will get processed after all prior entries complete.
296 if(log.isDebugEnabled()) {
297 log.debug("{}: Commit for Tx {} already in progress - skipping canCommit for {} for now",
298 name, currentCohortEntry.getTransactionID(), cohortEntry.getTransactionID());
304 // No Tx commit currently in progress - check if this entry is the next one in the queue, If so make
305 // it the current entry and proceed with canCommit.
306 // Purposely checking reference equality here.
307 if(queuedCohortEntries.peek() == cohortEntry) {
308 currentCohortEntry = queuedCohortEntries.poll();
309 doCanCommit(currentCohortEntry);
311 if(log.isDebugEnabled()) {
312 log.debug("{}: Tx {} is the next pending canCommit - skipping {} for now", name,
313 queuedCohortEntries.peek() != null ? queuedCohortEntries.peek().getTransactionID() : "???",
314 cohortEntry.getTransactionID());
320 * This method handles the canCommit phase for a transaction.
322 * @param transactionID the ID of the transaction to canCommit
323 * @param sender the actor to which to send the response
324 * @param shard the transaction's shard actor
326 void handleCanCommit(Identifier transactionID, final ActorRef sender, final Shard shard) {
327 // Lookup the cohort entry that was cached previously (or should have been) by
328 // transactionReady (via the ForwardedReadyTransaction message).
329 final CohortEntry cohortEntry = cohortCache.get(transactionID);
330 if(cohortEntry == null) {
331 // Either canCommit was invoked before ready(shouldn't happen) or a long time passed
332 // between canCommit and ready and the entry was expired from the cache.
333 IllegalStateException ex = new IllegalStateException(
334 String.format("%s: No cohort entry found for transaction %s", name, transactionID));
335 log.error(ex.getMessage());
336 sender.tell(new Failure(ex), shard.self());
340 cohortEntry.setReplySender(sender);
341 cohortEntry.setShard(shard);
343 handleCanCommit(cohortEntry);
346 private void doCanCommit(final CohortEntry cohortEntry) {
347 boolean canCommit = false;
349 canCommit = cohortEntry.canCommit();
351 log.debug("{}: canCommit for {}: {}", name, cohortEntry.getTransactionID(), canCommit);
353 if(cohortEntry.isDoImmediateCommit()) {
355 doCommit(cohortEntry);
357 cohortEntry.getReplySender().tell(new Failure(new TransactionCommitFailedException(
358 "Can Commit failed, no detailed cause available.")), cohortEntry.getShard().self());
361 cohortEntry.getReplySender().tell(
362 canCommit ? CanCommitTransactionReply.yes(cohortEntry.getClientVersion()).toSerializable() :
363 CanCommitTransactionReply.no(cohortEntry.getClientVersion()).toSerializable(),
364 cohortEntry.getShard().self());
366 } catch (Exception e) {
367 log.debug("{}: An exception occurred during canCommit", name, e);
369 Throwable failure = e;
370 if(e instanceof ExecutionException) {
371 failure = e.getCause();
374 cohortEntry.getReplySender().tell(new Failure(failure), cohortEntry.getShard().self());
377 // Remove the entry from the cache now.
378 currentTransactionComplete(cohortEntry.getTransactionID(), true);
383 private boolean doCommit(CohortEntry cohortEntry) {
384 log.debug("{}: Committing transaction {}", name, cohortEntry.getTransactionID());
386 boolean success = false;
388 // We perform the preCommit phase here atomically with the commit phase. This is an
389 // optimization to eliminate the overhead of an extra preCommit message. We lose front-end
390 // coordination of preCommit across shards in case of failure but preCommit should not
391 // normally fail since we ensure only one concurrent 3-phase commit.
394 cohortEntry.preCommit();
396 cohortEntry.getShard().continueCommit(cohortEntry);
398 cohortEntry.updateLastAccessTime();
401 } catch (Exception e) {
402 log.error("{} An exception occurred while preCommitting transaction {}",
403 name, cohortEntry.getTransactionID(), e);
404 cohortEntry.getReplySender().tell(new Failure(e), cohortEntry.getShard().self());
406 currentTransactionComplete(cohortEntry.getTransactionID(), true);
413 * This method handles the preCommit and commit phases for a transaction.
415 * @param transactionID the ID of the transaction to commit
416 * @param sender the actor to which to send the response
417 * @param shard the transaction's shard actor
418 * @return true if the transaction was successfully prepared, false otherwise.
420 boolean handleCommit(final Identifier transactionID, final ActorRef sender, final Shard shard) {
421 // Get the current in-progress cohort entry in the commitCoordinator if it corresponds to
423 final CohortEntry cohortEntry = getCohortEntryIfCurrent(transactionID);
424 if(cohortEntry == null) {
425 // We're not the current Tx - the Tx was likely expired b/c it took too long in
426 // between the canCommit and commit messages.
427 IllegalStateException ex = new IllegalStateException(
428 String.format("%s: Cannot commit transaction %s - it is not the current transaction",
429 name, transactionID));
430 log.error(ex.getMessage());
431 sender.tell(new Failure(ex), shard.self());
435 cohortEntry.setReplySender(sender);
436 return doCommit(cohortEntry);
439 void handleAbort(final Identifier transactionID, final ActorRef sender, final Shard shard) {
440 CohortEntry cohortEntry = getCohortEntryIfCurrent(transactionID);
441 if(cohortEntry != null) {
442 // We don't remove the cached cohort entry here (ie pass false) in case the Tx was
443 // aborted during replication in which case we may still commit locally if replication
445 currentTransactionComplete(transactionID, false);
447 cohortEntry = getAndRemoveCohortEntry(transactionID);
450 if(cohortEntry == null) {
454 log.debug("{}: Aborting transaction {}", name, transactionID);
456 final ActorRef self = shard.getSelf();
460 shard.getShardMBean().incrementAbortTransactionsCount();
463 sender.tell(AbortTransactionReply.instance(cohortEntry.getClientVersion()).toSerializable(), self);
465 } catch (Exception e) {
466 log.error("{}: An exception happened during abort", name, e);
469 sender.tell(new Failure(e), self);
474 void checkForExpiredTransactions(final long timeout, final Shard shard) {
475 CohortEntry cohortEntry = getCurrentCohortEntry();
476 if(cohortEntry != null) {
477 if(cohortEntry.isExpired(timeout)) {
478 log.warn("{}: Current transaction {} has timed out after {} ms - aborting",
479 name, cohortEntry.getTransactionID(), timeout);
481 handleAbort(cohortEntry.getTransactionID(), null, shard);
485 cleanupExpiredCohortEntries();
488 void abortPendingTransactions(final String reason, final Shard shard) {
489 if(currentCohortEntry == null && queuedCohortEntries.isEmpty()) {
493 List<CohortEntry> cohortEntries = getAndClearPendingCohortEntries();
495 log.debug("{}: Aborting {} pending queued transactions", name, cohortEntries.size());
497 for(CohortEntry cohortEntry: cohortEntries) {
498 if(cohortEntry.getReplySender() != null) {
499 cohortEntry.getReplySender().tell(new Failure(new RuntimeException(reason)), shard.self());
504 private List<CohortEntry> getAndClearPendingCohortEntries() {
505 List<CohortEntry> cohortEntries = new ArrayList<>();
507 if(currentCohortEntry != null) {
508 cohortEntries.add(currentCohortEntry);
509 cohortCache.remove(currentCohortEntry.getTransactionID());
510 currentCohortEntry = null;
513 for(CohortEntry cohortEntry: queuedCohortEntries) {
514 cohortEntries.add(cohortEntry);
515 cohortCache.remove(cohortEntry.getTransactionID());
518 queuedCohortEntries.clear();
519 return cohortEntries;
522 Collection<Object> convertPendingTransactionsToMessages(final int maxModificationsPerBatch) {
523 if(currentCohortEntry == null && queuedCohortEntries.isEmpty()) {
524 return Collections.emptyList();
527 Collection<Object> messages = new ArrayList<>();
528 List<CohortEntry> cohortEntries = getAndClearPendingCohortEntries();
529 for(CohortEntry cohortEntry: cohortEntries) {
530 if(cohortEntry.isExpired(cacheExpiryTimeoutInMillis) || cohortEntry.isAborted()) {
534 final LinkedList<BatchedModifications> newModifications = new LinkedList<>();
535 cohortEntry.getDataTreeModification().applyToCursor(new AbstractBatchedModificationsCursor() {
537 protected BatchedModifications getModifications() {
538 if(newModifications.isEmpty() ||
539 newModifications.getLast().getModifications().size() >= maxModificationsPerBatch) {
540 newModifications.add(new BatchedModifications(cohortEntry.getTransactionID(),
541 cohortEntry.getClientVersion()));
544 return newModifications.getLast();
548 if(!newModifications.isEmpty()) {
549 BatchedModifications last = newModifications.getLast();
550 last.setDoCommitOnReady(cohortEntry.isDoImmediateCommit());
552 last.setTotalMessagesSent(newModifications.size());
553 messages.addAll(newModifications);
555 if(!cohortEntry.isDoImmediateCommit() && cohortEntry.getState() == CohortEntry.State.CAN_COMMITTED) {
556 messages.add(new CanCommitTransaction(cohortEntry.getTransactionID(),
557 cohortEntry.getClientVersion()));
560 if(!cohortEntry.isDoImmediateCommit() && cohortEntry.getState() == CohortEntry.State.PRE_COMMITTED) {
561 messages.add(new CommitTransaction(cohortEntry.getTransactionID(),
562 cohortEntry.getClientVersion()));
571 * Returns the cohort entry for the Tx commit currently in progress if the given transaction ID
572 * matches the current entry.
574 * @param transactionID the ID of the transaction
575 * @return the current CohortEntry or null if the given transaction ID does not match the
578 CohortEntry getCohortEntryIfCurrent(Identifier transactionID) {
579 if(isCurrentTransaction(transactionID)) {
580 return currentCohortEntry;
586 CohortEntry getCurrentCohortEntry() {
587 return currentCohortEntry;
590 CohortEntry getAndRemoveCohortEntry(Identifier transactionID) {
591 return cohortCache.remove(transactionID);
594 boolean isCurrentTransaction(Identifier transactionID) {
595 return currentCohortEntry != null &&
596 currentCohortEntry.getTransactionID().equals(transactionID);
600 * This method is called when a transaction is complete, successful or not. If the given
601 * given transaction ID matches the current in-progress transaction, the next cohort entry,
602 * if any, is dequeued and processed.
604 * @param transactionID the ID of the completed transaction
605 * @param removeCohortEntry if true the CohortEntry for the transaction is also removed from
608 void currentTransactionComplete(Identifier transactionID, boolean removeCohortEntry) {
609 if(removeCohortEntry) {
610 cohortCache.remove(transactionID);
613 if(isCurrentTransaction(transactionID)) {
614 currentCohortEntry = null;
616 log.debug("{}: currentTransactionComplete: {}", name, transactionID);
618 maybeProcessNextCohortEntry();
622 private void maybeProcessNextCohortEntry() {
623 // Check if there's a next cohort entry waiting in the queue and if it is ready to commit. Also
624 // clean out expired entries.
625 final Iterator<CohortEntry> iter = queuedCohortEntries.iterator();
626 while(iter.hasNext()) {
627 final CohortEntry next = iter.next();
628 if(next.isReadyToCommit()) {
629 if(currentCohortEntry == null) {
630 if(log.isDebugEnabled()) {
631 log.debug("{}: Next entry to canCommit {}", name, next);
635 currentCohortEntry = next;
636 currentCohortEntry.updateLastAccessTime();
637 doCanCommit(currentCohortEntry);
641 } else if(next.isExpired(cacheExpiryTimeoutInMillis)) {
642 log.warn("{}: canCommit for transaction {} was not received within {} ms - entry removed from cache",
643 name, next.getTransactionID(), cacheExpiryTimeoutInMillis);
644 } else if(!next.isAborted()) {
649 cohortCache.remove(next.getTransactionID());
652 maybeRunOperationOnPendingTransactionsComplete();
655 void cleanupExpiredCohortEntries() {
656 maybeProcessNextCohortEntry();
659 void setRunOnPendingTransactionsComplete(Runnable operation) {
660 runOnPendingTransactionsComplete = operation;
661 maybeRunOperationOnPendingTransactionsComplete();
664 private void maybeRunOperationOnPendingTransactionsComplete() {
665 if(runOnPendingTransactionsComplete != null && currentCohortEntry == null && queuedCohortEntries.isEmpty()) {
666 log.debug("{}: Pending transactions complete - running operation {}", name, runOnPendingTransactionsComplete);
668 runOnPendingTransactionsComplete.run();
669 runOnPendingTransactionsComplete = null;
674 void setCohortDecorator(CohortDecorator cohortDecorator) {
675 this.cohortDecorator = cohortDecorator;
678 void processCohortRegistryCommand(ActorRef sender, CohortRegistryCommand message) {
679 cohortRegistry.process(sender, message);