2 * Copyright (c) 2016 Cisco 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.databroker.actors.dds;
10 import akka.actor.ActorRef;
11 import akka.actor.Status;
12 import com.google.common.base.Throwables;
13 import com.google.common.base.Verify;
14 import java.util.ArrayList;
15 import java.util.Collection;
17 import java.util.concurrent.ConcurrentHashMap;
18 import java.util.concurrent.atomic.AtomicLong;
19 import java.util.concurrent.locks.StampedLock;
20 import org.opendaylight.controller.cluster.access.client.BackendInfoResolver;
21 import org.opendaylight.controller.cluster.access.client.ClientActorBehavior;
22 import org.opendaylight.controller.cluster.access.client.ClientActorContext;
23 import org.opendaylight.controller.cluster.access.client.ConnectedClientConnection;
24 import org.opendaylight.controller.cluster.access.concepts.LocalHistoryIdentifier;
25 import org.opendaylight.yangtools.yang.data.api.YangInstanceIdentifier;
26 import org.slf4j.Logger;
27 import org.slf4j.LoggerFactory;
30 * {@link ClientActorBehavior} acting as an intermediary between the backend actors and the DistributedDataStore
34 * This class is not visible outside of this package because it breaks the actor containment. Services provided to
35 * Java world outside of actor containment are captured in {@link DataStoreClient}.
38 * IMPORTANT: this class breaks actor containment via methods implementing {@link DataStoreClient} contract.
39 * When touching internal state, be mindful of the execution context from which execution context, Actor
40 * or POJO, is the state being accessed or modified.
43 * THREAD SAFETY: this class must always be kept thread-safe, so that both the Actor System thread and the application
44 * threads can run concurrently. All state transitions must be made in a thread-safe manner. When in
45 * doubt, feel free to synchronize on this object.
48 * PERFORMANCE: this class lies in a performance-critical fast path. All code needs to be concise and efficient, but
49 * performance must not come at the price of correctness. Any optimizations need to be carefully analyzed
50 * for correctness and performance impact.
53 * TRADE-OFFS: part of the functionality runs in application threads without switching contexts, which makes it ideal
54 * for performing work and charging applications for it. That has two positive effects:
55 * - CPU usage is distributed across applications, minimizing work done in the actor thread
56 * - CPU usage provides back-pressure towards the application.
58 * @author Robert Varga
60 abstract class AbstractDataStoreClientBehavior extends ClientActorBehavior<ShardBackendInfo>
61 implements DataStoreClient {
62 private static final Logger LOG = LoggerFactory.getLogger(AbstractDataStoreClientBehavior.class);
64 private final Map<LocalHistoryIdentifier, ClientLocalHistory> histories = new ConcurrentHashMap<>();
65 private final AtomicLong nextHistoryId = new AtomicLong(1);
66 private final StampedLock lock = new StampedLock();
67 private final SingleClientHistory singleHistory;
69 private volatile Throwable aborted;
71 AbstractDataStoreClientBehavior(final ClientActorContext context,
72 final BackendInfoResolver<ShardBackendInfo> resolver) {
73 super(context, resolver);
74 singleHistory = new SingleClientHistory(this, new LocalHistoryIdentifier(getIdentifier(), 0));
79 // Methods below are invoked from the client actor thread
84 protected final void haltClient(final Throwable cause) {
85 // If we have encountered a previous problem there is no cleanup necessary, as we have already cleaned up
86 // Thread safely is not an issue, as both this method and any failures are executed from the same (client actor)
88 if (aborted != null) {
89 abortOperations(cause);
93 private void abortOperations(final Throwable cause) {
94 final long stamp = lock.writeLock();
96 // This acts as a barrier, application threads check this after they have added an entry in the maps,
97 // and if they observe aborted being non-null, they will perform their cleanup and not return the handle.
100 for (ClientLocalHistory h : histories.values()) {
105 lock.unlockWrite(stamp);
109 private AbstractDataStoreClientBehavior shutdown(final ClientActorBehavior<ShardBackendInfo> currentBehavior) {
110 abortOperations(new IllegalStateException("Client " + getIdentifier() + " has been shut down"));
115 protected final AbstractDataStoreClientBehavior onCommand(final Object command) {
116 if (command instanceof GetClientRequest) {
117 ((GetClientRequest) command).getReplyTo().tell(new Status.Success(this), ActorRef.noSender());
119 LOG.warn("{}: ignoring unhandled command {}", persistenceId(), command);
126 * The connection has resolved, which means we have to potentially perform message adaptation. This is a bit more
127 * involved, as the messages need to be replayed to the individual proxies.
130 protected final ConnectionConnectCohort connectionUp(final ConnectedClientConnection<ShardBackendInfo> newConn) {
131 final long stamp = lock.writeLock();
133 // Step 1: Freeze all AbstractProxyHistory instances pointing to that shard. This indirectly means that no
134 // further TransactionProxies can be created and we can safely traverse maps without risking
136 final Collection<HistoryReconnectCohort> cohorts = new ArrayList<>();
137 startReconnect(singleHistory, newConn, cohorts);
138 for (ClientLocalHistory h : histories.values()) {
139 startReconnect(h, newConn, cohorts);
142 return previousEntries -> {
144 // Step 2: Collect previous successful requests from the cohorts. We do not want to expose
145 // the non-throttling interface to the connection, hence we use a wrapper consumer
146 for (HistoryReconnectCohort c : cohorts) {
147 c.replayRequests(previousEntries);
150 // Step 3: Install a forwarder, which will forward requests back to affected cohorts. Any outstanding
151 // requests will be immediately sent to it and requests being sent concurrently will get
152 // forwarded once they hit the new connection.
153 return BouncingReconnectForwarder.forCohorts(newConn, cohorts);
156 // Step 4: Complete switchover of the connection. The cohorts can resume normal operations.
157 for (HistoryReconnectCohort c : cohorts) {
161 lock.unlockWrite(stamp);
167 private static void startReconnect(final AbstractClientHistory history,
168 final ConnectedClientConnection<ShardBackendInfo> newConn,
169 final Collection<HistoryReconnectCohort> cohorts) {
170 final HistoryReconnectCohort cohort = history.startReconnect(newConn);
171 if (cohort != null) {
178 // Methods below are invoked from application threads
183 public final ClientLocalHistory createLocalHistory() {
184 final LocalHistoryIdentifier historyId = new LocalHistoryIdentifier(getIdentifier(),
185 nextHistoryId.getAndIncrement());
187 final long stamp = lock.readLock();
189 if (aborted != null) {
190 throw Throwables.propagate(aborted);
193 final ClientLocalHistory history = new ClientLocalHistory(this, historyId);
194 LOG.debug("{}: creating a new local history {}", persistenceId(), history);
196 Verify.verify(histories.put(historyId, history) == null);
199 lock.unlockRead(stamp);
204 public final ClientTransaction createTransaction() {
205 return singleHistory.createTransaction();
209 public final ClientSnapshot createSnapshot() {
210 return singleHistory.takeSnapshot();
214 public final void close() {
215 context().executeInActor(this::shutdown);
218 abstract Long resolveShardForPath(YangInstanceIdentifier path);