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.mdsal.singleton.dom.impl;
10 import static com.google.common.base.Preconditions.checkArgument;
11 import static com.google.common.base.Preconditions.checkState;
12 import static com.google.common.base.Verify.verify;
13 import static com.google.common.base.Verify.verifyNotNull;
14 import static java.util.Objects.requireNonNull;
16 import com.google.common.annotations.VisibleForTesting;
17 import com.google.common.base.MoreObjects;
18 import com.google.common.collect.ImmutableList;
19 import com.google.common.collect.ImmutableSet;
20 import com.google.common.util.concurrent.FutureCallback;
21 import com.google.common.util.concurrent.Futures;
22 import com.google.common.util.concurrent.ListenableFuture;
23 import com.google.common.util.concurrent.MoreExecutors;
24 import com.google.common.util.concurrent.SettableFuture;
25 import java.util.Collection;
26 import java.util.HashMap;
27 import java.util.Iterator;
29 import java.util.Map.Entry;
31 import java.util.concurrent.ConcurrentHashMap;
32 import java.util.concurrent.atomic.AtomicIntegerFieldUpdater;
33 import java.util.concurrent.atomic.AtomicReference;
34 import javax.annotation.concurrent.GuardedBy;
35 import javax.annotation.concurrent.ThreadSafe;
36 import org.eclipse.jdt.annotation.NonNull;
37 import org.opendaylight.mdsal.eos.common.api.CandidateAlreadyRegisteredException;
38 import org.opendaylight.mdsal.eos.common.api.EntityOwnershipChangeState;
39 import org.opendaylight.mdsal.eos.common.api.GenericEntity;
40 import org.opendaylight.mdsal.eos.common.api.GenericEntityOwnershipCandidateRegistration;
41 import org.opendaylight.mdsal.eos.common.api.GenericEntityOwnershipChange;
42 import org.opendaylight.mdsal.eos.common.api.GenericEntityOwnershipListener;
43 import org.opendaylight.mdsal.eos.common.api.GenericEntityOwnershipService;
44 import org.opendaylight.mdsal.singleton.common.api.ClusterSingletonService;
45 import org.opendaylight.mdsal.singleton.common.api.ClusterSingletonServiceRegistration;
46 import org.opendaylight.yangtools.concepts.Path;
47 import org.slf4j.Logger;
48 import org.slf4j.LoggerFactory;
51 * Implementation of {@link ClusterSingletonServiceGroup} on top of the Entity Ownership Service. Since EOS is atomic
52 * in its operation and singleton services incur startup and most notably cleanup, we need to do something smart here.
55 * The implementation takes advantage of the fact that EOS provides stable ownership, i.e. owners are not moved as
56 * a result on new candidates appearing. We use two entities:
57 * - service entity, to which all nodes register
58 * - cleanup entity, which only the service entity owner registers to
61 * Once the cleanup entity ownership is acquired, services are started. As long as the cleanup entity is registered,
62 * it should remain the owner. In case a new service owner emerges, the old owner will start the cleanup process,
63 * eventually releasing the cleanup entity. The new owner registers for the cleanup entity -- but will not see it
64 * granted until the old owner finishes the cleanup.
66 * @param <P> the instance identifier path type
67 * @param <E> the GenericEntity type
68 * @param <C> the GenericEntityOwnershipChange type
69 * @param <G> the GenericEntityOwnershipListener type
70 * @param <S> the GenericEntityOwnershipService type
73 final class ClusterSingletonServiceGroupImpl<P extends Path<P>, E extends GenericEntity<P>,
74 C extends GenericEntityOwnershipChange<P, E>, G extends GenericEntityOwnershipListener<P, C>,
75 S extends GenericEntityOwnershipService<P, E, G>> extends ClusterSingletonServiceGroup<P, E, C> {
77 private enum EntityState {
79 * This entity was never registered.
83 * Registration exists, but we are waiting for it to resolve.
87 * Registration indicated we are the owner.
91 * Registration indicated we are the owner, but global state is uncertain -- meaning there can be owners in
92 * another partition, for example.
96 * Registration indicated we are not the owner. In this state we do not care about global state, therefore we
97 * do not need an UNOWNED_JEOPARDY state.
104 * Local service is up and running.
106 // FIXME: we should support async startup, which will require a STARTING state.
109 * Local service is being stopped.
114 private static final Logger LOG = LoggerFactory.getLogger(ClusterSingletonServiceGroupImpl.class);
116 private final S entityOwnershipService;
117 private final String identifier;
119 /* Entity instances */
120 private final E serviceEntity;
121 private final E cleanupEntity;
123 private final Set<ClusterSingletonServiceRegistration> members = ConcurrentHashMap.newKeySet();
125 private final Map<ClusterSingletonServiceRegistration, ServiceInfo> services = new HashMap<>();
127 // Marker for when any state changed
128 @SuppressWarnings("rawtypes")
129 private static final AtomicIntegerFieldUpdater<ClusterSingletonServiceGroupImpl> DIRTY_UPDATER =
130 AtomicIntegerFieldUpdater.newUpdater(ClusterSingletonServiceGroupImpl.class, "dirty");
131 private volatile int dirty;
133 // Simplified lock: non-reentrant, support tryLock() only
134 @SuppressWarnings("rawtypes")
135 private static final AtomicIntegerFieldUpdater<ClusterSingletonServiceGroupImpl> LOCK_UPDATER =
136 AtomicIntegerFieldUpdater.newUpdater(ClusterSingletonServiceGroupImpl.class, "lock");
137 @SuppressWarnings("unused")
138 private volatile int lock;
141 * State tracking is quite involved, as we are tracking up to four asynchronous sources of events:
142 * - user calling close()
143 * - service entity ownership
144 * - cleanup entity ownership
145 * - service shutdown future
147 * Absolutely correct solution would be a set of behaviors, which govern each state, remembering where we want to
148 * get to and what we are doing. That would result in ~15 classes which would quickly render this code unreadable
149 * due to boilerplate overhead.
151 * We therefore take a different approach, tracking state directly in this class and evaluate state transitions
152 * based on recorded bits -- without explicit representation of state machine state.
155 * Group close future. In can only go from null to non-null reference. Whenever it is non-null, it indicates that
156 * the user has closed the group and we are converging to termination.
158 // We are using volatile get-and-set to support non-blocking close(). It may be more efficient to inline it here,
159 // as we perform a volatile read after unlocking -- that volatile read may easier on L1 cache.
160 // XXX: above needs a microbenchmark contention ever becomes a problem.
161 private final AtomicReference<SettableFuture<Void>> closeFuture = new AtomicReference<>();
164 * Service (base) entity registration. This entity selects an owner candidate across nodes. Candidates proceed to
165 * acquire {@link #cleanupEntity}.
168 private GenericEntityOwnershipCandidateRegistration<P, E> serviceEntityReg = null;
170 * Service (base) entity last reported state.
173 private EntityState serviceEntityState = EntityState.UNREGISTERED;
176 * Cleanup (owner) entity registration. This entity guards access to service state and coordinates shutdown cleanup
180 private GenericEntityOwnershipCandidateRegistration<P, E> cleanupEntityReg;
182 * Cleanup (owner) entity last reported state.
185 private EntityState cleanupEntityState = EntityState.UNREGISTERED;
187 private volatile boolean initialized;
190 * Class constructor. Note: last argument is reused as-is.
192 * @param identifier non-empty string as identifier
193 * @param mainEntity as Entity instance
194 * @param closeEntity as Entity instance
195 * @param entityOwnershipService GenericEntityOwnershipService instance
196 * @param parent parent service
197 * @param services Services list
199 ClusterSingletonServiceGroupImpl(final String identifier, final S entityOwnershipService, final E mainEntity,
200 final E closeEntity, final Collection<ClusterSingletonServiceRegistration> services) {
201 checkArgument(!identifier.isEmpty(), "Identifier may not be empty");
202 this.identifier = identifier;
203 this.entityOwnershipService = requireNonNull(entityOwnershipService);
204 this.serviceEntity = requireNonNull(mainEntity);
205 this.cleanupEntity = requireNonNull(closeEntity);
206 members.addAll(services);
208 LOG.debug("Instantiated new service group for {}", identifier);
212 ClusterSingletonServiceGroupImpl(final String identifier, final E mainEntity,
213 final E closeEntity, final S entityOwnershipService) {
214 this(identifier, entityOwnershipService, mainEntity, closeEntity, ImmutableList.of());
218 public String getIdentifier() {
223 ListenableFuture<?> closeClusterSingletonGroup() {
224 final ListenableFuture<?> ret = destroyGroup();
231 LOG.debug("Service group {} postponing sync on close", identifier);
237 private boolean isClosed() {
238 return closeFuture.get() != null;
242 void initialize() throws CandidateAlreadyRegisteredException {
245 checkState(!initialized, "Singleton group %s was already initilized", identifier);
246 LOG.debug("Initializing service group {} with services {}", identifier, members);
247 synchronized (this) {
248 serviceEntityState = EntityState.REGISTERED;
249 serviceEntityReg = entityOwnershipService.registerCandidate(serviceEntity);
257 private void checkNotClosed() {
258 checkState(!isClosed(), "Service group %s has already been closed", identifier);
262 void registerService(final ClusterSingletonServiceRegistration reg) {
263 final ClusterSingletonService service = reg.getInstance();
264 verify(identifier.equals(service.getIdentifier().getValue()));
267 checkState(initialized, "Service group %s is not initialized yet", identifier);
269 // First put the service
270 LOG.debug("Adding service {} to service group {}", service, identifier);
271 verify(members.add(reg));
275 LOG.debug("Service group {} delayed register of {}", identifier, reg);
283 ListenableFuture<?> unregisterService(final ClusterSingletonServiceRegistration reg) {
284 final ClusterSingletonService service = reg.getInstance();
285 verify(identifier.equals(service.getIdentifier().getValue()));
288 verify(members.remove(reg));
290 if (members.isEmpty()) {
291 // We need to let AbstractClusterSingletonServiceProviderImpl know this group is to be shutdown
292 // before we start applying state, because while we do not re-enter, the user is free to do whatever,
293 // notably including registering a service with the same ID from the service shutdown hook. That
294 // registration request needs to hit the successor of this group.
295 return destroyGroup();
301 LOG.debug("Service group {} delayed unregister of {}", identifier, reg);
306 private synchronized @NonNull ListenableFuture<?> destroyGroup() {
307 final SettableFuture<Void> future = SettableFuture.create();
308 if (!closeFuture.compareAndSet(null, future)) {
309 return verifyNotNull(closeFuture.get());
312 if (serviceEntityReg != null) {
313 // We are still holding the service registration, close it now...
314 LOG.debug("Service group {} unregistering service entity {}", identifier, serviceEntity);
315 serviceEntityReg.close();
316 serviceEntityReg = null;
324 void ownershipChanged(final C ownershipChange) {
325 LOG.debug("Ownership change {} for ClusterSingletonServiceGroup {}", ownershipChange, identifier);
327 synchronized (this) {
328 lockedOwnershipChanged(ownershipChange);
333 LOG.debug("Service group {} postponing ownership change sync", identifier);
342 * Handle an ownership change with the lock held. Callers are expected to handle termination conditions, this method
343 * and anything it calls must not call {@link #lockedClose(SettableFuture)}.
345 * @param ownershipChange reported change
348 private void lockedOwnershipChanged(final C ownershipChange) {
349 final E entity = ownershipChange.getEntity();
350 if (serviceEntity.equals(entity)) {
351 serviceOwnershipChanged(ownershipChange.getState(), ownershipChange.inJeopardy());
353 } else if (cleanupEntity.equals(entity)) {
354 cleanupCandidateOwnershipChanged(ownershipChange.getState(), ownershipChange.inJeopardy());
357 LOG.warn("Group {} received unrecognized change {}", identifier, ownershipChange);
362 private void cleanupCandidateOwnershipChanged(final EntityOwnershipChangeState state, final boolean jeopardy) {
365 case LOCAL_OWNERSHIP_GRANTED:
366 case LOCAL_OWNERSHIP_RETAINED_WITH_NO_CHANGE:
367 LOG.warn("Service group {} cleanup entity owned without certainty", identifier);
368 cleanupEntityState = EntityState.OWNED_JEOPARDY;
370 case LOCAL_OWNERSHIP_LOST_NEW_OWNER:
371 case LOCAL_OWNERSHIP_LOST_NO_OWNER:
372 case REMOTE_OWNERSHIP_CHANGED:
373 case REMOTE_OWNERSHIP_LOST_NO_OWNER:
374 LOG.info("Service group {} cleanup entity ownership uncertain", identifier);
375 cleanupEntityState = EntityState.UNOWNED;
378 throw new IllegalStateException("Unhandled cleanup entity jeopardy change " + state);
384 if (cleanupEntityState == EntityState.OWNED_JEOPARDY) {
385 // Pair info message with previous jeopardy
386 LOG.info("Service group {} cleanup entity ownership ascertained", identifier);
390 case LOCAL_OWNERSHIP_GRANTED:
391 case LOCAL_OWNERSHIP_RETAINED_WITH_NO_CHANGE:
392 cleanupEntityState = EntityState.OWNED;
394 case LOCAL_OWNERSHIP_LOST_NEW_OWNER:
395 case LOCAL_OWNERSHIP_LOST_NO_OWNER:
396 case REMOTE_OWNERSHIP_LOST_NO_OWNER:
397 case REMOTE_OWNERSHIP_CHANGED:
398 cleanupEntityState = EntityState.UNOWNED;
401 LOG.warn("Service group {} ignoring unhandled cleanup entity change {}", identifier, state);
407 private void serviceOwnershipChanged(final EntityOwnershipChangeState state, final boolean jeopardy) {
409 LOG.info("Service group {} service entity ownership uncertain", identifier);
411 case LOCAL_OWNERSHIP_GRANTED:
412 case LOCAL_OWNERSHIP_RETAINED_WITH_NO_CHANGE:
413 serviceEntityState = EntityState.OWNED_JEOPARDY;
415 case LOCAL_OWNERSHIP_LOST_NEW_OWNER:
416 case LOCAL_OWNERSHIP_LOST_NO_OWNER:
417 case REMOTE_OWNERSHIP_CHANGED:
418 case REMOTE_OWNERSHIP_LOST_NO_OWNER:
419 serviceEntityState = EntityState.UNOWNED;
422 throw new IllegalStateException("Unhandled cleanup entity jeopardy change " + state);
427 if (serviceEntityState == EntityState.OWNED_JEOPARDY) {
428 // Pair info message with previous jeopardy
429 LOG.info("Service group {} service entity ownership ascertained", identifier);
433 case LOCAL_OWNERSHIP_GRANTED:
434 case LOCAL_OWNERSHIP_RETAINED_WITH_NO_CHANGE:
435 LOG.debug("Service group {} acquired service entity ownership", identifier);
436 serviceEntityState = EntityState.OWNED;
438 case LOCAL_OWNERSHIP_LOST_NEW_OWNER:
439 case LOCAL_OWNERSHIP_LOST_NO_OWNER:
440 case REMOTE_OWNERSHIP_CHANGED:
441 case REMOTE_OWNERSHIP_LOST_NO_OWNER:
442 LOG.debug("Service group {} lost service entity ownership", identifier);
443 serviceEntityState = EntityState.UNOWNED;
446 LOG.warn("Service group {} ignoring unhandled cleanup entity change {}", identifier, state);
450 // has to be called with lock asserted, which will be released prior to returning
451 private void reconcileState() {
452 // Always check if there is any state change to be applied.
455 if (conditionalClean()) {
459 // We may have ran a round of reconciliation, but the either one of may have happened asynchronously:
462 // - service future completed
463 // - entity state changed
465 // We are dropping the lock, but we need to recheck dirty and try to apply state again if it is found to
466 // be dirty again. This closes the following race condition:
468 // A: runs these checks holding the lock
469 // B: modifies them, fails to acquire lock
470 // A: releases lock -> noone takes care of reconciliation
477 LOG.debug("Service group {} re-running reconciliation", identifier);
481 LOG.debug("Service group {} will be reconciled by someone else", identifier);
483 LOG.debug("Service group {} is completely reconciled", identifier);
490 private void serviceTransitionCompleted() {
497 // Has to be called with lock asserted
498 private void tryReconcileState() {
499 // First take a safe snapshot of current state on which we will base our decisions.
500 final Set<ClusterSingletonServiceRegistration> localMembers;
501 final boolean haveCleanup;
502 final boolean haveService;
503 synchronized (this) {
504 if (serviceEntityReg != null) {
505 switch (serviceEntityState) {
516 throw new IllegalStateException("Unhandled service entity state " + serviceEntityState);
522 if (haveService && cleanupEntityReg == null) {
523 // We have the service entity but have not registered for cleanup entity. Do that now and retry.
524 LOG.debug("Service group {} registering cleanup entity", identifier);
526 cleanupEntityState = EntityState.REGISTERED;
527 cleanupEntityReg = entityOwnershipService.registerCandidate(cleanupEntity);
528 } catch (CandidateAlreadyRegisteredException e) {
529 LOG.error("Service group {} failed to take ownership, aborting", identifier, e);
530 if (serviceEntityReg != null) {
531 serviceEntityReg.close();
532 serviceEntityReg = null;
539 if (cleanupEntityReg != null) {
540 switch (cleanupEntityState) {
551 throw new IllegalStateException("Unhandled service entity state " + serviceEntityState);
557 localMembers = ImmutableSet.copyOf(members);
560 if (haveService && haveCleanup) {
561 ensureServicesStarting(localMembers);
565 ensureServicesStopping();
567 if (!haveService && services.isEmpty()) {
568 LOG.debug("Service group {} has no running services", identifier);
569 final boolean canFinishClose;
570 synchronized (this) {
571 if (cleanupEntityReg != null) {
572 LOG.debug("Service group {} releasing cleanup entity", identifier);
573 cleanupEntityReg.close();
574 cleanupEntityReg = null;
577 switch (cleanupEntityState) {
581 // When we are registered we need to wait for registration to resolve, otherwise
582 // the notification could be routed to the next incarnation of this group -- which could be
583 // confused by the fact it is not registered, but receives, for example, OWNED notification.
584 canFinishClose = false;
588 canFinishClose = true;
591 throw new IllegalStateException("Unhandled cleanup entity state " + cleanupEntityState);
595 if (canFinishClose) {
596 final SettableFuture<Void> localFuture = closeFuture.get();
597 if (localFuture != null && !localFuture.isDone()) {
598 LOG.debug("Service group {} completing termination", identifier);
599 localFuture.set(null);
605 // Has to be called with lock asserted
606 @SuppressWarnings("illegalCatch")
607 private void ensureServicesStarting(final Set<ClusterSingletonServiceRegistration> localConfig) {
608 LOG.debug("Service group {} starting services", identifier);
610 // This may look counter-intuitive, but the localConfig may be missing some services that are started -- for
611 // example when this method is executed as part of unregisterService() call. In that case we need to ensure
612 // services in the list are stopping
613 final Iterator<Entry<ClusterSingletonServiceRegistration, ServiceInfo>> it = services.entrySet().iterator();
614 while (it.hasNext()) {
615 final Entry<ClusterSingletonServiceRegistration, ServiceInfo> entry = it.next();
616 final ClusterSingletonServiceRegistration reg = entry.getKey();
617 if (!localConfig.contains(reg)) {
618 final ServiceInfo newInfo = ensureStopping(reg, entry.getValue());
619 if (newInfo != null) {
620 entry.setValue(newInfo);
627 // Now make sure member services are being juggled around
628 for (ClusterSingletonServiceRegistration reg : localConfig) {
629 if (!services.containsKey(reg)) {
630 final ClusterSingletonService service = reg.getInstance();
631 LOG.debug("Starting service {}", service);
634 service.instantiateServiceInstance();
635 } catch (Exception e) {
636 LOG.warn("Service group {} service {} failed to start, attempting to continue", identifier, service,
641 services.put(reg, ServiceInfo.started());
646 // Has to be called with lock asserted
647 private void ensureServicesStopping() {
648 final Iterator<Entry<ClusterSingletonServiceRegistration, ServiceInfo>> it = services.entrySet().iterator();
649 while (it.hasNext()) {
650 final Entry<ClusterSingletonServiceRegistration, ServiceInfo> entry = it.next();
651 final ServiceInfo newInfo = ensureStopping(entry.getKey(), entry.getValue());
652 if (newInfo != null) {
653 entry.setValue(newInfo);
660 @SuppressWarnings("illegalCatch")
661 private ServiceInfo ensureStopping(final ClusterSingletonServiceRegistration reg, final ServiceInfo info) {
662 switch (info.getState()) {
664 final ClusterSingletonService service = reg.getInstance();
666 LOG.debug("Service group {} stopping service {}", identifier, service);
667 final @NonNull ListenableFuture<?> future;
669 future = verifyNotNull(service.closeServiceInstance());
670 } catch (Exception e) {
671 LOG.warn("Service group {} service {} failed to stop, attempting to continue", identifier, service,
676 Futures.addCallback(future, new FutureCallback<Object>() {
678 public void onSuccess(final Object result) {
679 LOG.debug("Service group {} service {} stopped successfully", identifier, service);
680 serviceTransitionCompleted();
684 public void onFailure(final Throwable cause) {
685 LOG.debug("Service group {} service {} stopped with error", identifier, service, cause);
686 serviceTransitionCompleted();
688 }, MoreExecutors.directExecutor());
689 return info.toState(ServiceState.STOPPING, future);
691 if (info.getFuture().isDone()) {
692 LOG.debug("Service group {} removed stopped service {}", identifier, reg.getInstance());
697 throw new IllegalStateException("Unhandled state " + info.getState());
701 private void markDirty() {
705 private boolean isDirty() {
709 private boolean conditionalClean() {
710 return DIRTY_UPDATER.compareAndSet(this, 1, 0);
713 private boolean tryLock() {
714 return LOCK_UPDATER.compareAndSet(this, 0, 1);
717 private boolean unlock() {
718 verify(LOCK_UPDATER.compareAndSet(this, 1, 0));
723 public String toString() {
724 return MoreObjects.toStringHelper(this).add("identifier", identifier).toString();