Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 1 | package org.onlab.onos.store.cluster.impl; |
| 2 | |
| 3 | import static com.google.common.base.Preconditions.checkArgument; |
| 4 | import static com.google.common.base.Verify.verifyNotNull; |
| 5 | import static org.onlab.util.Tools.namedThreads; |
| 6 | import static org.slf4j.LoggerFactory.getLogger; |
| 7 | |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame^] | 8 | import java.io.IOException; |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 9 | import java.util.Map; |
| 10 | import java.util.Set; |
| 11 | import java.util.concurrent.Executors; |
| 12 | import java.util.concurrent.ScheduledExecutorService; |
| 13 | import java.util.concurrent.TimeUnit; |
| 14 | |
| 15 | import org.apache.felix.scr.annotations.Activate; |
| 16 | import org.apache.felix.scr.annotations.Component; |
| 17 | import org.apache.felix.scr.annotations.Deactivate; |
| 18 | import org.apache.felix.scr.annotations.Reference; |
| 19 | import org.apache.felix.scr.annotations.ReferenceCardinality; |
| 20 | import org.apache.felix.scr.annotations.Service; |
| 21 | import org.onlab.onos.cluster.ClusterService; |
| 22 | import org.onlab.onos.cluster.ControllerNode; |
| 23 | import org.onlab.onos.cluster.Leadership; |
| 24 | import org.onlab.onos.cluster.LeadershipEvent; |
| 25 | import org.onlab.onos.cluster.LeadershipEventListener; |
| 26 | import org.onlab.onos.cluster.LeadershipService; |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame^] | 27 | import org.onlab.onos.store.cluster.messaging.ClusterCommunicationService; |
| 28 | import org.onlab.onos.store.cluster.messaging.ClusterMessage; |
| 29 | import org.onlab.onos.store.cluster.messaging.ClusterMessageHandler; |
| 30 | import org.onlab.onos.store.cluster.messaging.MessageSubject; |
| 31 | import org.onlab.onos.store.serializers.KryoNamespaces; |
| 32 | import org.onlab.onos.store.serializers.KryoSerializer; |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 33 | import org.onlab.onos.store.service.Lock; |
| 34 | import org.onlab.onos.store.service.LockService; |
| 35 | import org.onlab.onos.store.service.impl.DistributedLockManager; |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame^] | 36 | import org.onlab.util.KryoNamespace; |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 37 | import org.slf4j.Logger; |
| 38 | |
| 39 | import com.google.common.collect.Maps; |
| 40 | import com.google.common.collect.Sets; |
| 41 | |
| 42 | /** |
| 43 | * Distributed implementation of LeadershipService that is based on the primitives exposed by |
| 44 | * LockService. |
| 45 | */ |
| 46 | @Component(immediate = true) |
| 47 | @Service |
| 48 | public class LeadershipManager implements LeadershipService { |
| 49 | |
| 50 | private final Logger log = getLogger(getClass()); |
| 51 | |
| 52 | // TODO: Remove this dependency |
| 53 | private static final int TERM_DURATION_MS = |
| 54 | DistributedLockManager.DEAD_LOCK_TIMEOUT_MS; |
| 55 | |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame^] | 56 | // Time to wait before retrying leadership after |
| 57 | // a unexpected error. |
| 58 | private static final int WAIT_BEFORE_RETRY_MS = 2000; |
| 59 | |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 60 | // TODO: Appropriate Thread pool sizing. |
| 61 | private static final ScheduledExecutorService THREAD_POOL = |
| 62 | Executors.newScheduledThreadPool(25, namedThreads("leadership-manager-%d")); |
| 63 | |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame^] | 64 | private static final MessageSubject LEADERSHIP_UPDATES = |
| 65 | new MessageSubject("leadership-contest-updates"); |
| 66 | |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 67 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
| 68 | private ClusterService clusterService; |
| 69 | |
| 70 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame^] | 71 | private ClusterCommunicationService clusterCommunicator; |
| 72 | |
| 73 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 74 | private LockService lockService; |
| 75 | |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame^] | 76 | private final Map<String, Leadership> leaderBoard = Maps.newHashMap(); |
| 77 | |
| 78 | private final Map<String, Lock> openContests = Maps.newHashMap(); |
| 79 | private final Set<LeadershipEventListener> listeners = Sets.newIdentityHashSet(); |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 80 | private ControllerNode localNode; |
| 81 | |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame^] | 82 | private final LeadershipEventListener peerAdvertiser = new PeerAdvertiser(); |
| 83 | private final LeadershipEventListener leaderBoardUpdater = new LeaderBoardUpdater(); |
| 84 | |
| 85 | public static final KryoSerializer SERIALIZER = new KryoSerializer() { |
| 86 | @Override |
| 87 | protected void setupKryoPool() { |
| 88 | serializerPool = KryoNamespace.newBuilder() |
| 89 | .register(KryoNamespaces.API) |
| 90 | .build() |
| 91 | .populate(1); |
| 92 | } |
| 93 | }; |
| 94 | |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 95 | @Activate |
| 96 | public void activate() { |
| 97 | localNode = clusterService.getLocalNode(); |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame^] | 98 | |
| 99 | addListener(peerAdvertiser); |
| 100 | addListener(leaderBoardUpdater); |
| 101 | |
| 102 | clusterCommunicator.addSubscriber( |
| 103 | LEADERSHIP_UPDATES, |
| 104 | new PeerAdvertisementHandler()); |
| 105 | |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 106 | log.info("Started."); |
| 107 | } |
| 108 | |
| 109 | @Deactivate |
| 110 | public void deactivate() { |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame^] | 111 | removeListener(peerAdvertiser); |
| 112 | removeListener(leaderBoardUpdater); |
| 113 | |
| 114 | clusterCommunicator.removeSubscriber(LEADERSHIP_UPDATES); |
| 115 | |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 116 | THREAD_POOL.shutdown(); |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame^] | 117 | |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 118 | log.info("Stopped."); |
| 119 | } |
| 120 | |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame^] | 121 | |
| 122 | @Override |
| 123 | public ControllerNode getLeader(String path) { |
| 124 | synchronized (leaderBoard) { |
| 125 | Leadership leadership = leaderBoard.get(path); |
| 126 | if (leadership != null) { |
| 127 | return leadership.leader(); |
| 128 | } |
| 129 | } |
| 130 | return null; |
| 131 | } |
| 132 | |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 133 | @Override |
| 134 | public void runForLeadership(String path) { |
| 135 | checkArgument(path != null); |
| 136 | if (openContests.containsKey(path)) { |
| 137 | log.info("Already in the leadership contest for {}", path); |
| 138 | return; |
| 139 | } else { |
| 140 | Lock lock = lockService.create(path); |
| 141 | openContests.put(path, lock); |
| 142 | tryAcquireLeadership(path); |
| 143 | } |
| 144 | } |
| 145 | |
| 146 | @Override |
| 147 | public void withdraw(String path) { |
| 148 | checkArgument(path != null); |
| 149 | Lock lock = openContests.remove(path); |
| 150 | |
| 151 | if (lock != null && lock.isLocked()) { |
| 152 | lock.unlock(); |
| 153 | notifyListeners( |
| 154 | new LeadershipEvent( |
| 155 | LeadershipEvent.Type.LEADER_BOOTED, |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame^] | 156 | new Leadership(lock.path(), localNode, lock.epoch()))); |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 157 | } |
| 158 | } |
| 159 | |
| 160 | @Override |
| 161 | public void addListener(LeadershipEventListener listener) { |
| 162 | checkArgument(listener != null); |
| 163 | listeners.add(listener); |
| 164 | } |
| 165 | |
| 166 | @Override |
| 167 | public void removeListener(LeadershipEventListener listener) { |
| 168 | checkArgument(listener != null); |
| 169 | listeners.remove(listener); |
| 170 | } |
| 171 | |
| 172 | private void notifyListeners(LeadershipEvent event) { |
| 173 | for (LeadershipEventListener listener : listeners) { |
| 174 | listener.event(event); |
| 175 | } |
| 176 | } |
| 177 | |
| 178 | private void tryAcquireLeadership(String path) { |
| 179 | Lock lock = openContests.get(path); |
| 180 | verifyNotNull(lock, "Lock should not be null"); |
| 181 | lock.lockAsync(TERM_DURATION_MS).whenComplete((response, error) -> { |
| 182 | if (error == null) { |
| 183 | THREAD_POOL.schedule( |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame^] | 184 | new ReelectionTask(lock), |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 185 | TERM_DURATION_MS / 2, |
| 186 | TimeUnit.MILLISECONDS); |
| 187 | notifyListeners( |
| 188 | new LeadershipEvent( |
| 189 | LeadershipEvent.Type.LEADER_ELECTED, |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame^] | 190 | new Leadership(lock.path(), localNode, lock.epoch()))); |
| 191 | return; |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 192 | } else { |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame^] | 193 | log.warn("Failed to acquire lock for {}. Will retry in {} sec", path, WAIT_BEFORE_RETRY_MS, error); |
| 194 | try { |
| 195 | Thread.sleep(WAIT_BEFORE_RETRY_MS); |
| 196 | tryAcquireLeadership(path); |
| 197 | } catch (InterruptedException e) { |
| 198 | Thread.currentThread().interrupt(); |
| 199 | } |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 200 | } |
| 201 | }); |
| 202 | } |
| 203 | |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame^] | 204 | private class ReelectionTask implements Runnable { |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 205 | |
| 206 | private final Lock lock; |
| 207 | |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame^] | 208 | public ReelectionTask(Lock lock) { |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 209 | this.lock = lock; |
| 210 | } |
| 211 | |
| 212 | @Override |
| 213 | public void run() { |
| 214 | if (lock.extendExpiration(TERM_DURATION_MS)) { |
| 215 | notifyListeners( |
| 216 | new LeadershipEvent( |
| 217 | LeadershipEvent.Type.LEADER_REELECTED, |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame^] | 218 | new Leadership(lock.path(), localNode, lock.epoch()))); |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 219 | THREAD_POOL.schedule(this, TERM_DURATION_MS / 2, TimeUnit.MILLISECONDS); |
| 220 | } else { |
| 221 | if (openContests.containsKey(lock.path())) { |
| 222 | notifyListeners( |
| 223 | new LeadershipEvent( |
| 224 | LeadershipEvent.Type.LEADER_BOOTED, |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame^] | 225 | new Leadership(lock.path(), localNode, lock.epoch()))); |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 226 | tryAcquireLeadership(lock.path()); |
| 227 | } |
| 228 | } |
| 229 | } |
| 230 | } |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame^] | 231 | |
| 232 | private class PeerAdvertiser implements LeadershipEventListener { |
| 233 | @Override |
| 234 | public void event(LeadershipEvent event) { |
| 235 | // publish events originating on this host. |
| 236 | if (event.subject().leader().equals(localNode)) { |
| 237 | try { |
| 238 | clusterCommunicator.broadcast( |
| 239 | new ClusterMessage( |
| 240 | localNode.id(), |
| 241 | LEADERSHIP_UPDATES, |
| 242 | SERIALIZER.encode(event))); |
| 243 | } catch (IOException e) { |
| 244 | log.error("Failed to broadcast leadership update message", e); |
| 245 | } |
| 246 | } |
| 247 | } |
| 248 | } |
| 249 | |
| 250 | private class PeerAdvertisementHandler implements ClusterMessageHandler { |
| 251 | @Override |
| 252 | public void handle(ClusterMessage message) { |
| 253 | LeadershipEvent event = SERIALIZER.decode(message.payload()); |
| 254 | log.debug("Received {} from {}", event, message.sender()); |
| 255 | notifyListeners(event); |
| 256 | } |
| 257 | } |
| 258 | |
| 259 | private class LeaderBoardUpdater implements LeadershipEventListener { |
| 260 | @Override |
| 261 | public void event(LeadershipEvent event) { |
| 262 | Leadership leadershipUpdate = event.subject(); |
| 263 | synchronized (leaderBoard) { |
| 264 | Leadership currentLeadership = leaderBoard.get(leadershipUpdate.topic()); |
| 265 | switch (event.type()) { |
| 266 | case LEADER_ELECTED: |
| 267 | case LEADER_REELECTED: |
| 268 | if (currentLeadership == null || currentLeadership.epoch() < leadershipUpdate.epoch()) { |
| 269 | leaderBoard.put(leadershipUpdate.topic(), leadershipUpdate); |
| 270 | } |
| 271 | break; |
| 272 | case LEADER_BOOTED: |
| 273 | if (currentLeadership != null && currentLeadership.epoch() <= leadershipUpdate.epoch()) { |
| 274 | leaderBoard.remove(leadershipUpdate.topic()); |
| 275 | } |
| 276 | break; |
| 277 | default: |
| 278 | break; |
| 279 | } |
| 280 | } |
| 281 | } |
| 282 | } |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 283 | } |