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; |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 4 | import static org.onlab.util.Tools.namedThreads; |
| 5 | import static org.slf4j.LoggerFactory.getLogger; |
| 6 | |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame] | 7 | import java.io.IOException; |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 8 | import java.util.Map; |
| 9 | import java.util.Set; |
| 10 | import java.util.concurrent.Executors; |
| 11 | import java.util.concurrent.ScheduledExecutorService; |
| 12 | import java.util.concurrent.TimeUnit; |
| 13 | |
| 14 | import org.apache.felix.scr.annotations.Activate; |
| 15 | import org.apache.felix.scr.annotations.Component; |
| 16 | import org.apache.felix.scr.annotations.Deactivate; |
| 17 | import org.apache.felix.scr.annotations.Reference; |
| 18 | import org.apache.felix.scr.annotations.ReferenceCardinality; |
| 19 | import org.apache.felix.scr.annotations.Service; |
| 20 | import org.onlab.onos.cluster.ClusterService; |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 21 | import org.onlab.onos.cluster.Leadership; |
| 22 | import org.onlab.onos.cluster.LeadershipEvent; |
| 23 | import org.onlab.onos.cluster.LeadershipEventListener; |
| 24 | import org.onlab.onos.cluster.LeadershipService; |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 25 | import org.onlab.onos.cluster.NodeId; |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame] | 26 | import org.onlab.onos.store.cluster.messaging.ClusterCommunicationService; |
| 27 | import org.onlab.onos.store.cluster.messaging.ClusterMessage; |
| 28 | import org.onlab.onos.store.cluster.messaging.ClusterMessageHandler; |
| 29 | import org.onlab.onos.store.cluster.messaging.MessageSubject; |
| 30 | import org.onlab.onos.store.serializers.KryoNamespaces; |
| 31 | import org.onlab.onos.store.serializers.KryoSerializer; |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 32 | import org.onlab.onos.store.service.Lock; |
| 33 | import org.onlab.onos.store.service.LockService; |
| 34 | import org.onlab.onos.store.service.impl.DistributedLockManager; |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame] | 35 | import org.onlab.util.KryoNamespace; |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 36 | import org.slf4j.Logger; |
| 37 | |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 38 | import com.google.common.collect.ImmutableMap; |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 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 | */ |
Jonathan Hart | 949c284 | 2014-11-28 23:44:09 -0800 | [diff] [blame] | 46 | @Component(enabled = false) |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 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. |
Madan Jampani | ddaffd0 | 2014-11-21 13:12:09 -0800 | [diff] [blame] | 61 | private final ScheduledExecutorService threadPool = |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 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 | |
Madan Jampani | d1a467f | 2014-11-24 15:11:09 -0800 | [diff] [blame] | 78 | private final Map<String, Lock> openContests = Maps.newConcurrentMap(); |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame] | 79 | private final Set<LeadershipEventListener> listeners = Sets.newIdentityHashSet(); |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 80 | private NodeId localNodeId; |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 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() { |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 97 | localNodeId = clusterService.getLocalNode().id(); |
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 | ddaffd0 | 2014-11-21 13:12:09 -0800 | [diff] [blame] | 116 | threadPool.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 |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 123 | public NodeId getLeader(String path) { |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame] | 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); |
Madan Jampani | d1a467f | 2014-11-24 15:11:09 -0800 | [diff] [blame] | 136 | |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 137 | if (openContests.containsKey(path)) { |
| 138 | log.info("Already in the leadership contest for {}", path); |
| 139 | return; |
| 140 | } else { |
| 141 | Lock lock = lockService.create(path); |
| 142 | openContests.put(path, lock); |
Madan Jampani | d1a467f | 2014-11-24 15:11:09 -0800 | [diff] [blame] | 143 | threadPool.schedule(new TryLeadership(lock), 0, TimeUnit.MILLISECONDS); |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 144 | } |
| 145 | } |
| 146 | |
| 147 | @Override |
| 148 | public void withdraw(String path) { |
| 149 | checkArgument(path != null); |
Madan Jampani | d1a467f | 2014-11-24 15:11:09 -0800 | [diff] [blame] | 150 | |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 151 | Lock lock = openContests.remove(path); |
| 152 | |
| 153 | if (lock != null && lock.isLocked()) { |
| 154 | lock.unlock(); |
| 155 | notifyListeners( |
| 156 | new LeadershipEvent( |
| 157 | LeadershipEvent.Type.LEADER_BOOTED, |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 158 | new Leadership(lock.path(), localNodeId, lock.epoch()))); |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 159 | } |
| 160 | } |
| 161 | |
| 162 | @Override |
Yuta HIGUCHI | c2bf3d8 | 2014-11-28 18:50:41 -0800 | [diff] [blame] | 163 | public Map<String, Leadership> getLeaderBoard() { |
| 164 | return ImmutableMap.copyOf(leaderBoard); |
| 165 | } |
| 166 | |
| 167 | @Override |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 168 | public void addListener(LeadershipEventListener listener) { |
| 169 | checkArgument(listener != null); |
| 170 | listeners.add(listener); |
| 171 | } |
| 172 | |
| 173 | @Override |
| 174 | public void removeListener(LeadershipEventListener listener) { |
| 175 | checkArgument(listener != null); |
| 176 | listeners.remove(listener); |
| 177 | } |
| 178 | |
| 179 | private void notifyListeners(LeadershipEvent event) { |
| 180 | for (LeadershipEventListener listener : listeners) { |
Madan Jampani | d1a467f | 2014-11-24 15:11:09 -0800 | [diff] [blame] | 181 | try { |
| 182 | listener.event(event); |
| 183 | } catch (Exception e) { |
| 184 | log.error("Notifying listener failed with exception.", e); |
| 185 | } |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 186 | } |
| 187 | } |
| 188 | |
| 189 | private void tryAcquireLeadership(String path) { |
| 190 | Lock lock = openContests.get(path); |
Madan Jampani | d1a467f | 2014-11-24 15:11:09 -0800 | [diff] [blame] | 191 | if (lock == null) { |
| 192 | // withdrew from race. |
| 193 | return; |
| 194 | } |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 195 | lock.lockAsync(TERM_DURATION_MS).whenComplete((response, error) -> { |
| 196 | if (error == null) { |
Madan Jampani | ddaffd0 | 2014-11-21 13:12:09 -0800 | [diff] [blame] | 197 | threadPool.schedule( |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame] | 198 | new ReelectionTask(lock), |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 199 | TERM_DURATION_MS / 2, |
| 200 | TimeUnit.MILLISECONDS); |
| 201 | notifyListeners( |
| 202 | new LeadershipEvent( |
| 203 | LeadershipEvent.Type.LEADER_ELECTED, |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 204 | new Leadership(lock.path(), localNodeId, lock.epoch()))); |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame] | 205 | return; |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 206 | } else { |
Madan Jampani | d1a467f | 2014-11-24 15:11:09 -0800 | [diff] [blame] | 207 | log.warn("Failed to acquire lock for {}. Will retry in {} ms", path, WAIT_BEFORE_RETRY_MS, error); |
| 208 | threadPool.schedule(new TryLeadership(lock), WAIT_BEFORE_RETRY_MS, TimeUnit.MILLISECONDS); |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 209 | } |
| 210 | }); |
| 211 | } |
| 212 | |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame] | 213 | private class ReelectionTask implements Runnable { |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 214 | |
| 215 | private final Lock lock; |
| 216 | |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame] | 217 | public ReelectionTask(Lock lock) { |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 218 | this.lock = lock; |
| 219 | } |
| 220 | |
| 221 | @Override |
| 222 | public void run() { |
Madan Jampani | d1a467f | 2014-11-24 15:11:09 -0800 | [diff] [blame] | 223 | if (!openContests.containsKey(lock.path())) { |
| 224 | log.debug("Node withdrew from leadership race for {}. Cancelling reelection task.", lock.path()); |
| 225 | return; |
| 226 | } |
| 227 | |
| 228 | boolean lockExtended = false; |
| 229 | try { |
| 230 | lockExtended = lock.extendExpiration(TERM_DURATION_MS); |
| 231 | } catch (Exception e) { |
| 232 | log.warn("Attempt to extend lock failed with an exception.", e); |
| 233 | } |
| 234 | |
| 235 | if (lockExtended) { |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 236 | notifyListeners( |
| 237 | new LeadershipEvent( |
| 238 | LeadershipEvent.Type.LEADER_REELECTED, |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 239 | new Leadership(lock.path(), localNodeId, lock.epoch()))); |
Madan Jampani | ddaffd0 | 2014-11-21 13:12:09 -0800 | [diff] [blame] | 240 | threadPool.schedule(this, TERM_DURATION_MS / 2, TimeUnit.MILLISECONDS); |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 241 | } else { |
Madan Jampani | d1a467f | 2014-11-24 15:11:09 -0800 | [diff] [blame] | 242 | // Check if this node already withdrew from the contest, in which case |
| 243 | // we don't need to notify here. |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 244 | if (openContests.containsKey(lock.path())) { |
| 245 | notifyListeners( |
| 246 | new LeadershipEvent( |
| 247 | LeadershipEvent.Type.LEADER_BOOTED, |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 248 | new Leadership(lock.path(), localNodeId, lock.epoch()))); |
Madan Jampani | d1a467f | 2014-11-24 15:11:09 -0800 | [diff] [blame] | 249 | // Retry leadership after a brief wait. |
| 250 | threadPool.schedule(new TryLeadership(lock), WAIT_BEFORE_RETRY_MS, TimeUnit.MILLISECONDS); |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 251 | } |
| 252 | } |
| 253 | } |
| 254 | } |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame] | 255 | |
Madan Jampani | d1a467f | 2014-11-24 15:11:09 -0800 | [diff] [blame] | 256 | private class TryLeadership implements Runnable { |
| 257 | private final Lock lock; |
| 258 | |
| 259 | public TryLeadership(Lock lock) { |
| 260 | this.lock = lock; |
| 261 | } |
| 262 | |
| 263 | @Override |
| 264 | public void run() { |
| 265 | tryAcquireLeadership(lock.path()); |
| 266 | } |
| 267 | } |
| 268 | |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame] | 269 | private class PeerAdvertiser implements LeadershipEventListener { |
| 270 | @Override |
| 271 | public void event(LeadershipEvent event) { |
| 272 | // publish events originating on this host. |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 273 | if (event.subject().leader().equals(localNodeId)) { |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame] | 274 | try { |
| 275 | clusterCommunicator.broadcast( |
| 276 | new ClusterMessage( |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 277 | localNodeId, |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame] | 278 | LEADERSHIP_UPDATES, |
| 279 | SERIALIZER.encode(event))); |
| 280 | } catch (IOException e) { |
| 281 | log.error("Failed to broadcast leadership update message", e); |
| 282 | } |
| 283 | } |
| 284 | } |
| 285 | } |
| 286 | |
| 287 | private class PeerAdvertisementHandler implements ClusterMessageHandler { |
| 288 | @Override |
| 289 | public void handle(ClusterMessage message) { |
| 290 | LeadershipEvent event = SERIALIZER.decode(message.payload()); |
Yuta HIGUCHI | 47b2f55 | 2014-11-28 20:13:15 -0800 | [diff] [blame] | 291 | log.trace("Received {} from {}", event, message.sender()); |
Madan Jampani | 1ee9178 | 2014-11-20 20:24:24 -0800 | [diff] [blame] | 292 | notifyListeners(event); |
| 293 | } |
| 294 | } |
| 295 | |
| 296 | private class LeaderBoardUpdater implements LeadershipEventListener { |
| 297 | @Override |
| 298 | public void event(LeadershipEvent event) { |
| 299 | Leadership leadershipUpdate = event.subject(); |
| 300 | synchronized (leaderBoard) { |
| 301 | Leadership currentLeadership = leaderBoard.get(leadershipUpdate.topic()); |
| 302 | switch (event.type()) { |
| 303 | case LEADER_ELECTED: |
| 304 | case LEADER_REELECTED: |
| 305 | if (currentLeadership == null || currentLeadership.epoch() < leadershipUpdate.epoch()) { |
| 306 | leaderBoard.put(leadershipUpdate.topic(), leadershipUpdate); |
| 307 | } |
| 308 | break; |
| 309 | case LEADER_BOOTED: |
| 310 | if (currentLeadership != null && currentLeadership.epoch() <= leadershipUpdate.epoch()) { |
| 311 | leaderBoard.remove(leadershipUpdate.topic()); |
| 312 | } |
| 313 | break; |
| 314 | default: |
| 315 | break; |
| 316 | } |
| 317 | } |
| 318 | } |
| 319 | } |
Madan Jampani | 1d3494e | 2014-11-20 11:24:22 -0800 | [diff] [blame] | 320 | } |