Thomas Vachuska | 4f1a60c | 2014-10-28 13:39:07 -0700 | [diff] [blame] | 1 | /* |
Brian O'Connor | a09fe5b | 2017-08-03 21:12:30 -0700 | [diff] [blame] | 2 | * Copyright 2014-present Open Networking Foundation |
Thomas Vachuska | 4f1a60c | 2014-10-28 13:39:07 -0700 | [diff] [blame] | 3 | * |
| 4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | * you may not use this file except in compliance with the License. |
| 6 | * You may obtain a copy of the License at |
| 7 | * |
| 8 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | * |
| 10 | * Unless required by applicable law or agreed to in writing, software |
| 11 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | * See the License for the specific language governing permissions and |
| 14 | * limitations under the License. |
| 15 | */ |
Brian O'Connor | abafb50 | 2014-12-02 22:26:20 -0800 | [diff] [blame] | 16 | package org.onosproject.cluster.impl; |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 17 | |
Thomas Vachuska | 1e68bdd | 2014-11-29 13:53:10 -0800 | [diff] [blame] | 18 | import com.codahale.metrics.Timer; |
| 19 | import com.codahale.metrics.Timer.Context; |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 20 | import com.google.common.collect.Lists; |
Claudine Chiu | dce0815 | 2016-03-09 18:19:28 +0000 | [diff] [blame] | 21 | import com.google.common.collect.Sets; |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 22 | import com.google.common.util.concurrent.Futures; |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 23 | import org.apache.felix.scr.annotations.Activate; |
tom | 4a5d171 | 2014-09-23 17:49:39 -0700 | [diff] [blame] | 24 | import org.apache.felix.scr.annotations.Component; |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 25 | import org.apache.felix.scr.annotations.Deactivate; |
Victor Silva | f2b9d03 | 2016-09-19 19:43:20 -0300 | [diff] [blame] | 26 | import org.apache.felix.scr.annotations.Modified; |
| 27 | import org.apache.felix.scr.annotations.Property; |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 28 | import org.apache.felix.scr.annotations.Reference; |
| 29 | import org.apache.felix.scr.annotations.ReferenceCardinality; |
tom | 4a5d171 | 2014-09-23 17:49:39 -0700 | [diff] [blame] | 30 | import org.apache.felix.scr.annotations.Service; |
Yuta HIGUCHI | 6a46283 | 2014-11-23 23:56:03 -0800 | [diff] [blame] | 31 | import org.onlab.metrics.MetricsService; |
Victor Silva | f2b9d03 | 2016-09-19 19:43:20 -0300 | [diff] [blame] | 32 | import org.onosproject.cfg.ComponentConfigService; |
| 33 | import org.onosproject.cfg.ConfigProperty; |
Brian O'Connor | abafb50 | 2014-12-02 22:26:20 -0800 | [diff] [blame] | 34 | import org.onosproject.cluster.ClusterService; |
| 35 | import org.onosproject.cluster.ControllerNode; |
| 36 | import org.onosproject.cluster.NodeId; |
| 37 | import org.onosproject.cluster.RoleInfo; |
| 38 | import org.onosproject.core.MetricsHelper; |
Thomas Vachuska | 7a8de84 | 2016-03-07 20:56:35 -0800 | [diff] [blame] | 39 | import org.onosproject.event.AbstractListenerManager; |
Brian O'Connor | abafb50 | 2014-12-02 22:26:20 -0800 | [diff] [blame] | 40 | import org.onosproject.mastership.MastershipAdminService; |
| 41 | import org.onosproject.mastership.MastershipEvent; |
| 42 | import org.onosproject.mastership.MastershipListener; |
| 43 | import org.onosproject.mastership.MastershipService; |
| 44 | import org.onosproject.mastership.MastershipStore; |
| 45 | import org.onosproject.mastership.MastershipStoreDelegate; |
| 46 | import org.onosproject.mastership.MastershipTerm; |
| 47 | import org.onosproject.mastership.MastershipTermService; |
| 48 | import org.onosproject.net.DeviceId; |
| 49 | import org.onosproject.net.MastershipRole; |
Claudine Chiu | dce0815 | 2016-03-09 18:19:28 +0000 | [diff] [blame] | 50 | import org.onosproject.net.region.Region; |
| 51 | import org.onosproject.net.region.RegionService; |
Jordan Halterman | 713830d | 2017-10-07 13:40:44 -0700 | [diff] [blame] | 52 | import org.onosproject.upgrade.UpgradeEvent; |
| 53 | import org.onosproject.upgrade.UpgradeEventListener; |
| 54 | import org.onosproject.upgrade.UpgradeService; |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 55 | import org.slf4j.Logger; |
| 56 | |
Claudine Chiu | dce0815 | 2016-03-09 18:19:28 +0000 | [diff] [blame] | 57 | import java.util.ArrayList; |
Thomas Vachuska | 1e68bdd | 2014-11-29 13:53:10 -0800 | [diff] [blame] | 58 | import java.util.Collection; |
Claudine Chiu | dce0815 | 2016-03-09 18:19:28 +0000 | [diff] [blame] | 59 | import java.util.Collections; |
Thomas Vachuska | 12dfdc3 | 2014-11-29 16:03:12 -0800 | [diff] [blame] | 60 | import java.util.HashMap; |
| 61 | import java.util.HashSet; |
Thomas Vachuska | 1e68bdd | 2014-11-29 13:53:10 -0800 | [diff] [blame] | 62 | import java.util.Iterator; |
| 63 | import java.util.List; |
Thomas Vachuska | 12dfdc3 | 2014-11-29 16:03:12 -0800 | [diff] [blame] | 64 | import java.util.Map; |
Thomas Vachuska | 1e68bdd | 2014-11-29 13:53:10 -0800 | [diff] [blame] | 65 | import java.util.Set; |
Madan Jampani | f7536ab | 2015-05-07 23:23:23 -0700 | [diff] [blame] | 66 | import java.util.concurrent.CompletableFuture; |
| 67 | |
Thomas Vachuska | 1e68bdd | 2014-11-29 13:53:10 -0800 | [diff] [blame] | 68 | import static com.google.common.base.Preconditions.checkNotNull; |
| 69 | import static com.google.common.collect.Lists.newArrayList; |
| 70 | import static org.onlab.metrics.MetricsUtil.startTimer; |
| 71 | import static org.onlab.metrics.MetricsUtil.stopTimer; |
Brian O'Connor | abafb50 | 2014-12-02 22:26:20 -0800 | [diff] [blame] | 72 | import static org.onosproject.net.MastershipRole.MASTER; |
Changhoon Yoon | 541ef71 | 2015-05-23 17:18:34 +0900 | [diff] [blame] | 73 | import static org.onosproject.security.AppGuard.checkPermission; |
Thomas Vachuska | 7a8de84 | 2016-03-07 20:56:35 -0800 | [diff] [blame] | 74 | import static org.onosproject.security.AppPermission.Type.CLUSTER_READ; |
| 75 | import static org.onosproject.security.AppPermission.Type.CLUSTER_WRITE; |
Thomas Vachuska | 42e8cce | 2015-07-29 19:25:18 -0700 | [diff] [blame] | 76 | import static org.slf4j.LoggerFactory.getLogger; |
Changhoon Yoon | b856b81 | 2015-08-10 03:47:19 +0900 | [diff] [blame] | 77 | |
Changhoon Yoon | 541ef71 | 2015-05-23 17:18:34 +0900 | [diff] [blame] | 78 | |
Yuta HIGUCHI | 6a46283 | 2014-11-23 23:56:03 -0800 | [diff] [blame] | 79 | |
tom | 4a5d171 | 2014-09-23 17:49:39 -0700 | [diff] [blame] | 80 | @Component(immediate = true) |
| 81 | @Service |
Ayaka Koshibe | 3eed2b0 | 2014-09-23 13:28:05 -0700 | [diff] [blame] | 82 | public class MastershipManager |
Thomas Vachuska | 42e8cce | 2015-07-29 19:25:18 -0700 | [diff] [blame] | 83 | extends AbstractListenerManager<MastershipEvent, MastershipListener> |
Yuta HIGUCHI | a22f69f | 2014-11-24 22:25:17 -0800 | [diff] [blame] | 84 | implements MastershipService, MastershipAdminService, MastershipTermService, |
| 85 | MetricsHelper { |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 86 | |
| 87 | private static final String NODE_ID_NULL = "Node ID cannot be null"; |
| 88 | private static final String DEVICE_ID_NULL = "Device ID cannot be null"; |
| 89 | private static final String ROLE_NULL = "Mastership role cannot be null"; |
| 90 | |
| 91 | private final Logger log = getLogger(getClass()); |
| 92 | |
alshabib | 339a3d9 | 2014-09-26 17:54:32 -0700 | [diff] [blame] | 93 | private final MastershipStoreDelegate delegate = new InternalDelegate(); |
Jordan Halterman | 713830d | 2017-10-07 13:40:44 -0700 | [diff] [blame] | 94 | private final UpgradeEventListener upgradeEventListener = new InternalUpgradeEventListener(); |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 95 | |
| 96 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
| 97 | protected MastershipStore store; |
| 98 | |
| 99 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
tom | 4a5d171 | 2014-09-23 17:49:39 -0700 | [diff] [blame] | 100 | protected ClusterService clusterService; |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 101 | |
Yuta HIGUCHI | 6a46283 | 2014-11-23 23:56:03 -0800 | [diff] [blame] | 102 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
| 103 | protected MetricsService metricsService; |
| 104 | |
Claudine Chiu | dce0815 | 2016-03-09 18:19:28 +0000 | [diff] [blame] | 105 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
| 106 | protected RegionService regionService; |
| 107 | |
Victor Silva | f2b9d03 | 2016-09-19 19:43:20 -0300 | [diff] [blame] | 108 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
| 109 | protected ComponentConfigService cfgService; |
| 110 | |
Jordan Halterman | 713830d | 2017-10-07 13:40:44 -0700 | [diff] [blame] | 111 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
| 112 | protected UpgradeService upgradeService; |
| 113 | |
Madan Jampani | c6e574f | 2015-05-29 13:41:52 -0700 | [diff] [blame] | 114 | private NodeId localNodeId; |
Yuta HIGUCHI | 6a46283 | 2014-11-23 23:56:03 -0800 | [diff] [blame] | 115 | private Timer requestRoleTimer; |
Victor Silva | f2b9d03 | 2016-09-19 19:43:20 -0300 | [diff] [blame] | 116 | |
| 117 | static final boolean DEFAULT_USE_REGION_FOR_BALANCE_ROLES = false; |
| 118 | @Property(name = "useRegionForBalanceRoles", boolValue = DEFAULT_USE_REGION_FOR_BALANCE_ROLES, |
| 119 | label = "Use Regions for balancing roles") |
Jordan Halterman | 713830d | 2017-10-07 13:40:44 -0700 | [diff] [blame] | 120 | protected boolean useRegionForBalanceRoles; |
| 121 | |
| 122 | private static final boolean DEFAULT_REBALANCE_ROLES_ON_UPGRADE = true; |
| 123 | @Property(name = "rebalanceRolesOnUpgrade", |
| 124 | boolValue = DEFAULT_REBALANCE_ROLES_ON_UPGRADE, |
| 125 | label = "Automatically rebalance roles following an upgrade") |
Jordan Halterman | 0676c97 | 2017-11-05 11:49:23 -0800 | [diff] [blame] | 126 | protected boolean rebalanceRolesOnUpgrade = DEFAULT_REBALANCE_ROLES_ON_UPGRADE; |
Ayaka Koshibe | 3de43ca | 2014-09-26 16:40:23 -0700 | [diff] [blame] | 127 | |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 128 | @Activate |
| 129 | public void activate() { |
Victor Silva | f2b9d03 | 2016-09-19 19:43:20 -0300 | [diff] [blame] | 130 | cfgService.registerProperties(getClass()); |
| 131 | modified(); |
| 132 | |
Yuta HIGUCHI | 6a46283 | 2014-11-23 23:56:03 -0800 | [diff] [blame] | 133 | requestRoleTimer = createTimer("Mastership", "requestRole", "responseTime"); |
Madan Jampani | c6e574f | 2015-05-29 13:41:52 -0700 | [diff] [blame] | 134 | localNodeId = clusterService.getLocalNode().id(); |
Jordan Halterman | 61aeb35 | 2017-10-18 16:22:17 -0700 | [diff] [blame] | 135 | upgradeService.addListener(upgradeEventListener); |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 136 | eventDispatcher.addSink(MastershipEvent.class, listenerRegistry); |
alshabib | 339a3d9 | 2014-09-26 17:54:32 -0700 | [diff] [blame] | 137 | store.setDelegate(delegate); |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 138 | log.info("Started"); |
| 139 | } |
| 140 | |
Victor Silva | f2b9d03 | 2016-09-19 19:43:20 -0300 | [diff] [blame] | 141 | @Modified |
| 142 | public void modified() { |
| 143 | Set<ConfigProperty> configProperties = cfgService.getProperties(getClass().getCanonicalName()); |
| 144 | for (ConfigProperty property : configProperties) { |
Jon Hall | a3fcf67 | 2017-03-28 16:53:22 -0700 | [diff] [blame] | 145 | if ("useRegionForBalanceRoles".equals(property.name())) { |
Victor Silva | f2b9d03 | 2016-09-19 19:43:20 -0300 | [diff] [blame] | 146 | useRegionForBalanceRoles = property.asBoolean(); |
| 147 | } |
| 148 | } |
| 149 | } |
| 150 | |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 151 | @Deactivate |
| 152 | public void deactivate() { |
| 153 | eventDispatcher.removeSink(MastershipEvent.class); |
Jordan Halterman | 61aeb35 | 2017-10-18 16:22:17 -0700 | [diff] [blame] | 154 | upgradeService.removeListener(upgradeEventListener); |
alshabib | 339a3d9 | 2014-09-26 17:54:32 -0700 | [diff] [blame] | 155 | store.unsetDelegate(delegate); |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 156 | log.info("Stopped"); |
Victor Silva | f2b9d03 | 2016-09-19 19:43:20 -0300 | [diff] [blame] | 157 | cfgService.unregisterProperties(getClass(), false); |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 158 | } |
| 159 | |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 160 | @Override |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 161 | public CompletableFuture<Void> setRole(NodeId nodeId, DeviceId deviceId, MastershipRole role) { |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 162 | checkNotNull(nodeId, NODE_ID_NULL); |
| 163 | checkNotNull(deviceId, DEVICE_ID_NULL); |
| 164 | checkNotNull(role, ROLE_NULL); |
Ayaka Koshibe | d9f693e | 2014-09-29 18:04:54 -0700 | [diff] [blame] | 165 | |
Madan Jampani | f7536ab | 2015-05-07 23:23:23 -0700 | [diff] [blame] | 166 | CompletableFuture<MastershipEvent> eventFuture = null; |
Ayaka Koshibe | e60d452 | 2014-10-28 15:07:00 -0700 | [diff] [blame] | 167 | |
| 168 | switch (role) { |
| 169 | case MASTER: |
Madan Jampani | f7536ab | 2015-05-07 23:23:23 -0700 | [diff] [blame] | 170 | eventFuture = store.setMaster(nodeId, deviceId); |
Ayaka Koshibe | e60d452 | 2014-10-28 15:07:00 -0700 | [diff] [blame] | 171 | break; |
| 172 | case STANDBY: |
Madan Jampani | f7536ab | 2015-05-07 23:23:23 -0700 | [diff] [blame] | 173 | eventFuture = store.setStandby(nodeId, deviceId); |
Ayaka Koshibe | e60d452 | 2014-10-28 15:07:00 -0700 | [diff] [blame] | 174 | break; |
| 175 | case NONE: |
Madan Jampani | f7536ab | 2015-05-07 23:23:23 -0700 | [diff] [blame] | 176 | eventFuture = store.relinquishRole(nodeId, deviceId); |
Ayaka Koshibe | e60d452 | 2014-10-28 15:07:00 -0700 | [diff] [blame] | 177 | break; |
| 178 | default: |
| 179 | log.info("Unknown role; ignoring"); |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 180 | return CompletableFuture.completedFuture(null); |
Ayaka Koshibe | 971a38a | 2014-09-30 11:56:23 -0700 | [diff] [blame] | 181 | } |
Ayaka Koshibe | d9f693e | 2014-09-29 18:04:54 -0700 | [diff] [blame] | 182 | |
Madan Jampani | c6e574f | 2015-05-29 13:41:52 -0700 | [diff] [blame] | 183 | return eventFuture.thenAccept(this::post) |
| 184 | .thenApply(v -> null); |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 185 | } |
| 186 | |
| 187 | @Override |
tom | b41d1ac | 2014-09-24 01:51:24 -0700 | [diff] [blame] | 188 | public MastershipRole getLocalRole(DeviceId deviceId) { |
Changhoon Yoon | b856b81 | 2015-08-10 03:47:19 +0900 | [diff] [blame] | 189 | checkPermission(CLUSTER_READ); |
Changhoon Yoon | 541ef71 | 2015-05-23 17:18:34 +0900 | [diff] [blame] | 190 | |
tom | b41d1ac | 2014-09-24 01:51:24 -0700 | [diff] [blame] | 191 | checkNotNull(deviceId, DEVICE_ID_NULL); |
| 192 | return store.getRole(clusterService.getLocalNode().id(), deviceId); |
| 193 | } |
| 194 | |
| 195 | @Override |
Madan Jampani | c6e574f | 2015-05-29 13:41:52 -0700 | [diff] [blame] | 196 | public CompletableFuture<Void> relinquishMastership(DeviceId deviceId) { |
Changhoon Yoon | b856b81 | 2015-08-10 03:47:19 +0900 | [diff] [blame] | 197 | checkPermission(CLUSTER_WRITE); |
Madan Jampani | c6e574f | 2015-05-29 13:41:52 -0700 | [diff] [blame] | 198 | return store.relinquishRole(localNodeId, deviceId) |
| 199 | .thenAccept(this::post) |
| 200 | .thenApply(v -> null); |
tom | b41d1ac | 2014-09-24 01:51:24 -0700 | [diff] [blame] | 201 | } |
| 202 | |
| 203 | @Override |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 204 | public CompletableFuture<MastershipRole> requestRoleFor(DeviceId deviceId) { |
Changhoon Yoon | b856b81 | 2015-08-10 03:47:19 +0900 | [diff] [blame] | 205 | checkPermission(CLUSTER_WRITE); |
Changhoon Yoon | 541ef71 | 2015-05-23 17:18:34 +0900 | [diff] [blame] | 206 | |
tom | b41d1ac | 2014-09-24 01:51:24 -0700 | [diff] [blame] | 207 | checkNotNull(deviceId, DEVICE_ID_NULL); |
Yuta HIGUCHI | 6a46283 | 2014-11-23 23:56:03 -0800 | [diff] [blame] | 208 | final Context timer = startTimer(requestRoleTimer); |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 209 | return store.requestRole(deviceId).whenComplete((result, error) -> stopTimer(timer)); |
| 210 | |
tom | b41d1ac | 2014-09-24 01:51:24 -0700 | [diff] [blame] | 211 | } |
| 212 | |
| 213 | @Override |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 214 | public NodeId getMasterFor(DeviceId deviceId) { |
Changhoon Yoon | b856b81 | 2015-08-10 03:47:19 +0900 | [diff] [blame] | 215 | checkPermission(CLUSTER_READ); |
Changhoon Yoon | 541ef71 | 2015-05-23 17:18:34 +0900 | [diff] [blame] | 216 | |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 217 | checkNotNull(deviceId, DEVICE_ID_NULL); |
| 218 | return store.getMaster(deviceId); |
| 219 | } |
| 220 | |
| 221 | @Override |
| 222 | public Set<DeviceId> getDevicesOf(NodeId nodeId) { |
Changhoon Yoon | b856b81 | 2015-08-10 03:47:19 +0900 | [diff] [blame] | 223 | checkPermission(CLUSTER_READ); |
Changhoon Yoon | 541ef71 | 2015-05-23 17:18:34 +0900 | [diff] [blame] | 224 | |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 225 | checkNotNull(nodeId, NODE_ID_NULL); |
| 226 | return store.getDevices(nodeId); |
| 227 | } |
| 228 | |
Ayaka Koshibe | 45503ce | 2014-10-14 11:26:45 -0700 | [diff] [blame] | 229 | @Override |
Ayaka Koshibe | abedb09 | 2014-10-20 17:01:31 -0700 | [diff] [blame] | 230 | public RoleInfo getNodesFor(DeviceId deviceId) { |
Changhoon Yoon | b856b81 | 2015-08-10 03:47:19 +0900 | [diff] [blame] | 231 | checkPermission(CLUSTER_READ); |
Changhoon Yoon | 541ef71 | 2015-05-23 17:18:34 +0900 | [diff] [blame] | 232 | |
Ayaka Koshibe | 45503ce | 2014-10-14 11:26:45 -0700 | [diff] [blame] | 233 | checkNotNull(deviceId, DEVICE_ID_NULL); |
| 234 | return store.getNodes(deviceId); |
| 235 | } |
Ayaka Koshibe | b70d34b | 2014-09-25 15:43:01 -0700 | [diff] [blame] | 236 | |
| 237 | @Override |
Yuta HIGUCHI | bcac499 | 2014-11-22 19:27:57 -0800 | [diff] [blame] | 238 | public MastershipTerm getMastershipTerm(DeviceId deviceId) { |
Heedo Kang | 4a47a30 | 2016-02-29 17:40:23 +0900 | [diff] [blame] | 239 | checkPermission(CLUSTER_READ); |
Yuta HIGUCHI | bcac499 | 2014-11-22 19:27:57 -0800 | [diff] [blame] | 240 | return store.getTermFor(deviceId); |
Ayaka Koshibe | b70d34b | 2014-09-25 15:43:01 -0700 | [diff] [blame] | 241 | } |
| 242 | |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 243 | @Override |
Yuta HIGUCHI | a22f69f | 2014-11-24 22:25:17 -0800 | [diff] [blame] | 244 | public MetricsService metricsService() { |
| 245 | return metricsService; |
| 246 | } |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 247 | |
Thomas Vachuska | 1e68bdd | 2014-11-29 13:53:10 -0800 | [diff] [blame] | 248 | @Override |
| 249 | public void balanceRoles() { |
| 250 | List<ControllerNode> nodes = newArrayList(clusterService.getNodes()); |
Thomas Vachuska | 12dfdc3 | 2014-11-29 16:03:12 -0800 | [diff] [blame] | 251 | Map<ControllerNode, Set<DeviceId>> controllerDevices = new HashMap<>(); |
Thomas Vachuska | 1e68bdd | 2014-11-29 13:53:10 -0800 | [diff] [blame] | 252 | int deviceCount = 0; |
| 253 | |
| 254 | // Create buckets reflecting current ownership. |
| 255 | for (ControllerNode node : nodes) { |
Thomas Vachuska | 7a8de84 | 2016-03-07 20:56:35 -0800 | [diff] [blame] | 256 | if (clusterService.getState(node.id()).isActive()) { |
Thomas Vachuska | 12dfdc3 | 2014-11-29 16:03:12 -0800 | [diff] [blame] | 257 | Set<DeviceId> devicesOf = new HashSet<>(getDevicesOf(node.id())); |
| 258 | deviceCount += devicesOf.size(); |
| 259 | controllerDevices.put(node, devicesOf); |
| 260 | log.info("Node {} has {} devices.", node.id(), devicesOf.size()); |
| 261 | } |
Thomas Vachuska | 1e68bdd | 2014-11-29 13:53:10 -0800 | [diff] [blame] | 262 | } |
| 263 | |
Claudine Chiu | dce0815 | 2016-03-09 18:19:28 +0000 | [diff] [blame] | 264 | if (useRegionForBalanceRoles && balanceRolesUsingRegions(controllerDevices)) { |
| 265 | return; |
| 266 | } |
| 267 | |
| 268 | // Now re-balance the buckets until they are roughly even. |
| 269 | List<CompletableFuture<Void>> balanceBucketsFutures = balanceControllerNodes(controllerDevices, deviceCount); |
| 270 | |
| 271 | CompletableFuture<Void> balanceRolesFuture = CompletableFuture.allOf( |
| 272 | balanceBucketsFutures.toArray(new CompletableFuture[balanceBucketsFutures.size()])); |
| 273 | |
| 274 | Futures.getUnchecked(balanceRolesFuture); |
| 275 | } |
| 276 | |
| 277 | /** |
| 278 | * Balances the nodes specified in controllerDevices. |
| 279 | * |
| 280 | * @param controllerDevices controller nodes to devices map |
| 281 | * @param deviceCount number of devices mastered by controller nodes |
| 282 | * @return list of setRole futures for "moved" devices |
| 283 | */ |
| 284 | private List<CompletableFuture<Void>> balanceControllerNodes( |
| 285 | Map<ControllerNode, Set<DeviceId>> controllerDevices, int deviceCount) { |
Thomas Vachuska | 12dfdc3 | 2014-11-29 16:03:12 -0800 | [diff] [blame] | 286 | // Now re-balance the buckets until they are roughly even. |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 287 | List<CompletableFuture<Void>> balanceBucketsFutures = Lists.newLinkedList(); |
Thomas Vachuska | 12dfdc3 | 2014-11-29 16:03:12 -0800 | [diff] [blame] | 288 | int rounds = controllerDevices.keySet().size(); |
Thomas Vachuska | 1e68bdd | 2014-11-29 13:53:10 -0800 | [diff] [blame] | 289 | for (int i = 0; i < rounds; i++) { |
| 290 | // Iterate over the buckets and find the smallest and the largest. |
Thomas Vachuska | 12dfdc3 | 2014-11-29 16:03:12 -0800 | [diff] [blame] | 291 | ControllerNode smallest = findBucket(true, controllerDevices); |
| 292 | ControllerNode largest = findBucket(false, controllerDevices); |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 293 | balanceBucketsFutures.add(balanceBuckets(smallest, largest, controllerDevices, deviceCount)); |
Thomas Vachuska | 1e68bdd | 2014-11-29 13:53:10 -0800 | [diff] [blame] | 294 | } |
Claudine Chiu | dce0815 | 2016-03-09 18:19:28 +0000 | [diff] [blame] | 295 | return balanceBucketsFutures; |
Thomas Vachuska | 1e68bdd | 2014-11-29 13:53:10 -0800 | [diff] [blame] | 296 | } |
| 297 | |
Claudine Chiu | dce0815 | 2016-03-09 18:19:28 +0000 | [diff] [blame] | 298 | /** |
| 299 | * Finds node with the minimum/maximum devices from a list of nodes. |
| 300 | * |
| 301 | * @param min true: minimum, false: maximum |
| 302 | * @param controllerDevices controller nodes to devices map |
| 303 | * @return controller node with minimum/maximum devices |
| 304 | */ |
Thomas Vachuska | 12dfdc3 | 2014-11-29 16:03:12 -0800 | [diff] [blame] | 305 | private ControllerNode findBucket(boolean min, |
| 306 | Map<ControllerNode, Set<DeviceId>> controllerDevices) { |
Thomas Vachuska | 1e68bdd | 2014-11-29 13:53:10 -0800 | [diff] [blame] | 307 | int xSize = min ? Integer.MAX_VALUE : -1; |
| 308 | ControllerNode xNode = null; |
Thomas Vachuska | 12dfdc3 | 2014-11-29 16:03:12 -0800 | [diff] [blame] | 309 | for (ControllerNode node : controllerDevices.keySet()) { |
Thomas Vachuska | 1e68bdd | 2014-11-29 13:53:10 -0800 | [diff] [blame] | 310 | int size = controllerDevices.get(node).size(); |
| 311 | if ((min && size < xSize) || (!min && size > xSize)) { |
| 312 | xSize = size; |
| 313 | xNode = node; |
| 314 | } |
| 315 | } |
| 316 | return xNode; |
| 317 | } |
| 318 | |
Claudine Chiu | dce0815 | 2016-03-09 18:19:28 +0000 | [diff] [blame] | 319 | /** |
| 320 | * Balance the node buckets by moving devices from largest to smallest node. |
| 321 | * |
| 322 | * @param smallest node that is master of the smallest number of devices |
| 323 | * @param largest node that is master of the largest number of devices |
| 324 | * @param controllerDevices controller nodes to devices map |
| 325 | * @param deviceCount number of devices mastered by controller nodes |
| 326 | * @return list of setRole futures for "moved" devices |
| 327 | */ |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 328 | private CompletableFuture<Void> balanceBuckets(ControllerNode smallest, ControllerNode largest, |
Thomas Vachuska | 12dfdc3 | 2014-11-29 16:03:12 -0800 | [diff] [blame] | 329 | Map<ControllerNode, Set<DeviceId>> controllerDevices, |
Thomas Vachuska | 1e68bdd | 2014-11-29 13:53:10 -0800 | [diff] [blame] | 330 | int deviceCount) { |
| 331 | Collection<DeviceId> minBucket = controllerDevices.get(smallest); |
| 332 | Collection<DeviceId> maxBucket = controllerDevices.get(largest); |
| 333 | int bucketCount = controllerDevices.keySet().size(); |
| 334 | |
| 335 | int delta = (maxBucket.size() - minBucket.size()) / 2; |
| 336 | delta = Math.min(deviceCount / bucketCount, delta); |
| 337 | |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 338 | List<CompletableFuture<Void>> setRoleFutures = Lists.newLinkedList(); |
| 339 | |
Thomas Vachuska | 1e68bdd | 2014-11-29 13:53:10 -0800 | [diff] [blame] | 340 | if (delta > 0) { |
| 341 | log.info("Attempting to move {} nodes from {} to {}...", delta, |
| 342 | largest.id(), smallest.id()); |
| 343 | |
| 344 | int i = 0; |
| 345 | Iterator<DeviceId> it = maxBucket.iterator(); |
| 346 | while (it.hasNext() && i < delta) { |
| 347 | DeviceId deviceId = it.next(); |
| 348 | log.info("Setting {} as the master for {}", smallest.id(), deviceId); |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 349 | setRoleFutures.add(setRole(smallest.id(), deviceId, MASTER)); |
Thomas Vachuska | 12dfdc3 | 2014-11-29 16:03:12 -0800 | [diff] [blame] | 350 | controllerDevices.get(smallest).add(deviceId); |
Thomas Vachuska | 1e68bdd | 2014-11-29 13:53:10 -0800 | [diff] [blame] | 351 | it.remove(); |
| 352 | i++; |
| 353 | } |
| 354 | } |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 355 | |
| 356 | return CompletableFuture.allOf(setRoleFutures.toArray(new CompletableFuture[setRoleFutures.size()])); |
Thomas Vachuska | 1e68bdd | 2014-11-29 13:53:10 -0800 | [diff] [blame] | 357 | } |
| 358 | |
Claudine Chiu | dce0815 | 2016-03-09 18:19:28 +0000 | [diff] [blame] | 359 | /** |
| 360 | * Balances the nodes considering Region information. |
| 361 | * |
| 362 | * @param allControllerDevices controller nodes to devices map |
| 363 | * @return true: nodes balanced; false: nodes not balanced |
| 364 | */ |
| 365 | private boolean balanceRolesUsingRegions(Map<ControllerNode, Set<DeviceId>> allControllerDevices) { |
| 366 | Set<Region> regions = regionService.getRegions(); |
| 367 | if (regions.isEmpty()) { |
| 368 | return false; // no balancing was done using regions. |
| 369 | } |
| 370 | |
| 371 | // handle nodes belonging to regions |
| 372 | Set<ControllerNode> nodesInRegions = Sets.newHashSet(); |
| 373 | for (Region region : regions) { |
| 374 | Map<ControllerNode, Set<DeviceId>> activeRegionControllers = |
| 375 | balanceRolesInRegion(region, allControllerDevices); |
| 376 | nodesInRegions.addAll(activeRegionControllers.keySet()); |
| 377 | } |
| 378 | |
| 379 | // handle nodes not belonging to any region |
| 380 | Set<ControllerNode> nodesNotInRegions = Sets.difference(allControllerDevices.keySet(), nodesInRegions); |
| 381 | if (!nodesNotInRegions.isEmpty()) { |
| 382 | int deviceCount = 0; |
| 383 | Map<ControllerNode, Set<DeviceId>> controllerDevicesNotInRegions = new HashMap<>(); |
| 384 | for (ControllerNode controllerNode: nodesNotInRegions) { |
| 385 | controllerDevicesNotInRegions.put(controllerNode, allControllerDevices.get(controllerNode)); |
| 386 | deviceCount += allControllerDevices.get(controllerNode).size(); |
| 387 | } |
| 388 | // Now re-balance the buckets until they are roughly even. |
| 389 | List<CompletableFuture<Void>> balanceBucketsFutures = |
| 390 | balanceControllerNodes(controllerDevicesNotInRegions, deviceCount); |
| 391 | |
| 392 | CompletableFuture<Void> balanceRolesFuture = CompletableFuture.allOf( |
| 393 | balanceBucketsFutures.toArray(new CompletableFuture[balanceBucketsFutures.size()])); |
| 394 | |
| 395 | Futures.getUnchecked(balanceRolesFuture); |
| 396 | } |
| 397 | return true; // balancing was done using regions. |
| 398 | } |
| 399 | |
| 400 | /** |
| 401 | * Balances the nodes in specified region. |
| 402 | * |
| 403 | * @param region region in which nodes are to be balanced |
| 404 | * @param allControllerDevices controller nodes to devices map |
| 405 | * @return controller nodes that were balanced |
| 406 | */ |
| 407 | private Map<ControllerNode, Set<DeviceId>> balanceRolesInRegion(Region region, |
| 408 | Map<ControllerNode, Set<DeviceId>> allControllerDevices) { |
| 409 | |
| 410 | // retrieve all devices associated with specified region |
| 411 | Set<DeviceId> devicesInRegion = regionService.getRegionDevices(region.id()); |
| 412 | log.info("Region {} has {} devices.", region.id(), devicesInRegion.size()); |
| 413 | if (devicesInRegion.isEmpty()) { |
| 414 | return new HashMap<>(); // no devices in this region, so nothing to balance. |
| 415 | } |
| 416 | |
| 417 | List<Set<NodeId>> mastersList = region.masters(); |
| 418 | log.info("Region {} has {} sets of masters.", region.id(), mastersList.size()); |
| 419 | if (mastersList.isEmpty()) { |
| 420 | // TODO handle devices that belong to a region, which has no masters defined |
| 421 | return new HashMap<>(); // for now just leave devices alone |
| 422 | } |
| 423 | |
| 424 | // get the region's preferred set of masters |
| 425 | Set<DeviceId> devicesInMasters = Sets.newHashSet(); |
| 426 | Map<ControllerNode, Set<DeviceId>> regionalControllerDevices = |
| 427 | getRegionsPreferredMasters(region, devicesInMasters, allControllerDevices); |
| 428 | |
| 429 | // Now re-balance the buckets until they are roughly even. |
| 430 | List<CompletableFuture<Void>> balanceBucketsFutures = |
| 431 | balanceControllerNodes(regionalControllerDevices, devicesInMasters.size()); |
| 432 | |
| 433 | // handle devices that are not currently mastered by the master node set |
| 434 | Set<DeviceId> devicesNotMasteredWithControllers = Sets.difference(devicesInRegion, devicesInMasters); |
| 435 | if (!devicesNotMasteredWithControllers.isEmpty()) { |
| 436 | // active controllers in master node set are already balanced, just |
| 437 | // assign device mastership in sequence |
| 438 | List<ControllerNode> sorted = new ArrayList<>(regionalControllerDevices.keySet()); |
| 439 | Collections.sort(sorted, (o1, o2) -> |
| 440 | ((Integer) (regionalControllerDevices.get(o1)).size()) |
| 441 | .compareTo((Integer) (regionalControllerDevices.get(o2)).size())); |
| 442 | int deviceIndex = 0; |
| 443 | for (DeviceId deviceId : devicesNotMasteredWithControllers) { |
| 444 | ControllerNode cnode = sorted.get(deviceIndex % sorted.size()); |
| 445 | balanceBucketsFutures.add(setRole(cnode.id(), deviceId, MASTER)); |
| 446 | regionalControllerDevices.get(cnode).add(deviceId); |
| 447 | deviceIndex++; |
| 448 | } |
| 449 | } |
| 450 | |
| 451 | CompletableFuture<Void> balanceRolesFuture = CompletableFuture.allOf( |
| 452 | balanceBucketsFutures.toArray(new CompletableFuture[balanceBucketsFutures.size()])); |
| 453 | |
| 454 | Futures.getUnchecked(balanceRolesFuture); |
| 455 | |
| 456 | // update the map before returning |
| 457 | regionalControllerDevices.forEach((controllerNode, deviceIds) -> { |
| 458 | regionalControllerDevices.put(controllerNode, new HashSet<>(getDevicesOf(controllerNode.id()))); |
| 459 | }); |
| 460 | |
| 461 | return regionalControllerDevices; |
| 462 | } |
| 463 | |
| 464 | /** |
| 465 | * Get region's preferred set of master nodes - the first master node set that has at |
| 466 | * least one active node. |
| 467 | * |
| 468 | * @param region region for which preferred set of master nodes is requested |
| 469 | * @param devicesInMasters device set to track devices in preferred set of master nodes |
| 470 | * @param allControllerDevices controller nodes to devices map |
| 471 | * @return region's preferred master nodes (and devices that use them as masters) |
| 472 | */ |
| 473 | private Map<ControllerNode, Set<DeviceId>> getRegionsPreferredMasters(Region region, |
| 474 | Set<DeviceId> devicesInMasters, |
| 475 | Map<ControllerNode, Set<DeviceId>> allControllerDevices) { |
| 476 | Map<ControllerNode, Set<DeviceId>> regionalControllerDevices = new HashMap<>(); |
| 477 | int listIndex = 0; |
| 478 | for (Set<NodeId> masterSet: region.masters()) { |
| 479 | log.info("Region {} masters set {} has {} nodes.", |
| 480 | region.id(), listIndex, masterSet.size()); |
| 481 | if (masterSet.isEmpty()) { // nothing on this level |
| 482 | listIndex++; |
| 483 | continue; |
| 484 | } |
| 485 | // Create buckets reflecting current ownership. |
| 486 | for (NodeId nodeId : masterSet) { |
Michele Santuari | 6ebb36e | 2016-03-28 10:12:04 -0700 | [diff] [blame] | 487 | if (clusterService.getState(nodeId).isActive()) { |
Claudine Chiu | dce0815 | 2016-03-09 18:19:28 +0000 | [diff] [blame] | 488 | ControllerNode controllerNode = clusterService.getNode(nodeId); |
| 489 | Set<DeviceId> devicesOf = new HashSet<>(allControllerDevices.get(controllerNode)); |
| 490 | regionalControllerDevices.put(controllerNode, devicesOf); |
| 491 | devicesInMasters.addAll(devicesOf); |
| 492 | log.info("Active Node {} has {} devices.", nodeId, devicesOf.size()); |
| 493 | } |
| 494 | } |
| 495 | if (!regionalControllerDevices.isEmpty()) { |
| 496 | break; // now have a set of >0 active controllers |
| 497 | } |
| 498 | listIndex++; // keep on looking |
| 499 | } |
| 500 | return regionalControllerDevices; |
| 501 | } |
Thomas Vachuska | 1e68bdd | 2014-11-29 13:53:10 -0800 | [diff] [blame] | 502 | |
alshabib | 339a3d9 | 2014-09-26 17:54:32 -0700 | [diff] [blame] | 503 | public class InternalDelegate implements MastershipStoreDelegate { |
alshabib | 339a3d9 | 2014-09-26 17:54:32 -0700 | [diff] [blame] | 504 | @Override |
| 505 | public void notify(MastershipEvent event) { |
Thomas Vachuska | 42e8cce | 2015-07-29 19:25:18 -0700 | [diff] [blame] | 506 | post(event); |
alshabib | 339a3d9 | 2014-09-26 17:54:32 -0700 | [diff] [blame] | 507 | } |
alshabib | 339a3d9 | 2014-09-26 17:54:32 -0700 | [diff] [blame] | 508 | } |
| 509 | |
Jordan Halterman | 713830d | 2017-10-07 13:40:44 -0700 | [diff] [blame] | 510 | private class InternalUpgradeEventListener implements UpgradeEventListener { |
| 511 | @Override |
| 512 | public void event(UpgradeEvent event) { |
| 513 | if (rebalanceRolesOnUpgrade && |
| 514 | (event.type() == UpgradeEvent.Type.COMMITTED || event.type() == UpgradeEvent.Type.RESET)) { |
| 515 | balanceRoles(); |
| 516 | } |
| 517 | } |
| 518 | } |
| 519 | |
Ayaka Koshibe | 1660969 | 2014-09-23 12:46:15 -0700 | [diff] [blame] | 520 | } |