/** {@inheritDoc} */
@Override
public Collection<ClusterNode> nodes(int p, AffinityTopologyVersion topVer) {
Collection<ClusterNode> affNodes = cctx.affinity().nodes(p, topVer);
lock.readLock().lock();
try {
assert node2part != null && node2part.valid()
: "Invalid node-to-partitions map [topVer1="
+ topVer
+ ", topVer2="
+ this.topVer
+ ", cache="
+ cctx.name()
+ ", node2part="
+ node2part
+ ']';
Collection<ClusterNode> nodes = null;
Collection<UUID> nodeIds = part2node.get(p);
if (!F.isEmpty(nodeIds)) {
Collection<UUID> affIds = new HashSet<>(F.viewReadOnly(affNodes, F.node2id()));
for (UUID nodeId : nodeIds) {
if (!affIds.contains(nodeId) && hasState(p, nodeId, OWNING, MOVING, RENTING)) {
ClusterNode n = cctx.discovery().node(nodeId);
if (n != null
&& (topVer.topologyVersion() < 0 || n.order() <= topVer.topologyVersion())) {
if (nodes == null) {
nodes = new ArrayList<>(affNodes.size() + 2);
nodes.addAll(affNodes);
}
nodes.add(n);
}
}
}
}
return nodes != null ? nodes : affNodes;
} finally {
lock.readLock().unlock();
}
}
/**
* @param p Partition.
* @param topVer Topology version ({@code -1} for all nodes).
* @param state Partition state.
* @param states Additional partition states.
* @return List of nodes for the partition.
*/
private List<ClusterNode> nodes(
int p,
AffinityTopologyVersion topVer,
GridDhtPartitionState state,
GridDhtPartitionState... states) {
Collection<UUID> allIds =
topVer.topologyVersion() > 0 ? F.nodeIds(CU.affinityNodes(cctx, topVer)) : null;
lock.readLock().lock();
try {
assert node2part != null && node2part.valid()
: "Invalid node-to-partitions map [topVer="
+ topVer
+ ", allIds="
+ allIds
+ ", node2part="
+ node2part
+ ", cache="
+ cctx.name()
+ ']';
Collection<UUID> nodeIds = part2node.get(p);
// Node IDs can be null if both, primary and backup, nodes disappear.
int size = nodeIds == null ? 0 : nodeIds.size();
if (size == 0) return Collections.emptyList();
List<ClusterNode> nodes = new ArrayList<>(size);
for (UUID id : nodeIds) {
if (topVer.topologyVersion() > 0 && !allIds.contains(id)) continue;
if (hasState(p, id, state, states)) {
ClusterNode n = cctx.discovery().node(id);
if (n != null && (topVer.topologyVersion() < 0 || n.order() <= topVer.topologyVersion()))
nodes.add(n);
}
}
return nodes;
} finally {
lock.readLock().unlock();
}
}
/**
* Starts activity.
*
* @throws IgniteInterruptedCheckedException If interrupted.
*/
public void init() throws IgniteInterruptedCheckedException {
if (isDone()) return;
if (init.compareAndSet(false, true)) {
if (isDone()) return;
try {
// Wait for event to occur to make sure that discovery
// will return corresponding nodes.
U.await(evtLatch);
assert discoEvt != null : this;
assert !dummy && !forcePreload : this;
ClusterNode oldest = CU.oldestAliveCacheServerNode(cctx, exchId.topologyVersion());
oldestNode.set(oldest);
startCaches();
// True if client node joined or failed.
boolean clientNodeEvt;
if (F.isEmpty(reqs)) {
int type = discoEvt.type();
assert type == EVT_NODE_JOINED || type == EVT_NODE_LEFT || type == EVT_NODE_FAILED
: discoEvt;
clientNodeEvt = CU.clientNode(discoEvt.eventNode());
} else {
assert discoEvt.type() == EVT_DISCOVERY_CUSTOM_EVT : discoEvt;
boolean clientOnlyStart = true;
for (DynamicCacheChangeRequest req : reqs) {
if (!req.clientStartOnly()) {
clientOnlyStart = false;
break;
}
}
clientNodeEvt = clientOnlyStart;
}
if (clientNodeEvt) {
ClusterNode node = discoEvt.eventNode();
// Client need to initialize affinity for local join event or for stated client caches.
if (!node.isLocal()) {
for (GridCacheContext cacheCtx : cctx.cacheContexts()) {
if (cacheCtx.isLocal()) continue;
GridDhtPartitionTopology top = cacheCtx.topology();
top.updateTopologyVersion(exchId, this, -1, stopping(cacheCtx.cacheId()));
if (cacheCtx.affinity().affinityTopologyVersion() == AffinityTopologyVersion.NONE) {
initTopology(cacheCtx);
top.beforeExchange(this);
} else
cacheCtx.affinity().clientEventTopologyChange(discoEvt, exchId.topologyVersion());
}
if (exchId.isLeft())
cctx.mvcc().removeExplicitNodeLocks(exchId.nodeId(), exchId.topologyVersion());
onDone(exchId.topologyVersion());
skipPreload = cctx.kernalContext().clientNode();
return;
}
}
if (cctx.kernalContext().clientNode()) {
skipPreload = true;
for (GridCacheContext cacheCtx : cctx.cacheContexts()) {
if (cacheCtx.isLocal()) continue;
GridDhtPartitionTopology top = cacheCtx.topology();
top.updateTopologyVersion(exchId, this, -1, stopping(cacheCtx.cacheId()));
}
for (GridCacheContext cacheCtx : cctx.cacheContexts()) {
if (cacheCtx.isLocal()) continue;
initTopology(cacheCtx);
}
if (oldestNode.get() != null) {
rmtNodes =
new ConcurrentLinkedQueue<>(
CU.aliveRemoteServerNodesWithCaches(cctx, exchId.topologyVersion()));
rmtIds = Collections.unmodifiableSet(new HashSet<>(F.nodeIds(rmtNodes)));
ready.set(true);
initFut.onDone(true);
if (log.isDebugEnabled()) log.debug("Initialized future: " + this);
sendPartitions();
} else onDone(exchId.topologyVersion());
return;
}
assert oldestNode.get() != null;
for (GridCacheContext cacheCtx : cctx.cacheContexts()) {
if (isCacheAdded(cacheCtx.cacheId(), exchId.topologyVersion())) {
if (cacheCtx
.discovery()
.cacheAffinityNodes(cacheCtx.name(), topologyVersion())
.isEmpty())
U.quietAndWarn(log, "No server nodes found for cache client: " + cacheCtx.namex());
}
cacheCtx.preloader().onExchangeFutureAdded();
}
List<String> cachesWithoutNodes = null;
if (exchId.isLeft()) {
for (String name : cctx.cache().cacheNames()) {
if (cctx.discovery().cacheAffinityNodes(name, topologyVersion()).isEmpty()) {
if (cachesWithoutNodes == null) cachesWithoutNodes = new ArrayList<>();
cachesWithoutNodes.add(name);
// Fire event even if there is no client cache started.
if (cctx.gridEvents().isRecordable(EventType.EVT_CACHE_NODES_LEFT)) {
Event evt =
new CacheEvent(
name,
cctx.localNode(),
cctx.localNode(),
"All server nodes have left the cluster.",
EventType.EVT_CACHE_NODES_LEFT,
0,
false,
null,
null,
null,
null,
false,
null,
false,
null,
null,
null);
cctx.gridEvents().record(evt);
}
}
}
}
if (cachesWithoutNodes != null) {
StringBuilder sb =
new StringBuilder(
"All server nodes for the following caches have left the cluster: ");
for (int i = 0; i < cachesWithoutNodes.size(); i++) {
String cache = cachesWithoutNodes.get(i);
sb.append('\'').append(cache).append('\'');
if (i != cachesWithoutNodes.size() - 1) sb.append(", ");
}
U.quietAndWarn(log, sb.toString());
U.quietAndWarn(log, "Must have server nodes for caches to operate.");
}
assert discoEvt != null;
assert exchId.nodeId().equals(discoEvt.eventNode().id());
for (GridCacheContext cacheCtx : cctx.cacheContexts()) {
GridClientPartitionTopology clientTop =
cctx.exchange().clearClientTopology(cacheCtx.cacheId());
long updSeq = clientTop == null ? -1 : clientTop.lastUpdateSequence();
// Update before waiting for locks.
if (!cacheCtx.isLocal())
cacheCtx
.topology()
.updateTopologyVersion(exchId, this, updSeq, stopping(cacheCtx.cacheId()));
}
// Grab all alive remote nodes with order of equal or less than last joined node.
rmtNodes =
new ConcurrentLinkedQueue<>(
CU.aliveRemoteServerNodesWithCaches(cctx, exchId.topologyVersion()));
rmtIds = Collections.unmodifiableSet(new HashSet<>(F.nodeIds(rmtNodes)));
for (Map.Entry<UUID, GridDhtPartitionsSingleMessage> m : singleMsgs.entrySet())
// If received any messages, process them.
onReceive(m.getKey(), m.getValue());
for (Map.Entry<UUID, GridDhtPartitionsFullMessage> m : fullMsgs.entrySet())
// If received any messages, process them.
onReceive(m.getKey(), m.getValue());
AffinityTopologyVersion topVer = exchId.topologyVersion();
for (GridCacheContext cacheCtx : cctx.cacheContexts()) {
if (cacheCtx.isLocal()) continue;
// Must initialize topology after we get discovery event.
initTopology(cacheCtx);
cacheCtx.preloader().updateLastExchangeFuture(this);
}
IgniteInternalFuture<?> partReleaseFut = cctx.partitionReleaseFuture(topVer);
// Assign to class variable so it will be included into toString() method.
this.partReleaseFut = partReleaseFut;
if (log.isDebugEnabled()) log.debug("Before waiting for partition release future: " + this);
while (true) {
try {
partReleaseFut.get(2 * cctx.gridConfig().getNetworkTimeout(), TimeUnit.MILLISECONDS);
break;
} catch (IgniteFutureTimeoutCheckedException ignored) {
// Print pending transactions and locks that might have led to hang.
dumpPendingObjects();
}
}
if (log.isDebugEnabled()) log.debug("After waiting for partition release future: " + this);
if (!F.isEmpty(reqs)) blockGateways();
if (exchId.isLeft())
cctx.mvcc().removeExplicitNodeLocks(exchId.nodeId(), exchId.topologyVersion());
IgniteInternalFuture<?> locksFut = cctx.mvcc().finishLocks(exchId.topologyVersion());
while (true) {
try {
locksFut.get(2 * cctx.gridConfig().getNetworkTimeout(), TimeUnit.MILLISECONDS);
break;
} catch (IgniteFutureTimeoutCheckedException ignored) {
U.warn(
log,
"Failed to wait for locks release future. "
+ "Dumping pending objects that might be the cause: "
+ cctx.localNodeId());
U.warn(log, "Locked entries:");
Map<IgniteTxKey, Collection<GridCacheMvccCandidate>> locks =
cctx.mvcc().unfinishedLocks(exchId.topologyVersion());
for (Map.Entry<IgniteTxKey, Collection<GridCacheMvccCandidate>> e : locks.entrySet())
U.warn(log, "Locked entry [key=" + e.getKey() + ", mvcc=" + e.getValue() + ']');
}
}
for (GridCacheContext cacheCtx : cctx.cacheContexts()) {
if (cacheCtx.isLocal()) continue;
// Notify replication manager.
GridCacheContext drCacheCtx =
cacheCtx.isNear() ? cacheCtx.near().dht().context() : cacheCtx;
if (drCacheCtx.isDrEnabled()) drCacheCtx.dr().beforeExchange(topVer, exchId.isLeft());
// Partition release future is done so we can flush the write-behind store.
cacheCtx.store().forceFlush();
// Process queued undeploys prior to sending/spreading map.
cacheCtx.preloader().unwindUndeploys();
GridDhtPartitionTopology top = cacheCtx.topology();
assert topVer.equals(top.topologyVersion())
: "Topology version is updated only in this class instances inside single ExchangeWorker thread.";
top.beforeExchange(this);
}
for (GridClientPartitionTopology top : cctx.exchange().clientTopologies()) {
top.updateTopologyVersion(exchId, this, -1, stopping(top.cacheId()));
top.beforeExchange(this);
}
} catch (IgniteInterruptedCheckedException e) {
onDone(e);
throw e;
} catch (Throwable e) {
U.error(
log,
"Failed to reinitialize local partitions (preloading will be stopped): " + exchId,
e);
onDone(e);
if (e instanceof Error) throw (Error) e;
return;
}
if (F.isEmpty(rmtIds)) {
onDone(exchId.topologyVersion());
return;
}
ready.set(true);
initFut.onDone(true);
if (log.isDebugEnabled()) log.debug("Initialized future: " + this);
// If this node is not oldest.
if (!oldestNode.get().id().equals(cctx.localNodeId())) sendPartitions();
else {
boolean allReceived = allReceived();
if (allReceived && replied.compareAndSet(false, true)) {
if (spreadPartitions()) onDone(exchId.topologyVersion());
}
}
scheduleRecheck();
} else assert false : "Skipped init future: " + this;
}
/** {@inheritDoc} */
@SuppressWarnings({"MismatchedQueryAndUpdateOfCollection"})
@Nullable
@Override
public GridDhtPartitionMap update(
@Nullable GridDhtPartitionExchangeId exchId, GridDhtPartitionFullMap partMap) {
if (log.isDebugEnabled())
log.debug(
"Updating full partition map [exchId=" + exchId + ", parts=" + fullMapString() + ']');
assert partMap != null;
lock.writeLock().lock();
try {
if (stopping) return null;
if (exchId != null && lastExchangeId != null && lastExchangeId.compareTo(exchId) >= 0) {
if (log.isDebugEnabled())
log.debug(
"Stale exchange id for full partition map update (will ignore) [lastExchId="
+ lastExchangeId
+ ", exchId="
+ exchId
+ ']');
return null;
}
if (node2part != null && node2part.compareTo(partMap) >= 0) {
if (log.isDebugEnabled())
log.debug(
"Stale partition map for full partition map update (will ignore) [lastExchId="
+ lastExchangeId
+ ", exchId="
+ exchId
+ ", curMap="
+ node2part
+ ", newMap="
+ partMap
+ ']');
return null;
}
long updateSeq = this.updateSeq.incrementAndGet();
if (exchId != null) lastExchangeId = exchId;
if (node2part != null) {
for (GridDhtPartitionMap part : node2part.values()) {
GridDhtPartitionMap newPart = partMap.get(part.nodeId());
// If for some nodes current partition has a newer map,
// then we keep the newer value.
if (newPart != null && newPart.updateSequence() < part.updateSequence()) {
if (log.isDebugEnabled())
log.debug(
"Overriding partition map in full update map [exchId="
+ exchId
+ ", curPart="
+ mapString(part)
+ ", newPart="
+ mapString(newPart)
+ ']');
partMap.put(part.nodeId(), part);
}
}
for (Iterator<UUID> it = partMap.keySet().iterator(); it.hasNext(); ) {
UUID nodeId = it.next();
if (!cctx.discovery().alive(nodeId)) {
if (log.isDebugEnabled())
log.debug(
"Removing left node from full map update [nodeId="
+ nodeId
+ ", partMap="
+ partMap
+ ']');
it.remove();
}
}
}
node2part = partMap;
Map<Integer, Set<UUID>> p2n = new HashMap<>(cctx.affinity().partitions(), 1.0f);
for (Map.Entry<UUID, GridDhtPartitionMap> e : partMap.entrySet()) {
for (Integer p : e.getValue().keySet()) {
Set<UUID> ids = p2n.get(p);
if (ids == null)
// Initialize HashSet to size 3 in anticipation that there won't be
// more than 3 nodes per partitions.
p2n.put(p, ids = U.newHashSet(3));
ids.add(e.getKey());
}
}
part2node = p2n;
boolean changed = checkEvictions(updateSeq);
consistencyCheck();
if (log.isDebugEnabled()) log.debug("Partition map after full update: " + fullMapString());
return changed ? localPartitionMap() : null;
} finally {
lock.writeLock().unlock();
}
}
/**
* @param updateSeq Update sequence.
* @return Checks if any of the local partitions need to be evicted.
*/
private boolean checkEvictions(long updateSeq) {
assert lock.isWriteLockedByCurrentThread();
boolean changed = false;
UUID locId = cctx.nodeId();
for (GridDhtLocalPartition part : locParts.values()) {
GridDhtPartitionState state = part.state();
if (state.active()) {
int p = part.id();
List<ClusterNode> affNodes = cctx.affinity().nodes(p, topVer);
if (!affNodes.contains(cctx.localNode())) {
Collection<UUID> nodeIds = F.nodeIds(nodes(p, topVer, OWNING));
// If all affinity nodes are owners, then evict partition from local node.
if (nodeIds.containsAll(F.nodeIds(affNodes))) {
part.rent(false);
updateLocal(part.id(), locId, part.state(), updateSeq);
changed = true;
if (log.isDebugEnabled())
log.debug("Evicted local partition (all affinity nodes are owners): " + part);
} else {
int ownerCnt = nodeIds.size();
int affCnt = affNodes.size();
if (ownerCnt > affCnt) {
List<ClusterNode> sorted = new ArrayList<>(cctx.discovery().nodes(nodeIds));
// Sort by node orders in ascending order.
Collections.sort(sorted, CU.nodeComparator(true));
int diff = sorted.size() - affCnt;
for (int i = 0; i < diff; i++) {
ClusterNode n = sorted.get(i);
if (locId.equals(n.id())) {
part.rent(false);
updateLocal(part.id(), locId, part.state(), updateSeq);
changed = true;
if (log.isDebugEnabled())
log.debug(
"Evicted local partition (this node is oldest non-affinity node): " + part);
break;
}
}
}
}
}
}
}
return changed;
}
/** {@inheritDoc} */
@SuppressWarnings({"MismatchedQueryAndUpdateOfCollection"})
@Nullable
@Override
public GridDhtPartitionMap update(
@Nullable GridDhtPartitionExchangeId exchId, GridDhtPartitionMap parts) {
if (log.isDebugEnabled())
log.debug(
"Updating single partition map [exchId=" + exchId + ", parts=" + mapString(parts) + ']');
if (!cctx.discovery().alive(parts.nodeId())) {
if (log.isDebugEnabled())
log.debug(
"Received partition update for non-existing node (will ignore) [exchId="
+ exchId
+ ", parts="
+ parts
+ ']');
return null;
}
lock.writeLock().lock();
try {
if (stopping) return null;
if (lastExchangeId != null && exchId != null && lastExchangeId.compareTo(exchId) > 0) {
if (log.isDebugEnabled())
log.debug(
"Stale exchange id for single partition map update (will ignore) [lastExchId="
+ lastExchangeId
+ ", exchId="
+ exchId
+ ']');
return null;
}
if (exchId != null) lastExchangeId = exchId;
if (node2part == null)
// Create invalid partition map.
node2part = new GridDhtPartitionFullMap();
GridDhtPartitionMap cur = node2part.get(parts.nodeId());
if (cur != null && cur.updateSequence() >= parts.updateSequence()) {
if (log.isDebugEnabled())
log.debug(
"Stale update sequence for single partition map update (will ignore) [exchId="
+ exchId
+ ", curSeq="
+ cur.updateSequence()
+ ", newSeq="
+ parts.updateSequence()
+ ']');
return null;
}
long updateSeq = this.updateSeq.incrementAndGet();
node2part = new GridDhtPartitionFullMap(node2part, updateSeq);
boolean changed = false;
if (cur == null || !cur.equals(parts)) changed = true;
node2part.put(parts.nodeId(), parts);
part2node = new HashMap<>(part2node);
// Add new mappings.
for (Integer p : parts.keySet()) {
Set<UUID> ids = part2node.get(p);
if (ids == null)
// Initialize HashSet to size 3 in anticipation that there won't be
// more than 3 nodes per partition.
part2node.put(p, ids = U.newHashSet(3));
changed |= ids.add(parts.nodeId());
}
// Remove obsolete mappings.
if (cur != null) {
for (Integer p : F.view(cur.keySet(), F0.notIn(parts.keySet()))) {
Set<UUID> ids = part2node.get(p);
if (ids != null) changed |= ids.remove(parts.nodeId());
}
}
changed |= checkEvictions(updateSeq);
consistencyCheck();
if (log.isDebugEnabled()) log.debug("Partition map after single update: " + fullMapString());
return changed ? localPartitionMap() : null;
} finally {
lock.writeLock().unlock();
}
}