private void sendPartitions() {
ClusterNode oldestNode = this.oldestNode.get();
try {
sendLocalPartitions(oldestNode, exchId);
} catch (ClusterTopologyCheckedException ignore) {
if (log.isDebugEnabled())
log.debug(
"Oldest node left during partition exchange [nodeId="
+ oldestNode.id()
+ ", exchId="
+ exchId
+ ']');
} catch (IgniteCheckedException e) {
scheduleRecheck();
U.error(
log,
"Failed to send local partitions to oldest node (will retry after timeout) [oldestNodeId="
+ oldestNode.id()
+ ", exchId="
+ exchId
+ ']',
e);
}
}
private void recheck() {
// If this is the oldest node.
if (oldestNode.get().id().equals(cctx.localNodeId())) {
Collection<UUID> remaining = remaining();
if (!remaining.isEmpty()) {
try {
cctx.io()
.safeSend(
cctx.discovery().nodes(remaining),
new GridDhtPartitionsSingleRequest(exchId),
SYSTEM_POOL,
null);
} catch (IgniteCheckedException e) {
U.error(
log,
"Failed to request partitions from nodes [exchangeId="
+ exchId
+ ", nodes="
+ remaining
+ ']',
e);
}
}
// Resend full partition map because last attempt failed.
else {
if (spreadPartitions()) onDone(exchId.topologyVersion());
}
} else sendPartitions();
// Schedule another send.
scheduleRecheck();
}
/** Clears swap entries for evicted partition. */
private void clearSwap() {
assert state() == EVICTED;
assert !GridQueryProcessor.isEnabled(cctx.config())
: "Indexing needs to have unswapped values.";
try {
GridCloseableIterator<Map.Entry<byte[], GridCacheSwapEntry>> it = cctx.swap().iterator(id);
boolean isLocStore = cctx.store().isLocal();
if (it != null) {
// We can safely remove these values because no entries will be created for evicted
// partition.
while (it.hasNext()) {
Map.Entry<byte[], GridCacheSwapEntry> entry = it.next();
byte[] keyBytes = entry.getKey();
KeyCacheObject key = cctx.toCacheKeyObject(keyBytes);
cctx.swap().remove(key);
if (isLocStore) cctx.store().remove(null, key.value(cctx.cacheObjectContext(), false));
}
}
} catch (IgniteCheckedException e) {
U.error(log, "Failed to clear swap for evicted partition: " + this, e);
}
}
/**
* @param nodeId Node ID.
* @param retryCnt Number of retries.
*/
private void sendAllPartitions(final UUID nodeId, final int retryCnt) {
ClusterNode n = cctx.node(nodeId);
try {
if (n != null) sendAllPartitions(F.asList(n), exchId);
} catch (IgniteCheckedException e) {
if (e instanceof ClusterTopologyCheckedException || !cctx.discovery().alive(n)) {
log.debug(
"Failed to send full partition map to node, node left grid "
+ "[rmtNode="
+ nodeId
+ ", exchangeId="
+ exchId
+ ']');
return;
}
if (retryCnt > 0) {
long timeout = cctx.gridConfig().getNetworkSendRetryDelay();
LT.error(
log,
e,
"Failed to send full partition map to node (will retry after timeout) "
+ "[node="
+ nodeId
+ ", exchangeId="
+ exchId
+ ", timeout="
+ timeout
+ ']');
cctx.time()
.addTimeoutObject(
new GridTimeoutObjectAdapter(timeout) {
@Override
public void onTimeout() {
sendAllPartitions(nodeId, retryCnt - 1);
}
});
} else
U.error(
log,
"Failed to send full partition map [node=" + n + ", exchangeId=" + exchId + ']',
e);
}
}
/** @return {@code True} if succeeded. */
private boolean spreadPartitions() {
try {
sendAllPartitions(rmtNodes, exchId);
return true;
} catch (IgniteCheckedException e) {
scheduleRecheck();
if (!X.hasCause(e, InterruptedException.class))
U.error(
log,
"Failed to send full partition map to nodes (will retry after timeout) [nodes="
+ F.nodeId8s(rmtNodes)
+ ", exchangeId="
+ exchId
+ ']',
e);
return false;
}
}
/**
* 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;
}
/** Clears values for this partition. */
private void clearAll() {
GridCacheVersion clearVer = cctx.versions().next();
boolean swap = cctx.isSwapOrOffheapEnabled();
boolean rec = cctx.events().isRecordable(EVT_CACHE_REBALANCE_OBJECT_UNLOADED);
Iterator<GridDhtCacheEntry> it = map.values().iterator();
GridCloseableIterator<Map.Entry<byte[], GridCacheSwapEntry>> swapIt = null;
if (swap
&& GridQueryProcessor.isEnabled(cctx.config())) { // Indexing needs to unswap cache values.
Iterator<GridDhtCacheEntry> unswapIt = null;
try {
swapIt = cctx.swap().iterator(id);
unswapIt = unswapIterator(swapIt);
} catch (Exception e) {
U.error(log, "Failed to clear swap for evicted partition: " + this, e);
}
if (unswapIt != null) it = F.concat(it, unswapIt);
}
try {
while (it.hasNext()) {
GridDhtCacheEntry cached = it.next();
try {
if (cached.clearInternal(clearVer, swap)) {
map.remove(cached.key(), cached);
if (!cached.isInternal()) {
mapPubSize.decrement();
if (rec)
cctx.events()
.addEvent(
cached.partition(),
cached.key(),
cctx.localNodeId(),
(IgniteUuid) null,
null,
EVT_CACHE_REBALANCE_OBJECT_UNLOADED,
null,
false,
cached.rawGet(),
cached.hasValue(),
null,
null,
null);
}
}
} catch (IgniteCheckedException e) {
U.error(log, "Failed to clear cache entry for evicted partition: " + cached, e);
}
}
} finally {
U.close(swapIt, log);
}
}