/** {@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();
}
}
/** {@inheritDoc} */
@Override
public GridDhtPartitionMap localPartitionMap() {
lock.readLock().lock();
try {
return new GridDhtPartitionMap(
cctx.nodeId(), updateSeq.get(), F.viewReadOnly(locParts, CU.part2state()), true);
} 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();
}
}
/** 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);
}
}
/** {@inheritDoc} */
@Override
public boolean afterExchange(GridDhtPartitionsExchangeFuture exchFut)
throws IgniteCheckedException {
boolean changed = waitForRent();
ClusterNode loc = cctx.localNode();
int num = cctx.affinity().partitions();
AffinityTopologyVersion topVer = exchFut.topologyVersion();
lock.writeLock().lock();
try {
if (stopping) return false;
assert topVer.equals(exchFut.topologyVersion())
: "Invalid topology version [topVer=" + topVer + ", exchId=" + exchFut.exchangeId() + ']';
if (log.isDebugEnabled())
log.debug(
"Partition map before afterExchange [exchId="
+ exchFut.exchangeId()
+ ", fullMap="
+ fullMapString()
+ ']');
long updateSeq = this.updateSeq.incrementAndGet();
for (int p = 0; p < num; p++) {
GridDhtLocalPartition locPart = localPartition(p, topVer, false, false);
if (cctx.affinity().localNode(p, topVer)) {
// This partition will be created during next topology event,
// which obviously has not happened at this point.
if (locPart == null) {
if (log.isDebugEnabled())
log.debug("Skipping local partition afterExchange (will not create): " + p);
continue;
}
GridDhtPartitionState state = locPart.state();
if (state == MOVING) {
if (cctx.rebalanceEnabled()) {
Collection<ClusterNode> owners = owners(p);
// If there are no other owners, then become an owner.
if (F.isEmpty(owners)) {
boolean owned = locPart.own();
assert owned
: "Failed to own partition [cacheName"
+ cctx.name()
+ ", locPart="
+ locPart
+ ']';
updateLocal(p, loc.id(), locPart.state(), updateSeq);
changed = true;
if (cctx.events().isRecordable(EVT_CACHE_REBALANCE_PART_DATA_LOST)) {
DiscoveryEvent discoEvt = exchFut.discoveryEvent();
cctx.events()
.addPreloadEvent(
p,
EVT_CACHE_REBALANCE_PART_DATA_LOST,
discoEvt.eventNode(),
discoEvt.type(),
discoEvt.timestamp());
}
if (log.isDebugEnabled()) log.debug("Owned partition: " + locPart);
} else if (log.isDebugEnabled())
log.debug(
"Will not own partition (there are owners to rebalance from) [locPart="
+ locPart
+ ", owners = "
+ owners
+ ']');
} else updateLocal(p, loc.id(), locPart.state(), updateSeq);
}
} else {
if (locPart != null) {
GridDhtPartitionState state = locPart.state();
if (state == MOVING) {
locPart.rent(false);
updateLocal(p, loc.id(), locPart.state(), updateSeq);
changed = true;
if (log.isDebugEnabled())
log.debug("Evicting moving partition (it does not belong to affinity): " + locPart);
}
}
}
}
consistencyCheck();
} finally {
lock.writeLock().unlock();
}
return changed;
}
/**
* @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();
}
}