/**
* Waits for renting partitions.
*
* @return {@code True} if mapping was changed.
* @throws IgniteCheckedException If failed.
*/
private boolean waitForRent() throws IgniteCheckedException {
boolean changed = false;
// Synchronously wait for all renting partitions to complete.
for (Iterator<GridDhtLocalPartition> it = locParts.values().iterator(); it.hasNext(); ) {
GridDhtLocalPartition p = it.next();
GridDhtPartitionState state = p.state();
if (state == RENTING || state == EVICTED) {
if (log.isDebugEnabled()) log.debug("Waiting for renting partition: " + p);
// Wait for partition to empty out.
p.rent(true).get();
if (log.isDebugEnabled()) log.debug("Finished waiting for renting partition: " + p);
// Remove evicted partition.
it.remove();
changed = true;
}
}
return changed;
}
/**
* @param updateSeq Update sequence.
* @return {@code True} if entry has been transitioned to state EVICTED.
*/
boolean tryEvict(boolean updateSeq) {
if (state.getReference() != RENTING || state.getStamp() != 0 || groupReserved()) return false;
// Attempt to evict partition entries from cache.
clearAll();
if (map.isEmpty() && state.compareAndSet(RENTING, EVICTED, 0, 0)) {
if (log.isDebugEnabled()) log.debug("Evicted partition: " + this);
if (!GridQueryProcessor.isEnabled(cctx.config())) clearSwap();
if (cctx.isDrEnabled()) cctx.dr().partitionEvicted(id);
cctx.dataStructures().onPartitionEvicted(id);
rent.onDone();
((GridDhtPreloader) cctx.preloader()).onPartitionEvicted(this, updateSeq);
clearDeferredDeletes();
return true;
}
return false;
}
/**
* @param updateSeq Update sequence.
* @return Future for evict attempt.
*/
IgniteInternalFuture<Boolean> tryEvictAsync(boolean updateSeq) {
if (map.isEmpty()
&& !GridQueryProcessor.isEnabled(cctx.config())
&& state.compareAndSet(RENTING, EVICTED, 0, 0)) {
if (log.isDebugEnabled()) log.debug("Evicted partition: " + this);
clearSwap();
if (cctx.isDrEnabled()) cctx.dr().partitionEvicted(id);
cctx.dataStructures().onPartitionEvicted(id);
rent.onDone();
((GridDhtPreloader) cctx.preloader()).onPartitionEvicted(this, updateSeq);
clearDeferredDeletes();
return new GridFinishedFuture<>(true);
}
return cctx.closures()
.callLocalSafe(
new GPC<Boolean>() {
@Override
public Boolean call() {
return tryEvict(true);
}
}, /*system pool*/
true);
}
/**
* @param p Partition number.
* @param topVer Topology version.
* @param create Create flag.
* @param updateSeq Update sequence.
* @return Local partition.
*/
private GridDhtLocalPartition localPartition(
int p, AffinityTopologyVersion topVer, boolean create, boolean updateSeq) {
while (true) {
boolean belongs = cctx.affinity().localNode(p, topVer);
GridDhtLocalPartition loc = locParts.get(p);
if (loc != null && loc.state() == EVICTED) {
locParts.remove(p, loc);
if (!create) return null;
if (!belongs)
throw new GridDhtInvalidPartitionException(
p,
"Adding entry to evicted partition [part="
+ p
+ ", topVer="
+ topVer
+ ", this.topVer="
+ this.topVer
+ ']');
continue;
}
if (loc == null && create) {
if (!belongs)
throw new GridDhtInvalidPartitionException(
p,
"Creating partition which does not belong [part="
+ p
+ ", topVer="
+ topVer
+ ", this.topVer="
+ this.topVer
+ ']');
lock.writeLock().lock();
try {
GridDhtLocalPartition old =
locParts.putIfAbsent(p, loc = new GridDhtLocalPartition(cctx, p));
if (old != null) loc = old;
else {
if (updateSeq) this.updateSeq.incrementAndGet();
if (log.isDebugEnabled()) log.debug("Created local partition: " + loc);
}
} finally {
lock.writeLock().unlock();
}
}
return loc;
}
}
/**
* @param updateSeq Update sequence.
* @return Future to signal that this node is no longer an owner or backup.
*/
IgniteInternalFuture<?> rent(boolean updateSeq) {
while (true) {
int reservations = state.getStamp();
GridDhtPartitionState s = state.getReference();
if (s == RENTING || s == EVICTED) return rent;
if (state.compareAndSet(s, RENTING, reservations, reservations)) {
if (log.isDebugEnabled()) log.debug("Moved partition to RENTING state: " + this);
// Evict asynchronously, as the 'rent' method may be called
// from within write locks on local partition.
tryEvictAsync(updateSeq);
break;
}
}
return rent;
}
/** @return {@code True} if transitioned to OWNING state. */
boolean own() {
while (true) {
int reservations = state.getStamp();
GridDhtPartitionState s = state.getReference();
if (s == RENTING || s == EVICTED) return false;
if (s == OWNING) return true;
assert s == MOVING;
if (state.compareAndSet(MOVING, OWNING, reservations, reservations)) {
if (log.isDebugEnabled()) log.debug("Owned partition: " + this);
// No need to keep history any more.
evictHist = null;
return true;
}
}
}
/** {@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();
}
}
/** {@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;
}
/** {@inheritDoc} */
@Override
public void beforeExchange(GridDhtPartitionsExchangeFuture exchFut)
throws IgniteCheckedException {
waitForRent();
ClusterNode loc = cctx.localNode();
int num = cctx.affinity().partitions();
lock.writeLock().lock();
try {
GridDhtPartitionExchangeId exchId = exchFut.exchangeId();
if (stopping) return;
assert topVer.equals(exchId.topologyVersion())
: "Invalid topology version [topVer=" + topVer + ", exchId=" + exchId + ']';
if (exchId.isLeft()) removeNode(exchId.nodeId());
// In case if node joins, get topology at the time of joining node.
ClusterNode oldest = CU.oldestAliveCacheServerNode(cctx.shared(), topVer);
assert oldest != null;
if (log.isDebugEnabled())
log.debug(
"Partition map beforeExchange [exchId="
+ exchId
+ ", fullMap="
+ fullMapString()
+ ']');
long updateSeq = this.updateSeq.incrementAndGet();
// If this is the oldest node.
if (oldest.id().equals(loc.id())
|| exchFut.isCacheAdded(cctx.cacheId(), exchId.topologyVersion())) {
if (node2part == null) {
node2part = new GridDhtPartitionFullMap(oldest.id(), oldest.order(), updateSeq);
if (log.isDebugEnabled())
log.debug(
"Created brand new full topology map on oldest node [exchId="
+ exchId
+ ", fullMap="
+ fullMapString()
+ ']');
} else if (!node2part.valid()) {
node2part =
new GridDhtPartitionFullMap(oldest.id(), oldest.order(), updateSeq, node2part, false);
if (log.isDebugEnabled())
log.debug(
"Created new full topology map on oldest node [exchId="
+ exchId
+ ", fullMap="
+ node2part
+ ']');
} else if (!node2part.nodeId().equals(loc.id())) {
node2part =
new GridDhtPartitionFullMap(oldest.id(), oldest.order(), updateSeq, node2part, false);
if (log.isDebugEnabled())
log.debug(
"Copied old map into new map on oldest node (previous oldest node left) [exchId="
+ exchId
+ ", fullMap="
+ fullMapString()
+ ']');
}
}
if (cctx.rebalanceEnabled()) {
for (int p = 0; p < num; p++) {
// If this is the first node in grid.
boolean added = exchFut.isCacheAdded(cctx.cacheId(), exchId.topologyVersion());
if ((oldest.id().equals(loc.id())
&& oldest.id().equals(exchId.nodeId())
&& exchId.isJoined())
|| added) {
assert exchId.isJoined() || added;
try {
GridDhtLocalPartition locPart = localPartition(p, topVer, true, false);
assert locPart != null;
boolean owned = locPart.own();
assert owned
: "Failed to own partition for oldest node [cacheName"
+ cctx.name()
+ ", part="
+ locPart
+ ']';
if (log.isDebugEnabled()) log.debug("Owned partition for oldest node: " + locPart);
updateLocal(p, loc.id(), locPart.state(), updateSeq);
} catch (GridDhtInvalidPartitionException e) {
if (log.isDebugEnabled())
log.debug(
"Ignoring invalid partition on oldest node (no need to create a partition "
+ "if it no longer belongs to local node: "
+ e.partition());
}
}
// If this is not the first node in grid.
else {
if (node2part != null && node2part.valid()) {
if (cctx.affinity().localNode(p, topVer)) {
try {
// This will make sure that all non-existing partitions
// will be created in MOVING state.
GridDhtLocalPartition locPart = localPartition(p, topVer, true, false);
updateLocal(p, loc.id(), locPart.state(), updateSeq);
} catch (GridDhtInvalidPartitionException e) {
if (log.isDebugEnabled())
log.debug(
"Ignoring invalid partition (no need to create a partition if it "
+ "no longer belongs to local node: "
+ e.partition());
}
}
}
// If this node's map is empty, we pre-create local partitions,
// so local map will be sent correctly during exchange.
else if (cctx.affinity().localNode(p, topVer)) {
try {
localPartition(p, topVer, true, false);
} catch (GridDhtInvalidPartitionException e) {
if (log.isDebugEnabled())
log.debug(
"Ignoring invalid partition (no need to pre-create a partition if it "
+ "no longer belongs to local node: "
+ e.partition());
}
}
}
}
} else {
// If preloader is disabled, then we simply clear out
// the partitions this node is not responsible for.
for (int p = 0; p < num; p++) {
GridDhtLocalPartition locPart = localPartition(p, topVer, false, false);
boolean belongs = cctx.affinity().localNode(p, topVer);
if (locPart != null) {
if (!belongs) {
GridDhtPartitionState state = locPart.state();
if (state.active()) {
locPart.rent(false);
updateLocal(p, loc.id(), locPart.state(), updateSeq);
if (log.isDebugEnabled())
log.debug(
"Evicting partition with rebalancing disabled "
+ "(it does not belong to affinity): "
+ locPart);
}
}
} else if (belongs) {
try {
// Pre-create partitions.
localPartition(p, topVer, true, false);
} catch (GridDhtInvalidPartitionException e) {
if (log.isDebugEnabled())
log.debug(
"Ignoring invalid partition with disabled rebalancer (no need to "
+ "pre-create a partition if it no longer belongs to local node: "
+ e.partition());
}
}
}
}
if (node2part != null && node2part.valid()) checkEvictions(updateSeq);
consistencyCheck();
if (log.isDebugEnabled())
log.debug(
"Partition map after beforeExchange [exchId="
+ exchId
+ ", fullMap="
+ fullMapString()
+ ']');
} finally {
lock.writeLock().unlock();
}
// Wait for evictions.
waitForRent();
}
/**
* @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();
}
}