/**
   * @param cctx Context.
   * @param id Partition ID.
   */
  @SuppressWarnings("ExternalizableWithoutPublicNoArgConstructor")
  GridDhtLocalPartition(GridCacheContext cctx, int id) {
    assert cctx != null;

    this.id = id;
    this.cctx = cctx;

    log = U.logger(cctx.kernalContext(), logRef, this);

    rent =
        new GridFutureAdapter<Object>() {
          @Override
          public String toString() {
            return "PartitionRentFuture [part=" + GridDhtLocalPartition.this + ", map=" + map + ']';
          }
        };

    map = new ConcurrentHashMap8<>(cctx.config().getStartSize() / cctx.affinity().partitions());

    int delQueueSize =
        CU.isSystemCache(cctx.name())
            ? 100
            : Math.max(MAX_DELETE_QUEUE_SIZE / cctx.affinity().partitions(), 20);

    rmvQueue = new GridCircularBuffer<>(U.ceilPow2(delQueueSize));
  }
  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);
    }
  }
  /** {@inheritDoc} */
  @Override
  public void finishUnmarshal(GridCacheSharedContext ctx, ClassLoader ldr)
      throws IgniteCheckedException {
    super.finishUnmarshal(ctx, ldr);

    if (writes != null) unmarshalTx(writes, false, ctx, ldr);

    if (reads != null) unmarshalTx(reads, false, ctx, ldr);

    if (grpLockKeyBytes != null && grpLockKey == null)
      grpLockKey = ctx.marshaller().unmarshal(grpLockKeyBytes, ldr);

    if (dhtVerKeys != null && dhtVers == null) {
      assert dhtVerVals != null;
      assert dhtVerKeys.size() == dhtVerVals.size();

      Iterator<IgniteTxKey> keyIt = dhtVerKeys.iterator();
      Iterator<GridCacheVersion> verIt = dhtVerVals.iterator();

      dhtVers = U.newHashMap(dhtVerKeys.size());

      while (keyIt.hasNext()) {
        IgniteTxKey key = keyIt.next();

        key.finishUnmarshal(ctx.cacheContext(key.cacheId()), ldr);

        dhtVers.put(key, verIt.next());
      }
    }

    if (txNodesBytes != null) txNodes = ctx.marshaller().unmarshal(txNodesBytes, ldr);
  }
  /**
   * @param cacheCtx Cache context.
   * @return {@code True} if local node can calculate affinity on it's own for this partition map
   *     exchange.
   */
  private boolean canCalculateAffinity(GridCacheContext cacheCtx) {
    AffinityFunction affFunc = cacheCtx.config().getAffinity();

    // Do not request affinity from remote nodes if affinity function is not centralized.
    if (!U.hasAnnotation(affFunc, AffinityCentralizedFunction.class)) return true;

    // If local node did not initiate exchange or local node is the only cache node in grid.
    Collection<ClusterNode> affNodes = CU.affinityNodes(cacheCtx, exchId.topologyVersion());

    return !exchId.nodeId().equals(cctx.localNodeId())
        || (affNodes.size() == 1 && affNodes.contains(cctx.localNode()));
  }
  /**
   * @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);
    }
  }
 /** {@inheritDoc} */
 @Override
 public String toString() {
   return S.toString(
       GridDhtLocalPartition.class,
       this,
       "state",
       state(),
       "reservations",
       reservations(),
       "empty",
       map.isEmpty(),
       "createTime",
       U.format(createTime),
       "mapPubSize",
       mapPubSize);
 }
  private void dumpPendingObjects() {
    U.warn(
        log,
        "Failed to wait for partition release future. Dumping pending objects that might be the cause: "
            + cctx.localNodeId());

    U.warn(log, "Pending transactions:");

    for (IgniteInternalTx tx : cctx.tm().activeTransactions()) U.warn(log, ">>> " + tx);

    U.warn(log, "Pending explicit locks:");

    for (GridCacheExplicitLockSpan lockSpan : cctx.mvcc().activeExplicitLocks())
      U.warn(log, ">>> " + lockSpan);

    U.warn(log, "Pending cache futures:");

    for (GridCacheFuture<?> fut : cctx.mvcc().activeFutures()) U.warn(log, ">>> " + fut);

    U.warn(log, "Pending atomic cache futures:");

    for (GridCacheFuture<?> fut : cctx.mvcc().atomicFutures()) U.warn(log, ">>> " + fut);
  }
  /** {@inheritDoc} */
  @Override
  public void finishUnmarshal(GridCacheSharedContext ctx, ClassLoader ldr)
      throws IgniteCheckedException {
    super.finishUnmarshal(ctx, ldr);

    if (ownedValKeys != null && ownedVals == null) {
      ownedVals = U.newHashMap(ownedValKeys.size());

      assert ownedValKeys.size() == ownedValVals.size();

      Iterator<IgniteTxKey> keyIter = ownedValKeys.iterator();

      Iterator<CacheVersionedValue> valIter = ownedValVals.iterator();

      while (keyIter.hasNext()) {
        IgniteTxKey key = keyIter.next();

        GridCacheContext cctx = ctx.cacheContext(key.cacheId());

        CacheVersionedValue val = valIter.next();

        key.finishUnmarshal(cctx, ldr);

        val.finishUnmarshal(cctx, ldr);

        ownedVals.put(key, val);
      }
    }

    if (retVal != null && retVal.cacheId() != 0) {
      GridCacheContext cctx = ctx.cacheContext(retVal.cacheId());

      assert cctx != null : retVal.cacheId();

      retVal.finishUnmarshal(cctx, ldr);
    }

    if (filterFailedKeys != null) {
      for (IgniteTxKey key : filterFailedKeys) {
        GridCacheContext cctx = ctx.cacheContext(key.cacheId());

        key.finishUnmarshal(cctx, ldr);
      }
    }
  }
  /** @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;
  }
Пример #13
0
  /** 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);
    }
  }
Пример #14
0
/** Key partition. */
public class GridDhtLocalPartition implements Comparable<GridDhtLocalPartition>, GridReservable {
  /** Maximum size for delete queue. */
  public static final int MAX_DELETE_QUEUE_SIZE =
      Integer.getInteger(IGNITE_ATOMIC_CACHE_DELETE_HISTORY_SIZE, 200_000);

  /** Static logger to avoid re-creation. */
  private static final AtomicReference<IgniteLogger> logRef = new AtomicReference<>();

  /** Logger. */
  private static volatile IgniteLogger log;

  /** Partition ID. */
  private final int id;

  /** State. */
  @GridToStringExclude
  private final AtomicStampedReference<GridDhtPartitionState> state =
      new AtomicStampedReference<>(MOVING, 0);

  /** Rent future. */
  @GridToStringExclude private final GridFutureAdapter<?> rent;

  /** Entries map. */
  private final ConcurrentMap<KeyCacheObject, GridDhtCacheEntry> map;

  /** Context. */
  private final GridCacheContext cctx;

  /** Create time. */
  @GridToStringExclude private final long createTime = U.currentTimeMillis();

  /** Eviction history. */
  private volatile Map<KeyCacheObject, GridCacheVersion> evictHist = new HashMap<>();

  /** Lock. */
  private final ReentrantLock lock = new ReentrantLock();

  /** Public size counter. */
  private final LongAdder8 mapPubSize = new LongAdder8();

  /** Remove queue. */
  private final GridCircularBuffer<T2<KeyCacheObject, GridCacheVersion>> rmvQueue;

  /** Group reservations. */
  private final CopyOnWriteArrayList<GridDhtPartitionsReservation> reservations =
      new CopyOnWriteArrayList<>();

  /**
   * @param cctx Context.
   * @param id Partition ID.
   */
  @SuppressWarnings("ExternalizableWithoutPublicNoArgConstructor")
  GridDhtLocalPartition(GridCacheContext cctx, int id) {
    assert cctx != null;

    this.id = id;
    this.cctx = cctx;

    log = U.logger(cctx.kernalContext(), logRef, this);

    rent =
        new GridFutureAdapter<Object>() {
          @Override
          public String toString() {
            return "PartitionRentFuture [part=" + GridDhtLocalPartition.this + ", map=" + map + ']';
          }
        };

    map = new ConcurrentHashMap8<>(cctx.config().getStartSize() / cctx.affinity().partitions());

    int delQueueSize =
        CU.isSystemCache(cctx.name())
            ? 100
            : Math.max(MAX_DELETE_QUEUE_SIZE / cctx.affinity().partitions(), 20);

    rmvQueue = new GridCircularBuffer<>(U.ceilPow2(delQueueSize));
  }

  /**
   * Adds group reservation to this partition.
   *
   * @param r Reservation.
   * @return {@code false} If such reservation already added.
   */
  public boolean addReservation(GridDhtPartitionsReservation r) {
    assert state.getReference() != EVICTED : "we can reserve only active partitions";
    assert state.getStamp() != 0
        : "partition must be already reserved before adding group reservation";

    return reservations.addIfAbsent(r);
  }

  /** @param r Reservation. */
  public void removeReservation(GridDhtPartitionsReservation r) {
    if (!reservations.remove(r))
      throw new IllegalStateException("Reservation was already removed.");
  }

  /** @return Partition ID. */
  public int id() {
    return id;
  }

  /** @return Create time. */
  long createTime() {
    return createTime;
  }

  /** @return Partition state. */
  public GridDhtPartitionState state() {
    return state.getReference();
  }

  /** @return Reservations. */
  public int reservations() {
    return state.getStamp();
  }

  /** @return Keys belonging to partition. */
  public Set<KeyCacheObject> keySet() {
    return map.keySet();
  }

  /** @return Entries belonging to partition. */
  public Collection<GridDhtCacheEntry> entries() {
    return map.values();
  }

  /** @return {@code True} if partition is empty. */
  public boolean isEmpty() {
    return map.isEmpty();
  }

  /** @return Number of entries in this partition (constant-time method). */
  public int size() {
    return map.size();
  }

  /** Increments public size of the map. */
  public void incrementPublicSize() {
    mapPubSize.increment();
  }

  /** Decrements public size of the map. */
  public void decrementPublicSize() {
    mapPubSize.decrement();
  }

  /** @return Number of public (non-internal) entries in this partition. */
  public int publicSize() {
    return mapPubSize.intValue();
  }

  /** @return If partition is moving or owning or renting. */
  public boolean valid() {
    GridDhtPartitionState state = state();

    return state == MOVING || state == OWNING || state == RENTING;
  }

  /** @param entry Entry to add. */
  void onAdded(GridDhtCacheEntry entry) {
    GridDhtPartitionState state = state();

    if (state == EVICTED)
      throw new GridDhtInvalidPartitionException(
          id, "Adding entry to invalid partition [part=" + id + ']');

    map.put(entry.key(), entry);

    if (!entry.isInternal()) mapPubSize.increment();
  }

  /** @param entry Entry to remove. */
  @SuppressWarnings("SynchronizationOnLocalVariableOrMethodParameter")
  void onRemoved(GridDhtCacheEntry entry) {
    assert entry.obsolete();

    // Make sure to remove exactly this entry.
    synchronized (entry) {
      map.remove(entry.key(), entry);

      if (!entry.isInternal() && !entry.deleted()) mapPubSize.decrement();
    }

    // Attempt to evict.
    tryEvict(true);
  }

  /**
   * @param key Removed key.
   * @param ver Removed version.
   * @throws IgniteCheckedException If failed.
   */
  public void onDeferredDelete(KeyCacheObject key, GridCacheVersion ver)
      throws IgniteCheckedException {
    try {
      T2<KeyCacheObject, GridCacheVersion> evicted = rmvQueue.add(new T2<>(key, ver));

      if (evicted != null) cctx.dht().removeVersionedEntry(evicted.get1(), evicted.get2());
    } catch (InterruptedException e) {
      Thread.currentThread().interrupt();

      throw new IgniteInterruptedCheckedException(e);
    }
  }

  /** Locks partition. */
  @SuppressWarnings({"LockAcquiredButNotSafelyReleased"})
  public void lock() {
    lock.lock();
  }

  /** Unlocks partition. */
  public void unlock() {
    lock.unlock();
  }

  /**
   * @param key Key.
   * @param ver Version.
   */
  public void onEntryEvicted(KeyCacheObject key, GridCacheVersion ver) {
    assert key != null;
    assert ver != null;
    assert lock.isHeldByCurrentThread(); // Only one thread can enter this method at a time.

    if (state() != MOVING) return;

    Map<KeyCacheObject, GridCacheVersion> evictHist0 = evictHist;

    if (evictHist0 != null) {
      GridCacheVersion ver0 = evictHist0.get(key);

      if (ver0 == null || ver0.isLess(ver)) {
        GridCacheVersion ver1 = evictHist0.put(key, ver);

        assert ver1 == ver0;
      }
    }
  }

  /**
   * Cache preloader should call this method within partition lock.
   *
   * @param key Key.
   * @param ver Version.
   * @return {@code True} if preloading is permitted.
   */
  public boolean preloadingPermitted(KeyCacheObject key, GridCacheVersion ver) {
    assert key != null;
    assert ver != null;
    assert lock.isHeldByCurrentThread(); // Only one thread can enter this method at a time.

    if (state() != MOVING) return false;

    Map<KeyCacheObject, GridCacheVersion> evictHist0 = evictHist;

    if (evictHist0 != null) {
      GridCacheVersion ver0 = evictHist0.get(key);

      // Permit preloading if version in history
      // is missing or less than passed in.
      return ver0 == null || ver0.isLess(ver);
    }

    return false;
  }

  /**
   * Reserves a partition so it won't be cleared.
   *
   * @return {@code True} if reserved.
   */
  @Override
  public boolean reserve() {
    while (true) {
      int reservations = state.getStamp();

      GridDhtPartitionState s = state.getReference();

      if (s == EVICTED) return false;

      if (state.compareAndSet(s, s, reservations, reservations + 1)) return true;
    }
  }

  /** Releases previously reserved partition. */
  @Override
  public void release() {
    while (true) {
      int reservations = state.getStamp();

      if (reservations == 0) return;

      GridDhtPartitionState s = state.getReference();

      assert s != EVICTED;

      // Decrement reservations.
      if (state.compareAndSet(s, s, reservations, --reservations)) {
        tryEvict(true);

        break;
      }
    }
  }

  /** @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;
      }
    }
  }

  /**
   * @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;
  }

  /**
   * @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);
  }

  /** @return {@code true} If there is a group reservation. */
  private boolean groupReserved() {
    for (GridDhtPartitionsReservation reservation : reservations) {
      if (!reservation.invalidate())
        return true; // Failed to invalidate reservation -> we are reserved.
    }

    return false;
  }

  /**
   * @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;
  }

  /** 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);
    }
  }

  /** */
  void onUnlock() {
    tryEvict(true);
  }

  /**
   * @param topVer Topology version.
   * @return {@code True} if local node is primary for this partition.
   */
  public boolean primary(AffinityTopologyVersion topVer) {
    return cctx.affinity().primary(cctx.localNode(), id, topVer);
  }

  /** 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);
    }
  }

  /**
   * @param it Swap iterator.
   * @return Unswapping iterator over swapped entries.
   */
  private Iterator<GridDhtCacheEntry> unswapIterator(
      final GridCloseableIterator<Map.Entry<byte[], GridCacheSwapEntry>> it) {
    if (it == null) return null;

    return new Iterator<GridDhtCacheEntry>() {
      /** */
      GridDhtCacheEntry lastEntry;

      @Override
      public boolean hasNext() {
        return it.hasNext();
      }

      @Override
      public GridDhtCacheEntry next() {
        Map.Entry<byte[], GridCacheSwapEntry> entry = it.next();

        byte[] keyBytes = entry.getKey();

        try {
          KeyCacheObject key = cctx.toCacheKeyObject(keyBytes);

          lastEntry = (GridDhtCacheEntry) cctx.cache().entryEx(key, false);

          lastEntry.unswap(true);

          return lastEntry;
        } catch (IgniteCheckedException e) {
          throw new CacheException(e);
        }
      }

      @Override
      public void remove() {
        map.remove(lastEntry.key(), lastEntry);
      }
    };
  }

  /** */
  private void clearDeferredDeletes() {
    rmvQueue.forEach(
        new CI1<T2<KeyCacheObject, GridCacheVersion>>() {
          @Override
          public void apply(T2<KeyCacheObject, GridCacheVersion> t) {
            cctx.dht().removeVersionedEntry(t.get1(), t.get2());
          }
        });
  }

  /** {@inheritDoc} */
  @Override
  public int hashCode() {
    return id;
  }

  /** {@inheritDoc} */
  @SuppressWarnings({"OverlyStrongTypeCast"})
  @Override
  public boolean equals(Object obj) {
    return obj instanceof GridDhtLocalPartition
        && (obj == this || ((GridDhtLocalPartition) obj).id() == id);
  }

  /** {@inheritDoc} */
  @Override
  public int compareTo(@NotNull GridDhtLocalPartition part) {
    if (part == null) return 1;

    return Integer.compare(id, part.id());
  }

  /** {@inheritDoc} */
  @Override
  public String toString() {
    return S.toString(
        GridDhtLocalPartition.class,
        this,
        "state",
        state(),
        "reservations",
        reservations(),
        "empty",
        map.isEmpty(),
        "createTime",
        U.format(createTime),
        "mapPubSize",
        mapPubSize);
  }
}