/**
   * @param nodes Nodes.
   * @param id ID.
   * @throws IgniteCheckedException If failed.
   */
  private void sendAllPartitions(
      Collection<? extends ClusterNode> nodes, GridDhtPartitionExchangeId id)
      throws IgniteCheckedException {
    GridDhtPartitionsFullMessage m =
        new GridDhtPartitionsFullMessage(id, lastVer.get(), id.topologyVersion());

    for (GridCacheContext cacheCtx : cctx.cacheContexts()) {
      if (!cacheCtx.isLocal()) {
        AffinityTopologyVersion startTopVer = cacheCtx.startTopologyVersion();

        boolean ready = startTopVer == null || startTopVer.compareTo(id.topologyVersion()) <= 0;

        if (ready)
          m.addFullPartitionsMap(cacheCtx.cacheId(), cacheCtx.topology().partitionMap(true));
      }
    }

    // It is important that client topologies be added after contexts.
    for (GridClientPartitionTopology top : cctx.exchange().clientTopologies())
      m.addFullPartitionsMap(top.cacheId(), top.partitionMap(true));

    if (log.isDebugEnabled())
      log.debug(
          "Sending full partition map [nodeIds="
              + F.viewReadOnly(nodes, F.node2id())
              + ", exchId="
              + exchId
              + ", msg="
              + m
              + ']');

    cctx.io().safeSend(nodes, m, SYSTEM_POOL, null);
  }
  /**
   * @param cctx Cache context.
   * @param busyLock Busy lock.
   * @param exchId Exchange ID.
   * @param reqs Cache change requests.
   */
  public GridDhtPartitionsExchangeFuture(
      GridCacheSharedContext cctx,
      ReadWriteLock busyLock,
      GridDhtPartitionExchangeId exchId,
      Collection<DynamicCacheChangeRequest> reqs) {
    assert busyLock != null;
    assert exchId != null;

    dummy = false;
    forcePreload = false;
    reassign = false;

    this.cctx = cctx;
    this.busyLock = busyLock;
    this.exchId = exchId;
    this.reqs = reqs;

    log = cctx.logger(getClass());

    initFut = new GridFutureAdapter<>();

    if (log.isDebugEnabled())
      log.debug(
          "Creating exchange future [localNode=" + cctx.localNodeId() + ", fut=" + this + ']');
  }
  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);
    }
  }
  /**
   * 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 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 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;
  }
  /** @return {@code true} if entered to busy state. */
  private boolean enterBusy() {
    if (busyLock.readLock().tryLock()) return true;

    if (log.isDebugEnabled())
      log.debug("Failed to enter busy state (exchanger is stopping): " + this);

    return false;
  }
  /**
   * @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 rmtReducer Optional reducer.
   * @param rmtTransform Optional transformer.
   * @param args Arguments.
   * @return Future.
   */
  @SuppressWarnings("IfMayBeConditional")
  private <R> CacheQueryFuture<R> execute(
      @Nullable IgniteReducer<T, R> rmtReducer,
      @Nullable IgniteClosure<T, R> rmtTransform,
      @Nullable Object... args) {
    Collection<ClusterNode> nodes = nodes();

    cctx.checkSecurity(SecurityPermission.CACHE_READ);

    if (nodes.isEmpty())
      return new GridCacheQueryErrorFuture<>(
          cctx.kernalContext(), new ClusterGroupEmptyCheckedException());

    if (log.isDebugEnabled())
      log.debug("Executing query [query=" + this + ", nodes=" + nodes + ']');

    if (cctx.deploymentEnabled()) {
      try {
        cctx.deploy().registerClasses(filter, rmtReducer, rmtTransform);
        cctx.deploy().registerClasses(args);
      } catch (IgniteCheckedException e) {
        return new GridCacheQueryErrorFuture<>(cctx.kernalContext(), e);
      }
    }

    if (subjId == null) subjId = cctx.localNodeId();

    taskHash = cctx.kernalContext().job().currentTaskNameHash();

    final GridCacheQueryBean bean =
        new GridCacheQueryBean(
            this,
            (IgniteReducer<Object, Object>) rmtReducer,
            (IgniteClosure<Object, Object>) rmtTransform,
            args);

    final GridCacheQueryManager qryMgr = cctx.queries();

    boolean loc = nodes.size() == 1 && F.first(nodes).id().equals(cctx.localNodeId());

    if (type == SQL_FIELDS || type == SPI)
      return (CacheQueryFuture<R>)
          (loc ? qryMgr.queryFieldsLocal(bean) : qryMgr.queryFieldsDistributed(bean, nodes));
    else if (type == SCAN && part != null && nodes.size() > 1)
      return new CacheQueryFallbackFuture<>(nodes, bean, qryMgr);
    else
      return (CacheQueryFuture<R>)
          (loc ? qryMgr.queryLocal(bean) : qryMgr.queryDistributed(bean, nodes));
  }
  /** {@inheritDoc} */
  @Override
  public boolean onDone(AffinityTopologyVersion res, Throwable err) {
    Map<Integer, Boolean> m = null;

    for (GridCacheContext cacheCtx : cctx.cacheContexts()) {
      if (cacheCtx.config().getTopologyValidator() != null && !CU.isSystemCache(cacheCtx.name())) {
        if (m == null) m = new HashMap<>();

        m.put(
            cacheCtx.cacheId(),
            cacheCtx.config().getTopologyValidator().validate(discoEvt.topologyNodes()));
      }
    }

    cacheValidRes = m != null ? m : Collections.<Integer, Boolean>emptyMap();

    cctx.cache().onExchangeDone(exchId.topologyVersion(), reqs, err);

    cctx.exchange().onExchangeDone(this, err);

    if (super.onDone(res, err) && !dummy && !forcePreload) {
      if (log.isDebugEnabled())
        log.debug(
            "Completed partition exchange [localNode="
                + cctx.localNodeId()
                + ", exchange= "
                + this
                + ']');

      initFut.onDone(err == null);

      GridTimeoutObject timeoutObj = this.timeoutObj;

      // Deschedule timeout object.
      if (timeoutObj != null) cctx.kernalContext().timeout().removeTimeoutObject(timeoutObj);

      if (exchId.isLeft()) {
        for (GridCacheContext cacheCtx : cctx.cacheContexts())
          cacheCtx.config().getAffinity().removeNode(exchId.nodeId());
      }

      return true;
    }

    return dummy;
  }
예제 #11
0
  /**
   * @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;
  }
예제 #12
0
  /** @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 node Node.
   * @param id ID.
   * @throws IgniteCheckedException If failed.
   */
  private void sendLocalPartitions(ClusterNode node, @Nullable GridDhtPartitionExchangeId id)
      throws IgniteCheckedException {
    GridDhtPartitionsSingleMessage m =
        new GridDhtPartitionsSingleMessage(
            id, cctx.kernalContext().clientNode(), cctx.versions().last());

    for (GridCacheContext cacheCtx : cctx.cacheContexts()) {
      if (!cacheCtx.isLocal())
        m.addLocalPartitionMap(cacheCtx.cacheId(), cacheCtx.topology().localPartitionMap());
    }

    if (log.isDebugEnabled())
      log.debug(
          "Sending local partitions [nodeId="
              + node.id()
              + ", exchId="
              + exchId
              + ", msg="
              + m
              + ']');

    cctx.io().send(node, m, SYSTEM_POOL);
  }
  /**
   * @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);
    }
  }
  /**
   * @param cacheCtx Cache context.
   * @throws IgniteCheckedException If failed.
   */
  private void initTopology(GridCacheContext cacheCtx) throws IgniteCheckedException {
    if (stopping(cacheCtx.cacheId())) return;

    if (canCalculateAffinity(cacheCtx)) {
      if (log.isDebugEnabled())
        log.debug(
            "Will recalculate affinity [locNodeId="
                + cctx.localNodeId()
                + ", exchId="
                + exchId
                + ']');

      cacheCtx.affinity().calculateAffinity(exchId.topologyVersion(), discoEvt);
    } else {
      if (log.isDebugEnabled())
        log.debug(
            "Will request affinity from remote node [locNodeId="
                + cctx.localNodeId()
                + ", exchId="
                + exchId
                + ']');

      // Fetch affinity assignment from remote node.
      GridDhtAssignmentFetchFuture fetchFut =
          new GridDhtAssignmentFetchFuture(
              cacheCtx,
              exchId.topologyVersion(),
              CU.affinityNodes(cacheCtx, exchId.topologyVersion()));

      fetchFut.init();

      List<List<ClusterNode>> affAssignment = fetchFut.get();

      if (log.isDebugEnabled())
        log.debug(
            "Fetched affinity from remote node, initializing affinity assignment [locNodeId="
                + cctx.localNodeId()
                + ", topVer="
                + exchId.topologyVersion()
                + ']');

      if (affAssignment == null) {
        affAssignment = new ArrayList<>(cacheCtx.affinity().partitions());

        List<ClusterNode> empty = Collections.emptyList();

        for (int i = 0; i < cacheCtx.affinity().partitions(); i++) affAssignment.add(empty);
      }

      cacheCtx.affinity().initializeAffinity(exchId.topologyVersion(), affAssignment);
    }
  }
  /** {@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();
    }
  }
  /** {@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();
  }
  /** {@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;
  }
  /**
   * 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 nodeId Sender node ID.
   * @param msg Full partition info.
   */
  public void onReceive(final UUID nodeId, final GridDhtPartitionsFullMessage msg) {
    assert msg != null;

    if (isDone()) {
      if (log.isDebugEnabled())
        log.debug("Received message for finished future [msg=" + msg + ", fut=" + this + ']');

      return;
    }

    if (log.isDebugEnabled())
      log.debug("Received full partition map from node [nodeId=" + nodeId + ", msg=" + msg + ']');

    assert exchId.topologyVersion().equals(msg.topologyVersion());

    initFut.listen(
        new CI1<IgniteInternalFuture<Boolean>>() {
          @Override
          public void apply(IgniteInternalFuture<Boolean> t) {
            ClusterNode curOldest = oldestNode.get();

            if (!nodeId.equals(curOldest.id())) {
              if (log.isDebugEnabled())
                log.debug(
                    "Received full partition map from unexpected node [oldest="
                        + curOldest.id()
                        + ", unexpectedNodeId="
                        + nodeId
                        + ']');

              ClusterNode snd = cctx.discovery().node(nodeId);

              if (snd == null) {
                if (log.isDebugEnabled())
                  log.debug(
                      "Sender node left grid, will ignore message from unexpected node [nodeId="
                          + nodeId
                          + ", exchId="
                          + msg.exchangeId()
                          + ']');

                return;
              }

              // Will process message later if sender node becomes oldest node.
              if (snd.order() > curOldest.order()) fullMsgs.put(nodeId, msg);

              return;
            }

            assert msg.exchangeId().equals(exchId);

            assert msg.lastVersion() != null;

            cctx.versions().onReceived(nodeId, msg.lastVersion());

            updatePartitionFullMap(msg);

            onDone(exchId.topologyVersion());
          }
        });
  }
  /**
   * @param nodeId Sender node id.
   * @param msg Single partition info.
   */
  public void onReceive(final UUID nodeId, final GridDhtPartitionsSingleMessage msg) {
    assert msg != null;

    assert msg.exchangeId().equals(exchId);

    // Update last seen version.
    while (true) {
      GridCacheVersion old = lastVer.get();

      if (old == null || old.compareTo(msg.lastVersion()) < 0) {
        if (lastVer.compareAndSet(old, msg.lastVersion())) break;
      } else break;
    }

    if (isDone()) {
      if (log.isDebugEnabled())
        log.debug(
            "Received message for finished future (will reply only to sender) [msg="
                + msg
                + ", fut="
                + this
                + ']');

      sendAllPartitions(nodeId, cctx.gridConfig().getNetworkSendRetryCount());
    } else {
      initFut.listen(
          new CI1<IgniteInternalFuture<Boolean>>() {
            @Override
            public void apply(IgniteInternalFuture<Boolean> t) {
              try {
                if (!t.get()) // Just to check if there was an error.
                return;

                ClusterNode loc = cctx.localNode();

                singleMsgs.put(nodeId, msg);

                boolean match = true;

                // Check if oldest node has changed.
                if (!oldestNode.get().equals(loc)) {
                  match = false;

                  synchronized (mux) {
                    // Double check.
                    if (oldestNode.get().equals(loc)) match = true;
                  }
                }

                if (match) {
                  boolean allReceived;

                  synchronized (rcvdIds) {
                    if (rcvdIds.add(nodeId)) updatePartitionSingleMap(msg);

                    allReceived = allReceived();
                  }

                  // If got all replies, and initialization finished, and reply has not been sent
                  // yet.
                  if (allReceived && ready.get() && replied.compareAndSet(false, true)) {
                    spreadPartitions();

                    onDone(exchId.topologyVersion());
                  } else if (log.isDebugEnabled())
                    log.debug(
                        "Exchange future full map is not sent [allReceived="
                            + allReceived()
                            + ", ready="
                            + ready
                            + ", replied="
                            + replied.get()
                            + ", init="
                            + init.get()
                            + ", fut="
                            + this
                            + ']');
                }
              } catch (IgniteCheckedException e) {
                U.error(log, "Failed to initialize exchange future: " + this, e);
              }
            }
          });
    }
  }
  /**
   * @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;
  }