Ejemplo n.º 1
0
  /**
   * @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));
  }
  /** {@inheritDoc} */
  @Override
  public GridDhtPartitionMap localPartitionMap() {
    lock.readLock().lock();

    try {
      return new GridDhtPartitionMap(
          cctx.nodeId(), updateSeq.get(), F.viewReadOnly(locParts, CU.part2state()), true);
    } finally {
      lock.readLock().unlock();
    }
  }
  /**
   * @param p Partition.
   * @param topVer Topology version ({@code -1} for all nodes).
   * @param state Partition state.
   * @param states Additional partition states.
   * @return List of nodes for the partition.
   */
  private List<ClusterNode> nodes(
      int p,
      AffinityTopologyVersion topVer,
      GridDhtPartitionState state,
      GridDhtPartitionState... states) {
    Collection<UUID> allIds =
        topVer.topologyVersion() > 0 ? F.nodeIds(CU.affinityNodes(cctx, topVer)) : null;

    lock.readLock().lock();

    try {
      assert node2part != null && node2part.valid()
          : "Invalid node-to-partitions map [topVer="
              + topVer
              + ", allIds="
              + allIds
              + ", node2part="
              + node2part
              + ", cache="
              + cctx.name()
              + ']';

      Collection<UUID> nodeIds = part2node.get(p);

      // Node IDs can be null if both, primary and backup, nodes disappear.
      int size = nodeIds == null ? 0 : nodeIds.size();

      if (size == 0) return Collections.emptyList();

      List<ClusterNode> nodes = new ArrayList<>(size);

      for (UUID id : nodeIds) {
        if (topVer.topologyVersion() > 0 && !allIds.contains(id)) continue;

        if (hasState(p, id, state, states)) {
          ClusterNode n = cctx.discovery().node(id);

          if (n != null && (topVer.topologyVersion() < 0 || n.order() <= topVer.topologyVersion()))
            nodes.add(n);
        }
      }

      return nodes;
    } finally {
      lock.readLock().unlock();
    }
  }
  /** @param nodeId Node to remove. */
  private void removeNode(UUID nodeId) {
    assert nodeId != null;
    assert lock.writeLock().isHeldByCurrentThread();

    ClusterNode oldest = CU.oldestAliveCacheServerNode(cctx.shared(), topVer);

    assert oldest != null;

    ClusterNode loc = cctx.localNode();

    if (node2part != null) {
      if (oldest.equals(loc) && !node2part.nodeId().equals(loc.id())) {
        updateSeq.setIfGreater(node2part.updateSequence());

        node2part =
            new GridDhtPartitionFullMap(
                loc.id(), loc.order(), updateSeq.incrementAndGet(), node2part, false);
      } else node2part = new GridDhtPartitionFullMap(node2part, node2part.updateSequence());

      part2node = new HashMap<>(part2node);

      GridDhtPartitionMap parts = node2part.remove(nodeId);

      if (parts != null) {
        for (Integer p : parts.keySet()) {
          Set<UUID> nodeIds = part2node.get(p);

          if (nodeIds != null) {
            nodeIds.remove(nodeId);

            if (nodeIds.isEmpty()) part2node.remove(p);
          }
        }
      }

      consistencyCheck();
    }
  }
  /** {@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();
  }
  /**
   * Updates value for single partition.
   *
   * @param p Partition.
   * @param nodeId Node ID.
   * @param state State.
   * @param updateSeq Update sequence.
   */
  @SuppressWarnings({"MismatchedQueryAndUpdateOfCollection"})
  private void updateLocal(int p, UUID nodeId, GridDhtPartitionState state, long updateSeq) {
    assert lock.isWriteLockedByCurrentThread();
    assert nodeId.equals(cctx.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 this node became the oldest node.
    if (oldest.id().equals(cctx.nodeId())) {
      long seq = node2part.updateSequence();

      if (seq != updateSeq) {
        if (seq > updateSeq) {
          if (this.updateSeq.get() < seq) {
            // Update global counter if necessary.
            boolean b = this.updateSeq.compareAndSet(this.updateSeq.get(), seq + 1);

            assert b
                : "Invalid update sequence [updateSeq="
                    + updateSeq
                    + ", seq="
                    + seq
                    + ", curUpdateSeq="
                    + this.updateSeq.get()
                    + ", node2part="
                    + node2part.toFullString()
                    + ']';

            updateSeq = seq + 1;
          } else updateSeq = seq;
        }

        node2part.updateSequence(updateSeq);
      }
    }

    GridDhtPartitionMap map = node2part.get(nodeId);

    if (map == null)
      node2part.put(
          nodeId,
          map =
              new GridDhtPartitionMap(
                  nodeId,
                  updateSeq,
                  Collections.<Integer, GridDhtPartitionState>emptyMap(),
                  false));

    map.updateSequence(updateSeq);

    map.put(p, state);

    Set<UUID> ids = part2node.get(p);

    if (ids == null) part2node.put(p, ids = U.newHashSet(3));

    ids.add(nodeId);
  }
  /**
   * @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;
  }