/** {@inheritDoc} */
  @Override
  public GridDhtPartitionFullMap partitionMap(boolean onlyActive) {
    lock.readLock().lock();

    try {
      assert node2part != null && node2part.valid()
          : "Invalid node2part [node2part: "
              + node2part
              + ", cache="
              + cctx.name()
              + ", started="
              + cctx.started()
              + ", stopping="
              + stopping
              + ", locNodeId="
              + cctx.localNode().id()
              + ", locName="
              + cctx.gridName()
              + ']';

      GridDhtPartitionFullMap m = node2part;

      return new GridDhtPartitionFullMap(
          m.nodeId(), m.nodeOrder(), m.updateSequence(), m, onlyActive);
    } finally {
      lock.readLock().unlock();
    }
  }
Ejemplo n.º 2
0
  /**
   * @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;
  }
Ejemplo n.º 3
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 Collection<ClusterNode> nodes(int p, AffinityTopologyVersion topVer) {
    Collection<ClusterNode> affNodes = cctx.affinity().nodes(p, topVer);

    lock.readLock().lock();

    try {
      assert node2part != null && node2part.valid()
          : "Invalid node-to-partitions map [topVer1="
              + topVer
              + ", topVer2="
              + this.topVer
              + ", cache="
              + cctx.name()
              + ", node2part="
              + node2part
              + ']';

      Collection<ClusterNode> nodes = null;

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

      if (!F.isEmpty(nodeIds)) {
        Collection<UUID> affIds = new HashSet<>(F.viewReadOnly(affNodes, F.node2id()));

        for (UUID nodeId : nodeIds) {
          if (!affIds.contains(nodeId) && hasState(p, nodeId, OWNING, MOVING, RENTING)) {
            ClusterNode n = cctx.discovery().node(nodeId);

            if (n != null
                && (topVer.topologyVersion() < 0 || n.order() <= topVer.topologyVersion())) {
              if (nodes == null) {
                nodes = new ArrayList<>(affNodes.size() + 2);

                nodes.addAll(affNodes);
              }

              nodes.add(n);
            }
          }
        }
      }

      return nodes != null ? nodes : affNodes;
    } finally {
      lock.readLock().unlock();
    }
  }
  /** {@inheritDoc} */
  @Override
  public void printMemoryStats(int threshold) {
    X.println(
        ">>>  Cache partition topology stats [grid="
            + cctx.gridName()
            + ", cache="
            + cctx.name()
            + ']');

    for (GridDhtLocalPartition part : locParts.values()) {
      int size = part.size();

      if (size >= threshold)
        X.println(">>>   Local partition [part=" + part.id() + ", size=" + size + ']');
    }
  }
  /** @param cctx Context. */
  GridDhtPartitionTopologyImpl(GridCacheContext<?, ?> cctx) {
    assert cctx != null;

    this.cctx = cctx;

    log = cctx.logger(getClass());
  }
  /**
   * @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 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;
    }
  }
Ejemplo n.º 9
0
  /** 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 GridDhtPartitionMap localPartitionMap() {
    lock.readLock().lock();

    try {
      return new GridDhtPartitionMap(
          cctx.nodeId(), updateSeq.get(), F.viewReadOnly(locParts, CU.part2state()), true);
    } finally {
      lock.readLock().unlock();
    }
  }
Ejemplo n.º 11
0
  /**
   * @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);
    }
  }
Ejemplo n.º 12
0
  /**
   * @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);
  }
  /** {@inheritDoc} */
  @Override
  public void onRemoved(GridDhtCacheEntry e) {
    /*
     * Make sure not to acquire any locks here as this method
     * may be called from sensitive synchronization blocks.
     * ===================================================
     */

    GridDhtLocalPartition loc =
        localPartition(cctx.affinity().partition(e.key()), topologyVersion(), false);

    if (loc != null) loc.onRemoved(e);
  }
  /** {@inheritDoc} */
  @Override
  public AffinityTopologyVersion topologyVersion() {
    lock.readLock().lock();

    try {
      assert topVer.topologyVersion() > 0
          : "Invalid topology version [topVer=" + topVer + ", cacheName=" + cctx.name() + ']';

      return topVer;
    } 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 GridDhtLocalPartition onAdded(AffinityTopologyVersion topVer, GridDhtCacheEntry e) {
    /*
     * Make sure not to acquire any locks here as this method
     * may be called from sensitive synchronization blocks.
     * ===================================================
     */

    int p = cctx.affinity().partition(e.key());

    GridDhtLocalPartition loc = localPartition(p, topVer, true);

    assert loc != null;

    loc.onAdded(e);

    return loc;
  }
  /** {@inheritDoc} */
  @Override
  public void onEvicted(GridDhtLocalPartition part, boolean updateSeq) {
    assert updateSeq || lock.isWriteLockedByCurrentThread();

    lock.writeLock().lock();

    try {
      if (stopping) return;

      assert part.state() == EVICTED;

      long seq = updateSeq ? this.updateSeq.incrementAndGet() : this.updateSeq.get();

      updateLocal(part.id(), cctx.localNodeId(), part.state(), seq);

      consistencyCheck();
    } finally {
      lock.writeLock().unlock();
    }
  }
  /** {@inheritDoc} */
  @Override
  public boolean own(GridDhtLocalPartition part) {
    ClusterNode loc = cctx.localNode();

    lock.writeLock().lock();

    try {
      if (part.own()) {
        updateLocal(part.id(), loc.id(), part.state(), updateSeq.incrementAndGet());

        consistencyCheck();

        return true;
      }

      consistencyCheck();

      return false;
    } finally {
      lock.writeLock().unlock();
    }
  }
  /** {@inheritDoc} */
  @SuppressWarnings({"MismatchedQueryAndUpdateOfCollection"})
  @Nullable
  @Override
  public GridDhtPartitionMap update(
      @Nullable GridDhtPartitionExchangeId exchId, GridDhtPartitionFullMap partMap) {
    if (log.isDebugEnabled())
      log.debug(
          "Updating full partition map [exchId=" + exchId + ", parts=" + fullMapString() + ']');

    assert partMap != null;

    lock.writeLock().lock();

    try {
      if (stopping) return null;

      if (exchId != null && lastExchangeId != null && lastExchangeId.compareTo(exchId) >= 0) {
        if (log.isDebugEnabled())
          log.debug(
              "Stale exchange id for full partition map update (will ignore) [lastExchId="
                  + lastExchangeId
                  + ", exchId="
                  + exchId
                  + ']');

        return null;
      }

      if (node2part != null && node2part.compareTo(partMap) >= 0) {
        if (log.isDebugEnabled())
          log.debug(
              "Stale partition map for full partition map update (will ignore) [lastExchId="
                  + lastExchangeId
                  + ", exchId="
                  + exchId
                  + ", curMap="
                  + node2part
                  + ", newMap="
                  + partMap
                  + ']');

        return null;
      }

      long updateSeq = this.updateSeq.incrementAndGet();

      if (exchId != null) lastExchangeId = exchId;

      if (node2part != null) {
        for (GridDhtPartitionMap part : node2part.values()) {
          GridDhtPartitionMap newPart = partMap.get(part.nodeId());

          // If for some nodes current partition has a newer map,
          // then we keep the newer value.
          if (newPart != null && newPart.updateSequence() < part.updateSequence()) {
            if (log.isDebugEnabled())
              log.debug(
                  "Overriding partition map in full update map [exchId="
                      + exchId
                      + ", curPart="
                      + mapString(part)
                      + ", newPart="
                      + mapString(newPart)
                      + ']');

            partMap.put(part.nodeId(), part);
          }
        }

        for (Iterator<UUID> it = partMap.keySet().iterator(); it.hasNext(); ) {
          UUID nodeId = it.next();

          if (!cctx.discovery().alive(nodeId)) {
            if (log.isDebugEnabled())
              log.debug(
                  "Removing left node from full map update [nodeId="
                      + nodeId
                      + ", partMap="
                      + partMap
                      + ']');

            it.remove();
          }
        }
      }

      node2part = partMap;

      Map<Integer, Set<UUID>> p2n = new HashMap<>(cctx.affinity().partitions(), 1.0f);

      for (Map.Entry<UUID, GridDhtPartitionMap> e : partMap.entrySet()) {
        for (Integer p : e.getValue().keySet()) {
          Set<UUID> ids = p2n.get(p);

          if (ids == null)
            // Initialize HashSet to size 3 in anticipation that there won't be
            // more than 3 nodes per partitions.
            p2n.put(p, ids = U.newHashSet(3));

          ids.add(e.getKey());
        }
      }

      part2node = p2n;

      boolean changed = checkEvictions(updateSeq);

      consistencyCheck();

      if (log.isDebugEnabled()) log.debug("Partition map after full update: " + fullMapString());

      return changed ? localPartitionMap() : null;
    } finally {
      lock.writeLock().unlock();
    }
  }
 /** {@inheritDoc} */
 @Override
 public GridDhtLocalPartition localPartition(Object key, boolean create) {
   return localPartition(cctx.affinity().partition(key), AffinityTopologyVersion.NONE, create);
 }
Ejemplo n.º 21
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);
    }
  }
  /** {@inheritDoc} */
  @Override
  public List<ClusterNode> owners(int p, AffinityTopologyVersion topVer) {
    if (!cctx.rebalanceEnabled()) return ownersAndMoving(p, topVer);

    return nodes(p, topVer, OWNING);
  }
  /** {@inheritDoc} */
  @Override
  public void beforeExchange(GridDhtPartitionsExchangeFuture exchFut)
      throws IgniteCheckedException {
    waitForRent();

    ClusterNode loc = cctx.localNode();

    int num = cctx.affinity().partitions();

    lock.writeLock().lock();

    try {
      GridDhtPartitionExchangeId exchId = exchFut.exchangeId();

      if (stopping) return;

      assert topVer.equals(exchId.topologyVersion())
          : "Invalid topology version [topVer=" + topVer + ", exchId=" + exchId + ']';

      if (exchId.isLeft()) removeNode(exchId.nodeId());

      // In case if node joins, get topology at the time of joining node.
      ClusterNode oldest = CU.oldestAliveCacheServerNode(cctx.shared(), topVer);

      assert oldest != null;

      if (log.isDebugEnabled())
        log.debug(
            "Partition map beforeExchange [exchId="
                + exchId
                + ", fullMap="
                + fullMapString()
                + ']');

      long updateSeq = this.updateSeq.incrementAndGet();

      // If this is the oldest node.
      if (oldest.id().equals(loc.id())
          || exchFut.isCacheAdded(cctx.cacheId(), exchId.topologyVersion())) {
        if (node2part == null) {
          node2part = new GridDhtPartitionFullMap(oldest.id(), oldest.order(), updateSeq);

          if (log.isDebugEnabled())
            log.debug(
                "Created brand new full topology map on oldest node [exchId="
                    + exchId
                    + ", fullMap="
                    + fullMapString()
                    + ']');
        } else if (!node2part.valid()) {
          node2part =
              new GridDhtPartitionFullMap(oldest.id(), oldest.order(), updateSeq, node2part, false);

          if (log.isDebugEnabled())
            log.debug(
                "Created new full topology map on oldest node [exchId="
                    + exchId
                    + ", fullMap="
                    + node2part
                    + ']');
        } else if (!node2part.nodeId().equals(loc.id())) {
          node2part =
              new GridDhtPartitionFullMap(oldest.id(), oldest.order(), updateSeq, node2part, false);

          if (log.isDebugEnabled())
            log.debug(
                "Copied old map into new map on oldest node (previous oldest node left) [exchId="
                    + exchId
                    + ", fullMap="
                    + fullMapString()
                    + ']');
        }
      }

      if (cctx.rebalanceEnabled()) {
        for (int p = 0; p < num; p++) {
          // If this is the first node in grid.
          boolean added = exchFut.isCacheAdded(cctx.cacheId(), exchId.topologyVersion());

          if ((oldest.id().equals(loc.id())
                  && oldest.id().equals(exchId.nodeId())
                  && exchId.isJoined())
              || added) {
            assert exchId.isJoined() || added;

            try {
              GridDhtLocalPartition locPart = localPartition(p, topVer, true, false);

              assert locPart != null;

              boolean owned = locPart.own();

              assert owned
                  : "Failed to own partition for oldest node [cacheName"
                      + cctx.name()
                      + ", part="
                      + locPart
                      + ']';

              if (log.isDebugEnabled()) log.debug("Owned partition for oldest node: " + locPart);

              updateLocal(p, loc.id(), locPart.state(), updateSeq);
            } catch (GridDhtInvalidPartitionException e) {
              if (log.isDebugEnabled())
                log.debug(
                    "Ignoring invalid partition on oldest node (no need to create a partition "
                        + "if it no longer belongs to local node: "
                        + e.partition());
            }
          }
          // If this is not the first node in grid.
          else {
            if (node2part != null && node2part.valid()) {
              if (cctx.affinity().localNode(p, topVer)) {
                try {
                  // This will make sure that all non-existing partitions
                  // will be created in MOVING state.
                  GridDhtLocalPartition locPart = localPartition(p, topVer, true, false);

                  updateLocal(p, loc.id(), locPart.state(), updateSeq);
                } catch (GridDhtInvalidPartitionException e) {
                  if (log.isDebugEnabled())
                    log.debug(
                        "Ignoring invalid partition (no need to create a partition if it "
                            + "no longer belongs to local node: "
                            + e.partition());
                }
              }
            }
            // If this node's map is empty, we pre-create local partitions,
            // so local map will be sent correctly during exchange.
            else if (cctx.affinity().localNode(p, topVer)) {
              try {
                localPartition(p, topVer, true, false);
              } catch (GridDhtInvalidPartitionException e) {
                if (log.isDebugEnabled())
                  log.debug(
                      "Ignoring invalid partition (no need to pre-create a partition if it "
                          + "no longer belongs to local node: "
                          + e.partition());
              }
            }
          }
        }
      } else {
        // If preloader is disabled, then we simply clear out
        // the partitions this node is not responsible for.
        for (int p = 0; p < num; p++) {
          GridDhtLocalPartition locPart = localPartition(p, topVer, false, false);

          boolean belongs = cctx.affinity().localNode(p, topVer);

          if (locPart != null) {
            if (!belongs) {
              GridDhtPartitionState state = locPart.state();

              if (state.active()) {
                locPart.rent(false);

                updateLocal(p, loc.id(), locPart.state(), updateSeq);

                if (log.isDebugEnabled())
                  log.debug(
                      "Evicting partition with rebalancing disabled "
                          + "(it does not belong to affinity): "
                          + locPart);
              }
            }
          } else if (belongs) {
            try {
              // Pre-create partitions.
              localPartition(p, topVer, true, false);
            } catch (GridDhtInvalidPartitionException e) {
              if (log.isDebugEnabled())
                log.debug(
                    "Ignoring invalid partition with disabled rebalancer (no need to "
                        + "pre-create a partition if it no longer belongs to local node: "
                        + e.partition());
            }
          }
        }
      }

      if (node2part != null && node2part.valid()) checkEvictions(updateSeq);

      consistencyCheck();

      if (log.isDebugEnabled())
        log.debug(
            "Partition map after beforeExchange [exchId="
                + exchId
                + ", fullMap="
                + fullMapString()
                + ']');
    } finally {
      lock.writeLock().unlock();
    }

    // Wait for evictions.
    waitForRent();
  }
  /**
   * @param updateSeq Update sequence.
   * @return Checks if any of the local partitions need to be evicted.
   */
  private boolean checkEvictions(long updateSeq) {
    assert lock.isWriteLockedByCurrentThread();

    boolean changed = false;

    UUID locId = cctx.nodeId();

    for (GridDhtLocalPartition part : locParts.values()) {
      GridDhtPartitionState state = part.state();

      if (state.active()) {
        int p = part.id();

        List<ClusterNode> affNodes = cctx.affinity().nodes(p, topVer);

        if (!affNodes.contains(cctx.localNode())) {
          Collection<UUID> nodeIds = F.nodeIds(nodes(p, topVer, OWNING));

          // If all affinity nodes are owners, then evict partition from local node.
          if (nodeIds.containsAll(F.nodeIds(affNodes))) {
            part.rent(false);

            updateLocal(part.id(), locId, part.state(), updateSeq);

            changed = true;

            if (log.isDebugEnabled())
              log.debug("Evicted local partition (all affinity nodes are owners): " + part);
          } else {
            int ownerCnt = nodeIds.size();
            int affCnt = affNodes.size();

            if (ownerCnt > affCnt) {
              List<ClusterNode> sorted = new ArrayList<>(cctx.discovery().nodes(nodeIds));

              // Sort by node orders in ascending order.
              Collections.sort(sorted, CU.nodeComparator(true));

              int diff = sorted.size() - affCnt;

              for (int i = 0; i < diff; i++) {
                ClusterNode n = sorted.get(i);

                if (locId.equals(n.id())) {
                  part.rent(false);

                  updateLocal(part.id(), locId, part.state(), updateSeq);

                  changed = true;

                  if (log.isDebugEnabled())
                    log.debug(
                        "Evicted local partition (this node is oldest non-affinity node): " + part);

                  break;
                }
              }
            }
          }
        }
      }
    }

    return changed;
  }
  /**
   * 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);
  }
Ejemplo n.º 26
0
 /**
  * @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);
 }
  /** {@inheritDoc} */
  @Override
  public List<ClusterNode> moving(int p) {
    if (!cctx.rebalanceEnabled()) return ownersAndMoving(p, AffinityTopologyVersion.NONE);

    return nodes(p, AffinityTopologyVersion.NONE, MOVING);
  }
  /** {@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 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;
  }