/** {@inheritDoc} */
  @Override
  public final void stop(boolean cancel) {
    if (!starting.get() || !stop.compareAndSet(false, true))
      // Ignoring attempt to stop manager that has never been started.
      return;

    stop0(cancel);

    if (log != null && log.isDebugEnabled()) log.debug(stopInfo());
  }
  /** {@inheritDoc} */
  @Override
  public final void onKernalStop(boolean cancel) {
    if (!starting.get())
      // Ignoring attempt to stop manager that has never been started.
      return;

    onKernalStop0(cancel);

    if (log != null && log.isDebugEnabled()) log.debug(kernalStopInfo());
  }
  /** {@inheritDoc} */
  @Override
  public final void start(GridCacheSharedContext<K, V> cctx) throws IgniteCheckedException {
    if (!starting.compareAndSet(false, true))
      assert false : "Method start is called more than once for manager: " + this;

    assert cctx != null;

    this.cctx = cctx;

    log = cctx.logger(getClass());

    start0();

    if (log.isDebugEnabled()) log.debug(startInfo());
  }
  /**
   * 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;
  }
 /** @return {@code True} */
 public boolean onAdded() {
   return added.compareAndSet(false, true);
 }
  /**
   * JUnit.
   *
   * @throws Exception If failed.
   */
  @SuppressWarnings({"TooBroadScope"})
  public void testRestarts() throws Exception {
    int duration = 60 * 1000;
    int qryThreadNum = 10;
    final long nodeLifeTime = 2 * 1000;
    final int logFreq = 20;

    final IgniteCache<Integer, Integer> cache = grid(0).cache(null);

    assert cache != null;

    for (int i = 0; i < KEY_CNT; i++) cache.put(i, i);

    assertEquals(KEY_CNT, cache.localSize());

    final AtomicInteger qryCnt = new AtomicInteger();

    final AtomicBoolean done = new AtomicBoolean();

    IgniteInternalFuture<?> fut1 =
        multithreadedAsync(
            new CAX() {
              @Override
              public void applyx() throws IgniteCheckedException {
                while (!done.get()) {
                  Collection<Cache.Entry<Integer, Integer>> res =
                      cache.query(new SqlQuery(Integer.class, "_val >= 0")).getAll();

                  assertFalse(res.isEmpty());

                  int c = qryCnt.incrementAndGet();

                  if (c % logFreq == 0) info("Executed queries: " + c);
                }
              }
            },
            qryThreadNum);

    final AtomicInteger restartCnt = new AtomicInteger();

    CollectingEventListener lsnr = new CollectingEventListener();

    for (int i = 0; i < GRID_CNT; i++)
      grid(i).events().localListen(lsnr, EventType.EVT_CACHE_REBALANCE_STOPPED);

    IgniteInternalFuture<?> fut2 =
        multithreadedAsync(
            new Callable<Object>() {
              @SuppressWarnings({"BusyWait"})
              @Override
              public Object call() throws Exception {
                while (!done.get()) {
                  int idx = GRID_CNT;

                  startGrid(idx);

                  Thread.sleep(nodeLifeTime);

                  stopGrid(idx);

                  int c = restartCnt.incrementAndGet();

                  if (c % logFreq == 0) info("Node restarts: " + c);
                }

                return true;
              }
            },
            1);

    Thread.sleep(duration);

    done.set(true);

    fut1.get();
    fut2.get();

    info("Awaiting rebalance events [restartCnt=" + restartCnt.get() + ']');

    boolean success = lsnr.awaitEvents(GRID_CNT * 2 * restartCnt.get(), 15000);

    for (int i = 0; i < GRID_CNT; i++)
      grid(i).events().stopLocalListen(lsnr, EventType.EVT_CACHE_REBALANCE_STOPPED);

    assert success;
  }
  /**
   * Grabs local events and detects if events was lost since last poll.
   *
   * @param ignite Target grid.
   * @param evtOrderKey Unique key to take last order key from node local map.
   * @param evtThrottleCntrKey Unique key to take throttle count from node local map.
   * @param evtTypes Event types to collect.
   * @param evtMapper Closure to map grid events to Visor data transfer objects.
   * @return Collections of node events
   */
  public static Collection<VisorGridEvent> collectEvents(
      Ignite ignite,
      String evtOrderKey,
      String evtThrottleCntrKey,
      final int[] evtTypes,
      IgniteClosure<Event, VisorGridEvent> evtMapper) {
    assert ignite != null;
    assert evtTypes != null && evtTypes.length > 0;

    ConcurrentMap<String, Long> nl = ignite.cluster().nodeLocalMap();

    final long lastOrder = getOrElse(nl, evtOrderKey, -1L);
    final long throttle = getOrElse(nl, evtThrottleCntrKey, 0L);

    // When we first time arrive onto a node to get its local events,
    // we'll grab only last those events that not older than given period to make sure we are
    // not grabbing GBs of data accidentally.
    final long notOlderThan = System.currentTimeMillis() - EVENTS_COLLECT_TIME_WINDOW;

    // Flag for detecting gaps between events.
    final AtomicBoolean lastFound = new AtomicBoolean(lastOrder < 0);

    IgnitePredicate<Event> p =
        new IgnitePredicate<Event>() {
          /** */
          private static final long serialVersionUID = 0L;

          @Override
          public boolean apply(Event e) {
            // Detects that events were lost.
            if (!lastFound.get() && (lastOrder == e.localOrder())) lastFound.set(true);

            // Retains events by lastOrder, period and type.
            return e.localOrder() > lastOrder
                && e.timestamp() > notOlderThan
                && F.contains(evtTypes, e.type());
          }
        };

    Collection<Event> evts = ignite.events().localQuery(p);

    // Update latest order in node local, if not empty.
    if (!evts.isEmpty()) {
      Event maxEvt = Collections.max(evts, EVTS_ORDER_COMPARATOR);

      nl.put(evtOrderKey, maxEvt.localOrder());
    }

    // Update throttle counter.
    if (!lastFound.get())
      nl.put(evtThrottleCntrKey, throttle == 0 ? EVENTS_LOST_THROTTLE : throttle - 1);

    boolean lost = !lastFound.get() && throttle == 0;

    Collection<VisorGridEvent> res = new ArrayList<>(evts.size() + (lost ? 1 : 0));

    if (lost) res.add(new VisorGridEventsLost(ignite.cluster().localNode().id()));

    for (Event e : evts) {
      VisorGridEvent visorEvt = evtMapper.apply(e);

      if (visorEvt != null) res.add(visorEvt);
    }

    return res;
  }