ndb - bug#22013

    Fix bug in event handling wrt early node shutdown

ndb/src/mgmsrv/MgmtSrvr.cpp

@@ -137,8 +137,11 @@ MgmtSrvr::logLevelThreadRun()
         m_started_nodes.erase(0, false);
         m_started_nodes.unlock();
 
-        setEventReportingLevelImpl(node, req);
-
+        if (setEventReportingLevelImpl(node, req))
+	{
+	  ndbout_c("setEventReportingLevelImpl(%d): failed", node);
+	}
+	
         SetLogLevelOrd ord;
         ord = m_nodeLogLevel[node];
         setNodeLogLevelImpl(node, ord);
@@ -155,10 +158,16 @@ MgmtSrvr::logLevelThreadRun()
       m_log_level_requests.erase(0, false);
       m_log_level_requests.unlock();
 
-      if(req.blockRef == 0){
+      if(req.blockRef == 0)
+      {
 	req.blockRef = _ownReference;
-	setEventReportingLevelImpl(0, req);
-      } else {
+	if (setEventReportingLevelImpl(0, req))
+	{
+	  ndbout_c("setEventReportingLevelImpl: failed 2!");
+	}
+      } 
+      else 
+      {
         SetLogLevelOrd ord;
         ord = req;
 	setNodeLogLevelImpl(req.blockRef, ord);
@@ -1376,9 +1385,6 @@ int MgmtSrvr::restartDB(bool nostart, bool initialStart,
   NodeId nodeId = 0;
   NDB_TICKS maxTime = NdbTick_CurrentMillisecond() + waitTime;
 
-  ndbout_c(" %d", nodes.get(1));
-  ndbout_c(" %d", nodes.get(2));
-
   while(getNextNodeId(&nodeId, NDB_MGM_NODE_TYPE_NDB)) {
     if (!nodes.get(nodeId))
       continue;
@@ -1584,6 +1590,11 @@ MgmtSrvr::setEventReportingLevelImpl(int nodeId,
     }
   }
 
+  if (nodes.isclear())
+  {
+    return SEND_OR_RECEIVE_FAILED;
+  }
+  
   int error = 0;
   while (!nodes.isclear())
   {
@@ -1600,16 +1611,24 @@ MgmtSrvr::setEventReportingLevelImpl(int nodeId,
       error = 1;
       break;
     }
+      // Since sending okToSend(true), 
+      // there is no guarantee that NF_COMPLETEREP will come
+      // i.e listen also to NODE_FAILREP
+    case GSN_NODE_FAILREP: {
+      const NodeFailRep * const rep =
+	CAST_CONSTPTR(NodeFailRep, signal->getDataPtr());
+      NdbNodeBitmask mask;
+      mask.assign(NdbNodeBitmask::Size, rep->theNodes);
+      nodes.bitANDC(mask);
+      break;
+    }
+      
     case GSN_NF_COMPLETEREP:{
       const NFCompleteRep * const rep =
 	CAST_CONSTPTR(NFCompleteRep, signal->getDataPtr());
       nodes.clear(rep->failedNodeId);
       break;
     }
-    case GSN_NODE_FAILREP:{
-      // ignore, NF_COMPLETEREP will arrive later
-      break;
-    }
     default:
       report_unknown_signal(signal);
       return SEND_OR_RECEIVE_FAILED;
@@ -1909,7 +1928,10 @@ MgmtSrvr::handleStatus(NodeId nodeId, bool alive, bool nfComplete)
 
   theData[1] = nodeId;
   if (alive) {
-    m_started_nodes.push_back(nodeId);
+    if (nodeTypes[nodeId] == NODE_TYPE_DB)
+    {
+      m_started_nodes.push_back(nodeId);
+    }
     rep->setEventType(NDB_LE_Connected);
   } else {
     rep->setEventType(NDB_LE_Disconnected);

ndb/src/ndbapi/ClusterMgr.cpp

@@ -507,6 +507,7 @@ ClusterMgr::reportConnected(NodeId nodeId){
   theNode.m_info.m_version = 0;
   theNode.compatible = true;
   theNode.nfCompleteRep = true;
+  theNode.m_state.startLevel = NodeState::SL_NOTHING;
   
   theFacade.ReportNodeAlive(nodeId);
 }
@@ -518,14 +519,13 @@ ClusterMgr::reportDisconnected(NodeId nodeId){
 
   noOfConnectedNodes--;
   theNodes[nodeId].connected = false;
-
   theNodes[nodeId].m_state.m_connected_nodes.clear();
 
-  reportNodeFailed(nodeId);
+  reportNodeFailed(nodeId, true);
 }
 
 void
-ClusterMgr::reportNodeFailed(NodeId nodeId){
+ClusterMgr::reportNodeFailed(NodeId nodeId, bool disconnect){
 
   Node & theNode = theNodes[nodeId];
  
@@ -536,10 +536,11 @@ ClusterMgr::reportNodeFailed(NodeId nodeId){
   {
     theFacade.doDisconnect(nodeId);
   }
+  
   const bool report = (theNode.m_state.startLevel != NodeState::SL_NOTHING);  
   theNode.m_state.startLevel = NodeState::SL_NOTHING;
   
-  if(report)
+  if(disconnect || report)
   {
     theFacade.ReportNodeDead(nodeId);
   }

ndb/src/ndbapi/ClusterMgr.hpp

@@ -97,8 +97,8 @@ private:
   NdbMutex*     clusterMgrThreadMutex;
   
   void showState(NodeId nodeId);
-  void reportNodeFailed(NodeId nodeId);
-
+  void reportNodeFailed(NodeId nodeId, bool disconnect = false);
+  
   /**
    * Signals received
    */

ndb/src/ndbapi/SignalSender.cpp

@@ -19,6 +19,14 @@
 #include <signaldata/NFCompleteRep.hpp>
 #include <signaldata/NodeFailRep.hpp>
 
+static
+void
+require(bool x)
+{
+  if (!x)
+    abort();
+}
+
 SimpleSignal::SimpleSignal(bool dealloc){
   memset(this, 0, sizeof(* this));
   deallocSections = dealloc;
@@ -145,6 +153,7 @@ SignalSender::waitFor(Uint32 timeOutMillis, T & t)
 {
   SimpleSignal * s = t.check(m_jobBuffer);
   if(s != 0){
+    m_usedBuffer.push_back(s);
     return s;
   }