ndb - bug#15695 bug#16447 bug#18612

  For various reasone have a partitioned cluster been created
  This patch makes sure that when they connect
  1) it's detected
  2) shutdown is forced


ndb/src/kernel/blocks/cmvmi/Cmvmi.cpp:
  New dump/error insert for simulating network failure
ndb/src/kernel/blocks/qmgr/Qmgr.hpp:
  1) Activly detect paritioned cluster(s)
  2) add some documentation
ndb/src/kernel/blocks/qmgr/QmgrMain.cpp:
  1) Activly detect paritioned cluster(s)
  2) add some documentation

ndb/src/kernel/blocks/cmvmi/Cmvmi.cpp

@@ -133,6 +133,9 @@ Cmvmi::~Cmvmi()
 {
 }
 
+#ifdef ERROR_INSERT
+NodeBitmask c_error_9000_nodes_mask;
+#endif
 
 void Cmvmi::execNDB_TAMPER(Signal* signal) 
 {
@@ -390,21 +393,33 @@ void Cmvmi::execOPEN_COMREQ(Signal* signal)
 
   const Uint32 len = signal->getLength();
   if(len == 2){
-    globalTransporterRegistry.do_connect(tStartingNode);
-    globalTransporterRegistry.setIOState(tStartingNode, HaltIO);
 
-    //-----------------------------------------------------
-    // Report that the connection to the node is opened
-    //-----------------------------------------------------
-    signal->theData[0] = EventReport::CommunicationOpened;
-    signal->theData[1] = tStartingNode;
-    sendSignal(CMVMI_REF, GSN_EVENT_REP, signal, 2, JBB);
-    //-----------------------------------------------------
+#ifdef ERROR_INSERT
+    if (! (ERROR_INSERTED(9000) && c_error_9000_nodes_mask.get(tStartingNode)))
+#endif
+    {
+      globalTransporterRegistry.do_connect(tStartingNode);
+      globalTransporterRegistry.setIOState(tStartingNode, HaltIO);
+      
+      //-----------------------------------------------------
+      // Report that the connection to the node is opened
+      //-----------------------------------------------------
+      signal->theData[0] = EventReport::CommunicationOpened;
+      signal->theData[1] = tStartingNode;
+      sendSignal(CMVMI_REF, GSN_EVENT_REP, signal, 2, JBB);
+      //-----------------------------------------------------
+    }
   } else {
     for(unsigned int i = 1; i < MAX_NODES; i++ ) {
       jam();
       if (i != getOwnNodeId() && getNodeInfo(i).m_type == tData2){
 	jam();
+
+#ifdef ERROR_INSERT
+	if (ERROR_INSERTED(9000) && c_error_9000_nodes_mask.get(i))
+	  continue;
+#endif
+	
 	globalTransporterRegistry.do_connect(i);
 	globalTransporterRegistry.setIOState(i, HaltIO);
 	
@@ -1010,7 +1025,8 @@ Cmvmi::execDUMP_STATE_ORD(Signal* signal)
   }
 
   DumpStateOrd * const & dumpState = (DumpStateOrd *)&signal->theData[0];
-  if (dumpState->args[0] == DumpStateOrd::CmvmiDumpConnections){
+  Uint32 arg = dumpState->args[0];
+  if (arg == DumpStateOrd::CmvmiDumpConnections){
     for(unsigned int i = 1; i < MAX_NODES; i++ ){
       const char* nodeTypeStr = "";
       switch(getNodeInfo(i).m_type){
@@ -1043,13 +1059,13 @@ Cmvmi::execDUMP_STATE_ORD(Signal* signal)
     }
   }
   
-  if (dumpState->args[0] == DumpStateOrd::CmvmiDumpLongSignalMemory){
+  if (arg == DumpStateOrd::CmvmiDumpLongSignalMemory){
     infoEvent("Cmvmi: g_sectionSegmentPool size: %d free: %d",
 	      g_sectionSegmentPool.getSize(),
 	      g_sectionSegmentPool.getNoOfFree());
   }
   
-  if (dumpState->args[0] == DumpStateOrd::CmvmiSetRestartOnErrorInsert)
+  if (arg == DumpStateOrd::CmvmiSetRestartOnErrorInsert)
   {
     if(signal->getLength() == 1)
     {
@@ -1069,7 +1085,7 @@ Cmvmi::execDUMP_STATE_ORD(Signal* signal)
     }
   }
 
-  if (dumpState->args[0] == DumpStateOrd::CmvmiTestLongSigWithDelay) {
+  if (arg == DumpStateOrd::CmvmiTestLongSigWithDelay) {
     unsigned i;
     Uint32 loopCount = dumpState->args[1];
     const unsigned len0 = 11;
@@ -1097,6 +1113,30 @@ Cmvmi::execDUMP_STATE_ORD(Signal* signal)
     sendSignal(reference(), GSN_TESTSIG, signal, 8, JBB, ptr, 2);
   }
 
+#ifdef ERROR_INSERT
+  if (arg == 9000)
+  {
+    SET_ERROR_INSERT_VALUE(9000);
+    for (Uint32 i = 1; i<signal->getLength(); i++)
+      c_error_9000_nodes_mask.set(signal->theData[i]);
+  }
+  
+  if (arg == 9001)
+  {
+    CLEAR_ERROR_INSERT_VALUE;
+    for (Uint32 i = 0; i<MAX_NODES; i++)
+    {
+      if (c_error_9000_nodes_mask.get(i))
+      {
+	signal->theData[0] = 0;
+	signal->theData[1] = i;
+	EXECUTE_DIRECT(CMVMI, GSN_OPEN_COMREQ, signal, 2);
+      }
+    }
+    c_error_9000_nodes_mask.clear();
+  }
+#endif
+
 #ifdef VM_TRACE
 #if 0
   {

ndb/src/kernel/blocks/qmgr/Qmgr.hpp

@@ -100,7 +100,12 @@ public:
   };
 
   struct StartRecord {
-    void reset(){ m_startKey++; m_startNode = 0;}
+    void reset(){ 
+      m_startKey++; 
+      m_startNode = 0; 
+      m_gsn = RNIL; 
+      m_nodes.clearWaitingFor();
+    }
     Uint32 m_startKey;
     Uint32 m_startNode;
     Uint64 m_startTimeout;
@@ -112,6 +117,14 @@ public:
   NdbNodeBitmask c_definedNodes; // DB nodes in config
   NdbNodeBitmask c_clusterNodes; // DB nodes in cluster
   NodeBitmask c_connectedNodes;  // All kinds of connected nodes
+
+  /**
+   * Nodes which we're checking for partitioned cluster
+   *
+   * i.e. nodes that connect to use, when we already have elected president
+   */
+  NdbNodeBitmask c_cmregreq_nodes;
+  
   Uint32 c_maxDynamicId;
   
   // Records
@@ -251,8 +264,10 @@ private:
 
   // Generated statement blocks
   void startphase1(Signal* signal);
-  void electionWon();
+  void electionWon(Signal* signal);
   void cmInfoconf010Lab(Signal* signal);
+  bool check_cmregreq_reply(Signal* signal, Uint32 nodeId, Uint32 gsn);
+  
   void apiHbHandlingLab(Signal* signal);
   void timerHandlingLab(Signal* signal);
   void hbReceivedLab(Signal* signal);