Merge perch.ndb.mysql.com:/home/jonas/src/50-work

into  perch.ndb.mysql.com:/home/jonas/src/50-jonas

ndb/include/kernel/signaldata/CmRegSignalData.hpp

@@ -30,12 +30,17 @@ class CmRegReq {
   friend class Qmgr;
   
 public:
-  STATIC_CONST( SignalLength = 3 );
+  STATIC_CONST( SignalLength = 5 + NdbNodeBitmask::Size );
 private:
   
   Uint32 blockRef;
   Uint32 nodeId;
   Uint32 version;    // See ndb_version.h
+
+  Uint32 start_type; // As specified by cmd-line or mgm, NodeState::StartType
+  Uint32 latest_gci; // 0 means no fs
+  Uint32 skip_nodes[NdbNodeBitmask::Size]; // Nodes that does not _need_ 
+                                           // to be part of restart
 };
 
 /**
@@ -59,7 +64,6 @@ private:
    * The dynamic id that the node reciving this signal has
    */
   Uint32 dynamicId;
-  
   Uint32 allNdbNodes[NdbNodeBitmask::Size];  
 };
 
@@ -73,7 +77,7 @@ class CmRegRef {
   friend class Qmgr;
   
 public:
-  STATIC_CONST( SignalLength = 4 );
+  STATIC_CONST( SignalLength = 7 + NdbNodeBitmask::Size );
   
   enum ErrorCode {
     ZBUSY = 0,          /* Only the president can send this */
@@ -85,14 +89,27 @@ public:
                          * as president. */
     ZNOT_PRESIDENT = 5, /* We are not president */
     ZNOT_DEAD = 6,       /* We are not dead when we are starting  */
-    ZINCOMPATIBLE_VERSION = 7
+    ZINCOMPATIBLE_VERSION = 7,
+    ZINCOMPATIBLE_START_TYPE = 8
   };
 private:
   
   Uint32 blockRef;
   Uint32 nodeId;
   Uint32 errorCode;
+  /**
+   * Applicable if ZELECTION
+   */
   Uint32 presidentCandidate;
+  Uint32 candidate_latest_gci; // 0 means non
+
+  /**
+   * Data for sending node sending node
+   */
+  Uint32 latest_gci; 
+  Uint32 start_type; 
+  Uint32 skip_nodes[NdbNodeBitmask::Size]; // Nodes that does not _need_ 
+                                           // to be part of restart
 };
 
 class CmAdd {

ndb/include/kernel/signaldata/DumpStateOrd.hpp

@@ -64,6 +64,7 @@ public:
     // 19 NDBFS Fipple with O_SYNC, O_CREATE etc.
     // 20-24 BACKUP
     NdbcntrTestStopOnError = 25,
+    NdbcntrStopNodes = 70,
     // 100-105 TUP and ACC  
     // 200-240 UTIL
     // 300-305 TRIX

ndb/include/kernel/signaldata/FailRep.hpp

@@ -18,6 +18,7 @@
 #define FAIL_REP_HPP
 
 #include "SignalData.hpp"
+#include <NodeBitmask.hpp>
 
 /**
  * 
@@ -27,6 +28,7 @@ class FailRep {
    * Sender(s) & Reciver(s)
    */
   friend class Qmgr;
+  friend class Ndbcntr;
   
   /**
    * For printing
@@ -35,6 +37,7 @@ class FailRep {
 
 public:
   STATIC_CONST( SignalLength = 2 );
+  STATIC_CONST( ExtraLength = 1 + NdbNodeBitmask::Size );
   
   enum FailCause {
     ZOWN_FAILURE=0,
@@ -43,13 +46,20 @@ public:
     ZSTART_IN_REGREQ=3,
     ZHEARTBEAT_FAILURE=4,
     ZLINK_FAILURE=5,
-    ZOTHERNODE_FAILED_DURING_START=6
+    ZOTHERNODE_FAILED_DURING_START=6,
+    ZMULTI_NODE_SHUTDOWN = 7,
+    ZPARTITIONED_CLUSTER = 8
   };
   
 private:
   
   Uint32 failNodeId;
   Uint32 failCause;
+  /**
+   * Used when failCause == ZPARTITIONED_CLUSTER
+   */
+  Uint32 president;
+  Uint32 partition[NdbNodeBitmask::Size];
 };
 
 

ndb/include/kernel/signaldata/StopReq.hpp

@@ -32,7 +32,7 @@ class StopReq
   friend class MgmtSrvr;
 
 public:
-  STATIC_CONST( SignalLength = 9 );
+  STATIC_CONST( SignalLength = 9 + NdbNodeBitmask::Size);
   
 public:
   Uint32 senderRef;
@@ -49,29 +49,34 @@ public:
   Int32 readOperationTimeout; // Timeout before read operations are aborted
   Int32 operationTimeout;     // Timeout before all operations are aborted
 
+  Uint32 nodes[NdbNodeBitmask::Size];
+
   static void setSystemStop(Uint32 & requestInfo, bool value);
   static void setPerformRestart(Uint32 & requestInfo, bool value);
   static void setNoStart(Uint32 & requestInfo, bool value);
   static void setInitialStart(Uint32 & requestInfo, bool value);
-  static void setEscalateOnNodeFail(Uint32 & requestInfo, bool value);
   /**
    * Don't perform "graceful" shutdown/restart...
    */
   static void setStopAbort(Uint32 & requestInfo, bool value);
+  static void setStopNodes(Uint32 & requestInfo, bool value);
 
   static bool getSystemStop(const Uint32 & requestInfo);
   static bool getPerformRestart(const Uint32 & requestInfo);
   static bool getNoStart(const Uint32 & requestInfo);
   static bool getInitialStart(const Uint32 & requestInfo);
-  static bool getEscalateOnNodeFail(const Uint32 & requestInfo);
   static bool getStopAbort(const Uint32 & requestInfo);
+  static bool getStopNodes(const Uint32 & requestInfo);
 };
 
 struct StopConf
 {
   STATIC_CONST( SignalLength = 2 );
   Uint32 senderData;
+  union {
     Uint32 nodeState;
+    Uint32 nodeId;
+  };
 };
 
 class StopRef 
@@ -94,7 +99,9 @@ public:
     NodeShutdownInProgress = 1,
     SystemShutdownInProgress = 2,
     NodeShutdownWouldCauseSystemCrash = 3,
-    TransactionAbortFailed = 4
+    TransactionAbortFailed = 4,
+    UnsupportedNodeShutdown = 5,
+    MultiNodeShutdownNotMaster = 6
   };
   
 public:
@@ -132,16 +139,16 @@ StopReq::getInitialStart(const Uint32 & requestInfo)
 
 inline
 bool
-StopReq::getEscalateOnNodeFail(const Uint32 & requestInfo)
+StopReq::getStopAbort(const Uint32 & requestInfo)
 {
-  return requestInfo & 16;
+  return requestInfo & 32;
 }
 
 inline
 bool
-StopReq::getStopAbort(const Uint32 & requestInfo)
+StopReq::getStopNodes(const Uint32 & requestInfo)
 {
-  return requestInfo & 32;
+  return requestInfo & 64;
 }
 
 
@@ -187,24 +194,23 @@ StopReq::setInitialStart(Uint32 & requestInfo, bool value)
 
 inline
 void
-StopReq::setEscalateOnNodeFail(Uint32 & requestInfo, bool value)
+StopReq::setStopAbort(Uint32 & requestInfo, bool value)
 {
   if(value)
-    requestInfo |= 16;
+    requestInfo |= 32;
   else
-    requestInfo &= ~16;
+    requestInfo &= ~32;
 }
 
 inline
 void
-StopReq::setStopAbort(Uint32 & requestInfo, bool value)
+StopReq::setStopNodes(Uint32 & requestInfo, bool value)
 {
   if(value)
-    requestInfo |= 32;
+    requestInfo |= 64;
   else
-    requestInfo &= ~32;
+    requestInfo &= ~64;
 }
 
-
 #endif
 

ndb/include/kernel/signaldata/WaitGCP.hpp

@@ -46,7 +46,9 @@ public:
     Complete = 1,           ///< Wait for a GCP to complete
     CompleteForceStart = 2, ///< Wait for a GCP to complete start one if needed
     CompleteIfRunning = 3,  ///< Wait for ongoing GCP
-    CurrentGCI        = 8   ///< Immediately return current GCI
+    CurrentGCI        = 8,  ///< Immediately return current GCI
+    BlockStartGcp     = 9,
+    UnblockStartGcp   = 10
   };
 
   Uint32 senderRef;
@@ -70,11 +72,12 @@ class WaitGCPConf {
   //friend class Grep::PSCoord;
 
 public:
-  STATIC_CONST( SignalLength = 2 );
+  STATIC_CONST( SignalLength = 3 );
   
 public:
   Uint32 senderData;
   Uint32 gcp;
+  Uint32 blockStatus;
 };
 
 class WaitGCPRef {

ndb/include/mgmapi/ndb_logevent.h

@@ -166,10 +166,14 @@ extern "C" {
     /** NDB_MGM_EVENT_CATEGORY_BACKUP */
     NDB_LE_BackupCompleted = 56,
     /** NDB_MGM_EVENT_CATEGORY_BACKUP */
-    NDB_LE_BackupAborted = 57
+    NDB_LE_BackupAborted = 57,
 
     /* 58 used in 5.1  */
     /* 59 used */
+
+    /** NDB_MGM_EVENT_CATEGORY_STARTUP */
+    NDB_LE_StartReport = 60
+
     /* 60 unused */
     /* 61 unused */
     /* 62 unused */
@@ -625,6 +629,13 @@ extern "C" {
         unsigned type;
         unsigned node_id;
       } SingleUser;
+      /** Log even data @ref NDB_LE_StartReport */
+      struct {
+	unsigned report_type;
+	unsigned remaining_time;
+	unsigned bitmask_size;
+	unsigned bitmask_data[1];
+      } StartReport;
 #ifndef DOXYGEN_FIX
     };
 #else

ndb/src/common/debugger/EventLogger.cpp

@@ -707,6 +707,90 @@ void getTextSingleUser(QQQQ) {
   }
 }
 
+void getTextStartReport(QQQQ) {
+  Uint32 time = theData[2];
+  Uint32 sz = theData[3];
+  char mask1[100];
+  char mask2[100];
+  char mask3[100];
+  char mask4[100];
+  BitmaskImpl::getText(sz, theData + 4 + (0 * sz), mask1);
+  BitmaskImpl::getText(sz, theData + 4 + (1 * sz), mask2);
+  BitmaskImpl::getText(sz, theData + 4 + (2 * sz), mask3);
+  BitmaskImpl::getText(sz, theData + 4 + (3 * sz), mask4);
+  switch(theData[1]){
+  case 1: // Wait initial
+    BaseString::snprintf
+      (m_text, m_text_len,
+       "Initial start, waiting for %s to connect, "
+       " nodes [ all: %s connected: %s no-wait: %s ]",
+       mask4, mask1, mask2, mask3);
+    break;
+  case 2: // Wait partial
+    BaseString::snprintf
+      (m_text, m_text_len,
+       "Waiting until nodes: %s connects, "
+       "nodes [ all: %s connected: %s no-wait: %s ]",
+       mask4, mask1, mask2, mask3);
+    break;
+  case 3: // Wait partial timeout
+    BaseString::snprintf
+      (m_text, m_text_len,
+       "Waiting %u sec for nodes %s to connect, "
+       "nodes [ all: %s connected: %s no-wait: %s ]",
+       
+       time, mask4, mask1, mask2, mask3);
+    break;
+  case 4: // Wait partioned
+    BaseString::snprintf
+      (m_text, m_text_len,
+       "Waiting for non partitioned start, "
+       "nodes [ all: %s connected: %s missing: %s no-wait: %s ]",
+       
+       mask1, mask2, mask4, mask3);
+    break;
+  case 5:
+    BaseString::snprintf
+      (m_text, m_text_len,
+       "Waiting %u sec for non partitioned start, "
+       "nodes [ all: %s connected: %s missing: %s no-wait: %s ]",
+       
+       time, mask1, mask2, mask4, mask3);
+    break;
+  case 0x8000: // Do initial
+    BaseString::snprintf
+      (m_text, m_text_len,
+       "Initial start with nodes %s [ missing: %s no-wait: %s ]",
+       mask2, mask4, mask3);
+    break;
+  case 0x8001: // Do start
+    BaseString::snprintf
+      (m_text, m_text_len,
+       "Start with all nodes %s",
+       mask2);
+    break;
+  case 0x8002: // Do partial
+    BaseString::snprintf
+      (m_text, m_text_len,
+       "Start with nodes %s [ missing: %s no-wait: %s ]",
+       mask2, mask4, mask3);
+    break;
+  case 0x8003: // Do partioned
+    BaseString::snprintf
+      (m_text, m_text_len,
+       "Start potentially partitioned with nodes %s "
+       " [ missing: %s no-wait: %s ]",
+       mask2, mask4, mask3);
+    break;
+  default:
+    BaseString::snprintf
+      (m_text, m_text_len,
+       "Unknown startreport: 0x%x [ %s %s %s %s ]", 
+       theData[1],
+       mask1, mask2, mask3, mask4);
+  }
+}
+
 #if 0
 BaseString::snprintf(m_text, 
 		     m_text_len, 
@@ -755,6 +839,7 @@ const EventLoggerBase::EventRepLogLevelMatrix EventLoggerBase::matrix[] = {
   ROW(StartREDOLog,            LogLevel::llStartUp,    10, Logger::LL_INFO ),
   ROW(StartLog,                LogLevel::llStartUp,    10, Logger::LL_INFO ),
   ROW(UNDORecordsExecuted,     LogLevel::llStartUp,    15, Logger::LL_INFO ),
+  ROW(StartReport,             LogLevel::llStartUp,     4, Logger::LL_INFO ),
   
   // NODERESTART
   ROW(NR_CopyDict,             LogLevel::llNodeRestart, 8, Logger::LL_INFO ),

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

@@ -134,6 +134,9 @@ Cmvmi::~Cmvmi()
 {
 }
 
+#ifdef ERROR_INSERT
+NodeBitmask c_error_9000_nodes_mask;
+#endif
 
 void Cmvmi::execNDB_TAMPER(Signal* signal) 
 {
@@ -419,6 +422,11 @@ void Cmvmi::execOPEN_COMREQ(Signal* signal)
 
   const Uint32 len = signal->getLength();
   if(len == 2){
+
+#ifdef ERROR_INSERT
+    if (! (ERROR_INSERTED(9000) && c_error_9000_nodes_mask.get(tStartingNode)))
+#endif
+    {
       globalTransporterRegistry.do_connect(tStartingNode);
       globalTransporterRegistry.setIOState(tStartingNode, HaltIO);
       
@@ -429,11 +437,18 @@ void Cmvmi::execOPEN_COMREQ(Signal* signal)
       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);
 	
@@ -1039,7 +1054,8 @@ 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){
@@ -1072,13 +1088,13 @@ 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)
     {
@@ -1098,7 +1114,7 @@ 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;
@@ -1126,6 +1142,30 @@ 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/dbdih/DbdihInit.cpp

@@ -71,6 +71,7 @@ void Dbdih::initData()
   cwaitLcpSr       = false;
   c_blockCommit    = false;
   c_blockCommitNo  = 1;
+  cntrlblockref    = RNIL;
 }//Dbdih::initData()
 
 void Dbdih::initRecords() 

ndb/src/kernel/blocks/dbdih/DbdihMain.cpp

@@ -11659,7 +11659,7 @@ void Dbdih::makeNodeGroups(Uint32 nodeArray[])
   Uint32 tmngNode;
   Uint32 tmngNodeGroup;
   Uint32 tmngLimit;
-  Uint32 i;
+  Uint32 i, j;
 
   /**-----------------------------------------------------------------------
    * ASSIGN ALL ACTIVE NODES INTO NODE GROUPS. HOT SPARE NODES ARE ASSIGNED 
@@ -11705,6 +11705,38 @@ void Dbdih::makeNodeGroups(Uint32 nodeArray[])
       Sysfile::setNodeGroup(mngNodeptr.i, SYSFILE->nodeGroups, mngNodeptr.p->nodeGroup);
     }//if
   }//for
+
+  for (i = 0; i<cnoOfNodeGroups; i++)
+  {
+    jam();
+    bool alive = false;
+    NodeGroupRecordPtr NGPtr;
+    NGPtr.i = i;
+    ptrCheckGuard(NGPtr, MAX_NDB_NODES, nodeGroupRecord);
+    for (j = 0; j<NGPtr.p->nodeCount; j++)
+    {
+      jam();
+      mngNodeptr.i = NGPtr.p->nodesInGroup[j];
+      ptrCheckGuard(mngNodeptr, MAX_NDB_NODES, nodeRecord);
+      if (checkNodeAlive(NGPtr.p->nodesInGroup[j]))
+      {
+	alive = true;
+	break;
+      }
+    }
+
+    if (!alive)
+    {
+      char buf[255];
+      BaseString::snprintf
+	(buf, sizeof(buf), 
+	 "Illegal initial start, no alive node in nodegroup %u", i);
+      progError(__LINE__, 
+		NDBD_EXIT_SR_RESTARTCONFLICT,
+		buf);
+      
+    }
+  }
 }//Dbdih::makeNodeGroups()
 
 /**
@@ -12513,7 +12545,6 @@ void Dbdih::sendStartFragreq(Signal* signal,
 void Dbdih::setInitialActiveStatus()
 {
   NodeRecordPtr siaNodeptr;
-  Uint32 tsiaNodeActiveStatus;
   Uint32 tsiaNoActiveNodes;
 
   tsiaNoActiveNodes = csystemnodes - cnoHotSpare;
@@ -12521,39 +12552,34 @@ void Dbdih::setInitialActiveStatus()
     SYSFILE->nodeStatus[i] = 0;
   for (siaNodeptr.i = 1; siaNodeptr.i < MAX_NDB_NODES; siaNodeptr.i++) {
     ptrAss(siaNodeptr, nodeRecord);
-    if (siaNodeptr.p->nodeStatus == NodeRecord::ALIVE) {
+    switch(siaNodeptr.p->nodeStatus){
+    case NodeRecord::ALIVE:
+    case NodeRecord::DEAD:
       if (tsiaNoActiveNodes == 0) {
         jam();
         siaNodeptr.p->activeStatus = Sysfile::NS_HotSpare;
       } else {
         jam();
         tsiaNoActiveNodes = tsiaNoActiveNodes - 1;
-        siaNodeptr.p->activeStatus = Sysfile::NS_Active;
-      }//if
-    } else {
-      jam();
-      siaNodeptr.p->activeStatus = Sysfile::NS_NotDefined;
-    }//if
-    switch (siaNodeptr.p->activeStatus) {
-    case Sysfile::NS_Active:
-      jam();
-      tsiaNodeActiveStatus = Sysfile::NS_Active;
-      break;
-    case Sysfile::NS_HotSpare:
-      jam();
-      tsiaNodeActiveStatus = Sysfile::NS_HotSpare;
-      break;
-    case Sysfile::NS_NotDefined:
+        if (siaNodeptr.p->nodeStatus == NodeRecord::ALIVE)
+	{
 	  jam();
-      tsiaNodeActiveStatus = Sysfile::NS_NotDefined;
+	  siaNodeptr.p->activeStatus = Sysfile::NS_Active;
+	} 
+	else
+	{
+	  siaNodeptr.p->activeStatus = Sysfile::NS_NotActive_NotTakenOver;
+	}
+      }
       break;
     default:
-      ndbrequire(false);
-      return;
+      jam();
+      siaNodeptr.p->activeStatus = Sysfile::NS_NotDefined;
       break;
-    }//switch
-    Sysfile::setNodeStatus(siaNodeptr.i, SYSFILE->nodeStatus,
-                           tsiaNodeActiveStatus);
+    }//if
+    Sysfile::setNodeStatus(siaNodeptr.i, 
+			   SYSFILE->nodeStatus,
+                           siaNodeptr.p->activeStatus);
   }//for
 }//Dbdih::setInitialActiveStatus()
 
@@ -14274,11 +14300,36 @@ void Dbdih::execWAIT_GCP_REQ(Signal* signal)
     jam();
     conf->senderData = senderData;
     conf->gcp = cnewgcp;
+    conf->blockStatus = cgcpOrderBlocked;
     sendSignal(senderRef, GSN_WAIT_GCP_CONF, signal, 
 	       WaitGCPConf::SignalLength, JBB);
     return;
   }//if
 
+  if (requestType == WaitGCPReq::BlockStartGcp)
+  {
+    jam();
+    conf->senderData = senderData;
+    conf->gcp = cnewgcp;
+    conf->blockStatus = cgcpOrderBlocked;
+    sendSignal(senderRef, GSN_WAIT_GCP_CONF, signal, 
+	       WaitGCPConf::SignalLength, JBB);
+    cgcpOrderBlocked = 1;
+    return;
+  }
+
+  if (requestType == WaitGCPReq::UnblockStartGcp)
+  {
+    jam();
+    conf->senderData = senderData;
+    conf->gcp = cnewgcp;
+    conf->blockStatus = cgcpOrderBlocked;
+    sendSignal(senderRef, GSN_WAIT_GCP_CONF, signal, 
+	       WaitGCPConf::SignalLength, JBB);
+    cgcpOrderBlocked = 0;
+    return;
+  }
+  
   if(isMaster()) {
     /**
      * Master
@@ -14290,6 +14341,7 @@ void Dbdih::execWAIT_GCP_REQ(Signal* signal)
       jam();
       conf->senderData = senderData;
       conf->gcp = coldgcp;
+      conf->blockStatus = cgcpOrderBlocked;
       sendSignal(senderRef, GSN_WAIT_GCP_CONF, signal, 
 		 WaitGCPConf::SignalLength, JBB);
       return;
@@ -14376,6 +14428,7 @@ void Dbdih::execWAIT_GCP_CONF(Signal* signal)
 
   conf->senderData = ptr.p->clientData;
   conf->gcp = gcp;
+  conf->blockStatus = cgcpOrderBlocked;
   sendSignal(ptr.p->clientRef, GSN_WAIT_GCP_CONF, signal,
 	     WaitGCPConf::SignalLength, JBB);
   
@@ -14443,6 +14496,7 @@ void Dbdih::emptyWaitGCPMasterQueue(Signal* signal)
 
     c_waitGCPMasterList.next(ptr);    
     conf->senderData = clientData;
+    conf->blockStatus = cgcpOrderBlocked;
     sendSignal(clientRef, GSN_WAIT_GCP_CONF, signal,
 	       WaitGCPConf::SignalLength, JBB);
     

ndb/src/kernel/blocks/ndbcntr/Ndbcntr.hpp

@@ -203,6 +203,7 @@ private:
   void execWAIT_GCP_CONF(Signal* signal);
 
   void execSTOP_REQ(Signal* signal);
+  void execSTOP_CONF(Signal* signal);
   void execRESUME_REQ(Signal* signal);
 
   void execCHANGE_NODE_STATE_CONF(Signal* signal);
@@ -338,6 +339,16 @@ public:
     void progError(int line, int cause, const char * extra) { 
       cntr.progError(line, cause, extra); 
     }
+
+    enum StopNodesStep {
+      SR_BLOCK_GCP_START_GCP = 0,
+      SR_WAIT_COMPLETE_GCP = 1,
+      SR_UNBLOCK_GCP_START_GCP = 2,
+      SR_QMGR_STOP_REQ = 3,
+      SR_WAIT_NODE_FAILURES = 4,
+      SR_CLUSTER_SHUTDOWN = 12
+    } m_state;
+    SignalCounter m_stop_req_counter;
   };
 private:
   StopRecord c_stopRec;

ndb/src/kernel/blocks/ndbcntr/NdbcntrInit.cpp

@@ -87,6 +87,7 @@ Ndbcntr::Ndbcntr(const class Configuration & conf):
   addRecSignal(GSN_STOP_ME_CONF, &Ndbcntr::execSTOP_ME_CONF);
 
   addRecSignal(GSN_STOP_REQ, &Ndbcntr::execSTOP_REQ);
+  addRecSignal(GSN_STOP_CONF, &Ndbcntr::execSTOP_CONF);
   addRecSignal(GSN_RESUME_REQ, &Ndbcntr::execRESUME_REQ);
 
   addRecSignal(GSN_WAIT_GCP_REF, &Ndbcntr::execWAIT_GCP_REF);

ndb/src/kernel/blocks/ndbcntr/NdbcntrMain.cpp

@@ -42,6 +42,8 @@
 #include <signaldata/FsRemoveReq.hpp>
 #include <signaldata/ReadConfig.hpp>
 
+#include <signaldata/FailRep.hpp>
+
 #include <AttributeHeader.hpp>
 #include <Configuration.hpp>
 #include <DebuggerNames.hpp>
@@ -818,17 +820,9 @@ Ndbcntr::trySystemRestart(Signal* signal){
     return false;
   }
   
-  if(!allNodes && c_start.m_startPartialTimeout > now){
-    jam();
-    return false;
-  }
-
   NodeState::StartType srType = NodeState::ST_SYSTEM_RESTART;
-  if(c_start.m_waiting.equal(c_start.m_withoutLog)){
-    if(!allNodes){
-      jam();
-      return false;
-    }
+  if(c_start.m_waiting.equal(c_start.m_withoutLog))
+  {
     jam();
     srType = NodeState::ST_INITIAL_START;
     c_start.m_starting = c_start.m_withoutLog; // Used for starting...
@@ -858,10 +852,6 @@ Ndbcntr::trySystemRestart(Signal* signal){
 	ndbrequire(false); // All nodes -> partitioning, which is not allowed
       }
       
-      if(c_start.m_startPartitionedTimeout > now){
-	jam();
-	return false;
-      }
       break;
     }    
     
@@ -1474,13 +1464,74 @@ void Ndbcntr::execNODE_FAILREP(Signal* signal)
   sendSignal(SUMA_REF, GSN_NODE_FAILREP, signal,
 	     NodeFailRep::SignalLength, JBB);
 
+  if (c_stopRec.stopReq.senderRef)
+  {
+    jam();
+    switch(c_stopRec.m_state){
+    case StopRecord::SR_WAIT_NODE_FAILURES:
+    {
+      jam();
+      NdbNodeBitmask tmp;
+      tmp.assign(NdbNodeBitmask::Size, c_stopRec.stopReq.nodes);
+      tmp.bitANDC(allFailed);      
+      tmp.copyto(NdbNodeBitmask::Size, c_stopRec.stopReq.nodes);
+      
+      if (tmp.isclear())
+      {
+	jam();
+	if (c_stopRec.stopReq.senderRef != RNIL)
+	{
+	  jam();
+	  StopConf * const stopConf = (StopConf *)&signal->theData[0];
+	  stopConf->senderData = c_stopRec.stopReq.senderData;
+	  stopConf->nodeState  = (Uint32) NodeState::SL_SINGLEUSER;
+	  sendSignal(c_stopRec.stopReq.senderRef, GSN_STOP_CONF, signal, 
+		     StopConf::SignalLength, JBB);
+	}
+
+	c_stopRec.stopReq.senderRef = 0;
+	WaitGCPReq * req = (WaitGCPReq*)&signal->theData[0];
+	req->senderRef = reference();
+	req->senderData = StopRecord::SR_UNBLOCK_GCP_START_GCP;
+	req->requestType = WaitGCPReq::UnblockStartGcp;
+	sendSignal(DBDIH_REF, GSN_WAIT_GCP_REQ, signal, 
+		   WaitGCPReq::SignalLength, JBA);
+      }
+      break;
+    }
+    case StopRecord::SR_QMGR_STOP_REQ:
+    {
+      NdbNodeBitmask tmp;
+      tmp.assign(NdbNodeBitmask::Size, c_stopRec.stopReq.nodes);
+      tmp.bitANDC(allFailed);      
+
+      if (tmp.isclear())
+      {
+	Uint32 nodeId = allFailed.find(0);
+	tmp.set(nodeId);
+
+	StopConf* conf = (StopConf*)signal->getDataPtrSend();
+	conf->senderData = c_stopRec.stopReq.senderData;
+	conf->nodeId = nodeId;
+	sendSignal(reference(), 
+		   GSN_STOP_CONF, signal, StopConf::SignalLength, JBB);
+      }
+
+      tmp.copyto(NdbNodeBitmask::Size, c_stopRec.stopReq.nodes);
+      
+      break;
+    }
+    }
+  }
+  
+  signal->theData[0] = NDB_LE_NODE_FAILREP;
+  signal->theData[2] = 0;
+  
   Uint32 nodeId = 0;
   while(!allFailed.isclear()){
     nodeId = allFailed.find(nodeId + 1);
     allFailed.clear(nodeId);
-    signal->theData[0] = NDB_LE_NODE_FAILREP;
     signal->theData[1] = nodeId;
-    signal->theData[2] = 0;
     sendSignal(CMVMI_REF, GSN_EVENT_REP, signal, 3, JBB);
   }//for
 
@@ -1924,13 +1975,15 @@ void
 Ndbcntr::execDUMP_STATE_ORD(Signal* signal)
 {
   DumpStateOrd * const & dumpState = (DumpStateOrd *)&signal->theData[0];
-  if(signal->theData[0] == 13){
+  Uint32 arg = dumpState->args[0];
+
+  if(arg == 13){
     infoEvent("Cntr: cstartPhase = %d, cinternalStartphase = %d, block = %d", 
 	      cstartPhase, cinternalStartphase, cndbBlocksCount);
     infoEvent("Cntr: cmasterNodeId = %d", cmasterNodeId);
   }
 
-  if (dumpState->args[0] == DumpStateOrd::NdbcntrTestStopOnError){
+  if (arg == DumpStateOrd::NdbcntrTestStopOnError){
     if (theConfiguration.stopOnError() == true)
       ((Configuration&)theConfiguration).stopOnError(false);
     
@@ -1943,6 +1996,28 @@ Ndbcntr::execDUMP_STATE_ORD(Signal* signal)
 	       SystemError::SignalLength, JBA);
   }
 
+  if (arg == DumpStateOrd::NdbcntrStopNodes)
+  {
+    NdbNodeBitmask mask;
+    for(Uint32 i = 1; i<signal->getLength(); i++)
+      mask.set(signal->theData[i]);
+
+    StopReq* req = (StopReq*)signal->getDataPtrSend();
+    req->senderRef = RNIL;
+    req->senderData = 123;
+    req->requestInfo = 0;
+    req->singleuser = 0;
+    req->singleUserApi = 0;
+    mask.copyto(NdbNodeBitmask::Size, req->nodes);
+    StopReq::setPerformRestart(req->requestInfo, 1);
+    StopReq::setNoStart(req->requestInfo, 1);
+    StopReq::setStopNodes(req->requestInfo, 1);
+    StopReq::setStopAbort(req->requestInfo, 1);
+    
+    sendSignal(reference(), GSN_STOP_REQ, signal,
+	       StopReq::SignalLength, JBB);
+    return;
+  }
 
 }//Ndbcntr::execDUMP_STATE_ORD()
 
@@ -2003,9 +2078,12 @@ Ndbcntr::execSTOP_REQ(Signal* signal){
   Uint32 senderData = req->senderData;
   BlockReference senderRef = req->senderRef;
   bool abort = StopReq::getStopAbort(req->requestInfo);
+  bool stopnodes = StopReq::getStopNodes(req->requestInfo);
 
-  if(getNodeState().startLevel < NodeState::SL_STARTED || 
-     abort && !singleuser){
+  if(!singleuser && 
+     (getNodeState().startLevel < NodeState::SL_STARTED || 
+      (abort && !stopnodes)))
+  {
     /**
      * Node is not started yet
      *
@@ -2047,6 +2125,28 @@ Ndbcntr::execSTOP_REQ(Signal* signal){
     else
       ref->errorCode = StopRef::NodeShutdownInProgress;
     ref->senderData = senderData;
+    
+    if (senderRef != RNIL)
+      sendSignal(senderRef, GSN_STOP_REF, signal, StopRef::SignalLength, JBB);
+    return;
+  }
+
+  if (stopnodes && !abort)
+  {
+    jam();
+    ref->errorCode = StopRef::UnsupportedNodeShutdown;
+    ref->senderData = senderData;
+    if (senderRef != RNIL)
+      sendSignal(senderRef, GSN_STOP_REF, signal, StopRef::SignalLength, JBB);
+    return;
+  }
+
+  if (stopnodes && cmasterNodeId != getOwnNodeId())
+  {
+    jam();
+    ref->errorCode = StopRef::MultiNodeShutdownNotMaster;
+    ref->senderData = senderData;
+    if (senderRef != RNIL)
       sendSignal(senderRef, GSN_STOP_REF, signal, StopRef::SignalLength, JBB);
     return;
   }
@@ -2054,14 +2154,42 @@ Ndbcntr::execSTOP_REQ(Signal* signal){
   c_stopRec.stopReq = * req;
   c_stopRec.stopInitiatedTime = NdbTick_CurrentMillisecond();
   
-  if(!singleuser) {
-    if(StopReq::getSystemStop(c_stopRec.stopReq.requestInfo)) {
+  if (stopnodes)
+  {
     jam();
-      if(StopReq::getPerformRestart(c_stopRec.stopReq.requestInfo)){
+
+    if(!c_stopRec.checkNodeFail(signal))
+    {
+      jam();
+      return;
+    }
+
+    char buf[100];
+    NdbNodeBitmask mask;
+    mask.assign(NdbNodeBitmask::Size, c_stopRec.stopReq.nodes);
+    infoEvent("Initiating shutdown abort of %s", mask.getText(buf));
+    ndbout_c("Initiating shutdown abort of %s", mask.getText(buf));    
+
+    WaitGCPReq * req = (WaitGCPReq*)&signal->theData[0];
+    req->senderRef = reference();
+    req->senderData = StopRecord::SR_BLOCK_GCP_START_GCP;
+    req->requestType = WaitGCPReq::BlockStartGcp;
+    sendSignal(DBDIH_REF, GSN_WAIT_GCP_REQ, signal, 
+	       WaitGCPReq::SignalLength, JBB);
+    return;
+  }
+  else if(!singleuser) 
+  {
+    if(StopReq::getSystemStop(c_stopRec.stopReq.requestInfo)) 
+    {
+      jam();
+      if(StopReq::getPerformRestart(c_stopRec.stopReq.requestInfo))
+      {
 	((Configuration&)theConfiguration).stopOnError(false);
       }
     }
-    if(!c_stopRec.checkNodeFail(signal)){
+    if(!c_stopRec.checkNodeFail(signal))
+    {
       jam();
       return;
     }
@@ -2131,7 +2259,17 @@ Ndbcntr::StopRecord::checkNodeFail(Signal* signal){
    */
   NodeBitmask ndbMask; 
   ndbMask.assign(cntr.c_startedNodes);
+
+  if (StopReq::getStopNodes(stopReq.requestInfo))
+  {
+    NdbNodeBitmask tmp;
+    tmp.assign(NdbNodeBitmask::Size, stopReq.nodes);
+    ndbMask.bitANDC(tmp);
+  }
+  else
+  {
     ndbMask.clear(cntr.getOwnNodeId());
+  }
   
   CheckNodeGroups* sd = (CheckNodeGroups*)&signal->theData[0];
   sd->blockRef = cntr.reference();
@@ -2153,6 +2291,7 @@ Ndbcntr::StopRecord::checkNodeFail(Signal* signal){
   ref->errorCode = StopRef::NodeShutdownWouldCauseSystemCrash;
   
   const BlockReference bref = stopReq.senderRef;
+  if (bref != RNIL)
     cntr.sendSignal(bref, GSN_STOP_REF, signal, StopRef::SignalLength, JBB);
   
   stopReq.senderRef = 0;
@@ -2215,7 +2354,7 @@ Ndbcntr::StopRecord::checkTcTimeout(Signal* signal){
       {
 	WaitGCPReq * req = (WaitGCPReq*)&signal->theData[0];
 	req->senderRef = cntr.reference();
-	  req->senderData = 12;
+	req->senderData = StopRecord::SR_CLUSTER_SHUTDOWN;
 	req->requestType = WaitGCPReq::CompleteForceStart;
 	cntr.sendSignal(DBDIH_REF, GSN_WAIT_GCP_REQ, signal, 
 			WaitGCPReq::SignalLength, JBB);
@@ -2381,7 +2520,7 @@ void Ndbcntr::execWAIT_GCP_REF(Signal* signal){
 
   WaitGCPReq * req = (WaitGCPReq*)&signal->theData[0];
   req->senderRef = reference();
-  req->senderData = 12;
+  req->senderData = StopRecord::SR_CLUSTER_SHUTDOWN;
   req->requestType = WaitGCPReq::CompleteForceStart;
   sendSignal(DBDIH_REF, GSN_WAIT_GCP_REQ, signal, 
 	     WaitGCPReq::SignalLength, JBB);
@@ -2390,6 +2529,63 @@ void Ndbcntr::execWAIT_GCP_REF(Signal* signal){
 void Ndbcntr::execWAIT_GCP_CONF(Signal* signal){
   jamEntry();
 
+  WaitGCPConf* conf = (WaitGCPConf*)signal->getDataPtr();
+
+  switch(conf->senderData){
+  case StopRecord::SR_BLOCK_GCP_START_GCP:
+  {
+    jam();
+    /**
+     * 
+     */
+    if(!c_stopRec.checkNodeFail(signal))
+    {
+      jam();
+      goto unblock;
+    }
+    
+    WaitGCPReq * req = (WaitGCPReq*)&signal->theData[0];
+    req->senderRef = reference();
+    req->senderData = StopRecord::SR_WAIT_COMPLETE_GCP;
+    req->requestType = WaitGCPReq::CompleteIfRunning;
+
+    sendSignal(DBDIH_REF, GSN_WAIT_GCP_REQ, signal, 
+	       WaitGCPReq::SignalLength, JBB);
+    return;
+  }
+  case StopRecord::SR_UNBLOCK_GCP_START_GCP:
+  {
+    jam();
+    return;
+  }
+  case StopRecord::SR_WAIT_COMPLETE_GCP:
+  {
+    jam();
+    if(!c_stopRec.checkNodeFail(signal))
+    {
+      jam();
+      goto unblock;
+    }
+
+    NdbNodeBitmask tmp;
+    tmp.assign(NdbNodeBitmask::Size, c_stopRec.stopReq.nodes);
+    c_stopRec.m_stop_req_counter = tmp;
+    NodeReceiverGroup rg(QMGR, tmp);
+    StopReq * stopReq = (StopReq *)&signal->theData[0];
+    * stopReq = c_stopRec.stopReq;
+    stopReq->senderRef = reference();
+    sendSignal(rg, GSN_STOP_REQ, signal, StopReq::SignalLength, JBA);
+    c_stopRec.m_state = StopRecord::SR_QMGR_STOP_REQ; 
+    return;
+  }
+  case StopRecord::SR_CLUSTER_SHUTDOWN:
+  {
+    jam();
+    break;
+  }
+  }
+  
+  {  
     ndbrequire(StopReq::getSystemStop(c_stopRec.stopReq.requestInfo));
     NodeState newState(NodeState::SL_STOPPING_3, true); 
     
@@ -2400,7 +2596,8 @@ void Ndbcntr::execWAIT_GCP_CONF(Signal* signal){
     rep->nodeState = newState;
     rep->nodeState.masterNodeId = cmasterNodeId;
     rep->nodeState.setNodeGroup(c_nodeGroup);
-  EXECUTE_DIRECT(QMGR, GSN_NODE_STATE_REP, signal, NodeStateRep::SignalLength);
+    EXECUTE_DIRECT(QMGR, GSN_NODE_STATE_REP, signal, 
+		   NodeStateRep::SignalLength);
     
     if(StopReq::getPerformRestart(c_stopRec.stopReq.requestInfo)){
       jam();
@@ -2413,6 +2610,48 @@ void Ndbcntr::execWAIT_GCP_CONF(Signal* signal){
       sendSignalWithDelay(CMVMI_REF, GSN_STOP_ORD, signal, 500, 1);
     }
     return;
+  }
+  
+unblock:
+  WaitGCPReq * req = (WaitGCPReq*)&signal->theData[0];
+  req->senderRef = reference();
+  req->senderData = StopRecord::SR_UNBLOCK_GCP_START_GCP;
+  req->requestType = WaitGCPReq::UnblockStartGcp;
+  sendSignal(DBDIH_REF, GSN_WAIT_GCP_REQ, signal, 
+	     WaitGCPReq::SignalLength, JBB);
+}
+
+void
+Ndbcntr::execSTOP_CONF(Signal* signal)
+{
+  jamEntry();
+  StopConf *conf = (StopConf*)signal->getDataPtr();
+  ndbrequire(c_stopRec.m_state == StopRecord::SR_QMGR_STOP_REQ);
+  c_stopRec.m_stop_req_counter.clearWaitingFor(conf->nodeId);
+  if (c_stopRec.m_stop_req_counter.done())
+  {
+    char buf[100];
+    NdbNodeBitmask mask;
+    mask.assign(NdbNodeBitmask::Size, c_stopRec.stopReq.nodes);
+    infoEvent("Stopping of %s", mask.getText(buf));
+    ndbout_c("Stopping of %s", mask.getText(buf));    
+
+    /**
+     * Kill any node...
+     */
+    FailRep * const failRep = (FailRep *)&signal->theData[0];
+    failRep->failCause = FailRep::ZMULTI_NODE_SHUTDOWN;
+    NodeReceiverGroup rg(QMGR, c_clusterNodes);
+    Uint32 nodeId = 0;
+    while ((nodeId = NdbNodeBitmask::find(c_stopRec.stopReq.nodes, nodeId+1))
+	   != NdbNodeBitmask::NotFound)
+    {
+      failRep->failNodeId = nodeId;
+      sendSignal(rg, GSN_FAIL_REP, signal, FailRep::SignalLength, JBA);
+    }
+    c_stopRec.m_state = StopRecord::SR_WAIT_NODE_FAILURES;
+    return;
+  }
 }
 
 void Ndbcntr::execSTTORRY(Signal* signal){

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

@@ -29,6 +29,7 @@
 #include <signaldata/CmRegSignalData.hpp>
 #include <signaldata/ApiRegSignalData.hpp>
 #include <signaldata/FailRep.hpp>
+#include <signaldata/StopReq.hpp>
 
 #include "timer.hpp"
 
@@ -49,6 +50,7 @@
 #define ZAPI_HB_HANDLING 3
 #define ZTIMER_HANDLING 4
 #define ZARBIT_HANDLING 5
+#define ZSTART_FAILURE_LIMIT 6
 
 /* Error Codes ------------------------------*/
 #define ZERRTOOMANY 1101
@@ -100,18 +102,42 @@ 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;
     
     Uint32 m_gsn;
     SignalCounter m_nodes;
+    Uint32 m_latest_gci;
+
+    Uint32 m_start_type;
+    NdbNodeBitmask m_skip_nodes;
+    NdbNodeBitmask m_starting_nodes;
+    NdbNodeBitmask m_starting_nodes_w_log;
+
+    Uint16 m_president_candidate;
+    Uint32 m_president_candidate_gci;
+    Uint16 m_regReqReqSent;
+    Uint16 m_regReqReqRecv;
   } c_start;
   
   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_readnodes_nodes;
+
   Uint32 c_maxDynamicId;
   
   // Records
@@ -204,6 +230,7 @@ private:
   void execPRES_TOCONF(Signal* signal);
   void execDISCONNECT_REP(Signal* signal);
   void execSYSTEM_ERROR(Signal* signal);
+  void execSTOP_REQ(Signal* signal);
 
   // Received signals
   void execDUMP_STATE_ORD(Signal* signal);
@@ -218,6 +245,11 @@ private:
   void execREAD_NODESREQ(Signal* signal);
   void execSET_VAR_REQ(Signal* signal);
 
+  void execREAD_NODESREF(Signal* signal);
+  void execREAD_NODESCONF(Signal* signal);
+
+  void execDIH_RESTARTREF(Signal* signal);
+  void execDIH_RESTARTCONF(Signal* signal);
   
   void execAPI_VERSION_REQ(Signal* signal);
   void execAPI_BROADCAST_REP(Signal* signal);
@@ -234,6 +266,9 @@ private:
   void execARBIT_STOPREP(Signal* signal);
 
   // Statement blocks
+  void check_readnodes_reply(Signal* signal, Uint32 nodeId, Uint32 gsn);
+  Uint32 check_startup(Signal* signal);
+
   void node_failed(Signal* signal, Uint16 aFailedNode);
   void checkStartInterface(Signal* signal);
   void failReport(Signal* signal,
@@ -251,8 +286,9 @@ private:
 
   // Generated statement blocks
   void startphase1(Signal* signal);
-  void electionWon();
+  void electionWon(Signal* signal);
   void cmInfoconf010Lab(Signal* signal);
+  
   void apiHbHandlingLab(Signal* signal);
   void timerHandlingLab(Signal* signal);
   void hbReceivedLab(Signal* signal);
@@ -354,12 +390,12 @@ private:
   /* Status flags ----------------------------------*/
 
   Uint32 c_restartPartialTimeout;
+  Uint32 c_restartPartionedTimeout;
+  Uint32 c_restartFailureTimeout;
+  Uint64 c_start_election_time;
 
   Uint16 creadyDistCom;
-  Uint16 c_regReqReqSent;
-  Uint16 c_regReqReqRecv;
-  Uint64 c_stopElectionTime;
-  Uint16 cpresidentCandidate;
+
   Uint16 cdelayRegreq;
   Uint16 cpresidentAlive;
   Uint16 cnoFailedNodes;
@@ -388,6 +424,8 @@ private:
   Uint16 cprepFailedNodes[MAX_NDB_NODES];
   Uint16 ccommitFailedNodes[MAX_NDB_NODES];
   
+  StopReq c_stopReq;
+  void check_multi_node_shutdown(Signal* signal);
 };
 
 #endif

ndb/src/kernel/blocks/qmgr/QmgrInit.cpp

@@ -35,9 +35,8 @@ void Qmgr::initData()
 
   Uint32 hbDBAPI = 500;
   setHbApiDelay(hbDBAPI);
-
-  c_connectedNodes.clear();
   c_connectedNodes.set(getOwnNodeId());
+  c_stopReq.senderRef = 0;
 }//Qmgr::initData()
 
 void Qmgr::initRecords() 
@@ -52,6 +51,7 @@ Qmgr::Qmgr(const class Configuration & conf)
 
   // Transit signals
   addRecSignal(GSN_DUMP_STATE_ORD, &Qmgr::execDUMP_STATE_ORD);
+  addRecSignal(GSN_STOP_REQ, &Qmgr::execSTOP_REQ);
   addRecSignal(GSN_DEBUG_SIG, &Qmgr::execDEBUG_SIG);
   addRecSignal(GSN_CONTINUEB, &Qmgr::execCONTINUEB);
   addRecSignal(GSN_CM_HEARTBEAT, &Qmgr::execCM_HEARTBEAT);
@@ -96,6 +96,12 @@ Qmgr::Qmgr(const class Configuration & conf)
   addRecSignal(GSN_ARBIT_CHOOSEREF, &Qmgr::execARBIT_CHOOSEREF);
   addRecSignal(GSN_ARBIT_STOPREP, &Qmgr::execARBIT_STOPREP);
 
+  addRecSignal(GSN_READ_NODESREF, &Qmgr::execREAD_NODESREF);
+  addRecSignal(GSN_READ_NODESCONF, &Qmgr::execREAD_NODESCONF);
+
+  addRecSignal(GSN_DIH_RESTARTREF, &Qmgr::execDIH_RESTARTREF);
+  addRecSignal(GSN_DIH_RESTARTCONF, &Qmgr::execDIH_RESTARTCONF);
+  
   initData();
 }//Qmgr::Qmgr()
 

ndb/src/kernel/blocks/qmgr/QmgrMain.cpp

@@ -56,6 +56,33 @@
 #define DEBUG_START3(signal, msg)
 #endif
 
+/**
+ * c_start.m_gsn = GSN_CM_REGREQ
+ *   Possible for all nodes
+ *   c_start.m_nodes contains all nodes in config
+ *
+ * c_start.m_gsn = GSN_CM_NODEINFOREQ;
+ *   Set when receiving CM_REGCONF
+ *   State possible for starting node only (not in cluster)
+ *
+ *   c_start.m_nodes contains all node in alive cluster that
+ *                   that has not replied to GSN_CM_NODEINFOREQ
+ *                   passed by president in GSN_CM_REGCONF
+ *
+ * c_start.m_gsn = GSN_CM_ADD
+ *   Possible for president only
+ *   Set when receiving and accepting CM_REGREQ (to include node)
+ *
+ *   c_start.m_nodes contains all nodes in alive cluster + starting node
+ *                   that has not replied to GSN_CM_ADD
+ *                   by sending GSN_CM_ACKADD
+ *
+ * c_start.m_gsn = GSN_CM_NODEINFOCONF
+ *   Possible for non presidents only
+ *     c_start.m_nodes contains a node that has been accepted by president
+ *     but has not connected to us yet
+ */
+
 // Signal entries and statement blocks
 /* 4  P R O G R A M        */
 /*******************************/
@@ -119,6 +146,30 @@ void Qmgr::execCONTINUEB(Signal* signal)
     runArbitThread(signal);
     return;
     break;
+  case ZSTART_FAILURE_LIMIT:{
+    if (cpresident != ZNIL)
+    {
+      jam();
+      return;
+    }
+    Uint64 now = NdbTick_CurrentMillisecond();
+    if (now > (c_start_election_time + c_restartFailureTimeout))
+    {
+      jam();
+      BaseString tmp;
+      tmp.append("Shutting down node as total restart time exceeds "
+		 " StartFailureTimeout as set in config file ");
+      if(c_restartFailureTimeout == ~0)
+	tmp.append(" 0 (inifinite)");
+      else
+	tmp.appfmt(" %d", c_restartFailureTimeout);
+      
+      progError(__LINE__, NDBD_EXIT_SYSTEM_ERROR, tmp.c_str());
+    }
+    signal->theData[0] = ZSTART_FAILURE_LIMIT;
+    sendSignalWithDelay(reference(), GSN_CONTINUEB, signal, 3000, 1);
+    return;
+  }
   default:
     jam();
     // ZCOULD_NOT_OCCUR_ERROR;
@@ -247,13 +298,27 @@ void Qmgr::startphase1(Signal* signal)
   nodePtr.p->blockRef = reference();
   c_connectedNodes.set(nodePtr.i);
   
-  signal->theData[0] = 0; // no answer
-  signal->theData[1] = 0; // no id
-  signal->theData[2] = NodeInfo::DB;
-  sendSignal(CMVMI_REF, GSN_OPEN_COMREQ, signal, 3, JBB);
+  signal->theData[0] = reference();
+  sendSignal(DBDIH_REF, GSN_DIH_RESTARTREQ, signal, 1, JBB);
+  return;
+}
+
+void
+Qmgr::execDIH_RESTARTREF(Signal*signal)
+{
+  jamEntry();
 
+  c_start.m_latest_gci = 0;
+  execCM_INFOCONF(signal);
+}
+
+void
+Qmgr::execDIH_RESTARTCONF(Signal*signal)
+{
+  jamEntry();
+  
+  c_start.m_latest_gci = signal->theData[1];
   execCM_INFOCONF(signal);
-  return;
 }
 
 void Qmgr::setHbDelay(UintR aHbDelay)
@@ -280,18 +345,24 @@ void Qmgr::execCONNECT_REP(Signal* signal)
 {
   jamEntry();
   const Uint32 nodeId = signal->theData[0];
+
+  if (ERROR_INSERTED(931))
+  {
+    jam();
+    ndbout_c("Discarding CONNECT_REP(%d)", nodeId);
+    infoEvent("Discarding CONNECT_REP(%d)", nodeId);
+    return;
+  }
+  
   c_connectedNodes.set(nodeId);
   NodeRecPtr nodePtr;
   nodePtr.i = getOwnNodeId();
   ptrCheckGuard(nodePtr, MAX_NODES, nodeRec);
   switch(nodePtr.p->phase){
-  case ZSTARTING:
   case ZRUNNING:
+    ndbrequire(!c_clusterNodes.get(nodeId));
+  case ZSTARTING:
     jam();
-    if(!c_start.m_nodes.isWaitingFor(nodeId)){
-      jam();
-      return;
-    }
     break;
   case ZPREPARE_FAIL:
   case ZFAIL_CLOSING:
@@ -304,19 +375,44 @@ void Qmgr::execCONNECT_REP(Signal* signal)
     return;
   }
 
+  if (getNodeInfo(nodeId).getType() != NodeInfo::DB)
+  {
+    jam();
+    return;
+  }
+
   switch(c_start.m_gsn){
   case GSN_CM_REGREQ:
     jam();
     sendCmRegReq(signal, nodeId);
+
+    /**
+     * We're waiting for CM_REGCONF c_start.m_nodes contains all configured
+     *   nodes
+     */
+    ndbrequire(nodePtr.p->phase == ZSTARTING);
+    ndbrequire(c_start.m_nodes.isWaitingFor(nodeId));
     return;
   case GSN_CM_NODEINFOREQ:
     jam();
+
+    if (c_start.m_nodes.isWaitingFor(nodeId))
+    {
+      jam();
+      ndbrequire(getOwnNodeId() != cpresident);
+      ndbrequire(nodePtr.p->phase == ZSTARTING);
       sendCmNodeInfoReq(signal, nodeId, nodePtr.p);
       return;
-  case GSN_CM_ADD:{
+    }
+    return;
+  case GSN_CM_NODEINFOCONF:{
     jam();
     
     ndbrequire(getOwnNodeId() != cpresident);
+    ndbrequire(nodePtr.p->phase == ZRUNNING);
+    if (c_start.m_nodes.isWaitingFor(nodeId))
+    {
+      jam();
       c_start.m_nodes.clearWaitingFor(nodeId);
       c_start.m_gsn = RNIL;
       
@@ -326,36 +422,78 @@ void Qmgr::execCONNECT_REP(Signal* signal)
       cmAddPrepare(signal, addNodePtr, nodePtr.p);
       return;
     }
+  }
   default:
-    return;
+    (void)1;
   }
+  
+  ndbrequire(!c_start.m_nodes.isWaitingFor(nodeId));
+  ndbrequire(!c_readnodes_nodes.get(nodeId));
+  c_readnodes_nodes.set(nodeId);
+  signal->theData[0] = reference();
+  sendSignal(calcQmgrBlockRef(nodeId), GSN_READ_NODESREQ, signal, 1, JBA);
   return;
 }//Qmgr::execCONNECT_REP()
 
+void
+Qmgr::execREAD_NODESCONF(Signal* signal)
+{
+  check_readnodes_reply(signal, 
+			refToNode(signal->getSendersBlockRef()),
+			GSN_READ_NODESCONF);
+}
+
+void
+Qmgr::execREAD_NODESREF(Signal* signal)
+{
+  check_readnodes_reply(signal, 
+			refToNode(signal->getSendersBlockRef()),
+			GSN_READ_NODESREF);
+}
+
 /*******************************/
 /* CM_INFOCONF                */
 /*******************************/
 void Qmgr::execCM_INFOCONF(Signal* signal) 
 {
+  /**
+   * Open communcation to all DB nodes
+   */
+  signal->theData[0] = 0; // no answer
+  signal->theData[1] = 0; // no id
+  signal->theData[2] = NodeInfo::DB;
+  sendSignal(CMVMI_REF, GSN_OPEN_COMREQ, signal, 3, JBB);
+
   cpresident = ZNIL;
-  cpresidentCandidate = getOwnNodeId();
   cpresidentAlive = ZFALSE;
-  c_stopElectionTime = NdbTick_CurrentMillisecond();
-  c_stopElectionTime += c_restartPartialTimeout;
+  c_start_election_time = NdbTick_CurrentMillisecond();
+  
+  signal->theData[0] = ZSTART_FAILURE_LIMIT;
+  sendSignalWithDelay(reference(), GSN_CONTINUEB, signal, 3000, 1);
+  
   cmInfoconf010Lab(signal);
   
   return;
 }//Qmgr::execCM_INFOCONF()
 
+Uint32 g_start_type = 0;
+NdbNodeBitmask g_nowait_nodes; // Set by clo
+
 void Qmgr::cmInfoconf010Lab(Signal* signal) 
 {
   c_start.m_startKey = 0;
   c_start.m_startNode = getOwnNodeId();
   c_start.m_nodes.clearWaitingFor();
   c_start.m_gsn = GSN_CM_REGREQ;
+  c_start.m_starting_nodes.clear();
+  c_start.m_starting_nodes_w_log.clear();
+  c_start.m_regReqReqSent = 0;
+  c_start.m_regReqReqRecv = 0;
+  c_start.m_skip_nodes = g_nowait_nodes;
+  c_start.m_skip_nodes.bitAND(c_definedNodes);
+  c_start.m_start_type = g_start_type;
 
   NodeRecPtr nodePtr;
-  c_regReqReqSent = c_regReqReqRecv = 0;
   cnoOfNodes = 0;
   for (nodePtr.i = 1; nodePtr.i < MAX_NDB_NODES; nodePtr.i++) {
     jam();
@@ -390,14 +528,18 @@ void Qmgr::cmInfoconf010Lab(Signal* signal)
 
 void
 Qmgr::sendCmRegReq(Signal * signal, Uint32 nodeId){
-  c_regReqReqSent++;
-  CmRegReq * const cmRegReq = (CmRegReq *)&signal->theData[0];
-  cmRegReq->blockRef = reference();
-  cmRegReq->nodeId   = getOwnNodeId();
-  cmRegReq->version  = NDB_VERSION;
+  CmRegReq * req = (CmRegReq *)&signal->theData[0];
+  req->blockRef = reference();
+  req->nodeId = getOwnNodeId();
+  req->version = NDB_VERSION;
+  req->latest_gci = c_start.m_latest_gci;
+  req->start_type = c_start.m_start_type;
+  c_start.m_skip_nodes.copyto(NdbNodeBitmask::Size, req->skip_nodes);
   const Uint32 ref = calcQmgrBlockRef(nodeId);
   sendSignal(ref, GSN_CM_REGREQ, signal, CmRegReq::SignalLength, JBB);
   DEBUG_START(GSN_CM_REGREQ, nodeId, "");
+  
+  c_start.m_regReqReqSent++;
 }
 
 /*
@@ -437,6 +579,18 @@ Qmgr::sendCmRegReq(Signal * signal, Uint32 nodeId){
 /*******************************/
 /* CM_REGREQ                  */
 /*******************************/
+static
+int
+check_start_type(Uint32 starting, Uint32 own)
+{
+  if (starting == (1 << NodeState::ST_INITIAL_START) &&
+      ((own & (1 << NodeState::ST_INITIAL_START)) == 0))
+  {
+    return 1;
+  }
+  return 0;
+}
+
 void Qmgr::execCM_REGREQ(Signal* signal) 
 {
   DEBUG_START3(signal, "");
@@ -448,6 +602,17 @@ void Qmgr::execCM_REGREQ(Signal* signal)
   const BlockReference Tblockref = cmRegReq->blockRef;
   const Uint32 startingVersion = cmRegReq->version;
   addNodePtr.i = cmRegReq->nodeId;
+  Uint32 gci = 1;
+  Uint32 start_type = ~0;
+  NdbNodeBitmask skip_nodes;
+
+  if (signal->getLength() == CmRegReq::SignalLength)
+  {
+    jam();
+    gci = cmRegReq->latest_gci;
+    start_type = cmRegReq->start_type;
+    skip_nodes.assign(NdbNodeBitmask::Size, cmRegReq->skip_nodes);
+  }
   
   if (creadyDistCom == ZFALSE) {
     jam();
@@ -461,11 +626,19 @@ void Qmgr::execCM_REGREQ(Signal* signal)
     return;
   }
 
-  ptrCheckGuard(addNodePtr, MAX_NDB_NODES, nodeRec);
+  if (check_start_type(start_type, c_start.m_start_type))
+  {
+    jam();
+    sendCmRegrefLab(signal, Tblockref, CmRegRef::ZINCOMPATIBLE_START_TYPE);
+    return;
+  }
   
-  if (cpresident != getOwnNodeId()){
+  if (cpresident != getOwnNodeId())
+  {
     jam();                      
-    if (cpresident == ZNIL) {
+
+    if (cpresident == ZNIL) 
+    {
       /***
        * We don't know the president. 
        * If the node to be added has lower node id 
@@ -473,13 +646,19 @@ void Qmgr::execCM_REGREQ(Signal* signal)
        * candidate
        */
       jam(); 
-      if (addNodePtr.i < cpresidentCandidate) {
+      if (gci > c_start.m_president_candidate_gci || 
+	  (gci == c_start.m_president_candidate_gci && 
+	   addNodePtr.i < c_start.m_president_candidate))
+      {
 	jam();
-	cpresidentCandidate = addNodePtr.i;
-      }//if
+	c_start.m_president_candidate = addNodePtr.i;
+	c_start.m_president_candidate_gci = gci;
+	ndbout_c("assign candidate: %u %u", addNodePtr.i, gci);
+      }
       sendCmRegrefLab(signal, Tblockref, CmRegRef::ZELECTION);
       return;
     }                          
+    
     /**
      * We are not the president.
      * We know the president.
@@ -489,7 +668,8 @@ void Qmgr::execCM_REGREQ(Signal* signal)
     return;
   }//if
 
-  if (c_start.m_startNode != 0){
+  if (c_start.m_startNode != 0)
+  {
     jam();
     /**
      * President busy by adding another node
@@ -498,7 +678,8 @@ void Qmgr::execCM_REGREQ(Signal* signal)
     return;
   }//if
   
-  if (ctoStatus == Q_ACTIVE) {   
+  if (ctoStatus == Q_ACTIVE) 
+  {   
     jam();
     /**
      * Active taking over as president
@@ -507,7 +688,8 @@ void Qmgr::execCM_REGREQ(Signal* signal)
     return;
   }//if
 
-  if (getNodeInfo(addNodePtr.i).m_type != NodeInfo::DB) {
+  if (getNodeInfo(addNodePtr.i).m_type != NodeInfo::DB) 
+  {
     jam(); 
     /** 
      * The new node is not in config file
@@ -516,13 +698,15 @@ void Qmgr::execCM_REGREQ(Signal* signal)
     return;
   } 
   
+  ptrCheckGuard(addNodePtr, MAX_NDB_NODES, nodeRec);
   Phase phase = addNodePtr.p->phase;
-  if (phase != ZINIT){
+  if (phase != ZINIT)
+  {
     jam();
     DEBUG("phase = " << phase);
     sendCmRegrefLab(signal, Tblockref, CmRegRef::ZNOT_DEAD);
     return;
-  }//if
+  }
   
   jam(); 
   /**
@@ -594,7 +778,12 @@ void Qmgr::sendCmRegrefLab(Signal* signal, BlockReference TBRef,
   ref->blockRef = reference();
   ref->nodeId = getOwnNodeId();
   ref->errorCode = Terror;
-  ref->presidentCandidate = (cpresident == ZNIL ? cpresidentCandidate : cpresident);
+  ref->presidentCandidate = 
+    (cpresident == ZNIL ? c_start.m_president_candidate : cpresident);
+  ref->candidate_latest_gci = c_start.m_president_candidate_gci;
+  ref->latest_gci = c_start.m_latest_gci;
+  ref->start_type = c_start.m_start_type;
+  c_start.m_skip_nodes.copyto(NdbNodeBitmask::Size, ref->skip_nodes);
   sendSignal(TBRef, GSN_CM_REGREF, signal, 
 	     CmRegRef::SignalLength, JBB);
   DEBUG_START(GSN_CM_REGREF, refToNode(TBRef), "");
@@ -622,15 +811,24 @@ void Qmgr::execCM_REGCONF(Signal* signal)
   jamEntry();
 
   const CmRegConf * const cmRegConf = (CmRegConf *)&signal->theData[0];
+  Uint32 presidentNodeId = cmRegConf->presidentNodeId;
 
   if (!ndbCompatible_ndb_ndb(NDB_VERSION, cmRegConf->presidentVersion)) {
     jam();
     char buf[128];
-    BaseString::snprintf(buf,sizeof(buf),"incompatible version own=0x%x other=0x%x, shutting down", NDB_VERSION, cmRegConf->presidentVersion);
+    BaseString::snprintf(buf,sizeof(buf), 
+			 "incompatible version own=0x%x other=0x%x, "
+			 " shutting down", 
+			 NDB_VERSION, cmRegConf->presidentVersion);
     systemErrorLab(signal, __LINE__, buf);
     return;
   }
 
+  myNodePtr.i = getOwnNodeId();
+  ptrCheckGuard(myNodePtr, MAX_NDB_NODES, nodeRec);
+  
+  ndbrequire(c_start.m_gsn == GSN_CM_REGREQ);
+  ndbrequire(myNodePtr.p->phase = ZSTARTING);
   
   cpdistref    = cmRegConf->presidentBlockRef;
   cpresident   = cmRegConf->presidentNodeId;
@@ -638,6 +836,8 @@ void Qmgr::execCM_REGCONF(Signal* signal)
   c_maxDynamicId = TdynamicId;
   c_clusterNodes.assign(NdbNodeBitmask::Size, cmRegConf->allNdbNodes);
 
+  myNodePtr.p->ndynamicId = TdynamicId;
+  
 /*--------------------------------------------------------------*/
 // Send this as an EVENT REPORT to inform about hearing about
 // other NDB node proclaiming to be president.
@@ -648,10 +848,6 @@ void Qmgr::execCM_REGCONF(Signal* signal)
   signal->theData[3] = TdynamicId;
   sendSignal(CMVMI_REF, GSN_EVENT_REP, signal, 4, JBB);
 
-  myNodePtr.i = getOwnNodeId();
-  ptrCheckGuard(myNodePtr, MAX_NDB_NODES, nodeRec);
-  myNodePtr.p->ndynamicId = TdynamicId;
-
   for (nodePtr.i = 1; nodePtr.i < MAX_NDB_NODES; nodePtr.i++) {
     jam();
     if (c_clusterNodes.get(nodePtr.i)){
@@ -674,6 +870,84 @@ void Qmgr::execCM_REGCONF(Signal* signal)
   return;
 }//Qmgr::execCM_REGCONF()
 
+void
+Qmgr::check_readnodes_reply(Signal* signal, Uint32 nodeId, Uint32 gsn)
+{
+  NodeRecPtr myNodePtr;
+  myNodePtr.i = getOwnNodeId();
+  ptrCheckGuard(myNodePtr, MAX_NDB_NODES, nodeRec);
+  
+  NodeRecPtr nodePtr;
+  nodePtr.i = nodeId;
+  ptrCheckGuard(nodePtr, MAX_NDB_NODES, nodeRec);
+
+  ndbrequire(c_readnodes_nodes.get(nodeId));
+  ReadNodesConf* conf = (ReadNodesConf*)signal->getDataPtr();
+  if (gsn == GSN_READ_NODESREF)
+  {
+    jam();
+retry:
+    signal->theData[0] = reference();
+    sendSignal(calcQmgrBlockRef(nodeId), GSN_READ_NODESREQ, signal, 1, JBA);
+    return;
+  }
+  
+  if (conf->masterNodeId == ZNIL)
+  {
+    jam();
+    goto retry;
+  }
+  
+  Uint32 president = conf->masterNodeId;
+  if (president == cpresident)
+  {
+    jam();
+    c_readnodes_nodes.clear(nodeId);
+    return;
+  }
+
+  char buf[255];
+  BaseString::snprintf(buf, sizeof(buf),
+		       "Partitioned cluster! check StartPartialTimeout, "
+		       " node %d thinks %d is president, "
+		       " I think president is: %d",
+		       nodeId, president, cpresident);
+
+  ndbout_c(buf);
+  CRASH_INSERTION(933);
+
+  if (getNodeState().startLevel == NodeState::SL_STARTED)
+  {
+    jam();
+    NdbNodeBitmask part;
+    part.assign(NdbNodeBitmask::Size, conf->clusterNodes);
+    FailRep* rep = (FailRep*)signal->getDataPtrSend();
+    rep->failCause = FailRep::ZPARTITIONED_CLUSTER;
+    rep->president = cpresident;
+    c_clusterNodes.copyto(NdbNodeBitmask::Size, rep->partition);
+    Uint32 ref = calcQmgrBlockRef(nodeId);
+    Uint32 i = 0;
+    while((i = part.find(i + 1)) != NdbNodeBitmask::NotFound)
+    {
+      if (i == nodeId)
+	continue;
+      rep->failNodeId = i;
+      sendSignal(ref, GSN_FAIL_REP, signal, FailRep::SignalLength, JBA);
+    }
+    rep->failNodeId = nodeId;
+    sendSignal(ref, GSN_FAIL_REP, signal, FailRep::SignalLength, JBB);
+    return;
+  }
+  
+  CRASH_INSERTION(932);
+  
+  progError(__LINE__, 
+	    NDBD_EXIT_ARBIT_SHUTDOWN,
+	    buf);
+  
+  ndbrequire(false);
+}
+
 void
 Qmgr::sendCmNodeInfoReq(Signal* signal, Uint32 nodeId, const NodeRec * self){
   CmNodeInfoReq * const req = (CmNodeInfoReq*)signal->getDataPtrSend();
@@ -703,26 +977,105 @@ Qmgr::sendCmNodeInfoReq(Signal* signal, Uint32 nodeId, const NodeRec * self){
 /*******************************/
 /* CM_REGREF                  */
 /*******************************/
+static
+const char *
+get_start_type_string(Uint32 st)
+{
+  static char buf[256];
+  
+  if (st == 0)
+  {
+    return "<ANY>";
+  }
+  else
+  {
+    buf[0] = 0;
+    for(Uint32 i = 0; i<NodeState::ST_ILLEGAL_TYPE; i++)
+    {
+      if (st & (1 << i))
+      {
+	if (buf[0])
+	  strcat(buf, "/");
+	switch(i){
+	case NodeState::ST_INITIAL_START:
+	  strcat(buf, "inital start");
+	  break;
+	case NodeState::ST_SYSTEM_RESTART:
+	  strcat(buf, "system restart");
+	  break;
+	case NodeState::ST_NODE_RESTART:
+	  strcat(buf, "node restart");
+	  break;
+	case NodeState::ST_INITIAL_NODE_RESTART:
+	  strcat(buf, "initial node restart");
+	  break;
+	}
+      }
+    }
+    return buf;
+  }
+}
+
 void Qmgr::execCM_REGREF(Signal* signal) 
 {
   jamEntry();
-  c_regReqReqRecv++;
 
-  // Ignore block reference in data[0]
-  UintR TaddNodeno = signal->theData[1];
-  UintR TrefuseReason = signal->theData[2];
-  Uint32 candidate = signal->theData[3];
+  CmRegRef* ref = (CmRegRef*)signal->getDataPtr();
+  UintR TaddNodeno = ref->nodeId;
+  UintR TrefuseReason = ref->errorCode;
+  Uint32 candidate = ref->presidentCandidate;
+  Uint32 node_gci = 1;
+  Uint32 candidate_gci = 1;
+  Uint32 start_type = ~0;
+  NdbNodeBitmask skip_nodes;
   DEBUG_START3(signal, TrefuseReason);
 
-  if(candidate != cpresidentCandidate){
+  if (signal->getLength() == CmRegRef::SignalLength)
+  {
+    jam();
+    node_gci = ref->latest_gci;
+    candidate_gci = ref->candidate_latest_gci;
+    start_type = ref->start_type;
+    skip_nodes.assign(NdbNodeBitmask::Size, ref->skip_nodes);
+  }
+  
+  c_start.m_regReqReqRecv++;
+
+  // Ignore block reference in data[0]
+  
+  if(candidate != c_start.m_president_candidate)
+  {
+    jam();
+    c_start.m_regReqReqRecv = ~0;
+  }
+  
+  c_start.m_starting_nodes.set(TaddNodeno);
+  if (node_gci)
+  {
     jam();
-    c_regReqReqRecv = ~0;
+    c_start.m_starting_nodes_w_log.set(TaddNodeno);
   }
   
+  skip_nodes.bitAND(c_definedNodes);
+  c_start.m_skip_nodes.bitOR(skip_nodes);
+  
+  char buf[100];
   switch (TrefuseReason) {
   case CmRegRef::ZINCOMPATIBLE_VERSION:
     jam();
-    systemErrorLab(signal, __LINE__, "incompatible version, connection refused by running ndb node");
+    systemErrorLab(signal, __LINE__, 
+		   "incompatible version, "
+		   "connection refused by running ndb node");
+  case CmRegRef::ZINCOMPATIBLE_START_TYPE:
+    jam();
+    BaseString::snprintf(buf, sizeof(buf),
+			 "incompatible start type detected: node %d"
+			 " reports %s(%d) my start type: %s(%d)",
+			 TaddNodeno,
+			 get_start_type_string(start_type), start_type,
+			 get_start_type_string(c_start.m_start_type),
+			 c_start.m_start_type);
+    progError(__LINE__, NDBD_EXIT_SR_RESTARTCONFLICT, buf);
     break;
   case CmRegRef::ZBUSY:
   case CmRegRef::ZBUSY_TO_PRES:
@@ -741,14 +1094,19 @@ void Qmgr::execCM_REGREF(Signal* signal)
     break;
   case CmRegRef::ZELECTION:
     jam();
-    if (cpresidentCandidate > TaddNodeno) {
+    if (candidate_gci > c_start.m_president_candidate_gci ||
+	(candidate_gci == c_start.m_president_candidate_gci &&
+	 candidate < c_start.m_president_candidate))
+    {
       jam();
       //----------------------------------------
       /* We may already have a candidate      */
       /* choose the lowest nodeno             */
       //----------------------------------------
       signal->theData[3] = 2;
-      cpresidentCandidate = TaddNodeno;
+      c_start.m_president_candidate = candidate;
+      c_start.m_president_candidate_gci = candidate_gci;
+      ndbout_c("assign candidate: %u %u", candidate, candidate_gci);
     } else {
       signal->theData[3] = 4;
     }//if
@@ -776,32 +1134,34 @@ void Qmgr::execCM_REGREF(Signal* signal)
 //-----------------------------------------
   sendSignal(CMVMI_REF, GSN_EVENT_REP, signal, 4, JBB);
 
-  if(cpresidentAlive == ZTRUE){
+  if(cpresidentAlive == ZTRUE)
+  {
     jam();
-    DEBUG("");
+    DEBUG("cpresidentAlive");
     return;
   }
   
-  if(c_regReqReqSent != c_regReqReqRecv){
+  if(c_start.m_regReqReqSent != c_start.m_regReqReqRecv)
+  {
     jam();
-    DEBUG( c_regReqReqSent << " != " << c_regReqReqRecv);
+    DEBUG(c_start.m_regReqReqSent << " != " << c_start.m_regReqReqRecv);
     return;
   }
   
-  if(cpresidentCandidate != getOwnNodeId()){
+  if(c_start.m_president_candidate != getOwnNodeId())
+  {
     jam();
-    DEBUG("");
+    DEBUG("i'm not the candidate");
     return;
   }
   
   /**
-   * All configured nodes has agreed
+   * All connected nodes has agreed
    */
-  Uint64 now = NdbTick_CurrentMillisecond();
-  if((c_regReqReqRecv == cnoOfNodes) || now > c_stopElectionTime){
+  if(check_startup(signal))
+  {
     jam();
-    electionWon();
-    sendSttorryLab(signal);
+    electionWon(signal);
     
     /**
      * Start timer handling 
@@ -813,8 +1173,192 @@ void Qmgr::execCM_REGREF(Signal* signal)
   return;
 }//Qmgr::execCM_REGREF()
 
+Uint32
+Qmgr::check_startup(Signal* signal)
+{
+  Uint64 now = NdbTick_CurrentMillisecond();
+  Uint64 partial_timeout = c_start_election_time + c_restartPartialTimeout;
+  Uint64 partitioned_timeout = partial_timeout + c_restartPartionedTimeout;
+
+  /**
+   * First see if we should wait more...
+   */
+  NdbNodeBitmask tmp;
+  tmp.bitOR(c_start.m_skip_nodes);
+  tmp.bitOR(c_start.m_starting_nodes);
+
+  NdbNodeBitmask wait;
+  wait.assign(c_definedNodes);
+  wait.bitANDC(tmp);
+
+  Uint32 retVal = 0;
+  NdbNodeBitmask report_mask;
+
+  if ((c_start.m_latest_gci == 0) || 
+      (c_start.m_start_type == (1 << NodeState::ST_INITIAL_START)))
+  {
+    if (!tmp.equal(c_definedNodes))
+    {
+      jam();
+      signal->theData[1] = 1;
+      signal->theData[2] = ~0;
+      report_mask.assign(wait);
+      retVal = 0;
+      goto start_report;
+    }
+    else
+    {
+      jam();
+      signal->theData[1] = 0x8000;
+      report_mask.assign(c_definedNodes);
+      report_mask.bitANDC(c_start.m_starting_nodes);
+      retVal = 1;
+      goto start_report;
+    }
+  }
+  const bool all = c_start.m_starting_nodes.equal(c_definedNodes);
+  CheckNodeGroups* sd = (CheckNodeGroups*)&signal->theData[0];
+  
+  {
+    /**
+     * Check for missing node group directly
+     */
+    char buf[100];
+    NdbNodeBitmask check;
+    check.assign(c_definedNodes);
+    check.bitANDC(c_start.m_starting_nodes);    // Not connected nodes
+    check.bitOR(c_start.m_starting_nodes_w_log);
+    
+    sd->blockRef = reference();
+    sd->requestType = CheckNodeGroups::Direct | CheckNodeGroups::ArbitCheck;
+    sd->mask = check;
+    EXECUTE_DIRECT(DBDIH, GSN_CHECKNODEGROUPSREQ, signal, 
+		   CheckNodeGroups::SignalLength);
+    
+    if (sd->output == CheckNodeGroups::Lose)
+    {
+      jam();
+      goto missing_nodegroup;
+    }
+  }
+  
+  sd->blockRef = reference();
+  sd->requestType = CheckNodeGroups::Direct | CheckNodeGroups::ArbitCheck;
+  sd->mask = c_start.m_starting_nodes;
+  EXECUTE_DIRECT(DBDIH, GSN_CHECKNODEGROUPSREQ, signal, 
+		 CheckNodeGroups::SignalLength);
+  
+  const Uint32 result = sd->output;
+  
+  sd->blockRef = reference();
+  sd->requestType = CheckNodeGroups::Direct | CheckNodeGroups::ArbitCheck;
+  sd->mask = c_start.m_starting_nodes_w_log;
+  EXECUTE_DIRECT(DBDIH, GSN_CHECKNODEGROUPSREQ, signal, 
+		 CheckNodeGroups::SignalLength);
+  
+  const Uint32 result_w_log = sd->output;
+
+  if (tmp.equal(c_definedNodes))
+  {
+    /**
+     * All nodes (wrt no-wait nodes) has connected...
+     *   this means that we will now start or die
+     */
+    jam();    
+    switch(result_w_log){
+    case CheckNodeGroups::Lose:
+    {
+      jam();
+      goto missing_nodegroup;
+    }
+    case CheckNodeGroups::Win:
+      signal->theData[1] = all ? 0x8001 : 0x8002;
+      report_mask.assign(c_definedNodes);
+      report_mask.bitANDC(c_start.m_starting_nodes);
+      retVal = 1;
+      goto start_report;
+    case CheckNodeGroups::Partitioning:
+      ndbrequire(result != CheckNodeGroups::Lose);
+      signal->theData[1] = 
+	all ? 0x8001 : (result == CheckNodeGroups::Win ? 0x8002 : 0x8003);
+      report_mask.assign(c_definedNodes);
+      report_mask.bitANDC(c_start.m_starting_nodes);
+      retVal = 1;
+      goto start_report;
+    }
+  }
+
+  if (now < partial_timeout)
+  {
+    jam();
+    signal->theData[1] = c_restartPartialTimeout == ~0 ? 2 : 3;
+    signal->theData[2] = Uint32((partial_timeout - now + 500) / 1000);
+    report_mask.assign(wait);
+    retVal = 0;
+    goto start_report;
+  }
+  
+  /**
+   * Start partial has passed...check for partitioning...
+   */  
+  switch(result_w_log){
+  case CheckNodeGroups::Lose:
+    jam();
+    goto missing_nodegroup;
+  case CheckNodeGroups::Partitioning:
+    if (now < partitioned_timeout && result != CheckNodeGroups::Win)
+    {
+      signal->theData[1] = c_restartPartionedTimeout == ~0 ? 4 : 5;
+      signal->theData[2] = Uint32((partitioned_timeout - now + 500) / 1000);
+      report_mask.assign(c_definedNodes);
+      report_mask.bitANDC(c_start.m_starting_nodes);
+      retVal = 0;
+      goto start_report;
+    }
+    // Fall through...
+  case CheckNodeGroups::Win:
+    signal->theData[1] = 
+      all ? 0x8001 : (result == CheckNodeGroups::Win ? 0x8002 : 0x8003);
+    report_mask.assign(c_definedNodes);
+    report_mask.bitANDC(c_start.m_starting_nodes);
+    retVal = 1;
+    goto start_report;
+  }
+
+  ndbrequire(false);
+
+start_report:
+  jam();
+  {
+    Uint32 sz = NdbNodeBitmask::Size;
+    signal->theData[0] = NDB_LE_StartReport;
+    signal->theData[3] = sz;
+    Uint32* ptr = signal->theData+4;
+    c_definedNodes.copyto(sz, ptr); ptr += sz;
+    c_start.m_starting_nodes.copyto(sz, ptr); ptr += sz;
+    c_start.m_skip_nodes.copyto(sz, ptr); ptr += sz;
+    report_mask.copyto(sz, ptr); ptr+= sz;
+    sendSignal(CMVMI_REF, GSN_EVENT_REP, signal, 
+	       4+4*NdbNodeBitmask::Size, JBB);
+  }
+  return retVal;
+  
+missing_nodegroup:
+  jam();
+  char buf[100], mask1[100], mask2[100];
+  c_start.m_starting_nodes.getText(mask1);
+  tmp.assign(c_start.m_starting_nodes);
+  tmp.bitANDC(c_start.m_starting_nodes_w_log);
+  tmp.getText(mask2);
+  BaseString::snprintf(buf, sizeof(buf),
+		       "Unable to start missing node group! "
+		       " starting: %s (missing fs for: %s)",
+		       mask1, mask2);
+  progError(__LINE__, NDBD_EXIT_SR_RESTARTCONFLICT, buf);
+}
+
 void
-Qmgr::electionWon(){
+Qmgr::electionWon(Signal* signal){
   NodeRecPtr myNodePtr;
   cpresident = getOwnNodeId(); /* This node becomes president. */
   myNodePtr.i = getOwnNodeId();
@@ -831,8 +1375,21 @@ Qmgr::electionWon(){
   c_clusterNodes.set(getOwnNodeId());
   
   cpresidentAlive = ZTRUE;
-  c_stopElectionTime = ~0;
+  c_start_election_time = ~0;
   c_start.reset();
+
+  signal->theData[0] = NDB_LE_CM_REGCONF;
+  signal->theData[1] = getOwnNodeId();
+  signal->theData[2] = cpresident;
+  signal->theData[3] = 1;
+  sendSignal(CMVMI_REF, GSN_EVENT_REP, signal, 4, JBB);
+
+  c_start.m_starting_nodes.clear(getOwnNodeId());
+  if (c_start.m_starting_nodes.isclear())
+  {
+    jam();
+    sendSttorryLab(signal);
+  }
 }
 
 /*
@@ -846,7 +1403,15 @@ Qmgr::electionWon(){
 /*--------------------------------------------------------------*/
 void Qmgr::regreqTimeLimitLab(Signal* signal) 
 {
-  if(cpresident == ZNIL){
+  if(cpresident == ZNIL)
+  {
+    if (c_start.m_president_candidate == ZNIL)
+    {
+      jam();
+      c_start.m_president_candidate = getOwnNodeId();
+      ndbout_c("Assigning candidate to self: %d", getOwnNodeId());
+    }
+    
     cmInfoconf010Lab(signal);
   }
 }//Qmgr::regreqTimelimitLab()
@@ -967,7 +1532,7 @@ Qmgr::cmAddPrepare(Signal* signal, NodeRecPtr nodePtr, const NodeRec * self){
     ndbrequire(signal->header.theVerId_signalNumber == GSN_CM_ADD);
     c_start.m_nodes.clearWaitingFor();
     c_start.m_nodes.setWaitingFor(nodePtr.i);
-    c_start.m_gsn = GSN_CM_ADD;
+    c_start.m_gsn = GSN_CM_NODEINFOCONF;
 #else
     warningEvent("Enabling communication to CM_ADD node %u state=%d", 
 		 nodePtr.i,
@@ -1256,6 +1821,17 @@ void Qmgr::execCM_ACKADD(Signal* signal)
      */
     handleArbitNdbAdd(signal, addNodePtr.i);
     c_start.reset();
+
+    if (c_start.m_starting_nodes.get(addNodePtr.i))
+    {
+      jam();
+      c_start.m_starting_nodes.clear(addNodePtr.i);
+      if (c_start.m_starting_nodes.isclear())
+      {
+	jam();
+	sendSttorryLab(signal);
+      }
+    }
     return;
   }//switch
   ndbrequire(false);
@@ -1409,7 +1985,8 @@ void Qmgr::initData(Signal* signal)
   cnoPrepFailedNodes = 0;
   creadyDistCom = ZFALSE;
   cpresident = ZNIL;
-  cpresidentCandidate = ZNIL;
+  c_start.m_president_candidate = ZNIL;
+  c_start.m_president_candidate_gci = 0;
   cpdistref = 0;
   cneighbourh = ZNIL;
   cneighbourl = ZNIL;
@@ -1437,15 +2014,33 @@ void Qmgr::initData(Signal* signal)
   Uint32 hbDBAPI = 1500;
   Uint32 arbitTimeout = 1000;
   c_restartPartialTimeout = 30000;
+  c_restartPartionedTimeout = 60000;
+  c_restartFailureTimeout = ~0;
   ndb_mgm_get_int_parameter(p, CFG_DB_HEARTBEAT_INTERVAL, &hbDBDB);
   ndb_mgm_get_int_parameter(p, CFG_DB_API_HEARTBEAT_INTERVAL, &hbDBAPI);
   ndb_mgm_get_int_parameter(p, CFG_DB_ARBIT_TIMEOUT, &arbitTimeout);
   ndb_mgm_get_int_parameter(p, CFG_DB_START_PARTIAL_TIMEOUT, 
 			    &c_restartPartialTimeout);
-  if(c_restartPartialTimeout == 0){
+  ndb_mgm_get_int_parameter(p, CFG_DB_START_PARTITION_TIMEOUT,
+			    &c_restartPartionedTimeout);
+  ndb_mgm_get_int_parameter(p, CFG_DB_START_FAILURE_TIMEOUT,
+			    &c_restartFailureTimeout);
+  
+  if(c_restartPartialTimeout == 0)
+  {
     c_restartPartialTimeout = ~0;
   }
   
+  if (c_restartPartionedTimeout ==0)
+  {
+    c_restartPartionedTimeout = ~0;
+  }
+  
+  if (c_restartFailureTimeout == 0)
+  {
+    c_restartFailureTimeout = ~0;
+  }
+
   setHbDelay(hbDBDB);
   setHbApiDelay(hbDBAPI);
   setArbitTimeout(arbitTimeout);
@@ -1872,10 +2467,23 @@ void Qmgr::execDISCONNECT_REP(Signal* signal)
   const Uint32 nodeId = rep->nodeId;
   const Uint32 err = rep->err;
   c_connectedNodes.clear(nodeId);
+  c_readnodes_nodes.clear(nodeId);
   
   NodeRecPtr nodePtr;
   nodePtr.i = getOwnNodeId();
   ptrCheckGuard(nodePtr, MAX_NODES, nodeRec);
+
+  char buf[100];
+  if (getNodeInfo(nodeId).getType() == NodeInfo::DB &&
+      getNodeState().startLevel < NodeState::SL_STARTED)
+  {
+    jam();
+    CRASH_INSERTION(932);
+    BaseString::snprintf(buf, 100, "Node %u disconected", nodeId);    
+    progError(__LINE__, NDBD_EXIT_SR_OTHERNODEFAILED, buf);
+    ndbrequire(false);
+  }
+  
   switch(nodePtr.p->phase){
   case ZRUNNING:
     jam();
@@ -1893,9 +2501,12 @@ void Qmgr::execDISCONNECT_REP(Signal* signal)
   case ZAPI_ACTIVE:
     ndbrequire(false);
   case ZAPI_INACTIVE:
+  {
+    BaseString::snprintf(buf, 100, "Node %u disconected", nodeId);    
+    progError(__LINE__, NDBD_EXIT_SR_OTHERNODEFAILED, buf);
     ndbrequire(false);
   }
-
+  }
   node_failed(signal, nodeId);
 }//DISCONNECT_REP
 
@@ -2150,10 +2761,16 @@ void Qmgr::failReportLab(Signal* signal, Uint16 aFailedNode,
 
   failedNodePtr.i = aFailedNode;
   ptrCheckGuard(failedNodePtr, MAX_NODES, nodeRec);
+  FailRep* rep = (FailRep*)signal->getDataPtr();
+
+  check_multi_node_shutdown(signal);
+  
   if (failedNodePtr.i == getOwnNodeId()) {
     jam();
 
+    Uint32 code = 0;
     const char * msg = 0;
+    char extra[100];
     switch(aFailCause){
     case FailRep::ZOWN_FAILURE: 
       msg = "Own failure"; 
@@ -2174,17 +2791,51 @@ void Qmgr::failReportLab(Signal* signal, Uint16 aFailedNode,
     case FailRep::ZLINK_FAILURE:
       msg = "Connection failure";
       break;
+    case FailRep::ZPARTITIONED_CLUSTER:
+    {
+      code = NDBD_EXIT_ARBIT_SHUTDOWN;
+      char buf1[100], buf2[100];
+      c_clusterNodes.getText(buf1);
+      if (signal->getLength()== FailRep::SignalLength + FailRep::ExtraLength &&
+	  signal->header.theVerId_signalNumber == GSN_FAIL_REP)
+      {
+	jam();
+	NdbNodeBitmask part;
+	part.assign(NdbNodeBitmask::Size, rep->partition);
+	part.getText(buf2);
+	BaseString::snprintf(extra, sizeof(extra),
+			     "Partitioned cluster!"
+			     " Our cluster: %s other cluster: %s",
+			     buf1, buf2);
+      }
+      else
+      {
+	jam();
+	BaseString::snprintf(extra, sizeof(extra),
+			     "Partitioned cluster!"
+			     " Our cluster: %s ", buf1);
+      }
+      msg = extra;
+      break;
+    }
+    case FailRep::ZMULTI_NODE_SHUTDOWN:
+      msg = "Multi node shutdown";
+      break;
+    default:
+      msg = "<UNKNOWN>";
     }
     
-    char buf[100];
-    BaseString::snprintf(buf, 100, 
+    CRASH_INSERTION(932);
+
+    char buf[255];
+    BaseString::snprintf(buf, sizeof(buf), 
 			 "We(%u) have been declared dead by %u reason: %s(%u)",
 			 getOwnNodeId(),
 			 refToNode(signal->getSendersBlockRef()),
 			 aFailCause,
 			 msg ? msg : "<Unknown>");
     
-    progError(__LINE__, 0, buf);
+    progError(__LINE__, code, buf);
     return;
   }//if
   
@@ -2242,6 +2893,8 @@ void Qmgr::execPREP_FAILREQ(Signal* signal)
   NodeRecPtr myNodePtr;
   jamEntry();
   
+  check_multi_node_shutdown(signal);
+  
   PrepFailReqRef * const prepFail = (PrepFailReqRef *)&signal->theData[0];
 
   BlockReference Tblockref  = prepFail->xxxBlockRef;
@@ -3893,6 +4546,7 @@ Qmgr::stateArbitCrash(Signal* signal)
   if (! (arbitRec.getTimediff() > getArbitTimeout()))
     return;
 #endif
+  CRASH_INSERTION(932);
   progError(__LINE__, NDBD_EXIT_ARBIT_SHUTDOWN,
             "Arbitrator decided to shutdown this node");
 }
@@ -3956,8 +4610,10 @@ Qmgr::execDUMP_STATE_ORD(Signal* signal)
   case 1:
     infoEvent("creadyDistCom = %d, cpresident = %d\n",
 	      creadyDistCom, cpresident);
-    infoEvent("cpresidentAlive = %d, cpresidentCand = %d\n",
-              cpresidentAlive, cpresidentCandidate);
+    infoEvent("cpresidentAlive = %d, cpresidentCand = %d (gci: %d)\n",
+              cpresidentAlive, 
+	      c_start.m_president_candidate, 
+	      c_start.m_president_candidate_gci);
     infoEvent("ctoStatus = %d\n", ctoStatus);
     for(Uint32 i = 1; i<MAX_NDB_NODES; i++){
       if(getNodeInfo(i).getType() == NodeInfo::DB){
@@ -4054,3 +4710,40 @@ Qmgr::execAPI_BROADCAST_REP(Signal* signal)
   NodeReceiverGroup rg(API_CLUSTERMGR, mask);
   sendSignal(rg, api.gsn, signal, len, JBB); // forward sections
 }
+
+void
+Qmgr::execSTOP_REQ(Signal* signal)
+{
+  jamEntry();
+  c_stopReq = * (StopReq*)signal->getDataPtr();
+
+  if (c_stopReq.senderRef)
+  {
+    ndbrequire(NdbNodeBitmask::get(c_stopReq.nodes, getOwnNodeId()));
+    
+    StopConf *conf = (StopConf*)signal->getDataPtrSend();
+    conf->senderData = c_stopReq.senderData;
+    conf->nodeState = getOwnNodeId();
+    sendSignal(c_stopReq.senderRef, 
+	       GSN_STOP_CONF, signal, StopConf::SignalLength, JBA);
+  }
+}
+
+void
+Qmgr::check_multi_node_shutdown(Signal* signal)
+{
+  if (c_stopReq.senderRef && 
+      NdbNodeBitmask::get(c_stopReq.nodes, getOwnNodeId()))
+  {
+    jam();
+    if(StopReq::getPerformRestart(c_stopReq.requestInfo))
+    {
+      jam();
+      StartOrd * startOrd = (StartOrd *)&signal->theData[0];
+      startOrd->restartInfo = c_stopReq.requestInfo;
+      EXECUTE_DIRECT(CMVMI, GSN_START_ORD, signal, 2);
+    } else {
+      EXECUTE_DIRECT(CMVMI, GSN_STOP_ORD, signal, 1);
+    }
+  }
+}

ndb/src/kernel/vm/Configuration.cpp

@@ -55,6 +55,12 @@ enum ndbd_options {
 NDB_STD_OPTS_VARS;
 // XXX should be my_bool ???
 static int _daemon, _no_daemon, _foreground,  _initial, _no_start;
+static int _initialstart;
+static const char* _nowait_nodes;
+
+extern Uint32 g_start_type;
+extern NdbNodeBitmask g_nowait_nodes;
+
 /**
  * Arguments to NDB process
  */ 
@@ -82,6 +88,14 @@ static struct my_option my_long_options[] =
     " (implies --nodaemon)",
     (gptr*) &_foreground, (gptr*) &_foreground, 0,
     GET_BOOL, NO_ARG, 0, 0, 0, 0, 0, 0 },
+  { "nowait-nodes", NO_ARG, 
+    "Nodes that will not be waited for during start",
+    (gptr*) &_nowait_nodes, (gptr*) &_nowait_nodes, 0,
+    GET_STR, REQUIRED_ARG, 0, 0, 0, 0, 0, 0 },
+  { "initial-start", NO_ARG, 
+    "Perform initial start",
+    (gptr*) &_initialstart, (gptr*) &_initialstart, 0,
+    GET_BOOL, NO_ARG, 0, 0, 0, 0, 0, 0 },
   { 0, 0, 0, 0, 0, 0, GET_NO_ARG, NO_ARG, 0, 0, 0, 0, 0, 0}
 };
 static void short_usage_sub(void)
@@ -150,6 +164,37 @@ Configuration::init(int argc, char** argv)
   
   globalData.ownId= 0;
 
+  if (_nowait_nodes)
+  {
+    BaseString str(_nowait_nodes);
+    Vector<BaseString> arr;
+    str.split(arr, ",");
+    for (Uint32 i = 0; i<arr.size(); i++)
+    {
+      char *endptr = 0;
+      long val = strtol(arr[i].c_str(), &endptr, 10);
+      if (*endptr)
+      {
+	ndbout_c("Unable to parse nowait-nodes argument: %s : %s", 
+		 arr[i].c_str(), _nowait_nodes);
+	exit(-1);
+      }
+      if (! (val > 0 && val < MAX_NDB_NODES))
+      {
+	ndbout_c("Invalid nodeid specified in nowait-nodes: %d : %s", 
+		 val, _nowait_nodes);
+	exit(-1);
+      }
+      g_nowait_nodes.set(val);
+    }
+  }
+
+  if (_initialstart)
+  {
+    _initialStart = true;
+    g_start_type |= (1 << NodeState::ST_INITIAL_START);
+  }
+  
   return true;
 }
 

ndb/test/ndbapi/testNodeRestart.cpp

@@ -22,7 +22,7 @@
 #include <NdbRestarts.hpp>
 #include <Vector.hpp>
 #include <signaldata/DumpStateOrd.hpp>
-
+#include <Bitmask.hpp>
 
 int runLoadTable(NDBT_Context* ctx, NDBT_Step* step){
 
@@ -669,6 +669,206 @@ err:
   return NDBT_FAILED;    
 }
 
+int 
+runBug18612(NDBT_Context* ctx, NDBT_Step* step){
+
+  // Assume two replicas
+  NdbRestarter restarter;
+  if (restarter.getNumDbNodes() < 2)
+  {
+    ctx->stopTest();
+    return NDBT_OK;
+  }
+
+  Uint32 cnt = restarter.getNumDbNodes();
+
+  for(int loop = 0; loop < ctx->getNumLoops(); loop++)
+  {
+    int partition0[256];
+    int partition1[256];
+    bzero(partition0, sizeof(partition0));
+    bzero(partition1, sizeof(partition1));
+    Bitmask<4> nodesmask;
+    
+    Uint32 node1 = restarter.getDbNodeId(rand()%cnt);
+    for (Uint32 i = 0; i<cnt/2; i++)
+    {
+      do { 
+	int tmp = restarter.getRandomNodeOtherNodeGroup(node1, rand());
+	if (tmp == -1)
+	  break;
+	node1 = tmp;
+      } while(nodesmask.get(node1));
+      
+      partition0[i] = node1;
+      partition1[i] = restarter.getRandomNodeSameNodeGroup(node1, rand());
+      
+      ndbout_c("nodes %d %d", node1, partition1[i]);
+      
+      assert(!nodesmask.get(node1));
+      assert(!nodesmask.get(partition1[i]));
+      nodesmask.set(node1);
+      nodesmask.set(partition1[i]);
+    } 
+    
+    ndbout_c("done");
+
+    int dump[255];
+    dump[0] = DumpStateOrd::NdbcntrStopNodes;
+    memcpy(dump + 1, partition0, sizeof(int)*cnt/2);
+    
+    Uint32 master = restarter.getMasterNodeId();
+    
+    if (restarter.dumpStateOneNode(master, dump, 1+cnt/2))
+      return NDBT_FAILED;
+    
+    if (restarter.waitNodesNoStart(partition0, cnt/2))
+      return NDBT_FAILED;
+
+    int val2[] = { DumpStateOrd::CmvmiSetRestartOnErrorInsert, 1 };
+    
+    if (restarter.dumpStateAllNodes(val2, 2))
+      return NDBT_FAILED;
+    
+    if (restarter.insertErrorInAllNodes(932))
+      return NDBT_FAILED;
+
+    dump[0] = 9000;
+    memcpy(dump + 1, partition0, sizeof(int)*cnt/2);    
+    for (Uint32 i = 0; i<cnt/2; i++)
+      if (restarter.dumpStateOneNode(partition1[i], dump, 1+cnt/2))
+	return NDBT_FAILED;
+
+    dump[0] = 9000;
+    memcpy(dump + 1, partition1, sizeof(int)*cnt/2);    
+    for (Uint32 i = 0; i<cnt/2; i++)
+      if (restarter.dumpStateOneNode(partition0[i], dump, 1+cnt/2))
+	return NDBT_FAILED;
+    
+    if (restarter.startNodes(partition0, cnt/2))
+      return NDBT_FAILED;
+    
+    if (restarter.waitNodesStartPhase(partition0, cnt/2, 2))
+      return NDBT_FAILED;
+    
+    dump[0] = 9001;
+    for (Uint32 i = 0; i<cnt/2; i++)
+      if (restarter.dumpStateAllNodes(dump, 2))
+	return NDBT_FAILED;
+
+    if (restarter.waitNodesNoStart(partition0, cnt/2))
+      return NDBT_FAILED;
+    
+    for (Uint32 i = 0; i<cnt/2; i++)
+      if (restarter.restartOneDbNode(partition0[i], true, true, true))
+	return NDBT_FAILED;
+    
+    if (restarter.waitNodesNoStart(partition0, cnt/2))
+      return NDBT_FAILED;
+    
+    if (restarter.startAll())
+      return NDBT_FAILED;
+
+    if (restarter.waitClusterStarted())
+      return NDBT_FAILED;
+  }
+  return NDBT_OK;
+}
+
+int 
+runBug18612SR(NDBT_Context* ctx, NDBT_Step* step){
+
+  // Assume two replicas
+  NdbRestarter restarter;
+  if (restarter.getNumDbNodes() < 2)
+  {
+    ctx->stopTest();
+    return NDBT_OK;
+  }
+
+  Uint32 cnt = restarter.getNumDbNodes();
+
+  for(int loop = 0; loop < ctx->getNumLoops(); loop++)
+  {
+    int partition0[256];
+    int partition1[256];
+    bzero(partition0, sizeof(partition0));
+    bzero(partition1, sizeof(partition1));
+    Bitmask<4> nodesmask;
+    
+    Uint32 node1 = restarter.getDbNodeId(rand()%cnt);
+    for (Uint32 i = 0; i<cnt/2; i++)
+    {
+      do { 
+	int tmp = restarter.getRandomNodeOtherNodeGroup(node1, rand());
+	if (tmp == -1)
+	  break;
+	node1 = tmp;
+      } while(nodesmask.get(node1));
+      
+      partition0[i] = node1;
+      partition1[i] = restarter.getRandomNodeSameNodeGroup(node1, rand());
+      
+      ndbout_c("nodes %d %d", node1, partition1[i]);
+      
+      assert(!nodesmask.get(node1));
+      assert(!nodesmask.get(partition1[i]));
+      nodesmask.set(node1);
+      nodesmask.set(partition1[i]);
+    } 
+    
+    ndbout_c("done");
+
+    if (restarter.restartAll(false, true, false))
+      return NDBT_FAILED;
+
+    int dump[255];
+    dump[0] = 9000;
+    memcpy(dump + 1, partition0, sizeof(int)*cnt/2);    
+    for (Uint32 i = 0; i<cnt/2; i++)
+      if (restarter.dumpStateOneNode(partition1[i], dump, 1+cnt/2))
+	return NDBT_FAILED;
+
+    dump[0] = 9000;
+    memcpy(dump + 1, partition1, sizeof(int)*cnt/2);    
+    for (Uint32 i = 0; i<cnt/2; i++)
+      if (restarter.dumpStateOneNode(partition0[i], dump, 1+cnt/2))
+	return NDBT_FAILED;
+
+    int val2[] = { DumpStateOrd::CmvmiSetRestartOnErrorInsert, 1 };
+    
+    if (restarter.dumpStateAllNodes(val2, 2))
+      return NDBT_FAILED;
+    
+    if (restarter.insertErrorInAllNodes(932))
+      return NDBT_FAILED;
+    
+    if (restarter.startAll())
+      return NDBT_FAILED;
+    
+    if (restarter.waitClusterStartPhase(2))
+      return NDBT_FAILED;
+    
+    dump[0] = 9001;
+    for (Uint32 i = 0; i<cnt/2; i++)
+      if (restarter.dumpStateAllNodes(dump, 2))
+	return NDBT_FAILED;
+
+    if (restarter.waitClusterNoStart(30))
+      if (restarter.waitNodesNoStart(partition0, cnt/2, 10))
+	if (restarter.waitNodesNoStart(partition1, cnt/2, 10))
+	  return NDBT_FAILED;
+    
+    if (restarter.startAll())
+      return NDBT_FAILED;
+    
+    if (restarter.waitClusterStarted())
+      return NDBT_FAILED;
+  }
+  return NDBT_OK;
+}
+
+
 NDBT_TESTSUITE(testNodeRestart);
 TESTCASE("NoLoad", 
 	 "Test that one node at a time can be stopped and then restarted "\
@@ -963,6 +1163,18 @@ TESTCASE("Bug18414",
   STEP(runBug18414);
   FINALIZER(runClearTable);
 }
+TESTCASE("Bug18612",
+	 "Test bug with partitioned clusters"){
+  INITIALIZER(runLoadTable);
+  STEP(runBug18612);
+  FINALIZER(runClearTable);
+}
+TESTCASE("Bug18612SR",
+	 "Test bug with partitioned clusters"){
+  INITIALIZER(runLoadTable);
+  STEP(runBug18612SR);
+  FINALIZER(runClearTable);
+}
 NDBT_TESTSUITE_END(testNodeRestart);
 
 int main(int argc, const char** argv){

ndb/test/run-test/daily-basic-tests.txt

@@ -433,10 +433,18 @@ args: -n Bug16772 T1
 #cmd: testSystemRestart
 #args: -n Bug18385 T1
 #
-max-time: 500
+max-time: 1000
 cmd: testNodeRestart
 args: -n Bug18414 T1
 
+max-time: 1000
+cmd: testNodeRestart
+args: -n Bug18612 T1
+
+max-time: 1000
+cmd: testNodeRestart
+args: -n Bug18612SR T1
+
 # OLD FLEX
 max-time: 500
 cmd: flexBench

ndb/test/src/NdbRestarts.cpp

@@ -445,8 +445,7 @@ int twoNodeFailure(NdbRestarter& _restarter,
 	 << ") secs " << endl;
   NdbSleep_SecSleep(seconds);
 
-  randomId = (rand() % _restarter.getNumDbNodes());
-  nodeId = _restarter.getDbNodeId(randomId);  
+  nodeId = _restarter.getRandomNodeOtherNodeGroup(nodeId, rand());
   g_info << _restart->m_name << ": node = "<< nodeId << endl;
 
   CHECK(_restarter.insertErrorInNode(nodeId, 9999) == 0,