ndb - bug#26481

  fix for killed node during initial node restart

ndb/src/kernel/blocks/dbdih/Dbdih.hpp

@@ -1368,6 +1368,7 @@ private:
   Uint32 csystemnodes;
   Uint32 currentgcp;
   Uint32 c_newest_restorable_gci;
+  Uint32 c_set_initial_start_flag;
 
   enum GcpMasterTakeOverState {
     GMTOS_IDLE = 0,

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

@@ -75,6 +75,7 @@ void Dbdih::initData()
   c_blockCommit    = false;
   c_blockCommitNo  = 1;
   cntrlblockref    = RNIL;
+  c_set_initial_start_flag = FALSE;
 }//Dbdih::initData()
 
 void Dbdih::initRecords() 

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

@@ -667,6 +667,12 @@ done:
   {
     jam();
     memcpy(sysfileData, cdata, sizeof(sysfileData));
+
+    if (c_set_initial_start_flag)
+    {
+      jam();
+      Sysfile::setInitialStartOngoing(SYSFILE->systemRestartBits);
+    }
   }
 
   c_copyGCISlave.m_copyReason = reason;
@@ -1260,6 +1266,11 @@ void Dbdih::execNDB_STTOR(Signal* signal)
     // The permission is given by the master node in the alive set.  
     /*-----------------------------------------------------------------------*/
     createMutexes(signal, 0);
+    if (cstarttype == NodeState::ST_INITIAL_NODE_RESTART)
+    {
+      jam();
+      c_set_initial_start_flag = TRUE; // In sysfile...
+    }
     break;
     
   case ZNDB_SPH3:
@@ -10277,6 +10288,17 @@ Dbdih::sendLCP_COMPLETE_REP(Signal* signal){
   
   sendSignal(c_lcpState.m_masterLcpDihRef, GSN_LCP_COMPLETE_REP, signal, 
 	     LcpCompleteRep::SignalLength, JBB);
+
+  /**
+   * Say that an initial node restart does not need to be redone
+   *   once node has been part of first LCP
+   */
+  if (c_set_initial_start_flag &&
+      c_lcpState.m_participatingLQH.get(getOwnNodeId()))
+  {
+    jam();
+    c_set_initial_start_flag = FALSE;
+  }
 }
 
 /*-------------------------------------------------------------------------- */

ndb/src/kernel/blocks/dblqh/DblqhMain.cpp

@@ -11673,7 +11673,8 @@ void Dblqh::execGCP_SAVEREQ(Signal* signal)
     return;
   }
 
-  if(getNodeState().getNodeRestartInProgress()){
+  if(getNodeState().getNodeRestartInProgress() && cstartRecReq == ZFALSE)
+  {
     GCPSaveRef * const saveRef = (GCPSaveRef*)&signal->theData[0];
     saveRef->dihPtr = dihPtr;
     saveRef->nodeId = getOwnNodeId();

ndb/test/ndbapi/testNodeRestart.cpp

@@ -1085,6 +1085,46 @@ retry:
   return NDBT_OK;
 }
 
+int 
+runBug26481(NDBT_Context* ctx, NDBT_Step* step)
+{
+  
+  int result = NDBT_OK;
+  int loops = ctx->getNumLoops();
+  int records = ctx->getNumRecords();
+  NdbRestarter res;
+  
+  int node = res.getRandomNotMasterNodeId(rand());
+  ndbout_c("node: %d", node);
+  if (res.restartOneDbNode(node, true, true, true))
+    return NDBT_FAILED;
+
+  if (res.waitNodesNoStart(&node, 1))
+    return NDBT_FAILED;
+
+  int val2[] = { DumpStateOrd::CmvmiSetRestartOnErrorInsert, 1 };
+  if (res.dumpStateOneNode(node, val2, 2))
+    return NDBT_FAILED;
+
+  if (res.insertErrorInNode(node, 7018))
+    return NDBT_FAILED;
+
+  if (res.startNodes(&node, 1))
+    return NDBT_FAILED;
+
+  res.waitNodesStartPhase(&node, 1, 3);
+  
+  if (res.waitNodesNoStart(&node, 1))
+    return NDBT_FAILED;
+
+  res.startNodes(&node, 1);
+  
+  if (res.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 "\
@@ -1409,6 +1449,9 @@ TESTCASE("Bug25554", ""){
 TESTCASE("Bug26457", ""){
   INITIALIZER(runBug26457);
 }
+TESTCASE("Bug26481", ""){
+  INITIALIZER(runBug26481);
+}
 NDBT_TESTSUITE_END(testNodeRestart);
 
 int main(int argc, const char** argv){