ndb - bug#26457

  master failure during master take over


ndb/src/kernel/blocks/ERROR_codes.txt:
  new error code
ndb/src/kernel/blocks/dbdih/DbdihMain.cpp:
  Make sure to clear NF_XX_LCP if master fails during master take-over
ndb/test/include/NdbRestarter.hpp:
  Add support for querying next master and node group
    (for multi node failure testing)
ndb/test/ndbapi/testNodeRestart.cpp:
  testcase
ndb/test/run-test/daily-basic-tests.txt:
  testcase
ndb/test/src/NdbRestarter.cpp:
  Add support for querying next master and node group
    (for multi node failure testing)

6 files changed, 139 insertions, 2 deletions

diff --git a/ndb/src/kernel/blocks/ERROR_codes.txt b/ndb/src/kernel/blocks/ERROR_codes.txt
index 0bcc99a6334..f7cb49014cb 100644
--- a/ndb/src/kernel/blocks/ERROR_codes.txt
+++ b/ndb/src/kernel/blocks/ERROR_codes.txt
@@ -5,7 +5,7 @@ Next DBACC 3002
 Next DBTUP 4014
 Next DBLQH 5043
 Next DBDICT 6007
-Next DBDIH 7178
+Next DBDIH 7181
 Next DBTC 8039
 Next CMVMI 9000
 Next BACKUP 10022
@@ -71,6 +71,8 @@ Delay GCP_SAVEREQ by 10 secs
 
 7177: Delay copying of sysfileData in execCOPY_GCIREQ
 
+7180: Crash master during master-take-over in execMASTER_LCPCONF
+
 ERROR CODES FOR TESTING NODE FAILURE, LOCAL CHECKPOINT HANDLING:
 -----------------------------------------------------------------
 
diff --git a/ndb/src/kernel/blocks/dbdih/DbdihMain.cpp b/ndb/src/kernel/blocks/dbdih/DbdihMain.cpp
index 5f573d40dfe..0e6fe4714b6 100644
--- a/ndb/src/kernel/blocks/dbdih/DbdihMain.cpp
+++ b/ndb/src/kernel/blocks/dbdih/DbdihMain.cpp
@@ -4612,6 +4612,8 @@ void
 Dbdih::startLcpMasterTakeOver(Signal* signal, Uint32 nodeId){
   jam();
 
+  Uint32 oldNode = c_lcpMasterTakeOverState.failedNodeId;
+
   c_lcpMasterTakeOverState.minTableId = ~0;
   c_lcpMasterTakeOverState.minFragId = ~0;
   c_lcpMasterTakeOverState.failedNodeId = nodeId;
@@ -4630,7 +4632,20 @@ Dbdih::startLcpMasterTakeOver(Signal* signal, Uint32 nodeId){
     /**
      * Node failure during master take over...
      */
-    ndbout_c("Nodefail during master take over");
+    ndbout_c("Nodefail during master take over (old: %d)", oldNode);
+  }
+  
+  NodeRecordPtr nodePtr;
+  nodePtr.i = oldNode;
+  if (oldNode > 0 && oldNode < MAX_NDB_NODES)
+  {
+    jam();
+    ptrCheckGuard(nodePtr, MAX_NDB_NODES, nodeRecord);
+    if (nodePtr.p->m_nodefailSteps.get(NF_LCP_TAKE_OVER))
+    {
+      jam();
+      checkLocalNodefailComplete(signal, oldNode, NF_LCP_TAKE_OVER);
+    }
   }
   
   setLocalNodefailHandling(signal, nodeId, NF_LCP_TAKE_OVER);
@@ -5646,6 +5661,14 @@ void Dbdih::execMASTER_LCPREQ(Signal* signal)
   jamEntry();
   const BlockReference newMasterBlockref = req->masterRef;
 
+  if (newMasterBlockref != cmasterdihref)
+  {
+    jam();
+    ndbout_c("resending GSN_MASTER_LCPREQ");
+    sendSignalWithDelay(reference(), GSN_MASTER_LCPREQ, signal,
+			signal->getLength(), 50);
+    return;
+  }
   Uint32 failedNodeId = req->failedNodeId;
 
   /**
@@ -5946,6 +5969,8 @@ void Dbdih::execMASTER_LCPCONF(Signal* signal)
   ptrCheckGuard(nodePtr, MAX_NDB_NODES, nodeRecord);
   nodePtr.p->lcpStateAtTakeOver = lcpState;
 
+  CRASH_INSERTION(7180);
+  
 #ifdef VM_TRACE
   ndbout_c("MASTER_LCPCONF");
   printMASTER_LCP_CONF(stdout, &signal->theData[0], 0, 0);
diff --git a/ndb/test/include/NdbRestarter.hpp b/ndb/test/include/NdbRestarter.hpp
index 3ec92ae786e..a2e6a4b3248 100644
--- a/ndb/test/include/NdbRestarter.hpp
+++ b/ndb/test/include/NdbRestarter.hpp
@@ -62,6 +62,8 @@ public:
   int dumpStateAllNodes(int * _args, int _num_args);
 
   int getMasterNodeId();
+  int getNextMasterNodeId(int nodeId);
+  int getNodeGroup(int nodeId);
   int getRandomNodeSameNodeGroup(int nodeId, int randomNumber);
   int getRandomNodeOtherNodeGroup(int nodeId, int randomNumber);
   int getRandomNotMasterNodeId(int randomNumber);
diff --git a/ndb/test/ndbapi/testNodeRestart.cpp b/ndb/test/ndbapi/testNodeRestart.cpp
index 43fb77342b5..e729e8179b5 100644
--- a/ndb/test/ndbapi/testNodeRestart.cpp
+++ b/ndb/test/ndbapi/testNodeRestart.cpp
@@ -1045,6 +1045,45 @@ int runBug25554(NDBT_Context* ctx, NDBT_Step* step){
   return NDBT_OK;
 }
 
+int
+runBug26457(NDBT_Context* ctx, NDBT_Step* step)
+{
+  NdbRestarter res;
+  if (res.getNumDbNodes() < 4)
+    return NDBT_OK;
+
+  int loops = ctx->getNumLoops();
+  while (loops --)
+  {
+retry:
+    int master = res.getMasterNodeId();
+    int next = res.getNextMasterNodeId(master);
+
+    ndbout_c("master: %d next: %d", master, next);
+
+    if (res.getNodeGroup(master) == res.getNodeGroup(next))
+    {
+      res.restartOneDbNode(next, false, false, true);
+      if (res.waitClusterStarted())
+	return NDBT_FAILED;
+      goto retry;
+    }
+
+    int val2[] = { DumpStateOrd::CmvmiSetRestartOnErrorInsert, 2 };
+    
+    if (res.dumpStateOneNode(next, val2, 2))
+      return NDBT_FAILED;
+    
+    if (res.insertErrorInNode(next, 7180))
+      return NDBT_FAILED;
+    
+    res.restartOneDbNode(master, false, false, true);
+    if (res.waitClusterStarted())
+      return NDBT_FAILED;
+  }
+  
+  return NDBT_OK;
+}
 
 NDBT_TESTSUITE(testNodeRestart);
 TESTCASE("NoLoad", 
@@ -1367,6 +1406,9 @@ TESTCASE("Bug25364", ""){
 TESTCASE("Bug25554", ""){
   INITIALIZER(runBug25554);
 }
+TESTCASE("Bug26457", ""){
+  INITIALIZER(runBug26457);
+}
 NDBT_TESTSUITE_END(testNodeRestart);
 
 int main(int argc, const char** argv){
diff --git a/ndb/test/run-test/daily-basic-tests.txt b/ndb/test/run-test/daily-basic-tests.txt
index 00fa14eea2c..9074ff145d8 100644
--- a/ndb/test/run-test/daily-basic-tests.txt
+++ b/ndb/test/run-test/daily-basic-tests.txt
@@ -477,6 +477,10 @@ max-time: 1000
 cmd: testNodeRestart
 args: -n Bug25554 T1
 
+max-time: 1000
+cmd: testNodeRestart
+args: -n Bug26457 T1
+
 # OLD FLEX
 max-time: 500
 cmd: flexBench
diff --git a/ndb/test/src/NdbRestarter.cpp b/ndb/test/src/NdbRestarter.cpp
index 2c16a05240d..4c7f52a8622 100644
--- a/ndb/test/src/NdbRestarter.cpp
+++ b/ndb/test/src/NdbRestarter.cpp
@@ -129,6 +129,68 @@ NdbRestarter::getMasterNodeId(){
 }
 
 int
+NdbRestarter::getNodeGroup(int nodeId){
+  if (!isConnected())
+    return -1;
+  
+  if (getStatus() != 0)
+    return -1;
+  
+  for(size_t i = 0; i < ndbNodes.size(); i++)
+  {
+    if(ndbNodes[i].node_id == nodeId)
+    {
+      return ndbNodes[i].node_group;
+    }
+  }
+  
+  return -1;
+}
+
+int
+NdbRestarter::getNextMasterNodeId(int nodeId){
+  if (!isConnected())
+    return -1;
+  
+  if (getStatus() != 0)
+    return -1;
+  
+  size_t i;
+  for(i = 0; i < ndbNodes.size(); i++)
+  {
+    if(ndbNodes[i].node_id == nodeId)
+    {
+      break;
+    }
+  }
+  assert(i < ndbNodes.size());
+  if (i == ndbNodes.size())
+    return -1;
+
+  int dynid = ndbNodes[i].dynamic_id;
+  int minid = dynid;
+  for (i = 0; i<ndbNodes.size(); i++)
+    if (ndbNodes[i].dynamic_id > minid)
+      minid = ndbNodes[i].dynamic_id;
+  
+  for (i = 0; i<ndbNodes.size(); i++)
+    if (ndbNodes[i].dynamic_id > dynid && 
+	ndbNodes[i].dynamic_id < minid)
+    {
+      minid = ndbNodes[i].dynamic_id;
+    }
+  
+  if (minid != ~0)
+  {
+    for (i = 0; i<ndbNodes.size(); i++)
+      if (ndbNodes[i].dynamic_id == minid)
+	return ndbNodes[i].node_id;
+  }
+  
+  return getMasterNodeId();
+}
+
+int
 NdbRestarter::getRandomNotMasterNodeId(int rand){
   int master = getMasterNodeId();
   if(master == -1)