Bug fix 3.4/cluster comm threads start stop (#6939)

* Start ClusterComm threads in `ClusterFeature::start`. Stop ClusterComm threads in `ClusterFeature::stop`.

* Do not free objects in `Scheduler::shutdown`. Let the `unique_ptr` do their job. Stop ClusterComm threads in `ClusterFeature::stop`, but free instance in `ClusterFeature::unprepare`.

* `io_context` may contains lambdas that hold `shared_ptr`s to `Tasks` the required a functional `VocBase` in their destructor.

* Clean up.

arangod/Cluster/ClusterComm.cpp

@@ -257,11 +257,7 @@ ClusterComm::ClusterComm(bool ignored)
 ////////////////////////////////////////////////////////////////////////////////
 
 ClusterComm::~ClusterComm() {
-  for (ClusterCommThread * thread: _backgroundThreads) {
+  stopBackgroundThreads();
-    thread->beginShutdown();
-    delete thread;
-  }
-
   cleanupAllQueues();
 }
 
@@ -317,6 +313,10 @@ void ClusterComm::initialize() {
 ////////////////////////////////////////////////////////////////////////////////
 
 void ClusterComm::cleanup() {
+  if (!_theInstance) {
+    return ;
+  }
+
   _theInstance.reset();    // no more operations will be started, but running
                            // ones have their copy of the shared_ptr
 }
@@ -341,6 +341,16 @@ void ClusterComm::startBackgroundThread() {
   } // for
 }
 
+void ClusterComm::stopBackgroundThreads() {
+  for (ClusterCommThread * thread: _backgroundThreads) {
+    thread->beginShutdown();
+    delete thread;
+  }
+
+  _backgroundThreads.clear();
+}
+
+
 ////////////////////////////////////////////////////////////////////////////////
 /// @brief choose next communicator via round robin
 ////////////////////////////////////////////////////////////////////////////////

arangod/Cluster/ClusterComm.h

@@ -497,6 +497,7 @@ class ClusterComm {
   //////////////////////////////////////////////////////////////////////////////
 
   void startBackgroundThread();
+  void stopBackgroundThreads();
 
   //////////////////////////////////////////////////////////////////////////////
   /// @brief submit an HTTP request to a shard asynchronously.

arangod/Cluster/ClusterFeature.cpp

@@ -272,11 +272,6 @@ void ClusterFeature::prepare() {
     }
   }
 
-  if (startClusterComm) {
-    // initialize ClusterComm library, must call initialize only once
-    ClusterComm::initialize();
-  }
-
   // return if cluster is disabled
   if (!_enableCluster) {
     reportRole(ServerState::instance()->getRole());
@@ -361,6 +356,11 @@ void ClusterFeature::prepare() {
 }
 
 void ClusterFeature::start() {
+
+  if (ServerState::instance()->isAgent() || _enableCluster) {
+    ClusterComm::initialize();
+  }
+
   // return if cluster is disabled
   if (!_enableCluster) {
     startHeartbeatThread(nullptr, 5000, 5, std::string());
@@ -440,6 +440,8 @@ void ClusterFeature::stop() {
       }
     }
   }
+
+  ClusterComm::instance()->stopBackgroundThreads();
 }
 
 
@@ -479,7 +481,6 @@ void ClusterFeature::unprepare() {
   // Try only once to unregister because maybe the agencycomm
   // is shutting down as well...
 
-
   // Remove from role list
   ServerState::RoleEnum role = ServerState::instance()->getRole();
   std::string alk = ServerState::roleToAgencyListKey(role);
@@ -501,7 +502,6 @@ void ClusterFeature::unprepare() {
   }
 
   AgencyCommManager::MANAGER->stop();
-  ClusterComm::cleanup();
 
   ClusterInfo::cleanup();
 }

arangod/Scheduler/Scheduler.cpp

@@ -191,9 +191,6 @@ Scheduler::~Scheduler() {
   _managerGuard.reset();
   _managerContext.reset();
 
-  _serviceGuard.reset();
-  _ioContext.reset();
-
   FifoJob* job = nullptr;
 
   for (int i = 0; i < NUMBER_FIFOS; ++i) {
@@ -524,6 +521,9 @@ void Scheduler::shutdown() {
     std::this_thread::sleep_for(std::chrono::microseconds(20000));
   }
 
+  // One has to clean up the ioContext here, because there could a lambda
+  // in its queue, that requires for it finalization some object (for example vocbase)
+  // that would already be destroyed
   _managerContext.reset();
   _ioContext.reset();
 }