New upstream version 8.1.0

fsck/CMakeLists.txt

@@ -0,0 +1,153 @@
+include_directories(
+	source
+)
+
+add_library(
+	fsck STATIC
+	./source/toolkit/DatabaseTk.cpp
+	./source/toolkit/FsckDefinitions.h
+	./source/toolkit/FsckException.h
+	./source/toolkit/FsckDefinitions.cpp
+	./source/toolkit/FsckTkEx.cpp
+	./source/toolkit/DatabaseTk.h
+	./source/toolkit/FsckTkEx.h
+	./source/database/VectorSource.h
+	./source/database/DirInode.h
+	./source/database/FsckDBTable.cpp
+	./source/database/Select.h
+	./source/database/EntryID.h
+	./source/database/Union.h
+	./source/database/SetFragmentCursor.h
+	./source/database/FsID.h
+	./source/database/Set.h
+	./source/database/FsckDB.cpp
+	./source/database/SetFragment.h
+	./source/database/Group.h
+	./source/database/FsckDB.h
+	./source/database/FsckDBChecks.cpp
+	./source/database/Table.h
+	./source/database/Cursor.h
+	./source/database/UsedTarget.h
+	./source/database/FsckDBException.cpp
+	./source/database/FileInode.h
+	./source/database/ModificationEvent.h
+	./source/database/LeftJoinEq.h
+	./source/database/Filter.h
+	./source/database/FsckDBException.h
+	./source/database/DirEntry.h
+	./source/database/Chunk.h
+	./source/database/FsckDBTable.h
+	./source/database/ContDir.h
+	./source/database/StripeTargets.h
+	./source/database/Buffer.h
+	./source/database/Distinct.h
+	./source/database/DiskList.h
+	./source/net/message/NetMessageFactory.h
+	./source/net/message/NetMessageFactory.cpp
+	./source/net/message/nodes/HeartbeatMsgEx.cpp
+	./source/net/message/nodes/HeartbeatMsgEx.h
+	./source/net/message/testing/DummyMsgEx.h
+	./source/net/message/testing/DummyMsgEx.cpp
+	./source/net/message/fsck/FsckModificationEventMsgEx.cpp
+	./source/net/message/fsck/FsckModificationEventMsgEx.h
+	./source/net/msghelpers/MsgHelperRepair.h
+	./source/net/msghelpers/MsgHelperRepair.cpp
+	./source/components/DatagramListener.h
+	./source/components/ModificationEventHandler.h
+	./source/components/InternodeSyncer.h
+	./source/components/InternodeSyncer.cpp
+	./source/components/DataFetcher.cpp
+	./source/components/ModificationEventHandler.cpp
+	./source/components/DatagramListener.cpp
+	./source/components/worker/RetrieveDirEntriesWork.h
+	./source/components/worker/RetrieveInodesWork.h
+	./source/components/worker/AdjustChunkPermissionsWork.h
+	./source/components/worker/AdjustChunkPermissionsWork.cpp
+	./source/components/worker/RetrieveFsIDsWork.cpp
+	./source/components/worker/RetrieveChunksWork.cpp
+	./source/components/worker/RetrieveChunksWork.h
+	./source/components/worker/RetrieveInodesWork.cpp
+	./source/components/worker/RetrieveFsIDsWork.h
+	./source/components/worker/RetrieveDirEntriesWork.cpp
+	./source/components/DataFetcher.h
+	./source/modes/ModeHelp.cpp
+	./source/modes/Mode.h
+	./source/modes/ModeCheckFS.cpp
+	./source/modes/Mode.cpp
+	./source/modes/ModeEnableQuota.cpp
+	./source/modes/ModeHelp.h
+	./source/modes/ModeEnableQuota.h
+	./source/modes/ModeCheckFS.h
+	./source/program/Program.h
+	./source/program/Program.cpp
+	./source/program/Main.cpp
+	./source/app/App.h
+	./source/app/App.cpp
+	./source/app/config/Config.h
+	./source/app/config/Config.cpp
+)
+
+target_link_libraries(
+	fsck
+	beegfs-common
+	dl
+	pthread
+)
+	
+add_executable(
+	fsck.beegfs
+	source/program/Main.cpp
+)
+
+target_link_libraries(
+	fsck.beegfs
+	fsck
+)
+
+if(NOT BEEGFS_SKIP_TESTS)
+	# This is a dirty workaround used for a hardcoded path in
+	# "fsck/tests/TestConfig.cpp". Once the "defaultConfigFile" test does
+	# not steal the config file from "client_module" anymore, this
+	# workaround can be removed.
+	set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DTHIS_IS_A_CMAKE_BUILD=1")
+
+	add_executable(
+		test-fsck
+		./tests/TestFsckTk.cpp
+		./tests/TestDatabase.cpp
+		./tests/TestConfig.h
+		./tests/TestSet.cpp
+		./tests/TestCursors.cpp
+		./tests/TestTable.cpp
+		./tests/FlatTest.h
+		./tests/TestSetFragment.cpp
+		./tests/TestSerialization.cpp
+		./tests/FlatTest.cpp
+		./tests/TestTable.h
+		./tests/TestConfig.cpp
+		./tests/TestDatabase.h
+	)
+
+	target_link_libraries(
+		test-fsck
+		fsck
+		gtest_main
+	)
+
+	# required for a test
+	file(
+		COPY ${CMAKE_CURRENT_SOURCE_DIR}/../client_module/build/dist/etc/beegfs-client.conf
+		DESTINATION dist/etc/
+	)
+
+	add_test(
+		NAME test-fsck
+		COMMAND test-fsck --compiler
+	)
+endif()
+
+install(
+	TARGETS fsck.beegfs
+	DESTINATION "sbin"
+	COMPONENT "utils"
+)

fsck/build/Makefile

@@ -0,0 +1,28 @@
+include ../../build/Makefile
+
+main := ../source/program/Main.cpp
+sources := $(filter-out $(main), $(shell find ../source -iname '*.cpp'))
+
+$(call build-static-library,\
+   Fsck,\
+   $(sources),\
+   common dl nl3-route,\
+   ../source)
+
+$(call define-dep-lib,\
+   Fsck,\
+   -I ../source,\
+   $(build_dir)/libFsck.a)
+
+$(call build-executable,\
+   beegfs-fsck,\
+   $(main),\
+   Fsck common dl nl3-route)
+
+$(call build-test,\
+   test-runner,\
+   $(shell find ../tests -name '*.cpp'),\
+   Fsck common dl nl3-route,\
+   ../tests)
+
+CXXFLAGS += -DBOOST_RESULT_OF_USE_DECLTYPE

fsck/source/app/App.cpp

@@ -0,0 +1,562 @@
+#include <common/components/RegistrationDatagramListener.h>
+#include <program/Program.h>
+
+#include <toolkit/FsckTkEx.h>
+
+#include <csignal>
+#include <syslog.h>
+
+#define APP_WORKERS_DIRECT_NUM   1
+#define APP_SYSLOG_IDENTIFIER    "beegfs-fsck"
+
+App::App(int argc, char** argv)
+{
+   this->argc = argc;
+   this->argv = argv;
+
+   this->appResult = APPCODE_NO_ERROR;
+
+   this->cfg = NULL;
+   this->netFilter = NULL;
+   this->tcpOnlyFilter = NULL;
+   this->log = NULL;
+   this->mgmtNodes = NULL;
+   this->metaNodes = NULL;
+   this->storageNodes = NULL;
+   this->internodeSyncer = NULL;
+   this->targetMapper = NULL;
+   this->targetStateStore = NULL;
+   this->buddyGroupMapper = NULL;
+   this->metaBuddyGroupMapper = NULL;
+   this->workQueue = NULL;
+   this->ackStore = NULL;
+   this->netMessageFactory = NULL;
+   this->dgramListener = NULL;
+   this->modificationEventHandler = NULL;
+   this->runMode = NULL;
+
+}
+
+App::~App()
+{
+   // Note: Logging of the common lib classes is not working here, because this is called
+   // from class Program (so the thread-specific app-pointer isn't set in this context).
+
+   workersDelete();
+
+   SAFE_DELETE(this->dgramListener);
+   SAFE_DELETE(this->netMessageFactory);
+   SAFE_DELETE(this->ackStore);
+   SAFE_DELETE(this->workQueue);
+   SAFE_DELETE(this->storageNodes);
+   SAFE_DELETE(this->metaNodes);
+   SAFE_DELETE(this->mgmtNodes);
+   this->localNode.reset();
+   SAFE_DELETE(this->buddyGroupMapper);
+   SAFE_DELETE(this->metaBuddyGroupMapper);
+   SAFE_DELETE(this->targetStateStore);
+   SAFE_DELETE(this->targetMapper);
+   SAFE_DELETE(this->internodeSyncer);
+   SAFE_DELETE(this->log);
+   SAFE_DELETE(this->tcpOnlyFilter);
+   SAFE_DELETE(this->netFilter);
+   SAFE_DELETE(this->cfg);
+   SAFE_DELETE(this->runMode);
+
+   Logger::destroyLogger();
+   closelog();
+}
+
+void App::run()
+{
+   try
+   {
+      openlog(APP_SYSLOG_IDENTIFIER, LOG_NDELAY | LOG_PID | LOG_CONS, LOG_DAEMON);
+
+      this->cfg = new Config(argc, argv);
+
+      runNormal();
+   }
+   catch (InvalidConfigException& e)
+   {
+      std::cerr << std::endl;
+      std::cerr << "Error: " << e.what() << std::endl;
+      std::cerr << std::endl;
+      std::cerr << std::endl;
+
+      if(this->log)
+         log->logErr(e.what() );
+
+      this->appResult = APPCODE_INVALID_CONFIG;
+      return;
+   }
+   catch (std::exception& e)
+   {
+      std::cerr << std::endl;
+      std::cerr << "Unrecoverable error: " << e.what() << std::endl;
+      std::cerr << std::endl;
+
+      if(this->log)
+         log->logErr(e.what() );
+
+      appResult = APPCODE_RUNTIME_ERROR;
+      return;
+   }
+}
+
+void App::runNormal()
+{
+   bool componentsStarted = false;
+
+   auto runModeEnum = Program::getApp()->getConfig()->determineRunMode();
+
+   if (runModeEnum == RunMode_CHECKFS || runModeEnum == RunMode_ENABLEQUOTA)
+   {
+      // check if running as root
+      if(geteuid())
+      {
+         std::cerr << std::endl 
+               << "Running beegfs-fsck requires root privileges." 
+               << std::endl << std::endl << std::endl;
+         ModeHelp().execute();
+         appResult = APPCODE_INITIALIZATION_ERROR;
+         return;
+      }
+
+      if (runModeEnum == RunMode_CHECKFS)
+         runMode = new ModeCheckFS();
+      else if (runModeEnum == RunMode_ENABLEQUOTA)
+         runMode = new ModeEnableQuota();
+   }
+   else if (runModeEnum == RunMode_HELP)
+   {
+      appResult = ModeHelp().execute();
+      return;
+   }
+   else
+   {
+      ModeHelp().execute();
+      appResult = APPCODE_INVALID_CONFIG;
+      return;
+   }
+
+   initDataObjects(argc, argv);
+
+   // wait for mgmtd
+   if ( !cfg->getSysMgmtdHost().length() )
+      throw InvalidConfigException("Management host undefined");
+
+   bool mgmtWaitRes = waitForMgmtNode();
+   if(!mgmtWaitRes)
+   { // typically user just pressed ctrl+c in this case
+      log->logErr("Waiting for beegfs-mgmtd canceled");
+      appResult = APPCODE_RUNTIME_ERROR;
+      return;
+   }
+
+   // init components
+   try
+   {
+      initComponents();
+   }
+   catch (ComponentInitException& e)
+   {
+      FsckTkEx::fsckOutput(e.what(), OutputOptions_DOUBLELINEBREAK | OutputOptions_STDERR);
+
+      FsckTkEx::fsckOutput("A hard error occurred. BeeGFS Fsck will abort.",
+         OutputOptions_LINEBREAK | OutputOptions_STDERR);
+      appResult = APPCODE_INITIALIZATION_ERROR;
+      return;
+   }
+
+   // log system and configuration info
+
+   logInfos();
+
+   // start component threads
+   startComponents();
+   componentsStarted = true;
+
+   try
+   {
+      appResult = runMode->execute();
+   }
+   catch (InvalidConfigException& e)
+   {
+      ModeHelp modeHelp;
+      modeHelp.execute();
+
+      appResult = APPCODE_INVALID_CONFIG;
+   }
+   catch (std::exception &e)
+   {
+      FsckTkEx::fsckOutput("Unrecoverable error. BeeGFS Fsck will abort.",
+         OutputOptions_LINEBREAK | OutputOptions_ADDLINEBREAKBEFORE | OutputOptions_STDERR);
+      FsckTkEx::fsckOutput(e.what(), OutputOptions_LINEBREAK | OutputOptions_STDERR);
+   }
+
+   // self-termination
+   if(componentsStarted)
+   {
+      stopComponents();
+      joinComponents();
+   }
+}
+
+void App::initDataObjects(int argc, char** argv)
+{
+   this->netFilter = new NetFilter(cfg->getConnNetFilterFile() );
+   this->tcpOnlyFilter = new NetFilter(cfg->getConnTcpOnlyFilterFile() );
+
+   Logger::createLogger(cfg->getLogLevel(), cfg->getLogType(), cfg->getLogNoDate(),
+         cfg->getLogStdFile(), cfg->getLogNumLines(), cfg->getLogNumRotatedFiles());
+
+   this->log = new LogContext("App");
+
+   std::string interfacesFilename = this->cfg->getConnInterfacesFile();
+   if ( interfacesFilename.length() )
+      Config::loadStringListFile(interfacesFilename.c_str(), this->allowedInterfaces);
+
+   this->targetMapper = new TargetMapper();
+   this->targetStateStore = new TargetStateStore(NODETYPE_Storage);
+   this->buddyGroupMapper = new MirrorBuddyGroupMapper(targetMapper);
+   this->metaBuddyGroupMapper = new MirrorBuddyGroupMapper();
+
+   this->mgmtNodes = new NodeStoreServers(NODETYPE_Mgmt, false);
+   this->metaNodes = new NodeStoreServers(NODETYPE_Meta, false);
+   this->storageNodes = new NodeStoreServers(NODETYPE_Storage, false);
+
+   this->workQueue = new MultiWorkQueue();
+   this->ackStore = new AcknowledgmentStore();
+
+   initLocalNodeInfo();
+
+   registerSignalHandler();
+
+   this->netMessageFactory = new NetMessageFactory();
+}
+
+void App::findAllowedRDMAInterfaces(NicAddressList& outList) const
+{
+   Config* cfg = this->getConfig();
+
+   if(cfg->getConnUseRDMA() && RDMASocket::rdmaDevicesExist() )
+   {
+      bool foundRdmaInterfaces = NetworkInterfaceCard::checkAndAddRdmaCapability(outList);
+      if (foundRdmaInterfaces)
+         outList.sort(NetworkInterfaceCard::NicAddrComp{&allowedInterfaces}); // re-sort the niclist
+   }
+}
+
+void App::findAllowedInterfaces(NicAddressList& outList) const
+{
+   // discover local NICs and filter them
+   NetworkInterfaceCard::findAllInterfaces(allowedInterfaces, outList);
+   outList.sort(NetworkInterfaceCard::NicAddrComp{&allowedInterfaces});
+}
+
+void App::initLocalNodeInfo()
+{
+
+   findAllowedInterfaces(localNicList);
+   findAllowedRDMAInterfaces(localNicList);
+
+   if ( this->localNicList.empty() )
+      throw InvalidConfigException("Couldn't find any usable NIC");
+
+   initRoutingTable();
+   updateRoutingTable();
+
+   std::string nodeID = System::getHostname();
+
+   this->localNode = std::make_shared<LocalNode>(NODETYPE_Client, nodeID, NumNodeID(), 0, 0,
+         this->localNicList);
+}
+
+void App::initComponents()
+{
+   this->log->log(Log_DEBUG, "Initializing components...");
+
+   // Note: We choose a random udp port here to avoid conflicts with the client
+   unsigned short udpListenPort = 0;
+
+   this->dgramListener = new DatagramListener(netFilter, localNicList, ackStore, udpListenPort,
+      this->cfg->getConnRestrictOutboundInterfaces());
+
+   // update the local node info with udp port
+   this->localNode->updateInterfaces(dgramListener->getUDPPort(), 0, this->localNicList);
+
+   this->internodeSyncer = new InternodeSyncer();
+
+   workersInit();
+
+   this->log->log(Log_DEBUG, "Components initialized.");
+}
+
+void App::startComponents()
+{
+   log->log(Log_SPAM, "Starting up components...");
+
+   // make sure child threads don't receive SIGINT/SIGTERM (blocked signals are inherited)
+   PThread::blockInterruptSignals();
+
+   dgramListener->start();
+
+   internodeSyncer->start();
+   internodeSyncer->waitForServers();
+
+   workersStart();
+
+   PThread::unblockInterruptSignals(); // main app thread may receive SIGINT/SIGTERM
+
+   log->log(Log_DEBUG, "Components running.");
+}
+
+void App::stopComponents()
+{
+   // note: this method may not wait for termination of the components, because that could
+   //    lead to a deadlock (when calling from signal handler)
+   workersStop();
+
+   if(this->internodeSyncer)
+      this->internodeSyncer->selfTerminate();
+
+   if ( dgramListener )
+   {
+      dgramListener->selfTerminate();
+      dgramListener->sendDummyToSelfUDP();
+   }
+
+   this->selfTerminate();
+}
+
+void App::joinComponents()
+{
+   log->log(4, "Joining component threads...");
+
+   this->internodeSyncer->join();
+
+   /* (note: we need one thread for which we do an untimed join, so this should be a quite reliably
+      terminating thread) */
+   this->dgramListener->join();
+
+   workersJoin();
+}
+
+void App::workersInit()
+{
+   unsigned numWorkers = cfg->getTuneNumWorkers();
+
+   for(unsigned i=0; i < numWorkers; i++)
+   {
+      Worker* worker = new Worker(
+         std::string("Worker") + StringTk::intToStr(i+1), workQueue, QueueWorkType_INDIRECT);
+      workerList.push_back(worker);
+   }
+
+   for(unsigned i=0; i < APP_WORKERS_DIRECT_NUM; i++)
+   {
+      Worker* worker = new Worker(
+         std::string("DirectWorker") + StringTk::intToStr(i+1), workQueue, QueueWorkType_DIRECT);
+      workerList.push_back(worker);
+   }
+}
+
+void App::workersStart()
+{
+   for ( WorkerListIter iter = workerList.begin(); iter != workerList.end(); iter++ )
+   {
+      (*iter)->start();
+   }
+}
+
+void App::workersStop()
+{
+   // need two loops because we don't know if the worker that handles the work will be the same that
+   // received the self-terminate-request
+   for(WorkerListIter iter = workerList.begin(); iter != workerList.end(); iter++)
+   {
+      (*iter)->selfTerminate();
+   }
+
+   for(WorkerListIter iter = workerList.begin(); iter != workerList.end(); iter++)
+   {
+      workQueue->addDirectWork(new DummyWork() );
+   }
+}
+
+void App::workersDelete()
+{
+   for(WorkerListIter iter = workerList.begin(); iter != workerList.end(); iter++)
+   {
+      delete(*iter);
+   }
+
+   workerList.clear();
+}
+
+void App::workersJoin()
+{
+   for(WorkerListIter iter = workerList.begin(); iter != workerList.end(); iter++)
+   {
+      waitForComponentTermination(*iter);
+   }
+}
+
+void App::logInfos()
+{
+   // print software version (BEEGFS_VERSION)
+   log->log(Log_NOTICE, std::string("Version: ") + BEEGFS_VERSION);
+
+   // print debug version info
+   LOG_DEBUG_CONTEXT(*log, Log_CRITICAL, "--DEBUG VERSION--");
+
+   // print local nodeID
+   log->log(Log_NOTICE, std::string("LocalNode: ") + localNode->getTypedNodeID() );
+
+   // list usable network interfaces
+   NicAddressList nicList(localNode->getNicList());
+   logUsableNICs(log, nicList);
+
+   // print net filters
+   if ( netFilter->getNumFilterEntries() )
+   {
+      this->log->log(2,
+         std::string("Net filters: ") + StringTk::uintToStr(netFilter->getNumFilterEntries()));
+   }
+
+   if(tcpOnlyFilter->getNumFilterEntries() )
+   {
+      this->log->log(Log_WARNING, std::string("TCP-only filters: ") +
+         StringTk::uintToStr(tcpOnlyFilter->getNumFilterEntries() ) );
+   }
+}
+
+/**
+ * Request mgmt heartbeat and wait for the mgmt node to appear in nodestore.
+ *
+ * @return true if mgmt heartbeat received, false on error or thread selftermination order
+ */
+bool App::waitForMgmtNode()
+{
+   const unsigned waitTimeoutMS = 0; // infinite wait
+   const unsigned nameResolutionRetries = 3;
+
+   // choose a random udp port here
+   unsigned udpListenPort = 0;
+   unsigned udpMgmtdPort = cfg->getConnMgmtdPort();
+   std::string mgmtdHost = cfg->getSysMgmtdHost();
+
+   RegistrationDatagramListener regDGramLis(this->netFilter, this->localNicList, this->ackStore,
+      udpListenPort, this->cfg->getConnRestrictOutboundInterfaces());
+
+   regDGramLis.start();
+
+   log->log(Log_CRITICAL, "Waiting for beegfs-mgmtd@" +
+      mgmtdHost + ":" + StringTk::uintToStr(udpMgmtdPort) + "...");
+
+   bool gotMgmtd = NodesTk::waitForMgmtHeartbeat(
+         this, &regDGramLis, this->mgmtNodes, mgmtdHost, udpMgmtdPort, waitTimeoutMS,
+         nameResolutionRetries);
+
+   regDGramLis.selfTerminate();
+   regDGramLis.sendDummyToSelfUDP(); // for faster termination
+
+   regDGramLis.join();
+
+   return gotMgmtd;
+}
+
+void App::updateLocalNicList(NicAddressList& localNicList)
+{
+   std::vector<AbstractNodeStore*> allNodes({ mgmtNodes, metaNodes, storageNodes});
+   updateLocalNicListAndRoutes(log, localNicList, allNodes);
+   localNode->updateInterfaces(0, 0, localNicList);
+   dgramListener->setLocalNicList(localNicList);
+}
+
+/*
+ * Handles expections that lead to the termination of a component.
+ * Initiates an application shutdown.
+ */
+
+void App::handleComponentException(std::exception& e)
+{
+   const char* logContext = "App (component exception handler)";
+   LogContext log(logContext);
+
+   const auto componentName = PThread::getCurrentThreadName();
+
+   log.logErr(
+         "The component [" + componentName + "] encountered an unrecoverable error. " +
+         std::string("[SysErr: ") + System::getErrString() + "] " +
+         std::string("Exception message: ") + e.what() );
+
+   log.log(2, "Shutting down...");
+
+   shallAbort.set(1);
+   stopComponents();
+}
+
+void App::handleNetworkInterfaceFailure(const std::string& devname)
+{
+   LOG(GENERAL, ERR, "Network interface failure.",
+      ("Device", devname));
+   internodeSyncer->setForceCheckNetwork();
+}
+
+void App::registerSignalHandler()
+{
+   signal(SIGINT, App::signalHandler);
+   signal(SIGTERM, App::signalHandler);
+}
+
+void App::signalHandler(int sig)
+{
+   App* app = Program::getApp();
+
+   Logger* log = Logger::getLogger();
+   const char* logContext = "App::signalHandler";
+
+   // note: this might deadlock if the signal was thrown while the logger mutex is locked by the
+   //    application thread (depending on whether the default mutex style is recursive). but
+   //    even recursive mutexes are not acceptable in this case.
+   //    we need something like a timed lock for the log mutex. if it succeeds within a
+   //    few seconds, we know that we didn't hold the mutex lock. otherwise we simply skip the
+   //    log message. this will only work if the mutex is non-recusive (which is unknown for
+   //    the default mutex style).
+   //    but it is very unlikely that the application thread holds the log mutex, because it
+   //    joins the component threads and so it doesn't do anything else but sleeping!
+
+   switch(sig)
+   {
+      case SIGINT:
+      {
+         signal(sig, SIG_DFL); // reset the handler to its default
+         log->log(1, logContext, "Received a SIGINT. Shutting down...");
+      }
+         break;
+
+      case SIGTERM:
+      {
+         signal(sig, SIG_DFL); // reset the handler to its default
+         log->log(1, logContext, "Received a SIGTERM. Shutting down...");
+      }
+         break;
+
+      default:
+      {
+         signal(sig, SIG_DFL); // reset the handler to its default
+         log->log(1, logContext, "Received an unknown signal. Shutting down...");
+      }
+         break;
+   }
+
+   app->abort();
+}
+
+void App::abort()
+{
+   shallAbort.set(1);
+   stopComponents();
+}

fsck/source/app/App.h

@@ -0,0 +1,223 @@
+#ifndef APP_H_
+#define APP_H_
+
+#include <app/config/Config.h>
+#include <common/app/log/LogContext.h>
+#include <common/app/log/Logger.h>
+#include <common/app/AbstractApp.h>
+#include <common/nodes/NodeStore.h>
+#include <common/toolkit/AcknowledgmentStore.h>
+#include <common/toolkit/NetFilter.h>
+#include <common/toolkit/NodesTk.h>
+#include <common/Common.h>
+#include <common/components/worker/Worker.h>
+#include <common/components/worker/DummyWork.h>
+#include <common/components/ComponentInitException.h>
+#include <common/net/sock/RDMASocket.h>
+#include <common/nodes/RootInfo.h>
+#include <common/nodes/TargetMapper.h>
+#include <common/nodes/LocalNode.h>
+#include <components/DatagramListener.h>
+#include <components/InternodeSyncer.h>
+#include <components/ModificationEventHandler.h>
+#include <modes/ModeHelp.h>
+#include <modes/ModeCheckFS.h>
+#include <modes/ModeEnableQuota.h>
+#include <net/message/NetMessageFactory.h>
+
+
+#ifndef BEEGFS_VERSION
+   #error BEEGFS_VERSION undefined
+#endif
+
+// program return codes
+#define APPCODE_NO_ERROR               0
+#define APPCODE_INVALID_CONFIG         1
+#define APPCODE_INITIALIZATION_ERROR   2
+#define APPCODE_COMMUNICATION_ERROR    3
+#define APPCODE_RUNTIME_ERROR          4
+#define APPCODE_USER_ABORTED           5
+
+typedef std::list<Worker*> WorkerList;
+typedef WorkerList::iterator WorkerListIter;
+
+
+// forward declarations
+class LogContext;
+
+class App : public AbstractApp
+{
+   public:
+      App(int argc, char** argv);
+      virtual ~App();
+
+      virtual void run() override;
+
+      void abort();
+      virtual void handleComponentException(std::exception& e) override;
+      virtual void handleNetworkInterfaceFailure(const std::string& devname) override;
+
+      bool getShallAbort()
+      {
+         return (shallAbort.read() != 0);
+      }
+
+   private:
+      int appResult;
+      int argc;
+      char** argv;
+
+      AtomicSizeT shallAbort;
+
+      Config* cfg;
+      LogContext* log;
+
+      NetFilter* netFilter;
+      NetFilter* tcpOnlyFilter; // for IPs that allow only plain TCP (no RDMA etc)
+      std::list<std::string> allowedInterfaces;
+
+      std::shared_ptr<Node> localNode;
+      NodeStore* mgmtNodes;
+      NodeStore* metaNodes;
+      NodeStore* storageNodes;
+      RootInfo metaRoot;
+      InternodeSyncer* internodeSyncer;
+      TargetMapper* targetMapper;
+      TargetStateStore* targetStateStore;
+      MirrorBuddyGroupMapper* buddyGroupMapper;
+      MirrorBuddyGroupMapper* metaBuddyGroupMapper;
+      MultiWorkQueue* workQueue;
+      AcknowledgmentStore* ackStore;
+      NetMessageFactory* netMessageFactory;
+
+      DatagramListener* dgramListener;
+      WorkerList workerList;
+
+      ModificationEventHandler* modificationEventHandler;
+
+      Mode* runMode;
+
+      void runNormal();
+
+      void workersInit();
+      void workersStart();
+      void workersStop();
+      void workersDelete();
+      void workersJoin();
+      void initDataObjects(int argc, char** argv);
+      void initLocalNodeInfo();
+      void initComponents();
+      bool initRunMode();
+      void startComponents();
+      void joinComponents();
+      virtual void stopComponents() override;
+      void logInfos();
+      bool waitForMgmtNode();
+      void registerSignalHandler();
+      static void signalHandler(int sig);
+
+   public:
+      virtual const ICommonConfig* getCommonConfig() const override
+      {
+         return cfg;
+      }
+
+      virtual const AbstractNetMessageFactory* getNetMessageFactory() const override
+      {
+         return netMessageFactory;
+      }
+
+      virtual const NetFilter* getNetFilter() const override
+      {
+         return netFilter;
+      }
+
+      virtual const NetFilter* getTcpOnlyFilter() const override
+      {
+         return tcpOnlyFilter;
+      }
+
+      Config* getConfig() const
+      {
+         return cfg;
+      }
+
+      DatagramListener* getDatagramListener() const
+      {
+         return dgramListener;
+      }
+
+      int getAppResult() const
+      {
+         return appResult;
+      }
+
+      NodeStore* getMgmtNodes() const
+      {
+         return mgmtNodes;
+      }
+
+      NodeStore* getMetaNodes()
+      {
+         return metaNodes;
+      }
+
+      NodeStore* getStorageNodes() const
+      {
+         return storageNodes;
+      }
+
+      Node& getLocalNode() const
+      {
+         return *localNode;
+      }
+
+      void updateLocalNicList(NicAddressList& localNicList);
+
+      MultiWorkQueue* getWorkQueue()
+      {
+         return workQueue;
+      }
+
+      InternodeSyncer* getInternodeSyncer()
+      {
+         return internodeSyncer;
+      }
+
+      TargetMapper* getTargetMapper()
+      {
+         return targetMapper;
+      }
+
+      TargetStateStore* getTargetStateStore()
+      {
+         return targetStateStore;
+      }
+
+      MirrorBuddyGroupMapper* getMirrorBuddyGroupMapper()
+      {
+         return buddyGroupMapper;
+      }
+
+      MirrorBuddyGroupMapper* getMetaMirrorBuddyGroupMapper()
+      {
+         return metaBuddyGroupMapper;
+      }
+
+      ModificationEventHandler* getModificationEventHandler()
+      {
+         return modificationEventHandler;
+      }
+
+      void setModificationEventHandler(ModificationEventHandler* handler)
+      {
+         modificationEventHandler = handler;
+      }
+
+      const RootInfo& getMetaRoot() const { return metaRoot; }
+      RootInfo& getMetaRoot() { return metaRoot; }
+      void findAllowedInterfaces(NicAddressList& outList) const;
+      void findAllowedRDMAInterfaces(NicAddressList& outList) const;
+};
+
+#endif // APP_H_

fsck/source/app/config/Config.cpp

@@ -0,0 +1,317 @@
+#include <common/toolkit/StringTk.h>
+#include "Config.h"
+
+#define IGNORE_CONFIG_CLIENT_VALUE(keyStr) /* to be used in applyConfigMap() */ \
+   if(testConfigMapKeyMatch(iter, keyStr, addDashes) ) \
+      ; \
+   else
+
+// Note: Keep in sync with enum RunMode
+RunModesElem const __RunModes[] =
+{
+   { "--checkfs", RunMode_CHECKFS },
+   { "--enablequota", RunMode_ENABLEQUOTA },
+   { NULL, RunMode_INVALID }
+};
+
+Config::Config(int argc, char** argv) :
+   AbstractConfig(argc, argv)
+{
+   initConfig(argc, argv, false, true);
+   logType = LogType_LOGFILE;
+}
+
+/**
+ * Determine RunMode from config.
+ * If a valid RunMode exists in the config, the corresponding config element will be erased.
+ */
+enum RunMode Config::determineRunMode()
+{
+   /* test for given help argument, e.g. in case the user wants to see mode-specific help with
+      arguments "--help --<mode>". */
+
+   StringMapIter iter = configMap.find(RUNMODE_HELP_KEY_STRING);
+   if(iter != configMap.end() )
+   { // user did specify "--help"
+      /* note: it's important to remove the help arg here, because mode help will call this again
+         to find out whether user wants to see mode-specific help. */
+      eraseFromConfigMap(iter);
+      return RunMode_HELP;
+   }
+
+   // walk all defined modes to check whether we find any of them in the config
+
+   for(int i=0; __RunModes[i].modeString != NULL; i++)
+   {
+      iter = configMap.find(__RunModes[i].modeString);
+      if(iter != configMap.end() )
+      { // we found a valid mode in the config
+         eraseFromConfigMap(iter);
+         return __RunModes[i].runMode;
+      }
+   }
+
+   // no valid mode found
+
+   return RunMode_INVALID;
+}
+
+/**
+ * Sets the default values for each configurable in the configMap.
+ *
+ * @param addDashes true to prepend "--" to all config keys.
+ */
+void Config::loadDefaults(bool addDashes)
+{
+   AbstractConfig::loadDefaults(addDashes);
+
+   // re-definitions
+   configMapRedefine("cfgFile", createDefaultCfgFilename(), addDashes);
+
+   // own definitions
+   configMapRedefine("connInterfacesFile", "", addDashes);
+
+   configMapRedefine("tuneNumWorkers", "32", addDashes);
+   configMapRedefine("tunePreferredNodesFile", "", addDashes);
+   configMapRedefine("tuneDbFragmentSize", "0", addDashes);
+   configMapRedefine("tuneDentryCacheSize", "0", addDashes);
+
+   configMapRedefine("runDaemonized", "false", addDashes);
+
+   configMapRedefine("databasePath", CONFIG_DEFAULT_DBPATH, addDashes);
+
+   configMapRedefine("overwriteDbFile", "false", addDashes);
+
+   configMapRedefine("testDatabasePath", CONFIG_DEFAULT_TESTDBPATH, addDashes);
+
+   configMapRedefine("DatabaseNumMaxConns", "16", addDashes);
+
+   configMapRedefine("overrideRootMDS", "", addDashes);
+
+   configMapRedefine("logStdFile", CONFIG_DEFAULT_LOGFILE, addDashes);
+   configMapRedefine("logOutFile", CONFIG_DEFAULT_OUTFILE, addDashes);
+   configMapRedefine("logNoDate", "false", addDashes);
+
+   configMapRedefine("readOnly", "false", addDashes);
+
+   configMapRedefine("noFetch", "false", addDashes);
+
+   configMapRedefine("automatic", "false", addDashes);
+
+   configMapRedefine("runOffline", "false", addDashes);
+
+   configMapRedefine("forceRestart", "false", addDashes);
+
+   configMapRedefine("quotaEnabled", "false", addDashes);
+
+   configMapRedefine("ignoreDBDiskSpace", "false", addDashes);
+}
+
+/**
+ * @param addDashes true to prepend "--" to tested config keys for matching.
+ */
+void Config::applyConfigMap(bool enableException, bool addDashes)
+{
+   AbstractConfig::applyConfigMap(false, addDashes);
+
+   for (StringMapIter iter = configMap.begin(); iter != configMap.end();)
+   {
+      bool unknownElement = false;
+
+      IGNORE_CONFIG_CLIENT_VALUE("logClientID")
+      IGNORE_CONFIG_CLIENT_VALUE("logType")
+      IGNORE_CONFIG_CLIENT_VALUE("connNumCommRetries")
+      IGNORE_CONFIG_CLIENT_VALUE("connUnmountRetries")
+      IGNORE_CONFIG_CLIENT_VALUE("connCommRetrySecs")
+      IGNORE_CONFIG_CLIENT_VALUE("connMaxConcurrentAttempts")
+      IGNORE_CONFIG_CLIENT_VALUE("connRDMAInterfacesFile")
+      IGNORE_CONFIG_CLIENT_VALUE("connTCPFallbackEnabled")
+      IGNORE_CONFIG_CLIENT_VALUE("connMessagingTimeouts")
+      IGNORE_CONFIG_CLIENT_VALUE("connInterfacesList")
+      IGNORE_CONFIG_CLIENT_VALUE("connRDMATimeouts")
+      IGNORE_CONFIG_CLIENT_VALUE("connRDMAFragmentSize")
+      IGNORE_CONFIG_CLIENT_VALUE("connRDMAKeyType")
+      IGNORE_CONFIG_CLIENT_VALUE("connRDMAMetaFragmentSize")
+      IGNORE_CONFIG_CLIENT_VALUE("connRDMAMetaBufNum")
+      IGNORE_CONFIG_CLIENT_VALUE("connRDMAMetaBufSize")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneFileCacheType")
+      IGNORE_CONFIG_CLIENT_VALUE("tunePagedIOBufSize")
+      IGNORE_CONFIG_CLIENT_VALUE("tunePagedIOBufNum")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneFileCacheBufSize")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneFileCacheBufNum")
+      IGNORE_CONFIG_CLIENT_VALUE("tunePageCacheValidityMS")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneAttribCacheValidityMS")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneMaxWriteWorks")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneMaxReadWorks")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneAllowMultiSetWrite")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneAllowMultiSetRead")
+      IGNORE_CONFIG_CLIENT_VALUE("tunePathBufSize")
+      IGNORE_CONFIG_CLIENT_VALUE("tunePathBufNum")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneMaxReadWriteNum")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneMaxReadWriteNodesNum")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneMsgBufSize")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneMsgBufNum")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneRemoteFSync")
+      IGNORE_CONFIG_CLIENT_VALUE("tunePreferredMetaFile")
+      IGNORE_CONFIG_CLIENT_VALUE("tunePreferredStorageFile")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneUseGlobalFileLocks")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneRefreshOnGetAttr")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneInodeBlockBits")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneInodeBlockSize")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneMaxClientMirrorSize") // was removed, kept here for compat
+      IGNORE_CONFIG_CLIENT_VALUE("tuneEarlyCloseResponse")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneUseGlobalAppendLocks")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneUseBufferedAppend")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneStatFsCacheSecs")
+      IGNORE_CONFIG_CLIENT_VALUE("sysCacheInvalidationVersion")
+      IGNORE_CONFIG_CLIENT_VALUE("sysCreateHardlinksAsSymlinks")
+      IGNORE_CONFIG_CLIENT_VALUE("sysMountSanityCheckMS")
+      IGNORE_CONFIG_CLIENT_VALUE("sysSyncOnClose")
+      IGNORE_CONFIG_CLIENT_VALUE("sysSessionCheckOnClose")
+      IGNORE_CONFIG_CLIENT_VALUE("sysSessionChecksEnabled")
+      IGNORE_CONFIG_CLIENT_VALUE("sysTargetOfflineTimeoutSecs")
+      IGNORE_CONFIG_CLIENT_VALUE("sysInodeIDStyle")
+      IGNORE_CONFIG_CLIENT_VALUE("sysACLsEnabled")
+      IGNORE_CONFIG_CLIENT_VALUE("sysXAttrsEnabled")
+      IGNORE_CONFIG_CLIENT_VALUE("sysBypassFileAccessCheckOnMeta")
+      IGNORE_CONFIG_CLIENT_VALUE("sysXAttrsCheckCapabilities")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneDirSubentryCacheValidityMS")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneFileSubentryCacheValidityMS")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneENOENTCacheValidityMS")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneCoherentBuffers")
+      IGNORE_CONFIG_CLIENT_VALUE("sysFileEventLogMask")
+      IGNORE_CONFIG_CLIENT_VALUE("sysRenameEbusyAsXdev")
+      IGNORE_CONFIG_CLIENT_VALUE("tuneNumRetryWorkers")
+      IGNORE_CONFIG_CLIENT_VALUE("connHelperdPortTCP") // was removed, kept here for compat
+
+      if (testConfigMapKeyMatch(iter, "connInterfacesFile", addDashes))
+         connInterfacesFile = iter->second;
+      else if (testConfigMapKeyMatch(iter, "tuneNumWorkers", addDashes))
+         tuneNumWorkers = StringTk::strToUInt(iter->second);
+      else if (testConfigMapKeyMatch(iter, "tunePreferredNodesFile", addDashes))
+         tunePreferredNodesFile = iter->second;
+      else if (testConfigMapKeyMatch(iter, "tuneDbFragmentSize", addDashes))
+         tuneDbFragmentSize = StringTk::strToUInt64(iter->second.c_str());
+      else if (testConfigMapKeyMatch(iter, "tuneDentryCacheSize", addDashes))
+         tuneDentryCacheSize = StringTk::strToUInt64(iter->second.c_str());
+      else if (testConfigMapKeyMatch(iter, "runDaemonized", addDashes))
+         runDaemonized = StringTk::strToBool(iter->second);
+      else if (testConfigMapKeyMatch(iter, "databasePath", addDashes))
+         databasePath = iter->second;
+      else if (testConfigMapKeyMatch(iter, "overwriteDbFile", addDashes))
+         overwriteDbFile = StringTk::strToBool(iter->second);
+      else if (testConfigMapKeyMatch(iter, "testDatabasePath", addDashes))
+         testDatabasePath = iter->second;
+      else if (testConfigMapKeyMatch(iter, "databaseNumMaxConns", addDashes))
+         databaseNumMaxConns = StringTk::strHexToUInt(iter->second);
+      else if (testConfigMapKeyMatch(iter, "overrideRootMDS", addDashes))
+         overrideRootMDS = iter->second;
+      else if (testConfigMapKeyMatch(iter, "logStdFile", addDashes))
+         logStdFile = iter->second;
+      else if (testConfigMapKeyMatch(iter, "logOutFile", addDashes))
+         logOutFile = iter->second;
+      else if (testConfigMapKeyMatch(iter, "readOnly", addDashes))
+         readOnly = StringTk::strToBool(iter->second);
+      else if (testConfigMapKeyMatch(iter, "noFetch", addDashes))
+         noFetch = StringTk::strToBool(iter->second);
+      else if (testConfigMapKeyMatch(iter, "automatic", addDashes))
+         automatic = StringTk::strToBool(iter->second);
+      else if (testConfigMapKeyMatch(iter, "runOffline", addDashes))
+         runOffline = StringTk::strToBool(iter->second);
+      else if (testConfigMapKeyMatch(iter, "forceRestart", addDashes))
+         forceRestart = StringTk::strToBool(iter->second);
+      else if (testConfigMapKeyMatch(iter, "quotaEnabled", addDashes))
+         quotaEnabled = StringTk::strToBool(iter->second);
+      else if (testConfigMapKeyMatch(iter, "ignoreDBDiskSpace", addDashes))
+         ignoreDBDiskSpace = StringTk::strToBool(iter->second);
+      else if (testConfigMapKeyMatch(iter, "checkMalformedChunk", addDashes))
+         checkFsActions.set(CHECK_MALFORMED_CHUNK);
+      else if (testConfigMapKeyMatch(iter, "checkFilesWithMissingTargets", addDashes))
+         checkFsActions.set(CHECK_FILES_WITH_MISSING_TARGETS);
+      else if (testConfigMapKeyMatch(iter, "checkOrphanedDentryByIDFiles", addDashes))
+         checkFsActions.set(CHECK_ORPHANED_DENTRY_BYIDFILES);
+      else if (testConfigMapKeyMatch(iter, "checkDirEntriesWithBrokenIDFile", addDashes))
+         checkFsActions.set(CHECK_DIRENTRIES_WITH_BROKENIDFILE);
+      else if (testConfigMapKeyMatch(iter, "checkOrphanedChunk", addDashes))
+         checkFsActions.set(CHECK_ORPHANED_CHUNK);
+      else if (testConfigMapKeyMatch(iter, "checkChunksInWrongPath", addDashes))
+         checkFsActions.set(CHECK_CHUNKS_IN_WRONGPATH);
+      else if (testConfigMapKeyMatch(iter, "checkWrongInodeOwner", addDashes))
+         checkFsActions.set(CHECK_WRONG_INODE_OWNER);
+      else if (testConfigMapKeyMatch(iter, "checkWrongOwnerInDentry", addDashes))
+         checkFsActions.set(CHECK_WRONG_OWNER_IN_DENTRY);
+      else if (testConfigMapKeyMatch(iter, "checkOrphanedContDir", addDashes))
+         checkFsActions.set(CHECK_ORPHANED_CONT_DIR);
+      else if (testConfigMapKeyMatch(iter, "checkOrphanedDirInode", addDashes))
+         checkFsActions.set(CHECK_ORPHANED_DIR_INODE);
+      else if (testConfigMapKeyMatch(iter, "checkOrphanedFileInode", addDashes))
+         checkFsActions.set(CHECK_ORPHANED_FILE_INODE);
+      else if (testConfigMapKeyMatch(iter, "checkDanglingDentry", addDashes))
+         checkFsActions.set(CHECK_DANGLING_DENTRY);
+      else if (testConfigMapKeyMatch(iter, "checkMissingContDir", addDashes))
+         checkFsActions.set(CHECK_MISSING_CONT_DIR);
+      else if (testConfigMapKeyMatch(iter, "checkWrongFileAttribs", addDashes))
+         checkFsActions.set(CHECK_WRONG_FILE_ATTRIBS);
+      else if (testConfigMapKeyMatch(iter, "checkWrongDirAttribs", addDashes))
+         checkFsActions.set(CHECK_WRONG_DIR_ATTRIBS);
+      else if (testConfigMapKeyMatch(iter, "checkOldStyledHardlinks", addDashes))
+         checkFsActions.set(CHECK_OLD_STYLED_HARDLINKS);
+      else
+      {
+         // unknown element occurred
+         unknownElement = true;
+
+         if (enableException)
+         {
+            throw InvalidConfigException("The config argument '" + iter->first + "' is invalid");
+         }
+      }
+
+      // advance iterator (and remove handled element)
+
+      if (unknownElement)
+      {
+         // just skip the unknown element
+         iter++;
+      }
+      else
+      {
+         // remove this element from the map
+         iter = eraseFromConfigMap(iter);
+      }
+   }
+}
+
+void Config::initImplicitVals()
+{
+   // tuneNumWorkers
+   if (!tuneNumWorkers)
+      tuneNumWorkers = BEEGFS_MAX(System::getNumOnlineCPUs() * 2, 4);
+
+   if (!tuneDbFragmentSize)
+      tuneDbFragmentSize = uint64_t(sysconf(_SC_PHYS_PAGES) ) * sysconf(_SC_PAGESIZE) / 2;
+
+   // just blindly assume that 384 bytes will be enough for a single cache entry. should be
+   if (!tuneDentryCacheSize)
+      tuneDentryCacheSize = tuneDbFragmentSize / 384;
+
+   // read in connAuthFile only if we are running as root.
+   // if not root, the program will abort anyway
+   if(!geteuid())
+   {
+      AbstractConfig::initConnAuthHash(connAuthFile, &connAuthHash);
+   }
+}
+
+std::string Config::createDefaultCfgFilename() const
+{
+   struct stat statBuf;
+
+   const int statRes = stat(CONFIG_DEFAULT_CFGFILENAME, &statBuf);
+
+   if (!statRes && S_ISREG(statBuf.st_mode))
+      return CONFIG_DEFAULT_CFGFILENAME; // there appears to be a config file
+
+   return ""; // no default file otherwise
+}
+

fsck/source/app/config/Config.h

@@ -0,0 +1,224 @@
+#ifndef CONFIG_H_
+#define CONFIG_H_
+
+#include <common/app/config/AbstractConfig.h>
+#include <bitset>
+
+#define CONFIG_DEFAULT_CFGFILENAME "/etc/beegfs/beegfs-client.conf"
+#define CONFIG_DEFAULT_LOGFILE     "/var/log/beegfs-fsck.log"
+#define CONFIG_DEFAULT_OUTFILE     "/var/log/beegfs-fsck.out"
+#define CONFIG_DEFAULT_DBPATH      "/var/lib/beegfs/"
+#define CONFIG_DEFAULT_TESTDBPATH  "/tmp/beegfs-fsck/"
+
+#define RUNMODE_HELP_KEY_STRING     "--help" /* key for usage help */
+#define __RUNMODES_SIZE \
+   ( (sizeof(__RunModes) ) / (sizeof(RunModesElem) ) - 1)
+   /* -1 because last elem is NULL */
+
+enum CheckFsActions
+{
+   CHECK_MALFORMED_CHUNK               = 0,
+   CHECK_FILES_WITH_MISSING_TARGETS    = 1,
+   CHECK_ORPHANED_DENTRY_BYIDFILES     = 2,
+   CHECK_DIRENTRIES_WITH_BROKENIDFILE  = 3,
+   CHECK_ORPHANED_CHUNK                = 4,
+   CHECK_CHUNKS_IN_WRONGPATH           = 5,
+   CHECK_WRONG_INODE_OWNER             = 6,
+   CHECK_WRONG_OWNER_IN_DENTRY         = 7,
+   CHECK_ORPHANED_CONT_DIR             = 8,
+   CHECK_ORPHANED_DIR_INODE            = 9,
+   CHECK_ORPHANED_FILE_INODE           = 10,
+   CHECK_DANGLING_DENTRY               = 11,
+   CHECK_MISSING_CONT_DIR              = 12,
+   CHECK_WRONG_FILE_ATTRIBS            = 13,
+   CHECK_WRONG_DIR_ATTRIBS             = 14,
+   CHECK_OLD_STYLED_HARDLINKS          = 15,
+   CHECK_FS_ACTIONS_COUNT              = 16
+};
+
+// Note: Keep in sync with __RunModes array
+enum RunMode
+{
+   RunMode_CHECKFS              =     0,
+   RunMode_ENABLEQUOTA          =     1,
+   RunMode_HELP                 =     2,
+   RunMode_INVALID              =     3 /* not valid as index in RunModes array */
+};
+
+struct RunModesElem
+{
+   const char* modeString;
+   enum RunMode runMode;
+};
+
+
+extern RunModesElem const __RunModes[];
+
+class Config : public AbstractConfig
+{
+   public:
+      Config(int argc, char** argv);
+
+      enum RunMode determineRunMode();
+
+   private:
+
+      // configurables
+
+      std::string connInterfacesFile;
+
+      unsigned    tuneNumWorkers;
+      std::string tunePreferredNodesFile;
+      size_t      tuneDbFragmentSize;
+      size_t      tuneDentryCacheSize;
+
+      bool        runDaemonized;
+
+      std::string databasePath;
+      bool        overwriteDbFile;
+
+      // only relevant for unit testing, to give the used databasePath
+      std::string testDatabasePath;
+
+      unsigned    databaseNumMaxConns;
+
+      std::string overrideRootMDS; // not tested well, should only be used by developers
+
+      // file for fsck output (not the log messages, but the output, which is also on the console)
+      std::string logOutFile;
+
+      bool        readOnly;
+      bool        noFetch;
+      bool        automatic;
+      bool        runOffline;
+      bool        forceRestart;
+      bool        quotaEnabled;
+      bool        ignoreDBDiskSpace;
+
+      std::bitset<CHECK_FS_ACTIONS_COUNT> checkFsActions;
+
+      // internals
+      virtual void loadDefaults(bool addDashes) override;
+      virtual void applyConfigMap(bool enableException, bool addDashes) override;
+      virtual void initImplicitVals() override;
+      std::string createDefaultCfgFilename() const;
+
+   public:
+      // getters & setters
+      const StringMap* getUnknownConfigArgs() const
+      {
+         return getConfigMap();
+      }
+
+      const std::string& getConnInterfacesFile() const
+      {
+         return connInterfacesFile;
+      }
+
+      unsigned getTuneNumWorkers() const
+      {
+         return tuneNumWorkers;
+      }
+
+      size_t getTuneDbFragmentSize() const
+      {
+         return tuneDbFragmentSize;
+      }
+
+      size_t getTuneDentryCacheSize() const
+      {
+         return tuneDentryCacheSize;
+      }
+
+      const std::string& getTunePreferredNodesFile() const
+      {
+         return tunePreferredNodesFile;
+      }
+
+      bool getRunDaemonized() const
+      {
+         return runDaemonized;
+      }
+
+      const std::string& getDatabasePath() const
+      {
+         return databasePath;
+      }
+
+      bool getOverwriteDbFile() const
+      {
+         return overwriteDbFile;
+      }
+
+      const std::string& getTestDatabasePath() const
+      {
+         return testDatabasePath;
+      }
+
+      unsigned getDatabaseNumMaxConns() const
+      {
+         return databaseNumMaxConns;
+      }
+
+      std::string getOverrideRootMDS() const
+      {
+         return overrideRootMDS;
+      }
+
+      bool getReadOnly() const
+      {
+         return readOnly;
+      }
+
+      bool getNoFetch() const
+      {
+         return noFetch;
+      }
+
+      bool getAutomatic() const
+      {
+         return automatic;
+      }
+
+      const std::string& getLogOutFile() const
+      {
+         return logOutFile;
+      }
+
+      bool getRunOffline() const
+      {
+         return runOffline;
+      }
+
+      bool getForceRestart() const
+      {
+         return forceRestart;
+      }
+
+      bool getQuotaEnabled() const
+      {
+         return quotaEnabled;
+      }
+
+      bool getIgnoreDBDiskSpace() const
+      {
+         return ignoreDBDiskSpace;
+      }
+
+      std::bitset<CHECK_FS_ACTIONS_COUNT> getCheckFsActions() const
+      {
+         return checkFsActions;
+      }
+
+      void disableAutomaticRepairMode()
+      {
+         this->automatic = false;
+      }
+
+      void setReadOnly()
+      {
+         this->readOnly = true;
+      }
+};
+
+#endif /*CONFIG_H_*/

fsck/source/components/DataFetcher.cpp

@@ -0,0 +1,203 @@
+#include "DataFetcher.h"
+
+#include <common/Common.h>
+#include <common/storage/Storagedata.h>
+#include <components/worker/RetrieveChunksWork.h>
+#include <components/worker/RetrieveDirEntriesWork.h>
+#include <components/worker/RetrieveFsIDsWork.h>
+#include <components/worker/RetrieveInodesWork.h>
+#include <program/Program.h>
+#include <toolkit/FsckTkEx.h>
+#include <toolkit/FsckException.h>
+
+DataFetcher::DataFetcher(FsckDB& db, bool forceRestart)
+   : database(&db),
+     workQueue(Program::getApp()->getWorkQueue() ),
+     generatedPackages(0),
+     forceRestart(forceRestart)
+{
+}
+
+FhgfsOpsErr DataFetcher::execute()
+{
+   FhgfsOpsErr retVal = FhgfsOpsErr_SUCCESS;
+
+   NodeStore* metaNodeStore = Program::getApp()->getMetaNodes();
+   auto metaNodeList = metaNodeStore->referenceAllNodes();
+
+   printStatus();
+
+   retrieveDirEntries(metaNodeList);
+   retrieveInodes(metaNodeList);
+   const bool retrieveRes = retrieveChunks();
+
+   if (!retrieveRes)
+   {
+      retVal = FhgfsOpsErr_INUSE;
+      Program::getApp()->abort();
+   }
+
+   // wait for all packages to finish, because we cannot proceed if not all data was fetched
+   // BUT : update output each OUTPUT_INTERVAL_MS ms
+   while (!finishedPackages.timedWaitForCount(generatedPackages, DATAFETCHER_OUTPUT_INTERVAL_MS))
+   {
+      printStatus();
+
+      if (retVal != FhgfsOpsErr_INUSE && Program::getApp()->getShallAbort())
+      {
+         // setting retVal to INTERRUPTED
+         // but still, we needed to wait for the workers to terminate, because of the
+         // SynchronizedCounter (this object cannnot be destroyed before all workers terminate)
+         retVal = FhgfsOpsErr_INTERRUPTED;
+      }
+   }
+
+   if (retVal == FhgfsOpsErr_SUCCESS && fatalErrorsFound.read() > 0)
+      retVal = FhgfsOpsErr_INTERNAL;
+
+   if(retVal == FhgfsOpsErr_SUCCESS)
+   {
+      std::set<FsckTargetID> allUsedTargets;
+
+      while(!this->usedTargets.empty() )
+      {
+         allUsedTargets.insert(this->usedTargets.front().begin(), this->usedTargets.front().end() );
+         this->usedTargets.pop_front();
+      }
+
+      std::list<FsckTargetID> usedTargetsList(allUsedTargets.begin(), allUsedTargets.end() );
+
+      this->database->getUsedTargetIDsTable()->insert(usedTargetsList,
+         this->database->getUsedTargetIDsTable()->newBulkHandle() );
+   }
+
+   printStatus(true);
+
+   FsckTkEx::fsckOutput(""); // just a new line
+
+   return retVal;
+}
+
+void DataFetcher::retrieveDirEntries(const std::vector<NodeHandle>& nodes)
+{
+   for (auto nodeIter = nodes.begin(); nodeIter != nodes.end(); nodeIter++)
+   {
+      Node& node = **nodeIter;
+
+      int requestsPerNode = 2;
+
+      unsigned hashDirsPerRequest = (unsigned)(META_DENTRIES_LEVEL1_SUBDIR_NUM/requestsPerNode);
+
+      unsigned hashDirStart = 0;
+      unsigned hashDirEnd = 0;
+
+      do
+      {
+         hashDirEnd = hashDirStart + hashDirsPerRequest;
+
+         // fetch DirEntries
+
+         // before we create a package we increment the generated packages counter
+         this->generatedPackages++;
+         this->usedTargets.insert(this->usedTargets.end(), std::set<FsckTargetID>() );
+
+         this->workQueue->addIndirectWork(
+            new RetrieveDirEntriesWork(database, node, &finishedPackages, fatalErrorsFound,
+               hashDirStart, BEEGFS_MIN(hashDirEnd, META_DENTRIES_LEVEL1_SUBDIR_NUM - 1),
+               &numDentriesFound, &numFileInodesFound, usedTargets.back()));
+
+         // fetch fsIDs
+
+         // before we create a package we increment the generated packages counter
+         this->generatedPackages++;
+
+         this->workQueue->addIndirectWork(
+            new RetrieveFsIDsWork(database, node, &finishedPackages, fatalErrorsFound, hashDirStart,
+               BEEGFS_MIN(hashDirEnd, META_DENTRIES_LEVEL1_SUBDIR_NUM - 1)));
+
+         hashDirStart = hashDirEnd + 1;
+      }  while (hashDirEnd < META_DENTRIES_LEVEL1_SUBDIR_NUM);
+   }
+}
+
+void DataFetcher::retrieveInodes(const std::vector<NodeHandle>& nodes)
+{
+   for (auto nodeIter = nodes.begin(); nodeIter != nodes.end(); nodeIter++)
+   {
+      Node& node = **nodeIter;
+
+      int requestsPerNode = 2;
+
+      unsigned hashDirsPerRequest = (unsigned)(META_INODES_LEVEL1_SUBDIR_NUM/ requestsPerNode);
+
+      unsigned hashDirStart = 0;
+      unsigned hashDirEnd = 0;
+
+      do
+      {
+         // before we create a package we increment the generated packages counter
+         this->generatedPackages++;
+         this->usedTargets.insert(this->usedTargets.end(), std::set<FsckTargetID>() );
+
+         hashDirEnd = hashDirStart + hashDirsPerRequest;
+
+         this->workQueue->addIndirectWork(
+            new RetrieveInodesWork(database, node, &finishedPackages, fatalErrorsFound,
+               hashDirStart, BEEGFS_MIN(hashDirEnd, META_INODES_LEVEL1_SUBDIR_NUM - 1),
+               &numFileInodesFound, &numDirInodesFound, usedTargets.back()));
+
+         hashDirStart = hashDirEnd + 1;
+      } while (hashDirEnd < META_INODES_LEVEL1_SUBDIR_NUM);
+   }
+}
+
+/**
+ * Add the RetrieveChunksWork for each storage node to the work queue.
+ * @returns true if successful, false if work can't be started because another fsck is already
+ *          running or was aborted prematurely
+ */
+bool DataFetcher::retrieveChunks()
+{
+   App* app = Program::getApp();
+
+   NodeStore* storageNodes = app->getStorageNodes();
+
+   // for each server create a work package to retrieve chunks
+   for (const auto& node : storageNodes->referenceAllNodes())
+   {
+      // before we create a package we increment the generated packages counter
+      this->generatedPackages++;
+
+      RetrieveChunksWork* retrieveWork = new RetrieveChunksWork(database, node, &finishedPackages,
+            &numChunksFound, forceRestart);
+
+      // node will be released inside of work package
+      workQueue->addIndirectWork(retrieveWork);
+
+      bool started;
+      retrieveWork->waitForStarted(&started);
+      if (!started)
+         return false;
+   }
+
+   return true;
+}
+
+void DataFetcher::printStatus(bool toLogFile)
+{
+   uint64_t dentryCount = numDentriesFound.read();
+   uint64_t fileInodeCount = numFileInodesFound.read();
+   uint64_t dirInodeCount = numDirInodesFound.read();
+   uint64_t chunkCount = numChunksFound.read();
+
+   std::string outputStr = "Fetched data > Directory entries: " + StringTk::uint64ToStr(dentryCount)
+      + " | Inodes: " + StringTk::uint64ToStr(fileInodeCount+dirInodeCount) + " | Chunks: " +
+      StringTk::uint64ToStr(chunkCount);
+
+   int outputFlags = OutputOptions_LINEDELETE;
+
+   if (!toLogFile)
+      outputFlags = outputFlags | OutputOptions_NOLOG;
+
+   FsckTkEx::fsckOutput(outputStr, outputFlags);
+}

fsck/source/components/DataFetcher.h

@@ -0,0 +1,43 @@
+#ifndef DATAFETCHER_H_
+#define DATAFETCHER_H_
+
+#include <common/components/worker/queue/MultiWorkQueue.h>
+#include <common/nodes/Node.h>
+#include <common/toolkit/SynchronizedCounter.h>
+#include <database/FsckDB.h>
+
+#define DATAFETCHER_OUTPUT_INTERVAL_MS 2000
+
+class DataFetcher
+{
+   private:
+      FsckDB* database;
+
+      MultiWorkQueue* workQueue;
+
+      AtomicUInt64 numDentriesFound;
+      AtomicUInt64 numFileInodesFound;
+      AtomicUInt64 numDirInodesFound;
+      AtomicUInt64 numChunksFound;
+
+   public:
+      DataFetcher(FsckDB& db, bool forceRestart);
+
+      FhgfsOpsErr execute();
+
+   private:
+      SynchronizedCounter finishedPackages;
+      AtomicUInt64 fatalErrorsFound;
+      unsigned generatedPackages;
+      bool forceRestart;
+
+      std::list<std::set<FsckTargetID> > usedTargets;
+
+      void retrieveDirEntries(const std::vector<NodeHandle>& nodes);
+      void retrieveInodes(const std::vector<NodeHandle>& nodes);
+      bool retrieveChunks();
+
+      void printStatus(bool toLogFile = false);
+};
+
+#endif /* DATAFETCHER_H_ */

fsck/source/components/DatagramListener.cpp

@@ -0,0 +1,53 @@
+#include "DatagramListener.h"
+
+#include <common/net/message/NetMessageLogHelper.h>
+
+DatagramListener::DatagramListener(NetFilter* netFilter, NicAddressList& localNicList,
+   AcknowledgmentStore* ackStore, unsigned short udpPort, bool restrictOutboundInterfaces) :
+   AbstractDatagramListener("DGramLis", netFilter, localNicList, ackStore, udpPort,
+      restrictOutboundInterfaces)
+{
+}
+
+DatagramListener::~DatagramListener()
+{
+}
+
+void DatagramListener::handleIncomingMsg(struct sockaddr_in* fromAddr, NetMessage* msg)
+{
+   HighResolutionStats stats; // currently ignored
+   std::shared_ptr<StandardSocket> sock = findSenderSock(fromAddr->sin_addr);
+   if (sock == nullptr)
+   {
+      log.log(Log_WARNING, "Could not handle incoming message: no socket");
+      return;
+   }
+
+   NetMessage::ResponseContext rctx(fromAddr, sock.get(), sendBuf, DGRAMMGR_SENDBUF_SIZE, &stats);
+
+   const auto messageType = netMessageTypeToStr(msg->getMsgType());
+
+   switch(msg->getMsgType() )
+   {
+      // valid messages within this context
+      case NETMSGTYPE_Heartbeat:
+      case NETMSGTYPE_FsckModificationEvent:
+      {
+         if(!msg->processIncoming(rctx) )
+         {
+            LOG(GENERAL, WARNING,
+                  "Problem encountered during handling of incoming message.", messageType);
+         }
+      } break;
+
+      default:
+      { // valid, but not within this context
+         log.logErr(
+            "Received a message that is invalid within the current context "
+            "from: " + Socket::ipaddrToStr(fromAddr->sin_addr) + "; "
+            "type: " + messageType );
+      } break;
+   };
+}
+
+

fsck/source/components/DatagramListener.h

@@ -0,0 +1,22 @@
+#ifndef DATAGRAMLISTENER_H_
+#define DATAGRAMLISTENER_H_
+
+#include <common/components/AbstractDatagramListener.h>
+
+class DatagramListener : public AbstractDatagramListener
+{
+   public:
+      DatagramListener(NetFilter* netFilter, NicAddressList& localNicList,
+         AcknowledgmentStore* ackStore, unsigned short udpPort,
+         bool restrictOutboundInterfaces);
+      virtual ~DatagramListener();
+
+
+   protected:
+      virtual void handleIncomingMsg(struct sockaddr_in* fromAddr, NetMessage* msg);
+
+   private:
+
+};
+
+#endif /*DATAGRAMLISTENER_H_*/

fsck/source/components/InternodeSyncer.cpp

@@ -0,0 +1,491 @@
+#include <app/config/Config.h>
+#include <app/App.h>
+#include <common/toolkit/MessagingTk.h>
+#include <common/toolkit/NodesTk.h>
+#include <common/toolkit/Time.h>
+#include <common/nodes/NodeStore.h>
+#include <common/nodes/TargetCapacityPools.h>
+#include <program/Program.h>
+#include "InternodeSyncer.h"
+
+#include <mutex>
+
+#include <boost/lexical_cast.hpp>
+
+InternodeSyncer::InternodeSyncer() :
+   PThread("XNodeSync"),
+   log("XNodeSync"),
+   serversDownloaded(false)
+{
+}
+
+InternodeSyncer::~InternodeSyncer()
+{
+}
+
+
+void InternodeSyncer::run()
+{
+   try
+   {
+      registerSignalHandler();
+
+      // download all nodes,mappings,states and buddy groups
+      NumNodeIDList addedStorageNodes;
+      NumNodeIDList removedStorageNodes;
+      NumNodeIDList addedMetaNodes;
+      NumNodeIDList removedMetaNodes;
+      bool syncRes = downloadAndSyncNodes(addedStorageNodes, removedStorageNodes, addedMetaNodes,
+         removedMetaNodes);
+      if (!syncRes)
+      {
+         log.logErr("Error downloading nodes from mgmtd.");
+         Program::getApp()->abort();
+         return;
+      }
+
+      syncRes = downloadAndSyncTargetMappings();
+      if (!syncRes)
+      {
+         log.logErr("Error downloading target mappings from mgmtd.");
+         Program::getApp()->abort();
+         return;
+      }
+
+      originalTargetMap = Program::getApp()->getTargetMapper()->getMapping();
+
+      syncRes = downloadAndSyncTargetStates();
+      if (!syncRes)
+      {
+         log.logErr("Error downloading target states from mgmtd.");
+         Program::getApp()->abort();
+         return;
+      }
+
+      syncRes = downloadAndSyncMirrorBuddyGroups();
+      if ( !syncRes )
+      {
+         log.logErr("Error downloading mirror buddy groups from mgmtd.");
+         Program::getApp()->abort();
+         return;
+      }
+
+      Program::getApp()->getMirrorBuddyGroupMapper()->getMirrorBuddyGroups(
+         originalMirrorBuddyGroupMap);
+
+      syncRes = downloadAndSyncMetaMirrorBuddyGroups();
+      if ( !syncRes )
+      {
+         log.logErr("Error downloading metadata mirror buddy groups from mgmtd.");
+         Program::getApp()->abort();
+         return;
+      }
+
+      Program::getApp()->getMetaMirrorBuddyGroupMapper()->getMirrorBuddyGroups(
+         originalMetaMirrorBuddyGroupMap);
+
+      {
+         std::lock_guard<Mutex> lock(serversDownloadedMutex);
+         serversDownloaded = true;
+         serversDownloadedCondition.signal();
+      }
+
+      syncLoop();
+
+      log.log(Log_DEBUG, "Component stopped.");
+   }
+   catch(std::exception& e)
+   {
+      PThread::getCurrentThreadApp()->handleComponentException(e);
+   }
+}
+
+void InternodeSyncer::syncLoop()
+{
+   const int sleepIntervalMS = 3*1000; // 3sec
+   const unsigned downloadNodesAndStatesIntervalMS = 30000; // 30 sec
+   const unsigned checkNetworkIntervalMS = 60*1000; // 1 minute
+
+   Time lastDownloadNodesAndStatesT;
+   Time lastCheckNetworkT;
+
+   while(!waitForSelfTerminateOrder(sleepIntervalMS) )
+   {
+      // download & sync nodes
+      if (lastDownloadNodesAndStatesT.elapsedMS() > downloadNodesAndStatesIntervalMS)
+      {
+         NumNodeIDList addedStorageNodes;
+         NumNodeIDList removedStorageNodes;
+         NumNodeIDList addedMetaNodes;
+         NumNodeIDList removedMetaNodes;
+         bool syncRes = downloadAndSyncNodes(addedStorageNodes, removedStorageNodes, addedMetaNodes,
+            removedMetaNodes);
+
+         if (!syncRes)
+         {
+            log.logErr("Error downloading nodes from mgmtd.");
+            Program::getApp()->abort();
+            break;
+         }
+
+         handleNodeChanges(NODETYPE_Meta, addedMetaNodes, removedMetaNodes);
+
+         handleNodeChanges(NODETYPE_Storage, addedStorageNodes, removedStorageNodes);
+
+         syncRes = downloadAndSyncTargetMappings();
+         if (!syncRes)
+         {
+            log.logErr("Error downloading target mappings from mgmtd.");
+            Program::getApp()->abort();
+            break;
+         }
+
+         handleTargetMappingChanges();
+
+         syncRes = downloadAndSyncTargetStates();
+         if (!syncRes)
+         {
+            log.logErr("Error downloading target states from mgmtd.");
+            Program::getApp()->abort();
+            break;
+         }
+
+         syncRes = downloadAndSyncMirrorBuddyGroups();
+         if ( !syncRes )
+         {
+            log.logErr("Error downloading mirror buddy groups from mgmtd.");
+            Program::getApp()->abort();
+            break;
+         }
+
+         handleBuddyGroupChanges();
+
+         lastDownloadNodesAndStatesT.setToNow();
+      }
+
+      bool checkNetworkForced = getAndResetForceCheckNetwork();
+
+      if( checkNetworkForced ||
+         (lastCheckNetworkT.elapsedMS() > checkNetworkIntervalMS))
+      {
+         checkNetwork();
+         lastCheckNetworkT.setToNow();
+      }
+
+   }
+}
+
+/**
+ * @return false on error.
+ */
+bool InternodeSyncer::downloadAndSyncTargetStates()
+{
+   App* app = Program::getApp();
+   NodeStore* mgmtNodes = app->getMgmtNodes();
+   TargetStateStore* targetStateStore = app->getTargetStateStore();
+
+   auto node = mgmtNodes->referenceFirstNode();
+   if(!node)
+      return false;
+
+   UInt16List targetIDs;
+   UInt8List reachabilityStates;
+   UInt8List consistencyStates;
+
+   bool downloadRes = NodesTk::downloadTargetStates(*node, NODETYPE_Storage,
+      &targetIDs, &reachabilityStates, &consistencyStates, false);
+
+   if(downloadRes)
+      targetStateStore->syncStatesFromLists(targetIDs, reachabilityStates,
+         consistencyStates);
+
+   return downloadRes;
+}
+
+/**
+ * @return false on error
+ */
+bool InternodeSyncer::downloadAndSyncNodes(NumNodeIDList& addedStorageNodes,
+   NumNodeIDList& removedStorageNodes, NumNodeIDList& addedMetaNodes,
+   NumNodeIDList& removedMetaNodes)
+{
+   const char* logContext = "Nodes sync";
+
+   App* app = Program::getApp();
+   NodeStoreServers* mgmtNodes = app->getMgmtNodes();
+   NodeStoreServers* metaNodes = app->getMetaNodes();
+   NodeStoreServers* storageNodes = app->getStorageNodes();
+   Node& localNode = app->getLocalNode();
+
+   auto mgmtNode = mgmtNodes->referenceFirstNode();
+   if (!mgmtNode)
+      return false;
+
+   { // storage nodes
+      std::vector<NodeHandle> storageNodesList;
+
+      bool storageRes =
+         NodesTk::downloadNodes(*mgmtNode, NODETYPE_Storage, storageNodesList, false);
+      if(!storageRes)
+         return false;
+
+      storageNodes->syncNodes(storageNodesList, &addedStorageNodes, &removedStorageNodes,
+         &localNode);
+      printSyncNodesResults(NODETYPE_Storage, &addedStorageNodes, &removedStorageNodes);
+   }
+
+   { // metadata nodes
+      std::vector<NodeHandle> metaNodesList;
+      NumNodeID rootNodeID;
+      bool rootIsBuddyMirrored;
+
+      bool metaRes =
+         NodesTk::downloadNodes(*mgmtNode, NODETYPE_Meta, metaNodesList, false, &rootNodeID,
+            &rootIsBuddyMirrored);
+      if(!metaRes)
+         return false;
+
+      metaNodes->syncNodes(metaNodesList, &addedMetaNodes, &removedMetaNodes, &localNode);
+
+      if (app->getMetaRoot().setIfDefault(rootNodeID, rootIsBuddyMirrored))
+      {
+         LogContext(logContext).log(Log_CRITICAL,
+            "Root NodeID (from sync results): " + rootNodeID.str() );
+      }
+
+      printSyncNodesResults(NODETYPE_Meta, &addedMetaNodes, &removedMetaNodes);
+   }
+
+   return true;
+}
+
+void InternodeSyncer::printSyncNodesResults(NodeType nodeType, NumNodeIDList* addedNodes,
+   NumNodeIDList* removedNodes)
+{
+   const char* logContext = "Sync results";
+
+   if (!addedNodes->empty())
+      LogContext(logContext).log(Log_WARNING, std::string("Nodes added: ") +
+         StringTk::uintToStr(addedNodes->size() ) +
+         " (Type: " + boost::lexical_cast<std::string>(nodeType) + ")");
+
+   if (!removedNodes->empty())
+      LogContext(logContext).log(Log_WARNING, std::string("Nodes removed: ") +
+         StringTk::uintToStr(removedNodes->size() ) +
+         " (Type: " + boost::lexical_cast<std::string>(nodeType) + ")");
+}
+
+/**
+ * @return false on error
+ */
+bool InternodeSyncer::downloadAndSyncTargetMappings()
+{
+   App* app = Program::getApp();
+   NodeStoreServers* mgmtNodes = app->getMgmtNodes();
+   TargetMapper* targetMapper = app->getTargetMapper();
+
+   bool retVal = true;
+
+   auto mgmtNode = mgmtNodes->referenceFirstNode();
+   if(!mgmtNode)
+      return false;
+
+   auto mappings = NodesTk::downloadTargetMappings(*mgmtNode, false);
+   if (mappings.first)
+      targetMapper->syncTargets(std::move(mappings.second));
+   else
+      retVal = false;
+
+   return retVal;
+}
+
+/**
+ * @return false on error
+ */
+bool InternodeSyncer::downloadAndSyncMirrorBuddyGroups()
+{
+   App* app = Program::getApp();
+   NodeStoreServers* mgmtNodes = app->getMgmtNodes();
+   MirrorBuddyGroupMapper* buddyGroupMapper = app->getMirrorBuddyGroupMapper();
+
+   bool retVal = true;
+
+   auto mgmtNode = mgmtNodes->referenceFirstNode();
+   if(!mgmtNode)
+      return false;
+
+   UInt16List buddyGroupIDs;
+   UInt16List primaryTargetIDs;
+   UInt16List secondaryTargetIDs;
+
+   bool downloadRes = NodesTk::downloadMirrorBuddyGroups(*mgmtNode, NODETYPE_Storage,
+         &buddyGroupIDs, &primaryTargetIDs, &secondaryTargetIDs, false);
+
+   if(downloadRes)
+      buddyGroupMapper->syncGroupsFromLists(buddyGroupIDs, primaryTargetIDs, secondaryTargetIDs,
+         NumNodeID() );
+   else
+      retVal = false;
+
+   return retVal;
+}
+
+/**
+ * @return false on error
+ */
+bool InternodeSyncer::downloadAndSyncMetaMirrorBuddyGroups()
+{
+   App* app = Program::getApp();
+   NodeStoreServers* mgmtNodes = app->getMgmtNodes();
+   MirrorBuddyGroupMapper* buddyGroupMapper = app->getMetaMirrorBuddyGroupMapper();
+
+   bool retVal = true;
+
+   auto mgmtNode = mgmtNodes->referenceFirstNode();
+   if(!mgmtNode)
+      return false;
+
+   UInt16List buddyGroupIDs;
+   UInt16List primaryTargetIDs;
+   UInt16List secondaryTargetIDs;
+
+   bool downloadRes = NodesTk::downloadMirrorBuddyGroups(*mgmtNode, NODETYPE_Meta, &buddyGroupIDs,
+      &primaryTargetIDs, &secondaryTargetIDs, false);
+
+   if(downloadRes)
+      buddyGroupMapper->syncGroupsFromLists(buddyGroupIDs, primaryTargetIDs, secondaryTargetIDs,
+         NumNodeID() );
+   else
+      retVal = false;
+
+   return retVal;
+}
+
+void InternodeSyncer::handleNodeChanges(NodeType nodeType, NumNodeIDList& addedNodes,
+   NumNodeIDList& removedNodes)
+{
+   const char* logContext = "handleNodeChanges";
+
+   if (!addedNodes.empty())
+      LogContext(logContext).log(Log_WARNING,
+         std::string("Nodes added while beegfs-fsck was running: ")
+            + StringTk::uintToStr(addedNodes.size())
+            + " (Type: " + boost::lexical_cast<std::string>(nodeType) + ")");
+
+   if (!removedNodes.empty())
+   {
+      // removed nodes must lead to fsck stoppage
+      LogContext(logContext).log(Log_WARNING,
+         std::string("Nodes removed while beegfs-fsck was running: ")
+            + StringTk::uintToStr(removedNodes.size())
+            + " (Type: " + boost::lexical_cast<std::string>(nodeType) + ")");
+
+      Program::getApp()->abort();
+   }
+}
+
+void InternodeSyncer::handleTargetMappingChanges()
+{
+   const char* logContext = "handleTargetMappingChanges";
+
+   TargetMap newTargetMap = Program::getApp()->getTargetMapper()->getMapping();
+
+   for ( TargetMapIter originalMapIter = originalTargetMap.begin();
+      originalMapIter != originalTargetMap.end(); originalMapIter++ )
+   {
+      uint16_t targetID = originalMapIter->first;
+      NumNodeID oldNodeID = originalMapIter->second;
+
+      TargetMapIter newMapIter = newTargetMap.find(targetID);
+
+      if ( newMapIter == newTargetMap.end() )
+      {
+         LogContext(logContext).log(Log_WARNING,
+            "Target removed while beegfs-fsck was running; beegfs-fsck can't continue; targetID: "
+               + StringTk::uintToStr(targetID));
+
+         Program::getApp()->abort();
+      }
+      else
+      {
+         NumNodeID newNodeID = newMapIter->second;
+         if ( oldNodeID != newNodeID )
+         {
+            LogContext(logContext).log(Log_WARNING,
+               "Target re-mapped while beegfs-fsck was running; beegfs-fsck can't continue; "
+                  "targetID: " + StringTk::uintToStr(targetID));
+
+            Program::getApp()->abort();
+         }
+      }
+   }
+}
+
+void InternodeSyncer::handleBuddyGroupChanges()
+{
+   const char* logContext = "handleBuddyGroupChanges";
+
+   MirrorBuddyGroupMap newMirrorBuddyGroupMap;
+   Program::getApp()->getMirrorBuddyGroupMapper()->getMirrorBuddyGroups(newMirrorBuddyGroupMap);
+
+   for ( MirrorBuddyGroupMapIter originalMapIter = originalMirrorBuddyGroupMap.begin();
+      originalMapIter != originalMirrorBuddyGroupMap.end(); originalMapIter++ )
+   {
+      uint16_t buddyGroupID = originalMapIter->first;
+      MirrorBuddyGroup oldBuddyGroup = originalMapIter->second;
+
+      MirrorBuddyGroupMapIter newMapIter = newMirrorBuddyGroupMap.find(buddyGroupID);
+
+      if ( newMapIter == newMirrorBuddyGroupMap.end() )
+      {
+         LogContext(logContext).log(Log_WARNING,
+            "Mirror buddy group removed while beegfs-fsck was running; beegfs-fsck can't continue; "
+            "groupID: " + StringTk::uintToStr(buddyGroupID));
+
+         Program::getApp()->abort();
+      }
+      else
+      {
+         MirrorBuddyGroup newBuddyGroup = newMapIter->second;
+         if ( oldBuddyGroup.firstTargetID != newBuddyGroup.firstTargetID )
+         {
+            LogContext(logContext).log(Log_WARNING,
+               "Primary of mirror buddy group changed while beegfs-fsck was running; beegfs-fsck "
+               "can't continue; groupID: " + StringTk::uintToStr(buddyGroupID));
+
+            Program::getApp()->abort();
+         }
+      }
+   }
+}
+
+/**
+ * Blocks until list of servers has been downloaded from management node
+ */
+void InternodeSyncer::waitForServers()
+{
+   std::lock_guard<Mutex> lock(serversDownloadedMutex);
+   while (!serversDownloaded)
+      serversDownloadedCondition.wait(&serversDownloadedMutex);
+}
+
+/**
+ * Inspect the available and allowed network interfaces for any changes.
+ */
+bool InternodeSyncer::checkNetwork()
+{
+   App* app = Program::getApp();
+   NicAddressList newLocalNicList;
+   bool res = false;
+
+   app->findAllowedInterfaces(newLocalNicList);
+   app->findAllowedRDMAInterfaces(newLocalNicList);
+   if (!std::equal(newLocalNicList.begin(), newLocalNicList.end(), app->getLocalNicList().begin()))
+   {
+      log.log(Log_NOTICE, "checkNetwork: local interfaces have changed");
+      app->updateLocalNicList(newLocalNicList);
+      res = true;
+   }
+
+   return res;
+}

fsck/source/components/InternodeSyncer.h

@@ -0,0 +1,95 @@
+#ifndef INTERNODESYNCER_H_
+#define INTERNODESYNCER_H_
+
+#include <common/app/log/LogContext.h>
+#include <common/components/ComponentInitException.h>
+#include <common/nodes/MirrorBuddyGroupMapper.h>
+#include <common/threading/PThread.h>
+#include <common/Common.h>
+
+#include <mutex>
+
+class InternodeSyncer : public PThread
+{
+   public:
+      InternodeSyncer();
+      virtual ~InternodeSyncer();
+
+      bool downloadAndSyncNodes(NumNodeIDList& addedStorageNodes,
+         NumNodeIDList& removedStorageNodes, NumNodeIDList& addedMetaNodes,
+         NumNodeIDList& removedMetaNodes);
+      bool downloadAndSyncTargetMappings();
+      bool downloadAndSyncMirrorBuddyGroups();
+      bool downloadAndSyncMetaMirrorBuddyGroups();
+      bool downloadAndSyncTargetStates();
+
+   private:
+      LogContext log;
+      Mutex forceNodesAndTargetStatesUpdateMutex;
+      bool forceNodesAndTargetStatesUpdate;
+      Mutex forceCheckNetworkMutex;
+      bool forceCheckNetwork;  // true to force check of network interfaces
+
+      TargetMap originalTargetMap;
+      MirrorBuddyGroupMap originalMirrorBuddyGroupMap;
+      MirrorBuddyGroupMap originalMetaMirrorBuddyGroupMap;
+
+      virtual void run();
+      void syncLoop();
+      void handleNodeChanges(NodeType nodeType, NumNodeIDList& addedNodes,
+         NumNodeIDList& removedNodes);
+      void handleTargetMappingChanges();
+      void handleBuddyGroupChanges();
+
+      bool getAndResetForceCheckNetwork()
+      {
+         const std::lock_guard<Mutex> lock(forceCheckNetworkMutex);
+
+         bool retVal = this->forceCheckNetwork;
+
+         this->forceCheckNetwork = false;
+
+         return retVal;
+      }
+
+      bool checkNetwork();
+
+      Condition serversDownloadedCondition;
+      Mutex serversDownloadedMutex;
+      bool serversDownloaded;
+
+   public:
+      void waitForServers();
+
+      void setForceCheckNetwork()
+      {
+         const std::lock_guard<Mutex> lock(forceCheckNetworkMutex);
+
+         this->forceCheckNetwork = true;
+      }
+
+   private:
+      static void printSyncNodesResults(NodeType nodeType, NumNodeIDList* addedNodes,
+         NumNodeIDList* removedNodes);
+
+      void setForceNodesAndTargetStatesUpdate()
+      {
+         const std::lock_guard<Mutex> lock(forceNodesAndTargetStatesUpdateMutex);
+
+         this->forceNodesAndTargetStatesUpdate = true;
+      }
+
+      bool getAndResetForceNodesAndTargetStatesUpdate()
+      {
+         const std::lock_guard<Mutex> lock(forceNodesAndTargetStatesUpdateMutex);
+
+         bool retVal = this->forceNodesAndTargetStatesUpdate;
+
+         this->forceNodesAndTargetStatesUpdate = false;
+
+         return retVal;
+      }
+};
+
+
+#endif /* INTERNODESYNCER_H_ */

fsck/source/components/ModificationEventHandler.cpp

@@ -0,0 +1,98 @@
+#include "ModificationEventHandler.h"
+
+#include <common/toolkit/ZipIterator.h>
+#include <database/FsckDBException.h>
+
+#include <program/Program.h>
+
+#include <mutex>
+
+ModificationEventHandler::ModificationEventHandler(FsckDBModificationEventsTable& table)
+   : PThread("ModificationEventHandler"),
+     table(&table)
+{
+}
+
+void ModificationEventHandler::run()
+{
+   FsckDBModificationEventsTable::BulkHandle bulkHandle(table->newBulkHandle() );
+   while ( !getSelfTerminate() )
+   {
+      std::unique_lock<Mutex> bufferListSafeLock(bufferListMutex); // LOCK BUFFER
+
+      // make sure to group at least MODHANDLER_MINSIZE_FLUSH flush elements (to not bother the DB
+      // with every single event)
+      if (bufferList.size() < MODHANDLER_MINSIZE_FLUSH)
+      {
+         bufferListSafeLock.unlock(); // UNLOCK BUFFER
+         const std::lock_guard<Mutex> lock(eventsAddedMutex);
+         eventsAddedCond.timedwait(&eventsAddedMutex, 2000);
+         continue;
+      }
+      else
+      {
+         // create a copy of the buffer list and flush this to DB, so that the buffer will become
+         // free immediately and the incoming messages do not have to wait for DB
+         FsckModificationEventList bufferListCopy;
+
+         bufferListCopy.splice(bufferListCopy.begin(), bufferList);
+
+         bufferListSafeLock.unlock(); // UNLOCK BUFFER
+
+         table->insert(bufferListCopy, bulkHandle);
+      }
+   }
+
+   // a last flush after component stopped
+   FsckModificationEventList bufferListCopy;
+
+   {
+      const std::lock_guard<Mutex> bufferListLock(bufferListMutex);
+      bufferListCopy.splice(bufferListCopy.begin(), bufferList);
+   }
+
+   table->insert(bufferListCopy, bulkHandle);
+}
+
+bool ModificationEventHandler::add(UInt8List& eventTypeList, StringList& entryIDList)
+{
+   const char* logContext = "ModificationEventHandler (add)";
+
+   if ( unlikely(eventTypeList.size() != entryIDList.size()) )
+   {
+      LogContext(logContext).logErr("Unable to add events. The lists do not have equal sizes.");
+      return false;
+   }
+
+   while ( true )
+   {
+      {
+         const std::lock_guard<Mutex> lock(bufferListMutex);
+         if (this->bufferList.size() < MODHANDLER_MAXSIZE_EVENTLIST)
+         {
+            break;
+         }
+      }
+      {
+         const std::lock_guard<Mutex> lock(eventsFlushedMutex);
+         this->eventsFlushedCond.timedwait(&eventsFlushedMutex, 2000);
+      }
+   }
+
+   ZipIterRange<UInt8List, StringList> eventTypeEntryIDIter(eventTypeList, entryIDList);
+
+   {
+      const std::lock_guard<Mutex> bufferListLock(bufferListMutex);
+
+      for ( ; !eventTypeEntryIDIter.empty(); ++eventTypeEntryIDIter)
+      {
+         FsckModificationEvent event((ModificationEventType)*(eventTypeEntryIDIter()->first),
+            *(eventTypeEntryIDIter()->second) );
+         this->bufferList.push_back(event);
+      }
+   }
+
+   this->eventsAddedCond.signal();
+
+   return true;
+}

fsck/source/components/ModificationEventHandler.h

@@ -0,0 +1,44 @@
+#ifndef MODIFICATIONEVENTHANDLER_H
+#define MODIFICATIONEVENTHANDLER_H
+
+#include <common/fsck/FsckModificationEvent.h>
+#include <common/threading/Condition.h>
+#include <database/FsckDBTable.h>
+
+
+#define MODHANDLER_MAXSIZE_EVENTLIST      50000
+#define MODHANDLER_MINSIZE_FLUSH          200
+
+class ModificationEventHandler: public PThread
+{
+   public:
+      ModificationEventHandler(FsckDBModificationEventsTable& table);
+
+      virtual void run();
+
+      bool add(UInt8List& eventTypeList, StringList& entryIDList);
+
+   private:
+      FsckDBModificationEventsTable* table;
+
+      FsckModificationEventList bufferList;
+
+      Mutex bufferListMutex;
+      Mutex bufferListCopyMutex;
+      Mutex flushMutex;
+
+      Mutex eventsAddedMutex;
+      Condition eventsAddedCond;
+      Mutex eventsFlushedMutex;
+      Condition eventsFlushedCond;
+
+   public:
+      void stop()
+      {
+         selfTerminate();
+         eventsAddedCond.signal();
+         join();
+      }
+};
+
+#endif /* MODIFICATIONEVENTHANDLER_H */

fsck/source/components/worker/AdjustChunkPermissionsWork.cpp

@@ -0,0 +1,97 @@
+#include "AdjustChunkPermissionsWork.h"
+#include <common/net/message/fsck/AdjustChunkPermissionsMsg.h>
+#include <common/net/message/fsck/AdjustChunkPermissionsRespMsg.h>
+#include <common/toolkit/MessagingTk.h>
+#include <toolkit/FsckException.h>
+#include <program/Program.h>
+
+AdjustChunkPermissionsWork::AdjustChunkPermissionsWork(Node& node, SynchronizedCounter* counter,
+   AtomicUInt64* fileCount, AtomicUInt64* errorCount)
+    : log("AdjustChunkPermissionsWork"),
+      node(node),
+      counter(counter),
+      fileCount(fileCount),
+      errorCount(errorCount)
+{
+}
+
+AdjustChunkPermissionsWork::~AdjustChunkPermissionsWork()
+{
+}
+
+void AdjustChunkPermissionsWork::process(char* bufIn, unsigned bufInLen, char* bufOut,
+   unsigned bufOutLen)
+{
+   log.log(4, "Processing AdjustChunkPermissionsWork");
+
+   try
+   {
+      doWork(false);
+      doWork(true);
+      // work package finished => increment counter
+      this->counter->incCount();
+   }
+   catch (std::exception &e)
+   {
+      // exception thrown, but work package is finished => increment counter
+      this->counter->incCount();
+
+      // after incrementing counter, re-throw exception
+      throw;
+   }
+
+   log.log(4, "Processed AdjustChunkPermissionsWork");
+}
+
+void AdjustChunkPermissionsWork::doWork(bool isBuddyMirrored)
+{
+   for ( unsigned firstLevelhashDirNum = 0;
+      firstLevelhashDirNum <= META_DENTRIES_LEVEL1_SUBDIR_NUM - 1; firstLevelhashDirNum++ )
+   {
+      for ( unsigned secondLevelhashDirNum = 0;
+         secondLevelhashDirNum < META_DENTRIES_LEVEL2_SUBDIR_NUM; secondLevelhashDirNum++ )
+      {
+         unsigned hashDirNum = StorageTk::mergeHashDirs(firstLevelhashDirNum,
+            secondLevelhashDirNum);
+
+         int64_t hashDirOffset = 0;
+         int64_t contDirOffset = 0;
+         std::string currentContDirID;
+         unsigned resultCount = 0;
+
+         do
+         {
+            AdjustChunkPermissionsMsg adjustChunkPermissionsMsg(hashDirNum, currentContDirID,
+               ADJUST_AT_ONCE, hashDirOffset, contDirOffset, isBuddyMirrored);
+
+            const auto respMsg = MessagingTk::requestResponse(node, adjustChunkPermissionsMsg,
+               NETMSGTYPE_AdjustChunkPermissionsResp);
+
+            if (respMsg)
+            {
+               auto* adjustChunkPermissionsRespMsg = (AdjustChunkPermissionsRespMsg*) respMsg.get();
+
+               // set new parameters
+               currentContDirID = adjustChunkPermissionsRespMsg->getCurrentContDirID();
+               hashDirOffset = adjustChunkPermissionsRespMsg->getNewHashDirOffset();
+               contDirOffset = adjustChunkPermissionsRespMsg->getNewContDirOffset();
+               resultCount = adjustChunkPermissionsRespMsg->getCount();
+
+               this->fileCount->increase(resultCount);
+
+               if (adjustChunkPermissionsRespMsg->getErrorCount() > 0)
+                  this->errorCount->increase(adjustChunkPermissionsRespMsg->getErrorCount());
+            }
+            else
+            {
+               throw FsckException("Communication error occured with node " + node.getAlias());
+            }
+
+            // if any of the worker threads threw an exception, we should stop now!
+            if ( Program::getApp()->getSelfTerminate() )
+               return;
+
+         } while ( resultCount > 0 );
+      }
+   }
+}

fsck/source/components/worker/AdjustChunkPermissionsWork.h

@@ -0,0 +1,32 @@
+#ifndef ADJUSTCHUNKPERMISSIONSWORK_H
+#define ADJUSTCHUNKPERMISSIONSWORK_H
+
+#include <common/app/log/LogContext.h>
+#include <common/components/worker/Work.h>
+#include <common/toolkit/SynchronizedCounter.h>
+
+#include <database/FsckDB.h>
+
+// the size of one packet, i.e. how many files are adjusted at once; basically just
+// limited to have some control and give the user feedback
+#define ADJUST_AT_ONCE 50
+
+class AdjustChunkPermissionsWork : public Work
+{
+   public:
+      AdjustChunkPermissionsWork(Node& node, SynchronizedCounter* counter, AtomicUInt64* fileCount,
+         AtomicUInt64* errorCount);
+      virtual ~AdjustChunkPermissionsWork();
+      void process(char* bufIn, unsigned bufInLen, char* bufOut, unsigned bufOutLen);
+
+   private:
+      LogContext log;
+      Node& node;
+      SynchronizedCounter* counter;
+      AtomicUInt64* fileCount;
+      AtomicUInt64* errorCount;
+
+      void doWork(bool isBuddyMirrored);
+};
+
+#endif /* ADJUSTCHUNKPERMISSIONSWORK_H */

fsck/source/components/worker/RetrieveChunksWork.cpp

@@ -0,0 +1,152 @@
+#include "RetrieveChunksWork.h"
+
+#include <common/net/message/fsck/FetchFsckChunkListMsg.h>
+#include <common/net/message/fsck/FetchFsckChunkListRespMsg.h>
+#include <common/storage/Storagedata.h>
+#include <common/toolkit/FsckTk.h>
+#include <common/toolkit/MessagingTk.h>
+#include <common/toolkit/StorageTk.h>
+
+#include <database/FsckDBException.h>
+#include <toolkit/FsckException.h>
+#include <program/Program.h>
+
+RetrieveChunksWork::RetrieveChunksWork(FsckDB* db, NodeHandle node, SynchronizedCounter* counter,
+      AtomicUInt64* numChunksFound, bool forceRestart) :
+   log("RetrieveChunksWork"), node(std::move(node)), counter(counter),
+   numChunksFound(numChunksFound),
+   chunks(db->getChunksTable()), chunksHandle(chunks->newBulkHandle()),
+   malformedChunks(db->getMalformedChunksList()),
+   forceRestart(forceRestart),
+   started(false), startedBarrier(2)
+{
+}
+
+RetrieveChunksWork::~RetrieveChunksWork()
+{
+}
+
+void RetrieveChunksWork::process(char* bufIn, unsigned bufInLen, char* bufOut, unsigned bufOutLen)
+{
+   log.log(Log_DEBUG, "Processing RetrieveChunksWork");
+
+   try
+   {
+      doWork();
+      // flush buffers before signaling completion
+      chunks->flush(chunksHandle);
+      // work package finished => increment counter
+      this->counter->incCount();
+
+   }
+   catch (std::exception& e)
+   {
+      // exception thrown, but work package is finished => increment counter
+      this->counter->incCount();
+
+      // after incrementing counter, re-throw exception
+      throw;
+   }
+
+   log.log(Log_DEBUG, "Processed RetrieveChunksWork");
+}
+
+void RetrieveChunksWork::doWork()
+{
+   // take the node associated with the current target and send a RetrieveChunksMsg to
+   // that node; the chunks are retrieved incrementally
+   if ( node )
+   {
+      std::string nodeID = node->getAlias();
+      FetchFsckChunkListStatus status = FetchFsckChunkListStatus_NOTSTARTED;
+      unsigned resultCount = 0;
+
+      do
+      {
+         FetchFsckChunkListMsg fetchFsckChunkListMsg(RETRIEVE_CHUNKS_PACKET_SIZE, status,
+               forceRestart);
+
+         const auto respMsg = MessagingTk::requestResponse(*node, fetchFsckChunkListMsg,
+            NETMSGTYPE_FetchFsckChunkListResp);
+
+         if (respMsg)
+         {
+            auto* fetchFsckChunkListRespMsg = (FetchFsckChunkListRespMsg*) respMsg.get();
+
+            FsckChunkList& chunks = fetchFsckChunkListRespMsg->getChunkList();
+            resultCount = chunks.size();
+
+            status = fetchFsckChunkListRespMsg->getStatus();
+
+            // check entry IDs
+            for (auto it = chunks.begin(); it != chunks.end(); )
+            {
+               if (db::EntryID::tryFromStr(it->getID()).first
+                     && it->getSavedPath()->str().size() <= db::Chunk::SAVED_PATH_SIZE)
+               {
+                  ++it;
+                  continue;
+               }
+
+               ++it;
+               malformedChunks->append(*std::prev(it));
+               chunks.erase(std::prev(it));
+            }
+
+            if (status == FetchFsckChunkListStatus_NOTSTARTED)
+            {
+               // Another fsck run is still in progress or was aborted, and --forceRestart was not
+               // set - this means we can't start a new chunk fetcher.
+               started = false;
+               startedBarrier.wait();
+               startedBarrier.wait();
+               return;
+            }
+            else if (status == FetchFsckChunkListStatus_READERROR)
+            {
+               throw FsckException("Read error occured while fetching chunks from node; nodeID: "
+                  + nodeID);
+            }
+
+            if (!started)
+            {
+               started = true;
+               startedBarrier.wait();
+               startedBarrier.wait();
+            }
+
+            this->chunks->insert(chunks, this->chunksHandle);
+
+            numChunksFound->increase(resultCount);
+         }
+         else
+         {
+            throw FsckException("Communication error occured with node; nodeID: " + nodeID);
+         }
+
+         if ( Program::getApp()->getShallAbort() )
+            break;
+
+      } while ( (resultCount > 0) || (status == FetchFsckChunkListStatus_RUNNING) );
+   }
+   else
+   {
+      // basically this should never ever happen
+      log.logErr("Requested node does not exist");
+      throw FsckException("Requested node does not exist");
+   }
+}
+
+/**
+ * Waits until started conditin is is signalled and returns the value of stared
+ * @param isStarted ptr to boolean which is set to whether the server replied it actually started
+ *                  the process. Note: this is a ptr and not a return to ensure the member variable
+ *                  started is no longer accessed after doWork is finished and the object possibly
+ *                  already deleted.
+ */
+void RetrieveChunksWork::waitForStarted(bool* isStarted)
+{
+   startedBarrier.wait();
+   *isStarted = started;
+   startedBarrier.wait();
+}

fsck/source/components/worker/RetrieveChunksWork.h

@@ -0,0 +1,57 @@
+#ifndef RETRIEVECHUNKSWORK_H
+#define RETRIEVECHUNKSWORK_H
+
+/*
+ * retrieve all chunks from one storage server and save them to DB
+ */
+
+#include <common/app/log/LogContext.h>
+#include <common/components/worker/Work.h>
+#include <common/threading/Barrier.h>
+#include <common/toolkit/SynchronizedCounter.h>
+#include <database/FsckDB.h>
+#include <database/FsckDBTable.h>
+
+// the size of one response packet, i.e. how many chunks are asked for at once
+#define RETRIEVE_CHUNKS_PACKET_SIZE 400
+
+class RetrieveChunksWork : public Work
+{
+   public:
+      /*
+       * @param db database instance
+       * @param node pointer to the node to retrieve data from
+       * @param counter a pointer to a Synchronized counter; this is incremented by one at the end
+       * and the calling thread can wait for the counter
+       * @param numChunksFound
+       * @param forceRestart In case the storage servers' chunk fetchers still have data from a
+       *                     previous run, force a restart instead of aborting with an error
+       */
+      RetrieveChunksWork(FsckDB* db, NodeHandle node, SynchronizedCounter* counter,
+         AtomicUInt64* numChunksFound, bool forceRestart);
+      virtual ~RetrieveChunksWork();
+      void process(char* bufIn, unsigned bufInLen, char* bufOut, unsigned bufOutLen);
+
+      void waitForStarted(bool* isStarted);
+
+
+   private:
+      LogContext log;
+      NodeHandle node;
+      SynchronizedCounter* counter;
+      AtomicUInt64* numChunksFound;
+
+      FsckDBChunksTable* chunks;
+      FsckDBChunksTable::BulkHandle chunksHandle;
+
+      DiskList<FsckChunk>* malformedChunks;
+
+      bool forceRestart;
+
+      void doWork();
+
+      bool started;
+      Barrier startedBarrier;
+};
+
+#endif /* RETRIEVECHUNKSWORK_H */

fsck/source/components/worker/RetrieveDirEntriesWork.cpp

@@ -0,0 +1,222 @@
+#include "RetrieveDirEntriesWork.h"
+#include <common/net/message/fsck/RetrieveDirEntriesMsg.h>
+#include <common/net/message/fsck/RetrieveDirEntriesRespMsg.h>
+#include <common/toolkit/MessagingTk.h>
+#include <common/toolkit/MetaStorageTk.h>
+#include <database/FsckDBException.h>
+#include <toolkit/FsckException.h>
+
+#include <program/Program.h>
+
+#include <set>
+
+RetrieveDirEntriesWork::RetrieveDirEntriesWork(FsckDB* db, Node& node, SynchronizedCounter* counter,
+      AtomicUInt64& errors, unsigned hashDirStart, unsigned hashDirEnd,
+      AtomicUInt64* numDentriesFound, AtomicUInt64* numFileInodesFound,
+      std::set<FsckTargetID>& usedTargets) :
+   log("RetrieveDirEntriesWork"), node(node), counter(counter), errors(&errors),
+   numDentriesFound(numDentriesFound), numFileInodesFound(numFileInodesFound),
+   usedTargets(&usedTargets), hashDirStart(hashDirStart), hashDirEnd(hashDirEnd),
+   dentries(db->getDentryTable()), dentriesHandle(dentries->newBulkHandle()),
+   files(db->getFileInodesTable()), filesHandle(files->newBulkHandle()),
+   contDirs(db->getContDirsTable()), contDirsHandle(contDirs->newBulkHandle())
+{
+}
+
+void RetrieveDirEntriesWork::process(char* bufIn, unsigned bufInLen, char* bufOut,
+   unsigned bufOutLen)
+{
+   log.log(4, "Processing RetrieveDirEntriesWork");
+
+   try
+   {
+      doWork(false);
+      doWork(true);
+      // flush buffers before signaling completion
+      dentries->flush(dentriesHandle);
+      files->flush(filesHandle);
+      contDirs->flush(contDirsHandle);
+      // work package finished => increment counter
+      this->counter->incCount();
+   }
+   catch (std::exception &e)
+   {
+      // exception thrown, but work package is finished => increment counter
+      this->counter->incCount();
+
+      // after incrementing counter, re-throw exception
+      throw;
+   }
+
+   log.log(4, "Processed RetrieveDirEntriesWork");
+}
+
+void RetrieveDirEntriesWork::doWork(bool isBuddyMirrored)
+{
+   for ( unsigned firstLevelhashDirNum = hashDirStart; firstLevelhashDirNum <= hashDirEnd;
+      firstLevelhashDirNum++ )
+   {
+      for ( unsigned secondLevelhashDirNum = 0;
+         secondLevelhashDirNum < META_DENTRIES_LEVEL2_SUBDIR_NUM; secondLevelhashDirNum++ )
+      {
+         unsigned hashDirNum = StorageTk::mergeHashDirs(firstLevelhashDirNum,
+            secondLevelhashDirNum);
+
+         int64_t hashDirOffset = 0;
+         int64_t contDirOffset = 0;
+         std::string currentContDirID;
+         int resultCount = 0;
+
+         do
+         {
+            RetrieveDirEntriesMsg retrieveDirEntriesMsg(hashDirNum, currentContDirID,
+               RETRIEVE_DIR_ENTRIES_PACKET_SIZE, hashDirOffset, contDirOffset, isBuddyMirrored);
+
+            const auto respMsg = MessagingTk::requestResponse(node, retrieveDirEntriesMsg,
+               NETMSGTYPE_RetrieveDirEntriesResp);
+
+            if (respMsg)
+            {
+               auto* retrieveDirEntriesRespMsg = (RetrieveDirEntriesRespMsg*) respMsg.get();
+
+               // set new parameters
+               currentContDirID = retrieveDirEntriesRespMsg->getCurrentContDirID();
+               hashDirOffset = retrieveDirEntriesRespMsg->getNewHashDirOffset();
+               contDirOffset = retrieveDirEntriesRespMsg->getNewContDirOffset();
+
+               // parse directory entries
+               FsckDirEntryList& dirEntries = retrieveDirEntriesRespMsg->getDirEntries();
+               // this is the actual result count we are interested in, because if no dirEntries
+               // were read, there is nothing left on the server
+
+               resultCount = dirEntries.size();
+
+               // check dentry entry IDs
+               for (auto it = dirEntries.begin(); it != dirEntries.end(); )
+               {
+                  if (db::EntryID::tryFromStr(it->getID()).first
+                        && db::EntryID::tryFromStr(it->getParentDirID()).first)
+                  {
+                     ++it;
+                     continue;
+                  }
+
+                  LOG(GENERAL, ERR, "Found dentry with invalid entry IDs.",
+                        ("node", it->getSaveNodeID()),
+                        ("isBuddyMirrored", it->getIsBuddyMirrored()),
+                        ("entryID", it->getID()),
+                        ("parentEntryID", it->getParentDirID()));
+
+                  ++it;
+                  errors->increase();
+                  dirEntries.erase(std::prev(it));
+               }
+
+               this->dentries->insert(dirEntries, this->dentriesHandle);
+
+               numDentriesFound->increase(resultCount);
+
+               // parse inlined file inodes
+               FsckFileInodeList& inlinedFileInodes =
+                  retrieveDirEntriesRespMsg->getInlinedFileInodes();
+
+               // check inode entry IDs
+               for (auto it = inlinedFileInodes.begin(); it != inlinedFileInodes.end(); )
+               {
+                  if (db::EntryID::tryFromStr(it->getID()).first
+                        && db::EntryID::tryFromStr(it->getParentDirID()).first
+                        && (!it->getPathInfo()->hasOrigFeature()
+                              || db::EntryID::tryFromStr(
+                                    it->getPathInfo()->getOrigParentEntryID()).first))
+                  {
+                     ++it;
+                     continue;
+                  }
+
+                  LOG(GENERAL, ERR, "Found inode with invalid entry IDs.",
+                        ("node", it->getSaveNodeID()),
+                        ("isBuddyMirrored", it->getIsBuddyMirrored()),
+                        ("entryID", it->getID()),
+                        ("parentEntryID", it->getParentDirID()),
+                        ("origParent", it->getPathInfo()->getOrigParentEntryID()));
+
+                  ++it;
+                  errors->increase();
+                  inlinedFileInodes.erase(std::prev(it));
+               }
+
+               struct ops
+               {
+                  static bool dentryCmp(const FsckDirEntry& a, const FsckDirEntry& b)
+                  {
+                     return a.getID() < b.getID();
+                  }
+
+                  static bool inodeCmp(const FsckFileInode& a, const FsckFileInode& b)
+                  {
+                     return a.getID() < b.getID();
+                  }
+               };
+
+               dirEntries.sort(ops::dentryCmp);
+               inlinedFileInodes.sort(ops::inodeCmp);
+
+               this->files->insert(inlinedFileInodes, this->filesHandle);
+
+               numFileInodesFound->increase(inlinedFileInodes.size());
+
+               // add used targetIDs
+               for ( FsckFileInodeListIter iter = inlinedFileInodes.begin();
+                  iter != inlinedFileInodes.end(); iter++ )
+               {
+                  FsckTargetIDType fsckTargetIDType;
+
+                  if (iter->getStripePatternType() == FsckStripePatternType_BUDDYMIRROR)
+                     fsckTargetIDType = FsckTargetIDType_BUDDYGROUP;
+                  else
+                     fsckTargetIDType = FsckTargetIDType_TARGET;
+
+                  for (auto targetsIter = iter->getStripeTargets().begin();
+                     targetsIter != iter->getStripeTargets().end(); targetsIter++)
+                  {
+                     this->usedTargets->insert(FsckTargetID(*targetsIter, fsckTargetIDType) );
+                  }
+               }
+
+               // parse all new cont. directories
+               FsckContDirList& contDirs = retrieveDirEntriesRespMsg->getContDirs();
+
+               // check entry IDs
+               for (auto it = contDirs.begin(); it != contDirs.end(); )
+               {
+                  if (db::EntryID::tryFromStr(it->getID()).first)
+                  {
+                     ++it;
+                     continue;
+                  }
+
+                  LOG(GENERAL, ERR, "Found content directory with invalid entry ID.",
+                        ("node", it->getSaveNodeID()),
+                        ("isBuddyMirrored", it->getIsBuddyMirrored()),
+                        ("entryID", it->getID()));
+
+                  ++it;
+                  errors->increase();
+                  contDirs.erase(std::prev(it));
+               }
+
+               this->contDirs->insert(contDirs, this->contDirsHandle);
+            }
+            else
+            {
+               throw FsckException("Communication error occured with node " + node.getAlias());
+            }
+
+            // if any of the worker threads threw an exception, we should stop now!
+            if ( Program::getApp()->getShallAbort() )
+               return;
+
+         } while ( resultCount > 0 );
+      }
+   }
+}

fsck/source/components/worker/RetrieveDirEntriesWork.h

@@ -0,0 +1,62 @@
+#ifndef RETRIEVEDIRENTRIESWORK_H
+#define RETRIEVEDIRENTRIESWORK_H
+
+/*
+ * retrieve all dir entries from one node, inside a specified range of hashDirs and save them to DB
+ */
+
+#include <common/app/log/LogContext.h>
+#include <common/components/worker/Work.h>
+#include <common/toolkit/SynchronizedCounter.h>
+
+#include <database/FsckDB.h>
+#include <database/FsckDBTable.h>
+
+// the size of one response packet, i.e. how many dentries are asked for at once
+#define RETRIEVE_DIR_ENTRIES_PACKET_SIZE 500
+
+class RetrieveDirEntriesWork : public Work
+{
+   public:
+      /*
+       * @param db database instance
+       * @param node the node to retrieve data from
+       * @param counter a pointer to a Synchronized counter; this is incremented by one at the end
+       * and the calling thread can wait for the counter
+       * @param hashDirStart the first top-level hashDir to open
+       * @param hashDirEnd the last top-level hashDir to open
+       * @param numDentriesFound
+       * @param numFileInodesFound
+       */
+      RetrieveDirEntriesWork(FsckDB* db, Node& node, SynchronizedCounter* counter,
+         AtomicUInt64& errors, unsigned hashDirStart, unsigned hashDirEnd,
+         AtomicUInt64* numDentriesFound, AtomicUInt64* numFileInodesFound,
+         std::set<FsckTargetID>& usedTargets);
+
+      void process(char* bufIn, unsigned bufInLen, char* bufOut, unsigned bufOutLen);
+
+   private:
+      LogContext log;
+      Node& node;
+      SynchronizedCounter* counter;
+      AtomicUInt64* errors;
+      AtomicUInt64* numDentriesFound;
+      AtomicUInt64* numFileInodesFound;
+      std::set<FsckTargetID>* usedTargets;
+
+      unsigned hashDirStart;
+      unsigned hashDirEnd;
+
+      FsckDBDentryTable* dentries;
+      FsckDBDentryTable::BulkHandle dentriesHandle;
+
+      FsckDBFileInodesTable* files;
+      FsckDBFileInodesTable::BulkHandle filesHandle;
+
+      FsckDBContDirsTable* contDirs;
+      FsckDBContDirsTable::BulkHandle contDirsHandle;
+
+      void doWork(bool isBuddyMirrored);
+};
+
+#endif /* RETRIEVEDIRENTRIESWORK_H */

fsck/source/components/worker/RetrieveFsIDsWork.cpp

@@ -0,0 +1,125 @@
+#include "RetrieveFsIDsWork.h"
+#include <common/net/message/fsck/RetrieveFsIDsMsg.h>
+#include <common/net/message/fsck/RetrieveFsIDsRespMsg.h>
+#include <common/toolkit/MessagingTk.h>
+#include <common/toolkit/MetaStorageTk.h>
+#include <database/FsckDBException.h>
+#include <toolkit/FsckException.h>
+
+#include <program/Program.h>
+
+
+RetrieveFsIDsWork::RetrieveFsIDsWork(FsckDB* db, Node& node, SynchronizedCounter* counter,
+      AtomicUInt64& errors, unsigned hashDirStart, unsigned hashDirEnd) :
+   log("RetrieveFsIDsWork"), node(node), counter(counter), errors(&errors),
+   hashDirStart(hashDirStart), hashDirEnd(hashDirEnd),
+   table(db->getFsIDsTable()), bulkHandle(table->newBulkHandle())
+{
+}
+
+RetrieveFsIDsWork::~RetrieveFsIDsWork()
+{
+}
+
+void RetrieveFsIDsWork::process(char* bufIn, unsigned bufInLen, char* bufOut,
+   unsigned bufOutLen)
+{
+   log.log(4, "Processing RetrieveFsIDsWork");
+
+   try
+   {
+      doWork(false);
+      doWork(true);
+      table->flush(bulkHandle);
+      // work package finished => increment counter
+      this->counter->incCount();
+
+   }
+   catch (std::exception &e)
+   {
+      // exception thrown, but work package is finished => increment counter
+      this->counter->incCount();
+
+      // after incrementing counter, re-throw exception
+      throw;
+   }
+
+   log.log(4, "Processed RetrieveFsIDsWork");
+}
+
+void RetrieveFsIDsWork::doWork(bool isBuddyMirrored)
+{
+   for ( unsigned firstLevelhashDirNum = hashDirStart; firstLevelhashDirNum <= hashDirEnd;
+      firstLevelhashDirNum++ )
+   {
+      for ( unsigned secondLevelhashDirNum = 0;
+         secondLevelhashDirNum < META_DENTRIES_LEVEL2_SUBDIR_NUM; secondLevelhashDirNum++ )
+      {
+         unsigned hashDirNum = StorageTk::mergeHashDirs(firstLevelhashDirNum,
+            secondLevelhashDirNum);
+
+         int64_t hashDirOffset = 0;
+         int64_t contDirOffset = 0;
+         std::string currentContDirID;
+         int resultCount = 0;
+
+         do
+         {
+            RetrieveFsIDsMsg retrieveFsIDsMsg(hashDirNum, isBuddyMirrored, currentContDirID,
+               RETRIEVE_FSIDS_PACKET_SIZE, hashDirOffset, contDirOffset);
+
+            const auto respMsg = MessagingTk::requestResponse(node, retrieveFsIDsMsg,
+               NETMSGTYPE_RetrieveFsIDsResp);
+
+            if (respMsg)
+            {
+               auto* retrieveFsIDsRespMsg = (RetrieveFsIDsRespMsg*) respMsg.get();
+
+               // set new parameters
+               currentContDirID = retrieveFsIDsRespMsg->getCurrentContDirID();
+               hashDirOffset = retrieveFsIDsRespMsg->getNewHashDirOffset();
+               contDirOffset = retrieveFsIDsRespMsg->getNewContDirOffset();
+
+               // parse FS-IDs
+               FsckFsIDList& fsIDs = retrieveFsIDsRespMsg->getFsIDs();
+
+               // this is the actual result count we are interested in, because if no fsIDs
+               // were read, there is nothing left on the server
+               resultCount = fsIDs.size();
+
+               // check entry IDs
+               for (auto it = fsIDs.begin(); it != fsIDs.end(); )
+               {
+                  if (db::EntryID::tryFromStr(it->getID()).first
+                        && db::EntryID::tryFromStr(it->getParentDirID()).first)
+                  {
+                     ++it;
+                     continue;
+                  }
+
+                  LOG(GENERAL, ERR, "Found fsid file with invalid entry IDs.",
+                        ("node", it->getSaveNodeID()),
+                        ("isBuddyMirrored", it->getIsBuddyMirrored()),
+                        ("entryID", it->getID()),
+                        ("parentEntryID", it->getParentDirID()));
+
+                  ++it;
+                  errors->increase();
+                  fsIDs.erase(std::prev(it));
+               }
+
+               this->table->insert(fsIDs, this->bulkHandle);
+            }
+            else
+            {
+               throw FsckException("Communication error occured with node " + node.getAlias());
+            }
+
+            // if any of the worker threads threw an exception, we should stop now!
+            if ( Program::getApp()->getShallAbort() )
+               return;
+
+         } while ( resultCount > 0 );
+      }
+   }
+}

fsck/source/components/worker/RetrieveFsIDsWork.h

@@ -0,0 +1,49 @@
+#ifndef RETRIEVEFSIDSWORK_H
+#define RETRIEVEFSIDSWORK_H
+
+/*
+ * retrieve all FS-IDs from one node, inside a specified range of hashDirs and save them to DB
+ */
+
+#include <common/app/log/LogContext.h>
+#include <common/components/worker/Work.h>
+#include <common/toolkit/SynchronizedCounter.h>
+
+#include <database/FsckDB.h>
+#include <database/FsckDBTable.h>
+
+// the size of one response packet, i.e. how many fsids are asked for at once
+#define RETRIEVE_FSIDS_PACKET_SIZE 1000
+
+class RetrieveFsIDsWork : public Work
+{
+   public:
+      /*
+       * @param db database instance
+       * @param node the node to retrieve data from
+       * @param counter a pointer to a Synchronized counter; this is incremented by one at the end
+       * and the calling thread can wait for the counter
+       * @param hashDirStart the first top-level hashDir to open
+       * @param hashDirEnd the last top-level hashDir to open
+       */
+      RetrieveFsIDsWork(FsckDB* db, Node& node, SynchronizedCounter* counter, AtomicUInt64& errors,
+         unsigned hashDirStart, unsigned hashDirEnd);
+      virtual ~RetrieveFsIDsWork();
+      void process(char* bufIn, unsigned bufInLen, char* bufOut, unsigned bufOutLen);
+
+   private:
+      LogContext log;
+      Node& node;
+      SynchronizedCounter* counter;
+      AtomicUInt64* errors;
+
+      unsigned hashDirStart;
+      unsigned hashDirEnd;
+
+      FsckDBFsIDsTable* table;
+      FsckDBFsIDsTable::BulkHandle bulkHandle;
+
+      void doWork(bool isBuddyMirrored);
+};
+
+#endif /* RETRIEVEFSIDSWORK_H */

fsck/source/components/worker/RetrieveInodesWork.cpp

@@ -0,0 +1,187 @@
+#include "RetrieveInodesWork.h"
+#include <common/net/message/fsck/RetrieveInodesMsg.h>
+#include <common/net/message/fsck/RetrieveInodesRespMsg.h>
+#include <common/toolkit/MetaStorageTk.h>
+#include <common/toolkit/MessagingTk.h>
+#include <database/FsckDBException.h>
+#include <toolkit/FsckException.h>
+
+#include <program/Program.h>
+
+#include <set>
+
+RetrieveInodesWork::RetrieveInodesWork(FsckDB* db, Node& node, SynchronizedCounter* counter,
+      AtomicUInt64& errors, unsigned hashDirStart, unsigned hashDirEnd,
+      AtomicUInt64* numFileInodesFound, AtomicUInt64* numDirInodesFound,
+      std::set<FsckTargetID>& usedTargets) :
+   log("RetrieveInodesWork"), node(node), counter(counter), errors(&errors),
+   usedTargets(&usedTargets), hashDirStart(hashDirStart), hashDirEnd(hashDirEnd),
+   numFileInodesFound(numFileInodesFound), numDirInodesFound(numDirInodesFound),
+   files(db->getFileInodesTable()), filesHandle(files->newBulkHandle()),
+   dirs(db->getDirInodesTable()), dirsHandle(dirs->newBulkHandle())
+{
+}
+
+void RetrieveInodesWork::process(char* bufIn, unsigned bufInLen, char* bufOut,
+   unsigned bufOutLen)
+{
+   log.log(4, "Processing RetrieveInodesWork");
+
+   try
+   {
+      doWork(false);
+      doWork(true);
+      // flush buffers before signaling completion
+      files->flush(filesHandle);
+      dirs->flush(dirsHandle);
+      // work package finished => increment counter
+      this->counter->incCount();
+
+   }
+   catch (std::exception &e)
+   {
+      // exception thrown, but work package is finished => increment counter
+      this->counter->incCount();
+
+      // after incrementing counter, re-throw exception
+      throw;
+   }
+
+   log.log(4, "Processed RetrieveInodesWork");
+}
+
+void RetrieveInodesWork::doWork(bool isBuddyMirrored)
+{
+   const NumNodeID& metaRootID = Program::getApp()->getMetaRoot().getID();
+   const NumNodeID& nodeID = node.getNumID();
+   const NumNodeID nodeBuddyGroupID = NumNodeID(Program::getApp()->getMetaMirrorBuddyGroupMapper()
+         ->getBuddyGroupID(node.getNumID().val()));
+
+   for ( unsigned firstLevelhashDirNum = hashDirStart; firstLevelhashDirNum <= hashDirEnd;
+      firstLevelhashDirNum++ )
+   {
+      for ( unsigned secondLevelhashDirNum = 0;
+         secondLevelhashDirNum < META_DENTRIES_LEVEL2_SUBDIR_NUM; secondLevelhashDirNum++ )
+      {
+         unsigned hashDirNum = StorageTk::mergeHashDirs(firstLevelhashDirNum,
+            secondLevelhashDirNum);
+
+         int64_t lastOffset = 0;
+         size_t fileInodeCount;
+         size_t dirInodeCount;
+
+         do
+         {
+            RetrieveInodesMsg retrieveInodesMsg(hashDirNum, lastOffset,
+               RETRIEVE_INODES_PACKET_SIZE, isBuddyMirrored);
+
+            const auto respMsg = MessagingTk::requestResponse(node, retrieveInodesMsg,
+               NETMSGTYPE_RetrieveInodesResp);
+            if (respMsg)
+            {
+               auto* retrieveInodesRespMsg = (RetrieveInodesRespMsg*) respMsg.get();
+
+               // set new parameters
+               lastOffset = retrieveInodesRespMsg->getLastOffset();
+
+               // parse all file inodes
+               FsckFileInodeList& fileInodes = retrieveInodesRespMsg->getFileInodes();
+
+               // check inode entry IDs
+               for (auto it = fileInodes.begin(); it != fileInodes.end(); )
+               {
+                  if (db::EntryID::tryFromStr(it->getID()).first
+                        && db::EntryID::tryFromStr(it->getParentDirID()).first
+                        && (!it->getPathInfo()->hasOrigFeature()
+                              || db::EntryID::tryFromStr(
+                                    it->getPathInfo()->getOrigParentEntryID()).first))
+                  {
+                     ++it;
+                     continue;
+                  }
+
+                  LOG(GENERAL, ERR, "Found inode with invalid entry IDs.",
+                        ("node", it->getSaveNodeID()),
+                        ("isBuddyMirrored", it->getIsBuddyMirrored()),
+                        ("entryID", it->getID()),
+                        ("parentEntryID", it->getParentDirID()),
+                        ("origParent", it->getPathInfo()->getOrigParentEntryID()));
+
+                  ++it;
+                  errors->increase();
+                  fileInodes.erase(std::prev(it));
+               }
+
+               // add targetIDs
+               for (auto iter = fileInodes.begin(); iter != fileInodes.end(); iter++)
+               {
+                  FsckTargetIDType fsckTargetIDType;
+
+                  if (iter->getStripePatternType() == FsckStripePatternType_BUDDYMIRROR)
+                     fsckTargetIDType = FsckTargetIDType_BUDDYGROUP;
+                  else
+                     fsckTargetIDType = FsckTargetIDType_TARGET;
+
+                  for (auto targetsIter = iter->getStripeTargets().begin();
+                     targetsIter != iter->getStripeTargets().end(); targetsIter++)
+                  {
+                     this->usedTargets->insert(FsckTargetID(*targetsIter, fsckTargetIDType) );
+                  }
+               }
+
+               // parse all directory inodes
+               FsckDirInodeList& dirInodes = retrieveInodesRespMsg->getDirInodes();
+
+               // check inode entry IDs
+               for (auto it = dirInodes.begin(); it != dirInodes.end(); )
+               {
+                  auto entryIDPair = db::EntryID::tryFromStr(it->getID());
+                  if (!entryIDPair.first ||
+                        !db::EntryID::tryFromStr(it->getParentDirID()).first)
+                  {
+                     LOG(GENERAL, ERR, "Found inode with invalid entry IDs.",
+                           ("node", it->getSaveNodeID()),
+                           ("isBuddyMirrored", it->getIsBuddyMirrored()),
+                           ("entryID", it->getID()),
+                           ("parentEntryID", it->getParentDirID()));
+
+                     ++it;
+                     errors->increase();
+                     dirInodes.erase(std::prev(it));
+                     continue;
+                  }
+
+                  // remove root inodes from non root metas
+                  if (entryIDPair.second.isRootDir() &&
+                        ((it->getIsBuddyMirrored() && nodeBuddyGroupID != metaRootID)
+                        || (!it->getIsBuddyMirrored() && nodeID != metaRootID)))
+                  {
+                     ++it;
+                     dirInodes.erase(std::prev(it));
+                     continue;
+                  }
+                  ++it;
+               }
+
+               fileInodeCount = fileInodes.size();
+               dirInodeCount = dirInodes.size();
+
+               this->files->insert(fileInodes, this->filesHandle);
+               this->dirs->insert(dirInodes, this->dirsHandle);
+
+               numFileInodesFound->increase(fileInodeCount);
+               numDirInodesFound->increase(dirInodeCount);
+            }
+            else
+            {
+               throw FsckException("Communication error occured with node " + node.getAlias());
+            }
+
+            // if any of the worker threads threw an exception, we should stop now!
+            if ( Program::getApp()->getShallAbort() )
+               return;
+
+         } while ( (fileInodeCount + dirInodeCount) > 0 );
+      }
+   }
+}

fsck/source/components/worker/RetrieveInodesWork.h

@@ -0,0 +1,58 @@
+#ifndef RETRIEVEINODESWORK_H
+#define RETRIEVEINODESWORK_H
+
+/*
+ * retrieve all inodes from one node, inside a specified range of hashDirs and save them to DB
+
+ */
+
+#include <common/app/log/LogContext.h>
+#include <common/components/worker/Work.h>
+#include <common/toolkit/SynchronizedCounter.h>
+#include <database/FsckDB.h>
+#include <database/FsckDBTable.h>
+
+// the size of one response packet, i.e. how many inodes are asked for at once
+#define RETRIEVE_INODES_PACKET_SIZE 500
+
+class RetrieveInodesWork : public Work
+{
+   public:
+      /*
+       * @param db database instance
+       * @param node the node to retrieve data from
+       * @param counter a pointer to a Synchronized counter; this is incremented by one at the end
+       * and the calling thread can wait for the counter
+       * @param hashDirStart the first top-level hashDir to open
+       * @param hashDirEnd the last top-level hashDir to open
+       */
+      RetrieveInodesWork(FsckDB* db, Node& node, SynchronizedCounter* counter,
+         AtomicUInt64& errors, unsigned hashDirStart, unsigned hashDirEnd,
+         AtomicUInt64* numFileInodesFound, AtomicUInt64* numDirInodesFound,
+         std::set<FsckTargetID>& usedTargets);
+
+      void process(char* bufIn, unsigned bufInLen, char* bufOut, unsigned bufOutLen);
+
+   private:
+      LogContext log;
+      Node& node;
+      SynchronizedCounter* counter;
+      AtomicUInt64* errors;
+      std::set<FsckTargetID>* usedTargets;
+
+      unsigned hashDirStart;
+      unsigned hashDirEnd;
+
+      AtomicUInt64* numFileInodesFound;
+      AtomicUInt64* numDirInodesFound;
+
+      FsckDBFileInodesTable* files;
+      FsckDBFileInodesTable::BulkHandle filesHandle;
+
+      FsckDBDirInodesTable* dirs;
+      FsckDBDirInodesTable::BulkHandle dirsHandle;
+
+      void doWork(bool isBuddyMirrored);
+};
+
+#endif /* RETRIEVEINODESWORK_H */

fsck/source/database/Buffer.h

@@ -0,0 +1,47 @@
+#ifndef BUFFER_H_
+#define BUFFER_H_
+
+#include <database/Set.h>
+
+template<typename Data>
+class Buffer {
+   private:
+      Set<Data>* set;
+      SetFragment<Data>* currentFragment;
+      size_t fragmentSize;
+
+      Buffer(const Buffer&);
+      Buffer& operator=(const Buffer&);
+
+   public:
+      Buffer(Set<Data>& set, size_t fragmentSize)
+         : set(&set), currentFragment(NULL), fragmentSize(fragmentSize)
+      {}
+
+      ~Buffer()
+      {
+         flush();
+      }
+
+      void flush()
+      {
+         if(currentFragment)
+         {
+            currentFragment->flush();
+            currentFragment = NULL;
+         }
+      }
+
+      void append(const Data& data)
+      {
+         if(!currentFragment)
+            currentFragment = set->newFragment();
+
+         currentFragment->append(data);
+
+         if(currentFragment->size() * sizeof(Data) >= fragmentSize)
+            flush();
+      }
+};
+
+#endif

fsck/source/database/Chunk.h

@@ -0,0 +1,55 @@
+#ifndef CHUNK_H_
+#define CHUNK_H_
+
+#include <common/fsck/FsckChunk.h>
+#include <database/EntryID.h>
+
+#include <boost/tuple/tuple.hpp>
+#include <boost/tuple/tuple_comparison.hpp>
+
+namespace db {
+
+struct Chunk {
+   static const unsigned SAVED_PATH_SIZE = 43;
+
+   EntryID id;             /* 12 */
+   uint32_t targetID;      /* 16 */
+   uint32_t buddyGroupID;  /* 20 */
+
+   /* savedPath layout is ... weird.
+    * we have two layouts:
+    *   - <u16 _b16>/<u16 _b16>/<id _string>
+    *   - u<origOwnerUID _b16>/<(ts >> 16) _b16>/<(ts >> 12) & 0xF _b16>/<id _string>
+    *
+    * where u16 are random 16 bit unsigneds each. that gives maximum path lengths of
+    *   - 2 + 1 + 2 + 26
+    *   - 9 + 1 + 4 + 1 + 1 + 1 + 26
+    * add a terminating NUL to each, the maximum path length becomes 44 bytes,
+    * which is also conveniently aligned for uint_64t followers */
+   char savedPath[SAVED_PATH_SIZE + 1]; /* 64 */
+
+   int64_t fileSize;       /* 72 */
+   int64_t usedBlocks;     /* 80 */
+
+   uint32_t uid;           /* 84 */
+   uint32_t gid;           /* 88 */
+
+   typedef boost::tuple<EntryID, uint32_t, uint32_t> KeyType;
+
+   KeyType pkey() const
+   {
+      return KeyType(id, targetID, buddyGroupID);
+   }
+
+   operator FsckChunk() const
+   {
+      Path path(savedPath);
+
+      return FsckChunk(id.str(), targetID, path, fileSize, usedBlocks, 0, 0, 0, uid,
+         gid, buddyGroupID);
+   }
+};
+
+}
+
+#endif

fsck/source/database/ContDir.h

@@ -0,0 +1,26 @@
+#ifndef CONTDIR_H_
+#define CONTDIR_H_
+
+#include <common/fsck/FsckContDir.h>
+#include <database/EntryID.h>
+
+namespace db {
+
+struct ContDir {
+   EntryID id;                 /* 12 */
+   uint32_t saveNodeID;        /* 16 */
+   uint32_t isBuddyMirrored:1; /* 20 */
+
+   typedef EntryID KeyType;
+
+   KeyType pkey() const { return id; }
+
+   operator FsckContDir() const
+   {
+      return FsckContDir(id.str(), NumNodeID(saveNodeID), isBuddyMirrored);
+   }
+};
+
+}
+
+#endif

fsck/source/database/Cursor.h

@@ -0,0 +1,67 @@
+#ifndef CURSOR_H_
+#define CURSOR_H_
+
+#include <boost/make_shared.hpp>
+#include <boost/shared_ptr.hpp>
+
+template<typename Obj>
+class Cursor
+{
+   public:
+      typedef Obj ElementType;
+
+   private:
+      class SourceBase
+      {
+         public:
+            virtual ~SourceBase() {}
+
+            virtual bool step() = 0;
+            virtual ElementType* get() = 0;
+      };
+
+      template<typename Inner>
+      class Source : public SourceBase
+      {
+         public:
+            Source(Inner inner)
+               : inner(inner)
+            {
+            }
+
+            bool step()
+            {
+               return inner.step();
+            }
+
+            ElementType* get()
+            {
+               return inner.get();
+            }
+
+         private:
+            Inner inner;
+      };
+
+   public:
+      template<typename Inner>
+      explicit Cursor(Inner inner)
+         : source(boost::make_shared<Source<Inner> >(inner) )
+      {
+      }
+
+      bool step()
+      {
+         return source->step();
+      }
+
+      ElementType* get()
+      {
+         return source->get();
+      }
+
+   private:
+      boost::shared_ptr<SourceBase> source;
+};
+
+#endif

fsck/source/database/DirEntry.h

@@ -0,0 +1,67 @@
+#ifndef DIRENTRY_H_
+#define DIRENTRY_H_
+
+#include <common/fsck/FsckDirEntry.h>
+#include <database/EntryID.h>
+
+namespace db {
+
+struct DirEntry
+{
+   EntryID id;                          /* 12 */
+   EntryID parentDirID;                 /* 24 */
+
+   uint32_t entryOwnerNodeID;           /* 28 */
+   uint32_t inodeOwnerNodeID;           /* 32 */
+
+   uint32_t saveNodeID;                 /* 36 */
+   int32_t saveDevice;                  /* 40 */
+   uint64_t saveInode;                  /* 48 */
+
+   union {
+      struct {
+         char text[16];               /* +16 */
+      } inlined;
+      struct {
+         /* take name from a specific offset in some other file,
+          * identified by a u64 for convenience */
+         uint64_t fileID;             /* +8 */
+         uint64_t fileOffset;         /* +16 */
+      } extended;
+   } name;                              /* 64 */
+
+   // the identifying key for dir entries is actually (parent, name). since storing the name as
+   // part of the key is very inefficient, and existence of duplicate names per parent can be
+   // easily excluded, we instead number all dir entries to create a surrogate key (seqNo), which
+   // will also work as (parent, seqNo).
+   //
+   // the actual primary key of the table is (id, parent, GFID, seqNo) though. this is because
+   // most querys only need the id, a few queries need the full fsid linking identifier
+   // (id, parent, GFID), and deletion always need (parent, seqNo). this also means that the
+   // deletion tracker set for dir entries will be rather large per entry, but it does maek queries
+   // more efficient.
+   uint64_t seqNo;                      /* 72 */
+
+   uint32_t entryType:16;
+   uint32_t fileNameInlined:1;
+   uint32_t hasInlinedInode:1;
+   uint32_t isBuddyMirrored:1;          /* 76 */
+
+   typedef boost::tuple<EntryID, EntryID, uint32_t, int32_t, uint32_t, uint64_t> KeyType;
+
+   KeyType pkey() const
+   {
+      return KeyType(id, parentDirID, saveNodeID, saveDevice, saveInode, seqNo);
+   }
+
+   operator FsckDirEntry() const
+   {
+      return FsckDirEntry(id.str(), "[<unresolved>]", parentDirID.str(),
+         NumNodeID(entryOwnerNodeID), NumNodeID(inodeOwnerNodeID), FsckDirEntryType(entryType),
+         hasInlinedInode, NumNodeID(saveNodeID), saveDevice, saveInode, isBuddyMirrored, seqNo);
+   }
+};
+
+}
+
+#endif

fsck/source/database/DirInode.h

@@ -0,0 +1,40 @@
+#ifndef DIRINODE_H_
+#define DIRINODE_H_
+
+#include <common/fsck/FsckDirInode.h>
+#include <database/EntryID.h>
+
+namespace db {
+
+struct DirInode {
+   EntryID id;                /* 12 */
+   EntryID parentDirID;       /* 24 */
+
+   uint32_t parentNodeID;       /* 28 */
+   uint32_t ownerNodeID;        /* 32 */
+   uint32_t saveNodeID;         /* 36 */
+
+   uint32_t stripePatternType:16;
+   uint32_t readable:1;
+   uint32_t isBuddyMirrored:1;
+   uint32_t isMismirrored:1;    /* 40 */
+
+   uint64_t size;               /* 48 */
+   uint64_t numHardlinks;       /* 56 */
+
+   typedef EntryID KeyType;
+
+   EntryID pkey() const { return id; }
+
+   operator FsckDirInode() const
+   {
+      return FsckDirInode(id.str(), parentDirID.str(), NumNodeID(parentNodeID),
+         NumNodeID(ownerNodeID), size, numHardlinks, UInt16Vector(),
+         FsckStripePatternType(stripePatternType), NumNodeID(saveNodeID), isBuddyMirrored,
+         readable, isMismirrored);
+   }
+};
+
+}
+
+#endif

fsck/source/database/DiskList.h

@@ -0,0 +1,149 @@
+#ifndef DISKLIST_H_
+#define DISKLIST_H_
+
+#include <common/threading/Mutex.h>
+#include <common/toolkit/serialization/Serialization.h>
+
+#include <cerrno>
+#include <stdexcept>
+#include <string>
+#include <mutex>
+#include <vector>
+
+#include <fcntl.h>
+#include <stdio.h>
+#include <unistd.h>
+
+struct DiskListDoesNotExist : public std::runtime_error
+{
+   DiskListDoesNotExist(const std::string& file)
+      : runtime_error(file)
+   {}
+};
+
+template<typename Data>
+class DiskList;
+
+template<typename Data>
+class DiskListCursor
+{
+   friend class DiskList<Data>;
+
+   public:
+      bool step()
+      {
+         uint64_t itemLen;
+
+         {
+            ssize_t readRes = ::pread(fd, &itemLen, sizeof(itemLen), offset);
+
+            if (readRes == 0)
+               return false;
+
+            if (readRes < (ssize_t) sizeof(itemLen))
+               throw std::runtime_error("could not read from disk list file: " + std::string(strerror(errno)));
+         }
+
+         itemLen = LE_TO_HOST_64(itemLen);
+
+         std::unique_ptr<char[]> itemBuf(new char[itemLen]);
+
+         ssize_t readRes = ::pread(fd, itemBuf.get(), itemLen, offset + sizeof(itemLen));
+
+         if (readRes < (ssize_t) itemLen)
+            throw std::runtime_error("could not read from disk list file: " + std::string(strerror(errno)));
+
+         Deserializer des(itemBuf.get(), itemLen);
+
+         des % item;
+         if (!des.good())
+            throw std::runtime_error("could not read from disk list file: " + std::string(strerror(errno)));
+
+         offset += sizeof(itemLen) + itemLen;
+         return true;
+      }
+
+      Data* get()
+      {
+         return &item;
+      }
+
+   private:
+      int fd;
+      size_t offset;
+      Data item;
+
+      DiskListCursor(int fd):
+         fd(fd), offset(0)
+      {
+      }
+};
+
+template<typename Data>
+class DiskList {
+   private:
+      std::string file;
+      int fd;
+      uint64_t itemCount;
+      Mutex mtx;
+
+      DiskList(const DiskList&);
+      DiskList& operator=(const DiskList&);
+
+   public:
+      DiskList(const std::string& file, bool allowCreate = true) :
+         file(file), itemCount(0)
+      {
+         fd = ::open(file.c_str(), O_RDWR | (allowCreate ? O_CREAT : 0), 0660);
+         if(fd < 0)
+         {
+            int eno = errno;
+            if(!allowCreate && errno == ENOENT)
+               throw DiskListDoesNotExist(file);
+            else
+               throw std::runtime_error("could not open disk list file " + file + ": " + strerror(eno));
+         }
+      }
+
+      ~DiskList()
+      {
+         if(fd >= 0)
+            close(fd);
+      }
+
+      const std::string filename() const { return file; }
+
+      void append(const Data& data)
+      {
+         std::lock_guard<Mutex> lock(mtx);
+
+         boost::scoped_array<char> buffer;
+
+         ssize_t size = serializeIntoNewBuffer(data, buffer);
+         if (size < 0)
+            throw std::runtime_error("error serializing disk list item");
+
+         uint64_t bufSize = HOST_TO_LE_64(size);
+         if (::write(fd, &bufSize, sizeof(bufSize)) < (ssize_t) sizeof(bufSize))
+            throw std::runtime_error("error writing disk list item: " + std::string(strerror(errno)));
+
+         if (::write(fd, buffer.get(), size) < size)
+            throw std::runtime_error("error writing disk list item: " + std::string(strerror(errno)));
+      }
+
+      DiskListCursor<Data> cursor()
+      {
+         return DiskListCursor<Data>(fd);
+      }
+
+      void clear()
+      {
+         if (fd >= 0)
+         {
+            if (ftruncate(fd, 0) < 0)
+               throw std::runtime_error("error clearing disk list: " + std::string(strerror(errno)));
+         }
+      }
+};
+
+#endif

fsck/source/database/Distinct.h

@@ -0,0 +1,88 @@
+#ifndef DISTINCT_H_
+#define DISTINCT_H_
+
+#include <boost/type_traits/decay.hpp>
+#include <boost/utility/result_of.hpp>
+
+template<typename Source, typename KeyExtract>
+class Distinct
+{
+   private:
+      typedef typename boost::decay<
+            typename boost::result_of<KeyExtract(typename Source::ElementType&)>::type
+         >::type KeyType;
+
+   public:
+      typedef typename Source::ElementType ElementType;
+      typedef typename Source::MarkerType MarkerType;
+
+   public:
+      Distinct(Source source, KeyExtract keyExtract)
+         : source(source), keyExtract(keyExtract), hasKey(false)
+      {}
+
+      bool step()
+      {
+         if(!this->source.step() )
+            return false;
+
+         while(this->hasKey && this->key == this->keyExtract(*this->source.get() ) )
+         {
+            if(!this->source.step() )
+               return false;
+         }
+
+         this->hasKey = true;
+         this->key = this->keyExtract(*this->source.get() );
+         return true;
+      }
+
+      ElementType* get()
+      {
+         return this->source.get();
+      }
+
+      MarkerType mark() const
+      {
+         return this->source.mark();
+      }
+
+      void restore(MarkerType mark)
+      {
+         this->source.restore(mark);
+         this->hasKey = true;
+         this->key = this->keyExtract(*this->source.get() );
+      }
+
+   private:
+      Source source;
+      KeyType key;
+      KeyExtract keyExtract;
+      bool hasKey;
+};
+
+
+namespace db {
+
+template<typename KeyExtract>
+struct DistinctOp
+{
+   KeyExtract keyExtract;
+
+   template<typename Source>
+   friend Distinct<Source, KeyExtract> operator|(Source source, const DistinctOp& op)
+   {
+      return Distinct<Source, KeyExtract>(source, op.keyExtract);
+   }
+};
+
+template<typename KeyExtract>
+inline DistinctOp<KeyExtract> distinctBy(KeyExtract keyExtract)
+{
+   DistinctOp<KeyExtract> result = {keyExtract};
+   return result;
+}
+
+}
+
+#endif

fsck/source/database/EntryID.h

@@ -0,0 +1,163 @@
+#ifndef ENTRYID_H_
+#define ENTRYID_H_
+
+#include <common/storage/Metadata.h>
+
+#include <algorithm>
+#include <cstdio>
+#include <exception>
+#include <string>
+
+#include <boost/tuple/tuple.hpp>
+#include <boost/tuple/tuple_comparison.hpp>
+
+namespace db {
+
+struct EntryID {
+   unsigned sequence;
+   unsigned timestamp;
+   unsigned nodeID;
+
+   // required for queries, since they use boost::tuple to avoid writing out huge comparison
+   // operators
+   EntryID()
+      : sequence(0), timestamp(0), nodeID(0)
+   {}
+
+   EntryID(unsigned sequence, unsigned timestamp, unsigned nodeID)
+      : sequence(sequence), timestamp(timestamp), nodeID(nodeID)
+   {
+      if(nodeID == 0 && (timestamp != 0 || sequence > 3) )
+         throw std::exception();
+   }
+
+   bool isSpecial() const { return nodeID == 0; }
+
+   bool operator<(const EntryID& other) const
+   {
+      return boost::make_tuple(sequence, timestamp, nodeID)
+         < boost::make_tuple(other.sequence, other.timestamp, other.nodeID);
+   }
+
+   bool operator==(const EntryID& other) const
+   {
+      return this->sequence == other.sequence
+         && this->timestamp == other.timestamp
+         && this->nodeID == other.nodeID;
+   }
+
+   bool operator!=(const EntryID& other) const
+   {
+      return !(*this == other);
+   }
+
+   bool isDisposalDir() const
+   {
+      return *this == disposal() || *this == mdisposal();
+   }
+
+   bool isRootDir() const
+   {
+      return *this == root();
+   }
+
+   std::string str() const
+   {
+      if(isSpecial() )
+      {
+         if(*this == anchor() )
+            return "";
+
+         if(*this == root() )
+            return META_ROOTDIR_ID_STR;
+
+         if(*this == disposal() )
+            return META_DISPOSALDIR_ID_STR;
+
+         if(*this == mdisposal())
+            return META_MIRRORDISPOSALDIR_ID_STR;
+
+         throw std::exception();
+      }
+
+      char buffer[3*8 + 2 + 1];
+      sprintf(buffer, "%X-%X-%X", sequence, timestamp, nodeID);
+      return buffer;
+   }
+
+   static EntryID anchor() { return EntryID(0, 0, 0); }
+   static EntryID root() { return EntryID(1, 0, 0); }
+   static EntryID disposal() { return EntryID(2, 0, 0); }
+   static EntryID mdisposal() { return EntryID(3, 0, 0); }
+
+   static EntryID fromStr(const std::string& str)
+   {
+      auto pair = tryFromStr(str);
+      if (!pair.first)
+         throw std::exception();
+
+      return pair.second;
+   }
+
+   static std::pair<bool, EntryID> tryFromStr(const std::string& str)
+   {
+      if (str.empty())
+         return {true, anchor()};
+
+      if (str == META_ROOTDIR_ID_STR)
+         return {true, root()};
+
+      if (str == META_DISPOSALDIR_ID_STR)
+         return {true, disposal()};
+
+      if (str == META_MIRRORDISPOSALDIR_ID_STR)
+         return {true, mdisposal()};
+
+      std::string::size_type sep1;
+      std::string::size_type sep2;
+
+      sep1 = str.find('-');
+      sep2 = str.find('-', sep1 + 1);
+
+      if (sep1 == str.npos || sep2 == str.npos)
+         return {false, {}};
+
+      std::string seqStr = str.substr(0, sep1);
+      std::string tsStr = str.substr(sep1 + 1, sep2 - (sep1 + 1));
+      std::string nodeStr = str.substr(sep2 + 1);
+
+      if (seqStr.empty() || tsStr.empty() || nodeStr.empty() ||
+            seqStr.size() > 8 || tsStr.size() > 8 || nodeStr.size() > 8)
+         return {false, {}};
+
+      struct
+      {
+         static bool notHex(char c)
+         {
+            return !std::isxdigit(c);
+         }
+
+         bool operator()(const std::string& str) const
+         {
+            return std::count_if(str.begin(), str.end(), notHex) == 0;
+         }
+      } onlyHex;
+
+      if (!onlyHex(seqStr) || !onlyHex(tsStr) || !onlyHex(nodeStr) )
+         return {false, {}};
+
+      unsigned seq;
+      unsigned ts;
+      unsigned node;
+
+      std::sscanf(seqStr.c_str(), "%x", &seq);
+      std::sscanf(tsStr.c_str(), "%x", &ts);
+      std::sscanf(nodeStr.c_str(), "%x", &node);
+
+      return {true, EntryID(seq, ts, node)};
+   }
+};
+
+}
+
+#endif

fsck/source/database/FileInode.h

@@ -0,0 +1,99 @@
+#ifndef FILEINODE_H_
+#define FILEINODE_H_
+
+#include <common/fsck/FsckFileInode.h>
+#include <database/EntryID.h>
+
+namespace db {
+
+struct FileInode {
+   EntryID id;                   /* 12 */
+   EntryID parentDirID;          /* 24 */
+   EntryID origParentEntryID;    /* 36 */
+
+   uint32_t parentNodeID;        /* 40 */
+   uint32_t saveNodeID;          /* 44 */
+   uint32_t origParentUID;       /* 48 */
+
+   uint32_t uid;                 /* 52 */
+   uint32_t gid;                 /* 56 */
+
+   uint64_t fileSize;            /* 64 */
+   uint64_t usedBlocks;          /* 72 */
+
+   uint64_t numHardlinks;        /* 80 */
+
+   uint64_t saveInode;           /* 88 */
+   int32_t saveDevice;           /* 92 */
+
+   uint32_t chunkSize;           /* 96 */
+
+   enum {
+      NTARGETS = 6
+   };
+   uint16_t targets[NTARGETS];   /* 108 */
+
+   uint32_t isInlined:1;
+   uint32_t pathInfoFlags:4;
+   uint32_t stripePatternType:4;
+   uint32_t stripePatternSize:20;
+   uint32_t readable:1;
+   uint32_t isBuddyMirrored:1;
+   uint32_t isMismirrored:1;     /* 112 */
+
+   typedef EntryID KeyType;
+
+   EntryID pkey() const { return id; }
+
+   FsckFileInode toInodeWithoutStripes() const
+   {
+      PathInfo info(origParentUID, origParentEntryID.str(), pathInfoFlags);
+
+      return FsckFileInode(id.str(), parentDirID.str(), NumNodeID(parentNodeID), info, uid, gid,
+            fileSize, numHardlinks, usedBlocks, {}, FsckStripePatternType(stripePatternType),
+            chunkSize, NumNodeID(saveNodeID), saveInode, saveDevice, isInlined, isBuddyMirrored,
+            readable, isMismirrored);
+   }
+
+   friend std::ostream& operator<<(std::ostream& os, FileInode const& obj)
+   {
+         os << "-------------" << "\n";
+         os << "FsckFileInode" << "\n";
+         os << "-------------" << "\n";
+         os << "EntryID: " << obj.id.str() << "\n";
+         os << "Parent dir's entryID: " << obj.parentDirID.str() << "\n";
+         os << "Orig parent dir's entryID: " << obj.origParentEntryID.str() << "\n";
+         os << "parent nodeID: " << obj.parentNodeID << "\n";
+         os << "save nodeID: " << obj.saveNodeID << "\n";
+         os << "origParentUID: " << obj.origParentUID << "\n";
+         os << "uid: " << obj.uid << "\n";
+         os << "gid: " << obj.gid << "\n";
+         os << "fileSize: " << obj.fileSize << "\n";
+         os << "usedBlocks: " << obj.usedBlocks << "\n";
+         os << "numHardlinks: " << obj.numHardlinks << "\n";
+         os << "saveInode: " << obj.saveInode << "\n";
+         os << "saveDevice: " << obj.saveDevice << "\n";
+         os << "chunkSize: " << obj.chunkSize << "\n";
+
+         os << "Stripe targets: [ ";
+
+         for (int i=0; i< NTARGETS; i++)
+         {
+            os << obj.targets[i] << " ";
+         }
+         os << "]\n";
+
+         os << "isInlined: " << obj.isInlined << "\n";
+         os << "pathInfoFlags: " << obj.pathInfoFlags << "\n";
+         os << "stripePatternType: " << obj.stripePatternType << "\n";
+         os << "stripePatternSize: " << obj.stripePatternSize << "\n";
+         os << "readable: " << obj.readable << "\n";
+         os << "isBuddyMirrored: " << obj.isBuddyMirrored << "\n";
+         os << "isMismirrored: " << obj.isMismirrored << "\n\n";
+         return os;
+   }
+};
+
+}
+
+#endif

fsck/source/database/Filter.h

@@ -0,0 +1,72 @@
+#ifndef FILTER_H_
+#define FILTER_H_
+
+template<typename Source, class Pred>
+class Filter
+{
+   public:
+      typedef typename Source::ElementType ElementType;
+      typedef typename Source::MarkerType MarkerType;
+
+   public:
+      Filter(Source source, Pred pred)
+         : source(source), pred(pred)
+      {
+      }
+
+      bool step()
+      {
+         while(this->source.step() )
+         {
+            if(this->pred(*get() ) )
+               return true;
+         }
+
+         return false;
+      }
+
+      ElementType* get()
+      {
+         return this->source.get();
+      }
+
+      MarkerType mark() const
+      {
+         return this->source.mark();
+      }
+
+      void restore(MarkerType mark)
+      {
+         this->source.restore(mark);
+      }
+
+   private:
+      Source source;
+      Pred pred;
+};
+
+
+namespace db {
+
+template<typename Pred>
+struct FilterOp
+{
+   Pred pred;
+
+   template<typename Source>
+   friend Filter<Source, Pred> operator|(Source source, const FilterOp& op)
+   {
+      return Filter<Source, Pred>(source, op.pred);
+   }
+};
+
+template<typename Pred>
+inline FilterOp<Pred> where(Pred pred)
+{
+   FilterOp<Pred> result = {pred};
+   return result;
+}
+
+}
+
+#endif

fsck/source/database/FsID.h

@@ -0,0 +1,38 @@
+#ifndef FSID_H_
+#define FSID_H_
+
+#include <common/fsck/FsckFsID.h>
+#include <database/EntryID.h>
+
+#include <boost/tuple/tuple.hpp>
+#include <boost/tuple/tuple_comparison.hpp>
+
+namespace db {
+
+struct FsID {
+   EntryID id;                  /* 12 */
+   EntryID parentDirID;         /* 24 */
+
+   uint32_t saveNodeID;         /* 28 */
+   int32_t saveDevice;          /* 32 */
+   uint64_t saveInode;          /* 40 */
+
+   uint64_t isBuddyMirrored:1;  /* 48 */
+
+   typedef boost::tuple<EntryID, EntryID, uint32_t, int32_t, uint32_t> KeyType;
+
+   KeyType pkey() const
+   {
+      return KeyType(id, parentDirID, saveNodeID, saveDevice, saveInode);
+   }
+
+   operator FsckFsID() const
+   {
+      return FsckFsID(id.str(), parentDirID.str(), NumNodeID(saveNodeID), saveDevice, saveInode,
+            isBuddyMirrored);
+   }
+};
+
+}
+
+#endif

fsck/source/database/FsckDB.cpp

@@ -0,0 +1,40 @@
+#include "FsckDB.h"
+
+#include <program/Program.h>
+#include <common/toolkit/StorageTk.h>
+#include <common/toolkit/StringTk.h>
+#include <database/FsckDBException.h>
+#include <database/FsckDBTable.h>
+#include <toolkit/FsckTkEx.h>
+
+#include <cstdio>
+
+FsckDB::FsckDB(const std::string& databasePath, size_t fragmentSize, size_t nameCacheLimit,
+      bool allowCreate)
+   : log("FsckDB"),
+     databasePath(databasePath),
+     dentryTable(new FsckDBDentryTable(databasePath, fragmentSize, nameCacheLimit, allowCreate) ),
+     fileInodesTable(new FsckDBFileInodesTable(databasePath, fragmentSize, allowCreate) ),
+     dirInodesTable(new FsckDBDirInodesTable(databasePath, fragmentSize, allowCreate) ),
+     chunksTable(new FsckDBChunksTable(databasePath, fragmentSize, allowCreate) ),
+     contDirsTable(new FsckDBContDirsTable(databasePath, fragmentSize, allowCreate) ),
+     fsIDsTable(new FsckDBFsIDsTable(databasePath, fragmentSize, allowCreate) ),
+     usedTargetIDsTable(new FsckDBUsedTargetIDsTable(databasePath, fragmentSize, allowCreate) ),
+     modificationEventsTable(new FsckDBModificationEventsTable(databasePath, fragmentSize,
+         allowCreate) ),
+     malformedChunks(databasePath + "/malformedChunks")
+{
+}
+
+void FsckDB::clear()
+{
+   this->dentryTable->clear();
+   this->fileInodesTable->clear();
+   this->dirInodesTable->clear();
+   this->chunksTable->clear();
+   this->contDirsTable->clear();
+   this->fsIDsTable->clear();
+   this->usedTargetIDsTable->clear();
+   this->modificationEventsTable->clear();
+   this->malformedChunks.clear();
+}

fsck/source/database/FsckDB.h

@@ -0,0 +1,178 @@
+/*
+ * This class is intended to be the interface to the sqlite DB. It can be created once in the
+ * applicationm, and can then be used in all threads, as sqlite claims to be thread-safe in
+ * "serialized" mode (which is the default, http://sqlite.org/threadsafe.html).
+ */
+
+#ifndef FSCKDB_H_
+#define FSCKDB_H_
+
+#include <common/app/log/LogContext.h>
+#include <common/Common.h>
+#include <common/fsck/FsckChunk.h>
+#include <common/fsck/FsckContDir.h>
+#include <common/fsck/FsckDirEntry.h>
+#include <common/fsck/FsckDirInode.h>
+#include <common/fsck/FsckFileInode.h>
+#include <common/fsck/FsckFsID.h>
+#include <common/fsck/FsckTargetID.h>
+#include <common/fsck/FsckDuplicateInodeInfo.h>
+#include <common/nodes/TargetMapper.h>
+#include <common/storage/striping/StripePattern.h>
+#include <common/storage/StorageDefinitions.h>
+#include <database/ContDir.h>
+#include <database/Cursor.h>
+#include <database/DirEntry.h>
+#include <database/DirInode.h>
+#include <database/DiskList.h>
+#include <database/EntryID.h>
+#include <database/FileInode.h>
+#include <database/Filter.h>
+#include <database/VectorSource.h>
+#include <toolkit/FsckDefinitions.h>
+
+#include <boost/scoped_ptr.hpp>
+
+class FsckDBDentryTable;
+class FsckDBFileInodesTable;
+class FsckDBDirInodesTable;
+class FsckDBChunksTable;
+class FsckDBContDirsTable;
+class FsckDBFsIDsTable;
+class FsckDBUsedTargetIDsTable;
+class FsckDBModificationEventsTable;
+class FsckDBDentryErrorTable;
+class FsckDBDirInodeErrorTable;
+class FsckDBFsIDErrorTable;
+class FsckDBFileInodeErrorTable;
+class FsckDBChunkErrorTable;
+class FsckDBContDirErrorTable;
+class FsckDBStripeTargetErrorTable;
+
+namespace checks {
+
+struct InodeAttribs
+{
+   uint64_t size;
+   uint64_t nlinks;
+};
+
+struct OptionalInodeAttribs
+{
+   boost::optional<uint64_t> size;
+   boost::optional<uint64_t> nlinks;
+
+   void reset() {
+       size.reset();
+       nlinks.reset();
+   }
+};
+
+typedef std::pair<db::EntryID, std::set<FsckDuplicateInodeInfo>> DuplicatedInode;
+}
+
+class FsckDB
+{
+   friend class TestDatabase;
+
+   public:
+      // FsckDB.cpp
+      FsckDB(const std::string& databasePath, size_t fragmentSize, size_t nameCacheLimit,
+         bool allowCreate);
+
+      void clear();
+
+      // FsckDBChecks.cpp
+      Cursor<checks::DuplicatedInode> findDuplicateInodeIDs();
+      Cursor<std::list<FsckChunk> > findDuplicateChunks();
+      Cursor<std::list<db::ContDir>> findDuplicateContDirs();
+
+      Cursor<db::DirEntry> findMismirroredDentries();
+      Cursor<db::DirInode> findMismirroredDirectories();
+      Cursor<db::FileInode> findMismirroredFiles();
+
+      Cursor<db::DirEntry> findDanglingDirEntries();
+      Cursor<db::DirEntry> findDirEntriesWithBrokenByIDFile();
+      Cursor<FsckFsID> findOrphanedFsIDFiles();
+      Cursor<FsckDirInode> findInodesWithWrongOwner();
+      Cursor<std::pair<db::DirEntry, NumNodeID> > findDirEntriesWithWrongOwner();
+      Cursor<FsckDirInode> findOrphanedDirInodes();
+      Cursor<FsckFileInode> findOrphanedFileInodes();
+      Cursor<FsckChunk> findOrphanedChunks();
+
+      Cursor<FsckDirInode> findInodesWithoutContDir();
+
+      Cursor<FsckContDir> findOrphanedContDirs();
+
+      Cursor<std::pair<FsckFileInode, checks::OptionalInodeAttribs> > findWrongInodeFileAttribs();
+
+      Cursor<std::pair<FsckDirInode, checks::InodeAttribs> > findWrongInodeDirAttribs();
+
+      Cursor<db::DirEntry> findFilesWithMissingStripeTargets(TargetMapper* targetMapper,
+         MirrorBuddyGroupMapper* buddyGroupMapper);
+      Cursor<std::pair<FsckChunk, FsckFileInode> > findChunksWithWrongPermissions();
+      Cursor<std::pair<FsckChunk, FsckFileInode> > findChunksInWrongPath();
+      Cursor<db::FileInode> findFilesWithMultipleHardlinks();
+
+   private:
+      LogContext log;
+      std::string databasePath;
+
+      boost::scoped_ptr<FsckDBDentryTable> dentryTable;
+      boost::scoped_ptr<FsckDBFileInodesTable> fileInodesTable;
+      boost::scoped_ptr<FsckDBDirInodesTable> dirInodesTable;
+      boost::scoped_ptr<FsckDBChunksTable> chunksTable;
+      boost::scoped_ptr<FsckDBContDirsTable> contDirsTable;
+      boost::scoped_ptr<FsckDBFsIDsTable> fsIDsTable;
+      boost::scoped_ptr<FsckDBUsedTargetIDsTable> usedTargetIDsTable;
+      boost::scoped_ptr<FsckDBModificationEventsTable> modificationEventsTable;
+      DiskList<FsckChunk> malformedChunks;
+
+   public:
+      FsckDBDentryTable* getDentryTable()
+      {
+         return this->dentryTable.get();
+      }
+
+      FsckDBFileInodesTable* getFileInodesTable()
+      {
+         return this->fileInodesTable.get();
+      }
+
+      FsckDBDirInodesTable* getDirInodesTable()
+      {
+         return this->dirInodesTable.get();
+      }
+
+      FsckDBChunksTable* getChunksTable()
+      {
+         return this->chunksTable.get();
+      }
+
+      FsckDBContDirsTable* getContDirsTable()
+      {
+         return this->contDirsTable.get();
+      }
+
+      FsckDBFsIDsTable* getFsIDsTable()
+      {
+         return this->fsIDsTable.get();
+      }
+
+      FsckDBUsedTargetIDsTable* getUsedTargetIDsTable()
+      {
+         return this->usedTargetIDsTable.get();
+      }
+
+      FsckDBModificationEventsTable* getModificationEventsTable()
+      {
+         return this->modificationEventsTable.get();
+      }
+
+      DiskList<FsckChunk>* getMalformedChunksList()
+      {
+         return &malformedChunks;
+      }
+};
+
+#endif /* FSCKDB_H_ */

fsck/source/database/FsckDBException.cpp

@@ -0,0 +1,2 @@
+#include "FsckDBException.h"
+

fsck/source/database/FsckDBException.h

@@ -0,0 +1,8 @@
+#ifndef FSCKDBEXCEPTION_H
+#define FSCKDBEXCEPTION_H
+
+#include <common/toolkit/NamedException.h>
+
+DECLARE_NAMEDEXCEPTION(FsckDBException, "FsckDBException")
+
+#endif /*FSCKDBEXCEPTION_H*/

fsck/source/database/FsckDBTable.cpp

@@ -0,0 +1,509 @@
+#include "FsckDBTable.h"
+
+#include <database/FsckDBException.h>
+#include <toolkit/FsckTkEx.h>
+
+#include <boost/make_shared.hpp>
+
+static db::DirEntry fsckDirEntryToDbDirEntry(const FsckDirEntry& dentry)
+{
+   db::DirEntry result = {
+      db::EntryID::fromStr(dentry.getID() ),
+      db::EntryID::fromStr(dentry.getParentDirID() ),
+
+      dentry.getEntryOwnerNodeID().val(),
+      dentry.getInodeOwnerNodeID().val(),
+
+      dentry.getSaveNodeID().val(), dentry.getSaveDevice(), dentry.getSaveInode(),
+
+      {},
+
+      dentry.getInternalID(),
+
+      dentry.getEntryType(),
+      dentry.getName().size() < sizeof(result.name.inlined.text),
+      dentry.getHasInlinedInode(),
+      dentry.getIsBuddyMirrored(),
+   };
+
+   if(result.fileNameInlined) {
+      ::memset(result.name.inlined.text, '\0', sizeof(result.name.inlined.text));
+      ::strncpy(result.name.inlined.text, dentry.getName().c_str(),
+            sizeof(result.name.inlined.text) - 1);
+   }
+
+   return result;
+}
+
+void FsckDBDentryTable::insert(FsckDirEntryList& dentries, const BulkHandle* handle)
+{
+   NameBuffer& names = handle ? *handle->nameBuffer : getNameBuffer(0);
+
+   for(FsckDirEntryListIter it = dentries.begin(), end = dentries.end(); it != end; ++it)
+   {
+      db::DirEntry dentry = fsckDirEntryToDbDirEntry(*it);
+      ByParent link = { dentry.parentDirID, dentry.id, dentry.entryType };
+
+      if(!dentry.fileNameInlined)
+      {
+         dentry.name.extended.fileID = names.id();
+         dentry.name.extended.fileOffset = names.put(it->getName() );
+      }
+
+      if(handle)
+      {
+         handle->dentries->append(dentry);
+         handle->byParent->append(link);
+      }
+      else
+      {
+         this->table.insert(dentry);
+         this->byParent.insert(link);
+      }
+   }
+}
+
+void FsckDBDentryTable::updateFieldsExceptParent(FsckDirEntryList& dentries)
+{
+   for(FsckDirEntryListIter it = dentries.begin(), end = dentries.end(); it != end; ++it)
+   {
+      db::DirEntry dentry = fsckDirEntryToDbDirEntry(*it);
+
+      this->table.remove(dentry.pkey() );
+      this->table.insert(dentry);
+   }
+}
+
+void FsckDBDentryTable::remove(FsckDirEntryList& dentries)
+{
+   for(FsckDirEntryListIter it = dentries.begin(), end = dentries.end(); it != end; ++it)
+      this->table.remove(fsckDirEntryToDbDirEntry(*it).pkey() );
+}
+
+Table<db::DirEntry>::QueryType FsckDBDentryTable::get()
+{
+   this->table.commitChanges();
+   return this->table.cursor();
+}
+
+Table<FsckDBDentryTable::ByParent>::QueryType FsckDBDentryTable::getByParent()
+{
+   this->byParent.commitChanges();
+   return this->byParent.cursor();
+}
+
+std::pair<bool, db::DirEntry> FsckDBDentryTable::getAnyFor(db::EntryID id)
+{
+   this->table.commitChanges();
+
+   struct ops
+   {
+      static db::EntryID onlyID(const db::DirEntry::KeyType& key) { return boost::get<0>(key); }
+   };
+
+   return this->table.getByKeyProjection(id, ops::onlyID);
+}
+
+std::string FsckDBDentryTable::getNameOf(const db::DirEntry& dentry)
+{
+   if(dentry.fileNameInlined)
+      return dentry.name.inlined.text;
+
+   return getNameBuffer(dentry.name.extended.fileID).get(dentry.name.extended.fileOffset);
+}
+
+std::string FsckDBDentryTable::getPathOf(const db::DirEntry& dentry)
+{
+   std::string result = getNameOf(dentry);
+
+   db::EntryID parent = dentry.parentDirID;
+   unsigned depth = 0;
+
+   while(!parent.isSpecial() )
+   {
+      depth++;
+      if(depth > 255)
+         return "[<unresolved>]/" + result;
+
+      const NameCacheEntry* cacheEntry = getFromCache(parent);
+      if(cacheEntry)
+      {
+         result = cacheEntry->name + "/" + result;
+         parent = cacheEntry->parent;
+         continue;
+      }
+
+      std::pair<bool, db::DirEntry> item = getAnyFor(parent);
+      if(!item.first)
+         return "[<unresolved>]/" + result;
+
+      std::string parentName = getNameOf(item.second);
+
+      result = parentName + "/" + result;
+      parent = item.second.parentDirID;
+
+      addToCache(item.second, parentName);
+   }
+
+   return getNameOf(parent) + "/" + result;
+}
+
+
+
+void FsckDBFileInodesTable::insert(FsckFileInodeList& fileInodes, const BulkHandle* handle)
+{
+   for(FsckFileInodeListIter it = fileInodes.begin(), end = fileInodes.end(); it != end; ++it)
+   {
+      const std::vector<uint16_t>& stripes = it->getStripeTargets();
+
+      db::FileInode inode = {
+         db::EntryID::fromStr(it->getID() ),
+         db::EntryID::fromStr(it->getParentDirID() ),
+         db::EntryID::fromStr(it->getPathInfo()->getOrigParentEntryID() ),
+
+         it->getParentNodeID().val(),
+         it->getSaveNodeID().val(),
+         it->getPathInfo()->getOrigUID(),
+
+         it->getUserID(), it->getGroupID(),
+
+         (uint64_t) it->getFileSize(), it->getUsedBlocks(),
+
+         it->getNumHardLinks(),
+
+         it->getSaveInode(),
+         it->getSaveDevice(),
+
+         it->getChunkSize(),
+
+         {},
+
+         it->getIsInlined(),
+         uint32_t(it->getPathInfo()->getFlags() ),
+         it->getStripePatternType(),
+         uint32_t(stripes.size() ),
+         it->getReadable(),
+         it->getIsBuddyMirrored(),
+         it->getIsMismirrored(),
+      };
+
+      db::StripeTargets extraTargets = { inode.id, {}, 0 };
+
+      for(size_t i = 0; i < stripes.size(); i++)
+      {
+         if(i < inode.NTARGETS)
+            inode.targets[i] = stripes[i];
+         else
+         {
+            size_t offsetInPattern = (i - inode.NTARGETS) % extraTargets.NTARGETS;
+
+            if(offsetInPattern == 0)
+               extraTargets.firstTargetIndex = i;
+
+            extraTargets.targets[offsetInPattern] = stripes[i];
+
+            if(offsetInPattern == extraTargets.NTARGETS - 1)
+            {
+               if(handle)
+                  std::get<1>(*handle)->append(extraTargets);
+               else
+                  this->targets.insert(extraTargets);
+
+               extraTargets.firstTargetIndex = 0;
+            }
+         }
+      }
+
+      if(handle)
+      {
+         std::get<0>(*handle)->append(inode);
+         if(extraTargets.firstTargetIndex != 0)
+            std::get<1>(*handle)->append(extraTargets);
+      }
+      else
+      {
+         this->inodes.insert(inode);
+         if(extraTargets.firstTargetIndex != 0)
+            this->targets.insert(extraTargets);
+      }
+   }
+}
+
+void FsckDBFileInodesTable::update(FsckFileInodeList& inodes)
+{
+   for(FsckFileInodeListIter it = inodes.begin(), end = inodes.end(); it != end; ++it)
+   {
+      FsckFileInodeList list(1, *it);
+
+      remove(list);
+      insert(list);
+   }
+}
+
+void FsckDBFileInodesTable::remove(FsckFileInodeList& fileInodes)
+{
+   for(FsckFileInodeListIter it = fileInodes.begin(), end = fileInodes.end(); it != end; ++it)
+   {
+      this->inodes.remove(db::EntryID::fromStr(it->getID() ) );
+      this->targets.remove(db::EntryID::fromStr(it->getID() ) );
+   }
+}
+
+FsckDBFileInodesTable::InodesQueryType FsckDBFileInodesTable::getInodes()
+{
+   this->inodes.commitChanges();
+   return
+      this->inodes.cursor()
+      | db::groupBy(UniqueInlinedInode() )
+      | db::select(SelectFirst() );
+}
+
+Table<db::StripeTargets, true>::QueryType FsckDBFileInodesTable::getTargets()
+{
+   this->targets.commitChanges();
+   return this->targets.cursor();
+}
+
+UInt16Vector FsckDBFileInodesTable::getStripeTargetsByKey(const db::EntryID& id)
+{
+   UInt16Vector allStripeTargets;
+   std::pair<bool, db::FileInode> result = this->get(id.str());
+
+   if (!result.first)
+      return allStripeTargets;
+
+   // firstly copy stripeTargets from FileInode object
+   auto inode = result.second;
+   uint32_t targetArrSize = inode.NTARGETS;
+   uint32_t numTargets = (inode.stripePatternSize > targetArrSize) ? targetArrSize
+      : inode.stripePatternSize;
+
+   std::copy(inode.targets, inode.targets + numTargets, std::back_inserter(allStripeTargets));
+
+   // if extraTargets are present then get them from targets table
+   uint32_t numExtraTargets = inode.stripePatternSize - numTargets;
+   if (numExtraTargets)
+   {
+      this->targets.commitChanges();
+      auto dbRes = this->targets.getAllByKey(id);
+
+      while (dbRes.step())
+      {
+         auto elem = dbRes.get();
+
+         for (int i=0; i<elem->NTARGETS; i++)
+         {
+            allStripeTargets.push_back(elem->targets[i]);
+            numExtraTargets -= 1;
+            if (numExtraTargets == 0) break;
+         }
+
+         if (numExtraTargets == 0) break;
+      }
+   }
+
+   return allStripeTargets;
+}
+
+std::pair<bool, db::FileInode> FsckDBFileInodesTable::get(std::string id)
+{
+   this->inodes.commitChanges();
+   return this->inodes.getByKey(db::EntryID::fromStr(id) );
+}
+
+
+
+void FsckDBDirInodesTable::insert(FsckDirInodeList& dirInodes, const BulkHandle* handle)
+{
+   for(FsckDirInodeListIter it = dirInodes.begin(), end = dirInodes.end(); it != end; ++it)
+   {
+      db::DirInode inode = {
+         db::EntryID::fromStr(it->getID() ),
+         db::EntryID::fromStr(it->getParentDirID() ),
+
+         it->getParentNodeID().val(),
+         it->getOwnerNodeID().val(),
+         it->getSaveNodeID().val(),
+
+         it->getStripePatternType(),
+         it->getReadable(),
+         it->getIsBuddyMirrored(),
+         it->getIsMismirrored(),
+
+         (uint64_t) it->getSize(),
+         it->getNumHardLinks(),
+      };
+
+      if(handle)
+         (*handle)->append(inode);
+      else
+         this->table.insert(inode);
+   }
+}
+
+void FsckDBDirInodesTable::update(FsckDirInodeList& inodes)
+{
+   for(FsckDirInodeListIter it = inodes.begin(), end = inodes.end(); it != end; ++it)
+   {
+      FsckDirInodeList list(1, *it);
+
+      remove(list);
+      insert(list);
+   }
+}
+
+void FsckDBDirInodesTable::remove(FsckDirInodeList& dirInodes)
+{
+   for(FsckDirInodeListIter it = dirInodes.begin(), end = dirInodes.end(); it != end; ++it)
+      this->table.remove(db::EntryID::fromStr(it->getID() ) );
+}
+
+Table<db::DirInode>::QueryType FsckDBDirInodesTable::get()
+{
+   this->table.commitChanges();
+   return this->table.cursor();
+}
+
+std::pair<bool, FsckDirInode> FsckDBDirInodesTable::get(std::string id)
+{
+   this->table.commitChanges();
+   return this->table.getByKey(db::EntryID::fromStr(id) );
+}
+
+
+
+static db::Chunk fsckChunkToDbChunk(FsckChunk& chunk)
+{
+   db::Chunk result = {
+      db::EntryID::fromStr(chunk.getID() ),
+      chunk.getTargetID(), chunk.getBuddyGroupID(),
+      {},
+      chunk.getFileSize(), chunk.getUsedBlocks(),
+      chunk.getUserID(), chunk.getGroupID(),
+   };
+
+   strncpy(result.savedPath, chunk.getSavedPath()->str().c_str(),
+      sizeof(result.savedPath) - 1);
+
+   return result;
+}
+
+void FsckDBChunksTable::insert(FsckChunkList& chunks, const BulkHandle* handle)
+{
+   for(FsckChunkListIter it = chunks.begin(), end = chunks.end(); it != end; ++it)
+   {
+      if(handle)
+         (*handle)->append(fsckChunkToDbChunk(*it) );
+      else
+         this->table.insert(fsckChunkToDbChunk(*it) );
+   }
+}
+
+void FsckDBChunksTable::update(FsckChunkList& chunks)
+{
+   for(FsckChunkListIter it = chunks.begin(), end = chunks.end(); it != end; ++it)
+   {
+      db::Chunk chunk = fsckChunkToDbChunk(*it);
+
+      this->table.remove(chunk.pkey() );
+      this->table.insert(chunk);
+   }
+}
+
+void FsckDBChunksTable::remove(const db::Chunk::KeyType& id)
+{
+   this->table.remove(id);
+}
+
+Table<db::Chunk>::QueryType FsckDBChunksTable::get()
+{
+   this->table.commitChanges();
+   return this->table.cursor();
+}
+
+
+
+static db::ContDir fsckContDirToDbContDir(FsckContDir& contDir)
+{
+   return {
+      db::EntryID::fromStr(contDir.getID()),
+      contDir.getSaveNodeID().val(),
+      contDir.getIsBuddyMirrored(),
+   };
+}
+
+void FsckDBContDirsTable::insert(FsckContDirList& contDirs, const BulkHandle* handle)
+{
+   for(FsckContDirListIter it = contDirs.begin(), end = contDirs.end(); it != end; ++it)
+   {
+      if(handle)
+         (*handle)->append(fsckContDirToDbContDir(*it) );
+      else
+         this->table.insert(fsckContDirToDbContDir(*it) );
+   }
+}
+
+Table<db::ContDir>::QueryType FsckDBContDirsTable::get()
+{
+   this->table.commitChanges();
+   return this->table.cursor();
+}
+
+
+
+static db::FsID fsckFsIDToDbFsID(FsckFsID& id)
+{
+   return {
+      db::EntryID::fromStr(id.getID() ),
+      db::EntryID::fromStr(id.getParentDirID() ),
+      id.getSaveNodeID().val(), id.getSaveDevice(), id.getSaveInode(),
+      id.getIsBuddyMirrored(),
+   };
+}
+
+void FsckDBFsIDsTable::insert(FsckFsIDList& fsIDs, const BulkHandle* handle)
+{
+   for(FsckFsIDListIter it = fsIDs.begin(), end = fsIDs.end(); it != end; ++it)
+   {
+      if(handle)
+         (*handle)->append(fsckFsIDToDbFsID(*it) );
+      else
+         this->table.insert(fsckFsIDToDbFsID(*it) );
+   }
+}
+
+void FsckDBFsIDsTable::remove(FsckFsIDList& fsIDs)
+{
+   for(FsckFsIDListIter it = fsIDs.begin(), end = fsIDs.end(); it != end; ++it)
+      this->table.remove(fsckFsIDToDbFsID(*it).pkey() );
+}
+
+Table<db::FsID>::QueryType FsckDBFsIDsTable::get()
+{
+   this->table.commitChanges();
+   return this->table.cursor();
+}
+
+
+
+void FsckDBUsedTargetIDsTable::insert(FsckTargetIDList& targetIDs, const BulkHandle& handle)
+{
+   for(FsckTargetIDListIter it = targetIDs.begin(), end = targetIDs.end(); it != end; ++it)
+   {
+      db::UsedTarget target = { it->getID(), it->getTargetIDType() };
+      handle->append(target);
+   }
+}
+
+
+
+void FsckDBModificationEventsTable::insert(const FsckModificationEventList& events,
+   const BulkHandle& handle)
+{
+   for(std::list<FsckModificationEvent>::const_iterator it = events.begin(), end = events.end();
+         it != end; ++it)
+   {
+      db::ModificationEvent event = { db::EntryID::fromStr(it->getEntryID() ) };
+      handle->append(event);
+   }
+}

fsck/source/database/FsckDBTable.h

@@ -0,0 +1,641 @@
+#ifndef FsckDBTable_h_drBbZVcUQetDB8UVnMd1Uj
+#define FsckDBTable_h_drBbZVcUQetDB8UVnMd1Uj
+
+#include <common/fsck/FsckModificationEvent.h>
+#include <database/Chunk.h>
+#include <database/ContDir.h>
+#include <database/DirEntry.h>
+#include <database/DirInode.h>
+#include <database/FileInode.h>
+#include <database/FsID.h>
+#include <database/Group.h>
+#include <database/Table.h>
+#include <database/ModificationEvent.h>
+#include <database/StripeTargets.h>
+#include <database/UsedTarget.h>
+
+class FsckDBDentryTable
+{
+   public:
+      struct ByParent
+      {
+         db::EntryID parent; /* 12 */
+         db::EntryID target; /* 24 */
+         uint32_t entryType; /* 28 */
+
+         typedef std::pair<db::EntryID, db::EntryID> KeyType;
+
+         KeyType pkey() const { return KeyType(parent, target); }
+      };
+
+      class NameBuffer
+      {
+         public:
+            NameBuffer(const std::string& path, unsigned id):
+               fileID(id)
+            {
+               fd = ::open(path.c_str(), O_RDWR | O_CREAT, 0660);
+               if(fd < 0)
+                  throw std::runtime_error("could not open name file");
+
+               streamBuf.reserve(4096 * 1024);
+
+               writePos = ::lseek(fd, 0, SEEK_END);
+               if (writePos < 0)
+               {
+                  close(fd);
+                  fd = -1;
+                  throw std::runtime_error("could not open name file");
+               }
+            }
+
+            NameBuffer(NameBuffer&&) = delete;
+            NameBuffer& operator=(NameBuffer&&) = delete;
+
+            ~NameBuffer()
+            {
+               if (fd >= 0)
+               {
+                  flush();
+                  close(fd);
+               }
+            }
+
+            uint64_t put(const std::string& name)
+            {
+               if (streamBuf.size() + name.size() + 1 > streamBuf.capacity())
+                  flush();
+
+               if (name.size() > streamBuf.capacity())
+                  streamBuf.reserve(name.size());
+
+               uint64_t result = writePos + streamBuf.size();
+               streamBuf.insert(streamBuf.end(), name.begin(), name.end());
+               streamBuf.push_back(0);
+
+               return result;
+            }
+
+            std::string get(uint64_t offset)
+            {
+               flush();
+
+               char buffer[255 + 1];
+               ssize_t readRes = ::pread(fd, buffer, sizeof(buffer), offset);
+               // if the read did not fill the full buffer, we can
+               //   a) have an error, or
+               //   b) have a short read.
+               // errors should be reported directly. short reads are likely caused be trying to
+               // read past the end of the file. since we cannot possible have a sane read that
+               // stretches beyond writePos here (as we have flushed the write buffer), we can
+               // reliably detect both.
+               if (readRes < 0 || ssize_t(offset) + readRes > writePos)
+                  throw std::runtime_error("could not read name file: " + std::string(strerror(errno)));
+
+               return buffer;
+            }
+
+            unsigned id() const { return fileID; }
+
+         private:
+            unsigned fileID;
+            std::vector<char> streamBuf;
+
+            int fd;
+            ssize_t writePos;
+
+            void flush()
+            {
+               ssize_t writeRes = ::pwrite(fd, &streamBuf[0], streamBuf.size(), writePos);
+               if (writeRes < 0 || size_t(writeRes) < streamBuf.size())
+                  throw std::runtime_error("error in flush: " + std::string(strerror(errno)));
+
+               writePos += writeRes;
+               streamBuf.resize(0);
+            }
+      };
+
+      struct BulkHandle
+      {
+         boost::shared_ptr<Buffer<db::DirEntry> > dentries;
+         boost::shared_ptr<Buffer<ByParent> > byParent;
+         NameBuffer* nameBuffer;
+      };
+
+   private:
+      struct NameCacheEntry
+      {
+         db::EntryID id;
+         db::EntryID parent;
+         std::string name;
+
+         std::list<NameCacheEntry*>::iterator lruLink;
+      };
+
+   public:
+      FsckDBDentryTable(const std::string& dbPath, size_t fragmentSize, size_t nameCacheLimit,
+            bool allowCreate)
+         : dbPath(dbPath),
+           table(dbPath + "/dentries", fragmentSize, allowCreate),
+           byParent(dbPath + "/dentriesbyparent", fragmentSize, allowCreate),
+           insertSeqNo(0),
+           nameCacheLimit(nameCacheLimit)
+      {
+      }
+
+      void insert(FsckDirEntryList& dentries, const BulkHandle* handle = NULL);
+      void updateFieldsExceptParent(FsckDirEntryList& dentries);
+      void remove(FsckDirEntryList& dentries);
+
+      Table<db::DirEntry>::QueryType get();
+      std::pair<bool, db::DirEntry> getAnyFor(db::EntryID id);
+
+      Table<ByParent>::QueryType getByParent();
+
+      std::string getNameOf(const db::DirEntry& dentry);
+      std::string getPathOf(const db::DirEntry& dentry);
+
+   private:
+      std::string dbPath;
+      Table<db::DirEntry> table;
+      Table<ByParent> byParent;
+
+      std::map<uint64_t, boost::shared_ptr<NameBuffer> > nameBuffers;
+      uint64_t insertSeqNo;
+
+      std::map<db::EntryID, NameCacheEntry> nameCache;
+      std::list<NameCacheEntry*> nameCacheLRU;
+      size_t nameCacheLimit;
+
+      NameBuffer& getNameBuffer(unsigned id)
+      {
+         if(this->nameBuffers.count(id) )
+            return *this->nameBuffers[id];
+
+         const std::string& nextNameFile = dbPath + "/dentrynames."
+            + StringTk::intToStr(id);
+
+         boost::shared_ptr<NameBuffer> buf = boost::make_shared<NameBuffer>(
+            nextNameFile, id);
+         this->nameBuffers[id] = buf;
+         return *buf;
+      }
+
+      void addToCache(const db::DirEntry& entry, const std::string& name)
+      {
+         if(nameCache.size() >= nameCacheLimit)
+         {
+            nameCache.erase(nameCacheLRU.front()->id);
+            nameCacheLRU.pop_front();
+         }
+
+         NameCacheEntry cacheEntry = { entry.id, entry.parentDirID, name, nameCacheLRU.end() };
+         std::pair<std::map<db::EntryID, NameCacheEntry>::iterator, bool> pos = nameCache.insert(
+            std::make_pair(entry.id, cacheEntry) );
+
+         if(!pos.second)
+            return;
+
+         pos.first->second.lruLink = nameCacheLRU.insert(nameCacheLRU.end(), &pos.first->second);
+      }
+
+      const NameCacheEntry* getFromCache(db::EntryID id)
+      {
+         std::map<db::EntryID, NameCacheEntry>::iterator pos = nameCache.find(id);
+         if(pos == nameCache.end() )
+            return NULL;
+
+         nameCacheLRU.splice(nameCacheLRU.end(), nameCacheLRU, pos->second.lruLink);
+         return &pos->second;
+      }
+
+   public:
+      void clear()
+      {
+         this->table.clear();
+         this->byParent.clear();
+         this->insertSeqNo = 0;
+         this->nameCache.clear();
+         this->nameCacheLRU.clear();
+      }
+
+      BulkHandle newBulkHandle()
+      {
+         BulkHandle result = {
+            this->table.bulkInsert(),
+            this->byParent.bulkInsert(),
+            &getNameBuffer(this->nameBuffers.size() ),
+         };
+         return result;
+      }
+
+      void flush(BulkHandle& handle)
+      {
+         handle = {{}, {}, nullptr};
+      }
+
+      void insert(FsckDirEntryList& dentries, const BulkHandle& handle)
+      {
+         insert(dentries, &handle);
+      }
+
+      std::string getNameOf(db::EntryID id)
+      {
+         if(id == db::EntryID::anchor() || id == db::EntryID::root() )
+            return "";
+         else
+         if(id == db::EntryID::disposal() )
+            return "[<disposal>]";
+         else
+         if(id == db::EntryID::mdisposal() )
+            return "[<mdisposal>]";
+
+         std::pair<bool, db::DirEntry> item = getAnyFor(id);
+         if(!item.first)
+            return "[<unresolved>]";
+         else
+            return getNameOf(item.second);
+      }
+
+      std::string getPathOf(db::EntryID id)
+      {
+         std::pair<bool, db::DirEntry> item = getAnyFor(id);
+         if(!item.first)
+            return "[<unresolved>]";
+         else
+            return getPathOf(item.second);
+      }
+
+      void commitChanges()
+      {
+         table.commitChanges();
+      }
+};
+
+class FsckDBFileInodesTable
+{
+   public:
+      typedef std::tuple<
+         boost::shared_ptr<Buffer<db::FileInode> >,
+         boost::shared_ptr<Buffer<db::StripeTargets> > > BulkHandle;
+
+      struct UniqueInlinedInode
+      {
+         typedef boost::tuple<db::EntryID, uint64_t, uint32_t, int32_t> KeyType;
+         typedef db::FileInode ProjType;
+         typedef int GroupType;
+
+         KeyType key(const db::FileInode& fi)
+         {
+            return boost::make_tuple(fi.id, fi.saveInode, fi.saveNodeID, fi.saveDevice);
+         }
+
+         db::FileInode project(const db::FileInode& fi)
+         {
+            return fi;
+         }
+
+         void step(const db::FileInode& fi) {}
+
+         GroupType finish()
+         {
+            return 0;
+         }
+      };
+
+      struct SelectFirst
+      {
+         typedef db::FileInode result_type;
+
+         db::FileInode operator()(std::pair<db::FileInode, int>& pair) const
+         {
+            return pair.first;
+         }
+      };
+
+      typedef
+         Select<
+            Group<
+               Table<db::FileInode>::QueryType,
+               UniqueInlinedInode>,
+            SelectFirst> InodesQueryType;
+
+   public:
+      FsckDBFileInodesTable(const std::string& dbPath, size_t fragmentSize, bool allowCreate)
+         : inodes(dbPath + "/fileinodes", fragmentSize, allowCreate),
+           targets(dbPath + "/filetargets", fragmentSize, allowCreate)
+      {
+      }
+
+      void insert(FsckFileInodeList& fileInodes, const BulkHandle* handle = NULL);
+      void update(FsckFileInodeList& inodes);
+      void remove(FsckFileInodeList& fileInodes);
+
+      InodesQueryType getInodes();
+      Table<db::StripeTargets, true>::QueryType getTargets();
+      UInt16Vector getStripeTargetsByKey(const db::EntryID& id);
+
+      std::pair<bool, db::FileInode> get(std::string id);
+
+   private:
+      Table<db::FileInode> inodes;
+      Table<db::StripeTargets, true> targets;
+
+   public:
+      void clear()
+      {
+         inodes.clear();
+         targets.clear();
+      }
+
+      void remove(db::EntryID id)
+      {
+         inodes.remove(id);
+         targets.remove(id);
+      }
+
+      BulkHandle newBulkHandle()
+      {
+         return BulkHandle(this->inodes.bulkInsert(), this->targets.bulkInsert() );
+      }
+
+      void flush(BulkHandle& handle)
+      {
+         handle = {};
+      }
+
+      void insert(FsckFileInodeList& fileInodes, const BulkHandle& handle)
+      {
+         insert(fileInodes, &handle);
+      }
+
+      void commitChanges()
+      {
+         inodes.commitChanges();
+         targets.commitChanges();
+      }
+};
+
+class FsckDBDirInodesTable
+{
+   public:
+      typedef boost::shared_ptr<Buffer<db::DirInode> > BulkHandle;
+
+   public:
+      FsckDBDirInodesTable(const std::string& dbPath, size_t fragmentSize, bool allowCreate)
+         : table(dbPath + "/dirinodes", fragmentSize, allowCreate)
+      {
+      }
+
+      void insert(FsckDirInodeList& dirInodes, const BulkHandle* handle = NULL);
+      void update(FsckDirInodeList& inodes);
+      void remove(FsckDirInodeList& dirInodes);
+
+      Table<db::DirInode>::QueryType get();
+      std::pair<bool, FsckDirInode> get(std::string id);
+
+   private:
+      Table<db::DirInode> table;
+
+   public:
+      void clear()
+      {
+         this->table.clear();
+      }
+
+      BulkHandle newBulkHandle()
+      {
+         return this->table.bulkInsert();
+      }
+
+      void flush(BulkHandle& handle)
+      {
+         handle = {};
+      }
+
+      void insert(FsckDirInodeList& fileInodes, const BulkHandle& handle)
+      {
+         insert(fileInodes, &handle);
+      }
+
+      void commitChanges()
+      {
+         table.commitChanges();
+      }
+};
+
+class FsckDBChunksTable
+{
+   public:
+      typedef boost::shared_ptr<Buffer<db::Chunk> > BulkHandle;
+
+   public:
+      FsckDBChunksTable(const std::string& dbPath, size_t fragmentSize, bool allowCreate)
+         : table(dbPath + "/chunks", fragmentSize, allowCreate)
+      {
+      }
+
+      void insert(FsckChunkList& chunks, const BulkHandle* handle = NULL);
+      void update(FsckChunkList& chunks);
+      void remove(const db::Chunk::KeyType& id);
+
+      Table<db::Chunk>::QueryType get();
+
+   private:
+      Table<db::Chunk> table;
+
+   public:
+      void clear()
+      {
+         table.clear();
+      }
+
+      BulkHandle newBulkHandle()
+      {
+         return this->table.bulkInsert();
+      }
+
+      void flush(BulkHandle& handle)
+      {
+         handle = {};
+      }
+
+      void insert(FsckChunkList& chunks, const BulkHandle& handle)
+      {
+         insert(chunks, &handle);
+      }
+
+      void commitChanges()
+      {
+         table.commitChanges();
+      }
+};
+
+class FsckDBContDirsTable
+{
+   public:
+      typedef boost::shared_ptr<Buffer<db::ContDir> > BulkHandle;
+
+   public:
+      FsckDBContDirsTable(const std::string& dbPath, size_t fragmentSize, bool allowCreate)
+         : table(dbPath + "/contdirs", fragmentSize, allowCreate)
+      {
+      }
+
+      void insert(FsckContDirList& contDirs, const BulkHandle* handle = NULL);
+
+      Table<db::ContDir>::QueryType get();
+
+   private:
+      Table<db::ContDir> table;
+
+   public:
+      void clear()
+      {
+         table.clear();
+      }
+
+      BulkHandle newBulkHandle()
+      {
+         return this->table.bulkInsert();
+      }
+
+      void flush(BulkHandle& handle)
+      {
+         handle = {};
+      }
+
+      void insert(FsckContDirList& contDirs, const BulkHandle& handle)
+      {
+         insert(contDirs, &handle);
+      }
+
+      void commitChanges()
+      {
+         table.commitChanges();
+      }
+};
+
+class FsckDBFsIDsTable
+{
+   public:
+      typedef boost::shared_ptr<Buffer<db::FsID> > BulkHandle;
+
+   public:
+      FsckDBFsIDsTable(const std::string& dbPath, size_t fragmentSize, bool allowCreate)
+         : table(dbPath + "/fsids", fragmentSize, allowCreate)
+      {
+      }
+
+      void insert(FsckFsIDList& fsIDs, const BulkHandle* handle = NULL);
+      void remove(FsckFsIDList& fsIDs);
+
+      Table<db::FsID>::QueryType get();
+
+   private:
+      Table<db::FsID> table;
+
+   public:
+      void clear()
+      {
+         this->table.clear();
+      }
+
+      BulkHandle newBulkHandle()
+      {
+         return this->table.bulkInsert();
+      }
+
+      void flush(BulkHandle& handle)
+      {
+         handle = {};
+      }
+
+      void insert(FsckFsIDList& fsIDs, const BulkHandle& handle)
+      {
+         insert(fsIDs, &handle);
+      }
+
+      void commitChanges()
+      {
+         table.commitChanges();
+      }
+};
+
+template<typename Data>
+class FsckUniqueTable
+{
+   public:
+      typedef boost::shared_ptr<Buffer<Data> > BulkHandle;
+
+   protected:
+      FsckUniqueTable(const std::string& name, size_t bufferSize, bool allowCreate)
+         : set(name, allowCreate), bufferSize(bufferSize)
+      {}
+
+   private:
+      Set<Data> set;
+      size_t bufferSize;
+      std::vector<boost::weak_ptr<Buffer<Data> > > bulkHandles;
+
+   public:
+      void clear()
+      {
+         set.clear();
+         bulkHandles.clear();
+      }
+
+      BulkHandle newBulkHandle()
+      {
+         BulkHandle result(new Buffer<Data>(set, bufferSize) );
+         bulkHandles.push_back(result);
+         return result;
+      }
+
+      void flush(BulkHandle& handle)
+      {
+         handle = {};
+      }
+
+      SetFragmentCursor<Data> get()
+      {
+         if(!bulkHandles.empty() )
+         {
+            for(size_t i = 0; i < bulkHandles.size(); i++)
+            {
+               if(!bulkHandles[i].expired() )
+                  throw std::runtime_error("still has open bulk handles");
+            }
+
+            bulkHandles.clear();
+            set.makeUnique();
+         }
+
+         return set.cursor();
+      }
+};
+
+class FsckDBUsedTargetIDsTable : public FsckUniqueTable<db::UsedTarget>
+{
+   public:
+      FsckDBUsedTargetIDsTable(const std::string& dbPath, size_t fragmentSize, bool allowCreate)
+         : FsckUniqueTable<db::UsedTarget>(dbPath + "/usedtargets", fragmentSize, allowCreate)
+      {
+      }
+
+      void insert(FsckTargetIDList& targetIDs, const BulkHandle& handle);
+};
+
+class FsckDBModificationEventsTable : public FsckUniqueTable<db::ModificationEvent>
+{
+   public:
+      FsckDBModificationEventsTable(const std::string& dbPath, size_t fragmentSize,
+         bool allowCreate)
+         : FsckUniqueTable<db::ModificationEvent>(dbPath + "/modevents", fragmentSize, allowCreate)
+      {
+      }
+
+      void insert(const FsckModificationEventList& events, const BulkHandle& handle);
+};
+
+#endif

fsck/source/database/Group.h

@@ -0,0 +1,131 @@
+#ifndef GROUP_H_
+#define GROUP_H_
+
+#include <boost/utility/result_of.hpp>
+
+#include <utility>
+
+/*
+ * requires Ops : struct {
+ *    typedef <key-type> KeyType;
+ *    typedef <proj-type> ProjType;
+ *    typedef <group-type> GroupType;
+ *
+ *    KeyType key(Source::ElementType&);
+ *    ProjType project(Source::ElementType&);
+ *    void step(Source::ElementType&);
+ *
+ *    GroupType finish();
+ * }
+ *
+ * yields sequence (Source::ElementType, <row-type>)
+ */
+template<typename Source, typename Ops>
+class Group
+{
+   private:
+      typedef typename Ops::KeyType KeyType;
+      typedef typename Ops::ProjType ProjType;
+      typedef typename Ops::GroupType GroupType;
+
+   public:
+      typedef std::pair<ProjType, GroupType> ElementType;
+      typedef typename Source::MarkerType MarkerType;
+
+   public:
+      Group(Source source, Ops ops)
+         : source(source), ops(ops), sourceIsActive(false)
+      {
+      }
+
+      bool step()
+      {
+         if(!this->sourceIsActive && !this->source.step() )
+            return false;
+
+         groupCurrentSet();
+         return true;
+      }
+
+      ElementType* get()
+      {
+         return &current;
+      }
+
+      MarkerType mark() const
+      {
+          return this->currentMark;
+      }
+
+      void restore(MarkerType mark)
+      {
+         this->source.restore(mark);
+         groupCurrentSet();
+      }
+
+   private:
+      void groupCurrentSet()
+      {
+         typename Source::ElementType* row = this->source.get();
+
+         this->currentMark = this->source.mark();
+         this->current.first = this->ops.project(*row);
+         this->ops.step(*row);
+
+         KeyType currentKey = this->ops.key(*row);
+
+         while(true)
+         {
+            if(!this->source.step() )
+            {
+               this->sourceIsActive = false;
+               break;
+            }
+
+            row = this->source.get();
+
+            if(!(currentKey == this->ops.key(*row) ) )
+            {
+               this->sourceIsActive = true;
+               break;
+            }
+
+            this->ops.step(*row);
+         }
+
+         this->current.second = this->ops.finish();
+      }
+
+   private:
+      Source source;
+      Ops ops;
+      ElementType current;
+      typename Source::MarkerType currentMark;
+      bool sourceIsActive;
+};
+
+
+namespace db {
+
+template<typename Ops>
+struct GroupOp
+{
+   Ops ops;
+
+   template<typename Source>
+   friend Group<Source, Ops> operator|(Source source, const GroupOp& fn)
+   {
+      return Group<Source, Ops>(source, fn.ops);
+   }
+};
+
+template<typename Ops>
+inline GroupOp<Ops> groupBy(Ops ops)
+{
+   GroupOp<Ops> result = {ops};
+   return result;
+}
+
+}
+
+#endif

fsck/source/database/LeftJoinEq.h

@@ -0,0 +1,189 @@
+#ifndef LEFTJOINEQ_H_
+#define LEFTJOINEQ_H_
+
+#include <common/Common.h>
+
+#include <cstddef>
+#include <utility>
+
+#include <boost/type_traits/decay.hpp>
+#include <boost/utility/result_of.hpp>
+
+template<typename Left, typename Right, typename KeyExtract>
+class LeftJoinEq
+{
+   private:
+      enum State {
+         s_next_left,
+         s_first_right,
+         s_next_right,
+         s_only_left,
+      };
+
+      typedef typename boost::decay<
+            typename boost::result_of<KeyExtract(typename Left::ElementType&)>::type
+         >::type LeftKey;
+      typedef typename boost::decay<
+            typename boost::result_of<KeyExtract(typename Right::ElementType&)>::type
+         >::type RightKey;
+
+   public:
+      typedef std::pair<typename Left::ElementType, typename Right::ElementType*> ElementType;
+
+      struct MarkerType
+      {
+         typename Left::MarkerType leftMark;
+         typename Right::MarkerType rightMark;
+         State state;
+
+         bool hasRightMark;
+         typename Right::MarkerType innerRightMark;
+      };
+
+   public:
+      LeftJoinEq(Left left, Right right, KeyExtract keyExtract)
+         : left(left), right(right), keyExtract(keyExtract), state(s_next_left),
+           hasRightMark(false)
+      {
+         if(!this->right.step() )
+            this->state = s_only_left;
+      }
+
+      bool step()
+      {
+         switch(this->state)
+         {
+            case s_only_left:
+               if(!this->left.step() )
+                  return false;
+
+               return makeCurrent(false);
+
+            case s_next_left:
+               if(!this->left.step() )
+               {
+                  this->state = s_only_left;
+                  return false;
+               }
+
+               if(this->hasRightMark &&
+                     this->keyExtract(*this->left.get() ) == this->rightKeyAtMark)
+                  this->right.restore(this->rightMark);
+               else
+                  this->hasRightMark = false;
+
+               BEEGFS_FALLTHROUGH;
+
+            case s_first_right: {
+               while(rightKey() < leftKey() )
+               {
+                  if(!this->right.step() )
+                  {
+                     this->state = s_only_left;
+                     return makeCurrent(false);
+                  }
+               }
+
+               if(leftKey() == rightKey() )
+               {
+                  this->state = s_next_right;
+                  this->hasRightMark = true;
+                  this->rightKeyAtMark = rightKey();
+                  this->rightMark = this->right.mark();
+                  return makeCurrent(true);
+               }
+
+               this->state = s_next_left;
+               return makeCurrent(false);
+            }
+
+            case s_next_right: {
+               if(!this->right.step() )
+               {
+                  this->state = s_next_left;
+                  return step();
+               }
+
+               if(leftKey() == rightKey() )
+               {
+                  this->state = s_next_right;
+                  return makeCurrent(true);
+               }
+
+               this->state = s_next_left;
+               return step();
+            }
+         }
+
+         return false;
+      }
+
+      ElementType* get()
+      {
+         return &current;
+      }
+
+      MarkerType mark() const
+      {
+         MarkerType result = { this->left.mark(), this->right.mark(), this->state,
+            this->hasRightMark, this->rightMark };
+         return result;
+      }
+
+      void restore(MarkerType mark)
+      {
+         this->hasRightMark = mark.hasRightMark;
+         if(mark.hasRightMark)
+         {
+            this->rightMark = mark.innerRightMark;
+            this->right.restore(mark.innerRightMark);
+            this->rightKeyAtMark = rightKey();
+         }
+
+         this->left.restore(mark.leftMark);
+         this->right.restore(mark.rightMark);
+         this->state = mark.state;
+
+         if(this->state == s_only_left)
+            makeCurrent(false);
+         else
+            makeCurrent(leftKey() == rightKey() );
+      }
+
+   private:
+      bool makeCurrent(bool hasRight)
+      {
+         if(hasRight)
+            this->current = ElementType(*this->left.get(), this->right.get() );
+         else
+            this->current = ElementType(*this->left.get(), nullptr);
+
+         return true;
+      }
+
+      LeftKey leftKey() { return this->keyExtract(*this->left.get() ); }
+      RightKey rightKey() { return this->keyExtract(*this->right.get() ); }
+
+   private:
+      Left left;
+      Right right;
+      KeyExtract keyExtract;
+      ElementType current;
+      State state;
+
+      bool hasRightMark;
+      RightKey rightKeyAtMark;
+      typename Right::MarkerType rightMark;
+};
+
+namespace db {
+
+template<typename KeyExtract, typename Left, typename Right>
+inline LeftJoinEq<Left, Right, KeyExtract> leftJoinBy(KeyExtract ex, Left left, Right right)
+{
+   return LeftJoinEq<Left, Right, KeyExtract>(left, right, ex);
+}
+
+}
+
+#endif

fsck/source/database/ModificationEvent.h

@@ -0,0 +1,18 @@
+#ifndef MODFICATIONEVENT_H_
+#define MODFICATIONEVENT_H_
+
+#include <database/EntryID.h>
+
+namespace db {
+
+struct ModificationEvent
+{
+   EntryID id;
+
+   typedef EntryID KeyType;
+   EntryID pkey() const { return id; }
+};
+
+}
+
+#endif

fsck/source/database/Select.h

@@ -0,0 +1,74 @@
+#ifndef SELECT_H_
+#define SELECT_H_
+
+#include <boost/utility/result_of.hpp>
+
+template<typename Source, typename Fn>
+class Select
+{
+   public:
+      typedef typename boost::result_of<Fn(typename Source::ElementType&)>::type ElementType;
+      typedef typename Source::MarkerType MarkerType;
+
+   public:
+      Select(Source source, Fn fn)
+         : source(source), fn(fn)
+      {
+      }
+
+      bool step()
+      {
+         if(!this->source.step() )
+            return false;
+
+         this->current = this->fn(*this->source.get() );
+         return true;
+      }
+
+      ElementType* get()
+      {
+         return &current;
+      }
+
+      MarkerType mark() const
+      {
+         return this->source.mark();
+      }
+
+      void restore(MarkerType mark)
+      {
+         this->source.restore(mark);
+         this->current = this->fn(*this->source.get() );
+      }
+
+   private:
+      Source source;
+      Fn fn;
+      ElementType current;
+};
+
+
+namespace db {
+
+template<typename Pred>
+struct SelectOp
+{
+   Pred pred;
+
+   template<typename Source>
+   friend Select<Source, Pred> operator|(Source source, const SelectOp& op)
+   {
+      return Select<Source, Pred>(source, op.pred);
+   }
+};
+
+template<typename Pred>
+inline SelectOp<Pred> select(Pred pred)
+{
+   SelectOp<Pred> result = {pred};
+   return result;
+}
+
+}
+
+#endif

fsck/source/database/Set.h

@@ -0,0 +1,380 @@
+#ifndef SET_H_
+#define SET_H_
+
+#include <common/threading/Mutex.h>
+#include <database/SetFragment.h>
+#include <database/SetFragmentCursor.h>
+#include <database/Union.h>
+
+#include <iomanip>
+#include <fstream>
+#include <limits>
+#include <map>
+#include <mutex>
+#include <sstream>
+
+#include <limits.h>
+
+#include <boost/format.hpp>
+
+template<typename Data>
+class Set {
+   public:
+      typedef SetFragment<Data> Fragment;
+      typedef SetFragmentCursor<Data> Cursor;
+
+   private:
+      static const unsigned VERSION = 1;
+
+      std::string basename;
+      unsigned nextID;
+      std::map<std::string, Fragment*> openFragments;
+      bool dropped;
+
+      Mutex mutex;
+
+      typedef typename std::map<std::string, Fragment*>::iterator FragmentIter;
+
+      Set(const Set&);
+      Set& operator=(const Set&);
+
+      std::string fragmentName(unsigned id)
+      {
+         std::stringstream ss;
+         ss << basename << '.' << std::setw(5) << std::setfill('0') << id;
+         return ss.str();
+      }
+
+      Fragment* getFragment(const std::string& file, bool allowCreate)
+      {
+         if(openFragments.count(file) )
+            return openFragments[file];
+
+         Fragment* fragment = new Fragment(file, allowCreate);
+         openFragments[file] = fragment;
+         return fragment;
+      }
+
+      void acceptFragment(Fragment* foreign)
+      {
+         const unsigned newID = nextID++;
+         const std::string& file = fragmentName(newID);
+
+         foreign->rename(file);
+
+         openFragments[file] = foreign;
+         saveConfig();
+      }
+
+      void removeFragment(const std::string& file)
+      {
+         Fragment* fragment = openFragments[file];
+         openFragments.erase(file);
+         fragment->drop();
+         delete fragment;
+      }
+
+      void loadConfig(bool allowCreate)
+      {
+         std::ifstream in( (basename + ".t").c_str() );
+
+         if(!in)
+         {
+            if(allowCreate)
+               return;
+
+            throw std::runtime_error("could not read set description file " + basename);
+         }
+
+         {
+            char code;
+            unsigned version;
+            in >> code >> version;
+            if (code != 'v' || version != VERSION) {
+               throw std::runtime_error(str(
+                        boost::format("cannot read set %1% with version %2%")
+                           % basename
+                           % version));
+            }
+         }
+
+         in >> nextID;
+
+         unsigned count = 0;
+         in >> count;
+
+         in.ignore(1, '\n');
+
+         while(count > 0)
+         {
+            std::string file;
+
+            std::getline(in, file);
+
+            if(!in)
+               break;
+
+            getFragment(basename + file, false);
+            count--;
+         }
+
+         if(!in || in.peek() != std::ifstream::traits_type::eof() )
+            throw std::runtime_error("bad set description for " + basename);
+      }
+
+      void saveConfig()
+      {
+         std::ofstream out((basename + ".t").c_str(), std::ostream::trunc);
+
+         out << 'v' << VERSION << '\n';
+         out << nextID << '\n';
+         out << openFragments.size() << '\n';
+
+         for (FragmentIter it = openFragments.begin(), end = openFragments.end(); it != end; ++it) {
+            if (it->second->filename().find(basename) != 0)
+               throw std::runtime_error("set corrupted: " + basename);
+
+            out << it->second->filename().substr(basename.size()) << '\n';
+         }
+      }
+
+      static std::string makeAbsolute(const std::string& path)
+      {
+         std::string::size_type spos = path.find_last_of('/');
+         if (spos == path.npos)
+            spos = 0;
+         else
+            spos += 1;
+
+         // can't/don't want to use ::basename because it may return static buffers
+         const std::string basename = path.substr(spos);
+
+         if (basename.empty() || basename == "." || basename == "..")
+            throw std::runtime_error("bad path");
+
+         if (path[0] == '/')
+            return path;
+
+         char cwd[PATH_MAX];
+
+         if (::getcwd(cwd, sizeof(cwd)) == NULL)
+            throw std::runtime_error("bad path");
+
+         return cwd + ('/' + path);
+      }
+
+   public:
+      Set(const std::string& basename, bool allowCreate = true)
+         : basename(makeAbsolute(basename) ), nextID(0), dropped(false)
+      {
+         loadConfig(allowCreate);
+      }
+
+      ~Set()
+      {
+         if(dropped)
+            return;
+
+         saveConfig();
+
+         for(FragmentIter it = openFragments.begin(), end = openFragments.end(); it != end; ++it)
+            delete it->second;
+      }
+
+      size_t size()
+      {
+         size_t result = 0;
+
+         for(FragmentIter it = openFragments.begin(), end = openFragments.end(); it != end; ++it)
+            result += it->second->size();
+
+         return result;
+      }
+
+      Fragment* newFragment()
+      {
+         const std::lock_guard<Mutex> lock(mutex);
+
+         Fragment* result = getFragment(fragmentName(nextID++), true);
+         saveConfig();
+
+         return result;
+      }
+
+      void sort()
+      {
+         if(openFragments.empty() )
+            newFragment();
+
+         std::multimap<size_t, Fragment*> sortedFragments;
+
+         for(FragmentIter it = openFragments.begin(), end = openFragments.end(); it != end; ++it)
+            it->second->flush();
+
+         for(FragmentIter it = openFragments.begin(), end = openFragments.end(); it != end; ++it)
+         {
+            it->second->sort();
+            sortedFragments.insert(std::make_pair(it->second->size(), it->second) );
+         }
+
+         static const unsigned MERGE_WIDTH = 4;
+
+         if(sortedFragments.size() == 1)
+            return;
+
+         saveConfig();
+
+         while(sortedFragments.size() > 1)
+         {
+            Fragment* inputs[MERGE_WIDTH] = {};
+            unsigned inputCount = 0;
+
+            for(inputCount = 0; inputCount < MERGE_WIDTH; inputCount++)
+            {
+               if(sortedFragments.empty() )
+                  break;
+
+               inputs[inputCount] = sortedFragments.begin()->second;
+               sortedFragments.erase(sortedFragments.begin());
+            }
+
+            SetFragment<Data>* merged = getFragment(fragmentName(nextID++), true);
+
+            struct op
+            {
+               static typename Data::KeyType key(const Data& d) { return d.pkey(); }
+            };
+            typedef typename Data::KeyType (*key_t)(const Data&);
+            typedef Union<Cursor, Cursor, key_t> L1Union;
+
+            switch(inputCount)
+            {
+            case 2: {
+               L1Union u(Cursor(*inputs[0]), (Cursor(*inputs[1]) ), op::key);
+               while(u.step() )
+                  merged->append(*u.get() );
+               break;
+            }
+
+            case 3: {
+               Union<L1Union, SetFragmentCursor<Data>, key_t> u(
+                  L1Union(Cursor(*inputs[0]), (Cursor(*inputs[1]) ), op::key),
+                  Cursor(*inputs[2]),
+                  op::key);
+               while(u.step() )
+                  merged->append(*u.get() );
+               break;
+            }
+
+            case 4: {
+               Union<L1Union, L1Union, key_t> u(
+                  L1Union(Cursor(*inputs[0]), (Cursor(*inputs[1]) ), op::key),
+                  L1Union(Cursor(*inputs[2]), (Cursor(*inputs[3]) ), op::key),
+                  op::key);
+               while(u.step() )
+                  merged->append(*u.get() );
+               break;
+            }
+
+            default:
+               throw std::runtime_error("");
+            }
+
+            merged->flush();
+
+            sortedFragments.insert(std::make_pair(merged->size(), merged) );
+
+            for (unsigned i = 0; i < inputCount; i++)
+               removeFragment(inputs[i]->filename() );
+
+            saveConfig();
+         }
+      }
+
+      Cursor cursor()
+      {
+         sort();
+
+         return Cursor(*openFragments.begin()->second);
+      }
+
+      template<typename Key, typename Fn>
+      std::pair<bool, Data> getByKeyProjection(const Key& key, Fn fn)
+      {
+         sort();
+
+         if(openFragments.empty() )
+            return std::make_pair(false, Data() );
+
+         return openFragments.begin()->second->getByKeyProjection(key, fn);
+      }
+
+      std::pair<bool, Data> getByKey(const typename Data::KeyType& key)
+      {
+         struct ops
+         {
+            static typename Data::KeyType key(const typename Data::KeyType& key) { return key; }
+         };
+
+         return getByKeyProjection(key, ops::key);
+      }
+
+      void clear()
+      {
+         while(!openFragments.empty() )
+            removeFragment(openFragments.begin()->first);
+
+         nextID = 0;
+         saveConfig();
+      }
+
+      void drop()
+      {
+         clear();
+         if(::unlink( (basename + ".t").c_str() ) )
+            throw std::runtime_error("could not unlink set file " + basename + ": " + std::string(strerror(errno)));
+
+         dropped = true;
+      }
+
+      void mergeInto(Set& other)
+      {
+         for(FragmentIter it = openFragments.begin(), end = openFragments.end(); it != end; ++it)
+            other.acceptFragment(it->second);
+
+         openFragments.clear();
+         saveConfig();
+      }
+
+      void makeUnique()
+      {
+         sort();
+
+         Fragment* input = openFragments.begin()->second;
+         Cursor cursor(*input);
+         typename Data::KeyType key;
+
+         if(!cursor.step() )
+            return;
+
+         Fragment* unique = newFragment();
+         unique->append(*cursor.get() );
+         key = cursor.get()->pkey();
+
+         while(cursor.step() )
+         {
+            if(key == cursor.get()->pkey() )
+               continue;
+
+            unique->append(*cursor.get() );
+            key = cursor.get()->pkey();
+         }
+
+         unique->flush();
+
+         removeFragment(input->filename() );
+         saveConfig();
+      }
+};
+
+#endif

fsck/source/database/SetFragment.h

@@ -0,0 +1,325 @@
+#ifndef SETFRAGMENT_H_
+#define SETFRAGMENT_H_
+
+#include <algorithm>
+#include <cerrno>
+#include <stdexcept>
+#include <string>
+#include <vector>
+
+#include <boost/scoped_array.hpp>
+#include <fcntl.h>
+#include <stdio.h>
+#include <string.h>
+#include <unistd.h>
+
+struct FragmentDoesNotExist : public std::runtime_error
+{
+   FragmentDoesNotExist(const std::string& file)
+      : runtime_error(file)
+   {}
+};
+
+template<typename Data>
+class SetFragment {
+   public:
+      static const unsigned CONFIG_AREA_SIZE = 4096;
+      static const size_t BUFFER_SIZE = 4ULL * 1024 * 1024;
+
+   private:
+      std::string file;
+      int fd;
+      size_t itemCount;
+      bool sorted;
+
+      std::vector<Data> buffer;
+      size_t firstBufferedItem;
+      ssize_t firstDirtyItem;
+      ssize_t lastDirtyItem;
+
+      typename Data::KeyType lastItemKey;
+
+      SetFragment(const SetFragment&);
+      SetFragment& operator=(const SetFragment&);
+
+      size_t readBlock(Data* into, size_t count, size_t from)
+      {
+         size_t total = 0;
+         count *= sizeof(Data);
+         from *= sizeof(Data);
+
+         from += CONFIG_AREA_SIZE;
+
+         char* buf = (char*) into;
+
+         while(total < count)
+         {
+            ssize_t current = ::pread(fd, buf + total, count - total, from + total);
+            total += current;
+
+            if (current < 0)
+               throw std::runtime_error("read failed: " + std::string(strerror(errno)));
+
+            if (current == 0)
+               break;
+         }
+
+         return total / sizeof(Data);
+      }
+
+      size_t writeBlock(const Data* source, size_t count, size_t from)
+      {
+         size_t total = 0;
+         count *= sizeof(Data);
+         from *= sizeof(Data);
+
+         from += CONFIG_AREA_SIZE;
+
+         const char* buf = (const char*) source;
+
+         while(total < count)
+         {
+            ssize_t current = ::pwrite(fd, buf + total, count - total, from + total);
+            total += current;
+
+            if (current < 0)
+               throw std::runtime_error("write failed: " + std::string(strerror(errno)));
+
+            if (current == 0)
+               break;
+         }
+
+         return total / sizeof(Data);
+      }
+
+      void flushBuffer()
+      {
+         if(firstDirtyItem < 0)
+            return;
+
+         const Data* first = &buffer[firstDirtyItem - firstBufferedItem];
+
+         writeBlock(first, lastDirtyItem - firstDirtyItem + 1, firstDirtyItem);
+
+         buffer.clear();
+         firstDirtyItem = -1;
+      }
+
+      void bufferFileRange(size_t begin, size_t end)
+      {
+         flushBuffer();
+
+         const size_t firstAddress = begin;
+         const size_t windowSize = end == size_t(-1) ? end : (end - begin + 1);
+         const size_t totalRead = std::min<size_t>(windowSize, BUFFER_SIZE / sizeof(Data) );
+
+         buffer.resize(totalRead);
+
+         size_t read = readBlock(&buffer[0], totalRead, firstAddress);
+
+         firstBufferedItem = begin;
+         buffer.resize(read);
+      }
+
+   public:
+      SetFragment(const std::string& file, bool allowCreate = true)
+         : file(file),
+           itemCount(0),
+           firstBufferedItem(0), firstDirtyItem(-1), lastDirtyItem(0)
+      {
+         fd = ::open(file.c_str(), O_RDWR | (allowCreate ? O_CREAT : 0), 0660);
+         if(fd < 0)
+         {
+            int eno = errno;
+            if(!allowCreate && errno == ENOENT)
+               throw FragmentDoesNotExist(file);
+            else
+               throw std::runtime_error("could not open fragment file " + file + ": " + strerror(eno));
+         }
+
+         off_t totalSize = ::lseek(fd, 0, SEEK_END);
+
+         if(totalSize > 0)
+         {
+            int eno = errno;
+            if (totalSize < (off_t) CONFIG_AREA_SIZE)
+               throw std::runtime_error("error while opening fragment file " + file + ": " + strerror(eno));
+            else
+            if ( (totalSize - CONFIG_AREA_SIZE) % sizeof(Data) )
+               throw std::runtime_error("error while opening fragment file " + file + ": " + strerror(eno));
+         }
+
+         if(totalSize == 0)
+         {
+            sorted = true;
+            itemCount = 0;
+         }
+         else
+         {
+            if (::pread(fd, &sorted, sizeof(sorted), 0) != sizeof(sorted)) {
+               int eno = errno;
+               throw std::runtime_error("error while opening fragment file " + file + ": " + strerror(eno));
+            }
+
+            itemCount = (totalSize - CONFIG_AREA_SIZE) / sizeof(Data);
+         }
+
+         if(itemCount > 0)
+            lastItemKey = (*this)[itemCount - 1].pkey();
+      }
+
+      ~SetFragment()
+      {
+         if(fd >= 0)
+         {
+            flush();
+            close(fd);
+         }
+      }
+
+      const std::string filename() const { return file; }
+      size_t size() const { return itemCount; }
+
+      void append(const Data& data)
+      {
+         if(itemCount > firstBufferedItem + buffer.size() )
+         {
+            flushBuffer();
+            firstBufferedItem = itemCount;
+         }
+
+         if(firstDirtyItem < 0)
+            firstDirtyItem = itemCount;
+
+         lastDirtyItem = itemCount;
+
+         if(itemCount > 0 && sorted)
+            sorted = lastItemKey < data.pkey();
+
+         buffer.push_back(data);
+         itemCount++;
+         lastItemKey = data.pkey();
+
+         if(buffer.size() * sizeof(Data) >= BUFFER_SIZE)
+            flushBuffer();
+      }
+
+      Data& operator[](size_t offset)
+      {
+         if(offset < firstBufferedItem || offset >= firstBufferedItem + buffer.size() )
+            bufferFileRange(offset == 0 ? 0 : offset - 1, -1);
+
+         return buffer[offset - firstBufferedItem];
+      }
+
+      void flush()
+      {
+         if(firstDirtyItem < 0)
+            return;
+
+         flushBuffer();
+         buffer = std::vector<Data>();
+
+         if (::pwrite(fd, &sorted, sizeof(sorted), 0) < 0) {
+            int eno = errno;
+            throw std::runtime_error("error in flush of " + file + ": " + strerror(eno));
+         }
+
+         // truncate to CONFIG_AREA_SIZE (for reopen)
+         if (itemCount == 0 && ::ftruncate(fd, CONFIG_AREA_SIZE) < 0) {
+            int eno = errno;
+            throw std::runtime_error("error in flush of " + file + ": " + strerror(eno));
+         }
+      }
+
+      void sort()
+      {
+         flush();
+
+         if(sorted)
+            return;
+
+         boost::scoped_array<Data> data(new Data[size()]);
+
+         if(readBlock(data.get(), size(), 0) < size() )
+            throw 42;
+
+         struct ops
+         {
+            static bool compare(const Data& l, const Data& r)
+            {
+               return l.pkey() < r.pkey();
+            }
+         };
+
+         std::sort(data.get(), data.get() + size(), ops::compare);
+
+         writeBlock(data.get(), size(), 0);
+
+         sorted = true;
+
+         flush();
+      }
+
+      void drop()
+      {
+         flush();
+         if (::unlink(file.c_str()) < 0) {
+            int eno = errno;
+            throw std::runtime_error("could not unlink fragment file " + file + ": " + strerror(eno));
+         }
+
+         close(fd);
+         fd = -1;
+      }
+
+      void rename(const std::string& to)
+      {
+         if (::rename(file.c_str(), to.c_str()) < 0) {
+            int eno = errno;
+            throw std::runtime_error("could not rename fragment file " + file + ": " + strerror(eno));
+         }
+
+         file = to;
+      }
+
+      template<typename Key, typename Fn>
+      std::pair<bool, Data> getByKeyProjection(const Key& key, Fn fn)
+      {
+         if(size() == 0)
+            return std::make_pair(false, Data() );
+
+         size_t first = 0;
+         size_t last = size() - 1;
+
+         while(first != last)
+         {
+            size_t midIdx = first + (last - first) / 2;
+
+            bufferFileRange(midIdx, midIdx + 1);
+            Data& mid = (*this)[midIdx];
+
+            if (fn(mid.pkey() ) < key)
+               first = midIdx + 1;
+            else
+               last = midIdx;
+         }
+
+         if(fn( (*this)[first].pkey() ) == key)
+            return std::make_pair(true, (*this)[first]);
+         else
+            return std::make_pair(false, Data() );
+      }
+
+      std::pair<bool, Data> getByKey(const typename Data::KeyType& key)
+      {
+         struct ops
+         {
+            static typename Data::KeyType key(const typename Data::KeyType& key) { return key; }
+         };
+
+         return getByKeyProjection(key, ops::key);
+      }
+};
+
+#endif

fsck/source/database/SetFragmentCursor.h

@@ -0,0 +1,47 @@
+#ifndef SETFRAGMENTCURSOR_H_
+#define SETFRAGMENTCURSOR_H_
+
+#include <database/SetFragment.h>
+
+template<typename Data>
+class SetFragmentCursor {
+   public:
+      typedef Data ElementType;
+      typedef size_t MarkerType;
+
+   private:
+      SetFragment<Data>* fragment;
+      size_t currentGetIndex;
+
+   public:
+      explicit SetFragmentCursor(SetFragment<Data>& fragment)
+         : fragment(&fragment), currentGetIndex(-1)
+      {}
+
+      bool step()
+      {
+         if (currentGetIndex + 1 >= fragment->size() )
+            return false;
+
+         currentGetIndex++;
+         return true;
+      }
+
+      Data* get()
+      {
+         return &(*fragment)[currentGetIndex];
+      }
+
+      MarkerType mark() const
+      {
+         return currentGetIndex;
+      }
+
+      void restore(MarkerType mark)
+      {
+         currentGetIndex = mark - 1;
+         step();
+      }
+};
+
+#endif

fsck/source/database/StripeTargets.h

@@ -0,0 +1,45 @@
+#ifndef STRIPETARGETS_H_
+#define STRIPETARGETS_H_
+
+#include <database/EntryID.h>
+
+#include <inttypes.h>
+
+namespace db {
+
+struct StripeTargets
+{
+   EntryID id;                 /* 12 */
+
+   enum {
+      NTARGETS = 7
+   };
+   uint16_t targets[NTARGETS]; /* 26 */
+   uint16_t firstTargetIndex;  /* 28 */
+
+   typedef EntryID KeyType;
+
+   EntryID pkey() const { return id; }
+
+   friend std::ostream& operator<<(std::ostream& os, StripeTargets const& obj)
+   {
+      os << "-------------" << "\n";
+      os << "StripeTargets" << "\n";
+      os << "-------------" << "\n";
+      os << "EntryID: " << obj.id.str() << "\n";
+      os << "Stripe targets: [ ";
+
+      for (int i=0; i<obj.NTARGETS; i++)
+      {
+         os <<  obj.targets[i] << " ";
+      }
+      os << "]\n";
+
+      os << "FirstTargetsIndex: " << obj.firstTargetIndex << "\n";
+      return os;
+   }
+};
+
+}
+
+#endif

fsck/source/database/Table.h

@@ -0,0 +1,430 @@
+#ifndef TABLE_H_
+#define TABLE_H_
+
+#include <database/Buffer.h>
+#include <database/Filter.h>
+#include <database/LeftJoinEq.h>
+#include <database/Select.h>
+#include <database/Set.h>
+
+#include <boost/make_shared.hpp>
+#include <boost/scoped_ptr.hpp>
+#include <boost/shared_ptr.hpp>
+#include <boost/ref.hpp>
+#include <boost/weak_ptr.hpp>
+
+template<typename Data, bool AllowDuplicateKeys = false>
+class Table {
+   private:
+      struct Key
+      {
+         typedef typename Data::KeyType KeyType;
+
+         KeyType value;
+
+         Key() {}
+
+         Key(const KeyType& key) : value(key) {}
+
+         KeyType pkey() const { return value; }
+
+         bool operator<(const Key& r) const { return value < r.value; }
+         bool operator==(const Key& r) const { return value == r.value; }
+      };
+
+      struct GetKey
+      {
+         typedef typename Data::KeyType result_type;
+
+         result_type operator()(const Data& data) const { return data.pkey(); }
+         result_type operator()(const Key& data) const { return data.value; }
+      };
+
+      struct SecondIsNull
+      {
+         typedef bool result_type;
+
+         bool operator()(std::pair<Data, Key*>& p) const { return p.second == NULL; }
+      };
+
+      struct First
+      {
+         typedef Data result_type;
+
+         Data operator()(std::pair<Data, Key*>& p) const { return p.first; }
+      };
+
+      struct HasMatchingKey
+      {
+         HasMatchingKey(const typename Data::KeyType& k) : key(k) {}
+         typename Data::KeyType key;
+
+         bool operator()(std::pair<Data, Key*>& p) const
+         {
+            return (p.first.pkey() == key);
+         }
+      };
+
+   private:
+      std::string basename;
+      size_t fragmentSize;
+
+      Set<Data> base;
+      Set<Data> inserts;
+      Set<Key> deletes;
+
+      typename Data::KeyType lastChangeKey;
+      Set<Data> insertsPending;
+      Set<Key> deletesPending;
+
+      boost::scoped_ptr<Buffer<Data> > insertBuffer;
+      boost::scoped_ptr<Buffer<Key> > deleteBuffer;
+
+      enum {
+         ms_none,
+         ms_ordered_only_insert,
+         ms_ordered_only_delete,
+         ms_ordered_at_insert,
+         ms_ordered_at_delete,
+         ms_unordered_insert,
+         ms_unordered_delete,
+         ms_bulk_insert,
+      } modificationState;
+
+      bool dropped;
+      bool insertsIntoBase;
+
+      std::vector<boost::weak_ptr<Buffer<Data> > > bulkInserters;
+
+      void assertNoBulkInsert()
+      {
+         for(size_t i = 0; i < bulkInserters.size(); i++)
+         {
+            if(!bulkInserters[i].expired() )
+               throw std::runtime_error("bulk insert still running");
+         }
+
+         bulkInserters.clear();
+      }
+
+      void collapseChanges()
+      {
+         deleteBuffer.reset();
+         insertBuffer.reset();
+         insertsIntoBase = false;
+
+         if(deletesPending.size() )
+         {
+            Set<Data> insertsTemp(basename + ".it");
+
+            LeftJoinEq<
+                  SetFragmentCursor<Data>,
+                  SetFragmentCursor<Key>,
+                  GetKey>
+               cleanedInserts = db::leftJoinBy(
+                  GetKey(),
+                  inserts.cursor(),
+                  deletesPending.cursor() );
+
+            // sequence will always be sorted, no need to fragment and sort again
+            SetFragment<Data>* fragment = insertsTemp.newFragment();
+            while(cleanedInserts.step() )
+            {
+               if (cleanedInserts.get()->second == nullptr)
+                  fragment->append(cleanedInserts.get()->first);
+            }
+
+            inserts.clear();
+            insertsTemp.mergeInto(inserts);
+            insertsTemp.drop();
+
+            deletesPending.mergeInto(deletes);
+            deletes.makeUnique();
+         }
+
+         if(insertsPending.size() )
+            insertsPending.mergeInto(inserts);
+
+         modificationState = ms_none;
+      }
+
+      void setupBaseBuffer()
+      {
+         insertBuffer.reset(new Buffer<Data>(base, fragmentSize));
+         insertsIntoBase = true;
+      }
+
+   public:
+      Table(const std::string& basename, size_t fragmentSize, bool allowCreate = true)
+         : basename(basename), fragmentSize(fragmentSize),
+           base(basename + ".b", allowCreate),
+           inserts(basename + ".i", allowCreate),
+           deletes(basename + ".d", allowCreate),
+           insertsPending(basename + ".ip", true),
+           deletesPending(basename + ".dp", true),
+           modificationState(ms_none),
+           dropped(false), insertsIntoBase(false)
+      {
+         if (base.size() == 0)
+            setupBaseBuffer();
+      }
+
+      ~Table()
+      {
+         if(dropped)
+            return;
+
+         collapseChanges();
+
+         deletesPending.drop();
+         insertsPending.drop();
+      }
+
+      boost::shared_ptr<Buffer<Data> > bulkInsert()
+      {
+         if(inserts.size() || deletes.size() || insertsPending.size() || deletesPending.size() )
+            throw std::runtime_error("unordered operation");
+
+         switch(modificationState)
+         {
+         case ms_none:
+         case ms_ordered_only_insert:
+         case ms_bulk_insert:
+            break;
+
+         default:
+            throw std::runtime_error("unordered operation");
+         }
+
+         modificationState = ms_bulk_insert;
+         boost::shared_ptr<Buffer<Data> > buffer = boost::make_shared<Buffer<Data> >(
+            boost::ref(base), fragmentSize);
+
+         bulkInserters.push_back(buffer);
+         return buffer;
+      }
+
+      void insert(const Data& data)
+      {
+         using namespace std::rel_ops;
+
+         switch(modificationState)
+         {
+         case ms_bulk_insert:
+            assertNoBulkInsert();
+            BEEGFS_FALLTHROUGH;
+
+         case ms_none:
+            modificationState = ms_ordered_only_insert;
+            break;
+
+         case ms_ordered_only_insert:
+            if(lastChangeKey >= data.pkey() )
+               modificationState = ms_unordered_insert;
+
+            break;
+
+         case ms_unordered_delete:
+            throw std::runtime_error("unordered operation");
+
+         case ms_unordered_insert:
+            break;
+
+         case ms_ordered_at_delete:
+         case ms_ordered_only_delete:
+            if(lastChangeKey <= data.pkey() )
+               modificationState = ms_ordered_at_insert;
+            else
+               throw std::runtime_error("unordered operation");
+
+            break;
+
+         case ms_ordered_at_insert:
+            if(lastChangeKey < data.pkey()
+                  || (!AllowDuplicateKeys && lastChangeKey == data.pkey() ) )
+               throw std::runtime_error("unordered operation");
+
+            break;
+         }
+
+         lastChangeKey = data.pkey();
+
+         if(!insertBuffer)
+            insertBuffer.reset(new Buffer<Data>(insertsPending, fragmentSize) );
+
+         insertBuffer->append(data);
+      }
+
+      void remove(const typename Data::KeyType& key)
+      {
+         using namespace std::rel_ops;
+
+         if(modificationState == ms_bulk_insert)
+            assertNoBulkInsert();
+
+         // if deletes happen during initial insert, complete the base set. delete tracking
+         // cannot cope otherwise
+         if(insertsIntoBase)
+         {
+            insertBuffer.reset();
+            insertsIntoBase = false;
+            modificationState = ms_none;
+         }
+
+         switch(modificationState)
+         {
+         case ms_none:
+         case ms_bulk_insert:
+            modificationState = ms_ordered_only_delete;
+            break;
+
+         case ms_ordered_only_delete:
+            if(lastChangeKey >= key)
+               modificationState = ms_unordered_delete;
+
+            break;
+
+         case ms_unordered_delete:
+            break;
+
+         case ms_unordered_insert:
+            throw std::runtime_error("unordered operation");
+
+         case ms_ordered_at_delete:
+         case ms_ordered_at_insert:
+         case ms_ordered_only_insert:
+            if(key <= lastChangeKey)
+               throw std::runtime_error("unordered operation");
+
+            modificationState = ms_ordered_at_delete;
+            break;
+         }
+
+         lastChangeKey = key;
+
+         if(!deleteBuffer)
+            deleteBuffer.reset(new Buffer<Key>(deletesPending, fragmentSize) );
+
+         deleteBuffer->append(key);
+      }
+
+      void commitChanges()
+      {
+         collapseChanges();
+         base.sort();
+         inserts.sort();
+         deletes.sort();
+      }
+
+      typedef Union<
+         Select<
+            Filter<
+               LeftJoinEq<
+                  SetFragmentCursor<Data>,
+                  SetFragmentCursor<Key>,
+                  GetKey>,
+               SecondIsNull>,
+            First>,
+         SetFragmentCursor<Data>,
+         GetKey> QueryType;
+
+      QueryType cursor()
+      {
+         return db::unionBy(
+            GetKey(),
+            db::leftJoinBy(
+               GetKey(),
+               base.cursor(),
+               deletes.cursor() )
+            | db::where(SecondIsNull() )
+            | db::select(First() ),
+            inserts.cursor() );
+      }
+
+      typedef Union<
+         Select<
+            Filter<
+               Filter<
+                  LeftJoinEq<
+                     SetFragmentCursor<Data>,
+                     SetFragmentCursor<Key>,
+                     GetKey>,
+                  SecondIsNull>,
+               HasMatchingKey>,
+            First>,
+         SetFragmentCursor<Data>,
+         GetKey> MultiRowResultType;
+
+      MultiRowResultType getAllByKey(const typename Data::KeyType& key)
+      {
+         return db::unionBy(
+            GetKey(),
+            db::leftJoinBy(
+               GetKey(),
+               base.cursor(),
+               deletes.cursor() )
+            | db::where(SecondIsNull() )
+            | db::where(HasMatchingKey(key) )
+            | db::select(First() ),
+            inserts.cursor() );
+      }
+
+      template<typename Key, typename Fn>
+      std::pair<bool, Data> getByKeyProjection(const Key& key, Fn fn)
+      {
+         Data dummy = {};
+
+         std::pair<bool, Data> insIdx = inserts.getByKeyProjection(key, fn);
+         if(insIdx.first)
+            return insIdx;
+
+         std::pair<bool, Data> baseIdx = base.getByKeyProjection(key, fn);
+         if(!baseIdx.first)
+            return std::make_pair(false, dummy);
+
+         std::pair<bool, typename Table::Key> delIdx = deletes.getByKey(baseIdx.second.pkey() );
+         if(delIdx.first)
+            return std::make_pair(false, dummy);
+
+         return baseIdx;
+      }
+
+      std::pair<bool, Data> getByKey(const typename Data::KeyType& key)
+      {
+         struct ops
+         {
+            static typename Data::KeyType key(const typename Data::KeyType& key) { return key; }
+         };
+
+         return getByKeyProjection(key, ops::key);
+      }
+
+      void drop()
+      {
+         insertBuffer.reset();
+         deleteBuffer.reset();
+
+         base.drop();
+         inserts.drop();
+         deletes.drop();
+         insertsPending.drop();
+         deletesPending.drop();
+         dropped = true;
+      }
+
+      void clear()
+      {
+         insertBuffer.reset();
+         deleteBuffer.reset();
+
+         base.clear();
+         inserts.clear();
+         deletes.clear();
+         insertsPending.clear();
+         deletesPending.clear();
+
+         setupBaseBuffer();
+      }
+};
+
+#endif

fsck/source/database/Union.h

@@ -0,0 +1,130 @@
+#ifndef UNION_H_
+#define UNION_H_
+
+#include <utility>
+
+template<typename Left, typename Right, typename KeyExtract>
+class Union
+{
+   public:
+      typedef typename Left::ElementType ElementType;
+
+      struct MarkerType
+      {
+         typename Left::MarkerType leftMark;
+         typename Right::MarkerType rightMark;
+         bool leftEnded;
+         bool rightEnded;
+         bool nextStepLeft;
+      };
+
+   public:
+      Union(Left left, Right right, KeyExtract keyExtract):
+         left(left), right(right), keyExtract(keyExtract), leftEnded(false), nextStepLeft(true),
+         currentAtLeft(false)
+      {
+         this->rightEnded = !this->right.step();
+      }
+
+      bool step()
+      {
+         if(this->leftEnded && this->rightEnded)
+            return false;
+
+         if(this->leftEnded)
+         {
+            this->rightEnded = !this->right.step();
+            return resetCurrent();
+         }
+
+         if(this->rightEnded)
+         {
+            this->leftEnded = !this->left.step();
+            return resetCurrent();
+         }
+
+         if(this->nextStepLeft)
+            this->leftEnded = !this->left.step();
+         else
+            this->rightEnded = !this->right.step();
+
+         return resetCurrent();
+      }
+
+      ElementType* get()
+      {
+         if(this->currentAtLeft)
+            return this->left.get();
+         else
+            return this->right.get();
+      }
+
+      MarkerType mark() const
+      {
+         MarkerType result = { this->left.mark(), this->right.mark(), this->leftEnded,
+            this->rightEnded, this->nextStepLeft };
+         return result;
+      }
+
+      void restore(MarkerType mark)
+      {
+         this->leftEnded = mark.leftEnded;
+         this->rightEnded = mark.rightEnded;
+         this->nextStepLeft = mark.nextStepLeft;
+
+         if(!this->leftEnded)
+            this->left.restore(mark.leftMark);
+
+         if(!this->rightEnded)
+            this->right.restore(mark.rightMark);
+
+         resetCurrent();
+      }
+
+   private:
+      bool resetCurrent()
+      {
+         if(this->leftEnded && this->rightEnded)
+            return false;
+
+         if(this->leftEnded)
+            this->currentAtLeft = false;
+         else
+         if(this->rightEnded)
+            this->currentAtLeft = true;
+         else
+         if(this->keyExtract(*this->left.get() ) < this->keyExtract(*this->right.get() ) )
+         {
+            this->currentAtLeft = true;
+            this->nextStepLeft = true;
+         }
+         else
+         {
+            this->currentAtLeft = false;
+            this->nextStepLeft = false;
+         }
+
+         return true;
+      }
+
+   private:
+      Left left;
+      Right right;
+      KeyExtract keyExtract;
+      bool leftEnded;
+      bool rightEnded;
+      bool nextStepLeft;
+      bool currentAtLeft;
+};
+
+namespace db {
+
+template<typename KeyExtract, typename Left, typename Right>
+inline Union<Left, Right, KeyExtract> unionBy(KeyExtract keyExtract, Left left, Right right)
+{
+   return Union<Left, Right, KeyExtract>(left, right, keyExtract);
+}
+
+}
+
+#endif

fsck/source/database/UsedTarget.h

@@ -0,0 +1,25 @@
+#ifndef USEDTARGET_H_
+#define USEDTARGET_H_
+
+#include <common/fsck/FsckTargetID.h>
+
+namespace db {
+
+struct UsedTarget
+{
+   uint32_t id;
+   uint32_t type;
+
+   typedef std::pair<uint32_t, uint32_t> KeyType;
+
+   KeyType pkey() const { return std::make_pair(id, type); }
+
+   operator FsckTargetID() const
+   {
+      return FsckTargetID(id, FsckTargetIDType(type) );
+   }
+};
+
+}
+
+#endif

fsck/source/database/VectorSource.h

@@ -0,0 +1,52 @@
+#ifndef VECTORSOURCE_H_
+#define VECTORSOURCE_H_
+
+#include <vector>
+
+#include <boost/shared_ptr.hpp>
+#include <boost/make_shared.hpp>
+
+template<typename Obj>
+class VectorSource
+{
+   public:
+      typedef Obj ElementType;
+      typedef typename std::vector<Obj>::const_iterator MarkerType;
+
+   public:
+      explicit VectorSource(std::vector<Obj>& data)
+         : content(boost::make_shared<std::vector<Obj> >() ), index(-1)
+      {
+         data.swap(*content);
+      }
+
+      bool step()
+      {
+         if(this->index == this->content->size() - 1)
+            return false;
+
+         this->index += 1;
+         return true;
+      }
+
+      ElementType* get()
+      {
+         return &(*this->content)[this->index];
+      }
+
+      MarkerType mark() const
+      {
+         return this->content->begin() + this->index;
+      }
+
+      void restore(MarkerType mark)
+      {
+         this->index = mark - this->content->begin();
+      }
+
+   private:
+      boost::shared_ptr<std::vector<Obj> > content;
+      size_t index;
+};
+
+#endif

fsck/source/modes/Mode.cpp

@@ -0,0 +1,23 @@
+#include "Mode.h"
+
+
+/**
+ * Check given config for a remaining arg and print it as invalid.
+ *
+ * @return true if a remaining invalid argument was found in the given config.
+ */
+bool Mode::checkInvalidArgs(const StringMap* cfg)
+{
+   if(cfg->empty() )
+      return false;
+
+   if(cfg->size() == 1)
+      std::cerr << "Invalid argument: " << cfg->begin()->first << std::endl;
+   else
+   { // multiple invalid args
+      std::cerr << "Found " << cfg->size() << " invalid arguments. One of them is: " <<
+         cfg->begin()->first << std::endl;
+   }
+
+   return true;
+}

fsck/source/modes/Mode.h

@@ -0,0 +1,24 @@
+#ifndef MODE_H_
+#define MODE_H_
+
+#include <common/nodes/Node.h>
+#include <common/Common.h>
+
+class Mode
+{
+   public:
+      virtual ~Mode() {}
+
+      /**
+       * Executes the mode inside the calling thread.
+       *
+       * @return APPCODE_...
+       */
+      virtual int execute() = 0;
+
+   protected:
+      Mode() {}
+      bool checkInvalidArgs(const StringMap* cfg);
+};
+
+#endif /*MODE_H_*/

fsck/source/modes/ModeCheckFS.h

@@ -0,0 +1,170 @@
+#ifndef MODECHECKFS_H
+#define MODECHECKFS_H
+
+#include <boost/scoped_ptr.hpp>
+#include <vector>
+
+#include <database/FsckDB.h>
+#include <database/FsckDBException.h>
+#include <modes/Mode.h>
+
+class UserPrompter
+{
+   public:
+      template<unsigned Actions>
+      UserPrompter(const FsckRepairAction (&possibleActions)[Actions],
+         FsckRepairAction defaultRepairAction);
+
+      FsckRepairAction chooseAction(const std::string& prompt);
+
+   private:
+      bool askForAction;
+      std::vector<FsckRepairAction> possibleActions;
+      FsckRepairAction repairAction;
+};
+
+struct RepairChunkState
+{
+   UserPrompter* prompt;
+   std::string lastID;
+   FsckRepairAction lastChunkAction;
+};
+
+struct FsckErrCount
+{
+   FsckErrCount() : unfixableErrors(0), fixableErrors(0)
+   {
+   }
+
+   FsckErrCount operator+(const FsckErrCount& other)
+   {
+      unfixableErrors += other.unfixableErrors;
+      fixableErrors += other.fixableErrors;
+      return *this;
+   }
+
+   void operator+=(const FsckErrCount& other)
+   {
+      *this = *this + other;
+   }
+
+   uint64_t getTotalErrors() const
+   {
+      return unfixableErrors + fixableErrors;
+   }
+
+   uint64_t unfixableErrors;
+   uint64_t fixableErrors;
+};
+
+class ModeCheckFS : public Mode
+{
+   public:
+      ModeCheckFS();
+      virtual ~ModeCheckFS();
+
+      static void printHelp();
+
+      virtual int execute();
+
+   private:
+      boost::scoped_ptr<FsckDB> database;
+
+      LogContext log;
+
+      NodeHandle lostAndFoundNode;
+      EntryInfo lostAndFoundInfo;
+      boost::shared_ptr<FsckDirInode> lostAndFoundInode;
+
+      std::set<NumNodeID> secondariesSetBad;
+
+      int initDatabase();
+      void printHeaderInformation();
+      void disposeUnusedFiles();
+      FhgfsOpsErr gatherData(bool forceRestart);
+
+      template<typename Obj, typename State>
+      FsckErrCount checkAndRepairGeneric(Cursor<Obj> cursor,
+         void (ModeCheckFS::*repair)(Obj&, FsckErrCount&, State&), State& state);
+
+      FsckErrCount checkAndRepairDanglingDentry();
+      FsckErrCount checkAndRepairWrongInodeOwner();
+      FsckErrCount checkAndRepairWrongOwnerInDentry();
+      FsckErrCount checkAndRepairOrphanedContDir();
+      FsckErrCount checkAndRepairOrphanedDirInode();
+      FsckErrCount checkAndRepairOrphanedFileInode();
+      FsckErrCount checkAndRepairDuplicateInodes();
+      FsckErrCount checkAndRepairOrphanedChunk();
+      FsckErrCount checkAndRepairMissingContDir();
+      FsckErrCount checkAndRepairWrongFileAttribs();
+      FsckErrCount checkAndRepairWrongDirAttribs();
+      FsckErrCount checkAndRepairFilesWithMissingTargets();
+      FsckErrCount checkAndRepairDirEntriesWithBrokeByIDFile();
+      FsckErrCount checkAndRepairOrphanedDentryByIDFiles();
+      FsckErrCount checkAndRepairChunksWithWrongPermissions();
+      FsckErrCount checkMissingMirrorChunks();
+      FsckErrCount checkMissingPrimaryChunks();
+      FsckErrCount checkDifferingChunkAttribs();
+      FsckErrCount checkAndRepairChunksInWrongPath();
+      FsckErrCount checkAndUpdateOldStyledHardlinks();
+
+      void logDuplicateInodeID(checks::DuplicatedInode& dups, int&);
+
+      FsckErrCount checkDuplicateChunks();
+      void logDuplicateChunk(std::list<FsckChunk>& dups, FsckErrCount& errCount, int&);
+
+      FsckErrCount checkDuplicateContDirs();
+      void logDuplicateContDir(std::list<db::ContDir>& dups, FsckErrCount& errCount, int&);
+
+      FsckErrCount checkMismirroredDentries();
+      void logMismirroredDentry(db::DirEntry& entry, FsckErrCount& errCount, int&);
+
+      FsckErrCount checkMismirroredDirectories();
+      void logMismirroredDirectory(db::DirInode& dir, FsckErrCount& errCount, int&);
+
+      FsckErrCount checkMismirroredFiles();
+      void logMismirroredFile(db::FileInode& file, FsckErrCount& errCount, int&);
+
+      FsckErrCount checkAndRepairMalformedChunk();
+      void repairMalformedChunk(FsckChunk& chunk, FsckErrCount& errCount, UserPrompter& prompt);
+
+      void checkAndRepair();
+
+      void repairDanglingDirEntry(db::DirEntry& entry, FsckErrCount& errCount,
+         std::pair<UserPrompter*, UserPrompter*>& prompt);
+      void repairWrongInodeOwner(FsckDirInode& inode, FsckErrCount& errCount,
+         UserPrompter& prompt);
+      void repairWrongInodeOwnerInDentry(std::pair<db::DirEntry, NumNodeID>& error,
+         FsckErrCount& errCount, UserPrompter& prompt);
+      void repairOrphanedDirInode(FsckDirInode& inode, FsckErrCount& errCount,
+         UserPrompter& prompt);
+      void repairOrphanedFileInode(FsckFileInode& inode, FsckErrCount& errCount,
+         UserPrompter& prompt);
+      void repairDuplicateInode(checks::DuplicatedInode& dupInode, FsckErrCount& errCount,
+         UserPrompter& prompt);
+      void repairOrphanedChunk(FsckChunk& chunk, FsckErrCount& errCount, RepairChunkState& state);
+      void repairMissingContDir(FsckDirInode& inode, FsckErrCount& errCount, UserPrompter& prompt);
+      void repairOrphanedContDir(FsckContDir& dir, FsckErrCount& errCount, UserPrompter& prompt);
+      void repairWrongFileAttribs(std::pair<FsckFileInode, checks::OptionalInodeAttribs>& error,
+         FsckErrCount& errCount, UserPrompter& prompt);
+      void repairWrongDirAttribs(std::pair<FsckDirInode, checks::InodeAttribs>& error,
+         FsckErrCount& errCount, UserPrompter& prompt);
+      void repairFileWithMissingTargets(db::DirEntry& entry, FsckErrCount& errCount,
+         UserPrompter& prompt);
+      void repairDirEntryWithBrokenByIDFile(db::DirEntry& entry, FsckErrCount& errCount,
+         UserPrompter& prompt);
+      void repairOrphanedDentryByIDFile(FsckFsID& id, FsckErrCount& errCount, UserPrompter& prompt);
+      void repairChunkWithWrongPermissions(std::pair<FsckChunk, FsckFileInode>& error,
+         FsckErrCount& errCount, UserPrompter& prompt);
+      void repairWrongChunkPath(std::pair<FsckChunk, FsckFileInode>& error, FsckErrCount& errCount,
+         UserPrompter& prompt);
+      void updateOldStyledHardlinks(db::FileInode& inode, FsckErrCount& errCount, UserPrompter& prompt);
+
+      void deleteFsIDsFromDB(FsckDirEntryList& dentries);
+      void deleteFilesFromDB(FsckDirEntryList& dentries);
+
+      bool ensureLostAndFoundExists();
+      void releaseLostAndFound();
+};
+
+#endif /* MODECHECKFS_H */

fsck/source/modes/ModeEnableQuota.cpp

@@ -0,0 +1,116 @@
+#include "ModeEnableQuota.h"
+
+#include <common/toolkit/ListTk.h>
+#include <common/toolkit/UnitTk.h>
+#include <components/worker/AdjustChunkPermissionsWork.h>
+#include <net/msghelpers/MsgHelperRepair.h>
+#include <toolkit/FsckTkEx.h>
+
+#include <program/Program.h>
+
+ModeEnableQuota::ModeEnableQuota()
+ : log("ModeEnableQuota")
+{
+}
+
+void ModeEnableQuota::printHelp()
+{
+   std::cout << "MODE ARGUMENTS:" << std::endl;
+   std::cout << " Optional:" << std::endl;
+   std::cout << "  --databasePath=<path>  Path to store the database files." << std::endl;
+   std::cout << "                         (Default: " << CONFIG_DEFAULT_DBPATH << ")" << std::endl;
+   std::cout << "  --overwriteDbFile      Overwrite an existing database file without prompt." << std::endl;
+   std::cout << "  --logOutFile=<path>    Path to the fsck output file, which contains a copy of" << std::endl;
+   std::cout << "                         the console output." << std::endl;
+   std::cout << "                         (Default: " << CONFIG_DEFAULT_OUTFILE << ")" << std::endl;
+   std::cout << "  --logStdFile=<path>    Path to the program error log file, which contains e.g." << std::endl;
+   std::cout << "                         network error messages." << std::endl;
+   std::cout << "                         (Default: " << CONFIG_DEFAULT_LOGFILE << ")" << std::endl;
+   std::cout << std::endl;
+   std::cout << "USAGE:" << std::endl;
+   std::cout << " This mode sets quota information on an existing file system." << std::endl;
+   std::cout << std::endl;
+   std::cout << " This is useful when quota support is being enabled on a file system instance" << std::endl;
+   std::cout << " that was previously used without quota support." << std::endl;
+   std::cout << std::endl;
+   std::cout << " Example: Set quota information" << std::endl;
+   std::cout << "  $ beegfs-fsck --enablequota" << std::endl;
+}
+
+int ModeEnableQuota::execute()
+{
+   App* app = Program::getApp();
+   Config *cfg = app->getConfig();
+
+   if(this->checkInvalidArgs(cfg->getUnknownConfigArgs()))
+      return APPCODE_INVALID_CONFIG;
+
+   FsckTkEx::printVersionHeader(false);
+   printHeaderInformation();
+
+   if ( !FsckTkEx::checkReachability() )
+      return APPCODE_COMMUNICATION_ERROR;
+
+   fixPermissions();
+
+   return APPCODE_NO_ERROR;
+}
+
+void ModeEnableQuota::printHeaderInformation()
+{
+   Config* cfg = Program::getApp()->getConfig();
+
+   // get the current time and create some nice output, so the user can see at which time the run
+   // was started (especially nice to find older runs in the log file)
+   time_t t;
+   time(&t);
+   std::string timeStr = std::string(ctime(&t));
+   FsckTkEx::fsckOutput(
+      "Started BeeGFS fsck in enableQuota mode [" + timeStr.substr(0, timeStr.length() - 1)
+         + "]\nLog will be written to " + cfg->getLogStdFile()
+         + "\nDatabase files will be saved in " + cfg->getDatabasePath(),
+         OutputOptions_LINEBREAK | OutputOptions_HEADLINE);
+}
+
+void ModeEnableQuota::fixPermissions()
+{
+   SynchronizedCounter finishedWork;
+   unsigned generatedWork = 0;
+
+   AtomicUInt64 fileCount;
+   AtomicUInt64 errorCount;
+
+   NodeStore* metaNodes = Program::getApp()->getMetaNodes();
+
+   auto nodes = metaNodes->referenceAllNodes();
+
+   for (auto iter = nodes.begin(); iter != nodes.end(); iter++)
+   {
+      generatedWork++;
+      Program::getApp()->getWorkQueue()->addIndirectWork(new AdjustChunkPermissionsWork(**iter,
+         &finishedWork, &fileCount, &errorCount));
+   }
+
+   FsckTkEx::fsckOutput("Processed 0 entries", OutputOptions_ADDLINEBREAKBEFORE |
+      OutputOptions_LINEDELETE | OutputOptions_NOLOG);
+
+   // wait for all packages to finish, because we cannot proceed if not all data was fetched
+   // BUT : update output each OUTPUT_INTERVAL_MS ms
+   while (! finishedWork.timedWaitForCount(generatedWork, MODEENABLEQUOTA_OUTPUT_INTERVAL_MS) )
+   {
+      FsckTkEx::fsckOutput("Processed " + StringTk::uint64ToStr(fileCount.read()) + " entries",
+         OutputOptions_LINEDELETE | OutputOptions_NOLOG);
+   }
+
+   FsckTkEx::fsckOutput("Processed " + StringTk::uint64ToStr(fileCount.read()) + " entries",
+      OutputOptions_LINEDELETE | OutputOptions_NOLOG | OutputOptions_LINEBREAK);
+
+   if ( errorCount.read() > 0 )
+   {
+      FsckTkEx::fsckOutput("", OutputOptions_LINEBREAK);
+      FsckTkEx::fsckOutput(
+         StringTk::uint64ToStr(errorCount.read())
+            + " errors occurred. Please see Metadata server logs for more details.",
+         OutputOptions_ADDLINEBREAKBEFORE | OutputOptions_LINEBREAK);
+   }
+}

fsck/source/modes/ModeEnableQuota.h

@@ -0,0 +1,27 @@
+#ifndef MODEENABLEQUOTA_H
+#define MODEENABLEQUOTA_H
+
+#include <app/config/Config.h>
+#include <database/FsckDB.h>
+#include <modes/Mode.h>
+
+#define MODEENABLEQUOTA_OUTPUT_INTERVAL_MS 5000
+
+class ModeEnableQuota : public Mode
+{
+   public:
+      ModeEnableQuota();
+
+      static void printHelp();
+
+      virtual int execute();
+
+   private:
+      LogContext log;
+
+      void printHeaderInformation();
+
+      void fixPermissions();
+};
+
+#endif /* MODEENABLEQUOTA_H */

fsck/source/modes/ModeHelp.cpp

@@ -0,0 +1,99 @@
+#include "ModeHelp.h"
+
+#include <program/Program.h>
+
+#include <modes/ModeCheckFS.h>
+#include <modes/ModeEnableQuota.h>
+
+int ModeHelp::execute()
+{
+   App* app = Program::getApp();
+   Config* cfg = app->getConfig();
+
+   RunMode runMode = cfg->determineRunMode();
+
+   if(runMode == RunMode_INVALID)
+      printGeneralHelp();
+   else
+      printSpecificHelp(runMode);
+
+   std::cout << std::endl;
+
+   return APPCODE_NO_ERROR;
+}
+
+void ModeHelp::printGeneralHelp()
+{
+   std::cout << "BeeGFS File System Check (http://www.beegfs.com)" << std::endl;
+   std::cout << "Version: " << BEEGFS_VERSION << std::endl;
+   std::cout << std::endl;
+   std::cout << "GENERAL USAGE:" << std::endl;
+   std::cout << " $ beegfs-fsck --<modename> --help" << std::endl;
+   std::cout << " $ beegfs-fsck --<modename> [mode_arguments] [client_arguments]" << std::endl;
+   std::cout << std::endl;
+   std::cout << "MODES:" << std::endl;
+   std::cout << " --checkfs     => Perform a full check and optional repair of " << std::endl;
+   std::cout << "                  a BeeGFS instance." << std::endl;
+   std::cout << " --enablequota => Set attributes needed for quota support in FhGFS." << std::endl;
+   std::cout << "                  Can be used to enable quota support on an existing" << std::endl;
+   std::cout << "                  system." << std::endl;
+   std::cout << std::endl;
+   std::cout << "USAGE:" << std::endl;
+   std::cout << " This is the BeeGFS file system consistency check and repair tool." << std::endl;
+   std::cout << std::endl;
+   std::cout << " Choose a mode from the list above and use the parameter \"--help\"" << std::endl;
+   std::cout << " to show arguments and usage examples for that particular mode." << std::endl;
+   std::cout << std::endl;
+   std::cout << " (Running beegfs-fsck requires root privileges.)" << std::endl;
+   std::cout << std::endl;
+   std::cout << " Example: Show help for mode \"--checkfs\"" << std::endl;
+   std::cout << "  $ beegfs-fsck --checkfs --help" << std::endl;
+}
+
+
+/**
+ * Note: don't call this with RunMode_INVALID.
+ */
+void ModeHelp::printSpecificHelp(RunMode runMode)
+{
+   printSpecificHelpHeader(); // print general usage and client options info
+
+   switch(runMode)
+   {
+      case RunMode_CHECKFS:
+      {
+         ModeCheckFS::printHelp();
+      } break;
+
+      case RunMode_ENABLEQUOTA:
+      {
+         ModeEnableQuota::printHelp();
+      } break;
+
+      default:
+      {
+         std::cerr << "Error: Unhandled mode specified. Mode number: " << runMode << std::endl;
+         std::cout << std::endl;
+
+         printGeneralHelp();
+      } break;
+   }
+
+}
+
+/**
+ * Print the help message header that applies to any specific mode help. Contains general usage
+ * and client options info.
+ */
+void ModeHelp::printSpecificHelpHeader()
+{
+   std::cout << "GENERAL USAGE:" << std::endl;
+   std::cout << "  $ beegfs-fsck --<modename> [mode_arguments] [client_arguments]" << std::endl;
+   std::cout << std::endl;
+   std::cout << "CLIENT ARGUMENTS:" << std::endl;
+   std::cout << " --cfgFile=<path>         Path to BeeGFS client config file." << std::endl;
+   std::cout << "                          (Default: " CONFIG_DEFAULT_CFGFILENAME ")" << std::endl;
+   std::cout << " --<key>=<value>          Any setting from the client config file to override" << std::endl;
+   std::cout << "                          the config file values (e.g. \"--logLevel=5\")." << std::endl;
+   std::cout << std::endl;
+}

fsck/source/modes/ModeHelp.h

@@ -0,0 +1,24 @@
+#ifndef MODEHELP_H_
+#define MODEHELP_H_
+
+#include <app/config/Config.h>
+#include <common/Common.h>
+
+#include "Mode.h"
+
+
+class ModeHelp : public Mode
+{
+   public:
+      ModeHelp() {};
+
+      virtual int execute();
+
+   private:
+      void printGeneralHelp();
+      void printSpecificHelp(RunMode runMode);
+      void printSpecificHelpHeader();
+};
+
+
+#endif /*MODEHELP_H_*/

fsck/source/net/message/NetMessageFactory.cpp

@@ -0,0 +1,125 @@
+// control messages
+#include <common/net/message/control/GenericResponseMsg.h>
+
+// fsck messages
+#include <common/net/message/fsck/AdjustChunkPermissionsRespMsg.h>
+#include <common/net/message/fsck/CreateDefDirInodesRespMsg.h>
+#include <common/net/message/fsck/CreateEmptyContDirsRespMsg.h>
+#include <common/net/message/fsck/DeleteChunksRespMsg.h>
+#include <common/net/message/fsck/DeleteDirEntriesRespMsg.h>
+#include <common/net/message/fsck/FetchFsckChunkListRespMsg.h>
+#include <common/net/message/fsck/FsckSetEventLoggingRespMsg.h>
+#include <common/net/message/fsck/FixInodeOwnersInDentryRespMsg.h>
+#include <common/net/message/fsck/FixInodeOwnersRespMsg.h>
+#include <common/net/message/fsck/FsckSetEventLoggingRespMsg.h>
+#include <common/net/message/fsck/LinkToLostAndFoundRespMsg.h>
+#include <common/net/message/fsck/MoveChunkFileRespMsg.h>
+#include <common/net/message/fsck/RecreateDentriesRespMsg.h>
+#include <common/net/message/fsck/RecreateFsIDsRespMsg.h>
+#include <common/net/message/fsck/RemoveInodesRespMsg.h>
+#include <common/net/message/fsck/RetrieveDirEntriesRespMsg.h>
+#include <common/net/message/fsck/RetrieveFsIDsRespMsg.h>
+#include <common/net/message/fsck/RetrieveInodesRespMsg.h>
+#include <common/net/message/fsck/UpdateDirAttribsRespMsg.h>
+#include <common/net/message/fsck/UpdateFileAttribsRespMsg.h>
+#include <net/message/fsck/FsckModificationEventMsgEx.h>
+#include <common/net/message/fsck/CheckAndRepairDupInodeRespMsg.h>
+
+// nodes messages
+#include <common/net/message/nodes/GetMirrorBuddyGroupsRespMsg.h>
+#include <common/net/message/nodes/GetNodesRespMsg.h>
+#include <common/net/message/nodes/GetTargetMappingsRespMsg.h>
+#include <common/net/message/nodes/GetTargetStatesRespMsg.h>
+#include <common/net/message/nodes/SetTargetConsistencyStatesRespMsg.h>
+#include <net/message/nodes/HeartbeatMsgEx.h>
+
+// storage messages
+#include <common/net/message/storage/attribs/GetEntryInfoRespMsg.h>
+#include <common/net/message/storage/attribs/SetLocalAttrRespMsg.h>
+#include <common/net/message/storage/attribs/StatRespMsg.h>
+#include <common/net/message/storage/creating/MkDirRespMsg.h>
+#include <common/net/message/storage/creating/RmDirEntryRespMsg.h>
+#include <common/net/message/storage/creating/UnlinkFileRespMsg.h>
+#include <common/net/message/storage/listing/ListDirFromOffsetRespMsg.h>
+#include <common/net/message/storage/lookup/FindOwnerRespMsg.h>
+#include <common/net/message/storage/StatStoragePathRespMsg.h>
+#include <common/net/message/storage/creating/MoveFileInodeRespMsg.h>
+
+// general includes
+#include <common/net/message/SimpleMsg.h>
+#include <net/message/testing/DummyMsgEx.h>
+
+#include "NetMessageFactory.h"
+
+/**
+ * @return NetMessage that must be deleted by the caller
+ * (msg->msgType is NETMSGTYPE_Invalid on error)
+ */
+std::unique_ptr<NetMessage> NetMessageFactory::createFromMsgType(unsigned short msgType) const
+{
+   NetMessage* msg;
+
+   switch(msgType)
+   {
+      // The following lines are grouped by "type of the message" and ordered alphabetically inside
+      // the groups. There should always be one message per line to keep a clear layout (although
+      // this might lead to lines that are longer than usual)
+
+      // control messages
+      case NETMSGTYPE_GenericResponse: { msg = new GenericResponseMsg(); } break;
+
+      // nodes messages
+      case NETMSGTYPE_GetNodesResp: { msg = new GetNodesRespMsg(); } break;
+      case NETMSGTYPE_Heartbeat: { msg = new HeartbeatMsgEx(); } break;
+      case NETMSGTYPE_GetMirrorBuddyGroupsResp: { msg = new GetMirrorBuddyGroupsRespMsg(); } break;
+      case NETMSGTYPE_GetTargetMappingsResp: { msg = new GetTargetMappingsRespMsg(); } break;
+      case NETMSGTYPE_GetTargetStatesResp: { msg = new GetTargetStatesRespMsg(); } break;
+      case NETMSGTYPE_SetTargetConsistencyStatesResp: { msg = new SetTargetConsistencyStatesRespMsg(); } break;
+
+      // storage messages
+      case NETMSGTYPE_FindOwnerResp: { msg = new FindOwnerRespMsg(); } break;
+      case NETMSGTYPE_ListDirFromOffsetResp: { msg = new ListDirFromOffsetRespMsg(); } break;
+      case NETMSGTYPE_RmDirEntryResp: { msg = new RmDirEntryRespMsg(); } break;
+      case NETMSGTYPE_GetEntryInfoResp: { msg = new GetEntryInfoRespMsg(); } break;
+      case NETMSGTYPE_StatResp: { msg = new StatRespMsg(); } break;
+      case NETMSGTYPE_StatStoragePathResp: { msg = new StatStoragePathRespMsg(); } break;
+      case NETMSGTYPE_MkDirResp: { msg = new MkDirRespMsg(); } break;
+      case NETMSGTYPE_SetLocalAttrResp: { msg = new SetLocalAttrRespMsg(); } break;
+      case NETMSGTYPE_UnlinkFileResp: { msg = new UnlinkFileRespMsg(); } break;
+      case NETMSGTYPE_MoveFileInodeResp: {msg = new MoveFileInodeRespMsg(); } break;
+
+      //fsck
+      case NETMSGTYPE_RetrieveDirEntriesResp: { msg = new RetrieveDirEntriesRespMsg(); } break;
+      case NETMSGTYPE_RetrieveInodesResp: { msg = new RetrieveInodesRespMsg(); } break;
+      case NETMSGTYPE_RetrieveFsIDsResp: { msg = new RetrieveFsIDsRespMsg(); } break;
+      case NETMSGTYPE_DeleteDirEntriesResp: { msg = new DeleteDirEntriesRespMsg(); } break;
+      case NETMSGTYPE_CreateDefDirInodesResp: { msg = new CreateDefDirInodesRespMsg(); } break;
+      case NETMSGTYPE_FixInodeOwnersResp: { msg = new FixInodeOwnersRespMsg(); } break;
+      case NETMSGTYPE_FixInodeOwnersInDentryResp: { msg = new FixInodeOwnersInDentryRespMsg(); } break;
+      case NETMSGTYPE_LinkToLostAndFoundResp : { msg = new LinkToLostAndFoundRespMsg(); } break;
+      case NETMSGTYPE_DeleteChunksResp: { msg = new DeleteChunksRespMsg(); } break;
+      case NETMSGTYPE_CreateEmptyContDirsResp: { msg = new CreateEmptyContDirsRespMsg(); } break;
+      case NETMSGTYPE_UpdateFileAttribsResp: { msg = new UpdateFileAttribsRespMsg(); } break;
+      case NETMSGTYPE_UpdateDirAttribsResp: { msg = new UpdateDirAttribsRespMsg(); } break;
+      case NETMSGTYPE_RecreateFsIDsResp: { msg = new RecreateFsIDsRespMsg(); } break;
+      case NETMSGTYPE_RecreateDentriesResp: { msg = new RecreateDentriesRespMsg(); } break;
+      case NETMSGTYPE_FsckSetEventLoggingResp: { msg = new FsckSetEventLoggingRespMsg(); } break;
+      case NETMSGTYPE_FsckModificationEvent: { msg = new FsckModificationEventMsgEx(); } break;
+      case NETMSGTYPE_AdjustChunkPermissionsResp: { msg = new AdjustChunkPermissionsRespMsg(); } break;
+      case NETMSGTYPE_FetchFsckChunkListResp: { msg = new FetchFsckChunkListRespMsg(); } break;
+      case NETMSGTYPE_MoveChunkFileResp: { msg = new MoveChunkFileRespMsg(); } break;
+      case NETMSGTYPE_RemoveInodesResp: { msg = new RemoveInodesRespMsg(); } break;
+      case NETMSGTYPE_CheckAndRepairDupInodeResp: { msg = new CheckAndRepairDupInodeRespMsg(); } break;
+
+      //testing
+      case NETMSGTYPE_Dummy: { msg = new DummyMsgEx(); } break;
+
+      default:
+      {
+         msg = new SimpleMsg(NETMSGTYPE_Invalid);
+      } break;
+   }
+
+   return std::unique_ptr<NetMessage>(msg);
+}
+

fsck/source/net/message/NetMessageFactory.h

@@ -0,0 +1,16 @@
+#ifndef NETMESSAGEFACTORY_H_
+#define NETMESSAGEFACTORY_H_
+
+#include <common/Common.h>
+#include <common/net/message/AbstractNetMessageFactory.h>
+
+class NetMessageFactory : public AbstractNetMessageFactory
+{
+   public:
+      NetMessageFactory() {}
+
+   protected:
+      virtual std::unique_ptr<NetMessage> createFromMsgType(unsigned short msgType) const override;
+} ;
+
+#endif /*NETMESSAGEFACTORY_H_*/

fsck/source/net/message/fsck/FsckModificationEventMsgEx.cpp

@@ -0,0 +1,26 @@
+#include <program/Program.h>
+
+#include "FsckModificationEventMsgEx.h"
+
+bool FsckModificationEventMsgEx::processIncoming(ResponseContext& ctx)
+{
+   LogContext log("FsckModificationEventMsg incoming");
+
+   App* app = Program::getApp();
+   ModificationEventHandler *eventHandler = app->getModificationEventHandler();
+
+   StringList& entryIDList = this->getEntryIDList();
+
+   bool addRes = eventHandler->add(getModificationEventTypeList(), entryIDList);
+   if (!addRes)
+   {
+      log.logErr("Unable to add modification event to database");
+      return false;
+   }
+
+   bool ackRes = acknowledge(ctx);
+   if (!ackRes)
+      log.logErr("Unable to send ack to metadata server");
+
+   return true;
+}

fsck/source/net/message/fsck/FsckModificationEventMsgEx.h

@@ -0,0 +1,12 @@
+#ifndef FSCKMODIFICATIONEVENTMSGEX_H
+#define FSCKMODIFICATIONEVENTMSGEX_H
+
+#include <common/net/message/fsck/FsckModificationEventMsg.h>
+
+class FsckModificationEventMsgEx : public FsckModificationEventMsg
+{
+   public:
+      virtual bool processIncoming(ResponseContext& ctx);
+};
+
+#endif /*FSCKMODIFICATIONEVENTMSGEX_H*/

fsck/source/net/message/nodes/HeartbeatMsgEx.cpp

@@ -0,0 +1,102 @@
+#include <common/net/sock/NetworkInterfaceCard.h>
+#include <program/Program.h>
+#include "HeartbeatMsgEx.h"
+
+#include <boost/lexical_cast.hpp>
+
+bool HeartbeatMsgEx::processIncoming(ResponseContext& ctx)
+{
+   LogContext log("Heartbeat incoming");
+
+   App* app = Program::getApp();
+
+   bool isNodeNew;
+
+   // construct node
+   NicAddressList& nicList = getNicList();
+
+   auto node = std::make_shared<Node>(getNodeType(), getNodeID(), getNodeNumID(), getPortUDP(),
+         getPortTCP(), nicList);
+
+   // set local nic capabilities
+   Node& localNode = app->getLocalNode();
+   NicAddressList localNicList(localNode.getNicList());
+   NicListCapabilities localNicCaps;
+
+   NetworkInterfaceCard::supportedCapabilities(&localNicList, &localNicCaps);
+   node->getConnPool()->setLocalNicList(localNicList, localNicCaps);
+
+   std::string nodeIDWithTypeStr = node->getNodeIDWithTypeStr();
+
+
+   // add/update node in store
+   NodeStore* nodes = NULL;
+
+   switch(getNodeType() )
+   {
+      case NODETYPE_Meta:
+         nodes = app->getMetaNodes(); break;
+
+      case NODETYPE_Storage:
+         nodes = app->getStorageNodes(); break;
+
+
+      case NODETYPE_Mgmt:
+         nodes = app->getMgmtNodes(); break;
+
+      default:
+      {
+         LOG(GENERAL, ERR, "Invalid node type.",
+               ("Node Type", getNodeType()),
+               ("Sender", ctx.peerName()),
+               ("NodeID", getNodeNumID()),
+               ("Port (UDP)", getPortUDP()),
+               ("Port (TCP)", getPortTCP())
+            );
+
+         goto ack_resp;
+      } break;
+   }
+
+
+   isNodeNew = (nodes->addOrUpdateNode(std::move(node)) == NodeStoreResult::Added);
+   if(isNodeNew)
+   {
+      bool supportsRDMA = NetworkInterfaceCard::supportsRDMA(&nicList);
+
+      log.log(Log_WARNING, std::string("New node: ") +
+         nodeIDWithTypeStr + "; " +
+         std::string(supportsRDMA ? "RDMA; " : "") +
+         std::string("Source: ") + ctx.peerName() );
+
+      log.log(Log_DEBUG, std::string("Number of nodes in the system: ") +
+         StringTk::intToStr(nodes->getSize() ) +
+         std::string(" (Type: ") + boost::lexical_cast<std::string>(getNodeType()) + ")");
+   }
+
+
+   processIncomingRoot();
+
+ack_resp:
+   acknowledge(ctx);
+
+   return true;
+}
+
+/**
+ * Handles the contained root information.
+ */
+void HeartbeatMsgEx::processIncomingRoot()
+{
+   LogContext log("Heartbeat incoming");
+
+   // check whether root info is defined
+   if( (getNodeType() != NODETYPE_Meta) || !getRootNumID() )
+      return;
+
+   // try to apply the contained root info
+   if(Program::getApp()->getMetaRoot().setIfDefault(getRootNumID(), getRootIsBuddyMirrored()))
+   {
+      log.log(Log_CRITICAL, "Root (by Heartbeat): " + getRootNumID().str() );
+   }
+}

fsck/source/net/message/nodes/HeartbeatMsgEx.h

@@ -0,0 +1,15 @@
+#ifndef HEARTBEATMSGEX_H_
+#define HEARTBEATMSGEX_H_
+
+#include <common/net/message/nodes/HeartbeatMsg.h>
+
+class HeartbeatMsgEx : public HeartbeatMsg
+{
+   public:
+      virtual bool processIncoming(ResponseContext& ctx);
+
+   private:
+      void processIncomingRoot();
+};
+
+#endif /*HEARTBEATMSGEX_H_*/

fsck/source/net/message/testing/DummyMsgEx.cpp

@@ -0,0 +1,9 @@
+#include <program/Program.h>
+#include "DummyMsgEx.h"
+
+bool DummyMsgEx::processIncoming(ResponseContext& ctx)
+{
+   ctx.sendResponse(DummyMsg() );
+
+   return true;
+}

fsck/source/net/message/testing/DummyMsgEx.h

@@ -0,0 +1,20 @@
+#ifndef DUMMYMSGEX_H
+#define DUMMYMSGEX_H
+
+#include <common/net/message/control/DummyMsg.h>
+
+/*
+ * this is intended for testing purposes only
+ * just sends another DummyMsg back to the sender
+ */
+
+class DummyMsgEx : public DummyMsg
+{
+   public:
+      virtual bool processIncoming(ResponseContext& ctx);
+
+   private:
+      void processIncomingRoot();
+};
+
+#endif /*DUMMYMSGEX_H*/

fsck/source/net/msghelpers/MsgHelperRepair.cpp

@@ -0,0 +1,951 @@
+#include <common/net/message/fsck/CreateDefDirInodesMsg.h>
+#include <common/net/message/fsck/CreateDefDirInodesRespMsg.h>
+#include <common/net/message/fsck/CreateEmptyContDirsMsg.h>
+#include <common/net/message/fsck/CreateEmptyContDirsRespMsg.h>
+#include <common/net/message/fsck/DeleteDirEntriesMsg.h>
+#include <common/net/message/fsck/DeleteDirEntriesRespMsg.h>
+#include <common/net/message/fsck/DeleteChunksMsg.h>
+#include <common/net/message/fsck/DeleteChunksRespMsg.h>
+#include <common/net/message/fsck/FixInodeOwnersMsg.h>
+#include <common/net/message/fsck/FixInodeOwnersRespMsg.h>
+#include <common/net/message/fsck/FixInodeOwnersInDentryMsg.h>
+#include <common/net/message/fsck/FixInodeOwnersInDentryRespMsg.h>
+#include <common/net/message/fsck/LinkToLostAndFoundMsg.h>
+#include <common/net/message/fsck/LinkToLostAndFoundRespMsg.h>
+#include <common/net/message/fsck/MoveChunkFileMsg.h>
+#include <common/net/message/fsck/MoveChunkFileRespMsg.h>
+#include <common/net/message/fsck/RecreateFsIDsMsg.h>
+#include <common/net/message/fsck/RecreateFsIDsRespMsg.h>
+#include <common/net/message/fsck/RecreateDentriesMsg.h>
+#include <common/net/message/fsck/RecreateDentriesRespMsg.h>
+#include <common/net/message/fsck/RemoveInodesMsg.h>
+#include <common/net/message/fsck/RemoveInodesRespMsg.h>
+#include <common/net/message/fsck/CheckAndRepairDupInodeMsg.h>
+#include <common/net/message/fsck/CheckAndRepairDupInodeRespMsg.h>
+
+#include <common/net/message/fsck/UpdateFileAttribsMsg.h>
+#include <common/net/message/fsck/UpdateFileAttribsRespMsg.h>
+#include <common/net/message/fsck/UpdateDirAttribsMsg.h>
+#include <common/net/message/fsck/UpdateDirAttribsRespMsg.h>
+
+#include <common/net/message/nodes/SetTargetConsistencyStatesMsg.h>
+#include <common/net/message/nodes/SetTargetConsistencyStatesRespMsg.h>
+
+#include <common/net/message/storage/creating/MkDirMsg.h>
+#include <common/net/message/storage/creating/MkDirRespMsg.h>
+#include <common/net/message/storage/creating/UnlinkFileMsg.h>
+#include <common/net/message/storage/creating/UnlinkFileRespMsg.h>
+#include <common/net/message/storage/attribs/SetLocalAttrMsg.h>
+#include <common/net/message/storage/attribs/SetLocalAttrRespMsg.h>
+#include <common/net/message/storage/creating/MoveFileInodeMsg.h>
+#include <common/net/message/storage/creating/MoveFileInodeRespMsg.h>
+
+#include <common/toolkit/MessagingTk.h>
+#include <common/toolkit/MetadataTk.h>
+
+#include <program/Program.h>
+
+#include "MsgHelperRepair.h"
+
+#include <boost/lexical_cast.hpp>
+
+template<typename RepairItemT>
+static bool setSecondaryBad(uint16_t groupID, std::set<NumNodeID>& secondariesWithRepair,
+      const std::list<RepairItemT>& args, std::list<RepairItemT>& failed)
+{
+   auto* bgm = Program::getApp()->getMetaMirrorBuddyGroupMapper();
+   NumNodeID secondary(bgm->getSecondaryTargetID(groupID));
+
+   if (secondariesWithRepair.count(secondary))
+      return true;
+
+   auto setRes = MsgHelperRepair::setNodeState(secondary, TargetConsistencyState_BAD);
+   if (setRes == FhgfsOpsErr_SUCCESS)
+   {
+      secondariesWithRepair.insert(NumNodeID(bgm->getSecondaryTargetID(groupID)));
+      return true;
+   }
+
+   LOG(GENERAL, ERR, "Failed to set secondary consistency state, not attempting repair action.",
+         groupID, setRes);
+   failed = args;
+   return false;
+}
+
+
+
+FhgfsOpsErr MsgHelperRepair::setNodeState(NumNodeID node, TargetConsistencyState state)
+{
+   std::list<uint16_t> targets(1, node.val());
+   std::list<uint8_t> states(1, state);
+   SetTargetConsistencyStatesMsg msg(NODETYPE_Meta, &targets, &states, false);
+
+   {
+      auto secondary = Program::getApp()->getMetaNodes()->referenceNode(node);
+
+      MessagingTk::requestResponse(*secondary, msg, NETMSGTYPE_SetTargetConsistencyStatesResp);
+   }
+
+   {
+      auto mgmt = Program::getApp()->getMgmtNodes()->referenceFirstNode();
+
+      const auto respMsg = MessagingTk::requestResponse(*mgmt, msg,
+            NETMSGTYPE_SetTargetConsistencyStatesResp);
+      if (!respMsg)
+         return FhgfsOpsErr_COMMUNICATION;
+
+      return static_cast<SetTargetConsistencyStatesRespMsg&>(*respMsg).getResult();
+   }
+}
+
+void MsgHelperRepair::deleteDanglingDirEntries(NumNodeID node, bool isBuddyMirrored,
+   FsckDirEntryList* dentries, FsckDirEntryList* failedDeletes,
+   std::set<NumNodeID>& secondariesWithRepair)
+{
+   DeleteDirEntriesMsg deleteDirEntriesMsg(dentries);
+
+   RequestResponseNode rrNode(node, Program::getApp()->getMetaNodes());
+   RequestResponseArgs rrArgs(nullptr, &deleteDirEntriesMsg, NETMSGTYPE_DeleteDirEntriesResp);
+
+   if (isBuddyMirrored)
+   {
+      rrNode.setMirrorInfo(Program::getApp()->getMetaMirrorBuddyGroupMapper(), false);
+      if (!setSecondaryBad(node.val(), secondariesWithRepair, *dentries, *failedDeletes))
+         return;
+   }
+
+   FhgfsOpsErr commRes = MessagingTk::requestResponseNode(&rrNode, &rrArgs);
+
+   if (commRes == FhgfsOpsErr_SUCCESS)
+   {
+      auto* deleteDirEntriesRespMsg = (DeleteDirEntriesRespMsg*) rrArgs.outRespMsg.get();
+
+      deleteDirEntriesRespMsg->getFailedEntries().swap(*failedDeletes);
+
+      if (! failedDeletes->empty())
+      {
+         for (auto iter = failedDeletes->begin(); iter != failedDeletes->end(); iter++)
+            LOG(GENERAL, CRITICAL, "Failed to delete directory entry from metadata node.",
+                 node, isBuddyMirrored, ("entryID", iter->getID()));
+      }
+   }
+   else
+   {
+      *failedDeletes = *dentries;
+
+      LOG(GENERAL, ERR, "Communication error occured.", node, isBuddyMirrored, commRes);
+   }
+}
+
+void MsgHelperRepair::createDefDirInodes(NumNodeID node, bool isBuddyMirrored,
+   const std::vector<std::tuple<std::string, bool>>& entries, FsckDirInodeList* createdInodes,
+   std::set<NumNodeID>& secondariesWithRepair)
+{
+   StringList failedInodeIDs;
+
+   CreateDefDirInodesMsg createDefDirInodesMsgEx(entries);
+
+   RequestResponseNode rrNode(node, Program::getApp()->getMetaNodes());
+   RequestResponseArgs rrArgs(nullptr, &createDefDirInodesMsgEx, NETMSGTYPE_CreateDefDirInodesResp);
+
+   if (isBuddyMirrored)
+   {
+      rrNode.setMirrorInfo(Program::getApp()->getMetaMirrorBuddyGroupMapper(), false);
+      if (!setSecondaryBad(node.val(), secondariesWithRepair, {}, *createdInodes))
+      {
+         for (auto it = entries.begin(); it != entries.end(); ++it)
+            failedInodeIDs.push_back(std::get<0>(*it));
+         return;
+      }
+   }
+
+   FhgfsOpsErr commRes = MessagingTk::requestResponseNode(&rrNode, &rrArgs);
+
+   if (commRes == FhgfsOpsErr_SUCCESS)
+   {
+      auto* createDefDirInodesRespMsg = (CreateDefDirInodesRespMsg*) rrArgs.outRespMsg.get();
+
+      createDefDirInodesRespMsg->getFailedInodeIDs().swap(failedInodeIDs);
+      createDefDirInodesRespMsg->getCreatedInodes().swap(*createdInodes);
+   }
+   else
+   {
+      for (auto it = entries.begin(); it != entries.end(); ++it)
+         failedInodeIDs.push_back(std::get<0>(*it));
+
+      LOG(GENERAL, ERR, "Communication error occured.", node, isBuddyMirrored, commRes);
+   }
+
+   if (! failedInodeIDs.empty())
+   {
+      for (StringListIter iter = failedInodeIDs.begin(); iter != failedInodeIDs.end(); iter++)
+         LOG(GENERAL, CRITICAL, "Failed to create default directory inode.", node, isBuddyMirrored,
+               ("entryID", *iter));
+   }
+}
+
+void MsgHelperRepair::correctInodeOwnersInDentry(NumNodeID node, bool isBuddyMirrored,
+   FsckDirEntryList* dentries, NumNodeIDList* owners, FsckDirEntryList* failedCorrections,
+   std::set<NumNodeID>& secondariesWithRepair)
+{
+   FixInodeOwnersInDentryMsg fixInodeOwnersMsg(*dentries, *owners);
+
+   RequestResponseNode rrNode(node, Program::getApp()->getMetaNodes());
+   RequestResponseArgs rrArgs(nullptr, &fixInodeOwnersMsg, NETMSGTYPE_FixInodeOwnersInDentryResp);
+
+   if (isBuddyMirrored)
+   {
+      rrNode.setMirrorInfo(Program::getApp()->getMetaMirrorBuddyGroupMapper(), false);
+      if (!setSecondaryBad(node.val(), secondariesWithRepair, *dentries, *failedCorrections))
+         return;
+   }
+
+   FhgfsOpsErr commRes = MessagingTk::requestResponseNode(&rrNode, &rrArgs);
+
+   if (commRes == FhgfsOpsErr_SUCCESS)
+   {
+      auto* fixInodeOwnersRespMsg = (FixInodeOwnersInDentryRespMsg*) rrArgs.outRespMsg.get();
+
+      fixInodeOwnersRespMsg->getFailedEntries().swap(*failedCorrections);
+   }
+   else
+   {
+      *failedCorrections = *dentries;
+
+      LOG(GENERAL, ERR, "Communication error occured.", node, isBuddyMirrored, commRes);
+   }
+
+   if (! failedCorrections->empty())
+   {
+      for (auto iter = failedCorrections->begin(); iter != failedCorrections->end(); iter++)
+         LOG(GENERAL, CRITICAL, "Failed to correct inode owner information in dentry.",
+               node, isBuddyMirrored, ("entryID", iter->getID()));
+   }
+}
+
+void MsgHelperRepair::correctInodeOwners(NumNodeID node, bool isBuddyMirrored,
+   FsckDirInodeList* dirInodes, FsckDirInodeList* failedCorrections,
+   std::set<NumNodeID>& secondariesWithRepair)
+{
+   FixInodeOwnersMsg fixInodeOwnersMsg(dirInodes);
+
+   RequestResponseNode rrNode(node, Program::getApp()->getMetaNodes());
+   RequestResponseArgs rrArgs(nullptr, &fixInodeOwnersMsg, NETMSGTYPE_FixInodeOwnersResp);
+
+   if (isBuddyMirrored)
+   {
+      rrNode.setMirrorInfo(Program::getApp()->getMetaMirrorBuddyGroupMapper(), false);
+      if (!setSecondaryBad(node.val(), secondariesWithRepair, *dirInodes, *failedCorrections))
+         return;
+   }
+
+   FhgfsOpsErr commRes = MessagingTk::requestResponseNode(&rrNode, &rrArgs);
+
+   if (commRes == FhgfsOpsErr_SUCCESS)
+   {
+      auto* fixInodeOwnersRespMsg = (FixInodeOwnersRespMsg*) rrArgs.outRespMsg.get();
+
+      fixInodeOwnersRespMsg->getFailedInodes().swap(*failedCorrections);
+   }
+   else
+   {
+      *failedCorrections = *dirInodes;
+
+      LOG(GENERAL, ERR, "Communication error occured.", node, isBuddyMirrored, commRes);
+   }
+
+   if (! failedCorrections->empty())
+   {
+      for (auto iter = failedCorrections->begin(); iter != failedCorrections->end(); iter++)
+         LOG(GENERAL, CRITICAL, "Failed to correct inode owner information.", node, isBuddyMirrored,
+              ("entryID", iter->getID()));
+   }
+}
+
+void MsgHelperRepair::deleteFiles(NumNodeID node, bool isBuddyMirrored, FsckDirEntryList* dentries,
+   FsckDirEntryList* failedDeletes)
+{
+   const char* logContext = "MsgHelperRepair (deleteFiles)";
+
+   for ( FsckDirEntryListIter iter = dentries->begin(); iter != dentries->end(); iter++ )
+   {
+      EntryInfo parentInfo(node, "", iter->getParentDirID(), "", DirEntryType_DIRECTORY, 0);
+
+      std::string entryName = iter->getName();
+      UnlinkFileMsg unlinkFileMsg(&parentInfo, entryName);
+
+      RequestResponseNode rrNode(node, Program::getApp()->getMetaNodes());
+      RequestResponseArgs rrArgs(nullptr, &unlinkFileMsg, NETMSGTYPE_UnlinkFileResp);
+
+      if (isBuddyMirrored)
+         rrNode.setMirrorInfo(Program::getApp()->getMetaMirrorBuddyGroupMapper(), false);
+
+      FhgfsOpsErr commRes = MessagingTk::requestResponseNode(&rrNode, &rrArgs);
+
+      if (commRes == FhgfsOpsErr_SUCCESS)
+      {
+         const auto unlinkFileRespMsg = (const UnlinkFileRespMsg*) rrArgs.outRespMsg.get();
+
+         // get result
+         int unlinkRes = unlinkFileRespMsg->getValue();
+
+         if ( unlinkRes )
+         {
+            LogContext(logContext).log(Log_CRITICAL,
+               "Failed to delete file; entryID: " + iter->getID());
+            failedDeletes->push_back(*iter);
+         }
+      }
+      else
+      {
+         failedDeletes->push_back(*iter);
+
+         LOG(GENERAL, ERR, "Communication error occured.", node, isBuddyMirrored, commRes);
+      }
+   }
+}
+
+void MsgHelperRepair::deleteChunks(Node& node, FsckChunkList* chunks, FsckChunkList* failedDeletes)
+{
+   const char* logContext = "MsgHelperRepair (deleteChunks)";
+
+   DeleteChunksMsg deleteChunksMsg(chunks);
+
+   const auto respMsg = MessagingTk::requestResponse(node, deleteChunksMsg,
+         NETMSGTYPE_DeleteChunksResp);
+
+   if (respMsg)
+   {
+      DeleteChunksRespMsg* deleteChunksRespMsg = (DeleteChunksRespMsg*) respMsg.get();
+
+      deleteChunksRespMsg->getFailedChunks().swap(*failedDeletes);
+
+      if (! failedDeletes->empty())
+      {
+         for (FsckChunkListIter iter = failedDeletes->begin(); iter != failedDeletes->end();
+            iter++)
+         {
+            LogContext(logContext).log(Log_CRITICAL, "Failed to delete chunk entry. targetID: " +
+               StringTk::uintToStr(iter->getTargetID()) + " chunkID: " + iter->getID());
+         }
+      }
+   }
+   else
+   {
+      *failedDeletes = *chunks;
+
+      LogContext(logContext).logErr("Communication error occured with node: " + node.getAlias());
+   }
+}
+
+NodeHandle MsgHelperRepair::referenceLostAndFoundOwner(EntryInfo* outLostAndFoundEntryInfo)
+{
+   const char* logContext = "MsgHelperRepair (referenceLostAndFoundOwner)";
+   App* app = Program::getApp();
+
+// find owner node
+   Path path("/" + std::string(META_LOSTANDFOUND_PATH));
+
+   NodeHandle ownerNode;
+
+   FhgfsOpsErr findRes = MetadataTk::referenceOwner(&path, app->getMetaNodes(), ownerNode,
+      outLostAndFoundEntryInfo, app->getMetaRoot(), app->getMetaMirrorBuddyGroupMapper());
+
+   if ( findRes != FhgfsOpsErr_SUCCESS )
+   {
+      LogContext(logContext).log(Log_DEBUG, "No owner node found for lost+found. Directory does not"
+         " seem to exist (yet)");
+   }
+
+   return ownerNode;
+}
+
+bool MsgHelperRepair::createLostAndFound(NodeHandle& outReferencedNode,
+   EntryInfo& outLostAndFoundEntryInfo)
+{
+   const char* logContext = "MsgHelperRepair (createLostAndFound)";
+   App* app = Program::getApp();
+   NodeStore* metaNodes = app->getMetaNodes();
+   bool retVal = false;
+
+   // get root owner node and entryInfo
+   NodeHandle rootNode;
+   Path rootPath("");
+   // rootPath.setAbsolute(true);
+   EntryInfo rootEntryInfo;
+
+   FhgfsOpsErr findRes = MetadataTk::referenceOwner(&rootPath, metaNodes,
+      rootNode, &rootEntryInfo, app->getMetaRoot(), app->getMetaMirrorBuddyGroupMapper());
+
+   if ( findRes != FhgfsOpsErr_SUCCESS )
+   {
+      LogContext(logContext).log(Log_CRITICAL, "Unable to reference metadata node for root "
+         "directory");
+      return false;
+   }
+
+   // create the directory
+   std::string lostFoundPathStr = META_LOSTANDFOUND_PATH;
+   UInt16List preferredNodes;
+
+   MkDirMsg mkDirMsg(&rootEntryInfo, lostFoundPathStr , 0, 0, S_IFDIR | S_IRWXU | S_IRWXG, 0000,
+      &preferredNodes);
+
+   RequestResponseNode rrNode(rootEntryInfo.getOwnerNodeID(), Program::getApp()->getMetaNodes());
+   RequestResponseArgs rrArgs(nullptr, &mkDirMsg, NETMSGTYPE_MkDirResp);
+
+   if (rootEntryInfo.getIsBuddyMirrored())
+      rrNode.setMirrorInfo(Program::getApp()->getMetaMirrorBuddyGroupMapper(), false);
+
+   FhgfsOpsErr commRes = MessagingTk::requestResponseNode(&rrNode, &rrArgs);
+
+   if (commRes != FhgfsOpsErr_SUCCESS)
+      return false;
+
+   auto* mkDirRespMsg = (MkDirRespMsg*) rrArgs.outRespMsg.get();
+
+   FhgfsOpsErr mkDirRes = (FhgfsOpsErr) mkDirRespMsg->getResult();
+   if ( mkDirRes != FhgfsOpsErr_SUCCESS )
+   {
+      LogContext(logContext).log(Log_CRITICAL,
+         "Node encountered an error: " + boost::lexical_cast<std::string>(mkDirRes));
+      return false;
+   }
+
+   // create seems to have succeeded
+   // copy is created because delete is called on mkDirRespMsg, but we still need this object
+   outLostAndFoundEntryInfo = *(mkDirRespMsg->getEntryInfo());
+
+   outReferencedNode = outLostAndFoundEntryInfo.getIsBuddyMirrored()
+      ? metaNodes->referenceNode(
+            NumNodeID(
+               app->getMetaMirrorBuddyGroupMapper()->getPrimaryTargetID(
+                  outLostAndFoundEntryInfo.getOwnerNodeID().val())))
+      : metaNodes->referenceNode(outLostAndFoundEntryInfo.getOwnerNodeID());
+   retVal = true;
+
+   return retVal;
+}
+
+void MsgHelperRepair::linkToLostAndFound(Node& lostAndFoundNode, EntryInfo* lostAndFoundInfo,
+   FsckDirInodeList* dirInodes, FsckDirInodeList* failedInodes, FsckDirEntryList* createdDentries,
+   std::set<NumNodeID>& secondariesWithRepair)
+{
+   const char* logContext = "MsgHelperRepair (linkToLostAndFound)";
+
+   if (lostAndFoundInfo->getIsBuddyMirrored() &&
+         !setSecondaryBad(lostAndFoundInfo->getOwnerNodeID().val(), secondariesWithRepair,
+            {}, *createdDentries))
+   {
+      return;
+   }
+
+   LinkToLostAndFoundMsg linkToLostAndFoundMsg(lostAndFoundInfo, dirInodes);
+
+   const auto respMsg = MessagingTk::requestResponse(lostAndFoundNode, linkToLostAndFoundMsg,
+      NETMSGTYPE_LinkToLostAndFoundResp);
+
+   if (respMsg)
+   {
+      auto* linkToLostAndFoundRespMsg = (LinkToLostAndFoundRespMsg*) respMsg.get();
+
+      linkToLostAndFoundRespMsg->getFailedDirInodes().swap(*failedInodes);
+      linkToLostAndFoundRespMsg->getCreatedDirEntries().swap(*createdDentries);
+   }
+   else
+   {
+      *failedInodes = *dirInodes;
+
+      LogContext(logContext).logErr("Communication error occured with node: " +
+         lostAndFoundNode.getAlias());
+   }
+
+   if (! failedInodes->empty())
+   {
+      for (FsckDirInodeListIter iter = failedInodes->begin(); iter != failedInodes->end();
+         iter++)
+      {
+         LogContext(logContext).log(Log_CRITICAL, "Failed to link directory inode to lost+found. "
+            "entryID: " + iter->getID());
+      }
+   }
+}
+
+void MsgHelperRepair::linkToLostAndFound(Node& lostAndFoundNode, EntryInfo* lostAndFoundInfo,
+   FsckFileInodeList* fileInodes, FsckFileInodeList* failedInodes,
+   FsckDirEntryList* createdDentries, std::set<NumNodeID>& secondariesWithRepair)
+{
+   const char* logContext = "MsgHelperRepair (linkToLostAndFound)";
+
+   if (lostAndFoundInfo->getIsBuddyMirrored() &&
+         !setSecondaryBad(lostAndFoundInfo->getOwnerNodeID().val(), secondariesWithRepair,
+            *fileInodes, *failedInodes))
+   {
+      return;
+   }
+
+   LinkToLostAndFoundMsg linkToLostAndFoundMsg(lostAndFoundInfo, fileInodes);
+
+   const auto respMsg = MessagingTk::requestResponse(lostAndFoundNode, linkToLostAndFoundMsg,
+      NETMSGTYPE_LinkToLostAndFoundResp);
+
+   if (respMsg)
+   {
+      auto* linkToLostAndFoundRespMsg = (LinkToLostAndFoundRespMsg*) respMsg.get();
+
+      linkToLostAndFoundRespMsg->getFailedFileInodes().swap(*failedInodes);
+      linkToLostAndFoundRespMsg->getCreatedDirEntries().swap(*createdDentries);
+   }
+   else
+   {
+      *failedInodes = *fileInodes;
+
+      LogContext(logContext).logErr("Communication error occured with node: " +
+         lostAndFoundNode.getAlias());
+   }
+
+   if (! failedInodes->empty())
+   {
+      for (FsckFileInodeListIter iter = failedInodes->begin(); iter != failedInodes->end();
+         iter++)
+      {
+         LogContext(logContext).log(Log_CRITICAL, "Failed to link file inode to lost+found. "
+            "entryID: " + iter->getID());
+      }
+   }
+}
+
+void MsgHelperRepair::createContDirs(NumNodeID node, bool isBuddyMirrored, FsckDirInodeList* inodes,
+   StringList* failedCreates, std::set<NumNodeID>& secondariesWithRepair)
+{
+   // create a string list with the IDs
+   std::vector<CreateEmptyContDirsMsg::Item> items;
+   for (FsckDirInodeListIter iter = inodes->begin(); iter != inodes->end(); iter++)
+      items.emplace_back(iter->getID(), iter->getIsBuddyMirrored());
+
+   CreateEmptyContDirsMsg createContDirsMsg(std::move(items));
+
+   RequestResponseNode rrNode(node, Program::getApp()->getMetaNodes());
+   RequestResponseArgs rrArgs(nullptr, &createContDirsMsg, NETMSGTYPE_CreateEmptyContDirsResp);
+
+   if (isBuddyMirrored)
+   {
+      rrNode.setMirrorInfo(Program::getApp()->getMetaMirrorBuddyGroupMapper(), false);
+      if (!setSecondaryBad(node.val(), secondariesWithRepair, {}, *failedCreates))
+      {
+         for (auto it = inodes->begin(); it != inodes->end(); ++it)
+            failedCreates->push_back(it->getID());
+         return;
+      }
+   }
+
+   FhgfsOpsErr commRes = MessagingTk::requestResponseNode(&rrNode, &rrArgs);
+
+   if (commRes == FhgfsOpsErr_SUCCESS)
+   {
+      auto* createContDirsRespMsg = (CreateEmptyContDirsRespMsg*) rrArgs.outRespMsg.get();
+
+      createContDirsRespMsg->getFailedIDs().swap(*failedCreates);
+   }
+   else
+   {
+      failedCreates->clear();
+      for (auto it = inodes->begin(); it != inodes->end(); ++it)
+         failedCreates->push_back(it->getID());
+
+      LOG(GENERAL, ERR, "Communication error occured.", node, isBuddyMirrored, commRes);
+   }
+
+   if ( !failedCreates->empty() )
+   {
+      for ( StringListIter iter = failedCreates->begin(); iter != failedCreates->end(); iter++ )
+         LOG(GENERAL, CRITICAL, "Failed to create empty content directory.", ("dirID", *iter));
+   }
+}
+
+void MsgHelperRepair::updateFileAttribs(NumNodeID node, bool isBuddyMirrored, FsckFileInodeList* inodes,
+   FsckFileInodeList* failedUpdates, std::set<NumNodeID>& secondariesWithRepair)
+{
+   UpdateFileAttribsMsg updateFileAttribsMsg(inodes);
+
+   RequestResponseNode rrNode(node, Program::getApp()->getMetaNodes());
+   RequestResponseArgs rrArgs(nullptr, &updateFileAttribsMsg, NETMSGTYPE_UpdateFileAttribsResp);
+
+   if (isBuddyMirrored)
+   {
+      rrNode.setMirrorInfo(Program::getApp()->getMetaMirrorBuddyGroupMapper(), false);
+      if (!setSecondaryBad(node.val(), secondariesWithRepair, *inodes, *failedUpdates))
+         return;
+   }
+
+   FhgfsOpsErr commRes = MessagingTk::requestResponseNode(&rrNode, &rrArgs);
+
+   if (commRes == FhgfsOpsErr_SUCCESS)
+   {
+      auto* updateFileAttribsRespMsg = (UpdateFileAttribsRespMsg*) rrArgs.outRespMsg.get();
+
+      updateFileAttribsRespMsg->getFailedInodes().swap(*failedUpdates);
+   }
+   else
+   {
+      *failedUpdates = *inodes;
+
+      LOG(GENERAL, ERR, "Communication error occured.", node, isBuddyMirrored, commRes);
+   }
+
+   if ( !failedUpdates->empty() )
+   {
+      for (auto iter = failedUpdates->begin(); iter != failedUpdates->end(); iter++)
+         LOG(GENERAL, CRITICAL, "Failed to update attributes of file inode.",
+             ("entryID", iter->getID()));
+   }
+}
+
+void MsgHelperRepair::updateDirAttribs(NumNodeID node, bool isBuddyMirrored,
+   FsckDirInodeList* inodes, FsckDirInodeList* failedUpdates,
+   std::set<NumNodeID>& secondariesWithRepair)
+{
+   UpdateDirAttribsMsg updateDirAttribsMsg(inodes);
+
+   RequestResponseNode rrNode(node, Program::getApp()->getMetaNodes());
+   RequestResponseArgs rrArgs(nullptr, &updateDirAttribsMsg, NETMSGTYPE_UpdateDirAttribsResp);
+
+   if (isBuddyMirrored)
+   {
+      rrNode.setMirrorInfo(Program::getApp()->getMetaMirrorBuddyGroupMapper(), false);
+      if (!setSecondaryBad(node.val(), secondariesWithRepair, *inodes, *failedUpdates))
+         return;
+   }
+
+   FhgfsOpsErr commRes = MessagingTk::requestResponseNode(&rrNode, &rrArgs);
+
+   if (commRes == FhgfsOpsErr_SUCCESS)
+   {
+      auto* updateDirAttribsRespMsg = (UpdateDirAttribsRespMsg*) rrArgs.outRespMsg.get();
+
+      updateDirAttribsRespMsg->getFailedInodes().swap(*failedUpdates);
+   }
+   else
+   {
+      *failedUpdates = *inodes;
+
+      LOG(GENERAL, ERR, "Communication error occured.", node, isBuddyMirrored, commRes);
+   }
+
+   if ( !failedUpdates->empty() )
+   {
+      for (auto iter = failedUpdates->begin(); iter != failedUpdates->end(); iter++)
+         LOG(GENERAL, CRITICAL, "Failed to update attributes of directory inode.",
+               ("entryID", iter->getID()));
+   }
+}
+
+void MsgHelperRepair::recreateFsIDs(NumNodeID node, bool isBuddyMirrored,
+   FsckDirEntryList* dentries, FsckDirEntryList* failedEntries,
+   std::set<NumNodeID>& secondariesWithRepair)
+{
+   const char* logContext = "MsgHelperRepair (recreateFsIDs)";
+
+   RecreateFsIDsMsg recreateFsIDsMsg(dentries);
+
+   RequestResponseNode rrNode(node, Program::getApp()->getMetaNodes());
+   RequestResponseArgs rrArgs(nullptr, &recreateFsIDsMsg, NETMSGTYPE_RecreateFsIDsResp);
+
+   if (isBuddyMirrored)
+   {
+      rrNode.setMirrorInfo(Program::getApp()->getMetaMirrorBuddyGroupMapper(), false);
+      if (!setSecondaryBad(node.val(), secondariesWithRepair, *dentries, *failedEntries))
+         return;
+   }
+
+   FhgfsOpsErr commRes = MessagingTk::requestResponseNode(&rrNode, &rrArgs);
+
+   if (commRes == FhgfsOpsErr_SUCCESS)
+   {
+      RecreateFsIDsRespMsg* recreateFsIDsRespMsg = (RecreateFsIDsRespMsg*) rrArgs.outRespMsg.get();
+
+      recreateFsIDsRespMsg->getFailedEntries().swap(*failedEntries);
+   }
+   else
+   {
+      *failedEntries = *dentries;
+
+      LOG(GENERAL, ERR, "Communication error occured.", node, isBuddyMirrored, commRes);
+   }
+
+   if ( !failedEntries->empty() )
+   {
+      for ( FsckDirEntryListIter iter = failedEntries->begin(); iter != failedEntries->end();
+         iter++ )
+      {
+         LogContext(logContext).log(Log_CRITICAL, "Failed to recreate dentry-by-ID file link."
+            " entryID: " + iter->getID());
+      }
+   }
+}
+
+void MsgHelperRepair::recreateDentries(NumNodeID node, bool isBuddyMirrored, FsckFsIDList* fsIDs,
+   FsckFsIDList* failedCreates, FsckDirEntryList* createdDentries, FsckFileInodeList* createdInodes,
+   std::set<NumNodeID>& secondariesWithRepair)
+{
+   const char* logContext = "MsgHelperRepair (recreateDentries)";
+
+   RecreateDentriesMsg recreateDentriesMsg(fsIDs);
+
+   RequestResponseNode rrNode(node, Program::getApp()->getMetaNodes());
+   RequestResponseArgs rrArgs(nullptr, &recreateDentriesMsg, NETMSGTYPE_RecreateDentriesResp);
+
+   if (isBuddyMirrored)
+   {
+      rrNode.setMirrorInfo(Program::getApp()->getMetaMirrorBuddyGroupMapper(), false);
+      if (!setSecondaryBad(node.val(), secondariesWithRepair, *fsIDs, *failedCreates))
+         return;
+   }
+
+   FhgfsOpsErr commRes = MessagingTk::requestResponseNode(&rrNode, &rrArgs);
+
+   if (commRes == FhgfsOpsErr_SUCCESS)
+   {
+      auto* recreateDentriesMsg = (RecreateDentriesRespMsg*) rrArgs.outRespMsg.get();
+
+      recreateDentriesMsg->getFailedCreates().swap(*failedCreates);
+      recreateDentriesMsg->getCreatedDentries().swap(*createdDentries);
+      recreateDentriesMsg->getCreatedInodes().swap(*createdInodes);
+   }
+   else
+   {
+      *failedCreates = *fsIDs;
+
+      LOG(GENERAL, ERR, "Communication error occured.", node, isBuddyMirrored, commRes);
+   }
+
+   if ( !failedCreates->empty() )
+   {
+      for ( FsckFsIDListIter iter = failedCreates->begin(); iter != failedCreates->end();
+         iter++ )
+      {
+         LogContext(logContext).log(Log_CRITICAL, "Failed to recreate dentry."
+            " entryID: " + iter->getID());
+      }
+   }
+}
+
+void MsgHelperRepair::fixChunkPermissions(Node& node, FsckChunkList& chunkList,
+   PathInfoList& pathInfoList, FsckChunkList& failedChunks)
+{
+   const char* logContext = "MsgHelperRepair (fixChunkPermissions)";
+
+   if ( chunkList.size() != pathInfoList.size() )
+   {
+      LogContext(logContext).logErr(
+         "Failed to set uid/gid for chunks. Size of lists does not match.");
+      return;
+   }
+
+   FsckChunkListIter chunksIter = chunkList.begin();
+   PathInfoListIter pathInfoIter = pathInfoList.begin();
+   for ( ; chunksIter != chunkList.end(); chunksIter++, pathInfoIter++ )
+   {
+      std::string chunkID = chunksIter->getID();
+      uint16_t targetID = chunksIter->getTargetID();
+      int validAttribs = SETATTR_CHANGE_USERID | SETATTR_CHANGE_GROUPID; // only interested in these
+      SettableFileAttribs attribs = {0, 0, 0, 0, 0};
+      attribs.userID = chunksIter->getUserID();
+      attribs.groupID = chunksIter->getGroupID();
+
+      bool enableCreation = false;
+
+      PathInfo pathInfo = *pathInfoIter;
+      SetLocalAttrMsg setLocalAttrMsg(chunkID, targetID, &pathInfo, validAttribs, &attribs,
+         enableCreation);
+      setLocalAttrMsg.addMsgHeaderFeatureFlag(SETLOCALATTRMSG_FLAG_USE_QUOTA);
+
+      const auto respMsg = MessagingTk::requestResponse(node, setLocalAttrMsg,
+            NETMSGTYPE_SetLocalAttrResp);
+
+      if (respMsg)
+      {
+         SetLocalAttrRespMsg* setLocalAttrRespMsg = (SetLocalAttrRespMsg*) respMsg.get();
+
+         if ( setLocalAttrRespMsg->getResult() != FhgfsOpsErr_SUCCESS )
+         {
+            LogContext(logContext).logErr(
+               "Failed to set uid/gid for chunk. chunkID: " + chunkID + "; targetID: "
+                  + StringTk::uintToStr(targetID));
+
+            failedChunks.push_back(*chunksIter);
+         }
+      }
+      else
+      {
+         LogContext(logContext).logErr("Communication error occured with node: " + node.getAlias());
+
+         failedChunks.push_back(*chunksIter);
+      }
+   }
+}
+
+/*
+ * NOTE: chunk gets modified! (new savedPath is set)
+ */
+bool MsgHelperRepair::moveChunk(Node& node, FsckChunk& chunk, const std::string& moveTo,
+   bool allowOverwrite)
+{
+   const char* logContext = "MsgHelperRepair (moveChunks)";
+   bool result;
+
+   MoveChunkFileMsg moveChunkFileMsg(chunk.getID(), chunk.getTargetID(),
+         chunk.getBuddyGroupID() != 0, chunk.getSavedPath()->str(), moveTo, allowOverwrite);
+
+   const auto respMsg = MessagingTk::requestResponse(node, moveChunkFileMsg,
+      NETMSGTYPE_MoveChunkFileResp);
+
+   if (respMsg)
+   {
+      MoveChunkFileRespMsg* moveChunkFileRespMsg = (MoveChunkFileRespMsg*) respMsg.get();
+
+      result = moveChunkFileRespMsg->getValue() == FhgfsOpsErr_SUCCESS;
+
+      if(!result)
+      {
+         LogContext(logContext).logErr(
+            "Failed to move chunk. chunkID: " + chunk.getID() + "; targetID: "
+               + StringTk::uintToStr(chunk.getTargetID() ) + "; fromPath: "
+               + chunk.getSavedPath()->str() + "; toPath: "
+               + moveTo);
+      }
+      else
+      {
+         // set newPath in chunk
+         chunk.setSavedPath(Path(moveTo));
+      }
+   }
+   else
+   {
+      LogContext(logContext).logErr("Communication error occured with node: " + node.getAlias());
+      result = false;
+   }
+
+   return result;
+}
+
+void MsgHelperRepair::deleteFileInodes(NumNodeID node, bool isBuddyMirrored,
+   FsckFileInodeList& inodes, StringList& failedDeletes, std::set<NumNodeID>& secondariesWithRepair)
+{
+   std::vector<RemoveInodesMsg::Item> items;
+
+   for (auto it = inodes.begin(); it != inodes.end(); ++it)
+      items.emplace_back(it->getID(), DirEntryType_REGULARFILE, it->getIsBuddyMirrored());
+
+   RemoveInodesMsg removeInodesMsg(std::move(items));
+
+   RequestResponseNode rrNode(node, Program::getApp()->getMetaNodes());
+   RequestResponseArgs rrArgs(nullptr, &removeInodesMsg, NETMSGTYPE_RemoveInodesResp);
+
+   if (isBuddyMirrored)
+   {
+      rrNode.setMirrorInfo(Program::getApp()->getMetaMirrorBuddyGroupMapper(), false);
+      if (!setSecondaryBad(node.val(), secondariesWithRepair, {}, failedDeletes))
+      {
+         for (auto it = inodes.begin(); it != inodes.end(); ++it)
+            failedDeletes.push_back(it->getID());
+         return;
+      }
+   }
+
+   FhgfsOpsErr commRes = MessagingTk::requestResponseNode(&rrNode, &rrArgs);
+
+   if (commRes == FhgfsOpsErr_SUCCESS)
+   {
+      RemoveInodesRespMsg* removeInodesRespMsg = (RemoveInodesRespMsg*) rrArgs.outRespMsg.get();
+      failedDeletes = removeInodesRespMsg->releaseFailedEntryIDList();
+
+      for ( StringListIter iter = failedDeletes.begin(); iter != failedDeletes.end(); iter++ )
+         LOG(GENERAL, ERR, "Failed to delete file inode.", node, isBuddyMirrored,
+             ("entryID", *iter));
+   }
+   else
+   {
+      LOG(GENERAL, ERR, "Communication error occured.", node, isBuddyMirrored, commRes);
+
+      failedDeletes.clear();
+      for (auto it = inodes.begin(); it != inodes.end(); ++it)
+         failedDeletes.push_back(it->getID());
+   }
+}
+
+void MsgHelperRepair::deleteDuplicateFileInodes(NumNodeID node, bool isBuddyMirrored,
+      FsckDuplicateInodeInfoVector& dupInodes, StringList& failedEntries, std::set<NumNodeID>& secondariesWithRepair)
+{
+   const char* logContext = "MsgHelperRepair (check and repair duplicate inode)";
+
+   CheckAndRepairDupInodeMsg repairDupInodeMsg(&dupInodes);
+
+   RequestResponseNode rrNode(node, Program::getApp()->getMetaNodes());
+   RequestResponseArgs rrArgs(nullptr, &repairDupInodeMsg, NETMSGTYPE_CheckAndRepairDupInodeResp);
+
+   if (isBuddyMirrored)
+   {
+      rrNode.setMirrorInfo(Program::getApp()->getMetaMirrorBuddyGroupMapper(), false);
+      if (!setSecondaryBad(node.val(), secondariesWithRepair, {}, failedEntries))
+      {
+         for (auto it = dupInodes.begin(); it != dupInodes.end(); ++it)
+            failedEntries.push_back(it->getID());
+         return;
+      }
+   }
+
+   FhgfsOpsErr commRes = MessagingTk::requestResponseNode(&rrNode, &rrArgs);
+
+   if (commRes == FhgfsOpsErr_SUCCESS)
+   {
+      CheckAndRepairDupInodeRespMsg* repairDupInodeRespMsg = (CheckAndRepairDupInodeRespMsg*) rrArgs.outRespMsg.get();
+      failedEntries = repairDupInodeRespMsg->releaseFailedEntryIDList();
+
+      for (StringListIter iter = failedEntries.begin(); iter != failedEntries.end(); ++iter)
+      {
+         LogContext(logContext).logErr("Failed to repair duplicate inode, entryID: " + *iter);
+      }
+   }
+   else
+   {
+      LogContext(logContext).logErr("Communication error occured with node: " + std::to_string(node.val()));
+   }
+}
+
+void MsgHelperRepair::deinlineFileInode(NumNodeID node, EntryInfo* entryInfo, StringList& failedEntries,
+   std::set<NumNodeID>& secondariesWithRepair)
+{
+   const char* logContext = "MsgHelperRepair (Deinline File Inode)";
+
+   MoveFileInodeMsg msg(entryInfo, MODE_DEINLINE);
+
+   RequestResponseNode rrNode(node, Program::getApp()->getMetaNodes());
+   RequestResponseArgs rrArgs(nullptr, &msg, NETMSGTYPE_MoveFileInodeResp);
+
+   if (entryInfo->getIsBuddyMirrored())
+   {
+      rrNode.setMirrorInfo(Program::getApp()->getMetaMirrorBuddyGroupMapper(), false);
+      if (!setSecondaryBad(node.val(), secondariesWithRepair, {}, failedEntries))
+      {
+         failedEntries.push_back(entryInfo->getEntryID());
+         return;
+      }
+   }
+
+   FhgfsOpsErr commRes = MessagingTk::requestResponseNode(&rrNode, &rrArgs);
+   if (commRes == FhgfsOpsErr_SUCCESS)
+   {
+      auto respMsg = (MoveFileInodeRespMsg*) rrArgs.outRespMsg.get();
+      if (respMsg->getResult() != FhgfsOpsErr_SUCCESS)
+      {
+         LogContext(logContext).logErr("Failed to migrate hardlink metadata to new version, entryID: " + entryInfo->getEntryID());
+         failedEntries.push_back(entryInfo->getEntryID());
+      }
+   }
+   else
+   {
+      failedEntries.push_back(entryInfo->getEntryID());
+      LogContext(logContext).logErr("Communication error occured with node: " + std::to_string(node.val()));
+   }
+}

fsck/source/net/msghelpers/MsgHelperRepair.h

@@ -0,0 +1,77 @@
+#ifndef MSGHELPERREPAIR_H_
+#define MSGHELPERREPAIR_H_
+
+#include <common/Common.h>
+#include <common/nodes/Node.h>
+#include <common/nodes/TargetStateInfo.h>
+#include <common/fsck/FsckChunk.h>
+#include <common/fsck/FsckDirEntry.h>
+#include <common/fsck/FsckDirInode.h>
+#include <common/fsck/FsckFsID.h>
+#include <common/fsck/FsckDuplicateInodeInfo.h>
+
+class MsgHelperRepair
+{
+   public:
+      static FhgfsOpsErr setNodeState(NumNodeID node, TargetConsistencyState state);
+
+      static void deleteDanglingDirEntries(NumNodeID node, bool isBuddyMirrored,
+         FsckDirEntryList* dentries, FsckDirEntryList* failedDeletes,
+         std::set<NumNodeID>& secondariesWithRepair);
+
+      static void createDefDirInodes(NumNodeID node, bool isBuddyMirrored,
+         const std::vector<std::tuple<std::string, bool>>& entries,
+         FsckDirInodeList* createdInodes,
+         std::set<NumNodeID>& secondariesWithRepair);
+
+      static void correctInodeOwnersInDentry(NumNodeID node, bool isBuddyMirrored,
+         FsckDirEntryList* dentries, NumNodeIDList* owners, FsckDirEntryList* failedCorrections,
+         std::set<NumNodeID>& secondariesWithRepair);
+      static void correctInodeOwners(NumNodeID node, bool isBuddyMirrored,
+         FsckDirInodeList* dirInodes, FsckDirInodeList* failedCorrections,
+         std::set<NumNodeID>& secondariesWithRepair);
+
+      static void deleteFiles(NumNodeID node, bool isBuddyMirrored, FsckDirEntryList* dentries,
+         FsckDirEntryList* failedDeletes);
+      static void deleteChunks(Node& node, FsckChunkList* chunks, FsckChunkList* failedDeletes);
+
+      static NodeHandle referenceLostAndFoundOwner(EntryInfo* outLostAndFoundEntryInfo);
+      static bool createLostAndFound(NodeHandle& outReferencedNode,
+         EntryInfo& outLostAndFoundEntryInfo);
+      static void linkToLostAndFound(Node& lostAndFoundNode, EntryInfo* lostAndFoundInfo,
+         FsckDirInodeList* dirInodes, FsckDirInodeList* failedInodes,
+         FsckDirEntryList* createdDentries, std::set<NumNodeID>& secondariesWithRepair);
+      static void linkToLostAndFound(Node& lostAndFoundNode, EntryInfo* lostAndFoundInfo,
+         FsckFileInodeList* fileInodes, FsckFileInodeList* failedInodes,
+         FsckDirEntryList* createdDentries, std::set<NumNodeID>& secondariesWithRepair);
+      static void createContDirs(NumNodeID node, bool isBuddyMirrored, FsckDirInodeList* inodes,
+         StringList* failedCreates, std::set<NumNodeID>& secondariesWithRepair);
+      static void updateFileAttribs(NumNodeID node, bool isBuddyMirrored, FsckFileInodeList* inodes,
+         FsckFileInodeList* failedUpdates, std::set<NumNodeID>& secondariesWithRepair);
+      static void updateDirAttribs(NumNodeID node, bool isBuddyMirrored, FsckDirInodeList* inodes,
+         FsckDirInodeList* failedUpdates, std::set<NumNodeID>& secondariesWithRepair);
+      static void recreateFsIDs(NumNodeID node, bool isBuddyMirrored,  FsckDirEntryList* dentries,
+         FsckDirEntryList* failedEntries, std::set<NumNodeID>& secondariesWithRepair);
+      static void recreateDentries(NumNodeID node, bool isBuddyMirrored, FsckFsIDList* fsIDs,
+         FsckFsIDList* failedCreates, FsckDirEntryList* createdDentries,
+         FsckFileInodeList* createdInodes, std::set<NumNodeID>& secondariesWithRepair);
+      static void fixChunkPermissions(Node& node, FsckChunkList& chunkList,
+         PathInfoList& pathInfoList, FsckChunkList& failedChunks);
+      static bool moveChunk(Node& node, FsckChunk& chunk, const std::string& moveTo,
+         bool allowOverwrite);
+      static void deleteFileInodes(NumNodeID node, bool isBuddyMirrored, FsckFileInodeList& inodes,
+         StringList& failedDeletes, std::set<NumNodeID>& secondariesWithRepair);
+      static void deleteDuplicateFileInodes(NumNodeID node, bool isBuddyMirrored,
+         FsckDuplicateInodeInfoVector& dupInodes, StringList& failedEntries, std::set<NumNodeID>& secondariesWithRepair);
+      static void deinlineFileInode(NumNodeID node, EntryInfo* entryInfo, StringList& failedEntries,
+         std::set<NumNodeID>& secondariesWithRepair);
+
+
+   private:
+      MsgHelperRepair() {}
+
+   public:
+      // inliners
+};
+
+#endif /* MSGHELPERREPAIR_H_ */

fsck/source/program/Main.cpp

@@ -0,0 +1,7 @@
+#include "Program.h"
+
+int main(int argc, char** argv)
+{
+   return Program::main(argc, argv);
+}
+

fsck/source/program/Program.cpp

@@ -0,0 +1,23 @@
+#include <common/toolkit/BuildTypeTk.h>
+#include "Program.h"
+
+App* Program::app = NULL;
+
+int Program::main(int argc, char** argv)
+{
+   BuildTypeTk::checkDebugBuildTypes();
+
+   AbstractApp::runTimeInitsAndChecks(); // must be called before creating a new App
+
+   app = new App(argc, argv);
+
+   app->startInCurrentThread();
+
+   int appRes = app->getAppResult();
+
+   delete app;
+
+   return appRes;
+}
+
+

fsck/source/program/Program.h

@@ -0,0 +1,27 @@
+#ifndef PROGRAM_H_
+#define PROGRAM_H_
+
+#include <app/App.h>
+
+class Program
+{
+   public:
+      static int main(int argc, char** argv);
+
+
+   private:
+      Program() {}
+
+      static App* app;
+
+
+   public:
+      // getters & setters
+      static App* getApp()
+      {
+         return app;
+      }
+
+};
+
+#endif /*PROGRAM_H_*/

fsck/source/toolkit/DatabaseTk.cpp

@@ -0,0 +1,273 @@
+#include "DatabaseTk.h"
+
+#include <common/app/log/LogContext.h>
+#include <common/storage/striping/Raid0Pattern.h>
+#include <common/toolkit/FsckTk.h>
+#include <common/toolkit/StringTk.h>
+#include <database/FsckDB.h>
+#include <database/FsckDBTable.h>
+#include <common/storage/Path.h>
+#include <common/toolkit/StorageTk.h>
+
+DatabaseTk::DatabaseTk()
+{
+}
+
+DatabaseTk::~DatabaseTk()
+{
+}
+
+FsckDirEntry DatabaseTk::createDummyFsckDirEntry(FsckDirEntryType entryType)
+{
+   FsckDirEntryList list;
+   createDummyFsckDirEntries(7, 1, &list, entryType);
+   return list.front();
+}
+
+void DatabaseTk::createDummyFsckDirEntries(uint amount, FsckDirEntryList* outList,
+   FsckDirEntryType entryType)
+{
+   createDummyFsckDirEntries(0,amount, outList, entryType);
+}
+
+void DatabaseTk::createDummyFsckDirEntries(uint from, uint amount, FsckDirEntryList* outList,
+   FsckDirEntryType entryType)
+{
+   for ( uint i = from; i < (from+amount); i++ )
+   {
+      std::string index = StringTk::uintToHexStr(i);
+
+      std::string id = index + "-1-1";
+      std::string name = "dentryName" + index;
+      std::string parentID = index + "-2-1";
+      NumNodeID entryOwnerNodeID(i + 1000);
+      NumNodeID inodeOwnerNodeID(i + 2000);
+      NumNodeID saveNodeID(i + 3000);
+      int saveDevice = i;
+      uint64_t saveInode = i;
+
+      FsckDirEntry dentry(id, name, parentID, entryOwnerNodeID, inodeOwnerNodeID, entryType, true,
+         saveNodeID, saveDevice, saveInode, false);
+
+      outList->push_back(dentry);
+   }
+}
+
+FsckFileInode DatabaseTk::createDummyFsckFileInode()
+{
+   FsckFileInodeList list;
+   createDummyFsckFileInodes(7, 1, &list);
+   return list.front();
+}
+
+void DatabaseTk::createDummyFsckFileInodes(uint amount, FsckFileInodeList* outList)
+{
+   createDummyFsckFileInodes(0,amount,outList);
+}
+
+void DatabaseTk::createDummyFsckFileInodes(uint from, uint amount, FsckFileInodeList* outList)
+{
+   for ( uint i = from; i < (from + amount); i++ )
+   {
+      std::string index = StringTk::uintToHexStr(i);
+      UInt16Vector targetIDs;
+      while (targetIDs.size() < i)
+         targetIDs.push_back(i + 1000 * (targetIDs.size() + 1) );
+      Raid0Pattern stripePatternIn(1024, targetIDs);
+
+      std::string id = index + "-1-1";
+      std::string parentDirID = index + "-2-1";
+      NumNodeID parentNodeID(i + 1000);
+
+      unsigned userID = 1000;
+      unsigned groupID = 1000;
+
+      uint64_t usedBlocks = 0;
+      int64_t fileSize = usedBlocks*512;
+      unsigned numHardLinks = 1;
+
+      PathInfo pathInfo(userID, parentDirID, PATHINFO_FEATURE_ORIG);
+
+      FsckStripePatternType stripePatternType = FsckTk::stripePatternToFsckStripePattern(
+         &stripePatternIn);
+      unsigned chunkSize = 524288;
+
+      NumNodeID saveNodeID(i + 2000);
+
+      FsckFileInode fileInode(id, parentDirID, parentNodeID, pathInfo, userID, groupID, fileSize,
+            numHardLinks, usedBlocks, targetIDs, stripePatternType, chunkSize, saveNodeID,
+            1000 + i, 42, true, false, true, false);
+
+      outList->push_back(fileInode);
+   }
+}
+
+FsckDirInode DatabaseTk::createDummyFsckDirInode()
+{
+   FsckDirInodeList list;
+   createDummyFsckDirInodes(7, 1, &list);
+   return list.front();
+}
+
+void DatabaseTk::createDummyFsckDirInodes(uint amount, FsckDirInodeList* outList)
+{
+   createDummyFsckDirInodes(0, amount, outList);
+}
+
+void DatabaseTk::createDummyFsckDirInodes(uint from, uint amount, FsckDirInodeList* outList)
+{
+   for ( uint i = from; i < (from + amount); i++ )
+   {
+      std::string index = StringTk::uintToHexStr(i);
+      UInt16Vector targetIDs;
+      Raid0Pattern stripePatternIn(1024, targetIDs);
+
+      std::string id = index + "-1-1";
+      std::string parentDirID = index + "-2-1";
+      NumNodeID parentNodeID(i + 1000);
+      NumNodeID ownerNodeID(i + 2000);
+
+      int64_t size = 10;
+      unsigned numHardLinks = 12;
+
+      FsckStripePatternType stripePatternType = FsckTk::stripePatternToFsckStripePattern(
+         &stripePatternIn);
+      NumNodeID saveNodeID(i + 3000);
+
+      FsckDirInode dirInode(id, parentDirID, parentNodeID, ownerNodeID, size, numHardLinks,
+         targetIDs, stripePatternType, saveNodeID, false, true, false);
+      outList->push_back(dirInode);
+   }
+}
+
+FsckChunk DatabaseTk::createDummyFsckChunk()
+{
+   FsckChunkList list;
+   createDummyFsckChunks(7, 1, &list);
+   return list.front();
+}
+
+void DatabaseTk::createDummyFsckChunks(uint amount, FsckChunkList* outList)
+{
+   createDummyFsckChunks(amount, 1, outList);
+}
+
+void DatabaseTk::createDummyFsckChunks(uint amount, uint numTargets, FsckChunkList* outList)
+{
+   createDummyFsckChunks(0, amount, numTargets, outList);
+}
+
+void DatabaseTk::createDummyFsckChunks(uint from, uint amount, uint numTargets,
+   FsckChunkList* outList)
+{
+   for ( uint i = from; i < (from + amount); i++ )
+   {
+      std::string index = StringTk::uintToHexStr(i);
+
+      for ( uint j = 0; j < numTargets; j++ )
+      {
+         std::string id = index + "-1-1";
+         uint16_t targetID = j;
+
+         PathInfo pathInfo(0, META_ROOTDIR_ID_STR,  PATHINFO_FEATURE_ORIG);
+         Path chunkDirPath; // ignored!
+         std::string chunkFilePathStr;
+         StorageTk::getChunkDirChunkFilePath(&pathInfo, id, true, chunkDirPath,
+            chunkFilePathStr);
+         Path savedPath(chunkFilePathStr);
+
+         uint64_t usedBlocks = 300;
+         int64_t fileSize = usedBlocks*512;
+
+         FsckChunk chunk(id, targetID, savedPath, fileSize, usedBlocks, 0, 0, 0, 0, 0);
+         outList->push_back(chunk);
+      }
+   }
+}
+
+FsckContDir DatabaseTk::createDummyFsckContDir()
+{
+   FsckContDirList list;
+   createDummyFsckContDirs(7, 1, &list);
+   return list.front();
+}
+
+void DatabaseTk::createDummyFsckContDirs(uint amount, FsckContDirList* outList)
+{
+   createDummyFsckContDirs(0, amount, outList);
+}
+
+void DatabaseTk::createDummyFsckContDirs(uint from, uint amount, FsckContDirList* outList)
+{
+   for ( uint i = from; i < (from + amount); i++ )
+   {
+      std::string index = StringTk::uintToHexStr(i);
+
+      std::string id = index + "-1-1";
+      NumNodeID saveNodeID(i);
+
+      FsckContDir contDir(id, saveNodeID, false);
+      outList->push_back(contDir);
+   }
+}
+
+FsckFsID DatabaseTk::createDummyFsckFsID()
+{
+   FsckFsIDList list;
+   createDummyFsckFsIDs(7, 1, &list);
+   return list.front();
+}
+
+void DatabaseTk::createDummyFsckFsIDs(uint amount, FsckFsIDList* outList)
+{
+   createDummyFsckFsIDs(0, amount, outList);
+}
+
+void DatabaseTk::createDummyFsckFsIDs(uint from, uint amount, FsckFsIDList* outList)
+{
+   for ( uint i = from; i < (from + amount); i++ )
+   {
+      std::string index = StringTk::uintToHexStr(i);
+
+      std::string id = index + "-1-1";
+      std::string parentDirID = index + "-2-1";
+      NumNodeID saveNodeID(i);
+      int saveDevice = i;
+      uint64_t saveInode = i;
+
+      FsckFsID fsID(id, parentDirID, saveNodeID, saveDevice, saveInode, false);
+      outList->push_back(fsID);
+   }
+}
+
+std::string DatabaseTk::calculateExpectedChunkPath(std::string entryID, unsigned origParentUID,
+   std::string origParentEntryID, int pathInfoFlags)
+{
+   std::string resStr;
+
+   bool hasOrigFeature;
+   hasOrigFeature = pathInfoFlags == PATHINFO_FEATURE_ORIG;
+
+   PathInfo pathInfo(origParentUID, origParentEntryID, pathInfoFlags);
+   Path chunkDirPath;
+   std::string chunkFilePath;
+   StorageTk::getChunkDirChunkFilePath(&pathInfo, entryID, hasOrigFeature, chunkDirPath,
+      chunkFilePath);
+
+   if (hasOrigFeature) // we can use the Path chunkDirPath
+      resStr = chunkDirPath.str();
+   else
+   {
+      /* chunk in old format => Path chunkDirPath is not set in getChunkDirChunkFilePath(),
+       * so it is a bit more tricky */
+
+      // we need to strip the filename itself (including the '/')
+      size_t startPos = 0;
+      size_t len = chunkFilePath.find_last_of('/');
+
+      // generate a substring as result
+      resStr = chunkFilePath.substr(startPos, len);
+   }
+
+   return resStr;
+}

fsck/source/toolkit/DatabaseTk.h

@@ -0,0 +1,62 @@
+#ifndef DATABASETK_H_
+#define DATABASETK_H_
+
+#include <common/fsck/FsckChunk.h>
+#include <common/fsck/FsckContDir.h>
+#include <common/fsck/FsckDirEntry.h>
+#include <common/fsck/FsckDirInode.h>
+#include <common/fsck/FsckFileInode.h>
+#include <common/fsck/FsckFsID.h>
+#include <common/fsck/FsckModificationEvent.h>
+#include <common/fsck/FsckTargetID.h>
+
+#include <database/Chunk.h>
+#include <database/ContDir.h>
+#include <database/DirEntry.h>
+#include <database/DirInode.h>
+#include <database/FileInode.h>
+#include <database/FsID.h>
+
+class FsckDB;
+
+class DatabaseTk
+{
+   public:
+      DatabaseTk();
+      virtual ~DatabaseTk();
+
+      static FsckDirEntry createDummyFsckDirEntry(
+         FsckDirEntryType entryType = FsckDirEntryType_REGULARFILE);
+      static void createDummyFsckDirEntries(uint amount, FsckDirEntryList* outList,
+         FsckDirEntryType entryType = FsckDirEntryType_REGULARFILE);
+      static void createDummyFsckDirEntries(uint from, uint amount, FsckDirEntryList* outList,
+         FsckDirEntryType entryType = FsckDirEntryType_REGULARFILE);
+
+
+      static FsckFileInode createDummyFsckFileInode();
+      static void createDummyFsckFileInodes(uint amount, FsckFileInodeList* outList);
+      static void createDummyFsckFileInodes(uint from, uint amount, FsckFileInodeList* outList);
+
+      static FsckDirInode createDummyFsckDirInode();
+      static void createDummyFsckDirInodes(uint amount, FsckDirInodeList* outList);
+      static void createDummyFsckDirInodes(uint from, uint amount, FsckDirInodeList* outList);
+
+      static FsckChunk createDummyFsckChunk();
+      static void createDummyFsckChunks(uint amount, FsckChunkList* outList);
+      static void createDummyFsckChunks(uint amount, uint numTargets, FsckChunkList* outList);
+      static void createDummyFsckChunks(uint from, uint amount, uint numTargets,
+         FsckChunkList* outList);
+
+      static FsckContDir createDummyFsckContDir();
+      static void createDummyFsckContDirs(uint amount, FsckContDirList* outList);
+      static void createDummyFsckContDirs(uint from, uint amount, FsckContDirList* outList);
+
+      static FsckFsID createDummyFsckFsID();
+      static void createDummyFsckFsIDs(uint amount, FsckFsIDList* outList);
+      static void createDummyFsckFsIDs(uint from, uint amount, FsckFsIDList* outList);
+
+      static std::string calculateExpectedChunkPath(std::string entryID, unsigned origParentUID,
+         std::string origParentEntryID, int pathInfoFlags);
+};
+
+#endif /* DATABASETK_H_ */

fsck/source/toolkit/FsckDefinitions.cpp

@@ -0,0 +1,2 @@
+#include "FsckDefinitions.h"
+

fsck/source/toolkit/FsckDefinitions.h

@@ -0,0 +1,75 @@
+#ifndef FSCKDBTYPES_H_
+#define FSCKDBTYPES_H_
+
+#include <common/Common.h>
+#include <common/storage/striping/StripePattern.h>
+#include <common/storage/StorageDefinitions.h>
+
+#define TABLE_NAME_DIR_ENTRIES "dirEntries"
+#define TABLE_NAME_FILE_INODES "fileInodes"
+#define TABLE_NAME_DIR_INODES "dirInodes"
+#define TABLE_NAME_CHUNKS "chunks"
+#define TABLE_NAME_STRIPE_PATTERNS "stripePatterns"
+#define TABLE_NAME_CONT_DIRS "contDirs"
+#define TABLE_NAME_FSIDS "fsIDs"
+#define TABLE_NAME_USED_TARGETS "usedTargets"
+
+#define TABLE_NAME_MODIFICATION_EVENTS "modificationEvents"
+
+enum FsckRepairAction
+{
+   FsckRepairAction_DELETEDENTRY = 0,
+   FsckRepairAction_DELETEFILE = 1,
+   FsckRepairAction_CREATEDEFAULTDIRINODE = 2,
+   FsckRepairAction_CORRECTOWNER = 3,
+   FsckRepairAction_LOSTANDFOUND = 4,
+   FsckRepairAction_CREATECONTDIR = 5,
+   FsckRepairAction_DELETEINODE = 7,
+   FsckRepairAction_DELETECHUNK = 8,
+   FsckRepairAction_DELETECONTDIR = 9,
+   FsckRepairAction_UPDATEATTRIBS = 10,
+   FsckRepairAction_CHANGETARGET = 11,
+   FsckRepairAction_RECREATEFSID = 12,
+   FsckRepairAction_RECREATEDENTRY = 13,
+   FsckRepairAction_FIXPERMISSIONS = 14,
+   FsckRepairAction_MOVECHUNK = 15,
+   FsckRepairAction_REPAIRDUPLICATEINODE = 16,
+   FsckRepairAction_UPDATEOLDTYLEDHARDLINKS = 17,
+   FsckRepairAction_NOTHING = 18,
+   FsckRepairAction_UNDEFINED = 19
+};
+
+struct FsckRepairActionElem
+{
+   const char* actionShortDesc;
+   const char* actionDesc;
+   enum FsckRepairAction action;
+};
+
+// Note: Keep in sync with enum FsckErrorCode
+FsckRepairActionElem const __FsckRepairActions[] =
+{
+      {"DeleteDentry", "Delete directory entry", FsckRepairAction_DELETEDENTRY},
+      {"DeleteFile", "Delete file", FsckRepairAction_DELETEFILE},
+      {"CreateDefDirInode", "Create a directory inode with default values",
+         FsckRepairAction_CREATEDEFAULTDIRINODE},
+      {"CorrectOwner", "Correct owner node", FsckRepairAction_CORRECTOWNER},
+      {"LostAndFound", "Link to lost+found", FsckRepairAction_LOSTANDFOUND},
+      {"CreateContDir", "Create an empty content directory", FsckRepairAction_CREATECONTDIR},
+      {"DeleteInode", "Delete inodes", FsckRepairAction_DELETEINODE},
+      {"DeleteChunk", "Delete chunk", FsckRepairAction_DELETECHUNK},
+      {"DeleteContDir", "Delete content directory", FsckRepairAction_DELETECONTDIR},
+      {"UpdateAttribs", "Update attributes", FsckRepairAction_UPDATEATTRIBS},
+      {"ChangeTarget", "Change target ID in stripe patterns", FsckRepairAction_CHANGETARGET},
+      {"RecreateFsID", "Recreate dentry-by-id file", FsckRepairAction_RECREATEFSID},
+      {"RecreateDentry", "Recreate directory entry file", FsckRepairAction_RECREATEDENTRY},
+      {"FixPermissions", "Fix permissions", FsckRepairAction_FIXPERMISSIONS},
+      {"MoveChunk", "Move chunk", FsckRepairAction_MOVECHUNK},
+      {"RepairDuplicateInode", "Repair duplicate inode", FsckRepairAction_REPAIRDUPLICATEINODE},
+      {"UpdateOldStyledHardlinks", "Update metadata of old styled hardlinks to new format",
+         FsckRepairAction_UPDATEOLDTYLEDHARDLINKS},
+      {"Nothing", "Do nothing", FsckRepairAction_NOTHING},
+      {0, 0, FsckRepairAction_UNDEFINED}
+};
+
+#endif /* FSCKDBTYPES_H_ */

fsck/source/toolkit/FsckException.h

@@ -0,0 +1,16 @@
+/*
+ * FsckException.h
+ *
+ *      This exception is intended for things that can happen in fsck, that prevents it from running
+ *      further,  e.g. if a server goes down fsck should abort
+ *
+ */
+
+#ifndef FSCKEXCEPTION_H
+#define FSCKEXCEPTION_H
+
+#include <common/toolkit/NamedException.h>
+
+DECLARE_NAMEDEXCEPTION(FsckException, "FsckException")
+
+#endif /* FSCKEXCEPTION_H */

fsck/source/toolkit/FsckTkEx.cpp

@@ -0,0 +1,604 @@
+#include "FsckTkEx.h"
+
+#include <common/net/message/fsck/FsckSetEventLoggingMsg.h>
+#include <common/net/message/fsck/FsckSetEventLoggingRespMsg.h>
+#include <common/net/message/storage/StatStoragePathMsg.h>
+#include <common/net/message/storage/StatStoragePathRespMsg.h>
+#include <common/toolkit/ListTk.h>
+#include <common/toolkit/FsckTk.h>
+#include <common/toolkit/UnitTk.h>
+#include <program/Program.h>
+
+#include <mutex>
+
+char FsckTkEx::progressChar = '-';
+Mutex FsckTkEx::outputMutex;
+
+/*
+ * check the reachability of all nodes in the system
+ */
+bool FsckTkEx::checkReachability()
+{
+   NodeStore* metaNodes = Program::getApp()->getMetaNodes();
+   NodeStore* storageNodes = Program::getApp()->getStorageNodes();
+
+   StringList errors;
+   bool commSuccess = true;
+
+   FsckTkEx::fsckOutput("Step 1: Check reachability of nodes: ", OutputOptions_FLUSH);
+
+   if ( metaNodes->getSize() == 0 )
+   {
+      errors.push_back("No metadata nodes found");
+      commSuccess = false;
+   }
+
+   if ( storageNodes->getSize() == 0 )
+   {
+      errors.push_back("No storage nodes found");
+      commSuccess = false;
+   }
+
+   for (const auto& node : metaNodes->referenceAllNodes())
+   {
+      if ( !FsckTkEx::checkReachability(*node, NODETYPE_Meta) )
+      {
+         errors.push_back("Communication with metadata node failed: " + node->getAlias());
+         commSuccess = false;
+      }
+   }
+
+   for (const auto& node : storageNodes->referenceAllNodes())
+   {
+      if ( !FsckTkEx::checkReachability(*node, NODETYPE_Storage) )
+      {
+         errors.push_back("Communication with storage node failed: " + node->getAlias());
+         commSuccess = false;
+      }
+   }
+
+   if ( commSuccess )
+      FsckTkEx::fsckOutput("Finished", OutputOptions_LINEBREAK);
+   else
+   {
+      for ( StringListIter iter = errors.begin(); iter != errors.end(); iter++ )
+      {
+         FsckTkEx::fsckOutput(*iter,
+            OutputOptions_NONE | OutputOptions_LINEBREAK | OutputOptions_STDERR);
+      }
+   }
+
+   return commSuccess;
+}
+
+/*
+ * check reachability of a single node
+ */
+bool FsckTkEx::checkReachability(Node& node, NodeType nodetype)
+{
+   bool retVal = false;
+   HeartbeatRequestMsg heartbeatRequestMsg;
+   std::string realNodeID = node.getAlias();
+
+   const auto respMsg = MessagingTk::requestResponse(node, heartbeatRequestMsg,
+         NETMSGTYPE_Heartbeat);
+   if (respMsg)
+   {
+      HeartbeatMsg *heartbeatMsg = (HeartbeatMsg *) respMsg.get();
+      std::string receivedNodeID = heartbeatMsg->getNodeID();
+      retVal = receivedNodeID.compare(realNodeID) == 0;
+   }
+
+   return retVal;
+}
+
+void FsckTkEx::fsckOutput(std::string text, int optionFlags)
+{
+   const std::lock_guard<Mutex> lock(outputMutex);
+
+   static bool fileErrLogged = false; // to make sure we print logfile open err only once
+
+   Config* cfg = Program::getApp()->getConfig(); // might be NULL on app init failure
+   bool toLog = cfg && (!(OutputOptions_NOLOG & optionFlags)); // true if write to log file
+
+   FILE *logFile = NULL;
+
+   if (likely(toLog))
+   {
+      std::string logFilePath = cfg->getLogOutFile();
+
+      logFile = fopen(logFilePath.c_str(),"a+");
+      if (logFile == NULL)
+      {
+         toLog = false;
+
+         if(!fileErrLogged)
+         {
+            std::cerr << "Cannot open output file for writing: '" << logFilePath << "'"
+               << std::endl;
+
+            fileErrLogged = true;
+         }
+      }
+   }
+
+   const char* colorNormal = OutputColor_NORMAL;
+   const char* color = OutputColor_NORMAL;
+
+   FILE *outFile = stdout;
+
+   if (OutputOptions_STDERR & optionFlags)
+   {
+      outFile = stderr;
+   }
+   else if (OutputOptions_NOSTDOUT & optionFlags)
+   {
+      outFile = NULL;
+   }
+
+   bool outFileIsTty;
+
+   if (outFile)
+      outFileIsTty = isatty(fileno(outFile));
+   else
+      outFileIsTty = false;
+
+   if (OutputOptions_COLORGREEN & optionFlags)
+   {
+      color = OutputColor_GREEN;
+   }
+   else if (OutputOptions_COLORRED & optionFlags)
+   {
+      color = OutputColor_RED;
+   }
+   else
+   {
+      color = OutputColor_NORMAL;
+   }
+
+   if (OutputOptions_LINEDELETE & optionFlags)
+   {
+      optionFlags = optionFlags | OutputOptions_FLUSH;
+      SAFE_FPRINTF(outFile, "\r");
+      SAFE_FPRINTF(outFile, "                                                                   ");
+      SAFE_FPRINTF(outFile, "\r");
+   }
+
+   if (OutputOptions_ADDLINEBREAKBEFORE & optionFlags)
+   {
+      SAFE_FPRINTF(outFile, "\n");
+      if (likely(toLog)) SAFE_FPRINTF(logFile,"\n");
+   }
+
+   if (OutputOptions_HEADLINE & optionFlags)
+   {
+      SAFE_FPRINTF(outFile, "\n--------------------------------------------------------------------\n");
+
+      if (likely(toLog))
+         SAFE_FPRINTF(logFile,"\n--------------------------------------------------------------------\n");
+
+      if (likely(outFileIsTty))
+         SAFE_FPRINTF(outFile, "%s%s%s",color,text.c_str(),colorNormal);
+      else
+         SAFE_FPRINTF(outFile, "%s",text.c_str());
+
+      if (likely(toLog))
+         SAFE_FPRINTF(logFile,"%s",text.c_str());
+
+      SAFE_FPRINTF(outFile, "\n--------------------------------------------------------------------\n") ;
+
+      if (likely(toLog))
+         SAFE_FPRINTF(logFile,"\n--------------------------------------------------------------------\n");
+   }
+   else
+   {
+      if (likely(outFileIsTty))
+         SAFE_FPRINTF(outFile, "%s%s%s",color,text.c_str(),colorNormal);
+      else
+         SAFE_FPRINTF(outFile, "%s",text.c_str());
+
+      if (likely(toLog)) SAFE_FPRINTF(logFile,"%s",text.c_str());
+   }
+
+   if (OutputOptions_LINEBREAK & optionFlags)
+   {
+      SAFE_FPRINTF(outFile, "\n");
+      if (likely(toLog))
+         SAFE_FPRINTF(logFile,"\n");
+   }
+
+   if (OutputOptions_DOUBLELINEBREAK & optionFlags)
+   {
+      SAFE_FPRINTF(outFile, "\n\n");
+      if (likely(toLog))
+         SAFE_FPRINTF(logFile,"\n\n");
+   }
+
+   if (OutputOptions_FLUSH & optionFlags)
+   {
+      fflush(outFile);
+      if (likely(toLog))
+         fflush(logFile);
+   }
+
+   if (logFile != NULL)
+   {
+      fclose(logFile);
+   }
+}
+
+void FsckTkEx::printVersionHeader(bool toStdErr, bool noLogFile)
+{
+   int optionFlags = OutputOptions_LINEBREAK;
+   if (toStdErr)
+   {
+      optionFlags = OutputOptions_LINEBREAK | OutputOptions_STDERR;
+   }
+   if (noLogFile)
+   {
+      optionFlags = optionFlags | OutputOptions_NOLOG;
+   }
+
+   FsckTkEx::fsckOutput("\n", optionFlags);
+   FsckTkEx::fsckOutput("BeeGFS File System Check Version : " + std::string(BEEGFS_VERSION),
+      optionFlags);
+   FsckTkEx::fsckOutput("----", optionFlags);
+}
+
+void FsckTkEx::progressMeter()
+{
+   const std::lock_guard<Mutex> lock(outputMutex);
+
+   printf("\b%c",progressChar);
+   fflush(stdout);
+
+   switch(progressChar)
+   {
+      case '-' :
+      {
+         progressChar = '\\';
+         break;
+      }
+      case '\\' :
+      {
+         progressChar = '|';
+         break;
+      }
+      case '|' :
+      {
+         progressChar = '/';
+         break;
+      }
+      case '/' :
+      {
+         progressChar = '-';
+         break;
+      }
+      default:
+      {
+         progressChar = '-';
+         break;
+      }
+
+   }
+}
+
+/*
+ * this is only a rough approximation
+ */
+int64_t FsckTkEx::calcNeededSpace()
+{
+   const char* logContext = "FsckTkEx (calcNeededSpace)";
+   int64_t neededSpace = 0;
+
+   // get used inodes from all meta data servers and sum them up
+   NodeStore* metaNodes = Program::getApp()->getMetaNodes();
+
+   for (const auto& metaNode : metaNodes->referenceAllNodes())
+   {
+      NumNodeID nodeID = metaNode->getNumID();
+
+      StatStoragePathMsg statStoragePathMsg(0);
+
+      const auto respMsg = MessagingTk::requestResponse(*metaNode, statStoragePathMsg,
+         NETMSGTYPE_StatStoragePathResp);
+      if (respMsg)
+      {
+         StatStoragePathRespMsg* statStoragePathRespMsg = (StatStoragePathRespMsg *) respMsg.get();
+         int64_t usedInodes = statStoragePathRespMsg->getInodesTotal()
+            - statStoragePathRespMsg->getInodesFree();
+         neededSpace += usedInodes * NEEDED_DISKSPACE_META_INODE;
+      }
+      else
+      {
+         LogContext(logContext).logErr(
+            "Unable to calculate needed disk space; Communication error with node: "
+            + nodeID.str());
+         return 0;
+      }
+   }
+
+   // get used inodes from all storage servers and sum them up
+   NodeStore* storageNodes = Program::getApp()->getStorageNodes();
+   TargetMapper* targetMapper = Program::getApp()->getTargetMapper();
+
+   for (const auto& storageNode : storageNodes->referenceAllNodes())
+   {
+      NumNodeID nodeID = storageNode->getNumID();
+      UInt16List targetIDs;
+      targetMapper->getTargetsByNode(nodeID, targetIDs);
+
+      for ( UInt16ListIter targetIDIter = targetIDs.begin(); targetIDIter != targetIDs.end();
+         targetIDIter++ )
+      {
+         uint16_t targetID = *targetIDIter;
+         StatStoragePathMsg statStoragePathMsg(targetID);
+
+         const auto respMsg = MessagingTk::requestResponse(*storageNode, statStoragePathMsg,
+            NETMSGTYPE_StatStoragePathResp);
+         if (respMsg)
+         {
+            auto* statStoragePathRespMsg = (StatStoragePathRespMsg *) respMsg.get();
+            int64_t usedInodes = statStoragePathRespMsg->getInodesTotal()
+               - statStoragePathRespMsg->getInodesFree();
+            neededSpace += usedInodes * NEEDED_DISKSPACE_STORAGE_INODE;
+         }
+         else
+         {
+            LogContext(logContext).logErr(
+               "Unable to calculate needed disk space; Communication error with node: "
+               + nodeID.str());
+            return -1;
+         }
+      }
+   }
+
+   // now we take the calculated approximation and double it to have a lot of space for errors
+   return neededSpace*2;
+}
+
+bool FsckTkEx::checkDiskSpace(Path& dbPath)
+{
+   int64_t neededDiskSpace = FsckTkEx::calcNeededSpace();
+
+   if ( unlikely(neededDiskSpace < 0) )
+   {
+      FsckTkEx::fsckOutput("Could not determine needed disk space. Aborting now.",
+         OutputOptions_LINEBREAK | OutputOptions_ADDLINEBREAKBEFORE);
+      return false;
+   }
+
+   int64_t sizeTotal;
+   int64_t sizeFree;
+   int64_t inodesTotal;
+   int64_t inodesFree;
+   bool statRes = StorageTk::statStoragePath(dbPath, true, &sizeTotal, &sizeFree,
+      &inodesTotal, &inodesFree);
+
+   if (!statRes)
+   {
+      FsckTkEx::fsckOutput(
+         "Could not stat database file path to determine free space; database file: "
+            + dbPath.str());
+      return false;
+   }
+
+   if ( neededDiskSpace >= sizeFree )
+   {
+      std::string neededDiskSpaceUnit;
+      double neededDiskSpaceValue = UnitTk::byteToXbyte(neededDiskSpace, &neededDiskSpaceUnit);
+
+      std::string sizeFreeUnit;
+      double sizeFreeValue = UnitTk::byteToXbyte(sizeFree, &sizeFreeUnit);
+
+      FsckTkEx::fsckOutput(
+         "Not enough disk space to create database file: " + dbPath.str()
+            + "; Recommended free space: " + StringTk::doubleToStr(neededDiskSpaceValue)
+            + neededDiskSpaceUnit + "; Free space: " + StringTk::doubleToStr(sizeFreeValue)
+            + sizeFreeUnit, OutputOptions_LINEBREAK | OutputOptions_ADDLINEBREAKBEFORE);
+
+      bool ignoreDBDiskSpace = Program::getApp()->getConfig()->getIgnoreDBDiskSpace();
+      if(!ignoreDBDiskSpace)
+         return false;
+   }
+
+   return true;
+}
+
+std::string FsckTkEx::getRepairActionDesc(FsckRepairAction repairAction, bool shortDesc)
+{
+   for (size_t i = 0; __FsckRepairActions[i].actionDesc != nullptr; i++)
+   {
+      if( repairAction == __FsckRepairActions[i].action )
+      { // we have a match
+         if (shortDesc)
+            return __FsckRepairActions[i].actionShortDesc;
+         else
+            return __FsckRepairActions[i].actionDesc;
+      }
+   }
+
+   return "";
+}
+
+FhgfsOpsErr FsckTkEx::startModificationLogging(NodeStore* metaNodes, Node& localNode,
+      bool forceRestart)
+{
+   const char* logContext = "FsckTkEx (startModificationLogging)";
+
+   FhgfsOpsErr retVal = FhgfsOpsErr_SUCCESS;
+
+   NicAddressList localNicList = localNode.getNicList();
+   unsigned localPortUDP = localNode.getPortUDP();
+
+   FsckTkEx::fsckOutput("-----",
+      OutputOptions_ADDLINEBREAKBEFORE | OutputOptions_FLUSH | OutputOptions_LINEBREAK);
+   FsckTkEx::fsckOutput(
+      "Waiting for metadata servers to start modification logging. This might take some time.",
+      OutputOptions_FLUSH | OutputOptions_LINEBREAK);
+   FsckTkEx::fsckOutput("-----", OutputOptions_FLUSH | OutputOptions_DOUBLELINEBREAK);
+
+   NumNodeIDList nodeIDs;
+
+   auto metaNodeList = metaNodes->referenceAllNodes();
+
+   for (auto iter = metaNodeList.begin(); iter != metaNodeList.end(); iter++)
+   {
+      auto node = metaNodes->referenceNode((*iter)->getNumID());
+
+      NicAddressList nicList;
+      FsckSetEventLoggingMsg fsckSetEventLoggingMsg(true, localPortUDP,
+         &localNicList, forceRestart);
+
+      const auto respMsg = MessagingTk::requestResponse(*node, fsckSetEventLoggingMsg,
+         NETMSGTYPE_FsckSetEventLoggingResp);
+
+      if (respMsg)
+      {
+         auto* fsckSetEventLoggingRespMsg = (FsckSetEventLoggingRespMsg*) respMsg.get();
+
+         bool started = fsckSetEventLoggingRespMsg->getLoggingEnabled();
+
+         if (!started) // EventFlusher was already running on this node!
+         {
+            LogContext(logContext).logErr("Modification logging already running on node: "
+                  + node->getAlias());
+
+            retVal = FhgfsOpsErr_INUSE;
+            break;
+         }
+      }
+      else
+      {
+         LogContext(logContext).logErr("Communication error occured with node: " + node->getAlias());
+         retVal = FhgfsOpsErr_COMMUNICATION;
+      }
+   }
+
+   return retVal;
+}
+
+bool FsckTkEx::stopModificationLogging(NodeStore* metaNodes)
+{
+   const char* logContext = "FsckTkEx (stopModificationLogging)";
+
+   bool retVal = true;
+
+   FsckTkEx::fsckOutput("-----",
+      OutputOptions_ADDLINEBREAKBEFORE | OutputOptions_FLUSH | OutputOptions_LINEBREAK);
+   FsckTkEx::fsckOutput(
+      "Waiting for metadata servers to stop modification logging. This might take some time.",
+      OutputOptions_FLUSH | OutputOptions_LINEBREAK);
+   FsckTkEx::fsckOutput("-----", OutputOptions_FLUSH | OutputOptions_DOUBLELINEBREAK);
+
+   NumNodeIDList nodeIDs;
+
+   auto metaNodeList = metaNodes->referenceAllNodes();
+
+   for (auto iter = metaNodeList.begin(); iter != metaNodeList.end(); iter++)
+      nodeIDs.push_back((*iter)->getNumID());
+
+   NumNodeIDListIter nodeIdIter = nodeIDs.begin();
+
+   while (! nodeIDs.empty())
+   {
+      NumNodeID nodeID = *nodeIdIter;
+
+      auto node = metaNodes->referenceNode(nodeID);
+
+      NicAddressList nicList;
+      FsckSetEventLoggingMsg fsckSetEventLoggingMsg(false, 0, &nicList, false);
+
+      const auto respMsg = MessagingTk::requestResponse(*node, fsckSetEventLoggingMsg,
+         NETMSGTYPE_FsckSetEventLoggingResp);
+
+      if (respMsg)
+      {
+         auto* fsckSetEventLoggingRespMsg = (FsckSetEventLoggingRespMsg*) respMsg.get();
+
+         bool result = fsckSetEventLoggingRespMsg->getResult();
+         bool missedEvents = fsckSetEventLoggingRespMsg->getMissedEvents();
+
+         if ( result )
+         {
+            nodeIdIter = nodeIDs.erase(nodeIdIter);
+            if ( missedEvents )
+            {
+               retVal = false;
+            }
+         }
+         else
+            nodeIdIter++; // keep in list and try again later
+      }
+      else
+      {
+         LogContext(logContext).logErr("Communication error occurred with node: " + node->getAlias());
+         retVal = false;
+      }
+
+      if (nodeIdIter == nodeIDs.end())
+      {
+         nodeIdIter = nodeIDs.begin();
+         sleep(5);
+      }
+   }
+
+   return retVal;
+}
+
+bool FsckTkEx::checkConsistencyStates()
+{
+   auto mgmtdNode = Program::getApp()->getMgmtNodes()->referenceFirstNode();
+   if (!mgmtdNode)
+      throw std::runtime_error("Management node not found");
+
+   std::list<uint8_t> storageReachabilityStates;
+   std::list<uint8_t> storageConsistencyStates;
+   std::list<uint16_t> storageTargetIDs;
+   std::list<uint8_t> metaReachabilityStates;
+   std::list<uint8_t> metaConsistencyStates;
+   std::list<uint16_t> metaTargetIDs;
+
+   if (!NodesTk::downloadTargetStates(*mgmtdNode, NODETYPE_Storage, &storageTargetIDs,
+           &storageReachabilityStates, &storageConsistencyStates, false)
+         || !NodesTk::downloadTargetStates(*mgmtdNode, NODETYPE_Meta, &metaTargetIDs,
+           &metaReachabilityStates, &metaConsistencyStates, false))
+   {
+      throw std::runtime_error("Could not download target states from management.");
+   }
+
+
+   bool result = true;
+
+   {
+      auto idIt = storageTargetIDs.begin();
+      auto stateIt = storageConsistencyStates.begin();
+      for (; idIt != storageTargetIDs.end() && stateIt != storageConsistencyStates.end();
+            idIt++, stateIt++)
+      {
+         if (*stateIt == TargetConsistencyState_NEEDS_RESYNC)
+         {
+            FsckTkEx::fsckOutput("Storage target " + StringTk::uintToStr(*idIt) + " is set to "
+                  "NEEDS_RESYNC.", OutputOptions_LINEBREAK);
+            result = false;
+         }
+      }
+   }
+
+   {
+      auto idIt = metaTargetIDs.begin();
+      auto stateIt = metaConsistencyStates.begin();
+      for (; idIt != metaTargetIDs.end() && stateIt != metaConsistencyStates.end(); 
+            idIt++, stateIt++)
+      {
+         if (*stateIt == TargetConsistencyState_NEEDS_RESYNC)
+         {
+            FsckTkEx::fsckOutput("Meta node " + StringTk::uintToStr(*idIt) + " is set to "
+                  "NEEDS_RESYNC.", OutputOptions_LINEBREAK);
+            result = false;
+         }
+      }
+   }
+
+   return result;
+}

fsck/source/toolkit/FsckTkEx.h

@@ -0,0 +1,103 @@
+#ifndef FSCKTKEX_H_
+#define FSCKTKEX_H_
+
+#include <common/app/log/LogContext.h>
+#include <common/fsck/FsckDirEntry.h>
+#include <common/fsck/FsckChunk.h>
+#include <common/fsck/FsckContDir.h>
+#include <common/fsck/FsckDirEntry.h>
+#include <common/fsck/FsckDirInode.h>
+#include <common/fsck/FsckFileInode.h>
+#include <common/fsck/FsckFsID.h>
+#include <common/net/message/nodes/HeartbeatRequestMsg.h>
+#include <common/net/message/nodes/HeartbeatMsg.h>
+#include <common/nodes/NodeStore.h>
+#include <common/threading/PThread.h>
+#include <common/toolkit/NodesTk.h>
+#include <common/toolkit/MessagingTk.h>
+#include <toolkit/FsckDefinitions.h>
+
+/* OutputOption Flags for fsckOutput */
+#define OutputOptions_NONE                0
+#define OutputOptions_LINEBREAK           1
+#define OutputOptions_DOUBLELINEBREAK     1 << 1
+#define OutputOptions_HEADLINE            1 << 2
+#define OutputOptions_FLUSH               1 << 3
+#define OutputOptions_ADDLINEBREAKBEFORE  1 << 4
+#define OutputOptions_COLORGREEN          1 << 5
+#define OutputOptions_COLORRED            1 << 6
+#define OutputOptions_LINEDELETE          1 << 7
+#define OutputOptions_NOSTDOUT            1 << 8
+#define OutputOptions_NOLOG               1 << 9
+#define OutputOptions_STDERR              1 << 10
+
+#define OutputColor_NORMAL "\033[0m";
+#define OutputColor_GREEN "\033[32m";
+#define OutputColor_RED "\033[31m";
+
+#define SAFE_FPRINTF(stream, fmt, args...) \
+   do{ if(stream) {fprintf(stream, fmt, ##args);} } while(0)
+
+/*
+ * calculating with:
+ * DirEntry    76+256+28 Byte (space for dentry + longest name + one index)
+ * FileInode   96 Byte    |
+ * DirInode    56 Byte    # only the larger of these two counts, even though files are inlined
+ * ContDir     16 Byte
+ * FsID        40 Byte
+ * chunk       88 Byte
+ *
+ */
+#define NEEDED_DISKSPACE_META_INODE 512
+#define NEEDED_DISKSPACE_STORAGE_INODE 88
+
+class FsckTkEx
+{
+   public:
+      // check the reachability of all nodes
+      static bool checkReachability();
+      // check the reachability of a given node by sending a heartbeat message
+      static bool checkReachability(Node& node, NodeType nodetype);
+
+      /*
+       * a formatted output for fsck
+       *
+       * @param text The text to be printed
+       * @param optionFlags OutputOptions_... flags (mainly for formatiing)
+       */
+      static void fsckOutput(std::string text, int optionFlags);
+      // just print a formatted header with the version to the console
+      static void printVersionHeader(bool toStdErr = false, bool noLogFile = false);
+      // print the progress meter which goes round and round (-\|/-)
+      static void progressMeter();
+
+      static int64_t calcNeededSpace();
+      static bool checkDiskSpace(Path& dbPath);
+
+      static std::string getRepairActionDesc(FsckRepairAction repairAction, bool shortDesc = false);
+
+      static FhgfsOpsErr startModificationLogging(NodeStore* metaNodes, Node& localNode,
+            bool forceRestart);
+      static bool stopModificationLogging(NodeStore* metaNodes);
+
+      static bool checkConsistencyStates();
+
+   private:
+      FsckTkEx() {}
+
+      // saves the last char output by the progress meter
+      static char progressChar;
+      // a mutex that is locked by output functions to make sure the output does not get messed up
+      // by two threads doing output at the same time
+      static Mutex outputMutex;
+
+
+   public:
+      // inliners
+      static void fsckOutput(std::string text)
+      {
+         fsckOutput(text, OutputOptions_LINEBREAK);
+      }
+};
+
+#endif /* FSCKTKEX_H_ */

fsck/tests/FlatTest.cpp

@@ -0,0 +1,21 @@
+#include "FlatTest.h"
+
+#include <program/Program.h>
+#include <common/toolkit/StorageTk.h>
+
+FlatTest::FlatTest()
+{
+   dirName = "./fsck.test.dir/";
+   this->fileName = this->dirName + "set";
+}
+
+void FlatTest::SetUp()
+{
+   if(::mkdir(this->dirName.c_str(), 0770) < 0 && errno != EEXIST)
+      throw std::runtime_error("could not create dir");
+}
+
+void FlatTest::TearDown()
+{
+   StorageTk::removeDirRecursive(this->dirName);
+}

fsck/tests/FlatTest.h

@@ -0,0 +1,31 @@
+#ifndef FLATTEST_H_
+#define FLATTEST_H_
+
+#include <string>
+
+#include <inttypes.h>
+
+#include <gtest/gtest.h>
+
+class FlatTest : public ::testing::Test
+{
+   public:
+      FlatTest();
+
+      void SetUp();
+      void TearDown();
+
+      struct Data {
+         uint64_t id;
+         char dummy[1024 - sizeof(uint64_t)];
+
+         typedef uint64_t KeyType;
+         uint64_t pkey() const { return id; }
+      };
+
+   protected:
+      std::string dirName;
+      std::string fileName;
+};
+
+#endif

fsck/tests/TestConfig.cpp

@@ -0,0 +1,64 @@
+#include "TestConfig.h"
+
+void TestConfig::SetUp()
+{
+   emptyConfigFile = DUMMY_EMPTY_CONFIG_FILE;
+   this->dummyConfigFile = DUMMY_NOEXIST_CONFIG_FILE;
+}
+
+void TestConfig::TearDown()
+{
+   // delete generated config file
+   if ( StorageTk::pathExists(this->emptyConfigFile) )
+   {
+      // return value of remove is ignored now
+      remove(this->emptyConfigFile.c_str());
+   }
+}
+
+TEST_F(TestConfig, defaultConfigFile)
+{
+   log.log(Log_DEBUG, "testDefaultConfigFile started");
+
+   // get the path where the binary resides
+   int BUFSIZE = 255;
+   char exePathBuf[BUFSIZE];
+   // read only BUFSIZE-1, as we need to terminate the string manually later
+   ssize_t len = readlink("/proc/self/exe", exePathBuf, BUFSIZE-1);
+
+   /* In case of an error, failure will indicate the error */
+   if (len < 0)
+      FAIL() << "Internal error";
+
+   /* in case of insufficient buffer size, failure will indicate the error */
+   if (len >= BUFSIZE)
+      FAIL() << "Internal error";
+
+   // readlink does NOT null terminate the string, so we do it here to be safe
+   exePathBuf[len] = '\0';
+
+   // construct the path to the default config file
+   std::string defaultFileName = std::string(dirname(exePathBuf))
+      + "/" + DEFAULT_CONFIG_FILE_RELATIVE;
+
+   // create config with the default file and see what happens while parsing
+   int argc = 2;
+   char* argv[2];
+
+   std::string appNameStr = APP_NAME;
+   appNameStr += '\0';
+
+   std::string cfgLineStr = "cfgFile=" + defaultFileName;
+   cfgLineStr += '\0';
+
+   argv[0] = &appNameStr[0];
+   argv[1] = &cfgLineStr[0];
+
+   try {
+      Config config(argc, argv);
+   } catch (ConnAuthFileException& e) {
+      return;
+   }
+
+   log.log(Log_DEBUG, "testDefaultConfigFile finished");
+}

fsck/tests/TestConfig.h

@@ -0,0 +1,29 @@
+#ifndef TESTCONFIG_H_
+#define TESTCONFIG_H_
+
+#include <app/config/Config.h>
+
+#include <libgen.h>
+#include <gtest/gtest.h>
+#include <common/app/config/ConnAuthFileException.h>
+
+#define DUMMY_NOEXIST_CONFIG_FILE "/tmp/nonExistantConfigFile.fsck"
+#define DUMMY_EMPTY_CONFIG_FILE "/tmp/emptyConfigFile.fsck"
+#define DEFAULT_CONFIG_FILE_RELATIVE "dist/etc/beegfs-client.conf"
+#define APP_NAME "beegfs-fsck";
+
+class TestConfig: public ::testing::Test
+{
+   protected:
+      void SetUp();
+      void TearDown();
+
+      LogContext log;
+
+      // we have these filenames as member variables because
+      // we might need to delete them in tearDown function
+      std::string dummyConfigFile;
+      std::string emptyConfigFile;
+};
+
+#endif /* TESTCONFIG_H_ */

fsck/tests/TestCursors.cpp

@@ -0,0 +1,254 @@
+#include <database/Distinct.h>
+#include <database/Filter.h>
+#include <database/Group.h>
+#include <database/LeftJoinEq.h>
+#include <database/Select.h>
+#include <database/Union.h>
+#include <database/VectorSource.h>
+
+#include <boost/static_assert.hpp>
+
+#include <gtest/gtest.h>
+
+template<typename Obj, size_t Size>
+static VectorSource<Obj> vectorSource(const Obj (&data)[Size])
+{
+   std::vector<Obj> vector(data, data + Size);
+
+   return VectorSource<Obj>(vector);
+}
+
+template<typename Source, typename Data, size_t Size, typename EqFn>
+static void expect(const Source& source, const Data (&data)[Size], EqFn eqFn)
+{
+   BOOST_STATIC_ASSERT(Size > 2);
+
+   for(size_t markPos = 0; markPos < Size; markPos++)
+   {
+      boost::shared_ptr<typename Source::MarkerType> mark;
+      Source current = source;
+
+      for(size_t i = 0; i < Size; i++)
+      {
+         ASSERT_TRUE(current.step());
+
+         if(i == markPos)
+            mark = boost::make_shared<typename Source::MarkerType>(current.mark() );
+
+         ASSERT_TRUE(eqFn(*current.get(), data[i]));
+      }
+
+      ASSERT_FALSE(current.step());
+
+      current.restore(*mark);
+
+      for(size_t i = markPos; i < Size; i++)
+      {
+         ASSERT_TRUE(eqFn(*current.get(), data[i]));
+
+         if(i < Size - 1) {
+            ASSERT_TRUE(current.step());
+         }
+      }
+
+      ASSERT_FALSE(current.step());
+   }
+}
+
+template<typename Source, typename Data, size_t Size>
+static void expect(Source source, const Data (&data)[Size])
+{
+   expect(source, data, std::equal_to<Data>() );
+}
+
+TEST(Cursors, vectorSource)
+{
+   const int data[] = { 1, 2, 3 };
+
+   VectorSource<int> source = vectorSource(data);
+
+   expect(source, data);
+}
+
+TEST(Cursors, union)
+{
+   const std::pair<int, int> data1[] = {
+      std::make_pair(1, 1),
+      std::make_pair(1, 2),
+      std::make_pair(2, 2)
+   };
+
+   const std::pair<int, int> data2[] = {
+      std::make_pair(1, 3),
+      std::make_pair(2, 3)
+   };
+
+   const std::pair<int, int> expected[] = {
+      std::make_pair(1, 3),
+      std::make_pair(1, 1),
+      std::make_pair(1, 2),
+      std::make_pair(2, 3),
+      std::make_pair(2, 2)
+   };
+
+   struct Fn
+   {
+      static int key(std::pair<int, int> pair)
+      {
+         return pair.first;
+      }
+   };
+
+   expect(
+      db::unionBy(
+         Fn::key,
+         vectorSource(data1),
+         vectorSource(data2) ),
+      expected);
+}
+
+TEST(Cursors, select)
+{
+   const int data[] = { 1, 2, 3, 4 };
+
+   const int expected[] = { -1, -2, -3, -4 };
+
+   struct Fn
+   {
+      static int op(int i)
+      {
+         return -i;
+      }
+   };
+
+   expect(
+      vectorSource(data)
+      | db::select(Fn::op),
+      expected);
+}
+
+TEST(Cursors, leftJoinEq)
+{
+   const int data1[] = { 1, 2, 2, 3, 4 };
+   const int data2[] = { 1, 2, 3, 3, 5 };
+
+   const std::pair<int, int> expected[] = {
+      std::make_pair(1, 1),
+      std::make_pair(2, 2),
+      std::make_pair(2, 2),
+      std::make_pair(3, 3),
+      std::make_pair(3, 3),
+      std::make_pair(4, 0),
+   };
+
+   struct Fn
+   {
+      static bool eq(std::pair<int, int*> left, std::pair<int, int> right)
+      {
+         return left.first == right.first &&
+            ( (right.second == 0 && left.second == NULL) ||
+              (right.second != 0 && left.second != NULL && right.second == *left.second) );
+      }
+
+      static int key(int i)
+      {
+         return i;
+      }
+   };
+
+   expect(
+      db::leftJoinBy(
+         Fn::key,
+         vectorSource(data1),
+         vectorSource(data2) ),
+      expected,
+      Fn::eq);
+}
+
+struct TestGroupOps
+{
+   typedef int KeyType;
+   typedef int ProjType;
+   typedef int GroupType;
+
+   int count;
+
+   TestGroupOps()
+      : count(0)
+   {}
+
+   KeyType key(int i)
+   {
+      return i;
+   }
+
+   ProjType project(int i)
+   {
+      return -i;
+   }
+
+   void step(int i)
+   {
+      this->count += 1;
+   }
+
+   int finish()
+   {
+      int result = this->count;
+      this->count = 0;
+      return result;
+   }
+};
+
+TEST(Cursors, group)
+{
+   const int data[] = { 1, 2, 2, 3 };
+   const std::pair<int, int> expected[] = {
+      std::make_pair(-1, 1),
+      std::make_pair(-2, 2),
+      std::make_pair(-3, 1)
+   };
+
+   expect(
+      vectorSource(data)
+      | db::groupBy(TestGroupOps() ),
+      expected);
+}
+
+TEST(Cursors, filter)
+{
+   const int data[] = { 1, 2, 2, 3 };
+   const int expected[] = { 2, 2, 3 };
+
+   struct Fn
+   {
+      static bool fn(int i)
+      {
+         return i > 1;
+      }
+   };
+
+   expect(
+      vectorSource(data)
+      | db::where(Fn::fn),
+      expected);
+}
+
+TEST(Cursors, distinct)
+{
+   const int data[] = { 1, 2, 2, 3 };
+   const int expected[] = { 1, 2, 3 };
+
+   struct Fn
+   {
+      static int fn(int i)
+      {
+         return i;
+      }
+   };
+
+   expect(
+      vectorSource(data)
+      | db::distinctBy(Fn::fn),
+      expected);
+}

fsck/tests/TestDatabase.h

@@ -0,0 +1,115 @@
+#ifndef TESTDATABASE_H_
+#define TESTDATABASE_H_
+
+#include <common/net/message/NetMessage.h>
+#include <database/FsckDB.h>
+
+#include <boost/scoped_ptr.hpp>
+#include <gtest/gtest.h>
+
+class TestDatabase: public ::testing::Test
+{
+   public:
+      TestDatabase();
+      virtual ~TestDatabase();
+
+      void SetUp();
+      void TearDown();
+
+   protected:
+      std::string databasePath;
+      boost::scoped_ptr<FsckDB> db;
+
+      void testInsertAndReadSingleDentry();
+      void testInsertAndReadDentries();
+
+      void testUpdateDentries();
+      void testUpdateDirInodes();
+      void testUpdateFileInodes();
+      void testUpdateChunks();
+
+      void testInsertAndReadSingleFileInode();
+
+      void testInsertAndReadSingleDirInode();
+      void testInsertAndReadDirInodes();
+
+      void testInsertAndReadSingleChunk();
+      void testInsertAndReadChunks();
+
+      void testInsertAndReadSingleContDir();
+      void testInsertAndReadContDirs();
+
+      void testInsertAndReadSingleFsID();
+      void testInsertAndReadFsIDs();
+
+      void testFindDuplicateInodeIDs();
+      void testFindDuplicateChunks();
+      void testFindDuplicateContDirs();
+      void testFindMismirroredDentries();
+      void testFindMismirroredDirectories();
+
+      void testCheckForAndInsertDanglingDirEntries();
+      void testCheckForAndInsertInodesWithWrongOwner();
+      void testCheckForAndInsertDirEntriesWithWrongOwner();
+      void testCheckForAndInsertMissingDentryByIDFile();
+      void testCheckForAndInsertOrphanedDentryByIDFiles();
+      void testCheckForAndInsertOrphanedDirInodes();
+      void testCheckForAndInsertOrphanedFileInodes();
+      void testCheckForAndInsertOrphanedChunks();
+      void testCheckForAndInsertInodesWithoutContDir();
+      void testCheckForAndInsertOrphanedContDirs();
+      void testCheckForAndInsertFileInodesWithWrongAttribs();
+      void testCheckForAndInsertDirInodesWithWrongAttribs();
+      void testCheckForAndInsertFilesWithMissingStripeTargets();
+      void testCheckForAndInsertChunksWithWrongPermissions();
+      void testCheckForAndInsertChunksInWrongPath();
+
+      void testDeleteDentries();
+      void testDeleteChunks();
+      void testDeleteFsIDs();
+
+      void testGetRowCount();
+
+      void testGetFullPath();
+
+      template<typename Table, typename Items>
+      static bool setRepairActions(Table* table, std::list<Items>& items, FsckRepairAction action)
+      {
+         for (typename std::list<Items>::iterator it = items.begin(), end = items.end();
+               it != end;
+               ++it)
+         {
+            if (!table->setRepairAction(*it, action) )
+               return false;
+         }
+
+         return true;
+      }
+
+      template<typename Item, typename Source>
+      static void drainToList(Source source, std::list<Item>& list)
+      {
+         list.clear();
+         while(source.step() )
+            list.push_back(*source.get() );
+      }
+
+      template<typename Item, typename Source>
+      static std::list<Item> drain(Source source)
+      {
+         std::list<Item> list;
+         drainToList(source, list);
+         return list;
+      }
+
+      template<typename Source>
+      static size_t countCursor(Source source)
+      {
+         size_t result = 0;
+         while(source.step() )
+            result++;
+         return result;
+      }
+};
+
+#endif /* TESTDATABASE_H_ */

fsck/tests/TestFsckTk.cpp

@@ -0,0 +1,35 @@
+#include <common/toolkit/FsckTk.h>
+#include <common/toolkit/StringTk.h>
+#include <toolkit/DatabaseTk.h>
+#include <toolkit/FsckTkEx.h>
+
+#include <gtest/gtest.h>
+
+TEST(FsckTk, dirEntryTypeConversion)
+{
+   FsckDirEntryType entryTypeOut;
+
+   entryTypeOut = FsckTk::DirEntryTypeToFsckDirEntryType(DirEntryType_INVALID);
+   ASSERT_EQ(entryTypeOut, FsckDirEntryType_INVALID);
+
+   entryTypeOut = FsckTk::DirEntryTypeToFsckDirEntryType(DirEntryType_REGULARFILE);
+   ASSERT_EQ(entryTypeOut, FsckDirEntryType_REGULARFILE);
+
+   entryTypeOut = FsckTk::DirEntryTypeToFsckDirEntryType(DirEntryType_SYMLINK);
+   ASSERT_EQ(entryTypeOut, FsckDirEntryType_SYMLINK);
+
+   entryTypeOut = FsckTk::DirEntryTypeToFsckDirEntryType(DirEntryType_DIRECTORY);
+   ASSERT_EQ(entryTypeOut, FsckDirEntryType_DIRECTORY);
+
+   entryTypeOut = FsckTk::DirEntryTypeToFsckDirEntryType(DirEntryType_BLOCKDEV);
+   ASSERT_EQ(entryTypeOut, FsckDirEntryType_SPECIAL);
+
+   entryTypeOut = FsckTk::DirEntryTypeToFsckDirEntryType(DirEntryType_CHARDEV);
+   ASSERT_EQ(entryTypeOut, FsckDirEntryType_SPECIAL);
+
+   entryTypeOut = FsckTk::DirEntryTypeToFsckDirEntryType(DirEntryType_FIFO);
+   ASSERT_EQ(entryTypeOut, FsckDirEntryType_SPECIAL);
+
+   entryTypeOut = FsckTk::DirEntryTypeToFsckDirEntryType(DirEntryType_SOCKET);
+   ASSERT_EQ(entryTypeOut, FsckDirEntryType_SPECIAL);
+}

fsck/tests/TestSerialization.cpp

@@ -0,0 +1,77 @@
+#include <common/net/message/nodes/HeartbeatRequestMsg.h>
+#include <common/net/message/nodes/HeartbeatMsg.h>
+#include <common/toolkit/ZipIterator.h>
+#include <net/message/NetMessageFactory.h>
+#include <toolkit/DatabaseTk.h>
+
+#include <gtest/gtest.h>
+
+template<typename Obj>
+static void testObjectRoundTrip(Obj& data)
+{
+   Serializer ser;
+   ser % data;
+
+   std::vector<char> buffer(ser.size());
+
+   ser = Serializer(&buffer[0], buffer.size());
+   ser % data;
+   ASSERT_TRUE(ser.good());
+   ASSERT_EQ(ser.size(), buffer.size());
+
+   Deserializer des(&buffer[0], buffer.size());
+
+   Obj read;
+
+   des % read;
+   ASSERT_TRUE(des.good());
+   ASSERT_EQ(des.size(), buffer.size());
+
+   ser = Serializer();
+   ser % read;
+   ASSERT_EQ(ser.size(), buffer.size());
+
+   std::vector<char> rtBuffer(buffer.size());
+
+   ser = Serializer(&rtBuffer[0], rtBuffer.size());
+   ser % read;
+   ASSERT_TRUE(ser.good());
+   ASSERT_EQ(ser.size(), rtBuffer.size());
+   ASSERT_EQ(buffer, rtBuffer);
+}
+
+TEST(Serialization, fsckDirEntrySerialization)
+{
+   FsckDirEntry dirEntryIn = DatabaseTk::createDummyFsckDirEntry();
+   testObjectRoundTrip(dirEntryIn);
+}
+
+TEST(Serialization, fsckDirInodeSerialization)
+{
+   FsckDirInode dirInodeIn = DatabaseTk::createDummyFsckDirInode();
+   testObjectRoundTrip(dirInodeIn);
+}
+
+TEST(Serialization, fsckFileInodeSerialization)
+{
+   FsckFileInode fileInodeIn = DatabaseTk::createDummyFsckFileInode();
+   testObjectRoundTrip(fileInodeIn);
+}
+
+TEST(Serialization, fsckChunkSerialization)
+{
+   FsckChunk chunkIn = DatabaseTk::createDummyFsckChunk();
+   testObjectRoundTrip(chunkIn);
+}
+
+TEST(Serialization, fsckContDirSerialization)
+{
+   FsckContDir contDirIn = DatabaseTk::createDummyFsckContDir();
+   testObjectRoundTrip(contDirIn);
+}
+
+TEST(Serialization, fsckFsIDSerialization)
+{
+   FsckFsID fsIDIn = DatabaseTk::createDummyFsckFsID();
+   testObjectRoundTrip(fsIDIn);
+}

fsck/tests/TestSet.cpp

@@ -0,0 +1,177 @@
+#include <database/Set.h>
+#include <program/Program.h>
+
+#include "FlatTest.h"
+
+class TestSet : public FlatTest {
+};
+
+TEST_F(TestSet, open)
+{
+   // break config
+   int cfgFile = ::creat((fileName + ".t").c_str(), 0660);
+   ASSERT_GE(cfgFile, 0);
+   ASSERT_EQ(::truncate( (this->fileName + ".t").c_str(), 5), 0);
+   do {
+      try {
+         Set<Data> set(this->fileName);
+      } catch (const std::runtime_error& e) {
+         ASSERT_NE(std::string(e.what()).find("cannot read set"), std::string::npos);
+         ASSERT_NE(std::string(e.what()).find("with version 0"), std::string::npos);
+         break;
+      }
+
+      FAIL();
+   } while (false);
+   ASSERT_EQ(::write(cfgFile, "v1\n1\n", 5), 5);
+   do {
+      try {
+         Set<Data> set(this->fileName);
+      } catch (const std::runtime_error& e) {
+         ASSERT_NE(std::string(e.what()).find("bad set description"), std::string::npos);
+         break;
+      }
+
+      FAIL();
+   } while (false);
+   ASSERT_EQ(::close(cfgFile), 0);
+   ASSERT_EQ(::unlink((this->fileName + ".t").c_str()), 0);
+
+   // reopen with saved config
+   {
+      Set<Data> set(this->fileName);
+   }
+   {
+      Set<Data> set(this->fileName);
+   }
+
+   ASSERT_EQ(::truncate((this->fileName + ".t").c_str(), 5000), 0);
+   do {
+      try {
+         Set<Data> set(this->fileName);
+      } catch (const std::runtime_error& e) {
+         ASSERT_NE(std::string(e.what()).find("bad set description"), std::string::npos);
+         break;
+      }
+
+      FAIL();
+   } while (false);
+}
+
+TEST_F(TestSet, drop)
+{
+   {
+      Set<Data> set(this->fileName);
+      set.newFragment();
+   }
+   {
+      Set<Data> set(this->fileName);
+      set.drop();
+   }
+
+   ASSERT_EQ(::rmdir(this->dirName.c_str()), 0);
+}
+
+TEST_F(TestSet, clear)
+{
+   {
+      Set<Data> set(this->fileName);
+      set.newFragment();
+   }
+   {
+      Set<Data> set(this->fileName);
+      set.clear();
+   }
+
+   ASSERT_EQ(::unlink( (this->fileName + ".t").c_str()), 0);
+   ASSERT_EQ(::rmdir(this->dirName.c_str()), 0);
+}
+
+TEST_F(TestSet, merge)
+{
+   Data d = {0, {}};
+
+   {
+      Set<Data> set1(this->fileName + "1");
+      set1.newFragment()->append(d);
+
+      Set<Data> set2(this->fileName + "2");
+      set2.newFragment()->append(d);
+   }
+
+   {
+      Set<Data> set1(this->fileName + "1");
+      Set<Data> set2(this->fileName + "2");
+      set2.mergeInto(set1);
+   }
+
+   {
+      Set<Data> set1(this->fileName + "1");
+      set1.drop();
+   }
+
+   ASSERT_EQ(::unlink((this->fileName + "2.t").c_str()), 0);
+   ASSERT_EQ(::rmdir(this->dirName.c_str()), 0);
+}
+
+TEST_F(TestSet, dataOps)
+{
+   static const unsigned FRAG_COUNT = 2 * 3 * 4;
+
+   Set<Data> set(this->fileName);
+
+   ASSERT_FALSE(set.cursor().step());
+
+   {
+      for(unsigned i = 0; i < FRAG_COUNT; i++)
+      {
+         Data d = { i, {} };
+         SetFragment<Data>* frag = set.newFragment();
+
+         frag->append(d);
+         frag->append(d);
+
+         ASSERT_EQ(set.size(), 2 * (i + 1));
+      }
+   }
+
+   {
+      Set<Data>::Cursor cursor = set.cursor();
+
+      for(unsigned i = 0; i < FRAG_COUNT; i++) {
+         ASSERT_TRUE(cursor.step());
+         ASSERT_EQ(cursor.get()->id, i);
+         ASSERT_TRUE(cursor.step());
+         ASSERT_EQ(cursor.get()->id, i);
+      }
+
+      ASSERT_FALSE(cursor.step());
+   }
+
+   set.makeUnique();
+   ASSERT_EQ(set.size(), FRAG_COUNT);
+
+   {
+      Set<Data>::Cursor cursor = set.cursor();
+
+      for(unsigned i = 0; i < FRAG_COUNT; i++) {
+         ASSERT_TRUE(cursor.step());
+         ASSERT_EQ(cursor.get()->id, i);
+      }
+
+      ASSERT_FALSE(cursor.step());
+   }
+
+   ASSERT_TRUE(set.getByKey(FRAG_COUNT / 2).first);
+   ASSERT_EQ(set.getByKey(FRAG_COUNT / 2).second.id, FRAG_COUNT / 2);
+
+   ASSERT_FALSE(set.getByKey(FRAG_COUNT * 2).first);
+
+   struct ops
+   {
+      static int64_t key(uint64_t key) { return key + 6; }
+   };
+
+   ASSERT_TRUE(set.getByKeyProjection(7, ops::key).first);
+   ASSERT_EQ(set.getByKeyProjection(7, ops::key).second.id, 1u);
+}

fsck/tests/TestSetFragment.cpp

@@ -0,0 +1,203 @@
+#include <database/SetFragment.h>
+#include <program/Program.h>
+
+#include "FlatTest.h"
+
+class TestSetFragment : public FlatTest {
+};
+
+TEST_F(TestSetFragment, open)
+{
+   if(::getuid() == 0)
+   {
+      std::cerr << "Test run as root, skipping DAC tests\n";
+   }
+   else
+   {
+      // bad directory permissions (can't open or create)
+      ASSERT_EQ(::chmod(this->dirName.c_str(), 0), 0);
+      do {
+         try {
+            SetFragment<Data> frag(this->fileName);
+         } catch (const std::runtime_error& e) {
+            ASSERT_NE(std::string(e.what()).find("could not open fragment file"), std::string::npos);
+            break;
+         }
+
+         FAIL();
+      } while (false);
+      ASSERT_EQ(::chmod(this->dirName.c_str(), 0770), 0);
+   }
+
+   // not a regular file
+   ASSERT_EQ(::mkfifo(this->fileName.c_str(), 0660), 0);
+   do {
+      try {
+         SetFragment<Data> frag(this->fileName);
+      } catch (const std::runtime_error& e) {
+         ASSERT_NE(std::string(e.what()).find("error while opening fragment file"), std::string::npos);
+         break;
+      }
+
+      FAIL();
+   } while (false);
+   ASSERT_EQ(::unlink(this->fileName.c_str()), 0);
+
+   // corrupted config area
+   ASSERT_EQ(::close(::creat(this->fileName.c_str(), 0660)), 0);
+   ASSERT_EQ(::truncate(this->fileName.c_str(), SetFragment<Data>::CONFIG_AREA_SIZE - 1), 0);
+   do {
+      try {
+         SetFragment<Data> frag(this->fileName);
+      } catch (const std::runtime_error& e) {
+         ASSERT_NE(std::string(e.what()).find("error while opening fragment file"), std::string::npos);
+         break;
+      }
+
+      FAIL();
+   } while (false);
+   ASSERT_EQ(::unlink(this->fileName.c_str()), 0);
+
+   // corrupted data
+   ASSERT_EQ(::close(::creat(this->fileName.c_str(), 0660)), 0);
+   ASSERT_EQ(::truncate(this->fileName.c_str(),
+      SetFragment<Data>::CONFIG_AREA_SIZE + sizeof(Data) - 1), 0);
+   do {
+      try {
+         SetFragment<Data> frag(this->fileName);
+      } catch (const std::runtime_error& e) {
+         ASSERT_NE(std::string(e.what()).find("error while opening fragment file"), std::string::npos);
+         break;
+      }
+
+      FAIL();
+   } while (false);
+   ASSERT_EQ(::unlink(this->fileName.c_str()), 0);
+
+   // should work
+   SetFragment<Data> frag(this->fileName);
+}
+
+TEST_F(TestSetFragment, drop)
+{
+   struct stat stat;
+
+   {
+      SetFragment<Data> frag(this->fileName);
+   }
+
+   ASSERT_EQ(::stat(this->fileName.c_str(), &stat), 0);
+
+   {
+      SetFragment<Data> frag(this->fileName);
+      frag.drop();
+   }
+
+   ASSERT_EQ(::stat(this->fileName.c_str(), &stat), -1);
+   ASSERT_EQ(errno, ENOENT);
+}
+
+TEST_F(TestSetFragment, flush)
+{
+   SetFragment<Data> frag(this->fileName);
+   Data d = {0, {}};
+   struct stat stat;
+
+   frag.append(d);
+
+   ASSERT_EQ(::stat(this->fileName.c_str(), &stat), 0);
+   ASSERT_EQ(stat.st_size, 0);
+
+   frag.flush();
+
+   ASSERT_EQ(::stat(this->fileName.c_str(), &stat), 0);
+   ASSERT_EQ(size_t(stat.st_size), frag.CONFIG_AREA_SIZE + sizeof(d));
+}
+
+TEST_F(TestSetFragment, appendAndAccess)
+{
+   struct ops
+   {
+      static void fill(SetFragment<Data>& frag, unsigned limit)
+      {
+         Data d = {0, {}};
+
+         for (unsigned i = 0; i < limit; i++) {
+            d.id = i;
+            frag.append(d);
+         }
+      }
+
+      static void check(SetFragment<Data>& frag, unsigned limit)
+      {
+         ASSERT_EQ(frag.size(), limit);
+
+         for (unsigned i = 0; i < limit; i++) {
+            ASSERT_EQ(frag[i].id, i);
+            ASSERT_EQ(
+               size_t(std::count(
+                  frag[i].dummy,
+                  frag[i].dummy + sizeof(frag[i].dummy),
+                  0)),
+               sizeof(frag[i].dummy));
+         }
+      }
+   };
+
+   static const unsigned LIMIT = 2 * SetFragment<Data>::BUFFER_SIZE / sizeof(Data);
+
+   {
+      SetFragment<Data> frag(this->fileName);
+
+      ops::fill(frag, LIMIT);
+      ops::check(frag, LIMIT);
+   }
+
+   // read everything again, with a new instance
+   {
+      SetFragment<Data> frag(this->fileName);
+
+      ops::check(frag, LIMIT);
+   }
+}
+
+TEST_F(TestSetFragment, sortAndGet)
+{
+   SetFragment<Data> frag(this->fileName);
+
+   // generate a pseudo-random sequence as {0, 1, ...} * 4001 mod 4003 (both prime)
+   for (unsigned i = 0; i < 4003; i++)
+   {
+      Data d = { i * 4001 % 4003, {} };
+      frag.append(d);
+   }
+
+   frag.sort();
+
+   for (unsigned i = 0; i < 4003; i++)
+      ASSERT_EQ(frag[i].id, i);
+
+   ASSERT_TRUE(frag.getByKey(17).first);
+   ASSERT_EQ(frag.getByKey(17).second.id, 17u);
+
+   ASSERT_TRUE(!frag.getByKey(170000).first);
+
+   struct ops
+   {
+      static int64_t key(uint64_t key) { return key + 6; }
+   };
+
+   ASSERT_TRUE(frag.getByKeyProjection(23, ops::key).first);
+   ASSERT_EQ(frag.getByKeyProjection(23, ops::key).second.id, 17u);
+}
+
+TEST_F(TestSetFragment, rename)
+{
+   SetFragment<Data> frag(this->fileName);
+
+   frag.rename(this->fileName + "1");
+
+   ASSERT_EQ(::close(::open( (this->fileName + "1").c_str(), O_RDWR)), 0);
+
+   frag.rename(this->fileName);
+}

fsck/tests/TestTable.cpp

@@ -0,0 +1,182 @@
+#include "TestTable.h"
+
+void TestTable::SetUp()
+{
+   FlatTest::SetUp();
+   this->table.reset(new Table<Data>(this->fileName, 4096) );
+}
+
+void TestTable::TearDown()
+{
+   if(this->table)
+      this->table->drop();
+
+   FlatTest::TearDown();
+}
+
+TEST_F(TestTable, dataHandling)
+{
+   Data d = {0, {}};
+
+   // first, only ordered inserts
+   d.id = 0;
+   this->table->insert(d);
+   d.id = 1;
+   this->table->insert(d);
+
+   // complete the base set
+   this->table->commitChanges();
+
+   // insert "in order"
+   d.id = 2;
+   this->table->insert(d);
+
+   // unordered delete should fail
+   ASSERT_THROW(this->table->remove(1), std::runtime_error);
+
+   // ordered delete should work
+   this->table->remove(3);
+
+   // reinsert of same id should work
+   d.id = 3;
+   this->table->insert(d);
+
+   // unordered insert should now fail
+   ASSERT_THROW(this->table->insert(d), std::runtime_error);
+
+   // reset changes
+   this->table->commitChanges();
+
+   // unordered insert should work
+   d.id = 0;
+   this->table->insert(d);
+   d.id = 4;
+   this->table->insert(d);
+   d.id = 1;
+   this->table->insert(d);
+
+   // delete should fail
+   ASSERT_THROW(this->table->remove(1), std::runtime_error);
+
+   // reset changes
+   this->table->commitChanges();
+
+   // ordered delete should work
+   this->table->remove(1);
+   this->table->remove(2);
+   this->table->remove(3);
+
+   // unordered insert should fail
+   d.id = 2;
+   ASSERT_THROW(this->table->insert(d), std::runtime_error);
+
+   // unordered delete should work
+   this->table->remove(3);
+
+   // insert should fail after unordered delete
+   d.id = 10;
+   ASSERT_THROW(this->table->insert(d), std::runtime_error);
+
+   // uncomitted changes should yield the sequence [0, 0, 1, 1, 2, 3, 4]
+   {
+      Table<Data>::QueryType cursor = this->table->cursor();
+
+      unsigned expect[7] = { 0, 0, 1, 1, 2, 3, 4 };
+      for(unsigned i = 0; i < 7; i++)
+      {
+         ASSERT_TRUE(cursor.step());
+         ASSERT_EQ(cursor.get()->id, expect[i]);
+      }
+
+      ASSERT_FALSE(cursor.step());
+   }
+
+   // commited should contain the sequence [0, 0, 4]
+   this->table->commitChanges();
+   {
+      Table<Data>::QueryType cursor = this->table->cursor();
+
+      ASSERT_TRUE(cursor.step());
+      ASSERT_EQ(cursor.get()->id, 0u);
+      ASSERT_TRUE(cursor.step());
+      ASSERT_EQ(cursor.get()->id, 0u);
+      ASSERT_TRUE(cursor.step());
+      ASSERT_EQ(cursor.get()->id, 4u);
+      ASSERT_FALSE(cursor.step());
+   }
+
+   ASSERT_TRUE(this->table->getByKey(4).first);
+   ASSERT_EQ(this->table->getByKey(4).second.id, 4u);
+
+   ASSERT_FALSE(this->table->getByKey(3).first);
+
+   struct ops
+   {
+      static int64_t key(uint64_t key) { return key + 6; }
+   };
+
+   ASSERT_TRUE(this->table->getByKeyProjection(10, ops::key).first);
+   ASSERT_EQ(this->table->getByKeyProjection(10, ops::key).second.id, 4u);
+}
+
+TEST_F(TestTable, bulkInsert)
+{
+   {
+      boost::shared_ptr<Buffer<Data> > buf1, buf2;
+
+      buf1 = this->table->bulkInsert();
+      buf2 = this->table->bulkInsert();
+
+      Data d = { 0, {} };
+
+      d.id = 1;
+      buf1->append(d);
+
+      d.id = 2;
+      buf2->append(d);
+   }
+
+   // all insert must have gone to base, queryable immediatly
+   {
+      Table<Data>::QueryType cursor = this->table->cursor();
+
+      ASSERT_TRUE(cursor.step());
+      ASSERT_EQ(cursor.get()->id, 1u);
+      ASSERT_TRUE(cursor.step());
+      ASSERT_EQ(cursor.get()->id, 2u);
+      ASSERT_FALSE(cursor.step());
+   }
+
+   // another bulk insert is fine, no modifications happened yet
+   {
+      boost::shared_ptr<Buffer<Data> > buf;
+
+      buf = this->table->bulkInsert();
+
+      Data d = { 3, {} };
+      buf->append(d);
+
+      // can't do normal insert or remove during bulk insert
+      ASSERT_THROW(this->table->insert(d), std::runtime_error);
+      ASSERT_THROW(this->table->remove(4), std::runtime_error);
+   }
+
+   // all insert must have gone to base, queryable immediatly
+   {
+      Table<Data>::QueryType cursor = this->table->cursor();
+
+      ASSERT_TRUE(cursor.step());
+      ASSERT_EQ(cursor.get()->id, 1u);
+      ASSERT_TRUE(cursor.step());
+      ASSERT_EQ(cursor.get()->id, 2u);
+      ASSERT_TRUE(cursor.step() );
+      ASSERT_EQ(cursor.get()->id, 3u);
+      ASSERT_FALSE(cursor.step());
+   }
+
+   this->table->remove(1);
+   ASSERT_THROW(this->table->bulkInsert(), std::runtime_error);
+
+   this->table->commitChanges();
+   ASSERT_THROW(this->table->bulkInsert(), std::runtime_error);
+}

fsck/tests/TestTable.h

@@ -0,0 +1,20 @@
+#ifndef TESTTABLE_H_
+#define TESTTABLE_H_
+
+#include <database/Table.h>
+#include "FlatTest.h"
+
+#include <boost/scoped_ptr.hpp>
+#include <gtest/gtest.h>
+
+class TestTable : public FlatTest
+{
+   public:
+      void SetUp();
+      void TearDown();
+
+   protected:
+      boost::scoped_ptr<Table<Data> > table;
+};
+
+#endif