Bug #28751 Lots of memory locked in memory causes high kswapd

- add odirect option for lcp+backup+redo log to lower CPU/kswapd usage
- writing odirect removes need for kernel write buffers avoiding kswapd to kick in

mysql-test/ndb/ndb_config_2_node.ini

@@ -12,6 +12,7 @@ MaxNoOfAttributes= CHOOSE_MaxNoOfAttributes
 TimeBetweenGlobalCheckpoints= 500
 NoOfFragmentLogFiles= 3
 DiskPageBufferMemory= CHOOSE_DiskPageBufferMemory
+ODirect= 1
 # the following parametes just function as a small regression
 # test that the parameter exists
 InitialNoOfOpenFiles= 27

storage/ndb/include/mgmapi/mgmapi_config_parameters.h

@@ -115,6 +115,8 @@
 
 #define CFG_DB_MEMREPORT_FREQUENCY    166
 
+#define CFG_DB_O_DIRECT               168
+
 #define CFG_DB_SGA                    198 /* super pool mem */
 #define CFG_DB_DATA_MEM_2             199 /* used in special build in 5.1 */
 

storage/ndb/include/ndb_global.h.in

@@ -146,4 +146,6 @@ extern "C" {
 #define MAX(x,y) (((x)>(y))?(x):(y))
 #endif
 
+#define NDB_O_DIRECT_WRITE_ALIGNMENT 512
+
 #endif

storage/ndb/src/kernel/blocks/backup/Backup.cpp

@@ -2761,6 +2761,8 @@ Backup::openFiles(Signal* signal, BackupRecordPtr ptr)
   c_backupFilePool.getPtr(filePtr, ptr.p->dataFilePtr);
   filePtr.p->m_flags |= BackupFile::BF_OPENING;
 
+  if (c_defaults.m_o_direct)
+    req->fileFlags |= FsOpenReq::OM_DIRECT;
   req->userPointer = filePtr.i;
   FsOpenReq::setVersion(req->fileNumber, 2);
   FsOpenReq::setSuffix(req->fileNumber, FsOpenReq::S_DATA);
@@ -3735,12 +3737,31 @@ Backup::OperationRecord::newFragment(Uint32 tableId, Uint32 fragNo)
 }
 
 bool
-Backup::OperationRecord::fragComplete(Uint32 tableId, Uint32 fragNo)
+Backup::OperationRecord::fragComplete(Uint32 tableId, Uint32 fragNo, bool fill_record)
 {
   Uint32 * tmp;
   const Uint32 footSz = sizeof(BackupFormat::DataFile::FragmentFooter) >> 2;
+  Uint32 sz = footSz + 1;
 
-  if(dataBuffer.getWritePtr(&tmp, footSz + 1)) {
+  if (fill_record)
+  {
+    Uint32 * new_tmp;
+    if (!dataBuffer.getWritePtr(&tmp, sz))
+      return false;
+    new_tmp = tmp + sz;
+
+    if ((UintPtr)new_tmp & (sizeof(Page32)-1))
+    {
+      /* padding is needed to get full write */
+      new_tmp += 2 /* to fit empty header minimum 2 words*/;
+      new_tmp = (Uint32 *)(((UintPtr)new_tmp + sizeof(Page32)-1) &
+                            ~(UintPtr)(sizeof(Page32)-1));
+      /* new write sz */
+      sz = new_tmp - tmp;
+    }
+  }
+
+  if(dataBuffer.getWritePtr(&tmp, sz)) {
     jam();
     * tmp = 0; // Finish record stream
     tmp++;
@@ -3752,7 +3773,17 @@ Backup::OperationRecord::fragComplete(Uint32 tableId, Uint32 fragNo)
     foot->FragmentNo    = htonl(fragNo);
     foot->NoOfRecords   = htonl(noOfRecords);
     foot->Checksum      = htonl(0);
-    dataBuffer.updateWritePtr(footSz + 1);
+
+    if (sz != footSz + 1)
+    {
+      tmp += footSz;
+      memset(tmp, 0, (sz - footSz - 1) * 4);
+      *tmp = htonl(BackupFormat::EMPTY_ENTRY);
+      tmp++;
+      *tmp = htonl(sz - footSz - 1);
+    }
+
+    dataBuffer.updateWritePtr(sz);
     return true;
   }//if
   return false;
@@ -3854,8 +3885,13 @@ Backup::fragmentCompleted(Signal* signal, BackupFilePtr filePtr)
     return;
   }//if
     
+  BackupRecordPtr ptr LINT_SET_PTR;
+  c_backupPool.getPtr(ptr, filePtr.p->backupPtr);
+
   OperationRecord & op = filePtr.p->operation;
-  if(!op.fragComplete(filePtr.p->tableId, filePtr.p->fragmentNo)) {
+  if(!op.fragComplete(filePtr.p->tableId, filePtr.p->fragmentNo,
+                      c_defaults.m_o_direct))
+  {
     jam();
     signal->theData[0] = BackupContinueB::BUFFER_FULL_FRAG_COMPLETE;
     signal->theData[1] = filePtr.i;
@@ -3865,9 +3901,6 @@ Backup::fragmentCompleted(Signal* signal, BackupFilePtr filePtr)
   
   filePtr.p->m_flags &= ~(Uint32)BackupFile::BF_SCAN_THREAD;
   
-  BackupRecordPtr ptr LINT_SET_PTR;
-  c_backupPool.getPtr(ptr, filePtr.p->backupPtr);
-
   if (ptr.p->is_lcp())
   {
     ptr.p->slaveState.setState(STOPPING);
@@ -4905,6 +4938,8 @@ Backup::lcp_open_file(Signal* signal, BackupRecordPtr ptr)
     FsOpenReq::OM_CREATE | 
     FsOpenReq::OM_APPEND |
     FsOpenReq::OM_AUTOSYNC;
+  if (c_defaults.m_o_direct)
+    req->fileFlags |= FsOpenReq::OM_DIRECT;
   FsOpenReq::v2_setCount(req->fileNumber, 0xFFFFFFFF);
   req->auto_sync_size = c_defaults.m_disk_synch_size;
   

storage/ndb/src/kernel/blocks/backup/Backup.hpp

@@ -240,7 +240,7 @@ public:
      * Once per fragment
      */
     bool newFragment(Uint32 tableId, Uint32 fragNo);
-    bool fragComplete(Uint32 tableId, Uint32 fragNo);
+    bool fragComplete(Uint32 tableId, Uint32 fragNo, bool fill_record);
     
     /**
      * Once per scan frag (next) req/conf
@@ -534,6 +534,7 @@ public:
     Uint32 m_disk_write_speed;
     Uint32 m_disk_synch_size;
     Uint32 m_diskless;
+    Uint32 m_o_direct;
   };
   
   /**

storage/ndb/src/kernel/blocks/backup/BackupFormat.hpp

@@ -32,7 +32,8 @@ struct BackupFormat {
     TABLE_LIST        = 4,
     TABLE_DESCRIPTION = 5,
     GCP_ENTRY         = 6,
-    FRAGMENT_INFO     = 7
+    FRAGMENT_INFO     = 7,
+    EMPTY_ENTRY       = 8
   };
 
   struct FileHeader {
@@ -93,6 +94,13 @@ struct BackupFormat {
       Uint32 NoOfRecords;
       Uint32 Checksum;
     };
+
+    /* optional padding for O_DIRECT */
+    struct EmptyEntry {
+      Uint32 SectionType;
+      Uint32 SectionLength;
+      /* not used data */
+    };
   };
 
   /**

storage/ndb/src/kernel/blocks/backup/BackupInit.cpp

@@ -148,10 +148,13 @@ Backup::execREAD_CONFIG_REQ(Signal* signal)
   c_defaults.m_disk_write_speed = 10 * (1024 * 1024);
   c_defaults.m_disk_write_speed_sr = 100 * (1024 * 1024);
   c_defaults.m_disk_synch_size = 4 * (1024 * 1024);
+  c_defaults.m_o_direct = true;
 
   Uint32 noBackups = 0, noTables = 0, noAttribs = 0, noFrags = 0;
   ndbrequire(!ndb_mgm_get_int_parameter(p, CFG_DB_DISCLESS, 
 					&c_defaults.m_diskless));
+  ndb_mgm_get_int_parameter(p, CFG_DB_O_DIRECT,
+                            &c_defaults.m_o_direct);
   ndb_mgm_get_int_parameter(p, CFG_DB_CHECKPOINT_SPEED_SR,
 			    &c_defaults.m_disk_write_speed_sr);
   ndb_mgm_get_int_parameter(p, CFG_DB_CHECKPOINT_SPEED,
@@ -204,7 +207,7 @@ Backup::execREAD_CONFIG_REQ(Signal* signal)
     / sizeof(Page32);
   // We need to allocate an additional of 2 pages. 1 page because of a bug in
   // ArrayPool and another one for DICTTAINFO.
-  c_pagePool.setSize(noPages + NO_OF_PAGES_META_FILE + 2); 
+  c_pagePool.setSize(noPages + NO_OF_PAGES_META_FILE + 2, true); 
   
   { // Init all tables
     SLList<Table> tables(c_tablePool);

storage/ndb/src/kernel/blocks/backup/FsBuffer.hpp

@@ -270,8 +270,8 @@ FsBuffer::getReadPtr(Uint32 ** ptr, Uint32 * sz, bool * _eof){
     
     * ptr = &Tp[Tr];
 
-    DEBUG(ndbout_c("getReadPtr() Tr: %d Tw: %d Ts: %d Tm: %d sz1: %d -> %d",
-		   Tr, Tw, Ts, Tm, sz1, * sz));
+    DEBUG(ndbout_c("getReadPtr() Tr: %d Tmw: %d Ts: %d Tm: %d sz1: %d -> %d",
+		   Tr, Tmw, Ts, Tm, sz1, * sz));
 
     return true;
   }
@@ -279,8 +279,8 @@ FsBuffer::getReadPtr(Uint32 ** ptr, Uint32 * sz, bool * _eof){
   if(!m_eof){
     * _eof = false;
     
-    DEBUG(ndbout_c("getReadPtr() Tr: %d Tw: %d Ts: %d Tm: %d sz1: %d -> false",
-		   Tr, Tw, Ts, Tm, sz1));
+    DEBUG(ndbout_c("getReadPtr() Tr: %d Tmw: %d Ts: %d Tm: %d sz1: %d -> false",
+		   Tr, Tmw, Ts, Tm, sz1));
     
     return false;
   }
@@ -289,8 +289,8 @@ FsBuffer::getReadPtr(Uint32 ** ptr, Uint32 * sz, bool * _eof){
   * _eof = true;
   * ptr = &Tp[Tr];
 
-  DEBUG(ndbout_c("getReadPtr() Tr: %d Tw: %d Ts: %d Tm: %d sz1: %d -> %d eof",
-		 Tr, Tw, Ts, Tm, sz1, * sz));
+  DEBUG(ndbout_c("getReadPtr() Tr: %d Tmw: %d Ts: %d Tm: %d sz1: %d -> %d eof",
+		 Tr, Tmw, Ts, Tm, sz1, * sz));
   
   return false;
 }
@@ -316,13 +316,13 @@ FsBuffer::getWritePtr(Uint32 ** ptr, Uint32 sz){
   if(sz1 > sz){ // Note at least 1 word of slack
     * ptr = &Tp[Tw];
 
-    DEBUG(ndbout_c("getWritePtr(%d) Tr: %d Tw: %d Ts: %d sz1: %d -> true",
-		   sz, Tr, Tw, Ts, sz1));
+    DEBUG(ndbout_c("getWritePtr(%d) Tw: %d sz1: %d -> true",
+		   sz, Tw, sz1));
     return true;
   }
 
-  DEBUG(ndbout_c("getWritePtr(%d) Tr: %d Tw: %d Ts: %d sz1: %d -> false",
-		 sz, Tr, Tw, Ts, sz1));
+  DEBUG(ndbout_c("getWritePtr(%d) Tw: %d sz1: %d -> false",
+		 sz, Tw, sz1));
 
   return false;
 }
@@ -339,11 +339,15 @@ FsBuffer::updateWritePtr(Uint32 sz){
   m_free -= sz;
   if(Tnew < Ts){
     m_writeIndex = Tnew;
+    DEBUG(ndbout_c("updateWritePtr(%d) m_writeIndex: %d",
+                   sz, m_writeIndex));
     return;
   }
 
   memcpy(Tp, &Tp[Ts], (Tnew - Ts) << 2);
   m_writeIndex = Tnew - Ts;
+  DEBUG(ndbout_c("updateWritePtr(%d) m_writeIndex: %d",
+                 sz, m_writeIndex));
 }
 
 inline

storage/ndb/src/kernel/blocks/dblqh/Dblqh.hpp

@@ -115,9 +115,6 @@ class Dbtup;
 /* ------------------------------------------------------------------------- */
 /*       VARIOUS CONSTANTS USED AS FLAGS TO THE FILE MANAGER.                */
 /* ------------------------------------------------------------------------- */
-#define ZOPEN_READ 0
-#define ZOPEN_WRITE 1
-#define ZOPEN_READ_WRITE 2
 #define ZVAR_NO_LOG_PAGE_WORD 1
 #define ZLIST_OF_PAIRS 0
 #define ZLIST_OF_PAIRS_SYNCH 16
@@ -2686,6 +2683,7 @@ private:
   UintR clfoFileSize;
 
   LogPageRecord *logPageRecord;
+  void *logPageRecordUnaligned;
   LogPageRecordPtr logPagePtr;
   UintR cfirstfreeLogPage;
   UintR clogPageFileSize;
@@ -2889,6 +2887,7 @@ private:
   UintR ctransidHash[1024];
   
   Uint32 c_diskless;
+  Uint32 c_o_direct;
   Uint32 c_error_insert_table_id;
   
 public:

storage/ndb/src/kernel/blocks/dblqh/DblqhInit.cpp

@@ -49,6 +49,7 @@ void Dblqh::initData()
   logFileRecord = 0;
   logFileOperationRecord = 0;
   logPageRecord = 0;
+  logPageRecordUnaligned= 0;
   pageRefRecord = 0;
   tablerec = 0;
   tcConnectionrec = 0;
@@ -105,9 +106,12 @@ void Dblqh::initRecords()
 		sizeof(LogFileOperationRecord), 
 		clfoFileSize);
 
-  logPageRecord = (LogPageRecord*)allocRecord("LogPageRecord",
+  logPageRecord =
+    (LogPageRecord*)allocRecordAligned("LogPageRecord",
                                        sizeof(LogPageRecord),
                                        clogPageFileSize,
+                                       &logPageRecordUnaligned,
+                                       NDB_O_DIRECT_WRITE_ALIGNMENT,
                                        false);
 
   pageRefRecord = (PageRefRecord*)allocRecord("PageRefRecord",
@@ -378,7 +382,7 @@ Dblqh::~Dblqh()
 		sizeof(LogFileOperationRecord), 
 		clfoFileSize);
   
-  deallocRecord((void**)&logPageRecord,
+  deallocRecord((void**)&logPageRecordUnaligned,
 		"LogPageRecord",
 		sizeof(LogPageRecord),
 		clogPageFileSize);

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

@@ -1015,6 +1015,8 @@ void Dblqh::execREAD_CONFIG_REQ(Signal* signal)
   cmaxAccOps = cscanrecFileSize * MAX_PARALLEL_OP_PER_SCAN;
 
   ndbrequire(!ndb_mgm_get_int_parameter(p, CFG_DB_DISCLESS, &c_diskless));
+  c_o_direct = true;
+  ndb_mgm_get_int_parameter(p, CFG_DB_O_DIRECT, &c_o_direct);
   
   Uint32 tmp= 0;
   ndbrequire(!ndb_mgm_get_int_parameter(p, CFG_LQH_FRAG, &tmp));
@@ -13243,7 +13245,9 @@ void Dblqh::openFileRw(Signal* signal, LogFileRecordPtr olfLogFilePtr)
   signal->theData[3] = olfLogFilePtr.p->fileName[1];
   signal->theData[4] = olfLogFilePtr.p->fileName[2];
   signal->theData[5] = olfLogFilePtr.p->fileName[3];
-  signal->theData[6] = ZOPEN_READ_WRITE | FsOpenReq::OM_AUTOSYNC;
+  signal->theData[6] = FsOpenReq::OM_READWRITE | FsOpenReq::OM_AUTOSYNC;
+  if (c_o_direct)
+    signal->theData[6] |= FsOpenReq::OM_DIRECT;
   req->auto_sync_size = MAX_REDO_PAGES_WITHOUT_SYNCH * sizeof(LogPageRecord);
   sendSignal(NDBFS_REF, GSN_FSOPENREQ, signal, FsOpenReq::SignalLength, JBA);
 }//Dblqh::openFileRw()
@@ -13263,7 +13267,9 @@ void Dblqh::openLogfileInit(Signal* signal)
   signal->theData[3] = logFilePtr.p->fileName[1];
   signal->theData[4] = logFilePtr.p->fileName[2];
   signal->theData[5] = logFilePtr.p->fileName[3];
-  signal->theData[6] = 0x302 | FsOpenReq::OM_AUTOSYNC;
+  signal->theData[6] = FsOpenReq::OM_READWRITE | FsOpenReq::OM_TRUNCATE | FsOpenReq::OM_CREATE | FsOpenReq::OM_AUTOSYNC;
+  if (c_o_direct)
+    signal->theData[6] |= FsOpenReq::OM_DIRECT;
   req->auto_sync_size = MAX_REDO_PAGES_WITHOUT_SYNCH * sizeof(LogPageRecord);
   sendSignal(NDBFS_REF, GSN_FSOPENREQ, signal, FsOpenReq::SignalLength, JBA);
 }//Dblqh::openLogfileInit()
@@ -13299,7 +13305,9 @@ void Dblqh::openNextLogfile(Signal* signal)
     signal->theData[3] = onlLogFilePtr.p->fileName[1];
     signal->theData[4] = onlLogFilePtr.p->fileName[2];
     signal->theData[5] = onlLogFilePtr.p->fileName[3];
-    signal->theData[6] = 2 | FsOpenReq::OM_AUTOSYNC;
+    signal->theData[6] = FsOpenReq::OM_READWRITE | FsOpenReq::OM_AUTOSYNC;
+    if (c_o_direct)
+      signal->theData[6] |= FsOpenReq::OM_DIRECT;
     req->auto_sync_size = MAX_REDO_PAGES_WITHOUT_SYNCH * sizeof(LogPageRecord);
     sendSignal(NDBFS_REF, GSN_FSOPENREQ, signal, FsOpenReq::SignalLength, JBA);
   }//if

storage/ndb/src/kernel/blocks/ndbfs/AsyncFile.cpp

@@ -163,7 +163,12 @@ AsyncFile::run()
   theStartFlag = true;
   // Create write buffer for bigger writes
   theWriteBufferSize = WRITEBUFFERSIZE;
-  theWriteBuffer = (char *) ndbd_malloc(theWriteBufferSize); 
+  theWriteBufferUnaligned = (char *) ndbd_malloc(theWriteBufferSize +
+                                                 NDB_O_DIRECT_WRITE_ALIGNMENT-1);
+  theWriteBuffer = (char *)
+    (((UintPtr)theWriteBufferUnaligned + NDB_O_DIRECT_WRITE_ALIGNMENT - 1) &
+     ~(UintPtr)(NDB_O_DIRECT_WRITE_ALIGNMENT - 1));
+
   NdbMutex_Unlock(theStartMutexPtr);
   NdbCondition_Signal(theStartConditionPtr);
   
@@ -247,6 +252,78 @@ AsyncFile::run()
 static char g_odirect_readbuf[2*GLOBAL_PAGE_SIZE -1];
 #endif
 
+int
+AsyncFile::check_odirect_write(Uint32 flags, int& new_flags, int mode)
+{
+  assert(new_flags & (O_CREAT | O_TRUNC));
+#ifdef O_DIRECT
+  int ret;
+  char * bufptr = (char*)((UintPtr(g_odirect_readbuf)+(GLOBAL_PAGE_SIZE - 1)) & ~(GLOBAL_PAGE_SIZE - 1));
+  while (((ret = ::write(theFd, bufptr, GLOBAL_PAGE_SIZE)) == -1) && 
+         (errno == EINTR));
+  if (ret == -1)
+  {
+    new_flags &= ~O_DIRECT;
+    ndbout_c("%s Failed to write using O_DIRECT, disabling", 
+             theFileName.c_str());
+  }
+  
+  close(theFd);
+  theFd = ::open(theFileName.c_str(), new_flags, mode);
+  if (theFd == -1)
+    return errno;
+#endif
+
+  return 0;
+}
+
+int
+AsyncFile::check_odirect_read(Uint32 flags, int &new_flags, int mode)
+{
+#ifdef O_DIRECT
+  int ret;
+  char * bufptr = (char*)((UintPtr(g_odirect_readbuf)+(GLOBAL_PAGE_SIZE - 1)) & ~(GLOBAL_PAGE_SIZE - 1));
+  while (((ret = ::read(theFd, bufptr, GLOBAL_PAGE_SIZE)) == -1) && 
+         (errno == EINTR));
+  if (ret == -1)
+  {
+    ndbout_c("%s Failed to read using O_DIRECT, disabling", 
+             theFileName.c_str());
+    goto reopen;
+  }
+  
+  if(lseek(theFd, 0, SEEK_SET) != 0)
+  {
+    return errno;
+  }
+  
+  if ((flags & FsOpenReq::OM_CHECK_SIZE) == 0)
+  {
+    struct stat buf;
+    if ((fstat(theFd, &buf) == -1))
+    {
+      return errno;
+    } 
+    else if ((buf.st_size % GLOBAL_PAGE_SIZE) != 0)
+    {
+      ndbout_c("%s filesize not a multiple of %d, disabling O_DIRECT", 
+               theFileName.c_str(), GLOBAL_PAGE_SIZE);
+      goto reopen;
+    }
+  }
+  
+  return 0;
+  
+reopen:
+  close(theFd);
+  new_flags &= ~O_DIRECT;
+  theFd = ::open(theFileName.c_str(), new_flags, mode);
+  if (theFd == -1)
+    return errno;  
+#endif
+  return 0;
+}
+
 void AsyncFile::openReq(Request* request)
 {  
   m_auto_sync_freq = 0;
@@ -312,7 +389,7 @@ void AsyncFile::openReq(Request* request)
   }
 #else
   Uint32 flags = request->par.open.flags;
-  Uint32 new_flags = 0;
+  int new_flags = 0;
 
   // Convert file open flags from Solaris to Liux
   if (flags & FsOpenReq::OM_CREATE)
@@ -343,10 +420,6 @@ void AsyncFile::openReq(Request* request)
   {
     new_flags |= O_DIRECT;
   }
-#elif defined O_SYNC
-  {
-    flags |= FsOpenReq::OM_SYNC;
-  }
 #endif
   
   if ((flags & FsOpenReq::OM_SYNC) && ! (flags & FsOpenReq::OM_INIT))
@@ -356,14 +429,18 @@ void AsyncFile::openReq(Request* request)
 #endif
   }
 
+  const char * rw = "";
   switch(flags & 0x3){
   case FsOpenReq::OM_READONLY:
+    rw = "r";
     new_flags |= O_RDONLY;
     break;
   case FsOpenReq::OM_WRITEONLY:
+    rw = "w";
     new_flags |= O_WRONLY;
     break;
   case FsOpenReq::OM_READWRITE:
+    rw = "rw";
     new_flags |= O_RDWR;
     break;
   default:
@@ -404,11 +481,6 @@ void AsyncFile::openReq(Request* request)
 	if (new_flags & O_DIRECT)
 	{
 	  new_flags &= ~O_DIRECT;
-	  flags |= FsOpenReq::OM_SYNC;
-#ifdef O_SYNC
-	  if (! (flags & FsOpenReq::OM_INIT))
-	    new_flags |= O_SYNC;
-#endif
 	  goto no_odirect;
 	}
 #endif
@@ -421,11 +493,6 @@ void AsyncFile::openReq(Request* request)
     else if (new_flags & O_DIRECT)
     {
       new_flags &= ~O_DIRECT;
-      flags |= FsOpenReq::OM_SYNC;
-#ifdef O_SYNC
-      if (! (flags & FsOpenReq::OM_INIT))
-	new_flags |= O_SYNC;
-#endif
       goto no_odirect;
     }
 #endif
@@ -512,7 +579,6 @@ void AsyncFile::openReq(Request* request)
 	{
 	  ndbout_c("error on first write(%d), disable O_DIRECT", err);
 	  new_flags &= ~O_DIRECT;
-	  flags |= FsOpenReq::OM_SYNC;
 	  close(theFd);
 	  theFd = ::open(theFileName.c_str(), new_flags, mode);
 	  if (theFd != -1)
@@ -532,26 +598,32 @@ void AsyncFile::openReq(Request* request)
   else if (flags & FsOpenReq::OM_DIRECT)
   {
 #ifdef O_DIRECT
-    do {
-      int ret;
-      char * bufptr = (char*)((UintPtr(g_odirect_readbuf)+(GLOBAL_PAGE_SIZE - 1)) & ~(GLOBAL_PAGE_SIZE - 1));
-      while (((ret = ::read(theFd, bufptr, GLOBAL_PAGE_SIZE)) == -1) && (errno == EINTR));
-      if (ret == -1)
+    if (flags & (FsOpenReq::OM_TRUNCATE | FsOpenReq::OM_CREATE))
     {
-	ndbout_c("%s Failed to read using O_DIRECT, disabling", theFileName.c_str());
-	flags |= FsOpenReq::OM_SYNC;
-	flags |= FsOpenReq::OM_INIT;
-	break;
+      request->error = check_odirect_write(flags, new_flags, mode);
     }
-      if(lseek(theFd, 0, SEEK_SET) != 0)
+    else
     {
-	request->error = errno;
+      request->error = check_odirect_read(flags, new_flags, mode);
+    }
+    
+    if (request->error)
       return;
+#endif
   }
-    } while (0);
+#ifdef VM_TRACE
+  if (flags & FsOpenReq::OM_DIRECT)
+  {
+#ifdef O_DIRECT
+    ndbout_c("%s %s O_DIRECT: %d",
+             theFileName.c_str(), rw,
+             !!(new_flags & O_DIRECT));
+#else
+    ndbout_c("%s %s O_DIRECT: 0",
+             theFileName.c_str(), rw);
 #endif
   }
-
+#endif  
   if ((flags & FsOpenReq::OM_SYNC) && (flags & FsOpenReq::OM_INIT))
   {
 #ifdef O_SYNC
@@ -562,6 +634,10 @@ void AsyncFile::openReq(Request* request)
     new_flags &= ~(O_CREAT | O_TRUNC);
     new_flags |= O_SYNC;
     theFd = ::open(theFileName.c_str(), new_flags, mode);
+    if (theFd == -1)
+    {
+      request->error = errno;
+    }
 #endif
   }
 #endif
@@ -1079,7 +1155,8 @@ AsyncFile::rmrfReq(Request * request, char * path, bool removePath){
 void AsyncFile::endReq()
 {
   // Thread is ended with return
-  if (theWriteBuffer) ndbd_free(theWriteBuffer, theWriteBufferSize);
+  if (theWriteBufferUnaligned)
+    ndbd_free(theWriteBufferUnaligned, theWriteBufferSize);
 }
 
 

storage/ndb/src/kernel/blocks/ndbfs/AsyncFile.hpp

@@ -232,9 +232,13 @@ private:
   bool   theStartFlag;
   int theWriteBufferSize;
   char* theWriteBuffer;
+  void* theWriteBufferUnaligned;
   
   size_t m_write_wo_sync;  // Writes wo/ sync
   size_t m_auto_sync_freq; // Auto sync freq in bytes
+
+  int check_odirect_read(Uint32 flags, int&new_flags, int mode);
+  int check_odirect_write(Uint32 flags, int&new_flags, int mode);
 public:
   SimulatedBlock& m_fs;
   Ptr<GlobalPage> m_page_ptr;

storage/ndb/src/kernel/blocks/restore.cpp

@@ -559,6 +559,9 @@ Restore::restore_next(Signal* signal, FilePtr file_ptr)
       case BackupFormat::GCP_ENTRY:
 	parse_gcp_entry(signal, file_ptr, data, len);
 	break;
+      case BackupFormat::EMPTY_ENTRY:
+        // skip
+        break;
       case 0x4e444242: // 'NDBB'
 	if (check_file_version(signal, ntohl(* (data+2))) == 0)
 	{

storage/ndb/src/kernel/vm/SimulatedBlock.cpp

@@ -39,6 +39,9 @@
 #include <AttributeDescriptor.hpp>
 #include <NdbSqlUtil.hpp>
 
+#include <EventLogger.hpp>
+extern EventLogger g_eventLogger;
+
 #define ljamEntry() jamEntryLine(30000 + __LINE__)
 #define ljam() jamLine(30000 + __LINE__)
 
@@ -656,13 +659,19 @@ SimulatedBlock::getBatSize(Uint16 blockNo){
   return sb->theBATSize;
 }
 
+void* SimulatedBlock::allocRecord(const char * type, size_t s, size_t n, bool clear, Uint32 paramId)
+{
+  return allocRecordAligned(type, s, n, 0, 0, clear, paramId);
+}
+
 void* 
-SimulatedBlock::allocRecord(const char * type, size_t s, size_t n, bool clear, Uint32 paramId) 
+SimulatedBlock::allocRecordAligned(const char * type, size_t s, size_t n, void **unaligned_buffer, Uint32 align, bool clear, Uint32 paramId)
 {
 
   void * p = NULL;
-  size_t size = n*s;
-  Uint64 real_size = (Uint64)((Uint64)n)*((Uint64)s);
+  Uint32 over_alloc = unaligned_buffer ? (align - 1) : 0;
+  size_t size = n*s + over_alloc;
+  Uint64 real_size = (Uint64)((Uint64)n)*((Uint64)s) + over_alloc;
   refresh_watch_dog(9);
   if (real_size > 0){
 #ifdef VM_TRACE_MEM
@@ -705,6 +714,16 @@ SimulatedBlock::allocRecord(const char * type, size_t s, size_t n, bool clear, U
       refresh_watch_dog(9);
       memset(ptr, 0, size);
     }
+    if (unaligned_buffer)
+    {
+      *unaligned_buffer = p;
+      p = (void *)(((UintPtr)p + over_alloc) & ~(UintPtr)(over_alloc));
+#ifdef VM_TRACE
+      g_eventLogger.info("'%s' (%u) %llu %llu, alignment correction %u bytes",
+                         type, align, (Uint64)p, (Uint64)p+n*s,
+                         (Uint32)((UintPtr)p - (UintPtr)*unaligned_buffer));
+#endif
+    }
   }
   return p;
 }

storage/ndb/src/kernel/vm/SimulatedBlock.hpp

@@ -378,6 +378,7 @@ protected:
    *
    */
   void* allocRecord(const char * type, size_t s, size_t n, bool clear = true, Uint32 paramId = 0);
+  void* allocRecordAligned(const char * type, size_t s, size_t n, void **unaligned_buffer, Uint32 align = NDB_O_DIRECT_WRITE_ALIGNMENT, bool clear = true, Uint32 paramId = 0);
   
   /**
    * Deallocate record

storage/ndb/src/mgmsrv/ConfigInfo.cpp

@@ -1313,6 +1313,18 @@ const ConfigInfo::ParamInfo ConfigInfo::m_ParamInfo[] = {
     "0",
     STR_VALUE(MAX_INT_RNIL) },
   
+  {
+    CFG_DB_O_DIRECT,
+    "ODirect",
+    DB_TOKEN,
+    "Use O_DIRECT file write/read when possible",
+    ConfigInfo::CI_USED,
+    true,
+    ConfigInfo::CI_BOOL,
+    "false",
+    "false",
+    "true"},
+
   /***************************************************************************
    * API
    ***************************************************************************/

storage/ndb/tools/restore/Restore.cpp

@@ -867,13 +867,32 @@ bool RestoreDataIterator::readFragmentHeader(int & ret, Uint32 *fragmentId)
   
   debug << "RestoreDataIterator::getNextFragment" << endl;
   
-  if (buffer_read(&Header, sizeof(Header), 1) != 1){
+  while (1)
+  {
+    /* read first part of header */
+    if (buffer_read(&Header, 8, 1) != 1)
+    {
       ret = 0;
       return false;
     } // if
 
+    /* skip if EMPTY_ENTRY */
     Header.SectionType  = ntohl(Header.SectionType);
     Header.SectionLength  = ntohl(Header.SectionLength);
+    if (Header.SectionType == BackupFormat::EMPTY_ENTRY)
+    {
+      void *tmp;
+      buffer_get_ptr(&tmp, Header.SectionLength*4-8, 1);
+      continue;
+    }
+    break;
+  }
+  /* read rest of header */
+  if (buffer_read(((char*)&Header)+8, sizeof(Header)-8, 1) != 1)
+  {
+    ret = 0;
+    return false;
+  }
   Header.TableId  = ntohl(Header.TableId);
   Header.FragmentNo  = ntohl(Header.FragmentNo);
   Header.ChecksumType  = ntohl(Header.ChecksumType);