diff --git a/include/client/taos.h b/include/client/taos.h
index a102629efa..da24136d80 100644
--- a/include/client/taos.h
+++ b/include/client/taos.h
@@ -85,11 +85,7 @@ typedef struct taosField {
   int32_t bytes;
 } TAOS_FIELD;
 
-#ifdef _TD_GO_DLL_
-#define DLL_EXPORT __declspec(dllexport)
-#else
 #define DLL_EXPORT
-#endif
 
 typedef void (*__taos_async_fn_t)(void *param, TAOS_RES *, int code);
 
diff --git a/include/common/tcommon.h b/include/common/tcommon.h
index 385c123fec..5308be72bb 100644
--- a/include/common/tcommon.h
+++ b/include/common/tcommon.h
@@ -54,25 +54,28 @@ typedef struct SColumnDataAgg {
 } SColumnDataAgg;
 
 typedef struct SDataBlockInfo {
-  STimeWindow    window;
-  int32_t        rows;
-  int32_t        rowSize;
-  int16_t        numOfCols;
-  int16_t        hasVarCol;
-  union {int64_t uid; int64_t blockId;};
+  STimeWindow window;
+  int32_t     rows;
+  int32_t     rowSize;
+  int16_t     numOfCols;
+  int16_t     hasVarCol;
+  union {
+    int64_t uid;
+    int64_t blockId;
+  };
 } SDataBlockInfo;
 
-//typedef struct SConstantItem {
-//  SColumnInfo info;
-//  int32_t     startRow;  // run-length-encoding to save the space for multiple rows
-//  int32_t     endRow;
-//  SVariant    value;
-//} SConstantItem;
+// typedef struct SConstantItem {
+//   SColumnInfo info;
+//   int32_t     startRow;  // run-length-encoding to save the space for multiple rows
+//   int32_t     endRow;
+//   SVariant    value;
+// } SConstantItem;
 
 // info.numOfCols = taosArrayGetSize(pDataBlock) + taosArrayGetSize(pConstantList);
 typedef struct SSDataBlock {
-  SColumnDataAgg *pBlockAgg;
-  SArray         *pDataBlock;    // SArray<SColumnInfoData>
+  SColumnDataAgg* pBlockAgg;
+  SArray*         pDataBlock;  // SArray<SColumnInfoData>
   SDataBlockInfo  info;
 } SSDataBlock;
 
@@ -98,23 +101,37 @@ void*   blockDataDestroy(SSDataBlock* pBlock);
 int32_t tEncodeDataBlock(void** buf, const SSDataBlock* pBlock);
 void*   tDecodeDataBlock(const void* buf, SSDataBlock* pBlock);
 
-static FORCE_INLINE void tDeleteSSDataBlock(SSDataBlock* pBlock) {
-  if (pBlock == NULL) {
-    return;
+static FORCE_INLINE void blockDestroyInner(SSDataBlock* pBlock) {
+  // WARNING: do not use info.numOfCols,
+  // sometimes info.numOfCols != array size
+  int32_t numOfOutput = taosArrayGetSize(pBlock->pDataBlock);
+  for (int32_t i = 0; i < numOfOutput; ++i) {
+    SColumnInfoData* pColInfoData = (SColumnInfoData*)taosArrayGet(pBlock->pDataBlock, i);
+    if (IS_VAR_DATA_TYPE(pColInfoData->info.type)) {
+      tfree(pColInfoData->varmeta.offset);
+    } else {
+      tfree(pColInfoData->nullbitmap);
+    }
+
+    tfree(pColInfoData->pData);
   }
-  blockDataDestroy(pBlock);
+
+  taosArrayDestroy(pBlock->pDataBlock);
+  tfree(pBlock->pBlockAgg);
 }
 
+static FORCE_INLINE void tDeleteSSDataBlock(SSDataBlock* pBlock) { blockDestroyInner(pBlock); }
+
 static FORCE_INLINE int32_t tEncodeSMqPollRsp(void** buf, const SMqPollRsp* pRsp) {
   int32_t tlen = 0;
   int32_t sz = 0;
-  tlen += taosEncodeFixedI64(buf, pRsp->consumerId);
+  // tlen += taosEncodeFixedI64(buf, pRsp->consumerId);
   tlen += taosEncodeFixedI64(buf, pRsp->reqOffset);
   tlen += taosEncodeFixedI64(buf, pRsp->rspOffset);
   tlen += taosEncodeFixedI32(buf, pRsp->skipLogNum);
   tlen += taosEncodeFixedI32(buf, pRsp->numOfTopics);
   if (pRsp->numOfTopics == 0) return tlen;
-  tlen += tEncodeSSchemaWrapper(buf, pRsp->schemas);
+  tlen += tEncodeSSchemaWrapper(buf, pRsp->schema);
   if (pRsp->pBlockData) {
     sz = taosArrayGetSize(pRsp->pBlockData);
   }
@@ -128,15 +145,15 @@ static FORCE_INLINE int32_t tEncodeSMqPollRsp(void** buf, const SMqPollRsp* pRsp
 
 static FORCE_INLINE void* tDecodeSMqPollRsp(void* buf, SMqPollRsp* pRsp) {
   int32_t sz;
-  buf = taosDecodeFixedI64(buf, &pRsp->consumerId);
+  // buf = taosDecodeFixedI64(buf, &pRsp->consumerId);
   buf = taosDecodeFixedI64(buf, &pRsp->reqOffset);
   buf = taosDecodeFixedI64(buf, &pRsp->rspOffset);
   buf = taosDecodeFixedI32(buf, &pRsp->skipLogNum);
   buf = taosDecodeFixedI32(buf, &pRsp->numOfTopics);
   if (pRsp->numOfTopics == 0) return buf;
-  pRsp->schemas = (SSchemaWrapper*)calloc(1, sizeof(SSchemaWrapper));
-  if (pRsp->schemas == NULL) return NULL;
-  buf = tDecodeSSchemaWrapper(buf, pRsp->schemas);
+  pRsp->schema = (SSchemaWrapper*)calloc(1, sizeof(SSchemaWrapper));
+  if (pRsp->schema == NULL) return NULL;
+  buf = tDecodeSSchemaWrapper(buf, pRsp->schema);
   buf = taosDecodeFixedI32(buf, &sz);
   pRsp->pBlockData = taosArrayInit(sz, sizeof(SSDataBlock));
   for (int32_t i = 0; i < sz; i++) {
@@ -148,13 +165,13 @@ static FORCE_INLINE void* tDecodeSMqPollRsp(void* buf, SMqPollRsp* pRsp) {
 }
 
 static FORCE_INLINE void tDeleteSMqConsumeRsp(SMqPollRsp* pRsp) {
-  if (pRsp->schemas) {
-    if (pRsp->schemas->nCols) {
-      tfree(pRsp->schemas->pSchema);
+  if (pRsp->schema) {
+    if (pRsp->schema->nCols) {
+      tfree(pRsp->schema->pSchema);
     }
-    free(pRsp->schemas);
+    free(pRsp->schema);
   }
-  taosArrayDestroyEx(pRsp->pBlockData, (void (*)(void*))tDeleteSSDataBlock);
+  taosArrayDestroyEx(pRsp->pBlockData, (void (*)(void*))blockDestroyInner);
   pRsp->pBlockData = NULL;
 }
 
@@ -196,7 +213,7 @@ typedef struct SGroupbyExpr {
 
 typedef struct SFunctParam {
   int32_t  type;
-  SColumn *pCol;
+  SColumn* pCol;
   SVariant param;
 } SFunctParam;
 
@@ -214,12 +231,12 @@ typedef struct SResSchame {
 typedef struct SExprBasicInfo {
   SResSchema   resSchema;
   int16_t      numOfParams;  // argument value of each function
-  SFunctParam *pParam;
+  SFunctParam* pParam;
 } SExprBasicInfo;
 
 typedef struct SExprInfo {
-  struct SExprBasicInfo  base;
-  struct tExprNode      *pExpr;
+  struct SExprBasicInfo base;
+  struct tExprNode*     pExpr;
 } SExprInfo;
 
 typedef struct SStateWindow {
diff --git a/include/common/tmsg.h b/include/common/tmsg.h
index 75c2be7932..ca4e6b97a4 100644
--- a/include/common/tmsg.h
+++ b/include/common/tmsg.h
@@ -1282,7 +1282,7 @@ static FORCE_INLINE SMqRebSubscribe* tNewSMqRebSubscribe(const char* key) {
   if (pRebSub == NULL) {
     goto _err;
   }
-  pRebSub->key = key;
+  pRebSub->key = strdup(key);
   pRebSub->lostConsumers = taosArrayInit(0, sizeof(int64_t));
   if (pRebSub->lostConsumers == NULL) {
     goto _err;
@@ -2116,25 +2116,16 @@ typedef struct {
   int8_t  mqMsgType;
   int32_t code;
   int32_t epoch;
+  int64_t consumerId;
 } SMqRspHead;
 
-typedef struct {
-  int64_t         consumerId;
-  SSchemaWrapper* schemas;
-  int64_t         reqOffset;
-  int64_t         rspOffset;
-  int32_t         skipLogNum;
-  int32_t         numOfTopics;
-  SArray*         pBlockData;  // SArray<SSDataBlock>
-} SMqPollRsp;
-
-// one req for one vg+topic
 typedef struct {
   SMsgHead head;
 
   int64_t consumerId;
   int64_t blockingTime;
   int32_t epoch;
+  int8_t  withSchema;
   char    cgroup[TSDB_CGROUP_LEN];
 
   int64_t currentOffset;
@@ -2153,19 +2144,22 @@ typedef struct {
 } SMqSubTopicEp;
 
 typedef struct {
-  int64_t consumerId;
-  char    cgroup[TSDB_CGROUP_LEN];
-  SArray* topics;  // SArray<SMqSubTopicEp>
-} SMqCMGetSubEpRsp;
+  SMqRspHead head;
+  int64_t    reqOffset;
+  int64_t    rspOffset;
+  int32_t    skipLogNum;
+  // TODO: replace with topic name
+  int32_t numOfTopics;
+  // TODO: remove from msg
+  SSchemaWrapper* schema;
+  SArray*         pBlockData;  // SArray<SSDataBlock>
+} SMqPollRsp;
 
 typedef struct {
   SMqRspHead head;
-  union {
-    SMqPollRsp       consumeRsp;
-    SMqCMGetSubEpRsp getEpRsp;
-  };
-  void* extra;
-} SMqMsgWrapper;
+  char       cgroup[TSDB_CGROUP_LEN];
+  SArray*    topics;  // SArray<SMqSubTopicEp>
+} SMqCMGetSubEpRsp;
 
 typedef struct {
   int32_t curBlock;
@@ -2173,11 +2167,13 @@ typedef struct {
   void**  uData;
 } SMqRowIter;
 
-struct tmq_message_t_v1 {
-  SMqPollRsp rsp;
+struct tmq_message_t {
+  SMqPollRsp msg;
+  void*      vg;
   SMqRowIter iter;
 };
 
+#if 0
 struct tmq_message_t {
   SMqRspHead head;
   union {
@@ -2189,6 +2185,7 @@ struct tmq_message_t {
   int32_t curRow;
   void**  uData;
 };
+#endif
 
 static FORCE_INLINE void tDeleteSMqSubTopicEp(SMqSubTopicEp* pSubTopicEp) { taosArrayDestroy(pSubTopicEp->vgs); }
 
@@ -2241,8 +2238,7 @@ static FORCE_INLINE void* tDecodeSMqSubTopicEp(void* buf, SMqSubTopicEp* pTopicE
 
 static FORCE_INLINE int32_t tEncodeSMqCMGetSubEpRsp(void** buf, const SMqCMGetSubEpRsp* pRsp) {
   int32_t tlen = 0;
-  tlen += taosEncodeFixedI64(buf, pRsp->consumerId);
-  tlen += taosEncodeString(buf, pRsp->cgroup);
+  // tlen += taosEncodeString(buf, pRsp->cgroup);
   int32_t sz = taosArrayGetSize(pRsp->topics);
   tlen += taosEncodeFixedI32(buf, sz);
   for (int32_t i = 0; i < sz; i++) {
@@ -2253,8 +2249,7 @@ static FORCE_INLINE int32_t tEncodeSMqCMGetSubEpRsp(void** buf, const SMqCMGetSu
 }
 
 static FORCE_INLINE void* tDecodeSMqCMGetSubEpRsp(void* buf, SMqCMGetSubEpRsp* pRsp) {
-  buf = taosDecodeFixedI64(buf, &pRsp->consumerId);
-  buf = taosDecodeStringTo(buf, pRsp->cgroup);
+  // buf = taosDecodeStringTo(buf, pRsp->cgroup);
   int32_t sz;
   buf = taosDecodeFixedI32(buf, &sz);
   pRsp->topics = taosArrayInit(sz, sizeof(SMqSubTopicEp));
@@ -2275,8 +2270,8 @@ enum {
 };
 
 typedef struct {
-  void*  inputHandle;
-  void** executor;
+  void* inputHandle;
+  void* executor[4];
 } SStreamTaskParRunner;
 
 typedef struct {
diff --git a/include/os/osAtomic.h b/include/os/osAtomic.h
index bd3d372fc0..5d45ced3ed 100644
--- a/include/os/osAtomic.h
+++ b/include/os/osAtomic.h
@@ -20,337 +20,96 @@
 extern "C" {
 #endif
 
-#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
-  #define atomic_load_8(ptr) (*(char volatile*)(ptr))
-  #define atomic_load_16(ptr) (*(short volatile*)(ptr))
-  #define atomic_load_32(ptr) (*(long volatile*)(ptr))
-  #define atomic_load_64(ptr) (*(__int64 volatile*)(ptr))
-  #define atomic_load_ptr(ptr) (*(void* volatile*)(ptr))
-
-  #define atomic_store_8(ptr, val) ((*(char volatile*)(ptr)) = (char)(val))
-  #define atomic_store_16(ptr, val) ((*(short volatile*)(ptr)) = (short)(val))
-  #define atomic_store_32(ptr, val) ((*(long volatile*)(ptr)) = (long)(val))
-  #define atomic_store_64(ptr, val) ((*(__int64 volatile*)(ptr)) = (__int64)(val))
-  #define atomic_store_ptr(ptr, val) ((*(void* volatile*)(ptr)) = (void*)(val))
-
-  #define atomic_exchange_8(ptr, val) _InterlockedExchange8((char volatile*)(ptr), (char)(val))
-  #define atomic_exchange_16(ptr, val) _InterlockedExchange16((short volatile*)(ptr), (short)(val))
-  #define atomic_exchange_32(ptr, val) _InterlockedExchange((long volatile*)(ptr), (long)(val))
-  #define atomic_exchange_64(ptr, val) _InterlockedExchange64((__int64 volatile*)(ptr), (__int64)(val))
-  #ifdef _WIN64 
-    #define atomic_exchange_ptr(ptr, val) _InterlockedExchangePointer((void* volatile*)(ptr), (void*)(val)) 
-  #else
-    #define atomic_exchange_ptr(ptr, val) _InlineInterlockedExchangePointer((void* volatile*)(ptr), (void*)(val))
-  #endif
-  
-  #ifdef _TD_GO_DLL_
-    #define atomic_val_compare_exchange_8 __sync_val_compare_and_swap
-  #else
-    #define atomic_val_compare_exchange_8(ptr, oldval, newval) _InterlockedCompareExchange8((char volatile*)(ptr), (char)(newval), (char)(oldval))
-  #endif
-  #define atomic_val_compare_exchange_16(ptr, oldval, newval) _InterlockedCompareExchange16((short volatile*)(ptr), (short)(newval), (short)(oldval))
-  #define atomic_val_compare_exchange_32(ptr, oldval, newval) _InterlockedCompareExchange((long volatile*)(ptr), (long)(newval), (long)(oldval))
-  #define atomic_val_compare_exchange_64(ptr, oldval, newval) _InterlockedCompareExchange64((__int64 volatile*)(ptr), (__int64)(newval), (__int64)(oldval))
-  #define atomic_val_compare_exchange_ptr(ptr, oldval, newval) _InterlockedCompareExchangePointer((void* volatile*)(ptr), (void*)(newval), (void*)(oldval))
-
-  char    interlocked_add_fetch_8(char volatile *ptr, char val);
-  short   interlocked_add_fetch_16(short volatile *ptr, short val);
-  long    interlocked_add_fetch_32(long volatile *ptr, long val);
-  __int64 interlocked_add_fetch_64(__int64 volatile *ptr, __int64 val);
-
-  char interlocked_and_fetch_8(char volatile* ptr, char val);
-  short interlocked_and_fetch_16(short volatile* ptr, short val);
-  long interlocked_and_fetch_32(long volatile* ptr, long val);
-  __int64 interlocked_and_fetch_64(__int64 volatile* ptr, __int64 val);
-
-  __int64 interlocked_fetch_and_64(__int64 volatile* ptr, __int64 val);
-
-  char interlocked_or_fetch_8(char volatile* ptr, char val);
-  short interlocked_or_fetch_16(short volatile* ptr, short val);
-  long interlocked_or_fetch_32(long volatile* ptr, long val);
-  __int64 interlocked_or_fetch_64(__int64 volatile* ptr, __int64 val);
-
-  char interlocked_xor_fetch_8(char volatile* ptr, char val);
-  short interlocked_xor_fetch_16(short volatile* ptr, short val);
-  long interlocked_xor_fetch_32(long volatile* ptr, long val);
-  __int64 interlocked_xor_fetch_64(__int64 volatile* ptr, __int64 val);
-
-  __int64 interlocked_fetch_xor_64(__int64 volatile* ptr, __int64 val);
-
-  #define atomic_add_fetch_8(ptr, val) interlocked_add_fetch_8((char volatile*)(ptr), (char)(val))
-  #define atomic_add_fetch_16(ptr, val) interlocked_add_fetch_16((short volatile*)(ptr), (short)(val))
-  #define atomic_add_fetch_32(ptr, val) interlocked_add_fetch_32((long volatile*)(ptr), (long)(val))
-  #define atomic_add_fetch_64(ptr, val) interlocked_add_fetch_64((__int64 volatile*)(ptr), (__int64)(val))
-  #ifdef _TD_GO_DLL_
-    #define atomic_fetch_add_8 __sync_fetch_and_ad
-    #define atomic_fetch_add_16 __sync_fetch_and_add
-  #else
-    #define atomic_fetch_add_8(ptr, val) _InterlockedExchangeAdd8((char volatile*)(ptr), (char)(val))
-    #define atomic_fetch_add_16(ptr, val) _InterlockedExchangeAdd16((short volatile*)(ptr), (short)(val))
-  #endif  
-  #define atomic_fetch_add_8(ptr, val) _InterlockedExchangeAdd8((char volatile*)(ptr), (char)(val))
-  #define atomic_fetch_add_16(ptr, val) _InterlockedExchangeAdd16((short volatile*)(ptr), (short)(val))
-  #define atomic_fetch_add_32(ptr, val) _InterlockedExchangeAdd((long volatile*)(ptr), (long)(val))
-  #define atomic_fetch_add_64(ptr, val) _InterlockedExchangeAdd64((__int64 volatile*)(ptr), (__int64)(val))
-  
-  #define atomic_sub_fetch_8(ptr, val) interlocked_add_fetch_8((char volatile*)(ptr), -(char)(val))
-  #define atomic_sub_fetch_16(ptr, val) interlocked_add_fetch_16((short volatile*)(ptr), -(short)(val))
-  #define atomic_sub_fetch_32(ptr, val) interlocked_add_fetch_32((long volatile*)(ptr), -(long)(val))
-  #define atomic_sub_fetch_64(ptr, val) interlocked_add_fetch_64((__int64 volatile*)(ptr), -(__int64)(val))
-
-  #define atomic_fetch_sub_8(ptr, val) _InterlockedExchangeAdd8((char volatile*)(ptr), -(char)(val))
-  #define atomic_fetch_sub_16(ptr, val) _InterlockedExchangeAdd16((short volatile*)(ptr), -(short)(val))
-  #define atomic_fetch_sub_32(ptr, val) _InterlockedExchangeAdd((long volatile*)(ptr), -(long)(val))
-  #define atomic_fetch_sub_64(ptr, val) _InterlockedExchangeAdd64((__int64 volatile*)(ptr), -(__int64)(val))
-
-  #define atomic_and_fetch_8(ptr, val) interlocked_and_fetch_8((char volatile*)(ptr), (char)(val))
-  #define atomic_and_fetch_16(ptr, val) interlocked_and_fetch_16((short volatile*)(ptr), (short)(val))
-  #define atomic_and_fetch_32(ptr, val) interlocked_and_fetch_32((long volatile*)(ptr), (long)(val))
-  #define atomic_and_fetch_64(ptr, val) interlocked_and_fetch_64((__int64 volatile*)(ptr), (__int64)(val))
-
-  #define atomic_fetch_and_8(ptr, val) _InterlockedAnd8((char volatile*)(ptr), (char)(val))
-  #define atomic_fetch_and_16(ptr, val) _InterlockedAnd16((short volatile*)(ptr), (short)(val))
-  #define atomic_fetch_and_32(ptr, val) _InterlockedAnd((long volatile*)(ptr), (long)(val))
-  #define atomic_fetch_and_64(ptr, val) interlocked_fetch_and_64((__int64 volatile*)(ptr), (__int64)(val))
-
-  #define atomic_or_fetch_8(ptr, val) interlocked_or_fetch_8((char volatile*)(ptr), (char)(val))
-  #define atomic_or_fetch_16(ptr, val) interlocked_or_fetch_16((short volatile*)(ptr), (short)(val))
-  #define atomic_or_fetch_32(ptr, val) interlocked_or_fetch_32((long volatile*)(ptr), (long)(val))
-  #define atomic_or_fetch_64(ptr, val) interlocked_or_fetch_64((__int64 volatile*)(ptr), (__int64)(val))
-
-  #define atomic_fetch_or_8(ptr, val) _InterlockedOr8((char volatile*)(ptr), (char)(val))
-  #define atomic_fetch_or_16(ptr, val) _InterlockedOr16((short volatile*)(ptr), (short)(val))
-  #define atomic_fetch_or_32(ptr, val) _InterlockedOr((long volatile*)(ptr), (long)(val))
-  #define atomic_fetch_or_64(ptr, val) interlocked_fetch_or_64((__int64 volatile*)(ptr), (__int64)(val))
-
-  #define atomic_xor_fetch_8(ptr, val) interlocked_xor_fetch_8((char volatile*)(ptr), (char)(val))
-  #define atomic_xor_fetch_16(ptr, val) interlocked_xor_fetch_16((short volatile*)(ptr), (short)(val))
-  #define atomic_xor_fetch_32(ptr, val) interlocked_xor_fetch_32((long volatile*)(ptr), (long)(val))
-  #define atomic_xor_fetch_64(ptr, val) interlocked_xor_fetch_64((__int64 volatile*)(ptr), (__int64)(val))
-
-  #define atomic_fetch_xor_8(ptr, val) _InterlockedXor8((char volatile*)(ptr), (char)(val))
-  #define atomic_fetch_xor_16(ptr, val) _InterlockedXor16((short volatile*)(ptr), (short)(val))
-  #define atomic_fetch_xor_32(ptr, val) _InterlockedXor((long volatile*)(ptr), (long)(val))
-  #define atomic_fetch_xor_64(ptr, val) interlocked_fetch_xor_64((__int64 volatile*)(ptr), (__int64)(val))
-
-  #ifdef _WIN64
-    #define atomic_add_fetch_ptr atomic_add_fetch_64
-    #define atomic_fetch_add_ptr atomic_fetch_add_64
-    #define atomic_sub_fetch_ptr atomic_sub_fetch_64
-    #define atomic_fetch_sub_ptr atomic_fetch_sub_64
-    #define atomic_and_fetch_ptr atomic_and_fetch_64
-    #define atomic_fetch_and_ptr atomic_fetch_and_64
-    #define atomic_or_fetch_ptr  atomic_or_fetch_64
-    #define atomic_fetch_or_ptr  atomic_fetch_or_64
-    #define atomic_xor_fetch_ptr atomic_xor_fetch_64
-    #define atomic_fetch_xor_ptr atomic_fetch_xor_64
-  #else
-    #define atomic_add_fetch_ptr atomic_add_fetch_32
-    #define atomic_fetch_add_ptr atomic_fetch_add_32
-    #define atomic_sub_fetch_ptr atomic_sub_fetch_32
-    #define atomic_fetch_sub_ptr atomic_fetch_sub_32
-    #define atomic_and_fetch_ptr atomic_and_fetch_32
-    #define atomic_fetch_and_ptr atomic_fetch_and_32
-    #define atomic_or_fetch_ptr  atomic_or_fetch_32
-    #define atomic_fetch_or_ptr  atomic_fetch_or_32
-    #define atomic_xor_fetch_ptr atomic_xor_fetch_32
-    #define atomic_fetch_xor_ptr atomic_fetch_xor_32
-  #endif
-#elif defined(_TD_NINGSI_60)
-  /*
-  * type __sync_fetch_and_add (type *ptr, type value);
-  * type __sync_fetch_and_sub (type *ptr, type value);
-  * type __sync_fetch_and_or (type *ptr, type value);
-  * type __sync_fetch_and_and (type *ptr, type value);
-  * type __sync_fetch_and_xor (type *ptr, type value);
-  * type __sync_fetch_and_nand (type *ptr, type value);
-  * type __sync_add_and_fetch (type *ptr, type value);
-  * type __sync_sub_and_fetch (type *ptr, type value);
-  * type __sync_or_and_fetch (type *ptr, type value);
-  * type __sync_and_and_fetch (type *ptr, type value);
-  * type __sync_xor_and_fetch (type *ptr, type value);
-  * type __sync_nand_and_fetch (type *ptr, type value);
-  *
-  * bool __sync_bool_compare_and_swap (type*ptr, type oldval, type newval, ...)
-  * type __sync_val_compare_and_swap (type *ptr, type oldval, ?type newval, ...)
-  * */
-
-  #define atomic_load_8(ptr)   __sync_fetch_and_add((ptr), 0)
-  #define atomic_load_16(ptr)  __sync_fetch_and_add((ptr), 0)
-  #define atomic_load_32(ptr)  __sync_fetch_and_add((ptr), 0)
-  #define atomic_load_64(ptr)  __sync_fetch_and_add((ptr), 0)
-  #define atomic_load_ptr(ptr) __sync_fetch_and_add((ptr), 0)
-    
-  #define atomic_store_8(ptr, val)   (*(ptr)=(val))
-  #define atomic_store_16(ptr, val)  (*(ptr)=(val))
-  #define atomic_store_32(ptr, val)  (*(ptr)=(val))
-  #define atomic_store_64(ptr, val)  (*(ptr)=(val))
-  #define atomic_store_ptr(ptr, val) (*(ptr)=(val))
-
-  int8_t  atomic_exchange_8_impl(int8_t* ptr, int8_t val );
-  int16_t atomic_exchange_16_impl(int16_t* ptr, int16_t val );
-  int32_t atomic_exchange_32_impl(int32_t* ptr, int32_t val );
-  int64_t atomic_exchange_64_impl(int64_t* ptr, int64_t val );
-  void*   atomic_exchange_ptr_impl( void **ptr, void *val );
-
-  #define atomic_exchange_8(ptr, val)   atomic_exchange_8_impl((int8_t*)ptr, (int8_t)val)
-  #define atomic_exchange_16(ptr, val)  atomic_exchange_16_impl((int16_t*)ptr, (int16_t)val)
-  #define atomic_exchange_32(ptr, val)  atomic_exchange_32_impl((int32_t*)ptr, (int32_t)val)
-  #define atomic_exchange_64(ptr, val)  atomic_exchange_64_impl((int64_t*)ptr, (int64_t)val)
-  #define atomic_exchange_ptr(ptr, val) atomic_exchange_ptr_impl((void **)ptr, (void*)val)
-      
-  #define atomic_val_compare_exchange_8   __sync_val_compare_and_swap
-  #define atomic_val_compare_exchange_16  __sync_val_compare_and_swap
-  #define atomic_val_compare_exchange_32  __sync_val_compare_and_swap
-  #define atomic_val_compare_exchange_64  __sync_val_compare_and_swap
-  #define atomic_val_compare_exchange_ptr __sync_val_compare_and_swap
-    
-  #define atomic_add_fetch_8(ptr, val) __sync_add_and_fetch((ptr), (val))
-  #define atomic_add_fetch_16(ptr, val) __sync_add_and_fetch((ptr), (val))
-  #define atomic_add_fetch_32(ptr, val) __sync_add_and_fetch((ptr), (val))
-  #define atomic_add_fetch_64(ptr, val) __sync_add_and_fetch((ptr), (val))
-  #define atomic_add_fetch_ptr(ptr, val) __sync_add_and_fetch((ptr), (val))
-    
-  #define atomic_fetch_add_8(ptr, val) __sync_fetch_and_add((ptr), (val))
-  #define atomic_fetch_add_16(ptr, val) __sync_fetch_and_add((ptr), (val))
-  #define atomic_fetch_add_32(ptr, val) __sync_fetch_and_add((ptr), (val))
-  #define atomic_fetch_add_64(ptr, val) __sync_fetch_and_add((ptr), (val))
-  #define atomic_fetch_add_ptr(ptr, val) __sync_fetch_and_add((ptr), (val))
-    
-  #define atomic_sub_fetch_8(ptr, val) __sync_sub_and_fetch((ptr), (val))
-  #define atomic_sub_fetch_16(ptr, val) __sync_sub_and_fetch((ptr), (val))
-  #define atomic_sub_fetch_32(ptr, val) __sync_sub_and_fetch((ptr), (val))
-  #define atomic_sub_fetch_64(ptr, val) __sync_sub_and_fetch((ptr), (val))
-  #define atomic_sub_fetch_ptr(ptr, val) __sync_sub_and_fetch((ptr), (val))
-    
-  #define atomic_fetch_sub_8(ptr, val) __sync_fetch_and_sub((ptr), (val))
-  #define atomic_fetch_sub_16(ptr, val) __sync_fetch_and_sub((ptr), (val))
-  #define atomic_fetch_sub_32(ptr, val) __sync_fetch_and_sub((ptr), (val))
-  #define atomic_fetch_sub_64(ptr, val) __sync_fetch_and_sub((ptr), (val))
-  #define atomic_fetch_sub_ptr(ptr, val) __sync_fetch_and_sub((ptr), (val))
-    
-  #define atomic_and_fetch_8(ptr, val) __sync_and_and_fetch((ptr), (val))
-  #define atomic_and_fetch_16(ptr, val) __sync_and_and_fetch((ptr), (val))
-  #define atomic_and_fetch_32(ptr, val) __sync_and_and_fetch((ptr), (val))
-  #define atomic_and_fetch_64(ptr, val) __sync_and_and_fetch((ptr), (val))
-  #define atomic_and_fetch_ptr(ptr, val) __sync_and_and_fetch((ptr), (val))
-    
-  #define atomic_fetch_and_8(ptr, val) __sync_fetch_and_and((ptr), (val))
-  #define atomic_fetch_and_16(ptr, val) __sync_fetch_and_and((ptr), (val))
-  #define atomic_fetch_and_32(ptr, val) __sync_fetch_and_and((ptr), (val))
-  #define atomic_fetch_and_64(ptr, val) __sync_fetch_and_and((ptr), (val))
-  #define atomic_fetch_and_ptr(ptr, val) __sync_fetch_and_and((ptr), (val))
-    
-  #define atomic_or_fetch_8(ptr, val) __sync_or_and_fetch((ptr), (val))
-  #define atomic_or_fetch_16(ptr, val) __sync_or_and_fetch((ptr), (val))
-  #define atomic_or_fetch_32(ptr, val) __sync_or_and_fetch((ptr), (val))
-  #define atomic_or_fetch_64(ptr, val) __sync_or_and_fetch((ptr), (val))
-  #define atomic_or_fetch_ptr(ptr, val) __sync_or_and_fetch((ptr), (val))
-    
-  #define atomic_fetch_or_8(ptr, val) __sync_fetch_and_or((ptr), (val))
-  #define atomic_fetch_or_16(ptr, val) __sync_fetch_and_or((ptr), (val))
-  #define atomic_fetch_or_32(ptr, val) __sync_fetch_and_or((ptr), (val))
-  #define atomic_fetch_or_64(ptr, val) __sync_fetch_and_or((ptr), (val))
-  #define atomic_fetch_or_ptr(ptr, val) __sync_fetch_and_or((ptr), (val))
-    
-  #define atomic_xor_fetch_8(ptr, val) __sync_xor_and_fetch((ptr), (val))
-  #define atomic_xor_fetch_16(ptr, val) __sync_xor_and_fetch((ptr), (val))
-  #define atomic_xor_fetch_32(ptr, val) __sync_xor_and_fetch((ptr), (val))
-  #define atomic_xor_fetch_64(ptr, val) __sync_xor_and_fetch((ptr), (val))
-  #define atomic_xor_fetch_ptr(ptr, val) __sync_xor_and_fetch((ptr), (val))
-    
-  #define atomic_fetch_xor_8(ptr, val) __sync_fetch_and_xor((ptr), (val))
-  #define atomic_fetch_xor_16(ptr, val) __sync_fetch_and_xor((ptr), (val))
-  #define atomic_fetch_xor_32(ptr, val) __sync_fetch_and_xor((ptr), (val))
-  #define atomic_fetch_xor_64(ptr, val) __sync_fetch_and_xor((ptr), (val))
-  #define atomic_fetch_xor_ptr(ptr, val) __sync_fetch_and_xor((ptr), (val))
-
-#else
-  #define atomic_load_8(ptr) __atomic_load_n((ptr), __ATOMIC_SEQ_CST)
-  #define atomic_load_16(ptr) __atomic_load_n((ptr), __ATOMIC_SEQ_CST)
-  #define atomic_load_32(ptr) __atomic_load_n((ptr), __ATOMIC_SEQ_CST)
-  #define atomic_load_64(ptr) __atomic_load_n((ptr), __ATOMIC_SEQ_CST)
-  #define atomic_load_ptr(ptr) __atomic_load_n((ptr), __ATOMIC_SEQ_CST)
-
-  #define atomic_store_8(ptr, val) __atomic_store_n((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_store_16(ptr, val) __atomic_store_n((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_store_32(ptr, val) __atomic_store_n((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_store_64(ptr, val) __atomic_store_n((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_store_ptr(ptr, val) __atomic_store_n((ptr), (val), __ATOMIC_SEQ_CST)
-
-  #define atomic_exchange_8(ptr, val) __atomic_exchange_n((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_exchange_16(ptr, val) __atomic_exchange_n((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_exchange_32(ptr, val) __atomic_exchange_n((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_exchange_64(ptr, val) __atomic_exchange_n((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_exchange_ptr(ptr, val) __atomic_exchange_n((ptr), (val), __ATOMIC_SEQ_CST)
-
-  #define atomic_val_compare_exchange_8 __sync_val_compare_and_swap
-  #define atomic_val_compare_exchange_16 __sync_val_compare_and_swap
-  #define atomic_val_compare_exchange_32 __sync_val_compare_and_swap
-  #define atomic_val_compare_exchange_64 __sync_val_compare_and_swap
-  #define atomic_val_compare_exchange_ptr __sync_val_compare_and_swap
-
-  #define atomic_add_fetch_8(ptr, val) __atomic_add_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_add_fetch_16(ptr, val) __atomic_add_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_add_fetch_32(ptr, val) __atomic_add_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_add_fetch_64(ptr, val) __atomic_add_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_add_fetch_ptr(ptr, val) __atomic_add_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-
-  #define atomic_fetch_add_8(ptr, val) __atomic_fetch_add((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_fetch_add_16(ptr, val) __atomic_fetch_add((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_fetch_add_32(ptr, val) __atomic_fetch_add((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_fetch_add_64(ptr, val) __atomic_fetch_add((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_fetch_add_ptr(ptr, val) __atomic_fetch_add((ptr), (val), __ATOMIC_SEQ_CST)
-
-  #define atomic_sub_fetch_8(ptr, val) __atomic_sub_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_sub_fetch_16(ptr, val) __atomic_sub_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_sub_fetch_32(ptr, val) __atomic_sub_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_sub_fetch_64(ptr, val) __atomic_sub_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_sub_fetch_ptr(ptr, val) __atomic_sub_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-
-  #define atomic_fetch_sub_8(ptr, val) __atomic_fetch_sub((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_fetch_sub_16(ptr, val) __atomic_fetch_sub((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_fetch_sub_32(ptr, val) __atomic_fetch_sub((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_fetch_sub_64(ptr, val) __atomic_fetch_sub((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_fetch_sub_ptr(ptr, val) __atomic_fetch_sub((ptr), (val), __ATOMIC_SEQ_CST)
-
-  #define atomic_and_fetch_8(ptr, val) __atomic_and_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_and_fetch_16(ptr, val) __atomic_and_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_and_fetch_32(ptr, val) __atomic_and_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_and_fetch_64(ptr, val) __atomic_and_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_and_fetch_ptr(ptr, val) __atomic_and_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-
-  #define atomic_fetch_and_8(ptr, val) __atomic_fetch_and((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_fetch_and_16(ptr, val) __atomic_fetch_and((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_fetch_and_32(ptr, val) __atomic_fetch_and((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_fetch_and_64(ptr, val) __atomic_fetch_and((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_fetch_and_ptr(ptr, val) __atomic_fetch_and((ptr), (val), __ATOMIC_SEQ_CST)
-
-  #define atomic_or_fetch_8(ptr, val) __atomic_or_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_or_fetch_16(ptr, val) __atomic_or_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_or_fetch_32(ptr, val) __atomic_or_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_or_fetch_64(ptr, val) __atomic_or_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_or_fetch_ptr(ptr, val) __atomic_or_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-
-  #define atomic_fetch_or_8(ptr, val) __atomic_fetch_or((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_fetch_or_16(ptr, val) __atomic_fetch_or((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_fetch_or_32(ptr, val) __atomic_fetch_or((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_fetch_or_64(ptr, val) __atomic_fetch_or((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_fetch_or_ptr(ptr, val) __atomic_fetch_or((ptr), (val), __ATOMIC_SEQ_CST)
-
-  #define atomic_xor_fetch_8(ptr, val) __atomic_xor_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_xor_fetch_16(ptr, val) __atomic_xor_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_xor_fetch_32(ptr, val) __atomic_xor_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_xor_fetch_64(ptr, val) __atomic_xor_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_xor_fetch_ptr(ptr, val) __atomic_xor_fetch((ptr), (val), __ATOMIC_SEQ_CST)
-
-  #define atomic_fetch_xor_8(ptr, val) __atomic_fetch_xor((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_fetch_xor_16(ptr, val) __atomic_fetch_xor((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_fetch_xor_32(ptr, val) __atomic_fetch_xor((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_fetch_xor_64(ptr, val) __atomic_fetch_xor((ptr), (val), __ATOMIC_SEQ_CST)
-  #define atomic_fetch_xor_ptr(ptr, val) __atomic_fetch_xor((ptr), (val), __ATOMIC_SEQ_CST)
+// If the error is in a third-party library, place this header file under the third-party library header file.
+// When you want to use this feature, you should find or add the same function in the following section.
+#ifndef ALLOW_FORBID_FUNC
+  #define __atomic_load_n __ATOMIC_LOAD_N_FUNC_TAOS_FORBID
+  #define __atomic_store_n __ATOMIC_STORE_N_FUNC_TAOS_FORBID
+  #define __atomic_exchange_n __ATOMIC_EXCHANGE_N_FUNC_TAOS_FORBID
+  #define __sync_val_compare_and_swap __SYNC_VAL_COMPARE_AND_SWAP_FUNC_TAOS_FORBID
+  #define __atomic_add_fetch __ATOMIC_ADD_FETCH_FUNC_TAOS_FORBID
+  #define __atomic_fetch_add __ATOMIC_FETCH_ADD_FUNC_TAOS_FORBID
+  #define __atomic_sub_fetch __ATOMIC_SUB_FETCH_FUNC_TAOS_FORBID
+  #define __atomic_fetch_sub __ATOMIC_FETCH_SUB_FUNC_TAOS_FORBID
+  #define __atomic_and_fetch __ATOMIC_AND_FETCH_FUNC_TAOS_FORBID
+  #define __atomic_fetch_and __ATOMIC_FETCH_AND_FUNC_TAOS_FORBID
+  #define __atomic_or_fetch __ATOMIC_OR_FETCH_FUNC_TAOS_FORBID
+  #define __atomic_fetch_or __ATOMIC_FETCH_OR_FUNC_TAOS_FORBID
+  #define __atomic_xor_fetch __ATOMIC_XOR_FETCH_FUNC_TAOS_FORBID
+  #define __atomic_fetch_xor __ATOMIC_FETCH_XOR_FUNC_TAOS_FORBID
 #endif
 
+int8_t atomic_load_8(int8_t volatile *ptr);
+int16_t atomic_load_16(int16_t volatile *ptr);
+int32_t atomic_load_32(int32_t volatile *ptr);
+int64_t atomic_load_64(int64_t volatile *ptr);
+void* atomic_load_ptr(void *ptr);
+void atomic_store_8(int8_t volatile *ptr, int8_t val);
+void atomic_store_16(int16_t volatile *ptr, int16_t val);
+void atomic_store_32(int32_t volatile *ptr, int32_t val);
+void atomic_store_64(int64_t volatile *ptr, int64_t val);
+void atomic_store_ptr(void *ptr, void *val);
+int8_t atomic_exchange_8(int8_t volatile *ptr, int8_t val);
+int16_t atomic_exchange_16(int16_t volatile *ptr, int16_t val);
+int32_t atomic_exchange_32(int32_t volatile *ptr, int32_t val);
+int64_t atomic_exchange_64(int64_t volatile *ptr, int64_t val);
+void* atomic_exchange_ptr(void *ptr, void *val);
+int8_t atomic_val_compare_exchange_8(int8_t volatile *ptr, int8_t oldval, int8_t newval);
+int16_t atomic_val_compare_exchange_16(int16_t volatile *ptr, int16_t oldval, int16_t newval);
+int32_t atomic_val_compare_exchange_32(int32_t volatile *ptr, int32_t oldval, int32_t newval);
+int64_t atomic_val_compare_exchange_64(int64_t volatile *ptr, int64_t oldval, int64_t newval);
+void* atomic_val_compare_exchange_ptr(void *ptr, void *oldval, void *newval);
+int8_t atomic_add_fetch_8(int8_t volatile *ptr, int8_t val);
+int16_t atomic_add_fetch_16(int16_t volatile *ptr, int16_t val);
+int32_t atomic_add_fetch_32(int32_t volatile *ptr, int32_t val);
+int64_t atomic_add_fetch_64(int64_t volatile *ptr, int64_t val);
+void* atomic_add_fetch_ptr(void *ptr, int32_t val);
+int8_t atomic_fetch_add_8(int8_t volatile *ptr, int8_t val);
+int16_t atomic_fetch_add_16(int16_t volatile *ptr, int16_t val);
+int32_t atomic_fetch_add_32(int32_t volatile *ptr, int32_t val);
+int64_t atomic_fetch_add_64(int64_t volatile *ptr, int64_t val);
+void* atomic_fetch_add_ptr(void *ptr, int32_t val);
+int8_t atomic_sub_fetch_8(int8_t volatile *ptr, int8_t val);
+int16_t atomic_sub_fetch_16(int16_t volatile *ptr, int16_t val);
+int32_t atomic_sub_fetch_32(int32_t volatile *ptr, int32_t val);
+int64_t atomic_sub_fetch_64(int64_t volatile *ptr, int64_t val);
+void* atomic_sub_fetch_ptr(void *ptr, int32_t val);
+int8_t atomic_fetch_sub_8(int8_t volatile *ptr, int8_t val);
+int16_t atomic_fetch_sub_16(int16_t volatile *ptr, int16_t val);
+int32_t atomic_fetch_sub_32(int32_t volatile *ptr, int32_t val);
+int64_t atomic_fetch_sub_64(int64_t volatile *ptr, int64_t val);
+void* atomic_fetch_sub_ptr(void *ptr, int32_t val);
+int8_t atomic_and_fetch_8(int8_t volatile *ptr, int8_t val);
+int16_t atomic_and_fetch_16(int16_t volatile *ptr, int16_t val);
+int32_t atomic_and_fetch_32(int32_t volatile *ptr, int32_t val);
+int64_t atomic_and_fetch_64(int64_t volatile *ptr, int64_t val);
+void* atomic_and_fetch_ptr(void *ptr, void *val);
+int8_t atomic_fetch_and_8(int8_t volatile *ptr, int8_t val);
+int16_t atomic_fetch_and_16(int16_t volatile *ptr, int16_t val);
+int32_t atomic_fetch_and_32(int32_t volatile *ptr, int32_t val);
+int64_t atomic_fetch_and_64(int64_t volatile *ptr, int64_t val);
+void* atomic_fetch_and_ptr(void *ptr, void *val);
+int8_t atomic_or_fetch_8(int8_t volatile *ptr, int8_t val);
+int16_t atomic_or_fetch_16(int16_t volatile *ptr, int16_t val);
+int32_t atomic_or_fetch_32(int32_t volatile *ptr, int32_t val);
+int64_t atomic_or_fetch_64(int64_t volatile *ptr, int64_t val);
+void* atomic_or_fetch_ptr(void *ptr, void *val);
+int8_t atomic_fetch_or_8(int8_t volatile *ptr, int8_t val);
+int16_t atomic_fetch_or_16(int16_t volatile *ptr, int16_t val);
+int32_t atomic_fetch_or_32(int32_t volatile *ptr, int32_t val);
+int64_t atomic_fetch_or_64(int64_t volatile *ptr, int64_t val);
+void* atomic_fetch_or_ptr(void *ptr, void *val);
+int8_t atomic_xor_fetch_8(int8_t volatile *ptr, int8_t val);
+int16_t atomic_xor_fetch_16(int16_t volatile *ptr, int16_t val);
+int32_t atomic_xor_fetch_32(int32_t volatile *ptr, int32_t val);
+int64_t atomic_xor_fetch_64(int64_t volatile *ptr, int64_t val);
+void* atomic_xor_fetch_ptr(void *ptr, void *val);
+int8_t atomic_fetch_xor_8(int8_t volatile *ptr, int8_t val);
+int16_t atomic_fetch_xor_16(int16_t volatile *ptr, int16_t val);
+int32_t atomic_fetch_xor_32(int32_t volatile *ptr, int32_t val);
+int64_t atomic_fetch_xor_64(int64_t volatile *ptr, int64_t val);
+void* atomic_fetch_xor_ptr(void *ptr, void *val);
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/include/os/osDir.h b/include/os/osDir.h
index 6cf28fb878..d3597cab36 100644
--- a/include/os/osDir.h
+++ b/include/os/osDir.h
@@ -17,6 +17,7 @@
 #define _TD_OS_DIR_H_
 
 // If the error is in a third-party library, place this header file under the third-party library header file.
+// When you want to use this feature, you should find or add the same function in the following section.
 #ifndef ALLOW_FORBID_FUNC
     #define opendir OPENDIR_FUNC_TAOS_FORBID
     #define readdir READDIR_FUNC_TAOS_FORBID
diff --git a/include/os/osFile.h b/include/os/osFile.h
index 508a522679..143b4bf2f8 100644
--- a/include/os/osFile.h
+++ b/include/os/osFile.h
@@ -23,6 +23,7 @@ extern "C" {
 #include "osSocket.h"
 
 // If the error is in a third-party library, place this header file under the third-party library header file.
+// When you want to use this feature, you should find or add the same function in the following sectio
 #ifndef ALLOW_FORBID_FUNC
     #define open OPEN_FUNC_TAOS_FORBID
     #define fopen FOPEN_FUNC_TAOS_FORBID
diff --git a/include/os/osLocale.h b/include/os/osLocale.h
index ddafd2e93c..74922cc0b9 100644
--- a/include/os/osLocale.h
+++ b/include/os/osLocale.h
@@ -23,6 +23,7 @@ extern "C" {
 #endif
 
 // If the error is in a third-party library, place this header file under the third-party library header file.
+// When you want to use this feature, you should find or add the same function in the following section.
 #ifndef ALLOW_FORBID_FUNC
     #define setlocale SETLOCALE_FUNC_TAOS_FORBID
 #endif
diff --git a/include/os/osRand.h b/include/os/osRand.h
index 09e1f1b41d..bce2c08a2e 100644
--- a/include/os/osRand.h
+++ b/include/os/osRand.h
@@ -21,6 +21,7 @@ extern "C" {
 #endif
 
 // If the error is in a third-party library, place this header file under the third-party library header file.
+// When you want to use this feature, you should find or add the same function in the following section.
 #ifndef ALLOW_FORBID_FUNC
     #define rand RAND_FUNC_TAOS_FORBID
     #define srand SRAND_FUNC_TAOS_FORBID
diff --git a/include/os/osSleep.h b/include/os/osSleep.h
index feb9472995..1d0ccf32bd 100644
--- a/include/os/osSleep.h
+++ b/include/os/osSleep.h
@@ -21,6 +21,7 @@ extern "C" {
 #endif
 
 // If the error is in a third-party library, place this header file under the third-party library header file.
+// When you want to use this feature, you should find or add the same function in the following section.
 #ifndef ALLOW_FORBID_FUNC
     #define Sleep SLEEP_FUNC_TAOS_FORBID
     #define sleep SLEEP_FUNC_TAOS_FORBID
diff --git a/include/os/osSocket.h b/include/os/osSocket.h
index fb330a9e2f..3c8cc8abd1 100644
--- a/include/os/osSocket.h
+++ b/include/os/osSocket.h
@@ -17,6 +17,7 @@
 #define _TD_OS_SOCKET_H_
 
 // If the error is in a third-party library, place this header file under the third-party library header file.
+// When you want to use this feature, you should find or add the same function in the following section.
 #ifndef ALLOW_FORBID_FUNC
     #define socket SOCKET_FUNC_TAOS_FORBID
     #define bind   BIND_FUNC_TAOS_FORBID
@@ -52,9 +53,6 @@ extern "C" {
 
 #if (defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)) 
   #define htobe64 htonll
-  #if defined(_TD_GO_DLL_)
-    uint64_t htonll(uint64_t val);
-  #endif
 #endif
 
 #if defined(_TD_DARWIN_64)
diff --git a/include/os/osString.h b/include/os/osString.h
index 5846d6b7ed..66d69a849c 100644
--- a/include/os/osString.h
+++ b/include/os/osString.h
@@ -24,6 +24,7 @@ typedef wchar_t TdWchar;
 typedef int32_t TdUcs4;
 
 // If the error is in a third-party library, place this header file under the third-party library header file.
+// When you want to use this feature, you should find or add the same function in the following section.
 #ifndef ALLOW_FORBID_FUNC
     #define iconv_open ICONV_OPEN_FUNC_TAOS_FORBID
     #define iconv_close ICONV_CLOSE_FUNC_TAOS_FORBID
diff --git a/include/os/osSystem.h b/include/os/osSystem.h
index 413dae8bfb..15959a2d8c 100644
--- a/include/os/osSystem.h
+++ b/include/os/osSystem.h
@@ -21,6 +21,7 @@ extern "C" {
 #endif
 
 // If the error is in a third-party library, place this header file under the third-party library header file.
+// When you want to use this feature, you should find or add the same function in the following section.
 #ifndef ALLOW_FORBID_FUNC
     #define popen POPEN_FUNC_TAOS_FORBID
     #define pclose PCLOSE_FUNC_TAOS_FORBID
diff --git a/include/os/osThread.h b/include/os/osThread.h
index 504814c1c6..6d8ff1de3e 100644
--- a/include/os/osThread.h
+++ b/include/os/osThread.h
@@ -37,6 +37,7 @@ typedef pthread_condattr_t TdThreadCondAttr;
 #define taosThreadCleanupPop pthread_cleanup_pop
 
 // If the error is in a third-party library, place this header file under the third-party library header file.
+// When you want to use this feature, you should find or add the same function in the following section.
 #ifndef ALLOW_FORBID_FUNC
     #define pthread_t PTHREAD_T_TYPE_TAOS_FORBID
     #define pthread_spinlock_t PTHREAD_SPINLOCK_T_TYPE_TAOS_FORBID
diff --git a/include/os/osTime.h b/include/os/osTime.h
index 4904eb9b51..031b9d28f9 100644
--- a/include/os/osTime.h
+++ b/include/os/osTime.h
@@ -21,6 +21,7 @@ extern "C" {
 #endif
 
 // If the error is in a third-party library, place this header file under the third-party library header file.
+// When you want to use this feature, you should find or add the same function in the following section.
 #ifndef ALLOW_FORBID_FUNC
     #define strptime STRPTIME_FUNC_TAOS_FORBID
     #define gettimeofday GETTIMEOFDAY_FUNC_TAOS_FORBID
@@ -33,11 +34,7 @@ extern "C" {
 
   #define CLOCK_REALTIME 	0
 
-  #ifdef _TD_GO_DLL_
-    #define MILLISECOND_PER_SECOND (1000LL)
-  #else
-    #define MILLISECOND_PER_SECOND (1000i64)
-  #endif
+  #define MILLISECOND_PER_SECOND (1000i64)
 #else
   #define MILLISECOND_PER_SECOND ((int64_t)1000L)
 #endif
diff --git a/include/os/osTimer.h b/include/os/osTimer.h
index 9040113b23..5fece43217 100644
--- a/include/os/osTimer.h
+++ b/include/os/osTimer.h
@@ -21,6 +21,7 @@ extern "C" {
 #endif
 
 // If the error is in a third-party library, place this header file under the third-party library header file.
+// When you want to use this feature, you should find or add the same function in the following section.
 #ifndef ALLOW_FORBID_FUNC
     #define timer_create TIMER_CREATE_FUNC_TAOS_FORBID
     #define timer_settime TIMER_SETTIME_FUNC_TAOS_FORBID
diff --git a/include/os/osTimezone.h b/include/os/osTimezone.h
index fa0e70f209..c8df8c3f3d 100644
--- a/include/os/osTimezone.h
+++ b/include/os/osTimezone.h
@@ -21,6 +21,7 @@ extern "C" {
 #endif
 
 // If the error is in a third-party library, place this header file under the third-party library header file.
+// When you want to use this feature, you should find or add the same function in the following section.
 #ifndef ALLOW_FORBID_FUNC
     #define tzset TZSET_FUNC_TAOS_FORBID
 #endif
diff --git a/source/client/src/clientEnv.c b/source/client/src/clientEnv.c
index 07f925bce8..6ebf9e71e0 100644
--- a/source/client/src/clientEnv.c
+++ b/source/client/src/clientEnv.c
@@ -48,8 +48,8 @@ static void registerRequest(SRequestObj *pRequest) {
   if (pTscObj->pAppInfo) {
     SInstanceSummary *pSummary = &pTscObj->pAppInfo->summary;
 
-    int32_t total = atomic_add_fetch_32(&pSummary->totalRequests, 1);
-    int32_t currentInst = atomic_add_fetch_32(&pSummary->currentRequests, 1);
+    int32_t total = atomic_add_fetch_64(&pSummary->totalRequests, 1);
+    int32_t currentInst = atomic_add_fetch_64(&pSummary->currentRequests, 1);
     tscDebug("0x%" PRIx64 " new Request from connObj:0x%" PRIx64
              ", current:%d, app current:%d, total:%d, reqId:0x%" PRIx64,
              pRequest->self, pRequest->pTscObj->id, num, currentInst, total, pRequest->requestId);
@@ -62,7 +62,7 @@ static void deregisterRequest(SRequestObj *pRequest) {
   STscObj *         pTscObj = pRequest->pTscObj;
   SInstanceSummary *pActivity = &pTscObj->pAppInfo->summary;
 
-  int32_t currentInst = atomic_sub_fetch_32(&pActivity->currentRequests, 1);
+  int32_t currentInst = atomic_sub_fetch_64(&pActivity->currentRequests, 1);
   int32_t num = atomic_sub_fetch_32(&pTscObj->numOfReqs, 1);
 
   int64_t duration = taosGetTimestampMs() - pRequest->metric.start;
diff --git a/source/client/src/tmq.c b/source/client/src/tmq.c
index a8fc394720..7c35b7c149 100644
--- a/source/client/src/tmq.c
+++ b/source/client/src/tmq.c
@@ -681,7 +681,7 @@ static char* formatTimestamp(char* buf, int64_t val, int precision) {
 
 int32_t tmqGetSkipLogNum(tmq_message_t* tmq_message) {
   if (tmq_message == NULL) return 0;
-  SMqPollRsp* pRsp = &tmq_message->consumeRsp;
+  SMqPollRsp* pRsp = &tmq_message->msg;
   return pRsp->skipLogNum;
 }
 
@@ -690,15 +690,15 @@ void tmqShowMsg(tmq_message_t* tmq_message) {
 
   static bool noPrintSchema;
   char        pBuf[128];
-  SMqPollRsp* pRsp = &tmq_message->consumeRsp;
-  int32_t     colNum = pRsp->schemas->nCols;
+  SMqPollRsp* pRsp = &tmq_message->msg;
+  int32_t     colNum = pRsp->schema->nCols;
   if (!noPrintSchema) {
     printf("|");
     for (int32_t i = 0; i < colNum; i++) {
       if (i == 0)
-        printf(" %25s |", pRsp->schemas->pSchema[i].name);
+        printf(" %25s |", pRsp->schema->pSchema[i].name);
       else
-        printf(" %15s |", pRsp->schemas->pSchema[i].name);
+        printf(" %15s |", pRsp->schema->pSchema[i].name);
     }
     printf("\n");
     printf("===============================================\n");
@@ -735,7 +735,7 @@ int32_t tmqPollCb(void* param, const SDataBuf* pMsg, int32_t code) {
   SMqClientVg*    pVg = pParam->pVg;
   tmq_t*          tmq = pParam->tmq;
   if (code != 0) {
-    printf("msg discard\n");
+    printf("msg discard %x\n", code);
     goto WRITE_QUEUE_FAIL;
   }
 
@@ -778,19 +778,19 @@ int32_t tmqPollCb(void* param, const SDataBuf* pMsg, int32_t code) {
     goto WRITE_QUEUE_FAIL;
   }
   memcpy(pRsp, pMsg->pData, sizeof(SMqRspHead));
-  tDecodeSMqPollRsp(POINTER_SHIFT(pMsg->pData, sizeof(SMqRspHead)), &pRsp->consumeRsp);
-  pRsp->curBlock = 0;
-  pRsp->curRow = 0;
+  tDecodeSMqPollRsp(POINTER_SHIFT(pMsg->pData, sizeof(SMqRspHead)), &pRsp->msg);
+  pRsp->iter.curBlock = 0;
+  pRsp->iter.curRow = 0;
   // TODO: alloc mem
   /*pRsp->*/
   /*printf("rsp commit off:%ld rsp off:%ld has data:%d\n", pRsp->committedOffset, pRsp->rspOffset, pRsp->numOfTopics);*/
-  if (pRsp->consumeRsp.numOfTopics == 0) {
+  if (pRsp->msg.numOfTopics == 0) {
     /*printf("no data\n");*/
     taosFreeQitem(pRsp);
     goto WRITE_QUEUE_FAIL;
   }
 
-  pRsp->extra = pParam->pVg;
+  pRsp->vg = pParam->pVg;
   taosWriteQitem(tmq->mqueue, pRsp);
   atomic_add_fetch_32(&tmq->readyRequest, 1);
   tsem_post(&tmq->rspSem);
@@ -860,14 +860,14 @@ int32_t tmqAskEpCb(void* param, const SDataBuf* pMsg, int32_t code) {
     }
     tDeleteSMqCMGetSubEpRsp(&rsp);
   } else {
-    tmq_message_t* pRsp = taosAllocateQitem(sizeof(tmq_message_t));
+    SMqCMGetSubEpRsp* pRsp = taosAllocateQitem(sizeof(SMqCMGetSubEpRsp));
     if (pRsp == NULL) {
       terrno = TSDB_CODE_OUT_OF_MEMORY;
       code = -1;
       goto END;
     }
     memcpy(pRsp, pMsg->pData, sizeof(SMqRspHead));
-    tDecodeSMqCMGetSubEpRsp(POINTER_SHIFT(pMsg->pData, sizeof(SMqRspHead)), &pRsp->getEpRsp);
+    tDecodeSMqCMGetSubEpRsp(POINTER_SHIFT(pMsg->pData, sizeof(SMqRspHead)), pRsp);
 
     taosWriteQitem(tmq->mqueue, pRsp);
     tsem_post(&tmq->rspSem);
@@ -983,6 +983,7 @@ SMqPollReq* tmqBuildConsumeReqImpl(tmq_t* tmq, int64_t blockingTime, SMqClientTo
   return pReq;
 }
 
+#if 0
 tmq_message_t* tmqSyncPollImpl(tmq_t* tmq, int64_t blockingTime) {
   tmq_message_t* msg = NULL;
   for (int i = 0; i < taosArrayGetSize(tmq->clientTopics); i++) {
@@ -1050,6 +1051,7 @@ tmq_message_t* tmqSyncPollImpl(tmq_t* tmq, int64_t blockingTime) {
   }
   return NULL;
 }
+#endif
 
 int32_t tmqPollImpl(tmq_t* tmq, int64_t blockingTime) {
   /*printf("call poll\n");*/
@@ -1111,11 +1113,12 @@ int32_t tmqPollImpl(tmq_t* tmq, int64_t blockingTime) {
 }
 
 // return
-int32_t tmqHandleRes(tmq_t* tmq, tmq_message_t* rspMsg, bool* pReset) {
-  if (rspMsg->head.mqMsgType == TMQ_MSG_TYPE__EP_RSP) {
+int32_t tmqHandleRes(tmq_t* tmq, SMqRspHead* rspHead, bool* pReset) {
+  if (rspHead->mqMsgType == TMQ_MSG_TYPE__EP_RSP) {
     /*printf("ep %d %d\n", rspMsg->head.epoch, tmq->epoch);*/
-    if (rspMsg->head.epoch > atomic_load_32(&tmq->epoch)) {
-      tmqUpdateEp(tmq, rspMsg->head.epoch, &rspMsg->getEpRsp);
+    if (rspHead->epoch > atomic_load_32(&tmq->epoch)) {
+      SMqCMGetSubEpRsp* rspMsg = (SMqCMGetSubEpRsp*)rspHead;
+      tmqUpdateEp(tmq, rspHead->epoch, rspMsg);
       tmqClearUnhandleMsg(tmq);
       *pReset = true;
     } else {
@@ -1129,21 +1132,22 @@ int32_t tmqHandleRes(tmq_t* tmq, tmq_message_t* rspMsg, bool* pReset) {
 
 tmq_message_t* tmqHandleAllRsp(tmq_t* tmq, int64_t blockingTime, bool pollIfReset) {
   while (1) {
-    tmq_message_t* rspMsg = NULL;
-    taosGetQitem(tmq->qall, (void**)&rspMsg);
-    if (rspMsg == NULL) {
+    SMqRspHead* rspHead = NULL;
+    taosGetQitem(tmq->qall, (void**)&rspHead);
+    if (rspHead == NULL) {
       taosReadAllQitems(tmq->mqueue, tmq->qall);
-      taosGetQitem(tmq->qall, (void**)&rspMsg);
-      if (rspMsg == NULL) return NULL;
+      taosGetQitem(tmq->qall, (void**)&rspHead);
+      if (rspHead == NULL) return NULL;
     }
 
-    if (rspMsg->head.mqMsgType == TMQ_MSG_TYPE__POLL_RSP) {
+    if (rspHead->mqMsgType == TMQ_MSG_TYPE__POLL_RSP) {
+      tmq_message_t* rspMsg = (tmq_message_t*)rspHead;
       atomic_sub_fetch_32(&tmq->readyRequest, 1);
       /*printf("handle poll rsp %d\n", rspMsg->head.mqMsgType);*/
-      if (rspMsg->head.epoch == atomic_load_32(&tmq->epoch)) {
+      if (rspMsg->msg.head.epoch == atomic_load_32(&tmq->epoch)) {
         /*printf("epoch match\n");*/
-        SMqClientVg* pVg = rspMsg->extra;
-        pVg->currentOffset = rspMsg->consumeRsp.rspOffset;
+        SMqClientVg* pVg = rspMsg->vg;
+        pVg->currentOffset = rspMsg->msg.rspOffset;
         atomic_store_32(&pVg->vgStatus, TMQ_VG_STATUS__IDLE);
         return rspMsg;
       } else {
@@ -1153,8 +1157,8 @@ tmq_message_t* tmqHandleAllRsp(tmq_t* tmq, int64_t blockingTime, bool pollIfRese
     } else {
       /*printf("handle ep rsp %d\n", rspMsg->head.mqMsgType);*/
       bool reset = false;
-      tmqHandleRes(tmq, rspMsg, &reset);
-      taosFreeQitem(rspMsg);
+      tmqHandleRes(tmq, rspHead, &reset);
+      taosFreeQitem(rspHead);
       if (pollIfReset && reset) {
         printf("reset and repoll\n");
         tmqPollImpl(tmq, blockingTime);
@@ -1163,6 +1167,7 @@ tmq_message_t* tmqHandleAllRsp(tmq_t* tmq, int64_t blockingTime, bool pollIfRese
   }
 }
 
+#if 0
 tmq_message_t* tmq_consumer_poll_v1(tmq_t* tmq, int64_t blocking_time) {
   tmq_message_t* rspMsg = NULL;
   int64_t        startTime = taosGetTimestampMs();
@@ -1185,6 +1190,7 @@ tmq_message_t* tmq_consumer_poll_v1(tmq_t* tmq, int64_t blocking_time) {
       return NULL;
   }
 }
+#endif
 
 tmq_message_t* tmq_consumer_poll(tmq_t* tmq, int64_t blocking_time) {
   tmq_message_t* rspMsg;
@@ -1350,7 +1356,7 @@ tmq_resp_err_t tmq_commit(tmq_t* tmq, const tmq_topic_vgroup_list_t* tmq_topic_v
 
 void tmq_message_destroy(tmq_message_t* tmq_message) {
   if (tmq_message == NULL) return;
-  SMqPollRsp* pRsp = &tmq_message->consumeRsp;
+  SMqPollRsp* pRsp = &tmq_message->msg;
   tDeleteSMqConsumeRsp(pRsp);
   /*free(tmq_message);*/
   taosFreeQitem(tmq_message);
@@ -1366,24 +1372,24 @@ const char* tmq_err2str(tmq_resp_err_t err) {
 }
 
 TAOS_ROW tmq_get_row(tmq_message_t* message) {
-  SMqPollRsp* rsp = &message->consumeRsp;
+  SMqPollRsp* rsp = &message->msg;
   while (1) {
-    if (message->curBlock < taosArrayGetSize(rsp->pBlockData)) {
-      SSDataBlock* pBlock = taosArrayGet(rsp->pBlockData, message->curBlock);
-      if (message->curRow < pBlock->info.rows) {
+    if (message->iter.curBlock < taosArrayGetSize(rsp->pBlockData)) {
+      SSDataBlock* pBlock = taosArrayGet(rsp->pBlockData, message->iter.curBlock);
+      if (message->iter.curRow < pBlock->info.rows) {
         for (int i = 0; i < pBlock->info.numOfCols; i++) {
           SColumnInfoData* pData = taosArrayGet(pBlock->pDataBlock, i);
-          if (colDataIsNull_s(pData, message->curRow))
-            message->uData[i] = NULL;
+          if (colDataIsNull_s(pData, message->iter.curRow))
+            message->iter.uData[i] = NULL;
           else {
-            message->uData[i] = colDataGetData(pData, message->curRow);
+            message->iter.uData[i] = colDataGetData(pData, message->iter.curRow);
           }
         }
-        message->curRow++;
-        return message->uData;
+        message->iter.curRow++;
+        return message->iter.uData;
       } else {
-        message->curBlock++;
-        message->curRow = 0;
+        message->iter.curBlock++;
+        message->iter.curRow = 0;
         continue;
       }
     }
diff --git a/source/common/src/tdatablock.c b/source/common/src/tdatablock.c
index 31328cc4b2..0d48d7cc14 100644
--- a/source/common/src/tdatablock.c
+++ b/source/common/src/tdatablock.c
@@ -18,14 +18,14 @@
 #include "tcompare.h"
 #include "tglobal.h"
 
-int32_t taosGetFqdnPortFromEp(const char *ep, SEp* pEp) {
+int32_t taosGetFqdnPortFromEp(const char* ep, SEp* pEp) {
   pEp->port = 0;
   strcpy(pEp->fqdn, ep);
 
-  char *temp = strchr(pEp->fqdn, ':');
+  char* temp = strchr(pEp->fqdn, ':');
   if (temp) {
     *temp = 0;
-    pEp->port = atoi(temp+1);
+    pEp->port = atoi(temp + 1);
   }
 
   if (pEp->port == 0) {
@@ -36,7 +36,7 @@ int32_t taosGetFqdnPortFromEp(const char *ep, SEp* pEp) {
   return 0;
 }
 
-void addEpIntoEpSet(SEpSet *pEpSet, const char* fqdn, uint16_t port) {
+void addEpIntoEpSet(SEpSet* pEpSet, const char* fqdn, uint16_t port) {
   if (pEpSet == NULL || fqdn == NULL || strlen(fqdn) == 0) {
     return;
   }
@@ -47,26 +47,25 @@ void addEpIntoEpSet(SEpSet *pEpSet, const char* fqdn, uint16_t port) {
   pEpSet->numOfEps += 1;
 }
 
-bool isEpsetEqual(const SEpSet *s1, const SEpSet *s2) {
+bool isEpsetEqual(const SEpSet* s1, const SEpSet* s2) {
   if (s1->numOfEps != s2->numOfEps || s1->inUse != s2->inUse) {
     return false;
   }
 
   for (int32_t i = 0; i < s1->numOfEps; i++) {
-    if (s1->eps[i].port != s2->eps[i].port
-        || strncmp(s1->eps[i].fqdn, s2->eps[i].fqdn, TSDB_FQDN_LEN) != 0)
+    if (s1->eps[i].port != s2->eps[i].port || strncmp(s1->eps[i].fqdn, s2->eps[i].fqdn, TSDB_FQDN_LEN) != 0)
       return false;
   }
   return true;
 }
 
-void updateEpSet_s(SCorEpSet *pEpSet, SEpSet *pNewEpSet) {
+void updateEpSet_s(SCorEpSet* pEpSet, SEpSet* pNewEpSet) {
   taosCorBeginWrite(&pEpSet->version);
   pEpSet->epSet = *pNewEpSet;
   taosCorEndWrite(&pEpSet->version);
 }
 
-SEpSet getEpSet_s(SCorEpSet *pEpSet) {
+SEpSet getEpSet_s(SCorEpSet* pEpSet) {
   SEpSet ep = {0};
   taosCorBeginRead(&pEpSet->version);
   ep = pEpSet->epSet;
@@ -75,7 +74,6 @@ SEpSet getEpSet_s(SCorEpSet *pEpSet) {
   return ep;
 }
 
-
 int32_t colDataGetLength(const SColumnInfoData* pColumnInfoData, int32_t numOfRows) {
   ASSERT(pColumnInfoData != NULL);
   if (IS_VAR_DATA_TYPE(pColumnInfoData->info.type)) {
@@ -95,7 +93,7 @@ int32_t colDataAppend(SColumnInfoData* pColumnInfoData, uint32_t currentRow, con
   if (isNull) {
     // There is a placehold for each NULL value of binary or nchar type.
     if (IS_VAR_DATA_TYPE(pColumnInfoData->info.type)) {
-      pColumnInfoData->varmeta.offset[currentRow] = -1; // it is a null value of VAR type.
+      pColumnInfoData->varmeta.offset[currentRow] = -1;  // it is a null value of VAR type.
     } else {
       colDataSetNull_f(pColumnInfoData->nullbitmap, currentRow);
     }
@@ -113,7 +111,7 @@ int32_t colDataAppend(SColumnInfoData* pColumnInfoData, uint32_t currentRow, con
         newSize = 8;
       }
 
-      while(newSize < pAttr->length + varDataTLen(pData)) {
+      while (newSize < pAttr->length + varDataTLen(pData)) {
         newSize = newSize * 1.5;
       }
 
@@ -133,19 +131,40 @@ int32_t colDataAppend(SColumnInfoData* pColumnInfoData, uint32_t currentRow, con
     pColumnInfoData->varmeta.length += varDataTLen(pData);
   } else {
     char* p = pColumnInfoData->pData + pColumnInfoData->info.bytes * currentRow;
-    switch(type) {
-      case TSDB_DATA_TYPE_BOOL: {*(bool*) p = *(bool*) pData;break;}
+    switch (type) {
+      case TSDB_DATA_TYPE_BOOL: {
+        *(bool*)p = *(bool*)pData;
+        break;
+      }
       case TSDB_DATA_TYPE_TINYINT:
-      case TSDB_DATA_TYPE_UTINYINT: {*(int8_t*) p = *(int8_t*) pData;break;}
+      case TSDB_DATA_TYPE_UTINYINT: {
+        *(int8_t*)p = *(int8_t*)pData;
+        break;
+      }
       case TSDB_DATA_TYPE_SMALLINT:
-      case TSDB_DATA_TYPE_USMALLINT: {*(int16_t*) p = *(int16_t*) pData;break;}
+      case TSDB_DATA_TYPE_USMALLINT: {
+        *(int16_t*)p = *(int16_t*)pData;
+        break;
+      }
       case TSDB_DATA_TYPE_INT:
-      case TSDB_DATA_TYPE_UINT: {*(int32_t*) p = *(int32_t*) pData;break;}
+      case TSDB_DATA_TYPE_UINT: {
+        *(int32_t*)p = *(int32_t*)pData;
+        break;
+      }
       case TSDB_DATA_TYPE_TIMESTAMP:
       case TSDB_DATA_TYPE_BIGINT:
-      case TSDB_DATA_TYPE_UBIGINT: {*(int64_t*) p = *(int64_t*) pData;break;}
-      case TSDB_DATA_TYPE_FLOAT: {*(float*) p = *(float*) pData;break;}
-      case TSDB_DATA_TYPE_DOUBLE: {*(double*) p = *(double*) pData;break;}
+      case TSDB_DATA_TYPE_UBIGINT: {
+        *(int64_t*)p = *(int64_t*)pData;
+        break;
+      }
+      case TSDB_DATA_TYPE_FLOAT: {
+        *(float*)p = *(float*)pData;
+        break;
+      }
+      case TSDB_DATA_TYPE_DOUBLE: {
+        *(double*)p = *(double*)pData;
+        break;
+      }
       default:
         assert(0);
     }
@@ -154,7 +173,8 @@ int32_t colDataAppend(SColumnInfoData* pColumnInfoData, uint32_t currentRow, con
   return 0;
 }
 
-static void doBitmapMerge(SColumnInfoData* pColumnInfoData, int32_t numOfRow1, const SColumnInfoData* pSource, int32_t numOfRow2) {
+static void doBitmapMerge(SColumnInfoData* pColumnInfoData, int32_t numOfRow1, const SColumnInfoData* pSource,
+                          int32_t numOfRow2) {
   uint32_t total = numOfRow1 + numOfRow2;
 
   if (BitmapLen(numOfRow1) < BitmapLen(total)) {
@@ -188,7 +208,8 @@ static void doBitmapMerge(SColumnInfoData* pColumnInfoData, int32_t numOfRow1, c
   }
 }
 
-int32_t colDataMergeCol(SColumnInfoData* pColumnInfoData, uint32_t numOfRow1, const SColumnInfoData* pSource, uint32_t numOfRow2) {
+int32_t colDataMergeCol(SColumnInfoData* pColumnInfoData, uint32_t numOfRow1, const SColumnInfoData* pSource,
+                        uint32_t numOfRow2) {
   ASSERT(pColumnInfoData != NULL && pSource != NULL && pColumnInfoData->info.type == pSource->info.type);
 
   if (numOfRow2 == 0) {
@@ -202,8 +223,8 @@ int32_t colDataMergeCol(SColumnInfoData* pColumnInfoData, uint32_t numOfRow1, co
       // TODO
     }
 
-    pColumnInfoData->varmeta.offset = (int32_t*) p;
-    for(int32_t i = 0; i < numOfRow2; ++i) {
+    pColumnInfoData->varmeta.offset = (int32_t*)p;
+    for (int32_t i = 0; i < numOfRow2; ++i) {
       pColumnInfoData->varmeta.offset[i + numOfRow1] = pSource->varmeta.offset[i] + pColumnInfoData->varmeta.length;
     }
 
@@ -244,9 +265,7 @@ size_t blockDataGetNumOfCols(const SSDataBlock* pBlock) {
   return pBlock->info.numOfCols;
 }
 
-size_t blockDataGetNumOfRows(const SSDataBlock* pBlock) {
-  return pBlock->info.rows;
-}
+size_t blockDataGetNumOfRows(const SSDataBlock* pBlock) { return pBlock->info.rows; }
 
 int32_t blockDataUpdateTsWindow(SSDataBlock* pDataBlock) {
   if (pDataBlock == NULL || pDataBlock->info.rows <= 0) {
@@ -263,8 +282,8 @@ int32_t blockDataUpdateTsWindow(SSDataBlock* pDataBlock) {
   }
 
   ASSERT(pColInfoData->nullbitmap == NULL);
-  pDataBlock->info.window.skey = *(TSKEY*) colDataGetData(pColInfoData, 0);
-  pDataBlock->info.window.ekey = *(TSKEY*) colDataGetData(pColInfoData, (pDataBlock->info.rows - 1));
+  pDataBlock->info.window.skey = *(TSKEY*)colDataGetData(pColInfoData, 0);
+  pDataBlock->info.window.ekey = *(TSKEY*)colDataGetData(pColInfoData, (pDataBlock->info.rows - 1));
   return 0;
 }
 
@@ -272,7 +291,7 @@ int32_t blockDataMerge(SSDataBlock* pDest, const SSDataBlock* pSrc) {
   assert(pSrc != NULL && pDest != NULL && pDest->info.numOfCols == pSrc->info.numOfCols);
 
   int32_t numOfCols = pSrc->info.numOfCols;
-  for(int32_t i = 0; i < numOfCols; ++i) {
+  for (int32_t i = 0; i < numOfCols; ++i) {
     SColumnInfoData* pCol2 = taosArrayGet(pDest->pDataBlock, i);
     SColumnInfoData* pCol1 = taosArrayGet(pSrc->pDataBlock, i);
 
@@ -280,7 +299,7 @@ int32_t blockDataMerge(SSDataBlock* pDest, const SSDataBlock* pSrc) {
     uint32_t newLen = colDataGetLength(pCol1, pSrc->info.rows);
 
     int32_t newSize = oldLen + newLen;
-    char* tmp = realloc(pCol2->pData, newSize);
+    char*   tmp = realloc(pCol2->pData, newSize);
     if (tmp != NULL) {
       pCol2->pData = tmp;
       colDataMergeCol(pCol2, pDest->info.rows, pCol1, pSrc->info.rows);
@@ -296,10 +315,10 @@ int32_t blockDataMerge(SSDataBlock* pDest, const SSDataBlock* pSrc) {
 size_t blockDataGetSize(const SSDataBlock* pBlock) {
   assert(pBlock != NULL);
 
-  size_t total = 0;
+  size_t  total = 0;
   int32_t numOfCols = pBlock->info.numOfCols;
 
-  for(int32_t i = 0; i < numOfCols; ++i) {
+  for (int32_t i = 0; i < numOfCols; ++i) {
     SColumnInfoData* pColInfoData = taosArrayGet(pBlock->pDataBlock, i);
     total += colDataGetLength(pColInfoData, pBlock->info.rows);
 
@@ -315,7 +334,8 @@ size_t blockDataGetSize(const SSDataBlock* pBlock) {
 
 // the number of tuples can be fit in one page.
 // Actual data rows pluses the corresponding meta data must fit in one memory buffer of the given page size.
-int32_t blockDataSplitRows(SSDataBlock* pBlock, bool hasVarCol, int32_t startIndex, int32_t* stopIndex, int32_t pageSize) {
+int32_t blockDataSplitRows(SSDataBlock* pBlock, bool hasVarCol, int32_t startIndex, int32_t* stopIndex,
+                           int32_t pageSize) {
   ASSERT(pBlock != NULL && stopIndex != NULL);
 
   int32_t numOfCols = pBlock->info.numOfCols;
@@ -323,13 +343,13 @@ int32_t blockDataSplitRows(SSDataBlock* pBlock, bool hasVarCol, int32_t startInd
 
   int32_t bitmapChar = 1;
 
-  size_t headerSize    = sizeof(int32_t);
+  size_t headerSize = sizeof(int32_t);
   size_t colHeaderSize = sizeof(int32_t) * numOfCols;
-  size_t payloadSize   = pageSize - (headerSize + colHeaderSize);
+  size_t payloadSize = pageSize - (headerSize + colHeaderSize);
 
   // TODO speedup by checking if the whole page can fit in firstly.
   if (!hasVarCol) {
-    size_t rowSize = blockDataGetRowSize(pBlock);
+    size_t  rowSize = blockDataGetRowSize(pBlock);
     int32_t capacity = (payloadSize / (rowSize * 8 + bitmapChar * numOfCols)) * 8;
 
     *stopIndex = startIndex + capacity;
@@ -342,7 +362,7 @@ int32_t blockDataSplitRows(SSDataBlock* pBlock, bool hasVarCol, int32_t startInd
     // iterate the rows that can be fit in this buffer page
     int32_t size = (headerSize + colHeaderSize);
 
-    for(int32_t j = startIndex; j < numOfRows; ++j) {
+    for (int32_t j = startIndex; j < numOfRows; ++j) {
       for (int32_t i = 0; i < numOfCols; ++i) {
         SColumnInfoData* pColInfoData = TARRAY_GET_ELEM(pBlock->pDataBlock, i);
         if (IS_VAR_DATA_TYPE(pColInfoData->info.type)) {
@@ -359,7 +379,7 @@ int32_t blockDataSplitRows(SSDataBlock* pBlock, bool hasVarCol, int32_t startInd
           size += pColInfoData->info.bytes;
 
           if (((j - startIndex) & 0x07) == 0) {
-            size += 1; // the space for null bitmap
+            size += 1;  // the space for null bitmap
           }
         }
       }
@@ -393,8 +413,8 @@ SSDataBlock* blockDataExtractBlock(SSDataBlock* pBlock, int32_t startIndex, int3
   pDst->info.rows = 0;
   pDst->pDataBlock = taosArrayInit(pBlock->info.numOfCols, sizeof(SColumnInfoData));
 
-  for(int32_t i = 0; i < pBlock->info.numOfCols; ++i) {
-    SColumnInfoData colInfo = {0};
+  for (int32_t i = 0; i < pBlock->info.numOfCols; ++i) {
+    SColumnInfoData  colInfo = {0};
     SColumnInfoData* pSrcCol = taosArrayGet(pBlock->pDataBlock, i);
     colInfo.info = pSrcCol->info;
 
@@ -414,7 +434,7 @@ SSDataBlock* blockDataExtractBlock(SSDataBlock* pBlock, int32_t startIndex, int3
     SColumnInfoData* pDstCol = taosArrayGet(pDst->pDataBlock, i);
 
     for (int32_t j = startIndex; j < (startIndex + rowCount); ++j) {
-      bool isNull = colDataIsNull(pColData, pBlock->info.rows, j, pBlock->pBlockAgg);
+      bool  isNull = colDataIsNull(pColData, pBlock->info.rows, j, pBlock->pBlockAgg);
       char* p = colDataGetData(pColData, j);
 
       colDataAppend(pDstCol, j - startIndex, p, isNull);
@@ -440,14 +460,14 @@ int32_t blockDataToBuf(char* buf, const SSDataBlock* pBlock) {
   ASSERT(pBlock != NULL);
 
   // write the number of rows
-  *(uint32_t*) buf = pBlock->info.rows;
+  *(uint32_t*)buf = pBlock->info.rows;
 
   int32_t numOfCols = pBlock->info.numOfCols;
   int32_t numOfRows = pBlock->info.rows;
 
   char* pStart = buf + sizeof(uint32_t);
 
-  for(int32_t i = 0; i < numOfCols; ++i) {
+  for (int32_t i = 0; i < numOfCols; ++i) {
     SColumnInfoData* pCol = taosArrayGet(pBlock->pDataBlock, i);
     if (IS_VAR_DATA_TYPE(pCol->info.type)) {
       memcpy(pStart, pCol->varmeta.offset, numOfRows * sizeof(int32_t));
@@ -459,7 +479,7 @@ int32_t blockDataToBuf(char* buf, const SSDataBlock* pBlock) {
 
     uint32_t dataSize = colDataGetLength(pCol, numOfRows);
 
-    *(int32_t*) pStart = dataSize;
+    *(int32_t*)pStart = dataSize;
     pStart += sizeof(int32_t);
 
     memcpy(pStart, pCol->pData, dataSize);
@@ -470,12 +490,12 @@ int32_t blockDataToBuf(char* buf, const SSDataBlock* pBlock) {
 }
 
 int32_t blockDataFromBuf(SSDataBlock* pBlock, const char* buf) {
-  pBlock->info.rows = *(int32_t*) buf;
+  pBlock->info.rows = *(int32_t*)buf;
 
-  int32_t numOfCols = pBlock->info.numOfCols;
+  int32_t     numOfCols = pBlock->info.numOfCols;
   const char* pStart = buf + sizeof(uint32_t);
 
-  for(int32_t i = 0; i < numOfCols; ++i) {
+  for (int32_t i = 0; i < numOfCols; ++i) {
     SColumnInfoData* pCol = taosArrayGet(pBlock->pDataBlock, i);
 
     size_t metaSize = pBlock->info.rows * sizeof(int32_t);
@@ -487,7 +507,7 @@ int32_t blockDataFromBuf(SSDataBlock* pBlock, const char* buf) {
       pStart += BitmapLen(pBlock->info.rows);
     }
 
-    int32_t colLength = *(int32_t*) pStart;
+    int32_t colLength = *(int32_t*)pStart;
     pStart += sizeof(int32_t);
 
     if (IS_VAR_DATA_TYPE(pCol->info.type)) {
@@ -517,7 +537,7 @@ size_t blockDataGetRowSize(const SSDataBlock* pBlock) {
   size_t rowSize = 0;
 
   size_t numOfCols = pBlock->info.numOfCols;
-  for(int32_t i = 0; i < numOfCols; ++i) {
+  for (int32_t i = 0; i < numOfCols; ++i) {
     SColumnInfoData* pColInfo = taosArrayGet(pBlock->pDataBlock, i);
     rowSize += pColInfo->info.bytes;
   }
@@ -537,10 +557,10 @@ size_t blockDataGetSerialMetaSize(const SSDataBlock* pBlock) {
 
 SSchema* blockDataExtractSchema(const SSDataBlock* pBlock, int32_t* numOfCols) {
   SSchema* pSchema = calloc(pBlock->info.numOfCols, sizeof(SSchema));
-  for(int32_t i = 0; i < pBlock->info.numOfCols; ++i) {
+  for (int32_t i = 0; i < pBlock->info.numOfCols; ++i) {
     SColumnInfoData* pColInfoData = taosArrayGet(pBlock->pDataBlock, i);
     pSchema[i].bytes = pColInfoData->info.bytes;
-    pSchema[i].type  = pColInfoData->info.type;
+    pSchema[i].type = pColInfoData->info.type;
     pSchema[i].colId = pColInfoData->info.colId;
   }
 
@@ -556,14 +576,14 @@ double blockDataGetSerialRowSize(const SSDataBlock* pBlock) {
   double rowSize = 0;
 
   size_t numOfCols = pBlock->info.numOfCols;
-  for(int32_t i = 0; i < numOfCols; ++i) {
+  for (int32_t i = 0; i < numOfCols; ++i) {
     SColumnInfoData* pColInfo = taosArrayGet(pBlock->pDataBlock, i);
     rowSize += pColInfo->info.bytes;
 
     if (IS_VAR_DATA_TYPE(pColInfo->info.type)) {
       rowSize += sizeof(int32_t);
     } else {
-      rowSize += 1/8.0;
+      rowSize += 1 / 8.0;
     }
   }
 
@@ -571,56 +591,56 @@ double blockDataGetSerialRowSize(const SSDataBlock* pBlock) {
 }
 
 typedef struct SSDataBlockSortHelper {
-  SArray      *orderInfo;   // SArray<SBlockOrderInfo>
-  SSDataBlock *pDataBlock;
+  SArray*      orderInfo;  // SArray<SBlockOrderInfo>
+  SSDataBlock* pDataBlock;
   bool         nullFirst;
 } SSDataBlockSortHelper;
 
 int32_t dataBlockCompar(const void* p1, const void* p2, const void* param) {
-  const SSDataBlockSortHelper* pHelper = (const SSDataBlockSortHelper*) param;
+  const SSDataBlockSortHelper* pHelper = (const SSDataBlockSortHelper*)param;
 
   SSDataBlock* pDataBlock = pHelper->pDataBlock;
 
-  int32_t left  = *(int32_t*) p1;
-  int32_t right = *(int32_t*) p2;
+  int32_t left = *(int32_t*)p1;
+  int32_t right = *(int32_t*)p2;
 
   SArray* pInfo = pHelper->orderInfo;
 
-  for(int32_t i = 0; i < pInfo->size; ++i) {
+  for (int32_t i = 0; i < pInfo->size; ++i) {
     SBlockOrderInfo* pOrder = TARRAY_GET_ELEM(pInfo, i);
-    SColumnInfoData* pColInfoData = pOrder->pColData;//TARRAY_GET_ELEM(pDataBlock->pDataBlock, pOrder->colIndex);
+    SColumnInfoData* pColInfoData = pOrder->pColData;  // TARRAY_GET_ELEM(pDataBlock->pDataBlock, pOrder->colIndex);
 
     if (pColInfoData->hasNull) {
-      bool leftNull  = colDataIsNull(pColInfoData, pDataBlock->info.rows, left, pDataBlock->pBlockAgg);
+      bool leftNull = colDataIsNull(pColInfoData, pDataBlock->info.rows, left, pDataBlock->pBlockAgg);
       bool rightNull = colDataIsNull(pColInfoData, pDataBlock->info.rows, right, pDataBlock->pBlockAgg);
       if (leftNull && rightNull) {
-        continue; // continue to next slot
+        continue;  // continue to next slot
       }
 
       if (rightNull) {
-        return pHelper->nullFirst? 1:-1;
+        return pHelper->nullFirst ? 1 : -1;
       }
 
       if (leftNull) {
-        return pHelper->nullFirst? -1:1;
+        return pHelper->nullFirst ? -1 : 1;
       }
     }
 
-    void* left1  = colDataGetData(pColInfoData, left);
+    void* left1 = colDataGetData(pColInfoData, left);
     void* right1 = colDataGetData(pColInfoData, right);
 
-    switch(pColInfoData->info.type) {
+    switch (pColInfoData->info.type) {
       case TSDB_DATA_TYPE_INT: {
-        int32_t leftx  = *(int32_t*) left1;
-        int32_t rightx = *(int32_t*) right1;
+        int32_t leftx = *(int32_t*)left1;
+        int32_t rightx = *(int32_t*)right1;
 
         if (leftx == rightx) {
           break;
         } else {
           if (pOrder->order == TSDB_ORDER_ASC) {
-            return (leftx < rightx)? -1:1;
+            return (leftx < rightx) ? -1 : 1;
           } else {
-            return (leftx < rightx)? 1:-1;
+            return (leftx < rightx) ? 1 : -1;
           }
         }
       }
@@ -632,7 +652,8 @@ int32_t dataBlockCompar(const void* p1, const void* p2, const void* param) {
   return 0;
 }
 
-static int32_t doAssignOneTuple(SColumnInfoData* pDstCols, int32_t numOfRows, const SSDataBlock* pSrcBlock, int32_t tupleIndex) {
+static int32_t doAssignOneTuple(SColumnInfoData* pDstCols, int32_t numOfRows, const SSDataBlock* pSrcBlock,
+                                int32_t tupleIndex) {
   int32_t code = 0;
   int32_t numOfCols = pSrcBlock->info.numOfCols;
 
@@ -666,17 +687,17 @@ static int32_t blockDataAssign(SColumnInfoData* pCols, const SSDataBlock* pDataB
     }
   }
 #else
-  for(int32_t i = 0; i < pDataBlock->info.numOfCols; ++i) {
+  for (int32_t i = 0; i < pDataBlock->info.numOfCols; ++i) {
     SColumnInfoData* pDst = &pCols[i];
     SColumnInfoData* pSrc = taosArrayGet(pDataBlock->pDataBlock, i);
 
     if (IS_VAR_DATA_TYPE(pSrc->info.type)) {
-     memcpy(pDst->pData, pSrc->pData, pSrc->varmeta.length);
-     pDst->varmeta.length = pSrc->varmeta.length;
+      memcpy(pDst->pData, pSrc->pData, pSrc->varmeta.length);
+      pDst->varmeta.length = pSrc->varmeta.length;
 
-     for(int32_t j = 0; j < pDataBlock->info.rows; ++j) {
-       pDst->varmeta.offset[j] = pSrc->varmeta.offset[index[j]];
-     }
+      for (int32_t j = 0; j < pDataBlock->info.rows; ++j) {
+        pDst->varmeta.offset[j] = pSrc->varmeta.offset[index[j]];
+      }
     } else {
       switch (pSrc->info.type) {
         case TSDB_DATA_TYPE_UINT:
@@ -749,7 +770,7 @@ static SColumnInfoData* createHelpColInfoData(const SSDataBlock* pDataBlock) {
     return NULL;
   }
 
-  for(int32_t i = 0; i < numOfCols; ++i) {
+  for (int32_t i = 0; i < numOfCols; ++i) {
     SColumnInfoData* pColInfoData = taosArrayGet(pDataBlock->pDataBlock, i);
     pCols[i].info = pColInfoData->info;
 
@@ -771,7 +792,7 @@ static SColumnInfoData* createHelpColInfoData(const SSDataBlock* pDataBlock) {
 static void copyBackToBlock(SSDataBlock* pDataBlock, SColumnInfoData* pCols) {
   int32_t numOfCols = pDataBlock->info.numOfCols;
 
-  for(int32_t i = 0; i < numOfCols; ++i) {
+  for (int32_t i = 0; i < numOfCols; ++i) {
     SColumnInfoData* pColInfoData = taosArrayGet(pDataBlock->pDataBlock, i);
     pColInfoData->info = pCols[i].info;
 
@@ -796,31 +817,39 @@ static int32_t* createTupleIndex(size_t rows) {
     return NULL;
   }
 
-  for(int32_t i = 0; i < rows; ++i) {
+  for (int32_t i = 0; i < rows; ++i) {
     index[i] = i;
   }
 
   return index;
 }
 
-static void destroyTupleIndex(int32_t* index) {
-  tfree(index);
-}
+static void destroyTupleIndex(int32_t* index) { tfree(index); }
 
 static __compar_fn_t getComparFn(int32_t type, int32_t order) {
-  switch(type) {
-    case TSDB_DATA_TYPE_TINYINT:  return order == TSDB_ORDER_ASC? compareInt8Val:compareInt8ValDesc;
-    case TSDB_DATA_TYPE_SMALLINT: return order == TSDB_ORDER_ASC? compareInt16Val:compareInt16ValDesc;
-    case TSDB_DATA_TYPE_INT:      return order == TSDB_ORDER_ASC? compareInt32Val:compareInt32ValDesc;
-    case TSDB_DATA_TYPE_BIGINT:   return order == TSDB_ORDER_ASC? compareInt64Val:compareInt64ValDesc;
-    case TSDB_DATA_TYPE_FLOAT:    return order == TSDB_ORDER_ASC? compareFloatVal:compareFloatValDesc;
-    case TSDB_DATA_TYPE_DOUBLE:   return order == TSDB_ORDER_ASC? compareDoubleVal:compareDoubleValDesc;
-    case TSDB_DATA_TYPE_UTINYINT: return order == TSDB_ORDER_ASC? compareUint8Val:compareUint8ValDesc;
-    case TSDB_DATA_TYPE_USMALLINT:return order == TSDB_ORDER_ASC? compareUint16Val:compareUint16ValDesc;
-    case TSDB_DATA_TYPE_UINT:     return order == TSDB_ORDER_ASC? compareUint32Val:compareUint32ValDesc;
-    case TSDB_DATA_TYPE_UBIGINT:  return order == TSDB_ORDER_ASC? compareUint64Val:compareUint64ValDesc;
+  switch (type) {
+    case TSDB_DATA_TYPE_TINYINT:
+      return order == TSDB_ORDER_ASC ? compareInt8Val : compareInt8ValDesc;
+    case TSDB_DATA_TYPE_SMALLINT:
+      return order == TSDB_ORDER_ASC ? compareInt16Val : compareInt16ValDesc;
+    case TSDB_DATA_TYPE_INT:
+      return order == TSDB_ORDER_ASC ? compareInt32Val : compareInt32ValDesc;
+    case TSDB_DATA_TYPE_BIGINT:
+      return order == TSDB_ORDER_ASC ? compareInt64Val : compareInt64ValDesc;
+    case TSDB_DATA_TYPE_FLOAT:
+      return order == TSDB_ORDER_ASC ? compareFloatVal : compareFloatValDesc;
+    case TSDB_DATA_TYPE_DOUBLE:
+      return order == TSDB_ORDER_ASC ? compareDoubleVal : compareDoubleValDesc;
+    case TSDB_DATA_TYPE_UTINYINT:
+      return order == TSDB_ORDER_ASC ? compareUint8Val : compareUint8ValDesc;
+    case TSDB_DATA_TYPE_USMALLINT:
+      return order == TSDB_ORDER_ASC ? compareUint16Val : compareUint16ValDesc;
+    case TSDB_DATA_TYPE_UINT:
+      return order == TSDB_ORDER_ASC ? compareUint32Val : compareUint32ValDesc;
+    case TSDB_DATA_TYPE_UBIGINT:
+      return order == TSDB_ORDER_ASC ? compareUint64Val : compareUint64ValDesc;
     default:
-      return order == TSDB_ORDER_ASC? compareInt32Val:compareInt32ValDesc;
+      return order == TSDB_ORDER_ASC ? compareInt32Val : compareInt32ValDesc;
   }
 }
 
@@ -865,10 +894,8 @@ int32_t blockDataSort(SSDataBlock* pDataBlock, SArray* pOrderInfo, bool nullFirs
 
         return TSDB_CODE_SUCCESS;
       } else {  // var data type
-
       }
     } else if (pDataBlock->info.numOfCols == 2) {
-
     }
   }
 
@@ -881,7 +908,7 @@ int32_t blockDataSort(SSDataBlock* pDataBlock, SArray* pOrderInfo, bool nullFirs
   int64_t p0 = taosGetTimestampUs();
 
   SSDataBlockSortHelper helper = {.nullFirst = nullFirst, .pDataBlock = pDataBlock, .orderInfo = pOrderInfo};
-  for(int32_t i = 0; i < taosArrayGetSize(helper.orderInfo); ++i) {
+  for (int32_t i = 0; i < taosArrayGetSize(helper.orderInfo); ++i) {
     struct SBlockOrderInfo* pInfo = taosArrayGet(helper.orderInfo, i);
     pInfo->pColData = taosArrayGet(pDataBlock->pDataBlock, pInfo->colIndex);
   }
@@ -909,7 +936,8 @@ int32_t blockDataSort(SSDataBlock* pDataBlock, SArray* pOrderInfo, bool nullFirs
   copyBackToBlock(pDataBlock, pCols);
   int64_t p4 = taosGetTimestampUs();
 
-  printf("sort:%" PRId64 ", create:%" PRId64 ", assign:%" PRId64 ", copyback:%" PRId64 ", rows:%d\n", p1-p0, p2 - p1, p3 - p2, p4-p3, rows);
+  printf("sort:%" PRId64 ", create:%" PRId64 ", assign:%" PRId64 ", copyback:%" PRId64 ", rows:%d\n", p1 - p0, p2 - p1,
+         p3 - p2, p4 - p3, rows);
   destroyTupleIndex(index);
 
   return TSDB_CODE_SUCCESS;
@@ -917,14 +945,18 @@ int32_t blockDataSort(SSDataBlock* pDataBlock, SArray* pOrderInfo, bool nullFirs
 
 typedef struct SHelper {
   int32_t index;
-  union {char *pData; int64_t i64; double d64;};
+  union {
+    char*   pData;
+    int64_t i64;
+    double  d64;
+  };
 } SHelper;
 
 SHelper* createTupleIndex_rv(int32_t numOfRows, SArray* pOrderInfo, SSDataBlock* pBlock) {
   int32_t sortValLengthPerRow = 0;
   int32_t numOfCols = taosArrayGetSize(pOrderInfo);
 
-  for(int32_t i = 0; i < numOfCols; ++i) {
+  for (int32_t i = 0; i < numOfCols; ++i) {
     SBlockOrderInfo* pInfo = taosArrayGet(pOrderInfo, i);
     SColumnInfoData* pColInfo = taosArrayGet(pBlock->pDataBlock, pInfo->colIndex);
     pInfo->pColData = pColInfo;
@@ -933,19 +965,20 @@ SHelper* createTupleIndex_rv(int32_t numOfRows, SArray* pOrderInfo, SSDataBlock*
 
   size_t len = sortValLengthPerRow * pBlock->info.rows;
 
-  char* buf = calloc(1, len);
+  char*    buf = calloc(1, len);
   SHelper* phelper = calloc(numOfRows, sizeof(SHelper));
-  for(int32_t i = 0; i < numOfRows; ++i) {
+  for (int32_t i = 0; i < numOfRows; ++i) {
     phelper[i].index = i;
     phelper[i].pData = buf + sortValLengthPerRow * i;
   }
 
   int32_t offset = 0;
-  for(int32_t i = 0; i < numOfCols; ++i) {
+  for (int32_t i = 0; i < numOfCols; ++i) {
     SBlockOrderInfo* pInfo = taosArrayGet(pOrderInfo, i);
-    for(int32_t j = 0; j < numOfRows; ++j) {
-      phelper[j].i64 = *(int32_t*) pInfo->pColData->pData + pInfo->pColData->info.bytes * j;
-//      memcpy(phelper[j].pData + offset, pInfo->pColData->pData + pInfo->pColData->info.bytes * j, pInfo->pColData->info.bytes);
+    for (int32_t j = 0; j < numOfRows; ++j) {
+      phelper[j].i64 = *(int32_t*)pInfo->pColData->pData + pInfo->pColData->info.bytes * j;
+      //      memcpy(phelper[j].pData + offset, pInfo->pColData->pData + pInfo->pColData->info.bytes * j,
+      //      pInfo->pColData->info.bytes);
     }
 
     offset += pInfo->pColData->info.bytes;
@@ -955,70 +988,68 @@ SHelper* createTupleIndex_rv(int32_t numOfRows, SArray* pOrderInfo, SSDataBlock*
 }
 
 int32_t dataBlockCompar_rv(const void* p1, const void* p2, const void* param) {
-  const SSDataBlockSortHelper* pHelper = (const SSDataBlockSortHelper*) param;
+  const SSDataBlockSortHelper* pHelper = (const SSDataBlockSortHelper*)param;
 
-//  SSDataBlock* pDataBlock = pHelper->pDataBlock;
+  //  SSDataBlock* pDataBlock = pHelper->pDataBlock;
 
-  SHelper* left  = (SHelper*) p1;
-  SHelper* right = (SHelper*) p2;
+  SHelper* left = (SHelper*)p1;
+  SHelper* right = (SHelper*)p2;
 
   SArray* pInfo = pHelper->orderInfo;
 
   int32_t offset = 0;
-//  for(int32_t i = 0; i < pInfo->size; ++i) {
-//    SBlockOrderInfo* pOrder = TARRAY_GET_ELEM(pInfo, 0);
-//    SColumnInfoData* pColInfoData = pOrder->pColData;//TARRAY_GET_ELEM(pDataBlock->pDataBlock, pOrder->colIndex);
+  //  for(int32_t i = 0; i < pInfo->size; ++i) {
+  //    SBlockOrderInfo* pOrder = TARRAY_GET_ELEM(pInfo, 0);
+  //    SColumnInfoData* pColInfoData = pOrder->pColData;//TARRAY_GET_ELEM(pDataBlock->pDataBlock, pOrder->colIndex);
 
-//    if (pColInfoData->hasNull) {
-//      bool leftNull  = colDataIsNull(pColInfoData, pDataBlock->info.rows, left, pDataBlock->pBlockAgg);
-//      bool rightNull = colDataIsNull(pColInfoData, pDataBlock->info.rows, right, pDataBlock->pBlockAgg);
-//      if (leftNull && rightNull) {
-//        continue; // continue to next slot
-//      }
-//
-//      if (rightNull) {
-//        return pHelper->nullFirst? 1:-1;
-//      }
-//
-//      if (leftNull) {
-//        return pHelper->nullFirst? -1:1;
-//      }
-//    }
+  //    if (pColInfoData->hasNull) {
+  //      bool leftNull  = colDataIsNull(pColInfoData, pDataBlock->info.rows, left, pDataBlock->pBlockAgg);
+  //      bool rightNull = colDataIsNull(pColInfoData, pDataBlock->info.rows, right, pDataBlock->pBlockAgg);
+  //      if (leftNull && rightNull) {
+  //        continue; // continue to next slot
+  //      }
+  //
+  //      if (rightNull) {
+  //        return pHelper->nullFirst? 1:-1;
+  //      }
+  //
+  //      if (leftNull) {
+  //        return pHelper->nullFirst? -1:1;
+  //      }
+  //    }
 
-//    void* left1  = colDataGetData(pColInfoData, left);
-//    void* right1 = colDataGetData(pColInfoData, right);
+  //    void* left1  = colDataGetData(pColInfoData, left);
+  //    void* right1 = colDataGetData(pColInfoData, right);
 
-//    switch(pColInfoData->info.type) {
-//      case TSDB_DATA_TYPE_INT: {
-        int32_t leftx  = *(int32_t*)left->pData;//*(int32_t*)(left->pData + offset);
-        int32_t rightx = *(int32_t*)right->pData;//*(int32_t*)(right->pData + offset);
+  //    switch(pColInfoData->info.type) {
+  //      case TSDB_DATA_TYPE_INT: {
+  int32_t leftx = *(int32_t*)left->pData;    //*(int32_t*)(left->pData + offset);
+  int32_t rightx = *(int32_t*)right->pData;  //*(int32_t*)(right->pData + offset);
 
-//        offset += pColInfoData->info.bytes;
-        if (leftx == rightx) {
-//          break;
-          return 0;
-        } else {
-//          if (pOrder->order == TSDB_ORDER_ASC) {
-            return (leftx < rightx)? -1:1;
-//          } else {
-//            return (leftx < rightx)? 1:-1;
-//          }
-        }
-//      }
-//      default:
-//        assert(0);
-//    }
-//  }
+  //        offset += pColInfoData->info.bytes;
+  if (leftx == rightx) {
+    //          break;
+    return 0;
+  } else {
+    //          if (pOrder->order == TSDB_ORDER_ASC) {
+    return (leftx < rightx) ? -1 : 1;
+    //          } else {
+    //            return (leftx < rightx)? 1:-1;
+    //          }
+  }
+  //      }
+  //      default:
+  //        assert(0);
+  //    }
+  //  }
 
   return 0;
 }
 
-int32_t varColSort(SColumnInfoData* pColumnInfoData, SBlockOrderInfo* pOrder) {
-    return 0;
-}
+int32_t varColSort(SColumnInfoData* pColumnInfoData, SBlockOrderInfo* pOrder) { return 0; }
 
 int32_t blockDataSort_rv(SSDataBlock* pDataBlock, SArray* pOrderInfo, bool nullFirst) {
-// Allocate the additional buffer.
+  // Allocate the additional buffer.
   int64_t p0 = taosGetTimestampUs();
 
   SSDataBlockSortHelper helper = {.nullFirst = nullFirst, .pDataBlock = pDataBlock, .orderInfo = pOrderInfo};
@@ -1032,7 +1063,7 @@ int32_t blockDataSort_rv(SSDataBlock* pDataBlock, SArray* pOrderInfo, bool nullF
 
   taosqsort(index, rows, sizeof(SHelper), &helper, dataBlockCompar_rv);
 
-  int64_t p1 = taosGetTimestampUs();
+  int64_t          p1 = taosGetTimestampUs();
   SColumnInfoData* pCols = createHelpColInfoData(pDataBlock);
   if (pCols == NULL) {
     terrno = TSDB_CODE_OUT_OF_MEMORY;
@@ -1052,7 +1083,8 @@ int32_t blockDataSort_rv(SSDataBlock* pDataBlock, SArray* pOrderInfo, bool nullF
   copyBackToBlock(pDataBlock, pCols);
   int64_t p4 = taosGetTimestampUs();
 
-  printf("sort:%" PRId64 ", create:%" PRId64", assign:%" PRId64 ", copyback:%" PRId64 ", rows:%d\n", p1 - p0, p2 - p1, p3 - p2, p4 - p3, rows);
+  printf("sort:%" PRId64 ", create:%" PRId64 ", assign:%" PRId64 ", copyback:%" PRId64 ", rows:%d\n", p1 - p0, p2 - p1,
+         p3 - p2, p4 - p3, rows);
   //  destroyTupleIndex(index);
   return 0;
 }
@@ -1110,8 +1142,8 @@ int32_t blockDataEnsureColumnCapacity(SColumnInfoData* pColumn, uint32_t numOfRo
 
 int32_t blockDataEnsureCapacity(SSDataBlock* pDataBlock, uint32_t numOfRows) {
   int32_t code = 0;
-  
-  for(int32_t i = 0; i < pDataBlock->info.numOfCols; ++i) {
+
+  for (int32_t i = 0; i < pDataBlock->info.numOfCols; ++i) {
     SColumnInfoData* p = taosArrayGet(pDataBlock->pDataBlock, i);
     code = blockDataEnsureColumnCapacity(p, numOfRows);
     if (code) {
@@ -1127,21 +1159,8 @@ void* blockDataDestroy(SSDataBlock* pBlock) {
     return NULL;
   }
 
-  int32_t numOfOutput = pBlock->info.numOfCols;
-  for(int32_t i = 0; i < numOfOutput; ++i) {
-    SColumnInfoData* pColInfoData = taosArrayGet(pBlock->pDataBlock, i);
-    if (IS_VAR_DATA_TYPE(pColInfoData->info.type)) {
-      tfree(pColInfoData->varmeta.offset);
-    } else {
-      tfree(pColInfoData->nullbitmap);
-    }
-
-    tfree(pColInfoData->pData);
-  }
-
-  taosArrayDestroy(pBlock->pDataBlock);
-  tfree(pBlock->pBlockAgg);
-  // tfree(pBlock);
+  blockDestroyInner(pBlock);
+  tfree(pBlock);
   return NULL;
 }
 
@@ -1154,7 +1173,7 @@ SSDataBlock* createOneDataBlock(const SSDataBlock* pDataBlock) {
   pBlock->info.numOfCols = numOfCols;
   pBlock->info.hasVarCol = pDataBlock->info.hasVarCol;
 
-  for(int32_t i = 0; i < numOfCols; ++i) {
+  for (int32_t i = 0; i < numOfCols; ++i) {
     SColumnInfoData  colInfo = {0};
     SColumnInfoData* p = taosArrayGet(pDataBlock->pDataBlock, i);
     colInfo.info = p->info;
@@ -1230,4 +1249,4 @@ void* tDecodeDataBlock(const void* buf, SSDataBlock* pBlock) {
     taosArrayPush(pBlock->pDataBlock, &data);
   }
   return (void*)buf;
-}
\ No newline at end of file
+}
diff --git a/source/common/src/tmsg.c b/source/common/src/tmsg.c
index a2382d1dd1..b4caf5ba97 100644
--- a/source/common/src/tmsg.c
+++ b/source/common/src/tmsg.c
@@ -2720,6 +2720,8 @@ int32_t tEncodeSStreamTask(SCoder *pEncoder, const SStreamTask *pTask) {
   if (tEncodeI32(pEncoder, pTask->taskId) < 0) return -1;
   if (tEncodeI32(pEncoder, pTask->level) < 0) return -1;
   if (tEncodeI8(pEncoder, pTask->status) < 0) return -1;
+  if (tEncodeI8(pEncoder, pTask->pipeEnd) < 0) return -1;
+  if (tEncodeI8(pEncoder, pTask->parallel) < 0) return -1;
   if (tEncodeSEpSet(pEncoder, &pTask->NextOpEp) < 0) return -1;
   if (tEncodeCStr(pEncoder, pTask->qmsg) < 0) return -1;
   tEndEncode(pEncoder);
@@ -2732,6 +2734,8 @@ int32_t tDecodeSStreamTask(SCoder *pDecoder, SStreamTask *pTask) {
   if (tDecodeI32(pDecoder, &pTask->taskId) < 0) return -1;
   if (tDecodeI32(pDecoder, &pTask->level) < 0) return -1;
   if (tDecodeI8(pDecoder, &pTask->status) < 0) return -1;
+  if (tDecodeI8(pDecoder, &pTask->pipeEnd) < 0) return -1;
+  if (tDecodeI8(pDecoder, &pTask->parallel) < 0) return -1;
   if (tDecodeSEpSet(pDecoder, &pTask->NextOpEp) < 0) return -1;
   if (tDecodeCStrAlloc(pDecoder, &pTask->qmsg) < 0) return -1;
   tEndDecode(pDecoder);
diff --git a/source/dnode/mgmt/vnode/src/vmWorker.c b/source/dnode/mgmt/vnode/src/vmWorker.c
index 52ee624103..d5f0a9d20e 100644
--- a/source/dnode/mgmt/vnode/src/vmWorker.c
+++ b/source/dnode/mgmt/vnode/src/vmWorker.c
@@ -77,8 +77,9 @@ static void vmProcessFetchQueue(SVnodeObj *pVnode, SNodeMsg *pMsg) {
   }
 
   dTrace("msg:%p, is freed, result:0x%04x:%s", pMsg, code & 0XFFFF, tstrerror(code));
-  rpcFreeCont(pMsg->rpcMsg.pCont);
-  taosFreeQitem(pMsg);
+  // TODO: handle invalid write
+  /*rpcFreeCont(pMsg->rpcMsg.pCont);*/
+  /*taosFreeQitem(pMsg);*/
 }
 
 static void vmProcessWriteQueue(SVnodeObj *pVnode, STaosQall *qall, int32_t numOfMsgs) {
diff --git a/source/dnode/mnode/impl/src/mndStream.c b/source/dnode/mnode/impl/src/mndStream.c
index e7cc12bb96..99aded0292 100644
--- a/source/dnode/mnode/impl/src/mndStream.c
+++ b/source/dnode/mnode/impl/src/mndStream.c
@@ -166,7 +166,7 @@ static int32_t mndStreamActionDelete(SSdb *pSdb, SStreamObj *pStream) {
 
 static int32_t mndStreamActionUpdate(SSdb *pSdb, SStreamObj *pOldStream, SStreamObj *pNewStream) {
   mTrace("stream:%s, perform update action", pOldStream->name);
-  atomic_exchange_32(&pOldStream->updateTime, pNewStream->updateTime);
+  atomic_exchange_64(&pOldStream->updateTime, pNewStream->updateTime);
   atomic_exchange_32(&pOldStream->version, pNewStream->version);
 
   taosWLockLatch(&pOldStream->lock);
diff --git a/source/dnode/mnode/impl/src/mndSubscribe.c b/source/dnode/mnode/impl/src/mndSubscribe.c
index 4fbf0352e5..a4909dfe37 100644
--- a/source/dnode/mnode/impl/src/mndSubscribe.c
+++ b/source/dnode/mnode/impl/src/mndSubscribe.c
@@ -272,7 +272,6 @@ static int32_t mndProcessGetSubEpReq(SNodeMsg *pMsg) {
   /*sdbWrite(pMnode->pSdb, pConsumerRaw);*/
 
   strcpy(rsp.cgroup, pReq->cgroup);
-  rsp.consumerId = consumerId;
   if (epoch != pConsumer->epoch) {
     mInfo("send new assignment to consumer, consumer epoch %d, server epoch %d", epoch, pConsumer->epoch);
     SArray *pTopics = pConsumer->currentTopics;
@@ -322,6 +321,7 @@ static int32_t mndProcessGetSubEpReq(SNodeMsg *pMsg) {
   }
   ((SMqRspHead *)buf)->mqMsgType = TMQ_MSG_TYPE__EP_RSP;
   ((SMqRspHead *)buf)->epoch = pConsumer->epoch;
+  ((SMqRspHead *)buf)->consumerId = pConsumer->consumerId;
 
   void *abuf = POINTER_SHIFT(buf, sizeof(SMqRspHead));
   tEncodeSMqCMGetSubEpRsp(&abuf, &rsp);
@@ -337,21 +337,21 @@ static int32_t mndSplitSubscribeKey(const char *key, char *topic, char *cgroup)
   while (key[i] != TMQ_SEPARATOR) {
     i++;
   }
-  memcpy(topic, key, i - 1);
-  topic[i] = 0;
-  strcpy(cgroup, &key[i + 1]);
+  memcpy(cgroup, key, i);
+  cgroup[i] = 0;
+  strcpy(topic, &key[i + 1]);
   return 0;
 }
 
 static SMqRebSubscribe *mndGetOrCreateRebSub(SHashObj *pHash, const char *key) {
-  SMqRebSubscribe *pRebSub = taosHashGet(pHash, key, strlen(key));
+  SMqRebSubscribe *pRebSub = taosHashGet(pHash, key, strlen(key) + 1);
   if (pRebSub == NULL) {
     pRebSub = tNewSMqRebSubscribe(key);
     if (pRebSub == NULL) {
       terrno = TSDB_CODE_OUT_OF_MEMORY;
       return NULL;
     }
-    taosHashPut(pHash, key, strlen(key), pRebSub, sizeof(SMqRebSubscribe));
+    taosHashPut(pHash, key, strlen(key) + 1, pRebSub, sizeof(SMqRebSubscribe));
   }
   return pRebSub;
 }
@@ -441,6 +441,7 @@ static int32_t mndProcessDoRebalanceMsg(SNodeMsg *pMsg) {
     if (pIter == NULL) break;
     SMqRebSubscribe *pRebSub = (SMqRebSubscribe *)pIter;
     SMqSubscribeObj *pSub = mndAcquireSubscribeByKey(pMnode, pRebSub->key);
+    tfree(pRebSub->key);
 
     mInfo("mq rebalance subscription: %s", pSub->key);
 
@@ -503,7 +504,8 @@ static int32_t mndProcessDoRebalanceMsg(SNodeMsg *pMsg) {
             atomic_store_32(&pRebConsumer->status, MQ_CONSUMER_STATUS__IDLE);
           }
 
-          mInfo("mq consumer:%" PRId64 ", status change from %d to %d", pRebConsumer->consumerId, status, pRebConsumer->status);
+          mInfo("mq consumer:%" PRId64 ", status change from %d to %d", pRebConsumer->consumerId, status,
+                pRebConsumer->status);
 
           SSdbRaw *pConsumerRaw = mndConsumerActionEncode(pRebConsumer);
           sdbSetRawStatus(pConsumerRaw, SDB_STATUS_READY);
@@ -537,14 +539,14 @@ static int32_t mndProcessDoRebalanceMsg(SNodeMsg *pMsg) {
               mndSplitSubscribeKey(pSub->key, topic, cgroup);
               SMqTopicObj *pTopic = mndAcquireTopic(pMnode, topic);
 
-              mInfo("mq set conn: assign vgroup %d of topic %s to consumer %" PRId64 "", pConsumerEp->vgId, topic,
-                    pConsumerEp->consumerId);
+              mInfo("mq set conn: assign vgroup %d of topic %s to consumer %" PRId64 " cgroup: %s", pConsumerEp->vgId,
+                    topic, pConsumerEp->consumerId, cgroup);
 
               mndPersistMqSetConnReq(pMnode, pTrans, pTopic, cgroup, pConsumerEp);
               mndReleaseTopic(pMnode, pTopic);
             } else {
-              mInfo("mq rebalance: assign vgroup %d, from consumer %" PRId64 " to consumer %" PRId64 "", pConsumerEp->vgId,
-                    pConsumerEp->oldConsumerId, pConsumerEp->consumerId);
+              mInfo("mq rebalance: assign vgroup %d, from consumer %" PRId64 " to consumer %" PRId64 "",
+                    pConsumerEp->vgId, pConsumerEp->oldConsumerId, pConsumerEp->consumerId);
 
               mndPersistRebalanceMsg(pMnode, pTrans, pConsumerEp);
             }
@@ -1099,7 +1101,8 @@ static int32_t mndProcessSubscribeReq(SNodeMsg *pMsg) {
       SMqSubscribeObj *pSub = mndAcquireSubscribe(pMnode, cgroup, newTopicName);
       bool             createSub = false;
       if (pSub == NULL) {
-        mDebug("create new subscription by consumer %" PRId64 ", group: %s, topic %s", consumerId, cgroup, newTopicName);
+        mDebug("create new subscription by consumer %" PRId64 ", group: %s, topic %s", consumerId, cgroup,
+               newTopicName);
         pSub = mndCreateSubscription(pMnode, pTopic, cgroup);
         createSub = true;
 
diff --git a/source/dnode/mnode/impl/src/mndTopic.c b/source/dnode/mnode/impl/src/mndTopic.c
index e94ae4f8ec..bd0fd0e612 100644
--- a/source/dnode/mnode/impl/src/mndTopic.c
+++ b/source/dnode/mnode/impl/src/mndTopic.c
@@ -176,7 +176,7 @@ static int32_t mndTopicActionDelete(SSdb *pSdb, SMqTopicObj *pTopic) {
 
 static int32_t mndTopicActionUpdate(SSdb *pSdb, SMqTopicObj *pOldTopic, SMqTopicObj *pNewTopic) {
   mTrace("topic:%s, perform update action", pOldTopic->name);
-  atomic_exchange_32(&pOldTopic->updateTime, pNewTopic->updateTime);
+  atomic_exchange_64(&pOldTopic->updateTime, pNewTopic->updateTime);
   atomic_exchange_32(&pOldTopic->version, pNewTopic->version);
 
   taosWLockLatch(&pOldTopic->lock);
@@ -241,12 +241,12 @@ static int32_t mndGetPlanString(SCMCreateTopicReq *pCreate, char **pStr) {
     return TSDB_CODE_SUCCESS;
   }
 
-  SNode* pAst = NULL;
+  SNode  *pAst = NULL;
   int32_t code = nodesStringToNode(pCreate->ast, &pAst);
 
-  SQueryPlan* pPlan = NULL;
+  SQueryPlan *pPlan = NULL;
   if (TSDB_CODE_SUCCESS == code) {
-    SPlanContext cxt = { .pAstRoot = pAst, .topicQuery = true };
+    SPlanContext cxt = {.pAstRoot = pAst, .topicQuery = true};
     code = qCreateQueryPlan(&cxt, &pPlan, NULL);
   }
 
@@ -274,7 +274,7 @@ static int32_t mndCreateTopic(SMnode *pMnode, SNodeMsg *pReq, SCMCreateTopicReq
   topicObj.logicalPlan = "";
   topicObj.sqlLen = strlen(pCreate->sql);
 
-  char* pPlanStr = NULL;
+  char *pPlanStr = NULL;
   if (TSDB_CODE_SUCCESS != mndGetPlanString(pCreate, &pPlanStr)) {
     mError("topic:%s, failed to get plan since %s", pCreate->name, terrstr());
     return -1;
diff --git a/source/dnode/snode/src/snode.c b/source/dnode/snode/src/snode.c
index a29baca3bd..147f44b260 100644
--- a/source/dnode/snode/src/snode.c
+++ b/source/dnode/snode/src/snode.c
@@ -57,7 +57,9 @@ void sndMetaDelete(SStreamMeta *pMeta) {
 }
 
 int32_t sndMetaDeployTask(SStreamMeta *pMeta, SStreamTask *pTask) {
-  pTask->runner.executor = qCreateStreamExecTaskInfo(pTask->qmsg, NULL);
+  for (int i = 0; i < pTask->parallel; i++) {
+    pTask->runner.executor[i] = qCreateStreamExecTaskInfo(pTask->qmsg, NULL);
+  }
   return taosHashPut(pMeta->pHash, &pTask->taskId, sizeof(int32_t), pTask, sizeof(void *));
 }
 
@@ -95,6 +97,7 @@ void sndProcessUMsg(SSnode *pSnode, SRpcMsg *pMsg) {
     SStreamTask *pTask = malloc(sizeof(SStreamTask));
     if (pTask == NULL) {
       ASSERT(0);
+      return;
     }
     SCoder decoder;
     tCoderInit(&decoder, TD_LITTLE_ENDIAN, msg, pMsg->contLen - sizeof(SMsgHead), TD_DECODER);
diff --git a/source/dnode/vnode/src/tq/tq.c b/source/dnode/vnode/src/tq/tq.c
index b94e7d7c03..87c670a4e9 100644
--- a/source/dnode/vnode/src/tq/tq.c
+++ b/source/dnode/vnode/src/tq/tq.c
@@ -245,7 +245,7 @@ int32_t tqProcessPollReq(STQ* pTq, SRpcMsg* pMsg) {
   }
 
   SMqPollRsp rsp = {
-      .consumerId = consumerId,
+      /*.consumerId = consumerId,*/
       .numOfTopics = 0,
       .pBlockData = NULL,
   };
@@ -298,7 +298,7 @@ int32_t tqProcessPollReq(STQ* pTq, SRpcMsg* pMsg) {
         }
 
         taosArrayPush(pRes, pDataBlock);
-        rsp.schemas = pTopic->buffer.output[pos].pReadHandle->pSchemaWrapper;
+        rsp.schema = pTopic->buffer.output[pos].pReadHandle->pSchemaWrapper;
         rsp.rspOffset = fetchOffset;
 
         rsp.numOfTopics = 1;
@@ -312,10 +312,11 @@ int32_t tqProcessPollReq(STQ* pTq, SRpcMsg* pMsg) {
         }
         ((SMqRspHead*)buf)->mqMsgType = TMQ_MSG_TYPE__POLL_RSP;
         ((SMqRspHead*)buf)->epoch = pReq->epoch;
+        ((SMqRspHead*)buf)->consumerId = consumerId;
 
         void* abuf = POINTER_SHIFT(buf, sizeof(SMqRspHead));
         tEncodeSMqPollRsp(&abuf, &rsp);
-        taosArrayDestroyEx(rsp.pBlockData, (void (*)(void*))tDeleteSSDataBlock);
+        /*taosArrayDestroyEx(rsp.pBlockData, (void (*)(void*))tDeleteSSDataBlock);*/
         pMsg->pCont = buf;
         pMsg->contLen = tlen;
         pMsg->code = 0;
diff --git a/source/dnode/vnode/src/vnd/vnodeInt.c b/source/dnode/vnode/src/vnd/vnodeInt.c
index 7d0b594e95..a64e834ff8 100644
--- a/source/dnode/vnode/src/vnd/vnodeInt.c
+++ b/source/dnode/vnode/src/vnd/vnodeInt.c
@@ -14,8 +14,8 @@
  */
 
 #define _DEFAULT_SOURCE
-#include "vnd.h"
 #include "sync.h"
+#include "vnd.h"
 // #include "vnodeInt.h"
 
 int32_t vnodeAlter(SVnode *pVnode, const SVnodeCfg *pCfg) { return 0; }
@@ -41,6 +41,6 @@ int32_t vnodeGetLoad(SVnode *pVnode, SVnodeLoad *pLoad) {
 }
 
 int vnodeProcessSyncReq(SVnode *pVnode, SRpcMsg *pMsg, SRpcMsg **pRsp) {
-  vInfo("sync message is processed");
+  /*vInfo("sync message is processed");*/
   return 0;
 }
diff --git a/source/libs/catalog/inc/catalogInt.h b/source/libs/catalog/inc/catalogInt.h
index 60c5886250..88a7662a76 100644
--- a/source/libs/catalog/inc/catalogInt.h
+++ b/source/libs/catalog/inc/catalogInt.h
@@ -278,7 +278,7 @@ typedef struct SCtgAction {
 #define CTG_ERR_JRET(c) do { code = c; if (code != TSDB_CODE_SUCCESS) { terrno = code; goto _return; } } while (0)
 
 #define CTG_API_LEAVE(c) do { int32_t __code = c; CTG_UNLOCK(CTG_READ, &gCtgMgmt.lock); CTG_API_DEBUG("CTG API leave %s", __FUNCTION__); CTG_RET(__code); } while (0)
-#define CTG_API_ENTER() do { CTG_API_DEBUG("CTG API enter %s", __FUNCTION__); CTG_LOCK(CTG_READ, &gCtgMgmt.lock); if (atomic_load_8(&gCtgMgmt.exit)) { CTG_API_LEAVE(TSDB_CODE_CTG_OUT_OF_SERVICE); }  } while (0)
+#define CTG_API_ENTER() do { CTG_API_DEBUG("CTG API enter %s", __FUNCTION__); CTG_LOCK(CTG_READ, &gCtgMgmt.lock); if (atomic_load_8((int8_t*)&gCtgMgmt.exit)) { CTG_API_LEAVE(TSDB_CODE_CTG_OUT_OF_SERVICE); }  } while (0)
 
 
 
diff --git a/source/libs/catalog/src/catalog.c b/source/libs/catalog/src/catalog.c
index fbb2903a8e..3209b1d80f 100644
--- a/source/libs/catalog/src/catalog.c
+++ b/source/libs/catalog/src/catalog.c
@@ -2044,7 +2044,7 @@ void* ctgUpdateThreadFunc(void* param) {
   while (true) {
     tsem_wait(&gCtgMgmt.sem);
     
-    if (atomic_load_8(&gCtgMgmt.exit)) {
+    if (atomic_load_8((int8_t*)&gCtgMgmt.exit)) {
       break;
     }
 
@@ -2090,7 +2090,7 @@ int32_t catalogInit(SCatalogCfg *cfg) {
     CTG_ERR_RET(TSDB_CODE_CTG_INVALID_INPUT);
   }
 
-  atomic_store_8(&gCtgMgmt.exit, false);
+  atomic_store_8((int8_t*)&gCtgMgmt.exit, false);
 
   if (cfg) {
     memcpy(&gCtgMgmt.cfg, cfg, sizeof(*cfg));
@@ -2705,11 +2705,11 @@ int32_t catalogGetExpiredDBs(SCatalog* pCtg, SDbVgVersion **dbs, uint32_t *num)
 void catalogDestroy(void) {
   qInfo("start to destroy catalog");
   
-  if (NULL == gCtgMgmt.pCluster || atomic_load_8(&gCtgMgmt.exit)) {
+  if (NULL == gCtgMgmt.pCluster || atomic_load_8((int8_t*)&gCtgMgmt.exit)) {
     return;
   }
 
-  atomic_store_8(&gCtgMgmt.exit, true);
+  atomic_store_8((int8_t*)&gCtgMgmt.exit, true);
 
   tsem_post(&gCtgMgmt.sem);
 
diff --git a/source/libs/executor/src/tsimplehash.c b/source/libs/executor/src/tsimplehash.c
index 4012c3926e..65067f58e3 100644
--- a/source/libs/executor/src/tsimplehash.c
+++ b/source/libs/executor/src/tsimplehash.c
@@ -40,7 +40,7 @@ typedef struct SHNode {
 typedef struct SSHashObj {
   SHNode         **hashList;
   size_t           capacity;     // number of slots
-  size_t           size;         // number of elements in hash table
+  int64_t          size;         // number of elements in hash table
   _hash_fn_t       hashFp;       // hash function
   _equal_fn_t      equalFp;      // equal function
   int32_t          keyLen;
@@ -91,7 +91,7 @@ int32_t tSimpleHashGetSize(const SSHashObj *pHashObj) {
   if (pHashObj == NULL) {
     return 0;
   }
-  return (int32_t)atomic_load_64(&pHashObj->size);
+  return (int32_t)atomic_load_64((int64_t*)&pHashObj->size);
 }
 
 static SHNode *doCreateHashNode(const void *key, size_t keyLen, const void *pData, size_t dsize, uint32_t hashVal) {
diff --git a/source/libs/function/src/functionMgt.c b/source/libs/function/src/functionMgt.c
index f9f2e90770..10ac8bbf43 100644
--- a/source/libs/function/src/functionMgt.c
+++ b/source/libs/function/src/functionMgt.c
@@ -97,7 +97,7 @@ bool fmIsAggFunc(int32_t funcId) {
 
 void fmFuncMgtDestroy() {
   void* m = gFunMgtService.pFuncNameHashTable;
-  if (m != NULL && atomic_val_compare_exchange_ptr(&gFunMgtService.pFuncNameHashTable, m, 0) == m) {
+  if (m != NULL && atomic_val_compare_exchange_ptr((void**)&gFunMgtService.pFuncNameHashTable, m, 0) == m) {
     taosHashCleanup(m);
   }
 }
diff --git a/source/libs/parser/test/mockCatalogService.cpp b/source/libs/parser/test/mockCatalogService.cpp
index 1f333e49e7..9abaf5b14f 100644
--- a/source/libs/parser/test/mockCatalogService.cpp
+++ b/source/libs/parser/test/mockCatalogService.cpp
@@ -16,8 +16,8 @@
 #include <iomanip>
 #include <iostream>
 #include <map>
-#include "tdatablock.h"
 #include "mockCatalogService.h"
+#include "tdatablock.h"
 
 #include "tname.h"
 #include "ttypes.h"
diff --git a/source/libs/qworker/test/qworkerTests.cpp b/source/libs/qworker/test/qworkerTests.cpp
index 30c0e2fca2..7d4a6d10db 100644
--- a/source/libs/qworker/test/qworkerTests.cpp
+++ b/source/libs/qworker/test/qworkerTests.cpp
@@ -50,7 +50,7 @@ namespace {
 int32_t qwtTestMaxExecTaskUsec = 2;
 int32_t qwtTestReqMaxDelayUsec = 2;
 
-uint64_t qwtTestQueryId = 0;
+int64_t qwtTestQueryId = 0;
 bool qwtTestEnableSleep = true;
 bool qwtTestStop = false;
 bool qwtTestDeadLoop = false;
diff --git a/source/libs/sync/inc/syncIO.h b/source/libs/sync/inc/syncIO.h
index 0185d5f211..99f9deb99e 100644
--- a/source/libs/sync/inc/syncIO.h
+++ b/source/libs/sync/inc/syncIO.h
@@ -35,7 +35,7 @@ extern "C" {
 typedef struct SSyncIO {
   STaosQueue *pMsgQ;
   STaosQset * pQset;
-  TdThread   consumerTid;
+  TdThread    consumerTid;
 
   void * serverRpc;
   void * clientRpc;
diff --git a/source/libs/sync/inc/syncInt.h b/source/libs/sync/inc/syncInt.h
index ee009d6428..eaf4d56f67 100644
--- a/source/libs/sync/inc/syncInt.h
+++ b/source/libs/sync/inc/syncInt.h
@@ -217,10 +217,12 @@ int32_t syncNodeStartHeartbeatTimer(SSyncNode* pSyncNode);
 int32_t syncNodeStopHeartbeatTimer(SSyncNode* pSyncNode);
 
 // utils --------------
-int32_t syncNodeSendMsgById(const SRaftId* destRaftId, SSyncNode* pSyncNode, SRpcMsg* pMsg);
-int32_t syncNodeSendMsgByInfo(const SNodeInfo* nodeInfo, SSyncNode* pSyncNode, SRpcMsg* pMsg);
-cJSON*  syncNode2Json(const SSyncNode* pSyncNode);
-char*   syncNode2Str(const SSyncNode* pSyncNode);
+int32_t    syncNodeSendMsgById(const SRaftId* destRaftId, SSyncNode* pSyncNode, SRpcMsg* pMsg);
+int32_t    syncNodeSendMsgByInfo(const SNodeInfo* nodeInfo, SSyncNode* pSyncNode, SRpcMsg* pMsg);
+cJSON*     syncNode2Json(const SSyncNode* pSyncNode);
+char*      syncNode2Str(const SSyncNode* pSyncNode);
+SSyncNode* syncNodeAcquire(int64_t rid);
+void       syncNodeRelease(SSyncNode* pNode);
 
 // raft state change --------------
 void syncNodeUpdateTerm(SSyncNode* pSyncNode, SyncTerm term);
diff --git a/source/libs/sync/src/syncMain.c b/source/libs/sync/src/syncMain.c
index c4adb4762b..facb0180f9 100644
--- a/source/libs/sync/src/syncMain.c
+++ b/source/libs/sync/src/syncMain.c
@@ -575,6 +575,17 @@ char* syncNode2Str(const SSyncNode* pSyncNode) {
   return serialized;
 }
 
+SSyncNode* syncNodeAcquire(int64_t rid) {
+  SSyncNode* pNode = taosAcquireRef(tsNodeRefId, rid);
+  if (pNode == NULL) {
+    sTrace("failed to acquire node from refId:%" PRId64, rid);
+  }
+
+  return pNode;
+}
+
+void syncNodeRelease(SSyncNode* pNode) { taosReleaseRef(tsNodeRefId, pNode->rid); }
+
 // raft state change --------------
 void syncNodeUpdateTerm(SSyncNode* pSyncNode, SyncTerm term) {
   if (term > pSyncNode->pRaftStore->currentTerm) {
diff --git a/source/libs/sync/test/CMakeLists.txt b/source/libs/sync/test/CMakeLists.txt
index 3fcb9087ab..f02ba29218 100644
--- a/source/libs/sync/test/CMakeLists.txt
+++ b/source/libs/sync/test/CMakeLists.txt
@@ -28,6 +28,7 @@ add_executable(syncRpcMsgTest "")
 add_executable(syncPingTimerTest2 "")
 add_executable(syncPingSelfTest "")
 add_executable(syncElectTest "")
+add_executable(syncElectTest2 "")
 add_executable(syncEncodeTest "")
 add_executable(syncWriteTest "")
 add_executable(syncReplicateTest "")
@@ -154,6 +155,10 @@ target_sources(syncElectTest
     PRIVATE
     "syncElectTest.cpp"
 )
+target_sources(syncElectTest2
+    PRIVATE
+    "syncElectTest2.cpp"
+)
 target_sources(syncEncodeTest
     PRIVATE
     "syncEncodeTest.cpp"
@@ -322,7 +327,7 @@ target_include_directories(syncElectTest
     "${CMAKE_SOURCE_DIR}/include/libs/sync"
     "${CMAKE_CURRENT_SOURCE_DIR}/../inc"
 )
-target_include_directories(syncElectTest
+target_include_directories(syncElectTest2
     PUBLIC
     "${CMAKE_SOURCE_DIR}/include/libs/sync"
     "${CMAKE_CURRENT_SOURCE_DIR}/../inc"
@@ -469,6 +474,10 @@ target_link_libraries(syncElectTest
     sync
     gtest_main
 )
+target_link_libraries(syncElectTest2
+    sync
+    gtest_main
+)
 target_link_libraries(syncEncodeTest
     sync
     gtest_main
diff --git a/source/libs/sync/test/syncElectTest.cpp b/source/libs/sync/test/syncElectTest.cpp
index 47f55aa660..99cfe75867 100644
--- a/source/libs/sync/test/syncElectTest.cpp
+++ b/source/libs/sync/test/syncElectTest.cpp
@@ -120,6 +120,7 @@ int main(int argc, char** argv) {
         "elect sleep, state: %d, %s, term:%lu electTimerLogicClock:%lu, electTimerLogicClockUser:%lu, electTimerMS:%d",
         gSyncNode->state, syncUtilState2String(gSyncNode->state), gSyncNode->pRaftStore->currentTerm,
         gSyncNode->electTimerLogicClock, gSyncNode->electTimerLogicClockUser, gSyncNode->electTimerMS);
+    taosMsleep(1000);
   }
 
   return 0;
diff --git a/source/libs/sync/test/syncElectTest2.cpp b/source/libs/sync/test/syncElectTest2.cpp
new file mode 100644
index 0000000000..8463ffd448
--- /dev/null
+++ b/source/libs/sync/test/syncElectTest2.cpp
@@ -0,0 +1,135 @@
+#include <gtest/gtest.h>
+#include <stdio.h>
+#include "syncEnv.h"
+#include "syncIO.h"
+#include "syncInt.h"
+#include "syncRaftLog.h"
+#include "syncRaftStore.h"
+#include "syncUtil.h"
+#include "tref.h"
+
+void logTest() {
+  sTrace("--- sync log test: trace");
+  sDebug("--- sync log test: debug");
+  sInfo("--- sync log test: info");
+  sWarn("--- sync log test: warn");
+  sError("--- sync log test: error");
+  sFatal("--- sync log test: fatal");
+}
+
+uint16_t ports[] = {7010, 7110, 7210, 7310, 7410};
+int32_t  replicaNum = 3;
+int32_t  myIndex = 0;
+
+SRaftId   ids[TSDB_MAX_REPLICA];
+SSyncInfo syncInfo;
+SSyncFSM* pFsm;
+SWal*     pWal;
+
+int64_t syncNodeInit() {
+  syncInfo.vgId = 1234;
+  syncInfo.rpcClient = gSyncIO->clientRpc;
+  syncInfo.FpSendMsg = syncIOSendMsg;
+  syncInfo.queue = gSyncIO->pMsgQ;
+  syncInfo.FpEqMsg = syncIOEqMsg;
+  syncInfo.pFsm = pFsm;
+  snprintf(syncInfo.path, sizeof(syncInfo.path), "./elect2_test_%d", myIndex);
+
+  int code = walInit();
+  assert(code == 0);
+  SWalCfg walCfg;
+  memset(&walCfg, 0, sizeof(SWalCfg));
+  walCfg.vgId = syncInfo.vgId;
+  walCfg.fsyncPeriod = 1000;
+  walCfg.retentionPeriod = 1000;
+  walCfg.rollPeriod = 1000;
+  walCfg.retentionSize = 1000;
+  walCfg.segSize = 1000;
+  walCfg.level = TAOS_WAL_FSYNC;
+
+  char tmpdir[128];
+  snprintf(tmpdir, sizeof(tmpdir), "./elect2_test_wal_%d", myIndex);
+  pWal = walOpen(tmpdir, &walCfg);
+  assert(pWal != NULL);
+
+  syncInfo.pWal = pWal;
+
+  SSyncCfg* pCfg = &syncInfo.syncCfg;
+  pCfg->myIndex = myIndex;
+  pCfg->replicaNum = replicaNum;
+
+  for (int i = 0; i < replicaNum; ++i) {
+    pCfg->nodeInfo[i].nodePort = ports[i];
+    snprintf(pCfg->nodeInfo[i].nodeFqdn, sizeof(pCfg->nodeInfo[i].nodeFqdn), "%s", "127.0.0.1");
+    // taosGetFqdn(pCfg->nodeInfo[0].nodeFqdn);
+  }
+
+  int64_t rid = syncStart(&syncInfo);
+  assert(rid > 0);
+
+  SSyncNode* pSyncNode = (SSyncNode*)syncNodeAcquire(rid);
+  assert(pSyncNode != NULL);
+
+  gSyncIO->FpOnSyncPing = pSyncNode->FpOnPing;
+  gSyncIO->FpOnSyncPingReply = pSyncNode->FpOnPingReply;
+  gSyncIO->FpOnSyncRequestVote = pSyncNode->FpOnRequestVote;
+  gSyncIO->FpOnSyncRequestVoteReply = pSyncNode->FpOnRequestVoteReply;
+  gSyncIO->FpOnSyncAppendEntries = pSyncNode->FpOnAppendEntries;
+  gSyncIO->FpOnSyncAppendEntriesReply = pSyncNode->FpOnAppendEntriesReply;
+  gSyncIO->FpOnSyncPing = pSyncNode->FpOnPing;
+  gSyncIO->FpOnSyncPingReply = pSyncNode->FpOnPingReply;
+  gSyncIO->FpOnSyncTimeout = pSyncNode->FpOnTimeout;
+  gSyncIO->pSyncNode = pSyncNode;
+
+  syncNodeRelease(pSyncNode);
+
+  return rid;
+}
+
+void initRaftId(SSyncNode* pSyncNode) {
+  for (int i = 0; i < replicaNum; ++i) {
+    ids[i] = pSyncNode->replicasId[i];
+    char* s = syncUtilRaftId2Str(&ids[i]);
+    printf("raftId[%d] : %s\n", i, s);
+    free(s);
+  }
+}
+
+int main(int argc, char** argv) {
+  // taosInitLog((char *)"syncTest.log", 100000, 10);
+  tsAsyncLog = 0;
+  sDebugFlag = 143 + 64;
+
+  myIndex = 0;
+  if (argc >= 2) {
+    myIndex = atoi(argv[1]);
+  }
+
+  int32_t ret = syncIOStart((char*)"127.0.0.1", ports[myIndex]);
+  assert(ret == 0);
+
+  ret = syncInit();
+  assert(ret == 0);
+
+  int64_t rid = syncNodeInit();
+  assert(rid > 0);
+
+  SSyncNode* pSyncNode = (SSyncNode*)syncNodeAcquire(rid);
+  assert(pSyncNode != NULL);
+
+  syncNodePrint2((char*)"", pSyncNode);
+  initRaftId(pSyncNode);
+
+  //---------------------------
+  while (1) {
+    sTrace(
+        "elect sleep, state: %d, %s, term:%lu electTimerLogicClock:%lu, electTimerLogicClockUser:%lu, electTimerMS:%d",
+        pSyncNode->state, syncUtilState2String(pSyncNode->state), pSyncNode->pRaftStore->currentTerm,
+        pSyncNode->electTimerLogicClock, pSyncNode->electTimerLogicClockUser, pSyncNode->electTimerMS);
+    taosMsleep(1000);
+  }
+
+  syncNodeRelease(pSyncNode);
+
+  return 0;
+}
diff --git a/source/os/src/osAtomic.c b/source/os/src/osAtomic.c
index 6429cdc9d3..8b452947d3 100644
--- a/source/os/src/osAtomic.c
+++ b/source/os/src/osAtomic.c
@@ -13,25 +13,18 @@
  * along with this program. If not, see <http://www.gnu.org/licenses/>.
  */
 
+#define ALLOW_FORBID_FUNC
 #include "os.h"
 
 #if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
 
 // add
 char interlocked_add_fetch_8(char volatile* ptr, char val) {
-#ifdef _TD_GO_DLL_
-  return __sync_fetch_and_add(ptr, val) + val;
-#else
   return _InterlockedExchangeAdd8(ptr, val) + val;
-#endif
 }
 
 short interlocked_add_fetch_16(short volatile* ptr, short val) {
-#ifdef _TD_GO_DLL_
-  return __sync_fetch_and_add(ptr, val) + val;
-#else
   return _InterlockedExchangeAdd16(ptr, val) + val;
-#endif
 }
 
 long interlocked_add_fetch_32(long volatile* ptr, long val) {
@@ -39,11 +32,7 @@ long interlocked_add_fetch_32(long volatile* ptr, long val) {
 }
 
 __int64 interlocked_add_fetch_64(__int64 volatile* ptr, __int64 val) {
-//#ifdef _WIN64
   return InterlockedExchangeAdd64(ptr, val) + val;
-//#else
-//  return _InterlockedExchangeAdd(ptr, val) + val;
-//#endif
 }
 
 char interlocked_and_fetch_8(char volatile* ptr, char val) {
@@ -197,3 +186,709 @@ int64_t atomic_exchange_64_impl(int64_t* ptr, int64_t val ) {
   return old;
 }
 #endif
+
+
+int8_t atomic_load_8(int8_t volatile *ptr) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return (*(int8_t volatile*)(ptr));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_add((ptr), 0);
+#else
+  return __atomic_load_n((ptr), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int16_t atomic_load_16(int16_t volatile *ptr) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return (*(int16_t volatile*)(ptr));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_add((ptr), 0);
+#else
+  return __atomic_load_n((ptr), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int32_t atomic_load_32(int32_t volatile *ptr) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return (*(int32_t volatile*)(ptr));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_add((ptr), 0);
+#else
+  return __atomic_load_n((ptr), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int64_t atomic_load_64(int64_t volatile *ptr) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return (*(int64_t volatile*)(ptr));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_add((ptr), 0);
+#else
+  return __atomic_load_n((ptr), __ATOMIC_SEQ_CST);
+#endif
+}
+
+void* atomic_load_ptr(void *ptr) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return (*(void* volatile*)(ptr));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_add((ptr), 0);
+#else
+  return __atomic_load_n((void**)(ptr), __ATOMIC_SEQ_CST);
+#endif
+}
+
+void atomic_store_8(int8_t volatile *ptr, int8_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  ((*(int8_t volatile*)(ptr)) = (int8_t)(val));
+#elif defined(_TD_NINGSI_60)
+  (*(ptr)=(val));
+#else
+  __atomic_store_n((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+void atomic_store_16(int16_t volatile *ptr, int16_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  ((*(int16_t volatile*)(ptr)) = (int16_t)(val));
+#elif defined(_TD_NINGSI_60)
+  (*(ptr)=(val));
+#else
+  __atomic_store_n((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+void atomic_store_32(int32_t volatile *ptr, int32_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  ((*(int32_t volatile*)(ptr)) = (int32_t)(val));
+#elif defined(_TD_NINGSI_60)
+  (*(ptr)=(val));
+#else
+  __atomic_store_n((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+void atomic_store_64(int64_t volatile *ptr, int64_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  ((*(int64_t volatile*)(ptr)) = (int64_t)(val));
+#elif defined(_TD_NINGSI_60)
+  (*(ptr)=(val));
+#else
+  __atomic_store_n((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+void atomic_store_ptr(void *ptr, void *val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  ((*(void* volatile*)(ptr)) = (void*)(val));
+#elif defined(_TD_NINGSI_60)
+  (*(ptr)=(val));
+#else
+  __atomic_store_n((void **)(ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int8_t atomic_exchange_8(int8_t volatile *ptr, int8_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedExchange8((int8_t volatile*)(ptr), (int8_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return atomic_exchange_8_impl((int8_t*)ptr, (int8_t)val);
+#else
+  return __atomic_exchange_n((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int16_t atomic_exchange_16(int16_t volatile *ptr, int16_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedExchange16((int16_t volatile*)(ptr), (int16_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return atomic_exchange_16_impl((int16_t*)ptr, (int16_t)val);
+#else
+  return __atomic_exchange_n((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int32_t atomic_exchange_32(int32_t volatile *ptr, int32_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedExchange((int32_t volatile*)(ptr), (int32_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return atomic_exchange_32_impl((int32_t*)ptr, (int32_t)val);
+#else
+  return __atomic_exchange_n((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int64_t atomic_exchange_64(int64_t volatile *ptr, int64_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedExchange64((int64_t volatile*)(ptr), (int64_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return atomic_exchange_64_impl((int64_t*)ptr, (int64_t)val);
+#else
+  return __atomic_exchange_n((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+void* atomic_exchange_ptr(void *ptr, void *val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  #ifdef _WIN64 
+  return _InterlockedExchangePointer((void* volatile*)(ptr), (void*)(val));
+  #else
+  return _InlineInterlockedExchangePointer((void* volatile*)(ptr), (void*)(val));
+  #endif
+#elif defined(_TD_NINGSI_60)
+  return atomic_exchange_ptr_impl((void *)ptr, (void*)val);
+#else
+  return __atomic_exchange_n((void **)(ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int8_t atomic_val_compare_exchange_8(int8_t volatile *ptr, int8_t oldval, int8_t newval) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedCompareExchange8((int8_t volatile*)(ptr), (int8_t)(newval), (int8_t)(oldval));
+#elif defined(_TD_NINGSI_60)
+  return __sync_val_compare_and_swap(ptr, oldval, newval);
+#else
+  return __sync_val_compare_and_swap(ptr, oldval, newval);
+#endif
+}
+
+int16_t atomic_val_compare_exchange_16(int16_t volatile *ptr, int16_t oldval, int16_t newval) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedCompareExchange16((int16_t volatile*)(ptr), (int16_t)(newval), (int16_t)(oldval));
+#elif defined(_TD_NINGSI_60)
+  return __sync_val_compare_and_swap(ptr, oldval, newval);
+#else
+  return __sync_val_compare_and_swap(ptr, oldval, newval);
+#endif
+}
+
+int32_t atomic_val_compare_exchange_32(int32_t volatile *ptr, int32_t oldval, int32_t newval) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedCompareExchange((int32_t volatile*)(ptr), (int32_t)(newval), (int32_t)(oldval));
+#elif defined(_TD_NINGSI_60)
+  return __sync_val_compare_and_swap(ptr, oldval, newval);
+#else
+  return __sync_val_compare_and_swap(ptr, oldval, newval);
+#endif
+}
+
+int64_t atomic_val_compare_exchange_64(int64_t volatile *ptr, int64_t oldval, int64_t newval) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedCompareExchange64((int64_t volatile*)(ptr), (int64_t)(newval), (int64_t)(oldval));
+#elif defined(_TD_NINGSI_60)
+  return __sync_val_compare_and_swap(ptr, oldval, newval);
+#else
+  return __sync_val_compare_and_swap(ptr, oldval, newval);
+#endif
+}
+
+void* atomic_val_compare_exchange_ptr(void *ptr, void *oldval, void *newval) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedCompareExchangePointer((void* volatile*)(ptr), (void*)(newval), (void*)(oldval));
+#elif defined(_TD_NINGSI_60)
+  return __sync_val_compare_and_swap(ptr, oldval, newval);
+#else
+  return __sync_val_compare_and_swap((void **)ptr, oldval, newval);
+#endif
+}
+
+int8_t atomic_add_fetch_8(int8_t volatile *ptr, int8_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_add_fetch_8((int8_t volatile*)(ptr), (int8_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_add_and_fetch((ptr), (val));
+#else
+  return __atomic_add_fetch((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int16_t atomic_add_fetch_16(int16_t volatile *ptr, int16_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_add_fetch_16((int16_t volatile*)(ptr), (int16_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_add_and_fetch((ptr), (val));
+#else
+  return __atomic_add_fetch((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int32_t atomic_add_fetch_32(int32_t volatile *ptr, int32_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_add_fetch_32((int32_t volatile*)(ptr), (int32_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_add_and_fetch((ptr), (val));
+#else
+  return __atomic_add_fetch((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int64_t atomic_add_fetch_64(int64_t volatile *ptr, int64_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_add_fetch_64((int64_t volatile*)(ptr), (int64_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_add_and_fetch((ptr), (val));
+#else
+  return __atomic_add_fetch((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+void* atomic_add_fetch_ptr(void *ptr, int32_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_add_fetch_ptr((void* volatile*)(ptr), (void*)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_add_and_fetch((ptr), (val));
+#else
+  return __atomic_add_fetch((void **)(ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int8_t atomic_fetch_add_8(int8_t volatile *ptr, int8_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedExchangeAdd8((int8_t volatile*)(ptr), (int8_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_add((ptr), (val));
+#else
+  return __atomic_fetch_add((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int16_t atomic_fetch_add_16(int16_t volatile *ptr, int16_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedExchangeAdd16((int16_t volatile*)(ptr), (int16_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_add((ptr), (val));
+#else
+  return __atomic_fetch_add((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int32_t atomic_fetch_add_32(int32_t volatile *ptr, int32_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedExchangeAdd((int32_t volatile*)(ptr), (int32_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_add((ptr), (val));
+#else
+  return __atomic_fetch_add((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int64_t atomic_fetch_add_64(int64_t volatile *ptr, int64_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedExchangeAdd64((int64_t volatile*)(ptr), (int64_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_add((ptr), (val));
+#else
+  return __atomic_fetch_add((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+void* atomic_fetch_add_ptr(void *ptr, int32_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedExchangeAddptr((void* volatile*)(ptr), (void*)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_add((ptr), (val));
+#else
+  return __atomic_fetch_add((void **)(ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int8_t atomic_sub_fetch_8(int8_t volatile *ptr, int8_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_add_fetch_8((int8_t volatile*)(ptr), -(int8_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_sub_and_fetch((ptr), (val));
+#else
+  return __atomic_sub_fetch((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int16_t atomic_sub_fetch_16(int16_t volatile *ptr, int16_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_add_fetch_16((int16_t volatile*)(ptr), -(int16_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_sub_and_fetch((ptr), (val));
+#else
+  return __atomic_sub_fetch((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int32_t atomic_sub_fetch_32(int32_t volatile *ptr, int32_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_add_fetch_32((int32_t volatile*)(ptr), -(int32_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_sub_and_fetch((ptr), (val));
+#else
+  return __atomic_sub_fetch((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int64_t atomic_sub_fetch_64(int64_t volatile *ptr, int64_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_add_fetch_64((int64_t volatile*)(ptr), -(int64_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_sub_and_fetch((ptr), (val));
+#else
+  return __atomic_sub_fetch((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+void* atomic_sub_fetch_ptr(void *ptr, int32_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_add_fetch_ptr((void* volatile*)(ptr), -(void*)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_sub_and_fetch((ptr), (val));
+#else
+  return __atomic_sub_fetch((void **)(ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int8_t atomic_fetch_sub_8(int8_t volatile *ptr, int8_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedExchangeAdd8((int8_t volatile*)(ptr), -(int8_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_sub((ptr), (val));
+#else
+  return __atomic_fetch_sub((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int16_t atomic_fetch_sub_16(int16_t volatile *ptr, int16_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedExchangeAdd16((int16_t volatile*)(ptr), -(int16_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_sub((ptr), (val));
+#else
+  return __atomic_fetch_sub((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int32_t atomic_fetch_sub_32(int32_t volatile *ptr, int32_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedExchangeAdd((int32_t volatile*)(ptr), -(int32_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_sub((ptr), (val));
+#else
+  return __atomic_fetch_sub((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int64_t atomic_fetch_sub_64(int64_t volatile *ptr, int64_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedExchangeAdd64((int64_t volatile*)(ptr), -(int64_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_sub((ptr), (val));
+#else
+  return __atomic_fetch_sub((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+void* atomic_fetch_sub_ptr(void *ptr, int32_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedExchangeAddptr((void* volatile*)(ptr), -(void*)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_sub((ptr), (val));
+#else
+  return __atomic_fetch_sub((void **)(ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int8_t atomic_and_fetch_8(int8_t volatile *ptr, int8_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_and_fetch_8((int8_t volatile*)(ptr), (int8_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_and_and_fetch((ptr), (val));
+#else
+  return __atomic_and_fetch((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int16_t atomic_and_fetch_16(int16_t volatile *ptr, int16_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_and_fetch_16((int16_t volatile*)(ptr), (int16_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_and_and_fetch((ptr), (val));
+#else
+  return __atomic_and_fetch((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int32_t atomic_and_fetch_32(int32_t volatile *ptr, int32_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_and_fetch_32((int32_t volatile*)(ptr), (int32_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_and_and_fetch((ptr), (val));
+#else
+  return __atomic_and_fetch((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int64_t atomic_and_fetch_64(int64_t volatile *ptr, int64_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_and_fetch_64((int64_t volatile*)(ptr), (int64_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_and_and_fetch((ptr), (val));
+#else
+  return __atomic_and_fetch((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+void* atomic_and_fetch_ptr(void *ptr, void *val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_and_fetch_ptr((void* volatile*)(ptr), (void*)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_and_and_fetch((ptr), (val));
+#else
+  return __atomic_and_fetch((void **)(ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int8_t atomic_fetch_and_8(int8_t volatile *ptr, int8_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedAnd8((int8_t volatile*)(ptr), (int8_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_and((ptr), (val));
+#else
+  return __atomic_fetch_and((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int16_t atomic_fetch_and_16(int16_t volatile *ptr, int16_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedAnd16((int16_t volatile*)(ptr), (int16_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_and((ptr), (val));
+#else
+  return __atomic_fetch_and((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int32_t atomic_fetch_and_32(int32_t volatile *ptr, int32_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedAnd((int32_t volatile*)(ptr), (int32_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_and((ptr), (val));
+#else
+  return __atomic_fetch_and((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int64_t atomic_fetch_and_64(int64_t volatile *ptr, int64_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_fetch_and_64((int64_t volatile*)(ptr), (int64_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_and((ptr), (val));
+#else
+  return __atomic_fetch_and((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+void* atomic_fetch_and_ptr(void *ptr, void *val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_fetch_and_ptr((void* volatile*)(ptr), (void*)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_and((ptr), (val));
+#else
+  return __atomic_fetch_and((void **)(ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int8_t atomic_or_fetch_8(int8_t volatile *ptr, int8_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_or_fetch_8((int8_t volatile*)(ptr), (int8_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_or_and_fetch((ptr), (val));
+#else
+  return __atomic_or_fetch((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int16_t atomic_or_fetch_16(int16_t volatile *ptr, int16_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_or_fetch_16((int16_t volatile*)(ptr), (int16_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_or_and_fetch((ptr), (val));
+#else
+  return __atomic_or_fetch((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int32_t atomic_or_fetch_32(int32_t volatile *ptr, int32_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_or_fetch_32((int32_t volatile*)(ptr), (int32_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_or_and_fetch((ptr), (val));
+#else
+  return __atomic_or_fetch((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int64_t atomic_or_fetch_64(int64_t volatile *ptr, int64_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_or_fetch_64((int64_t volatile*)(ptr), (int64_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_or_and_fetch((ptr), (val));
+#else
+  return __atomic_or_fetch((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+void* atomic_or_fetch_ptr(void *ptr, void *val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_or_fetch_ptr((void* volatile*)(ptr), (void*)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_or_and_fetch((ptr), (val));
+#else
+  return __atomic_or_fetch((void **)(ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int8_t atomic_fetch_or_8(int8_t volatile *ptr, int8_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedOr8((int8_t volatile*)(ptr), (int8_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_or((ptr), (val));
+#else
+  return __atomic_fetch_or((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int16_t atomic_fetch_or_16(int16_t volatile *ptr, int16_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedOr16((int16_t volatile*)(ptr), (int16_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_or((ptr), (val));
+#else
+  return __atomic_fetch_or((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int32_t atomic_fetch_or_32(int32_t volatile *ptr, int32_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedOr((int32_t volatile*)(ptr), (int32_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_or((ptr), (val));
+#else
+  return __atomic_fetch_or((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int64_t atomic_fetch_or_64(int64_t volatile *ptr, int64_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_fetch_or_64((int64_t volatile*)(ptr), (int64_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_or((ptr), (val));
+#else
+  return __atomic_fetch_or((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+void* atomic_fetch_or_ptr(void *ptr, void *val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_fetch_or_ptr((void* volatile*)(ptr), (void*)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_or((ptr), (val));
+#else
+  return __atomic_fetch_or((void **)(ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int8_t atomic_xor_fetch_8(int8_t volatile *ptr, int8_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_xor_fetch_8((int8_t volatile*)(ptr), (int8_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_xor_and_fetch((ptr), (val));
+#else
+  return __atomic_xor_fetch((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int16_t atomic_xor_fetch_16(int16_t volatile *ptr, int16_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_xor_fetch_16((int16_t volatile*)(ptr), (int16_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_xor_and_fetch((ptr), (val));
+#else
+  return __atomic_xor_fetch((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int32_t atomic_xor_fetch_32(int32_t volatile *ptr, int32_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_xor_fetch_32((int32_t volatile*)(ptr), (int32_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_xor_and_fetch((ptr), (val));
+#else
+  return __atomic_xor_fetch((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int64_t atomic_xor_fetch_64(int64_t volatile *ptr, int64_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_xor_fetch_64((int64_t volatile*)(ptr), (int64_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_xor_and_fetch((ptr), (val));
+#else
+  return __atomic_xor_fetch((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+void* atomic_xor_fetch_ptr(void *ptr, void *val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_xor_fetch_ptr((void* volatile*)(ptr), (void*)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_xor_and_fetch((ptr), (val));
+#else
+  return __atomic_xor_fetch((void **)(ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int8_t atomic_fetch_xor_8(int8_t volatile *ptr, int8_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedXor8((int8_t volatile*)(ptr), (int8_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_xor((ptr), (val));
+#else
+  return __atomic_fetch_xor((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int16_t atomic_fetch_xor_16(int16_t volatile *ptr, int16_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedXor16((int16_t volatile*)(ptr), (int16_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_xor((ptr), (val));
+#else
+  return __atomic_fetch_xor((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int32_t atomic_fetch_xor_32(int32_t volatile *ptr, int32_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return _InterlockedXor((int32_t volatile*)(ptr), (int32_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_xor((ptr), (val));
+#else
+  return __atomic_fetch_xor((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+int64_t atomic_fetch_xor_64(int64_t volatile *ptr, int64_t val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_fetch_xor_64((int64_t volatile*)(ptr), (int64_t)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_xor((ptr), (val));
+#else
+  return __atomic_fetch_xor((ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
+void* atomic_fetch_xor_ptr(void *ptr, void *val) {
+#if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
+  return interlocked_fetch_xor_ptr((void* volatile*)(ptr), (void*)(val));
+#elif defined(_TD_NINGSI_60)
+  return __sync_fetch_and_xor((ptr), (val));
+#else
+  return __atomic_fetch_xor((void **)(ptr), (val), __ATOMIC_SEQ_CST);
+#endif
+}
+
diff --git a/source/os/src/osSocket.c b/source/os/src/osSocket.c
index e5f83fa167..0e6abb5785 100644
--- a/source/os/src/osSocket.c
+++ b/source/os/src/osSocket.c
@@ -216,12 +216,6 @@ int32_t taosShutDownSocketServerRDWR(TdSocketServerPtr pSocketServer) {
 #endif
 }
 
-#if (defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32))
-  #if defined(_TD_GO_DLL_)
-    uint64_t htonll(uint64_t val) { return (((uint64_t)htonl(val)) << 32) + htonl(val >> 32); }
-  #endif
-#endif
-
 void taosWinSocketInit1() {
 #if defined(_TD_WINDOWS_64) || defined(_TD_WINDOWS_32)
   static char flag = 0;
diff --git a/source/util/src/tcache.c b/source/util/src/tcache.c
index 9a1b213534..cf5b623678 100644
--- a/source/util/src/tcache.c
+++ b/source/util/src/tcache.c
@@ -456,7 +456,7 @@ void *taosCacheAcquireByKey(SCacheObj *pCacheObj, const void *key, size_t keyLen
   }
 
   if (pCacheObj->numOfElems == 0) {
-    atomic_add_fetch_32(&pCacheObj->statistics.missCount, 1);
+    atomic_add_fetch_64(&pCacheObj->statistics.missCount, 1);
     return NULL;
   }
 
@@ -475,15 +475,15 @@ void *taosCacheAcquireByKey(SCacheObj *pCacheObj, const void *key, size_t keyLen
 
   void *pData = (pNode != NULL) ? pNode->data : NULL;
   if (pData != NULL) {
-    atomic_add_fetch_32(&pCacheObj->statistics.hitCount, 1);
+    atomic_add_fetch_64(&pCacheObj->statistics.hitCount, 1);
     uDebug("cache:%s, key:%p, %p is retrieved from cache, refcnt:%d", pCacheObj->name, key, pData,
            T_REF_VAL_GET(pNode));
   } else {
-    atomic_add_fetch_32(&pCacheObj->statistics.missCount, 1);
+    atomic_add_fetch_64(&pCacheObj->statistics.missCount, 1);
     uDebug("cache:%s, key:%p, not in cache, retrieved failed", pCacheObj->name, key);
   }
 
-  atomic_add_fetch_32(&pCacheObj->statistics.totalAccess, 1);
+  atomic_add_fetch_64(&pCacheObj->statistics.totalAccess, 1);
   return pData;
 }
 
diff --git a/source/util/src/thash.c b/source/util/src/thash.c
index 8b437b9797..a2e20f1762 100644
--- a/source/util/src/thash.c
+++ b/source/util/src/thash.c
@@ -55,7 +55,7 @@ typedef struct SHashEntry {
 struct SHashObj {
   SHashEntry **     hashList;
   size_t            capacity;      // number of slots
-  size_t            size;          // number of elements in hash table
+  int64_t           size;          // number of elements in hash table
   _hash_fn_t        hashFp;        // hash function
   _equal_fn_t       equalFp;       // equal function
   _hash_free_fn_t   freeFp;        // hash node free callback function
@@ -186,7 +186,7 @@ static SHashNode *doCreateHashNode(const void *key, size_t keyLen, const void *p
 static FORCE_INLINE void doUpdateHashNode(SHashObj *pHashObj, SHashEntry* pe, SHashNode* prev, SHashNode *pNode, SHashNode *pNewNode) {
   assert(pNode->keyLen == pNewNode->keyLen);
 
-  atomic_sub_fetch_32(&pNode->refCount, 1);
+  atomic_sub_fetch_16(&pNode->refCount, 1);
   if (prev != NULL) {
     prev->next = pNewNode;
   } else {
@@ -302,7 +302,7 @@ int32_t taosHashGetSize(const SHashObj *pHashObj) {
   if (pHashObj == NULL) {
     return 0;
   }
-  return (int32_t)atomic_load_64(&pHashObj->size);
+  return (int32_t)atomic_load_64((int64_t*)&pHashObj->size);
 }
 
 int32_t taosHashPut(SHashObj *pHashObj, const void *key, size_t keyLen, void *data, size_t size) {
@@ -508,7 +508,7 @@ int32_t taosHashRemove(SHashObj *pHashObj, const void *key, size_t keyLen) {
         pNode->removed == 0) {
       code = 0;  // it is found
 
-      atomic_sub_fetch_32(&pNode->refCount, 1);
+      atomic_sub_fetch_16(&pNode->refCount, 1);
       pNode->removed = 1;
       if (pNode->refCount <= 0) {
         if (prevNode == NULL) {
@@ -755,7 +755,7 @@ static void *taosHashReleaseNode(SHashObj *pHashObj, void *p, int *slot) {
       pNode = pNode->next;
     }
 
-    atomic_sub_fetch_32(&pOld->refCount, 1);
+    atomic_sub_fetch_16(&pOld->refCount, 1);
     if (pOld->refCount <=0) {
       if (prevNode) {
         prevNode->next = pOld->next;
diff --git a/source/util/src/ttimer.c b/source/util/src/ttimer.c
index cc31d9c94a..5e59e61ea1 100644
--- a/source/util/src/ttimer.c
+++ b/source/util/src/ttimer.c
@@ -132,7 +132,7 @@ static timer_map_t timerMap;
 static uintptr_t getNextTimerId() {
   uintptr_t id;
   do {
-    id = atomic_add_fetch_ptr(&nextTimerId, 1);
+    id = (uintptr_t)atomic_add_fetch_ptr((void **)&nextTimerId, 1);
   } while (id == 0);
   return id;
 }
diff --git a/tests/test/c/tmqDemo.c b/tests/test/c/tmqDemo.c
index 08e49a7efe..f8759f703b 100644
--- a/tests/test/c/tmqDemo.c
+++ b/tests/test/c/tmqDemo.c
@@ -65,11 +65,11 @@ typedef struct {
 static SConfInfo g_stConfInfo = {
     "tmqdb",
     "stb",
-	"./tmqResult.txt",	// output_file
-	"/data2/dnode/data/vnode/vnode2/wal",
+    "./tmqResult.txt",	// output_file
+    "/data2/dnode/data/vnodes/vnode2/wal",
     1, // threads
     1, // tables
-	1, // vgroups
+    1, // vgroups
     0, // run mode
     1, // columns
     1, // ratio
@@ -306,7 +306,7 @@ int32_t init_env() {
   }
 
   //const char* sql = "select * from tu1";
-  sprintf(sqlStr, "create topic test_stb_topic_1 as select * from %s%d", g_stConfInfo.stbName, 0);
+  sprintf(sqlStr, "create topic test_stb_topic_1 as select * from %s", g_stConfInfo.stbName);
   /*pRes = tmq_create_topic(pConn, "test_stb_topic_1", sqlStr, strlen(sqlStr));*/
   pRes = taos_query(pConn, sqlStr);
   if (taos_errno(pRes) != 0) {