merge from 3.0

2.0/src/query/src/queryMain.c

@@ -461,7 +461,7 @@ int32_t qKillQuery(qinfo_t qinfo) {
   }
 
   qDebug("QInfo:0x%"PRIx64" query killed", pQInfo->qId);
-  setQueryKilled(pQInfo);
+  setTaskKilled(pQInfo);
 
   // Wait for the query executing thread being stopped/
   // Once the query is stopped, the owner of qHandle will be cleared immediately.
@@ -634,7 +634,7 @@ int32_t qKillQueryByQId(void* pMgmt, int64_t qId, int32_t waitMs, int32_t waitCo
     return TSDB_CODE_QRY_INVALID_QHANDLE;
   }
   qWarn("QId:0x%"PRIx64" be killed(no memory commit).", pQInfo->qId);
-  setQueryKilled(pQInfo);
+  setTaskKilled(pQInfo);
 
   // wait query stop
   int32_t loop = 0;

include/libs/executor/executor.h

@@ -24,6 +24,7 @@ extern "C" {
 
 typedef void* qTaskInfo_t;
 typedef void* DataSinkHandle;
+struct SRpcMsg;
 struct SSubplan;
 
  /**
@@ -45,7 +46,7 @@ int32_t     qSetStreamInput(qTaskInfo_t tinfo, void* input);
   * @param qId
   * @return
   */
-int32_t qCreateExecTask(void* readHandle, int32_t vgId, struct SSubplan* pPlan, qTaskInfo_t* pTaskInfo, DataSinkHandle* handle);
+int32_t qCreateExecTask(void* readHandle, int32_t vgId, uint64_t taskId, struct SSubplan* pPlan, qTaskInfo_t* pTaskInfo, DataSinkHandle* handle);
 
 /**
  * The main task execution function, including query on both table and multiple tables,
@@ -208,6 +209,8 @@ void** qReleaseTask(void* pMgmt, void* pQInfo, bool freeHandle);
  */
 void** qDeregisterQInfo(void* pMgmt, void* pQInfo);
 
+void qProcessFetchRsp(void* parent, struct SRpcMsg* pMsg, struct SEpSet* pEpSet);
+
 #ifdef __cplusplus
 }
 #endif

include/libs/qworker/qworker.h

@@ -49,9 +49,10 @@ typedef struct {
 } SQWorkerStat;
 
 typedef int32_t (*putReqToQueryQFp)(void *, struct SRpcMsg *);
+typedef int32_t (*sendReqToDnodeFp)(void *, struct SEpSet *, struct SRpcMsg *);
 
-
-int32_t qWorkerInit(int8_t nodeType, int32_t nodeId, SQWorkerCfg *cfg, void **qWorkerMgmt, void *nodeObj, putReqToQueryQFp fp);
+int32_t qWorkerInit(int8_t nodeType, int32_t nodeId, SQWorkerCfg *cfg, void **qWorkerMgmt, void *nodeObj,
+                    putReqToQueryQFp fp1, sendReqToDnodeFp fp2);
 
 int32_t qWorkerProcessQueryMsg(void *node, void *qWorkerMgmt, SRpcMsg *pMsg);
 
@@ -65,6 +66,8 @@ int32_t qWorkerProcessStatusMsg(void *node, void *qWorkerMgmt, SRpcMsg *pMsg);
 
 int32_t qWorkerProcessFetchMsg(void *node, void *qWorkerMgmt, SRpcMsg *pMsg);
 
+int32_t qWorkerProcessFetchRsp(void *node, void *qWorkerMgmt, SRpcMsg *pMsg);
+
 int32_t qWorkerProcessCancelMsg(void *node, void *qWorkerMgmt, SRpcMsg *pMsg);
 
 int32_t qWorkerProcessDropMsg(void *node, void *qWorkerMgmt, SRpcMsg *pMsg);

include/os/os.h

@@ -51,6 +51,7 @@ extern "C" {
 #include <libgen.h>
 
 #include <sys/mman.h>
+#include <sys/prctl.h>
 
 #include "osAtomic.h"
 #include "osDef.h"

include/os/osDef.h

@@ -181,7 +181,8 @@ extern "C" {
   #endif
 #else
   // Windows
-  #define setThreadName(name)
+//  #define setThreadName(name)
+#define setThreadName(name) do { prctl(PR_SET_NAME, (name)); } while (0)
 #endif
 
 #if defined(_WIN32)

include/util/tqueue.h

@@ -15,6 +15,7 @@
 
 #ifndef _TD_UTIL_QUEUE_H
 #define _TD_UTIL_QUEUE_H
+#include "os.h"
 
 #ifdef __cplusplus
 extern "C" {
@@ -40,13 +41,13 @@ shall be used to set up the protection.
 typedef struct STaosQueue STaosQueue;
 typedef struct STaosQset  STaosQset;
 typedef struct STaosQall  STaosQall;
-typedef void (*FProcessItem)(void *ahandle, void *pItem);
-typedef void (*FProcessItems)(void *ahandle, STaosQall *qall, int32_t numOfItems);
+typedef void (*FItem)(void *ahandle, void *pItem);
+typedef void (*FItems)(void *ahandle, STaosQall *qall, int32_t numOfItems);
 
 STaosQueue *taosOpenQueue();
 void        taosCloseQueue(STaosQueue *queue);
-void        taosSetQueueFp(STaosQueue *queue, FProcessItem itemFp, FProcessItems itemsFp);
-void       *taosAllocateQitem(int32_t size);
+void        taosSetQueueFp(STaosQueue *queue, FItem itemFp, FItems itemsFp);
+void *      taosAllocateQitem(int32_t size);
 void        taosFreeQitem(void *pItem);
 int32_t     taosWriteQitem(STaosQueue *queue, void *pItem);
 int32_t     taosReadQitem(STaosQueue *queue, void **ppItem);
@@ -66,8 +67,11 @@ int32_t    taosAddIntoQset(STaosQset *qset, STaosQueue *queue, void *ahandle);
 void       taosRemoveFromQset(STaosQset *qset, STaosQueue *queue);
 int32_t    taosGetQueueNumber(STaosQset *qset);
 
-int32_t taosReadQitemFromQset(STaosQset *qset, void **ppItem, void **ahandle, FProcessItem *itemFp);
-int32_t taosReadAllQitemsFromQset(STaosQset *qset, STaosQall *qall, void **ahandle, FProcessItems *itemsFp);
+int32_t taosReadQitemFromQset(STaosQset *qset, void **ppItem, void **ahandle, FItem *itemFp);
+int32_t taosReadAllQitemsFromQset(STaosQset *qset, STaosQall *qall, void **ahandle, FItems *itemsFp);
+
+int32_t taosReadQitemFromQsetByThread(STaosQset *qset, void **ppItem, void **ahandle, FItem *itemFp, int32_t threadId);
+void    taosResetQsetThread(STaosQset *qset, void *pItem);
 
 int32_t taosGetQueueItemsNumber(STaosQueue *queue);
 int32_t taosGetQsetItemsNumber(STaosQset *qset);

include/util/tworker.h

@@ -15,57 +15,61 @@
 
 #ifndef _TD_UTIL_WORKER_H
 #define _TD_UTIL_WORKER_H
-
 #include "tqueue.h"
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-typedef struct SWorkerPool  SWorkerPool;
-typedef struct SMWorkerPool SMWorkerPool;
+typedef struct SQWorkerPool SQWorkerPool;
+typedef struct SWWorkerPool SWWorkerPool;
 
-typedef struct SWorker {
-  int32_t      id;      // worker ID
-  pthread_t    thread;  // thread
-  SWorkerPool *pool;
-} SWorker;
+typedef struct SQWorker {
+  int32_t       id;      // worker ID
+  pthread_t     thread;  // thread
+  SQWorkerPool *pool;
+} SQWorker, SFWorker;
 
-typedef struct SWorkerPool {
+typedef struct SQWorkerPool {
   int32_t         max;  // max number of workers
   int32_t         min;  // min number of workers
   int32_t         num;  // current number of workers
-  STaosQset      *qset;
-  const char     *name;
-  SWorker        *workers;
+  STaosQset *     qset;
+  const char *    name;
+  SQWorker *      workers;
   pthread_mutex_t mutex;
-} SWorkerPool;
+} SQWorkerPool, SFWorkerPool;
 
-typedef struct SMWorker {
+typedef struct SWWorker {
   int32_t       id;      // worker id
   pthread_t     thread;  // thread
-  STaosQall    *qall;
-  STaosQset    *qset;  // queue set
-  SMWorkerPool *pool;
-} SMWorker;
+  STaosQall *   qall;
+  STaosQset *   qset;  // queue set
+  SWWorkerPool *pool;
+} SWWorker;
 
-typedef struct SMWorkerPool {
+typedef struct SWWorkerPool {
   int32_t         max;     // max number of workers
   int32_t         nextId;  // from 0 to max-1, cyclic
-  const char     *name;
-  SMWorker       *workers;
+  const char *    name;
+  SWWorker *      workers;
   pthread_mutex_t mutex;
-} SMWorkerPool;
+} SWWorkerPool;
+
+int32_t     tQWorkerInit(SQWorkerPool *pool);
+void        tQWorkerCleanup(SQWorkerPool *pool);
+STaosQueue *tQWorkerAllocQueue(SQWorkerPool *pool, void *ahandle, FItem fp);
+void        tQWorkerFreeQueue(SQWorkerPool *pool, STaosQueue *queue);
 
-int32_t     tWorkerInit(SWorkerPool *pool);
-void        tWorkerCleanup(SWorkerPool *pool);
-STaosQueue *tWorkerAllocQueue(SWorkerPool *pool, void *ahandle, FProcessItem fp);
-void        tWorkerFreeQueue(SWorkerPool *pool, STaosQueue *queue);
+int32_t     tFWorkerInit(SFWorkerPool *pool);
+void        tFWorkerCleanup(SFWorkerPool *pool);
+STaosQueue *tFWorkerAllocQueue(SFWorkerPool *pool, void *ahandle, FItem fp);
+void        tFWorkerFreeQueue(SFWorkerPool *pool, STaosQueue *queue);
 
-int32_t     tMWorkerInit(SMWorkerPool *pool);
-void        tMWorkerCleanup(SMWorkerPool *pool);
-STaosQueue *tMWorkerAllocQueue(SMWorkerPool *pool, void *ahandle, FProcessItems fp);
-void        tMWorkerFreeQueue(SMWorkerPool *pool, STaosQueue *queue);
+int32_t     tWWorkerInit(SWWorkerPool *pool);
+void        tWWorkerCleanup(SWWorkerPool *pool);
+STaosQueue *tWWorkerAllocQueue(SWWorkerPool *pool, void *ahandle, FItems fp);
+void        tWWorkerFreeQueue(SWWorkerPool *pool, STaosQueue *queue);
 
 #ifdef __cplusplus
 }

source/client/test/clientTests.cpp

@@ -570,6 +570,7 @@ TEST(testCase, create_topic_Test) {
   //taos_close(pConn);
 //}
 
+#if 0
 TEST(testCase, tmq_subscribe_Test) {
   TAOS* pConn = taos_connect("localhost", "root", "taosdata", NULL, 0);
   assert(pConn != NULL);
@@ -594,6 +595,7 @@ TEST(testCase, tmq_subscribe_Test) {
     //if (msg == NULL) break;
   }
 }
+#endif
 
 TEST(testCase, tmq_consume_Test) {
 }

source/dnode/mgmt/impl/inc/dndEnv.h

@@ -31,8 +31,8 @@ typedef struct {
   SDnode     *pDnode;
   STaosQueue *queue;
   union {
-    SWorkerPool  pool;
-    SMWorkerPool mpool;
+    SQWorkerPool  pool;
+    SWWorkerPool mpool;
   };
 } SDnodeWorker;
 
@@ -109,10 +109,10 @@ typedef struct {
   int32_t      openVnodes;
   int32_t      totalVnodes;
   SRWLatch     latch;
-  SWorkerPool  queryPool;
-  SWorkerPool  fetchPool;
-  SMWorkerPool syncPool;
-  SMWorkerPool writePool;
+  SQWorkerPool queryPool;
+  SFWorkerPool fetchPool;
+  SWWorkerPool syncPool;
+  SWWorkerPool writePool;
 } SVnodesMgmt;
 
 typedef struct {

source/dnode/mgmt/impl/src/dndEnv.c

@@ -289,6 +289,7 @@ int32_t dndInit(const SDnodeEnvCfg *pCfg) {
       .charset = pCfg->charset,
       .nthreads = pCfg->numOfCommitThreads,
       .putReqToVQueryQFp = dndPutReqToVQueryQ,
+      .sendReqToDnodeFp = dndSendReqToDnode
   };
 
   if (vnodeInit(&vnodeOpt) != 0) {

source/dnode/mgmt/impl/src/dndMgmt.c

@@ -487,6 +487,8 @@ static void *dnodeThreadRoutine(void *param) {
   SDnodeMgmt *pMgmt = &pDnode->dmgmt;
   int32_t     ms = pDnode->cfg.statusInterval * 1000;
 
+  setThreadName("dnode-hb");
+
   while (true) {
     pthread_testcancel();
     taosMsleep(ms);
@@ -527,7 +529,7 @@ int32_t dndInitMgmt(SDnode *pDnode) {
   }
 
   if (pMgmt->dropped) {
-    dError("dnode will not start for its already dropped");
+    dError("dnode not start since its already dropped");
     return -1;
   }
 

source/dnode/mgmt/impl/src/dndTransport.c

@@ -25,9 +25,9 @@
 #include "dndMnode.h"
 #include "dndVnodes.h"
 
-#define INTERNAL_USER "_internal"
+#define INTERNAL_USER "_dnd"
 #define INTERNAL_CKEY "_key"
-#define INTERNAL_SECRET "_secret"
+#define INTERNAL_SECRET "_pwd"
 
 static void dndInitMsgFp(STransMgmt *pMgmt) {
   // Requests handled by DNODE
@@ -120,8 +120,9 @@ static void dndInitMsgFp(STransMgmt *pMgmt) {
   // Requests handled by VNODE
   pMgmt->msgFp[TMSG_INDEX(TDMT_VND_SUBMIT)] = dndProcessVnodeWriteMsg;
   pMgmt->msgFp[TMSG_INDEX(TDMT_VND_QUERY)] = dndProcessVnodeQueryMsg;
-  pMgmt->msgFp[TMSG_INDEX(TDMT_VND_QUERY_CONTINUE)] = dndProcessVnodeQueryMsg;  
+  pMgmt->msgFp[TMSG_INDEX(TDMT_VND_QUERY_CONTINUE)] = dndProcessVnodeQueryMsg;
   pMgmt->msgFp[TMSG_INDEX(TDMT_VND_FETCH)] = dndProcessVnodeFetchMsg;
+  pMgmt->msgFp[TMSG_INDEX(TDMT_VND_FETCH_RSP)] = dndProcessVnodeFetchMsg;
   pMgmt->msgFp[TMSG_INDEX(TDMT_VND_ALTER_TABLE)] = dndProcessVnodeWriteMsg;
   pMgmt->msgFp[TMSG_INDEX(TDMT_VND_UPDATE_TAG_VAL)] = dndProcessVnodeWriteMsg;
   pMgmt->msgFp[TMSG_INDEX(TDMT_VND_TABLE_META)] = dndProcessVnodeFetchMsg;
@@ -159,17 +160,18 @@ static void dndProcessResponse(void *parent, SRpcMsg *pRsp, SEpSet *pEpSet) {
 
   if (dndGetStat(pDnode) == DND_STAT_STOPPED) {
     if (pRsp == NULL || pRsp->pCont == NULL) return;
-    dTrace("RPC %p, rsp:%s is ignored since dnode is stopping", pRsp->handle, TMSG_INFO(msgType));
+    dTrace("RPC %p, rsp:%s ignored since dnode exiting, app:%p", pRsp->handle, TMSG_INFO(msgType), pRsp->ahandle);
     rpcFreeCont(pRsp->pCont);
     return;
   }
 
   DndMsgFp fp = pMgmt->msgFp[TMSG_INDEX(msgType)];
   if (fp != NULL) {
-    dTrace("RPC %p, rsp:%s will be processed, code:0x%x", pRsp->handle, TMSG_INFO(msgType), pRsp->code & 0XFFFF);
+    dTrace("RPC %p, rsp:%s will be processed, code:0x%x app:%p", pRsp->handle, TMSG_INFO(msgType), pRsp->code & 0XFFFF,
+           pRsp->ahandle);
     (*fp)(pDnode, pRsp, pEpSet);
   } else {
-    dError("RPC %p, rsp:%s not processed", pRsp->handle, TMSG_INFO(msgType));
+    dError("RPC %p, rsp:%s not processed, app:%p", pRsp->handle, TMSG_INFO(msgType), pRsp->ahandle);
     rpcFreeCont(pRsp->pCont);
   }
 }
@@ -196,7 +198,7 @@ static int32_t dndInitClient(SDnode *pDnode) {
 
   pMgmt->clientRpc = rpcOpen(&rpcInit);
   if (pMgmt->clientRpc == NULL) {
-    dError("failed to init rpc client");
+    dError("failed to init dnode rpc client");
     return -1;
   }
 
@@ -219,40 +221,39 @@ static void dndProcessRequest(void *param, SRpcMsg *pReq, SEpSet *pEpSet) {
 
   tmsg_t msgType = pReq->msgType;
   if (msgType == TDMT_DND_NETWORK_TEST) {
-    dTrace("RPC %p, network test req, app:%p will be processed, code:0x%x", pReq->handle, pReq->ahandle, pReq->code);
+    dTrace("RPC %p, network test req will be processed, app:%p", pReq->handle, pReq->ahandle);
     dndProcessStartupReq(pDnode, pReq);
     return;
   }
 
   if (dndGetStat(pDnode) == DND_STAT_STOPPED) {
-    dError("RPC %p, req:%s app:%p is ignored since dnode exiting", pReq->handle, TMSG_INFO(msgType), pReq->ahandle);
-    SRpcMsg rspMsg = {.handle = pReq->handle, .code = TSDB_CODE_DND_OFFLINE};
+    dError("RPC %p, req:%s ignored since dnode exiting, app:%p", pReq->handle, TMSG_INFO(msgType), pReq->ahandle);
+    SRpcMsg rspMsg = {.handle = pReq->handle, .code = TSDB_CODE_DND_OFFLINE, .ahandle = pReq->ahandle};
     rpcSendResponse(&rspMsg);
     rpcFreeCont(pReq->pCont);
     return;
   } else if (dndGetStat(pDnode) != DND_STAT_RUNNING) {
-    dError("RPC %p, req:%s app:%p is ignored since dnode not running", pReq->handle, TMSG_INFO(msgType), pReq->ahandle);
-    SRpcMsg rspMsg = {.handle = pReq->handle, .code = TSDB_CODE_APP_NOT_READY};
+    dError("RPC %p, req:%s ignored since dnode not running, app:%p", pReq->handle, TMSG_INFO(msgType), pReq->ahandle);
+    SRpcMsg rspMsg = {.handle = pReq->handle, .code = TSDB_CODE_APP_NOT_READY, .ahandle = pReq->ahandle};
     rpcSendResponse(&rspMsg);
     rpcFreeCont(pReq->pCont);
     return;
   }
 
   if (pReq->pCont == NULL) {
-    dTrace("RPC %p, req:%s app:%p not processed since content is null", pReq->handle, TMSG_INFO(msgType),
-           pReq->ahandle);
-    SRpcMsg rspMsg = {.handle = pReq->handle, .code = TSDB_CODE_DND_INVALID_MSG_LEN};
+    dTrace("RPC %p, req:%s not processed since its empty, app:%p", pReq->handle, TMSG_INFO(msgType), pReq->ahandle);
+    SRpcMsg rspMsg = {.handle = pReq->handle, .code = TSDB_CODE_DND_INVALID_MSG_LEN, .ahandle = pReq->ahandle};
     rpcSendResponse(&rspMsg);
     return;
   }
 
   DndMsgFp fp = pMgmt->msgFp[TMSG_INDEX(msgType)];
   if (fp != NULL) {
-    dTrace("RPC %p, req:%s app:%p will be processed", pReq->handle, TMSG_INFO(msgType), pReq->ahandle);
+    dTrace("RPC %p, req:%s will be processed, app:%p", pReq->handle, TMSG_INFO(msgType), pReq->ahandle);
     (*fp)(pDnode, pReq, pEpSet);
   } else {
-    dError("RPC %p, req:%s app:%p is not processed since no handle", pReq->handle, TMSG_INFO(msgType), pReq->ahandle);
-    SRpcMsg rspMsg = {.handle = pReq->handle, .code = TSDB_CODE_MSG_NOT_PROCESSED};
+    dError("RPC %p, req:%s not processed since no handle, app:%p", pReq->handle, TMSG_INFO(msgType), pReq->ahandle);
+    SRpcMsg rspMsg = {.handle = pReq->handle, .code = TSDB_CODE_MSG_NOT_PROCESSED, .ahandle = pReq->ahandle};
     rpcSendResponse(&rspMsg);
     rpcFreeCont(pReq->pCont);
   }
@@ -292,24 +293,24 @@ static int32_t dndRetrieveUserAuthInfo(void *parent, char *user, char *spi, char
   SDnode *pDnode = parent;
 
   if (dndAuthInternalReq(parent, user, spi, encrypt, secret, ckey) == 0) {
-    dTrace("user:%s, get auth from internal mnode, spi:%d encrypt:%d", user, *spi, *encrypt);
+    dTrace("user:%s, get auth from mnode, spi:%d encrypt:%d", user, *spi, *encrypt);
     return 0;
   }
 
   if (dndGetUserAuthFromMnode(pDnode, user, spi, encrypt, secret, ckey) == 0) {
-    dTrace("user:%s, get auth from internal mnode, spi:%d encrypt:%d", user, *spi, *encrypt);
+    dTrace("user:%s, get auth from mnode, spi:%d encrypt:%d", user, *spi, *encrypt);
     return 0;
   }
 
   if (terrno != TSDB_CODE_APP_NOT_READY) {
-    dTrace("failed to get user auth from internal mnode since %s", terrstr());
+    dTrace("failed to get user auth from mnode since %s", terrstr());
     return -1;
   }
 
   SAuthReq *pReq = rpcMallocCont(sizeof(SAuthReq));
   tstrncpy(pReq->user, user, TSDB_USER_LEN);
 
-  SRpcMsg rpcMsg = {.pCont = pReq, .contLen = sizeof(SAuthReq), .msgType = TDMT_MND_AUTH};
+  SRpcMsg rpcMsg = {.pCont = pReq, .contLen = sizeof(SAuthReq), .msgType = TDMT_MND_AUTH, .ahandle = (void *)9528};
   SRpcMsg rpcRsp = {0};
   dTrace("user:%s, send user auth req to other mnodes, spi:%d encrypt:%d", user, pReq->spi, pReq->encrypt);
   dndSendMsgToMnodeRecv(pDnode, &rpcMsg, &rpcRsp);
@@ -353,7 +354,7 @@ static int32_t dndInitServer(SDnode *pDnode) {
 
   pMgmt->serverRpc = rpcOpen(&rpcInit);
   if (pMgmt->serverRpc == NULL) {
-    dError("failed to init rpc server");
+    dError("failed to init dnode rpc server");
     return -1;
   }
 

source/dnode/mgmt/impl/src/dndVnodes.c

@@ -253,7 +253,7 @@ static int32_t dndGetVnodesFromFile(SDnode *pDnode, SWrapperCfg **ppCfgs, int32_
     }
 
     for (int32_t i = 0; i < vnodesNum; ++i) {
-      cJSON       *vnode = cJSON_GetArrayItem(vnodes, i);
+      cJSON *      vnode = cJSON_GetArrayItem(vnodes, i);
       SWrapperCfg *pCfg = &pCfgs[i];
 
       cJSON *vgId = cJSON_GetObjectItem(vnode, "vgId");
@@ -382,7 +382,7 @@ static void *dnodeOpenVnodeFunc(void *param) {
     dndReportStartup(pDnode, "open-vnodes", stepDesc);
 
     SVnodeCfg cfg = {.pDnode = pDnode, .pTfs = pDnode->pTfs, .vgId = pCfg->vgId};
-    SVnode   *pImpl = vnodeOpen(pCfg->path, &cfg);
+    SVnode *  pImpl = vnodeOpen(pCfg->path, &cfg);
     if (pImpl == NULL) {
       dError("vgId:%d, failed to open vnode by thread:%d", pCfg->vgId, pThread->threadIndex);
       pThread->failed++;
@@ -737,15 +737,9 @@ int32_t dndProcessCompactVnodeReq(SDnode *pDnode, SRpcMsg *pReq) {
   return 0;
 }
 
-static void dndProcessVnodeQueryQueue(SVnodeObj *pVnode, SRpcMsg *pMsg) {
-  SRpcMsg *pRsp = NULL;
-  vnodeProcessQueryReq(pVnode->pImpl, pMsg, &pRsp);
-}
+static void dndProcessVnodeQueryQueue(SVnodeObj *pVnode, SRpcMsg *pMsg) { vnodeProcessQueryMsg(pVnode->pImpl, pMsg); }
 
-static void dndProcessVnodeFetchQueue(SVnodeObj *pVnode, SRpcMsg *pMsg) {
-  SRpcMsg *pRsp = NULL;
-  vnodeProcessFetchReq(pVnode->pImpl, pMsg, &pRsp);
-}
+static void dndProcessVnodeFetchQueue(SVnodeObj *pVnode, SRpcMsg *pMsg) { vnodeProcessFetchMsg(pVnode->pImpl, pMsg); }
 
 static void dndProcessVnodeWriteQueue(SVnodeObj *pVnode, STaosQall *qall, int32_t numOfMsgs) {
   SArray *pArray = taosArrayInit(numOfMsgs, sizeof(SRpcMsg *));
@@ -916,27 +910,27 @@ static int32_t dndInitVnodeWorkers(SDnode *pDnode) {
   int32_t maxWriteThreads = TMAX(pDnode->env.numOfCores, 1);
   int32_t maxSyncThreads = TMAX(pDnode->env.numOfCores / 2, 1);
 
-  SWorkerPool *pPool = &pMgmt->queryPool;
-  pPool->name = "vnode-query";
-  pPool->min = minQueryThreads;
-  pPool->max = maxQueryThreads;
-  if (tWorkerInit(pPool) != 0) return -1;
+  SQWorkerPool *pQPool = &pMgmt->queryPool;
+  pQPool->name = "vnode-query";
+  pQPool->min = minQueryThreads;
+  pQPool->max = maxQueryThreads;
+  if (tQWorkerInit(pQPool) != 0) return -1;
 
-  pPool = &pMgmt->fetchPool;
-  pPool->name = "vnode-fetch";
-  pPool->min = minFetchThreads;
-  pPool->max = maxFetchThreads;
-  if (tWorkerInit(pPool) != 0) return -1;
+  SFWorkerPool *pFPool = &pMgmt->fetchPool;
+  pFPool->name = "vnode-fetch";
+  pFPool->min = minFetchThreads;
+  pFPool->max = maxFetchThreads;
+  if (tFWorkerInit(pFPool) != 0) return -1;
 
-  SMWorkerPool *pMPool = &pMgmt->writePool;
-  pMPool->name = "vnode-write";
-  pMPool->max = maxWriteThreads;
-  if (tMWorkerInit(pMPool) != 0) return -1;
+  SWWorkerPool *pWPool = &pMgmt->writePool;
+  pWPool->name = "vnode-write";
+  pWPool->max = maxWriteThreads;
+  if (tWWorkerInit(pWPool) != 0) return -1;
 
-  pMPool = &pMgmt->syncPool;
-  pMPool->name = "vnode-sync";
-  pMPool->max = maxSyncThreads;
-  if (tMWorkerInit(pMPool) != 0) return -1;
+  pWPool = &pMgmt->syncPool;
+  pWPool->name = "vnode-sync";
+  pWPool->max = maxSyncThreads;
+  if (tWWorkerInit(pWPool) != 0) return -1;
 
   dDebug("vnode workers is initialized");
   return 0;
@@ -944,21 +938,21 @@ static int32_t dndInitVnodeWorkers(SDnode *pDnode) {
 
 static void dndCleanupVnodeWorkers(SDnode *pDnode) {
   SVnodesMgmt *pMgmt = &pDnode->vmgmt;
-  tWorkerCleanup(&pMgmt->fetchPool);
-  tWorkerCleanup(&pMgmt->queryPool);
-  tMWorkerCleanup(&pMgmt->writePool);
-  tMWorkerCleanup(&pMgmt->syncPool);
+  tFWorkerCleanup(&pMgmt->fetchPool);
+  tQWorkerCleanup(&pMgmt->queryPool);
+  tWWorkerCleanup(&pMgmt->writePool);
+  tWWorkerCleanup(&pMgmt->syncPool);
   dDebug("vnode workers is closed");
 }
 
 static int32_t dndAllocVnodeQueue(SDnode *pDnode, SVnodeObj *pVnode) {
   SVnodesMgmt *pMgmt = &pDnode->vmgmt;
 
-  pVnode->pWriteQ = tMWorkerAllocQueue(&pMgmt->writePool, pVnode, (FProcessItems)dndProcessVnodeWriteQueue);
-  pVnode->pApplyQ = tMWorkerAllocQueue(&pMgmt->writePool, pVnode, (FProcessItems)dndProcessVnodeApplyQueue);
-  pVnode->pSyncQ = tMWorkerAllocQueue(&pMgmt->syncPool, pVnode, (FProcessItems)dndProcessVnodeSyncQueue);
-  pVnode->pFetchQ = tWorkerAllocQueue(&pMgmt->fetchPool, pVnode, (FProcessItem)dndProcessVnodeFetchQueue);
-  pVnode->pQueryQ = tWorkerAllocQueue(&pMgmt->queryPool, pVnode, (FProcessItem)dndProcessVnodeQueryQueue);
+  pVnode->pWriteQ = tWWorkerAllocQueue(&pMgmt->writePool, pVnode, (FItems)dndProcessVnodeWriteQueue);
+  pVnode->pApplyQ = tWWorkerAllocQueue(&pMgmt->writePool, pVnode, (FItems)dndProcessVnodeApplyQueue);
+  pVnode->pSyncQ = tWWorkerAllocQueue(&pMgmt->syncPool, pVnode, (FItems)dndProcessVnodeSyncQueue);
+  pVnode->pFetchQ = tFWorkerAllocQueue(&pMgmt->fetchPool, pVnode, (FItem)dndProcessVnodeFetchQueue);
+  pVnode->pQueryQ = tQWorkerAllocQueue(&pMgmt->queryPool, pVnode, (FItem)dndProcessVnodeQueryQueue);
 
   if (pVnode->pApplyQ == NULL || pVnode->pWriteQ == NULL || pVnode->pSyncQ == NULL || pVnode->pFetchQ == NULL ||
       pVnode->pQueryQ == NULL) {
@@ -971,11 +965,11 @@ static int32_t dndAllocVnodeQueue(SDnode *pDnode, SVnodeObj *pVnode) {
 
 static void dndFreeVnodeQueue(SDnode *pDnode, SVnodeObj *pVnode) {
   SVnodesMgmt *pMgmt = &pDnode->vmgmt;
-  tWorkerFreeQueue(&pMgmt->queryPool, pVnode->pQueryQ);
-  tWorkerFreeQueue(&pMgmt->fetchPool, pVnode->pFetchQ);
-  tMWorkerFreeQueue(&pMgmt->writePool, pVnode->pWriteQ);
-  tMWorkerFreeQueue(&pMgmt->writePool, pVnode->pApplyQ);
-  tMWorkerFreeQueue(&pMgmt->syncPool, pVnode->pSyncQ);
+  tQWorkerFreeQueue(&pMgmt->queryPool, pVnode->pQueryQ);
+  tFWorkerFreeQueue(&pMgmt->fetchPool, pVnode->pFetchQ);
+  tWWorkerFreeQueue(&pMgmt->writePool, pVnode->pWriteQ);
+  tWWorkerFreeQueue(&pMgmt->writePool, pVnode->pApplyQ);
+  tWWorkerFreeQueue(&pMgmt->syncPool, pVnode->pSyncQ);
   pVnode->pWriteQ = NULL;
   pVnode->pApplyQ = NULL;
   pVnode->pSyncQ = NULL;

source/dnode/mgmt/impl/src/dndWorker.c

@@ -31,28 +31,28 @@ int32_t dndInitWorker(SDnode *pDnode, SDnodeWorker *pWorker, EWorkerType type, c
   pWorker->pDnode = pDnode;
 
   if (pWorker->type == DND_WORKER_SINGLE) {
-    SWorkerPool *pPool = &pWorker->pool;
+    SQWorkerPool *pPool = &pWorker->pool;
     pPool->name = name;
     pPool->min = minNum;
     pPool->max = maxNum;
-    if (tWorkerInit(pPool) != 0) {
+    if (tQWorkerInit(pPool) != 0) {
       terrno = TSDB_CODE_OUT_OF_MEMORY;
       return -1;
     }
-    pWorker->queue = tWorkerAllocQueue(pPool, pDnode, (FProcessItem)queueFp);
+    pWorker->queue = tQWorkerAllocQueue(pPool, pDnode, (FItem)queueFp);
     if (pWorker->queue == NULL) {
       terrno = TSDB_CODE_OUT_OF_MEMORY;
       return -1;
     }
   } else if (pWorker->type == DND_WORKER_MULTI) {
-    SMWorkerPool *pPool = &pWorker->mpool;
+    SWWorkerPool *pPool = &pWorker->mpool;
     pPool->name = name;
     pPool->max = maxNum;
-    if (tMWorkerInit(pPool) != 0) {
+    if (tWWorkerInit(pPool) != 0) {
       terrno = TSDB_CODE_OUT_OF_MEMORY;
       return -1;
     }
-    pWorker->queue = tMWorkerAllocQueue(pPool, pDnode, (FProcessItems)queueFp);
+    pWorker->queue = tWWorkerAllocQueue(pPool, pDnode, (FItems)queueFp);
     if (pWorker->queue == NULL) {
       terrno = TSDB_CODE_OUT_OF_MEMORY;
       return -1;
@@ -70,11 +70,11 @@ void dndCleanupWorker(SDnodeWorker *pWorker) {
   }
 
   if (pWorker->type == DND_WORKER_SINGLE) {
-    tWorkerCleanup(&pWorker->pool);
-    tWorkerFreeQueue(&pWorker->pool, pWorker->queue);
+    tQWorkerCleanup(&pWorker->pool);
+    tQWorkerFreeQueue(&pWorker->pool, pWorker->queue);
   } else if (pWorker->type == DND_WORKER_MULTI) {
-    tMWorkerCleanup(&pWorker->mpool);
-    tMWorkerFreeQueue(&pWorker->mpool, pWorker->queue);
+    tWWorkerCleanup(&pWorker->mpool);
+    tWWorkerFreeQueue(&pWorker->mpool, pWorker->queue);
   } else {
   }
 }

source/dnode/mnode/impl/src/mndTrans.c

@@ -921,7 +921,7 @@ static int32_t mndProcessTransMsg(SMnodeMsg *pMsg) {
 
 void mndTransPullup(SMnode *pMnode) {
   STrans *pTrans = NULL;
-  void   *pIter = NULL;
+  void *  pIter = NULL;
 
   while (1) {
     pIter = sdbFetch(pMnode->pSdb, SDB_TRANS, pIter, (void **)&pTrans);
@@ -930,4 +930,6 @@ void mndTransPullup(SMnode *pMnode) {
     mndTransExecute(pMnode, pTrans);
     sdbRelease(pMnode->pSdb, pTrans);
   }
+
+  sdbWriteFile(pMnode->pSdb);
 }

source/dnode/mnode/sdb/src/sdb.c

@@ -64,11 +64,7 @@ SSdb *sdbInit(SSdbOpt *pOption) {
 void sdbCleanup(SSdb *pSdb) {
   mDebug("start to cleanup sdb");
 
-  if (pSdb->curVer > pSdb->lastCommitVer) {
-    mDebug("write sdb file for current ver:%" PRId64 " larger than last commit ver:%" PRId64, pSdb->curVer,
-           pSdb->lastCommitVer);
-    sdbWriteFile(pSdb);
-  }
+  sdbWriteFile(pSdb);
 
   if (pSdb->currDir != NULL) {
     tfree(pSdb->currDir);

source/dnode/mnode/sdb/src/sdbFile.c

@@ -221,7 +221,7 @@ PARSE_SDB_DATA_ERROR:
   return code;
 }
 
-int32_t sdbWriteFile(SSdb *pSdb) {
+static int32_t sdbWriteFileImp(SSdb *pSdb) {
   int32_t code = 0;
 
   char tmpfile[PATH_MAX] = {0};
@@ -229,7 +229,8 @@ int32_t sdbWriteFile(SSdb *pSdb) {
   char curfile[PATH_MAX] = {0};
   snprintf(curfile, sizeof(curfile), "%s%ssdb.data", pSdb->currDir, TD_DIRSEP);
 
-  mDebug("start to write file:%s", curfile);
+  mDebug("start to write file:%s, current ver:%" PRId64 ", commit ver:%" PRId64, curfile, pSdb->curVer,
+         pSdb->lastCommitVer);
 
   FileFd fd = taosOpenFileCreateWriteTrunc(tmpfile);
   if (fd <= 0) {
@@ -323,12 +324,20 @@ int32_t sdbWriteFile(SSdb *pSdb) {
   return code;
 }
 
+int32_t sdbWriteFile(SSdb *pSdb) {
+  if (pSdb->curVer == pSdb->lastCommitVer) {
+    return 0;
+  }
+
+  return sdbWriteFileImp(pSdb);
+}
+
 int32_t sdbDeploy(SSdb *pSdb) {
   if (sdbRunDeployFp(pSdb) != 0) {
     return -1;
   }
 
-  if (sdbWriteFile(pSdb) != 0) {
+  if (sdbWriteFileImp(pSdb) != 0) {
     return -1;
   }
 

source/dnode/vnode/inc/tq.h

@@ -318,8 +318,8 @@ int       tqRegisterContext(STqGroup*, void* ahandle);
 int       tqSendLaunchQuery(STqMsgItem*, int64_t offset);
 #endif
 
-int32_t tqProcessConsumeReq(STQ* pTq, SRpcMsg* pMsg, SRpcMsg** ppRsp);
-int32_t tqProcessSetConnReq(STQ* pTq, char* msg, SRpcMsg** ppRsp);
+int32_t tqProcessConsumeReq(STQ* pTq, SRpcMsg* pMsg);
+int32_t tqProcessSetConnReq(STQ* pTq, char* msg);
 
 #ifdef __cplusplus
 }

source/dnode/vnode/inc/tsdb.h

@@ -75,7 +75,6 @@ typedef struct STsdbQueryCond {
 } STsdbQueryCond;
 
 typedef struct {
-  void    *pTable;
   TSKEY    lastKey;
   uint64_t uid;
 } STableKeyInfo;
@@ -104,8 +103,7 @@ typedef void* tsdbReadHandleT;
  * @param qinfo      query info handle from query processor
  * @return
  */
-tsdbReadHandleT *tsdbQueryTables(STsdb *tsdb, STsdbQueryCond *pCond, STableGroupInfo *tableInfoGroup, uint64_t qId,
-                                  void *pRef);
+tsdbReadHandleT *tsdbQueryTables(STsdb *tsdb, STsdbQueryCond *pCond, STableGroupInfo *tableInfoGroup, uint64_t qId, uint64_t taskId);
 
 /**
  * Get the last row of the given query time window for all the tables in STableGroupInfo object.
@@ -142,7 +140,7 @@ bool isTsdbCacheLastRow(tsdbReadHandleT* pTsdbReadHandle);
  */
 int32_t tsdbQuerySTableByTagCond(STsdb* tsdb, uint64_t uid, TSKEY skey, const char* pTagCond, size_t len,
                                  int16_t tagNameRelType, const char* tbnameCond, STableGroupInfo* pGroupInfo,
-                                 SColIndex* pColIndex, int32_t numOfCols, uint64_t reqId);
+                                 SColIndex* pColIndex, int32_t numOfCols, uint64_t reqId, uint64_t taskId);
 /**
  * get num of rows in mem table
  *

source/dnode/vnode/inc/vnode.h

@@ -33,6 +33,7 @@ extern "C" {
 typedef struct SVnode SVnode;
 typedef struct SDnode SDnode;
 typedef int32_t (*PutReqToVQueryQFp)(SDnode *pDnode, struct SRpcMsg *pReq);
+typedef int32_t (*SendReqToDnodeFp)(SDnode *pDnode, struct SEpSet *epSet, struct SRpcMsg *rpcMsg);
 
 typedef struct STqCfg {
   // TODO
@@ -64,6 +65,7 @@ typedef struct {
   const char       *charset;
   uint16_t          nthreads;  // number of commit threads. 0 for no threads and a schedule queue should be given (TODO)
   PutReqToVQueryQFp putReqToVQueryQFp;
+  SendReqToDnodeFp  sendReqToDnodeFp;
 } SVnodeOpt;
 
 typedef struct STqReadHandle {
@@ -159,20 +161,18 @@ int vnodeProcessSyncReq(SVnode *pVnode, SRpcMsg *pMsg, SRpcMsg **pRsp);
  *
  * @param pVnode The vnode object.
  * @param pMsg The request message
- * @param pRsp The response message
  * @return int 0 for success, -1 for failure
  */
-int vnodeProcessQueryReq(SVnode *pVnode, SRpcMsg *pMsg, SRpcMsg **pRsp);
+int vnodeProcessQueryMsg(SVnode *pVnode, SRpcMsg *pMsg);
 
 /**
  * @brief Process a fetch message.
  *
  * @param pVnode The vnode object.
  * @param pMsg The request message
- * @param pRsp The response message
  * @return int 0 for success, -1 for failure
  */
-int vnodeProcessFetchReq(SVnode *pVnode, SRpcMsg *pMsg, SRpcMsg **pRsp);
+int vnodeProcessFetchMsg(SVnode *pVnode, SRpcMsg *pMsg);
 
 /* ------------------------ SVnodeCfg ------------------------ */
 /**

source/dnode/vnode/src/inc/vnd.h

@@ -55,6 +55,7 @@ typedef struct SVnodeMgr {
   // For vnode Mgmt
   SDnode*           pDnode;
   PutReqToVQueryQFp putReqToVQueryQFp;
+  SendReqToDnodeFp  sendReqToDnodeFp;
 } SVnodeMgr;
 
 extern SVnodeMgr vnodeMgr;
@@ -85,6 +86,7 @@ struct SVnode {
 int vnodeScheduleTask(SVnodeTask* task);
 
 int32_t vnodePutReqToVQueryQ(SVnode* pVnode, struct SRpcMsg* pReq);
+void    vnodeSendReqToDnode(SVnode* pVnode, struct SEpSet* epSet, struct SRpcMsg* pReq);
 
 // For Log
 extern int32_t vDebugFlag;

source/dnode/vnode/src/tq/tq.c

@@ -667,8 +667,8 @@ int tqItemSSize() {
 }
 #endif
 
-int32_t tqProcessConsumeReq(STQ* pTq, SRpcMsg* pMsg, SRpcMsg** ppRsp) {
-  SMqConsumeReq* pReq = pMsg->pCont;
+int32_t tqProcessConsumeReq(STQ* pTq, SRpcMsg* pMsg) {
+  SMqConsumeReq*   pReq = pMsg->pCont;
   SRpcMsg          rpcMsg;
   int64_t          reqId = pReq->reqId;
   int64_t          consumerId = pReq->consumerId;
@@ -767,7 +767,7 @@ int32_t tqProcessConsumeReq(STQ* pTq, SRpcMsg* pMsg, SRpcMsg** ppRsp) {
   return 0;
 }
 
-int32_t tqProcessSetConnReq(STQ* pTq, char* msg, SRpcMsg** ppRsp) {
+int32_t tqProcessSetConnReq(STQ* pTq, char* msg) {
   SMqSetCVgReq req;
   tDecodeSMqSetCVgReq(msg, &req);
   STqConsumerHandle* pConsumer = calloc(sizeof(STqConsumerHandle), 1);

source/dnode/vnode/src/tsdb/tsdbRead.c

@@ -81,7 +81,6 @@ enum {
   CHECKINFO_CHOSEN_BOTH = 2    //for update=2(merge case)
 };
 
-
 typedef struct STableCheckInfo {
   uint64_t      tableId;
   TSKEY         lastKey;
@@ -132,7 +131,7 @@ typedef struct STsdbReadHandle {
   bool           loadExternalRow;  // load time window external data rows
   bool           currentLoadExternalRows; // current load external rows
   int32_t        loadType;         // block load type
-  uint64_t       qId;              // query info handle, for debug purpose
+  char          *idStr;            // query info handle, for debug purpose
   int32_t        type;             // query type: retrieve all data blocks, 2. retrieve only last row, 3. retrieve direct prev|next rows
   SDFileSet*     pFileGroup;
   SFSIter        fileIter;
@@ -209,31 +208,6 @@ static SArray* getDefaultLoadColumns(STsdbReadHandle* pTsdbReadHandle, bool load
   return pLocalIdList;
 }
 
-static void tsdbMayTakeMemSnapshot(STsdbReadHandle* pTsdbReadHandle, SArray* psTable) {
-//  assert(pTsdbReadHandle != NULL && pTsdbReadHandle->pMemRef != NULL);
-//
-//  STsdbMemTable* pMemRef = pTsdbReadHandle->pMemRef;
-//  if (pTsdbReadHandle->pMemRef->ref++ == 0) {
-//    tsdbTakeMemSnapshot(pTsdbReadHandle->pTsdb, &(pMemRef->snapshot), psTable);
-//  }
-//
-//  taosArrayDestroy(psTable);
-}
-
-static void tsdbMayUnTakeMemSnapshot(STsdbReadHandle* pTsdbReadHandle) {
-//  assert(pTsdbReadHandle != NULL);
-//  STsdbMemTable* pMemRef = pTsdbReadHandle->pMemRef;
-//  if (pMemRef == NULL) { // it has been freed
-//    return;
-//  }
-//
-//  if (--pMemRef->ref == 0) {
-//    tsdbUnTakeMemSnapShot(pTsdbReadHandle->pTsdb, &(pMemRef->snapshot));
-//  }
-//
-//  pTsdbReadHandle->pMemRef = NULL;
-}
-
 //int64_t tsdbGetNumOfRowsInMemTable(tsdbReadHandleT* pHandle) {
 //  STsdbReadHandle* pTsdbReadHandle = (STsdbReadHandle*) pHandle;
 //
@@ -263,9 +237,9 @@ static void tsdbMayUnTakeMemSnapshot(STsdbReadHandle* pTsdbReadHandle) {
 //  return rows;
 //}
 
-static SArray* createCheckInfoFromTableGroup(STsdbReadHandle* pTsdbReadHandle, STableGroupInfo* pGroupList, SArray** psTable) {
-  size_t sizeOfGroup = taosArrayGetSize(pGroupList->pGroupList);
-  assert(sizeOfGroup >= 1);
+static SArray* createCheckInfoFromTableGroup(STsdbReadHandle* pTsdbReadHandle, STableGroupInfo* pGroupList) {
+  size_t numOfGroup = taosArrayGetSize(pGroupList->pGroupList);
+  assert(numOfGroup >= 1);
 
   // allocate buffer in order to load data blocks from file
   SArray* pTableCheckInfo = taosArrayInit(pGroupList->numOfTables, sizeof(STableCheckInfo));
@@ -273,14 +247,8 @@ static SArray* createCheckInfoFromTableGroup(STsdbReadHandle* pTsdbReadHandle, S
     return NULL;
   }
 
-  SArray* pTable = taosArrayInit(4, sizeof(STable*));
-  if (pTable == NULL) {
-    taosArrayDestroy(pTableCheckInfo);
-    return NULL;
-  }
-
   // todo apply the lastkey of table check to avoid to load header file
-  for (int32_t i = 0; i < sizeOfGroup; ++i) {
+  for (int32_t i = 0; i < numOfGroup; ++i) {
     SArray* group = *(SArray**) taosArrayGet(pGroupList->pGroupList, i);
 
     size_t gsize = taosArrayGetSize(group);
@@ -289,12 +257,7 @@ static SArray* createCheckInfoFromTableGroup(STsdbReadHandle* pTsdbReadHandle, S
     for (int32_t j = 0; j < gsize; ++j) {
       STableKeyInfo* pKeyInfo = (STableKeyInfo*) taosArrayGet(group, j);
 
-      STableCheckInfo info = { .lastKey = pKeyInfo->lastKey};
-//      assert(info.pTableObj != NULL && (info.pTableObj->type == TSDB_NORMAL_TABLE ||
-//                                        info.pTableObj->type == TSDB_CHILD_TABLE || info.pTableObj->type == TSDB_STREAM_TABLE));
-
-      info.tableId = pKeyInfo->uid;
-
+      STableCheckInfo info = { .lastKey = pKeyInfo->lastKey, .tableId = pKeyInfo->uid};
       if (ASCENDING_TRAVERSE(pTsdbReadHandle->order)) {
         if (info.lastKey == INT64_MIN || info.lastKey < pTsdbReadHandle->window.skey) {
           info.lastKey = pTsdbReadHandle->window.skey;
@@ -306,17 +269,12 @@ static SArray* createCheckInfoFromTableGroup(STsdbReadHandle* pTsdbReadHandle, S
       }
 
       taosArrayPush(pTableCheckInfo, &info);
-      tsdbDebug("%p check table uid:%"PRId64" from lastKey:%"PRId64" 0x%"PRIx64, pTsdbReadHandle, info.tableId, info.lastKey, pTsdbReadHandle->qId);
+      tsdbDebug("%p check table uid:%"PRId64" from lastKey:%"PRId64" %s", pTsdbReadHandle, info.tableId, info.lastKey, pTsdbReadHandle->idStr);
     }
   }
 
-//  taosArraySort(pTableCheckInfo, tsdbCheckInfoCompar);
-  size_t gsize = taosArrayGetSize(pTableCheckInfo);
-//  for (int32_t i = 0; i < gsize; ++i) {
-//    STableCheckInfo* pInfo = (STableCheckInfo*) taosArrayGet(pTableCheckInfo, i);
-//  }
-
-  *psTable = pTable;
+  // TODO  group table according to the tag value.
+  taosArraySort(pTableCheckInfo, tsdbCheckInfoCompar);
   return pTableCheckInfo;
 }
 
@@ -389,13 +347,13 @@ static void setQueryTimewindow(STsdbReadHandle* pTsdbReadHandle, STsdbQueryCond*
   }
 
   if (updateTs) {
-    tsdbDebug("%p update the query time window, old:%" PRId64 " - %" PRId64 ", new:%" PRId64 " - %" PRId64
-              ", 0x%" PRIx64, pTsdbReadHandle, pCond->twindow.skey, pCond->twindow.ekey, pTsdbReadHandle->window.skey,
-              pTsdbReadHandle->window.ekey, pTsdbReadHandle->qId);
+    tsdbDebug("%p update the query time window, old:%" PRId64 " - %" PRId64 ", new:%" PRId64 " - %" PRId64 ", %s",
+              pTsdbReadHandle, pCond->twindow.skey, pCond->twindow.ekey, pTsdbReadHandle->window.skey,
+              pTsdbReadHandle->window.ekey, pTsdbReadHandle->idStr);
   }
 }
 
-static STsdbReadHandle* tsdbQueryTablesImpl(STsdb* tsdb, STsdbQueryCond* pCond, uint64_t qId, STsdbMemTable* pMemRef) {
+static STsdbReadHandle* tsdbQueryTablesImpl(STsdb* tsdb, STsdbQueryCond* pCond, uint64_t qId, uint64_t taskId) {
   STsdbReadHandle* pReadHandle = calloc(1, sizeof(STsdbReadHandle));
   if (pReadHandle == NULL) {
     goto _end;
@@ -408,7 +366,6 @@ static STsdbReadHandle* tsdbQueryTablesImpl(STsdb* tsdb, STsdbQueryCond* pCond,
   pReadHandle->cur.win     = TSWINDOW_INITIALIZER;
   pReadHandle->checkFiles  = true;
   pReadHandle->activeIndex = 0;   // current active table index
-  pReadHandle->qId         = qId;
   pReadHandle->allocSize   = 0;
   pReadHandle->locateStart = false;
   pReadHandle->loadType    = pCond->type;
@@ -417,6 +374,10 @@ static STsdbReadHandle* tsdbQueryTablesImpl(STsdb* tsdb, STsdbQueryCond* pCond,
   pReadHandle->loadExternalRow = pCond->loadExternalRows;
   pReadHandle->currentLoadExternalRows = pCond->loadExternalRows;
 
+  char buf[128] = {0};
+  snprintf(buf, tListLen(buf), "TID:0x%"PRIx64" QID:0x%"PRIx64, taskId, qId);
+  pReadHandle->idStr = strdup(buf);
+
   if (tsdbInitReadH(&pReadHandle->rhelper, (STsdb*)tsdb) != 0) {
     goto _end;
   }
@@ -455,7 +416,7 @@ static STsdbReadHandle* tsdbQueryTablesImpl(STsdb* tsdb, STsdbQueryCond* pCond,
 
   pReadHandle->pDataCols = tdNewDataCols(1000, pReadHandle->pTsdb->config.maxRowsPerFileBlock);
   if (pReadHandle->pDataCols == NULL) {
-    tsdbError("%p failed to malloc buf for pDataCols, %"PRIu64, pReadHandle, pReadHandle->qId);
+    tsdbError("%p failed to malloc buf for pDataCols, %s", pReadHandle, pReadHandle->idStr);
     terrno = TSDB_CODE_TDB_OUT_OF_MEMORY;
     goto _end;
   }
@@ -471,8 +432,8 @@ static STsdbReadHandle* tsdbQueryTablesImpl(STsdb* tsdb, STsdbQueryCond* pCond,
   return NULL;
 }
 
-tsdbReadHandleT* tsdbQueryTables(STsdb* tsdb, STsdbQueryCond* pCond, STableGroupInfo* groupList, uint64_t qId, void* pRef) {
-  STsdbReadHandle* pTsdbReadHandle = tsdbQueryTablesImpl(tsdb, pCond, qId, pRef);
+tsdbReadHandleT* tsdbQueryTables(STsdb* tsdb, STsdbQueryCond* pCond, STableGroupInfo* groupList, uint64_t qId, uint64_t taskId) {
+  STsdbReadHandle* pTsdbReadHandle = tsdbQueryTablesImpl(tsdb, pCond, qId, taskId);
   if (pTsdbReadHandle == NULL) {
     return NULL;
   }
@@ -481,20 +442,17 @@ tsdbReadHandleT* tsdbQueryTables(STsdb* tsdb, STsdbQueryCond* pCond, STableGroup
     return (tsdbReadHandleT*) pTsdbReadHandle;
   }
 
-  SArray* psTable = NULL;
-
   // todo apply the lastkey of table check to avoid to load header file
-  pTsdbReadHandle->pTableCheckInfo = createCheckInfoFromTableGroup(pTsdbReadHandle, groupList, &psTable);
+  pTsdbReadHandle->pTableCheckInfo = createCheckInfoFromTableGroup(pTsdbReadHandle, groupList);
   if (pTsdbReadHandle->pTableCheckInfo == NULL) {
 //    tsdbCleanupQueryHandle(pTsdbReadHandle);
-    taosArrayDestroy(psTable);
     terrno = TSDB_CODE_TDB_OUT_OF_MEMORY;
     return NULL;
   }
 
-//  tsdbMayTakeMemSnapshot(pTsdbReadHandle, psTable);
+  tsdbDebug("%p total numOfTable:%" PRIzu " in this query, group %"PRIzu" %s", pTsdbReadHandle, taosArrayGetSize(pTsdbReadHandle->pTableCheckInfo),
+      taosArrayGetSize(groupList->pGroupList), pTsdbReadHandle->idStr);
 
-  tsdbDebug("%p total numOfTable:%" PRIzu " in query, 0x%"PRIx64, pTsdbReadHandle, taosArrayGetSize(pTsdbReadHandle->pTableCheckInfo), pTsdbReadHandle->qId);
   return (tsdbReadHandleT) pTsdbReadHandle;
 }
 
@@ -575,7 +533,7 @@ void tsdbResetQueryHandleForNewTable(tsdbReadHandleT queryHandle, STsdbQueryCond
 //  pTsdbReadHandle->next = doFreeColumnInfoData(pTsdbReadHandle->next);
 }
 
-tsdbReadHandleT tsdbQueryLastRow(STsdb *tsdb, STsdbQueryCond *pCond, STableGroupInfo *groupList, uint64_t qId, STsdbMemTable* pMemRef) {
+tsdbReadHandleT tsdbQueryLastRow(STsdb *tsdb, STsdbQueryCond *pCond, STableGroupInfo *groupList, uint64_t qId, uint64_t taskId) {
   pCond->twindow = updateLastrowForEachGroup(groupList);
 
   // no qualified table
@@ -583,7 +541,7 @@ tsdbReadHandleT tsdbQueryLastRow(STsdb *tsdb, STsdbQueryCond *pCond, STableGroup
     return NULL;
   }
 
-  STsdbReadHandle *pTsdbReadHandle = (STsdbReadHandle*) tsdbQueryTables(tsdb, pCond, groupList, qId, pMemRef);
+  STsdbReadHandle *pTsdbReadHandle = (STsdbReadHandle*) tsdbQueryTables(tsdb, pCond, groupList, qId, taskId);
   if (pTsdbReadHandle == NULL) {
     return NULL;
   }
@@ -666,7 +624,7 @@ static STableGroupInfo* trimTableGroup(STimeWindow* window, STableGroupInfo* pGr
   return pNew;
 }
 
-tsdbReadHandleT tsdbQueryRowsInExternalWindow(STsdb *tsdb, STsdbQueryCond* pCond, STableGroupInfo *groupList, uint64_t qId, STsdbMemTable* pRef) {
+tsdbReadHandleT tsdbQueryRowsInExternalWindow(STsdb *tsdb, STsdbQueryCond* pCond, STableGroupInfo *groupList, uint64_t qId, uint64_t taskId) {
   STableGroupInfo* pNew = trimTableGroup(&pCond->twindow, groupList);
 
   if (pNew->numOfTables == 0) {
@@ -681,7 +639,7 @@ tsdbReadHandleT tsdbQueryRowsInExternalWindow(STsdb *tsdb, STsdbQueryCond* pCond
     }
   }
 
-  STsdbReadHandle *pTsdbReadHandle = (STsdbReadHandle*) tsdbQueryTables(tsdb, pCond, pNew, qId, pRef);
+  STsdbReadHandle *pTsdbReadHandle = (STsdbReadHandle*) tsdbQueryTables(tsdb, pCond, pNew, qId, taskId);
   pTsdbReadHandle->loadExternalRow = true;
   pTsdbReadHandle->currentLoadExternalRows = true;
 
@@ -734,8 +692,8 @@ static bool initTableMemIterator(STsdbReadHandle* pHandle, STableCheckInfo* pChe
     SMemRow row = (SMemRow)SL_GET_NODE_DATA(node);
     TSKEY   key = memRowKey(row);  // first timestamp in buffer
     tsdbDebug("%p uid:%" PRId64 ", check data in mem from skey:%" PRId64 ", order:%d, ts range in buf:%" PRId64
-              "-%" PRId64 ", lastKey:%" PRId64 ", numOfRows:%"PRId64", 0x%"PRIx64,
-              pHandle, pCheckInfo->tableId, key, order, (*pMem)->keyMin, (*pMem)->keyMax, pCheckInfo->lastKey, (*pMem)->nrows, pHandle->qId);
+              "-%" PRId64 ", lastKey:%" PRId64 ", numOfRows:%"PRId64", %s",
+              pHandle, pCheckInfo->tableId, key, order, (*pMem)->keyMin, (*pMem)->keyMax, pCheckInfo->lastKey, (*pMem)->nrows, pHandle->idStr);
 
     if (ASCENDING_TRAVERSE(order)) {
       assert(pCheckInfo->lastKey <= key);
@@ -744,7 +702,7 @@ static bool initTableMemIterator(STsdbReadHandle* pHandle, STableCheckInfo* pChe
     }
 
   } else {
-    tsdbDebug("%p uid:%"PRId64", no data in mem, 0x%"PRIx64, pHandle, pCheckInfo->tableId, pHandle->qId);
+    tsdbDebug("%p uid:%"PRId64", no data in mem, %s", pHandle, pCheckInfo->tableId, pHandle->idStr);
   }
 
   if (!imemEmpty) {
@@ -754,8 +712,8 @@ static bool initTableMemIterator(STsdbReadHandle* pHandle, STableCheckInfo* pChe
     SMemRow row = (SMemRow)SL_GET_NODE_DATA(node);
     TSKEY   key = memRowKey(row);  // first timestamp in buffer
     tsdbDebug("%p uid:%" PRId64 ", check data in imem from skey:%" PRId64 ", order:%d, ts range in buf:%" PRId64
-              "-%" PRId64 ", lastKey:%" PRId64 ", numOfRows:%"PRId64", 0x%"PRIx64,
-              pHandle, pCheckInfo->tableId, key, order, (*pIMem)->keyMin, (*pIMem)->keyMax, pCheckInfo->lastKey, (*pIMem)->nrows, pHandle->qId);
+              "-%" PRId64 ", lastKey:%" PRId64 ", numOfRows:%"PRId64", %s",
+              pHandle, pCheckInfo->tableId, key, order, (*pIMem)->keyMin, (*pIMem)->keyMax, pCheckInfo->lastKey, (*pIMem)->nrows, pHandle->idStr);
 
     if (ASCENDING_TRAVERSE(order)) {
       assert(pCheckInfo->lastKey <= key);
@@ -763,7 +721,7 @@ static bool initTableMemIterator(STsdbReadHandle* pHandle, STableCheckInfo* pChe
       assert(pCheckInfo->lastKey >= key);
     }
   } else {
-    tsdbDebug("%p uid:%"PRId64", no data in imem, 0x%"PRIx64, pHandle, pCheckInfo->tableId, pHandle->qId);
+    tsdbDebug("%p uid:%"PRId64", no data in imem, %s", pHandle, pCheckInfo->tableId, pHandle->idStr);
   }
 
   return true;
@@ -952,8 +910,8 @@ static bool hasMoreDataInCache(STsdbReadHandle* pHandle) {
   }
 
   pCheckInfo->lastKey = memRowKey(row);  // first timestamp in buffer
-  tsdbDebug("%p uid:%" PRId64", check data in buffer from skey:%" PRId64 ", order:%d, 0x%"PRIx64, pHandle,
-      pCheckInfo->tableId,  pCheckInfo->lastKey, pHandle->order, pHandle->qId);
+  tsdbDebug("%p uid:%" PRId64", check data in buffer from skey:%" PRId64 ", order:%d, %s", pHandle,
+      pCheckInfo->tableId,  pCheckInfo->lastKey, pHandle->order, pHandle->idStr);
 
   // all data in mem are checked already.
   if ((pCheckInfo->lastKey > pHandle->window.ekey && ASCENDING_TRAVERSE(pHandle->order)) ||
@@ -1138,21 +1096,21 @@ static int32_t doLoadFileDataBlock(STsdbReadHandle* pTsdbReadHandle, SBlock* pBl
   STSchema *pSchema = metaGetTbTSchema(pTsdbReadHandle->pTsdb->pMeta, pCheckInfo->tableId, 0);
   int32_t   code = tdInitDataCols(pTsdbReadHandle->pDataCols, pSchema);
   if (code != TSDB_CODE_SUCCESS) {
-    tsdbError("%p failed to malloc buf for pDataCols, 0x%"PRIx64, pTsdbReadHandle, pTsdbReadHandle->qId);
+    tsdbError("%p failed to malloc buf for pDataCols, %s", pTsdbReadHandle, pTsdbReadHandle->idStr);
     terrno = TSDB_CODE_TDB_OUT_OF_MEMORY;
     goto _error;
   }
 
   code = tdInitDataCols(pTsdbReadHandle->rhelper.pDCols[0], pSchema);
   if (code != TSDB_CODE_SUCCESS) {
-    tsdbError("%p failed to malloc buf for rhelper.pDataCols[0], 0x%"PRIx64, pTsdbReadHandle, pTsdbReadHandle->qId);
+    tsdbError("%p failed to malloc buf for rhelper.pDataCols[0], %s", pTsdbReadHandle, pTsdbReadHandle->idStr);
     terrno = TSDB_CODE_TDB_OUT_OF_MEMORY;
     goto _error;
   }
 
   code = tdInitDataCols(pTsdbReadHandle->rhelper.pDCols[1], pSchema);
   if (code != TSDB_CODE_SUCCESS) {
-    tsdbError("%p failed to malloc buf for rhelper.pDataCols[1], 0x%"PRIx64, pTsdbReadHandle, pTsdbReadHandle->qId);
+    tsdbError("%p failed to malloc buf for rhelper.pDataCols[1], %s", pTsdbReadHandle, pTsdbReadHandle->idStr);
     terrno = TSDB_CODE_TDB_OUT_OF_MEMORY;
     goto _error;
   }
@@ -1188,15 +1146,15 @@ static int32_t doLoadFileDataBlock(STsdbReadHandle* pTsdbReadHandle, SBlock* pBl
   int64_t elapsedTime = (taosGetTimestampUs() - st);
   pTsdbReadHandle->cost.blockLoadTime += elapsedTime;
 
-  tsdbDebug("%p load file block into buffer, index:%d, brange:%"PRId64"-%"PRId64", rows:%d, elapsed time:%"PRId64 " us, 0x%"PRIx64,
-      pTsdbReadHandle, slotIndex, pBlock->keyFirst, pBlock->keyLast, pBlock->numOfRows, elapsedTime, pTsdbReadHandle->qId);
+  tsdbDebug("%p load file block into buffer, index:%d, brange:%"PRId64"-%"PRId64", rows:%d, elapsed time:%"PRId64 " us, %s",
+      pTsdbReadHandle, slotIndex, pBlock->keyFirst, pBlock->keyLast, pBlock->numOfRows, elapsedTime, pTsdbReadHandle->idStr);
   return TSDB_CODE_SUCCESS;
 
 _error:
   pBlock->numOfRows = 0;
 
-  tsdbError("%p error occurs in loading file block, index:%d, brange:%"PRId64"-%"PRId64", rows:%d, 0x%"PRIx64,
-            pTsdbReadHandle, slotIndex, pBlock->keyFirst, pBlock->keyLast, pBlock->numOfRows, pTsdbReadHandle->qId);
+  tsdbError("%p error occurs in loading file block, index:%d, brange:%"PRId64"-%"PRId64", rows:%d, %s",
+            pTsdbReadHandle, slotIndex, pBlock->keyFirst, pBlock->keyLast, pBlock->numOfRows, pTsdbReadHandle->idStr);
   return terrno;
 }
 
@@ -1219,9 +1177,9 @@ static int32_t handleDataMergeIfNeeded(STsdbReadHandle* pTsdbReadHandle, SBlock*
   key = extractFirstTraverseKey(pCheckInfo, pTsdbReadHandle->order, pCfg->update);
 
   if (key != TSKEY_INITIAL_VAL) {
-    tsdbDebug("%p key in mem:%"PRId64", 0x%"PRIx64, pTsdbReadHandle, key, pTsdbReadHandle->qId);
+    tsdbDebug("%p key in mem:%"PRId64", %s", pTsdbReadHandle, key, pTsdbReadHandle->idStr);
   } else {
-    tsdbDebug("%p no data in mem, 0x%"PRIx64, pTsdbReadHandle, pTsdbReadHandle->qId);
+    tsdbDebug("%p no data in mem, %s", pTsdbReadHandle, pTsdbReadHandle->idStr);
   }
 
   if ((ASCENDING_TRAVERSE(pTsdbReadHandle->order) && (key != TSKEY_INITIAL_VAL && key <= binfo.window.ekey)) ||
@@ -1288,11 +1246,11 @@ static int32_t handleDataMergeIfNeeded(STsdbReadHandle* pTsdbReadHandle, SBlock*
 
     assert(cur->blockCompleted);
     if (cur->rows == binfo.rows) {
-      tsdbDebug("%p whole file block qualified, brange:%"PRId64"-%"PRId64", rows:%d, lastKey:%"PRId64", %"PRIx64,
-                pTsdbReadHandle, cur->win.skey, cur->win.ekey, cur->rows, cur->lastKey, pTsdbReadHandle->qId);
+      tsdbDebug("%p whole file block qualified, brange:%"PRId64"-%"PRId64", rows:%d, lastKey:%"PRId64", %s",
+                pTsdbReadHandle, cur->win.skey, cur->win.ekey, cur->rows, cur->lastKey, pTsdbReadHandle->idStr);
     } else {
-      tsdbDebug("%p create data block from remain file block, brange:%"PRId64"-%"PRId64", rows:%d, total:%d, lastKey:%"PRId64", %"PRIx64,
-                pTsdbReadHandle, cur->win.skey, cur->win.ekey, cur->rows, binfo.rows, cur->lastKey, pTsdbReadHandle->qId);
+      tsdbDebug("%p create data block from remain file block, brange:%"PRId64"-%"PRId64", rows:%d, total:%d, lastKey:%"PRId64", %s",
+                pTsdbReadHandle, cur->win.skey, cur->win.ekey, cur->rows, binfo.rows, cur->lastKey, pTsdbReadHandle->idStr);
     }
 
   }
@@ -1815,8 +1773,8 @@ static void copyAllRemainRowsFromFileBlock(STsdbReadHandle* pTsdbReadHandle, STa
   updateInfoAfterMerge(pTsdbReadHandle, pCheckInfo, numOfRows, pos);
   doCheckGeneratedBlockRange(pTsdbReadHandle);
 
-  tsdbDebug("%p uid:%" PRIu64", data block created, mixblock:%d, brange:%"PRIu64"-%"PRIu64" rows:%d, 0x%"PRIx64,
-            pTsdbReadHandle, pCheckInfo->tableId, cur->mixBlock, cur->win.skey, cur->win.ekey, cur->rows, pTsdbReadHandle->qId);
+  tsdbDebug("%p uid:%" PRIu64", data block created, mixblock:%d, brange:%"PRIu64"-%"PRIu64" rows:%d, %s",
+            pTsdbReadHandle, pCheckInfo->tableId, cur->mixBlock, cur->win.skey, cur->win.ekey, cur->rows, pTsdbReadHandle->idStr);
 }
 
 int32_t getEndPosInDataBlock(STsdbReadHandle* pTsdbReadHandle, SDataBlockInfo* pBlockInfo) {
@@ -1868,9 +1826,9 @@ static void doMergeTwoLevelData(STsdbReadHandle* pTsdbReadHandle, STableCheckInf
   int32_t endPos = getEndPosInDataBlock(pTsdbReadHandle, &blockInfo);
 
   tsdbDebug("%p uid:%" PRIu64" start merge data block, file block range:%"PRIu64"-%"PRIu64" rows:%d, start:%d,"
-            "end:%d, 0x%"PRIx64,
+            "end:%d, %s",
             pTsdbReadHandle, pCheckInfo->tableId, blockInfo.window.skey, blockInfo.window.ekey,
-            blockInfo.rows, cur->pos, endPos, pTsdbReadHandle->qId);
+            blockInfo.rows, cur->pos, endPos, pTsdbReadHandle->idStr);
 
   // compared with the data from in-memory buffer, to generate the correct timestamp array list
   int32_t numOfRows = 0;
@@ -2027,8 +1985,8 @@ static void doMergeTwoLevelData(STsdbReadHandle* pTsdbReadHandle, STableCheckInf
   updateInfoAfterMerge(pTsdbReadHandle, pCheckInfo, numOfRows, pos);
   doCheckGeneratedBlockRange(pTsdbReadHandle);
 
-  tsdbDebug("%p uid:%" PRIu64", data block created, mixblock:%d, brange:%"PRIu64"-%"PRIu64" rows:%d, 0x%"PRIx64,
-      pTsdbReadHandle, pCheckInfo->tableId, cur->mixBlock, cur->win.skey, cur->win.ekey, cur->rows, pTsdbReadHandle->qId);
+  tsdbDebug("%p uid:%" PRIu64", data block created, mixblock:%d, brange:%"PRIu64"-%"PRIu64" rows:%d, %s",
+      pTsdbReadHandle, pCheckInfo->tableId, cur->mixBlock, cur->win.skey, cur->win.ekey, cur->rows, pTsdbReadHandle->idStr);
 }
 
 int32_t binarySearchForKey(char* pValue, int num, TSKEY key, int order) {
@@ -2202,13 +2160,13 @@ static int32_t createDataBlocksInfo(STsdbReadHandle* pTsdbReadHandle, int32_t nu
     memcpy(pTsdbReadHandle->pDataBlockInfo, sup.pDataBlockInfo[0], sizeof(STableBlockInfo) * numOfBlocks);
     cleanBlockOrderSupporter(&sup, numOfQualTables);
 
-    tsdbDebug("%p create data blocks info struct completed for 1 table, %d blocks not sorted 0x%"PRIx64, pTsdbReadHandle, cnt,
-        pTsdbReadHandle->qId);
+    tsdbDebug("%p create data blocks info struct completed for 1 table, %d blocks not sorted %s", pTsdbReadHandle, cnt,
+        pTsdbReadHandle->idStr);
     return TSDB_CODE_SUCCESS;
   }
 
-  tsdbDebug("%p create data blocks info struct completed, %d blocks in %d tables 0x%"PRIx64, pTsdbReadHandle, cnt,
-      numOfQualTables, pTsdbReadHandle->qId);
+  tsdbDebug("%p create data blocks info struct completed, %d blocks in %d tables %s", pTsdbReadHandle, cnt,
+      numOfQualTables, pTsdbReadHandle->idStr);
 
   assert(cnt <= numOfBlocks && numOfQualTables <= numOfTables);  // the pTableQueryInfo[j]->numOfBlocks may be 0
   sup.numOfTables = numOfQualTables;
@@ -2244,7 +2202,7 @@ static int32_t createDataBlocksInfo(STsdbReadHandle* pTsdbReadHandle, int32_t nu
    * }
    */
 
-  tsdbDebug("%p %d data blocks sort completed, 0x%"PRIx64, pTsdbReadHandle, cnt, pTsdbReadHandle->qId);
+  tsdbDebug("%p %d data blocks sort completed, %s", pTsdbReadHandle, cnt, pTsdbReadHandle->idStr);
   cleanBlockOrderSupporter(&sup, numOfTables);
   free(pTree);
 
@@ -2302,8 +2260,8 @@ static int32_t getFirstFileDataBlock(STsdbReadHandle* pTsdbReadHandle, bool* exi
     if ((ASCENDING_TRAVERSE(pTsdbReadHandle->order) && win.skey > pTsdbReadHandle->window.ekey) ||
         (!ASCENDING_TRAVERSE(pTsdbReadHandle->order) && win.ekey < pTsdbReadHandle->window.ekey)) {
       tsdbUnLockFS(REPO_FS(pTsdbReadHandle->pTsdb));
-      tsdbDebug("%p remain files are not qualified for qrange:%" PRId64 "-%" PRId64 ", ignore, 0x%"PRIx64, pTsdbReadHandle,
-                pTsdbReadHandle->window.skey, pTsdbReadHandle->window.ekey, pTsdbReadHandle->qId);
+      tsdbDebug("%p remain files are not qualified for qrange:%" PRId64 "-%" PRId64 ", ignore, %s", pTsdbReadHandle,
+                pTsdbReadHandle->window.skey, pTsdbReadHandle->window.ekey, pTsdbReadHandle->idStr);
       pTsdbReadHandle->pFileGroup = NULL;
       assert(pTsdbReadHandle->numOfBlocks == 0);
       break;
@@ -2326,8 +2284,8 @@ static int32_t getFirstFileDataBlock(STsdbReadHandle* pTsdbReadHandle, bool* exi
       break;
     }
 
-    tsdbDebug("%p %d blocks found in file for %d table(s), fid:%d, 0x%"PRIx64, pTsdbReadHandle, numOfBlocks, numOfTables,
-              pTsdbReadHandle->pFileGroup->fid, pTsdbReadHandle->qId);
+    tsdbDebug("%p %d blocks found in file for %d table(s), fid:%d, %s", pTsdbReadHandle, numOfBlocks, numOfTables,
+              pTsdbReadHandle->pFileGroup->fid, pTsdbReadHandle->idStr);
 
     assert(numOfBlocks >= 0);
     if (numOfBlocks == 0) {
@@ -2420,8 +2378,8 @@ int32_t tsdbGetFileBlocksDistInfo(tsdbReadHandleT* queryHandle, STableBlockDist*
     if ((ASCENDING_TRAVERSE(pTsdbReadHandle->order) && win.skey > pTsdbReadHandle->window.ekey) ||
     (!ASCENDING_TRAVERSE(pTsdbReadHandle->order) && win.ekey < pTsdbReadHandle->window.ekey)) {
       tsdbUnLockFS(REPO_FS(pTsdbReadHandle->pTsdb));
-      tsdbDebug("%p remain files are not qualified for qrange:%" PRId64 "-%" PRId64 ", ignore, 0x%"PRIx64, pTsdbReadHandle,
-                pTsdbReadHandle->window.skey, pTsdbReadHandle->window.ekey, pTsdbReadHandle->qId);
+      tsdbDebug("%p remain files are not qualified for qrange:%" PRId64 "-%" PRId64 ", ignore, %s", pTsdbReadHandle,
+                pTsdbReadHandle->window.skey, pTsdbReadHandle->window.ekey, pTsdbReadHandle->idStr);
       pTsdbReadHandle->pFileGroup = NULL;
       break;
     }
@@ -2444,8 +2402,8 @@ int32_t tsdbGetFileBlocksDistInfo(tsdbReadHandleT* queryHandle, STableBlockDist*
       break;
     }
 
-    tsdbDebug("%p %d blocks found in file for %d table(s), fid:%d, 0x%"PRIx64, pTsdbReadHandle, numOfBlocks, numOfTables,
-              pTsdbReadHandle->pFileGroup->fid, pTsdbReadHandle->qId);
+    tsdbDebug("%p %d blocks found in file for %d table(s), fid:%d, %s", pTsdbReadHandle, numOfBlocks, numOfTables,
+              pTsdbReadHandle->pFileGroup->fid, pTsdbReadHandle->idStr);
 
     if (numOfBlocks == 0) {
       continue;
@@ -2499,8 +2457,8 @@ static int32_t getDataBlocksInFiles(STsdbReadHandle* pTsdbReadHandle, bool* exis
     if ((!cur->mixBlock) || cur->blockCompleted) {
       // all data blocks in current file has been checked already, try next file if exists
     } else {
-      tsdbDebug("%p continue in current data block, index:%d, pos:%d, 0x%"PRIx64, pTsdbReadHandle, cur->slot, cur->pos,
-                pTsdbReadHandle->qId);
+      tsdbDebug("%p continue in current data block, index:%d, pos:%d, %s", pTsdbReadHandle, cur->slot, cur->pos,
+                pTsdbReadHandle->idStr);
       int32_t code = handleDataMergeIfNeeded(pTsdbReadHandle, pBlockInfo->compBlock, pCheckInfo);
       *exists = (pTsdbReadHandle->realNumOfRows > 0);
 
@@ -2624,8 +2582,8 @@ static int tsdbReadRowsFromCache(STableCheckInfo* pCheckInfo, TSKEY maxKey, int
   }
 
   int64_t elapsedTime = taosGetTimestampUs() - st;
-  tsdbDebug("%p build data block from cache completed, elapsed time:%"PRId64" us, numOfRows:%d, numOfCols:%d, 0x%"PRIx64, pTsdbReadHandle,
-            elapsedTime, numOfRows, numOfCols, pTsdbReadHandle->qId);
+  tsdbDebug("%p build data block from cache completed, elapsed time:%"PRId64" us, numOfRows:%d, numOfCols:%d, %s", pTsdbReadHandle,
+            elapsedTime, numOfRows, numOfCols, pTsdbReadHandle->idStr);
 
   return numOfRows;
 }
@@ -2639,7 +2597,7 @@ static int32_t getAllTableList(SMeta* pMeta, uint64_t uid, SArray* list) {
       break;
     }
 
-    STableKeyInfo info = {.pTable = NULL, .lastKey = TSKEY_INITIAL_VAL, uid = id};
+    STableKeyInfo info = {.lastKey = TSKEY_INITIAL_VAL, uid = id};
     taosArrayPush(list, &info);
   }
 
@@ -2937,7 +2895,7 @@ bool tsdbNextDataBlock(tsdbReadHandleT pHandle) {
   STsdbReadHandle* pTsdbReadHandle = (STsdbReadHandle*) pHandle;
 
   if (emptyQueryTimewindow(pTsdbReadHandle)) {
-    tsdbDebug("%p query window not overlaps with the data set, no result returned, 0x%"PRIx64, pTsdbReadHandle, pTsdbReadHandle->qId);
+    tsdbDebug("%p query window not overlaps with the data set, no result returned, %s", pTsdbReadHandle, pTsdbReadHandle->idStr);
     return false;
   }
 
@@ -3052,7 +3010,7 @@ bool tsdbNextDataBlock(tsdbReadHandleT pHandle) {
 //    memcpy(&cond.colList[i], &pColInfoData->info, sizeof(SColumnInfo));
 //  }
 //
-//  pSecQueryHandle = tsdbQueryTablesImpl(pTsdbReadHandle->pTsdb, &cond, pTsdbReadHandle->qId, pMemRef);
+//  pSecQueryHandle = tsdbQueryTablesImpl(pTsdbReadHandle->pTsdb, &cond, pTsdbReadHandle->idStr, pMemRef);
 //  tfree(cond.colList);
 //
 //  // current table, only one table
@@ -3230,7 +3188,7 @@ STimeWindow updateLastrowForEachGroup(STableGroupInfo *groupList) {
       if (key < lastKey) {
         key = lastKey;
 
-        keyInfo.pTable  = pInfo->pTable;
+//        keyInfo.pTable  = pInfo->pTable;
         keyInfo.lastKey = key;
         pInfo->lastKey  = key;
 
@@ -3244,29 +3202,19 @@ STimeWindow updateLastrowForEachGroup(STableGroupInfo *groupList) {
       }
     }
 
-    // clear current group, unref unused table
-    for (int32_t i = 0; i < numOfTables; ++i) {
-      STableKeyInfo* pInfo = (STableKeyInfo*)taosArrayGet(pGroup, i);
-
-      // keyInfo.pTable may be NULL here.
-      if (pInfo->pTable != keyInfo.pTable) {
-//        tsdbUnRefTable(pInfo->pTable);
-      }
-    }
-
     // more than one table in each group, only one table left for each group
-    if (keyInfo.pTable != NULL) {
-      totalNumOfTable++;
-      if (taosArrayGetSize(pGroup) == 1) {
-        // do nothing
-      } else {
-        taosArrayClear(pGroup);
-        taosArrayPush(pGroup, &keyInfo);
-      }
-    } else {  // mark all the empty groups, and remove it later
-      taosArrayDestroy(pGroup);
-      taosArrayPush(emptyGroup, &j);
-    }
+//    if (keyInfo.pTable != NULL) {
+//      totalNumOfTable++;
+//      if (taosArrayGetSize(pGroup) == 1) {
+//        // do nothing
+//      } else {
+//        taosArrayClear(pGroup);
+//        taosArrayPush(pGroup, &keyInfo);
+//      }
+//    } else {  // mark all the empty groups, and remove it later
+//      taosArrayDestroy(pGroup);
+//      taosArrayPush(emptyGroup, &j);
+//    }
   }
 
   // window does not being updated, so set the original
@@ -3456,11 +3404,13 @@ void filterPrepare(void* expr, void* param) {
   }
 }
 
+#endif
 
 static int32_t tableGroupComparFn(const void *p1, const void *p2, const void *param) {
+#if 0
   STableGroupSupporter* pTableGroupSupp = (STableGroupSupporter*) param;
-  STable* pTable1 = ((STableKeyInfo*) p1)->pTable;
-  STable* pTable2 = ((STableKeyInfo*) p2)->pTable;
+  STable* pTable1 = ((STableKeyInfo*) p1)->uid;
+  STable* pTable2 = ((STableKeyInfo*) p2)->uid;
 
   for (int32_t i = 0; i < pTableGroupSupp->numOfCols; ++i) {
     SColIndex* pColIndex = &pTableGroupSupp->pCols[i];
@@ -3508,10 +3458,9 @@ static int32_t tableGroupComparFn(const void *p1, const void *p2, const void *pa
       return ret;
     }
   }
-
+#endif
   return 0;
 }
-#endif
 
 static int tsdbCheckInfoCompar(const void* key1, const void* key2) {
   if (((STableCheckInfo*)key1)->tableId < ((STableCheckInfo*)key2)->tableId) {
@@ -3527,10 +3476,9 @@ static int tsdbCheckInfoCompar(const void* key1, const void* key2) {
 void createTableGroupImpl(SArray* pGroups, SArray* pTableList, size_t numOfTables, TSKEY skey,
                           STableGroupSupporter* pSupp, __ext_compar_fn_t compareFn) {
   STable* pTable = taosArrayGetP(pTableList, 0);
-
   SArray* g = taosArrayInit(16, sizeof(STableKeyInfo));
 
-  STableKeyInfo info = {.pTable = pTable, .lastKey = skey};
+  STableKeyInfo info = {.lastKey = skey};
   taosArrayPush(g, &info);
 
   for (int32_t i = 1; i < numOfTables; ++i) {
@@ -3541,13 +3489,13 @@ void createTableGroupImpl(SArray* pGroups, SArray* pTableList, size_t numOfTable
     assert(ret == 0 || ret == -1);
 
     if (ret == 0) {
-      STableKeyInfo info1 = {.pTable = *p, .lastKey = skey};
+      STableKeyInfo info1 = {.lastKey = skey};
       taosArrayPush(g, &info1);
     } else {
       taosArrayPush(pGroups, &g);  // current group is ended, start a new group
       g = taosArrayInit(16, sizeof(STableKeyInfo));
 
-      STableKeyInfo info1 = {.pTable = *p, .lastKey = skey};
+      STableKeyInfo info1 = {.lastKey = skey};
       taosArrayPush(g, &info1);
     }
   }
@@ -3580,8 +3528,8 @@ SArray* createTableGroup(SArray* pTableList, SSchemaWrapper* pTagSchema, SColInd
     sup.pTagSchema = pTagSchema->pSchema;
     sup.pCols = pCols;
 
-//    taosqsort(pTableList->pData, size, sizeof(STableKeyInfo), &sup, tableGroupComparFn);
-//    createTableGroupImpl(pTableGroup, pTableList, size, skey, &sup, tableGroupComparFn);
+    taosqsort(pTableList->pData, size, sizeof(STableKeyInfo), &sup, tableGroupComparFn);
+    createTableGroupImpl(pTableGroup, pTableList, size, skey, &sup, tableGroupComparFn);
   }
 
   return pTableGroup;
@@ -3689,16 +3637,16 @@ SArray* createTableGroup(SArray* pTableList, SSchemaWrapper* pTagSchema, SColInd
 
 int32_t tsdbQuerySTableByTagCond(STsdb* tsdb, uint64_t uid, TSKEY skey, const char* pTagCond, size_t len,
                                  int16_t tagNameRelType, const char* tbnameCond, STableGroupInfo* pGroupInfo,
-                                 SColIndex* pColIndex, int32_t numOfCols, uint64_t reqId) {
+                                 SColIndex* pColIndex, int32_t numOfCols, uint64_t reqId, uint64_t taskId) {
   STbCfg* pTbCfg = metaGetTbInfoByUid(tsdb->pMeta, uid);
   if (pTbCfg == NULL) {
-    tsdbError("%p failed to get stable, uid:%"PRIu64", reqId:0x%"PRIx64, tsdb, uid, reqId);
+    tsdbError("%p failed to get stable, uid:%"PRIu64", TID:0x%"PRIx64" QID:0x%"PRIx64, tsdb, uid, taskId, reqId);
     terrno = TSDB_CODE_TDB_INVALID_TABLE_ID;
     goto _error;
   }
 
   if (pTbCfg->type != META_SUPER_TABLE) {
-    tsdbError("%p query normal tag not allowed, uid:%" PRIu64 ", reId:0x%"PRIx64, tsdb, uid, reqId);
+    tsdbError("%p query normal tag not allowed, uid:%" PRIu64 ", TID:0x%"PRIx64" QID:0x%"PRIx64, tsdb, uid, taskId, reqId);
     terrno = TSDB_CODE_OPS_NOT_SUPPORT; //basically, this error is caused by invalid sql issued by client
     goto _error;
   }
@@ -3717,8 +3665,8 @@ int32_t tsdbQuerySTableByTagCond(STsdb* tsdb, uint64_t uid, TSKEY skey, const ch
     pGroupInfo->numOfTables = (uint32_t) taosArrayGetSize(res);
     pGroupInfo->pGroupList  = createTableGroup(res, pTagSchema, pColIndex, numOfCols, skey);
 
-    tsdbDebug("%p no table name/tag condition, all tables qualified, numOfTables:%u, group:%zu", tsdb,
-              pGroupInfo->numOfTables, taosArrayGetSize(pGroupInfo->pGroupList));
+    tsdbDebug("%p no table name/tag condition, all tables qualified, numOfTables:%u, group:%zu, TID:0x%"PRIx64" QID:0x%"PRIx64, tsdb,
+              pGroupInfo->numOfTables, taosArrayGetSize(pGroupInfo->pGroupList), taskId, reqId);
 
     taosArrayDestroy(res);
     return ret;
@@ -3891,7 +3839,7 @@ void tsdbCleanupQueryHandle(tsdbReadHandleT queryHandle) {
   tfree(pTsdbReadHandle->statis);
 
   if (!emptyQueryTimewindow(pTsdbReadHandle)) {
-    tsdbMayUnTakeMemSnapshot(pTsdbReadHandle);
+//    tsdbMayUnTakeMemSnapshot(pTsdbReadHandle);
   } else {
     assert(pTsdbReadHandle->pTableCheckInfo == NULL);
   }
@@ -3910,8 +3858,8 @@ void tsdbCleanupQueryHandle(tsdbReadHandleT queryHandle) {
 
   SIOCostSummary* pCost = &pTsdbReadHandle->cost;
 
-  tsdbDebug("%p :io-cost summary: head-file read cnt:%"PRIu64", head-file time:%"PRIu64" us, statis-info:%"PRId64" us, datablock:%" PRId64" us, check data:%"PRId64" us, 0x%"PRIx64,
-      pTsdbReadHandle, pCost->headFileLoad, pCost->headFileLoadTime, pCost->statisInfoLoadTime, pCost->blockLoadTime, pCost->checkForNextTime, pTsdbReadHandle->qId);
+  tsdbDebug("%p :io-cost summary: head-file read cnt:%"PRIu64", head-file time:%"PRIu64" us, statis-info:%"PRId64" us, datablock:%" PRId64" us, check data:%"PRId64" us, %s",
+      pTsdbReadHandle, pCost->headFileLoad, pCost->headFileLoadTime, pCost->statisInfoLoadTime, pCost->blockLoadTime, pCost->checkForNextTime, pTsdbReadHandle->idStr);
 
   tfree(pTsdbReadHandle);
 }

source/dnode/vnode/src/vnd/vnodeMgr.c

@@ -26,6 +26,7 @@ int vnodeInit(const SVnodeOpt *pOption) {
 
   vnodeMgr.stop = false;
   vnodeMgr.putReqToVQueryQFp = pOption->putReqToVQueryQFp;
+  vnodeMgr.sendReqToDnodeFp = pOption->sendReqToDnodeFp;
 
   // Start commit handers
   if (pOption->nthreads > 0) {
@@ -96,8 +97,14 @@ int32_t vnodePutReqToVQueryQ(SVnode* pVnode, struct SRpcMsg* pReq) {
   return (*vnodeMgr.putReqToVQueryQFp)(pVnode->pDnode, pReq);
 }
 
+void vnodeSendReqToDnode(SVnode* pVnode, struct SEpSet* epSet, struct SRpcMsg* pReq) {
+  (*vnodeMgr.sendReqToDnodeFp)(pVnode->pDnode, epSet, pReq);
+}
+
 /* ------------------------ STATIC METHODS ------------------------ */
 static void* loop(void* arg) {
+  setThreadName("vnode-commit");
+
   SVnodeTask* pTask;
   for (;;) {
     pthread_mutex_lock(&(vnodeMgr.mutex));

source/dnode/vnode/src/vnd/vnodeQuery.c

@@ -17,14 +17,14 @@
 #include "vnd.h"
 
 static int32_t vnodeGetTableList(SVnode *pVnode, SRpcMsg *pMsg);
-static int     vnodeGetTableMeta(SVnode *pVnode, SRpcMsg *pMsg, SRpcMsg **pRsp);
+static int     vnodeGetTableMeta(SVnode *pVnode, SRpcMsg *pMsg);
 
 int vnodeQueryOpen(SVnode *pVnode) {
   return qWorkerInit(NODE_TYPE_VNODE, pVnode->vgId, NULL, (void **)&pVnode->pQuery, pVnode,
-                     (putReqToQueryQFp)vnodePutReqToVQueryQ);
+                     (putReqToQueryQFp)vnodePutReqToVQueryQ, (sendReqToDnodeFp)vnodeSendReqToDnode);
 }
 
-int vnodeProcessQueryReq(SVnode *pVnode, SRpcMsg *pMsg, SRpcMsg **pRsp) {
+int vnodeProcessQueryMsg(SVnode *pVnode, SRpcMsg *pMsg) {
   vTrace("message in query queue is processing");
 
   switch (pMsg->msgType) {
@@ -38,11 +38,13 @@ int vnodeProcessQueryReq(SVnode *pVnode, SRpcMsg *pMsg, SRpcMsg **pRsp) {
   }
 }
 
-int vnodeProcessFetchReq(SVnode *pVnode, SRpcMsg *pMsg, SRpcMsg **pRsp) {
+int vnodeProcessFetchMsg(SVnode *pVnode, SRpcMsg *pMsg) {
   vTrace("message in fetch queue is processing");
   switch (pMsg->msgType) {
     case TDMT_VND_FETCH:
       return qWorkerProcessFetchMsg(pVnode, pVnode->pQuery, pMsg);
+    case TDMT_VND_FETCH_RSP:
+      return qWorkerProcessFetchRsp(pVnode, pVnode->pQuery, pMsg);
     case TDMT_VND_RES_READY:
       return qWorkerProcessReadyMsg(pVnode, pVnode->pQuery, pMsg);
     case TDMT_VND_TASKS_STATUS:
@@ -57,24 +59,24 @@ int vnodeProcessFetchReq(SVnode *pVnode, SRpcMsg *pMsg, SRpcMsg **pRsp) {
       return vnodeGetTableList(pVnode, pMsg);
       //      return qWorkerProcessShowFetchMsg(pVnode->pMeta, pVnode->pQuery, pMsg);
     case TDMT_VND_TABLE_META:
-      return vnodeGetTableMeta(pVnode, pMsg, pRsp);
+      return vnodeGetTableMeta(pVnode, pMsg);
     case TDMT_VND_CONSUME:
-      return tqProcessConsumeReq(pVnode->pTq, pMsg, pRsp);
+      return tqProcessConsumeReq(pVnode->pTq, pMsg);
     default:
       vError("unknown msg type:%d in fetch queue", pMsg->msgType);
       return TSDB_CODE_VND_APP_ERROR;
   }
 }
 
-static int vnodeGetTableMeta(SVnode *pVnode, SRpcMsg *pMsg, SRpcMsg **pRsp) {
+static int vnodeGetTableMeta(SVnode *pVnode, SRpcMsg *pMsg) {
   STableInfoReq * pReq = (STableInfoReq *)(pMsg->pCont);
   STbCfg *        pTbCfg = NULL;
   STbCfg *        pStbCfg = NULL;
   tb_uid_t        uid;
   int32_t         nCols;
   int32_t         nTagCols;
-  SSchemaWrapper *pSW;
-  STableMetaRsp * pTbMetaMsg = NULL;
+  SSchemaWrapper *pSW = NULL;
+  STableMetaRsp  *pTbMetaMsg = NULL;
   SSchema *       pTagSchema;
   SRpcMsg         rpcMsg;
   int             msgLen = 0;
@@ -145,15 +147,22 @@ static int vnodeGetTableMeta(SVnode *pVnode, SRpcMsg *pMsg, SRpcMsg **pRsp) {
 
 _exit:
 
-  free(pSW->pSchema);
-  free(pSW);
-  free(pTbCfg->name);
-  free(pTbCfg);
-  if (pTbCfg->type == META_SUPER_TABLE) {
-    free(pTbCfg->stbCfg.pTagSchema);
-  } else if (pTbCfg->type == META_SUPER_TABLE) {
-    kvRowFree(pTbCfg->ctbCfg.pTag);
+  if (pSW != NULL) {
+    tfree(pSW->pSchema);
+    tfree(pSW);
   }
+
+  if (pTbCfg) {
+    tfree(pTbCfg->name);
+    if (pTbCfg->type == META_SUPER_TABLE) {
+      free(pTbCfg->stbCfg.pTagSchema);
+    } else if (pTbCfg->type == META_SUPER_TABLE) {
+      kvRowFree(pTbCfg->ctbCfg.pTag);
+    }
+
+    tfree(pTbCfg);
+  }
+
   rpcMsg.handle = pMsg->handle;
   rpcMsg.ahandle = pMsg->ahandle;
   rpcMsg.pCont = pTbMetaMsg;

source/dnode/vnode/src/vnd/vnodeWrite.c

@@ -115,7 +115,7 @@ int vnodeApplyWMsg(SVnode *pVnode, SRpcMsg *pMsg, SRpcMsg **pRsp) {
       }
       break;
     case TDMT_VND_MQ_SET_CONN: {
-      if (tqProcessSetConnReq(pVnode->pTq, POINTER_SHIFT(ptr, sizeof(SMsgHead)), NULL) < 0) {
+      if (tqProcessSetConnReq(pVnode->pTq, POINTER_SHIFT(ptr, sizeof(SMsgHead))) < 0) {
         // TODO: handle error
       }
     } break;

source/libs/executor/inc/executil.h

@@ -38,7 +38,7 @@
 #define GET_RES_WINDOW_KEY_LEN(_l) ((_l) + sizeof(uint64_t))
 #define GET_RES_EXT_WINDOW_KEY_LEN(_l) ((_l) + sizeof(uint64_t) + POINTER_BYTES)
 
-#define GET_TASKID(_t)  (((SExecTaskInfo*)(_t))->id.queryId)
+#define GET_TASKID(_t)  (((SExecTaskInfo*)(_t))->id.str)
 
 #define curTimeWindowIndex(_winres)        ((_winres)->curIndex)
 

source/libs/executor/inc/executorimpl.h

@@ -239,7 +239,7 @@ typedef struct STaskIdInfo {
   uint64_t       queryId;    // this is also a request id
   uint64_t       subplanId;
   uint64_t       templateId;
-  uint64_t       taskId;     // this is a subplan id
+  char          *str;
 } STaskIdInfo;
 
 typedef struct SExecTaskInfo {
@@ -250,9 +250,8 @@ typedef struct SExecTaskInfo {
   STaskCostInfo   cost;
   int64_t         owner;       // if it is in execution
   int32_t         code;
-
+  uint64_t        totalRows;   // total number of rows
   STableGroupInfo tableqinfoGroupInfo;  // this is a group array list, including SArray<STableQueryInfo*> structure
-  pthread_mutex_t lock;        // used to synchronize the rsp/query threads
   char           *sql;         // query sql string
   jmp_buf         env;         //
   struct SOperatorInfo  *pRoot;
@@ -378,8 +377,10 @@ typedef struct SExchangeInfo {
   SSDataBlock       *pResult;
   int32_t            current;
   uint64_t           rowsOfCurrentSource;
-  uint64_t           bytes;   // total load bytes from remote
-  uint64_t           totalRows;
+
+  uint64_t           totalSize;   // total load bytes from remote
+  uint64_t           totalRows;   // total number of rows
+  uint64_t           totalElapsed;// total elapsed time
 } SExchangeInfo;
 
 typedef struct STableScanInfo {
@@ -622,8 +623,6 @@ int32_t createIndirectQueryFuncExprFromMsg(SQueryTableReq *pQueryMsg, int32_t nu
 int32_t createQueryFilter(char *data, uint16_t len, SFilterInfo** pFilters);
 
 SGroupbyExpr *createGroupbyExprFromMsg(SQueryTableReq *pQueryMsg, SColIndex *pColIndex, int32_t *code);
-SQInfo *createQInfoImpl(SQueryTableReq *pQueryMsg, SGroupbyExpr *pGroupbyExpr, SExprInfo *pExprs,
-                        SExprInfo *pSecExprs, STableGroupInfo *pTableGroupInfo, SColumnInfo* pTagCols, SFilterInfo* pFilters, int32_t vgId, char* sql, uint64_t qId, struct SUdfInfo* pUdfInfo);
 
 int32_t initQInfo(STsBufInfo* pTsBufInfo, void* tsdb, void* sourceOptr, SQInfo* pQInfo, STaskParam* param, char* start,
                   int32_t prevResultLen, void* merger);
@@ -639,32 +638,29 @@ int32_t buildArithmeticExprFromMsg(SExprInfo *pArithExprInfo, void *pQueryMsg);
 bool isTaskKilled(SExecTaskInfo *pTaskInfo);
 int32_t checkForQueryBuf(size_t numOfTables);
 bool checkNeedToCompressQueryCol(SQInfo *pQInfo);
-bool doBuildResCheck(SQInfo* pQInfo);
 void setQueryStatus(STaskRuntimeEnv *pRuntimeEnv, int8_t status);
 
 bool onlyQueryTags(STaskAttr* pQueryAttr);
 //void destroyUdfInfo(struct SUdfInfo* pUdfInfo);
 
-bool isValidQInfo(void *param);
-
 int32_t doDumpQueryResult(SQInfo *pQInfo, char *data, int8_t compressed, int32_t *compLen);
 
 size_t getResultSize(SQInfo *pQInfo, int64_t *numOfRows);
-void setQueryKilled(SQInfo *pQInfo);
+void setTaskKilled(SExecTaskInfo *pTaskInfo);
 
 void publishOperatorProfEvent(SOperatorInfo* operatorInfo, EQueryProfEventType eventType);
 void publishQueryAbortEvent(SExecTaskInfo * pTaskInfo, int32_t code);
 
 void calculateOperatorProfResults(SQInfo* pQInfo);
-void queryCostStatis(SQInfo *pQInfo);
+void queryCostStatis(SExecTaskInfo *pTaskInfo);
 
-void doDestroyTask(SQInfo *pQInfo);
+void doDestroyTask(SExecTaskInfo *pTaskInfo);
 void freeQueryAttr(STaskAttr *pQuery);
 
 int32_t getMaximumIdleDurationSec();
 
 void doInvokeUdf(struct SUdfInfo* pUdfInfo, SQLFunctionCtx *pCtx, int32_t idx, int32_t type);
 void setTaskStatus(SExecTaskInfo *pTaskInfo, int8_t status);
-int32_t createExecTaskInfoImpl(SSubplan* pPlan, SExecTaskInfo** pTaskInfo, void* readerHandle);
+int32_t createExecTaskInfoImpl(SSubplan* pPlan, SExecTaskInfo** pTaskInfo, void* readerHandle, uint64_t taskId);
 
 #endif  // TDENGINE_EXECUTORIMPL_H

source/libs/executor/src/dataDispatcher.c

@@ -121,11 +121,19 @@ static void toDataCacheEntry(const SDataDispatchHandle* pHandle, const SInputDat
 }
 
 static bool allocBuf(SDataDispatchHandle* pDispatcher, const SInputData* pInput, SDataDispatchBuf* pBuf) {
-  if (taosQueueSize(pDispatcher->pDataBlocks) >= pDispatcher->pManager->cfg.maxDataBlockNumPerQuery) {
+  uint32_t capacity = pDispatcher->pManager->cfg.maxDataBlockNumPerQuery;
+  if (taosQueueSize(pDispatcher->pDataBlocks) > capacity) {
+    qError("SinkNode queue is full, no capacity, max:%d, current:%d, no capacity", capacity,
+           taosQueueSize(pDispatcher->pDataBlocks));
     return false;
   }
+
   pBuf->allocSize = DATA_META_LENGTH(pInput->pTableRetrieveTsMap) + pDispatcher->schema.resultRowSize * pInput->pData->info.rows;
   pBuf->pData = malloc(pBuf->allocSize);
+  if (pBuf->pData == NULL) {
+    qError("SinkNode failed to malloc memory, size:%d, code:%d", pBuf->allocSize, TAOS_SYSTEM_ERROR(errno));
+  }
+
   return NULL != pBuf->pData;
 }
 

source/libs/executor/src/executor.c

@@ -18,20 +18,20 @@
 #include "executorimpl.h"
 #include "planner.h"
 
-static int32_t doSetStreamBlock(SOperatorInfo* pOperator, void* input, uint64_t reqId) {
+static int32_t doSetStreamBlock(SOperatorInfo* pOperator, void* input, char* id) {
   ASSERT(pOperator != NULL);
   if (pOperator->operatorType != OP_StreamScan) {
     if (pOperator->numOfDownstream == 0) {
-      qError("failed to find stream scan operator to set the input data block, reqId:0x%" PRIx64, reqId);
+      qError("failed to find stream scan operator to set the input data block, %s" PRIx64, id);
       return TSDB_CODE_QRY_APP_ERROR;
     }
 
     if (pOperator->numOfDownstream > 1) {  // not handle this in join query
-      qError("join not supported for stream block scan, reqId:0x%" PRIx64, reqId);
+      qError("join not supported for stream block scan, %s" PRIx64, id);
       return TSDB_CODE_QRY_APP_ERROR;
     }
 
-    return doSetStreamBlock(pOperator->pDownstream[0], input, reqId);
+    return doSetStreamBlock(pOperator->pDownstream[0], input, id);
   } else {
     SStreamBlockScanInfo* pInfo = pOperator->info;
     tqReadHandleSetMsg(pInfo->readerHandle, input, 0);
@@ -52,9 +52,9 @@ int32_t qSetStreamInput(qTaskInfo_t tinfo, void* input) {
 
   int32_t code = doSetStreamBlock(pTaskInfo->pRoot, input, GET_TASKID(pTaskInfo));
   if (code != TSDB_CODE_SUCCESS) {
-    qError("failed to set the stream block data, reqId:0x%"PRIx64, GET_TASKID(pTaskInfo));
+    qError("%s failed to set the stream block data", GET_TASKID(pTaskInfo));
   } else {
-    qDebug("set the stream block successfully, reqId:0x%"PRIx64, GET_TASKID(pTaskInfo));
+    qDebug("%s set the stream block successfully", GET_TASKID(pTaskInfo));
   }
 
   return code;
@@ -81,7 +81,7 @@ qTaskInfo_t qCreateStreamExecTaskInfo(void* msg, void* streamReadHandle) {
   }
 
   qTaskInfo_t pTaskInfo = NULL;
-  code = qCreateExecTask(streamReadHandle, 0, plan, &pTaskInfo, NULL);
+  code = qCreateExecTask(streamReadHandle, 0, 0, plan, &pTaskInfo, NULL);
   if (code != TSDB_CODE_SUCCESS) {
     // TODO: destroy SSubplan & pTaskInfo
     terrno = code;

source/libs/executor/src/executorMain.c

@@ -69,11 +69,11 @@ void freeParam(STaskParam *param) {
   tfree(param->prevResult);
 }
 
-int32_t qCreateExecTask(void* readHandle, int32_t vgId, SSubplan* pSubplan, qTaskInfo_t* pTaskInfo, DataSinkHandle* handle) {
+int32_t qCreateExecTask(void* readHandle, int32_t vgId, uint64_t taskId, SSubplan* pSubplan, qTaskInfo_t* pTaskInfo, DataSinkHandle* handle) {
   assert(readHandle != NULL && pSubplan != NULL);
   SExecTaskInfo** pTask = (SExecTaskInfo**)pTaskInfo;
 
-  int32_t code = createExecTaskInfoImpl(pSubplan, pTask, readHandle);
+  int32_t code = createExecTaskInfoImpl(pSubplan, pTask, readHandle, taskId);
   if (code != TSDB_CODE_SUCCESS) {
     goto _error;
   }
@@ -141,16 +141,10 @@ int32_t qExecTask(qTaskInfo_t tinfo, SSDataBlock** pRes, uint64_t *useconds) {
   SExecTaskInfo* pTaskInfo = (SExecTaskInfo*)tinfo;
   int64_t        threadId = taosGetSelfPthreadId();
 
-  // todo: remove it.
-  if (tinfo == NULL) {
-    return TSDB_CODE_SUCCESS;
-  }
-
   *pRes = NULL;
-
   int64_t curOwner = 0;
   if ((curOwner = atomic_val_compare_exchange_64(&pTaskInfo->owner, 0, threadId)) != 0) {
-    qError("QInfo:0x%" PRIx64 "-%p qhandle is now executed by thread:%p", GET_TASKID(pTaskInfo), pTaskInfo,
+    qError("%s-%p execTask is now executed by thread:%p", GET_TASKID(pTaskInfo), pTaskInfo,
            (void*)curOwner);
     pTaskInfo->code = TSDB_CODE_QRY_IN_EXEC;
     return pTaskInfo->code;
@@ -161,7 +155,7 @@ int32_t qExecTask(qTaskInfo_t tinfo, SSDataBlock** pRes, uint64_t *useconds) {
   }
 
   if (isTaskKilled(pTaskInfo)) {
-    qDebug("QInfo:0x%" PRIx64 " it is already killed, abort", GET_TASKID(pTaskInfo));
+    qDebug("%s already killed, abort", GET_TASKID(pTaskInfo));
     return TSDB_CODE_SUCCESS;
   }
 
@@ -170,12 +164,12 @@ int32_t qExecTask(qTaskInfo_t tinfo, SSDataBlock** pRes, uint64_t *useconds) {
   if (ret != TSDB_CODE_SUCCESS) {
     publishQueryAbortEvent(pTaskInfo, ret);
     pTaskInfo->code = ret;
-    qDebug("QInfo:0x%" PRIx64 " query abort due to error/cancel occurs, code:%s", GET_TASKID(pTaskInfo),
+    qDebug("%s task abort due to error/cancel occurs, code:%s", GET_TASKID(pTaskInfo),
            tstrerror(pTaskInfo->code));
     return pTaskInfo->code;
   }
 
-  qDebug("QInfo:0x%" PRIx64 " query task is launched", GET_TASKID(pTaskInfo));
+  qDebug("%s execTask is launched", GET_TASKID(pTaskInfo));
 
   bool newgroup = false;
   publishOperatorProfEvent(pTaskInfo->pRoot, QUERY_PROF_BEFORE_OPERATOR_EXEC);
@@ -184,66 +178,25 @@ int32_t qExecTask(qTaskInfo_t tinfo, SSDataBlock** pRes, uint64_t *useconds) {
   st = taosGetTimestampUs();
   *pRes = pTaskInfo->pRoot->exec(pTaskInfo->pRoot, &newgroup);
 
-  pTaskInfo->cost.elapsedTime += (taosGetTimestampUs() - st);
+  uint64_t el = (taosGetTimestampUs() - st);
+  pTaskInfo->cost.elapsedTime += el;
+
   publishOperatorProfEvent(pTaskInfo->pRoot, QUERY_PROF_AFTER_OPERATOR_EXEC);
 
   if (NULL == *pRes) {
     *useconds = pTaskInfo->cost.elapsedTime;
   }
 
-  qDebug("QInfo:0x%" PRIx64 " query paused, %d rows returned, total:%" PRId64 " rows, in sinkNode:%d",
-         GET_TASKID(pTaskInfo), 0, 0L, 0);
+  int32_t current = (*pRes != NULL)? (*pRes)->info.rows:0;
+  pTaskInfo->totalRows += current;
+
+  qDebug("%s task suspended, %d rows returned, total:%" PRId64 " rows, in sinkNode:%d, elapsed:%.2f ms",
+         GET_TASKID(pTaskInfo), current, pTaskInfo->totalRows, 0, el/1000.0);
 
   atomic_store_64(&pTaskInfo->owner, 0);
   return pTaskInfo->code;
 }
 
-int32_t qRetrieveQueryResultInfo(qTaskInfo_t qinfo, bool* buildRes, void* pRspContext) {
-  SQInfo *pQInfo = (SQInfo *)qinfo;
-
-  if (pQInfo == NULL || !isValidQInfo(pQInfo)) {
-    qError("QInfo invalid qhandle");
-    return TSDB_CODE_QRY_INVALID_QHANDLE;
-  }
-
-  *buildRes = false;
-  if (IS_QUERY_KILLED(pQInfo)) {
-    qDebug("QInfo:0x%"PRIx64" query is killed, code:0x%08x", pQInfo->qId, pQInfo->code);
-    return pQInfo->code;
-  }
-
-  int32_t code = TSDB_CODE_SUCCESS;
-
-  if (tsRetrieveBlockingModel) {
-    pQInfo->rspContext = pRspContext;
-    tsem_wait(&pQInfo->ready);
-    *buildRes = true;
-    code = pQInfo->code;
-  } else {
-    STaskRuntimeEnv* pRuntimeEnv = &pQInfo->runtimeEnv;
-    STaskAttr *pQueryAttr = pQInfo->runtimeEnv.pQueryAttr;
-
-    pthread_mutex_lock(&pQInfo->lock);
-
-    assert(pQInfo->rspContext == NULL);
-    if (pQInfo->dataReady == QUERY_RESULT_READY) {
-      *buildRes = true;
-      qDebug("QInfo:0x%"PRIx64" retrieve result info, rowsize:%d, rows:%d, code:%s", pQInfo->qId, pQueryAttr->resultRowSize,
-             GET_NUM_OF_RESULTS(pRuntimeEnv), tstrerror(pQInfo->code));
-    } else {
-      *buildRes = false;
-      qDebug("QInfo:0x%"PRIx64" retrieve req set query return result after paused", pQInfo->qId);
-      pQInfo->rspContext = pRspContext;
-      assert(pQInfo->rspContext != NULL);
-    }
-
-    code = pQInfo->code;
-    pthread_mutex_unlock(&pQInfo->lock);
-  }
-
-  return code;
-}
-
 void* qGetResultRetrieveMsg(qTaskInfo_t qinfo) {
   SQInfo* pQInfo = (SQInfo*) qinfo;
   assert(pQInfo != NULL);
@@ -252,18 +205,18 @@ void* qGetResultRetrieveMsg(qTaskInfo_t qinfo) {
 }
 
 int32_t qKillTask(qTaskInfo_t qinfo) {
-  SQInfo *pQInfo = (SQInfo *)qinfo;
+  SExecTaskInfo *pTaskInfo = (SExecTaskInfo *)qinfo;
 
-  if (pQInfo == NULL || !isValidQInfo(pQInfo)) {
+  if (pTaskInfo == NULL) {
     return TSDB_CODE_QRY_INVALID_QHANDLE;
   }
 
-  qDebug("QInfo:0x%"PRIx64" query killed", pQInfo->qId);
-  setQueryKilled(pQInfo);
+  qDebug("%s execTask killed", GET_TASKID(pTaskInfo));
+  setTaskKilled(pTaskInfo);
 
   // Wait for the query executing thread being stopped/
   // Once the query is stopped, the owner of qHandle will be cleared immediately.
-  while (pQInfo->owner != 0) {
+  while (pTaskInfo->owner != 0) {
     taosMsleep(100);
   }
 
@@ -271,14 +224,14 @@ int32_t qKillTask(qTaskInfo_t qinfo) {
 }
 
 int32_t qAsyncKillTask(qTaskInfo_t qinfo) {
-  SQInfo *pQInfo = (SQInfo *)qinfo;
+  SExecTaskInfo *pTaskInfo = (SExecTaskInfo *)qinfo;
 
-  if (pQInfo == NULL || !isValidQInfo(pQInfo)) {
+  if (pTaskInfo == NULL) {
     return TSDB_CODE_QRY_INVALID_QHANDLE;
   }
 
-  qDebug("QInfo:0x%"PRIx64" query async killed", pQInfo->qId);
-  setQueryKilled(pQInfo);
+  qDebug("%s execTask async killed", GET_TASKID(pTaskInfo));
+  setTaskKilled(pTaskInfo);
 
   return TSDB_CODE_SUCCESS;
 }
@@ -293,145 +246,12 @@ int32_t qIsTaskCompleted(qTaskInfo_t qinfo) {
   return isTaskKilled(pTaskInfo) || Q_STATUS_EQUAL(pTaskInfo->status, TASK_OVER);
 }
 
-void qDestroyTask(qTaskInfo_t qHandle) {
-  SQInfo* pQInfo = (SQInfo*) qHandle;
-  if (!isValidQInfo(pQInfo)) {
-    return;
-  }
-
-  qDebug("QInfo:0x%"PRIx64" query completed", pQInfo->qId);
-  queryCostStatis(pQInfo);   // print the query cost summary
-  doDestroyTask(pQInfo);
-}
-
-void* qOpenTaskMgmt(int32_t vgId) {
-  const int32_t refreshHandleInterval = 30; // every 30 seconds, refresh handle pool
-
-  char cacheName[128] = {0};
-  sprintf(cacheName, "qhandle_%d", vgId);
-
-  STaskMgmt* pTaskMgmt = calloc(1, sizeof(STaskMgmt));
-  if (pTaskMgmt == NULL) {
-    terrno = TSDB_CODE_QRY_OUT_OF_MEMORY;
-    return NULL;
-  }
-
-  pTaskMgmt->qinfoPool = taosCacheInit(TSDB_CACHE_PTR_KEY, refreshHandleInterval, true, freeqinfoFn, cacheName);
-  pTaskMgmt->closed    = false;
-  pTaskMgmt->vgId      = vgId;
-
-  pthread_mutex_init(&pTaskMgmt->lock, NULL);
-
-  qDebug("vgId:%d, open queryTaskMgmt success", vgId);
-  return pTaskMgmt;
-}
-
-void qTaskMgmtNotifyClosing(void* pQMgmt) {
-  if (pQMgmt == NULL) {
-    return;
-  }
-
-  STaskMgmt* pQueryMgmt = pQMgmt;
-  qInfo("vgId:%d, set querymgmt closed, wait for all queries cancelled", pQueryMgmt->vgId);
-
-  pthread_mutex_lock(&pQueryMgmt->lock);
-  pQueryMgmt->closed = true;
-  pthread_mutex_unlock(&pQueryMgmt->lock);
-
-  taosCacheRefresh(pQueryMgmt->qinfoPool, taskMgmtKillTaskFn, NULL);
-}
-
-void qQueryMgmtReOpen(void *pQMgmt) {
-  if (pQMgmt == NULL) {
-    return;
-  }
-
-  STaskMgmt *pQueryMgmt = pQMgmt;
-  qInfo("vgId:%d, set querymgmt reopen", pQueryMgmt->vgId);
-
-  pthread_mutex_lock(&pQueryMgmt->lock);
-  pQueryMgmt->closed = false;
-  pthread_mutex_unlock(&pQueryMgmt->lock);
-}
-
-void qCleanupTaskMgmt(void* pQMgmt) {
-  if (pQMgmt == NULL) {
-    return;
-  }
-
-  STaskMgmt* pQueryMgmt = pQMgmt;
-  int32_t vgId = pQueryMgmt->vgId;
-
-  assert(pQueryMgmt->closed);
-
-  SCacheObj* pqinfoPool = pQueryMgmt->qinfoPool;
-  pQueryMgmt->qinfoPool = NULL;
-
-  taosCacheCleanup(pqinfoPool);
-  pthread_mutex_destroy(&pQueryMgmt->lock);
-  tfree(pQueryMgmt);
-
-  qDebug("vgId:%d, queryMgmt cleanup completed", vgId);
-}
-
-void** qRegisterTask(void* pMgmt, uint64_t qId, void *qInfo) {
-  if (pMgmt == NULL) {
-    terrno = TSDB_CODE_VND_INVALID_VGROUP_ID;
-    return NULL;
-  }
-
-  STaskMgmt *pQueryMgmt = pMgmt;
-  if (pQueryMgmt->qinfoPool == NULL) {
-    qError("QInfo:0x%"PRIx64"-%p failed to add qhandle into qMgmt, since qMgmt is closed", qId, (void*)qInfo);
-    terrno = TSDB_CODE_VND_INVALID_VGROUP_ID;
-    return NULL;
-  }
-
-  pthread_mutex_lock(&pQueryMgmt->lock);
-  if (pQueryMgmt->closed) {
-    pthread_mutex_unlock(&pQueryMgmt->lock);
-    qError("QInfo:0x%"PRIx64"-%p failed to add qhandle into cache, since qMgmt is colsing", qId, (void*)qInfo);
-    terrno = TSDB_CODE_VND_INVALID_VGROUP_ID;
-    return NULL;
-  } else {
-    void** handle = taosCachePut(pQueryMgmt->qinfoPool, &qId, sizeof(qId), &qInfo, sizeof(TSDB_CACHE_PTR_TYPE),
-                                 (getMaximumIdleDurationSec()*1000));
-    pthread_mutex_unlock(&pQueryMgmt->lock);
-
-    return handle;
-  }
-}
-
-void** qAcquireTask(void* pMgmt, uint64_t _key) {
-  STaskMgmt *pQueryMgmt = pMgmt;
-
-  if (pQueryMgmt->closed) {
-    terrno = TSDB_CODE_VND_INVALID_VGROUP_ID;
-    return NULL;
-  }
-
-  if (pQueryMgmt->qinfoPool == NULL) {
-    terrno = TSDB_CODE_QRY_INVALID_QHANDLE;
-    return NULL;
-  }
-
-  void** handle = taosCacheAcquireByKey(pQueryMgmt->qinfoPool, &_key, sizeof(_key));
-  if (handle == NULL || *handle == NULL) {
-    terrno = TSDB_CODE_QRY_INVALID_QHANDLE;
-    return NULL;
-  } else {
-    return handle;
-  }
-}
-
-void** qReleaseTask(void* pMgmt, void* pQInfo, bool freeHandle) {
-  STaskMgmt *pQueryMgmt = pMgmt;
-  if (pQueryMgmt->qinfoPool == NULL) {
-    return NULL;
-  }
+void qDestroyTask(qTaskInfo_t qTaskHandle) {
+  SExecTaskInfo* pTaskInfo = (SExecTaskInfo*) qTaskHandle;
+  qDebug("%s execTask completed, numOfRows:%"PRId64, GET_TASKID(pTaskInfo), pTaskInfo->totalRows);
 
-  taosCacheRelease(pQueryMgmt->qinfoPool, pQInfo, freeHandle);
-  return 0;
+  queryCostStatis(pTaskInfo);   // print the query cost summary
+  doDestroyTask(pTaskInfo);
 }
 
 #if 0
@@ -445,8 +265,8 @@ int32_t qKillQueryByQId(void* pMgmt, int64_t qId, int32_t waitMs, int32_t waitCo
   if (pQInfo == NULL || !isValidQInfo(pQInfo)) {
     return TSDB_CODE_QRY_INVALID_QHANDLE;
   }
-  qWarn("QId:0x%"PRIx64" be killed(no memory commit).", pQInfo->qId);
-  setQueryKilled(pQInfo);
+  qWarn("%s be killed(no memory commit).", pQInfo->qId);
+  setTaskKilled(pQInfo);
 
   // wait query stop
   int32_t loop = 0;

source/libs/executor/src/executorimpl.c

@@ -2432,10 +2432,6 @@ static bool needBuildResAfterQueryComplete(SQInfo* pQInfo) {
 }
 
 bool isTaskKilled(SExecTaskInfo *pTaskInfo) {
-  if (IS_QUERY_KILLED(pTaskInfo)) {
-    return true;
-  }
-
   // query has been executed more than tsShellActivityTimer, and the retrieve has not arrived
   // abort current query execution.
   if (pTaskInfo->owner != 0 && ((taosGetTimestampSec() - pTaskInfo->cost.start/1000) > 10*getMaximumIdleDurationSec())
@@ -2444,13 +2440,13 @@ bool isTaskKilled(SExecTaskInfo *pTaskInfo) {
     assert(pTaskInfo->cost.start != 0);
 //    qDebug("QInfo:%" PRIu64 " retrieve not arrive beyond %d ms, abort current query execution, start:%" PRId64
 //           ", current:%d", pQInfo->qId, 1, pQInfo->startExecTs, taosGetTimestampSec());
-    return true;
+//    return true;
   }
 
   return false;
 }
 
-void setQueryKilled(SQInfo *pQInfo) { pQInfo->code = TSDB_CODE_TSC_QUERY_CANCELLED;}
+void setTaskKilled(SExecTaskInfo *pTaskInfo) { pTaskInfo->code = TSDB_CODE_TSC_QUERY_CANCELLED;}
 
 //static bool isFixedOutputQuery(STaskAttr* pQueryAttr) {
 //  if (QUERY_IS_INTERVAL_QUERY(pQueryAttr)) {
@@ -4420,33 +4416,32 @@ void calculateOperatorProfResults(SQInfo* pQInfo) {
   taosArrayDestroy(opStack);
 }
 
-void queryCostStatis(SQInfo *pQInfo) {
-  STaskRuntimeEnv *pRuntimeEnv = &pQInfo->runtimeEnv;
-  STaskCostInfo *pSummary = &pQInfo->summary;
+void queryCostStatis(SExecTaskInfo *pTaskInfo) {
+  STaskCostInfo *pSummary = &pTaskInfo->cost;
 
-  uint64_t hashSize = taosHashGetMemSize(pQInfo->runtimeEnv.pResultRowHashTable);
-  hashSize += taosHashGetMemSize(pRuntimeEnv->tableqinfoGroupInfo.map);
-  pSummary->hashSize = hashSize;
+//  uint64_t hashSize = taosHashGetMemSize(pQInfo->runtimeEnv.pResultRowHashTable);
+//  hashSize += taosHashGetMemSize(pRuntimeEnv->tableqinfoGroupInfo.map);
+//  pSummary->hashSize = hashSize;
 
   // add the merge time
   pSummary->elapsedTime += pSummary->firstStageMergeTime;
 
-  SResultRowPool* p = pQInfo->runtimeEnv.pool;
-  if (p != NULL) {
-    pSummary->winInfoSize = getResultRowPoolMemSize(p);
-    pSummary->numOfTimeWindows = getNumOfAllocatedResultRows(p);
-  } else {
-    pSummary->winInfoSize = 0;
-    pSummary->numOfTimeWindows = 0;
-  }
-
-  calculateOperatorProfResults(pQInfo);
-
-  //qDebug("QInfo:0x%"PRIx64" :cost summary: elapsed time:%"PRId64" us, first merge:%"PRId64" us, total blocks:%d, "
-//         "load block statis:%d, load data block:%d, total rows:%"PRId64 ", check rows:%"PRId64,
-//         pQInfo->qId, pSummary->elapsedTime, pSummary->firstStageMergeTime, pSummary->totalBlocks, pSummary->loadBlockStatis,
-//         pSummary->loadBlocks, pSummary->totalRows, pSummary->totalCheckedRows);
+//  SResultRowPool* p = pTaskInfo->pool;
+//  if (p != NULL) {
+//    pSummary->winInfoSize = getResultRowPoolMemSize(p);
+//    pSummary->numOfTimeWindows = getNumOfAllocatedResultRows(p);
+//  } else {
+//    pSummary->winInfoSize = 0;
+//    pSummary->numOfTimeWindows = 0;
+//  }
+//
+//  calculateOperatorProfResults(pQInfo);
 
+  qDebug("%s :cost summary: elapsed time:%"PRId64" us, first merge:%"PRId64" us, total blocks:%d, "
+         "load block statis:%d, load data block:%d, total rows:%"PRId64 ", check rows:%"PRId64,
+         GET_TASKID(pTaskInfo), pSummary->elapsedTime, pSummary->firstStageMergeTime, pSummary->totalBlocks, pSummary->loadBlockStatis,
+         pSummary->loadBlocks, pSummary->totalRows, pSummary->totalCheckedRows);
+//
   //qDebug("QInfo:0x%"PRIx64" :cost summary: winResPool size:%.2f Kb, numOfWin:%"PRId64", tableInfoSize:%.2f Kb, hashTable:%.2f Kb", pQInfo->qId, pSummary->winInfoSize/1024.0,
 //      pSummary->numOfTimeWindows, pSummary->tableInfoSize/1024.0, pSummary->hashSize/1024.0);
 
@@ -4995,7 +4990,7 @@ static SSDataBlock* doTableScan(void* param, bool *newgroup) {
       pResultRowInfo->curPos = 0;
     }
 
-    qDebug("QInfo:0x%"PRIx64" start to repeat scan data blocks due to query func required, qrange:%" PRId64 "-%" PRId64,
+    qDebug("%s start to repeat scan data blocks due to query func required, qrange:%" PRId64 "-%" PRId64,
            GET_TASKID(pTaskInfo), pTaskInfo->window.skey, pTaskInfo->window.ekey);
   }
 
@@ -5006,7 +5001,7 @@ static SSDataBlock* doTableScan(void* param, bool *newgroup) {
 //    STsdbQueryCond cond = createTsdbQueryCond(pQueryAttr, &pQueryAttr->window);
 //    tsdbResetQueryHandle(pTableScanInfo->pTsdbReadHandle, &cond);
 
-    qDebug("QInfo:0x%"PRIx64" start to reverse scan data blocks due to query func required, qrange:%" PRId64 "-%" PRId64,
+    qDebug("%s start to reverse scan data blocks due to query func required, qrange:%" PRId64 "-%" PRId64,
            GET_TASKID(pTaskInfo), pTaskInfo->window.skey, pTaskInfo->window.ekey);
 
     if (pResultRowInfo->size > 0) {
@@ -5120,7 +5115,7 @@ static void destroySendMsgInfo(SMsgSendInfo* pMsgBody) {
   tfree(pMsgBody);
 }
 
-void processRspMsg(void* parent, SRpcMsg* pMsg, SEpSet* pEpSet) {
+void qProcessFetchRsp(void* parent, SRpcMsg* pMsg, SEpSet* pEpSet) {
   SMsgSendInfo *pSendInfo = (SMsgSendInfo *) pMsg->ahandle;
   assert(pMsg->ahandle != NULL);
 
@@ -5151,6 +5146,8 @@ static SSDataBlock* doLoadRemoteData(void* param, bool* newgroup) {
 
   size_t totalSources = taosArrayGetSize(pExchangeInfo->pSources);
   if (pExchangeInfo->current >= totalSources) {
+    qDebug("%s all %"PRIzu" source(s) are exhausted, total rows:%"PRIu64" bytes:%"PRIu64", elapsed:%.2f ms", GET_TASKID(pTaskInfo), totalSources,
+           pExchangeInfo->totalRows, pExchangeInfo->totalSize, pExchangeInfo->totalElapsed/1000.0);
     return NULL;
   }
 
@@ -5171,7 +5168,8 @@ static SSDataBlock* doLoadRemoteData(void* param, bool* newgroup) {
     epSet.port[0] = pSource->addr.epAddr[0].port;
     tstrncpy(epSet.fqdn[0], pSource->addr.epAddr[0].fqdn, tListLen(epSet.fqdn[0]));
 
-    qDebug("QID:0x%" PRIx64 " build fetch msg and send to vgId:%d, ep:%s, taskId:0x%" PRIx64 ", %d/%" PRIzu,
+    int64_t startTs = taosGetTimestampUs();
+    qDebug("%s build fetch msg and send to vgId:%d, ep:%s, taskId:0x%" PRIx64 ", %d/%" PRIzu,
            GET_TASKID(pTaskInfo), pSource->addr.nodeId, epSet.fqdn[0], pSource->taskId, pExchangeInfo->current, totalSources);
 
     pMsg->header.vgId = htonl(pSource->addr.nodeId);
@@ -5182,7 +5180,7 @@ static SSDataBlock* doLoadRemoteData(void* param, bool* newgroup) {
     // send the fetch remote task result reques
     pMsgSendInfo = calloc(1, sizeof(SMsgSendInfo));
     if (NULL == pMsgSendInfo) {
-      qError("QID:0x%" PRIx64 " prepare message %d failed", GET_TASKID(pTaskInfo), (int32_t)sizeof(SMsgSendInfo));
+      qError("%s prepare message %d failed", GET_TASKID(pTaskInfo), (int32_t)sizeof(SMsgSendInfo));
       pTaskInfo->code = TSDB_CODE_QRY_OUT_OF_MEMORY;
       goto _error;
     }
@@ -5199,7 +5197,7 @@ static SSDataBlock* doLoadRemoteData(void* param, bool* newgroup) {
 
     SRetrieveTableRsp* pRsp = pExchangeInfo->pRsp;
     if (pRsp->numOfRows == 0) {
-      qDebug("QID:0x%"PRIx64" vgId:%d, taskID:0x%"PRIx64" %d of total completed, rowsOfSource:%"PRIu64", totalRows:%"PRIu64" try next",
+      qDebug("%s vgId:%d, taskID:0x%"PRIx64" %d of total completed, rowsOfSource:%"PRIu64", totalRows:%"PRIu64" try next",
           GET_TASKID(pTaskInfo), pSource->addr.nodeId, pSource->taskId, pExchangeInfo->current + 1,
              pExchangeInfo->rowsOfCurrentSource, pExchangeInfo->totalRows);
 
@@ -5207,6 +5205,11 @@ static SSDataBlock* doLoadRemoteData(void* param, bool* newgroup) {
       pExchangeInfo->current += 1;
 
       if (pExchangeInfo->current >= totalSources) {
+        int64_t el = taosGetTimestampUs() - startTs;
+        pExchangeInfo->totalElapsed += el;
+
+        qDebug("%s all %"PRIzu" sources are exhausted, total rows: %"PRIu64" bytes:%"PRIu64", elapsed:%.2f ms", GET_TASKID(pTaskInfo), totalSources,
+               pExchangeInfo->totalRows, pExchangeInfo->totalSize, pExchangeInfo->totalElapsed/1000.0);
         return NULL;
       } else {
         continue;
@@ -5233,21 +5236,24 @@ static SSDataBlock* doLoadRemoteData(void* param, bool* newgroup) {
     pRes->info.numOfCols = pOperator->numOfOutput;
     pRes->info.rows = pRsp->numOfRows;
 
+    int64_t el = taosGetTimestampUs() - startTs;
+
     pExchangeInfo->totalRows += pRsp->numOfRows;
-    pExchangeInfo->bytes += pRsp->compLen;
+    pExchangeInfo->totalSize += pRsp->compLen;
     pExchangeInfo->rowsOfCurrentSource += pRsp->numOfRows;
+    pExchangeInfo->totalElapsed += el;
 
     if (pRsp->completed == 1) {
-      qDebug("QID:0x%" PRIx64 " fetch msg rsp from vgId:%d, taskId:0x%" PRIx64 " numOfRows:%d, rowsOfSource:%" PRIu64
+      qDebug("%s fetch msg rsp from vgId:%d, taskId:0x%" PRIx64 " numOfRows:%d, rowsOfSource:%" PRIu64
              ", totalRows:%" PRIu64 ", totalBytes:%" PRIu64 " try next %d/%" PRIzu,
-             GET_TASKID(pTaskInfo), pSource->addr.nodeId, pSource->taskId, pRes->info.rows, pExchangeInfo->rowsOfCurrentSource, pExchangeInfo->totalRows, pExchangeInfo->bytes,
+             GET_TASKID(pTaskInfo), pSource->addr.nodeId, pSource->taskId, pRes->info.rows, pExchangeInfo->rowsOfCurrentSource, pExchangeInfo->totalRows, pExchangeInfo->totalSize,
              pExchangeInfo->current + 1, totalSources);
 
       pExchangeInfo->rowsOfCurrentSource = 0;
       pExchangeInfo->current += 1;
     } else {
-      qDebug("QID:0x%" PRIx64 " fetch msg rsp from vgId:%d, taskId:0x%" PRIx64 " numOfRows:%d, totalRows:%" PRIu64 ", totalBytes:%" PRIu64,
-             GET_TASKID(pTaskInfo), pSource->addr.nodeId, pSource->taskId, pRes->info.rows, pExchangeInfo->totalRows, pExchangeInfo->bytes);
+      qDebug("%s fetch msg rsp from vgId:%d, taskId:0x%" PRIx64 " numOfRows:%d, totalRows:%" PRIu64 ", totalBytes:%" PRIu64,
+             GET_TASKID(pTaskInfo), pSource->addr.nodeId, pSource->taskId, pRes->info.rows, pExchangeInfo->totalRows, pExchangeInfo->totalSize);
     }
 
     return pExchangeInfo->pResult;
@@ -5290,13 +5296,14 @@ SOperatorInfo* createExchangeOperatorInfo(const SArray* pSources, const SArray*
   pOperator->exec         = doLoadRemoteData;
   pOperator->pTaskInfo    = pTaskInfo;
 
+#if 1
   { // todo refactor
     SRpcInit rpcInit;
     memset(&rpcInit, 0, sizeof(rpcInit));
     rpcInit.localPort = 0;
-    rpcInit.label = "TSC";
+    rpcInit.label = "EX";
     rpcInit.numOfThreads = 1;
-    rpcInit.cfp = processRspMsg;
+    rpcInit.cfp = qProcessFetchRsp;
     rpcInit.sessions = tsMaxConnections;
     rpcInit.connType = TAOS_CONN_CLIENT;
     rpcInit.user = (char *)"root";
@@ -5310,7 +5317,7 @@ SOperatorInfo* createExchangeOperatorInfo(const SArray* pSources, const SArray*
       return NULL; // todo
     }
   }
-
+#endif
   return pOperator;
 }
 
@@ -7729,33 +7736,37 @@ static int32_t deserializeColFilterInfo(SColumnFilterInfo* pColFilters, int16_t
   return TSDB_CODE_SUCCESS;
 }
 
-static SExecTaskInfo* createExecTaskInfo(uint64_t queryId) {
+static SExecTaskInfo* createExecTaskInfo(uint64_t queryId, uint64_t taskId) {
   SExecTaskInfo* pTaskInfo = calloc(1, sizeof(SExecTaskInfo));
   setTaskStatus(pTaskInfo, TASK_NOT_COMPLETED);
 
-  pthread_mutex_init(&pTaskInfo->lock, NULL);
   pTaskInfo->cost.created = taosGetTimestampMs();
   pTaskInfo->id.queryId = queryId;
+
+  char* p = calloc(1, 128);
+  snprintf(p, 128, "TID:0x%"PRIx64" QID:0x%"PRIx64, taskId, queryId);
+  pTaskInfo->id.str = strdup(p);
+
   return pTaskInfo;
 }
 
-static tsdbReadHandleT doCreateDataReadHandle(STableScanPhyNode* pTableScanNode, void* readerHandle, uint64_t queryId);
+static tsdbReadHandleT doCreateDataReadHandle(STableScanPhyNode* pTableScanNode, void* readerHandle, uint64_t queryId, uint64_t taskId);
 
-SOperatorInfo* doCreateOperatorTreeNode(SPhyNode* pPhyNode, SExecTaskInfo* pTaskInfo, void* readerHandle, uint64_t queryId) {
+SOperatorInfo* doCreateOperatorTreeNode(SPhyNode* pPhyNode, SExecTaskInfo* pTaskInfo, void* readerHandle, uint64_t queryId, uint64_t taskId) {
   if (pPhyNode->pChildren == NULL || taosArrayGetSize(pPhyNode->pChildren) == 0) {
     if (pPhyNode->info.type == OP_TableScan) {
 
       SScanPhyNode*  pScanPhyNode = (SScanPhyNode*)pPhyNode;
       size_t         numOfCols = taosArrayGetSize(pPhyNode->pTargets);
 
-      tsdbReadHandleT tReaderHandle = doCreateDataReadHandle((STableScanPhyNode*) pPhyNode, readerHandle, (uint64_t) queryId);
+      tsdbReadHandleT tReaderHandle = doCreateDataReadHandle((STableScanPhyNode*) pPhyNode, readerHandle, (uint64_t) queryId, taskId);
 
       return createTableScanOperatorInfo(tReaderHandle, pScanPhyNode->order, numOfCols, pScanPhyNode->count, pTaskInfo);
     } else if (pPhyNode->info.type == OP_DataBlocksOptScan) {
       SScanPhyNode*  pScanPhyNode = (SScanPhyNode*)pPhyNode;
       size_t         numOfCols = taosArrayGetSize(pPhyNode->pTargets);
 
-      tsdbReadHandleT tReaderHandle = doCreateDataReadHandle((STableScanPhyNode*) pPhyNode, readerHandle, (uint64_t) queryId);
+      tsdbReadHandleT tReaderHandle = doCreateDataReadHandle((STableScanPhyNode*) pPhyNode, readerHandle, (uint64_t) queryId, taskId);
 
       return createDataBlocksOptScanInfo(tReaderHandle, pScanPhyNode->order, numOfCols, pScanPhyNode->count, pScanPhyNode->reverse, pTaskInfo);
     } else if (pPhyNode->info.type == OP_Exchange) {
@@ -7773,13 +7784,13 @@ SOperatorInfo* doCreateOperatorTreeNode(SPhyNode* pPhyNode, SExecTaskInfo* pTask
 
     for (int32_t i = 0; i < size; ++i) {
       SPhyNode*      pChildNode = taosArrayGetP(pPhyNode->pChildren, i);
-      SOperatorInfo* op = doCreateOperatorTreeNode(pChildNode, pTaskInfo, readerHandle, queryId);
+      SOperatorInfo* op = doCreateOperatorTreeNode(pChildNode, pTaskInfo, readerHandle, queryId, taskId);
       return createAggregateOperatorInfo(op, pPhyNode->pTargets, pTaskInfo);
     }
   }
 }
 
-static tsdbReadHandleT createDataReadHandle(STableScanPhyNode* pTableScanNode, STableGroupInfo* pGroupInfo, void* readerHandle, uint64_t queryId) {
+static tsdbReadHandleT createDataReadHandle(STableScanPhyNode* pTableScanNode, STableGroupInfo* pGroupInfo, void* readerHandle, uint64_t queryId, uint64_t taskId) {
   STsdbQueryCond cond = {.loadExternalRows = false};
 
   cond.order = pTableScanNode->scan.order;
@@ -7798,10 +7809,10 @@ static tsdbReadHandleT createDataReadHandle(STableScanPhyNode* pTableScanNode, S
     cond.colList[i].colId = pSchema->colId;
   }
 
-  return tsdbQueryTables(readerHandle, &cond, pGroupInfo, queryId, NULL);
+  return tsdbQueryTables(readerHandle, &cond, pGroupInfo, queryId, taskId);
 }
 
-static tsdbReadHandleT doCreateDataReadHandle(STableScanPhyNode* pTableScanNode, void* readerHandle, uint64_t queryId) {
+static tsdbReadHandleT doCreateDataReadHandle(STableScanPhyNode* pTableScanNode, void* readerHandle, uint64_t queryId, uint64_t taskId) {
   int32_t         code = 0;
   STableGroupInfo groupInfo = {0};
 
@@ -7811,7 +7822,7 @@ static tsdbReadHandleT doCreateDataReadHandle(STableScanPhyNode* pTableScanNode,
 
   if (tableType == TSDB_SUPER_TABLE) {
     code =
-        tsdbQuerySTableByTagCond(readerHandle, uid, window.skey, NULL, 0, 0, NULL, &groupInfo, NULL, 0, queryId);
+        tsdbQuerySTableByTagCond(readerHandle, uid, window.skey, NULL, 0, 0, NULL, &groupInfo, NULL, 0, queryId, taskId);
     if (code != TSDB_CODE_SUCCESS) {
       goto _error;
     }
@@ -7821,35 +7832,35 @@ static tsdbReadHandleT doCreateDataReadHandle(STableScanPhyNode* pTableScanNode,
 
     SArray* pa = taosArrayInit(1, sizeof(STableKeyInfo));
 
-    STableKeyInfo info = {.pTable = NULL, .lastKey = 0, .uid = uid};
+    STableKeyInfo info = {.lastKey = 0, .uid = uid};
     taosArrayPush(pa, &info);
     taosArrayPush(groupInfo.pGroupList, &pa);
   }
 
   if (groupInfo.numOfTables == 0) {
     code = 0;
-    qDebug("no table qualified for query, reqId:0x%"PRIx64, queryId);
+    qDebug("no table qualified for query, TID:0x%"PRIx64", QID:0x%"PRIx64, taskId, queryId);
     goto _error;
   }
 
-  return createDataReadHandle(pTableScanNode, &groupInfo, readerHandle, queryId);
+  return createDataReadHandle(pTableScanNode, &groupInfo, readerHandle, queryId, taskId);
 
   _error:
   terrno = code;
   return NULL;
 }
 
-int32_t createExecTaskInfoImpl(SSubplan* pPlan, SExecTaskInfo** pTaskInfo, void* readerHandle) {
+int32_t createExecTaskInfoImpl(SSubplan* pPlan, SExecTaskInfo** pTaskInfo, void* readerHandle, uint64_t taskId) {
   uint64_t queryId = pPlan->id.queryId;
 
   int32_t code = TSDB_CODE_SUCCESS;
-  *pTaskInfo = createExecTaskInfo(queryId);
+  *pTaskInfo = createExecTaskInfo(queryId, taskId);
   if (*pTaskInfo == NULL) {
     code = TSDB_CODE_QRY_OUT_OF_MEMORY;
     goto _complete;
   }
 
-  (*pTaskInfo)->pRoot = doCreateOperatorTreeNode(pPlan->pNode, *pTaskInfo, readerHandle, queryId);
+  (*pTaskInfo)->pRoot = doCreateOperatorTreeNode(pPlan->pNode, *pTaskInfo, readerHandle, queryId, taskId);
   if ((*pTaskInfo)->pRoot == NULL) {
     code = TSDB_CODE_QRY_OUT_OF_MEMORY;
     goto _complete;
@@ -8673,229 +8684,6 @@ FORCE_INLINE bool checkQIdEqual(void *qHandle, uint64_t qId) {
   return ((SQInfo *)qHandle)->qId == qId;
 }
 
-SQInfo* createQInfoImpl(SQueryTableReq* pQueryMsg, SGroupbyExpr* pGroupbyExpr, SExprInfo* pExprs,
-                        SExprInfo* pSecExprs, STableGroupInfo* pTableGroupInfo, SColumnInfo* pTagCols, SFilterInfo* pFilters, int32_t vgId,
-                        char* sql, uint64_t qId, struct SUdfInfo* pUdfInfo) {
-  int16_t numOfCols = pQueryMsg->numOfCols;
-  int16_t numOfOutput = pQueryMsg->numOfOutput;
-
-  SQInfo *pQInfo = (SQInfo *)calloc(1, sizeof(SQInfo));
-  if (pQInfo == NULL) {
-    goto _cleanup_qinfo;
-  }
-
-  pQInfo->qId = qId;
-  pQInfo->startExecTs = 0;
-
-  pQInfo->runtimeEnv.pUdfInfo = pUdfInfo;
-
-  // to make sure third party won't overwrite this structure
-  pQInfo->signature = pQInfo;
-  STaskAttr* pQueryAttr = &pQInfo->query;
-  pQInfo->runtimeEnv.pQueryAttr = pQueryAttr;
-
-  pQueryAttr->tableGroupInfo  = *pTableGroupInfo;
-  pQueryAttr->numOfCols       = numOfCols;
-  pQueryAttr->numOfOutput     = numOfOutput;
-  pQueryAttr->limit.limit     = pQueryMsg->limit;
-  pQueryAttr->limit.offset    = pQueryMsg->offset;
-  pQueryAttr->order.order     = pQueryMsg->order;
-  pQueryAttr->order.col.info.colId = pQueryMsg->orderColId;
-  pQueryAttr->pExpr1          = pExprs;
-  pQueryAttr->pExpr2          = pSecExprs;
-  pQueryAttr->numOfExpr2      = pQueryMsg->secondStageOutput;
-  pQueryAttr->pGroupbyExpr    = pGroupbyExpr;
-  memcpy(&pQueryAttr->interval, &pQueryMsg->interval, sizeof(pQueryAttr->interval));
-  pQueryAttr->fillType        = pQueryMsg->fillType;
-  pQueryAttr->numOfTags       = pQueryMsg->numOfTags;
-  pQueryAttr->tagColList      = pTagCols;
-  pQueryAttr->prjInfo.vgroupLimit = pQueryMsg->vgroupLimit;
-  pQueryAttr->prjInfo.ts      = (pQueryMsg->order == TSDB_ORDER_ASC)? INT64_MIN:INT64_MAX;
-//  pQueryAttr->sw              = pQueryMsg->sw;
-  pQueryAttr->vgId            = vgId;
-
-  pQueryAttr->stableQuery     = pQueryMsg->stableQuery;
-  pQueryAttr->topBotQuery     = pQueryMsg->topBotQuery;
-  pQueryAttr->groupbyColumn   = pQueryMsg->groupbyColumn;
-  pQueryAttr->hasTagResults   = pQueryMsg->hasTagResults;
-  pQueryAttr->timeWindowInterpo = pQueryMsg->timeWindowInterpo;
-  pQueryAttr->queryBlockDist  = pQueryMsg->queryBlockDist;
-  pQueryAttr->stabledev       = pQueryMsg->stabledev;
-  pQueryAttr->tsCompQuery     = pQueryMsg->tsCompQuery;
-  pQueryAttr->simpleAgg       = pQueryMsg->simpleAgg;
-  pQueryAttr->pointInterpQuery = pQueryMsg->pointInterpQuery;
-  pQueryAttr->needReverseScan  = pQueryMsg->needReverseScan;
-  pQueryAttr->stateWindow      = pQueryMsg->stateWindow;
-  pQueryAttr->vgId            = vgId;
-//  pQueryAttr->pFilters        = pFilters;
-  
-  pQueryAttr->tableCols = calloc(numOfCols, sizeof(SSingleColumnFilterInfo));
-  if (pQueryAttr->tableCols == NULL) {
-    goto _cleanup;
-  }
-
-  pQueryAttr->srcRowSize = 0;
-  pQueryAttr->maxTableColumnWidth = 0;
-  for (int16_t i = 0; i < numOfCols; ++i) {
-    pQueryAttr->tableCols[i] = pQueryMsg->tableCols[i];
-//    pQueryAttr->tableCols[i].flist.filterInfo = tFilterInfoDup(pQueryMsg->tableCols[i].flist.filterInfo, pQueryAttr->tableCols[i].flist.numOfFilters);
-
-    pQueryAttr->srcRowSize += pQueryAttr->tableCols[i].bytes;
-    if (pQueryAttr->maxTableColumnWidth < pQueryAttr->tableCols[i].bytes) {
-      pQueryAttr->maxTableColumnWidth = pQueryAttr->tableCols[i].bytes;
-    }
-  }
-
-  for (int16_t col = 0; col < numOfOutput; ++col) {
-    assert(pExprs[col].base.resSchema.bytes > 0);
-    pQueryAttr->resultRowSize += pExprs[col].base.resSchema.bytes;
-
-    // keep the tag length
-    if (TSDB_COL_IS_TAG(pExprs[col].base.pColumns->flag)) {
-      pQueryAttr->tagLen += pExprs[col].base.resSchema.bytes;
-    }
-
-//    if (pExprs[col].base.flist.filterInfo) {
-//      ++pQueryAttr->havingNum;
-//    }
-  }
-
-  doUpdateExprColumnIndex(pQueryAttr);
-
-  if (pSecExprs != NULL) {
-    int32_t resultRowSize = 0;
-
-    // calculate the result row size
-    for (int16_t col = 0; col < pQueryAttr->numOfExpr2; ++col) {
-      assert(pSecExprs[col].base.resSchema.bytes > 0);
-      resultRowSize += pSecExprs[col].base.resSchema.bytes;
-    }
-
-    if (resultRowSize > pQueryAttr->resultRowSize) {
-      pQueryAttr->resultRowSize = resultRowSize;
-    }
-  }
-
-  if (pQueryAttr->fillType != TSDB_FILL_NONE) {
-    pQueryAttr->fillVal = malloc(sizeof(int64_t) * pQueryAttr->numOfOutput);
-    if (pQueryAttr->fillVal == NULL) {
-      goto _cleanup;
-    }
-
-    // the first column is the timestamp
-    memcpy(pQueryAttr->fillVal, (char *)pQueryMsg->fillVal, pQueryAttr->numOfOutput * sizeof(int64_t));
-  }
-
-  size_t numOfGroups = 0;
-  if (pTableGroupInfo->pGroupList != NULL) {
-    numOfGroups = taosArrayGetSize(pTableGroupInfo->pGroupList);
-    STableGroupInfo* pTableqinfo = &pQInfo->runtimeEnv.tableqinfoGroupInfo;
-
-    pTableqinfo->pGroupList = taosArrayInit(numOfGroups, POINTER_BYTES);
-    pTableqinfo->numOfTables = pTableGroupInfo->numOfTables;
-    pTableqinfo->map = taosHashInit(pTableGroupInfo->numOfTables, taosGetDefaultHashFunction(TSDB_DATA_TYPE_INT), true, HASH_NO_LOCK);
-  }
-
-  pQInfo->pBuf = calloc(pTableGroupInfo->numOfTables, sizeof(STableQueryInfo));
-  if (pQInfo->pBuf == NULL) {
-    goto _cleanup;
-  }
-
-  pQInfo->dataReady = QUERY_RESULT_NOT_READY;
-  pQInfo->rspContext = NULL;
-  pQInfo->sql = sql;
-  pthread_mutex_init(&pQInfo->lock, NULL);
-  tsem_init(&pQInfo->ready, 0, 0);
-
-  pQueryAttr->window = pQueryMsg->window;
-  updateDataCheckOrder(pQInfo, pQueryMsg, pQueryAttr->stableQuery);
-
-  STaskRuntimeEnv* pRuntimeEnv = &pQInfo->runtimeEnv;
-  STimeWindow window = pQueryAttr->window;
-
-  int32_t index = 0;
-  for(int32_t i = 0; i < numOfGroups; ++i) {
-    SArray* pa = taosArrayGetP(pQueryAttr->tableGroupInfo.pGroupList, i);
-
-    size_t s = taosArrayGetSize(pa);
-    SArray* p1 = taosArrayInit(s, POINTER_BYTES);
-    if (p1 == NULL) {
-      goto _cleanup;
-    }
-
-    taosArrayPush(pRuntimeEnv->tableqinfoGroupInfo.pGroupList, &p1);
-
-    for(int32_t j = 0; j < s; ++j) {
-//      STableKeyInfo* info = taosArrayGet(pa, j);
-//      window.skey = info->lastKey;
-//
-//      void* buf = (char*) pQInfo->pBuf + index * sizeof(STableQueryInfo);
-//      STableQueryInfo* item = createTableQueryInfo(pQueryAttr, info->pTable, pQueryAttr->groupbyColumn, window, buf);
-//      if (item == NULL) {
-//        goto _cleanup;
-//      }
-//
-//      item->groupIndex = i;
-//      taosArrayPush(p1, &item);
-
-//      STableId* id = TSDB_TABLEID(info->pTable);
-//      taosHashPut(pRuntimeEnv->tableqinfoGroupInfo.map, &id->tid, sizeof(id->tid), &item, POINTER_BYTES);
-//      index += 1;
-    }
-  }
-
-  colIdCheck(pQueryAttr, pQInfo->qId);
-
-//  int32_t functionId = getExprFunctionId(&pExpr[0]);
-//  pQInfo->query.queryBlockDist = (functionId == FUNCTION_BLKINFO);
-
-  //qDebug("qmsg:%p vgId:%d, QInfo:0x%" PRIx64 "-%p created", pQueryMsg, pQInfo->query.vgId, pQInfo->qId, pQInfo);
-  return pQInfo;
-
-_cleanup_qinfo:
-//  tsdbDestroyTableGroup(pTableGroupInfo);
-
-  if (pGroupbyExpr != NULL) {
-    taosArrayDestroy(pGroupbyExpr->columnInfo);
-    free(pGroupbyExpr);
-  }
-
-  tfree(pTagCols);
-  for (int32_t i = 0; i < numOfOutput; ++i) {
-    SExprInfo* pExprInfo = &pExprs[i];
-    if (pExprInfo->pExpr != NULL) {
-      tExprTreeDestroy(pExprInfo->pExpr, NULL);
-      pExprInfo->pExpr = NULL;
-    }
-
-//    if (pExprInfo->base.flist.filterInfo) {
-//      freeColumnFilterInfo(pExprInfo->base.flist.filterInfo, pExprInfo->base.flist.numOfFilters);
-//    }
-  }
-
-  tfree(pExprs);
-
-//  filterFreeInfo(pFilters);
-
-_cleanup:
-  doDestroyTask(pQInfo);
-  return NULL;
-}
-
-bool isValidQInfo(void *param) {
-  SQInfo *pQInfo = (SQInfo *)param;
-  if (pQInfo == NULL) {
-    return false;
-  }
-
-  /*
-   * pQInfo->signature may be changed by another thread, so we assign value of signature
-   * into local variable, then compare by using local variable
-   */
-  uint64_t sig = (uint64_t)pQInfo->signature;
-  return (sig == (uint64_t)pQInfo);
-}
-
 int32_t initQInfo(STsBufInfo* pTsBufInfo, void* tsdb, void* sourceOptr, SQInfo* pQInfo, STaskParam* param, char* start,
                   int32_t prevResultLen, void* merger) {
   int32_t code = TSDB_CODE_SUCCESS;
@@ -8957,7 +8745,7 @@ int32_t initQInfo(STsBufInfo* pTsBufInfo, void* tsdb, void* sourceOptr, SQInfo*
 
 _error:
   // table query ref will be decrease during error handling
-  doDestroyTask(pQInfo);
+//  doDestroyTask(pQInfo);
   return code;
 }
 
@@ -9038,113 +8826,16 @@ void* freeColumnInfo(SColumnInfo* pColumnInfo, int32_t numOfCols) {
   return NULL;
 }
 
-void doDestroyTask(SQInfo *pQInfo) {
-  if (!isValidQInfo(pQInfo)) {
-    return;
-  }
-
-  //qDebug("QInfo:0x%"PRIx64" start to free QInfo", pQInfo->qId);
-
-  STaskRuntimeEnv* pRuntimeEnv = &pQInfo->runtimeEnv;
-  releaseQueryBuf(pRuntimeEnv->tableqinfoGroupInfo.numOfTables);
-
-  doDestroyTableQueryInfo(&pRuntimeEnv->tableqinfoGroupInfo);
-  teardownQueryRuntimeEnv(&pQInfo->runtimeEnv);
-
-  STaskAttr *pQueryAttr = pQInfo->runtimeEnv.pQueryAttr;
-  freeQueryAttr(pQueryAttr);
-
-//  tsdbDestroyTableGroup(&pQueryAttr->tableGroupInfo);
-
-  tfree(pQInfo->pBuf);
-  tfree(pQInfo->sql);
-
-  taosArrayDestroy(pQInfo->summary.queryProfEvents);
-  taosHashCleanup(pQInfo->summary.operatorProfResults);
-
-  taosArrayDestroy(pRuntimeEnv->groupResInfo.pRows);
-  pQInfo->signature = 0;
-
-  //qDebug("QInfo:0x%"PRIx64" QInfo is freed", pQInfo->qId);
-
-  tfree(pQInfo);
-}
-
-int32_t doDumpQueryResult(SQInfo *pQInfo, char *data, int8_t compressed, int32_t *compLen) {
-  // the remained number of retrieved rows, not the interpolated result
-  STaskRuntimeEnv* pRuntimeEnv = &pQInfo->runtimeEnv;
-  STaskAttr *pQueryAttr = pQInfo->runtimeEnv.pQueryAttr;
-
-  // load data from file to msg buffer
-  if (pQueryAttr->tsCompQuery) {
-    SColumnInfoData* pColInfoData = taosArrayGet(pRuntimeEnv->outputBuf->pDataBlock, 0);
-    FILE *f = *(FILE **)pColInfoData->pData;  // TODO refactor
-
-    // make sure file exist
-    if (f) {
-      off_t s = lseek(fileno(f), 0, SEEK_END);
-      assert(s == pRuntimeEnv->outputBuf->info.rows);
-
-      //qDebug("QInfo:0x%"PRIx64" ts comp data return, file:%p, size:%"PRId64, pQInfo->qId, f, (uint64_t)s);
-      if (fseek(f, 0, SEEK_SET) >= 0) {
-        size_t sz = fread(data, 1, s, f);
-        if(sz < s) {  // todo handle error
-          //qError("fread(f:%p,%d) failed, rsize:%" PRId64 ", expect size:%" PRId64, f, fileno(f), (uint64_t)sz, (uint64_t)s);
-          assert(0);
-        }
-      } else {
-        UNUSED(s);
-        //qError("fseek(f:%p,%d) failed, error:%s", f, fileno(f), strerror(errno));
-        assert(0);
-      }
-
-      // dump error info
-      if (s <= (sizeof(STSBufFileHeader) + sizeof(STSGroupBlockInfo) + 6 * sizeof(int32_t))) {
-//        qDump(data, s);
-        assert(0);
-      }
-
-      fclose(f);
-      *(FILE **)pColInfoData->pData = NULL;
-    }
-
-    // all data returned, set query over
-    if (Q_STATUS_EQUAL(pRuntimeEnv->status, TASK_COMPLETED)) {
-//      setTaskStatus(pOperator->pTaskInfo, QUERY_OVER);
-    }
-  } else {
-    doCopyQueryResultToMsg(pQInfo, (int32_t)pRuntimeEnv->outputBuf->info.rows, data, compressed, compLen);
-  }
-
-  //qDebug("QInfo:0x%"PRIx64" current numOfRes rows:%d, total:%" PRId64, pQInfo->qId,
-//         pRuntimeEnv->outputBuf->info.rows, pRuntimeEnv->resultInfo.total);
-
-  if (pQueryAttr->limit.limit > 0 && pQueryAttr->limit.limit == pRuntimeEnv->resultInfo.total) {
-    //qDebug("QInfo:0x%"PRIx64" results limitation reached, limitation:%"PRId64, pQInfo->qId, pQueryAttr->limit.limit);
-//    setTaskStatus(pOperator->pTaskInfo, QUERY_OVER);
-  }
-
-  return TSDB_CODE_SUCCESS;
-}
-
-bool doBuildResCheck(SQInfo* pQInfo) {
-  bool buildRes = false;
-
-  pthread_mutex_lock(&pQInfo->lock);
-
-  pQInfo->dataReady = QUERY_RESULT_READY;
-  buildRes = needBuildResAfterQueryComplete(pQInfo);
-
-  // clear qhandle owner, it must be in the secure area. other thread may run ahead before current, after it is
-  // put into task to be executed.
-  assert(pQInfo->owner == taosGetSelfPthreadId());
-  pQInfo->owner = 0;
+void doDestroyTask(SExecTaskInfo *pTaskInfo) {
+  doDestroyTableQueryInfo(&pTaskInfo->tableqinfoGroupInfo);
+//  taosArrayDestroy(pTaskInfo->summary.queryProfEvents);
+//  taosHashCleanup(pTaskInfo->summary.operatorProfResults);
 
-  pthread_mutex_unlock(&pQInfo->lock);
+  tfree(pTaskInfo->sql);
+  tfree(pTaskInfo->id.str);
+  qDebug("%s execTask is freed", GET_TASKID(pTaskInfo));
 
-  // used in retrieve blocking model.
-  tsem_post(&pQInfo->ready);
-  return buildRes;
+  tfree(pTaskInfo);
 }
 
 static void doSetTagValueToResultBuf(char* output, const char* val, int16_t type, int16_t bytes) {

source/libs/executor/test/executorTests.cpp

@@ -219,7 +219,7 @@ TEST(testCase, build_executor_tree_Test) {
 
   SExecTaskInfo* pTaskInfo = nullptr;
   DataSinkHandle sinkHandle = nullptr;
-  int32_t code = qCreateExecTask((void*) 1, 2, NULL, (void**) &pTaskInfo, &sinkHandle);
+  int32_t code = qCreateExecTask((void*) 1, 2, 1, NULL, (void**) &pTaskInfo, &sinkHandle);
 }
 
 #pragma GCC diagnostic pop
\ No newline at end of file

source/libs/planner/src/logicPlan.c

@@ -413,7 +413,6 @@ static void doDestroyQueryNode(SQueryPlanNode* pQueryNode) {
     tfree(pQueryTableInfo->tableName);
   }
 
-  printf("----------->Free:%p\n", pQueryNode->pExpr);
   taosArrayDestroy(pQueryNode->pExpr);
 
   tfree(pQueryNode->pExtInfo);

source/libs/planner/src/physicalPlanJson.c

@@ -1125,8 +1125,8 @@ int32_t subPlanToString(const SSubplan* subplan, char** str, int32_t* len) {
   *str = cJSON_Print(json);
   cJSON_Delete(json);
 
-  /*printf("====Physical plan:====\n");*/
-  /*printf("%s\n", *str);*/
+//  printf("====Physical plan:====\n");
+//  printf("%s\n", *str);
   *len = strlen(*str) + 1;
   return TSDB_CODE_SUCCESS;
 }

source/libs/qworker/inc/qworkerInt.h

@@ -137,6 +137,7 @@ typedef struct SQWorkerMgmt {
   SHashObj        *ctxHash;       //key: queryId+taskId, value: SQWTaskCtx
   void            *nodeObj;
   putReqToQueryQFp putToQueueFp;
+  sendReqToDnodeFp sendReqFp;
 } SQWorkerMgmt;
 
 #define QW_FPARAMS_DEF SQWorkerMgmt *mgmt, uint64_t sId, uint64_t qId, uint64_t tId
@@ -173,17 +174,17 @@ typedef struct SQWorkerMgmt {
 #define QW_SCH_ELOG(param, ...) qError("QW:%p SID:%"PRIx64" " param, mgmt, sId, __VA_ARGS__)
 #define QW_SCH_DLOG(param, ...) qDebug("QW:%p SID:%"PRIx64" " param, mgmt, sId, __VA_ARGS__)
 
-#define QW_TASK_ELOG(param, ...) qError("QW:%p QID:0x%"PRIx64",TID:%"PRIx64" " param, mgmt, qId, tId, __VA_ARGS__)
-#define QW_TASK_WLOG(param, ...) qWarn("QW:%p QID:0x%"PRIx64",TID:%"PRIx64" " param, mgmt, qId, tId, __VA_ARGS__)
-#define QW_TASK_DLOG(param, ...) qDebug("QW:%p QID:0x%"PRIx64",TID:%"PRIx64" " param, mgmt, qId, tId, __VA_ARGS__)
+#define QW_TASK_ELOG(param, ...) qError("QW:%p QID:0x%"PRIx64",TID:0x%"PRIx64" " param, mgmt, qId, tId, __VA_ARGS__)
+#define QW_TASK_WLOG(param, ...) qWarn("QW:%p QID:0x%"PRIx64",TID:0x%"PRIx64" " param, mgmt, qId, tId, __VA_ARGS__)
+#define QW_TASK_DLOG(param, ...) qDebug("QW:%p QID:0x%"PRIx64",TID:0x%"PRIx64" " param, mgmt, qId, tId, __VA_ARGS__)
 
-#define QW_TASK_ELOG_E(param) qError("QW:%p QID:0x%"PRIx64",TID:%"PRIx64" " param, mgmt, qId, tId)
-#define QW_TASK_WLOG_E(param) qWarn("QW:%p QID:0x%"PRIx64",TID:%"PRIx64" " param, mgmt, qId, tId)
-#define QW_TASK_DLOG_E(param) qDebug("QW:%p QID:0x%"PRIx64",TID:%"PRIx64" " param, mgmt, qId, tId)
+#define QW_TASK_ELOG_E(param) qError("QW:%p QID:0x%"PRIx64",TID:0x%"PRIx64" " param, mgmt, qId, tId)
+#define QW_TASK_WLOG_E(param) qWarn("QW:%p QID:0x%"PRIx64",TID:0x%"PRIx64" " param, mgmt, qId, tId)
+#define QW_TASK_DLOG_E(param) qDebug("QW:%p QID:0x%"PRIx64",TID:0x%"PRIx64" " param, mgmt, qId, tId)
 
-#define QW_SCH_TASK_ELOG(param, ...) qError("QW:%p SID:%"PRIx64",QID:0x%"PRIx64",TID:%"PRIx64" " param, mgmt, sId, qId, tId, __VA_ARGS__)
+#define QW_SCH_TASK_ELOG(param, ...) qError("QW:%p SID:%"PRIx64",QID:0x%"PRIx64",TID:0x%"PRIx64" " param, mgmt, sId, qId, tId, __VA_ARGS__)
 #define QW_SCH_TASK_WLOG(param, ...) qWarn("QW:%p SID:%"PRIx64",QID:0x%"PRIx64",TID:%"PRIx64" " param, mgmt, sId, qId, tId, __VA_ARGS__)
-#define QW_SCH_TASK_DLOG(param, ...) qDebug("QW:%p SID:%"PRIx64",QID:0x%"PRIx64",TID:%"PRIx64" " param, mgmt, sId, qId, tId, __VA_ARGS__)
+#define QW_SCH_TASK_DLOG(param, ...) qDebug("QW:%p SID:%"PRIx64",QID:0x%"PRIx64",TID:0x%"PRIx64" " param, mgmt, sId, qId, tId, __VA_ARGS__)
 
 #define QW_LOCK_DEBUG(...) do { if (gQWDebug.lockDebug) { qDebug(__VA_ARGS__); } } while (0)
 

source/libs/qworker/inc/qworkerMsg.h

@@ -32,7 +32,7 @@ int32_t qwProcessDrop(SQWorkerMgmt *mgmt, uint64_t sId, uint64_t qId, uint64_t t
 int32_t qwBuildAndSendDropRsp(void *connection, int32_t code);
 int32_t qwBuildAndSendCancelRsp(SRpcMsg *pMsg, int32_t code);
 int32_t qwBuildAndSendFetchRsp(void *connection, SRetrieveTableRsp *pRsp, int32_t dataLength, int32_t code);
-void qwBuildFetchRsp(void *msg, SOutputData *input, int32_t len);
+void qwBuildFetchRsp(void *msg, SOutputData *input, int32_t len, bool qComplete);
 int32_t qwBuildAndSendCQueryMsg(SQWorkerMgmt *mgmt, uint64_t sId, uint64_t qId, uint64_t tId, void *connection);
 int32_t qwBuildAndSendSchSinkMsg(SQWorkerMgmt *mgmt, uint64_t sId, uint64_t qId, uint64_t tId, void *connection);
 int32_t qwBuildAndSendReadyRsp(void *connection, int32_t code);

source/libs/qworker/src/qworker.c

@@ -456,7 +456,7 @@ _return:
   QW_RET(code);
 }
 
-int32_t qwExecTask(QW_FPARAMS_DEF, SQWTaskCtx *ctx) {
+int32_t qwExecTask(QW_FPARAMS_DEF, SQWTaskCtx *ctx, bool *queryEnd) {
   int32_t code = 0;
   bool  qcontinue = true;
   SSDataBlock* pRes = NULL;
@@ -467,11 +467,11 @@ int32_t qwExecTask(QW_FPARAMS_DEF, SQWTaskCtx *ctx) {
   DataSinkHandle sinkHandle = ctx->sinkHandle;
  
   while (true) {
-    QW_TASK_DLOG("start to execTask in executor, loopIdx:%d", i++);
+    QW_TASK_DLOG("start to execTask, loopIdx:%d", i++);
     
     code = qExecTask(*taskHandle, &pRes, &useconds);
     if (code) {
-      QW_TASK_ELOG("qExecTask failed, code:%x", code);
+      QW_TASK_ELOG("qExecTask failed, code:%s", tstrerror(code));
       QW_ERR_JRET(code);
     }
 
@@ -485,6 +485,10 @@ int32_t qwExecTask(QW_FPARAMS_DEF, SQWTaskCtx *ctx) {
       if (TASK_TYPE_TEMP == ctx->taskType) {
         qwFreeTaskHandle(QW_FPARAMS(), taskHandle);
       }
+
+      if (queryEnd) {
+        *queryEnd = true;
+      }
       
       break;
     }
@@ -492,7 +496,7 @@ int32_t qwExecTask(QW_FPARAMS_DEF, SQWTaskCtx *ctx) {
     SInputData inputData = {.pData = pRes, .pTableRetrieveTsMap = NULL};
     code = dsPutDataBlock(sinkHandle, &inputData, &qcontinue);
     if (code) {
-      QW_TASK_ELOG("dsPutDataBlock failed, code:%x", code);
+      QW_TASK_ELOG("dsPutDataBlock failed, code:%s", tstrerror(code));
       QW_ERR_JRET(code);
     }
 
@@ -587,12 +591,7 @@ int32_t qwGetResFromSink(QW_FPARAMS_DEF, SQWTaskCtx *ctx, int32_t *dataLen, void
     QW_ERR_RET(code);
   }
 
-  queryEnd = pOutput->queryEnd;
-  pOutput->queryEnd = false;
-
-  if (DS_BUF_EMPTY == pOutput->bufStatus && queryEnd) {
-    pOutput->queryEnd = true;
-    
+  if (DS_BUF_EMPTY == pOutput->bufStatus && pOutput->queryEnd) {
     QW_SCH_TASK_DLOG("task all fetched, status:%d", JOB_TASK_STATUS_SUCCEED);
     QW_ERR_RET(qwUpdateTaskStatus(QW_FPARAMS(), JOB_TASK_STATUS_SUCCEED));
   }
@@ -974,7 +973,7 @@ int32_t qwProcessQuery(QW_FPARAMS_DEF, SQWMsg *qwMsg, int8_t taskType) {
     QW_ERR_JRET(code);
   }
   
-  code = qCreateExecTask(qwMsg->node, 0, (struct SSubplan *)plan, &pTaskInfo, &sinkHandle);
+  code = qCreateExecTask(qwMsg->node, 0, tId, (struct SSubplan *)plan, &pTaskInfo, &sinkHandle);
   if (code) {
     QW_TASK_ELOG("qCreateExecTask failed, code:%s", tstrerror(code));
     QW_ERR_JRET(code);
@@ -996,7 +995,7 @@ int32_t qwProcessQuery(QW_FPARAMS_DEF, SQWMsg *qwMsg, int8_t taskType) {
   atomic_store_ptr(&ctx->sinkHandle, sinkHandle);
 
   if (pTaskInfo && sinkHandle) {
-    QW_ERR_JRET(qwExecTask(QW_FPARAMS(), ctx));
+    QW_ERR_JRET(qwExecTask(QW_FPARAMS(), ctx, NULL));
   }
   
 _return:
@@ -1083,6 +1082,7 @@ int32_t qwProcessCQuery(QW_FPARAMS_DEF, SQWMsg *qwMsg) {
   SQWPhaseOutput output = {0};
   void *rsp = NULL;
   int32_t dataLen = 0;
+  bool queryEnd = false;
   
   do {
     QW_ERR_JRET(qwHandlePrePhaseEvents(QW_FPARAMS(), QW_PHASE_PRE_CQUERY, &input, &output));
@@ -1102,7 +1102,7 @@ int32_t qwProcessCQuery(QW_FPARAMS_DEF, SQWMsg *qwMsg) {
 
     DataSinkHandle  sinkHandle = ctx->sinkHandle;
 
-    QW_ERR_JRET(qwExecTask(QW_FPARAMS(), ctx));
+    QW_ERR_JRET(qwExecTask(QW_FPARAMS(), ctx, &queryEnd));
 
     if (QW_IS_EVENT_RECEIVED(ctx, QW_EVENT_FETCH)) {
       SOutputData sOutput = {0};
@@ -1114,13 +1114,10 @@ int32_t qwProcessCQuery(QW_FPARAMS_DEF, SQWMsg *qwMsg) {
         // RC WARNING
         atomic_store_8(&ctx->queryContinue, 1);
       }
-
-      if (sOutput.queryEnd) {
-        needStop = true;
-      }
       
       if (rsp) {
-        qwBuildFetchRsp(rsp, &sOutput, dataLen); 
+        bool qComplete = (DS_BUF_EMPTY == sOutput.bufStatus && sOutput.queryEnd);
+        qwBuildFetchRsp(rsp, &sOutput, dataLen, qComplete);
         
         QW_SET_EVENT_PROCESSED(ctx, QW_EVENT_FETCH);            
         
@@ -1131,6 +1128,10 @@ int32_t qwProcessCQuery(QW_FPARAMS_DEF, SQWMsg *qwMsg) {
       }
     }
 
+    if (queryEnd) {
+      needStop = true;
+    }
+
   _return:
 
     if (NULL == ctx) {
@@ -1196,7 +1197,8 @@ int32_t qwProcessFetch(QW_FPARAMS_DEF, SQWMsg *qwMsg) {
   if (NULL == rsp) {
     QW_SET_EVENT_RECEIVED(ctx, QW_EVENT_FETCH);
   } else {
-    qwBuildFetchRsp(rsp, &sOutput, dataLen);
+    bool qComplete = (DS_BUF_EMPTY == sOutput.bufStatus && sOutput.queryEnd);
+    qwBuildFetchRsp(rsp, &sOutput, dataLen, qComplete);
   }
 
   if ((!sOutput.queryEnd) && (DS_BUF_LOW == sOutput.bufStatus || DS_BUF_EMPTY == sOutput.bufStatus)) {    
@@ -1311,12 +1313,13 @@ _return:
   QW_RET(code);
 }
 
-int32_t qWorkerInit(int8_t nodeType, int32_t nodeId, SQWorkerCfg *cfg, void **qWorkerMgmt, void *nodeObj, putReqToQueryQFp fp) {
-  if (NULL == qWorkerMgmt || NULL == nodeObj || NULL == fp) {
+int32_t qWorkerInit(int8_t nodeType, int32_t nodeId, SQWorkerCfg *cfg, void **qWorkerMgmt, void *nodeObj,
+                    putReqToQueryQFp fp1, sendReqToDnodeFp fp2) {
+  if (NULL == qWorkerMgmt || NULL == nodeObj || NULL == fp1 || NULL == fp2) {
     qError("invalid param to init qworker");
     QW_RET(TSDB_CODE_QRY_INVALID_INPUT);
   }
-  
+
   SQWorkerMgmt *mgmt = calloc(1, sizeof(SQWorkerMgmt));
   if (NULL == mgmt) {
     qError("calloc %d failed", (int32_t)sizeof(SQWorkerMgmt));
@@ -1359,7 +1362,8 @@ int32_t qWorkerInit(int8_t nodeType, int32_t nodeId, SQWorkerCfg *cfg, void **qW
   mgmt->nodeType = nodeType;
   mgmt->nodeId = nodeId;
   mgmt->nodeObj = nodeObj;
-  mgmt->putToQueueFp = fp;
+  mgmt->putToQueueFp = fp1;
+  mgmt->sendReqFp = fp2;
 
   *qWorkerMgmt = mgmt;
 

source/libs/qworker/src/qworkerMsg.c

@@ -26,11 +26,11 @@ int32_t qwMallocFetchRsp(int32_t length, SRetrieveTableRsp **rsp) {
   return TSDB_CODE_SUCCESS;
 }
 
-void qwBuildFetchRsp(void *msg, SOutputData *input, int32_t len) {
+void qwBuildFetchRsp(void *msg, SOutputData *input, int32_t len, bool qComplete) {
   SRetrieveTableRsp *rsp = (SRetrieveTableRsp *)msg;
   
   rsp->useconds = htobe64(input->useconds);
-  rsp->completed = input->queryEnd;
+  rsp->completed = qComplete;
   rsp->precision = input->precision;
   rsp->compressed = input->compressed;
   rsp->compLen = htonl(len);
@@ -258,12 +258,12 @@ int32_t qwBuildAndSendCQueryMsg(SQWorkerMgmt *mgmt, uint64_t sId, uint64_t qId,
 
   int32_t code = (*mgmt->putToQueueFp)(mgmt->nodeObj, &pNewMsg);
   if (TSDB_CODE_SUCCESS != code) {
-    QW_SCH_TASK_ELOG("put query continue msg to queue failed, code:%x", code);
+    QW_SCH_TASK_ELOG("put query continue msg to queue failed, vgId:%d, code:%s", mgmt->nodeId, tstrerror(code));
     rpcFreeCont(req);
     QW_ERR_RET(code);
   }
 
-  QW_SCH_TASK_DLOG("put query continue msg to query queue, vgId:%d", mgmt->nodeId);
+  QW_SCH_TASK_DLOG("put task continue exec msg to query queue, vgId:%d", mgmt->nodeId);
 
   return TSDB_CODE_SUCCESS;
 }
@@ -421,6 +421,11 @@ int32_t qWorkerProcessFetchMsg(void *node, void *qWorkerMgmt, SRpcMsg *pMsg) {
   return TSDB_CODE_SUCCESS;  
 }
 
+int32_t qWorkerProcessFetchRsp(void *node, void *qWorkerMgmt, SRpcMsg *pMsg) {
+  qProcessFetchRsp(NULL, pMsg, NULL);
+  return TSDB_CODE_SUCCESS;
+}
+
 int32_t qWorkerProcessCancelMsg(void *node, void *qWorkerMgmt, SRpcMsg *pMsg) {
   if (NULL == node || NULL == qWorkerMgmt || NULL == pMsg) {
     return TSDB_CODE_QRY_INVALID_INPUT;

source/libs/qworker/test/qworkerTests.cpp

@@ -1081,7 +1081,7 @@ TEST(rcTest, shortExecshortDelay) {
   qwtTestStop = false;
   qwtTestQuitThreadNum = 0;
 
-  code = qWorkerInit(NODE_TYPE_VNODE, 1, NULL, &mgmt, mockPointer, qwtPutReqToQueue);
+  code = qWorkerInit(NODE_TYPE_VNODE, 1, NULL, &mgmt, mockPointer, qwtPutReqToQueue, NULL);
   ASSERT_EQ(code, 0);
 
   qwtTestMaxExecTaskUsec = 0;
@@ -1162,7 +1162,7 @@ TEST(rcTest, longExecshortDelay) {
   qwtTestStop = false;
   qwtTestQuitThreadNum = 0;
 
-  code = qWorkerInit(NODE_TYPE_VNODE, 1, NULL, &mgmt, mockPointer, qwtPutReqToQueue);
+  code = qWorkerInit(NODE_TYPE_VNODE, 1, NULL, &mgmt, mockPointer, qwtPutReqToQueue, NULL);
   ASSERT_EQ(code, 0);
 
   qwtTestMaxExecTaskUsec = 1000000;
@@ -1245,7 +1245,7 @@ TEST(rcTest, shortExeclongDelay) {
   qwtTestStop = false;
   qwtTestQuitThreadNum = 0;
 
-  code = qWorkerInit(NODE_TYPE_VNODE, 1, NULL, &mgmt, mockPointer, qwtPutReqToQueue);
+  code = qWorkerInit(NODE_TYPE_VNODE, 1, NULL, &mgmt, mockPointer, qwtPutReqToQueue, NULL);
   ASSERT_EQ(code, 0);
 
   qwtTestMaxExecTaskUsec = 0;

source/libs/transport/inc/transComm.h

@@ -182,7 +182,7 @@ typedef struct {
 
 #define TRANS_RESERVE_SIZE (sizeof(STranConnCtx))
 
-#define TRANS_MSG_OVERHEAD (sizeof(STransMsgHead) + sizeof(STransDigestMsg))
+#define TRANS_MSG_OVERHEAD (sizeof(STransMsgHead))
 #define transHeadFromCont(cont) ((STransMsgHead*)((char*)cont - sizeof(STransMsgHead)))
 #define transContFromHead(msg) (msg + sizeof(STransMsgHead))
 #define transMsgLenFromCont(contLen) (contLen + sizeof(STransMsgHead))
@@ -201,6 +201,9 @@ bool transDecompressMsg(char* msg, int32_t len, int32_t* flen);
 
 void transConnCtxDestroy(STransConnCtx* ctx);
 
+void transFreeMsg(void* msg);
+
+//
 typedef struct SConnBuffer {
   char* buf;
   int   len;
@@ -208,4 +211,9 @@ typedef struct SConnBuffer {
   int   left;
 } SConnBuffer;
 
+int transInitBuffer(SConnBuffer* buf);
+int transClearBuffer(SConnBuffer* buf);
+int transDestroyBuffer(SConnBuffer* buf);
+int transAllocBuffer(SConnBuffer* connBuf, uv_buf_t* uvBuf);
+
 #endif

source/libs/transport/src/rpcMain.c

@@ -242,11 +242,14 @@ void *rpcOpen(const SRpcInit *pInit) {
   pRpc = (SRpcInfo *)calloc(1, sizeof(SRpcInfo));
   if (pRpc == NULL) return NULL;
 
-  if (pInit->label) tstrncpy(pRpc->label, pInit->label, strlen(pInit->label));
+  if (pInit->label) tstrncpy(pRpc->label, pInit->label, tListLen(pInit->label));
 
   pRpc->connType = pInit->connType;
   if (pRpc->connType == TAOS_CONN_CLIENT) {
     pRpc->numOfThreads = pInit->numOfThreads;
+    if (pRpc->numOfThreads >= 10) {
+      pRpc->numOfThreads = 10;
+    }
   } else {
     pRpc->numOfThreads = pInit->numOfThreads > TSDB_MAX_RPC_THREADS ? TSDB_MAX_RPC_THREADS : pInit->numOfThreads;
   }
@@ -769,8 +772,8 @@ static SRpcConn *rpcAllocateServerConn(SRpcInfo *pRpc, SRecvInfo *pRecv) {
     }
 
     taosHashPut(pRpc->hash, hashstr, size, (char *)&pConn, POINTER_BYTES);
-    tDebug("%s %p server connection is allocated, uid:0x%x sid:%d key:%s spi:%d", pRpc->label, pConn, pConn->linkUid, sid,
-           hashstr, pConn->spi);
+    tDebug("%s %p server connection is allocated, uid:0x%x sid:%d key:%s spi:%d", pRpc->label, pConn, pConn->linkUid,
+           sid, hashstr, pConn->spi);
   }
 
   return pConn;
@@ -1357,9 +1360,14 @@ static void rpcSendMsgToPeer(SRpcConn *pConn, void *msg, int msgLen) {
     tDebug("%s, %s is sent to %s:%hu, len:%d sig:0x%08x:0x%08x:%d", pConn->info, TMSG_INFO(pHead->msgType),
            pConn->peerFqdn, pConn->peerPort, msgLen, pHead->sourceId, pHead->destId, pHead->tranId);
   } else {
-    if (pHead->code == 0) pConn->secured = 1;  // for success response, set link as secured
-    tDebug("%s, %s is sent to 0x%x:%hu, code:0x%x len:%d sig:0x%08x:0x%08x:%d", pConn->info, TMSG_INFO(pHead->msgType),
-           pConn->peerIp, pConn->peerPort, htonl(pHead->code), msgLen, pHead->sourceId, pHead->destId, pHead->tranId);
+    if (pHead->code == 0) {
+      pConn->secured = 1;  // for success response, set link as secured
+    }
+
+    char ipport[40] = {0};
+    taosIpPort2String(pConn->peerIp, pConn->peerPort, ipport);
+    tDebug("%s, %s is sent to %s, code:0x%x len:%d sig:0x%08x:0x%08x:%d", pConn->info, TMSG_INFO(pHead->msgType),
+           ipport, htonl(pHead->code), msgLen, pHead->sourceId, pHead->destId, pHead->tranId);
   }
 
   // tTrace("connection type is: %d", pConn->connType);

source/libs/transport/src/trans.c

@@ -30,7 +30,8 @@ void* rpcOpen(const SRpcInit* pInit) {
     tstrncpy(pRpc->label, pInit->label, strlen(pInit->label));
   }
   pRpc->cfp = pInit->cfp;
-  pRpc->numOfThreads = pInit->numOfThreads > TSDB_MAX_RPC_THREADS ? TSDB_MAX_RPC_THREADS : pInit->numOfThreads;
+  // pRpc->numOfThreads = pInit->numOfThreads > TSDB_MAX_RPC_THREADS ? TSDB_MAX_RPC_THREADS : pInit->numOfThreads;
+  pRpc->numOfThreads = pInit->numOfThreads;
   pRpc->connType = pInit->connType;
   pRpc->idleTime = pInit->idleTime;
   pRpc->tcphandle = (*taosInitHandle[pRpc->connType])(0, pInit->localPort, pRpc->label, pRpc->numOfThreads, NULL, pRpc);
@@ -55,7 +56,13 @@ void* rpcMallocCont(int contLen) {
   }
   return start + sizeof(STransMsgHead);
 }
-void  rpcFreeCont(void* cont) { return; }
+void rpcFreeCont(void* cont) {
+  // impl
+  if (cont == NULL) {
+    return;
+  }
+  free((char*)cont - TRANS_MSG_OVERHEAD);
+}
 void* rpcReallocCont(void* ptr, int contLen) { return NULL; }
 
 void rpcSendRedirectRsp(void* pConn, const SEpSet* pEpSet) {}

source/libs/transport/src/transCli.c

@@ -31,6 +31,7 @@ typedef struct SCliConn {
   char         secured;
   uint64_t     expireTime;
   int8_t       notifyCount;  // timers already notify to client
+  int32_t      ref;
 } SCliConn;
 
 typedef struct SCliMsg {
@@ -105,15 +106,28 @@ static void* clientThread(void* arg);
 static void clientHandleResp(SCliConn* conn) {
   STransConnCtx* pCtx = ((SCliMsg*)conn->data)->ctx;
   SRpcInfo*      pRpc = pCtx->pTransInst;
-  SRpcMsg        rpcMsg;
 
-  rpcMsg.pCont = conn->readBuf.buf;
-  rpcMsg.contLen = conn->readBuf.len;
+  STransMsgHead* pHead = (STransMsgHead*)(conn->readBuf.buf);
+  pHead->code = htonl(pHead->code);
+  pHead->msgLen = htonl(pHead->msgLen);
+
+  SRpcMsg rpcMsg;
+  rpcMsg.contLen = transContLenFromMsg(pHead->msgLen);
+  rpcMsg.pCont = transContFromHead((char*)pHead);
+  rpcMsg.code = pHead->code;
+  rpcMsg.msgType = pHead->msgType;
   rpcMsg.ahandle = pCtx->ahandle;
+
+  tDebug("conn %p handle resp", conn);
   (pRpc->cfp)(NULL, &rpcMsg, NULL);
   conn->notifyCount += 1;
 
   SCliThrdObj* pThrd = conn->hostThrd;
+  tfree(conn->data);
+  // buf alread translated to rpcMsg.pCont
+  transClearBuffer(&conn->readBuf);
+
+  uv_read_start((uv_stream_t*)conn->stream, clientAllocBufferCb, clientReadCb);
   addConnToPool(pThrd->pool, pCtx->ip, pCtx->port, conn);
 
   // start thread's timer of conn pool if not active
@@ -123,16 +137,24 @@ static void clientHandleResp(SCliConn* conn) {
   destroyTransConnCtx(pCtx);
 }
 static void clientHandleExcept(SCliConn* pConn) {
+  if (pConn->data == NULL) {
+    clientConnDestroy(pConn, true);
+    return;
+  }
+  tDebug("conn %p destroy", pConn);
   SCliMsg* pMsg = pConn->data;
+  transFreeMsg((pMsg->msg.pCont));
+  pMsg->msg.pCont = NULL;
 
   STransConnCtx* pCtx = pMsg->ctx;
   SRpcInfo*      pRpc = pCtx->pTransInst;
 
-  SRpcMsg rpcMsg;
+  SRpcMsg rpcMsg = {0};
   rpcMsg.ahandle = pCtx->ahandle;
+  rpcMsg.code = -1;
   // SRpcInfo* pRpc = pMsg->ctx->pRpc;
   (pRpc->cfp)(NULL, &rpcMsg, NULL);
-
+  tfree(pConn->data);
   pConn->notifyCount += 1;
   destroyTransConnCtx(pCtx);
   clientConnDestroy(pConn, true);
@@ -201,12 +223,17 @@ static SCliConn* getConnFromPool(void* pool, char* ip, uint32_t port) {
   }
   queue* h = QUEUE_HEAD(&plist->conn);
   QUEUE_REMOVE(h);
-  return QUEUE_DATA(h, SCliConn, conn);
+
+  SCliConn* conn = QUEUE_DATA(h, SCliConn, conn);
+  QUEUE_INIT(&conn->conn);
+  return conn;
 }
 static void addConnToPool(void* pool, char* ip, uint32_t port, SCliConn* conn) {
   char key[128] = {0};
+
   tstrncpy(key, ip, strlen(ip));
   tstrncpy(key + strlen(key), (char*)(&port), sizeof(port));
+  tDebug("conn %p added to conn pool, read buf cap: %d", conn, conn->readBuf.cap);
 
   SRpcInfo* pRpc = ((SCliThrdObj*)conn->hostThrd)->pTransInst;
   conn->expireTime = taosGetTimestampMs() + CONN_PERSIST_TIME(pRpc->idleTime);
@@ -225,40 +252,18 @@ static bool clientReadComplete(SConnBuffer* data) {
     if (msgLen > data->len) {
       data->left = msgLen - data->len;
       return false;
-    } else {
+    } else if (msgLen == data->len) {
+      data->left = 0;
       return true;
     }
   } else {
     return false;
   }
 }
-static void clientAllocReadBufferCb(uv_handle_t* handle, size_t suggested_size, uv_buf_t* buf) {
-  // impl later
-  static const int CAPACITY = 512;
-
+static void clientAllocBufferCb(uv_handle_t* handle, size_t suggested_size, uv_buf_t* buf) {
   SCliConn*    conn = handle->data;
   SConnBuffer* pBuf = &conn->readBuf;
-  if (pBuf->cap == 0) {
-    pBuf->buf = (char*)calloc(1, CAPACITY * sizeof(char));
-    pBuf->len = 0;
-    pBuf->cap = CAPACITY;
-    pBuf->left = -1;
-
-    buf->base = pBuf->buf;
-    buf->len = CAPACITY;
-  } else {
-    if (pBuf->len >= pBuf->cap) {
-      if (pBuf->left == -1) {
-        pBuf->cap *= 2;
-        pBuf->buf = realloc(pBuf->buf, pBuf->cap);
-      } else if (pBuf->len + pBuf->left > pBuf->cap) {
-        pBuf->cap = pBuf->len + pBuf->left;
-        pBuf->buf = realloc(pBuf->buf, pBuf->cap);
-      }
-    }
-    buf->base = pBuf->buf + pBuf->len;
-    buf->len = pBuf->cap - pBuf->len;
-  }
+  transAllocBuffer(pBuf, buf);
 }
 static void clientReadCb(uv_stream_t* handle, ssize_t nread, const uv_buf_t* buf) {
   // impl later
@@ -267,6 +272,7 @@ static void clientReadCb(uv_stream_t* handle, ssize_t nread, const uv_buf_t* buf
   if (nread > 0) {
     pBuf->len += nread;
     if (clientReadComplete(pBuf)) {
+      uv_read_stop((uv_stream_t*)conn->stream);
       tDebug("conn %p read complete", conn);
       clientHandleResp(conn);
     } else {
@@ -276,10 +282,9 @@ static void clientReadCb(uv_stream_t* handle, ssize_t nread, const uv_buf_t* buf
   }
   assert(nread <= 0);
   if (nread == 0) {
-    tError("conn %p closed", conn);
     return;
   }
-  if (nread < 0) {
+  if (nread < 0 || nread == UV_EOF) {
     tError("conn %p read error: %s", conn, uv_err_name(nread));
     clientHandleExcept(conn);
   }
@@ -288,39 +293,46 @@ static void clientReadCb(uv_stream_t* handle, ssize_t nread, const uv_buf_t* buf
 }
 
 static void clientConnDestroy(SCliConn* conn, bool clear) {
-  tDebug("conn %p destroy", conn);
-  if (clear) {
-    uv_close((uv_handle_t*)conn->stream, NULL);
+  //
+  conn->ref--;
+  if (conn->ref == 0) {
+    tDebug("conn %p remove from conn pool", conn);
+    QUEUE_REMOVE(&conn->conn);
+    tDebug("conn %p remove from conn pool successfully", conn);
+    if (clear) {
+      uv_close((uv_handle_t*)conn->stream, clientDestroy);
+    }
   }
-  free(conn->stream);
-  free(conn->readBuf.buf);
-  free(conn->writeReq);
-  free(conn);
 }
 static void clientDestroy(uv_handle_t* handle) {
   SCliConn* conn = handle->data;
-  // QUEUE_REMOVE(&conn->conn);
-  clientConnDestroy(conn, false);
+  // transDestroyBuffer(&conn->readBuf);
+
+  free(conn->stream);
+  free(conn->writeReq);
+  tDebug("conn %p destroy successfully", conn);
+  free(conn);
+
+  // clientConnDestroy(conn, false);
 }
 
 static void clientWriteCb(uv_write_t* req, int status) {
   SCliConn* pConn = req->data;
-
   if (status == 0) {
     tDebug("conn %p data already was written out", pConn);
+    SCliMsg* pMsg = pConn->data;
+    if (pMsg != NULL) {
+      transFreeMsg((pMsg->msg.pCont));
+      pMsg->msg.pCont = NULL;
+    }
+
   } else {
     tError("conn %p failed to write: %s", pConn, uv_err_name(status));
     clientHandleExcept(pConn);
     return;
   }
   SCliThrdObj* pThrd = pConn->hostThrd;
-  // if (pConn->stream == NULL) {
-  //  pConn->stream = (uv_stream_t*)malloc(sizeof(uv_tcp_t));
-  //  uv_tcp_init(pThrd->loop, (uv_tcp_t*)pConn->stream);
-  //  pConn->stream->data = pConn;
-  //}
-  uv_read_start((uv_stream_t*)pConn->stream, clientAllocReadBufferCb, clientReadCb);
-  // impl later
+  uv_read_start((uv_stream_t*)pConn->stream, clientAllocBufferCb, clientReadCb);
 }
 
 static void clientWrite(SCliConn* pConn) {
@@ -365,14 +377,11 @@ static void clientHandleReq(SCliMsg* pMsg, SCliThrdObj* pThrd) {
     tDebug("conn %p get from conn pool", conn);
     conn->data = pMsg;
     conn->writeReq->data = conn;
-
-    conn->readBuf.len = 0;
-    memset(conn->readBuf.buf, 0, conn->readBuf.cap);
-    conn->readBuf.left = -1;
+    transDestroyBuffer(&conn->readBuf);
     clientWrite(conn);
   } else {
     SCliConn* conn = calloc(1, sizeof(SCliConn));
-
+    conn->ref++;
     // read/write stream handle
     conn->stream = (uv_stream_t*)malloc(sizeof(uv_tcp_t));
     uv_tcp_init(pThrd->loop, (uv_tcp_t*)(conn->stream));
@@ -381,6 +390,7 @@ static void clientHandleReq(SCliMsg* pMsg, SCliThrdObj* pThrd) {
     // write req handle
     conn->writeReq = malloc(sizeof(uv_write_t));
     conn->writeReq->data = conn;
+
     QUEUE_INIT(&conn->conn);
 
     conn->connReq.data = conn;

source/libs/transport/src/transComm.c

@@ -191,4 +191,58 @@ void transConnCtxDestroy(STransConnCtx* ctx) {
   free(ctx->ip);
   free(ctx);
 }
+
+void transFreeMsg(void* msg) {
+  if (msg == NULL) {
+    return;
+  }
+  free((char*)msg - sizeof(STransMsgHead));
+}
+
+int transInitBuffer(SConnBuffer* buf) {
+  transClearBuffer(buf);
+  return 0;
+}
+int transClearBuffer(SConnBuffer* buf) {
+  memset(buf, 0, sizeof(*buf));
+  return 0;
+}
+int transAllocBuffer(SConnBuffer* connBuf, uv_buf_t* uvBuf) {
+  /*
+   * formate of data buffer:
+   * |<--------------------------data from socket------------------------------->|
+   * |<------STransMsgHead------->|<-------------------other data--------------->|
+   */
+  static const int CAPACITY = 1024;
+
+  SConnBuffer* p = connBuf;
+  if (p->cap == 0) {
+    p->buf = (char*)calloc(CAPACITY, sizeof(char));
+    p->len = 0;
+    p->cap = CAPACITY;
+    p->left = -1;
+
+    uvBuf->base = p->buf;
+    uvBuf->len = CAPACITY;
+  } else {
+    if (p->len >= p->cap) {
+      if (p->left == -1) {
+        p->cap *= 2;
+        p->buf = realloc(p->buf, p->cap);
+      } else if (p->len + p->left > p->cap) {
+        p->cap = p->len + p->left;
+        p->buf = realloc(p->buf, p->len + p->left);
+      }
+    }
+    uvBuf->base = p->buf + p->len;
+    uvBuf->len = p->cap - p->len;
+  }
+  return 0;
+}
+int transDestroyBuffer(SConnBuffer* buf) {
+  if (buf->cap > 0) {
+    tfree(buf->buf);
+  }
+  transClearBuffer(buf);
+}
 #endif

source/libs/transport/src/transSrv.c

@@ -17,7 +17,7 @@
 
 #include "transComm.h"
 
-typedef struct SConn {
+typedef struct SSrvConn {
   uv_tcp_t*   pTcp;
   uv_write_t* pWriter;
   uv_timer_t* pTimer;
@@ -26,13 +26,14 @@ typedef struct SConn {
   queue       queue;
   int         ref;
   int         persist;  // persist connection or not
-  SConnBuffer connBuf;  // read buf,
+  SConnBuffer readBuf;  // read buf,
   int         inType;
   void*       pTransInst;  // rpc init
   void*       ahandle;     //
   void*       hostThrd;
+  void*       pSrvMsg;
 
-  SRpcMsg sendMsg;
+  // SRpcMsg sendMsg;
   // del later
   char secured;
   int  spi;
@@ -40,7 +41,13 @@ typedef struct SConn {
   char user[TSDB_UNI_LEN];  // user ID for the link
   char secret[TSDB_PASSWORD_LEN];
   char ckey[TSDB_PASSWORD_LEN];  // ciphering key
-} SConn;
+} SSrvConn;
+
+typedef struct SSrvMsg {
+  SSrvConn* pConn;
+  SRpcMsg   msg;
+  queue     q;
+} SSrvMsg;
 
 typedef struct SWorkThrdObj {
   pthread_t       thread;
@@ -48,8 +55,8 @@ typedef struct SWorkThrdObj {
   int             fd;
   uv_loop_t*      loop;
   uv_async_t*     workerAsync;  //
-  queue           conn;
-  pthread_mutex_t connMtx;
+  queue           msg;
+  pthread_mutex_t msgMtx;
   void*           pTransInst;
 } SWorkThrdObj;
 
@@ -68,9 +75,9 @@ typedef struct SServerObj {
 static const char* notify = "a";
 
 // refactor later
-static int transAddAuthPart(SConn* pConn, char* msg, int msgLen);
+static int transAddAuthPart(SSrvConn* pConn, char* msg, int msgLen);
 
-static int uvAuthMsg(SConn* pConn, char* msg, int msgLen);
+static int uvAuthMsg(SSrvConn* pConn, char* msg, int msgLen);
 
 static void uvAllocConnBufferCb(uv_handle_t* handle, size_t suggested_size, uv_buf_t* buf);
 static void uvAllocReadBufferCb(uv_handle_t* handle, size_t suggested_size, uv_buf_t* buf);
@@ -82,12 +89,13 @@ static void uvOnAcceptCb(uv_stream_t* stream, int status);
 static void uvOnConnectionCb(uv_stream_t* q, ssize_t nread, const uv_buf_t* buf);
 static void uvWorkerAsyncCb(uv_async_t* handle);
 
-static void uvPrepareSendData(SConn* conn, uv_buf_t* wb);
-
+static void uvPrepareSendData(SSrvMsg* msg, uv_buf_t* wb);
+static void uvStartSendResp(SSrvMsg* msg);
+static void destroySrvMsg(SSrvConn* conn);
 // check whether already read complete packet
-static bool   readComplete(SConnBuffer* buf);
-static SConn* createConn();
-static void   destroyConn(SConn* conn, bool clear /*clear handle or not*/);
+static bool      readComplete(SConnBuffer* buf);
+static SSrvConn* createConn();
+static void      destroyConn(SSrvConn* conn, bool clear /*clear handle or not*/);
 
 static void uvDestroyConn(uv_handle_t* handle);
 
@@ -105,31 +113,9 @@ void uvAllocReadBufferCb(uv_handle_t* handle, size_t suggested_size, uv_buf_t* b
    * |<--------------------------data from socket------------------------------->|
    * |<------STransMsgHead------->|<-------------------other data--------------->|
    */
-  static const int CAPACITY = 1024;
-
-  SConn*       conn = handle->data;
-  SConnBuffer* pBuf = &conn->connBuf;
-  if (pBuf->cap == 0) {
-    pBuf->buf = (char*)calloc(CAPACITY, sizeof(char));
-    pBuf->len = 0;
-    pBuf->cap = CAPACITY;
-    pBuf->left = -1;
-
-    buf->base = pBuf->buf;
-    buf->len = CAPACITY;
-  } else {
-    if (pBuf->len >= pBuf->cap) {
-      if (pBuf->left == -1) {
-        pBuf->cap *= 2;
-        pBuf->buf = realloc(pBuf->buf, pBuf->cap);
-      } else if (pBuf->len + pBuf->left > pBuf->cap) {
-        pBuf->cap = pBuf->len + pBuf->left;
-        pBuf->buf = realloc(pBuf->buf, pBuf->len + pBuf->left);
-      }
-    }
-    buf->base = pBuf->buf + pBuf->len;
-    buf->len = pBuf->cap - pBuf->len;
-  }
+  SSrvConn*    conn = handle->data;
+  SConnBuffer* pBuf = &conn->readBuf;
+  transAllocBuffer(pBuf, buf);
 }
 
 // check data read from socket completely or not
@@ -159,7 +145,7 @@ static bool readComplete(SConnBuffer* data) {
 //  // impl later
 //  STransMsgHead* pHead = (STransMsgHead*)pRecv->msg;
 //  SRpcInfo*      pRpc = (SRpcInfo*)pRecv->shandle;
-//  SConn*         pConn = pRecv->thandle;
+//  SSrvConn*         pConn = pRecv->thandle;
 //  tDump(pRecv->msg, pRecv->msgLen);
 //  terrno = 0;
 //  // SRpcReqContext* pContest;
@@ -167,7 +153,7 @@ static bool readComplete(SConnBuffer* data) {
 //  // do auth and check
 //}
 
-static int uvAuthMsg(SConn* pConn, char* msg, int len) {
+static int uvAuthMsg(SSrvConn* pConn, char* msg, int len) {
   STransMsgHead* pHead = (STransMsgHead*)msg;
 
   int code = 0;
@@ -222,14 +208,14 @@ static int uvAuthMsg(SConn* pConn, char* msg, int len) {
 
 // refers specifically to query or insert timeout
 static void uvHandleActivityTimeout(uv_timer_t* handle) {
-  SConn* conn = handle->data;
+  SSrvConn* conn = handle->data;
   tDebug("%p timeout since no activity", conn);
 }
 
-static void uvHandleReq(SConn* pConn) {
+static void uvHandleReq(SSrvConn* pConn) {
   SRecvInfo    info;
   SRecvInfo*   p = &info;
-  SConnBuffer* pBuf = &pConn->connBuf;
+  SConnBuffer* pBuf = &pConn->readBuf;
   p->msg = pBuf->buf;
   p->msgLen = pBuf->len;
   p->ip = 0;
@@ -255,7 +241,6 @@ static void uvHandleReq(SConn* pConn) {
   pHead->code = htonl(pHead->code);
 
   int32_t dlen = 0;
-  SRpcMsg rpcMsg;
   if (transDecompressMsg(NULL, 0, NULL)) {
     // add compress later
     // pHead = rpcDecompressRpcMsg(pHead);
@@ -264,6 +249,8 @@ static void uvHandleReq(SConn* pConn) {
     // impl later
     //
   }
+
+  SRpcMsg rpcMsg;
   rpcMsg.contLen = transContLenFromMsg(pHead->msgLen);
   rpcMsg.pCont = pHead->content;
   rpcMsg.msgType = pHead->msgType;
@@ -271,6 +258,7 @@ static void uvHandleReq(SConn* pConn) {
   rpcMsg.ahandle = NULL;
   rpcMsg.handle = pConn;
 
+  transClearBuffer(&pConn->readBuf);
   pConn->ref++;
   (*(pRpc->cfp))(pRpc->parent, &rpcMsg, NULL);
   // uv_timer_start(pConn->pTimer, uvHandleActivityTimeout, pRpc->idleTime * 10000, 0);
@@ -280,8 +268,8 @@ static void uvHandleReq(SConn* pConn) {
 
 void uvOnReadCb(uv_stream_t* cli, ssize_t nread, const uv_buf_t* buf) {
   // opt
-  SConn*       conn = cli->data;
-  SConnBuffer* pBuf = &conn->connBuf;
+  SSrvConn*    conn = cli->data;
+  SConnBuffer* pBuf = &conn->readBuf;
   if (nread > 0) {
     pBuf->len += nread;
     tDebug("conn %p read summroy, total read: %d, current read: %d", conn, pBuf->len, (int)nread);
@@ -294,11 +282,12 @@ void uvOnReadCb(uv_stream_t* cli, ssize_t nread, const uv_buf_t* buf) {
     return;
   }
   if (nread == 0) {
-    tDebug("conn %p except read", conn);
-    // destroyConn(conn, true);
     return;
   }
-  if (nread != UV_EOF) {
+  if (nread < 0 || nread != UV_EOF) {
+    if (conn->ref > 1) {
+      conn->ref++;  // ref > 1 signed that write is in progress
+    }
     tDebug("conn %p read error: %s", conn, uv_err_name(nread));
     destroyConn(conn, true);
   }
@@ -310,21 +299,21 @@ void uvAllocConnBufferCb(uv_handle_t* handle, size_t suggested_size, uv_buf_t* b
 
 void uvOnTimeoutCb(uv_timer_t* handle) {
   // opt
-  SConn* pConn = handle->data;
+  SSrvConn* pConn = handle->data;
   tDebug("conn %p time out", pConn);
 }
 
 void uvOnWriteCb(uv_write_t* req, int status) {
-  SConn* conn = req->data;
+  SSrvConn* conn = req->data;
+  SSrvMsg*  smsg = conn->pSrvMsg;
+  destroySrvMsg(conn);
 
-  SConnBuffer* buf = &conn->connBuf;
-  buf->len = 0;
-  memset(buf->buf, 0, buf->cap);
-  buf->left = -1;
+  transClearBuffer(&conn->readBuf);
   if (status == 0) {
     tDebug("conn %p data already was written on stream", conn);
   } else {
     tDebug("conn %p failed to write data, %s", conn, uv_err_name(status));
+    //
     destroyConn(conn, true);
   }
   // opt
@@ -337,16 +326,16 @@ static void uvOnPipeWriteCb(uv_write_t* req, int status) {
   }
 }
 
-static void uvPrepareSendData(SConn* conn, uv_buf_t* wb) {
+static void uvPrepareSendData(SSrvMsg* smsg, uv_buf_t* wb) {
   // impl later;
-  tDebug("conn %p prepare to send resp", conn);
-  SRpcMsg* pMsg = &conn->sendMsg;
+  tDebug("conn %p prepare to send resp", smsg->pConn);
+  SRpcMsg* pMsg = &smsg->msg;
   if (pMsg->pCont == 0) {
     pMsg->pCont = (void*)rpcMallocCont(0);
     pMsg->contLen = 0;
   }
   STransMsgHead* pHead = transHeadFromCont(pMsg->pCont);
-  pHead->msgType = conn->inType + 1;
+  pHead->msgType = smsg->pConn->inType + 1;
   // add more info
   char*   msg = (char*)pHead;
   int32_t len = transMsgLenFromCont(pMsg->contLen);
@@ -357,28 +346,55 @@ static void uvPrepareSendData(SConn* conn, uv_buf_t* wb) {
   wb->base = msg;
   wb->len = len;
 }
+static void uvStartSendResp(SSrvMsg* smsg) {
+  // impl
+  uv_buf_t wb;
+  uvPrepareSendData(smsg, &wb);
+
+  SSrvConn* pConn = smsg->pConn;
+  uv_timer_stop(pConn->pTimer);
+
+  pConn->pSrvMsg = smsg;
+  // conn->pWriter->data = smsg;
+  uv_write(pConn->pWriter, (uv_stream_t*)pConn->pTcp, &wb, 1, uvOnWriteCb);
+
+  // SRpcMsg* rpcMsg = smsg->msg;
+
+  return;
+}
+static void destroySrvMsg(SSrvConn* conn) {
+  SSrvMsg* smsg = conn->pSrvMsg;
+  if (smsg == NULL) {
+    return;
+  }
+  transFreeMsg(smsg->msg.pCont);
+  free(conn->pSrvMsg);
+  conn->pSrvMsg = NULL;
+}
 void uvWorkerAsyncCb(uv_async_t* handle) {
   SWorkThrdObj* pThrd = handle->data;
-  SConn*        conn = NULL;
+  SSrvConn*     conn = NULL;
   queue         wq;
   // batch process to avoid to lock/unlock frequently
-  pthread_mutex_lock(&pThrd->connMtx);
-  QUEUE_MOVE(&pThrd->conn, &wq);
-  pthread_mutex_unlock(&pThrd->connMtx);
+  pthread_mutex_lock(&pThrd->msgMtx);
+  QUEUE_MOVE(&pThrd->msg, &wq);
+  pthread_mutex_unlock(&pThrd->msgMtx);
 
   while (!QUEUE_IS_EMPTY(&wq)) {
     queue* head = QUEUE_HEAD(&wq);
     QUEUE_REMOVE(head);
-    SConn* conn = QUEUE_DATA(head, SConn, queue);
-    if (conn == NULL) {
-      tError("except occurred, do nothing");
-      return;
+
+    SSrvMsg* msg = QUEUE_DATA(head, SSrvMsg, q);
+    if (msg == NULL) {
+      tError("except occurred, continue");
+      continue;
     }
-    uv_buf_t wb;
-    uvPrepareSendData(conn, &wb);
-    uv_timer_stop(conn->pTimer);
+    uvStartSendResp(msg);
+    // uv_buf_t wb;
+    // uvPrepareSendData(msg, &wb);
+    // uv_timer_stop(conn->pTimer);
 
-    uv_write(conn->pWriter, (uv_stream_t*)conn->pTcp, &wb, 1, uvOnWriteCb);
+    // uv_write(conn->pWriter, (uv_stream_t*)conn->pTcp, &wb, 1, uvOnWriteCb);
   }
 }
 
@@ -431,7 +447,7 @@ void uvOnConnectionCb(uv_stream_t* q, ssize_t nread, const uv_buf_t* buf) {
   uv_handle_type pending = uv_pipe_pending_type(pipe);
   assert(pending == UV_TCP);
 
-  SConn* pConn = createConn();
+  SSrvConn* pConn = createConn();
 
   pConn->pTransInst = pThrd->pTransInst;
   /* init conn timer*/
@@ -480,8 +496,8 @@ static bool addHandleToWorkloop(void* arg) {
 
   pThrd->pipe->data = pThrd;
 
-  QUEUE_INIT(&pThrd->conn);
-  pthread_mutex_init(&pThrd->connMtx, NULL);
+  QUEUE_INIT(&pThrd->msg);
+  pthread_mutex_init(&pThrd->msgMtx, NULL);
 
   pThrd->workerAsync = malloc(sizeof(uv_async_t));
   uv_async_init(pThrd->loop, pThrd->workerAsync, uvWorkerAsyncCb);
@@ -519,34 +535,42 @@ void* workerThread(void* arg) {
   uv_run(pThrd->loop, UV_RUN_DEFAULT);
 }
 
-static SConn* createConn() {
-  SConn* pConn = (SConn*)calloc(1, sizeof(SConn));
+static SSrvConn* createConn() {
+  SSrvConn* pConn = (SSrvConn*)calloc(1, sizeof(SSrvConn));
+  tDebug("conn %p created", pConn);
   ++pConn->ref;
   return pConn;
 }
 
-static void destroyConn(SConn* conn, bool clear) {
+static void destroyConn(SSrvConn* conn, bool clear) {
   if (conn == NULL) {
     return;
   }
-  if (--conn->ref == 0) {
+  // SRpcMsg* pMsg = &conn->sendMsg;
+  // transFreeMsg(pMsg->pCont);
+  // pMsg->pCont = NULL;
+
+  tDebug("conn %p try to destroy", conn);
+  if (--conn->ref > 0) {
     return;
   }
+  transDestroyBuffer(&conn->readBuf);
+  destroySrvMsg(conn);
+
   if (clear) {
-    uv_close((uv_handle_t*)conn->pTcp, NULL);
+    uv_close((uv_handle_t*)conn->pTcp, uvDestroyConn);
   }
+}
+static void uvDestroyConn(uv_handle_t* handle) {
+  SSrvConn* conn = handle->data;
+  tDebug("conn %p destroy", conn);
   uv_timer_stop(conn->pTimer);
   free(conn->pTimer);
-  free(conn->pTcp);
-  free(conn->connBuf.buf);
+  // free(conn->pTcp);
   free(conn->pWriter);
   free(conn);
 }
-static void uvDestroyConn(uv_handle_t* handle) {
-  SConn* conn = handle->data;
-  destroyConn(conn, false);
-}
-static int transAddAuthPart(SConn* pConn, char* msg, int msgLen) {
+static int transAddAuthPart(SSrvConn* pConn, char* msg, int msgLen) {
   STransMsgHead* pHead = (STransMsgHead*)msg;
 
   if (pConn->spi && pConn->secured == 0) {
@@ -628,6 +652,7 @@ void destroyWorkThrd(SWorkThrdObj* pThrd) {
   pthread_join(pThrd->thread, NULL);
   // free(srv->pipe[i]);
   free(pThrd->loop);
+  pthread_mutex_destroy(&pThrd->msgMtx);
   free(pThrd);
 }
 void taosCloseServer(void* arg) {
@@ -644,17 +669,20 @@ void taosCloseServer(void* arg) {
 }
 
 void rpcSendResponse(const SRpcMsg* pMsg) {
-  SConn*        pConn = pMsg->handle;
+  SSrvConn*     pConn = pMsg->handle;
   SWorkThrdObj* pThrd = pConn->hostThrd;
 
-  // opt later
-  pConn->sendMsg = *pMsg;
-  pthread_mutex_lock(&pThrd->connMtx);
-  QUEUE_PUSH(&pThrd->conn, &pConn->queue);
-  pthread_mutex_unlock(&pThrd->connMtx);
+  SSrvMsg* srvMsg = calloc(1, sizeof(SSrvMsg));
+  srvMsg->pConn = pConn;
+  srvMsg->msg = *pMsg;
+
+  pthread_mutex_lock(&pThrd->msgMtx);
+  QUEUE_PUSH(&pThrd->msg, &srvMsg->q);
+  pthread_mutex_unlock(&pThrd->msgMtx);
+
   tDebug("conn %p start to send resp", pConn);
 
-  uv_async_send(pConn->pWorkerAsync);
+  uv_async_send(pThrd->workerAsync);
 }
 
 #endif

source/libs/transport/test/rclient.c

@@ -64,6 +64,7 @@ static void *sendRequest(void *param) {
     // tsem_wait(&pInfo->rspSem);
     tsem_wait(&pInfo->rspSem);
     tDebug("recv response succefully");
+
     // usleep(100000000);
   }
 

source/util/src/tqueue.c

@@ -13,35 +13,36 @@
  * along with this program. If not, see <http://www.gnu.org/licenses/>.
  */
 
-#include "os.h"
-
-#include "taoserror.h"
+#define _DEFAULT_SOURCE
 #include "tqueue.h"
+#include "taoserror.h"
 #include "ulog.h"
 
 typedef struct STaosQnode STaosQnode;
 
 typedef struct STaosQnode {
   STaosQnode *next;
+  STaosQueue *queue;
   char        item[];
 } STaosQnode;
 
 typedef struct STaosQueue {
   int32_t         itemSize;
   int32_t         numOfItems;
-  STaosQnode     *head;
-  STaosQnode     *tail;
-  STaosQueue     *next;     // for queue set
-  STaosQset      *qset;     // for queue set
-  void           *ahandle;  // for queue set
-  FProcessItem    itemFp;
-  FProcessItems   itemsFp;
+  int32_t         threadId;
+  STaosQnode *    head;
+  STaosQnode *    tail;
+  STaosQueue *    next;     // for queue set
+  STaosQset *     qset;     // for queue set
+  void *          ahandle;  // for queue set
+  FItem           itemFp;
+  FItems          itemsFp;
   pthread_mutex_t mutex;
 } STaosQueue;
 
 typedef struct STaosQset {
-  STaosQueue     *head;
-  STaosQueue     *current;
+  STaosQueue *    head;
+  STaosQueue *    current;
   pthread_mutex_t mutex;
   int32_t         numOfQueues;
   int32_t         numOfItems;
@@ -56,19 +57,23 @@ typedef struct STaosQall {
 } STaosQall;
 
 STaosQueue *taosOpenQueue() {
-  STaosQueue *queue = calloc(sizeof(STaosQueue), 1);
+  STaosQueue *queue = calloc(1, sizeof(STaosQueue));
   if (queue == NULL) {
     terrno = TSDB_CODE_OUT_OF_MEMORY;
     return NULL;
   }
 
-  pthread_mutex_init(&queue->mutex, NULL);
+  if (pthread_mutex_init(&queue->mutex, NULL) != 0) {
+    terrno = TSDB_CODE_OUT_OF_MEMORY;
+    return NULL;
+  }
 
-  uTrace("queue:%p is opened", queue);
+  queue->threadId = -1;
+  uDebug("queue:%p is opened", queue);
   return queue;
 }
 
-void taosSetQueueFp(STaosQueue *queue, FProcessItem itemFp, FProcessItems itemsFp) {
+void taosSetQueueFp(STaosQueue *queue, FItem itemFp, FItems itemsFp) {
   if (queue == NULL) return;
   queue->itemFp = itemFp;
   queue->itemsFp = itemsFp;
@@ -77,7 +82,7 @@ void taosSetQueueFp(STaosQueue *queue, FProcessItem itemFp, FProcessItems itemsF
 void taosCloseQueue(STaosQueue *queue) {
   if (queue == NULL) return;
   STaosQnode *pTemp;
-  STaosQset  *qset;
+  STaosQset * qset;
 
   pthread_mutex_lock(&queue->mutex);
   STaosQnode *pNode = queue->head;
@@ -85,7 +90,9 @@ void taosCloseQueue(STaosQueue *queue) {
   qset = queue->qset;
   pthread_mutex_unlock(&queue->mutex);
 
-  if (queue->qset) taosRemoveFromQset(qset, queue);
+  if (queue->qset) {
+    taosRemoveFromQset(qset, queue);
+  }
 
   while (pNode) {
     pTemp = pNode;
@@ -96,7 +103,7 @@ void taosCloseQueue(STaosQueue *queue) {
   pthread_mutex_destroy(&queue->mutex);
   free(queue);
 
-  uTrace("queue:%p is closed", queue);
+  uDebug("queue:%p is closed", queue);
 }
 
 bool taosQueueEmpty(STaosQueue *queue) {
@@ -120,19 +127,23 @@ int32_t taosQueueSize(STaosQueue *queue) {
 }
 
 void *taosAllocateQitem(int32_t size) {
-  STaosQnode *pNode = (STaosQnode *)calloc(sizeof(STaosQnode) + size, 1);
+  STaosQnode *pNode = calloc(1, sizeof(STaosQnode) + size);
+
+  if (pNode == NULL) {
+    terrno = TSDB_CODE_OUT_OF_MEMORY;
+    return NULL;
+  }
 
-  if (pNode == NULL) return NULL;
   uTrace("item:%p, node:%p is allocated", pNode->item, pNode);
   return (void *)pNode->item;
 }
 
-void taosFreeQitem(void *param) {
-  if (param == NULL) return;
+void taosFreeQitem(void *pItem) {
+  if (pItem == NULL) return;
 
-  char *temp = (char *)param;
+  char *temp = pItem;
   temp -= sizeof(STaosQnode);
-  uTrace("item:%p, node:%p is freed", param, temp);
+  uTrace("item:%p, node:%p is freed", pItem, temp);
   free(temp);
 }
 
@@ -175,7 +186,7 @@ int32_t taosReadQitem(STaosQueue *queue, void **ppItem) {
     queue->numOfItems--;
     if (queue->qset) atomic_sub_fetch_32(&queue->qset->numOfItems, 1);
     code = 1;
-    uDebug("item:%p is read out from queue:%p, items:%d", *ppItem, queue, queue->numOfItems);
+    uTrace("item:%p is read out from queue:%p, items:%d", *ppItem, queue, queue->numOfItems);
   }
 
   pthread_mutex_unlock(&queue->mutex);
@@ -183,7 +194,7 @@ int32_t taosReadQitem(STaosQueue *queue, void **ppItem) {
   return code;
 }
 
-STaosQall *taosAllocateQall() { return calloc(sizeof(STaosQall), 1); }
+STaosQall *taosAllocateQall() { return calloc(1, sizeof(STaosQall)); }
 
 void taosFreeQall(STaosQall *qall) { free(qall); }
 
@@ -238,7 +249,7 @@ int32_t taosGetQitem(STaosQall *qall, void **ppItem) {
 void taosResetQitems(STaosQall *qall) { qall->current = qall->start; }
 
 STaosQset *taosOpenQset() {
-  STaosQset *qset = (STaosQset *)calloc(sizeof(STaosQset), 1);
+  STaosQset *qset = calloc(sizeof(STaosQset), 1);
   if (qset == NULL) {
     terrno = TSDB_CODE_OUT_OF_MEMORY;
     return NULL;
@@ -247,7 +258,7 @@ STaosQset *taosOpenQset() {
   pthread_mutex_init(&qset->mutex, NULL);
   tsem_init(&qset->sem, 0, 0);
 
-  uTrace("qset:%p is opened", qset);
+  uDebug("qset:%p is opened", qset);
   return qset;
 }
 
@@ -268,7 +279,7 @@ void taosCloseQset(STaosQset *qset) {
   pthread_mutex_destroy(&qset->mutex);
   tsem_destroy(&qset->sem);
   free(qset);
-  uTrace("qset:%p is closed", qset);
+  uDebug("qset:%p is closed", qset);
 }
 
 // tsem_post 'qset->sem', so that reader threads waiting for it
@@ -338,12 +349,12 @@ void taosRemoveFromQset(STaosQset *qset, STaosQueue *queue) {
 
   pthread_mutex_unlock(&qset->mutex);
 
-  uTrace("queue:%p is removed from qset:%p", queue, qset);
+  uDebug("queue:%p is removed from qset:%p", queue, qset);
 }
 
 int32_t taosGetQueueNumber(STaosQset *qset) { return qset->numOfQueues; }
 
-int32_t taosReadQitemFromQset(STaosQset *qset, void **ppItem, void **ahandle, FProcessItem *itemFp) {
+int32_t taosReadQitemFromQset(STaosQset *qset, void **ppItem, void **ahandle, FItem *itemFp) {
   STaosQnode *pNode = NULL;
   int32_t     code = 0;
 
@@ -365,6 +376,7 @@ int32_t taosReadQitemFromQset(STaosQset *qset, void **ppItem, void **ahandle, FP
       *ppItem = pNode->item;
       if (ahandle) *ahandle = queue->ahandle;
       if (itemFp) *itemFp = queue->itemFp;
+
       queue->head = pNode->next;
       if (queue->head == NULL) queue->tail = NULL;
       queue->numOfItems--;
@@ -382,7 +394,7 @@ int32_t taosReadQitemFromQset(STaosQset *qset, void **ppItem, void **ahandle, FP
   return code;
 }
 
-int32_t taosReadAllQitemsFromQset(STaosQset *qset, STaosQall *qall, void **ahandle, FProcessItems *itemsFp) {
+int32_t taosReadAllQitemsFromQset(STaosQset *qset, STaosQall *qall, void **ahandle, FItems *itemsFp) {
   STaosQueue *queue;
   int32_t     code = 0;
 
@@ -411,7 +423,9 @@ int32_t taosReadAllQitemsFromQset(STaosQset *qset, STaosQall *qall, void **ahand
       queue->tail = NULL;
       queue->numOfItems = 0;
       atomic_sub_fetch_32(&qset->numOfItems, qall->numOfItems);
-      for (int32_t j = 1; j < qall->numOfItems; ++j) tsem_wait(&qset->sem);
+      for (int32_t j = 1; j < qall->numOfItems; ++j) {
+        tsem_wait(&qset->sem);
+      }
     }
 
     pthread_mutex_unlock(&queue->mutex);
@@ -423,6 +437,65 @@ int32_t taosReadAllQitemsFromQset(STaosQset *qset, STaosQall *qall, void **ahand
   return code;
 }
 
+int32_t taosReadQitemFromQsetByThread(STaosQset *qset, void **ppItem, void **ahandle, FItem *itemFp, int32_t threadId) {
+  STaosQnode *pNode = NULL;
+  int32_t     code = -1;
+
+  tsem_wait(&qset->sem);
+
+  pthread_mutex_lock(&qset->mutex);
+
+  for (int32_t i = 0; i < qset->numOfQueues; ++i) {
+    if (qset->current == NULL) qset->current = qset->head;
+    STaosQueue *queue = qset->current;
+    if (queue) qset->current = queue->next;
+    if (queue == NULL) break;
+    if (queue->head == NULL) continue;
+    if (queue->threadId != -1 && queue->threadId != threadId) {
+      code = 0;
+      continue;
+    }
+
+    pthread_mutex_lock(&queue->mutex);
+
+    if (queue->head) {
+      pNode = queue->head;
+      pNode->queue = queue;
+      queue->threadId = threadId;
+      *ppItem = pNode->item;
+
+      if (ahandle) *ahandle = queue->ahandle;
+      if (itemFp) *itemFp = queue->itemFp;
+
+      queue->head = pNode->next;
+      if (queue->head == NULL) queue->tail = NULL;
+      queue->numOfItems--;
+      atomic_sub_fetch_32(&qset->numOfItems, 1);
+      code = 1;
+      uTrace("item:%p is read out from queue:%p, items:%d", *ppItem, queue, queue->numOfItems);
+    }
+
+    pthread_mutex_unlock(&queue->mutex);
+    if (pNode) break;
+  }
+
+  pthread_mutex_unlock(&qset->mutex);
+
+  return code;
+}
+
+void taosResetQsetThread(STaosQset *qset, void *pItem) {
+  if (pItem == NULL) return;
+  STaosQnode *pNode = (STaosQnode *)((char *)pItem - sizeof(STaosQnode));
+
+  pthread_mutex_lock(&qset->mutex);
+  pNode->queue->threadId = -1;
+  for (int32_t i = 0; i < pNode->queue->numOfItems; ++i) {
+    tsem_post(&qset->sem);
+  }
+  pthread_mutex_unlock(&qset->mutex);
+}
+
 int32_t taosGetQueueItemsNumber(STaosQueue *queue) {
   if (!queue) return 0;
 

source/util/src/tworker.c

@@ -14,38 +14,46 @@
  */
 
 #define _DEFAULT_SOURCE
-#include "os.h"
-#include "ulog.h"
-#include "tqueue.h"
 #include "tworker.h"
+#include "taoserror.h"
+#include "ulog.h"
 
-typedef void* (*ThreadFp)(void *param);
+typedef void *(*ThreadFp)(void *param);
 
-int32_t tWorkerInit(SWorkerPool *pool) {
+int32_t tQWorkerInit(SQWorkerPool *pool) {
   pool->qset = taosOpenQset();
-  pool->workers = calloc(sizeof(SWorker), pool->max);
-  pthread_mutex_init(&pool->mutex, NULL);
-  for (int i = 0; i < pool->max; ++i) {
-    SWorker *worker = pool->workers + i;
+  pool->workers = calloc(sizeof(SQWorker), pool->max);
+  if (pool->workers == NULL) {
+    terrno = TSDB_CODE_OUT_OF_MEMORY;
+    return -1;
+  }
+
+  if (pthread_mutex_init(&pool->mutex, NULL)) {
+    terrno = TSDB_CODE_OUT_OF_MEMORY;
+    return -1;
+  }
+
+  for (int32_t i = 0; i < pool->max; ++i) {
+    SQWorker *worker = pool->workers + i;
     worker->id = i;
     worker->pool = pool;
   }
 
-  uInfo("worker:%s is initialized, min:%d max:%d", pool->name, pool->min, pool->max);
+  uDebug("worker:%s is initialized, min:%d max:%d", pool->name, pool->min, pool->max);
   return 0;
 }
 
-void tWorkerCleanup(SWorkerPool *pool) {
-  for (int i = 0; i < pool->max; ++i) {
-    SWorker *worker = pool->workers + i;
+void tQWorkerCleanup(SQWorkerPool *pool) {
+  for (int32_t i = 0; i < pool->max; ++i) {
+    SQWorker *worker = pool->workers + i;
     if (worker == NULL) continue;
     if (taosCheckPthreadValid(worker->thread)) {
       taosQsetThreadResume(pool->qset);
     }
   }
 
-  for (int i = 0; i < pool->max; ++i) {
-    SWorker *worker = pool->workers + i;
+  for (int32_t i = 0; i < pool->max; ++i) {
+    SQWorker *worker = pool->workers + i;
     if (worker == NULL) continue;
     if (taosCheckPthreadValid(worker->thread)) {
       pthread_join(worker->thread, NULL);
@@ -56,15 +64,15 @@ void tWorkerCleanup(SWorkerPool *pool) {
   taosCloseQset(pool->qset);
   pthread_mutex_destroy(&pool->mutex);
 
-  uInfo("worker:%s is closed", pool->name);
+  uDebug("worker:%s is closed", pool->name);
 }
 
-static void *tWorkerThreadFp(SWorker *worker) {
-  SWorkerPool *pool = worker->pool;
-  FProcessItem fp = NULL;
+static void *tQWorkerThreadFp(SQWorker *worker) {
+  SQWorkerPool *pool = worker->pool;
+  FItem         fp = NULL;
 
-  void   *msg = NULL;
-  void   *ahandle = NULL;
+  void *  msg = NULL;
+  void *  ahandle = NULL;
   int32_t code = 0;
 
   taosBlockSIGPIPE();
@@ -77,7 +85,7 @@ static void *tWorkerThreadFp(SWorker *worker) {
       break;
     }
 
-    if (fp) {
+    if (fp != NULL) {
       (*fp)(ahandle, msg);
     }
   }
@@ -85,11 +93,12 @@ static void *tWorkerThreadFp(SWorker *worker) {
   return NULL;
 }
 
-STaosQueue *tWorkerAllocQueue(SWorkerPool *pool, void *ahandle, FProcessItem fp) {
+STaosQueue *tWorkerAllocQueue(SQWorkerPool *pool, void *ahandle, FItem fp, ThreadFp threadFp) {
   pthread_mutex_lock(&pool->mutex);
   STaosQueue *queue = taosOpenQueue();
   if (queue == NULL) {
     pthread_mutex_unlock(&pool->mutex);
+    terrno = TSDB_CODE_OUT_OF_MEMORY;
     return NULL;
   }
 
@@ -99,14 +108,18 @@ STaosQueue *tWorkerAllocQueue(SWorkerPool *pool, void *ahandle, FProcessItem fp)
   // spawn a thread to process queue
   if (pool->num < pool->max) {
     do {
-      SWorker *worker = pool->workers + pool->num;
+      SQWorker *worker = pool->workers + pool->num;
 
       pthread_attr_t thAttr;
       pthread_attr_init(&thAttr);
       pthread_attr_setdetachstate(&thAttr, PTHREAD_CREATE_JOINABLE);
 
-      if (pthread_create(&worker->thread, &thAttr, (ThreadFp)tWorkerThreadFp, worker) != 0) {
+      if (pthread_create(&worker->thread, &thAttr, threadFp, worker) != 0) {
         uError("worker:%s:%d failed to create thread to process since %s", pool->name, worker->id, strerror(errno));
+        taosCloseQueue(queue);
+        terrno = TSDB_CODE_OUT_OF_MEMORY;
+        queue = NULL;
+        break;
       }
 
       pthread_attr_destroy(&thAttr);
@@ -121,19 +134,73 @@ STaosQueue *tWorkerAllocQueue(SWorkerPool *pool, void *ahandle, FProcessItem fp)
   return queue;
 }
 
-void tWorkerFreeQueue(SWorkerPool *pool, STaosQueue *queue) {
+STaosQueue *tQWorkerAllocQueue(SQWorkerPool *pool, void *ahandle, FItem fp) {
+  return tWorkerAllocQueue(pool, ahandle, fp, (ThreadFp)tQWorkerThreadFp);
+}
+
+void tQWorkerFreeQueue(SQWorkerPool *pool, STaosQueue *queue) {
   taosCloseQueue(queue);
   uDebug("worker:%s, queue:%p is freed", pool->name, queue);
 }
 
-int32_t tMWorkerInit(SMWorkerPool *pool) {
+int32_t tFWorkerInit(SFWorkerPool *pool) { return tQWorkerInit((SQWorkerPool *)pool); }
+
+void tFWorkerCleanup(SFWorkerPool *pool) { tQWorkerCleanup(pool); }
+
+static void *tFWorkerThreadFp(SQWorker *worker) {
+  SQWorkerPool *pool = worker->pool;
+
+  FItem   fp = NULL;
+  void *  msg = NULL;
+  void *  ahandle = NULL;
+  int32_t code = 0;
+
+  taosBlockSIGPIPE();
+  setThreadName(pool->name);
+  uDebug("worker:%s:%d is running", pool->name, worker->id);
+
+  while (1) {
+    code = taosReadQitemFromQsetByThread(pool->qset, (void **)&msg, &ahandle, &fp, worker->id);
+
+    if (code < 0) {
+      uDebug("worker:%s:%d qset:%p, got no message and exiting", pool->name, worker->id, pool->qset);
+      break;
+    } else if (code == 0) {
+      // uTrace("worker:%s:%d qset:%p, got no message and continue", pool->name, worker->id, pool->qset);
+      continue;
+    }
+
+    if (fp != NULL) {
+      (*fp)(ahandle, msg);
+    }
+
+    taosResetQsetThread(pool->qset, msg);
+  }
+
+  return NULL;
+}
+
+STaosQueue *tFWorkerAllocQueue(SQWorkerPool *pool, void *ahandle, FItem fp) {
+  return tWorkerAllocQueue(pool, ahandle, fp, (ThreadFp)tFWorkerThreadFp);
+}
+
+void tFWorkerFreeQueue(SFWorkerPool *pool, STaosQueue *queue) { tQWorkerFreeQueue(pool, queue); }
+
+int32_t tWWorkerInit(SWWorkerPool *pool) {
   pool->nextId = 0;
-  pool->workers = calloc(sizeof(SMWorker), pool->max);
-  if (pool->workers == NULL) return -1;
+  pool->workers = calloc(sizeof(SWWorker), pool->max);
+  if (pool->workers == NULL) {
+    terrno = TSDB_CODE_OUT_OF_MEMORY;
+    return -1;
+  }
+
+  if (pthread_mutex_init(&pool->mutex, NULL) != 0) {
+    terrno = TSDB_CODE_OUT_OF_MEMORY;
+    return -1;
+  }
 
-  pthread_mutex_init(&pool->mutex, NULL);
   for (int32_t i = 0; i < pool->max; ++i) {
-    SMWorker *worker = pool->workers + i;
+    SWWorker *worker = pool->workers + i;
     worker->id = i;
     worker->qall = NULL;
     worker->qset = NULL;
@@ -144,16 +211,18 @@ int32_t tMWorkerInit(SMWorkerPool *pool) {
   return 0;
 }
 
-void tMWorkerCleanup(SMWorkerPool *pool) {
+void tWWorkerCleanup(SWWorkerPool *pool) {
   for (int32_t i = 0; i < pool->max; ++i) {
-    SMWorker *worker = pool->workers + i;
+    SWWorker *worker = pool->workers + i;
     if (taosCheckPthreadValid(worker->thread)) {
-      if (worker->qset) taosQsetThreadResume(worker->qset);
+      if (worker->qset) {
+        taosQsetThreadResume(worker->qset);
+      }
     }
   }
 
   for (int32_t i = 0; i < pool->max; ++i) {
-    SMWorker *worker = pool->workers + i;
+    SWWorker *worker = pool->workers + i;
     if (taosCheckPthreadValid(worker->thread)) {
       pthread_join(worker->thread, NULL);
       taosFreeQall(worker->qall);
@@ -167,12 +236,12 @@ void tMWorkerCleanup(SMWorkerPool *pool) {
   uInfo("worker:%s is closed", pool->name);
 }
 
-static void *tWriteWorkerThreadFp(SMWorker *worker) {
-  SMWorkerPool *pool = worker->pool;
-  FProcessItems fp = NULL;
+static void *tWWorkerThreadFp(SWWorker *worker) {
+  SWWorkerPool *pool = worker->pool;
+  FItems        fp = NULL;
 
-  void   *msg = NULL;
-  void   *ahandle = NULL;
+  void *  msg = NULL;
+  void *  ahandle = NULL;
   int32_t numOfMsgs = 0;
   int32_t qtype = 0;
 
@@ -187,7 +256,7 @@ static void *tWriteWorkerThreadFp(SMWorker *worker) {
       break;
     }
 
-    if (fp) {
+    if (fp != NULL) {
       (*fp)(ahandle, worker->qall, numOfMsgs);
     }
   }
@@ -195,13 +264,14 @@ static void *tWriteWorkerThreadFp(SMWorker *worker) {
   return NULL;
 }
 
-STaosQueue *tMWorkerAllocQueue(SMWorkerPool *pool, void *ahandle, FProcessItems fp) {
+STaosQueue *tWWorkerAllocQueue(SWWorkerPool *pool, void *ahandle, FItems fp) {
   pthread_mutex_lock(&pool->mutex);
-  SMWorker *worker = pool->workers + pool->nextId;
+  SWWorker *worker = pool->workers + pool->nextId;
 
   STaosQueue *queue = taosOpenQueue();
   if (queue == NULL) {
     pthread_mutex_unlock(&pool->mutex);
+    terrno = TSDB_CODE_OUT_OF_MEMORY;
     return NULL;
   }
 
@@ -221,17 +291,19 @@ STaosQueue *tMWorkerAllocQueue(SMWorkerPool *pool, void *ahandle, FProcessItems
       taosCloseQset(worker->qset);
       taosCloseQueue(queue);
       pthread_mutex_unlock(&pool->mutex);
+      terrno = TSDB_CODE_OUT_OF_MEMORY;
       return NULL;
     }
     pthread_attr_t thAttr;
     pthread_attr_init(&thAttr);
     pthread_attr_setdetachstate(&thAttr, PTHREAD_CREATE_JOINABLE);
 
-    if (pthread_create(&worker->thread, &thAttr, (ThreadFp)tWriteWorkerThreadFp, worker) != 0) {
+    if (pthread_create(&worker->thread, &thAttr, (ThreadFp)tWWorkerThreadFp, worker) != 0) {
       uError("worker:%s:%d failed to create thread to process since %s", pool->name, worker->id, strerror(errno));
       taosFreeQall(worker->qall);
       taosCloseQset(worker->qset);
       taosCloseQueue(queue);
+      terrno = TSDB_CODE_OUT_OF_MEMORY;
       queue = NULL;
     } else {
       uDebug("worker:%s:%d is launched, max:%d", pool->name, worker->id, pool->max);
@@ -250,7 +322,7 @@ STaosQueue *tMWorkerAllocQueue(SMWorkerPool *pool, void *ahandle, FProcessItems
   return queue;
 }
 
-void tMWorkerFreeQueue(SMWorkerPool *pool, STaosQueue *queue) {
+void tWWorkerFreeQueue(SWWorkerPool *pool, STaosQueue *queue) {
   taosCloseQueue(queue);
   uDebug("worker:%s, queue:%p is freed", pool->name, queue);
 }

tests/script/tsim/dnode/basic1.sim

@@ -178,9 +178,9 @@ if $rows != 3 then
 endi
 
 sql select * from st
-#if $rows != 15 then 
-#  return -1
-#endi
+if $rows != 15 then 
+  return -1
+endi
 
 print =============== drop dnode
 sql drop dnode 2;

tests/script/tsim/table/basic1.sim

@@ -139,9 +139,9 @@ endi
 
 print =============== query data frpm st
 sql select * from st
-#if $rows != 21 then 
-#  return -1
-#endi
+if $rows != 21 then 
+  return -1
+endi
 
 system sh/exec.sh -n dnode1 -s stop -x SIGINT
 system sh/exec.sh -n dnode1 -s start
@@ -200,8 +200,8 @@ endi
 
 print =============== query data frpm st
 sql select * from st
-#if $rows != 21 then 
-#  return -1
-#endi
+if $rows != 21 then 
+  return -1
+endi
 
 system sh/exec.sh -n dnode1 -s stop -x SIGINT
\ No newline at end of file