Merge remote-tracking branch 'origin/feature/qnode' into feature/3.0_wxy

include/libs/qcom/query.h

@@ -91,9 +91,10 @@ enum {
   META_TYPE_NON_TABLE = 1,
   META_TYPE_CTABLE,
   META_TYPE_TABLE,
-  META_TYPE_BOTH_TABLE,
+  META_TYPE_BOTH_TABLE
 };
 
+
 typedef struct STableMetaOutput {
   int32_t     metaType;
   char        ctbFname[TSDB_TABLE_FNAME_LEN];

include/libs/scheduler/scheduler.h

@@ -24,7 +24,7 @@ extern "C" {
 #include "catalog.h"
 
 typedef struct SSchedulerCfg {
-  int32_t maxJobNum;
+  uint32_t maxJobNum;
 } SSchedulerCfg;
 
 typedef struct SQueryProfileSummary {

source/libs/catalog/src/catalog.c

@@ -1226,6 +1226,16 @@ int32_t catalogGetTableDistVgroup(struct SCatalog* pCatalog, void *pRpc, const S
   tNameGetFullDbName(pTableName, db);
   CTG_ERR_JRET(ctgGetDBVgroup(pCatalog, pRpc, pMgmtEps, db, false, &dbVgroup));
 
+  // REMOEV THIS ....
+  if (0 == tbMeta->vgId) {
+    SVgroupInfo vgroup = {0};
+    
+    catalogGetTableHashVgroup(pCatalog, pRpc, pMgmtEps, pTableName, &vgroup);
+
+    tbMeta->vgId = vgroup.vgId;
+  }
+  // REMOVE THIS ....
+
   if (tbMeta->tableType == TSDB_SUPER_TABLE) {
     CTG_ERR_JRET(ctgGetVgInfoFromDB(pCatalog, pRpc, pMgmtEps, dbVgroup, pVgroupList));
   } else {

source/libs/parser/src/astValidate.c

@@ -3628,6 +3628,33 @@ int32_t evaluateSqlNode(SSqlNode* pNode, int32_t tsPrecision, SMsgBuf* pMsgBuf)
   return TSDB_CODE_SUCCESS;
 }
 
+int32_t setTableVgroupList(SParseBasicCtx *pCtx, SName* name, SVgroupsInfo **pVgList) {
+  SArray* vgroupList = NULL;
+  int32_t code = catalogGetTableDistVgroup(pCtx->pCatalog, pCtx->pTransporter, &pCtx->mgmtEpSet, name, &vgroupList);
+  if (code != TSDB_CODE_SUCCESS) {
+    return code;
+  }
+  
+  int32_t vgroupNum = taosArrayGetSize(vgroupList);
+
+  SVgroupsInfo *vgList = calloc(1, sizeof(SVgroupsInfo) + sizeof(SVgroupMsg) * vgroupNum);
+  
+  vgList->numOfVgroups = vgroupNum;
+  
+  for (int32_t i = 0; i < vgroupNum; ++i) {
+    SVgroupInfo *vg = taosArrayGet(vgroupList, i);
+    vgList->vgroups[i].vgId = vg->vgId;
+    vgList->vgroups[i].numOfEps = vg->numOfEps;
+    memcpy(vgList->vgroups[i].epAddr, vg->epAddr, sizeof(vgList->vgroups[i].epAddr));
+  }
+
+  *pVgList = vgList;
+
+  taosArrayDestroy(vgroupList);
+
+  return TSDB_CODE_SUCCESS;
+}
+
 int32_t qParserValidateSqlNode(SParseBasicCtx *pCtx, SSqlInfo* pInfo, SQueryStmtInfo* pQueryInfo, char* msgBuf, int32_t msgBufLen) {
   assert(pCtx != NULL && pInfo != NULL);
   int32_t code = 0;
@@ -3916,7 +3943,7 @@ int32_t qParserValidateSqlNode(SParseBasicCtx *pCtx, SSqlInfo* pInfo, SQueryStmt
   if (code != TSDB_CODE_SUCCESS) {
     return code;
   }
-
+  
   data.pTableMeta = taosArrayInit(1, POINTER_BYTES);
   taosArrayPush(data.pTableMeta, &pmt);
 
@@ -3926,6 +3953,12 @@ int32_t qParserValidateSqlNode(SParseBasicCtx *pCtx, SSqlInfo* pInfo, SQueryStmt
   pQueryInfo->pTableMetaInfo[0]->name = *name;
   pQueryInfo->numOfTables = 1;
 
+  code = setTableVgroupList(pCtx, name, &pQueryInfo->pTableMetaInfo[0]->vgroupList);
+  if (code != TSDB_CODE_SUCCESS) {
+    taosArrayDestroy(data.pTableMeta);
+    return code;
+  }
+
   // evaluate the sqlnode
   STableMeta* pTableMeta = (STableMeta*) taosArrayGetP(data.pTableMeta, 0);
   assert(pTableMeta != NULL);

source/libs/planner/src/physicalPlan.c

@@ -160,9 +160,6 @@ static SPhyNode* createUserTableScanNode(SQueryPlanNode* pPlanNode, SQueryTableI
   return (SPhyNode*)node;
 }
 
-static SPhyNode* createSingleTableScanNode(SQueryPlanNode* pPlanNode, SQueryTableInfo* pTable) {
-  return createUserTableScanNode(pPlanNode, pTable, OP_TableScan);
-}
 
 static bool isSystemTable(SQueryTableInfo* pTable) {
   // todo
@@ -259,12 +256,20 @@ static bool needMultiNodeScan(SQueryTableInfo* pTable) {
   return (TSDB_SUPER_TABLE == pTable->pMeta->pTableMeta->tableType);
 }
 
+static SPhyNode* createSingleTableScanNode(SQueryPlanNode* pPlanNode, SQueryTableInfo* pTable, SSubplan* subplan) {
+  vgroupMsgToEpSet(&(pTable->pMeta->vgroupList->vgroups[0]), &subplan->execNode);
+
+  return createUserTableScanNode(pPlanNode, pTable, OP_TableScan);
+}
+
+
 static SPhyNode* createTableScanNode(SPlanContext* pCxt, SQueryPlanNode* pPlanNode) {
   SQueryTableInfo* pTable = (SQueryTableInfo*)pPlanNode->pExtInfo;
+    
   if (needMultiNodeScan(pTable)) {
     return createExchangeNode(pCxt, pPlanNode, splitSubplanByTable(pCxt, pPlanNode, pTable));
   }
-  return createSingleTableScanNode(pPlanNode, pTable);
+  return createSingleTableScanNode(pPlanNode, pTable, pCxt->pCurrentSubplan);
 }
 
 static SPhyNode* createPhyNode(SPlanContext* pCxt, SQueryPlanNode* pPlanNode) {
@@ -322,12 +327,12 @@ static void createSubplanByLevel(SPlanContext* pCxt, SQueryPlanNode* pRoot) {
   if (QNODE_MODIFY == pRoot->info.type) {
     splitModificationOpSubPlan(pCxt, pRoot);
   } else {
-    SSubplan* subplan  = initSubplan(pCxt, QUERY_TYPE_MERGE);
+    SSubplan* subplan  = initSubplan(pCxt, QUERY_TYPE_SCAN);
     ++(pCxt->nextId.templateId);
 
     subplan->msgType   = TDMT_VND_QUERY;
     subplan->pNode     = createPhyNode(pCxt, pRoot);
-    subplan->pDataSink = createDataDispatcher(pCxt, pRoot);
+    subplan->pDataSink = createDataDispatcher(pCxt, pRoot);    
   }
   // todo deal subquery
 }

source/libs/scheduler/inc/schedulerInt.h

@@ -37,10 +37,10 @@ enum {
 };
 
 typedef struct SSchedulerMgmt {
-  uint64_t  taskId; 
-  uint64_t  sId;
+  uint64_t      taskId; // sequential taksId
+  uint64_t      sId;    // schedulerId
   SSchedulerCfg cfg;
-  SHashObj *jobs;  // key: queryId, value: SQueryJob*
+  SHashObj     *jobs;   // key: queryId, value: SQueryJob*
 } SSchedulerMgmt;
 
 typedef struct SSchCallbackParam {
@@ -83,52 +83,61 @@ typedef struct SSchJobAttr {
 
 typedef struct SSchJob {
   uint64_t         queryId;
-  int32_t          levelNum;
-  int32_t          levelIdx;
-  int8_t           status;
   SSchJobAttr      attr;
-  SEpSet           dataSrcEps;
-  SEpAddr          resEp;
+  int32_t          levelNum;
   void            *transport;
   SArray          *nodeList;   // qnode/vnode list, element is SQueryNodeAddr
+  SArray          *levels;    // Element is SQueryLevel, starting from 0. SArray<SSchLevel>
+  SArray          *subPlans;  // subplan pointer copied from DAG, no need to free it in scheduler
+
+  int32_t          levelIdx;
+  SEpSet           dataSrcEps;
+  SHashObj        *execTasks; // executing tasks, key:taskid, value:SQueryTask*
+  SHashObj        *succTasks; // succeed tasks, key:taskid, value:SQueryTask*
+  SHashObj        *failTasks; // failed tasks, key:taskid, value:SQueryTask*
+
+  int8_t           status;  
+  SQueryNodeAddr   resNode;
   tsem_t           rspSem;
   int32_t          userFetch;
   int32_t          remoteFetch;
   SSchTask        *fetchTask;
-
   int32_t          errCode;
   void            *res;
   int32_t          resNumOfRows;
-
-  SHashObj        *execTasks; // executing tasks, key:taskid, value:SQueryTask*
-  SHashObj        *succTasks; // succeed tasks, key:taskid, value:SQueryTask*
-  SHashObj        *failTasks; // failed tasks, key:taskid, value:SQueryTask*
-
-  SArray          *levels;    // Element is SQueryLevel, starting from 0. SArray<SSchLevel>
-  SArray          *subPlans;  // Element is SArray*, and nested element is SSubplan. The execution level of subplan, starting from 0. SArray<void*>
-
   SQueryProfileSummary summary;
 } SSchJob;
 
 #define SCH_HAS_QNODE_IN_CLUSTER(type) (false) //TODO CLUSTER TYPE
-#define SCH_TASK_READY_TO_LUNCH(task) ((task)->childReady >= taosArrayGetSize((task)->children))   // MAY NEED TO ENHANCE
+#define SCH_TASK_READY_TO_LUNCH(task) (atomic_load_32(&(task)->childReady) >= taosArrayGetSize((task)->children))
 #define SCH_IS_DATA_SRC_TASK(task) ((task)->plan->type == QUERY_TYPE_SCAN)
 #define SCH_TASK_NEED_WAIT_ALL(task) ((task)->plan->type == QUERY_TYPE_MODIFY)
 
-#define SCH_JOB_ERR_LOG(param, ...) qError("QID:%"PRIx64 param, job->queryId, __VA_ARGS__)
-#define SCH_TASK_ERR_LOG(param, ...) qError("QID:%"PRIx64",TID:%"PRIx64 param, job->queryId, task->taskId, __VA_ARGS__)
+#define SCH_SET_TASK_STATUS(task, st) atomic_store_8(&(task)->status, st) 
+#define SCH_GET_TASK_STATUS(task) atomic_load_8(&(task)->status)
+
+#define SCH_SET_JOB_STATUS(job, st) atomic_store_8(&(job)->status, st) 
+#define SCH_GET_JOB_STATUS(job) atomic_load_8(&(job)->status)
+
+#define SCH_SET_JOB_TYPE(pAttr, type) (pAttr)->queryJob = ((type) != QUERY_TYPE_MODIFY)
+#define SCH_JOB_NEED_FETCH(pAttr) ((pAttr)->queryJob)
+
+#define SCH_JOB_ELOG(param, ...) qError("QID:%"PRIx64" " param, pJob->queryId, __VA_ARGS__)
+#define SCH_JOB_DLOG(param, ...) qDebug("QID:%"PRIx64" " param, pJob->queryId, __VA_ARGS__)
+
+#define SCH_TASK_ELOG(param, ...) qError("QID:%"PRIx64",TID:%"PRIx64" " param, pJob->queryId, pTask->taskId, __VA_ARGS__)
+#define SCH_TASK_DLOG(param, ...) qDebug("QID:%"PRIx64",TID:%"PRIx64" " param, pJob->queryId, pTask->taskId, __VA_ARGS__)
 
 #define SCH_ERR_RET(c) do { int32_t _code = c; if (_code != TSDB_CODE_SUCCESS) { terrno = _code; return _code; } } while (0)
 #define SCH_RET(c) do { int32_t _code = c; if (_code != TSDB_CODE_SUCCESS) { terrno = _code; } return _code; } while (0)
-#define SCH_ERR_LRET(c,...) do { int32_t _code = c; if (_code != TSDB_CODE_SUCCESS) { qError(__VA_ARGS__); terrno = _code; return _code; } } while (0)
 #define SCH_ERR_JRET(c) do { code = c; if (code != TSDB_CODE_SUCCESS) { terrno = code; goto _return; } } while (0)
 
 #define SCH_LOCK(type, _lock) (SCH_READ == (type) ? taosRLockLatch(_lock) : taosWLockLatch(_lock))
 #define SCH_UNLOCK(type, _lock) (SCH_READ == (type) ? taosRUnLockLatch(_lock) : taosWUnLockLatch(_lock))
 
 
-extern int32_t schLaunchTask(SSchJob *job, SSchTask *task);
-extern int32_t schBuildAndSendMsg(SSchJob *job, SSchTask *task, int32_t msgType);
+static int32_t schLaunchTask(SSchJob *job, SSchTask *task);
+static int32_t schBuildAndSendMsg(SSchJob *job, SSchTask *task, int32_t msgType);
 
 #ifdef __cplusplus
 }