Merge remote-tracking branch 'origin/3.0' into feature/dnode3

CMakeLists.txt

@@ -50,6 +50,11 @@ if(${BUILD_WITH_LUCENE})
     cat("${CMAKE_SUPPORT_DIR}/lucene_CMakeLists.txt.in" ${DEPS_TMP_FILE})
 endif(${BUILD_WITH_LUCENE})
 
+## NuRaft
+if(${BUILD_WITH_NURAFT})
+    cat("${CMAKE_SUPPORT_DIR}/nuraft_CMakeLists.txt.in" ${DEPS_TMP_FILE})
+endif(${BUILD_WITH_NURAFT})
+
 ## download dependencies
 configure_file(${DEPS_TMP_FILE} "${CMAKE_SOURCE_DIR}/deps/deps-download/CMakeLists.txt")
 execute_process(COMMAND "${CMAKE_COMMAND}" -G "${CMAKE_GENERATOR}" .
@@ -69,4 +74,7 @@ target_include_directories(api INTERFACE "include/client")
 # src
 add_subdirectory(source)
 
+# docs 
+add_subdirectory(docs)
+
 # tests (TODO)

Jenkinsfile

@@ -142,7 +142,7 @@ pipeline {
           }
       }
   //     stage('Parallel test stage') {
-  //       //only build pr
+  //     skip defaultCheckout 
   //       options { skipDefaultCheckout() } 
   //       when {
   //         allOf{

cmake/cmake.options

@@ -16,7 +16,7 @@ option(
 option(
     BUILD_WITH_ROCKSDB 
     "If build with rocksdb" 
-    OFF
+    ON
 )
 
 option(
@@ -25,8 +25,20 @@ option(
     OFF
 )
 
+option(
+    BUILD_WITH_NURAFT
+    "If build with NuRaft" 
+    OFF
+)
+
 option(
     BUILD_DEPENDENCY_TESTS 
     "If build dependency tests" 
     OFF
+)
+
+option(
+    BUILD_DOCS
+    "If use doxygen build documents" 
+    ON
 )

cmake/nuraft_CMakeLists.txt.in

@@ -0,0 +1,12 @@
+
+# NuRaft
+ExternalProject_Add(NuRaft
+        GIT_REPOSITORY https://github.com/eBay/NuRaft.git
+        GIT_TAG v1.3.0
+        SOURCE_DIR "${CMAKE_SOURCE_DIR}/deps/nuraft"
+        BINARY_DIR "${CMAKE_SOURCE_DIR}/deps/nuraft"
+        CONFIGURE_COMMAND "./prepare.sh"
+        BUILD_COMMAND ""
+        INSTALL_COMMAND ""
+        TEST_COMMAND ""
+        )

deps/CMakeLists.txt

@@ -67,6 +67,12 @@ if(${BUILD_WITH_LUCENE})
     add_subdirectory(lucene)
 endif(${BUILD_WITH_LUCENE})
 
+# NuRaft
+if(${BUILD_WITH_NURAFT})
+    add_subdirectory(nuraft)
+endif(${BUILD_WITH_NURAFT})
+
+
 # ================================================================================================
 # DEPENDENCY TEST
 # ================================================================================================

docs/CMakeLists.txt

@@ -0,0 +1,23 @@
+# Generate API documentation
+## https://vicrucann.github.io/tutorials/quick-cmake-doxygen/
+if(${BUILD_DOCS})
+    find_package(Doxygen)
+    if (DOXYGEN_FOUND)
+        # Build the doc
+        set(DOXYGEN_IN ${CMAKE_SOURCE_DIR}/docs/Doxyfile.in)
+        set(DOXYGEN_OUT ${CMAKE_BINARY_DIR}/Doxyfile)
+
+        configure_file(${DOXYGEN_IN} ${DOXYGEN_OUT} @ONLY)
+        message("Doxygen build start")
+
+        add_custom_target(
+            tdengine_doxygen ALL
+            COMMAND ${DOXYGEN_EXECUTABLE} ${DOXYGEN_OUT}
+            WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR}
+            COMMENT "Generating API doxumentation with Doxygen"
+            VERBATIM
+        )
+    else(DOXYGEN_FOUND)
+        message("Doxygen need to be installed to generate the doxygen documentation")
+    endif(DOXYGEN_FOUND)
+endif(${BUILD_DOCS})

include/common/common.h

@@ -19,7 +19,7 @@
 #include "taosdef.h"
 #include "taosmsg.h"
 #include "tarray.h"
-
+#include "tvariant.h"
 //typedef struct STimeWindow {
 //  TSKEY skey;
 //  TSKEY ekey;
@@ -66,4 +66,59 @@ typedef struct SColumnInfoData {
   char       *pData;    // the corresponding block data in memory
 } SColumnInfoData;
 
+//======================================================================================================================
+// the following structure shared by parser and executor
+typedef struct SColumn {
+  uint64_t     uid;
+  char         name[TSDB_COL_NAME_LEN];
+  int8_t       flag;      // column type: normal column, tag, or user-input column (integer/float/string)
+  SColumnInfo  info;
+} SColumn;
+
+typedef struct SLimit {
+  int64_t   limit;
+  int64_t   offset;
+} SLimit;
+
+typedef struct SOrder {
+  uint32_t  order;
+  int32_t   orderColId;
+} SOrder;
+
+typedef struct SGroupbyExpr {
+  SArray*  columnInfo;  // SArray<SColIndex>, group by columns information
+  bool     groupbyTag;  // group by tag or column
+} SGroupbyExpr;
+
+// the structure for sql function in select clause
+typedef struct SSqlExpr {
+  char      token[TSDB_COL_NAME_LEN];      // original token
+  SSchema   resSchema;
+
+  int32_t   numOfCols;
+  SColumn*  pColumns;       // data columns that are required by query
+  int32_t   interBytes;     // inter result buffer size
+  int16_t   numOfParams;    // argument value of each function
+  SVariant  param[3];       // parameters are not more than 3
+} SSqlExpr;
+
+typedef struct SExprInfo {
+  struct SSqlExpr    base;
+  struct tExprNode  *pExpr;
+} SExprInfo;
+
+typedef struct SStateWindow {
+  SColumn col;
+} SStateWindow;
+
+typedef struct SSessionWindow {
+  int64_t gap;             // gap between two session window(in microseconds)
+  SColumn col;
+} SSessionWindow;
+
+#define QUERY_ASC_FORWARD_STEP   1
+#define QUERY_DESC_FORWARD_STEP -1
+
+#define GET_FORWARD_DIRECTION_FACTOR(ord) (((ord) == TSDB_ORDER_ASC) ? QUERY_ASC_FORWARD_STEP : QUERY_DESC_FORWARD_STEP)
+
 #endif  // TDENGINE_COMMON_H

include/common/taosmsg.h

@@ -204,14 +204,14 @@ enum _mgmt_table {
 #define TSDB_COL_NORMAL             0x0u    // the normal column of the table
 #define TSDB_COL_TAG                0x1u    // the tag column type
 #define TSDB_COL_UDC                0x2u    // the user specified normal string column, it is a dummy column
-#define TSDB_COL_NULL               0x4u    // the column filter NULL or not
+#define TSDB_COL_TMP                0x4u    // internal column generated by the previous operators
+#define TSDB_COL_NULL               0x8u    // the column filter NULL or not
 
 #define TSDB_COL_IS_TAG(f)          (((f&(~(TSDB_COL_NULL)))&TSDB_COL_TAG) != 0)
 #define TSDB_COL_IS_NORMAL_COL(f)   ((f&(~(TSDB_COL_NULL))) == TSDB_COL_NORMAL)
 #define TSDB_COL_IS_UD_COL(f)       ((f&(~(TSDB_COL_NULL))) == TSDB_COL_UDC)
 #define TSDB_COL_REQ_NULL(f)        (((f)&TSDB_COL_NULL) != 0)
 
-
 extern char *taosMsg[];
 
 #pragma pack(push, 1)
@@ -489,11 +489,6 @@ typedef struct SInterval {
   int64_t offset;
 } SInterval;
 
-typedef struct SSessionWindow {
-  int64_t gap;             // gap between two session window(in microseconds)
-  int32_t primaryColId;    // primary timestamp column
-} SSessionWindow;
-
 typedef struct {
   SMsgHead    head;
   char        version[TSDB_VERSION_LEN];
@@ -518,7 +513,7 @@ typedef struct {
   int16_t     orderColId;
   int16_t     numOfCols;        // the number of columns will be load from vnode
   SInterval   interval;
-  SSessionWindow sw;            // session window
+//  SSessionWindow sw;            // session window
   uint16_t    tagCondLen;       // tag length in current query
   uint16_t    colCondLen;       // column length in current query
   int16_t     numOfGroupCols;   // num of group by columns

include/common/tname.h

@@ -16,6 +16,8 @@
 #ifndef TDENGINE_TNAME_H
 #define TDENGINE_TNAME_H
 
+#include "taosmsg.h"
+
 #define TSDB_DB_NAME_T     1
 #define TSDB_TABLE_NAME_T  2
 
@@ -52,6 +54,8 @@ int32_t tNameFromString(SName* dst, const char* str, uint32_t type);
 
 int32_t tNameSetAcctId(SName* dst, const char* acct);
 
+SSchema* tGetTbnameColumnSchema();
+
 #if 0
 int32_t tNameSetDbName(SName* dst, const char* acct, SToken* dbToken);
 #endif

include/common/trequest.h

@@ -0,0 +1,47 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#ifndef _TD_TREQUEST_H_
+#define _TD_TREQUEST_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* ------------------------ TYPES EXPOSED ------------------------ */
+typedef struct SRequest      SRequest;
+typedef struct SReqBatch     SReqBatch;
+typedef struct SReqBatchIter SReqBatchIter;
+
+// SRequest
+
+// SReqBatch
+
+// SReqBatchIter
+void            tdInitRBIter(SReqBatchIter *pIter, SReqBatch *pReqBatch);
+const SRequest *tdRBIterNext(SReqBatchIter *pIter);
+void            tdClearRBIter(SReqBatchIter *pIter);
+
+/* ------------------------ TYPES DEFINITION ------------------------ */
+struct SReqBatchIter {
+  int        iReq;
+  SReqBatch *pReqBatch;
+};
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /*_TD_TREQUEST_H_*/

include/common/ttszip.h

@@ -21,7 +21,7 @@ extern "C" {
 #endif
 
 #include "os.h"
-#include "taosdef.h"
+#include "tdef.h"
 #include "tvariant.h"
 
 #define MEM_BUF_SIZE (1 << 20)

include/dnode/vnode/meta/impl/metaImpl.h

@@ -0,0 +1,75 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#ifndef _TD_META_IMPL_H_
+#define _TD_META_IMPL_H_
+
+#include "os.h"
+
+#include "taosmsg.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+typedef uint64_t tb_uid_t;
+
+/* ------------------------ SMetaOptions ------------------------ */
+struct SMetaOptions {
+  size_t lruCacheSize;  // LRU cache size
+};
+
+/* ------------------------ STbOptions ------------------------ */
+#define META_NORMAL_TABLE ((uint8_t)1)
+#define META_SUPER_TABLE ((uint8_t)2)
+#define META_CHILD_TABLE ((uint8_t)3)
+
+typedef struct {
+} SSMAOptions;
+
+// super table options
+typedef struct {
+  tb_uid_t  uid;
+  STSchema* pSchema;
+  STSchema* pTagSchema;
+} SSTbOptions;
+
+// child table options
+typedef struct {
+  tb_uid_t suid;
+  SKVRow   tags;
+} SCTbOptions;
+
+// normal table options
+typedef struct {
+  STSchema* pSchame;
+} SNTbOptions;
+
+struct STbOptions {
+  uint8_t     type;
+  char*       name;
+  uint32_t    ttl;   // time to live in (SECONDS)
+  SSMAOptions bsma;  // Block-wise sma
+  union {
+    SSTbOptions stbOptions;
+    SNTbOptions ntbOptions;
+    SCTbOptions ctbOptions;
+  };
+};
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /*_TD_META_IMPL_H_*/

include/dnode/vnode/meta/meta.h

@@ -16,81 +16,36 @@
 #ifndef _TD_META_H_
 #define _TD_META_H_
 
-#include "taosmsg.h"
+#include "impl/metaImpl.h"
-
-#include "os.h"
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-/* ------------------------ APIs Exposed ------------------------ */
-
 // Types exported
-typedef uint64_t                tb_uid_t;
+typedef struct SMeta        SMeta;
-typedef struct SMeta            SMeta;
+typedef struct SMetaOptions SMetaOptions;
-typedef struct SMetaOpts        SMetaOpts;
+typedef struct STbOptions   STbOptions;
-typedef struct SMetaQueryHandle SMetaQueryHandle;
-typedef struct SMetaQueryOpts   SMetaQueryOpts;
-typedef struct STableOpts       STableOpts;
 
 // SMeta operations
-int    metaCreate(const char *path);
+SMeta *metaOpen(const char *path, const SMetaOptions *);
-void   metaDestroy(const char *path);
-SMeta *metaOpen(SMetaOpts *);
 void   metaClose(SMeta *);
-int    metaCreateTable(SMeta *, const STableOpts *);
+void   metaRemove(const char *path);
-int    metaDropTable(SMeta *, uint64_t tuid_t);
+int    metaCreateTable(SMeta *pMeta, const STbOptions *);
-int    metaAlterTable(SMeta *, void *);
+int    metaDropTable(SMeta *pMeta, tb_uid_t uid);
 int    metaCommit(SMeta *);
 
 // Options
-SMetaOpts *metaOptionsCreate();
+void metaOptionsInit(SMetaOptions *);
-void       metaOptionsDestroy(SMetaOpts *);
+void metaOptionsClear(SMetaOptions *);
-void       metaOptionsSetCache(SMetaOpts *, size_t capacity);
-
-// SMetaQueryHandle
-SMetaQueryHandle *metaQueryHandleCreate(SMetaQueryOpts *);
-void              metaQueryHandleDestroy(SMetaQueryHandle *);
-
-// SMetaQueryOpts
-SMetaQueryOpts *metaQueryOptionsCreate();
-void            metaQueryOptionsDestroy(SMetaQueryOpts *);
 
 // STableOpts
 #define META_TABLE_OPTS_DECLARE(name) STableOpts name = {0}
-void metaNormalTableOptsInit(STableOpts *, const char *name, const STSchema *pSchema);
+void metaNormalTableOptsInit(STbOptions *, const char *name, const STSchema *pSchema);
-void metaSuperTableOptsInit(STableOpts *, const char *name, tb_uid_t uid, const STSchema *pSchema,
+void metaSuperTableOptsInit(STbOptions *, const char *name, tb_uid_t uid, const STSchema *pSchema,
                             const STSchema *pTagSchema);
-void metaChildTableOptsInit(STableOpts *, const char *name, tb_uid_t suid, const SKVRow tags);
+void metaChildTableOptsInit(STbOptions *, const char *name, tb_uid_t suid, const SKVRow tags);
-void metaTableOptsClear(STableOpts *);
+void metaTableOptsClear(STbOptions *);
-
-/* ------------------------ Impl should hidden ------------------------ */
-typedef enum { META_INIT_TABLE = 0, META_SUPER_TABLE = 1, META_CHILD_TABLE = 2, META_NORMAL_TABLE = 3 } EMetaTableT;
-typedef struct SSuperTableOpts {
-  tb_uid_t  uid;
-  STSchema *pSchema;     // (ts timestamp, a int)
-  STSchema *pTagSchema;  // (tag1 binary(10), tag2 int)
-} SSuperTableOpts;
-
-typedef struct SChildTableOpts {
-  tb_uid_t suid;  // super table uid
-  SKVRow   tags;  // tag value of the child table
-} SChildTableOpts;
-
-typedef struct SNormalTableOpts {
-  STSchema *pSchema;
-} SNormalTableOpts;
-
-struct STableOpts {
-  int8_t type;
-  char * name;
-  union {
-    SSuperTableOpts  superOpts;
-    SChildTableOpts  childOpts;
-    SNormalTableOpts normalOpts;
-  };
-};
 
 #ifdef __cplusplus
 }

include/dnode/vnode/tq/tq.h

@@ -19,9 +19,6 @@
 #include "os.h"
 #include "tutil.h"
 
-#define TQ_ACTION_INSERT 0x7f7f7f7fULL
-#define TQ_ACTION_DELETE 0x80808080ULL
-
 #ifdef __cplusplus
 extern "C" {
 #endif

include/dnode/vnode/tsdb/impl/tsdbImpl.h

@@ -23,6 +23,7 @@ extern "C" {
 #endif
 
 struct STsdbOptions {
+  size_t lruCacheSize;
   /* TODO */
 };
 

include/dnode/vnode/vnode.h

@@ -17,14 +17,80 @@
 #define _TD_VNODE_H_
 
 #include "os.h"
-#include "taosmsg.h"
+#include "trequest.h"
-#include "trpc.h"
+
+#include "meta.h"
+#include "tq.h"
+#include "tsdb.h"
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-typedef struct SVnode SVnode;
+/* ------------------------ TYPES EXPOSED ------------------------ */
+typedef struct SVnode        SVnode;
+typedef struct SVnodeOptions SVnodeOptions;
+
+/* ------------------------ SVnode ------------------------ */
+SVnode *vnodeOpen(const char *path, const SVnodeOptions *pVnodeOptions);
+void    vnodeClose(SVnode *pVnode);
+void    vnodeDestroy(const char *path);
+int     vnodeProcessWriteReqs(SVnode *pVnode, SReqBatch *pReqBatch);
+int     vnodeApplyWriteRequest(SVnode *pVnode, const SRequest *pRequest);
+int     vnodeProcessReadReq(SVnode *pVnode, SRequest *pReq);
+int     vnodeProcessSyncReq(SVnode *pVnode, SRequest *pReq);
+
+/* ------------------------ SVnodeOptions ------------------------ */
+void vnodeOptionsInit(SVnodeOptions *);
+void vnodeOptionsClear(SVnodeOptions *);
+
+/* ------------------------ STRUCT DEFINITIONS ------------------------ */
+struct SVnodeOptions {
+  /**
+   * @brief write buffer size in BYTES
+   *
+   */
+  uint64_t wsize;
+
+  /**
+   * @brief time to live of tables in this vnode
+   * in SECONDS
+   *
+   */
+  uint32_t ttl;
+
+  /**
+   * @brief if time-series requests eventual consistency
+   *
+   */
+  bool isWeak;
+
+  /**
+   * @brief if the allocator is heap allcator or arena allocator
+   *
+   */
+  bool isHeapAllocator;
+
+  /**
+   * @brief TSDB options
+   *
+   */
+  STsdbOptions tsdbOptions;
+
+  /**
+   * @brief META options
+   *
+   */
+  SMetaOptions metaOptions;
+  // STqOptions   tqOptions; // TODO
+};
+
+/* ------------------------ FOR COMPILE ------------------------ */
+
+#if 1
+
+#include "taosmsg.h"
+#include "trpc.h"
 
 typedef struct {
   char     db[TSDB_FULL_DB_NAME_LEN];
@@ -71,8 +137,6 @@ typedef struct {
 int32_t vnodeInit(SVnodePara);
 void    vnodeCleanup();
 
-SVnode *vnodeOpen(int32_t vgId, const char *path);
-void    vnodeClose(SVnode *pVnode);
 int32_t vnodeAlter(SVnode *pVnode, const SVnodeCfg *pCfg);
 SVnode *vnodeCreate(int32_t vgId, const char *path, const SVnodeCfg *pCfg);
 void    vnodeDrop(SVnode *pVnode);
@@ -86,6 +150,8 @@ int32_t    vnodeAppendMsg(SVnodeMsg *pMsg, SRpcMsg *pRpcMsg);
 void       vnodeCleanupMsg(SVnodeMsg *pMsg);
 void       vnodeProcessMsg(SVnode *pVnode, SVnodeMsg *pMsg, EVnMsgType msgType);
 
+#endif
+
 #ifdef __cplusplus
 }
 #endif

include/libs/function/function.h

@@ -26,8 +26,8 @@ extern "C" {
 
 #define MAX_INTERVAL_TIME_WINDOW 1000000  // maximum allowed time windows in final results
 
-#define FUNCTION_SCALAR       1
+#define FUNCTION_TYPE_SCALAR       1
-#define FUNCTION_AGG          2
+#define FUNCTION_TYPE_AGG          2
 
 #define TOP_BOTTOM_QUERY_LIMIT    100
 #define FUNCTIONS_NAME_MAX_LENGTH 16
@@ -78,13 +78,30 @@ extern "C" {
 #define FUNCTION_MODE         36
 #define FUNCTION_SAMPLE       37
 
+#define FUNCTION_COV          38
+
+// determine the real data need to calculated the result
+enum {
+  BLK_DATA_NO_NEEDED     = 0x0,
+  BLK_DATA_STATIS_NEEDED = 0x1,
+  BLK_DATA_ALL_NEEDED    = 0x3,
+  BLK_DATA_DISCARD       = 0x4,   // discard current data block since it is not qualified for filter
+};
+
+enum {
+  MASTER_SCAN   = 0x0u,
+  REVERSE_SCAN  = 0x1u,
+  REPEAT_SCAN   = 0x2u,  //repeat scan belongs to the master scan
+  MERGE_STAGE   = 0x20u,
+};
+
 typedef struct SPoint1 {
   int64_t   key;
   union{double  val; char* ptr;};
 } SPoint1;
 
 struct SQLFunctionCtx;
-struct SResultRowCellInfo;
+struct SResultRowEntryInfo;
 
 //for selectivity query, the corresponding tag value is assigned if the data is qualified
 typedef struct SExtTagsInfo {
@@ -93,6 +110,23 @@ typedef struct SExtTagsInfo {
   struct SQLFunctionCtx **pTagCtxList;
 } SExtTagsInfo;
 
+typedef struct SResultDataInfo {
+  int16_t type;
+  int16_t bytes;
+  int32_t intermediateBytes;
+} SResultDataInfo;
+
+#define GET_RES_INFO(ctx) ((ctx)->resultInfo)
+
+typedef struct SFunctionFpSet {
+  bool (*init)(struct SQLFunctionCtx *pCtx, struct SResultRowEntryInfo* pResultCellInfo);  // setup the execute environment
+  void (*addInput)(struct SQLFunctionCtx *pCtx);
+
+  // finalizer must be called after all exec has been executed to generated final result.
+  void (*finalize)(struct SQLFunctionCtx *pCtx);
+  void (*combine)(struct SQLFunctionCtx *pCtx);
+} SFunctionFpSet;
+
 // sql function runtime context
 typedef struct SQLFunctionCtx {
   int32_t      size;      // number of rows
@@ -101,9 +135,7 @@ typedef struct SQLFunctionCtx {
   int16_t      inputType;
   int16_t      inputBytes;
 
-  int16_t      outputType;
+  SResultDataInfo resDataInfo;
-  int16_t      outputBytes;   // size of results, determined by function and input column data type
-  int32_t      interBufBytes; // internal buffer size
   bool         hasNull;       // null value exist in current block
   bool         requireNull;   // require null in some function
   bool         stableQuery;
@@ -117,18 +149,21 @@ typedef struct SQLFunctionCtx {
   void        *ptsOutputBuf;  // corresponding output buffer for timestamp of each result, e.g., top/bottom*/
   SVariant     tag;
 
-  bool         isSmaSet;
+  bool        isAggSet;
-  SColumnDataAgg sma;
+  SColumnDataAgg agg;
-  struct  SResultRowCellInfo *resultInfo;
+  struct  SResultRowEntryInfo *resultInfo;
   SExtTagsInfo tagInfo;
   SPoint1      start;
   SPoint1      end;
+
+  SFunctionFpSet* fpSet;
 } SQLFunctionCtx;
 
 enum {
   TEXPR_NODE_DUMMY     = 0x0,
   TEXPR_BINARYEXPR_NODE= 0x1,
   TEXPR_UNARYEXPR_NODE = 0x2,
+  TEXPR_FUNCTION_NODE  = 0x3,
   TEXPR_COL_NODE       = 0x4,
   TEXPR_VALUE_NODE     = 0x8,
 };
@@ -137,10 +172,7 @@ typedef struct tExprNode {
   uint8_t nodeType;
   union {
     struct {
-      union {
+      int32_t           optr;   // binary operator
-        int32_t         optr;   // binary operator
-        int32_t         functionId;// unary operator
-      };
       void             *info;   // support filter operation on this expression only available for leaf node
       struct tExprNode *pLeft;  // left child pointer
       struct tExprNode *pRight; // right child pointer
@@ -148,44 +180,52 @@ typedef struct tExprNode {
 
     SSchema            *pSchema;// column node
     struct SVariant    *pVal;   // value node
+
+    struct {// function node
+      char              functionName[FUNCTIONS_NAME_MAX_LENGTH];
+//      int32_t           functionId;
+      int32_t           num;
+
+      // Note that the attribute of pChild is not the parameter of function, it is the columns that involved in the
+      // calculation instead.
+      // E.g., Cov(col1, col2), the column information, w.r.t. the col1 and col2, is kept in pChild nodes.
+      //  The concat function, concat(col1, col2), is a binary scalar
+      //  operator and is kept in the attribute of _node.
+      struct tExprNode **pChild;
+    } _function;
   };
 } tExprNode;
 
+//TODO create?
 void exprTreeToBinary(SBufferWriter* bw, tExprNode* pExprTree);
 void tExprTreeDestroy(tExprNode *pNode, void (*fp)(void *));
 
 typedef struct SAggFunctionInfo {
-  char     name[FUNCTIONS_NAME_MAX_LENGTH];
+  char      name[FUNCTIONS_NAME_MAX_LENGTH];
-  int8_t   type;         // Scalar function or aggregation function
+  int8_t    type;         // Scalar function or aggregation function
-  uint8_t  functionId;   // Function Id
+  uint32_t  functionId;   // Function Id
-  int8_t   sFunctionId;  // Transfer function for super table query
+  int8_t    sFunctionId;  // Transfer function for super table query
-  uint16_t status;
+  uint16_t  status;
 
-  bool (*init)(SQLFunctionCtx *pCtx, struct SResultRowCellInfo* pResultCellInfo);  // setup the execute environment
+  bool (*init)(SQLFunctionCtx *pCtx, struct SResultRowEntryInfo* pResultCellInfo);  // setup the execute environment
-  void (*exec)(SQLFunctionCtx *pCtx);
+  void (*addInput)(SQLFunctionCtx *pCtx);
 
   // finalizer must be called after all exec has been executed to generated final result.
-  void (*xFinalize)(SQLFunctionCtx *pCtx);
+  void (*finalize)(SQLFunctionCtx *pCtx);
-  void (*mergeFunc)(SQLFunctionCtx *pCtx);
+  void (*combine)(SQLFunctionCtx *pCtx);
 
   int32_t (*dataReqFunc)(SQLFunctionCtx *pCtx, STimeWindow* w, int32_t colId);
 } SAggFunctionInfo;
 
+struct SScalarFuncParam;
+
 typedef struct SScalarFunctionInfo {
-  char     name[FUNCTIONS_NAME_MAX_LENGTH];
+  char      name[FUNCTIONS_NAME_MAX_LENGTH];
-  int8_t   type;              // scalar function or aggregation function
+  int8_t    type;              // scalar function or aggregation function
-  uint8_t  functionId;        // index of scalar function
+  uint32_t  functionId;        // index of scalar function
-
+  void     (*process)(struct SScalarFuncParam* pOutput, size_t numOfInput, const struct SScalarFuncParam *pInput);
-  bool    (*init)(SQLFunctionCtx *pCtx, struct SResultRowCellInfo* pResultCellInfo);  // setup the execute environment
-  void    (*exec)(SQLFunctionCtx *pCtx);
 } SScalarFunctionInfo;
 
-typedef struct SResultDataInfo {
-  int16_t type;
-  int16_t bytes;
-  int32_t intermediateBytes;
-} SResultDataInfo;
-
 typedef struct SMultiFunctionsDesc {
   bool stableQuery;
   bool groupbyColumn;
@@ -195,11 +235,12 @@ typedef struct SMultiFunctionsDesc {
   bool hasFilter;
   bool onlyTagQuery;
   bool orderProjectQuery;
-  bool stateWindow;
   bool globalMerge;
   bool multigroupResult;
   bool blockDistribution;
+  bool stateWindow;
   bool timewindow;
+  bool sessionWindow;
   bool topbotQuery;
   bool interpQuery;
   bool distinct;
@@ -215,16 +256,61 @@ int32_t getResultDataInfo(int32_t dataType, int32_t dataBytes, int32_t functionI
  * @param len
  * @return
  */
-int32_t qIsBuiltinFunction(const char* name, int32_t len);
+int32_t qIsBuiltinFunction(const char* name, int32_t len, bool* scalarFunction);
 
 bool qIsValidUdf(SArray* pUdfInfo, const char* name, int32_t len, int32_t* functionId);
 
-const char* qGetFunctionName(int32_t functionId);
+bool qIsAggregateFunction(const char* functionName);
+
+tExprNode* exprTreeFromBinary(const void* data, size_t size);
 
 void extractFunctionDesc(SArray* pFunctionIdList, SMultiFunctionsDesc* pDesc);
 
 tExprNode* exprdup(tExprNode* pTree);
 
+void resetResultRowEntryResult(SQLFunctionCtx* pCtx, int32_t num);
+void cleanupResultRowEntry(struct SResultRowEntryInfo* pCell);
+int32_t getNumOfResult(SQLFunctionCtx* pCtx, int32_t num);
+bool isRowEntryCompleted(struct SResultRowEntryInfo* pEntry);
+bool isRowEntryInitialized(struct SResultRowEntryInfo* pEntry);
+
+struct SScalarFunctionSupport* createScalarFuncSupport(int32_t num);
+void destroyScalarFuncSupport(struct SScalarFunctionSupport* pSupport, int32_t num);
+struct SScalarFunctionSupport* getScalarFuncSupport(struct SScalarFunctionSupport* pSupport, int32_t index);
+
+///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// fill api
+struct SFillInfo;
+struct SFillColInfo;
+
+typedef struct SPoint {
+  int64_t key;
+  void *  val;
+} SPoint;
+
+void taosFillSetStartInfo(struct SFillInfo* pFillInfo, int32_t numOfRows, TSKEY endKey);
+void taosResetFillInfo(struct SFillInfo* pFillInfo, TSKEY startTimestamp);
+void taosFillSetInputDataBlock(struct SFillInfo* pFillInfo, const struct SSDataBlock* pInput);
+struct SFillColInfo* createFillColInfo(SExprInfo* pExpr, int32_t numOfOutput, const int64_t* fillVal);
+bool taosFillHasMoreResults(struct SFillInfo* pFillInfo);
+
+struct SFillInfo* taosCreateFillInfo(int32_t order, TSKEY skey, int32_t numOfTags, int32_t capacity, int32_t numOfCols,
+                              int64_t slidingTime, int8_t slidingUnit, int8_t precision, int32_t fillType,
+                              struct SFillColInfo* pFillCol, void* handle);
+
+void* taosDestroyFillInfo(struct SFillInfo *pFillInfo);
+int64_t taosFillResultDataBlock(struct SFillInfo* pFillInfo, void** output, int32_t capacity);
+int64_t getFillInfoStart(struct SFillInfo *pFillInfo);
+
+int32_t taosGetLinearInterpolationVal(SPoint* point, int32_t outputType, SPoint* point1, SPoint* point2, int32_t inputType);
+
+///////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
+// udf api
+struct SUdfInfo;
+
+void qAddUdfInfo(uint64_t id, struct SUdfInfo* pUdfInfo);
+void qRemoveUdfInfo(uint64_t id, struct SUdfInfo* pUdfInfo);
+
 #ifdef __cplusplus
 }
 #endif

include/libs/parser/parser.h

@@ -26,50 +26,6 @@ extern "C" {
 #include "tvariant.h"
 #include "function.h"
 
-typedef struct SColumn {
-  uint64_t     tableUid;
-  int32_t      columnIndex;
-  SColumnInfo  info;
-} SColumn;
-
-// the structure for sql function in select clause
-typedef struct SSqlExpr {
-  char      token[TSDB_COL_NAME_LEN];      // original token
-  SSchema   resSchema;
-  SColIndex colInfo;
-  uint64_t  uid;            // table uid, todo refactor use the pointer
-  int32_t   interBytes;     // inter result buffer size
-  int16_t   numOfParams;    // argument value of each function
-  SVariant  param[3];       // parameters are not more than 3
-} SSqlExpr;
-
-typedef struct SExprInfo {
-  SSqlExpr           base;
-  struct tExprNode  *pExpr;
-} SExprInfo;
-
-//typedef struct SInterval {
-//  int32_t  tz;            // query client timezone
-//  char     intervalUnit;
-//  char     slidingUnit;
-//  char     offsetUnit;
-//  int64_t  interval;
-//  int64_t  sliding;
-//  int64_t  offset;
-//} SInterval;
-//
-//typedef struct SSessionWindow {
-//  int64_t  gap;             // gap between two session window(in microseconds)
-//  int32_t  primaryColId;    // primary timestamp column
-//} SSessionWindow;
-
-typedef struct SGroupbyExpr {
-  int16_t  tableIndex;
-  SArray*  columnInfo;  // SArray<SColIndex>, group by columns information
-  int16_t  orderIndex;  // order by column index
-  int16_t  orderType;   // order by type: asc/desc
-} SGroupbyExpr;
-
 typedef struct SField {
   char     name[TSDB_COL_NAME_LEN];
   uint8_t  type;
@@ -82,16 +38,6 @@ typedef struct SFieldInfo {
   SArray     *internalField; // SArray<SInternalField>
 } SFieldInfo;
 
-typedef struct SLimit {
-  int64_t   limit;
-  int64_t   offset;
-} SLimit;
-
-typedef struct SOrder {
-  uint32_t  order;
-  int32_t   orderColId;
-} SOrder;
-
 typedef struct SCond {
   uint64_t uid;
   int32_t  len;  // length of tag query condition data
@@ -120,12 +66,6 @@ typedef struct STagCond {
 typedef struct STableMetaInfo {
   STableMeta    *pTableMeta;      // table meta, cached in client side and acquired by name
   SVgroupsInfo  *vgroupList;
-
-  /*
-   * 1. keep the vgroup index during the multi-vnode super table projection query
-   * 2. keep the vgroup index for multi-vnode insertion
-   */
-  int32_t       vgroupIndex;
   SName         name;
   char          aliasName[TSDB_TABLE_NAME_LEN];    // alias name of table specified in query sql
   SArray       *tagColList;                        // SArray<SColumn*>, involved tag columns
@@ -137,11 +77,11 @@ typedef struct SQueryStmtInfo {
   STimeWindow      window;        // the whole query time window
   SInterval        interval;      // tumble time window
   SSessionWindow   sessionWindow; // session time window
+  SStateWindow     stateWindow;   // state window query
   SGroupbyExpr     groupbyExpr;   // groupby tags info
   SArray *         colList;       // SArray<SColumn*>
   SFieldInfo       fieldsInfo;
-  SArray *         exprList;      // SArray<SExprInfo*>
+  SArray**         exprList;      // SArray<SExprInfo*>
-  SArray *         exprList1;     // final exprlist in case of arithmetic expression exists
   SLimit           limit;
   SLimit           slimit;
   STagCond         tagCond;
@@ -172,6 +112,7 @@ typedef struct SQueryStmtInfo {
   struct SQueryStmtInfo *pDownstream;
   int32_t            havingFieldNum;
   SMultiFunctionsDesc     info;
+  int32_t            exprListLevelIndex;
 } SQueryStmtInfo;
 
 typedef struct SColumnIndex {
@@ -225,14 +166,23 @@ void assignExprInfo(SExprInfo* dst, const SExprInfo* src);
 void columnListCopy(SArray* dst, const SArray* src, uint64_t uid);
 void columnListDestroy(SArray* pColumnList);
 
-void dropAllExprInfo(SArray* pExprInfo);
+void dropAllExprInfo(SArray** pExprInfo, int32_t numOfLevel);
-SExprInfo* createExprInfo(STableMetaInfo* pTableMetaInfo, int16_t functionId, SColumnIndex* pColIndex, struct tExprNode* pParamExpr, SSchema* pResSchema, int16_t interSize);
+
+typedef struct SSourceParam {
+  SArray            *pExprNodeList; //Array<struct tExprNode*>
+  SArray            *pColumnList;   //Array<struct SColumn>
+  int32_t    num;
+} SSourceParam;
+
+SExprInfo* createExprInfo(STableMetaInfo* pTableMetaInfo, const char* funcName, SSourceParam* pSource, SSchema* pResSchema, int16_t interSize);
 int32_t copyExprInfoList(SArray* dst, const SArray* src, uint64_t uid, bool deepcopy);
 
 STableMetaInfo* getMetaInfo(SQueryStmtInfo* pQueryInfo, int32_t tableIndex);
 SSchema *getOneColumnSchema(const STableMeta* pTableMeta, int32_t colIndex);
+SSchema createSchema(uint8_t type, int16_t bytes, int16_t colId, const char* name);
 
 int32_t getNewResColId();
+void addIntoSourceParam(SSourceParam* pSourceParam, tExprNode* pNode, SColumn* pColumn);
 
 #ifdef __cplusplus
 }

include/libs/sync/sync.h

@@ -26,13 +26,13 @@ extern "C" {
 typedef int32_t  SyncNodeId;
 typedef int32_t  SyncGroupId;
 typedef int64_t  SyncIndex;
-typedef uint64_t SSyncTerm;
+typedef uint64_t SyncTerm;
 
 typedef enum {
-  TAOS_SYNC_ROLE_FOLLOWER = 0,
+  TAOS_SYNC_STATE_FOLLOWER = 0,
-  TAOS_SYNC_ROLE_CANDIDATE = 1,
+  TAOS_SYNC_STATE_CANDIDATE = 1,
-  TAOS_SYNC_ROLE_LEADER = 2,
+  TAOS_SYNC_STATE_LEADER = 2,
-} ESyncRole;
+} ESyncState;
 
 typedef struct {
   void*  data;
@@ -55,19 +55,19 @@ typedef struct {
   int32_t   selfIndex;
   int32_t   replica;
   SNodeInfo node[TSDB_MAX_REPLICA];
-  ESyncRole role[TSDB_MAX_REPLICA];
+  ESyncState role[TSDB_MAX_REPLICA];
 } SNodesRole;
 
 typedef struct SSyncFSM {
   void* pData;
 
-  // apply committed log, bufs will be free by raft module
+  // apply committed log, bufs will be free by sync module
   int32_t (*applyLog)(struct SSyncFSM* fsm, SyncIndex index, const SSyncBuffer* buf, void* pData);
 
   // cluster commit callback
   int32_t (*onClusterChanged)(struct SSyncFSM* fsm, const SSyncCluster* cluster, void* pData);
 
-  // fsm return snapshot in ppBuf, bufs will be free by raft module
+  // fsm return snapshot in ppBuf, bufs will be free by sync module
   // TODO: getSnapshot SHOULD be async?
   int32_t (*getSnapshot)(struct SSyncFSM* fsm, SSyncBuffer** ppBuf, int32_t* objId, bool* isLast);
 
@@ -89,19 +89,30 @@ typedef struct SSyncLogStore {
   // write log with given index
   int32_t (*logWrite)(struct SSyncLogStore* logStore, SyncIndex index, SSyncBuffer* pBuf);
 
+  /** 
+   * read log from given index(included) with limit, return the actual num in nBuf,
+   * pBuf will be free in sync module
+   **/
+  int32_t (*logRead)(struct SSyncLogStore* logStore, SyncIndex index, int limit,
+                      SSyncBuffer* pBuf, int* nBuf);
+
   // mark log with given index has been commtted
   int32_t (*logCommit)(struct SSyncLogStore* logStore, SyncIndex index);
 
-  // prune log before given index
+  // prune log before given index(not included)
   int32_t (*logPrune)(struct SSyncLogStore* logStore, SyncIndex index);
 
-  // rollback log after given index
+  // rollback log after given index(included)
   int32_t (*logRollback)(struct SSyncLogStore* logStore, SyncIndex index);
+
+  // return last index of log
+  SyncIndex (*logLastIndex)(struct SSyncLogStore* logStore);
 } SSyncLogStore;
 
 typedef struct SSyncServerState {
   SyncNodeId voteFor;
-  SSyncTerm  term;
+  SyncTerm  term;
+  SyncIndex  commitIndex;
 } SSyncServerState;
 
 typedef struct SSyncClusterConfig {
@@ -122,9 +133,9 @@ typedef struct SStateManager {
 
   int32_t (*readServerState)(struct SStateManager* stateMng, SSyncServerState* state);
 
-  // void (*saveCluster)(struct SStateManager* stateMng, const SSyncClusterConfig* cluster);
+  void (*saveCluster)(struct SStateManager* stateMng, const SSyncClusterConfig* cluster);
 
-  // const SSyncClusterConfig* (*readCluster)(struct SStateManager* stateMng);
+  const SSyncClusterConfig* (*readCluster)(struct SStateManager* stateMng);
 } SStateManager;
 
 typedef struct {
@@ -146,13 +157,13 @@ SSyncNode* syncStart(const SSyncInfo*);
 void       syncReconfig(const SSyncNode*, const SSyncCluster*);
 void       syncStop(const SSyncNode*);
 
-int32_t syncPropose(SSyncNode* syncNode, SSyncBuffer buffer, void* pData, bool isWeak);
+int32_t syncPropose(SSyncNode* syncNode, const SSyncBuffer* pBuf, void* pData, bool isWeak);
 
-// int32_t syncAddNode(SSyncNode syncNode, const SNodeInfo *pNode);
+int32_t syncAddNode(SSyncNode syncNode, const SNodeInfo *pNode);
 
-// int32_t syncRemoveNode(SSyncNode syncNode, const SNodeInfo *pNode);
+int32_t syncRemoveNode(SSyncNode syncNode, const SNodeInfo *pNode);
 
-extern int32_t syncDebugFlag;
+extern int32_t sDebugFlag;
 
 #ifdef __cplusplus
 }

include/os/os.h

@@ -45,6 +45,8 @@ extern "C" {
 #include <float.h>
 #include <math.h>
 #include <sys/stat.h>
+#include <sys/types.h>
+#include <unistd.h>
 
 #include "osAtomic.h"
 #include "osDef.h"

include/os/osDir.h

@@ -22,7 +22,7 @@ extern "C" {
 
 void taosRemoveDir(const char *dirname);
 bool taosDirExist(char *dirname);
-bool taosMkDir(char *dirname);
+bool taosMkDir(const char *dirname);
 void taosRemoveOldFiles(char *dirname, int32_t keepDays);
 bool taosExpandDir(char *dirname, char *outname, int32_t maxlen);
 bool taosRealPath(char *dirname, int32_t maxlen);

include/util/amalloc.h

@@ -1,52 +0,0 @@
-/*
- * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
- *
- * This program is free software: you can use, redistribute, and/or modify
- * it under the terms of the GNU Affero General Public License, version 3
- * or later ("AGPL"), as published by the Free Software Foundation.
- *
- * This program is distributed in the hope that it will be useful, but WITHOUT
- * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
- * FITNESS FOR A PARTICULAR PURPOSE.
- *
- * You should have received a copy of the GNU Affero General Public License
- * along with this program. If not, see <http://www.gnu.org/licenses/>.
- */
-
-#ifndef _TD_AMALLOC_H_
-#define _TD_AMALLOC_H_
-
-#include "os.h"
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#define AMALLOC_APIS                                  \
-  void *(*malloc)(void *, size_t size);               \
-  void *(*calloc)(void *, size_t nmemb, size_t size); \
-  void *(*realloc)(void *, size_t size);              \
-  void (*free)(void *ptr);
-
-// Interfaces to implement
-typedef struct {
-  AMALLOC_APIS
-} SMemAllocatorIf;
-
-typedef struct {
-  void *impl;
-  AMALLOC_APIS
-} SMemAllocator;
-
-#define amalloc(allocator, size) ((allocator) ? (*((allocator)->malloc))((allocator)->impl, (size)) : malloc(size))
-#define acalloc(allocator, nmemb, size) \
-  ((allocator) ? (*((allocator)->calloc))((allocator)->impl, (nmemb), (size)) : calloc((nmemb), (size)))
-#define arealloc(allocator, ptr, size) \
-  ((allocator) ? (*((allocator)->realloc))((allocator)->impl, (ptr), (size)) : realloc((ptr), (size)))
-#define afree(allocator, ptr, size) ((allocator) ? (*((allocator)->free))((allocator)->impl, (ptr), (size)) : free(ptr))
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif /*_TD_AMALLOC_H_*/

include/util/mallocator.h

@@ -0,0 +1,63 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#ifndef _TD_MALLOCATOR_H_
+#define _TD_MALLOCATOR_H_
+
+#include "os.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct SMemAllocator SMemAllocator;
+
+#define MALLOCATOR_APIS                                        \
+  void *(*malloc)(SMemAllocator *, size_t size);               \
+  void *(*calloc)(SMemAllocator *, size_t nmemb, size_t size); \
+  void *(*realloc)(SMemAllocator *, void *ptr, size_t size);   \
+  void (*free)(SMemAllocator *, void *ptr);                    \
+  size_t (*usage)(SMemAllocator *);
+
+// Interfaces to implement
+typedef struct {
+  MALLOCATOR_APIS
+} SMemAllocatorIf;
+
+struct SMemAllocator {
+  void * impl;
+  size_t usize;
+  MALLOCATOR_APIS
+};
+
+// heap allocator
+SMemAllocator *tdCreateHeapAllocator();
+void           tdDestroyHeapAllocator(SMemAllocator *pMemAllocator);
+
+// arena allocator
+SMemAllocator *tdCreateArenaAllocator(size_t size);
+void           tdDestroyArenaAllocator(SMemAllocator *);
+
+#define mMalloc(pMemAllocator, size) (*(pMemAllocator->malloc))(pMemAllocator, size)
+#define mCalloc(pMemAllocator, nmemb, size) (*(pMemAllocator->calloc))(pMemAllocator, nmemb, size)
+#define mRealloc(pMemAllocator, ptr, size) (*(pMemAllocator->realloc))(pMemAllocator, ptr, size)
+#define mFree(pMemAllocator, ptr) (*(pMemAllocator->free))(pMemAllocator, ptr)
+#define mUsage(pMemAllocator) (*(pMemAllocator->usage))(pMemAllocator)
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /*_TD_MALLOCATOR_H_*/

include/util/tcompression.h

@@ -20,7 +20,7 @@
 extern "C" {
 #endif
 
-#include "tdef.h"
+#include "taos.h"
 #include "tutil.h"
 
 #define COMP_OVERFLOW_BYTES 2

include/util/tdef.h

@@ -134,6 +134,15 @@ do { \
 #define TSDB_BINARY_OP_REMAINDER  4004
 #define TSDB_BINARY_OP_CONCAT     4005
 
+#define FUNCTION_CEIL        4500
+#define FUNCTION_FLOOR       4501
+#define FUNCTION_ABS         4502
+#define FUNCTION_ROUND       4503
+
+#define FUNCTION_LENGTH      4800
+#define FUNCTION_CONCAT      4801
+#define FUNCTION_LTRIM       4802
+#define FUNCTION_RTRIM       4803
 
 #define IS_RELATION_OPTR(op) (((op) >= TSDB_RELATION_LESS) && ((op) < TSDB_RELATION_IN))
 #define IS_ARITHMETIC_OPTR(op) (((op) >= TSDB_BINARY_OP_ADD) && ((op) <= TSDB_BINARY_OP_REMAINDER))
@@ -311,10 +320,6 @@ do { \
 #define TSDB_QUERY_TYPE_NON_TYPE               0x00u     // none type
 #define TSDB_QUERY_TYPE_FREE_RESOURCE          0x01u     // free qhandle at vnode
 
-#define TSDB_UDF_TYPE_SCALAR       1
-#define TSDB_UDF_TYPE_AGGREGATE    2
-
-
 /*
  * 1. ordinary sub query for select * from super_table
  * 2. all sqlobj generated by createSubqueryObj with this flag
@@ -367,21 +372,6 @@ do { \
 #define TSDB_MAX_DISKS_PER_TIER 16
 #define TSDB_MAX_DISKS           (TSDB_MAX_TIERS * TSDB_MAX_DISKS_PER_TIER)
 
-#define TSDB_DATA_TYPE_NULL       0     // 1 bytes
-#define TSDB_DATA_TYPE_BOOL       1     // 1 bytes
-#define TSDB_DATA_TYPE_TINYINT    2     // 1 byte
-#define TSDB_DATA_TYPE_SMALLINT   3     // 2 bytes
-#define TSDB_DATA_TYPE_INT        4     // 4 bytes
-#define TSDB_DATA_TYPE_BIGINT     5     // 8 bytes
-#define TSDB_DATA_TYPE_FLOAT      6     // 4 bytes
-#define TSDB_DATA_TYPE_DOUBLE     7     // 8 bytes
-#define TSDB_DATA_TYPE_BINARY     8     // string
-#define TSDB_DATA_TYPE_TIMESTAMP  9     // 8 bytes
-#define TSDB_DATA_TYPE_NCHAR      10    // unicode string
-#define TSDB_DATA_TYPE_UTINYINT   11    // 1 byte
-#define TSDB_DATA_TYPE_USMALLINT  12    // 2 bytes
-#define TSDB_DATA_TYPE_UINT       13    // 4 bytes
-#define TSDB_DATA_TYPE_UBIGINT    14    // 8 bytes
 
 enum { TRANS_STAT_INIT = 0, TRANS_STAT_EXECUTING, TRANS_STAT_EXECUTED, TRANS_STAT_ROLLBACKING, TRANS_STAT_ROLLBACKED };
 enum { TRANS_OPER_INIT = 0, TRANS_OPER_EXECUTE, TRANS_OPER_ROLLBACK };

include/util/tskiplist.h

@@ -21,7 +21,7 @@ extern "C" {
 #endif
 
 #include "os.h"
-//#include "tdef.h"
+#include "taos.h"
 #include "tarray.h"
 #include "tfunctional.h"
 

source/dnode/mgmt/src/dnodeVnodes.c

@@ -376,7 +376,7 @@ static void *dnodeOpenVnodeFunc(void *param) {
 
     char path[PATH_MAX + 20] = {0};
     snprintf(path, sizeof(path),"%s/vnode%d", tsVnodeDir, pVnode->vgId);
-    SVnode *pImpl = vnodeOpen(pVnode->vgId, path);
+    SVnode *pImpl = vnodeOpen(path, NULL);
     if (pImpl == NULL) {
       dError("vgId:%d, failed to open vnode by thread:%d", pVnode->vgId, pThread->threadIndex);
       pThread->failed++;

source/dnode/vnode/impl/CMakeLists.txt

@@ -7,6 +7,7 @@ target_include_directories(
 )
 target_link_libraries(
     vnode
+    PUBLIC os
     PUBLIC transport
     PUBLIC meta
     PUBLIC tq

source/dnode/vnode/impl/inc/vnodeAllocatorPool.h

@@ -13,20 +13,27 @@
  * along with this program. If not, see <http://www.gnu.org/licenses/>.
  */
 
-#ifndef _TD_COMMON_QARITHMETICOPERATOR_H_
+#ifndef _TD_VNODE_ALLOCATOR_POOL_H_
-#define _TD_COMMON_QARITHMETICOPERATOR_H_
+#define _TD_VNODE_ALLOCATOR_POOL_H_
+
+#include "vnode.h"
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-typedef void (*_arithmetic_operator_fn_t)(void *left, int32_t numLeft, int32_t leftType, void *right, int32_t numRight,
+typedef struct {
-                                          int32_t rightType, void *output, int32_t order);
+  int            nexta;
+  int            enda;
+  SMemAllocator *free[3];
+  SMemAllocator *used[3];
+} SVAllocatorPool;
 
-_arithmetic_operator_fn_t getArithmeticOperatorFn(int32_t arithmeticOptr);
+int  vnodeOpenAllocatorPool(SVnode *pVnode);
+void vnodeCloseAllocatorPool(SVnode *pVnode);
 
 #ifdef __cplusplus
 }
 #endif
 
-#endif  /*_TD_COMMON_QARITHMETICOPERATOR_H_*/
+#endif /*_TD_VNODE_ALLOCATOR_POOL_H_*/

source/dnode/vnode/impl/inc/vnodeCommit.h

@@ -16,13 +16,14 @@
 #ifndef _TD_VNODE_COMMIT_H_
 #define _TD_VNODE_COMMIT_H_
 
-#include "vnodeInt.h"
+#include "vnode.h"
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-int vnodeAsyncCommit(SVnode *pVnode);
+bool vnodeShouldCommit(SVnode *pVnode);
+int  vnodeAsyncCommit(SVnode *pVnode);
 
 #ifdef __cplusplus
 }

source/dnode/vnode/impl/inc/vnodeDef.h

@@ -0,0 +1,52 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#ifndef _TD_VNODE_DEF_H_
+#define _TD_VNODE_DEF_H_
+
+#include "mallocator.h"
+#include "sync.h"
+#include "tlockfree.h"
+
+#include "vnode.h"
+#include "vnodeAllocatorPool.h"
+#include "vnodeCommit.h"
+#include "vnodeFileSystem.h"
+#include "vnodeOptions.h"
+#include "vnodeStateMgr.h"
+#include "vnodeSync.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct SVnode {
+  char*            path;
+  SVnodeOptions    options;
+  SVState          state;
+  SVAllocatorPool* pool;
+  SMemAllocator*   inuse;
+  SMeta*           pMeta;
+  STsdb*           pTsdb;
+  STQ*             pTq;
+  SVnodeSync*      pSync;
+  SVnodeFS*        pFs;
+};
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /*_TD_VNODE_DEF_H_*/

source/dnode/vnode/impl/inc/vnodeFileSystem.h

@@ -0,0 +1,30 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#ifndef _TD_VNODE_FILE_SYSTEM_H_
+#define _TD_VNODE_FILE_SYSTEM_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+} SVnodeFS;
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /*_TD_VNODE_FILE_SYSTEM_H_*/

source/dnode/vnode/impl/inc/vnodeInt.h

@@ -18,7 +18,6 @@
 
 #include "vnode.h"
 
-#include "amalloc.h"
 #include "meta.h"
 #include "sync.h"
 #include "tlog.h"
@@ -39,16 +38,6 @@ extern int32_t vDebugFlag;
 #define vDebug(...) { if (vDebugFlag & DEBUG_DEBUG) { taosPrintLog("VND ", vDebugFlag, __VA_ARGS__); }}
 #define vTrace(...) { if (vDebugFlag & DEBUG_TRACE) { taosPrintLog("VND ", vDebugFlag, __VA_ARGS__); }}
 
-typedef struct SVnode {
-  int32_t    vgId;
-  SVnodeCfg  cfg;
-  SMeta     *pMeta;
-  STsdb     *pTsdb;
-  STQ       *pTQ;
-  SWal      *pWal;
-  SSyncNode *pSync;
-} SVnode;
-
 #ifdef __cplusplus
 }
 #endif

source/dnode/vnode/impl/inc/vnodeMemAllocator.h

@@ -16,40 +16,12 @@
 #ifndef _TD_VNODE_MEM_ALLOCATOR_H_
 #define _TD_VNODE_MEM_ALLOCATOR_H_
 
-#include "vnodeInt.h"
-
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-typedef struct SVnodeMemAllocator SVnodeMemAllocator;
-
-SVnodeMemAllocator *VMACreate(size_t size /* base size */, size_t ssize /* step size */,
-                              size_t threshold /* threshold size when full*/);
-void                VMADestroy(SVnodeMemAllocator *pvma);
-void                VMAReset(SVnodeMemAllocator *pvma);
-void *              VMAMalloc(SVnodeMemAllocator *pvma, size_t size);
-void                VMAFree(SVnodeMemAllocator *pvma, void *ptr);
-bool                VMAIsFull(SVnodeMemAllocator *pvma);
-
-// ------------------ FOR TEST ONLY ------------------
-typedef struct SVMANode {
-  struct SVMANode *prev;
-  size_t           tsize;
-  size_t           used;
-  char             data[];
-} SVMANode;
-
-struct SVnodeMemAllocator {
-  bool      full;       // if allocator is full
-  size_t    threshold;  // threshold;
-  size_t    ssize;      // step size to allocate
-  SVMANode *inuse;      // inuse node to allocate
-  SVMANode  node;       // basic node to use
-};
-
 #ifdef __cplusplus
 }
 #endif
 
-#endif /*_TD_VNODE_MEM_ALLOCATOR_H_*/
+#endif /*_TD_VNODE_MEM_ALLOCATOR_H_*/

source/dnode/vnode/impl/inc/vnodeOptions.h

@@ -0,0 +1,34 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#ifndef _TD_VNODE_OPTIONS_H_
+#define _TD_VNODE_OPTIONS_H_
+
+#include "vnode.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern const SVnodeOptions defaultVnodeOptions;
+
+int  vnodeValidateOptions(const SVnodeOptions *);
+void vnodeOptionsCopy(SVnodeOptions *pDest, const SVnodeOptions *pSrc);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /*_TD_VNODE_OPTIONS_H_*/

source/dnode/vnode/impl/inc/vnodeRequest.h

@@ -0,0 +1,27 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#ifndef _TD_VNODE_REQUEST_H_
+#define _TD_VNODE_REQUEST_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /*_TD_VNODE_REQUEST_H_*/

source/dnode/vnode/impl/inc/vnodeStateMgr.h

@@ -0,0 +1,30 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#ifndef _TD_VNODE_STATE_MGR_H_
+#define _TD_VNODE_STATE_MGR_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+} SVState;
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /*_TD_VNODE_STATE_MGR_H_*/

source/dnode/vnode/impl/inc/vnodeSync.h

@@ -0,0 +1,33 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#ifndef _TD_VNODE_SYNC_H_
+#define _TD_VNODE_SYNC_H_
+
+#include "sync.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+  /* data */
+} SVnodeSync;
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /*_TD_VNODE_SYNC_H_*/

source/dnode/vnode/impl/inc/vnodeWrite.h

@@ -16,10 +16,11 @@
 #ifndef _TD_VNODE_WRITE_H_
 #define _TD_VNODE_WRITE_H_
 
+#include "vnode.h"
+
 #ifdef __cplusplus
 extern "C" {
 #endif
-#include "vnodeInt.h"
 
 void vnodeProcessWriteMsg(SVnode* pVnode, SVnodeMsg* pMsg);
 

source/dnode/vnode/impl/src/vnodeAllocatorPool.c

@@ -0,0 +1,37 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include "vnodeDef.h"
+
+int vnodeOpenAllocatorPool(SVnode *pVnode) {
+  // TODO
+  return 0;
+}
+
+void vnodeCloseAllocatorPool(SVnode *pVnode) {
+  if (pVnode->pool) {
+  }
+}
+
+/* ------------------------ STATIC METHODS ------------------------ */
+static SVAllocatorPool *vapCreate() {
+  SVAllocatorPool *pPool = NULL;
+  /* TODO */
+  return pPool;
+}
+
+static void vapDestroy() {
+  // TODO
+}

source/dnode/vnode/impl/src/vnodeCommit.c

@@ -13,13 +13,15 @@
  * along with this program. If not, see <http://www.gnu.org/licenses/>.
  */
 
-#include "vnodeInt.h"
+#include "vnodeDef.h"
 
 static int vnodeStartCommit(SVnode *pVnode);
 static int vnodeEndCommit(SVnode *pVnode);
 
+bool vnodeShouldCommit(SVnode *pVnode) { return false; }
+
 int vnodeAsyncCommit(SVnode *pVnode) {
-  #if 0
+#if 0
   if (vnodeStartCommit(pVnode) < 0) {
     // TODO
   }

source/dnode/vnode/impl/src/vnodeFileSystem.c

@@ -0,0 +1,16 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include "vnodeDef.h"

source/dnode/vnode/impl/src/vnodeInt.c

@@ -20,8 +20,6 @@
 int32_t vnodeInit(SVnodePara para) { return 0; }
 void    vnodeCleanup() {}
 
-SVnode *vnodeOpen(int32_t vgId, const char *path) { return NULL; }
-void    vnodeClose(SVnode *pVnode) {}
 int32_t vnodeAlter(SVnode *pVnode, const SVnodeCfg *pCfg) { return 0; }
 SVnode *vnodeCreate(int32_t vgId, const char *path, const SVnodeCfg *pCfg) { return NULL; }
 void    vnodeDrop(SVnode *pVnode) {}
@@ -69,4 +67,4 @@ void vnodeProcessMsg(SVnode *pVnode, SVnodeMsg *pMsg, EVnMsgType msgType) {
     case VN_MSG_TYPE_FETCH:
       break;
   }
-}
+}

source/dnode/vnode/impl/src/vnodeMain.c

@@ -0,0 +1,125 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include "vnodeDef.h"
+
+static SVnode *vnodeNew(const char *path, const SVnodeOptions *pVnodeOptions);
+static void    vnodeFree(SVnode *pVnode);
+static int     vnodeOpenImpl(SVnode *pVnode);
+static void    vnodeCloseImpl(SVnode *pVnode);
+
+SVnode *vnodeOpen(const char *path, const SVnodeOptions *pVnodeOptions) {
+  SVnode *pVnode = NULL;
+
+  // Set default options
+  if (pVnodeOptions == NULL) {
+    pVnodeOptions = &defaultVnodeOptions;
+  }
+
+  // Validate options
+  if (vnodeValidateOptions(pVnodeOptions) < 0) {
+    // TODO
+    return NULL;
+  }
+
+  // Create the handle
+  pVnode = vnodeNew(path, pVnodeOptions);
+  if (pVnode == NULL) {
+    // TODO: handle error
+    return NULL;
+  }
+
+  taosMkDir(path);
+
+  // Open the vnode
+  if (vnodeOpenImpl(pVnode) < 0) {
+    // TODO: handle error
+    return NULL;
+  }
+
+  return pVnode;
+}
+
+void vnodeClose(SVnode *pVnode) {
+  if (pVnode) {
+    vnodeCloseImpl(pVnode);
+    vnodeFree(pVnode);
+  }
+}
+
+void vnodeDestroy(const char *path) { taosRemoveDir(path); }
+
+/* ------------------------ STATIC METHODS ------------------------ */
+static SVnode *vnodeNew(const char *path, const SVnodeOptions *pVnodeOptions) {
+  SVnode *pVnode = NULL;
+
+  pVnode = (SVnode *)calloc(1, sizeof(*pVnode));
+  if (pVnode == NULL) {
+    // TODO
+    return NULL;
+  }
+
+  pVnode->path = strdup(path);
+  vnodeOptionsCopy(&(pVnode->options), pVnodeOptions);
+
+  return pVnode;
+}
+
+static void vnodeFree(SVnode *pVnode) {
+  if (pVnode) {
+    tfree(pVnode->path);
+    free(pVnode);
+  }
+}
+
+static int vnodeOpenImpl(SVnode *pVnode) {
+  char dir[TSDB_FILENAME_LEN];
+
+  // Open allocator pool
+  if (vnodeOpenAllocatorPool(pVnode) < 0) {
+    // TODO: handle error
+    return -1;
+  }
+
+  // Open meta
+  sprintf(dir, "%s/meta", pVnode->path);
+  pVnode->pMeta = metaOpen(dir, &(pVnode->options.metaOptions));
+  if (pVnode->pMeta == NULL) {
+    // TODO: handle error
+    return -1;
+  }
+
+  // Open tsdb
+  sprintf(dir, "%s/tsdb", pVnode->path);
+  pVnode->pTsdb = tsdbOpen(dir, &(pVnode->options.tsdbOptions));
+  if (pVnode->pTsdb == NULL) {
+    // TODO: handle error
+    return -1;
+  }
+
+  // TODO: Open TQ
+
+  // TODO
+  return 0;
+}
+
+static void vnodeCloseImpl(SVnode *pVnode) {
+  if (pVnode) {
+    vnodeCloseAllocatorPool(pVnode);
+    // TODO: Close TQ
+    tsdbClose(pVnode->pTsdb);
+    metaClose(pVnode->pMeta);
+  }
+}

source/dnode/vnode/impl/src/vnodeMemAllocator.c

@@ -13,112 +13,88 @@
  * along with this program. If not, see <http://www.gnu.org/licenses/>.
  */
 
-#include "vnodeMemAllocator.h"
+#include "vnodeDef.h"
 
-#define VMA_IS_FULL(pvma) \
+#define VNODE_HEAP_ALLOCATOR 0
-  (((pvma)->inuse != &((pvma)->node)) || ((pvma)->inuse->tsize - (pvma)->inuse->used < (pvma)->threshold))
+#define VNODE_ARENA_ALLOCATOR 1
 
-static SVMANode *VMANodeNew(size_t size);
+typedef struct {
-static void      VMANodeFree(SVMANode *node);
+  uint64_t tsize;
+  uint64_t used;
+} SVHeapAllocator;
 
-SVnodeMemAllocator *VMACreate(size_t size, size_t ssize, size_t threshold) {
+typedef struct SVArenaNode {
-  SVnodeMemAllocator *pvma = NULL;
+  struct SVArenaNode *prev;
+  void *              nptr;
+  char                data[];
+} SVArenaNode;
 
-  if (size < threshold) {
+typedef struct {
+  SVArenaNode *inuse;
+  SVArenaNode  node;
+} SVArenaAllocator;
+
+typedef struct {
+  int8_t   type;
+  uint64_t tsize;
+  T_REF_DECLARE()
+  union {
+    SVHeapAllocator  vha;
+    SVArenaAllocator vaa;
+  };
+} SVMemAllocator;
+
+SMemAllocator *vnodeCreateMemAllocator(int8_t type, uint64_t tsize, uint64_t ssize /* step size only for arena */) {
+  SMemAllocator * pma;
+  uint64_t        msize;
+  SVMemAllocator *pva;
+
+  msize = sizeof(*pma) + sizeof(SVMemAllocator);
+  if (type == VNODE_ARENA_ALLOCATOR) {
+    msize += tsize;
+  }
+
+  pma = (SMemAllocator *)calloc(1, msize);
+  if (pma == NULL) {
     return NULL;
   }
 
-  pvma = (SVnodeMemAllocator *)malloc(sizeof(*pvma) + size);
+  pma->impl = POINTER_SHIFT(pma, sizeof(*pma));
-  if (pvma) {
+  pva = (SVMemAllocator *)(pma->impl);
-    pvma->full = false;
+  pva->type = type;
-    pvma->threshold = threshold;
+  pva->tsize = tsize;
-    pvma->ssize = ssize;
-    pvma->inuse = &(pvma->node);
 
-    pvma->inuse->prev = NULL;
+  if (type == VNODE_HEAP_ALLOCATOR) {
-    pvma->inuse->tsize = size;
+    pma->malloc = NULL;
-    pvma->inuse->used = 0;
+    pma->calloc = NULL;
+    pma->realloc = NULL;
+    pma->free = NULL;
+    pma->usage = NULL;
+  } else if (type == VNODE_ARENA_ALLOCATOR) {
+    pma->malloc = NULL;
+    pma->calloc = NULL;
+    pma->realloc = NULL;
+    pma->free = NULL;
+    pma->usage = NULL;
+  } else {
+    ASSERT(0);
   }
 
-  return pvma;
+  return pma;
 }
 
-void VMADestroy(SVnodeMemAllocator *pvma) {
+void vnodeDestroyMemAllocator(SMemAllocator *pma) {
-  if (pvma) {
+  // TODO
-    VMAReset(pvma);
-    free(pvma);
-  }
 }
 
-void VMAReset(SVnodeMemAllocator *pvma) {
+void vnodeRefMemAllocator(SMemAllocator *pma) {
-  while (pvma->inuse != &(pvma->node)) {
+  // TODO
-    SVMANode *node = pvma->inuse;
-    pvma->inuse = node->prev;
-    VMANodeFree(node);
-  }
-
-  pvma->inuse->used = 0;
-  pvma->full = false;
 }
 
-void *VMAMalloc(SVnodeMemAllocator *pvma, size_t size) {
+void vnodeUnrefMemAllocator(SMemAllocator *pma) {
-  void * ptr = NULL;
+  // TODO
-  size_t tsize = size + sizeof(size_t);
-
-  if (pvma->inuse->tsize - pvma->inuse->used < tsize) {
-    SVMANode *pNode = VMANodeNew(MAX(pvma->ssize, tsize));
-    if (pNode == NULL) {
-      return NULL;
-    }
-
-    pNode->prev = pvma->inuse;
-    pvma->inuse = pNode;
-  }
-
-  ptr = pvma->inuse->data + pvma->inuse->used;
-  pvma->inuse->used += tsize;
-  *(size_t *)ptr = size;
-  ptr = POINTER_SHIFT(ptr, sizeof(size_t));
-
-  pvma->full = VMA_IS_FULL(pvma);
-
-  return ptr;
 }
 
-void VMAFree(SVnodeMemAllocator *pvma, void *ptr) {
+/* ------------------------ Heap Allocator IMPL ------------------------ */
-  if (ptr) {
-    size_t size = *(size_t *)POINTER_SHIFT(ptr, -sizeof(size_t));
-    if (POINTER_SHIFT(ptr, size) == pvma->inuse->data + pvma->inuse->used) {
-      pvma->inuse->used -= (size + sizeof(size_t));
 
-      if ((pvma->inuse->used == 0) && (pvma->inuse != &(pvma->node))) {
+/* ------------------------ Arena Allocator IMPL ------------------------ */
-        SVMANode *node = pvma->inuse;
-        pvma->inuse = node->prev;
-        VMANodeFree(node);
-      }
-
-      pvma->full = VMA_IS_FULL(pvma);
-    }
-  }
-}
-
-bool VMAIsFull(SVnodeMemAllocator *pvma) { return pvma->full; }
-
-static SVMANode *VMANodeNew(size_t size) {
-  SVMANode *node = NULL;
-
-  node = (SVMANode *)malloc(sizeof(*node) + size);
-  if (node) {
-    node->prev = NULL;
-    node->tsize = size;
-    node->used = 0;
-  }
-
-  return node;
-}
-
-static void VMANodeFree(SVMANode *node) {
-  if (node) {
-    free(node);
-  }
-}

source/dnode/vnode/impl/src/vnodeOptions.c

@@ -0,0 +1,34 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include "vnodeDef.h"
+
+const SVnodeOptions defaultVnodeOptions = {0}; /* TODO */
+
+void vnodeOptionsInit(SVnodeOptions *pVnodeOptions) { /* TODO */
+  vnodeOptionsCopy(pVnodeOptions, &defaultVnodeOptions);
+}
+
+void vnodeOptionsClear(SVnodeOptions *pVnodeOptions) { /* TODO */
+}
+
+int vnodeValidateOptions(const SVnodeOptions *pVnodeOptions) {
+  // TODO
+  return 0;
+}
+
+void vnodeOptionsCopy(SVnodeOptions *pDest, const SVnodeOptions *pSrc) {
+  memcpy((void *)pDest, (void *)pSrc, sizeof(SVnodeOptions));
+}

source/dnode/vnode/impl/src/vnodeRead.c

@@ -13,5 +13,4 @@
  * along with this program. If not, see <http://www.gnu.org/licenses/>.
  */
 
-#define _DEFAULT_SOURCE
+#include "vnodeDef.h"
-#include "vnodeRead.h"

source/dnode/vnode/impl/src/vnodeRequest.c

@@ -0,0 +1,14 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */

source/dnode/vnode/impl/src/vnodeStateMgr.c

@@ -0,0 +1,14 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */

source/dnode/vnode/impl/src/vnodeSync.c

@@ -0,0 +1,14 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */

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

@@ -13,5 +13,50 @@
  * along with this program. If not, see <http://www.gnu.org/licenses/>.
  */
 
-#define _DEFAULT_SOURCE
+#include "vnodeDef.h"
-#include "vnodeWrite.h"
+
+int vnodeProcessWriteReqs(SVnode *pVnode, SReqBatch *pReqBatch) {
+  /* TODO */
+  return 0;
+}
+
+int vnodeApplyWriteRequest(SVnode *pVnode, const SRequest *pRequest) {
+  int      reqType;        /* TODO */
+  size_t   reqSize;        /* TODO */
+  uint64_t reqVersion = 0; /* TODO */
+  int      code = 0;
+
+  // Copy the request to vnode buffer
+  SRequest *pReq = mMalloc(pVnode->inuse, reqSize);
+  if (pReq == NULL) {
+    // TODO: handle error
+  }
+
+  memcpy(pReq, pRequest, reqSize);
+
+  // Push the request to TQ so consumers can consume
+  tqPushMsg(pVnode->pTq, pReq, 0);
+
+  // Process the request
+  switch (reqType) {
+    case TSDB_MSG_TYPE_CREATE_TABLE:
+      code = metaCreateTable(pVnode->pMeta, NULL /* TODO */);
+      break;
+    case TSDB_MSG_TYPE_DROP_TABLE:
+      code = metaDropTable(pVnode->pMeta, 0 /* TODO */);
+      break;
+      /* TODO */
+    default:
+      break;
+  }
+
+  if (vnodeShouldCommit(pVnode)) {
+    if (vnodeAsyncCommit(pVnode) < 0) {
+      // TODO: handle error
+    }
+  }
+
+  return code;
+}
+
+/* ------------------------ STATIC METHODS ------------------------ */

source/dnode/vnode/impl/test/CMakeLists.txt

@@ -1,9 +1,7 @@
-# vnodeMemAllocatorTest
+# Vnode API test
-add_executable(VMATest "")
+add_executable(vnodeApiTests "")
-target_sources(VMATest
+target_sources(vnodeApiTests
     PRIVATE
-    "../src/vnodeMemAllocator.c"
+    "vnodeApiTests.cpp"
-    "vnodeMemAllocatorTest.cpp"
 )
-target_include_directories(VMATest PUBLIC "${CMAKE_CURRENT_SOURCE_DIR}/../inc")
+target_link_libraries(vnodeApiTests vnode gtest gtest_main)
-target_link_libraries(VMATest os gtest_main vnode)

source/dnode/vnode/impl/test/vnodeApiTests.cpp

@@ -0,0 +1,13 @@
+#include <gtest/gtest.h>
+#include <iostream>
+
+#include "vnode.h"
+
+TEST(vnodeApiTest, vnodeOpen_vnodeClose_test) {
+  // Create and open a vnode
+  SVnode *pVnode = vnodeOpen("vnode1", NULL);
+  ASSERT_NE(pVnode, nullptr);
+
+  // Close the vnode
+  vnodeClose(pVnode);
+}

source/dnode/vnode/impl/test/vnodeMemAllocatorTest.cpp

@@ -1,22 +0,0 @@
-#include <gtest/gtest.h>
-#include <iostream>
-
-#include "vnodeMemAllocator.h"
-
-TEST(VMATest, basic_create_and_destroy_test) {
-  SVnodeMemAllocator *vma = VMACreate(1024, 512, 64);
-  EXPECT_TRUE(vma != nullptr);
-  EXPECT_EQ(vma->full, false);
-  EXPECT_EQ(vma->ssize, 512);
-  EXPECT_EQ(vma->threshold, 64);
-  EXPECT_EQ(vma->inuse->tsize, 1024);
-  VMADestroy(vma);
-
-  vma = VMACreate(1024, 512, 1024);
-  EXPECT_TRUE(vma != nullptr);
-  VMADestroy(vma);
-
-  vma = VMACreate(1024, 512, 1025);
-  EXPECT_TRUE(vma == nullptr);
-  VMADestroy(vma);
-}

source/dnode/vnode/meta/inc/metaCache.h

@@ -0,0 +1,36 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#ifndef _TD_META_CACHE_H_
+#define _TD_META_CACHE_H_
+
+#include "rocksdb/c.h"
+
+#include "meta.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef rocksdb_cache_t meta_cache_t;
+
+int  metaOpenCache(SMeta *pMeta);
+void metaCloseCache(SMeta *pMeta);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /*_TD_META_CACHE_H_*/

source/dnode/vnode/meta/inc/metaDB.h

@@ -0,0 +1,44 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#ifndef _TD_META_DB_H_
+#define _TD_META_DB_H_
+
+#include "rocksdb/c.h"
+
+#include "meta.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef struct {
+  rocksdb_t *tbDb;      // uid -> tb obj
+  rocksdb_t *nameDb;    // name -> uid
+  rocksdb_t *tagDb;     // uid -> tag
+  rocksdb_t *schemaDb;  // uid+version -> schema
+  rocksdb_t *mapDb;     // suid -> uid_list
+} meta_db_t;
+
+int  metaOpenDB(SMeta *pMeta);
+void metaCloseDB(SMeta *pMeta);
+int  metaSaveTableToDB(SMeta *pMeta, const STbOptions *pTbOptions);
+int  metaRemoveTableFromDb(SMeta *pMeta, tb_uid_t uid);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /*_TD_META_DB_H_*/

source/dnode/vnode/meta/inc/metaDef.h

@@ -16,22 +16,26 @@
 #ifndef _TD_META_DEF_H_
 #define _TD_META_DEF_H_
 
-#include "metaUid.h"
+#include "meta.h"
-#include "tkv.h"
+#include "metaCache.h"
+#include "metaDB.h"
+#include "metaIdx.h"
+#include "metaOptions.h"
+#include "metaTbOptions.h"
+#include "metaTbTag.h"
+#include "metaTbUid.h"
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
 struct SMeta {
-  STableUidGenerator uidGenerator;
+  char*           path;     // path of current meta
-
+  SMetaOptions    options;  // meta option
-  STkvDb* tableDb;   // uid->table obj
+  meta_db_t*      pDB;      // raw data db
-  STkvDb* tbnameDb;  // tbname --> uid
+  meta_index_t*   pIdx;     // tag index
-  STkvDb* schemaDb;  // uid+version --> schema
+  meta_cache_t*   pCache;   // LRU cache
-  STkvDb* tagDb;     // uid --> tag
+  STbUidGenerator uidGnrt;  // meta table UID generator
-  STkvDb* tagIdx;    // TODO: need to integrate lucene or our own
-  // STkvCache* metaCache; // TODO: add a global cache here
 };
 
 #ifdef __cplusplus

source/dnode/vnode/meta/inc/metaIdx.h

@@ -0,0 +1,38 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#ifndef _TD_META_IDX_H_
+#define _TD_META_IDX_H_
+
+#include "rocksdb/c.h"
+
+#include "meta.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef rocksdb_t meta_index_t;
+
+int  metaOpenIdx(SMeta *pMeta);
+void metaCloseIdx(SMeta *pMeta);
+int  metaSaveTableToIdx(SMeta *pMeta, const STbOptions *pTbOptions);
+int  metaRemoveTableFromIdx(SMeta *pMeta, tb_uid_t uid);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /*_TD_META_IDX_H_*/

source/dnode/vnode/meta/inc/metaOptions.h

@@ -0,0 +1,34 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#ifndef _TD_META_OPTIONS_H_
+#define _TD_META_OPTIONS_H_
+
+#include "meta.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+extern const SMetaOptions defaultMetaOptions;
+
+int  metaValidateOptions(const SMetaOptions *);
+void metaOptionsCopy(SMetaOptions *pDest, const SMetaOptions *pSrc);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /*_TD_META_OPTIONS_H_*/

source/dnode/vnode/meta/inc/metaQuery.h

@@ -0,0 +1,27 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#ifndef _TD_META_QUERY_H_
+#define _TD_META_QUERY_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /*_TD_META_QUERY_H_*/

source/dnode/vnode/meta/inc/metaTbOptions.h

@@ -0,0 +1,32 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#ifndef _TD_META_TABLE_OPTIONS_H_
+#define _TD_META_TABLE_OPTIONS_H_
+
+#include "meta.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int    metaValidateTbOptions(SMeta *pMeta, const STbOptions *);
+size_t metaEncodeTbObjFromTbOptions(const STbOptions *, void *pBuf, size_t bsize);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /*_TD_META_TABLE_OPTIONS_H_*/

source/dnode/vnode/meta/inc/metaTbTag.h

@@ -0,0 +1,27 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#ifndef _TD_META_TB_TAG_H_
+#define _TD_META_TB_TAG_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /*_TD_META_TB_TAG_H_*/

source/dnode/vnode/meta/inc/metaTbUid.h

@@ -23,20 +23,17 @@ extern "C" {
 #endif
 
 /* ------------------------ APIS EXPOSED ------------------------ */
-typedef struct STableUidGenerator STableUidGenerator;
+typedef struct STbUidGenerator {
+  tb_uid_t nextUid;
+} STbUidGenerator;
+
+// STableUidGenerator
+int  metaOpenUidGnrt(SMeta *pMeta);
+void metaCloseUidGnrt(SMeta *pMeta);
 
 // tb_uid_t
 #define IVLD_TB_UID 0
-tb_uid_t generateUid(STableUidGenerator *);
+tb_uid_t metaGenerateUid(SMeta *pMeta);
-
-// STableUidGenerator
-void tableUidGeneratorInit(STableUidGenerator *, tb_uid_t suid);
-#define tableUidGeneratorClear(ug)
-
-/* ------------------------ FOR TEST AND COMPILE ONLY ------------------------ */
-struct STableUidGenerator {
-  tb_uid_t nextUid;
-};
 
 #ifdef __cplusplus
 }

source/dnode/vnode/meta/src/metaCache.c

@@ -0,0 +1,37 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include "meta.h"
+#include "metaDef.h"
+
+int metaOpenCache(SMeta *pMeta) {
+  // TODO
+  if (pMeta->options.lruCacheSize) {
+    pMeta->pCache = rocksdb_cache_create_lru(pMeta->options.lruCacheSize);
+    if (pMeta->pCache == NULL) {
+      // TODO: handle error
+      return -1;
+    }
+  }
+
+  return 0;
+}
+
+void metaCloseCache(SMeta *pMeta) {
+  if (pMeta->pCache) {
+    rocksdb_cache_destroy(pMeta->pCache);
+    pMeta->pCache = NULL;
+  }
+}

source/dnode/vnode/meta/src/metaCommit.c

@@ -0,0 +1,21 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include "meta.h"
+
+int metaCommit(SMeta *pMeta) {
+  // TODO
+  return 0;
+}

source/dnode/vnode/meta/src/metaDB.c

@@ -0,0 +1,199 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include "metaDef.h"
+
+static void metaSaveSchemaDB(SMeta *pMeta, tb_uid_t uid, STSchema *pSchema);
+static void metaGetSchemaDBKey(char key[], tb_uid_t uid, int sversion);
+static int  metaSaveMapDB(SMeta *pMeta, tb_uid_t suid, tb_uid_t uid);
+
+#define SCHEMA_KEY_LEN (sizeof(tb_uid_t) + sizeof(int))
+
+#define META_OPEN_DB_IMPL(pDB, options, dir, err) \
+  do {                                            \
+    pDB = rocksdb_open(options, dir, &err);       \
+    if (pDB == NULL) {                            \
+      metaCloseDB(pMeta);                         \
+      rocksdb_options_destroy(options);           \
+      return -1;                                  \
+    }                                             \
+  } while (0)
+
+int metaOpenDB(SMeta *pMeta) {
+  char               dir[128];
+  char *             err = NULL;
+  rocksdb_options_t *options = rocksdb_options_create();
+
+  if (pMeta->pCache) {
+    rocksdb_options_set_row_cache(options, pMeta->pCache);
+  }
+  rocksdb_options_set_create_if_missing(options, 1);
+
+  pMeta->pDB = (meta_db_t *)calloc(1, sizeof(*(pMeta->pDB)));
+  if (pMeta->pDB == NULL) {
+    // TODO: handle error
+    rocksdb_options_destroy(options);
+    return -1;
+  }
+
+  // tbDb
+  sprintf(dir, "%s/tb_db", pMeta->path);
+  META_OPEN_DB_IMPL(pMeta->pDB->tbDb, options, dir, err);
+
+  // nameDb
+  sprintf(dir, "%s/name_db", pMeta->path);
+  META_OPEN_DB_IMPL(pMeta->pDB->nameDb, options, dir, err);
+
+  // tagDb
+  sprintf(dir, "%s/tag_db", pMeta->path);
+  META_OPEN_DB_IMPL(pMeta->pDB->tagDb, options, dir, err);
+
+  // schemaDb
+  sprintf(dir, "%s/schema_db", pMeta->path);
+  META_OPEN_DB_IMPL(pMeta->pDB->schemaDb, options, dir, err);
+
+  // mapDb
+  sprintf(dir, "%s/map_db", pMeta->path);
+  META_OPEN_DB_IMPL(pMeta->pDB->mapDb, options, dir, err);
+
+  rocksdb_options_destroy(options);
+  return 0;
+}
+
+#define META_CLOSE_DB_IMPL(pDB) \
+  do {                          \
+    if (pDB) {                  \
+      rocksdb_close(pDB);       \
+      pDB = NULL;               \
+    }                           \
+  } while (0)
+
+void metaCloseDB(SMeta *pMeta) {
+  if (pMeta->pDB) {
+    META_CLOSE_DB_IMPL(pMeta->pDB->mapDb);
+    META_CLOSE_DB_IMPL(pMeta->pDB->schemaDb);
+    META_CLOSE_DB_IMPL(pMeta->pDB->tagDb);
+    META_CLOSE_DB_IMPL(pMeta->pDB->nameDb);
+    META_CLOSE_DB_IMPL(pMeta->pDB->tbDb);
+    free(pMeta->pDB);
+    pMeta->pDB = NULL;
+  }
+}
+
+int metaSaveTableToDB(SMeta *pMeta, const STbOptions *pTbOptions) {
+  tb_uid_t uid;
+  char *   err = NULL;
+  size_t   size;
+  char     pBuf[1024];  // TODO
+
+  rocksdb_writeoptions_t *wopt = rocksdb_writeoptions_create();
+
+  // Generate a uid for child and normal table
+  if (pTbOptions->type == META_SUPER_TABLE) {
+    uid = pTbOptions->stbOptions.uid;
+  } else {
+    uid = metaGenerateUid(pMeta);
+  }
+
+  // Save tbname -> uid to tbnameDB
+  rocksdb_put(pMeta->pDB->nameDb, wopt, pTbOptions->name, strlen(pTbOptions->name), (char *)(&uid), sizeof(uid), &err);
+
+  // Save uid -> tb_obj to tbDB
+  size = metaEncodeTbObjFromTbOptions(pTbOptions, pBuf, 1024);
+  rocksdb_put(pMeta->pDB->tbDb, wopt, (char *)(&uid), sizeof(uid), pBuf, size, &err);
+
+  switch (pTbOptions->type) {
+    case META_NORMAL_TABLE:
+      // save schemaDB
+      metaSaveSchemaDB(pMeta, uid, pTbOptions->ntbOptions.pSchame);
+      break;
+    case META_SUPER_TABLE:
+      // save schemaDB
+      metaSaveSchemaDB(pMeta, uid, pTbOptions->stbOptions.pSchema);
+
+      // save mapDB (really need?)
+      rocksdb_put(pMeta->pDB->mapDb, wopt, (char *)(&uid), sizeof(uid), "", 0, &err);
+      break;
+    case META_CHILD_TABLE:
+      // save tagDB
+      rocksdb_put(pMeta->pDB->tagDb, wopt, (char *)(&uid), sizeof(uid), pTbOptions->ctbOptions.tags,
+                  kvRowLen(pTbOptions->ctbOptions.tags), &err);
+
+      // save mapDB
+      metaSaveMapDB(pMeta, pTbOptions->ctbOptions.suid, uid);
+      break;
+    default:
+      ASSERT(0);
+  }
+
+  rocksdb_writeoptions_destroy(wopt);
+
+  return 0;
+}
+
+int metaRemoveTableFromDb(SMeta *pMeta, tb_uid_t uid) {
+  /* TODO */
+  return 0;
+}
+
+/* ------------------------ STATIC METHODS ------------------------ */
+static void metaSaveSchemaDB(SMeta *pMeta, tb_uid_t uid, STSchema *pSchema) {
+  char   key[64];
+  char   pBuf[1024];
+  char * ppBuf = pBuf;
+  size_t vsize;
+  char * err = NULL;
+
+  rocksdb_writeoptions_t *wopt = rocksdb_writeoptions_create();
+
+  metaGetSchemaDBKey(key, uid, schemaVersion(pSchema));
+  vsize = tdEncodeSchema((void **)(&ppBuf), pSchema);
+  rocksdb_put(pMeta->pDB->schemaDb, wopt, key, SCHEMA_KEY_LEN, pBuf, vsize, &err);
+
+  rocksdb_writeoptions_destroy(wopt);
+}
+
+static void metaGetSchemaDBKey(char *key, tb_uid_t uid, int sversion) {
+  *(tb_uid_t *)key = uid;
+  *(int *)POINTER_SHIFT(key, sizeof(tb_uid_t)) = sversion;
+}
+
+static int metaSaveMapDB(SMeta *pMeta, tb_uid_t suid, tb_uid_t uid) {
+  size_t vlen;
+  char * val;
+  char * err = NULL;
+
+  rocksdb_readoptions_t *ropt = rocksdb_readoptions_create();
+  val = rocksdb_get(pMeta->pDB->mapDb, ropt, (char *)(&suid), sizeof(suid), &vlen, &err);
+  rocksdb_readoptions_destroy(ropt);
+
+  void *nval = malloc(vlen + sizeof(uid));
+  if (nval == NULL) {
+    return -1;
+  }
+
+  if (vlen) {
+    memcpy(nval, val, vlen);
+  }
+  memcpy(POINTER_SHIFT(nval, vlen), (void *)(&uid), sizeof(uid));
+
+  rocksdb_writeoptions_t *wopt = rocksdb_writeoptions_create();
+
+  rocksdb_put(pMeta->pDB->mapDb, wopt, (char *)(&suid), sizeof(suid), nval, vlen + sizeof(uid), &err);
+
+  rocksdb_writeoptions_destroy(wopt);
+
+  return 0;
+}

source/dnode/vnode/meta/src/metaIdx.c

@@ -0,0 +1,53 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include "metaDef.h"
+
+int metaOpenIdx(SMeta *pMeta) {
+  char               idxDir[128];  // TODO
+  char *             err = NULL;
+  rocksdb_options_t *options = rocksdb_options_create();
+
+  // TODO
+  sprintf(idxDir, "%s/index", pMeta->path);
+
+  if (pMeta->pCache) {
+    rocksdb_options_set_row_cache(options, pMeta->pCache);
+  }
+  rocksdb_options_set_create_if_missing(options, 1);
+
+  pMeta->pIdx = rocksdb_open(options, idxDir, &err);
+  if (pMeta->pIdx == NULL) {
+    // TODO: handle error
+    rocksdb_options_destroy(options);
+    return -1;
+  }
+
+  rocksdb_options_destroy(options);
+
+  return 0;
+}
+
+void metaCloseIdx(SMeta *pMeta) { /* TODO */
+  if (pMeta->pIdx) {
+    rocksdb_close(pMeta->pIdx);
+    pMeta->pIdx = NULL;
+  }
+}
+
+int metaSaveTableToIdx(SMeta *pMeta, const STbOptions *pTbOptions) {
+  // TODO
+  return 0;
+}

source/dnode/vnode/meta/src/metaMain.c

@@ -13,75 +13,134 @@
  * along with this program. If not, see <http://www.gnu.org/licenses/>.
  */
 
-#include "meta.h"
-#include "metaDef.h"
 #include "tcoding.h"
 
-static int metaCreateSuperTable(SMeta *pMeta, const char *tbname, const SSuperTableOpts *pSuperTableOpts);
+#include "meta.h"
-static int metaCreateChildTable(SMeta *pMeta, const char *tbname, const SChildTableOpts *pChildTableOpts);
+#include "metaDB.h"
-static int metaCreateNormalTable(SMeta *pMeta, const char *tbname, const SNormalTableOpts *pNormalTableOpts);
+#include "metaDef.h"
+#include "metaOptions.h"
 
-SMeta *metaOpen(SMetaOpts *pMetaOpts) {
+static SMeta *metaNew(const char *path, const SMetaOptions *pMetaOptions);
+static void   metaFree(SMeta *pMeta);
+static int    metaOpenImpl(SMeta *pMeta);
+static void   metaCloseImpl(SMeta *pMeta);
+
+SMeta *metaOpen(const char *path, const SMetaOptions *pMetaOptions) {
   SMeta *pMeta = NULL;
 
+  // Set default options
+  if (pMetaOptions == NULL) {
+    pMetaOptions = &defaultMetaOptions;
+  }
+
+  // Validate the options
+  if (metaValidateOptions(pMetaOptions) < 0) {
+    // TODO: deal with error
+    return NULL;
+  }
+
+  // Allocate handle
+  pMeta = metaNew(path, pMetaOptions);
+  if (pMeta == NULL) {
+    // TODO: handle error
+    return NULL;
+  }
+
+  // Create META path (TODO)
+  taosMkDir(path);
+
+  // Open meta
+  if (metaOpenImpl(pMeta) < 0) {
+    metaFree(pMeta);
+    return NULL;
+  }
+
+  return pMeta;
+}
+
+void metaClose(SMeta *pMeta) {
+  if (pMeta) {
+    metaCloseImpl(pMeta);
+    metaFree(pMeta);
+  }
+}
+
+void metaRemove(const char *path) { taosRemoveDir(path); }
+
+/* ------------------------ STATIC METHODS ------------------------ */
+static SMeta *metaNew(const char *path, const SMetaOptions *pMetaOptions) {
+  SMeta *pMeta;
+  size_t psize = strlen(path);
+
   pMeta = (SMeta *)calloc(1, sizeof(*pMeta));
   if (pMeta == NULL) {
     return NULL;
   }
 
-  // TODO: check if file exists and handle the error
+  pMeta->path = strdup(path);
-  taosMkDir("meta");
+  if (pMeta->path == NULL) {
+    metaFree(pMeta);
+    return NULL;
+  }
 
-  // Open tableDb
+  metaOptionsCopy(&(pMeta->options), pMetaOptions);
-  STkvOpts *tableDbOpts = tkvOptsCreate();
-  tkvOptsSetCreateIfMissing(tableDbOpts, 1);
-  pMeta->tableDb = tkvOpen(tableDbOpts, "meta/table_db");
-  tkvOptsDestroy(tableDbOpts);
 
-  // Open tbnameDb
-  STkvOpts *tbnameDbOpts = tkvOptsCreate();
-  tkvOptsSetCreateIfMissing(tbnameDbOpts, 1);
-  pMeta->tbnameDb = tkvOpen(tbnameDbOpts, "meta/tbname_db");
-  tkvOptsDestroy(tbnameDbOpts);
-
-  // Open schemaDb
-  STkvOpts *schemaDbOpts = tkvOptsCreate();
-  tkvOptsSetCreateIfMissing(schemaDbOpts, 1);
-  pMeta->schemaDb = tkvOpen(schemaDbOpts, "meta/schema_db");
-  tkvOptsDestroy(schemaDbOpts);
-
-  // Open tagDb
-  STkvOpts *tagDbOpts = tkvOptsCreate();
-  tkvOptsSetCreateIfMissing(tagDbOpts, 1);
-  pMeta->tagDb = tkvOpen(tagDbOpts, "meta/tag_db");
-  tkvOptsDestroy(tagDbOpts);
-
-  // Open tagIdx
-  STkvOpts *tagIdxDbOpts = tkvOptsCreate();
-  tkvOptsSetCreateIfMissing(tagIdxDbOpts, 1);
-  pMeta->tagIdx = tkvOpen(tagIdxDbOpts, "meta/tag_idx_db");
-  tkvOptsDestroy(tagIdxDbOpts);
-
-  // TODO: need to figure out how to persist the START UID
-  tableUidGeneratorInit(&(pMeta->uidGenerator), IVLD_TB_UID);
   return pMeta;
-}
+};
 
-void metaClose(SMeta *pMeta) {
+static void metaFree(SMeta *pMeta) {
   if (pMeta) {
-    tableUidGeneratorClear(&pMeta->uidGenerator);
+    tfree(pMeta->path);
-
-    tkvClose(pMeta->tagIdx);
-    tkvClose(pMeta->tagDb);
-    tkvClose(pMeta->schemaDb);
-    tkvClose(pMeta->tbnameDb);
-    tkvClose(pMeta->tableDb);
-
     free(pMeta);
   }
 }
 
-int metaCreateTable(SMeta *pMeta, const STableOpts *pTableOpts) {
+static int metaOpenImpl(SMeta *pMeta) {
+  // Open meta cache
+  if (metaOpenCache(pMeta) < 0) {
+    // TODO: handle error
+    metaCloseImpl(pMeta);
+    return -1;
+  }
+
+  // Open meta db
+  if (metaOpenDB(pMeta) < 0) {
+    // TODO: handle error
+    metaCloseImpl(pMeta);
+    return -1;
+  }
+
+  // Open meta index
+  if (metaOpenIdx(pMeta) < 0) {
+    // TODO: handle error
+    metaCloseImpl(pMeta);
+    return -1;
+  }
+
+  // Open meta table uid generator
+  if (metaOpenUidGnrt(pMeta) < 0) {
+    // TODO: handle error
+    metaCloseImpl(pMeta);
+    return -1;
+  }
+
+  return 0;
+}
+
+static void metaCloseImpl(SMeta *pMeta) {
+  metaCloseUidGnrt(pMeta);
+  metaCloseIdx(pMeta);
+  metaCloseDB(pMeta);
+  metaCloseCache(pMeta);
+}
+
+// OLD -------------------------------------------------------------------
+#if 0
+static int    metaCreateSuperTable(SMeta *pMeta, const char *tbname, const SSuperTableOpts *pSuperTableOpts);
+static int    metaCreateChildTable(SMeta *pMeta, const char *tbname, const SChildTableOpts *pChildTableOpts);
+static int    metaCreateNormalTable(SMeta *pMeta, const char *tbname, const SNormalTableOpts *pNormalTableOpts);
+
+int metaCreateTable(SMeta *pMeta, const STableOptions *pTableOpts) {
   size_t vallen;
   char * pUid;
 
@@ -107,7 +166,6 @@ int metaCreateTable(SMeta *pMeta, const STableOpts *pTableOpts) {
   return 0;
 }
 
-/* ------------------------ STATIC METHODS ------------------------ */
 static int metaCreateSuperTable(SMeta *pMeta, const char *tbname, const SSuperTableOpts *pSuperTableOpts) {
   size_t vallen;
   size_t keylen;
@@ -213,13 +271,13 @@ static int metaCreateNormalTable(SMeta *pMeta, const char *tbname, const SNormal
   return 0;
 }
 
-void metaNormalTableOptsInit(STableOpts *pTableOpts, const char *name, const STSchema *pSchema) {
+void metaNormalTableOptsInit(STableOptions *pTableOpts, const char *name, const STSchema *pSchema) {
   pTableOpts->type = META_NORMAL_TABLE;
   pTableOpts->name = strdup(name);
   pTableOpts->normalOpts.pSchema = tdDupSchema(pSchema);
 }
 
-void metaSuperTableOptsInit(STableOpts *pTableOpts, const char *name, tb_uid_t uid, const STSchema *pSchema,
+void metaSuperTableOptsInit(STableOptions *pTableOpts, const char *name, tb_uid_t uid, const STSchema *pSchema,
                             const STSchema *pTagSchema) {
   pTableOpts->type = META_SUPER_TABLE;
   pTableOpts->name = strdup(name);
@@ -228,14 +286,14 @@ void metaSuperTableOptsInit(STableOpts *pTableOpts, const char *name, tb_uid_t u
   pTableOpts->superOpts.pTagSchema = tdDupSchema(pTagSchema);
 }
 
-void metaChildTableOptsInit(STableOpts *pTableOpts, const char *name, tb_uid_t suid, const SKVRow tags) {
+void metaChildTableOptsInit(STableOptions *pTableOpts, const char *name, tb_uid_t suid, const SKVRow tags) {
   pTableOpts->type = META_CHILD_TABLE;
   pTableOpts->name = strdup(name);
   pTableOpts->childOpts.suid = suid;
   pTableOpts->childOpts.tags = tdKVRowDup(tags);
 }
 
-void metaTableOptsClear(STableOpts *pTableOpts) {
+void metaTableOptsClear(STableOptions *pTableOpts) {
   switch (pTableOpts->type) {
     case META_NORMAL_TABLE:
       tfree(pTableOpts->name);
@@ -258,3 +316,5 @@ void metaTableOptsClear(STableOpts *pTableOpts) {
 }
 
 void metaDestroy(const char *path) { taosRemoveDir(path); }
+
+#endif

source/dnode/vnode/meta/src/metaOptions.c

@@ -0,0 +1,34 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include "metaDef.h"
+
+const SMetaOptions defaultMetaOptions = {.lruCacheSize = 0};
+
+/* ------------------------ EXPOSED METHODS ------------------------ */
+void metaOptionsInit(SMetaOptions *pMetaOptions) { metaOptionsCopy(pMetaOptions, &defaultMetaOptions); }
+
+void metaOptionsClear(SMetaOptions *pMetaOptions) {
+  // TODO
+}
+
+int metaValidateOptions(const SMetaOptions *pMetaOptions) {
+  // TODO
+  return 0;
+}
+
+void metaOptionsCopy(SMetaOptions *pDest, const SMetaOptions *pSrc) { memcpy(pDest, pSrc, sizeof(*pSrc)); }
+
+/* ------------------------ STATIC METHODS ------------------------ */

source/dnode/vnode/meta/src/metaQuery.c

@@ -0,0 +1,14 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */

source/dnode/vnode/meta/src/metaTable.c

@@ -0,0 +1,52 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include "metaDef.h"
+
+int metaCreateTable(SMeta *pMeta, const STbOptions *pTbOptions) {
+  // Validate the tbOptions
+  if (metaValidateTbOptions(pMeta, pTbOptions) < 0) {
+    // TODO: handle error
+    return -1;
+  }
+
+  // TODO: add atomicity
+
+  if (metaSaveTableToDB(pMeta, pTbOptions) < 0) {
+    // TODO: handle error
+    return -1;
+  }
+
+  if (metaSaveTableToIdx(pMeta, pTbOptions) < 0) {
+    // TODO: handle error
+    return -1;
+  }
+
+  return 0;
+}
+
+int metaDropTable(SMeta *pMeta, tb_uid_t uid) {
+  if (metaRemoveTableFromIdx(pMeta, uid) < 0) {
+    // TODO: handle error
+    return -1;
+  }
+
+  if (metaRemoveTableFromIdx(pMeta, uid) < 0) {
+    // TODO
+    return -1;
+  }
+
+  return 0;
+}

source/dnode/vnode/meta/src/metaTbOptions.c

@@ -0,0 +1,49 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#include "metaDef.h"
+#include "tcoding.h"
+
+int metaValidateTbOptions(SMeta *pMeta, const STbOptions *pTbOptions) {
+  // TODO
+  return 0;
+}
+
+size_t metaEncodeTbObjFromTbOptions(const STbOptions *pTbOptions, void *pBuf, size_t bsize) {
+  void **ppBuf = &pBuf;
+  int    tlen = 0;
+
+  tlen += taosEncodeFixedU8(ppBuf, pTbOptions->type);
+  tlen += taosEncodeString(ppBuf, pTbOptions->name);
+  tlen += taosEncodeFixedU32(ppBuf, pTbOptions->ttl);
+
+  switch (pTbOptions->type) {
+    case META_SUPER_TABLE:
+      tlen += taosEncodeFixedU64(ppBuf, pTbOptions->stbOptions.uid);
+      tlen += tdEncodeSchema(ppBuf, pTbOptions->stbOptions.pTagSchema);
+      // TODO: encode schema version array
+      break;
+    case META_CHILD_TABLE:
+      tlen += taosEncodeFixedU64(ppBuf, pTbOptions->ctbOptions.suid);
+      break;
+    case META_NORMAL_TABLE:
+      // TODO: encode schema version array
+      break;
+    default:
+      break;
+  }
+
+  return tlen;
+}

source/dnode/vnode/meta/src/metaTbTag.c

@@ -0,0 +1,14 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */

source/dnode/vnode/meta/src/metaTbUid.c

@@ -13,14 +13,18 @@
  * along with this program. If not, see <http://www.gnu.org/licenses/>.
  */
 
-#include "metaUid.h"
+#include "meta.h"
+#include "metaDef.h"
 
-tb_uid_t generateUid(STableUidGenerator *pGen) {
+int metaOpenUidGnrt(SMeta *pMeta) {
-  // Generate a new table UID
+  // Init a generator
-  return ++(pGen->nextUid);
+  pMeta->uidGnrt.nextUid = IVLD_TB_UID;
+  return 0;
 }
 
-void tableUidGeneratorInit(STableUidGenerator *pGen, tb_uid_t suid) {
+void metaCloseUidGnrt(SMeta *pMeta) { /* TODO */ }
-  // Init a generator
+
-  pGen->nextUid = suid;
+tb_uid_t metaGenerateUid(SMeta *pMeta) {
+  // Generate a new table UID
+  return ++(pMeta->uidGnrt.nextUid);
 }

source/dnode/vnode/meta/test/CMakeLists.txt

@@ -1,24 +1,24 @@
-add_executable(metaTest "")
+# add_executable(metaTest "")
-target_sources(metaTest
+# target_sources(metaTest
-    PRIVATE
+#     PRIVATE
-    "../src/metaMain.c"
+#     "../src/metaMain.c"
-    "../src/metaUid.c"
+#     "../src/metaUid.c"
-    "metaTests.cpp"
+#     "metaTests.cpp"
-)
+# )
-target_include_directories(metaTest 
+# target_include_directories(metaTest 
-    PUBLIC
+#     PUBLIC
-    "${CMAKE_SOURCE_DIR}/include/dnode/vnode/meta"
+#     "${CMAKE_SOURCE_DIR}/include/server/vnode/meta"
-    "${CMAKE_CURRENT_SOURCE_DIR}/../inc"
+#     "${CMAKE_CURRENT_SOURCE_DIR}/../inc"
-)
+# )
-target_link_libraries(metaTest
+# target_link_libraries(metaTest
-    os
+#     os
-    util
+#     util
-    common
+#     common
-    gtest_main
+#     gtest_main
-    tkv
+#     tkv
-)
+# )
-enable_testing()
+# enable_testing()
-add_test(
+# add_test(
-    NAME meta_test
+#     NAME meta_test
-    COMMAND metaTest
+#     COMMAND metaTest
-)
+# )

source/dnode/vnode/meta/test/metaAPITest.cpp

@@ -1,3 +1,4 @@
+#if 0
 #include <gtest/gtest.h>
 #include <string.h>
 #include <iostream>
@@ -38,7 +39,7 @@ static SKVRow metaGetSimpleTags() {
 
 TEST(MetaTest, DISABLED_meta_create_1m_normal_tables_test) {
   // Open Meta
-  SMeta *meta = metaOpen(NULL);
+  SMeta *meta = metaOpen(NULL, NULL);
   std::cout << "Meta is opened!" << std::endl;
 
   // Create 1000000 normal tables
@@ -100,4 +101,5 @@ TEST(MetaTest, meta_create_1m_child_tables_test) {
   // Destroy Meta
   metaDestroy("meta");
   std::cout << "Meta is destroyed!" << std::endl;
-}
+}
+#endif

source/dnode/vnode/tq/CMakeLists.txt

@@ -12,3 +12,7 @@ target_link_libraries(
     PUBLIC os
     PUBLIC util
 )
+
+if(${BUILD_TEST})
+    add_subdirectory(test)
+endif(${BUILD_TEST})

source/dnode/vnode/tq/inc/tqMetaStore.h

@@ -19,23 +19,42 @@
 #include "os.h"
 #include "tq.h"
 
-#define TQ_BUCKET_SIZE 0xFF
-#define TQ_PAGE_SIZE 4096
-//key + offset + size
-#define TQ_IDX_ENTRY_SIZE 24
-
-inline static int TqMaxEntryOnePage() { //170
-  return TQ_PAGE_SIZE / TQ_IDX_ENTRY_SIZE;
-}
-
-inline static int TqEmptyTail() { //16
-  return TQ_PAGE_SIZE - TqMaxEntryOnePage();
-}
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
+#define TQ_BUCKET_MASK 0xFF
+#define TQ_BUCKET_SIZE 256
+
+#define TQ_PAGE_SIZE 4096
+//key + offset + size
+#define TQ_IDX_SIZE 24
+//4096 / 24
+#define TQ_MAX_IDX_ONE_PAGE 170
+//24 * 170
+#define TQ_IDX_PAGE_BODY_SIZE 4080
+//4096 - 4080
+#define TQ_IDX_PAGE_HEAD_SIZE 16
+
+#define TQ_ACTION_CONST      0
+#define TQ_ACTION_INUSE      1
+#define TQ_ACTION_INUSE_CONT 2
+#define TQ_ACTION_INTXN      3
+
+#define TQ_SVER              0
+
+static const int8_t TQ_CONST_DELETE = TQ_ACTION_CONST;
+#define TQ_DELETE_TOKEN  (void*)&TQ_CONST_DELETE
+
+typedef struct TqSerializedHead {
+  int16_t ver;
+  int16_t action;
+  int32_t checksum;
+  int64_t ssize;
+  char    content[];
+} TqSerializedHead;
+
 typedef struct TqMetaHandle {
   int64_t key;
   int64_t offset;
@@ -59,30 +78,31 @@ typedef struct TqMetaStore {
   TqMetaList* unpersistHead;
   int fileFd; //TODO:temporaral use, to be replaced by unified tfile
   int idxFd;  //TODO:temporaral use, to be replaced by unified tfile
-  int (*serializer)(TqGroupHandle*, void**);
+  char* dirPath;
-  const void* (*deserializer)(const void*, TqGroupHandle*);
+  int (*serializer)(const void* pObj, TqSerializedHead** ppHead);
+  const void* (*deserializer)(const TqSerializedHead* pHead, void** ppObj);
   void  (*deleter)(void*);
 } TqMetaStore;
 
 TqMetaStore*  tqStoreOpen(const char* path,
-    int serializer(TqGroupHandle*, void**),
+    int serializer(const void* pObj, TqSerializedHead** ppHead),
-    const void* deserializer(const void*, TqGroupHandle*),
+    const void* deserializer(const TqSerializedHead* pHead, void** ppObj),
-    void deleter(void*));
+    void deleter(void* pObj));
 int32_t       tqStoreClose(TqMetaStore*);
 //int32_t       tqStoreDelete(TqMetaStore*);
 //int32_t       TqStoreCommitAll(TqMetaStore*);
 int32_t       tqStorePersist(TqMetaStore*);
+//clean deleted idx and data from persistent file
+int32_t       tqStoreCompact(TqMetaStore*);
 
-TqMetaHandle* tqHandleGet(TqMetaStore*, int64_t key);
+void*   tqHandleGet(TqMetaStore*, int64_t key);
-int32_t       tqHandlePut(TqMetaStore*, int64_t key, void* value);
+int32_t tqHandleMovePut(TqMetaStore*, int64_t key, void* value);
-//do commit
+int32_t tqHandleCopyPut(TqMetaStore*, int64_t key, void* value, size_t vsize);
-int32_t       tqHandleCommit(TqMetaStore*, int64_t key);
+//delete committed kv pair
-//delete uncommitted
+//notice that a delete action still needs to be committed
-int32_t       tqHandleAbort(TqMetaStore*, int64_t key);
+int32_t tqHandleDel(TqMetaStore*, int64_t key);
-//delete committed
+int32_t tqHandleCommit(TqMetaStore*, int64_t key);
-int32_t       tqHandleDel(TqMetaStore*, int64_t key);
+int32_t tqHandleAbort(TqMetaStore*, int64_t key);
-//delete both committed and uncommitted
-int32_t       tqHandleClear(TqMetaStore*, int64_t key);
 
 #ifdef __cplusplus
 }

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

@@ -22,6 +22,9 @@
 //
 //handle management message
 //
+
+int tqGetgHandleSSize(const TqGroupHandle *gHandle);
+
 static int tqProtoCheck(TmqMsgHead *pMsg) {
   return pMsg->protoVer == 0;
 }

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

@@ -14,6 +14,7 @@
  */
 #include "tqMetaStore.h"
 //TODO:replace by an abstract file layer
+#include "osDir.h"
 #include <fcntl.h>
 #include <string.h>
 #include <unistd.h>
@@ -22,32 +23,80 @@
 #define TQ_IDX_NAME  "tq.idx"
 
 
-static int32_t       tqHandlePutCommitted(TqMetaStore*, int64_t key, void* value);
+static int32_t tqHandlePutCommitted(TqMetaStore*, int64_t key, void* value);
-static TqMetaHandle* tqHandleGetUncommitted(TqMetaStore*, int64_t key);
+static void*   tqHandleGetUncommitted(TqMetaStore*, int64_t key);
 
-typedef struct TqMetaPageBuf {
+static inline void tqLinkUnpersist(TqMetaStore *pMeta, TqMetaList* pNode) {
-  int16_t offset;
+  if(pNode->unpersistNext == NULL) {
-  char buffer[TQ_PAGE_SIZE];
+    pNode->unpersistNext = pMeta->unpersistHead->unpersistNext;
-} TqMetaPageBuf;
+    pNode->unpersistPrev = pMeta->unpersistHead;
+    pMeta->unpersistHead->unpersistNext->unpersistPrev = pNode;
+    pMeta->unpersistHead->unpersistNext = pNode;
+  }
+}
+
+static inline int tqSeekLastPage(int fd) {
+  int offset = lseek(fd, 0, SEEK_END);
+  int pageNo = offset / TQ_PAGE_SIZE;
+  int curPageOffset = pageNo * TQ_PAGE_SIZE;
+  return lseek(fd, curPageOffset, SEEK_SET);
+}
+
+//TODO: the struct is tightly coupled with index entry
+typedef struct TqIdxPageHead {
+  int16_t writeOffset;
+  int8_t  unused[14];
+} TqIdxPageHead;
+
+typedef struct TqIdxPageBuf {
+  TqIdxPageHead head;
+  char buffer[TQ_IDX_PAGE_BODY_SIZE];
+} TqIdxPageBuf;
+
+static inline int tqReadLastPage(int fd, TqIdxPageBuf* pBuf) {
+  int offset = tqSeekLastPage(fd);
+  int nBytes;
+  if((nBytes = read(fd, pBuf, TQ_PAGE_SIZE)) == -1) {
+    return -1;
+  }
+  if(nBytes == 0) {
+    memset(pBuf, 0, TQ_PAGE_SIZE);
+    pBuf->head.writeOffset = TQ_IDX_PAGE_HEAD_SIZE;
+  }
+  ASSERT(nBytes == 0 || nBytes == pBuf->head.writeOffset);
+
+  return lseek(fd, offset, SEEK_SET);
+}
 
 TqMetaStore* tqStoreOpen(const char* path,
-    int serializer(TqGroupHandle*, void**),
+    int serializer(const void* pObj, TqSerializedHead** ppHead),
-    const void* deserializer(const void*, TqGroupHandle*),
+    const void* deserializer(const TqSerializedHead* pHead, void** ppObj),
-    void deleter(void*)) {
+    void deleter(void* pObj)) {
   TqMetaStore* pMeta = malloc(sizeof(TqMetaStore)); 
   if(pMeta == NULL) {
     //close
     return NULL;
   }
+  memset(pMeta, 0, sizeof(TqMetaStore));
 
   //concat data file name and index file name
   size_t pathLen = strlen(path);
+  pMeta->dirPath = malloc(pathLen+1);
+  if(pMeta->dirPath != NULL) {
+    //TODO: memory insufficient
+  }
+  strcpy(pMeta->dirPath, path);
+  
   char name[pathLen+10];
 
   strcpy(name, path);
+  if(!taosDirExist(name) && !taosMkDir(name)) {
+    ASSERT(false);
+  }
   strcat(name, "/" TQ_IDX_NAME);
-  int idxFd = open(name, O_WRONLY | O_CREAT | O_EXCL, 0755);
+  int idxFd = open(name, O_RDWR | O_CREAT, 0755);
   if(idxFd < 0) {
+    ASSERT(false);
     //close file
     //free memory
     return NULL;
@@ -56,17 +105,24 @@ TqMetaStore* tqStoreOpen(const char* path,
   pMeta->idxFd = idxFd;
   pMeta->unpersistHead = malloc(sizeof(TqMetaList));
   if(pMeta->unpersistHead == NULL) {
+    ASSERT(false);
     //close file
     //free memory
     return NULL;
   }
+  memset(pMeta->unpersistHead, 0, sizeof(TqMetaList));
+  pMeta->unpersistHead->unpersistNext
+    = pMeta->unpersistHead->unpersistPrev
+    = pMeta->unpersistHead;
 
   strcpy(name, path);
   strcat(name, "/" TQ_META_NAME);
-  int fileFd = open(name, O_WRONLY | O_CREAT | O_EXCL, 0755);
+  int fileFd = open(name, O_RDWR | O_CREAT, 0755);
-  if(fileFd < 0) return NULL;
+  if(fileFd < 0){
+    ASSERT(false);
+    return NULL;
+  }
 
-  memset(pMeta, 0, sizeof(TqMetaStore));
   pMeta->fileFd = fileFd;
   
   pMeta->serializer = serializer;
@@ -74,23 +130,109 @@ TqMetaStore* tqStoreOpen(const char* path,
   pMeta->deleter = deleter;
 
   //read idx file and load into memory
-  char readBuf[TQ_PAGE_SIZE];
+  TqIdxPageBuf idxBuf;
-  int readSize;
+  TqSerializedHead* serializedObj = malloc(TQ_PAGE_SIZE);
-  while((readSize = read(idxFd, readBuf, TQ_PAGE_SIZE)) != -1) {
+  if(serializedObj == NULL) {
+    //TODO:memory insufficient
+  }
+  int idxRead;
+  int allocated = TQ_PAGE_SIZE;
+  bool readEnd = false;
+  while((idxRead = read(idxFd, &idxBuf, TQ_PAGE_SIZE))) {
+    if(idxRead == -1) {
+      //TODO: handle error
+      ASSERT(false);
+    }
+    ASSERT(idxBuf.head.writeOffset == idxRead);
     //loop read every entry
-    for(int i = 0; i < readSize; i += TQ_IDX_ENTRY_SIZE) {
+    for(int i = 0; i < idxBuf.head.writeOffset - TQ_IDX_PAGE_HEAD_SIZE; i += TQ_IDX_SIZE) {
-      TqMetaList *pNode = malloc(sizeof(TqMetaHandle));
+      TqMetaList *pNode = malloc(sizeof(TqMetaList));
-      memset(pNode, 0, sizeof(TqMetaList));
       if(pNode == NULL) {
         //TODO: free memory and return error
       }
-      memcpy(&pNode->handle, &readBuf[i], TQ_IDX_ENTRY_SIZE);
+      memset(pNode, 0, sizeof(TqMetaList));
-      int bucketKey = pNode->handle.key & TQ_BUCKET_SIZE;
+      memcpy(&pNode->handle, &idxBuf.buffer[i], TQ_IDX_SIZE);
-      pNode->next = pMeta->bucket[bucketKey];
+
-      pMeta->bucket[bucketKey] = pNode;
+      lseek(fileFd, pNode->handle.offset, SEEK_SET);
+      if(allocated < pNode->handle.serializedSize) {
+        void *ptr = realloc(serializedObj, pNode->handle.serializedSize);
+        if(ptr == NULL) {
+          //TODO: memory insufficient
+        }
+        serializedObj = ptr;
+        allocated = pNode->handle.serializedSize;
+      }
+      serializedObj->ssize = pNode->handle.serializedSize;
+      if(read(fileFd, serializedObj, pNode->handle.serializedSize) != pNode->handle.serializedSize) {
+        //TODO: read error
+      }
+      if(serializedObj->action == TQ_ACTION_INUSE) {
+        if(serializedObj->ssize != sizeof(TqSerializedHead)) {
+          pMeta->deserializer(serializedObj, &pNode->handle.valueInUse);
+        } else {
+          pNode->handle.valueInUse = TQ_DELETE_TOKEN;
+        }
+      } else if(serializedObj->action == TQ_ACTION_INTXN) {
+        if(serializedObj->ssize != sizeof(TqSerializedHead)) {
+          pMeta->deserializer(serializedObj, &pNode->handle.valueInTxn);
+        } else {
+          pNode->handle.valueInTxn = TQ_DELETE_TOKEN;
+        }
+      } else if(serializedObj->action == TQ_ACTION_INUSE_CONT) {
+        if(serializedObj->ssize != sizeof(TqSerializedHead)) {
+          pMeta->deserializer(serializedObj, &pNode->handle.valueInUse);
+        } else {
+          pNode->handle.valueInUse = TQ_DELETE_TOKEN;
+        }
+        TqSerializedHead* ptr = POINTER_SHIFT(serializedObj, serializedObj->ssize);
+        if(ptr->ssize != sizeof(TqSerializedHead)) {
+          pMeta->deserializer(ptr, &pNode->handle.valueInTxn);
+        } else {
+          pNode->handle.valueInTxn = TQ_DELETE_TOKEN;
+        }
+      } else {
+        ASSERT(0);
+      }
+
+      //put into list
+      int bucketKey = pNode->handle.key & TQ_BUCKET_MASK;
+      TqMetaList* pBucketNode = pMeta->bucket[bucketKey];
+      if(pBucketNode == NULL) {
+        pMeta->bucket[bucketKey] = pNode;
+      } else if(pBucketNode->handle.key == pNode->handle.key) {
+        pNode->next = pBucketNode->next;
+        pMeta->bucket[bucketKey] = pNode;
+      } else {
+        while(pBucketNode->next &&
+            pBucketNode->next->handle.key != pNode->handle.key) {
+          pBucketNode = pBucketNode->next; 
+        }
+        if(pBucketNode->next) {
+          ASSERT(pBucketNode->next->handle.key == pNode->handle.key);
+          TqMetaList *pNodeFound = pBucketNode->next;
+          pNode->next = pNodeFound->next;
+          pBucketNode->next = pNode;
+          pBucketNode = pNodeFound;
+        } else {
+          pNode->next = pMeta->bucket[bucketKey];
+          pMeta->bucket[bucketKey] = pNode;
+          pBucketNode = NULL;
+        }
+      }
+      if(pBucketNode) {
+        if(pBucketNode->handle.valueInUse
+            && pBucketNode->handle.valueInUse != TQ_DELETE_TOKEN) {
+          pMeta->deleter(pBucketNode->handle.valueInUse);
+        }
+        if(pBucketNode->handle.valueInTxn
+            && pBucketNode->handle.valueInTxn != TQ_DELETE_TOKEN) {
+          pMeta->deleter(pBucketNode->handle.valueInTxn);
+        }
+        free(pBucketNode);
+      }
     }
   }
-
+  free(serializedObj);
   return pMeta;
 }
 
@@ -102,107 +244,167 @@ int32_t tqStoreClose(TqMetaStore* pMeta) {
   close(pMeta->idxFd);
   //free memory
   for(int i = 0; i < TQ_BUCKET_SIZE; i++) {
-    TqMetaList* node = pMeta->bucket[i];
+    TqMetaList* pNode = pMeta->bucket[i];
-    pMeta->bucket[i] = NULL;
+    while(pNode) {
-    while(node) {
+      ASSERT(pNode->unpersistNext == NULL);
-      ASSERT(node->unpersistNext == NULL);
+      ASSERT(pNode->unpersistPrev == NULL);
-      ASSERT(node->unpersistPrev == NULL);
+      if(pNode->handle.valueInTxn
-      if(node->handle.valueInTxn) {
+          && pNode->handle.valueInTxn != TQ_DELETE_TOKEN) {
-        pMeta->deleter(node->handle.valueInTxn);
+        pMeta->deleter(pNode->handle.valueInTxn);
       }
-      if(node->handle.valueInUse) {
+      if(pNode->handle.valueInUse
-        pMeta->deleter(node->handle.valueInUse);
+          && pNode->handle.valueInUse != TQ_DELETE_TOKEN) {
+        pMeta->deleter(pNode->handle.valueInUse);
       }
-      TqMetaList* next = node->next;
+      TqMetaList* next = pNode->next;
-      free(node);
+      free(pNode);
-      node = next;
+      pNode = next;
     }
   }
+  free(pMeta->dirPath);
+  free(pMeta->unpersistHead);
   free(pMeta);
   return 0;
 }
 
 int32_t tqStoreDelete(TqMetaStore* pMeta) {
-  //close file
+  close(pMeta->fileFd);
-  //delete file
+  close(pMeta->idxFd);
   //free memory
+  for(int i = 0; i < TQ_BUCKET_SIZE; i++) {
+    TqMetaList* pNode = pMeta->bucket[i];
+    pMeta->bucket[i] = NULL;
+    while(pNode) {
+      if(pNode->handle.valueInTxn
+          && pNode->handle.valueInTxn != TQ_DELETE_TOKEN) {
+        pMeta->deleter(pNode->handle.valueInTxn);
+      }
+      if(pNode->handle.valueInUse
+          && pNode->handle.valueInUse != TQ_DELETE_TOKEN) {
+        pMeta->deleter(pNode->handle.valueInUse);
+      }
+      TqMetaList* next = pNode->next;
+      free(pNode);
+      pNode = next;
+    }
+  }
+  free(pMeta->unpersistHead);
+  taosRemoveDir(pMeta->dirPath);
+  free(pMeta->dirPath);
+  free(pMeta);
   return 0;
 }
 
 //TODO: wrap in tfile
 int32_t tqStorePersist(TqMetaStore* pMeta) {
-  char writeBuf[TQ_PAGE_SIZE];
+  TqIdxPageBuf idxBuf;
-  int64_t* bufPtr = (int64_t*)writeBuf;
+  int64_t* bufPtr = (int64_t*)idxBuf.buffer;
   TqMetaList *pHead = pMeta->unpersistHead;
   TqMetaList *pNode = pHead->unpersistNext;
-  while(pHead != pNode) {
+  TqSerializedHead *pSHead = malloc(sizeof(TqSerializedHead));
-    if(pNode->handle.valueInUse == NULL) {
+  if(pSHead == NULL) {
-      //put delete token in data file
+    //TODO: memory error
-      uint32_t delete = TQ_ACTION_DELETE;
+    return -1;
-      int nBytes = write(pMeta->fileFd, &delete, sizeof(uint32_t));
+  }
-      ASSERT(nBytes == sizeof(uint32_t));
+  pSHead->ver = TQ_SVER;
+  pSHead->checksum = 0;
+  pSHead->ssize = sizeof(TqSerializedHead);
+  int allocatedSize = sizeof(TqSerializedHead);
+  int offset = lseek(pMeta->fileFd, 0, SEEK_CUR);
 
-      //remove from list
+  tqReadLastPage(pMeta->idxFd, &idxBuf);
-      int bucketKey = pNode->handle.key & TQ_BUCKET_SIZE;
+
+  if(idxBuf.head.writeOffset == TQ_PAGE_SIZE) {
+    lseek(pMeta->idxFd, 0, SEEK_END);
+    memset(&idxBuf, 0, TQ_PAGE_SIZE);
+    idxBuf.head.writeOffset = TQ_IDX_PAGE_HEAD_SIZE;
+  } else {
+    bufPtr = POINTER_SHIFT(&idxBuf, idxBuf.head.writeOffset);
+  }
+
+  while(pHead != pNode) {
+    int nBytes = 0;
+
+    if(pNode->handle.valueInUse) {
+      if(pNode->handle.valueInTxn) {
+        pSHead->action = TQ_ACTION_INUSE_CONT;
+      } else {
+        pSHead->action = TQ_ACTION_INUSE;
+      }
+
+      if(pNode->handle.valueInUse == TQ_DELETE_TOKEN) {
+        pSHead->ssize = sizeof(TqSerializedHead);
+      } else {
+        pMeta->serializer(pNode->handle.valueInUse, &pSHead);
+      }
+      nBytes = write(pMeta->fileFd, pSHead, pSHead->ssize);
+      ASSERT(nBytes == pSHead->ssize);
+    }
+
+    if(pNode->handle.valueInTxn) {
+      pSHead->action = TQ_ACTION_INTXN;
+      if(pNode->handle.valueInTxn == TQ_DELETE_TOKEN) {
+        pSHead->ssize = sizeof(TqSerializedHead);
+      } else {
+        pMeta->serializer(pNode->handle.valueInTxn, &pSHead);
+      }
+      int nBytesTxn = write(pMeta->fileFd, pSHead, pSHead->ssize);
+      ASSERT(nBytesTxn == pSHead->ssize);
+      nBytes += nBytesTxn;
+    }
+    pNode->handle.offset = offset;
+    offset += nBytes;
+
+    //write idx file
+    //TODO: endian check and convert
+    *(bufPtr++) = pNode->handle.key;
+    *(bufPtr++) = pNode->handle.offset;
+    *(bufPtr++) = (int64_t)nBytes;
+    idxBuf.head.writeOffset += TQ_IDX_SIZE;
+
+    if(idxBuf.head.writeOffset >= TQ_PAGE_SIZE) {
+      nBytes = write(pMeta->idxFd, &idxBuf, TQ_PAGE_SIZE);
+      //TODO: handle error with tfile
+      ASSERT(nBytes == TQ_PAGE_SIZE);
+      memset(&idxBuf, 0, TQ_PAGE_SIZE);
+      idxBuf.head.writeOffset = TQ_IDX_PAGE_HEAD_SIZE;
+      bufPtr = (int64_t*)&idxBuf.buffer;
+    }
+    //remove from unpersist list
+    pHead->unpersistNext = pNode->unpersistNext;
+    pHead->unpersistNext->unpersistPrev = pHead;
+    pNode->unpersistPrev = pNode->unpersistNext = NULL;
+    pNode = pHead->unpersistNext;
+
+    //remove from bucket
+    if(pNode->handle.valueInUse == TQ_DELETE_TOKEN &&
+        pNode->handle.valueInTxn == NULL
+        ) {
+      int bucketKey = pNode->handle.key & TQ_BUCKET_MASK;
       TqMetaList* pBucketHead = pMeta->bucket[bucketKey];
       if(pBucketHead == pNode) {
-        pMeta->bucket[bucketKey] = pBucketHead->next;
+        pMeta->bucket[bucketKey] = pNode->next;
       } else {
         TqMetaList* pBucketNode = pBucketHead;
         while(pBucketNode->next != NULL
             && pBucketNode->next != pNode) {
           pBucketNode = pBucketNode->next; 
         }
-        if(pBucketNode->next != NULL) {
+        //impossible for pBucket->next == NULL
-          ASSERT(pBucketNode->next == pNode);
+        ASSERT(pBucketNode->next == pNode);
-          pBucketNode->next = pNode->next;
+        pBucketNode->next = pNode->next;
-          if(pNode->handle.valueInUse) {
-            pMeta->deleter(pNode->handle.valueInUse);
-          }
-          free(pNode);
-        }
       }
+      free(pNode);
     }
-    //serialize
-    void* pBytes = NULL;
-    int sz = pMeta->serializer(pNode->handle.valueInUse, &pBytes);
-    ASSERT(pBytes != NULL);
-    //get current offset
-    //append data
-    int64_t offset = lseek(pMeta->fileFd, 0, SEEK_CUR);
-    int nBytes = write(pMeta->fileFd, pBytes, sz);
-    //TODO: handle error in tfile
-    ASSERT(nBytes == sz);
-
-    pNode->handle.offset = offset;
-    pNode->handle.serializedSize = sz;
-
-    //write idx
-    //TODO: endian check and convert
-    *(bufPtr++) = pNode->handle.key;
-    *(bufPtr++) = pNode->handle.offset;
-    *(bufPtr++) = (int64_t)sz;
-    if((char*)(bufPtr + 3) > writeBuf + TQ_PAGE_SIZE) {
-      nBytes = write(pMeta->idxFd, writeBuf, sizeof(writeBuf));
-      //TODO: handle error in tfile
-      ASSERT(nBytes == sizeof(writeBuf));
-      memset(writeBuf, 0, TQ_PAGE_SIZE);
-      bufPtr = (int64_t*)writeBuf;
-    }
-
-    //remove from unpersist list
-    pHead->unpersistNext = pNode->unpersistNext;
-    pHead->unpersistNext->unpersistPrev = pHead;
-
-    pNode->unpersistPrev = pNode->unpersistNext = NULL;
-    pNode = pHead->unpersistNext;
   }
+
   //write left bytes
-  if((char*)bufPtr != writeBuf) {
+  free(pSHead);
-    int used = (char*)bufPtr - writeBuf;
+  //TODO: write new version in tfile
-    int nBytes = write(pMeta->idxFd, writeBuf, used);
+  if((char*)bufPtr != idxBuf.buffer) {
+    int nBytes = write(pMeta->idxFd, &idxBuf, idxBuf.head.writeOffset);
     //TODO: handle error in tfile
-    ASSERT(nBytes == used);
+    ASSERT(nBytes == idxBuf.head.writeOffset);
   }
   //TODO: using fsync in tfile
   fsync(pMeta->idxFd);
@@ -211,12 +413,15 @@ int32_t tqStorePersist(TqMetaStore* pMeta) {
 }
 
 static int32_t tqHandlePutCommitted(TqMetaStore* pMeta, int64_t key, void* value) {
-  int64_t bucketKey = key & TQ_BUCKET_SIZE;
+  int64_t bucketKey = key & TQ_BUCKET_MASK;
   TqMetaList* pNode = pMeta->bucket[bucketKey];
   while(pNode) {
     if(pNode->handle.key == key) {
       //TODO: think about thread safety
-      pMeta->deleter(pNode->handle.valueInUse);
+      if(pNode->handle.valueInUse
+          && pNode->handle.valueInUse != TQ_DELETE_TOKEN) {
+        pMeta->deleter(pNode->handle.valueInUse);
+      }
       //change pointer ownership
       pNode->handle.valueInUse = value;
       return 0;
@@ -240,13 +445,14 @@ static int32_t tqHandlePutCommitted(TqMetaStore* pMeta, int64_t key, void* value
   return 0;
 }
 
-TqMetaHandle* tqHandleGet(TqMetaStore* pMeta, int64_t key) {
+void* tqHandleGet(TqMetaStore* pMeta, int64_t key) {
-  int64_t bucketKey = key & TQ_BUCKET_SIZE;
+  int64_t bucketKey = key & TQ_BUCKET_MASK;
   TqMetaList* pNode = pMeta->bucket[bucketKey];
   while(pNode) {
     if(pNode->handle.key == key) {
-      if(pNode->handle.valueInUse != NULL) {
+      if(pNode->handle.valueInUse != NULL
-        return &pNode->handle;
+          && pNode->handle.valueInUse != TQ_DELETE_TOKEN) {
+        return pNode->handle.valueInUse;
       } else {
         return NULL;
       }
@@ -257,15 +463,19 @@ TqMetaHandle* tqHandleGet(TqMetaStore* pMeta, int64_t key) {
   return NULL;
 }
 
-int32_t tqHandlePut(TqMetaStore* pMeta, int64_t key, void* value) {
+int32_t tqHandleMovePut(TqMetaStore* pMeta, int64_t key, void* value) {
-  int64_t bucketKey = key & TQ_BUCKET_SIZE;
+  int64_t bucketKey = key & TQ_BUCKET_MASK;
   TqMetaList* pNode = pMeta->bucket[bucketKey];
   while(pNode) {
     if(pNode->handle.key == key) {
       //TODO: think about thread safety
-      pMeta->deleter(pNode->handle.valueInTxn);
+      if(pNode->handle.valueInTxn
+          && pNode->handle.valueInTxn != TQ_DELETE_TOKEN) {
+        pMeta->deleter(pNode->handle.valueInTxn);
+      }
       //change pointer ownership
       pNode->handle.valueInTxn = value;
+      tqLinkUnpersist(pMeta, pNode);
       return 0;
     } else {
       pNode = pNode->next;
@@ -279,16 +489,58 @@ int32_t tqHandlePut(TqMetaStore* pMeta, int64_t key, void* value) {
   memset(pNewNode, 0, sizeof(TqMetaList));
   pNewNode->handle.key = key;
   pNewNode->handle.valueInTxn = value;
+  pNewNode->next = pMeta->bucket[bucketKey];
+  pMeta->bucket[bucketKey] = pNewNode;
+  tqLinkUnpersist(pMeta, pNewNode);
   return 0;
 }
 
-static TqMetaHandle* tqHandleGetUncommitted(TqMetaStore* pMeta, int64_t key) {
+int32_t tqHandleCopyPut(TqMetaStore* pMeta, int64_t key, void* value, size_t vsize) {
-  int64_t bucketKey = key & TQ_BUCKET_SIZE;
+  void *vmem = malloc(vsize);
+  if(vmem == NULL) {
+    //TODO: memory error
+    return -1;
+  }
+  memcpy(vmem, value, vsize);
+  int64_t bucketKey = key & TQ_BUCKET_MASK;
   TqMetaList* pNode = pMeta->bucket[bucketKey];
   while(pNode) {
     if(pNode->handle.key == key) {
-      if(pNode->handle.valueInTxn != NULL) {
+      //TODO: think about thread safety
-        return &pNode->handle;
+      if(pNode->handle.valueInTxn
+          && pNode->handle.valueInTxn != TQ_DELETE_TOKEN) {
+        pMeta->deleter(pNode->handle.valueInTxn);
+      }
+      //change pointer ownership
+      pNode->handle.valueInTxn = vmem;
+      tqLinkUnpersist(pMeta, pNode);
+      return 0;
+    } else {
+      pNode = pNode->next;
+    }
+  }
+  TqMetaList *pNewNode = malloc(sizeof(TqMetaList));
+  if(pNewNode == NULL) {
+    //TODO: memory error
+    return -1;
+  }
+  memset(pNewNode, 0, sizeof(TqMetaList));
+  pNewNode->handle.key = key;
+  pNewNode->handle.valueInTxn = vmem;
+  pNewNode->next = pMeta->bucket[bucketKey];
+  pMeta->bucket[bucketKey] = pNewNode;
+  tqLinkUnpersist(pMeta, pNewNode);
+  return 0;
+}
+
+static void* tqHandleGetUncommitted(TqMetaStore* pMeta, int64_t key) {
+  int64_t bucketKey = key & TQ_BUCKET_MASK;
+  TqMetaList* pNode = pMeta->bucket[bucketKey];
+  while(pNode) {
+    if(pNode->handle.key == key) {
+      if(pNode->handle.valueInTxn != NULL
+          && pNode->handle.valueInTxn != TQ_DELETE_TOKEN) {
+        return pNode->handle.valueInTxn;
       } else {
         return NULL;
       }
@@ -300,36 +552,39 @@ static TqMetaHandle* tqHandleGetUncommitted(TqMetaStore* pMeta, int64_t key) {
 }
 
 int32_t tqHandleCommit(TqMetaStore* pMeta, int64_t key) {
-  int64_t bucketKey = key & TQ_BUCKET_SIZE;
+  int64_t bucketKey = key & TQ_BUCKET_MASK;
   TqMetaList* pNode = pMeta->bucket[bucketKey];
   while(pNode) {
     if(pNode->handle.key == key) {
-      if(pNode->handle.valueInUse != NULL) {
+      if(pNode->handle.valueInTxn == NULL) {
+        return -1;
+      }
+      if(pNode->handle.valueInUse
+          && pNode->handle.valueInUse != TQ_DELETE_TOKEN) {
         pMeta->deleter(pNode->handle.valueInUse);
       }
       pNode->handle.valueInUse = pNode->handle.valueInTxn;
-      if(pNode->unpersistNext == NULL) {
+      pNode->handle.valueInTxn = NULL;
-        pNode->unpersistNext = pMeta->unpersistHead->unpersistNext;
+      tqLinkUnpersist(pMeta, pNode);
-        pNode->unpersistPrev = pMeta->unpersistHead;
-        pMeta->unpersistHead->unpersistNext->unpersistPrev = pNode;
-        pMeta->unpersistHead->unpersistNext = pNode;
-      }
       return 0;
     } else {
       pNode = pNode->next;
     }
   }
-  return -1;
+  return -2;
 }
 
 int32_t tqHandleAbort(TqMetaStore* pMeta, int64_t key) {
-  int64_t bucketKey = key & TQ_BUCKET_SIZE;
+  int64_t bucketKey = key & TQ_BUCKET_MASK;
   TqMetaList* pNode = pMeta->bucket[bucketKey];
   while(pNode) {
     if(pNode->handle.key == key) {
-      if(pNode->handle.valueInTxn != NULL) {
+      if(pNode->handle.valueInTxn) {
-        pMeta->deleter(pNode->handle.valueInTxn);
+        if(pNode->handle.valueInTxn != TQ_DELETE_TOKEN) {
+          pMeta->deleter(pNode->handle.valueInTxn);
+        }
         pNode->handle.valueInTxn = NULL;
+        tqLinkUnpersist(pMeta, pNode);
         return 0;
       }
       return -1;
@@ -341,12 +596,15 @@ int32_t tqHandleAbort(TqMetaStore* pMeta, int64_t key) {
 }
 
 int32_t tqHandleDel(TqMetaStore* pMeta, int64_t key) {
-  int64_t bucketKey = key & TQ_BUCKET_SIZE;
+  int64_t bucketKey = key & TQ_BUCKET_MASK;
   TqMetaList* pNode = pMeta->bucket[bucketKey];
   while(pNode) {
-    if(pNode->handle.key == key) {
+    if(pNode->handle.valueInTxn != TQ_DELETE_TOKEN) {
-      pMeta->deleter(pNode->handle.valueInTxn);
+      if(pNode->handle.valueInTxn) {
-      pNode->handle.valueInTxn = NULL;
+        pMeta->deleter(pNode->handle.valueInTxn);
+      }
+      pNode->handle.valueInTxn = TQ_DELETE_TOKEN;
+      tqLinkUnpersist(pMeta, pNode);
       return 0;
     } else {
       pNode = pNode->next;
@@ -356,35 +614,7 @@ int32_t tqHandleDel(TqMetaStore* pMeta, int64_t key) {
   return -1;
 }
 
-int32_t tqHandleClear(TqMetaStore* pMeta, int64_t key) {
+//TODO: clean deleted idx and data from persistent file
-  int64_t bucketKey = key & TQ_BUCKET_SIZE;
+int32_t tqStoreCompact(TqMetaStore *pMeta) {
-  TqMetaList* pNode = pMeta->bucket[bucketKey];
+  return 0;
-  bool exist = false;
-  while(pNode) {
-    if(pNode->handle.key == key) {
-      if(pNode->handle.valueInUse != NULL) {
-        exist = true;
-        pMeta->deleter(pNode->handle.valueInUse);
-        pNode->handle.valueInUse = NULL;
-      }
-      if(pNode->handle.valueInTxn != NULL) {
-        exist = true;
-        pMeta->deleter(pNode->handle.valueInTxn);
-        pNode->handle.valueInTxn = NULL;
-      }
-      if(exist) {
-        if(pNode->unpersistNext == NULL) {
-          pNode->unpersistNext = pMeta->unpersistHead->unpersistNext;
-          pNode->unpersistPrev = pMeta->unpersistHead;
-          pMeta->unpersistHead->unpersistNext->unpersistPrev = pNode;
-          pMeta->unpersistHead->unpersistNext = pNode;
-        }
-        return 0;
-      }
-      return -1;
-    } else {
-      pNode = pNode->next;
-    }
-  }
-  return -2;
 }

source/dnode/vnode/tq/test/CMakeLists.txt

@@ -0,0 +1,20 @@
+add_executable(tqTest "")
+target_sources(tqTest
+    PRIVATE
+    "tqMetaTest.cpp"
+)
+target_include_directories(tqTest
+    PUBLIC
+    "${CMAKE_SOURCE_DIR}/include/server/vnode/tq"
+    "${CMAKE_CURRENT_SOURCE_DIR}/../inc"
+)
+
+target_link_libraries(tqTest
+    tq
+    gtest_main
+)
+enable_testing()
+add_test(
+    NAME tq_test
+    COMMAND tqTest
+)

source/dnode/vnode/tq/test/tqMetaTest.cpp

@@ -0,0 +1,294 @@
+#include <gtest/gtest.h>
+#include <cstring>
+#include <iostream>
+#include <queue>
+
+#include "tqMetaStore.h"
+
+struct Foo {
+  int32_t a;
+};
+
+int FooSerializer(const void* pObj, TqSerializedHead** ppHead) {
+  Foo* foo = (Foo*) pObj;
+  if((*ppHead) == NULL || (*ppHead)->ssize < sizeof(TqSerializedHead) + sizeof(int32_t)) {
+    *ppHead = (TqSerializedHead*)realloc(*ppHead, sizeof(TqSerializedHead) + sizeof(int32_t));
+    (*ppHead)->ssize = sizeof(TqSerializedHead) + sizeof(int32_t);
+  }
+  *(int32_t*)(*ppHead)->content = foo->a;
+  return (*ppHead)->ssize;
+}
+
+const void* FooDeserializer(const TqSerializedHead* pHead, void** ppObj) {
+  if(*ppObj == NULL) {
+    *ppObj = realloc(*ppObj, sizeof(int32_t));
+  }
+  Foo* pFoo = *(Foo**)ppObj;
+  pFoo->a = *(int32_t*)pHead->content; 
+  return NULL;
+}
+
+void FooDeleter(void* pObj) {
+  free(pObj); 
+}
+
+class TqMetaTest : public ::testing::Test {
+  protected:
+
+    void SetUp() override {
+      taosRemoveDir(pathName);
+      pMeta = tqStoreOpen(pathName,
+          FooSerializer, FooDeserializer, FooDeleter);
+      ASSERT(pMeta);
+    }
+
+    void TearDown() override {
+      tqStoreClose(pMeta);
+    }
+
+    TqMetaStore* pMeta;
+    const char* pathName = "/tmp/tq_test";
+};
+
+TEST_F(TqMetaTest, copyPutTest) {
+  Foo foo;
+  foo.a = 3;
+  tqHandleCopyPut(pMeta, 1, &foo, sizeof(Foo));
+
+  Foo* pFoo = (Foo*) tqHandleGet(pMeta, 1);
+  EXPECT_EQ(pFoo == NULL, true);
+
+  tqHandleCommit(pMeta, 1);
+  pFoo = (Foo*) tqHandleGet(pMeta, 1);
+  EXPECT_EQ(pFoo->a, 3);
+}
+
+TEST_F(TqMetaTest, persistTest) {
+  Foo* pFoo = (Foo*)malloc(sizeof(Foo));
+  pFoo->a = 2;
+  tqHandleMovePut(pMeta, 1, pFoo);
+  Foo* pBar = (Foo*)tqHandleGet(pMeta, 1);
+  EXPECT_EQ(pBar == NULL, true);
+  tqHandleCommit(pMeta, 1);
+  pBar = (Foo*)tqHandleGet(pMeta, 1);
+  EXPECT_EQ(pBar->a, pFoo->a);
+  pBar = (Foo*)tqHandleGet(pMeta, 2);
+  EXPECT_EQ(pBar == NULL, true);
+
+  tqStoreClose(pMeta);
+  pMeta = tqStoreOpen(pathName,
+      FooSerializer, FooDeserializer, FooDeleter);
+  ASSERT(pMeta);
+
+  pBar = (Foo*)tqHandleGet(pMeta, 1);
+  ASSERT_EQ(pBar != NULL, true);
+  EXPECT_EQ(pBar->a, 2);
+
+  pBar = (Foo*)tqHandleGet(pMeta, 2);
+  EXPECT_EQ(pBar == NULL, true);
+}
+
+TEST_F(TqMetaTest, uncommittedTest) {
+  Foo* pFoo = (Foo*)malloc(sizeof(Foo));
+  pFoo->a = 3;
+  tqHandleMovePut(pMeta, 1, pFoo);
+
+  pFoo = (Foo*) tqHandleGet(pMeta, 1);
+  EXPECT_EQ(pFoo == NULL, true);
+}
+
+TEST_F(TqMetaTest, abortTest) {
+  Foo* pFoo = (Foo*)malloc(sizeof(Foo));
+  pFoo->a = 3;
+  tqHandleMovePut(pMeta, 1, pFoo);
+
+  pFoo = (Foo*) tqHandleGet(pMeta, 1);
+  EXPECT_EQ(pFoo == NULL, true);
+
+  tqHandleAbort(pMeta, 1);
+  pFoo = (Foo*) tqHandleGet(pMeta, 1);
+  EXPECT_EQ(pFoo == NULL, true);
+}
+
+TEST_F(TqMetaTest, deleteTest) {
+  Foo* pFoo = (Foo*)malloc(sizeof(Foo));
+  pFoo->a = 3;
+  tqHandleMovePut(pMeta, 1, pFoo);
+
+  pFoo = (Foo*) tqHandleGet(pMeta, 1);
+  EXPECT_EQ(pFoo == NULL, true);
+
+  tqHandleCommit(pMeta, 1);
+
+  pFoo = (Foo*) tqHandleGet(pMeta, 1);
+  ASSERT_EQ(pFoo != NULL, true);
+  EXPECT_EQ(pFoo->a, 3);
+
+  tqHandleDel(pMeta, 1);
+  pFoo = (Foo*) tqHandleGet(pMeta, 1);
+  ASSERT_EQ(pFoo != NULL, true);
+  EXPECT_EQ(pFoo->a, 3);
+
+  tqHandleCommit(pMeta, 1);
+  pFoo = (Foo*) tqHandleGet(pMeta, 1);
+  EXPECT_EQ(pFoo == NULL, true);
+
+  tqStoreClose(pMeta);
+  pMeta = tqStoreOpen(pathName,
+      FooSerializer, FooDeserializer, FooDeleter);
+  ASSERT(pMeta);
+
+  pFoo = (Foo*) tqHandleGet(pMeta, 1);
+  EXPECT_EQ(pFoo == NULL, true);
+}
+
+TEST_F(TqMetaTest, intxnPersist) {
+  Foo* pFoo = (Foo*)malloc(sizeof(Foo));
+  pFoo->a = 3;
+  tqHandleMovePut(pMeta, 1, pFoo);
+  tqHandleCommit(pMeta, 1);
+
+  Foo* pBar = (Foo*)malloc(sizeof(Foo));
+  pBar->a = 4;
+  tqHandleMovePut(pMeta, 1, pBar);
+
+  Foo* pFoo1 = (Foo*)tqHandleGet(pMeta, 1);
+  EXPECT_EQ(pFoo1->a, 3);
+
+  tqStoreClose(pMeta);
+  pMeta = tqStoreOpen(pathName,
+      FooSerializer, FooDeserializer, FooDeleter);
+  ASSERT(pMeta);
+
+  pFoo1 = (Foo*)tqHandleGet(pMeta, 1);
+  EXPECT_EQ(pFoo1->a, 3);
+
+  tqHandleCommit(pMeta, 1);
+
+  pFoo1 = (Foo*)tqHandleGet(pMeta, 1);
+  EXPECT_EQ(pFoo1->a, 4);
+
+  tqStoreClose(pMeta);
+  pMeta = tqStoreOpen(pathName,
+      FooSerializer, FooDeserializer, FooDeleter);
+  ASSERT(pMeta);
+
+  pFoo1 = (Foo*)tqHandleGet(pMeta, 1);
+  EXPECT_EQ(pFoo1->a, 4);
+}
+
+TEST_F(TqMetaTest, multiplePage) {
+  srand(0);
+  std::vector<int> v;
+  for(int i = 0; i < 1000; i++) {
+    v.push_back(rand());
+    Foo foo;
+    foo.a = v[i];
+    tqHandleCopyPut(pMeta, i, &foo, sizeof(Foo));
+  }
+  for(int i = 0; i < 500; i++) {
+    tqHandleCommit(pMeta, i);
+    Foo* pFoo = (Foo*)tqHandleGet(pMeta, i);
+    ASSERT_EQ(pFoo != NULL, true) << " at idx " << i << "\n";
+    EXPECT_EQ(pFoo->a, v[i]);
+  }
+
+  tqStoreClose(pMeta);
+  pMeta = tqStoreOpen(pathName,
+      FooSerializer, FooDeserializer, FooDeleter);
+  ASSERT(pMeta);
+  
+  for(int i = 500; i < 1000; i++) {
+    tqHandleCommit(pMeta, i);
+    Foo* pFoo = (Foo*)tqHandleGet(pMeta, i);
+    ASSERT_EQ(pFoo != NULL, true) << " at idx " << i << "\n";
+    EXPECT_EQ(pFoo->a, v[i]);
+  }
+
+  for(int i = 0; i < 1000; i++) {
+    Foo* pFoo = (Foo*)tqHandleGet(pMeta, i);
+    ASSERT_EQ(pFoo != NULL, true) << " at idx " << i << "\n";
+    EXPECT_EQ(pFoo->a, v[i]);
+  }
+
+}
+
+TEST_F(TqMetaTest, multipleRewrite) {
+  srand(0);
+  std::vector<int> v;
+  for(int i = 0; i < 1000; i++) {
+    v.push_back(rand());
+    Foo foo;
+    foo.a = v[i];
+    tqHandleCopyPut(pMeta, i, &foo, sizeof(Foo));
+  }
+
+  for(int i = 0; i < 500; i++) {
+    tqHandleCommit(pMeta, i);
+    v[i] = rand();
+    Foo foo;
+    foo.a = v[i];
+    tqHandleCopyPut(pMeta, i, &foo, sizeof(Foo));
+  }
+
+  for(int i = 500; i < 1000; i++) {
+    v[i] = rand();
+    Foo foo;
+    foo.a = v[i];
+    tqHandleCopyPut(pMeta, i, &foo, sizeof(Foo));
+  }
+
+  for(int i = 0; i < 1000; i++) {
+    tqHandleCommit(pMeta, i);
+  }
+
+  tqStoreClose(pMeta);
+  pMeta = tqStoreOpen(pathName,
+      FooSerializer, FooDeserializer, FooDeleter);
+  ASSERT(pMeta);
+  
+  for(int i = 500; i < 1000; i++) {
+    v[i] = rand();
+    Foo foo;
+    foo.a = v[i];
+    tqHandleCopyPut(pMeta, i, &foo, sizeof(Foo));
+    tqHandleCommit(pMeta, i);
+  }
+
+  for(int i = 0; i < 1000; i++) {
+    Foo* pFoo = (Foo*)tqHandleGet(pMeta, i);
+    ASSERT_EQ(pFoo != NULL, true) << " at idx " << i << "\n";
+    EXPECT_EQ(pFoo->a, v[i]);
+  }
+
+}
+
+TEST_F(TqMetaTest, dupCommit) {
+  srand(0);
+  std::vector<int> v;
+  for(int i = 0; i < 1000; i++) {
+    v.push_back(rand());
+    Foo foo;
+    foo.a = v[i];
+    tqHandleCopyPut(pMeta, i, &foo, sizeof(Foo));
+  }
+
+  for(int i = 0; i < 1000; i++) {
+    int ret = tqHandleCommit(pMeta, i);
+    EXPECT_EQ(ret, 0);
+    ret = tqHandleCommit(pMeta, i);
+    EXPECT_EQ(ret, -1);
+  }
+
+  for(int i = 0; i < 1000; i++) {
+    int ret = tqHandleCommit(pMeta, i);
+    EXPECT_EQ(ret, -1);
+  }
+
+  for(int i = 0; i < 1000; i++) {
+    Foo* pFoo = (Foo*)tqHandleGet(pMeta, i);
+    ASSERT_EQ(pFoo != NULL, true) << " at idx " << i << "\n";
+    EXPECT_EQ(pFoo->a, v[i]);
+  }
+
+}

source/dnode/vnode/tsdb/inc/tsdbIdx.h

@@ -13,8 +13,8 @@
  * along with this program. If not, see <http://www.gnu.org/licenses/>.
  */
 
-#ifndef _TD_RAFT_INT_H_
+#ifndef _TD_TSDB_IDX_H_
-#define _TD_RAFT_INT_H_
+#define _TD_TSDB_IDX_H_
 
 #ifdef __cplusplus
 extern "C" {
@@ -24,4 +24,4 @@ extern "C" {
 }
 #endif
 
-#endif /*_TD_RAFT_INT_H_*/
+#endif /*_TD_TSDB_IDX_H_*/

source/dnode/vnode/tsdb/inc/tsdbOptions.h

@@ -20,6 +20,11 @@
 extern "C" {
 #endif
 
+extern const STsdbOptions defautlTsdbOptions;
+
+int  tsdbValidateOptions(const STsdbOptions *);
+void tsdbOptionsCopy(STsdbOptions *pDest, const STsdbOptions *pSrc);
+
 #ifdef __cplusplus
 }
 #endif

source/dnode/vnode/tsdb/inc/tsdbSMA.h

@@ -0,0 +1,27 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+
+#ifndef _TD_TSDB_SMA_H_
+#define _TD_TSDB_SMA_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /*_TD_TSDB_SMA_H_*/

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

@@ -0,0 +1,14 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */

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

@@ -13,4 +13,82 @@
  * along with this program. If not, see <http://www.gnu.org/licenses/>.
  */
 
-#include "tsdbDef.h"
+#include "tsdbDef.h"
+
+static STsdb *tsdbNew(const char *path, const STsdbOptions *pTsdbOptions);
+static void   tsdbFree(STsdb *pTsdb);
+static int    tsdbOpenImpl(STsdb *pTsdb);
+static void   tsdbCloseImpl(STsdb *pTsdb);
+
+STsdb *tsdbOpen(const char *path, const STsdbOptions *pTsdbOptions) {
+  STsdb *pTsdb = NULL;
+
+  // Set default TSDB Options
+  if (pTsdbOptions == NULL) {
+    pTsdbOptions = &defautlTsdbOptions;
+  }
+
+  // Validate the options
+  if (tsdbValidateOptions(pTsdbOptions) < 0) {
+    // TODO: handle error
+    return NULL;
+  }
+
+  // Create the handle
+  pTsdb = tsdbNew(path, pTsdbOptions);
+  if (pTsdb == NULL) {
+    // TODO: handle error
+    return NULL;
+  }
+
+  taosMkDir(path);
+
+  // Open the TSDB
+  if (tsdbOpenImpl(pTsdb) < 0) {
+    // TODO: handle error
+    return NULL;
+  }
+
+  return pTsdb;
+}
+
+void tsdbClose(STsdb *pTsdb) {
+  if (pTsdb) {
+    tsdbCloseImpl(pTsdb);
+    tsdbFree(pTsdb);
+  }
+}
+
+void tsdbRemove(const char *path) { taosRemoveDir(path); }
+
+/* ------------------------ STATIC METHODS ------------------------ */
+static STsdb *tsdbNew(const char *path, const STsdbOptions *pTsdbOptions) {
+  STsdb *pTsdb = NULL;
+
+  pTsdb = (STsdb *)calloc(1, sizeof(STsdb));
+  if (pTsdb == NULL) {
+    // TODO: handle error
+    return NULL;
+  }
+
+  pTsdb->path = strdup(path);
+  tsdbOptionsCopy(&(pTsdb->options), pTsdbOptions);
+
+  return pTsdb;
+}
+
+static void tsdbFree(STsdb *pTsdb) {
+  if (pTsdb) {
+    tfree(pTsdb->path);
+    free(pTsdb);
+  }
+}
+
+static int tsdbOpenImpl(STsdb *pTsdb) {
+  // TODO
+  return 0;
+}
+
+static void tsdbCloseImpl(STsdb *pTsdb) {
+  // TODO
+}

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

@@ -11,4 +11,24 @@
  *
  * You should have received a copy of the GNU Affero General Public License
  * along with this program. If not, see <http://www.gnu.org/licenses/>.
- */
+ */
+
+#include "tsdbDef.h"
+
+const STsdbOptions defautlTsdbOptions = {.lruCacheSize = 0};
+
+int tsdbOptionsInit(STsdbOptions *pTsdbOptions) {
+  // TODO
+  return 0;
+}
+
+void tsdbOptionsClear(STsdbOptions *pTsdbOptions) {
+  // TODO
+}
+
+int tsdbValidateOptions(const STsdbOptions *pTsdbOptions) {
+  // TODO
+  return 0;
+}
+
+void tsdbOptionsCopy(STsdbOptions *pDest, const STsdbOptions *pSrc) { memcpy(pDest, pSrc, sizeof(STsdbOptions)); }

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

@@ -0,0 +1,14 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */

source/libs/executor/CMakeLists.txt

@@ -8,5 +8,5 @@ target_include_directories(
 
 target_link_libraries(
         executor
-        PRIVATE os util common
+        PRIVATE os util common function parser
 )

source/libs/executor/inc/executil.h

@@ -15,6 +15,8 @@
 #ifndef TDENGINE_QUERYUTIL_H
 #define TDENGINE_QUERYUTIL_H
 
+#include "common.h"
+#include "tpagedfile.h"
 #include "tbuffer.h"
 
 #define SET_RES_WINDOW_KEY(_k, _ori, _len, _uid)     \
@@ -40,42 +42,92 @@
 
 #define curTimeWindowIndex(_winres)        ((_winres)->curIndex)
 
-int32_t getOutputInterResultBufSize(SQueryAttr* pQueryAttr);
+struct SColumnFilterElem;
 
-size_t  getResultRowSize(SQueryRuntimeEnv* pRuntimeEnv);
+typedef bool (*__filter_func_t)(struct SColumnFilterElem* pFilter, const char* val1, const char* val2, int16_t type);
+
+typedef struct SGroupResInfo {
+  int32_t totalGroup;
+  int32_t currentGroup;
+  int32_t index;
+  SArray* pRows;      // SArray<SResultRow*>
+  bool    ordered;
+  int32_t position;
+} SGroupResInfo;
+
+typedef struct SResultRow {
+  int32_t       pageId;      // pageId & rowId is the position of current result in disk-based output buffer
+  int32_t       offset:29;   // row index in buffer page
+  bool          startInterp; // the time window start timestamp has done the interpolation already.
+  bool          endInterp;   // the time window end timestamp has done the interpolation already.
+  bool          closed;      // this result status: closed or opened
+  uint32_t      numOfRows;   // number of rows of current time window
+  struct SResultRowEntryInfo* pEntryInfo;  // For each result column, there is a resultInfo
+  STimeWindow   win;
+  char         *key;               // start key of current result row
+} SResultRow;
+
+typedef struct SResultRowInfo {
+  SResultRow** pResult;    // result list
+  int16_t      type:8;     // data type for hash key
+  int32_t      size:24;    // number of result set
+  int32_t      capacity;   // max capacity
+  int32_t      curPos;     // current active result row index of pResult list
+} SResultRowInfo;
+
+typedef struct SResultRowPool {
+  int32_t elemSize;
+  int32_t blockSize;
+  int32_t numOfElemPerBlock;
+
+  struct {
+    int32_t blockIndex;
+    int32_t pos;
+  } position;
+
+  SArray* pData;    // SArray<void*>
+} SResultRowPool;
+
+struct SQueryAttr;
+struct SQueryRuntimeEnv;
+struct SUdfInfo;
+
+int32_t getOutputInterResultBufSize(struct SQueryAttr* pQueryAttr);
+
+size_t  getResultRowSize(struct SQueryRuntimeEnv* pRuntimeEnv);
 int32_t initResultRowInfo(SResultRowInfo* pResultRowInfo, int32_t size, int16_t type);
 void    cleanupResultRowInfo(SResultRowInfo* pResultRowInfo);
 
-void    resetResultRowInfo(SQueryRuntimeEnv* pRuntimeEnv, SResultRowInfo* pResultRowInfo);
+void    resetResultRowInfo(struct SQueryRuntimeEnv* pRuntimeEnv, SResultRowInfo* pResultRowInfo);
 int32_t numOfClosedResultRows(SResultRowInfo* pResultRowInfo);
 void    closeAllResultRows(SResultRowInfo* pResultRowInfo);
 
 int32_t initResultRow(SResultRow *pResultRow);
 void    closeResultRow(SResultRowInfo* pResultRowInfo, int32_t slot);
 bool    isResultRowClosed(SResultRowInfo *pResultRowInfo, int32_t slot);
-void    clearResultRow(SQueryRuntimeEnv* pRuntimeEnv, SResultRow* pResultRow, int16_t type);
+void    clearResultRow(struct SQueryRuntimeEnv* pRuntimeEnv, SResultRow* pResultRow, int16_t type);
 
-SResultRowCellInfo* getResultCell(const SResultRow* pRow, int32_t index, int32_t* offset);
+struct SResultRowEntryInfo* getResultCell(const SResultRow* pRow, int32_t index, int32_t* offset);
 
 void* destroyQueryFuncExpr(SExprInfo* pExprInfo, int32_t numOfExpr);
 void* freeColumnInfo(SColumnInfo* pColumnInfo, int32_t numOfCols);
-int32_t getRowNumForMultioutput(SQueryAttr* pQueryAttr, bool topBottomQuery, bool stable);
+int32_t getRowNumForMultioutput(struct SQueryAttr* pQueryAttr, bool topBottomQuery, bool stable);
 
 static FORCE_INLINE SResultRow *getResultRow(SResultRowInfo *pResultRowInfo, int32_t slot) {
   assert(pResultRowInfo != NULL && slot >= 0 && slot < pResultRowInfo->size);
   return pResultRowInfo->pResult[slot];
 }
 
-static FORCE_INLINE char* getPosInResultPage(SQueryAttr* pQueryAttr, tFilePage* page, int32_t rowOffset,
+static FORCE_INLINE char* getPosInResultPage(struct SQueryAttr* pQueryAttr, SFilePage* page, int32_t rowOffset,
                                              int32_t offset) {
   assert(rowOffset >= 0 && pQueryAttr != NULL);
 
-  int32_t numOfRows = (int32_t)getRowNumForMultioutput(pQueryAttr, pQueryAttr->topBotQuery, pQueryAttr->stableQuery);
+//  int32_t numOfRows = (int32_t)getRowNumForMultioutput(pQueryAttr, pQueryAttr->topBotQuery, pQueryAttr->stableQuery);
-  return ((char *)page->data) + rowOffset + offset * numOfRows;
+//  return ((char *)page->data) + rowOffset + offset * numOfRows;
 }
 
-bool isNullOperator(SColumnFilterElem *pFilter, const char* minval, const char* maxval, int16_t type);
+//bool isNullOperator(SColumnFilterElem *pFilter, const char* minval, const char* maxval, int16_t type);
-bool notNullOperator(SColumnFilterElem *pFilter, const char* minval, const char* maxval, int16_t type);
+//bool notNullOperator(SColumnFilterElem *pFilter, const char* minval, const char* maxval, int16_t type);
 
 __filter_func_t getFilterOperator(int32_t lowerOptr, int32_t upperOptr);
 
@@ -103,8 +155,8 @@ bool    hasRemainData(SGroupResInfo* pGroupResInfo);
 bool    incNextGroup(SGroupResInfo* pGroupResInfo);
 int32_t getNumOfTotalRes(SGroupResInfo* pGroupResInfo);
 
-int32_t mergeIntoGroupResult(SGroupResInfo* pGroupResInfo, SQueryRuntimeEnv *pRuntimeEnv, int32_t* offset);
+int32_t mergeIntoGroupResult(SGroupResInfo* pGroupResInfo, struct SQueryRuntimeEnv *pRuntimeEnv, int32_t* offset);
 
-int32_t initUdfInfo(SUdfInfo* pUdfInfo);
+int32_t initUdfInfo(struct SUdfInfo* pUdfInfo);
 
 #endif  // TDENGINE_QUERYUTIL_H

source/libs/executor/inc/executorimpl.h

@@ -0,0 +1,645 @@
+/*
+ * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
+ *
+ * This program is free software: you can use, redistribute, and/or modify
+ * it under the terms of the GNU Affero General Public License, version 3
+ * or later ("AGPL"), as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program. If not, see <http://www.gnu.org/licenses/>.
+ */
+#ifndef TDENGINE_EXECUTORIMPL_H
+#define TDENGINE_EXECUTORIMPL_H
+
+#include "os.h"
+#include "common.h"
+#include "ttszip.h"
+#include "tvariant.h"
+
+#include "thash.h"
+//#include "parser.h"
+#include "executil.h"
+#include "taosdef.h"
+#include "tarray.h"
+#include "tfilter.h"
+#include "tlockfree.h"
+#include "tpagedfile.h"
+
+struct SColumnFilterElem;
+
+typedef struct {
+  uint32_t  numOfTables;
+  SArray   *pGroupList;
+  SHashObj *map;  // speedup acquire the tableQueryInfo by table uid
+} STableGroupInfo;
+
+typedef int32_t (*__block_search_fn_t)(char* data, int32_t num, int64_t key, int32_t order);
+
+#define IS_QUERY_KILLED(_q) ((_q)->code == TSDB_CODE_TSC_QUERY_CANCELLED)
+#define Q_STATUS_EQUAL(p, s)  (((p) & (s)) != 0u)
+#define QUERY_IS_ASC_QUERY(q) (GET_FORWARD_DIRECTION_FACTOR((q)->order.order) == QUERY_ASC_FORWARD_STEP)
+
+#define GET_TABLEGROUP(q, _index)   ((SArray*) taosArrayGetP((q)->tableqinfoGroupInfo.pGroupList, (_index)))
+
+#define GET_NUM_OF_RESULTS(_r) (((_r)->outputBuf) == NULL? 0:((_r)->outputBuf)->info.rows)
+
+#define NEEDTO_COMPRESS_QUERY(size) ((size) > tsCompressColData? 1 : 0)
+
+enum {
+  // when query starts to execute, this status will set
+      QUERY_NOT_COMPLETED = 0x1u,
+
+  /* query is over
+   * 1. this status is used in one row result query process, e.g., count/sum/first/last/ avg...etc.
+   * 2. when all data within queried time window, it is also denoted as query_completed
+   */
+      QUERY_COMPLETED = 0x2u,
+
+  /* when the result is not completed return to client, this status will be
+   * usually used in case of interval query with interpolation option
+   */
+      QUERY_OVER = 0x4u,
+};
+
+
+typedef struct SResultRowCell {
+  uint64_t     groupId;
+  SResultRow  *pRow;
+} SResultRowCell;
+
+/**
+ * If the number of generated results is greater than this value,
+ * query query will be halt and return results to client immediate.
+ */
+typedef struct SRspResultInfo {
+  int64_t total;      // total generated result size in rows
+  int32_t capacity;   // capacity of current result output buffer
+  int32_t threshold;  // result size threshold in rows.
+} SRspResultInfo;
+
+typedef struct SColumnFilterElem {
+  int16_t           bytes;  // column length
+  __filter_func_t   fp;
+  SColumnFilterInfo filterInfo;
+  void              *q;
+} SColumnFilterElem;
+
+typedef struct SSingleColumnFilterInfo {
+  void*              pData;
+  void*              pData2;  //used for nchar column
+  int32_t            numOfFilters;
+  SColumnInfo        info;
+  SColumnFilterElem* pFilters;
+} SSingleColumnFilterInfo;
+
+typedef struct STableQueryInfo {
+  TSKEY       lastKey;
+  int32_t     groupIndex;     // group id in table list
+  SVariant    tag;
+  STimeWindow win;
+  STSCursor   cur;
+  void*       pTable;         // for retrieve the page id list
+  SResultRowInfo resInfo;
+} STableQueryInfo;
+
+typedef enum {
+  QUERY_PROF_BEFORE_OPERATOR_EXEC = 0,
+  QUERY_PROF_AFTER_OPERATOR_EXEC,
+  QUERY_PROF_QUERY_ABORT
+} EQueryProfEventType;
+
+typedef struct {
+  EQueryProfEventType eventType;
+  int64_t eventTime;
+
+  union {
+    uint8_t operatorType; //for operator event
+    int32_t abortCode; //for query abort event
+  };
+} SQueryProfEvent;
+
+typedef struct {
+  uint8_t operatorType;
+  int64_t sumSelfTime;
+  int64_t sumRunTimes;
+} SOperatorProfResult;
+
+typedef struct SQueryCostInfo {
+  uint64_t loadStatisTime;
+  uint64_t loadFileBlockTime;
+  uint64_t loadDataInCacheTime;
+  uint64_t loadStatisSize;
+  uint64_t loadFileBlockSize;
+  uint64_t loadDataInCacheSize;
+  
+  uint64_t loadDataTime;
+  uint64_t totalRows;
+  uint64_t totalCheckedRows;
+  uint32_t totalBlocks;
+  uint32_t loadBlocks;
+  uint32_t loadBlockStatis;
+  uint32_t discardBlocks;
+  uint64_t elapsedTime;
+  uint64_t firstStageMergeTime;
+  uint64_t winInfoSize;
+  uint64_t tableInfoSize;
+  uint64_t hashSize;
+  uint64_t numOfTimeWindows;
+
+  SArray*   queryProfEvents;  //SArray<SQueryProfEvent>
+  SHashObj* operatorProfResults; //map<operator_type, SQueryProfEvent>
+} SQueryCostInfo;
+
+typedef struct {
+  int64_t vgroupLimit;
+  int64_t ts;
+} SOrderedPrjQueryInfo;
+
+typedef struct {
+  char*   tags;
+  SArray* pResult;  // SArray<SStddevInterResult>
+} SInterResult;
+
+// The basic query information extracted from the SQueryInfo tree to support the
+// execution of query in a data node.
+typedef struct SQueryAttr {
+  SLimit           limit;
+  SLimit           slimit;
+
+  // todo comment it
+  bool             stableQuery;      // super table query or not
+  bool             topBotQuery;      // TODO used bitwise flag
+  bool             groupbyColumn;    // denote if this is a groupby normal column query
+  bool             hasTagResults;    // if there are tag values in final result or not
+  bool             timeWindowInterpo;// if the time window start/end required interpolation
+  bool             queryBlockDist;    // if query data block distribution
+  bool             stabledev;        // super table stddev query
+  bool             tsCompQuery;      // is tscomp query
+  bool             diffQuery;        // is diff query
+  bool             simpleAgg;
+  bool             pointInterpQuery; // point interpolation query
+  bool             needReverseScan;  // need reverse scan
+  bool             distinct;         // distinct  query or not
+  bool             stateWindow;       // window State on sub/normal table
+  bool             createFilterOperator; // if filter operator is needed
+  bool             multigroupResult; // multigroup result can exist in one SSDataBlock
+  int32_t          interBufSize;     // intermediate buffer sizse
+
+  int32_t          havingNum;        // having expr number
+
+  SOrder           order;
+  int16_t          numOfCols;
+  int16_t          numOfTags;
+
+  STimeWindow      window;
+  SInterval        interval;
+  SSessionWindow   sw;
+  int16_t          precision;
+  int16_t          numOfOutput;
+  int16_t          fillType;
+
+  int32_t          srcRowSize;       // todo extract struct
+  int32_t          resultRowSize;
+  int32_t          intermediateResultRowSize; // intermediate result row size, in case of top-k query.
+  int32_t          maxTableColumnWidth;
+  int32_t          tagLen;           // tag value length of current query
+  SGroupbyExpr    *pGroupbyExpr;
+
+  SExprInfo*       pExpr1;
+  SExprInfo*       pExpr2;
+  int32_t          numOfExpr2;
+  SExprInfo*       pExpr3;
+  int32_t          numOfExpr3;
+
+  SColumnInfo*     tableCols;
+  SColumnInfo*     tagColList;
+  int32_t          numOfFilterCols;
+  int64_t*         fillVal;
+  SOrderedPrjQueryInfo prjInfo;        // limit value for each vgroup, only available in global order projection query.
+
+  SSingleColumnFilterInfo* pFilterInfo;
+//  SFilterInfo     *pFilters;
+  
+  void*            tsdb;
+//  SMemRef          memRef;
+  STableGroupInfo  tableGroupInfo;       // table <tid, last_key> list  SArray<STableKeyInfo>
+  int32_t          vgId;
+  SArray          *pUdfInfo;             // no need to free
+} SQueryAttr;
+
+typedef SSDataBlock* (*__operator_fn_t)(void* param, bool* newgroup);
+typedef void (*__optr_cleanup_fn_t)(void* param, int32_t num);
+
+struct SOperatorInfo;
+
+typedef struct SQueryRuntimeEnv {
+  jmp_buf               env;
+  SQueryAttr*           pQueryAttr;
+  uint32_t              status;           // query status
+  void*                 qinfo;
+  uint8_t               scanFlag;         // denotes reversed scan of data or not
+  void*                 pQueryHandle;
+
+  int32_t               prevGroupId;      // previous executed group id
+  bool                  enableGroupData;
+  SDiskbasedResultBuf*  pResultBuf;       // query result buffer based on blocked-wised disk file
+  SHashObj*             pResultRowHashTable; // quick locate the window object for each result
+  SHashObj*             pResultRowListSet;   // used to check if current ResultRowInfo has ResultRow object or not
+  SArray*               pResultRowArrayList; // The array list that contains the Result rows
+  char*                 keyBuf;           // window key buffer
+  SResultRowPool*       pool;             // The window result objects pool, all the resultRow Objects are allocated and managed by this object.
+  char**                prevRow;
+
+  SArray*               prevResult;       // intermediate result, SArray<SInterResult>
+  STSBuf*               pTsBuf;           // timestamp filter list
+  STSCursor             cur;
+
+  char*                 tagVal;           // tag value of current data block
+  struct SScalarFunctionSupport   * scalarSup;
+
+  SSDataBlock          *outputBuf;
+  STableGroupInfo       tableqinfoGroupInfo;  // this is a group array list, including SArray<STableQueryInfo*> structure
+  struct SOperatorInfo *proot;
+  SGroupResInfo         groupResInfo;
+  int64_t               currentOffset;   // dynamic offset value
+
+  STableQueryInfo      *current;
+  SRspResultInfo        resultInfo;
+  SHashObj             *pTableRetrieveTsMap;
+  struct SUdfInfo      *pUdfInfo;
+} SQueryRuntimeEnv;
+
+enum {
+  OP_IN_EXECUTING   = 1,
+  OP_RES_TO_RETURN  = 2,
+  OP_EXEC_DONE      = 3,
+};
+
+enum OPERATOR_TYPE_E {
+  OP_TableScan         = 1,
+  OP_DataBlocksOptScan = 2,
+  OP_TableSeqScan      = 3,
+  OP_TagScan           = 4,
+  OP_TableBlockInfoScan= 5,
+  OP_Aggregate         = 6,
+  OP_Project           = 7,
+  OP_Groupby           = 8,
+  OP_Limit             = 9,
+  OP_SLimit            = 10,
+  OP_TimeWindow        = 11,
+  OP_SessionWindow     = 12,
+  OP_Fill              = 13,
+  OP_MultiTableAggregate     = 14,
+  OP_MultiTableTimeInterval  = 15,
+  OP_DummyInput        = 16,   //TODO remove it after fully refactor.
+  OP_MultiwayMergeSort = 17,   // multi-way data merge into one input stream.
+  OP_GlobalAggregate   = 18,   // global merge for the multi-way data sources.
+  OP_Filter            = 19,
+  OP_Distinct          = 20,
+  OP_Join              = 21,
+  OP_StateWindow       = 22,
+  OP_AllTimeWindow     = 23,
+  OP_AllMultiTableTimeInterval = 24,
+  OP_Order             = 25,
+};
+
+typedef struct SOperatorInfo {
+  uint8_t               operatorType;
+  bool                  blockingOptr;  // block operator or not
+  uint8_t               status;        // denote if current operator is completed
+  int32_t               numOfOutput;   // number of columns of the current operator results
+  char                 *name;          // name, used to show the query execution plan
+  void                 *info;          // extension attribution
+  SExprInfo            *pExpr;
+  SQueryRuntimeEnv     *pRuntimeEnv;
+
+  struct SOperatorInfo **upstream;     // upstream pointer list
+  int32_t               numOfUpstream; // number of upstream. The value is always ONE expect for join operator
+  __operator_fn_t       exec;
+  __optr_cleanup_fn_t   cleanup;
+} SOperatorInfo;
+
+enum {
+  QUERY_RESULT_NOT_READY = 1,
+  QUERY_RESULT_READY     = 2,
+};
+
+typedef struct {
+  int32_t      numOfTags;
+  int32_t      numOfCols;
+  SColumnInfo *colList;
+} SQueriedTableInfo;
+
+typedef struct SQInfo {
+  void*            signature;
+  uint64_t         qId;
+  int32_t          code;        // error code to returned to client
+  int64_t          owner;       // if it is in execution
+
+  SQueryRuntimeEnv runtimeEnv;
+  SQueryAttr       query;
+  void*            pBuf;        // allocated buffer for STableQueryInfo, sizeof(STableQueryInfo)*numOfTables;
+
+  pthread_mutex_t  lock;        // used to synchronize the rsp/query threads
+  tsem_t           ready;
+  int32_t          dataReady;   // denote if query result is ready or not
+  void*            rspContext;  // response context
+  int64_t          startExecTs; // start to exec timestamp
+  char*            sql;         // query sql string
+  SQueryCostInfo   summary;
+} SQInfo;
+
+typedef struct SQueryParam {
+  char            *sql;
+  char            *tagCond;
+  char            *colCond;
+  char            *tbnameCond;
+  char            *prevResult;
+  SArray          *pTableIdList;
+  SSqlExpr       **pExpr;
+  SSqlExpr       **pSecExpr;
+  SExprInfo       *pExprs;
+  SExprInfo       *pSecExprs;
+
+  SFilterInfo     *pFilters;
+
+  SColIndex       *pGroupColIndex;
+  SColumnInfo     *pTagColumnInfo;
+  SGroupbyExpr *pGroupbyExpr;
+  int32_t          tableScanOperator;
+  SArray          *pOperator;
+  struct SUdfInfo        *pUdfInfo;
+} SQueryParam;
+
+typedef struct STableScanInfo {
+  void           *pQueryHandle;
+  int32_t         numOfBlocks;
+  int32_t         numOfSkipped;
+  int32_t         numOfBlockStatis;
+  int64_t         numOfRows;
+                 
+  int32_t         order;        // scan order
+  int32_t         times;        // repeat counts
+  int32_t         current;
+  int32_t         reverseTimes; // 0 by default
+
+  SQLFunctionCtx *pCtx;         // next operator query context
+  SResultRowInfo *pResultRowInfo;
+  int32_t        *rowCellInfoOffset;
+  SExprInfo      *pExpr;
+  SSDataBlock     block;
+  int32_t         numOfOutput;
+  int64_t         elapsedTime;
+
+  int32_t         tableIndex;
+  int32_t         prevGroupId;     // previous table group id
+} STableScanInfo;
+
+typedef struct STagScanInfo {
+  SColumnInfo* pCols;
+  SSDataBlock* pRes;
+  int32_t      totalTables;
+  int32_t      curPos;
+} STagScanInfo;
+
+typedef struct SOptrBasicInfo {
+  SResultRowInfo    resultRowInfo;
+  int32_t          *rowCellInfoOffset;  // offset value for each row result cell info
+  SQLFunctionCtx   *pCtx;
+  SSDataBlock      *pRes;
+} SOptrBasicInfo;
+
+typedef struct SOptrBasicInfo STableIntervalOperatorInfo;
+
+typedef struct SAggOperatorInfo {
+  SOptrBasicInfo binfo;
+  uint32_t       seed;
+} SAggOperatorInfo;
+
+typedef struct SProjectOperatorInfo {
+  SOptrBasicInfo binfo;
+  int32_t        bufCapacity;
+  uint32_t       seed;
+
+  SSDataBlock   *existDataBlock;
+} SProjectOperatorInfo;
+
+typedef struct SLimitOperatorInfo {
+  int64_t   limit;
+  int64_t   total;
+} SLimitOperatorInfo;
+
+typedef struct SSLimitOperatorInfo {
+  int64_t      groupTotal;
+  int64_t      currentGroupOffset;
+
+  int64_t      rowsTotal;
+  int64_t      currentOffset;
+  SLimit       limit;
+  SLimit       slimit;
+
+  char       **prevRow;
+  SArray      *orderColumnList;
+  bool         hasPrev;
+  bool         ignoreCurrentGroup;
+  bool         multigroupResult;
+  SSDataBlock *pRes;   // result buffer
+  SSDataBlock *pPrevBlock;
+  int64_t      capacity;
+  int64_t      threshold;
+} SSLimitOperatorInfo;
+
+typedef struct SFilterOperatorInfo {
+  SSingleColumnFilterInfo *pFilterInfo;
+  int32_t numOfFilterCols;
+} SFilterOperatorInfo;
+
+typedef struct SFillOperatorInfo {
+  struct SFillInfo   *pFillInfo;
+  SSDataBlock *pRes;
+  int64_t      totalInputRows;
+  void       **p;
+  SSDataBlock *existNewGroupBlock;
+  bool         multigroupResult;
+} SFillOperatorInfo;
+
+typedef struct SGroupbyOperatorInfo {
+  SOptrBasicInfo binfo;
+  int32_t        colIndex;
+  char          *prevData;   // previous group by value
+} SGroupbyOperatorInfo;
+
+typedef struct SSWindowOperatorInfo {
+  SOptrBasicInfo binfo;
+  STimeWindow    curWindow;  // current time window
+  TSKEY          prevTs;     // previous timestamp
+  int32_t        numOfRows;  // number of rows
+  int32_t        start;      // start row index
+  bool           reptScan;    // next round scan
+} SSWindowOperatorInfo;
+
+typedef struct SStateWindowOperatorInfo {
+  SOptrBasicInfo binfo;
+  STimeWindow    curWindow;  // current time window
+  int32_t        numOfRows;  // number of rows
+  int32_t        colIndex;      // start row index
+  int32_t        start;
+  char*          prevData;    // previous data 
+  bool           reptScan;
+} SStateWindowOperatorInfo;
+
+typedef struct SDistinctDataInfo {
+  int32_t index;
+  int32_t type;
+  int32_t bytes;
+} SDistinctDataInfo; 
+
+typedef struct SDistinctOperatorInfo {
+  SHashObj         *pSet;
+  SSDataBlock      *pRes;
+  bool              recordNullVal;  //has already record the null value, no need to try again
+  int64_t           threshold;
+  int64_t           outputCapacity;
+  int32_t           totalBytes; 
+  char*             buf;
+  SArray*           pDistinctDataInfo; 
+} SDistinctOperatorInfo;
+
+struct SGlobalMerger;
+
+typedef struct SMultiwayMergeInfo {
+  struct SGlobalMerger *pMerge;
+  SOptrBasicInfo       binfo;
+  int32_t              bufCapacity;
+  int64_t              seed;
+  char               **prevRow;
+  SArray              *orderColumnList;
+  int32_t              resultRowFactor;
+
+  bool                 hasGroupColData;
+  char               **currentGroupColData;
+  SArray              *groupColumnList;
+  bool                 hasDataBlockForNewGroup;
+  SSDataBlock         *pExistBlock;
+
+  SArray              *udfInfo;
+  bool                 hasPrev;
+  bool                 multiGroupResults;
+} SMultiwayMergeInfo;
+
+// todo support the disk-based sort
+typedef struct SOrderOperatorInfo {
+  int32_t      colIndex;
+  int32_t      order;
+  SSDataBlock *pDataBlock;
+} SOrderOperatorInfo;
+
+void appendUpstream(SOperatorInfo* p, SOperatorInfo* pUpstream);
+
+SOperatorInfo* createDataBlocksOptScanInfo(void* pTsdbQueryHandle, SQueryRuntimeEnv* pRuntimeEnv, int32_t repeatTime, int32_t reverseTime);
+SOperatorInfo* createTableScanOperator(void* pTsdbQueryHandle, SQueryRuntimeEnv* pRuntimeEnv, int32_t repeatTime);
+SOperatorInfo* createTableSeqScanOperator(void* pTsdbQueryHandle, SQueryRuntimeEnv* pRuntimeEnv);
+
+SOperatorInfo* createAggregateOperatorInfo(SQueryRuntimeEnv* pRuntimeEnv, SOperatorInfo* upstream, SExprInfo* pExpr, int32_t numOfOutput);
+SOperatorInfo* createProjectOperatorInfo(SQueryRuntimeEnv* pRuntimeEnv, SOperatorInfo* upstream, SExprInfo* pExpr, int32_t numOfOutput);
+SOperatorInfo* createLimitOperatorInfo(SQueryRuntimeEnv* pRuntimeEnv, SOperatorInfo* upstream);
+SOperatorInfo* createTimeIntervalOperatorInfo(SQueryRuntimeEnv* pRuntimeEnv, SOperatorInfo* upstream, SExprInfo* pExpr, int32_t numOfOutput);
+SOperatorInfo* createAllTimeIntervalOperatorInfo(SQueryRuntimeEnv* pRuntimeEnv, SOperatorInfo* upstream, SExprInfo* pExpr, int32_t numOfOutput);
+SOperatorInfo* createSWindowOperatorInfo(SQueryRuntimeEnv* pRuntimeEnv, SOperatorInfo* upstream, SExprInfo* pExpr, int32_t numOfOutput);
+SOperatorInfo* createFillOperatorInfo(SQueryRuntimeEnv* pRuntimeEnv, SOperatorInfo* upstream, SExprInfo* pExpr, int32_t numOfOutput, bool multigroupResult);
+SOperatorInfo* createGroupbyOperatorInfo(SQueryRuntimeEnv* pRuntimeEnv, SOperatorInfo* upstream, SExprInfo* pExpr, int32_t numOfOutput);
+SOperatorInfo* createMultiTableAggOperatorInfo(SQueryRuntimeEnv* pRuntimeEnv, SOperatorInfo* upstream, SExprInfo* pExpr, int32_t numOfOutput);
+SOperatorInfo* createMultiTableTimeIntervalOperatorInfo(SQueryRuntimeEnv* pRuntimeEnv, SOperatorInfo* upstream, SExprInfo* pExpr, int32_t numOfOutput);
+SOperatorInfo* createAllMultiTableTimeIntervalOperatorInfo(SQueryRuntimeEnv* pRuntimeEnv, SOperatorInfo* upstream, SExprInfo* pExpr, int32_t numOfOutput);
+SOperatorInfo* createTagScanOperatorInfo(SQueryRuntimeEnv* pRuntimeEnv, SExprInfo* pExpr, int32_t numOfOutput);
+SOperatorInfo* createDistinctOperatorInfo(SQueryRuntimeEnv* pRuntimeEnv, SOperatorInfo* upstream, SExprInfo* pExpr, int32_t numOfOutput);
+SOperatorInfo* createTableBlockInfoScanOperator(void* pTsdbQueryHandle, SQueryRuntimeEnv* pRuntimeEnv);
+SOperatorInfo* createMultiwaySortOperatorInfo(SQueryRuntimeEnv* pRuntimeEnv, SExprInfo* pExpr, int32_t numOfOutput,
+                                              int32_t numOfRows, void* merger);
+SOperatorInfo* createGlobalAggregateOperatorInfo(SQueryRuntimeEnv* pRuntimeEnv, SOperatorInfo* upstream, SExprInfo* pExpr, int32_t numOfOutput, void* param, SArray* pUdfInfo, bool groupResultMixedUp);
+SOperatorInfo* createStatewindowOperatorInfo(SQueryRuntimeEnv* pRuntimeEnv, SOperatorInfo* upstream, SExprInfo* pExpr, int32_t numOfOutput);
+SOperatorInfo* createSLimitOperatorInfo(SQueryRuntimeEnv* pRuntimeEnv, SOperatorInfo* upstream, SExprInfo* pExpr, int32_t numOfOutput, void* merger, bool multigroupResult);
+SOperatorInfo* createFilterOperatorInfo(SQueryRuntimeEnv* pRuntimeEnv, SOperatorInfo* upstream, SExprInfo* pExpr,
+                                        int32_t numOfOutput, SColumnInfo* pCols, int32_t numOfFilter);
+
+SOperatorInfo* createJoinOperatorInfo(SOperatorInfo** pUpstream, int32_t numOfUpstream, SSchema* pSchema, int32_t numOfOutput);
+SOperatorInfo* createOrderOperatorInfo(SQueryRuntimeEnv* pRuntimeEnv, SOperatorInfo* upstream, SExprInfo* pExpr, int32_t numOfOutput, SOrder* pOrderVal);
+
+SSDataBlock* doGlobalAggregate(void* param, bool* newgroup);
+SSDataBlock* doMultiwayMergeSort(void* param, bool* newgroup);
+SSDataBlock* doSLimit(void* param, bool* newgroup);
+
+int32_t doCreateFilterInfo(SColumnInfo* pCols, int32_t numOfCols, int32_t numOfFilterCols, SSingleColumnFilterInfo** pFilterInfo, uint64_t qId);
+void doSetFilterColumnInfo(SSingleColumnFilterInfo* pFilterInfo, int32_t numOfFilterCols, SSDataBlock* pBlock);
+bool doFilterDataBlock(SSingleColumnFilterInfo* pFilterInfo, int32_t numOfFilterCols, int32_t numOfRows, int8_t* p);
+void doCompactSDataBlock(SSDataBlock* pBlock, int32_t numOfRows, int8_t* p);
+
+SSDataBlock* createOutputBuf(SExprInfo* pExpr, int32_t numOfOutput, int32_t numOfRows);
+
+void* destroyOutputBuf(SSDataBlock* pBlock);
+void* doDestroyFilterInfo(SSingleColumnFilterInfo* pFilterInfo, int32_t numOfFilterCols);
+
+void setInputDataBlock(SOperatorInfo* pOperator, SQLFunctionCtx* pCtx, SSDataBlock* pBlock, int32_t order);
+void finalizeQueryResult(SOperatorInfo* pOperator, SQLFunctionCtx* pCtx, SResultRowInfo* pResultRowInfo, int32_t* rowCellInfoOffset);
+void updateOutputBuf(SOptrBasicInfo* pBInfo, int32_t *bufCapacity, int32_t numOfInputRows);
+void clearOutputBuf(SOptrBasicInfo* pBInfo, int32_t *bufCapacity);
+void copyTsColoum(SSDataBlock* pRes, SQLFunctionCtx* pCtx, int32_t numOfOutput);
+
+void freeParam(SQueryParam *param);
+int32_t convertQueryMsg(SQueryTableMsg *pQueryMsg, SQueryParam* param);
+int32_t createQueryFunc(SQueriedTableInfo* pTableInfo, int32_t numOfOutput, SExprInfo** pExprInfo,
+                        SSqlExpr** pExprMsg, SColumnInfo* pTagCols, int32_t queryType, void* pMsg, struct SUdfInfo* pUdfInfo);
+
+int32_t createIndirectQueryFuncExprFromMsg(SQueryTableMsg *pQueryMsg, int32_t numOfOutput, SExprInfo **pExprInfo,
+                                           SSqlExpr **pExpr, SExprInfo *prevExpr, struct SUdfInfo *pUdfInfo);
+
+int32_t createQueryFilter(char *data, uint16_t len, SFilterInfo** pFilters);
+
+SGroupbyExpr *createGroupbyExprFromMsg(SQueryTableMsg *pQueryMsg, SColIndex *pColIndex, int32_t *code);
+SQInfo *createQInfoImpl(SQueryTableMsg *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, SQueryParam* param, char* start,
+                  int32_t prevResultLen, void* merger);
+
+int32_t createFilterInfo(SQueryAttr* pQueryAttr, uint64_t qId);
+void freeColumnFilterInfo(SColumnFilterInfo* pFilter, int32_t numOfFilters);
+
+STableQueryInfo *createTableQueryInfo(SQueryAttr* pQueryAttr, void* pTable, bool groupbyColumn, STimeWindow win, void* buf);
+STableQueryInfo* createTmpTableQueryInfo(STimeWindow win);
+
+int32_t buildArithmeticExprFromMsg(SExprInfo *pArithExprInfo, void *pQueryMsg);
+
+bool isQueryKilled(SQInfo *pQInfo);
+int32_t checkForQueryBuf(size_t numOfTables);
+bool checkNeedToCompressQueryCol(SQInfo *pQInfo);
+bool doBuildResCheck(SQInfo* pQInfo);
+void setQueryStatus(SQueryRuntimeEnv *pRuntimeEnv, int8_t status);
+
+bool onlyQueryTags(SQueryAttr* 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 publishOperatorProfEvent(SOperatorInfo* operatorInfo, EQueryProfEventType eventType);
+void publishQueryAbortEvent(SQInfo* pQInfo, int32_t code);
+void calculateOperatorProfResults(SQInfo* pQInfo);
+void queryCostStatis(SQInfo *pQInfo);
+
+void freeQInfo(SQInfo *pQInfo);
+void freeQueryAttr(SQueryAttr *pQuery);
+
+int32_t getMaximumIdleDurationSec();
+
+void doInvokeUdf(struct SUdfInfo* pUdfInfo, SQLFunctionCtx *pCtx, int32_t idx, int32_t type);
+
+#endif  // TDENGINE_EXECUTORIMPL_H

source/libs/executor/inc/tfilter.h

@@ -20,9 +20,9 @@
 extern "C" {
 #endif
 
-#include "texpr.h"
+#include "thash.h"
-#include "hash.h"
 #include "tname.h"
+#include "function.h"
 
 #define FILTER_DEFAULT_GROUP_SIZE 4
 #define FILTER_DEFAULT_UNIT_SIZE 4
@@ -105,7 +105,7 @@ typedef struct SFilterColRange {
 
 typedef bool (*rangeCompFunc) (const void *, const void *, const void *, const void *, __compar_fn_t);
 typedef int32_t(*filter_desc_compare_func)(const void *, const void *);
-typedef bool(*filter_exec_func)(void *, int32_t, int8_t**, SDataStatis *, int16_t);
+typedef bool(*filter_exec_func)(void *, int32_t, int8_t**, SColumnDataAgg *, int16_t);
 
 typedef struct SFilterRangeCompare {
   int64_t s;
@@ -324,13 +324,13 @@ typedef struct SFilterInfo {
 
 
 extern int32_t filterInitFromTree(tExprNode* tree, SFilterInfo **pinfo, uint32_t options);
-extern bool filterExecute(SFilterInfo *info, int32_t numOfRows, int8_t** p, SDataStatis *statis, int16_t numOfCols);
+extern bool filterExecute(SFilterInfo *info, int32_t numOfRows, int8_t** p, SColumnDataAgg *statis, int16_t numOfCols);
 extern int32_t filterSetColFieldData(SFilterInfo *info, int32_t numOfCols, SArray* pDataBlock);
 extern int32_t filterGetTimeRange(SFilterInfo *info, STimeWindow *win);
 extern int32_t filterConverNcharColumns(SFilterInfo* pFilterInfo, int32_t rows, bool *gotNchar);
 extern int32_t filterFreeNcharColumns(SFilterInfo* pFilterInfo);
 extern void filterFreeInfo(SFilterInfo *info);
-extern bool filterRangeExecute(SFilterInfo *info, SDataStatis *pDataStatis, int32_t numOfCols, int32_t numOfRows);
+extern bool filterRangeExecute(SFilterInfo *info, SColumnDataAgg *pDataStatis, int32_t numOfCols, int32_t numOfRows);
 
 #ifdef __cplusplus
 }

source/libs/executor/src/executil.c

@@ -15,11 +15,11 @@
 
 #include "os.h"
 #include "taosmsg.h"
-#include "hash.h"
+#include "thash.h"
 
-#include "qExecutor.h"
+#include "executil.h"
-#include "qUtil.h"
+#include "executorimpl.h"
-#include "queryLog.h"
+//#include "queryLog.h"
 #include "tbuffer.h"
 #include "tcompression.h"
 #include "tlosertree.h"
@@ -33,9 +33,9 @@ typedef struct SCompSupporter {
 int32_t getRowNumForMultioutput(SQueryAttr* pQueryAttr, bool topBottomQuery, bool stable) {
   if (pQueryAttr && (!stable)) {
     for (int16_t i = 0; i < pQueryAttr->numOfOutput; ++i) {
-      if (pQueryAttr->pExpr1[i].base.functionId == TSDB_FUNC_TOP || pQueryAttr->pExpr1[i].base.functionId == TSDB_FUNC_BOTTOM) {
+//      if (pQueryAttr->pExpr1[i].base. == FUNCTION_TOP || pQueryAttr->pExpr1[i].base.functionId == FUNCTION_BOTTOM) {
-        return (int32_t)pQueryAttr->pExpr1[i].base.param[0].i64;
+//        return (int32_t)pQueryAttr->pExpr1[i].base.param[0].i;
-      }
+//      }
     }
   }
 
@@ -143,18 +143,18 @@ void clearResultRow(SQueryRuntimeEnv *pRuntimeEnv, SResultRow *pResultRow, int16
 
   // the result does not put into the SDiskbasedResultBuf, ignore it.
   if (pResultRow->pageId >= 0) {
-    tFilePage *page = getResBufPage(pRuntimeEnv->pResultBuf, pResultRow->pageId);
+    SFilePage *page = getResBufPage(pRuntimeEnv->pResultBuf, pResultRow->pageId);
 
     int16_t offset = 0;
     for (int32_t i = 0; i < pRuntimeEnv->pQueryAttr->numOfOutput; ++i) {
-      SResultRowCellInfo *pResultInfo = &pResultRow->pCellInfo[i];
+      struct SResultRowEntryInfo *pEntryInfo = NULL;//pResultRow->pEntryInfo[i];
 
-      int16_t size = pRuntimeEnv->pQueryAttr->pExpr1[i].base.resType;
+      int16_t size = pRuntimeEnv->pQueryAttr->pExpr1[i].base.resSchema.bytes;
       char * s = getPosInResultPage(pRuntimeEnv->pQueryAttr, page, pResultRow->offset, offset);
       memset(s, 0, size);
 
       offset += size;
-      RESET_RESULT_INFO(pResultInfo);
+      cleanupResultRowEntry(pEntryInfo);
     }
   }
 
@@ -168,14 +168,16 @@ void clearResultRow(SQueryRuntimeEnv *pRuntimeEnv, SResultRow *pResultRow, int16
 }
 
 // TODO refactor: use macro
-SResultRowCellInfo* getResultCell(const SResultRow* pRow, int32_t index, int32_t* offset) {
+struct SResultRowEntryInfo* getResultCell(const SResultRow* pRow, int32_t index, int32_t* offset) {
   assert(index >= 0 && offset != NULL);
-  return (SResultRowCellInfo*)((char*) pRow->pCellInfo + offset[index]);
+//  return (SResultRowEntryInfo*)((char*) pRow->pCellInfo + offset[index]);
+return NULL;
 }
 
 size_t getResultRowSize(SQueryRuntimeEnv* pRuntimeEnv) {
   SQueryAttr* pQueryAttr = pRuntimeEnv->pQueryAttr;
-  return (pQueryAttr->numOfOutput * sizeof(SResultRowCellInfo)) + pQueryAttr->interBufSize + sizeof(SResultRow);
+  return 0;
+//  return (pQueryAttr->numOfOutput * sizeof(SResultRowEntryInfo)) + pQueryAttr->interBufSize + sizeof(SResultRow);
 }
 
 SResultRowPool* initResultRowPool(size_t size) {
@@ -271,9 +273,9 @@ void interResToBinary(SBufferWriter* bw, SArray* pRes, int32_t tagLen) {
       tbufWriteUint32(bw, numOfRows);
 
       for(int32_t k = 0; k < numOfRows; ++k) {
-        SResPair v = *(SResPair*) taosArrayGet(p->pResult, k);
+//        SResPair v = *(SResPair*) taosArrayGet(p->pResult, k);
-        tbufWriteDouble(bw, v.avg);
+//        tbufWriteDouble(bw, v.avg);
-        tbufWriteInt64(bw, v.key);
+//        tbufWriteInt64(bw, v.key);
       }
     }
   }
@@ -301,19 +303,19 @@ SArray* interResFromBinary(const char* data, int32_t len) {
 
     SArray* p = taosArrayInit(numOfCols, sizeof(SStddevInterResult));
     for(int32_t j = 0; j < numOfCols; ++j) {
-      int16_t colId = tbufReadUint16(&br);
+//      int16_t colId = tbufReadUint16(&br);
       int32_t numOfRows = tbufReadUint32(&br);
 
-      SStddevInterResult interRes = {.colId = colId, .pResult = taosArrayInit(4, sizeof(struct SResPair)),};
+//      SStddevInterResult interRes = {.colId = colId, .pResult = taosArrayInit(4, sizeof(struct SResPair)),};
       for(int32_t k = 0; k < numOfRows; ++k) {
-        SResPair px = {0};
+//        SResPair px = {0};
-        px.avg = tbufReadDouble(&br);
+//        px.avg = tbufReadDouble(&br);
-        px.key = tbufReadInt64(&br);
+//        px.key = tbufReadInt64(&br);
-
+//
-        taosArrayPush(interRes.pResult, &px);
+//        taosArrayPush(interRes.pResult, &px);
       }
 
-      taosArrayPush(p, &interRes);
+//      taosArrayPush(p, &interRes);
     }
 
     char* p1 = NULL;
@@ -395,22 +397,22 @@ static int64_t getNumOfResultWindowRes(SQueryRuntimeEnv* pRuntimeEnv, SResultRow
   SQueryAttr* pQueryAttr = pRuntimeEnv->pQueryAttr;
 
   for (int32_t j = 0; j < pQueryAttr->numOfOutput; ++j) {
-    int32_t functionId = pQueryAttr->pExpr1[j].base.functionId;
+    int32_t functionId = 0;//pQueryAttr->pExpr1[j].base.functionId;
 
     /*
      * ts, tag, tagprj function can not decide the output number of current query
      * the number of output result is decided by main output
      */
-    if (functionId == TSDB_FUNC_TS || functionId == TSDB_FUNC_TAG || functionId == TSDB_FUNC_TAGPRJ) {
+    if (functionId == FUNCTION_TS || functionId == FUNCTION_TAG || functionId == FUNCTION_TAGPRJ) {
       continue;
     }
 
-    SResultRowCellInfo *pResultInfo = getResultCell(pResultRow, j, rowCellInfoOffset);
+//    SResultRowEntryInfo *pResultInfo = getResultCell(pResultRow, j, rowCellInfoOffset);
-    assert(pResultInfo != NULL);
+//    assert(pResultInfo != NULL);
-
+//
-    if (pResultInfo->numOfRes > 0) {
+//    if (pResultInfo->numOfRes > 0) {
-      return pResultInfo->numOfRes;
+//      return pResultInfo->numOfRes;
-    }
+//    }
   }
 
   return 0;
@@ -545,7 +547,7 @@ static UNUSED_FUNC int32_t mergeIntoGroupResultImpl(SQueryRuntimeEnv *pRuntimeEn
   pTableQueryInfoList = malloc(POINTER_BYTES * size);
 
   if (pTableQueryInfoList == NULL || posList == NULL || pGroupResInfo->pRows == NULL || pGroupResInfo->pRows == NULL) {
-    qError("QInfo:%"PRIu64" failed alloc memory", GET_QID(pRuntimeEnv));
+//    qError("QInfo:%"PRIu64" failed alloc memory", GET_QID(pRuntimeEnv));
     code = TSDB_CODE_QRY_OUT_OF_MEMORY;
     goto _end;
   }
@@ -617,8 +619,8 @@ static UNUSED_FUNC int32_t mergeIntoGroupResultImpl(SQueryRuntimeEnv *pRuntimeEn
 
   int64_t endt = taosGetTimestampMs();
 
-  qDebug("QInfo:%"PRIx64" result merge completed for group:%d, elapsed time:%" PRId64 " ms", GET_QID(pRuntimeEnv),
+//  qDebug("QInfo:%"PRIx64" result merge completed for group:%d, elapsed time:%" PRId64 " ms", GET_QID(pRuntimeEnv),
-         pGroupResInfo->currentGroup, endt - startt);
+//         pGroupResInfo->currentGroup, endt - startt);
 
   _end:
   tfree(pTableQueryInfoList);
@@ -639,90 +641,90 @@ int32_t mergeIntoGroupResult(SGroupResInfo* pGroupResInfo, SQueryRuntimeEnv* pRu
       break;
     }
 
-    qDebug("QInfo:%"PRIu64" no result in group %d, continue", GET_QID(pRuntimeEnv), pGroupResInfo->currentGroup);
+//    qDebug("QInfo:%"PRIu64" no result in group %d, continue", GET_QID(pRuntimeEnv), pGroupResInfo->currentGroup);
     cleanupGroupResInfo(pGroupResInfo);
     incNextGroup(pGroupResInfo);
   }
 
-  int64_t elapsedTime = taosGetTimestampUs() - st;
+//  int64_t elapsedTime = taosGetTimestampUs() - st;
-  qDebug("QInfo:%"PRIu64" merge res data into group, index:%d, total group:%d, elapsed time:%" PRId64 "us", GET_QID(pRuntimeEnv),
+//  qDebug("QInfo:%"PRIu64" merge res data into group, index:%d, total group:%d, elapsed time:%" PRId64 "us", GET_QID(pRuntimeEnv),
-         pGroupResInfo->currentGroup, pGroupResInfo->totalGroup, elapsedTime);
+//         pGroupResInfo->currentGroup, pGroupResInfo->totalGroup, elapsedTime);
 
   return TSDB_CODE_SUCCESS;
 }
 
-void blockDistInfoToBinary(STableBlockDist* pDist, struct SBufferWriter* bw) {
+//void blockDistInfoToBinary(STableBlockDist* pDist, struct SBufferWriter* bw) {
-  tbufWriteUint32(bw, pDist->numOfTables);
+//  tbufWriteUint32(bw, pDist->numOfTables);
-  tbufWriteUint16(bw, pDist->numOfFiles);
+//  tbufWriteUint16(bw, pDist->numOfFiles);
-  tbufWriteUint64(bw, pDist->totalSize);
+//  tbufWriteUint64(bw, pDist->totalSize);
-  tbufWriteUint64(bw, pDist->totalRows);
+//  tbufWriteUint64(bw, pDist->totalRows);
-  tbufWriteInt32(bw, pDist->maxRows);
+//  tbufWriteInt32(bw, pDist->maxRows);
-  tbufWriteInt32(bw, pDist->minRows);
+//  tbufWriteInt32(bw, pDist->minRows);
-  tbufWriteUint32(bw, pDist->numOfRowsInMemTable);
+//  tbufWriteUint32(bw, pDist->numOfRowsInMemTable);
-  tbufWriteUint32(bw, pDist->numOfSmallBlocks);
+//  tbufWriteUint32(bw, pDist->numOfSmallBlocks);
-  tbufWriteUint64(bw, taosArrayGetSize(pDist->dataBlockInfos));
+//  tbufWriteUint64(bw, taosArrayGetSize(pDist->dataBlockInfos));
+//
+//  // compress the binary string
+//  char* p = TARRAY_GET_START(pDist->dataBlockInfos);
+//
+//  // compress extra bytes
+//  size_t x = taosArrayGetSize(pDist->dataBlockInfos) * pDist->dataBlockInfos->elemSize;
+//  char* tmp = malloc(x + 2);
+//
+//  bool comp = false;
+//  int32_t len = tsCompressString(p, (int32_t)x, 1, tmp, (int32_t)x, ONE_STAGE_COMP, NULL, 0);
+//  if (len == -1 || len >= x) { // compress failed, do not compress this binary data
+//    comp = false;
+//    len = (int32_t)x;
+//  } else {
+//    comp = true;
+//  }
+//
+//  tbufWriteUint8(bw, comp);
+//  tbufWriteUint32(bw, len);
+//  if (comp) {
+//    tbufWriteBinary(bw, tmp, len);
+//  } else {
+//    tbufWriteBinary(bw, p, len);
+//  }
+//  tfree(tmp);
+//}
 
-  // compress the binary string
+//void blockDistInfoFromBinary(const char* data, int32_t len, STableBlockDist* pDist) {
-  char* p = TARRAY_GET_START(pDist->dataBlockInfos);
+//  SBufferReader br = tbufInitReader(data, len, false);
-
+//
-  // compress extra bytes
+//  pDist->numOfTables = tbufReadUint32(&br);
-  size_t x = taosArrayGetSize(pDist->dataBlockInfos) * pDist->dataBlockInfos->elemSize;
+//  pDist->numOfFiles  = tbufReadUint16(&br);
-  char* tmp = malloc(x + 2);
+//  pDist->totalSize   = tbufReadUint64(&br);
-
+//  pDist->totalRows   = tbufReadUint64(&br);
-  bool comp = false;
+//  pDist->maxRows     = tbufReadInt32(&br);
-  int32_t len = tsCompressString(p, (int32_t)x, 1, tmp, (int32_t)x, ONE_STAGE_COMP, NULL, 0);
+//  pDist->minRows     = tbufReadInt32(&br);
-  if (len == -1 || len >= x) { // compress failed, do not compress this binary data
+//  pDist->numOfRowsInMemTable = tbufReadUint32(&br);
-    comp = false;
+//  pDist->numOfSmallBlocks = tbufReadUint32(&br);
-    len = (int32_t)x;
+//  int64_t numSteps = tbufReadUint64(&br);
-  } else {
+//
-    comp = true;
+//  bool comp = tbufReadUint8(&br);
-  }
+//  uint32_t compLen = tbufReadUint32(&br);
-
+//
-  tbufWriteUint8(bw, comp);
+//  size_t originalLen = (size_t) (numSteps *sizeof(SFileBlockInfo));
-  tbufWriteUint32(bw, len);
+//
-  if (comp) {
+//  char* outputBuf = NULL;
-    tbufWriteBinary(bw, tmp, len);
+//  if (comp) {
-  } else {
+//    outputBuf = malloc(originalLen);
-    tbufWriteBinary(bw, p, len);
+//
-  }
+//    size_t actualLen = compLen;
-  tfree(tmp);
+//    const char* compStr = tbufReadBinary(&br, &actualLen);
-}
+//
-
+//    int32_t orignalLen = tsDecompressString(compStr, compLen, 1, outputBuf,
-void blockDistInfoFromBinary(const char* data, int32_t len, STableBlockDist* pDist) {
+//                                            (int32_t)originalLen , ONE_STAGE_COMP, NULL, 0);
-  SBufferReader br = tbufInitReader(data, len, false);
+//    assert(orignalLen == numSteps *sizeof(SFileBlockInfo));
-
+//  } else {
-  pDist->numOfTables = tbufReadUint32(&br);
+//    outputBuf = (char*) tbufReadBinary(&br, &originalLen);
-  pDist->numOfFiles  = tbufReadUint16(&br);
+//  }
-  pDist->totalSize   = tbufReadUint64(&br);
+//
-  pDist->totalRows   = tbufReadUint64(&br);
+//  pDist->dataBlockInfos = taosArrayFromList(outputBuf, (uint32_t)numSteps, sizeof(SFileBlockInfo));
-  pDist->maxRows     = tbufReadInt32(&br);
+//  if (comp) {
-  pDist->minRows     = tbufReadInt32(&br);
+//    tfree(outputBuf);
-  pDist->numOfRowsInMemTable = tbufReadUint32(&br);
+//  }
-  pDist->numOfSmallBlocks = tbufReadUint32(&br);
+//}
-  int64_t numSteps = tbufReadUint64(&br);
-
-  bool comp = tbufReadUint8(&br);
-  uint32_t compLen = tbufReadUint32(&br);
-
-  size_t originalLen = (size_t) (numSteps *sizeof(SFileBlockInfo));
-
-  char* outputBuf = NULL;
-  if (comp) {
-    outputBuf = malloc(originalLen);
-
-    size_t actualLen = compLen;
-    const char* compStr = tbufReadBinary(&br, &actualLen);
-
-    int32_t orignalLen = tsDecompressString(compStr, compLen, 1, outputBuf,
-                                            (int32_t)originalLen , ONE_STAGE_COMP, NULL, 0);
-    assert(orignalLen == numSteps *sizeof(SFileBlockInfo));
-  } else {
-    outputBuf = (char*) tbufReadBinary(&br, &originalLen);
-  }
-
-  pDist->dataBlockInfos = taosArrayFromList(outputBuf, (uint32_t)numSteps, sizeof(SFileBlockInfo));
-  if (comp) {
-    tfree(outputBuf);
-  }
-}
 

source/libs/executor/src/tarithoperator.c

@@ -1,411 +0,0 @@
-/*
- * Copyright (c) 2019 TAOS Data, Inc. <jhtao@taosdata.com>
- *
- * This program is free software: you can use, redistribute, and/or modify
- * it under the terms of the GNU Affero General Public License, version 3
- * or later ("AGPL"), as published by the Free Software Foundation.
- *
- * This program is distributed in the hope that it will be useful, but WITHOUT
- * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
- * FITNESS FOR A PARTICULAR PURPOSE.
- *
- * You should have received a copy of the GNU Affero General Public License
- * along with this program. If not, see <http://www.gnu.org/licenses/>.
- */
-
-#include "os.h"
-
-#include "ttypes.h"
-#include "tarithoperator.h"
-#include "tcompare.h"
-
-//GET_TYPED_DATA(v, double, _right_type, (char *)&((right)[i]));                                
-
-void calc_i32_i32_add(void *left, void *right, int32_t numLeft, int32_t numRight, void *output, int32_t order) {
-  int32_t *pLeft = (int32_t *)left;
-  int32_t *pRight = (int32_t *)right;
-  double * pOutput = (double *)output;
-
-  int32_t i = (order == TSDB_ORDER_ASC) ? 0 : MAX(numLeft, numRight) - 1;
-  int32_t step = (order == TSDB_ORDER_ASC) ? 1 : -1;
-
-  if (numLeft == numRight) {
-    for (; i >= 0 && i < numRight; i += step, pOutput += 1) {
-      if (isNull((char *)&(pLeft[i]), TSDB_DATA_TYPE_INT) || isNull((char *)&(pRight[i]), TSDB_DATA_TYPE_INT)) {
-        SET_DOUBLE_NULL(pOutput);
-        continue;
-      }
-
-      *pOutput = (double)pLeft[i] + pRight[i];
-    }
-  } else if (numLeft == 1) {
-    for (; i >= 0 && i < numRight; i += step, pOutput += 1) {
-      if (isNull((char *)(pLeft), TSDB_DATA_TYPE_INT) || isNull((char *)&(pRight[i]), TSDB_DATA_TYPE_INT)) {
-        SET_DOUBLE_NULL(pOutput);
-        continue;
-      }
-
-      *pOutput = (double)pLeft[0] + pRight[i];
-    }
-  } else if (numRight == 1) {
-    for (; i >= 0 && i < numLeft; i += step, pOutput += 1) {
-      if (isNull((char *)&(pLeft[i]), TSDB_DATA_TYPE_INT) || isNull((char *)(pRight), TSDB_DATA_TYPE_INT)) {
-        SET_DOUBLE_NULL(pOutput);
-        continue;
-      }
-      *pOutput = (double)pLeft[i] + pRight[0];
-    }
-  }
-}
-
-typedef double (*_arithmetic_getVectorDoubleValue_fn_t)(void *src, int32_t index);
-
-double getVectorDoubleValue_TINYINT(void *src, int32_t index) {
-  return (double)*((int8_t *)src + index);
-}
-double getVectorDoubleValue_UTINYINT(void *src, int32_t index) {
-  return (double)*((uint8_t *)src + index);
-}
-double getVectorDoubleValue_SMALLINT(void *src, int32_t index) {
-  return (double)*((int16_t *)src + index);
-}
-double getVectorDoubleValue_USMALLINT(void *src, int32_t index) {
-  return (double)*((uint16_t *)src + index);
-}
-double getVectorDoubleValue_INT(void *src, int32_t index) {
-  return (double)*((int32_t *)src + index);
-}
-double getVectorDoubleValue_UINT(void *src, int32_t index) {
-  return (double)*((uint32_t *)src + index);
-}
-double getVectorDoubleValue_BIGINT(void *src, int32_t index) {
-  return (double)*((int64_t *)src + index);
-}
-double getVectorDoubleValue_UBIGINT(void *src, int32_t index) {
-  return (double)*((uint64_t *)src + index);
-}
-double getVectorDoubleValue_FLOAT(void *src, int32_t index) {
-  return (double)*((float *)src + index);
-}
-double getVectorDoubleValue_DOUBLE(void *src, int32_t index) {
-  return (double)*((double *)src + index);
-}
-_arithmetic_getVectorDoubleValue_fn_t getVectorDoubleValueFn(int32_t srcType) {
-    _arithmetic_getVectorDoubleValue_fn_t p = NULL;
-    if(srcType==TSDB_DATA_TYPE_TINYINT) {
-        p = getVectorDoubleValue_TINYINT;
-    }else if(srcType==TSDB_DATA_TYPE_UTINYINT) {
-        p = getVectorDoubleValue_UTINYINT;
-    }else if(srcType==TSDB_DATA_TYPE_SMALLINT) {
-        p = getVectorDoubleValue_SMALLINT;
-    }else if(srcType==TSDB_DATA_TYPE_USMALLINT) {
-        p = getVectorDoubleValue_USMALLINT;
-    }else if(srcType==TSDB_DATA_TYPE_INT) {
-        p = getVectorDoubleValue_INT;
-    }else if(srcType==TSDB_DATA_TYPE_UINT) {
-        p = getVectorDoubleValue_UINT;
-    }else if(srcType==TSDB_DATA_TYPE_BIGINT) {
-        p = getVectorDoubleValue_BIGINT;
-    }else if(srcType==TSDB_DATA_TYPE_UBIGINT) {
-        p = getVectorDoubleValue_UBIGINT;
-    }else if(srcType==TSDB_DATA_TYPE_FLOAT) {
-        p = getVectorDoubleValue_FLOAT;
-    }else if(srcType==TSDB_DATA_TYPE_DOUBLE) {
-        p = getVectorDoubleValue_DOUBLE;
-    }else {
-        assert(0);
-    }
-    return p;
-}
-
-
-typedef void* (*_arithmetic_getVectorValueAddr_fn_t)(void *src, int32_t index);
-
-void* getVectorValueAddr_TINYINT(void *src, int32_t index) {
-  return (void*)((int8_t *)src + index);
-}
-void* getVectorValueAddr_UTINYINT(void *src, int32_t index) {
-  return (void*)((uint8_t *)src + index);
-}
-void* getVectorValueAddr_SMALLINT(void *src, int32_t index) {
-  return (void*)((int16_t *)src + index);
-}
-void* getVectorValueAddr_USMALLINT(void *src, int32_t index) {
-  return (void*)((uint16_t *)src + index);
-}
-void* getVectorValueAddr_INT(void *src, int32_t index) {
-  return (void*)((int32_t *)src + index);
-}
-void* getVectorValueAddr_UINT(void *src, int32_t index) {
-  return (void*)((uint32_t *)src + index);
-}
-void* getVectorValueAddr_BIGINT(void *src, int32_t index) {
-  return (void*)((int64_t *)src + index);
-}
-void* getVectorValueAddr_UBIGINT(void *src, int32_t index) {
-  return (void*)((uint64_t *)src + index);
-}
-void* getVectorValueAddr_FLOAT(void *src, int32_t index) {
-  return (void*)((float *)src + index);
-}
-void* getVectorValueAddr_DOUBLE(void *src, int32_t index) {
-  return (void*)((double *)src + index);
-}
-
-_arithmetic_getVectorValueAddr_fn_t getVectorValueAddrFn(int32_t srcType) {
-    _arithmetic_getVectorValueAddr_fn_t p = NULL;
-    if(srcType==TSDB_DATA_TYPE_TINYINT) {
-        p = getVectorValueAddr_TINYINT;
-    }else if(srcType==TSDB_DATA_TYPE_UTINYINT) {
-        p = getVectorValueAddr_UTINYINT;
-    }else if(srcType==TSDB_DATA_TYPE_SMALLINT) {
-        p = getVectorValueAddr_SMALLINT;
-    }else if(srcType==TSDB_DATA_TYPE_USMALLINT) {
-        p = getVectorValueAddr_USMALLINT;
-    }else if(srcType==TSDB_DATA_TYPE_INT) {
-        p = getVectorValueAddr_INT;
-    }else if(srcType==TSDB_DATA_TYPE_UINT) {
-        p = getVectorValueAddr_UINT;
-    }else if(srcType==TSDB_DATA_TYPE_BIGINT) {
-        p = getVectorValueAddr_BIGINT;
-    }else if(srcType==TSDB_DATA_TYPE_UBIGINT) {
-        p = getVectorValueAddr_UBIGINT;
-    }else if(srcType==TSDB_DATA_TYPE_FLOAT) {
-        p = getVectorValueAddr_FLOAT;
-    }else if(srcType==TSDB_DATA_TYPE_DOUBLE) {
-        p = getVectorValueAddr_DOUBLE;
-    }else {
-        assert(0);
-    }
-    return p;
-}
-
-void vectorAdd(void *left, int32_t len1, int32_t _left_type, void *right, int32_t len2, int32_t _right_type, void *out, int32_t _ord) {
-  int32_t i = ((_ord) == TSDB_ORDER_ASC) ? 0 : MAX(len1, len2) - 1;                                 
-  int32_t step = ((_ord) == TSDB_ORDER_ASC) ? 1 : -1;                                               
-  double *output=(double*)out;
-  _arithmetic_getVectorValueAddr_fn_t getVectorValueAddrFnLeft = getVectorValueAddrFn(_left_type);
-  _arithmetic_getVectorValueAddr_fn_t getVectorValueAddrFnRight = getVectorValueAddrFn(_right_type);
-  _arithmetic_getVectorDoubleValue_fn_t getVectorDoubleValueFnLeft = getVectorDoubleValueFn(_left_type);
-  _arithmetic_getVectorDoubleValue_fn_t getVectorDoubleValueFnRight = getVectorDoubleValueFn(_right_type);
-                                                                                                 
-  if ((len1) == (len2)) {                                                                           
-    for (; i < (len2) && i >= 0; i += step, output += 1) {                                           
-      if (isNull(getVectorValueAddrFnLeft(left,i), _left_type) || isNull(getVectorValueAddrFnRight(right,i), _right_type)) { 
-        SET_DOUBLE_NULL(output);                                                                       
-        continue;                                                                                   
-      }                                                                                             
-      SET_DOUBLE_VAL(output,getVectorDoubleValueFnLeft(left,i) + getVectorDoubleValueFnRight(right,i));                                                       
-    }                                                                                               
-  } else if ((len1) == 1) {                                                                         
-    for (; i >= 0 && i < (len2); i += step, output += 1) {                                           
-      if (isNull(getVectorValueAddrFnLeft(left,0), _left_type) || isNull(getVectorValueAddrFnRight(right,i), _right_type)) {         
-        SET_DOUBLE_NULL(output);                                                                       
-        continue;                                                                                   
-      }                                                                                             
-      SET_DOUBLE_VAL(output,getVectorDoubleValueFnLeft(left,0) + getVectorDoubleValueFnRight(right,i));                                                       
-    }                                                                                               
-  } else if ((len2) == 1) {                                                                         
-    for (; i >= 0 && i < (len1); i += step, output += 1) {                                           
-      if (isNull(getVectorValueAddrFnLeft(left,i), _left_type) || isNull(getVectorValueAddrFnRight(right,0), _right_type)) {         
-        SET_DOUBLE_NULL(output);                                                                       
-        continue;                                                                                   
-      }                                                                                             
-      SET_DOUBLE_VAL(output,getVectorDoubleValueFnLeft(left,i) + getVectorDoubleValueFnRight(right,0));                                                        
-    }                                                                                               
-  }                                                                                                 
-}
-void vectorSub(void *left, int32_t len1, int32_t _left_type, void *right, int32_t len2, int32_t _right_type, void *out, int32_t _ord) {
-  int32_t i = ((_ord) == TSDB_ORDER_ASC) ? 0 : MAX(len1, len2) - 1;                                 
-  int32_t step = ((_ord) == TSDB_ORDER_ASC) ? 1 : -1;                                               
-  double *output=(double*)out;
-  _arithmetic_getVectorValueAddr_fn_t getVectorValueAddrFnLeft = getVectorValueAddrFn(_left_type);
-  _arithmetic_getVectorValueAddr_fn_t getVectorValueAddrFnRight = getVectorValueAddrFn(_right_type);
-  _arithmetic_getVectorDoubleValue_fn_t getVectorDoubleValueFnLeft = getVectorDoubleValueFn(_left_type);
-  _arithmetic_getVectorDoubleValue_fn_t getVectorDoubleValueFnRight = getVectorDoubleValueFn(_right_type);
-                                                                                                  
-  if ((len1) == (len2)) {                                                                           
-    for (; i < (len2) && i >= 0; i += step, output += 1) {                                           
-      if (isNull(getVectorValueAddrFnLeft(left,i), _left_type) || isNull(getVectorValueAddrFnRight(right,i), _right_type)) { 
-        SET_DOUBLE_NULL(output);                                                                       
-        continue;                                                                                   
-      }                                                                                             
-      SET_DOUBLE_VAL(output,getVectorDoubleValueFnLeft(left,i) - getVectorDoubleValueFnRight(right,i));                                                       
-    }                                                                                               
-  } else if ((len1) == 1) {                                                                         
-    for (; i >= 0 && i < (len2); i += step, output += 1) {                                           
-      if (isNull(getVectorValueAddrFnLeft(left,0), _left_type) || isNull(getVectorValueAddrFnRight(right,i), _right_type)) {         
-        SET_DOUBLE_NULL(output);                                                                       
-        continue;                                                                                   
-      }                                                                                             
-      SET_DOUBLE_VAL(output,getVectorDoubleValueFnLeft(left,0) - getVectorDoubleValueFnRight(right,i));                                                       
-    }                                                                                               
-  } else if ((len2) == 1) {                                                                         
-    for (; i >= 0 && i < (len1); i += step, output += 1) {                                           
-      if (isNull(getVectorValueAddrFnLeft(left,i), _left_type) || isNull(getVectorValueAddrFnRight(right,0), _right_type)) {         
-        SET_DOUBLE_NULL(output);                                                                       
-        continue;                                                                                   
-      }                                                                                             
-      SET_DOUBLE_VAL(output,getVectorDoubleValueFnLeft(left,i) - getVectorDoubleValueFnRight(right,0));                                                        
-    }                                                                                               
-  }                                                                                                 
-}
-void vectorMultiply(void *left, int32_t len1, int32_t _left_type, void *right, int32_t len2, int32_t _right_type, void *out, int32_t _ord) {
-  int32_t i = ((_ord) == TSDB_ORDER_ASC) ? 0 : MAX(len1, len2) - 1;                                 
-  int32_t step = ((_ord) == TSDB_ORDER_ASC) ? 1 : -1;                                               
-  double *output=(double*)out;
-  _arithmetic_getVectorValueAddr_fn_t getVectorValueAddrFnLeft = getVectorValueAddrFn(_left_type);
-  _arithmetic_getVectorValueAddr_fn_t getVectorValueAddrFnRight = getVectorValueAddrFn(_right_type);
-  _arithmetic_getVectorDoubleValue_fn_t getVectorDoubleValueFnLeft = getVectorDoubleValueFn(_left_type);
-  _arithmetic_getVectorDoubleValue_fn_t getVectorDoubleValueFnRight = getVectorDoubleValueFn(_right_type);
-                                                                                                  
-  if ((len1) == (len2)) {                                                                           
-    for (; i < (len2) && i >= 0; i += step, output += 1) {                                           
-      if (isNull(getVectorValueAddrFnLeft(left,i), _left_type) || isNull(getVectorValueAddrFnRight(right,i), _right_type)) { 
-        SET_DOUBLE_NULL(output);                                                                       
-        continue;                                                                                   
-      }                                                                                             
-      SET_DOUBLE_VAL(output,getVectorDoubleValueFnLeft(left,i) * getVectorDoubleValueFnRight(right,i));    
-    }                                                                                               
-  } else if ((len1) == 1) {                                                                         
-    for (; i >= 0 && i < (len2); i += step, output += 1) {                                           
-      if (isNull(getVectorValueAddrFnLeft(left,0), _left_type) || isNull(getVectorValueAddrFnRight(right,i), _right_type)) {         
-        SET_DOUBLE_NULL(output);                                                                       
-        continue;                                                                                   
-      }                                                                                             
-      SET_DOUBLE_VAL(output,getVectorDoubleValueFnLeft(left,0) * getVectorDoubleValueFnRight(right,i));                                                       
-    }                                                                                               
-  } else if ((len2) == 1) {                                                                         
-    for (; i >= 0 && i < (len1); i += step, output += 1) {                                           
-      if (isNull(getVectorValueAddrFnLeft(left,i), _left_type) || isNull(getVectorValueAddrFnRight(right,0), _right_type)) {         
-        SET_DOUBLE_NULL(output);                                                                       
-        continue;                                                                                   
-      }                                                                                             
-      SET_DOUBLE_VAL(output,getVectorDoubleValueFnLeft(left,i) * getVectorDoubleValueFnRight(right,0));                                                        
-    }                                                                                               
-  }                                                                                                 
-}
-void vectorDivide(void *left, int32_t len1, int32_t _left_type, void *right, int32_t len2, int32_t _right_type, void *out, int32_t _ord) {                
-  int32_t i = ((_ord) == TSDB_ORDER_ASC) ? 0 : MAX(len1, len2) - 1;                                 
-  int32_t step = ((_ord) == TSDB_ORDER_ASC) ? 1 : -1;                                               
-  double *output=(double*)out;
-  _arithmetic_getVectorValueAddr_fn_t getVectorValueAddrFnLeft = getVectorValueAddrFn(_left_type);
-  _arithmetic_getVectorValueAddr_fn_t getVectorValueAddrFnRight = getVectorValueAddrFn(_right_type);
-  _arithmetic_getVectorDoubleValue_fn_t getVectorDoubleValueFnLeft = getVectorDoubleValueFn(_left_type);
-  _arithmetic_getVectorDoubleValue_fn_t getVectorDoubleValueFnRight = getVectorDoubleValueFn(_right_type);
-                                                                                                    
-  if ((len1) == (len2)) {                                                                           
-    for (; i < (len2) && i >= 0; i += step, output += 1) {                                           
-      if (isNull(getVectorValueAddrFnLeft(left,i), _left_type) || isNull(getVectorValueAddrFnRight(right,i), _right_type)) { 
-        SET_DOUBLE_NULL(output);                                                                       
-        continue;                                                                                   
-      }                                                                                             
-      double v, u = 0.0;                                                                            
-      GET_TYPED_DATA(v, double, _right_type, getVectorValueAddrFnRight(right,i));                                
-      if (getComparFunc(TSDB_DATA_TYPE_DOUBLE, 0)(&v, &u) == 0) {                                   
-        SET_DOUBLE_NULL(output);                                                                       
-        continue;                                                                                   
-      }                                                                                             
-      SET_DOUBLE_VAL(output,getVectorDoubleValueFnLeft(left,i) /getVectorDoubleValueFnRight(right,i));                                                       
-    }                                                                                               
-  } else if ((len1) == 1) {                                                                         
-    for (; i >= 0 && i < (len2); i += step, output += 1) {                                           
-      if (isNull(getVectorValueAddrFnLeft(left,0), _left_type) || isNull(getVectorValueAddrFnRight(right,i), _right_type)) {         
-        SET_DOUBLE_NULL(output);                                                                       
-        continue;                                                                                   
-      }                                                                                             
-      double v, u = 0.0;                                                                            
-      GET_TYPED_DATA(v, double, _right_type, getVectorValueAddrFnRight(right,i));                                
-      if (getComparFunc(TSDB_DATA_TYPE_DOUBLE, 0)(&v, &u) == 0) {                                   
-        SET_DOUBLE_NULL(output);                                                                       
-        continue;                                                                                   
-      }                                                                                             
-      SET_DOUBLE_VAL(output,getVectorDoubleValueFnLeft(left,0) /getVectorDoubleValueFnRight(right,i));                                                       
-    }                                                                                               
-  } else if ((len2) == 1) {                                                                         
-    for (; i >= 0 && i < (len1); i += step, output += 1) {                                           
-      if (isNull(getVectorValueAddrFnLeft(left,i), _left_type) || isNull(getVectorValueAddrFnRight(right,0), _right_type)) {         
-        SET_DOUBLE_NULL(output);                                                                       
-        continue;                                                                                   
-      }                                                                                             
-      double v, u = 0.0;                                                                            
-      GET_TYPED_DATA(v, double, _right_type, getVectorValueAddrFnRight(right,0));                                
-      if (getComparFunc(TSDB_DATA_TYPE_DOUBLE, 0)(&v, &u) == 0) {                                   
-        SET_DOUBLE_NULL(output);                                                                       
-        continue;                                                                                   
-      }                                                                                             
-      SET_DOUBLE_VAL(output,getVectorDoubleValueFnLeft(left,i) /getVectorDoubleValueFnRight(right,0));                                                       
-    }                                                                                               
-  }                                                                                                 
-} 
-void vectorRemainder(void *left, int32_t len1, int32_t _left_type, void *right, int32_t len2, int32_t _right_type, void *out, int32_t _ord) {
-  int32_t i = (_ord == TSDB_ORDER_ASC) ? 0 : MAX(len1, len2) - 1;                                   
-  int32_t step = (_ord == TSDB_ORDER_ASC) ? 1 : -1;                                                 
-  double *output=(double*)out;
-  _arithmetic_getVectorValueAddr_fn_t getVectorValueAddrFnLeft = getVectorValueAddrFn(_left_type);
-  _arithmetic_getVectorValueAddr_fn_t getVectorValueAddrFnRight = getVectorValueAddrFn(_right_type);
-  _arithmetic_getVectorDoubleValue_fn_t getVectorDoubleValueFnLeft = getVectorDoubleValueFn(_left_type);
-  _arithmetic_getVectorDoubleValue_fn_t getVectorDoubleValueFnRight = getVectorDoubleValueFn(_right_type);
-                                                                                                    
-  if (len1 == (len2)) {                                                                             
-    for (; i >= 0 && i < (len2); i += step, output += 1) {                                           
-      if (isNull(getVectorValueAddrFnLeft(left,i), _left_type) || isNull(getVectorValueAddrFnRight(right,i), _right_type)) {     
-        SET_DOUBLE_NULL(output);                                                                       
-        continue;                                                                                   
-      }                                                                                             
-      double v, u = 0.0;                                                                            
-      GET_TYPED_DATA(v, double, _right_type, getVectorValueAddrFnRight(right,i));                                
-      if (getComparFunc(TSDB_DATA_TYPE_DOUBLE, 0)(&v, &u) == 0) {                                   
-        SET_DOUBLE_NULL(output);                                                                       
-        continue;                                                                                   
-      }                                                                                             
-      SET_DOUBLE_VAL(output,getVectorDoubleValueFnLeft(left,i) - ((int64_t)(getVectorDoubleValueFnLeft(left,i) / getVectorDoubleValueFnRight(right,i))) * getVectorDoubleValueFnRight(right,i));      
-    }                                                                                               
-  } else if (len1 == 1) {                                                                           
-    for (; i >= 0 && i < (len2); i += step, output += 1) {                                           
-      if (isNull(getVectorValueAddrFnLeft(left,0), _left_type) || isNull(getVectorValueAddrFnRight(right,i), _right_type)) {       
-        SET_DOUBLE_NULL(output);                                                                       
-        continue;                                                                                   
-      }                                                                                             
-      double v, u = 0.0;                                                                            
-      GET_TYPED_DATA(v, double, _right_type, getVectorValueAddrFnRight(right,i));                                
-      if (getComparFunc(TSDB_DATA_TYPE_DOUBLE, 0)(&v, &u) == 0) {                                   
-        SET_DOUBLE_NULL(output);                                                                       
-        continue;                                                                                   
-      }                                                                                             
-      SET_DOUBLE_VAL(output,getVectorDoubleValueFnLeft(left,0) - ((int64_t)(getVectorDoubleValueFnLeft(left,0) / getVectorDoubleValueFnRight(right,i))) * getVectorDoubleValueFnRight(right,i));      
-    }                                                                                               
-  } else if ((len2) == 1) {                                                                         
-    for (; i >= 0 && i < len1; i += step, output += 1) {                                             
-      if (isNull(getVectorValueAddrFnLeft(left,i), _left_type) || isNull(getVectorValueAddrFnRight(right,0), _right_type)) {       
-        SET_DOUBLE_NULL(output);                                                                       
-        continue;                                                                                   
-      }                                                                                             
-      double v, u = 0.0;                                                                            
-      GET_TYPED_DATA(v, double, _right_type, getVectorValueAddrFnRight(right,0));                                
-      if (getComparFunc(TSDB_DATA_TYPE_DOUBLE, 0)(&v, &u) == 0) {                                   
-        SET_DOUBLE_NULL(output);                                                                       
-        continue;                                                                                   
-      }                                                                                             
-      SET_DOUBLE_VAL(output,getVectorDoubleValueFnLeft(left,i) - ((int64_t)(getVectorDoubleValueFnLeft(left,i) / getVectorDoubleValueFnRight(right,0))) * getVectorDoubleValueFnRight(right,0));      
-    }                                                                                               
-  }                                                                                                 
-}
-
-_arithmetic_operator_fn_t getArithmeticOperatorFn(int32_t arithmeticOptr) {
-  switch (arithmeticOptr) {
-    case TSDB_BINARY_OP_ADD:
-      return vectorAdd;
-    case TSDB_BINARY_OP_SUBTRACT:
-      return vectorSub;
-    case TSDB_BINARY_OP_MULTIPLY:
-      return vectorMultiply;
-    case TSDB_BINARY_OP_DIVIDE:
-      return vectorDivide;
-    case TSDB_BINARY_OP_REMAINDER:
-      return vectorRemainder;
-    default:
-      assert(0);
-      return NULL;
-  }
-}