Merge pull request #21699 from taosdata/fix/TD-24717

enh(tdb/recycle): first round implemenation of page recycling

source/libs/tdb/src/db/tdbBtree.c

@@ -233,7 +233,11 @@ int tdbBtreeDelete(SBTree *pBt, const void *pKey, int kLen, TXN *pTxn) {
   int  ret;
 
   tdbBtcOpen(&btc, pBt, pTxn);
-
+  /*
+  btc.coder.ofps = taosArrayInit(8, sizeof(SPage *));
+  // btc.coder.ofps = taosArrayInit(8, sizeof(SPgno));
+  //pBtc->coder.ofps = taosArrayInit(8, sizeof(SPage *));
+  */
   tdbTrace("tdb delete, btc: %p, pTxn: %p", &btc, pTxn);
 
   // move the cursor
@@ -254,7 +258,18 @@ int tdbBtreeDelete(SBTree *pBt, const void *pKey, int kLen, TXN *pTxn) {
     tdbBtcClose(&btc);
     return -1;
   }
+  /*
+  SArray *ofps = btc.coder.ofps;
+  if (ofps) {
+    for (int i = 0; i < TARRAY_SIZE(ofps); ++i) {
+      SPage *ofp = *(SPage **)taosArrayGet(ofps, i);
+      tdbPagerInsertFreePage(btc.pBt->pPager, ofp, btc.pTxn);
+    }
 
+    taosArrayDestroy(ofps);
+    btc.coder.ofps = NULL;
+  }
+  */
   tdbBtcClose(&btc);
   return 0;
 }
@@ -563,6 +578,7 @@ static int tdbBtreeBalanceNonRoot(SBTree *pBt, SPage *pParent, int idx, TXN *pTx
       }
     }
     // copy the parent key out if child pages are not leaf page
+    // childNotLeaf = !(TDB_BTREE_PAGE_IS_LEAF(pOlds[0]) || TDB_BTREE_PAGE_IS_OVFL(pOlds[0]));
     childNotLeaf = !TDB_BTREE_PAGE_IS_LEAF(pOlds[0]);
     if (childNotLeaf) {
       for (int i = 0; i < nOlds; i++) {
@@ -592,7 +608,30 @@ static int tdbBtreeBalanceNonRoot(SBTree *pBt, SPage *pParent, int idx, TXN *pTx
     for (int i = 0; i < nOlds; i++) {
       nCells = TDB_PAGE_TOTAL_CELLS(pParent);
       if (sIdx < nCells) {
+        bool destroyOfps = false;
+        if (!childNotLeaf) {
+          if (!pParent->pPager->ofps) {
+            pParent->pPager->ofps = taosArrayInit(8, sizeof(SPage *));
+            destroyOfps = true;
+          }
+        }
+
         tdbPageDropCell(pParent, sIdx, pTxn, pBt);
+
+        if (!childNotLeaf) {
+          SArray *ofps = pParent->pPager->ofps;
+          if (ofps) {
+            for (int i = 0; i < TARRAY_SIZE(ofps); ++i) {
+              SPage *ofp = *(SPage **)taosArrayGet(ofps, i);
+              tdbPagerInsertFreePage(pParent->pPager, ofp, pTxn);
+            }
+
+            if (destroyOfps) {
+              taosArrayDestroy(ofps);
+              pParent->pPager->ofps = NULL;
+            }
+          }
+        }
       } else {
         ((SIntHdr *)pParent->pData)->pgno = 0;
       }
@@ -861,6 +900,8 @@ static int tdbBtreeBalanceNonRoot(SBTree *pBt, SPage *pParent, int idx, TXN *pTx
     if (!TDB_BTREE_PAGE_IS_LEAF(pNews[0])) {
       ((SIntHdr *)(pParent->pData))->pgno = ((SIntHdr *)(pNews[0]->pData))->pgno;
     }
+
+    tdbPagerInsertFreePage(pBt->pPager, pNews[0], pTxn);
   }
 
   for (int i = 0; i < 3; i++) {
@@ -870,6 +911,9 @@ static int tdbBtreeBalanceNonRoot(SBTree *pBt, SPage *pParent, int idx, TXN *pTx
   }
 
   for (pageIdx = 0; pageIdx < nOlds; ++pageIdx) {
+    if (pageIdx >= nNews) {
+      tdbPagerInsertFreePage(pBt->pPager, pOlds[pageIdx], pTxn);
+    }
     tdbPagerReturnPage(pBt->pPager, pOlds[pageIdx], pTxn);
   }
   for (; pageIdx < nNews; ++pageIdx) {
@@ -1311,7 +1355,11 @@ static int tdbBtreeDecodePayload(SPage *pPage, const SCell *pCell, int nHeader,
         if (ret < 0) {
           return -1;
         }
-
+        /*
+        if (pDecoder->ofps) {
+          taosArrayPush(pDecoder->ofps, &ofp);
+        }
+        */
         ofpCell = tdbPageGetCell(ofp, 0);
 
         if (nLeft <= ofp->maxLocal - sizeof(SPgno)) {
@@ -1346,11 +1394,17 @@ static int tdbBtreeDecodePayload(SPage *pPage, const SCell *pCell, int nHeader,
       int lastKeyPageSpace = 0;
       // load left key & val to ovpages
       while (pgno != 0) {
+        tdbTrace("tdb decode-ofp, pTxn: %p, pgno:%u by cell:%p", pTxn, pgno, pCell);
+        // printf("tdb decode-ofp, pTxn: %p, pgno:%u by cell:%p\n", pTxn, pgno, pCell);
         ret = tdbLoadOvflPage(&pgno, &ofp, pTxn, pBt);
         if (ret < 0) {
           return -1;
         }
-
+        /*
+        if (pDecoder->ofps) {
+          taosArrayPush(pDecoder->ofps, &ofp);
+        }
+        */
         ofpCell = tdbPageGetCell(ofp, 0);
 
         int lastKeyPage = 0;
@@ -1518,8 +1572,8 @@ static int tdbBtreeCellSize(const SPage *pPage, SCell *pCell, int dropOfp, TXN *
 
   if (pPage->vLen == TDB_VARIANT_LEN) {
     if (!leaf) {
-      tdbError("tdb/btree-cell-size: not a leaf page.");
-      return -1;
+      tdbError("tdb/btree-cell-size: not a leaf page:%p, pgno:%" PRIu32 ".", pPage, TDB_PAGE_PGNO(pPage));
+      // return -1;
     }
     nHeader += tdbGetVarInt(pCell + nHeader, &vLen);
   } else if (leaf) {
@@ -1559,8 +1613,27 @@ static int tdbBtreeCellSize(const SPage *pPage, SCell *pCell, int dropOfp, TXN *
           bytes = ofp->maxLocal - sizeof(SPgno);
         }
 
+        // SPgno origPgno = pgno;
         memcpy(&pgno, ofpCell + bytes, sizeof(pgno));
 
+        ret = tdbPagerWrite(pBt->pPager, ofp);
+        if (ret < 0) {
+          tdbError("failed to write page since %s", terrstr());
+          return -1;
+        }
+        /*
+        tdbPageDropCell(ofp, 0, pTxn, pBt);
+        */
+        // SIntHdr *pIntHdr = (SIntHdr *)(ofp->pData);
+        //  pIntHdr->flags = TDB_FLAG_ADD(0, TDB_BTREE_OVFL);
+        // pIntHdr->pgno = 0;
+        // ofp->pPager = NULL;
+
+        SArray *ofps = pPage->pPager->ofps;
+        if (ofps) {
+          taosArrayPush(ofps, &ofp);
+        }
+
         tdbPagerReturnPage(pPage->pPager, ofp, pTxn);
 
         nLeft -= bytes;
@@ -1980,6 +2053,11 @@ static int tdbBtcMoveDownward(SBTC *pBtc) {
     return -1;
   }
 
+  if (TDB_BTREE_PAGE_IS_OVFL(pBtc->pPage)) {
+    tdbError("tdb/btc-move-downward: should not be a ovfl page here.");
+    return -1;
+  }
+
   if (pBtc->idx < TDB_PAGE_TOTAL_CELLS(pBtc->pPage)) {
     pCell = tdbPageGetCell(pBtc->pPage, pBtc->idx);
     pgno = ((SPgno *)pCell)[0];
@@ -2068,8 +2146,27 @@ int tdbBtcDelete(SBTC *pBtc) {
     return -1;
   }
 
+  bool destroyOfps = false;
+  if (!pBtc->pPage->pPager->ofps) {
+    pBtc->pPage->pPager->ofps = taosArrayInit(8, sizeof(SPage *));
+    destroyOfps = true;
+  }
+
   tdbPageDropCell(pBtc->pPage, idx, pBtc->pTxn, pBtc->pBt);
 
+  SArray *ofps = pBtc->pPage->pPager->ofps;
+  if (ofps) {
+    for (int i = 0; i < TARRAY_SIZE(ofps); ++i) {
+      SPage *ofp = *(SPage **)taosArrayGet(ofps, i);
+      tdbPagerInsertFreePage(pBtc->pPage->pPager, ofp, pBtc->pTxn);
+    }
+
+    if (destroyOfps) {
+      taosArrayDestroy(ofps);
+      pBtc->pPage->pPager->ofps = NULL;
+    }
+  }
+
   // update interior page or do balance
   if (idx == nCells - 1) {
     if (idx) {
@@ -2113,6 +2210,8 @@ int tdbBtcDelete(SBTC *pBtc) {
         return -1;
       }
 
+      // printf("tdb/btc-delete: btree balance delete pgno: %d.\n", TDB_PAGE_PGNO(pBtc->pPage));
+
       ret = tdbBtreeBalance(pBtc);
       if (ret < 0) {
         tdbError("tdb/btc-delete: btree balance failed with ret: %d.", ret);
@@ -2181,7 +2280,13 @@ int tdbBtcUpsert(SBTC *pBtc, const void *pKey, int kLen, const void *pData, int
     tdbError("tdb/btc-upsert: page insert/update cell failed with ret: %d.", ret);
     return -1;
   }
-
+  /*
+  bool destroyOfps = false;
+  if (!pBtc->pPage->pPager->ofps) {
+    pBtc->pPage->pPager->ofps = taosArrayInit(8, sizeof(SPage *));
+    destroyOfps = true;
+  }
+  */
   // check balance
   if (pBtc->pPage->nOverflow > 0) {
     ret = tdbBtreeBalance(pBtc);
@@ -2190,7 +2295,20 @@ int tdbBtcUpsert(SBTC *pBtc, const void *pKey, int kLen, const void *pData, int
       return -1;
     }
   }
+  /*
+  SArray *ofps = pBtc->pPage->pPager->ofps;
+  if (ofps) {
+    for (int i = 0; i < TARRAY_SIZE(ofps); ++i) {
+      SPage *ofp = *(SPage **)taosArrayGet(ofps, i);
+      tdbPagerInsertFreePage(pBtc->pPage->pPager, ofp, pBtc->pTxn);
+    }
 
+    if (destroyOfps) {
+      taosArrayDestroy(ofps);
+      pBtc->pPage->pPager->ofps = NULL;
+    }
+  }
+  */
   return 0;
 }
 

source/libs/tdb/src/db/tdbDb.c

@@ -70,6 +70,11 @@ int32_t tdbOpen(const char *dbname, int32_t szPage, int32_t pages, TDB **ppDb, i
   if (ret < 0) {
     return -1;
   }
+
+  ret = tdbTbOpen(TDB_FREEDB_NAME, sizeof(SPgno), 0, NULL, pDb, &pDb->pFreeDb, rollback);
+  if (ret < 0) {
+    return -1;
+  }
 #endif
 
   *ppDb = pDb;
@@ -82,6 +87,7 @@ int tdbClose(TDB *pDb) {
   if (pDb) {
 #ifdef USE_MAINDB
     if (pDb->pMainDb) tdbTbClose(pDb->pMainDb);
+    if (pDb->pFreeDb) tdbTbClose(pDb->pFreeDb);
 #endif
 
     for (pPager = pDb->pgrList; pPager; pPager = pDb->pgrList) {

source/libs/tdb/src/db/tdbPager.c

@@ -292,7 +292,23 @@ int tdbPagerBegin(SPager *pPager, TXN *pTxn) {
   */
   return 0;
 }
+/*
+int tdbPagerCancelDirty(SPager *pPager, SPage *pPage, TXN *pTxn) {
+  SRBTreeNode *pNode = tRBTreeGet(&pPager->rbt, (SRBTreeNode *)pPage);
+  if (pNode) {
+    pPage->isDirty = 0;
 
+    tRBTreeDrop(&pPager->rbt, (SRBTreeNode *)pPage);
+    if (pTxn->jPageSet) {
+      hashset_remove(pTxn->jPageSet, (void *)((long)TDB_PAGE_PGNO(pPage)));
+    }
+
+    tdbPCacheRelease(pPager->pCache, pPage, pTxn);
+  }
+
+  return 0;
+}
+*/
 int tdbPagerCommit(SPager *pPager, TXN *pTxn) {
   SPage *pPage;
   int    ret;
@@ -338,10 +354,13 @@ int tdbPagerCommit(SPager *pPager, TXN *pTxn) {
     if (pTxn->jPageSet) {
       hashset_remove(pTxn->jPageSet, (void *)((long)TDB_PAGE_PGNO(pPage)));
     }
+
+    tdbTrace("tdb/pager-commit: remove page: %p %d from dirty tree: %p", pPage, TDB_PAGE_PGNO(pPage), &pPager->rbt);
+
     tdbPCacheRelease(pPager->pCache, pPage, pTxn);
   }
 
-  tdbTrace("pager/commit reset dirty tree: %p", &pPager->rbt);
+  tdbTrace("tdb/pager-commit reset dirty tree: %p", &pPager->rbt);
   tRBTreeCreate(&pPager->rbt, pageCmpFn);
 
   // sync the db file
@@ -629,6 +648,8 @@ int tdbPagerFlushPage(SPager *pPager, TXN *pTxn) {
   return 0;
 }
 
+static int tdbPagerAllocPage(SPager *pPager, SPgno *ppgno, TXN *pTxn);
+
 int tdbPagerFetchPage(SPager *pPager, SPgno *ppgno, SPage **ppPage, int (*initPage)(SPage *, void *, int), void *arg,
                       TXN *pTxn) {
   SPage *pPage;
@@ -643,7 +664,7 @@ int tdbPagerFetchPage(SPager *pPager, SPgno *ppgno, SPage **ppPage, int (*initPa
   // alloc new page
   if (pgno == 0) {
     loadPage = 0;
-    ret = tdbPagerAllocPage(pPager, &pgno);
+    ret = tdbPagerAllocPage(pPager, &pgno, pTxn);
     if (ret < 0) {
       tdbError("tdb/pager: %p, ret: %d pgno: %" PRIu32 ", alloc page failed.", pPager, ret, pgno);
       return -1;
@@ -695,23 +716,86 @@ void tdbPagerReturnPage(SPager *pPager, SPage *pPage, TXN *pTxn) {
   //        TDB_PAGE_PGNO(pPage), pPage);
 }
 
-static int tdbPagerAllocFreePage(SPager *pPager, SPgno *ppgno) {
-  // TODO: Allocate a page from the free list
+int tdbPagerInsertFreePage(SPager *pPager, SPage *pPage, TXN *pTxn) {
+  int   code = 0;
+  SPgno pgno = TDB_PAGE_PGNO(pPage);
+
+  // memset(pPage->pData, 0, pPage->pageSize);
+  tdbTrace("tdb/insert-free-page: tbc recycle page: %d.", pgno);
+  // printf("tdb/insert-free-page: tbc recycle page: %d.\n", pgno);
+  code = tdbTbInsert(pPager->pEnv->pFreeDb, &pgno, sizeof(pgno), NULL, 0, pTxn);
+  if (code < 0) {
+    tdbError("tdb/insert-free-page: tb insert failed with ret: %d.", code);
+    return -1;
+  }
+
+  pPage->pPager = NULL;
+
+  return code;
+}
+
+static int tdbPagerRemoveFreePage(SPager *pPager, SPgno *pPgno, TXN *pTxn) {
+  int  code = 0;
+  TBC *pCur;
+
+  if (!pPager->pEnv->pFreeDb) {
+    return 0;
+  }
+
+  code = tdbTbcOpen(pPager->pEnv->pFreeDb, &pCur, pTxn);
+  if (code < 0) {
     return 0;
+  }
+
+  code = tdbTbcMoveToFirst(pCur);
+  if (code) {
+    tdbError("tdb/remove-free-page: moveto first failed with ret: %d.", code);
+    tdbTbcClose(pCur);
+    return 0;
+  }
+
+  void *pKey = NULL;
+  int   nKey = 0;
+
+  code = tdbTbcGet(pCur, (const void **)&pKey, &nKey, NULL, NULL);
+  if (code < 0) {
+    // tdbError("tdb/remove-free-page: tbc get failed with ret: %d.", code);
+    tdbTbcClose(pCur);
+    return 0;
+  }
+
+  *pPgno = *(SPgno *)pKey;
+  tdbTrace("tdb/remove-free-page: tbc get page: %d.", *pPgno);
+  // printf("tdb/remove-free-page: tbc get page: %d.\n", *pPgno);
+
+  code = tdbTbcDelete(pCur);
+  if (code < 0) {
+    tdbError("tdb/remove-free-page: tbc delete failed with ret: %d.", code);
+    tdbTbcClose(pCur);
+    return 0;
+  }
+  tdbTbcClose(pCur);
+  return 0;
+}
+
+static int tdbPagerAllocFreePage(SPager *pPager, SPgno *ppgno, TXN *pTxn) {
+  // Allocate a page from the free list
+  return tdbPagerRemoveFreePage(pPager, ppgno, pTxn);
 }
 
 static int tdbPagerAllocNewPage(SPager *pPager, SPgno *ppgno) {
   *ppgno = ++pPager->dbFileSize;
+  // tdbError("tdb/alloc-new-page: %d.", *ppgno);
   return 0;
 }
 
-int tdbPagerAllocPage(SPager *pPager, SPgno *ppgno) {
+static int tdbPagerAllocPage(SPager *pPager, SPgno *ppgno, TXN *pTxn) {
   int ret;
 
   *ppgno = 0;
 
   // Try to allocate from the free list of the pager
-  ret = tdbPagerAllocFreePage(pPager, ppgno);
+  ret = tdbPagerAllocFreePage(pPager, ppgno, pTxn);
   if (ret < 0) {
     return -1;
   }

source/libs/tdb/src/inc/tdbInt.h

@@ -138,6 +138,7 @@ typedef struct {
   SPgno   pgno;
   u8     *pBuf;
   u8      freeKV;
+  SArray *ofps;
 } SCellDecoder;
 
 struct SBTC {
@@ -198,7 +199,8 @@ int  tdbPagerAbort(SPager *pPager, TXN *pTxn);
 int  tdbPagerFetchPage(SPager *pPager, SPgno *ppgno, SPage **ppPage, int (*initPage)(SPage *, void *, int), void *arg,
                        TXN *pTxn);
 void tdbPagerReturnPage(SPager *pPager, SPage *pPage, TXN *pTxn);
-int  tdbPagerAllocPage(SPager *pPager, SPgno *ppgno);
+int  tdbPagerInsertFreePage(SPager *pPager, SPage *pPage, TXN *pTxn);
+// int  tdbPagerAllocPage(SPager *pPager, SPgno *ppgno);
 int tdbPagerRestoreJournals(SPager *pPager);
 int tdbPagerRollback(SPager *pPager);
 
@@ -373,6 +375,7 @@ static inline SCell *tdbPageGetCell(SPage *pPage, int idx) {
 
 #ifdef USE_MAINDB
 #define TDB_MAINDB_NAME "main.tdb"
+#define TDB_FREEDB_NAME "_free.db"
 #endif
 
 struct STDB {
@@ -386,6 +389,7 @@ struct STDB {
   SPager **pgrHash;
 #ifdef USE_MAINDB
   TTB *pMainDb;
+  TTB *pFreeDb;
 #endif
   int64_t txnId;
 };
@@ -403,6 +407,7 @@ struct SPager {
   SRBTree rbt;
   // u8        inTran;
   TXN    *pActiveTxn;
+  SArray *ofps;
   SPager *pNext;      // used by TDB
   SPager *pHashNext;  // used by TDB
 #ifdef USE_MAINDB

source/libs/tdb/test/CMakeLists.txt

@@ -14,3 +14,7 @@ target_link_libraries(tdbExOVFLTest tdb gtest gtest_main)
 add_executable(tdbPageDefragmentTest "tdbPageDefragmentTest.cpp")
 target_link_libraries(tdbPageDefragmentTest tdb gtest gtest_main)
 
+# page recycling testing
+add_executable(tdbPageRecycleTest "tdbPageRecycleTest.cpp")
+target_link_libraries(tdbPageRecycleTest tdb gtest gtest_main)
+

source/libs/tdb/test/tdbExOVFLTest.cpp

@@ -190,6 +190,15 @@ static void insertOfp(void) {
   // commit current transaction
   tdbCommit(pEnv, txn);
   tdbPostCommit(pEnv, txn);
+
+  closePool(pPool);
+
+  // Close a database
+  tdbTbClose(pDb);
+
+  // Close Env
+  ret = tdbClose(pEnv);
+  GTEST_ASSERT_EQ(ret, 0);
 }
 
 // TEST(TdbOVFLPagesTest, DISABLED_TbInsertTest) {
@@ -233,6 +242,13 @@ TEST(TdbOVFLPagesTest, TbGetTest) {
 
     tdbFree(pVal);
   }
+
+  // Close a database
+  tdbTbClose(pDb);
+
+  // Close Env
+  ret = tdbClose(pEnv);
+  GTEST_ASSERT_EQ(ret, 0);
 }
 
 // TEST(TdbOVFLPagesTest, DISABLED_TbDeleteTest) {
@@ -334,6 +350,15 @@ tdbBegin(pEnv, &txn);
   // commit current transaction
   tdbCommit(pEnv, txn);
   tdbPostCommit(pEnv, txn);
+
+  closePool(pPool);
+
+  // Close a database
+  tdbTbClose(pDb);
+
+  // Close Env
+  ret = tdbClose(pEnv);
+  GTEST_ASSERT_EQ(ret, 0);
 }
 
 // TEST(tdb_test, DISABLED_simple_insert1) {
@@ -407,6 +432,8 @@ TEST(tdb_test, simple_insert1) {
     tdbCommit(pEnv, txn);
     tdbPostCommit(pEnv, txn);
 
+    closePool(pPool);
+
     {  // Query the data
       void *pVal = NULL;
       int   vLen;

tests/script/tsim/sma/drop_sma.sim

@@ -129,6 +129,7 @@ sql DROP INDEX sma_index_3 ;
 
 print ========== step8
 sql drop database if exists db;
+sleep 2000
 sql create database db duration 300;
 sql use db;
 sql create table stb1(ts timestamp, c_int int, c_bint bigint, c_sint smallint, c_tint tinyint,c_float float, c_double double, c_bool bool,c_binary binary(16), c_nchar nchar(32), c_ts timestamp,c_tint_un tinyint unsigned, c_sint_un smallint unsigned,c_int_un int unsigned, c_bint_un bigint unsigned) tags (t_int int);