SQLite

** Boston, MA  02111-1307, USA.
**
** Author contact information:
**   drh@hwaci.com
**   http://www.hwaci.com/drh/
**
*************************************************************************
** $Id: btree.c,v 1.20 2001/07/02 17:51:46 drh Exp $
**
** This file implements a external (disk-based) database using BTrees.
** For a detailed discussion of BTrees, refer to
**
**     Donald E. Knuth, THE ART OF COMPUTER PROGRAMMING, Volume 3:
**     "Sorting And Searching", pages 473-480. Addison-Wesley
**     Publishing Company, Reading, Massachusetts.

/*
** The maximum number of database entries that can be held in a single
** page of the database. 
*/
#define MX_CELL ((SQLITE_PAGE_SIZE-sizeof(PageHdr))/MIN_CELL_SIZE)

/*
** The amount of usable space on a single page of the BTree.  This is the
** page size minus the overhead of the page header.
*/
#define USABLE_SPACE  (SQLITE_PAGE_SIZE - sizeof(PageHdr))

/*
** The maximum amount of payload (in bytes) that can be stored locally for
** a database entry.  If the entry contains more data than this, the
** extra goes onto overflow pages.
**
** This number is chosen so that at least 4 cells will fit on every page.
*/
#define MX_LOCAL_PAYLOAD ((USABLE_SPACE/4-(sizeof(CellHdr)+sizeof(Pgno)))&~3)


/*
** Data on a database page is stored as a linked list of Cell structures.
** Both the key and the data are stored in aPayload[].  The key always comes
** first.  The aPayload[] field grows as necessary to hold the key and data,
** up to a maximum of MX_LOCAL_PAYLOAD bytes.  If the size of the key and
** data combined exceeds MX_LOCAL_PAYLOAD bytes, then Cell.ovfl is the

** into one big FreeBlk at the end of the page.
*/
static void defragmentPage(MemPage *pPage){
  int pc, i, n;
  FreeBlk *pFBlk;
  char newPage[SQLITE_PAGE_SIZE];

  assert( sqlitepager_iswriteable(pPage) );
  pc = sizeof(PageHdr);
  pPage->u.hdr.firstCell = pc;
  memcpy(newPage, pPage->u.aDisk, pc);
  for(i=0; i<pPage->nCell; i++){
    Cell *pCell = pPage->apCell[i];

    /* This routine should never be called on an overfull page.  The

*/
static int allocateSpace(MemPage *pPage, int nByte){
  FreeBlk *p;
  u16 *pIdx;
  int start;
  int cnt = 0;

  assert( sqlitepager_iswriteable(pPage) );
  assert( nByte==ROUNDUP(nByte) );
  if( pPage->nFree<nByte || pPage->isOverfull ) return 0;
  pIdx = &pPage->u.hdr.firstFree;
  p = (FreeBlk*)&pPage->u.aDisk[*pIdx];
  while( p->iSize<nByte ){
    assert( cnt++ < SQLITE_PAGE_SIZE/4 );
    if( p->iNext==0 ){

static void freeSpace(MemPage *pPage, int start, int size){
  int end = start + size;
  u16 *pIdx, idx;
  FreeBlk *pFBlk;
  FreeBlk *pNew;
  FreeBlk *pNext;

  assert( sqlitepager_iswriteable(pPage) );
  assert( size == ROUNDUP(size) );
  assert( start == ROUNDUP(start) );
  pIdx = &pPage->u.hdr.firstFree;
  idx = *pIdx;
  while( idx!=0 && idx<start ){
    pFBlk = (FreeBlk*)&pPage->u.aDisk[idx];
    if( idx + pFBlk->iSize == start ){

  if( pParent ){
    pPage->pParent = pParent;
    sqlitepager_ref(pParent);
  }
  if( pPage->isInit ) return SQLITE_OK;
  pPage->isInit = 1;
  pPage->nCell = 0;
  freeSpace = USABLE_SPACE;
  idx = pPage->u.hdr.firstCell;
  while( idx!=0 ){
    if( idx>SQLITE_PAGE_SIZE-MIN_CELL_SIZE ) goto page_format_error;
    if( idx<sizeof(PageHdr) ) goto page_format_error;
    if( idx!=ROUNDUP(idx) ) goto page_format_error;
    pCell = (Cell*)&pPage->u.aDisk[idx];
    sz = cellSize(pCell);

/*
** Set up a raw page so that it looks like a database page holding
** no entries.
*/
static void zeroPage(MemPage *pPage){
  PageHdr *pHdr;
  FreeBlk *pFBlk;
  assert( sqlitepager_iswriteable(pPage) );
  memset(pPage, 0, SQLITE_PAGE_SIZE);
  pHdr = &pPage->u.hdr;
  pHdr->firstCell = 0;
  pHdr->firstFree = sizeof(*pHdr);
  pFBlk = (FreeBlk*)&pHdr[1];
  pFBlk->iNext = 0;
  pFBlk->iSize = SQLITE_PAGE_SIZE - sizeof(*pHdr);

/*
** Open a new database.
**
** Actually, this routine just sets up the internal data structures
** for accessing the database.  We do not open the database file 
** until the first page is loaded.
*/
int sqliteBtreeOpen(
  const char *zFilename,    /* Name of the file containing the BTree database */
  int mode,                 /* Not currently used */
  int nCache,               /* How many pages in the page cache */
  Btree **ppBtree           /* Pointer to new Btree object written here */
){
  Btree *pBt;
  int rc;

  pBt = sqliteMalloc( sizeof(*pBt) );
  if( pBt==0 ){
    *ppBtree = 0;
    return SQLITE_NOMEM;
  }
  if( nCache<10 ) nCache = 10;
  rc = sqlitepager_open(&pBt->pPager, zFilename, nCache, EXTRA_SIZE);
  if( rc!=SQLITE_OK ){
    if( pBt->pPager ) sqlitepager_close(pBt->pPager);
    sqliteFree(pBt);
    *ppBtree = 0;
    return rc;
  }
  sqlitepager_set_destructor(pBt->pPager, pageDestructor);

** Move the cursor down to a new child page.
*/
static int moveToChild(BtCursor *pCur, int newPgno){
  int rc;
  MemPage *pNewPage;

  rc = sqlitepager_get(pCur->pBt->pPager, newPgno, (void**)&pNewPage);
  if( rc ) return rc;


  rc = initPage(pNewPage, newPgno, pCur->pPage);
  if( rc ) return rc;
  sqlitepager_unref(pCur->pPage);
  pCur->pPage = pNewPage;
  pCur->idx = 0;
  return SQLITE_OK;
}

/*

** Move the cursor to the root page
*/
static int moveToRoot(BtCursor *pCur){
  MemPage *pNew;
  int rc;

  rc = sqlitepager_get(pCur->pBt->pPager, pCur->pgnoRoot, (void**)&pNew);
  if( rc ) return rc;
  rc = initPage(pNew, pCur->pgnoRoot, 0);
  if( rc ) return rc;
  sqlitepager_unref(pCur->pPage);
  pCur->pPage = pNew;
  pCur->idx = 0;
  return SQLITE_OK;
}

*/
static void reparentPage(Pager *pPager, Pgno pgno, MemPage *pNewParent){
  MemPage *pThis;

  if( pgno==0 ) return;
  assert( pPager!=0 );
  pThis = sqlitepager_lookup(pPager, pgno);
  if( pThis && pThis->isInit ){
    if( pThis->pParent!=pNewParent ){
      if( pThis->pParent ) sqlitepager_unref(pThis->pParent);
      pThis->pParent = pNewParent;
      if( pNewParent ) sqlitepager_ref(pNewParent);
    }
    sqlitepager_unref(pThis);
  }

** routine will be called soon after this routine in order to rebuild 
** the linked list.
*/
static void dropCell(MemPage *pPage, int idx, int sz){
  int j;
  assert( idx>=0 && idx<pPage->nCell );
  assert( sz==cellSize(pPage->apCell[idx]) );
  assert( sqlitepager_iswriteable(pPage) );
  freeSpace(pPage, Addr(pPage->apCell[idx]) - Addr(pPage), sz);
  for(j=idx; j<pPage->nCell-1; j++){
    pPage->apCell[j] = pPage->apCell[j+1];
  }
  pPage->nCell--;
}

** routine will be called soon after this routine in order to rebuild 
** the linked list.
*/
static void insertCell(MemPage *pPage, int i, Cell *pCell, int sz){
  int idx, j;
  assert( i>=0 && i<=pPage->nCell );
  assert( sz==cellSize(pCell) );
  assert( sqlitepager_iswriteable(pPage) );
  idx = allocateSpace(pPage, sz);
  for(j=pPage->nCell; j>i; j--){
    pPage->apCell[j] = pPage->apCell[j-1];
  }
  pPage->nCell++;
  if( idx<=0 ){
    pPage->isOverfull = 1;

** occur in the order specified by the pPage->apCell[] array.  
** Invoke this routine once to repair damage after one or more
** invocations of either insertCell() or dropCell().
*/
static void relinkCellList(MemPage *pPage){
  int i;
  u16 *pIdx;
  assert( sqlitepager_iswriteable(pPage) );
  pIdx = &pPage->u.hdr.firstCell;
  for(i=0; i<pPage->nCell; i++){
    int idx = Addr(pPage->apCell[i]) - Addr(pPage);
    assert( idx>0 && idx<SQLITE_PAGE_SIZE );
    *pIdx = idx;
    pIdx = &pPage->apCell[i]->h.iNext;
  }

  int nNew;                    /* Number of pages in apNew[] */
  int nDiv;                    /* Number of cells in apDiv[] */
  int i, j, k;                 /* Loop counters */
  int idx;                     /* Index of pPage in pParent->apCell[] */
  int nxDiv;                   /* Next divider slot in pParent->apCell[] */
  int rc;                      /* The return code */
  int iCur;                    /* apCell[iCur] is the cell of the cursor */


  int totalSize;               /* Total bytes for all cells */
  int subtotal;                /* Subtotal of bytes in cells on one page */
  int cntNew[4];               /* Index in apCell[] of cell after i-th page */
  int szNew[4];                /* Combined size of cells place on i-th page */
  MemPage *extraUnref = 0;     /* A page that needs to be unref-ed */
  Pgno pgno;                   /* Page number */
  Cell *apCell[MX_CELL*3+5];   /* All cells from pages being balanceed */
  int szCell[MX_CELL*3+5];     /* Local size of all cells */
  Cell aTemp[2];               /* Temporary holding area for apDiv[] */
  MemPage aOld[3];             /* Temporary copies of pPage and its siblings */

  /* 
  ** Return without doing any work if pPage is neither overfull nor
  ** underfull.
  */
  assert( sqlitepager_iswriteable(pPage) );
  if( !pPage->isOverfull && pPage->nFree<SQLITE_PAGE_SIZE/3 ){
    relinkCellList(pPage);
    return SQLITE_OK;
  }

  /*
  ** Find the parent of the page to be balanceed.
  ** If there is no parent, it means this page is the root page and
  ** special rules apply.
  */
  pParent = pPage->pParent;
  if( pParent==0 ){
    Pgno pgnoChild;
    MemPage *pChild;
    if( pPage->nCell==0 ){
      if( pPage->u.hdr.rightChild ){
        /*
        ** The root page is empty.  Copy the one child page
        ** into the root page and return.  This reduces the depth
        ** of the BTree by one.
        */


        pgnoChild = pPage->u.hdr.rightChild;
        rc = sqlitepager_get(pBt->pPager, pgnoChild, (void**)&pChild);
        if( rc ) return rc;
        memcpy(pPage, pChild, SQLITE_PAGE_SIZE);
        pPage->isInit = 0;
        rc = initPage(pPage, sqlitepager_pagenumber(pPage), 0);
        assert( rc==SQLITE_OK );
        reparentChildPages(pBt->pPager, pPage);
        freePage(pBt, pChild, pgnoChild);
        sqlitepager_unref(pChild);
      }else{
        relinkCellList(pPage);
      }
      return SQLITE_OK;

    ** child.  Then fall thru to the code below which will cause
    ** the overfull child page to be split.
    */
    rc = sqlitepager_write(pPage);
    if( rc ) return rc;
    rc = allocatePage(pBt, &pChild, &pgnoChild);
    if( rc ) return rc;
    assert( sqlitepager_iswriteable(pChild) );
    copyPage(pChild, pPage);
    pChild->pParent = pPage;
    sqlitepager_ref(pPage);
    pChild->isOverfull = 1;
    if( pCur ){
      sqlitepager_unref(pCur->pPage);
      pCur->pPage = pChild;
    }else{
      extraUnref = pChild;
    }
    zeroPage(pPage);
    pPage->u.hdr.rightChild = pgnoChild;
    pParent = pPage;
    pPage = pChild;

  }
  rc = sqlitepager_write(pParent);
  if( rc ) return rc;

  
  /*
  ** Find the Cell in the parent page whose h.leftChild points back
  ** to pPage.  The "idx" variable is the index of that cell.  If pPage
  ** is the rightmost child of pParent then set idx to pParent->nCell 
  */
  idx = -1;

      pgnoOld[i] = apDiv[i]->h.leftChild;
    }else if( k==pParent->nCell ){
      pgnoOld[i] = pParent->u.hdr.rightChild;
    }else{
      break;
    }
    rc = sqlitepager_get(pBt->pPager, pgnoOld[i], (void**)&apOld[i]);
    if( rc ) goto balance_cleanup;
    rc = initPage(apOld[i], pgnoOld[i], pParent);
    if( rc ) goto balance_cleanup;
    nOld++;
  }

  /*
  ** Set iCur to be the index in apCell[] of the cell that the cursor
  ** is pointing to.  We will need this later on in order to keep the

      assert( apCell[nCell]->h.leftChild==pgnoOld[i] );
      apCell[nCell]->h.leftChild = pOld->u.hdr.rightChild;
      nCell++;
    }
  }

  /*
  ** Figure out the number of pages needed to hold all nCell cells.
  ** Store this number in "k".  Also compute szNew[] which is the total
  ** size of all cells on the i-th page and cntNew[] which is the index
  ** in apCell[] of the cell that divides path i from path i+1.  
  ** cntNew[k] should equal nCell.
  **
  ** This little patch of code is critical for keeping the tree
  ** balanced. 
  */
  totalSize = 0;
  for(i=0; i<nCell; i++){
    totalSize += szCell[i];
  }
  for(subtotal=k=i=0; i<nCell; i++){
    subtotal += szCell[i];
    if( subtotal > USABLE_SPACE ){
      szNew[k] = subtotal - szCell[i];
      cntNew[k] = i;
      subtotal = 0;
      k++;
    }
  }
  szNew[k] = subtotal;
  cntNew[k] = nCell;
  k++;
  for(i=k-1; i>0; i--){
    while( szNew[i]<USABLE_SPACE/2 ){
      cntNew[i-1]--;
      assert( cntNew[i-1]>0 );
      szNew[i] += szCell[cntNew[i-1]];
      szNew[i-1] -= szCell[cntNew[i-1]-1];
    }
  }
  assert( cntNew[0]>0 );

  /*
  ** Allocate k new pages
  */
  for(i=0; i<k; i++){
    rc = allocatePage(pBt, &apNew[i], &pgnoNew[i]);
    if( rc ) goto balance_cleanup;
    nNew++;
    zeroPage(apNew[i]);
    apNew[i]->isInit = 1;
  }

  /*
  ** Evenly distribute the data in apCell[] across the new pages.
  ** Insert divider cells into pParent as necessary.
  */
  j = 0;
  for(i=0; i<nNew; i++){
    MemPage *pNew = apNew[i];
    while( j<cntNew[i] ){
      assert( pNew->nFree>=szCell[j] );
      if( pCur && iCur==j ){ pCur->pPage = pNew; pCur->idx = pNew->nCell; }
      insertCell(pNew, pNew->nCell, apCell[j], szCell[j]);
      j++;
    }
    assert( pNew->nCell>0 );
    assert( !pNew->isOverfull );
    relinkCellList(pNew);
    if( i<nNew-1 && j<nCell ){
      pNew->u.hdr.rightChild = apCell[j]->h.leftChild;
      apCell[j]->h.leftChild = pgnoNew[i];
      if( pCur && iCur==j ){ pCur->pPage = pParent; pCur->idx = nxDiv; }
      insertCell(pParent, nxDiv, apCell[j], szCell[j]);
      j++;
      nxDiv++;
    }
  }
  assert( j==nCell );
  apNew[nNew-1]->u.hdr.rightChild = apOld[nOld-1]->u.hdr.rightChild;
  if( nxDiv==pParent->nCell ){
    pParent->u.hdr.rightChild = pgnoNew[nNew-1];
  }else{
    pParent->apCell[nxDiv]->h.leftChild = pgnoNew[nNew-1];
  }
  if( pCur ){

    Cell *pNext;
    int szNext;
    getTempCursor(pCur, &leafCur);
    rc = sqliteBtreeNext(&leafCur, 0);
    if( rc!=SQLITE_OK ){
      return SQLITE_CORRUPT;
    }
    rc = sqlitepager_write(leafCur.pPage);
    if( rc ) return rc;
    dropCell(pPage, pCur->idx, cellSize(pCell));
    pNext = leafCur.pPage->apCell[leafCur.idx];
    szNext = cellSize(pNext);
    pNext->h.leftChild = pgnoChild;
    insertCell(pPage, pCur->idx, pNext, szNext);
    rc = balance(pCur->pBt, pPage, pCur);
    if( rc ) return rc;
    pCur->bSkipNext = 1;
    dropCell(leafCur.pPage, leafCur.idx, szNext);
    rc = balance(pCur->pBt, leafCur.pPage, 0);
    releaseTempCursor(&leafCur);
  }else{
    dropCell(pPage, pCur->idx, cellSize(pCell));
    if( pCur->idx>=pPage->nCell && pCur->idx>0 ){
      pCur->idx--;
    }else{
      pCur->bSkipNext = 1;
    }
    rc = balance(pCur->pBt, pPage, pCur);
  }
  return rc;
}

/*
** Create a new BTree in the same file.  Write into *piTable the index
** of the root page of the new table.
*/
int sqliteBtreeCreateTable(Btree *pBt, int *piTable){
  MemPage *pRoot;
  Pgno pgnoRoot;
  int rc;
  if( !pBt->inTrans ){
    return SQLITE_ERROR;  /* Must start a transaction first */
  }
  rc = allocatePage(pBt, &pRoot, &pgnoRoot);
  if( rc ) return rc;
  assert( sqlitepager_iswriteable(pRoot) );
  zeroPage(pRoot);
  sqlitepager_unref(pRoot);
  *piTable = (int)pgnoRoot;
  return SQLITE_OK;
}

/*
** Erase the given database page and all its children.  Return
** the page to the freelist.
*/
static int clearDatabasePage(Btree *pBt, Pgno pgno, int freePageFlag){
  MemPage *pPage;
  int rc;
  Cell *pCell;
  int idx;

  rc = sqlitepager_get(pBt->pPager, pgno, (void**)&pPage);
  if( rc ) return rc;
  rc = sqlitepager_write(pPage);
  if( rc ) return rc;
  idx = pPage->u.hdr.firstCell;
  while( idx>0 ){
    pCell = (Cell*)&pPage->u.aDisk[idx];
    idx = pCell->h.iNext;
    if( pCell->h.leftChild ){
      rc = clearDatabasePage(pBt, pCell->h.leftChild, 1);

******************************************************************************/
#ifdef SQLITE_TEST

/*
** Print a disassembly of the given page on standard output.  This routine
** is used for debugging and testing only.
*/
int sqliteBtreePageDump(Btree *pBt, int pgno, int recursive){
  int rc;
  MemPage *pPage;
  int i, j;
  int nFree;
  u16 idx;
  char range[20];
  unsigned char payload[20];
  rc = sqlitepager_get(pBt->pPager, (Pgno)pgno, (void**)&pPage);
  if( rc ){
    return rc;
  }
  if( recursive ) printf("PAGE %d:\n", pgno);
  i = 0;
  idx = pPage->u.hdr.firstCell;
  while( idx>0 && idx<=SQLITE_PAGE_SIZE-MIN_CELL_SIZE ){
    Cell *pCell = (Cell*)&pPage->u.aDisk[idx];
    int sz = cellSize(pCell);
    sprintf(range,"%d..%d", idx, idx+sz-1);
    sz = pCell->h.nKey + pCell->h.nData;
    if( sz>sizeof(payload)-1 ) sz = sizeof(payload)-1;
    memcpy(payload, pCell->aPayload, sz);
    for(j=0; j<sz; j++){
      if( payload[j]<0x20 || payload[j]>0x7f ) payload[j] = '.';
    }
    payload[sz] = 0;
    printf(
      "cell %2d: i=%-10s chld=%-4d nk=%-4d nd=%-4d payload=%s\n",
      i, range, (int)pCell->h.leftChild, pCell->h.nKey, pCell->h.nData,
      payload
    );
    if( pPage->isInit && pPage->apCell[i]!=pCell ){
      printf("**** apCell[%d] does not match on prior entry ****\n", i);
    }
    i++;
    idx = pCell->h.iNext;
  }
  if( idx!=0 ){
    printf("ERROR: next cell index out of range: %d\n", idx);

       i, range, p->iSize, nFree);
    idx = p->iNext;
    i++;
  }
  if( idx!=0 ){
    printf("ERROR: next freeblock index out of range: %d\n", idx);
  }
  if( recursive && pPage->u.hdr.rightChild!=0 ){
    idx = pPage->u.hdr.firstCell;
    while( idx>0 && idx<SQLITE_PAGE_SIZE-MIN_CELL_SIZE ){
      Cell *pCell = (Cell*)&pPage->u.aDisk[idx];
      sqliteBtreePageDump(pBt, pCell->h.leftChild, 1);
      idx = pCell->h.iNext;
    }
    sqliteBtreePageDump(pBt, pPage->u.hdr.rightChild, 1);
  }
  sqlitepager_unref(pPage);
  return SQLITE_OK;
}

/*
** Fill aResult[] with information about the entry and page that the
** cursor is pointing to.

*/
typedef struct SanityCheck SanityCheck;
struct SanityCheck {
  Btree *pBt;    // The tree being checked out
  Pager *pPager; // The associated pager.  Also accessible by pBt->pPager
  int nPage;     // Number of pages in the database
  int *anRef;    // Number of times each page is referenced
  int nTreePage; // Number of BTree pages
  int nByte;     // Number of bytes of data stored on BTree pages
  char *zErrMsg; // An error message.  NULL of no errors seen.
};

/*
** Append a message to the error message string.
*/
static void checkAppendMsg(SanityCheck *pCheck, char *zMsg1, char *zMsg2){

**          but combine to completely cover the page.
**      2.  Make sure cell keys are in order.
**      3.  Make sure no key is less than or equal to zLowerBound.
**      4.  Make sure no key is greater than or equal to zUpperBound.
**      5.  Check the integrity of overflow pages.
**      6.  Recursively call checkTreePage on all children.
**      7.  Verify that the depth of all children is the same.
**      8.  Make sure this page is at least 33% full or else it is
**          the root of the tree.
*/
static int checkTreePage(
  SanityCheck *pCheck,  /* Context for the sanity check */
  int iPage,            /* Page number of the page to check */
  MemPage *pParent,     /* Parent page */
  char *zParentContext, /* Parent context */

      checkAppendMsg(pCheck, zMsg, 0);
      break;
    }
  }

  /* Check that free space is kept to a minimum
  */
#if 0
  if( pParent && pParent->nCell>2 && pPage->nFree>3*SQLITE_PAGE_SIZE/4 ){
    sprintf(zMsg, "free space (%d) greater than max (%d)", pPage->nFree,
       SQLITE_PAGE_SIZE/3);
    checkAppendMsg(pCheck, zContext, zMsg);
  }
#endif

  /* Update freespace totals.
  */
  pCheck->nTreePage++;
  pCheck->nByte += USABLE_SPACE - pPage->nFree;

  sqlitepager_unref(pPage);
  return depth;
}

/*
** This routine does a complete check of the given BTree file.  aRoot[] is

**   drh@hwaci.com
**   http://www.hwaci.com/drh/
**
*************************************************************************
** This header file defines the interface that the sqlite B-Tree file
** subsystem.
**
** @(#) $Id: btree.h,v 1.9 2001/07/02 17:51:46 drh Exp $
*/

typedef struct Btree Btree;
typedef struct BtCursor BtCursor;

int sqliteBtreeOpen(const char *zFilename, int mode, int nPg, Btree **ppBtree);
int sqliteBtreeClose(Btree*);

int sqliteBtreeBeginTrans(Btree*);
int sqliteBtreeCommit(Btree*);
int sqliteBtreeRollback(Btree*);

int sqliteBtreeCreateTable(Btree*, int*);

#define SQLITE_N_BTREE_META 4
int sqliteBtreeGetMeta(Btree*, int*);
int sqliteBtreeUpdateMeta(Btree*, int*);


#ifdef SQLITE_TEST
int sqliteBtreePageDump(Btree*, int, int);
int sqliteBtreeCursorDump(BtCursor*, int*);
Pager *sqliteBtreePager(Btree*);
char *sqliteBtreeSanityCheck(Btree*, int*, int);
#endif

*************************************************************************
** This is the implementation of the page cache subsystem.
** 
** The page cache is used to access a database file.  The pager journals
** all writes in order to support rollback.  Locking is used to limit
** access to one or more reader or one writer.
**
** @(#) $Id: pager.c,v 1.13 2001/07/02 17:51:46 drh Exp $
*/
#include "sqliteInt.h"
#include "pager.h"
#include <fcntl.h>
#include <sys/stat.h>
#include <unistd.h>
#include <assert.h>

  void (*xDestructor)(void*); /* Call this routine when freeing pages */
  int nPage;                  /* Total number of in-memory pages */
  int nRef;                   /* Number of in-memory pages with PgHdr.nRef>0 */
  int mxPage;                 /* Maximum number of pages to hold in cache */
  int nHit, nMiss, nOvfl;     /* Cache hits, missing, and LRU overflows */
  unsigned char state;        /* SQLITE_UNLOCK, _READLOCK or _WRITELOCK */
  unsigned char errMask;      /* One of several kinds of errors */
  unsigned char *aInJournal;  /* One bit for each page in the database file */
  PgHdr *pFirst, *pLast;      /* List of free pages */
  PgHdr *pAll;                /* List of all pages */
  PgHdr *aHash[N_PG_HASH];    /* Hash table to map page number of PgHdr */
};

/*
** These are bits that can be set in Pager.errMask.

}

/*
** Move the cursor for file descriptor fd to the point whereto from
** the beginning of the file.
*/
static int pager_seek(int fd, off_t whereto){
  /*printf("SEEK to page %d\n", whereto/SQLITE_PAGE_SIZE + 1);*/
  lseek(fd, whereto, SEEK_SET);
  return SQLITE_OK;
}

/*
** Truncate the given file so that it contains exactly mxPg pages
** of data.

** Read nBytes of data from fd into pBuf.  If the data cannot be
** read or only a partial read occurs, then the unread parts of
** pBuf are filled with zeros and this routine returns SQLITE_IOERR.
** If the read is completely successful, return SQLITE_OK.
*/
static int pager_read(int fd, void *pBuf, int nByte){
  int rc;
  /* printf("READ\n");*/
  rc = read(fd, pBuf, nByte);
  if( rc<0 ){
    memset(pBuf, 0, nByte);
    return SQLITE_IOERR;
  }
  if( rc<nByte ){
    memset(&((char*)pBuf)[rc], 0, nByte - rc);

/*
** Write nBytes of data into fd.  If any problem occurs or if the
** write is incomplete, SQLITE_IOERR is returned.  SQLITE_OK is
** returned upon complete success.
*/
static int pager_write(int fd, const void *pBuf, int nByte){
  int rc;
  /*printf("WRITE\n");*/
  rc = write(fd, pBuf, nByte);
  if( rc<nByte ){
    return SQLITE_FULL;
  }else{
    return SQLITE_OK;
  }
}

  PgHdr *pPg;
  if( pPager->state!=SQLITE_WRITELOCK ) return SQLITE_OK;
  pager_unlock(pPager->fd);
  rc = pager_lock(pPager->fd, 0);
  unlink(pPager->zJournal);
  close(pPager->jfd);
  pPager->jfd = -1;
  sqliteFree( pPager->aInJournal );
  pPager->aInJournal = 0;
  for(pPg=pPager->pAll; pPg; pPg=pPg->pNextAll){
    pPg->inJournal = 0;
    pPg->dirty = 0;
  }
  if( rc!=SQLITE_OK ){
    pPager->state = SQLITE_UNLOCK;
    rc = SQLITE_PROTOCOL;

    /* Playback the page.  Update the in-memory copy of the page
    ** at the same time, if there is one.
    */
    pPg = pager_lookup(pPager, pgRec.pgno);
    if( pPg ){
      memcpy(PGHDR_TO_DATA(pPg), pgRec.aData, SQLITE_PAGE_SIZE);
      memset(PGHDR_TO_EXTRA(pPg), 0, pPager->nExtra);
    }
    rc = pager_seek(pPager->fd, (pgRec.pgno-1)*SQLITE_PAGE_SIZE);
    if( rc!=SQLITE_OK ) break;
    rc = pager_write(pPager->fd, pgRec.aData, SQLITE_PAGE_SIZE);
    if( rc!=SQLITE_OK ) break;
  }
  if( rc!=SQLITE_OK ){

      pPg->pPrevAll = 0;
      pPager->pAll = pPg;
      pPager->nPage++;
    }else{
      /* Recycle an older page.  First locate the page to be recycled.
      ** Try to find one that is not dirty and is near the head of
      ** of the free list */
      int cnt = pPager->mxPage/2;
      pPg = pPager->pFirst;
      while( pPg->dirty && 0<cnt-- && pPg->pNextFree ){
        pPg = pPg->pNextFree;
      }
      if( pPg==0 || pPg->dirty ) pPg = pPager->pFirst;
      assert( pPg->nRef==0 );

      /* If the page to be recycled is dirty, sync the journal and write 
      ** the old page into the database. */

          if( rc==SQLITE_OK ) rc = SQLITE_FULL;
          return rc;
        }
      }

      /* Unlink the old page from the free list and the hash table
      */
      if( pPg->pPrevFree ){
        pPg->pPrevFree->pNextFree = pPg->pNextFree;
      }else{
        assert( pPager->pFirst==pPg );
        pPager->pFirst = pPg->pNextFree;
      }
      if( pPg->pNextFree ){
        pPg->pNextFree->pPrevFree = pPg->pPrevFree;
      }else{
        assert( pPager->pLast==pPg );
        pPager->pLast = pPg->pPrevFree;
      }
      pPg->pNextFree = pPg->pPrevFree = 0;
      if( pPg->pNextHash ){
        pPg->pNextHash->pPrevHash = pPg->pPrevHash;
      }
      if( pPg->pPrevHash ){
        pPg->pPrevHash->pNextHash = pPg->pNextHash;
      }else{
        h = pager_hash(pPg->pgno);
        assert( pPager->aHash[h]==pPg );
        pPager->aHash[h] = pPg->pNextHash;
      }
      pPg->pNextHash = pPg->pPrevHash = 0;
      pPager->nOvfl++;
    }
    pPg->pgno = pgno;
    if( pPager->aInJournal && pgno<=pPager->origDbSize ){
      pPg->inJournal = (pPager->aInJournal[pgno/8] & (1<<(pgno&7)))!=0;
    }else{
      pPg->inJournal = 0;
    }
    pPg->dirty = 0;
    pPg->nRef = 1;
    REFINFO(pPg);
    pPager->nRef++;
    h = pager_hash(pgno);
    pPg->pNextHash = pPager->aHash[h];
    pPager->aHash[h] = pPg;

  if( pPager->errMask ){ 
    return pager_errcode(pPager);
  }
  pPg->dirty = 1;
  if( pPg->inJournal ){ return SQLITE_OK; }
  assert( pPager->state!=SQLITE_UNLOCK );
  if( pPager->state==SQLITE_READLOCK ){
    assert( pPager->aInJournal==0 );
    pPager->aInJournal = sqliteMalloc( pPager->dbSize/8 + 1 );
    if( pPager->aInJournal==0 ){
      return SQLITE_NOMEM;
    }
    pPager->jfd = open(pPager->zJournal, O_RDWR|O_CREAT, 0644);
    if( pPager->jfd<0 ){
      return SQLITE_CANTOPEN;
    }
    if( pager_lock(pPager->jfd, 1) ){
      close(pPager->jfd);
      pPager->jfd = -1;

      rc = pager_write(pPager->jfd, pData, SQLITE_PAGE_SIZE);
    }
    if( rc!=SQLITE_OK ){
      sqlitepager_rollback(pPager);
      pPager->errMask |= PAGER_ERR_FULL;
      return rc;
    }
    assert( pPager->aInJournal!=0 );
    pPager->aInJournal[pPg->pgno/8] |= 1<<(pPg->pgno&7);
  }
  pPg->inJournal = 1;
  if( pPager->dbSize<pPg->pgno ){
    pPager->dbSize = pPg->pgno;
  }
  return rc;
}

/*
** Return TRUE if the page given in the argument was previous passed
** to sqlitepager_write().  In other words, return TRUE if it is ok
** to change the content of the page.
*/
int sqlitepager_iswriteable(void *pData){
  PgHdr *pPg = DATA_TO_PGHDR(pData);
  return pPg->dirty;
}

/*
** Commit all changes to the database and release the write lock.
**
** If the commit fails for any reason, a rollback attempt is made
** and an error code is returned.  If the commit worked, SQLITE_OK
** is returned.

**   http://www.hwaci.com/drh/
**
*************************************************************************
** This header file defines the interface that the sqlite page cache
** subsystem.  The page cache subsystem reads and writes a file a page
** at a time and provides a journal for rollback.
**
** @(#) $Id: pager.h,v 1.7 2001/07/02 17:51:46 drh Exp $
*/

/*
** The size of one page
*/
#define SQLITE_PAGE_SIZE 1024

int sqlitepager_close(Pager *pPager);
int sqlitepager_get(Pager *pPager, Pgno pgno, void **ppPage);
void *sqlitepager_lookup(Pager *pPager, Pgno pgno);
int sqlitepager_ref(void*);
int sqlitepager_unref(void*);
Pgno sqlitepager_pagenumber(void*);
int sqlitepager_write(void*);
int sqlitepager_iswriteable(void*);
int sqlitepager_pagecount(Pager*);
int sqlitepager_commit(Pager*);
int sqlitepager_rollback(Pager*);
int *sqlitepager_stats(Pager*);

#ifdef SQLITE_TEST
void sqlitepager_refdump(Pager*);
int pager_refinfo_enable;
#endif

**   http://www.hwaci.com/drh/
**
*************************************************************************
** Code for testing the btree.c module in SQLite.  This code
** is not included in the SQLite library.  It is used for automated
** testing of the SQLite library.
**
** $Id: test3.c,v 1.7 2001/07/02 17:51:46 drh Exp $
*/
#include "sqliteInt.h"
#include "pager.h"
#include "btree.h"
#include "tcl.h"
#include <stdlib.h>
#include <string.h>

  int rc;
  char zBuf[100];
  if( argc!=2 ){
    Tcl_AppendResult(interp, "wrong # args: should be \"", argv[0],
       " FILENAME\"", 0);
    return TCL_ERROR;
  }
  rc = sqliteBtreeOpen(argv[1], 0666, 10, &pBt);
  if( rc!=SQLITE_OK ){
    Tcl_AppendResult(interp, errorName(rc), 0);
    return TCL_ERROR;
  }
  sprintf(zBuf,"0x%x",(int)pBt);
  Tcl_AppendResult(interp, zBuf, 0);
  return TCL_OK;

  if( argc!=3 ){
    Tcl_AppendResult(interp, "wrong # args: should be \"", argv[0],
       " ID\"", 0);
    return TCL_ERROR;
  }
  if( Tcl_GetInt(interp, argv[1], (int*)&pBt) ) return TCL_ERROR;
  if( Tcl_GetInt(interp, argv[2], &iPage) ) return TCL_ERROR;
  rc = sqliteBtreePageDump(pBt, iPage, 0);
  if( rc!=SQLITE_OK ){
    Tcl_AppendResult(interp, errorName(rc), 0);
    return TCL_ERROR;
  }
  return TCL_OK;
}

/*
** Usage:   btree_tree_dump ID PAGENUM
**
** Print a disassembly of a page and all its child pages on standard output
*/
static int btree_tree_dump(
  void *NotUsed,
  Tcl_Interp *interp,    /* The TCL interpreter that invoked this command */
  int argc,              /* Number of arguments */
  char **argv            /* Text of each argument */
){
  Btree *pBt;
  int iPage;
  int rc;

  if( argc!=3 ){
    Tcl_AppendResult(interp, "wrong # args: should be \"", argv[0],
       " ID\"", 0);
    return TCL_ERROR;
  }
  if( Tcl_GetInt(interp, argv[1], (int*)&pBt) ) return TCL_ERROR;
  if( Tcl_GetInt(interp, argv[2], &iPage) ) return TCL_ERROR;
  rc = sqliteBtreePageDump(pBt, iPage, 1);
  if( rc!=SQLITE_OK ){
    Tcl_AppendResult(interp, errorName(rc), 0);
    return TCL_ERROR;
  }
  return TCL_OK;
}

  Tcl_CreateCommand(interp, "btree_rollback", btree_rollback, 0, 0);
  Tcl_CreateCommand(interp, "btree_create_table", btree_create_table, 0, 0);
  Tcl_CreateCommand(interp, "btree_drop_table", btree_drop_table, 0, 0);
  Tcl_CreateCommand(interp, "btree_clear_table", btree_clear_table, 0, 0);
  Tcl_CreateCommand(interp, "btree_get_meta", btree_get_meta, 0, 0);
  Tcl_CreateCommand(interp, "btree_update_meta", btree_update_meta, 0, 0);
  Tcl_CreateCommand(interp, "btree_page_dump", btree_page_dump, 0, 0);
  Tcl_CreateCommand(interp, "btree_tree_dump", btree_tree_dump, 0, 0);
  Tcl_CreateCommand(interp, "btree_pager_stats", btree_pager_stats, 0, 0);
  Tcl_CreateCommand(interp, "btree_pager_ref_dump", btree_pager_ref_dump, 0, 0);
  Tcl_CreateCommand(interp, "btree_cursor", btree_cursor, 0, 0);
  Tcl_CreateCommand(interp, "btree_close_cursor", btree_close_cursor, 0, 0);
  Tcl_CreateCommand(interp, "btree_move_to", btree_move_to, 0, 0);
  Tcl_CreateCommand(interp, "btree_delete", btree_delete, 0, 0);
  Tcl_CreateCommand(interp, "btree_insert", btree_insert, 0, 0);

#   drh@hwaci.com
#   http://www.hwaci.com/drh/
#
#***********************************************************************
# This file implements regression tests for SQLite library.  The
# focus of this script is btree database backend
#
# $Id: btree.test,v 1.6 2001/07/02 17:51:47 drh Exp $


set testdir [file dirname $argv0]
source $testdir/tester.tcl

if {$dbprefix!="memory:" && [info commands btree_open]!=""} {

  btree_drop_table $::b1 2
  set ::c1 [btree_cursor $::b1 2]
  lindex [btree_get_meta $::b1] 0
} {4}
do_test btree-8.24 {
  lindex [btree_pager_stats $::b1] 1
} {2}
#btree_pager_ref_dump $::b1

# Check page splitting logic
#
do_test btree-9.1 {
  for {set i 1} {$i<=19} {incr i} {
    set key [format %03d $i]
    set data "*** $key *** $key *** $key *** $key ***"
    btree_insert $::c1 $key $data
  }
} {}
#btree_tree_dump $::b1 2
#btree_pager_ref_dump $::b1
#set pager_refinfo_enable 1
do_test btree-9.2 {
  btree_insert $::c1 020 {*** 020 *** 020 *** 020 *** 020 ***}
  select_keys $::c1
} {001 002 003 004 005 006 007 008 009 010 011 012 013 014 015 016 017 018 019 020}
#btree_page_dump $::b1 5

#btree_page_dump $::b1 2
#btree_page_dump $::b1 6
do_test btree-10.4 {
  btree_move_to $::c1 011
  btree_delete $::c1
  select_keys $::c1
} {001 002 003 004 005 006 007 008 009 010 012 013 014 015 016 017 018 019 020}
#btree_tree_dump $::b1 2
#btree_pager_ref_dump $::b1
for {set i 1} {$i<=20} {incr i} {
  do_test btree-10.5.$i {
    btree_move_to $::c1 [format %03d $i]
    lindex [btree_pager_stats $::b1] 1
  } {2}
  #btree_pager_ref_dump $::b1
  #btree_tree_dump $::b1 2
}

# Create a tree with lots more pages
#
catch {unset ::data}
catch {unset ::key}
for {set i 21} {$i<=1000} {incr i} {

} {258}
do_test btree-11.4.3 {
  btree_move_to $::c1 259
  btree_key $::c1
} {259}
do_test btree-11.4.4 {
  btree_move_to $::c1 257
  set n [btree_key $::c1]
  expr {$n==256||$n==258}
} {1}
do_test btree-11.5 {
  btree_move_to $::c1 513
  btree_delete $::c1
  btree_next $::c1
  btree_key $::c1
} {514}
do_test btree-11.5.1 {
  btree_move_to $::c1 512
  btree_key $::c1
} {512}
do_test btree-11.5.2 {
  btree_move_to $::c1 514
  btree_key $::c1
} {514}
do_test btree-11.5.3 {
  btree_move_to $::c1 515
  btree_key $::c1
} {515}
do_test btree-11.5.4 {
  btree_move_to $::c1 513
  set n [btree_key $::c1]
  expr {$n==512||$n==514}
} {1}
do_test btree-11.6 {
  btree_move_to $::c1 769
  btree_delete $::c1
  btree_next $::c1
  btree_key $::c1
} {770}
do_test btree-11.6.1 {
  btree_move_to $::c1 768
  btree_key $::c1
} {768}
do_test btree-11.6.2 {
  btree_move_to $::c1 771
  btree_key $::c1
} {771}
do_test btree-11.6.3 {
  btree_move_to $::c1 770
  btree_key $::c1
} {770}
do_test btree-11.6.4 {
  btree_move_to $::c1 769
  set n [btree_key $::c1]
  expr {$n==768||$n==770}
} {1}
#btree_page_dump $::b1 2
#btree_page_dump $::b1 25

# Change the data on an intermediate node such that the node becomes overfull
# and has to split.  We happen to know that intermediate nodes exist on
# 337, 401 and 465 by the btree_page_dumps above
#

#   drh@hwaci.com
#   http://www.hwaci.com/drh/
#
#***********************************************************************
# This file implements regression tests for SQLite library.  The
# focus of this script is btree database backend
#
# $Id: btree2.test,v 1.2 2001/07/02 17:51:47 drh Exp $


set testdir [file dirname $argv0]
source $testdir/tester.tcl

if {$dbprefix!="memory:" && [info commands btree_open]!=""} {

# Create a new database file containing no entries.  The database should
# contain 5 tables:
#
#     2   The descriptor table
#     3   The foreground table
#     4   The background table
#     5   The long key table
#     6   The long data table
#
# An explanation for what all these tables are used for is provided below.
#
do_test btree2-1.1 {
  expr srand(1)
  file delete -force test2.bt
  file delete -force test2.bt-journal
  set ::b [btree_open test2.bt]
  btree_begin_transaction $::b
  btree_create_table $::b
} {3}
do_test btree2-1.2 {

  return [string range $r 0 [expr {$len-1}]]
}

# Verify the invariants on the database.  Return an empty string on 
# success or an error message if something is amiss.
#
proc check_invariants {} {
  set ck [btree_sanity_check $::b 2 3 4 5 6]
  if {$ck!=""} {
    puts "\n*** SANITY:\n$ck"
    exit
    return $ck
  }
  btree_move_to $::c3 {}
  btree_move_to $::c4 {}
  btree_move_to $::c2 N
  set N [btree_data $::c2]
  btree_move_to $::c2 L
  set L [btree_data $::c2]
  set LM1 [expr {$L-1}]
  for {set i 1} {$i<=$N} {incr i} {
    set key [btree_key $::c3]
    if {[scan $key %d k]<1} {set k 0}
    if {$k!=$i} {
      set key [btree_key $::c4]
      if {[scan $key %d k]<1} {set k 0}
      if {$k!=$i} {
        # puts "MISSING $i"
        # puts {Page 3:}; btree_page_dump $::b 3
        # puts {Page 4:}; btree_page_dump $::b 4
        # exit
        return "Key $i is missing from both foreground and background"
      }
      set data [btree_data $::c4]
      btree_next $::c4
    } else {
      set data [btree_data $::c3]
      btree_next $::c3
    }

# Make random changes to the database such that each change preserves
# the invariants.  The number of changes is $n*N where N is the parameter
# from the descriptor table.  Each changes begins with a random key.
# the entry with that key is put in the foreground table with probability
# $I and it is put in background with probability (1.0-$I).  It gets
# a long key with probability $K and long data with probability $D.  
# 
set chngcnt 0
proc random_changes {n I K D} {
  btree_move_to $::c2 N
  set N [btree_data $::c2]
  btree_move_to $::c2 L
  set L [btree_data $::c2]
  set LM1 [expr {$L-1}]
  set total [expr {int($N*$n)}]
  set format %0${L}d
  for {set i 0} {$i<$total} {incr i} {
    set k [expr {int(rand()*$N)+1}]
    set insert [expr {rand()<=$I}]
    set longkey [expr {rand()<=$K}]
    set longdata [expr {rand()<=$D}]
    # incr ::chngcnt
    # if {$::chngcnt==251} {btree_tree_dump $::b 3} 
    # puts "CHANGE $::chngcnt: $k $insert $longkey $longdata"
    if {$longkey} {
      set x [expr {rand()}]
      set keylen [expr {int($x*$x*$x*$x*3000)+10}]
    } else {
      set keylen $L
    }
    set key [make_payload $k $L $keylen]
    if {$longdata} {
      set x [expr {rand()}]
      set datalen [expr {int($x*$x*$x*$x*3000)+10}]
    } else {
      set datalen $L
    }
    set data [make_payload $k $L $datalen]
    set basekey [format $format $k]

    if {[set c [btree_move_to $::c3 $basekey]]==0} {
      btree_delete $::c3
    } else {
      if {$c<0} {btree_next $::c3}
      if {[string match $basekey* [btree_key $::c3]]} {
        btree_delete $::c3
      }
    }
    if {[set c [btree_move_to $::c4 $basekey]]==0} {
      btree_delete $::c4
    } else {
      if {$c<0} {btree_next $::c4}
      if {[string match $basekey* [btree_key $::c4]]} {
        btree_delete $::c4
      }
    }
    if {[scan [btree_key $::c4] %d kx]<1} {set kx -1}
    if {$kx==$k} {
      btree_delete $::c4
    }
    if {$insert} {
      btree_insert $::c3 $key $data
    } else {
      btree_insert $::c4 $key $data
    }
    if {$longkey} {
      btree_insert $::c5 $basekey $keylen
    } elseif {[btree_move_to $::c5 $basekey]==0} {
      btree_delete $::c5
    }
    if {$longdata} {
      btree_insert $::c6 $basekey $datalen
    } elseif {[btree_move_to $::c6 $basekey]==0} {
      btree_delete $::c6
    }
    # set ck [btree_sanity_check $::b 2 3 4 5 6]
    # if {$ck!=""} {
    #   puts "\nSANITY CHECK FAILED!\n$ck"
    #   exit
    # }
    # puts "PAGE 3:"; btree_page_dump $::b 3
    # puts "PAGE 4:"; btree_page_dump $::b 4
  }

}

# Repeat this test sequence on database of various sizes
#
set testno 2
foreach {N L} {
  10 2
  50 2
  200 3
  2000 5
} {
  puts "**** N=$N L=$L ****"
  set hash [md5file test2.bt]
  do_test btree2-$testno.1 [subst -nocommands {
    set ::c2 [btree_cursor $::b 2]
    set ::c3 [btree_cursor $::b 3]
    set ::c4 [btree_cursor $::b 4]

    btree_close_cursor $::c5
    btree_close_cursor $::c6
    lindex [btree_pager_stats $::b] 1
  } {0}
  do_test btree2-$testno.7 {
    btree_close $::b
    set ::b [btree_open test2.bt]
    set ::c2 [btree_cursor $::b 2]
    set ::c3 [btree_cursor $::b 3]
    set ::c4 [btree_cursor $::b 4]
    set ::c5 [btree_cursor $::b 5]
    set ::c6 [btree_cursor $::b 6]
    check_invariants
  } {}

  # For each database size, run various changes tests.
  #
  set num2 1
  foreach {n I K D} {
    0.5 0.5 0.1 0.1
    1.0 0.2 0.1 0.1
    1.0 0.8 0.1 0.1
    2.0 0.0 0.1 0.1
    2.0 1.0 0.1 0.1
    2.0 0.0 0.0 0.0
    2.0 1.0 0.0 0.0
  } {
    set testid btree2-$testno.8.$num2
    set cnt 6
    for {set i 2} {$i<=6} {incr i} {
      if {[lindex [btree_cursor_dump [set ::c$i]] 0]!=$i} {incr cnt}
    }
    do_test $testid.1 {





      btree_begin_transaction $::b
      lindex [btree_pager_stats $::b] 1
    } $cnt
    set hash [md5file test2.bt]
    # exec cp test2.bt test2.bt.bu1
    do_test $testid.2 [subst {
      random_changes $n $I $K $D
    }] {}
    do_test $testid.3 {
      check_invariants
    } {}
    do_test $testid.4 {
      btree_close_cursor $::c2
      btree_close_cursor $::c3
      btree_close_cursor $::c4
      btree_close_cursor $::c5
      btree_close_cursor $::c6
      btree_rollback $::b
      md5file test2.bt
    } $hash
    # exec cp test2.bt test2.bt.bu2
    btree_begin_transaction $::b
    set ::c2 [btree_cursor $::b 2]
    set ::c3 [btree_cursor $::b 3]
    set ::c4 [btree_cursor $::b 4]
    set ::c5 [btree_cursor $::b 5]
    set ::c6 [btree_cursor $::b 6]
    do_test $testid.5 [subst {
      random_changes $n $I $K $D
    }] {}
    do_test $testid.6 {
      check_invariants
    } {}
    do_test $testid.7 {
      btree_commit $::b
      check_invariants
    } {}
    set hash [md5file test2.bt]
    do_test $testid.8 {
      btree_close_cursor $::c2
      btree_close_cursor $::c3
      btree_close_cursor $::c4
      btree_close_cursor $::c5
      btree_close_cursor $::c6
      lindex [btree_pager_stats $::b] 1
    } {0}
    do_test $testid.9 {
      btree_close $::b
      set ::b [btree_open test2.bt]
      set ::c2 [btree_cursor $::b 2]
      set ::c3 [btree_cursor $::b 3]
      set ::c4 [btree_cursor $::b 4]
      set ::c5 [btree_cursor $::b 5]
      set ::c6 [btree_cursor $::b 6]
      check_invariants
    } {}
    incr num2
  }
  btree_close_cursor $::c2
  btree_close_cursor $::c3
  btree_close_cursor $::c4
  btree_close_cursor $::c5
  btree_close_cursor $::c6
  incr testno
}  

# Testing is complete.  Shut everything down.
#
do_test btree-999.1 {
  lindex [btree_pager_stats $::b] 1