SQLite

  }
  put2byte(&data[hdr+3], nCell);
  put2byte(&data[hdr+5], cellbody);
  pPage->nFree -= (nCell*2 + nUsable - cellbody);
  pPage->nCell = (u16)nCell;
}


static void rebuildPage(
  MemPage *pPg,                   /* Edit this page */
  int nRemove,                    /* Cells to remove from start of page */
  int nCell,                      /* Final number of cells on page */
  u8 **apCell,                    /* Array of nCell final cells */
  u16 *szCell                     /* Array of nCell cell sizes */
){
  const int hdr = pPg->hdrOffset;          /* Offset of header on pPg */
  u8 * const aData = pPg->aData;           /* Pointer to data for pPg */
  const int usableSize = pPg->pBt->usableSize;
  u8 * const pEnd = &aData[usableSize];
  int i;
  u8 *pCellptr = pPg->aCellIdx;
  u8 *pTmp = sqlite3PagerTempSpace(pPg->pBt->pPager);
  u8 *pData;

  i = get2byte(&aData[hdr+5]);
  memcpy(&pTmp[i], &aData[i], usableSize - i);
  pData = &aData[usableSize];

  for(i=0; i<nCell; i++){
    u8 *pCell = apCell[i];
    if( pCell>aData && pCell<pEnd ){
      pCell = &pTmp[pCell - aData];
    }
    pData -= szCell[i];
    memcpy(pData, pCell, szCell[i]);
    put2byte(pCellptr, (pData - aData));
    pCellptr += 2;
    assert( szCell[i]==cellSizePtr(pPg, pCell) );
  }

  pPg->nFree = (pData - pCellptr);
  pPg->nCell = nCell;
  pPg->nOverflow = 0;

  put2byte(&aData[hdr+1], 0);
  put2byte(&aData[hdr+3], pPg->nCell);
  put2byte(&aData[hdr+5], pData - aData);
  aData[hdr+7] = 0x00;
}

/*
** The following parameters determine how many adjacent pages get involved
** in a balancing operation.  NN is the number of neighbors on either side
** of the page that participate in the balancing operation.  NB is the
** total number of pages that participate, including the target page and
** NN neighbors on either side.
**

    releasePage(pNew);
  }

  return rc;
}
#endif /* SQLITE_OMIT_QUICKBALANCE */

#if 1
/*
** This function does not contribute anything to the operation of SQLite.
** it is sometimes activated temporarily while debugging code responsible 
** for setting pointer-map entries.
*/
static int ptrmapCheckPages(MemPage **apPage, int nPage){
  int i, j;

  int usableSpace;             /* Bytes in pPage beyond the header */
  int pageFlags;               /* Value of pPage->aData[0] */
  int subtotal;                /* Subtotal of bytes in cells on one page */
  int iSpace1 = 0;             /* First unused byte of aSpace1[] */
  int iOvflSpace = 0;          /* First unused byte of aOvflSpace[] */
  int szScratch;               /* Size of scratch memory requested */
  MemPage *apOld[NB];          /* pPage and up to two siblings */

  MemPage *apNew[NB+2];        /* pPage and up to NB siblings after balancing */
  u8 *pRight;                  /* Location in parent of right-sibling pointer */
  u8 *apDiv[NB-1];             /* Divider cells in pParent */
  int cntNew[NB+2];            /* Index in aCell[] of cell after i-th page */
  int szNew[NB+2];             /* Combined size of cells place on i-th page */
  u8 **apCell = 0;             /* All cells begin balanced */
  u16 *szCell;                 /* Local size of all cells in apCell[] */
  u8 *aSpace1;                 /* Space for copies of dividers cells */
  Pgno pgno;                   /* Temp var to store a page number in */

  int aShiftLeft[NB+2];
  int aShiftRight[NB+2];
  u8 abDone[NB+2];
  Pgno aPgno[NB+2];
  u16 aPgFlags[NB+2];

  memset(abDone, 0, sizeof(abDone));
  pBt = pParent->pBt;
  assert( sqlite3_mutex_held(pBt->mutex) );
  assert( sqlite3PagerIswriteable(pParent->pDbPage) );

#if 0
  TRACE(("BALANCE: begin page %d child of %d\n", pPage->pgno, pParent->pgno));
#endif

  /* Make nMaxCells a multiple of 4 in order to preserve 8-byte
  ** alignment */
  nMaxCells = (nMaxCells + 3)&~3;

  /*
  ** Allocate space for memory structures
  */

  szScratch =
       nMaxCells*sizeof(u8*)                       /* apCell */
     + nMaxCells*sizeof(u16)                       /* szCell */

     + pBt->pageSize;                              /* aSpace1 */
  apCell = sqlite3ScratchMalloc( szScratch ); 
  if( apCell==0 ){
    rc = SQLITE_NOMEM;
    goto balance_cleanup;
  }
  szCell = (u16*)&apCell[nMaxCells];
  aSpace1 = (u8*)&szCell[nMaxCells];
  assert( EIGHT_BYTE_ALIGNMENT(aSpace1) );

  /*
  ** Load pointers to all cells on sibling pages and the divider cells
  ** into the local apCell[] array.  Make copies of the divider cells
  ** into space obtained from aSpace1[]. The divider cells have already
  ** been removed from pParent.
  **
  ** If the siblings are on leaf pages, then the child pointers of the
  ** divider cells are stripped from the cells before they are copied
  ** into aSpace1[].  In this way, all cells in apCell[] are without
  ** child pointers.  If siblings are not leaves, then all cell in
  ** apCell[] include child pointers.  Either way, all cells in apCell[]
  ** are alike.
  **
  ** leafCorrection:  4 if pPage is a leaf.  0 if pPage is not a leaf.
  **       leafData:  1 if pPage holds key+data and pParent holds only keys.
  */
  leafCorrection = apOld[0]->leaf*4;
  leafData = apOld[0]->intKeyLeaf;
  for(i=0; i<nOld; i++){
    int limit;





    MemPage *pOld = apOld[i];




    limit = pOld->nCell+pOld->nOverflow;
    if( pOld->nOverflow>0 ){
      for(j=0; j<limit; j++){
        assert( nCell<nMaxCells );
        apCell[nCell] = findOverflowCell(pOld, j);
        szCell[nCell] = cellSizePtr(pOld, apCell[nCell]);

  ** file is corrupt.  The corruption will be detected and reported later
  ** in this procedure so there is no need to act upon it now.
  */
#if 0
  assert( cntNew[0]>0 || (pParent->pgno==1 && pParent->nCell==0) );
#endif

  TRACE(("BALANCE: old: %d(nc=%d) %d(nc=%d) %d(nc=%d)\n",
    apOld[0]->pgno, apOld[0]->nCell,
    nOld>=2 ? apOld[1]->pgno : 0, nOld>=2 ? apOld[1]->nCell : 0,
    nOld>=3 ? apOld[2]->pgno : 0, nOld>=3 ? apOld[2]->nCell : 0
  ));

  /*
  ** Allocate k new pages.  Reuse old pages where possible.
  */
  if( apOld[0]->pgno<=1 ){
    rc = SQLITE_CORRUPT_BKPT;

      rc = sqlite3PagerWrite(pNew->pDbPage);
      nNew++;
      if( rc ) goto balance_cleanup;
    }else{
      assert( i>0 );
      rc = allocateBtreePage(pBt, &pNew, &pgno, (bBulk ? 1 : pgno), 0);
      if( rc ) goto balance_cleanup;
      zeroPage(pNew, pageFlags);
      apNew[i] = pNew;
      nNew++;

      /* Set the pointer-map entry for the new sibling page. */
      if( ISAUTOVACUUM ){
        ptrmapPut(pBt, pNew->pgno, PTRMAP_BTREE, pParent->pgno, &rc);
        if( rc!=SQLITE_OK ){
          goto balance_cleanup;
        }
      }
    }
  }











  /*
  ** Reassign page numbers so that the new pages are in ascending order. 
  ** This helps to keep entries in the disk file in order so that a scan
  ** of the table is closer to a linear scan through the file. That in turn 
  ** helps the operating system to deliver pages from the disk more rapidly.

  **
  ** An O(n^2) insertion sort algorithm is used, but since n is never more 
  ** than (NB+2) (a small constant), that should not be a problem.

  **
  ** When NB==3, this one optimization makes the database about 25% faster 
  ** for large insertions and deletions.
  */
  for(i=0; i<nNew; i++){
    aPgno[i] = apNew[i]->pgno;
    aPgFlags[i] = apNew[i]->pDbPage->flags;
  }
  for(i=0; i<nNew; i++){
    Pgno iGt = (i==0 ? 0 : apNew[i-1]->pgno);
    Pgno iMin = 0;
    u16 flags = 0;
    for(j=0; j<nNew; j++){
      Pgno iPgno = aPgno[j];
      if( iPgno>iGt && (iMin==0 || iPgno<iMin) ){
        iMin = iPgno;
        flags = aPgFlags[j];
      }
    }
    if( apNew[i]->pgno!=iMin ){

      apNew[i]->pDbPage->flags = flags;
      sqlite3PagerRekey(apNew[i]->pDbPage, iMin);
      apNew[i]->pgno = iMin;
    }
  }

  TRACE(("BALANCE: new: %d(%d nc=%d) %d(%d nc=%d) %d(%d nc=%d) "
         "%d(%d nc=%d) %d(%d nc=%d)\n",
    apNew[0]->pgno, szNew[0], cntNew[0],
    nNew>=2 ? apNew[1]->pgno : 0, nNew>=2 ? szNew[1] : 0,
    nNew>=2 ? cntNew[1] - cntNew[0] - !leafData : 0,
    nNew>=3 ? apNew[2]->pgno : 0, nNew>=3 ? szNew[2] : 0,
    nNew>=3 ? cntNew[2] - cntNew[1] - !leafData : 0,
    nNew>=4 ? apNew[3]->pgno : 0, nNew>=4 ? szNew[3] : 0,
    nNew>=4 ? cntNew[3] - cntNew[2] - !leafData : 0,
    nNew>=5 ? apNew[4]->pgno : 0, nNew>=5 ? szNew[4] : 0,
    nNew>=5 ? cntNew[4] - cntNew[3] - !leafData : 0
  ));

  assert( sqlite3PagerIswriteable(pParent->pDbPage) );
  put4byte(pRight, apNew[nNew-1]->pgno);

  j = 0;
  for(i=0; i<nNew; i++){
    /* At this point, "j" is the apCell[] index of the first cell currently
    ** stored on page apNew[i]. Or, if apNew[i] was not one of the original 
    ** sibling pages, "j" should be set to nCell. Variable iFirst is set
    ** to the apCell[] index of the first cell that will appear on the
    ** page following this balancing operation.  */
    int iFirst = (i==0 ? 0 : cntNew[i-1] + !leafData);     /* new first cell */
    assert( i<nOld || j==nCell );
    aShiftLeft[i] = j - iFirst;
    j += apNew[i]->nCell + apNew[i]->nOverflow;
    aShiftRight[i] = cntNew[i] - j;
    assert( i!=nOld-1 || j==nCell );
    if( j<nCell ) j += !leafData;
  }

  /* If the sibling pages are not leaves, ensure that the right-child pointer
  ** of the right-most new sibling page is set to the value that was 
  ** originally in the same field of the right-most old sibling page. */
  if( (pageFlags & PTF_LEAF)==0 && nOld!=nNew ){
    MemPage *pOld = (nNew>nOld ? apNew : apOld)[nOld-1];
    memcpy(&apNew[nNew-1]->aData[8], &pOld->aData[8], 4);
  }

  /* Make any required updates to pointer map entries associated with 
  ** cells stored on sibling pages following the balance operation. Pointer
  ** map entries associated with divider cells are set by the insertCell()
  ** routine. The associated pointer map entries are:
  **
  **   a) if the cell contains a reference to an overflow chain, the
  **      entry associated with the first page in the overflow chain, and
  **
  **   b) if the sibling pages are not leaves, the child page associated
  **      with the cell.
  **
  ** If the sibling pages are not leaves, then the pointer map entry 
  ** associated with the right-child of each sibling may also need to be 
  ** updated. This happens below, after the sibling pages have been 
  ** populated, not here.
  */
  if( ISAUTOVACUUM ){
    MemPage *pNew = apNew[0];
    u8 *aOld = pNew->aData;
    int cntOldNext = pNew->nCell + pNew->nOverflow;
    int usableSize = pBt->usableSize;
    int iNew = 0;
    int iOld = 0;

    for(i=0; i<nCell; i++){

      u8 *pCell = apCell[i];
      if( i==cntOldNext ){
        MemPage *pOld = (++iOld)<nNew ? apNew[iOld] : apOld[iOld];
        cntOldNext += pOld->nCell + pOld->nOverflow + !leafData;
        aOld = pOld->aData;
      }
      if( i==cntNew[iNew] ){
        pNew = apNew[++iNew];
        if( !leafData ) continue;
      }

      /* Cell pCell is destined for new sibling page pNew. Originally, it
      ** was either part of sibling page iOld (possibly an overflow page), 
      ** or else the divider cell to the left of sibling page iOld. So,
      ** if sibling page iOld had the same page number as pNew, and if

      ** pCell really was a part of sibling page iOld (not a divider or
      ** overflow cell), we can skip updating the pointer map entries.  */
      if( pNew->pgno!=aPgno[iOld] || pCell<aOld || pCell>=&aOld[usableSize] ){
        if( !leafCorrection ){
          ptrmapPut(pBt, get4byte(pCell), PTRMAP_BTREE, pNew->pgno, &rc);
        }
        if( szCell[i]>pNew->minLocal ){
          ptrmapPutOvflPtr(pNew, pCell, &rc);
        }
      }
    }
  }

  /* Insert new divider cells into pParent. */
  for(i=0; i<nNew-1; i++){
    u8 *pCell;
    u8 *pTemp;
    int sz;
    MemPage *pNew = apNew[i];
    j = cntNew[i];










    assert( j<nMaxCells );
    pCell = apCell[j];
    sz = szCell[j] + leafCorrection;
    pTemp = &aOvflSpace[iOvflSpace];
    if( !pNew->leaf ){
      memcpy(&pNew->aData[8], pCell, 4);
    }else if( leafData ){
      /* If the tree is a leaf-data tree, and the siblings are leaves, 
      ** then there is no divider cell in apCell[]. Instead, the divider 
      ** cell consists of the integer key for the right-most cell of 
      ** the sibling-page assembled above only.
      */
      CellInfo info;
      j--;
      btreeParseCellPtr(pNew, apCell[j], &info);
      pCell = pTemp;
      sz = 4 + putVarint(&pCell[4], info.nKey);
      pTemp = 0;
    }else{
      pCell -= 4;
      /* Obscure case for non-leaf-data trees: If the cell at pCell was
      ** previously stored on a leaf node, and its reported size was 4
      ** bytes, then it may actually be smaller than this 
      ** (see btreeParseCellPtr(), 4 bytes is the minimum size of
      ** any cell). But it is important to pass the correct size to 
      ** insertCell(), so reparse the cell now.
      **
      ** Note that this can never happen in an SQLite data file, as all
      ** cells are at least 4 bytes. It only happens in b-trees used
      ** to evaluate "IN (SELECT ...)" and similar clauses.
      */
      if( szCell[j]==4 ){
        assert(leafCorrection==4);
        sz = cellSizePtr(pParent, pCell);
      }
    }
    iOvflSpace += sz;
    assert( sz<=pBt->maxLocal+23 );
    assert( iOvflSpace <= (int)pBt->pageSize );
    insertCell(pParent, nxDiv+i, pCell, sz, pTemp, pNew->pgno, &rc);
    if( rc!=SQLITE_OK ) goto balance_cleanup;
    assert( sqlite3PagerIswriteable(pParent->pDbPage) );
  }

  /* Now update the actual sibling pages. The order in which they are updated
  ** is important, as this code needs to avoid disrupting any page from which
  ** cells may still to be read. In practice, this means:
  **
  **   1) If the aShiftLeft[] entry is less than 0, it is not safe to 
  **      update the page until the page to the left of the current page
  **      (apNew[i-1]) has already been updated.
  **
  **   2) If the aShiftRight[] entry is less than 0, it is not safe to 
  **      update the page until the page to the right of the current page
  **      (apNew[i+1]) has already been updated.
  **
  ** If neither of the above apply, the page is safe to update.
  */
  assert( aShiftRight[nNew-1]>=0 && aShiftLeft[0]==0 );
  for(i=0; i<nNew*2; i++){
    int iPg = (i>=nNew ? i-nNew : nNew-1-i);
    if( abDone[iPg]==0 
     && (aShiftLeft[iPg]>=0 || abDone[iPg-1])
     && (aShiftRight[iPg]>=0 || abDone[iPg+1])
    ){
      MemPage *pNew = apNew[iPg];
      int iLeft = ((iPg==0) ? 0 : cntNew[iPg-1] + !leafData);
      rebuildPage(pNew, 
          aShiftLeft[iPg] < 0 ? (aShiftLeft[iPg]*-1) : 0,
          cntNew[iPg] - iLeft,
          &apCell[iLeft],
          &szCell[iLeft]
      );
      abDone[iPg] = 1;
      assert( pNew->nOverflow==0 );
      assert( pNew->nCell==(cntNew[iPg] - (iPg==0?0:cntNew[iPg-1]+!leafData)) );
    }
  }
  assert( memcmp(abDone, "\01\01\01\01\01", nNew)==0 );

  assert( nOld>0 );
  assert( nNew>0 );





  if( isRoot && pParent->nCell==0 && pParent->hdrOffset<=apNew[0]->nFree ){
    /* The root page of the b-tree now contains no cells. The only sibling
    ** page is the right-child of the parent. Copy the contents of the
    ** child page into the parent, decreasing the overall height of the
    ** b-tree structure by one. This is described as the "balance-shallower"
    ** sub-algorithm in some documentation.

    ** image.  */
    assert( nNew==1 );
    assert( apNew[0]->nFree == 
        (get2byte(&apNew[0]->aData[5])-apNew[0]->cellOffset-apNew[0]->nCell*2) 
    );
    copyNodeContent(apNew[0], pParent, &rc);
    freePage(apNew[0], &rc);
  }else if( ISAUTOVACUUM && !leafCorrection ){











    /* Fix the pointer map entries associated with the right-child of each










































    ** sibling page. All other pointer map entries have already been taken









    ** care of.  */

























    for(i=0; i<nNew; i++){
      u32 key = get4byte(&apNew[i]->aData[8]);
      ptrmapPut(pBt, key, PTRMAP_BTREE, apNew[i]->pgno, &rc);
    }
  }

  assert( pParent->isInit );
  TRACE(("BALANCE: finished: old=%d new=%d cells=%d\n",
          nOld, nNew, nCell));

  /* Free any old pages that were not reused as new pages.
  */
  for(i=nNew; i<nOld; i++){
    freePage(apOld[i], &rc);
  }

#if 1
  if( ISAUTOVACUUM && rc==SQLITE_OK && apNew[0]->isInit ){
    /* The ptrmapCheckPages() contains assert() statements that verify that
    ** all pointer map pages are set correctly. This is helpful while 
    ** debugging. This is usually disabled because a corrupt database may
    ** cause an assert() statement to fail.  */
    ptrmapCheckPages(apNew, nNew);
    ptrmapCheckPages(&pParent, 1);

  }
#endif




  /*
  ** Cleanup before returning.
  */
balance_cleanup:
  sqlite3ScratchFree(apCell);
  for(i=0; i<nOld; i++){

    pPgHdr->flags |= PGHDR_NEED_SYNC;
    sqlite3PcacheMakeDirty(pPgHdr);
    sqlite3PagerUnrefNotNull(pPgHdr);
  }

  return SQLITE_OK;
}

void sqlite3PagerRekey(DbPage *pPage, Pgno iNew){
  PgHdr *pPg = (PgHdr*)pPage;
  assert( pPg->flags & PGHDR_DIRTY );
  assert( !subjRequiresPage(pPg) );
  sqlite3PcacheMove(pPg, iNew);
}

#endif

/*
** Return a pointer to the data for the specified page.
*/
void *sqlite3PagerGetData(DbPage *pPg){
  assert( pPg->nRef>0 || pPg->pPager->memDb );

** is empty, return 0.
*/
int sqlite3PagerWalFramesize(Pager *pPager){
  assert( pPager->eState>=PAGER_READER );
  return sqlite3WalFramesize(pPager->pWal);
}
#endif


#endif /* SQLITE_OMIT_DISKIO */

int sqlite3PagerIsMemdb(Pager*);
void sqlite3PagerCacheStat(Pager *, int, int, int *);
void sqlite3PagerClearCache(Pager *);
int sqlite3SectorSize(sqlite3_file *);

/* Functions used to truncate the database file. */
void sqlite3PagerTruncateImage(Pager*,Pgno);

void sqlite3PagerRekey(DbPage*, Pgno);

#if defined(SQLITE_HAS_CODEC) && !defined(SQLITE_OMIT_WAL)
void *sqlite3PagerCodec(DbPage *);
#endif

/* Functions to support testing and debugging. */
#if !defined(NDEBUG) || defined(SQLITE_TEST)