SQLite

** pages that do contain overflow cells.
*/
static u8 *findOverflowCell(MemPage *pPage, int iCell){
  int i;
  assert( sqlite3_mutex_held(pPage->pBt->mutex) );
  for(i=pPage->nOverflow-1; i>=0; i--){
    int k;
    struct _OvflCell *pOvfl;
    pOvfl = &pPage->aOvfl[i];
    k = pPage->aiOvfl[i];
    k = pOvfl->idx;
    if( k<=iCell ){
      if( k==iCell ){
        return pOvfl->pCell;
        return pPage->apOvfl[i];
      }
      iCell--;
    }
  }
  return findCell(pPage, iCell);
}

/*
** Insert a new cell on pPage at cell index "i".  pCell points to the
** content of the cell.
**
** If the cell content will fit on the page, then put it there.  If it
** will not fit, then make a copy of the cell content into pTemp if
** pTemp is not null.  Regardless of pTemp, allocate a new entry
** in pPage->aOvfl[] and make it point to the cell content (either
** in pPage->apOvfl[] and make it point to the cell content (either
** in pTemp or the original pCell) and also record its index. 
** Allocating a new entry in pPage->aCell[] implies that 
** pPage->nOverflow is incremented.
**
** If nSkip is non-zero, then do not copy the first nSkip bytes of the
** cell. The caller will overwrite them after this function returns. If
** nSkip is non-zero, then pCell may not point to an invalid memory location 

  int nSkip = (iChild ? 4 : 0);

  if( *pRC ) return;

  assert( i>=0 && i<=pPage->nCell+pPage->nOverflow );
  assert( pPage->nCell<=MX_CELL(pPage->pBt) && MX_CELL(pPage->pBt)<=10921 );
  assert( pPage->nOverflow<=ArraySize(pPage->aOvfl) );
  assert( pPage->nOverflow<=ArraySize(pPage->apOvfl) );
  assert( ArraySize(pPage->apOvfl)==ArraySize(pPage->aiOvfl) );
  assert( sqlite3_mutex_held(pPage->pBt->mutex) );
  /* The cell should normally be sized correctly.  However, when moving a
  ** malformed cell from a leaf page to an interior page, if the cell size
  ** wanted to be less than 4 but got rounded up to 4 on the leaf, then size
  ** might be less than 8 (leaf-size + pointer) on the interior node.  Hence
  ** the term after the || in the following assert(). */
  assert( sz==cellSizePtr(pPage, pCell) || (sz==8 && iChild>0) );
  if( pPage->nOverflow || sz+2>pPage->nFree ){
    if( pTemp ){
      memcpy(pTemp+nSkip, pCell+nSkip, sz-nSkip);
      pCell = pTemp;
    }
    if( iChild ){
      put4byte(pCell, iChild);
    }
    j = pPage->nOverflow++;
    assert( j<(int)(sizeof(pPage->aOvfl)/sizeof(pPage->aOvfl[0])) );
    pPage->aOvfl[j].pCell = pCell;
    pPage->aOvfl[j].idx = (u16)i;
    assert( j<(int)(sizeof(pPage->apOvfl)/sizeof(pPage->apOvfl[0])) );
    pPage->apOvfl[j] = pCell;
    pPage->aiOvfl[j] = (u16)i;
  }else{
    int rc = sqlite3PagerWrite(pPage->pDbPage);
    if( rc!=SQLITE_OK ){
      *pRC = rc;
      return;
    }
    assert( sqlite3PagerIswriteable(pPage->pDbPage) );

  ** may be inserted. If both these operations are successful, proceed.
  */
  rc = allocateBtreePage(pBt, &pNew, &pgnoNew, 0, 0);

  if( rc==SQLITE_OK ){

    u8 *pOut = &pSpace[4];
    u8 *pCell = pPage->aOvfl[0].pCell;
    u8 *pCell = pPage->apOvfl[0];
    u16 szCell = cellSizePtr(pPage, pCell);
    u8 *pStop;

    assert( sqlite3PagerIswriteable(pNew->pDbPage) );
    assert( pPage->aData[0]==(PTF_INTKEY|PTF_LEAFDATA|PTF_LEAF) );
    zeroPage(pNew, PTF_INTKEY|PTF_LEAFDATA|PTF_LEAF);
    assemblePage(pNew, 1, &pCell, &szCell);

** on page pFrom to page pTo. If page pFrom was not a leaf page, then
** the pointer-map entries for each child page are updated so that the
** parent page stored in the pointer map is page pTo. If pFrom contained
** any cells with overflow page pointers, then the corresponding pointer
** map entries are also updated so that the parent page is page pTo.
**
** If pFrom is currently carrying any overflow cells (entries in the
** MemPage.aOvfl[] array), they are not copied to pTo. 
** MemPage.apOvfl[] array), they are not copied to pTo. 
**
** Before returning, page pTo is reinitialized using btreeInitPage().
**
** The performance of this function is not critical. It is only used by 
** the balance_shallower() and balance_deeper() procedures, neither of
** which are called often under normal circumstances.
*/

  /* At this point pParent may have at most one overflow cell. And if
  ** this overflow cell is present, it must be the cell with 
  ** index iParentIdx. This scenario comes about when this function
  ** is called (indirectly) from sqlite3BtreeDelete().
  */
  assert( pParent->nOverflow==0 || pParent->nOverflow==1 );
  assert( pParent->nOverflow==0 || pParent->aOvfl[0].idx==iParentIdx );
  assert( pParent->nOverflow==0 || pParent->aiOvfl[0]==iParentIdx );

  if( !aOvflSpace ){
    return SQLITE_NOMEM;
  }

  /* Find the sibling pages to balance. Also locate the cells in pParent 
  ** that divide the siblings. An attempt is made to find NN siblings on 

    if( rc ){
      memset(apOld, 0, (i+1)*sizeof(MemPage*));
      goto balance_cleanup;
    }
    nMaxCells += 1+apOld[i]->nCell+apOld[i]->nOverflow;
    if( (i--)==0 ) break;

    if( i+nxDiv==pParent->aOvfl[0].idx && pParent->nOverflow ){
      apDiv[i] = pParent->aOvfl[0].pCell;
    if( i+nxDiv==pParent->aiOvfl[0] && pParent->nOverflow ){
      apDiv[i] = pParent->apOvfl[0];
      pgno = get4byte(apDiv[i]);
      szNew[i] = cellSizePtr(pParent, apDiv[i]);
      pParent->nOverflow = 0;
    }else{
      apDiv[i] = findCell(pParent, i+nxDiv-pParent->nOverflow);
      pgno = get4byte(apDiv[i]);
      szNew[i] = cellSizePtr(pParent, apDiv[i]);

    ** setting a pointer map entry is a relatively expensive operation, this
    ** code only sets pointer map entries for child or overflow pages that have
    ** actually moved between pages.  */
    MemPage *pNew = apNew[0];
    MemPage *pOld = apCopy[0];
    int nOverflow = pOld->nOverflow;
    int iNextOld = pOld->nCell + nOverflow;
    int iOverflow = (nOverflow ? pOld->aOvfl[0].idx : -1);
    int iOverflow = (nOverflow ? pOld->aiOvfl[0] : -1);
    j = 0;                             /* Current 'old' sibling page */
    k = 0;                             /* Current 'new' sibling page */
    for(i=0; i<nCell; i++){
      int isDivider = 0;
      while( i==iNextOld ){
        /* Cell i is the cell immediately following the last cell on old
        ** sibling page j. If the siblings are not leaf pages of an
        ** intkey b-tree, then cell i was a divider cell. */
        assert( j+1 < ArraySize(apCopy) );
        pOld = apCopy[++j];
        iNextOld = i + !leafData + pOld->nCell + pOld->nOverflow;
        if( pOld->nOverflow ){
          nOverflow = pOld->nOverflow;
          iOverflow = i + !leafData + pOld->aOvfl[0].idx;
          iOverflow = i + !leafData + pOld->aiOvfl[0];
        }
        isDivider = !leafData;  
      }

      assert(nOverflow>0 || iOverflow<i );
      assert(nOverflow<2 || pOld->aOvfl[0].idx==pOld->aOvfl[1].idx-1);
      assert(nOverflow<3 || pOld->aOvfl[1].idx==pOld->aOvfl[2].idx-1);
      assert(nOverflow<2 || pOld->aiOvfl[0]==pOld->aiOvfl[1]-1);
      assert(nOverflow<3 || pOld->aiOvfl[1]==pOld->aiOvfl[2]-1);
      if( i==iOverflow ){
        isDivider = 1;
        if( (--nOverflow)>0 ){
          iOverflow++;
        }
      }

  assert( sqlite3PagerIswriteable(pChild->pDbPage) );
  assert( sqlite3PagerIswriteable(pRoot->pDbPage) );
  assert( pChild->nCell==pRoot->nCell );

  TRACE(("BALANCE: copy root %d into %d\n", pRoot->pgno, pChild->pgno));

  /* Copy the overflow cells from pRoot to pChild */
  memcpy(pChild->aiOvfl, pRoot->aiOvfl,
  memcpy(pChild->aOvfl, pRoot->aOvfl, pRoot->nOverflow*sizeof(pRoot->aOvfl[0]));
         pRoot->nOverflow*sizeof(pRoot->aiOvfl[0]));
  memcpy(pChild->apOvfl, pRoot->apOvfl,
         pRoot->nOverflow*sizeof(pRoot->apOvfl[0]));
  pChild->nOverflow = pRoot->nOverflow;

  /* Zero the contents of pRoot. Then install pChild as the right-child. */
  zeroPage(pRoot, pChild->aData[0] & ~PTF_LEAF);
  put4byte(&pRoot->aData[pRoot->hdrOffset+8], pgnoChild);

  *ppChild = pChild;

      int const iIdx = pCur->aiIdx[iPage-1];

      rc = sqlite3PagerWrite(pParent->pDbPage);
      if( rc==SQLITE_OK ){
#ifndef SQLITE_OMIT_QUICKBALANCE
        if( pPage->hasData
         && pPage->nOverflow==1
         && pPage->aOvfl[0].idx==pPage->nCell
         && pPage->aiOvfl[0]==pPage->nCell
         && pParent->pgno!=1
         && pParent->nCell==iIdx
        ){
          /* Call balance_quick() to create a new sibling of pPage on which
          ** to store the overflow cell. balance_quick() inserts a new cell
          ** into pParent, which may cause pParent overflow. If this
          ** happens, the next interation of the do-loop will balance pParent 

  u8 max1bytePayload;  /* min(maxLocal,127) */
  u16 maxLocal;        /* Copy of BtShared.maxLocal or BtShared.maxLeaf */
  u16 minLocal;        /* Copy of BtShared.minLocal or BtShared.minLeaf */
  u16 cellOffset;      /* Index in aData of first cell pointer */
  u16 nFree;           /* Number of free bytes on the page */
  u16 nCell;           /* Number of cells on this page, local and ovfl */
  u16 maskPage;        /* Mask for page offset */
  struct _OvflCell {   /* Cells that will not fit on aData[] */
    u8 *pCell;          /* Pointers to the body of the overflow cell */
    u16 idx;            /* Insert this cell before idx-th non-overflow cell */
  } aOvfl[5];
  u16 aiOvfl[5];       /* Insert the i-th overflow cell before the aiOvfl-th
                       ** non-overflow cell */
  u8 *apOvfl[5];       /* Pointers to the body of overflow cells */
  BtShared *pBt;       /* Pointer to BtShared that this page is part of */
  u8 *aData;           /* Pointer to disk image of the page data */
  u8 *aDataEnd;        /* One byte past the end of usable data */
  u8 *aCellIdx;        /* The cell index area */
  DbPage *pDbPage;     /* Pager page handle */
  Pgno pgno;           /* Page number for this page */
};

** found at self->pBt->mutex. 
*/
struct BtCursor {
  Btree *pBtree;            /* The Btree to which this cursor belongs */
  BtShared *pBt;            /* The BtShared this cursor points to */
  BtCursor *pNext, *pPrev;  /* Forms a linked list of all cursors */
  struct KeyInfo *pKeyInfo; /* Argument passed to comparison function */
#ifndef SQLITE_OMIT_INCRBLOB
  Pgno *aOverflow;          /* Cache of overflow page locations */
#endif
  Pgno pgnoRoot;            /* The root page of this tree */
  sqlite3_int64 cachedRowid; /* Next rowid cache.  0 means not valid */
  CellInfo info;            /* A parse of the cell we are pointing at */
  i64 nKey;        /* Size of pKey, or last integer key */
  void *pKey;      /* Saved key that was cursor's last known position */
  int skipNext;    /* Prev() is noop if negative. Next() is noop if positive */
  u8 wrFlag;                /* True if writable */
  u8 atLast;                /* Cursor pointing to the last entry */
  u8 validNKey;             /* True if info.nKey is valid */
  u8 eState;                /* One of the CURSOR_XXX constants (see below) */
#ifndef SQLITE_OMIT_INCRBLOB
  Pgno *aOverflow;          /* Cache of overflow page locations */
  u8 isIncrblobHandle;      /* True if this cursor is an incr. io handle */
#endif
  i16 iPage;                            /* Index of current page in apPage */
  u16 aiIdx[BTCURSOR_MAX_DEPTH];        /* Current index in apPage[i] */
  MemPage *apPage[BTCURSOR_MAX_DEPTH];  /* Pages from root to current page */
};

** This structure is passed around through all the sanity checking routines
** in order to keep track of some global state information.
*/
typedef struct IntegrityCk IntegrityCk;
struct IntegrityCk {
  BtShared *pBt;    /* The tree being checked out */
  Pager *pPager;    /* The associated pager.  Also accessible by pBt->pPager */
  Pgno nPage;       /* Number of pages in the database */
  int *anRef;       /* Number of times each page is referenced */
  Pgno nPage;       /* Number of pages in the database */
  int mxErr;        /* Stop accumulating errors when this reaches zero */
  int nErr;         /* Number of messages written to zErrMsg so far */
  int mallocFailed; /* A memory allocation error has occurred */
  StrAccum errMsg;  /* Accumulate the error message text here */
};

/*
** Routines to read or write a two- and four-byte big-endian integer values.
*/
#define get2byte(x)   ((x)[0]<<8 | (x)[1])
#define put2byte(p,v) ((p)[0] = (u8)((v)>>8), (p)[1] = (u8)(v))
#define get4byte sqlite3Get4byte
#define put4byte sqlite3Put4byte

  PGroup *pGroup;                     /* PGroup this cache belongs to */
  int szPage;                         /* Size of allocated pages in bytes */
  int szExtra;                        /* Size of extra space in bytes */
  int bPurgeable;                     /* True if cache is purgeable */
  unsigned int nMin;                  /* Minimum number of pages reserved */
  unsigned int nMax;                  /* Configured "cache_size" value */
  unsigned int n90pct;                /* nMax*9/10 */
  unsigned int iMaxKey;               /* Largest key seen since xTruncate() */

  /* Hash table of all pages. The following variables may only be accessed
  ** when the accessor is holding the PGroup mutex.
  */
  unsigned int nRecyclable;           /* Number of pages in the LRU list */
  unsigned int nPage;                 /* Total number of pages in apHash */
  unsigned int nHash;                 /* Number of slots in apHash[] */
  PgHdr1 **apHash;                    /* Hash table for fast lookup by key */

  unsigned int iMaxKey;               /* Largest key seen since xTruncate() */
};

/*
** Each cache entry is represented by an instance of the following 
** structure. Unless SQLITE_PCACHE_SEPARATE_HEADER is defined, a buffer of
** PgHdr1.pCache->szPage bytes is allocated directly before this structure 
** in memory.

  int isInit;                    /* True if initialized */
  int szSlot;                    /* Size of each free slot */
  int nSlot;                     /* The number of pcache slots */
  int nReserve;                  /* Try to keep nFreeSlot above this */
  void *pStart, *pEnd;           /* Bounds of pagecache malloc range */
  /* Above requires no mutex.  Use mutex below for variable that follow. */
  sqlite3_mutex *mutex;          /* Mutex for accessing the following: */
  int nFreeSlot;                 /* Number of unused pcache slots */
  PgFreeslot *pFree;             /* Free page blocks */
  int nFreeSlot;                 /* Number of unused pcache slots */
  /* The following value requires a mutex to change.  We skip the mutex on
  ** reading because (1) most platforms read a 32-bit integer atomically and
  ** (2) even if an incorrect value is read, no great harm is done since this
  ** is really just an optimization. */
  int bUnderPressure;            /* True if low on PAGECACHE memory */
} pcache1_g;