DBD-SQLite2

 view release on metacpan or  search on metacpan

btree.c  view on Meta::CPAN

struct FreelistInfo {
  int nFree;
  Pgno aFree[(OVERFLOW_SIZE-sizeof(int))/sizeof(Pgno)];
};

/*
** For every page in the database file, an instance of the following structure
** is stored in memory.  The u.aDisk[] array contains the raw bits read from
** the disk.  The rest is auxiliary information held in memory only. The
** auxiliary info is only valid for regular database pages - it is not
** used for overflow pages and pages on the freelist.
**
** Of particular interest in the auxiliary info is the apCell[] entry.  Each
** apCell[] entry is a pointer to a Cell structure in u.aDisk[].  The cells are
** put in this array so that they can be accessed in constant time, rather
** than in linear time which would be needed if we had to walk the linked 
** list on every access.
**
** Note that apCell[] contains enough space to hold up to two more Cells
** than can possibly fit on one page.  In the steady state, every apCell[]
** points to memory inside u.aDisk[].  But in the middle of an insert
** operation, some apCell[] entries may temporarily point to data space
** outside of u.aDisk[].  This is a transient situation that is quickly
** resolved.  But while it is happening, it is possible for a database
** page to hold as many as two more cells than it might otherwise hold.
** The extra two entries in apCell[] are an allowance for this situation.
**
** The pParent field points back to the parent page.  This allows us to
** walk up the BTree from any leaf to the root.  Care must be taken to
** unref() the parent page pointer when this page is no longer referenced.
** The pageDestructor() routine handles that chore.
*/
struct MemPage {
  union u_page_data {
    char aDisk[SQLITE_PAGE_SIZE];  /* Page data stored on disk */
    PageHdr hdr;                   /* Overlay page header */
  } u;
  u8 isInit;                     /* True if auxiliary data is initialized */
  u8 idxShift;                   /* True if apCell[] indices have changed */
  u8 isOverfull;                 /* Some apCell[] points outside u.aDisk[] */
  MemPage *pParent;              /* The parent of this page.  NULL for root */
  int idxParent;                 /* Index in pParent->apCell[] of this node */
  int nFree;                     /* Number of free bytes in u.aDisk[] */
  int nCell;                     /* Number of entries on this page */
  Cell *apCell[MX_CELL+2];       /* All data entires in sorted order */
};

/*
** The in-memory image of a disk page has the auxiliary information appended
** to the end.  EXTRA_SIZE is the number of bytes of space needed to hold
** that extra information.
*/
#define EXTRA_SIZE (sizeof(MemPage)-sizeof(union u_page_data))

/*
** Everything we need to know about an open database
*/
struct Btree {
  BtOps *pOps;          /* Function table */
  Pager *pPager;        /* The page cache */
  BtCursor *pCursor;    /* A list of all open cursors */
  PageOne *page1;       /* First page of the database */
  u8 inTrans;           /* True if a transaction is in progress */
  u8 inCkpt;            /* True if there is a checkpoint on the transaction */
  u8 readOnly;          /* True if the underlying file is readonly */
  u8 needSwab;          /* Need to byte-swapping */
};
typedef Btree Bt;

/*
** A cursor is a pointer to a particular entry in the BTree.
** The entry is identified by its MemPage and the index in
** MemPage.apCell[] of the entry.
*/
struct BtCursor {
  BtCursorOps *pOps;        /* Function table */
  Btree *pBt;               /* The Btree to which this cursor belongs */
  BtCursor *pNext, *pPrev;  /* Forms a linked list of all cursors */
  BtCursor *pShared;        /* Loop of cursors with the same root page */
  Pgno pgnoRoot;            /* The root page of this tree */
  MemPage *pPage;           /* Page that contains the entry */
  int idx;                  /* Index of the entry in pPage->apCell[] */
  u8 wrFlag;                /* True if writable */
  u8 eSkip;                 /* Determines if next step operation is a no-op */
  u8 iMatch;                /* compare result from last sqliteBtreeMoveto() */
};

/*
** Legal values for BtCursor.eSkip.
*/
#define SKIP_NONE     0   /* Always step the cursor */
#define SKIP_NEXT     1   /* The next sqliteBtreeNext() is a no-op */
#define SKIP_PREV     2   /* The next sqliteBtreePrevious() is a no-op */
#define SKIP_INVALID  3   /* Calls to Next() and Previous() are invalid */

/* Forward declarations */
static int fileBtreeCloseCursor(BtCursor *pCur);

/*
** Routines for byte swapping.
*/
u16 swab16(u16 x){
  return ((x & 0xff)<<8) | ((x>>8)&0xff);
}
u32 swab32(u32 x){
  return ((x & 0xff)<<24) | ((x & 0xff00)<<8) |
         ((x>>8) & 0xff00) | ((x>>24)&0xff);
}

/*
** Compute the total number of bytes that a Cell needs on the main
** database page.  The number returned includes the Cell header,
** local payload storage, and the pointer to overflow pages (if
** applicable).  Additional space allocated on overflow pages
** is NOT included in the value returned from this routine.
*/
static int cellSize(Btree *pBt, Cell *pCell){
  int n = NKEY(pBt, pCell->h) + NDATA(pBt, pCell->h);
  if( n>MX_LOCAL_PAYLOAD ){
    n = MX_LOCAL_PAYLOAD + sizeof(Pgno);
  }else{
    n = ROUNDUP(n);
  }
  n += sizeof(CellHdr);
  return n;
}

/*
** Defragment the page given.  All Cells are moved to the
** beginning of the page and all free space is collected 
** into one big FreeBlk at the end of the page.
*/
static void defragmentPage(Btree *pBt, MemPage *pPage){
  int pc, i, n;
  FreeBlk *pFBlk;
  char newPage[SQLITE_USABLE_SIZE];

  assert( sqlitepager_iswriteable(pPage) );
  assert( pPage->isInit );
  pc = sizeof(PageHdr);
  pPage->u.hdr.firstCell = SWAB16(pBt, 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
    ** following asserts verify that constraint. */
    assert( Addr(pCell) > Addr(pPage) );
    assert( Addr(pCell) < Addr(pPage) + SQLITE_USABLE_SIZE );

    n = cellSize(pBt, pCell);

btree.c  view on Meta::CPAN

/*
** 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.
**
** zFilename is the name of the database file.  If zFilename is NULL
** a new database with a random name is created.  This randomly named
** database file will be deleted when sqliteBtreeClose() is called.
*/
int sqliteBtreeOpen(
  const char *zFilename,    /* Name of the file containing the BTree database */
  int omitJournal,          /* if TRUE then do not journal this file */
  int nCache,               /* How many pages in the page cache */
  Btree **ppBtree           /* Pointer to new Btree object written here */
){
  Btree *pBt;
  int rc;

  /*
  ** The following asserts make sure that structures used by the btree are
  ** the right size.  This is to guard against size changes that result
  ** when compiling on a different architecture.
  */
  assert( sizeof(u32)==4 );
  assert( sizeof(u16)==2 );
  assert( sizeof(Pgno)==4 );
  assert( sizeof(PageHdr)==8 );
  assert( sizeof(CellHdr)==12 );
  assert( sizeof(FreeBlk)==4 );
  assert( sizeof(OverflowPage)==SQLITE_USABLE_SIZE );
  assert( sizeof(FreelistInfo)==OVERFLOW_SIZE );
  assert( sizeof(ptr)==sizeof(char*) );
  assert( sizeof(uptr)==sizeof(ptr) );

  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,
                        !omitJournal);
  if( rc!=SQLITE_OK ){
    if( pBt->pPager ) sqlitepager_close(pBt->pPager);
    sqliteFree(pBt);
    *ppBtree = 0;
    return rc;
  }
  sqlitepager_set_destructor(pBt->pPager, pageDestructor);
  pBt->pCursor = 0;
  pBt->page1 = 0;
  pBt->readOnly = sqlitepager_isreadonly(pBt->pPager);
  pBt->pOps = &sqliteBtreeOps;
  *ppBtree = pBt;
  return SQLITE_OK;
}

/*
** Close an open database and invalidate all cursors.
*/
static int fileBtreeClose(Btree *pBt){
  while( pBt->pCursor ){
    fileBtreeCloseCursor(pBt->pCursor);
  }
  sqlitepager_close(pBt->pPager);
  sqliteFree(pBt);
  return SQLITE_OK;
}

/*
** Change the limit on the number of pages allowed in the cache.
**
** The maximum number of cache pages is set to the absolute
** value of mxPage.  If mxPage is negative, the pager will
** operate asynchronously - it will not stop to do fsync()s
** to insure data is written to the disk surface before
** continuing.  Transactions still work if synchronous is off,
** and the database cannot be corrupted if this program
** crashes.  But if the operating system crashes or there is
** an abrupt power failure when synchronous is off, the database
** could be left in an inconsistent and unrecoverable state.
** Synchronous is on by default so database corruption is not
** normally a worry.
*/
static int fileBtreeSetCacheSize(Btree *pBt, int mxPage){
  sqlitepager_set_cachesize(pBt->pPager, mxPage);
  return SQLITE_OK;
}

/*
** Change the way data is synced to disk in order to increase or decrease
** how well the database resists damage due to OS crashes and power
** failures.  Level 1 is the same as asynchronous (no syncs() occur and
** there is a high probability of damage)  Level 2 is the default.  There
** is a very low but non-zero probability of damage.  Level 3 reduces the
** probability of damage to near zero but with a write performance reduction.
*/
static int fileBtreeSetSafetyLevel(Btree *pBt, int level){
  sqlitepager_set_safety_level(pBt->pPager, level);
  return SQLITE_OK;
}

/*
** Get a reference to page1 of the database file.  This will
** also acquire a readlock on that file.
**
** SQLITE_OK is returned on success.  If the file is not a
** well-formed database file, then SQLITE_CORRUPT is returned.
** SQLITE_BUSY is returned if the database is locked.  SQLITE_NOMEM
** is returned if we run out of memory.  SQLITE_PROTOCOL is returned
** if there is a locking protocol violation.
*/
static int lockBtree(Btree *pBt){
  int rc;
  if( pBt->page1 ) return SQLITE_OK;
  rc = sqlitepager_get(pBt->pPager, 1, (void**)&pBt->page1);
  if( rc!=SQLITE_OK ) return rc;

  /* Do some checking to help insure the file we opened really is
  ** a valid database file. 
  */
  if( sqlitepager_pagecount(pBt->pPager)>0 ){
    PageOne *pP1 = pBt->page1;
    if( strcmp(pP1->zMagic,zMagicHeader)!=0 ||
          (pP1->iMagic!=MAGIC && swab32(pP1->iMagic)!=MAGIC) ){
      rc = SQLITE_NOTADB;
      goto page1_init_failed;
    }
    pBt->needSwab = pP1->iMagic!=MAGIC;
  }
  return rc;

page1_init_failed:
  sqlitepager_unref(pBt->page1);
  pBt->page1 = 0;
  return rc;
}

/*
** If there are no outstanding cursors and we are not in the middle
** of a transaction but there is a read lock on the database, then
** this routine unrefs the first page of the database file which 
** has the effect of releasing the read lock.
**
** If there are any outstanding cursors, this routine is a no-op.
**
** If there is a transaction in progress, this routine is a no-op.
*/
static void unlockBtreeIfUnused(Btree *pBt){
  if( pBt->inTrans==0 && pBt->pCursor==0 && pBt->page1!=0 ){
    sqlitepager_unref(pBt->page1);
    pBt->page1 = 0;
    pBt->inTrans = 0;
    pBt->inCkpt = 0;
  }
}

/*
** Create a new database by initializing the first two pages of the
** file.
*/
static int newDatabase(Btree *pBt){
  MemPage *pRoot;
  PageOne *pP1;
  int rc;
  if( sqlitepager_pagecount(pBt->pPager)>1 ) return SQLITE_OK;
  pP1 = pBt->page1;
  rc = sqlitepager_write(pBt->page1);
  if( rc ) return rc;
  rc = sqlitepager_get(pBt->pPager, 2, (void**)&pRoot);
  if( rc ) return rc;
  rc = sqlitepager_write(pRoot);
  if( rc ){
    sqlitepager_unref(pRoot);
    return rc;
  }
  strcpy(pP1->zMagic, zMagicHeader);
  if( btree_native_byte_order ){
    pP1->iMagic = MAGIC;
    pBt->needSwab = 0;
  }else{
    pP1->iMagic = swab32(MAGIC);
    pBt->needSwab = 1;
  }
  zeroPage(pBt, pRoot);
  sqlitepager_unref(pRoot);
  return SQLITE_OK;
}

/*
** Attempt to start a new transaction.
**
** A transaction must be started before attempting any changes
** to the database.  None of the following routines will work
** unless a transaction is started first:
**
**      sqliteBtreeCreateTable()
**      sqliteBtreeCreateIndex()
**      sqliteBtreeClearTable()
**      sqliteBtreeDropTable()
**      sqliteBtreeInsert()
**      sqliteBtreeDelete()
**      sqliteBtreeUpdateMeta()
*/
static int fileBtreeBeginTrans(Btree *pBt){
  int rc;
  if( pBt->inTrans ) return SQLITE_ERROR;
  if( pBt->readOnly ) return SQLITE_READONLY;
  if( pBt->page1==0 ){
    rc = lockBtree(pBt);
    if( rc!=SQLITE_OK ){
      return rc;
    }
  }
  rc = sqlitepager_begin(pBt->page1);
  if( rc==SQLITE_OK ){
    rc = newDatabase(pBt);
  }
  if( rc==SQLITE_OK ){
    pBt->inTrans = 1;
    pBt->inCkpt = 0;
  }else{
    unlockBtreeIfUnused(pBt);
  }
  return rc;
}

/*
** Commit the transaction currently in progress.
**
** This will release the write lock on the database file.  If there
** are no active cursors, it also releases the read lock.
*/
static int fileBtreeCommit(Btree *pBt){
  int rc;
  rc = pBt->readOnly ? SQLITE_OK : sqlitepager_commit(pBt->pPager);
  pBt->inTrans = 0;
  pBt->inCkpt = 0;
  unlockBtreeIfUnused(pBt);
  return rc;
}

/*
** Rollback the transaction in progress.  All cursors will be
** invalided by this operation.  Any attempt to use a cursor
** that was open at the beginning of this operation will result
** in an error.
**
** This will release the write lock on the database file.  If there
** are no active cursors, it also releases the read lock.
*/
static int fileBtreeRollback(Btree *pBt){
  int rc;
  BtCursor *pCur;
  if( pBt->inTrans==0 ) return SQLITE_OK;
  pBt->inTrans = 0;
  pBt->inCkpt = 0;
  rc = pBt->readOnly ? SQLITE_OK : sqlitepager_rollback(pBt->pPager);
  for(pCur=pBt->pCursor; pCur; pCur=pCur->pNext){
    if( pCur->pPage && pCur->pPage->isInit==0 ){
      sqlitepager_unref(pCur->pPage);
      pCur->pPage = 0;
    }
  }
  unlockBtreeIfUnused(pBt);
  return rc;
}

/*
** Set the checkpoint for the current transaction.  The checkpoint serves
** as a sub-transaction that can be rolled back independently of the
** main transaction.  You must start a transaction before starting a
** checkpoint.  The checkpoint is ended automatically if the transaction
** commits or rolls back.
**
** Only one checkpoint may be active at a time.  It is an error to try
** to start a new checkpoint if another checkpoint is already active.
*/
static int fileBtreeBeginCkpt(Btree *pBt){
  int rc;
  if( !pBt->inTrans || pBt->inCkpt ){
    return pBt->readOnly ? SQLITE_READONLY : SQLITE_ERROR;
  }
  rc = pBt->readOnly ? SQLITE_OK : sqlitepager_ckpt_begin(pBt->pPager);
  pBt->inCkpt = 1;
  return rc;
}


/*
** Commit a checkpoint to transaction currently in progress.  If no
** checkpoint is active, this is a no-op.
*/
static int fileBtreeCommitCkpt(Btree *pBt){
  int rc;
  if( pBt->inCkpt && !pBt->readOnly ){
    rc = sqlitepager_ckpt_commit(pBt->pPager);
  }else{
    rc = SQLITE_OK;
  }
  pBt->inCkpt = 0;
  return rc;
}

/*
** Rollback the checkpoint to the current transaction.  If there
** is no active checkpoint or transaction, this routine is a no-op.
**
** All cursors will be invalided by this operation.  Any attempt
** to use a cursor that was open at the beginning of this operation
** will result in an error.
*/
static int fileBtreeRollbackCkpt(Btree *pBt){
  int rc;
  BtCursor *pCur;
  if( pBt->inCkpt==0 || pBt->readOnly ) return SQLITE_OK;
  rc = sqlitepager_ckpt_rollback(pBt->pPager);
  for(pCur=pBt->pCursor; pCur; pCur=pCur->pNext){
    if( pCur->pPage && pCur->pPage->isInit==0 ){
      sqlitepager_unref(pCur->pPage);
      pCur->pPage = 0;
    }
  }
  pBt->inCkpt = 0;
  return rc;
}

/*
** Create a new cursor for the BTree whose root is on the page
** iTable.  The act of acquiring a cursor gets a read lock on 
** the database file.
**
** If wrFlag==0, then the cursor can only be used for reading.
** If wrFlag==1, then the cursor can be used for reading or for
** writing if other conditions for writing are also met.  These
** are the conditions that must be met in order for writing to
** be allowed:
**
** 1:  The cursor must have been opened with wrFlag==1
**
** 2:  No other cursors may be open with wrFlag==0 on the same table
**
** 3:  The database must be writable (not on read-only media)
**
** 4:  There must be an active transaction.
**
** Condition 2 warrants further discussion.  If any cursor is opened
** on a table with wrFlag==0, that prevents all other cursors from
** writing to that table.  This is a kind of "read-lock".  When a cursor
** is opened with wrFlag==0 it is guaranteed that the table will not
** change as long as the cursor is open.  This allows the cursor to
** do a sequential scan of the table without having to worry about
** entries being inserted or deleted during the scan.  Cursors should
** be opened with wrFlag==0 only if this read-lock property is needed.
** That is to say, cursors should be opened with wrFlag==0 only if they
** intend to use the sqliteBtreeNext() system call.  All other cursors
** should be opened with wrFlag==1 even if they never really intend
** to write.
** 
** No checking is done to make sure that page iTable really is the
** root page of a b-tree.  If it is not, then the cursor acquired
** will not work correctly.
*/
static 
int fileBtreeCursor(Btree *pBt, int iTable, int wrFlag, BtCursor **ppCur){
  int rc;
  BtCursor *pCur, *pRing;

  if( pBt->readOnly && wrFlag ){
    *ppCur = 0;
    return SQLITE_READONLY;
  }
  if( pBt->page1==0 ){
    rc = lockBtree(pBt);
    if( rc!=SQLITE_OK ){
      *ppCur = 0;
      return rc;
    }
  }
  pCur = sqliteMalloc( sizeof(*pCur) );
  if( pCur==0 ){
    rc = SQLITE_NOMEM;
    goto create_cursor_exception;
  }
  pCur->pgnoRoot = (Pgno)iTable;
  rc = sqlitepager_get(pBt->pPager, pCur->pgnoRoot, (void**)&pCur->pPage);
  if( rc!=SQLITE_OK ){
    goto create_cursor_exception;
  }
  rc = initPage(pBt, pCur->pPage, pCur->pgnoRoot, 0);
  if( rc!=SQLITE_OK ){
    goto create_cursor_exception;
  }
  pCur->pOps = &sqliteBtreeCursorOps;
  pCur->pBt = pBt;
  pCur->wrFlag = wrFlag;
  pCur->idx = 0;
  pCur->eSkip = SKIP_INVALID;
  pCur->pNext = pBt->pCursor;
  if( pCur->pNext ){
    pCur->pNext->pPrev = pCur;
  }
  pCur->pPrev = 0;
  pRing = pBt->pCursor;
  while( pRing && pRing->pgnoRoot!=pCur->pgnoRoot ){ pRing = pRing->pNext; }
  if( pRing ){
    pCur->pShared = pRing->pShared;
    pRing->pShared = pCur;
  }else{
    pCur->pShared = pCur;
  }
  pBt->pCursor = pCur;
  *ppCur = pCur;
  return SQLITE_OK;

create_cursor_exception:
  *ppCur = 0;
  if( pCur ){
    if( pCur->pPage ) sqlitepager_unref(pCur->pPage);
    sqliteFree(pCur);
  }
  unlockBtreeIfUnused(pBt);
  return rc;
}

/*
** Close a cursor.  The read lock on the database file is released
** when the last cursor is closed.
*/
static int fileBtreeCloseCursor(BtCursor *pCur){
  Btree *pBt = pCur->pBt;
  if( pCur->pPrev ){
    pCur->pPrev->pNext = pCur->pNext;
  }else{
    pBt->pCursor = pCur->pNext;
  }
  if( pCur->pNext ){
    pCur->pNext->pPrev = pCur->pPrev;
  }
  if( pCur->pPage ){
    sqlitepager_unref(pCur->pPage);
  }
  if( pCur->pShared!=pCur ){
    BtCursor *pRing = pCur->pShared;
    while( pRing->pShared!=pCur ){ pRing = pRing->pShared; }
    pRing->pShared = pCur->pShared;
  }
  unlockBtreeIfUnused(pBt);
  sqliteFree(pCur);
  return SQLITE_OK;
}

/*
** Make a temporary cursor by filling in the fields of pTempCur.
** The temporary cursor is not on the cursor list for the Btree.
*/
static void getTempCursor(BtCursor *pCur, BtCursor *pTempCur){
  memcpy(pTempCur, pCur, sizeof(*pCur));
  pTempCur->pNext = 0;
  pTempCur->pPrev = 0;
  if( pTempCur->pPage ){
    sqlitepager_ref(pTempCur->pPage);
  }
}

/*
** Delete a temporary cursor such as was made by the CreateTemporaryCursor()
** function above.
*/
static void releaseTempCursor(BtCursor *pCur){
  if( pCur->pPage ){
    sqlitepager_unref(pCur->pPage);
  }
}

/*
** Set *pSize to the number of bytes of key in the entry the
** cursor currently points to.  Always return SQLITE_OK.
** Failure is not possible.  If the cursor is not currently
** pointing to an entry (which can happen, for example, if
** the database is empty) then *pSize is set to 0.
*/
static int fileBtreeKeySize(BtCursor *pCur, int *pSize){
  Cell *pCell;
  MemPage *pPage;

  pPage = pCur->pPage;
  assert( pPage!=0 );
  if( pCur->idx >= pPage->nCell ){
    *pSize = 0;
  }else{
    pCell = pPage->apCell[pCur->idx];
    *pSize = NKEY(pCur->pBt, pCell->h);
  }
  return SQLITE_OK;
}

/*
** Read payload information from the entry that the pCur cursor is
** pointing to.  Begin reading the payload at "offset" and read
** a total of "amt" bytes.  Put the result in zBuf.
**
** This routine does not make a distinction between key and data.
** It just reads bytes from the payload area.
*/
static int getPayload(BtCursor *pCur, int offset, int amt, char *zBuf){
  char *aPayload;
  Pgno nextPage;
  int rc;
  Btree *pBt = pCur->pBt;
  assert( pCur!=0 && pCur->pPage!=0 );
  assert( pCur->idx>=0 && pCur->idx<pCur->pPage->nCell );
  aPayload = pCur->pPage->apCell[pCur->idx]->aPayload;
  if( offset<MX_LOCAL_PAYLOAD ){
    int a = amt;
    if( a+offset>MX_LOCAL_PAYLOAD ){
      a = MX_LOCAL_PAYLOAD - offset;
    }
    memcpy(zBuf, &aPayload[offset], a);
    if( a==amt ){
      return SQLITE_OK;
    }
    offset = 0;
    zBuf += a;
    amt -= a;
  }else{
    offset -= MX_LOCAL_PAYLOAD;
  }
  if( amt>0 ){
    nextPage = SWAB32(pBt, pCur->pPage->apCell[pCur->idx]->ovfl);
  }
  while( amt>0 && nextPage ){
    OverflowPage *pOvfl;
    rc = sqlitepager_get(pBt->pPager, nextPage, (void**)&pOvfl);
    if( rc!=0 ){
      return rc;
    }
    nextPage = SWAB32(pBt, pOvfl->iNext);
    if( offset<OVERFLOW_SIZE ){
      int a = amt;
      if( a + offset > OVERFLOW_SIZE ){
        a = OVERFLOW_SIZE - offset;
      }
      memcpy(zBuf, &pOvfl->aPayload[offset], a);
      offset = 0;
      amt -= a;
      zBuf += a;
    }else{
      offset -= OVERFLOW_SIZE;
    }
    sqlitepager_unref(pOvfl);
  }
  if( amt>0 ){
    return SQLITE_CORRUPT;
  }
  return SQLITE_OK;
}

/*
** Read part of the key associated with cursor pCur.  A maximum
** of "amt" bytes will be transfered into zBuf[].  The transfer
** begins at "offset".  The number of bytes actually read is
** returned. 
**
** Change:  It used to be that the amount returned will be smaller
** than the amount requested if there are not enough bytes in the key
** to satisfy the request.  But now, it must be the case that there
** is enough data available to satisfy the request.  If not, an exception
** is raised.  The change was made in an effort to boost performance
** by eliminating unneeded tests.
*/
static int fileBtreeKey(BtCursor *pCur, int offset, int amt, char *zBuf){
  MemPage *pPage;

  assert( amt>=0 );
  assert( offset>=0 );
  assert( pCur->pPage!=0 );
  pPage = pCur->pPage;
  if( pCur->idx >= pPage->nCell ){
    return 0;
  }
  assert( amt+offset <= NKEY(pCur->pBt, pPage->apCell[pCur->idx]->h) );
  getPayload(pCur, offset, amt, zBuf);
  return amt;
}

/*
** Set *pSize to the number of bytes of data in the entry the
** cursor currently points to.  Always return SQLITE_OK.
** Failure is not possible.  If the cursor is not currently
** pointing to an entry (which can happen, for example, if
** the database is empty) then *pSize is set to 0.
*/
static int fileBtreeDataSize(BtCursor *pCur, int *pSize){
  Cell *pCell;
  MemPage *pPage;

  pPage = pCur->pPage;
  assert( pPage!=0 );
  if( pCur->idx >= pPage->nCell ){
    *pSize = 0;
  }else{
    pCell = pPage->apCell[pCur->idx];
    *pSize = NDATA(pCur->pBt, pCell->h);
  }
  return SQLITE_OK;
}

/*
** Read part of the data associated with cursor pCur.  A maximum
** of "amt" bytes will be transfered into zBuf[].  The transfer
** begins at "offset".  The number of bytes actually read is
** returned.  The amount returned will be smaller than the
** amount requested if there are not enough bytes in the data
** to satisfy the request.
*/
static int fileBtreeData(BtCursor *pCur, int offset, int amt, char *zBuf){
  Cell *pCell;
  MemPage *pPage;

  assert( amt>=0 );
  assert( offset>=0 );
  assert( pCur->pPage!=0 );
  pPage = pCur->pPage;
  if( pCur->idx >= pPage->nCell ){
    return 0;
  }
  pCell = pPage->apCell[pCur->idx];
  assert( amt+offset <= NDATA(pCur->pBt, pCell->h) );
  getPayload(pCur, offset + NKEY(pCur->pBt, pCell->h), amt, zBuf);
  return amt;
}

/*
** Compare an external key against the key on the entry that pCur points to.
**
** The external key is pKey and is nKey bytes long.  The last nIgnore bytes
** of the key associated with pCur are ignored, as if they do not exist.
** (The normal case is for nIgnore to be zero in which case the entire
** internal key is used in the comparison.)
**
** The comparison result is written to *pRes as follows:
**
**    *pRes<0    This means pCur<pKey
**
**    *pRes==0   This means pCur==pKey for all nKey bytes
**
**    *pRes>0    This means pCur>pKey
**
** When one key is an exact prefix of the other, the shorter key is
** considered less than the longer one.  In order to be equal the
** keys must be exactly the same length. (The length of the pCur key
** is the actual key length minus nIgnore bytes.)
*/
static int fileBtreeKeyCompare(
  BtCursor *pCur,       /* Pointer to entry to compare against */
  const void *pKey,     /* Key to compare against entry that pCur points to */
  int nKey,             /* Number of bytes in pKey */
  int nIgnore,          /* Ignore this many bytes at the end of pCur */
  int *pResult          /* Write the result here */
){
  Pgno nextPage;
  int n, c, rc, nLocal;
  Cell *pCell;
  Btree *pBt = pCur->pBt;
  const char *zKey  = (const char*)pKey;

  assert( pCur->pPage );
  assert( pCur->idx>=0 && pCur->idx<pCur->pPage->nCell );
  pCell = pCur->pPage->apCell[pCur->idx];
  nLocal = NKEY(pBt, pCell->h) - nIgnore;
  if( nLocal<0 ) nLocal = 0;
  n = nKey<nLocal ? nKey : nLocal;
  if( n>MX_LOCAL_PAYLOAD ){
    n = MX_LOCAL_PAYLOAD;
  }
  c = memcmp(pCell->aPayload, zKey, n);
  if( c!=0 ){
    *pResult = c;
    return SQLITE_OK;
  }
  zKey += n;
  nKey -= n;
  nLocal -= n;
  nextPage = SWAB32(pBt, pCell->ovfl);
  while( nKey>0 && nLocal>0 ){
    OverflowPage *pOvfl;
    if( nextPage==0 ){
      return SQLITE_CORRUPT;
    }
    rc = sqlitepager_get(pBt->pPager, nextPage, (void**)&pOvfl);
    if( rc ){
      return rc;
    }
    nextPage = SWAB32(pBt, pOvfl->iNext);
    n = nKey<nLocal ? nKey : nLocal;
    if( n>OVERFLOW_SIZE ){
      n = OVERFLOW_SIZE;
    }
    c = memcmp(pOvfl->aPayload, zKey, n);
    sqlitepager_unref(pOvfl);
    if( c!=0 ){
      *pResult = c;
      return SQLITE_OK;
    }
    nKey -= n;
    nLocal -= n;
    zKey += n;
  }
  if( c==0 ){
    c = nLocal - nKey;
  }
  *pResult = c;
  return SQLITE_OK;
}

/*
** Move the cursor down to a new child page.  The newPgno argument is the
** page number of the child page in the byte order of the disk image.
*/
static int moveToChild(BtCursor *pCur, int newPgno){
  int rc;
  MemPage *pNewPage;
  Btree *pBt = pCur->pBt;

  newPgno = SWAB32(pBt, newPgno);
  rc = sqlitepager_get(pBt->pPager, newPgno, (void**)&pNewPage);
  if( rc ) return rc;
  rc = initPage(pBt, pNewPage, newPgno, pCur->pPage);
  if( rc ) return rc;
  assert( pCur->idx>=pCur->pPage->nCell
          || pCur->pPage->apCell[pCur->idx]->h.leftChild==SWAB32(pBt,newPgno) );
  assert( pCur->idx<pCur->pPage->nCell
          || pCur->pPage->u.hdr.rightChild==SWAB32(pBt,newPgno) );
  pNewPage->idxParent = pCur->idx;
  pCur->pPage->idxShift = 0;
  sqlitepager_unref(pCur->pPage);
  pCur->pPage = pNewPage;
  pCur->idx = 0;
  if( pNewPage->nCell<1 ){
    return SQLITE_CORRUPT;
  }
  return SQLITE_OK;
}

/*
** Move the cursor up to the parent page.
**
** pCur->idx is set to the cell index that contains the pointer
** to the page we are coming from.  If we are coming from the
** right-most child page then pCur->idx is set to one more than
** the largest cell index.
*/
static void moveToParent(BtCursor *pCur){
  Pgno oldPgno;
  MemPage *pParent;
  MemPage *pPage;
  int idxParent;
  pPage = pCur->pPage;
  assert( pPage!=0 );
  pParent = pPage->pParent;
  assert( pParent!=0 );
  idxParent = pPage->idxParent;
  sqlitepager_ref(pParent);
  sqlitepager_unref(pPage);
  pCur->pPage = pParent;
  assert( pParent->idxShift==0 );
  if( pParent->idxShift==0 ){
    pCur->idx = idxParent;
#ifndef NDEBUG  
    /* Verify that pCur->idx is the correct index to point back to the child
    ** page we just came from 
    */
    oldPgno = SWAB32(pCur->pBt, sqlitepager_pagenumber(pPage));
    if( pCur->idx<pParent->nCell ){
      assert( pParent->apCell[idxParent]->h.leftChild==oldPgno );
    }else{
      assert( pParent->u.hdr.rightChild==oldPgno );
    }
#endif
  }else{
    /* The MemPage.idxShift flag indicates that cell indices might have 
    ** changed since idxParent was set and hence idxParent might be out
    ** of date.  So recompute the parent cell index by scanning all cells
    ** and locating the one that points to the child we just came from.
    */
    int i;
    pCur->idx = pParent->nCell;
    oldPgno = SWAB32(pCur->pBt, sqlitepager_pagenumber(pPage));
    for(i=0; i<pParent->nCell; i++){
      if( pParent->apCell[i]->h.leftChild==oldPgno ){
        pCur->idx = i;
        break;
      }
    }
  }
}

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

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

/*
** Move the cursor down to the left-most leaf entry beneath the
** entry to which it is currently pointing.
*/
static int moveToLeftmost(BtCursor *pCur){
  Pgno pgno;
  int rc;

  while( (pgno = pCur->pPage->apCell[pCur->idx]->h.leftChild)!=0 ){
    rc = moveToChild(pCur, pgno);
    if( rc ) return rc;
  }
  return SQLITE_OK;
}

/*
** Move the cursor down to the right-most leaf entry beneath the
** page to which it is currently pointing.  Notice the difference
** between moveToLeftmost() and moveToRightmost().  moveToLeftmost()
** finds the left-most entry beneath the *entry* whereas moveToRightmost()
** finds the right-most entry beneath the *page*.
*/
static int moveToRightmost(BtCursor *pCur){
  Pgno pgno;
  int rc;

  while( (pgno = pCur->pPage->u.hdr.rightChild)!=0 ){
    pCur->idx = pCur->pPage->nCell;
    rc = moveToChild(pCur, pgno);
    if( rc ) return rc;
  }
  pCur->idx = pCur->pPage->nCell - 1;
  return SQLITE_OK;
}

/* Move the cursor to the first entry in the table.  Return SQLITE_OK
** on success.  Set *pRes to 0 if the cursor actually points to something
** or set *pRes to 1 if the table is empty.
*/
static int fileBtreeFirst(BtCursor *pCur, int *pRes){
  int rc;
  if( pCur->pPage==0 ) return SQLITE_ABORT;
  rc = moveToRoot(pCur);
  if( rc ) return rc;
  if( pCur->pPage->nCell==0 ){
    *pRes = 1;
    return SQLITE_OK;
  }
  *pRes = 0;
  rc = moveToLeftmost(pCur);
  pCur->eSkip = SKIP_NONE;
  return rc;
}

/* Move the cursor to the last entry in the table.  Return SQLITE_OK
** on success.  Set *pRes to 0 if the cursor actually points to something
** or set *pRes to 1 if the table is empty.
*/
static int fileBtreeLast(BtCursor *pCur, int *pRes){
  int rc;
  if( pCur->pPage==0 ) return SQLITE_ABORT;
  rc = moveToRoot(pCur);
  if( rc ) return rc;
  assert( pCur->pPage->isInit );
  if( pCur->pPage->nCell==0 ){
    *pRes = 1;
    return SQLITE_OK;
  }
  *pRes = 0;
  rc = moveToRightmost(pCur);
  pCur->eSkip = SKIP_NONE;
  return rc;
}

/* Move the cursor so that it points to an entry near pKey.
** Return a success code.
**
** If an exact match is not found, then the cursor is always
** left pointing at a leaf page which would hold the entry if it
** were present.  The cursor might point to an entry that comes
** before or after the key.
**
** The result of comparing the key with the entry to which the
** cursor is left pointing is stored in pCur->iMatch.  The same
** value is also written to *pRes if pRes!=NULL.  The meaning of
** this value is as follows:
**
**     *pRes<0      The cursor is left pointing at an entry that
**                  is smaller than pKey or if the table is empty
**                  and the cursor is therefore left point to nothing.
**
**     *pRes==0     The cursor is left pointing at an entry that
**                  exactly matches pKey.
**
**     *pRes>0      The cursor is left pointing at an entry that
**                  is larger than pKey.
*/
static
int fileBtreeMoveto(BtCursor *pCur, const void *pKey, int nKey, int *pRes){
  int rc;
  if( pCur->pPage==0 ) return SQLITE_ABORT;
  pCur->eSkip = SKIP_NONE;
  rc = moveToRoot(pCur);
  if( rc ) return rc;
  for(;;){
    int lwr, upr;
    Pgno chldPg;
    MemPage *pPage = pCur->pPage;
    int c = -1;  /* pRes return if table is empty must be -1 */
    lwr = 0;
    upr = pPage->nCell-1;
    while( lwr<=upr ){
      pCur->idx = (lwr+upr)/2;
      rc = fileBtreeKeyCompare(pCur, pKey, nKey, 0, &c);
      if( rc ) return rc;
      if( c==0 ){
        pCur->iMatch = c;
        if( pRes ) *pRes = 0;
        return SQLITE_OK;
      }
      if( c<0 ){
        lwr = pCur->idx+1;
      }else{
        upr = pCur->idx-1;
      }
    }
    assert( lwr==upr+1 );
    assert( pPage->isInit );
    if( lwr>=pPage->nCell ){
      chldPg = pPage->u.hdr.rightChild;
    }else{
      chldPg = pPage->apCell[lwr]->h.leftChild;
    }
    if( chldPg==0 ){
      pCur->iMatch = c;
      if( pRes ) *pRes = c;
      return SQLITE_OK;
    }
    pCur->idx = lwr;
    rc = moveToChild(pCur, chldPg);
    if( rc ) return rc;
  }
  /* NOT REACHED */
}

/*
** Advance the cursor to the next entry in the database.  If
** successful then set *pRes=0.  If the cursor
** was already pointing to the last entry in the database before
** this routine was called, then set *pRes=1.
*/
static int fileBtreeNext(BtCursor *pCur, int *pRes){
  int rc;
  MemPage *pPage = pCur->pPage;
  assert( pRes!=0 );
  if( pPage==0 ){
    *pRes = 1;
    return SQLITE_ABORT;
  }
  assert( pPage->isInit );
  assert( pCur->eSkip!=SKIP_INVALID );
  if( pPage->nCell==0 ){
    *pRes = 1;
    return SQLITE_OK;
  }
  assert( pCur->idx<pPage->nCell );
  if( pCur->eSkip==SKIP_NEXT ){
    pCur->eSkip = SKIP_NONE;
    *pRes = 0;
    return SQLITE_OK;
  }
  pCur->eSkip = SKIP_NONE;
  pCur->idx++;
  if( pCur->idx>=pPage->nCell ){
    if( pPage->u.hdr.rightChild ){
      rc = moveToChild(pCur, pPage->u.hdr.rightChild);
      if( rc ) return rc;
      rc = moveToLeftmost(pCur);
      *pRes = 0;
      return rc;
    }
    do{
      if( pPage->pParent==0 ){
        *pRes = 1;
        return SQLITE_OK;
      }
      moveToParent(pCur);
      pPage = pCur->pPage;
    }while( pCur->idx>=pPage->nCell );
    *pRes = 0;
    return SQLITE_OK;
  }
  *pRes = 0;
  if( pPage->u.hdr.rightChild==0 ){
    return SQLITE_OK;
  }
  rc = moveToLeftmost(pCur);
  return rc;
}

/*
** Step the cursor to the back to the previous entry in the database.  If
** successful then set *pRes=0.  If the cursor
** was already pointing to the first entry in the database before
** this routine was called, then set *pRes=1.
*/
static int fileBtreePrevious(BtCursor *pCur, int *pRes){
  int rc;
  Pgno pgno;
  MemPage *pPage;
  pPage = pCur->pPage;
  if( pPage==0 ){
    *pRes = 1;
    return SQLITE_ABORT;
  }
  assert( pPage->isInit );
  assert( pCur->eSkip!=SKIP_INVALID );
  if( pPage->nCell==0 ){
    *pRes = 1;
    return SQLITE_OK;
  }
  if( pCur->eSkip==SKIP_PREV ){
    pCur->eSkip = SKIP_NONE;
    *pRes = 0;
    return SQLITE_OK;
  }
  pCur->eSkip = SKIP_NONE;
  assert( pCur->idx>=0 );
  if( (pgno = pPage->apCell[pCur->idx]->h.leftChild)!=0 ){
    rc = moveToChild(pCur, pgno);
    if( rc ) return rc;
    rc = moveToRightmost(pCur);
  }else{
    while( pCur->idx==0 ){
      if( pPage->pParent==0 ){
        if( pRes ) *pRes = 1;
        return SQLITE_OK;
      }
      moveToParent(pCur);
      pPage = pCur->pPage;
    }
    pCur->idx--;
    rc = SQLITE_OK;
  }
  *pRes = 0;
  return rc;
}

/*
** Allocate a new page from the database file.
**
** The new page is marked as dirty.  (In other words, sqlitepager_write()
** has already been called on the new page.)  The new page has also
** been referenced and the calling routine is responsible for calling
** sqlitepager_unref() on the new page when it is done.
**
** SQLITE_OK is returned on success.  Any other return value indicates
** an error.  *ppPage and *pPgno are undefined in the event of an error.
** Do not invoke sqlitepager_unref() on *ppPage if an error is returned.
**
** If the "nearby" parameter is not 0, then a (feeble) effort is made to 
** locate a page close to the page number "nearby".  This can be used in an
** attempt to keep related pages close to each other in the database file,

btree.c  view on Meta::CPAN

** NN neighbors on either side.
**
** The minimum value of NN is 1 (of course).  Increasing NN above 1
** (to 2 or 3) gives a modest improvement in SELECT and DELETE performance
** in exchange for a larger degradation in INSERT and UPDATE performance.
** The value of NN appears to give the best results overall.
*/
#define NN 1             /* Number of neighbors on either side of pPage */
#define NB (NN*2+1)      /* Total pages involved in the balance */

/*
** This routine redistributes Cells on pPage and up to two siblings
** of pPage so that all pages have about the same amount of free space.
** Usually one sibling on either side of pPage is used in the balancing,
** though both siblings might come from one side if pPage is the first
** or last child of its parent.  If pPage has fewer than two siblings
** (something which can only happen if pPage is the root page or a 
** child of root) then all available siblings participate in the balancing.
**
** The number of siblings of pPage might be increased or decreased by
** one in an effort to keep pages between 66% and 100% full. The root page
** is special and is allowed to be less than 66% full. If pPage is 
** the root page, then the depth of the tree might be increased
** or decreased by one, as necessary, to keep the root page from being
** overfull or empty.
**
** This routine calls relinkCellList() on its input page regardless of
** whether or not it does any real balancing.  Client routines will typically
** invoke insertCell() or dropCell() before calling this routine, so we
** need to call relinkCellList() to clean up the mess that those other
** routines left behind.
**
** pCur is left pointing to the same cell as when this routine was called
** even if that cell gets moved to a different page.  pCur may be NULL.
** Set the pCur parameter to NULL if you do not care about keeping track
** of a cell as that will save this routine the work of keeping track of it.
**
** Note that when this routine is called, some of the Cells on pPage
** might not actually be stored in pPage->u.aDisk[].  This can happen
** if the page is overfull.  Part of the job of this routine is to
** make sure all Cells for pPage once again fit in pPage->u.aDisk[].
**
** In the course of balancing the siblings of pPage, the parent of pPage
** might become overfull or underfull.  If that happens, then this routine
** is called recursively on the parent.
**
** If this routine fails for any reason, it might leave the database
** in a corrupted state.  So if this routine fails, the database should
** be rolled back.
*/
static int balance(Btree *pBt, MemPage *pPage, BtCursor *pCur){
  MemPage *pParent;            /* The parent of pPage */
  int nCell;                   /* Number of cells in apCell[] */
  int nOld;                    /* Number of pages in apOld[] */
  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 */
  MemPage *pOldCurPage;        /* The cursor originally points to this page */
  int subtotal;                /* Subtotal of bytes in cells on one page */
  MemPage *extraUnref = 0;     /* A page that needs to be unref-ed */
  MemPage *apOld[NB];          /* pPage and up to two siblings */
  Pgno pgnoOld[NB];            /* Page numbers for each page in apOld[] */
  MemPage *apNew[NB+1];        /* pPage and up to NB siblings after balancing */
  Pgno pgnoNew[NB+1];          /* Page numbers for each page in apNew[] */
  int idxDiv[NB];              /* Indices of divider cells in pParent */
  Cell *apDiv[NB];             /* Divider cells in pParent */
  Cell aTemp[NB];              /* Temporary holding area for apDiv[] */
  int cntNew[NB+1];            /* Index in apCell[] of cell after i-th page */
  int szNew[NB+1];             /* Combined size of cells place on i-th page */
  MemPage aOld[NB];            /* Temporary copies of pPage and its siblings */
  Cell *apCell[(MX_CELL+2)*NB]; /* All cells from pages being balanced */
  int szCell[(MX_CELL+2)*NB];  /* Local size of all cells */

  /* 
  ** Return without doing any work if pPage is neither overfull nor
  ** underfull.
  */
  assert( sqlitepager_iswriteable(pPage) );
  if( !pPage->isOverfull && pPage->nFree<SQLITE_USABLE_SIZE/2 
        && pPage->nCell>=2){
    relinkCellList(pBt, 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;
    assert( pPage->isInit );
    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 = SWAB32(pBt, pPage->u.hdr.rightChild);
        rc = sqlitepager_get(pBt->pPager, pgnoChild, (void**)&pChild);
        if( rc ) return rc;
        memcpy(pPage, pChild, SQLITE_USABLE_SIZE);
        pPage->isInit = 0;
        rc = initPage(pBt, pPage, sqlitepager_pagenumber(pPage), 0);
        assert( rc==SQLITE_OK );
        reparentChildPages(pBt, pPage);
        if( pCur && pCur->pPage==pChild ){
          sqlitepager_unref(pChild);
          pCur->pPage = pPage;
          sqlitepager_ref(pPage);
        }
        freePage(pBt, pChild, pgnoChild);
        sqlitepager_unref(pChild);
      }else{
        relinkCellList(pBt, pPage);

btree.c  view on Meta::CPAN

  ** 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 
  */
  if( pParent->idxShift ){
    Pgno pgno, swabPgno;
    pgno = sqlitepager_pagenumber(pPage);
    swabPgno = SWAB32(pBt, pgno);
    for(idx=0; idx<pParent->nCell; idx++){
      if( pParent->apCell[idx]->h.leftChild==swabPgno ){
        break;
      }
    }
    assert( idx<pParent->nCell || pParent->u.hdr.rightChild==swabPgno );
  }else{
    idx = pPage->idxParent;
  }

  /*
  ** Initialize variables so that it will be safe to jump
  ** directly to balance_cleanup at any moment.
  */
  nOld = nNew = 0;
  sqlitepager_ref(pParent);

  /*
  ** Find sibling pages to pPage and the Cells in pParent that divide
  ** the siblings.  An attempt is made to find NN siblings on either
  ** side of pPage.  More siblings are taken from one side, however, if
  ** pPage there are fewer than NN siblings on the other side.  If pParent
  ** has NB or fewer children then all children of pParent are taken.
  */
  nxDiv = idx - NN;
  if( nxDiv + NB > pParent->nCell ){
    nxDiv = pParent->nCell - NB + 1;
  }
  if( nxDiv<0 ){
    nxDiv = 0;
  }
  nDiv = 0;
  for(i=0, k=nxDiv; i<NB; i++, k++){
    if( k<pParent->nCell ){
      idxDiv[i] = k;
      apDiv[i] = pParent->apCell[k];
      nDiv++;
      pgnoOld[i] = SWAB32(pBt, apDiv[i]->h.leftChild);
    }else if( k==pParent->nCell ){
      pgnoOld[i] = SWAB32(pBt, pParent->u.hdr.rightChild);
    }else{
      break;
    }
    rc = sqlitepager_get(pBt->pPager, pgnoOld[i], (void**)&apOld[i]);
    if( rc ) goto balance_cleanup;
    rc = initPage(pBt, apOld[i], pgnoOld[i], pParent);
    if( rc ) goto balance_cleanup;
    apOld[i]->idxParent = k;
    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
  ** cursor pointing at the same cell.  If pCur points to a page that
  ** has no involvement with this rebalancing, then set iCur to a large
  ** number so that the iCur==j tests always fail in the main cell
  ** distribution loop below.
  */
  if( pCur ){
    iCur = 0;
    for(i=0; i<nOld; i++){
      if( pCur->pPage==apOld[i] ){
        iCur += pCur->idx;
        break;
      }
      iCur += apOld[i]->nCell;
      if( i<nOld-1 && pCur->pPage==pParent && pCur->idx==idxDiv[i] ){
        break;
      }
      iCur++;
    }
    pOldCurPage = pCur->pPage;
  }

  /*
  ** Make copies of the content of pPage and its siblings into aOld[].
  ** The rest of this function will use data from the copies rather
  ** that the original pages since the original pages will be in the
  ** process of being overwritten.
  */
  for(i=0; i<nOld; i++){
    copyPage(&aOld[i], apOld[i]);
  }

  /*
  ** Load pointers to all cells on sibling pages and the divider cells
  ** into the local apCell[] array.  Make copies of the divider cells
  ** into aTemp[] and remove the the divider Cells from pParent.
  */
  nCell = 0;
  for(i=0; i<nOld; i++){
    MemPage *pOld = &aOld[i];
    for(j=0; j<pOld->nCell; j++){
      apCell[nCell] = pOld->apCell[j];
      szCell[nCell] = cellSize(pBt, apCell[nCell]);
      nCell++;
    }
    if( i<nOld-1 ){
      szCell[nCell] = cellSize(pBt, apDiv[i]);
      memcpy(&aTemp[i], apDiv[i], szCell[nCell]);
      apCell[nCell] = &aTemp[i];
      dropCell(pBt, pParent, nxDiv, szCell[nCell]);
      assert( SWAB32(pBt, 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.

btree.c  view on Meta::CPAN

      if( pCur && iCur==j ){ pCur->pPage = pParent; pCur->idx = nxDiv; }
      insertCell(pBt, pParent, nxDiv, apCell[j], szCell[j]);
      j++;
      nxDiv++;
    }
  }
  assert( j==nCell );
  apNew[nNew-1]->u.hdr.rightChild = aOld[nOld-1].u.hdr.rightChild;
  if( nxDiv==pParent->nCell ){
    pParent->u.hdr.rightChild = SWAB32(pBt, pgnoNew[nNew-1]);
  }else{
    pParent->apCell[nxDiv]->h.leftChild = SWAB32(pBt, pgnoNew[nNew-1]);
  }
  if( pCur ){
    if( j<=iCur && pCur->pPage==pParent && pCur->idx>idxDiv[nOld-1] ){
      assert( pCur->pPage==pOldCurPage );
      pCur->idx += nNew - nOld;
    }else{
      assert( pOldCurPage!=0 );
      sqlitepager_ref(pCur->pPage);
      sqlitepager_unref(pOldCurPage);
    }
  }

  /*
  ** Reparent children of all cells.
  */
  for(i=0; i<nNew; i++){
    reparentChildPages(pBt, apNew[i]);
  }
  reparentChildPages(pBt, pParent);

  /*
  ** balance the parent page.
  */
  rc = balance(pBt, pParent, pCur);

  /*
  ** Cleanup before returning.
  */
balance_cleanup:
  if( extraUnref ){
    sqlitepager_unref(extraUnref);
  }
  for(i=0; i<nOld; i++){
    if( apOld[i]!=0 && apOld[i]!=&aOld[i] ) sqlitepager_unref(apOld[i]);
  }
  for(i=0; i<nNew; i++){
    sqlitepager_unref(apNew[i]);
  }
  if( pCur && pCur->pPage==0 ){
    pCur->pPage = pParent;
    pCur->idx = 0;
  }else{
    sqlitepager_unref(pParent);
  }
  return rc;
}

/*
** This routine checks all cursors that point to the same table
** as pCur points to.  If any of those cursors were opened with
** wrFlag==0 then this routine returns SQLITE_LOCKED.  If all
** cursors point to the same table were opened with wrFlag==1
** then this routine returns SQLITE_OK.
**
** In addition to checking for read-locks (where a read-lock 
** means a cursor opened with wrFlag==0) this routine also moves
** all cursors other than pCur so that they are pointing to the 
** first Cell on root page.  This is necessary because an insert 
** or delete might change the number of cells on a page or delete
** a page entirely and we do not want to leave any cursors 
** pointing to non-existant pages or cells.
*/
static int checkReadLocks(BtCursor *pCur){
  BtCursor *p;
  assert( pCur->wrFlag );
  for(p=pCur->pShared; p!=pCur; p=p->pShared){
    assert( p );
    assert( p->pgnoRoot==pCur->pgnoRoot );
    if( p->wrFlag==0 ) return SQLITE_LOCKED;
    if( sqlitepager_pagenumber(p->pPage)!=p->pgnoRoot ){
      moveToRoot(p);
    }
  }
  return SQLITE_OK;
}

/*
** Insert a new record into the BTree.  The key is given by (pKey,nKey)
** and the data is given by (pData,nData).  The cursor is used only to
** define what database the record should be inserted into.  The cursor
** is left pointing at the new record.
*/
static int fileBtreeInsert(
  BtCursor *pCur,                /* Insert data into the table of this cursor */
  const void *pKey, int nKey,    /* The key of the new record */
  const void *pData, int nData   /* The data of the new record */
){
  Cell newCell;
  int rc;
  int loc;
  int szNew;
  MemPage *pPage;
  Btree *pBt = pCur->pBt;

  if( pCur->pPage==0 ){
    return SQLITE_ABORT;  /* A rollback destroyed this cursor */
  }
  if( !pBt->inTrans || nKey+nData==0 ){
    /* Must start a transaction before doing an insert */
    return pBt->readOnly ? SQLITE_READONLY : SQLITE_ERROR;
  }
  assert( !pBt->readOnly );
  if( !pCur->wrFlag ){
    return SQLITE_PERM;   /* Cursor not open for writing */
  }
  if( checkReadLocks(pCur) ){
    return SQLITE_LOCKED; /* The table pCur points to has a read lock */
  }
  rc = fileBtreeMoveto(pCur, pKey, nKey, &loc);
  if( rc ) return rc;
  pPage = pCur->pPage;
  assert( pPage->isInit );
  rc = sqlitepager_write(pPage);
  if( rc ) return rc;
  rc = fillInCell(pBt, &newCell, pKey, nKey, pData, nData);
  if( rc ) return rc;
  szNew = cellSize(pBt, &newCell);
  if( loc==0 ){
    newCell.h.leftChild = pPage->apCell[pCur->idx]->h.leftChild;
    rc = clearCell(pBt, pPage->apCell[pCur->idx]);
    if( rc ) return rc;
    dropCell(pBt, pPage, pCur->idx, cellSize(pBt, pPage->apCell[pCur->idx]));
  }else if( loc<0 && pPage->nCell>0 ){
    assert( pPage->u.hdr.rightChild==0 );  /* Must be a leaf page */
    pCur->idx++;
  }else{
    assert( pPage->u.hdr.rightChild==0 );  /* Must be a leaf page */
  }
  insertCell(pBt, pPage, pCur->idx, &newCell, szNew);
  rc = balance(pCur->pBt, pPage, pCur);
  /* sqliteBtreePageDump(pCur->pBt, pCur->pgnoRoot, 1); */
  /* fflush(stdout); */
  pCur->eSkip = SKIP_INVALID;
  return rc;
}

/*
** Delete the entry that the cursor is pointing to.
**
** The cursor is left pointing at either the next or the previous
** entry.  If the cursor is left pointing to the next entry, then 
** the pCur->eSkip flag is set to SKIP_NEXT which forces the next call to 
** sqliteBtreeNext() to be a no-op.  That way, you can always call
** sqliteBtreeNext() after a delete and the cursor will be left
** pointing to the first entry after the deleted entry.  Similarly,
** pCur->eSkip is set to SKIP_PREV is the cursor is left pointing to
** the entry prior to the deleted entry so that a subsequent call to
** sqliteBtreePrevious() will always leave the cursor pointing at the
** entry immediately before the one that was deleted.
*/
static int fileBtreeDelete(BtCursor *pCur){
  MemPage *pPage = pCur->pPage;
  Cell *pCell;
  int rc;
  Pgno pgnoChild;
  Btree *pBt = pCur->pBt;

  assert( pPage->isInit );
  if( pCur->pPage==0 ){
    return SQLITE_ABORT;  /* A rollback destroyed this cursor */
  }
  if( !pBt->inTrans ){
    /* Must start a transaction before doing a delete */
    return pBt->readOnly ? SQLITE_READONLY : SQLITE_ERROR;
  }
  assert( !pBt->readOnly );
  if( pCur->idx >= pPage->nCell ){
    return SQLITE_ERROR;  /* The cursor is not pointing to anything */
  }
  if( !pCur->wrFlag ){
    return SQLITE_PERM;   /* Did not open this cursor for writing */
  }
  if( checkReadLocks(pCur) ){
    return SQLITE_LOCKED; /* The table pCur points to has a read lock */
  }
  rc = sqlitepager_write(pPage);
  if( rc ) return rc;
  pCell = pPage->apCell[pCur->idx];
  pgnoChild = SWAB32(pBt, pCell->h.leftChild);
  clearCell(pBt, pCell);
  if( pgnoChild ){
    /*
    ** The entry we are about to delete is not a leaf so if we do not
    ** do something we will leave a hole on an internal page.
    ** We have to fill the hole by moving in a cell from a leaf.  The
    ** next Cell after the one to be deleted is guaranteed to exist and
    ** to be a leaf so we can use it.
    */
    BtCursor leafCur;
    Cell *pNext;
    int szNext;
    int notUsed;
    getTempCursor(pCur, &leafCur);
    rc = fileBtreeNext(&leafCur, &notUsed);
    if( rc!=SQLITE_OK ){
      if( rc!=SQLITE_NOMEM ) rc = SQLITE_CORRUPT;
      return rc;
    }
    rc = sqlitepager_write(leafCur.pPage);
    if( rc ) return rc;
    dropCell(pBt, pPage, pCur->idx, cellSize(pBt, pCell));
    pNext = leafCur.pPage->apCell[leafCur.idx];
    szNext = cellSize(pBt, pNext);
    pNext->h.leftChild = SWAB32(pBt, pgnoChild);
    insertCell(pBt, pPage, pCur->idx, pNext, szNext);
    rc = balance(pBt, pPage, pCur);
    if( rc ) return rc;
    pCur->eSkip = SKIP_NEXT;
    dropCell(pBt, leafCur.pPage, leafCur.idx, szNext);
    rc = balance(pBt, leafCur.pPage, pCur);
    releaseTempCursor(&leafCur);
  }else{
    dropCell(pBt, pPage, pCur->idx, cellSize(pBt, pCell));
    if( pCur->idx>=pPage->nCell ){
      pCur->idx = pPage->nCell-1;
      if( pCur->idx<0 ){ 
        pCur->idx = 0;
        pCur->eSkip = SKIP_NEXT;
      }else{
        pCur->eSkip = SKIP_PREV;
      }
    }else{
      pCur->eSkip = SKIP_NEXT;
    }
    rc = balance(pBt, pPage, pCur);
  }
  return rc;
}

/*
** Create a new BTree table.  Write into *piTable the page

btree.c  view on Meta::CPAN

  idx = SWAB16(pBt, pPage->u.hdr.firstCell);
  while( idx>0 ){
    pCell = (Cell*)&pPage->u.aDisk[idx];
    idx = SWAB16(pBt, pCell->h.iNext);
    if( pCell->h.leftChild ){
      rc = clearDatabasePage(pBt, SWAB32(pBt, pCell->h.leftChild), 1);
      if( rc ) return rc;
    }
    rc = clearCell(pBt, pCell);
    if( rc ) return rc;
  }
  if( pPage->u.hdr.rightChild ){
    rc = clearDatabasePage(pBt, SWAB32(pBt, pPage->u.hdr.rightChild), 1);
    if( rc ) return rc;
  }
  if( freePageFlag ){
    rc = freePage(pBt, pPage, pgno);
  }else{
    zeroPage(pBt, pPage);
  }
  sqlitepager_unref(pPage);
  return rc;
}

/*
** Delete all information from a single table in the database.
*/
static int fileBtreeClearTable(Btree *pBt, int iTable){
  int rc;
  BtCursor *pCur;
  if( !pBt->inTrans ){
    return pBt->readOnly ? SQLITE_READONLY : SQLITE_ERROR;
  }
  for(pCur=pBt->pCursor; pCur; pCur=pCur->pNext){
    if( pCur->pgnoRoot==(Pgno)iTable ){
      if( pCur->wrFlag==0 ) return SQLITE_LOCKED;
      moveToRoot(pCur);
    }
  }
  rc = clearDatabasePage(pBt, (Pgno)iTable, 0);
  if( rc ){
    fileBtreeRollback(pBt);
  }
  return rc;
}

/*
** Erase all information in a table and add the root of the table to
** the freelist.  Except, the root of the principle table (the one on
** page 2) is never added to the freelist.
*/
static int fileBtreeDropTable(Btree *pBt, int iTable){
  int rc;
  MemPage *pPage;
  BtCursor *pCur;
  if( !pBt->inTrans ){
    return pBt->readOnly ? SQLITE_READONLY : SQLITE_ERROR;
  }
  for(pCur=pBt->pCursor; pCur; pCur=pCur->pNext){
    if( pCur->pgnoRoot==(Pgno)iTable ){
      return SQLITE_LOCKED;  /* Cannot drop a table that has a cursor */
    }
  }
  rc = sqlitepager_get(pBt->pPager, (Pgno)iTable, (void**)&pPage);
  if( rc ) return rc;
  rc = fileBtreeClearTable(pBt, iTable);
  if( rc ) return rc;
  if( iTable>2 ){
    rc = freePage(pBt, pPage, iTable);
  }else{
    zeroPage(pBt, pPage);
  }
  sqlitepager_unref(pPage);
  return rc;  
}

#if 0 /* UNTESTED */
/*
** Copy all cell data from one database file into another.
** pages back the freelist.
*/
static int copyCell(Btree *pBtFrom, BTree *pBtTo, Cell *pCell){
  Pager *pFromPager = pBtFrom->pPager;
  OverflowPage *pOvfl;
  Pgno ovfl, nextOvfl;
  Pgno *pPrev;
  int rc = SQLITE_OK;
  MemPage *pNew, *pPrevPg;
  Pgno new;

  if( NKEY(pBtTo, pCell->h) + NDATA(pBtTo, pCell->h) <= MX_LOCAL_PAYLOAD ){
    return SQLITE_OK;
  }
  pPrev = &pCell->ovfl;
  pPrevPg = 0;
  ovfl = SWAB32(pBtTo, pCell->ovfl);
  while( ovfl && rc==SQLITE_OK ){
    rc = sqlitepager_get(pFromPager, ovfl, (void**)&pOvfl);
    if( rc ) return rc;
    nextOvfl = SWAB32(pBtFrom, pOvfl->iNext);
    rc = allocatePage(pBtTo, &pNew, &new, 0);
    if( rc==SQLITE_OK ){
      rc = sqlitepager_write(pNew);
      if( rc==SQLITE_OK ){
        memcpy(pNew, pOvfl, SQLITE_USABLE_SIZE);
        *pPrev = SWAB32(pBtTo, new);
        if( pPrevPg ){
          sqlitepager_unref(pPrevPg);
        }
        pPrev = &pOvfl->iNext;
        pPrevPg = pNew;
      }
    }
    sqlitepager_unref(pOvfl);
    ovfl = nextOvfl;
  }
  if( pPrevPg ){
    sqlitepager_unref(pPrevPg);
  }
  return rc;
}

btree.c  view on Meta::CPAN

  idx = SWAB16(pBt, pPage->u.hdr.firstCell);
  while( idx>0 && idx<=SQLITE_USABLE_SIZE-MIN_CELL_SIZE ){
    Cell *pCell = (Cell*)&pPage->u.aDisk[idx];
    int sz = cellSize(pBt, pCell);
    sprintf(range,"%d..%d", idx, idx+sz-1);
    sz = NKEY(pBt, pCell->h) + NDATA(pBt, pCell->h);
    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, 
      NKEY(pBt, pCell->h), NDATA(pBt, pCell->h),
      payload
    );
    if( pPage->isInit && pPage->apCell[i]!=pCell ){
      printf("**** apCell[%d] does not match on prior entry ****\n", i);
    }
    i++;
    idx = SWAB16(pBt, pCell->h.iNext);
  }
  if( idx!=0 ){
    printf("ERROR: next cell index out of range: %d\n", idx);
  }
  printf("right_child: %d\n", SWAB32(pBt, pPage->u.hdr.rightChild));
  nFree = 0;
  i = 0;
  idx = SWAB16(pBt, pPage->u.hdr.firstFree);
  while( idx>0 && idx<SQLITE_USABLE_SIZE ){
    FreeBlk *p = (FreeBlk*)&pPage->u.aDisk[idx];
    sprintf(range,"%d..%d", idx, idx+p->iSize-1);
    nFree += SWAB16(pBt, p->iSize);
    printf("freeblock %2d: i=%-10s size=%-4d total=%d\n",
       i, range, SWAB16(pBt, p->iSize), nFree);
    idx = SWAB16(pBt, 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 = SWAB16(pBt, pPage->u.hdr.firstCell);
    while( idx>0 && idx<SQLITE_USABLE_SIZE-MIN_CELL_SIZE ){
      Cell *pCell = (Cell*)&pPage->u.aDisk[idx];
      fileBtreePageDump(pBt, SWAB32(pBt, pCell->h.leftChild), 1);
      idx = SWAB16(pBt, pCell->h.iNext);
    }
    fileBtreePageDump(pBt, SWAB32(pBt, pPage->u.hdr.rightChild), 1);
  }
  sqlitepager_unref(pPage);
  return SQLITE_OK;
}
#endif

#ifdef SQLITE_TEST
/*
** Fill aResult[] with information about the entry and page that the
** cursor is pointing to.
** 
**   aResult[0] =  The page number
**   aResult[1] =  The entry number
**   aResult[2] =  Total number of entries on this page
**   aResult[3] =  Size of this entry
**   aResult[4] =  Number of free bytes on this page
**   aResult[5] =  Number of free blocks on the page
**   aResult[6] =  Page number of the left child of this entry
**   aResult[7] =  Page number of the right child for the whole page
**
** This routine is used for testing and debugging only.
*/
static int fileBtreeCursorDump(BtCursor *pCur, int *aResult){
  int cnt, idx;
  MemPage *pPage = pCur->pPage;
  Btree *pBt = pCur->pBt;
  aResult[0] = sqlitepager_pagenumber(pPage);
  aResult[1] = pCur->idx;
  aResult[2] = pPage->nCell;
  if( pCur->idx>=0 && pCur->idx<pPage->nCell ){
    aResult[3] = cellSize(pBt, pPage->apCell[pCur->idx]);
    aResult[6] = SWAB32(pBt, pPage->apCell[pCur->idx]->h.leftChild);
  }else{
    aResult[3] = 0;
    aResult[6] = 0;
  }
  aResult[4] = pPage->nFree;
  cnt = 0;
  idx = SWAB16(pBt, pPage->u.hdr.firstFree);
  while( idx>0 && idx<SQLITE_USABLE_SIZE ){
    cnt++;
    idx = SWAB16(pBt, ((FreeBlk*)&pPage->u.aDisk[idx])->iNext);
  }
  aResult[5] = cnt;
  aResult[7] = SWAB32(pBt, pPage->u.hdr.rightChild);
  return SQLITE_OK;
}
#endif

/*
** Return the pager associated with a BTree.  This routine is used for
** testing and debugging only.
*/
static Pager *fileBtreePager(Btree *pBt){
  return pBt->pPager;
}

/*
** 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 {
  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 */
  char *zErrMsg; /* An error message.  NULL of no errors seen. */
};



( run in 0.499 second using v1.01-cache-2.11-cpan-13bb782fe5a )