SQLite

    ** The caller detects this and calls this function again. This is
    ** required as the version of page 1 currently in the page1 buffer
    ** may not be the latest version - there may be a newer one in the log
    ** file.
    */
    if( page1[19]==2 && pBt->doNotUseWAL==0 ){
      int isOpen = 0;
      rc = sqlite3PagerOpenLog(pBt->pPager, &isOpen);
      if( rc!=SQLITE_OK ){
        goto page1_init_failed;
      }else if( isOpen==0 ){
        releasePage(pPage1);
        return SQLITE_OK;
      }
      rc = SQLITE_NOTADB;

void *sqlite3_wal_hook(
  sqlite3 *db,                    /* Attach the hook to this db handle */
  int(*xCallback)(void *, sqlite3*, const char*, int),
  void *pArg                      /* First argument passed to xCallback() */
){
  void *pRet;
  sqlite3_mutex_enter(db->mutex);
  pRet = db->pLogArg;
  db->xLogCallback = xCallback;
  db->pLogArg = pArg;
  sqlite3_mutex_leave(db->mutex);
  return pRet;
}

/*
** This function returns true if main-memory should be used instead of
** a temporary file for transient pager files and statement journals.

*/
static int unixShmClose(sqlite3_shm *pSharedMem){
  unixShm *p;            /* The connection to be closed */
  unixShmFile *pFile;    /* The underlying shared-memory file */
  unixShm **pp;          /* For looping over sibling connections */
  int nRef;              /* Number of connections to pFile */


  p = (struct unixShm*)pSharedMem;
  pFile = p->pFile;

  /* Verify that the connection being closed holds no locks */
  assert( p->exclMask==0 );
  assert( p->sharedMask==0 );

** shared memory segment.  xShmRelease() must be called to release the
** lock.
*/
static int unixShmSize(
  sqlite3_shm *pSharedMem,  /* Pointer returned by unixShmOpen() */
  int reqSize,              /* Requested size.  -1 for query only */
  int *pNewSize,            /* Write new size here */
  char **ppBuf              /* Write new buffer origin here */
){
  unixShm *p = (unixShm*)pSharedMem;
  unixShmFile *pFile = p->pFile;
  int rc = SQLITE_OK;

  sqlite3_mutex_enter(pFile->mutexBuf);
  sqlite3_mutex_enter(pFile->mutex);

  /* Return directly if this is just a lock state query, or if
  ** the connection is already in the desired locking state.
  */
  if( desiredLock==SQLITE_SHM_QUERY
   || desiredLock==p->lockState
   || (desiredLock==SQLITE_SHM_READ && p->lockState==SQLITE_SHM_READ_FULL)
  ){
    *pGotLock = p->lockState;
    return SQLITE_OK;
  }

  sqlite3_mutex_enter(pFile->mutex);
  switch( desiredLock ){
    case SQLITE_SHM_UNLOCK: {
      assert( p->lockState!=SQLITE_SHM_RECOVER );

      if( rc==SQLITE_OK ){
        p->lockState = SQLITE_SHM_RECOVER;
      }
      break;
    }
  }
  sqlite3_mutex_leave(pFile->mutex);
  *pGotLock = p->lockState;
  return rc;
}

/*
** Delete a shared-memory segment from the system.
*/
static int unixShmDelete(sqlite3_vfs *pVfs, const char *zName){

  void (*xCodecSizeChng)(void*,int,int); /* Notify of page size changes */
  void (*xCodecFree)(void*);             /* Destructor for the codec */
  void *pCodec;               /* First argument to xCodec... methods */
#endif
  char *pTmpSpace;            /* Pager.pageSize bytes of space for tmp use */
  PCache *pPCache;            /* Pointer to page cache object */
  sqlite3_backup *pBackup;    /* Pointer to list of ongoing backup processes */
  Log *pLog;                  /* Log used by "journal_mode=wal" */
};

/*
** The following global variables hold counters used for
** testing purposes only.  These variables do not exist in
** a non-testing build.  These variables are not thread-safe.
*/

  return rc;
}

/*
** Return true if this pager uses a write-ahead log instead of the usual
** rollback journal. Otherwise false.
*/
static int pagerUseLog(Pager *pPager){
  return (pPager->pLog!=0);
}

/*
** Unlock the database file. This function is a no-op if the pager
** is in exclusive mode.
**
** If the pager is currently in error state, discard the contents of 

    ** values stored in Pager.dbSize etc. might become invalid if
    ** this happens.  One can argue that this doesn't need to be cleared
    ** until the change-counter check fails in PagerSharedLock().
    ** Clearing the page size cache here is being conservative.
    */
    pPager->dbSizeValid = 0;

    if( pagerUseLog(pPager) ){
      sqlite3WalCloseSnapshot(pPager->pLog);
    }else{
      rc = osUnlock(pPager->fd, NO_LOCK);
    }
    if( rc ){
      pPager->errCode = rc;
    }
    IOTRACE(("UNLOCK %p\n", pPager))

  if( pPager->state<PAGER_RESERVED ){
    return SQLITE_OK;
  }
  releaseAllSavepoints(pPager);

  assert( isOpen(pPager->jfd) || pPager->pInJournal==0 );
  if( isOpen(pPager->jfd) ){
    assert( !pagerUseLog(pPager) );

    /* Finalize the journal file. */
    if( sqlite3IsMemJournal(pPager->jfd) ){
      assert( pPager->journalMode==PAGER_JOURNALMODE_MEMORY );
      sqlite3OsClose(pPager->jfd);
    }else if( pPager->journalMode==PAGER_JOURNALMODE_TRUNCATE ){
      if( pPager->journalOff==0 ){

#endif
  }
  sqlite3BitvecDestroy(pPager->pInJournal);
  pPager->pInJournal = 0;
  pPager->nRec = 0;
  sqlite3PcacheCleanAll(pPager->pPCache);

  if( pagerUseLog(pPager) ){
    rc2 = sqlite3WalWriteLock(pPager->pLog, 0);
    pPager->state = PAGER_SHARED;
  }else if( !pPager->exclusiveMode ){
    rc2 = osUnlock(pPager->fd, SHARED_LOCK);
    pPager->state = PAGER_SHARED;
    pPager->changeCountDone = 0;
  }else if( pPager->state==PAGER_SYNCED ){
    pPager->state = PAGER_EXCLUSIVE;

  assert( (isMainJrnl&~1)==0 );      /* isMainJrnl is 0 or 1 */
  assert( (isSavepnt&~1)==0 );       /* isSavepnt is 0 or 1 */
  assert( isMainJrnl || pDone );     /* pDone always used on sub-journals */
  assert( isSavepnt || pDone==0 );   /* pDone never used on non-savepoint */

  aData = pPager->pTmpSpace;
  assert( aData );         /* Temp storage must have already been allocated */
  assert( pagerUseLog(pPager)==0 || (!isMainJrnl && isSavepnt) );

  /* Read the page number and page data from the journal or sub-journal
  ** file. Return an error code to the caller if an IO error occurs.
  */
  jfd = isMainJrnl ? pPager->jfd : pPager->sjfd;
  rc = read32bits(jfd, *pOffset, &pgno);
  if( rc!=SQLITE_OK ) return rc;

  ** in the main journal either because the page is not in cache or else
  ** the page is marked as needSync==0.
  **
  ** 2008-04-14:  When attempting to vacuum a corrupt database file, it
  ** is possible to fail a statement on a database that does not yet exist.
  ** Do not attempt to write if database file has never been opened.
  */
  if( pagerUseLog(pPager) ){
    pPg = 0;
  }else{
    pPg = pager_lookup(pPager, pgno);
  }
  assert( pPg || !MEMDB );
  PAGERTRACE(("PLAYBACK %d page %d hash(%08x) %s\n",
           PAGERID(pPager), pgno, pager_datahash(pPager->pageSize, (u8*)aData),
           (isMainJrnl?"main-journal":"sub-journal")
  ));
  if( isMainJrnl ){
    isSynced = pPager->noSync || (*pOffset <= pPager->journalHdr);
  }else{
    isSynced = (pPg==0 || 0==(pPg->flags & PGHDR_NEED_SYNC));
  }
  if( (pPager->state>=PAGER_EXCLUSIVE)
   && isOpen(pPager->fd)
   && isSynced
  ){
    i64 ofst = (pgno-1)*(i64)pPager->pageSize;
    testcase( !isSavepnt && pPg!=0 && (pPg->flags&PGHDR_NEED_SYNC)!=0 );
    assert( !pagerUseLog(pPager) );
    rc = sqlite3OsWrite(pPager->fd, (u8*)aData, pPager->pageSize, ofst);
    if( pgno>pPager->dbFileSize ){
      pPager->dbFileSize = pgno;
    }
    if( pPager->pBackup ){
      CODEC1(pPager, aData, pgno, 3, rc=SQLITE_NOMEM);
      sqlite3BackupUpdate(pPager->pBackup, pgno, (u8*)aData);

      ** the PGHDR_NEED_SYNC flag is cleared, if the page is written to
      ** again within this transaction, it will be marked as dirty but
      ** the PGHDR_NEED_SYNC flag will not be set. It could then potentially
      ** be written out into the database file before its journal file
      ** segment is synced. If a crash occurs during or following this,
      ** database corruption may ensue.
      */
      assert( !pagerUseLog(pPager) );
      sqlite3PcacheMakeClean(pPg);
    }
#ifdef SQLITE_CHECK_PAGES
    pPg->pageHash = pager_pagehash(pPg);
#endif
    /* If this was page 1, then restore the value of Pager.dbFileVers.
    ** Do this before any decoding. */

** Otherwise, SQLITE_OK is returned.
*/
static int readDbPage(PgHdr *pPg){
  Pager *pPager = pPg->pPager; /* Pager object associated with page pPg */
  Pgno pgno = pPg->pgno;       /* Page number to read */
  int rc = SQLITE_OK;          /* Return code */
  i64 iOffset;                 /* Byte offset of file to read from */
  int isInLog = 0;             /* True if page is in log file */

  assert( pPager->state>=PAGER_SHARED && !MEMDB );
  assert( isOpen(pPager->fd) );

  if( NEVER(!isOpen(pPager->fd)) ){
    assert( pPager->tempFile );
    memset(pPg->pData, 0, pPager->pageSize);
    return SQLITE_OK;
  }

  if( pagerUseLog(pPager) ){
    /* Try to pull the page from the write-ahead log. */
    rc = sqlite3WalRead(pPager->pLog, pgno, &isInLog, pPg->pData);
  }
  if( rc==SQLITE_OK && !isInLog ){
    iOffset = (pgno-1)*(i64)pPager->pageSize;
    rc = sqlite3OsRead(pPager->fd, pPg->pData, pPager->pageSize, iOffset);
    if( rc==SQLITE_IOERR_SHORT_READ ){
      rc = SQLITE_OK;
    }
  }

  return rc;
}

/*
** This function is called to rollback a transaction on a WAL database.
*/
static int pagerRollbackLog(Pager *pPager){
  int rc;                         /* Return Code */
  PgHdr *pList;                   /* List of dirty pages to revert */

  /* Normally, if a transaction is rolled back, any backup processes are
  ** updated as data is copied out of the rollback journal and into the
  ** database. This is not generally possible with a WAL database, as
  ** rollback involves simply truncating the log file. Therefore, if one
  ** or more frames have already been written to the log (and therefore 
  ** also copied into the backup databases) as part of this transaction,
  ** the backups must be restarted.
  */
  if( sqlite3WalDirty(pPager->pLog) ){
    sqlite3BackupRestart(pPager->pBackup);
  }

  /* For all pages in the cache that are currently dirty or have already
  ** been written (but not committed) to the log file, do one of the 
  ** following:
  **
  **   + Discard the cached page (if refcount==0), or
  **   + Reload page content from the database (if refcount>0).
  */
  pPager->dbSize = pPager->dbOrigSize;
  rc = sqlite3WalUndo(pPager->pLog, pagerUndoCallback, (void *)pPager);
  pList = sqlite3PcacheDirtyList(pPager->pPCache);
  while( pList && rc==SQLITE_OK ){
    PgHdr *pNext = pList->pDirty;
    rc = pagerUndoCallback((void *)pPager, pList->pgno);
    pList = pNext;
  }

  }

  /* Set the database size back to the value it was before the savepoint 
  ** being reverted was opened.
  */
  pPager->dbSize = pSavepoint ? pSavepoint->nOrig : pPager->dbOrigSize;

  if( !pSavepoint && pagerUseLog(pPager) ){
    return pagerRollbackLog(pPager);
  }

  /* Use pPager->journalOff as the effective size of the main rollback
  ** journal.  The actual file might be larger than this in
  ** PAGER_JOURNALMODE_TRUNCATE or PAGER_JOURNALMODE_PERSIST.  But anything
  ** past pPager->journalOff is off-limits to us.
  */
  szJ = pPager->journalOff;
  assert( pagerUseLog(pPager)==0 || szJ==0 );

  /* Begin by rolling back records from the main journal starting at
  ** PagerSavepoint.iOffset and continuing to the next journal header.
  ** There might be records in the main journal that have a page number
  ** greater than the current database size (pPager->dbSize) but those
  ** will be skipped automatically.  Pages are added to pDone as they
  ** are played back.
  */
  if( pSavepoint && !pagerUseLog(pPager) ){
    iHdrOff = pSavepoint->iHdrOffset ? pSavepoint->iHdrOffset : szJ;
    pPager->journalOff = pSavepoint->iOffset;
    while( rc==SQLITE_OK && pPager->journalOff<iHdrOff ){
      rc = pager_playback_one_page(pPager, &pPager->journalOff, pDone, 1, 1);
    }
    assert( rc!=SQLITE_DONE );
  }else{

  ** previously rolled back out of the main journal (and are hence in pDone)
  ** will be skipped.  Out-of-range pages are also skipped.
  */
  if( pSavepoint ){
    u32 ii;            /* Loop counter */
    i64 offset = pSavepoint->iSubRec*(4+pPager->pageSize);

    if( pagerUseLog(pPager) ){
      rc = sqlite3WalSavepointUndo(pPager->pLog, pSavepoint->iFrame);
    }
    for(ii=pSavepoint->iSubRec; rc==SQLITE_OK && ii<pPager->nSubRec; ii++){
      assert( offset==ii*(4+pPager->pageSize) );
      rc = pager_playback_one_page(pPager, &offset, pDone, 0, 1);
    }
    assert( rc!=SQLITE_DONE );
  }

  /* Determine the number of pages in the file. Store this in nPage. */
  if( pPager->dbSizeValid ){
    nPage = pPager->dbSize;
  }else{
    int rc;                 /* Error returned by OsFileSize() */
    i64 n = 0;              /* File size in bytes returned by OsFileSize() */

    if( pagerUseLog(pPager) ){
      sqlite3WalDbsize(pPager->pLog, &nPage);
    }

    if( nPage==0 ){
      assert( isOpen(pPager->fd) || pPager->tempFile );
      if( isOpen(pPager->fd) ){
        if( SQLITE_OK!=(rc = sqlite3OsFileSize(pPager->fd, &n)) ){
          pager_error(pPager, rc);

int sqlite3PagerClose(Pager *pPager){
  u8 *pTmp = (u8 *)pPager->pTmpSpace;

  disable_simulated_io_errors();
  sqlite3BeginBenignMalloc();
  pPager->errCode = 0;
  pPager->exclusiveMode = 0;
  sqlite3WalClose(pPager->pLog, pPager->fd, 
    (pPager->noSync ? 0 : pPager->sync_flags), pTmp
  );
  pPager->pLog = 0;
  pager_reset(pPager);
  if( MEMDB ){
    pager_unlock(pPager);
  }else{
    /* Set Pager.journalHdr to -1 for the benefit of the pager_playback() 
    ** call which may be made from within pagerUnlockAndRollback(). If it
    ** is not -1, then the unsynced portion of an open journal file may

  **
  ** While the pager is in the RESERVED state, the original database file
  ** is unchanged and we can rollback without having to playback the
  ** journal into the original database file.  Once we transition to
  ** EXCLUSIVE, it means the database file has been changed and any rollback
  ** will require a journal playback.
  */
  assert( !pagerUseLog(pList->pPager) );
  assert( pPager->state>=PAGER_RESERVED );
  rc = pager_wait_on_lock(pPager, EXCLUSIVE_LOCK);

  /* If the file is a temp-file has not yet been opened, open it now. It
  ** is not possible for rc to be other than SQLITE_OK if this branch
  ** is taken, as pager_wait_on_lock() is a no-op for temp-files.
  */

    void *pData = pPg->pData;
    i64 offset = pPager->nSubRec*(4+pPager->pageSize);
    char *pData2;

    CODEC2(pPager, pData, pPg->pgno, 7, return SQLITE_NOMEM, pData2);
    PAGERTRACE(("STMT-JOURNAL %d page %d\n", PAGERID(pPager), pPg->pgno));
  
    assert( pagerUseLog(pPager) 
         || pageInJournal(pPg) 
         || pPg->pgno>pPager->dbOrigSize 
    );
    rc = write32bits(pPager->sjfd, offset, pPg->pgno);
    if( rc==SQLITE_OK ){
      rc = sqlite3OsWrite(pPager->sjfd, pData2, pPager->pageSize, offset+4);
    }

/*
** This function is a wrapper around sqlite3WalFrames(). As well as logging
** the contents of the list of pages headed by pList (connected by pDirty),
** this function notifies any active backup processes that the pages have
** changed. 
*/ 
static int pagerLogFrames(
  Pager *pPager,                  /* Pager object */
  PgHdr *pList,                   /* List of frames to log */
  Pgno nTruncate,                 /* Database size after this commit */
  int isCommit,                   /* True if this is a commit */
  int sync_flags                  /* Flags to pass to OsSync() (or 0) */
){
  int rc;                         /* Return code */

  assert( pPager->pLog );
  rc = sqlite3WalFrames(pPager->pLog, 
      pPager->pageSize, pList, nTruncate, isCommit, sync_flags
  );
  if( rc==SQLITE_OK && pPager->pBackup ){
    PgHdr *p;
    for(p=pList; p; p=p->pDirty){
      sqlite3BackupUpdate(pPager->pBackup, p->pgno, (u8 *)p->pData);
    }

  Pager *pPager = (Pager *)p;
  int rc = SQLITE_OK;

  assert( pPg->pPager==pPager );
  assert( pPg->flags&PGHDR_DIRTY );

  pPg->pDirty = 0;
  if( pagerUseLog(pPager) ){
    /* Write a single frame for this page to the log. */
    if( subjRequiresPage(pPg) ){ 
      rc = subjournalPage(pPg); 
    }
    if( rc==SQLITE_OK ){
      rc = pagerLogFrames(pPager, pPg, 0, 0, 0);
    }
  }else{
    /* The doNotSync flag is set by the sqlite3PagerWrite() function while it
    ** is journalling a set of two or more database pages that are stored
    ** on the same disk sector. Syncing the journal is not allowed while
    ** this is happening as it is important that all members of such a
    ** set of pages are synced to disk together. So, if the page this function

** a WAL, this ensures there is no race condition between the xAccess() 
** below and an xDelete() being executed by some other connection.
*/
static int pagerHasWAL(Pager *pPager, int *pExists){
  int rc;                         /* Return code */

  if( !pPager->tempFile ){
    char *zLog = sqlite3_mprintf("%s-wal", pPager->zFilename);
    if( !zLog ){
      rc = SQLITE_NOMEM;
    }else{
      rc = sqlite3OsAccess(pPager->pVfs, zLog, SQLITE_ACCESS_EXISTS, pExists);
      sqlite3_free(zLog);
    }
  }else{
    rc = SQLITE_OK;
    *pExists = 0;
  }
  return rc;
}

static int pagerOpenSnapshot(Pager *pPager){
  int rc;                         /* Return code */
  int changed = 0;                /* True if cache must be reset */

  assert( pagerUseLog(pPager) );

  rc = sqlite3WalOpenSnapshot(pPager->pLog, &changed);
  if( rc==SQLITE_OK ){
    int dummy;
    if( changed ){
      pager_reset(pPager);
      assert( pPager->errCode || pPager->dbSizeValid==0 );
    }
    rc = sqlite3PagerPagecount(pPager, &dummy);

    if( isOpen(pPager->jfd) || pPager->zJournal ){
      isErrorReset = 1;
    }
    pPager->errCode = SQLITE_OK;
    pager_reset(pPager);
  }

  if( pagerUseLog(pPager) ){
    rc = pagerOpenSnapshot(pPager);
  }else if( pPager->state==PAGER_UNLOCK || isErrorReset ){
    sqlite3_vfs * const pVfs = pPager->pVfs;
    int isHotJournal = 0;
    int isWal = 0;
    assert( !MEMDB );
    assert( sqlite3PcacheRefCount(pPager->pPCache)==0 );

    rc = pagerHasWAL(pPager, &isWal);
    if( rc!=SQLITE_OK ){
      goto failed;
    }
    if( isWal ){
      pager_reset(pPager);
      rc = sqlite3PagerOpenLog(pPager, 0);
      if( rc==SQLITE_OK ){
        rc = pagerOpenSnapshot(pPager);
      }
    }else if( pPager->journalMode==PAGER_JOURNALMODE_WAL ){
      pPager->journalMode = PAGER_JOURNALMODE_DELETE;
    }
  }

**
** SQLITE_OK is returned if everything goes according to plan. 
** An SQLITE_IOERR_XXX error code is returned if a call to 
** sqlite3OsOpen() fails.
*/
static int openSubJournal(Pager *pPager){
  int rc = SQLITE_OK;
  if( (pagerUseLog(pPager) || isOpen(pPager->jfd)) && !isOpen(pPager->sjfd) ){
    if( pPager->journalMode==PAGER_JOURNALMODE_MEMORY || pPager->subjInMemory ){
      sqlite3MemJournalOpen(pPager->sjfd);
    }else{
      rc = pagerOpentemp(pPager, pPager->sjfd, SQLITE_OPEN_SUBJOURNAL);
    }
  }
  return rc;

  int rc = SQLITE_OK;
  assert( pPager->state!=PAGER_UNLOCK );
  pPager->subjInMemory = (u8)subjInMemory;
  if( pPager->state==PAGER_SHARED ){
    assert( pPager->pInJournal==0 );
    assert( !MEMDB && !pPager->tempFile );

    if( pagerUseLog(pPager) ){
      /* Grab the write lock on the log file. If successful, upgrade to
      ** PAGER_EXCLUSIVE state. Otherwise, return an error code to the caller.
      ** The busy-handler is not invoked if another connection already
      ** holds the write-lock. If possible, the upper layer will call it.
      */
      rc = sqlite3WalWriteLock(pPager->pLog, 1);
      if( rc==SQLITE_OK ){
        pPager->dbOrigSize = pPager->dbSize;
        pPager->state = PAGER_RESERVED;
        pPager->journalOff = 0;
      }
    }else{
      /* Obtain a RESERVED lock on the database file. If the exFlag parameter

  }else if( isOpen(pPager->jfd) && pPager->journalOff==0 ){
    /* This happens when the pager was in exclusive-access mode the last
    ** time a (read or write) transaction was successfully concluded
    ** by this connection. Instead of deleting the journal file it was 
    ** kept open and either was truncated to 0 bytes or its header was
    ** overwritten with zeros.
    */
    assert( pagerUseLog(pPager)==0 );
    assert( pPager->nRec==0 );
    assert( pPager->dbOrigSize==0 );
    assert( pPager->pInJournal==0 );
    rc = pager_open_journal(pPager);
  }

  PAGERTRACE(("TRANSACTION %d\n", PAGERID(pPager)));

  CHECK_PAGE(pPg);

  /* Mark the page as dirty.  If the page has already been written
  ** to the journal then we can return right away.
  */
  sqlite3PcacheMakeDirty(pPg);
  if( pageInJournal(pPg) && !subjRequiresPage(pPg) ){
    assert( !pagerUseLog(pPager) );
    pPager->dbModified = 1;
  }else{

    /* If we get this far, it means that the page needs to be
    ** written to the transaction journal or the ckeckpoint journal
    ** or both.
    **
    ** Higher level routines should have already started a transaction,
    ** which means they have acquired the necessary locks but the rollback
    ** journal might not yet be open.
    */
    rc = sqlite3PagerBegin(pPager, 0, pPager->subjInMemory);
    if( rc!=SQLITE_OK ){
      return rc;
    }
    if( !isOpen(pPager->jfd) 
     && pPager->journalMode!=PAGER_JOURNALMODE_OFF 
     && !pagerUseLog(pPager)
    ){
      assert( pPager->useJournal );
      rc = pager_open_journal(pPager);
      if( rc!=SQLITE_OK ) return rc;
    }
    pPager->dbModified = 1;
  
    /* The transaction journal now exists and we have a RESERVED or an
    ** EXCLUSIVE lock on the main database file.  Write the current page to
    ** the transaction journal if it is not there already.
    */
    if( !pageInJournal(pPg) && isOpen(pPager->jfd) ){
      assert( !pagerUseLog(pPager) );
      if( pPg->pgno<=pPager->dbOrigSize ){
        u32 cksum;
        char *pData2;

        /* We should never write to the journal file the page that
        ** contains the database locks.  The following assert verifies
        ** that we do not. */

  if( MEMDB && pPager->dbModified ){
    /* If this is an in-memory db, or no pages have been written to, or this
    ** function has already been called, it is mostly a no-op.  However, any
    ** backup in progress needs to be restarted.
    */
    sqlite3BackupRestart(pPager->pBackup);
  }else if( pPager->state!=PAGER_SYNCED && pPager->dbModified ){
    if( pagerUseLog(pPager) ){
      PgHdr *pList = sqlite3PcacheDirtyList(pPager->pPCache);
      if( pList ){
        rc = pagerLogFrames(pPager, pList, pPager->dbSize, 1, 
            (pPager->fullSync ? pPager->sync_flags : 0)
        );
      }
      sqlite3PcacheCleanAll(pPager->pPCache);
    }else{
      /* The following block updates the change-counter. Exactly how it
      ** does this depends on whether or not the atomic-update optimization

**   the journal file. It needs to be left in the file-system so that
**   some other process can use it to restore the database state (by
**   hot-journal rollback).
*/
int sqlite3PagerRollback(Pager *pPager){
  int rc = SQLITE_OK;                  /* Return code */
  PAGERTRACE(("ROLLBACK %d\n", PAGERID(pPager)));
  if( pagerUseLog(pPager) ){
    int rc2;

    rc = sqlite3PagerSavepoint(pPager, SAVEPOINT_ROLLBACK, -1);
    rc2 = pager_end_transaction(pPager, pPager->setMaster);
    if( rc==SQLITE_OK ) rc = rc2;
  }else if( !pPager->dbModified || !isOpen(pPager->jfd) ){
    rc = pager_end_transaction(pPager, pPager->setMaster);

        aNew[ii].iOffset = JOURNAL_HDR_SZ(pPager);
      }
      aNew[ii].iSubRec = pPager->nSubRec;
      aNew[ii].pInSavepoint = sqlite3BitvecCreate(nPage);
      if( !aNew[ii].pInSavepoint ){
        return SQLITE_NOMEM;
      }
      if( pagerUseLog(pPager) ){
        aNew[ii].iFrame = sqlite3WalSavepoint(pPager->pLog);
      }
    }

    /* Open the sub-journal, if it is not already opened. */
    rc = openSubJournal(pPager);
    assertTruncateConstraint(pPager);
  }

      }
    }
    /* Else this is a rollback operation, playback the specified savepoint.
    ** If this is a temp-file, it is possible that the journal file has
    ** not yet been opened. In this case there have been no changes to
    ** the database file, so the playback operation can be skipped.
    */
    else if( pagerUseLog(pPager) || isOpen(pPager->jfd) ){
      PagerSavepoint *pSavepoint = (nNew==0)?0:&pPager->aSavepoint[nNew-1];
      rc = pagerPlaybackSavepoint(pPager, pSavepoint);
      assert(rc!=SQLITE_DONE);
    }
  
  }
  return rc;

}

/*
** This function is called when the user invokes "PRAGMA checkpoint".
*/
int sqlite3PagerCheckpoint(Pager *pPager){
  int rc = SQLITE_OK;
  if( pPager->pLog ){
    u8 *zBuf = (u8 *)pPager->pTmpSpace;
    rc = sqlite3WalCheckpoint(pPager->pLog, pPager->fd, 
        (pPager->noSync ? 0 : pPager->sync_flags),
        zBuf, pPager->xBusyHandler, pPager->pBusyHandlerArg
    );
  }
  return rc;
}

int sqlite3PagerLogCallback(Pager *pPager){
  return sqlite3WalCallback(pPager->pLog);
}

/*
** Open a connection to the write-ahead log file for pager pPager. If
** the log connection is already open, this function is a no-op.
**
** The caller must be holding a SHARED lock on the database file to call
** this function.
*/
int sqlite3PagerOpenLog(Pager *pPager, int *pisOpen){
  int rc = SQLITE_OK;             /* Return code */

  assert( pPager->state>=PAGER_SHARED );
  if( !pPager->pLog ){

    /* Open the connection to the log file. If this operation fails, 
    ** (e.g. due to malloc() failure), unlock the database file and 
    ** return an error code.
    */
    rc = sqlite3WalOpen(pPager->pVfs, pPager->zFilename, &pPager->pLog);
    if( rc==SQLITE_OK ){
      pPager->journalMode = PAGER_JOURNALMODE_WAL;
    }
  }else{
    *pisOpen = 1;
  }

  return rc;
}


/*
** This function is called to close the connection to the log file prior
** to switching from WAL to rollback mode.
**
** Before closing the log file, this function attempts to take an 
** EXCLUSIVE lock on the database file. If this cannot be obtained, an
** error (SQLITE_BUSY) is returned and the log connection is not closed.
** If successful, the EXCLUSIVE lock is not released before returning.
*/
int sqlite3PagerCloseLog(Pager *pPager){
  int rc = SQLITE_OK;

  assert( pPager->journalMode==PAGER_JOURNALMODE_WAL );

  /* If the log file is not already open, but does exist in the file-system,
  ** it may need to be checkpointed before the connection can switch to
  ** rollback mode. Open it now so this can happen.
  */
  if( !pPager->pLog ){
    int logexists = 0;
    rc = sqlite3OsLock(pPager->fd, SQLITE_LOCK_SHARED);
    if( rc==SQLITE_OK ){
      rc = pagerHasWAL(pPager, &logexists);
    }
    if( rc==SQLITE_OK && logexists ){
      rc = sqlite3WalOpen(pPager->pVfs, pPager->zFilename, &pPager->pLog);
    }
  }
    
  /* Checkpoint and close the log. Because an EXCLUSIVE lock is held on
  ** the database file, the log and log-summary files will be deleted.
  */
  if( rc==SQLITE_OK && pPager->pLog ){
    rc = sqlite3OsLock(pPager->fd, SQLITE_LOCK_EXCLUSIVE);
    if( rc==SQLITE_OK ){
      rc = sqlite3WalClose(pPager->pLog, pPager->fd,
        (pPager->noSync ? 0 : pPager->sync_flags), 
        (u8*)pPager->pTmpSpace
      );
      pPager->pLog = 0;
    }
  }
  return rc;
}

#endif /* SQLITE_OMIT_DISKIO */

int sqlite3PagerCommitPhaseTwo(Pager*);
int sqlite3PagerRollback(Pager*);
int sqlite3PagerOpenSavepoint(Pager *pPager, int n);
int sqlite3PagerSavepoint(Pager *pPager, int op, int iSavepoint);
int sqlite3PagerSharedLock(Pager *pPager);

int sqlite3PagerCheckpoint(Pager *pPager);
int sqlite3PagerLogCallback(Pager *pPager);
int sqlite3PagerOpenLog(Pager *pPager, int *pisOpen);
int sqlite3PagerCloseLog(Pager *pPager);

/* Functions used to query pager state and configuration. */
u8 sqlite3PagerIsreadonly(Pager*);
int sqlite3PagerRefcount(Pager*);
int sqlite3PagerMemUsed(Pager*);
const char *sqlite3PagerFilename(Pager*);
const sqlite3_vfs *sqlite3PagerVfs(Pager*);

  int (*xCurrentTime)(sqlite3_vfs*, double*);
  int (*xGetLastError)(sqlite3_vfs*, int, char *);
  /*
  ** The methods above are in version 1 of the sqlite_vfs object
  ** definition.  Those that follow are added in version 2 or later
  */
  int (*xShmOpen)(sqlite3_vfs*, const char *zName, sqlite3_shm**);
  int (*xShmSize)(sqlite3_shm*, int reqSize, int *pNewSize, char**);
  int (*xShmRelease)(sqlite3_shm*);
  int (*xShmPush)(sqlite3_shm*);
  int (*xShmPull)(sqlite3_shm*);
  int (*xShmLock)(sqlite3_shm*, int desiredLock, int *gotLock);
  int (*xShmClose)(sqlite3_shm*);
  int (*xShmDelete)(sqlite3_vfs*, const char *zName);
  int (*xRename)(sqlite3_vfs*, const char *zOld, const char *zNew, int dirSync);

  void *pProfileArg;                        /* Argument to profile function */
  void *pCommitArg;                 /* Argument to xCommitCallback() */   
  int (*xCommitCallback)(void*);    /* Invoked at every commit. */
  void *pRollbackArg;               /* Argument to xRollbackCallback() */   
  void (*xRollbackCallback)(void*); /* Invoked at every commit. */
  void *pUpdateArg;
  void (*xUpdateCallback)(void*,int, const char*,const char*,sqlite_int64);
  int (*xLogCallback)(void *, sqlite3 *, const char *, int);
  void *pLogArg;
  void(*xCollNeeded)(void*,sqlite3*,int eTextRep,const char*);
  void(*xCollNeeded16)(void*,sqlite3*,int eTextRep,const void*);
  void *pCollNeededArg;
  sqlite3_value *pErr;          /* Most recent error message */
  char *zErrMsg;                /* Most recent error message (UTF-8 encoded) */
  char *zErrMsg16;              /* Most recent error message (UTF-16 encoded) */
  union {

  extern int sqlite3WhereTrace;
  extern int sqlite3OSTrace;
  extern int sqlite3VdbeAddopTrace;
#endif
#ifdef SQLITE_TEST
  extern char sqlite3_query_plan[];
  static char *query_plan = sqlite3_query_plan;
  extern int sqlite3_walsummary_mmap_incr;                 /* In wal.c */
#ifdef SQLITE_ENABLE_FTS3
  extern int sqlite3_fts3_enable_parentheses;
#endif
#endif

  for(i=0; i<sizeof(aCmd)/sizeof(aCmd[0]); i++){
    Tcl_CreateCommand(interp, aCmd[i].zName, aCmd[i].xProc, 0, 0);

#if SQLITE_OS_WIN
  Tcl_LinkVar(interp, "sqlite_os_type",
      (char*)&sqlite3_os_type, TCL_LINK_INT);
#endif
#ifdef SQLITE_TEST
  Tcl_LinkVar(interp, "sqlite_query_plan",
      (char*)&query_plan, TCL_LINK_STRING|TCL_LINK_READ_ONLY);
  Tcl_LinkVar(interp, "sqlite_walsummary_mmap_incr",
      (char*)&sqlite3_walsummary_mmap_incr, TCL_LINK_INT);
#endif
#ifdef SQLITE_DEBUG
  Tcl_LinkVar(interp, "sqlite_addop_trace",
      (char*)&sqlite3VdbeAddopTrace, TCL_LINK_INT);
  Tcl_LinkVar(interp, "sqlite_where_trace",
      (char*)&sqlite3WhereTrace, TCL_LINK_INT);
  Tcl_LinkVar(interp, "sqlite_os_trace",

            "cannot change %s wal mode from within a transaction",
            (eNew==PAGER_JOURNALMODE_WAL ? "into" : "out of")
        );
      }else{
  
        if( eOld==PAGER_JOURNALMODE_WAL ){
          /* If leaving WAL mode, close the log file. If successful, the call
          ** to PagerCloseLog() checkpoints and deletes the write-ahead-log 
          ** file. An EXCLUSIVE lock may still be held on the database file 
          ** after a successful return. 
          */
          rc = sqlite3PagerCloseLog(pPager);
          if( rc!=SQLITE_OK ) goto abort_due_to_error;
          sqlite3PagerJournalMode(pPager, eNew);
        }else{
          sqlite3PagerJournalMode(pPager, PAGER_JOURNALMODE_DELETE);
        }
  
        /* Open a transaction on the database file. Regardless of the journal

void sqlite3_result_error_nomem(sqlite3_context *pCtx){
  assert( sqlite3_mutex_held(pCtx->s.db->mutex) );
  sqlite3VdbeMemSetNull(&pCtx->s);
  pCtx->isError = SQLITE_NOMEM;
  pCtx->s.db->mallocFailed = 1;
}

static int doLogCallbacks(sqlite3 *db){
  int i;
  int rc = SQLITE_OK;
  for(i=0; i<db->nDb; i++){
    Btree *pBt = db->aDb[i].pBt;
    if( pBt ){
      int nEntry = sqlite3PagerLogCallback(sqlite3BtreePager(pBt));
      if( db->xLogCallback && nEntry>0 && rc==SQLITE_OK
       && db->xLogCallback(db->pLogArg, db, db->aDb[i].zName, nEntry)
      ){
        rc = sqlite3PagerCheckpoint(sqlite3BtreePager(pBt));
      }
    }
  }
  return rc;
}

    elapseTime -= p->startTime;
    db->xProfile(db->pProfileArg, p->zSql, elapseTime);
  }
#endif

  if( rc==SQLITE_DONE ){
    assert( p->rc==SQLITE_OK );
    p->rc = doLogCallbacks(db);
    if( p->rc!=SQLITE_OK ){
      rc = SQLITE_ERROR;
    }
  }

  db->errCode = rc;
  if( SQLITE_NOMEM==sqlite3ApiExit(p->db, p->rc) ){

/*











** This file contains the implementation of a log file used in 
** "journal_mode=wal" mode.
*/



/*
** LOG FILE FORMAT
**
** A log file consists of a header followed by zero or more log frames.
** The log header is 12 bytes in size and consists of the following three
** big-endian 32-bit unsigned integer values:
**
**     0: Database page size,
**     4: Randomly selected salt value 1,
**     8: Randomly selected salt value 2.
**
** Immediately following the log header are zero or more log frames. Each
** frame itself consists of a 16-byte header followed by a <page-size> bytes
** of page data. The header is broken into 4 big-endian 32-bit unsigned 
** integer values, as follows:
**
**     0: Page number.
**     4: For commit records, the size of the database image in pages 
**        after the commit. For all other records, zero.
**     8: Checksum value 1.
**    12: Checksum value 2.
*/

/* 
** LOG SUMMARY FILE FORMAT
**
** The log-summary file consists of a header region, followed by an 
** region that contains no useful data (used to apply byte-range locks
** to), followed by the data region. 
**
** The contents of both the header and data region are specified in terms
** of 1, 2 and 4 byte unsigned integers. All integers are stored in 
** machine-endian order.

**
** A log-summary file is essentially a shadow-pager map. It contains a
** mapping from database page number to the set of locations in the log 
** file that contain versions of the database page. When a database 
** client needs to read a page of data, it first queries the log-summary
** file to determine if the required version of the page is stored in
** the log. If so, it is read from the log file. If not, it is read from
** the database file.
**
** Whenever a transaction is appended to the log or a checkpoint transfers
** data from the log file into the database file, the log-summary is 
** updated accordingly.
**
** The fields in the log-summary file header are described in the comment 
** directly above the definition of struct LogSummaryHdr (see below). 
** Immediately following the fields in the LogSummaryHdr structure is
** an 8 byte checksum based on the contents of the header. This field is
** not the same as the iCheck1 and iCheck2 fields of the LogSummaryHdr.
*/

#include "wal.h"

#include <unistd.h>
#include <fcntl.h>
#include <sys/mman.h>

typedef struct LogSummaryHdr LogSummaryHdr;
typedef struct LogSummary LogSummary;
typedef struct LogIterator LogIterator;
typedef struct LogLock LogLock;


/*
** The following structure may be used to store the same data that
** is stored in the log-summary header.
**
** Member variables iCheck1 and iCheck2 contain the checksum for the
** last frame written to the log, or 2 and 3 respectively if the log 
** is currently empty.
*/
struct LogSummaryHdr {
  u32 iChange;                    /* Counter incremented each transaction */
  u32 pgsz;                       /* Database page size in bytes */
  u32 iLastPg;                    /* Address of last valid frame in log */
  u32 nPage;                      /* Size of database in pages */
  u32 iCheck1;                    /* Checkpoint value 1 */
  u32 iCheck2;                    /* Checkpoint value 2 */
};

/* Size of serialized LogSummaryHdr object. */
#define LOGSUMMARY_HDR_NFIELD (sizeof(LogSummaryHdr) / sizeof(u32))

/* A block of 16 bytes beginning at LOGSUMMARY_LOCK_OFFSET is reserved
** for locks. Since some systems only feature mandatory file-locks, we
** do not read or write data from the region of the file on which locks
** are applied.
*/
#define LOGSUMMARY_LOCK_OFFSET   ((sizeof(LogSummaryHdr))+2*sizeof(u32))
#define LOGSUMMARY_LOCK_RESERVED 16

/* Size of header before each frame in log file */
#define LOG_FRAME_HDRSIZE 16

/* Size of log header */
#define LOG_HDRSIZE 12

/*
** Return the offset of frame iFrame in the log file, assuming a database
** page size of pgsz bytes. The offset returned is to the start of the
** log frame-header.
*/
#define logFrameOffset(iFrame, pgsz) (                               \
  LOG_HDRSIZE + ((iFrame)-1)*((pgsz)+LOG_FRAME_HDRSIZE)        \
)

/*
** If using mmap() to access a shared (or otherwise) log-summary file, then
** the mapping size is incremented in units of the following size.
**
** A 64 KB log-summary mapping corresponds to a log file containing over
** 13000 frames, so the mapping size does not need to be increased often.
*/
#ifdef SQLITE_TEST
int sqlite3_walsummary_mmap_incr = 128;
# define LOGSUMMARY_MMAP_INCREMENT sqlite3_walsummary_mmap_incr
#else
# define LOGSUMMARY_MMAP_INCREMENT (64*1024)
#endif

/*
** There is one instance of this structure for each log-summary object
** that this process has a connection to. They are stored in a linked
** list starting at pLogSummary (global variable).
**
** TODO: LogSummary.fd is a unix file descriptor. Unix APIs are used 
**       directly in this implementation because the VFS does not support
**       the required blocking file-locks.
*/
struct LogSummary {
  sqlite3_mutex *mutex;           /* Mutex used to protect this object */
  int nRef;                       /* Number of pointers to this structure */
  int fd;                         /* File descriptor open on log-summary */
  char *zPath;                    /* Path to associated WAL file */
  LogLock *pLock;                 /* Linked list of locks on this object */
  LogSummary *pNext;              /* Next in global list */

  int nData;                      /* Size of aData allocation/mapping */
  u32 *aData;                     /* File body */
};

/*
** This module uses three different types of file-locks. All are taken
** on the log-summary file. The three types of locks are as follows:
**
** MUTEX:  The MUTEX lock is used as a robust inter-process mutex. It
**         is held while the log-summary header is modified, and 
**         sometimes when it is read. It is also held while a new client
**         obtains the DMH lock (see below), and while log recovery is
**         being run.
**
** DMS:    The DMS (Dead Mans Switch mechanism) lock is used to ensure
**         that log-recovery is always run following a system restart.
**         When it first opens a log-summary file, a process takes a
**         SHARED lock on the DMH region. This lock is not released until
**         the log-summary file is closed. 
**
**         The process then attempts to upgrade to an EXCLUSIVE lock. If 
**         successful, then the contents of the log-summary file are deemed 
**         suspect and the log-summary header zeroed. This forces the
**         first process that reads the log-summary file to run log 
**         recovery. After zeroing the log-summary header, the process
**         downgrades to a SHARED lock on the DMH region.
**

**         If the attempt to obtain the EXCLUSIVE lock fails, then the
**         process concludes that some other process is already using the
**         log-summary file, and it can therefore be trusted.
**
**         The procedure described in the previous three paragraphs (taking
**         a SHARED lock and then upgrading to an EXCLUSIVE lock to check
**         if the process is the only one to have an open connection to the 
**         log file) is protected by holding the MUTEX lock. This avoids the
**         race condition wherein the first two clients connect almost 
**         simultaneously following a system restart and each prevents 
**         the other from obtaining the EXCLUSIVE lock.
**
**
** REGION: There are 4 different region locks, regions A, B, C and D.
**         Various EXCLUSIVE and SHARED locks on these regions are obtained
**         when a client reads, writes or checkpoints the database.
**

**    To obtain a reader lock:
**
**         1. Attempt a SHARED lock on regions A and B.
**         2. If step 1 is successful, drop the lock on region B. Or, if
**            it is unsuccessful, attempt a SHARED lock on region D.
**         3. Repeat the above until the lock attempt in step 1 or 2 is 
**            successful.
**
**         The reader lock is released when the read transaction is finished.
**
**    To obtain a writer lock:
**
**         1. Take (wait for) an EXCLUSIVE lock on regions C and D.
**
**         The locks are released after the write transaction is finished
**         and, if any frames were committed to the log, the log-summary
**         file updated.
**
**    To obtain a checkpointer lock:
**
**         1. Take (wait for) an EXCLUSIVE lock on regions B and C.
**         2. Take (wait for) an EXCLUSIVE lock on region A.
**
**         Step 1 waits until any existing writer has finished. And forces
**         all new readers to become "region D" readers.
**
**         Step 2 causes the checkpointer to wait until all existing region A
**         readers have finished their transactions. Once the exclusive lock
**         on region A has been obtained, only "region D" readers exist.
**         These readers are operating on the snapshot at the head of the
**         log. As such, the log can be safely copied into the database file
**         without interfering with the readers.
**
**         Once the checkpoint has finished and the log-summary header
**         updated (to indicate the log contents can now be ignored), all
**         locks are released.
**
**         However, there may still exist region D readers using data in 
**         the body of the log file, so the log file itself cannot be 
**         truncated or overwritten until all region D readers have finished.
**         That requirement is satisfied, because writers (the clients that
**         write to the log file) require an exclusive lock on region D.
**         Which they cannot get until all region D readers have finished.
*/
#define LOG_LOCK_MUTEX  (LOGSUMMARY_LOCK_OFFSET)
#define LOG_LOCK_DMH    (LOG_LOCK_MUTEX+1)
#define LOG_LOCK_REGION (LOG_LOCK_DMH+1)

/*
** The four lockable regions associated with each log-summary. A connection
** may take either a SHARED or EXCLUSIVE lock on each. An ORed combination
** of the following bitmasks is passed as the second argument to the
** logLockRegion() function.
*/
#define LOG_REGION_A 0x01
#define LOG_REGION_B 0x02
#define LOG_REGION_C 0x04
#define LOG_REGION_D 0x08

/*
** Values for the third parameter to logLockRegion().
*/
#define LOG_UNLOCK  0             /* Unlock a range of bytes */
#define LOG_RDLOCK  1             /* Put a SHARED lock on a range of bytes */
#define LOG_WRLOCK  2             /* Put an EXCLUSIVE lock on a byte-range */
#define LOG_WRLOCKW 3             /* Block on EXCLUSIVE lock on a byte-range */

/*
** A single instance of this structure is allocated as part of each 
** connection to a database log. All structures associated with the 
** same log file are linked together into a list using LogLock.pNext
** starting at LogSummary.pLock.
**
** The mLock field of the structure describes the locks (if any) 
** currently held by the connection. If a SHARED lock is held on
** any of the four locking regions, then the associated LOG_REGION_X
** bit (see above) is set. If an EXCLUSIVE lock is held on the region,
** then the (LOG_REGION_X << 8) bit is set.
*/
struct LogLock {
  LogLock *pNext;                 /* Next lock on the same log */
  u32 mLock;                      /* Mask of locks */
};

struct Log {
  LogSummary *pSummary;           /* Log file summary data */
  sqlite3_vfs *pVfs;              /* The VFS used to create pFd */
  sqlite3_file *pFd;              /* File handle for log file */
  int isLocked;                   /* Non-zero if a snapshot is held open */

  int isWriteLocked;              /* True if this is the writer connection */
  u32 iCallback;                  /* Value to pass to log callback (or 0) */
  LogSummaryHdr hdr;              /* Log summary header for current snapshot */
  LogLock lock;                   /* Lock held by this connection (if any) */

};


/*
** This structure is used to implement an iterator that iterates through
** all frames in the log in database page order. Where two or more frames
** correspond to the same database page, the iterator visits only the 
** frame most recently written to the log.
**
** The internals of this structure are only accessed by:
**
**   logIteratorInit() - Create a new iterator,
**   logIteratorNext() - Step an iterator,
**   logIteratorFree() - Free an iterator.
**
** This functionality is used by the checkpoint code (see logCheckpoint()).
*/
struct LogIterator {
  int nSegment;                   /* Size of LogIterator.aSegment[] array */
  int nFinal;                     /* Elements in segment nSegment-1 */
  struct LogSegment {
    int iNext;                    /* Next aIndex index */
    u8 *aIndex;                   /* Pointer to index array */
    u32 *aDbPage;                 /* Pointer to db page array */
  } aSegment[1];
};



/*
** List of all LogSummary objects created by this process. Protected by
** static mutex LOG_SUMMARY_MUTEX. TODO: Should have a dedicated mutex
** here instead of borrowing the LRU mutex.
*/
#define LOG_SUMMARY_MUTEX SQLITE_MUTEX_STATIC_LRU
static LogSummary *pLogSummary = 0;

/*
** Generate an 8 byte checksum based on the data in array aByte[] and the
** initial values of aCksum[0] and aCksum[1]. The checksum is written into
** aCksum[] before returning.
**
** The range of bytes to checksum is treated as an array of 32-bit 
** little-endian unsigned integers. For each integer X in the array, from
** start to finish, do the following:
**
**   aCksum[0] += X;
**   aCksum[1] += aCksum[0];
**
** For the calculation above, use 64-bit unsigned accumulators. Before
** returning, truncate the values to 32-bits as follows: 
**
**   aCksum[0] = (u32)(aCksum[0] + (aCksum[0]>>24));
**   aCksum[1] = (u32)(aCksum[1] + (aCksum[1]>>24));
*/
static void logChecksumBytes(u8 *aByte, int nByte, u32 *aCksum){
  u64 sum1 = aCksum[0];
  u64 sum2 = aCksum[1];
  u32 *a32 = (u32 *)aByte;
  u32 *aEnd = (u32 *)&aByte[nByte];

  assert( (nByte&0x00000003)==0 );

  }

  aCksum[0] = sum1 + (sum1>>24);
  aCksum[1] = sum2 + (sum2>>24);
}

/*
** Argument zPath must be a nul-terminated string containing a path-name.
** This function modifies the string in-place by removing any "./" or "../" 
** elements in the path. For example, the following input:
**
**   "/home/user/plans/good/../evil/./world_domination.txt"
**
** is overwritten with the 'normalized' version:


**
**   "/home/user/plans/evil/world_domination.txt"
*/
static void logNormalizePath(char *zPath){
  int i, j;
  char *z = zPath;
  int n = strlen(z);

  while( n>1 && z[n-1]=='/' ){ n--; }
  for(i=j=0; i<n; i++){
    if( z[i]=='/' ){
      if( z[i+1]=='/' ) continue;
      if( z[i+1]=='.' && i+2<n && z[i+2]=='/' ){
        i += 1;
        continue;
      }
      if( z[i+1]=='.' && i+3<n && z[i+2]=='.' && z[i+3]=='/' ){
        while( j>0 && z[j-1]!='/' ){ j--; }
        if( j>0 ){ j--; }
        i += 2;
        continue;
      }
    }
    z[j++] = z[i];
  }
  z[j] = 0;
}

/*
** Unmap the log-summary mapping and close the file-descriptor. If
** the isTruncate argument is non-zero, truncate the log-summary file
** region to zero bytes.
**
** Regardless of the value of isTruncate, close the file-descriptor
** opened on the log-summary file.
*/
static int logSummaryUnmap(LogSummary *pSummary, int isUnlink){
  int rc = SQLITE_OK;
  if( pSummary->aData ){
    assert( pSummary->fd>0 );
    munmap(pSummary->aData, pSummary->nData);
    pSummary->aData = 0;
    if( isUnlink ){
      char *zFile = sqlite3_mprintf("%s-summary", pSummary->zPath);
      if( !zFile ){
        rc = SQLITE_NOMEM;
      }
      unlink(zFile);
      sqlite3_free(zFile);
    }
  }
  if( pSummary->fd>0 ){
    close(pSummary->fd);
    pSummary->fd = -1;
  }
  return rc;
}




static void logSummaryWriteHdr(LogSummary *pSummary, LogSummaryHdr *pHdr){
  u32 *aHdr = pSummary->aData;                   /* Write header here */
  u32 *aCksum = &aHdr[LOGSUMMARY_HDR_NFIELD];    /* Write header cksum here */

  assert( LOGSUMMARY_HDR_NFIELD==sizeof(LogSummaryHdr)/4 );

  memcpy(aHdr, pHdr, sizeof(LogSummaryHdr));
  aCksum[0] = aCksum[1] = 1;
  logChecksumBytes((u8 *)aHdr, sizeof(LogSummaryHdr), aCksum);
}

/*
** This function encodes a single frame header and writes it to a buffer
** supplied by the caller. A log frame-header is made up of a series of 
** 4-byte big-endian integers, as follows:
**
**     0: Database page size in bytes.
**     4: Page number.
**     8: New database size (for commit frames, otherwise zero).
**    12: Frame checksum 1.
**    16: Frame checksum 2.
*/
static void logEncodeFrame(
  u32 *aCksum,                    /* IN/OUT: Checksum values */
  u32 iPage,                      /* Database page number for frame */
  u32 nTruncate,                  /* New db size (or 0 for non-commit frames) */
  int nData,                      /* Database page size (size of aData[]) */
  u8 *aData,                      /* Pointer to page data (for checksum) */
  u8 *aFrame                      /* OUT: Write encoded frame here */
){
  assert( LOG_FRAME_HDRSIZE==16 );

  sqlite3Put4byte(&aFrame[0], iPage);
  sqlite3Put4byte(&aFrame[4], nTruncate);

  logChecksumBytes(aFrame, 8, aCksum);
  logChecksumBytes(aData, nData, aCksum);

  sqlite3Put4byte(&aFrame[8], aCksum[0]);
  sqlite3Put4byte(&aFrame[12], aCksum[1]);
}

/*
** Return 1 and populate *piPage, *pnTruncate and aCksum if the 
** frame checksum looks Ok. Otherwise return 0.
*/
static int logDecodeFrame(
  u32 *aCksum,                    /* IN/OUT: Checksum values */
  u32 *piPage,                    /* OUT: Database page number for frame */
  u32 *pnTruncate,                /* OUT: New db size (or 0 if not commit) */
  int nData,                      /* Database page size (size of aData[]) */
  u8 *aData,                      /* Pointer to page data (for checksum) */
  u8 *aFrame                      /* Frame data */
){
  assert( LOG_FRAME_HDRSIZE==16 );

  logChecksumBytes(aFrame, 8, aCksum);
  logChecksumBytes(aData, nData, aCksum);

  if( aCksum[0]!=sqlite3Get4byte(&aFrame[8]) 
   || aCksum[1]!=sqlite3Get4byte(&aFrame[12]) 
  ){
    /* Checksum failed. */
    return 0;
  }

  *piPage = sqlite3Get4byte(&aFrame[0]);
  *pnTruncate = sqlite3Get4byte(&aFrame[4]);
  return 1;
}

static void logMergesort8(
  Pgno *aContent,                 /* Pages in log */
  u8 *aBuffer,                    /* Buffer of at least *pnList items to use */
  u8 *aList,                      /* IN/OUT: List to sort */
  int *pnList                     /* IN/OUT: Number of elements in aList[] */
){
  int nList = *pnList;
  if( nList>1 ){
    int nLeft = nList / 2;        /* Elements in left list */
    int nRight = nList - nLeft;   /* Elements in right list */
    u8 *aLeft = aList;            /* Left list */
    u8 *aRight = &aList[nLeft];   /* Right list */
    int iLeft = 0;                /* Current index in aLeft */
    int iRight = 0;               /* Current index in aright */
    int iOut = 0;                 /* Current index in output buffer */

    /* TODO: Change to non-recursive version. */
    logMergesort8(aContent, aBuffer, aLeft, &nLeft);
    logMergesort8(aContent, aBuffer, aRight, &nRight);

    while( iRight<nRight || iLeft<nLeft ){
      u8 logpage;
      Pgno dbpage;

      if( (iLeft<nLeft) 
       && (iRight>=nRight || aContent[aLeft[iLeft]]<aContent[aRight[iRight]])

    }
  }
#endif
}


/*
** Memory map the first nByte bytes of the summary file opened with 
** pSummary->fd at pSummary->aData. If the summary file is smaller than
** nByte bytes in size when this function is called, ftruncate() is
** used to expand it before it is mapped.
**
** It is assumed that an exclusive lock is held on the summary file
** by the caller (to protect the ftruncate()).
*/
static int logSummaryMap(LogSummary *pSummary, int nByte){
  struct stat sStat;
  int rc;
  int fd = pSummary->fd;
  void *pMap;

  assert( pSummary->aData==0 );

  /* If the file is less than nByte bytes in size, cause it to grow. */
  rc = fstat(fd, &sStat);
  if( rc!=0 ) return SQLITE_IOERR;
  if( sStat.st_size<nByte ){
    rc = ftruncate(fd, nByte);
    if( rc!=0 ) return SQLITE_IOERR;
  }else{
    nByte = sStat.st_size;
  }

  /* Map the file. */
  pMap = mmap(0, nByte, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
  if( pMap==MAP_FAILED ){
    return SQLITE_IOERR;
  }
  pSummary->aData = (u32 *)pMap;
  pSummary->nData = nByte/4;

  return SQLITE_OK;
}

/*
** The log-summary file is already mapped to pSummary->aData[], but the
** mapping needs to be resized. Unmap and remap the file so that the mapping 
** is at least nByte bytes in size, or the size of the entire file if it 
** is larger than nByte bytes.
*/
static int logSummaryRemap(LogSummary *pSummary, int nByte){
  int rc;
  sqlite3_mutex_enter(pSummary->mutex);
  munmap(pSummary->aData, pSummary->nData*4);
  pSummary->aData = 0;
  rc = logSummaryMap(pSummary, nByte);
  sqlite3_mutex_leave(pSummary->mutex);
  return rc;
}

/*
** Return the index in the LogSummary.aData array that corresponds to 
** frame iFrame. The log-summary file consists of a header, followed by
** alternating "map" and "index" blocks.
*/
static int logSummaryEntry(u32 iFrame){
  return (
      (LOGSUMMARY_LOCK_OFFSET+LOGSUMMARY_LOCK_RESERVED)/sizeof(u32)
    + (((iFrame-1)>>8)<<6)        /* Indexes that occur before iFrame */
    + iFrame-1                    /* Db page numbers that occur before iFrame */
  );
}











/*




























** Set an entry in the log-summary map to map log frame iFrame to db 
** page iPage. Values are always appended to the log-summary (i.e. the
** value of iFrame is always exactly one more than the value passed to
** the previous call), but that restriction is not enforced or asserted
** here.
*/
static int logSummaryAppend(LogSummary *pSummary, u32 iFrame, u32 iPage){
  u32 iSlot = logSummaryEntry(iFrame);


  while( (iSlot+128)>=pSummary->nData ){
    int rc;
    int nByte = pSummary->nData*4 + LOGSUMMARY_MMAP_INCREMENT;

    /* Unmap and remap the log-summary file. */
    rc = logSummaryRemap(pSummary, nByte);
    if( rc!=SQLITE_OK ){
      return rc;
    }
  }

  /* Set the log-summary entry itself */
  pSummary->aData[iSlot] = iPage;

  /* If the frame number is a multiple of 256 (frames are numbered starting
  ** at 1), build an index of the most recently added 256 frames.
  */
  if( (iFrame&0x000000FF)==0 ){
    int i;                        /* Iterator used while initializing aIndex */
    u32 *aFrame;                  /* Pointer to array of 256 frames */
    int nIndex;                   /* Number of entries in index */
    u8 *aIndex;                   /* 256 bytes to build index in */
    u8 *aTmp;                     /* Scratch space to use while sorting */

    aFrame = &pSummary->aData[iSlot-255];
    aIndex = (u8 *)&pSummary->aData[iSlot+1];
    aTmp = &aIndex[256];

    nIndex = 256;
    for(i=0; i<256; i++) aIndex[i] = (u8)i;
    logMergesort8(aFrame, aTmp, aIndex, &nIndex);
    memset(&aIndex[nIndex], aIndex[nIndex-1], 256-nIndex);
  }

  return SQLITE_OK;
}


/*
** Recover the log-summary by reading the log file. The caller must hold 
** an exclusive lock on the log-summary file.
*/
static int logSummaryRecover(LogSummary *pSummary, sqlite3_file *pFd){
  int rc;                         /* Return Code */
  i64 nSize;                      /* Size of log file */
  LogSummaryHdr hdr;              /* Recovered log-summary header */


  memset(&hdr, 0, sizeof(hdr));

  rc = sqlite3OsFileSize(pFd, &nSize);
  if( rc!=SQLITE_OK ){
    return rc;
  }

  if( nSize>LOG_FRAME_HDRSIZE ){
    u8 aBuf[LOG_FRAME_HDRSIZE];   /* Buffer to load first frame header into */
    u8 *aFrame = 0;               /* Malloc'd buffer to load entire frame */
    int nFrame;                   /* Number of bytes at aFrame */
    u8 *aData;                    /* Pointer to data part of aFrame buffer */
    int iFrame;                   /* Index of last frame read */
    i64 iOffset;                  /* Next offset to read from log file */
    int nPgsz;                    /* Page size according to the log */
    u32 aCksum[2];                /* Running checksum */

    /* Read in the first frame header in the file (to determine the 
    ** database page size).
    */
    rc = sqlite3OsRead(pFd, aBuf, LOG_HDRSIZE, 0);
    if( rc!=SQLITE_OK ){
      return rc;
    }

    /* If the database page size is not a power of two, or is greater than
    ** SQLITE_MAX_PAGE_SIZE, conclude that the log file contains no valid data.
    */
    nPgsz = sqlite3Get4byte(&aBuf[0]);
    if( nPgsz&(nPgsz-1) || nPgsz>SQLITE_MAX_PAGE_SIZE || nPgsz<512 ){
      goto finished;
    }
    aCksum[0] = sqlite3Get4byte(&aBuf[4]);
    aCksum[1] = sqlite3Get4byte(&aBuf[8]);

    /* Malloc a buffer to read frames into. */
    nFrame = nPgsz + LOG_FRAME_HDRSIZE;
    aFrame = (u8 *)sqlite3_malloc(nFrame);
    if( !aFrame ){
      return SQLITE_NOMEM;
    }
    aData = &aFrame[LOG_FRAME_HDRSIZE];

    /* Read all frames from the log file. */
    iFrame = 0;
    for(iOffset=LOG_HDRSIZE; (iOffset+nFrame)<=nSize; iOffset+=nFrame){
      u32 pgno;                   /* Database page number for frame */
      u32 nTruncate;              /* dbsize field from frame header */
      int isValid;                /* True if this frame is valid */

      /* Read and decode the next log frame. */
      rc = sqlite3OsRead(pFd, aFrame, nFrame, iOffset);
      if( rc!=SQLITE_OK ) break;
      isValid = logDecodeFrame(aCksum, &pgno, &nTruncate, nPgsz, aData, aFrame);
      if( !isValid ) break;
      logSummaryAppend(pSummary, ++iFrame, pgno);

      /* If nTruncate is non-zero, this is a commit record. */
      if( nTruncate ){
        hdr.iCheck1 = aCksum[0];
        hdr.iCheck2 = aCksum[1];
        hdr.iLastPg = iFrame;
        hdr.nPage = nTruncate;
        hdr.pgsz = nPgsz;
      }
    }

    sqlite3_free(aFrame);
  }else{
    hdr.iCheck1 = 2;
    hdr.iCheck2 = 3;
  }

finished:
  logSummaryWriteHdr(pSummary, &hdr);
  return rc;
}

/*
** Place, modify or remove a lock on the log-summary file associated 
** with pSummary.
**
** The locked byte-range should be inside the region dedicated to 
** locking. This region of the log-summary file is never read or written.
*/
static int logLockFd(
  LogSummary *pSummary,           /* The log-summary object to lock */
  int iStart,                     /* First byte to lock */
  int nByte,                      /* Number of bytes to lock */
  int op                          /* LOG_UNLOCK, RDLOCK, WRLOCK or WRLOCKW */
){
  int aType[4] = { 
    F_UNLCK,                      /* LOG_UNLOCK */
    F_RDLCK,                      /* LOG_RDLOCK */
    F_WRLCK,                      /* LOG_WRLOCK */
    F_WRLCK                       /* LOG_WRLOCKW */
  };
  int aOp[4] = { 
    F_SETLK,                      /* LOG_UNLOCK */
    F_SETLK,                      /* LOG_RDLOCK */
    F_SETLK,                      /* LOG_WRLOCK */
    F_SETLKW                      /* LOG_WRLOCKW */
  };
  struct flock f;                 /* Locking operation */
  int rc;                         /* Value returned by fcntl() */

  assert( ArraySize(aType)==ArraySize(aOp) );
  assert( op>=0 && op<ArraySize(aType) );
  assert( nByte>0 );
  assert( iStart>=LOGSUMMARY_LOCK_OFFSET 
       && iStart+nByte<=LOGSUMMARY_LOCK_OFFSET+LOGSUMMARY_LOCK_RESERVED
  );
#if defined(SQLITE_DEBUG) && defined(SQLITE_OS_UNIX)
  if( pSummary->aData ) memset(&((u8*)pSummary->aData)[iStart], op, nByte);
#endif

  memset(&f, 0, sizeof(f));
  f.l_type = aType[op];
  f.l_whence = SEEK_SET;
  f.l_start = iStart;
  f.l_len = nByte;
  rc = fcntl(pSummary->fd, aOp[op], &f);
  return (rc==0) ? SQLITE_OK : SQLITE_BUSY;
}

static int logLockRegion(Log *pLog, u32 mRegion, int op){
  LogSummary *pSummary = pLog->pSummary;
  LogLock *p;                     /* Used to iterate through in-process locks */
  u32 mOther;                     /* Locks held by other connections */
  u32 mNew;                       /* New mask for pLog */

  assert( 
       /* Writer lock operations */
          (op==LOG_WRLOCK && mRegion==(LOG_REGION_C|LOG_REGION_D))
       || (op==LOG_UNLOCK && mRegion==(LOG_REGION_C|LOG_REGION_D))

       /* Normal reader lock operations */
       || (op==LOG_RDLOCK && mRegion==(LOG_REGION_A|LOG_REGION_B))
       || (op==LOG_UNLOCK && mRegion==(LOG_REGION_A))
       || (op==LOG_UNLOCK && mRegion==(LOG_REGION_B))

       /* Region D reader lock operations */
       || (op==LOG_RDLOCK && mRegion==(LOG_REGION_D))
       || (op==LOG_RDLOCK && mRegion==(LOG_REGION_A))
       || (op==LOG_UNLOCK && mRegion==(LOG_REGION_D))

       /* Checkpointer lock operations */
       || (op==LOG_WRLOCK && mRegion==(LOG_REGION_B|LOG_REGION_C))
       || (op==LOG_WRLOCK && mRegion==(LOG_REGION_A))
       || (op==LOG_UNLOCK && mRegion==(LOG_REGION_B|LOG_REGION_C))
       || (op==LOG_UNLOCK && mRegion==(LOG_REGION_A|LOG_REGION_B|LOG_REGION_C))
  );

  /* Assert that a connection never tries to go from an EXCLUSIVE to a 
  ** SHARED lock on a region. Moving from SHARED to EXCLUSIVE sometimes
  ** happens though (when a region D reader upgrades to a writer).
  */
  assert( op!=LOG_RDLOCK || 0==(pLog->lock.mLock & (mRegion<<8)) );

  sqlite3_mutex_enter(pSummary->mutex);

  /* Calculate a mask of logs held by all connections in this process apart
  ** from this one. The least significant byte of the mask contains a mask
  ** of the SHARED logs held. The next least significant byte of the mask
  ** indicates the EXCLUSIVE locks held. For example, to test if some other
  ** connection is holding a SHARED lock on region A, or an EXCLUSIVE lock
  ** on region C, do:
  **
  **   hasSharedOnA    = (mOther & (LOG_REGION_A<<0));
  **   hasExclusiveOnC = (mOther & (LOG_REGION_C<<8));
  **
  ** In all masks, if the bit in the EXCLUSIVE byte mask is set, so is the 
  ** corresponding bit in the SHARED mask.
  */
  mOther = 0;
  for(p=pSummary->pLock; p; p=p->pNext){
    assert( (p->mLock & (p->mLock<<8))==(p->mLock&0x0000FF00) );
    if( p!=&pLog->lock ){
      mOther |= p->mLock;
    }
  }

  /* If this call is to lock a region (not to unlock one), test if locks held
  ** by any other connection in this process prevent the new locks from
  ** begin granted. If so, exit the summary mutex and return SQLITE_BUSY.
  */
  if( op && (mOther & (mRegion << (op==LOG_RDLOCK ? 8 : 0))) ){
    sqlite3_mutex_leave(pSummary->mutex);
    return SQLITE_BUSY;
  }

  /* Figure out the new log mask for this connection. */
  switch( op ){
    case LOG_UNLOCK: 
      mNew = (pLog->lock.mLock & ~(mRegion|(mRegion<<8)));
      break;
    case LOG_RDLOCK:
      mNew = (pLog->lock.mLock | mRegion);
      break;
    default:
      assert( op==LOG_WRLOCK );
      mNew = (pLog->lock.mLock | (mRegion<<8) | mRegion);
      break;
  }

  /* Now modify the locks held on the log-summary file descriptor. This
  ** file descriptor is shared by all log connections in this process. 
  ** Therefore:
  **
  **   + If one or more log connections in this process hold a SHARED lock
  **     on a region, the file-descriptor should hold a SHARED lock on
  **     the file region.
  **
  **   + If a log connection in this process holds an EXCLUSIVE lock on a
  **     region, the file-descriptor should also hold an EXCLUSIVE lock on
  **     the region in question.
  **
  ** If this is an LOG_UNLOCK operation, only regions for which no other
  ** connection holds a lock should actually be unlocked. And if this
  ** is a LOG_RDLOCK operation and other connections already hold all
  ** the required SHARED locks, then no system call is required.
  */
  if( op==LOG_UNLOCK ){
    mRegion = (mRegion & ~mOther);
  }
  if( (op==LOG_WRLOCK)
   || (op==LOG_UNLOCK && mRegion) 
   || (op==LOG_RDLOCK && (mOther&mRegion)!=mRegion)
  ){
    struct LockMap {
      int iStart;                 /* Byte offset to start locking operation */
      int iLen;                   /* Length field for locking operation */
    } aMap[] = {
      /* 0000 */ {0, 0},                    /* 0001 */ {3+LOG_LOCK_REGION, 1}, 
      /* 0010 */ {2+LOG_LOCK_REGION, 1},    /* 0011 */ {2+LOG_LOCK_REGION, 2},
      /* 0100 */ {1+LOG_LOCK_REGION, 1},    /* 0101 */ {0, 0}, 
      /* 0110 */ {1+LOG_LOCK_REGION, 2},    /* 0111 */ {1+LOG_LOCK_REGION, 3},
      /* 1000 */ {0+LOG_LOCK_REGION, 1},    /* 1001 */ {0, 0}, 
      /* 1010 */ {0, 0},                    /* 1011 */ {0, 0},
      /* 1100 */ {0+LOG_LOCK_REGION, 2},    /* 1101 */ {0, 0}, 
      /* 1110 */ {0, 0},                    /* 1111 */ {0, 0}
    };
    int rc;                       /* Return code of logLockFd() */

    assert( mRegion<ArraySize(aMap) && aMap[mRegion].iStart!=0 );

    rc = logLockFd(pSummary, aMap[mRegion].iStart, aMap[mRegion].iLen, op);
    if( rc!=0 ){
      sqlite3_mutex_leave(pSummary->mutex);
      return rc;
    }
  }

  pLog->lock.mLock = mNew;
  sqlite3_mutex_leave(pSummary->mutex);
  return SQLITE_OK;
}

/*
** Lock the DMH region, either with an EXCLUSIVE or SHARED lock. This
** function is never called with LOG_UNLOCK - the only way the DMH region
** is every completely unlocked is by by closing the file descriptor.
*/
static int logLockDMH(LogSummary *pSummary, int eLock){
  assert( sqlite3_mutex_held(pSummary->mutex) );
  assert( eLock==LOG_RDLOCK || eLock==LOG_WRLOCK );
  return logLockFd(pSummary, LOG_LOCK_DMH, 1, eLock);
}

/*
** Lock (or unlock) the MUTEX region. It is always locked using an
** EXCLUSIVE, blocking lock.
*/
static int logLockMutex(LogSummary *pSummary, int eLock){
  assert( sqlite3_mutex_held(pSummary->mutex) );
  assert( eLock==LOG_WRLOCKW || eLock==LOG_UNLOCK );
  logLockFd(pSummary, LOG_LOCK_MUTEX, 1, eLock);
  return SQLITE_OK;
}

/*
** This function intializes the connection to the log-summary identified
** by struct pSummary.
*/
static int logSummaryInit(
  LogSummary *pSummary,           /* Log summary object to initialize */
  sqlite3_file *pFd               /* File descriptor open on log file */
){
  int rc;                         /* Return Code */
  char *zFile;                    /* File name for summary file */

  assert( pSummary->fd<0 );
  assert( pSummary->aData==0 );
  assert( pSummary->nRef>0 );
  assert( pSummary->zPath );

  /* Open a file descriptor on the summary file. */
  zFile = sqlite3_mprintf("%s-summary", pSummary->zPath);
  if( !zFile ){
    return SQLITE_NOMEM;
  }
  pSummary->fd = open(zFile, O_RDWR|O_CREAT, S_IWUSR|S_IRUSR);
  sqlite3_free(zFile);
  if( pSummary->fd<0 ){
    return SQLITE_IOERR;
  }

  /* Grab an exclusive lock the summary file. Then mmap() it. 
  **
  ** TODO: This code needs to be enhanced to support a growable mapping. 
  ** For now, just make the mapping very large to start with. The 
  ** pages should not be allocated until they are first accessed anyhow,
  ** so using a large mapping consumes no more resources than a smaller
  ** one would.
  */
  assert( sqlite3_mutex_held(pSummary->mutex) );
  rc = logLockMutex(pSummary, LOG_WRLOCKW);
  if( rc!=SQLITE_OK ) return rc;
  rc = logSummaryMap(pSummary, LOGSUMMARY_MMAP_INCREMENT);
  if( rc!=SQLITE_OK ) goto out;

  /* Try to obtain an EXCLUSIVE lock on the dead-mans-hand region. If this
  ** is possible, the contents of the log-summary file (if any) may not
  ** be trusted. Zero the log-summary header before continuing.
  */
  rc = logLockDMH(pSummary, LOG_WRLOCK);
  if( rc==SQLITE_OK ){
    memset(pSummary->aData, 0, (LOGSUMMARY_HDR_NFIELD+2)*sizeof(u32) );
  }
  rc = logLockDMH(pSummary, LOG_RDLOCK);
  if( rc!=SQLITE_OK ){
    rc = SQLITE_IOERR;
  }

 out:
  logLockMutex(pSummary, LOG_UNLOCK);
  return rc;
}

/* 
** Open a connection to the log file associated with database zDb. The
** database file does not actually have to exist. zDb is used only to
** figure out the name of the log file to open. If the log file does not 
** exist it is created by this call.
**
** A SHARED lock should be held on the database file when this function
** is called. The purpose of this SHARED lock is to prevent any other
** client from unlinking the log or log-summary file. If another process
** were to do this just after this client opened one of these files, the
** system would be badly broken.
*/
int sqlite3WalOpen(
  sqlite3_vfs *pVfs,              /* vfs module to open log file with */
  const char *zDb,                /* Name of database file */
  Log **ppLog                     /* OUT: Allocated Log handle */
){
  int rc = SQLITE_OK;             /* Return Code */
  Log *pRet;                      /* Object to allocate and return */
  LogSummary *pSummary = 0;       /* Summary object */
  sqlite3_mutex *mutex = 0;       /* LOG_SUMMARY_MUTEX mutex */
  int flags;                      /* Flags passed to OsOpen() */
  char *zWal = 0;                 /* Path to WAL file */
  int nWal;                       /* Length of zWal in bytes */

  assert( zDb );


  /* Allocate an instance of struct Log to return. */
  *ppLog = 0;

  pRet = (Log *)sqlite3MallocZero(sizeof(Log) + pVfs->szOsFile);
  if( !pRet ) goto out;
  pRet->pVfs = pVfs;
  pRet->pFd = (sqlite3_file *)&pRet[1];





  /* Normalize the path name. */
  zWal = sqlite3_mprintf("%s-wal", zDb);
  if( !zWal ) goto out;
  logNormalizePath(zWal);
  flags = (SQLITE_OPEN_READWRITE|SQLITE_OPEN_CREATE|SQLITE_OPEN_MAIN_JOURNAL);
  nWal = sqlite3Strlen30(zWal);

  /* Enter the mutex that protects the linked-list of LogSummary structures */
  if( sqlite3GlobalConfig.bCoreMutex ){
    mutex = sqlite3_mutex_alloc(LOG_SUMMARY_MUTEX);
  }
  sqlite3_mutex_enter(mutex);

  /* Search for an existing log summary object in the linked list. If one 
  ** cannot be found, allocate and initialize a new object.
  */
  for(pSummary=pLogSummary; pSummary; pSummary=pSummary->pNext){
    int nPath = sqlite3Strlen30(pSummary->zPath);
    if( nWal==nPath && 0==memcmp(pSummary->zPath, zWal, nPath) ) break;
  }
  if( !pSummary ){
    int nByte = sizeof(LogSummary) + nWal + 1;
    pSummary = (LogSummary *)sqlite3MallocZero(nByte);
    if( !pSummary ){
      rc = SQLITE_NOMEM;
      goto out;
    }
    if( sqlite3GlobalConfig.bCoreMutex ){
      pSummary->mutex = sqlite3_mutex_alloc(SQLITE_MUTEX_RECURSIVE);
    }
    pSummary->zPath = (char *)&pSummary[1];
    pSummary->fd = -1;
    memcpy(pSummary->zPath, zWal, nWal);
    pSummary->pNext = pLogSummary;
    pLogSummary = pSummary;
  }
  pSummary->nRef++;
  pRet->pSummary = pSummary;

  /* Exit the mutex protecting the linked-list of LogSummary objects. */
  sqlite3_mutex_leave(mutex);
  mutex = 0;

  /* Open file handle on the log file. */
  rc = sqlite3OsOpen(pVfs, pSummary->zPath, pRet->pFd, flags, &flags);
  if( rc!=SQLITE_OK ) goto out;

  /* Object pSummary is shared between all connections to the database made
  ** by this process. So at this point it may or may not be connected to
  ** the log-summary. If it is not, connect it.
  */
  sqlite3_mutex_enter(pSummary->mutex);
  mutex = pSummary->mutex;
  if( pSummary->fd<0 ){
    rc = logSummaryInit(pSummary, pRet->pFd);
  }

  pRet->lock.pNext = pSummary->pLock;
  pSummary->pLock = &pRet->lock;

 out:
  sqlite3_mutex_leave(mutex);
  sqlite3_free(zWal);
  if( rc!=SQLITE_OK ){
    assert(0);
    if( pRet ){

      sqlite3OsClose(pRet->pFd);
      sqlite3_free(pRet);
    }
    assert( !pSummary || pSummary->nRef==0 );
    sqlite3_free(pSummary);
  }
  *ppLog = pRet;
  return rc;
}

static int logIteratorNext(
  LogIterator *p,               /* Iterator */
  u32 *piPage,                    /* OUT: Next db page to write */
  u32 *piFrame                    /* OUT: Log frame to read from */
){
  u32 iMin = *piPage;
  u32 iRet = 0xFFFFFFFF;
  int i;
  int nBlock = p->nFinal;

  for(i=p->nSegment-1; i>=0; i--){
    struct LogSegment *pSegment = &p->aSegment[i];
    while( pSegment->iNext<nBlock ){
      u32 iPg = pSegment->aDbPage[pSegment->aIndex[pSegment->iNext]];
      if( iPg>iMin ){
        if( iPg<iRet ){
          iRet = iPg;
          *piFrame = i*256 + 1 + pSegment->aIndex[pSegment->iNext];
        }
        break;
      }
      pSegment->iNext++;
    }

    nBlock = 256;
  }

  *piPage = iRet;
  return (iRet==0xFFFFFFFF);
}

static LogIterator *logIteratorInit(Log *pLog){
  u32 *aData = pLog->pSummary->aData;
  LogIterator *p;                 /* Return value */
  int nSegment;                   /* Number of segments to merge */
  u32 iLast;                      /* Last frame in log */
  int nByte;                      /* Number of bytes to allocate */
  int i;                          /* Iterator variable */
  int nFinal;                     /* Number of unindexed entries */
  struct LogSegment *pFinal;      /* Final (unindexed) segment */
  u8 *aTmp;                       /* Temp space used by merge-sort */



  iLast = pLog->hdr.iLastPg;
  nSegment = (iLast >> 8) + 1;
  nFinal = (iLast & 0x000000FF);

  nByte = sizeof(LogIterator) + (nSegment-1)*sizeof(struct LogSegment) + 512;
  p = (LogIterator *)sqlite3_malloc(nByte);
  if( p ){
    memset(p, 0, nByte);
    p->nSegment = nSegment;
    p->nFinal = nFinal;
  }

  for(i=0; i<nSegment-1; i++){
    p->aSegment[i].aDbPage = &aData[logSummaryEntry(i*256+1)];
    p->aSegment[i].aIndex = (u8 *)&aData[logSummaryEntry(i*256+1)+256];
  }
  pFinal = &p->aSegment[nSegment-1];

  pFinal->aDbPage = &aData[logSummaryEntry((nSegment-1)*256+1)];
  pFinal->aIndex = (u8 *)&pFinal[1];
  aTmp = &pFinal->aIndex[256];
  for(i=0; i<nFinal; i++){
    pFinal->aIndex[i] = i;
  }
  logMergesort8(pFinal->aDbPage, aTmp, pFinal->aIndex, &nFinal);
  p->nFinal = nFinal;

  return p;
}

/* 
** Free a log iterator allocated by logIteratorInit().
*/
static void logIteratorFree(LogIterator *p){
  sqlite3_free(p);
}

/*
** Checkpoint the contents of the log file.
*/
static int logCheckpoint(
  Log *pLog,                      /* Log connection */
  sqlite3_file *pFd,              /* File descriptor open on db file */
  int sync_flags,                 /* Flags for OsSync() (or 0) */
  u8 *zBuf                        /* Temporary buffer to use */
){
  int rc;                         /* Return code */
  int pgsz = pLog->hdr.pgsz;      /* Database page-size */
  LogIterator *pIter = 0;         /* Log iterator context */
  u32 iDbpage = 0;                /* Next database page to write */
  u32 iFrame = 0;                 /* Log frame containing data for iDbpage */

  if( pLog->hdr.iLastPg==0 ){
    return SQLITE_OK;
  }

  /* Allocate the iterator */
  pIter = logIteratorInit(pLog);
  if( !pIter ) return SQLITE_NOMEM;

  /* Sync the log file to disk */
  if( sync_flags ){
    rc = sqlite3OsSync(pLog->pFd, sync_flags);
    if( rc!=SQLITE_OK ) goto out;
  }

  /* Iterate through the contents of the log, copying data to the db file. */
  while( 0==logIteratorNext(pIter, &iDbpage, &iFrame) ){
    rc = sqlite3OsRead(pLog->pFd, zBuf, pgsz, 
        logFrameOffset(iFrame, pgsz) + LOG_FRAME_HDRSIZE
    );
    if( rc!=SQLITE_OK ) goto out;
    rc = sqlite3OsWrite(pFd, zBuf, pgsz, (iDbpage-1)*pgsz);
    if( rc!=SQLITE_OK ) goto out;
  }

  /* Truncate the database file */
  rc = sqlite3OsTruncate(pFd, ((i64)pLog->hdr.nPage*(i64)pgsz));
  if( rc!=SQLITE_OK ) goto out;

  /* Sync the database file. If successful, update the log-summary. */
  if( sync_flags ){
    rc = sqlite3OsSync(pFd, sync_flags);
    if( rc!=SQLITE_OK ) goto out;
  }
  pLog->hdr.iLastPg = 0;
  pLog->hdr.iCheck1 = 2;
  pLog->hdr.iCheck2 = 3;
  logSummaryWriteHdr(pLog->pSummary, &pLog->hdr);

  /* TODO: If a crash occurs and the current log is copied into the 
  ** database there is no problem. However, if a crash occurs while
  ** writing the next transaction into the start of the log, such that:
  **
  **   * The first transaction currently in the log is left intact, but
  **   * The second (or subsequent) transaction is damaged,
  **
  ** then the database could become corrupt.
  **
  ** The easiest thing to do would be to write and sync a dummy header
  ** into the log at this point. Unfortunately, that turns out to be
  ** an unwelcome performance hit. Alternatives are...
  */
#if 0 
  memset(zBuf, 0, LOG_FRAME_HDRSIZE);
  rc = sqlite3OsWrite(pLog->pFd, zBuf, LOG_FRAME_HDRSIZE, 0);
  if( rc!=SQLITE_OK ) goto out;
  rc = sqlite3OsSync(pLog->pFd, pLog->sync_flags);
#endif

 out:
  logIteratorFree(pIter);
  return rc;
}

/*
** Close a connection to a log file.
*/
int sqlite3WalClose(
  Log *pLog,                      /* Log to close */
  sqlite3_file *pFd,              /* Database file */
  int sync_flags,                 /* Flags to pass to OsSync() (or 0) */
  u8 *zBuf                        /* Buffer of at least page-size bytes */
){
  int rc = SQLITE_OK;
  if( pLog ){
    LogLock **ppL;
    LogSummary *pSummary = pLog->pSummary;
    sqlite3_mutex *mutex = 0;

    sqlite3_mutex_enter(pSummary->mutex);
    for(ppL=&pSummary->pLock; *ppL!=&pLog->lock; ppL=&(*ppL)->pNext);
    *ppL = pLog->lock.pNext;
    sqlite3_mutex_leave(pSummary->mutex);

    if( sqlite3GlobalConfig.bCoreMutex ){
      mutex = sqlite3_mutex_alloc(LOG_SUMMARY_MUTEX);
    }
    sqlite3_mutex_enter(mutex);

    /* Decrement the reference count on the log summary. If this is the last
    ** reference to the log summary object in this process, the object will
    ** be freed. If this is also the last connection to the database, then
    ** checkpoint the database and truncate the log and log-summary files
    ** to zero bytes in size.
    **/
    pSummary->nRef--;
    if( pSummary->nRef==0 ){
      int rc;
      LogSummary **pp;
      for(pp=&pLogSummary; *pp!=pSummary; pp=&(*pp)->pNext);
      *pp = (*pp)->pNext;

      sqlite3_mutex_leave(mutex);

      rc = sqlite3OsLock(pFd, SQLITE_LOCK_EXCLUSIVE);
      if( rc==SQLITE_OK ){

        /* This is the last connection to the database (including other
        ** processes). Do three things:
        **
        **   1. Checkpoint the db.
        **   2. Truncate the log file.
        **   3. Unlink the log-summary file.
        */
        rc = logCheckpoint(pLog, pFd, sync_flags, zBuf);
        if( rc==SQLITE_OK ){
          rc = sqlite3OsDelete(pLog->pVfs, pSummary->zPath, 0);
        }

        logSummaryUnmap(pSummary, 1);
      }else{
        if( rc==SQLITE_BUSY ){
          rc = SQLITE_OK;
        }
        logSummaryUnmap(pSummary, 0);
      }

      sqlite3_mutex_free(pSummary->mutex);
      sqlite3_free(pSummary);
    }else{
      sqlite3_mutex_leave(mutex);
    }

    /* Close the connection to the log file and free the Log handle. */
    sqlite3OsClose(pLog->pFd);
    sqlite3_free(pLog);
  }
  return rc;
}

/*
** Enter and leave the log-summary mutex. In this context, entering the
** log-summary mutex means:
**
**   1. Obtaining mutex pLog->pSummary->mutex, and
**   2. Taking an exclusive lock on the log-summary file.
**
** i.e. this mutex locks out other processes as well as other threads
** hosted in this address space.
*/
static int logEnterMutex(Log *pLog){
  LogSummary *pSummary = pLog->pSummary;
  int rc;

  sqlite3_mutex_enter(pSummary->mutex);
  rc = logLockMutex(pSummary, LOG_WRLOCKW);
  if( rc!=SQLITE_OK ){
    sqlite3_mutex_leave(pSummary->mutex);
  }
  return rc;
}
static void logLeaveMutex(Log *pLog){
  LogSummary *pSummary = pLog->pSummary;
  logLockMutex(pSummary, LOG_UNLOCK);
  sqlite3_mutex_leave(pSummary->mutex);
}

/*
** Try to read the log-summary header. Attempt to verify the header
** checksum. If the checksum can be verified, copy the log-summary
** header into structure pLog->hdr. If the contents of pLog->hdr are
** modified by this and pChanged is not NULL, set *pChanged to 1. 
** Otherwise leave *pChanged unmodified.
**
** If the checksum cannot be verified return SQLITE_ERROR.
*/
int logSummaryTryHdr(Log *pLog, int *pChanged){
  u32 aCksum[2] = {1, 1};
  u32 aHdr[LOGSUMMARY_HDR_NFIELD+2];

  /* Read the header. The caller may or may not have locked the log-summary
  ** file, meaning it is possible that an inconsistent snapshot is read
  ** from the file. If this happens, return SQLITE_ERROR. The caller will
  ** retry. Or, if the caller has already locked the file and the header
  ** still looks inconsistent, it will run recovery.
  */
  memcpy(aHdr, pLog->pSummary->aData, sizeof(aHdr));
  logChecksumBytes((u8*)aHdr, sizeof(u32)*LOGSUMMARY_HDR_NFIELD, aCksum);
  if( aCksum[0]!=aHdr[LOGSUMMARY_HDR_NFIELD]
   || aCksum[1]!=aHdr[LOGSUMMARY_HDR_NFIELD+1]
  ){
    return SQLITE_ERROR;
  }

  if( memcmp(&pLog->hdr, aHdr, sizeof(LogSummaryHdr)) ){
    if( pChanged ){
      *pChanged = 1;
    }
    memcpy(&pLog->hdr, aHdr, sizeof(LogSummaryHdr));
  }
  return SQLITE_OK;
}

/*
** Read the log-summary header from the log-summary file into structure 
** pLog->hdr. If attempting to verify the header checksum fails, try
** to recover the log before returning.
**
** If the log-summary header is successfully read, return SQLITE_OK. 
** Otherwise an SQLite error code.
*/
int logSummaryReadHdr(Log *pLog, int *pChanged){
  int rc;




  /* First try to read the header without a lock. Verify the checksum
  ** before returning. This will almost always work.  
  **
  ** TODO: Doing this causes a race-condition with the code that resizes
  ** the mapping. Unless Log.pSummary->mutex is held, it is possible that 
  ** LogSummary.aData is invalid.
  */
#if 0
  if( SQLITE_OK==logSummaryTryHdr(pLog, pChanged) ){
    return SQLITE_OK;
  }
#endif

  /* If the first attempt to read the header failed, lock the log-summary
  ** file and try again. If the header checksum verification fails this
  ** time as well, run log recovery.
  */
  if( SQLITE_OK==(rc = logEnterMutex(pLog)) ){
    if( SQLITE_OK!=logSummaryTryHdr(pLog, pChanged) ){
      if( pChanged ){
        *pChanged = 1;
      }
      rc = logSummaryRecover(pLog->pSummary, pLog->pFd);
      if( rc==SQLITE_OK ){
        rc = logSummaryTryHdr(pLog, 0);
      }
    }
    logLeaveMutex(pLog);
  }

  return rc;
}

/*
** Lock a snapshot.
**
** If this call obtains a new read-lock and the database contents have been
** modified since the most recent call to LogCloseSnapshot() on this Log
** connection, then *pChanged is set to 1 before returning. Otherwise, it 
** is left unmodified. This is used by the pager layer to determine whether 
** or not any cached pages may be safely reused.
*/
int sqlite3WalOpenSnapshot(Log *pLog, int *pChanged){
  int rc = SQLITE_OK;
  if( pLog->isLocked==0 ){
    int nAttempt;

    /* Obtain a snapshot-lock on the log-summary file. The procedure
    ** for obtaining the snapshot log is:
    **
    **    1. Attempt a SHARED lock on regions A and B.
    **    2a. If step 1 is successful, drop the lock on region B.
    **    2b. If step 1 is unsuccessful, attempt a SHARED lock on region D.
    **    3. Repeat the above until the lock attempt in step 1 or 2b is 
    **       successful.
    **
    ** If neither of the locks can be obtained after 5 tries, presumably
    ** something is wrong (i.e. a process not following the locking protocol). 
    ** Return an error code in this case.
    */
    rc = SQLITE_BUSY;
    for(nAttempt=0; nAttempt<5 && rc==SQLITE_BUSY; nAttempt++){
      rc = logLockRegion(pLog, LOG_REGION_A|LOG_REGION_B, LOG_RDLOCK);
      if( rc==SQLITE_BUSY ){
        rc = logLockRegion(pLog, LOG_REGION_D, LOG_RDLOCK);
        if( rc==SQLITE_OK ) pLog->isLocked = LOG_REGION_D;
      }else{
        logLockRegion(pLog, LOG_REGION_B, LOG_UNLOCK);
        pLog->isLocked = LOG_REGION_A;
      }
    }
    if( rc!=SQLITE_OK ){
      return rc;
    }

    rc = logSummaryReadHdr(pLog, pChanged);
    if( rc!=SQLITE_OK ){
      /* An error occured while attempting log recovery. */
      sqlite3WalCloseSnapshot(pLog);
    }else{
      /* Check if the mapping needs to grow. */
      LogSummary *pSummary = pLog->pSummary;

      if( pLog->hdr.iLastPg 
       && logSummaryEntry(pLog->hdr.iLastPg)>=pSummary->nData 
      ){
        rc = logSummaryRemap(pSummary, 0);
        assert( rc || logSummaryEntry(pLog->hdr.iLastPg)<pSummary->nData );
      }
    }
  }
  return rc;
}

/*
** Unlock the current snapshot.
*/
void sqlite3WalCloseSnapshot(Log *pLog){
  if( pLog->isLocked ){
    assert( pLog->isLocked==LOG_REGION_A || pLog->isLocked==LOG_REGION_D );
    logLockRegion(pLog, pLog->isLocked, LOG_UNLOCK);
  }
  pLog->isLocked = 0;
}

/*
** Read a page from the log, if it is present. 
*/
int sqlite3WalRead(Log *pLog, Pgno pgno, int *pInLog, u8 *pOut){
  LogSummary *pSummary = pLog->pSummary;
  u32 iRead = 0;
  u32 *aData; 
  int iFrame = (pLog->hdr.iLastPg & 0xFFFFFF00);

  assert( pLog->isLocked );
  sqlite3_mutex_enter(pSummary->mutex);


  /* Do a linear search of the unindexed block of page-numbers (if any) 
  ** at the end of the log-summary. An alternative to this would be to
  ** build an index in private memory each time a read transaction is
  ** opened on a new snapshot.
  */
  aData = pSummary->aData;
  if( pLog->hdr.iLastPg ){
    u32 *pi = &aData[logSummaryEntry(pLog->hdr.iLastPg)];
    u32 *piStop = pi - (pLog->hdr.iLastPg & 0xFF);
    while( *pi!=pgno && pi!=piStop ) pi--;
    if( pi!=piStop ){
      iRead = (pi-piStop) + iFrame;
    }
  }
  assert( iRead==0 || aData[logSummaryEntry(iRead)]==pgno );

  while( iRead==0 && iFrame>0 ){
    int iLow = 0;
    int iHigh = 255;
    u32 *aFrame;
    u8 *aIndex;

    iFrame -= 256;
    aFrame = &aData[logSummaryEntry(iFrame+1)];
    aIndex = (u8 *)&aFrame[256];

    while( iLow<=iHigh ){
      int iTest = (iLow+iHigh)>>1;
      u32 iPg = aFrame[aIndex[iTest]];

      if( iPg==pgno ){
        iRead = iFrame + 1 + aIndex[iTest];
        break;
      }
      else if( iPg<pgno ){
        iLow = iTest+1;
      }else{
        iHigh = iTest-1;
      }
    }
  }
  assert( iRead==0 || aData[logSummaryEntry(iRead)]==pgno );

  sqlite3_mutex_leave(pLog->pSummary->mutex);

  /* If iRead is non-zero, then it is the log frame number that contains the
  ** required page. Read and return data from the log file.
  */
  if( iRead ){
    i64 iOffset = logFrameOffset(iRead, pLog->hdr.pgsz) + LOG_FRAME_HDRSIZE;
    *pInLog = 1;
    return sqlite3OsRead(pLog->pFd, pOut, pLog->hdr.pgsz, iOffset);
  }

  *pInLog = 0;
  return SQLITE_OK;
}


/* 
** Set *pPgno to the size of the database file (or zero, if unknown).
*/
void sqlite3WalDbsize(Log *pLog, Pgno *pPgno){
  assert( pLog->isLocked );

  *pPgno = pLog->hdr.nPage;
}

/* 
** This function returns SQLITE_OK if the caller may write to the database.
** Otherwise, if the caller is operating on a snapshot that has already
** been overwritten by another writer, SQLITE_BUSY is returned.
*/
int sqlite3WalWriteLock(Log *pLog, int op){
  assert( pLog->isLocked );
  if( op ){

    /* Obtain the writer lock */
    int rc = logLockRegion(pLog, LOG_REGION_C|LOG_REGION_D, LOG_WRLOCK);
    if( rc!=SQLITE_OK ){
      return rc;
    }

    /* If this is connection is a region D reader, then the SHARED lock on 
    ** region D has just been upgraded to EXCLUSIVE. But no lock at all is 
    ** held on region A. This means that if the write-transaction is committed
    ** and this connection downgrades to a reader, it will be left with no
    ** lock at all. And so its snapshot could get clobbered by a checkpoint
    ** operation. 
    **
    ** To stop this from happening, grab a SHARED lock on region A now.
    ** This should always be successful, as the only time a client holds
    ** an EXCLUSIVE lock on region A, it must also be holding an EXCLUSIVE
    ** lock on region C (a checkpointer does this). This is not possible,
    ** as this connection currently has the EXCLUSIVE lock on region C.
    */
    if( pLog->isLocked==LOG_REGION_D ){
      logLockRegion(pLog, LOG_REGION_A, LOG_RDLOCK);
      pLog->isLocked = LOG_REGION_A;
    }

    /* If this connection is not reading the most recent database snapshot,
    ** it is not possible to write to the database. In this case release
    ** the write locks and return SQLITE_BUSY.
    */
    if( memcmp(&pLog->hdr, pLog->pSummary->aData, sizeof(pLog->hdr)) ){
      logLockRegion(pLog, LOG_REGION_C|LOG_REGION_D, LOG_UNLOCK);
      return SQLITE_BUSY;
    }
    pLog->isWriteLocked = 1;

  }else if( pLog->isWriteLocked ){
    logLockRegion(pLog, LOG_REGION_C|LOG_REGION_D, LOG_UNLOCK);
    memcpy(&pLog->hdr, pLog->pSummary->aData, sizeof(pLog->hdr));
    pLog->isWriteLocked = 0;
  }
  return SQLITE_OK;
}

/*
** The log handle passed to this function must be holding the write-lock.
**
** If any data has been written (but not committed) to the log file, this
** function moves the write-pointer back to the start of the transaction.
**
** Additionally, the callback function is invoked for each frame written
** to the log since the start of the transaction. If the callback returns
** other than SQLITE_OK, it is not invoked again and the error code is
** returned to the caller.
**
** Otherwise, if the callback function does not return an error, this
** function returns SQLITE_OK.
*/
int sqlite3WalUndo(Log *pLog, int (*xUndo)(void *, Pgno), void *pUndoCtx){
  int rc = SQLITE_OK;
  Pgno iMax = pLog->hdr.iLastPg;
  Pgno iFrame;

  assert( pLog->isWriteLocked );
  logSummaryReadHdr(pLog, 0);
  for(iFrame=pLog->hdr.iLastPg+1; iFrame<=iMax && rc==SQLITE_OK; iFrame++){
    rc = xUndo(pUndoCtx, pLog->pSummary->aData[logSummaryEntry(iFrame)]);
  }

  return rc;
}




u32 sqlite3WalSavepoint(Log *pLog){
  assert( pLog->isWriteLocked );
  return pLog->hdr.iLastPg;
}




int sqlite3WalSavepointUndo(Log *pLog, u32 iFrame){
  int rc = SQLITE_OK;
  u8 aCksum[8];
  assert( pLog->isWriteLocked );

  pLog->hdr.iLastPg = iFrame;
  if( iFrame>0 ){
    i64 iOffset = logFrameOffset(iFrame, pLog->hdr.pgsz) + sizeof(u32)*2;
    rc = sqlite3OsRead(pLog->pFd, aCksum, sizeof(aCksum), iOffset);
    pLog->hdr.iCheck1 = sqlite3Get4byte(&aCksum[0]);
    pLog->hdr.iCheck2 = sqlite3Get4byte(&aCksum[4]);
  }

  return rc;
}

/* 
** Return true if data has been written but not committed to the log file. 
*/
int sqlite3WalDirty(Log *pLog){
  assert( pLog->isWriteLocked );
  return( pLog->hdr.iLastPg!=((LogSummaryHdr*)pLog->pSummary->aData)->iLastPg );
}

/* 
** Write a set of frames to the log. The caller must hold the write-lock
** on the log file (obtained using sqlite3WalWriteLock()).
*/
int sqlite3WalFrames(
  Log *pLog,                      /* Log handle to write to */
  int nPgsz,                      /* Database page-size in bytes */
  PgHdr *pList,                   /* List of dirty pages to write */
  Pgno nTruncate,                 /* Database size after this commit */
  int isCommit,                   /* True if this is a commit */
  int sync_flags                  /* Flags to pass to OsSync() (or 0) */
){
  int rc;                         /* Used to catch return codes */
  u32 iFrame;                     /* Next frame address */
  u8 aFrame[LOG_FRAME_HDRSIZE];   /* Buffer to assemble frame-header in */
  PgHdr *p;                       /* Iterator to run through pList with. */
  u32 aCksum[2];                  /* Checksums */
  PgHdr *pLast;                   /* Last frame in list */
  int nLast = 0;                  /* Number of extra copies of last page */

  assert( LOG_FRAME_HDRSIZE==(4 * 2 + 2*sizeof(u32)) );
  assert( pList );


  /* If this is the first frame written into the log, write the log 
  ** header to the start of the log file. See comments at the top of
  ** this file for a description of the log-header format.
  */
  assert( LOG_FRAME_HDRSIZE>=LOG_HDRSIZE );
  iFrame = pLog->hdr.iLastPg;
  if( iFrame==0 ){
    sqlite3Put4byte(aFrame, nPgsz);
    sqlite3_randomness(8, &aFrame[4]);
    pLog->hdr.iCheck1 = sqlite3Get4byte(&aFrame[4]);
    pLog->hdr.iCheck2 = sqlite3Get4byte(&aFrame[8]);
    rc = sqlite3OsWrite(pLog->pFd, aFrame, LOG_HDRSIZE, 0);
    if( rc!=SQLITE_OK ){
      return rc;
    }
  }

  aCksum[0] = pLog->hdr.iCheck1;
  aCksum[1] = pLog->hdr.iCheck2;

  /* Write the log file. */
  for(p=pList; p; p=p->pDirty){
    u32 nDbsize;                  /* Db-size field for frame header */
    i64 iOffset;                  /* Write offset in log file */

    iOffset = logFrameOffset(++iFrame, nPgsz);
    
    /* Populate and write the frame header */
    nDbsize = (isCommit && p->pDirty==0) ? nTruncate : 0;
    logEncodeFrame(aCksum, p->pgno, nDbsize, nPgsz, p->pData, aFrame);
    rc = sqlite3OsWrite(pLog->pFd, aFrame, sizeof(aFrame), iOffset);
    if( rc!=SQLITE_OK ){
      return rc;
    }

    /* Write the page data */
    rc = sqlite3OsWrite(pLog->pFd, p->pData, nPgsz, iOffset + sizeof(aFrame));
    if( rc!=SQLITE_OK ){
      return rc;
    }
    pLast = p;
  }

  /* Sync the log file if the 'isSync' flag was specified. */
  if( sync_flags ){
    i64 iSegment = sqlite3OsSectorSize(pLog->pFd);
    i64 iOffset = logFrameOffset(iFrame+1, nPgsz);

    assert( isCommit );

    if( iSegment<SQLITE_DEFAULT_SECTOR_SIZE ){
      iSegment = SQLITE_DEFAULT_SECTOR_SIZE;
    }
    iSegment = (((iOffset+iSegment-1)/iSegment) * iSegment);
    while( iOffset<iSegment ){
      logEncodeFrame(aCksum,pLast->pgno,nTruncate,nPgsz,pLast->pData,aFrame);
      rc = sqlite3OsWrite(pLog->pFd, aFrame, sizeof(aFrame), iOffset);
      if( rc!=SQLITE_OK ){
        return rc;
      }

      iOffset += LOG_FRAME_HDRSIZE;
      rc = sqlite3OsWrite(pLog->pFd, pLast->pData, nPgsz, iOffset); 
      if( rc!=SQLITE_OK ){
        return rc;
      }
      nLast++;
      iOffset += nPgsz;
    }

    rc = sqlite3OsSync(pLog->pFd, sync_flags);
    if( rc!=SQLITE_OK ){
      return rc;
    }
  }

  /* Append data to the log summary. It is not necessary to lock the 
  ** log-summary to do this as the RESERVED lock held on the db file
  ** guarantees that there are no other writers, and no data that may
  ** be in use by existing readers is being overwritten.
  */
  iFrame = pLog->hdr.iLastPg;
  for(p=pList; p; p=p->pDirty){
    iFrame++;
    logSummaryAppend(pLog->pSummary, iFrame, p->pgno);
  }
  while( nLast>0 ){
    iFrame++;
    nLast--;
    logSummaryAppend(pLog->pSummary, iFrame, pLast->pgno);
  }

  /* Update the private copy of the header. */
  pLog->hdr.pgsz = nPgsz;
  pLog->hdr.iLastPg = iFrame;
  if( isCommit ){
    pLog->hdr.iChange++;
    pLog->hdr.nPage = nTruncate;
  }
  pLog->hdr.iCheck1 = aCksum[0];
  pLog->hdr.iCheck2 = aCksum[1];

  /* If this is a commit, update the log-summary header too. */
  if( isCommit && SQLITE_OK==(rc = logEnterMutex(pLog)) ){
    logSummaryWriteHdr(pLog->pSummary, &pLog->hdr);
    logLeaveMutex(pLog);
    pLog->iCallback = iFrame;
  }


  return rc;
}

/* 
** Checkpoint the database:
**
**   1. Wait for an EXCLUSIVE lock on regions B and C.
**   2. Wait for an EXCLUSIVE lock on region A.
**   3. Copy the contents of the log into the database file.
**   4. Zero the log-summary header (so new readers will ignore the log).
**   5. Drop the locks obtained in steps 1 and 2.
*/
int sqlite3WalCheckpoint(
  Log *pLog,                      /* Log connection */
  sqlite3_file *pFd,              /* File descriptor open on db file */
  int sync_flags,                 /* Flags to sync db file with (or 0) */
  u8 *zBuf,                       /* Temporary buffer to use */
  int (*xBusyHandler)(void *),    /* Pointer to busy-handler function */
  void *pBusyHandlerArg           /* Argument to pass to xBusyHandler */
){
  int rc;                         /* Return code */
  int isChanged = 0;              /* True if a new wal-index header is loaded */

  assert( !pLog->isLocked );

  /* Wait for an EXCLUSIVE lock on regions B and C. */
  do {
    rc = logLockRegion(pLog, LOG_REGION_B|LOG_REGION_C, LOG_WRLOCK);
  }while( rc==SQLITE_BUSY && xBusyHandler(pBusyHandlerArg) );
  if( rc!=SQLITE_OK ) return rc;

  /* Wait for an EXCLUSIVE lock on region A. */
  do {
    rc = logLockRegion(pLog, LOG_REGION_A, LOG_WRLOCK);
  }while( rc==SQLITE_BUSY && xBusyHandler(pBusyHandlerArg) );
  if( rc!=SQLITE_OK ){
    logLockRegion(pLog, LOG_REGION_B|LOG_REGION_C, LOG_UNLOCK);
    return rc;
  }

  /* Copy data from the log to the database file. */
  rc = logSummaryReadHdr(pLog, &isChanged);
  if( rc==SQLITE_OK ){
    rc = logCheckpoint(pLog, pFd, sync_flags, zBuf);
  }
  if( isChanged ){
    /* If a new wal-index header was loaded before the checkpoint was 
    ** performed, then the pager-cache associated with log pLog is now
    ** out of date. So zero the cached wal-index header to ensure that
    ** next time the pager opens a snapshot on this database it knows that
    ** the cache needs to be reset.
    */
    memset(&pLog->hdr, 0, sizeof(LogSummaryHdr));
  }

  /* Release the locks. */
  logLockRegion(pLog, LOG_REGION_A|LOG_REGION_B|LOG_REGION_C, LOG_UNLOCK);
  return rc;
}






int sqlite3WalCallback(Log *pLog){
  u32 ret = 0;
  if( pLog ){
    ret = pLog->iCallback;
    pLog->iCallback = 0;
  }
  return (int)ret;
}

#define _WAL_H_

#include "sqliteInt.h"

/* Connection to a write-ahead log (WAL) file. 
** There is one object of this type for each pager. 
*/
typedef struct Log Log;

/* Open and close a connection to a write-ahead log. */
int sqlite3WalOpen(sqlite3_vfs*, const char *zDb, Log **ppLog);
int sqlite3WalClose(Log *pLog, sqlite3_file *pFd, int sync_flags, u8 *zBuf);

/* Used by readers to open (lock) and close (unlock) a snapshot.  A 
** snapshot is like a read-transaction.  It is the state of the database
** at an instant in time.  sqlite3WalOpenSnapshot gets a read lock and
** preserves the current state even if the other threads or processes
** write to or checkpoint the WAL.  sqlite3WalCloseSnapshot() closes the
** transaction and releases the lock.
*/
int sqlite3WalOpenSnapshot(Log *pLog, int *);
void sqlite3WalCloseSnapshot(Log *pLog);

/* Read a page from the write-ahead log, if it is present. */
int sqlite3WalRead(Log *pLog, Pgno pgno, int *pInLog, u8 *pOut);

/* Return the size of the database as it existed at the beginning
** of the snapshot */
void sqlite3WalDbsize(Log *pLog, Pgno *pPgno);

/* Obtain or release the WRITER lock. */
int sqlite3WalWriteLock(Log *pLog, int op);

/* Undo any frames written (but not committed) to the log */
int sqlite3WalUndo(Log *pLog, int (*xUndo)(void *, Pgno), void *pUndoCtx);

/* Return an integer that records the current (uncommitted) write
** position in the WAL */
u32 sqlite3WalSavepoint(Log *pLog);

/* Move the write position of the WAL back to iFrame.  Called in
** response to a ROLLBACK TO command. */
int sqlite3WalSavepointUndo(Log *pLog, u32 iFrame);

/* Return true if data has been written but not committed to the log file. */
int sqlite3WalDirty(Log *pLog);

/* Write a frame or frames to the log. */
int sqlite3WalFrames(Log *pLog, int, PgHdr *, Pgno, int, int);

/* Copy pages from the log to the database file */ 
int sqlite3WalCheckpoint(
  Log *pLog,                      /* Log connection */
  sqlite3_file *pFd,              /* File descriptor open on db file */
  int sync_flags,                 /* Flags to sync db file with (or 0) */
  u8 *zBuf,                       /* Temporary buffer to use */
  int (*xBusyHandler)(void *),    /* Pointer to busy-handler function */
  void *pBusyHandlerArg           /* Argument to pass to xBusyHandler */
);

/* Return the value to pass to a sqlite3_wal_hook callback, the
** number of frames in the WAL at the point of the last commit since
** sqlite3WalCallback() was called.  If no commits have occurred since
** the last call, then return 0.
*/
int sqlite3WalCallback(Log *pLog);

#endif /* _WAL_H_ */