SQLite

** Function unixMutexHeld() is used to assert() that the global mutex 
** is held when required. This function is only used as part of assert() 
** statements. e.g.
**
**   unixEnterMutex()
**     assert( unixMutexHeld() );
**   unixEnterLeave()
**
** To prevent deadlock, the global unixBigLock must must be acquired
** before the unixInodeInfo.pLockMutex mutex, if both are held.  It is
** OK to get the pLockMutex without holding unixBigLock first, but if
** that happens, the unixBigLock mutex must not be acquired until after
** pLockMutex is released.
**
**      OK:     enter(unixBigLock),  enter(pLockInfo)
**      OK:     enter(unixBigLock)
**      OK:     enter(pLockInfo)
**   ERROR:     enter(pLockInfo), enter(unixBigLock)
*/
static sqlite3_mutex *unixBigLock = 0;
static void unixEnterMutex(void){
  assert( sqlite3_mutex_notheld(unixBigLock) );  /* Not a recursive mutex */
  sqlite3_mutex_enter(unixBigLock);
}
static void unixLeaveMutex(void){
  assert( sqlite3_mutex_held(unixBigLock) );
  sqlite3_mutex_leave(unixBigLock);
}
#ifdef SQLITE_DEBUG
static int unixMutexHeld(void) {
  return sqlite3_mutex_held(unixBigLock);
}
#endif

**
** A single inode can have multiple file descriptors, so each unixFile
** structure contains a pointer to an instance of this object and this
** object keeps a count of the number of unixFile pointing to it.
**
** Mutex rules:
**
**  (1) Only the pLockMutex mutex must be held in order to read or write
**      any of the locking fields:
**          nShared, nLock, eFileLock, or bProcessLock
**
**  (2) When nRef>0, then the following fields are unchanging and can
**      be read (but not written) without holding any mutex:
**          fileId, pLockMutex
**
**  (3) The pUnused field may only be changed while holding bo the
**      pLockMutex and the bigUnixLock mutex.  But it may be read
**      while holding either.
**
**  (4) With the exceptions above, all the fields may only be read
**      or written while holding the global unixBigLock mutex.
**
** Deadlock prevention:  The global unixBigLock mutex may not
** be acquired while holding the pLockMutex mutex.  If both unixBigLock
** and pLockMutex are needed, then unixBigLock must be acquired first.
*/
struct unixInodeInfo {
  struct unixFileId fileId;       /* The lookup key */
  sqlite3_mutex *pLockMutex;      /* Hold this mutex for... */
  int nShared;                      /* Number of SHARED locks held */
  int nLock;                        /* Number of outstanding file locks */
  unsigned char eFileLock;          /* One of SHARED_LOCK, RESERVED_LOCK etc. */
  unsigned char bProcessLock;       /* An exclusive process lock is held */
  UnixUnusedFd *pUnused;          /* Unused file descriptors to close */
  int nRef;                       /* Number of pointers to this structure */
  unixShmNode *pShmNode;          /* Shared memory associated with this inode */

  unixInodeInfo *pNext;           /* List of all unixInodeInfo objects */
  unixInodeInfo *pPrev;           /*    .... doubly linked */
#if SQLITE_ENABLE_LOCKING_STYLE
  unsigned long long sharedByte;  /* for AFP simulated shared lock */
#endif
#if OS_VXWORKS
  sem_t *pSem;                    /* Named POSIX semaphore */
  char aSemName[MAX_PATHNAME+2];  /* Name of that semaphore */
#endif
};

/*
** A lists of all unixInodeInfo objects.
*/
static unixInodeInfo *inodeList = 0;  /* All unixInodeInfo objects */

#ifdef SQLITE_DEBUG
/*
** True if the inode mutex is held, or not.  Used only within assert()
** to help verify correct mutex usage.
*/
int unixFileMutexHeld(unixFile *pFile){
  assert( pFile->pInode );
  return sqlite3_mutex_held(pFile->pInode->pLockMutex);
}
int unixFileMutexNotheld(unixFile *pFile){
  assert( pFile->pInode );
  return sqlite3_mutex_notheld(pFile->pInode->pLockMutex);
}
#endif

/*
**
** This function - unixLogErrorAtLine(), is only ever called via the macro
** unixLogError().
**
** It is invoked after an error occurs in an OS function and errno has been

/*
** Close all file descriptors accumuated in the unixInodeInfo->pUnused list.
*/ 
static void closePendingFds(unixFile *pFile){
  unixInodeInfo *pInode = pFile->pInode;
  UnixUnusedFd *p;
  UnixUnusedFd *pNext;
  assert( unixMutexHeld() );
  assert( unixFileMutexNotheld(pFile) );
  sqlite3_mutex_enter(pInode->pLockMutex);
  for(p=pInode->pUnused; p; p=pNext){
    pNext = p->pNext;
    robust_close(pFile, p->fd, __LINE__);
    sqlite3_free(p);

  }
  pInode->pUnused = 0;
  sqlite3_mutex_leave(pInode->pLockMutex);
}

/*
** Release a unixInodeInfo structure previously allocated by findInodeInfo().
**
** The mutex entered using the unixEnterMutex() function must be held
** when this function is called.
*/
static void releaseInodeInfo(unixFile *pFile){
  unixInodeInfo *pInode = pFile->pInode;
  assert( unixMutexHeld() );
  assert( unixFileMutexNotheld(pFile) );
  if( ALWAYS(pInode) ){
    pInode->nRef--;
    if( pInode->nRef==0 ){
      assert( pInode->pShmNode==0 );
      closePendingFds(pFile);
      if( pInode->pPrev ){
        assert( pInode->pPrev->pNext==pInode );
        pInode->pPrev->pNext = pInode->pNext;
      }else{
        assert( inodeList==pInode );
        inodeList = pInode->pNext;
      }
      if( pInode->pNext ){
        assert( pInode->pNext->pPrev==pInode );
        pInode->pNext->pPrev = pInode->pPrev;
      }
      sqlite3_mutex_free(pInode->pLockMutex);
      sqlite3_free(pInode);
    }
  }

}

/*
** Given a file descriptor, locate the unixInodeInfo object that
** describes that file descriptor.  Create a new one if necessary.  The
** return value might be uninitialized if an error occurs.
**

  memset(&fileId, 0, sizeof(fileId));
  fileId.dev = statbuf.st_dev;
#if OS_VXWORKS
  fileId.pId = pFile->pId;
#else
  fileId.ino = (u64)statbuf.st_ino;
#endif

  pInode = inodeList;
  while( pInode && memcmp(&fileId, &pInode->fileId, sizeof(fileId)) ){
    pInode = pInode->pNext;
  }
  if( pInode==0 ){
    pInode = sqlite3_malloc64( sizeof(*pInode) );
    if( pInode==0 ){

/*
** Add the file descriptor used by file handle pFile to the corresponding
** pUnused list.
*/
static void setPendingFd(unixFile *pFile){
  unixInodeInfo *pInode = pFile->pInode;
  UnixUnusedFd *p = pFile->pPreallocatedUnused;
  assert( unixMutexHeld() );
  assert( unixFileMutexNotheld(pFile) );
  sqlite3_mutex_enter(pInode->pLockMutex);
  p->pNext = pInode->pUnused;
  pInode->pUnused = p;
  sqlite3_mutex_leave(pInode->pLockMutex);
  pFile->h = -1;
  pFile->pPreallocatedUnused = 0;

}

/*
** Lower the locking level on file descriptor pFile to eFileLock.  eFileLock
** must be either NO_LOCK or SHARED_LOCK.
**
** If the locking level of the file descriptor is already at or below
** the requested locking level, this routine is a no-op.
** 
** If handleNFSUnlock is true, then on downgrading an EXCLUSIVE_LOCK to SHARED
** the byte range is divided into 2 parts and the first part is unlocked then
** set to a read lock, then the other part is simply unlocked.  This works 
** around a bug in BSD NFS lockd (also seen on MacOSX 10.3+) that fails to 
** remove the write lock on a region when a read lock is set.
*/
static int posixUnlock(sqlite3_file *id, int eFileLock, int handleNFSUnlock){
  unixFile *pFile = (unixFile*)id;
  unixInodeInfo *pInode;
  struct flock lock;
  int rc = SQLITE_OK;
  int wantToClosePending = 0;  /* True to try to close file old descriptors */

  assert( pFile );
  OSTRACE(("UNLOCK  %d %d was %d(%d,%d) pid=%d (unix)\n", pFile->h, eFileLock,
      pFile->eFileLock, pFile->pInode->eFileLock, pFile->pInode->nShared,
      osGetpid(0)));

  assert( eFileLock<=SHARED_LOCK );

    /* Decrement the count of locks against this same file.  When the
    ** count reaches zero, close any other file descriptors whose close
    ** was deferred because of outstanding locks.
    */
    pInode->nLock--;
    assert( pInode->nLock>=0 );
    if( pInode->nLock==0 && pInode->pUnused!=0 ){
      wantToClosePending = 1;
    }
  }

end_unlock:
  sqlite3_mutex_leave(pInode->pLockMutex);
  if( rc==SQLITE_OK ){
    pFile->eFileLock = eFileLock;
    if( wantToClosePending ){
      unixEnterMutex();
      if( pInode->nLock==0 ) closePendingFds(pFile);
      unixLeaveMutex();
    }
  }
  return rc;
}

/*
** Lower the locking level on file descriptor pFile to eFileLock.  eFileLock
** must be either NO_LOCK or SHARED_LOCK.
**

** Close a file.
*/
static int unixClose(sqlite3_file *id){
  int rc = SQLITE_OK;
  unixFile *pFile = (unixFile *)id;
  verifyDbFile(pFile);
  unixUnlock(id, NO_LOCK);
  assert( unixFileMutexNotheld(pFile) );
  unixEnterMutex();

  /* unixFile.pInode is always valid here. Otherwise, a different close
  ** routine (e.g. nolockClose()) would be called instead.
  */
  assert( pFile->pInode->nLock>0 || pFile->pInode->bProcessLock==0 );
  if( ALWAYS(pFile->pInode) && pFile->pInode->nLock ){

 ** Close a file.
 */
static int semXClose(sqlite3_file *id) {
  if( id ){
    unixFile *pFile = (unixFile*)id;
    semXUnlock(id, NO_LOCK);
    assert( pFile );
    assert( unixFileMutexNotheld(pFile) );
    unixEnterMutex();
    releaseInodeInfo(pFile);
    unixLeaveMutex();
    closeUnixFile(id);
  }
  return SQLITE_OK;
}

*/
static int afpUnlock(sqlite3_file *id, int eFileLock) {
  int rc = SQLITE_OK;
  unixFile *pFile = (unixFile*)id;
  unixInodeInfo *pInode;
  afpLockingContext *context = (afpLockingContext *) pFile->lockingContext;
  int skipShared = 0;
  int wantToClosePending = 0;
#ifdef SQLITE_TEST
  int h = pFile->h;
#endif

  assert( pFile );
  OSTRACE(("UNLOCK  %d %d was %d(%d,%d) pid=%d (afp)\n", pFile->h, eFileLock,
           pFile->eFileLock, pFile->pInode->eFileLock, pFile->pInode->nShared,

        pInode->eFileLock = NO_LOCK;
        pFile->eFileLock = NO_LOCK;
      }
    }
    if( rc==SQLITE_OK ){
      pInode->nLock--;
      assert( pInode->nLock>=0 );
      if( pInode->nLock==0 && pInode->pUnused!=0 ) wantToClosePending = 1;


    }
  }
  
  sqlite3_mutex_leave(pInode->pLockMutex);
  if( rc==SQLITE_OK ){
    pFile->eFileLock = eFileLock;
    if( wantToClosePending ){
      unixEnterMutex();
      if( pInode->nLock==0 ) closePendingFds(pFile);
      unixLeaveMutex();
    }
  }
  return rc;
}

/*
** Close a file & cleanup AFP specific locking context 
*/
static int afpClose(sqlite3_file *id) {
  int rc = SQLITE_OK;
  unixFile *pFile = (unixFile*)id;
  assert( id!=0 );
  afpUnlock(id, NO_LOCK);
  assert( unixFileMutexNotheld(pFile) );
  unixEnterMutex();
  if( pFile->pInode && pFile->pInode->nLock ){
    /* If there are outstanding locks, do not actually close the file just
    ** yet because that would clear those locks.  Instead, add the file
    ** descriptor to pInode->aPending.  It will be automatically closed when
    ** the last lock is cleared.
    */

  if( p==0 ) return SQLITE_NOMEM_BKPT;
  memset(p, 0, sizeof(*p));
  assert( pDbFd->pShm==0 );

  /* Check to see if a unixShmNode object already exists. Reuse an existing
  ** one if present. Create a new one if necessary.
  */
  assert( unixFileMutexNotheld(pDbFd) );
  unixEnterMutex();
  pInode = pDbFd->pInode;
  pShmNode = pInode->pShmNode;
  if( pShmNode==0 ){
    struct stat sStat;                 /* fstat() info for database file */
#ifndef SQLITE_SHM_DIRECTORY
    const char *zBasePath = pDbFd->zPath;

** any load or store begun after the barrier.
*/
static void unixShmBarrier(
  sqlite3_file *fd                /* Database file holding the shared memory */
){
  UNUSED_PARAMETER(fd);
  sqlite3MemoryBarrier();         /* compiler-defined memory barrier */
  assert( unixFileMutexNotheld((unixFile*)fd) );
  unixEnterMutex();               /* Also mutex, for redundancy */
  unixLeaveMutex();
}

/*
** Close a connection to shared-memory.  Delete the underlying 
** storage if deleteFlag is true.

  /* Free the connection p */
  sqlite3_free(p);
  pDbFd->pShm = 0;
  sqlite3_mutex_leave(pShmNode->mutex);

  /* If pShmNode->nRef has reached 0, then close the underlying
  ** shared-memory file, too */
  assert( unixFileMutexNotheld(pDbFd) );
  unixEnterMutex();
  assert( pShmNode->nRef>0 );
  pShmNode->nRef--;
  if( pShmNode->nRef==0 ){
    if( deleteFlag && pShmNode->h>=0 ){
      osUnlink(pShmNode->zFilename);
    }

  ** almost certain that an open() call on the same path will also fail.
  ** For this reason, if an error occurs in the stat() call here, it is
  ** ignored and -1 is returned. The caller will try to open a new file
  ** descriptor on the same path, fail, and return an error to SQLite.
  **
  ** Even if a subsequent open() call does succeed, the consequences of
  ** not searching for a reusable file descriptor are not dire.  */
  if( inodeList!=0 && 0==osStat(zPath, &sStat) ){
    unixInodeInfo *pInode;

    pInode = inodeList;
    while( pInode && (pInode->fileId.dev!=sStat.st_dev
                     || pInode->fileId.ino!=(u64)sStat.st_ino) ){
       pInode = pInode->pNext;
    }
    if( pInode ){
      UnixUnusedFd **pp;
      assert( sqlite3_mutex_notheld(pInode->pLockMutex) );
      sqlite3_mutex_enter(pInode->pLockMutex);
      for(pp=&pInode->pUnused; *pp && (*pp)->flags!=flags; pp=&((*pp)->pNext));
      pUnused = *pp;
      if( pUnused ){

        *pp = pUnused->pNext;
      }
      sqlite3_mutex_leave(pInode->pLockMutex);
    }
  }
  unixLeaveMutex();
#endif    /* if !OS_VXWORKS */
  return pUnused;
}