os_unix.c

最新的sqlite3.6.2源代码

struct threadTestData {
  int fd;                /* File to be locked */
  struct flock lock;     /* The locking operation */
  int result;            /* Result of the locking operation */
};

#ifdef SQLITE_LOCK_TRACE
/*
** Print out information about all locking operations.
**
** This routine is used for troubleshooting locks on multithreaded
** platforms.  Enable by compiling with the -DSQLITE_LOCK_TRACE
** command-line option on the compiler.  This code is normally
** turned off.
*/
static int lockTrace(int fd, int op, struct flock *p){
  char *zOpName, *zType;
  int s;
  int savedErrno;
  if( op==F_GETLK ){
    zOpName = "GETLK";
  }else if( op==F_SETLK ){
    zOpName = "SETLK";
  }else{
    s = fcntl(fd, op, p);
    sqlite3DebugPrintf("fcntl unknown %d %d %d\n", fd, op, s);
    return s;
  }
  if( p->l_type==F_RDLCK ){
    zType = "RDLCK";
  }else if( p->l_type==F_WRLCK ){
    zType = "WRLCK";
  }else if( p->l_type==F_UNLCK ){
    zType = "UNLCK";
  }else{
    assert( 0 );
  }
  assert( p->l_whence==SEEK_SET );
  s = fcntl(fd, op, p);
  savedErrno = errno;
  sqlite3DebugPrintf("fcntl %d %d %s %s %d %d %d %d\n",
     threadid, fd, zOpName, zType, (int)p->l_start, (int)p->l_len,
     (int)p->l_pid, s);
  if( s==(-1) && op==F_SETLK && (p->l_type==F_RDLCK || p->l_type==F_WRLCK) ){
    struct flock l2;
    l2 = *p;
    fcntl(fd, F_GETLK, &l2);
    if( l2.l_type==F_RDLCK ){
      zType = "RDLCK";
    }else if( l2.l_type==F_WRLCK ){
      zType = "WRLCK";
    }else if( l2.l_type==F_UNLCK ){
      zType = "UNLCK";
    }else{
      assert( 0 );
    }
    sqlite3DebugPrintf("fcntl-failure-reason: %s %d %d %d\n",
       zType, (int)l2.l_start, (int)l2.l_len, (int)l2.l_pid);
  }
  errno = savedErrno;
  return s;
}
#define fcntl lockTrace
#endif /* SQLITE_LOCK_TRACE */

/*
** The testThreadLockingBehavior() routine launches two separate
** threads on this routine.  This routine attempts to lock a file
** descriptor then returns.  The success or failure of that attempt
** allows the testThreadLockingBehavior() procedure to determine
** whether or not threads can override each others locks.
*/
static void *threadLockingTest(void *pArg){
  struct threadTestData *pData = (struct threadTestData*)pArg;
  pData->result = fcntl(pData->fd, F_SETLK, &pData->lock);
  return pArg;
}

/*
** This procedure attempts to determine whether or not threads
** can override each others locks then sets the 
** threadsOverrideEachOthersLocks variable appropriately.
*/
static void testThreadLockingBehavior(int fd_orig){
  int fd;
  struct threadTestData d[2];
  pthread_t t[2];

  fd = dup(fd_orig);
  if( fd<0 ) return;
  memset(d, 0, sizeof(d));
  d[0].fd = fd;
  d[0].lock.l_type = F_RDLCK;
  d[0].lock.l_len = 1;
  d[0].lock.l_start = 0;
  d[0].lock.l_whence = SEEK_SET;
  d[1] = d[0];
  d[1].lock.l_type = F_WRLCK;
  pthread_create(&t[0], 0, threadLockingTest, &d[0]);
  pthread_create(&t[1], 0, threadLockingTest, &d[1]);
  pthread_join(t[0], 0);
  pthread_join(t[1], 0);
  close(fd);
  threadsOverrideEachOthersLocks =  d[0].result==0 && d[1].result==0;
}
#endif /* SQLITE_THREADSAFE */

/*
** Release a lockInfo structure previously allocated by findLockInfo().
*/
static void releaseLockInfo(struct lockInfo *pLock){
  if( pLock ){
    pLock->nRef--;
    if( pLock->nRef==0 ){
      if( pLock->pPrev ){
        assert( pLock->pPrev->pNext==pLock );
        pLock->pPrev->pNext = pLock->pNext;
      }else{
        assert( lockList==pLock );
        lockList = pLock->pNext;
      }
      if( pLock->pNext ){
        assert( pLock->pNext->pPrev==pLock );
        pLock->pNext->pPrev = pLock->pPrev;
      }
      sqlite3_free(pLock);
    }
  }
}

/*
** Release a openCnt structure previously allocated by findLockInfo().
*/
static void releaseOpenCnt(struct openCnt *pOpen){
  if( pOpen ){
    pOpen->nRef--;
    if( pOpen->nRef==0 ){
      if( pOpen->pPrev ){
        assert( pOpen->pPrev->pNext==pOpen );
        pOpen->pPrev->pNext = pOpen->pNext;
      }else{
        assert( openList==pOpen );
        openList = pOpen->pNext;
      }
      if( pOpen->pNext ){
        assert( pOpen->pNext->pPrev==pOpen );
        pOpen->pNext->pPrev = pOpen->pPrev;
      }
      sqlite3_free(pOpen->aPending);
      sqlite3_free(pOpen);
    }
  }
}

#ifdef SQLITE_ENABLE_LOCKING_STYLE
/*
** Tests a byte-range locking query to see if byte range locks are 
** supported, if not we fall back to dotlockLockingStyle.
*/
static int testLockingStyle(int fd){
  struct flock lockInfo;

  /* Test byte-range lock using fcntl(). If the call succeeds, 
  ** assume that the file-system supports POSIX style locks. 
  */
  lockInfo.l_len = 1;
  lockInfo.l_start = 0;
  lockInfo.l_whence = SEEK_SET;
  lockInfo.l_type = F_RDLCK;
  if( fcntl(fd, F_GETLK, &lockInfo)!=-1 ) {
    return LOCKING_STYLE_POSIX;
  }
  
  /* Testing for flock() can give false positives.  So if if the above 
  ** test fails, then we fall back to using dot-file style locking.
  */  
  return LOCKING_STYLE_DOTFILE;
}
#endif

/* 
** If SQLITE_ENABLE_LOCKING_STYLE is defined, this function Examines the 
** f_fstypename entry in the statfs structure as returned by stat() for 
** the file system hosting the database file and selects  the appropriate
** locking style based on its value.  These values and assignments are 
** based on Darwin/OSX behavior and have not been thoroughly tested on 
** other systems.
**
** If SQLITE_ENABLE_LOCKING_STYLE is not defined, this function always
** returns LOCKING_STYLE_POSIX.
*/
static int detectLockingStyle(
  sqlite3_vfs *pVfs,
  const char *filePath, 
  int fd
){
#ifdef SQLITE_ENABLE_LOCKING_STYLE
  struct Mapping {
    const char *zFilesystem;
    int eLockingStyle;
  } aMap[] = {
    { "hfs",    LOCKING_STYLE_POSIX },
    { "ufs",    LOCKING_STYLE_POSIX },
    { "afpfs",  LOCKING_STYLE_AFP },
#ifdef SQLITE_ENABLE_AFP_LOCKING_SMB
    { "smbfs",  LOCKING_STYLE_AFP },
#else
    { "smbfs",  LOCKING_STYLE_FLOCK },
#endif
    { "msdos",  LOCKING_STYLE_DOTFILE },
    { "webdav", LOCKING_STYLE_NONE },
    { 0, 0 }
  };
  int i;
  struct statfs fsInfo;

  if( !filePath ){
    return LOCKING_STYLE_NONE;
  }
  if( pVfs->pAppData ){
    return SQLITE_PTR_TO_INT(pVfs->pAppData);
  }

  if( statfs(filePath, &fsInfo) != -1 ){
    if( fsInfo.f_flags & MNT_RDONLY ){
      return LOCKING_STYLE_NONE;
    }
    for(i=0; aMap[i].zFilesystem; i++){
      if( strcmp(fsInfo.f_fstypename, aMap[i].zFilesystem)==0 ){
        return aMap[i].eLockingStyle;
      }
    }
  }

  /* Default case. Handles, amongst others, "nfs". */
  return testLockingStyle(fd);  
#endif
  return LOCKING_STYLE_POSIX;
}

/*
** Given a file descriptor, locate lockInfo and openCnt structures that
** describes that file descriptor.  Create new ones if necessary.  The
** return values might be uninitialized if an error occurs.
**
** Return an appropriate error code.
*/
static int findLockInfo(
  int fd,                      /* The file descriptor used in the key */
  struct lockInfo **ppLock,    /* Return the lockInfo structure here */
  struct openCnt **ppOpen      /* Return the openCnt structure here */
){
  int rc;
  struct lockKey key1;
  struct openKey key2;
  struct stat statbuf;
  struct lockInfo *pLock;
  struct openCnt *pOpen;
  rc = fstat(fd, &statbuf);
  if( rc!=0 ){
#ifdef EOVERFLOW
    if( errno==EOVERFLOW ) return SQLITE_NOLFS;
#endif
    return SQLITE_IOERR;
  }

  /* On OS X on an msdos filesystem, the inode number is reported
  ** incorrectly for zero-size files.  See ticket #3260.  To work
  ** around this problem (we consider it a bug in OS X, not SQLite)
  ** we always increase the file size to 1 by writing a single byte
  ** prior to accessing the inode number.  The one byte written is
  ** an ASCII 'S' character which also happens to be the first byte
  ** in the header of every SQLite database.  In this way, if there
  ** is a race condition such that another thread has already populated
  ** the first page of the database, no damage is done.
  */
  if( statbuf.st_size==0 ){
    write(fd, "S", 1);
    rc = fstat(fd, &statbuf);
    if( rc!=0 ){
      return SQLITE_IOERR;
    }
  }

  memset(&key1, 0, sizeof(key1));
  key1.dev = statbuf.st_dev;
  key1.ino = statbuf.st_ino;
#if SQLITE_THREADSAFE
  if( threadsOverrideEachOthersLocks<0 ){
    testThreadLockingBehavior(fd);
  }
  key1.tid = threadsOverrideEachOthersLocks ? 0 : pthread_self();
#endif
  memset(&key2, 0, sizeof(key2));
  key2.dev = statbuf.st_dev;
  key2.ino = statbuf.st_ino;
  pLock = lockList;
  while( pLock && memcmp(&key1, &pLock->key, sizeof(key1)) ){
    pLock = pLock->pNext;
  }
  if( pLock==0 ){
    pLock = sqlite3_malloc( sizeof(*pLock) );
    if( pLock==0 ){
      rc = SQLITE_NOMEM;
      goto exit_findlockinfo;
    }
    pLock->key = key1;
    pLock->nRef = 1;
    pLock->cnt = 0;
    pLock->locktype = 0;
    pLock->pNext = lockList;
    pLock->pPrev = 0;
    if( lockList ) lockList->pPrev = pLock;
    lockList = pLock;
  }else{
    pLock->nRef++;
  }
  *ppLock = pLock;
  if( ppOpen!=0 ){
    pOpen = openList;
    while( pOpen && memcmp(&key2, &pOpen->key, sizeof(key2)) ){
      pOpen = pOpen->pNext;
    }
    if( pOpen==0 ){
      pOpen = sqlite3_malloc( sizeof(*pOpen) );
      if( pOpen==0 ){
        releaseLockInfo(pLock);
        rc = SQLITE_NOMEM;
        goto exit_findlockinfo;
      }
      pOpen->key = key2;
      pOpen->nRef = 1;
      pOpen->nLock = 0;
      pOpen->nPending = 0;
      pOpen->aPending = 0;
      pOpen->pNext = openList;
      pOpen->pPrev = 0;
      if( openList ) openList->pPrev = pOpen;
      openList = pOpen;
    }else{
      pOpen->nRef++;
    }
    *ppOpen = pOpen;
  }

exit_findlockinfo:
  return rc;
}

#ifdef SQLITE_DEBUG
/*
** Helper function for printing out trace information from debugging
** binaries. This returns the string represetation of the supplied
** integer lock-type.
*/
static const char *locktypeName(int locktype){
  switch( locktype ){
  case NO_LOCK: return "NONE";
  case SHARED_LOCK: return "SHARED";
  case RESERVED_LOCK: return "RESERVED";
  case PENDING_LOCK: return "PENDING";
  case EXCLUSIVE_LOCK: return "EXCLUSIVE";
  }
  return "ERROR";
}
#endif

/*
** If we are currently in a different thread than the thread that the
** unixFile argument belongs to, then transfer ownership of the unixFile
** over to the current thread.
**
** A unixFile is only owned by a thread on systems where one thread is
** unable to override locks created by a different thread.  RedHat9 is
** an example of such a system.
**
** Ownership transfer is only allowed if the unixFile is currently unlocked.
** If the unixFile is locked and an ownership is wrong, then return
** SQLITE_MISUSE.  SQLITE_OK is returned if everything works.
*/
#if SQLITE_THREADSAFE
static int transferOwnership(unixFile *pFile){
  int rc;
  pthread_t hSelf;
  if( threadsOverrideEachOthersLocks ){
    /* Ownership transfers not needed on this system */
    return SQLITE_OK;
  }
  hSelf = pthread_self();
  if( pthread_equal(pFile->tid, hSelf) ){
    /* We are still in the same thread */
    OSTRACE1("No-transfer, same thread\n");
    return SQLITE_OK;
  }
  if( pFile->locktype!=NO_LOCK ){
    /* We cannot change ownership while we are holding a lock! */
    return SQLITE_MISUSE;
  }
  OSTRACE4("Transfer ownership of %d from %d to %d\n",
            pFile->h, pFile->tid, hSelf);
  pFile->tid = hSelf;
  if (pFile->pLock != NULL) {
    releaseLockInfo(pFile->pLock);
    rc = findLockInfo(pFile->h, &pFile->pLock, 0);
    OSTRACE5("LOCK    %d is now %s(%s,%d)\n", pFile->h,
           locktypeName(pFile->locktype),
           locktypeName(pFile->pLock->locktype), pFile->pLock->cnt);
    return rc;
  } else {
    return SQLITE_OK;
  }
}
#else
  /* On single-threaded builds, ownership transfer is a no-op */
# define transferOwnership(X) SQLITE_OK
#endif

/*
** Seek to the offset passed as the second argument, then read cnt 
** bytes into pBuf. Return the number of bytes actually read.
**