os_unix.c

最新的sqlite3.6.2源代码

** NB:  If you define USE_PREAD or USE_PREAD64, then it might also
** be necessary to define _XOPEN_SOURCE to be 500.  This varies from
** one system to another.  Since SQLite does not define USE_PREAD
** any any form by default, we will not attempt to define _XOPEN_SOURCE.
** See tickets #2741 and #2681.
*/
static int seekAndRead(unixFile *id, sqlite3_int64 offset, void *pBuf, int cnt){
  int got;
  i64 newOffset;
  TIMER_START;
#if defined(USE_PREAD)
  got = pread(id->h, pBuf, cnt, offset);
  SimulateIOError( got = -1 );
#elif defined(USE_PREAD64)
  got = pread64(id->h, pBuf, cnt, offset);
  SimulateIOError( got = -1 );
#else
  newOffset = lseek(id->h, offset, SEEK_SET);
  SimulateIOError( newOffset-- );
  if( newOffset!=offset ){
    return -1;
  }
  got = read(id->h, pBuf, cnt);
#endif
  TIMER_END;
  OSTRACE5("READ    %-3d %5d %7lld %llu\n", id->h, got, offset, TIMER_ELAPSED);
  return got;
}

/*
** Read data from a file into a buffer.  Return SQLITE_OK if all
** bytes were read successfully and SQLITE_IOERR if anything goes
** wrong.
*/
static int unixRead(
  sqlite3_file *id, 
  void *pBuf, 
  int amt,
  sqlite3_int64 offset
){
  int got;
  assert( id );
  got = seekAndRead((unixFile*)id, offset, pBuf, amt);
  if( got==amt ){
    return SQLITE_OK;
  }else if( got<0 ){
    return SQLITE_IOERR_READ;
  }else{
    memset(&((char*)pBuf)[got], 0, amt-got);
    return SQLITE_IOERR_SHORT_READ;
  }
}

/*
** Seek to the offset in id->offset then read cnt bytes into pBuf.
** Return the number of bytes actually read.  Update the offset.
*/
static int seekAndWrite(unixFile *id, i64 offset, const void *pBuf, int cnt){
  int got;
  i64 newOffset;
  TIMER_START;
#if defined(USE_PREAD)
  got = pwrite(id->h, pBuf, cnt, offset);
#elif defined(USE_PREAD64)
  got = pwrite64(id->h, pBuf, cnt, offset);
#else
  newOffset = lseek(id->h, offset, SEEK_SET);
  if( newOffset!=offset ){
    return -1;
  }
  got = write(id->h, pBuf, cnt);
#endif
  TIMER_END;
  OSTRACE5("WRITE   %-3d %5d %7lld %llu\n", id->h, got, offset, TIMER_ELAPSED);
  return got;
}


/*
** Write data from a buffer into a file.  Return SQLITE_OK on success
** or some other error code on failure.
*/
static int unixWrite(
  sqlite3_file *id, 
  const void *pBuf, 
  int amt,
  sqlite3_int64 offset 
){
  int wrote = 0;
  assert( id );
  assert( amt>0 );
  while( amt>0 && (wrote = seekAndWrite((unixFile*)id, offset, pBuf, amt))>0 ){
    amt -= wrote;
    offset += wrote;
    pBuf = &((char*)pBuf)[wrote];
  }
  SimulateIOError(( wrote=(-1), amt=1 ));
  SimulateDiskfullError(( wrote=0, amt=1 ));
  if( amt>0 ){
    if( wrote<0 ){
      return SQLITE_IOERR_WRITE;
    }else{
      return SQLITE_FULL;
    }
  }
  return SQLITE_OK;
}

#ifdef SQLITE_TEST
/*
** Count the number of fullsyncs and normal syncs.  This is used to test
** that syncs and fullsyncs are occuring at the right times.
*/
int sqlite3_sync_count = 0;
int sqlite3_fullsync_count = 0;
#endif

/*
** Use the fdatasync() API only if the HAVE_FDATASYNC macro is defined.
** Otherwise use fsync() in its place.
*/
#ifndef HAVE_FDATASYNC
# define fdatasync fsync
#endif

/*
** Define HAVE_FULLFSYNC to 0 or 1 depending on whether or not
** the F_FULLFSYNC macro is defined.  F_FULLFSYNC is currently
** only available on Mac OS X.  But that could change.
*/
#ifdef F_FULLFSYNC
# define HAVE_FULLFSYNC 1
#else
# define HAVE_FULLFSYNC 0
#endif


/*
** The fsync() system call does not work as advertised on many
** unix systems.  The following procedure is an attempt to make
** it work better.
**
** The SQLITE_NO_SYNC macro disables all fsync()s.  This is useful
** for testing when we want to run through the test suite quickly.
** You are strongly advised *not* to deploy with SQLITE_NO_SYNC
** enabled, however, since with SQLITE_NO_SYNC enabled, an OS crash
** or power failure will likely corrupt the database file.
*/
static int full_fsync(int fd, int fullSync, int dataOnly){
  int rc;

  /* Record the number of times that we do a normal fsync() and 
  ** FULLSYNC.  This is used during testing to verify that this procedure
  ** gets called with the correct arguments.
  */
#ifdef SQLITE_TEST
  if( fullSync ) sqlite3_fullsync_count++;
  sqlite3_sync_count++;
#endif

  /* If we compiled with the SQLITE_NO_SYNC flag, then syncing is a
  ** no-op
  */
#ifdef SQLITE_NO_SYNC
  rc = SQLITE_OK;
#else

#if HAVE_FULLFSYNC
  if( fullSync ){
    rc = fcntl(fd, F_FULLFSYNC, 0);
  }else{
    rc = 1;
  }
  /* If the FULLFSYNC failed, fall back to attempting an fsync().
   * It shouldn't be possible for fullfsync to fail on the local 
   * file system (on OSX), so failure indicates that FULLFSYNC
   * isn't supported for this file system. So, attempt an fsync 
   * and (for now) ignore the overhead of a superfluous fcntl call.  
   * It'd be better to detect fullfsync support once and avoid 
   * the fcntl call every time sync is called.
   */
  if( rc ) rc = fsync(fd);

#else 
  if( dataOnly ){
    rc = fdatasync(fd);
  }else{
    rc = fsync(fd);
  }
#endif /* HAVE_FULLFSYNC */
#endif /* defined(SQLITE_NO_SYNC) */

  return rc;
}

/*
** Make sure all writes to a particular file are committed to disk.
**
** If dataOnly==0 then both the file itself and its metadata (file
** size, access time, etc) are synced.  If dataOnly!=0 then only the
** file data is synced.
**
** Under Unix, also make sure that the directory entry for the file
** has been created by fsync-ing the directory that contains the file.
** If we do not do this and we encounter a power failure, the directory
** entry for the journal might not exist after we reboot.  The next
** SQLite to access the file will not know that the journal exists (because
** the directory entry for the journal was never created) and the transaction
** will not roll back - possibly leading to database corruption.
*/
static int unixSync(sqlite3_file *id, int flags){
  int rc;
  unixFile *pFile = (unixFile*)id;

  int isDataOnly = (flags&SQLITE_SYNC_DATAONLY);
  int isFullsync = (flags&0x0F)==SQLITE_SYNC_FULL;

  /* Check that one of SQLITE_SYNC_NORMAL or FULL was passed */
  assert((flags&0x0F)==SQLITE_SYNC_NORMAL
      || (flags&0x0F)==SQLITE_SYNC_FULL
  );

  assert( pFile );
  OSTRACE2("SYNC    %-3d\n", pFile->h);
  rc = full_fsync(pFile->h, isFullsync, isDataOnly);
  SimulateIOError( rc=1 );
  if( rc ){
    return SQLITE_IOERR_FSYNC;
  }
  if( pFile->dirfd>=0 ){
    OSTRACE4("DIRSYNC %-3d (have_fullfsync=%d fullsync=%d)\n", pFile->dirfd,
            HAVE_FULLFSYNC, isFullsync);
#ifndef SQLITE_DISABLE_DIRSYNC
    /* The directory sync is only attempted if full_fsync is
    ** turned off or unavailable.  If a full_fsync occurred above,
    ** then the directory sync is superfluous.
    */
    if( (!HAVE_FULLFSYNC || !isFullsync) && full_fsync(pFile->dirfd,0,0) ){
       /*
       ** We have received multiple reports of fsync() returning
       ** errors when applied to directories on certain file systems.
       ** A failed directory sync is not a big deal.  So it seems
       ** better to ignore the error.  Ticket #1657
       */
       /* return SQLITE_IOERR; */
    }
#endif
    close(pFile->dirfd);  /* Only need to sync once, so close the directory */
    pFile->dirfd = -1;    /* when we are done. */
  }
  return SQLITE_OK;
}

/*
** Truncate an open file to a specified size
*/
static int unixTruncate(sqlite3_file *id, i64 nByte){
  int rc;
  assert( id );
  SimulateIOError( return SQLITE_IOERR_TRUNCATE );
  rc = ftruncate(((unixFile*)id)->h, (off_t)nByte);
  if( rc ){
    return SQLITE_IOERR_TRUNCATE;
  }else{
    return SQLITE_OK;
  }
}

/*
** Determine the current size of a file in bytes
*/
static int unixFileSize(sqlite3_file *id, i64 *pSize){
  int rc;
  struct stat buf;
  assert( id );
  rc = fstat(((unixFile*)id)->h, &buf);
  SimulateIOError( rc=1 );
  if( rc!=0 ){
    return SQLITE_IOERR_FSTAT;
  }
  *pSize = buf.st_size;

  /* When opening a zero-size database, the findLockInfo() procedure
  ** writes a single byte into that file in order to work around a bug
  ** in the OS-X msdos filesystem.  In order to avoid problems with upper
  ** layers, we need to report this file size as zero even though it is
  ** really 1.   Ticket #3260.
  */
  if( *pSize==1 ) *pSize = 0;


  return SQLITE_OK;
}

/*
** This routine translates a standard POSIX errno code into something
** useful to the clients of the sqlite3 functions.  Specifically, it is
** intended to translate a variety of "try again" errors into SQLITE_BUSY
** and a variety of "please close the file descriptor NOW" errors into 
** SQLITE_IOERR
** 
** Errors during initialization of locks, or file system support for locks,
** should handle ENOLCK, ENOTSUP, EOPNOTSUPP separately.
*/
static int sqliteErrorFromPosixError(int posixError, int sqliteIOErr) {
  switch (posixError) {
  case 0: 
    return SQLITE_OK;
    
  case EAGAIN:
  case ETIMEDOUT:
  case EBUSY:
  case EINTR:
  case ENOLCK:  
    /* random NFS retry error, unless during file system support 
     * introspection, in which it actually means what it says */
    return SQLITE_BUSY;
    
  case EACCES: 
    /* EACCES is like EAGAIN during locking operations, but not any other time*/
    if( (sqliteIOErr == SQLITE_IOERR_LOCK) || 
	(sqliteIOErr == SQLITE_IOERR_UNLOCK) || 
	(sqliteIOErr == SQLITE_IOERR_RDLOCK) ||
	(sqliteIOErr == SQLITE_IOERR_CHECKRESERVEDLOCK) ){
      return SQLITE_BUSY;
    }
    /* else fall through */
  case EPERM: 
    return SQLITE_PERM;
    
  case EDEADLK:
    return SQLITE_IOERR_BLOCKED;
    
#if EOPNOTSUPP!=ENOTSUP
  case EOPNOTSUPP: 
    /* something went terribly awry, unless during file system support 
     * introspection, in which it actually means what it says */
#endif
  case ENOTSUP: 
    /* invalid fd, unless during file system support introspection, in which 
     * it actually means what it says */
  case EIO:
  case EBADF:
  case EINVAL:
  case ENOTCONN:
  case ENODEV:
  case ENXIO:
  case ENOENT:
  case ESTALE:
  case ENOSYS:
    /* these should force the client to close the file and reconnect */
    
  default: 
    return sqliteIOErr;
  }
}

/*
** This routine checks if there is a RESERVED lock held on the specified
** file by this or any other process. If such a lock is held, set *pResOut
** to a non-zero value otherwise *pResOut is set to zero.  The return value
** is set to SQLITE_OK unless an I/O error occurs during lock checking.
*/
static int unixCheckReservedLock(sqlite3_file *id, int *pResOut){
  int rc = SQLITE_OK;
  int reserved = 0;
  unixFile *pFile = (unixFile*)id;

  SimulateIOError( return SQLITE_IOERR_CHECKRESERVEDLOCK; );

  assert( pFile );
  enterMutex(); /* Because pFile->pLock is shared across threads */

  /* Check if a thread in this process holds such a lock */
  if( pFile->pLock->locktype>SHARED_LOCK ){
    reserved = 1;
  }

  /* Otherwise see if some other process holds it.
  */
  if( !reserved ){
    struct flock lock;
    lock.l_whence = SEEK_SET;
    lock.l_start = RESERVED_BYTE;
    lock.l_len = 1;
    lock.l_type = F_WRLCK;
    if (-1 == fcntl(pFile->h, F_GETLK, &lock)) {
      int tErrno = errno;
      rc = sqliteErrorFromPosixError(tErrno, SQLITE_IOERR_CHECKRESERVEDLOCK);
      pFile->lastErrno = tErrno;
    } else if( lock.l_type!=F_UNLCK ){
      reserved = 1;
    }
  }
  
  leaveMutex();
  OSTRACE4("TEST WR-LOCK %d %d %d\n", pFile->h, rc, reserved);

  *pResOut = reserved;
  return rc;
}

/*
** Lock the file with the lock specified by parameter locktype - one
** of the following:
**
**     (1) SHARED_LOCK
**     (2) RESERVED_LOCK
**     (3) PENDING_LOCK
**     (4) EXCLUSIVE_LOCK
**
** Sometimes when requesting one lock state, additional lock states
** are inserted in between.  The locking might fail on one of the later
** transitions leaving the lock state different from what it started but
** still short of its goal.  The following chart shows the allowed
** transitions and the inserted intermediate states:
**
**    UNLOCKED -> SHARED
**    SHARED -> RESERVED
**    SHARED -> (PENDING) -> EXCLUSIVE
**    RESERVED -> (PENDING) -> EXCLUSIVE