aio_fsync() a directory ?

[Xavier Roche <xavier dot roche at 3ds dot com>]

Hi folks!

As a followup to a discussion 
(<https://groups.google.com/forum/?fromgroups#!topic/comp.unix.programmer/AM2V83RCOVE>) 
regarding synchronization of file operations (file creation, renaming, 
symlink, etc.), it seems that the only way to "commit" a filename change 
is to open() the containing directory in O_RDONLY, and fsync() the given 
file descriptor.

This behavior seems to work with fsync() [even if it is not really 
documented], and appears to have the expected side effects.

However, aio_fsync() [as tested in glibc 2.17) refuses to operate on a 
file descriptor not opened in write mode:

Extract from the glibc-2.17 sysdeps/pthread/aio_fsync.c:

int
aio_fsync (int op, struct aiocb *aiocbp)
{
...
  flags = fcntl (aiocbp->aio_fildes, F_GETFL);
  if (__builtin_expect (flags == -1, 0)
      || __builtin_expect ((flags & O_ACCMODE) == O_RDONLY, 0))
    {
      __set_errno (EBADF);
      return -1;
    }

This behavior is inconsistent with fsync() ; but OTOH it is consistent 
with the standard ; aio_fsync() "shall fail if the aio_fildes member of 
the aiocb structure referenced by the aiocbp argument is not a valid 
file descriptor open for writing"

Strangely, fsync() does not have this restriction (neither in the 
standard nor on the implementation) - was it on purpose ?

http://pubs.opengroup.org/onlinepubs/009695399/functions/fsync.html
http://pubs.opengroup.org/onlinepubs/009696699/functions/aio_fsync.html

At last, is fsync'ing a directory the right way to commit operations on 
files entries ? If so, isn't it inconsistent not to accept this on the 
aio side ?