Why does _dbus_exit call _exit?

Havoc Pennington hp at redhat.com
Tue Oct 3 07:37:08 PDT 2006


Padraig O'Briain wrote:
> I have been looking at why ORBit2 leaves stale sockets in 
> $TMPDIR/orbit-padraigo when I log out.
> 
> I noticed that /usr/lib/gnome-vfs-daemon always left its stale ORBit 
> socket behind.
> 
> This program exits when dbus_connection_dispatch calls _dbus_exit. The 
> function _dbus-exit calls _exit instead of exit.
> 
> When _exit is called instead of exit the at_exit handler is not called 
> so ORBit does not clean up.
> 

I believe it was originally because I was using dbus_exit after a fork, 
so didn't want the child to call the parent's atexit handlers. (One of 
the many nasty aspects of atexit handlers[1]). dbus_exit then happened 
to be used also for the dbus_connection_set_exit_on_disconnect() feature.

The way dbus addresses the stale sockets problem on linux is to use 
abstract sockets, and on other platforms it has dbus-cleanup-sockets 
which can be run in a cron job or on logout or whenever. 
dbus-cleanup-sockets is copied from orbit, I think the orbit version 
isn't installed though so can't be run in cron. That's what I would 
suggest doing with orbit.

It's probably tough to argue that 
dbus_connection_set_exit_on_disconnect() should use _exit, though.

Havoc

[1]
  - they run in semi-undefined order (order of registration)
  - there's an arbitrary max number of them and no way to really respond
    if it's exceeded or know it's exceeded
  - you can't remove them if you want to uninit or unload what added them
  - they run in child processes also if you fork and screw up the parent
  - the big one - they don't run on abnormal exit, so can't possibly do
    anything that genuinely needs doing

the only real benefit of exit() afaik is that it flushes stdout.


More information about the dbus mailing list