Re: May close() return any error code?

[Takashi Iwai <tiwai@suse.de>]

On Wed, 29 Jul 2015 16:45:11 +0200,
Dmitry Torokhov wrote:
> 
> HI Takashi,
> 
> On Wed, Jul 29, 2015 at 12:46:59PM +0200, Takashi Iwai wrote:
> > Hi,
> > 
> > while debugging a problem of X and gdm with the old systemd-210, we
> > encountered a sudden death of systemd-logind, and this turned out to
> > be an unexpected errno from close().  The close() call for input
> > devices returns ENODEV error.  The logind in systemd-210 treats this
> > error code as fatal, triggers assert() and eventually kills itself.
> > The details are found in an openSUSE bugzilla thread:
> >   https://bugzilla.opensuse.org/show_bug.cgi?id=939571
> > 
> > This seems coming from evdev_flush().  As there is no fd leak, it's no
> > big problem per se.  But, now the question is whether returning such
> > an error code is correct behavior at all.  At least, it doesn't seem
> > defined in POSIX:
> >   http://pubs.opengroup.org/onlinepubs/009695399/functions/close.html
> 
> Hmm, if I checked the right version of the code close_nointr_nofail()
> expects only 0 as the return code so even if we change the kernel to
> use more conforming -EIO instead of -ENODEV systemd will still die...

Yes, it can be seen rather as a bug in systemd-210, and the later
systemd version doesn't seem to trigger assert() with that condition.

> The question is whether we really need to propagate return value from
> f_op->flush() up to userspace in filp_close(). Why don't we ask Al?

Oh yes, of course, Al is the best person to ask :)

Looking back at this problem now, I find it even dangerous to
propagate an error, especially -EINTR.  Then user-space may interpret
it as if the close didn't succeed and the fd were alive, although the
kernel performed the close itself.


thanks,

Takashi