bug#13540: 24.3.50; Regression in trunk: auto-revert-mode fails on files accessed via samba

[Eli Zaretskii <eliz@gnu.org>]

> Date: Fri, 25 Jan 2013 09:57:17 +0200
> From: Eli Zaretskii <eliz@gnu.org>
> Cc: michael.albinus@gmx.de, 13540@debbugs.gnu.org
> 
> > > The lack of any messages suggests that the threads continue to run,
> > > but file notifications never come in, which probably means we cannot
> > > detect this problem automatically, and must rely on the users who know
> > > which volumes can and cannot be supported.
> > 
> > Ok, if Emacs cannot determine itself whether file notifications work,
> > I'll set `auto-revert-use-notify' to nil on that machine.
> > 
> > > Also, I presume that you tried to modify files on the mounted volumes,
> > > while you had GDB attached to Emacs.
> > 
> > Of course.  I started and "emacs -Q" process under GDB, and from that
> > emacs session I visited the file on the mounted volume and turned on
> > auto-revert-mode.  Then, if I modify the file with another program,
> > the emacs buffer isn't reverted and the GDB console doesn't show any
> > message.
> 
> Thanks.
> 
> I guess autorevert should allow users to specify whether notifications
> should be used on a per filesystem basis, rather than just globally
> for the entire session.

Actually, there's one more thing I'd ask you to check.  Can you put a
breakpoint inside the watch_completion function, say on this line:

  if (!io_info) <<<<<<<<<<<<<<<<<<<<<
    return;

and see if it ever breaks when you modify files in the directory that
is being watched?

The issue here is: do any notifications, even invalid ones, come in,
or not?  If we get invalid notifications, we could perhaps signal to
the main thread that the watch is not working, and autorevert could
then fall back on the old methods automatically.