bug#21432: 25.0.50; file-notify-rm-watch (inotify) errors if watched dir is deleted

[Eli Zaretskii <eliz@gnu.org>]

> From: Tassilo Horn <tsdh@gnu.org>
> Cc: michael.albinus@gmx.de,  21432@debbugs.gnu.org
> Date: Mon, 14 Sep 2015 09:56:12 +0200
> 
> >> AFAIR, filenotify.el watches the parent directory of the
> >> file/directory you asked it to watch.  So to see the invalid-p method
> >> in action you need to remove the parent, not the directory itself.
> >
> > Ah, I see.  I'll adapt the tests accordingly.
> 
> Done, however the descriptors still don't become invalid when deleting
> the parent directory of the watched file or directory.  I guess, that's
> a problem in the inotify and TRAMP implementations then.

No, I think it's a problem with our mental model of what happens.  The
file notifications use the Emacs event loop, and Emacs won't check for
events until it's idle.  So calling file-notify-valid-p as part of the
test ends up doing that _before_ the directory deletion notification
is read by Emacs and invalidates the watch.  I actually see the
message saying the watch is valid before the notification comes in and
its message is inserted into *Messages*.

The following simple test case works with w32notify:

  (setq mydir "/tmp/x")
  (make-directory mydir t)
  (setq myfile "/tmp/x/y")
  (write-region "foo\n" nil myfile)
  (setq w (w32notify-add-watch mydir '(file-name attributes size last-write-time)
			       (lambda (event) (message "%s" event))))
  (message "valid: %s" (w32notify-valid-p w))
  (delete-directory mydir t)
  (run-with-idle-timer 1 nil
		       (lambda ()
			 (message "valid: %s" (w32notify-valid-p w))))

When I run the above snippet with eval-region, I see these messages in
*Messages*:

  Wrote d:/tmp/x/y
  valid: t
  (100161480 modified y)
  (100161480 removed y)
  valid: nil

Try something similar with inotify and see if you see the same basic
issue.  If you do, I trust you will think of a way to modify the tests
so that validation does do its thing.