bug#26126: 26.0.50; file-notify-rm-watch removes arbitrary watches

[Michael Albinus <michael.albinus@gmx.de>]

Andreas Politz <politza@hochschule-trier.de> writes:

Hi Andreas,

> inotify seems to be inode-based, so it emits notifications independent
> of the filename used.  I think that hard-links are not commonly used
> under win32, but I don't know.

Yes, inotify watches inodes.

> I we stay with string-equal, we are watching filenames, not files.
> Which is probably sufficient/OK.  Maybe this should be mentioned in the
> manual, I made a note of it.

The docstring speaks about filesystems. This is a hint, that the
behaviour of the libraries might differ.

>>> + Watching a /dir/file may receive events (e.g. touch /dir) for dir.
>>
>> Could you pls give an example?
>
> With inotify: 
>
> (let ((desc (file-notify-add-watch
>              "/tmp/file" '(attribute-change)
>              (lambda (_) (cl-assert nil)))))
>   (unwind-protect
>       (progn
>         (shell-command "touch /tmp")
>         (sit-for 3))
>     (file-notify-rm-watch desc)))

Yes, I remember. For all backends except kqueue, we watch the
directory. This is intended.

>>> + Why is the existence of kqueue checked for the handler in
>>>   file-notify-add-watch ? After all we don't know how this handler will
>>>   operate.
>>
>> Why don't we know what kqueue does?
>
> This:
>
>     (if handler
> 	;; A file name handler could exist even if there is no local
> 	;; file notification support.
> 	(setq desc (funcall
> 		    handler 'file-notify-add-watch
>                     ;; kqueue does not report file changes in
>                     ;; directory monitor.  So we must watch the file
>                     ;; itself.
>                     (if (eq file-notify--library 'kqueue) file dir)
>                     flags callback))
>
> Why should we assume that handler is somehow related to kqueue ? I think
> its just a copy and paste error.  The comment should be moved to the
> actual library call below this.

This looks like an error, indeed. I will check.

> I also wonder, if the passed argument should not always be the filename
> for which the watch was requested, as opposed to its directory.  After
> all we should not make assumptions about the abilities of the underlying
> mechanism.  For example it could work similar to kqueue, i.e. with an
> inability to watch directories.

We've discussed this years ago, maybe you find it in the archives. There
are problems when you watch only the file. This doesn't work for example
when you want to watch a file which does not exist yet. Or which
disappears, and reappears.

The agreement was to watch the upper directory. This works for all
backends except kqueue.

> Thanks for you response,
> -ap

Best regards, Michael.