Re: [PATCH -v3 5/8] fsnotify: unified filesystem notification backend

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Fri, 2008-11-28 at 18:22 -0500, Eric Paris wrote:
> On Thu, 2008-11-27 at 17:20 +0100, Peter Zijlstra wrote:
> > On Tue, 2008-11-25 at 12:21 -0500, Eric Paris wrote:
> > > +int fsnotify_check_notif_queue(struct fsnotify_group *group)
> > > +{
> > > +       mutex_lock(&group->notification_mutex);
> > > +       if (!list_empty(&group->notification_list))
> > > +               return 1;
> > > +       mutex_unlock(&group->notification_mutex);
> > > +       return 0;
> > > +}
> > 
> > > +void fsnotify_clear_notif(struct fsnotify_group *group)
> > > +{
> > > +       struct fsnotify_event *event;
> > > +
> > > +       while (fsnotify_check_notif_queue(group)) {
> > > +               event = get_event_from_notif(group);
> > > +               fsnotify_put_event(event);
> > > +               /* fsnotify_check_notif_queue() took this lock */
> > > +               mutex_unlock(&group->notification_mutex);
> > > +       }
> > > +}
> > 
> > That is quite horrible, please just open code that to keep the locking
> > symmetric.
> 
> While horrible, I use fsnotify_check_notif_queue in my fsnotify (not in
> this series as this only includes dnotify) has
> 
> wait_event_interruptible(group->notification_waitq, fanotify_check_notif_queue(group));
> 
> So I wouldn't know how to open code that...  I can open code this
> instance, but it's going to mean redoing all of that other code to
> handle having thing not be present when we return.  Since I didn't
> submit that as well I guess I'm not allowed to use it as a reason...

Or you add a lock parameter to wait_event*() which gets unlocked before
schedule and locks again afterwards.

That would allow you to write it like so:

 mutex_lock(&group->notification_mutex);
 wait_event_interruptible_lock(group->notification_waitq,
                               !list_empty(&group_notificatioin_list), 
                               &group_notification_mutex);

 /* handle the !empty list */
 mutex_unlock(&group->notification_mutex);

You could use the type matching magic we have to select between
spinlock/mutex operations for the lock argument.

I've come across such a pattern a few times, most of the times we end up
open coding the wait_event stuff.