Re: [PATCH 2.6] 2/3 Serio: possible race in handle_events

[Dmitry Torokhov <dtor_core@ameritech.net>]

On Saturday 23 August 2003 02:19 am, Andrew Morton wrote:
> Dmitry Torokhov <dtor_core@ameritech.net> wrote:
> > On Saturday 23 August 2003 02:00 am, Andrew Morton wrote:
> > > Dmitry Torokhov <dtor_core@ameritech.net> wrote:
> > > > +static int is_known_serio(struct serio *serio)
> > > >  +{
> > > >  +	struct serio *s;
> > > >  +
> > > >  +	list_for_each_entry(s, &serio_list, node)
> > > >  +		if (s == serio)
> > > >  +			return 1;
> > > >  +	return 0;
> > > >  +}
> > >
> > > Could this just be
> > >
> > > 	return !list_empty(&serio->node);
> > >
> > > ?
> >
> > The serio could be free()d, I dont think we want to call list_empty with
> > a dangling pointer. Or am I missing something?
>
> Well if we're playing around with a freed pointer then something is
> seriously wrong.  Like, someone could have allocated a new one and got the
> same address.
>
> If event->serio can point at freed memory and there's any doubt over it
> then we should be nulling out event->serio to indicate that.

Right now we can't as events are queued in an event list and are processed by 
other thread; serio does not know that it's queued and even existence of the
event list is not known outside of serio.c module.

Do you think we should introduce allocate_serio/free_serio pair for dynamically 
allocated serios; free_serio would scan event_list and invalidate events that 
refer to freed serio?