On Tue, 13 Mar 2007, Rusty Russell wrote:
> Hi Trent,
>
> 	Patch looks good, just one comment:
>
> On Mon, 2007-03-12 at 07:07 -0700, Trent Piepho wrote:
> > +       use = already_uses(a, b);
> > +       if (!use) {
> > +           printk(KERN_ERR "module %s trying to un-use a module, %s, which "
> > +                  "it is not using", a->name, b->name);
> > +            return 0;
> > +       }
>
> s/return 0/BUG()/.  This is potentially quite a nasty bug.

Ok, I did that before, I'll change it back.

Note that the reference counting isn't perfect when it comes to catching
mistakes.

The fundamental problem is that when a module is loaded and linked, all the
modules that it used symbols from gain a "use".  To be symmetric, when a
module is unloaded all the modules it used symbols from should lose a
"use".  Except, there is no record of what modules gained a "use" at link
time.  Suppose module 1 uses a symbol from module 2.  At link time, a
module_use that "1 uses 2" is created.  Now say 1 does a symbol_put() on
something in 2, with no matching get.  The "1 uses 2" goes away.  When 1 is
unloaded, there is no way to tell that "1 uses 2", deleted by the extra
put, is missing.

If it's wanted, I think I could fix this.  I'd have a separate count of
static uses vs dynamic uses.