Re: [PATCH] module: Fix race condition between load and unload module

[Al Viro <viro@ZenIV.linux.org.uk>]

On Fri, Apr 12, 2013 at 04:47:50PM -0700, Linus Torvalds wrote:
> This is a much more generic bug in kobjects, and I would hate to add
> some random workaround for just one case of this bug like you do. The
> more fundamental bug needs to be fixed too.
> 
> I think the more fundamental bugfix is to just fix kobject_get() to
> return NULL if the refcount was zero, because in that case the kobject
> no longer really exists.
> 
> So instead of having
> 
>     kref_get(&kobj->kref);
> 
> it should do
> 
>     if (!atomic_inc_not_zero(&kobj->kref.refcount))
>         kobj = NULL;
> 
> and I think that should fix your race automatically, no? Proper patch
> attached (but TOTALLY UNTESTED - it seems to compile, though).
> 
> The problem is that we lose the warning for when the refcount is zero
> and somebody does a kobject_get(), but that is ok *assuming* that
> people actually check the return value of kobject_get() rather than
> just "know" that if they passed in a non-NULL kobj, they'll get it
> right back.
> 
> Greg - please take a look... I'm adding Al to the discussion too,
> because Al just *loooves* these kinds of races ;)

Unless I'm misreading what's going on, we have the following to thank for that:
        /* remove from sysfs if the caller did not do it */
        if (kobj->state_in_sysfs) {
                pr_debug("kobject: '%s' (%p): auto cleanup kobject_del\n",
                         kobject_name(kobj), kobj);
                kobject_del(kobj);
        }
in kobject_cleanup().  Why don't we require kobject_del() before the final
kobject_put(), if the sucker had been added?  FWIW, I thought it *was*
required all along...