Re: [PATCH v2 03/10] cgroup: net_cls: Protect access to task_cls_classid() when built as module

[Daniel Wagner <wagi-kQCPcA+X3s7YtjvyW6yDsg@public.gmane.org>]

On Tue, Aug 28, 2012 at 07:47:25AM -0700, Paul E. McKenney wrote:
> On Sat, Aug 25, 2012 at 06:56:29PM +0200, Daniel Wagner wrote:
> > On 25.08.2012 01:26, Tejun Heo wrote:
> > >On Fri, Aug 24, 2012 at 04:01:37PM +0200, Daniel Wagner wrote:
> > >>@@ -306,6 +312,11 @@ static void __exit exit_cgroup_cls(void)
> > >>  	synchronize_rcu();
> > >>  #endif
> > >>
> > >>+#if IS_MODULE(CONFIG_NET_CLS_CGROUP)
> > >>+	static_key_slow_dec(&cgroup_cls_enabled);
> > >>+	rcu_barrier();
> > >
> > >Why is this rcu_barrier() necessary?
> > 
> > I have read the rcubarrier.txt document and I got from that that an
> > rcu_barrier() is needed when unloading a module. But maybe I got it
> > wrong.
> > 
> > So the idea after disabling the jump lables all pending readers in
> > task_cls_classid() have left. THe same thing is done in the old code
> > with the dynamic id part. With the difference that synchronize_rcu()
> > is used.
> 
> FWIW, the rcu_barrier() is needed only if the module uses call_rcu().
> In that case it is required to ensure that all the resulting callbacks
> execute before the module's .text is freed up.

Thanks for the clarification. After reading the documentations again
and I think this here should do the trick:

static inline u32 task_cls_classid(struct task_struct *p)
{
	struct cgroup_cls_state *cs;
	u32 classid = 0;

	if (!clscg_enabled || in_interrupt())
		return 0;

	rcu_read_lock();
	cs = container_of(task_subsys_state(p, net_cls_subsys_id),
			struct cgroup_cls_state, css);
	if (cs)
		classid = cs->classid;
	rcu_read_unlock();

	return classid;
}

static void __exit exit_cgroup_cls(void)
{
	unregister_tcf_proto_ops(&cls_cgroup_ops);

#if IS_MODULE(CONFIG_NET_CLS_CGROUP)
	static_key_slow_dec(&cgroup_cls_enabled);
	synchronize_rcu();
#endif

	cgroup_unload_subsys(&net_cls_subsys);
}

So when unloading the module, we first disable the static branch, then
we wait for all old readers leaving the reader side issuing a
synchronize_rcu(). New readers might already have passed the static
branch and now entering the reader side. So we need to test if the
pointer we retrieve is valid. Basically, this change is using the same
approach we had before. Instead looking at the id is valid we look at
the pointer if it is valid.