Re: [PATCH net-2.6 v2] pkt_sched: gen_estimator: add a new lock

[Eric Dumazet <eric.dumazet@gmail.com>]

Le mercredi 09 juin 2010 à 11:33 +0000, Jarek Poplawski a écrit :
> On Wed, Jun 09, 2010 at 11:39:10AM +0200, Eric Dumazet wrote:
> ...
> > Here is v2 of patch, adding protection in gen_estimator_active() as
> > well.
> > 
> > [PATCH net-2.6 v2] pkt_sched: gen_estimator: add a new lock
> > 
> > gen_kill_estimator() / gen_new_estimator() is not always called with
> > RTNL held.
> > 
> > net/netfilter/xt_RATEEST.c is one user of these API that do not hold
> > RTNL, so random corruptions can occur between "tc" and "iptables".
> > 
> > Add a new fine grained lock instead of trying to use RTNL in netfilter.
> 
> Btw, maybe it's a different case, but RTNL happens in netfilter already
> (nf_conntrack_helper.c, nf_conntrack_proto.c).
> 
> It would be nice to mention Changli's argument that
> gen_replace_estimator isn't atomic operation after this change.
> 

See my next comment. This function is still used in qdisc domain only.
We could add ASSERT_RTNL() here, but its not a bug fix...

> > @@ -312,8 +315,14 @@ EXPORT_SYMBOL(gen_replace_estimator);
> >  bool gen_estimator_active(const struct gnet_stats_basic_packed *bstats,
> >  			  const struct gnet_stats_rate_est *rate_est)
> >  {
> > +	bool res;
> > +
> >  	ASSERT_RTNL();
> 
> This line should go away as well.
> 

Nope. This is really needed for safety. Of course, its a debugging knob
and you can remove it if you trust all callers to be good.

Caller should still use RTNL here, or else, as soon as we exit from
gen_estimator_active() with a 'true' result, some other qdisc user could
destroy the estimator.

Fortunatly up to 2.6.35, xt_RATEEST cannot reuse/delete an estimator
created by a qdisc user, and it doesnt use gen_estimator_active().

In a future patch, we should add RCU safety here instead of ASSERT_RTNL.
Thats to make sure caller of gen_estimator_active() is inside a
rcu_read_lock() section, and estimator cannot disappear under it, even
if another cpu/thread calls gen_kill_estimator() on same estimator.