Re: Question on lockdep and MAX_LOCK_DEPTH

[Ingo Molnar <mingo@kernel.org>]

* Steven Rostedt <rostedt@goodmis.org> wrote:

> On Wed, 2013-02-06 at 07:56 -0800, Ben Greear wrote:
> 
> > > I'm 99% sure that the bug is in your modifications.
> > 
> > I'm sorry, I tried to make that clear.
> 
> You said it was an out of tree module, I didn't realize it had changes
> to the core Linux as well.
> 
> > My tree is here, minus a few debugging patches related to 
> > this bug:
> > 
> > http://dmz2.candelatech.com/git/gitweb.cgi?p=linux-3.7.dev.y/.git;a=summary
> 
> Here's your bug:
> 
> 3319         list_for_each_entry_rcu(ptype, &ptype_all, list) {
> 3320                 if (!ptype->dev || ptype->dev == skb->dev) {
> 3321                         if (pt_prev) {
> 3322                                 ret = deliver_skb(skb, pt_prev, orig_dev);
> 3323                                 if (ret == NET_RX_CONSUMED) {
> 3324                                         kfree_skb(skb); /* we made a copy in deliver_skb */
> 3325                                         goto out;
> 3326                                 }
> 3327                         }
> 
> One of many. All those "goto out;"s skip the rcu_read_unlock().

Wondering whether we could improve lockdep to detect 'excessive' 
lock nesting and treat it as a bug - the 'ran out of lockdep 
entries' assert might be too detached and it might even trigger 
in innocent, low lesting count contexts.

Thanks,

	Ingo