Re: [PATCH v3 net 11/16] ipv6: input: convert to dev_net_rcu()

[Jakub Kicinski <kuba@kernel.org>]

On Tue, 4 Feb 2025 13:17:08 -0800 Paul E. McKenney wrote:
> > > TBH I'm slightly confused by this, and the previous warnings.
> > >
> > > The previous one was from a timer callback.
> > >
> > > This one is with BH disabled.
> > >
> > > I thought BH implies RCU protection. We certainly depend on that
> > > in NAPI for XDP. And threaded NAPI does the exact same thing as
> > > xfrm_trans_reinject(), a bare local_bh_disable().
> > >
> > > RCU folks, did something change or is just holes in my brain again?  
> > 
> > Nope, BH does not imply rcu_read_lock()  
> 
> You are both right?  ;-)
> 
> The synchronize_rcu() function will wait for all types of RCU readers,
> including BH-disabled regions of code.  However, lockdep can distinguish
> between the various sorts of readers.  So for example
> 
> 	lockdep_assert_in_rcu_read_lock_bh();
> 
> will complain unless you did rcu_read_lock_bh(), even if you did something
> like disable_bh().  If you don't want to distinguish and are happy with
> any type of RCU reader, you can use
> 
> 	lockdep_assert_in_rcu_reader();
> 
> I have been expecting that CONFIG_PREEMPT_RT=y kernels will break this
> any day now, but so far so good.  ;-)

Thanks Paul! So IIUC in this case we could:

diff --git a/include/net/net_namespace.h b/include/net/net_namespace.h
index 0f5eb9db0c62..58ec1eb9ae6a 100644
--- a/include/net/net_namespace.h
+++ b/include/net/net_namespace.h
@@ -401,7 +401,7 @@ static inline struct net *read_pnet(const possible_net_t *pnet)
 static inline struct net *read_pnet_rcu(possible_net_t *pnet)
 {
 #ifdef CONFIG_NET_NS
-	return rcu_dereference(pnet->net);
+	return rcu_dereference_check(pnet->net, rcu_read_lock_bh_held());
 #else
 	return &init_net;
 #endif

Sorry for the sideline, Eric, up to you how to proceed..
I'll try to remember the details better next time :)