Re: [2.6.14-rt1] slowdown / oops.

[Ingo Molnar <mingo@elte.hu>]

* Ingo Molnar <mingo@elte.hu> wrote:

> more updates: NETFILTER_DEBUG catches the situation too - the problem 
> seems to be wrong reference counts on the skb. [...]

ok, could you check whether the patch below fixes the problem for you?  
(I have also put it into -rt4)

local_bh_disable()/enable() is a NOP under PREEMPT_RT, and the 
ip_ct_deliver_cached_events PER_CPU code relies on not being preempted 
by the net_rx_action softirq handler. So this is a bug in PREEMPT_RT and 
the upstream code should be fine.

	Ingo

Index: linux/net/ipv4/netfilter/ip_conntrack_core.c
===================================================================
--- linux.orig/net/ipv4/netfilter/ip_conntrack_core.c
+++ linux/net/ipv4/netfilter/ip_conntrack_core.c
@@ -105,11 +105,11 @@ void ip_ct_deliver_cached_events(const s
 {
 	struct ip_conntrack_ecache *ecache;
 	
-	local_bh_disable();
+	read_lock_bh(&ip_conntrack_lock);
 	ecache = &__get_cpu_var(ip_conntrack_ecache);
 	if (ecache->ct == ct)
 		__ip_ct_deliver_cached_events(ecache);
-	local_bh_enable();
+	read_unlock_bh(&ip_conntrack_lock);
 }
 
 void __ip_ct_event_cache_init(struct ip_conntrack *ct)