Re: [bug] __nf_ct_refresh_acct(): WARNING: at lib/list_debug.c:30 __list_add+0x7d/0xad()

Patrick McHardy a écrit :

Patrick McHardy wrote:

quoted

Eric Dumazet wrote:

quoted

quoted

a test-box still triggered this crash overnight:

[  252.433471] ------------[ cut here ]------------
[  252.436031] WARNING: at lib/list_debug.c:30 __list_add+0x95/0xa0()
...
With your patch (repeated below) applied. Is Patrick's alternative
patch supposed to fix something that yours does not?

Hmm, not really, Patrick patch should fix same problem, but without
extra locking
as mine.

This new stack trace is somewhat different, as corruption is detected
in the add_timer()
call in __nf_conntrack_confirm()

So there is another problem. CCed Pablo Neira Ayuso who added some stuff
in netfilter and timeout logic recently.

That timeout logic shouldn't be relevant in this case, its only
activated when netlink event delivery is used, a userspace process
is actually listening and it has the socket marked for reliable
delivery.

I think its still the same problem, the detection is just noticed
at a different point.

I can't find anything wrong with the conntrack timer use itself.
Since its the timer base that is corrupted, its possible that
something else is using timers incorrectly and conntrack just
happens to hit the problem afterwards, but it seems rather
unlikely since the first backtrace originated in a network driver,
the second one in the socket layer.

But I've noticed a different problem, the lockless lookup might
incorrectly return an conntrack entry that is supposed to be
invisible. This can happen because we only check for equal tuples
and !atomic_inc_not_zero(refcnt), but the conntrack can be removed
from the lists and still have references to it. It should never
have an active timer at that point however, so all following
mod_timer_pending() calls *should* do nothing unless I'm missing
something.

Ingo, could you please try whether this patch (combined with the
last one) makes any difference? Enabling CONFIG_NETFILTER_DEBUG
could also help.

Interesting !

In my own analysis, I found death_by_timeout() might be problematic,
with RCU and lockless lookups.

static void death_by_timeout(unsigned long ul_conntrack)
{
        struct nf_conn *ct = (void *)ul_conntrack;

        if (!test_bit(IPS_DYING_BIT, &ct->status) &&
            unlikely(nf_conntrack_event(IPCT_DESTROY, ct) < 0)) {
                /* destroy event was not delivered */
                nf_ct_delete_from_lists(ct);
<< HERE >>

                nf_ct_insert_dying_list(ct);
                return;
        }
        set_bit(IPS_DYING_BIT, &ct->status);
        nf_ct_delete_from_lists(ct);
        nf_ct_put(ct);
}


We delete ct from a list and insert it in a new list.

I believe a reader could "*catch*" ct while doing a lookup and miss the end
of its chain. (nulls algo check the null value at the end of lookup and can
decide to restart the lookup if the null value is not the expected one)

We need to change nf_conntrack_init_net() and use a different "null" value,
guaranteed not being used in regular lists

Patch follows :

diff --git a/net/netfilter/nf_conntrack_core.c b/net/netfilter/nf_conntrack_core.c
index 5f72b94..5276a2d 100644
--- a/net/netfilter/nf_conntrack_core.c
+++ b/net/netfilter/nf_conntrack_core.c

@@ -1267,13 +1267,19 @@ err_cache:
 	return ret;
 }
 
+/*
+ * We need to use special "null" values, not used in hash table
+ */
+#define UNCONFIRMED_NULLS_VAL	((1<<30)+0)
+#define DYING_NULLS_VAL		((1<<30)+1)
+
 static int nf_conntrack_init_net(struct net *net)
 {
 	int ret;
 
 	atomic_set(&net->ct.count, 0);
-	INIT_HLIST_NULLS_HEAD(&net->ct.unconfirmed, 0);
-	INIT_HLIST_NULLS_HEAD(&net->ct.dying, 0);
+	INIT_HLIST_NULLS_HEAD(&net->ct.unconfirmed, UNCONFIRMED_NULLS_VAL);
+	INIT_HLIST_NULLS_HEAD(&net->ct.dying, DYING_NULLS_VAL);
 	net->ct.stat = alloc_percpu(struct ip_conntrack_stat);
 	if (!net->ct.stat) {
 		ret = -ENOMEM;