Re: Packets being dropped somewhere in the kernel, between iptables and packet capture layers

[Florian Westphal <fw@strlen.de>]

Niklas Hambüchen <mail@nh2.me> wrote:
> I'm sending this to netdev@vger.kernel.org even though http://vger.kernel.org/lkml/ still suggests linux-net@vger.kernel.org, because the latter seems to be inactive since 2011 and full of spam, and I got "unresolvable address" for it. Perhaps somebody should update the page that recommends it.
> Nevertheless, please let me know if here is the wrong place.

This problem is known; I asked for test feedback on this patch but never
got a response:

    netfilter: nf_nat: return the same reply tuple for matching CTs
    
    It is possible that two concurrent packets originating from the same
    socket of a connection-less protocol (e.g. UDP) can end up having
    different IP_CT_DIR_REPLY tuples which results in one of the packets
    being dropped.
    
    To illustrate this, consider the following simplified scenario:
    
    1. No DNAT/SNAT/MASQUEARADE rules are installed, but the nf_nat module
       is loaded.
    2. Packet A and B are sent at the same time from two different threads
       via the same UDP socket which hasn't been used before (=no CT has
       been created before). Both packets have the same IP_CT_DIR_ORIGINAL
       tuple.
    3. CT of A has been created and confirmed, afterwards get_unique_tuple
       is called for B. Because IP_CT_DIR_REPLY tuple (the inverse of
       the IP_CT_DIR_ORIGINAL tuple) is already taken by the A's confirmed
       CT (nf_nat_used_tuple finds it), get_unique_tuple calls UDP's
       unique_tuple which returns a different IP_CT_DIR_REPLY tuple (usually
       with src port = 1024)
    4. B's CT cannot get confirmed in __nf_conntrack_confirm due to
       the found IP_CT_DIR_ORIGINAL tuple of A and the different
       IP_CT_DIR_REPLY tuples, thus the packet B gets dropped.
    
    This patch modifies nf_conntrack_tuple_taken so it doesn't consider
    colliding reply tuples if the IP_CT_DIR_ORIGINAL tuples are equal.
    
    Then, at insert time, either clash resolution is possible (new packet
    has the existing/older conntrack assigned to it), or it has to be
    dropped.
    
diff --git a/net/netfilter/nf_conntrack_core.c b/net/netfilter/nf_conntrack_core.c
index 741b533148ba..07847a612adf 100644
--- a/net/netfilter/nf_conntrack_core.c
+++ b/net/netfilter/nf_conntrack_core.c
@@ -1007,6 +1007,22 @@ nf_conntrack_tuple_taken(const struct nf_conntrack_tuple *tuple,
 		}
 
 		if (nf_ct_key_equal(h, tuple, zone, net)) {
+			/* If the origin tuples are identical, we can ignore
+			 * this clashing entry: they refer to the same flow.
+			 * Do not apply nat clash resolution in this case and
+			 * let nf_ct_resolve_clash() deal with this.
+			 *
+			 * This can happen with UDP in particular, e.g. when
+			 * more than one packet is sent from same socket in
+			 * different threads.
+			 *
+			 * We would now mangle our entry and would then have to
+			 * discard it at conntrack confirm time.
+			 */
+			if (nf_ct_tuple_equal(&ignored_conntrack->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
+					      &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple))
+				continue;
+
 			NF_CT_STAT_INC_ATOMIC(net, found);
 			rcu_read_unlock();
 			return 1;