Re: [PATCH net 2/3] udp: check encap socket in __udp_lib_err

On Mon, 2021-07-12 at 15:05 +0100, Vadim Fedorenko wrote:

On 12.07.2021 14:37, Paolo Abeni wrote:

quoted

On Mon, 2021-07-12 at 13:45 +0100, Vadim Fedorenko wrote:

quoted

quoted

After this patch, the above chunk will not clear 'sk' for packets
targeting ESP in UDP sockets, but AFAICS we will still enter the
following conditional, preserving the current behavior - no ICMP
processing.

We will not enter following conditional for ESP in UDP case because
there is no more check for encap_type or encap_enabled.

I see. You have a bug in the ipv6 code-path. With your patch applied:

---
  	sk = __udp6_lib_lookup(net, daddr, uh->dest, saddr, uh->source,
                                inet6_iif(skb), inet6_sdif(skb), udptable, NULL);
         if (sk && udp_sk(sk)->encap_enabled) {
		//...
         }

         if (!sk || udp_sk(sk)->encap_enabled) {
	// can still enter here...
---	

Oh, my bad, thanks for catching this!

quoted

quoted

I maybe missing something but d26796ae5894 doesn't actually explain
which particular situation should be avoided by this additional check
and no tests were added to simply reproduce the problem. If you can
explain it a bit more it would greatly help me to improve the fix.

Xin knows better, but AFAICS it used to cover the situation you
explicitly tests in patch 3/3 - incoming packet with src-port == dst-
port == tunnel port - for e.g. vxlan tunnels.

Ok, so my assumption was like yours, that's good.

quoted

quoted

quoted

Why can't you use something alike the following instead?

---

diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
index c0f9f3260051..96a3b640e4da 100644
--- a/net/ipv4/udp.c
+++ b/net/ipv4/udp.c

@@ -707,7 +707,7 @@ int __udp4_lib_err(struct sk_buff *skb, u32 info, struct udp_table *udptable)
          sk = __udp4_lib_lookup(net, iph->daddr, uh->dest,
                                 iph->saddr, uh->source, skb->dev->ifindex,
                                 inet_sdif(skb), udptable, NULL);
-       if (!sk || udp_sk(sk)->encap_type) {
+       if (!sk || READ_ONCE(udp_sk(sk)->encap_err_lookup)) {
                  /* No socket for error: try tunnels before discarding */
                  sk = ERR_PTR(-ENOENT);
                  if (static_branch_unlikely(&udp_encap_needed_key)) {

---

Could you please have a look at the above ?

Sure. The main problem I see here is that udp4_lib_lookup in udp_lib_err_encap
could return different socket because of different source and destination port
and in this case we will never check for correctness of originally found socket,
i.e. encap_err_lookup will never be called and the ICMP notification will never
be applied to that socket even if it passes checks.
My point is that it's simplier to explicitly check socket that was found than
rely on the result of udp4_lib_lookup with different inputs and leave the case
of no socket as it was before d26796ae5894.

If it's ok, I will unify the code for check as Willem suggested and resend v2.

If the final code is small enough, please go ahead with that.

Thanks!

Paolo