[PATCH v2] netfilter: nf_nat: Fix stale outer UDP checksum on VXLAN encapsulated packets

[lvjunyu <lvjunyu@cmss.chinamobile.com>]

When MASQUERADE --random-fully rewrites the outer UDP source port and
address of a VXLAN-encapsulated packet whose inner header has
CHECKSUM_PARTIAL, the outer UDP checksum is not updated correctly.

udp_set_csum() takes the LCO (Local Checksum Offload) branch for
encapsulated CHECKSUM_PARTIAL non-GSO packets, writing a complete
(folded and complemented) checksum into uh->check while leaving
csum_start pointing at the inner transport header.  In this state:

  - inet_proto_csum_replace2() is a no-op (pseudohdr=false,
    CHECKSUM_PARTIAL)
  - nf_csum_update() -> inet_proto_csum_replace4(pseudohdr=true)
    takes the pseudo-header branch which treats uh->check as an
    uncomplemented seed, applying the address delta with the wrong
    sign

Fix this by using the offload target (csum_start + csum_offset) to
distinguish the LCO state from the seed/offload state, and temporarily
flipping ip_summed to CHECKSUM_NONE so both the address and port updates
use csum_replace*() (the complete-checksum path).

Test results (10 connections, single netns vxlan + veth environment):
  Before fix: ~1080ms avg first-connection latency, 10 SYN retransmits,
              UdpInCsumErrors incremented per connection
  After fix:  ~29ms avg first-connection latency, 0 SYN retransmits,
              UdpInCsumErrors unchanged

Fixes: faec18dbb0405 ("netfilter: nat: remove l4proto->manip_pkt")
Cc: stable@vger.kernel.org
Signed-off-by: lvjunyu <lvjunyu@cmss.chinamobile.com>

---

Changes in v2:
- Fix both address and port delta (v1 only fixed port delta)
- Use offload-target test instead of skb->encapsulation as the LCO
  discriminator (addresses Sashiko AI review feedback)
- Update comment to describe both CHECKSUM_PARTIAL states

v1: https://lore.kernel.org/netdev/20260920074543.525572-1-lvjunyu@cmss.chinamobile.com/
---
 net/netfilter/nf_nat_proto.c | 26 ++++++++++++++++++++++++++
 1 file changed, 26 insertions(+)

diff --git a/net/netfilter/nf_nat_proto.c b/net/netfilter/nf_nat_proto.c
index 64b9bac228e..cb70d85716b 100644
--- a/net/netfilter/nf_nat_proto.c
+++ b/net/netfilter/nf_nat_proto.c
@@ -54,9 +54,35 @@ __udp_manip_pkt(struct sk_buff *skb,
 		portptr = &hdr->dest;
 	}
 	if (do_csum) {
+		/* When udp_set_csum() takes the LCO branch (encapsulated
+		 * CHECKSUM_PARTIAL, non-GSO), uh->check holds a complete
+		 * (folded and complemented) checksum while csum_start
+		 * points at the inner transport header, not at this outer
+		 * udphdr.  In that state, inet_proto_csum_replace*() and
+		 * nf_csum_update() take the pseudo-header branch which
+		 * operates on an uncomplemented seed, applying the address
+		 * and port deltas with the wrong sign.  Temporarily flip
+		 * ip_summed so they use csum_replace*() instead.
+		 *
+		 * The seed/offload branch of udp_set_csum() sets csum_start
+		 * to the outer UDP header, so the offload-target test
+		 * below distinguishes the two states.
+		 */
+		bool lco = skb->ip_summed == CHECKSUM_PARTIAL &&
+			   !skb_is_gso(skb) &&
+			   (skb->head + skb->csum_start + skb->csum_offset) !=
+			   (unsigned char *)&hdr->check;
+
+		if (lco)
+			skb->ip_summed = CHECKSUM_NONE;
+
 		nf_csum_update(skb, iphdroff, &hdr->check, tuple, maniptype);
 		inet_proto_csum_replace2(&hdr->check, skb, *portptr, newport,
 					 false);
+
+		if (lco)
+			skb->ip_summed = CHECKSUM_PARTIAL;
+
 		if (!hdr->check)
 			hdr->check = CSUM_MANGLED_0;
 	}
--
2.43.0