- userspace wrote epoch ns while the kernel compares ban expiry against
bpf_ktime_get_ns(), so gateway/detector bans never expired (v4 and v6).
xdp::monotonic_ns() now reads CLOCK_MONOTONIC.
- universal_filter.c required sizeof(tcphdr) (20 bytes) of L4 for every
protocol, so short UDP datagrams took an early XDP_PASS and bypassed
the blacklist (v4 and v6). Minimum now depends on the protocol.
- tests/xdp_ban_clock.rs guards the clock choice.
- xdp_ipv6_smoke.sh: v6 pass/ban/unban, short datagrams, v4 short
datagrams and real 2s ban expiry on a live kernel.
- C: lpm_key6 + blacklist6_map (LPM_TRIE, 20-byte key); native IPv6 no
longer returns XDP_PASS before filtering: same L4/port gates, then
blacklist6, UDP drop-policy, TCP flag checks and /64 prefix stats.
- Rust: blacklist6_key/blacklist6_update; ban_cidr handles V4 and V6;
gateway xdp_ban uses ban_cidr (/32 or /128); dead ban_ip removed.
- tests/xdp_ban_v6.rs (key layout) and scripts/xdp_ipv6_smoke.sh
(real kernel: pass -> ban drops -> unban passes).
- TODO: record done items and two findings (userspace ban expiry clock
mismatch; short-L4 bypass of blacklist).
Found on real kernel 5.15 (netns+veth, challenge enabled):
1. truncated IP: after bpf_xdp_adjust_tail the IPv4 tot_len / IPv6
payload_len still claimed the original SYN size — peer silently
dropped the malformed SYN-ACK, challenge never seen
2. endianness: marker/bad_ack written as raw host-order halfwords into
network-order tcpw[] — client echoed swap16halves(bad_ack) in its
RST.seq, verification always failed (CHAL_FAILED == CHAL_SENT)
Post-fix e2e: SYN -> bad-ACK SYN-ACK (XDP_TX) -> client RST ->
verified -> retry connects transparently; counters SENT=1 VERIFIED=1
FAILED=0