- Inventory nodes/servers use cursor SCAN plus MGET batches instead of KEYS
and per-key GET; a Redis error returns 503 instead of an empty fleet.
- GET /blacklist no longer writes: the legacy metadata backfill moved to a
one-shot startup task. The list endpoint paginates (limit 100, max 1000,
offset max 10000) instead of dumping SMEMBERS.
- Router-wide DefaultBodyLimit 8 KiB and TimeoutLayer 10 s; every mutating
endpoint is rate limited (30/min per client IP) with idle eviction, and the
login limiter resets on success. reason > 256 bytes and malformed targets
are rejected with 400.
- SIGTERM/Ctrl-C trigger graceful shutdown; the heartbeat and eviction tasks
are cancelled and joined so in-flight ban edits reach Redis.
- MANAGER_TRUSTED_PROXIES enables X-Forwarded-For (right-to-left walk) only
for configured proxy networks; otherwise the TCP peer is used.
- tests/manager_api.rs covers 401/400/413/429/503 and CORS preflight through
the real router; manager_security gains limiter and proxy-IP tests.
- bind.address is parsed as IpAddr (no ":::port"), the socket domain follows
the family, and "::" binds dual-stack. IPv6 was impossible before.
- limits.relay_idle_timeout_secs (default 300, 1..=86400) replaces the
hardcoded relay constant; the deploy examples document it.
- The connection budget is an Arc<Semaphore> owned by Gateway instead of a
process-wide OnceLock, so tests and multiple gateways stay independent.
- PROXY header identity is validated: unspecified, loopback, link-local,
multicast, broadcast and addresses inside trusted_cidrs are rejected, and
IPv4-mapped IPv6 is normalised first. An invalid header falls back to the
transport peer. Bans targeting trusted_cidrs are suppressed, closing the
ban-amplification path through a trusted load balancer.
- TrustedProxyCidr::contains no longer overflows its mask for IPv6 prefixes
shorter than /96.
Security:
- An empty API_PASSWORD passed validation (only the literal "changeme" was
rejected) and verify_password("", "") returned true, so a login with an
empty password minted an admin token. Require >= 16 characters and >= 32
non-whitespace bytes for JWT_SECRET; verify rejects an empty stored secret.
- syn_challenge_enabled required a real challenge_secret_hex: the public
placeholder from config.h is now rejected in every spelling.
- docker-compose: Redis/ClickHouse/Grafana are loopback-only and every
password uses ${VAR:?} instead of a default. An exposed Redis was a way to
lift bans and publish unbans to the whole fleet with one redis-cli.
- Empty CORS_ORIGIN no longer means "allow any origin".
- ban_cidr / userspace blacklist validate and mask prefixes; CIDR bans are
stored as ranges, so a /24 or /64 ban actually covers the subnet (before,
only the network address was blocked).
Availability:
- ClickHouse client has connect/request timeouts and no mutex held across
await: a hung analytics backend can no longer freeze the ban path. Events
go through a bounded buffer with a drop counter.
- Redis uses a ConnectionManager with per-command timeouts; pubsub and
reconcile reset the backoff after success; publish cannot block shutdown.
- Ban durations from pubsub are clamped and use checked_add; the sync task
survives panics instead of silently dying.
- Blacklist TOCTOU (remove_if) and replace_redis no longer empties the table
during a Redis reconnect.
- RateLimiter and reputation have hard caps with incremental eviction instead
of an unbounded DashMap and a stop-the-world retain.
- The attack detector scores before updating its baseline and needs several
quiet windows to leave UnderAttack, so a sustained flood keeps banning new
offenders instead of ending after one tick.
- accept() backs off on EMFILE/ENFILE, shutdown Err is treated as shutdown,
a global semaphore bounds in-flight connections, rate limiting runs before
PoW, whitelisted IPs are never banned, and connect/handshake/relay have
timeouts with an idle budget.
- PoW uses the configured difficulty as a floor, the clamp is enforced, and
nonce framing survives split segments and coalesced handshake bytes.
- Subnet tracker memory is capped (prefixes and per-prefix sources) and the
monitor bans with the configured TTL.
CI and docs:
- CI runs the test suite with --no-default-features --features protocol-http;
make ebpf fails on errors; the secrets gate catches weak compose defaults.
- doctor skips protected endpoints without a token and exits non-zero on
failure; a drain timeout returns exit code 1.
- TODO §2b records the review findings, what is fixed and what is deferred;
configuration/testing docs describe the new XDP and secret semantics.
BPF program (the first two made the filter wrong on real hardware):
- Compute the TCP payload length from ip->tot_len (IPv4) / payload_len
(IPv6) instead of data_end. Ethernet padding on a real NIC added 6 bogus
payload bytes, broke expected_seq and dropped every connection after the
fourth segment. scripts/xdp_pkt_test.sh reproduces it on synthetic frames
and fails against the old code.
- Read map values (conn->fails, sent_at) before bpf_map_delete_elem; the LRU
element is freed by the delete.
- Parse up to two VLAN/QinQ tags before L3; fragments are checked against the
blacklist and passed instead of bypassing all filtering; first fragments are
no longer dropped.
- Unknown non-SYN flows PASS: the program is a stateless filter fed by the
conntrack state machine, not a firewall, so spoofed SYNs cannot evict
established flows into a drop and replies to the daemon's own outbound TCP
keep working.
- Reset conn->fails on in-order segments, honour SYN payload (TFO), fix the
IPv6 SYN-ACK address swap, count every DROP path, drop SYN with IP options.
- Expired prefix bans are no longer deleted with the wrong LPM key; userspace
reaps them with the exact key.
Loader:
- Remember the effective attach flags (native vs generic) and use them for
detach/query, verifying the interface is clean afterwards. The old code
always detached the native slot, leaving a generic program attached after
shutdown.
- DetachGuard rolls back the attach if anything after it fails.
- ban_cidr validates the prefix length and masks host bits.
- Whitelist maps (v4 and the new v6 LPM trie) are populated from the config at
startup and fail fast on error.
- Stats read all 13 counters, and a test asserts every map's key/value size
against the Rust encodings.
Tests: xdp_pkt_test (padding, VLAN/QinQ, fragments), xdp_loader (attach flags,
rollback, detach), xdp_ban_keys, xdp_whitelist_keys, xdp_services, live
xdp_attach_smoke (ignored, needs root).
XDP behaviour:
- udp_rate_limit6 (LRU, full 128-bit source key) plus throttle_hit6, so
udp_policy = "rate-limit" applies to native IPv6 as well.
- scripts/xdp_ipv6_smoke.sh: phase 4 patches .rodata before load and
verifies policy=2 on a real kernel (3 pass / 3 rate-limited).
Loader (XDP could never load before this):
- .rodata is one struct rampart_config with frozen offsets enforced by
_Static_assert; XdpGlobals is #[repr(C)] and the image is built with
offset_of!. Before, a 96-byte image was written into a 76-byte map at
offsets clang had reordered, so load() failed with EINVAL.
- patch_rodata fails loudly when the map size and image size diverge.
- build.rs tracks xdp/core and xdp/hooks, not just universal_filter.c.
- build_ringbuf looked for "events_map"; the map is "events_ringbuf".
- tests/xdp_object_maps.rs checks map names and .rodata size without root.
Docs and style:
- Code comments are English, JavaDoc/Rustdoc style, without emoji or
banners (new project rule, TODO §0.8).
- README rewritten bilingual: Russian first, English below, nav buttons,
compact TOC and copy-paste quick start.
- docs/README.md: documentation index and maintenance rules;
configuration.md documents the whole [xdp] section and IPv4/IPv6
coverage; testing.md covers the real-kernel smoke; stale links removed.