Ignore duplicate or counter-regressed ReceiverReports before updating
RTT, loss, goodput, or ETX, so a delayed or reordered report can no
longer poison link metrics. Compute the RTT-from-echo sample with
checked timestamp arithmetic and reject zero, negative, or out-of-range
results instead of risking wrap or underflow on untrusted wire values.
On the sender side, when receiver dwell time overflows the u16 wire
field, suppress the timestamp echo (send 0) and saturate dwell to
u16::MAX rather than truncating, so a bogus small RTT cannot be formed.
Adds duplicate, out-of-order, wrapped-add, and future-dated (checked_sub)
sample tests, asserts loss and goodput stay unchanged on a dropped
duplicate, and covers the dwell-overflow echo suppression. Documents the
behavior in the MMP design note and CHANGELOG.
Co-authored-by: Johnathan Corgan <johnathan@corganlabs.com>
Add scoped mDNS / DNS-SD discovery for peers on the same local link,
giving sub-second pairing without a relay or NAT-traversal roundtrip.
A node advertises its npub, protocol version, and an optional network
scope over link-local multicast, and browses for matching adverts to
initiate Noise handshakes against same-LAN peers.
LAN discovery is disabled by default; operators enable it with
node.discovery.lan.enabled: true. Default-off avoids reintroducing a
per-LAN identity broadcast on nodes that have deliberately disabled
other discovery channels, and avoids any multicast surprise on upgrade.
The startup advertised-port picker now excludes bootstrap transports
and selects a non-bootstrap operational UDP transport with a stable
lowest-id selector, so the advertised port is deterministic across
restarts rather than dependent on HashMap iteration order. This
matches the per-dial transport selection used for discovered peers.
Co-authored-by: Johnathan Corgan <johnathan@corganlabs.com>
Add Node::update_peers for runtime peer-list refresh. It re-derives the
active peer connections from a new peer configuration, adding newly
configured peers and removing those no longer present, while keeping
links to peers that remain in the set rather than tearing every
connection down. The call returns an UpdatePeersOutcome summarizing the
added, removed, and retained peers.
PeerAddress gains a seen_at_ms recency field (with_seen_at_ms). Active
path selection now sorts address candidates by recency so the most
recently observed address wins when concurrent path probes race.
complete_rekey_msg2 now returns the remote peer's startup epoch
alongside the new Noise session, letting the rekey path detect a peer
restart and clear stale session state. A stale FSP session is cleared
when a peer restart is detected during FMP rekey or cross-connection
promotion, so the session-layer map no longer lingers out of sync with
the freshly promoted peer.
Per-tick work budgets bound the connection churn in a single node tick
(MAX_DISCOVERY_CONNECTS_PER_TICK, MAX_RETRY_CONNECTIONS_PER_TICK,
MAX_PARALLEL_PATH_CANDIDATES_PER_PEER); work beyond a tick's budget is
deferred to the next tick rather than discarded.
Co-authored-by: Johnathan Corgan <johnathan@corganlabs.com>
Moves both AEAD layers (ChaCha20-Poly1305, one round per layer per
packet) plus the sendmsg syscall off the rx_loop task onto a per-shard
worker pool, adds per-peer connect(2)-ed UDP with SO_REUSEPORT, and
uses Linux UDP GSO (sendmsg+UDP_SEGMENT — kernel splits one super-skb
into N on-the-wire datagrams in a single TX-stack walk) when packets
in a batch are uniform-size. Same kernel primitive WireGuard's
in-kernel module and BoringTun use to hit 2.5–3.2 Gbps single-stream.
Single TCP stream on a 5-node docker-bridge mesh, 5 x 15 s x P=1:
A→D: 1379 → 2708 Mbps (1.96x, RTT +0.12 ms)
A→E: 1394 → 2663 Mbps (1.91x, RTT +0.11 ms)
E→A: 1406 → 2624 Mbps (1.87x, RTT +0.19 ms)
Static-peer pairs only — every CoV under 3%, 0 outliers, 0% ICMP
loss. The ~+100 µs RTT is the worker queue handoff cost; AEAD +
sendmmsg now run on a separate core in exchange.
What lands:
- src/node/encrypt_worker.rs: std::thread + crossbeam_channel
workers; hash-by-destination dispatch pins a TCP flow to one
worker so wire ordering is preserved; per-worker sendmmsg(2)
batching up to 32; Linux uses sendmsg(2)+UDP_SEGMENT when
packets in a group are uniform-size.
- src/node/decrypt_worker.rs: receive-side mirror. Each shard owns
its session's recv cipher + replay window in a thread-local
HashMap (no shared RwLock/Mutex). Sessions are handed off at
promote_connection and re-registered on K-bit flip / rekey
cutover.
- src/node/handlers/session.rs try_send_session_data_pipelined:
FSP+FMP both seal in-place in the worker on one wire-buffer
alloc; no intermediate inner_plaintext / fsp_payload Vecs.
- src/transport/udp/connected_peer.rs + peer_drain.rs: per-peer
connect(2)-ed UDP socket with SO_REUSEPORT (set on the listen
socket too — without that, EADDRINUSE on activation and every
packet falls back to the wildcard path); the worker sends with
msg_name=NULL and the kernel uses its cached 5-tuple. Tick-
driven activation in handlers/connected_udp.rs, idempotent.
- src/transport/udp/mod.rs: mem::replace the recvmmsg backing buffer
instead of buf.to_vec() per packet — single pointer swap, no
MTU-sized memcpy.
- src/protocol/link.rs SessionDatagramRef: zero-copy borrowed view
used by handle_session_datagram for the bulk local-delivery
path; handle_session_payload takes the borrowed payload
directly (no payload[35..].to_vec()).
- src/transport/mod.rs TransportAddr::from_socket_addr: collapses
the two-alloc from_string(addr.to_string()) pattern to one.
- src/node/handlers/rx_loop.rs: decrypt-fallback drain promoted
ahead of packet_rx in the select! (TCP ACK starvation fix);
interleaved fallback drain every 32 packets inside the rx burst
loop.
- noise::Session: send_cipher_clone / recv_cipher_clone /
recv_replay_snapshot_owned / take_send_counter / accept_replay
so off-task workers can hold a cloned cipher + reserved counter
while the dispatcher keeps replay/counter sequencing serial.
CipherState::cipher_clone returns a refcount-bumped LessSafeKey.
AsyncUdpSocket: AsRawFd so workers issue raw sendmmsg / sendmsg
without going through the tokio reactor.
- Worker pool sizing: both default to num_cpus, overridable via
FIPS_ENCRYPT_WORKERS=N / FIPS_DECRYPT_WORKERS=N. Per-peer
connected UDP can be disabled via FIPS_CONNECTED_UDP=0.
- src/perf_profile.rs: optional per-stage timing reporter under
FIPS_PERF=1 (or FIPS_PIPELINE_TRACE=1). Off by default; zero
overhead when disabled.
- All cfg(unix)-gated. Windows continues on the existing tokio-
based send/recv.
Decrypt worker session lifecycle:
- Node::unregister_decrypt_worker_session mirrors the existing
register helper. Wired at the two natural sites that already
iterate peers_by_index: the rekey drain-completion block in
handlers/rekey.rs (drops the worker entry for the old our_index
once the drain window has expired and the cache_key is
unreachable to any in-flight OLD-K packet), and remove_active_peer
in handlers/dispatch.rs (drops the worker entry for each of the
four index slots: current, rekey, pending, previous). Only
our_index is normally registered; unregister_session is fire-
and-forget for missing entries, so calling unconditionally on
all four slots is correct and bounds the cleanup without per-
slot accounting. Without these callers the per-worker sessions
HashMap and the Node's decrypt_registered_sessions set would
grow monotonically per rekey on long-lived peers.
Testing:
- testing/static/scripts/bench-multirun.sh: multi-run iperf3 +
ping bench. N reruns (default 5), median / min / max / CoV % /
per-run outlier flag, avg ping RTT, ICMP loss %, TCP retransmit
total. Plain client→dest labels + topology header. Pre-bench
peer-convergence check (FIPS_BENCH_CONVERGE_SECS, default 15);
per-path route verification via stats.bytes_sent deltas — fails
fast if traffic exits via a non-static-peer link.
- testing/static/docker-compose.yml: passes FIPS_ENCRYPT_WORKERS /
FIPS_DECRYPT_WORKERS / FIPS_PERF through to containers for A/B
benchmarking without rebuilds.
- testing/static/scripts/iperf-test.sh: same plain client→dest
labels + topology header (was multihop/direct/N hop, which
conflated topology distance with on-wire path).
- .config/nextest.toml: synthetic UDP node tests serialized
through a max-threads=1 test group. Localhost handshakes drop
on shared CI runners under parallel load; one-at-a-time keeps
assertions reliable.
- src/node/tests/spanning_tree.rs: repair_missing_edge_handshakes
— retries up to 5 times for synthetic edges whose msg1 was
dropped, with a drain after each edge retry instead of after
each attempt's full burst.
- src/node/decrypt_worker.rs::tests: two unit tests asserting
WorkerMsg::UnregisterSession removes the worker-thread session
HashMap entry (handle_msg_unregister_session_removes_entry) and
is a no-op for never-seen cache_keys
(handle_msg_unregister_session_idempotent_on_unknown_key), which
is the safety invariant the unconditional unregister calls at
the four index slots in remove_active_peer rely on.
- src/node/encrypt_worker.rs::unix_tests
pipelined_send_wire_layout_roundtrips_canonical_decoders: mirrors
the encoder geometry of try_send_session_data_pipelined (no
coords, the common established-session path), runs the worker's
real seal + send via flush_direct_batch_sync, and decodes the
resulting wire packet using only canonical receive-side decoders
(EncryptedHeader::parse, SessionDatagramRef::decode, FSP header
parse, noise::open). Any divergence between the hand-rolled
encoder offsets (fsp_aad_offset, fsp_plaintext_offset) and the
decoders fails at one of the parse / open / decode steps before
the inner-plaintext assertion fires. Complements the existing
fsp_preseal_runs_before_outer_fmp_seal test which covers the
seal-ordering invariant with synthetic headers but does not
exercise the wire-layout invariant.
CHANGELOG.md [Unreleased] # Changed entry added describing the
worker-pool threading model, hash-by-destination dispatch,
sendmmsg/UDP_GSO, per-peer connected UDP, the operator-facing env
vars, and the bench numbers above.
Cherry-picks from mmalmi/master (paths translated from
crates/fips-core/src/ to src/): 9b7c723, 0deb5cb, 13f7339, e036c0e,
3740a68, 3792f83, 8510193, 4910b07, e53f545, e4e2896, 5fe4af5,
1d01ada, 8c37008, e12469e, 6eb2860.
Co-authored-by: Johnathan Corgan <johnathan@corganlabs.com>
The non-Linux test build was emitting warnings from code that is
intentionally platform-specific: the nftables firewall parser is
Linux-only, the utun address-family helper is only used in macOS
TUN paths, and one macOS Ethernet test module trips a clippy
layout lint. These warnings made focused test runs noisy and
encouraged bundling unrelated warning fixes into behavioral PRs.
- Gate the firewall parser dead-code allowance to non-Linux
targets, where the parser is compiled but not used.
- Mark the macOS utun helper and long TUN reader entry point with
narrow allowances.
- Rewrite the small MAC-copy loop to satisfy clippy and mark the
macOS Ethernet test module layout explicitly.
No runtime behavior change.
- Borrowed SessionDatagramRef decoder is used in the forwarding
handler so local delivery and coordinate-cache warming no longer
allocate or copy the session payload. The owned SessionDatagram is
materialized only when re-encoding for the next hop.
- Owned SessionDatagram::decode is reimplemented as Ref::decode +
into_owned, so the two decoders cannot drift.
- recvmmsg / recvmsg_x (Linux + macOS) receive loop moves each filled
slot buffer into ReceivedPacket via mem::replace instead of cloning
it; a fresh empty buffer is installed for the next syscall.
- TransportAddr is formatted directly from the SocketAddr without
going through an intermediate String.
Focused decode bench: ref 1.6 ns/op vs owned 34.7 ns/op (21.4x).
End-to-end iperf is neutral as expected for a ~30 ns saving per
packet.
Unit tests added:
- test_session_datagram_ref_decode_borrows_payload (verifies the
payload slice pointer equals the input slice's offset 35, a real
zero-copy invariant guard against accidental future to_vec)
- bench_session_datagram_decode_owned_vs_ref (ignored, run with
--ignored --nocapture)
- test_transport_addr_from_socket_addr
The Linux recv path drains up to 32 datagrams per kernel wakeup via
recvmmsg(2), amortising the per-syscall + per-task-wakeup cost across
the burst. macOS still fell through to single-packet recv_from, so
the same overhead capped inbound rate on Apple builds.
Add an equivalent batch path for Darwin using recvmsg_x(2). It is a
xnu-private syscall (not in the public SDK) but is the canonical
amortisation primitive on macOS — same shape used by quinn-udp for
the same reason. ABI is the public msghdr layout plus a trailing
msg_datalen (per-datagram bytes-received output), declared via
`unsafe extern "C"` against a local repr(C) `msghdr_x`.
Same `(count, kernel_drops)` contract as the Linux `recv_batch`. macOS
has no SO_RXQ_OVFL equivalent, so `kernel_drops` is always 0 — the
1Hz `sample_transport_congestion()` detector simply sees no kernel
drop signal on Apple hosts (it already tolerates that, since the
field has been 0 there pre-batching too).
cmsg buffer is intentionally null: we never consume ancillary data on
this path, and quinn-udp documents that `recvmsg_x` does not overwrite
`msg_controllen` on macOS 10.15+ (zeroed init is the only safe state).
The udp_receive_loop dispatch widens from cfg(linux) to
cfg(any(linux, macos)); the per-packet recv_from path is now used
only on the remaining unix targets (BSDs etc.) and Windows.
Add test_burst_recv_batch exercising 10 in-flight datagrams to
verify per-datagram boundaries and arrival order across the batch.
Add an ignored bench_udp_recv_amortization measuring recv-side
syscall amortization across 1/2/4/8 sender threads on dedicated
blocking std threads (kernel rx queue stays saturated regardless of
tokio scheduling). Sample numbers on aarch64-apple-darwin (100B
payloads, 3s windows):
senders=1: recv_from 398k pps recv_batch 432k pps 1.09x
senders=2: recv_from 353k pps recv_batch 608k pps 1.72x
senders=4: recv_from 322k pps recv_batch 503k pps 1.56x
senders=8: recv_from 353k pps recv_batch 515k pps 1.46x
Gate the Linux-only IpAddr import in control::listening behind a
cfg(target_os = "linux") so the macOS test build is warning-clean
now that test code paths there compile.
New testing/boringtun/ harness runs two Cloudflare BoringTun userspace
WireGuard containers with iperf3 between them, giving a single-hop
userspace tunnel baseline for comparison against FIPS throughput
numbers. Local WG key generation runs through the harness image so the
host needs no wireguard-tools.
New testing/static/scripts/iperf-compare-refs.sh builds two git refs
into separate fips-test:* images via git worktree and runs the same
static iperf topology against both, with RUNS-based repetition and
aggregate avg/min/max reporting.
testing/static/scripts/iperf-test.sh gains DURATION, PARALLEL,
SETTLE_SECONDS, IPERF_TIMEOUT env knobs and a per-path iperf timeout.
testing/static/docker-compose.yml selects the image under test via
FIPS_TEST_IMAGE; testing/scripts/build.sh respects CARGO_TARGET_DIR.
Author benchmark on aarch64 Docker Desktop:
boringtun bob -> alice : 1000.13 Mbits/sec
Skip BootstrapEvent::Established and BootstrapEvent::Failed dispatch
in poll_nostr_discovery for peers that are already connected or
actively handshaking. Without these guards, stale traversal events
arriving after a peer connected through a different path would
either attempt to adopt a redundant socket against the live
connection (Established) or poison the per-peer failure-state
cooldown and trigger redundant retraversal via schedule_retry /
try_peer_addresses (Failed).
The four guard sites use a new is_connecting_to_peer helper extracted
from the existing closure inside initiate_peer_connection; the helper
checks for an in-flight outbound handshake state. adopt_established_traversal
gains a defense-in-depth check returning PeerAlreadyExists when called
against an already-promoted peer, so the invariant holds if a future
caller bypasses the outer dispatch guard.
Side benefit: narrows a cooldown-poisoning vector previously available
to an attacker injecting stale failure events for an active peer.
Test coverage for the new behavior:
- test_try_peer_addresses_skips_connected_peer
- test_try_peer_addresses_skips_connecting_peer
- test_nostr_traversal_failure_skips_connected_peer (Failed-arm event
injection)
- test_nostr_traversal_established_skips_connected_peer (Established-arm
event injection, mirror of the Failed test)
- test_adopted_traversal_skips_already_connected_peer
(adopt_established_traversal defense-in-depth)
CHANGELOG entry under [Unreleased] / Fixed.
Closes#87
Investigated the mipsel-unknown-linux-musl build of this branch on a
Linux/x86_64 host. ring 0.17, portable-atomic, and the fips codebase
itself all compile cleanly for that target — the AtomicU64 portability
work is already done. The actual blocker is in the nostr-relay-pool 0.44
transitive dep, which uses std::sync::atomic::AtomicU64 directly in
src/relay/{stats,ping,flags}.rs. Verified fixable with a 6-line
portable-atomic patch via a [patch.crates-io] shim during local testing.
Updating the comment so the next person looking at this matrix has the
right starting point.
No functional change.
The chacha20 crate (RustCrypto) ships SSE2 + soft backends only — on
aarch64 (Apple Silicon, ARM Linux servers, Docker on M-series Macs) it
falls through to a portable software impl at ~600–800 MB/s/core. ring
0.17 wraps BoringSSL's hand-tuned ChaCha20-Poly1305, which dispatches
to NEON on aarch64 and AVX2/AVX-512 on x86_64 — typically 3-5 GB/s/core
on the same hardware.
Same wire format. ChaCha20-Poly1305 is byte-deterministic for a given
(key, nonce, plaintext, aad), so any correct AEAD implementation
produces identical ciphertext. The full noise test suite covers this
implicitly: IK and XK roundtrip handshakes, replay window correctness,
multi-message nonce sequencing, and 100-message stress all pass at
1129/1129 (the lib's full `cargo test` count) — these only succeed if
ring's output matches what the receiver's existing replay-window
decrypt path expects.
Implementation notes:
* `LessSafeKey` (and `UnboundKey`) deliberately do not implement
Clone for safety. `CipherState`'s manual Clone impl rebuilds it
from the retained 32-byte key — cheap for ChaCha20-Poly1305 since
construction is essentially a key copy + a constant-time check.
* The keyed AEAD is now cached in `CipherState.cipher` instead of
being re-derived per packet. This was already a perf win for the
chacha20poly1305 backend (`new_from_slice` per packet was hot in
profiles); for ring it's a bigger win because `LessSafeKey`
construction also derives the Poly1305 key.
* Public `Vec<u8>`-returning API preserved. New module-private
`seal`/`open` helpers wrap ring's `seal_in_place_append_tag` /
`open_in_place` so the per-packet allocation pattern is local to
one place.
* `EndToEndState::Established` triggers `clippy::large_enum_variant`
after the swap (`NoiseSession` grew from ~600 to ~1.5 KB because
ring precomputes the Poly1305 key state at construction). That
precomputation is the win — boxing the variant would re-add an
indirection per packet and work against it. `#[allow]`'d at the
enum decl with a justifying comment.
ring is widely deployed (rustls, hyper-rustls, AWS SDK, …) and a
pure-Rust crate (uses BoringSSL's asm via a vendored build). It
introduces no new C toolchain requirements that aren't already there
for any rustls user.
Bench data from a downstream consumer of this crate (Docker e2e,
DURATION=10, identical hardware before/after, aarch64 Linux on
Apple Silicon):
2-node direct (A↔B):
TCP 1-stream 437 → 1097 Mbps (2.51×)
TCP 4-stream 439 → 1109 Mbps (2.53×)
TCP 8-stream 445 → 1069 Mbps (2.40×)
UDP @1000 Mbit 599/40% loss → 1000 Mbps lossless
ping under load ~0.6 ms (unchanged)
3-node forced transit (A → C → B):
TCP 1-stream 438 → 1019 Mbps (2.33×)
TCP 4-stream 421 → 982 Mbps (2.33×)
TCP 8-stream 443 → 1031 Mbps (2.33×)
UDP @1000 Mbit 475/52% loss → 1000 Mbps lossless
ping under load 7.68 ms / 215 ms max → 0.72 ms / 3.6 ms max
The relay-path lift is the cleanest tell on the bottleneck: the
transit node was crypto-bound (single-threaded soft chacha couldn't
keep up with offered rate), so the queue accumulated under load. With
NEON the relay isn't crypto-bound and the queue stops accumulating —
the 215ms ping-tail collapses to 3.6ms.
The advert cache inside fetch_advert is read-only on hit — once a peer's
overlay advert is cached, every subsequent lookup returns the same
endpoints regardless of whether they still work. So when a peer rebinds
its NAT (or its STUN-discovered port flaps), connection retries to that
peer dial the same dead address forever, even with exponential backoff
firing at the right cadence.
Observed in deployment: macOS daemon's view of a Linux peer would
"regress" — peer marked rch=False after a brief link-dead window, then
hours of "Retry connection initiation failed: no operational transport
for any of <npub>'s addresses" with no recovery. Manual pause+resume of
the daemon (which restarts the FIPS endpoint and forces fresh advert
fetches) was the only way out.
When initiate_peer_connection / a retry tick returns
NodeError::NoTransportForType, fire-and-forget refetch_advert_for_stale_check
on the peer's npub. This re-fetches kind 37195 from advert_relays; if
the relay has a newer advert it replaces the cached entry, if it has
nothing it evicts the cached entry. Either way the next retry tick goes
to fresh data instead of looping on the same dead endpoint.
Mirrors the existing stale-advert sweep that runs from the
BootstrapEvent::Failed (NAT-traversal-streak) path, but covers the
direct-UDP-retry path which never crosses that streak threshold.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
When initiate_peer_connections() runs at boot, address resolution can
fail for an entire peer (no operational transport for the configured
transport types, all addresses unreachable, NAT rebind invalidated cached
endpoints, etc.). Before this change the failure was logged and silently
forgotten — the peer entry stayed in a dead state forever, accepting
incoming pings but unable to answer them, until the daemon was manually
restarted.
The retry plumbing (schedule_retry / process_pending_retries with
exponential backoff) already exists and is wired into the post-handshake
failure paths (BootstrapEvent::Failed, MMP dead-link timeout, handshake
timeout). The startup loop just wasn't calling it. Mirror the
BootstrapEvent::Failed path: on a startup peer-init error, parse the
peer's npub and call schedule_retry so the peer recovers without
operator intervention.
Includes a regression test that asserts retry_pending is populated when
initiate_peer_connections() fails for a peer with no operational
transport.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The previous fix (6ebca3e) only refetched the advert when retry returned
NodeError::NoTransportForType (cache returned no addresses at all). But
the much more common stale-cache failure mode is: cache returns an
endpoint that LOOKS valid (the address it had last week, before the
peer's NAT rebound), the dial succeeds at the IP layer, the handshake
times out, MMP fires, schedule_reconnect adds the entry back to
retry_pending, next retry hits the same cached endpoint, dials it
again, times out again. Loop forever — no NoTransportForType ever
fires because the cache has data, just dead data.
Move refetch_advert_for_stale_check to before each retry attempt
unconditionally. Cheap (one Filter query against advert_relays with
a 2s timeout, bounded by the retry backoff cadence), and replaces the
cache only if the relay has a newer advert or evicts if the relay has
nothing. Keeps the retry loop pinned to relay ground truth instead of
whatever the cache happened to learn at startup.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
When a node was the smallest-NodeAddr peer it could see (no smaller
neighbor available as a parent), the spanning-tree state was promoting
it to root. But the ancestry it advertised on the next TreeAnnounce
still referenced its previous parent's path, so receiving peers
rejected the announce with `invalid ancestry: advertised root X is
not the minimum path entry Y`, blocking mesh transit on any path that
needed to traverse this node.
Detect the self-root transition explicitly in `TreeState::become_root`
and rebuild the advertised ancestry to start from self. Also surface
the same path through the MMP receive handler so a stale ancestry
inherited across reconnect is corrected eagerly rather than waiting
for the next observation tick.
Adds 80 unit tests in `tree::tests` covering self-root transitions,
mid-chain ancestor disappearance, and ancestry validation against the
new root, plus a regression in `node::tests::spanning_tree` for a
3-node chain where the middle node's only parent (the smallest-addr
peer) goes away — previously it would advertise an ancestry rejected
by both endpoints; now it self-roots cleanly.
The run_rx_loop's `tokio::select!` was costing one full scheduler hop
+ futex per inbound packet and per outbound TUN packet. Under
sustained load that capped throughput at one event per scheduler
quantum — independent of CPU (which sat near-idle) because every
iteration parked the worker, woke it via futex, processed one event,
then parked again.
After the await on `packet_rx.recv()` / `tun_outbound_rx.recv()`
fires, drain up to 256 additional ready items via `try_recv()` in a
tight inner loop before yielding back to `select!`. `biased` ordering
gives the data-plane branches priority over tick / control / DNS
under sustained load.
The 256 cap is empirically tuned to keep the worker on a busy stream
between yield points (a contiguous burst of ~256 MTU-sized packets
≈ 400 KB of contiguous traffic) while still bounding the inner loop
so a flood on one branch can't starve the periodic tick or control
socket. Lower caps (64) left perf on the table; higher caps (1024+)
delayed tick handling visibly under stress.
Pairs with the recvmmsg(2) change in the previous commit: the kernel
UDP queue now hands packets to `packet_rx` in 32-batches, and the
rx_loop drains them without a per-packet scheduler hop.
The UDP recv loop drained the kernel queue one packet per recvmsg(2).
Each call paid full per-syscall + per-task-wakeup overhead (~50us avg
including a futex-based scheduler hop), so under sustained load the
loop ran at one rx event per scheduler quantum — the dominant cap on
inbound packet rate.
On Linux, switch the steady-state path to recvmmsg(2) with a 32-packet
batch. A single readable() wakeup drains up to 32 datagrams in one
syscall before yielding back to the reactor. Stack-allocated mmsghdr
arrays sized to a module-level `BATCH_SIZE` constant.
`SO_RXQ_OVFL` is sampled once per batch off the cmsg chain of `msgs[0]`
and plumbed through `AsyncUdpSocket::recv_batch` as `(count, drops)`.
The counter is socket-wide and monotonic, so a single sample per batch
gives the 1Hz `sample_transport_congestion()` detector ample fresh
values under load (one batch = up to 32 datagrams). Cost is one
stack-allocated CMSG_SPACE(4) buffer + one CMSG_FIRSTHDR walk per
batch syscall.
macOS / Windows fall through to the per-packet recv_from loop —
recvmmsg is Linux-specific and the per-packet API is fast enough on
those platforms for now (recvmsg_x for Darwin can be added later).
The slice-array build also drops the `MaybeUninit::uninit().assume_init()`
+ `transmute` pair for `std::array::from_fn` over a single shared
`backing.iter_mut()` — same disjoint mutable borrows, no `unsafe`.
`PeerIdentity::pubkey_full()` falls through to
`self.pubkey.public_key(Parity::Even)` whenever the parity-aware full
key wasn't passed at construction (i.e. for every peer constructed
from an npub or x-only key). Underneath, that runs a secp256k1 EC
point parse — `fe_sqrt` + `fe_mul` + `ge_set_xo_var` — which is ~6%
of per-packet CPU on the bulk-data send path for a value that never
changes after construction.
Compute it eagerly. The same EC point parse already runs at
construction inside `NodeAddr::from_pubkey`, so the cost is paid once
where it would be paid anyway.
`Node::adopt_established_traversal` was constructing the adopted UDP
transport with `UdpConfig::default()` — MTU 1280, default recv/send
buffer sizes, default accept/advertise flags. If the operator had
configured a higher MTU on the primary `[transports.udp]` listener
(e.g. 1500 on a path where larger frames are known viable), full-sized
tunnel datagrams sent over the NAT-traversed link would exceed the
adopted socket's MTU and get dropped at the socket layer with no
visibility into why throughput collapsed.
Inherit the primary UDP config (MTU + recv/send buffer sizes + accept
/ advertise flags) and clear the bind / external-address fields since
the adopted socket is already bound. Lookup tries `transport_name`
first so operators with multiple named `[transports.udp.<name>]`
listeners pick up inheritance from the matching listener, and falls
back to the unnamed `Single` listener so single-instance configs work
unchanged.
The previous default of MTU 1280 was deliberately the IPv6 minimum,
the only value guaranteed to survive arbitrary middlebox paths. With
this change, operators who set their primary listener higher (based
on known-clean LAN topology) will have NAT-traversed flows initially
attempting that higher MTU and possibly black-holing on tighter paths
until reactive `MtuExceeded` recovery kicks in. Documented in the
adoption call-site comment so future readers understand why the
conservative default went away.
Discovered in a downstream consumer where a `MESH_TUNNEL_MTU=1320` /
encrypted wire ~1426B produced silent packet drop on every session
that had been promoted onto a NAT-traversed link.
Adds two sibling tests in `src/node/tests/bootstrap.rs` pinning the
new behaviour for the `Single` and `Named` config variants.
Martti MalmiandJohnathan Corgan
