The guard collapsed all thirteen chaos legs to the token "chaos" and all five deb-install legs to "deb-install", so sixteen of the thirty-four integration legs were invisible to it. Deleting twelve chaos legs and four deb-install legs from a copy of the workflow left it still reporting "CI parity OK". The fold existed because the guard compared the cosmetic suite name, which really does differ between the runners; both sides have carried a directly comparable scenario field all along, so the fix is to compare that instead, along with the chaos flags.
The local suite set is now discovered by sweeping ci-local.sh for suite arrays rather than reading a hardcoded list of variable names, and the deb-install distro list is read from the suite's own script. A suite dispatched with no backing array is invisible to either approach, so every run_suite arm is now checked to have one and the guard names any that does not. That is not hypothetical: next dispatches its mixed-profile suite without an array, and the guard could only report it as missing from the local runner without being able to say why.
The guard also now refuses to run rather than failing obscurely when python3 or pyyaml is absent, since a parity check that cannot run must not look like one that passed. Verified by breaking each thing it guards: legs deleted from the workflow, drifted chaos flags, a fabricated suite array, and a dispatch arm whose array was removed. That last case initially did not fire because the arm pattern was pinned to the wrong indentation, which the test caught.
The scenario loader read every section and member through raw.get with a default, so a mistyped key was silently ignored and the block it belonged to stayed default-constructed while the YAML still looked correct on the page. Writing "assertion:" for "assertions:" disarmed a scenario's only assertions and left it unable to return a non-zero exit code; "link_flap:" for "link_flaps:" turned off the churn the scenario existed to inject, and a calm mesh then ran under the name of a chaos test.
Keys are now checked at load time, at the top level and inside every section with a fixed schema, including the netem policy bodies, the {min, max} ranges and the assertion bodies themselves. The error names the offending key, the section it appeared in, and the keys that section does understand. Three mappings are deliberately exempt because their keys are names the author chooses rather than a schema, and two sub-trees are passed through whole; all five are listed at the key sets with the reason.
Adding an assertion type now means registering it, so the next person to add one gets a rejection rather than a silent no-op. That coupling is the point and it is written down where they will see it. Verified against the seventeen scenarios in the tree, all of which still load, and against fixtures for each failure mode: a top-level typo, a nested typo, a typo inside an assertion body, and a section left empty, which previously raised an unrelated AttributeError from the next line of the loader.
The e2e scenarios build fips and fips-gateway in a builder image and extract them with two docker cp calls whose stderr went to /dev/null and whose exit status was never tested. Because build_fips_for_e2e ended in a log call it returned 0 on every path, so the caller's guard could not fire. A failed extraction left the previous run's binaries in the cache, they satisfied the caller's executable check, and the five e2e legs then exercised the previous commit's code and reported green.
The extraction now deletes the cached binaries before it starts, which is what makes the stale case impossible rather than merely detected, and it checks each docker cp independently, reports the failure with docker's own message, verifies each extracted file is non-empty and executable, and returns an explicit status. This mirrors what the deb-install suite already does.
Confirmed by inducing the failure rather than by observing a green run. With both extractions broken and a stale binary in place, the old code logged "Cached fips (26936 bytes)" and went on to build the runtime image around it; the new code reports both docker errors, deletes the stale binaries and fails the suite. Breaking either extraction on its own is caught separately, and an unmodified run passes 7 of 7.
Compose commands run with their output captured and with check set, so a failure raised a CalledProcessError and nothing ever looked at what the daemon had said. That exception reports the argv and the exit status and nothing else, so every scenario that died bringing its containers up left a runner log saying only that docker compose returned non-zero exit status 1, and the reason was captured and then discarded. In the run that prompted this work twelve scenarios failed that way and why they failed can no longer be established at all. Log the captured stderr and stdout before the error goes anywhere, keeping the tail of each so a verbose failure cannot fill the log file. A command that times out raises before its exit status is ever read, so that path records whatever it had emitted too, rather than leaving the same hole one shape over. Reproduced against both of the shapes this is meant to catch: an unusable network prefix now records the daemon's parse error, and a name already held by another container now records the conflict and names the container holding it. The success path logs nothing new, and a completed scenario's runner log is unchanged line for line.
The raising behaviour is deliberately untouched, since the abort flag and the exit code that came with it depend on it. The check is applied here rather than by subprocess so that a caller who passed check false gets the same record: both such callers are the compose down in teardown, which today can leave a mesh behind without saying so.
The same omission was costing the veth path its reasons. Failing ip commands had their stderr logged at debug, which the runner does not emit unless asked for verbose output, so a failed pair creation reached the log as nothing more than the names of the two interfaces it could not create. Raise that to a warning. The deletes issued to clear a stale pair are expected to fail and still say nothing.
Teardown runs from a finally, so it also runs after a failed setup, and it was guarded only on the topology and the compose file, both of which are set well before any container exists. A scenario that died bringing its containers up therefore ran the whole harvest anyway: final snapshots, docker logs, the analysis, the assertions and the metadata, all addressing containers by names that are global to the host. What it left behind was not an empty directory an investigator would notice but a full and plausible one, in the recorded case ten node logs and an analysis reporting two promotions and two parent switches, every byte of it from a different scenario's mesh. Gate the harvest on whether the containers ever started, and leave the compose down outside that gate so a partly successful start still gets cleaned up.
Record the outcome in a status file in every result directory, naming the run as completed, interrupted, aborted, setup-failed or teardown-failed, alongside the scenario, the seed and the container names it used. With the harvest gated, the presence of an analysis file is now itself proof that the scenario's own mesh existed, and the status file says which of the several ways a run can end applies. A run cut short by a signal keeps the existing exit codes rather than gaining one of its own: what it collected before stopping is real and still worth reporting, the status file records that it was truncated, and every wrapper already reports a Ctrl-C of its own.
Log collection never looked at the status of docker logs. It concatenated stdout and stderr unconditionally, so collecting from containers that were not there wrote the daemon's "No such container" reply into each node log and analysed the result as a mesh with no panics, no errors and no sessions, which reads exactly like a clean run and exits zero. Check the return code, and treat a harvest that cannot read every container, or that reads none, as a failed teardown. A teardown that raises now also reports on the same footing as a run that never started, instead of escaping past the exit codes as a bare traceback and being read as a malformed command line.
The simulation caught every exception a scenario raised, logged it, and exited 0, so a scenario that died during setup was recorded as a pass. Local CI discarded the exit code in any case: run_chaos ends in a call whose own last statement is an echo, so it returned 0 whatever the scenario did, and the parallel launcher's wait therefore always succeeded and always recorded a pass. Between them no chaos failure of any kind could turn a local run red, and none could since the script was written.
Carry the abort out of the run and give it its own exit code, ordered ahead of the two content codes so a run that never produced a mesh cannot report that mesh's panic and assertion counts. Return the scenario's status from run_chaos so the launcher sees it. The documented exit codes were stale before this change and are rewritten in full, including that the argument parser rejects a malformed command line with the same code that reports panics.
Expect scenarios that have been reporting green to start reporting red. In the last recorded local run twelve of the thirteen scenarios never got past starting their containers, and all thirteen were counted as passes; those that now execute have not run for months, and their assertions have never been evaluated against a mesh of their own.
Ethernet edges get a veth pair created in the host namespace and then moved into the containers, named from the node ids alone. ethernet-mesh and ethernet-only both connect n01 to n04 and always run together, so one tore down the other's live link. Hash the scenario suffix to four hex characters and insert those after the vh prefix; 15 characters is far too few to carry the suffix itself.
The names are now run-specific and never regenerated, so the delete-before-create no longer reclaims an interface an earlier run abandoned. ci-cleanup.sh takes over, deriving the names from the suffixes a run's teardown hands it. It is the first thing that script removes that is neither a docker object nor labelled, so an unscoped reap can sever a running bare simulation's links; testing/README.md and the --reap help say so.
Chaos scenarios run four at a time under local CI, but every one of them claimed the container names fips-node-nNN and wrote its generated configs and compose file to the same generated-configs/sim directory. Container names are global in Docker and are not scoped by the compose project, so concurrent scenarios collided on both, and a scenario could start containers from a compose file another had overwritten.
Thread the existing FIPS_CI_NAME_SUFFIX into the simulation. run_chaos narrows the run-wide suffix to the scenario, and the sim reads it once when the topology is built, applying it to the container names and to the config directory basename. The compose template renders the name through the topology accessor instead of duplicating the literal, so one expression produces every chaos container name.
The suffix is empty when the variable is unset, so a bare chaos.sh run and the hosted CI jobs render byte-identical names and paths. Verified by rendering every scenario's compose file before and after with the variable unset and diffing.
The isolation loop and the failure-log dump both built container names
without the run suffix, so under a suffixed run they addressed containers
that do not exist.
The consequence was worse than the visible failure. Two of the three
checks in that loop assert a ping FAILS, and a ping into a non-existent
container fails for the wrong reason - so both passed vacuously, and the
security assertion the loop exists for was silently a no-op. Only the
third check, which expects a ping to succeed, noticed anything was wrong.
Add a guard that fails loudly when the container is absent, so a future
naming slip cannot quietly turn these back into no-ops rather than
surfacing as one confusing failure among two false passes.
These sites were missed because the suite passes when run by hand: with no
suffix set the unsuffixed names are correct, and the defect only appears
under the automation that sets one. Verified this time with the suffix
set, which is the condition that matters: 18 of 18, container names
resolving to real containers.
Making the sidecar network name unique per run exposed a latent problem
rather than causing a new one. With a fixed name, compose found any
existing network and reused it, which quietly masked the fact that the
address range was hardcoded. With a unique name there is nothing to
reuse, so the range is requested every time and the run fails outright if
anything already holds it.
Claim the range instead of assuming it. The suite now tries to create its
network on a candidate range and, if the daemon reports an overlap, moves
to the next candidate. The daemon's own address pool becomes the arbiter,
so two concurrent runs cannot land on the same range at all - as opposed
to a range derived from the run identifier, which only makes a collision
unlikely, and a collision is the precise failure being avoided.
The node addresses are derived from whichever range is claimed rather than
written alongside it, including the gateway address the isolation check
probes. Moved to a range that nothing in the tree claims and that sits
outside the daemon's default pool.
Only an overlap is worth advancing on. Any other creation failure is real
and is reported with the daemon's own message, rather than being retried
sixty-four times and buried - a lesson from a failure elsewhere in the
harness this week that was undiagnosable because its error output was
discarded.
All three nodes now attach to the pre-created network as external, so
teardown removes it explicitly.
Verified: three concurrent claims take three distinct ranges; a
non-overlap error fails fast rather than looping; the suite passes 18/18;
and the claimed gateway address is confirmed pingable when not firewalled,
so the isolation assertions still mean something.
The suite red-ed with an empty final peer_count while every substantive
assertion passed: both denied peers showed inbound handshake attempts and
zero outbound responses, which is the gate this suite exists to verify.
Two pre-existing defects combined. The final count was sampled once, in
the same second the load driver issues its last round of peer restarts, so
the read can land on a daemon busy with those restarts and return nothing.
And the fallback meant to cover that could never fire, because the shell
applies it to the last stage of the pipeline, which succeeds on empty
input - so an unanswered query was indistinguishable from a genuine zero,
and reported as a cap violation.
Poll the final read for a short window instead, and give the two cases
distinct output. A real regression still fails, just after the retry
window, since the loop exits early only on the expected value. Extract the
read the two phases share rather than leaving them to drift.
Two runs on one host destroyed each other's containers, producing
mid-test "No such container" failures that look like real defects. The
automated builder runs a full local CI on the same box every few minutes,
so the machine is contended almost always and this has red-ed both a hand
run and an automated gate.
Two independent causes are fixed here. Container names were hardcoded, and
docker names are global rather than scoped by compose project, so two runs
collided on the same name; every name now takes an optional suffix that
the harness sets from the run id. And the cleanup sweep matched a label
shared by every run, so one run's teardown force-removed another's
containers; resources now also carry a per-run label and the sweep can be
narrowed to it.
A third hazard turned up that was not in the original report: the sidecar
suite passes explicit compose project names, which override the run-scoped
project and put it outside the shared prefix entirely. Its project names,
network, and derived container references are now scoped too.
Both are default-off. With the suffix unset, names render exactly as they
do today and a bare compose invocation is unchanged, which is what keeps
the hosted CI and the documentation correct. A cleanup run with no run id
still reaps everything, which is what a manual "clear the box" wants.
The literal-name sweep was not sufficient: six scripts build container
names dynamically from node labels, and two suites create their own
containers outside compose. Those are handled at their construction sites.
Verified: syntax check on all modified scripts; compose validation on
every modified file with the suffix both set and unset; and a synthetic
two-run reproduction that shows the old cleanup destroying a bystander run
and the new one leaving it alone.
Known gap: subnets are still hardcoded, so two concurrent full runs will
still collide on address-pool overlap. That fix reaches into topology
configs, chaos scenarios, diagrams and production source, so it is left
for its own change rather than half-done here.
The rekey suites' Phase 4 polled until the initiator-cutover count
reached its value at phase entry plus two, while Phase 6 asserts an
absolute minimum of four. On an unloaded host the ping phases can
outrun the jittered first rekey cycle, so Phase 4 can begin with a
single cutover on the books, release at three, and leave Phase 6 to
lose a sub-second race against the remaining first-cycle cutovers.
A rekey-outbound-only run failed exactly this way: five counted
cutover lines by t=32.6s of container life, with the Phase-6 count
snapshot at t~32.4s seeing three.
Floor the Phase-4 wait target at four so a released wait guarantees
the Phase-6 assertion — the docker-logs count is monotone over a
container's lifetime. A genuine rekey stall still times out the
bounded wait and fails Phase 6 with the real count.
Also drop the second-cycle framing from the comment, the phase
banner, and the assertion description: with rekey timers resetting at
each cutover a second cycle cannot begin inside this test's window,
so the counted events are first-cycle cutovers spread across the
topology's links, emitted once or twice per completed link rekey
depending on which side's promotion path logs them.
Phase 4 of the rekey suite waited for only one more initiator cutover
(guaranteeing three) while Phase 6 asserts at least four. The fourth
cutover then had to land in the brief window between the wait returning
and the Phase 6 log snapshot, so host load could push it past the window
and fail the run even though every cutover completed correctly. Wait for
two more cutovers, the full second rekey cycle, matching the Phase 6
threshold, so the asserted count is guaranteed before it is checked.
Add a "Running CI locally" section to testing/README.md covering the
ci-local.sh orchestrator: the per-run FIPS_CI_RUN_ID override and how it
scopes compose projects, image tags, and per-chaos-child subnets so
simultaneous runs on one host never collide; the cancellation exit codes
(130/143 vs 0/1) and how a preempting worker interprets them; and the
label-based cleanup via --reap / ci-cleanup.sh.
A CI worker may preempt an in-flight ci-local.sh run (SIGTERM, then SIGKILL
after a grace period) to restart on a newer commit. For that kill to be safe,
the script must clean up after itself and never let a dying run collide with
its restart. It previously had no signal handling, shared the default compose
project name across runs, and tore down each suite only at the suite end.
- Derive a per-run id (honoring FIPS_CI_RUN_ID, else short-sha+random) and
namespace every docker resource to it: a fipsci_<run>_<suite> compose project
per suite and per parallel chaos child, and per-run image tags
(fips-test:<run>, fips-test-app:<run>) retagged to :latest only after both
builds succeed so :latest never points at a half-built image.
- Install a bounded, idempotent teardown trap on SIGTERM/SIGINT (+ EXIT): reap
parallel chaos children, then force-remove this run's docker resources via
the new ci-cleanup.sh, wrapped in timeout so a stuck down cannot wedge it.
- Exit 143 (SIGTERM) / 130 (SIGINT), distinct from 0 (pass) / 1 (failed), so a
preempting worker tells a cancelled run from a real failure.
- Add ci-cleanup.sh (also ci-local.sh --reap): force-removes leftover CI
resources by the com.corganlabs.fips-ci=1 label and the fipsci_ project
prefix, robust to however a prior run died.
- Label every per-suite docker resource so the label sweep reaps it after a
SIGKILL regardless of network name: direct docker run/network resources, the
sidecar compose services, and every per-suite compose network (acl-allowlist,
boringtun, firewall, nat, static, both tor suites, and the chaos generator
template). Parametrize the static/sidecar compose image refs so the per-run
tags are honored.
- Give each parallel chaos child a unique /24 from 10.30.x (a new --subnet
override on the sim CLI, assigned per-child in ci-local.sh) so parallel
children never collide on a shared docker subnet, and a chaos net can never
span a fixed-subnet suite (sidecar/static in 172.20.x). 10.30.x sits outside
docker's default-address-pool range, so an auto-assigned net cannot land on
it either; node IPs derive from the subnet, so no scenario config changes.
Replace the fixed post-first-rekey and post-second-rekey settle sleeps in
the rekey integration test, and the fixed-deadline baseline wait in the
interop test, with the deterministic wait_until_connected progress-aware
polling helper already used for the initial baseline convergence.
Each converted site fails fast with structured diagnostics when the mesh
is stuck and extends its deadline only while pairwise reachability is
still climbing, removing the wall-clock settle windows that produced
intermittent connectivity failures after a rekey.
The GitHub deb-install matrix ran debian12/ubuntu24/ubuntu26, but the
local harness (testing/deb-install/test.sh) runs five distros. Add the
missing debian13 (trixie) and ubuntu22 legs so the cloud gate covers the
same distro set as local CI. Each new leg invokes the existing test.sh
scenario, so per-distro behavior is identical to the local run.
Update the granularity-only parity notes in ci.yml, ci-local.sh, and
check-ci-parity.sh to list the full distro set.
The sidecar chain test intends node-a to be a standalone root ("node-a: no
outbound peers"), but started node-a without clearing FIPS_PEER_*, so it
inherited the external peer default from testing/sidecar/.env (a real public
mesh node, test-us01.fips.network) and auto-connected to the live mesh. The
whole a-b-c chain then attached under an external root, inflating tree depth and
breaking test isolation; this also produced spurious multi-hop failures.
Set FIPS_PEER_NPUB/FIPS_PEER_ADDR empty for node-a, matching how node-b and
node-c already get explicit inline peers, so the suite is hermetic regardless of
the .env default. Validated against the unmodified .env: node-a comes up with a
single link to node-b, no external attachment, multi-hop passes 18/18.
wait_until_connected fail-fasted on stall even when the mesh was all but
converged, abandoning the run with budget still unspent because one hard
pair straggled to come up. On the rekey-outbound-only topology this turned a
rare deep-node timing straggle (stacked discovery backoff + late bloom
propagation, which clears well inside the budget) into a false baseline RED.
Add a near_converged_slack threshold (default 2): when the number of
still-failing pairs is at or below the slack and the stall window elapses,
keep polling toward max_secs instead of bailing. A mesh genuinely far from
convergence still fast-bails on stall, and a never-converging pair still
hits the hard cap, so a real regression is never masked.
Add a self-contained behavioral test (wait-converge-test.sh) covering the
four contract cases: near-converged hold (with a slack=0 contrast that
proves the slack is what saves the run), far-from-converged fast-bail,
never-converging hard cap, and four-argument backward compatibility.
The interop harness only tested a full mesh, where every pair is a direct
one-hop FMP link. That left multi-hop forwarding, routing, coordinate
handling, and mesh-size estimation untested across versions.
Add an opt-in multi-hop topology and three new checks:
- generate-configs.sh: FIPS_INTEROP_EDGES selects an explicit undirected
edge list instead of the full mesh, with symmetric peering, connectivity
validation (no isolated node, graph connected), and per-node degree plus
edge metadata in nodes.env. Unset means full mesh, unchanged.
- interop-test.sh: a --topology flag with a built-in multihop-3v-cycle
(six nodes, two of each version, one cycle, two leaves). The per-node
peer check now expects each node's adjacency degree rather than N-1, and
the all-pairs ping now also exercises cross-version forwarding over
non-adjacent pairs. Adds a control-differential data-plane continuity
stream across the rekey window (Phase 5b) and a strict plus/minus 25
percent mesh-size convergence check (Phase 7).
- interop-stress.sh: forward --topology through to the per-rep driver
(mutually exclusive with a positional node-spec).
Phase 1 makes the strict retrying ping authoritative. The convergence
detector (one fully-clean, no-retry sweep of every directed pair) is now an
advisory settle-wait whose timeout is non-fatal; the strict retrying ping
decides pass/fail, matching how the post-rekey phases already work. Under
packet loss a clean no-retry sweep of a large mesh is statistically unlikely
even when the mesh is healthy (30 pairs at 2 percent loss leaves only about
55 percent of sweeps clean), so the detector otherwise falsely failed runs
whose strict ping reported full reachability. CONVERGENCE_TIMEOUT is now
env-overridable and defaults to 90s for larger meshes under loss.
Full-mesh runs are unaffected. Validated end-to-end with the
multihop-3v-cycle topology across three versions: all 30 directed pairs
reachable including multi-hop routed pairs, zero data-plane loss through
both rekey cutovers, and mesh-size converging to the true node count on
every node.
Bring the local runner (testing/ci-local.sh) and GitHub CI into agreement
on two axes.
Suite coverage: the admission-cap integration suite ran only locally, so a
regression in it could never turn the GitHub gate red. Add an admission-cap
leg to the integration matrix, add testing/check-ci-parity.sh (wired as
'ci-local.sh --check-parity') to diff the two suite sets and fail on
unexpected drift, and document the deliberate local-only (live-Tor) and
granularity-only differences in a comment block atop both runners.
Toolchain selection: every CI and packaging job installed its toolchain
with dtolnay/rust-toolchain@stable, but the rust-toolchain.toml channel pin
overrode it for all compilation, wasting an install and printing a
misleading rustc version. Switch the stable call sites to
actions-rust-lang/setup-rust-toolchain, which reads rust-toolchain.toml as
the single source of truth. Keep the explicit cache steps (cache: false on
the new action), set rustflags empty so the action does not impose a global
-D warnings the previous setup never applied, fold the macOS cross-compile
target into the action input, and leave the OpenWrt nightly Tier-3 leg on
dtolnay/rust-toolchain@nightly.
Add wait_until_connected to the shared convergence helpers: it polls a
suite's own pairwise pings (the signal it actually asserts on), returns
as soon as every pair is reachable, extends its deadline while the
reachable-pair count is still climbing, and gives up only when progress
stalls.
Use it in the rekey, static-mesh, and sidecar suites in place of the
fixed wall-clock baseline timeout and the blind sleep, which timed out
under concurrent CI load while the mesh was still converging.
Remove the duplicated immutable fields (config, identity, startup_epoch,
started_at, is_leaf_only, max_connections/peers/links) from the Node
struct so the Arc<NodeContext> bundle is the single source of truth.
Previously Node owned these fields and a parallel context copy, kept in
lockstep by rebuild_context() at every mutation site — pure overhead that
existed only because of the duplication.
- Replace rebuild_context() with replace_context(): a clone-edit-swap of
the whole Arc. The per-instance context stays immutable; mutation swaps
the Arc. This is the sole runtime mutation path (constructors, leaf_only,
update_peers).
- Add Copy-returning accessors startup_epoch() and max_connections()/
max_peers()/max_links(); migrate the remaining direct field readers onto
the accessors. node_addr()/npub()/Debug now read identity/is_leaf_only
from the context.
- update_peers reads the pre-update peer set from the live context Arc
before building a fresh Config + context and swapping — preserving the
read-before-write ordering its mutation-window test depends on.
- Remove the test-only set_max_* setters; tests set the limits on Config at
construction instead (new make_node_with_max_peers/links helpers).
- Add a ci-local guard that fails if the Node struct re-declares a bundled
field, so the single-store invariant can't silently regress.
cargo test --lib 1291/0; clippy -D warnings and release build clean.
New integration scenario verifying the early-gate silent-drop behavior
of the inbound max_peers admission check at sustained scale, using the
existing 5-node mesh topology with one node's node.limits.max_peers
lowered to 1. This forces 2 of the cap'd node's 3 configured peers
into a sustained denied state, and asserts via tcpdump that no Msg2
responses go back to those denied peers across a 60s capture window.
A background load-driver restarts the denied peer containers every 15s
to reset their auto-reconnect exponential backoff (5s base / 300s cap),
producing fresh Msg1 bursts each cycle. Without this loop the gate
fires ~3-4 times per denied peer in a 60s window; with restarts the
observed rate is 15 per denied peer (~30 total firings), high enough
that any Msg2 leakage would be caught with strong statistical
confidence.
Local run on this branch: cap'd node-c converged to peer_count=1 with
node-b admitted; nodes d and e sustained-retried as denied; tcpdump
captured 30 inbound Msg1 (len 84) packets from the denied pair and 0
outbound Msg2 (len 104) packets, with final peer_count unchanged.
Files:
testing/static/scripts/admission-cap-test.sh — new test script with
inject-config subcommand (sets node.limits.max_peers) and a
3-phase test driver (converge, capture-with-load, per-peer assert)
testing/ci-local.sh — register admission-cap as a new suite category
(ADMISSION_SUITES), wire run_admission_cap function, add to
run_suite dispatch, list_suites, and the default integration sweep
Together with the existing unit-level coverage in src/node/tests/unit.rs
(handle_msg1_silent_drops_at_cap_for_new_peer with mock-transport Msg2
discriminator, and handle_msg1_admits_existing_peer_at_cap as the
bypass regression guard), the gate's silent-drop behavior is now
verified both at single-firing wire-observable resolution and at
sustained multi-firing cross-process scale.
Adds a tree/mmp-targeted compose overlay (compose-trace-tree.yml) and
a new FIPS_MESH_LAB_TRACE_TREE env-var gate in run_rekey_family,
layered independently of the existing FIPS_MESH_LAB_TRACE rekey-class
overlay. Trace targets: fips::node::tree, fips::tree,
fips::node::handlers::mmp, fips::node::handlers::handshake.
Used for tree-partition race investigations during multi-peer startup
where evaluate_parent inputs, send_tree_announce_to_all recipient
enumeration, and process_receiver_report first-RTT triggers all need
to be visible together to bracket the loss window.
Extend parse_rekey to emit phase1_status / phase1_baseline_passed /
phase1_baseline_total fields in signature.json by scraping rekey-test.sh's
"Best observed baseline before timeout: N/M passed" line (the timeout
path) and "Pre-rekey baseline (all 20 pairs): N/M passed" (the success
path). Phase-5-shape parsing is unchanged.
Extend mechanism_match_rekey to also fire on the Phase 1 characteristic
12/20 split — the multi-hop-routing-failure shape where direct-peer pairs
pass and the four multi-hop pairs (x 2 directions) fail. The Phase 5
predicate remains intact for the pre-existing flake class.
Add FIPS_MESH_LAB_NO_RESOURCE_LIMITS=1 to run_rekey_family for
unconstrained characterization runs where the goal is to surface a race
or scheduling artefact rather than reproduce GHA pressure. Default
behaviour (variable unset) keeps the compose-resource-limits.yml overlay
engaged.
Documented in README.md and the in-script env-var header block.
Closes the eventually-consistent gap in spanning-tree state
distribution. Every existing send_tree_announce_to_all call site
gates on a local state-change event (parent switch, self-root
promotion, ancestry change, peer promotion, parent loss). Once a
partition latches — for example a parent-switch announce stranded
in the brief cross-init handshake swap window, where the announce
arrives on a session-index whose decrypt-worker entry has been
unregistered — neither side's state changes again, so neither
side ever re-broadcasts. The existing 60 s check_periodic_parent_reeval
was a re-evaluation, not a re-broadcast: it short-circuited
silently on no-change. Production-side healing depended on
incidental link churn; lab harnesses with stable docker-bridge
links had no equivalent path.
Add a final else branch that fires send_tree_announce_to_all
unconditionally on the no-change path, alongside the existing
switch and self-promote arms. Receivers coalesce by sequence
comparison (ParentDeclaration::is_fresher_than) and short-circuit
at the `if !updated` gate in handle_tree_announce; same-sequence
repeats drop silently with no cascade. The per-peer 500 ms
rate-limiter is well below this 60 s cadence and does not suppress
the heartbeat broadcast.
The fix is a general protocol-robustness improvement: it addresses
any in-flight TreeAnnounce loss class, not only the specific
cross-init swap-window drop site.
testing/static/scripts/rekey-test.sh BASELINE_CONVERGENCE_TIMEOUT
60 -> 65 so a partition healed by the periodic broadcast at T+60
lands inside the convergence window. wait_for_full_baseline
early-exits on PASS, so successful reps see no extra wall-clock.
Three deltas to the mesh-lab nat-lan suite for stall characterization:
- FIPS_NAT_LAN_CPUSET env-var-driven CPU-pinning sidecar in
run_nat_lan, mirroring the bloom-storm pattern. Pinning is needed
because the mesh-lab compose-resource-limits.yml override is
rekey-family service-name specific (rekey-* / rekey-accept-off-* /
rekey-outbound-only-*), so it does not constrain the nat-lan
containers. Default cpuset 0,1 mimics a GHA 2-core runner; empty
disables the sidecar.
- New compose-trace-nat.yml overlay that bumps RUST_LOG to trace on
discovery::nostr, transport::udp, node::lifecycle,
handlers::handshake, handlers::forwarding — the modules covering
the cross-init / adoption / handshake path. Picked up by the
nat-test.sh COMPOSE array via a new FIPS_NAT_EXTRA_COMPOSE
colon-separated env-var hook. run_nat_lan sets this hook when
FIPS_MESH_LAB_TRACE is non-empty; the README env-var section
updated to reflect that FIPS_MESH_LAB_TRACE now applies to nat-lan
in addition to the rekey-family.
- parse_nat_lan extended with a per-node stall_signature emitting
last-occurrence timestamps for eight event categories (startup,
discovery, adoption, handshake_init, msg2_sent, cross_init_ignore_*,
handshake_failed) plus derived last_meaningful_event_ts,
last_event_category, silent_gap_s. Top-level stall_class binned as
no_timeout / silent / localized / distributed / incomplete from
the per-node categories. Aggregation phase consumes the per-rep
signatures across a characterization run to classify stall
mechanism.
Wired support in nat-test.sh: FIPS_NAT_EXTRA_COMPOSE colon-separated
list of repo-relative or absolute compose files layered onto the
base via the COMPOSE array; FIPS_NAT_SKIP_FINAL_CLEANUP gates the
success-path teardown so the mesh-lab harness can capture docker
logs before tearing down (failure paths already returned without
cleanup, leaving stall-state containers intact for capture).
Smoke-tested on idle profile with TRACE on: 1 rep PASS, 32/36 TRACE
lines per node, signature.json events all populated with the
expected category timestamps.
The bloom-storm scenario's bloom_send_rate ceiling has been bumped
from 30 to 40 sends per node over the trailing 30 s window. A 59-rep
characterization run under `github-runner-equivalent` pressure with
per-container CPU pinning to `cpuset=0,1` (mimicking a 2-core
`ubuntu-latest` runner) measured n04 (the structural max-spike node)
at mean 24.4, P99 29, max 30. The original ceiling of 30 sat at the
lab's structural max, leaving no headroom for the asymmetric transient
spikes observed on GitHub Actions (n04=34 on master CI run 25933972365,
re-fired on run 26008950865). GHA fires do not reproduce on this lab
host even with the cpuset sidecar applied.
Rationale: lab max + ~2σ ≈ 39.4 → round to 40, giving 33 % margin over
the lab maximum while staying well below the deployment-scale storm
rate (~480× steady state). The companion `min_parent_switches` guard
is unchanged.
See README.md alongside this file for the updated threshold derivation.
Wires the bloom-storm chaos scenario into the mesh-lab harness as
a first-class suite, with optional per-container CPU pinning to
mimic GitHub Actions' 2-core ubuntu-latest budget.
Dispatch path — three new run-loop.sh functions plus the
dispatch_suite and dispatch_mechanism_match case-arm additions:
- `run_bloom_storm` invokes `bash testing/chaos/scripts/chaos.sh
bloom-storm` and captures stdout+stderr into the rep's
test-output.log. Chaos uses its own python sim runner
(`python3 -m sim`), not docker-compose, so this suite gets no
per-container compose override, no separate `docker logs`
capture, and no in-container netem injection — the chaos
scenario yaml owns its own netem and link-swap config.
- `parse_bloom_storm` extracts the bloom_send_rate result
(pass/fail/unknown), ceiling, max-observed per-node delta,
offenders list, full per-node delta distribution, the companion
min_parent_switches result, and panic + error counts. Lands in
the rep's signature.json. Two parser details: assertion greps
are anchored on `^(PASS|FAIL)` so they only match the bare
end-of-run summary line, not python-logger-prefixed lines that
contain the same substring; and `grep -c` panic/error counts
use `; true` + a defensive empty-string check instead of the
common `|| echo 0` fallback (`grep -c` exits 1 on zero matches
while also printing "0", so the fallback would corrupt the
count to "0\\n0").
- `mechanism_match_bloom_storm` returns true when a rep both
fails the bloom_send_rate assertion and the FAIL line carries
a named offender (filtering the harness-side "failed to sample
window endpoints" sub-failure out of the mechanism count).
CPU-pinning sidecar — bloom-storm's chaos sim spawns containers
directly via the docker SDK, so the mesh-lab compose-resource-
limits override does not apply. A poll-and-pin loop around the
chaos.sh invocation lists \`fips-*\` containers every 0.5 s and
applies \`docker update --cpuset-cpus <set>\` to each. Pinning is
idempotent (re-applying the same cpuset is a no-op). Default
cpuset \`0,1\` mimics the GHA 2-core budget; override via
\`FIPS_BLOOM_STORM_CPUSET=<set>\` (any comma-separated CPU list),
or set to the empty string to disable. Only applies to the
bloom-storm suite; other suites' dispatch paths are unchanged.
README's "Suites supported" entry covers the assertion class, and
the \`FIPS_BLOOM_STORM_CPUSET\` knob is documented alongside the
other mesh-lab env-var knobs.
An FSP session rekey could leave the two endpoints holding different
key sets for a brief window: if a handshake message was lost in
transit, one side rotated to the new keys while the other did not.
Traffic sealed in one key epoch then reached a peer still on the
other epoch and failed to decrypt, producing bursts of AEAD
decryption failures and dropped connectivity until a later rekey
cycle reconverged the pair. Choreographing the cutover order cannot
close this window: any fixed ordering still leaves a skew that
packet reordering widens.
Make rekey correctness independent of cutover timing by overlapping
the key epochs on the receive path. During a rekey transition the
receiver trial-decrypts each frame against every live session it
holds: current, the not-yet-promoted pending session, and the
draining previous session. The K-bit becomes a hint that orders the
trial-decrypt cascade rather than a hard gate, and a frame that
authenticates against the pending session is itself the cutover
signal. No rotation ordering and no packet reordering can then cause
a decryption failure.
The pre-rekey Noise session is held in the `previous` slot until the
peer has demonstrably moved off it. Its drain deadline is anchored
on the most recent frame the peer authenticated against that slot,
refreshed each time the trial-decrypt cascade lands there, rather
than on a fixed wall-clock timer started unilaterally at the local
cutover. A peer that never received the new keys keeps authenticating
against `previous` and the slot stays live; without this, a fixed
timer would erase the only key set that could decrypt the peer's
frames, producing a permanent silent decrypt failure on a live data
path. A peer that never catches up is handled by the existing FSP
session liveness path rather than by silent decrypt failure.
The lost-handshake liveness gap is closed separately by retransmitting
the third rekey handshake message until the peer is confirmed on the
new keys, with a bounded retry budget after which the rekey cycle is
cleanly abandoned and retried on the next timer.
Adds unit tests covering the trial-decrypt cascade (epoch selection,
promotion on pending decrypt, reordered old-epoch stragglers after
cutover, per-slot replay-window integrity), the msg3 retransmission
lifecycle, and the peer-progress-aware drain retirement.
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 Phase 1, Phase 3, and Phase 5 strict asserts each fire a
single ping per directed pair. Under low-level packet loss
(e.g. 1% i.i.d. per-direction loss from a CI runner under
pressure), a single-shot round-trip fails at ~2% per pair, so a
20-pair strict assert misses with probability
1 - (0.98)^20 = ~33% per phase from ICMP noise alone, well above
the routing-state signal the asserts are meant to catch.
ping_one gains a max_attempts parameter (default 1, preserving
existing call sites). On failure it retries up to
MAX_PING_ATTEMPTS-1 additional times with PING_RETRY_DELAY
seconds between attempts. Per-pair worst case under the defaults
(4 attempts, 1 s spacing, 5 s ping6 -W timeout) is 4*5 + 3 = 23 s;
per-rep worst case scales with the failing-pair count.
Successful retries log "OK (RTT, attempt N)"; exhausted retries
log "FAIL (after N attempts)".
The retry budget is wired into all three strict asserts:
- Phase 1 final ping_all (after wait_for_full_baseline converges)
- Phase 3 ping_all (post-first-rekey)
- Phase 5 ping_all (post-second-rekey)
The wait_for_full_baseline convergence loop itself stays
single-shot. Its job is to detect when the mesh first sees a
fully clean 20-pair batch, and retries inside the loop would
conflate transient ping loss with still-converging routing
state.
No daemon code changes.
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
Phase 5's per-pair connectivity check ran immediately after the second
rekey cycle, with no settle for routing reconvergence. Under
GitHub-runner CPU contention, post-rekey parent-switches and
coord-cache flushes can take longer than the per-ping 5s timeout for
a small fraction of pairs (1-3 of 20 typically), even though the
rekey mechanism itself completes cleanly. Phase 6 log analysis stays
all-green on these failed runs; the failure is purely connectivity
timing.
Mirror Phase 3's existing 12-second settle pattern: reuse REKEY_SETTLE
and emit the same banner shape. Two-line change. Cost on the success
path is a fixed 12s per suite run.
Convert assert_log_contains from a one-shot grep snapshot into a
bounded poll that retries until the pattern appears or the timeout
elapses (default 15s). Same wait-with-timeout shape as
wait_for_peers_exact above it in the file.
The pre-existing flake on next-branch CI is structural: under XX
handshake, the cross-connection tie-breaker selects which side
reaches its ACL-check point first. When container-a wins the
tie-breaker on the first attempt against c and d, only the
outbound-handshake-context rejection fires immediately, and the
inbound-handshake-context rejection only emits on a later retry
when c or d's msg1 lands while a has no pending outbound. On one
2026-05-14 run the inbound rejection appeared 63ms after the test
had given up. The race window is small but real.
Polling the log instead of one-shot reading absorbs the
millisecond-to-second variance without slowing the success path
(the helper returns as soon as the pattern appears).
The Phase 1 pre-rekey baseline in `wait_for_full_baseline`
occasionally times out on GitHub-hosted runners with one ping pair
failing to converge inside the BASELINE_CONVERGENCE_TIMEOUT window.
Phases 2–6 always pass cleanly when this happens — the rekey itself
is fine, the mesh just hasn't finished spanning-tree + bloom-filter
convergence by the time Phase 1 starts pinging.
The wait loop returns as soon as all 20 pairs converge, so the cost
on the success path is unchanged (typical local CI returns well
under the old 36s). The bump only adds headroom on the failure
path. Operators previously worked around this with
`gh run rerun --failed`; this aims to retire that workaround for
the IK/XK lines.
Distinct from the next-branch XX rekey dual-init race, which is a
real protocol bug tracked separately.
The sidecar entrypoint hardcoded `udp.mtu: 1472`, the Docker-bridge
IPv4 maximum (1500 MTU - 8 UDP - 20 IPv4 header). Promote it to
`FIPS_UDP_MTU` (defaulting to 1472, preserving behavior) so non-Docker
reuses of the example can override without editing the script. Plumb
the env var through `docker-compose.yml` so a host-level setting
reaches the container, and add a row to the README's env-var table.
Also annotate the static-CI node template with a comment explaining
the same 1472 rationale and the daemon's 1280 default. The template
keeps 1472 as the literal value since the CI suite runs on a Docker
bridge where that's correct.
No behavior change unless the host explicitly sets FIPS_UDP_MTU.
Operator-facing IPs in user-visible configs/docs (examples, tutorials,
packaging, sidecar templates) are now the resolvable hostnames of the
public test fleet (test-us01.fips.network, etc.) so they keep working
without baking specific addresses into examples.
Doc-comment and test fixtures in src/config/transport.rs use RFC 5737
TEST-NET-2 (198.51.100.1) so they cannot accidentally point at a real
host.
Also resyncs the openwrt-ipk fips.yaml with the common reference
(merge from master) and applies the same DNS-name swap there.
Johnathan Corgan
Martti MalmiandJohnathan Corgan
Arjen