mirror of https://github.com/jmcorgan/fips.git synced 2026-10-30 07:37:36 +00:00

Files

T

Johnathan Corgan c6609f06d9 Stop the NAT lab start from outrunning its own interface waits

The cone and symmetric starts failed now and then with
`Invalid "netns" value`, a node log ending "Timed out waiting for IPv4
on eth0" and, often, a router that had exited with an empty log. Both
entrypoints gave the harness 30 s from container start to deliver their
interface, and setup-topology.sh delivered it with ten serial helper
containers, five per pair. Starting a container is normally 0.16 s, but
while another lab starts on the same Docker daemon it stalls for 3 to
15 s at a time (measured in create and start; exec, inspect and logs
stay fast, and the kernel and disk were not the bottleneck). Ten such
starts overran the window, so the move landed on an exited container.

Reproduced before the change, with nothing injected: two NAT labs
started together on one daemon failed in 4 of 4 starts:

    Error: argument "2219471" is wrong: Invalid "netns" value
    LAB START FAILED: start_cone exited non-zero on attempt 1; ...
    fips-nat-cone-b: Timed out waiting for IPv4 on eth0
    fips-nat-router-b: (empty log)

A docker shim that delays each `docker run ... --pid=host` by 8 s also
failed both of two runs the same way. A single idle lab passes either
way.

The change:

- setup-topology.sh runs one helper container per pair (delete, add and
  both moves in one `sh -c`, each step echoed) instead of five, checks
  that both containers are still running before it, and after a failed
  move or configure step names any container that exited, with its exit
  code and last 20 log lines. wait_for_pid stops early once a container
  has exited.
- Both entrypoints wait IFACE_WAIT_SECS (default 120 s) for their
  interface, separate from WAIT_TIMEOUT_SECS (30 s), which still bounds
  the relay and STUN waits. The default is set in the scripts, not the
  compose file, so the change applies unchanged to the other branches.
- The router entrypoint now says which interface it timed out on and
  lists its links, instead of exiting silently under set -e.
- testing/nat/README.md describes the start order and the bound.

The longer wait bounds harness setup, not product behaviour; a stuck
setup still fails, naming the container.

After the change: the two-labs-together start passed 4 of 4, with
router B configured at 31 to 34 s after compose up in the symmetric
lab, past the old 30 s window, so both halves of the change are
needed. The 8 s shim passed twice, and the symmetric scenario alone
passed. Breaking the guard with a 130 s shim delay fails with
"fips-nat-cone-a exited (code 1) before its interface was moved",
followed by its log, "Timed out after 120s waiting for IPv4 on eth0".

There is no committed regression test: reproducing this needs Docker
and a second lab or a delaying shim, so it stays a captured run.

2026-10-15 21:52:38 +00:00

4.9 KiB

Raw Permalink Blame History

NAT Lab Harness

Real Docker-based NAT traversal integration tests for the mainline FIPS Nostr/STUN bootstrap path.

This harness spins up:

two FIPS nodes
a local Nostr relay
a local STUN server
one or two Linux router containers performing NAT with iptables

For the NAT scenarios, the node LAN interfaces are not attached to Docker bridge networks. The harness creates explicit veth pairs and moves them into the node and router namespaces after docker compose up so every packet must traverse the router namespace.

It covers three scenarios:

cone: both peers behind explicit namespace/veth full-cone emulation, UDP traversal succeeds
symmetric: both peers behind symmetric-style NAT, UDP traversal fails, TCP fallback succeeds
lan: both peers share a LAN subnet, LAN targets are preferred over reflexive addresses

NAT model notes

The harness does not rely on plain Docker MASQUERADE for the cone case.

cone
- uses explicit full-cone emulation in the router namespace
- outbound UDP is SNATed to the router WAN address while preserving the source port
- inbound UDP to the router WAN address is DNATed back to the single LAN host regardless of remote source
symmetric
- uses UDP MASQUERADE --random-fully
- outbound mappings may be port-randomized and are only reopened by matching conntrack state

This distinction matters because plain MASQUERADE is convenient source NAT, but it does not by itself model the "accept from any remote once mapped" behavior expected from a full-cone NAT.

Prerequisites

Docker with Compose support
locally built fips-test:latest

Build the test image with:

./testing/scripts/build.sh

Run

Run all scenarios:

./testing/nat/scripts/nat-test.sh

Run one scenario:

./testing/nat/scripts/nat-test.sh cone
./testing/nat/scripts/nat-test.sh symmetric
./testing/nat/scripts/nat-test.sh lan

Layout

docker-compose.yml
- relay/STUN/WAN topology plus container definitions
node/
- node bootstrap wrapper that waits for the injected veth interface (IFACE_WAIT_SECS)
router/
- NAT router image and iptables setup
stun/
- minimal STUN binding responder
relay/
- local strfry config. The relay image's strfry build is pinned by digest in docker-compose.yml (STRFRY_IMAGE); bump it there deliberately.
scripts/generate-configs.sh
- derives ephemeral identities and writes per-scenario FIPS configs
scripts/setup-topology.sh
- injects and configures the NAT LAN veth pairs in the container namespaces; see "Start order and interface waits"
scripts/nat-test.sh
- boots the lab, waits for convergence, and asserts the resulting path
scripts/nostr-relay-test.sh
- exercises the Nostr overlay advert publish/consume round-trip, including rejection of a malformed advert event
scripts/stun-faults-test.sh
- cycles the daemon through STUN drop, delay and outage faults and asserts graceful behavior at each step

Start order and interface waits

The cone and symmetric starts run in this order:

docker compose up -d starts the relay, STUN, both routers and both nodes. The routers have only their WAN interface and the nodes have none (network_mode: none).
Each entrypoint then waits for the interface the harness will give it: a node for an IPv4 address on eth0 (node/entrypoint.sh), a router for eth1 (router/entrypoint.sh).
scripts/setup-topology.sh wires the pairs one after the other. For each pair it checks that both containers are still running, then runs one helper container (the router image, --net=host --pid=host) that deletes any stale interfaces, creates the veth pair and moves each end into its container's namespace. It then renames and addresses the ends with docker exec, which is what ends both waits.

The interface waits are bounded by IFACE_WAIT_SECS, 120 s by default and set in each entrypoint rather than in docker-compose.yml. The relay and STUN waits in the node entrypoint keep their own WAIT_TIMEOUT_SECS (30 s). The interface bound covers harness setup, not product behaviour: starting a container has been measured at 3 to 15 s at a time while other labs start on the same Docker daemon, and the earlier 30 s bound with five helper containers per pair was overrun often enough to fail the suite.

When a wait does run out, the container says so in its log and exits, and setup-topology.sh fails with <container> exited (code N) before its interface was moved, followed by the end of that container's log.

Assertions

cone
- both nodes connect
- connected transport is UDP
- active link remote addresses are on the WAN NAT subnet
symmetric
- NAT bootstrap does not establish a UDP link
- fallback converges
- connected transport is TCP via router-published WAN addresses
lan
- both nodes connect
- connected transport is UDP
- active link remote addresses stay on the shared LAN subnet

4.9 KiB Raw Permalink Blame History