The runtime/probe changes from #642 and GNU C17/build changes from #643 were squash-merged into #641's source branch, which made the result-normalization review include later stages. This follow-up preserves those approved changes while restoring #641's focused diff against develop. No force-push or approved-code deletion is required. Merge after #641 and before #644. #644 will target this branch. This combines the already reviewed steps 5 and 6; it introduces no new worker-stack or module-extraction changes. The UDP follow-up also corrects a validated finding: EHOSTUNREACH on a connected socket is ambiguous and can represent an unreachable host or failed neighbour lookup. It now reports down; ECONNREFUSED/ECONNRESET and actual peer datagrams retain their existing alive behavior. The production-linked test rejects the original EHOSTUNREACH behavior. Validation: Ubuntu 24.04/aarch64, clean GCC build and make check: 17 passed, one existing setuid prerequisite skip, zero failures. The full tracked source tree is byte-identical to the independently validated #641 functional-fix snapshot. The UDP and poll-host regressions fail against the original code and pass after the fixes. The isolated step-4 tree independently passes 15 tests with one existing skip. No live endpoints, host installation, privilege changes, or test suppression. Repository squash merges still require restacking/retargeting the next child after each merge; clean ordinary merge simulations are insufficient by themselves. --------- Signed-off-by: Thomas Vincent <thomasvincent@gmail.com>
Spine test harness
make check runs the cmocka unit tests. The Docker-backed integration fixture
uses a real Net-SNMP agent and MariaDB or MySQL to exercise the complete poll
and persistence path.
Run the standard end-to-end test with:
SPINE_TEST_DATABASE=mariadb tests/snmpv3/scripts/run-integration.sh
Run the concurrency contract with:
SPINE_TEST_HOSTS=24 \
SPINE_TEST_THREAD_COUNTS="1 4 12" \
SPINE_TEST_REPEATS=1 \
tests/integration/test_concurrency.sh
Exercise timeout and recovery behavior with real UDP packet loss:
tests/integration/test_network_faults.sh
The concurrency test creates an isolated Compose project, clones independent
SNMPv3 host records, sweeps worker counts, and verifies exact output and host
accounting after every run. It is intentionally transport-neutral so it can
also test a future libuv engine and reactor-count matrix without changing its
assertions. Fixture services communicate only on their project-local Compose
network, so independent test projects can run without competing for host ports.
The fault test uses tc netem inside the test-only SNMP agent container to
drop every response, verifies bounded completion without stale output, removes
the fault, and verifies the next poll recovers.
Useful controls:
SPINE_TEST_TIMEOUT_SECONDS: per-Spine-run timeout (default120).SPINE_TEST_ARTIFACT_DIR: destination for logs and TAP results.SPINE_TEST_KEEP_ENV_ON_FAILURE=1: preserve containers after a failure.SPINE_TEST_PROJECT: explicit Compose project name for reproducibility.SPINE_TEST_EXTRA_ARGS: additional whitespace-separated Spine arguments. This provides the hook for future--snmp-engineand reactor-count sweeps; arguments are split but never evaluated as shell code.
Keep CI workloads bounded. Larger host counts and repeat counts belong in a scheduled soak workflow rather than the pull-request critical path.