897f50e430
Two real Alpine containers on an isolated instance, addressed the way a GNS3 user does it (the node's persistent /etc/network/interfaces, applied by init.sh) and driven through the Docker daemon socket the server itself uses. Four scenarios: * fast path: the per-link kernel bridge enslaves exactly the two veth host ends, real ICMP crosses, the L2-anchor spec E.1/E.2 assertions hold (idle silence measured with the guests' links down -- a container's own boot-time IGMP report re-fires ~10 s later, so silencing is not just IPv6), netem delay, suspend/resume, capture, link delete/re-create, container stop/start (the veth host ends die with the adapter and are rebuilt from the NIO), zero residue on project delete; * filter matrix, one type at a time, each asserted through real traffic and its tc state: netem core (delay, jitter+distribution, loss, corrupt, duplicate), the extensions (reorder+correl+gap, seed+limit, gemodel, rate), cls_bpf match-drop including a byte-size discriminator, the eBPF classifier (frequency_drop every-Nth and -1, quota byte-cap cutoff, window_drop single outage with per-seq pass/drop/pass evidence), plus available_filters offering the kernel-only types; * relay control: no per-link bridge, the kernel-only types hidden; * relay->kernel reopen upgrade: the datapath choice is flipped across a server restart and the project reopened -- the link is rebuilt on the kernel datapath and forwards. The harness gains a Docker guest driver (a stdlib Engine API client over the daemon socket: exec with stream demux, container id by deterministic name -- the controller's node payload does not carry it, image inspect), a digest-pinned image helper (cache hit = no network, miss = pull through the server's own route, neither = skip) and Server.restart() for the reopen scenario (which also needs the local compute's capabilities to be up before a project open is accepted). Live: 4/4 scenarios, 142 s for the file; full e2e suite 10/10 (1503 s); unit suite 2149 passed; ruff clean. The matrix measures loss with 20 packets at 0.2 s spacing (143 s -> 58 s) and the jitter RTT assertion keeps statistical margin (per-direction N(50,20); 10 packets, >= 65 ms).
Live end-to-end tests
Tests under this directory run against a real server: real uBridge, real
kernel interfaces (privileged), real emulator binaries and real images. They
verify what unit tests cannot — that a datapath actually forwards traffic,
that kernel bridges actually enslave anchors, that tc impairments actually
shape packets. They are marked e2e and excluded from normal test runs
(see tests/pytest.ini); run them explicitly:
pytest -m e2e tests/e2e -s # -s shows the progress prints
pytest -m e2e tests/e2e -k dynamips # one scenario
Two run modes, chosen per test through harness.live_server(kernel=...):
- Target an existing server — set
GNS3_E2E_URL(e.g.http://127.0.0.1:3080/v3, plusGNS3_E2E_USER/GNS3_E2E_PASSWORD, default admin/admin). The server must already be configured for the datapath under test (kernel links needenable_kernel_datapath = True). - Isolated instance (the default) — the harness starts its own server
process on a free port with its own config directory (the controller DB
and every path follow the config), so nothing on the machine is touched.
It needs
ubridge,dynamipsand an images directory (GNS3_E2E_IMAGES, default~/GNS3/images) available locally; the Docker scenarios additionally need the Docker daemon the server talks to.
Either way each test creates its own project and deletes it again. Set
GNS3_E2E_KEEP=1 to keep the project (and a started instance) on failure
and print where to look at it.
The dynamips scenarios give their routers an idle-PC: detected once per
image through the controller's auto_idlepc (a throwaway project and a
real CPU-usage measurement — the GUI's idle-PC finder path) and cached in
~/.cache/gns3-e2e/idlepc.json keyed by the image checksum. Without one,
each c7200 burns a full CPU core for the whole scenario — the e2e
creates raw nodes, so no template supplies a value. Set GNS3_E2E_IDLEPC
to skip the detection and use the given value (detection failing is never
fatal: the routers simply run without one).
harness.py holds the shared pieces — REST client, server lifecycle, the
IOS console driver, the Docker guest driver (a tiny Docker Engine API
client over the daemon socket, plus the digest-pinned image helper),
host-side kernel inspections (bridge membership, tap admin state, tc
qdiscs, the L2-anchor spec §E assertions) — so a new node type adds only
its own scenario module.
Current scenarios:
test_dynamips_kernel_datapath.py— two real c7200 routers: kernel link (anchors, per-link bridge, netem delay, suspend, capture, delete/re-create, stop/start) plus a serial link staying on the relay, and a negative control on an isolated relay-configured instance.test_ethernet_switch_kernel_datapath.py— a real Ethernet switch (kernel bridge) between two c7200 routers: deferred joins on boot, VLAN isolation, absorbed-anchor filters/suspend/capture, VPCS peers staying on the relay, and a relay negative control.test_iol_docker_kernel_datapath.py— two real iol-xe containers (Cisco CML iol-runner images): kernel link through the port bridge's swappable TAP leg (anchors at start, netem delay, suspend, capture, link delete/re-create swap, node stop/start rewiring) and a relay negative control.test_docker_kernel_datapath.py— two real Alpine containers (the plain veth datapath): the per-link kernel bridge enslaves exactly the two veth host ends, real ICMP crosses, the filter matrix runs one type at a time (netem core and extensions, cls_bpf match-drop, the eBPF classifier — each measured through real traffic), plus a relay negative control and a relay→kernel reopen upgrade (the server's datapath choice is flipped across a restart).
The Docker scenarios run a digest-pinned alpine:3, so every developer
tests the same bytes: a cached copy in the local Docker daemon costs no
network, a cache miss pulls through the server's own image-pull route, and
with neither cache nor registry the scenario skips (GNS3_E2E_DOCKER_IMAGE
overrides the reference). Their guests are addressed the way a GNS3 user
does it — the node's persistent /etc/network/interfaces, which the
container's init.sh applies at boot — and driven through the Docker
daemon socket the server itself uses.
Every kernel scenario also asserts the L2-anchor spec's server-side
guarantees (docs/design/ubridge-l2-anchor-spec.md §E): anchors and bridges
carry no L3 identity (assert_pure_l2), joined ports are FORWARDING
(assert_forwarding), and — in the Dynamips suite — an idle link with the
guests silenced stays silent for 5 s (assert_idle_silence). These skip on
a uBridge without link l2only (probed once by harness.l2only_supported()).
Set GNS3_E2E_DEBUG=1 to run the isolated instance with debug logging (the
uBridge command stream lands in its server.log).