mirror of
https://github.com/GNS3/gns3-server.git
synced 2026-11-21 05:35:15 +02:00
feat: cross-compute VXLAN kernel datapath
Cross-compute links can ride kernel VXLAN tunnels instead of the uBridge
UDP relay: each endpoint creates one VXLAN device (uBridge's rtnetlink
vxlan module) enslaved into the same per-link Linux bridge the same-host
datapath uses, so filters, markers, capture and suspend keep running on
the anchor unchanged. Opt-in (Server.enable_vxlan_cross_compute, default
off) because the tunnel caps the usable MTU at the underlay minus the
outer headers; the relay stays the default and the fallback.
The controller derives what the pair must agree on: one VNI per link
(upper VNI space, per-link random salt) and deterministic device names
(gvl{link}{node}) that make crashed-run leftovers claimable
(show -> delete -> create). Same-host compute pairs, switch endpoints
and serial/unix-socket containers stay on the relay.
A VNI the kernel refuses self-heals on every path it can surface:
interactive create retries a fresh salt then degrades to the relay;
project open re-dispatches refused batch entries the same way; node
start recognises the refused device, stops the half-started node,
re-wires both ends on a fresh VNI (or degrades the link) and retries
the start. The device MTU is derived from the lower device (dev=) and
the per-link bridge inherits it - no configuration. The capability
probe (ubridge_vxlan in /capabilities) runs one scratch create/delete
on a scratch port that never equals the configured dstport.
The e2e scenario runs two computes on one machine (a network namespace
joined by a veth pair as the fabric) and needs root.
This commit is contained in:
24 files changed
+2107
-59
No files matched your search
@@ -27,6 +27,7 @@ Beyond a Linux kernel with the datapath features (bridges, veth/TAP devices, tc
|
||||
* **`tc` (iproute2)** for the filter scenarios: their qdisc and classifier assertions read it from `PATH`. Without it those assertions cannot run.
|
||||
* **A Docker daemon** for the Docker scenarios (the harness's own guest driver talks to the same daemon as the server; socket override: `GNS3_E2E_DOCKER_SOCKET`).
|
||||
* **A C compiler that can link statically** (`cc -static`) for the raw-frame probe of the link-local scenario; without it that probe skips.
|
||||
* **Root (or passwordless sudo) for the cross-compute scenario** — a second network stack on one machine is a network namespace, and Linux confines creating and entering one to CAP_SYS_ADMIN (uBridge's file capabilities do not help there). The scenario skips with the exact run command otherwise; it also writes the second compute's node files, which the harness's passwordless-sudo path cannot do as the test user.
|
||||
* **Emulator images** — see the matrix below.
|
||||
|
||||
## Test images
|
||||
@@ -93,3 +94,5 @@ Every kernel scenario also asserts the L2-anchor spec's server-side guarantees (
|
||||
* `GNS3_E2E_DOCKER_SOCKET` — the Docker daemon socket the harness's guest driver talks to (default `/var/run/docker.sock` — the same daemon the server uses).
|
||||
* `GNS3_E2E_IDLEPC` — skip idle-PC detection for the Dynamips scenarios and use the given value.
|
||||
* `GNS3_E2E_DEBUG=1` — run the isolated instance with debug logging (the uBridge command stream lands in its `server.log`).
|
||||
|
||||
* `test_vxlan_cross_compute.py` — the cross-compute VXLAN datapath, with two computes on one machine: the isolated instance on the host and a second isolated instance inside a network namespace joined to the host by a veth pair (`harness.cross_compute_fabric` — two computes cannot tunnel from the host namespace, since the kernel keeps one VXLAN socket per (namespace, dstport) and the pair must share one VNI on it). One real Alpine container per compute, linked across the tunnel: both ends agree on the VNI and the shared dstport, each per-link bridge enslaves exactly `[anchor, tunnel]`, the tunnel device's MTU is the underlay minus 50 with the bridges inheriting it (a frame at it crosses, one past it is dropped at the bridge egress), dual-stack ICMP crosses, a `delay 100` filter shapes both directions, a stale device squatting on the link's deterministic name is replaced on node restart, a device squatting the link's VNI at node start self-heals (both ends re-wired on a fresh VNI, the start succeeding), link delete leaves nothing behind on either machine, and a project close/reopen rebuilds the tunnel through the batch path. **Needs root** (see host requirements) — run it with `sudo env GNS3_E2E_IMAGES=$HOME/GNS3/images pytest -m e2e tests/e2e/test_vxlan_cross_compute.py -s`.
|
||||
+257
-16
@@ -67,6 +67,7 @@ import time
|
||||
import urllib.error
|
||||
import urllib.parse
|
||||
import urllib.request
|
||||
import uuid
|
||||
|
||||
import pytest
|
||||
from websockets.sync.client import connect as ws_connect
|
||||
@@ -145,14 +146,15 @@ class Compute:
|
||||
{"compute_id": "local", "name": name, "node_type": "dynamips", "properties": properties},
|
||||
)
|
||||
|
||||
def create_docker_node(self, project_id, name, image, adapters=1):
|
||||
def create_docker_node(self, project_id, name, image, adapters=1, compute_id="local"):
|
||||
"""A standard Docker node (the veth datapath): the scenario addresses
|
||||
its eth0 from the host and pings through it."""
|
||||
its eth0 from the host and pings through it. *compute_id* places it
|
||||
on a remote compute (the cross-compute VXLAN scenario's peer)."""
|
||||
return self.call(
|
||||
"POST",
|
||||
f"/projects/{project_id}/nodes",
|
||||
{
|
||||
"compute_id": "local",
|
||||
"compute_id": compute_id,
|
||||
"name": name,
|
||||
"node_type": "docker",
|
||||
"properties": {"image": image, "adapters": adapters, "console_type": "telnet"},
|
||||
@@ -295,10 +297,11 @@ default_admin_username = admin
|
||||
default_admin_password = admin
|
||||
|
||||
[Server]
|
||||
host = 127.0.0.1
|
||||
host = {host}
|
||||
port = {port}
|
||||
ubridge_path = {ubridge}
|
||||
enable_kernel_datapath = {kernel}
|
||||
enable_vxlan_cross_compute = {vxlan}
|
||||
report_errors = False
|
||||
projects_path = {tmp}/projects
|
||||
configs_path = {tmp}/configs
|
||||
@@ -330,10 +333,14 @@ class Server:
|
||||
instance (own config dir — the controller DB and every path follow it —
|
||||
under a temporary directory removed on stop)."""
|
||||
|
||||
def __init__(self, compute, process=None, tmpdir=None):
|
||||
def __init__(self, compute, process=None, tmpdir=None, privilege=()):
|
||||
self.compute = compute
|
||||
self.process = process
|
||||
self.tmpdir = tmpdir
|
||||
# Command prefix for cleaning up the temp directory when the process
|
||||
# ran privileged (a compute spawned inside a netns as root writes
|
||||
# root-owned files the test user cannot remove itself).
|
||||
self.privilege = privilege
|
||||
|
||||
@property
|
||||
def isolated(self):
|
||||
@@ -361,6 +368,11 @@ class Server:
|
||||
self._stop_process()
|
||||
if self.tmpdir:
|
||||
shutil.rmtree(self.tmpdir, ignore_errors=True)
|
||||
if self.privilege and os.path.exists(self.tmpdir):
|
||||
# S603: fixed argv (the privilege prefix is sudo -n or empty)
|
||||
subprocess.run( # noqa: S603
|
||||
[*self.privilege, "rm", "-rf", self.tmpdir], capture_output=True, check=False
|
||||
)
|
||||
|
||||
def restart(self, kernel=True):
|
||||
"""Restart an isolated instance in place: stop the process — the
|
||||
@@ -376,10 +388,15 @@ class Server:
|
||||
return self
|
||||
|
||||
|
||||
def _spawn_isolated(tmpdir, kernel):
|
||||
def _spawn_isolated(tmpdir, kernel, host="127.0.0.1", vxlan=False, netns=None, privilege=()):
|
||||
"""Start one isolated instance inside *tmpdir* (its config is written
|
||||
there; a fresh free port each boot) and return ``(process, compute)``,
|
||||
logged in. Shared by the first boot and by ``Server.restart``."""
|
||||
logged in. Shared by the first boot and by ``Server.restart``.
|
||||
|
||||
With *netns* the instance is spawned inside that network namespace
|
||||
(through *privilege*, e.g. ``sudo -n``) and binds *host* — the pattern
|
||||
the cross-compute VXLAN scenario uses for its second "machine".
|
||||
"""
|
||||
ubridge = shutil.which("ubridge")
|
||||
dynamips = shutil.which("dynamips")
|
||||
if not ubridge:
|
||||
@@ -402,7 +419,15 @@ def _spawn_isolated(tmpdir, kernel):
|
||||
fork_build = os.path.expanduser("~/myCode/GNS3/vpcs/src/vpcs")
|
||||
vpcs = fork_build if os.path.isfile(fork_build) else (shutil.which("vpcs") or "vpcs")
|
||||
config = ISOLATED_CONFIG.format(
|
||||
port=port, ubridge=ubridge, dynamips=dynamips, images=images, tmp=tmpdir, kernel=kernel, vpcs=vpcs
|
||||
port=port,
|
||||
ubridge=ubridge,
|
||||
dynamips=dynamips,
|
||||
images=images,
|
||||
tmp=tmpdir,
|
||||
kernel=kernel,
|
||||
vpcs=vpcs,
|
||||
vxlan=vxlan,
|
||||
host=host,
|
||||
)
|
||||
config_path = os.path.join(tmpdir, "gns3_server.conf")
|
||||
with open(config_path, "w") as f:
|
||||
@@ -410,7 +435,15 @@ def _spawn_isolated(tmpdir, kernel):
|
||||
|
||||
log = open(os.path.join(tmpdir, "server.log"), "a")
|
||||
# S603: fixed argv, this interpreter, no shell involvement
|
||||
argv = [sys.executable, "-m", "gns3server", "--config", config_path]
|
||||
argv = [
|
||||
*privilege,
|
||||
*(["ip", "netns", "exec", netns] if netns else []),
|
||||
sys.executable,
|
||||
"-m",
|
||||
"gns3server",
|
||||
"--config",
|
||||
config_path,
|
||||
]
|
||||
if os.environ.get("GNS3_E2E_DEBUG"):
|
||||
argv.append("-d") # debug logging — the uBridge command stream lands in server.log
|
||||
process = subprocess.Popen( # noqa: S603
|
||||
@@ -420,13 +453,13 @@ def _spawn_isolated(tmpdir, kernel):
|
||||
stderr=subprocess.STDOUT,
|
||||
)
|
||||
process.log = log # keep the file object alive with the process handle
|
||||
compute = Compute(f"http://127.0.0.1:{port}/v3")
|
||||
compute = Compute(f"http://{host}:{port}/v3")
|
||||
|
||||
deadline = time.time() + 60
|
||||
while time.time() < deadline:
|
||||
try:
|
||||
# S310: the port is ours, the host loopback
|
||||
with urllib.request.urlopen(f"http://127.0.0.1:{port}/v3/version", timeout=2):
|
||||
# S310: the address is ours (loopback or the scenario's veth fabric)
|
||||
with urllib.request.urlopen(f"http://{host}:{port}/v3/version", timeout=2):
|
||||
break
|
||||
except Exception:
|
||||
if process.poll() is not None:
|
||||
@@ -446,13 +479,13 @@ def _spawn_isolated(tmpdir, kernel):
|
||||
return process, compute
|
||||
|
||||
|
||||
def _start_isolated(kernel):
|
||||
def _start_isolated(kernel, vxlan=False):
|
||||
tmpdir = tempfile.mkdtemp(prefix="gns3-e2e-")
|
||||
process, compute = _spawn_isolated(tmpdir, kernel)
|
||||
process, compute = _spawn_isolated(tmpdir, kernel, vxlan=vxlan)
|
||||
return Server(compute, process, tmpdir)
|
||||
|
||||
|
||||
def live_server(kernel=True):
|
||||
def live_server(kernel=True, vxlan=False):
|
||||
"""The server under test.
|
||||
|
||||
With ``GNS3_E2E_URL`` set, that server is used for kernel mode (it must
|
||||
@@ -475,7 +508,7 @@ def live_server(kernel=True):
|
||||
).login()
|
||||
server = Server(compute)
|
||||
else:
|
||||
server = _start_isolated(kernel)
|
||||
server = _start_isolated(kernel, vxlan=vxlan)
|
||||
if not wait_until(lambda: (server.compute.capabilities() or {}).get("version"), timeout=45):
|
||||
raise AssertionError(
|
||||
f"{server.compute.url}: the controller's local compute did not report capabilities within 45s"
|
||||
@@ -1505,3 +1538,211 @@ def pcap_icmp_rtt(path, ethertype):
|
||||
deltas.append((ts - pending) * 1000)
|
||||
pending = None
|
||||
return sum(deltas) / len(deltas) if deltas else None
|
||||
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Cross-compute fabric: a second "machine" for VXLAN links
|
||||
# ---------------------------------------------------------------------------
|
||||
|
||||
# The /24 both ends of the veth "fabric" are addressed in. RFC 2544's
|
||||
# benchmarking range: never routed on a real network, so it cannot collide
|
||||
# with anything the host legitimately talks to.
|
||||
FABRIC_SUBNET_PREFIX = "198.18.7"
|
||||
|
||||
|
||||
def fabric_privilege():
|
||||
"""The command prefix for network-namespace operations: nothing when the
|
||||
suite runs as root, ``sudo -n`` when passwordless sudo is configured,
|
||||
None otherwise (the caller skips with the explanation).
|
||||
|
||||
A second network stack — the thing a cross-compute VXLAN link needs on
|
||||
a single machine — is a network namespace, and Linux confines both
|
||||
creating one and entering one to CAP_SYS_ADMIN of its owning user
|
||||
namespace. uBridge's file capabilities do not help here (they are caps
|
||||
for netlink *operations*, not for joining a namespace), so either the
|
||||
whole run is root or sudo carries the two calls.
|
||||
"""
|
||||
|
||||
if os.geteuid() == 0:
|
||||
return []
|
||||
sudo = shutil.which("sudo")
|
||||
# S603: resolved binary, fixed argv, no shell
|
||||
if sudo and subprocess.run([sudo, "-n", "true"], capture_output=True, check=False).returncode == 0: # noqa: S603
|
||||
return [sudo, "-n"]
|
||||
return None
|
||||
|
||||
|
||||
class CrossComputeFabric:
|
||||
"""A second "machine" on one host: a network namespace joined to the
|
||||
host by a veth pair, both ends addressed inside ``198.18.7.0/24``.
|
||||
|
||||
Two computes on one machine cannot tunnel to each other from the host
|
||||
namespace: the kernel keeps one VXLAN socket per (namespace, dstport)
|
||||
and both ends must register the same VNI on it — the pair collides by
|
||||
design. Distinct network namespaces give the pair its two sockets, and
|
||||
the veth pair is the L2 fabric the tunnels ride. uBridge's file
|
||||
capabilities make the *devices* work inside the namespace; only the
|
||||
namespace itself needs privilege (see ``fabric_privilege``).
|
||||
"""
|
||||
|
||||
def __init__(self, netns, host_ifc, peer_ifc, host_ip, peer_ip, prefix):
|
||||
self.netns = netns
|
||||
self.host_ifc = host_ifc
|
||||
self.peer_ifc = peer_ifc
|
||||
self.host_ip = host_ip
|
||||
self.peer_ip = peer_ip
|
||||
self.prefix = prefix
|
||||
|
||||
def netns_run(self, *argv, **kwargs):
|
||||
"""Run a command inside the namespace (privileged, like every netns
|
||||
access)."""
|
||||
kwargs.setdefault("capture_output", True)
|
||||
kwargs.setdefault("text", True)
|
||||
# S603: fixed argv (the caller builds the netns-side command)
|
||||
return subprocess.run( # noqa: S603
|
||||
[*self.prefix, "ip", "netns", "exec", self.netns, *argv], check=False, **kwargs
|
||||
)
|
||||
|
||||
def iface_mtu(self, name):
|
||||
"""The MTU of an interface inside the namespace (None when absent)."""
|
||||
out = self.netns_run("ip", "-o", "link", "show", name).stdout
|
||||
match = re.search(r"\bmtu (\d+)", out)
|
||||
return int(match.group(1)) if match else None
|
||||
|
||||
def iface_exists(self, name):
|
||||
return self.netns_run("ip", "link", "show", name).returncode == 0
|
||||
|
||||
def vxlan_show(self, name):
|
||||
"""``ip -d link show`` details of an interface inside the namespace,
|
||||
as one string (None when absent) — carries the ``vxlan id <vni>
|
||||
... dstport <port>`` attributes."""
|
||||
result = self.netns_run("ip", "-d", "link", "show", name)
|
||||
return result.stdout if result.returncode == 0 and name in result.stdout else None
|
||||
|
||||
def bridge_members(self, bridge):
|
||||
"""Interfaces enslaved to a bridge inside the namespace, or None
|
||||
when the bridge does not exist."""
|
||||
result = self.netns_run("ls", f"/sys/class/net/{bridge}/brif")
|
||||
if result.returncode != 0:
|
||||
return None
|
||||
return sorted(result.stdout.split())
|
||||
|
||||
def brport_state(self, name):
|
||||
"""An interface's bridge-port STP state inside the namespace (None
|
||||
when it is not enslaved; the kernel reports 3 for forwarding)."""
|
||||
result = self.netns_run("cat", f"/sys/class/net/{name}/brport/state")
|
||||
if result.returncode != 0:
|
||||
return None
|
||||
return result.stdout.strip() or None
|
||||
|
||||
def teardown(self):
|
||||
"""Delete the namespace (all its devices go with it) and any host
|
||||
end a half-created pair left behind."""
|
||||
# S603: fixed argv, resolved ip binary, no shell
|
||||
subprocess.run( # noqa: S603
|
||||
[*self.prefix, "ip", "netns", "del", self.netns], capture_output=True, check=False
|
||||
)
|
||||
subprocess.run( # noqa: S603
|
||||
[*self.prefix, "ip", "link", "del", self.host_ifc], capture_output=True, check=False
|
||||
)
|
||||
|
||||
|
||||
def cross_compute_fabric():
|
||||
"""Create the fabric, or skip the calling test when this host cannot.
|
||||
|
||||
The names carry a random suffix so concurrent runs (and leftovers of a
|
||||
crashed one) cannot collide; the caller must call ``teardown()`` in a
|
||||
finally."""
|
||||
|
||||
prefix = fabric_privilege()
|
||||
if prefix is None:
|
||||
pytest.skip(
|
||||
"cross-compute VXLAN needs a second network namespace: run the suite as root, e.g. "
|
||||
"`sudo env GNS3_E2E_IMAGES=$HOME/GNS3/images pytest -m e2e "
|
||||
"tests/e2e/test_vxlan_cross_compute.py -s`, or configure passwordless sudo"
|
||||
)
|
||||
|
||||
suffix = uuid.uuid4().hex[:6]
|
||||
fabric = CrossComputeFabric(
|
||||
netns=f"gns3vx{suffix}",
|
||||
host_ifc=f"vxh{suffix}",
|
||||
peer_ifc=f"vxp{suffix}",
|
||||
host_ip=f"{FABRIC_SUBNET_PREFIX}.1",
|
||||
peer_ip=f"{FABRIC_SUBNET_PREFIX}.2",
|
||||
prefix=prefix,
|
||||
)
|
||||
try:
|
||||
# S603: fixed argv, resolved ip binary, no shell
|
||||
subprocess.run( # noqa: S603
|
||||
[*prefix, "ip", "netns", "add", fabric.netns], check=True, capture_output=True
|
||||
)
|
||||
subprocess.run( # noqa: S603
|
||||
[*prefix, "ip", "link", "add", fabric.host_ifc, "type", "veth", "peer", "name", fabric.peer_ifc],
|
||||
check=True,
|
||||
capture_output=True,
|
||||
)
|
||||
subprocess.run( # noqa: S603
|
||||
[*prefix, "ip", "link", "set", fabric.peer_ifc, "netns", fabric.netns], check=True, capture_output=True
|
||||
)
|
||||
subprocess.run( # noqa: S603
|
||||
[*prefix, "ip", "addr", "add", f"{fabric.host_ip}/24", "dev", fabric.host_ifc],
|
||||
check=True,
|
||||
capture_output=True,
|
||||
)
|
||||
subprocess.run( # noqa: S603
|
||||
[*prefix, "ip", "link", "set", fabric.host_ifc, "up"], check=True, capture_output=True
|
||||
)
|
||||
for argv in (
|
||||
("ip", "addr", "add", f"{fabric.peer_ip}/24", "dev", fabric.peer_ifc),
|
||||
("ip", "link", "set", fabric.peer_ifc, "up"),
|
||||
("ip", "link", "set", "lo", "up"),
|
||||
):
|
||||
fabric.netns_run(*argv, check=True)
|
||||
except subprocess.CalledProcessError:
|
||||
fabric.teardown()
|
||||
raise
|
||||
return fabric
|
||||
|
||||
|
||||
def start_isolated_peer(fabric, kernel=True):
|
||||
"""An isolated instance on the other side of the fabric: spawned inside
|
||||
the namespace and bound to its address, so the first server can register
|
||||
it as a remote compute whose network stack is the second "machine"."""
|
||||
|
||||
tmpdir = tempfile.mkdtemp(prefix="gns3-e2e-peer-")
|
||||
process, compute = _spawn_isolated(tmpdir, kernel, host=fabric.peer_ip, netns=fabric.netns, privilege=fabric.prefix)
|
||||
return Server(compute, process, tmpdir, privilege=fabric.prefix)
|
||||
|
||||
|
||||
def register_remote_compute(server, peer, name="vxlan-peer"):
|
||||
"""Register the peer's server on the first server's controller, so links
|
||||
between nodes on the two computes are cross-compute links. Waits until
|
||||
the controller connected it (capabilities reported — the vxlan probe in
|
||||
particular: the controller gates tunnel eligibility on it)."""
|
||||
|
||||
compute = server.compute.call(
|
||||
"POST",
|
||||
"/computes",
|
||||
{
|
||||
"name": name,
|
||||
"protocol": "http",
|
||||
"host": peer.compute.url.split("://", 1)[1].split(":", 1)[0],
|
||||
"port": int(peer.compute.url.rsplit(":", 1)[1].split("/", 1)[0]),
|
||||
"user": peer.compute.username,
|
||||
"password": peer.compute.password,
|
||||
},
|
||||
)
|
||||
compute_id = compute["compute_id"]
|
||||
if not wait_until(lambda: (server.compute.call("GET", f"/computes/{compute_id}") or {}).get("version"), timeout=60):
|
||||
raise AssertionError(f"controller did not connect the remote compute {compute_id} within 60s")
|
||||
if not wait_until(
|
||||
lambda: (
|
||||
((server.compute.call("GET", f"/computes/{compute_id}") or {}).get("capabilities") or {}).get(
|
||||
"ubridge_vxlan"
|
||||
)
|
||||
is True
|
||||
),
|
||||
timeout=60,
|
||||
):
|
||||
raise AssertionError(f"remote compute {compute_id} never reported ubridge_vxlan (its netns probe failed)")
|
||||
return compute_id
|
||||
@@ -0,0 +1,305 @@
|
||||
#
|
||||
# Copyright (C) 2026 GNS3 Technologies Inc.
|
||||
#
|
||||
# This program is free software: you can redistribute it and/or modify
|
||||
# it under the terms of the GNU General Public License as published by
|
||||
# the Free Software Foundation, either version 3 of the License, or
|
||||
# (at your option) any later version.
|
||||
#
|
||||
# This program is distributed in the hope that it will be useful,
|
||||
# but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||||
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||||
# GNU General Public License for more details.
|
||||
#
|
||||
# You should have received a copy of the GNU General Public License
|
||||
# along with this program. If not, see <http://www.gnu.org/licenses/>.
|
||||
|
||||
"""
|
||||
Live end-to-end test for cross-compute VXLAN links (pytest marker ``e2e``):
|
||||
two computes on one machine — the harness's isolated instance on the host and
|
||||
a second isolated instance inside a network namespace joined to the host by a
|
||||
veth pair (``harness.cross_compute_fabric``) — with one real Alpine container
|
||||
per compute, linked across the tunnel.
|
||||
|
||||
Two computes on one machine cannot tunnel from the host namespace: the kernel
|
||||
keeps one VXLAN socket per (namespace, dstport), and both ends of a link must
|
||||
register the same VNI on it, so the pair collides by design (the controller
|
||||
falls back to the relay for same-host pairs). Separate network namespaces
|
||||
give the pair its two sockets, and the veth pair is the L2 fabric the tunnels
|
||||
ride — the closest a single machine gets to two hosts. The same-host
|
||||
fallback itself is asserted by the controller unit tests; this scenario is
|
||||
the real-kernel proof of the tunnel.
|
||||
|
||||
Needs root: a network namespace is privileged to create and to enter, and the
|
||||
second compute's node files (its project directories) are written by the
|
||||
test. Run it as::
|
||||
|
||||
sudo env GNS3_E2E_IMAGES=$HOME/GNS3/images pytest -m e2e \\
|
||||
tests/e2e/test_vxlan_cross_compute.py -s
|
||||
|
||||
What it asserts:
|
||||
|
||||
* both ends report the same VNI on the shared dstport, each pointing at the
|
||||
other machine's address, and each per-link bridge enslaves exactly
|
||||
``[anchor, tunnel]``;
|
||||
* the tunnel device's MTU is the underlay MTU minus 50 (the kernel derives
|
||||
it from the lower device the peer route resolves to), the bridges inherit
|
||||
it, and a frame past it is dropped at the bridge egress while one at it
|
||||
crosses — the reason the datapath is opt-in on non-jumbo fabrics;
|
||||
* real dual-stack ICMP crosses, a ``delay 100`` filter shapes it (one end
|
||||
per direction), link delete leaves no device or bridge on either machine;
|
||||
* a stale tunnel device squatting on the link's deterministic name is
|
||||
replaced by the attach path, and the tunnel recovers;
|
||||
* a device squatting the link's VNI at node start self-heals: the start
|
||||
re-wires the link on a fresh VNI — the peer re-wired too — and succeeds.
|
||||
"""
|
||||
|
||||
import os
|
||||
import re
|
||||
import subprocess
|
||||
import time
|
||||
|
||||
import pytest
|
||||
|
||||
from gns3server.utils.vxlan import VXLAN_VNI_MAX, VXLAN_VNI_MIN, vxlan_device_name
|
||||
from tests.e2e import harness
|
||||
|
||||
pytestmark = pytest.mark.e2e
|
||||
|
||||
# every standard Docker adapter has a single port
|
||||
ETH = (0, 0)
|
||||
G1_IP = "10.66.1.1"
|
||||
G2_IP = "10.66.1.2"
|
||||
# the dual-stack half: a ULA mirroring the v4 addressing
|
||||
G1_IP6 = "fd00:66:1::1"
|
||||
G2_IP6 = "fd00:66:1::2"
|
||||
|
||||
# ICMP payloads against the derived tunnel MTU (underlay 1500 -> tunnel
|
||||
# 1450): 1400 + 42 bytes of headers is a 1442-byte frame, under the bridge
|
||||
# egress gate; 1500 + 42 is a 1542-byte frame, over it — and over the
|
||||
# guest's own 1500 MTU too, so the IP stack may fragment it; either way the
|
||||
# pieces cannot cross the 1450-wide egress and the ping must fail silently.
|
||||
FIT_PAYLOAD = 1400
|
||||
TOO_BIG_PAYLOAD = 1500
|
||||
|
||||
|
||||
def _ip(*argv):
|
||||
"""Run an iproute2 command on the host (the scenario runs as root)."""
|
||||
ip = harness.IP
|
||||
assert ip, "ip binary not found"
|
||||
# S603: resolved binary, fixed argv from the test itself
|
||||
return subprocess.run([ip, *argv], capture_output=True, text=True, check=True) # noqa: S603
|
||||
|
||||
|
||||
def _host_vxlan_attrs(name):
|
||||
"""(vni, dstport, remote) of a host-side tunnel device, from
|
||||
``ip -d link show``."""
|
||||
details = _ip("-d", "link", "show", name).stdout
|
||||
vni = re.search(r"\bvxlan id (\d+)", details)
|
||||
dstport = re.search(r"\bdstport (\d+)", details)
|
||||
remote = re.search(r"\bremote (\S+)", details)
|
||||
assert vni and dstport and remote, details
|
||||
return int(vni.group(1)), int(dstport.group(1)), remote.group(1)
|
||||
|
||||
|
||||
def _peer_vxlan_attrs(fabric, name):
|
||||
details = fabric.vxlan_show(name)
|
||||
assert details, f"tunnel device {name} missing inside {fabric.netns}"
|
||||
vni = re.search(r"\bvxlan id (\d+)", details)
|
||||
dstport = re.search(r"\bdstport (\d+)", details)
|
||||
remote = re.search(r"\bremote (\S+)", details)
|
||||
assert vni and dstport and remote, details
|
||||
return int(vni.group(1)), int(dstport.group(1)), remote.group(1)
|
||||
|
||||
|
||||
def _container_id(daemon, node):
|
||||
container_id = daemon.container_id(f"GNS3.{node['name']}.{node['project_id']}")
|
||||
assert container_id, f"container for {node['name']} not found on the Docker daemon"
|
||||
node["container_id"] = container_id
|
||||
return container_id
|
||||
|
||||
|
||||
def _start_pair(compute, pid, daemon, n1, n2):
|
||||
"""Start both containers and wait for their userspace to be reachable:
|
||||
the v4 address from the interfaces file (init.sh), the v6 twin from the
|
||||
harness (the daemon's disable_ipv6 on bridge networks makes the
|
||||
interfaces-file route fail). The container ids are re-resolved after the
|
||||
starts: a project reopen can replace a stale container with a fresh one
|
||||
under the same deterministic name (a new id)."""
|
||||
for node in (n1, n2):
|
||||
compute.call("POST", f"/projects/{pid}/nodes/{node['node_id']}/start")
|
||||
_container_id(daemon, node)
|
||||
for node, address in ((n1, G1_IP), (n2, G2_IP)):
|
||||
harness.docker_wait_address(daemon, node["container_id"], address)
|
||||
for node, address6 in ((n1, G1_IP6), (n2, G2_IP6)):
|
||||
harness.docker_bring_up_ipv6(daemon, node["container_id"], address6)
|
||||
|
||||
|
||||
def test_vxlan_cross_compute_link():
|
||||
if os.geteuid() != 0:
|
||||
pytest.skip(
|
||||
"cross-compute VXLAN needs a second network stack (a netns) and writes the second "
|
||||
"compute's node files: run as root, e.g. `sudo env GNS3_E2E_IMAGES=$HOME/GNS3/images "
|
||||
"pytest -m e2e tests/e2e/test_vxlan_cross_compute.py -s`"
|
||||
)
|
||||
|
||||
fabric = harness.cross_compute_fabric()
|
||||
server = harness.live_server(kernel=True, vxlan=True)
|
||||
peer = harness.start_isolated_peer(fabric)
|
||||
compute = server.compute
|
||||
daemon = harness.DockerDaemon()
|
||||
failed = True
|
||||
pid = None
|
||||
try:
|
||||
peer_id = harness.register_remote_compute(server, peer)
|
||||
image = harness.ensure_docker_image(server)
|
||||
project = compute.create_project("vxlan-e2e")
|
||||
pid = project["project_id"]
|
||||
|
||||
print(".. creating one container per compute")
|
||||
n1 = compute.create_docker_node(pid, "VXE1", image)
|
||||
harness.configure_docker_interfaces(n1, G1_IP)
|
||||
_container_id(daemon, n1)
|
||||
n2 = compute.create_docker_node(pid, "VXE2", image, compute_id=peer_id)
|
||||
harness.configure_docker_interfaces(n2, G2_IP)
|
||||
_container_id(daemon, n2)
|
||||
assert n1["compute_id"] == "local", n1["compute_id"]
|
||||
assert n2["compute_id"] == peer_id, n2["compute_id"]
|
||||
|
||||
print(".. linking them across the computes")
|
||||
link = compute.create_link(pid, (n1["node_id"], *ETH), (n2["node_id"], *ETH))
|
||||
lid = link["link_id"]
|
||||
assert link["kernel_datapath"] is True, link
|
||||
bridge = harness.link_bridge_name(lid)
|
||||
a1 = harness.docker_anchor_name(n1["node_id"], *ETH)
|
||||
a2 = harness.docker_anchor_name(n2["node_id"], *ETH)
|
||||
vx1 = vxlan_device_name(lid, n1["node_id"])
|
||||
vx2 = vxlan_device_name(lid, n2["node_id"])
|
||||
assert not harness.tap_exists(vx1), "the tunnel lands with node start"
|
||||
|
||||
print(".. starting the containers (the wiring lands with node start)")
|
||||
_start_pair(compute, pid, daemon, n1, n2)
|
||||
|
||||
print(".. asserting the tunnel on both machines")
|
||||
assert harness.tap_exists(vx1), f"tunnel device {vx1} missing on the host"
|
||||
assert harness.bridge_members(bridge) == sorted([a1, vx1]), harness.bridge_members(bridge)
|
||||
assert fabric.bridge_members(bridge) == sorted([a2, vx2]), fabric.bridge_members(bridge)
|
||||
vni1, port1, remote1 = _host_vxlan_attrs(vx1)
|
||||
vni2, port2, remote2 = _peer_vxlan_attrs(fabric, vx2)
|
||||
assert vni1 == vni2, (vni1, vni2)
|
||||
assert VXLAN_VNI_MIN <= vni1 <= VXLAN_VNI_MAX, vni1
|
||||
assert port1 == port2, (port1, port2)
|
||||
assert remote1 == fabric.peer_ip, (remote1, fabric.peer_ip)
|
||||
assert remote2 == fabric.host_ip, (remote2, fabric.host_ip)
|
||||
|
||||
# MTU: the kernel derives the tunnel device's MTU from the lower
|
||||
# device the peer route resolves to (underlay - 50 for the outer
|
||||
# headers), and the per-link bridge — no MTU of its own — reports
|
||||
# the minimum of its ports, i.e. the honest "what this fabric can
|
||||
# carry".
|
||||
underlay_mtu = int(harness.sysfs(f"/sys/class/net/{fabric.host_ifc}/mtu"))
|
||||
assert underlay_mtu == 1500, underlay_mtu # the veth default; the derived MTU follows from it
|
||||
assert int(harness.sysfs(f"/sys/class/net/{vx1}/mtu")) == underlay_mtu - 50
|
||||
assert int(harness.sysfs(f"/sys/class/net/{bridge}/mtu")) == underlay_mtu - 50
|
||||
assert fabric.iface_mtu(vx2) == underlay_mtu - 50
|
||||
assert fabric.iface_mtu(bridge) == underlay_mtu - 50
|
||||
assert fabric.iface_mtu(fabric.peer_ifc) == underlay_mtu
|
||||
|
||||
# The tunnel device is an L2 anchor like any other: hardened
|
||||
# (lenient — the shared with plain `ip` in the doc's example) and
|
||||
# FORWARDING on both bridges.
|
||||
if harness.l2only_supported():
|
||||
harness.assert_pure_l2(vx1)
|
||||
harness.assert_forwarding(vx1)
|
||||
assert fabric.brport_state(vx2) == "3", fabric.brport_state(vx2)
|
||||
|
||||
print(".. pinging across the tunnel")
|
||||
ping = harness.docker_wait_ping(daemon, n1["container_id"], G2_IP)
|
||||
assert ping["loss"] == 0, ping["raw"]
|
||||
assert ping["avg"] is not None and ping["avg"] < 50, ping
|
||||
ping6 = harness.docker_wait_ping(daemon, n1["container_id"], G2_IP6)
|
||||
assert ping6["loss"] == 0, ping6["raw"]
|
||||
|
||||
# The derived MTU is the hard limit: a frame at it crosses, one past
|
||||
# it is dropped at the bridge egress — silently, which is why the
|
||||
# deployment has to size the underlay for the frames its guests send.
|
||||
fit = harness.docker_ping(daemon, n1["container_id"], G2_IP, size=FIT_PAYLOAD)
|
||||
assert fit["loss"] == 0, fit["raw"]
|
||||
too_big = harness.docker_ping(daemon, n1["container_id"], G2_IP, count=2, size=TOO_BIG_PAYLOAD)
|
||||
assert too_big["loss"] > 0 or too_big["exit"] != 0, too_big["raw"]
|
||||
|
||||
print(".. delay 100 shapes the tunnel on both ends (one direction each)")
|
||||
compute.call("PUT", f"/projects/{pid}/links/{lid}", {"filters": {"delay": [100]}})
|
||||
time.sleep(1)
|
||||
delayed = harness.docker_ping(daemon, n1["container_id"], G2_IP, count=5, interval=None, timeout=2)
|
||||
assert delayed["loss"] == 0, delayed["raw"]
|
||||
assert 150 <= (delayed["avg"] or 0) <= 450, delayed
|
||||
compute.call("PUT", f"/projects/{pid}/links/{lid}", {"filters": {}})
|
||||
|
||||
print(".. a stale device on the link's name is replaced on node restart")
|
||||
compute.call("POST", f"/projects/{pid}/nodes/{n1['node_id']}/stop")
|
||||
_ip("link", "del", vx1)
|
||||
# A stale tunnel: same deterministic name, some other VNI/port —
|
||||
# what a crashed run or a hand-made device would leave behind.
|
||||
_ip("link", "add", vx1, "type", "vxlan", "id", "4242", "dstport", "47999", "remote", "127.0.0.1")
|
||||
compute.call("POST", f"/projects/{pid}/nodes/{n1['node_id']}/start")
|
||||
harness.docker_wait_address(daemon, n1["container_id"], G1_IP)
|
||||
harness.docker_bring_up_ipv6(daemon, n1["container_id"], G1_IP6)
|
||||
stale_vni, stale_port, stale_remote = _host_vxlan_attrs(vx1)
|
||||
assert (stale_vni, stale_port, stale_remote) == (vni1, port1, remote1), (stale_vni, stale_port, stale_remote)
|
||||
assert harness.bridge_members(bridge) == sorted([a1, vx1]), harness.bridge_members(bridge)
|
||||
assert int(harness.sysfs(f"/sys/class/net/{bridge}/mtu")) == underlay_mtu - 50
|
||||
assert harness.docker_wait_ping(daemon, n1["container_id"], G2_IP, attempts=5)["loss"] == 0
|
||||
|
||||
print(".. a VNI squatted at node start self-heals on a fresh VNI")
|
||||
compute.call("POST", f"/projects/{pid}/nodes/{n1['node_id']}/stop")
|
||||
# A device holding the link's exact VNI on the shared dstport socket
|
||||
# — what a colliding link (or an operator) would have registered.
|
||||
# The node's start replays the stored NIO, the kernel refuses the
|
||||
# VNI, and the controller must re-wire the link on a fresh one and
|
||||
# retry the start instead of failing it.
|
||||
_ip("link", "add", "gvxsquat", "type", "vxlan", "id", str(vni1), "dstport", str(port1), "remote", "127.0.0.1")
|
||||
try:
|
||||
compute.call("POST", f"/projects/{pid}/nodes/{n1['node_id']}/start")
|
||||
harness.docker_wait_address(daemon, n1["container_id"], G1_IP)
|
||||
harness.docker_bring_up_ipv6(daemon, n1["container_id"], G1_IP6)
|
||||
healed_vni, healed_port, healed_remote = _host_vxlan_attrs(vx1)
|
||||
assert healed_vni != vni1, (healed_vni, vni1)
|
||||
assert VXLAN_VNI_MIN <= healed_vni <= VXLAN_VNI_MAX, healed_vni
|
||||
assert healed_port == port1, (healed_port, port1)
|
||||
assert healed_remote == remote1, (healed_remote, remote1)
|
||||
# the peer was re-wired to the fresh VNI too — the pair must agree
|
||||
peer_vni, _, _ = _peer_vxlan_attrs(fabric, vx2)
|
||||
assert peer_vni == healed_vni, (peer_vni, healed_vni)
|
||||
assert harness.docker_wait_ping(daemon, n1["container_id"], G2_IP, attempts=5)["loss"] == 0
|
||||
finally:
|
||||
_ip("link", "del", "gvxsquat")
|
||||
|
||||
print(".. deleting the link leaves no tunnel or bridge behind on either machine")
|
||||
compute.call("DELETE", f"/projects/{pid}/links/{lid}")
|
||||
assert not harness.tap_exists(vx1)
|
||||
assert not harness.tap_exists(bridge)
|
||||
assert not fabric.iface_exists(vx2), f"tunnel device {vx2} survived on the peer"
|
||||
assert fabric.bridge_members(bridge) is None
|
||||
|
||||
print(".. closing and reopening the project rebuilds the tunnel through the batch path")
|
||||
compute.call("POST", f"/projects/{pid}/close")
|
||||
compute.call("POST", f"/projects/{pid}/open")
|
||||
_start_pair(compute, pid, daemon, n1, n2)
|
||||
assert harness.tap_exists(vx1), f"tunnel device {vx1} not rebuilt on reopen"
|
||||
assert fabric.iface_exists(vx2), f"tunnel device {vx2} not rebuilt on reopen"
|
||||
assert harness.docker_wait_ping(daemon, n1["container_id"], G2_IP, attempts=5)["loss"] == 0
|
||||
|
||||
failed = False
|
||||
finally:
|
||||
keep = failed and harness.keep_on_failure()
|
||||
if not keep:
|
||||
peer.stop()
|
||||
if pid is not None:
|
||||
harness.release(server, pid, failed=failed)
|
||||
else:
|
||||
server.stop()
|
||||
if keep:
|
||||
print(f"\n!! kept the fabric {fabric.netns} (and its peer compute) for inspection")
|
||||
else:
|
||||
fabric.teardown()
|
||||
Reference in new issue
Block a user