Files
linux-cloud-router/README.md
T
2026-10-23 22:35:53 +02:00

527 lines
26 KiB
Markdown

# vpn-router
A Debian package that configures a Linux host or virtual machine as a VPN router:
- site-to-site IKEv2 IPSec with pre-shared key authentication (strongSwan swanctl)
- road-warrior (P2S) access using IKEv2 EAP-TLS, with a bundled certificate authority
- OpenVPN (P2S) access with a certificate and a password per user, the users kept in a local
LDAP directory and managed with `vpn-router-user` or the web service
- optional WireGuard endpoint
- routing and NAT for a protected subnet, TCP MSS clamping, and UFW firewall rules
- firewall rules for the ports Docker publishes, when Docker is installed
Nothing in the package is specific to a cloud provider. It is built on the external and
internal NIC names, and a `mode` setting says how much of the rest you are supplying and
how much is read from the system. Detection happens only in the mode that asks for it, so
a router never quietly guesses which interface to NAT out of unless you told it to.
Platform-specific additions live in separate modules and are selected explicitly.
Installing the package configures nothing by default. The shipped platform is `none`,
which means the files are laid down and the machine is left alone until you say
otherwise.
## Installing
The package can be installed by hand or as a step in automated provisioning. Both are
supported equally, and both end up with the same two things: a configuration file and a
service that applies it.
```
/etc/vpn-router/vpn-router.json the configuration
systemctl restart vpn-router-setup apply it
```
If the host runs Docker, the recommended order is Docker first and the package second. The
package secures the ports Docker publishes when it finds the Docker service; the steps for a
host that already runs Docker, and for installing Docker later, are under [Docker](#docker).
### By hand
```sh
apt-get install vpn-router # answer "none" to the platform question
$EDITOR /etc/vpn-router/vpn-router.json
systemctl restart vpn-router-setup
```
Answering `none` to the platform question defers everything: no further questions are
asked and nothing on the machine is configured. The only change an install makes is the
package's own `/etc/sysctl.d/` drop-in, which enables IP forwarding and relaxes
`rp_filter` as a router needs. Set `platform` in the file when you are ready.
### Automated
Preseed the answers and the router comes up configured, with no follow-up command. Note
that `platform` must be something other than `none`, and `mode` says how much you are
supplying:
```shell
debconf-set-selections <<'EOF'
vpn-router vpn-router/platform select generic
vpn-router vpn-router/mode select manual
vpn-router vpn-router/external_interface string eth0
vpn-router vpn-router/internal_interface string eth1
vpn-router vpn-router/int_addr string 10.1.1.4
vpn-router vpn-router/int_gateway_ip string 10.1.1.1
vpn-router vpn-router/local_fqdn string router.example.com
vpn-router vpn-router/local_cidrs string 10.0.0.0/24
vpn-router vpn-router/remote_addrs string peer.example.net
vpn-router vpn-router/remote_id string peer.example.net
vpn-router vpn-router/remote_cidrs string 192.168.0.0/24
vpn-router vpn-router/psk password s3cr3t
EOF
DEBIAN_FRONTEND=noninteractive apt-get install -y vpn-router
```
With mode `interfaces` the `int_addr` and `int_gateway_ip` lines are dropped and read
from the internal NIC instead; with mode `auto` the interface names go too.
Alternatively write `/etc/vpn-router/vpn-router.json` directly before or after installing.
A configuration-management tool should do that and then restart `vpn-router-setup`.
See [examples/](examples/) for a cloud-init template and a shell installer, and
[examples/azure/](examples/azure/) / [examples/gcp/](examples/gcp/) for full Terraform examples
that deploy a router on Azure or GCP.
## Configuration
`/etc/vpn-router/vpn-router.json` is JSON, mode 0600, and is created on first install. It
is managed with `ucf`, so it is yours to edit: a later `dpkg-reconfigure` applies what
changed and asks before touching anything you altered by hand. A complete example is
installed at `/usr/share/doc/vpn-router/vpn-router.json.example`.
The file is an object of sections, each an object of settings. A section or setting that
is left out takes its default, and an empty string or empty list means unset. Every value
must have the type shown below; an unknown section or setting is an error.
| Section | Setting | Type | Default | Meaning |
|---|---|---|---|---|
| `general` | `platform` | string | `none` | `none`, `generic`, `azure` or `gcp`. `none` defers configuration entirely |
| `general` | `mode` | string | `manual` | How much you supply: `manual`, `interfaces` or `auto` |
| `interfaces` | `external` | string | | Name of the NIC facing the untrusted network |
| `interfaces` | `internal` | string | | Name of the NIC facing the protected network |
| `wan` | `local_fqdn` | string | | This router's name or IP address, used as its IKE identity in `fqdn` mode |
| `wan` | `local_id_mode` | string | `fqdn` | IKE identity source: `fqdn`, `public_ip` or `internal_ip` |
| `local` | `cidrs` | list | | Local subnets advertised into the tunnel |
| `local` | `int_addr` | string | | This host's address on the internal network |
| `local` | `int_gateway_ip` | string | | Next hop on the internal side |
| `remote` | `addrs` | list | | Remote gateway addresses or FQDNs |
| `remote` | `id` | string | | Remote IKE identity, an FQDN or an IPv4 or IPv6 address |
| `remote` | `cidrs` | list | | Remote subnets reachable through the tunnel |
| `remote` | `psk` | string | | Pre-shared key |
| `remote` | `psk_file` | string | | Path to a file holding the key; wins over `psk` |
| `p2s` | `enabled` | boolean | `false` | Road-warrior access |
| `p2s` | `address_pool` | string | | CIDR assigned to road-warrior clients |
| `p2s` | `ca_name` | string | `VPN Router CA` | Common name for a locally created CA |
| `wireguard` | `enabled` | boolean | `false` | WireGuard endpoint |
| `wireguard` | `address` | string | | Address and prefix for `wg0`, for example `192.168.200.1/24` |
| `wireguard` | `listen_port` | integer | `51820` | UDP port |
| `openvpn` | `enabled` | boolean | `false` | OpenVPN access for users |
| `openvpn` | `address_pool` | string | | IPv4 CIDR assigned to OpenVPN clients |
| `openvpn` | `port` | integer | `1194` | Port of the OpenVPN server on the external interface |
| `openvpn` | `protocol` | string | `udp` | `udp` or `tcp` |
| `docker` | `mode` | string | `auto` | `auto` secures Docker when its service is installed, `on` requires it, `off` removes the rules |
| `docker` | `exposed_ports` | list | | TCP ports and port ranges (`8000:8100`) reachable from the external interface; at most 15, a range counting as two |
| `web` | `enabled` | boolean | `true` | The status web service on the internal address |
| `web` | `port` | integer | `8080` | TCP port of the status web service |
| `web` | `jwt_secret` | string | | Secret, at least 32 characters, that signs the tokens for `/config`; without it `/config` is not served |
A minimal site-to-site configuration in `auto` mode:
```json
{
"general": { "platform": "generic", "mode": "auto" },
"wan": { "local_fqdn": "router.example.com" },
"local": { "cidrs": ["10.0.0.0/24"] },
"remote": {
"addrs": ["peer.example.net"],
"id": "peer.example.net",
"cidrs": ["192.168.0.0/24"],
"psk": "s3cr3t"
}
}
```
The router's IKE identity (`wan.local_fqdn` in `fqdn` mode) and the remote one (`remote.id`) are
an FQDN or an IPv4 or IPv6 address. They are written to the strongSwan configuration as given;
strongSwan reads an address as an IP identity and a name as an FQDN, so a leading `@` is not needed
and is refused. Road-warrior and OpenVPN access need a name in `wan.local_fqdn`.
WireGuard peers are added by hand in `/etc/wireguard/wg0.conf`. The package owns the
`[Interface]` section of that file and rewrites it when the settings above change, but it
carries every `[Peer]` section across untouched.
Every feature is independent. Leaving a group empty simply means that feature is not
configured, and its generated files are removed.
### What `mode` decides
`mode` states how much of the network configuration you are supplying, and therefore how
much the package reads from the system:
| `mode` | You provide | The package works out |
|---|---|---|
| `manual` | both interface names, `int_addr`, `int_gateway_ip` | nothing |
| `interfaces` | both interface names | `int_addr` and `int_gateway_ip`, from those interfaces |
| `auto` | nothing | the interface names too: the external one from the routing table, the internal one as the only other NIC with an address |
`manual` is the only mode in which no detection code runs at all. `auto` is the
convenient one and can pick the wrong interface, which is a trade you make deliberately
by choosing it rather than something the package does behind your back.
Detection ignores virtual interfaces, those the kernel lists under `/sys/devices/virtual`:
loopback, bridges such as `docker0`, veth pairs and tunnels. Only the NICs of the machine
count, so Docker does not change the result. If the host has more than two NICs with an
address, `auto` stops and says so; use mode `interfaces` and name them.
The addresses strongSwan binds on always come from the external NIC, in every mode, so
they are never configured twice and cannot drift out of step with the host. If a named
interface does not exist the service stops and says so. In `interfaces` and `auto` mode
the internal gateway is taken from the routing table, falling back to the first host
address of the connected subnet with a warning when the routing table has nothing to say.
## Docker
Docker publishes container ports in its own firewall rules, which bypass UFW and can expose
ports on the external interface. Docker evaluates the `DOCKER-USER` chain before its own
rules, so the package filters there. With `docker.mode` set to `auto`, the package secures
Docker when `docker.service` is installed, and does nothing otherwise.
### On a host that already runs Docker
Installing Docker first is the recommended order. After installing the package:
1. Set `general.platform` in `/etc/vpn-router/vpn-router.json`. While it is `none`, nothing is
configured, the Docker rules included.
2. List the TCP ports the running containers must expose to the internet, for example
`"docker": { "exposed_ports": ["443"] }`. The default is none: once the rules are applied,
every published port that is not listed is unreachable from the external interface, and
remains reachable through the IPsec tunnel.
3. Apply the configuration with `systemctl restart vpn-router-setup`. Docker is running, so
this also restarts `vpn-router-docker`, which installs the rules.
4. Check the result:
```sh
iptables -S DOCKER-USER # "-j VPN-ROUTER-DOCKER" is the first rule
iptables -S VPN-ROUTER-DOCKER # the rules of the external interface
```
### When Docker is installed after the package
The Docker rules are not applied when Docker starts for the first time, so run
`systemctl restart vpn-router-setup` after installing Docker, then check as above. A reboot or a
restart of Docker applies them as well.
The rules sit in their own chain, `VPN-ROUTER-DOCKER`, entered by one jump at the top of
`DOCKER-USER`; other rules in `DOCKER-USER` are left alone. For traffic that arrives on the
external interface, in this order:
1. IPsec protected traffic is accepted.
2. TCP to the ports in `docker.exposed_ports` is accepted; none by default.
3. Packets of established and related connections are accepted.
4. Everything else is dropped.
Traffic from every other interface, including the internal one and the Docker bridges, is
not affected. The rules cover IPv4 only.
`vpn-router-setup` starts after `docker.service`, so Docker is up when it looks for it; the
ordering has no effect on a host without Docker. The `vpn-router-docker` service applies the
rules after `docker.service` starts and removes
them when Docker stops, so restarting Docker restores them. `systemctl restart
vpn-router-setup` applies a changed configuration while Docker runs. With `docker.mode` set to
`on`, the service fails when Docker is not installed or its daemon is not running; with
`off` it removes the rules.
## Applying changes
```sh
systemctl restart vpn-router-setup
```
The service also runs at boot. With platform `none` it does nothing and says so.
Otherwise it regenerates the configuration files, resolves the interfaces and their
addresses as `mode` directs, applies routes, generates WireGuard keys and injects
firewall rules. It provisions the user directory in `slapd` while OpenVPN or road-warrior
access is configured. It then starts `strongswan`, `wg-quick@wg0` and
`openvpn-server@vpn-router` if the corresponding feature is configured, reloading or
restarting them only when their inputs changed, and
restarts `systemd-resolved` only when the road-warrior DNS drop-in changed. Turning a
feature off in the configuration stops its service and removes its files. Every step is
idempotent, so restarting with no changes does nothing.
To regenerate the configuration files and PKI without starting, stopping or reloading
anything:
```sh
/usr/lib/vpn-router/configure
```
This does still write to the system: it renders into `/etc/swanctl/`,
`/etc/systemd/resolved.conf.d/` and `/etc/wireguard/`, and if road-warrior access is
enabled with an empty PKI directory it creates the certificate authority and adds it
to the host trust store. What it leaves alone is routes, firewall rules and services -
that is what `systemctl restart vpn-router-setup` adds.
## Certificates
`/etc/vpn-router/pki` is the PKI directory, laid out the way `simple-ca` expects:
```
ca_cert.pem ca_key.pem ca_bundle.pem <label>_cert.pem <label>_key.pem crl.pem
```
`<label>` is the first component of the router FQDN, so `router.example.com` gives
`router_cert.pem`.
If road-warrior or OpenVPN access is enabled and this directory is empty, a certificate
authority, a server certificate and a CRL valid for ten years are created there
automatically; a CA without a CRL gets one. User certificates are issued by
`vpn-router-user` into `users/`. To use your own instead, place the
files in that layout before enabling the feature; nothing is generated or overwritten when
material is already present.
Either way, the package copies what each consumer needs into place: the CA into
`/etc/swanctl/x509ca/`, the server certificate and key into `/etc/swanctl/x509/` and
`/etc/swanctl/private/`, a CRL into `/etc/swanctl/x509crl/`, and the CA into the system
trust store.
Issue a client certificate with the bundled tool:
```sh
/usr/lib/vpn-router/simple-ca make-cert --ca-dir /etc/vpn-router/pki alice
/usr/lib/vpn-router/simple-ca make-pfx --ca-dir /etc/vpn-router/pki \
--password s3cr3t /etc/vpn-router/pki/alice_cert.pem
```
`simple-ca` comes from a separate project, <https://gitea.koszewscy.waw.pl/slawek/simple-ca>,
and also supports `make-crl` and `revoke-cert`.
## OpenVPN and users
With `openvpn.enabled`, an `address_pool` and `wan.local_fqdn` set, `vpn-router-setup` runs
`openvpn-server@vpn-router` with `/etc/openvpn/server/vpn-router.conf`. The server uses the
router's CA, certificate and CRL (the copies under `/etc/swanctl`), pushes routes to
`local.cidrs` and the router as DNS server, and opens `openvpn.port` on the external interface.
A client needs both a certificate of the router CA and the password of the user the
certificate names:
- the certificate is checked against the CA and the CRL (`crl-verify`, read at every
connection);
- the password is checked by `openvpn-auth-ldap`, which binds to the user directory as the
user;
- `/usr/lib/vpn-router/openvpn-same-user` refuses a login whose user name is not the common
name of the certificate.
### The user directory
While OpenVPN or road-warrior access is configured, `vpn-router-setup` keeps a directory of
users in the local slapd:
- its own database, `dc=vpn-router,dc=local` in `/var/lib/ldap/vpn-router`, next to the one the
slapd package creates; root manages it through the local socket, with no administrator
password;
- `ou=users` holds one `inetOrgPerson` per user, named by `uid`;
- `cn=user-manager,ou=services` is the account that changes users; it may write `ou=users` and
set passwords, never read them, and its password is in
`/etc/vpn-router/ldap/user-manager.secret` (root only);
- passwords are stored as ARGON2 hashes, and a disabled user is locked with the password policy
(`pwdAccountLockedTime`);
- slapd listens on its local socket only (`SLAPD_SERVICES` in `/etc/default/slapd`).
The directory is kept when OpenVPN and road-warrior access are switched off, and on purge.
### Managing users
`vpn-router-user` manages the users. A user is the directory entry and a client certificate
with the user name as its common name, in `/etc/vpn-router/pki/users`. Passwords are read from
the first line of standard input. The answer is JSON.
```sh
echo 's3cr3t-pass' | vpn-router-user add alice
vpn-router-user list
vpn-router-user profile alice
echo 'n3w-pass' | vpn-router-user passwd alice
vpn-router-user disable alice
vpn-router-user enable alice
vpn-router-user remove alice
```
| Command | Effect |
| --- | --- |
| `list` | The users, whether each is enabled and has a certificate |
| `add <name>` | Adds the user and issues the certificate |
| `passwd <name>` | Sets the password |
| `disable <name>`, `enable <name>` | Locks the user out of OpenVPN, and lets the user in again |
| `remove <name>` | Revokes the certificate, renews the CRL and removes the user |
| `profile <name>` | The OpenVPN client profile, with the CA, the certificate and the key inline |
A user name is 1 to 32 lowercase letters, digits and hyphens, starting with a letter; a
password has 8 to 128 characters. The exit status is 0 on success, 2 for an invalid request,
3 when the user does not exist, 4 when the user exists and 1 for any other failure.
The certificate is also accepted by road-warrior (IKEv2 EAP-TLS) access, which checks
certificates only, so `disable` does not affect it; `remove` revokes the certificate for both.
## Platform modules
Platform-specific values do not belong in the core, so they are Python modules in
`/usr/lib/vpn-router/vpnrouter_platforms/`, imported at run time and selected by
`general.platform`. The platform is never probed: the caller already knows where it is
deploying.
`none` is not a module. It is the shipped default and means "configure nothing at all",
which is how a deferred install is expressed. `generic` is the module to choose when the
router should be configured but the platform adds nothing.
The shipped modules currently add nothing. Road-warrior clients are given the router
itself as their DNS server and the router forwards through its own `systemd-resolved`, so
no provider resolver has to be advertised or routed through the tunnel on any platform.
The modules exist as the documented place for platform specifics if that changes.
A module may define any of the following, and takes the default for anything it omits:
| Name | Type | Effect |
|---|---|---|
| `EXTRA_P2S_DNS` | `str` | added to the road-warrior pool `dns` list |
| `EXTRA_LOCAL_TS` | `str` | added to `local_ts` on both connections |
| `apply(ctx)` | function | called once the system is configured, to apply platform-specific state; must be idempotent |
`ctx` is the template context, so a module reads what it needs from it -
`ctx['wan_iface']`, `ctx['int_iface']`, `ctx['local_cidrs']` and the rest.
Adding a platform is adding one file. Modules are discovered by listing the package, so
nothing in the core needs editing:
```python
# /usr/lib/vpn-router/vpnrouter_platforms/example.py
"""Example provider."""
EXTRA_P2S_DNS = '192.0.2.53'
EXTRA_LOCAL_TS = '192.0.2.53/32'
```
Set `general.platform` to `example` and restart the service. The debconf question offers only the
modules shipped with the package; the configuration file accepts any module installed.
## Status web service
The `vpn-router-web` service shows the state of the router on its internal address and
`web.port` (8080 by default). It uses only the Python standard library, serves HTTP without
authentication, and shows no configuration: no key, identity, address or subnet.
| Page | Content |
| --- | --- |
| `/` | The status for people: plain text for `curl` and `wget`, a page styled with Tailwind CSS for a browser, in the browser's light or dark scheme, that reloads itself every minute (the style sheet is loaded from a CDN, so the browser needs internet access; the text is readable without it) |
| `/status` | The same status as JSON |
| `/config` | The configuration with its secrets masked, and what the router resolved and runs, as JSON; needs a token |
| `/users` | The users of OpenVPN and road-warrior access; needs a token |
The status holds the state of `vpn-router-setup`, of `strongswan` when site-to-site or road-warrior
access is configured, of `wg-quick@wg0` when WireGuard is, of `openvpn-server@vpn-router` when
OpenVPN is, and of `slapd` when either OpenVPN or road-warrior access is; the number of connected
OpenVPN clients, from the server's status file, which OpenVPN rewrites every minute; the package version; whether
Docker is installed, and when it is, the state of `docker.service` and of `vpn-router-docker`
(the Docker rules), which do not change `status`; whether the
site-to-site tunnel is established; the number of connected road-warrior (IPsec P2S) clients; and
the WireGuard peers with the age of their last handshake and their traffic. `status` is `ok` when
all these services are active, and `degraded` otherwise.
```bash
curl http://10.0.0.4:8080/
curl http://10.0.0.4:8080/status
```
### The /config page
`/config` shows what `/` and `/status` leave out, for a client such as manage-ui:
| Key | Content |
| --- | --- |
| `configuration` | `vpn-router.json` with the defaults applied and the secrets masked (`psk`, `jwt_secret`, and any option named like a password, token or key) |
| `network` | The external and internal interface, their addresses, the internal gateway, the local and remote IKE identities and subnets, and the IPv4 routes as `ip -j -4 route` prints them |
| `ipsec` | `connections` and `sas`: one object per connection and per IKE SA, read from `swanctl --list-conns --pretty` and `swanctl --list-sas --pretty`, with the connection's name under `name` and swanctl's keys and string values |
| `wireguard` | The public key and port of `wg0`, and each peer's public key, endpoint, allowed IPs, handshake age and traffic; no private or preshared key |
| `docker` | `docker.exposed_ports` and the rules of the `VPN-ROUTER-DOCKER` chain |
The page needs a JWT in an `Authorization: Bearer` header, signed with HS256 by `web.jwt_secret`.
The token must have an `exp` claim, not more than ten minutes from now (a minute of clock
difference is allowed); `nbf` is checked when present. Other algorithms are refused, with 401 and
`WWW-Authenticate: Bearer`. While `web.jwt_secret` is not set the page is not served (404). The
token and the page travel over plain HTTP on the internal network, and the token can be used
until it expires, so keep the lifetime short. A client makes a token from the secret with only
HMAC-SHA256; with Python:
```python
import base64, hashlib, hmac, json, time
def b64(data):
return base64.urlsafe_b64encode(data).rstrip(b'=').decode()
secret = 'the web.jwt_secret value'
head = b64(json.dumps({'alg': 'HS256', 'typ': 'JWT'}).encode())
body = b64(json.dumps({'exp': int(time.time()) + 300}).encode())
sign = b64(hmac.new(secret.encode(), f'{head}.{body}'.encode(), hashlib.sha256).digest())
print(f'{head}.{body}.{sign}')
```
```bash
curl -H "Authorization: Bearer $TOKEN" http://10.0.0.4:8080/config
```
### The /users API
`/users` manages users with the same token as `/config`, by running `vpn-router-user`. Request
and answer bodies are JSON; a request body is at most 4096 bytes.
| Request | Body | Answer |
| --- | --- | --- |
| `GET /users` | | `{"users": [{"name", "enabled", "certificate"}, ...]}` |
| `POST /users` | `{"name": ..., "password": ...}` | 201 and the user |
| `PUT /users/<name>/password` | `{"password": ...}` | the user |
| `POST /users/<name>/disable` | | the user |
| `POST /users/<name>/enable` | | the user |
| `DELETE /users/<name>` | | `{"name", "removed": true}` |
| `GET /users/<name>/profile` | | `{"name", "profile"}`, the OpenVPN client profile |
A failure is `{"error": ...}` with 400 for an invalid request, 404 for a user that does not
exist, 409 for a user that exists and 500 otherwise. While neither OpenVPN nor road-warrior
access is configured, `/users` answers 404. Passwords and private keys cross the internal
network in plain HTTP, like the token.
```bash
curl -H "Authorization: Bearer $TOKEN" -H 'Content-Type: application/json' \
-d '{"name": "alice", "password": "s3cr3t-pass"}' http://10.0.0.4:8080/users
```
The service starts on its own, without waiting for `vpn-router-setup`, and keeps running. Every
30 seconds it works out the internal address from `vpn-router.json` as `general.mode` directs,
and reads `web.port` and `web.enabled`; it listens there, moves when the address or the port
changes, and stops listening while `web.enabled` is `false` or the address is not known. Each
request reads the token secret from `vpn-router.json`, and the features and the facts for
`/config` from `/run/vpn-router/web.json` (readable by root only), which `vpn-router-setup` writes;
before that file exists the page shows no features. `vpn-router-setup` opens `web.port` on the
internal interface in the firewall.
## Removal
`apt-get remove` stops `vpn-router-docker`, which removes the Docker rules. It terminates the
tunnels and takes `wg-quick@wg0` and `openvpn-server@vpn-router` down, but leaves the
configuration in place. `strongswan` is left enabled, since it is a shared service that
may be in use by something else.
`apt-get purge` additionally removes the generated files and the firewall rules, restoring
`/etc/ufw/before.rules` to its original content, and hands `vpn-router.json` back to `ucf`
before deleting it so a later reinstall starts clean. Key material and users are never
deleted: `/etc/vpn-router/pki`, `/etc/vpn-router/ldap`, the user directory in slapd, the
WireGuard key pair and everything under `/etc/swanctl` are left alone.
## Building
```sh
cd debian-package
./build.sh
```
The build runs in a container and writes the package to `debian-package/out/`.
`./publish.sh` uploads it to the configured repository.