live and let live.与民争利的政府必亡
Friday, 24 June 2016
build an OpenVPN image for Docker.
OpenVPN for Docker
OpenVPN server in a Docker container complete with an EasyRSA PKI CA.
Retrieve the client configuration with embedded certificates
docker run --volumes-from $OVPN_DATA --rm kylemanna/openvpn ovpn_getclient CLIENTNAME > CLIENTNAME.ovpn
Create an environment variable with the name DEBUG and value of 1 to enable debug output (using "docker -e").
docker run --volumes-from $OVPN_DATA -p 1194:1194/udp --privileged -e DEBUG=1 kylemanna/openvpn
Test using a client that has openvpn installed correctly
$ openvpn --config CLIENTNAME.ovpn
Run through a barrage of debugging checks on the client if things don't just work
$ ping 188.8.131.52 # checks connectivity without touching name resolution
$ dig google.com # won't use the search directives in resolv.conf
$ nslookup google.com # will use search
How Does It Work?
Initialize the volume container using the kylemanna/openvpn image with the included scripts to automatically generate:
a private key
a self-certificate matching the private key for the OpenVPN server
an EasyRSA CA key and certificate
a TLS auth key from HMAC security
The OpenVPN server is started with the default run cmd of ovpn_run
The configuration is located in /etc/openvpn, and the Dockerfile declares that directory as a volume. It means that you can start another container with the --volumes-from flag, and access the configuration. The volume also holds the PKI keys and certs so that it could be backed up.
To generate a client certificate, kylemanna/openvpn uses EasyRSA via the easyrsa command in the container's path. TheEASYRSA_* environmental variables place the PKI CA under /etc/openvpn/pki.
Conveniently, kylemanna/openvpn comes with a script called ovpn_getclient, which dumps an inline OpenVPN client configuration file. This single file can then be given to a client for access to the VPN.
To enable Two Factor Authentication for clients (a.k.a. OTP) see this document.
We use tun mode, because it works on the widest range of devices. tap mode, for instance, does not work on Android, except if the device is rooted.
The topology used is net30, because it works on the widest range of OS. p2p, for instance, does not work on Windows.
The UDP server uses192.168.255.0/24 for dynamic clients by default.
The client profile specifies redirect-gateway def1, meaning that after establishing the VPN connection, all traffic will go through the VPN. This might cause problems if you use local DNS recursors which are not directly reachable, since you will try to reach them through the VPN and they might not answer to you. If that happens, use public DNS resolvers like those of Google (184.108.40.206 and 220.127.116.11) or OpenDNS (18.104.22.168 and 22.214.171.124).
The Docker container runs its own EasyRSA PKI Certificate Authority. This was chosen as a good way to compromise on security and convenience. The container runs under the assumption that the OpenVPN container is running on a secure host, that is to say that an adversary does not have access to the PKI files under /etc/openvpn/pki. This is a fairly reasonable compromise because if an adversary had access to these files, the adversary could manipulate the function of the OpenVPN server itself (sniff packets, create a new PKI CA, MITM packets, etc).
The certificate authority key is kept in the container by default for simplicity. It's highly recommended to secure the CA key with some passphrase to protect against a filesystem compromise. A more secure system would put the EasyRSA PKI CA on an offline system (can use the same Docker image and the script ovpn_copy_server_files to accomplish this).
It would be impossible for an adversary to sign bad or forged certificates without first cracking the key's passphase should the adversary have root access to the filesystem.
The EasyRSA build-client-full command will generate and leave keys on the server, again possible to compromise and steal the keys. The keys generated need to be signed by the CA which the user hopefully configured with a passphrase as described above.
Assuming the rest of the Docker container's filesystem is secure, TLS + PKI security should prevent any malicious host from using the VPN.
Benefits of Running Inside a Docker Container
The Entire Daemon and Dependencies are in the Docker Image
This means that it will function correctly (after Docker itself is setup) on all distributions Linux distributions such as: Ubuntu, Arch, Debian, Fedora, etc. Furthermore, an old stable server can run a bleeding edge OpenVPN server without having to install/muck with library dependencies (i.e. run latest OpenVPN with latest OpenSSL on Ubuntu 12.04 LTS).
It Doesn't Stomp All Over the Server's Filesystem
Everything for the Docker container is contained in two images: the ephemeral run time image (kylemanna/openvpn) and the data image (using busybox as a base). To remove it, remove the two Docker images and corresponding containers and it's all gone. This also makes it easier to run multiple servers since each lives in the bubble of the container (of course multiple IPs or separate ports are needed to communicate with the world).
Some (arguable) Security Benefits
At the simplest level compromising the container may prevent additional compromise of the server. There are many arguments surrounding this, but the take away is that it certainly makes it more difficult to break out of the container. People are actively working on Linux containers to make this more of a guarantee in the future.
Differences from jpetazzo/dockvpn
No longer uses serveconfig to distribute the configuration via https
Proper PKI support integrated into image
OpenVPN config files, PKI keys and certs are stored on a storage volume for re-use across containers
Addition of tls-auth for HMAC security
server a Digital Ocean Droplet with 512 MB RAM running Ubuntu 14.04
Android App OpenVPN Connect 1.1.14 (built 56)
OpenVPN core 3.0 android armv7a thumb2 32-bit
OS X Mavericks with Tunnelblick 3.4beta26 (build 3828) using openvpn-2.3.4