Nemo's Home

Migrating DNSCrypt Server to Docker ⚓

18 May 2019

I’ve been running a personal DNSCrypt server in Bangalore for the last 2 years. When I set it up, it was just a compiled version of dnscrypt-wrapper, which was the bare minimum setup I could do.

Since then, I’ve upgraded it to a distribution supported version, but recent changes in dnscrypt key rotation, I’ve been wanting to setup something automated as well.

The easiest way was to switch to the official DNSCrypt Docker image, which does both key generation and certificate rotation. Since my public key was already present in the DNSCrypt Server lists, I was not too keen to regenerate a new key.

The primary challenge was ensuring that the docker container picks up my existing keys without trying to generate new ones from scratch. It was basically 2 steps:

  1. Match the directory structure that the container expects.
  2. Invoke the container directly into start mode while passing existing keys.

Directory Structure

I copied my keys (public.key, secret.key) to /etc/dnscrypt-keys and ran the following:

echo 2.dnscrypt-cert.captnemo.in > provider_name
touch provider_info.txt # I couldn't figure out how to output the same info, so kept it blank
hexdump -ve '1/1 "%.2x"' < public.key > public.key.txt

Then I ensured that the file permissions are matching what the container expects:

chmod 640 secret.key
chmod 644 public.key
chown root:1002 public.key secret.key
chmod 644 provider_name

This is how the final permissions looked for the directory (/etc/dnscrypt-keys)

-rw-r-----   1 root 1002    64 May 18 07:15 secret.key
-rw-r--r--   1 root 1002    32 May 18 07:15 public.key
-rw-r--r--   1 root root    28 May 18 07:19 provider_name
-rw-r--r--   1 root root     0 May 18 07:23 provider_info.txt
-rw-r--r--   1 root root    64 May 18 07:25 public.key.txt
drwxr-xr-x   2 root root  4096 May 18 07:26 .

Running the Container

Then, I directly ran dnscrypt-wrapper container:

docker run --detatched --restart=unless-stopped --volume /etc/dnscrypt-keys:/opt/dnscrypt-wrapper/etc/keys --publish 10.47.0.5:4434:443/tcp --publish 10.47.0.5:4434:443/udp jedisct1/dnscrypt-server start

I pass a host path mount instead of creating a Docker Volume, since they can get deleted in regular docker prune.

Here, 10.47.0.5 is the “Anchor IP”, which Digital Ocean internally maps to my Floating IP.

The container comes up, generates new short-term keys and goes live:

Starting DNSCrypt service for provider:
2.dnscrypt-cert.captnemo.in
Starting pre-service scripts in /etc/runit_init.d
setup in directory /opt/unbound/etc/unbound
generating unbound_server.key
Generating RSA private key, 3072 bit long modulus (2 primes)
.......++++
...................++++
e is 65537 (0x010001)
generating unbound_control.key
Generating RSA private key, 3072 bit long modulus (2 primes)
.........................++++
........................................++++
e is 65537 (0x010001)
create unbound_server.pem (self signed certificate)
create unbound_control.pem (signed client certificate)
Signature ok
subject=CN = unbound-control
Getting CA Private Key
Setup success. Certificates created. Enable in unbound.conf file to use
ok: run: unbound: (pid 28) 300s
ok: run: dnscrypt-wrapper: (pid 31) 300s
ok: run: unbound: (pid 28) 600s
ok: run: dnscrypt-wrapper: (pid 31) 600s

Once the server was up, I verified connectivity with dnscrypt-proxy and it worked perfectly.

Future Scope

Right now, I have a single container that does 2 things:

  1. Certificate Rotation via a service that checks it every 30 minutes.
  2. DNSCrypt Service, which is accessible over the internet.

For (1) to work, it needs access to the Private Keys that are used to sign the temporary certificates that last 24 hours. Since both things are managed within the same container, the container ends up with both network and long-term keys access. This means, any RCE on the service can result in the long-term keys being compromised.

A simple fix for this would be to separate out the Certificate Rotation part into a separate “mode” on the Docker Image, which can be called independently. This would allow someone to run certificate rotation on a second container using a scheduler, but with far more limitations (such as no network access). A common file-mount between both the containers can take care of sharing the temporary keys between the containers, and a simple unix socket on the shared-file-mount can be used to signal a certificate rotation (this triggers the dnscrypt service restart, so it picks the new cert).

Published on May 18, 2019
By