wesher creates and manages a mesh overlay network across a group of nodes, using wireguard.
Its main use-case is adding low-maintenance security to public-cloud networks or connecting different cloud providers.
⚠ WARNING: since mesh membership is controlled by a mesh-wide pre-shared key, this effectively downgrades some of the security benefits from wireguard. See security considerations below for more details.
Before starting, make sure wireguard is installed on all nodes.
The following ports must be accessible between all nodes (see configuration options to change these):
- 51820 UDP
- 7946 UDP and TCP
wesher on all nodes with go >= 1.11:
$ GO111MODULE=on go get github.com/costela/wesher
On the first node (assuming
$GOPATH/bin is in the
Running the command above on a terminal will currently output a generated cluster key, like:
new cluster key generated: XXXXX
Then, on any further node:
# wesher --cluster-key XXXXX --join x.x.x.x
XXXXX is the base64 encoded 256 bit key printed by the step above and
x.x.x.x is the hostname or IP of any of
the nodes already joined to the mesh cluster.
wireguard - and therefore
wesher - need root access.
wesher tool builds a cluster and manages the configuration of wireguard on each node to create peer-to-peer
connections between all nodes, thus forming a full mesh VPN.
This approach may not scale for hundreds of nodes (benchmarks accepted
Automatic Key management
The wireguard private keys are created on startup for each node and the respective public keys are then broadcast across the cluster.
The control-plane cluster communication is secured with a pre-shared AES-256 key. This key can be be automatically created during startup of the first node in a cluster, or it can be provided (see configuration). The cluster key must then be sent to other nodes via a out-of-band secure channel (e.g. ssh, cloud-init, etc). Once set, the cluster key is saved locally and reused on the next startup.
Automatic IP address management
The overlay IP address of each node is selected out of a private network (
10.0.0.0/8 by default) and is consistently
hashed based on the hostname, meaning a host will always receive the same overlay IP address (see limitations
of this approach below). The hostname is also used by the underlying cluster management (using memberlist)
to identify nodes and must therefore be unique in the cluster.
To ease intra-node communication,
wesher also adds entries to
/etc/hosts for each other node. See configuration
below for how to disable this behavior.
If a node in the cluster is restarted, it will attempt to re-join the last-known nodes using the same cluster key. This means a restart requires no manual intervention.
All options can be passed either as command-line flags or environment variables:
||WESHER_CLUSTER_KEY||shared key for cluster membership; must be 32 bytes base64 encoded; will be generated if not provided||autogenerated/loaded|
||WESHER_JOIN||comma separated list of hostnames or IP addresses to existing cluster members; if not provided, will attempt resuming any known state or otherwise wait for further members|
||WESHER_INIT||whether to explicitly (re)initialize the cluster; any known state from previous runs will be forgotten||
||WESHER_BIND_ADDR||IP address to bind to for cluster membership||autodetected|
||WESHER_CLUSTER_PORT||port used for membership gossip traffic (both TCP and UDP); must be the same across cluster||
||WESHER_WIREGUARD_PORT||port used for wireguard traffic (UDP); must be the same across cluster||
||WESHER_OVERLAY_NET||the network in which to allocate addresses for the overlay mesh network (CIDR format); smaller networks increase the chance of IP collision||
||WESHER_INTERFACE||name of the wireguard interface to create and manage||
||WESHER_NO_ETC_HOSTS||whether to skip writing hosts entries for each node in mesh||
||WESHER_LOG_LEVEL||set the verbosity (one of debug/info/warn/error)||
The decision of whom to allow in the mesh is made by memberlist and is secured by a cluster-wide pre-shared key. Compromise of this key will allow an attacker to:
- access services exposed on the overlay network
- impersonate and/or disrupt traffic to/from other nodes It will not, however, allow the attacker access to decrypt the traffic between other nodes.
This pre-shared key is currently static, set up during cluster bootstrapping, but will - in a future version - be rotated for improved security.
Current known limitations
Overlay IP collisions
Since the assignment of IPs on the overlay network is currently decided by the individual node and implemented as a naive hashing of the hostname, there can be no guarantee two hosts will not generate the same overlay IPs. This limitation may be worked around in a future version.
Once a cluster is joined, there is currently no way to distinguish a failed node from an intentionally removed one. This is partially by design: growing and shrinking your cluster dynamically (e.g. via autoscaling) should be as easy as possible.
However, this does mean longer connection loss between any two parts of the cluster (e.g. across a WAN link between different cloud providers) can lead to a split-brain scenario where each side thinks the other side is simply "gone".
There is currently no clean solution for this problem, but one could work around it by designating edge nodes which
wesher with the
--join option pointing to the other side.
Future versions might include the notion of a "static" node to more cleanly avoid this.
Broken connections on join/leave
wesher uses wireguard's management commands (
wg-quick) to talk to wireguard. The current
implementation is very naive and recreates the wireguard interface for every configuration change (e.g. nodes' joining
and leavning). Since wireguard's underlying traffic is essentially stateless, this does not directly impact TCP
connections on the overlay network. The connection should treat it as normal packet drops.
However, if your application is sensitive to ICMP errors (e.g. "no route to host"), or if you are using UDP, you might experience connection resets and message loss, respectively.
This behavior will be improved in future versions.