<i>All credit goes to the WireGuard project, [zx2c4](https://www.zx2c4.com/) and the [open source contributors](https://github.com/WireGuard/WireGuard/graphs/contributors) for the original software,<br/> this is my solo unofficial attempt at providing more comprehensive documentation, API references, and examples.</i>
[WireGuard](https://www.wireguard.com/) is an open-source VPN solution written in C by [Jason Donenfeld](https://www.jasondonenfeld.com) and [others](https://github.com/WireGuard/WireGuard/graphs/contributors), aiming to fix many of the problems that have plagued other modern server-to-server VPN offerings like IPSec/IKEv2, OpenVPN, or L2TP. It shares some similarities with other modern VPN offerings like [Tinc](https://www.tinc-vpn.org/) and [MeshBird](https://github.com/meshbird/meshbird), namely good cipher suites and minimal config. As of 2020-01 [it's been merged into the 5.6 version of the Linux kernel](https://arstechnica.com/gadgets/2020/01/linus-torvalds-pulled-wireguard-vpn-into-the-5-6-kernel-source-tree/), meaning it will ship with most Linux systems out-of-the-box.
Whether living behind the Great Wall of China or just trying to form a network between your servers, WireGuard is a great option and serves as a "lego block" for building networks (much in the same way that ZFS is a lego block for building filesystems).
But you can write your own solutions for these problems using WireGuard under the hood (like [Tailscale](https://github.com/tailscale/tailscale) or [AltheaNet](https://althea.net/)).
- [IPSec (IKEv2)](https://github.com/jawj/IKEv2-setup)/strongSwan: in my experience, there was lots of brittle config that was different for each OS, the NAT busting setup is very manual and involves updating the central server and starting all the others in the correct order, it wasn't great at becoming stable again after network downtime, had to be manually restarted often. your mileage may vary.
- [OpenVPN](https://openvpn.net/vpn-server-resources/site-to-site-routing-explained-in-detail/): can work over UDP or be disguised as HTTPS traffic over TCP
- StealthVPN: haven't tried it, should I?
- [DsVPN](https://github.com/jedisct1/dsvpn): I think it does TCP-over-TCP which usually doesn't end well...
- [SoftEther](https://www.softether.org/) ([SSTP](https://en.wikipedia.org/wiki/Secure_Socket_Tunneling_Protocol)): haven't tried it yet, should I? (also does TCP-over-TCP?)
- IP addresses & ranges: `192.0.2.1/24`, `192.0.2.3`, `192.0.2.3/32`, `2001:DB8::/64` can be replaced with your preferred subnets and addresses (e.g. `192.168.5.1/24`)
A host that connects to the VPN and registers a VPN subnet address such as `192.0.2.3` for itself. It can also optionally route traffic for more than its own address(es) by specifying subnet ranges in comma-separated CIDR notation.
A publicly reachable peer/node that serves as a fallback to relay traffic for other VPN peers behind NATs. A bounce server is not a special type of server, it's a normal peer just like all the others, the only difference is that it has a public IP and has kernel-level IP forwarding turned on which allows it to bounce traffic back down the VPN to other clients.
A group of IPs separate from the public internet, e.g. 192.0.2.1-255 or 192.168.1.1/24. Generally behind a NAT provided by a router, e.g. in office internet LAN or a home Wi-Fi network.
A way of defining a subnet and its size with a "mask", a smaller mask = more address bits usable by the subnet & more IPs in the range. Most common ones:
To people just getting started `192.0.2.1/32` may seem like a weird and confusing way to refer to a single IP. This design is nice though because it allows peers to expose multiple IPs if needed without needing multiple notations. Just know that anywhere you see something like `192.0.2.3/32`, it really just means `192.0.2.3`.
A subnet with private IPs provided by a router standing in front of them doing Network Address Translation, individual nodes are not publicly accessible from the internet, instead the router keeps track of outgoing connections and forwards responses to the correct internal IP (e.g. standard office networks, home Wi-Fi networks, free public Wi-Fi networks, etc)
The publicly accessible address:port for a node, e.g. `123.124.125.126:1234` or `some.domain.tld:1234` (must be accessible via the public internet, generally can't be a private IP like `192.0.2.1` or `192.168.1.1` unless it's directly accessible using that address by other peers on the same subnet).
Domain Name Server, used to resolve hostnames to IPs for VPN clients, instead of allowing DNS requests to leak outside the VPN and reveal traffic. Leaks are testable with http://dnsleak.com.
Public relays are just normal VPN peers that are able to act as an intermediate relay server between any VPN clients behind NATs, they can forward any VPN subnet traffic they receive to the correct peer at the system level (WireGuard doesn't care how this happens, it's handled by the kernel `net.ipv4.ip_forward = 1` and the iptables routing rules).
If all peers are publicly accessible, you don't have to worry about special treatment to make one of them a relay server, it's only needed if you have any peers connecting from behind a NAT.
Each client only needs to define the publicly accessible servers/peers in its config, any traffic bound to other peers behind NATs will go to the catchall VPN subnet (e.g. `192.0.2.1/24`) in the public relays `AllowedIPs` route and will be forwarded accordingly once it hits the relay server.
In summary: only direct connections between clients should be configured, any connections that need to be bounced should not be defined as peers, as they should head to the bounce server first and be routed from there back down the vpn to the correct client.
### How WireGuard Routes Packets
More complex topologies are definitely achievable, but these are the basic routing methods used in typical WireGuard setups:
In the simplest case, the nodes will either be on the same LAN or both be publicly accessible. Define directly accessible nodes with hardcoded `Endpoint` addresses and ports so that WireGuard can connect straight to the open port and route UDP packets without intermediate hops.
When 1 of the 2 parties is behind remote NAT (e.g. when a laptop behind NAT connects to `public-server2`), define the publicly accessible node with a hardcoded `Endpoint` and the NAT-ed node without. The connection will be opened from NAT client -> public client, then traffic will route directly between them in both directions as long as the connection is kept alive by outgoing `PersistentKeepalive` pings from the NAT-ed client.
Most of the time when both parties are behind NATs, the NATs do source port randomization making direct connections infeasible, so they will both have to open a connection to `public-server1`, and traffic will forward through the intermediary bounce server as long as the connections are kept alive.
While sometimes possible, it's generally infeasible to do direct NAT-to-NAT connections on modern networks, because most NAT routers are quite strict about randomizing the source port, making it impossible to coordinate an open port for both sides ahead of time. Instead, a signaling server (STUN) must be used that stands in the middle and communicates which random source ports are assigned to the other side. Both clients make an initial connection to the public signaling server, then it records the random source ports and sends them back to the clients. This is how WebRTC works in modern P2P web apps. Even with a signalling server and known source ports for both ends, sometimes direct connections are not possible because the NAT routers are strict about only accepting traffic from the original destination address (the signalling server), and will require a new random source port to be opened to accept traffic from other IPs (e.g. the other client attempting to use the originally communicated source port). This is especially true for "carrier-grade NATs" like cellular networks and some enterprise networks, which are designed specifically to prevent this sort of hole-punching connection. See the full section below on [**NAT to NAT Connections**](#NAT-to-NAT-Connections) for more information.
More specific (also usually more direct) routes provided by other peers will take precedence when available, otherwise traffic will fall back to the least specific route and use the `192.0.2.1/24` catchall to forward traffic to the bounce server, where it will in turn be routed by the relay server's system routing table (`net.ipv4.ip_forward = 1`) back down the VPN to the specific peer that's accepting routes for that traffic. WireGuard does not automatically find the fastest route or attempt to form direct connections between peers if not already defined, it just goes from the most specific route in `[Peers]` to least specific.
You can figure out which routing method WireGuard is using for a given address by measuring the ping times to figure out the unique length of each hop, and by inspecting the output of:
WireGuard uses encrypted UDP packets for all traffic, it does not provide guarantees around packet delivery or ordering, as that is handled by TCP connections within the encrypted tunnel.
WireGuard claims faster performance than most other competing VPN solutions, though the exact numbers are sometimes debated and may depend on whether hardware-level acceleration is available for certain cryptographic ciphers.
WireGuard's performance gains are achieved by handling routing at the kernel level, and by using modern cipher suites running on all cores to encrypt traffic. WireGuard also gains a significant advantage by using UDP with no delivery/ordering guarantees (compared to VPNs that run over TCP or implement their own guaranteed delivery mechanisms).
WireGuard uses the following protocols and primitives to secure traffic:
- ChaCha20 for symmetric encryption, authenticated with Poly1305, using RFC7539’s AEAD construction
- Curve25519 for ECDH
- BLAKE2s for hashing and keyed hashing, described in RFC7693
- SipHash24 for hashtable keys
- HKDF for key derivation, as described in RFC5869
> WireGuard's cryptography is essentially an instantiation of Trevor Perrin's Noise framework. It's modern and, again, simple. Every other VPN option is a mess of negotiation and handshaking and complicated state machines. WireGuard is like the Signal/Axolotl of VPNs, except it's much simpler and easier to reason about (cryptographically, in this case) than double ratchet messaging protocols.
> It is basically the qmail of VPN software.
> And it's ~4000 lines of code. It is plural orders of magnitude smaller than its competitors.
Authentication in both directions is achieved with a simple public/private key pair for each peer. Each peer generates these keys during the setup phase, and shares only the public key with other peers.
You can also read in keys from a file or via command if you don't want to hardcode them in `wg0.conf`, this makes managing keys via 3rd party service much easier:
```ini
[Interface]
...
PostUp = wg set %i private-key /etc/wireguard/wg0.key <(cat /some/path/%i/privkey)
Technically, multiple servers can share the same private key as long as clients arent connected to two servers with the same key simulatenously.
An example of a scenario where this is a reasonable setup is if you're using round-robin DNS to load-balance connections between two servers that are pretending to be a single server.
Most of the time however, every peer should have its own public/private keypair so that peers can't read eachothers traffic and can be individually revoked.
-`[Interface]` Make sure to specify a CIDR range for the entire VPN subnet when defining the address the server accepts routes for `Address = 192.0.2.1/24`
-`[Interface]` Make sure to specify only a single IP for client peers that don't relay traffic `Address = 192.0.2.3/32`.
-`[Peer]` Create a peer section for each public peer not behind a NAT, make sure to specify a CIDR range for the entire VPN subnet when defining the remote peer acting as the bounce server `AllowedIPs = 192.0.2.1/24`. Make sure to specify individual IPs for remote peers that don't relay traffic and only act as simple clients `AllowedIPs = 192.0.2.3/32`.
7. Traffic is routed from peer to peer using most specific route first over the WireGuard interface, e.g. `ping 192.0.2.3` checks for a direct route to a peer with `AllowedIPs = 192.0.2.3/32` first, then falls back to a relay server that's accepting IPs in the whole subnet
WireGuard config is in [INI syntax](https://en.wikipedia.org/wiki/INI_file), defined in a file usually called `wg0.conf`. It can be placed anywhere on the system, but is often placed in `/etc/wireguard/wg0.conf`.
The config file name must be in the format `${name of the new WireGuard interface}.conf`. WireGuard interface names are typically prefixed with `wg` and numbered starting at `0`, but you can use any name that matches the regex `^[a-zA-Z0-9_=+.-]{1,15}$`.
Config files can opt to use the limited set of `wg` config options, or the more extended `wg-quick` options, depending on what command is preferred to start WireGuard. These docs recommend sticking to `wg-quick` as it provides a more powerful and user-friendly config experience.
This is just a standard comment in INI syntax used to help keep track of which config section belongs to which node, it's completely ignored by WireGuard and has no effect on VPN behavior.
Defines what address range the local node should route traffic for. Depending on whether the node is a simple client joining the VPN subnet, or a bounce server that's relaying traffic between multiple clients, this can be set to a single IP of the node itself (specified with CIDR notation), e.g. 192.0.2.3/32), or a range of IPv4/IPv6 subnets that the node can route traffic for.
When the node is acting as the public bounce server, it should set this to be the entire subnet that it can route traffic, not just a single IP for itself.
When the node is acting as a public bounce server, it should hardcode a port to listen for incoming VPN connections from the public internet. Clients not acting as relays should not set this value.
The DNS server(s) to announce to VPN clients via DHCP, most clients will use this server for DNS requests over the VPN, but clients can also override this value locally on their nodes
Optionally defines which routing table to use for the WireGuard routes, not necessary to configure for most setups.
There are two special values: ‘off’ disables the creation of routes altogether, and ‘auto’ (the default) adds routes to the default table and enables special handling of default routes.
Optionally defines the maximum transmission unit (MTU, aka packet/frame size) to use when connecting to the peer, not necessary to configure for most setups.
Defines the VPN settings for a remote peer capable of routing traffic for one or more addresses (itself and/or other peers). Peers can be either a public bounce server that relays traffic to other peers, or a directly accessible client via LAN/internet that is not behind a NAT and only routes traffic for itself.
All clients must be defined as peers on the public bounce server. Simple clients that only route traffic for themselves, only need to define peers for the public relay, and any other nodes directly accessible. Nodes that are behind separate NATs should _not_ be defined as peers outside of the public server config, as no direct route is available between separate NATs. Instead, nodes behind NATs should only define the public relay servers and other public clients as their peers, and should specify `AllowedIPs = 192.0.2.1/24` on the public server that accept routes and bounce traffic for the VPN subnet to the remote NAT-ed peers.
In summary, all nodes must be defined on the main bounce server. On client servers, only peers that are directly accessible from a node should be defined as peers of that node, any peers that must be relayed by a bounce server should be left out and will be handled by the relay server's catchall route.
In the configuration outlined in the docs below, a single server `public-server1` acts as the relay bounce server for a mix of publicly accessible and NAT-ed clients, and peers are configured on each node accordingly:
This is just a standard comment in INI syntax used to help keep track of which config section belongs to which node, it's completely ignored by WireGuard and has no effect on VPN behavior.
#### `Endpoint`
Defines the publicly accessible address for a remote peer. This should be left out for peers behind a NAT or peers that don't have a stable publicly accessible IP:PORT pair. Typically, this only needs to be defined on the main bounce server, but it can also be defined on other public nodes with stable IPs like `public-server2` in the example config below.
This defines the IP ranges for which a peer will route traffic. On simple clients, this is usually a single address (the VPN address of the simple client itself). For bounce servers this will be a range of the IPs or subnets that the relay server is capable of routing traffic for. Multiple IPs and subnets may be specified using comma-separated IPv4 or IPv6 CIDR notation (from a single /32 or /128 address, all the way up to `0.0.0.0/0` and `::/0` to indicate a default route to send all internet and VPN traffic through that peer). This option may be specified multiple times.
When deciding how to route a packet, the system chooses the most specific route first, and falls back to broader routes. So for a packet destined to `192.0.2.3`, the system would first look for a peer advertising `192.0.2.3/32` specifically, and would fall back to a peer advertising `192.0.2.1/24` or a larger range like `0.0.0.0/0` as a last resort.
If the connection is going from a NAT-ed peer to a public peer, the node behind the NAT must regularly send an outgoing ping in order to keep the bidirectional connection alive in the NAT router's connection table.
This value should be left undefined as it's the client's responsibility to keep the connection alive because the server cannot reopen a dead connection to the client if it times out.
The examples in these docs primarily use IPv4, but WireGuard natively supports IPv6 CIDR notation and addresses everywhere that it supports IPv4, simply add them as you would any other subnet range or address.
If you want to forward *all* internet traffic through the VPN, and not just use it as a server-to-server subnet, you can add `0.0.0.0/0, ::/0` to the `AllowedIPs` definition of the peer you want to pipe your traffic through.
WireGuard can sometimes natively make connections between two clients behind NATs without the need for a public relay server, but in most cases this is not possible. NAT-to-NAT connections are only possible if at least one host has a stable, publicly-accessible IP address:port pair that can be hardcoded ahead of time, whether that's using a FQDN updated with Dynamic DNS, or a static public IP with a non-randomized NAT port opened by outgoing packets, anything works as long as all peers can communicate it beforehand and it doesn't change once the connection is initiated.
A known port and address need to be configured ahead of time because WireGuard doesn't have a signalling layer or public STUN servers that can be used to search for other hosts dynamically. WebRTC is an example of a protocol that can dynamically configure a connection between two NATs, but it does this by using an out-of-band signaling server to detect the IP:port combo of each host. WireGuard doesn't have this, so it only works with a hardcoded `Endpoint` + `ListenPort` (and `PersistentKeepalive` so it doesn't drop after inactivity).
- At least one peer has to have to have a hardcoded, directly-accessible `Endpoint` defined. If they're both behind NATs without stable IP addresses, then you'll need to use Dynamic DNS or another solution to have a stable, publicly accessibly domain/IP for at least one peer
- At least one peer has to have a hardcoded UDP `ListenPort` defined, and it's NAT router must not do UDP source port randomization, otherwise return packets will be sent to the hardcoded `ListenPort` and dropped by the router, instead of using the random port assigned by the NAT on the outgoing packet
- All NAT'ed peers must have `PersistentKeepalive` enabled on all other peers, so that they continually send outgoing pings to keep connections persisted in their NAT's routing table
1. Peer1 sends a UDP packet to Peer2, it's rejected Peer2's NAT router immediately, but that's ok, the only purpose was to get Peer1's NAT to start forwarding any expected UDP responses back to Peer1 behind its NAT
2. Peer2 sends a UDP packet to Peer1, it's accepted and forwarded to Peer1 as Peer1's NAT server is already expecting responses from Peer2 because of the initial outgoing packet
3. Peer1 sends a UDP response to Peer2's packet, it's accepted and forwarded by Peer2's NAT server as it's also expecting responses because of the initial outgoing packet
This process of sending an initial packet that gets rejected, then using the fact that the router has now created a forwarding rule to accept responses is called "UDP hole-punching".
When you send a UDP packet out, the router (usually) creates a temporary rule mapping your source address and port to the destination address and port, and vice versa. UDP packets returning from the destination address and port (and no other) are passed through to the original source address and port (and no other). This is how most UDP applications function behind NATs (e.g. BitTorrent, Skype, etc). This rule will timeout after some minutes of inactivity, so the client behind the NAT must send regular outgoing packets to keep it open (see `PersistentKeepalive`).
Getting this to work when both end-points are behind NATs or firewalls requires that both end-points send packets to each-other at about the same time. This means that both sides need to know each-other's public IP addresses and port numbers ahead of time, in WireGuard's case this is achieved by hard-coding pre-defined ports for both sides in `wg0.conf`.
As of 2019, many of the old hole-punching methods used that used to work are no longer effective. One example was a novel method pioneered by [pwnat](https://github.com/samyk/pwnat) that faked an ICMP Time Exceeded response from outside the NAT to get a packet back through to a NAT'ed peer, thereby leaking its own source port. Hardcoding UDP ports and public IPs for both sides of a NAT-to-NAT connection (as described above) still works on a small percentage of networks. Generally the more "enterprisey" a network is, the less likely you'll be able to hole punch public UDP ports (commercial public Wi-Fi and cell data NATs often don't work for example).
NAT-to-NAT connections are not possible if all endpoints are behind NAT's with strict UDP source port randomization (e.g. most cellular data networks). Since neither side is able to hardcode a `ListenPort` and guarantee that their NAT will accept traffic on that port after the outgoing ping, you cannot coordinate a port for the initial hole-punch between peers and connections will fail. For this reason, you generally cannot do phone-to-phone connections on LTE/3g networks, but you might be able to do phone-to-office or phone-to-home where the office or home has a stable public IP and doesn't do source port randomization.
NAT-to-NAT connections from behind NATs with strict source-port randomization is possible, you just need a signaling server to tell each side the other's IP:port tuple. Here are a few implementations that achieve this with WireGuard:
Many users report having to restart WireGuard whenever a dynamic IP changes, as it only resolves hostnames on startup. To force WireGuard to re-resolve dynamic DNS `Endpoint` hostnames more often, you may want to use a `PostUp` hook to restart WireGuard every few minutes or hours.
You can see if a hole-punching setup is feasible by using netcat on the client and server to see what ports and connection order work to get a bidirectional connection open: run `nc -v -u -p 51820 <address of peer2> 51820` (on peer1) and `nc -v -u -l 0.0.0.0 51820` (on peer2), then type in both windows to see if you can get bidirectional traffic going. If it doesn't work regardless of which peer sends the initial packet, then WireGuard won't be unable to work between the peers without a public relay server.
All of the userspace implementations are slower than the native C version that runs in kernel-land, but provide other benefits by running in userland (e.g. easier containerization, compatibility, etc.).
WireGuard will ignore a peer whose public key matches the interface's private key. So you can distribute a single list of peers everywhere, and only define the `[Interface]` separately on each server.
It's up to you to decide how you want to share the `peers.conf`, be it via a proper orchestration platform, something much more pedestrian like Dropbox, or something kinda wild like Ceph. I dunno, but it's pretty great that you can just wildly fling a peer section around, without worrying whether it's the same as the interface.
#### Setting config values from files or command outputs
You can set config values from arbitrary commands or by reading in values from files, this makes key management and deployment much easier as you can read in keys at runtime from a 3rd party service like Kubernetes Secrets or AWS KMS.
WireGuard can be run in Docker with varying degrees of ease. In the simplest case, `--privileged` and `--cap-add=all` arguments can be added to the docker commands to enable the loading of the kernel module.
Setups can get somewhat complex and are highly dependent on what you're trying to achieve. You can have WireGuard itself run in a container and expose a network interface to the host, or you can have WireGuard running on the host exposing an interface to specific containers.
For more detailed instructions, see the [QuickStart](#QuickStart) guide and API reference above. You can also download the complete example setup here: https://github.com/pirate/wireguard-example.
