Every ZeroTier virtual network has a *network controller*. This is our reference implementation and is the same one we use to power our own hosted services at [my.zerotier.com](https://my.zerotier.com/). Network controllers act as configuration servers and certificate authorities for the members of networks. Controllers are located on the network by simply parsing out the first 10 digits of a network's 16-digit network ID: these are the address of the controller.
As of ZeroTier One version 1.2.0 this code is included in normal builds for desktop, laptop, and server (Linux, etc.) targets, allowing any device to create virtual networks without having to be rebuilt from source with special flags to enable this feature. While this does offer a convenient way to create ad-hoc networks or experiment, we recommend running a dedicated controller somewhere secure and stable for any "serious" use case.
Controller data is stored in JSON format under `controller.d` in the ZeroTier working directory. It can be copied, rsync'd, placed in `git`, etc. The files under `controller.d` should not be modified in place while the controller is running or data loss may result, and if they are edited directly take care not to save corrupt JSON since that can also lead to data loss when the controller is restarted. Going through the API is strongly preferred to directly modifying these files.
Older versions of this code used a SQLite database instead of in-filesystem JSON. A migration utility called `migrate-sqlite` is included here and *must* be used to migrate this data to the new format. If the controller is started with an old `controller.db` in its working directory it will terminate after printing an error to *stderr*. This is done to prevent "surprises" for those running DIY controllers using the old code.
The migration tool is written in nodeJS and can be used like this:
Controllers can in theory host up to 2^24 networks and serve many millions of devices (or more), but we recommend spreading large numbers of networks across many controllers for load balancing and fault tolerance reasons. Since the controller uses the filesystem as its data store we recommend fast filesystems and fast SSD drives for heavily loaded controllers.
Since ZeroTier nodes are mobile and do not need static IPs, implementing high availability fail-over for controllers is easy. Just replicate their working directories from master to backup and have something automatically fire up the backup if the master goes down. Many modern orchestration tools have built-in support for this. It would also be possible in theory to run controllers on a replicated or distributed filesystem, but we haven't tested this yet.
ZeroTier network controllers can easily be run in Docker or other container systems. Since containers do not need to actually join networks, extra privilege options like "--device=/dev/net/tun --privileged" are not needed. You'll just need to map the local JSON API port of the running controller and allow it to access the Internet (over UDP/9993 at a minimum) so things can reach and query it.
The controller API is hosted via the same JSON API endpoint that ZeroTier One uses for local control (usually at 127.0.0.1 port 9993). All controller options are routed under the `/controller` base path.
The controller microservice does not implement any fine-grained access control (authentication is via authtoken.secret just like the regular JSON API) or other complex mangement features. It just takes network and network member configurations and reponds to controller queries. We have an enterprise product called [ZeroTier Central](https://my.zerotier.com/) that we host as a service (and that companies can license to self-host) that does this.
All working network IDs on a controller must begin with the controller's ZeroTier address. The API will *allow* "foreign" networks to be added but the controller will have no way of doing anything with them since nobody will know to query it. (In the future we might support secondaries, which would make this relevant.)
The JSON API is *very* sensitive about types. Integers must be integers and strings strings, etc. Incorrectly typed and unrecognized fields may result in ignored fields or a 400 (bad request) error.
| clock | integer | Current clock on controller, ms since epoch | no |
#### `/controller/network`
* Purpose: List all networks hosted by this controller
* Methods: GET
* Returns: [ string, ... ]
This returns an array of 16-digit hexadecimal network IDs.
#### `/controller/network/<network ID>`
* Purpose: Create, configure, and delete hosted networks
* Methods: GET, POST, DELETE
* Returns: { object }
By making queries to this path you can create, configure, and delete networks. DELETE is final, so don't do it unless you really mean it.
When POSTing new networks take care that their IDs are not in use, otherwise you may overwrite an existing one. To create a new network with a random unused ID, POST to `/controller/network/##########______`. The #'s are the controller's 10-digit ZeroTier address and they're followed by six underscores. Check the `nwid` field of the returned JSON object for your network's newly allocated ID. Subsequent POSTs to this network must refer to its actual path.
* Networks without rules won't carry any traffic. If you don't specify any on network creation an "accept anything" rule set will automatically be added.
* Managed IP address assignments and IP assignment pools that do not fall within a route configured in `routes` are ignored and won't be used or sent to members.
* The default for `private` is `true` and this is probably what you want. Turning `private` off means *anyone* can join your network with only its 16-digit network ID. It's also impossible to de-authorize a member as these networks don't issue or enforce certificates. Such "party line" networks are used for decentralized app backplanes, gaming, and testing but are otherwise not common.
For IPv4 the only valid setting is `zt` which, if true, causes IPv4 addresses to be auto-assigned from `ipAssignmentPools` to members that do not have an IPv4 assignment. Note that active bridges are exempt and will not get auto-assigned IPs since this can interfere with bridging. (You can still manually assign one if you want.)
IPv6 includes this option and two others: `6plane` and `rfc4193`. These assign private IPv6 addresses to each member based on a deterministic assignment scheme that allows members to emulate IPv6 NDP to skip multicast for better performance and scalability. The `rfc4193` mode gives every member a /128 on a /88 network, while `6plane` gives every member a /80 within a /40 network but uses NDP emulation to route *all* IPs under that /80 to its owner. The `6plane` mode is great for use cases like Docker since it allows every member to assign IPv6 addresses within its /80 that just work instantly and globally across the network.
(You can POST a shortened-form IPv6 address but the API will always report back un-shortened canonical form addresses.)
That defines a range within network `fd00:feed:feed:beef::/64` that contains up to 2^64 addresses. If an IPv6 range is large enough, the controller will assign addresses by placing each member's device ID into the address in a manner similar to the RFC4193 and 6PLANE modes. Otherwise it will assign addresses at random.
Each rule is actually a sequence of zero or more `MATCH_` entries in the rule array followed by an `ACTION_` entry that describes what to do if all the preceding entries match. An `ACTION_` without any preceding `MATCH_` entries is always taken, so setting a single `ACTION_ACCEPT` rule yields a network that allows all traffic. If no rules are present the default action is `ACTION_DROP`.
Rules are evaluated in the order in which they appear in the array. There is currently a limit of 256 entries per network. Capabilities should be used if a larger and more complex rule set is needed since they allow rules to be grouped by purpose and only shipped to members that need them.
* IPv4 and IPv6 IP address rules do not match for frames that are not IPv4 or IPv6 respectively.
*`ACTION_DEBUG_LOG` is a no-op on nodes not built with `ZT_RULES_ENGINE_DEBUGGING` enabled (see Network.cpp). If that is enabled nodes will dump a trace of rule evaluation results to *stdout* when this action is encountered but will otherwise keep evaluating rules. This is used for basic "smoke testing" of the rules engine.
* Multicast packets and packets destined for bridged devices treated a little differently. They are matched more than once. They are matched at the point of send with a NULL ZeroTier destination address, meaning that `MATCH_DEST_ZEROTIER_ADDRESS` is useless. That's because the true VL1 destination is not yet known. Then they are matched again for each true VL1 destination. On these later subsequent matches TEE actions are ignored and REDIRECT rules are interpreted as DROPs. This prevents multiple TEE or REDIRECT packets from being sent to third party devices.
* Rules in capabilities are always matched as if the current device is the sender (inbound == false). A capability specifies sender side rules that can be enforced on both sides.
