Traffic on ZeroTier networks can be observed and controlled with a system of globally applied network rules. These are enforced in a distributed fashion by both the senders and the receivers of packets. To escape the rules engine a malicious attacker would need to fully compromise both sides of any conversation.
The ZeroTier VL2 rules engine differs from most other firewalls and SDN rules engines in several ways. The most immediately relevant of these is that the ZeroTier rules engine is stateless, meaning it lacks connection tracking. This means that bidirectional whitelisting can’t be accomplished by simply whitelisting reply packets to established connections. Instead some thought must be put into how to allow both sides of a desired flow. Rule patterns to achieve the most common desired objectives are included in this manual.
The decision to make our rules engine stateless was a design trade-off driven by several concerns. First we wanted to keep complexity, code footprint, and memory use very low to support small embedded devices. The second and more fundamental reason is that distributed stateful filtering requires distributed state synchronization. This would have added a large volume of additional sync traffic as well as introducing inescapablenew sources of instability and failure and a lot of surface area for security vulnerabilities.
While ZeroTier lacks state tracking, its rules engine includes something not found anywhere else in the enterprise networking space: capability-based security and device tagging. Capabilities and tags allow extremely complex micro-segmented network rule schemes to be implemented in a sane, conceptual way that is both easier for human beings to understand and more efficient for machines to handle.
This section assumes some level of familiarity with network rules as they’re commonly used on firewalls and routers, etc. While the rules engine is part of VL2, it’s been given its own section in this manual due to the depth and cross-cutting nature of the topic.
Rule sets are ordered lists of one or more rules, with each rule consisting of one or more match conditions followed by one action. As a rule set is evaluated, each match is tested in order and is then ANDed or ORed with the previous match result state. When an action is encountered it is taken if the result of the preceding matches is true. An action with no preceding matches is always taken. If no permissive actions are taken by any rule set the packet is discarded.
Here is a simple rule set that constrains Ethernet traffic on a network to only IPv4, ARP, or IPv6 as it would appear in the raw JSON format used by ZeroTier One’s built-in network controller implementation. (Don’t worry if this seems verbose and difficult. We have a more human-friendly way of writing rule sets, but before we introduce it it’s important to understand what is really happening.)
This checks whether an Ethernet level packet is not IPv4 (ethertype 2048) and not IPv4 ARP (ethertype 2054) and not IPv6 (ethertype 34525). If all three matches evaluate to true (meaning the ethertype is none of these) then the drop action is taken. Otherwise the accept action is taken.
Networks have one base rule set that is applied to all traffic. Its size is constrained to 1024 entries (each match or action is an entry). It should be used to set the overall policies for all members of the network, and for most common use cases it’s all you’ll need. For more complex scenarios, both capabilities and tags provide methods of both managing complexity and scaling the overall size of a network’s rule system.
These are the available matches and actions in raw form. The rule definition language outlined in section 3.4 provides a friendlier way for human beings to specify rules.
|ACTION_DROP||Drop packet and terminate all rule evaluation, including capabilities.|
|ACTION_BREAK||Terminate evaluation of this rule set but continue evaluating capabilities.|
|ACTION_ACCEPT||Accept packet and terminate further evaluation.|
|ACTION_TEE||length,address||Send a copy of up to the first length bytes (-1 for all) to a ZeroTier address.|
|ACTION_REDIRECT||address||Transparently redirect this packet to a ZeroTier address without changing its headers.|
|MATCH_SOURCE_ZEROTIER_ADDRESS||zt||VL1 source address.|
|MATCH_DEST_ZEROTIER_ADDRESS||zt||VL1 destination address.|
|MATCH_MAC_SOURCE||mac||Ethernet source address.|
|MATCH_MAC_DEST||mac||Ethernet destination address.|
|MATCH_IPV4_SOURCE||ip||IPv4 source address.|
|MATCH_IPV4_DEST||ip||IPv4 destination address.|
|MATCH_IPV6_SOURCE||ip||IPv6 source address.|
|MATCH_IPV6_DEST||ip||IPv6 destination address.|
|MATCH_IP_TOS||mask,start,end||IP TOS field bitwise-ANDed with mask is within range.|
|MATCH_IP_PROTOCOL||ipProtocol||IP protocol number.|
|MATCH_ETHERTYPE||etherType||Ethernet frame type.|
|MATCH_ICMP||icmpType,icmpCode||ICMP type and code, if applicable. (V4 or V6)|
|MATCH_IP_SOURCE_PORT_RANGE||start,end||IP (V4 or V6) source port range (inclusive).|
|MATCH_IP_DEST_PORT_RANGE||start,end||IP (V4 or V6) destination port range (inclusive).|
|MATCH_CHARACTERISTICS||mask||Bitwise AND characteristic bits with mask, true if result is nonzero.|
|MATCH_FRAME_SIZE_RANGE||start,end||Ethernet frame size range (inclusive).|
|MATCH_RANDOM||probability||Match if a random 32-bit number is less than or equal to the probability.|
|MATCH_TAGS_DIFFERENCE||id,value||Difference between tags with this ID is less than or equal to the value.|
|MATCH_TAGS_BITWISE_AND||id,value||Tags ANDed together equal value.|
|MATCH_TAGS_BITWISE_OR||id,value||Tags ORed together equal value.|
|MATCH_TAGS_BITWISE_XOR||id,value||Tags XORed together equal value.|
|MATCH_TAGS_EQUAL||id,value||Both tags equal the same value.|
|MATCH_TAG_SENDER||id,value||Sending side’s tag equals this value.|
|MATCH_TAG_RECEIVER||id,value||Receiving side’s tag equals this value.|
A capability is a small rule set that is bundled into a credential object, signed by the network controller, and issued to only those member(s) permitted to exercise it. When a member detects that outgoing traffic does not match the base rule set but is allowed by one of its capabilities, it periodically pushes the matching capability credential to the recipient ahead of the packet(s) in question. Peer to peer capability distribution is automatic and is triggered by capability match.
When the recipient receives the capability it authenticates it by checking its signature and timestamp and, provided the capability is valid, adds it to the set of capabilities to apply to incoming traffic from the capability’s owner. The sender has effectively told the recipient “I can too send this packet! Teacher says so!”
Capabilities allow large systems of rules to be broken down into functional aspects and then distributed intelligently only to those members with a need to know. This avoids the bandwidth and storage overhead of distributing huge monolithic rule sets and organizes rules conceptually to make them easier for administrators to understand.
There are three terminating actions that can be taken in a rule set: accept, break, and drop. The accept action terminates rule evaluation and accepts the packet. The break action terminates the evaluation of the current rule set but permits the further evaluation of capabilities. The drop action terminates rule evaluation and drops the packet without checking capabilities in the base rule set, but is equivalent to break in capability rule sets. In most cases break should be used unless certain traffic must be absolutely prohibited under any circumstance.
In the simple base rule set example in section 3.1 the drop action is
taken in the unapproved case. This means that ethernet whitelisting
cannot be overridden by a capability. If we change
ACTION_DROP in our
ACTION_BREAK, then it becomes possible to issue the
Ethertype 2114 is wake-on-LAN, a special packet that can cause some systems to wake from sleep mode. If we place the above tiny rule set into a capability and issue it to a device, this device but no others will now be permitted to send wake-on-LAN magic packets. (Wake-on-LAN requires hardware support so it would only work to target devices plugged into a physical network bridged to a ZeroTier network, but don’t worry about that here. It’s just an example of special traffic.)
Capability rule sets are limited to only 64 entries. The idea is to keep them small and simple. A capability should grant one thing or one small set of conceptually related things.
ZeroTier provides a second mechanism to control rule set complexity. Tags are 32-bit numeric key-value pair credentials that are issued to network members and signed by the controller. They are then distributed peer to peer on a need to know basis in a similar manner to capabilities.
Tags provide a way to conditionally drop or allow traffic between members by member classification. They allow very detailed network micro-segmentation by member role, permission, function, etc. without resulting in a combinatorial explosion in rules table size.
Let’s say we want to permit traffic on TCP ports 139 and 445 (netbios/CIFS file sharing) only between systems that belong to the same department. Our company has 12,000 devices and 10 departments. Without tags this would require 144,000,000 rules, but with tags it can be accomplished by only a few.
First a tag is created to represent the department. Let’s give it tag ID 100. Each member system receives the tag with a value from 1 to 10 indicating which department it belongs to. We can then add the following rules to our network’s base rule set (or to a capability if so desired):
This tells members in our network to accept TCP packets on ports 139 or 445 if the difference between tags with tag ID 10 is zero, meaning they match. (If a member does not have a value for this tag, it does not match.) Now all members of the same department can access CIFS file shares, but CIFS sharing between departments could still be prohibited. (TCP whitelisting requires some additional rules due to the stateless nature of our rules engine. See the section below on rule design patterns.)
Tags can be compared on numeric value or as bit fields via several different bit mask operations allowing many different systems of member classification to be implemented.
Raw rule sets are verbose and difficult to write, so we created a minimal rule definition langugage that’s easier for human beings.
This creates a network that can pass IPv4 (and ARP) and IPv6 traffic but
no other Ethernet frame types. In addition the
not chr ipauth
condition drops traffic between IP addresses that have not been assigned
by ZeroTier to their respective sources or destinations, blocking all IP
spoofing. These are enforced with a hard
drop, preventing them from
being overridden by any capability.
All UDP is allowed, but all non-whitelisted new TCP connections
(SYN/!ACK packets) are blocked. This is done with a soft
allowing it to be overridden by capabilities. New TCP connections are
allowed on ports 22, 80, and 443 for everyone. Windows file sharing is
allowed only between computers in the same department by way of a tag. A
super-user capability that can be assigned to administrative nodes
allows the sender to initiate any kind of connection.
See section 3.5.1 for a discussion of how we accomplish TCP whitelisting here.
Each action, tag, capability, or macro block ends with a semicolon. White space separates things. Indentation is not significant. Hash symbols indicate that the remainder of a line is a comment.
As described in sections 3.1 through 3.3, a rule set is composed of one or more sequences of match,[match],…,action in which the action is taken if the chain of matches evaluates to true. Capabilities are small bundles of rules that can be assigned to nodes to give them special abilities, and tags are key/value pairs that can be assigned to nodes to allow rules to be selectively applied.
Macros are just a convenient way of defining more complex actions that can then be applied in multiple places. An example of a macro might be:
You could then allow TCP ports in a standard TCP whitelisting scheme by just saying:
flag directives have no meaning to the actual ZeroTier
rules engine, but they can be used by user interfaces like ZeroTier
Central to make it easier to assign and search tags.
|drop||Drop packet and terminate all rule evaluation, including capabilities.|
|break||Terminate evaluation of this rule set but continue evaluating capabilities.|
|accept||Accept packet and terminate further evaluation.|
|tee||<length> <address>||Send a copy of up to the first length bytes (-1 for all) to a ZeroTier address.|
|redirect||<address>||Transparently redirect this packet to a ZeroTier address without changing its headers.|
|ztsrc||<address>||VL1 source address.|
|ztdest||<address>||VL1 destination address.|
|macsrc||<address>||Ethernet source address.|
|macdest||<address>||Ethernet destination address.|
|ipsrc||<address/prefix>||Source IP address or network (V4 or V6 auto-detected).|
|ipdest||<address/prefix>||Destination IP address or network (V4 or V6 auto-detected).|
|iptos||<mask> <start[-end]>||IP TOS field bitwise-ANDed with mask is within range.|
|ipprotocol||<protocol>||IP protocol number.|
|ethertype||<type>||Ethernet frame type.|
|icmp||<type> <code>||ICMP type (V4 or V6) and code. Use -1 for code if not applicable.|
|sport||<start[-end]>||IP (V4 or V6) source port range (inclusive).|
|dport||<start[-end]>||IP (V4 or V6) destination port range (inclusive).|
|chr||<mask/name>||Bitwise AND characteristic bits with mask, true if result is nonzero.|
|framesize||<start[-end]>||Ethernet frame size range (inclusive).|
|random||<probability>||Match randomly with probability 0.0 (0%) to 1.0 (100%).|
|tdiff||<id> <value>||Difference between tags with this ID is less than or equal to the value.|
|tand||<id> <value>||Tags ANDed together equal value.|
|tor||<id> <value>||Tags ORed together equal value.|
|txor||<id> <value>||Tags XORed together equal value.|
|teq||<id> <value>||Both tags equal the same value.|
|tseq||<id> <value>||Sending side’s tag equals this value.|
|treq||<id> <value>||Receiving side’s tag equals this value.|
Match conditions may be joined by and (default if none specified) or or and may be modified by not.
Match conditions are evaluated from left to right;
a and b or c is a different rule than
b or c and a.
The rules language doesn't support parens, but for example purposes the above would evaluate as:
(a and b) or c vs
(b or c) and a
For convenience the following symbols can be used when matching on certain packet attributes:
- IP protocols: icmp (for IPv4), igmp, ipip, tcp, egp, igp, udp, rdp, esp, ah, icmp6 (for IPv6), l2tp, sctp, and udplite.
- Ethernet frame types: ipv4, arp, ipv6, wol (wake on LAN), rarp, atalk, aarp, ipx*a**, ipx***b**.
- Packet characteristics (bit masks for chr):
- inbound: packet is being filtered on the receiving side (use
not inboundfor sending side)
- multicast: destination is a multicast or broadcast MAC
- broadcast: destination is the broadcast MAC
- ipauth: sender IP is assigned by ZeroTier to the sending node
- tcp_fin: packet is TCP with FIN flag set
- tcp_syn: packet is TCP with SYN flag set
- tcp_rst: packet is TCP with RST flag set
- tcp_psh: packet is TCP with PSH flag set
- tcp_ack: packet is TCP with ACK flag set
- tcp_urg: packet is TCP with URG flag set
- tcp_ece: packet is TCP with ECE flag set
- tcp_cwr: packet is TCP with CWR flag set
- tcp_ns: packet is TCP with NS flag set
- tcp_rs2: packet is TCP with RS2 (reserved bit 2) flag set
- tcp_rs1: packet is TCP with RS1 (reserved bit 1) flag set
- tcp_rs0: packet is TCP with RS0 (reserved bit 0) flag set
- inbound: packet is being filtered on the receiving side (use
First, add this at or near the bottom of your rules:
Then above the SYN,!ACK break (or in a capability) add rules to allow TCP packets with permitted destination ports:
ZeroTier’s filter is stateless. If we block all TCP packets except those with the correct destination port, this will prevent reply packets from returning to their senders.
Allowing reply packets with the correct (inverse) source port seems like a simple fix, but this is insecure as it allows anyone to initiate any TCP connection if they bind to whitelisted source ports.
The best solution is to block all non-whitelisted TCP packets with flags SYN,!ACK (SYN and not ACK) and and then whitelist desired destination ports. This way TCP replies and control traffic are allowed, but new connections cannot be opened unless they are permitted.
UDP is tougher to deal with in a stateless paradigm. It’s connectionless so there is no way to specifically select a new session vs. an existing session. The best way to lock down UDP on a network is to use tags to allow it to and from things like DNS servers that need to speak it.
First we define a tag called
udpserver with a default value of 0. We
don’t set any enums or flags for this tag since it will be used as a
boolean. For servers that need to respond to DNS queries, set the
Then we accept UDP traffic if the value of the
udpserver tag is
when both sender and receiver tags are ORed together, or if UDP traffic
is multicast. This allows multicast mDNS and Netbios announcements and
allows UDP traffic to and from UDP servers, but prohibits other
horizontal UDP traffic.
Here’s a simple rule to monitor everything:
That’s going to flood
deadbeef11 with two full copies of every single
packet, since it will match on both the sender and the recipient side. A
less bandwidth-intensive security monitor setup might look like this:
This is a bit more clever. It sends the first 128 bytes of every TCP SYN, RST, or FIN packet (TCP connection open and close) to one observer on the inbound side and another observer on the outbound side. It also sends the first 128 bytes of other packets with a probability of one in ten. The first 128 bytes of a packet will be enough to see the Ethernet and IP headers as well as layer 4+ information about many protocols.
This would allow observers to watch every new TCP connection on the network and also passively monitor other traffic in a “fuzzy” probabilistic fashion without using very much bandwidth. We split sending and receiving observers to prevent duplicate packets and also to allow us to detect cases where one side is failing to tee traffic more frequently than would be easily explained by packet loss.
There are many, many variations on the above that are possible but hopefully these will be enough to get you started.
Passive out-of-line monitoring with tee is not perfect. If bandwidth between two parties exceeds the bandwidth of one or both parties to the observer, packets will get dropped. It also provides no mechanism to filter traffic in depth since the observer cannot directly intervene (though it could de-authorize a member via the controller API).
Active in-line monitoring can be achieved with redirect:
This will pipe all HTTP traffic through
deadbeef11. The target node
will receive each Ethernet frame intact including its original Ethernet
source and destination MAC address. To forward the packet on to its
final destination, simply re-send it unmodified via the same interface
on which it was received. The intermediate node could scan, modify, and
interrupt HTTP traffic across the network. To regular network
participants this is completely invisible as it occurs at layer 2
(Ethernet), though some performance degradation may be noticed
especially if the link is running across WAN.
One useful pattern for access control resembles the way military organizations classify data. Information is deemed classified, and only those who have the required level of classification are allowed to access it.
A model resembling this can be implemented in ZeroTier network rules as follows:
Initially members will be assigned a default classification of 0 (“no”). These can freely communicate since the bitwise OR of their classified tags will be zero. Neither member possesses a classification requirement.
To restrict access to a member, set its classified tag to secret or
top. (In this example there is no difference, but two levels are
included in case you want to implement some kind of more detailed
segmentation based on these.) Now the first match
not tor classified 0) will be true and the packet will be dropped
unless the two communicating members have at least once clearance bit
in common (
tand clearance 0).