Skip to content
A testbed for NAT traversal software.
Go Other
  1. Go 99.2%
  2. Other 0.8%
Branch: main
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.

NATlab is a testbed for NAT traversal software. It intercepts packets before Linux's NAT implementation can process them, and does its own translation according to configurable policies. This enables NATlab to emulate any NAT behavior you desire.

Combined with something like docker-compose, you can construct complex network topologies, featuring multiple NAT gateways with different behaviors, and see how well your software is able to traverse them.


We write ip:port combinations as X:x, X1:x1, Y2:y2 and so forth. We write a NAT mapping as X1:x1 <> X1':x1', meaning "the LAN source ip:port X1:x1 is rewritten to X1':x1' as packets go through the NAT box".

NAT behaviors

NATlab implements a configurable NAT gateway, with configuration knobs for each of the behaviors identified by RFC 4787. By varying these knobs, you can emulate a wide variety of NAT devices.

In general, the questions we're trying to answer are:

  • When a packet shows up on the LAN interface, what do its IPs and ports get translated to when leaving on the WAN interface?
  • When a packet shows up on the WAN interface, where does it go on the LAN interface - if anywhere? (it might get dropped)

The specific knobs offered by NATlab are as follows. We'll use REQ-x to reference each behavior, although the RFC uses REQ-x to specify what a nice NAT box should do, we use it to describe all possible behaviors.

REQ-1: Mapping reuse behavior

When a LAN client sends packets from X1:x1, the NAT box assigns a NAT mapping that rewrites the source to X1':x1' on the way out. The LAN client then continues to send from X1:x1, but varies the destination IP and/or port.

  1. Endpoint-Independent: The mapping X1:x1 <> X1':x1' is reused regardless of the destination address.

  2. Address-Dependent: The mapping X1:x1 <> X1':x1 is reused as long as the destination IP address is the same.

  3. Address-And-Port-Dependent: The mapping X1:x1 <> X1':x1' is reused as long as both the destination IP and destination port are the same.

RFC recommends Endpoint-Independent behavior.

REQ-2: IP address pooling

Assume the NAT box has multiple public IP addresses it can choose from when creating new NAT mappings. A LAN client sends packets from X1, on various source ports and to various destinations. What public IP(s) will be used for mappings?

  1. Arbitrary: different sessions from X1 may get different public IPs X1', X2', ....
  2. Paired: all sessions from X1 use a single public IP X1'. If X1' has no free ports to use for a new mapping, new sessions are dropped.
  3. Soft-Paired: all sessions from X1 use a single public IP X1'. If X1' has no free ports to use for a new mapping, ports on other IPs may be used.

RFC recommends Paired behavior, but notes that many enterprise NAT gateways use Arbitrary behavior "for security reasons" (i.e. wooly thinking).

REQ-3: Port assignment

When a new NAT mapping X1:x1 <> X1':x1' needs to be created, how does the NAT box pick x1'?

  1. Port-Overloading: The NAT box will set x1' = x1. If this results in a collision, the previous mapping is deleted.
  2. Port-Preserving: The NAT box will attempt to set x1' = x1. In case of collision, the NAT box will pick some other x1'.
  3. Arbitrary: The NAT box will not attempt to keep x1' = x1 at all.

REQ-3 does not specify what source IP to use. It's allowable to pick x1' on any available public IP, subject to the constraints of the other settings.

RFC recommends anything but Port-Overloading, for the obvious reason that it breaks stuff hilariously.

REQ-4: Port Parity

When picking a port x1' to map X1:x1, does the NAT box attempt to keep LSB(x1') == LSB(x1), i.e. map even ports to even ports and odd ports to odd ports?

This is a silly distinction designed to placate ancient RTP clients, so NATlab doesn't implement it.

REQ-5: Mapping refresh timer

How long can a mapping go without seeing "qualifying traffic" (see REQ-6) before it gets deleted?

RFC recommends 5 minutes, and begs vendors to not set it lower than 2 minutes. Exceptions exist for certain types of traffic, e.g. DNS can have much shorter timeouts. NATlab doesn't (yet?) support overriding the timer by port.

REQ-6: Qualifying packets for mapping refresh

What packets trigger a renewal of the NAT mapping's lease?

  1. Outbound-Only: only packets going from LAN to WAN refresh the mapping.
  2. Inbound-Only: only packets going from WAN to LAN refresh the mapping.
  3. Both: packets going in either direction refresh the mapping.

RFC recommends Outbound-Only at minimum, ideally Both (but points out that Both may enable a resource DoS on the NAT box, so "for security reasons" expect Outbound-Only to be the norm)

REQ-7: Internal/External address conflicts

This specifies that if LAN and WAN have address collisions, the NAT box should handle this gracefully. NATlab doesn't implement this at all, as afaict this is a theoretical concern rather than a practical one.

REQ-8: Filtering Behavior

When packets are traversing the NAT from WAN to LAN and match a NAT mapping, what packets are permitted to flow back to the LAN client?

  1. Endpoint-Independent: all packets that matched the NAT mapping can flow.
  2. Address-Dependent: packets from Y:* can only flow if the LAN client has previously sent packets to Y.
  3. Address-And-Port-Dependent: packets from Y:y can only flow if the LAN client has previously sent packets to Y:y.

RFC recommends either Endpoint-Independent or Address-Dependent, depending on paranoia levels.

REQ-9: Hairpinning behavior

If two clients X1:x1 and X2:x2 are on the same LAN, can they use each other's public addresses X1':x1' and X2':x2' and hairpin through the NAT box?

  1. No: hairpinning is not allowed, packets that attempt to hairpin are dropped.
  2. Internal-Source: hairpinning is allowed. Packets are delivered with the internal ip:port as the source.
  3. External-Source: hairpinning is allowed. Packets are delivered with the external ip:port as the source.

RFC requires External-Source behavior.

REQ-10: ALG behavior

This requirement has to do with NATs having explicit behavioral support for certain protocols (e.g. VOIP). The RFC says to disable all ALGs, and NATlab doesn't implement any.

REQ-11: Determinism

Roughly, this section says NATs shouldn't vary one REQ- behavior based on the outcome of another REQ- behavior. Some specific examples are given, but each one would have to be encoded as another sub-behavior of the appropriate REQ-.

For now, NATlab's implementation is deterministic according to this section.

REQ-12: ICMP support

Does the NAT box correctly rewrite and forward ICMP messages that pertain to an active NATed session?

  1. Yes: ICMP messages whose payload contains fragments of UDP packets are correctly rewritten and forwarded across the NAT.
  2. No: ICMP messages are eaten by the NAT box.

RFC recommends Yes.

REQ-13, REQ-14: IP Fragmentation

These are just a requirement that the NAT gateway should handle IP fragmentation correctly. DF=1 packets should result in ICMP Fragmentation Needed + drop, DF=0 should result in fragmentation.

NATlab relies on the linux kernel to do this right.

XXX-1: NAT helper protocols

This isn't from the RFC, but there are a variety of "NAT helper" protocols that a client can use to explicitly create a port mapping. Eventually, we'd like NATlab to support them. They are:

  1. UPnP IGDP: a horror in XML, SOAP and UDP. Never seen implemented cleanly, because one does not implement cursed protocols cleanly.
  2. NAT-PMP: Apple's surprising contribution to not being completely impossible to implement. It's really quite nice.
  3. PCP: Evolution of NAT-PMP, similarly nice. Also includes logic for NAT64 and other esoterica.
You can’t perform that action at this time.