This is a reimplementation of the client- and server-side rule processing mechanisms of the Geneva project.
Geneva is both a method to describe ways of manipulating packets to attempt to circumvent censorship, and a genetic algoritm (GENetic EVAsion) that one can deploy to discover new circumventions. (This package does not implement the genetic algorithm.) More broadly, though, one can encode arbitrary instructions for packet manipulation using Geneva rules as a sort of "standard syntax", although the use case outside of censorship circumvention may be somewhat tenuous.
This package aims to implement the same triggers and actions that the Geneva project's canonical Python package does. Please note: this package is a work-in-progress, and there are still things left to implement.
Geneva rules are called strategies. A strategy consists of zero or more action trees that can be applied to inbound or outbound packets. The actions trees define both a trigger and a tree of actions to take on a packet if the trigger matches. The result of an action tree will be zero or more packets that should replace the original packet, which then can be reinjected into the host OS' network stack.
Let's work from the top down. A strategy, conceptually, looks like this:
outbound-forest \/ inbound-forest
outbound-forest and inbound-forest are ordered lists of (trigger, action tree) pairs. The Geneva paper calls these
ordered lists forests. The outbound and inbound forests are separated by the \/ characters (that is a backslash
followed by a forward-slash); if the strategy omits one or the other, then that side of the \/ is left empty. For
example, a strategy that only includes an outbound forest would take the form outbound \/, whereas an inbound-only
strategy would be \/ inbound.
The original Geneva paper does not have a name for these (trigger, action tree) pairs. In practice, however, the Python code actually defines an action tree as a (trigger, action) pair, where the "action" is the root of a tree of actions. This package follows this nomenclature as well.
A real example, taken from the original paper (pg 2202), would look like this:
[TCP:flags:S]-
duplicate(
tamper{TCP:flags:replace:SA}(
send),
send)-| \/
[TCP:flags:R]-drop-|
In this example, the outbound forest would trigger on TCP packets that have just the SYN flag set, and would perform a
few different actions on those packets. The inbound forest would only apply to TCP packets with the RST flag set, and
would simply drop them. Each of the forests in the example are made up of a single (trigger, action tree) pair.
The outbound forest for the above action in graph form looks like this:
In a forest, each action tree must adhere to the syntax [trigger]-action-|. Currently, the parser for this package is
stricter than the original; this is a bug.
A trigger defines a way to match packets so that an action tree can be applied to them. In the example above, the first
trigger is [TCP:flags:S]. This is a trigger that matches on the TCP segment's flags field, and requires that only
the SYN flag be set. (Note that this trigger will not fire for packets that have, i.e., both SYN and ACK set.) If
the packet is not a TCP packet, or the flags do not match exactly, then this trigger will not fire.
An action simply encodes steps to manipulate a packet. There are a number of actions described in the Geneva paper:
The "send" action simply yields the given packet. (A quirk—what the paper calls canonical syntax—is to elide any "send" actions in the action tree. For instance, the action "duplicate(,)" is equivalent to "duplicate(send,send)". Bear this in mind when reading Geneva strategies!)
The "drop" action discards the given packet.
The "duplicate" action copies the original packet, then applies action a1 to the original and a2 to the copy. For
example, if a1 and a2 are both "send" actions, then the action will yield two packets identical to the first.
The "fragment" action takes the original packet and fragments it, applying a1 to one of the fragments and a2 to the
other. Since both the IP and TCP layers support fragmentation, the rule must specify which layer's payload to
fragment. The first fragment will include up to offset bytes of the layer's payload; the second fragment will contain
the rest. As an example, given an IPv4 packet with a 60-byte payload and an 8-byte offset, the first fragment will have
the same IP header as the original packet (aside from the fields that must be fixed) and then the first eight bytes of
the payload. The second fragment will contain the other 52 bytes. (You can also indicate that the fragments be returned
out-of-order; i.e., reversed, by specifying "False" for the inOrder argument.)
The "tamper" action takes the original packet and modifies it in some fashion, depending on the protocol, field, and mode given. There are two modes: replace and corrupt. The "replace" mode will replace the value of the given field with newValue, while the "corrupt" mode will replace the value with random data. (Note that there are other modes that the Python code supports that are not defined in the original Geneva paper.)
Additionally, note that not all actions are valid for both inbound and outbound directions. The Python code mentions that "branching actions are not supported on inbound trees". Practically, this means that the duplicate and fragment actions can only be applied to outbound packets, while the sleep, drop, and tamper actions can apply to packets of either direction. (Note that this package does not currently enforce this; this is also a bug.)
Currently only IPv4 and TCP are supported. There are plans to add support for UDP in the future (although pull requests
are welcome! Look at TCPTamperAction and IPv4TamperAction in actions/tamper_action.go as examples.
UDPTamperAction must implement the actions.Action interface). There are no plans at the moment to add support
for IPv6.
See https://censorship.ai for more information about Geneva itself.