go-flows

README

go-flows

This is a highy customizable general-purpose flow exporter. For more information have a look at godoc.

This flow exporter implementation is featured along with a comparison against other flow exporters and related performance measurements in the IEEE Surveys & Tutorials journal article cited below. Please use this citation whenever using or referencing go-flows:

G. Vormayr, J. Fabini and T. Zseby, "Why are My Flows Different? A Tutorial on Flow Exporters," in IEEE Communications Surveys & Tutorials, vol. 22, no. 3, pp. 2064-2103, thirdquarter 2020, doi: 10.1109/COMST.2020.2989695.

Documentation

Overview

This contains a highly customizable general-purpose flow exporter.

Building

For building either "go build", "go install", or the program provided in go-flows-build can be used. The latter allows for customizing builtin modules and can help with building modules as plugins or combined binaries.

Overview

This flow exporter can convert network packets into flows, extract features, and export those via various file formats.

In general the flow exporter reads the feature definitions (i.e. a description of which features to extract, and the flow key), and builds an execution graph (this can be viewed with the option callgraph) from this definition.

Packets are then read from a source and processed in the following pipeline:

source -> filter -> (parse) -> label -> [key] -> (table) -> (flow) -> (record) -> {feature} -> (merge) -> export

() parts in the pipeline are fixed, [] parts can be configured via the specification, {} can be configured via the specification and provided via modules, and everything else can be provided from a module and configured from the command line.

source is a packet source, which must provide single packets as []byte sequences and metadata like capture time, and dropped/filtered packets. The []byte-buffer can be reused for the next packet. For examples look at modules/sources.

filter is a packet filter, which must return true for a given packet if it should be filtered out. For examples look at modules/filters.

parse is a fixed step that parses the packet with gopacket.

label is an optional step, that can provide an arbitrary label for every packet. For examples look at modules/labels.

key is a fixed step that calculates the flow key. Key parameters can be configured via the specification.

table, flow, record are fixed steps that are described in more detail in the flows package.

merge merges the output from every table into one record stream.

The feature step calculates the actual feature values. Features can be provided via modules, and the selection of which features to calculate must be provided via the specification. Features are described in more detail in the flows package.

If a flow ends (e.g. because of timeout, or tcp-rst) it gets exported via the exported, which must be provided as a module and configured via the command line. For examples look at modules/exporters.

The whole pipeline is executed concurrently with the following four subpipelines running concurrently:

source -> filter

parse -> label -> key

n times:
	table -> flow -> record -> feature

log n times:
	merge

export

The "table"-pipeline exists n times, where n can be configured on the command line. Packets are divided onto the different "table"-pipelines according to flow-key.

WARNING: Due to this concurrent processing flow output is neither order nor deterministic (without sorting)! To ensure deterministic output, flow output can be order by start time, stop time (default), or export time.

Specification

Specification files are JSON files based on the NTARC format (https://nta-meta-analysis.readthedocs.io/en/latest/). Only version 2 files can be used. It is also possible to use a simpler format, if a paper specification is not needed.

Simpleformat specification:

{
	"active_timeout": <Number>,
	"idle_timeout": <Number>,
	"bidirectional": <bool>,
	"features": [...],
	"key_features": [...],
	"_control_features": [...],
	"_filter_features": [...],
	"_per_packet": <bool>,
	"_allow_zero": <bool>,
	"_expire_TCP": <bool>
}

V2-formated file:

{
  "version": "v2",
  "preprocessing": {
    "flows": [
      <simpleformat>
    ]
  }
}

Unlike in the NTARC specification active_timeout and idle_timeout MUST be specified (there are no defaults). If bidirectional is true, every flow contains packets from both directions.

key features give a list of features, which are used to compute a flow key. features is a formated list of features to export. This list can also contain combinations of features and operations (https://nta-meta-analysis.readthedocs.io/en/latest/features.html). Only single pass operations can ever be supported due to design restrictions in the flow exporter.

In addition to the features specified in the nta-meta-analysis, two addional types of features are present: Filter features which can exclude packets from a whole flow, and control features which can change flow behaviour like exporting the flow before the end, restarting the flow, or discarding the flow. _per_packet allows exporting one flow per packet. If _allow_zero is true, then packets are accepted, where one of the parts of the flow key would be zero (e.g. non-IP packets for flow keys that contain IP-Addresses). If _expire_TCP is set to false, no TCP-based expiry is carried out (e.g. RST packets). TCP expiry is only carried out if at least the five-tuple is part of the flow key.

A list of supported features can be queried with "./go-flows features"

Example usage

The examples directory contains several example flow specifications that can be used. The general syntax on the command line is "go-flows run <commands>" where <commands> is a list of "<verb> <which> [options] [--]" sequences. <verb> can be one of features, export, source filter, or label, and <which> is the actual module. The options can be queried from the help of the different modules (e.g. go-flows <verb>s <which>; e.g. go-flows exporters ipfix).

Example:

go-flows run features examples/complex_simple.json export ipfix out.ipfix source libpcap input.pcap

Contents

The following list describes all the different things contained in the subdirectories.

• examples: example specifications
• flows: flows package; Contains base flow functionality, which is not dependent on packets
• packet: packet package; Packet-part of the flow implementation
• modules: implementation of exporters, filters, labels, sources, and features
• util: package wit utility functions
• go-flows-build: build script for customizing binaries and compiling plugins

Implementing Features

Features most follow the conventions in https://nta-meta-analysis.readthedocs.io/en/latest/features.html, which states that names must follow the ipfix iana assignments (https://www.iana.org/assignments/ipfix/ipfix.xhtml), or start with an _ for common features or __ for uncommon ones. Feature names must be camelCase. The flow exporter has the full list of ipfix iana assignments already builtin which means that for these features one needs to only specifiy the name - all type information is automatically added by the flow extractor.

For implementing features most of the time flows.BaseFeature is a good start point. Features need to override the needed methods:

• Start(*EventContext) gets called when a flow starts. Do cleanup here (features might be reused!). MUST call flows.BaseFeature.Start from this function!

• Event(interface{}, *EventContext, interface{}) gets called for every packet belonging to the current flow

• Stop(FlowEndReason, *EventContext) gets called when a flow finishes (before export)

• SetValue(new interface{}, when *EventContext, self interface{}) Call this one for setting a value. It stores the new value and forwards it to all dependent features.

Less commonly used functions

• Variant() gets called to determine which variant to use, if a feature can different types depending on data (see sourceIPAddress for an example)
• SetArguments(args []int, all []Feature) gets called during instantiation if a feature expects constant arguments (args contains indizes of all; see select_slice for an example)

See also documentation of subpackage flows for more details about which base to choose.

A simple example is the protocolIdentifier:

type protocolIdentifier struct {
	flows.BaseFeature
}

func (f *protocolIdentifier) Event(new interface{}, context *flows.EventContext, src interface{}) {
	if f.Value() == nil {
		f.SetValue(new.(packet.Buffer).Proto(), context, f)
	}
}

This feature doesn't need a Start or Stop (since both functions don't provide a packet). For every packet, it checks, if the protocolIdentifier has already been set, and if it hasn't been, it sets a new value. The new value provided to Event will always be a packet.Buffer for features that expect a raw packet. For other features, this will be the actual value emitted from other features. E.g. for the specification

{"eq": ["protocolIdentifier", 17]}

the minfeature will receive the uint8 emitted by this feature.

The final component missing from the code is the feature registration. This has to be done in init with one of the Register* functions from the flows packet. For the protocolIdentifier this looks like the following:

func init() {
	flows.RegisterStandardFeature("protocolIdentifier", flows.FlowFeature, func() flows.Feature { return &protocolIdentifier{} }, flows.RawPacket)
}

Since protocolIdentifier is one of the iana assigned ipfix features, RegisterStandardFeature can be used, which automatically adds the rest of the ipfix information element specification. The second argument is what this feature implementation returns which in this case is a single value per flow - a FlowFeature. The third argument must be a function that returns a new feature instance. The last argument specifies the input to this features, which is a raw packet. The flows package contains a list of implemented types and Register functions.

For more examples have a look at the provided features.

Common part of sources/filters/labels/exporters

Sources, filters, labels, and exportes must register themselves with the matching Register* function:

func init() {
	packet.RegisterX("name", "short description", newX, helpX)
}

where a name and a short description have to be provideded. The helpX function gets called if the help for this module is invoked and must write the help to os.Stderr. The newX function must parse the given arguments and return a new X. This function must have the following signature:

func newXThing(name string, opts interface{}, args []string) (arguments []string, ret util.Module, err error)

name can be a provided name for the id, but can be empty. opts holds the parameters from a JSON specification or util.UseStringOption if args need to be parsed. args holds the rest of the arguments in case it is a command line invocation. Needed arguments must be parsed from this array and the remaining ones returned (arguments). If successful the created module must be returned as ret - otherwise an error. This function must only parse arguments and prepare the state of the module. Opening files etc. must happen in Init()

All modules must fulfill the util.Module interface which contains an Init and an ID function. ID must return a string for the callgraph representation (most of the time a combination of modulename|parameter). Init will be called during intialization. Side effects like creating files must happen in Init and not during the new function!

Examples of the different modules can be found in the modules directory.

Implementing sources

Sources must implement the packet.Source interface:

ReadPacket() (lt gopacket.LayerType, data []byte, ci gopacket.CaptureInfo, skipped uint64, filtered uint64, err error)
Stop()

ReadPacket gets called for reading the next packet. This function must return the layer type, the raw data of a single packet, capture information, how many packets have been skipped and filtered since the last invocation, or an error.

Stop might be called asynchronously (be careful with races) to stop an ongoing capture. After or during this happening ReadPacket must return io.EOF as error. This function is only called to stop the flow exporter early (e.g. ctrl+c).

data is not kept around by the flow exported which means, the source an reuse the same data buffer for every ReadPacket.

Implementing filters

Filters must implement the packet.Filter interface:

Matches(ci gopacket.CaptureInfo, data []byte) bool

Matches will be called for every packet with the capture info and the raw data as argument. If this function returns false, then the current packet gets filtered out (i.e. processing of this packet stops and the next one is used).

Don't hold on to data! This will be reused for the next packet.

Implementing labels

Labels must implement the packet.Label interface:

GetLabel(packet Buffer) (interface{}, error)

This function can return an arbitrary value as label for the packet (can also be nil for no label). If the label source is empty io.EOF must be returned.

Implementing exporters

Exporters must implement the flow.Exporter interface:

Fields([]string)
Export(Template, []Feature, DateTimeNanoseconds)
Finish()

Fields, Export, Finish will never be called concurrently, are expected to be blocking until finished, and, therefore, don't need to take care about synchronization.

The Fields function gets called before processing starts and provides a list of feature names that will be exported (e.g. the csv exporter uses this to create the csv header).

Export gets called for every record that must be exported. Arguments are a template for this list of features, the actual features values, and an export time.

Finish will be called after all packets and flows have been processed. This function must flush data and wait for this process to finish.