lnd/docs/go-fuzz

How to fuzz the Lightning Network Daemon's wire protocol using go-fuzz

This document will describe how to use the fuzz-testing library go-fuzz on the lnd wire protocol.

Introduction

Lnd uses its own wire protocol to send and receive messages of all types. There are 22 different message types, each with their own specific format. If a message is not in the correct format, lnd should logically reject the message and throw an error. But what if it doesn't? What if we could sneakily craft a custom message that could pass all the necessary checks and cause an error to go undetected? Chaos would ensue. However, crafting such a message would require an in-depth understanding of the many different cogs that make the wire protocol tick.

A better solution is fuzz-testing. Fuzz-testing or fuzzing is when a program known as a fuzzer generates many, many inputs to a function or program in an attempt to cause it to crash. Fuzzing is surprisingly effective at finding bugs and a particular fuzzing program AFL is well-known for the amount of bugs it has found with its learned approach. The library we will be using, go-fuzz, is based on AFL and has quite a track record of finding bugs in a diverse set of go programs. go-fuzz takes a coverage-guided approach in an attempt to cover as many code paths as possible on an attack surface. We give go-fuzz real, valid inputs and it will essentially change bits until it achieves a crash! After reading this document, you too may be able to find errors in lnd with go-fuzz!

Setup and Installation

This section will cover setup and installation of go-fuzz.

• First, we must get go-fuzz:

$ go get github.com/dvyukov/go-fuzz/go-fuzz
$ go get github.com/dvyukov/go-fuzz/go-fuzz-build

• Next, create a folder in the lnwire package. You can name it whatever.

$ mkdir lnwire/<folder name here>

• Unzip corpus.tar.gz in the docs/go-fuzz folder and move it to the folder you just made.

$ tar -xzf docs/go-fuzz/corpus.tar.gz
$ mv corpus lnwire/<folder name here>

• Now, move wirefuzz.go to the same folder you just created.

$ mv docs/go-fuzz/wirefuzz.go lnwire/<folder name here>

• Change the package name in wirefuzz.go from wirefuzz to <folder name here>.
• Build the test program - this produces a <folder name here>-fuzz.zip (archive) file.

$ go-fuzz-build github.com/lightningnetwork/lnd/lnwire/<folder name here>

• Now, run go-fuzz!!!

$ go-fuzz -bin=<.zip archive here> -workdir=lnwire/<folder name here>

go-fuzz will print out log lines every couple of seconds. Example output:

2017/09/19 17:44:23 slaves: 8, corpus: 23 (3s ago), crashers: 1, restarts: 1/748, execs: 400690 (16694/sec), cover: 394, uptime: 24s

Corpus is the number of items in the corpus. go-fuzz may add valid inputs to the corpus in an attempt to gain more coverage. Crashers is the number of inputs resulting in a crash. The inputs, and their outputs are logged in: <folder name here>/crashers. go-fuzz also creates a suppressions directory of stacktraces to ignore so that it doesn't create duplicate stacktraces. Cover is a number representing coverage of the program being fuzzed. When I ran this earlier, go-fuzz found two bugs (#310 and #312) within minutes!

Corpus Notes

You may wonder how I made the corpus that you unzipped in the previous step. It's quite simple really. For every message type that lnwire_test.go processed in TestLightningWireProtocol, I logged it (in []byte format) to a .txt file. Within minutes, I had a corpus of valid lnwire messages that I could use with go-fuzz! go-fuzz will alter these valid messages to create the sneakily crafted message that I described in the introduction that manages to bypass validation checks and crash the program. I ran go-fuzz for several hours on the corpus I generated and found two bugs. I believe I have exhausted the current corpus, but there are still perhaps possible malicious inputs that go-fuzz has not yet reached and could reach with a slightly different generated corpus.

Test Harness

If you take a look at the test harness that I used, wirefuzz.go, you will see that it consists of one function: func Fuzz(data []byte) int. go-fuzz requires that each input in the corpus is in []byte format. The test harness is also quite simple. It reads in []byte messages into lnwire.Message objects, serializes them into a buffer, deserializes them back into lnwire.Message objects and asserts their equality. If the pre-serialization and post-deserialization lnwire.Message objects are not equal, the wire protocol has encountered a bug. Wherever a 0 is returned, go-fuzz will ignore that input as it has reached an unimportant code path caused by the parser catching the error. If a 1 is returned, the []byte input was parsed successfully and the two lnwire.Message objects were indeed equal. This []byte input is then added to the corpus as a valid message. If a panic is reached, serialization or deserialization failed and go-fuzz may have found a bug.

Conclusion

Fuzzing is a powerful and quick way to find bugs in programs that works especially well with protocols where there is a strict format with validation rules. Fuzzing is important as an automated security tool and can find real bugs in real-world software. The fuzzing of lnd is by no means complete and there exist probably many more bugs in the software that may go undetected if left unfuzzed. Citizens, do your part and go-fuzz lnd today!