Turing Completeness and Smart Contract Security
At Kadena, we talk a lot about blockchain design decisions. One particularly profound design decision that I’ve come across recently revolves around the property of Turing completeness, which sparked debate in a recent post by one of our developers. As a non-technical member of the team, I wanted to learn what Turing completeness actually means, why the blockchain community seems to be divided on the matter, and why our engineers are working hard to educate others on its drawbacks for safe smart contract programming.
What follows is a brief summary of my exploration to demystify and understand this core property of smart contract security. So if you’re somebody like me who is captivated by the development of blockchain systems, and particularly smart contract languages, then hopefully this story will improve your understanding of how Turing completeness impacts smart contract security.
The Turing Machine Leads to Turing Completeness
In the 1930’s, Alan Turing built the concept of a Universal Turing Machine. Note that the Turing machine is different from the device used to solve the Enigma code during WWII. That was called the Bombe machine. The Turing machine is actually a theoretical device, a mathematical model of computation that describes the most powerful mechanical computer that we know how to build to this day.
So what is the relationship between the Universal Turing Machine and smart contracts?
It comes down to something called “Turing completeness,” which is a property used to describe programming languages that can be used to simulate a Turing machine. In fact, the vast majority of programming languages that exist are Turing complete. The most well-known example of a Turing complete smart contract machine is the Ethereum Virtual Machine (EVM).
“A programming language that is Turing complete is theoretically capable of expressing all tasks accomplishable by computers; nearly all programming languages are Turing complete if the limitations of finite memory are ignored.”
-Wikipedia, Turing Machine
So what does this actually mean in practice? Is it good or bad for a programming language to be able to express every computable algorithm?
At first, this sounds like a great and powerful property. There are certainly benefits to a smart contract language being universally applicable and flexible for building any type of application. Let us consider the implications.
Turing Completeness Applied to Blockchain Smart Contracts
Unfortunately, I am not a thought leader in the field of smart contract design, so I turned to the internet for guidance. Preliminary research returned wholly unhelpful answers from StackOverflow like this or in-depth papers on Turing machines like this. Neither of which examine the property in question through the context of programming smart contracts.
Fortunately, I had the privilege of asking a highly qualified and leading resource on the subject, Emily Pillmore who is a senior programmer with Kadena. I asked her, “What is the difference between Turing completeness and Turing incompleteness in the context of blockchain?” Her response:
“Turing incompleteness doesn’t really have much to do directly with blockchains. It is a property of abstract rewrite systems in computer science that ensures that all expressions in the system can be expressed in a normal form (i.e. can be reduced)… Turing incompleteness manifests as expressions in a language which are irreducible.”
And that’s only a fraction of the full answer! The remainder of the response is equally impenetrable to someone like me. I asked a “How-does-this-thing-work” question and got a comprehensive and technical answer. Fair enough. So I thought I’d come at the problem from a different angle. Maybe if I ask a “What-can-this-thing-do” question, then I’ll get a more concrete picture of its behaviors.
I asked, “Are there examples of things you can do with Turing completeness that would be useful on the blockchain?” Emily’s answer:
“No. I cannot come up with a single use-case for the blockchain that requires Turing completeness. The EVM does not use any of the properties of Turing completeness because it restricts all recursion via the gas model anyway, which enforces that recursion either terminate prior to gas running out, or terminate the program when gas runs out. So in effect, the EVM’s gas model sort of simulates Turing incompleteness, but not really. This is why I say they adopt all the flaws of Turing completeness (side effecting, illegible and unreasonable code, arbitrary looping), but use none of its benefits (no infinite recursion allowed).”
Aha, this I can follow! When you have the power to do more things, it also means that more things can go wrong. Turing completeness is inherently more powerful, but if you can’t utilize the benefits of that power, then you’re needlessly exposing yourself to more attack vectors due to the increased surface area that comes with so many additional features.
To further clarify the distinguishing characteristic:
- Turing complete languages always have a form of conditional repetition or conditional jump (while, for, goto)
- By their design, blockchains will always stop these endless loops with mechanisms like gas.
- Therefore, Turing completeness carries with it an unnecessary and burdensome attack surface. It is high risk, with low reward, as illustrated below.
At the time of publishing, there are 16 known attack vectors on Solidity, the most popular domain-specific programming language for smart contracts, which happens to be Turing complete. If Solidity were Turing incomplete, then the most costly of these attacks would not even be possible. Most notably, the DAO attack in 2016 — a reentrancy attack, enabled by Turing completeness — drained 3.6M ETH (~$50M USD) from the fund, resulting in the Ethereum community hard-forking the entire network to restore their preferred version of history. More recently in 2019, there was a pause on the Constantinople upgrade as it introduced a new form of reentrancy attacks.
When designing a risk-averse technology like smart contracts, it is best to avoid exposing developers and users to an unnecessary amount of potential risks just so they might be able to perhaps someday come up with a use case to enjoy its perceived benefits. After all, smart contracts are nothing more than computer programs that run on a blockchain. They do what they’re told to do. In using them, you are exposing yourself to the risk of trusting someone else’s sound logic for all possible inputs. Instead of trusting an unknown programmer’s logic, consider trusting a Turing incomplete programming language with which several dangerous bugs & exploits are not even possible. For even more safety consider trusting a Turing incomplete language designed with smart contract security in mind.
As I continue to learn about this subject, I would love to hear whether you think there are key implications of Turing completeness that I failed to address. Have you found benefits from using a Turing complete language in your smart contracts? Share your thoughts below!