Rust testing or verifying: Why not both?

Rust logo Dijkstra famously dissed testing by saying “Program testing can be used to show the presence of bugs, but never to show their absence!” As if you should choose one over the other. I don’t see them as opposites but as complementary techniques that should both be used to improve the quality of your code.

I am a big fan of formal verification. Formal verification tools can be the best bug-finding tools you have ever used. And they can completely eliminate entire classes of bugs from your code. But, formal verification won’t find bugs in your specification; or in your assumptions about your dependencies; or in your build/CI harness, etc. (See Fonseca for more examples of where testing has found bugs in formally verified systems.)

And, we all reluctantly agree that Dijkstra was right: even a thorough, perfectly executed test plan can miss bugs.

So, for the last few months, I have been trying to have both. We (my team and I at Google) have been reimplementing Jason Lingle’s proptest property-testing library for use with Rust formal verification tools. The original proptest lets you write test harnesses to test that your code (probably) satisfies properties. Today, we are releasing a reimplementation of the proptest interface that enables you to use exactly the same test harnesses to formally verify that the properties hold. So far, we have only tried this with the KLEE symbolic execution engine but our implementation is based on the verification interface used in verification competitions so it should be possible to port our library to many other verification tools.1

[Before I go any further, I should mention that what we are releasing this week is a very early research prototype. It is not ready for serious use and it is definitely not an official, supported Google product. We are releasing it now, in its current, immature state because we want to have a conversation about how programmers want to formally verify Rust code and we think that it is helpful to have something to push against. We welcome pull requests that add support for other verifiers, or that push the design in a better direction.]

A proptest test harness

To get an idea for what property testing looks like in proptest, we’ll look at an example from the proptest book.

fn add(a: i32, b: i32) -> i32 {
    a + b
}

proptest! {
    #[test]
    fn test_add(a in 0..1000i32, b in 0..1000i32) {
        let sum = add(a, b);
        assert!(sum >= a);
        assert!(sum >= b);
    }
}

This example defines a property called test_add that tests a function called add. In particular, it checks that, for non-negative values a and b, the result of sum(a, b) is at least as large as a and as b.

The notation 0..1000i32 represents the set of all values in the range [0 .. 1000) and the notation a in 0..1000i32 says that a value a should be chosen from that set.

The proptest! macro converts this property into a test function that repeatedly generates random values for a and b and executes the body of the property.

(After adding in some additional glue code) we can run this example with the command

cargo test

which produces the following output

running 1 test
test test_add ... ok

test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out

and, if we deliberately write a property that doesn’t hold, then cargo test produces output more like this

running 1 test
test test_add ... FAILED

failures:

---- test_add stdout ----
thread 'test_add' panicked at 'assertion failed: sum >= a + 100', src/main.rs:11:9
thread 'test_add' panicked at 'Test failed: assertion failed: sum >= a + 100; minimal failing input: a = 0, b = 0
        successes: 0
        local rejects: 0
        global rejects: 0
', src/main.rs:7:1


failures:
    test_add

test result: FAILED. 0 passed; 1 failed; 0 ignored; 0 measured; 0 filtered out

Verifying with propverify

Proptest checks the above property by generating random values and testing. But we can also interpret the property as saying that for all values a and b in the sets, the body of the property will not panic. That is, we can interpret it as a universally quantified specification.

To verify the above property, we use a script that compiles the Rust code, invokes the KLEE symbolic execution engine and filters the output of KLEE to determine whether there is any choice of a and b that can cause the body of the property to panic.

cargo-verify . --tests

This produces output like this: confirming that the property does hold

Running 1 test(s)
test test_add ... ok

test result: ok. 1 passed; 0 failed
VERIFICATION_RESULT: VERIFIED

And, if we change the example property so that the property does not hold, cargo-verify produces this output.

thread 'test_add' panicked at 'assertion failed: sum >= a + 100', src/main.rs:11:9

running 1 test
  Value a = 0
  Value b = 0
test test_add ... FAILED

failures:

failures:
    test_add

test result: FAILED. 0 passed; 1 failed; 0 ignored; 0 measured; 0 filtered out

On this example, random testing and formal verification produced similar results. I could have chosen an example where formal verification found a bug that random testing misses. Or I could have chosen an example where random testing easily finds a bug but formal verification spins forever. But, in this post, I wanted to focus on the similarity between the two approaches, not the differences.

Learning more

In this article, we used an example that was easy to explain but does not really show off the power of propverify. Some slightly better examples are

  • collections.rs where we check properties involving standard Rust collection types such as vectors and B-trees. (This should be read in conjunction with the proptest documentation.)

  • dynamic.rs where we check properties involving trait objects by quantifying over a subset of the possible instances of the trait.

And, we have more documentation about

The latter two are mostly for the benefit of verification tool developers.

Our next steps

[You can either read this as a sketch of what we plan to do or as an admission of what we have not yet done. As Fred and Ginger said, “Tomato. Tomato.”]

  • The tools have a horrifically complicated set of dependencies.

    This is partly because formal verification is not quite popular enough for enough Debian/Homebrew packages to exist and partly because we need some very specific versions. (It may also be possible to remove some dependencies!)

  • We are in the process of adding Crux-MIR support to the library and tool. Crux-MIR is a new part of Galois’ Software Analysis Workbench (SAW) that verifies the MIR code generated by the Rust compiler.

    This is taking us a bit longer than using KLEE because the functionality of our cargo-verify script overlaps with the functionality of Crux-MIR but they have slightly different approaches – we’re still working on the best way to handle the resulting conflicts.

  • It can be useful to focus our attention on a single crate at a time. We have some ideas for how to do that with fuzzing and verification but we have not had a chance to try them yet.

  • We have not looked seriously at how well this approach scales.

    The collections support in the propverify library has a few tricks to avoid some obvious scaling issues, but we don’t know if those tricks work well for all verification tools and they probably only scratch the surface of what needs to be done.

Summary

Testing and formal verification are usually portrayed as mortal enemies. I think that misses a huge opportunity to use your existing familiarity and comfort with testing to let you get value out of formal verification tools.

I encourage you to download our library and tool, try it out and give us feedback. If you are working on a Rust verification tool, we would love it if you tried to use our library with your tool. And we would love it even more if you sent us a pull request.

It’s probably not a good idea to commit to using the library at this stage: it is not very robust at the moment and I expect that it will change a lot as we gain experience from porting the library to other types of formal verification tools.

Enjoy!


More information

If you found this article interesting, you might also enjoy these related posts

And these papers:

  • The style of test used is a lot like property-based testing that started with claessen:icfp:2000 for Haskell and has since spread to many other languages.

    In strongly-typed languages, property-based testing usually uses the type to determine how to generate values but proptest is inspired by MacIver’s “Hypothesis” for Python maciver:ecoop:2020 that provides a set of fuzzing-constructors (“strategies”) for describing how to construct values and also to control “shrinking” (simplification) of random values.

  • The general approach of using test harnesses with symbolic execution goes back (at least) as far as tillmann:fse:2005 with more recent work by garg:icse:2013 and dimjasevic:ifm:2018.

  • goodman:ndss:2018 applied some very similar ideas to C++ testing in 2018 using the GoogleTest DSL as a starting point and providing support for angr, Manticore and Dr. Fuzz.

    They ran into and solved problems related to logging (causing a path explosion), symbolic loop bounds and symbolic array indices (causing a path explosion) and how to combine swarm testing with verification.

    Highly recommended!


  1. Ok, I have to admit that there are not many verification tools for Rust. Most of our time so far has been spent submitting patches to KLEE so that KLEE could be used to verify Rust programs. 

Written on September 3, 2020.
The opinions expressed are my own views and not my employer's.