Verifying concurrent, crash-safe systems with Perennial

Perennial is a system for verifying correctness for systems with both concurrency and crash-safety requirements, including recovery procedures. For example, think of file systems, concurrent write-ahead logging like Linux's jbd2 layer, and persistent key-value stores like RocksDB.

Perennial uses Goose to enable verification of programs written in (a subset of) Go.

This repository also includes a proof of correctness for GoJournal, the verified journaling system used in go-nfsd, including a simple NFS server built on top.

Compiling

We develop Perennial using Coq master and maintain compatibility with Coq 8.15. If CI (updated dependencies) is broken above Perennial should still compile but is currently incompatible with an upstream change to one of our dependencies. We try to avoid this situation.

This project uses git submodules to include several dependencies. You should run git submodule update --init --recursive to set that up.

To compile just run make with Coq on your $PATH.

We compile with coqc.py, a Python wrapper around coqc to get timing information. The wrapper requires Python3 and the argparse library. You can also compile without timing information with make TIMED=false.

Compilation times

Perennial takes about 120 CPU minutes to compile. Compiling in parallel with make -j4 is significantly faster, and can cut the time down to 45-50 minutes.

Incremental builds are better, after Iris and some core libraries are compiled.

When you make a change to a dependency, you can keep working without fully compiling the dependency by compiling vos interface files, which skips proofs. The simplest way to do this is just to run make vos, but it's fastest to pass a specific target, like make src/program_proof/wal/proof.required_vos, which only builds the vos dependencies to work on the wal/proof.v file.

If you're working on Goose and only need to re-check the Goose output, you can run make interpreter to run the interpreter's semantics tests, or directly compile just the Goose output files.

Coq also has a feature called vok files, where coqc compiles a vos file without requiring its dependencies to be built. The process does not produce a self-contained vo file, but emits an empty vok file to record that checking is complete. This allows checking individual files completely and in parallel. Using vos and vok files can significantly speed up the edit-compile-debug cycle. Note that technically vok checking isn't the same as regular compilation - it doesn't check universe constraints in the same way.

Updating Goose output

This repo has committed versions of the output of Goose, to avoid making Go and Goose a dependency for compilation. You can update these using the ./etc/update-goose.py script, which records exactly how to generate the output for the various Goose projects we have. Use ./etc/update-goose.py --help to get all the options. The script only translates the projects you pass a path to.

Source organization

src/

• program_logic/ The main library that implements the crash safety reasoning in Perennial. This includes crash weakest preconditions, crash invariants, idempotence, and crash refinement reasoning.

• goose_lang/ A lambda calculus with memory, concurrency, and an "FFI" to some external world, which can be instantiated to provide system calls specified using a relation over states.

  This language is the target of Goose and thus models Go and also implements the Iris language and Perennial crash_lang interfaces for proofs using Perennial.

  This directory includes a lifting to ghost state that supports more standard points-to facts, compared to the semantics which specifies transitions over the entire state. It also proves Hoare triples using these resources for the primitives of the language.

  • typing.v A type system for GooseLang. This type system is used as part of the typed_mem/ machinery. TODO: there's much more to say here.

  • lib/ GooseLang is partly a shallow embedding - many features of Go are implemented as implementations. These features are divided into several libraries. Each library has an implementation and a proof file. For example, map/impl.v implements operations maps using GooseLang's sums while map/map.v proves Hoare triples for the implementation. Separating the implementation allows us to run the implementation in the GooseLang interpreter without compiling the proofs.

    • typed_mem/ Implements support for flattening products over several contiguous locations, which is the foundation for supporting struct fields as independent entities. The proofs build up reasoning about the l ↦[t] v assertion, which says l points to a value v of type t (in the GooseLang type system). If t is a composite type, this assertion owns multiple physical locations.
    • struct/ Implements struct support. Structs are essentially tuples with names for the fields. The theorems proven here culminate in a way to split a typed points-to for a struct into its individual fields.
    • slice/ Implements Go slices using a tuple of an array, length, and capacity.
    • map/ Implements Go maps as a linked list of key-value pairs, terminating in a default value.
    • loop/ Implements a combinator for loops on top of the basic recursion support.
    • lock/ Implements locks and condition variables using a spin lock, which is implemented using CmpXchg.
    • encoding/ Implements uint64 and uint32 little-endian encoding.

  • examples/ The Goose unit tests; these are auto-generated from the Goose repo, from internal/examples/.

  • interpreter/ An interpreter for sequential GooseLang, equipped with a proof of partial correctness: if the interpreter produces a result, the semantics also can.

    This is used to implement tests in generated_test.v. These tests are Go functions which return booleans that should be true; we check this using Go test, and compare against the interpreter's behavior.

  • ffi/

    Two different FFIs to plug into GooseLang - disk.v is the one we actually use, while append_log.v is the FFI-based specification for the append_log example.

• program_proof/

  The proofs about programs we have so far.

  • append_log_proof.v Hoare triples about the append_log example, which is implemented in the Goose repo at internal/examples/append_log/.

  • examples/ Examples we wrote for POPL

  • wal/, txn/, and buftxn/ proof of the transaction system library in goose-nfsd

  • simple/ proof of a simple NFS server

• Helpers/

  • Integers.v Wrapper around coqutil's word library for u64, u32, and u8.
  • Transitions.v A library for writing relations in a monadic, combinator style.
  • NamedProps.v An Iris library for naming hypotheses within definitions and using them to automatically destruct propositions.
  • other files are basically standard library extensions

• algebra/

  Additional CMRAs for Iris ghost variables

It's also worth noting that external/Goose contains committed copies of the Goose output on some real code we have. This includes github.com/tchajed/marshal and github.com/mit-pdos/goose-nfsd. The directory structure here mirrors the way Go import paths work.

Publications

Perennial 1 is described in our SOSP paper, "Verifying concurrent, crash-safe systems with Perennial". The actual codebase was quite different at the time of this paper; it notably used a shallow embedding of Goose and did not have WPCs or any of the associated program logic infrastructure. See the tag sosp2019 or the shallow branch.

Goose is briefly described in a CoqPL extended abstract and associated talk, "Verifying concurrent Go code in Coq with Goose".

The verified interpreter and test framework for Goose is described in Sydney Gibson's masters thesis, "Waddle: A proven interpreter and test framework for a subset of the Go semantics".

The proof of GoJournal's correctness is described in the OSDI paper, "GoJournal: a verified, concurrent, crash-safe journaling system". The framework has evolved in several ways since then. See the tag osdi21 for the version used there.