WIP: Add fuzzer tests

[Kobzol]

Hi, I wanted to experiment with fuzzing in Rust, so I thought that I would start with this crate, since you have suggested that some fuzz tests would be nice (https://github.com/jonhoo/inferno/issues/63).

I simply added cargo fuzz and created a single fuzz target for each Collapse implemetation. Since the trait is so simple, this should be enough. Is that ok?

We should also define what do you actually consider a bug in the library and what do you consider a "bug" in the input. For example immediately after I started fuzzing perf, it crashed, because it fails on an assertion (https://github.com/jonhoo/inferno/blob/master/src/collapse/perf.rs#L173) if the input (file) is empty. For cmd usage this is probably not that much different from a controlled notification about a bad input with a Result, but for a library usage a crash like that on an invalid input is probably not something that you want. If I comment out the assert, the fuzzer doesn't seem to find any other crash in perf, so maybe the assert is superfluous?

I didn't run the fuzzers for long, they also found one other crash in dtrace (related to overflow in subtraction).

Cargo fuzz reguires nightly though, so the fuzz tests would have to be executed in CI using nightly.

[jonhoo]

That's awesome, thank you!

I think it would probably be better to return errors about malformed input in those cases rather than rely on assertions (cc @jasonrhansen), which would also let you fuzz better.

Running fuzzing only on nightly seems fine to me!

[jasonrhansen]

As for the assertion in perf.rs, we could replace it with something like this:

        if self.event_filter.is_none() {
            return Err(io::Error::new(
                io::ErrorKind::InvalidData,
                "no event filter found while processing input",
            ));
        }

[jasonrhansen]

In dtrace.rs maybe we should check if the stack is empty at the beginning of on_stack_end and log a warning with an early return? So something like this:

        if self.stack.is_empty() {
            warn!("on_stack_end called with an empty stack");
            return;
        }

[Kobzol]

@jonhoo Do you want to run fuzzing on CI? Cirrus CI is only used for FreeBSD, the rest is tested on Azure, which uses some predefined templates?

I could add cargo +nightly install cargo-fuzz && cargo +nightly fuzz run <targets> to Cirrus, but I'm not sure how to modify the Azure templates :)

[jonhoo]

I think we probably only need to run fuzzing on one of the CIs. Probably Azure. What you'll want to do is add another "stage" before the empty line before the line that reads resources:

 - stage: fuzz
   displayName: fuzzing tests
   jobs:
     - job: fuzz
       displayName: cargo +nightly fuzz
       pool:
         vmImage: ubuntu-16.04
       steps:
        - template: azure/install-rust.yml@templates
          parameters:
            rust: nightly
        - checkout: self
          submodules: recursive
        - script: cargo install cargo-fuzz
          displayName: Install cargo fuzz
        - script: cargo fuzz run <targets>
          displayName: Run fuzzer

[Kobzol]

Is Azure currently integrated into this GitHub (like for example in tokio) or do you only see the CI results privately? We have to choose some timeout for the fuzzing and maybe some conditions for the fuzzing stage (for example only for merge commits).

[jonhoo]

Looks like it was misconfigured in some way — should be fixed now! You may want to merge with master.

[Kobzol]

Ok, I rebased onto master. However there is a problem with clippy that will possibly require an update of crossbeam-channel.

[jonhoo]

Ah, yes, that seems to be https://github.com/crossbeam-rs/crossbeam/issues/404. I've pushed changes to master that should make things work again. Try merging again?

[jonhoo]

Looks like coverage sometimes crashes due to https://github.com/xd009642/tarpaulin/issues/190. Since your fuzzing stage now comes after coverage testing, it might not even run if that triggers. Try adding the following to the new stage you added:

   dependsOn: test

[jonhoo]

Now we're getting somewhere! At least one fuzz failure: https://dev.azure.com/jonhoo/jonhoo/_build/results?buildId=179&view=logs&s=921c3b4a-2264-5a5b-3600-c937c6937e79&j=4181b88e-0661-5860-37d0-7e60733b20ba

[Kobzol]

Cool :) Do you have access to the artifact directory? The test was written to /home/vsts/work/1/s/fuzz/artifacts/perf/crash-da39a3ee5e6b4b0d3255bfef95601890afd80709, but it probably has to be picked up manually. Maybe it would be possible to recreate the same input locally by running the fuzzer with the same random seed, but I didn't find such option in the documentation of cargo fuzz nor libfuzzer (it expects the generated corpus, but that is also on the Azure filesystem).

[jonhoo]

Hmm, I think for that you need to add another step at the end that publishes a Pipeline Artifact:

- publish: $(System.DefaultWorkingDirectory)/fuzz/artifacts/
  condition: failed()
  artifact: fuzz

I think that step should still execute even if the previous once failed (due to the condition), but not entirely sure. I also don't know where to download that artifact from once it's been published, though the docs seem to hint at it being possible.

[Kobzol]

Uhh, the terminology here is atrocious, multiple types of artifacts.. in GitLab it's pretty easy, in Azure I'm a bit lost :-) I think that you might be able to click on the pipeline run and then it shows Artifacts: -, maybe it will become clickable if there are any build artifacts exported.

I half expect that the failed() condition will only fire when the immediate predecessor step fails (let's hope not :-) ). Let's see.

[jonhoo]

Huh, that seems to not even have triggered a build...? :/