change from ninja and custom cmake target run_test to using cmake & ctest.

[Martyrshot]

change from ninja and custom cmake target run_test to using cmake & ctest.

Upon reading issue #1494 I figured I'd give this a go.
This PR aims to remove the need for ninja, and make tests platform agnostic.

The new build process will work as follows:

cmake -B build
cmake --build build --parallel
ctest --test-dir build/tests --output-on-failure

Since the run_tests rule was removed, and cmake doesn't have a way to skip tests by default,
the OQS_ENABLE_LONG_TESTS variable is introduced. If you would like to run the long tests
(currently only test_kat_all) you would do the following:

cmake -B build -DOQS_ENABLE_LONG_TESTS=on
cmake --build build --parallel
ctest --test-dir build/tests --output-on-failure

[baentsch]

Yikes -- we still suggest using OpenSSL1.1??

[Martyrshot]

oof. Good catch. I'll make another pr updating that to openssl@3.

[baentsch]

Thanks @Martyrshot for this improvement. I like the idea -- but I think quite a bit of our CI machinery depends on the (now) "old" way of selecting tests: That'd surely needs to be changed over, too.

[Martyrshot]

Looking at the liboqs ci (I haven't looked at downstream projects) the majority of ci calls pytest directly.

I was trying to avoid having to re-write all the tests in order to remove pytest, so as a nice side benefit ci should work. In a perfect world we would switch from pytest so everything is using a single testing platform... But I wanted to get initial feedback before I committed to that whole mess. :-)

[baentsch]

But I wanted to get initial feedback before I committed to that whole mess. :-)

My support (for resolving that mess :-) you have.

FWIW, the reason the latest oqs-provider release test fails is a combination of pytest fixture inflexibility (or misunderstanding of "that mess" on my side) and openssl crashing when dealing with too many KEXs :-(

[Martyrshot]

Is the inflexibility you're referring to being that it tests all of the algorithms? Or are the other issues that I can look at addressing when I remove pytest?

[baentsch]

Is the inflexibility you're referring to being that it tests all of the algorithms?

Yes and No: The "inflexibility" I mean is that the test fixture for server startup only parameterizes on the sigalgs (not on the KEMs) and thus has to enable all KEMs to prepare/allow for the subsequent (all-)KEM tests (against each sigalg test server). This is conceptually fine and logical (why create different servers for different KEMs?). Unfortunately, openssl obviously cannot handle more than 44 default groups without crashing (still need to understand why and when) -- and oqsprovider now has 52...

Add/edit: All that said, it just dawns on me that I can use OpenSSL's config mechanism (via env var) activating only the KEM the client wants to test... But that then requires each test to fire up both server and client for the SIG/KEM combination under test (right now, I'd hope only count(SIG) openssl s_server instances get started -- but then again, I don't really understand the pytest mechanism :-(

[Martyrshot]

Ah, ok. I understand. Well, I'll see what I can do.

[Martyrshot]

Add/edit: All that said, it just dawns on me that I can use OpenSSL's config mechanism (via env var) activating only the KEM the client wants to test... But that then requires each test to fire up both server and client for the SIG/KEM combination under test (right now, I'd hope only count(SIG) openssl s_server instances get started -- but then again, I don't really understand the pytest mechanism :-(

IIRC you can specify multiple algorithms at once. So, although not perfect, you could minimize the number of server/client spin ups to COUNT(algs)/44 ~= 2.

Anyway, I'm going to focus on removing pytest from liboqs. This will probably turn into converting the pytest scripts into vanilla python scripts and having ctest call them. I feel like this is more portable than writing bash scripts, but also far less painful then converting them to c binaries that need to be compiled, and that exec the test binaries with the right arguments...

[Martyrshot]

Current tasks/conversations to have after yesterday's call:

Investigate if we can still generate vscode projects with out build system. If not, do we care?

• I will look into this and report back later this week/early next week.

Do we want to keep pytest as the underlying test framework and just have cmake sit on top of it?

• In the meeting @dstebila showed support for keeping pytest for now. At the very least we can do the change incrementally. @baentsch's comments seem to prefer removing pytest at the same time as switching to cmake. I am open to either, but I think converting from pytest is going to be a decently large task on its own, so it might be worth doing in a separate PR.

[baentsch]

converting from pytest is going to be a decently large task on its own

It definitely is. Particularly if the goal is to obtain the same level of performance/parallelization as afforded by pytest-xdist...

[baentsch]

Can I ask whether this PR can be shelved (closed) for the time being, @Martyrshot ?

[Martyrshot]

Can I ask whether this PR can be shelved (closed) for the time being, @Martyrshot ?

Sorry I haven't had much time to work on personal projects. :( If others think this PR still makes sense then I can look into wrapping it up. But if there is no demand for it, then we can close it.

[baentsch]

If others think this PR still makes sense then I can look into wrapping it up. But if there is no demand for it, then we can close it.

The one big benefit I see this PR would bring is much better testing (coverage) on non-UNIX'ish platforms. But as I don't see strong demand for liboqs there I'd think it'd be OK to close this -- maybe create an issue pointing to this instead for someone (or yourself) to pick it up as time/renewed (or external) interest permits/allows.