Looking at replacing ATF/Kyua (in a limited fashion) with Google Test/shunit2

[Alan Somers]

On Wed, Apr 5, 2017 at 12:44 PM, Ngie Cooper (yaneurabeya)
<yaneurabeya at gmail.com> wrote:
> (Apologies for taking so long to reply; the longer the reply, the longer it takes for me to write a thoughtful response sometimes ;)..)
>
>> On Jan 17, 2017, at 16:09, Craig Rodrigues <rodrigc at freebsd.org> wrote:
>>
>> On Tue, Jan 17, 2017 at 3:01 PM, Ngie Cooper (yaneurabeya) <
>> yaneurabeya at gmail.com> wrote:
>>
>>> Hello all,
>>>
>>>        I had an initial discussion with a handful of other test
>>> stakeholders and due to the number of caveats with ATF/Kyua and a variety
>>> of issues in contributing back to the ATF/Kyua projects (time on
>>> maintainer’s end, legal issues, technical issues, etc), I'm seriously
>>> considering replacing parts of ATF/Kyua with GoogleTest and shunit2. In
>>> particular, I want to do these things [better]:
>>
>> I'll provide my perspective on your proposal.
>> For me, the nice thing about ATF/Kyua was that Julio pushed unit tests to
>> live along-side the code in FreeBSD,
>> and provided a consistent structure by which these tests could run and
>> provide test output.
>> The fact that he was able to modify the Makefile system to get this working
>> in FreeBSD *and* NetBSD
>> is amazing.  I never would have had the energy to do that!!
>>
>> I was hoping that ATF/Kyua would attract interest from projects other than
>> FreeBSD and NetBSD,
>> so that there would be an ecosystem of projects that would use this stuff
>> and drive this forward.
>> I updated the Mac Homebrew port of Kyua a few times, and also tried to
>> submit a Debian package
>> for ATF and Kyua but got stalled there.
>> Unfortunately, ATF/Kyua never really took off outside of FreeBSD and NetBSD.
>
> Yes. That’s a good summary of where we’ve come from.
>
>> Also, the implementation details of modern C++ and Lua for Kyua, and shell
>> script and C for ATF
>> is a high barrier of entry for people wanting to work on the code and
>> extend it.
>> There are enough developers in FreeBSD who could probably pick up the C and
>> ATF parts.
>> For the C++ and Lua parts, I think the barrier is quite high, and there are
>> not that many developers
>> in the FreeBSD community with the skills or interest to work on that.
>
> I’d have to look at how easy divorcing libatf-c from libatf-c++ would be, but I suspect it’s non-trivial. The ATF runners in kyua seem to depend on both libraries (although they can be obtained from ports):
>
> $ ldd /usr/local/tests/kyua/engine/atf_* /usr/local/tests/kyua/bootstrap/atf_helpers  | grep c++
>         libatf-c++.so.2 => /usr/local/lib/libatf-c++.so.2 (0x80117b000)
>         libc++.so.1 => /usr/lib/libc++.so.1 (0x8015ae000)
>         libatf-c++.so.2 => /usr/local/lib/libatf-c++.so.2 (0x80119f000)
>         libc++.so.1 => /usr/lib/libc++.so.1 (0x8015d2000)
>         libatf-c++.so.2 => /usr/local/lib/libatf-c++.so.2 (0x8011a2000)
>         libc++.so.1 => /usr/lib/libc++.so.1 (0x8015d5000)
>         libatf-c++.so.2 => /usr/local/lib/libatf-c++.so.2 (0x8011db000)
>         libc++.so.1 => /usr/lib/libc++.so.1 (0x80160e000)
>         libatf-c++.so.2 => /usr/local/lib/libatf-c++.so.2 (0x800823000)
>         libc++.so.1 => /usr/lib/libc++.so.1 (0x800c56000)

It's probably easier than you think.  Those programs are just Kyua's
own tests, and they aren't needed to run atf-c test programs.  For
example, this atf-c test program doesn't link against
libprivateatf-c++:
$ ldd /usr/tests/sbin/devd/client_test
/usr/tests/sbin/devd/client_test:
        libprivateatf-c.so.1 => /usr/lib/libprivateatf-c.so.1 (0x800822000)
        libc.so.7 => /lib/libc.so.7 (0x800a39000)

>
>> I have a selfish perspective in that I like to invest time in libraries and
>> technologies that
>> help me get jobs, even in companies that are not FreeBSD.
>> In my last job at a company making a FreeBSD appliance, I tried to push
>> adoption of ATF/Kyua
>> but wasn't able to convince folks to pick it up.  The developers gave me
>> feedback that the ATF/Kyua
>> documentation didn't give them confidence.  It didn't help that I couldn't
>> point to other projects
>> outside of FreeBSD/NetBSD using this stuff.
>> At least for me, I cannot say that ATF/Kyua is a marketable skillset which
>> opens doors for jobs. :(
>>
>> Moving forward, if you try to steer towards libraries and technologies that
>> have active
>> user communities *even outside of FreeBSD*, that would be nice.
>> Testing software isn't unique to FreeBSD, and there is a lot of good work
>> going on in these
>> other communities.
>>
>> You mentioned that GoogleTest is actively worked on by Google,
>> but there at last one non-Google community (LLVM) is using it, and has a
>> selfish interest to
>> keep it going.  That is one plus in my opinion.
>
> Bingo. Hence, one of the motivators behind GoogleTest (it has more mindshare in the open source community than ATF/Kyua does).
>
>> GoogleTest uses a 2-clause BSD license, so that is OK.
>> GoogleTest requires signing a Contributor License Agreement for people who
>> want to submit patches upstream to them.  That's no different from Kyua.
>> That's a bit annoying, but maybe something we can live with.
>
> Yes.
>
>> So I have no problem with GoogleTest on those points.
>>
>> My questions for you are:
>>   (1)  GoogleTest is very C++-focused.   ATF has API's for C, C++, and
>> shell.
>>          How successful will you be at testing a pile of C code with a
>> C++-focused library?
>>          It is technically possible (
>> https://meekrosoft.wordpress.com/2009/11/09/unit-testing-c-code-with-the-googletest-framework/
>> )
>>          but what are the gotchas?
>
> That I need to quite frankly investigate. It might come down to the fact that tests need to be written in C++ — which I’m not really keen on, but if the C++ is simple, and the boilerplate is simple enough, then the C++ shouldn’t need to get too fancy, and hence will look at lot like the existing, simple GoogleTest examples.

This, though, is probably harder than you think.  Googletest can
indeed be used to test C code -- sometimes.  But there are numerous C
constructs which are not valid C++.  For example, _Noreturn, static
array sizes, and anything with identifiers that are C++ reserved
words.  But I suspect that the hardest thing to do in Googletest will
be to replicate atf's test case isolation.  It'll be easy to do
per-test program, but not per-testcase.  For that reason, I don't
think we should attempt to convert existing tests, and I don't think
that Googletest is necessarily the best framework for new tests of C
code.

>
>>  (2)  What do we do with existing tests?  Do they need to be rewritten to
>> the GoogleTest API?
>>         Or do we have infrastructure that allows existing ATF tests to run
>> with Google Test?
>
> Same thing as before:
> 1. Write new tests in the preferred target framework.
> 2. Don’t rewrite existing tests, unless there’s a good reason to do so (needless refactoring introduces risk).
>
> As far as the second question (“Or do we…”), yes, my goal was to write equivalent macros, if at all possible, for using GoogleTest in lieu of ATF. The ownership for this task falls on me. I might need to write some compatibility functions/macros to ease the transition away from ATF, if my investigation demonstrates it's feasible/possible.

Why bother obfuscating Googletest?  It's not going to work exactly the
same way as ATF, so users will have to learn the differences anyway.
We may as well allow them to use Googletest as it's intended to be
used.

>
>>  (3)  If this is adopted, are new tests written in GoogleTest API, or in
>> the ATF API?
>
> The GoogleTest APIs.
>
>>  (4)  Kyua is the tool currently used  for running the tests, and
>> providing test result output.
>>         Kyua has test-case isolation, and the potential to parallelize the
>>         execution of test-cases.
>>         Are the Kyua features of test-case isolation and parallelization
>> of test-cases must-haves,
>>         or can we live without them?
>>         Do you want to move away from using Kyua, or modify it to work
>> with GoogleTest?
>
> As a first step, I want to mock ATF via GoogleTest. Although contributing back to Google (as a corporation — thanks $lawyers) has proved an impedance, we can always patch the port to add in the helper.
>
>>         If you want to move away from Kyua, how do you want people to run
>> the GoogleTests and
>>         get the test results?
>
> I really don’t want to change the inputs/outputs if at all possible — just add a bit more “window dressing” to denote setup/teardown fixtures and other gaps that ATF doesn’t have.
>
> If it doesn’t prove fruitful, I’ll send out an update email describing the work I’ve done, the roadblocks/tradeoffs I’ve made in getting to the decision, and ask for some consensus in terms of driving the effort forward, further.
>
>> I have no objection to your proposal.  Using a more mainstream testing
>> framework with a larger user community,
>> and some momentum to keep things going would be nice.
>
> Thanks!
> -Ngie

I like the idea of using Googletest, but I don't think it can
completely displace atf-C.  IMHO, Kyua should have runners for as many
test frameworks as people want, and isolation should be up to the
framework.  Different frameworks can provide different levels of
isolation.  I think that first we should add a Googletest runner to
Kyua, then import Googletest into /contrib and rewrite the paltry
number of atf-c++ tests to use Googletest.  Then we can delete
atf-c++.  If there's demand, we can add runners for other C testing
frameworks (though personally I'm satisfied with atf-c).

BTW, FreeBSD already includes one Googletest program in the base, but
it's not connected to the build because it depends on something from
ports: cddl/usr.sbin/zfsd/tests/zfsd_unittest.cc.

-Alan