Support Windows?

[DirectXMan12]

We may wish to consider supporting Windows at some point

[filmor]

What needs to be done to get there?

[DirectXMan12]

1. Compilation needs to be tested in Windows with MIT Kerberos for Windows on MinGW.
2. The conditional compilation logic needs to be adjusted to work for Windows
3. We need a way to build wheel files for Windows (optional, but useful for anyone who actually wants to use this on Windows).

[frozencemetery]

Other "nice to have" - some way to run our test suite, since I don't think k5test will work there.

[ibigpapa]

So using the MIT KfW. I see the source is available as well as precompiled sources could we just wrap these to make it work?

[DirectXMan12]

yeah, it really comes down to finding time and a Windows box to test on (plus setting up CI to make sure we don't break anything), plus tweaking setup.py to work properly on Windows.

[ibigpapa]

I'm working on trying to hack this in myself let you know how i do.

[DirectXMan12]

cool. Feel free to ping us on #python-gssapi on freenode if you have any questions best answered in a real-time manner (I'm usually on from 10:00-17:00EST)

[frozencemetery]

2a2b47b adds support for conditional compilation. Still missing are building wheels and running buildjob CI. Much like with Heimdal, a stretch goal is to port k5test so we can run the test suite there.

[aiudirog]

I've been working on Windows support the past couple of days to get my company to move off of requests-kerberos (I had previously patched ccs-kerberos to work on Windows) and got everything to compile properly under VS Build Tools, MinGW, and MSYS. Unfortunately I won't be able to test until I'm back at work on Monday since k5test isn't ported yet. My current progress can be found here.

Changes:

• I created a predefined macro (__MSYS__) to allow the header files to detect MinGW under MSYS so that when it isn't under it, they can target MIT Kerberos for Windows instead (which also seems to have better extension support).
• To get it compile under VS:
  • I had to cast all slicing of gss_buffer_desc.values to char* because many platforms seem to declare it as a void* instead and cl.exe doesn't like void* pointer arithmetic. This can be found a separate commit.
  • libraries is now passed separately to Extension like library_dirs so that Extension itself can determine how to pass them to the linker.
• link_args and compile_args are now always lists. This is to support arguments which may have spaces in them, like include paths on Windows.

[aiudirog]

Everything seems to be working as expected. I was able to connect to our internal Apache Livy and Phoenix servers on Windows and use them as normal through requests-gssapi.

[frozencemetery]

That's awesome, well done! Do you have any idea how this could be tested in CI? (Presumably just checking that the build works, as we do with Heimdal, since the krb5 server parts don't run on Windows, but don't let me stop you if you're feeling ambitious.)

[aiudirog]

I've been looking into it and Travis does support Windows now so I'll put some work into at the very least getting it to build and generate wheels.

[aiudirog]

I've opened a pull request to start playing with the Travis build. Hopefully this goes more smoothly than I'm expecting.

[aiudirog]

@frozencemetery I have it building on 2.7, 3.6, and 3.7 on Windows and I think I configured the deployment right, though I'm not sure. This is my first time using Travis (only experience with Jenkins, though I'm liking Travis a lot) so I might have missed something. Is there any way to do a mock deployment to test it?

[frozencemetery]

Either you figured it out, or I'm not understanding the question :) Please post a comment in that PR when it's ready for review.

[aiudirog]

My question was purely about deploying the wheels. Not sure if I set it up right since you can't test deployment from a PR. I guess we can test once it's merged.

[frozencemetery]

@aiudirog So I ran into some problems trying to cut a release for this. I made some fixes - a6e3f08 and 5c80107 - but then I ran into failure trying to use pypi - first second and third subjobs.

It seems to me - and please let me know if this isn't right - that we can't build wheels from Travis right now. That being the case, how would you feel about using Appveyor for this instead? (I think that works...)

[neirbowj]

I just have to say---and I know that this isn't necessarily appropriate for an issue thread---that I really appreciate the work each of you is doing to help achieve this. I wish there were more I could do to help move things along.

[aiudirog]

This seems to be the issue to keep an eye on.

For now, we can probably just manually upload using a script deployment and twine. I'm going to give that a try before redoing the whole thing in Appveyor.