August and September 2021 round-up

One of the hardest tasks with doing a port, whether it’s your first or your tenth, is figuring out the kernel configuration options that needs to be set, and those that should not be set.

With the changes in #410, it is now entirely automated and part of both the normalization step, and the usual build steps.

It means that when you normalize a kernel configuration with the tooling, the options that are required to be set or unset are set or unset accordingly automatically for the developer. Since it is also validated at build time, it means that any changes in the configurations that has not been handled by the maintainer will cause an obvious build failure with appropriate error messages.

Hopefully this means that future ports are this much more trivial.

While working on an upcoming feature requiring further use of the early user interface for Mobile NixOS, it was found that the FFI bindings for them were a dead-end.

These bindings worked seemingly well enough, but had issues. The more easily observable issue was the lack of proper ownership tracking for the FFI object, which meant that anything made through the mruby bindings needed to be tracked manually or else garbage collection would free up resources it shouldn’t have freed.

The workarounds seemed enough. They were not, when tested against more and more complex applications. Further causing issues, the FFI bindings were basically not debuggable. Whenever a crash happened, it crashed way too late to allow debugging the root cause of the issues.

With these issues in mind, I circled back to the approach tentatively used initially, and generated native code bindings for the library. But this time, with a slight twist. I already had a list of desired bindings with the FFI bindings. So I used the same API as the FFI bindings used. In the end, this rewrite needed some light correctness fixes outside of the bindings, but was almost entirely a drop-in replacement.

And this time, it can be debugged. I had to do so to figure out a few teething issues. Additionally, the untraceable segfaults I had with the new usage has simply vanished. It wasn’t as trivial as a garbage collection issue, as disabling the garbage collection didn’t solve any of the issues.

With these exhausting changes, I can finally put my efforts towards actually authoring the new feature!

A somewhat major change is that the default in-repository builds now use a pinned Nixpkgs revision. This should help providing a known working starting point for end-users. The pin will be manually moved as time goes on to known good revisions with good Hydra coverage.

Mobile NixOS outputs are now presented in the outputs attribute. There is a backward compatibility layer added to ensure previous incantations still work, but it will be removed at some point in the future.

Image builds are now faster under some conditions.