nixos/switch-to-configuration: Document and test handling of socket-activated services

[dasJ]

Motivation for this change

More tests, more docs

The manual should explain what this is doing exactly.

Things done

• Built on platform(s)
  • x86_64-linux
  • aarch64-linux
  • x86_64-darwin
  • aarch64-darwin
• For non-Linux: Is sandbox = true set in nix.conf? (See Nix manual)
• Tested, as applicable:
  • NixOS test(s) (look inside nixos/tests) → switchTest
  • and/or package tests
  • or, for functions and "core" functionality, tests in lib/tests or pkgs/test
  • made sure NixOS tests are linked to the relevant packages
• Tested compilation of all packages that depend on this change using nix-shell -p nixpkgs-review --run "nixpkgs-review rev HEAD". Note: all changes have to be committed, also see nixpkgs-review usage
• Tested basic functionality of all binary files (usually in ./result/bin/)
• 22.05 Release Notes (or backporting 21.11 Release notes)
  • (Package updates) Added a release notes entry if the change is major or breaking
  • (Module updates) Added a release notes entry if the change is significant
  • (Module addition) Added a release notes entry if adding a new NixOS module
  • (Release notes changes) Ran nixos/doc/manual/md-to-db.sh to update generated release notes
• Fits CONTRIBUTING.md.

[ncfavier]

@ofborg test switchTest

[dasJ]

Also cc @aanderse to remind people that you want the script rewritten ;)

[roberth]

Found a flaw 😢 This will break services where ExecStop relies on still having the old etc.

I'm not confident that this can be fixed without fixing systemd or mutating the old units. We really need to inhibit the old service start; otherwise it seems we'll have to introduce new corner cases.

[roberth]

This switches from a non-ExecStop service to an ExecStop one, so it doesn't need to stop the socket yet, but does need to do so on the next switch, even when ExecStop is removed.

[roberth]

This means that despite stopIfChanged being enabled, the old service will run until after etc has been switched.
I don't think that's acceptable.

[dasJ]

@roberth thank you for reviewing the change.
The whole "ExecStop is taken from the new generation" is fixable but I won't do that until we agreed what to do with the other points you mentioned.
Running the service in the new environment is imo acceptable because it's only very short and it increases general service availability.
I do however not know each and every service in NixOS and I don't know if services could suffer from this behavior.

Do you have an alternative suggestion I can implement? If we don't come to an agreement I can redo this PR to only clean the code, add test cases, and add docs to at least have the current behavior properly specified and tested.

[dasJ]

I removed most of the logic after a longer discussion with Robert. We came to the conclusion that there was no way we can keep up with the race-free nature of the current behavior (please correct me if I got your points wrong).
This would be possible if systemd blessed us with some magic way, but as of writing this comment, it's not.

So this PR mostly adds test cases and proper docs for handling of socket-activated units and leaves the current logic intact.

[cheriimoya]

why is this snake case when all the other variables are camel case?

[dasJ]

https://metacpan.org/pod/Perl::Critic::Policy::NamingConventions::Capitalization

[cheriimoya]

makes sense, but what is unitsToReload?

[dasJ]

There is currently a mix between camel case (not yet refactored) and snake case (already refactored)

[cheriimoya]

why not throw it out then?

[dasJ]

To make it clear that this check was left out explicitly and not just forgotten

[cheriimoya]

why aren't we stripping away the spaces from each out variable (or rather in the switch_to_specialisation function)? is it important to check the newline everywhere?

[dasJ]

Because restarting the following units and starting the following units must somehow be differentiated and using regex would be pretty expensive to run hundreds of times

[cheriimoya]

ah, i didn't know out was a string, thought it was an array, my bad

[dasJ]

If there are no more strong opinions, I'd merge this today or so. It's mostly docs and tests and doesn't touch the logic

[roberth]

Just the suggestions from my comment just now. Lgtm!

[stigtsp]

LGTM