cargo: Generate a Cargo.lock if one wasn't found

[nanonyme]

Spec from https://gitlab.com/BuildStream/bst-plugins-experimental/-/issues/42

Will behave like this:

• track: Cargo.lock will be generated by the host's cargo IF (generate_lock is true) OR (generate_lock is 'if-missing' and Cargo.lock doesn't exist). If generate_lock is false and Cargo.lock doesn't exist, raise an error. Then the existing logic for tracking cargo sources will happen. The ref will now encode everything necessary to reconstruct the vendored libraries, so the generated Cargo.lock is discarded
• fetch: The sources are fetched according to the ref, as they are now
• stage: The sources are staged according to the ref, as they are now
• Buildstream will default to generate_lock: false, and projects can (if they wish) globally override that default to generate_lock: if-missing to make the cargo plugin work in all scenarios. Then generate_lock: true can be used if Cargo.lock needs to be overwritten/ignored for whatever reason (i.e. a hypothetical repo that hasn't been updated in a while and Cargo.lock specifies dependencies w/ known security issues, so you want to force it to use newer libraries)
• If a project doesn't want to grow a dependency on cargo from the host to track, it can keep generate_lock: false as the default, and then maintain its own Cargo.lock files & stage them before the cargo source:

sources:
- kind: git_tag
  url: whatever:foobar.git
- kind: local
  path: files/cargo/foobar/Cargo.lock
- kind: cargo
  generate_lock: false # <-- from project defaults

[nanonyme]

Another option is to create separate plugin cargo-lock-generator. Then you have

• git
• cargo-lock-generator
• cargo

and cargo can consume lock as normal. cargo-lock-generator only generates cargo.lock when tracking but otherwise does nothing. It shall depend on previous source both for track and fetch.

[nanonyme]

@gtristan any opinions?

[nanonyme]

Creating new plugin does not change API or cache keys. Adding functionality to existing plugin does not necessarily need cache key change since the new functionality kicks in only with track.

[AdrianVovk]

cargo-lock-generator only generates cargo.lock when tracking but otherwise does nothing

This is rather ugly, in my experience.

Here's how it ends up working out:

• if we generate the Cargo.lock file only during tracking, this ends up w/ a broken project for people who haven't tracked that element. Basically anyone who wants to be able to just clone & build will be unsuccessful because, simply, the Cargo.lock file doesn't exist!
• if we generate the Cargo.lock file during both tracking and fetching, the problem is that we cannot reliably generate an identical Cargo.lock file. Basically this would turn the fetching step into tracking, which breaks Buildstream again in a different way
• The only solution is to store the Cargo.lock file in the project ref. This is very ugly

Building this into the cargo plugin is optimal because it alleviates the need for the Cargo.lock file if not tracking

Edit: Actually I suppose that Cargo.lock only existing during/after a track can be fine if the cargo plugin is willing to handle that. I seem to recall it blowing up for whatever reason.

[nanonyme]

I don't see any reason why cargo plugin would need Cargo.lock after tracking. It generates things to download based on it during tracking and records them in ref.

[AdrianVovk]

An issue with the separate plugin: We have to actually store the Cargo.lock file and then stage it, so that the cargo plugin can pick it up and use it. But this means that later, during the element's build, we actually place a Cargo.lock file into the sources. On a separate build machine, we would instead build with the Cargo.lock file missing (since it isn't really necessary). However, this behavior introduces a different build environment without changing the cache keys!

[nanonyme]

IMHO any generated cargo lock must be the ref of said plugin.

[AdrianVovk]

So you mean that the generated cargo.lock file should be base64 encoded and then stored in the ref?

That generates a massive ref. I don't mind since it lives in project.refs, but it it ends up being necessary in freedesktop-sdk you'll end up with loads of chunk in your element

[nanonyme]

I don't think it's that big a price to pay for determinism to store this information in ref. Also, freedesktop-sdk is not really needing this functionality in the first place as all the projects we use have Cargo.lock. This is more aimed for various downstreams and developers.

[AdrianVovk]

store this information in ref

It's already stored in the ref, at track time, by the cargo plugin.

Cargo.lock isn't necessary during the actual build, is it? I'm actually not sure now...

for determinism

Simply not including a Cargo.lock file isn't less deterministic than storing it in project.refs

Also, freedesktop-sdk is not really needing this functionality

For now...

[nanonyme]

Oh, never mind, I misread your point. Ok, I guess I'm convinced with adding it to cargo plugin itself as long as this does not add dependency by default. I don't think you even need to change add generate_lock into cache keys since it's for tracking only here so this should be doable as a mostly backwards-compat change.

[nanonyme]

@AdrianVovk freedesktop-sdk has been cooperating with any upstreams not shipping Cargo.lock and convincing them to do that. We plan to keep this strategy. Not shipping Cargo.lock with releases is a basically a bug.

[AdrianVovk]

freedesktop-sdk has been cooperating with any upstreams not shipping Cargo.lock and convincing them to do that

Some upstreams feel strongly about not shipping Cargo.lock... So unfortunately there's no escaping the need to generate it sometimes

i.e. systemd/zram-generator#65

[nanonyme]

Yeah, I guess we live in an imperfect world where people sometimes do not follow best practices. Anyway, point stands. What that project is doing is not a general thing for people to do. I am not saying that Cargo.lock should always exist, it is very common for it not to. However, when generating tags, it usually does. You can even have same commit retagged multiple times with different Cargo.lock files if you so want. Once you have tested that this combination actually builds.

[nanonyme]

* if we generate the Cargo.lock file only during tracking, this ends up w/ a broken project for people who haven't tracked that element. Basically anyone who wants to be able to just clone & build will be unsuccessful because, simply, the Cargo.lock file doesn't exist!

This is essentially the thing which made me believe you claimed Cargo.lock needs to exist during build. It is perfectly fine to expect an element has to be tracked at least once and outputs committed into git for it to be buildable. However, refs stored in git must be sufficient for instrumenting source fetching and building.