User control over cargo warnings

[epage]

Problem

Today, cargo maintainers are very cautious as to what warnings are created because they cannot be turned off by the user when they are intended. The problem is we've not had a mechanism to control it. With #12115, we'll now have a mechanism

Turning existing warnings into lints:

• [patch] in config.toml triggers unhelpful warnings #11782
• Publish size warnings (Bundle local images in rustdoc output rfcs#3397 (comment))
• The "unused manifest key" warning doesn't print or halt the publishing with cargo publish #7380
• Control the warning from Warn when an edition 2021 crate is in a virtual workspace with default resolver #10910
• Allow turning Cargo.toml warnings into errors #2568
• warning: doc tests are not supported for crate type(s) cdylib even with explicit doctest = false #15964

Porting clippy lints to cargo

• Expand manifest validation for all settings #4377 (see also clippy::cargo_common_metadata)
• Lint against non semver-open dependencies #5340 (see also clippy::wildcard_dependencies, deny::wildcard)
• Warn when new duplicate dependencies are introduced #13899 (see also clippy::multiple_crate_versions, deny::duplicate)
• Lint for negative feature names #15581
• Lint for redundant feature names #15580

Examples of new lints:

• cargo package should warn binary files by default #4413 (deny::detected-executable, deny::detected-executable-script)
• cargo publish should warn on invalid categories/keywords #4300
• Warning when large binary files are included into the bundle #9058
• Warn on old build-directive syntax (Cannot extend build directives within cargo #11461)
• Warn in more cases than Warn when an edition 2021 crate is in a virtual workspace with default resolver #10910
• Maybe "cargo publish" should error if there are path overrides #2489
• Disable warning from path dependencies dependencies warning control, overriding path heuristic #8546
• Warn on mixed case package names (Publishing a crate is case-sensitive #6043)
• deprecate or warn about downloads over unencrypted HTTP #7081
• SPDX dual-license inconsistency #2039 (see also Restriction Lint: Crate License rust-clippy#3867, deny's licenses)
• Cargo always rebuildsd if rerun-if-changed=PATH and PATH points to none-existing file #6003
• Add a warning when enabling LTO for crates with both cargo-types that support LTO and ones that don't. #4611
• Cargo will silently ignore invalid .cargo/config options #5890
• add a warning when compiling a user crate named core #7760
• Cargo allows you to have a different name in [lib] and [package] when publishing #6827
• Redundant package.homepage (see Add guidance on setting homepage in manifest #13293)
• Lint if workspace dependency inheritance isn't used #15578
• Lint if workspace lints are available but not inherited #15579
• non-latest edition
• include patterns should report invalid paths #14390
• warn if two dep info files within the same package point to the same source file (Spurious "error[E0583]: file not found for module" error. #7730).
• deny missing metadata when running cargo package cargo package (and publish) includes target folder #15398 (comment)

Proposed Solution

As package and workspace level lint control, turning existing warnings into diagnostics where possible

Tasks

• rustc-like diagnostics for parsing Cargo.toml (feat: Add rustc style errors for manifest parsing #13172)
• Report what set the lint level (see Unused dependencies cleanup #13778 for context)
• im_a_teapot lint should be unstable (see Cleanup linting system #13797 for context) (implemented in Error when unstable lints are specified but not enabled #13805)
• Clean up forbid handling for default/edition level (see Cleanup linting system #13797 for context)
• Documentation of all supported lints similar to rustc and clippy (see Add tooling to document lints #14025)
• Needs a stable lint at the time of stabilization
  • Add text-direction-codepoint-in-comment lint #16374, Add text-direction-codepoint-in-literal lint #16373
  • Lint for redundant package.homepage #15871
  • Lint that binary names should be kebab case #15586
  • Lint if workspace lints are available but not inherited #15579
• Warning, error count
  • Report as part of existing summaries?
• build.warnings control (feat(diag): Support build.warnings for cargo lints #17112)
• report all errors, not just first, in parse pass (fix(diag): Report all errors, in order #17095)
• Which sub-diagnostic the lint level reason goes into?
• Audit unwrap
• Make paths relative in the same way as rustc diagnostics (fix(diag): Give diagnostics the same display path behavior as rustc #17101)
• Incorrect suggestion-diff for lint unused_workspace_dependencies #16982
• Lints for unstable features (made difficult by lack of runtime iteration of -Z) (what should be the behavior of lints for runtime controlled unstable features like hint-mostly-unused, always run or only with -Z)
• Document writing of lints' (docs(diag): Provide jumping off points for writing diagnostics and passes #17090)
  • First group is primary, all others are secondary
  • Emitted reason is last element of primary group
  • Emitted reason is only emitted once per package
  • (TBD) Helps without snippets may be elements of the primary group, otherwise make them secondary groups
    • Never do a help as an annotation's label
  • When OR-ing of help messages, use a common prefix to help clarify that for users
  • If can't provide a Snippet (e.g. in-memory manifests), at least provide an Origin
  • Follow rustc's dev guide on grammar, etc
  • Rules on padding?
• Deterministic order of linting packages
• How to handle false positives from unused_dependencies (feat(lints): Emit unused_dependencies lint #16600). Started with an ignored = ["dep_name], then ignoring direct & transitive (fix(compile): Ignore unused deps if also transitive  #16935, Revert "feat(lints): Add unused deps ignore list" #16937)

Non-blocking

• rustc-like diagnostics for parsing Cargo.lock
• rustc-like diagnostics for parsing .cargo/config.toml
• Deciding how to handle edition compatibility groups and groups in general (since refactor: Remove rust_2024_compatibility lint group #13740 skipped it for 2024)
  • Should we reuse rustc groups or not?
• rustc-like diagnostics for processing Cargo.toml (ie port warnings off of shell.warn for consistency)
• Diagnostic summary
• Paths are not relative to the same root as rustc's

Rejected

• _::warnings dynamic group (rust or cargo?): specialized warning for specific workflows that is being superseded by other mechanisms

Open questions

• How many of the lints should be run as part of every cargo invocation vs run as part of an explicit "lint" step?
  • See https://rust-lang.zulipchat.com/#narrow/stream/246057-t-cargo/topic/Cargo.20Lints/near/389419039
  • Every invocation for now as decided at the All Hands
• Do we share namespaces when appropriate?
  • See https://rust-lang.zulipchat.com/#narrow/stream/246057-t-cargo/topic/Sharing.20of.20lints/near/475872847
  • No, at All Hands we decided that Cargo should operate on its own lints only
• Where should workspace-level lints run? On the workspace or on each package?
  • Run them on the workspace, subject to the workspace.lints (with a sometimes fallback to lints similar to *.resolver)

Notes

There will still be lints that are not related to the package but related to the user or / invocation. At a later stage, we can consider user-level lint levels.

Showing lint reason once per package:

• When inherited, we could potentially show it once per workspace just like rustc would do for a crate level attribute
• Starting with this now as an MVP, we can always re-evaluate later

Deferred reducing boilerplate

Deferred json messages

Deferred: whether to make lint docs more like clippy

Out of scope: CLI control. That would be evaluated for [lints] as a whole, potentially through config. Note that existing runtime lint control is not officially supported in Cargo but done through an escape hatch

[epage]

From #10160 (comment)

Just had a talk with @Muscraft. This is mostly my idea but it would be great if we have this kind of error overhaul. It will help things like JSON format message (#8283) or developing Cargo's own lint rules with -Zlints (#12115). It would also open a door to diagonostics translation (rust-lang/rust#100717). Probably the modularization of Cargo could benefit from it as well.

[epage]

imo we need structured diagnostics which we pass to the diagnostic system which will then be turned into errors or something else

[epage]

An obvious way to implement this is to shove the lint config in and our of Config as we enter and leave the relevant contexts. This is a bit more brittle and lowers our chances for doing some operations with threading as Config needs to be mutable everywhere

One thought for how we could implement this

• Rename Config to Context (we might want to do this anyways)
• Refactor Context to allow decorating it (wrapping it)
• Whenever we start down a workspace or package specific code path, we decorate the Context with that workspace or package's lint configuration