Skip to content

Fix nondeterministic SVH from sandbox-local OUT_DIR/CARGO_MANIFEST_DIR#20

Open
walter-zeromatter wants to merge 24 commits intohermeticbuild:mainfrom
walter-zeromatter:user/wgray/outdir-svh-fix
Open

Fix nondeterministic SVH from sandbox-local OUT_DIR/CARGO_MANIFEST_DIR#20
walter-zeromatter wants to merge 24 commits intohermeticbuild:mainfrom
walter-zeromatter:user/wgray/outdir-svh-fix

Conversation

@walter-zeromatter
Copy link
Copy Markdown

Under sandboxing, ${pwd} resolves to a per-pid sandbox path that differs across actions. rustc's env!() bakes the raw string into the crate SVH and --remap-path-prefix does not normalize env!() values, so crates like pyo3-build-config that do PathBuf::from(env!("OUT_DIR")) produce divergent SVHs between hollow-metadata and full rlib actions, breaking pipelined compilation and poisoning the cache.

Route CARGO_MANIFEST_DIR and OUT_DIR through ${exec_root}, and resolve ${exec_root} in the process wrapper by walking up from cwd to the first ancestor containing Bazel's DO_NOT_BUILD_HERE sentinel, then appending execroot/<workspace_name>. Falls back to the prior cwd-derived value if the sentinel is not found. ${output_base} is left untouched so that --remap-path-prefix=${output_base}=. keeps stripping sandbox-local prefixes from paths embedded in rmeta/debuginfo.

dzbarsky and others added 24 commits April 16, 2026 10:12
…hollow rlibs: the RustcMetadata action runs rustc to completion with -Zno-codegen, emitting a .rlib archive. This approach mirrors the one used by buck2 and avoids needing to kill rustc mid-output in order to produce metadata.

While not fixing problems with SVH mismatches when non-determinism, this does simplify the codepath and uses a production tested technique that doesn't have any of the dangers associated with killing the rustc process while it's still active.
Port the sharding wrapper feature from bazelbuild#3774 into the hermeticbuild fork. The implementation wraps rust_test executables when experimental_enable_sharding is set while keeping rustc_compile_action's existing provider-list API for internal and extension callers.

rust_test now scans the returned providers to replace DefaultInfo for the wrapper, so extensions such as prost and wasm-bindgen continue to consume rustc_compile_action without API churn.

Co-authored-by: Brian Duff <bduff@linkedin.com>

Co-authored-by: Codex <noreply@openai.com>
…tc's -Ldependency scan can pick up the _meta.rlib alongside the full '.rlbi', producing undefined-symbol link errors. To avoid this, emit the hollow metadata rlibs into a '_meta/' subdirectory, and add an explicit link to that directory for any metadata-consuming actions. (bazelbuild#17)
Under sandboxing, ${pwd} resolves to a per-pid sandbox path that differs
across actions. rustc's env!() bakes the raw string into the crate SVH
and --remap-path-prefix does not normalize env!() values, so crates like
pyo3-build-config that do PathBuf::from(env!("OUT_DIR")) produce
divergent SVHs between hollow-metadata and full rlib actions, breaking
pipelined compilation and poisoning the cache.

Route CARGO_MANIFEST_DIR and OUT_DIR through ${exec_root}, and resolve
${exec_root} in the process wrapper by walking up from cwd to the first
ancestor containing Bazel's DO_NOT_BUILD_HERE sentinel, then appending
execroot/<workspace_name>. Falls back to the prior cwd-derived value if
the sentinel is not found. ${output_base} is left untouched so that
--remap-path-prefix=\${output_base}=. keeps stripping sandbox-local
prefixes from paths embedded in rmeta/debuginfo.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants