Skip to content

macOS: bundled oils shell intermittently fails fast external commands (getpgid ESRCH race) #702

Description

@lifeiscontent

Symptom

A cached task on macOS intermittently reports exit 2 with empty stdout and oils I/O error (main): No such process on stderr. Seen in the wild running parallel test suites that spawn vp run (each task runs under the bundled osh, which fspy substitutes for /bin/sh).

Root cause

Not fspy's exec interposition — the bundled oils 0.37.0 itself. Process::RunProcess (core/process.py) calls getpgid(child_pid) right after fork() and before waitpid(). On macOS, getpgid() of a zombie child fails with ESRCH, so when the parent is descheduled past a fast command's exit (CPU contention), the OSError escapes to oils' main() → exit 2.

Reproduces with the raw oils binary, no fspy involved: ~5% under 6-8-way parallel loops of osh -c 'echo x > f; cat f'.

Fix

Upstream PR: oils-for-unix/oils#2689 — gate the getpgid() on job control being enabled (MaybeGiveTerminal() no-ops otherwise, so batch shells never needed the pgid).

Once an oils release includes it: rebuild the bundled binary and bump MACOS_BINARY_DOWNLOADS in crates/fspy/build.rs.

Verification of the fix

Patched 0.37.0 tarball build (same one-liner in the generated C++): 2000 concurrent runs of the failing shapes, 0 failures (vs ~5% before); smoke script output byte-identical to the unpatched binary.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Fields

    Priority

    None yet

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions