Skip to content

Linux 6.17 on Intel can cause checksum errors, SIGSEGVs, and silent corruption #6

Description

@filippo-claude

Some workloads run by sandboxed-step can fail intermittently on GitHub-hosted Ubuntu 24.04 Intel runners using Linux 6.17. These are sandbox memory-corruption symptoms, not necessarily bugs in the command that failed.

Examples seen in affected jobs:

zip: checksum error
signal: segmentation fault (core dumped)
fatal error: traceback did not unwind completely
found pointer to free object
invalid character '\x00' looking for beginning of value

Other possible signs include compiler SIGSEGVs, corrupted downloads or archives, impossible Go runtime failures, and output containing unexpected NUL bytes. Retrying may appear to fix the job, especially if the retry lands on a different CPU vendor.

The affected combination is:

  • the release-20251020.0 gVisor runsc bundled by current releases of this action;
  • Linux 6.17 GitHub runner kernels, observed on 6.17.0-1018-azure and 6.17.0-1020-azure; and
  • Intel CPUs. AMD controls did not reproduce the problem.

The root cause is google/gvisor#12741, fixed upstream by google/gvisor#12994.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions