systemg

Search docs

/
Install

Kernel Mode

Kernel-Enforced Sandboxing

sysg configures kernel enforcement mechanisms per service: rlimits, capabilities, namespaces, cgroups, seccomp, no_new_privs, and Landlock. All of them are Linux, root, and enforced by the kernel — sysg only configures them.

Limits, capabilities, and namespaces

services:
  worker:
    command: "./worker"
    user: "worker"
    limits:
      nofile: 4096
      cgroup:
        memory_max: "512M"
    capabilities: ["CAP_NET_BIND_SERVICE"]
    isolation:
      network: true

Syscall and filesystem confinement (schema v3)

Under version: "3":

  • isolation.landlock: { ro_paths, rw_paths } — a Landlock filesystem sandbox requiring ABI 3 or newer (normally Linux 6.2+, including supported backports). Truncation is granted only under rw_paths; listing a path in ro_paths does not grant it. The remaining handled filesystem rights retain their original ABI 1 policy. Landlock does not revoke access through descriptors opened before confinement.
  • isolation.seccomp: "baseline-v1" — a frozen, deny-by-default syscall allowlist; unlisted syscalls return EPERM.

sysg sets no_new_privs only when a service has an effective Landlock or seccomp policy. It can't be undone and every child inherits it, so setuid helpers like sudo can't gain privileges under it. For any other service sysg leaves the inherited state alone. If the supervisor starts without no_new_privs, those services can use sudo -u postgres ... and similar, subject to the helper's own policy.

version: "3"
services:
  worker:
    command: "./worker"
    isolation:
      seccomp: "baseline-v1"
      landlock:
        ro_paths: ["/usr", "/lib", "/etc/worker"]
        rw_paths: ["/var/lib/worker"]

An unsupported kernel or a malformed policy refuses the service (SG0722/SG0724) rather than running it unprotected. baseline-v1 is frozen; a different or stricter policy will ship as baseline-v2, never as a silent edit.

Keys that do not enforce

isolation.pid: true refuses the service under both schema versions. PID namespace lifecycle support is not implemented; sysg does not launch the service in the host PID namespace while claiming it is isolated.

isolation.apparmor_profile, isolation.selinux_context, isolation.private_devices, and isolation.private_tmp are accepted by the schema but cannot be enforced. Schema version 3 refuses services that request them; version 2 runs the service with a warning.

Warning

Each of these keys emits SG0721 at service start. A version 2 warning is not protection. Remove the keys only if running without the requested isolation is acceptable.

Diagnostics

  • SG0721 — a security key was accepted but cannot be enforced; PID isolation is refused under both schema versions, and other unsupported keys are refused under version 3
  • SG0722 — seccomp filter could not be built, compiled, or applied; service refused
  • SG0723 — no_new_privs could not be set; service refused
  • SG0724 — Landlock requested but the kernel ABI is unavailable or insufficient; service refused
  • SG0725 — unknown seccomp profile; the only built-in is baseline-v1
  • SG0726 — seccomp is unsupported on this CPU architecture
← Container-Init (PID 1)Kernel-Assisted Observation →