Capability-free sandbox fails to start on kernels < 5.19 (RHEL 9.x / 5.14): seccomp WAIT_KILLABLE_RECV EINVAL

从 issue 中描述的 capability-probe seccomp 通知设置和 supervisor 的启动确认处理开始；将它们与提议的 fallback 分支进行比较。确定旧 kernel 上的行为是否可接受，然后验证无 capability 的 sandbox 能在 kernel 5.14 上达到 Ready，同时在更新的 kernel 上保留 WAIT_KILLABLE_RECV，并记录降级语义。

User Story

As an operator running OpenShell on OpenShift / RHEL 9.x nodes, I want the RFC-0012 capability-free sandbox to start on my existing fleet, so that I can adopt the new isolation model without waiting for a kernel bump.

Problem Statement

At startup the sandbox installs its seccomp notification listener with SECCOMP_FILTER_FLAG_WAIT_KILLABLE_RECV, introduced in Linux 5.19. On older kernels the seccomp() call fails:

seccomp(SET_MODE_FILTER, NEW_LISTENER|WAIT_KILLABLE_RECV, …) = -1 EINVAL

The listener thread then dies (notification launcher disappeared) and the supervisor fails confirmation (sandbox confirmation evidence is incomplete or mismatched). This appears intentional — the code maps this EINVAL to "seccomp WAIT_KILLABLE_RECV is required (Linux 5.19 or newer)" — so this is as much a "should we relax it?" as a bug report.

Impact / Why This Matters

RHEL 9.x ships kernel 5.14 for its lifecycle, and OpenShift/RHCOS nodes run the RHEL kernel. So the capability-free sandbox has no working path on current OpenShift until RHEL 10 nodes. There is no in-cluster workaround (it is a kernel-feature gap, not a config/SCC issue). This blocks OpenShift adoption of RFC-0012 in the near term.

Reproduction Steps

1. On a node with kernel < 5.19 (e.g. RHEL 9.8, 5.14.0-687.35.1.el9_8), create a capability-free sandbox (or run openshell-sandbox capability-probe in a zero-cap, no_new_privs pod).
2. The probe fails at the seccomp notification step; strace shows the NEW_LISTENER|WAIT_KILLABLE_RECV call returning EINVAL.

Environment

• OpenShift / RHCOS, RHEL 9.8, kernel 5.14.0-687.35.1.el9_8
• RFC-0012 capability-free model (#2942), Kubernetes compute driver

Ruled out (measured on the same 5.14 node, zero capabilities)

• Not CAP_SYS_ADMIN / SCC: a plain pod at zero caps installs a NEW_LISTENER fine under both RuntimeDefault and Unconfined.
• Not the seccomp profile, user namespaces, SELinux, or struct-size mismatch (GET_NOTIF_SIZES = 80/24/64 on both kernels). The only differentiator is the WAIT_KILLABLE_RECV flag (5.19).

Proposed Design

Make WAIT_KILLABLE_RECV optional: attempt it, and on EINVAL fall back to a plain NEW_LISTENER; and stop gating launch-confirmation on the cancellation evidence (which is exactly this flag). Implications:

• We lose the killable-receive semantics (the workload-side notify wait becomes interruptible rather than kill-only).
• Zero-cap containment is unchanged — the listener still mediates every syscall; the supervisor keeps rejecting stale notifications via NOTIF_ID_VALID.
• Validated: with this change a capability-free sandbox reaches Ready on OpenShift / kernel 5.14, end to end (workload + supervisor).

A branch implementing this (two small commits, off current main) is available: akram:fix/seccomp-wait-killable-fallback-main — https://github.com/NVIDIA/OpenShell/compare/main...akram:OpenShell:fix/seccomp-wait-killable-fallback-main . Happy to open it as a PR if the direction is acceptable.

Alternatives Considered

• Require nodes on kernel ≥ 5.19 (RHEL 10 / newer RHCOS) — leaves current OpenShift users blocked.
• Runtime-installed listener via linux.seccomp.listenerPath — heavier, and not needed for this specific gap.

Acceptance Criteria

• The capability-free sandbox starts and reaches Ready on a kernel-5.14 node.
• On kernels ≥ 5.19 behavior is unchanged (WAIT_KILLABLE_RECV still used).
• The degraded semantics on < 5.19 are documented.

Open question for maintainers

Is the 5.19 floor a hard requirement (a load-bearing property of the isolation model), or the convenient baseline? If a fallback is acceptable, I have the branch above and can open a PR.