Skip to main content
FlexInfer docs

Gaming Image Canary (Sunshine security + exact-hardware gate)

Every gaming session activation is gated on the runtime image being listed in gaming.approvedImages (deploy/system/values-k3s.yaml → the flexinfer-gaming-policy ConfigMap). The list ships empty and only a digest that passed this canary may be added. There is deliberately no in-session override: a GamingSession cannot approve its own image.

Why the gate exists

  • The last known-streaming pin, Sunshine 2025.924.154138 (build/Dockerfile.runtime), is affected by GHSA-ph75-mgxh-mv57 / CVE-2026-32253, a critical remote authentication bypass fixed in v2026.516.143833.
  • The patched 2026.516 release was already tried here and reverted the same day: its wlroots capture path (broken from v2026.407, Sunshine #5018) segfaults on stream start on the exact cblevins-7900xtx headless sway + AMD VAAPI stack (f88f8678b → de205c44a). Upstream merged PR #5030 for wlroots/multi-GPU/AMD DMA-BUF problems, but a release note is not proof on this host.

So: the old image is insecure, the patched image is (last we checked) unusable. Candidates are one of:

  1. a newer upstream release carrying both the auth fix and the #5030 capture fixes;
  2. a changed capture/encoding path (e.g. different capture backend) on a secure release; or
  3. the known-working 2025.924 build with the authentication fix backported (last resort; maintenance burden).

Building a candidate

Off-CI on the GPU builder, like every heavy runtime build:

SUNSHINE_VERSION=<candidate-version> ./build/build-runtime.sh gfx1100-gaming --push

SUNSHINE_VERSION overrides the Dockerfile's safe-window pin for this build only (leave the Dockerfile pin + its window comment intact until a candidate passes). Use docker --context 7900xtx. Record the pushed digest.

Fail-fast sanity check before the full build: install the candidate .deb in a bare ubuntu:24.04 container to catch packaging breaks early.

The 30-minute kill test (exact hardware, isolated)

Run against an isolated runtime container on cblevins-7900xtx (not the live DaemonSet), LAN-restricted:

  1. Start the candidate image privileged with /dev/dri, /dev/uinput, /dev/input, /run/udev, host network, on the node.
  2. Confirm the Sunshine version reports the security fix, and that capture binds headless sway on renderD128 (NAVI31, not the iGPU) with h264/hevc/av1 VAAPI encoders probed.
  3. Pair and connect the Steam Deck via Moonlight at 1280x800/60.
  4. Play 15 minutes of real motion with controller input and audio; disconnect and reconnect once.
  5. PASS requires: no Sunshine segfault (check node dmesg, not container logs — the crash logs nothing), no backend restart (flexinfer_runtime_gaming_backend_restarts_total flat), no black frames, no capture-device switch, no sustained decoder failures on the Deck.
  6. Security check: an unpaired client must not be able to establish an authenticated control session against the web UI/API.

Prior art: .loom/killtest-gaming-sunshine-gfx1100-2026-06-30.md (original substrate proof), docs/user/gaming-mode-runbook.md (diagnosis table).

On pass

  1. Add the digest to gaming.approvedImages in deploy/system/values-k3s.yaml with a dated comment naming the Sunshine version and this canary. Merge; Flux updates the policy ConfigMap.
  2. Update the SUNSHINE_VERSION pin + window comment in build/Dockerfile.runtime to the passing version.
  3. Restore the gaming stanza on the gfx1100 runtime profile (image digest, hostNetwork: true, gaming: true, input/udev/gaming-data mounts, GAMING_* env) — shape in commit 0de40f610. From then on the node is statically gaming-capable and sessions are purely flexinfer game start/stop.

On fail

Keep approvedImages empty. Sessions stay refused with CapabilityReady=False/ImageNotApproved — the lease/CLI UX must not ship on top of a vulnerable or crashing streaming runtime. File the failure evidence next to this doc and try the next candidate class.

Gaming Image Canary (Sunshine security + exact-hardware gate) | FlexInfer docs