S3: Blacklist qxl to prevent shutdown deadlock (ENG-500) #8

Merged

demi.dev 1w ago

targeting main from stoopkidddd/eng-500-executor-vm-hangs-on-reboot-the-qxl-graphics-driver

Merged opened by and assigned to

demi.dev 1w ago targeting main from stoopkidddd/eng-500-executor-vm-hangs-on-reboot-the-qxl-graphics-driver

Linear: ENG-500. It blocks the S3 gate, ENG-472.

The ENG-472 reboot test hung for 40 minutes. The VM's qxl DRM driver logged [TTM] Buffer eviction failed every 15 s. At shutdown, PID 1 blocked on the DRM modeset lock, which was held by a kworker stuck in qxl_fence_wait. With PID 1 in D state, no stop timeout could fire.

This blacklists qxl in executor/hardware-configuration.nix. The console stays in VGA text mode and on ttyS0.

Checks: mise run check exit 0 (hadolint, shellcheck, nix flake check, toplevel eval → sdv11jmb…). boot.blacklistedKernelModules evaluates to ["ip_tables","qxl"].

Deploy (lead, after merge): build and switch on the VM, reboot twice, check lsmod has no qxl, and confirm the executor reconnects with no manual step.

🤖 Generated with Claude Code

Commits 0

Could not load commits.

Comments

v0 1w ago

demi.dev 1w ago

pneumo-reviewer: APPROVE (round 1, head 0c71f0d) #

Change: one comment line and boot.blacklistedKernelModules = [ "qxl" ]; in executor/hardware-configuration.nix. Nothing else changed.

Findings: none. No FIX, no FILE.

What the reviewer checked against the pinned nixpkgs (kernel 6.18.60, systemd initrd):

DRM_QXL builds as a module, not built in, so the blacklist applies. /etc/modprobe.d/nixos.conf is blacklist ip_tables / blacklist qxl.
No NixOS module in this config loads qxl. xserver, plymouth and kmscon are all off. The initrd modules (dm_mod virtio_balloon virtio_console virtio_gpu virtio_rng) don't pull qxl in, and the initrd also copies the blacklist.
grub.gfxpayloadBios = "text", so vgacon stays the tty1 console and Unraid VNC still shows it. console=ttyS0 is unchanged.
The boot path is unchanged: root on virtio_blk, GRUB on /dev/vda.

DoD: mise run check exit 0 on 0c71f0d (hadolint, shellcheck, nix flake check, toplevel eval). git diff --check is clean.

Deploy notes (lead):

The switch doesn't unload the qxl module that is already loaded, so the first reboot can still hang. Reboot soon after the switch, and be ready to force-reset. The second reboot is the real test.
Check lspci -k (no driver on the VGA device) and ls /sys/class/drm (empty) as well as lsmod.

Versions

pneumo-reviewer: APPROVE (round 1, head 0c71f0d) #