proxmox-system-safety

name	proxmox-system-safety
description	Proxmox system safety operations and hardware detection patterns. Use when managing system operations, hardware detection, or safety-critical tasks on Proxmox hosts.

Proxmox System Safety Rules

LVM Operations on Root Volumes

1. Do NOT create LVM snapshots of the Proxmox root volume (pve/root). Merging snapshots on a live root volume is unreliable and can leave the system in a stuck merge state requiring reboot.

2. Use file-based config backups (tar) and vzdump for VMs instead.

Reboot Awareness

3. If a playbook changes GRUB, initramfs, or kernel modules, a reboot may be needed.

4. Set pci_passthrough_allow_reboot: true in host vars to allow automated reboots.

5. After reboot, wait for SSH to come back with wait_for_connection.

Hardware Detection Requirements

6. iGPU: every modern Intel CPU has one. proxmox_igpu MUST hard-fail if absent.

7. WiFi + VT-d/IOMMU: required for PCI passthrough. proxmox_pci_passthrough MUST hard-fail if IOMMU is not active after reboot or groups are invalid.

8. NIC count: OK to handle dynamically (hardware legitimately varies).

Hardware Failure Requirements

9. NEVER add "graceful skip" for hardware expected on every host. Silent skips mask fixable BIOS settings (VT-d disabled) behind warnings that are easy to miss.

10. Previous bug: proxmox_pci_passthrough silently skipped WiFi passthrough when IOMMU groups were invalid on mesh1. Root cause was VT-d disabled in BIOS — a 30-second fix masked for an entire test cycle.

System Safety Decision Tree

11. Use this decision tree:

    Is it modifying LVM on root?
    ├── YES → BLOCK. Use tar + vzdump instead.
    └── NO → SAFE. Proceed.
    

Hardware Detection Pattern

12. For expected hardware (iGPU, WiFi with VT-d/IOMMU), always hard-fail when absent rather than graceful skip. This ensures critical issues are caught immediately rather than silently ignored.

PCI Passthrough Prerequisites

13. WiFi PCIe passthrough requires the q35 machine type. Set machine: q35 when wifi_pci_devices is non-empty.

14. IOMMU group isolation is mandatory. ALWAYS verify before binding to vfio-pci.

15. WiFi NICs must be excluded from bridge creation — they're passed through via PCIe, not bridged.

Missing DRI Devices Recovery

17. Between test cycles, after VM passthrough, or hookscript operations, the iGPU can lack /dev/dri nodes even though the native driver module is loaded. Two causes: (a) device bound to vfio-pci, (b) device unbound from native driver but module still loaded (PCI rescan alone won't re-bind).

18. proxmox_igpu MUST check /dev/dri/renderD128 existence BEFORE device detection. If missing: unbind from whatever driver holds the device, clear driver_override, PCI rescan, then EXPLICITLY bind to the native driver (echo PCI_ADDR > /sys/bus/pci/drivers/i915/bind). PCI rescan alone does NOT auto-bind when the module is already loaded.

19. Previous bug: molecule converge (no cleanup) after a passthrough test left the iGPU without DRI nodes on home. proxmox_igpu saw i915 in lsmod, skipped recovery, then hard-failed on "DRI devices missing." Explicit driver bind after PCI rescan is the fix.

Proxmox firmware package conflicts

20. On Proxmox VE, pve-firmware bundles all Intel/AMD firmware including iwlwifi. NEVER install standalone firmware-iwlwifi — it conflicts with pve-firmware and triggers the Proxmox apt hook to block removal of the proxmox-ve meta-package.

21. proxmox_pci_passthrough MUST check for pve-firmware before attempting to install firmware-iwlwifi. If pve-firmware is present, skip the install — the firmware is already available.

22. Previous bug: firmware-iwlwifi install failed on mesh1, bridge-1, bridge-2, and home with pve-apt-hook returned error code (1). All hosts had pve-firmware installed, which already provides iwlwifi firmware.

Enterprise Repository Management

23. Proxmox enterprise repo disabling MUST happen in pre_tasks of infrastructure plays, BEFORE any role that calls apt update. The proxmox_pci_passthrough role needs apt for firmware packages and runs before proxmox_igpu.

24. NEVER put repo management inside a role that isn't the FIRST role in the play. If any earlier role needs apt, the repos won't be ready.

25. Previous bug: Enterprise repo disabling was inside proxmox_igpu (third role in infra play). proxmox_pci_passthrough (second role) ran apt update for firmware packages and failed with 401 Unauthorized on mesh2. Fix: moved to pre_tasks of both infra plays in site.yml.

Package Name Verification

16. NEVER assume a package name is correct without checking. Package names vary between Debian releases, architectures, and distributions. ALWAYS verify with apt-cache search <keyword> or apt list <name>.

17. Previous bug: intel-media-va-driver-non-free was correct on Debian Bullseye but does not exist on Debian Trixie. The correct package is intel-media-va-driver. The task failed with "No package matching" and required manual investigation.

Dynamic Device Detection

18. NEVER hardcode device paths like /dev/dri/card0. The card number depends on driver probe order and can change across reboots or kernel updates.

19. ALWAYS detect devices dynamically by querying sysfs driver bindings: iterate /dev/dri/card*, check readlink -f /sys/class/drm/cardN/device/driver, and match on the driver name.

20. Previous bug: /dev/dri/card0 was assumed to be the Intel iGPU, but on a multi-GPU system card0 was the discrete GPU. Sysfs-based detection finds the correct device regardless of probe order.

iGPU PCI Passthrough (vfio-pci)

21. Prefer runtime sysfs manipulation over persistent modprobe configs for iGPU passthrough. Writing to /sys/bus/pci/drivers/vfio-pci/new_id and /sys/bus/pci/drivers/vfio-pci/bind is reversible without initramfs updates. Modprobe blacklists require update-initramfs and a reboot.

22. Single-GPU passthrough is supported for Intel iGPUs and discrete GPUs via Proxmox hookscripts. The hookscript pattern: pre-start unbinds the GPU from the native driver via sysfs and binds to vfio-pci; post-stop reverses the operation. The host runs headless (SSH/web only) while the VM has the GPU. NEVER attempt GPU passthrough on AMD APU iGPUs (Raven Ridge, etc.) — the GPU shares the SoC die and ANY unbind path (sysfs or modprobe -r) hangs the entire system. This is a hardware limitation, not fixable with hookscripts.

23. NEVER run modprobe -r amdgpu or modprobe -r i915 as GPU cleanup. ALWAYS use sysfs operations: unbind from vfio-pci, clear driver_override, PCI rescan. The rescan triggers the kernel to auto-bind the native driver. This is safe on ANY host regardless of GPU count.

24. GPU passthrough hookscript (/var/lib/vz/snippets/gpu-passthrough-hook.sh) manages the full lifecycle: discovers hostpci devices from VM config, stops GPU-consuming LXC containers, suspends conflicting VMs, binds/unbinds via sysfs, persists state in /run/gpu-passthrough/vm-<VMID>.state for post-stop recovery. The hookscript is generalized — works with any GPU vendor (Intel, AMD, NVIDIA) and any VM that uses --hostpci.

25. Previous bug: cleanup ran modprobe -r amdgpu on ALL hosts via E2E cleanup. On ai (single AMD GPU, USB ethernet), this caused a kernel panic. Fix: replaced all GPU modprobe -r with sysfs unbind + PCI rescan.

26. Previous bug: sysfs unbind of AMD Raven Ridge APU iGPU (1002:15dd) via hookscript pre-start hung the entire system. Unlike discrete GPUs, APU iGPUs share the SoC die with the CPU — even sysfs unbind triggers a GPU reset that freezes the NBIO, killing the entire system including USB ethernet (EHOSTUNREACH). This is the same class of failure as modprobe -r amdgpu but at the hardware level.

27. Previous bug: PCI rescan after vfio-pci unbind did NOT auto-bind the native driver when the module was already loaded. DRI devices (/dev/dri/renderD128) did not reappear. Fix: explicitly bind to the native driver after rescan (echo PCI_ADDR > /sys/bus/pci/drivers/i915/bind). The cleanup and hookscript post-stop both must do explicit rebinding, not rely on auto-binding.

Hookscript Attachment Ordering

28. Display-exclusive hookscripts that stop DRI-sharing containers MUST NOT be attached during provisioning. Attaching during kiosk_lxc (Phase 2.5) causes the hookscript to fire when desktop_lxc starts in Phase 3, stopping all DRI containers (Kodi, Moonlight, Kiosk) before their configure plays run.

29. Pattern: deploy the hookscript FILE in the provisioning role, attach it to containers/VMs in a dedicated play AFTER all configure plays finish. This ensures all containers are configured before the hookscript can stop them.

30. Configure roles for DRI-sharing containers should include a defensive "ensure container is running" guard at the top (check pct status, start if stopped, wait for readiness). This handles re-runs where hookscripts may already be attached from a previous cycle.

31. Previous bug: kiosk_lxc deployed AND attached the display-exclusive hookscript during Phase 2.5. When desktop_lxc started in Phase 3, the hookscript's pre-start(400) stopped Kodi (301), Moonlight (302), and Kiosk (401). Configure Kodi then failed with "container '301' not running!" Fix: split hookscript deployment (provisioning) from attachment (post-configure play).

32. Verify assertions for DRI-sharing containers (Kodi, Kiosk, Moonlight) MUST skip systemctl is-active checks when the Desktop LXC is running and has grabbed the DRI render node. The display-exclusive hookscript manages this conflict. Check pct status {{ desktop_ct_id }} and gate the service-active assertion.

33. Previous bug: Kodi container was running but systemctl is-active kodi returned inactive during verify. Root cause: Desktop VM held the iGPU, DRI devices absent from container. Fix: added "'running' not in (_desktop_status.stdout | default(''))" condition to the Kodi service-active assertion.

34. Not all services run as systemd daemons. Moonlight-embedded is an on-demand streaming client binary (/usr/local/bin/moonlight), not a persistent service. Verify assertions for such services should check binary existence and config deployment, NOT systemctl is-active.

35. Previous bug: Moonlight verify assertion checked systemctl is-active moonlight but there IS no moonlight.service — moonlight-embedded is compiled from source as a CLI binary with no systemd unit file. The assertion always failed. Fix: changed to check binary existence and config file presence.

Host Recoverability

28. Every host MUST declare wol_capable (true/false) in host_vars. This tracks whether the host can be remotely recovered via Wake-on-LAN after a crash or shutdown.

29. USB ethernet adapters do NOT support WoL. The USB host controller powers down in S5 (standby) and lacks magic packet detection circuitry. Hosts connected exclusively via USB ethernet (e.g., ai) MUST have wol_capable: false.

30. NEVER run operations that could crash a non-WoL host from automation. This includes modprobe -r of the sole GPU driver, shutdown, poweroff, or any operation that could trigger a kernel panic. Non-WoL hosts require physical intervention to recover.

31. scripts/wol.sh MUST NOT include non-WoL hosts. Unit tests in tests/test_wol.py enforce this. The E2E verify playbook also asserts wol_capable is defined for every host and that non-WoL hosts don't appear in wol.sh.

32. Previous bug: ai was listed in wol.sh with its PCIe NIC MAC, but ai is connected via USB ethernet only. The PCIe NIC is not connected to the network, making WoL impossible.