nixos-servers/docs/plans/pn51-stability.md

ASUS PN51 Stability Testing

Overview

Two ASUS PN51-E1 mini PCs (Ryzen 7 5700U) purchased years ago but shelved due to stability issues. Revisiting them to potentially add to the homelab.

Hardware

	pn01 (10.69.12.60)	pn02 (10.69.12.61)
CPU	AMD Ryzen 7 5700U (8C/16T)	AMD Ryzen 7 5700U (8C/16T)
RAM	2x 32GB DDR4 SO-DIMM (64GB)	1x 32GB DDR4 SO-DIMM (32GB)
Storage	1TB NVMe	1TB Samsung 870 EVO (SATA SSD)
BIOS	0508 (2023-11-08)	Updated 2026-02-21 (latest from ASUS)

Original Issues

• pn01: Would boot but freeze randomly after some time. No console errors, completely unresponsive. memtest86 passed.
• pn02: Had trouble booting — would start loading kernel from installer USB then instantly reboot. When it did boot, would also freeze randomly.

Debugging Steps

2026-02-21: Initial Setup

1. Disabled fTPM (labeled "Security Device" in ASUS BIOS) on both units
   • AMD Ryzen 5000 series had a known fTPM bug causing random hard freezes with no console output
   • Both units booted the NixOS installer successfully after this change
2. Installed NixOS on both, added to repo as pn01 and pn02 on VLAN 12
3. Configured monitoring (node-exporter, promtail, nixos-exporter)

2026-02-21: pn02 First Freeze

• pn02 froze approximately 1 hour after boot
• All three Prometheus targets went down simultaneously — hard freeze, not graceful shutdown
• Journal on next boot: system.journal corrupted or uncleanly shut down
• Kernel warnings from boot log before freeze:
  • TSC clocksource unstable: Marking clocksource 'tsc' as unstable because the skew is too large — TSC skewing ~3.8ms over 500ms relative to HPET watchdog
  • AMD PSP error: psp gfx command LOAD_TA(0x1) failed and response status is (0x7) — Platform Security Processor failing to load trusted application
• pn01 did not show these warnings on this particular boot, but has shown them historically (see below)

2026-02-21: pn02 BIOS Update

• Updated pn02 BIOS to latest version from ASUS website
• TSC still unstable after BIOS update — same ~3.8ms skew
• PSP LOAD_TA still failing after BIOS update
• Monitoring back up, letting it run to see if freeze recurs

2026-02-22: TSC/PSP Confirmed on Both Units

• Checked kernel logs after ~9 hours uptime — both units still running
• pn01 now shows TSC unstable and PSP LOAD_TA failure on this boot (same ~3.8ms TSC skew, same PSP error)
• pn01 had these same issues historically when tested years ago — the earlier clean boot was just lucky TSC calibration timing
• Conclusion: TSC instability and PSP LOAD_TA are platform-level quirks of the PN51-E1 / Ryzen 5700U, present on both units
• The kernel handles TSC instability gracefully (falls back to HPET), and PSP LOAD_TA is non-fatal
• Neither issue is likely the cause of the hard freezes — the fTPM bug remains the primary suspect

2026-02-22: Stress Test (1 hour)

• Ran stress-ng --cpu 16 --vm 2 --vm-bytes 8G --timeout 1h on both units
• CPU temps peaked at ~85°C, settled to ~80°C sustained (throttle limit is 105°C)
• Both survived the full hour with no freezes, no MCE errors, no kernel issues
• No concerning log entries during or after the test

2026-02-22: TSC Runtime Switch Test

• Attempted to switch clocksource back to TSC at runtime on pn01:

  echo tsc > /sys/devices/system/clocksource/clocksource0/current_clocksource
  

• Kernel watchdog immediately reverted to HPET — TSC skew is ongoing, not just a boot-time issue
• Conclusion: TSC is genuinely unstable on the PN51-E1 platform. HPET is the correct clocksource.
• For virtualization (Incus), this means guest VMs will use HPET-backed timing. Performance impact is minimal for typical server workloads (DNS, monitoring, light services) but would matter for latency-sensitive applications.

2026-02-22: BIOS Tweaks (Both Units)

• Disabled ErP Ready on both (EU power efficiency mode — aggressively cuts power in idle)
• Disabled WiFi and Bluetooth in BIOS on both
• TSC still unstable after these changes — same ~3.8ms skew on both units
• ErP/power states are not the cause of the TSC issue

2026-02-22: pn02 Second Freeze

• pn02 froze again ~5.5 hours after boot (at idle, not under load)
• All Prometheus targets down simultaneously — same hard freeze pattern
• Last log entry was normal nix-daemon activity — zero warning/error logs before crash
• Survived the 1h stress test earlier but froze at idle later — not thermal
• pn01 remains stable throughout
• Action: Blacklisted amdgpu kernel module on pn02 (boot.blacklistedKernelModules = [ "amdgpu" ]) to eliminate GPU/PSP firmware interactions as a cause. No console output but managed via SSH.
• Action: Added diagnostic/recovery config to pn02:
  • panic=10 + nmi_watchdog=1 kernel params — auto-reboot after 10s on panic
  • softlockup_panic + hardlockup_panic sysctls — convert lockups to panics with stack traces
  • hardware.rasdaemon with recording — logs hardware errors (MCE, PCIe AER, memory) to sqlite database, survives reboots
  • Check recorded errors: ras-mc-ctl --summary, ras-mc-ctl --errors

Benign Kernel Errors (Both Units)

These appear on both units and can be ignored:

• clocksource: Marking clocksource 'tsc' as unstable — TSC skew vs HPET, kernel falls back gracefully. Platform-level quirk on PN51-E1, not always reproducible on every boot.
• psp gfx command LOAD_TA(0x1) failed — AMD PSP firmware error, non-fatal. Present on both units across all BIOS versions.
• pcie_mp2_amd: amd_sfh_hid_client_init failed err -95 — AMD Sensor Fusion Hub, no sensors connected
• Bluetooth: hci0: Reading supported features failed — Bluetooth init quirk
• Serial bus multi instantiate pseudo device driver INT3515:00: error -ENXIO — unused serial bus device
• snd_hda_intel: no codecs found — no audio device connected, headless server
• ata2.00: supports DRM functions and may not be fully accessible — Samsung SSD DRM quirk (pn02 only)

2026-02-23: processor.max_cstate=1 and Proxmox Forums

• Found a thread on the Proxmox forums about PN51 units with similar freeze issues
  • Many users reporting identical symptoms — random hard freezes, no log evidence
  • No conclusive fix. Some have frequent freezes, others only a few times a month
  • Some reported BIOS updates helped, but results inconsistent
• Added processor.max_cstate=1 kernel parameter to pn02 — limits CPU to C1 halt state, preventing deep C-state sleep transitions that may trigger freezes on AMD mobile chips
• Also applied: amdgpu blacklist, panic=10, nmi_watchdog=1, softlockup/hardlockup panic, rasdaemon

2026-02-23: logind D-Bus Deadlock (pn02)

• node-exporter alert fired — but host was NOT frozen
• logind was running (PID 871) but deadlocked on D-Bus — not responding to org.freedesktop.login1 requests
• Every node-exporter scrape blocked for 25s waiting for logind, causing scrape timeouts
• Likely related to amdgpu blacklist — no DRM device means no graphical seat, logind may have deadlocked during seat enumeration at boot
• Fix: systemctl restart systemd-logind + systemctl restart prometheus-node-exporter
• After restart, logind responded normally and reported seat0

2026-02-27: pn02 Third Freeze

• pn02 crashed again after ~2 days 21 hours uptime (longest run so far)
• Evidence of crash:
  • Journal file corrupted: system.journal corrupted or uncleanly shut down
  • Boot partition fsck: Dirty bit is set. Fs was not properly unmounted
  • No orderly shutdown logs from previous boot
  • No auto-upgrade triggered
• NMI watchdog did NOT fire — no kernel panic logged. This is a true hard lockup below NMI level
• rasdaemon recorded nothing — no MCE, AER, or memory errors in the sqlite database
• Positive: The system auto-rebooted this time (likely hardware watchdog), unlike previous freezes that required manual power cycle
• processor.max_cstate=1 may have extended uptime (2d21h vs previous 1h and 5.5h) but did not prevent the freeze

2026-02-27 to 2026-03-03: Relative Stability

• pn02 ran without crashes for approximately one week after the third freeze
• pn01 continued to be completely stable throughout this period
• Auto-upgrade reboots continued daily (~4am) on both units — these are planned and healthy

2026-03-04: pn02 Fourth Crash — sched_ext Kernel Oops (pstore captured)

• pn02 crashed after ~5.8 days uptime (504566s)
• First crash captured by pstore — kernel oops and panic stack traces preserved across reboot
• Journal corruption confirmed: system.journal corrupted or uncleanly shut down
• Crash location: RIP: 0010:set_next_task_scx+0x6e/0x210 — crash in the sched_ext (SCX) scheduler subsystem
• Call trace: sysvec_apic_timer_interrupt → cpuidle_enter_state — crashed during CPU idle, triggered by APIC timer interrupt
• CR2: ffffffffffffff89 — dereferencing an obviously invalid kernel pointer
• Kernel: 6.12.74 (NixOS 25.11)
• Significance: This is the first crash with actual diagnostic output. Previous crashes were silent sub-NMI freezes. The sched_ext scheduler path is a new finding — earlier crashes were assumed to be hardware-level.

2026-03-06: pn02 Fifth Crash

• pn02 crashed again — journal corruption on next boot
• No pstore data captured for this crash

2026-03-07: pn02 Sixth and Seventh Crashes — Two in One Day

First crash (~11:06 UTC):

• ~26.6 hours uptime (95994s)
• pstore captured both Oops and Panic
• Crash location: Scheduler code path — pick_next_task_fair → __pick_next_task
• CR2: 000000c000726000 — invalid pointer dereference
• Notable: dbus-daemon segfaulted ~50 minutes before the kernel crash (segfault at 0 in libdbus-1.so.3.32.4 on CPU 0) — may indicate memory corruption preceding the kernel crash

Second crash (~21:15 UTC):

• Journal corruption confirmed on next boot
• No pstore data captured

2026-03-07: pn01 Status

• pn01 has had zero crashes since initial setup on Feb 21
• Zero journal corruptions, zero pstore dumps in 30 days
• Same BOOT_ID maintained between daily auto-upgrade reboots — consistently clean shutdown/reboot cycles
• All 8 reboots in 30 days are planned auto-upgrade reboots
• pn01 is fully stable

Crash Summary

Date	Uptime Before Crash	Crash Type	Diagnostic Data
Feb 21	~1h	Silent freeze	None — sub-NMI
Feb 22	~5.5h	Silent freeze	None — sub-NMI
Feb 27	~2d 21h	Silent freeze	None — sub-NMI, rasdaemon empty
Mar 4	~5.8d	Kernel oops	pstore: set_next_task_scx (sched_ext)
Mar 6	Unknown	Crash	Journal corruption only
Mar 7	~26.6h	Kernel oops + panic	pstore: pick_next_task_fair (scheduler) + dbus segfault
Mar 7	Unknown	Crash	Journal corruption only

Conclusion

pn02 is unreliable. After exhausting mitigations (fTPM disabled, BIOS updated, WiFi/BT disabled, ErP disabled, amdgpu blacklisted, processor.max_cstate=1, NMI watchdog, rasdaemon), the unit still crashes every few days. 26 reboots in 30 days (7 unclean crashes + daily auto-upgrade reboots).

The pstore crash dumps from March reveal a new dimension: at least some crashes are kernel scheduler bugs in sched_ext, not just silent hardware-level freezes. The set_next_task_scx and pick_next_task_fair crash sites, combined with the dbus-daemon segfault before one crash, suggest possible memory corruption that manifests in the scheduler. It's unclear whether this is:

1. A sched_ext kernel bug exposed by the PN51's hardware quirks (unstable TSC, C-state behavior)
2. Hardware-induced memory corruption that happens to hit scheduler data structures
3. A pure software bug in the 6.12.74 kernel's sched_ext implementation

pn01 is stable — zero crashes in 30 days of continuous operation. Both units have identical kernel and NixOS configuration (minus pn02's diagnostic mitigations), so the difference points toward a hardware defect specific to the pn02 board.

Next Steps

• pn02 memtest: Run memtest86 for 24h+ (available in systemd-boot menu). The crash signatures (userspace segfaults before kernel panics, corrupted pointers in scheduler structures) are consistent with intermittent RAM errors that a quick pass wouldn't catch. If memtest finds errors, swap the DIMM.
• pn02: Consider scrapping or repurposing for non-critical workloads that tolerate random reboots (auto-recovery via hardware watchdog is now working)
• pn02 investigation: Could try disabling sched_ext (boot.kernelParams = [ "sched_ext.enabled=0" ] or equivalent) to test whether the crashes stop — would help distinguish kernel bug from hardware defect
• pn01: Continue monitoring. If it remains stable long-term, it is viable for light workloads
• If pn01 eventually crashes, apply the same mitigations (amdgpu blacklist, max_cstate=1) to see if they help
• For the Incus hypervisor plan: likely need different hardware. Evaluating GMKtec G3 (Intel) as an alternative. Note: mixed Intel/AMD cluster complicates live migration

Diagnostics and Auto-Recovery (pn02)

Currently deployed on pn02:

boot.blacklistedKernelModules = [ "amdgpu" ];
boot.kernelParams = [ "panic=10" "nmi_watchdog=1" "processor.max_cstate=1" ];
boot.kernel.sysctl."kernel.softlockup_panic" = 1;
boot.kernel.sysctl."kernel.hardlockup_panic" = 1;
hardware.rasdaemon.enable = true;
hardware.rasdaemon.record = true;

Crash recovery is working: pstore now captures kernel oops/panic data, and the system auto-reboots via panic=10 or SP5100 TCO hardware watchdog.

After reboot, check:

• ras-mc-ctl --summary — overview of hardware errors
• ras-mc-ctl --errors — detailed error list
• journalctl -b -1 -p err — kernel logs from crashed boot (if panic was logged)
• pstore data is automatically archived by systemd-pstore.service and forwarded to Loki via promtail