Running a homelab on repurposed laptop hardware is a great idea right up until the fan starts screaming and the machine vanishes offline in the middle of the night. I chased two hardware gremlins on my Proxmox host for weeks, and I want to document exactly what I tried, because most of the answers came from ruling things out.
The setup #
The host is a 2015 HP notebook (board 80C1, Insyde BIOS F.11 from July 2015, i3-5005U, 12 GB) repurposed as a headless Proxmox VE host. The display and lid are gone, and it sits in a Thermaltake case for airflow. It runs the three k3s VMs (nodes 101, 102, 103).
Issue 1: The fan was loud from boot, loud under load and, before I cleaned it, loud at idle too.
Issue 2: The machine powered off abruptly, with no shutdown sequence in the journal. Just a hard stop, mid-log.
Both could be hardware. Both could be software. I wanted to know which.
The fan problem #
First instinct: OS fan control #
The natural place to start on Linux is hp-wmi, the kernel module that handles HP fan and thermal management. Every boot logs the same thing:
hp-wmi hp-wmi: Failed to apply initial fan settings: -22-22 is EINVAL, and reloading the module reproduces it exactly. That’s not a transient glitch; it’s a firmware and driver mismatch. sensors agrees: pwm1: N/A.
The root cause in the ACPI tables #
The SSDTs reference an HP embedded controller device named \_SB.PCI0.LPCB.H_EC. Every method that touches the fan (CPUP, PCAP, PECC, the thermal zones TZ00/TZ01, the whole lot) routes through H_EC. But H_EC is declared External in every table and never actually defined. The OS has no path to the fan controller at all.
That’s why hp-wmi fails. It’s also why the kernel logs AE_NOT_FOUND aborts on _TZ.TZ00._TMP and _TZ.TZ01._TMP: those methods can’t find H_EC either.
The EC that is reachable, \_SB.PCI0.LPCB.EC0 at the standard ports 0x62/0x66, is a generic helper with raw byte access (FANG(n) / FANW(n,v)) and no fan curve. Dead end.
What I tried and ruled out #
| # | What I tried | Result |
|---|---|---|
| 1 | hp_wmi reload |
Same -22 (EINVAL). Persistent, not a glitch. |
| 2 | /sys/firmware/acpi/platform_profile |
Doesn’t exist. This BIOS predates ACPI platform profiles. |
| 3 | hp_bioscfg BIOS attributes |
Only Sure_Start exposed. No fan or thermal toggle. |
| 4 | nbfc (NoteBook FanControl) | No config for this board. I checked all 313 configs in nbfc-linux/configs. No prior art. |
| 5 | EC register 0x2E/0x2F |
Used by every Broadwell-era HP ProBook/EliteBook nbfc config. 0x2F silently discards writes. Not the fan register here. |
| 6 | EC register 0x58 |
Writable, holds 103. Proven not to be the fan register (see the thermal test below). |
| 7 | Extended EC space (ERIB/ERBD), indices 0x0000 to 0x07FF |
No live registers, no plausible fan or RPM values. Effectively unimplemented. |
| 8 | HP EC shared-memory window at 0xFF000000 |
Readable via /dev/mem, but returns a repeating 1C D4 1E FA pattern. Not the EC block on this platform. |
| 9 | DPTF Fan cooling devices (cooling_device3 to 7) |
Registered, but their AML calls the undefined H_EC.ECMD(0x1A). Writing cur_state=0 vs 1 moved CPU temperature by 1 °C, which is noise. |
The EC register thermal test #
Register 0x58 looked promising: writable and holding a non-trivial value. With no tachometer anywhere on the board, “did the fan change?” had to be answered thermally. I ran a sustained single-core load and forced the register to each extreme, with a watchdog ready to restore the maximum at 82 °C:
PHASE 1 0x58 = 103 (BIOS max) + load -> plateau 64 °C
PHASE 2 0x58 = 0 (min/off) + load -> plateau 64 °C
delta = 0 °CA register that really commanded the fan couldn’t produce a 0 °C delta under sustained load. 0x58 is an unrelated config or telemetry byte.
The answer: a worn fan, not a control fault #
The temperatures were never alarming:
| State | CPU package |
|---|---|
| Idle | 50 to 53 °C |
| Two threads pinned | 55 to 58 °C |
| Normal k3s working load | 54 to 57 °C |
| Sustained single-core pin test | 63 to 64 °C |
For a 15 W Broadwell-U part with a critical threshold of 105 °C, those are unremarkable. The BIOS-side fan controller was doing its job. The OS just has no window into it.
So on 27 September I cleaned the fan and re-seated it. The behaviour changed immediately: loud during startup, then slowing down at idle, which is what a working fan curve sounds like. Before the clean it stayed loud whatever the load.
That settled it. The noise was mechanical. The fan is 12 years old, and cleaning restores airflow but can’t restore a worn bearing. A replacement fan and fresh thermal paste are next, in the same teardown. There’s no RPM readout on this machine (pwm1: N/A, no fan1_input anywhere), and a new fan won’t change that, so the result will be judged by ear.
One sensor note if you have similar hardware: acpitz (thermal_zone0) sat at exactly 27.8 °C while the package climbed nine degrees. It’s either dead or measuring something that doesn’t change. B0D4 (thermal_zone4) is the EC CPU sensor that tracks the package faithfully.
The phantom shutdown problem #
What the journal says #
journalctl -b -1 is unambiguous: this is a hard power loss, not a software shutdown.
- The log ends abruptly at
17:17:01on acron.hourlyentry. - No
systemd-shutdown, noReached target poweroff.target, noSyncing filesystems, no kernel messages. Nothing. - Clean shutdowns are recorded on this machine when they happen, so their absence here means something.
It isn’t new #
When I first wrote this up, I assumed the shutdowns were recent. They aren’t. The journal reaches back to November 2025, and boots from July, August and early September all end the same way: cut off between ordinary ten-second pvestatd lines, with zero shutdown markers.
boot -20 Sep 01 19:44 -> Sep 03 10:37 (46 h) ends mid-stream, 0 shutdown markers
boot -11 Sep 11 16:35 -> Sep 11 18:25 ends mid-stream, 0 shutdown markers
boot -30 Aug 25 10:12 -> Aug 25 14:45 ends mid-stream, 0 shutdown markers
boot -50 Jul 18 12:51 -> Jul 18 12:52 clean poweroff, 2 shutdown markersThose boots have thousands of journal lines each, so it isn’t log rotation. Boot -50 is a genuine clean poweroff on the same system and firmware, and it shows the check works.
k3s didn’t cause this either. It was installed on 17 September, long after the first truncated boots. What k3s changed is the duty cycle. The fault had been there all along; it just started killing running VMs instead of an idle host.
The Proxmox watchdog is ruled out #
pve-guests.service (PVE 9’s watchdog) is enabled and active. A watchdog trip panics and reboots with a log entry, and there’s no such record.
The leading suspect: power with nothing behind it #
The kernel says it outright:
ACPI: AC: AC Adapter [ACAD] (on-line)
ACPI: battery: Slot [BAT1] (battery absent)There’s no battery. The only power source the kernel sees is the adapter. A laptop running this way has nothing between the wall and the CPU, so when the adapter dips there’s no ride-through: the machine dies instantly with nothing written to disk. That’s exactly the signature in the logs. An adapter that coped with an idle host may not hold up under sustained VM load.
The log can’t tell the brick from the DC connector from the board, because all three produce the same truncation. Only AC-side monitoring can narrow it down.
What I got wrong about panics #
In the first version of this post I argued that because the machine didn’t come back on its own, a kernel panic with auto-reboot was ruled out. That reasoning was wrong. Proxmox ships with kernel.panic = 0, which means “after a panic, wait forever”. A panicking host never reboots itself.
I found that out the hard way on 10 November (INC-2026-012). The host took a general protection fault in the CPU idle loop at 02:39 local time, panicked, and sat there powered on, with no video and no network, for about nine hours until I held the power button. That looks nothing like the power losses. A panic leaves the machine on and frozen, and the kernel saved its last messages to firmware storage (pstore), which survived the reset and made the diagnosis possible. A power loss leaves the machine off with a log that simply stops.
The host now reboots 10 seconds after any panic, oops or hard lockup (kernel.panic = 10, panic_on_oops = 1, hardlockup_panic = 1), set by an Ansible playbook that reads the values back and fails if they don’t hold. A fault in idle code with nothing running usually points to hardware, so a memtest86+ run is on the list too.
Where this stands #
The fan is understood: a worn 12-year-old fan, with replacement and fresh paste planned. The shutdown is understood well enough to act on: a hard power loss with no battery to ride through it.
Next steps #
- Replace the fan and repaste, and blow out the heatsink fins while it’s open. If the new fan is also loud, the suspect moves to the fins or the duct. Shut the VMs down cleanly first: with no battery and an unreliable supply, an accidental loss during the work would kill them the same ungraceful way.
- Fit a UPS or a logged AC-side power meter. It’s the only way to tell “the wall lost power” from “the machine dropped power”.
- Try a different charger brick and outlet, and reseat the DC connector.
- Capture
journalctl -b -1 -eright after any recurrence. - Flash a newer BIOS, but only in person. It’s the only fix for the missing
H_EC. Board ID080C1selects the image, and HP shipped refreshes for 2015 platforms as late as 2023. A failed flash on a headless machine with no IPMI is a brick, so this waits for physical access and a USB recovery stick.
A cheaper mitigation is possible now: a small unit that watches the ACAD ACPI event could shut the VMs down cleanly on AC loss. That limits the damage from a fault that’s already confirmed.
What I learned #
- Absence of evidence is evidence when you’ve proven the check works. A clean poweroff in boot
-50is what made “zero shutdown markers” mean something. - When nothing can measure the thing directly, measure its effect. No tach exists on this board, so a temperature delta under controlled load was the only way to prove a register didn’t drive the fan.
- Check your assumptions about timing. I assumed the shutdowns were new because I’d only just started paying attention. Eleven months of journal said otherwise.
- Know what a host does after a crash before you depend on it. Proxmox’s default after a panic is to wait for a human.
Architecture takeaways #
A laptop without a battery is a server without a power supply buffer #
The original appeal of a laptop as a server is the built-in UPS. Remove the battery and you have the opposite: a machine with no ride-through at all, where any dip at the adapter is an instant hard stop. Power delivery is part of the architecture, and on this host it’s the weakest part.
One physical host is one failure domain #
All three cluster nodes are VMs on this one machine, so a power blip or a panic takes the whole cluster down together. Proxmox snapshots and Terraform make the VMs easy to rebuild, but they don’t help when the hardware underneath them stops. That’s a deliberate trade-off at this budget, and it’s why recovery (nightly etcd snapshots with an offsite copy) matters more here than redundancy.
Remote recovery has to be designed in #
With no IPMI and no out-of-band management, the only remote recovery path is the one the OS sets up for itself. kernel.panic = 10 turns a nine-hour outage into a ten-second reboot, and Persistent=true on the backup timers turned the missed night into a late one rather than a lost one. Both cost nothing, and both only work if they’re in place before the failure.
What’s next #
Next in the series: the tag that meant two things, a week in which fixing my monitoring opened my admin tools to the least-trusted machine on my network. If you’re new to the series, start with the architecture overview.