Fan control

Code: FanCurve, Zone, Daemon.decide/apply/step, failsafe() in usr/sbin/qnapd. Tests: TestCurve, TestDaemon, TestQm2 in tests/test_qnapd.py.

Zones and sources

Zone	Fans	Inputs	Floor
chassis	every qnap8528 pwmN group	CPU, EC system temps, HDD (S.M.A.R.T.), NVMe	MIN_PWM_CHASSIS = 40
one per QM2 card	the card’s pwm1	its slot sensors + NVMe SSDs behind its PCIe switch	MIN_PWM_QM2 = 64

• CPU: coretemp Package id 0 (or k10temp Tctl/Tdie), median over 60 s (CPU_SMOOTH_SECONDS) for the curve; the raw value for the critical check.
• System: every qnap8528 tempN_input within 5–110 °C (junk filtered).
• Disk: smartctl -n standby -A -H -j per non-removable sdX, every diskinterval s (default 60); standby disks are skipped. An awake disk whose read/write counters have not moved since its last query keeps its last reading and is re-queried only every 30 min (DiskTemps.IDLE_REFRESH_SECONDS): a S.M.A.R.T. read restarts the drive’s standby timer, so polling idle disks every minute kept them from ever spinning down (measured on the reference box, 2026-10-17). USB disks (sysfs path through a usbN controller) are left out: they sit outside the chassis airflow, so the fans cannot cool them. The reference box’s USB boot SSD reports itself as non-removable, hence the path check.
• NVMe: each nvme hwmon temp1 (Composite). The chassis zone sees every NVMe; a QM2 zone sees only the ones under its switch (by sysfs path).
• QM2 slot sensors use the NVMe range and critical limit.

Algorithm (every interval s, default 10)

1. Critical (raw readings): CPU 90, system 55, disk 55, NVMe/QM2 75 °C → PWM 255, in every mode.
2. Manual mode → manualpwm.
3. No sensors for an enabled zone → 255 and a fault.
4. Curve: per source demand = clamp((T − low)/(high − low), 0, 1); pwm = minpwm + max(demand) × (maxpwm − minpwm).
5. Ramp up immediately, remembering which source raised the fan and its temperature then.
6. Ramp down only once that source has cooled by hysteresis °C, then by at most 16 PWM per cycle.
7. Floor: never below the zone’s floor, whatever the mode.
8. Write PWM when changed, at least every 60 s, and at once if the PWM read back differs from the last write by more than 3 (qnap8528 rounds to percent).
9. Fan failure: 0 RPM at PWM ≥ 64 (or 65535) for 3 cycles → fault → status LED red.

Presets

Mode	CPU °C	System °C	Disk °C	NVMe/QM2 °C	PWM range
Quiet	60→90	35→50	40→50	50→70	48–200
Normal	55→85	30→45	38→48	45→65	64–255
Performance	50→80	28→40	35→45	40→60	100–255

Custom mode uses the datamodel fields (cpulow…nvmehigh, minpwm, maxpwm, hysteresis); defaults match Normal.

Failsafe

Event	Chassis	QM2 cards (useqm2 on)	Status LED
Clean stop/restart (SERVICE_RESULT=success)	255	card’s own controller (fan_default)	unchanged
Crash / kill / failure	255	255	blinking red
Exception in a loop iteration	255 for that cycle	255	—
qnapd failsafe by hand	255	255	—
useqm2 off	—	handed to the card at start, never written	—

The unit has no start limit (StartLimitIntervalSec=0): every Apply restarts the daemon, and a limit of 5 starts per 5 minutes left it dead during testing.

Lessons (each was a real bug or complaint on hardware)

• Hysteresis must belong to the source that raised the fan. Adding the hysteresis to every source pinned the fan at PWM 140 after a CPU burst: +3 °C on the 10 °C disk range is ≈57 PWM, and a disk at 39 °C that never caused the increase held it up.
• Smooth the CPU, and use a median. Raw J4125 readings spike for single samples; the fan sawtoothed between PWM 128 and 217. A 30 s mean still let frequent spikes through; a 60 s median fixed it.
• Every fan needs a floor in every mode, from its measured restart PWM plus margin. Quiet at 40 would have stalled the QM2 fan (stall 48).
• Do not blast a fan on routine restarts. Full speed on every Apply was audible; QM2 fans now return to the card’s controller on clean stops.
• A loaded UI skews measurements. Close the OMV dashboard before tuning.
• Verify writes. In the kill -9 failsafe test the QM2 PIC ignored the first PWM write after the restart (the I2C write reported success) and stayed at full speed until the 60 s refresh. The daemon now compares the read-back PWM every cycle and rewrites on a mismatch.

Measurements behind these numbers: ts-253d.md.