Very likely the same bug, on a different laptop (Lenovo Legion Pro 5 16ARX8, RTX 4060 Laptop, driver 615.78.08), reproduced 3/3. I traced it at the nvidia-powerd ioctl level and with its own verbose log. On the DC->AC transition JpacWorker logs ‘AC event 1’ and goes quiet, but its stop path (JpacWorker::postStop / stop) runs only on process exit. The 0x2080e0af calls with field[3]=0 happen inside that stop window. Details below, logs attached. All times are CET (UTC+1).
System
- Lenovo Legion Pro 5 16ARX8 (82WM), BIOS LPCN65WW (03/25/2026), BIOS graphics mode: dGPU only (panel on the NVIDIA GPU)
- AMD Ryzen 7 7745HX, NVIDIA GeForce RTX 4060 Laptop GPU, VBIOS 95.07.18.00.e0
- Driver 615.78.08, open kernel modules; nvidia-powerd 2.0 (build 1), enabled
- NVreg_DynamicPowerManagement=3; kernel 7.3.1 (CachyOS); KDE Plasma 6.8.2, Wayland
- Platform: cTGP 80 W + Dynamic Boost offset 55 W (enforced.power.limit 135-140 W on AC, 55 W on battery)
- Platform profile: performance
Steps to reproduce
- On AC, run a sustained 3D load (100% GPU utilization). GPU is in P0, memory clock 8001 MHz.
- Unplug AC for ~20 s while the load keeps running.
- Plug AC back in, keep the load running.
Result: 3/3 attempts (22:34, 22:48 and 22:58 on 2026-11-07; in the third, the load started while on battery). Earlier the same day the state also persisted for 4.5 hours
(from an unplug at 17:50 until nvidia-powerd was restarted at 22:23), through later AC plug/unplug cycles.
Not tested here: plug-in with the GPU idle (the linked thread reports that it recovers in that case).
Expected
After AC returns, the GPU goes back to P0 with the AC power policy.
Actual
- GPU stays in P3: memory clock 7001 MHz instead of 8001.
clocks_event_reasons = 0x24 (SW Power Cap + SW Thermal Slowdown), permanently. Both counters grow exactly 1:1 with
wall-clock time.
- enforced.power.limit goes back to 140 W.
- In the three repros the load was thermally bound (86-88 °C in both states), so power draw was similar stuck and healthy
(~60-68 W). The measurable difference was P3 / 7001 MHz / 0x24 vs P0 / 8001 MHz / 0x20.
- The 4.5 h episode, at lower temperature, showed the full impact: 100% utilization at 1155 MHz and ~24 W, 57-61 °C
(HW Slowdown, HW Thermal, Power Brake inactive).
sudo systemctl restart nvidia-powerd fixes it within ~1 s, even under load:
P3 / 7001 MHz / 0x24 → P0 / 8001 MHz / 0x20 (normal thermal target 87 °C).
Mechanism (ioctl trace of nvidia-powerd, repro2-sequence.txt)
All NV_ESC_RM_CONTROL calls from nvidia-powerd were captured with params before and after each call. Command IDs
not in the open headers are given as raw IDs.
| Phase |
What nvidia-powerd sends |
| AC, healthy |
only the periodic sampling calls; PERF_GET_POWERSTATE ~0.5/s |
| DC entry (22:48:00) |
0x2080e0d0 params {1, 9} (status 0x56). 0x2080e0af {3/2, 1, 0xf4/0xf8, 2, …}, then every 2 s. 0x20802068 PERF_GET_CURRENT_PSTATE 2/s. 0x2080a0a8 1/s. 0x13b NVPCF_GET_POWER_MODE_INFO every 4 s. PERF_GET_POWERSTATE 10/s |
| AC return (22:48:21) |
0x13b NVPCF_GET_POWER_MODE_INFO x5 (driver re-reads NVPCF func 1, 2, 9 from SBIOS: AC values). After this, 0x2080e0af / 0x20802068 / 0x2080a0a8 stop. No 0x2080e0af with field[3]=0, no 0x2080e0d0 {0, 9}, no 0x216. PERF_GET_POWERSTATE stays at 10/s with the driver answering AC |
| Stuck (22:48:21 to 22:49:59) |
nothing but the periodic sampling calls; GPU in P3 / 0x24 |
| nvidia-powerd SIGTERM (22:49:59.95, old PID, between “Stopping” and “Quit successfully”) |
0x216 x4, 0x2080e0d0 {0, 9}, 0x2080e0af {3/2, 1, 0xf4/0xf8, 0, …} → next 1 Hz sample (22:50:00.9) shows P0 / 8001 MHz / 0x20 |
| New instance (22:50:00) |
normal AC init |
- The last 0x2080e0af with field[3]=2 is at 22:48:20.79.
- There is no 0x2080e0af call at all between AC return (22:48:21) and exit (22:49:59.95).
- The DC->AC transition path does not send the exit-time release (0x2080e0af with field[3]=0, 0x2080e0d0 {0, 9}).
- With 1 Hz GPU sampling I cannot separate the exit-time release from the new instance’s init (~50 ms apart) as
what restores P0. But see repro3 below: the 0x2080e0af calls at exit fall inside JpacWorker’s stop window.
- 0x2080e0d0 returns NV_ERR_NOT_SUPPORTED (0x56) on this system in every phase. Possibly harmless, but maybe part of
why the AC path takes a different branch.
nvidia-powerd’s own verbose log (repro3-powerd-verbose.log)
nvidia-powerd only logs to syslog, and its log level is a global set to 2 in main() (non-PIE binary, address 0x6983a8;
76 messages are gated at level >= 3/4). I found no env var or command-line switch that raises it. For the third repro (3/3, 22:57:56 unplug,
22:58:21 plug-in, GPU load started on battery and kept through the plug-in) I set it to 4 with gdb in the running process.
- Battery:
INFO: Switched from AC to battery, JpacWorker AC event 0., JpacWorker NVPCF event 1., then one
INFO: JPAC prevGpcclkMHz: ..., fpsCap ..., igp.getTgpPwrAvgW(): ..., ljp2.m_limit / 1000.0f: ... line every second.
- AC return:
INFO: Switched from battery to AC, JpacWorker AC event 1., JpacWorker NVPCF event 1.
- After this, the JPAC lines stop; only
JpacWorker::update() once per second.
- No
JpacWorker::postStop() / JpacWorker stop... and no other message about releasing the limit.
- GPU stays in P3 / 0x24.
- Exit (systemctl restart, 22:59:32):
systemState stop 1, Uninitialize RM Client, JpacWorker::postStop(),
JpacWorker stop..., JpacWorker stop completed., JpacWorker::postUninitialize(), QBoost is STOPPED now.
→ GPU back to P0 / 8001 MHz.
In the same run the ioctl trace was active:
- JpacWorker::postStop() at 22:59:32.5394, then “stop completed” at 22:59:32.5473.
- The two 0x2080e0af calls at exit fall inside that window, at 22:59:32.5397 and 22:59:32.5415.
So on the AC event JPAC stops updating, but its stop path, which issues those 0x2080e0af calls, runs only on process exit.
Frame-rate limiter is released correctly (repro3-frl-packets.txt, D-Bus nvidia.powerd.frl.limit / Frlpacket):
- On battery, a packet every 2 s with the frame-time limit, for example 16881 us = 59.2 FPS =
fpsCap 59.2 in the log.
- At AC return (22:58:22.324), a
0 0 packet is sent.
- So the FRL side is released on AC. The RM-side request (0x2080e0af) is what stays behind.
Other evidence
- PERF_GET_POWERSTATE (0x2080205a) answers (repro-timeline.txt): BATTERY on unplug, AC from ~1 s after plug-in.
The driver learns AC correctly.
- ACPI path is complete on both transitions:
- EC query → Notify(ADP0, 0x80) → nv_acpi_powersource_hotplug_event(0x80) → Notify(NPCF, 0xC0) → nv_acpi_nvpcf_event(0xC0).
- No GPU D-notifier (Notify PEGP 0xD1-0xD5) on either transition; EC D-notifier level is D1.
- SBIOS NVPCF replies (nv_acpi_dsm_method, repro2-sequence.txt). Func 2 dynamic params:
- DC: TGPA 0x0000, TPPA 0x0280, MAGA 0x00c8.
- AC: TGPA 0x0208, TPPA 0x01b8, MAGA 0x00a0.
- DROP 0x88b8 in both.
- Func 9 polled every 4 s on DC.
- Lenovo WMI: cTGP 80 W, DB offset 55 W. Per the DSDT logic (custom mode off), GPS func 0x2A gives a thermal target of 87 °C. This was not captured at runtime.
- At its default log level nvidia-powerd logs nothing beyond “nvidia-powerd version:2.0 (build 1)” /
“DBus Connection is established”, also with __RM_ENABLE_VERBOSE_OUTPUT=1 in its environment.
/var/log/nvtopps/ exists but stays empty; strace shows the daemon only stat()s it and logs via /dev/log.
- At driver load, on every boot: “PlatformRequestHandler failed to get target temp from SBIOS” and
“failed to get platform power mode from SBIOS” (NV_ERR_INVALID_DATA). The GPU works normally before the AC event,
so possibly unrelated.
Notes
- nvidia-powerd.service ships with StartLimitIntervalSec=infinity and StartLimitBurst=5.
- So restart-based workarounds (such as the udev rule suggested in the linked thread) hit start-limit-hit after
five starts per boot.
- We override it with StartLimitIntervalSec=60.
- Workaround in use here: a udev rule on power_supply Mains “change” that starts a oneshot unit. The unit waits 5 s
(to coalesce adapter flapping) and runs systemctl try-restart nvidia-powerd.
Attachments
All bug reports anonymized:
- Hostname, user name, serial numbers, UUIDs, MAC/IP addresses and sudo lines replaced.
- Monitor EDIDs, the plasmashell journal and the encrypted NVML log removed.
- nvidia-debugdump archive kept as is.
All files are in the attached nvidia-powerd-ac-bug-report.tar.gz (folder nvidia-report/, also contains this text as REPORT.md):
nvidia-bug-report-stuck.log.gz: taken while stuck (first repro, 22:35, between plug-in and restart).
nvidia-bug-report-healthy.log.gz: taken in the healthy state after the second repro (22:50), for diffing.
repro-timeline.txt: first repro. 1 Hz GPU state, ACPI/EC/driver notify trace, PERF_GET_POWERSTATE rate and answers.
repro3-powerd-verbose.log: third repro, nvidia-powerd journal at log level 4, from unplug to exit.
repro3-frl-packets.txt: third repro, decoded Frlpacket D-Bus signals.
repro2-sequence.txt: second repro. Every non-periodic nvidia-powerd RM control with params in/out and status,
SBIOS _DSM replies, ACPI events, 1 Hz GPU state, including the exit-time release sequence.
nvidia-powerd-ac-bug-report.tar.gz (1.3 MB)