dGPU silently removed from PCI bus while in D0 — driver never notified (no nv_pci_remove, no Xid), stale state forever

[Kiraprint]

NVIDIA Open GPU Kernel Modules Version

610.57.04 (release build, notroot@, Aug 5 2026)

Please confirm this issue does not happen with the proprietary driver (of the same version).

• I confirm this does not happen with the proprietary driver package.
  (Not tested — only the open module is installed on this system.)

Operating System and Version

CachyOS (Arch-based), kernel 7.2.0-rc6-1-cachyos-rc

Hardware: GPU

NVIDIA GeForce RTX 5090 Laptop GPU (10de:2c58) on ASUS ROG Zephyrus G16 (GU605CX), at PCI 0000:01:00.0 under root port 0000:00:01.0. Intel Arc iGPU is the display GPU.

Describe the bug

Switching the GPU mode to "integrated" (via cardwire, a supergfxctl successor that only eBPF-blocks process access to the dGPU — it writes nothing to sysfs) makes the dGPU disappear from the PCI bus within 1–2 seconds, while fully in D0. The driver is never notified:

• nv_pci_remove is never called (kprobe-verified), no Xid, no kernel message of any kind (dmesg stays empty even with NVreg_EnableStreamDebug=1 and NVreg_EnableGPUResetLogging=1)
• nv_pmops_runtime_suspend / nv_pmops_runtime_resume are never called — this is not a runtime-D3 transition (kprobe-verified; power/control=on does not prevent it)
• nv_acpi_d3cold_dsm_for_upstream_port is never called
• cardwire's power tracker polls power_state throughout the event: it goes D0 → device gone, never D3Cold
• Result: /proc/driver/nvidia/gpus/0000:01:00.0 persists forever, power state frozen at "Video Memory: Active", /dev/nvidia0..nvidiactl persist, and the rail stays powered (GPU temp ~53C, ~20W on AC) while the device is "gone"

The device is revived (re-enumerates, driver re-binds) by switching back to "hybrid" or replugging AC — no reboot needed.

To Reproduce

1. Boot hybrid mode (dGPU on the bus, driver bound, NVreg_DynamicPowerManagement=0x02)
2. cardwire set integrated (or unplug AC — cardwire auto-switches to integrated)
3. Within ~1–2 s: lspci no longer lists 0000:01:00.0; driver frozen as described
4. cardwire set hybrid → device returns

Same result with NVreg_EnableS0ixPowerManagement=1 and with power/control=on pinned.

Bug Incidence

Always (10+ reproductions, identical).

nvidia-bug-report.log.gz

Full bug report + strace of the daemon + evidence write-up (kprobe setup/results, dmesg, timeline):
https://gist.github.com/Kiraprint/d7c4441a1d3f3bf58f6bb1a84fe4fb9e

More Info

The removal is performed by the platform/EC with no software-layer participation: not the driver (0 kprobe events), not the RM/GSP firmware (StreamDebug dmesg empty), not kernel PCI PM (control=on irrelevant). The trigger is the fully-blocked+idle state. The defect we'd like fixed: the driver does not handle device loss — no remove callback, no Xid, no cleanup, perpetual stale device state — whereas a graceful "GPU has fallen off the bus" (Xid 79-style) path would let the driver clear the device and the user recover without reboot.