Re: [PATCH v3] PCI: Disable spurious s2idle wake from the NVMe root port on ThinkPad P1 Gen 9
From: Kean Ren
Date: Tue Nov 17 2026 - 01:38:14 EST
Hi Bjorn,
Thanks — and thanks for Cc'ing the Legion folks.
I don't think it is the same bug, although the pattern is similar.
Both cases are firmware firing a GPE on the parent PCIe port of a device
that is entering a low-power state, and both workarounds target the parent
port's wake handling. The details differ:
- Legion: the trigger is the NVIDIA dGPU entering D3cold during runtime
suspend; the port (and then the GPU) is immediately woken again, in a
loop. The v4 patch removes the ACPI PM notifier from the parent port's
ACPI device, which turns the GPE handling into a no-op.
- ThinkPad P1 Gen 9 (21UF): the trigger is the primary NVMe SSD entering
D3 during system suspend (s2idle). Firmware leaves RP05's _PRW
enabled; the spurious PME sets GPE 0x6F, and the SCI arrives during the
noirq suspend phase and aborts s2idle after ~150 us. The quirk disables
wakeup on RP05 at late_initcall, so the port is not armed as a wake
source at suspend time. On this machine, masking GPE 0x6F or disabling
RP05 via /proc/acpi/wakeup keeps s2idle asleep; without the quirk every
suspend wakes after ~2.5 s.
So the two quirks are DMI/device specific and target different sleep paths
(runtime PM vs system s2idle). I think they should stay separate; I can add
a Link: to the Legion series in the changelog if you prefer a
cross-reference.
For the Legion folks: does acpi_remove_pm_notifier() also keep the port from
being armed as a wake source for system suspend, or does it only make the
runtime-PM GPE handling a no-op? In my case I don't think a notifier-only
change would have prevented the s2idle abort.
Thanks,
Kean