Root cause found (pstore + sysrq-w traces, attached): After S3/s2idle resume, pciehp on the dGPU root port 0000:00:01.0 (soldered RTX 3060, but the port advertises HotPlug+ Surprise+) reports a spurious "Slot(1): Card not present" and tries to remove the GPU functions while dpm_resume() is still running and user space is frozen. This deadlocks with async resume of the same device:
- nvidia loaded: pciehp -> nv_pci_remove -> nvkms down_read (blocked) - nvidia blacklisted: pciehp -> azx_remove [snd_hda_intel 01:00.1] -> snd_card_free -> wait_for_completion (blocked) - in both cases: async_resume -> device_resume -> mutex_lock (blocked) and systemd-sleep in dpm_resume -> async_synchronize_full The GPU is actually present and works after resume when PCIe hotplug is disabled. Working workaround (both s2idle and deep S3 work, nvidia-smi OK): pcie_ports=compat acpiphp.disable=1 Not sufficient: pcie_ports=compat alone (acpiphp then removes the device -> same deadlock). Earlier note about D3hot vs D3cold was wrong: d3cold_allowed on ports has no effect on the ports' own D3 state. ** Attachment added: "pstore log (sysrq-w), nvidia loaded: pciehp -> nv_pci_remove deadlock" https://bugs.launchpad.net/ubuntu/+source/linux/+bug/2174812/+attachment/6019427/+files/pstore-nvidia-loaded.txt -- You received this bug notification because you are a member of Ubuntu Bugs, which is subscribed to Ubuntu. https://bugs.launchpad.net/bugs/2174812 Title: Lenovo Legion 5 15IAH7 (82RB): suspend hangs at pm_test=platform, fixed by pcie_port_pm=off To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/linux/+bug/2174812/+subscriptions -- ubuntu-bugs mailing list [email protected] https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
