https://bugzilla.kernel.org/show_bug.cgi?id=214963
Bug ID: 214963 Summary: [amdgpu] resuming from suspend fails when IOMMU is missing Product: Drivers Version: 2.5 Kernel Version: 5.15.0, 5.15.1 Hardware: All OS: Linux Tree: Mainline Status: NEW Severity: normal Priority: P1 Component: Video(DRI - non Intel) Assignee: drivers_video-dri@kernel-bugs.osdl.org Reporter: spasswolf@web.de Regression: No
On the HP bw064-ng Laptop the BIOS does not correctly initialize the IOMMU. This brakes resuming from suspend: Nov 7 23:23:02 bart kernel: [ 96.378265] kfd kfd: amdgpu: Failed to resume IOMMU for device 1002:9874 Nov 7 23:23:02 bart kernel: [ 96.378271] amdgpu 0000:00:01.0: amdgpu: amdgpu_device_ip_resume failed (-6). Nov 7 23:23:02 bart kernel: [ 96.378274] PM: dpm_run_callback(): pci_pm_resume+0x0/0x110 returns -6 Nov 7 23:23:02 bart kernel: [ 96.378283] amdgpu 0000:00:01.0: PM: failed to resume async: error -6
leading to the follow up errors: Nov 7 23:23:03 bart kernel: [ 97.580799] [drm] perform_link_training_with_retries: Link training attempt 1 of 4 failed ... Nov 7 23:23:04 bart kernel: [ 98.646241] [drm] perform_link_training_with_retries: Link training attempt 2 of 4 failed ... Nov 7 23:23:05 bart kernel: [ 99.709472] [drm] perform_link_training_with_retries: Link training attempt 3 of 4 failed Nov 7 23:23:06 bart kernel: [ 100.772840] [drm] enabling link 0 failed: 15 Nov 7 23:23:22 bart kernel: [ 106.160854] [drm:amdgpu_dm_atomic_commit_tail [amdgpu]] *ERROR* Waiting for fences timed out! Nov 7 23:23:22 bart kernel: [ 116.397484] [drm:drm_atomic_helper_wait_for_flip_done] *ERROR* [CRTC:47:crtc-0] flip_done timed out
Simply changing the amdgpu_device_ip function in drivers/gpu/drm/amd/amdgpu/amdgpu_device.c to static int amdgpu_device_ip_resume(struct amdgpu_device *adev) { int r;
#if 0 r = amdgpu_amdkfd_resume_iommu(adev); if (r) return r; #endif
r = amdgpu_device_ip_resume_phase1(adev); if (r) return r;
r = amdgpu_device_fw_loading(adev); if (r) return r;
r = amdgpu_device_ip_resume_phase2(adev);
return r; } makes resuming from suspend work for me again. But of course this is not a viable solution as on systems with working IOMMU this is needed. One could probably add a flag so that in case of missing IOMMU amdgpu_amdkfd_resume_iommu returns 0.