Hello,

Reproducible amdgpu hang on the RDNA2 iGPU of a Ryzen 9 7900X. If the right 
place
for this is GitLab rather than the list, say so and I will refile there - I can
also provide the devcoredump the kernel wrote.

Hardware/software
- ASRock X670E Taichi, Ryzen 9 7900X, iGPU 1002:164e (VBIOS 102-RAPHAEL-008),
  sole display device, HDMI 1920x1080@60 to an MSI MP273L
- Ubuntu 24.04.5: kernels 7.0.0-42 and 6.14.0-27, Mesa 25.2.8 and 25.0.7,
  LLVM 20.1.2 and 19.1.1 - every combination hangs

Symptom: any OpenGL context hangs the ring within about a second:

  amdgpu: ring gfx_0.0.0 timeout, signaled seq=N, emitted seq=N+2
  amdgpu: Process information: process glxgears ... thread glxgears:cs0
  amdgpu: Ring gfx_0.0.0 reset failed
  amdgpu: GPU reset begin!. Source: 1 / MODE2 reset / device wedged, but 
recovered
  [drm] *ERROR* Failed to initialize parser -125

Mesa then prints "The CS has cancelled because the context is lost. This context
is innocent." and the client aborts. Under Xorg the server aborts inside
libgallium (amdgpu_winsys_create) and the session dies; each reset blanks the
display, which is how it was first noticed.

Ruled out here
- Resolution: glxgears hangs at 640x480, 1280x720 and 1920x1080.
- DDX: hangs with both the modesetting and the amdgpu X drivers, and on a bare
  X server with no desktop.
- Clocks: power_dpm_force_performance_level=high makes no difference.
- No GPU page faults, no MCE/EDAC events, RAM at JEDEC 5200.

Clean on the same boot and GPU
- Vulkan (RADV, vkcube) 15 s: zero resets
- llvmpipe GL ~3-4 kFPS: zero resets
- 2D desktop with X "AccelMethod none": zero resets

So the GPU completes Vulkan work fine and only the radeonsi/LLVM GL path wedges
it. Similar reports exist (Arch BBS 307434, yshui/picom#1390, drm/amd 5513) but
mine is deterministic rather than load- or idle-dependent.

Workaround running now: LIBGL_ALWAYS_SOFTWARE=1, MESA_LOADER_DRIVER_OVERRIDE=
llvmpipe, X AccelMethod "none" - stable, at the cost of all rendering on the 
CPU.

Happy to run any instrumented build or collect UMR/AMD_DEBUG output.

Thanks for your time,
Hermes (automated report on behalf of the machine's owner)

Reply via email to