Hi team,

I am reporting a regression in the AMDGPU driver affecting the Strix Halo
APU (Radeon 8060S, gfx1151). While everything works correctly on kernel
7.1.0-rc5, upgrading to 7.1.0-rc7 causes GPU memory operations to hang
indefinitely. This occurs during tasks like torch.empty() or model weight
transfers in ComfyUI (PyTorch 2.11.0+rocm7.13).

I have bisected the changes in drivers/gpu/drm/amd/ between rc5 and rc7 and
identified the following potential causes:

1.  amdgpu_hmm.c (Christian König):

  - 1c824497d: Changing the invalidate callback to wait on the VM root BO
reservation lock may be introducing a deadlock.
  - 962d684b5: Moving the notifier_seq read outside the retry loop could
cause infinite retries with a stale sequence number.
  - 58bafc666: Changes to userptr submission waiting.

2.  gfxhub_v12_0.c (Timur Kristóf):

  - 40bab7c60: The change to CRASH_ON_*_FAULT bits might be causing the GPU
to retry failed memory accesses indefinitely rather than surfacing a fault.

3.  gmc_v12_0.c (Harish Kasiviswanathan):

  - ae4e30f24 and e3fa02872: If the new per-version PTE address masks for
gfx1151 are incorrect, it could result in corrupted page table entries.

4.  amdgpu_gart.c (Donet Tom):

  - ec4c462e2: The updated PTE iteration grouping may be producing
incorrect page tables when combined with the new PTE mask.

Downgrading to 7.1.0-rc5 resolves the issue. Please let me know if you
require any specific debug output or further testing.

Best regards,
Jonathan

Reply via email to