drm/amdkfd: relax checks for over allocation of save area

[Benjamin Wheeler]

This is a tracking bug for inclusion of the upstream Linux kernel commit "drm/amdkfd: relax checks for over allocation of save area" into [Q/R] linux packages.

SRU Justification:

[Impact]

This will fix issues with certain AMD Strix Halo APUs running into hangs when paired with particular versions of AMD ROCm.

[Fix]

Include commit d15deafab5d722afb9e2f83c5edcdef9d9d98bd1 from the upstream mainline linux kernel into our generic kernels.

[Test Plan]

The reproducer script below should be sufficient to confirm the bug is no longer present:

export PYTORCH_ROCM_ARCH=gfx1151
export TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1
uv init --python 3.13
uv add --index therock_nightly=https://rocm.nightlies.amd.com/v2/gfx1151/ --index-strategy unsafe-best-match --prerelease allow "rocm-sdk-core==7.10.0a20251015" "rocm[libraries]==7.10.0a20251015" "torch==2.10.0a0+rocm7.10.0a20251015" "torchvision==0.25.0a0+rocm7.10.0a20251015" "torchaudio==2.8.0a0+rocm7.10.0a20251015" "pytorch-triton-rocm==3.5.0+gitb0cf18f2.rocm7.10.0a20251015"
uv add git+https://github.com/huggingface/diffusers.git
uv add git+https://github.com/ivanfioravanti/qwen-image-mps.git
uv run qwen-image-mps edit -i mushroom1.png mouse1.png -p "The mouse is under the mushroom."

(source: https://github.com/ROCm/ROCm/issues/5590#issuecomment-3481580910)

We don't have any Strix Halo hardware to test this fix on, so there currently is no test plan. I am currently investigating some Kraken Point APUs that may be suitable, if I am able to provision them with a Questing kernel. Currently, because they install the OEM kernel, they always install noble. WIP.

[Where problems could occur]

This could introduce regressions for folks utilizing Strix Halo APUs with ROCm. They may experience issues with version desync, as this patch is intended to be paired with a user-space update of ROCm. (See https://github.com/ROCm/rocm-systems/commit/770f30bc4c72d763742e39932e2c0583813d531f for details on the exact version.)

[Benjamin Wheeler (benjaminwheeler)]

Resolute will already have this by virtue of using a new enough kernel version. This patch will land in 6.18.