In a range-restricted allocation, the strict pass of __alloc_range_bias() still visits every free block of the other class in the range: a dirty request among many free cleared blocks steps through each of them. amdgpu restricts every VRAM allocation to its memory partition under NPS2/NPS4, so it takes this path for every KFD VRAM allocation there.
Patch 1 replaces the walk with a lookup in the offset-sorted free trees, on top of Arun's range guard code, and drops subtree_max_free_order, which nothing reads afterwards. Patch 2 adds tests for an inverted range and for the NPS4 fill pattern. v5: squash v4 patches 1 and 2 (Arun); no code change. v4: https://patchwork.freedesktop.org/series/178969/ v3: https://patchwork.freedesktop.org/series/178108/#rev2 v2: https://patchwork.freedesktop.org/series/178108/#rev1 v1: https://patchwork.freedesktop.org/series/176957/ William Palacek (2): gpu/buddy: find range-biased blocks in the free trees gpu/tests/buddy: add range-biased allocation tests drivers/gpu/buddy.c | 334 +++++++++++++++++++---------- drivers/gpu/tests/gpu_buddy_test.c | 67 ++++++ include/linux/gpu_buddy.h | 7 - 3 files changed, 291 insertions(+), 117 deletions(-) -- 2.43.0
