A range-restricted allocation walks every split block in the range, so its cost grows linearly with the number of live allocations there. amdgpu restricts every VRAM allocation to its memory partition under NPS2/NPS4, so on MI300X in CPX/NPS4 KFD allocation latency grows with every live BO, under a VRAM manager lock shared by the compute partitions.
Patch 1 narrows the walk to the highest free block wholly inside the range, found in the offset-sorted free trees. Patch 2 adds a KUnit test for the NPS4 pattern (range-biased, cleared, top-down, 64 KiB) that also checks placement: 1190 ms before, 3 ms after. v2: - rebased onto drm-misc-next (single free tree) (Arun) - take the highest contained block, keeping the walk's top-down placement, including for GPU_BUDDY_TOPDOWN_ALLOCATION (Arun) - the test checks top-down placement - placement differs from the full walk only for blocks straddling the upper range edge and for clear steering across blocks (see patch 1) 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 performance test drivers/gpu/buddy.c | 35 +++++++++++++++++++++++++++++- drivers/gpu/tests/gpu_buddy_test.c | 27 +++++++++++++++++++++++ 2 files changed, 61 insertions(+), 1 deletion(-) base-commit: 6f7098738fff3ec7393f924343188cb8e6ebc5d2 -- 2.43.0
