Claude review: gpu/buddy: Per-order free and used block scoreboards

[Claude Code Review Bot <claude-review@example.com>]

Overall Series Review

Subject: gpu/buddy: Per-order free and used block scoreboards
Author: Francois Dugast <francois.dugast@intel.com>
Patches: 9
Reviewed: 2026-05-25T19:06:37.067104

---

This is a well-structured 5-patch series that adds O(1) per-order free and used block counting to the gpu_buddy allocator. The series has good logical decomposition: bug fix first (patch 1), dead-code cleanup (patch 2), refactoring (patch 3), then the two feature patches (patches 4 and 5). Patches 1, 3, and 5 already carry `Reviewed-by: Matthew Auld`.

**Patch 1 is a genuine bug fix** — the existing error paths call `__gpu_buddy_free()` on a block that is still linked in an rbtree, which can corrupt the tree or cause use-after-free during coalescing. This is the most important patch in the series and should be considered for stable.

**The scoreboard approach is sound.** The incremental maintenance at the four state-transition points (mark_free, mark_allocated, mark_split, and the direct rbtree_remove sites) correctly mirrors the existing tree operations. The print functions become trivial O(1) lookups, which is a real win on large VRAM heaps.

**One concern worth raising:** the `free` vs `block_size` threshold change in patch 5 (from `free < SZ_1M` to `block_size < SZ_1M`) changes the KiB/MiB display logic — this is intentional and correct (avoids showing "0 MiB" for orders with zero blocks), but it changes debugfs output format which could break any userspace parsing.

Overall the series looks good. A few observations per patch below.

---
Generated by Claude Code Patch Reviewer