Claude review: drm/buddy: Integrate lockdep annotations for gpu buddy manager

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

Patch Review

**Positive aspects:**
- Pattern is well-established — directly follows `drm_gpusvm` precedent
- `lock_dep_map` field is `#ifdef CONFIG_LOCKDEP` guarded, so zero overhead on production builds
- The macro uses a local variable `__mm` to avoid double-evaluation, which is slightly better than the gpusvm original
- Correct use of `lock_is_held_type(map, 0)` for exclusive (mutex) lock assertion

**Issue 1 (Major): Missing driver callers**

Only xe gets `gpu_buddy_driver_set_lock()`, but amdgpu and i915 also use the buddy allocator with their own mutexes. Without wiring those up, the lockdep annotation does nothing for those drivers and `gpu_buddy_driver_lock_held()` will silently skip the assertion (because `lock_dep_map` is NULL).

```c
// xe_ttm_vram_mgr.c — wired up (good)
gpu_buddy_driver_set_lock(&mgr->mm, &mgr->lock);
```

Missing from `amdgpu_vram_mgr.c` (after `gpu_buddy_init` at line 934):
```c
err = gpu_buddy_init(&mgr->mm, man->size, PAGE_SIZE);
if (err)
    return err;
// needs: gpu_buddy_driver_set_lock(&mgr->mm, &mgr->lock);
```

Missing from `i915_ttm_buddy_manager.c` (after `gpu_buddy_init` at line 297):
```c
err = gpu_buddy_init(&bman->mm, size, chunk_size);
if (err)
    goto err_free_bman;
mutex_init(&bman->lock);
// needs: gpu_buddy_driver_set_lock(&bman->mm, &bman->lock);
```

Missing from `ttm_mock_manager.c` (test code, lower priority but still should be consistent).

**Issue 2 (Medium): Missing annotations on exported functions**

`gpu_buddy_fini()` and `gpu_buddy_block_print()` are exported and access manager state, but are not annotated:

```c
// buddy.c line 461 — no gpu_buddy_driver_lock_held(mm)
void gpu_buddy_fini(struct gpu_buddy *mm)
{
    ...
}
EXPORT_SYMBOL(gpu_buddy_fini);

// buddy.c line 1458 — no gpu_buddy_driver_lock_held(mm)
void gpu_buddy_block_print(struct gpu_buddy *mm, ...)
{
    ...
}
EXPORT_SYMBOL(gpu_buddy_block_print);
```

Similarly, `drm_buddy_block_print()` in `drm_buddy.c` is exported and reads buddy state but is not annotated (while `drm_buddy_print()` in the same file *is* annotated).

For `gpu_buddy_fini`, the annotation is debatable — typically teardown happens after the lock is no longer needed. But `gpu_buddy_fini` calls `__force_merge` which modifies the buddy tree, so either annotate it or add a comment explaining why it's intentionally excluded.

**Issue 3 (Minor): Inconsistent no-LOCKDEP stub style**

```c
// gpu_buddy_driver_set_lock non-lockdep stub:
#define gpu_buddy_driver_set_lock(mm, lock) do { (void)(mm); (void)(lock); } while (0)

// drm_gpusvm_driver_set_lock non-lockdep stub for comparison:
#define drm_gpusvm_driver_set_lock(gpusvm, lock) do {} while (0)
```

The `(void)` casts are not wrong (they suppress unused-variable warnings), but the gpusvm template the code is modeled after doesn't use them. This is a matter of style consistency — either approach works.

**Issue 4 (Minor): `gpu_buddy_init` NULL initialization placement**

```c
+#ifdef CONFIG_LOCKDEP
+	mm->lock_dep_map = NULL;
+#endif
 	return 0;
```

This is placed at the end of the success path in `gpu_buddy_init`, which is fine. However, the `drm_gpusvm` version also places it at the end of initialization. Correct and consistent.

**Issue 5 (Nit): Commit message says "Assisted-by: Claude"**

```
Assisted-by: Claude:claude-opus-4.6
```

The standard tag for AI assistance is not established upstream. This may need adjustment based on maintainer preferences. At minimum the format should probably be `Assisted-by: Claude (claude-opus-4.6)` or similar human-readable form rather than the colon-separated identifier.

**Summary**: The core design is correct and follows established kernel patterns. The main issue is that the patch is incomplete — it needs to wire up all gpu_buddy consumers (amdgpu, i915, test mock) and annotate all exported functions, or explicitly justify why some are excluded. I'd suggest splitting this into a series: patch 1 for the infrastructure + buddy.c/drm_buddy.c annotations, then one patch per driver wiring up the lock registration.

---
Generated by Claude Code Patch Reviewer