Claude review: drm/amdgpu: fix sleeping allocation under spinlock in PASID IDR

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

Patch Review

**Analysis of the bug:** Confirmed by reading the current tree at `amdgpu_ids.c:70-73` — the code currently calls `idr_alloc_cyclic()` with `GFP_KERNEL` while holding `amdgpu_pasid_idr_lock` (a spinlock). This is indeed a sleeping-in-atomic-context bug.

**Analysis of the fix:**

```c
+	idr_preload(GFP_KERNEL);
 	spin_lock(&amdgpu_pasid_idr_lock);
 	pasid = idr_alloc_cyclic(&amdgpu_pasid_idr, NULL, 1,
-				 1U << bits, GFP_KERNEL);
+				 1U << bits, GFP_NOWAIT);
 	spin_unlock(&amdgpu_pasid_idr_lock);
+	idr_preload_end();
```

This is textbook correct:
1. `idr_preload(GFP_KERNEL)` pre-allocates radix tree nodes (can sleep, called outside spinlock) and disables preemption.
2. `GFP_NOWAIT` inside the spinlock ensures the allocator uses the preloaded pool and never sleeps.
3. `idr_preload_end()` re-enables preemption after the spinlock is released.

**One minor ordering note:** `idr_preload_end()` is called after `spin_unlock()`, which is correct — `idr_preload()` disables preemption, so you must hold the preload context across the allocation. The unlock-before-preload_end ordering is the standard pattern (see e.g., `drivers/infiniband/`, `net/core/`).

**Fixes tag and Cc: stable** are both appropriate since this is a straightforward bugfix for a commit that introduced the regression.

**No issues found.** 

Reviewed-by recommendation: This patch is correct and follows established kernel patterns.

---
Generated by Claude Code Patch Reviewer