Claude review: drm/amdgpu: introduce SVM range migration decision layer

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

Patch Review

**Overall:** Clean, minimal patch. The decision layer is well-factored.

**Issue: `range_needs_migrate_to_vram` doesn't check for `AMDGPU_SVM_MIGRATE_TO_VRAM` explicitly:**
```c
+static bool
+range_needs_migrate_to_vram(struct drm_gpusvm_range *range,
+			    enum amdgpu_svm_migrate_mode mode)
+{
+	if (!range->pages.flags.migrate_devmem)
+		return false;
+
+	if (mode == AMDGPU_SVM_MIGRATE_TO_SYSMEM)
+		return false;
```
This means if `mode == AMDGPU_SVM_MIGRATE_NONE` (which is excluded by the WARN_ON caller, but still), it would return true. The function is only called from `amdgpu_svm_range_migrate_range()` which validates mode, so this is safe in practice but the logic could be clearer.

**Nit:** `amdgpu_pagemap_capable()` checks `adev->gmc.is_app_apu` — the name `is_app_apu` seems like an unusual field. Confirm this is the right check for "this device can do VRAM migration."

**Question:** The `AMDGPU_SVM_MIGRATE_PREFERRED` mode is defined in the enum but never handled in `amdgpu_svm_range_migrate_range()` — it triggers the WARN_ON. The resolution happens in the caller (`amdgpu_svm_range_map` in patch 5). This layering is correct but the docstring on the enum should make it even more explicit that PREFERRED must be resolved before calling migrate_range.

---
Generated by Claude Code Patch Reviewer