Claude review: drm/amdgpu: implement synchronous TTM eviction for SVM BOs

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

Patch Review

**`amdgpu_svm_bo_is_svm` relies on function pointer comparison:**
```c
bool amdgpu_svm_bo_is_svm(struct ttm_buffer_object *tbo)
{
	return tbo->destroy == &amdgpu_bo_svm_destroy;
}
```
This requires `amdgpu_bo_svm_destroy` to be visible across translation units, but it's declared `static` in `amdgpu_migrate.c`. This means `amdgpu_svm_bo_is_svm` must also be in `amdgpu_migrate.c`, which it is (added in the same file's diff). However, this coupling is fragile — if someone moves the function or adds LTO optimizations that merge identical functions, the pointer comparison breaks. Consider using a flag in the BO instead.

**`amdgpu_bo_is_amdgpu_bo` now calls `amdgpu_svm_bo_is_svm` unconditionally:**
```c
if (bo->destroy == &amdgpu_bo_destroy ||
    bo->destroy == &amdgpu_bo_user_destroy ||
    amdgpu_svm_bo_is_svm(bo))
	return true;
```
This adds a function call overhead on the hot path for every BO check, even when SVM is disabled. The `!IS_ENABLED(CONFIG_DRM_AMDGPU_SVM)` stub returns `false`, so this is just a minor performance nit — the compiler should inline and eliminate it.

**Eviction in `amdgpu_bo_move` — no error message on failure:**
```c
r = amdgpu_svm_bo_evict(abo);
if (!r)
	ttm_bo_move_null(bo, new_mem);
return r;
```
If `amdgpu_svm_bo_evict` fails, `r` is returned directly with no logging. TTM will retry eviction, potentially in an infinite loop. Consider at least a `dev_warn`.

---
Generated by Claude Code Patch Reviewer