Claude review: drm/amdgpu: add SVM range migration helpers for drm_pagemap

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

Patch Review

**`amdgpu_svm_get_dpagemap` doesn't NULL-check `apagemap`:**
```c
static struct drm_pagemap *
amdgpu_svm_get_dpagemap(struct amdgpu_svm *svm)
{
	struct amdgpu_pagemap *apagemap = svm->adev->apagemap;

	if (!apagemap->initialized)
		return NULL;

	return &apagemap->dpagemap;
}
```
If `apagemap` is NULL (SVM migration not initialized), this will dereference NULL. Needs:
```c
if (!apagemap || !apagemap->initialized)
    return NULL;
```
This is called from `amdgpu_pagemap_capable()` which is used in the fault path, so it's reachable when migration_init wasn't called (e.g., old GPU, APU, or init failure).

**`range_needs_migrate_to_vram` accesses `range->pages.flags.migrate_devmem` without READ_ONCE:**
```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;
```
But `range_in_vram()` carefully uses `READ_ONCE` on the flags:
```c
struct drm_gpusvm_pages_flags flags = {
	.__flags = READ_ONCE(range->pages.flags.__flags),
};
```
Inconsistent synchronization — if `flags.__flags` can change concurrently (as the comment "Pairs with WRITE_ONCE" suggests), then `range_needs_migrate_to_vram` should also use `READ_ONCE`.

**`amdgpu_svm_range_migrate_range` ignores migration errors:**
```c
if (range_needs_migrate_to_vram(range, migrate_mode))
	return amdgpu_svm_range_migrate_to_vram(svm, range);
```
The return value propagates here, but callers in patch 6 (fault path) ignore it — see patch 6 review.

**SPDX comment style:** Same issue as patch 1 — `.c` file uses `/* SPDX */` instead of `// SPDX`.

---
Generated by Claude Code Patch Reviewer