Claude review: drm/amdgpu: add VRAM migration infrastructure for drm_pagemap

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

Patch Review

**Overall:** Good foundational patch. Sets up ZONE_DEVICE registration and drm_pagemap initialization.

**Issue: SPDX vs. license boilerplate mismatch.** The file uses `/* SPDX-License-Identifier: GPL-2.0 OR MIT */` but then includes the full MIT permissive license text in the header block. The SPDX tag should be sufficient; the verbose boilerplate is unnecessary and adds noise. This applies to all new files in this series.

**Issue: `pgmap->flags = 0` is redundant** after `devm_kzalloc`:
```c
+	pgmap->nr_range = 1;
+	pgmap->flags = 0;
```
The struct was zero-initialized already.

**Issue: Error path leak in `drm_pagemap_init` failure.** If `drm_pagemap_init()` fails, the `devm_memremap_pages()` region is not cleaned up:
```c
+	if (drm_pagemap_init(&svm_dm->dpagemap, pgmap, adev_to_drm(adev),
+							&amdgpu_svm_drm_pagemap_ops)) {
+		dev_err(adev->dev, "SVM: failed to init drm_pagemap\n");
+		return -EINVAL;
+	}
```
After a successful `devm_memremap_pages()`, failing to init the pagemap should unwind the memremap. Although devres will eventually clean up, returning `-EINVAL` here leaves `adev->apagemap` as NULL with a phantom ZONE_DEVICE region registered.

**Issue: Alignment of `drm_pagemap_init` arguments:**
```c
+	if (drm_pagemap_init(&svm_dm->dpagemap, pgmap, adev_to_drm(adev),
+							&amdgpu_svm_drm_pagemap_ops)) {
```
The continuation is over-indented — should align after the opening paren per kernel style.

**Nit: `struct dev_pagemap` flex-array comment.** The comment says `pgmap` "must be last — flex-array" but `struct dev_pagemap` in current kernels uses `struct range ranges[]` which is indeed a flex-array. This constraint should be documented in the `amdgpu_pagemap` definition with `static_assert` or a build-time check, not just a comment.

**Question: `AMDGPU_SVM_PGMAP_OWNER` and `adev->hive`.** The macro:
```c
+#define AMDGPU_SVM_PGMAP_OWNER(adev) \
+	((adev)->hive ? (void *)(adev)->hive : (void *)(adev))
```
The `hive` field on `struct amdgpu_device` is `struct amdgpu_hive_info *hive`. This is valid, but note that `hive` can be set/cleared during runtime (XGMI topology changes). If the owner changes between registration and migration, `migrate_vma_setup()` will misidentify pages. Is this a concern for hot-plug scenarios?

**Nit:** `dpagemap_to_adev()` and `amdgpu_svm_page_to_apagemap()` are defined here but not used until patch 2. Unused code in patch 1 could trigger compiler warnings depending on config.

---
Generated by Claude Code Patch Reviewer