Claude review: drm/amdgpu: SVM VRAM migration via drm_pagemap (XNACK-on)

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

Overall Series Review

Subject: drm/amdgpu: SVM VRAM migration via drm_pagemap (XNACK-on)
Author: Junhua Shen <Junhua.Shen@amd.com>
Patches: 9
Reviewed: 2026-05-16T12:15:37.387075

---

This v4 series adds VRAM migration support to amdgpu's new drm_gpusvm-based SVM implementation, enabling GPU fault-driven (XNACK-on) migration of pages between system RAM and VRAM using SDMA. The series is well-structured, with each patch building logically on the previous one, and the architecture — wrapping `dev_pagemap` + `drm_pagemap` in `amdgpu_pagemap`, introducing `amdgpu_bo_svm` as a lightweight BO subtype — is sound and follows the Xe driver precedent.

**Strengths:**
- Good documentation: the address-space diagram in patch 1 and per-function kerneldoc are helpful
- Clean layering: infrastructure → SDMA callbacks → TTM eviction → range helpers → hookup
- Proper use of existing APIs (devm_memremap_pages, drm_pagemap framework, GART windowed SDMA)

**Concerns:**
- The `drm_pagemap_init()` error path in patch 1 leaks the `devm_memremap_pages` registration
- The SDMA copy functions have significant code duplication between `copy_to_devmem` and `copy_to_ram`
- `pre_migrate_fence` parameter is unused in both copy callbacks
- The `#if IS_ENABLED(CONFIG_DRM_AMDGPU_SVM)` guards in patch 5 are unnecessary since `amdgpu_svm_migration_init()` already has a static inline stub
- Migration failure in the fault path is silently ignored, which may lead to infinite fault loops

---

---
Generated by Claude Code Patch Reviewer