Claude review: drm/amdgpu: integrate VRAM migration into SVM range map and fault paths

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

Patch Review

**Migration failure silently ignored in fault path:**
```c
if (need_vram_migration)
	amdgpu_svm_range_migrate_range(svm, &range->base,
				      AMDGPU_SVM_MIGRATE_TO_VRAM);
```
The return value is discarded. If migration fails, execution continues and `amdgpu_svm_range_get_pages()` will map system RAM pages instead. This is reasonable as a fallback, but:
1. If `devmem_only` is true in `map_ctx`, the subsequent `get_pages` may fail because it expects VRAM pages
2. No logging means silent performance degradation is invisible
3. If migration fails due to VRAM pressure, the GPU will keep faulting on the same pages — potential infinite retry loop

At minimum, log the failure. Consider also clearing `map_ctx.devmem_only` after migration failure.

**Same issue in the prefetch path:**
```c
if (need_vram_migration) {
	AMDGPU_SVM_RANGE_DEBUG(range, "PREFETCH - MIGRATION PAGES");
	amdgpu_svm_range_migrate_range(svm, &range->base,
				      AMDGPU_SVM_MIGRATE_TO_VRAM);
}
```
Return value ignored again.

**`devmem_possible = false` TODO removal is the right cleanup**, but the declaration-after-statement pattern (`bool devmem_possible = ...; bool need_vram_migration = ...; struct drm_gpusvm_ctx map_ctx = { ... };`) mixing declarations and assigned variables should work fine with modern C standards but worth noting that some kernel subsystems still prefer C89-style declarations. The amdgpu driver generally allows mixed declarations, so this is fine.

**VRAM PTE flags handling looks correct:**
```c
if (entry->proto == AMDGPU_INTERCONNECT_VRAM)
	seg_pte_flags &= ~(AMDGPU_PTE_SYSTEM | AMDGPU_PTE_SNOOPED);
```
Clearing SYSTEM and SNOOPED for local VRAM pages is the right thing — these are direct VRAM accesses through MMHUB, not system memory through IOMMU.

---
Generated by Claude Code Patch Reviewer