Claude review: drm/shmem-helper: Fix Map huge page mapping in fault handler

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

Patch Review

**Race condition analysis is correct.** The commit message clearly explains the problem and reproducer. The Fixes tag is appropriate.

**Issue 1: Missing fallback from huge_fault to PTE mapping.**

When `.huge_fault` is called and the folio is PMD-mappable but alignment fails, the function returns `VM_FAULT_FALLBACK`. The MM layer will then call `.fault` which does PTE insertion. This is correct. However, if the folio is PMD-mappable and aligned but `vmf_insert_pfn_pmd()` itself fails (returns something other than `VM_FAULT_NOPAGE`), the error is propagated correctly too. This path looks fine.

**Issue 2: `pmd_none()` check was dropped.**

The original code had:
```c
if (aligned &&
    pmd_none(*vmf->pmd) &&
    folio_test_pmd_mappable(page_folio(page))) {
```

The new code removes the `pmd_none(*vmf->pmd)` check:
```c
if (aligned &&
    folio_test_pmd_mappable(page_folio(pages[page_offset]))) {
```

In the `.huge_fault` path, the PMD should be guaranteed to be in a suitable state by the MM layer, so dropping this check is likely fine — `vmf_insert_pfn_pmd()` handles it internally. But it's worth confirming this is intentional and that `vmf_insert_pfn_pmd` handles the non-none PMD case gracefully (returning an appropriate error rather than corrupting the page table).

**Issue 3: Needs rebase on drm-next.**

The current drm-next tree at `drivers/gpu/drm/drm_gem_shmem_helper.c:557-612` has substantially different code:
- The helper is now called `drm_gem_shmem_try_insert_pfn_pmd()` (not `drm_gem_shmem_try_map_pmd`)
- There's a `folio_mark_accessed()` call on success (line 605-606)
- A `pfn_mkwrite` handler has been added (line 648-664)
- Variable declarations differ (`page`, `folio` locals; `drm_dev` used for warnings)

The patch needs to be rebased to target the current tree.

**Issue 4: `folio_mark_accessed()` is missing from the patched code.**

The current drm-next code calls `folio_mark_accessed(folio)` on successful mapping. The patch (based on an older tree) does not include this. If this patch is meant to be backported to the older tree that didn't have it, this is fine; but when rebased on drm-next, it should be preserved.

**Minor: Subject line capitalization.** "Fix Map huge page" — the "Map" should be lowercase: "Fix map huge page mapping in fault handler" or perhaps rephrased to "Fix race between PMD and PTE fault paths".

**Overall:** The fix is correct in approach. Using `.huge_fault` for PMD-order mappings and `.fault` for PTE-order is the right way to let the MM layer serialize concurrent faults at different granularities. The patch needs a rebase onto drm-next before it can be applied.

---
Generated by Claude Code Patch Reviewer