Claude review: memcg: Track exported dma-buffers

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

Patch Review

This patch (from T.J. Mercier's earlier series) adds `MEMCG_DMABUF` as a memcg stat counter, charges at `dma_buf_export()` time, and uncharges at `dma_buf_release()` time.

**Issue 1 - Bisectability:** This patch charges in `dma_buf_export()` for ALL dma-buf exports, but patch 2 immediately removes that code and moves charging to `dma_heap_buffer_alloc()` only. Between patches 1 and 2, every DRM driver, V4L2 exporter, etc. that calls `dma_buf_export()` incurs a memcg charge. This creates a window where behavior differs from both the pre-series state and the final intended state. The infrastructure (stat counter, charge/uncharge functions) should be added in patch 1 without hooking into `dma_buf_export()`, and patch 2 should add the charging at the heap level.

```c
+	dmabuf->memcg = get_mem_cgroup_from_mm(current->mm);
+	if (!mem_cgroup_charge_dmabuf(dmabuf->memcg, PAGE_ALIGN(dmabuf->size) / PAGE_SIZE,
+				      GFP_KERNEL)) {
```
This charges unconditionally in `dma_buf_export()`, then patch 2 removes it entirely.

**Issue 2 - Uncharge before ops->release():** The uncharge is placed before `dmabuf->ops->release()`:
```c
+	mem_cgroup_uncharge_dmabuf(dmabuf->memcg, PAGE_ALIGN(dmabuf->size) / PAGE_SIZE);
+	mem_cgroup_put(dmabuf->memcg);
+
 	dmabuf->ops->release(dmabuf);
```
This means the memcg charge is removed before the backing pages are actually freed. If `ops->release()` takes significant time or blocks, there's a window where the pages are still allocated but not charged. This is a minor accuracy concern but worth noting.

**Issue 3 - doc tag:** The documentation adds `dmabuf (npn)` — it would be helpful to confirm that `(npn)` is the correct format tag. Looking at the existing doc entries, most stats don't use this tag pattern, so this might be incorrect formatting or an undefined abbreviation.

---
Generated by Claude Code Patch Reviewer