Claude review: net: devmem: allow rx-buf-size > PAGE_SIZE per binding

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

Overall Series Review

Subject: net: devmem: allow rx-buf-size > PAGE_SIZE per binding
Author: Bobby Eshleman <bobbyeshleman@gmail.com>
Patches: 5
Reviewed: 2026-06-05T06:49:28.207081

---

This is a well-motivated 4-patch series by Bobby Eshleman (Meta) that allows devmem TCP dmabuf bindings to use niov sizes larger than PAGE_SIZE. The ~19% reduction in app sys% CPU at 64K niovs vs 4K niovs is compelling. The architecture is clean: a new `NETDEV_A_DMABUF_RX_BUF_SIZE` netlink attribute lets userspace request a larger niov size at bind time, stored as `niov_shift` in the binding, with all PAGE_SIZE/PAGE_SHIFT arithmetic generalized. The udmabuf patch coalesces sg entries to match folio granularity, enabling the selftest to use MFD_HUGETLB backing for large niovs.

**Key concerns:**
1. **No kernel-side upper bound on `rx_buf_size`** -- the selftest enforces a 2MB cap, but the kernel accepts anything up to 2^31 (the largest power-of-two u32). An absurdly large niov_shift will fail later at the sg alignment check, but a sensible upper bound (e.g., aligned with the 2MB hugepage limit for current hardware) would provide a better user error.
2. **`netdev-genl-gen.c` is typically auto-generated** -- the hand edits should be verified against regeneration from the updated YAML spec.
3. The series is based on a tree that includes the `vdev` parameter to `net_devmem_bind_dmabuf()` and the folio/offsets refactoring of udmabuf, neither of which is present on drm-next. These are dependencies, not bugs, but worth noting for merge ordering.

Overall the series is in good shape. The issues below are mostly nits and one substantive validation gap.

---

---
Generated by Claude Code Patch Reviewer