Claude review: Add dmabuf read/write via io_uring

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

Overall Series Review

Subject: Add dmabuf read/write via io_uring
Author: Pavel Begunkov <asml.silence@gmail.com>
Patches: 17
Reviewed: 2026-05-05T11:25:59.755642

---

This is a well-structured v3 series from Pavel Begunkov that enables dmabuf read/write via io_uring, targeting GPU <-> NVMe transfers with impressive IOMMU cost improvements (570 KIOPS to 5.01 MIOPS in strict mode). The architecture is clean: a token/map abstraction in `lib/` handles dmabuf lifecycle and invalidation via percpu refcounts and dma fences, a new iov_iter type carries the opaque map through the block layer, and NVMe PCI is the first driver to implement the backend. The io_uring integration extends the registered buffer infrastructure with a new `IO_REGBUF_TYPE_DMABUF` type.

There is one clear **bug** (wrong `sizeof` in the NVMe driver), a **resource leak** on error path in the same function, a **pre-existing parenthesization bug** that the series perpetuates in `iov_iter_restore`, and several design/style points across the patches.

---
Generated by Claude Code Patch Reviewer