Claude review: vfio/pci: Add mmap() for DMABUFs

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

Overall Series Review

Subject: vfio/pci: Add mmap() for DMABUFs
Author: Matt Evans <mattev@meta.com>
Patches: 15
Reviewed: 2026-05-28T12:28:51.795261

---

This is a well-structured v2 series from Matt Evans (Meta) that converts VFIO PCI BAR mmap() to be backed by DMABUFs, then adds mmap() support for explicitly-exported DMABUFs, plus ioctl-based permanent revocation and memory attribute control. The goal is to allow userspace drivers to safely delegate BAR sub-ranges to subordinate processes via DMABUF fds.

**Architecture**: The approach of unifying the BAR mmap path and explicit DMABUF mmap path under common fault-handling code is sound and follows prior review feedback. The lifecycle management -- with the BAR mmap DMABUF taking ownership of the device fd -- is a clever way to preserve existing behavior.

**Concerns**:
1. The `#ifdef CONFIG_PCI_P2PDMA` / `#ifdef CONFIG_VFIO_PCI_DMABUF` guards sprinkled through `vfio_pci_dmabuf.c` (patch 4) create a file that compiles in multiple configurations with interleaved `#ifdef` blocks, which is fragile.
2. Patch 7's fault handler synchronization protocol (resv lock -> try_get_registration -> unlock resv -> memory_lock) is complex with 5 numbered steps. While thoroughly commented, this is a subtle ordering that future maintainers need to preserve.
3. The `bar_needs_zap` flag (patch 6) for nvgrace-gpu is a reasonable interim hack but deserves an actual timeline for removal.
4. Patch 9's use of `-ENOTSUPP` (kernel-internal) instead of `-EOPNOTSUPP` (POSIX) in a UAPI-facing path is a bug.

Overall the series is in good shape for a v2. The code is well-commented, the cover letter is thorough, and the incremental approach is easy to follow.

---

---
Generated by Claude Code Patch Reviewer