On Mon, Oct 13, 2025 at 06:26:11PM +0300, Leon Romanovsky wrote:
From: Leon Romanovsky leonro@nvidia.com
Add support for exporting PCI device MMIO regions through dma-buf, enabling safe sharing of non-struct page memory with controlled lifetime management. This allows RDMA and other subsystems to import dma-buf FDs and build them into memory regions for PCI P2P operations.
The implementation provides a revocable attachment mechanism using dma-buf move operations. MMIO regions are normally pinned as BARs don't change physical addresses, but access is revoked when the VFIO device is closed or a PCI reset is issued. This ensures kernel self-defense against potentially hostile userspace.
Let's enhance this:
Currently VFIO can take MMIO regions from the device's BAR and map them into a PFNMAP VMA with special PTEs. This mapping type ensures the memory cannot be used with things like pin_user_pages(), hmm, and so on. In practice only the user process CPU and KVM can safely make use of these VMA. When VFIO shuts down these VMAs are cleaned by unmap_mapping_range() to prevent any UAF of the MMIO beyond driver unbind.
However, VFIO type 1 has an insecure behavior where it uses follow_pfnmap_*() to fish a MMIO PFN out of a VMA and program it back into the IOMMU. This has a long history of enabling P2P DMA inside VMs, but has serious lifetime problems by allowing a UAF of the MMIO after the VFIO driver has been unbound.
Introduce DMABUF as a new safe way to export a FD based handle for the MMIO regions. This can be consumed by existing DMABUF importers like RDMA or DRM without opening an UAF. A following series will add an importer to iommufd to obsolete the type 1 code and allow safe UAF-free MMIO P2P in VM cases.
DMABUF has a built in synchronous invalidation mechanism called move_notify. VFIO keeps track of all drivers importing its MMIO and can invoke a synchronous invalidation callback to tell the importing drivers to DMA unmap and forget about the MMIO pfns. This process is being called revoke. This synchronous invalidation fully prevents any lifecycle problems. VFIO will do this before unbinding its driver ensuring there is no UAF of the MMIO beyond the driver lifecycle.
Further, VFIO has additional behavior to block access to the MMIO during things like Function Level Reset. This is because some poor platforms may experience a MCE type crash when touching MMIO of a PCI device that is undergoing a reset. Today this is done by using unmap_mapping_range() on the VMAs. Extend that into the DMABUF world and temporarily revoke the MMIO from the DMABUF importers during FLR as well. This will more robustly prevent an errant P2P from possibly upsetting the platform.
A DMABUF FD is a prefered handle for MMIO compared to using something like a pgmap because: - VFIO is supported, including its P2P feature, on archs that don't support pgmap - PCI devices have all sorts of BAR sizes, including ones smaller than a section so a pgmap cannot always be created - It is undesirable to waste alot of memory for struct pages, especially for a case like a GPU with ~100GB of BAR size - We want a synchronous revoke semantic to support FLR with light hardware requirements
Use the P2P subsystem to help generate the DMA mapping. This is a significant upgrade over the abuse of dma_map_resource() that has historically been used by DMABUF exporters. Experience with an OOT version of this patch shows that real systems do need this. This approach deals with all the P2P scenarios: - Non-zero PCI bus_offset - ACS flags routing traffic to the IOMMU - ACS flags that bypass the IOMMU - though vfio noiommu is required to hit this.
There will be further work to formalize the revoke semantic in DMABUF. For now this acts like a move_notify dynamic exporter where importer fault handling will get a failure when they attempt to map. This means that only fully restartable fault capable importers can import the VFIO DMABUFs. A future revoke semantic should open this up to more HW as the HW only needs to invalidate, not handle restartable faults.
Jason