On Tue, 28 Apr 2020 21:29:03 -0300 Jason Gunthorpe jgg@ziepe.ca wrote:
On Tue, Apr 28, 2020 at 02:12:23PM -0600, Alex Williamson wrote:
Maybe I was just getting lucky before this commit. For a VM_PFNMAP, vaddr_get_pfn() only needs pin_user_pages_remote() to return error and the vma information that we setup in vfio_pci_mmap().
I've written on this before, vfio should not be passing pages to the iommu that it cannot pin eg it should not touch VM_PFNMAP vma's in the first place.
It is a use-after-free security issue the way it is..
Where is the user after free? Here I'm trying to map device mmio space through the iommu, which we need to enable p2p when the user owns multiple devices.
Yes, I gathered what the intent was..
The device is owned by the user, bound to vfio-pci, and can't be unbound while the user has it open. The iommu mappings are torn down on release. I guess I don't understand the problem.
For PFNMAP VMAs the lifecycle rule is basically that the PFN inside the VMA can only be used inside the mmap_sem that read it. Ie you cannot take a PFN outside the mmap_sem and continue to use it.
This is because the owner of the VMA owns the lifetime of that PFN, and under the write side of the mmap_sem it can zap the PFN, or close the VMA. Afterwards the VMA owner knows that there are no active reference to the PFN in the system and can reclaim the PFN
ie the PFNMAP has no per-page pin counter. All lifetime revolves around the mmap_sem and the vma.
What vfio does is take the PFN out of the mmap_sem and program it into the iommu.
So when the VMA owner decides the PFN has no references, it actually doesn't: vfio continues to access it beyond its permitted lifetime.
HW like mlx5 and GPUs have BAR pages which have security properties. Once the PFN is returned to the driver the security context of the PFN can be reset and re-assigned to another process. Using VFIO a hostile user space can retain access to the BAR page and upon its reassignment access a security context they were not permitted to access.
This is why GUP does not return PFNMAP pages and vfio should not carry a reference outside the mmap_sem. It breaks all the lifetime rules.
Thanks for the explanation. I'm inferring that there is no solution to this, but why can't we use mmu notifiers to invalidate the iommu on zap or close? I know that at least QEMU won't consider these sorts of mapping fatal, so we could possibly change the default and make support for such mappings opt-in, but I don't know if I'd break DPDK, or potentially users within QEMU that make use of p2p between devices. Thanks,
Alex