CVE-2025-39961

Quoted from the kernel CNA's own scoring file, word for word. A highlighted metric carries the most severe value CVSS v3.1 defines for it, which is what pushes the score up. source

AV:LAttack vector, local

The race is reached through local DMA-mapping activity — either VFIO/iommufd map/unmap ioctls or ordinary syscall-driven I/O that exercises dma_map/dma_unmap on a device behind the AMD IOMMU. No network protocol handler processes remote data on this path.

AC:LAttack complexity, low

The attacker controls both sides of the race: one thread issues a high-IOVA map that forces increase_address_space() while other threads issue concurrent unmaps on the same domain, and the attempt can be repeated indefinitely because each newly created protection_domain restarts at PAGE_MODE_3_LEVEL. No condition depends on state outside the attacker's influence, and CONFIG_AMD_IOMMU with v1 host page tables is the default on all AMD x86_64 systems.

PR:LPrivileges required, low

An unprivileged local user can drive the concurrent map/unmap traffic that exhausts the 32-bit IOVA space and triggers the page-table level increase; in device-assignment deployments the VFIO/iommufd ioctl path is exercised by an unprivileged QEMU process. No root or capability in the init namespace is required.

UI:NUser interaction, none

The race is triggered entirely by the attacker's own concurrent DMA map/unmap operations. No victim action is needed.

S:CScope, changed

A silently-failed iommu_v1_unmap_pages() leaves a live IOMMU translation after __iommu_dma_unmap() frees the IOVA and the driver recycles the buffer, so a DMA-capable device — including a guest-controlled assigned device — retains read/write access to host physical memory it must no longer reach. This is an IOMMU/DMA containment bypass that crosses out of the kernel's own security authority.

C:HConfidentiality impact, high

The stale mapping lets the device read physical pages that have been returned to the allocator and refilled with arbitrary kernel or other-tenant data, giving an unbounded read primitive over reused memory; the torn read in alloc_pte() can additionally alias a mapping onto another buffer in the domain.

I:HIntegrity impact, high

The same stale or mis-installed translation is writable, so the device can DMA arbitrary attacker-chosen data into freed-and-reused kernel memory, yielding a write primitive suitable for control-flow hijacking. The corruption is invisible to the kernel because the unmap was reported as complete.

A:HAvailability impact, high

The failed unmap immediately produces WARN_ON splats in __iommu_dma_unmap()/iommu_unmap_nofail() (a panic with panic_on_warn), and the resulting DMA into recycled memory or the mis-built page-table subtree causes IO page faults, page-table page leaks, and kernel crashes.