Linux Kernel EUVD-2026-38821

Requires a host using the Intel VT-d (iommu/vt-d) driver with VFIO/iommufd PCI device passthrough configured in scalable mode with nested/second-stage translation, and a local actor with privileges to bind and release such a device (e.g. via QEMU). The specific trigger is releasing the device - the call path pcie_flr -> __iommu_set_group_pasid -> intel_nested_set_dev_pasid -> domain_remove_dev_pasid runs on device close/reset, e.g. when the QEMU process is terminated. Limiting factors: it is not remotely reachable (AV:L), needs pre-existing privilege/delegation to operate VFIO (PR:L), only manifests on Intel IOMMU hardware with passthrough actually in use, and only on unpatched 6.6-through-7.0 kernels; hosts without VFIO passthrough or without VT-d are not exposed.

Signals are internally consistent toward a moderate, local-only availability bug rather than a high-urgency RCE. The CVSS 3.1 vector (AV:L/AC:L/PR:L/UI:N/S:U/C:N/I:H/A:H, base 7.1) requires local access and existing privileges, which matches the call path: only a process able to drive VFIO/iommufd passthrough (typically root or a delegated, capability-granted user) can reach the vulnerable teardown. EPSS of 0.17% (7th percentile), absence from CISA KEV, and no identified POC all indicate low real-world exploitation likelihood. Note a conflict to verify: the intelligence tags label this 'Information Disclosure', yet the CVSS sets C:N and the kernel description describes a crash (oops/GPF) with no read-out primitive - the practical impact is denial of service, not confidentiality loss, so the 'Information Disclosure' tag appears mislabeled. The I:H in the supplied vector is also debatable, as the resolved behavior is a fault rather than demonstrated memory corruption. Net: a real correctness/stability bug worth patching on virtualization hosts, but not an emergency for general fleets.

A local user permitted to run QEMU guests with VFIO PCI passthrough on an Intel VT-d host binds a device using nested/scalable-mode translation, then kills the QEMU process. The resulting device close triggers an FLR reset whose PASID teardown dereferences the dummy blocked domain, faulting the kernel and crashing the host (denial of service). No public POC is identified; exploitation requires local access and existing privileges (AV:L/PR:L), so impact is host availability rather than remote compromise.