CVE-2026-89959

A logic error in the Linux kernel's s390 VFIO-AP driver causes the vfio_ap_config_remove function to use bitmap_andnot instead of bitmap_and when clearing a control-domain bit from the mediated device's matrix.adm bitmap, so control domains explicitly unplugged by the administrator are never actually removed from the KVM guest. It is triggered on IBM Z (s390) hosts when an operator unplugs or removes an assigned control domain from a vfio-ap mediated device passed through to a running KVM guest. The guest retains access to cryptographic control domains the host intended to revoke, so a low-privileged local attacker inside the guest can continue performing cryptographic operations against resources that should no longer be reachable (CVSS 8.8, scope changed). Only Linux-on-IBM-Z deployments that use the vfio-ap driver for AP (adjunct processor) crypto passthrough to KVM guests are affected; systems that do not use vfio-ap are unaffected. No public proof-of-concept or in-the-wild exploitation is known, and the issue is not listed in CISA KEV.

What to do: Update s390 Linux kernels to a build containing the upstream vfio-ap fix (bitmap_and in vfio_ap_config_remove); no fixed version number is given in the advisory, so track the stable kernel that carries this commit. As an interim mitigation, avoid removing control domains from active vfio-ap mdevs while guests are running — instead update the mdev's adapter/domain matrix and restart the guest so the AP masks are rebuilt. Audit current guest assignments of control domains against intended configuration to identify any domains that were 'unplugged' but remain available to guests.