GPU VulnDB

Database/Kernel, userspace & hypervisor

Linux KVM guest_memfd: unbind race with memory-failure handling leaves a guest hung

UnscoredCVE-2026-99336Kernel, userspace & hypervisorcurated

Impact

When a guest_memfd file is being torn down, kvm_gmem_unbind() removes the memslot binding without taking the invalidate lock. If that races with memory-failure handling (for example a poisoned page), the MMU invalidation counter is left raised and never comes back down. The guest's page faults then retry forever and the VM hangs. guest_memfd backs private memory for confidential VMs (SEV-SNP/TDX), so hosts running confidential GPU guests are the ones exposed. The record describes a hang of the affected VM only. It does not mention any memory disclosure or host compromise.

Who can reach it

Host-side race between a memory-failure event and a memslot delete on a dying guest_memfd file. The record does not describe a way for a guest or tenant to trigger it on demand. Only hosts whose VMM uses guest_memfd are affected.

What to do

Install a kernel that carries the stable fix (commits linked below) and reboot the host. Migrate or drain the VMs before the reboot. Livepatch may also be an option if your vendor ships one. The record does not list fixed version numbers, so check your distribution's kernel advisory.

References

Related entries

All Kernel, userspace & hypervisor entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.