GPU VulnDB

Database/Kernel, userspace & hypervisor

Linux kernel iommufd: dma_buf_unpin() without dma_resv lock warns on every DMABUF IOAS teardown

CVE-2026-80633Kernel, userspace & hypervisorcurated

Impact

iopt_release_pages() calls dma_buf_unpin() without holding the exporter's dma_resv lock, which that function asserts. Every iommufd_ioas_destroy() or iommufd_ioas_unmap() that drops the last reference on a DMABUF-backed iopt_pages therefore trips a kernel WARN - reproduced by the in-tree iommufd selftests, so it fires on a normal close(2) of the iommufd file descriptor. This is the path GPU and DPU passthrough uses when device memory is handed to a VM as a dma-buf, so on a virtualized GPU host it triggers on ordinary VM shutdown, not on an unusual sequence. No memory corruption is described in the record; the exposure is a missing lock and the resulting warning, which on a fleet running panic_on_warn means any tenant able to tear down a passthrough VM can panic the host and take every co-resident VM with it.

Who can reach it

Local process holding an iommufd file descriptor with a DMABUF-backed mapping - in practice the VMM on a host doing GPU or DPU passthrough, or any local user permitted to open /dev/iommu. No remote access and no root required beyond device permissions.

What to do

Apply the stable kernel fix and reboot the hypervisor hosts, which means live-migrating or draining guests first. As an interim measure, confirm whether panic_on_warn is set on your GPU hosts and clear it there if it is, since that setting is what turns this from a log splat into a node outage. The record carries stable commits only and no distro-fixed version - check your vendor's kernel errata.

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.