GPU VulnDB

Database/Kernel, userspace & hypervisor

Linux iommu/tegra241-cmdqv: multi-stream devices get incomplete vSID mapping, zero-stream devices read out of bounds

CVSS 8.8CVE-2026-90425Kernel, userspace & hypervisorcurated

Impact

tegra241_vintf_init_vsid() mapped a guest vSID to master->streams[0] while only warning if the device did not have exactly one stream. A device with several streams got only its first one mapped, so guest-issued vSID invalidations never reached the ATC and IOTLB entries of the others - stale IOMMU translations surviving a guest's invalidation, which breaks the isolation guarantee virtualized SMMU command queues exist to provide. A device with no streams made master->streams a ZERO_SIZE_PTR that was then read out of bounds. This is the Grace/Tegra241 CMDQ-V path, so it matters on NVIDIA Grace-based nodes that pass accelerators or NICs into guests; it is not reachable on x86 hosts or without the tegra241-cmdqv driver in use.

Who can reach it

Local, privileged: the host path that sets up a virtualized SMMU command-queue interface for a guest - device assignment to a VM on a Grace/Tegra241 host. Not reachable by an ordinary tenant without device-passthrough rights, and not reachable at all on hardware without this IOMMU.

What to do

Take the stable kernel containing the fix (mapping is now rejected with -EOPNOTSUPP unless master->num_streams is exactly one) and reboot affected Grace/Tegra241 nodes; drain each before rebooting. No runtime mitigation other than not assigning multi-stream devices into guests on these hosts.

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.