GPU VulnDB

Database/Kernel, userspace & hypervisor

Linux kernel ice: unchecked VF-supplied ptype writes past a bitmap and faults the host kernel

CVSS 8.8CVE-2026-68128Kernel, userspace & hypervisorcurated

Impact

ice_parser_profile_init() called set_bit(rslt->ptype, prof->ptypes) on a 1024-bit bitmap without checking the ptype a VF supplied over VIRTCHNL, so a ptype >= 1024 writes past the end of the bitmap. The reporter demonstrated a kernel oops on an E810-C host. On a GPU node this matters where Intel E810 NICs are used for tenant networking with SR-IOV: a guest or tenant that owns a VF reaches host kernel memory it should not, giving at least a host crash that takes every workload on the node with it, and an out-of-bounds write that may be steerable toward privilege escalation. Any node that crashes this way is a node whose in-flight training or inference work is lost, not something you drain cheaply.

Who can reach it

A VF owner - a guest VM or tenant handed an E810 virtual function - sending a crafted VIRTCHNL request such as VIRTCHNL_OP_ADD_RSS_CFG to the host PF driver. No host credentials needed beyond control of the VF; hosts not using ice SR-IOV with untrusted VF consumers are not exposed.

What to do

Update to a stable kernel containing the range check (four stable commits linked) and reboot each host using the ice driver with SR-IOV; a driver-level fix in the PF path cannot be hot-patched. Until then, the exposure is bounded by who holds VFs - not assigning E810 VFs to untrusted guests removes the attack path without a reboot. No vendor advisory with fixed version strings is in the record.

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.