GPU VulnDB

Database/Kernel, userspace & hypervisor

Linux kernel nvme-pci: use-after-free when Host Memory Buffer setup fails during device probe

CVSS 7.8CVE-2026-64071Kernel, userspace & hypervisorcurated

Impact

nvme_free_host_mem() frees dev->hmb_sgt but never clears the pointer, so the nvme_probe() error path can free it a second time. The reported crash is a NULL dereference inside iommu_dma_free_noncontiguous() that takes the machine down during probe. Reaching it requires an NVMe device that advertises a Host Memory Buffer and then fails nvme_set_host_mem() with an I/O error - HMB is a DRAM-less SSD feature, and the reporter hit it on a Thunderbolt-attached drive behind a dock, not on enterprise U.2 or E1.S media. On a GPU node the realistic exposure is boot or hot-plug of such a device rather than anything a tenant can drive, but a probe-time crash is an unplanned reboot on a node whose jobs are expensive to drain.

Who can reach it

Local and hardware-adjacent: needs an NVMe device advertising HMB that errors during probe, so boot time or a hot-plugged or link-unstable drive. The record describes no authenticated remote path and no tenant-reachable trigger.

What to do

Take the stable fix (the three linked commits), which sets dev->hmb_sgt to NULL after freeing so the second call takes the safe multi-descriptor path. Kernel change: drain and reboot each node onto a patched kernel. No fixed release version is named in the record - follow your distribution's stable backport. Fleets running only enterprise NVMe that does not advertise HMB are in practice unaffected.

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.