GPU VulnDB

Database/Kernel, userspace & hypervisor

Linux kernel mm/hugetlb: page fault touches freed folio, unprivileged user can corrupt host memory

UnscoredCVE-2026-99506Kernel, userspace & hypervisorcurated

Impact

For a fault in a private hugetlb mapping, hugetlb_no_page() keeps working on the new folio after dropping the page-table lock. A concurrent MADV_DONTNEED or hole punch can free the folio in that gap. The kernel then sets flags on, and unlocks, a page that is free or that another file now owns. That gives bad page state, kernel BUG crashes and possible cross-file page corruption. The record says an unprivileged process is enough, using MAP_PRIVATE | MAP_NORESERVE hugetlb mappings. GPU and HPC nodes often preallocate hugepage pools for training and RDMA, so a tenant on such a node can crash the host or corrupt another user's pages.

Who can reach it

Local, unprivileged. Any process that can mmap hugetlb memory (MAP_HUGETLB or hugetlbfs) on a host with a hugepage pool or surplus hugepages configured. Containers that can reach hugepages count.

What to do

Install a kernel with the stable fix, then drain and reboot each node. If no hugepages are configured or exposed to workloads, the path cannot be reached. The record gives no fixed version numbers.

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.