GPU VulnDB

Database/Kernel, userspace & hypervisor

Linux kernel BPF: stale rbtree node linkage after bpf_rb_root_free allows erase on a freed root

CVSS 7.8CVE-2026-74344Kernel, userspace & hypervisorcurated

Impact

bpf_rb_root_free() detached the root and dropped the contained objects without clearing each node's rb linkage or owner pointer. A refcounted object that survives the drop keeps an owner pointer into the freed root and stale rb pointers; if a later bpf_rb_root allocation lands on the same address, bpf_rbtree_remove() passes the owner check and calls rb_erase_cached() on a node whose links belong to the dead tree. That is attacker-influenced kernel memory corruption reachable from a loaded BPF program using rbtree collection kfuncs - the usual path from there is local privilege escalation on the host, which on a shared GPU node means escape from whatever the BPF loader was confined to.

Who can reach it

Local, requires the ability to load BPF programs that use rbtree collections (CAP_BPF/CAP_SYS_ADMIN). Not reachable from an unprivileged tenant workload unless BPF loading has been delegated to it.

What to do

Install a stable kernel with the fix that marks nodes busy while the root is detached and clears node linkage and owner before dropping the object (backported across stable branches; see the linked commits for the release on yours), then drain and reboot the node. No vendor-published mitigation beyond keeping BPF loading privileged.

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.