GPU VulnDB

Database/Kernel, userspace & hypervisor

Linux kernel perf/core: use-after-free on a freed group leader after a sibling is detached during CPU hotplug

CVE-2026-74637Kernel, userspace & hypervisorcurated

Impact

When a sibling perf event is detached with DETACH_GROUP - the CPU hot-unplug path - it is removed from the leader's sibling_list but keeps a group_leader pointer at the old leader. If the leader is then closed and freed while the sibling stays alive through its file descriptor, a PERF_IOC_FLAG_GROUP ioctl on the sibling follows the stale pointer into freed memory. The reproducer in the record is the perf fuzzer plus CPU hotplug plus load, and the trace shows a kernel paging fault in perf_ioctl. On a GPU node this is a local kernel use-after-free reachable by anyone permitted to open grouped perf events, which on fleets that relax perf_event_paranoid for profiling includes tenants; it becomes an availability and potential privilege-escalation risk on a shared node.

Who can reach it

Local user able to create perf event groups (governed by perf_event_paranoid and CAP_PERFMON) plus concurrent CPU hotplug activity. Not remote. On nodes where perf_event_paranoid is at the default restrictive setting and unprivileged perf is disabled, unprivileged tenants cannot reach it.

What to do

Apply the stable fix, which promotes a detached sibling to a singleton and corrects the __event_disable() cgroup accounting; three stable commits are listed. Rollout is a patched kernel and a per-node reboot, so drain GPU workloads first. As an interim measure, tightening kernel.perf_event_paranoid to disallow unprivileged perf removes tenant reachability.

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.