Linux kernel BPF LSM: two hooks reachable in atomic context were allowed for sleepable programs

Impact

bpf_lsm_task_to_inode() runs under rcu_read_lock() and bpf_lsm_inet_conn_established() runs in softirq context, yet both were listed as attachable by sleepable BPF LSM programs. A sleepable program attached to either hook can block where blocking is forbidden, which on a GPU node means a kernel crash or memory corruption that takes the whole host down - every tenant pod and every in-flight training job on that node with it. Exposure is narrow: loading a BPF LSM program needs privilege (CAP_BPF/CAP_SYS_ADMIN plus BPF LSM enabled in the kernel's lsm= list), so a fleet that does not hand tenants BPF loading rights or run a third-party eBPF security agent is not practically reachable here. The high NVD score reflects the privileged-local worst case, not a tenant-reachable path.

Who can reach it

Local, and effectively privileged: a user or agent that can load and attach sleepable BPF LSM programs (CAP_BPF plus a kernel built with BPF LSM active). Not reachable over the network and not reachable from an unprivileged tenant container.

What to do

Take the stable kernel update that drops both hooks from the sleepable allowlist (five stable branches carry the fix; the commits are linked in the record). Applying it means a kernel package update and a node reboot, so drain the node first. Interim mitigation if a reboot window is far off: audit which eBPF/LSM agents are permitted to load sleepable programs and keep BPF loading privileges off tenant workloads.

References

Related entries

• Linux kernel (net/xfrm): Transport-mode reinjection stashes a network-namespace pointer in the socket buffer's controlCVE-2026-63919 · Linux kernel (net/xfrm)High
• Linux kernel (arch/x86/kvm/svm): KVM read Page State Change entries and indices out of a guest-writable buffer moreCVE-2026-63937 · Linux kernel (arch/x86/kvm/svm)High
• KVM arm64: guest hypervisor using ZCR_EL1 escapes its SVE vector-length cap on nested restoreCVE-2026-63941 · Linux kernel KVM arm64 (ZCR_EL2 vector-length capping for nested guests)High
• Linux kernel (drivers/vfio/pci): Vfio-pci exports a dma-buf over BAR memory without confirming those BAR resources wereCVE-2026-64042 · Linux kernel (drivers/vfio/pci)High
• Linux kernel crypto/ecc: missing carry in 128-bit accumulation corrupts ECC arithmetic at a boundaryCVE-2026-64313 · Linux kernel crypto/ecc (vli multiplication carry handling)High
• Linux kernel QAT: use-after-free tearing down SR-IOV while VF2PF response work is in flightCVE-2026-64438 · Linux kernel Intel QAT driver (VF2PF work teardown in adf_disable_sriov)High