Linux kernel BPF: tail calls between programs on different hooks bypass ctx-access and return-value checks

Impact

The verifier checks each BPF program against the prototype and return-value rules of the kernel function or LSM hook it is attached to. It did not stop a program on one hook from tail-calling a program attached to a different hook. The callee could then read context arguments the caller's hook never supplied, or return values the caller's hook forbids, such as a positive return from a BPF LSM hook. The result is out-of-bounds context reads and broken LSM policy decisions. This matters on nodes where BPF loading is delegated to observability, security or CNI agents (Cilium, Tetragon-style tools) rather than kept with root.

Who can reach it

Local. Requires the ability to load and attach tracing or BPF LSM programs, which normally means root or CAP_BPF with CAP_PERFMON/CAP_SYS_ADMIN. It is not reachable by an unprivileged tenant unless such capabilities are granted to their container.

What to do

Update to a stable kernel that includes the fix (commits in the record; Debian LTS has shipped it) and reboot each node. Drain GPU workloads first. Until then, do not grant CAP_BPF or CAP_SYS_ADMIN to tenant containers, and keep unprivileged BPF disabled (kernel.unprivileged_bpf_disabled).

References

Related entries

• Linux kernel (drivers/gpu/drm/xe): The observation/OA path reuses one batch buffer and appends a batch-end command onCVE-2024-50090 · Linux kernel (drivers/gpu/drm/xe)High
• Linux kernel (drivers/iommu/intel): VT-d walked the PCI DMA-alias list for devices that are not PCI at all whileCVE-2024-50101 · Linux kernel (drivers/iommu/intel)High
• Linux kernel (drivers/gpu/drm/xe): Xe freed a job from inside timeout-detection-and-recovery while the submissionCVE-2024-50149 · Linux kernel (drivers/gpu/drm/xe)High
• Linux kernel cxl/port: use-after-free when CXL region decoders are torn down out of orderCVE-2024-50226 · Linux kernel CXL port driver (out-of-order switch decoder shutdown)High
• Linux kernel sunrpc: use-after-free when a network namespace is destroyed with an NFS TCP socket still armedCVE-2024-53168 · Linux kernel sunrpc (NFS client TCP socket lifetime vs. network namespace teardown)High
• Linux kernel (drivers/vfio/pci): Out-of-bounds read past the ecap_perms table when a tenant touches emulated PCIeCVE-2024-53214 · Linux kernel (drivers/vfio/pci)High