Linux kernel arm64 BPF JIT: misaligned PLT target can tear under concurrent text poke

CVSS 7.8CVE-2026-23383Kernel, userspace & hypervisorcurated

Impact

The JIT buffer was allocated with 4-byte alignment, so the u64 target field in struct bpf_plt can land off an 8-byte boundary. On arm64 a 64-bit load is only single-copy atomic when 64-bit aligned, so a WRITE_ONCE from bpf_arch_text_poke concurrent with JIT'd code executing the ldr can be read torn, sending the JIT to a corrupted address. On arm64 GPU hosts - Grace Hopper, GB200, Ampere-based nodes - BPF is in the hot path of the CNI, of security agents and of fleet observability, and trampoline patching happens whenever those attach or detach. The realistic outcome is a kernel panic on a node holding running training jobs, not a clean fault.

Who can reach it

Local, arm64 only. Requires the ability to load and attach BPF programs, i.e. CAP_BPF/CAP_SYS_ADMIN on the host, or an unprivileged-BPF configuration. A tenant pod without those capabilities cannot reach it; the common trigger is ordinary privileged agent churn rather than an attacker.

What to do

Take a stable kernel carrying the listed commits. The fix is in the JIT allocator, so it cannot be applied to a running kernel by reloading a module: drain the node and reboot it, or apply a vendor livepatch if one covers this commit. No mitigation short of not running BPF text-poking agents on arm64 nodes.

References

https://nvd.nist.gov/vuln/detail/CVE-2026-23383 https://git.kernel.org/stable/c/519b1ad91de5bf7a496f2b858e9212db6328e1de https://git.kernel.org/stable/c/f5a41b61b5faae1965a5a2cb55322fe8f17e3b2d

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.