Linux kernel io_uring: per-task restrictions are freed across exec, so post-exec rings are unrestricted

Impact

Per-task io_uring restrictions are meant to apply to every ring a task creates. The exec cancellation path called __io_uring_free(), which released the restriction along with the task context, so any ring created after an exec ran with no restrictions at all. Where io_uring restrictions are part of a sandbox or runtime confinement around tenant workloads, a local process can shed that confinement simply by exec'ing and opening a new ring. The CVSS vector records a scope change, consistent with escaping the confinement boundary rather than just corrupting state inside it. The practical footprint depends on whether io_uring is enabled for tenant workloads on your nodes - many GPU clusters disable io_uring outright for untrusted pods, which removes this entirely.

Who can reach it

Local, unprivileged user or container process that is allowed to call io_uring_setup. No authentication beyond code execution on the node; not reachable over the network.

What to do

Pick up a stable kernel containing the fix (commits bc0e8faf90e7 and fcef9325afee); no fixed release number is stated in the record, so check your vendor kernel's changelog. Applying it means installing a kernel and rebooting - drain the node's GPU workloads first, since this cannot be hot-patched into a running io_uring implementation. Interim mitigation on multi-tenant nodes is to block io_uring for untrusted workloads (seccomp profile or kernel.io_uring_disabled), which is cheap and needs no reboot.

References

Related entries

• Linux kernel qla2xxx: MSI-X vector count truncates to zero and the driver dereferences ZERO_SIZE_PTRCVE-2026-89856 · Linux kernel qla2xxx (MSI-X derived queue counts)High
• Linux kernel arm64 ptrace: tracer-modified x0 escapes seccomp re-evaluationCVE-2026-90347 · Linux kernel arm64 ptrace (orig_x0 vs x0 sync on syscall entry)High
• Linux kernel virtio-fs: double free of fs->vqs and fs->mq_map when queue setup fails during probeCVE-2026-93827 · Linux kernel virtio-fs (double free on virtio_fs_setup_vqs() probe failure)High
• Intel Data Center GPU driver for VMware ESXi (out-of-bounds read): Out-of-bounds read in the ESXi GPU driver exposingCVE-2026-20751 · Intel Data Center GPU driver for VMware ESXi (out-of-bounds read)High
• Intel Data Center GPU driver for VMware ESXi (out-of-bounds write): Out-of-bounds write in the ESXi GPU driver causingCVE-2026-20879 · Intel Data Center GPU driver for VMware ESXi (out-of-bounds write)High
• wolfSSL: trusted-peer matching ignores the public key, so a forged CA clone verifiesCVE-2026-93302 · wolfSSL MatchTrustedPeer (trusted peer certificate verification)High