GPU VulnDB

Database/Kernel, userspace & hypervisor

Linux arm64: SVE signal-context restore can leave a task invalid and read out of bounds

CVSS 7.1CVE-2026-23102Kernel, userspace & hypervisorcurated

Impact

On arm64 CPUs that implement SME, restoring a crafted or foreign SVE signal context could leave a task with SVCR.SM set while TIF_SME was clear. In that state the kernel restores streaming-mode SVE registers without configuring SMCR_ELx.LEN, so it may read past the task's allocated sve_state - unrelated kernel memory, or a fatal fault - and the task can later be killed with SIGKILL when the live vector length does not match. The commit message is explicit that the values loaded are never exposed to EL0 and are never written back, so this is a stability and kernel-memory-read bug, not a disclosure path. It matters on arm64 GPU hosts (Grace-class and Arm server nodes) because a fault here is a host-side crash that costs a drain, and because CRIU-style checkpoint/restore of jobs is one of the legitimate ways to reach the state.

Who can reach it

Local, unprivileged user code on an arm64 host with SME: a process modifies its own SVE signal frame, restores a frame saved in another task's context (CRIU), or uses ptrace to set SVCR.SM after sigreturn entry. Not reachable remotely and not reachable on x86 nodes.

What to do

Update to a stable kernel with the fpsimd signal fix (set TIF_SME on restore, clear SVCR.SM when the flag is absent); four stable commits are linked. This is a kernel update plus a reboot of each arm64 node, so drain the GPU workloads first. x86_64 fleets are unaffected and need no action.

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.