CVE-2026-46309: Linux xe GPU Driver Memory Disclosure via Incoherent Cache Access

Technical summary

The vulnerability exists in the xe_vm_madvise_ioctl() function of the Linux kernel's xe GPU driver. When userspace applies XE_COH_NONE (no coherency) PAT indices to CPU-cached memory regions via madvise, the GPU gains direct DRAM access that circumvents CPU cache coherence. During kernel page clearing (a security measure before reallocation), cleared pages remain dirty in CPU cache; a GPU with coh_none coherency can read these stale contents directly from DRAM, extracting sensitive data from other processes' freed allocations. The patch adds validation logic to reject this configuration on integrated GPUs, aligning with existing protections in the vm_bind path.

Business impact

Data exfiltration from freed memory across process boundaries poses a significant confidentiality risk in multi-tenant environments and systems handling sensitive data. An unprivileged local attacker with GPU access can read kernel buffers, other process memory, or cryptographic material. On shared infrastructure (cloud VMs, containerized workloads), this enables lateral privilege escalation and sensitive data theft. Remediation is straightforward and involves kernel patching with no functional impact to legitimate GPU operations.

Affected systems

Linux systems with Intel integrated GPUs (iGPU) running kernel versions prior to the patch. The restriction applies specifically to iGPU deployments; discrete GPU scenarios have different coherency semantics. Systems without Intel integrated graphics are unaffected. Verify your kernel version and GPU configuration against the vendor advisory.

Exploitability

Exploitation requires local code execution with sufficient privilege to invoke the madvise GPU memory ioctl. The attack is reliable once conditions are met but not remotely exploitable. No active exploitation is tracked in the KEV catalog. An attacker must understand GPU memory management APIs and timing to reliably capture stale data, but the underlying flaw is deterministic.

Remediation

Apply the kernel patch that adds validation to xe_vm_madvise_ioctl() to reject coh_none PAT indices on CPU-cached memory. Rebuild and deploy the updated kernel. No configuration changes or workarounds are required post-patch. Verify the patch commit 016ccdb674b8c899940b3944952c96a6a490d10a or later is included in your kernel release.

Patch guidance

Update your Linux kernel to a version containing the fix for CVE-2026-46309. Consult your distribution's advisory for specific patch version numbers and availability. For environments using Intel iGPUs in VMs or containers, prioritize this update to prevent cross-process memory leakage. Patch deployment can typically occur during standard maintenance windows as the fix carries no breaking changes.

Detection guidance

Monitor for unusual GPU madvise ioctl invocations, particularly those combining coh_none coherency flags with CPU-cached memory regions. Kernel audit logs (auditd) can capture madvise syscall patterns. GPU driver logs may reveal rejected PAT configuration attempts post-patch. Look for process behaviors attempting to enumerate or manipulate GPU memory coherency settings outside normal application scope. YARA rules targeting GPU memory API abuse patterns would be effective.

Why prioritize this

This vulnerability merits urgent attention due to its high CVSS (7.0), silent data exfiltration capability, and applicability to common deployment scenarios (cloud VMs, containers with iGPU passthrough). The privileged access requirement moderates risk slightly, but the lack of detection and deterministic nature of the flaw justify swift patching. It is not yet in active exploitation (KEV status false) but fits attacker profiles targeting multi-tenant isolation.

Risk score, explained

CVSS 7.0 (HIGH) reflects a high confidentiality impact with moderate complexity and low attack privilege. The local attack vector, requirement for higher privilege (PR:L), and moderate barriers to exploitability (AC:H) prevent a critical rating, but the integrity and availability impacts (writing stale data and potential DoS via resource exhaustion) elevate severity. The ability to leak across process boundaries and the silent nature of the attack justify the score.

Frequently asked questions