Linux kernel drm/xe: madvise accepts a non-coherent PAT index on CPU-cached memory, leaking freed page contents

CVSS 7.0CVE-2026-46309NVIDIA / GPU stackcurated

Impact

The xe driver's madvise ioctl did not reject PAT indices with XE_COH_NONE coherency for CPU-cached buffers, a check the vm_bind path already performed. When the kernel clears a page before handing it to a new owner, the zeroing can still sit dirty in CPU cache; a GPU mapping with coherency disabled bypasses the cache and reads the old DRAM contents, so a local process can pull data out of pages freed by other processes. On a shared accelerator node that is a cross-tenant read primitive that needs nothing but a render/compute context. The upstream fix limits the restriction to integrated GPUs, so the practical exposure is on hosts using Intel Xe iGPU contexts rather than discrete cards.

Who can reach it

Local user with an Xe GPU context (access to the DRM render node) who can issue the madvise ioctl. No special privilege beyond GPU device access; not reachable remotely.

What to do

Pick up the stable kernel containing commit 016ccdb674b8 (backports 4e5591c2fc1b, 87f9b1528e1f, fea04cf6f234) or your distribution's patched kernel, then reboot each affected node - there is no runtime toggle and the driver cannot be reloaded under live GPU workloads. Schedule it with normal kernel maintenance; drain the node first.

References

https://nvd.nist.gov/vuln/detail/CVE-2026-46309 https://git.kernel.org/stable/c/4e5591c2fc1b30f4ea5e2eab4c3a695acc404e39 https://access.redhat.com/security/cve/CVE-2026-46309 https://bugzilla.redhat.com/show_bug.cgi?id=2486468

Related entries

All NVIDIA / GPU stack 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.