Intel Xe Linux Driver Adds VRAM Health and Bad-Page Handling

Intel Xe graphics-driver changes queued for the Linux 7.4 development cycle add two related but distinct mechanisms around video-memory reliability: an early-probe VRAM health check for catching layout mistakes, and memory-page offlining that can keep identified bad GPU-memory pages out of future allocations.

The distinction matters because neither feature means Linux can repair physically failing VRAM. The probe-time check is primarily a development safeguard against driver or platform layout errors. Page offlining is a containment mechanism: when supported hardware reports a bad location, the driver can retire that memory instead of returning it to applications.

Intel engineer Matthew Auld describes the final VRAM page as a canary. During early device probing, the Xe driver can exercise that page through the PCI BAR mapping, Compression Control Surface sizing and identity-map setup. If one of those calculations is wrong, the edge of the usable address range is a useful place to expose the mistake.

Auld specifically points to CCS sizing on some Battlemage hardware, where a misaligned compression-storage offset exposed broken driver handling and allowed part of the reserved CCS storage to enter the allocator as if it were normal VRAM. The new check is therefore best understood as regression detection for memory-layout assumptions, not as a full memory test comparable to hours-long hardware diagnostics.

The larger reliability change lets the Xe driver permanently carve identified faulty video-memory pages out of service so later allocations do not reuse them. That is a familiar reliability principle in systems with error-aware memory: preserve operation where possible by shrinking the usable pool instead of repeatedly touching a known-bad location.

The driver also exposes information about offlined and pending pages through DebugFS. That can help kernel developers and operators establish whether a reported memory event was isolated. It should not be confused with Intel Xe’s separate GPU-health sysfs interface, which can expose broader ok, warning or critical device-health states on supported platforms.

For a typical gaming PC, these patches are more about making the graphics stack robust and diagnosable than delivering a new performance feature. A memory-layout regression can masquerade as corruption, hangs or a dead display, while genuine hardware degradation can produce similarly confusing symptoms. Giving the driver better ways to detect boundary mistakes and quarantine known-bad memory makes those cases easier to separate.

The changes are queued for the Linux 7.4 cycle rather than guaranteed to be present in a currently installed distribution kernel. Exact availability will depend on upstream integration, kernel release timing and distribution packaging. And if a GPU continues accumulating bad pages, offlining should be treated as evidence of degrading hardware, not as a reason to ignore the underlying fault.