Linux kernel (drivers/iommu/intel): SVA bind and unbind are asymmetric on VT-d hardware without PCI/PRI - bind skips

Impact

SVA bind and unbind are asymmetric on VT-d hardware without PCI/PRI - bind skips enabling I/O page faulting, unbind tries to disable it anyway. The unbind path hits a kernel WARNING every time a tenant closes an SVA-using GPU context, which on any node booted with panic_on_warn is an immediate node kill, and it leaves the device's IOPF enablement state out of step with reality.

Who can reach it

A tenant holding /dev/dri/renderD* on an Intel Xe GPU (the upstream report came from xe_vm_close_and_put) or any SVA-capable accelerator simply closes its GPU VM. Conditional on VT-d with SVA enabled and the device lacking PRI support. Repeatable at will from inside the container, with no host privilege.

What to do

Update to 6.18.39 / 6.20 or later. Interim: do not boot tenant nodes with panic_on_warn, which is what turns this from a log line into an outage, and disable SVA where the workload does not need it.

References