Linux kernel (drivers/gpu/drm/nouveau): Nouveau's VM_BIND remap path miscalculates the address and range of the unmap

Impact

Nouveau's VM_BIND remap path miscalculates the address and range of the unmap that accompanies a partial rebind, so it tears down GPU page-table entries well outside the region the tenant actually asked to remap. The reported effect is corrupted page tables and a kernel oops - meaning a tenant's bind operations can scribble on GPU mappings that are not theirs and take the node down.

Who can reach it

Driven from userspace by a container holding /dev/dri/renderD* on a nouveau-driven NVIDIA GPU: it was found by a standard Vulkan sparse-resources conformance test, i.e. ordinary sparse-binding workloads hit it without trying. Conditional on the open nouveau driver with the uvmm/VM_BIND interface in use.

What to do

Update to a kernel with the fix commits below. Interim: blacklist nouveau on nodes that run the proprietary NVIDIA driver, or block sparse-binding tenants from the affected hosts.

References