Linux kernel amdkfd (KFD compute driver, /dev/kfd) (drm/amdkfd): A memory or reference-count leak in the amdkfd (KFD

Impact

A memory or reference-count leak in the amdkfd (KFD compute driver, /dev/kfd). Each pass through the affected path drops an allocation or a refcount on the floor. A tenant that loops the operation drives the node into memory exhaustion or pins objects that can never be freed, which on a long-lived GPU host shows up as creeping unreclaimable memory, failed allocations for other tenants, and eventually an OOM kill or a driver that will not unbind. Refcount leaks that wrap can also degrade into use-after-free. Upstream fix: drm/amdkfd: Fix memory leak in kfd_mem_dmamap_userptr()

Who can reach it

Local. Reachable by any process or container with /dev/kfd and /dev/dri/renderD* mapped in - which is every ROCm workload, including an unprivileged tenant pod. Not reachable over the network and not reachable from a container that has no GPU device node mapped in.

What to do

Kernel-side fix: this lands in mainline Linux and flows into distro kernels (RHEL/Rocky, Ubuntu HWE, SLES) and into AMD's out-of-tree DKMS amdgpu package shipped with ROCm. Patch the kernel or the DKMS module, then **reload the amdgpu module or reboot the node** - you cannot fix a running driver in place. Reloading amdgpu requires no process holding /dev/kfd or a render node, so in practice this is a cordon + drain + reboot per node. Plan it as a rolling maintenance across the fleet; there is no VBIOS flash, no SBIOS/AGESA step and no firmware update involved. Nodes running the ROCm DKMS stack often lag mainline by a release or two, so confirm the fix is actually present in the AMD driver version you deploy rather than assuming a new distro kernel covers it. Until the reboot window, the only real mitigation is to stop handing the render node to untrusted workloads - the device plugin has to be mapping /dev/dri/renderD* and /dev/kfd into the container for a tenant to reach this at all.

References