CVE-2026-99632: drm/amdkfd: use the queue's process info for the WPTR BO eviction fence
In the Linux kernel, the following vulnerability has been resolved:
drm/amdkfd: use the queue's process info for the WPTR BO eviction fence
amdgpuamdkfdmapgttbotogart() removes the eviction fence of bo->vmbo->vm->processinfo from the WPTR BO of a new queue. bo->vmbo is the VM which opened the BO last, not necessarily the KFD process' VM, and a render node VM has no process info.
On navi31, an unprivileged process which exports its WPTR BO with AMDKFDIOCEXPORTDMABUF and imports it on a second render node fd oopses in AMDKFDIOCCREATEQUEUE:
BUG: kernel NULL pointer dereference, address: 0000000000000060 RIP: 0010:amdgpuamdkfdmapgttbotogart+0x86/0x1a0 [amdgpu] inituserqueue.part.0+0x180/0x250 [amdgpu] pqmcreatequeue+0x32d/0x730 [amdgpu] kfdioctlcreatequeue+0x583/0x9f0 [amdgpu]
Pass the process info of the queue instead. With the fix, the queue is created and destroyed normally.
(cherry picked from commit 6b4157498a176bd1c6083f6bb5aaf8a257a2027e)
Affected Software
Frequently Asked Questions
Who is exposed to this issue?
Systems using the Linux kernel AMDGPU/KFD stack on Navi31 are exposed when an unprivileged process can use the KFD queue interfaces and render-node file descriptors.
What does an attacker need to do to trigger the failure?
An unprivileged local process must export its WPTR buffer object with AMDKFD_IOC_EXPORT_DMABUF, import it through a second render-node file descriptor, and then create a queue using AMDKFD_IOC_CREATE_QUEUE.
What is the practical impact of successful exploitation?
The described result is a kernel NULL pointer dereference during queue creation, causing an oops. The supplied information does not establish privilege escalation or data disclosure.
How can administrators tell whether the issue is being triggered?
Look for kernel oopses showing a NULL pointer dereference in amdgpu_amdkfd_map_gtt_bo_to_gart, with call frames including init_user_queue, pqm_create_queue, and kfd_ioctl_create_queue.
What changes after applying the fix?
The fix uses the queue's process information rather than process information from the VM that most recently opened the buffer object. Under the described cross-render-node import scenario, queues can then be created and destroyed normally.