In the Linux kernel, the following vulnerability has been resolved:
drm/amdgpu: clamp the isolation index for rings outside a partition
adev->isolation[] has one slot per partition, but a ring that is not
assigned to one keeps AMDGPU_XCP_NO_PARTITION, which is ~0, so indexing
the array with it is out of bounds. SDMA submissions hit this on both
the isolation enforcement and the VM flush path and trip UBSAN.
Fall back to the first slot the way the cleaner shader path already
does, and stop taking the address before the ring type check that makes
it relevant.
CVSS Vector: CVSS:3.1/AV:L/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H
CVSS Score: 7.8
AV:L - The OOB is in amdgpu_device_enforce_isolation() and amdgpu_vm_flush(), reached only via local DRM ioctls on /dev/dri/renderD* (AMDGPU_CS and GEM_VA/VM updates that schedule SDMA jobs through amdgpu_job_prepare_job/amdgpu_ib_schedule). There is no network, adjacent-radio, or physical packet path.
AC:L - Unassigned rings already have xcp_id=AMDGPU_XCP_NO_PARTITION (~0), so indexing adev->isolation[MAX_XCP] is deterministic on each SDMA or GFX/compute job. The attacker only needs to submit ordinary command buffers or VA maps; no race or uninfluenced layout is required.
PR:L - AMDGPU_CS, GEM_VA, and related ioctls are DRM_AUTH|DRM_RENDER_ALLOW with no capability check, so any unprivileged local user with render-node access (render/video group, Android GPU clients, multi-tenant cloud GPU VMs) can reach the scheduler without init-namespace root.
UI:N - The attacker opens the render node and submits their own command buffers or GEM_VA maps; no separate victim action such as mounting a filesystem, opening a file, or interacting with another user's GPU context is required.
S:U - The isolation[] out-of-bounds access corrupts host kernel memory inside the amdgpu driver. It does not cross a VM, IOMMU/DMA, or guest-to-host boundary into a different security authority.
C:H - Indexing isolation[ring->xcp_id] with ~0 is an unbounded kernel out-of-bounds access. amdgpu_vm_flush reads isolation->spearhead and enforce_isolation reads spearhead/owner/sync fence objects, which can disclose kernel pointers; CNA guidance scores OOB reads High unless limited to a few bytes.
I:H - For GFX/compute rings, amdgpu_device_enforce_isolation writes isolation->spearhead, owner, and amdgpu_sync hash tables at the OOB address and dma_fence_put()s a wild fence pointer, which is an out-of-bounds write and UAF/control-flow primitive scored High.
A:H - The ~0 index yields a wildly invalid pointer whose dereference oopses the host, and UBSAN array-index-out-of-bounds aborts under CONFIG_UBSAN_TRAP or panic_on_warn. Either outcome is a kernel crash that fully denies GPU/host availability.
| Attack Vector |
Local |
Scope |
Unchanged |
| Attack Complexity |
Low |
Confidentiality Impact |
High |
| Privileges Required |
Low |
Integrity Impact |
High |
| User Interaction |
None |
Availability Impact |
High |
AV:L - The OOB is in amdgpu_device_enforce_isolation() and amdgpu_vm_flush(), reached only via local DRM ioctls on /dev/dri/renderD* (AMDGPU_CS and GEM_VA/VM updates that schedule SDMA jobs through amdgpu_job_prepare_job/amdgpu_ib_schedule). There is no network, adjacent-radio, or physical packet path.
AC:L - Unassigned rings already have xcp_id=AMDGPU_XCP_NO_PARTITION (~0), so indexing adev->isolation[MAX_XCP] is deterministic on each SDMA or GFX/compute job. The attacker only needs to submit ordinary command buffers or VA maps; no race or uninfluenced layout is required.
PR:L - AMDGPU_CS, GEM_VA, and related ioctls are DRM_AUTH|DRM_RENDER_ALLOW with no capability check, so any unprivileged local user with render-node access (render/video group, Android GPU clients, multi-tenant cloud GPU VMs) can reach the scheduler without init-namespace root.
UI:N - The attacker opens the render node and submits their own command buffers or GEM_VA maps; no separate victim action such as mounting a filesystem, opening a file, or interacting with another user's GPU context is required.
S:U - The isolation[] out-of-bounds access corrupts host kernel memory inside the amdgpu driver. It does not cross a VM, IOMMU/DMA, or guest-to-host boundary into a different security authority.
C:H - Indexing isolation[ring->xcp_id] with ~0 is an unbounded kernel out-of-bounds access. amdgpu_vm_flush reads isolation->spearhead and enforce_isolation reads spearhead/owner/sync fence objects, which can disclose kernel pointers; CNA guidance scores OOB reads High unless limited to a few bytes.
I:H - For GFX/compute rings, amdgpu_device_enforce_isolation writes isolation->spearhead, owner, and amdgpu_sync hash tables at the OOB address and dma_fence_put()s a wild fence pointer, which is an out-of-bounds write and UAF/control-flow primitive scored High.
A:H - The ~0 index yields a wildly invalid pointer whose dereference oopses the host, and UBSAN array-index-out-of-bounds aborts under CONFIG_UBSAN_TRAP or panic_on_warn. Either outcome is a kernel crash that fully denies GPU/host availability.
CVSS 3.1