drm/amdkfd: Don't assign dGPUs to APU topology devices

dGPUs need their own topology devices. Don't assign them to APU topology
devices with CPU cores.

Bug: https://github.com/RadeonOpenCompute/ROCK-Kernel-Driver/issues/66
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Tested-by: Elias Konstantinidis <ekondis@gmail.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
index 4734f1a35516ade4846fb36ff1476be6ea1fcd5e..09da91644f9fc6e15d29c628d6e16f1a2ddd0f69 100644 (file)
--- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
@@ -1093,8 +1093,6 @@ static uint32_t kfd_generate_gpu_id(struct kfd_dev *gpu)
  *             the GPU device is not already present in the topology device
  *             list then return NULL. This means a new topology device has to
  *             be created for this GPU.
- * TODO: Rather than assiging @gpu to first topology device withtout
- *             gpu attached, it will better to have more stringent check.
  */
 static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)
 {
@@ -1102,12 +1100,20 @@ static struct kfd_topology_device *kfd_assign_gpu(struct kfd_dev *gpu)
        struct kfd_topology_device *out_dev = NULL;
 
        down_write(&topology_lock);
-       list_for_each_entry(dev, &topology_device_list, list)
+       list_for_each_entry(dev, &topology_device_list, list) {
+               /* Discrete GPUs need their own topology device list
+                * entries. Don't assign them to CPU/APU nodes.
+                */
+               if (!gpu->device_info->needs_iommu_device &&
+                   dev->node_props.cpu_cores_count)
+                       continue;
+
                if (!dev->gpu && (dev->node_props.simd_count > 0)) {
                        dev->gpu = gpu;
                        out_dev = dev;
                        break;
                }
+       }
        up_write(&topology_lock);
        return out_dev;
 }