drm/amdgpu: Force signal hw_fences that are embedded in non-sched jobs

From YuBiao Wang
b4efb6af93ca755f5469114adfd7f26bb944962c in linux-6.1.y/6.1.25
033c56474acf567a450f8bafca50e0b610f2b716 in mainline linux

1 files changed, 9 insertions, 0 deletions

diff --git a/sys/dev/pci/drm/amd/amdgpu/amdgpu_fence.c b/sys/dev/pci/drm/amd/amdgpu/amdgpu_fence.c
index b06a0f06800..605b4c1b25e 100644
--- a/sys/dev/pci/drm/amd/amdgpu/amdgpu_fence.c
+++ b/sys/dev/pci/drm/amd/amdgpu/amdgpu_fence.c
@@ -644,6 +644,15 @@ void amdgpu_fence_driver_clear_job_fences(struct amdgpu_ring *ring)
 		ptr = &ring->fence_drv.fences[i];
 		old = rcu_dereference_protected(*ptr, 1);
 		if (old && old->ops == &amdgpu_job_fence_ops) {
+			struct amdgpu_job *job;
+
+			/* For non-scheduler bad job, i.e. failed ib test, we need to signal
+			 * it right here or we won't be able to track them in fence_drv
+			 * and they will remain unsignaled during sa_bo free.
+			 */
+			job = container_of(old, struct amdgpu_job, hw_fence);
+			if (!job->base.s_fence && !dma_fence_is_signaled(old))
+				dma_fence_signal(old);
 			RCU_INIT_POINTER(*ptr, NULL);
 			dma_fence_put(old);
 		}