/** @seqno: Sequence number of the last initialized fence. */
                atomic64_t seqno;
 
+               /**
+                * @last_fence: Fence of the last submitted job.
+                *
+                * We return this fence when we get an empty command stream.
+                * This way, we are guaranteed that all earlier jobs have completed
+                * when drm_sched_job::s_fence::finished without having to feed
+                * the CS ring buffer with a dummy job that only signals the fence.
+                */
+               struct dma_fence *last_fence;
+
                /**
                 * @in_flight_jobs: List containing all in-flight jobs.
                 *
        panthor_kernel_bo_destroy(queue->ringbuf);
        panthor_kernel_bo_destroy(queue->iface.mem);
 
+       /* Release the last_fence we were holding, if any. */
+       dma_fence_put(queue->fence_ctx.last_fence);
+
        kfree(queue);
 }
 
 
                spin_lock(&queue->fence_ctx.lock);
                list_for_each_entry_safe(job, job_tmp, &queue->fence_ctx.in_flight_jobs, node) {
-                       if (!job->call_info.size)
-                               continue;
-
                        if (syncobj->seqno < job->done_fence->seqno)
                                break;
 
        static_assert(sizeof(call_instrs) % 64 == 0,
                      "call_instrs is not aligned on a cacheline");
 
-       /* Stream size is zero, nothing to do => return a NULL fence and let
-        * drm_sched signal the parent.
+       /* Stream size is zero, nothing to do except making sure all previously
+        * submitted jobs are done before we signal the
+        * drm_sched_job::s_fence::finished fence.
         */
-       if (!job->call_info.size)
-               return NULL;
+       if (!job->call_info.size) {
+               job->done_fence = dma_fence_get(queue->fence_ctx.last_fence);
+               return dma_fence_get(job->done_fence);
+       }
 
        ret = pm_runtime_resume_and_get(ptdev->base.dev);
        if (drm_WARN_ON(&ptdev->base, ret))
                }
        }
 
+       /* Update the last fence. */
+       dma_fence_put(queue->fence_ctx.last_fence);
+       queue->fence_ctx.last_fence = dma_fence_get(job->done_fence);
+
        done_fence = dma_fence_get(job->done_fence);
 
 out_unlock:
                goto err_put_job;
        }
 
-       job->done_fence = kzalloc(sizeof(*job->done_fence), GFP_KERNEL);
-       if (!job->done_fence) {
-               ret = -ENOMEM;
-               goto err_put_job;
+       /* Empty command streams don't need a fence, they'll pick the one from
+        * the previously submitted job.
+        */
+       if (job->call_info.size) {
+               job->done_fence = kzalloc(sizeof(*job->done_fence), GFP_KERNEL);
+               if (!job->done_fence) {
+                       ret = -ENOMEM;
+                       goto err_put_job;
+               }
        }
 
        ret = drm_sched_job_init(&job->base,
 
         * Must be 64-bit/8-byte aligned (the size of a CS instruction)
         *
         * Can be zero if stream_addr is zero too.
+        *
+        * When the stream size is zero, the queue submit serves as a
+        * synchronization point.
         */
        __u32 stream_size;
 
         * ensure the GPU doesn't get garbage when reading the indirect command
         * stream buffers. If you want the cache flush to happen
         * unconditionally, pass a zero here.
+        *
+        * Ignored when stream_size is zero.
         */
        __u32 latest_flush;