drm/nouveau: implement 1:1 scheduler - entity relationship

Recent patches to the DRM scheduler [1][2] allow for a variable number
of run-queues and add support for (shared) workqueues rather than
dedicated kthreads per scheduler. This allows us to create a 1:1
relationship between a GPU scheduler and a scheduler entity, in order to
properly support firmware schedulers being able to handle an arbitrary
amount of dynamically allocated command ring buffers. This perfectly
matches Nouveau's needs, hence make use of it.

Topology wise we create one scheduler instance per client (handling
VM_BIND jobs) and one scheduler instance per channel (handling EXEC
jobs).

All channel scheduler instances share a workqueue, but every client
scheduler instance has a dedicated workqueue. The latter is required to
ensure that for VM_BIND job's free_job() work and run_job() work can
always run concurrently and hence, free_job() work can never stall
run_job() work. For EXEC jobs we don't have this requirement, since EXEC
job's free_job() does not require to take any locks which indirectly or
directly are held for allocations elsewhere.

[1] https://lore.kernel.org/all/8f53f7ef-7621-4f0b-bdef-d8d20bc497ff@redhat.com/T/
[2] https://lore.kernel.org/all/20231031032439.1558703-1-matthew.brost@intel.com/T/

Signed-off-by: Danilo Krummrich <dakr@redhat.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20231114002728.3491-1-dakr@redhat.com

1 files changed, 109 insertions, 88 deletions

diff --git a/drivers/gpu/drm/nouveau/nouveau_sched.c b/drivers/gpu/drm/nouveau/nouveau_sched.c
index f69f4cbf251b..0f4583f6fcad 100644
--- a/drivers/gpu/drm/nouveau/nouveau_sched.c
+++ b/drivers/gpu/drm/nouveau/nouveau_sched.c
@@ -12,30 +12,29 @@
 #include "nouveau_abi16.h"
 #include "nouveau_sched.h"
 
-/* FIXME
- *
- * We want to make sure that jobs currently executing can't be deferred by
- * other jobs competing for the hardware. Otherwise we might end up with job
- * timeouts just because of too many clients submitting too many jobs. We don't
- * want jobs to time out because of system load, but because of the job being
- * too bulky.
- *
- * For now allow for up to 16 concurrent jobs in flight until we know how many
- * rings the hardware can process in parallel.
- */
-#define NOUVEAU_SCHED_HW_SUBMISSIONS		16
+#define NOUVEAU_SCHED_HW_SUBMISSIONS		1
 #define NOUVEAU_SCHED_JOB_TIMEOUT_MS		10000
 
+/* Starts at 0, since the DRM scheduler interprets those parameters as (initial)
+ * index to the run-queue array.
+ */
+enum nouveau_sched_priority {
+	NOUVEAU_SCHED_PRIORITY_SINGLE = DRM_SCHED_PRIORITY_MIN,
+	NOUVEAU_SCHED_PRIORITY_COUNT,
+};
+
 int
 nouveau_job_init(struct nouveau_job *job,
 		 struct nouveau_job_args *args)
 {
-	struct nouveau_sched_entity *entity = args->sched_entity;
+	struct nouveau_sched *sched = args->sched;
 	int ret;
 
+	INIT_LIST_HEAD(&job->entry);
+
 	job->file_priv = args->file_priv;
 	job->cli = nouveau_cli(args->file_priv);
-	job->entity = entity;
+	job->sched = sched;
 
 	job->sync = args->sync;
 	job->resv_usage = args->resv_usage;
@@ -89,7 +88,7 @@ nouveau_job_init(struct nouveau_job *job,
 
 	}
 
-	ret = drm_sched_job_init(&job->base, &entity->base, 1, NULL);
+	ret = drm_sched_job_init(&job->base, &sched->entity, 1, NULL);
 	if (ret)
 		goto err_free_chains;
 
@@ -109,6 +108,27 @@ return ret;
 }
 
 void
+nouveau_job_fini(struct nouveau_job *job)
+{
+	dma_fence_put(job->done_fence);
+	drm_sched_job_cleanup(&job->base);
+
+	job->ops->free(job);
+}
+
+void
+nouveau_job_done(struct nouveau_job *job)
+{
+	struct nouveau_sched *sched = job->sched;
+
+	spin_lock(&sched->job.list.lock);
+	list_del(&job->entry);
+	spin_unlock(&sched->job.list.lock);
+
+	wake_up(&sched->job.wq);
+}
+
+void
 nouveau_job_free(struct nouveau_job *job)
 {
 	kfree(job->in_sync.data);
@@ -117,13 +137,6 @@ nouveau_job_free(struct nouveau_job *job)
 	kfree(job->out_sync.chains);
 }
 
-void nouveau_job_fini(struct nouveau_job *job)
-{
-	dma_fence_put(job->done_fence);
-	drm_sched_job_cleanup(&job->base);
-	job->ops->free(job);
-}
-
 static int
 sync_find_fence(struct nouveau_job *job,
 		struct drm_nouveau_sync *sync,
@@ -261,7 +274,7 @@ nouveau_job_fence_attach(struct nouveau_job *job)
 int
 nouveau_job_submit(struct nouveau_job *job)
 {
-	struct nouveau_sched_entity *entity = to_nouveau_sched_entity(job->base.entity);
+	struct nouveau_sched *sched = job->sched;
 	struct dma_fence *done_fence = NULL;
 	struct drm_gpuvm_exec vm_exec = {
 		.vm = &nouveau_cli_uvmm(job->cli)->base,
@@ -281,7 +294,7 @@ nouveau_job_submit(struct nouveau_job *job)
 	/* Make sure the job appears on the sched_entity's queue in the same
 	 * order as it was submitted.
 	 */
-	mutex_lock(&entity->mutex);
+	mutex_lock(&sched->mutex);
 
 	/* Guarantee we won't fail after the submit() callback returned
 	 * successfully.
@@ -292,33 +305,16 @@ nouveau_job_submit(struct nouveau_job *job)
 			goto err_cleanup;
 	}
 
+	/* Submit was successful; add the job to the schedulers job list. */
+	spin_lock(&sched->job.list.lock);
+	list_add(&job->entry, &sched->job.list.head);
+	spin_unlock(&sched->job.list.lock);
+
 	drm_sched_job_arm(&job->base);
 	job->done_fence = dma_fence_get(&job->base.s_fence->finished);
 	if (job->sync)
 		done_fence = dma_fence_get(job->done_fence);
 
-	/* If a sched job depends on a dma-fence from a job from the same GPU
-	 * scheduler instance, but a different scheduler entity, the GPU
-	 * scheduler does only wait for the particular job to be scheduled,
-	 * rather than for the job to fully complete. This is due to the GPU
-	 * scheduler assuming that there is a scheduler instance per ring.
-	 * However, the current implementation, in order to avoid arbitrary
-	 * amounts of kthreads, has a single scheduler instance while scheduler
-	 * entities represent rings.
-	 *
-	 * As a workaround, set the DRM_SCHED_FENCE_DONT_PIPELINE for all
-	 * out-fences in order to force the scheduler to wait for full job
-	 * completion for dependent jobs from different entities and same
-	 * scheduler instance.
-	 *
-	 * There is some work in progress [1] to address the issues of firmware
-	 * schedulers; once it is in-tree the scheduler topology in Nouveau
-	 * should be re-worked accordingly.
-	 *
-	 * [1] https://lore.kernel.org/dri-devel/20230801205103.627779-1-matthew.brost@intel.com/
-	 */
-	set_bit(DRM_SCHED_FENCE_DONT_PIPELINE, &job->done_fence->flags);
-
 	if (job->ops->armed_submit)
 		job->ops->armed_submit(job, &vm_exec);
 
@@ -331,7 +327,7 @@ nouveau_job_submit(struct nouveau_job *job)
 
 	drm_sched_entity_push_job(&job->base);
 
-	mutex_unlock(&entity->mutex);
+	mutex_unlock(&sched->mutex);
 
 	if (done_fence) {
 		dma_fence_wait(done_fence, true);
@@ -341,20 +337,13 @@ nouveau_job_submit(struct nouveau_job *job)
 	return 0;
 
 err_cleanup:
-	mutex_unlock(&entity->mutex);
+	mutex_unlock(&sched->mutex);
 	nouveau_job_fence_attach_cleanup(job);
 err:
 	job->state = NOUVEAU_JOB_SUBMIT_FAILED;
 	return ret;
 }
 
-bool
-nouveau_sched_entity_qwork(struct nouveau_sched_entity *entity,
-			   struct work_struct *work)
-{
-	return queue_work(entity->sched_wq, work);
-}
-
 static struct dma_fence *
 nouveau_job_run(struct nouveau_job *job)
 {
@@ -404,50 +393,82 @@ nouveau_sched_free_job(struct drm_sched_job *sched_job)
 	nouveau_job_fini(job);
 }
 
-int nouveau_sched_entity_init(struct nouveau_sched_entity *entity,
-			      struct drm_gpu_scheduler *sched,
-			      struct workqueue_struct *sched_wq)
-{
-	mutex_init(&entity->mutex);
-	spin_lock_init(&entity->job.list.lock);
-	INIT_LIST_HEAD(&entity->job.list.head);
-	init_waitqueue_head(&entity->job.wq);
-
-	entity->sched_wq = sched_wq;
-	return drm_sched_entity_init(&entity->base,
-				     DRM_SCHED_PRIORITY_NORMAL,
-				     &sched, 1, NULL);
-}
-
-void
-nouveau_sched_entity_fini(struct nouveau_sched_entity *entity)
-{
-	drm_sched_entity_destroy(&entity->base);
-}
-
 static const struct drm_sched_backend_ops nouveau_sched_ops = {
 	.run_job = nouveau_sched_run_job,
 	.timedout_job = nouveau_sched_timedout_job,
 	.free_job = nouveau_sched_free_job,
 };
 
-int nouveau_sched_init(struct nouveau_drm *drm)
+int
+nouveau_sched_init(struct nouveau_sched *sched, struct nouveau_drm *drm,
+		   struct workqueue_struct *wq)
 {
-	struct drm_gpu_scheduler *sched = &drm->sched;
+	struct drm_gpu_scheduler *drm_sched = &sched->base;
+	struct drm_sched_entity *entity = &sched->entity;
 	long job_hang_limit = msecs_to_jiffies(NOUVEAU_SCHED_JOB_TIMEOUT_MS);
+	int ret;
 
-	drm->sched_wq = create_singlethread_workqueue("nouveau_sched_wq");
-	if (!drm->sched_wq)
-		return -ENOMEM;
+	if (!wq) {
+		wq = alloc_workqueue("nouveau_sched_wq_%d", 0, WQ_MAX_ACTIVE,
+				     current->pid);
+		if (!wq)
+			return -ENOMEM;
+
+		sched->wq = wq;
+	}
 
-	return drm_sched_init(sched, &nouveau_sched_ops, NULL,
-			      DRM_SCHED_PRIORITY_COUNT,
-			      NOUVEAU_SCHED_HW_SUBMISSIONS, 0, job_hang_limit,
-			      NULL, NULL, "nouveau_sched", drm->dev->dev);
+	ret = drm_sched_init(drm_sched, &nouveau_sched_ops, wq,
+			     NOUVEAU_SCHED_PRIORITY_COUNT,
+			     NOUVEAU_SCHED_HW_SUBMISSIONS, 0, job_hang_limit,
+			     NULL, NULL, "nouveau_sched", drm->dev->dev);
+	if (ret)
+		goto fail_wq;
+
+	/* Using DRM_SCHED_PRIORITY_MIN, since that's what we're required to use
+	 * when we want to have a single run-queue only.
+	 *
+	 * It's not documented, but one will find out when trying to use any
+	 * other priority running into faults, because the scheduler uses the
+	 * priority as array index.
+	 *
+	 * Can't use NOUVEAU_SCHED_PRIORITY_SINGLE either, because it's not
+	 * matching the enum type used in drm_sched_entity_init().
+	 */
+	ret = drm_sched_entity_init(entity, DRM_SCHED_PRIORITY_MIN,
+				    &drm_sched, 1, NULL);
+	if (ret)
+		goto fail_sched;
+
+	mutex_init(&sched->mutex);
+	spin_lock_init(&sched->job.list.lock);
+	INIT_LIST_HEAD(&sched->job.list.head);
+	init_waitqueue_head(&sched->job.wq);
+
+	return 0;
+
+fail_sched:
+	drm_sched_fini(drm_sched);
+fail_wq:
+	if (sched->wq)
+		destroy_workqueue(sched->wq);
+	return ret;
 }
 
-void nouveau_sched_fini(struct nouveau_drm *drm)
+void
+nouveau_sched_fini(struct nouveau_sched *sched)
 {
-	destroy_workqueue(drm->sched_wq);
-	drm_sched_fini(&drm->sched);
+	struct drm_gpu_scheduler *drm_sched = &sched->base;
+	struct drm_sched_entity *entity = &sched->entity;
+
+	rmb(); /* for list_empty to work without lock */
+	wait_event(sched->job.wq, list_empty(&sched->job.list.head));
+
+	drm_sched_entity_fini(entity);
+	drm_sched_fini(drm_sched);
+
+	/* Destroy workqueue after scheduler tear down, otherwise it might still
+	 * be in use.
+	 */
+	if (sched->wq)
+		destroy_workqueue(sched->wq);
 }