drm/nouveau: enable dynamic job-flow control

Make use of the scheduler's credit limit and scheduler job's credit
count to account for the actual size of a job, such that we fill up the
ring efficiently.
Signed-off-by: Danilo Krummrich <dakr@redhat.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20231114002728.3491-2-dakr@redhat.com

drivers/gpu/drm/nouveau/nouveau_abi16.c

@@ -337,7 +337,8 @@ nouveau_abi16_ioctl_channel_alloc(ABI16_IOCTL_ARGS)
 	if (ret)
 		goto done;
 
-	ret = nouveau_sched_init(&chan->sched, drm, drm->sched_wq);
+	ret = nouveau_sched_init(&chan->sched, drm, drm->sched_wq,
+				 chan->chan->dma.ib_max);
 	if (ret)
 		goto done;
 

drivers/gpu/drm/nouveau/nouveau_drm.c

@@ -320,7 +320,7 @@ nouveau_cli_init(struct nouveau_drm *drm, const char *sname,
 	 * locks which indirectly or directly are held for allocations
 	 * elsewhere.
 	 */
-	ret = nouveau_sched_init(&cli->sched, drm, NULL);
+	ret = nouveau_sched_init(&cli->sched, drm, NULL, 1);
 	if (ret)
 		goto done;
 

drivers/gpu/drm/nouveau/nouveau_exec.c

@@ -231,10 +231,12 @@ nouveau_exec_job_init(struct nouveau_exec_job **pjob,
 		}
 	}
 
+	args.file_priv = __args->file_priv;
 	job->chan = __args->chan;
 
 	args.sched = __args->sched;
-	args.file_priv = __args->file_priv;
+	/* Plus one to account for the HW fence. */
+	args.credits = job->push.count + 1;
 
 	args.in_sync.count = __args->in_sync.count;
 	args.in_sync.s = __args->in_sync.s;

drivers/gpu/drm/nouveau/nouveau_sched.c

@@ -12,7 +12,6 @@
 #include "nouveau_abi16.h"
 #include "nouveau_sched.h"
 
-#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)
@@ -85,10 +84,10 @@ nouveau_job_init(struct nouveau_job *job,
 			ret = -ENOMEM;
 			goto err_free_objs;
 		}
-
 	}
 
-	ret = drm_sched_job_init(&job->base, &sched->entity, 1, NULL);
+	ret = drm_sched_job_init(&job->base, &sched->entity,
+				 args->credits, NULL);
 	if (ret)
 		goto err_free_chains;
 
@@ -401,7 +400,7 @@ static const struct drm_sched_backend_ops nouveau_sched_ops = {
 
 int
 nouveau_sched_init(struct nouveau_sched *sched, struct nouveau_drm *drm,
-		   struct workqueue_struct *wq)
+		   struct workqueue_struct *wq, u32 credit_limit)
 {
 	struct drm_gpu_scheduler *drm_sched = &sched->base;
 	struct drm_sched_entity *entity = &sched->entity;
@@ -419,7 +418,7 @@ nouveau_sched_init(struct nouveau_sched *sched, struct nouveau_drm *drm,
 
 	ret = drm_sched_init(drm_sched, &nouveau_sched_ops, wq,
 			     NOUVEAU_SCHED_PRIORITY_COUNT,
-			     NOUVEAU_SCHED_HW_SUBMISSIONS, 0, job_hang_limit,
+			     credit_limit, 0, job_hang_limit,
 			     NULL, NULL, "nouveau_sched", drm->dev->dev);
 	if (ret)
 		goto fail_wq;

drivers/gpu/drm/nouveau/nouveau_sched.h

@@ -27,6 +27,7 @@ enum nouveau_job_state {
 struct nouveau_job_args {
 	struct drm_file *file_priv;
 	struct nouveau_sched *sched;
+	u32 credits;
 
 	enum dma_resv_usage resv_usage;
 	bool sync;
@@ -111,7 +112,7 @@ struct nouveau_sched {
 };
 
 int nouveau_sched_init(struct nouveau_sched *sched, struct nouveau_drm *drm,
-		       struct workqueue_struct *wq);
+		       struct workqueue_struct *wq, u32 credit_limit);
 void nouveau_sched_fini(struct nouveau_sched *sched);
 
 #endif

drivers/gpu/drm/nouveau/nouveau_uvmm.c

@@ -1607,9 +1607,11 @@ nouveau_uvmm_bind_job_init(struct nouveau_uvmm_bind_job **pjob,
 
 	init_completion(&job->complete);
 
-	args.sched = __args->sched;
 	args.file_priv = __args->file_priv;
 
+	args.sched = __args->sched;
+	args.credits = 1;
+
 	args.in_sync.count = __args->in_sync.count;
 	args.in_sync.s = __args->in_sync.s;