drm/lima: enable runtime pm

Enable runtime pm by default so GPU suspend when idle
for 200ms. This value can be changed by
autosuspend_delay_ms in device's power sysfs dir.

On Allwinner H3 lima_device_resume takes ~40us and
lima_device_suspend takes ~20us.

Tested-by: Bhushan Shah <bshah@kde.org>
Reviewed-by: Vasily Khoruzhick <anarsoul@gmail.com>
Signed-off-by: Qiang Yu <yuq825@gmail.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200421133551.31481-11-yuq825@gmail.com

drivers/gpu/drm/lima/lima_drv.c

@@ -404,6 +404,12 @@ static int lima_pdev_probe(struct platform_device *pdev)
 		goto err_out2;
 	}
 
+	pm_runtime_set_active(ldev->dev);
+	pm_runtime_mark_last_busy(ldev->dev);
+	pm_runtime_set_autosuspend_delay(ldev->dev, 200);
+	pm_runtime_use_autosuspend(ldev->dev);
+	pm_runtime_enable(ldev->dev);
+
 	/*
 	 * Register the DRM device with the core and the connectors with
 	 * sysfs.
@@ -412,17 +418,16 @@ static int lima_pdev_probe(struct platform_device *pdev)
 	if (err < 0)
 		goto err_out3;
 
-	platform_set_drvdata(pdev, ldev);
-
 	if (sysfs_create_bin_file(&ldev->dev->kobj, &lima_error_state_attr))
 		dev_warn(ldev->dev, "fail to create error state sysfs\n");
 
 	return 0;
 
 err_out3:
-	lima_device_fini(ldev);
-err_out2:
+	pm_runtime_disable(ldev->dev);
 	lima_devfreq_fini(ldev);
+err_out2:
+	lima_device_fini(ldev);
 err_out1:
 	drm_dev_put(ddev);
 err_out0:
@@ -436,10 +441,16 @@ static int lima_pdev_remove(struct platform_device *pdev)
 	struct drm_device *ddev = ldev->ddev;
 
 	sysfs_remove_bin_file(&ldev->dev->kobj, &lima_error_state_attr);
-	platform_set_drvdata(pdev, NULL);
+
 	drm_dev_unregister(ddev);
+
+	/* stop autosuspend to make sure device is in active state */
+	pm_runtime_set_autosuspend_delay(ldev->dev, -1);
+	pm_runtime_disable(ldev->dev);
+
 	lima_devfreq_fini(ldev);
 	lima_device_fini(ldev);
+
 	drm_dev_put(ddev);
 	lima_sched_slab_fini();
 	return 0;

drivers/gpu/drm/lima/lima_sched.c

@@ -4,6 +4,7 @@
 #include <linux/kthread.h>
 #include <linux/slab.h>
 #include <linux/vmalloc.h>
+#include <linux/pm_runtime.h>
 
 #include "lima_devfreq.h"
 #include "lima_drv.h"
@@ -194,13 +195,36 @@ static struct dma_fence *lima_sched_dependency(struct drm_sched_job *job,
 	return NULL;
 }
 
+static int lima_pm_busy(struct lima_device *ldev)
+{
+	int ret;
+
+	/* resume GPU if it has been suspended by runtime PM */
+	ret = pm_runtime_get_sync(ldev->dev);
+	if (ret < 0)
+		return ret;
+
+	lima_devfreq_record_busy(&ldev->devfreq);
+	return 0;
+}
+
+static void lima_pm_idle(struct lima_device *ldev)
+{
+	lima_devfreq_record_idle(&ldev->devfreq);
+
+	/* GPU can do auto runtime suspend */
+	pm_runtime_mark_last_busy(ldev->dev);
+	pm_runtime_put_autosuspend(ldev->dev);
+}
+
 static struct dma_fence *lima_sched_run_job(struct drm_sched_job *job)
 {
 	struct lima_sched_task *task = to_lima_task(job);
 	struct lima_sched_pipe *pipe = to_lima_pipe(job->sched);
+	struct lima_device *ldev = pipe->ldev;
 	struct lima_fence *fence;
 	struct dma_fence *ret;
-	int i;
+	int i, err;
 
 	/* after GPU reset */
 	if (job->s_fence->finished.error < 0)
@@ -209,15 +233,20 @@ static struct dma_fence *lima_sched_run_job(struct drm_sched_job *job)
 	fence = lima_fence_create(pipe);
 	if (!fence)
 		return NULL;
+
+	err = lima_pm_busy(ldev);
+	if (err < 0) {
+		dma_fence_put(&fence->base);
+		return NULL;
+	}
+
 	task->fence = &fence->base;
 
 	/* for caller usage of the fence, otherwise irq handler
 	 * may consume the fence before caller use it
 	 */
 	ret = dma_fence_get(task->fence);
 
-	lima_devfreq_record_busy(&pipe->ldev->devfreq);
-
 	pipe->current_task = task;
 
 	/* this is needed for MMU to work correctly, otherwise GP/PP
@@ -388,6 +417,7 @@ static void lima_sched_timedout_job(struct drm_sched_job *job)
 {
 	struct lima_sched_pipe *pipe = to_lima_pipe(job->sched);
 	struct lima_sched_task *task = to_lima_task(job);
+	struct lima_device *ldev = pipe->ldev;
 
 	if (!pipe->error)
 		DRM_ERROR("lima job timeout\n");
@@ -413,7 +443,7 @@ static void lima_sched_timedout_job(struct drm_sched_job *job)
 	pipe->current_vm = NULL;
 	pipe->current_task = NULL;
 
-	lima_devfreq_record_idle(&pipe->ldev->devfreq);
+	lima_pm_idle(ldev);
 
 	drm_sched_resubmit_jobs(&pipe->base);
 	drm_sched_start(&pipe->base, true);
@@ -485,6 +515,7 @@ void lima_sched_pipe_fini(struct lima_sched_pipe *pipe)
 void lima_sched_pipe_task_done(struct lima_sched_pipe *pipe)
 {
 	struct lima_sched_task *task = pipe->current_task;
+	struct lima_device *ldev = pipe->ldev;
 
 	if (pipe->error) {
 		if (task && task->recoverable)
@@ -495,6 +526,6 @@ void lima_sched_pipe_task_done(struct lima_sched_pipe *pipe)
 		pipe->task_fini(pipe);
 		dma_fence_signal(task->fence);
 
-		lima_devfreq_record_idle(&pipe->ldev->devfreq);
+		lima_pm_idle(ldev);
 	}
 }