[RFC PATCH] drm/msm: tune the devfreq to take into account the load history

[Dmitry Baryshkov]

Partially restore the handling of the GPU load history. Accumulate the
busy_time and and total_time measured in active state during the polling
period. This results in slightly smoother picture of the GPU frequencies
(measured on the a530/msm8996, using kmscube in different resolutions).

A call to msm_devfreq_get_dev_status() from msm_devfreq_active() was
removed in 69f06a5d854f ("drm/msm: remove explicit devfreq status
reset"), because dev_pm_qos_update_request() triggered that internally.
As the commit fadcc3ab1302 ("drm/msm/gpu: Bypass PM QoS constraint for
idle clamp") removed the calls to dev_pm_qos_update_request(), this
removal was also reverted.

The code doesn't take the frequency into account while accumulating the
data to keep the code simple for the RFC.

Cc: Chia-I Wu <olvaffe at gmail.com>
Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov at linaro.org>
---
 drivers/gpu/drm/msm/msm_gpu.h         |  3 +++
 drivers/gpu/drm/msm/msm_gpu_devfreq.c | 30 ++++++++++++++++++++++++++-
 2 files changed, 32 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/msm/msm_gpu.h b/drivers/gpu/drm/msm/msm_gpu.h
index fc1c0d8611a8..9d1783375dcc 100644
--- a/drivers/gpu/drm/msm/msm_gpu.h
+++ b/drivers/gpu/drm/msm/msm_gpu.h
@@ -156,6 +156,9 @@ struct msm_gpu_devfreq {
 
 	/** suspended: tracks if we're suspended */
 	bool suspended;
+
+	/* stats for the current period */
+	struct devfreq_dev_status status;
 };
 
 struct msm_gpu {
diff --git a/drivers/gpu/drm/msm/msm_gpu_devfreq.c b/drivers/gpu/drm/msm/msm_gpu_devfreq.c
index e27dbf12b5e8..92cb022c8628 100644
--- a/drivers/gpu/drm/msm/msm_gpu_devfreq.c
+++ b/drivers/gpu/drm/msm/msm_gpu_devfreq.c
@@ -74,7 +74,7 @@ static unsigned long get_freq(struct msm_gpu *gpu)
 	return clk_get_rate(gpu->core_clk);
 }
 
-static int msm_devfreq_get_dev_status(struct device *dev,
+static int msm_devfreq_get_dev_status_int(struct device *dev,
 		struct devfreq_dev_status *status)
 {
 	struct msm_gpu *gpu = dev_to_gpu(dev);
@@ -112,6 +112,22 @@ static int msm_devfreq_get_dev_status(struct device *dev,
 	return 0;
 }
 
+static int msm_devfreq_get_dev_status(struct device *dev,
+		struct devfreq_dev_status *status)
+{
+	struct msm_gpu *gpu = dev_to_gpu(dev);
+	struct msm_gpu_devfreq *df = &gpu->devfreq;
+
+	msm_devfreq_get_dev_status_int(&gpu->pdev->dev, status);
+	status->busy_time += df->status.busy_time;
+	status->total_time += df->status.total_time;
+
+	df->status.busy_time = 0;
+	df->status.total_time = 0;
+
+	return 0;
+}
+
 static int msm_devfreq_get_cur_freq(struct device *dev, unsigned long *freq)
 {
 	*freq = get_freq(dev_to_gpu(dev));
@@ -290,6 +306,7 @@ void msm_devfreq_active(struct msm_gpu *gpu)
 	struct msm_gpu_devfreq *df = &gpu->devfreq;
 	unsigned int idle_time;
 	unsigned long target_freq;
+	struct devfreq_dev_status status;
 
 	if (!has_devfreq(gpu))
 		return;
@@ -319,6 +336,12 @@ void msm_devfreq_active(struct msm_gpu *gpu)
 	if (target_freq)
 		msm_devfreq_target(&gpu->pdev->dev, &target_freq, 0);
 
+	/*
+	 * Reset the polling interval so we aren't inconsistent
+	 * about freq vs busy/total cycles
+	 */
+	msm_devfreq_get_dev_status_int(&gpu->pdev->dev, &status);
+
 	mutex_unlock(&df->devfreq->lock);
 
 	/*
@@ -339,6 +362,7 @@ static void msm_devfreq_idle_work(struct kthread_work *work)
 	struct msm_gpu *gpu = container_of(df, struct msm_gpu, devfreq);
 	struct msm_drm_private *priv = gpu->dev->dev_private;
 	unsigned long idle_freq, target_freq = 0;
+	struct devfreq_dev_status status;
 
 	/*
 	 * Hold devfreq lock to synchronize with get_dev_status()/
@@ -346,6 +370,10 @@ static void msm_devfreq_idle_work(struct kthread_work *work)
 	 */
 	mutex_lock(&df->devfreq->lock);
 
+	msm_devfreq_get_dev_status_int(&gpu->pdev->dev, &status);
+	df->status.busy_time += status.busy_time;
+	df->status.total_time += status.total_time;
+
 	idle_freq = get_freq(gpu);
 
 	if (priv->gpu_clamp_to_idle)
-- 
2.30.2