drm/msm: return the average load over the polling period

simple_ondemand interacts poorly with clamp_to_idle.  It only looks at
the load since the last get_dev_status call, while it should really look
at the load over polling_ms.  When clamp_to_idle true, it almost always
picks the lowest frequency on active because the gpu is idle between
msm_devfreq_idle/msm_devfreq_active.

This logic could potentially be moved into devfreq core.

Fixes: 7c0ffcd40b16 ("drm/msm/gpu: Respect PM QoS constraints")
Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Cc: Rob Clark <robdclark@chromium.org>
Link: https://lore.kernel.org/r/20220416003314.59211-3-olvaffe@gmail.com
Signed-off-by: Rob Clark <robdclark@chromium.org>

2 files changed, 62 insertions, 1 deletions

diff --git a/drivers/gpu/drm/msm/msm_gpu.h b/drivers/gpu/drm/msm/msm_gpu.h
index 7418f99e77d9..6def00883046 100644
--- a/drivers/gpu/drm/msm/msm_gpu.h
+++ b/drivers/gpu/drm/msm/msm_gpu.h
@@ -9,6 +9,7 @@
 
 #include <linux/adreno-smmu-priv.h>
 #include <linux/clk.h>
+#include <linux/devfreq.h>
 #include <linux/interconnect.h>
 #include <linux/pm_opp.h>
 #include <linux/regulator/consumer.h>
@@ -118,6 +119,8 @@ struct msm_gpu_devfreq {
 	/** idle_time: Time of last transition to idle: */
 	ktime_t idle_time;
 
+	struct devfreq_dev_status average_status;
+
 	/**
 	 * idle_work:
 	 *
diff --git a/drivers/gpu/drm/msm/msm_gpu_devfreq.c b/drivers/gpu/drm/msm/msm_gpu_devfreq.c
index f531015107c3..d2539ca78c29 100644
--- a/drivers/gpu/drm/msm/msm_gpu_devfreq.c
+++ b/drivers/gpu/drm/msm/msm_gpu_devfreq.c
@@ -9,6 +9,7 @@
 
 #include <linux/devfreq.h>
 #include <linux/devfreq_cooling.h>
+#include <linux/math64.h>
 #include <linux/units.h>
 
 /*
@@ -75,12 +76,69 @@ static void get_raw_dev_status(struct msm_gpu *gpu,
 	status->busy_time = busy_time;
 }
 
+static void update_average_dev_status(struct msm_gpu *gpu,
+		const struct devfreq_dev_status *raw)
+{
+	struct msm_gpu_devfreq *df = &gpu->devfreq;
+	const u32 polling_ms = df->devfreq->profile->polling_ms;
+	const u32 max_history_ms = polling_ms * 11 / 10;
+	struct devfreq_dev_status *avg = &df->average_status;
+	u64 avg_freq;
+
+	/* simple_ondemand governor interacts poorly with gpu->clamp_to_idle.
+	 * When we enforce the constraint on idle, it calls get_dev_status
+	 * which would normally reset the stats.  When we remove the
+	 * constraint on active, it calls get_dev_status again where busy_time
+	 * would be 0.
+	 *
+	 * To remedy this, we always return the average load over the past
+	 * polling_ms.
+	 */
+
+	/* raw is longer than polling_ms or avg has no history */
+	if (div_u64(raw->total_time, USEC_PER_MSEC) >= polling_ms ||
+	    !avg->total_time) {
+		*avg = *raw;
+		return;
+	}
+
+	/* Truncate the oldest history first.
+	 *
+	 * Because we keep the history with a single devfreq_dev_status,
+	 * rather than a list of devfreq_dev_status, we have to assume freq
+	 * and load are the same over avg->total_time.  We can scale down
+	 * avg->busy_time and avg->total_time by the same factor to drop
+	 * history.
+	 */
+	if (div_u64(avg->total_time + raw->total_time, USEC_PER_MSEC) >=
+			max_history_ms) {
+		const u32 new_total_time = polling_ms * USEC_PER_MSEC -
+			raw->total_time;
+		avg->busy_time = div_u64(
+				mul_u32_u32(avg->busy_time, new_total_time),
+				avg->total_time);
+		avg->total_time = new_total_time;
+	}
+
+	/* compute the average freq over avg->total_time + raw->total_time */
+	avg_freq = mul_u32_u32(avg->current_frequency, avg->total_time);
+	avg_freq += mul_u32_u32(raw->current_frequency, raw->total_time);
+	do_div(avg_freq, avg->total_time + raw->total_time);
+
+	avg->current_frequency = avg_freq;
+	avg->busy_time += raw->busy_time;
+	avg->total_time += raw->total_time;
+}
+
 static int msm_devfreq_get_dev_status(struct device *dev,
 		struct devfreq_dev_status *status)
 {
 	struct msm_gpu *gpu = dev_to_gpu(dev);
+	struct devfreq_dev_status raw;
 
-	get_raw_dev_status(gpu, status);
+	get_raw_dev_status(gpu, &raw);
+	update_average_dev_status(gpu, &raw);
+	*status = gpu->devfreq.average_status;
 
 	return 0;
 }