drm/msm/gpu: Respect PM QoS constraints

Re-work the boost and idle clamping to use PM QoS requests instead, so
they get aggreggated with other requests (such as cooling device).

This does have the minor side-effect that devfreq sysfs min_freq/
max_freq files now reflect the boost and idle clamping, as they show
(despite what they are documented to show) the aggregated min/max freq.
Fixing that in devfreq does not look straightforward after considering
that OPPs can be dynamically added/removed.  However writes to the
sysfs files still behave as expected.

v2: Use 64b math to avoid potential 32b overflow
Signed-off-by: Rob Clark <robdclark@chromium.org>
Link: https://lore.kernel.org/r/20211120200103.1051459-3-robdclark@gmail.comSigned-off-by: Rob Clark <robdclark@chromium.org>

drivers/gpu/drm/msm/msm_gpu.h

@@ -87,6 +87,21 @@ struct msm_gpu_devfreq {
 	/** devfreq: devfreq instance */
 	struct devfreq *devfreq;
 
+	/**
+	 * idle_constraint:
+	 *
+	 * A PM QoS constraint to limit max freq while the GPU is idle.
+	 */
+	struct dev_pm_qos_request idle_freq;
+
+	/**
+	 * boost_constraint:
+	 *
+	 * A PM QoS constraint to boost min freq for a period of time
+	 * until the boost expires.
+	 */
+	struct dev_pm_qos_request boost_freq;
+
 	/**
 	 * busy_cycles:
 	 *
@@ -103,22 +118,19 @@ struct msm_gpu_devfreq {
 	ktime_t idle_time;
 
 	/**
-	 * idle_freq:
+	 * idle_work:
 	 *
-	 * Shadow frequency used while the GPU is idle.  From the PoV of
-	 * the devfreq governor, we are continuing to sample busyness and
-	 * adjust frequency while the GPU is idle, but we use this shadow
-	 * value as the GPU is actually clamped to minimum frequency while
-	 * it is inactive.
+	 * Used to delay clamping to idle freq on active->idle transition.
 	 */
-	unsigned long idle_freq;
+	struct msm_hrtimer_work idle_work;
 
 	/**
-	 * idle_work:
+	 * boost_work:
 	 *
-	 * Used to delay clamping to idle freq on active->idle transition.
+	 * Used to reset the boost_constraint after the boost period has
+	 * elapsed
 	 */
-	struct msm_hrtimer_work idle_work;
+	struct msm_hrtimer_work boost_work;
 };
 
 struct msm_gpu {
@@ -522,6 +534,7 @@ void msm_devfreq_init(struct msm_gpu *gpu);
 void msm_devfreq_cleanup(struct msm_gpu *gpu);
 void msm_devfreq_resume(struct msm_gpu *gpu);
 void msm_devfreq_suspend(struct msm_gpu *gpu);
+void msm_devfreq_boost(struct msm_gpu *gpu, unsigned factor);
 void msm_devfreq_active(struct msm_gpu *gpu);
 void msm_devfreq_idle(struct msm_gpu *gpu);
 

drivers/gpu/drm/msm/msm_gpu_devfreq.c

@@ -9,6 +9,7 @@
 
 #include <linux/devfreq.h>
 #include <linux/devfreq_cooling.h>
+#include <linux/units.h>
 
 /*
  * Power Management:
@@ -25,17 +26,6 @@ static int msm_devfreq_target(struct device *dev, unsigned long *freq,
 	 * to something that actually is in the opp table:
 	 */
 	opp = devfreq_recommended_opp(dev, freq, flags);
-
-	/*
-	 * If the GPU is idle, devfreq is not aware, so just ignore
-	 * it's requests
-	 */
-	if (gpu->devfreq.idle_freq) {
-		gpu->devfreq.idle_freq = *freq;
-		dev_pm_opp_put(opp);
-		return 0;
-	}
-
 	if (IS_ERR(opp))
 		return PTR_ERR(opp);
 
@@ -53,9 +43,6 @@ static int msm_devfreq_target(struct device *dev, unsigned long *freq,
 
 static unsigned long get_freq(struct msm_gpu *gpu)
 {
-	if (gpu->devfreq.idle_freq)
-		return gpu->devfreq.idle_freq;
-
 	if (gpu->funcs->gpu_get_freq)
 		return gpu->funcs->gpu_get_freq(gpu);
 
@@ -93,6 +80,7 @@ static struct devfreq_dev_profile msm_devfreq_profile = {
 	.get_cur_freq = msm_devfreq_get_cur_freq,
 };
 
+static void msm_devfreq_boost_work(struct kthread_work *work);
 static void msm_devfreq_idle_work(struct kthread_work *work);
 
 void msm_devfreq_init(struct msm_gpu *gpu)
@@ -103,6 +91,12 @@ void msm_devfreq_init(struct msm_gpu *gpu)
 	if (!gpu->funcs->gpu_busy)
 		return;
 
+	dev_pm_qos_add_request(&gpu->pdev->dev, &df->idle_freq,
+			       DEV_PM_QOS_MAX_FREQUENCY,
+			       PM_QOS_MAX_FREQUENCY_DEFAULT_VALUE);
+	dev_pm_qos_add_request(&gpu->pdev->dev, &df->boost_freq,
+			       DEV_PM_QOS_MIN_FREQUENCY, 0);
+
 	msm_devfreq_profile.initial_freq = gpu->fast_rate;
 
 	/*
@@ -133,13 +127,19 @@ void msm_devfreq_init(struct msm_gpu *gpu)
 		gpu->cooling = NULL;
 	}
 
+	msm_hrtimer_work_init(&df->boost_work, gpu->worker, msm_devfreq_boost_work,
+			      CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 	msm_hrtimer_work_init(&df->idle_work, gpu->worker, msm_devfreq_idle_work,
 			      CLOCK_MONOTONIC, HRTIMER_MODE_REL);
 }
 
 void msm_devfreq_cleanup(struct msm_gpu *gpu)
 {
+	struct msm_gpu_devfreq *df = &gpu->devfreq;
+
 	devfreq_cooling_unregister(gpu->cooling);
+	dev_pm_qos_remove_request(&df->boost_freq);
+	dev_pm_qos_remove_request(&df->idle_freq);
 }
 
 void msm_devfreq_resume(struct msm_gpu *gpu)
@@ -155,12 +155,40 @@ void msm_devfreq_suspend(struct msm_gpu *gpu)
 	devfreq_suspend_device(gpu->devfreq.devfreq);
 }
 
+static void msm_devfreq_boost_work(struct kthread_work *work)
+{
+	struct msm_gpu_devfreq *df = container_of(work,
+			struct msm_gpu_devfreq, boost_work.work);
+
+	dev_pm_qos_update_request(&df->boost_freq, 0);
+}
+
+void msm_devfreq_boost(struct msm_gpu *gpu, unsigned factor)
+{
+	struct msm_gpu_devfreq *df = &gpu->devfreq;
+	uint64_t freq;
+
+	freq = get_freq(gpu);
+	freq *= factor;
+
+	/*
+	 * A nice little trap is that PM QoS operates in terms of KHz,
+	 * while devfreq operates in terms of Hz:
+	 */
+	do_div(freq, HZ_PER_KHZ);
+
+	dev_pm_qos_update_request(&df->boost_freq, freq);
+
+	msm_hrtimer_queue_work(&df->boost_work,
+			       ms_to_ktime(msm_devfreq_profile.polling_ms),
+			       HRTIMER_MODE_REL);
+}
+
 void msm_devfreq_active(struct msm_gpu *gpu)
 {
 	struct msm_gpu_devfreq *df = &gpu->devfreq;
 	struct devfreq_dev_status status;
 	unsigned int idle_time;
-	unsigned long target_freq = df->idle_freq;
 
 	if (!df->devfreq)
 		return;
@@ -170,12 +198,6 @@ void msm_devfreq_active(struct msm_gpu *gpu)
 	 */
 	hrtimer_cancel(&df->idle_work.timer);
 
-	/*
-	 * Hold devfreq lock to synchronize with get_dev_status()/
-	 * target() callbacks
-	 */
-	mutex_lock(&df->devfreq->lock);
-
 	idle_time = ktime_to_ms(ktime_sub(ktime_get(), df->idle_time));
 
 	/*
@@ -184,20 +206,17 @@ void msm_devfreq_active(struct msm_gpu *gpu)
 	 * the governor to ramp up the freq.. so give some boost
 	 */
 	if (idle_time > msm_devfreq_profile.polling_ms) {
-		target_freq *= 2;
+		msm_devfreq_boost(gpu, 2);
 	}
 
-	df->idle_freq = 0;
-
-	msm_devfreq_target(&gpu->pdev->dev, &target_freq, 0);
+	dev_pm_qos_update_request(&df->idle_freq,
+				  PM_QOS_MAX_FREQUENCY_DEFAULT_VALUE);
 
 	/*
 	 * Reset the polling interval so we aren't inconsistent
 	 * about freq vs busy/total cycles
 	 */
 	msm_devfreq_get_dev_status(&gpu->pdev->dev, &status);
-
-	mutex_unlock(&df->devfreq->lock);
 }
 
 
@@ -206,23 +225,11 @@ static void msm_devfreq_idle_work(struct kthread_work *work)
 	struct msm_gpu_devfreq *df = container_of(work,
 			struct msm_gpu_devfreq, idle_work.work);
 	struct msm_gpu *gpu = container_of(df, struct msm_gpu, devfreq);
-	unsigned long idle_freq, target_freq = 0;
-
-	/*
-	 * Hold devfreq lock to synchronize with get_dev_status()/
-	 * target() callbacks
-	 */
-	mutex_lock(&df->devfreq->lock);
-
-	idle_freq = get_freq(gpu);
-
-	if (gpu->clamp_to_idle)
-		msm_devfreq_target(&gpu->pdev->dev, &target_freq, 0);
 
 	df->idle_time = ktime_get();
-	df->idle_freq = idle_freq;
 
-	mutex_unlock(&df->devfreq->lock);
+	if (gpu->clamp_to_idle)
+		dev_pm_qos_update_request(&df->idle_freq, 0);
 }
 
 void msm_devfreq_idle(struct msm_gpu *gpu)