[2/2] drm/msm/gpu: Respect PM QoS constraints

Commit Message

Rob Clark Nov. 19, 2021, 10:51 p.m. UTC

From: Rob Clark <[e-mail]>

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.

Signed-off-by: Rob Clark <[e-mail]>
---
 drivers/gpu/drm/msm/msm_gpu.h         | 33 +++++++----
 drivers/gpu/drm/msm/msm_gpu_devfreq.c | 82 ++++++++++++++-------------
 2 files changed, 66 insertions(+), 49 deletions(-)

Comments

Doug Anderson Nov. 20, 2021, 12:21 a.m. UTC | #1

Hi,

On Fri, Nov 19, 2021 at 2:47 PM Rob Clark <[e-mail]> wrote:
>
> +void msm_devfreq_boost(struct msm_gpu *gpu, unsigned factor)
> +{
> +       struct msm_gpu_devfreq *df = &gpu->devfreq;
> +       unsigned long freq;
> +
> +       freq = get_freq(gpu);
> +       freq *= factor;
> +       freq /= HZ_PER_KHZ;

Should it do the divide first? I don't know for sure, but it feels
like GPU frequency could conceivably be near-ish the u32 overflow? (~4
GHz). Better to be safe and do the / 1000 first?


> @@ -201,26 +217,14 @@ 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;
>
>         if (!df->devfreq)
>                 return;

Why does the msm_devfreq_idle_work() need a check for "!df->devfreq"
but the boost work doesn't? Maybe you don't need it anymore now that
you're not reaching into the mutex? ...or maybe the boost work does
need it?

...and if "df->devfreq" is NULL then doesn't it mean that
msm_hrtimer_work_init() was never called? That seems bad...


-Doug

Rob Clark Nov. 20, 2021, 6:08 p.m. UTC | #2

On Fri, Nov 19, 2021 at 4:21 PM Doug Anderson <[e-mail]> wrote:
>
> Hi,
>
> On Fri, Nov 19, 2021 at 2:47 PM Rob Clark <[e-mail]> wrote:
> >
> > +void msm_devfreq_boost(struct msm_gpu *gpu, unsigned factor)
> > +{
> > +       struct msm_gpu_devfreq *df = &gpu->devfreq;
> > +       unsigned long freq;
> > +
> > +       freq = get_freq(gpu);
> > +       freq *= factor;
> > +       freq /= HZ_PER_KHZ;
>
> Should it do the divide first? I don't know for sure, but it feels
> like GPU frequency could conceivably be near-ish the u32 overflow? (~4
> GHz). Better to be safe and do the / 1000 first?
>

It looks like on 8998 we have some GPU OPPs that are not integer # of
KHz.. although that would not change the integer math result unless
factor > 10.

We are a bit aways for 32b overflow (highest freq for current things
is 825MHz, but I guess we could see things closer to 1GHz in the
future.. generally GPUs aren't clocked nearly as high as CPUs.. slow
but wide, and all that).. but maybe this should just be 64b math
instead to be safe?

>
> > @@ -201,26 +217,14 @@ 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;
> >
> >         if (!df->devfreq)
> >                 return;
>
> Why does the msm_devfreq_idle_work() need a check for "!df->devfreq"
> but the boost work doesn't? Maybe you don't need it anymore now that
> you're not reaching into the mutex? ...or maybe the boost work does
> need it?
>
> ...and if "df->devfreq" is NULL then doesn't it mean that
> msm_hrtimer_work_init() was never called? That seems bad...

Looks like 658f4c829688 ("drm/msm/devfreq: Add 1ms delay before
clamping freq") was badly rebased on top of efb8a170a367 ("drm/msm:
Fix devfreq NULL pointer dereference on a3xx").. I'll send a separate
patch to fix that

BR,
-R

Patch

diff --git a/drivers/gpu/drm/msm/msm_gpu.h b/drivers/gpu/drm/msm/msm_gpu.h
index 59cdd00b69d0..96d8d37dd5b7 100644
--- a/drivers/gpu/drm/msm/msm_gpu.h
+++ b/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 {
@@ -498,6 +510,7 @@  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);
 
diff --git a/drivers/gpu/drm/msm/msm_gpu_devfreq.c b/drivers/gpu/drm/msm/msm_gpu_devfreq.c
index 7285041c737e..ff668e431cee 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/units.h>
 
 /*
  * Power Management:
@@ -22,15 +23,6 @@  static int msm_devfreq_target(struct device *dev, unsigned long *freq,
 
 	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;
-		return 0;
-	}
-
 	if (IS_ERR(opp))
 		return PTR_ERR(opp);
 
@@ -48,9 +40,6 @@  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);
 
@@ -88,6 +77,7 @@  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)
@@ -98,6 +88,12 @@  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;
 
 	/*
@@ -128,13 +124,19 @@  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)
@@ -150,12 +152,35 @@  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;
+	unsigned long freq;
+
+	freq = get_freq(gpu);
+	freq *= factor;
+	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;
@@ -165,12 +190,6 @@  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));
 
 	/*
@@ -179,20 +198,17 @@  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);
 }
 
 
@@ -201,26 +217,14 @@  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;
 
 	if (!df->devfreq)
 		return;
 
-	/*
-	 * 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)