linux-pm.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
* [RFC] PM / devfreq: add busfreq governor
@ 2020-03-25 15:22 ` Martin Kepplinger
  2020-03-25 22:34   ` Chanwoo Choi
  2020-03-25 22:35   ` Leonard Crestez
  0 siblings, 2 replies; 4+ messages in thread
From: Martin Kepplinger @ 2020-03-25 15:22 UTC (permalink / raw)
  To: linux-imx, leonard.crestez, myungjoo.ham, kyungmin.park, cw00.choi
  Cc: linux-pm, Martin Kepplinger

A devfreq governor to be used by drivers to request ("get") a frequency
setting they need. Currently only "HIGH" (performance) is available.

ATTENTION: This is a first draft to serve merely as a basis for discussions!
ONLY USE FOR TESTING!
---

I wanted to get early feedback on an idea that AFAIK is not yet available
in the kernel (but something similar via "busfreq" in NXP's tree):

Let drivers request high (dram) frequencies at runtime if they know they need
them. In our case the display stack on imx8mq would be the first user,
looking like so: #include <linux/devfreq.h> and:

	--- a/drivers/gpu/drm/bridge/nwl-dsi.c
	+++ b/drivers/gpu/drm/bridge/nwl-dsi.c
	@@ -1116,6 +1117,9 @@ static int imx8mq_dsi_poweron(struct nwl_dsi *dsi)
			ret = reset_control_deassert(dsi->rstc);
	 
	+       devfreq_busfreq_request(DEVFREQ_BUSFREQ_HIGH);
	+
		return ret;
	@@ -1125,6 +1129,10 @@ static int imx8mq_dsi_poweroff(struct nwl_dsi *dsi)
		if (dsi->rstc)
			ret = reset_control_assert(dsi->rstc);
	+
	+       devfreq_busfreq_release(DEVFREQ_BUSFREQ_HIGH);
	+
		return ret;


Could be called in pm_runtime() calls too.

_If_ the idea of such or a similar governor is viable, there are at least
some problems with this implemenation still:

* The governor saves its data as a global variable (which I don't yet like)
  but the driver (user) API becomes minimal.

* the name: In order to add to the devfreq/busfreq confusion, I named the
  devfreq governor "busfreq" -.-
  I just wanted to grab NXPs' attention because they do something similar
  in their tree.

* ATM we switch between "performace" and "powersave, which is
  not bad at all, but still limited in case a driver would really only
  need a medium frequecy for a device to work. doable?

* I doubt locking is done correctly and the code is overall of bad quality
  still. Again, It's a first idea.

I'm glad about any thought or feedback,

thanks,
                                   martin




 drivers/devfreq/Kconfig            |   5 +
 drivers/devfreq/Makefile           |   1 +
 drivers/devfreq/governor_busfreq.c | 172 +++++++++++++++++++++++++++++
 include/linux/devfreq.h            |  22 ++++
 4 files changed, 200 insertions(+)
 create mode 100644 drivers/devfreq/governor_busfreq.c

diff --git a/drivers/devfreq/Kconfig b/drivers/devfreq/Kconfig
index fc2ea336ef4b..3575cbdcd29e 100644
--- a/drivers/devfreq/Kconfig
+++ b/drivers/devfreq/Kconfig
@@ -74,6 +74,11 @@ config DEVFREQ_GOV_PASSIVE
 	  through sysfs entries. The passive governor recommends that
 	  devfreq device uses the OPP table to get the frequency/voltage.
 
+config DEVFREQ_GOV_BUSFREQ
+	tristate "Busfreq"
+	help
+	  Sets the frequency that compatible drivers request it to set.
+
 comment "DEVFREQ Drivers"
 
 config ARM_EXYNOS_BUS_DEVFREQ
diff --git a/drivers/devfreq/Makefile b/drivers/devfreq/Makefile
index 61d0edee16f7..657b0b6d92c4 100644
--- a/drivers/devfreq/Makefile
+++ b/drivers/devfreq/Makefile
@@ -6,6 +6,7 @@ obj-$(CONFIG_DEVFREQ_GOV_PERFORMANCE)	+= governor_performance.o
 obj-$(CONFIG_DEVFREQ_GOV_POWERSAVE)	+= governor_powersave.o
 obj-$(CONFIG_DEVFREQ_GOV_USERSPACE)	+= governor_userspace.o
 obj-$(CONFIG_DEVFREQ_GOV_PASSIVE)	+= governor_passive.o
+obj-$(CONFIG_DEVFREQ_GOV_BUSFREQ)	+= governor_busfreq.o
 
 # DEVFREQ Drivers
 obj-$(CONFIG_ARM_EXYNOS_BUS_DEVFREQ)	+= exynos-bus.o
diff --git a/drivers/devfreq/governor_busfreq.c b/drivers/devfreq/governor_busfreq.c
new file mode 100644
index 000000000000..e12d64e9a09b
--- /dev/null
+++ b/drivers/devfreq/governor_busfreq.c
@@ -0,0 +1,172 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ *  linux/drivers/devfreq/governor_busfreq.c
+ *
+ *  Copyright (C) 2020 Purism SPC
+ */
+
+#include <linux/slab.h>
+#include <linux/device.h>
+#include <linux/devfreq.h>
+#include <linux/pm.h>
+#include <linux/mutex.h>
+#include <linux/module.h>
+#include "governor.h"
+
+struct busfreq_data {
+	unsigned long user_frequency;
+	bool valid;
+	struct devfreq *df;
+	unsigned int high_count;
+};
+
+static struct busfreq_data *bfdata = NULL;
+
+static int devfreq_busfreq_func(struct devfreq *df, unsigned long *freq)
+{
+	if (bfdata->valid)
+		*freq = bfdata->user_frequency;
+	else
+		*freq = df->previous_freq; /* No user freq specified yet */
+
+	return 0;
+}
+
+int devfreq_busfreq_request(unsigned long freq)
+{
+	struct devfreq *devfreq;
+	int err = 0;
+
+	if (!bfdata) {
+		pr_info("%s: governor not available\n", __func__);
+		return -ENODEV;
+	}
+
+	if (freq != DEVFREQ_BUSFREQ_HIGH) {
+		pr_err("%s: undefined frequency\n", __func__);
+		return -EINVAL;
+	}
+
+	if (freq == DEVFREQ_BUSFREQ_HIGH)
+		bfdata->high_count++;
+
+	devfreq = bfdata->df;
+
+	mutex_lock(&devfreq->lock);
+
+	bfdata->user_frequency = freq;
+	bfdata->valid = true;
+	err = update_devfreq(devfreq); /* calls our get_target_freq */
+	if (err)
+		dev_err(&devfreq->dev, "update_devfreq failed: %d\n", err);
+
+	mutex_unlock(&devfreq->lock);
+
+	return err;
+}
+EXPORT_SYMBOL(devfreq_busfreq_request);
+
+void devfreq_busfreq_release(unsigned long freq)
+{
+	struct devfreq *devfreq;
+	int err = 0;
+
+	if (!bfdata) {
+		pr_info("%s: governor not available\n", __func__);
+		return;
+	}
+
+	if (freq != DEVFREQ_BUSFREQ_HIGH) {
+		pr_err("%s: undefined frequency\n", __func__);
+		return;
+	}
+
+	if (freq == DEVFREQ_BUSFREQ_HIGH && bfdata->high_count > 0)
+		bfdata->high_count--;
+
+	if (bfdata->high_count)
+		return;
+
+	devfreq = bfdata->df;
+
+	mutex_lock(&devfreq->lock);
+
+	bfdata->user_frequency = DEVFREQ_BUSFREQ_LOW;
+	bfdata->valid = true;
+	err = update_devfreq(devfreq); /* calls our get_target_freq */
+	if (err)
+		dev_err(&devfreq->dev, "update_devfreq failed: %d\n", err);
+
+	mutex_unlock(&devfreq->lock);
+}
+EXPORT_SYMBOL(devfreq_busfreq_release);
+
+static int busfreq_init(struct devfreq *devfreq)
+{
+	int err = 0;
+
+	bfdata = kzalloc(sizeof(struct busfreq_data),
+		      GFP_KERNEL);
+	if (!bfdata) {
+		err = -ENOMEM;
+		goto out;
+	}
+	bfdata->valid = false;
+	devfreq->data = bfdata;
+	bfdata->df = devfreq;
+
+out:
+	return err;
+}
+
+static void busfreq_exit(struct devfreq *devfreq)
+{
+	if (bfdata)
+		kfree(bfdata);
+
+	devfreq->data = NULL;
+}
+
+static int devfreq_busfreq_handler(struct devfreq *devfreq,
+				   unsigned int event, void *data)
+{
+	int ret = 0;
+
+	switch (event) {
+	case DEVFREQ_GOV_START:
+		ret = busfreq_init(devfreq);
+		break;
+	case DEVFREQ_GOV_STOP:
+		busfreq_exit(devfreq);
+		break;
+	default:
+		break;
+	}
+
+	return ret;
+}
+
+static struct devfreq_governor devfreq_busfreq = {
+	.name = "busfreq",
+	.get_target_freq = devfreq_busfreq_func,
+	.event_handler = devfreq_busfreq_handler,
+};
+
+static int __init devfreq_busfreq_init(void)
+{
+	return devfreq_add_governor(&devfreq_busfreq);
+}
+subsys_initcall(devfreq_busfreq_init);
+
+static void __exit devfreq_busfreq_exit(void)
+{
+	int ret;
+
+	ret = devfreq_remove_governor(&devfreq_busfreq);
+	if (ret)
+		pr_err("%s: failed remove governor %d\n", __func__, ret);
+
+	return;
+}
+module_exit(devfreq_busfreq_exit);
+MODULE_LICENSE("GPL");
diff --git a/include/linux/devfreq.h b/include/linux/devfreq.h
index 678391c1bb0f..df1bf6928d87 100644
--- a/include/linux/devfreq.h
+++ b/include/linux/devfreq.h
@@ -23,6 +23,7 @@
 #define DEVFREQ_GOV_POWERSAVE		"powersave"
 #define DEVFREQ_GOV_USERSPACE		"userspace"
 #define DEVFREQ_GOV_PASSIVE		"passive"
+#define DEVFREQ_GOV_BUSFREQ		"busfreq"
 
 /* DEVFREQ notifier interface */
 #define DEVFREQ_TRANSITION_NOTIFIER	(0)
@@ -310,6 +311,27 @@ struct devfreq_passive_data {
 };
 #endif
 
+/*
+ * "powersave" by default
+ * giving drivers the option to require "performance"
+ */
+#define DEVFREQ_BUSFREQ_HIGH			ULONG_MAX
+#define DEVFREQ_BUSFREQ_LOW			0
+#if IS_ENABLED(CONFIG_DEVFREQ_GOV_BUSFREQ)
+extern int devfreq_busfreq_request(unsigned long freq);
+extern void devfreq_busfreq_release(unsigned long freq);
+#else
+static inline int devfreq_busfreq_request(unsigned long freq)
+{
+	return -ENOSYS;
+}
+
+static inline void devfreq_busfreq_release(unsigned long freq)
+{
+	return;
+}
+#endif
+
 #else /* !CONFIG_PM_DEVFREQ */
 static inline struct devfreq *devfreq_add_device(struct device *dev,
 					  struct devfreq_dev_profile *profile,
-- 
2.20.1


^ permalink raw reply related	[flat|nested] 4+ messages in thread

* Re: [RFC] PM / devfreq: add busfreq governor
  2020-03-25 15:22 ` [RFC] PM / devfreq: add busfreq governor Martin Kepplinger
@ 2020-03-25 22:34   ` Chanwoo Choi
  2020-03-25 22:35   ` Leonard Crestez
  1 sibling, 0 replies; 4+ messages in thread
From: Chanwoo Choi @ 2020-03-25 22:34 UTC (permalink / raw)
  To: Martin Kepplinger, linux-imx, leonard.crestez, myungjoo.ham,
	kyungmin.park
  Cc: linux-pm

Hi Martin,

On 3/26/20 12:22 AM, Martin Kepplinger wrote:
> A devfreq governor to be used by drivers to request ("get") a frequency
> setting they need. Currently only "HIGH" (performance) is available.
> 
> ATTENTION: This is a first draft to serve merely as a basis for discussions!
> ONLY USE FOR TESTING!
> ---
> 
> I wanted to get early feedback on an idea that AFAIK is not yet available
> in the kernel (but something similar via "busfreq" in NXP's tree):
> 
> Let drivers request high (dram) frequencies at runtime if they know they need
> them. In our case the display stack on imx8mq would be the first user,
> looking like so: #include <linux/devfreq.h> and:
> 
> 	--- a/drivers/gpu/drm/bridge/nwl-dsi.c
> 	+++ b/drivers/gpu/drm/bridge/nwl-dsi.c
> 	@@ -1116,6 +1117,9 @@ static int imx8mq_dsi_poweron(struct nwl_dsi *dsi)
> 			ret = reset_control_deassert(dsi->rstc);
> 	 
> 	+       devfreq_busfreq_request(DEVFREQ_BUSFREQ_HIGH);

It seems that the nwl-dsi.c requires the minimum frequency to DRAM driver
in order to guarantee the at least performance.

The devfreq framework already support this feature via PM QoS.
You can refer to following the merged patches[1][2]

[1] 05d7ae15cfb1 ("PM / devfreq: Add PM QoS support")
[2] 27dbc542f651 ("PM / devfreq: Use PM QoS for sysfs min/max_freq")


> 	+
> 		return ret;
> 	@@ -1125,6 +1129,10 @@ static int imx8mq_dsi_poweroff(struct nwl_dsi *dsi)
> 		if (dsi->rstc)
> 			ret = reset_control_assert(dsi->rstc);
> 	+
> 	+       devfreq_busfreq_release(DEVFREQ_BUSFREQ_HIGH);

ditto.

> 	+
> 		return ret;
> 
> 
> Could be called in pm_runtime() calls too.
> 
> _If_ the idea of such or a similar governor is viable, there are at least
> some problems with this implemenation still:
> 
> * The governor saves its data as a global variable (which I don't yet like)
>   but the driver (user) API becomes minimal.
> 
> * the name: In order to add to the devfreq/busfreq confusion, I named the
>   devfreq governor "busfreq" -.-
>   I just wanted to grab NXPs' attention because they do something similar
>   in their tree.
> 
> * ATM we switch between "performace" and "powersave, which is
>   not bad at all, but still limited in case a driver would really only
>   need a medium frequecy for a device to work. doable?
> 
> * I doubt locking is done correctly and the code is overall of bad quality
>   still. Again, It's a first idea.
> 
> I'm glad about any thought or feedback,
> 
> thanks,
>                                    martin
> 
> 
> 
> 
>  drivers/devfreq/Kconfig            |   5 +
>  drivers/devfreq/Makefile           |   1 +
>  drivers/devfreq/governor_busfreq.c | 172 +++++++++++++++++++++++++++++
>  include/linux/devfreq.h            |  22 ++++
>  4 files changed, 200 insertions(+)
>  create mode 100644 drivers/devfreq/governor_busfreq.c
> 
> diff --git a/drivers/devfreq/Kconfig b/drivers/devfreq/Kconfig
> index fc2ea336ef4b..3575cbdcd29e 100644
> --- a/drivers/devfreq/Kconfig
> +++ b/drivers/devfreq/Kconfig
> @@ -74,6 +74,11 @@ config DEVFREQ_GOV_PASSIVE
>  	  through sysfs entries. The passive governor recommends that
>  	  devfreq device uses the OPP table to get the frequency/voltage.
>  
> +config DEVFREQ_GOV_BUSFREQ
> +	tristate "Busfreq"
> +	help
> +	  Sets the frequency that compatible drivers request it to set.
> +
>  comment "DEVFREQ Drivers"
>  
>  config ARM_EXYNOS_BUS_DEVFREQ
> diff --git a/drivers/devfreq/Makefile b/drivers/devfreq/Makefile
> index 61d0edee16f7..657b0b6d92c4 100644
> --- a/drivers/devfreq/Makefile
> +++ b/drivers/devfreq/Makefile
> @@ -6,6 +6,7 @@ obj-$(CONFIG_DEVFREQ_GOV_PERFORMANCE)	+= governor_performance.o
>  obj-$(CONFIG_DEVFREQ_GOV_POWERSAVE)	+= governor_powersave.o
>  obj-$(CONFIG_DEVFREQ_GOV_USERSPACE)	+= governor_userspace.o
>  obj-$(CONFIG_DEVFREQ_GOV_PASSIVE)	+= governor_passive.o
> +obj-$(CONFIG_DEVFREQ_GOV_BUSFREQ)	+= governor_busfreq.o

governor_busfreq.c seems that change the clock rate according to
the user requirement. On that, governor_busfreq only uses
two frequencies level like the high and low frequencies.

I think that it is possible to be implemented by using following combination:
- NXP DRM devfreq driver with userspace governor + PMQoS on nwl-dsi.c

IMO, instead of governor_busfreq.c, better to add the NXP's DRM devfreq driver.

>  
>  # DEVFREQ Drivers
>  obj-$(CONFIG_ARM_EXYNOS_BUS_DEVFREQ)	+= exynos-bus.o
> diff --git a/drivers/devfreq/governor_busfreq.c b/drivers/devfreq/governor_busfreq.c
> new file mode 100644
> index 000000000000..e12d64e9a09b
> --- /dev/null
> +++ b/drivers/devfreq/governor_busfreq.c
> @@ -0,0 +1,172 @@
> +// SPDX-License-Identifier: GPL-2.0-only
> +/*
> + *  linux/drivers/devfreq/governor_busfreq.c
> + *
> + *  Copyright (C) 2020 Purism SPC
> + */
> +
> +#include <linux/slab.h>
> +#include <linux/device.h>
> +#include <linux/devfreq.h>
> +#include <linux/pm.h>
> +#include <linux/mutex.h>
> +#include <linux/module.h>
> +#include "governor.h"
> +
> +struct busfreq_data {
> +	unsigned long user_frequency;
> +	bool valid;
> +	struct devfreq *df;
> +	unsigned int high_count;
> +};
> +
> +static struct busfreq_data *bfdata = NULL;
> +
> +static int devfreq_busfreq_func(struct devfreq *df, unsigned long *freq)
> +{
> +	if (bfdata->valid)
> +		*freq = bfdata->user_frequency;
> +	else
> +		*freq = df->previous_freq; /* No user freq specified yet */
> +
> +	return 0;
> +}
> +
> +int devfreq_busfreq_request(unsigned long freq)
> +{
> +	struct devfreq *devfreq;
> +	int err = 0;
> +
> +	if (!bfdata) {
> +		pr_info("%s: governor not available\n", __func__);
> +		return -ENODEV;
> +	}
> +
> +	if (freq != DEVFREQ_BUSFREQ_HIGH) {
> +		pr_err("%s: undefined frequency\n", __func__);
> +		return -EINVAL;
> +	}
> +
> +	if (freq == DEVFREQ_BUSFREQ_HIGH)
> +		bfdata->high_count++;
> +
> +	devfreq = bfdata->df;
> +
> +	mutex_lock(&devfreq->lock);
> +
> +	bfdata->user_frequency = freq;
> +	bfdata->valid = true;
> +	err = update_devfreq(devfreq); /* calls our get_target_freq */
> +	if (err)
> +		dev_err(&devfreq->dev, "update_devfreq failed: %d\n", err);
> +
> +	mutex_unlock(&devfreq->lock);
> +
> +	return err;
> +}
> +EXPORT_SYMBOL(devfreq_busfreq_request);
> +
> +void devfreq_busfreq_release(unsigned long freq)
> +{
> +	struct devfreq *devfreq;
> +	int err = 0;
> +
> +	if (!bfdata) {
> +		pr_info("%s: governor not available\n", __func__);
> +		return;
> +	}
> +
> +	if (freq != DEVFREQ_BUSFREQ_HIGH) {
> +		pr_err("%s: undefined frequency\n", __func__);
> +		return;
> +	}
> +
> +	if (freq == DEVFREQ_BUSFREQ_HIGH && bfdata->high_count > 0)
> +		bfdata->high_count--;
> +
> +	if (bfdata->high_count)
> +		return;
> +
> +	devfreq = bfdata->df;
> +
> +	mutex_lock(&devfreq->lock);
> +
> +	bfdata->user_frequency = DEVFREQ_BUSFREQ_LOW;
> +	bfdata->valid = true;
> +	err = update_devfreq(devfreq); /* calls our get_target_freq */
> +	if (err)
> +		dev_err(&devfreq->dev, "update_devfreq failed: %d\n", err);
> +
> +	mutex_unlock(&devfreq->lock);
> +}
> +EXPORT_SYMBOL(devfreq_busfreq_release);
> +
> +static int busfreq_init(struct devfreq *devfreq)
> +{
> +	int err = 0;
> +
> +	bfdata = kzalloc(sizeof(struct busfreq_data),
> +		      GFP_KERNEL);
> +	if (!bfdata) {
> +		err = -ENOMEM;
> +		goto out;
> +	}
> +	bfdata->valid = false;
> +	devfreq->data = bfdata;
> +	bfdata->df = devfreq;
> +
> +out:
> +	return err;
> +}
> +
> +static void busfreq_exit(struct devfreq *devfreq)
> +{
> +	if (bfdata)
> +		kfree(bfdata);
> +
> +	devfreq->data = NULL;
> +}
> +
> +static int devfreq_busfreq_handler(struct devfreq *devfreq,
> +				   unsigned int event, void *data)
> +{
> +	int ret = 0;
> +
> +	switch (event) {
> +	case DEVFREQ_GOV_START:
> +		ret = busfreq_init(devfreq);
> +		break;
> +	case DEVFREQ_GOV_STOP:
> +		busfreq_exit(devfreq);
> +		break;
> +	default:
> +		break;
> +	}
> +
> +	return ret;
> +}
> +
> +static struct devfreq_governor devfreq_busfreq = {
> +	.name = "busfreq",
> +	.get_target_freq = devfreq_busfreq_func,
> +	.event_handler = devfreq_busfreq_handler,
> +};
> +
> +static int __init devfreq_busfreq_init(void)
> +{
> +	return devfreq_add_governor(&devfreq_busfreq);
> +}
> +subsys_initcall(devfreq_busfreq_init);
> +
> +static void __exit devfreq_busfreq_exit(void)
> +{
> +	int ret;
> +
> +	ret = devfreq_remove_governor(&devfreq_busfreq);
> +	if (ret)
> +		pr_err("%s: failed remove governor %d\n", __func__, ret);
> +
> +	return;
> +}
> +module_exit(devfreq_busfreq_exit);
> +MODULE_LICENSE("GPL");
> diff --git a/include/linux/devfreq.h b/include/linux/devfreq.h
> index 678391c1bb0f..df1bf6928d87 100644
> --- a/include/linux/devfreq.h
> +++ b/include/linux/devfreq.h
> @@ -23,6 +23,7 @@
>  #define DEVFREQ_GOV_POWERSAVE		"powersave"
>  #define DEVFREQ_GOV_USERSPACE		"userspace"
>  #define DEVFREQ_GOV_PASSIVE		"passive"
> +#define DEVFREQ_GOV_BUSFREQ		"busfreq"
>  
>  /* DEVFREQ notifier interface */
>  #define DEVFREQ_TRANSITION_NOTIFIER	(0)
> @@ -310,6 +311,27 @@ struct devfreq_passive_data {
>  };
>  #endif
>  
> +/*
> + * "powersave" by default
> + * giving drivers the option to require "performance"
> + */
> +#define DEVFREQ_BUSFREQ_HIGH			ULONG_MAX
> +#define DEVFREQ_BUSFREQ_LOW			0
> +#if IS_ENABLED(CONFIG_DEVFREQ_GOV_BUSFREQ)
> +extern int devfreq_busfreq_request(unsigned long freq);
> +extern void devfreq_busfreq_release(unsigned long freq);
> +#else
> +static inline int devfreq_busfreq_request(unsigned long freq)
> +{
> +	return -ENOSYS;
> +}
> +
> +static inline void devfreq_busfreq_release(unsigned long freq)
> +{
> +	return;
> +}
> +#endif
> +
>  #else /* !CONFIG_PM_DEVFREQ */
>  static inline struct devfreq *devfreq_add_device(struct device *dev,
>  					  struct devfreq_dev_profile *profile,
> 


-- 
Best Regards,
Chanwoo Choi
Samsung Electronics

^ permalink raw reply	[flat|nested] 4+ messages in thread

* Re: [RFC] PM / devfreq: add busfreq governor
  2020-03-25 15:22 ` [RFC] PM / devfreq: add busfreq governor Martin Kepplinger
  2020-03-25 22:34   ` Chanwoo Choi
@ 2020-03-25 22:35   ` Leonard Crestez
  2020-03-25 22:51     ` Chanwoo Choi
  1 sibling, 1 reply; 4+ messages in thread
From: Leonard Crestez @ 2020-03-25 22:35 UTC (permalink / raw)
  To: Martin Kepplinger, cw00.choi
  Cc: dl-linux-imx, myungjoo.ham, kyungmin.park, linux-pm

On 2020-03-25 5:23 PM, Martin Kepplinger wrote:
> A devfreq governor to be used by drivers to request ("get") a frequency
> setting they need. Currently only "HIGH" (performance) is available.
> 
> ATTENTION: This is a first draft to serve merely as a basis for discussions!
> ONLY USE FOR TESTING!
> ---
> 
> I wanted to get early feedback on an idea that AFAIK is not yet available
> in the kernel (but something similar via "busfreq" in NXP's tree):

Kernel already provides soc-agnostic mechanism via pm_qos and 
interconnect frameworks. The interconnect framework is more powerful 
because devices express requests in kbps and provides can make better 
decisions based on that.

> Let drivers request high (dram) frequencies at runtime if they know they need
> them. In our case the display stack on imx8mq would be the first user,
> looking like so: #include <linux/devfreq.h> and:
> 
> 	--- a/drivers/gpu/drm/bridge/nwl-dsi.c
> 	+++ b/drivers/gpu/drm/bridge/nwl-dsi.c
> 	@@ -1116,6 +1117,9 @@ static int imx8mq_dsi_poweron(struct nwl_dsi *dsi)
> 			ret = reset_control_deassert(dsi->rstc);
> 	
> 	+       devfreq_busfreq_request(DEVFREQ_BUSFREQ_HIGH);
> 	+
> 		return ret;
> 	@@ -1125,6 +1129,10 @@ static int imx8mq_dsi_poweroff(struct nwl_dsi *dsi)
> 		if (dsi->rstc)
> 			ret = reset_control_assert(dsi->rstc);
> 	+
> 	+       devfreq_busfreq_release(DEVFREQ_BUSFREQ_HIGH);
> 	+
> 		return ret;
> 
> 
> Could be called in pm_runtime() calls too.
> 
> _If_ the idea of such or a similar governor is viable, there are at least
> some problems with this implemenation still:
There's an older RFC which I have to resend which adds support an 
interconnect provider for imx which then makes PM_QOS_MIN_FREQ requests 
to devfreq. It has several advantages:

https://patchwork.kernel.org/project/linux-arm-kernel/list/?series=203215

> * The governor saves its data as a global variable (which I don't yet like)
>    but the driver (user) API becomes minimal.
> 
> * the name: In order to add to the devfreq/busfreq confusion, I named the
>    devfreq governor "busfreq" -.-
>    I just wanted to grab NXPs' attention because they do something similar
>    in their tree.

Interconnect API already exists and is not SOC-specific.

> 
> * ATM we switch between "performace" and "powersave, which is
>    not bad at all, but still limited in case a driver would really only
>    need a medium frequecy for a device to work. doable?

Interconnect requests are made for *bandwdith* and the provider can 
compute minimum frequency based on that. This means that if a video 
device at low resolution requires bandwidth that can be satisfied by a 
"medium" frequency it will work as expected.

> 
> * I doubt locking is done correctly and the code is overall of bad quality
>    still. Again, It's a first idea.


^ permalink raw reply	[flat|nested] 4+ messages in thread

* Re: [RFC] PM / devfreq: add busfreq governor
  2020-03-25 22:35   ` Leonard Crestez
@ 2020-03-25 22:51     ` Chanwoo Choi
  0 siblings, 0 replies; 4+ messages in thread
From: Chanwoo Choi @ 2020-03-25 22:51 UTC (permalink / raw)
  To: Leonard Crestez, Martin Kepplinger
  Cc: dl-linux-imx, myungjoo.ham, kyungmin.park, linux-pm

Hi Leonard,

On 3/26/20 7:35 AM, Leonard Crestez wrote:
> On 2020-03-25 5:23 PM, Martin Kepplinger wrote:
>> A devfreq governor to be used by drivers to request ("get") a frequency
>> setting they need. Currently only "HIGH" (performance) is available.
>>
>> ATTENTION: This is a first draft to serve merely as a basis for discussions!
>> ONLY USE FOR TESTING!
>> ---
>>
>> I wanted to get early feedback on an idea that AFAIK is not yet available
>> in the kernel (but something similar via "busfreq" in NXP's tree):
> 
> Kernel already provides soc-agnostic mechanism via pm_qos and 
> interconnect frameworks. The interconnect framework is more powerful 
> because devices express requests in kbps and provides can make better 
> decisions based on that.
> 
>> Let drivers request high (dram) frequencies at runtime if they know they need
>> them. In our case the display stack on imx8mq would be the first user,
>> looking like so: #include <linux/devfreq.h> and:
>>
>> 	--- a/drivers/gpu/drm/bridge/nwl-dsi.c
>> 	+++ b/drivers/gpu/drm/bridge/nwl-dsi.c
>> 	@@ -1116,6 +1117,9 @@ static int imx8mq_dsi_poweron(struct nwl_dsi *dsi)
>> 			ret = reset_control_deassert(dsi->rstc);
>> 	
>> 	+       devfreq_busfreq_request(DEVFREQ_BUSFREQ_HIGH);
>> 	+
>> 		return ret;
>> 	@@ -1125,6 +1129,10 @@ static int imx8mq_dsi_poweroff(struct nwl_dsi *dsi)
>> 		if (dsi->rstc)
>> 			ret = reset_control_assert(dsi->rstc);
>> 	+
>> 	+       devfreq_busfreq_release(DEVFREQ_BUSFREQ_HIGH);
>> 	+
>> 		return ret;
>>
>>
>> Could be called in pm_runtime() calls too.
>>
>> _If_ the idea of such or a similar governor is viable, there are at least
>> some problems with this implemenation still:
> There's an older RFC which I have to resend which adds support an 
> interconnect provider for imx which then makes PM_QOS_MIN_FREQ requests 
> to devfreq. It has several advantages:
> 
> https://patchwork.kernel.org/project/linux-arm-kernel/list/?series=203215

I'm expecting your patches as your patches are good use-case of PM_QoS
on devfreq. Thanks.

[snip]

-- 
Best Regards,
Chanwoo Choi
Samsung Electronics

^ permalink raw reply	[flat|nested] 4+ messages in thread

end of thread, other threads:[~2020-03-25 22:42 UTC | newest]

Thread overview: 4+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
     [not found] <CGME20200325152302epcas1p3ba57cfef70d9d8b8dff6d9e2bb09b321@epcas1p3.samsung.com>
2020-03-25 15:22 ` [RFC] PM / devfreq: add busfreq governor Martin Kepplinger
2020-03-25 22:34   ` Chanwoo Choi
2020-03-25 22:35   ` Leonard Crestez
2020-03-25 22:51     ` Chanwoo Choi

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).