Re: [PATCH v4 5/6] PM / devfreq: imx8m-ddrc: Measure bandwidth with perf

[Leonard Crestez <leonard.crestez@nxp.com>]

On 11.11.2019 07:13, Chanwoo Choi wrote:
> Hi Leonard,
> 
> On 11/9/19 7:39 AM, Leonard Crestez wrote:
>> The imx8m ddrc has a performance monitoring block attached which can
>> be used to measure bandwidth usage and automatically adjust frequency.
>>
>> There is already a perf driver for that block so instead of implementing
>> a devfreq events driver use the in-kernel perf API to implement
>> get_dev_status directly.
>>
>> Signed-off-by: Leonard Crestez <leonard.crestez@nxp.com>
>> ---
>>   drivers/devfreq/imx8m-ddrc.c | 153 +++++++++++++++++++++++++++++++++++
>>   1 file changed, 153 insertions(+)
>>
>> diff --git a/drivers/devfreq/imx8m-ddrc.c b/drivers/devfreq/imx8m-ddrc.c
>> index 51903fee21a7..6372191f72d7 100644
>> --- a/drivers/devfreq/imx8m-ddrc.c
>> +++ b/drivers/devfreq/imx8m-ddrc.c
>> @@ -11,10 +11,13 @@
>>   #include <linux/pm_opp.h>
>>   #include <linux/clk.h>
>>   #include <linux/clk-provider.h>
>>   #include <linux/arm-smccc.h>
>>   
>> +#include <asm/perf_event.h>
>> +#include <linux/perf_event.h>
>> +
>>   #define IMX_SIP_DDR_DVFS			0xc2000004
>>   
>>   /* Values starting from 0 switch to specific frequency */
>>   #define IMX_SIP_DDR_FREQ_SET_HIGH		0x00
>>   
>> @@ -78,10 +81,22 @@ struct imx8m_ddrc {
>>   	struct clk *dram_alt;
>>   	struct clk *dram_apb;
>>   
>>   	int freq_count;
>>   	struct imx8m_ddrc_freq freq_table[IMX8M_DDRC_MAX_FREQ_COUNT];
>> +
>> +	/* For measuring load with perf events: */
>> +	struct platform_device *pmu_pdev;
>> +	struct pmu *pmu;
>> +
>> +	struct perf_event_attr rd_event_attr;
>> +	struct perf_event_attr wr_event_attr;
>> +	struct perf_event *rd_event;
>> +	struct perf_event *wr_event;
>> +
>> +	u64 last_rd_val, last_rd_ena, last_rd_run;
>> +	u64 last_wr_val, last_wr_ena, last_wr_run;
>>   };
>>   
>>   static struct imx8m_ddrc_freq *imx8m_ddrc_find_freq(struct imx8m_ddrc *priv,
>>   						    unsigned long rate)
>>   {
>> @@ -245,10 +260,131 @@ static int imx8m_ddrc_get_cur_freq(struct device *dev, unsigned long *freq)
>>   	*freq = clk_get_rate(priv->dram_core);
>>   
>>   	return 0;
>>   }
>>   
>> +static int imx8m_ddrc_get_dev_status(struct device *dev,
>> +				     struct devfreq_dev_status *stat)
>> +{
>> +	struct imx8m_ddrc *priv = dev_get_drvdata(dev);
>> +
>> +	stat->current_frequency = clk_get_rate(priv->dram_core);
>> +
>> +	if (priv->rd_event && priv->wr_event) {
>> +		u64 rd_delta, rd_val, rd_ena, rd_run;
>> +		u64 wr_delta, wr_val, wr_ena, wr_run;
>> +
>> +		rd_val = perf_event_read_value(priv->rd_event,
>> +					       &rd_ena, &rd_run);
>> +		wr_val = perf_event_read_value(priv->wr_event,
>> +					       &wr_ena, &wr_run);
>> +
>> +		rd_delta = (rd_val - priv->last_rd_val) *
>> +			   (rd_ena - priv->last_rd_ena);
>> +		do_div(rd_delta, rd_run - priv->last_rd_run);
>> +		priv->last_rd_val = rd_val;
>> +		priv->last_rd_ena = rd_ena;
>> +		priv->last_rd_run = rd_run;
>> +
>> +		wr_delta = (wr_val - priv->last_wr_val) *
>> +			   (wr_ena - priv->last_wr_ena);
>> +		do_div(wr_delta, wr_run - priv->last_wr_run);
>> +		priv->last_wr_val = wr_val;
>> +		priv->last_wr_ena = wr_ena;
>> +		priv->last_wr_run = wr_run;
>> +
>> +		/* magic numbers, possibly wrong */
>> +		stat->busy_time = 4 * (rd_delta + wr_delta);
>> +		stat->total_time = stat->current_frequency;
>> +	} else {
>> +		stat->busy_time = 0;
>> +		stat->total_time = 0;
>> +	}
>> +
>> +	return 0;
>> +}
>> +
>> +static int imx8m_ddrc_perf_disable(struct imx8m_ddrc *priv)
>> +{
>> +	/* release and set to NULL */
>> +	if (!IS_ERR_OR_NULL(priv->rd_event))
>> +		perf_event_release_kernel(priv->rd_event);
>> +	if (!IS_ERR_OR_NULL(priv->wr_event))
>> +		perf_event_release_kernel(priv->wr_event);
>> +	priv->rd_event = NULL;
>> +	priv->wr_event = NULL;
>> +
>> +	return 0;
>> +}
>> +
>> +static int imx8m_ddrc_perf_enable(struct imx8m_ddrc *priv)
> 
> Actually, this function have to call the just function for enabling
> the bandwidth monitoring instead of writing the detailed something.
> It looks like that you move the part of the different device driver into here.

This is the code for enabling bandwith monitoring: it creates perf 
counters using in-kernel API. The perf api doesn't seem to expose 
anything else to enable/disable the counter after creation and honestly 
just create/destroy seems fine.

As far as I can tell bandwidth monitoring in devfreq is always enabled 
on probe anyway, no matter which governor is in use. It would be nice if 
devfreq drivers could receive a callback and dynamically acquire/release 
monitoring resources only when the ondemand governor is in used.

Until then this driver will permanently allocate 2 (out of 3) counters 
in ddr pmu hardware.

>> +{
>> +	int ret;
>> +
>> +	priv->rd_event_attr.size = sizeof(priv->rd_event_attr);
>> +	priv->rd_event_attr.type = priv->pmu->type;
>> +	priv->rd_event_attr.config = 0x2a;
>> +
>> +	priv->rd_event = perf_event_create_kernel_counter(
>> +			&priv->rd_event_attr, 0, NULL, NULL, NULL);
>> +	if (IS_ERR(priv->rd_event)) {
>> +		ret = PTR_ERR(priv->rd_event);
>> +		goto err;
>> +	}
>> +
>> +	priv->wr_event_attr.size = sizeof(priv->wr_event_attr);
>> +	priv->wr_event_attr.type = priv->pmu->type;
>> +	priv->wr_event_attr.config = 0x2b;
>> +
>> +	priv->wr_event = perf_event_create_kernel_counter(
>> +			&priv->wr_event_attr, 0, NULL, NULL, NULL);
>> +	if (IS_ERR(priv->wr_event)) {
>> +		ret = PTR_ERR(priv->wr_event);
>> +		goto err;
>> +	}
>> +
>> +	return 0;
>> +
>> +err:
>> +	imx8m_ddrc_perf_disable(priv);
>> +	return ret;
>> +}
>> +
>> +static int imx8m_ddrc_init_events(struct device *dev,
>> +				  struct device_node *events_node)
> 
> ditto.
> 
>> +{
>> +	struct imx8m_ddrc *priv = dev_get_drvdata(dev);
>> +	struct device_driver *driver;
>> +
>> +	/*
>> +	 * We need pmu->type for perf_event_attr but there is no API for
>> +	 * mapping device_node to pmu. Fetch private data for imx-ddr-pmu and
>> +	 * cast that to a struct pmu instead.
>> +	 */
>> +	priv->pmu_pdev = of_find_device_by_node(events_node);
>> +	if (!priv->pmu_pdev)
>> +		return -EPROBE_DEFER;
>> +	driver = priv->pmu_pdev->dev.driver;
>> +	if (!driver)
>> +		return -EPROBE_DEFER;
>> +	if (strcmp(driver->name, "imx-ddr-pmu")) {
>> +		dev_warn(dev, "devfreq-events node %pOF has unexpected driver %s\n",
>> +				events_node, driver->name);
>> +		return -ENODEV;
>> +	}
>> +
>> +	priv->pmu = platform_get_drvdata(priv->pmu_pdev);
> 
> It seems that you access the different device driver without
> any standard interface from some linux kernel subsystem.
> 
> For the communication or control between different device drivers,
> we have to use the standard interface instead of direct access or call.
> I think that it make it too tightly coupled driver between them.
> 
> So, I developed the devfreq-event subsystem for this reason
> in order to remove the non-standard direct access to other device driver.
> 
> Unfortunately, I can't agree this approach. I don't prefer to use
> the direct access of other device driver(imx-ddr-pmu). You need to
> use standard interface provided from subsystem. or You need to make
> the new device driver with devfreq-event subsystem.

This could be cleaned-up by adding a new API to "fetch struct pmu* by 
struct device_node*" as available for many other subsystems. The perf 
api is otherwise standard/generic and has a few other in-kernel users.

The perf driver for DDR PMU is already functional and useful for 
profiling and reusing it seem very worthwhile. If you're suggesting I 
implemented an unrelated "devfreq-event" driver then how would it get 
probed? There's only one PMU node in DT.

I wouldn't mind to delay the monitoring part into a second series.

>> +	if (!priv->pmu) {
>> +		dev_err(dev, "devfreq-events device missing private data\n");
>> +		return -EINVAL;
>> +	}
>> +
>> +	dev_dbg(dev, "events from pmu %s\n", priv->pmu->name);
>> +
>> +	return imx8m_ddrc_perf_enable(priv);
>> +}
>> +
>>   static int imx8m_ddrc_init_freq_info(struct device *dev)
>>   {
>>   	struct imx8m_ddrc *priv = dev_get_drvdata(dev);
>>   	struct arm_smccc_res res;
>>   	int index;
>> @@ -328,17 +464,23 @@ static int imx8m_ddrc_check_opps(struct device *dev)
>>   	return 0;
>>   }
>>   
>>   static void imx8m_ddrc_exit(struct device *dev)
>>   {
>> +	struct imx8m_ddrc *priv = dev_get_drvdata(dev);
>> +
>> +	imx8m_ddrc_perf_disable(priv);
>> +	platform_device_put(priv->pmu_pdev);
>> +
>>   	dev_pm_opp_of_remove_table(dev);
>>   }
>>   
>>   static int imx8m_ddrc_probe(struct platform_device *pdev)
>>   {
>>   	struct device *dev = &pdev->dev;
>>   	struct imx8m_ddrc *priv;
>> +	struct device_node *events_node;
>>   	const char *gov = DEVFREQ_GOV_USERSPACE;
>>   	int ret;
>>   
>>   	priv = devm_kzalloc(dev, sizeof(*priv), GFP_KERNEL);
>>   	if (!priv)
>> @@ -350,10 +492,19 @@ static int imx8m_ddrc_probe(struct platform_device *pdev)
>>   	if (ret) {
>>   		dev_err(dev, "failed to init firmware freq info: %d\n", ret);
>>   		return ret;
>>   	}
>>   
>> +	events_node = of_parse_phandle(dev->of_node, "devfreq-events", 0);
>> +	if (events_node) {
>> +		ret = imx8m_ddrc_init_events(dev, events_node);
>> +		of_node_put(events_node);
>> +		if (ret)
>> +			goto err;
>> +		gov = DEVFREQ_GOV_SIMPLE_ONDEMAND;
>> +	}
>> +
>>   	priv->dram_core = devm_clk_get(dev, "core");
>>   	priv->dram_pll = devm_clk_get(dev, "pll");
>>   	priv->dram_alt = devm_clk_get(dev, "alt");
>>   	priv->dram_apb = devm_clk_get(dev, "apb");
>>   	if (IS_ERR(priv->dram_core) ||
>> @@ -390,10 +541,12 @@ static int imx8m_ddrc_probe(struct platform_device *pdev)
>>   	}
>>   
>>   	return 0;
>>   
>>   err:
>> +	imx8m_ddrc_perf_disable(priv);
>> +	platform_device_put(priv->pmu_pdev);
>>   	dev_pm_opp_of_remove_table(dev);
>>   	return ret;
>>   }
>>   
>>   static const struct of_device_id imx8m_ddrc_of_match[] = {