multiple and conflicting implementation of cpu_load_get

[nashif]

Describe the bug

we have two conflicting implementations of cpu_load_get:

include/zephyr/cpu_load/cpu_load.h:int cpu_load_get(int cpu_id);
include/zephyr/debug/cpu_load.h:int cpu_load_get(bool reset);

The recent one added to support cpu_freq expands hw support and also supports multicore, the original one we had is tied into tracing and very limited in HW support.

Regression

• This is a regression.

Steps to reproduce

git grep cpu_load_get

Relevant log output

include/zephyr/cpu_load/cpu_load.h:int cpu_load_get(int cpu_id);
include/zephyr/debug/cpu_load.h:int cpu_load_get(bool reset);

Impact

Major – Severely degrades functionality; workaround is difficult or unavailable.

Environment

No response

Additional Context

We should just have one. My preference is to remove the one under subsys/debug and see how we can enhance the new interface with some of the features that came with the debug variant, i.e. reset and counter support...

[nashif]

@nordic-krch @seankyer FYI

[nordic-krch]

I haven't seen this second cpu_load that was added later. It's also important to note that cpu_load added by @seankyer is just a wrapper around thread runtime states and it does not include time spent in interrupts.

This simple demo based on hello_world shows 10% cpu_load with CONFIG_CPU_LOAD and 0% with CONFIG_CPU_LOAD_METRIC:

#include <stdio.h>
#include <zephyr/kernel.h>
#ifdef CONFIG_CPU_LOAD
#include <zephyr/debug/cpu_load.h>
#elif defined(CONFIG_CPU_LOAD_METRIC)
#include <zephyr/cpu_load/cpu_load.h>
#endif

static void timeout(struct k_timer *timer)
{
	k_busy_wait(20);
}

K_TIMER_DEFINE(tmr, timeout, NULL);

int main(void)
{
	printf("Hello World! %s\n", CONFIG_BOARD_TARGET);

#ifdef CONFIG_CPU_LOAD
	(void)cpu_load_get(true);
#elif defined(CONFIG_CPU_LOAD_METRIC)
	(void)cpu_load_get(0);
#endif
	k_timer_start(&tmr, K_USEC(500), K_USEC(500));
	k_msleep(5);
#ifdef CONFIG_CPU_LOAD
	uint32_t load = cpu_load_get(true);
	printk("load:%d.%d\n", load / 10, load % 10);
#elif defined(CONFIG_CPU_LOAD_METRIC)
	uint32_t load = cpu_load_get(0);
	printk("load:%d\n", load);
#endif
	k_timer_stop(&tmr);

	return 0;
}

[nordic-krch]

Let me just mention that some of our SoC have HW which allows to get cycle accurate CPU active time and in that example it shows 10.168% vs 10.2% load shown by the software solution so it proves that it is accurate.

[nashif]

@peter-mitsis Can you please look into why cpu stats from kernel are showing different numbers? Are we taking interrupts into account?

getting:

west build -p -b mps2/an385  samples/hello_world/ -t run -- -DCONFIG_CPU_LOAD_METRIC=y
...
..

*** Booting Zephyr OS build v4.2.0-6152-gfd51dde8f5ca ***
Hello World! mps2/an385
load:0

and

west build -p -b mps2/an385  samples/hello_world/ -t run -- -DCONFIG_CPU_LOAD=y

*** Booting Zephyr OS build v4.2.0-6152-gfd51dde8f5ca ***
Hello World! mps2/an385
load:1.8

[peter-mitsis]

Looking into this.