Donate to e Foundation | Murena handsets with /e/OS | Own a part of Murena! Learn more

Commit 858d5751 authored by Joonwoo Park's avatar Joonwoo Park
Browse files

sched: introduce per CPU sched_load_boost knob 



Introduce a new knob sched_load_boost that allows userspace to
provide additional CPU capacity margin to work running on CPUs.  The
applied margin is linearly proportional to the actual load on the CPU
and sched_load_boost.  The applied margin is not affected by the max
capacity of the CPUs in the system.

Userspace could increase this for exceedingly sporadic workloads in
case the existing margin isn't sufficient.

Also, while at it, refactor cpu_util_freq().

Change-Id: I07379aa3c5d9e670f67886323e2e32225a1bdef9
Signed-off-by: default avatarJoonwoo Park <joonwoop@codeaurora.org>
parent b040b115

drivers/base/cpu.c

+55 −0
Original line number Diff line number Diff line
@@ -208,6 +208,59 @@ static struct attribute_group cpu_isolated_attr_group = { 


#endif



static ssize_t show_sched_load_boost(struct device *dev,

				struct device_attribute *attr, char *buf)

{

	ssize_t rc;

	unsigned int boost;

	struct cpu *cpu = container_of(dev, struct cpu, dev);

	int cpuid = cpu->dev.id;



	boost = per_cpu(sched_load_boost, cpuid);

	rc = snprintf(buf, PAGE_SIZE-2, "%d\n", boost);



	return rc;

}



static ssize_t __ref store_sched_load_boost(struct device *dev,

				struct device_attribute *attr,

				const char *buf, size_t count)

{

	int err;

	int boost;

	struct cpu *cpu = container_of(dev, struct cpu, dev);

	int cpuid = cpu->dev.id;



	err = kstrtoint(strstrip((char *)buf), 0, &boost);

	if (err)

		return err;



	/*

	 * -100 is low enough to cancel out CPU's load and make it near zro.

	 * 1000 is close to the maximum value that cpu_util_freq_{walt,pelt}

	 * can take without overflow.

	 */

	if (boost < -100 || boost > 1000)

		return -EINVAL;



	per_cpu(sched_load_boost, cpuid) = boost;



	return count;

}



static DEVICE_ATTR(sched_load_boost, 0644,

		   show_sched_load_boost,

		   store_sched_load_boost);



static struct attribute *sched_cpu_attrs[] = {

	&dev_attr_sched_load_boost.attr,

	NULL

};



static struct attribute_group sched_cpu_attr_group = {

	.attrs = sched_cpu_attrs,

};



static const struct attribute_group *common_cpu_attr_groups[] = {

#ifdef CONFIG_KEXEC

	&crash_note_cpu_attr_group,
@@ -215,6 +268,7 @@ static const struct attribute_group *common_cpu_attr_groups[] = { 
#ifdef CONFIG_HOTPLUG_CPU

	&cpu_isolated_attr_group,

#endif

	&sched_cpu_attr_group,

	NULL

};
@@ -225,6 +279,7 @@ static const struct attribute_group *hotplugable_cpu_attr_groups[] = { 
#ifdef CONFIG_HOTPLUG_CPU

	&cpu_isolated_attr_group,

#endif

	&sched_cpu_attr_group,

	NULL

};

include/linux/sched.h

+2 −0
Original line number Diff line number Diff line
@@ -3902,4 +3902,6 @@ void cpufreq_add_update_util_hook(int cpu, struct update_util_data *data, 
void cpufreq_remove_update_util_hook(int cpu);

#endif /* CONFIG_CPU_FREQ */



extern DEFINE_PER_CPU_READ_MOSTLY(int, sched_load_boost);



#endif

kernel/sched/core.c

+3 −3
Original line number Diff line number Diff line
@@ -3232,9 +3232,9 @@ unsigned int capacity_margin_freq = 1280; /* ~20% margin */ 


static inline

unsigned long sum_capacity_reqs(unsigned long cfs_cap,

				struct sched_capacity_reqs *scr)

				struct sched_capacity_reqs *scr, int cpu)

{

	unsigned long total = add_capacity_margin(cfs_cap + scr->rt);

	unsigned long total = add_capacity_margin(cfs_cap + scr->rt, cpu);

	return total += scr->dl;

}
@@ -3245,7 +3245,7 @@ static void sched_freq_tick_pelt(int cpu) 
	struct sched_capacity_reqs *scr;



	scr = &per_cpu(cpu_sched_capacity_reqs, cpu);

	if (sum_capacity_reqs(cpu_utilization, scr) < capacity_curr)

	if (sum_capacity_reqs(cpu_utilization, scr, cpu) < capacity_curr)

		return;



	/*

kernel/sched/fair.c

+2 −1
Original line number Diff line number Diff line
@@ -89,6 +89,7 @@ unsigned int sysctl_sched_is_big_little = 1; 
unsigned int sysctl_sched_sync_hint_enable = 1;

unsigned int sysctl_sched_initial_task_util = 0;

unsigned int sysctl_sched_cstate_aware = 1;

DEFINE_PER_CPU_READ_MOSTLY(int, sched_load_boost);



#ifdef CONFIG_SCHED_WALT

unsigned int sysctl_sched_use_walt_cpu_util = 1;
@@ -6750,7 +6751,7 @@ static int energy_aware_wake_cpu(struct task_struct *p, int target, int sync) 
			cpu_idle_idx = idle_get_state_idx(cpu_rq(i));



			if (!need_idle &&

			    add_capacity_margin(new_util_cum) <

			    add_capacity_margin(new_util_cum, i) <

			    capacity_curr_of(i)) {

				if (sysctl_sched_cstate_aware) {

					if (cpu_idle_idx < min_idle_idx) {

kernel/sched/rt.c

+1 −1
Original line number Diff line number Diff line
@@ -1811,7 +1811,7 @@ static int find_lowest_rq(struct task_struct *task) 
					cpu_idle_idx =

					     idle_get_state_idx(cpu_rq(cpu));



				if (add_capacity_margin(new_util_cum) <

				if (add_capacity_margin(new_util_cum, cpu) <

				    capacity_curr_of(cpu)) {

					if (cpu_idle_idx < best_cpu_idle_idx ||

					    (best_cpu != task_cpu(task) &&