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

Commit 9ba1e5d8 authored by Joonwoo Park's avatar Joonwoo Park
Browse files

sched: kill unnecessary divisions on fast path 



The max_possible_efficiency and CPU's efficiency are fixed values which
are determined at cluster allocation time.  Avoid division on the fast
by using precomputed scale factor.

Also update_cpu_busy_time() doesn't need to know how many full windows
have elapsed.  Thus replace unneeded division with simple comparison.

Change-Id: I2be1aad3fb9b895e4f0917d05bd8eade985bbccf
Suggested-by: default avatarSyed Rameez Mustafa <rameezmustafa@codeaurora.org>
Signed-off-by: default avatarJoonwoo Park <joonwoop@codeaurora.org>
parent 8ac49070

kernel/sched/core.c

+13 −13
Original line number Diff line number Diff line
@@ -1330,6 +1330,7 @@ static struct sched_cluster init_cluster = { 
	.dstate			=	0,

	.dstate_wakeup_energy	=	0,

	.dstate_wakeup_latency	=	0,

	.exec_scale_factor	=	1024,

};



void update_all_clusters_stats(void)
@@ -1347,6 +1348,10 @@ void update_all_clusters_stats(void) 
			compute_max_possible_capacity(cluster);

		cluster->load_scale_factor = compute_load_scale_factor(cluster);



		cluster->exec_scale_factor =

			DIV_ROUND_UP(cluster->efficiency * 1024,

				     max_possible_efficiency);



		if (mpc > highest_mpc)

			highest_mpc = mpc;
@@ -1864,15 +1869,11 @@ update_window_start(struct rq *rq, u64 wallclock) 


static inline u64 scale_exec_time(u64 delta, struct rq *rq)

{

	int cpu = cpu_of(rq);

	int sf;

	u32 freq;



	freq = cpu_cycles_to_freq(rq->cc.cycles, rq->cc.time);

	delta = DIV64_U64_ROUNDUP(delta * freq, max_possible_freq);

	sf = DIV_ROUND_UP(cpu_efficiency(cpu) * 1024, max_possible_efficiency);



	delta *= sf;

	delta *= rq->cluster->exec_scale_factor;

	delta >>= 10;



	return delta;
@@ -2315,7 +2316,7 @@ void update_task_pred_demand(struct rq *rq, struct task_struct *p, int event) 
static void update_cpu_busy_time(struct task_struct *p, struct rq *rq,

				 int event, u64 wallclock, u64 irqtime)

{

	int new_window, nr_full_windows = 0;

	int new_window, full_window = 0;

	int p_is_curr_task = (p == rq->curr);

	u64 mark_start = p->ravg.mark_start;

	u64 window_start = rq->window_start;
@@ -2332,8 +2333,7 @@ static void update_cpu_busy_time(struct task_struct *p, struct rq *rq, 


	new_window = mark_start < window_start;

	if (new_window) {

		nr_full_windows = div64_u64((window_start - mark_start),

						window_size);

		full_window = (window_start - mark_start) >= window_size;

		if (p->ravg.active_windows < USHRT_MAX)

			p->ravg.active_windows++;

	}
@@ -2368,7 +2368,7 @@ static void update_cpu_busy_time(struct task_struct *p, struct rq *rq, 
			u64 curr_sum = rq->curr_runnable_sum;

			u64 nt_curr_sum = rq->nt_curr_runnable_sum;



			if (nr_full_windows)

			if (full_window)

				curr_sum = nt_curr_sum = 0;



			rq->prev_runnable_sum = curr_sum;
@@ -2380,7 +2380,7 @@ static void update_cpu_busy_time(struct task_struct *p, struct rq *rq, 
	} else {

		if (p_is_curr_task && new_window) {

			flip_counters = 1;

			if (nr_full_windows)

			if (full_window)

				prev_sum_reset = 1;

		}

	}
@@ -2390,7 +2390,7 @@ static void update_cpu_busy_time(struct task_struct *p, struct rq *rq, 
	if (new_window && !is_idle_task(p) && !exiting_task(p)) {

		u32 curr_window = 0;



		if (!nr_full_windows)

		if (!full_window)

			curr_window = p->ravg.curr_window;



		p->ravg.prev_window = curr_window;
@@ -2469,7 +2469,7 @@ static void update_cpu_busy_time(struct task_struct *p, struct rq *rq, 
		 * Irqtime can't be accounted by a task that isn't the

		 * currently running task. */



		if (!nr_full_windows) {

		if (!full_window) {

			/* A full window hasn't elapsed, account partial

			 * contribution to previous completed window. */

			delta = scale_exec_time(window_start - mark_start, rq);
@@ -2513,7 +2513,7 @@ static void update_cpu_busy_time(struct task_struct *p, struct rq *rq, 
		 * An example of this would be a task that starts execution

		 * and then sleeps once a new window has begun. */



		if (!nr_full_windows) {

		if (!full_window) {

			/* A full window hasn't elapsed, account partial

			 * contribution to previous completed window. */

			delta = scale_exec_time(window_start - mark_start, rq);

kernel/sched/debug.c

+1 −0
Original line number Diff line number Diff line
@@ -332,6 +332,7 @@ do { \ 
	P(cluster->efficiency);

	P(cluster->cur_freq);

	P(cluster->max_freq);

	P(cluster->exec_scale_factor);

#endif

#endif

#ifdef CONFIG_SCHED_HMP

kernel/sched/sched.h

+1 −0
Original line number Diff line number Diff line
@@ -353,6 +353,7 @@ struct sched_cluster { 
	int capacity;

	int efficiency; /* Differentiate cpus with different IPC capability */

	int load_scale_factor;

	unsigned int exec_scale_factor;

	/*

	 * max_freq = user maximum

	 * max_mitigated_freq = thermal defined maximum