[COMMON] sched: ems: Change ontime policy and algorithm
authorDaeyeong Lee <daeyeong.lee@samsung.com>
Mon, 3 Sep 2018 06:07:27 +0000 (15:07 +0900)
committerCosmin Tanislav <demonsingur@gmail.com>
Mon, 22 Apr 2024 17:24:57 +0000 (20:24 +0300)
- If the task is above upper_boudnary of task's cpu,
  Any coregroup with a higher mips than task's cpu can be a candidates.
- If there are multiple candidate coregroups,
  calculate energy and select the most efficient CPU.
- If the task is below lower_boundary of task's cpu,
  leave it to the core selection afterwards.

Change-Id: I38667d4382692115d74ec27eaa4169e6124bd749
Signed-off-by: Daeyeong Lee <daeyeong.lee@samsung.com>
kernel/sched/ems/ontime.c

index a8326a20cad29eb2fc363bfc13222d0df9936880..c48b7c8bdf078f57cca6919edbe9d74adbfbdbf2 100644 (file)
@@ -124,100 +124,118 @@ static int
 ontime_select_fit_cpus(struct task_struct *p, struct cpumask *fit_cpus)
 {
        struct ontime_cond *curr;
-       struct cpumask cpus;
-       int cpu = task_cpu(p);
-
-       cpumask_and(fit_cpus, cpu_coregroup_mask(cpu), tsk_cpus_allowed(p));
-       if (cpumask_empty(fit_cpus))
-               return -ENODEV;
+       int src_cpu = task_cpu(p);
 
-       curr = get_current_cond(cpu);
+       curr = get_current_cond(src_cpu);
        if (!curr)
                return -EINVAL;
 
+       cpumask_clear(fit_cpus);
+
        if (ontime_load_avg(p) >= curr->upper_boundary) {
                /*
-                * 1. If task's load is bigger than upper boundary,
-                * find fit_cpus among next coregroup.
+                * If task's load is above upper boundary of source,
+                * find fit_cpus that have higher mips than source.
                 */
-               list_for_each_entry_from(curr, &cond_list, list) {
-                       cpumask_and(&cpus, &curr->cpus, tsk_cpus_allowed(p));
-                       if (cpumask_empty(&cpus))
-                               break;
-
-                       cpumask_copy(fit_cpus, &cpus);
+               list_for_each_entry(curr, &cond_list, list) {
+                       int dst_cpu = cpumask_first(&curr->cpus);
 
-                       if (ontime_load_avg(p) < curr->upper_boundary)
-                               break;
+                       if (get_cpu_mips(src_cpu) < get_cpu_mips(dst_cpu))
+                               cpumask_or(fit_cpus, fit_cpus, &curr->cpus);
                }
-       } else if (ontime_load_avg(p) < curr->lower_boundary) {
+       } else if (ontime_load_avg(p) >= curr->lower_boundary) {
                /*
-                * 2. If task's load is smaller than lower boundary,
-                * find fit_cpus among prev coregroup.
+                * If task's load is between upper boundary and lower boundary of source,
+                * fit cpus is the coregroup of source.
                 */
-               list_for_each_entry_from_reverse(curr, &cond_list, list) {
-                       cpumask_and(&cpus, &curr->cpus, tsk_cpus_allowed(p));
-                       if (cpumask_empty(&cpus))
-                               break;
-
-                       cpumask_copy(fit_cpus, &cpus);
-
-                       if (ontime_load_avg(p) >= curr->lower_boundary)
-                               break;
-               }
+               cpumask_copy(fit_cpus, cpu_coregroup_mask(src_cpu));
+       } else {
+               /*
+                * If task's load is below lower boundary,
+                * don't need to do ontime migration or wakeup.
+                */
+               return -1;
        }
 
+       if (cpumask_empty(fit_cpus))
+               return -1;
+
        return 0;
 }
 
 static int
-ontime_select_target_cpu(struct task_struct *p, struct cpumask *dst_cpus)
+ontime_select_target_cpu(struct task_struct *p, struct cpumask *fit_cpus)
 {
-       int cpu;
-       int best_cpu = -1, backup_cpu = -1;
-       unsigned int min_exit_latency = UINT_MAX;
-       unsigned long min_util = ULONG_MAX;
+       struct cpumask candidates;
+       int cpu, min_energy_cpu = -1;
+       int candidate_count = 0;
 
-       rcu_read_lock();
+       cpumask_clear(&candidates);
 
-       for_each_cpu_and(cpu, dst_cpus, cpu_active_mask) {
-               if (cpu_rq(cpu)->ontime_migrating)
+       /*
+        * First) Find min_util_cpu for each coregroup in fit cpus and candidate it.
+        */
+       for_each_cpu(cpu, fit_cpus) {
+               int i, min_util_cpu = -1;
+               unsigned long coverage_util, min_util = ULONG_MAX;
+
+               if (cpu != cpumask_first(cpu_coregroup_mask(cpu)))
                        continue;
 
-               if (idle_cpu(cpu)) {
-                       /* 1. Find shallowest idle cpu. */
-                       struct cpuidle_state *idle = idle_get_state(cpu_rq(cpu));
+               coverage_util = capacity_orig_of(cpu) * get_coverage_ratio(cpu);
 
-                       if (!idle) {
-                               rcu_read_unlock();
-                               return cpu;
-                       }
+               for_each_cpu_and(i, cpu_coregroup_mask(cpu), cpu_active_mask) {
+                       unsigned long new_util;
 
-                       if (idle->exit_latency < min_exit_latency) {
-                               min_exit_latency = idle->exit_latency;
-                               best_cpu = cpu;
-                       }
-               } else {
-                       /* 2. Find cpu that have to spare */
-                       unsigned long new_util = task_util(p) + cpu_util_wake(cpu, p);
+                       if (!cpumask_test_cpu(i, tsk_cpus_allowed(p)))
+                               continue;
 
-                       if (new_util * 100 >=
-                                       capacity_orig_of(cpu) * get_coverage_ratio(cpu))
+                       if (cpu_rq(i)->ontime_migrating)
+                               continue;
+
+                       new_util = task_util(p) + cpu_util_wake(i, p);
+
+                       if (new_util * 100 >= coverage_util)
                                continue;
 
                        if (new_util < min_util) {
                                min_util = new_util;
-                               backup_cpu = cpu;
+                               min_util_cpu = i;
                        }
                }
+
+               if (cpu_selected(min_util_cpu)) {
+                       cpumask_set_cpu(min_util_cpu, &candidates);
+                       candidate_count++;
+               }
        }
 
-       rcu_read_unlock();
+       /*
+        * Second) Find min_energy_cpu among the candidates and return it.
+        */
+       if (candidate_count > 1) {
+               /*
+                * If there is more than one candidate,
+                * calculate each energy and choose min_energy_cpu.
+                */
+               unsigned int min_energy = UINT_MAX;
+
+               for_each_cpu(cpu, &candidates) {
+                       unsigned int new_energy = calculate_energy(p, cpu);
 
-       if (cpu_selected(best_cpu))
-               return best_cpu;
+                       if (min_energy > new_energy) {
+                               min_energy = new_energy;
+                               min_energy_cpu = cpu;
+                       }
+               }
+       } else if (candidate_count == 1) {
+               /*
+                * If there is just one candidate, this will be min_energy_cpu.
+                */
+               min_energy_cpu = cpumask_first(&candidates);
+       }
 
-       return backup_cpu;
+       return min_energy_cpu;
 }
 
 extern struct sched_entity *__pick_next_entity(struct sched_entity *se);