sched: Extend sched_group_energy to test load-balancing decisions
authorMorten Rasmussen <morten.rasmussen@arm.com>
Fri, 2 Jan 2015 14:21:56 +0000 (14:21 +0000)
committerPunit Agrawal <punit.agrawal@arm.com>
Mon, 21 Mar 2016 12:34:30 +0000 (12:34 +0000)
Extended sched_group_energy() to support energy prediction with usage
(tasks) added/removed from a specific cpu or migrated between a pair of
cpus. Useful for load-balancing decision making.

cc: Ingo Molnar <mingo@redhat.com>
cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Morten Rasmussen <morten.rasmussen@arm.com>
kernel/sched/fair.c

index 49f7567c6b07457c0718b3a0bff7e93180a3c47f..f3de37414a664e69f3ab61003676ff27289f011c 100644 (file)
@@ -4688,12 +4688,21 @@ static unsigned long capacity_curr_of(int cpu)
  * capacity_orig) as it useful for predicting the capacity required after task
  * migrations (scheduler-driven DVFS).
  */
-static unsigned long cpu_util(int cpu)
+static unsigned long __cpu_util(int cpu, int delta)
 {
        unsigned long util = cpu_rq(cpu)->cfs.avg.util_avg;
        unsigned long capacity = capacity_orig_of(cpu);
 
-       return (util >= capacity) ? capacity : util;
+       delta += util;
+       if (delta < 0)
+               return 0;
+
+       return (delta >= capacity) ? capacity : delta;
+}
+
+static unsigned long cpu_util(int cpu)
+{
+       return __cpu_util(cpu, 0);
 }
 
 static inline bool energy_aware(void)
@@ -4701,8 +4710,18 @@ static inline bool energy_aware(void)
        return sched_feat(ENERGY_AWARE);
 }
 
+struct energy_env {
+       struct sched_group      *sg_top;
+       struct sched_group      *sg_cap;
+       int                     cap_idx;
+       int                     util_delta;
+       int                     src_cpu;
+       int                     dst_cpu;
+       int                     energy;
+};
+
 /*
- * cpu_norm_util() returns the cpu util relative to a specific capacity,
+ * __cpu_norm_util() returns the cpu util relative to a specific capacity,
  * i.e. it's busy ratio, in the range [0..SCHED_LOAD_SCALE] which is useful for
  * energy calculations. Using the scale-invariant util returned by
  * cpu_util() and approximating scale-invariant util by:
@@ -4715,9 +4734,9 @@ static inline bool energy_aware(void)
  *
  *   norm_util = running_time/time ~ util/capacity
  */
-static unsigned long cpu_norm_util(int cpu, unsigned long capacity)
+static unsigned long __cpu_norm_util(int cpu, unsigned long capacity, int delta)
 {
-       int util = cpu_util(cpu);
+       int util = __cpu_util(cpu, delta);
 
        if (util >= capacity)
                return SCHED_CAPACITY_SCALE;
@@ -4725,13 +4744,25 @@ static unsigned long cpu_norm_util(int cpu, unsigned long capacity)
        return (util << SCHED_CAPACITY_SHIFT)/capacity;
 }
 
-static unsigned long group_max_util(struct sched_group *sg)
+static int calc_util_delta(struct energy_env *eenv, int cpu)
 {
-       int i;
+       if (cpu == eenv->src_cpu)
+               return -eenv->util_delta;
+       if (cpu == eenv->dst_cpu)
+               return eenv->util_delta;
+       return 0;
+}
+
+static
+unsigned long group_max_util(struct energy_env *eenv)
+{
+       int i, delta;
        unsigned long max_util = 0;
 
-       for_each_cpu(i, sched_group_cpus(sg))
-               max_util = max(max_util, cpu_util(i));
+       for_each_cpu(i, sched_group_cpus(eenv->sg_cap)) {
+               delta = calc_util_delta(eenv, i);
+               max_util = max(max_util, __cpu_util(i, delta));
+       }
 
        return max_util;
 }
@@ -4745,31 +4776,36 @@ static unsigned long group_max_util(struct sched_group *sg)
  * latter is used as the estimate as it leads to a more pessimistic energy
  * estimate (more busy).
  */
-static unsigned long group_norm_util(struct sched_group *sg, int cap_idx)
+static unsigned
+long group_norm_util(struct energy_env *eenv, struct sched_group *sg)
 {
-       int i;
+       int i, delta;
        unsigned long util_sum = 0;
-       unsigned long capacity = sg->sge->cap_states[cap_idx].cap;
+       unsigned long capacity = sg->sge->cap_states[eenv->cap_idx].cap;
 
-       for_each_cpu(i, sched_group_cpus(sg))
-               util_sum += cpu_norm_util(i, capacity);
+       for_each_cpu(i, sched_group_cpus(sg)) {
+               delta = calc_util_delta(eenv, i);
+               util_sum += __cpu_norm_util(i, capacity, delta);
+       }
 
        if (util_sum > SCHED_CAPACITY_SCALE)
                return SCHED_CAPACITY_SCALE;
        return util_sum;
 }
 
-static int find_new_capacity(struct sched_group *sg,
+static int find_new_capacity(struct energy_env *eenv,
        const struct sched_group_energy const *sge)
 {
        int idx;
-       unsigned long util = group_max_util(sg);
+       unsigned long util = group_max_util(eenv);
 
        for (idx = 0; idx < sge->nr_cap_states; idx++) {
                if (sge->cap_states[idx].cap >= util)
-                       return idx;
+                       break;
        }
 
+       eenv->cap_idx = idx;
+
        return idx;
 }
 
@@ -4783,16 +4819,16 @@ static int find_new_capacity(struct sched_group *sg,
  * This can probably be done in a faster but more complex way.
  * Note: sched_group_energy() may fail when racing with sched_domain updates.
  */
-static int sched_group_energy(struct sched_group *sg_top)
+static int sched_group_energy(struct energy_env *eenv)
 {
        struct sched_domain *sd;
        int cpu, total_energy = 0;
        struct cpumask visit_cpus;
        struct sched_group *sg;
 
-       WARN_ON(!sg_top->sge);
+       WARN_ON(!eenv->sg_top->sge);
 
-       cpumask_copy(&visit_cpus, sched_group_cpus(sg_top));
+       cpumask_copy(&visit_cpus, sched_group_cpus(eenv->sg_top));
 
        while (!cpumask_empty(&visit_cpus)) {
                struct sched_group *sg_shared_cap = NULL;
@@ -4824,17 +4860,16 @@ static int sched_group_energy(struct sched_group *sg_top)
                                break;
 
                        do {
-                               struct sched_group *sg_cap_util;
                                unsigned long group_util;
                                int sg_busy_energy, sg_idle_energy, cap_idx;
 
                                if (sg_shared_cap && sg_shared_cap->group_weight >= sg->group_weight)
-                                       sg_cap_util = sg_shared_cap;
+                                       eenv->sg_cap = sg_shared_cap;
                                else
-                                       sg_cap_util = sg;
+                                       eenv->sg_cap = sg;
 
-                               cap_idx = find_new_capacity(sg_cap_util, sg->sge);
-                               group_util = group_norm_util(sg, cap_idx);
+                               cap_idx = find_new_capacity(eenv, sg->sge);
+                               group_util = group_norm_util(eenv, sg);
                                sg_busy_energy = (group_util * sg->sge->cap_states[cap_idx].power)
                                                                                >> SCHED_CAPACITY_SHIFT;
                                sg_idle_energy = ((SCHED_LOAD_SCALE-group_util) * sg->sge->idle_states[0].power)
@@ -4845,7 +4880,7 @@ static int sched_group_energy(struct sched_group *sg_top)
                                if (!sd->child)
                                        cpumask_xor(&visit_cpus, &visit_cpus, sched_group_cpus(sg));
 
-                               if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(sg_top)))
+                               if (cpumask_equal(sched_group_cpus(sg), sched_group_cpus(eenv->sg_top)))
                                        goto next_cpu;
 
                        } while (sg = sg->next, sg != sd->groups);
@@ -4854,7 +4889,8 @@ next_cpu:
                continue;
        }
 
-       return total_energy;
+       eenv->energy = total_energy;
+       return 0;
 }
 
 /*