Merge branch 'linus' into sched/core
[linux-2.6.git] / kernel / sched_fair.c
index 8b3eddb..e7d67a9 100644 (file)
  */
 
 #include <linux/latencytop.h>
+#include <linux/sched.h>
+#include <linux/cpumask.h>
 
 /*
  * Targeted preemption latency for CPU-bound tasks:
- * (default: 5ms * (1 + ilog(ncpus)), units: nanoseconds)
+ * (default: 6ms * (1 + ilog(ncpus)), units: nanoseconds)
  *
  * NOTE: this latency value is not the same as the concept of
  * 'timeslice length' - timeslices in CFS are of variable length
  * (to see the precise effective timeslice length of your workload,
  *  run vmstat and monitor the context-switches (cs) field)
  */
-unsigned int sysctl_sched_latency = 5000000ULL;
+unsigned int sysctl_sched_latency = 6000000ULL;
+unsigned int normalized_sysctl_sched_latency = 6000000ULL;
+
+/*
+ * The initial- and re-scaling of tunables is configurable
+ * (default SCHED_TUNABLESCALING_LOG = *(1+ilog(ncpus))
+ *
+ * Options are:
+ * SCHED_TUNABLESCALING_NONE - unscaled, always *1
+ * SCHED_TUNABLESCALING_LOG - scaled logarithmical, *1+ilog(ncpus)
+ * SCHED_TUNABLESCALING_LINEAR - scaled linear, *ncpus
+ */
+enum sched_tunable_scaling sysctl_sched_tunable_scaling
+       = SCHED_TUNABLESCALING_LOG;
 
 /*
  * Minimal preemption granularity for CPU-bound tasks:
- * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
+ * (default: 0.75 msec * (1 + ilog(ncpus)), units: nanoseconds)
  */
-unsigned int sysctl_sched_min_granularity = 1000000ULL;
+unsigned int sysctl_sched_min_granularity = 750000ULL;
+unsigned int normalized_sysctl_sched_min_granularity = 750000ULL;
 
 /*
  * is kept at sysctl_sched_latency / sysctl_sched_min_granularity
  */
-static unsigned int sched_nr_latency = 5;
+static unsigned int sched_nr_latency = 8;
 
 /*
  * After fork, child runs first. If set to 0 (default) then
@@ -54,14 +70,6 @@ static unsigned int sched_nr_latency = 5;
 unsigned int sysctl_sched_child_runs_first __read_mostly;
 
 /*
- * sys_sched_yield() compat mode
- *
- * This option switches the agressive yield implementation of the
- * old scheduler back on.
- */
-unsigned int __read_mostly sysctl_sched_compat_yield;
-
-/*
  * SCHED_OTHER wake-up granularity.
  * (default: 1 msec * (1 + ilog(ncpus)), units: nanoseconds)
  *
@@ -70,9 +78,17 @@ unsigned int __read_mostly sysctl_sched_compat_yield;
  * have immediate wakeup/sleep latencies.
  */
 unsigned int sysctl_sched_wakeup_granularity = 1000000UL;
+unsigned int normalized_sysctl_sched_wakeup_granularity = 1000000UL;
 
 const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
 
+/*
+ * The exponential sliding  window over which load is averaged for shares
+ * distribution.
+ * (default: 10msec)
+ */
+unsigned int __read_mostly sysctl_sched_shares_window = 10000000UL;
+
 static const struct sched_class fair_sched_class;
 
 /**************************************************************
@@ -127,6 +143,36 @@ static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
        return cfs_rq->tg->cfs_rq[this_cpu];
 }
 
+static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
+{
+       if (!cfs_rq->on_list) {
+               /*
+                * Ensure we either appear before our parent (if already
+                * enqueued) or force our parent to appear after us when it is
+                * enqueued.  The fact that we always enqueue bottom-up
+                * reduces this to two cases.
+                */
+               if (cfs_rq->tg->parent &&
+                   cfs_rq->tg->parent->cfs_rq[cpu_of(rq_of(cfs_rq))]->on_list) {
+                       list_add_rcu(&cfs_rq->leaf_cfs_rq_list,
+                               &rq_of(cfs_rq)->leaf_cfs_rq_list);
+               } else {
+                       list_add_tail_rcu(&cfs_rq->leaf_cfs_rq_list,
+                               &rq_of(cfs_rq)->leaf_cfs_rq_list);
+               }
+
+               cfs_rq->on_list = 1;
+       }
+}
+
+static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
+{
+       if (cfs_rq->on_list) {
+               list_del_rcu(&cfs_rq->leaf_cfs_rq_list);
+               cfs_rq->on_list = 0;
+       }
+}
+
 /* Iterate thr' all leaf cfs_rq's on a runqueue */
 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
        list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
@@ -230,6 +276,14 @@ static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
        return &cpu_rq(this_cpu)->cfs;
 }
 
+static inline void list_add_leaf_cfs_rq(struct cfs_rq *cfs_rq)
+{
+}
+
+static inline void list_del_leaf_cfs_rq(struct cfs_rq *cfs_rq)
+{
+}
+
 #define for_each_leaf_cfs_rq(rq, cfs_rq) \
                for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
 
@@ -304,6 +358,10 @@ static void update_min_vruntime(struct cfs_rq *cfs_rq)
        }
 
        cfs_rq->min_vruntime = max_vruntime(cfs_rq->min_vruntime, vruntime);
+#ifndef CONFIG_64BIT
+       smp_wmb();
+       cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
+#endif
 }
 
 /*
@@ -358,7 +416,7 @@ static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
        rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
 }
 
-static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
+static struct sched_entity *__pick_first_entity(struct cfs_rq *cfs_rq)
 {
        struct rb_node *left = cfs_rq->rb_leftmost;
 
@@ -368,6 +426,17 @@ static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
        return rb_entry(left, struct sched_entity, run_node);
 }
 
+static struct sched_entity *__pick_next_entity(struct sched_entity *se)
+{
+       struct rb_node *next = rb_next(&se->run_node);
+
+       if (!next)
+               return NULL;
+
+       return rb_entry(next, struct sched_entity, run_node);
+}
+
+#ifdef CONFIG_SCHED_DEBUG
 static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
 {
        struct rb_node *last = rb_last(&cfs_rq->tasks_timeline);
@@ -382,12 +451,12 @@ static struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
  * Scheduling class statistics methods:
  */
 
-#ifdef CONFIG_SCHED_DEBUG
-int sched_nr_latency_handler(struct ctl_table *table, int write,
-               struct file *filp, void __user *buffer, size_t *lenp,
+int sched_proc_update_handler(struct ctl_table *table, int write,
+               void __user *buffer, size_t *lenp,
                loff_t *ppos)
 {
-       int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
+       int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
+       int factor = get_update_sysctl_factor();
 
        if (ret || !write)
                return ret;
@@ -395,6 +464,13 @@ int sched_nr_latency_handler(struct ctl_table *table, int write,
        sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
                                        sysctl_sched_min_granularity);
 
+#define WRT_SYSCTL(name) \
+       (normalized_sysctl_##name = sysctl_##name / (factor))
+       WRT_SYSCTL(sched_min_granularity);
+       WRT_SYSCTL(sched_latency);
+       WRT_SYSCTL(sched_wakeup_granularity);
+#undef WRT_SYSCTL
+
        return 0;
 }
 #endif
@@ -470,6 +546,9 @@ static u64 sched_vslice(struct cfs_rq *cfs_rq, struct sched_entity *se)
        return calc_delta_fair(sched_slice(cfs_rq, se), se);
 }
 
+static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update);
+static void update_cfs_shares(struct cfs_rq *cfs_rq);
+
 /*
  * Update the current task's runtime statistics. Skip current tasks that
  * are not in our scheduling class.
@@ -480,19 +559,25 @@ __update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
 {
        unsigned long delta_exec_weighted;
 
-       schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
+       schedstat_set(curr->statistics.exec_max,
+                     max((u64)delta_exec, curr->statistics.exec_max));
 
        curr->sum_exec_runtime += delta_exec;
        schedstat_add(cfs_rq, exec_clock, delta_exec);
        delta_exec_weighted = calc_delta_fair(delta_exec, curr);
+
        curr->vruntime += delta_exec_weighted;
        update_min_vruntime(cfs_rq);
+
+#if defined CONFIG_SMP && defined CONFIG_FAIR_GROUP_SCHED
+       cfs_rq->load_unacc_exec_time += delta_exec;
+#endif
 }
 
 static void update_curr(struct cfs_rq *cfs_rq)
 {
        struct sched_entity *curr = cfs_rq->curr;
-       u64 now = rq_of(cfs_rq)->clock;
+       u64 now = rq_of(cfs_rq)->clock_task;
        unsigned long delta_exec;
 
        if (unlikely(!curr))
@@ -513,6 +598,7 @@ static void update_curr(struct cfs_rq *cfs_rq)
        if (entity_is_task(curr)) {
                struct task_struct *curtask = task_of(curr);
 
+               trace_sched_stat_runtime(curtask, delta_exec, curr->vruntime);
                cpuacct_charge(curtask, delta_exec);
                account_group_exec_runtime(curtask, delta_exec);
        }
@@ -521,7 +607,7 @@ static void update_curr(struct cfs_rq *cfs_rq)
 static inline void
 update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
 {
-       schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
+       schedstat_set(se->statistics.wait_start, rq_of(cfs_rq)->clock);
 }
 
 /*
@@ -540,18 +626,18 @@ static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
 static void
 update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
 {
-       schedstat_set(se->wait_max, max(se->wait_max,
-                       rq_of(cfs_rq)->clock - se->wait_start));
-       schedstat_set(se->wait_count, se->wait_count + 1);
-       schedstat_set(se->wait_sum, se->wait_sum +
-                       rq_of(cfs_rq)->clock - se->wait_start);
+       schedstat_set(se->statistics.wait_max, max(se->statistics.wait_max,
+                       rq_of(cfs_rq)->clock - se->statistics.wait_start));
+       schedstat_set(se->statistics.wait_count, se->statistics.wait_count + 1);
+       schedstat_set(se->statistics.wait_sum, se->statistics.wait_sum +
+                       rq_of(cfs_rq)->clock - se->statistics.wait_start);
 #ifdef CONFIG_SCHEDSTATS
        if (entity_is_task(se)) {
                trace_sched_stat_wait(task_of(se),
-                       rq_of(cfs_rq)->clock - se->wait_start);
+                       rq_of(cfs_rq)->clock - se->statistics.wait_start);
        }
 #endif
-       schedstat_set(se->wait_start, 0);
+       schedstat_set(se->statistics.wait_start, 0);
 }
 
 static inline void
@@ -574,7 +660,7 @@ update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
        /*
         * We are starting a new run period:
         */
-       se->exec_start = rq_of(cfs_rq)->clock;
+       se->exec_start = rq_of(cfs_rq)->clock_task;
 }
 
 /**************************************************
@@ -605,7 +691,6 @@ account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
                list_add(&se->group_node, &cfs_rq->tasks);
        }
        cfs_rq->nr_running++;
-       se->on_rq = 1;
 }
 
 static void
@@ -619,9 +704,164 @@ account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
                list_del_init(&se->group_node);
        }
        cfs_rq->nr_running--;
-       se->on_rq = 0;
 }
 
+#ifdef CONFIG_FAIR_GROUP_SCHED
+# ifdef CONFIG_SMP
+static void update_cfs_rq_load_contribution(struct cfs_rq *cfs_rq,
+                                           int global_update)
+{
+       struct task_group *tg = cfs_rq->tg;
+       long load_avg;
+
+       load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
+       load_avg -= cfs_rq->load_contribution;
+
+       if (global_update || abs(load_avg) > cfs_rq->load_contribution / 8) {
+               atomic_add(load_avg, &tg->load_weight);
+               cfs_rq->load_contribution += load_avg;
+       }
+}
+
+static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
+{
+       u64 period = sysctl_sched_shares_window;
+       u64 now, delta;
+       unsigned long load = cfs_rq->load.weight;
+
+       if (cfs_rq->tg == &root_task_group)
+               return;
+
+       now = rq_of(cfs_rq)->clock_task;
+       delta = now - cfs_rq->load_stamp;
+
+       /* truncate load history at 4 idle periods */
+       if (cfs_rq->load_stamp > cfs_rq->load_last &&
+           now - cfs_rq->load_last > 4 * period) {
+               cfs_rq->load_period = 0;
+               cfs_rq->load_avg = 0;
+               delta = period - 1;
+       }
+
+       cfs_rq->load_stamp = now;
+       cfs_rq->load_unacc_exec_time = 0;
+       cfs_rq->load_period += delta;
+       if (load) {
+               cfs_rq->load_last = now;
+               cfs_rq->load_avg += delta * load;
+       }
+
+       /* consider updating load contribution on each fold or truncate */
+       if (global_update || cfs_rq->load_period > period
+           || !cfs_rq->load_period)
+               update_cfs_rq_load_contribution(cfs_rq, global_update);
+
+       while (cfs_rq->load_period > period) {
+               /*
+                * Inline assembly required to prevent the compiler
+                * optimising this loop into a divmod call.
+                * See __iter_div_u64_rem() for another example of this.
+                */
+               asm("" : "+rm" (cfs_rq->load_period));
+               cfs_rq->load_period /= 2;
+               cfs_rq->load_avg /= 2;
+       }
+
+       if (!cfs_rq->curr && !cfs_rq->nr_running && !cfs_rq->load_avg)
+               list_del_leaf_cfs_rq(cfs_rq);
+}
+
+static long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
+{
+       long load_weight, load, shares;
+
+       load = cfs_rq->load.weight;
+
+       load_weight = atomic_read(&tg->load_weight);
+       load_weight += load;
+       load_weight -= cfs_rq->load_contribution;
+
+       shares = (tg->shares * load);
+       if (load_weight)
+               shares /= load_weight;
+
+       if (shares < MIN_SHARES)
+               shares = MIN_SHARES;
+       if (shares > tg->shares)
+               shares = tg->shares;
+
+       return shares;
+}
+
+static void update_entity_shares_tick(struct cfs_rq *cfs_rq)
+{
+       if (cfs_rq->load_unacc_exec_time > sysctl_sched_shares_window) {
+               update_cfs_load(cfs_rq, 0);
+               update_cfs_shares(cfs_rq);
+       }
+}
+# else /* CONFIG_SMP */
+static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
+{
+}
+
+static inline long calc_cfs_shares(struct cfs_rq *cfs_rq, struct task_group *tg)
+{
+       return tg->shares;
+}
+
+static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
+{
+}
+# endif /* CONFIG_SMP */
+static void reweight_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
+                           unsigned long weight)
+{
+       if (se->on_rq) {
+               /* commit outstanding execution time */
+               if (cfs_rq->curr == se)
+                       update_curr(cfs_rq);
+               account_entity_dequeue(cfs_rq, se);
+       }
+
+       update_load_set(&se->load, weight);
+
+       if (se->on_rq)
+               account_entity_enqueue(cfs_rq, se);
+}
+
+static void update_cfs_shares(struct cfs_rq *cfs_rq)
+{
+       struct task_group *tg;
+       struct sched_entity *se;
+       long shares;
+
+       tg = cfs_rq->tg;
+       se = tg->se[cpu_of(rq_of(cfs_rq))];
+       if (!se)
+               return;
+#ifndef CONFIG_SMP
+       if (likely(se->load.weight == tg->shares))
+               return;
+#endif
+       shares = calc_cfs_shares(cfs_rq, tg);
+
+       reweight_entity(cfs_rq_of(se), se, shares);
+}
+#else /* CONFIG_FAIR_GROUP_SCHED */
+static void update_cfs_load(struct cfs_rq *cfs_rq, int global_update)
+{
+}
+
+static inline void update_cfs_shares(struct cfs_rq *cfs_rq)
+{
+}
+
+static inline void update_entity_shares_tick(struct cfs_rq *cfs_rq)
+{
+}
+#endif /* CONFIG_FAIR_GROUP_SCHED */
+
 static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
 {
 #ifdef CONFIG_SCHEDSTATS
@@ -630,39 +870,39 @@ static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
        if (entity_is_task(se))
                tsk = task_of(se);
 
-       if (se->sleep_start) {
-               u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
+       if (se->statistics.sleep_start) {
+               u64 delta = rq_of(cfs_rq)->clock - se->statistics.sleep_start;
 
                if ((s64)delta < 0)
                        delta = 0;
 
-               if (unlikely(delta > se->sleep_max))
-                       se->sleep_max = delta;
+               if (unlikely(delta > se->statistics.sleep_max))
+                       se->statistics.sleep_max = delta;
 
-               se->sleep_start = 0;
-               se->sum_sleep_runtime += delta;
+               se->statistics.sleep_start = 0;
+               se->statistics.sum_sleep_runtime += delta;
 
                if (tsk) {
                        account_scheduler_latency(tsk, delta >> 10, 1);
                        trace_sched_stat_sleep(tsk, delta);
                }
        }
-       if (se->block_start) {
-               u64 delta = rq_of(cfs_rq)->clock - se->block_start;
+       if (se->statistics.block_start) {
+               u64 delta = rq_of(cfs_rq)->clock - se->statistics.block_start;
 
                if ((s64)delta < 0)
                        delta = 0;
 
-               if (unlikely(delta > se->block_max))
-                       se->block_max = delta;
+               if (unlikely(delta > se->statistics.block_max))
+                       se->statistics.block_max = delta;
 
-               se->block_start = 0;
-               se->sum_sleep_runtime += delta;
+               se->statistics.block_start = 0;
+               se->statistics.sum_sleep_runtime += delta;
 
                if (tsk) {
                        if (tsk->in_iowait) {
-                               se->iowait_sum += delta;
-                               se->iowait_count++;
+                               se->statistics.iowait_sum += delta;
+                               se->statistics.iowait_count++;
                                trace_sched_stat_iowait(tsk, delta);
                        }
 
@@ -709,24 +949,18 @@ place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
        if (initial && sched_feat(START_DEBIT))
                vruntime += sched_vslice(cfs_rq, se);
 
+       /* sleeps up to a single latency don't count. */
        if (!initial) {
-               /* sleeps upto a single latency don't count. */
-               if (sched_feat(NEW_FAIR_SLEEPERS)) {
-                       unsigned long thresh = sysctl_sched_latency;
+               unsigned long thresh = sysctl_sched_latency;
 
-                       /*
-                        * Convert the sleeper threshold into virtual time.
-                        * SCHED_IDLE is a special sub-class.  We care about
-                        * fairness only relative to other SCHED_IDLE tasks,
-                        * all of which have the same weight.
-                        */
-                       if (sched_feat(NORMALIZED_SLEEPER) &&
-                                       (!entity_is_task(se) ||
-                                        task_of(se)->policy != SCHED_IDLE))
-                               thresh = calc_delta_fair(thresh, se);
+               /*
+                * Halve their sleep time's effect, to allow
+                * for a gentler effect of sleepers:
+                */
+               if (sched_feat(GENTLE_FAIR_SLEEPERS))
+                       thresh >>= 1;
 
-                       vruntime -= thresh;
-               }
+               vruntime -= thresh;
        }
 
        /* ensure we never gain time by being placed backwards. */
@@ -736,15 +970,24 @@ place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
 }
 
 static void
-enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
+enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
 {
        /*
+        * Update the normalized vruntime before updating min_vruntime
+        * through callig update_curr().
+        */
+       if (!(flags & ENQUEUE_WAKEUP) || (flags & ENQUEUE_WAKING))
+               se->vruntime += cfs_rq->min_vruntime;
+
+       /*
         * Update run-time statistics of the 'current'.
         */
        update_curr(cfs_rq);
+       update_cfs_load(cfs_rq, 0);
        account_entity_enqueue(cfs_rq, se);
+       update_cfs_shares(cfs_rq);
 
-       if (wakeup) {
+       if (flags & ENQUEUE_WAKEUP) {
                place_entity(cfs_rq, se, 0);
                enqueue_sleeper(cfs_rq, se);
        }
@@ -753,25 +996,59 @@ enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
        check_spread(cfs_rq, se);
        if (se != cfs_rq->curr)
                __enqueue_entity(cfs_rq, se);
+       se->on_rq = 1;
+
+       if (cfs_rq->nr_running == 1)
+               list_add_leaf_cfs_rq(cfs_rq);
 }
 
-static void __clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
+static void __clear_buddies_last(struct sched_entity *se)
 {
-       if (cfs_rq->last == se)
-               cfs_rq->last = NULL;
+       for_each_sched_entity(se) {
+               struct cfs_rq *cfs_rq = cfs_rq_of(se);
+               if (cfs_rq->last == se)
+                       cfs_rq->last = NULL;
+               else
+                       break;
+       }
+}
 
-       if (cfs_rq->next == se)
-               cfs_rq->next = NULL;
+static void __clear_buddies_next(struct sched_entity *se)
+{
+       for_each_sched_entity(se) {
+               struct cfs_rq *cfs_rq = cfs_rq_of(se);
+               if (cfs_rq->next == se)
+                       cfs_rq->next = NULL;
+               else
+                       break;
+       }
+}
+
+static void __clear_buddies_skip(struct sched_entity *se)
+{
+       for_each_sched_entity(se) {
+               struct cfs_rq *cfs_rq = cfs_rq_of(se);
+               if (cfs_rq->skip == se)
+                       cfs_rq->skip = NULL;
+               else
+                       break;
+       }
 }
 
 static void clear_buddies(struct cfs_rq *cfs_rq, struct sched_entity *se)
 {
-       for_each_sched_entity(se)
-               __clear_buddies(cfs_rq_of(se), se);
+       if (cfs_rq->last == se)
+               __clear_buddies_last(se);
+
+       if (cfs_rq->next == se)
+               __clear_buddies_next(se);
+
+       if (cfs_rq->skip == se)
+               __clear_buddies_skip(se);
 }
 
 static void
-dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
+dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int flags)
 {
        /*
         * Update run-time statistics of the 'current'.
@@ -779,15 +1056,15 @@ dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
        update_curr(cfs_rq);
 
        update_stats_dequeue(cfs_rq, se);
-       if (sleep) {
+       if (flags & DEQUEUE_SLEEP) {
 #ifdef CONFIG_SCHEDSTATS
                if (entity_is_task(se)) {
                        struct task_struct *tsk = task_of(se);
 
                        if (tsk->state & TASK_INTERRUPTIBLE)
-                               se->sleep_start = rq_of(cfs_rq)->clock;
+                               se->statistics.sleep_start = rq_of(cfs_rq)->clock;
                        if (tsk->state & TASK_UNINTERRUPTIBLE)
-                               se->block_start = rq_of(cfs_rq)->clock;
+                               se->statistics.block_start = rq_of(cfs_rq)->clock;
                }
 #endif
        }
@@ -796,8 +1073,20 @@ dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
 
        if (se != cfs_rq->curr)
                __dequeue_entity(cfs_rq, se);
+       se->on_rq = 0;
+       update_cfs_load(cfs_rq, 0);
        account_entity_dequeue(cfs_rq, se);
+
+       /*
+        * Normalize the entity after updating the min_vruntime because the
+        * update can refer to the ->curr item and we need to reflect this
+        * movement in our normalized position.
+        */
+       if (!(flags & DEQUEUE_SLEEP))
+               se->vruntime -= cfs_rq->min_vruntime;
+
        update_min_vruntime(cfs_rq);
+       update_cfs_shares(cfs_rq);
 }
 
 /*
@@ -817,6 +1106,29 @@ check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
                 * re-elected due to buddy favours.
                 */
                clear_buddies(cfs_rq, curr);
+               return;
+       }
+
+       /*
+        * Ensure that a task that missed wakeup preemption by a
+        * narrow margin doesn't have to wait for a full slice.
+        * This also mitigates buddy induced latencies under load.
+        */
+       if (!sched_feat(WAKEUP_PREEMPT))
+               return;
+
+       if (delta_exec < sysctl_sched_min_granularity)
+               return;
+
+       if (cfs_rq->nr_running > 1) {
+               struct sched_entity *se = __pick_first_entity(cfs_rq);
+               s64 delta = curr->vruntime - se->vruntime;
+
+               if (delta < 0)
+                       return;
+
+               if (delta > ideal_runtime)
+                       resched_task(rq_of(cfs_rq)->curr);
        }
 }
 
@@ -843,7 +1155,7 @@ set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
         * when there are only lesser-weight tasks around):
         */
        if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
-               se->slice_max = max(se->slice_max,
+               se->statistics.slice_max = max(se->statistics.slice_max,
                        se->sum_exec_runtime - se->prev_sum_exec_runtime);
        }
 #endif
@@ -853,15 +1165,41 @@ set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
 static int
 wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se);
 
+/*
+ * Pick the next process, keeping these things in mind, in this order:
+ * 1) keep things fair between processes/task groups
+ * 2) pick the "next" process, since someone really wants that to run
+ * 3) pick the "last" process, for cache locality
+ * 4) do not run the "skip" process, if something else is available
+ */
 static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
 {
-       struct sched_entity *se = __pick_next_entity(cfs_rq);
+       struct sched_entity *se = __pick_first_entity(cfs_rq);
+       struct sched_entity *left = se;
+
+       /*
+        * Avoid running the skip buddy, if running something else can
+        * be done without getting too unfair.
+        */
+       if (cfs_rq->skip == se) {
+               struct sched_entity *second = __pick_next_entity(se);
+               if (second && wakeup_preempt_entity(second, left) < 1)
+                       se = second;
+       }
+
+       /*
+        * Prefer last buddy, try to return the CPU to a preempted task.
+        */
+       if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, left) < 1)
+               se = cfs_rq->last;
 
-       if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, se) < 1)
-               return cfs_rq->next;
+       /*
+        * Someone really wants this to run. If it's not unfair, run it.
+        */
+       if (cfs_rq->next && wakeup_preempt_entity(cfs_rq->next, left) < 1)
+               se = cfs_rq->next;
 
-       if (cfs_rq->last && wakeup_preempt_entity(cfs_rq->last, se) < 1)
-               return cfs_rq->last;
+       clear_buddies(cfs_rq, se);
 
        return se;
 }
@@ -892,6 +1230,11 @@ entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
         */
        update_curr(cfs_rq);
 
+       /*
+        * Update share accounting for long-running entities.
+        */
+       update_entity_shares_tick(cfs_rq);
+
 #ifdef CONFIG_SCHED_HRTICK
        /*
         * queued ticks are scheduled to match the slice, so don't bother
@@ -978,7 +1321,8 @@ static inline void hrtick_update(struct rq *rq)
  * increased. Here we update the fair scheduling stats and
  * then put the task into the rbtree:
  */
-static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
+static void
+enqueue_task_fair(struct rq *rq, struct task_struct *p, int flags)
 {
        struct cfs_rq *cfs_rq;
        struct sched_entity *se = &p->se;
@@ -987,83 +1331,83 @@ static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
                if (se->on_rq)
                        break;
                cfs_rq = cfs_rq_of(se);
-               enqueue_entity(cfs_rq, se, wakeup);
-               wakeup = 1;
+               enqueue_entity(cfs_rq, se, flags);
+               flags = ENQUEUE_WAKEUP;
+       }
+
+       for_each_sched_entity(se) {
+               struct cfs_rq *cfs_rq = cfs_rq_of(se);
+
+               update_cfs_load(cfs_rq, 0);
+               update_cfs_shares(cfs_rq);
        }
 
        hrtick_update(rq);
 }
 
+static void set_next_buddy(struct sched_entity *se);
+
 /*
  * The dequeue_task method is called before nr_running is
  * decreased. We remove the task from the rbtree and
  * update the fair scheduling stats:
  */
-static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
+static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int flags)
 {
        struct cfs_rq *cfs_rq;
        struct sched_entity *se = &p->se;
+       int task_sleep = flags & DEQUEUE_SLEEP;
 
        for_each_sched_entity(se) {
                cfs_rq = cfs_rq_of(se);
-               dequeue_entity(cfs_rq, se, sleep);
+               dequeue_entity(cfs_rq, se, flags);
+
                /* Don't dequeue parent if it has other entities besides us */
-               if (cfs_rq->load.weight)
+               if (cfs_rq->load.weight) {
+                       /*
+                        * Bias pick_next to pick a task from this cfs_rq, as
+                        * p is sleeping when it is within its sched_slice.
+                        */
+                       if (task_sleep && parent_entity(se))
+                               set_next_buddy(parent_entity(se));
                        break;
-               sleep = 1;
+               }
+               flags |= DEQUEUE_SLEEP;
+       }
+
+       for_each_sched_entity(se) {
+               struct cfs_rq *cfs_rq = cfs_rq_of(se);
+
+               update_cfs_load(cfs_rq, 0);
+               update_cfs_shares(cfs_rq);
        }
 
        hrtick_update(rq);
 }
 
-/*
- * sched_yield() support is very simple - we dequeue and enqueue.
- *
- * If compat_yield is turned on then we requeue to the end of the tree.
- */
-static void yield_task_fair(struct rq *rq)
-{
-       struct task_struct *curr = rq->curr;
-       struct cfs_rq *cfs_rq = task_cfs_rq(curr);
-       struct sched_entity *rightmost, *se = &curr->se;
-
-       /*
-        * Are we the only task in the tree?
-        */
-       if (unlikely(cfs_rq->nr_running == 1))
-               return;
+#ifdef CONFIG_SMP
 
-       clear_buddies(cfs_rq, se);
+static void task_waking_fair(struct task_struct *p)
+{
+       struct sched_entity *se = &p->se;
+       struct cfs_rq *cfs_rq = cfs_rq_of(se);
+       u64 min_vruntime;
 
-       if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
-               update_rq_clock(rq);
-               /*
-                * Update run-time statistics of the 'current'.
-                */
-               update_curr(cfs_rq);
+#ifndef CONFIG_64BIT
+       u64 min_vruntime_copy;
 
-               return;
-       }
-       /*
-        * Find the rightmost entry in the rbtree:
-        */
-       rightmost = __pick_last_entity(cfs_rq);
-       /*
-        * Already in the rightmost position?
-        */
-       if (unlikely(!rightmost || entity_before(rightmost, se)))
-               return;
+       do {
+               min_vruntime_copy = cfs_rq->min_vruntime_copy;
+               smp_rmb();
+               min_vruntime = cfs_rq->min_vruntime;
+       } while (min_vruntime != min_vruntime_copy);
+#else
+       min_vruntime = cfs_rq->min_vruntime;
+#endif
 
-       /*
-        * Minimally necessary key value to be last in the tree:
-        * Upon rescheduling, sched_class::put_prev_task() will place
-        * 'current' within the tree based on its new key value.
-        */
-       se->vruntime = rightmost->vruntime + 1;
+       se->vruntime -= min_vruntime;
 }
 
-#ifdef CONFIG_SMP
-
 #ifdef CONFIG_FAIR_GROUP_SCHED
 /*
  * effective_load() calculates the load change as seen from the root_task_group
@@ -1071,67 +1415,36 @@ static void yield_task_fair(struct rq *rq)
  * Adding load to a group doesn't make a group heavier, but can cause movement
  * of group shares between cpus. Assuming the shares were perfectly aligned one
  * can calculate the shift in shares.
- *
- * The problem is that perfectly aligning the shares is rather expensive, hence
- * we try to avoid doing that too often - see update_shares(), which ratelimits
- * this change.
- *
- * We compensate this by not only taking the current delta into account, but
- * also considering the delta between when the shares were last adjusted and
- * now.
- *
- * We still saw a performance dip, some tracing learned us that between
- * cgroup:/ and cgroup:/foo balancing the number of affine wakeups increased
- * significantly. Therefore try to bias the error in direction of failing
- * the affine wakeup.
- *
  */
-static long effective_load(struct task_group *tg, int cpu,
-               long wl, long wg)
+static long effective_load(struct task_group *tg, int cpu, long wl, long wg)
 {
        struct sched_entity *se = tg->se[cpu];
 
        if (!tg->parent)
                return wl;
 
-       /*
-        * By not taking the decrease of shares on the other cpu into
-        * account our error leans towards reducing the affine wakeups.
-        */
-       if (!wl && sched_feat(ASYM_EFF_LOAD))
-               return wl;
-
        for_each_sched_entity(se) {
-               long S, rw, s, a, b;
-               long more_w;
+               long lw, w;
 
-               /*
-                * Instead of using this increment, also add the difference
-                * between when the shares were last updated and now.
-                */
-               more_w = se->my_q->load.weight - se->my_q->rq_weight;
-               wl += more_w;
-               wg += more_w;
-
-               S = se->my_q->tg->shares;
-               s = se->my_q->shares;
-               rw = se->my_q->rq_weight;
+               tg = se->my_q->tg;
+               w = se->my_q->load.weight;
 
-               a = S*(rw + wl);
-               b = S*rw + s*wg;
+               /* use this cpu's instantaneous contribution */
+               lw = atomic_read(&tg->load_weight);
+               lw -= se->my_q->load_contribution;
+               lw += w + wg;
 
-               wl = s*(a-b);
+               wl += w;
 
-               if (likely(b))
-                       wl /= b;
+               if (lw > 0 && wl < lw)
+                       wl = (wl * tg->shares) / lw;
+               else
+                       wl = tg->shares;
 
-               /*
-                * Assume the group is already running and will
-                * thus already be accounted for in the weight.
-                *
-                * That is, moving shares between CPUs, does not
-                * alter the group weight.
-                */
+               /* zero point is MIN_SHARES */
+               if (wl < MIN_SHARES)
+                       wl = MIN_SHARES;
+               wl -= se->load.weight;
                wg = 0;
        }
 
@@ -1150,11 +1463,9 @@ static inline unsigned long effective_load(struct task_group *tg, int cpu,
 
 static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
 {
-       struct task_struct *curr = current;
-       unsigned long this_load, load;
+       s64 this_load, load;
        int idx, this_cpu, prev_cpu;
        unsigned long tl_per_task;
-       unsigned int imbalance;
        struct task_group *tg;
        unsigned long weight;
        int balanced;
@@ -1165,10 +1476,6 @@ static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
        load      = source_load(prev_cpu, idx);
        this_load = target_load(this_cpu, idx);
 
-       if (sync && (curr->se.avg_overlap > sysctl_sched_migration_cost ||
-                       p->se.avg_overlap > sysctl_sched_migration_cost))
-               sync = 0;
-
        /*
         * If sync wakeup then subtract the (maximum possible)
         * effect of the currently running task from the load
@@ -1185,8 +1492,6 @@ static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
        tg = task_group(p);
        weight = p->se.load.weight;
 
-       imbalance = 100 + (sd->imbalance_pct - 100) / 2;
-
        /*
         * In low-load situations, where prev_cpu is idle and this_cpu is idle
         * due to the sync cause above having dropped this_load to 0, we'll
@@ -1196,9 +1501,21 @@ static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
         * Otherwise check if either cpus are near enough in load to allow this
         * task to be woken on this_cpu.
         */
-       balanced = !this_load ||
-               100*(this_load + effective_load(tg, this_cpu, weight, weight)) <=
-               imbalance*(load + effective_load(tg, prev_cpu, 0, weight));
+       if (this_load > 0) {
+               s64 this_eff_load, prev_eff_load;
+
+               this_eff_load = 100;
+               this_eff_load *= power_of(prev_cpu);
+               this_eff_load *= this_load +
+                       effective_load(tg, this_cpu, weight, weight);
+
+               prev_eff_load = 100 + (sd->imbalance_pct - 100) / 2;
+               prev_eff_load *= power_of(this_cpu);
+               prev_eff_load *= load + effective_load(tg, prev_cpu, 0, weight);
+
+               balanced = this_eff_load <= prev_eff_load;
+       } else
+               balanced = true;
 
        /*
         * If the currently running task will sleep within
@@ -1208,7 +1525,7 @@ static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
        if (sync && balanced)
                return 1;
 
-       schedstat_inc(p, se.nr_wakeups_affine_attempts);
+       schedstat_inc(p, se.statistics.nr_wakeups_affine_attempts);
        tl_per_task = cpu_avg_load_per_task(this_cpu);
 
        if (balanced ||
@@ -1220,7 +1537,7 @@ static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
                 * there is no bad imbalance.
                 */
                schedstat_inc(sd, ttwu_move_affine);
-               schedstat_inc(p, se.nr_wakeups_affine);
+               schedstat_inc(p, se.statistics.nr_wakeups_affine);
 
                return 1;
        }
@@ -1232,11 +1549,11 @@ static int wake_affine(struct sched_domain *sd, struct task_struct *p, int sync)
  * domain.
  */
 static struct sched_group *
-find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
+find_idlest_group(struct sched_domain *sd, struct task_struct *p,
+                 int this_cpu, int load_idx)
 {
-       struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
+       struct sched_group *idlest = NULL, *group = sd->groups;
        unsigned long min_load = ULONG_MAX, this_load = 0;
-       int load_idx = sd->forkexec_idx;
        int imbalance = 100 + (sd->imbalance_pct-100)/2;
 
        do {
@@ -1266,11 +1583,10 @@ find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
                }
 
                /* Adjust by relative CPU power of the group */
-               avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
+               avg_load = (avg_load * SCHED_POWER_SCALE) / group->sgp->power;
 
                if (local_group) {
                        this_load = avg_load;
-                       this = group;
                } else if (avg_load < min_load) {
                        min_load = avg_load;
                        idlest = group;
@@ -1306,6 +1622,58 @@ find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
 }
 
 /*
+ * Try and locate an idle CPU in the sched_domain.
+ */
+static int select_idle_sibling(struct task_struct *p, int target)
+{
+       int cpu = smp_processor_id();
+       int prev_cpu = task_cpu(p);
+       struct sched_domain *sd;
+       int i;
+
+       /*
+        * If the task is going to be woken-up on this cpu and if it is
+        * already idle, then it is the right target.
+        */
+       if (target == cpu && idle_cpu(cpu))
+               return cpu;
+
+       /*
+        * If the task is going to be woken-up on the cpu where it previously
+        * ran and if it is currently idle, then it the right target.
+        */
+       if (target == prev_cpu && idle_cpu(prev_cpu))
+               return prev_cpu;
+
+       /*
+        * Otherwise, iterate the domains and find an elegible idle cpu.
+        */
+       rcu_read_lock();
+       for_each_domain(target, sd) {
+               if (!(sd->flags & SD_SHARE_PKG_RESOURCES))
+                       break;
+
+               for_each_cpu_and(i, sched_domain_span(sd), &p->cpus_allowed) {
+                       if (idle_cpu(i)) {
+                               target = i;
+                               break;
+                       }
+               }
+
+               /*
+                * Lets stop looking for an idle sibling when we reached
+                * the domain that spans the current cpu and prev_cpu.
+                */
+               if (cpumask_test_cpu(cpu, sched_domain_span(sd)) &&
+                   cpumask_test_cpu(prev_cpu, sched_domain_span(sd)))
+                       break;
+       }
+       rcu_read_unlock();
+
+       return target;
+}
+
+/*
  * sched_balance_self: balance the current task (running on cpu) in domains
  * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
  * SD_BALANCE_EXEC.
@@ -1316,27 +1684,33 @@ find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
  *
  * preempt must be disabled.
  */
-static int select_task_rq_fair(struct task_struct *p, int flag, int sync)
+static int
+select_task_rq_fair(struct task_struct *p, int sd_flag, int wake_flags)
 {
-       struct sched_domain *tmp, *sd = NULL;
+       struct sched_domain *tmp, *affine_sd = NULL, *sd = NULL;
        int cpu = smp_processor_id();
        int prev_cpu = task_cpu(p);
        int new_cpu = cpu;
        int want_affine = 0;
+       int want_sd = 1;
+       int sync = wake_flags & WF_SYNC;
 
-       if (flag & SD_BALANCE_WAKE) {
-               if (sched_feat(AFFINE_WAKEUPS))
+       if (sd_flag & SD_BALANCE_WAKE) {
+               if (cpumask_test_cpu(cpu, &p->cpus_allowed))
                        want_affine = 1;
                new_cpu = prev_cpu;
        }
 
        rcu_read_lock();
        for_each_domain(cpu, tmp) {
+               if (!(tmp->flags & SD_LOAD_BALANCE))
+                       continue;
+
                /*
                 * If power savings logic is enabled for a domain, see if we
                 * are not overloaded, if so, don't balance wider.
                 */
-               if (tmp->flags & SD_POWERSAVINGS_BALANCE) {
+               if (tmp->flags & (SD_POWERSAVINGS_BALANCE|SD_PREFER_LOCAL)) {
                        unsigned long power = 0;
                        unsigned long nr_running = 0;
                        unsigned long capacity;
@@ -1347,52 +1721,57 @@ static int select_task_rq_fair(struct task_struct *p, int flag, int sync)
                                nr_running += cpu_rq(i)->cfs.nr_running;
                        }
 
-                       capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
+                       capacity = DIV_ROUND_CLOSEST(power, SCHED_POWER_SCALE);
 
-                       if (nr_running/2 < capacity)
-                               break;
-               }
+                       if (tmp->flags & SD_POWERSAVINGS_BALANCE)
+                               nr_running /= 2;
 
-               switch (flag) {
-               case SD_BALANCE_WAKE:
-                       if (!sched_feat(LB_WAKEUP_UPDATE))
-                               break;
-               case SD_BALANCE_FORK:
-               case SD_BALANCE_EXEC:
-                       if (root_task_group_empty())
-                               break;
-                       update_shares(tmp);
-               default:
-                       break;
+                       if (nr_running < capacity)
+                               want_sd = 0;
                }
 
+               /*
+                * If both cpu and prev_cpu are part of this domain,
+                * cpu is a valid SD_WAKE_AFFINE target.
+                */
                if (want_affine && (tmp->flags & SD_WAKE_AFFINE) &&
                    cpumask_test_cpu(prev_cpu, sched_domain_span(tmp))) {
-
-                       if (wake_affine(tmp, p, sync)) {
-                               new_cpu = cpu;
-                               goto out;
-                       }
-
+                       affine_sd = tmp;
                        want_affine = 0;
                }
 
-               if (!(tmp->flags & flag))
+               if (!want_sd && !want_affine)
+                       break;
+
+               if (!(tmp->flags & sd_flag))
                        continue;
 
-               sd = tmp;
+               if (want_sd)
+                       sd = tmp;
+       }
+
+       if (affine_sd) {
+               if (cpu == prev_cpu || wake_affine(affine_sd, p, sync))
+                       prev_cpu = cpu;
+
+               new_cpu = select_idle_sibling(p, prev_cpu);
+               goto unlock;
        }
 
        while (sd) {
+               int load_idx = sd->forkexec_idx;
                struct sched_group *group;
                int weight;
 
-               if (!(sd->flags & flag)) {
+               if (!(sd->flags & sd_flag)) {
                        sd = sd->child;
                        continue;
                }
 
-               group = find_idlest_group(sd, p, cpu);
+               if (sd_flag & SD_BALANCE_WAKE)
+                       load_idx = sd->wake_idx;
+
+               group = find_idlest_group(sd, p, cpu, load_idx);
                if (!group) {
                        sd = sd->child;
                        continue;
@@ -1407,82 +1786,42 @@ static int select_task_rq_fair(struct task_struct *p, int flag, int sync)
 
                /* Now try balancing at a lower domain level of new_cpu */
                cpu = new_cpu;
-               weight = cpumask_weight(sched_domain_span(sd));
+               weight = sd->span_weight;
                sd = NULL;
                for_each_domain(cpu, tmp) {
-                       if (weight <= cpumask_weight(sched_domain_span(tmp)))
+                       if (weight <= tmp->span_weight)
                                break;
-                       if (tmp->flags & flag)
+                       if (tmp->flags & sd_flag)
                                sd = tmp;
                }
                /* while loop will break here if sd == NULL */
        }
-
-out:
+unlock:
        rcu_read_unlock();
+
        return new_cpu;
 }
 #endif /* CONFIG_SMP */
 
-/*
- * Adaptive granularity
- *
- * se->avg_wakeup gives the average time a task runs until it does a wakeup,
- * with the limit of wakeup_gran -- when it never does a wakeup.
- *
- * So the smaller avg_wakeup is the faster we want this task to preempt,
- * but we don't want to treat the preemptee unfairly and therefore allow it
- * to run for at least the amount of time we'd like to run.
- *
- * NOTE: we use 2*avg_wakeup to increase the probability of actually doing one
- *
- * NOTE: we use *nr_running to scale with load, this nicely matches the
- *       degrading latency on load.
- */
-static unsigned long
-adaptive_gran(struct sched_entity *curr, struct sched_entity *se)
-{
-       u64 this_run = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
-       u64 expected_wakeup = 2*se->avg_wakeup * cfs_rq_of(se)->nr_running;
-       u64 gran = 0;
-
-       if (this_run < expected_wakeup)
-               gran = expected_wakeup - this_run;
-
-       return min_t(s64, gran, sysctl_sched_wakeup_granularity);
-}
-
 static unsigned long
 wakeup_gran(struct sched_entity *curr, struct sched_entity *se)
 {
        unsigned long gran = sysctl_sched_wakeup_granularity;
 
-       if (cfs_rq_of(curr)->curr && sched_feat(ADAPTIVE_GRAN))
-               gran = adaptive_gran(curr, se);
-
        /*
         * Since its curr running now, convert the gran from real-time
         * to virtual-time in his units.
+        *
+        * By using 'se' instead of 'curr' we penalize light tasks, so
+        * they get preempted easier. That is, if 'se' < 'curr' then
+        * the resulting gran will be larger, therefore penalizing the
+        * lighter, if otoh 'se' > 'curr' then the resulting gran will
+        * be smaller, again penalizing the lighter task.
+        *
+        * This is especially important for buddies when the leftmost
+        * task is higher priority than the buddy.
         */
-       if (sched_feat(ASYM_GRAN)) {
-               /*
-                * By using 'se' instead of 'curr' we penalize light tasks, so
-                * they get preempted easier. That is, if 'se' < 'curr' then
-                * the resulting gran will be larger, therefore penalizing the
-                * lighter, if otoh 'se' > 'curr' then the resulting gran will
-                * be smaller, again penalizing the lighter task.
-                *
-                * This is especially important for buddies when the leftmost
-                * task is higher priority than the buddy.
-                */
-               if (unlikely(se->load.weight != NICE_0_LOAD))
-                       gran = calc_delta_fair(gran, se);
-       } else {
-               if (unlikely(curr->load.weight != NICE_0_LOAD))
-                       gran = calc_delta_fair(gran, curr);
-       }
-
-       return gran;
+       return calc_delta_fair(gran, se);
 }
 
 /*
@@ -1516,55 +1855,46 @@ wakeup_preempt_entity(struct sched_entity *curr, struct sched_entity *se)
 
 static void set_last_buddy(struct sched_entity *se)
 {
-       if (likely(task_of(se)->policy != SCHED_IDLE)) {
-               for_each_sched_entity(se)
-                       cfs_rq_of(se)->last = se;
-       }
+       if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
+               return;
+
+       for_each_sched_entity(se)
+               cfs_rq_of(se)->last = se;
 }
 
 static void set_next_buddy(struct sched_entity *se)
 {
-       if (likely(task_of(se)->policy != SCHED_IDLE)) {
-               for_each_sched_entity(se)
-                       cfs_rq_of(se)->next = se;
-       }
+       if (entity_is_task(se) && unlikely(task_of(se)->policy == SCHED_IDLE))
+               return;
+
+       for_each_sched_entity(se)
+               cfs_rq_of(se)->next = se;
+}
+
+static void set_skip_buddy(struct sched_entity *se)
+{
+       for_each_sched_entity(se)
+               cfs_rq_of(se)->skip = se;
 }
 
 /*
  * Preempt the current task with a newly woken task if needed:
  */
-static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int sync)
+static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
 {
        struct task_struct *curr = rq->curr;
        struct sched_entity *se = &curr->se, *pse = &p->se;
        struct cfs_rq *cfs_rq = task_cfs_rq(curr);
-
-       update_curr(cfs_rq);
-
-       if (unlikely(rt_prio(p->prio))) {
-               resched_task(curr);
-               return;
-       }
-
-       if (unlikely(p->sched_class != &fair_sched_class))
-               return;
+       int scale = cfs_rq->nr_running >= sched_nr_latency;
+       int next_buddy_marked = 0;
 
        if (unlikely(se == pse))
                return;
 
-       /*
-        * Only set the backward buddy when the current task is still on the
-        * rq. This can happen when a wakeup gets interleaved with schedule on
-        * the ->pre_schedule() or idle_balance() point, either of which can
-        * drop the rq lock.
-        *
-        * Also, during early boot the idle thread is in the fair class, for
-        * obvious reasons its a bad idea to schedule back to the idle thread.
-        */
-       if (sched_feat(LAST_BUDDY) && likely(se->on_rq && curr != rq->idle))
-               set_last_buddy(se);
-       if (sched_feat(NEXT_BUDDY))
+       if (sched_feat(NEXT_BUDDY) && scale && !(wake_flags & WF_FORK)) {
                set_next_buddy(pse);
+               next_buddy_marked = 1;
+       }
 
        /*
         * We can come here with TIF_NEED_RESCHED already set from new task
@@ -1573,36 +1903,53 @@ static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int sync)
        if (test_tsk_need_resched(curr))
                return;
 
+       /* Idle tasks are by definition preempted by non-idle tasks. */
+       if (unlikely(curr->policy == SCHED_IDLE) &&
+           likely(p->policy != SCHED_IDLE))
+               goto preempt;
+
        /*
-        * Batch and idle tasks do not preempt (their preemption is driven by
-        * the tick):
+        * Batch and idle tasks do not preempt non-idle tasks (their preemption
+        * is driven by the tick):
         */
        if (unlikely(p->policy != SCHED_NORMAL))
                return;
 
-       /* Idle tasks are by definition preempted by everybody. */
-       if (unlikely(curr->policy == SCHED_IDLE)) {
-               resched_task(curr);
-               return;
-       }
 
        if (!sched_feat(WAKEUP_PREEMPT))
                return;
 
-       if ((sched_feat(WAKEUP_SYNC) && sync) ||
-           (sched_feat(WAKEUP_OVERLAP) &&
-            (se->avg_overlap < sysctl_sched_migration_cost &&
-             pse->avg_overlap < sysctl_sched_migration_cost))) {
-               resched_task(curr);
-               return;
+       update_curr(cfs_rq);
+       find_matching_se(&se, &pse);
+       BUG_ON(!pse);
+       if (wakeup_preempt_entity(se, pse) == 1) {
+               /*
+                * Bias pick_next to pick the sched entity that is
+                * triggering this preemption.
+                */
+               if (!next_buddy_marked)
+                       set_next_buddy(pse);
+               goto preempt;
        }
 
-       find_matching_se(&se, &pse);
+       return;
 
-       BUG_ON(!pse);
+preempt:
+       resched_task(curr);
+       /*
+        * Only set the backward buddy when the current task is still
+        * on the rq. This can happen when a wakeup gets interleaved
+        * with schedule on the ->pre_schedule() or idle_balance()
+        * point, either of which can * drop the rq lock.
+        *
+        * Also, during early boot the idle thread is in the fair class,
+        * for obvious reasons its a bad idea to schedule back to it.
+        */
+       if (unlikely(!se->on_rq || curr == rq->idle))
+               return;
 
-       if (wakeup_preempt_entity(se, pse) == 1)
-               resched_task(curr);
+       if (sched_feat(LAST_BUDDY) && scale && entity_is_task(se))
+               set_last_buddy(se);
 }
 
 static struct task_struct *pick_next_task_fair(struct rq *rq)
@@ -1611,16 +1958,11 @@ static struct task_struct *pick_next_task_fair(struct rq *rq)
        struct cfs_rq *cfs_rq = &rq->cfs;
        struct sched_entity *se;
 
-       if (unlikely(!cfs_rq->nr_running))
+       if (!cfs_rq->nr_running)
                return NULL;
 
        do {
                se = pick_next_entity(cfs_rq);
-               /*
-                * If se was a buddy, clear it so that it will have to earn
-                * the favour again.
-                */
-               __clear_buddies(cfs_rq, se);
                set_next_entity(cfs_rq, se);
                cfs_rq = group_cfs_rq(se);
        } while (cfs_rq);
@@ -1645,71 +1987,258 @@ static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
        }
 }
 
+/*
+ * sched_yield() is very simple
+ *
+ * The magic of dealing with the ->skip buddy is in pick_next_entity.
+ */
+static void yield_task_fair(struct rq *rq)
+{
+       struct task_struct *curr = rq->curr;
+       struct cfs_rq *cfs_rq = task_cfs_rq(curr);
+       struct sched_entity *se = &curr->se;
+
+       /*
+        * Are we the only task in the tree?
+        */
+       if (unlikely(rq->nr_running == 1))
+               return;
+
+       clear_buddies(cfs_rq, se);
+
+       if (curr->policy != SCHED_BATCH) {
+               update_rq_clock(rq);
+               /*
+                * Update run-time statistics of the 'current'.
+                */
+               update_curr(cfs_rq);
+       }
+
+       set_skip_buddy(se);
+}
+
+static bool yield_to_task_fair(struct rq *rq, struct task_struct *p, bool preempt)
+{
+       struct sched_entity *se = &p->se;
+
+       if (!se->on_rq)
+               return false;
+
+       /* Tell the scheduler that we'd really like pse to run next. */
+       set_next_buddy(se);
+
+       yield_task_fair(rq);
+
+       return true;
+}
+
 #ifdef CONFIG_SMP
 /**************************************************
  * Fair scheduling class load-balancing methods:
  */
 
 /*
- * Load-balancing iterator. Note: while the runqueue stays locked
- * during the whole iteration, the current task might be
- * dequeued so the iterator has to be dequeue-safe. Here we
- * achieve that by always pre-iterating before returning
- * the current task:
+ * pull_task - move a task from a remote runqueue to the local runqueue.
+ * Both runqueues must be locked.
  */
-static struct task_struct *
-__load_balance_iterator(struct cfs_rq *cfs_rq, struct list_head *next)
+static void pull_task(struct rq *src_rq, struct task_struct *p,
+                     struct rq *this_rq, int this_cpu)
 {
-       struct task_struct *p = NULL;
-       struct sched_entity *se;
+       deactivate_task(src_rq, p, 0);
+       set_task_cpu(p, this_cpu);
+       activate_task(this_rq, p, 0);
+       check_preempt_curr(this_rq, p, 0);
+}
 
-       if (next == &cfs_rq->tasks)
-               return NULL;
+/*
+ * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
+ */
+static
+int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
+                    struct sched_domain *sd, enum cpu_idle_type idle,
+                    int *all_pinned)
+{
+       int tsk_cache_hot = 0;
+       /*
+        * We do not migrate tasks that are:
+        * 1) running (obviously), or
+        * 2) cannot be migrated to this CPU due to cpus_allowed, or
+        * 3) are cache-hot on their current CPU.
+        */
+       if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
+               schedstat_inc(p, se.statistics.nr_failed_migrations_affine);
+               return 0;
+       }
+       *all_pinned = 0;
 
-       se = list_entry(next, struct sched_entity, group_node);
-       p = task_of(se);
-       cfs_rq->balance_iterator = next->next;
+       if (task_running(rq, p)) {
+               schedstat_inc(p, se.statistics.nr_failed_migrations_running);
+               return 0;
+       }
 
-       return p;
-}
+       /*
+        * Aggressive migration if:
+        * 1) task is cache cold, or
+        * 2) too many balance attempts have failed.
+        */
 
-static struct task_struct *load_balance_start_fair(void *arg)
-{
-       struct cfs_rq *cfs_rq = arg;
+       tsk_cache_hot = task_hot(p, rq->clock_task, sd);
+       if (!tsk_cache_hot ||
+               sd->nr_balance_failed > sd->cache_nice_tries) {
+#ifdef CONFIG_SCHEDSTATS
+               if (tsk_cache_hot) {
+                       schedstat_inc(sd, lb_hot_gained[idle]);
+                       schedstat_inc(p, se.statistics.nr_forced_migrations);
+               }
+#endif
+               return 1;
+       }
 
-       return __load_balance_iterator(cfs_rq, cfs_rq->tasks.next);
+       if (tsk_cache_hot) {
+               schedstat_inc(p, se.statistics.nr_failed_migrations_hot);
+               return 0;
+       }
+       return 1;
 }
 
-static struct task_struct *load_balance_next_fair(void *arg)
+/*
+ * move_one_task tries to move exactly one task from busiest to this_rq, as
+ * part of active balancing operations within "domain".
+ * Returns 1 if successful and 0 otherwise.
+ *
+ * Called with both runqueues locked.
+ */
+static int
+move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
+             struct sched_domain *sd, enum cpu_idle_type idle)
 {
-       struct cfs_rq *cfs_rq = arg;
+       struct task_struct *p, *n;
+       struct cfs_rq *cfs_rq;
+       int pinned = 0;
+
+       for_each_leaf_cfs_rq(busiest, cfs_rq) {
+               list_for_each_entry_safe(p, n, &cfs_rq->tasks, se.group_node) {
+
+                       if (!can_migrate_task(p, busiest, this_cpu,
+                                               sd, idle, &pinned))
+                               continue;
 
-       return __load_balance_iterator(cfs_rq, cfs_rq->balance_iterator);
+                       pull_task(busiest, p, this_rq, this_cpu);
+                       /*
+                        * Right now, this is only the second place pull_task()
+                        * is called, so we can safely collect pull_task()
+                        * stats here rather than inside pull_task().
+                        */
+                       schedstat_inc(sd, lb_gained[idle]);
+                       return 1;
+               }
+       }
+
+       return 0;
 }
 
 static unsigned long
-__load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
-               unsigned long max_load_move, struct sched_domain *sd,
-               enum cpu_idle_type idle, int *all_pinned, int *this_best_prio,
-               struct cfs_rq *cfs_rq)
+balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
+             unsigned long max_load_move, struct sched_domain *sd,
+             enum cpu_idle_type idle, int *all_pinned,
+             struct cfs_rq *busiest_cfs_rq)
 {
-       struct rq_iterator cfs_rq_iterator;
+       int loops = 0, pulled = 0;
+       long rem_load_move = max_load_move;
+       struct task_struct *p, *n;
 
-       cfs_rq_iterator.start = load_balance_start_fair;
-       cfs_rq_iterator.next = load_balance_next_fair;
-       cfs_rq_iterator.arg = cfs_rq;
+       if (max_load_move == 0)
+               goto out;
 
-       return balance_tasks(this_rq, this_cpu, busiest,
-                       max_load_move, sd, idle, all_pinned,
-                       this_best_prio, &cfs_rq_iterator);
+       list_for_each_entry_safe(p, n, &busiest_cfs_rq->tasks, se.group_node) {
+               if (loops++ > sysctl_sched_nr_migrate)
+                       break;
+
+               if ((p->se.load.weight >> 1) > rem_load_move ||
+                   !can_migrate_task(p, busiest, this_cpu, sd, idle,
+                                     all_pinned))
+                       continue;
+
+               pull_task(busiest, p, this_rq, this_cpu);
+               pulled++;
+               rem_load_move -= p->se.load.weight;
+
+#ifdef CONFIG_PREEMPT
+               /*
+                * NEWIDLE balancing is a source of latency, so preemptible
+                * kernels will stop after the first task is pulled to minimize
+                * the critical section.
+                */
+               if (idle == CPU_NEWLY_IDLE)
+                       break;
+#endif
+
+               /*
+                * We only want to steal up to the prescribed amount of
+                * weighted load.
+                */
+               if (rem_load_move <= 0)
+                       break;
+       }
+out:
+       /*
+        * Right now, this is one of only two places pull_task() is called,
+        * so we can safely collect pull_task() stats here rather than
+        * inside pull_task().
+        */
+       schedstat_add(sd, lb_gained[idle], pulled);
+
+       return max_load_move - rem_load_move;
 }
 
 #ifdef CONFIG_FAIR_GROUP_SCHED
+/*
+ * update tg->load_weight by folding this cpu's load_avg
+ */
+static int update_shares_cpu(struct task_group *tg, int cpu)
+{
+       struct cfs_rq *cfs_rq;
+       unsigned long flags;
+       struct rq *rq;
+
+       if (!tg->se[cpu])
+               return 0;
+
+       rq = cpu_rq(cpu);
+       cfs_rq = tg->cfs_rq[cpu];
+
+       raw_spin_lock_irqsave(&rq->lock, flags);
+
+       update_rq_clock(rq);
+       update_cfs_load(cfs_rq, 1);
+
+       /*
+        * We need to update shares after updating tg->load_weight in
+        * order to adjust the weight of groups with long running tasks.
+        */
+       update_cfs_shares(cfs_rq);
+
+       raw_spin_unlock_irqrestore(&rq->lock, flags);
+
+       return 0;
+}
+
+static void update_shares(int cpu)
+{
+       struct cfs_rq *cfs_rq;
+       struct rq *rq = cpu_rq(cpu);
+
+       rcu_read_lock();
+       for_each_leaf_cfs_rq(rq, cfs_rq)
+               update_shares_cpu(cfs_rq->tg, cpu);
+       rcu_read_unlock();
+}
+
 static unsigned long
 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
                  unsigned long max_load_move,
                  struct sched_domain *sd, enum cpu_idle_type idle,
-                 int *all_pinned, int *this_best_prio)
+                 int *all_pinned)
 {
        long rem_load_move = max_load_move;
        int busiest_cpu = cpu_of(busiest);
@@ -1733,9 +2262,9 @@ load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
                rem_load = (u64)rem_load_move * busiest_weight;
                rem_load = div_u64(rem_load, busiest_h_load + 1);
 
-               moved_load = __load_balance_fair(this_rq, this_cpu, busiest,
-                               rem_load, sd, idle, all_pinned, this_best_prio,
-                               tg->cfs_rq[busiest_cpu]);
+               moved_load = balance_tasks(this_rq, this_cpu, busiest,
+                               rem_load, sd, idle, all_pinned,
+                               busiest_cfs_rq);
 
                if (!moved_load)
                        continue;
@@ -1752,41 +2281,1829 @@ load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
        return max_load_move - rem_load_move;
 }
 #else
+static inline void update_shares(int cpu)
+{
+}
+
 static unsigned long
 load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
                  unsigned long max_load_move,
                  struct sched_domain *sd, enum cpu_idle_type idle,
-                 int *all_pinned, int *this_best_prio)
+                 int *all_pinned)
 {
-       return __load_balance_fair(this_rq, this_cpu, busiest,
+       return balance_tasks(this_rq, this_cpu, busiest,
                        max_load_move, sd, idle, all_pinned,
-                       this_best_prio, &busiest->cfs);
+                       &busiest->cfs);
 }
 #endif
 
-static int
-move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
-                  struct sched_domain *sd, enum cpu_idle_type idle)
+/*
+ * move_tasks tries to move up to max_load_move weighted load from busiest to
+ * this_rq, as part of a balancing operation within domain "sd".
+ * Returns 1 if successful and 0 otherwise.
+ *
+ * Called with both runqueues locked.
+ */
+static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
+                     unsigned long max_load_move,
+                     struct sched_domain *sd, enum cpu_idle_type idle,
+                     int *all_pinned)
+{
+       unsigned long total_load_moved = 0, load_moved;
+
+       do {
+               load_moved = load_balance_fair(this_rq, this_cpu, busiest,
+                               max_load_move - total_load_moved,
+                               sd, idle, all_pinned);
+
+               total_load_moved += load_moved;
+
+#ifdef CONFIG_PREEMPT
+               /*
+                * NEWIDLE balancing is a source of latency, so preemptible
+                * kernels will stop after the first task is pulled to minimize
+                * the critical section.
+                */
+               if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
+                       break;
+
+               if (raw_spin_is_contended(&this_rq->lock) ||
+                               raw_spin_is_contended(&busiest->lock))
+                       break;
+#endif
+       } while (load_moved && max_load_move > total_load_moved);
+
+       return total_load_moved > 0;
+}
+
+/********** Helpers for find_busiest_group ************************/
+/*
+ * sd_lb_stats - Structure to store the statistics of a sched_domain
+ *             during load balancing.
+ */
+struct sd_lb_stats {
+       struct sched_group *busiest; /* Busiest group in this sd */
+       struct sched_group *this;  /* Local group in this sd */
+       unsigned long total_load;  /* Total load of all groups in sd */
+       unsigned long total_pwr;   /*   Total power of all groups in sd */
+       unsigned long avg_load;    /* Average load across all groups in sd */
+
+       /** Statistics of this group */
+       unsigned long this_load;
+       unsigned long this_load_per_task;
+       unsigned long this_nr_running;
+       unsigned long this_has_capacity;
+       unsigned int  this_idle_cpus;
+
+       /* Statistics of the busiest group */
+       unsigned int  busiest_idle_cpus;
+       unsigned long max_load;
+       unsigned long busiest_load_per_task;
+       unsigned long busiest_nr_running;
+       unsigned long busiest_group_capacity;
+       unsigned long busiest_has_capacity;
+       unsigned int  busiest_group_weight;
+
+       int group_imb; /* Is there imbalance in this sd */
+#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
+       int power_savings_balance; /* Is powersave balance needed for this sd */
+       struct sched_group *group_min; /* Least loaded group in sd */
+       struct sched_group *group_leader; /* Group which relieves group_min */
+       unsigned long min_load_per_task; /* load_per_task in group_min */
+       unsigned long leader_nr_running; /* Nr running of group_leader */
+       unsigned long min_nr_running; /* Nr running of group_min */
+#endif
+};
+
+/*
+ * sg_lb_stats - stats of a sched_group required for load_balancing
+ */
+struct sg_lb_stats {
+       unsigned long avg_load; /*Avg load across the CPUs of the group */
+       unsigned long group_load; /* Total load over the CPUs of the group */
+       unsigned long sum_nr_running; /* Nr tasks running in the group */
+       unsigned long sum_weighted_load; /* Weighted load of group's tasks */
+       unsigned long group_capacity;
+       unsigned long idle_cpus;
+       unsigned long group_weight;
+       int group_imb; /* Is there an imbalance in the group ? */
+       int group_has_capacity; /* Is there extra capacity in the group? */
+};
+
+/**
+ * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
+ * @group: The group whose first cpu is to be returned.
+ */
+static inline unsigned int group_first_cpu(struct sched_group *group)
+{
+       return cpumask_first(sched_group_cpus(group));
+}
+
+/**
+ * get_sd_load_idx - Obtain the load index for a given sched domain.
+ * @sd: The sched_domain whose load_idx is to be obtained.
+ * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
+ */
+static inline int get_sd_load_idx(struct sched_domain *sd,
+                                       enum cpu_idle_type idle)
+{
+       int load_idx;
+
+       switch (idle) {
+       case CPU_NOT_IDLE:
+               load_idx = sd->busy_idx;
+               break;
+
+       case CPU_NEWLY_IDLE:
+               load_idx = sd->newidle_idx;
+               break;
+       default:
+               load_idx = sd->idle_idx;
+               break;
+       }
+
+       return load_idx;
+}
+
+
+#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
+/**
+ * init_sd_power_savings_stats - Initialize power savings statistics for
+ * the given sched_domain, during load balancing.
+ *
+ * @sd: Sched domain whose power-savings statistics are to be initialized.
+ * @sds: Variable containing the statistics for sd.
+ * @idle: Idle status of the CPU at which we're performing load-balancing.
+ */
+static inline void init_sd_power_savings_stats(struct sched_domain *sd,
+       struct sd_lb_stats *sds, enum cpu_idle_type idle)
+{
+       /*
+        * Busy processors will not participate in power savings
+        * balance.
+        */
+       if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
+               sds->power_savings_balance = 0;
+       else {
+               sds->power_savings_balance = 1;
+               sds->min_nr_running = ULONG_MAX;
+               sds->leader_nr_running = 0;
+       }
+}
+
+/**
+ * update_sd_power_savings_stats - Update the power saving stats for a
+ * sched_domain while performing load balancing.
+ *
+ * @group: sched_group belonging to the sched_domain under consideration.
+ * @sds: Variable containing the statistics of the sched_domain
+ * @local_group: Does group contain the CPU for which we're performing
+ *             load balancing ?
+ * @sgs: Variable containing the statistics of the group.
+ */
+static inline void update_sd_power_savings_stats(struct sched_group *group,
+       struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
+{
+
+       if (!sds->power_savings_balance)
+               return;
+
+       /*
+        * If the local group is idle or completely loaded
+        * no need to do power savings balance at this domain
+        */
+       if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
+                               !sds->this_nr_running))
+               sds->power_savings_balance = 0;
+
+       /*
+        * If a group is already running at full capacity or idle,
+        * don't include that group in power savings calculations
+        */
+       if (!sds->power_savings_balance ||
+               sgs->sum_nr_running >= sgs->group_capacity ||
+               !sgs->sum_nr_running)
+               return;
+
+       /*
+        * Calculate the group which has the least non-idle load.
+        * This is the group from where we need to pick up the load
+        * for saving power
+        */
+       if ((sgs->sum_nr_running < sds->min_nr_running) ||
+           (sgs->sum_nr_running == sds->min_nr_running &&
+            group_first_cpu(group) > group_first_cpu(sds->group_min))) {
+               sds->group_min = group;
+               sds->min_nr_running = sgs->sum_nr_running;
+               sds->min_load_per_task = sgs->sum_weighted_load /
+                                               sgs->sum_nr_running;
+       }
+
+       /*
+        * Calculate the group which is almost near its
+        * capacity but still has some space to pick up some load
+        * from other group and save more power
+        */
+       if (sgs->sum_nr_running + 1 > sgs->group_capacity)
+               return;
+
+       if (sgs->sum_nr_running > sds->leader_nr_running ||
+           (sgs->sum_nr_running == sds->leader_nr_running &&
+            group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
+               sds->group_leader = group;
+               sds->leader_nr_running = sgs->sum_nr_running;
+       }
+}
+
+/**
+ * check_power_save_busiest_group - see if there is potential for some power-savings balance
+ * @sds: Variable containing the statistics of the sched_domain
+ *     under consideration.
+ * @this_cpu: Cpu at which we're currently performing load-balancing.
+ * @imbalance: Variable to store the imbalance.
+ *
+ * Description:
+ * Check if we have potential to perform some power-savings balance.
+ * If yes, set the busiest group to be the least loaded group in the
+ * sched_domain, so that it's CPUs can be put to idle.
+ *
+ * Returns 1 if there is potential to perform power-savings balance.
+ * Else returns 0.
+ */
+static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
+                                       int this_cpu, unsigned long *imbalance)
+{
+       if (!sds->power_savings_balance)
+               return 0;
+
+       if (sds->this != sds->group_leader ||
+                       sds->group_leader == sds->group_min)
+               return 0;
+
+       *imbalance = sds->min_load_per_task;
+       sds->busiest = sds->group_min;
+
+       return 1;
+
+}
+#else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
+static inline void init_sd_power_savings_stats(struct sched_domain *sd,
+       struct sd_lb_stats *sds, enum cpu_idle_type idle)
+{
+       return;
+}
+
+static inline void update_sd_power_savings_stats(struct sched_group *group,
+       struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
+{
+       return;
+}
+
+static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
+                                       int this_cpu, unsigned long *imbalance)
+{
+       return 0;
+}
+#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
+
+
+unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
+{
+       return SCHED_POWER_SCALE;
+}
+
+unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
+{
+       return default_scale_freq_power(sd, cpu);
+}
+
+unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
+{
+       unsigned long weight = sd->span_weight;
+       unsigned long smt_gain = sd->smt_gain;
+
+       smt_gain /= weight;
+
+       return smt_gain;
+}
+
+unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
+{
+       return default_scale_smt_power(sd, cpu);
+}
+
+unsigned long scale_rt_power(int cpu)
+{
+       struct rq *rq = cpu_rq(cpu);
+       u64 total, available;
+
+       total = sched_avg_period() + (rq->clock - rq->age_stamp);
+
+       if (unlikely(total < rq->rt_avg)) {
+               /* Ensures that power won't end up being negative */
+               available = 0;
+       } else {
+               available = total - rq->rt_avg;
+       }
+
+       if (unlikely((s64)total < SCHED_POWER_SCALE))
+               total = SCHED_POWER_SCALE;
+
+       total >>= SCHED_POWER_SHIFT;
+
+       return div_u64(available, total);
+}
+
+static void update_cpu_power(struct sched_domain *sd, int cpu)
+{
+       unsigned long weight = sd->span_weight;
+       unsigned long power = SCHED_POWER_SCALE;
+       struct sched_group *sdg = sd->groups;
+
+       if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
+               if (sched_feat(ARCH_POWER))
+                       power *= arch_scale_smt_power(sd, cpu);
+               else
+                       power *= default_scale_smt_power(sd, cpu);
+
+               power >>= SCHED_POWER_SHIFT;
+       }
+
+       sdg->sgp->power_orig = power;
+
+       if (sched_feat(ARCH_POWER))
+               power *= arch_scale_freq_power(sd, cpu);
+       else
+               power *= default_scale_freq_power(sd, cpu);
+
+       power >>= SCHED_POWER_SHIFT;
+
+       power *= scale_rt_power(cpu);
+       power >>= SCHED_POWER_SHIFT;
+
+       if (!power)
+               power = 1;
+
+       cpu_rq(cpu)->cpu_power = power;
+       sdg->sgp->power = power;
+}
+
+static void update_group_power(struct sched_domain *sd, int cpu)
+{
+       struct sched_domain *child = sd->child;
+       struct sched_group *group, *sdg = sd->groups;
+       unsigned long power;
+
+       if (!child) {
+               update_cpu_power(sd, cpu);
+               return;
+       }
+
+       power = 0;
+
+       group = child->groups;
+       do {
+               power += group->sgp->power;
+               group = group->next;
+       } while (group != child->groups);
+
+       sdg->sgp->power = power;
+}
+
+/*
+ * Try and fix up capacity for tiny siblings, this is needed when
+ * things like SD_ASYM_PACKING need f_b_g to select another sibling
+ * which on its own isn't powerful enough.
+ *
+ * See update_sd_pick_busiest() and check_asym_packing().
+ */
+static inline int
+fix_small_capacity(struct sched_domain *sd, struct sched_group *group)
+{
+       /*
+        * Only siblings can have significantly less than SCHED_POWER_SCALE
+        */
+       if (!(sd->flags & SD_SHARE_CPUPOWER))
+               return 0;
+
+       /*
+        * If ~90% of the cpu_power is still there, we're good.
+        */
+       if (group->sgp->power * 32 > group->sgp->power_orig * 29)
+               return 1;
+
+       return 0;
+}
+
+/**
+ * update_sg_lb_stats - Update sched_group's statistics for load balancing.
+ * @sd: The sched_domain whose statistics are to be updated.
+ * @group: sched_group whose statistics are to be updated.
+ * @this_cpu: Cpu for which load balance is currently performed.
+ * @idle: Idle status of this_cpu
+ * @load_idx: Load index of sched_domain of this_cpu for load calc.
+ * @local_group: Does group contain this_cpu.
+ * @cpus: Set of cpus considered for load balancing.
+ * @balance: Should we balance.
+ * @sgs: variable to hold the statistics for this group.
+ */
+static inline void update_sg_lb_stats(struct sched_domain *sd,
+                       struct sched_group *group, int this_cpu,
+                       enum cpu_idle_type idle, int load_idx,
+                       int local_group, const struct cpumask *cpus,
+                       int *balance, struct sg_lb_stats *sgs)
+{
+       unsigned long load, max_cpu_load, min_cpu_load, max_nr_running;
+       int i;
+       unsigned int balance_cpu = -1, first_idle_cpu = 0;
+       unsigned long avg_load_per_task = 0;
+
+       if (local_group)
+               balance_cpu = group_first_cpu(group);
+
+       /* Tally up the load of all CPUs in the group */
+       max_cpu_load = 0;
+       min_cpu_load = ~0UL;
+       max_nr_running = 0;
+
+       for_each_cpu_and(i, sched_group_cpus(group), cpus) {
+               struct rq *rq = cpu_rq(i);
+
+               /* Bias balancing toward cpus of our domain */
+               if (local_group) {
+                       if (idle_cpu(i) && !first_idle_cpu) {
+                               first_idle_cpu = 1;
+                               balance_cpu = i;
+                       }
+
+                       load = target_load(i, load_idx);
+               } else {
+                       load = source_load(i, load_idx);
+                       if (load > max_cpu_load) {
+                               max_cpu_load = load;
+                               max_nr_running = rq->nr_running;
+                       }
+                       if (min_cpu_load > load)
+                               min_cpu_load = load;
+               }
+
+               sgs->group_load += load;
+               sgs->sum_nr_running += rq->nr_running;
+               sgs->sum_weighted_load += weighted_cpuload(i);
+               if (idle_cpu(i))
+                       sgs->idle_cpus++;
+       }
+
+       /*
+        * First idle cpu or the first cpu(busiest) in this sched group
+        * is eligible for doing load balancing at this and above
+        * domains. In the newly idle case, we will allow all the cpu's
+        * to do the newly idle load balance.
+        */
+       if (idle != CPU_NEWLY_IDLE && local_group) {
+               if (balance_cpu != this_cpu) {
+                       *balance = 0;
+                       return;
+               }
+               update_group_power(sd, this_cpu);
+       }
+
+       /* Adjust by relative CPU power of the group */
+       sgs->avg_load = (sgs->group_load*SCHED_POWER_SCALE) / group->sgp->power;
+
+       /*
+        * Consider the group unbalanced when the imbalance is larger
+        * than the average weight of a task.
+        *
+        * APZ: with cgroup the avg task weight can vary wildly and
+        *      might not be a suitable number - should we keep a
+        *      normalized nr_running number somewhere that negates
+        *      the hierarchy?
+        */
+       if (sgs->sum_nr_running)
+               avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
+
+       if ((max_cpu_load - min_cpu_load) >= avg_load_per_task && max_nr_running > 1)
+               sgs->group_imb = 1;
+
+       sgs->group_capacity = DIV_ROUND_CLOSEST(group->sgp->power,
+                                               SCHED_POWER_SCALE);
+       if (!sgs->group_capacity)
+               sgs->group_capacity = fix_small_capacity(sd, group);
+       sgs->group_weight = group->group_weight;
+
+       if (sgs->group_capacity > sgs->sum_nr_running)
+               sgs->group_has_capacity = 1;
+}
+
+/**
+ * update_sd_pick_busiest - return 1 on busiest group
+ * @sd: sched_domain whose statistics are to be checked
+ * @sds: sched_domain statistics
+ * @sg: sched_group candidate to be checked for being the busiest
+ * @sgs: sched_group statistics
+ * @this_cpu: the current cpu
+ *
+ * Determine if @sg is a busier group than the previously selected
+ * busiest group.
+ */
+static bool update_sd_pick_busiest(struct sched_domain *sd,
+                                  struct sd_lb_stats *sds,
+                                  struct sched_group *sg,
+                                  struct sg_lb_stats *sgs,
+                                  int this_cpu)
+{
+       if (sgs->avg_load <= sds->max_load)
+               return false;
+
+       if (sgs->sum_nr_running > sgs->group_capacity)
+               return true;
+
+       if (sgs->group_imb)
+               return true;
+
+       /*
+        * ASYM_PACKING needs to move all the work to the lowest
+        * numbered CPUs in the group, therefore mark all groups
+        * higher than ourself as busy.
+        */
+       if ((sd->flags & SD_ASYM_PACKING) && sgs->sum_nr_running &&
+           this_cpu < group_first_cpu(sg)) {
+               if (!sds->busiest)
+                       return true;
+
+               if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
+                       return true;
+       }
+
+       return false;
+}
+
+/**
+ * update_sd_lb_stats - Update sched_group's statistics for load balancing.
+ * @sd: sched_domain whose statistics are to be updated.
+ * @this_cpu: Cpu for which load balance is currently performed.
+ * @idle: Idle status of this_cpu
+ * @cpus: Set of cpus considered for load balancing.
+ * @balance: Should we balance.
+ * @sds: variable to hold the statistics for this sched_domain.
+ */
+static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
+                       enum cpu_idle_type idle, const struct cpumask *cpus,
+                       int *balance, struct sd_lb_stats *sds)
+{
+       struct sched_domain *child = sd->child;
+       struct sched_group *sg = sd->groups;
+       struct sg_lb_stats sgs;
+       int load_idx, prefer_sibling = 0;
+
+       if (child && child->flags & SD_PREFER_SIBLING)
+               prefer_sibling = 1;
+
+       init_sd_power_savings_stats(sd, sds, idle);
+       load_idx = get_sd_load_idx(sd, idle);
+
+       do {
+               int local_group;
+
+               local_group = cpumask_test_cpu(this_cpu, sched_group_cpus(sg));
+               memset(&sgs, 0, sizeof(sgs));
+               update_sg_lb_stats(sd, sg, this_cpu, idle, load_idx,
+                               local_group, cpus, balance, &sgs);
+
+               if (local_group && !(*balance))
+                       return;
+
+               sds->total_load += sgs.group_load;
+               sds->total_pwr += sg->sgp->power;
+
+               /*
+                * In case the child domain prefers tasks go to siblings
+                * first, lower the sg capacity to one so that we'll try
+                * and move all the excess tasks away. We lower the capacity
+                * of a group only if the local group has the capacity to fit
+                * these excess tasks, i.e. nr_running < group_capacity. The
+                * extra check prevents the case where you always pull from the
+                * heaviest group when it is already under-utilized (possible
+                * with a large weight task outweighs the tasks on the system).
+                */
+               if (prefer_sibling && !local_group && sds->this_has_capacity)
+                       sgs.group_capacity = min(sgs.group_capacity, 1UL);
+
+               if (local_group) {
+                       sds->this_load = sgs.avg_load;
+                       sds->this = sg;
+                       sds->this_nr_running = sgs.sum_nr_running;
+                       sds->this_load_per_task = sgs.sum_weighted_load;
+                       sds->this_has_capacity = sgs.group_has_capacity;
+                       sds->this_idle_cpus = sgs.idle_cpus;
+               } else if (update_sd_pick_busiest(sd, sds, sg, &sgs, this_cpu)) {
+                       sds->max_load = sgs.avg_load;
+                       sds->busiest = sg;
+                       sds->busiest_nr_running = sgs.sum_nr_running;
+                       sds->busiest_idle_cpus = sgs.idle_cpus;
+                       sds->busiest_group_capacity = sgs.group_capacity;
+                       sds->busiest_load_per_task = sgs.sum_weighted_load;
+                       sds->busiest_has_capacity = sgs.group_has_capacity;
+                       sds->busiest_group_weight = sgs.group_weight;
+                       sds->group_imb = sgs.group_imb;
+               }
+
+               update_sd_power_savings_stats(sg, sds, local_group, &sgs);
+               sg = sg->next;
+       } while (sg != sd->groups);
+}
+
+int __weak arch_sd_sibling_asym_packing(void)
+{
+       return 0*SD_ASYM_PACKING;
+}
+
+/**
+ * check_asym_packing - Check to see if the group is packed into the
+ *                     sched doman.
+ *
+ * This is primarily intended to used at the sibling level.  Some
+ * cores like POWER7 prefer to use lower numbered SMT threads.  In the
+ * case of POWER7, it can move to lower SMT modes only when higher
+ * threads are idle.  When in lower SMT modes, the threads will
+ * perform better since they share less core resources.  Hence when we
+ * have idle threads, we want them to be the higher ones.
+ *
+ * This packing function is run on idle threads.  It checks to see if
+ * the busiest CPU in this domain (core in the P7 case) has a higher
+ * CPU number than the packing function is being run on.  Here we are
+ * assuming lower CPU number will be equivalent to lower a SMT thread
+ * number.
+ *
+ * Returns 1 when packing is required and a task should be moved to
+ * this CPU.  The amount of the imbalance is returned in *imbalance.
+ *
+ * @sd: The sched_domain whose packing is to be checked.
+ * @sds: Statistics of the sched_domain which is to be packed
+ * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
+ * @imbalance: returns amount of imbalanced due to packing.
+ */
+static int check_asym_packing(struct sched_domain *sd,
+                             struct sd_lb_stats *sds,
+                             int this_cpu, unsigned long *imbalance)
+{
+       int busiest_cpu;
+
+       if (!(sd->flags & SD_ASYM_PACKING))
+               return 0;
+
+       if (!sds->busiest)
+               return 0;
+
+       busiest_cpu = group_first_cpu(sds->busiest);
+       if (this_cpu > busiest_cpu)
+               return 0;
+
+       *imbalance = DIV_ROUND_CLOSEST(sds->max_load * sds->busiest->sgp->power,
+                                      SCHED_POWER_SCALE);
+       return 1;
+}
+
+/**
+ * fix_small_imbalance - Calculate the minor imbalance that exists
+ *                     amongst the groups of a sched_domain, during
+ *                     load balancing.
+ * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
+ * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
+ * @imbalance: Variable to store the imbalance.
+ */
+static inline void fix_small_imbalance(struct sd_lb_stats *sds,
+                               int this_cpu, unsigned long *imbalance)
+{
+       unsigned long tmp, pwr_now = 0, pwr_move = 0;
+       unsigned int imbn = 2;
+       unsigned long scaled_busy_load_per_task;
+
+       if (sds->this_nr_running) {
+               sds->this_load_per_task /= sds->this_nr_running;
+               if (sds->busiest_load_per_task >
+                               sds->this_load_per_task)
+                       imbn = 1;
+       } else
+               sds->this_load_per_task =
+                       cpu_avg_load_per_task(this_cpu);
+
+       scaled_busy_load_per_task = sds->busiest_load_per_task
+                                        * SCHED_POWER_SCALE;
+       scaled_busy_load_per_task /= sds->busiest->sgp->power;
+
+       if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
+                       (scaled_busy_load_per_task * imbn)) {
+               *imbalance = sds->busiest_load_per_task;
+               return;
+       }
+
+       /*
+        * OK, we don't have enough imbalance to justify moving tasks,
+        * however we may be able to increase total CPU power used by
+        * moving them.
+        */
+
+       pwr_now += sds->busiest->sgp->power *
+                       min(sds->busiest_load_per_task, sds->max_load);
+       pwr_now += sds->this->sgp->power *
+                       min(sds->this_load_per_task, sds->this_load);
+       pwr_now /= SCHED_POWER_SCALE;
+
+       /* Amount of load we'd subtract */
+       tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
+               sds->busiest->sgp->power;
+       if (sds->max_load > tmp)
+               pwr_move += sds->busiest->sgp->power *
+                       min(sds->busiest_load_per_task, sds->max_load - tmp);
+
+       /* Amount of load we'd add */
+       if (sds->max_load * sds->busiest->sgp->power <
+               sds->busiest_load_per_task * SCHED_POWER_SCALE)
+               tmp = (sds->max_load * sds->busiest->sgp->power) /
+                       sds->this->sgp->power;
+       else
+               tmp = (sds->busiest_load_per_task * SCHED_POWER_SCALE) /
+                       sds->this->sgp->power;
+       pwr_move += sds->this->sgp->power *
+                       min(sds->this_load_per_task, sds->this_load + tmp);
+       pwr_move /= SCHED_POWER_SCALE;
+
+       /* Move if we gain throughput */
+       if (pwr_move > pwr_now)
+               *imbalance = sds->busiest_load_per_task;
+}
+
+/**
+ * calculate_imbalance - Calculate the amount of imbalance present within the
+ *                      groups of a given sched_domain during load balance.
+ * @sds: statistics of the sched_domain whose imbalance is to be calculated.
+ * @this_cpu: Cpu for which currently load balance is being performed.
+ * @imbalance: The variable to store the imbalance.
+ */
+static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
+               unsigned long *imbalance)
 {
-       struct cfs_rq *busy_cfs_rq;
-       struct rq_iterator cfs_rq_iterator;
+       unsigned long max_pull, load_above_capacity = ~0UL;
+
+       sds->busiest_load_per_task /= sds->busiest_nr_running;
+       if (sds->group_imb) {
+               sds->busiest_load_per_task =
+                       min(sds->busiest_load_per_task, sds->avg_load);
+       }
+
+       /*
+        * In the presence of smp nice balancing, certain scenarios can have
+        * max load less than avg load(as we skip the groups at or below
+        * its cpu_power, while calculating max_load..)
+        */
+       if (sds->max_load < sds->avg_load) {
+               *imbalance = 0;
+               return fix_small_imbalance(sds, this_cpu, imbalance);
+       }
+
+       if (!sds->group_imb) {
+               /*
+                * Don't want to pull so many tasks that a group would go idle.
+                */
+               load_above_capacity = (sds->busiest_nr_running -
+                                               sds->busiest_group_capacity);
+
+               load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_POWER_SCALE);
+
+               load_above_capacity /= sds->busiest->sgp->power;
+       }
+
+       /*
+        * We're trying to get all the cpus to the average_load, so we don't
+        * want to push ourselves above the average load, nor do we wish to
+        * reduce the max loaded cpu below the average load. At the same time,
+        * we also don't want to reduce the group load below the group capacity
+        * (so that we can implement power-savings policies etc). Thus we look
+        * for the minimum possible imbalance.
+        * Be careful of negative numbers as they'll appear as very large values
+        * with unsigned longs.
+        */
+       max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
+
+       /* How much load to actually move to equalise the imbalance */
+       *imbalance = min(max_pull * sds->busiest->sgp->power,
+               (sds->avg_load - sds->this_load) * sds->this->sgp->power)
+                       / SCHED_POWER_SCALE;
+
+       /*
+        * if *imbalance is less than the average load per runnable task
+        * there is no guarantee that any tasks will be moved so we'll have
+        * a think about bumping its value to force at least one task to be
+        * moved
+        */
+       if (*imbalance < sds->busiest_load_per_task)
+               return fix_small_imbalance(sds, this_cpu, imbalance);
+
+}
+
+/******* find_busiest_group() helpers end here *********************/
+
+/**
+ * find_busiest_group - Returns the busiest group within the sched_domain
+ * if there is an imbalance. If there isn't an imbalance, and
+ * the user has opted for power-savings, it returns a group whose
+ * CPUs can be put to idle by rebalancing those tasks elsewhere, if
+ * such a group exists.
+ *
+ * Also calculates the amount of weighted load which should be moved
+ * to restore balance.
+ *
+ * @sd: The sched_domain whose busiest group is to be returned.
+ * @this_cpu: The cpu for which load balancing is currently being performed.
+ * @imbalance: Variable which stores amount of weighted load which should
+ *             be moved to restore balance/put a group to idle.
+ * @idle: The idle status of this_cpu.
+ * @cpus: The set of CPUs under consideration for load-balancing.
+ * @balance: Pointer to a variable indicating if this_cpu
+ *     is the appropriate cpu to perform load balancing at this_level.
+ *
+ * Returns:    - the busiest group if imbalance exists.
+ *             - If no imbalance and user has opted for power-savings balance,
+ *                return the least loaded group whose CPUs can be
+ *                put to idle by rebalancing its tasks onto our group.
+ */
+static struct sched_group *
+find_busiest_group(struct sched_domain *sd, int this_cpu,
+                  unsigned long *imbalance, enum cpu_idle_type idle,
+                  const struct cpumask *cpus, int *balance)
+{
+       struct sd_lb_stats sds;
+
+       memset(&sds, 0, sizeof(sds));
+
+       /*
+        * Compute the various statistics relavent for load balancing at
+        * this level.
+        */
+       update_sd_lb_stats(sd, this_cpu, idle, cpus, balance, &sds);
+
+       /*
+        * this_cpu is not the appropriate cpu to perform load balancing at
+        * this level.
+        */
+       if (!(*balance))
+               goto ret;
+
+       if ((idle == CPU_IDLE || idle == CPU_NEWLY_IDLE) &&
+           check_asym_packing(sd, &sds, this_cpu, imbalance))
+               return sds.busiest;
+
+       /* There is no busy sibling group to pull tasks from */
+       if (!sds.busiest || sds.busiest_nr_running == 0)
+               goto out_balanced;
+
+       sds.avg_load = (SCHED_POWER_SCALE * sds.total_load) / sds.total_pwr;
+
+       /*
+        * If the busiest group is imbalanced the below checks don't
+        * work because they assumes all things are equal, which typically
+        * isn't true due to cpus_allowed constraints and the like.
+        */
+       if (sds.group_imb)
+               goto force_balance;
+
+       /* SD_BALANCE_NEWIDLE trumps SMP nice when underutilized */
+       if (idle == CPU_NEWLY_IDLE && sds.this_has_capacity &&
+                       !sds.busiest_has_capacity)
+               goto force_balance;
 
-       cfs_rq_iterator.start = load_balance_start_fair;
-       cfs_rq_iterator.next = load_balance_next_fair;
+       /*
+        * If the local group is more busy than the selected busiest group
+        * don't try and pull any tasks.
+        */
+       if (sds.this_load >= sds.max_load)
+               goto out_balanced;
 
-       for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
+       /*
+        * Don't pull any tasks if this group is already above the domain
+        * average load.
+        */
+       if (sds.this_load >= sds.avg_load)
+               goto out_balanced;
+
+       if (idle == CPU_IDLE) {
                /*
-                * pass busy_cfs_rq argument into
-                * load_balance_[start|next]_fair iterators
+                * This cpu is idle. If the busiest group load doesn't
+                * have more tasks than the number of available cpu's and
+                * there is no imbalance between this and busiest group
+                * wrt to idle cpu's, it is balanced.
                 */
-               cfs_rq_iterator.arg = busy_cfs_rq;
-               if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
-                                      &cfs_rq_iterator))
-                   return 1;
+               if ((sds.this_idle_cpus <= sds.busiest_idle_cpus + 1) &&
+                   sds.busiest_nr_running <= sds.busiest_group_weight)
+                       goto out_balanced;
+       } else {
+               /*
+                * In the CPU_NEWLY_IDLE, CPU_NOT_IDLE cases, use
+                * imbalance_pct to be conservative.
+                */
+               if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
+                       goto out_balanced;
        }
 
+force_balance:
+       /* Looks like there is an imbalance. Compute it */
+       calculate_imbalance(&sds, this_cpu, imbalance);
+       return sds.busiest;
+
+out_balanced:
+       /*
+        * There is no obvious imbalance. But check if we can do some balancing
+        * to save power.
+        */
+       if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
+               return sds.busiest;
+ret:
+       *imbalance = 0;
+       return NULL;
+}
+
+/*
+ * find_busiest_queue - find the busiest runqueue among the cpus in group.
+ */
+static struct rq *
+find_busiest_queue(struct sched_domain *sd, struct sched_group *group,
+                  enum cpu_idle_type idle, unsigned long imbalance,
+                  const struct cpumask *cpus)
+{
+       struct rq *busiest = NULL, *rq;
+       unsigned long max_load = 0;
+       int i;
+
+       for_each_cpu(i, sched_group_cpus(group)) {
+               unsigned long power = power_of(i);
+               unsigned long capacity = DIV_ROUND_CLOSEST(power,
+                                                          SCHED_POWER_SCALE);
+               unsigned long wl;
+
+               if (!capacity)
+                       capacity = fix_small_capacity(sd, group);
+
+               if (!cpumask_test_cpu(i, cpus))
+                       continue;
+
+               rq = cpu_rq(i);
+               wl = weighted_cpuload(i);
+
+               /*
+                * When comparing with imbalance, use weighted_cpuload()
+                * which is not scaled with the cpu power.
+                */
+               if (capacity && rq->nr_running == 1 && wl > imbalance)
+                       continue;
+
+               /*
+                * For the load comparisons with the other cpu's, consider
+                * the weighted_cpuload() scaled with the cpu power, so that
+                * the load can be moved away from the cpu that is potentially
+                * running at a lower capacity.
+                */
+               wl = (wl * SCHED_POWER_SCALE) / power;
+
+               if (wl > max_load) {
+                       max_load = wl;
+                       busiest = rq;
+               }
+       }
+
+       return busiest;
+}
+
+/*
+ * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
+ * so long as it is large enough.
+ */
+#define MAX_PINNED_INTERVAL    512
+
+/* Working cpumask for load_balance and load_balance_newidle. */
+static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
+
+static int need_active_balance(struct sched_domain *sd, int idle,
+                              int busiest_cpu, int this_cpu)
+{
+       if (idle == CPU_NEWLY_IDLE) {
+
+               /*
+                * ASYM_PACKING needs to force migrate tasks from busy but
+                * higher numbered CPUs in order to pack all tasks in the
+                * lowest numbered CPUs.
+                */
+               if ((sd->flags & SD_ASYM_PACKING) && busiest_cpu > this_cpu)
+                       return 1;
+
+               /*
+                * The only task running in a non-idle cpu can be moved to this
+                * cpu in an attempt to completely freeup the other CPU
+                * package.
+                *
+                * The package power saving logic comes from
+                * find_busiest_group(). If there are no imbalance, then
+                * f_b_g() will return NULL. However when sched_mc={1,2} then
+                * f_b_g() will select a group from which a running task may be
+                * pulled to this cpu in order to make the other package idle.
+                * If there is no opportunity to make a package idle and if
+                * there are no imbalance, then f_b_g() will return NULL and no
+                * action will be taken in load_balance_newidle().
+                *
+                * Under normal task pull operation due to imbalance, there
+                * will be more than one task in the source run queue and
+                * move_tasks() will succeed.  ld_moved will be true and this
+                * active balance code will not be triggered.
+                */
+               if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
+                       return 0;
+       }
+
+       return unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2);
+}
+
+static int active_load_balance_cpu_stop(void *data);
+
+/*
+ * Check this_cpu to ensure it is balanced within domain. Attempt to move
+ * tasks if there is an imbalance.
+ */
+static int load_balance(int this_cpu, struct rq *this_rq,
+                       struct sched_domain *sd, enum cpu_idle_type idle,
+                       int *balance)
+{
+       int ld_moved, all_pinned = 0, active_balance = 0;
+       struct sched_group *group;
+       unsigned long imbalance;
+       struct rq *busiest;
+       unsigned long flags;
+       struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
+
+       cpumask_copy(cpus, cpu_active_mask);
+
+       schedstat_inc(sd, lb_count[idle]);
+
+redo:
+       group = find_busiest_group(sd, this_cpu, &imbalance, idle,
+                                  cpus, balance);
+
+       if (*balance == 0)
+               goto out_balanced;
+
+       if (!group) {
+               schedstat_inc(sd, lb_nobusyg[idle]);
+               goto out_balanced;
+       }
+
+       busiest = find_busiest_queue(sd, group, idle, imbalance, cpus);
+       if (!busiest) {
+               schedstat_inc(sd, lb_nobusyq[idle]);
+               goto out_balanced;
+       }
+
+       BUG_ON(busiest == this_rq);
+
+       schedstat_add(sd, lb_imbalance[idle], imbalance);
+
+       ld_moved = 0;
+       if (busiest->nr_running > 1) {
+               /*
+                * Attempt to move tasks. If find_busiest_group has found
+                * an imbalance but busiest->nr_running <= 1, the group is
+                * still unbalanced. ld_moved simply stays zero, so it is
+                * correctly treated as an imbalance.
+                */
+               all_pinned = 1;
+               local_irq_save(flags);
+               double_rq_lock(this_rq, busiest);
+               ld_moved = move_tasks(this_rq, this_cpu, busiest,
+                                     imbalance, sd, idle, &all_pinned);
+               double_rq_unlock(this_rq, busiest);
+               local_irq_restore(flags);
+
+               /*
+                * some other cpu did the load balance for us.
+                */
+               if (ld_moved && this_cpu != smp_processor_id())
+                       resched_cpu(this_cpu);
+
+               /* All tasks on this runqueue were pinned by CPU affinity */
+               if (unlikely(all_pinned)) {
+                       cpumask_clear_cpu(cpu_of(busiest), cpus);
+                       if (!cpumask_empty(cpus))
+                               goto redo;
+                       goto out_balanced;
+               }
+       }
+
+       if (!ld_moved) {
+               schedstat_inc(sd, lb_failed[idle]);
+               /*
+                * Increment the failure counter only on periodic balance.
+                * We do not want newidle balance, which can be very
+                * frequent, pollute the failure counter causing
+                * excessive cache_hot migrations and active balances.
+                */
+               if (idle != CPU_NEWLY_IDLE)
+                       sd->nr_balance_failed++;
+
+               if (need_active_balance(sd, idle, cpu_of(busiest), this_cpu)) {
+                       raw_spin_lock_irqsave(&busiest->lock, flags);
+
+                       /* don't kick the active_load_balance_cpu_stop,
+                        * if the curr task on busiest cpu can't be
+                        * moved to this_cpu
+                        */
+                       if (!cpumask_test_cpu(this_cpu,
+                                             &busiest->curr->cpus_allowed)) {
+                               raw_spin_unlock_irqrestore(&busiest->lock,
+                                                           flags);
+                               all_pinned = 1;
+                               goto out_one_pinned;
+                       }
+
+                       /*
+                        * ->active_balance synchronizes accesses to
+                        * ->active_balance_work.  Once set, it's cleared
+                        * only after active load balance is finished.
+                        */
+                       if (!busiest->active_balance) {
+                               busiest->active_balance = 1;
+                               busiest->push_cpu = this_cpu;
+                               active_balance = 1;
+                       }
+                       raw_spin_unlock_irqrestore(&busiest->lock, flags);
+
+                       if (active_balance)
+                               stop_one_cpu_nowait(cpu_of(busiest),
+                                       active_load_balance_cpu_stop, busiest,
+                                       &busiest->active_balance_work);
+
+                       /*
+                        * We've kicked active balancing, reset the failure
+                        * counter.
+                        */
+                       sd->nr_balance_failed = sd->cache_nice_tries+1;
+               }
+       } else
+               sd->nr_balance_failed = 0;
+
+       if (likely(!active_balance)) {
+               /* We were unbalanced, so reset the balancing interval */
+               sd->balance_interval = sd->min_interval;
+       } else {
+               /*
+                * If we've begun active balancing, start to back off. This
+                * case may not be covered by the all_pinned logic if there
+                * is only 1 task on the busy runqueue (because we don't call
+                * move_tasks).
+                */
+               if (sd->balance_interval < sd->max_interval)
+                       sd->balance_interval *= 2;
+       }
+
+       goto out;
+
+out_balanced:
+       schedstat_inc(sd, lb_balanced[idle]);
+
+       sd->nr_balance_failed = 0;
+
+out_one_pinned:
+       /* tune up the balancing interval */
+       if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
+                       (sd->balance_interval < sd->max_interval))
+               sd->balance_interval *= 2;
+
+       ld_moved = 0;
+out:
+       return ld_moved;
+}
+
+/*
+ * idle_balance is called by schedule() if this_cpu is about to become
+ * idle. Attempts to pull tasks from other CPUs.
+ */
+static void idle_balance(int this_cpu, struct rq *this_rq)
+{
+       struct sched_domain *sd;
+       int pulled_task = 0;
+       unsigned long next_balance = jiffies + HZ;
+
+       this_rq->idle_stamp = this_rq->clock;
+
+       if (this_rq->avg_idle < sysctl_sched_migration_cost)
+               return;
+
+       /*
+        * Drop the rq->lock, but keep IRQ/preempt disabled.
+        */
+       raw_spin_unlock(&this_rq->lock);
+
+       update_shares(this_cpu);
+       rcu_read_lock();
+       for_each_domain(this_cpu, sd) {
+               unsigned long interval;
+               int balance = 1;
+
+               if (!(sd->flags & SD_LOAD_BALANCE))
+                       continue;
+
+               if (sd->flags & SD_BALANCE_NEWIDLE) {
+                       /* If we've pulled tasks over stop searching: */
+                       pulled_task = load_balance(this_cpu, this_rq,
+                                                  sd, CPU_NEWLY_IDLE, &balance);
+               }
+
+               interval = msecs_to_jiffies(sd->balance_interval);
+               if (time_after(next_balance, sd->last_balance + interval))
+                       next_balance = sd->last_balance + interval;
+               if (pulled_task) {
+                       this_rq->idle_stamp = 0;
+                       break;
+               }
+       }
+       rcu_read_unlock();
+
+       raw_spin_lock(&this_rq->lock);
+
+       if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
+               /*
+                * We are going idle. next_balance may be set based on
+                * a busy processor. So reset next_balance.
+                */
+               this_rq->next_balance = next_balance;
+       }
+}
+
+/*
+ * active_load_balance_cpu_stop is run by cpu stopper. It pushes
+ * running tasks off the busiest CPU onto idle CPUs. It requires at
+ * least 1 task to be running on each physical CPU where possible, and
+ * avoids physical / logical imbalances.
+ */
+static int active_load_balance_cpu_stop(void *data)
+{
+       struct rq *busiest_rq = data;
+       int busiest_cpu = cpu_of(busiest_rq);
+       int target_cpu = busiest_rq->push_cpu;
+       struct rq *target_rq = cpu_rq(target_cpu);
+       struct sched_domain *sd;
+
+       raw_spin_lock_irq(&busiest_rq->lock);
+
+       /* make sure the requested cpu hasn't gone down in the meantime */
+       if (unlikely(busiest_cpu != smp_processor_id() ||
+                    !busiest_rq->active_balance))
+               goto out_unlock;
+
+       /* Is there any task to move? */
+       if (busiest_rq->nr_running <= 1)
+               goto out_unlock;
+
+       /*
+        * This condition is "impossible", if it occurs
+        * we need to fix it. Originally reported by
+        * Bjorn Helgaas on a 128-cpu setup.
+        */
+       BUG_ON(busiest_rq == target_rq);
+
+       /* move a task from busiest_rq to target_rq */
+       double_lock_balance(busiest_rq, target_rq);
+
+       /* Search for an sd spanning us and the target CPU. */
+       rcu_read_lock();
+       for_each_domain(target_cpu, sd) {
+               if ((sd->flags & SD_LOAD_BALANCE) &&
+                   cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
+                               break;
+       }
+
+       if (likely(sd)) {
+               schedstat_inc(sd, alb_count);
+
+               if (move_one_task(target_rq, target_cpu, busiest_rq,
+                                 sd, CPU_IDLE))
+                       schedstat_inc(sd, alb_pushed);
+               else
+                       schedstat_inc(sd, alb_failed);
+       }
+       rcu_read_unlock();
+       double_unlock_balance(busiest_rq, target_rq);
+out_unlock:
+       busiest_rq->active_balance = 0;
+       raw_spin_unlock_irq(&busiest_rq->lock);
        return 0;
 }
+
+#ifdef CONFIG_NO_HZ
+
+static DEFINE_PER_CPU(struct call_single_data, remote_sched_softirq_cb);
+
+static void trigger_sched_softirq(void *data)
+{
+       raise_softirq_irqoff(SCHED_SOFTIRQ);
+}
+
+static inline void init_sched_softirq_csd(struct call_single_data *csd)
+{
+       csd->func = trigger_sched_softirq;
+       csd->info = NULL;
+       csd->flags = 0;
+       csd->priv = 0;
+}
+
+/*
+ * idle load balancing details
+ * - One of the idle CPUs nominates itself as idle load_balancer, while
+ *   entering idle.
+ * - This idle load balancer CPU will also go into tickless mode when
+ *   it is idle, just like all other idle CPUs
+ * - When one of the busy CPUs notice that there may be an idle rebalancing
+ *   needed, they will kick the idle load balancer, which then does idle
+ *   load balancing for all the idle CPUs.
+ */
+static struct {
+       atomic_t load_balancer;
+       atomic_t first_pick_cpu;
+       atomic_t second_pick_cpu;
+       cpumask_var_t idle_cpus_mask;
+       cpumask_var_t grp_idle_mask;
+       unsigned long next_balance;     /* in jiffy units */
+} nohz ____cacheline_aligned;
+
+int get_nohz_load_balancer(void)
+{
+       return atomic_read(&nohz.load_balancer);
+}
+
+#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
+/**
+ * lowest_flag_domain - Return lowest sched_domain containing flag.
+ * @cpu:       The cpu whose lowest level of sched domain is to
+ *             be returned.
+ * @flag:      The flag to check for the lowest sched_domain
+ *             for the given cpu.
+ *
+ * Returns the lowest sched_domain of a cpu which contains the given flag.
+ */
+static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
+{
+       struct sched_domain *sd;
+
+       for_each_domain(cpu, sd)
+               if (sd && (sd->flags & flag))
+                       break;
+
+       return sd;
+}
+
+/**
+ * for_each_flag_domain - Iterates over sched_domains containing the flag.
+ * @cpu:       The cpu whose domains we're iterating over.
+ * @sd:                variable holding the value of the power_savings_sd
+ *             for cpu.
+ * @flag:      The flag to filter the sched_domains to be iterated.
+ *
+ * Iterates over all the scheduler domains for a given cpu that has the 'flag'
+ * set, starting from the lowest sched_domain to the highest.
+ */
+#define for_each_flag_domain(cpu, sd, flag) \
+       for (sd = lowest_flag_domain(cpu, flag); \
+               (sd && (sd->flags & flag)); sd = sd->parent)
+
+/**
+ * is_semi_idle_group - Checks if the given sched_group is semi-idle.
+ * @ilb_group: group to be checked for semi-idleness
+ *
+ * Returns:    1 if the group is semi-idle. 0 otherwise.
+ *
+ * We define a sched_group to be semi idle if it has atleast one idle-CPU
+ * and atleast one non-idle CPU. This helper function checks if the given
+ * sched_group is semi-idle or not.
+ */
+static inline int is_semi_idle_group(struct sched_group *ilb_group)
+{
+       cpumask_and(nohz.grp_idle_mask, nohz.idle_cpus_mask,
+                                       sched_group_cpus(ilb_group));
+
+       /*
+        * A sched_group is semi-idle when it has atleast one busy cpu
+        * and atleast one idle cpu.
+        */
+       if (cpumask_empty(nohz.grp_idle_mask))
+               return 0;
+
+       if (cpumask_equal(nohz.grp_idle_mask, sched_group_cpus(ilb_group)))
+               return 0;
+
+       return 1;
+}
+/**
+ * find_new_ilb - Finds the optimum idle load balancer for nomination.
+ * @cpu:       The cpu which is nominating a new idle_load_balancer.
+ *
+ * Returns:    Returns the id of the idle load balancer if it exists,
+ *             Else, returns >= nr_cpu_ids.
+ *
+ * This algorithm picks the idle load balancer such that it belongs to a
+ * semi-idle powersavings sched_domain. The idea is to try and avoid
+ * completely idle packages/cores just for the purpose of idle load balancing
+ * when there are other idle cpu's which are better suited for that job.
+ */
+static int find_new_ilb(int cpu)
+{
+       struct sched_domain *sd;
+       struct sched_group *ilb_group;
+       int ilb = nr_cpu_ids;
+
+       /*
+        * Have idle load balancer selection from semi-idle packages only
+        * when power-aware load balancing is enabled
+        */
+       if (!(sched_smt_power_savings || sched_mc_power_savings))
+               goto out_done;
+
+       /*
+        * Optimize for the case when we have no idle CPUs or only one
+        * idle CPU. Don't walk the sched_domain hierarchy in such cases
+        */
+       if (cpumask_weight(nohz.idle_cpus_mask) < 2)
+               goto out_done;
+
+       rcu_read_lock();
+       for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
+               ilb_group = sd->groups;
+
+               do {
+                       if (is_semi_idle_group(ilb_group)) {
+                               ilb = cpumask_first(nohz.grp_idle_mask);
+                               goto unlock;
+                       }
+
+                       ilb_group = ilb_group->next;
+
+               } while (ilb_group != sd->groups);
+       }
+unlock:
+       rcu_read_unlock();
+
+out_done:
+       return ilb;
+}
+#else /*  (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
+static inline int find_new_ilb(int call_cpu)
+{
+       return nr_cpu_ids;
+}
+#endif
+
+/*
+ * Kick a CPU to do the nohz balancing, if it is time for it. We pick the
+ * nohz_load_balancer CPU (if there is one) otherwise fallback to any idle
+ * CPU (if there is one).
+ */
+static void nohz_balancer_kick(int cpu)
+{
+       int ilb_cpu;
+
+       nohz.next_balance++;
+
+       ilb_cpu = get_nohz_load_balancer();
+
+       if (ilb_cpu >= nr_cpu_ids) {
+               ilb_cpu = cpumask_first(nohz.idle_cpus_mask);
+               if (ilb_cpu >= nr_cpu_ids)
+                       return;
+       }
+
+       if (!cpu_rq(ilb_cpu)->nohz_balance_kick) {
+               struct call_single_data *cp;
+
+               cpu_rq(ilb_cpu)->nohz_balance_kick = 1;
+               cp = &per_cpu(remote_sched_softirq_cb, cpu);
+               __smp_call_function_single(ilb_cpu, cp, 0);
+       }
+       return;
+}
+
+/*
+ * This routine will try to nominate the ilb (idle load balancing)
+ * owner among the cpus whose ticks are stopped. ilb owner will do the idle
+ * load balancing on behalf of all those cpus.
+ *
+ * When the ilb owner becomes busy, we will not have new ilb owner until some
+ * idle CPU wakes up and goes back to idle or some busy CPU tries to kick
+ * idle load balancing by kicking one of the idle CPUs.
+ *
+ * Ticks are stopped for the ilb owner as well, with busy CPU kicking this
+ * ilb owner CPU in future (when there is a need for idle load balancing on
+ * behalf of all idle CPUs).
+ */
+void select_nohz_load_balancer(int stop_tick)
+{
+       int cpu = smp_processor_id();
+
+       if (stop_tick) {
+               if (!cpu_active(cpu)) {
+                       if (atomic_read(&nohz.load_balancer) != cpu)
+                               return;
+
+                       /*
+                        * If we are going offline and still the leader,
+                        * give up!
+                        */
+                       if (atomic_cmpxchg(&nohz.load_balancer, cpu,
+                                          nr_cpu_ids) != cpu)
+                               BUG();
+
+                       return;
+               }
+
+               cpumask_set_cpu(cpu, nohz.idle_cpus_mask);
+
+               if (atomic_read(&nohz.first_pick_cpu) == cpu)
+                       atomic_cmpxchg(&nohz.first_pick_cpu, cpu, nr_cpu_ids);
+               if (atomic_read(&nohz.second_pick_cpu) == cpu)
+                       atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
+
+               if (atomic_read(&nohz.load_balancer) >= nr_cpu_ids) {
+                       int new_ilb;
+
+                       /* make me the ilb owner */
+                       if (atomic_cmpxchg(&nohz.load_balancer, nr_cpu_ids,
+                                          cpu) != nr_cpu_ids)
+                               return;
+
+                       /*
+                        * Check to see if there is a more power-efficient
+                        * ilb.
+                        */
+                       new_ilb = find_new_ilb(cpu);
+                       if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
+                               atomic_set(&nohz.load_balancer, nr_cpu_ids);
+                               resched_cpu(new_ilb);
+                               return;
+                       }
+                       return;
+               }
+       } else {
+               if (!cpumask_test_cpu(cpu, nohz.idle_cpus_mask))
+                       return;
+
+               cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
+
+               if (atomic_read(&nohz.load_balancer) == cpu)
+                       if (atomic_cmpxchg(&nohz.load_balancer, cpu,
+                                          nr_cpu_ids) != cpu)
+                               BUG();
+       }
+       return;
+}
+#endif
+
+static DEFINE_SPINLOCK(balancing);
+
+static unsigned long __read_mostly max_load_balance_interval = HZ/10;
+
+/*
+ * Scale the max load_balance interval with the number of CPUs in the system.
+ * This trades load-balance latency on larger machines for less cross talk.
+ */
+static void update_max_interval(void)
+{
+       max_load_balance_interval = HZ*num_online_cpus()/10;
+}
+
+/*
+ * It checks each scheduling domain to see if it is due to be balanced,
+ * and initiates a balancing operation if so.
+ *
+ * Balancing parameters are set up in arch_init_sched_domains.
+ */
+static void rebalance_domains(int cpu, enum cpu_idle_type idle)
+{
+       int balance = 1;
+       struct rq *rq = cpu_rq(cpu);
+       unsigned long interval;
+       struct sched_domain *sd;
+       /* Earliest time when we have to do rebalance again */
+       unsigned long next_balance = jiffies + 60*HZ;
+       int update_next_balance = 0;
+       int need_serialize;
+
+       update_shares(cpu);
+
+       rcu_read_lock();
+       for_each_domain(cpu, sd) {
+               if (!(sd->flags & SD_LOAD_BALANCE))
+                       continue;
+
+               interval = sd->balance_interval;
+               if (idle != CPU_IDLE)
+                       interval *= sd->busy_factor;
+
+               /* scale ms to jiffies */
+               interval = msecs_to_jiffies(interval);
+               interval = clamp(interval, 1UL, max_load_balance_interval);
+
+               need_serialize = sd->flags & SD_SERIALIZE;
+
+               if (need_serialize) {
+                       if (!spin_trylock(&balancing))
+                               goto out;
+               }
+
+               if (time_after_eq(jiffies, sd->last_balance + interval)) {
+                       if (load_balance(cpu, rq, sd, idle, &balance)) {
+                               /*
+                                * We've pulled tasks over so either we're no
+                                * longer idle.
+                                */
+                               idle = CPU_NOT_IDLE;
+                       }
+                       sd->last_balance = jiffies;
+               }
+               if (need_serialize)
+                       spin_unlock(&balancing);
+out:
+               if (time_after(next_balance, sd->last_balance + interval)) {
+                       next_balance = sd->last_balance + interval;
+                       update_next_balance = 1;
+               }
+
+               /*
+                * Stop the load balance at this level. There is another
+                * CPU in our sched group which is doing load balancing more
+                * actively.
+                */
+               if (!balance)
+                       break;
+       }
+       rcu_read_unlock();
+
+       /*
+        * next_balance will be updated only when there is a need.
+        * When the cpu is attached to null domain for ex, it will not be
+        * updated.
+        */
+       if (likely(update_next_balance))
+               rq->next_balance = next_balance;
+}
+
+#ifdef CONFIG_NO_HZ
+/*
+ * In CONFIG_NO_HZ case, the idle balance kickee will do the
+ * rebalancing for all the cpus for whom scheduler ticks are stopped.
+ */
+static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle)
+{
+       struct rq *this_rq = cpu_rq(this_cpu);
+       struct rq *rq;
+       int balance_cpu;
+
+       if (idle != CPU_IDLE || !this_rq->nohz_balance_kick)
+               return;
+
+       for_each_cpu(balance_cpu, nohz.idle_cpus_mask) {
+               if (balance_cpu == this_cpu)
+                       continue;
+
+               /*
+                * If this cpu gets work to do, stop the load balancing
+                * work being done for other cpus. Next load
+                * balancing owner will pick it up.
+                */
+               if (need_resched()) {
+                       this_rq->nohz_balance_kick = 0;
+                       break;
+               }
+
+               raw_spin_lock_irq(&this_rq->lock);
+               update_rq_clock(this_rq);
+               update_cpu_load(this_rq);
+               raw_spin_unlock_irq(&this_rq->lock);
+
+               rebalance_domains(balance_cpu, CPU_IDLE);
+
+               rq = cpu_rq(balance_cpu);
+               if (time_after(this_rq->next_balance, rq->next_balance))
+                       this_rq->next_balance = rq->next_balance;
+       }
+       nohz.next_balance = this_rq->next_balance;
+       this_rq->nohz_balance_kick = 0;
+}
+
+/*
+ * Current heuristic for kicking the idle load balancer
+ * - first_pick_cpu is the one of the busy CPUs. It will kick
+ *   idle load balancer when it has more than one process active. This
+ *   eliminates the need for idle load balancing altogether when we have
+ *   only one running process in the system (common case).
+ * - If there are more than one busy CPU, idle load balancer may have
+ *   to run for active_load_balance to happen (i.e., two busy CPUs are
+ *   SMT or core siblings and can run better if they move to different
+ *   physical CPUs). So, second_pick_cpu is the second of the busy CPUs
+ *   which will kick idle load balancer as soon as it has any load.
+ */
+static inline int nohz_kick_needed(struct rq *rq, int cpu)
+{
+       unsigned long now = jiffies;
+       int ret;
+       int first_pick_cpu, second_pick_cpu;
+
+       if (time_before(now, nohz.next_balance))
+               return 0;
+
+       if (rq->idle_at_tick)
+               return 0;
+
+       first_pick_cpu = atomic_read(&nohz.first_pick_cpu);
+       second_pick_cpu = atomic_read(&nohz.second_pick_cpu);
+
+       if (first_pick_cpu < nr_cpu_ids && first_pick_cpu != cpu &&
+           second_pick_cpu < nr_cpu_ids && second_pick_cpu != cpu)
+               return 0;
+
+       ret = atomic_cmpxchg(&nohz.first_pick_cpu, nr_cpu_ids, cpu);
+       if (ret == nr_cpu_ids || ret == cpu) {
+               atomic_cmpxchg(&nohz.second_pick_cpu, cpu, nr_cpu_ids);
+               if (rq->nr_running > 1)
+                       return 1;
+       } else {
+               ret = atomic_cmpxchg(&nohz.second_pick_cpu, nr_cpu_ids, cpu);
+               if (ret == nr_cpu_ids || ret == cpu) {
+                       if (rq->nr_running)
+                               return 1;
+               }
+       }
+       return 0;
+}
+#else
+static void nohz_idle_balance(int this_cpu, enum cpu_idle_type idle) { }
+#endif
+
+/*
+ * run_rebalance_domains is triggered when needed from the scheduler tick.
+ * Also triggered for nohz idle balancing (with nohz_balancing_kick set).
+ */
+static void run_rebalance_domains(struct softirq_action *h)
+{
+       int this_cpu = smp_processor_id();
+       struct rq *this_rq = cpu_rq(this_cpu);
+       enum cpu_idle_type idle = this_rq->idle_at_tick ?
+                                               CPU_IDLE : CPU_NOT_IDLE;
+
+       rebalance_domains(this_cpu, idle);
+
+       /*
+        * If this cpu has a pending nohz_balance_kick, then do the
+        * balancing on behalf of the other idle cpus whose ticks are
+        * stopped.
+        */
+       nohz_idle_balance(this_cpu, idle);
+}
+
+static inline int on_null_domain(int cpu)
+{
+       return !rcu_dereference_sched(cpu_rq(cpu)->sd);
+}
+
+/*
+ * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
+ */
+static inline void trigger_load_balance(struct rq *rq, int cpu)
+{
+       /* Don't need to rebalance while attached to NULL domain */
+       if (time_after_eq(jiffies, rq->next_balance) &&
+           likely(!on_null_domain(cpu)))
+               raise_softirq(SCHED_SOFTIRQ);
+#ifdef CONFIG_NO_HZ
+       else if (nohz_kick_needed(rq, cpu) && likely(!on_null_domain(cpu)))
+               nohz_balancer_kick(cpu);
+#endif
+}
+
+static void rq_online_fair(struct rq *rq)
+{
+       update_sysctl();
+}
+
+static void rq_offline_fair(struct rq *rq)
+{
+       update_sysctl();
+}
+
+#else  /* CONFIG_SMP */
+
+/*
+ * on UP we do not need to balance between CPUs:
+ */
+static inline void idle_balance(int cpu, struct rq *rq)
+{
+}
+
 #endif /* CONFIG_SMP */
 
 /*
@@ -1804,28 +4121,35 @@ static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
 }
 
 /*
- * Share the fairness runtime between parent and child, thus the
- * total amount of pressure for CPU stays equal - new tasks
- * get a chance to run but frequent forkers are not allowed to
- * monopolize the CPU. Note: the parent runqueue is locked,
- * the child is not running yet.
+ * called on fork with the child task as argument from the parent's context
+ *  - child not yet on the tasklist
+ *  - preemption disabled
  */
-static void task_new_fair(struct rq *rq, struct task_struct *p)
+static void task_fork_fair(struct task_struct *p)
 {
-       struct cfs_rq *cfs_rq = task_cfs_rq(p);
+       struct cfs_rq *cfs_rq = task_cfs_rq(current);
        struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
        int this_cpu = smp_processor_id();
+       struct rq *rq = this_rq();
+       unsigned long flags;
 
-       sched_info_queued(p);
+       raw_spin_lock_irqsave(&rq->lock, flags);
+
+       update_rq_clock(rq);
+
+       if (unlikely(task_cpu(p) != this_cpu)) {
+               rcu_read_lock();
+               __set_task_cpu(p, this_cpu);
+               rcu_read_unlock();
+       }
 
        update_curr(cfs_rq);
+
        if (curr)
                se->vruntime = curr->vruntime;
        place_entity(cfs_rq, se, 1);
 
-       /* 'curr' will be NULL if the child belongs to a different group */
-       if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) &&
-                       curr && entity_before(curr, se)) {
+       if (sysctl_sched_child_runs_first && curr && entity_before(curr, se)) {
                /*
                 * Upon rescheduling, sched_class::put_prev_task() will place
                 * 'current' within the tree based on its new key value.
@@ -1834,40 +4158,71 @@ static void task_new_fair(struct rq *rq, struct task_struct *p)
                resched_task(rq->curr);
        }
 
-       enqueue_task_fair(rq, p, 0);
+       se->vruntime -= cfs_rq->min_vruntime;
+
+       raw_spin_unlock_irqrestore(&rq->lock, flags);
 }
 
 /*
  * Priority of the task has changed. Check to see if we preempt
  * the current task.
  */
-static void prio_changed_fair(struct rq *rq, struct task_struct *p,
-                             int oldprio, int running)
+static void
+prio_changed_fair(struct rq *rq, struct task_struct *p, int oldprio)
 {
+       if (!p->se.on_rq)
+               return;
+
        /*
         * Reschedule if we are currently running on this runqueue and
         * our priority decreased, or if we are not currently running on
         * this runqueue and our priority is higher than the current's
         */
-       if (running) {
+       if (rq->curr == p) {
                if (p->prio > oldprio)
                        resched_task(rq->curr);
        } else
                check_preempt_curr(rq, p, 0);
 }
 
+static void switched_from_fair(struct rq *rq, struct task_struct *p)
+{
+       struct sched_entity *se = &p->se;
+       struct cfs_rq *cfs_rq = cfs_rq_of(se);
+
+       /*
+        * Ensure the task's vruntime is normalized, so that when its
+        * switched back to the fair class the enqueue_entity(.flags=0) will
+        * do the right thing.
+        *
+        * If it was on_rq, then the dequeue_entity(.flags=0) will already
+        * have normalized the vruntime, if it was !on_rq, then only when
+        * the task is sleeping will it still have non-normalized vruntime.
+        */
+       if (!se->on_rq && p->state != TASK_RUNNING) {
+               /*
+                * Fix up our vruntime so that the current sleep doesn't
+                * cause 'unlimited' sleep bonus.
+                */
+               place_entity(cfs_rq, se, 0);
+               se->vruntime -= cfs_rq->min_vruntime;
+       }
+}
+
 /*
  * We switched to the sched_fair class.
  */
-static void switched_to_fair(struct rq *rq, struct task_struct *p,
-                            int running)
+static void switched_to_fair(struct rq *rq, struct task_struct *p)
 {
+       if (!p->se.on_rq)
+               return;
+
        /*
         * We were most likely switched from sched_rt, so
         * kick off the schedule if running, otherwise just see
         * if we can still preempt the current task.
         */
-       if (running)
+       if (rq->curr == p)
                resched_task(rq->curr);
        else
                check_preempt_curr(rq, p, 0);
@@ -1887,15 +4242,44 @@ static void set_curr_task_fair(struct rq *rq)
 }
 
 #ifdef CONFIG_FAIR_GROUP_SCHED
-static void moved_group_fair(struct task_struct *p)
+static void task_move_group_fair(struct task_struct *p, int on_rq)
 {
-       struct cfs_rq *cfs_rq = task_cfs_rq(p);
-
-       update_curr(cfs_rq);
-       place_entity(cfs_rq, &p->se, 1);
+       /*
+        * If the task was not on the rq at the time of this cgroup movement
+        * it must have been asleep, sleeping tasks keep their ->vruntime
+        * absolute on their old rq until wakeup (needed for the fair sleeper
+        * bonus in place_entity()).
+        *
+        * If it was on the rq, we've just 'preempted' it, which does convert
+        * ->vruntime to a relative base.
+        *
+        * Make sure both cases convert their relative position when migrating
+        * to another cgroup's rq. This does somewhat interfere with the
+        * fair sleeper stuff for the first placement, but who cares.
+        */
+       if (!on_rq)
+               p->se.vruntime -= cfs_rq_of(&p->se)->min_vruntime;
+       set_task_rq(p, task_cpu(p));
+       if (!on_rq)
+               p->se.vruntime += cfs_rq_of(&p->se)->min_vruntime;
 }
 #endif
 
+static unsigned int get_rr_interval_fair(struct rq *rq, struct task_struct *task)
+{
+       struct sched_entity *se = &task->se;
+       unsigned int rr_interval = 0;
+
+       /*
+        * Time slice is 0 for SCHED_OTHER tasks that are on an otherwise
+        * idle runqueue:
+        */
+       if (rq->cfs.load.weight)
+               rr_interval = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
+
+       return rr_interval;
+}
+
 /*
  * All the scheduling class methods:
  */
@@ -1904,6 +4288,7 @@ static const struct sched_class fair_sched_class = {
        .enqueue_task           = enqueue_task_fair,
        .dequeue_task           = dequeue_task_fair,
        .yield_task             = yield_task_fair,
+       .yield_to_task          = yield_to_task_fair,
 
        .check_preempt_curr     = check_preempt_wakeup,
 
@@ -1913,19 +4298,24 @@ static const struct sched_class fair_sched_class = {
 #ifdef CONFIG_SMP
        .select_task_rq         = select_task_rq_fair,
 
-       .load_balance           = load_balance_fair,
-       .move_one_task          = move_one_task_fair,
+       .rq_online              = rq_online_fair,
+       .rq_offline             = rq_offline_fair,
+
+       .task_waking            = task_waking_fair,
 #endif
 
        .set_curr_task          = set_curr_task_fair,
        .task_tick              = task_tick_fair,
-       .task_new               = task_new_fair,
+       .task_fork              = task_fork_fair,
 
        .prio_changed           = prio_changed_fair,
+       .switched_from          = switched_from_fair,
        .switched_to            = switched_to_fair,
 
+       .get_rr_interval        = get_rr_interval_fair,
+
 #ifdef CONFIG_FAIR_GROUP_SCHED
-       .moved_group            = moved_group_fair,
+       .task_move_group        = task_move_group_fair,
 #endif
 };