sched: cfs core, kernel/sched_fair.c
authorIngo Molnar <mingo@elte.hu>
Mon, 9 Jul 2007 16:51:58 +0000 (18:51 +0200)
committerIngo Molnar <mingo@elte.hu>
Mon, 9 Jul 2007 16:51:58 +0000 (18:51 +0200)
add kernel/sched_fair.c - which implements the bulk of CFS's
behavioral changes for SCHED_OTHER tasks.

see Documentation/sched-design-CFS.txt about details.

Authors:

 Ingo Molnar <mingo@elte.hu>
 Dmitry Adamushko <dmitry.adamushko@gmail.com>
 Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
 Mike Galbraith <efault@gmx.de>

Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Dmitry Adamushko <dmitry.adamushko@gmail.com>
Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
include/linux/sched.h
kernel/sched_fair.c [new file with mode: 0644]

index d5084e7c48cfec17503ea17aa040d73ba40d0b76..90420321994f75db576e4b34900489e3b0b31ac3 100644 (file)
@@ -1243,6 +1243,14 @@ static inline void idle_task_exit(void) {}
 
 extern void sched_idle_next(void);
 
+extern unsigned int sysctl_sched_granularity;
+extern unsigned int sysctl_sched_wakeup_granularity;
+extern unsigned int sysctl_sched_batch_wakeup_granularity;
+extern unsigned int sysctl_sched_stat_granularity;
+extern unsigned int sysctl_sched_runtime_limit;
+extern unsigned int sysctl_sched_child_runs_first;
+extern unsigned int sysctl_sched_features;
+
 #ifdef CONFIG_RT_MUTEXES
 extern int rt_mutex_getprio(struct task_struct *p);
 extern void rt_mutex_setprio(struct task_struct *p, int prio);
diff --git a/kernel/sched_fair.c b/kernel/sched_fair.c
new file mode 100644 (file)
index 0000000..6971db0
--- /dev/null
@@ -0,0 +1,1131 @@
+/*
+ * Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
+ *
+ *  Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
+ *
+ *  Interactivity improvements by Mike Galbraith
+ *  (C) 2007 Mike Galbraith <efault@gmx.de>
+ *
+ *  Various enhancements by Dmitry Adamushko.
+ *  (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
+ *
+ *  Group scheduling enhancements by Srivatsa Vaddagiri
+ *  Copyright IBM Corporation, 2007
+ *  Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
+ *
+ *  Scaled math optimizations by Thomas Gleixner
+ *  Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
+ */
+
+/*
+ * Preemption granularity:
+ * (default: 2 msec, units: nanoseconds)
+ *
+ * NOTE: this granularity value is not the same as the concept of
+ * 'timeslice length' - timeslices in CFS will typically be somewhat
+ * larger than this value. (to see the precise effective timeslice
+ * length of your workload, run vmstat and monitor the context-switches
+ * field)
+ *
+ * On SMP systems the value of this is multiplied by the log2 of the
+ * number of CPUs. (i.e. factor 2x on 2-way systems, 3x on 4-way
+ * systems, 4x on 8-way systems, 5x on 16-way systems, etc.)
+ */
+unsigned int sysctl_sched_granularity __read_mostly = 2000000000ULL/HZ;
+
+/*
+ * SCHED_BATCH wake-up granularity.
+ * (default: 10 msec, units: nanoseconds)
+ *
+ * This option delays the preemption effects of decoupled workloads
+ * and reduces their over-scheduling. Synchronous workloads will still
+ * have immediate wakeup/sleep latencies.
+ */
+unsigned int sysctl_sched_batch_wakeup_granularity __read_mostly =
+                                                       10000000000ULL/HZ;
+
+/*
+ * SCHED_OTHER wake-up granularity.
+ * (default: 1 msec, units: nanoseconds)
+ *
+ * This option delays the preemption effects of decoupled workloads
+ * and reduces their over-scheduling. Synchronous workloads will still
+ * have immediate wakeup/sleep latencies.
+ */
+unsigned int sysctl_sched_wakeup_granularity __read_mostly = 1000000000ULL/HZ;
+
+unsigned int sysctl_sched_stat_granularity __read_mostly;
+
+/*
+ * Initialized in sched_init_granularity():
+ */
+unsigned int sysctl_sched_runtime_limit __read_mostly;
+
+/*
+ * Debugging: various feature bits
+ */
+enum {
+       SCHED_FEAT_FAIR_SLEEPERS        = 1,
+       SCHED_FEAT_SLEEPER_AVG          = 2,
+       SCHED_FEAT_SLEEPER_LOAD_AVG     = 4,
+       SCHED_FEAT_PRECISE_CPU_LOAD     = 8,
+       SCHED_FEAT_START_DEBIT          = 16,
+       SCHED_FEAT_SKIP_INITIAL         = 32,
+};
+
+unsigned int sysctl_sched_features __read_mostly =
+               SCHED_FEAT_FAIR_SLEEPERS        *1 |
+               SCHED_FEAT_SLEEPER_AVG          *1 |
+               SCHED_FEAT_SLEEPER_LOAD_AVG     *1 |
+               SCHED_FEAT_PRECISE_CPU_LOAD     *1 |
+               SCHED_FEAT_START_DEBIT          *1 |
+               SCHED_FEAT_SKIP_INITIAL         *0;
+
+extern struct sched_class fair_sched_class;
+
+/**************************************************************
+ * CFS operations on generic schedulable entities:
+ */
+
+#ifdef CONFIG_FAIR_GROUP_SCHED
+
+/* cpu runqueue to which this cfs_rq is attached */
+static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
+{
+       return cfs_rq->rq;
+}
+
+/* currently running entity (if any) on this cfs_rq */
+static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq)
+{
+       return cfs_rq->curr;
+}
+
+/* An entity is a task if it doesn't "own" a runqueue */
+#define entity_is_task(se)     (!se->my_q)
+
+static inline void
+set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se)
+{
+       cfs_rq->curr = se;
+}
+
+#else  /* CONFIG_FAIR_GROUP_SCHED */
+
+static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
+{
+       return container_of(cfs_rq, struct rq, cfs);
+}
+
+static inline struct sched_entity *cfs_rq_curr(struct cfs_rq *cfs_rq)
+{
+       struct rq *rq = rq_of(cfs_rq);
+
+       if (unlikely(rq->curr->sched_class != &fair_sched_class))
+               return NULL;
+
+       return &rq->curr->se;
+}
+
+#define entity_is_task(se)     1
+
+static inline void
+set_cfs_rq_curr(struct cfs_rq *cfs_rq, struct sched_entity *se) { }
+
+#endif /* CONFIG_FAIR_GROUP_SCHED */
+
+static inline struct task_struct *task_of(struct sched_entity *se)
+{
+       return container_of(se, struct task_struct, se);
+}
+
+
+/**************************************************************
+ * Scheduling class tree data structure manipulation methods:
+ */
+
+/*
+ * Enqueue an entity into the rb-tree:
+ */
+static inline void
+__enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
+{
+       struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
+       struct rb_node *parent = NULL;
+       struct sched_entity *entry;
+       s64 key = se->fair_key;
+       int leftmost = 1;
+
+       /*
+        * Find the right place in the rbtree:
+        */
+       while (*link) {
+               parent = *link;
+               entry = rb_entry(parent, struct sched_entity, run_node);
+               /*
+                * We dont care about collisions. Nodes with
+                * the same key stay together.
+                */
+               if (key - entry->fair_key < 0) {
+                       link = &parent->rb_left;
+               } else {
+                       link = &parent->rb_right;
+                       leftmost = 0;
+               }
+       }
+
+       /*
+        * Maintain a cache of leftmost tree entries (it is frequently
+        * used):
+        */
+       if (leftmost)
+               cfs_rq->rb_leftmost = &se->run_node;
+
+       rb_link_node(&se->run_node, parent, link);
+       rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
+       update_load_add(&cfs_rq->load, se->load.weight);
+       cfs_rq->nr_running++;
+       se->on_rq = 1;
+}
+
+static inline void
+__dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
+{
+       if (cfs_rq->rb_leftmost == &se->run_node)
+               cfs_rq->rb_leftmost = rb_next(&se->run_node);
+       rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
+       update_load_sub(&cfs_rq->load, se->load.weight);
+       cfs_rq->nr_running--;
+       se->on_rq = 0;
+}
+
+static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq)
+{
+       return cfs_rq->rb_leftmost;
+}
+
+static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
+{
+       return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node);
+}
+
+/**************************************************************
+ * Scheduling class statistics methods:
+ */
+
+/*
+ * We rescale the rescheduling granularity of tasks according to their
+ * nice level, but only linearly, not exponentially:
+ */
+static long
+niced_granularity(struct sched_entity *curr, unsigned long granularity)
+{
+       u64 tmp;
+
+       /*
+        * Negative nice levels get the same granularity as nice-0:
+        */
+       if (likely(curr->load.weight >= NICE_0_LOAD))
+               return granularity;
+       /*
+        * Positive nice level tasks get linearly finer
+        * granularity:
+        */
+       tmp = curr->load.weight * (u64)granularity;
+
+       /*
+        * It will always fit into 'long':
+        */
+       return (long) (tmp >> NICE_0_SHIFT);
+}
+
+static inline void
+limit_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se)
+{
+       long limit = sysctl_sched_runtime_limit;
+
+       /*
+        * Niced tasks have the same history dynamic range as
+        * non-niced tasks:
+        */
+       if (unlikely(se->wait_runtime > limit)) {
+               se->wait_runtime = limit;
+               schedstat_inc(se, wait_runtime_overruns);
+               schedstat_inc(cfs_rq, wait_runtime_overruns);
+       }
+       if (unlikely(se->wait_runtime < -limit)) {
+               se->wait_runtime = -limit;
+               schedstat_inc(se, wait_runtime_underruns);
+               schedstat_inc(cfs_rq, wait_runtime_underruns);
+       }
+}
+
+static inline void
+__add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
+{
+       se->wait_runtime += delta;
+       schedstat_add(se, sum_wait_runtime, delta);
+       limit_wait_runtime(cfs_rq, se);
+}
+
+static void
+add_wait_runtime(struct cfs_rq *cfs_rq, struct sched_entity *se, long delta)
+{
+       schedstat_add(cfs_rq, wait_runtime, -se->wait_runtime);
+       __add_wait_runtime(cfs_rq, se, delta);
+       schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
+}
+
+/*
+ * Update the current task's runtime statistics. Skip current tasks that
+ * are not in our scheduling class.
+ */
+static inline void
+__update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr, u64 now)
+{
+       unsigned long delta, delta_exec, delta_fair;
+       long delta_mine;
+       struct load_weight *lw = &cfs_rq->load;
+       unsigned long load = lw->weight;
+
+       if (unlikely(!load))
+               return;
+
+       delta_exec = curr->delta_exec;
+#ifdef CONFIG_SCHEDSTATS
+       if (unlikely(delta_exec > curr->exec_max))
+               curr->exec_max = delta_exec;
+#endif
+
+       curr->sum_exec_runtime += delta_exec;
+       cfs_rq->exec_clock += delta_exec;
+
+       delta_fair = calc_delta_fair(delta_exec, lw);
+       delta_mine = calc_delta_mine(delta_exec, curr->load.weight, lw);
+
+       if (cfs_rq->sleeper_bonus > sysctl_sched_stat_granularity) {
+               delta = calc_delta_mine(cfs_rq->sleeper_bonus,
+                                       curr->load.weight, lw);
+               if (unlikely(delta > cfs_rq->sleeper_bonus))
+                       delta = cfs_rq->sleeper_bonus;
+
+               cfs_rq->sleeper_bonus -= delta;
+               delta_mine -= delta;
+       }
+
+       cfs_rq->fair_clock += delta_fair;
+       /*
+        * We executed delta_exec amount of time on the CPU,
+        * but we were only entitled to delta_mine amount of
+        * time during that period (if nr_running == 1 then
+        * the two values are equal)
+        * [Note: delta_mine - delta_exec is negative]:
+        */
+       add_wait_runtime(cfs_rq, curr, delta_mine - delta_exec);
+}
+
+static void update_curr(struct cfs_rq *cfs_rq, u64 now)
+{
+       struct sched_entity *curr = cfs_rq_curr(cfs_rq);
+       unsigned long delta_exec;
+
+       if (unlikely(!curr))
+               return;
+
+       /*
+        * Get the amount of time the current task was running
+        * since the last time we changed load (this cannot
+        * overflow on 32 bits):
+        */
+       delta_exec = (unsigned long)(now - curr->exec_start);
+
+       curr->delta_exec += delta_exec;
+
+       if (unlikely(curr->delta_exec > sysctl_sched_stat_granularity)) {
+               __update_curr(cfs_rq, curr, now);
+               curr->delta_exec = 0;
+       }
+       curr->exec_start = now;
+}
+
+static inline void
+update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 now)
+{
+       se->wait_start_fair = cfs_rq->fair_clock;
+       se->wait_start = now;
+}
+
+/*
+ * We calculate fair deltas here, so protect against the random effects
+ * of a multiplication overflow by capping it to the runtime limit:
+ */
+#if BITS_PER_LONG == 32
+static inline unsigned long
+calc_weighted(unsigned long delta, unsigned long weight, int shift)
+{
+       u64 tmp = (u64)delta * weight >> shift;
+
+       if (unlikely(tmp > sysctl_sched_runtime_limit*2))
+               return sysctl_sched_runtime_limit*2;
+       return tmp;
+}
+#else
+static inline unsigned long
+calc_weighted(unsigned long delta, unsigned long weight, int shift)
+{
+       return delta * weight >> shift;
+}
+#endif
+
+/*
+ * Task is being enqueued - update stats:
+ */
+static void
+update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 now)
+{
+       s64 key;
+
+       /*
+        * Are we enqueueing a waiting task? (for current tasks
+        * a dequeue/enqueue event is a NOP)
+        */
+       if (se != cfs_rq_curr(cfs_rq))
+               update_stats_wait_start(cfs_rq, se, now);
+       /*
+        * Update the key:
+        */
+       key = cfs_rq->fair_clock;
+
+       /*
+        * Optimize the common nice 0 case:
+        */
+       if (likely(se->load.weight == NICE_0_LOAD)) {
+               key -= se->wait_runtime;
+       } else {
+               u64 tmp;
+
+               if (se->wait_runtime < 0) {
+                       tmp = -se->wait_runtime;
+                       key += (tmp * se->load.inv_weight) >>
+                                       (WMULT_SHIFT - NICE_0_SHIFT);
+               } else {
+                       tmp = se->wait_runtime;
+                       key -= (tmp * se->load.weight) >> NICE_0_SHIFT;
+               }
+       }
+
+       se->fair_key = key;
+}
+
+/*
+ * Note: must be called with a freshly updated rq->fair_clock.
+ */
+static inline void
+__update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 now)
+{
+       unsigned long delta_fair = se->delta_fair_run;
+
+#ifdef CONFIG_SCHEDSTATS
+       {
+               s64 delta_wait = now - se->wait_start;
+               if (unlikely(delta_wait > se->wait_max))
+                       se->wait_max = delta_wait;
+       }
+#endif
+
+       if (unlikely(se->load.weight != NICE_0_LOAD))
+               delta_fair = calc_weighted(delta_fair, se->load.weight,
+                                                       NICE_0_SHIFT);
+
+       add_wait_runtime(cfs_rq, se, delta_fair);
+}
+
+static void
+update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 now)
+{
+       unsigned long delta_fair;
+
+       delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
+                       (u64)(cfs_rq->fair_clock - se->wait_start_fair));
+
+       se->delta_fair_run += delta_fair;
+       if (unlikely(abs(se->delta_fair_run) >=
+                               sysctl_sched_stat_granularity)) {
+               __update_stats_wait_end(cfs_rq, se, now);
+               se->delta_fair_run = 0;
+       }
+
+       se->wait_start_fair = 0;
+       se->wait_start = 0;
+}
+
+static inline void
+update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 now)
+{
+       update_curr(cfs_rq, now);
+       /*
+        * Mark the end of the wait period if dequeueing a
+        * waiting task:
+        */
+       if (se != cfs_rq_curr(cfs_rq))
+               update_stats_wait_end(cfs_rq, se, now);
+}
+
+/*
+ * We are picking a new current task - update its stats:
+ */
+static inline void
+update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 now)
+{
+       /*
+        * We are starting a new run period:
+        */
+       se->exec_start = now;
+}
+
+/*
+ * We are descheduling a task - update its stats:
+ */
+static inline void
+update_stats_curr_end(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 now)
+{
+       se->exec_start = 0;
+}
+
+/**************************************************
+ * Scheduling class queueing methods:
+ */
+
+static void
+__enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 now)
+{
+       unsigned long load = cfs_rq->load.weight, delta_fair;
+       long prev_runtime;
+
+       if (sysctl_sched_features & SCHED_FEAT_SLEEPER_LOAD_AVG)
+               load = rq_of(cfs_rq)->cpu_load[2];
+
+       delta_fair = se->delta_fair_sleep;
+
+       /*
+        * Fix up delta_fair with the effect of us running
+        * during the whole sleep period:
+        */
+       if (sysctl_sched_features & SCHED_FEAT_SLEEPER_AVG)
+               delta_fair = div64_likely32((u64)delta_fair * load,
+                                               load + se->load.weight);
+
+       if (unlikely(se->load.weight != NICE_0_LOAD))
+               delta_fair = calc_weighted(delta_fair, se->load.weight,
+                                                       NICE_0_SHIFT);
+
+       prev_runtime = se->wait_runtime;
+       __add_wait_runtime(cfs_rq, se, delta_fair);
+       delta_fair = se->wait_runtime - prev_runtime;
+
+       /*
+        * Track the amount of bonus we've given to sleepers:
+        */
+       cfs_rq->sleeper_bonus += delta_fair;
+
+       schedstat_add(cfs_rq, wait_runtime, se->wait_runtime);
+}
+
+static void
+enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 now)
+{
+       struct task_struct *tsk = task_of(se);
+       unsigned long delta_fair;
+
+       if ((entity_is_task(se) && tsk->policy == SCHED_BATCH) ||
+                        !(sysctl_sched_features & SCHED_FEAT_FAIR_SLEEPERS))
+               return;
+
+       delta_fair = (unsigned long)min((u64)(2*sysctl_sched_runtime_limit),
+               (u64)(cfs_rq->fair_clock - se->sleep_start_fair));
+
+       se->delta_fair_sleep += delta_fair;
+       if (unlikely(abs(se->delta_fair_sleep) >=
+                               sysctl_sched_stat_granularity)) {
+               __enqueue_sleeper(cfs_rq, se, now);
+               se->delta_fair_sleep = 0;
+       }
+
+       se->sleep_start_fair = 0;
+
+#ifdef CONFIG_SCHEDSTATS
+       if (se->sleep_start) {
+               u64 delta = now - se->sleep_start;
+
+               if ((s64)delta < 0)
+                       delta = 0;
+
+               if (unlikely(delta > se->sleep_max))
+                       se->sleep_max = delta;
+
+               se->sleep_start = 0;
+               se->sum_sleep_runtime += delta;
+       }
+       if (se->block_start) {
+               u64 delta = now - se->block_start;
+
+               if ((s64)delta < 0)
+                       delta = 0;
+
+               if (unlikely(delta > se->block_max))
+                       se->block_max = delta;
+
+               se->block_start = 0;
+               se->sum_sleep_runtime += delta;
+       }
+#endif
+}
+
+static void
+enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
+              int wakeup, u64 now)
+{
+       /*
+        * Update the fair clock.
+        */
+       update_curr(cfs_rq, now);
+
+       if (wakeup)
+               enqueue_sleeper(cfs_rq, se, now);
+
+       update_stats_enqueue(cfs_rq, se, now);
+       __enqueue_entity(cfs_rq, se);
+}
+
+static void
+dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se,
+              int sleep, u64 now)
+{
+       update_stats_dequeue(cfs_rq, se, now);
+       if (sleep) {
+               se->sleep_start_fair = cfs_rq->fair_clock;
+#ifdef CONFIG_SCHEDSTATS
+               if (entity_is_task(se)) {
+                       struct task_struct *tsk = task_of(se);
+
+                       if (tsk->state & TASK_INTERRUPTIBLE)
+                               se->sleep_start = now;
+                       if (tsk->state & TASK_UNINTERRUPTIBLE)
+                               se->block_start = now;
+               }
+               cfs_rq->wait_runtime -= se->wait_runtime;
+#endif
+       }
+       __dequeue_entity(cfs_rq, se);
+}
+
+/*
+ * Preempt the current task with a newly woken task if needed:
+ */
+static void
+__check_preempt_curr_fair(struct cfs_rq *cfs_rq, struct sched_entity *se,
+                         struct sched_entity *curr, unsigned long granularity)
+{
+       s64 __delta = curr->fair_key - se->fair_key;
+
+       /*
+        * Take scheduling granularity into account - do not
+        * preempt the current task unless the best task has
+        * a larger than sched_granularity fairness advantage:
+        */
+       if (__delta > niced_granularity(curr, granularity))
+               resched_task(rq_of(cfs_rq)->curr);
+}
+
+static inline void
+set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, u64 now)
+{
+       /*
+        * Any task has to be enqueued before it get to execute on
+        * a CPU. So account for the time it spent waiting on the
+        * runqueue. (note, here we rely on pick_next_task() having
+        * done a put_prev_task_fair() shortly before this, which
+        * updated rq->fair_clock - used by update_stats_wait_end())
+        */
+       update_stats_wait_end(cfs_rq, se, now);
+       update_stats_curr_start(cfs_rq, se, now);
+       set_cfs_rq_curr(cfs_rq, se);
+}
+
+static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq, u64 now)
+{
+       struct sched_entity *se = __pick_next_entity(cfs_rq);
+
+       set_next_entity(cfs_rq, se, now);
+
+       return se;
+}
+
+static void
+put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev, u64 now)
+{
+       /*
+        * If still on the runqueue then deactivate_task()
+        * was not called and update_curr() has to be done:
+        */
+       if (prev->on_rq)
+               update_curr(cfs_rq, now);
+
+       update_stats_curr_end(cfs_rq, prev, now);
+
+       if (prev->on_rq)
+               update_stats_wait_start(cfs_rq, prev, now);
+       set_cfs_rq_curr(cfs_rq, NULL);
+}
+
+static void entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
+{
+       struct rq *rq = rq_of(cfs_rq);
+       struct sched_entity *next;
+       u64 now = __rq_clock(rq);
+
+       /*
+        * Dequeue and enqueue the task to update its
+        * position within the tree:
+        */
+       dequeue_entity(cfs_rq, curr, 0, now);
+       enqueue_entity(cfs_rq, curr, 0, now);
+
+       /*
+        * Reschedule if another task tops the current one.
+        */
+       next = __pick_next_entity(cfs_rq);
+       if (next == curr)
+               return;
+
+       __check_preempt_curr_fair(cfs_rq, next, curr, sysctl_sched_granularity);
+}
+
+/**************************************************
+ * CFS operations on tasks:
+ */
+
+#ifdef CONFIG_FAIR_GROUP_SCHED
+
+/* Walk up scheduling entities hierarchy */
+#define for_each_sched_entity(se) \
+               for (; se; se = se->parent)
+
+static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
+{
+       return p->se.cfs_rq;
+}
+
+/* runqueue on which this entity is (to be) queued */
+static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
+{
+       return se->cfs_rq;
+}
+
+/* runqueue "owned" by this group */
+static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
+{
+       return grp->my_q;
+}
+
+/* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
+ * another cpu ('this_cpu')
+ */
+static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
+{
+       /* A later patch will take group into account */
+       return &cpu_rq(this_cpu)->cfs;
+}
+
+/* Iterate thr' all leaf cfs_rq's on a runqueue */
+#define for_each_leaf_cfs_rq(rq, cfs_rq) \
+       list_for_each_entry(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
+
+/* Do the two (enqueued) tasks belong to the same group ? */
+static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
+{
+       if (curr->se.cfs_rq == p->se.cfs_rq)
+               return 1;
+
+       return 0;
+}
+
+#else  /* CONFIG_FAIR_GROUP_SCHED */
+
+#define for_each_sched_entity(se) \
+               for (; se; se = NULL)
+
+static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
+{
+       return &task_rq(p)->cfs;
+}
+
+static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
+{
+       struct task_struct *p = task_of(se);
+       struct rq *rq = task_rq(p);
+
+       return &rq->cfs;
+}
+
+/* runqueue "owned" by this group */
+static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
+{
+       return NULL;
+}
+
+static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
+{
+       return &cpu_rq(this_cpu)->cfs;
+}
+
+#define for_each_leaf_cfs_rq(rq, cfs_rq) \
+               for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
+
+static inline int is_same_group(struct task_struct *curr, struct task_struct *p)
+{
+       return 1;
+}
+
+#endif /* CONFIG_FAIR_GROUP_SCHED */
+
+/*
+ * The enqueue_task method is called before nr_running is
+ * 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, u64 now)
+{
+       struct cfs_rq *cfs_rq;
+       struct sched_entity *se = &p->se;
+
+       for_each_sched_entity(se) {
+               if (se->on_rq)
+                       break;
+               cfs_rq = cfs_rq_of(se);
+               enqueue_entity(cfs_rq, se, wakeup, now);
+       }
+}
+
+/*
+ * 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, u64 now)
+{
+       struct cfs_rq *cfs_rq;
+       struct sched_entity *se = &p->se;
+
+       for_each_sched_entity(se) {
+               cfs_rq = cfs_rq_of(se);
+               dequeue_entity(cfs_rq, se, sleep, now);
+               /* Don't dequeue parent if it has other entities besides us */
+               if (cfs_rq->load.weight)
+                       break;
+       }
+}
+
+/*
+ * sched_yield() support is very simple - we dequeue and enqueue
+ */
+static void yield_task_fair(struct rq *rq, struct task_struct *p)
+{
+       struct cfs_rq *cfs_rq = task_cfs_rq(p);
+       u64 now = __rq_clock(rq);
+
+       /*
+        * Dequeue and enqueue the task to update its
+        * position within the tree:
+        */
+       dequeue_entity(cfs_rq, &p->se, 0, now);
+       enqueue_entity(cfs_rq, &p->se, 0, now);
+}
+
+/*
+ * Preempt the current task with a newly woken task if needed:
+ */
+static void check_preempt_curr_fair(struct rq *rq, struct task_struct *p)
+{
+       struct task_struct *curr = rq->curr;
+       struct cfs_rq *cfs_rq = task_cfs_rq(curr);
+       unsigned long gran;
+
+       if (unlikely(rt_prio(p->prio))) {
+               update_curr(cfs_rq, rq_clock(rq));
+               resched_task(curr);
+               return;
+       }
+
+       gran = sysctl_sched_wakeup_granularity;
+       /*
+        * Batch tasks prefer throughput over latency:
+        */
+       if (unlikely(p->policy == SCHED_BATCH))
+               gran = sysctl_sched_batch_wakeup_granularity;
+
+       if (is_same_group(curr, p))
+               __check_preempt_curr_fair(cfs_rq, &p->se, &curr->se, gran);
+}
+
+static struct task_struct *pick_next_task_fair(struct rq *rq, u64 now)
+{
+       struct cfs_rq *cfs_rq = &rq->cfs;
+       struct sched_entity *se;
+
+       if (unlikely(!cfs_rq->nr_running))
+               return NULL;
+
+       do {
+               se = pick_next_entity(cfs_rq, now);
+               cfs_rq = group_cfs_rq(se);
+       } while (cfs_rq);
+
+       return task_of(se);
+}
+
+/*
+ * Account for a descheduled task:
+ */
+static void put_prev_task_fair(struct rq *rq, struct task_struct *prev, u64 now)
+{
+       struct sched_entity *se = &prev->se;
+       struct cfs_rq *cfs_rq;
+
+       for_each_sched_entity(se) {
+               cfs_rq = cfs_rq_of(se);
+               put_prev_entity(cfs_rq, se, now);
+       }
+}
+
+/**************************************************
+ * 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:
+ */
+static inline struct task_struct *
+__load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
+{
+       struct task_struct *p;
+
+       if (!curr)
+               return NULL;
+
+       p = rb_entry(curr, struct task_struct, se.run_node);
+       cfs_rq->rb_load_balance_curr = rb_next(curr);
+
+       return p;
+}
+
+static struct task_struct *load_balance_start_fair(void *arg)
+{
+       struct cfs_rq *cfs_rq = arg;
+
+       return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
+}
+
+static struct task_struct *load_balance_next_fair(void *arg)
+{
+       struct cfs_rq *cfs_rq = arg;
+
+       return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
+}
+
+static int cfs_rq_best_prio(struct cfs_rq *cfs_rq)
+{
+       struct sched_entity *curr;
+       struct task_struct *p;
+
+       if (!cfs_rq->nr_running)
+               return MAX_PRIO;
+
+       curr = __pick_next_entity(cfs_rq);
+       p = task_of(curr);
+
+       return p->prio;
+}
+
+static int
+load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
+                       unsigned long max_nr_move, unsigned long max_load_move,
+                       struct sched_domain *sd, enum cpu_idle_type idle,
+                       int *all_pinned, unsigned long *total_load_moved)
+{
+       struct cfs_rq *busy_cfs_rq;
+       unsigned long load_moved, total_nr_moved = 0, nr_moved;
+       long rem_load_move = max_load_move;
+       struct rq_iterator cfs_rq_iterator;
+
+       cfs_rq_iterator.start = load_balance_start_fair;
+       cfs_rq_iterator.next = load_balance_next_fair;
+
+       for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
+               struct cfs_rq *this_cfs_rq;
+               long imbalance;
+               unsigned long maxload;
+               int this_best_prio, best_prio, best_prio_seen = 0;
+
+               this_cfs_rq = cpu_cfs_rq(busy_cfs_rq, this_cpu);
+
+               imbalance = busy_cfs_rq->load.weight -
+                                                this_cfs_rq->load.weight;
+               /* Don't pull if this_cfs_rq has more load than busy_cfs_rq */
+               if (imbalance <= 0)
+                       continue;
+
+               /* Don't pull more than imbalance/2 */
+               imbalance /= 2;
+               maxload = min(rem_load_move, imbalance);
+
+               this_best_prio = cfs_rq_best_prio(this_cfs_rq);
+               best_prio = cfs_rq_best_prio(busy_cfs_rq);
+
+               /*
+                * Enable handling of the case where there is more than one task
+                * with the best priority. If the current running task is one
+                * of those with prio==best_prio we know it won't be moved
+                * and therefore it's safe to override the skip (based on load)
+                * of any task we find with that prio.
+                */
+               if (cfs_rq_curr(busy_cfs_rq) == &busiest->curr->se)
+                       best_prio_seen = 1;
+
+               /* pass busy_cfs_rq argument into
+                * load_balance_[start|next]_fair iterators
+                */
+               cfs_rq_iterator.arg = busy_cfs_rq;
+               nr_moved = balance_tasks(this_rq, this_cpu, busiest,
+                               max_nr_move, maxload, sd, idle, all_pinned,
+                               &load_moved, this_best_prio, best_prio,
+                               best_prio_seen, &cfs_rq_iterator);
+
+               total_nr_moved += nr_moved;
+               max_nr_move -= nr_moved;
+               rem_load_move -= load_moved;
+
+               if (max_nr_move <= 0 || rem_load_move <= 0)
+                       break;
+       }
+
+       *total_load_moved = max_load_move - rem_load_move;
+
+       return total_nr_moved;
+}
+
+/*
+ * scheduler tick hitting a task of our scheduling class:
+ */
+static void task_tick_fair(struct rq *rq, struct task_struct *curr)
+{
+       struct cfs_rq *cfs_rq;
+       struct sched_entity *se = &curr->se;
+
+       for_each_sched_entity(se) {
+               cfs_rq = cfs_rq_of(se);
+               entity_tick(cfs_rq, se);
+       }
+}
+
+/*
+ * 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.
+ */
+static void task_new_fair(struct rq *rq, struct task_struct *p)
+{
+       struct cfs_rq *cfs_rq = task_cfs_rq(p);
+       struct sched_entity *se = &p->se;
+       u64 now = rq_clock(rq);
+
+       sched_info_queued(p);
+
+       update_stats_enqueue(cfs_rq, se, now);
+       /*
+        * Child runs first: we let it run before the parent
+        * until it reschedules once. We set up the key so that
+        * it will preempt the parent:
+        */
+       p->se.fair_key = current->se.fair_key -
+               niced_granularity(&rq->curr->se, sysctl_sched_granularity) - 1;
+       /*
+        * The first wait is dominated by the child-runs-first logic,
+        * so do not credit it with that waiting time yet:
+        */
+       if (sysctl_sched_features & SCHED_FEAT_SKIP_INITIAL)
+               p->se.wait_start_fair = 0;
+
+       /*
+        * The statistical average of wait_runtime is about
+        * -granularity/2, so initialize the task with that:
+        */
+       if (sysctl_sched_features & SCHED_FEAT_START_DEBIT)
+               p->se.wait_runtime = -(sysctl_sched_granularity / 2);
+
+       __enqueue_entity(cfs_rq, se);
+       inc_nr_running(p, rq, now);
+}
+
+#ifdef CONFIG_FAIR_GROUP_SCHED
+/* Account for a task changing its policy or group.
+ *
+ * This routine is mostly called to set cfs_rq->curr field when a task
+ * migrates between groups/classes.
+ */
+static void set_curr_task_fair(struct rq *rq)
+{
+       struct task_struct *curr = rq->curr;
+       struct sched_entity *se = &curr->se;
+       u64 now = rq_clock(rq);
+       struct cfs_rq *cfs_rq;
+
+       for_each_sched_entity(se) {
+               cfs_rq = cfs_rq_of(se);
+               set_next_entity(cfs_rq, se, now);
+       }
+}
+#else
+static void set_curr_task_fair(struct rq *rq)
+{
+}
+#endif
+
+/*
+ * All the scheduling class methods:
+ */
+struct sched_class fair_sched_class __read_mostly = {
+       .enqueue_task           = enqueue_task_fair,
+       .dequeue_task           = dequeue_task_fair,
+       .yield_task             = yield_task_fair,
+
+       .check_preempt_curr     = check_preempt_curr_fair,
+
+       .pick_next_task         = pick_next_task_fair,
+       .put_prev_task          = put_prev_task_fair,
+
+       .load_balance           = load_balance_fair,
+
+       .set_curr_task          = set_curr_task_fair,
+       .task_tick              = task_tick_fair,
+       .task_new               = task_new_fair,
+};
+
+#ifdef CONFIG_SCHED_DEBUG
+void print_cfs_stats(struct seq_file *m, int cpu, u64 now)
+{
+       struct rq *rq = cpu_rq(cpu);
+       struct cfs_rq *cfs_rq;
+
+       for_each_leaf_cfs_rq(rq, cfs_rq)
+               print_cfs_rq(m, cpu, cfs_rq, now);
+}
+#endif