* miss some of their deadlines), and won't affect any other task.
  *
  * Copyright (C) 2012 Dario Faggioli <raistlin@linux.it>,
+ *                    Juri Lelli <juri.lelli@gmail.com>,
  *                    Michael Trimarchi <michael@amarulasolutions.com>,
  *                    Fabio Checconi <fchecconi@gmail.com>
  */
        return (s64)(a - b) < 0;
 }
 
+/*
+ * Tells if entity @a should preempt entity @b.
+ */
+static inline
+int dl_entity_preempt(struct sched_dl_entity *a, struct sched_dl_entity *b)
+{
+       return dl_time_before(a->deadline, b->deadline);
+}
+
 static inline struct task_struct *dl_task_of(struct sched_dl_entity *dl_se)
 {
        return container_of(dl_se, struct task_struct, dl);
 void init_dl_rq(struct dl_rq *dl_rq, struct rq *rq)
 {
        dl_rq->rb_root = RB_ROOT;
+
+#ifdef CONFIG_SMP
+       /* zero means no -deadline tasks */
+       dl_rq->earliest_dl.curr = dl_rq->earliest_dl.next = 0;
+
+       dl_rq->dl_nr_migratory = 0;
+       dl_rq->overloaded = 0;
+       dl_rq->pushable_dl_tasks_root = RB_ROOT;
+#endif
+}
+
+#ifdef CONFIG_SMP
+
+static inline int dl_overloaded(struct rq *rq)
+{
+       return atomic_read(&rq->rd->dlo_count);
+}
+
+static inline void dl_set_overload(struct rq *rq)
+{
+       if (!rq->online)
+               return;
+
+       cpumask_set_cpu(rq->cpu, rq->rd->dlo_mask);
+       /*
+        * Must be visible before the overload count is
+        * set (as in sched_rt.c).
+        *
+        * Matched by the barrier in pull_dl_task().
+        */
+       smp_wmb();
+       atomic_inc(&rq->rd->dlo_count);
+}
+
+static inline void dl_clear_overload(struct rq *rq)
+{
+       if (!rq->online)
+               return;
+
+       atomic_dec(&rq->rd->dlo_count);
+       cpumask_clear_cpu(rq->cpu, rq->rd->dlo_mask);
+}
+
+static void update_dl_migration(struct dl_rq *dl_rq)
+{
+       if (dl_rq->dl_nr_migratory && dl_rq->dl_nr_total > 1) {
+               if (!dl_rq->overloaded) {
+                       dl_set_overload(rq_of_dl_rq(dl_rq));
+                       dl_rq->overloaded = 1;
+               }
+       } else if (dl_rq->overloaded) {
+               dl_clear_overload(rq_of_dl_rq(dl_rq));
+               dl_rq->overloaded = 0;
+       }
+}
+
+static void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
+{
+       struct task_struct *p = dl_task_of(dl_se);
+       dl_rq = &rq_of_dl_rq(dl_rq)->dl;
+
+       dl_rq->dl_nr_total++;
+       if (p->nr_cpus_allowed > 1)
+               dl_rq->dl_nr_migratory++;
+
+       update_dl_migration(dl_rq);
+}
+
+static void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
+{
+       struct task_struct *p = dl_task_of(dl_se);
+       dl_rq = &rq_of_dl_rq(dl_rq)->dl;
+
+       dl_rq->dl_nr_total--;
+       if (p->nr_cpus_allowed > 1)
+               dl_rq->dl_nr_migratory--;
+
+       update_dl_migration(dl_rq);
+}
+
+/*
+ * The list of pushable -deadline task is not a plist, like in
+ * sched_rt.c, it is an rb-tree with tasks ordered by deadline.
+ */
+static void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
+{
+       struct dl_rq *dl_rq = &rq->dl;
+       struct rb_node **link = &dl_rq->pushable_dl_tasks_root.rb_node;
+       struct rb_node *parent = NULL;
+       struct task_struct *entry;
+       int leftmost = 1;
+
+       BUG_ON(!RB_EMPTY_NODE(&p->pushable_dl_tasks));
+
+       while (*link) {
+               parent = *link;
+               entry = rb_entry(parent, struct task_struct,
+                                pushable_dl_tasks);
+               if (dl_entity_preempt(&p->dl, &entry->dl))
+                       link = &parent->rb_left;
+               else {
+                       link = &parent->rb_right;
+                       leftmost = 0;
+               }
+       }
+
+       if (leftmost)
+               dl_rq->pushable_dl_tasks_leftmost = &p->pushable_dl_tasks;
+
+       rb_link_node(&p->pushable_dl_tasks, parent, link);
+       rb_insert_color(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
 }
 
+static void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
+{
+       struct dl_rq *dl_rq = &rq->dl;
+
+       if (RB_EMPTY_NODE(&p->pushable_dl_tasks))
+               return;
+
+       if (dl_rq->pushable_dl_tasks_leftmost == &p->pushable_dl_tasks) {
+               struct rb_node *next_node;
+
+               next_node = rb_next(&p->pushable_dl_tasks);
+               dl_rq->pushable_dl_tasks_leftmost = next_node;
+       }
+
+       rb_erase(&p->pushable_dl_tasks, &dl_rq->pushable_dl_tasks_root);
+       RB_CLEAR_NODE(&p->pushable_dl_tasks);
+}
+
+static inline int has_pushable_dl_tasks(struct rq *rq)
+{
+       return !RB_EMPTY_ROOT(&rq->dl.pushable_dl_tasks_root);
+}
+
+static int push_dl_task(struct rq *rq);
+
+#else
+
+static inline
+void enqueue_pushable_dl_task(struct rq *rq, struct task_struct *p)
+{
+}
+
+static inline
+void dequeue_pushable_dl_task(struct rq *rq, struct task_struct *p)
+{
+}
+
+static inline
+void inc_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
+{
+}
+
+static inline
+void dec_dl_migration(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
+{
+}
+
+#endif /* CONFIG_SMP */
+
 static void enqueue_task_dl(struct rq *rq, struct task_struct *p, int flags);
 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags);
 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
                        check_preempt_curr_dl(rq, p, 0);
                else
                        resched_task(rq->curr);
+#ifdef CONFIG_SMP
+               /*
+                * Queueing this task back might have overloaded rq,
+                * check if we need to kick someone away.
+                */
+               if (has_pushable_dl_tasks(rq))
+                       push_dl_task(rq);
+#endif
        }
 unlock:
        raw_spin_unlock(&rq->lock);
        }
 }
 
+#ifdef CONFIG_SMP
+
+static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu);
+
+static inline u64 next_deadline(struct rq *rq)
+{
+       struct task_struct *next = pick_next_earliest_dl_task(rq, rq->cpu);
+
+       if (next && dl_prio(next->prio))
+               return next->dl.deadline;
+       else
+               return 0;
+}
+
+static void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
+{
+       struct rq *rq = rq_of_dl_rq(dl_rq);
+
+       if (dl_rq->earliest_dl.curr == 0 ||
+           dl_time_before(deadline, dl_rq->earliest_dl.curr)) {
+               /*
+                * If the dl_rq had no -deadline tasks, or if the new task
+                * has shorter deadline than the current one on dl_rq, we
+                * know that the previous earliest becomes our next earliest,
+                * as the new task becomes the earliest itself.
+                */
+               dl_rq->earliest_dl.next = dl_rq->earliest_dl.curr;
+               dl_rq->earliest_dl.curr = deadline;
+       } else if (dl_rq->earliest_dl.next == 0 ||
+                  dl_time_before(deadline, dl_rq->earliest_dl.next)) {
+               /*
+                * On the other hand, if the new -deadline task has a
+                * a later deadline than the earliest one on dl_rq, but
+                * it is earlier than the next (if any), we must
+                * recompute the next-earliest.
+                */
+               dl_rq->earliest_dl.next = next_deadline(rq);
+       }
+}
+
+static void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline)
+{
+       struct rq *rq = rq_of_dl_rq(dl_rq);
+
+       /*
+        * Since we may have removed our earliest (and/or next earliest)
+        * task we must recompute them.
+        */
+       if (!dl_rq->dl_nr_running) {
+               dl_rq->earliest_dl.curr = 0;
+               dl_rq->earliest_dl.next = 0;
+       } else {
+               struct rb_node *leftmost = dl_rq->rb_leftmost;
+               struct sched_dl_entity *entry;
+
+               entry = rb_entry(leftmost, struct sched_dl_entity, rb_node);
+               dl_rq->earliest_dl.curr = entry->deadline;
+               dl_rq->earliest_dl.next = next_deadline(rq);
+       }
+}
+
+#else
+
+static inline void inc_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
+static inline void dec_dl_deadline(struct dl_rq *dl_rq, u64 deadline) {}
+
+#endif /* CONFIG_SMP */
+
+static inline
+void inc_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
+{
+       int prio = dl_task_of(dl_se)->prio;
+       u64 deadline = dl_se->deadline;
+
+       WARN_ON(!dl_prio(prio));
+       dl_rq->dl_nr_running++;
+
+       inc_dl_deadline(dl_rq, deadline);
+       inc_dl_migration(dl_se, dl_rq);
+}
+
+static inline
+void dec_dl_tasks(struct sched_dl_entity *dl_se, struct dl_rq *dl_rq)
+{
+       int prio = dl_task_of(dl_se)->prio;
+
+       WARN_ON(!dl_prio(prio));
+       WARN_ON(!dl_rq->dl_nr_running);
+       dl_rq->dl_nr_running--;
+
+       dec_dl_deadline(dl_rq, dl_se->deadline);
+       dec_dl_migration(dl_se, dl_rq);
+}
+
 static void __enqueue_dl_entity(struct sched_dl_entity *dl_se)
 {
        struct dl_rq *dl_rq = dl_rq_of_se(dl_se);
        rb_link_node(&dl_se->rb_node, parent, link);
        rb_insert_color(&dl_se->rb_node, &dl_rq->rb_root);
 
-       dl_rq->dl_nr_running++;
+       inc_dl_tasks(dl_se, dl_rq);
 }
 
 static void __dequeue_dl_entity(struct sched_dl_entity *dl_se)
        rb_erase(&dl_se->rb_node, &dl_rq->rb_root);
        RB_CLEAR_NODE(&dl_se->rb_node);
 
-       dl_rq->dl_nr_running--;
+       dec_dl_tasks(dl_se, dl_rq);
 }
 
 static void
                return;
 
        enqueue_dl_entity(&p->dl, flags);
+
+       if (!task_current(rq, p) && p->nr_cpus_allowed > 1)
+               enqueue_pushable_dl_task(rq, p);
+
        inc_nr_running(rq);
 }
 
 static void __dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
 {
        dequeue_dl_entity(&p->dl);
+       dequeue_pushable_dl_task(rq, p);
 }
 
 static void dequeue_task_dl(struct rq *rq, struct task_struct *p, int flags)
        update_curr_dl(rq);
 }
 
+#ifdef CONFIG_SMP
+
+static int find_later_rq(struct task_struct *task);
+static int latest_cpu_find(struct cpumask *span,
+                          struct task_struct *task,
+                          struct cpumask *later_mask);
+
+static int
+select_task_rq_dl(struct task_struct *p, int cpu, int sd_flag, int flags)
+{
+       struct task_struct *curr;
+       struct rq *rq;
+
+       if (sd_flag != SD_BALANCE_WAKE && sd_flag != SD_BALANCE_FORK)
+               goto out;
+
+       rq = cpu_rq(cpu);
+
+       rcu_read_lock();
+       curr = ACCESS_ONCE(rq->curr); /* unlocked access */
+
+       /*
+        * If we are dealing with a -deadline task, we must
+        * decide where to wake it up.
+        * If it has a later deadline and the current task
+        * on this rq can't move (provided the waking task
+        * can!) we prefer to send it somewhere else. On the
+        * other hand, if it has a shorter deadline, we
+        * try to make it stay here, it might be important.
+        */
+       if (unlikely(dl_task(curr)) &&
+           (curr->nr_cpus_allowed < 2 ||
+            !dl_entity_preempt(&p->dl, &curr->dl)) &&
+           (p->nr_cpus_allowed > 1)) {
+               int target = find_later_rq(p);
+
+               if (target != -1)
+                       cpu = target;
+       }
+       rcu_read_unlock();
+
+out:
+       return cpu;
+}
+
+static void check_preempt_equal_dl(struct rq *rq, struct task_struct *p)
+{
+       /*
+        * Current can't be migrated, useless to reschedule,
+        * let's hope p can move out.
+        */
+       if (rq->curr->nr_cpus_allowed == 1 ||
+           latest_cpu_find(rq->rd->span, rq->curr, NULL) == -1)
+               return;
+
+       /*
+        * p is migratable, so let's not schedule it and
+        * see if it is pushed or pulled somewhere else.
+        */
+       if (p->nr_cpus_allowed != 1 &&
+           latest_cpu_find(rq->rd->span, p, NULL) != -1)
+               return;
+
+       resched_task(rq->curr);
+}
+
+#endif /* CONFIG_SMP */
+
 /*
  * Only called when both the current and waking task are -deadline
  * tasks.
 static void check_preempt_curr_dl(struct rq *rq, struct task_struct *p,
                                  int flags)
 {
-       if (dl_time_before(p->dl.deadline, rq->curr->dl.deadline))
+       if (dl_entity_preempt(&p->dl, &rq->curr->dl)) {
                resched_task(rq->curr);
+               return;
+       }
+
+#ifdef CONFIG_SMP
+       /*
+        * In the unlikely case current and p have the same deadline
+        * let us try to decide what's the best thing to do...
+        */
+       if ((s64)(p->dl.deadline - rq->curr->dl.deadline) == 0 &&
+           !need_resched())
+               check_preempt_equal_dl(rq, p);
+#endif /* CONFIG_SMP */
 }
 
 #ifdef CONFIG_SCHED_HRTICK
 
        p = dl_task_of(dl_se);
        p->se.exec_start = rq_clock_task(rq);
+
+       /* Running task will never be pushed. */
+       if (p)
+               dequeue_pushable_dl_task(rq, p);
+
 #ifdef CONFIG_SCHED_HRTICK
        if (hrtick_enabled(rq))
                start_hrtick_dl(rq, p);
 #endif
+
+#ifdef CONFIG_SMP
+       rq->post_schedule = has_pushable_dl_tasks(rq);
+#endif /* CONFIG_SMP */
+
        return p;
 }
 
 static void put_prev_task_dl(struct rq *rq, struct task_struct *p)
 {
        update_curr_dl(rq);
+
+       if (on_dl_rq(&p->dl) && p->nr_cpus_allowed > 1)
+               enqueue_pushable_dl_task(rq, p);
 }
 
 static void task_tick_dl(struct rq *rq, struct task_struct *p, int queued)
        struct task_struct *p = rq->curr;
 
        p->se.exec_start = rq_clock_task(rq);
+
+       /* You can't push away the running task */
+       dequeue_pushable_dl_task(rq, p);
+}
+
+#ifdef CONFIG_SMP
+
+/* Only try algorithms three times */
+#define DL_MAX_TRIES 3
+
+static int pick_dl_task(struct rq *rq, struct task_struct *p, int cpu)
+{
+       if (!task_running(rq, p) &&
+           (cpu < 0 || cpumask_test_cpu(cpu, &p->cpus_allowed)) &&
+           (p->nr_cpus_allowed > 1))
+               return 1;
+
+       return 0;
+}
+
+/* Returns the second earliest -deadline task, NULL otherwise */
+static struct task_struct *pick_next_earliest_dl_task(struct rq *rq, int cpu)
+{
+       struct rb_node *next_node = rq->dl.rb_leftmost;
+       struct sched_dl_entity *dl_se;
+       struct task_struct *p = NULL;
+
+next_node:
+       next_node = rb_next(next_node);
+       if (next_node) {
+               dl_se = rb_entry(next_node, struct sched_dl_entity, rb_node);
+               p = dl_task_of(dl_se);
+
+               if (pick_dl_task(rq, p, cpu))
+                       return p;
+
+               goto next_node;
+       }
+
+       return NULL;
+}
+
+static int latest_cpu_find(struct cpumask *span,
+                          struct task_struct *task,
+                          struct cpumask *later_mask)
+{
+       const struct sched_dl_entity *dl_se = &task->dl;
+       int cpu, found = -1, best = 0;
+       u64 max_dl = 0;
+
+       for_each_cpu(cpu, span) {
+               struct rq *rq = cpu_rq(cpu);
+               struct dl_rq *dl_rq = &rq->dl;
+
+               if (cpumask_test_cpu(cpu, &task->cpus_allowed) &&
+                   (!dl_rq->dl_nr_running || dl_time_before(dl_se->deadline,
+                    dl_rq->earliest_dl.curr))) {
+                       if (later_mask)
+                               cpumask_set_cpu(cpu, later_mask);
+                       if (!best && !dl_rq->dl_nr_running) {
+                               best = 1;
+                               found = cpu;
+                       } else if (!best &&
+                                  dl_time_before(max_dl,
+                                                 dl_rq->earliest_dl.curr)) {
+                               max_dl = dl_rq->earliest_dl.curr;
+                               found = cpu;
+                       }
+               } else if (later_mask)
+                       cpumask_clear_cpu(cpu, later_mask);
+       }
+
+       return found;
+}
+
+static DEFINE_PER_CPU(cpumask_var_t, local_cpu_mask_dl);
+
+static int find_later_rq(struct task_struct *task)
+{
+       struct sched_domain *sd;
+       struct cpumask *later_mask = __get_cpu_var(local_cpu_mask_dl);
+       int this_cpu = smp_processor_id();
+       int best_cpu, cpu = task_cpu(task);
+
+       /* Make sure the mask is initialized first */
+       if (unlikely(!later_mask))
+               return -1;
+
+       if (task->nr_cpus_allowed == 1)
+               return -1;
+
+       best_cpu = latest_cpu_find(task_rq(task)->rd->span, task, later_mask);
+       if (best_cpu == -1)
+               return -1;
+
+       /*
+        * If we are here, some target has been found,
+        * the most suitable of which is cached in best_cpu.
+        * This is, among the runqueues where the current tasks
+        * have later deadlines than the task's one, the rq
+        * with the latest possible one.
+        *
+        * Now we check how well this matches with task's
+        * affinity and system topology.
+        *
+        * The last cpu where the task run is our first
+        * guess, since it is most likely cache-hot there.
+        */
+       if (cpumask_test_cpu(cpu, later_mask))
+               return cpu;
+       /*
+        * Check if this_cpu is to be skipped (i.e., it is
+        * not in the mask) or not.
+        */
+       if (!cpumask_test_cpu(this_cpu, later_mask))
+               this_cpu = -1;
+
+       rcu_read_lock();
+       for_each_domain(cpu, sd) {
+               if (sd->flags & SD_WAKE_AFFINE) {
+
+                       /*
+                        * If possible, preempting this_cpu is
+                        * cheaper than migrating.
+                        */
+                       if (this_cpu != -1 &&
+                           cpumask_test_cpu(this_cpu, sched_domain_span(sd))) {
+                               rcu_read_unlock();
+                               return this_cpu;
+                       }
+
+                       /*
+                        * Last chance: if best_cpu is valid and is
+                        * in the mask, that becomes our choice.
+                        */
+                       if (best_cpu < nr_cpu_ids &&
+                           cpumask_test_cpu(best_cpu, sched_domain_span(sd))) {
+                               rcu_read_unlock();
+                               return best_cpu;
+                       }
+               }
+       }
+       rcu_read_unlock();
+
+       /*
+        * At this point, all our guesses failed, we just return
+        * 'something', and let the caller sort the things out.
+        */
+       if (this_cpu != -1)
+               return this_cpu;
+
+       cpu = cpumask_any(later_mask);
+       if (cpu < nr_cpu_ids)
+               return cpu;
+
+       return -1;
+}
+
+/* Locks the rq it finds */
+static struct rq *find_lock_later_rq(struct task_struct *task, struct rq *rq)
+{
+       struct rq *later_rq = NULL;
+       int tries;
+       int cpu;
+
+       for (tries = 0; tries < DL_MAX_TRIES; tries++) {
+               cpu = find_later_rq(task);
+
+               if ((cpu == -1) || (cpu == rq->cpu))
+                       break;
+
+               later_rq = cpu_rq(cpu);
+
+               /* Retry if something changed. */
+               if (double_lock_balance(rq, later_rq)) {
+                       if (unlikely(task_rq(task) != rq ||
+                                    !cpumask_test_cpu(later_rq->cpu,
+                                                      &task->cpus_allowed) ||
+                                    task_running(rq, task) || !task->on_rq)) {
+                               double_unlock_balance(rq, later_rq);
+                               later_rq = NULL;
+                               break;
+                       }
+               }
+
+               /*
+                * If the rq we found has no -deadline task, or
+                * its earliest one has a later deadline than our
+                * task, the rq is a good one.
+                */
+               if (!later_rq->dl.dl_nr_running ||
+                   dl_time_before(task->dl.deadline,
+                                  later_rq->dl.earliest_dl.curr))
+                       break;
+
+               /* Otherwise we try again. */
+               double_unlock_balance(rq, later_rq);
+               later_rq = NULL;
+       }
+
+       return later_rq;
+}
+
+static struct task_struct *pick_next_pushable_dl_task(struct rq *rq)
+{
+       struct task_struct *p;
+
+       if (!has_pushable_dl_tasks(rq))
+               return NULL;
+
+       p = rb_entry(rq->dl.pushable_dl_tasks_leftmost,
+                    struct task_struct, pushable_dl_tasks);
+
+       BUG_ON(rq->cpu != task_cpu(p));
+       BUG_ON(task_current(rq, p));
+       BUG_ON(p->nr_cpus_allowed <= 1);
+
+       BUG_ON(!p->se.on_rq);
+       BUG_ON(!dl_task(p));
+
+       return p;
+}
+
+/*
+ * See if the non running -deadline tasks on this rq
+ * can be sent to some other CPU where they can preempt
+ * and start executing.
+ */
+static int push_dl_task(struct rq *rq)
+{
+       struct task_struct *next_task;
+       struct rq *later_rq;
+
+       if (!rq->dl.overloaded)
+               return 0;
+
+       next_task = pick_next_pushable_dl_task(rq);
+       if (!next_task)
+               return 0;
+
+retry:
+       if (unlikely(next_task == rq->curr)) {
+               WARN_ON(1);
+               return 0;
+       }
+
+       /*
+        * If next_task preempts rq->curr, and rq->curr
+        * can move away, it makes sense to just reschedule
+        * without going further in pushing next_task.
+        */
+       if (dl_task(rq->curr) &&
+           dl_time_before(next_task->dl.deadline, rq->curr->dl.deadline) &&
+           rq->curr->nr_cpus_allowed > 1) {
+               resched_task(rq->curr);
+               return 0;
+       }
+
+       /* We might release rq lock */
+       get_task_struct(next_task);
+
+       /* Will lock the rq it'll find */
+       later_rq = find_lock_later_rq(next_task, rq);
+       if (!later_rq) {
+               struct task_struct *task;
+
+               /*
+                * We must check all this again, since
+                * find_lock_later_rq releases rq->lock and it is
+                * then possible that next_task has migrated.
+                */
+               task = pick_next_pushable_dl_task(rq);
+               if (task_cpu(next_task) == rq->cpu && task == next_task) {
+                       /*
+                        * The task is still there. We don't try
+                        * again, some other cpu will pull it when ready.
+                        */
+                       dequeue_pushable_dl_task(rq, next_task);
+                       goto out;
+               }
+
+               if (!task)
+                       /* No more tasks */
+                       goto out;
+
+               put_task_struct(next_task);
+               next_task = task;
+               goto retry;
+       }
+
+       deactivate_task(rq, next_task, 0);
+       set_task_cpu(next_task, later_rq->cpu);
+       activate_task(later_rq, next_task, 0);
+
+       resched_task(later_rq->curr);
+
+       double_unlock_balance(rq, later_rq);
+
+out:
+       put_task_struct(next_task);
+
+       return 1;
+}
+
+static void push_dl_tasks(struct rq *rq)
+{
+       /* Terminates as it moves a -deadline task */
+       while (push_dl_task(rq))
+               ;
 }
 
+static int pull_dl_task(struct rq *this_rq)
+{
+       int this_cpu = this_rq->cpu, ret = 0, cpu;
+       struct task_struct *p;
+       struct rq *src_rq;
+       u64 dmin = LONG_MAX;
+
+       if (likely(!dl_overloaded(this_rq)))
+               return 0;
+
+       /*
+        * Match the barrier from dl_set_overloaded; this guarantees that if we
+        * see overloaded we must also see the dlo_mask bit.
+        */
+       smp_rmb();
+
+       for_each_cpu(cpu, this_rq->rd->dlo_mask) {
+               if (this_cpu == cpu)
+                       continue;
+
+               src_rq = cpu_rq(cpu);
+
+               /*
+                * It looks racy, abd it is! However, as in sched_rt.c,
+                * we are fine with this.
+                */
+               if (this_rq->dl.dl_nr_running &&
+                   dl_time_before(this_rq->dl.earliest_dl.curr,
+                                  src_rq->dl.earliest_dl.next))
+                       continue;
+
+               /* Might drop this_rq->lock */
+               double_lock_balance(this_rq, src_rq);
+
+               /*
+                * If there are no more pullable tasks on the
+                * rq, we're done with it.
+                */
+               if (src_rq->dl.dl_nr_running <= 1)
+                       goto skip;
+
+               p = pick_next_earliest_dl_task(src_rq, this_cpu);
+
+               /*
+                * We found a task to be pulled if:
+                *  - it preempts our current (if there's one),
+                *  - it will preempt the last one we pulled (if any).
+                */
+               if (p && dl_time_before(p->dl.deadline, dmin) &&
+                   (!this_rq->dl.dl_nr_running ||
+                    dl_time_before(p->dl.deadline,
+                                   this_rq->dl.earliest_dl.curr))) {
+                       WARN_ON(p == src_rq->curr);
+                       WARN_ON(!p->se.on_rq);
+
+                       /*
+                        * Then we pull iff p has actually an earlier
+                        * deadline than the current task of its runqueue.
+                        */
+                       if (dl_time_before(p->dl.deadline,
+                                          src_rq->curr->dl.deadline))
+                               goto skip;
+
+                       ret = 1;
+
+                       deactivate_task(src_rq, p, 0);
+                       set_task_cpu(p, this_cpu);
+                       activate_task(this_rq, p, 0);
+                       dmin = p->dl.deadline;
+
+                       /* Is there any other task even earlier? */
+               }
+skip:
+               double_unlock_balance(this_rq, src_rq);
+       }
+
+       return ret;
+}
+
+static void pre_schedule_dl(struct rq *rq, struct task_struct *prev)
+{
+       /* Try to pull other tasks here */
+       if (dl_task(prev))
+               pull_dl_task(rq);
+}
+
+static void post_schedule_dl(struct rq *rq)
+{
+       push_dl_tasks(rq);
+}
+
+/*
+ * Since the task is not running and a reschedule is not going to happen
+ * anytime soon on its runqueue, we try pushing it away now.
+ */
+static void task_woken_dl(struct rq *rq, struct task_struct *p)
+{
+       if (!task_running(rq, p) &&
+           !test_tsk_need_resched(rq->curr) &&
+           has_pushable_dl_tasks(rq) &&
+           p->nr_cpus_allowed > 1 &&
+           dl_task(rq->curr) &&
+           (rq->curr->nr_cpus_allowed < 2 ||
+            dl_entity_preempt(&rq->curr->dl, &p->dl))) {
+               push_dl_tasks(rq);
+       }
+}
+
+static void set_cpus_allowed_dl(struct task_struct *p,
+                               const struct cpumask *new_mask)
+{
+       struct rq *rq;
+       int weight;
+
+       BUG_ON(!dl_task(p));
+
+       /*
+        * Update only if the task is actually running (i.e.,
+        * it is on the rq AND it is not throttled).
+        */
+       if (!on_dl_rq(&p->dl))
+               return;
+
+       weight = cpumask_weight(new_mask);
+
+       /*
+        * Only update if the process changes its state from whether it
+        * can migrate or not.
+        */
+       if ((p->nr_cpus_allowed > 1) == (weight > 1))
+               return;
+
+       rq = task_rq(p);
+
+       /*
+        * The process used to be able to migrate OR it can now migrate
+        */
+       if (weight <= 1) {
+               if (!task_current(rq, p))
+                       dequeue_pushable_dl_task(rq, p);
+               BUG_ON(!rq->dl.dl_nr_migratory);
+               rq->dl.dl_nr_migratory--;
+       } else {
+               if (!task_current(rq, p))
+                       enqueue_pushable_dl_task(rq, p);
+               rq->dl.dl_nr_migratory++;
+       }
+
+       update_dl_migration(&rq->dl);
+}
+
+/* Assumes rq->lock is held */
+static void rq_online_dl(struct rq *rq)
+{
+       if (rq->dl.overloaded)
+               dl_set_overload(rq);
+}
+
+/* Assumes rq->lock is held */
+static void rq_offline_dl(struct rq *rq)
+{
+       if (rq->dl.overloaded)
+               dl_clear_overload(rq);
+}
+
+void init_sched_dl_class(void)
+{
+       unsigned int i;
+
+       for_each_possible_cpu(i)
+               zalloc_cpumask_var_node(&per_cpu(local_cpu_mask_dl, i),
+                                       GFP_KERNEL, cpu_to_node(i));
+}
+
+#endif /* CONFIG_SMP */
+
 static void switched_from_dl(struct rq *rq, struct task_struct *p)
 {
-       if (hrtimer_active(&p->dl.dl_timer))
+       if (hrtimer_active(&p->dl.dl_timer) && !dl_policy(p->policy))
                hrtimer_try_to_cancel(&p->dl.dl_timer);
+
+#ifdef CONFIG_SMP
+       /*
+        * Since this might be the only -deadline task on the rq,
+        * this is the right place to try to pull some other one
+        * from an overloaded cpu, if any.
+        */
+       if (!rq->dl.dl_nr_running)
+               pull_dl_task(rq);
+#endif
 }
 
+/*
+ * When switching to -deadline, we may overload the rq, then
+ * we try to push someone off, if possible.
+ */
 static void switched_to_dl(struct rq *rq, struct task_struct *p)
 {
+       int check_resched = 1;
+
        /*
         * If p is throttled, don't consider the possibility
         * of preempting rq->curr, the check will be done right
                return;
 
        if (p->on_rq || rq->curr != p) {
-               if (task_has_dl_policy(rq->curr))
+#ifdef CONFIG_SMP
+               if (rq->dl.overloaded && push_dl_task(rq) && rq != task_rq(p))
+                       /* Only reschedule if pushing failed */
+                       check_resched = 0;
+#endif /* CONFIG_SMP */
+               if (check_resched && task_has_dl_policy(rq->curr))
                        check_preempt_curr_dl(rq, p, 0);
-               else
-                       resched_task(rq->curr);
        }
 }
 
+/*
+ * If the scheduling parameters of a -deadline task changed,
+ * a push or pull operation might be needed.
+ */
 static void prio_changed_dl(struct rq *rq, struct task_struct *p,
                            int oldprio)
 {
-       switched_to_dl(rq, p);
-}
-
+       if (p->on_rq || rq->curr == p) {
 #ifdef CONFIG_SMP
-static int
-select_task_rq_dl(struct task_struct *p, int prev_cpu, int sd_flag, int flags)
-{
-       return task_cpu(p);
+               /*
+                * This might be too much, but unfortunately
+                * we don't have the old deadline value, and
+                * we can't argue if the task is increasing
+                * or lowering its prio, so...
+                */
+               if (!rq->dl.overloaded)
+                       pull_dl_task(rq);
+
+               /*
+                * If we now have a earlier deadline task than p,
+                * then reschedule, provided p is still on this
+                * runqueue.
+                */
+               if (dl_time_before(rq->dl.earliest_dl.curr, p->dl.deadline) &&
+                   rq->curr == p)
+                       resched_task(p);
+#else
+               /*
+                * Again, we don't know if p has a earlier
+                * or later deadline, so let's blindly set a
+                * (maybe not needed) rescheduling point.
+                */
+               resched_task(p);
+#endif /* CONFIG_SMP */
+       } else
+               switched_to_dl(rq, p);
 }
-#endif
 
 const struct sched_class dl_sched_class = {
        .next                   = &rt_sched_class,
 
 #ifdef CONFIG_SMP
        .select_task_rq         = select_task_rq_dl,
+       .set_cpus_allowed       = set_cpus_allowed_dl,
+       .rq_online              = rq_online_dl,
+       .rq_offline             = rq_offline_dl,
+       .pre_schedule           = pre_schedule_dl,
+       .post_schedule          = post_schedule_dl,
+       .task_woken             = task_woken_dl,
 #endif
 
        .set_curr_task          = set_curr_task_dl,