sched: Enable might_sleep before initializing drivers.
[linux-2.6.git] / kernel / sched.c
index bb504e1..d0f600c 100644 (file)
@@ -75,6 +75,9 @@
 #include <asm/tlb.h>
 #include <asm/irq_regs.h>
 #include <asm/mutex.h>
+#ifdef CONFIG_PARAVIRT
+#include <asm/paravirt.h>
+#endif
 
 #include "sched_cpupri.h"
 #include "workqueue_sched.h"
 
 static inline int rt_policy(int policy)
 {
-       if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
+       if (policy == SCHED_FIFO || policy == SCHED_RR)
                return 1;
        return 0;
 }
@@ -292,8 +295,8 @@ static DEFINE_SPINLOCK(task_group_lock);
  * (The default weight is 1024 - so there's no practical
  *  limitation from this.)
  */
-#define MIN_SHARES     2
-#define MAX_SHARES     (1UL << (18 + SCHED_LOAD_RESOLUTION))
+#define MIN_SHARES     (1UL <<  1)
+#define MAX_SHARES     (1UL << 18)
 
 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
 #endif
@@ -422,6 +425,7 @@ struct rt_rq {
  */
 struct root_domain {
        atomic_t refcount;
+       atomic_t rto_count;
        struct rcu_head rcu;
        cpumask_var_t span;
        cpumask_var_t online;
@@ -431,7 +435,6 @@ struct root_domain {
         * one runnable RT task.
         */
        cpumask_var_t rto_mask;
-       atomic_t rto_count;
        struct cpupri cpupri;
 };
 
@@ -528,6 +531,12 @@ struct rq {
 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
        u64 prev_irq_time;
 #endif
+#ifdef CONFIG_PARAVIRT
+       u64 prev_steal_time;
+#endif
+#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
+       u64 prev_steal_time_rq;
+#endif
 
        /* calc_load related fields */
        unsigned long calc_load_update;
@@ -581,7 +590,6 @@ static inline int cpu_of(struct rq *rq)
 
 #define rcu_dereference_check_sched_domain(p) \
        rcu_dereference_check((p), \
-                             rcu_read_lock_held() || \
                              lockdep_is_held(&sched_domains_mutex))
 
 /*
@@ -605,10 +613,10 @@ static inline int cpu_of(struct rq *rq)
 /*
  * Return the group to which this tasks belongs.
  *
- * We use task_subsys_state_check() and extend the RCU verification
- * with lockdep_is_held(&p->pi_lock) because cpu_cgroup_attach()
- * holds that lock for each task it moves into the cgroup. Therefore
- * by holding that lock, we pin the task to the current cgroup.
+ * We use task_subsys_state_check() and extend the RCU verification with
+ * pi->lock and rq->lock because cpu_cgroup_attach() holds those locks for each
+ * task it moves into the cgroup. Therefore by holding either of those locks,
+ * we pin the task to the current cgroup.
  */
 static inline struct task_group *task_group(struct task_struct *p)
 {
@@ -616,7 +624,8 @@ static inline struct task_group *task_group(struct task_struct *p)
        struct cgroup_subsys_state *css;
 
        css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
-                       lockdep_is_held(&p->pi_lock));
+                       lockdep_is_held(&p->pi_lock) ||
+                       lockdep_is_held(&task_rq(p)->lock));
        tg = container_of(css, struct task_group, css);
 
        return autogroup_task_group(p, tg);
@@ -1567,38 +1576,6 @@ static unsigned long cpu_avg_load_per_task(int cpu)
        return rq->avg_load_per_task;
 }
 
-#ifdef CONFIG_FAIR_GROUP_SCHED
-
-/*
- * Compute the cpu's hierarchical load factor for each task group.
- * This needs to be done in a top-down fashion because the load of a child
- * group is a fraction of its parents load.
- */
-static int tg_load_down(struct task_group *tg, void *data)
-{
-       unsigned long load;
-       long cpu = (long)data;
-
-       if (!tg->parent) {
-               load = cpu_rq(cpu)->load.weight;
-       } else {
-               load = tg->parent->cfs_rq[cpu]->h_load;
-               load *= tg->se[cpu]->load.weight;
-               load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
-       }
-
-       tg->cfs_rq[cpu]->h_load = load;
-
-       return 0;
-}
-
-static void update_h_load(long cpu)
-{
-       walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
-}
-
-#endif
-
 #ifdef CONFIG_PREEMPT
 
 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
@@ -1952,10 +1929,28 @@ void account_system_vtime(struct task_struct *curr)
 }
 EXPORT_SYMBOL_GPL(account_system_vtime);
 
-static void update_rq_clock_task(struct rq *rq, s64 delta)
+#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
+
+#ifdef CONFIG_PARAVIRT
+static inline u64 steal_ticks(u64 steal)
 {
-       s64 irq_delta;
+       if (unlikely(steal > NSEC_PER_SEC))
+               return div_u64(steal, TICK_NSEC);
 
+       return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
+}
+#endif
+
+static void update_rq_clock_task(struct rq *rq, s64 delta)
+{
+/*
+ * In theory, the compile should just see 0 here, and optimize out the call
+ * to sched_rt_avg_update. But I don't trust it...
+ */
+#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
+       s64 steal = 0, irq_delta = 0;
+#endif
+#ifdef CONFIG_IRQ_TIME_ACCOUNTING
        irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
 
        /*
@@ -1978,12 +1973,35 @@ static void update_rq_clock_task(struct rq *rq, s64 delta)
 
        rq->prev_irq_time += irq_delta;
        delta -= irq_delta;
+#endif
+#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
+       if (static_branch((&paravirt_steal_rq_enabled))) {
+               u64 st;
+
+               steal = paravirt_steal_clock(cpu_of(rq));
+               steal -= rq->prev_steal_time_rq;
+
+               if (unlikely(steal > delta))
+                       steal = delta;
+
+               st = steal_ticks(steal);
+               steal = st * TICK_NSEC;
+
+               rq->prev_steal_time_rq += steal;
+
+               delta -= steal;
+       }
+#endif
+
        rq->clock_task += delta;
 
-       if (irq_delta && sched_feat(NONIRQ_POWER))
-               sched_rt_avg_update(rq, irq_delta);
+#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
+       if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
+               sched_rt_avg_update(rq, irq_delta + steal);
+#endif
 }
 
+#ifdef CONFIG_IRQ_TIME_ACCOUNTING
 static int irqtime_account_hi_update(void)
 {
        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
@@ -2018,12 +2036,7 @@ static int irqtime_account_si_update(void)
 
 #define sched_clock_irqtime    (0)
 
-static void update_rq_clock_task(struct rq *rq, s64 delta)
-{
-       rq->clock_task += delta;
-}
-
-#endif /* CONFIG_IRQ_TIME_ACCOUNTING */
+#endif
 
 #include "sched_idletask.c"
 #include "sched_fair.c"
@@ -2200,6 +2213,16 @@ void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
                        !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
 
 #ifdef CONFIG_LOCKDEP
+       /*
+        * The caller should hold either p->pi_lock or rq->lock, when changing
+        * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
+        *
+        * sched_move_task() holds both and thus holding either pins the cgroup,
+        * see set_task_rq().
+        *
+        * Furthermore, all task_rq users should acquire both locks, see
+        * task_rq_lock().
+        */
        WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
                                      lockdep_is_held(&task_rq(p)->lock)));
 #endif
@@ -2209,7 +2232,7 @@ void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
 
        if (task_cpu(p) != new_cpu) {
                p->se.nr_migrations++;
-               perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
+               perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
        }
 
        __set_task_cpu(p, new_cpu);
@@ -2447,6 +2470,10 @@ ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
                }
                rcu_read_unlock();
        }
+
+       if (wake_flags & WF_MIGRATED)
+               schedstat_inc(p, se.statistics.nr_wakeups_migrate);
+
 #endif /* CONFIG_SMP */
 
        schedstat_inc(rq, ttwu_count);
@@ -2455,9 +2482,6 @@ ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
        if (wake_flags & WF_SYNC)
                schedstat_inc(p, se.statistics.nr_wakeups_sync);
 
-       if (cpu != task_cpu(p))
-               schedstat_inc(p, se.statistics.nr_wakeups_migrate);
-
 #endif /* CONFIG_SCHEDSTATS */
 }
 
@@ -2485,7 +2509,7 @@ ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
        if (p->sched_class->task_woken)
                p->sched_class->task_woken(rq, p);
 
-       if (unlikely(rq->idle_stamp)) {
+       if (rq->idle_stamp) {
                u64 delta = rq->clock - rq->idle_stamp;
                u64 max = 2*sysctl_sched_migration_cost;
 
@@ -2532,13 +2556,9 @@ static int ttwu_remote(struct task_struct *p, int wake_flags)
 }
 
 #ifdef CONFIG_SMP
-static void sched_ttwu_pending(void)
+static void sched_ttwu_do_pending(struct task_struct *list)
 {
        struct rq *rq = this_rq();
-       struct task_struct *list = xchg(&rq->wake_list, NULL);
-
-       if (!list)
-               return;
 
        raw_spin_lock(&rq->lock);
 
@@ -2551,9 +2571,45 @@ static void sched_ttwu_pending(void)
        raw_spin_unlock(&rq->lock);
 }
 
+#ifdef CONFIG_HOTPLUG_CPU
+
+static void sched_ttwu_pending(void)
+{
+       struct rq *rq = this_rq();
+       struct task_struct *list = xchg(&rq->wake_list, NULL);
+
+       if (!list)
+               return;
+
+       sched_ttwu_do_pending(list);
+}
+
+#endif /* CONFIG_HOTPLUG_CPU */
+
 void scheduler_ipi(void)
 {
-       sched_ttwu_pending();
+       struct rq *rq = this_rq();
+       struct task_struct *list = xchg(&rq->wake_list, NULL);
+
+       if (!list)
+               return;
+
+       /*
+        * Not all reschedule IPI handlers call irq_enter/irq_exit, since
+        * traditionally all their work was done from the interrupt return
+        * path. Now that we actually do some work, we need to make sure
+        * we do call them.
+        *
+        * Some archs already do call them, luckily irq_enter/exit nest
+        * properly.
+        *
+        * Arguably we should visit all archs and update all handlers,
+        * however a fair share of IPIs are still resched only so this would
+        * somewhat pessimize the simple resched case.
+        */
+       irq_enter();
+       sched_ttwu_do_pending(list);
+       irq_exit();
 }
 
 static void ttwu_queue_remote(struct task_struct *p, int cpu)
@@ -2573,14 +2629,34 @@ static void ttwu_queue_remote(struct task_struct *p, int cpu)
        if (!next)
                smp_send_reschedule(cpu);
 }
-#endif
+
+#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
+static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
+{
+       struct rq *rq;
+       int ret = 0;
+
+       rq = __task_rq_lock(p);
+       if (p->on_cpu) {
+               ttwu_activate(rq, p, ENQUEUE_WAKEUP);
+               ttwu_do_wakeup(rq, p, wake_flags);
+               ret = 1;
+       }
+       __task_rq_unlock(rq);
+
+       return ret;
+
+}
+#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
+#endif /* CONFIG_SMP */
 
 static void ttwu_queue(struct task_struct *p, int cpu)
 {
        struct rq *rq = cpu_rq(cpu);
 
-#if defined(CONFIG_SMP) && defined(CONFIG_SCHED_TTWU_QUEUE)
+#if defined(CONFIG_SMP)
        if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
+               sched_clock_cpu(cpu); /* sync clocks x-cpu */
                ttwu_queue_remote(p, cpu);
                return;
        }
@@ -2631,17 +2707,17 @@ try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
        while (p->on_cpu) {
 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
                /*
-                * If called from interrupt context we could have landed in the
-                * middle of schedule(), in this case we should take care not
-                * to spin on ->on_cpu if p is current, since that would
-                * deadlock.
+                * In case the architecture enables interrupts in
+                * context_switch(), we cannot busy wait, since that
+                * would lead to deadlocks when an interrupt hits and
+                * tries to wake up @prev. So bail and do a complete
+                * remote wakeup.
                 */
-               if (p == current) {
-                       ttwu_queue(p, cpu);
+               if (ttwu_activate_remote(p, wake_flags))
                        goto stat;
-               }
-#endif
+#else
                cpu_relax();
+#endif
        }
        /*
         * Pairs with the smp_wmb() in finish_lock_switch().
@@ -2655,8 +2731,10 @@ try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
                p->sched_class->task_waking(p);
 
        cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
-       if (task_cpu(p) != cpu)
+       if (task_cpu(p) != cpu) {
+               wake_flags |= WF_MIGRATED;
                set_task_cpu(p, cpu);
+       }
 #endif /* CONFIG_SMP */
 
        ttwu_queue(p, cpu);
@@ -2820,7 +2898,7 @@ void sched_fork(struct task_struct *p)
 #if defined(CONFIG_SMP)
        p->on_cpu = 0;
 #endif
-#ifdef CONFIG_PREEMPT
+#ifdef CONFIG_PREEMPT_COUNT
        /* Want to start with kernel preemption disabled. */
        task_thread_info(p)->preempt_count = 1;
 #endif
@@ -2987,7 +3065,7 @@ static void finish_task_switch(struct rq *rq, struct task_struct *prev)
 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
        local_irq_disable();
 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
-       perf_event_task_sched_in(current);
+       perf_event_task_sched_in(prev, current);
 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
        local_irq_enable();
 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
@@ -3647,30 +3725,6 @@ unsigned long long task_sched_runtime(struct task_struct *p)
 }
 
 /*
- * Return sum_exec_runtime for the thread group.
- * In case the task is currently running, return the sum plus current's
- * pending runtime that have not been accounted yet.
- *
- * Note that the thread group might have other running tasks as well,
- * so the return value not includes other pending runtime that other
- * running tasks might have.
- */
-unsigned long long thread_group_sched_runtime(struct task_struct *p)
-{
-       struct task_cputime totals;
-       unsigned long flags;
-       struct rq *rq;
-       u64 ns;
-
-       rq = task_rq_lock(p, &flags);
-       thread_group_cputime(p, &totals);
-       ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
-       task_rq_unlock(rq, p, &flags);
-
-       return ns;
-}
-
-/*
  * Account user cpu time to a process.
  * @p: the process that the cpu time gets accounted to
  * @cputime: the cpu time spent in user space since the last update
@@ -3811,6 +3865,25 @@ void account_idle_time(cputime_t cputime)
                cpustat->idle = cputime64_add(cpustat->idle, cputime64);
 }
 
+static __always_inline bool steal_account_process_tick(void)
+{
+#ifdef CONFIG_PARAVIRT
+       if (static_branch(&paravirt_steal_enabled)) {
+               u64 steal, st = 0;
+
+               steal = paravirt_steal_clock(smp_processor_id());
+               steal -= this_rq()->prev_steal_time;
+
+               st = steal_ticks(steal);
+               this_rq()->prev_steal_time += st * TICK_NSEC;
+
+               account_steal_time(st);
+               return st;
+       }
+#endif
+       return false;
+}
+
 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
 
 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
@@ -3842,6 +3915,9 @@ static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
        cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
        struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
 
+       if (steal_account_process_tick())
+               return;
+
        if (irqtime_account_hi_update()) {
                cpustat->irq = cputime64_add(cpustat->irq, tmp);
        } else if (irqtime_account_si_update()) {
@@ -3895,6 +3971,9 @@ void account_process_tick(struct task_struct *p, int user_tick)
                return;
        }
 
+       if (steal_account_process_tick())
+               return;
+
        if (user_tick)
                account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
        else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
@@ -4176,9 +4255,9 @@ pick_next_task(struct rq *rq)
 }
 
 /*
- * schedule() is the main scheduler function.
+ * __schedule() is the main scheduler function.
  */
-asmlinkage void __sched schedule(void)
+static void __sched __schedule(void)
 {
        struct task_struct *prev, *next;
        unsigned long *switch_count;
@@ -4219,16 +4298,6 @@ need_resched:
                                if (to_wakeup)
                                        try_to_wake_up_local(to_wakeup);
                        }
-
-                       /*
-                        * If we are going to sleep and we have plugged IO
-                        * queued, make sure to submit it to avoid deadlocks.
-                        */
-                       if (blk_needs_flush_plug(prev)) {
-                               raw_spin_unlock(&rq->lock);
-                               blk_schedule_flush_plug(prev);
-                               raw_spin_lock(&rq->lock);
-                       }
                }
                switch_count = &prev->nvcsw;
        }
@@ -4266,17 +4335,34 @@ need_resched:
        if (need_resched())
                goto need_resched;
 }
+
+static inline void sched_submit_work(struct task_struct *tsk)
+{
+       if (!tsk->state)
+               return;
+       /*
+        * If we are going to sleep and we have plugged IO queued,
+        * make sure to submit it to avoid deadlocks.
+        */
+       if (blk_needs_flush_plug(tsk))
+               blk_schedule_flush_plug(tsk);
+}
+
+asmlinkage void __sched schedule(void)
+{
+       struct task_struct *tsk = current;
+
+       sched_submit_work(tsk);
+       __schedule();
+}
 EXPORT_SYMBOL(schedule);
 
 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
 
 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
 {
-       bool ret = false;
-
-       rcu_read_lock();
        if (lock->owner != owner)
-               goto fail;
+               return false;
 
        /*
         * Ensure we emit the owner->on_cpu, dereference _after_ checking
@@ -4286,11 +4372,7 @@ static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
         */
        barrier();
 
-       ret = owner->on_cpu;
-fail:
-       rcu_read_unlock();
-
-       return ret;
+       return owner->on_cpu;
 }
 
 /*
@@ -4302,21 +4384,21 @@ int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
        if (!sched_feat(OWNER_SPIN))
                return 0;
 
+       rcu_read_lock();
        while (owner_running(lock, owner)) {
                if (need_resched())
-                       return 0;
+                       break;
 
                arch_mutex_cpu_relax();
        }
+       rcu_read_unlock();
 
        /*
-        * If the owner changed to another task there is likely
-        * heavy contention, stop spinning.
+        * We break out the loop above on need_resched() and when the
+        * owner changed, which is a sign for heavy contention. Return
+        * success only when lock->owner is NULL.
         */
-       if (lock->owner)
-               return 0;
-
-       return 1;
+       return lock->owner == NULL;
 }
 #endif
 
@@ -4339,7 +4421,7 @@ asmlinkage void __sched notrace preempt_schedule(void)
 
        do {
                add_preempt_count_notrace(PREEMPT_ACTIVE);
-               schedule();
+               __schedule();
                sub_preempt_count_notrace(PREEMPT_ACTIVE);
 
                /*
@@ -4367,7 +4449,7 @@ asmlinkage void __sched preempt_schedule_irq(void)
        do {
                add_preempt_count(PREEMPT_ACTIVE);
                local_irq_enable();
-               schedule();
+               __schedule();
                local_irq_disable();
                sub_preempt_count(PREEMPT_ACTIVE);
 
@@ -5492,7 +5574,7 @@ static inline int should_resched(void)
 static void __cond_resched(void)
 {
        add_preempt_count(PREEMPT_ACTIVE);
-       schedule();
+       __schedule();
        sub_preempt_count(PREEMPT_ACTIVE);
 }
 
@@ -5841,7 +5923,7 @@ void __cpuinit init_idle(struct task_struct *idle, int cpu)
        idle->state = TASK_RUNNING;
        idle->se.exec_start = sched_clock();
 
-       cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
+       do_set_cpus_allowed(idle, cpumask_of(cpu));
        /*
         * We're having a chicken and egg problem, even though we are
         * holding rq->lock, the cpu isn't yet set to this cpu so the
@@ -5929,6 +6011,16 @@ static inline void sched_init_granularity(void)
 }
 
 #ifdef CONFIG_SMP
+void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
+{
+       if (p->sched_class && p->sched_class->set_cpus_allowed)
+               p->sched_class->set_cpus_allowed(p, new_mask);
+       else {
+               cpumask_copy(&p->cpus_allowed, new_mask);
+               p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
+       }
+}
+
 /*
  * This is how migration works:
  *
@@ -5974,12 +6066,7 @@ int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
                goto out;
        }
 
-       if (p->sched_class->set_cpus_allowed)
-               p->sched_class->set_cpus_allowed(p, new_mask);
-       else {
-               cpumask_copy(&p->cpus_allowed, new_mask);
-               p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
-       }
+       do_set_cpus_allowed(p, new_mask);
 
        /* Can the task run on the task's current CPU? If so, we're done */
        if (cpumask_test_cpu(task_cpu(p), new_mask))
@@ -6518,7 +6605,7 @@ static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
                        break;
                }
 
-               if (!group->cpu_power) {
+               if (!group->sgp->power) {
                        printk(KERN_CONT "\n");
                        printk(KERN_ERR "ERROR: domain->cpu_power not "
                                        "set\n");
@@ -6542,9 +6629,9 @@ static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
                cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
 
                printk(KERN_CONT " %s", str);
-               if (group->cpu_power != SCHED_POWER_SCALE) {
+               if (group->sgp->power != SCHED_POWER_SCALE) {
                        printk(KERN_CONT " (cpu_power = %d)",
-                               group->cpu_power);
+                               group->sgp->power);
                }
 
                group = group->next;
@@ -6735,11 +6822,39 @@ static struct root_domain *alloc_rootdomain(void)
        return rd;
 }
 
+static void free_sched_groups(struct sched_group *sg, int free_sgp)
+{
+       struct sched_group *tmp, *first;
+
+       if (!sg)
+               return;
+
+       first = sg;
+       do {
+               tmp = sg->next;
+
+               if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
+                       kfree(sg->sgp);
+
+               kfree(sg);
+               sg = tmp;
+       } while (sg != first);
+}
+
 static void free_sched_domain(struct rcu_head *rcu)
 {
        struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
-       if (atomic_dec_and_test(&sd->groups->ref))
+
+       /*
+        * If its an overlapping domain it has private groups, iterate and
+        * nuke them all.
+        */
+       if (sd->flags & SD_OVERLAP) {
+               free_sched_groups(sd->groups, 1);
+       } else if (atomic_dec_and_test(&sd->groups->ref)) {
+               kfree(sd->groups->sgp);
                kfree(sd->groups);
+       }
        kfree(sd);
 }
 
@@ -6906,6 +7021,7 @@ int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
 struct sd_data {
        struct sched_domain **__percpu sd;
        struct sched_group **__percpu sg;
+       struct sched_group_power **__percpu sgp;
 };
 
 struct s_data {
@@ -6925,15 +7041,73 @@ struct sched_domain_topology_level;
 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
 
+#define SDTL_OVERLAP   0x01
+
 struct sched_domain_topology_level {
        sched_domain_init_f init;
        sched_domain_mask_f mask;
+       int                 flags;
        struct sd_data      data;
 };
 
-/*
- * Assumes the sched_domain tree is fully constructed
- */
+static int
+build_overlap_sched_groups(struct sched_domain *sd, int cpu)
+{
+       struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
+       const struct cpumask *span = sched_domain_span(sd);
+       struct cpumask *covered = sched_domains_tmpmask;
+       struct sd_data *sdd = sd->private;
+       struct sched_domain *child;
+       int i;
+
+       cpumask_clear(covered);
+
+       for_each_cpu(i, span) {
+               struct cpumask *sg_span;
+
+               if (cpumask_test_cpu(i, covered))
+                       continue;
+
+               sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
+                               GFP_KERNEL, cpu_to_node(i));
+
+               if (!sg)
+                       goto fail;
+
+               sg_span = sched_group_cpus(sg);
+
+               child = *per_cpu_ptr(sdd->sd, i);
+               if (child->child) {
+                       child = child->child;
+                       cpumask_copy(sg_span, sched_domain_span(child));
+               } else
+                       cpumask_set_cpu(i, sg_span);
+
+               cpumask_or(covered, covered, sg_span);
+
+               sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
+               atomic_inc(&sg->sgp->ref);
+
+               if (cpumask_test_cpu(cpu, sg_span))
+                       groups = sg;
+
+               if (!first)
+                       first = sg;
+               if (last)
+                       last->next = sg;
+               last = sg;
+               last->next = first;
+       }
+       sd->groups = groups;
+
+       return 0;
+
+fail:
+       free_sched_groups(first, 0);
+
+       return -ENOMEM;
+}
+
 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
 {
        struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
@@ -6942,24 +7116,24 @@ static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
        if (child)
                cpu = cpumask_first(sched_domain_span(child));
 
-       if (sg)
+       if (sg) {
                *sg = *per_cpu_ptr(sdd->sg, cpu);
+               (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
+               atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
+       }
 
        return cpu;
 }
 
 /*
- * build_sched_groups takes the cpumask we wish to span, and a pointer
- * to a function which identifies what group(along with sched group) a CPU
- * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
- * (due to the fact that we keep track of groups covered with a struct cpumask).
- *
  * build_sched_groups will build a circular linked list of the groups
  * covered by the given span, and will set each group's ->cpumask correctly,
  * and ->cpu_power to 0.
+ *
+ * Assumes the sched_domain tree is fully constructed
  */
-static void
-build_sched_groups(struct sched_domain *sd)
+static int
+build_sched_groups(struct sched_domain *sd, int cpu)
 {
        struct sched_group *first = NULL, *last = NULL;
        struct sd_data *sdd = sd->private;
@@ -6967,6 +7141,12 @@ build_sched_groups(struct sched_domain *sd)
        struct cpumask *covered;
        int i;
 
+       get_group(cpu, sdd, &sd->groups);
+       atomic_inc(&sd->groups->ref);
+
+       if (cpu != cpumask_first(sched_domain_span(sd)))
+               return 0;
+
        lockdep_assert_held(&sched_domains_mutex);
        covered = sched_domains_tmpmask;
 
@@ -6981,7 +7161,7 @@ build_sched_groups(struct sched_domain *sd)
                        continue;
 
                cpumask_clear(sched_group_cpus(sg));
-               sg->cpu_power = 0;
+               sg->sgp->power = 0;
 
                for_each_cpu(j, span) {
                        if (get_group(j, sdd, NULL) != group)
@@ -6998,6 +7178,8 @@ build_sched_groups(struct sched_domain *sd)
                last = sg;
        }
        last->next = first;
+
+       return 0;
 }
 
 /*
@@ -7012,12 +7194,17 @@ build_sched_groups(struct sched_domain *sd)
  */
 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
 {
-       WARN_ON(!sd || !sd->groups);
+       struct sched_group *sg = sd->groups;
 
-       if (cpu != group_first_cpu(sd->groups))
-               return;
+       WARN_ON(!sd || !sg);
 
-       sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups));
+       do {
+               sg->group_weight = cpumask_weight(sched_group_cpus(sg));
+               sg = sg->next;
+       } while (sg != sd->groups);
+
+       if (cpu != group_first_cpu(sg))
+               return;
 
        update_group_power(sd, cpu);
 }
@@ -7138,15 +7325,15 @@ static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
 static void claim_allocations(int cpu, struct sched_domain *sd)
 {
        struct sd_data *sdd = sd->private;
-       struct sched_group *sg = sd->groups;
 
        WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
        *per_cpu_ptr(sdd->sd, cpu) = NULL;
 
-       if (cpu == cpumask_first(sched_group_cpus(sg))) {
-               WARN_ON_ONCE(*per_cpu_ptr(sdd->sg, cpu) != sg);
+       if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
                *per_cpu_ptr(sdd->sg, cpu) = NULL;
-       }
+
+       if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
+               *per_cpu_ptr(sdd->sgp, cpu) = NULL;
 }
 
 #ifdef CONFIG_SCHED_SMT
@@ -7171,7 +7358,7 @@ static struct sched_domain_topology_level default_topology[] = {
 #endif
        { sd_init_CPU, cpu_cpu_mask, },
 #ifdef CONFIG_NUMA
-       { sd_init_NODE, cpu_node_mask, },
+       { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
        { sd_init_ALLNODES, cpu_allnodes_mask, },
 #endif
        { NULL, },
@@ -7195,9 +7382,14 @@ static int __sdt_alloc(const struct cpumask *cpu_map)
                if (!sdd->sg)
                        return -ENOMEM;
 
+               sdd->sgp = alloc_percpu(struct sched_group_power *);
+               if (!sdd->sgp)
+                       return -ENOMEM;
+
                for_each_cpu(j, cpu_map) {
                        struct sched_domain *sd;
                        struct sched_group *sg;
+                       struct sched_group_power *sgp;
 
                        sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
                                        GFP_KERNEL, cpu_to_node(j));
@@ -7212,6 +7404,13 @@ static int __sdt_alloc(const struct cpumask *cpu_map)
                                return -ENOMEM;
 
                        *per_cpu_ptr(sdd->sg, j) = sg;
+
+                       sgp = kzalloc_node(sizeof(struct sched_group_power),
+                                       GFP_KERNEL, cpu_to_node(j));
+                       if (!sgp)
+                               return -ENOMEM;
+
+                       *per_cpu_ptr(sdd->sgp, j) = sgp;
                }
        }
 
@@ -7227,11 +7426,16 @@ static void __sdt_free(const struct cpumask *cpu_map)
                struct sd_data *sdd = &tl->data;
 
                for_each_cpu(j, cpu_map) {
+                       struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
+                       if (sd && (sd->flags & SD_OVERLAP))
+                               free_sched_groups(sd->groups, 0);
                        kfree(*per_cpu_ptr(sdd->sd, j));
                        kfree(*per_cpu_ptr(sdd->sg, j));
+                       kfree(*per_cpu_ptr(sdd->sgp, j));
                }
                free_percpu(sdd->sd);
                free_percpu(sdd->sg);
+               free_percpu(sdd->sgp);
        }
 }
 
@@ -7277,8 +7481,13 @@ static int build_sched_domains(const struct cpumask *cpu_map,
                struct sched_domain_topology_level *tl;
 
                sd = NULL;
-               for (tl = sched_domain_topology; tl->init; tl++)
+               for (tl = sched_domain_topology; tl->init; tl++) {
                        sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
+                       if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
+                               sd->flags |= SD_OVERLAP;
+                       if (cpumask_equal(cpu_map, sched_domain_span(sd)))
+                               break;
+               }
 
                while (sd->child)
                        sd = sd->child;
@@ -7290,13 +7499,13 @@ static int build_sched_domains(const struct cpumask *cpu_map,
        for_each_cpu(i, cpu_map) {
                for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
                        sd->span_weight = cpumask_weight(sched_domain_span(sd));
-                       get_group(i, sd->private, &sd->groups);
-                       atomic_inc(&sd->groups->ref);
-
-                       if (i != cpumask_first(sched_domain_span(sd)))
-                               continue;
-
-                       build_sched_groups(sd);
+                       if (sd->flags & SD_OVERLAP) {
+                               if (build_overlap_sched_groups(sd, i))
+                                       goto error;
+                       } else {
+                               if (build_sched_groups(sd, i))
+                                       goto error;
+                       }
                }
        }
 
@@ -7706,18 +7915,14 @@ int in_sched_functions(unsigned long addr)
                && addr < (unsigned long)__sched_text_end);
 }
 
-static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
+static void init_cfs_rq(struct cfs_rq *cfs_rq)
 {
        cfs_rq->tasks_timeline = RB_ROOT;
        INIT_LIST_HEAD(&cfs_rq->tasks);
-#ifdef CONFIG_FAIR_GROUP_SCHED
-       cfs_rq->rq = rq;
-       /* allow initial update_cfs_load() to truncate */
-#ifdef CONFIG_SMP
-       cfs_rq->load_stamp = 1;
-#endif
-#endif
        cfs_rq->min_vruntime = (u64)(-(1LL << 20));
+#ifndef CONFIG_64BIT
+       cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
+#endif
 }
 
 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
@@ -7733,27 +7938,18 @@ static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
        /* delimiter for bitsearch: */
        __set_bit(MAX_RT_PRIO, array->bitmap);
 
-#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
+#if defined CONFIG_SMP
        rt_rq->highest_prio.curr = MAX_RT_PRIO;
-#ifdef CONFIG_SMP
        rt_rq->highest_prio.next = MAX_RT_PRIO;
-#endif
-#endif
-#ifdef CONFIG_SMP
        rt_rq->rt_nr_migratory = 0;
        rt_rq->overloaded = 0;
-       plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
+       plist_head_init(&rt_rq->pushable_tasks);
 #endif
 
        rt_rq->rt_time = 0;
        rt_rq->rt_throttled = 0;
        rt_rq->rt_runtime = 0;
        raw_spin_lock_init(&rt_rq->rt_runtime_lock);
-
-#ifdef CONFIG_RT_GROUP_SCHED
-       rt_rq->rt_nr_boosted = 0;
-       rt_rq->rq = rq;
-#endif
 }
 
 #ifdef CONFIG_FAIR_GROUP_SCHED
@@ -7762,11 +7958,17 @@ static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
                                struct sched_entity *parent)
 {
        struct rq *rq = cpu_rq(cpu);
-       tg->cfs_rq[cpu] = cfs_rq;
-       init_cfs_rq(cfs_rq, rq);
+
        cfs_rq->tg = tg;
+       cfs_rq->rq = rq;
+#ifdef CONFIG_SMP
+       /* allow initial update_cfs_load() to truncate */
+       cfs_rq->load_stamp = 1;
+#endif
 
+       tg->cfs_rq[cpu] = cfs_rq;
        tg->se[cpu] = se;
+
        /* se could be NULL for root_task_group */
        if (!se)
                return;
@@ -7789,12 +7991,14 @@ static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
 {
        struct rq *rq = cpu_rq(cpu);
 
-       tg->rt_rq[cpu] = rt_rq;
-       init_rt_rq(rt_rq, rq);
+       rt_rq->highest_prio.curr = MAX_RT_PRIO;
+       rt_rq->rt_nr_boosted = 0;
+       rt_rq->rq = rq;
        rt_rq->tg = tg;
-       rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
 
+       tg->rt_rq[cpu] = rt_rq;
        tg->rt_se[cpu] = rt_se;
+
        if (!rt_se)
                return;
 
@@ -7876,7 +8080,7 @@ void __init sched_init(void)
                rq->nr_running = 0;
                rq->calc_load_active = 0;
                rq->calc_load_update = jiffies + LOAD_FREQ;
-               init_cfs_rq(&rq->cfs, rq);
+               init_cfs_rq(&rq->cfs);
                init_rt_rq(&rq->rt, rq);
 #ifdef CONFIG_FAIR_GROUP_SCHED
                root_task_group.shares = root_task_group_load;
@@ -7947,7 +8151,7 @@ void __init sched_init(void)
 #endif
 
 #ifdef CONFIG_RT_MUTEXES
-       plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
+       plist_head_init(&init_task.pi_waiters);
 #endif
 
        /*
@@ -7990,7 +8194,7 @@ void __init sched_init(void)
        scheduler_running = 1;
 }
 
-#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
+#ifdef CONFIG_DEBUG_ATOMIC_SLEEP
 static inline int preempt_count_equals(int preempt_offset)
 {
        int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
@@ -7998,13 +8202,23 @@ static inline int preempt_count_equals(int preempt_offset)
        return (nested == preempt_offset);
 }
 
+static int __might_sleep_init_called;
+int __init __might_sleep_init(void)
+{
+       __might_sleep_init_called = 1;
+       return 0;
+}
+early_initcall(__might_sleep_init);
+
 void __might_sleep(const char *file, int line, int preempt_offset)
 {
-#ifdef in_atomic
        static unsigned long prev_jiffy;        /* ratelimiting */
 
        if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
-           system_state != SYSTEM_RUNNING || oops_in_progress)
+           oops_in_progress)
+               return;
+       if (system_state != SYSTEM_RUNNING &&
+           (!__might_sleep_init_called || system_state != SYSTEM_BOOTING))
                return;
        if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
                return;
@@ -8022,7 +8236,6 @@ void __might_sleep(const char *file, int line, int preempt_offset)
        if (irqs_disabled())
                print_irqtrace_events(current);
        dump_stack();
-#endif
 }
 EXPORT_SYMBOL(__might_sleep);
 #endif
@@ -8181,6 +8394,7 @@ int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
                if (!se)
                        goto err_free_rq;
 
+               init_cfs_rq(cfs_rq);
                init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
        }
 
@@ -8208,7 +8422,7 @@ static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
        list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
        raw_spin_unlock_irqrestore(&rq->lock, flags);
 }
-#else /* !CONFG_FAIR_GROUP_SCHED */
+#else /* !CONFIG_FAIR_GROUP_SCHED */
 static inline void free_fair_sched_group(struct task_group *tg)
 {
 }
@@ -8229,7 +8443,8 @@ static void free_rt_sched_group(struct task_group *tg)
 {
        int i;
 
-       destroy_rt_bandwidth(&tg->rt_bandwidth);
+       if (tg->rt_se)
+               destroy_rt_bandwidth(&tg->rt_bandwidth);
 
        for_each_possible_cpu(i) {
                if (tg->rt_rq)
@@ -8270,6 +8485,8 @@ int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
                if (!rt_se)
                        goto err_free_rq;
 
+               init_rt_rq(rt_rq, cpu_rq(i));
+               rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
                init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
        }
 
@@ -8411,10 +8628,7 @@ int sched_group_set_shares(struct task_group *tg, unsigned long shares)
        if (!tg->se[0])
                return -EINVAL;
 
-       if (shares < MIN_SHARES)
-               shares = MIN_SHARES;
-       else if (shares > MAX_SHARES)
-               shares = MAX_SHARES;
+       shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
 
        mutex_lock(&shares_mutex);
        if (tg->shares == shares)
@@ -8764,42 +8978,10 @@ cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
        return 0;
 }
 
-static int
-cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
-                     struct task_struct *tsk, bool threadgroup)
-{
-       int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
-       if (retval)
-               return retval;
-       if (threadgroup) {
-               struct task_struct *c;
-               rcu_read_lock();
-               list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
-                       retval = cpu_cgroup_can_attach_task(cgrp, c);
-                       if (retval) {
-                               rcu_read_unlock();
-                               return retval;
-                       }
-               }
-               rcu_read_unlock();
-       }
-       return 0;
-}
-
 static void
-cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
-                 struct cgroup *old_cont, struct task_struct *tsk,
-                 bool threadgroup)
+cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
 {
        sched_move_task(tsk);
-       if (threadgroup) {
-               struct task_struct *c;
-               rcu_read_lock();
-               list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
-                       sched_move_task(c);
-               }
-               rcu_read_unlock();
-       }
 }
 
 static void
@@ -8887,8 +9069,8 @@ struct cgroup_subsys cpu_cgroup_subsys = {
        .name           = "cpu",
        .create         = cpu_cgroup_create,
        .destroy        = cpu_cgroup_destroy,
-       .can_attach     = cpu_cgroup_can_attach,
-       .attach         = cpu_cgroup_attach,
+       .can_attach_task = cpu_cgroup_can_attach_task,
+       .attach_task    = cpu_cgroup_attach_task,
        .exit           = cpu_cgroup_exit,
        .populate       = cpu_cgroup_populate,
        .subsys_id      = cpu_cgroup_subsys_id,