Merge branch 'linus' into tracing/mmiotrace-mergefixups
[linux-2.6.git] / kernel / sched.c
index f9c8da7..2a7ad35 100644 (file)
 #include <linux/hrtimer.h>
 #include <linux/tick.h>
 #include <linux/bootmem.h>
+#include <linux/debugfs.h>
+#include <linux/ctype.h>
+#include <linux/ftrace.h>
 
 #include <asm/tlb.h>
 #include <asm/irq_regs.h>
 
 /*
- * Scheduler clock - returns current time in nanosec units.
- * This is default implementation.
- * Architectures and sub-architectures can override this.
- */
-unsigned long long __attribute__((weak)) sched_clock(void)
-{
-       return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
-}
-
-/*
  * Convert user-nice values [ -20 ... 0 ... 19 ]
  * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
  * and back.
@@ -144,7 +137,7 @@ static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
 
 static inline int rt_policy(int policy)
 {
-       if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
+       if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
                return 1;
        return 0;
 }
@@ -240,6 +233,12 @@ static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
 }
 #endif
 
+/*
+ * sched_domains_mutex serializes calls to arch_init_sched_domains,
+ * detach_destroy_domains and partition_sched_domains.
+ */
+static DEFINE_MUTEX(sched_domains_mutex);
+
 #ifdef CONFIG_GROUP_SCHED
 
 #include <linux/cgroup.h>
@@ -271,9 +270,21 @@ struct task_group {
 
        struct rcu_head rcu;
        struct list_head list;
+
+       struct task_group *parent;
+       struct list_head siblings;
+       struct list_head children;
 };
 
 #ifdef CONFIG_USER_SCHED
+
+/*
+ * Root task group.
+ *     Every UID task group (including init_task_group aka UID-0) will
+ *     be a child to this group.
+ */
+struct task_group root_task_group;
+
 #ifdef CONFIG_FAIR_GROUP_SCHED
 /* Default task group's sched entity on each cpu */
 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
@@ -285,6 +296,8 @@ static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
 #endif
+#else
+#define root_task_group init_task_group
 #endif
 
 /* task_group_lock serializes add/remove of task groups and also changes to
@@ -292,9 +305,6 @@ static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
  */
 static DEFINE_SPINLOCK(task_group_lock);
 
-/* doms_cur_mutex serializes access to doms_cur[] array */
-static DEFINE_MUTEX(doms_cur_mutex);
-
 #ifdef CONFIG_FAIR_GROUP_SCHED
 #ifdef CONFIG_USER_SCHED
 # define INIT_TASK_GROUP_LOAD  (2*NICE_0_LOAD)
@@ -302,6 +312,17 @@ static DEFINE_MUTEX(doms_cur_mutex);
 # define INIT_TASK_GROUP_LOAD  NICE_0_LOAD
 #endif
 
+/*
+ * A weight of 0 or 1 can cause arithmetics problems.
+ * A weight of a cfs_rq is the sum of weights of which entities
+ * are queued on this cfs_rq, so a weight of a entity should not be
+ * too large, so as the shares value of a task group.
+ * (The default weight is 1024 - so there's no practical
+ *  limitation from this.)
+ */
+#define MIN_SHARES     2
+#define MAX_SHARES     (1UL << 18)
+
 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
 #endif
 
@@ -340,21 +361,9 @@ static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
 #endif
 }
 
-static inline void lock_doms_cur(void)
-{
-       mutex_lock(&doms_cur_mutex);
-}
-
-static inline void unlock_doms_cur(void)
-{
-       mutex_unlock(&doms_cur_mutex);
-}
-
 #else
 
 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
-static inline void lock_doms_cur(void) { }
-static inline void unlock_doms_cur(void) { }
 
 #endif /* CONFIG_GROUP_SCHED */
 
@@ -368,8 +377,12 @@ struct cfs_rq {
 
        struct rb_root tasks_timeline;
        struct rb_node *rb_leftmost;
-       struct rb_node *rb_load_balance_curr;
-       /* 'curr' points to currently running entity on this cfs_rq.
+
+       struct list_head tasks;
+       struct list_head *balance_iterator;
+
+       /*
+        * 'curr' points to currently running entity on this cfs_rq.
         * It is set to NULL otherwise (i.e when none are currently running).
         */
        struct sched_entity *curr, *next;
@@ -501,13 +514,7 @@ struct rq {
        unsigned long next_balance;
        struct mm_struct *prev_mm;
 
-       u64 clock, prev_clock_raw;
-       s64 clock_max_delta;
-
-       unsigned int clock_warps, clock_overflows, clock_underflows;
-       u64 idle_clock;
-       unsigned int clock_deep_idle_events;
-       u64 tick_timestamp;
+       u64 clock;
 
        atomic_t nr_iowait;
 
@@ -572,82 +579,6 @@ static inline int cpu_of(struct rq *rq)
 #endif
 }
 
-#ifdef CONFIG_NO_HZ
-static inline bool nohz_on(int cpu)
-{
-       return tick_get_tick_sched(cpu)->nohz_mode != NOHZ_MODE_INACTIVE;
-}
-
-static inline u64 max_skipped_ticks(struct rq *rq)
-{
-       return nohz_on(cpu_of(rq)) ? jiffies - rq->last_tick_seen + 2 : 1;
-}
-
-static inline void update_last_tick_seen(struct rq *rq)
-{
-       rq->last_tick_seen = jiffies;
-}
-#else
-static inline u64 max_skipped_ticks(struct rq *rq)
-{
-       return 1;
-}
-
-static inline void update_last_tick_seen(struct rq *rq)
-{
-}
-#endif
-
-/*
- * Update the per-runqueue clock, as finegrained as the platform can give
- * us, but without assuming monotonicity, etc.:
- */
-static void __update_rq_clock(struct rq *rq)
-{
-       u64 prev_raw = rq->prev_clock_raw;
-       u64 now = sched_clock();
-       s64 delta = now - prev_raw;
-       u64 clock = rq->clock;
-
-#ifdef CONFIG_SCHED_DEBUG
-       WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
-#endif
-       /*
-        * Protect against sched_clock() occasionally going backwards:
-        */
-       if (unlikely(delta < 0)) {
-               clock++;
-               rq->clock_warps++;
-       } else {
-               /*
-                * Catch too large forward jumps too:
-                */
-               u64 max_jump = max_skipped_ticks(rq) * TICK_NSEC;
-               u64 max_time = rq->tick_timestamp + max_jump;
-
-               if (unlikely(clock + delta > max_time)) {
-                       if (clock < max_time)
-                               clock = max_time;
-                       else
-                               clock++;
-                       rq->clock_overflows++;
-               } else {
-                       if (unlikely(delta > rq->clock_max_delta))
-                               rq->clock_max_delta = delta;
-                       clock += delta;
-               }
-       }
-
-       rq->prev_clock_raw = now;
-       rq->clock = clock;
-}
-
-static void update_rq_clock(struct rq *rq)
-{
-       if (likely(smp_processor_id() == cpu_of(rq)))
-               __update_rq_clock(rq);
-}
-
 /*
  * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
  * See detach_destroy_domains: synchronize_sched for details.
@@ -663,6 +594,11 @@ static void update_rq_clock(struct rq *rq)
 #define task_rq(p)             cpu_rq(task_cpu(p))
 #define cpu_curr(cpu)          (cpu_rq(cpu)->curr)
 
+static inline void update_rq_clock(struct rq *rq)
+{
+       rq->clock = sched_clock_cpu(cpu_of(rq));
+}
+
 /*
  * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
  */
@@ -672,33 +608,158 @@ static void update_rq_clock(struct rq *rq)
 # define const_debug static const
 #endif
 
+/**
+ * runqueue_is_locked
+ *
+ * Returns true if the current cpu runqueue is locked.
+ * This interface allows printk to be called with the runqueue lock
+ * held and know whether or not it is OK to wake up the klogd.
+ */
+int runqueue_is_locked(void)
+{
+       int cpu = get_cpu();
+       struct rq *rq = cpu_rq(cpu);
+       int ret;
+
+       ret = spin_is_locked(&rq->lock);
+       put_cpu();
+       return ret;
+}
+
 /*
  * Debugging: various feature bits
  */
+
+#define SCHED_FEAT(name, enabled)      \
+       __SCHED_FEAT_##name ,
+
 enum {
-       SCHED_FEAT_NEW_FAIR_SLEEPERS    = 1,
-       SCHED_FEAT_WAKEUP_PREEMPT       = 2,
-       SCHED_FEAT_START_DEBIT          = 4,
-       SCHED_FEAT_AFFINE_WAKEUPS       = 8,
-       SCHED_FEAT_CACHE_HOT_BUDDY      = 16,
-       SCHED_FEAT_SYNC_WAKEUPS         = 32,
-       SCHED_FEAT_HRTICK               = 64,
-       SCHED_FEAT_DOUBLE_TICK          = 128,
-       SCHED_FEAT_NORMALIZED_SLEEPER   = 256,
+#include "sched_features.h"
 };
 
+#undef SCHED_FEAT
+
+#define SCHED_FEAT(name, enabled)      \
+       (1UL << __SCHED_FEAT_##name) * enabled |
+
 const_debug unsigned int sysctl_sched_features =
-               SCHED_FEAT_NEW_FAIR_SLEEPERS    * 1 |
-               SCHED_FEAT_WAKEUP_PREEMPT       * 1 |
-               SCHED_FEAT_START_DEBIT          * 1 |
-               SCHED_FEAT_AFFINE_WAKEUPS       * 1 |
-               SCHED_FEAT_CACHE_HOT_BUDDY      * 1 |
-               SCHED_FEAT_SYNC_WAKEUPS         * 1 |
-               SCHED_FEAT_HRTICK               * 1 |
-               SCHED_FEAT_DOUBLE_TICK          * 0 |
-               SCHED_FEAT_NORMALIZED_SLEEPER   * 1;
+#include "sched_features.h"
+       0;
 
-#define sched_feat(x) (sysctl_sched_features & SCHED_FEAT_##x)
+#undef SCHED_FEAT
+
+#ifdef CONFIG_SCHED_DEBUG
+#define SCHED_FEAT(name, enabled)      \
+       #name ,
+
+static __read_mostly char *sched_feat_names[] = {
+#include "sched_features.h"
+       NULL
+};
+
+#undef SCHED_FEAT
+
+static int sched_feat_open(struct inode *inode, struct file *filp)
+{
+       filp->private_data = inode->i_private;
+       return 0;
+}
+
+static ssize_t
+sched_feat_read(struct file *filp, char __user *ubuf,
+               size_t cnt, loff_t *ppos)
+{
+       char *buf;
+       int r = 0;
+       int len = 0;
+       int i;
+
+       for (i = 0; sched_feat_names[i]; i++) {
+               len += strlen(sched_feat_names[i]);
+               len += 4;
+       }
+
+       buf = kmalloc(len + 2, GFP_KERNEL);
+       if (!buf)
+               return -ENOMEM;
+
+       for (i = 0; sched_feat_names[i]; i++) {
+               if (sysctl_sched_features & (1UL << i))
+                       r += sprintf(buf + r, "%s ", sched_feat_names[i]);
+               else
+                       r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
+       }
+
+       r += sprintf(buf + r, "\n");
+       WARN_ON(r >= len + 2);
+
+       r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
+
+       kfree(buf);
+
+       return r;
+}
+
+static ssize_t
+sched_feat_write(struct file *filp, const char __user *ubuf,
+               size_t cnt, loff_t *ppos)
+{
+       char buf[64];
+       char *cmp = buf;
+       int neg = 0;
+       int i;
+
+       if (cnt > 63)
+               cnt = 63;
+
+       if (copy_from_user(&buf, ubuf, cnt))
+               return -EFAULT;
+
+       buf[cnt] = 0;
+
+       if (strncmp(buf, "NO_", 3) == 0) {
+               neg = 1;
+               cmp += 3;
+       }
+
+       for (i = 0; sched_feat_names[i]; i++) {
+               int len = strlen(sched_feat_names[i]);
+
+               if (strncmp(cmp, sched_feat_names[i], len) == 0) {
+                       if (neg)
+                               sysctl_sched_features &= ~(1UL << i);
+                       else
+                               sysctl_sched_features |= (1UL << i);
+                       break;
+               }
+       }
+
+       if (!sched_feat_names[i])
+               return -EINVAL;
+
+       filp->f_pos += cnt;
+
+       return cnt;
+}
+
+static struct file_operations sched_feat_fops = {
+       .open   = sched_feat_open,
+       .read   = sched_feat_read,
+       .write  = sched_feat_write,
+};
+
+static __init int sched_init_debug(void)
+{
+       debugfs_create_file("sched_features", 0644, NULL, NULL,
+                       &sched_feat_fops);
+
+       return 0;
+}
+late_initcall(sched_init_debug);
+
+#endif
+
+#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
 
 /*
  * Number of tasks to iterate in a single balance run.
@@ -733,7 +794,7 @@ static inline u64 global_rt_runtime(void)
        return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
 }
 
-static const unsigned long long time_sync_thresh = 100000;
+unsigned long long time_sync_thresh = 100000;
 
 static DEFINE_PER_CPU(unsigned long long, time_offset);
 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
@@ -747,11 +808,14 @@ static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
 static DEFINE_SPINLOCK(time_sync_lock);
 static unsigned long long prev_global_time;
 
-static unsigned long long __sync_cpu_clock(cycles_t time, int cpu)
+static unsigned long long __sync_cpu_clock(unsigned long long time, int cpu)
 {
-       unsigned long flags;
-
-       spin_lock_irqsave(&time_sync_lock, flags);
+       /*
+        * We want this inlined, to not get tracer function calls
+        * in this critical section:
+        */
+       spin_acquire(&time_sync_lock.dep_map, 0, 0, _THIS_IP_);
+       __raw_spin_lock(&time_sync_lock.raw_lock);
 
        if (time < prev_global_time) {
                per_cpu(time_offset, cpu) += prev_global_time - time;
@@ -760,7 +824,8 @@ static unsigned long long __sync_cpu_clock(cycles_t time, int cpu)
                prev_global_time = time;
        }
 
-       spin_unlock_irqrestore(&time_sync_lock, flags);
+       __raw_spin_unlock(&time_sync_lock.raw_lock);
+       spin_release(&time_sync_lock.dep_map, 1, _THIS_IP_);
 
        return time;
 }
@@ -768,8 +833,6 @@ static unsigned long long __sync_cpu_clock(cycles_t time, int cpu)
 static unsigned long long __cpu_clock(int cpu)
 {
        unsigned long long now;
-       unsigned long flags;
-       struct rq *rq;
 
        /*
         * Only call sched_clock() if the scheduler has already been
@@ -778,11 +841,7 @@ static unsigned long long __cpu_clock(int cpu)
        if (unlikely(!scheduler_running))
                return 0;
 
-       local_irq_save(flags);
-       rq = cpu_rq(cpu);
-       update_rq_clock(rq);
-       now = rq->clock;
-       local_irq_restore(flags);
+       now = sched_clock_cpu(cpu);
 
        return now;
 }
@@ -794,13 +853,18 @@ static unsigned long long __cpu_clock(int cpu)
 unsigned long long cpu_clock(int cpu)
 {
        unsigned long long prev_cpu_time, time, delta_time;
+       unsigned long flags;
 
+       local_irq_save(flags);
        prev_cpu_time = per_cpu(prev_cpu_time, cpu);
        time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
        delta_time = time-prev_cpu_time;
 
-       if (unlikely(delta_time > time_sync_thresh))
+       if (unlikely(delta_time > time_sync_thresh)) {
                time = __sync_cpu_clock(time, cpu);
+               per_cpu(prev_cpu_time, cpu) = time;
+       }
+       local_irq_restore(flags);
 
        return time;
 }
@@ -951,43 +1015,6 @@ static struct rq *this_rq_lock(void)
        return rq;
 }
 
-/*
- * We are going deep-idle (irqs are disabled):
- */
-void sched_clock_idle_sleep_event(void)
-{
-       struct rq *rq = cpu_rq(smp_processor_id());
-
-       spin_lock(&rq->lock);
-       __update_rq_clock(rq);
-       spin_unlock(&rq->lock);
-       rq->clock_deep_idle_events++;
-}
-EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
-
-/*
- * We just idled delta nanoseconds (called with irqs disabled):
- */
-void sched_clock_idle_wakeup_event(u64 delta_ns)
-{
-       struct rq *rq = cpu_rq(smp_processor_id());
-       u64 now = sched_clock();
-
-       rq->idle_clock += delta_ns;
-       /*
-        * Override the previous timestamp and ignore all
-        * sched_clock() deltas that occured while we idled,
-        * and use the PM-provided delta_ns to advance the
-        * rq clock:
-        */
-       spin_lock(&rq->lock);
-       rq->prev_clock_raw = now;
-       rq->clock += delta_ns;
-       spin_unlock(&rq->lock);
-       touch_softlockup_watchdog();
-}
-EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
-
 static void __resched_task(struct task_struct *p, int tif_bit);
 
 static inline void resched_task(struct task_struct *p)
@@ -1023,6 +1050,7 @@ static inline void resched_rq(struct rq *rq)
 enum {
        HRTICK_SET,             /* re-programm hrtick_timer */
        HRTICK_RESET,           /* not a new slice */
+       HRTICK_BLOCK,           /* stop hrtick operations */
 };
 
 /*
@@ -1034,6 +1062,8 @@ static inline int hrtick_enabled(struct rq *rq)
 {
        if (!sched_feat(HRTICK))
                return 0;
+       if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags)))
+               return 0;
        return hrtimer_is_hres_active(&rq->hrtick_timer);
 }
 
@@ -1109,14 +1139,72 @@ static enum hrtimer_restart hrtick(struct hrtimer *timer)
        WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
 
        spin_lock(&rq->lock);
-       __update_rq_clock(rq);
+       update_rq_clock(rq);
        rq->curr->sched_class->task_tick(rq, rq->curr, 1);
        spin_unlock(&rq->lock);
 
        return HRTIMER_NORESTART;
 }
 
-static inline void init_rq_hrtick(struct rq *rq)
+#ifdef CONFIG_SMP
+static void hotplug_hrtick_disable(int cpu)
+{
+       struct rq *rq = cpu_rq(cpu);
+       unsigned long flags;
+
+       spin_lock_irqsave(&rq->lock, flags);
+       rq->hrtick_flags = 0;
+       __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
+       spin_unlock_irqrestore(&rq->lock, flags);
+
+       hrtick_clear(rq);
+}
+
+static void hotplug_hrtick_enable(int cpu)
+{
+       struct rq *rq = cpu_rq(cpu);
+       unsigned long flags;
+
+       spin_lock_irqsave(&rq->lock, flags);
+       __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
+       spin_unlock_irqrestore(&rq->lock, flags);
+}
+
+static int
+hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
+{
+       int cpu = (int)(long)hcpu;
+
+       switch (action) {
+       case CPU_UP_CANCELED:
+       case CPU_UP_CANCELED_FROZEN:
+       case CPU_DOWN_PREPARE:
+       case CPU_DOWN_PREPARE_FROZEN:
+       case CPU_DEAD:
+       case CPU_DEAD_FROZEN:
+               hotplug_hrtick_disable(cpu);
+               return NOTIFY_OK;
+
+       case CPU_UP_PREPARE:
+       case CPU_UP_PREPARE_FROZEN:
+       case CPU_DOWN_FAILED:
+       case CPU_DOWN_FAILED_FROZEN:
+       case CPU_ONLINE:
+       case CPU_ONLINE_FROZEN:
+               hotplug_hrtick_enable(cpu);
+               return NOTIFY_OK;
+       }
+
+       return NOTIFY_DONE;
+}
+
+static void init_hrtick(void)
+{
+       hotcpu_notifier(hotplug_hrtick, 0);
+}
+#endif /* CONFIG_SMP */
+
+static void init_rq_hrtick(struct rq *rq)
 {
        rq->hrtick_flags = 0;
        hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
@@ -1153,6 +1241,10 @@ static inline void init_rq_hrtick(struct rq *rq)
 void hrtick_resched(void)
 {
 }
+
+static inline void init_hrtick(void)
+{
+}
 #endif
 
 /*
@@ -1269,8 +1361,13 @@ calc_delta_mine(unsigned long delta_exec, unsigned long weight,
 {
        u64 tmp;
 
-       if (unlikely(!lw->inv_weight))
-               lw->inv_weight = (WMULT_CONST-lw->weight/2) / (lw->weight+1);
+       if (!lw->inv_weight) {
+               if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
+                       lw->inv_weight = 1;
+               else
+                       lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
+                               / (lw->weight+1);
+       }
 
        tmp = (u64)delta_exec * weight;
        /*
@@ -1388,11 +1485,29 @@ static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
 #endif
 
+static inline void inc_cpu_load(struct rq *rq, unsigned long load)
+{
+       update_load_add(&rq->load, load);
+}
+
+static inline void dec_cpu_load(struct rq *rq, unsigned long load)
+{
+       update_load_sub(&rq->load, load);
+}
+
 #ifdef CONFIG_SMP
 static unsigned long source_load(int cpu, int type);
 static unsigned long target_load(int cpu, int type);
 static unsigned long cpu_avg_load_per_task(int cpu);
 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
+#else /* CONFIG_SMP */
+
+#ifdef CONFIG_FAIR_GROUP_SCHED
+static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
+{
+}
+#endif
+
 #endif /* CONFIG_SMP */
 
 #include "sched_stats.h"
@@ -2053,6 +2168,9 @@ out_activate:
        success = 1;
 
 out_running:
+       trace_mark(kernel_sched_wakeup,
+               "pid %d state %ld ## rq %p task %p rq->curr %p",
+               p->pid, p->state, rq, p, rq->curr);
        check_preempt_curr(rq, p);
 
        p->state = TASK_RUNNING;
@@ -2105,6 +2223,7 @@ static void __sched_fork(struct task_struct *p)
 
        INIT_LIST_HEAD(&p->rt.run_list);
        p->se.on_rq = 0;
+       INIT_LIST_HEAD(&p->se.group_node);
 
 #ifdef CONFIG_PREEMPT_NOTIFIERS
        INIT_HLIST_HEAD(&p->preempt_notifiers);
@@ -2182,6 +2301,9 @@ void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
                p->sched_class->task_new(rq, p);
                inc_nr_running(p, rq);
        }
+       trace_mark(kernel_sched_wakeup_new,
+               "pid %d state %ld ## rq %p task %p rq->curr %p",
+               p->pid, p->state, rq, p, rq->curr);
        check_preempt_curr(rq, p);
 #ifdef CONFIG_SMP
        if (p->sched_class->task_wake_up)
@@ -2354,6 +2476,11 @@ context_switch(struct rq *rq, struct task_struct *prev,
        struct mm_struct *mm, *oldmm;
 
        prepare_task_switch(rq, prev, next);
+       trace_mark(kernel_sched_schedule,
+               "prev_pid %d next_pid %d prev_state %ld "
+               "## rq %p prev %p next %p",
+               prev->pid, next->pid, prev->state,
+               rq, prev, next);
        mm = next->mm;
        oldmm = prev->active_mm;
        /*
@@ -3842,8 +3969,10 @@ void account_system_time(struct task_struct *p, int hardirq_offset,
        struct rq *rq = this_rq();
        cputime64_t tmp;
 
-       if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
-               return account_guest_time(p, cputime);
+       if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
+               account_guest_time(p, cputime);
+               return;
+       }
 
        p->stime = cputime_add(p->stime, cputime);
 
@@ -3907,19 +4036,11 @@ void scheduler_tick(void)
        int cpu = smp_processor_id();
        struct rq *rq = cpu_rq(cpu);
        struct task_struct *curr = rq->curr;
-       u64 next_tick = rq->tick_timestamp + TICK_NSEC;
+
+       sched_clock_tick();
 
        spin_lock(&rq->lock);
-       __update_rq_clock(rq);
-       /*
-        * Let rq->clock advance by at least TICK_NSEC:
-        */
-       if (unlikely(rq->clock < next_tick)) {
-               rq->clock = next_tick;
-               rq->clock_underflows++;
-       }
-       rq->tick_timestamp = rq->clock;
-       update_last_tick_seen(rq);
+       update_rq_clock(rq);
        update_cpu_load(rq);
        curr->sched_class->task_tick(rq, curr, 0);
        spin_unlock(&rq->lock);
@@ -3930,26 +4051,44 @@ void scheduler_tick(void)
 #endif
 }
 
-#if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
+#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
+                               defined(CONFIG_PREEMPT_TRACER))
+
+static inline unsigned long get_parent_ip(unsigned long addr)
+{
+       if (in_lock_functions(addr)) {
+               addr = CALLER_ADDR2;
+               if (in_lock_functions(addr))
+                       addr = CALLER_ADDR3;
+       }
+       return addr;
+}
 
 void __kprobes add_preempt_count(int val)
 {
+#ifdef CONFIG_DEBUG_PREEMPT
        /*
         * Underflow?
         */
        if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
                return;
+#endif
        preempt_count() += val;
+#ifdef CONFIG_DEBUG_PREEMPT
        /*
         * Spinlock count overflowing soon?
         */
        DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
                                PREEMPT_MASK - 10);
+#endif
+       if (preempt_count() == val)
+               trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
 }
 EXPORT_SYMBOL(add_preempt_count);
 
 void __kprobes sub_preempt_count(int val)
 {
+#ifdef CONFIG_DEBUG_PREEMPT
        /*
         * Underflow?
         */
@@ -3961,7 +4100,10 @@ void __kprobes sub_preempt_count(int val)
        if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
                        !(preempt_count() & PREEMPT_MASK)))
                return;
+#endif
 
+       if (preempt_count() == val)
+               trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
        preempt_count() -= val;
 }
 EXPORT_SYMBOL(sub_preempt_count);
@@ -3998,7 +4140,7 @@ static inline void schedule_debug(struct task_struct *prev)
         * schedule() atomically, we ignore that path for now.
         * Otherwise, whine if we are scheduling when we should not be.
         */
-       if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
+       if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
                __schedule_bug(prev);
 
        profile_hit(SCHED_PROFILING, __builtin_return_address(0));
@@ -4073,17 +4215,15 @@ need_resched_nonpreemptible:
         * Do the rq-clock update outside the rq lock:
         */
        local_irq_disable();
-       __update_rq_clock(rq);
+       update_rq_clock(rq);
        spin_lock(&rq->lock);
        clear_tsk_need_resched(prev);
 
        if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
-               if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
-                               signal_pending(prev))) {
+               if (unlikely(signal_pending_state(prev->state, prev)))
                        prev->state = TASK_RUNNING;
-               } else {
+               else
                        deactivate_task(rq, prev, 1);
-               }
                switch_count = &prev->nvcsw;
        }
 
@@ -4098,9 +4238,9 @@ need_resched_nonpreemptible:
        prev->sched_class->put_prev_task(rq, prev);
        next = pick_next_task(rq, prev);
 
-       sched_info_switch(prev, next);
-
        if (likely(prev != next)) {
+               sched_info_switch(prev, next);
+
                rq->nr_switches++;
                rq->curr = next;
                ++*switch_count;
@@ -4135,8 +4275,6 @@ EXPORT_SYMBOL(schedule);
 asmlinkage void __sched preempt_schedule(void)
 {
        struct thread_info *ti = current_thread_info();
-       struct task_struct *task = current;
-       int saved_lock_depth;
 
        /*
         * If there is a non-zero preempt_count or interrupts are disabled,
@@ -4147,16 +4285,7 @@ asmlinkage void __sched preempt_schedule(void)
 
        do {
                add_preempt_count(PREEMPT_ACTIVE);
-
-               /*
-                * We keep the big kernel semaphore locked, but we
-                * clear ->lock_depth so that schedule() doesnt
-                * auto-release the semaphore:
-                */
-               saved_lock_depth = task->lock_depth;
-               task->lock_depth = -1;
                schedule();
-               task->lock_depth = saved_lock_depth;
                sub_preempt_count(PREEMPT_ACTIVE);
 
                /*
@@ -4177,26 +4306,15 @@ EXPORT_SYMBOL(preempt_schedule);
 asmlinkage void __sched preempt_schedule_irq(void)
 {
        struct thread_info *ti = current_thread_info();
-       struct task_struct *task = current;
-       int saved_lock_depth;
 
        /* Catch callers which need to be fixed */
        BUG_ON(ti->preempt_count || !irqs_disabled());
 
        do {
                add_preempt_count(PREEMPT_ACTIVE);
-
-               /*
-                * We keep the big kernel semaphore locked, but we
-                * clear ->lock_depth so that schedule() doesnt
-                * auto-release the semaphore:
-                */
-               saved_lock_depth = task->lock_depth;
-               task->lock_depth = -1;
                local_irq_enable();
                schedule();
                local_irq_disable();
-               task->lock_depth = saved_lock_depth;
                sub_preempt_count(PREEMPT_ACTIVE);
 
                /*
@@ -4331,22 +4449,20 @@ do_wait_for_common(struct completion *x, long timeout, int state)
                             signal_pending(current)) ||
                            (state == TASK_KILLABLE &&
                             fatal_signal_pending(current))) {
-                               __remove_wait_queue(&x->wait, &wait);
-                               return -ERESTARTSYS;
+                               timeout = -ERESTARTSYS;
+                               break;
                        }
                        __set_current_state(state);
                        spin_unlock_irq(&x->wait.lock);
                        timeout = schedule_timeout(timeout);
                        spin_lock_irq(&x->wait.lock);
-                       if (!timeout) {
-                               __remove_wait_queue(&x->wait, &wait);
-                               return timeout;
-                       }
-               } while (!x->done);
+               } while (!x->done && timeout);
                __remove_wait_queue(&x->wait, &wait);
+               if (!x->done)
+                       return timeout;
        }
        x->done--;
-       return timeout;
+       return timeout ?: 1;
 }
 
 static long __sched
@@ -5118,7 +5234,6 @@ static void __cond_resched(void)
        } while (need_resched());
 }
 
-#if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
 int __sched _cond_resched(void)
 {
        if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
@@ -5129,7 +5244,6 @@ int __sched _cond_resched(void)
        return 0;
 }
 EXPORT_SYMBOL(_cond_resched);
-#endif
 
 /*
  * cond_resched_lock() - if a reschedule is pending, drop the given lock,
@@ -5321,7 +5435,7 @@ out_unlock:
        return retval;
 }
 
-static const char stat_nam[] = "RSDTtZX";
+static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
 
 void sched_show_task(struct task_struct *p)
 {
@@ -5424,8 +5538,11 @@ void __cpuinit init_idle(struct task_struct *idle, int cpu)
        spin_unlock_irqrestore(&rq->lock, flags);
 
        /* Set the preempt count _outside_ the spinlocks! */
+#if defined(CONFIG_PREEMPT)
+       task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
+#else
        task_thread_info(idle)->preempt_count = 0;
-
+#endif
        /*
         * The idle tasks have their own, simple scheduling class:
         */
@@ -6498,6 +6615,7 @@ static int find_next_best_node(int node, nodemask_t *used_nodes)
 /**
  * sched_domain_node_span - get a cpumask for a node's sched_domain
  * @node: node whose cpumask we're constructing
+ * @span: resulting cpumask
  *
  * Given a node, construct a good cpumask for its sched_domain to span. It
  * should be one that prevents unnecessary balancing, but also spreads tasks
@@ -6757,6 +6875,7 @@ static noinline void sd_init_##type(struct sched_domain *sd)      \
 {                                                              \
        memset(sd, 0, sizeof(*sd));                             \
        *sd = SD_##type##_INIT;                                 \
+       sd->level = SD_LV_##type;                               \
 }
 
 SD_INIT_FUNC(CPU)
@@ -6805,11 +6924,47 @@ struct allmasks {
 #define        SCHED_CPUMASK_VAR(v, a)         cpumask_t *v = (cpumask_t *) \
                        ((unsigned long)(a) + offsetof(struct allmasks, v))
 
+static int default_relax_domain_level = -1;
+
+static int __init setup_relax_domain_level(char *str)
+{
+       unsigned long val;
+
+       val = simple_strtoul(str, NULL, 0);
+       if (val < SD_LV_MAX)
+               default_relax_domain_level = val;
+
+       return 1;
+}
+__setup("relax_domain_level=", setup_relax_domain_level);
+
+static void set_domain_attribute(struct sched_domain *sd,
+                                struct sched_domain_attr *attr)
+{
+       int request;
+
+       if (!attr || attr->relax_domain_level < 0) {
+               if (default_relax_domain_level < 0)
+                       return;
+               else
+                       request = default_relax_domain_level;
+       } else
+               request = attr->relax_domain_level;
+       if (request < sd->level) {
+               /* turn off idle balance on this domain */
+               sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
+       } else {
+               /* turn on idle balance on this domain */
+               sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
+       }
+}
+
 /*
  * Build sched domains for a given set of cpus and attach the sched domains
  * to the individual cpus
  */
-static int build_sched_domains(const cpumask_t *cpu_map)
+static int __build_sched_domains(const cpumask_t *cpu_map,
+                                struct sched_domain_attr *attr)
 {
        int i;
        struct root_domain *rd;
@@ -6873,6 +7028,7 @@ static int build_sched_domains(const cpumask_t *cpu_map)
                                SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
                        sd = &per_cpu(allnodes_domains, i);
                        SD_INIT(sd, ALLNODES);
+                       set_domain_attribute(sd, attr);
                        sd->span = *cpu_map;
                        cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
                        p = sd;
@@ -6882,6 +7038,7 @@ static int build_sched_domains(const cpumask_t *cpu_map)
 
                sd = &per_cpu(node_domains, i);
                SD_INIT(sd, NODE);
+               set_domain_attribute(sd, attr);
                sched_domain_node_span(cpu_to_node(i), &sd->span);
                sd->parent = p;
                if (p)
@@ -6892,6 +7049,7 @@ static int build_sched_domains(const cpumask_t *cpu_map)
                p = sd;
                sd = &per_cpu(phys_domains, i);
                SD_INIT(sd, CPU);
+               set_domain_attribute(sd, attr);
                sd->span = *nodemask;
                sd->parent = p;
                if (p)
@@ -6902,6 +7060,7 @@ static int build_sched_domains(const cpumask_t *cpu_map)
                p = sd;
                sd = &per_cpu(core_domains, i);
                SD_INIT(sd, MC);
+               set_domain_attribute(sd, attr);
                sd->span = cpu_coregroup_map(i);
                cpus_and(sd->span, sd->span, *cpu_map);
                sd->parent = p;
@@ -6913,6 +7072,7 @@ static int build_sched_domains(const cpumask_t *cpu_map)
                p = sd;
                sd = &per_cpu(cpu_domains, i);
                SD_INIT(sd, SIBLING);
+               set_domain_attribute(sd, attr);
                sd->span = per_cpu(cpu_sibling_map, i);
                cpus_and(sd->span, sd->span, *cpu_map);
                sd->parent = p;
@@ -7110,8 +7270,15 @@ error:
 #endif
 }
 
+static int build_sched_domains(const cpumask_t *cpu_map)
+{
+       return __build_sched_domains(cpu_map, NULL);
+}
+
 static cpumask_t *doms_cur;    /* current sched domains */
 static int ndoms_cur;          /* number of sched domains in 'doms_cur' */
+static struct sched_domain_attr *dattr_cur;
+                               /* attribues of custom domains in 'doms_cur' */
 
 /*
  * Special case: If a kmalloc of a doms_cur partition (array of
@@ -7125,6 +7292,18 @@ void __attribute__((weak)) arch_update_cpu_topology(void)
 }
 
 /*
+ * Free current domain masks.
+ * Called after all cpus are attached to NULL domain.
+ */
+static void free_sched_domains(void)
+{
+       ndoms_cur = 0;
+       if (doms_cur != &fallback_doms)
+               kfree(doms_cur);
+       doms_cur = &fallback_doms;
+}
+
+/*
  * Set up scheduler domains and groups. Callers must hold the hotplug lock.
  * For now this just excludes isolated cpus, but could be used to
  * exclude other special cases in the future.
@@ -7139,6 +7318,7 @@ static int arch_init_sched_domains(const cpumask_t *cpu_map)
        if (!doms_cur)
                doms_cur = &fallback_doms;
        cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
+       dattr_cur = NULL;
        err = build_sched_domains(doms_cur);
        register_sched_domain_sysctl();
 
@@ -7168,6 +7348,22 @@ static void detach_destroy_domains(const cpumask_t *cpu_map)
        arch_destroy_sched_domains(cpu_map, &tmpmask);
 }
 
+/* handle null as "default" */
+static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
+                       struct sched_domain_attr *new, int idx_new)
+{
+       struct sched_domain_attr tmp;
+
+       /* fast path */
+       if (!new && !cur)
+               return 1;
+
+       tmp = SD_ATTR_INIT;
+       return !memcmp(cur ? (cur + idx_cur) : &tmp,
+                       new ? (new + idx_new) : &tmp,
+                       sizeof(struct sched_domain_attr));
+}
+
 /*
  * Partition sched domains as specified by the 'ndoms_new'
  * cpumasks in the array doms_new[] of cpumasks. This compares
@@ -7189,11 +7385,12 @@ static void detach_destroy_domains(const cpumask_t *cpu_map)
  *
  * Call with hotplug lock held
  */
-void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
+void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
+                            struct sched_domain_attr *dattr_new)
 {
        int i, j;
 
-       lock_doms_cur();
+       mutex_lock(&sched_domains_mutex);
 
        /* always unregister in case we don't destroy any domains */
        unregister_sched_domain_sysctl();
@@ -7202,12 +7399,14 @@ void partition_sched_domains(int ndoms_new, cpumask_t *doms_new)
                ndoms_new = 1;
                doms_new = &fallback_doms;
                cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
+               dattr_new = NULL;
        }
 
        /* Destroy deleted domains */
        for (i = 0; i < ndoms_cur; i++) {
                for (j = 0; j < ndoms_new; j++) {
-                       if (cpus_equal(doms_cur[i], doms_new[j]))
+                       if (cpus_equal(doms_cur[i], doms_new[j])
+                           && dattrs_equal(dattr_cur, i, dattr_new, j))
                                goto match1;
                }
                /* no match - a current sched domain not in new doms_new[] */
@@ -7219,11 +7418,13 @@ match1:
        /* Build new domains */
        for (i = 0; i < ndoms_new; i++) {
                for (j = 0; j < ndoms_cur; j++) {
-                       if (cpus_equal(doms_new[i], doms_cur[j]))
+                       if (cpus_equal(doms_new[i], doms_cur[j])
+                           && dattrs_equal(dattr_new, i, dattr_cur, j))
                                goto match2;
                }
                /* no match - add a new doms_new */
-               build_sched_domains(doms_new + i);
+               __build_sched_domains(doms_new + i,
+                                       dattr_new ? dattr_new + i : NULL);
 match2:
                ;
        }
@@ -7231,12 +7432,14 @@ match2:
        /* Remember the new sched domains */
        if (doms_cur != &fallback_doms)
                kfree(doms_cur);
+       kfree(dattr_cur);       /* kfree(NULL) is safe */
        doms_cur = doms_new;
+       dattr_cur = dattr_new;
        ndoms_cur = ndoms_new;
 
        register_sched_domain_sysctl();
 
-       unlock_doms_cur();
+       mutex_unlock(&sched_domains_mutex);
 }
 
 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
@@ -7245,8 +7448,11 @@ int arch_reinit_sched_domains(void)
        int err;
 
        get_online_cpus();
+       mutex_lock(&sched_domains_mutex);
        detach_destroy_domains(&cpu_online_map);
+       free_sched_domains();
        err = arch_init_sched_domains(&cpu_online_map);
+       mutex_unlock(&sched_domains_mutex);
        put_online_cpus();
 
        return err;
@@ -7330,6 +7536,7 @@ static int update_sched_domains(struct notifier_block *nfb,
        case CPU_DOWN_PREPARE:
        case CPU_DOWN_PREPARE_FROZEN:
                detach_destroy_domains(&cpu_online_map);
+               free_sched_domains();
                return NOTIFY_OK;
 
        case CPU_UP_CANCELED:
@@ -7348,8 +7555,16 @@ static int update_sched_domains(struct notifier_block *nfb,
                return NOTIFY_DONE;
        }
 
+#ifndef CONFIG_CPUSETS
+       /*
+        * Create default domain partitioning if cpusets are disabled.
+        * Otherwise we let cpusets rebuild the domains based on the
+        * current setup.
+        */
+
        /* The hotplug lock is already held by cpu_up/cpu_down */
        arch_init_sched_domains(&cpu_online_map);
+#endif
 
        return NOTIFY_OK;
 }
@@ -7364,13 +7579,16 @@ void __init sched_init_smp(void)
        BUG_ON(sched_group_nodes_bycpu == NULL);
 #endif
        get_online_cpus();
+       mutex_lock(&sched_domains_mutex);
        arch_init_sched_domains(&cpu_online_map);
        cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
        if (cpus_empty(non_isolated_cpus))
                cpu_set(smp_processor_id(), non_isolated_cpus);
+       mutex_unlock(&sched_domains_mutex);
        put_online_cpus();
        /* XXX: Theoretical race here - CPU may be hotplugged now */
        hotcpu_notifier(update_sched_domains, 0);
+       init_hrtick();
 
        /* Move init over to a non-isolated CPU */
        if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
@@ -7380,11 +7598,6 @@ void __init sched_init_smp(void)
 #else
 void __init sched_init_smp(void)
 {
-#if defined(CONFIG_NUMA)
-       sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
-                                                               GFP_KERNEL);
-       BUG_ON(sched_group_nodes_bycpu == NULL);
-#endif
        sched_init_granularity();
 }
 #endif /* CONFIG_SMP */
@@ -7399,6 +7612,7 @@ int in_sched_functions(unsigned long addr)
 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
 {
        cfs_rq->tasks_timeline = RB_ROOT;
+       INIT_LIST_HEAD(&cfs_rq->tasks);
 #ifdef CONFIG_FAIR_GROUP_SCHED
        cfs_rq->rq = rq;
 #endif
@@ -7461,7 +7675,7 @@ static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
 
        se->my_q = cfs_rq;
        se->load.weight = tg->shares;
-       se->load.inv_weight = div64_64(1ULL<<32, se->load.weight);
+       se->load.inv_weight = 0;
        se->parent = parent;
 }
 #endif
@@ -7490,7 +7704,6 @@ static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
        else
                rt_se->rt_rq = parent->my_q;
 
-       rt_se->rt_rq = &rq->rt;
        rt_se->my_q = rt_rq;
        rt_se->parent = parent;
        INIT_LIST_HEAD(&rt_se->run_list);
@@ -7508,12 +7721,15 @@ void __init sched_init(void)
 #ifdef CONFIG_RT_GROUP_SCHED
        alloc_size += 2 * nr_cpu_ids * sizeof(void **);
 #endif
+#ifdef CONFIG_USER_SCHED
+       alloc_size *= 2;
+#endif
        /*
         * As sched_init() is called before page_alloc is setup,
         * we use alloc_bootmem().
         */
        if (alloc_size) {
-               ptr = (unsigned long)alloc_bootmem_low(alloc_size);
+               ptr = (unsigned long)alloc_bootmem(alloc_size);
 
 #ifdef CONFIG_FAIR_GROUP_SCHED
                init_task_group.se = (struct sched_entity **)ptr;
@@ -7521,12 +7737,29 @@ void __init sched_init(void)
 
                init_task_group.cfs_rq = (struct cfs_rq **)ptr;
                ptr += nr_cpu_ids * sizeof(void **);
+
+#ifdef CONFIG_USER_SCHED
+               root_task_group.se = (struct sched_entity **)ptr;
+               ptr += nr_cpu_ids * sizeof(void **);
+
+               root_task_group.cfs_rq = (struct cfs_rq **)ptr;
+               ptr += nr_cpu_ids * sizeof(void **);
+#endif
 #endif
 #ifdef CONFIG_RT_GROUP_SCHED
                init_task_group.rt_se = (struct sched_rt_entity **)ptr;
                ptr += nr_cpu_ids * sizeof(void **);
 
                init_task_group.rt_rq = (struct rt_rq **)ptr;
+               ptr += nr_cpu_ids * sizeof(void **);
+
+#ifdef CONFIG_USER_SCHED
+               root_task_group.rt_se = (struct sched_rt_entity **)ptr;
+               ptr += nr_cpu_ids * sizeof(void **);
+
+               root_task_group.rt_rq = (struct rt_rq **)ptr;
+               ptr += nr_cpu_ids * sizeof(void **);
+#endif
 #endif
        }
 
@@ -7540,10 +7773,21 @@ void __init sched_init(void)
 #ifdef CONFIG_RT_GROUP_SCHED
        init_rt_bandwidth(&init_task_group.rt_bandwidth,
                        global_rt_period(), global_rt_runtime());
+#ifdef CONFIG_USER_SCHED
+       init_rt_bandwidth(&root_task_group.rt_bandwidth,
+                       global_rt_period(), RUNTIME_INF);
+#endif
 #endif
 
 #ifdef CONFIG_GROUP_SCHED
        list_add(&init_task_group.list, &task_groups);
+       INIT_LIST_HEAD(&init_task_group.children);
+
+#ifdef CONFIG_USER_SCHED
+       INIT_LIST_HEAD(&root_task_group.children);
+       init_task_group.parent = &root_task_group;
+       list_add(&init_task_group.siblings, &root_task_group.children);
+#endif
 #endif
 
        for_each_possible_cpu(i) {
@@ -7553,8 +7797,6 @@ void __init sched_init(void)
                spin_lock_init(&rq->lock);
                lockdep_set_class(&rq->lock, &rq->rq_lock_key);
                rq->nr_running = 0;
-               rq->clock = 1;
-               update_last_tick_seen(rq);
                init_cfs_rq(&rq->cfs, rq);
                init_rt_rq(&rq->rt, rq);
 #ifdef CONFIG_FAIR_GROUP_SCHED
@@ -7582,6 +7824,8 @@ void __init sched_init(void)
                 */
                init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
 #elif defined CONFIG_USER_SCHED
+               root_task_group.shares = NICE_0_LOAD;
+               init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
                /*
                 * In case of task-groups formed thr' the user id of tasks,
                 * init_task_group represents tasks belonging to root user.
@@ -7595,7 +7839,8 @@ void __init sched_init(void)
                 */
                init_tg_cfs_entry(&init_task_group,
                                &per_cpu(init_cfs_rq, i),
-                               &per_cpu(init_sched_entity, i), i, 1, NULL);
+                               &per_cpu(init_sched_entity, i), i, 1,
+                               root_task_group.se[i]);
 
 #endif
 #endif /* CONFIG_FAIR_GROUP_SCHED */
@@ -7606,9 +7851,11 @@ void __init sched_init(void)
 #ifdef CONFIG_CGROUP_SCHED
                init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
 #elif defined CONFIG_USER_SCHED
+               init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
                init_tg_rt_entry(&init_task_group,
                                &per_cpu(init_rt_rq, i),
-                               &per_cpu(init_sched_rt_entity, i), i, 1, NULL);
+                               &per_cpu(init_sched_rt_entity, i), i, 1,
+                               root_task_group.rt_se[i]);
 #endif
 #endif
 
@@ -7693,6 +7940,7 @@ EXPORT_SYMBOL(__might_sleep);
 static void normalize_task(struct rq *rq, struct task_struct *p)
 {
        int on_rq;
+
        update_rq_clock(rq);
        on_rq = p->se.on_rq;
        if (on_rq)
@@ -7724,7 +7972,6 @@ void normalize_rt_tasks(void)
                p->se.sleep_start               = 0;
                p->se.block_start               = 0;
 #endif
-               task_rq(p)->clock               = 0;
 
                if (!rt_task(p)) {
                        /*
@@ -8000,6 +8247,12 @@ struct task_group *sched_create_group(struct task_group *parent)
                register_rt_sched_group(tg, i);
        }
        list_add_rcu(&tg->list, &task_groups);
+
+       WARN_ON(!parent); /* root should already exist */
+
+       tg->parent = parent;
+       list_add_rcu(&tg->siblings, &parent->children);
+       INIT_LIST_HEAD(&tg->children);
        spin_unlock_irqrestore(&task_group_lock, flags);
 
        return tg;
@@ -8028,6 +8281,7 @@ void sched_destroy_group(struct task_group *tg)
                unregister_rt_sched_group(tg, i);
        }
        list_del_rcu(&tg->list);
+       list_del_rcu(&tg->siblings);
        spin_unlock_irqrestore(&task_group_lock, flags);
 
        /* wait for possible concurrent references to cfs_rqs complete */
@@ -8087,7 +8341,7 @@ static void set_se_shares(struct sched_entity *se, unsigned long shares)
                dequeue_entity(cfs_rq, se, 0);
 
        se->load.weight = shares;
-       se->load.inv_weight = div64_64((1ULL<<32), shares);
+       se->load.inv_weight = 0;
 
        if (on_rq)
                enqueue_entity(cfs_rq, se, 0);
@@ -8108,13 +8362,10 @@ int sched_group_set_shares(struct task_group *tg, unsigned long shares)
        if (!tg->se[0])
                return -EINVAL;
 
-       /*
-        * A weight of 0 or 1 can cause arithmetics problems.
-        * (The default weight is 1024 - so there's no practical
-        *  limitation from this.)
-        */
-       if (shares < 2)
-               shares = 2;
+       if (shares < MIN_SHARES)
+               shares = MIN_SHARES;
+       else if (shares > MAX_SHARES)
+               shares = MAX_SHARES;
 
        mutex_lock(&shares_mutex);
        if (tg->shares == shares)
@@ -8123,6 +8374,7 @@ int sched_group_set_shares(struct task_group *tg, unsigned long shares)
        spin_lock_irqsave(&task_group_lock, flags);
        for_each_possible_cpu(i)
                unregister_fair_sched_group(tg, i);
+       list_del_rcu(&tg->siblings);
        spin_unlock_irqrestore(&task_group_lock, flags);
 
        /* wait for any ongoing reference to this group to finish */
@@ -8143,6 +8395,7 @@ int sched_group_set_shares(struct task_group *tg, unsigned long shares)
        spin_lock_irqsave(&task_group_lock, flags);
        for_each_possible_cpu(i)
                register_fair_sched_group(tg, i);
+       list_add_rcu(&tg->siblings, &tg->parent->children);
        spin_unlock_irqrestore(&task_group_lock, flags);
 done:
        mutex_unlock(&shares_mutex);
@@ -8166,9 +8419,41 @@ static unsigned long to_ratio(u64 period, u64 runtime)
        if (runtime == RUNTIME_INF)
                return 1ULL << 16;
 
-       return div64_64(runtime << 16, period);
+       return div64_u64(runtime << 16, period);
 }
 
+#ifdef CONFIG_CGROUP_SCHED
+static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
+{
+       struct task_group *tgi, *parent = tg ? tg->parent : NULL;
+       unsigned long total = 0;
+
+       if (!parent) {
+               if (global_rt_period() < period)
+                       return 0;
+
+               return to_ratio(period, runtime) <
+                       to_ratio(global_rt_period(), global_rt_runtime());
+       }
+
+       if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
+               return 0;
+
+       rcu_read_lock();
+       list_for_each_entry_rcu(tgi, &parent->children, siblings) {
+               if (tgi == tg)
+                       continue;
+
+               total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
+                               tgi->rt_bandwidth.rt_runtime);
+       }
+       rcu_read_unlock();
+
+       return total + to_ratio(period, runtime) <
+               to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
+                               parent->rt_bandwidth.rt_runtime);
+}
+#elif defined CONFIG_USER_SCHED
 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
 {
        struct task_group *tgi;
@@ -8188,6 +8473,7 @@ static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
 
        return total + to_ratio(period, runtime) < global_ratio;
 }
+#endif
 
 /* Must be called with tasklist_lock held */
 static inline int tg_has_rt_tasks(struct task_group *tg)
@@ -8403,13 +8689,13 @@ cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
 }
 
 #ifdef CONFIG_FAIR_GROUP_SCHED
-static int cpu_shares_write_uint(struct cgroup *cgrp, struct cftype *cftype,
+static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
                                u64 shareval)
 {
        return sched_group_set_shares(cgroup_tg(cgrp), shareval);
 }
 
-static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
+static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
 {
        struct task_group *tg = cgroup_tg(cgrp);
 
@@ -8418,49 +8704,15 @@ static u64 cpu_shares_read_uint(struct cgroup *cgrp, struct cftype *cft)
 #endif
 
 #ifdef CONFIG_RT_GROUP_SCHED
-static ssize_t cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
-                               struct file *file,
-                               const char __user *userbuf,
-                               size_t nbytes, loff_t *unused_ppos)
+static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
+                               s64 val)
 {
-       char buffer[64];
-       int retval = 0;
-       s64 val;
-       char *end;
-
-       if (!nbytes)
-               return -EINVAL;
-       if (nbytes >= sizeof(buffer))
-               return -E2BIG;
-       if (copy_from_user(buffer, userbuf, nbytes))
-               return -EFAULT;
-
-       buffer[nbytes] = 0;     /* nul-terminate */
-
-       /* strip newline if necessary */
-       if (nbytes && (buffer[nbytes-1] == '\n'))
-               buffer[nbytes-1] = 0;
-       val = simple_strtoll(buffer, &end, 0);
-       if (*end)
-               return -EINVAL;
-
-       /* Pass to subsystem */
-       retval = sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
-       if (!retval)
-               retval = nbytes;
-       return retval;
+       return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
 }
 
-static ssize_t cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft,
-                                  struct file *file,
-                                  char __user *buf, size_t nbytes,
-                                  loff_t *ppos)
+static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
 {
-       char tmp[64];
-       long val = sched_group_rt_runtime(cgroup_tg(cgrp));
-       int len = sprintf(tmp, "%ld\n", val);
-
-       return simple_read_from_buffer(buf, nbytes, ppos, tmp, len);
+       return sched_group_rt_runtime(cgroup_tg(cgrp));
 }
 
 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
@@ -8479,20 +8731,20 @@ static struct cftype cpu_files[] = {
 #ifdef CONFIG_FAIR_GROUP_SCHED
        {
                .name = "shares",
-               .read_uint = cpu_shares_read_uint,
-               .write_uint = cpu_shares_write_uint,
+               .read_u64 = cpu_shares_read_u64,
+               .write_u64 = cpu_shares_write_u64,
        },
 #endif
 #ifdef CONFIG_RT_GROUP_SCHED
        {
                .name = "rt_runtime_us",
-               .read = cpu_rt_runtime_read,
-               .write = cpu_rt_runtime_write,
+               .read_s64 = cpu_rt_runtime_read,
+               .write_s64 = cpu_rt_runtime_write,
        },
        {
                .name = "rt_period_us",
-               .read_uint = cpu_rt_period_read_uint,
-               .write_uint = cpu_rt_period_write_uint,
+               .read_u64 = cpu_rt_period_read_uint,
+               .write_u64 = cpu_rt_period_write_uint,
        },
 #endif
 };
@@ -8623,8 +8875,8 @@ out:
 static struct cftype files[] = {
        {
                .name = "usage",
-               .read_uint = cpuusage_read,
-               .write_uint = cpuusage_write,
+               .read_u64 = cpuusage_read,
+               .write_u64 = cpuusage_write,
        },
 };