sched: Enable might_sleep before initializing drivers.
[linux-2.6.git] / kernel / sched.c
index 3b3826e..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"
@@ -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))
 
 /*
@@ -1921,10 +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;
 
        /*
@@ -1947,12 +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;
@@ -1987,12 +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"
@@ -2188,7 +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);
@@ -3021,7 +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 */
@@ -3681,30 +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
@@ -3845,6 +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
@@ -3876,6 +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()) {
@@ -3929,6 +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))
@@ -4210,9 +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;
@@ -4253,16 +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;
        }
@@ -4300,6 +4335,26 @@ 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
@@ -4366,7 +4421,7 @@ asmlinkage void __sched notrace preempt_schedule(void)
 
        do {
                add_preempt_count_notrace(PREEMPT_ACTIVE);
-               schedule();
+               __schedule();
                sub_preempt_count_notrace(PREEMPT_ACTIVE);
 
                /*
@@ -4394,7 +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);
 
@@ -5519,7 +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);
 }
 
@@ -7374,6 +7429,7 @@ static void __sdt_free(const struct cpumask *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));
                }
@@ -7859,17 +7915,10 @@ 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;
@@ -7889,27 +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
@@ -7918,11 +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;
@@ -7945,12 +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;
 
@@ -8032,7 +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;
@@ -8103,7 +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
 
        /*
@@ -8154,12 +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)
 {
        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;
@@ -8335,6 +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]);
        }
 
@@ -8383,7 +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)
@@ -8424,6 +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]);
        }