4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/stop_machine.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
74 #include <linux/slab.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
80 #include "workqueue_sched.h"
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/sched.h>
86 * Convert user-nice values [ -20 ... 0 ... 19 ]
87 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
90 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
91 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
92 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
95 * 'User priority' is the nice value converted to something we
96 * can work with better when scaling various scheduler parameters,
97 * it's a [ 0 ... 39 ] range.
99 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
100 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
101 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
104 * Helpers for converting nanosecond timing to jiffy resolution
106 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
108 #define NICE_0_LOAD SCHED_LOAD_SCALE
109 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
112 * These are the 'tuning knobs' of the scheduler:
114 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
115 * Timeslices get refilled after they expire.
117 #define DEF_TIMESLICE (100 * HZ / 1000)
120 * single value that denotes runtime == period, ie unlimited time.
122 #define RUNTIME_INF ((u64)~0ULL)
124 static inline int rt_policy(int policy)
126 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
131 static inline int task_has_rt_policy(struct task_struct *p)
133 return rt_policy(p->policy);
137 * This is the priority-queue data structure of the RT scheduling class:
139 struct rt_prio_array {
140 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
141 struct list_head queue[MAX_RT_PRIO];
144 struct rt_bandwidth {
145 /* nests inside the rq lock: */
146 raw_spinlock_t rt_runtime_lock;
149 struct hrtimer rt_period_timer;
152 static struct rt_bandwidth def_rt_bandwidth;
154 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
156 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
158 struct rt_bandwidth *rt_b =
159 container_of(timer, struct rt_bandwidth, rt_period_timer);
165 now = hrtimer_cb_get_time(timer);
166 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
171 idle = do_sched_rt_period_timer(rt_b, overrun);
174 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
178 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
180 rt_b->rt_period = ns_to_ktime(period);
181 rt_b->rt_runtime = runtime;
183 raw_spin_lock_init(&rt_b->rt_runtime_lock);
185 hrtimer_init(&rt_b->rt_period_timer,
186 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
187 rt_b->rt_period_timer.function = sched_rt_period_timer;
190 static inline int rt_bandwidth_enabled(void)
192 return sysctl_sched_rt_runtime >= 0;
195 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
199 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
202 if (hrtimer_active(&rt_b->rt_period_timer))
205 raw_spin_lock(&rt_b->rt_runtime_lock);
210 if (hrtimer_active(&rt_b->rt_period_timer))
213 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
214 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
216 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
217 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
218 delta = ktime_to_ns(ktime_sub(hard, soft));
219 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
220 HRTIMER_MODE_ABS_PINNED, 0);
222 raw_spin_unlock(&rt_b->rt_runtime_lock);
225 #ifdef CONFIG_RT_GROUP_SCHED
226 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
228 hrtimer_cancel(&rt_b->rt_period_timer);
233 * sched_domains_mutex serializes calls to arch_init_sched_domains,
234 * detach_destroy_domains and partition_sched_domains.
236 static DEFINE_MUTEX(sched_domains_mutex);
238 #ifdef CONFIG_CGROUP_SCHED
240 #include <linux/cgroup.h>
244 static LIST_HEAD(task_groups);
246 /* task group related information */
248 struct cgroup_subsys_state css;
250 #ifdef CONFIG_FAIR_GROUP_SCHED
251 /* schedulable entities of this group on each cpu */
252 struct sched_entity **se;
253 /* runqueue "owned" by this group on each cpu */
254 struct cfs_rq **cfs_rq;
255 unsigned long shares;
258 #ifdef CONFIG_RT_GROUP_SCHED
259 struct sched_rt_entity **rt_se;
260 struct rt_rq **rt_rq;
262 struct rt_bandwidth rt_bandwidth;
266 struct list_head list;
268 struct task_group *parent;
269 struct list_head siblings;
270 struct list_head children;
273 #define root_task_group init_task_group
275 /* task_group_lock serializes add/remove of task groups and also changes to
276 * a task group's cpu shares.
278 static DEFINE_SPINLOCK(task_group_lock);
280 #ifdef CONFIG_FAIR_GROUP_SCHED
283 static int root_task_group_empty(void)
285 return list_empty(&root_task_group.children);
289 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
292 * A weight of 0 or 1 can cause arithmetics problems.
293 * A weight of a cfs_rq is the sum of weights of which entities
294 * are queued on this cfs_rq, so a weight of a entity should not be
295 * too large, so as the shares value of a task group.
296 * (The default weight is 1024 - so there's no practical
297 * limitation from this.)
300 #define MAX_SHARES (1UL << 18)
302 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
305 /* Default task group.
306 * Every task in system belong to this group at bootup.
308 struct task_group init_task_group;
310 #endif /* CONFIG_CGROUP_SCHED */
312 /* CFS-related fields in a runqueue */
314 struct load_weight load;
315 unsigned long nr_running;
320 struct rb_root tasks_timeline;
321 struct rb_node *rb_leftmost;
323 struct list_head tasks;
324 struct list_head *balance_iterator;
327 * 'curr' points to currently running entity on this cfs_rq.
328 * It is set to NULL otherwise (i.e when none are currently running).
330 struct sched_entity *curr, *next, *last;
332 unsigned int nr_spread_over;
334 #ifdef CONFIG_FAIR_GROUP_SCHED
335 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
338 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
339 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
340 * (like users, containers etc.)
342 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
343 * list is used during load balance.
345 struct list_head leaf_cfs_rq_list;
346 struct task_group *tg; /* group that "owns" this runqueue */
350 * the part of load.weight contributed by tasks
352 unsigned long task_weight;
355 * h_load = weight * f(tg)
357 * Where f(tg) is the recursive weight fraction assigned to
360 unsigned long h_load;
363 * this cpu's part of tg->shares
365 unsigned long shares;
368 * load.weight at the time we set shares
370 unsigned long rq_weight;
375 /* Real-Time classes' related field in a runqueue: */
377 struct rt_prio_array active;
378 unsigned long rt_nr_running;
379 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
381 int curr; /* highest queued rt task prio */
383 int next; /* next highest */
388 unsigned long rt_nr_migratory;
389 unsigned long rt_nr_total;
391 struct plist_head pushable_tasks;
396 /* Nests inside the rq lock: */
397 raw_spinlock_t rt_runtime_lock;
399 #ifdef CONFIG_RT_GROUP_SCHED
400 unsigned long rt_nr_boosted;
403 struct list_head leaf_rt_rq_list;
404 struct task_group *tg;
411 * We add the notion of a root-domain which will be used to define per-domain
412 * variables. Each exclusive cpuset essentially defines an island domain by
413 * fully partitioning the member cpus from any other cpuset. Whenever a new
414 * exclusive cpuset is created, we also create and attach a new root-domain
421 cpumask_var_t online;
424 * The "RT overload" flag: it gets set if a CPU has more than
425 * one runnable RT task.
427 cpumask_var_t rto_mask;
429 struct cpupri cpupri;
433 * By default the system creates a single root-domain with all cpus as
434 * members (mimicking the global state we have today).
436 static struct root_domain def_root_domain;
438 #endif /* CONFIG_SMP */
441 * This is the main, per-CPU runqueue data structure.
443 * Locking rule: those places that want to lock multiple runqueues
444 * (such as the load balancing or the thread migration code), lock
445 * acquire operations must be ordered by ascending &runqueue.
452 * nr_running and cpu_load should be in the same cacheline because
453 * remote CPUs use both these fields when doing load calculation.
455 unsigned long nr_running;
456 #define CPU_LOAD_IDX_MAX 5
457 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
458 unsigned long last_load_update_tick;
461 unsigned char nohz_balance_kick;
463 unsigned int skip_clock_update;
465 /* capture load from *all* tasks on this cpu: */
466 struct load_weight load;
467 unsigned long nr_load_updates;
473 #ifdef CONFIG_FAIR_GROUP_SCHED
474 /* list of leaf cfs_rq on this cpu: */
475 struct list_head leaf_cfs_rq_list;
477 #ifdef CONFIG_RT_GROUP_SCHED
478 struct list_head leaf_rt_rq_list;
482 * This is part of a global counter where only the total sum
483 * over all CPUs matters. A task can increase this counter on
484 * one CPU and if it got migrated afterwards it may decrease
485 * it on another CPU. Always updated under the runqueue lock:
487 unsigned long nr_uninterruptible;
489 struct task_struct *curr, *idle, *stop;
490 unsigned long next_balance;
491 struct mm_struct *prev_mm;
499 struct root_domain *rd;
500 struct sched_domain *sd;
502 unsigned long cpu_power;
504 unsigned char idle_at_tick;
505 /* For active balancing */
509 struct cpu_stop_work active_balance_work;
510 /* cpu of this runqueue: */
514 unsigned long avg_load_per_task;
522 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
526 /* calc_load related fields */
527 unsigned long calc_load_update;
528 long calc_load_active;
530 #ifdef CONFIG_SCHED_HRTICK
532 int hrtick_csd_pending;
533 struct call_single_data hrtick_csd;
535 struct hrtimer hrtick_timer;
538 #ifdef CONFIG_SCHEDSTATS
540 struct sched_info rq_sched_info;
541 unsigned long long rq_cpu_time;
542 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
544 /* sys_sched_yield() stats */
545 unsigned int yld_count;
547 /* schedule() stats */
548 unsigned int sched_switch;
549 unsigned int sched_count;
550 unsigned int sched_goidle;
552 /* try_to_wake_up() stats */
553 unsigned int ttwu_count;
554 unsigned int ttwu_local;
557 unsigned int bkl_count;
561 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
564 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
566 static inline int cpu_of(struct rq *rq)
575 #define rcu_dereference_check_sched_domain(p) \
576 rcu_dereference_check((p), \
577 rcu_read_lock_sched_held() || \
578 lockdep_is_held(&sched_domains_mutex))
581 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
582 * See detach_destroy_domains: synchronize_sched for details.
584 * The domain tree of any CPU may only be accessed from within
585 * preempt-disabled sections.
587 #define for_each_domain(cpu, __sd) \
588 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
590 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
591 #define this_rq() (&__get_cpu_var(runqueues))
592 #define task_rq(p) cpu_rq(task_cpu(p))
593 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
594 #define raw_rq() (&__raw_get_cpu_var(runqueues))
596 #ifdef CONFIG_CGROUP_SCHED
599 * Return the group to which this tasks belongs.
601 * We use task_subsys_state_check() and extend the RCU verification
602 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
603 * holds that lock for each task it moves into the cgroup. Therefore
604 * by holding that lock, we pin the task to the current cgroup.
606 static inline struct task_group *task_group(struct task_struct *p)
608 struct cgroup_subsys_state *css;
610 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
611 lockdep_is_held(&task_rq(p)->lock));
612 return container_of(css, struct task_group, css);
615 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
616 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
618 #ifdef CONFIG_FAIR_GROUP_SCHED
619 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
620 p->se.parent = task_group(p)->se[cpu];
623 #ifdef CONFIG_RT_GROUP_SCHED
624 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
625 p->rt.parent = task_group(p)->rt_se[cpu];
629 #else /* CONFIG_CGROUP_SCHED */
631 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
632 static inline struct task_group *task_group(struct task_struct *p)
637 #endif /* CONFIG_CGROUP_SCHED */
639 static u64 irq_time_cpu(int cpu);
640 static void sched_irq_time_avg_update(struct rq *rq, u64 irq_time);
642 inline void update_rq_clock(struct rq *rq)
644 int cpu = cpu_of(rq);
647 if (rq->skip_clock_update)
650 rq->clock = sched_clock_cpu(cpu);
651 irq_time = irq_time_cpu(cpu);
652 if (rq->clock - irq_time > rq->clock_task)
653 rq->clock_task = rq->clock - irq_time;
655 sched_irq_time_avg_update(rq, irq_time);
659 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
661 #ifdef CONFIG_SCHED_DEBUG
662 # define const_debug __read_mostly
664 # define const_debug static const
669 * @cpu: the processor in question.
671 * Returns true if the current cpu runqueue is locked.
672 * This interface allows printk to be called with the runqueue lock
673 * held and know whether or not it is OK to wake up the klogd.
675 int runqueue_is_locked(int cpu)
677 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
681 * Debugging: various feature bits
684 #define SCHED_FEAT(name, enabled) \
685 __SCHED_FEAT_##name ,
688 #include "sched_features.h"
693 #define SCHED_FEAT(name, enabled) \
694 (1UL << __SCHED_FEAT_##name) * enabled |
696 const_debug unsigned int sysctl_sched_features =
697 #include "sched_features.h"
702 #ifdef CONFIG_SCHED_DEBUG
703 #define SCHED_FEAT(name, enabled) \
706 static __read_mostly char *sched_feat_names[] = {
707 #include "sched_features.h"
713 static int sched_feat_show(struct seq_file *m, void *v)
717 for (i = 0; sched_feat_names[i]; i++) {
718 if (!(sysctl_sched_features & (1UL << i)))
720 seq_printf(m, "%s ", sched_feat_names[i]);
728 sched_feat_write(struct file *filp, const char __user *ubuf,
729 size_t cnt, loff_t *ppos)
739 if (copy_from_user(&buf, ubuf, cnt))
745 if (strncmp(buf, "NO_", 3) == 0) {
750 for (i = 0; sched_feat_names[i]; i++) {
751 if (strcmp(cmp, sched_feat_names[i]) == 0) {
753 sysctl_sched_features &= ~(1UL << i);
755 sysctl_sched_features |= (1UL << i);
760 if (!sched_feat_names[i])
768 static int sched_feat_open(struct inode *inode, struct file *filp)
770 return single_open(filp, sched_feat_show, NULL);
773 static const struct file_operations sched_feat_fops = {
774 .open = sched_feat_open,
775 .write = sched_feat_write,
778 .release = single_release,
781 static __init int sched_init_debug(void)
783 debugfs_create_file("sched_features", 0644, NULL, NULL,
788 late_initcall(sched_init_debug);
792 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
795 * Number of tasks to iterate in a single balance run.
796 * Limited because this is done with IRQs disabled.
798 const_debug unsigned int sysctl_sched_nr_migrate = 32;
801 * ratelimit for updating the group shares.
804 unsigned int sysctl_sched_shares_ratelimit = 250000;
805 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
808 * Inject some fuzzyness into changing the per-cpu group shares
809 * this avoids remote rq-locks at the expense of fairness.
812 unsigned int sysctl_sched_shares_thresh = 4;
815 * period over which we average the RT time consumption, measured
820 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
823 * period over which we measure -rt task cpu usage in us.
826 unsigned int sysctl_sched_rt_period = 1000000;
828 static __read_mostly int scheduler_running;
831 * part of the period that we allow rt tasks to run in us.
834 int sysctl_sched_rt_runtime = 950000;
836 static inline u64 global_rt_period(void)
838 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
841 static inline u64 global_rt_runtime(void)
843 if (sysctl_sched_rt_runtime < 0)
846 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
849 #ifndef prepare_arch_switch
850 # define prepare_arch_switch(next) do { } while (0)
852 #ifndef finish_arch_switch
853 # define finish_arch_switch(prev) do { } while (0)
856 static inline int task_current(struct rq *rq, struct task_struct *p)
858 return rq->curr == p;
861 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
862 static inline int task_running(struct rq *rq, struct task_struct *p)
864 return task_current(rq, p);
867 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
871 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
873 #ifdef CONFIG_DEBUG_SPINLOCK
874 /* this is a valid case when another task releases the spinlock */
875 rq->lock.owner = current;
878 * If we are tracking spinlock dependencies then we have to
879 * fix up the runqueue lock - which gets 'carried over' from
882 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
884 raw_spin_unlock_irq(&rq->lock);
887 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
888 static inline int task_running(struct rq *rq, struct task_struct *p)
893 return task_current(rq, p);
897 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
901 * We can optimise this out completely for !SMP, because the
902 * SMP rebalancing from interrupt is the only thing that cares
907 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
908 raw_spin_unlock_irq(&rq->lock);
910 raw_spin_unlock(&rq->lock);
914 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
918 * After ->oncpu is cleared, the task can be moved to a different CPU.
919 * We must ensure this doesn't happen until the switch is completely
925 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
929 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
932 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
935 static inline int task_is_waking(struct task_struct *p)
937 return unlikely(p->state == TASK_WAKING);
941 * __task_rq_lock - lock the runqueue a given task resides on.
942 * Must be called interrupts disabled.
944 static inline struct rq *__task_rq_lock(struct task_struct *p)
951 raw_spin_lock(&rq->lock);
952 if (likely(rq == task_rq(p)))
954 raw_spin_unlock(&rq->lock);
959 * task_rq_lock - lock the runqueue a given task resides on and disable
960 * interrupts. Note the ordering: we can safely lookup the task_rq without
961 * explicitly disabling preemption.
963 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
969 local_irq_save(*flags);
971 raw_spin_lock(&rq->lock);
972 if (likely(rq == task_rq(p)))
974 raw_spin_unlock_irqrestore(&rq->lock, *flags);
978 static void __task_rq_unlock(struct rq *rq)
981 raw_spin_unlock(&rq->lock);
984 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
987 raw_spin_unlock_irqrestore(&rq->lock, *flags);
991 * this_rq_lock - lock this runqueue and disable interrupts.
993 static struct rq *this_rq_lock(void)
1000 raw_spin_lock(&rq->lock);
1005 #ifdef CONFIG_SCHED_HRTICK
1007 * Use HR-timers to deliver accurate preemption points.
1009 * Its all a bit involved since we cannot program an hrt while holding the
1010 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1013 * When we get rescheduled we reprogram the hrtick_timer outside of the
1019 * - enabled by features
1020 * - hrtimer is actually high res
1022 static inline int hrtick_enabled(struct rq *rq)
1024 if (!sched_feat(HRTICK))
1026 if (!cpu_active(cpu_of(rq)))
1028 return hrtimer_is_hres_active(&rq->hrtick_timer);
1031 static void hrtick_clear(struct rq *rq)
1033 if (hrtimer_active(&rq->hrtick_timer))
1034 hrtimer_cancel(&rq->hrtick_timer);
1038 * High-resolution timer tick.
1039 * Runs from hardirq context with interrupts disabled.
1041 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1043 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1045 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1047 raw_spin_lock(&rq->lock);
1048 update_rq_clock(rq);
1049 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1050 raw_spin_unlock(&rq->lock);
1052 return HRTIMER_NORESTART;
1057 * called from hardirq (IPI) context
1059 static void __hrtick_start(void *arg)
1061 struct rq *rq = arg;
1063 raw_spin_lock(&rq->lock);
1064 hrtimer_restart(&rq->hrtick_timer);
1065 rq->hrtick_csd_pending = 0;
1066 raw_spin_unlock(&rq->lock);
1070 * Called to set the hrtick timer state.
1072 * called with rq->lock held and irqs disabled
1074 static void hrtick_start(struct rq *rq, u64 delay)
1076 struct hrtimer *timer = &rq->hrtick_timer;
1077 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1079 hrtimer_set_expires(timer, time);
1081 if (rq == this_rq()) {
1082 hrtimer_restart(timer);
1083 } else if (!rq->hrtick_csd_pending) {
1084 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1085 rq->hrtick_csd_pending = 1;
1090 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1092 int cpu = (int)(long)hcpu;
1095 case CPU_UP_CANCELED:
1096 case CPU_UP_CANCELED_FROZEN:
1097 case CPU_DOWN_PREPARE:
1098 case CPU_DOWN_PREPARE_FROZEN:
1100 case CPU_DEAD_FROZEN:
1101 hrtick_clear(cpu_rq(cpu));
1108 static __init void init_hrtick(void)
1110 hotcpu_notifier(hotplug_hrtick, 0);
1114 * Called to set the hrtick timer state.
1116 * called with rq->lock held and irqs disabled
1118 static void hrtick_start(struct rq *rq, u64 delay)
1120 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1121 HRTIMER_MODE_REL_PINNED, 0);
1124 static inline void init_hrtick(void)
1127 #endif /* CONFIG_SMP */
1129 static void init_rq_hrtick(struct rq *rq)
1132 rq->hrtick_csd_pending = 0;
1134 rq->hrtick_csd.flags = 0;
1135 rq->hrtick_csd.func = __hrtick_start;
1136 rq->hrtick_csd.info = rq;
1139 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1140 rq->hrtick_timer.function = hrtick;
1142 #else /* CONFIG_SCHED_HRTICK */
1143 static inline void hrtick_clear(struct rq *rq)
1147 static inline void init_rq_hrtick(struct rq *rq)
1151 static inline void init_hrtick(void)
1154 #endif /* CONFIG_SCHED_HRTICK */
1157 * resched_task - mark a task 'to be rescheduled now'.
1159 * On UP this means the setting of the need_resched flag, on SMP it
1160 * might also involve a cross-CPU call to trigger the scheduler on
1165 #ifndef tsk_is_polling
1166 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1169 static void resched_task(struct task_struct *p)
1173 assert_raw_spin_locked(&task_rq(p)->lock);
1175 if (test_tsk_need_resched(p))
1178 set_tsk_need_resched(p);
1181 if (cpu == smp_processor_id())
1184 /* NEED_RESCHED must be visible before we test polling */
1186 if (!tsk_is_polling(p))
1187 smp_send_reschedule(cpu);
1190 static void resched_cpu(int cpu)
1192 struct rq *rq = cpu_rq(cpu);
1193 unsigned long flags;
1195 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1197 resched_task(cpu_curr(cpu));
1198 raw_spin_unlock_irqrestore(&rq->lock, flags);
1203 * In the semi idle case, use the nearest busy cpu for migrating timers
1204 * from an idle cpu. This is good for power-savings.
1206 * We don't do similar optimization for completely idle system, as
1207 * selecting an idle cpu will add more delays to the timers than intended
1208 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1210 int get_nohz_timer_target(void)
1212 int cpu = smp_processor_id();
1214 struct sched_domain *sd;
1216 for_each_domain(cpu, sd) {
1217 for_each_cpu(i, sched_domain_span(sd))
1224 * When add_timer_on() enqueues a timer into the timer wheel of an
1225 * idle CPU then this timer might expire before the next timer event
1226 * which is scheduled to wake up that CPU. In case of a completely
1227 * idle system the next event might even be infinite time into the
1228 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1229 * leaves the inner idle loop so the newly added timer is taken into
1230 * account when the CPU goes back to idle and evaluates the timer
1231 * wheel for the next timer event.
1233 void wake_up_idle_cpu(int cpu)
1235 struct rq *rq = cpu_rq(cpu);
1237 if (cpu == smp_processor_id())
1241 * This is safe, as this function is called with the timer
1242 * wheel base lock of (cpu) held. When the CPU is on the way
1243 * to idle and has not yet set rq->curr to idle then it will
1244 * be serialized on the timer wheel base lock and take the new
1245 * timer into account automatically.
1247 if (rq->curr != rq->idle)
1251 * We can set TIF_RESCHED on the idle task of the other CPU
1252 * lockless. The worst case is that the other CPU runs the
1253 * idle task through an additional NOOP schedule()
1255 set_tsk_need_resched(rq->idle);
1257 /* NEED_RESCHED must be visible before we test polling */
1259 if (!tsk_is_polling(rq->idle))
1260 smp_send_reschedule(cpu);
1263 #endif /* CONFIG_NO_HZ */
1265 static u64 sched_avg_period(void)
1267 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1270 static void sched_avg_update(struct rq *rq)
1272 s64 period = sched_avg_period();
1274 while ((s64)(rq->clock - rq->age_stamp) > period) {
1276 * Inline assembly required to prevent the compiler
1277 * optimising this loop into a divmod call.
1278 * See __iter_div_u64_rem() for another example of this.
1280 asm("" : "+rm" (rq->age_stamp));
1281 rq->age_stamp += period;
1286 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1288 rq->rt_avg += rt_delta;
1289 sched_avg_update(rq);
1292 #else /* !CONFIG_SMP */
1293 static void resched_task(struct task_struct *p)
1295 assert_raw_spin_locked(&task_rq(p)->lock);
1296 set_tsk_need_resched(p);
1299 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1303 static void sched_avg_update(struct rq *rq)
1306 #endif /* CONFIG_SMP */
1308 #if BITS_PER_LONG == 32
1309 # define WMULT_CONST (~0UL)
1311 # define WMULT_CONST (1UL << 32)
1314 #define WMULT_SHIFT 32
1317 * Shift right and round:
1319 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1322 * delta *= weight / lw
1324 static unsigned long
1325 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1326 struct load_weight *lw)
1330 if (!lw->inv_weight) {
1331 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1334 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1338 tmp = (u64)delta_exec * weight;
1340 * Check whether we'd overflow the 64-bit multiplication:
1342 if (unlikely(tmp > WMULT_CONST))
1343 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1346 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1348 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1351 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1357 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1364 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1365 * of tasks with abnormal "nice" values across CPUs the contribution that
1366 * each task makes to its run queue's load is weighted according to its
1367 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1368 * scaled version of the new time slice allocation that they receive on time
1372 #define WEIGHT_IDLEPRIO 3
1373 #define WMULT_IDLEPRIO 1431655765
1376 * Nice levels are multiplicative, with a gentle 10% change for every
1377 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1378 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1379 * that remained on nice 0.
1381 * The "10% effect" is relative and cumulative: from _any_ nice level,
1382 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1383 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1384 * If a task goes up by ~10% and another task goes down by ~10% then
1385 * the relative distance between them is ~25%.)
1387 static const int prio_to_weight[40] = {
1388 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1389 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1390 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1391 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1392 /* 0 */ 1024, 820, 655, 526, 423,
1393 /* 5 */ 335, 272, 215, 172, 137,
1394 /* 10 */ 110, 87, 70, 56, 45,
1395 /* 15 */ 36, 29, 23, 18, 15,
1399 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1401 * In cases where the weight does not change often, we can use the
1402 * precalculated inverse to speed up arithmetics by turning divisions
1403 * into multiplications:
1405 static const u32 prio_to_wmult[40] = {
1406 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1407 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1408 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1409 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1410 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1411 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1412 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1413 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1416 /* Time spent by the tasks of the cpu accounting group executing in ... */
1417 enum cpuacct_stat_index {
1418 CPUACCT_STAT_USER, /* ... user mode */
1419 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1421 CPUACCT_STAT_NSTATS,
1424 #ifdef CONFIG_CGROUP_CPUACCT
1425 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1426 static void cpuacct_update_stats(struct task_struct *tsk,
1427 enum cpuacct_stat_index idx, cputime_t val);
1429 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1430 static inline void cpuacct_update_stats(struct task_struct *tsk,
1431 enum cpuacct_stat_index idx, cputime_t val) {}
1434 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1436 update_load_add(&rq->load, load);
1439 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1441 update_load_sub(&rq->load, load);
1444 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1445 typedef int (*tg_visitor)(struct task_group *, void *);
1448 * Iterate the full tree, calling @down when first entering a node and @up when
1449 * leaving it for the final time.
1451 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1453 struct task_group *parent, *child;
1457 parent = &root_task_group;
1459 ret = (*down)(parent, data);
1462 list_for_each_entry_rcu(child, &parent->children, siblings) {
1469 ret = (*up)(parent, data);
1474 parent = parent->parent;
1483 static int tg_nop(struct task_group *tg, void *data)
1490 /* Used instead of source_load when we know the type == 0 */
1491 static unsigned long weighted_cpuload(const int cpu)
1493 return cpu_rq(cpu)->load.weight;
1497 * Return a low guess at the load of a migration-source cpu weighted
1498 * according to the scheduling class and "nice" value.
1500 * We want to under-estimate the load of migration sources, to
1501 * balance conservatively.
1503 static unsigned long source_load(int cpu, int type)
1505 struct rq *rq = cpu_rq(cpu);
1506 unsigned long total = weighted_cpuload(cpu);
1508 if (type == 0 || !sched_feat(LB_BIAS))
1511 return min(rq->cpu_load[type-1], total);
1515 * Return a high guess at the load of a migration-target cpu weighted
1516 * according to the scheduling class and "nice" value.
1518 static unsigned long target_load(int cpu, int type)
1520 struct rq *rq = cpu_rq(cpu);
1521 unsigned long total = weighted_cpuload(cpu);
1523 if (type == 0 || !sched_feat(LB_BIAS))
1526 return max(rq->cpu_load[type-1], total);
1529 static unsigned long power_of(int cpu)
1531 return cpu_rq(cpu)->cpu_power;
1534 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1536 static unsigned long cpu_avg_load_per_task(int cpu)
1538 struct rq *rq = cpu_rq(cpu);
1539 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1542 rq->avg_load_per_task = rq->load.weight / nr_running;
1544 rq->avg_load_per_task = 0;
1546 return rq->avg_load_per_task;
1549 #ifdef CONFIG_FAIR_GROUP_SCHED
1551 static __read_mostly unsigned long __percpu *update_shares_data;
1553 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1556 * Calculate and set the cpu's group shares.
1558 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1559 unsigned long sd_shares,
1560 unsigned long sd_rq_weight,
1561 unsigned long *usd_rq_weight)
1563 unsigned long shares, rq_weight;
1566 rq_weight = usd_rq_weight[cpu];
1569 rq_weight = NICE_0_LOAD;
1573 * \Sum_j shares_j * rq_weight_i
1574 * shares_i = -----------------------------
1575 * \Sum_j rq_weight_j
1577 shares = (sd_shares * rq_weight) / sd_rq_weight;
1578 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1580 if (abs(shares - tg->se[cpu]->load.weight) >
1581 sysctl_sched_shares_thresh) {
1582 struct rq *rq = cpu_rq(cpu);
1583 unsigned long flags;
1585 raw_spin_lock_irqsave(&rq->lock, flags);
1586 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1587 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1588 __set_se_shares(tg->se[cpu], shares);
1589 raw_spin_unlock_irqrestore(&rq->lock, flags);
1594 * Re-compute the task group their per cpu shares over the given domain.
1595 * This needs to be done in a bottom-up fashion because the rq weight of a
1596 * parent group depends on the shares of its child groups.
1598 static int tg_shares_up(struct task_group *tg, void *data)
1600 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1601 unsigned long *usd_rq_weight;
1602 struct sched_domain *sd = data;
1603 unsigned long flags;
1609 local_irq_save(flags);
1610 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1612 for_each_cpu(i, sched_domain_span(sd)) {
1613 weight = tg->cfs_rq[i]->load.weight;
1614 usd_rq_weight[i] = weight;
1616 rq_weight += weight;
1618 * If there are currently no tasks on the cpu pretend there
1619 * is one of average load so that when a new task gets to
1620 * run here it will not get delayed by group starvation.
1623 weight = NICE_0_LOAD;
1625 sum_weight += weight;
1626 shares += tg->cfs_rq[i]->shares;
1630 rq_weight = sum_weight;
1632 if ((!shares && rq_weight) || shares > tg->shares)
1633 shares = tg->shares;
1635 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1636 shares = tg->shares;
1638 for_each_cpu(i, sched_domain_span(sd))
1639 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1641 local_irq_restore(flags);
1647 * Compute the cpu's hierarchical load factor for each task group.
1648 * This needs to be done in a top-down fashion because the load of a child
1649 * group is a fraction of its parents load.
1651 static int tg_load_down(struct task_group *tg, void *data)
1654 long cpu = (long)data;
1657 load = cpu_rq(cpu)->load.weight;
1659 load = tg->parent->cfs_rq[cpu]->h_load;
1660 load *= tg->cfs_rq[cpu]->shares;
1661 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1664 tg->cfs_rq[cpu]->h_load = load;
1669 static void update_shares(struct sched_domain *sd)
1674 if (root_task_group_empty())
1677 now = local_clock();
1678 elapsed = now - sd->last_update;
1680 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1681 sd->last_update = now;
1682 walk_tg_tree(tg_nop, tg_shares_up, sd);
1686 static void update_h_load(long cpu)
1688 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1693 static inline void update_shares(struct sched_domain *sd)
1699 #ifdef CONFIG_PREEMPT
1701 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1704 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1705 * way at the expense of forcing extra atomic operations in all
1706 * invocations. This assures that the double_lock is acquired using the
1707 * same underlying policy as the spinlock_t on this architecture, which
1708 * reduces latency compared to the unfair variant below. However, it
1709 * also adds more overhead and therefore may reduce throughput.
1711 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1712 __releases(this_rq->lock)
1713 __acquires(busiest->lock)
1714 __acquires(this_rq->lock)
1716 raw_spin_unlock(&this_rq->lock);
1717 double_rq_lock(this_rq, busiest);
1724 * Unfair double_lock_balance: Optimizes throughput at the expense of
1725 * latency by eliminating extra atomic operations when the locks are
1726 * already in proper order on entry. This favors lower cpu-ids and will
1727 * grant the double lock to lower cpus over higher ids under contention,
1728 * regardless of entry order into the function.
1730 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1731 __releases(this_rq->lock)
1732 __acquires(busiest->lock)
1733 __acquires(this_rq->lock)
1737 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1738 if (busiest < this_rq) {
1739 raw_spin_unlock(&this_rq->lock);
1740 raw_spin_lock(&busiest->lock);
1741 raw_spin_lock_nested(&this_rq->lock,
1742 SINGLE_DEPTH_NESTING);
1745 raw_spin_lock_nested(&busiest->lock,
1746 SINGLE_DEPTH_NESTING);
1751 #endif /* CONFIG_PREEMPT */
1754 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1756 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1758 if (unlikely(!irqs_disabled())) {
1759 /* printk() doesn't work good under rq->lock */
1760 raw_spin_unlock(&this_rq->lock);
1764 return _double_lock_balance(this_rq, busiest);
1767 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1768 __releases(busiest->lock)
1770 raw_spin_unlock(&busiest->lock);
1771 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1775 * double_rq_lock - safely lock two runqueues
1777 * Note this does not disable interrupts like task_rq_lock,
1778 * you need to do so manually before calling.
1780 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1781 __acquires(rq1->lock)
1782 __acquires(rq2->lock)
1784 BUG_ON(!irqs_disabled());
1786 raw_spin_lock(&rq1->lock);
1787 __acquire(rq2->lock); /* Fake it out ;) */
1790 raw_spin_lock(&rq1->lock);
1791 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1793 raw_spin_lock(&rq2->lock);
1794 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1800 * double_rq_unlock - safely unlock two runqueues
1802 * Note this does not restore interrupts like task_rq_unlock,
1803 * you need to do so manually after calling.
1805 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1806 __releases(rq1->lock)
1807 __releases(rq2->lock)
1809 raw_spin_unlock(&rq1->lock);
1811 raw_spin_unlock(&rq2->lock);
1813 __release(rq2->lock);
1818 #ifdef CONFIG_FAIR_GROUP_SCHED
1819 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1822 cfs_rq->shares = shares;
1827 static void calc_load_account_idle(struct rq *this_rq);
1828 static void update_sysctl(void);
1829 static int get_update_sysctl_factor(void);
1830 static void update_cpu_load(struct rq *this_rq);
1832 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1834 set_task_rq(p, cpu);
1837 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1838 * successfuly executed on another CPU. We must ensure that updates of
1839 * per-task data have been completed by this moment.
1842 task_thread_info(p)->cpu = cpu;
1846 static const struct sched_class rt_sched_class;
1848 #define sched_class_highest (&stop_sched_class)
1849 #define for_each_class(class) \
1850 for (class = sched_class_highest; class; class = class->next)
1852 #include "sched_stats.h"
1854 static void inc_nr_running(struct rq *rq)
1859 static void dec_nr_running(struct rq *rq)
1864 static void set_load_weight(struct task_struct *p)
1867 * SCHED_IDLE tasks get minimal weight:
1869 if (p->policy == SCHED_IDLE) {
1870 p->se.load.weight = WEIGHT_IDLEPRIO;
1871 p->se.load.inv_weight = WMULT_IDLEPRIO;
1875 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1876 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1879 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1881 update_rq_clock(rq);
1882 sched_info_queued(p);
1883 p->sched_class->enqueue_task(rq, p, flags);
1887 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1889 update_rq_clock(rq);
1890 sched_info_dequeued(p);
1891 p->sched_class->dequeue_task(rq, p, flags);
1896 * activate_task - move a task to the runqueue.
1898 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1900 if (task_contributes_to_load(p))
1901 rq->nr_uninterruptible--;
1903 enqueue_task(rq, p, flags);
1908 * deactivate_task - remove a task from the runqueue.
1910 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1912 if (task_contributes_to_load(p))
1913 rq->nr_uninterruptible++;
1915 dequeue_task(rq, p, flags);
1919 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1922 * There are no locks covering percpu hardirq/softirq time.
1923 * They are only modified in account_system_vtime, on corresponding CPU
1924 * with interrupts disabled. So, writes are safe.
1925 * They are read and saved off onto struct rq in update_rq_clock().
1926 * This may result in other CPU reading this CPU's irq time and can
1927 * race with irq/account_system_vtime on this CPU. We would either get old
1928 * or new value (or semi updated value on 32 bit) with a side effect of
1929 * accounting a slice of irq time to wrong task when irq is in progress
1930 * while we read rq->clock. That is a worthy compromise in place of having
1931 * locks on each irq in account_system_time.
1933 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1934 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1936 static DEFINE_PER_CPU(u64, irq_start_time);
1937 static int sched_clock_irqtime;
1939 void enable_sched_clock_irqtime(void)
1941 sched_clock_irqtime = 1;
1944 void disable_sched_clock_irqtime(void)
1946 sched_clock_irqtime = 0;
1949 static u64 irq_time_cpu(int cpu)
1951 if (!sched_clock_irqtime)
1954 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1957 void account_system_vtime(struct task_struct *curr)
1959 unsigned long flags;
1963 if (!sched_clock_irqtime)
1966 local_irq_save(flags);
1968 cpu = smp_processor_id();
1969 now = sched_clock_cpu(cpu);
1970 delta = now - per_cpu(irq_start_time, cpu);
1971 per_cpu(irq_start_time, cpu) = now;
1973 * We do not account for softirq time from ksoftirqd here.
1974 * We want to continue accounting softirq time to ksoftirqd thread
1975 * in that case, so as not to confuse scheduler with a special task
1976 * that do not consume any time, but still wants to run.
1978 if (hardirq_count())
1979 per_cpu(cpu_hardirq_time, cpu) += delta;
1980 else if (in_serving_softirq() && !(curr->flags & PF_KSOFTIRQD))
1981 per_cpu(cpu_softirq_time, cpu) += delta;
1983 local_irq_restore(flags);
1985 EXPORT_SYMBOL_GPL(account_system_vtime);
1987 static void sched_irq_time_avg_update(struct rq *rq, u64 curr_irq_time)
1989 if (sched_clock_irqtime && sched_feat(NONIRQ_POWER)) {
1990 u64 delta_irq = curr_irq_time - rq->prev_irq_time;
1991 rq->prev_irq_time = curr_irq_time;
1992 sched_rt_avg_update(rq, delta_irq);
1998 static u64 irq_time_cpu(int cpu)
2003 static void sched_irq_time_avg_update(struct rq *rq, u64 curr_irq_time) { }
2007 #include "sched_idletask.c"
2008 #include "sched_fair.c"
2009 #include "sched_rt.c"
2010 #include "sched_stoptask.c"
2011 #ifdef CONFIG_SCHED_DEBUG
2012 # include "sched_debug.c"
2015 void sched_set_stop_task(int cpu, struct task_struct *stop)
2017 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2018 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2022 * Make it appear like a SCHED_FIFO task, its something
2023 * userspace knows about and won't get confused about.
2025 * Also, it will make PI more or less work without too
2026 * much confusion -- but then, stop work should not
2027 * rely on PI working anyway.
2029 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2031 stop->sched_class = &stop_sched_class;
2034 cpu_rq(cpu)->stop = stop;
2038 * Reset it back to a normal scheduling class so that
2039 * it can die in pieces.
2041 old_stop->sched_class = &rt_sched_class;
2046 * __normal_prio - return the priority that is based on the static prio
2048 static inline int __normal_prio(struct task_struct *p)
2050 return p->static_prio;
2054 * Calculate the expected normal priority: i.e. priority
2055 * without taking RT-inheritance into account. Might be
2056 * boosted by interactivity modifiers. Changes upon fork,
2057 * setprio syscalls, and whenever the interactivity
2058 * estimator recalculates.
2060 static inline int normal_prio(struct task_struct *p)
2064 if (task_has_rt_policy(p))
2065 prio = MAX_RT_PRIO-1 - p->rt_priority;
2067 prio = __normal_prio(p);
2072 * Calculate the current priority, i.e. the priority
2073 * taken into account by the scheduler. This value might
2074 * be boosted by RT tasks, or might be boosted by
2075 * interactivity modifiers. Will be RT if the task got
2076 * RT-boosted. If not then it returns p->normal_prio.
2078 static int effective_prio(struct task_struct *p)
2080 p->normal_prio = normal_prio(p);
2082 * If we are RT tasks or we were boosted to RT priority,
2083 * keep the priority unchanged. Otherwise, update priority
2084 * to the normal priority:
2086 if (!rt_prio(p->prio))
2087 return p->normal_prio;
2092 * task_curr - is this task currently executing on a CPU?
2093 * @p: the task in question.
2095 inline int task_curr(const struct task_struct *p)
2097 return cpu_curr(task_cpu(p)) == p;
2100 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2101 const struct sched_class *prev_class,
2102 int oldprio, int running)
2104 if (prev_class != p->sched_class) {
2105 if (prev_class->switched_from)
2106 prev_class->switched_from(rq, p, running);
2107 p->sched_class->switched_to(rq, p, running);
2109 p->sched_class->prio_changed(rq, p, oldprio, running);
2112 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2114 const struct sched_class *class;
2116 if (p->sched_class == rq->curr->sched_class) {
2117 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2119 for_each_class(class) {
2120 if (class == rq->curr->sched_class)
2122 if (class == p->sched_class) {
2123 resched_task(rq->curr);
2130 * A queue event has occurred, and we're going to schedule. In
2131 * this case, we can save a useless back to back clock update.
2133 if (rq->curr->se.on_rq && test_tsk_need_resched(rq->curr))
2134 rq->skip_clock_update = 1;
2139 * Is this task likely cache-hot:
2142 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2146 if (p->sched_class != &fair_sched_class)
2149 if (unlikely(p->policy == SCHED_IDLE))
2153 * Buddy candidates are cache hot:
2155 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2156 (&p->se == cfs_rq_of(&p->se)->next ||
2157 &p->se == cfs_rq_of(&p->se)->last))
2160 if (sysctl_sched_migration_cost == -1)
2162 if (sysctl_sched_migration_cost == 0)
2165 delta = now - p->se.exec_start;
2167 return delta < (s64)sysctl_sched_migration_cost;
2170 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2172 #ifdef CONFIG_SCHED_DEBUG
2174 * We should never call set_task_cpu() on a blocked task,
2175 * ttwu() will sort out the placement.
2177 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2178 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2181 trace_sched_migrate_task(p, new_cpu);
2183 if (task_cpu(p) != new_cpu) {
2184 p->se.nr_migrations++;
2185 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2188 __set_task_cpu(p, new_cpu);
2191 struct migration_arg {
2192 struct task_struct *task;
2196 static int migration_cpu_stop(void *data);
2199 * The task's runqueue lock must be held.
2200 * Returns true if you have to wait for migration thread.
2202 static bool migrate_task(struct task_struct *p, int dest_cpu)
2204 struct rq *rq = task_rq(p);
2207 * If the task is not on a runqueue (and not running), then
2208 * the next wake-up will properly place the task.
2210 return p->se.on_rq || task_running(rq, p);
2214 * wait_task_inactive - wait for a thread to unschedule.
2216 * If @match_state is nonzero, it's the @p->state value just checked and
2217 * not expected to change. If it changes, i.e. @p might have woken up,
2218 * then return zero. When we succeed in waiting for @p to be off its CPU,
2219 * we return a positive number (its total switch count). If a second call
2220 * a short while later returns the same number, the caller can be sure that
2221 * @p has remained unscheduled the whole time.
2223 * The caller must ensure that the task *will* unschedule sometime soon,
2224 * else this function might spin for a *long* time. This function can't
2225 * be called with interrupts off, or it may introduce deadlock with
2226 * smp_call_function() if an IPI is sent by the same process we are
2227 * waiting to become inactive.
2229 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2231 unsigned long flags;
2238 * We do the initial early heuristics without holding
2239 * any task-queue locks at all. We'll only try to get
2240 * the runqueue lock when things look like they will
2246 * If the task is actively running on another CPU
2247 * still, just relax and busy-wait without holding
2250 * NOTE! Since we don't hold any locks, it's not
2251 * even sure that "rq" stays as the right runqueue!
2252 * But we don't care, since "task_running()" will
2253 * return false if the runqueue has changed and p
2254 * is actually now running somewhere else!
2256 while (task_running(rq, p)) {
2257 if (match_state && unlikely(p->state != match_state))
2263 * Ok, time to look more closely! We need the rq
2264 * lock now, to be *sure*. If we're wrong, we'll
2265 * just go back and repeat.
2267 rq = task_rq_lock(p, &flags);
2268 trace_sched_wait_task(p);
2269 running = task_running(rq, p);
2270 on_rq = p->se.on_rq;
2272 if (!match_state || p->state == match_state)
2273 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2274 task_rq_unlock(rq, &flags);
2277 * If it changed from the expected state, bail out now.
2279 if (unlikely(!ncsw))
2283 * Was it really running after all now that we
2284 * checked with the proper locks actually held?
2286 * Oops. Go back and try again..
2288 if (unlikely(running)) {
2294 * It's not enough that it's not actively running,
2295 * it must be off the runqueue _entirely_, and not
2298 * So if it was still runnable (but just not actively
2299 * running right now), it's preempted, and we should
2300 * yield - it could be a while.
2302 if (unlikely(on_rq)) {
2303 schedule_timeout_uninterruptible(1);
2308 * Ahh, all good. It wasn't running, and it wasn't
2309 * runnable, which means that it will never become
2310 * running in the future either. We're all done!
2319 * kick_process - kick a running thread to enter/exit the kernel
2320 * @p: the to-be-kicked thread
2322 * Cause a process which is running on another CPU to enter
2323 * kernel-mode, without any delay. (to get signals handled.)
2325 * NOTE: this function doesnt have to take the runqueue lock,
2326 * because all it wants to ensure is that the remote task enters
2327 * the kernel. If the IPI races and the task has been migrated
2328 * to another CPU then no harm is done and the purpose has been
2331 void kick_process(struct task_struct *p)
2337 if ((cpu != smp_processor_id()) && task_curr(p))
2338 smp_send_reschedule(cpu);
2341 EXPORT_SYMBOL_GPL(kick_process);
2342 #endif /* CONFIG_SMP */
2345 * task_oncpu_function_call - call a function on the cpu on which a task runs
2346 * @p: the task to evaluate
2347 * @func: the function to be called
2348 * @info: the function call argument
2350 * Calls the function @func when the task is currently running. This might
2351 * be on the current CPU, which just calls the function directly
2353 void task_oncpu_function_call(struct task_struct *p,
2354 void (*func) (void *info), void *info)
2361 smp_call_function_single(cpu, func, info, 1);
2367 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2369 static int select_fallback_rq(int cpu, struct task_struct *p)
2372 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2374 /* Look for allowed, online CPU in same node. */
2375 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2376 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2379 /* Any allowed, online CPU? */
2380 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2381 if (dest_cpu < nr_cpu_ids)
2384 /* No more Mr. Nice Guy. */
2385 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2386 dest_cpu = cpuset_cpus_allowed_fallback(p);
2388 * Don't tell them about moving exiting tasks or
2389 * kernel threads (both mm NULL), since they never
2392 if (p->mm && printk_ratelimit()) {
2393 printk(KERN_INFO "process %d (%s) no "
2394 "longer affine to cpu%d\n",
2395 task_pid_nr(p), p->comm, cpu);
2403 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2406 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2408 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2411 * In order not to call set_task_cpu() on a blocking task we need
2412 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2415 * Since this is common to all placement strategies, this lives here.
2417 * [ this allows ->select_task() to simply return task_cpu(p) and
2418 * not worry about this generic constraint ]
2420 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2422 cpu = select_fallback_rq(task_cpu(p), p);
2427 static void update_avg(u64 *avg, u64 sample)
2429 s64 diff = sample - *avg;
2434 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2435 bool is_sync, bool is_migrate, bool is_local,
2436 unsigned long en_flags)
2438 schedstat_inc(p, se.statistics.nr_wakeups);
2440 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2442 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2444 schedstat_inc(p, se.statistics.nr_wakeups_local);
2446 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2448 activate_task(rq, p, en_flags);
2451 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2452 int wake_flags, bool success)
2454 trace_sched_wakeup(p, success);
2455 check_preempt_curr(rq, p, wake_flags);
2457 p->state = TASK_RUNNING;
2459 if (p->sched_class->task_woken)
2460 p->sched_class->task_woken(rq, p);
2462 if (unlikely(rq->idle_stamp)) {
2463 u64 delta = rq->clock - rq->idle_stamp;
2464 u64 max = 2*sysctl_sched_migration_cost;
2469 update_avg(&rq->avg_idle, delta);
2473 /* if a worker is waking up, notify workqueue */
2474 if ((p->flags & PF_WQ_WORKER) && success)
2475 wq_worker_waking_up(p, cpu_of(rq));
2479 * try_to_wake_up - wake up a thread
2480 * @p: the thread to be awakened
2481 * @state: the mask of task states that can be woken
2482 * @wake_flags: wake modifier flags (WF_*)
2484 * Put it on the run-queue if it's not already there. The "current"
2485 * thread is always on the run-queue (except when the actual
2486 * re-schedule is in progress), and as such you're allowed to do
2487 * the simpler "current->state = TASK_RUNNING" to mark yourself
2488 * runnable without the overhead of this.
2490 * Returns %true if @p was woken up, %false if it was already running
2491 * or @state didn't match @p's state.
2493 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2496 int cpu, orig_cpu, this_cpu, success = 0;
2497 unsigned long flags;
2498 unsigned long en_flags = ENQUEUE_WAKEUP;
2501 this_cpu = get_cpu();
2504 rq = task_rq_lock(p, &flags);
2505 if (!(p->state & state))
2515 if (unlikely(task_running(rq, p)))
2519 * In order to handle concurrent wakeups and release the rq->lock
2520 * we put the task in TASK_WAKING state.
2522 * First fix up the nr_uninterruptible count:
2524 if (task_contributes_to_load(p)) {
2525 if (likely(cpu_online(orig_cpu)))
2526 rq->nr_uninterruptible--;
2528 this_rq()->nr_uninterruptible--;
2530 p->state = TASK_WAKING;
2532 if (p->sched_class->task_waking) {
2533 p->sched_class->task_waking(rq, p);
2534 en_flags |= ENQUEUE_WAKING;
2537 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2538 if (cpu != orig_cpu)
2539 set_task_cpu(p, cpu);
2540 __task_rq_unlock(rq);
2543 raw_spin_lock(&rq->lock);
2546 * We migrated the task without holding either rq->lock, however
2547 * since the task is not on the task list itself, nobody else
2548 * will try and migrate the task, hence the rq should match the
2549 * cpu we just moved it to.
2551 WARN_ON(task_cpu(p) != cpu);
2552 WARN_ON(p->state != TASK_WAKING);
2554 #ifdef CONFIG_SCHEDSTATS
2555 schedstat_inc(rq, ttwu_count);
2556 if (cpu == this_cpu)
2557 schedstat_inc(rq, ttwu_local);
2559 struct sched_domain *sd;
2560 for_each_domain(this_cpu, sd) {
2561 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2562 schedstat_inc(sd, ttwu_wake_remote);
2567 #endif /* CONFIG_SCHEDSTATS */
2570 #endif /* CONFIG_SMP */
2571 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2572 cpu == this_cpu, en_flags);
2575 ttwu_post_activation(p, rq, wake_flags, success);
2577 task_rq_unlock(rq, &flags);
2584 * try_to_wake_up_local - try to wake up a local task with rq lock held
2585 * @p: the thread to be awakened
2587 * Put @p on the run-queue if it's not alredy there. The caller must
2588 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2589 * the current task. this_rq() stays locked over invocation.
2591 static void try_to_wake_up_local(struct task_struct *p)
2593 struct rq *rq = task_rq(p);
2594 bool success = false;
2596 BUG_ON(rq != this_rq());
2597 BUG_ON(p == current);
2598 lockdep_assert_held(&rq->lock);
2600 if (!(p->state & TASK_NORMAL))
2604 if (likely(!task_running(rq, p))) {
2605 schedstat_inc(rq, ttwu_count);
2606 schedstat_inc(rq, ttwu_local);
2608 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2611 ttwu_post_activation(p, rq, 0, success);
2615 * wake_up_process - Wake up a specific process
2616 * @p: The process to be woken up.
2618 * Attempt to wake up the nominated process and move it to the set of runnable
2619 * processes. Returns 1 if the process was woken up, 0 if it was already
2622 * It may be assumed that this function implies a write memory barrier before
2623 * changing the task state if and only if any tasks are woken up.
2625 int wake_up_process(struct task_struct *p)
2627 return try_to_wake_up(p, TASK_ALL, 0);
2629 EXPORT_SYMBOL(wake_up_process);
2631 int wake_up_state(struct task_struct *p, unsigned int state)
2633 return try_to_wake_up(p, state, 0);
2637 * Perform scheduler related setup for a newly forked process p.
2638 * p is forked by current.
2640 * __sched_fork() is basic setup used by init_idle() too:
2642 static void __sched_fork(struct task_struct *p)
2644 p->se.exec_start = 0;
2645 p->se.sum_exec_runtime = 0;
2646 p->se.prev_sum_exec_runtime = 0;
2647 p->se.nr_migrations = 0;
2649 #ifdef CONFIG_SCHEDSTATS
2650 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2653 INIT_LIST_HEAD(&p->rt.run_list);
2655 INIT_LIST_HEAD(&p->se.group_node);
2657 #ifdef CONFIG_PREEMPT_NOTIFIERS
2658 INIT_HLIST_HEAD(&p->preempt_notifiers);
2663 * fork()/clone()-time setup:
2665 void sched_fork(struct task_struct *p, int clone_flags)
2667 int cpu = get_cpu();
2671 * We mark the process as running here. This guarantees that
2672 * nobody will actually run it, and a signal or other external
2673 * event cannot wake it up and insert it on the runqueue either.
2675 p->state = TASK_RUNNING;
2678 * Revert to default priority/policy on fork if requested.
2680 if (unlikely(p->sched_reset_on_fork)) {
2681 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2682 p->policy = SCHED_NORMAL;
2683 p->normal_prio = p->static_prio;
2686 if (PRIO_TO_NICE(p->static_prio) < 0) {
2687 p->static_prio = NICE_TO_PRIO(0);
2688 p->normal_prio = p->static_prio;
2693 * We don't need the reset flag anymore after the fork. It has
2694 * fulfilled its duty:
2696 p->sched_reset_on_fork = 0;
2700 * Make sure we do not leak PI boosting priority to the child.
2702 p->prio = current->normal_prio;
2704 if (!rt_prio(p->prio))
2705 p->sched_class = &fair_sched_class;
2707 if (p->sched_class->task_fork)
2708 p->sched_class->task_fork(p);
2711 * The child is not yet in the pid-hash so no cgroup attach races,
2712 * and the cgroup is pinned to this child due to cgroup_fork()
2713 * is ran before sched_fork().
2715 * Silence PROVE_RCU.
2718 set_task_cpu(p, cpu);
2721 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2722 if (likely(sched_info_on()))
2723 memset(&p->sched_info, 0, sizeof(p->sched_info));
2725 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2728 #ifdef CONFIG_PREEMPT
2729 /* Want to start with kernel preemption disabled. */
2730 task_thread_info(p)->preempt_count = 1;
2732 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2738 * wake_up_new_task - wake up a newly created task for the first time.
2740 * This function will do some initial scheduler statistics housekeeping
2741 * that must be done for every newly created context, then puts the task
2742 * on the runqueue and wakes it.
2744 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2746 unsigned long flags;
2748 int cpu __maybe_unused = get_cpu();
2751 rq = task_rq_lock(p, &flags);
2752 p->state = TASK_WAKING;
2755 * Fork balancing, do it here and not earlier because:
2756 * - cpus_allowed can change in the fork path
2757 * - any previously selected cpu might disappear through hotplug
2759 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2760 * without people poking at ->cpus_allowed.
2762 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2763 set_task_cpu(p, cpu);
2765 p->state = TASK_RUNNING;
2766 task_rq_unlock(rq, &flags);
2769 rq = task_rq_lock(p, &flags);
2770 activate_task(rq, p, 0);
2771 trace_sched_wakeup_new(p, 1);
2772 check_preempt_curr(rq, p, WF_FORK);
2774 if (p->sched_class->task_woken)
2775 p->sched_class->task_woken(rq, p);
2777 task_rq_unlock(rq, &flags);
2781 #ifdef CONFIG_PREEMPT_NOTIFIERS
2784 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2785 * @notifier: notifier struct to register
2787 void preempt_notifier_register(struct preempt_notifier *notifier)
2789 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2791 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2794 * preempt_notifier_unregister - no longer interested in preemption notifications
2795 * @notifier: notifier struct to unregister
2797 * This is safe to call from within a preemption notifier.
2799 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2801 hlist_del(¬ifier->link);
2803 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2805 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2807 struct preempt_notifier *notifier;
2808 struct hlist_node *node;
2810 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2811 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2815 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2816 struct task_struct *next)
2818 struct preempt_notifier *notifier;
2819 struct hlist_node *node;
2821 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2822 notifier->ops->sched_out(notifier, next);
2825 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2827 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2832 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2833 struct task_struct *next)
2837 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2840 * prepare_task_switch - prepare to switch tasks
2841 * @rq: the runqueue preparing to switch
2842 * @prev: the current task that is being switched out
2843 * @next: the task we are going to switch to.
2845 * This is called with the rq lock held and interrupts off. It must
2846 * be paired with a subsequent finish_task_switch after the context
2849 * prepare_task_switch sets up locking and calls architecture specific
2853 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2854 struct task_struct *next)
2856 fire_sched_out_preempt_notifiers(prev, next);
2857 prepare_lock_switch(rq, next);
2858 prepare_arch_switch(next);
2862 * finish_task_switch - clean up after a task-switch
2863 * @rq: runqueue associated with task-switch
2864 * @prev: the thread we just switched away from.
2866 * finish_task_switch must be called after the context switch, paired
2867 * with a prepare_task_switch call before the context switch.
2868 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2869 * and do any other architecture-specific cleanup actions.
2871 * Note that we may have delayed dropping an mm in context_switch(). If
2872 * so, we finish that here outside of the runqueue lock. (Doing it
2873 * with the lock held can cause deadlocks; see schedule() for
2876 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2877 __releases(rq->lock)
2879 struct mm_struct *mm = rq->prev_mm;
2885 * A task struct has one reference for the use as "current".
2886 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2887 * schedule one last time. The schedule call will never return, and
2888 * the scheduled task must drop that reference.
2889 * The test for TASK_DEAD must occur while the runqueue locks are
2890 * still held, otherwise prev could be scheduled on another cpu, die
2891 * there before we look at prev->state, and then the reference would
2893 * Manfred Spraul <manfred@colorfullife.com>
2895 prev_state = prev->state;
2896 finish_arch_switch(prev);
2897 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2898 local_irq_disable();
2899 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2900 perf_event_task_sched_in(current);
2901 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2903 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2904 finish_lock_switch(rq, prev);
2906 fire_sched_in_preempt_notifiers(current);
2909 if (unlikely(prev_state == TASK_DEAD)) {
2911 * Remove function-return probe instances associated with this
2912 * task and put them back on the free list.
2914 kprobe_flush_task(prev);
2915 put_task_struct(prev);
2921 /* assumes rq->lock is held */
2922 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2924 if (prev->sched_class->pre_schedule)
2925 prev->sched_class->pre_schedule(rq, prev);
2928 /* rq->lock is NOT held, but preemption is disabled */
2929 static inline void post_schedule(struct rq *rq)
2931 if (rq->post_schedule) {
2932 unsigned long flags;
2934 raw_spin_lock_irqsave(&rq->lock, flags);
2935 if (rq->curr->sched_class->post_schedule)
2936 rq->curr->sched_class->post_schedule(rq);
2937 raw_spin_unlock_irqrestore(&rq->lock, flags);
2939 rq->post_schedule = 0;
2945 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2949 static inline void post_schedule(struct rq *rq)
2956 * schedule_tail - first thing a freshly forked thread must call.
2957 * @prev: the thread we just switched away from.
2959 asmlinkage void schedule_tail(struct task_struct *prev)
2960 __releases(rq->lock)
2962 struct rq *rq = this_rq();
2964 finish_task_switch(rq, prev);
2967 * FIXME: do we need to worry about rq being invalidated by the
2972 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2973 /* In this case, finish_task_switch does not reenable preemption */
2976 if (current->set_child_tid)
2977 put_user(task_pid_vnr(current), current->set_child_tid);
2981 * context_switch - switch to the new MM and the new
2982 * thread's register state.
2985 context_switch(struct rq *rq, struct task_struct *prev,
2986 struct task_struct *next)
2988 struct mm_struct *mm, *oldmm;
2990 prepare_task_switch(rq, prev, next);
2991 trace_sched_switch(prev, next);
2993 oldmm = prev->active_mm;
2995 * For paravirt, this is coupled with an exit in switch_to to
2996 * combine the page table reload and the switch backend into
2999 arch_start_context_switch(prev);
3002 next->active_mm = oldmm;
3003 atomic_inc(&oldmm->mm_count);
3004 enter_lazy_tlb(oldmm, next);
3006 switch_mm(oldmm, mm, next);
3009 prev->active_mm = NULL;
3010 rq->prev_mm = oldmm;
3013 * Since the runqueue lock will be released by the next
3014 * task (which is an invalid locking op but in the case
3015 * of the scheduler it's an obvious special-case), so we
3016 * do an early lockdep release here:
3018 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3019 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3022 /* Here we just switch the register state and the stack. */
3023 switch_to(prev, next, prev);
3027 * this_rq must be evaluated again because prev may have moved
3028 * CPUs since it called schedule(), thus the 'rq' on its stack
3029 * frame will be invalid.
3031 finish_task_switch(this_rq(), prev);
3035 * nr_running, nr_uninterruptible and nr_context_switches:
3037 * externally visible scheduler statistics: current number of runnable
3038 * threads, current number of uninterruptible-sleeping threads, total
3039 * number of context switches performed since bootup.
3041 unsigned long nr_running(void)
3043 unsigned long i, sum = 0;
3045 for_each_online_cpu(i)
3046 sum += cpu_rq(i)->nr_running;
3051 unsigned long nr_uninterruptible(void)
3053 unsigned long i, sum = 0;
3055 for_each_possible_cpu(i)
3056 sum += cpu_rq(i)->nr_uninterruptible;
3059 * Since we read the counters lockless, it might be slightly
3060 * inaccurate. Do not allow it to go below zero though:
3062 if (unlikely((long)sum < 0))
3068 unsigned long long nr_context_switches(void)
3071 unsigned long long sum = 0;
3073 for_each_possible_cpu(i)
3074 sum += cpu_rq(i)->nr_switches;
3079 unsigned long nr_iowait(void)
3081 unsigned long i, sum = 0;
3083 for_each_possible_cpu(i)
3084 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3089 unsigned long nr_iowait_cpu(int cpu)
3091 struct rq *this = cpu_rq(cpu);
3092 return atomic_read(&this->nr_iowait);
3095 unsigned long this_cpu_load(void)
3097 struct rq *this = this_rq();
3098 return this->cpu_load[0];
3102 /* Variables and functions for calc_load */
3103 static atomic_long_t calc_load_tasks;
3104 static unsigned long calc_load_update;
3105 unsigned long avenrun[3];
3106 EXPORT_SYMBOL(avenrun);
3108 static long calc_load_fold_active(struct rq *this_rq)
3110 long nr_active, delta = 0;
3112 nr_active = this_rq->nr_running;
3113 nr_active += (long) this_rq->nr_uninterruptible;
3115 if (nr_active != this_rq->calc_load_active) {
3116 delta = nr_active - this_rq->calc_load_active;
3117 this_rq->calc_load_active = nr_active;
3123 static unsigned long
3124 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3127 load += active * (FIXED_1 - exp);
3128 load += 1UL << (FSHIFT - 1);
3129 return load >> FSHIFT;
3134 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3136 * When making the ILB scale, we should try to pull this in as well.
3138 static atomic_long_t calc_load_tasks_idle;
3140 static void calc_load_account_idle(struct rq *this_rq)
3144 delta = calc_load_fold_active(this_rq);
3146 atomic_long_add(delta, &calc_load_tasks_idle);
3149 static long calc_load_fold_idle(void)
3154 * Its got a race, we don't care...
3156 if (atomic_long_read(&calc_load_tasks_idle))
3157 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3163 * fixed_power_int - compute: x^n, in O(log n) time
3165 * @x: base of the power
3166 * @frac_bits: fractional bits of @x
3167 * @n: power to raise @x to.
3169 * By exploiting the relation between the definition of the natural power
3170 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3171 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3172 * (where: n_i \elem {0, 1}, the binary vector representing n),
3173 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3174 * of course trivially computable in O(log_2 n), the length of our binary
3177 static unsigned long
3178 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3180 unsigned long result = 1UL << frac_bits;
3185 result += 1UL << (frac_bits - 1);
3186 result >>= frac_bits;
3192 x += 1UL << (frac_bits - 1);
3200 * a1 = a0 * e + a * (1 - e)
3202 * a2 = a1 * e + a * (1 - e)
3203 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3204 * = a0 * e^2 + a * (1 - e) * (1 + e)
3206 * a3 = a2 * e + a * (1 - e)
3207 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3208 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3212 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3213 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3214 * = a0 * e^n + a * (1 - e^n)
3216 * [1] application of the geometric series:
3219 * S_n := \Sum x^i = -------------
3222 static unsigned long
3223 calc_load_n(unsigned long load, unsigned long exp,
3224 unsigned long active, unsigned int n)
3227 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3231 * NO_HZ can leave us missing all per-cpu ticks calling
3232 * calc_load_account_active(), but since an idle CPU folds its delta into
3233 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3234 * in the pending idle delta if our idle period crossed a load cycle boundary.
3236 * Once we've updated the global active value, we need to apply the exponential
3237 * weights adjusted to the number of cycles missed.
3239 static void calc_global_nohz(unsigned long ticks)
3241 long delta, active, n;
3243 if (time_before(jiffies, calc_load_update))
3247 * If we crossed a calc_load_update boundary, make sure to fold
3248 * any pending idle changes, the respective CPUs might have
3249 * missed the tick driven calc_load_account_active() update
3252 delta = calc_load_fold_idle();
3254 atomic_long_add(delta, &calc_load_tasks);
3257 * If we were idle for multiple load cycles, apply them.
3259 if (ticks >= LOAD_FREQ) {
3260 n = ticks / LOAD_FREQ;
3262 active = atomic_long_read(&calc_load_tasks);
3263 active = active > 0 ? active * FIXED_1 : 0;
3265 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3266 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3267 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3269 calc_load_update += n * LOAD_FREQ;
3273 * Its possible the remainder of the above division also crosses
3274 * a LOAD_FREQ period, the regular check in calc_global_load()
3275 * which comes after this will take care of that.
3277 * Consider us being 11 ticks before a cycle completion, and us
3278 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3279 * age us 4 cycles, and the test in calc_global_load() will
3280 * pick up the final one.
3284 static void calc_load_account_idle(struct rq *this_rq)
3288 static inline long calc_load_fold_idle(void)
3293 static void calc_global_nohz(unsigned long ticks)
3299 * get_avenrun - get the load average array
3300 * @loads: pointer to dest load array
3301 * @offset: offset to add
3302 * @shift: shift count to shift the result left
3304 * These values are estimates at best, so no need for locking.
3306 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3308 loads[0] = (avenrun[0] + offset) << shift;
3309 loads[1] = (avenrun[1] + offset) << shift;
3310 loads[2] = (avenrun[2] + offset) << shift;
3314 * calc_load - update the avenrun load estimates 10 ticks after the
3315 * CPUs have updated calc_load_tasks.
3317 void calc_global_load(unsigned long ticks)
3321 calc_global_nohz(ticks);
3323 if (time_before(jiffies, calc_load_update + 10))
3326 active = atomic_long_read(&calc_load_tasks);
3327 active = active > 0 ? active * FIXED_1 : 0;
3329 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3330 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3331 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3333 calc_load_update += LOAD_FREQ;
3337 * Called from update_cpu_load() to periodically update this CPU's
3340 static void calc_load_account_active(struct rq *this_rq)
3344 if (time_before(jiffies, this_rq->calc_load_update))
3347 delta = calc_load_fold_active(this_rq);
3348 delta += calc_load_fold_idle();
3350 atomic_long_add(delta, &calc_load_tasks);
3352 this_rq->calc_load_update += LOAD_FREQ;
3356 * The exact cpuload at various idx values, calculated at every tick would be
3357 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3359 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3360 * on nth tick when cpu may be busy, then we have:
3361 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3362 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3364 * decay_load_missed() below does efficient calculation of
3365 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3366 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3368 * The calculation is approximated on a 128 point scale.
3369 * degrade_zero_ticks is the number of ticks after which load at any
3370 * particular idx is approximated to be zero.
3371 * degrade_factor is a precomputed table, a row for each load idx.
3372 * Each column corresponds to degradation factor for a power of two ticks,
3373 * based on 128 point scale.
3375 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3376 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3378 * With this power of 2 load factors, we can degrade the load n times
3379 * by looking at 1 bits in n and doing as many mult/shift instead of
3380 * n mult/shifts needed by the exact degradation.
3382 #define DEGRADE_SHIFT 7
3383 static const unsigned char
3384 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3385 static const unsigned char
3386 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3387 {0, 0, 0, 0, 0, 0, 0, 0},
3388 {64, 32, 8, 0, 0, 0, 0, 0},
3389 {96, 72, 40, 12, 1, 0, 0},
3390 {112, 98, 75, 43, 15, 1, 0},
3391 {120, 112, 98, 76, 45, 16, 2} };
3394 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3395 * would be when CPU is idle and so we just decay the old load without
3396 * adding any new load.
3398 static unsigned long
3399 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3403 if (!missed_updates)
3406 if (missed_updates >= degrade_zero_ticks[idx])
3410 return load >> missed_updates;
3412 while (missed_updates) {
3413 if (missed_updates % 2)
3414 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3416 missed_updates >>= 1;
3423 * Update rq->cpu_load[] statistics. This function is usually called every
3424 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3425 * every tick. We fix it up based on jiffies.
3427 static void update_cpu_load(struct rq *this_rq)
3429 unsigned long this_load = this_rq->load.weight;
3430 unsigned long curr_jiffies = jiffies;
3431 unsigned long pending_updates;
3434 this_rq->nr_load_updates++;
3436 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3437 if (curr_jiffies == this_rq->last_load_update_tick)
3440 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3441 this_rq->last_load_update_tick = curr_jiffies;
3443 /* Update our load: */
3444 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3445 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3446 unsigned long old_load, new_load;
3448 /* scale is effectively 1 << i now, and >> i divides by scale */
3450 old_load = this_rq->cpu_load[i];
3451 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3452 new_load = this_load;
3454 * Round up the averaging division if load is increasing. This
3455 * prevents us from getting stuck on 9 if the load is 10, for
3458 if (new_load > old_load)
3459 new_load += scale - 1;
3461 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3464 sched_avg_update(this_rq);
3467 static void update_cpu_load_active(struct rq *this_rq)
3469 update_cpu_load(this_rq);
3471 calc_load_account_active(this_rq);
3477 * sched_exec - execve() is a valuable balancing opportunity, because at
3478 * this point the task has the smallest effective memory and cache footprint.
3480 void sched_exec(void)
3482 struct task_struct *p = current;
3483 unsigned long flags;
3487 rq = task_rq_lock(p, &flags);
3488 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3489 if (dest_cpu == smp_processor_id())
3493 * select_task_rq() can race against ->cpus_allowed
3495 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3496 likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
3497 struct migration_arg arg = { p, dest_cpu };
3499 task_rq_unlock(rq, &flags);
3500 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3504 task_rq_unlock(rq, &flags);
3509 DEFINE_PER_CPU(struct kernel_stat, kstat);
3511 EXPORT_PER_CPU_SYMBOL(kstat);
3514 * Return any ns on the sched_clock that have not yet been accounted in
3515 * @p in case that task is currently running.
3517 * Called with task_rq_lock() held on @rq.
3519 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3523 if (task_current(rq, p)) {
3524 update_rq_clock(rq);
3525 ns = rq->clock_task - p->se.exec_start;
3533 unsigned long long task_delta_exec(struct task_struct *p)
3535 unsigned long flags;
3539 rq = task_rq_lock(p, &flags);
3540 ns = do_task_delta_exec(p, rq);
3541 task_rq_unlock(rq, &flags);
3547 * Return accounted runtime for the task.
3548 * In case the task is currently running, return the runtime plus current's
3549 * pending runtime that have not been accounted yet.
3551 unsigned long long task_sched_runtime(struct task_struct *p)
3553 unsigned long flags;
3557 rq = task_rq_lock(p, &flags);
3558 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3559 task_rq_unlock(rq, &flags);
3565 * Return sum_exec_runtime for the thread group.
3566 * In case the task is currently running, return the sum plus current's
3567 * pending runtime that have not been accounted yet.
3569 * Note that the thread group might have other running tasks as well,
3570 * so the return value not includes other pending runtime that other
3571 * running tasks might have.
3573 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3575 struct task_cputime totals;
3576 unsigned long flags;
3580 rq = task_rq_lock(p, &flags);
3581 thread_group_cputime(p, &totals);
3582 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3583 task_rq_unlock(rq, &flags);
3589 * Account user cpu time to a process.
3590 * @p: the process that the cpu time gets accounted to
3591 * @cputime: the cpu time spent in user space since the last update
3592 * @cputime_scaled: cputime scaled by cpu frequency
3594 void account_user_time(struct task_struct *p, cputime_t cputime,
3595 cputime_t cputime_scaled)
3597 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3600 /* Add user time to process. */
3601 p->utime = cputime_add(p->utime, cputime);
3602 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3603 account_group_user_time(p, cputime);
3605 /* Add user time to cpustat. */
3606 tmp = cputime_to_cputime64(cputime);
3607 if (TASK_NICE(p) > 0)
3608 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3610 cpustat->user = cputime64_add(cpustat->user, tmp);
3612 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3613 /* Account for user time used */
3614 acct_update_integrals(p);
3618 * Account guest cpu time to a process.
3619 * @p: the process that the cpu time gets accounted to
3620 * @cputime: the cpu time spent in virtual machine since the last update
3621 * @cputime_scaled: cputime scaled by cpu frequency
3623 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3624 cputime_t cputime_scaled)
3627 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3629 tmp = cputime_to_cputime64(cputime);
3631 /* Add guest time to process. */
3632 p->utime = cputime_add(p->utime, cputime);
3633 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3634 account_group_user_time(p, cputime);
3635 p->gtime = cputime_add(p->gtime, cputime);
3637 /* Add guest time to cpustat. */
3638 if (TASK_NICE(p) > 0) {
3639 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3640 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3642 cpustat->user = cputime64_add(cpustat->user, tmp);
3643 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3648 * Account system cpu time to a process.
3649 * @p: the process that the cpu time gets accounted to
3650 * @hardirq_offset: the offset to subtract from hardirq_count()
3651 * @cputime: the cpu time spent in kernel space since the last update
3652 * @cputime_scaled: cputime scaled by cpu frequency
3654 void account_system_time(struct task_struct *p, int hardirq_offset,
3655 cputime_t cputime, cputime_t cputime_scaled)
3657 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3660 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3661 account_guest_time(p, cputime, cputime_scaled);
3665 /* Add system time to process. */
3666 p->stime = cputime_add(p->stime, cputime);
3667 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3668 account_group_system_time(p, cputime);
3670 /* Add system time to cpustat. */
3671 tmp = cputime_to_cputime64(cputime);
3672 if (hardirq_count() - hardirq_offset)
3673 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3674 else if (in_serving_softirq())
3675 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3677 cpustat->system = cputime64_add(cpustat->system, tmp);
3679 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3681 /* Account for system time used */
3682 acct_update_integrals(p);
3686 * Account for involuntary wait time.
3687 * @steal: the cpu time spent in involuntary wait
3689 void account_steal_time(cputime_t cputime)
3691 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3692 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3694 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3698 * Account for idle time.
3699 * @cputime: the cpu time spent in idle wait
3701 void account_idle_time(cputime_t cputime)
3703 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3704 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3705 struct rq *rq = this_rq();
3707 if (atomic_read(&rq->nr_iowait) > 0)
3708 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3710 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3713 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3716 * Account a single tick of cpu time.
3717 * @p: the process that the cpu time gets accounted to
3718 * @user_tick: indicates if the tick is a user or a system tick
3720 void account_process_tick(struct task_struct *p, int user_tick)
3722 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3723 struct rq *rq = this_rq();
3726 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3727 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3728 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3731 account_idle_time(cputime_one_jiffy);
3735 * Account multiple ticks of steal time.
3736 * @p: the process from which the cpu time has been stolen
3737 * @ticks: number of stolen ticks
3739 void account_steal_ticks(unsigned long ticks)
3741 account_steal_time(jiffies_to_cputime(ticks));
3745 * Account multiple ticks of idle time.
3746 * @ticks: number of stolen ticks
3748 void account_idle_ticks(unsigned long ticks)
3750 account_idle_time(jiffies_to_cputime(ticks));
3756 * Use precise platform statistics if available:
3758 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3759 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3765 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3767 struct task_cputime cputime;
3769 thread_group_cputime(p, &cputime);
3771 *ut = cputime.utime;
3772 *st = cputime.stime;
3776 #ifndef nsecs_to_cputime
3777 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3780 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3782 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3785 * Use CFS's precise accounting:
3787 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3793 do_div(temp, total);
3794 utime = (cputime_t)temp;
3799 * Compare with previous values, to keep monotonicity:
3801 p->prev_utime = max(p->prev_utime, utime);
3802 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3804 *ut = p->prev_utime;
3805 *st = p->prev_stime;
3809 * Must be called with siglock held.
3811 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3813 struct signal_struct *sig = p->signal;
3814 struct task_cputime cputime;
3815 cputime_t rtime, utime, total;
3817 thread_group_cputime(p, &cputime);
3819 total = cputime_add(cputime.utime, cputime.stime);
3820 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3825 temp *= cputime.utime;
3826 do_div(temp, total);
3827 utime = (cputime_t)temp;
3831 sig->prev_utime = max(sig->prev_utime, utime);
3832 sig->prev_stime = max(sig->prev_stime,
3833 cputime_sub(rtime, sig->prev_utime));
3835 *ut = sig->prev_utime;
3836 *st = sig->prev_stime;
3841 * This function gets called by the timer code, with HZ frequency.
3842 * We call it with interrupts disabled.
3844 * It also gets called by the fork code, when changing the parent's
3847 void scheduler_tick(void)
3849 int cpu = smp_processor_id();
3850 struct rq *rq = cpu_rq(cpu);
3851 struct task_struct *curr = rq->curr;
3855 raw_spin_lock(&rq->lock);
3856 update_rq_clock(rq);
3857 update_cpu_load_active(rq);
3858 curr->sched_class->task_tick(rq, curr, 0);
3859 raw_spin_unlock(&rq->lock);
3861 perf_event_task_tick();
3864 rq->idle_at_tick = idle_cpu(cpu);
3865 trigger_load_balance(rq, cpu);
3869 notrace unsigned long get_parent_ip(unsigned long addr)
3871 if (in_lock_functions(addr)) {
3872 addr = CALLER_ADDR2;
3873 if (in_lock_functions(addr))
3874 addr = CALLER_ADDR3;
3879 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3880 defined(CONFIG_PREEMPT_TRACER))
3882 void __kprobes add_preempt_count(int val)
3884 #ifdef CONFIG_DEBUG_PREEMPT
3888 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3891 preempt_count() += val;
3892 #ifdef CONFIG_DEBUG_PREEMPT
3894 * Spinlock count overflowing soon?
3896 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3899 if (preempt_count() == val)
3900 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3902 EXPORT_SYMBOL(add_preempt_count);
3904 void __kprobes sub_preempt_count(int val)
3906 #ifdef CONFIG_DEBUG_PREEMPT
3910 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3913 * Is the spinlock portion underflowing?
3915 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3916 !(preempt_count() & PREEMPT_MASK)))
3920 if (preempt_count() == val)
3921 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3922 preempt_count() -= val;
3924 EXPORT_SYMBOL(sub_preempt_count);
3929 * Print scheduling while atomic bug:
3931 static noinline void __schedule_bug(struct task_struct *prev)
3933 struct pt_regs *regs = get_irq_regs();
3935 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3936 prev->comm, prev->pid, preempt_count());
3938 debug_show_held_locks(prev);
3940 if (irqs_disabled())
3941 print_irqtrace_events(prev);
3950 * Various schedule()-time debugging checks and statistics:
3952 static inline void schedule_debug(struct task_struct *prev)
3955 * Test if we are atomic. Since do_exit() needs to call into
3956 * schedule() atomically, we ignore that path for now.
3957 * Otherwise, whine if we are scheduling when we should not be.
3959 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3960 __schedule_bug(prev);
3962 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3964 schedstat_inc(this_rq(), sched_count);
3965 #ifdef CONFIG_SCHEDSTATS
3966 if (unlikely(prev->lock_depth >= 0)) {
3967 schedstat_inc(this_rq(), bkl_count);
3968 schedstat_inc(prev, sched_info.bkl_count);
3973 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3976 update_rq_clock(rq);
3977 prev->sched_class->put_prev_task(rq, prev);
3981 * Pick up the highest-prio task:
3983 static inline struct task_struct *
3984 pick_next_task(struct rq *rq)
3986 const struct sched_class *class;
3987 struct task_struct *p;
3990 * Optimization: we know that if all tasks are in
3991 * the fair class we can call that function directly:
3993 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3994 p = fair_sched_class.pick_next_task(rq);
3999 for_each_class(class) {
4000 p = class->pick_next_task(rq);
4005 BUG(); /* the idle class will always have a runnable task */
4009 * schedule() is the main scheduler function.
4011 asmlinkage void __sched schedule(void)
4013 struct task_struct *prev, *next;
4014 unsigned long *switch_count;
4020 cpu = smp_processor_id();
4022 rcu_note_context_switch(cpu);
4025 release_kernel_lock(prev);
4026 need_resched_nonpreemptible:
4028 schedule_debug(prev);
4030 if (sched_feat(HRTICK))
4033 raw_spin_lock_irq(&rq->lock);
4035 switch_count = &prev->nivcsw;
4036 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4037 if (unlikely(signal_pending_state(prev->state, prev))) {
4038 prev->state = TASK_RUNNING;
4041 * If a worker is going to sleep, notify and
4042 * ask workqueue whether it wants to wake up a
4043 * task to maintain concurrency. If so, wake
4046 if (prev->flags & PF_WQ_WORKER) {
4047 struct task_struct *to_wakeup;
4049 to_wakeup = wq_worker_sleeping(prev, cpu);
4051 try_to_wake_up_local(to_wakeup);
4053 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4055 switch_count = &prev->nvcsw;
4058 pre_schedule(rq, prev);
4060 if (unlikely(!rq->nr_running))
4061 idle_balance(cpu, rq);
4063 put_prev_task(rq, prev);
4064 next = pick_next_task(rq);
4065 clear_tsk_need_resched(prev);
4066 rq->skip_clock_update = 0;
4068 if (likely(prev != next)) {
4069 sched_info_switch(prev, next);
4070 perf_event_task_sched_out(prev, next);
4075 WARN_ON_ONCE(test_tsk_need_resched(next));
4077 context_switch(rq, prev, next); /* unlocks the rq */
4079 * The context switch have flipped the stack from under us
4080 * and restored the local variables which were saved when
4081 * this task called schedule() in the past. prev == current
4082 * is still correct, but it can be moved to another cpu/rq.
4084 cpu = smp_processor_id();
4087 raw_spin_unlock_irq(&rq->lock);
4091 if (unlikely(reacquire_kernel_lock(prev)))
4092 goto need_resched_nonpreemptible;
4094 preempt_enable_no_resched();
4098 EXPORT_SYMBOL(schedule);
4100 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4102 * Look out! "owner" is an entirely speculative pointer
4103 * access and not reliable.
4105 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
4110 if (!sched_feat(OWNER_SPIN))
4113 #ifdef CONFIG_DEBUG_PAGEALLOC
4115 * Need to access the cpu field knowing that
4116 * DEBUG_PAGEALLOC could have unmapped it if
4117 * the mutex owner just released it and exited.
4119 if (probe_kernel_address(&owner->cpu, cpu))
4126 * Even if the access succeeded (likely case),
4127 * the cpu field may no longer be valid.
4129 if (cpu >= nr_cpumask_bits)
4133 * We need to validate that we can do a
4134 * get_cpu() and that we have the percpu area.
4136 if (!cpu_online(cpu))
4143 * Owner changed, break to re-assess state.
4145 if (lock->owner != owner) {
4147 * If the lock has switched to a different owner,
4148 * we likely have heavy contention. Return 0 to quit
4149 * optimistic spinning and not contend further:
4157 * Is that owner really running on that cpu?
4159 if (task_thread_info(rq->curr) != owner || need_resched())
4169 #ifdef CONFIG_PREEMPT
4171 * this is the entry point to schedule() from in-kernel preemption
4172 * off of preempt_enable. Kernel preemptions off return from interrupt
4173 * occur there and call schedule directly.
4175 asmlinkage void __sched notrace preempt_schedule(void)
4177 struct thread_info *ti = current_thread_info();
4180 * If there is a non-zero preempt_count or interrupts are disabled,
4181 * we do not want to preempt the current task. Just return..
4183 if (likely(ti->preempt_count || irqs_disabled()))
4187 add_preempt_count_notrace(PREEMPT_ACTIVE);
4189 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4192 * Check again in case we missed a preemption opportunity
4193 * between schedule and now.
4196 } while (need_resched());
4198 EXPORT_SYMBOL(preempt_schedule);
4201 * this is the entry point to schedule() from kernel preemption
4202 * off of irq context.
4203 * Note, that this is called and return with irqs disabled. This will
4204 * protect us against recursive calling from irq.
4206 asmlinkage void __sched preempt_schedule_irq(void)
4208 struct thread_info *ti = current_thread_info();
4210 /* Catch callers which need to be fixed */
4211 BUG_ON(ti->preempt_count || !irqs_disabled());
4214 add_preempt_count(PREEMPT_ACTIVE);
4217 local_irq_disable();
4218 sub_preempt_count(PREEMPT_ACTIVE);
4221 * Check again in case we missed a preemption opportunity
4222 * between schedule and now.
4225 } while (need_resched());
4228 #endif /* CONFIG_PREEMPT */
4230 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4233 return try_to_wake_up(curr->private, mode, wake_flags);
4235 EXPORT_SYMBOL(default_wake_function);
4238 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4239 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4240 * number) then we wake all the non-exclusive tasks and one exclusive task.
4242 * There are circumstances in which we can try to wake a task which has already
4243 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4244 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4246 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4247 int nr_exclusive, int wake_flags, void *key)
4249 wait_queue_t *curr, *next;
4251 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4252 unsigned flags = curr->flags;
4254 if (curr->func(curr, mode, wake_flags, key) &&
4255 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4261 * __wake_up - wake up threads blocked on a waitqueue.
4263 * @mode: which threads
4264 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4265 * @key: is directly passed to the wakeup function
4267 * It may be assumed that this function implies a write memory barrier before
4268 * changing the task state if and only if any tasks are woken up.
4270 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4271 int nr_exclusive, void *key)
4273 unsigned long flags;
4275 spin_lock_irqsave(&q->lock, flags);
4276 __wake_up_common(q, mode, nr_exclusive, 0, key);
4277 spin_unlock_irqrestore(&q->lock, flags);
4279 EXPORT_SYMBOL(__wake_up);
4282 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4284 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4286 __wake_up_common(q, mode, 1, 0, NULL);
4288 EXPORT_SYMBOL_GPL(__wake_up_locked);
4290 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4292 __wake_up_common(q, mode, 1, 0, key);
4296 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4298 * @mode: which threads
4299 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4300 * @key: opaque value to be passed to wakeup targets
4302 * The sync wakeup differs that the waker knows that it will schedule
4303 * away soon, so while the target thread will be woken up, it will not
4304 * be migrated to another CPU - ie. the two threads are 'synchronized'
4305 * with each other. This can prevent needless bouncing between CPUs.
4307 * On UP it can prevent extra preemption.
4309 * It may be assumed that this function implies a write memory barrier before
4310 * changing the task state if and only if any tasks are woken up.
4312 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4313 int nr_exclusive, void *key)
4315 unsigned long flags;
4316 int wake_flags = WF_SYNC;
4321 if (unlikely(!nr_exclusive))
4324 spin_lock_irqsave(&q->lock, flags);
4325 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4326 spin_unlock_irqrestore(&q->lock, flags);
4328 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4331 * __wake_up_sync - see __wake_up_sync_key()
4333 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4335 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4337 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4340 * complete: - signals a single thread waiting on this completion
4341 * @x: holds the state of this particular completion
4343 * This will wake up a single thread waiting on this completion. Threads will be
4344 * awakened in the same order in which they were queued.
4346 * See also complete_all(), wait_for_completion() and related routines.
4348 * It may be assumed that this function implies a write memory barrier before
4349 * changing the task state if and only if any tasks are woken up.
4351 void complete(struct completion *x)
4353 unsigned long flags;
4355 spin_lock_irqsave(&x->wait.lock, flags);
4357 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4358 spin_unlock_irqrestore(&x->wait.lock, flags);
4360 EXPORT_SYMBOL(complete);
4363 * complete_all: - signals all threads waiting on this completion
4364 * @x: holds the state of this particular completion
4366 * This will wake up all threads waiting on this particular completion event.
4368 * It may be assumed that this function implies a write memory barrier before
4369 * changing the task state if and only if any tasks are woken up.
4371 void complete_all(struct completion *x)
4373 unsigned long flags;
4375 spin_lock_irqsave(&x->wait.lock, flags);
4376 x->done += UINT_MAX/2;
4377 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4378 spin_unlock_irqrestore(&x->wait.lock, flags);
4380 EXPORT_SYMBOL(complete_all);
4382 static inline long __sched
4383 do_wait_for_common(struct completion *x, long timeout, int state)
4386 DECLARE_WAITQUEUE(wait, current);
4388 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4390 if (signal_pending_state(state, current)) {
4391 timeout = -ERESTARTSYS;
4394 __set_current_state(state);
4395 spin_unlock_irq(&x->wait.lock);
4396 timeout = schedule_timeout(timeout);
4397 spin_lock_irq(&x->wait.lock);
4398 } while (!x->done && timeout);
4399 __remove_wait_queue(&x->wait, &wait);
4404 return timeout ?: 1;
4408 wait_for_common(struct completion *x, long timeout, int state)
4412 spin_lock_irq(&x->wait.lock);
4413 timeout = do_wait_for_common(x, timeout, state);
4414 spin_unlock_irq(&x->wait.lock);
4419 * wait_for_completion: - waits for completion of a task
4420 * @x: holds the state of this particular completion
4422 * This waits to be signaled for completion of a specific task. It is NOT
4423 * interruptible and there is no timeout.
4425 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4426 * and interrupt capability. Also see complete().
4428 void __sched wait_for_completion(struct completion *x)
4430 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4432 EXPORT_SYMBOL(wait_for_completion);
4435 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4436 * @x: holds the state of this particular completion
4437 * @timeout: timeout value in jiffies
4439 * This waits for either a completion of a specific task to be signaled or for a
4440 * specified timeout to expire. The timeout is in jiffies. It is not
4443 unsigned long __sched
4444 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4446 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4448 EXPORT_SYMBOL(wait_for_completion_timeout);
4451 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4452 * @x: holds the state of this particular completion
4454 * This waits for completion of a specific task to be signaled. It is
4457 int __sched wait_for_completion_interruptible(struct completion *x)
4459 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4460 if (t == -ERESTARTSYS)
4464 EXPORT_SYMBOL(wait_for_completion_interruptible);
4467 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4468 * @x: holds the state of this particular completion
4469 * @timeout: timeout value in jiffies
4471 * This waits for either a completion of a specific task to be signaled or for a
4472 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4474 unsigned long __sched
4475 wait_for_completion_interruptible_timeout(struct completion *x,
4476 unsigned long timeout)
4478 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4480 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4483 * wait_for_completion_killable: - waits for completion of a task (killable)
4484 * @x: holds the state of this particular completion
4486 * This waits to be signaled for completion of a specific task. It can be
4487 * interrupted by a kill signal.
4489 int __sched wait_for_completion_killable(struct completion *x)
4491 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4492 if (t == -ERESTARTSYS)
4496 EXPORT_SYMBOL(wait_for_completion_killable);
4499 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4500 * @x: holds the state of this particular completion
4501 * @timeout: timeout value in jiffies
4503 * This waits for either a completion of a specific task to be
4504 * signaled or for a specified timeout to expire. It can be
4505 * interrupted by a kill signal. The timeout is in jiffies.
4507 unsigned long __sched
4508 wait_for_completion_killable_timeout(struct completion *x,
4509 unsigned long timeout)
4511 return wait_for_common(x, timeout, TASK_KILLABLE);
4513 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4516 * try_wait_for_completion - try to decrement a completion without blocking
4517 * @x: completion structure
4519 * Returns: 0 if a decrement cannot be done without blocking
4520 * 1 if a decrement succeeded.
4522 * If a completion is being used as a counting completion,
4523 * attempt to decrement the counter without blocking. This
4524 * enables us to avoid waiting if the resource the completion
4525 * is protecting is not available.
4527 bool try_wait_for_completion(struct completion *x)
4529 unsigned long flags;
4532 spin_lock_irqsave(&x->wait.lock, flags);
4537 spin_unlock_irqrestore(&x->wait.lock, flags);
4540 EXPORT_SYMBOL(try_wait_for_completion);
4543 * completion_done - Test to see if a completion has any waiters
4544 * @x: completion structure
4546 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4547 * 1 if there are no waiters.
4550 bool completion_done(struct completion *x)
4552 unsigned long flags;
4555 spin_lock_irqsave(&x->wait.lock, flags);
4558 spin_unlock_irqrestore(&x->wait.lock, flags);
4561 EXPORT_SYMBOL(completion_done);
4564 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4566 unsigned long flags;
4569 init_waitqueue_entry(&wait, current);
4571 __set_current_state(state);
4573 spin_lock_irqsave(&q->lock, flags);
4574 __add_wait_queue(q, &wait);
4575 spin_unlock(&q->lock);
4576 timeout = schedule_timeout(timeout);
4577 spin_lock_irq(&q->lock);
4578 __remove_wait_queue(q, &wait);
4579 spin_unlock_irqrestore(&q->lock, flags);
4584 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4586 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4588 EXPORT_SYMBOL(interruptible_sleep_on);
4591 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4593 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4595 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4597 void __sched sleep_on(wait_queue_head_t *q)
4599 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4601 EXPORT_SYMBOL(sleep_on);
4603 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4605 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4607 EXPORT_SYMBOL(sleep_on_timeout);
4609 #ifdef CONFIG_RT_MUTEXES
4612 * rt_mutex_setprio - set the current priority of a task
4614 * @prio: prio value (kernel-internal form)
4616 * This function changes the 'effective' priority of a task. It does
4617 * not touch ->normal_prio like __setscheduler().
4619 * Used by the rt_mutex code to implement priority inheritance logic.
4621 void rt_mutex_setprio(struct task_struct *p, int prio)
4623 unsigned long flags;
4624 int oldprio, on_rq, running;
4626 const struct sched_class *prev_class;
4628 BUG_ON(prio < 0 || prio > MAX_PRIO);
4630 rq = task_rq_lock(p, &flags);
4632 trace_sched_pi_setprio(p, prio);
4634 prev_class = p->sched_class;
4635 on_rq = p->se.on_rq;
4636 running = task_current(rq, p);
4638 dequeue_task(rq, p, 0);
4640 p->sched_class->put_prev_task(rq, p);
4643 p->sched_class = &rt_sched_class;
4645 p->sched_class = &fair_sched_class;
4650 p->sched_class->set_curr_task(rq);
4652 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4654 check_class_changed(rq, p, prev_class, oldprio, running);
4656 task_rq_unlock(rq, &flags);
4661 void set_user_nice(struct task_struct *p, long nice)
4663 int old_prio, delta, on_rq;
4664 unsigned long flags;
4667 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4670 * We have to be careful, if called from sys_setpriority(),
4671 * the task might be in the middle of scheduling on another CPU.
4673 rq = task_rq_lock(p, &flags);
4675 * The RT priorities are set via sched_setscheduler(), but we still
4676 * allow the 'normal' nice value to be set - but as expected
4677 * it wont have any effect on scheduling until the task is
4678 * SCHED_FIFO/SCHED_RR:
4680 if (task_has_rt_policy(p)) {
4681 p->static_prio = NICE_TO_PRIO(nice);
4684 on_rq = p->se.on_rq;
4686 dequeue_task(rq, p, 0);
4688 p->static_prio = NICE_TO_PRIO(nice);
4691 p->prio = effective_prio(p);
4692 delta = p->prio - old_prio;
4695 enqueue_task(rq, p, 0);
4697 * If the task increased its priority or is running and
4698 * lowered its priority, then reschedule its CPU:
4700 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4701 resched_task(rq->curr);
4704 task_rq_unlock(rq, &flags);
4706 EXPORT_SYMBOL(set_user_nice);
4709 * can_nice - check if a task can reduce its nice value
4713 int can_nice(const struct task_struct *p, const int nice)
4715 /* convert nice value [19,-20] to rlimit style value [1,40] */
4716 int nice_rlim = 20 - nice;
4718 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4719 capable(CAP_SYS_NICE));
4722 #ifdef __ARCH_WANT_SYS_NICE
4725 * sys_nice - change the priority of the current process.
4726 * @increment: priority increment
4728 * sys_setpriority is a more generic, but much slower function that
4729 * does similar things.
4731 SYSCALL_DEFINE1(nice, int, increment)
4736 * Setpriority might change our priority at the same moment.
4737 * We don't have to worry. Conceptually one call occurs first
4738 * and we have a single winner.
4740 if (increment < -40)
4745 nice = TASK_NICE(current) + increment;
4751 if (increment < 0 && !can_nice(current, nice))
4754 retval = security_task_setnice(current, nice);
4758 set_user_nice(current, nice);
4765 * task_prio - return the priority value of a given task.
4766 * @p: the task in question.
4768 * This is the priority value as seen by users in /proc.
4769 * RT tasks are offset by -200. Normal tasks are centered
4770 * around 0, value goes from -16 to +15.
4772 int task_prio(const struct task_struct *p)
4774 return p->prio - MAX_RT_PRIO;
4778 * task_nice - return the nice value of a given task.
4779 * @p: the task in question.
4781 int task_nice(const struct task_struct *p)
4783 return TASK_NICE(p);
4785 EXPORT_SYMBOL(task_nice);
4788 * idle_cpu - is a given cpu idle currently?
4789 * @cpu: the processor in question.
4791 int idle_cpu(int cpu)
4793 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4797 * idle_task - return the idle task for a given cpu.
4798 * @cpu: the processor in question.
4800 struct task_struct *idle_task(int cpu)
4802 return cpu_rq(cpu)->idle;
4806 * find_process_by_pid - find a process with a matching PID value.
4807 * @pid: the pid in question.
4809 static struct task_struct *find_process_by_pid(pid_t pid)
4811 return pid ? find_task_by_vpid(pid) : current;
4814 /* Actually do priority change: must hold rq lock. */
4816 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4818 BUG_ON(p->se.on_rq);
4821 p->rt_priority = prio;
4822 p->normal_prio = normal_prio(p);
4823 /* we are holding p->pi_lock already */
4824 p->prio = rt_mutex_getprio(p);
4825 if (rt_prio(p->prio))
4826 p->sched_class = &rt_sched_class;
4828 p->sched_class = &fair_sched_class;
4833 * check the target process has a UID that matches the current process's
4835 static bool check_same_owner(struct task_struct *p)
4837 const struct cred *cred = current_cred(), *pcred;
4841 pcred = __task_cred(p);
4842 match = (cred->euid == pcred->euid ||
4843 cred->euid == pcred->uid);
4848 static int __sched_setscheduler(struct task_struct *p, int policy,
4849 struct sched_param *param, bool user)
4851 int retval, oldprio, oldpolicy = -1, on_rq, running;
4852 unsigned long flags;
4853 const struct sched_class *prev_class;
4857 /* may grab non-irq protected spin_locks */
4858 BUG_ON(in_interrupt());
4860 /* double check policy once rq lock held */
4862 reset_on_fork = p->sched_reset_on_fork;
4863 policy = oldpolicy = p->policy;
4865 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4866 policy &= ~SCHED_RESET_ON_FORK;
4868 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4869 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4870 policy != SCHED_IDLE)
4875 * Valid priorities for SCHED_FIFO and SCHED_RR are
4876 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4877 * SCHED_BATCH and SCHED_IDLE is 0.
4879 if (param->sched_priority < 0 ||
4880 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4881 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4883 if (rt_policy(policy) != (param->sched_priority != 0))
4887 * Allow unprivileged RT tasks to decrease priority:
4889 if (user && !capable(CAP_SYS_NICE)) {
4890 if (rt_policy(policy)) {
4891 unsigned long rlim_rtprio =
4892 task_rlimit(p, RLIMIT_RTPRIO);
4894 /* can't set/change the rt policy */
4895 if (policy != p->policy && !rlim_rtprio)
4898 /* can't increase priority */
4899 if (param->sched_priority > p->rt_priority &&
4900 param->sched_priority > rlim_rtprio)
4904 * Like positive nice levels, dont allow tasks to
4905 * move out of SCHED_IDLE either:
4907 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4910 /* can't change other user's priorities */
4911 if (!check_same_owner(p))
4914 /* Normal users shall not reset the sched_reset_on_fork flag */
4915 if (p->sched_reset_on_fork && !reset_on_fork)
4920 retval = security_task_setscheduler(p);
4926 * make sure no PI-waiters arrive (or leave) while we are
4927 * changing the priority of the task:
4929 raw_spin_lock_irqsave(&p->pi_lock, flags);
4931 * To be able to change p->policy safely, the apropriate
4932 * runqueue lock must be held.
4934 rq = __task_rq_lock(p);
4937 * Changing the policy of the stop threads its a very bad idea
4939 if (p == rq->stop) {
4940 __task_rq_unlock(rq);
4941 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4945 #ifdef CONFIG_RT_GROUP_SCHED
4948 * Do not allow realtime tasks into groups that have no runtime
4951 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4952 task_group(p)->rt_bandwidth.rt_runtime == 0) {
4953 __task_rq_unlock(rq);
4954 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4960 /* recheck policy now with rq lock held */
4961 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4962 policy = oldpolicy = -1;
4963 __task_rq_unlock(rq);
4964 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4967 on_rq = p->se.on_rq;
4968 running = task_current(rq, p);
4970 deactivate_task(rq, p, 0);
4972 p->sched_class->put_prev_task(rq, p);
4974 p->sched_reset_on_fork = reset_on_fork;
4977 prev_class = p->sched_class;
4978 __setscheduler(rq, p, policy, param->sched_priority);
4981 p->sched_class->set_curr_task(rq);
4983 activate_task(rq, p, 0);
4985 check_class_changed(rq, p, prev_class, oldprio, running);
4987 __task_rq_unlock(rq);
4988 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4990 rt_mutex_adjust_pi(p);
4996 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4997 * @p: the task in question.
4998 * @policy: new policy.
4999 * @param: structure containing the new RT priority.
5001 * NOTE that the task may be already dead.
5003 int sched_setscheduler(struct task_struct *p, int policy,
5004 struct sched_param *param)
5006 return __sched_setscheduler(p, policy, param, true);
5008 EXPORT_SYMBOL_GPL(sched_setscheduler);
5011 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5012 * @p: the task in question.
5013 * @policy: new policy.
5014 * @param: structure containing the new RT priority.
5016 * Just like sched_setscheduler, only don't bother checking if the
5017 * current context has permission. For example, this is needed in
5018 * stop_machine(): we create temporary high priority worker threads,
5019 * but our caller might not have that capability.
5021 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5022 struct sched_param *param)
5024 return __sched_setscheduler(p, policy, param, false);
5028 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5030 struct sched_param lparam;
5031 struct task_struct *p;
5034 if (!param || pid < 0)
5036 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5041 p = find_process_by_pid(pid);
5043 retval = sched_setscheduler(p, policy, &lparam);
5050 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5051 * @pid: the pid in question.
5052 * @policy: new policy.
5053 * @param: structure containing the new RT priority.
5055 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5056 struct sched_param __user *, param)
5058 /* negative values for policy are not valid */
5062 return do_sched_setscheduler(pid, policy, param);
5066 * sys_sched_setparam - set/change the RT priority of a thread
5067 * @pid: the pid in question.
5068 * @param: structure containing the new RT priority.
5070 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5072 return do_sched_setscheduler(pid, -1, param);
5076 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5077 * @pid: the pid in question.
5079 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5081 struct task_struct *p;
5089 p = find_process_by_pid(pid);
5091 retval = security_task_getscheduler(p);
5094 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5101 * sys_sched_getparam - get the RT priority of a thread
5102 * @pid: the pid in question.
5103 * @param: structure containing the RT priority.
5105 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5107 struct sched_param lp;
5108 struct task_struct *p;
5111 if (!param || pid < 0)
5115 p = find_process_by_pid(pid);
5120 retval = security_task_getscheduler(p);
5124 lp.sched_priority = p->rt_priority;
5128 * This one might sleep, we cannot do it with a spinlock held ...
5130 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5139 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5141 cpumask_var_t cpus_allowed, new_mask;
5142 struct task_struct *p;
5148 p = find_process_by_pid(pid);
5155 /* Prevent p going away */
5159 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5163 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5165 goto out_free_cpus_allowed;
5168 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5171 retval = security_task_setscheduler(p);
5175 cpuset_cpus_allowed(p, cpus_allowed);
5176 cpumask_and(new_mask, in_mask, cpus_allowed);
5178 retval = set_cpus_allowed_ptr(p, new_mask);
5181 cpuset_cpus_allowed(p, cpus_allowed);
5182 if (!cpumask_subset(new_mask, cpus_allowed)) {
5184 * We must have raced with a concurrent cpuset
5185 * update. Just reset the cpus_allowed to the
5186 * cpuset's cpus_allowed
5188 cpumask_copy(new_mask, cpus_allowed);
5193 free_cpumask_var(new_mask);
5194 out_free_cpus_allowed:
5195 free_cpumask_var(cpus_allowed);
5202 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5203 struct cpumask *new_mask)
5205 if (len < cpumask_size())
5206 cpumask_clear(new_mask);
5207 else if (len > cpumask_size())
5208 len = cpumask_size();
5210 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5214 * sys_sched_setaffinity - set the cpu affinity of a process
5215 * @pid: pid of the process
5216 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5217 * @user_mask_ptr: user-space pointer to the new cpu mask
5219 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5220 unsigned long __user *, user_mask_ptr)
5222 cpumask_var_t new_mask;
5225 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5228 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5230 retval = sched_setaffinity(pid, new_mask);
5231 free_cpumask_var(new_mask);
5235 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5237 struct task_struct *p;
5238 unsigned long flags;
5246 p = find_process_by_pid(pid);
5250 retval = security_task_getscheduler(p);
5254 rq = task_rq_lock(p, &flags);
5255 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5256 task_rq_unlock(rq, &flags);
5266 * sys_sched_getaffinity - get the cpu affinity of a process
5267 * @pid: pid of the process
5268 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5269 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5271 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5272 unsigned long __user *, user_mask_ptr)
5277 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5279 if (len & (sizeof(unsigned long)-1))
5282 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5285 ret = sched_getaffinity(pid, mask);
5287 size_t retlen = min_t(size_t, len, cpumask_size());
5289 if (copy_to_user(user_mask_ptr, mask, retlen))
5294 free_cpumask_var(mask);
5300 * sys_sched_yield - yield the current processor to other threads.
5302 * This function yields the current CPU to other tasks. If there are no
5303 * other threads running on this CPU then this function will return.
5305 SYSCALL_DEFINE0(sched_yield)
5307 struct rq *rq = this_rq_lock();
5309 schedstat_inc(rq, yld_count);
5310 current->sched_class->yield_task(rq);
5313 * Since we are going to call schedule() anyway, there's
5314 * no need to preempt or enable interrupts:
5316 __release(rq->lock);
5317 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5318 do_raw_spin_unlock(&rq->lock);
5319 preempt_enable_no_resched();
5326 static inline int should_resched(void)
5328 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5331 static void __cond_resched(void)
5333 add_preempt_count(PREEMPT_ACTIVE);
5335 sub_preempt_count(PREEMPT_ACTIVE);
5338 int __sched _cond_resched(void)
5340 if (should_resched()) {
5346 EXPORT_SYMBOL(_cond_resched);
5349 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5350 * call schedule, and on return reacquire the lock.
5352 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5353 * operations here to prevent schedule() from being called twice (once via
5354 * spin_unlock(), once by hand).
5356 int __cond_resched_lock(spinlock_t *lock)
5358 int resched = should_resched();
5361 lockdep_assert_held(lock);
5363 if (spin_needbreak(lock) || resched) {
5374 EXPORT_SYMBOL(__cond_resched_lock);
5376 int __sched __cond_resched_softirq(void)
5378 BUG_ON(!in_softirq());
5380 if (should_resched()) {
5388 EXPORT_SYMBOL(__cond_resched_softirq);
5391 * yield - yield the current processor to other threads.
5393 * This is a shortcut for kernel-space yielding - it marks the
5394 * thread runnable and calls sys_sched_yield().
5396 void __sched yield(void)
5398 set_current_state(TASK_RUNNING);
5401 EXPORT_SYMBOL(yield);
5404 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5405 * that process accounting knows that this is a task in IO wait state.
5407 void __sched io_schedule(void)
5409 struct rq *rq = raw_rq();
5411 delayacct_blkio_start();
5412 atomic_inc(&rq->nr_iowait);
5413 current->in_iowait = 1;
5415 current->in_iowait = 0;
5416 atomic_dec(&rq->nr_iowait);
5417 delayacct_blkio_end();
5419 EXPORT_SYMBOL(io_schedule);
5421 long __sched io_schedule_timeout(long timeout)
5423 struct rq *rq = raw_rq();
5426 delayacct_blkio_start();
5427 atomic_inc(&rq->nr_iowait);
5428 current->in_iowait = 1;
5429 ret = schedule_timeout(timeout);
5430 current->in_iowait = 0;
5431 atomic_dec(&rq->nr_iowait);
5432 delayacct_blkio_end();
5437 * sys_sched_get_priority_max - return maximum RT priority.
5438 * @policy: scheduling class.
5440 * this syscall returns the maximum rt_priority that can be used
5441 * by a given scheduling class.
5443 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5450 ret = MAX_USER_RT_PRIO-1;
5462 * sys_sched_get_priority_min - return minimum RT priority.
5463 * @policy: scheduling class.
5465 * this syscall returns the minimum rt_priority that can be used
5466 * by a given scheduling class.
5468 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5486 * sys_sched_rr_get_interval - return the default timeslice of a process.
5487 * @pid: pid of the process.
5488 * @interval: userspace pointer to the timeslice value.
5490 * this syscall writes the default timeslice value of a given process
5491 * into the user-space timespec buffer. A value of '0' means infinity.
5493 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5494 struct timespec __user *, interval)
5496 struct task_struct *p;
5497 unsigned int time_slice;
5498 unsigned long flags;
5508 p = find_process_by_pid(pid);
5512 retval = security_task_getscheduler(p);
5516 rq = task_rq_lock(p, &flags);
5517 time_slice = p->sched_class->get_rr_interval(rq, p);
5518 task_rq_unlock(rq, &flags);
5521 jiffies_to_timespec(time_slice, &t);
5522 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5530 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5532 void sched_show_task(struct task_struct *p)
5534 unsigned long free = 0;
5537 state = p->state ? __ffs(p->state) + 1 : 0;
5538 printk(KERN_INFO "%-13.13s %c", p->comm,
5539 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5540 #if BITS_PER_LONG == 32
5541 if (state == TASK_RUNNING)
5542 printk(KERN_CONT " running ");
5544 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5546 if (state == TASK_RUNNING)
5547 printk(KERN_CONT " running task ");
5549 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5551 #ifdef CONFIG_DEBUG_STACK_USAGE
5552 free = stack_not_used(p);
5554 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5555 task_pid_nr(p), task_pid_nr(p->real_parent),
5556 (unsigned long)task_thread_info(p)->flags);
5558 show_stack(p, NULL);
5561 void show_state_filter(unsigned long state_filter)
5563 struct task_struct *g, *p;
5565 #if BITS_PER_LONG == 32
5567 " task PC stack pid father\n");
5570 " task PC stack pid father\n");
5572 read_lock(&tasklist_lock);
5573 do_each_thread(g, p) {
5575 * reset the NMI-timeout, listing all files on a slow
5576 * console might take alot of time:
5578 touch_nmi_watchdog();
5579 if (!state_filter || (p->state & state_filter))
5581 } while_each_thread(g, p);
5583 touch_all_softlockup_watchdogs();
5585 #ifdef CONFIG_SCHED_DEBUG
5586 sysrq_sched_debug_show();
5588 read_unlock(&tasklist_lock);
5590 * Only show locks if all tasks are dumped:
5593 debug_show_all_locks();
5596 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5598 idle->sched_class = &idle_sched_class;
5602 * init_idle - set up an idle thread for a given CPU
5603 * @idle: task in question
5604 * @cpu: cpu the idle task belongs to
5606 * NOTE: this function does not set the idle thread's NEED_RESCHED
5607 * flag, to make booting more robust.
5609 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5611 struct rq *rq = cpu_rq(cpu);
5612 unsigned long flags;
5614 raw_spin_lock_irqsave(&rq->lock, flags);
5617 idle->state = TASK_RUNNING;
5618 idle->se.exec_start = sched_clock();
5620 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5622 * We're having a chicken and egg problem, even though we are
5623 * holding rq->lock, the cpu isn't yet set to this cpu so the
5624 * lockdep check in task_group() will fail.
5626 * Similar case to sched_fork(). / Alternatively we could
5627 * use task_rq_lock() here and obtain the other rq->lock.
5632 __set_task_cpu(idle, cpu);
5635 rq->curr = rq->idle = idle;
5636 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5639 raw_spin_unlock_irqrestore(&rq->lock, flags);
5641 /* Set the preempt count _outside_ the spinlocks! */
5642 #if defined(CONFIG_PREEMPT)
5643 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5645 task_thread_info(idle)->preempt_count = 0;
5648 * The idle tasks have their own, simple scheduling class:
5650 idle->sched_class = &idle_sched_class;
5651 ftrace_graph_init_task(idle);
5655 * In a system that switches off the HZ timer nohz_cpu_mask
5656 * indicates which cpus entered this state. This is used
5657 * in the rcu update to wait only for active cpus. For system
5658 * which do not switch off the HZ timer nohz_cpu_mask should
5659 * always be CPU_BITS_NONE.
5661 cpumask_var_t nohz_cpu_mask;
5664 * Increase the granularity value when there are more CPUs,
5665 * because with more CPUs the 'effective latency' as visible
5666 * to users decreases. But the relationship is not linear,
5667 * so pick a second-best guess by going with the log2 of the
5670 * This idea comes from the SD scheduler of Con Kolivas:
5672 static int get_update_sysctl_factor(void)
5674 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5675 unsigned int factor;
5677 switch (sysctl_sched_tunable_scaling) {
5678 case SCHED_TUNABLESCALING_NONE:
5681 case SCHED_TUNABLESCALING_LINEAR:
5684 case SCHED_TUNABLESCALING_LOG:
5686 factor = 1 + ilog2(cpus);
5693 static void update_sysctl(void)
5695 unsigned int factor = get_update_sysctl_factor();
5697 #define SET_SYSCTL(name) \
5698 (sysctl_##name = (factor) * normalized_sysctl_##name)
5699 SET_SYSCTL(sched_min_granularity);
5700 SET_SYSCTL(sched_latency);
5701 SET_SYSCTL(sched_wakeup_granularity);
5702 SET_SYSCTL(sched_shares_ratelimit);
5706 static inline void sched_init_granularity(void)
5713 * This is how migration works:
5715 * 1) we invoke migration_cpu_stop() on the target CPU using
5717 * 2) stopper starts to run (implicitly forcing the migrated thread
5719 * 3) it checks whether the migrated task is still in the wrong runqueue.
5720 * 4) if it's in the wrong runqueue then the migration thread removes
5721 * it and puts it into the right queue.
5722 * 5) stopper completes and stop_one_cpu() returns and the migration
5727 * Change a given task's CPU affinity. Migrate the thread to a
5728 * proper CPU and schedule it away if the CPU it's executing on
5729 * is removed from the allowed bitmask.
5731 * NOTE: the caller must have a valid reference to the task, the
5732 * task must not exit() & deallocate itself prematurely. The
5733 * call is not atomic; no spinlocks may be held.
5735 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5737 unsigned long flags;
5739 unsigned int dest_cpu;
5743 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5744 * drop the rq->lock and still rely on ->cpus_allowed.
5747 while (task_is_waking(p))
5749 rq = task_rq_lock(p, &flags);
5750 if (task_is_waking(p)) {
5751 task_rq_unlock(rq, &flags);
5755 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5760 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5761 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5766 if (p->sched_class->set_cpus_allowed)
5767 p->sched_class->set_cpus_allowed(p, new_mask);
5769 cpumask_copy(&p->cpus_allowed, new_mask);
5770 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5773 /* Can the task run on the task's current CPU? If so, we're done */
5774 if (cpumask_test_cpu(task_cpu(p), new_mask))
5777 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5778 if (migrate_task(p, dest_cpu)) {
5779 struct migration_arg arg = { p, dest_cpu };
5780 /* Need help from migration thread: drop lock and wait. */
5781 task_rq_unlock(rq, &flags);
5782 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5783 tlb_migrate_finish(p->mm);
5787 task_rq_unlock(rq, &flags);
5791 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5794 * Move (not current) task off this cpu, onto dest cpu. We're doing
5795 * this because either it can't run here any more (set_cpus_allowed()
5796 * away from this CPU, or CPU going down), or because we're
5797 * attempting to rebalance this task on exec (sched_exec).
5799 * So we race with normal scheduler movements, but that's OK, as long
5800 * as the task is no longer on this CPU.
5802 * Returns non-zero if task was successfully migrated.
5804 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5806 struct rq *rq_dest, *rq_src;
5809 if (unlikely(!cpu_active(dest_cpu)))
5812 rq_src = cpu_rq(src_cpu);
5813 rq_dest = cpu_rq(dest_cpu);
5815 double_rq_lock(rq_src, rq_dest);
5816 /* Already moved. */
5817 if (task_cpu(p) != src_cpu)
5819 /* Affinity changed (again). */
5820 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5824 * If we're not on a rq, the next wake-up will ensure we're
5828 deactivate_task(rq_src, p, 0);
5829 set_task_cpu(p, dest_cpu);
5830 activate_task(rq_dest, p, 0);
5831 check_preempt_curr(rq_dest, p, 0);
5836 double_rq_unlock(rq_src, rq_dest);
5841 * migration_cpu_stop - this will be executed by a highprio stopper thread
5842 * and performs thread migration by bumping thread off CPU then
5843 * 'pushing' onto another runqueue.
5845 static int migration_cpu_stop(void *data)
5847 struct migration_arg *arg = data;
5850 * The original target cpu might have gone down and we might
5851 * be on another cpu but it doesn't matter.
5853 local_irq_disable();
5854 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5859 #ifdef CONFIG_HOTPLUG_CPU
5861 * Figure out where task on dead CPU should go, use force if necessary.
5863 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5865 struct rq *rq = cpu_rq(dead_cpu);
5866 int needs_cpu, uninitialized_var(dest_cpu);
5867 unsigned long flags;
5869 local_irq_save(flags);
5871 raw_spin_lock(&rq->lock);
5872 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5874 dest_cpu = select_fallback_rq(dead_cpu, p);
5875 raw_spin_unlock(&rq->lock);
5877 * It can only fail if we race with set_cpus_allowed(),
5878 * in the racer should migrate the task anyway.
5881 __migrate_task(p, dead_cpu, dest_cpu);
5882 local_irq_restore(flags);
5886 * While a dead CPU has no uninterruptible tasks queued at this point,
5887 * it might still have a nonzero ->nr_uninterruptible counter, because
5888 * for performance reasons the counter is not stricly tracking tasks to
5889 * their home CPUs. So we just add the counter to another CPU's counter,
5890 * to keep the global sum constant after CPU-down:
5892 static void migrate_nr_uninterruptible(struct rq *rq_src)
5894 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5895 unsigned long flags;
5897 local_irq_save(flags);
5898 double_rq_lock(rq_src, rq_dest);
5899 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5900 rq_src->nr_uninterruptible = 0;
5901 double_rq_unlock(rq_src, rq_dest);
5902 local_irq_restore(flags);
5905 /* Run through task list and migrate tasks from the dead cpu. */
5906 static void migrate_live_tasks(int src_cpu)
5908 struct task_struct *p, *t;
5910 read_lock(&tasklist_lock);
5912 do_each_thread(t, p) {
5916 if (task_cpu(p) == src_cpu)
5917 move_task_off_dead_cpu(src_cpu, p);
5918 } while_each_thread(t, p);
5920 read_unlock(&tasklist_lock);
5924 * Schedules idle task to be the next runnable task on current CPU.
5925 * It does so by boosting its priority to highest possible.
5926 * Used by CPU offline code.
5928 void sched_idle_next(void)
5930 int this_cpu = smp_processor_id();
5931 struct rq *rq = cpu_rq(this_cpu);
5932 struct task_struct *p = rq->idle;
5933 unsigned long flags;
5935 /* cpu has to be offline */
5936 BUG_ON(cpu_online(this_cpu));
5939 * Strictly not necessary since rest of the CPUs are stopped by now
5940 * and interrupts disabled on the current cpu.
5942 raw_spin_lock_irqsave(&rq->lock, flags);
5944 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5946 activate_task(rq, p, 0);
5948 raw_spin_unlock_irqrestore(&rq->lock, flags);
5952 * Ensures that the idle task is using init_mm right before its cpu goes
5955 void idle_task_exit(void)
5957 struct mm_struct *mm = current->active_mm;
5959 BUG_ON(cpu_online(smp_processor_id()));
5962 switch_mm(mm, &init_mm, current);
5966 /* called under rq->lock with disabled interrupts */
5967 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5969 struct rq *rq = cpu_rq(dead_cpu);
5971 /* Must be exiting, otherwise would be on tasklist. */
5972 BUG_ON(!p->exit_state);
5974 /* Cannot have done final schedule yet: would have vanished. */
5975 BUG_ON(p->state == TASK_DEAD);
5980 * Drop lock around migration; if someone else moves it,
5981 * that's OK. No task can be added to this CPU, so iteration is
5984 raw_spin_unlock_irq(&rq->lock);
5985 move_task_off_dead_cpu(dead_cpu, p);
5986 raw_spin_lock_irq(&rq->lock);
5991 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5992 static void migrate_dead_tasks(unsigned int dead_cpu)
5994 struct rq *rq = cpu_rq(dead_cpu);
5995 struct task_struct *next;
5998 if (!rq->nr_running)
6000 next = pick_next_task(rq);
6003 next->sched_class->put_prev_task(rq, next);
6004 migrate_dead(dead_cpu, next);
6010 * remove the tasks which were accounted by rq from calc_load_tasks.
6012 static void calc_global_load_remove(struct rq *rq)
6014 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6015 rq->calc_load_active = 0;
6017 #endif /* CONFIG_HOTPLUG_CPU */
6019 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6021 static struct ctl_table sd_ctl_dir[] = {
6023 .procname = "sched_domain",
6029 static struct ctl_table sd_ctl_root[] = {
6031 .procname = "kernel",
6033 .child = sd_ctl_dir,
6038 static struct ctl_table *sd_alloc_ctl_entry(int n)
6040 struct ctl_table *entry =
6041 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6046 static void sd_free_ctl_entry(struct ctl_table **tablep)
6048 struct ctl_table *entry;
6051 * In the intermediate directories, both the child directory and
6052 * procname are dynamically allocated and could fail but the mode
6053 * will always be set. In the lowest directory the names are
6054 * static strings and all have proc handlers.
6056 for (entry = *tablep; entry->mode; entry++) {
6058 sd_free_ctl_entry(&entry->child);
6059 if (entry->proc_handler == NULL)
6060 kfree(entry->procname);
6068 set_table_entry(struct ctl_table *entry,
6069 const char *procname, void *data, int maxlen,
6070 mode_t mode, proc_handler *proc_handler)
6072 entry->procname = procname;
6074 entry->maxlen = maxlen;
6076 entry->proc_handler = proc_handler;
6079 static struct ctl_table *
6080 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6082 struct ctl_table *table = sd_alloc_ctl_entry(13);
6087 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6088 sizeof(long), 0644, proc_doulongvec_minmax);
6089 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6090 sizeof(long), 0644, proc_doulongvec_minmax);
6091 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6092 sizeof(int), 0644, proc_dointvec_minmax);
6093 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6094 sizeof(int), 0644, proc_dointvec_minmax);
6095 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6096 sizeof(int), 0644, proc_dointvec_minmax);
6097 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6098 sizeof(int), 0644, proc_dointvec_minmax);
6099 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6100 sizeof(int), 0644, proc_dointvec_minmax);
6101 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6102 sizeof(int), 0644, proc_dointvec_minmax);
6103 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6104 sizeof(int), 0644, proc_dointvec_minmax);
6105 set_table_entry(&table[9], "cache_nice_tries",
6106 &sd->cache_nice_tries,
6107 sizeof(int), 0644, proc_dointvec_minmax);
6108 set_table_entry(&table[10], "flags", &sd->flags,
6109 sizeof(int), 0644, proc_dointvec_minmax);
6110 set_table_entry(&table[11], "name", sd->name,
6111 CORENAME_MAX_SIZE, 0444, proc_dostring);
6112 /* &table[12] is terminator */
6117 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6119 struct ctl_table *entry, *table;
6120 struct sched_domain *sd;
6121 int domain_num = 0, i;
6124 for_each_domain(cpu, sd)
6126 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6131 for_each_domain(cpu, sd) {
6132 snprintf(buf, 32, "domain%d", i);
6133 entry->procname = kstrdup(buf, GFP_KERNEL);
6135 entry->child = sd_alloc_ctl_domain_table(sd);
6142 static struct ctl_table_header *sd_sysctl_header;
6143 static void register_sched_domain_sysctl(void)
6145 int i, cpu_num = num_possible_cpus();
6146 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6149 WARN_ON(sd_ctl_dir[0].child);
6150 sd_ctl_dir[0].child = entry;
6155 for_each_possible_cpu(i) {
6156 snprintf(buf, 32, "cpu%d", i);
6157 entry->procname = kstrdup(buf, GFP_KERNEL);
6159 entry->child = sd_alloc_ctl_cpu_table(i);
6163 WARN_ON(sd_sysctl_header);
6164 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6167 /* may be called multiple times per register */
6168 static void unregister_sched_domain_sysctl(void)
6170 if (sd_sysctl_header)
6171 unregister_sysctl_table(sd_sysctl_header);
6172 sd_sysctl_header = NULL;
6173 if (sd_ctl_dir[0].child)
6174 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6177 static void register_sched_domain_sysctl(void)
6180 static void unregister_sched_domain_sysctl(void)
6185 static void set_rq_online(struct rq *rq)
6188 const struct sched_class *class;
6190 cpumask_set_cpu(rq->cpu, rq->rd->online);
6193 for_each_class(class) {
6194 if (class->rq_online)
6195 class->rq_online(rq);
6200 static void set_rq_offline(struct rq *rq)
6203 const struct sched_class *class;
6205 for_each_class(class) {
6206 if (class->rq_offline)
6207 class->rq_offline(rq);
6210 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6216 * migration_call - callback that gets triggered when a CPU is added.
6217 * Here we can start up the necessary migration thread for the new CPU.
6219 static int __cpuinit
6220 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6222 int cpu = (long)hcpu;
6223 unsigned long flags;
6224 struct rq *rq = cpu_rq(cpu);
6228 case CPU_UP_PREPARE:
6229 case CPU_UP_PREPARE_FROZEN:
6230 rq->calc_load_update = calc_load_update;
6234 case CPU_ONLINE_FROZEN:
6235 /* Update our root-domain */
6236 raw_spin_lock_irqsave(&rq->lock, flags);
6238 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6242 raw_spin_unlock_irqrestore(&rq->lock, flags);
6245 #ifdef CONFIG_HOTPLUG_CPU
6247 case CPU_DEAD_FROZEN:
6248 migrate_live_tasks(cpu);
6249 /* Idle task back to normal (off runqueue, low prio) */
6250 raw_spin_lock_irq(&rq->lock);
6251 deactivate_task(rq, rq->idle, 0);
6252 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6253 rq->idle->sched_class = &idle_sched_class;
6254 migrate_dead_tasks(cpu);
6255 raw_spin_unlock_irq(&rq->lock);
6256 migrate_nr_uninterruptible(rq);
6257 BUG_ON(rq->nr_running != 0);
6258 calc_global_load_remove(rq);
6262 case CPU_DYING_FROZEN:
6263 /* Update our root-domain */
6264 raw_spin_lock_irqsave(&rq->lock, flags);
6266 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6269 raw_spin_unlock_irqrestore(&rq->lock, flags);
6277 * Register at high priority so that task migration (migrate_all_tasks)
6278 * happens before everything else. This has to be lower priority than
6279 * the notifier in the perf_event subsystem, though.
6281 static struct notifier_block __cpuinitdata migration_notifier = {
6282 .notifier_call = migration_call,
6283 .priority = CPU_PRI_MIGRATION,
6286 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6287 unsigned long action, void *hcpu)
6289 switch (action & ~CPU_TASKS_FROZEN) {
6291 case CPU_DOWN_FAILED:
6292 set_cpu_active((long)hcpu, true);
6299 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6300 unsigned long action, void *hcpu)
6302 switch (action & ~CPU_TASKS_FROZEN) {
6303 case CPU_DOWN_PREPARE:
6304 set_cpu_active((long)hcpu, false);
6311 static int __init migration_init(void)
6313 void *cpu = (void *)(long)smp_processor_id();
6316 /* Initialize migration for the boot CPU */
6317 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6318 BUG_ON(err == NOTIFY_BAD);
6319 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6320 register_cpu_notifier(&migration_notifier);
6322 /* Register cpu active notifiers */
6323 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6324 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6328 early_initcall(migration_init);
6333 #ifdef CONFIG_SCHED_DEBUG
6335 static __read_mostly int sched_domain_debug_enabled;
6337 static int __init sched_domain_debug_setup(char *str)
6339 sched_domain_debug_enabled = 1;
6343 early_param("sched_debug", sched_domain_debug_setup);
6345 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6346 struct cpumask *groupmask)
6348 struct sched_group *group = sd->groups;
6351 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6352 cpumask_clear(groupmask);
6354 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6356 if (!(sd->flags & SD_LOAD_BALANCE)) {
6357 printk("does not load-balance\n");
6359 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6364 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6366 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6367 printk(KERN_ERR "ERROR: domain->span does not contain "
6370 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6371 printk(KERN_ERR "ERROR: domain->groups does not contain"
6375 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6379 printk(KERN_ERR "ERROR: group is NULL\n");
6383 if (!group->cpu_power) {
6384 printk(KERN_CONT "\n");
6385 printk(KERN_ERR "ERROR: domain->cpu_power not "
6390 if (!cpumask_weight(sched_group_cpus(group))) {
6391 printk(KERN_CONT "\n");
6392 printk(KERN_ERR "ERROR: empty group\n");
6396 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6397 printk(KERN_CONT "\n");
6398 printk(KERN_ERR "ERROR: repeated CPUs\n");
6402 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6404 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6406 printk(KERN_CONT " %s", str);
6407 if (group->cpu_power != SCHED_LOAD_SCALE) {
6408 printk(KERN_CONT " (cpu_power = %d)",
6412 group = group->next;
6413 } while (group != sd->groups);
6414 printk(KERN_CONT "\n");
6416 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6417 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6420 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6421 printk(KERN_ERR "ERROR: parent span is not a superset "
6422 "of domain->span\n");
6426 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6428 cpumask_var_t groupmask;
6431 if (!sched_domain_debug_enabled)
6435 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6439 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6441 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6442 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6447 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6454 free_cpumask_var(groupmask);
6456 #else /* !CONFIG_SCHED_DEBUG */
6457 # define sched_domain_debug(sd, cpu) do { } while (0)
6458 #endif /* CONFIG_SCHED_DEBUG */
6460 static int sd_degenerate(struct sched_domain *sd)
6462 if (cpumask_weight(sched_domain_span(sd)) == 1)
6465 /* Following flags need at least 2 groups */
6466 if (sd->flags & (SD_LOAD_BALANCE |
6467 SD_BALANCE_NEWIDLE |
6471 SD_SHARE_PKG_RESOURCES)) {
6472 if (sd->groups != sd->groups->next)
6476 /* Following flags don't use groups */
6477 if (sd->flags & (SD_WAKE_AFFINE))
6484 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6486 unsigned long cflags = sd->flags, pflags = parent->flags;
6488 if (sd_degenerate(parent))
6491 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6494 /* Flags needing groups don't count if only 1 group in parent */
6495 if (parent->groups == parent->groups->next) {
6496 pflags &= ~(SD_LOAD_BALANCE |
6497 SD_BALANCE_NEWIDLE |
6501 SD_SHARE_PKG_RESOURCES);
6502 if (nr_node_ids == 1)
6503 pflags &= ~SD_SERIALIZE;
6505 if (~cflags & pflags)
6511 static void free_rootdomain(struct root_domain *rd)
6513 synchronize_sched();
6515 cpupri_cleanup(&rd->cpupri);
6517 free_cpumask_var(rd->rto_mask);
6518 free_cpumask_var(rd->online);
6519 free_cpumask_var(rd->span);
6523 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6525 struct root_domain *old_rd = NULL;
6526 unsigned long flags;
6528 raw_spin_lock_irqsave(&rq->lock, flags);
6533 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6536 cpumask_clear_cpu(rq->cpu, old_rd->span);
6539 * If we dont want to free the old_rt yet then
6540 * set old_rd to NULL to skip the freeing later
6543 if (!atomic_dec_and_test(&old_rd->refcount))
6547 atomic_inc(&rd->refcount);
6550 cpumask_set_cpu(rq->cpu, rd->span);
6551 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6554 raw_spin_unlock_irqrestore(&rq->lock, flags);
6557 free_rootdomain(old_rd);
6560 static int init_rootdomain(struct root_domain *rd)
6562 memset(rd, 0, sizeof(*rd));
6564 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6566 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6568 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6571 if (cpupri_init(&rd->cpupri) != 0)
6576 free_cpumask_var(rd->rto_mask);
6578 free_cpumask_var(rd->online);
6580 free_cpumask_var(rd->span);
6585 static void init_defrootdomain(void)
6587 init_rootdomain(&def_root_domain);
6589 atomic_set(&def_root_domain.refcount, 1);
6592 static struct root_domain *alloc_rootdomain(void)
6594 struct root_domain *rd;
6596 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6600 if (init_rootdomain(rd) != 0) {
6609 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6610 * hold the hotplug lock.
6613 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6615 struct rq *rq = cpu_rq(cpu);
6616 struct sched_domain *tmp;
6618 for (tmp = sd; tmp; tmp = tmp->parent)
6619 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6621 /* Remove the sched domains which do not contribute to scheduling. */
6622 for (tmp = sd; tmp; ) {
6623 struct sched_domain *parent = tmp->parent;
6627 if (sd_parent_degenerate(tmp, parent)) {
6628 tmp->parent = parent->parent;
6630 parent->parent->child = tmp;
6635 if (sd && sd_degenerate(sd)) {
6641 sched_domain_debug(sd, cpu);
6643 rq_attach_root(rq, rd);
6644 rcu_assign_pointer(rq->sd, sd);
6647 /* cpus with isolated domains */
6648 static cpumask_var_t cpu_isolated_map;
6650 /* Setup the mask of cpus configured for isolated domains */
6651 static int __init isolated_cpu_setup(char *str)
6653 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6654 cpulist_parse(str, cpu_isolated_map);
6658 __setup("isolcpus=", isolated_cpu_setup);
6661 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6662 * to a function which identifies what group(along with sched group) a CPU
6663 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6664 * (due to the fact that we keep track of groups covered with a struct cpumask).
6666 * init_sched_build_groups will build a circular linked list of the groups
6667 * covered by the given span, and will set each group's ->cpumask correctly,
6668 * and ->cpu_power to 0.
6671 init_sched_build_groups(const struct cpumask *span,
6672 const struct cpumask *cpu_map,
6673 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6674 struct sched_group **sg,
6675 struct cpumask *tmpmask),
6676 struct cpumask *covered, struct cpumask *tmpmask)
6678 struct sched_group *first = NULL, *last = NULL;
6681 cpumask_clear(covered);
6683 for_each_cpu(i, span) {
6684 struct sched_group *sg;
6685 int group = group_fn(i, cpu_map, &sg, tmpmask);
6688 if (cpumask_test_cpu(i, covered))
6691 cpumask_clear(sched_group_cpus(sg));
6694 for_each_cpu(j, span) {
6695 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6698 cpumask_set_cpu(j, covered);
6699 cpumask_set_cpu(j, sched_group_cpus(sg));
6710 #define SD_NODES_PER_DOMAIN 16
6715 * find_next_best_node - find the next node to include in a sched_domain
6716 * @node: node whose sched_domain we're building
6717 * @used_nodes: nodes already in the sched_domain
6719 * Find the next node to include in a given scheduling domain. Simply
6720 * finds the closest node not already in the @used_nodes map.
6722 * Should use nodemask_t.
6724 static int find_next_best_node(int node, nodemask_t *used_nodes)
6726 int i, n, val, min_val, best_node = 0;
6730 for (i = 0; i < nr_node_ids; i++) {
6731 /* Start at @node */
6732 n = (node + i) % nr_node_ids;
6734 if (!nr_cpus_node(n))
6737 /* Skip already used nodes */
6738 if (node_isset(n, *used_nodes))
6741 /* Simple min distance search */
6742 val = node_distance(node, n);
6744 if (val < min_val) {
6750 node_set(best_node, *used_nodes);
6755 * sched_domain_node_span - get a cpumask for a node's sched_domain
6756 * @node: node whose cpumask we're constructing
6757 * @span: resulting cpumask
6759 * Given a node, construct a good cpumask for its sched_domain to span. It
6760 * should be one that prevents unnecessary balancing, but also spreads tasks
6763 static void sched_domain_node_span(int node, struct cpumask *span)
6765 nodemask_t used_nodes;
6768 cpumask_clear(span);
6769 nodes_clear(used_nodes);
6771 cpumask_or(span, span, cpumask_of_node(node));
6772 node_set(node, used_nodes);
6774 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6775 int next_node = find_next_best_node(node, &used_nodes);
6777 cpumask_or(span, span, cpumask_of_node(next_node));
6780 #endif /* CONFIG_NUMA */
6782 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6785 * The cpus mask in sched_group and sched_domain hangs off the end.
6787 * ( See the the comments in include/linux/sched.h:struct sched_group
6788 * and struct sched_domain. )
6790 struct static_sched_group {
6791 struct sched_group sg;
6792 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6795 struct static_sched_domain {
6796 struct sched_domain sd;
6797 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6803 cpumask_var_t domainspan;
6804 cpumask_var_t covered;
6805 cpumask_var_t notcovered;
6807 cpumask_var_t nodemask;
6808 cpumask_var_t this_sibling_map;
6809 cpumask_var_t this_core_map;
6810 cpumask_var_t this_book_map;
6811 cpumask_var_t send_covered;
6812 cpumask_var_t tmpmask;
6813 struct sched_group **sched_group_nodes;
6814 struct root_domain *rd;
6818 sa_sched_groups = 0,
6824 sa_this_sibling_map,
6826 sa_sched_group_nodes,
6836 * SMT sched-domains:
6838 #ifdef CONFIG_SCHED_SMT
6839 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6840 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6843 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6844 struct sched_group **sg, struct cpumask *unused)
6847 *sg = &per_cpu(sched_groups, cpu).sg;
6850 #endif /* CONFIG_SCHED_SMT */
6853 * multi-core sched-domains:
6855 #ifdef CONFIG_SCHED_MC
6856 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6857 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6860 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6861 struct sched_group **sg, struct cpumask *mask)
6864 #ifdef CONFIG_SCHED_SMT
6865 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6866 group = cpumask_first(mask);
6871 *sg = &per_cpu(sched_group_core, group).sg;
6874 #endif /* CONFIG_SCHED_MC */
6877 * book sched-domains:
6879 #ifdef CONFIG_SCHED_BOOK
6880 static DEFINE_PER_CPU(struct static_sched_domain, book_domains);
6881 static DEFINE_PER_CPU(struct static_sched_group, sched_group_book);
6884 cpu_to_book_group(int cpu, const struct cpumask *cpu_map,
6885 struct sched_group **sg, struct cpumask *mask)
6888 #ifdef CONFIG_SCHED_MC
6889 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6890 group = cpumask_first(mask);
6891 #elif defined(CONFIG_SCHED_SMT)
6892 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6893 group = cpumask_first(mask);
6896 *sg = &per_cpu(sched_group_book, group).sg;
6899 #endif /* CONFIG_SCHED_BOOK */
6901 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6902 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6905 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6906 struct sched_group **sg, struct cpumask *mask)
6909 #ifdef CONFIG_SCHED_BOOK
6910 cpumask_and(mask, cpu_book_mask(cpu), cpu_map);
6911 group = cpumask_first(mask);
6912 #elif defined(CONFIG_SCHED_MC)
6913 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6914 group = cpumask_first(mask);
6915 #elif defined(CONFIG_SCHED_SMT)
6916 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6917 group = cpumask_first(mask);
6922 *sg = &per_cpu(sched_group_phys, group).sg;
6928 * The init_sched_build_groups can't handle what we want to do with node
6929 * groups, so roll our own. Now each node has its own list of groups which
6930 * gets dynamically allocated.
6932 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6933 static struct sched_group ***sched_group_nodes_bycpu;
6935 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6936 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6938 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6939 struct sched_group **sg,
6940 struct cpumask *nodemask)
6944 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6945 group = cpumask_first(nodemask);
6948 *sg = &per_cpu(sched_group_allnodes, group).sg;
6952 static void init_numa_sched_groups_power(struct sched_group *group_head)
6954 struct sched_group *sg = group_head;
6960 for_each_cpu(j, sched_group_cpus(sg)) {
6961 struct sched_domain *sd;
6963 sd = &per_cpu(phys_domains, j).sd;
6964 if (j != group_first_cpu(sd->groups)) {
6966 * Only add "power" once for each
6972 sg->cpu_power += sd->groups->cpu_power;
6975 } while (sg != group_head);
6978 static int build_numa_sched_groups(struct s_data *d,
6979 const struct cpumask *cpu_map, int num)
6981 struct sched_domain *sd;
6982 struct sched_group *sg, *prev;
6985 cpumask_clear(d->covered);
6986 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6987 if (cpumask_empty(d->nodemask)) {
6988 d->sched_group_nodes[num] = NULL;
6992 sched_domain_node_span(num, d->domainspan);
6993 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6995 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6998 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
7002 d->sched_group_nodes[num] = sg;
7004 for_each_cpu(j, d->nodemask) {
7005 sd = &per_cpu(node_domains, j).sd;
7010 cpumask_copy(sched_group_cpus(sg), d->nodemask);
7012 cpumask_or(d->covered, d->covered, d->nodemask);
7015 for (j = 0; j < nr_node_ids; j++) {
7016 n = (num + j) % nr_node_ids;
7017 cpumask_complement(d->notcovered, d->covered);
7018 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
7019 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
7020 if (cpumask_empty(d->tmpmask))
7022 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
7023 if (cpumask_empty(d->tmpmask))
7025 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7029 "Can not alloc domain group for node %d\n", j);
7033 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
7034 sg->next = prev->next;
7035 cpumask_or(d->covered, d->covered, d->tmpmask);
7042 #endif /* CONFIG_NUMA */
7045 /* Free memory allocated for various sched_group structures */
7046 static void free_sched_groups(const struct cpumask *cpu_map,
7047 struct cpumask *nodemask)
7051 for_each_cpu(cpu, cpu_map) {
7052 struct sched_group **sched_group_nodes
7053 = sched_group_nodes_bycpu[cpu];
7055 if (!sched_group_nodes)
7058 for (i = 0; i < nr_node_ids; i++) {
7059 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7061 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7062 if (cpumask_empty(nodemask))
7072 if (oldsg != sched_group_nodes[i])
7075 kfree(sched_group_nodes);
7076 sched_group_nodes_bycpu[cpu] = NULL;
7079 #else /* !CONFIG_NUMA */
7080 static void free_sched_groups(const struct cpumask *cpu_map,
7081 struct cpumask *nodemask)
7084 #endif /* CONFIG_NUMA */
7087 * Initialize sched groups cpu_power.
7089 * cpu_power indicates the capacity of sched group, which is used while
7090 * distributing the load between different sched groups in a sched domain.
7091 * Typically cpu_power for all the groups in a sched domain will be same unless
7092 * there are asymmetries in the topology. If there are asymmetries, group
7093 * having more cpu_power will pickup more load compared to the group having
7096 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7098 struct sched_domain *child;
7099 struct sched_group *group;
7103 WARN_ON(!sd || !sd->groups);
7105 if (cpu != group_first_cpu(sd->groups))
7108 sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups));
7112 sd->groups->cpu_power = 0;
7115 power = SCHED_LOAD_SCALE;
7116 weight = cpumask_weight(sched_domain_span(sd));
7118 * SMT siblings share the power of a single core.
7119 * Usually multiple threads get a better yield out of
7120 * that one core than a single thread would have,
7121 * reflect that in sd->smt_gain.
7123 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
7124 power *= sd->smt_gain;
7126 power >>= SCHED_LOAD_SHIFT;
7128 sd->groups->cpu_power += power;
7133 * Add cpu_power of each child group to this groups cpu_power.
7135 group = child->groups;
7137 sd->groups->cpu_power += group->cpu_power;
7138 group = group->next;
7139 } while (group != child->groups);
7143 * Initializers for schedule domains
7144 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7147 #ifdef CONFIG_SCHED_DEBUG
7148 # define SD_INIT_NAME(sd, type) sd->name = #type
7150 # define SD_INIT_NAME(sd, type) do { } while (0)
7153 #define SD_INIT(sd, type) sd_init_##type(sd)
7155 #define SD_INIT_FUNC(type) \
7156 static noinline void sd_init_##type(struct sched_domain *sd) \
7158 memset(sd, 0, sizeof(*sd)); \
7159 *sd = SD_##type##_INIT; \
7160 sd->level = SD_LV_##type; \
7161 SD_INIT_NAME(sd, type); \
7166 SD_INIT_FUNC(ALLNODES)
7169 #ifdef CONFIG_SCHED_SMT
7170 SD_INIT_FUNC(SIBLING)
7172 #ifdef CONFIG_SCHED_MC
7175 #ifdef CONFIG_SCHED_BOOK
7179 static int default_relax_domain_level = -1;
7181 static int __init setup_relax_domain_level(char *str)
7185 val = simple_strtoul(str, NULL, 0);
7186 if (val < SD_LV_MAX)
7187 default_relax_domain_level = val;
7191 __setup("relax_domain_level=", setup_relax_domain_level);
7193 static void set_domain_attribute(struct sched_domain *sd,
7194 struct sched_domain_attr *attr)
7198 if (!attr || attr->relax_domain_level < 0) {
7199 if (default_relax_domain_level < 0)
7202 request = default_relax_domain_level;
7204 request = attr->relax_domain_level;
7205 if (request < sd->level) {
7206 /* turn off idle balance on this domain */
7207 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7209 /* turn on idle balance on this domain */
7210 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7214 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7215 const struct cpumask *cpu_map)
7218 case sa_sched_groups:
7219 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
7220 d->sched_group_nodes = NULL;
7222 free_rootdomain(d->rd); /* fall through */
7224 free_cpumask_var(d->tmpmask); /* fall through */
7225 case sa_send_covered:
7226 free_cpumask_var(d->send_covered); /* fall through */
7227 case sa_this_book_map:
7228 free_cpumask_var(d->this_book_map); /* fall through */
7229 case sa_this_core_map:
7230 free_cpumask_var(d->this_core_map); /* fall through */
7231 case sa_this_sibling_map:
7232 free_cpumask_var(d->this_sibling_map); /* fall through */
7234 free_cpumask_var(d->nodemask); /* fall through */
7235 case sa_sched_group_nodes:
7237 kfree(d->sched_group_nodes); /* fall through */
7239 free_cpumask_var(d->notcovered); /* fall through */
7241 free_cpumask_var(d->covered); /* fall through */
7243 free_cpumask_var(d->domainspan); /* fall through */
7250 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7251 const struct cpumask *cpu_map)
7254 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
7256 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
7257 return sa_domainspan;
7258 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
7260 /* Allocate the per-node list of sched groups */
7261 d->sched_group_nodes = kcalloc(nr_node_ids,
7262 sizeof(struct sched_group *), GFP_KERNEL);
7263 if (!d->sched_group_nodes) {
7264 printk(KERN_WARNING "Can not alloc sched group node list\n");
7265 return sa_notcovered;
7267 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
7269 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
7270 return sa_sched_group_nodes;
7271 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
7273 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
7274 return sa_this_sibling_map;
7275 if (!alloc_cpumask_var(&d->this_book_map, GFP_KERNEL))
7276 return sa_this_core_map;
7277 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
7278 return sa_this_book_map;
7279 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
7280 return sa_send_covered;
7281 d->rd = alloc_rootdomain();
7283 printk(KERN_WARNING "Cannot alloc root domain\n");
7286 return sa_rootdomain;
7289 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
7290 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
7292 struct sched_domain *sd = NULL;
7294 struct sched_domain *parent;
7297 if (cpumask_weight(cpu_map) >
7298 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
7299 sd = &per_cpu(allnodes_domains, i).sd;
7300 SD_INIT(sd, ALLNODES);
7301 set_domain_attribute(sd, attr);
7302 cpumask_copy(sched_domain_span(sd), cpu_map);
7303 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
7308 sd = &per_cpu(node_domains, i).sd;
7310 set_domain_attribute(sd, attr);
7311 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7312 sd->parent = parent;
7315 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7320 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7321 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7322 struct sched_domain *parent, int i)
7324 struct sched_domain *sd;
7325 sd = &per_cpu(phys_domains, i).sd;
7327 set_domain_attribute(sd, attr);
7328 cpumask_copy(sched_domain_span(sd), d->nodemask);
7329 sd->parent = parent;
7332 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7336 static struct sched_domain *__build_book_sched_domain(struct s_data *d,
7337 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7338 struct sched_domain *parent, int i)
7340 struct sched_domain *sd = parent;
7341 #ifdef CONFIG_SCHED_BOOK
7342 sd = &per_cpu(book_domains, i).sd;
7344 set_domain_attribute(sd, attr);
7345 cpumask_and(sched_domain_span(sd), cpu_map, cpu_book_mask(i));
7346 sd->parent = parent;
7348 cpu_to_book_group(i, cpu_map, &sd->groups, d->tmpmask);
7353 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7354 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7355 struct sched_domain *parent, int i)
7357 struct sched_domain *sd = parent;
7358 #ifdef CONFIG_SCHED_MC
7359 sd = &per_cpu(core_domains, i).sd;
7361 set_domain_attribute(sd, attr);
7362 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7363 sd->parent = parent;
7365 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7370 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7371 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7372 struct sched_domain *parent, int i)
7374 struct sched_domain *sd = parent;
7375 #ifdef CONFIG_SCHED_SMT
7376 sd = &per_cpu(cpu_domains, i).sd;
7377 SD_INIT(sd, SIBLING);
7378 set_domain_attribute(sd, attr);
7379 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7380 sd->parent = parent;
7382 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7387 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7388 const struct cpumask *cpu_map, int cpu)
7391 #ifdef CONFIG_SCHED_SMT
7392 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7393 cpumask_and(d->this_sibling_map, cpu_map,
7394 topology_thread_cpumask(cpu));
7395 if (cpu == cpumask_first(d->this_sibling_map))
7396 init_sched_build_groups(d->this_sibling_map, cpu_map,
7398 d->send_covered, d->tmpmask);
7401 #ifdef CONFIG_SCHED_MC
7402 case SD_LV_MC: /* set up multi-core groups */
7403 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7404 if (cpu == cpumask_first(d->this_core_map))
7405 init_sched_build_groups(d->this_core_map, cpu_map,
7407 d->send_covered, d->tmpmask);
7410 #ifdef CONFIG_SCHED_BOOK
7411 case SD_LV_BOOK: /* set up book groups */
7412 cpumask_and(d->this_book_map, cpu_map, cpu_book_mask(cpu));
7413 if (cpu == cpumask_first(d->this_book_map))
7414 init_sched_build_groups(d->this_book_map, cpu_map,
7416 d->send_covered, d->tmpmask);
7419 case SD_LV_CPU: /* set up physical groups */
7420 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7421 if (!cpumask_empty(d->nodemask))
7422 init_sched_build_groups(d->nodemask, cpu_map,
7424 d->send_covered, d->tmpmask);
7427 case SD_LV_ALLNODES:
7428 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7429 d->send_covered, d->tmpmask);
7438 * Build sched domains for a given set of cpus and attach the sched domains
7439 * to the individual cpus
7441 static int __build_sched_domains(const struct cpumask *cpu_map,
7442 struct sched_domain_attr *attr)
7444 enum s_alloc alloc_state = sa_none;
7446 struct sched_domain *sd;
7452 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7453 if (alloc_state != sa_rootdomain)
7455 alloc_state = sa_sched_groups;
7458 * Set up domains for cpus specified by the cpu_map.
7460 for_each_cpu(i, cpu_map) {
7461 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7464 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7465 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7466 sd = __build_book_sched_domain(&d, cpu_map, attr, sd, i);
7467 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7468 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7471 for_each_cpu(i, cpu_map) {
7472 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7473 build_sched_groups(&d, SD_LV_BOOK, cpu_map, i);
7474 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7477 /* Set up physical groups */
7478 for (i = 0; i < nr_node_ids; i++)
7479 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7482 /* Set up node groups */
7484 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7486 for (i = 0; i < nr_node_ids; i++)
7487 if (build_numa_sched_groups(&d, cpu_map, i))
7491 /* Calculate CPU power for physical packages and nodes */
7492 #ifdef CONFIG_SCHED_SMT
7493 for_each_cpu(i, cpu_map) {
7494 sd = &per_cpu(cpu_domains, i).sd;
7495 init_sched_groups_power(i, sd);
7498 #ifdef CONFIG_SCHED_MC
7499 for_each_cpu(i, cpu_map) {
7500 sd = &per_cpu(core_domains, i).sd;
7501 init_sched_groups_power(i, sd);
7504 #ifdef CONFIG_SCHED_BOOK
7505 for_each_cpu(i, cpu_map) {
7506 sd = &per_cpu(book_domains, i).sd;
7507 init_sched_groups_power(i, sd);
7511 for_each_cpu(i, cpu_map) {
7512 sd = &per_cpu(phys_domains, i).sd;
7513 init_sched_groups_power(i, sd);
7517 for (i = 0; i < nr_node_ids; i++)
7518 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7520 if (d.sd_allnodes) {
7521 struct sched_group *sg;
7523 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7525 init_numa_sched_groups_power(sg);
7529 /* Attach the domains */
7530 for_each_cpu(i, cpu_map) {
7531 #ifdef CONFIG_SCHED_SMT
7532 sd = &per_cpu(cpu_domains, i).sd;
7533 #elif defined(CONFIG_SCHED_MC)
7534 sd = &per_cpu(core_domains, i).sd;
7535 #elif defined(CONFIG_SCHED_BOOK)
7536 sd = &per_cpu(book_domains, i).sd;
7538 sd = &per_cpu(phys_domains, i).sd;
7540 cpu_attach_domain(sd, d.rd, i);
7543 d.sched_group_nodes = NULL; /* don't free this we still need it */
7544 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7548 __free_domain_allocs(&d, alloc_state, cpu_map);
7552 static int build_sched_domains(const struct cpumask *cpu_map)
7554 return __build_sched_domains(cpu_map, NULL);
7557 static cpumask_var_t *doms_cur; /* current sched domains */
7558 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7559 static struct sched_domain_attr *dattr_cur;
7560 /* attribues of custom domains in 'doms_cur' */
7563 * Special case: If a kmalloc of a doms_cur partition (array of
7564 * cpumask) fails, then fallback to a single sched domain,
7565 * as determined by the single cpumask fallback_doms.
7567 static cpumask_var_t fallback_doms;
7570 * arch_update_cpu_topology lets virtualized architectures update the
7571 * cpu core maps. It is supposed to return 1 if the topology changed
7572 * or 0 if it stayed the same.
7574 int __attribute__((weak)) arch_update_cpu_topology(void)
7579 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7582 cpumask_var_t *doms;
7584 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7587 for (i = 0; i < ndoms; i++) {
7588 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7589 free_sched_domains(doms, i);
7596 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7599 for (i = 0; i < ndoms; i++)
7600 free_cpumask_var(doms[i]);
7605 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7606 * For now this just excludes isolated cpus, but could be used to
7607 * exclude other special cases in the future.
7609 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7613 arch_update_cpu_topology();
7615 doms_cur = alloc_sched_domains(ndoms_cur);
7617 doms_cur = &fallback_doms;
7618 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7620 err = build_sched_domains(doms_cur[0]);
7621 register_sched_domain_sysctl();
7626 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7627 struct cpumask *tmpmask)
7629 free_sched_groups(cpu_map, tmpmask);
7633 * Detach sched domains from a group of cpus specified in cpu_map
7634 * These cpus will now be attached to the NULL domain
7636 static void detach_destroy_domains(const struct cpumask *cpu_map)
7638 /* Save because hotplug lock held. */
7639 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7642 for_each_cpu(i, cpu_map)
7643 cpu_attach_domain(NULL, &def_root_domain, i);
7644 synchronize_sched();
7645 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7648 /* handle null as "default" */
7649 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7650 struct sched_domain_attr *new, int idx_new)
7652 struct sched_domain_attr tmp;
7659 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7660 new ? (new + idx_new) : &tmp,
7661 sizeof(struct sched_domain_attr));
7665 * Partition sched domains as specified by the 'ndoms_new'
7666 * cpumasks in the array doms_new[] of cpumasks. This compares
7667 * doms_new[] to the current sched domain partitioning, doms_cur[].
7668 * It destroys each deleted domain and builds each new domain.
7670 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7671 * The masks don't intersect (don't overlap.) We should setup one
7672 * sched domain for each mask. CPUs not in any of the cpumasks will
7673 * not be load balanced. If the same cpumask appears both in the
7674 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7677 * The passed in 'doms_new' should be allocated using
7678 * alloc_sched_domains. This routine takes ownership of it and will
7679 * free_sched_domains it when done with it. If the caller failed the
7680 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7681 * and partition_sched_domains() will fallback to the single partition
7682 * 'fallback_doms', it also forces the domains to be rebuilt.
7684 * If doms_new == NULL it will be replaced with cpu_online_mask.
7685 * ndoms_new == 0 is a special case for destroying existing domains,
7686 * and it will not create the default domain.
7688 * Call with hotplug lock held
7690 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7691 struct sched_domain_attr *dattr_new)
7696 mutex_lock(&sched_domains_mutex);
7698 /* always unregister in case we don't destroy any domains */
7699 unregister_sched_domain_sysctl();
7701 /* Let architecture update cpu core mappings. */
7702 new_topology = arch_update_cpu_topology();
7704 n = doms_new ? ndoms_new : 0;
7706 /* Destroy deleted domains */
7707 for (i = 0; i < ndoms_cur; i++) {
7708 for (j = 0; j < n && !new_topology; j++) {
7709 if (cpumask_equal(doms_cur[i], doms_new[j])
7710 && dattrs_equal(dattr_cur, i, dattr_new, j))
7713 /* no match - a current sched domain not in new doms_new[] */
7714 detach_destroy_domains(doms_cur[i]);
7719 if (doms_new == NULL) {
7721 doms_new = &fallback_doms;
7722 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7723 WARN_ON_ONCE(dattr_new);
7726 /* Build new domains */
7727 for (i = 0; i < ndoms_new; i++) {
7728 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7729 if (cpumask_equal(doms_new[i], doms_cur[j])
7730 && dattrs_equal(dattr_new, i, dattr_cur, j))
7733 /* no match - add a new doms_new */
7734 __build_sched_domains(doms_new[i],
7735 dattr_new ? dattr_new + i : NULL);
7740 /* Remember the new sched domains */
7741 if (doms_cur != &fallback_doms)
7742 free_sched_domains(doms_cur, ndoms_cur);
7743 kfree(dattr_cur); /* kfree(NULL) is safe */
7744 doms_cur = doms_new;
7745 dattr_cur = dattr_new;
7746 ndoms_cur = ndoms_new;
7748 register_sched_domain_sysctl();
7750 mutex_unlock(&sched_domains_mutex);
7753 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7754 static void arch_reinit_sched_domains(void)
7758 /* Destroy domains first to force the rebuild */
7759 partition_sched_domains(0, NULL, NULL);
7761 rebuild_sched_domains();
7765 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7767 unsigned int level = 0;
7769 if (sscanf(buf, "%u", &level) != 1)
7773 * level is always be positive so don't check for
7774 * level < POWERSAVINGS_BALANCE_NONE which is 0
7775 * What happens on 0 or 1 byte write,
7776 * need to check for count as well?
7779 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7783 sched_smt_power_savings = level;
7785 sched_mc_power_savings = level;
7787 arch_reinit_sched_domains();
7792 #ifdef CONFIG_SCHED_MC
7793 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7794 struct sysdev_class_attribute *attr,
7797 return sprintf(page, "%u\n", sched_mc_power_savings);
7799 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7800 struct sysdev_class_attribute *attr,
7801 const char *buf, size_t count)
7803 return sched_power_savings_store(buf, count, 0);
7805 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7806 sched_mc_power_savings_show,
7807 sched_mc_power_savings_store);
7810 #ifdef CONFIG_SCHED_SMT
7811 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7812 struct sysdev_class_attribute *attr,
7815 return sprintf(page, "%u\n", sched_smt_power_savings);
7817 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7818 struct sysdev_class_attribute *attr,
7819 const char *buf, size_t count)
7821 return sched_power_savings_store(buf, count, 1);
7823 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7824 sched_smt_power_savings_show,
7825 sched_smt_power_savings_store);
7828 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7832 #ifdef CONFIG_SCHED_SMT
7834 err = sysfs_create_file(&cls->kset.kobj,
7835 &attr_sched_smt_power_savings.attr);
7837 #ifdef CONFIG_SCHED_MC
7838 if (!err && mc_capable())
7839 err = sysfs_create_file(&cls->kset.kobj,
7840 &attr_sched_mc_power_savings.attr);
7844 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7847 * Update cpusets according to cpu_active mask. If cpusets are
7848 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7849 * around partition_sched_domains().
7851 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7854 switch (action & ~CPU_TASKS_FROZEN) {
7856 case CPU_DOWN_FAILED:
7857 cpuset_update_active_cpus();
7864 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7867 switch (action & ~CPU_TASKS_FROZEN) {
7868 case CPU_DOWN_PREPARE:
7869 cpuset_update_active_cpus();
7876 static int update_runtime(struct notifier_block *nfb,
7877 unsigned long action, void *hcpu)
7879 int cpu = (int)(long)hcpu;
7882 case CPU_DOWN_PREPARE:
7883 case CPU_DOWN_PREPARE_FROZEN:
7884 disable_runtime(cpu_rq(cpu));
7887 case CPU_DOWN_FAILED:
7888 case CPU_DOWN_FAILED_FROZEN:
7890 case CPU_ONLINE_FROZEN:
7891 enable_runtime(cpu_rq(cpu));
7899 void __init sched_init_smp(void)
7901 cpumask_var_t non_isolated_cpus;
7903 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7904 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7906 #if defined(CONFIG_NUMA)
7907 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7909 BUG_ON(sched_group_nodes_bycpu == NULL);
7912 mutex_lock(&sched_domains_mutex);
7913 arch_init_sched_domains(cpu_active_mask);
7914 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7915 if (cpumask_empty(non_isolated_cpus))
7916 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7917 mutex_unlock(&sched_domains_mutex);
7920 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7921 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7923 /* RT runtime code needs to handle some hotplug events */
7924 hotcpu_notifier(update_runtime, 0);
7928 /* Move init over to a non-isolated CPU */
7929 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7931 sched_init_granularity();
7932 free_cpumask_var(non_isolated_cpus);
7934 init_sched_rt_class();
7937 void __init sched_init_smp(void)
7939 sched_init_granularity();
7941 #endif /* CONFIG_SMP */
7943 const_debug unsigned int sysctl_timer_migration = 1;
7945 int in_sched_functions(unsigned long addr)
7947 return in_lock_functions(addr) ||
7948 (addr >= (unsigned long)__sched_text_start
7949 && addr < (unsigned long)__sched_text_end);
7952 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7954 cfs_rq->tasks_timeline = RB_ROOT;
7955 INIT_LIST_HEAD(&cfs_rq->tasks);
7956 #ifdef CONFIG_FAIR_GROUP_SCHED
7959 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7962 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7964 struct rt_prio_array *array;
7967 array = &rt_rq->active;
7968 for (i = 0; i < MAX_RT_PRIO; i++) {
7969 INIT_LIST_HEAD(array->queue + i);
7970 __clear_bit(i, array->bitmap);
7972 /* delimiter for bitsearch: */
7973 __set_bit(MAX_RT_PRIO, array->bitmap);
7975 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7976 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7978 rt_rq->highest_prio.next = MAX_RT_PRIO;
7982 rt_rq->rt_nr_migratory = 0;
7983 rt_rq->overloaded = 0;
7984 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7988 rt_rq->rt_throttled = 0;
7989 rt_rq->rt_runtime = 0;
7990 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7992 #ifdef CONFIG_RT_GROUP_SCHED
7993 rt_rq->rt_nr_boosted = 0;
7998 #ifdef CONFIG_FAIR_GROUP_SCHED
7999 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8000 struct sched_entity *se, int cpu, int add,
8001 struct sched_entity *parent)
8003 struct rq *rq = cpu_rq(cpu);
8004 tg->cfs_rq[cpu] = cfs_rq;
8005 init_cfs_rq(cfs_rq, rq);
8008 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8011 /* se could be NULL for init_task_group */
8016 se->cfs_rq = &rq->cfs;
8018 se->cfs_rq = parent->my_q;
8021 se->load.weight = tg->shares;
8022 se->load.inv_weight = 0;
8023 se->parent = parent;
8027 #ifdef CONFIG_RT_GROUP_SCHED
8028 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8029 struct sched_rt_entity *rt_se, int cpu, int add,
8030 struct sched_rt_entity *parent)
8032 struct rq *rq = cpu_rq(cpu);
8034 tg->rt_rq[cpu] = rt_rq;
8035 init_rt_rq(rt_rq, rq);
8037 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8039 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8041 tg->rt_se[cpu] = rt_se;
8046 rt_se->rt_rq = &rq->rt;
8048 rt_se->rt_rq = parent->my_q;
8050 rt_se->my_q = rt_rq;
8051 rt_se->parent = parent;
8052 INIT_LIST_HEAD(&rt_se->run_list);
8056 void __init sched_init(void)
8059 unsigned long alloc_size = 0, ptr;
8061 #ifdef CONFIG_FAIR_GROUP_SCHED
8062 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8064 #ifdef CONFIG_RT_GROUP_SCHED
8065 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8067 #ifdef CONFIG_CPUMASK_OFFSTACK
8068 alloc_size += num_possible_cpus() * cpumask_size();
8071 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8073 #ifdef CONFIG_FAIR_GROUP_SCHED
8074 init_task_group.se = (struct sched_entity **)ptr;
8075 ptr += nr_cpu_ids * sizeof(void **);
8077 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8078 ptr += nr_cpu_ids * sizeof(void **);
8080 #endif /* CONFIG_FAIR_GROUP_SCHED */
8081 #ifdef CONFIG_RT_GROUP_SCHED
8082 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8083 ptr += nr_cpu_ids * sizeof(void **);
8085 init_task_group.rt_rq = (struct rt_rq **)ptr;
8086 ptr += nr_cpu_ids * sizeof(void **);
8088 #endif /* CONFIG_RT_GROUP_SCHED */
8089 #ifdef CONFIG_CPUMASK_OFFSTACK
8090 for_each_possible_cpu(i) {
8091 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8092 ptr += cpumask_size();
8094 #endif /* CONFIG_CPUMASK_OFFSTACK */
8098 init_defrootdomain();
8101 init_rt_bandwidth(&def_rt_bandwidth,
8102 global_rt_period(), global_rt_runtime());
8104 #ifdef CONFIG_RT_GROUP_SCHED
8105 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8106 global_rt_period(), global_rt_runtime());
8107 #endif /* CONFIG_RT_GROUP_SCHED */
8109 #ifdef CONFIG_CGROUP_SCHED
8110 list_add(&init_task_group.list, &task_groups);
8111 INIT_LIST_HEAD(&init_task_group.children);
8113 #endif /* CONFIG_CGROUP_SCHED */
8115 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
8116 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
8117 __alignof__(unsigned long));
8119 for_each_possible_cpu(i) {
8123 raw_spin_lock_init(&rq->lock);
8125 rq->calc_load_active = 0;
8126 rq->calc_load_update = jiffies + LOAD_FREQ;
8127 init_cfs_rq(&rq->cfs, rq);
8128 init_rt_rq(&rq->rt, rq);
8129 #ifdef CONFIG_FAIR_GROUP_SCHED
8130 init_task_group.shares = init_task_group_load;
8131 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8132 #ifdef CONFIG_CGROUP_SCHED
8134 * How much cpu bandwidth does init_task_group get?
8136 * In case of task-groups formed thr' the cgroup filesystem, it
8137 * gets 100% of the cpu resources in the system. This overall
8138 * system cpu resource is divided among the tasks of
8139 * init_task_group and its child task-groups in a fair manner,
8140 * based on each entity's (task or task-group's) weight
8141 * (se->load.weight).
8143 * In other words, if init_task_group has 10 tasks of weight
8144 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8145 * then A0's share of the cpu resource is:
8147 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8149 * We achieve this by letting init_task_group's tasks sit
8150 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8152 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8154 #endif /* CONFIG_FAIR_GROUP_SCHED */
8156 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8157 #ifdef CONFIG_RT_GROUP_SCHED
8158 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8159 #ifdef CONFIG_CGROUP_SCHED
8160 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8164 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8165 rq->cpu_load[j] = 0;
8167 rq->last_load_update_tick = jiffies;
8172 rq->cpu_power = SCHED_LOAD_SCALE;
8173 rq->post_schedule = 0;
8174 rq->active_balance = 0;
8175 rq->next_balance = jiffies;
8180 rq->avg_idle = 2*sysctl_sched_migration_cost;
8181 rq_attach_root(rq, &def_root_domain);
8183 rq->nohz_balance_kick = 0;
8184 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8188 atomic_set(&rq->nr_iowait, 0);
8191 set_load_weight(&init_task);
8193 #ifdef CONFIG_PREEMPT_NOTIFIERS
8194 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8198 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8201 #ifdef CONFIG_RT_MUTEXES
8202 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
8206 * The boot idle thread does lazy MMU switching as well:
8208 atomic_inc(&init_mm.mm_count);
8209 enter_lazy_tlb(&init_mm, current);
8212 * Make us the idle thread. Technically, schedule() should not be
8213 * called from this thread, however somewhere below it might be,
8214 * but because we are the idle thread, we just pick up running again
8215 * when this runqueue becomes "idle".
8217 init_idle(current, smp_processor_id());
8219 calc_load_update = jiffies + LOAD_FREQ;
8222 * During early bootup we pretend to be a normal task:
8224 current->sched_class = &fair_sched_class;
8226 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8227 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8230 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8231 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8232 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8233 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8234 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8236 /* May be allocated at isolcpus cmdline parse time */
8237 if (cpu_isolated_map == NULL)
8238 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8243 scheduler_running = 1;
8246 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8247 static inline int preempt_count_equals(int preempt_offset)
8249 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8251 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
8254 void __might_sleep(const char *file, int line, int preempt_offset)
8257 static unsigned long prev_jiffy; /* ratelimiting */
8259 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8260 system_state != SYSTEM_RUNNING || oops_in_progress)
8262 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8264 prev_jiffy = jiffies;
8267 "BUG: sleeping function called from invalid context at %s:%d\n",
8270 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8271 in_atomic(), irqs_disabled(),
8272 current->pid, current->comm);
8274 debug_show_held_locks(current);
8275 if (irqs_disabled())
8276 print_irqtrace_events(current);
8280 EXPORT_SYMBOL(__might_sleep);
8283 #ifdef CONFIG_MAGIC_SYSRQ
8284 static void normalize_task(struct rq *rq, struct task_struct *p)
8288 on_rq = p->se.on_rq;
8290 deactivate_task(rq, p, 0);
8291 __setscheduler(rq, p, SCHED_NORMAL, 0);
8293 activate_task(rq, p, 0);
8294 resched_task(rq->curr);
8298 void normalize_rt_tasks(void)
8300 struct task_struct *g, *p;
8301 unsigned long flags;
8304 read_lock_irqsave(&tasklist_lock, flags);
8305 do_each_thread(g, p) {
8307 * Only normalize user tasks:
8312 p->se.exec_start = 0;
8313 #ifdef CONFIG_SCHEDSTATS
8314 p->se.statistics.wait_start = 0;
8315 p->se.statistics.sleep_start = 0;
8316 p->se.statistics.block_start = 0;
8321 * Renice negative nice level userspace
8324 if (TASK_NICE(p) < 0 && p->mm)
8325 set_user_nice(p, 0);
8329 raw_spin_lock(&p->pi_lock);
8330 rq = __task_rq_lock(p);
8332 normalize_task(rq, p);
8334 __task_rq_unlock(rq);
8335 raw_spin_unlock(&p->pi_lock);
8336 } while_each_thread(g, p);
8338 read_unlock_irqrestore(&tasklist_lock, flags);
8341 #endif /* CONFIG_MAGIC_SYSRQ */
8343 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8345 * These functions are only useful for the IA64 MCA handling, or kdb.
8347 * They can only be called when the whole system has been
8348 * stopped - every CPU needs to be quiescent, and no scheduling
8349 * activity can take place. Using them for anything else would
8350 * be a serious bug, and as a result, they aren't even visible
8351 * under any other configuration.
8355 * curr_task - return the current task for a given cpu.
8356 * @cpu: the processor in question.
8358 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8360 struct task_struct *curr_task(int cpu)
8362 return cpu_curr(cpu);
8365 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8369 * set_curr_task - set the current task for a given cpu.
8370 * @cpu: the processor in question.
8371 * @p: the task pointer to set.
8373 * Description: This function must only be used when non-maskable interrupts
8374 * are serviced on a separate stack. It allows the architecture to switch the
8375 * notion of the current task on a cpu in a non-blocking manner. This function
8376 * must be called with all CPU's synchronized, and interrupts disabled, the
8377 * and caller must save the original value of the current task (see
8378 * curr_task() above) and restore that value before reenabling interrupts and
8379 * re-starting the system.
8381 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8383 void set_curr_task(int cpu, struct task_struct *p)
8390 #ifdef CONFIG_FAIR_GROUP_SCHED
8391 static void free_fair_sched_group(struct task_group *tg)
8395 for_each_possible_cpu(i) {
8397 kfree(tg->cfs_rq[i]);
8407 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8409 struct cfs_rq *cfs_rq;
8410 struct sched_entity *se;
8414 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8417 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8421 tg->shares = NICE_0_LOAD;
8423 for_each_possible_cpu(i) {
8426 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8427 GFP_KERNEL, cpu_to_node(i));
8431 se = kzalloc_node(sizeof(struct sched_entity),
8432 GFP_KERNEL, cpu_to_node(i));
8436 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8447 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8449 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8450 &cpu_rq(cpu)->leaf_cfs_rq_list);
8453 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8455 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8457 #else /* !CONFG_FAIR_GROUP_SCHED */
8458 static inline void free_fair_sched_group(struct task_group *tg)
8463 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8468 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8472 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8475 #endif /* CONFIG_FAIR_GROUP_SCHED */
8477 #ifdef CONFIG_RT_GROUP_SCHED
8478 static void free_rt_sched_group(struct task_group *tg)
8482 destroy_rt_bandwidth(&tg->rt_bandwidth);
8484 for_each_possible_cpu(i) {
8486 kfree(tg->rt_rq[i]);
8488 kfree(tg->rt_se[i]);
8496 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8498 struct rt_rq *rt_rq;
8499 struct sched_rt_entity *rt_se;
8503 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8506 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8510 init_rt_bandwidth(&tg->rt_bandwidth,
8511 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8513 for_each_possible_cpu(i) {
8516 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8517 GFP_KERNEL, cpu_to_node(i));
8521 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8522 GFP_KERNEL, cpu_to_node(i));
8526 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8537 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8539 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8540 &cpu_rq(cpu)->leaf_rt_rq_list);
8543 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8545 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8547 #else /* !CONFIG_RT_GROUP_SCHED */
8548 static inline void free_rt_sched_group(struct task_group *tg)
8553 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8558 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8562 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8565 #endif /* CONFIG_RT_GROUP_SCHED */
8567 #ifdef CONFIG_CGROUP_SCHED
8568 static void free_sched_group(struct task_group *tg)
8570 free_fair_sched_group(tg);
8571 free_rt_sched_group(tg);
8575 /* allocate runqueue etc for a new task group */
8576 struct task_group *sched_create_group(struct task_group *parent)
8578 struct task_group *tg;
8579 unsigned long flags;
8582 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8584 return ERR_PTR(-ENOMEM);
8586 if (!alloc_fair_sched_group(tg, parent))
8589 if (!alloc_rt_sched_group(tg, parent))
8592 spin_lock_irqsave(&task_group_lock, flags);
8593 for_each_possible_cpu(i) {
8594 register_fair_sched_group(tg, i);
8595 register_rt_sched_group(tg, i);
8597 list_add_rcu(&tg->list, &task_groups);
8599 WARN_ON(!parent); /* root should already exist */
8601 tg->parent = parent;
8602 INIT_LIST_HEAD(&tg->children);
8603 list_add_rcu(&tg->siblings, &parent->children);
8604 spin_unlock_irqrestore(&task_group_lock, flags);
8609 free_sched_group(tg);
8610 return ERR_PTR(-ENOMEM);
8613 /* rcu callback to free various structures associated with a task group */
8614 static void free_sched_group_rcu(struct rcu_head *rhp)
8616 /* now it should be safe to free those cfs_rqs */
8617 free_sched_group(container_of(rhp, struct task_group, rcu));
8620 /* Destroy runqueue etc associated with a task group */
8621 void sched_destroy_group(struct task_group *tg)
8623 unsigned long flags;
8626 spin_lock_irqsave(&task_group_lock, flags);
8627 for_each_possible_cpu(i) {
8628 unregister_fair_sched_group(tg, i);
8629 unregister_rt_sched_group(tg, i);
8631 list_del_rcu(&tg->list);
8632 list_del_rcu(&tg->siblings);
8633 spin_unlock_irqrestore(&task_group_lock, flags);
8635 /* wait for possible concurrent references to cfs_rqs complete */
8636 call_rcu(&tg->rcu, free_sched_group_rcu);
8639 /* change task's runqueue when it moves between groups.
8640 * The caller of this function should have put the task in its new group
8641 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8642 * reflect its new group.
8644 void sched_move_task(struct task_struct *tsk)
8647 unsigned long flags;
8650 rq = task_rq_lock(tsk, &flags);
8652 running = task_current(rq, tsk);
8653 on_rq = tsk->se.on_rq;
8656 dequeue_task(rq, tsk, 0);
8657 if (unlikely(running))
8658 tsk->sched_class->put_prev_task(rq, tsk);
8660 #ifdef CONFIG_FAIR_GROUP_SCHED
8661 if (tsk->sched_class->task_move_group)
8662 tsk->sched_class->task_move_group(tsk, on_rq);
8665 set_task_rq(tsk, task_cpu(tsk));
8667 if (unlikely(running))
8668 tsk->sched_class->set_curr_task(rq);
8670 enqueue_task(rq, tsk, 0);
8672 task_rq_unlock(rq, &flags);
8674 #endif /* CONFIG_CGROUP_SCHED */
8676 #ifdef CONFIG_FAIR_GROUP_SCHED
8677 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8679 struct cfs_rq *cfs_rq = se->cfs_rq;
8684 dequeue_entity(cfs_rq, se, 0);
8686 se->load.weight = shares;
8687 se->load.inv_weight = 0;
8690 enqueue_entity(cfs_rq, se, 0);
8693 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8695 struct cfs_rq *cfs_rq = se->cfs_rq;
8696 struct rq *rq = cfs_rq->rq;
8697 unsigned long flags;
8699 raw_spin_lock_irqsave(&rq->lock, flags);
8700 __set_se_shares(se, shares);
8701 raw_spin_unlock_irqrestore(&rq->lock, flags);
8704 static DEFINE_MUTEX(shares_mutex);
8706 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8709 unsigned long flags;
8712 * We can't change the weight of the root cgroup.
8717 if (shares < MIN_SHARES)
8718 shares = MIN_SHARES;
8719 else if (shares > MAX_SHARES)
8720 shares = MAX_SHARES;
8722 mutex_lock(&shares_mutex);
8723 if (tg->shares == shares)
8726 spin_lock_irqsave(&task_group_lock, flags);
8727 for_each_possible_cpu(i)
8728 unregister_fair_sched_group(tg, i);
8729 list_del_rcu(&tg->siblings);
8730 spin_unlock_irqrestore(&task_group_lock, flags);
8732 /* wait for any ongoing reference to this group to finish */
8733 synchronize_sched();
8736 * Now we are free to modify the group's share on each cpu
8737 * w/o tripping rebalance_share or load_balance_fair.
8739 tg->shares = shares;
8740 for_each_possible_cpu(i) {
8744 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8745 set_se_shares(tg->se[i], shares);
8749 * Enable load balance activity on this group, by inserting it back on
8750 * each cpu's rq->leaf_cfs_rq_list.
8752 spin_lock_irqsave(&task_group_lock, flags);
8753 for_each_possible_cpu(i)
8754 register_fair_sched_group(tg, i);
8755 list_add_rcu(&tg->siblings, &tg->parent->children);
8756 spin_unlock_irqrestore(&task_group_lock, flags);
8758 mutex_unlock(&shares_mutex);
8762 unsigned long sched_group_shares(struct task_group *tg)
8768 #ifdef CONFIG_RT_GROUP_SCHED
8770 * Ensure that the real time constraints are schedulable.
8772 static DEFINE_MUTEX(rt_constraints_mutex);
8774 static unsigned long to_ratio(u64 period, u64 runtime)
8776 if (runtime == RUNTIME_INF)
8779 return div64_u64(runtime << 20, period);
8782 /* Must be called with tasklist_lock held */
8783 static inline int tg_has_rt_tasks(struct task_group *tg)
8785 struct task_struct *g, *p;
8787 do_each_thread(g, p) {
8788 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8790 } while_each_thread(g, p);
8795 struct rt_schedulable_data {
8796 struct task_group *tg;
8801 static int tg_schedulable(struct task_group *tg, void *data)
8803 struct rt_schedulable_data *d = data;
8804 struct task_group *child;
8805 unsigned long total, sum = 0;
8806 u64 period, runtime;
8808 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8809 runtime = tg->rt_bandwidth.rt_runtime;
8812 period = d->rt_period;
8813 runtime = d->rt_runtime;
8817 * Cannot have more runtime than the period.
8819 if (runtime > period && runtime != RUNTIME_INF)
8823 * Ensure we don't starve existing RT tasks.
8825 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8828 total = to_ratio(period, runtime);
8831 * Nobody can have more than the global setting allows.
8833 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8837 * The sum of our children's runtime should not exceed our own.
8839 list_for_each_entry_rcu(child, &tg->children, siblings) {
8840 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8841 runtime = child->rt_bandwidth.rt_runtime;
8843 if (child == d->tg) {
8844 period = d->rt_period;
8845 runtime = d->rt_runtime;
8848 sum += to_ratio(period, runtime);
8857 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8859 struct rt_schedulable_data data = {
8861 .rt_period = period,
8862 .rt_runtime = runtime,
8865 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8868 static int tg_set_bandwidth(struct task_group *tg,
8869 u64 rt_period, u64 rt_runtime)
8873 mutex_lock(&rt_constraints_mutex);
8874 read_lock(&tasklist_lock);
8875 err = __rt_schedulable(tg, rt_period, rt_runtime);
8879 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8880 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8881 tg->rt_bandwidth.rt_runtime = rt_runtime;
8883 for_each_possible_cpu(i) {
8884 struct rt_rq *rt_rq = tg->rt_rq[i];
8886 raw_spin_lock(&rt_rq->rt_runtime_lock);
8887 rt_rq->rt_runtime = rt_runtime;
8888 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8890 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8892 read_unlock(&tasklist_lock);
8893 mutex_unlock(&rt_constraints_mutex);
8898 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8900 u64 rt_runtime, rt_period;
8902 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8903 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8904 if (rt_runtime_us < 0)
8905 rt_runtime = RUNTIME_INF;
8907 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8910 long sched_group_rt_runtime(struct task_group *tg)
8914 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8917 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8918 do_div(rt_runtime_us, NSEC_PER_USEC);
8919 return rt_runtime_us;
8922 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8924 u64 rt_runtime, rt_period;
8926 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8927 rt_runtime = tg->rt_bandwidth.rt_runtime;
8932 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8935 long sched_group_rt_period(struct task_group *tg)
8939 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8940 do_div(rt_period_us, NSEC_PER_USEC);
8941 return rt_period_us;
8944 static int sched_rt_global_constraints(void)
8946 u64 runtime, period;
8949 if (sysctl_sched_rt_period <= 0)
8952 runtime = global_rt_runtime();
8953 period = global_rt_period();
8956 * Sanity check on the sysctl variables.
8958 if (runtime > period && runtime != RUNTIME_INF)
8961 mutex_lock(&rt_constraints_mutex);
8962 read_lock(&tasklist_lock);
8963 ret = __rt_schedulable(NULL, 0, 0);
8964 read_unlock(&tasklist_lock);
8965 mutex_unlock(&rt_constraints_mutex);
8970 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8972 /* Don't accept realtime tasks when there is no way for them to run */
8973 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8979 #else /* !CONFIG_RT_GROUP_SCHED */
8980 static int sched_rt_global_constraints(void)
8982 unsigned long flags;
8985 if (sysctl_sched_rt_period <= 0)
8989 * There's always some RT tasks in the root group
8990 * -- migration, kstopmachine etc..
8992 if (sysctl_sched_rt_runtime == 0)
8995 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8996 for_each_possible_cpu(i) {
8997 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8999 raw_spin_lock(&rt_rq->rt_runtime_lock);
9000 rt_rq->rt_runtime = global_rt_runtime();
9001 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9003 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9007 #endif /* CONFIG_RT_GROUP_SCHED */
9009 int sched_rt_handler(struct ctl_table *table, int write,
9010 void __user *buffer, size_t *lenp,
9014 int old_period, old_runtime;
9015 static DEFINE_MUTEX(mutex);
9018 old_period = sysctl_sched_rt_period;
9019 old_runtime = sysctl_sched_rt_runtime;
9021 ret = proc_dointvec(table, write, buffer, lenp, ppos);
9023 if (!ret && write) {
9024 ret = sched_rt_global_constraints();
9026 sysctl_sched_rt_period = old_period;
9027 sysctl_sched_rt_runtime = old_runtime;
9029 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9030 def_rt_bandwidth.rt_period =
9031 ns_to_ktime(global_rt_period());
9034 mutex_unlock(&mutex);
9039 #ifdef CONFIG_CGROUP_SCHED
9041 /* return corresponding task_group object of a cgroup */
9042 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9044 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9045 struct task_group, css);
9048 static struct cgroup_subsys_state *
9049 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9051 struct task_group *tg, *parent;
9053 if (!cgrp->parent) {
9054 /* This is early initialization for the top cgroup */
9055 return &init_task_group.css;
9058 parent = cgroup_tg(cgrp->parent);
9059 tg = sched_create_group(parent);
9061 return ERR_PTR(-ENOMEM);
9067 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9069 struct task_group *tg = cgroup_tg(cgrp);
9071 sched_destroy_group(tg);
9075 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9077 #ifdef CONFIG_RT_GROUP_SCHED
9078 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9081 /* We don't support RT-tasks being in separate groups */
9082 if (tsk->sched_class != &fair_sched_class)
9089 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9090 struct task_struct *tsk, bool threadgroup)
9092 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
9096 struct task_struct *c;
9098 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
9099 retval = cpu_cgroup_can_attach_task(cgrp, c);
9111 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9112 struct cgroup *old_cont, struct task_struct *tsk,
9115 sched_move_task(tsk);
9117 struct task_struct *c;
9119 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
9126 #ifdef CONFIG_FAIR_GROUP_SCHED
9127 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9130 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9133 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9135 struct task_group *tg = cgroup_tg(cgrp);
9137 return (u64) tg->shares;
9139 #endif /* CONFIG_FAIR_GROUP_SCHED */
9141 #ifdef CONFIG_RT_GROUP_SCHED
9142 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9145 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9148 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9150 return sched_group_rt_runtime(cgroup_tg(cgrp));
9153 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9156 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9159 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9161 return sched_group_rt_period(cgroup_tg(cgrp));
9163 #endif /* CONFIG_RT_GROUP_SCHED */
9165 static struct cftype cpu_files[] = {
9166 #ifdef CONFIG_FAIR_GROUP_SCHED
9169 .read_u64 = cpu_shares_read_u64,
9170 .write_u64 = cpu_shares_write_u64,
9173 #ifdef CONFIG_RT_GROUP_SCHED
9175 .name = "rt_runtime_us",
9176 .read_s64 = cpu_rt_runtime_read,
9177 .write_s64 = cpu_rt_runtime_write,
9180 .name = "rt_period_us",
9181 .read_u64 = cpu_rt_period_read_uint,
9182 .write_u64 = cpu_rt_period_write_uint,
9187 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9189 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9192 struct cgroup_subsys cpu_cgroup_subsys = {
9194 .create = cpu_cgroup_create,
9195 .destroy = cpu_cgroup_destroy,
9196 .can_attach = cpu_cgroup_can_attach,
9197 .attach = cpu_cgroup_attach,
9198 .populate = cpu_cgroup_populate,
9199 .subsys_id = cpu_cgroup_subsys_id,
9203 #endif /* CONFIG_CGROUP_SCHED */
9205 #ifdef CONFIG_CGROUP_CPUACCT
9208 * CPU accounting code for task groups.
9210 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9211 * (balbir@in.ibm.com).
9214 /* track cpu usage of a group of tasks and its child groups */
9216 struct cgroup_subsys_state css;
9217 /* cpuusage holds pointer to a u64-type object on every cpu */
9218 u64 __percpu *cpuusage;
9219 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9220 struct cpuacct *parent;
9223 struct cgroup_subsys cpuacct_subsys;
9225 /* return cpu accounting group corresponding to this container */
9226 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9228 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9229 struct cpuacct, css);
9232 /* return cpu accounting group to which this task belongs */
9233 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9235 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9236 struct cpuacct, css);
9239 /* create a new cpu accounting group */
9240 static struct cgroup_subsys_state *cpuacct_create(
9241 struct cgroup_subsys *ss, struct cgroup *cgrp)
9243 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9249 ca->cpuusage = alloc_percpu(u64);
9253 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9254 if (percpu_counter_init(&ca->cpustat[i], 0))
9255 goto out_free_counters;
9258 ca->parent = cgroup_ca(cgrp->parent);
9264 percpu_counter_destroy(&ca->cpustat[i]);
9265 free_percpu(ca->cpuusage);
9269 return ERR_PTR(-ENOMEM);
9272 /* destroy an existing cpu accounting group */
9274 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9276 struct cpuacct *ca = cgroup_ca(cgrp);
9279 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9280 percpu_counter_destroy(&ca->cpustat[i]);
9281 free_percpu(ca->cpuusage);
9285 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9287 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9290 #ifndef CONFIG_64BIT
9292 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9294 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9296 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9304 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9306 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9308 #ifndef CONFIG_64BIT
9310 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9312 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9314 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9320 /* return total cpu usage (in nanoseconds) of a group */
9321 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9323 struct cpuacct *ca = cgroup_ca(cgrp);
9324 u64 totalcpuusage = 0;
9327 for_each_present_cpu(i)
9328 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9330 return totalcpuusage;
9333 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9336 struct cpuacct *ca = cgroup_ca(cgrp);
9345 for_each_present_cpu(i)
9346 cpuacct_cpuusage_write(ca, i, 0);
9352 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9355 struct cpuacct *ca = cgroup_ca(cgroup);
9359 for_each_present_cpu(i) {
9360 percpu = cpuacct_cpuusage_read(ca, i);
9361 seq_printf(m, "%llu ", (unsigned long long) percpu);
9363 seq_printf(m, "\n");
9367 static const char *cpuacct_stat_desc[] = {
9368 [CPUACCT_STAT_USER] = "user",
9369 [CPUACCT_STAT_SYSTEM] = "system",
9372 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9373 struct cgroup_map_cb *cb)
9375 struct cpuacct *ca = cgroup_ca(cgrp);
9378 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9379 s64 val = percpu_counter_read(&ca->cpustat[i]);
9380 val = cputime64_to_clock_t(val);
9381 cb->fill(cb, cpuacct_stat_desc[i], val);
9386 static struct cftype files[] = {
9389 .read_u64 = cpuusage_read,
9390 .write_u64 = cpuusage_write,
9393 .name = "usage_percpu",
9394 .read_seq_string = cpuacct_percpu_seq_read,
9398 .read_map = cpuacct_stats_show,
9402 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9404 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9408 * charge this task's execution time to its accounting group.
9410 * called with rq->lock held.
9412 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9417 if (unlikely(!cpuacct_subsys.active))
9420 cpu = task_cpu(tsk);
9426 for (; ca; ca = ca->parent) {
9427 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9428 *cpuusage += cputime;
9435 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9436 * in cputime_t units. As a result, cpuacct_update_stats calls
9437 * percpu_counter_add with values large enough to always overflow the
9438 * per cpu batch limit causing bad SMP scalability.
9440 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9441 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9442 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9445 #define CPUACCT_BATCH \
9446 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9448 #define CPUACCT_BATCH 0
9452 * Charge the system/user time to the task's accounting group.
9454 static void cpuacct_update_stats(struct task_struct *tsk,
9455 enum cpuacct_stat_index idx, cputime_t val)
9458 int batch = CPUACCT_BATCH;
9460 if (unlikely(!cpuacct_subsys.active))
9467 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9473 struct cgroup_subsys cpuacct_subsys = {
9475 .create = cpuacct_create,
9476 .destroy = cpuacct_destroy,
9477 .populate = cpuacct_populate,
9478 .subsys_id = cpuacct_subsys_id,
9480 #endif /* CONFIG_CGROUP_CPUACCT */
9484 void synchronize_sched_expedited(void)
9488 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9490 #else /* #ifndef CONFIG_SMP */
9492 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
9494 static int synchronize_sched_expedited_cpu_stop(void *data)
9497 * There must be a full memory barrier on each affected CPU
9498 * between the time that try_stop_cpus() is called and the
9499 * time that it returns.
9501 * In the current initial implementation of cpu_stop, the
9502 * above condition is already met when the control reaches
9503 * this point and the following smp_mb() is not strictly
9504 * necessary. Do smp_mb() anyway for documentation and
9505 * robustness against future implementation changes.
9507 smp_mb(); /* See above comment block. */
9512 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9513 * approach to force grace period to end quickly. This consumes
9514 * significant time on all CPUs, and is thus not recommended for
9515 * any sort of common-case code.
9517 * Note that it is illegal to call this function while holding any
9518 * lock that is acquired by a CPU-hotplug notifier. Failing to
9519 * observe this restriction will result in deadlock.
9521 void synchronize_sched_expedited(void)
9523 int snap, trycount = 0;
9525 smp_mb(); /* ensure prior mod happens before capturing snap. */
9526 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
9528 while (try_stop_cpus(cpu_online_mask,
9529 synchronize_sched_expedited_cpu_stop,
9532 if (trycount++ < 10)
9533 udelay(trycount * num_online_cpus());
9535 synchronize_sched();
9538 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
9539 smp_mb(); /* ensure test happens before caller kfree */
9544 atomic_inc(&synchronize_sched_expedited_count);
9545 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9548 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9550 #endif /* #else #ifndef CONFIG_SMP */