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 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
566 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
569 * A queue event has occurred, and we're going to schedule. In
570 * this case, we can save a useless back to back clock update.
572 if (test_tsk_need_resched(p))
573 rq->skip_clock_update = 1;
576 static inline int cpu_of(struct rq *rq)
585 #define rcu_dereference_check_sched_domain(p) \
586 rcu_dereference_check((p), \
587 rcu_read_lock_sched_held() || \
588 lockdep_is_held(&sched_domains_mutex))
591 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
592 * See detach_destroy_domains: synchronize_sched for details.
594 * The domain tree of any CPU may only be accessed from within
595 * preempt-disabled sections.
597 #define for_each_domain(cpu, __sd) \
598 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
600 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
601 #define this_rq() (&__get_cpu_var(runqueues))
602 #define task_rq(p) cpu_rq(task_cpu(p))
603 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
604 #define raw_rq() (&__raw_get_cpu_var(runqueues))
606 #ifdef CONFIG_CGROUP_SCHED
609 * Return the group to which this tasks belongs.
611 * We use task_subsys_state_check() and extend the RCU verification
612 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
613 * holds that lock for each task it moves into the cgroup. Therefore
614 * by holding that lock, we pin the task to the current cgroup.
616 static inline struct task_group *task_group(struct task_struct *p)
618 struct cgroup_subsys_state *css;
620 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
621 lockdep_is_held(&task_rq(p)->lock));
622 return container_of(css, struct task_group, css);
625 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
626 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
628 #ifdef CONFIG_FAIR_GROUP_SCHED
629 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
630 p->se.parent = task_group(p)->se[cpu];
633 #ifdef CONFIG_RT_GROUP_SCHED
634 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
635 p->rt.parent = task_group(p)->rt_se[cpu];
639 #else /* CONFIG_CGROUP_SCHED */
641 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
642 static inline struct task_group *task_group(struct task_struct *p)
647 #endif /* CONFIG_CGROUP_SCHED */
649 static u64 irq_time_cpu(int cpu);
650 static void sched_irq_time_avg_update(struct rq *rq, u64 irq_time);
652 inline void update_rq_clock(struct rq *rq)
654 if (!rq->skip_clock_update) {
655 int cpu = cpu_of(rq);
658 rq->clock = sched_clock_cpu(cpu);
659 irq_time = irq_time_cpu(cpu);
660 if (rq->clock - irq_time > rq->clock_task)
661 rq->clock_task = rq->clock - irq_time;
663 sched_irq_time_avg_update(rq, irq_time);
668 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
670 #ifdef CONFIG_SCHED_DEBUG
671 # define const_debug __read_mostly
673 # define const_debug static const
678 * @cpu: the processor in question.
680 * Returns true if the current cpu runqueue is locked.
681 * This interface allows printk to be called with the runqueue lock
682 * held and know whether or not it is OK to wake up the klogd.
684 int runqueue_is_locked(int cpu)
686 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
690 * Debugging: various feature bits
693 #define SCHED_FEAT(name, enabled) \
694 __SCHED_FEAT_##name ,
697 #include "sched_features.h"
702 #define SCHED_FEAT(name, enabled) \
703 (1UL << __SCHED_FEAT_##name) * enabled |
705 const_debug unsigned int sysctl_sched_features =
706 #include "sched_features.h"
711 #ifdef CONFIG_SCHED_DEBUG
712 #define SCHED_FEAT(name, enabled) \
715 static __read_mostly char *sched_feat_names[] = {
716 #include "sched_features.h"
722 static int sched_feat_show(struct seq_file *m, void *v)
726 for (i = 0; sched_feat_names[i]; i++) {
727 if (!(sysctl_sched_features & (1UL << i)))
729 seq_printf(m, "%s ", sched_feat_names[i]);
737 sched_feat_write(struct file *filp, const char __user *ubuf,
738 size_t cnt, loff_t *ppos)
748 if (copy_from_user(&buf, ubuf, cnt))
754 if (strncmp(buf, "NO_", 3) == 0) {
759 for (i = 0; sched_feat_names[i]; i++) {
760 if (strcmp(cmp, sched_feat_names[i]) == 0) {
762 sysctl_sched_features &= ~(1UL << i);
764 sysctl_sched_features |= (1UL << i);
769 if (!sched_feat_names[i])
777 static int sched_feat_open(struct inode *inode, struct file *filp)
779 return single_open(filp, sched_feat_show, NULL);
782 static const struct file_operations sched_feat_fops = {
783 .open = sched_feat_open,
784 .write = sched_feat_write,
787 .release = single_release,
790 static __init int sched_init_debug(void)
792 debugfs_create_file("sched_features", 0644, NULL, NULL,
797 late_initcall(sched_init_debug);
801 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
804 * Number of tasks to iterate in a single balance run.
805 * Limited because this is done with IRQs disabled.
807 const_debug unsigned int sysctl_sched_nr_migrate = 32;
810 * ratelimit for updating the group shares.
813 unsigned int sysctl_sched_shares_ratelimit = 250000;
814 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
817 * Inject some fuzzyness into changing the per-cpu group shares
818 * this avoids remote rq-locks at the expense of fairness.
821 unsigned int sysctl_sched_shares_thresh = 4;
824 * period over which we average the RT time consumption, measured
829 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
832 * period over which we measure -rt task cpu usage in us.
835 unsigned int sysctl_sched_rt_period = 1000000;
837 static __read_mostly int scheduler_running;
840 * part of the period that we allow rt tasks to run in us.
843 int sysctl_sched_rt_runtime = 950000;
845 static inline u64 global_rt_period(void)
847 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
850 static inline u64 global_rt_runtime(void)
852 if (sysctl_sched_rt_runtime < 0)
855 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
858 #ifndef prepare_arch_switch
859 # define prepare_arch_switch(next) do { } while (0)
861 #ifndef finish_arch_switch
862 # define finish_arch_switch(prev) do { } while (0)
865 static inline int task_current(struct rq *rq, struct task_struct *p)
867 return rq->curr == p;
870 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
871 static inline int task_running(struct rq *rq, struct task_struct *p)
873 return task_current(rq, p);
876 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
880 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
882 #ifdef CONFIG_DEBUG_SPINLOCK
883 /* this is a valid case when another task releases the spinlock */
884 rq->lock.owner = current;
887 * If we are tracking spinlock dependencies then we have to
888 * fix up the runqueue lock - which gets 'carried over' from
891 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
893 raw_spin_unlock_irq(&rq->lock);
896 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
897 static inline int task_running(struct rq *rq, struct task_struct *p)
902 return task_current(rq, p);
906 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
910 * We can optimise this out completely for !SMP, because the
911 * SMP rebalancing from interrupt is the only thing that cares
916 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
917 raw_spin_unlock_irq(&rq->lock);
919 raw_spin_unlock(&rq->lock);
923 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
927 * After ->oncpu is cleared, the task can be moved to a different CPU.
928 * We must ensure this doesn't happen until the switch is completely
934 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
938 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
941 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
944 static inline int task_is_waking(struct task_struct *p)
946 return unlikely(p->state == TASK_WAKING);
950 * __task_rq_lock - lock the runqueue a given task resides on.
951 * Must be called interrupts disabled.
953 static inline struct rq *__task_rq_lock(struct task_struct *p)
960 raw_spin_lock(&rq->lock);
961 if (likely(rq == task_rq(p)))
963 raw_spin_unlock(&rq->lock);
968 * task_rq_lock - lock the runqueue a given task resides on and disable
969 * interrupts. Note the ordering: we can safely lookup the task_rq without
970 * explicitly disabling preemption.
972 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
978 local_irq_save(*flags);
980 raw_spin_lock(&rq->lock);
981 if (likely(rq == task_rq(p)))
983 raw_spin_unlock_irqrestore(&rq->lock, *flags);
987 static void __task_rq_unlock(struct rq *rq)
990 raw_spin_unlock(&rq->lock);
993 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
996 raw_spin_unlock_irqrestore(&rq->lock, *flags);
1000 * this_rq_lock - lock this runqueue and disable interrupts.
1002 static struct rq *this_rq_lock(void)
1003 __acquires(rq->lock)
1007 local_irq_disable();
1009 raw_spin_lock(&rq->lock);
1014 #ifdef CONFIG_SCHED_HRTICK
1016 * Use HR-timers to deliver accurate preemption points.
1018 * Its all a bit involved since we cannot program an hrt while holding the
1019 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1022 * When we get rescheduled we reprogram the hrtick_timer outside of the
1028 * - enabled by features
1029 * - hrtimer is actually high res
1031 static inline int hrtick_enabled(struct rq *rq)
1033 if (!sched_feat(HRTICK))
1035 if (!cpu_active(cpu_of(rq)))
1037 return hrtimer_is_hres_active(&rq->hrtick_timer);
1040 static void hrtick_clear(struct rq *rq)
1042 if (hrtimer_active(&rq->hrtick_timer))
1043 hrtimer_cancel(&rq->hrtick_timer);
1047 * High-resolution timer tick.
1048 * Runs from hardirq context with interrupts disabled.
1050 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1052 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1054 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1056 raw_spin_lock(&rq->lock);
1057 update_rq_clock(rq);
1058 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1059 raw_spin_unlock(&rq->lock);
1061 return HRTIMER_NORESTART;
1066 * called from hardirq (IPI) context
1068 static void __hrtick_start(void *arg)
1070 struct rq *rq = arg;
1072 raw_spin_lock(&rq->lock);
1073 hrtimer_restart(&rq->hrtick_timer);
1074 rq->hrtick_csd_pending = 0;
1075 raw_spin_unlock(&rq->lock);
1079 * Called to set the hrtick timer state.
1081 * called with rq->lock held and irqs disabled
1083 static void hrtick_start(struct rq *rq, u64 delay)
1085 struct hrtimer *timer = &rq->hrtick_timer;
1086 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1088 hrtimer_set_expires(timer, time);
1090 if (rq == this_rq()) {
1091 hrtimer_restart(timer);
1092 } else if (!rq->hrtick_csd_pending) {
1093 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1094 rq->hrtick_csd_pending = 1;
1099 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1101 int cpu = (int)(long)hcpu;
1104 case CPU_UP_CANCELED:
1105 case CPU_UP_CANCELED_FROZEN:
1106 case CPU_DOWN_PREPARE:
1107 case CPU_DOWN_PREPARE_FROZEN:
1109 case CPU_DEAD_FROZEN:
1110 hrtick_clear(cpu_rq(cpu));
1117 static __init void init_hrtick(void)
1119 hotcpu_notifier(hotplug_hrtick, 0);
1123 * Called to set the hrtick timer state.
1125 * called with rq->lock held and irqs disabled
1127 static void hrtick_start(struct rq *rq, u64 delay)
1129 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1130 HRTIMER_MODE_REL_PINNED, 0);
1133 static inline void init_hrtick(void)
1136 #endif /* CONFIG_SMP */
1138 static void init_rq_hrtick(struct rq *rq)
1141 rq->hrtick_csd_pending = 0;
1143 rq->hrtick_csd.flags = 0;
1144 rq->hrtick_csd.func = __hrtick_start;
1145 rq->hrtick_csd.info = rq;
1148 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1149 rq->hrtick_timer.function = hrtick;
1151 #else /* CONFIG_SCHED_HRTICK */
1152 static inline void hrtick_clear(struct rq *rq)
1156 static inline void init_rq_hrtick(struct rq *rq)
1160 static inline void init_hrtick(void)
1163 #endif /* CONFIG_SCHED_HRTICK */
1166 * resched_task - mark a task 'to be rescheduled now'.
1168 * On UP this means the setting of the need_resched flag, on SMP it
1169 * might also involve a cross-CPU call to trigger the scheduler on
1174 #ifndef tsk_is_polling
1175 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1178 static void resched_task(struct task_struct *p)
1182 assert_raw_spin_locked(&task_rq(p)->lock);
1184 if (test_tsk_need_resched(p))
1187 set_tsk_need_resched(p);
1190 if (cpu == smp_processor_id())
1193 /* NEED_RESCHED must be visible before we test polling */
1195 if (!tsk_is_polling(p))
1196 smp_send_reschedule(cpu);
1199 static void resched_cpu(int cpu)
1201 struct rq *rq = cpu_rq(cpu);
1202 unsigned long flags;
1204 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1206 resched_task(cpu_curr(cpu));
1207 raw_spin_unlock_irqrestore(&rq->lock, flags);
1212 * In the semi idle case, use the nearest busy cpu for migrating timers
1213 * from an idle cpu. This is good for power-savings.
1215 * We don't do similar optimization for completely idle system, as
1216 * selecting an idle cpu will add more delays to the timers than intended
1217 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1219 int get_nohz_timer_target(void)
1221 int cpu = smp_processor_id();
1223 struct sched_domain *sd;
1225 for_each_domain(cpu, sd) {
1226 for_each_cpu(i, sched_domain_span(sd))
1233 * When add_timer_on() enqueues a timer into the timer wheel of an
1234 * idle CPU then this timer might expire before the next timer event
1235 * which is scheduled to wake up that CPU. In case of a completely
1236 * idle system the next event might even be infinite time into the
1237 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1238 * leaves the inner idle loop so the newly added timer is taken into
1239 * account when the CPU goes back to idle and evaluates the timer
1240 * wheel for the next timer event.
1242 void wake_up_idle_cpu(int cpu)
1244 struct rq *rq = cpu_rq(cpu);
1246 if (cpu == smp_processor_id())
1250 * This is safe, as this function is called with the timer
1251 * wheel base lock of (cpu) held. When the CPU is on the way
1252 * to idle and has not yet set rq->curr to idle then it will
1253 * be serialized on the timer wheel base lock and take the new
1254 * timer into account automatically.
1256 if (rq->curr != rq->idle)
1260 * We can set TIF_RESCHED on the idle task of the other CPU
1261 * lockless. The worst case is that the other CPU runs the
1262 * idle task through an additional NOOP schedule()
1264 set_tsk_need_resched(rq->idle);
1266 /* NEED_RESCHED must be visible before we test polling */
1268 if (!tsk_is_polling(rq->idle))
1269 smp_send_reschedule(cpu);
1272 #endif /* CONFIG_NO_HZ */
1274 static u64 sched_avg_period(void)
1276 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1279 static void sched_avg_update(struct rq *rq)
1281 s64 period = sched_avg_period();
1283 while ((s64)(rq->clock - rq->age_stamp) > period) {
1285 * Inline assembly required to prevent the compiler
1286 * optimising this loop into a divmod call.
1287 * See __iter_div_u64_rem() for another example of this.
1289 asm("" : "+rm" (rq->age_stamp));
1290 rq->age_stamp += period;
1295 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1297 rq->rt_avg += rt_delta;
1298 sched_avg_update(rq);
1301 #else /* !CONFIG_SMP */
1302 static void resched_task(struct task_struct *p)
1304 assert_raw_spin_locked(&task_rq(p)->lock);
1305 set_tsk_need_resched(p);
1308 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1312 static void sched_avg_update(struct rq *rq)
1315 #endif /* CONFIG_SMP */
1317 #if BITS_PER_LONG == 32
1318 # define WMULT_CONST (~0UL)
1320 # define WMULT_CONST (1UL << 32)
1323 #define WMULT_SHIFT 32
1326 * Shift right and round:
1328 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1331 * delta *= weight / lw
1333 static unsigned long
1334 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1335 struct load_weight *lw)
1339 if (!lw->inv_weight) {
1340 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1343 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1347 tmp = (u64)delta_exec * weight;
1349 * Check whether we'd overflow the 64-bit multiplication:
1351 if (unlikely(tmp > WMULT_CONST))
1352 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1355 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1357 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1360 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1366 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1373 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1374 * of tasks with abnormal "nice" values across CPUs the contribution that
1375 * each task makes to its run queue's load is weighted according to its
1376 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1377 * scaled version of the new time slice allocation that they receive on time
1381 #define WEIGHT_IDLEPRIO 3
1382 #define WMULT_IDLEPRIO 1431655765
1385 * Nice levels are multiplicative, with a gentle 10% change for every
1386 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1387 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1388 * that remained on nice 0.
1390 * The "10% effect" is relative and cumulative: from _any_ nice level,
1391 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1392 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1393 * If a task goes up by ~10% and another task goes down by ~10% then
1394 * the relative distance between them is ~25%.)
1396 static const int prio_to_weight[40] = {
1397 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1398 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1399 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1400 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1401 /* 0 */ 1024, 820, 655, 526, 423,
1402 /* 5 */ 335, 272, 215, 172, 137,
1403 /* 10 */ 110, 87, 70, 56, 45,
1404 /* 15 */ 36, 29, 23, 18, 15,
1408 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1410 * In cases where the weight does not change often, we can use the
1411 * precalculated inverse to speed up arithmetics by turning divisions
1412 * into multiplications:
1414 static const u32 prio_to_wmult[40] = {
1415 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1416 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1417 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1418 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1419 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1420 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1421 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1422 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1425 /* Time spent by the tasks of the cpu accounting group executing in ... */
1426 enum cpuacct_stat_index {
1427 CPUACCT_STAT_USER, /* ... user mode */
1428 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1430 CPUACCT_STAT_NSTATS,
1433 #ifdef CONFIG_CGROUP_CPUACCT
1434 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1435 static void cpuacct_update_stats(struct task_struct *tsk,
1436 enum cpuacct_stat_index idx, cputime_t val);
1438 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1439 static inline void cpuacct_update_stats(struct task_struct *tsk,
1440 enum cpuacct_stat_index idx, cputime_t val) {}
1443 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1445 update_load_add(&rq->load, load);
1448 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1450 update_load_sub(&rq->load, load);
1453 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1454 typedef int (*tg_visitor)(struct task_group *, void *);
1457 * Iterate the full tree, calling @down when first entering a node and @up when
1458 * leaving it for the final time.
1460 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1462 struct task_group *parent, *child;
1466 parent = &root_task_group;
1468 ret = (*down)(parent, data);
1471 list_for_each_entry_rcu(child, &parent->children, siblings) {
1478 ret = (*up)(parent, data);
1483 parent = parent->parent;
1492 static int tg_nop(struct task_group *tg, void *data)
1499 /* Used instead of source_load when we know the type == 0 */
1500 static unsigned long weighted_cpuload(const int cpu)
1502 return cpu_rq(cpu)->load.weight;
1506 * Return a low guess at the load of a migration-source cpu weighted
1507 * according to the scheduling class and "nice" value.
1509 * We want to under-estimate the load of migration sources, to
1510 * balance conservatively.
1512 static unsigned long source_load(int cpu, int type)
1514 struct rq *rq = cpu_rq(cpu);
1515 unsigned long total = weighted_cpuload(cpu);
1517 if (type == 0 || !sched_feat(LB_BIAS))
1520 return min(rq->cpu_load[type-1], total);
1524 * Return a high guess at the load of a migration-target cpu weighted
1525 * according to the scheduling class and "nice" value.
1527 static unsigned long target_load(int cpu, int type)
1529 struct rq *rq = cpu_rq(cpu);
1530 unsigned long total = weighted_cpuload(cpu);
1532 if (type == 0 || !sched_feat(LB_BIAS))
1535 return max(rq->cpu_load[type-1], total);
1538 static unsigned long power_of(int cpu)
1540 return cpu_rq(cpu)->cpu_power;
1543 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1545 static unsigned long cpu_avg_load_per_task(int cpu)
1547 struct rq *rq = cpu_rq(cpu);
1548 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1551 rq->avg_load_per_task = rq->load.weight / nr_running;
1553 rq->avg_load_per_task = 0;
1555 return rq->avg_load_per_task;
1558 #ifdef CONFIG_FAIR_GROUP_SCHED
1560 static __read_mostly unsigned long __percpu *update_shares_data;
1562 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1565 * Calculate and set the cpu's group shares.
1567 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1568 unsigned long sd_shares,
1569 unsigned long sd_rq_weight,
1570 unsigned long *usd_rq_weight)
1572 unsigned long shares, rq_weight;
1575 rq_weight = usd_rq_weight[cpu];
1578 rq_weight = NICE_0_LOAD;
1582 * \Sum_j shares_j * rq_weight_i
1583 * shares_i = -----------------------------
1584 * \Sum_j rq_weight_j
1586 shares = (sd_shares * rq_weight) / sd_rq_weight;
1587 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1589 if (abs(shares - tg->se[cpu]->load.weight) >
1590 sysctl_sched_shares_thresh) {
1591 struct rq *rq = cpu_rq(cpu);
1592 unsigned long flags;
1594 raw_spin_lock_irqsave(&rq->lock, flags);
1595 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1596 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1597 __set_se_shares(tg->se[cpu], shares);
1598 raw_spin_unlock_irqrestore(&rq->lock, flags);
1603 * Re-compute the task group their per cpu shares over the given domain.
1604 * This needs to be done in a bottom-up fashion because the rq weight of a
1605 * parent group depends on the shares of its child groups.
1607 static int tg_shares_up(struct task_group *tg, void *data)
1609 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1610 unsigned long *usd_rq_weight;
1611 struct sched_domain *sd = data;
1612 unsigned long flags;
1618 local_irq_save(flags);
1619 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1621 for_each_cpu(i, sched_domain_span(sd)) {
1622 weight = tg->cfs_rq[i]->load.weight;
1623 usd_rq_weight[i] = weight;
1625 rq_weight += weight;
1627 * If there are currently no tasks on the cpu pretend there
1628 * is one of average load so that when a new task gets to
1629 * run here it will not get delayed by group starvation.
1632 weight = NICE_0_LOAD;
1634 sum_weight += weight;
1635 shares += tg->cfs_rq[i]->shares;
1639 rq_weight = sum_weight;
1641 if ((!shares && rq_weight) || shares > tg->shares)
1642 shares = tg->shares;
1644 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1645 shares = tg->shares;
1647 for_each_cpu(i, sched_domain_span(sd))
1648 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1650 local_irq_restore(flags);
1656 * Compute the cpu's hierarchical load factor for each task group.
1657 * This needs to be done in a top-down fashion because the load of a child
1658 * group is a fraction of its parents load.
1660 static int tg_load_down(struct task_group *tg, void *data)
1663 long cpu = (long)data;
1666 load = cpu_rq(cpu)->load.weight;
1668 load = tg->parent->cfs_rq[cpu]->h_load;
1669 load *= tg->cfs_rq[cpu]->shares;
1670 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1673 tg->cfs_rq[cpu]->h_load = load;
1678 static void update_shares(struct sched_domain *sd)
1683 if (root_task_group_empty())
1686 now = local_clock();
1687 elapsed = now - sd->last_update;
1689 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1690 sd->last_update = now;
1691 walk_tg_tree(tg_nop, tg_shares_up, sd);
1695 static void update_h_load(long cpu)
1697 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1702 static inline void update_shares(struct sched_domain *sd)
1708 #ifdef CONFIG_PREEMPT
1710 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1713 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1714 * way at the expense of forcing extra atomic operations in all
1715 * invocations. This assures that the double_lock is acquired using the
1716 * same underlying policy as the spinlock_t on this architecture, which
1717 * reduces latency compared to the unfair variant below. However, it
1718 * also adds more overhead and therefore may reduce throughput.
1720 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1721 __releases(this_rq->lock)
1722 __acquires(busiest->lock)
1723 __acquires(this_rq->lock)
1725 raw_spin_unlock(&this_rq->lock);
1726 double_rq_lock(this_rq, busiest);
1733 * Unfair double_lock_balance: Optimizes throughput at the expense of
1734 * latency by eliminating extra atomic operations when the locks are
1735 * already in proper order on entry. This favors lower cpu-ids and will
1736 * grant the double lock to lower cpus over higher ids under contention,
1737 * regardless of entry order into the function.
1739 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1740 __releases(this_rq->lock)
1741 __acquires(busiest->lock)
1742 __acquires(this_rq->lock)
1746 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1747 if (busiest < this_rq) {
1748 raw_spin_unlock(&this_rq->lock);
1749 raw_spin_lock(&busiest->lock);
1750 raw_spin_lock_nested(&this_rq->lock,
1751 SINGLE_DEPTH_NESTING);
1754 raw_spin_lock_nested(&busiest->lock,
1755 SINGLE_DEPTH_NESTING);
1760 #endif /* CONFIG_PREEMPT */
1763 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1765 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1767 if (unlikely(!irqs_disabled())) {
1768 /* printk() doesn't work good under rq->lock */
1769 raw_spin_unlock(&this_rq->lock);
1773 return _double_lock_balance(this_rq, busiest);
1776 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1777 __releases(busiest->lock)
1779 raw_spin_unlock(&busiest->lock);
1780 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1784 * double_rq_lock - safely lock two runqueues
1786 * Note this does not disable interrupts like task_rq_lock,
1787 * you need to do so manually before calling.
1789 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1790 __acquires(rq1->lock)
1791 __acquires(rq2->lock)
1793 BUG_ON(!irqs_disabled());
1795 raw_spin_lock(&rq1->lock);
1796 __acquire(rq2->lock); /* Fake it out ;) */
1799 raw_spin_lock(&rq1->lock);
1800 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1802 raw_spin_lock(&rq2->lock);
1803 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1809 * double_rq_unlock - safely unlock two runqueues
1811 * Note this does not restore interrupts like task_rq_unlock,
1812 * you need to do so manually after calling.
1814 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1815 __releases(rq1->lock)
1816 __releases(rq2->lock)
1818 raw_spin_unlock(&rq1->lock);
1820 raw_spin_unlock(&rq2->lock);
1822 __release(rq2->lock);
1827 #ifdef CONFIG_FAIR_GROUP_SCHED
1828 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1831 cfs_rq->shares = shares;
1836 static void calc_load_account_idle(struct rq *this_rq);
1837 static void update_sysctl(void);
1838 static int get_update_sysctl_factor(void);
1839 static void update_cpu_load(struct rq *this_rq);
1841 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1843 set_task_rq(p, cpu);
1846 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1847 * successfuly executed on another CPU. We must ensure that updates of
1848 * per-task data have been completed by this moment.
1851 task_thread_info(p)->cpu = cpu;
1855 static const struct sched_class rt_sched_class;
1857 #define sched_class_highest (&stop_sched_class)
1858 #define for_each_class(class) \
1859 for (class = sched_class_highest; class; class = class->next)
1861 #include "sched_stats.h"
1863 static void inc_nr_running(struct rq *rq)
1868 static void dec_nr_running(struct rq *rq)
1873 static void set_load_weight(struct task_struct *p)
1876 * SCHED_IDLE tasks get minimal weight:
1878 if (p->policy == SCHED_IDLE) {
1879 p->se.load.weight = WEIGHT_IDLEPRIO;
1880 p->se.load.inv_weight = WMULT_IDLEPRIO;
1884 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1885 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1888 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1890 update_rq_clock(rq);
1891 sched_info_queued(p);
1892 p->sched_class->enqueue_task(rq, p, flags);
1896 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1898 update_rq_clock(rq);
1899 sched_info_dequeued(p);
1900 p->sched_class->dequeue_task(rq, p, flags);
1905 * activate_task - move a task to the runqueue.
1907 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1909 if (task_contributes_to_load(p))
1910 rq->nr_uninterruptible--;
1912 enqueue_task(rq, p, flags);
1917 * deactivate_task - remove a task from the runqueue.
1919 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1921 if (task_contributes_to_load(p))
1922 rq->nr_uninterruptible++;
1924 dequeue_task(rq, p, flags);
1928 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1931 * There are no locks covering percpu hardirq/softirq time.
1932 * They are only modified in account_system_vtime, on corresponding CPU
1933 * with interrupts disabled. So, writes are safe.
1934 * They are read and saved off onto struct rq in update_rq_clock().
1935 * This may result in other CPU reading this CPU's irq time and can
1936 * race with irq/account_system_vtime on this CPU. We would either get old
1937 * or new value (or semi updated value on 32 bit) with a side effect of
1938 * accounting a slice of irq time to wrong task when irq is in progress
1939 * while we read rq->clock. That is a worthy compromise in place of having
1940 * locks on each irq in account_system_time.
1942 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1943 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1945 static DEFINE_PER_CPU(u64, irq_start_time);
1946 static int sched_clock_irqtime;
1948 void enable_sched_clock_irqtime(void)
1950 sched_clock_irqtime = 1;
1953 void disable_sched_clock_irqtime(void)
1955 sched_clock_irqtime = 0;
1958 static u64 irq_time_cpu(int cpu)
1960 if (!sched_clock_irqtime)
1963 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1966 void account_system_vtime(struct task_struct *curr)
1968 unsigned long flags;
1972 if (!sched_clock_irqtime)
1975 local_irq_save(flags);
1977 now = sched_clock();
1978 cpu = smp_processor_id();
1979 delta = now - per_cpu(irq_start_time, cpu);
1980 per_cpu(irq_start_time, cpu) = now;
1982 * We do not account for softirq time from ksoftirqd here.
1983 * We want to continue accounting softirq time to ksoftirqd thread
1984 * in that case, so as not to confuse scheduler with a special task
1985 * that do not consume any time, but still wants to run.
1987 if (hardirq_count())
1988 per_cpu(cpu_hardirq_time, cpu) += delta;
1989 else if (in_serving_softirq() && !(curr->flags & PF_KSOFTIRQD))
1990 per_cpu(cpu_softirq_time, cpu) += delta;
1992 local_irq_restore(flags);
1995 static void sched_irq_time_avg_update(struct rq *rq, u64 curr_irq_time)
1997 if (sched_clock_irqtime && sched_feat(NONIRQ_POWER)) {
1998 u64 delta_irq = curr_irq_time - rq->prev_irq_time;
1999 rq->prev_irq_time = curr_irq_time;
2000 sched_rt_avg_update(rq, delta_irq);
2006 static u64 irq_time_cpu(int cpu)
2011 static void sched_irq_time_avg_update(struct rq *rq, u64 curr_irq_time) { }
2015 #include "sched_idletask.c"
2016 #include "sched_fair.c"
2017 #include "sched_rt.c"
2018 #include "sched_stoptask.c"
2019 #ifdef CONFIG_SCHED_DEBUG
2020 # include "sched_debug.c"
2023 void sched_set_stop_task(int cpu, struct task_struct *stop)
2025 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2026 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2030 * Make it appear like a SCHED_FIFO task, its something
2031 * userspace knows about and won't get confused about.
2033 * Also, it will make PI more or less work without too
2034 * much confusion -- but then, stop work should not
2035 * rely on PI working anyway.
2037 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2039 stop->sched_class = &stop_sched_class;
2042 cpu_rq(cpu)->stop = stop;
2046 * Reset it back to a normal scheduling class so that
2047 * it can die in pieces.
2049 old_stop->sched_class = &rt_sched_class;
2054 * __normal_prio - return the priority that is based on the static prio
2056 static inline int __normal_prio(struct task_struct *p)
2058 return p->static_prio;
2062 * Calculate the expected normal priority: i.e. priority
2063 * without taking RT-inheritance into account. Might be
2064 * boosted by interactivity modifiers. Changes upon fork,
2065 * setprio syscalls, and whenever the interactivity
2066 * estimator recalculates.
2068 static inline int normal_prio(struct task_struct *p)
2072 if (task_has_rt_policy(p))
2073 prio = MAX_RT_PRIO-1 - p->rt_priority;
2075 prio = __normal_prio(p);
2080 * Calculate the current priority, i.e. the priority
2081 * taken into account by the scheduler. This value might
2082 * be boosted by RT tasks, or might be boosted by
2083 * interactivity modifiers. Will be RT if the task got
2084 * RT-boosted. If not then it returns p->normal_prio.
2086 static int effective_prio(struct task_struct *p)
2088 p->normal_prio = normal_prio(p);
2090 * If we are RT tasks or we were boosted to RT priority,
2091 * keep the priority unchanged. Otherwise, update priority
2092 * to the normal priority:
2094 if (!rt_prio(p->prio))
2095 return p->normal_prio;
2100 * task_curr - is this task currently executing on a CPU?
2101 * @p: the task in question.
2103 inline int task_curr(const struct task_struct *p)
2105 return cpu_curr(task_cpu(p)) == p;
2108 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2109 const struct sched_class *prev_class,
2110 int oldprio, int running)
2112 if (prev_class != p->sched_class) {
2113 if (prev_class->switched_from)
2114 prev_class->switched_from(rq, p, running);
2115 p->sched_class->switched_to(rq, p, running);
2117 p->sched_class->prio_changed(rq, p, oldprio, running);
2122 * Is this task likely cache-hot:
2125 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2129 if (p->sched_class != &fair_sched_class)
2132 if (unlikely(p->policy == SCHED_IDLE))
2136 * Buddy candidates are cache hot:
2138 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2139 (&p->se == cfs_rq_of(&p->se)->next ||
2140 &p->se == cfs_rq_of(&p->se)->last))
2143 if (sysctl_sched_migration_cost == -1)
2145 if (sysctl_sched_migration_cost == 0)
2148 delta = now - p->se.exec_start;
2150 return delta < (s64)sysctl_sched_migration_cost;
2153 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2155 #ifdef CONFIG_SCHED_DEBUG
2157 * We should never call set_task_cpu() on a blocked task,
2158 * ttwu() will sort out the placement.
2160 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2161 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2164 trace_sched_migrate_task(p, new_cpu);
2166 if (task_cpu(p) != new_cpu) {
2167 p->se.nr_migrations++;
2168 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2171 __set_task_cpu(p, new_cpu);
2174 struct migration_arg {
2175 struct task_struct *task;
2179 static int migration_cpu_stop(void *data);
2182 * The task's runqueue lock must be held.
2183 * Returns true if you have to wait for migration thread.
2185 static bool migrate_task(struct task_struct *p, int dest_cpu)
2187 struct rq *rq = task_rq(p);
2190 * If the task is not on a runqueue (and not running), then
2191 * the next wake-up will properly place the task.
2193 return p->se.on_rq || task_running(rq, p);
2197 * wait_task_inactive - wait for a thread to unschedule.
2199 * If @match_state is nonzero, it's the @p->state value just checked and
2200 * not expected to change. If it changes, i.e. @p might have woken up,
2201 * then return zero. When we succeed in waiting for @p to be off its CPU,
2202 * we return a positive number (its total switch count). If a second call
2203 * a short while later returns the same number, the caller can be sure that
2204 * @p has remained unscheduled the whole time.
2206 * The caller must ensure that the task *will* unschedule sometime soon,
2207 * else this function might spin for a *long* time. This function can't
2208 * be called with interrupts off, or it may introduce deadlock with
2209 * smp_call_function() if an IPI is sent by the same process we are
2210 * waiting to become inactive.
2212 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2214 unsigned long flags;
2221 * We do the initial early heuristics without holding
2222 * any task-queue locks at all. We'll only try to get
2223 * the runqueue lock when things look like they will
2229 * If the task is actively running on another CPU
2230 * still, just relax and busy-wait without holding
2233 * NOTE! Since we don't hold any locks, it's not
2234 * even sure that "rq" stays as the right runqueue!
2235 * But we don't care, since "task_running()" will
2236 * return false if the runqueue has changed and p
2237 * is actually now running somewhere else!
2239 while (task_running(rq, p)) {
2240 if (match_state && unlikely(p->state != match_state))
2246 * Ok, time to look more closely! We need the rq
2247 * lock now, to be *sure*. If we're wrong, we'll
2248 * just go back and repeat.
2250 rq = task_rq_lock(p, &flags);
2251 trace_sched_wait_task(p);
2252 running = task_running(rq, p);
2253 on_rq = p->se.on_rq;
2255 if (!match_state || p->state == match_state)
2256 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2257 task_rq_unlock(rq, &flags);
2260 * If it changed from the expected state, bail out now.
2262 if (unlikely(!ncsw))
2266 * Was it really running after all now that we
2267 * checked with the proper locks actually held?
2269 * Oops. Go back and try again..
2271 if (unlikely(running)) {
2277 * It's not enough that it's not actively running,
2278 * it must be off the runqueue _entirely_, and not
2281 * So if it was still runnable (but just not actively
2282 * running right now), it's preempted, and we should
2283 * yield - it could be a while.
2285 if (unlikely(on_rq)) {
2286 schedule_timeout_uninterruptible(1);
2291 * Ahh, all good. It wasn't running, and it wasn't
2292 * runnable, which means that it will never become
2293 * running in the future either. We're all done!
2302 * kick_process - kick a running thread to enter/exit the kernel
2303 * @p: the to-be-kicked thread
2305 * Cause a process which is running on another CPU to enter
2306 * kernel-mode, without any delay. (to get signals handled.)
2308 * NOTE: this function doesnt have to take the runqueue lock,
2309 * because all it wants to ensure is that the remote task enters
2310 * the kernel. If the IPI races and the task has been migrated
2311 * to another CPU then no harm is done and the purpose has been
2314 void kick_process(struct task_struct *p)
2320 if ((cpu != smp_processor_id()) && task_curr(p))
2321 smp_send_reschedule(cpu);
2324 EXPORT_SYMBOL_GPL(kick_process);
2325 #endif /* CONFIG_SMP */
2328 * task_oncpu_function_call - call a function on the cpu on which a task runs
2329 * @p: the task to evaluate
2330 * @func: the function to be called
2331 * @info: the function call argument
2333 * Calls the function @func when the task is currently running. This might
2334 * be on the current CPU, which just calls the function directly
2336 void task_oncpu_function_call(struct task_struct *p,
2337 void (*func) (void *info), void *info)
2344 smp_call_function_single(cpu, func, info, 1);
2350 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2352 static int select_fallback_rq(int cpu, struct task_struct *p)
2355 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2357 /* Look for allowed, online CPU in same node. */
2358 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2359 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2362 /* Any allowed, online CPU? */
2363 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2364 if (dest_cpu < nr_cpu_ids)
2367 /* No more Mr. Nice Guy. */
2368 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2369 dest_cpu = cpuset_cpus_allowed_fallback(p);
2371 * Don't tell them about moving exiting tasks or
2372 * kernel threads (both mm NULL), since they never
2375 if (p->mm && printk_ratelimit()) {
2376 printk(KERN_INFO "process %d (%s) no "
2377 "longer affine to cpu%d\n",
2378 task_pid_nr(p), p->comm, cpu);
2386 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2389 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2391 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2394 * In order not to call set_task_cpu() on a blocking task we need
2395 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2398 * Since this is common to all placement strategies, this lives here.
2400 * [ this allows ->select_task() to simply return task_cpu(p) and
2401 * not worry about this generic constraint ]
2403 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2405 cpu = select_fallback_rq(task_cpu(p), p);
2410 static void update_avg(u64 *avg, u64 sample)
2412 s64 diff = sample - *avg;
2417 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2418 bool is_sync, bool is_migrate, bool is_local,
2419 unsigned long en_flags)
2421 schedstat_inc(p, se.statistics.nr_wakeups);
2423 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2425 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2427 schedstat_inc(p, se.statistics.nr_wakeups_local);
2429 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2431 activate_task(rq, p, en_flags);
2434 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2435 int wake_flags, bool success)
2437 trace_sched_wakeup(p, success);
2438 check_preempt_curr(rq, p, wake_flags);
2440 p->state = TASK_RUNNING;
2442 if (p->sched_class->task_woken)
2443 p->sched_class->task_woken(rq, p);
2445 if (unlikely(rq->idle_stamp)) {
2446 u64 delta = rq->clock - rq->idle_stamp;
2447 u64 max = 2*sysctl_sched_migration_cost;
2452 update_avg(&rq->avg_idle, delta);
2456 /* if a worker is waking up, notify workqueue */
2457 if ((p->flags & PF_WQ_WORKER) && success)
2458 wq_worker_waking_up(p, cpu_of(rq));
2462 * try_to_wake_up - wake up a thread
2463 * @p: the thread to be awakened
2464 * @state: the mask of task states that can be woken
2465 * @wake_flags: wake modifier flags (WF_*)
2467 * Put it on the run-queue if it's not already there. The "current"
2468 * thread is always on the run-queue (except when the actual
2469 * re-schedule is in progress), and as such you're allowed to do
2470 * the simpler "current->state = TASK_RUNNING" to mark yourself
2471 * runnable without the overhead of this.
2473 * Returns %true if @p was woken up, %false if it was already running
2474 * or @state didn't match @p's state.
2476 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2479 int cpu, orig_cpu, this_cpu, success = 0;
2480 unsigned long flags;
2481 unsigned long en_flags = ENQUEUE_WAKEUP;
2484 this_cpu = get_cpu();
2487 rq = task_rq_lock(p, &flags);
2488 if (!(p->state & state))
2498 if (unlikely(task_running(rq, p)))
2502 * In order to handle concurrent wakeups and release the rq->lock
2503 * we put the task in TASK_WAKING state.
2505 * First fix up the nr_uninterruptible count:
2507 if (task_contributes_to_load(p)) {
2508 if (likely(cpu_online(orig_cpu)))
2509 rq->nr_uninterruptible--;
2511 this_rq()->nr_uninterruptible--;
2513 p->state = TASK_WAKING;
2515 if (p->sched_class->task_waking) {
2516 p->sched_class->task_waking(rq, p);
2517 en_flags |= ENQUEUE_WAKING;
2520 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2521 if (cpu != orig_cpu)
2522 set_task_cpu(p, cpu);
2523 __task_rq_unlock(rq);
2526 raw_spin_lock(&rq->lock);
2529 * We migrated the task without holding either rq->lock, however
2530 * since the task is not on the task list itself, nobody else
2531 * will try and migrate the task, hence the rq should match the
2532 * cpu we just moved it to.
2534 WARN_ON(task_cpu(p) != cpu);
2535 WARN_ON(p->state != TASK_WAKING);
2537 #ifdef CONFIG_SCHEDSTATS
2538 schedstat_inc(rq, ttwu_count);
2539 if (cpu == this_cpu)
2540 schedstat_inc(rq, ttwu_local);
2542 struct sched_domain *sd;
2543 for_each_domain(this_cpu, sd) {
2544 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2545 schedstat_inc(sd, ttwu_wake_remote);
2550 #endif /* CONFIG_SCHEDSTATS */
2553 #endif /* CONFIG_SMP */
2554 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2555 cpu == this_cpu, en_flags);
2558 ttwu_post_activation(p, rq, wake_flags, success);
2560 task_rq_unlock(rq, &flags);
2567 * try_to_wake_up_local - try to wake up a local task with rq lock held
2568 * @p: the thread to be awakened
2570 * Put @p on the run-queue if it's not alredy there. The caller must
2571 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2572 * the current task. this_rq() stays locked over invocation.
2574 static void try_to_wake_up_local(struct task_struct *p)
2576 struct rq *rq = task_rq(p);
2577 bool success = false;
2579 BUG_ON(rq != this_rq());
2580 BUG_ON(p == current);
2581 lockdep_assert_held(&rq->lock);
2583 if (!(p->state & TASK_NORMAL))
2587 if (likely(!task_running(rq, p))) {
2588 schedstat_inc(rq, ttwu_count);
2589 schedstat_inc(rq, ttwu_local);
2591 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2594 ttwu_post_activation(p, rq, 0, success);
2598 * wake_up_process - Wake up a specific process
2599 * @p: The process to be woken up.
2601 * Attempt to wake up the nominated process and move it to the set of runnable
2602 * processes. Returns 1 if the process was woken up, 0 if it was already
2605 * It may be assumed that this function implies a write memory barrier before
2606 * changing the task state if and only if any tasks are woken up.
2608 int wake_up_process(struct task_struct *p)
2610 return try_to_wake_up(p, TASK_ALL, 0);
2612 EXPORT_SYMBOL(wake_up_process);
2614 int wake_up_state(struct task_struct *p, unsigned int state)
2616 return try_to_wake_up(p, state, 0);
2620 * Perform scheduler related setup for a newly forked process p.
2621 * p is forked by current.
2623 * __sched_fork() is basic setup used by init_idle() too:
2625 static void __sched_fork(struct task_struct *p)
2627 p->se.exec_start = 0;
2628 p->se.sum_exec_runtime = 0;
2629 p->se.prev_sum_exec_runtime = 0;
2630 p->se.nr_migrations = 0;
2632 #ifdef CONFIG_SCHEDSTATS
2633 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2636 INIT_LIST_HEAD(&p->rt.run_list);
2638 INIT_LIST_HEAD(&p->se.group_node);
2640 #ifdef CONFIG_PREEMPT_NOTIFIERS
2641 INIT_HLIST_HEAD(&p->preempt_notifiers);
2646 * fork()/clone()-time setup:
2648 void sched_fork(struct task_struct *p, int clone_flags)
2650 int cpu = get_cpu();
2654 * We mark the process as running here. This guarantees that
2655 * nobody will actually run it, and a signal or other external
2656 * event cannot wake it up and insert it on the runqueue either.
2658 p->state = TASK_RUNNING;
2661 * Revert to default priority/policy on fork if requested.
2663 if (unlikely(p->sched_reset_on_fork)) {
2664 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2665 p->policy = SCHED_NORMAL;
2666 p->normal_prio = p->static_prio;
2669 if (PRIO_TO_NICE(p->static_prio) < 0) {
2670 p->static_prio = NICE_TO_PRIO(0);
2671 p->normal_prio = p->static_prio;
2676 * We don't need the reset flag anymore after the fork. It has
2677 * fulfilled its duty:
2679 p->sched_reset_on_fork = 0;
2683 * Make sure we do not leak PI boosting priority to the child.
2685 p->prio = current->normal_prio;
2687 if (!rt_prio(p->prio))
2688 p->sched_class = &fair_sched_class;
2690 if (p->sched_class->task_fork)
2691 p->sched_class->task_fork(p);
2694 * The child is not yet in the pid-hash so no cgroup attach races,
2695 * and the cgroup is pinned to this child due to cgroup_fork()
2696 * is ran before sched_fork().
2698 * Silence PROVE_RCU.
2701 set_task_cpu(p, cpu);
2704 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2705 if (likely(sched_info_on()))
2706 memset(&p->sched_info, 0, sizeof(p->sched_info));
2708 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2711 #ifdef CONFIG_PREEMPT
2712 /* Want to start with kernel preemption disabled. */
2713 task_thread_info(p)->preempt_count = 1;
2715 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2721 * wake_up_new_task - wake up a newly created task for the first time.
2723 * This function will do some initial scheduler statistics housekeeping
2724 * that must be done for every newly created context, then puts the task
2725 * on the runqueue and wakes it.
2727 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2729 unsigned long flags;
2731 int cpu __maybe_unused = get_cpu();
2734 rq = task_rq_lock(p, &flags);
2735 p->state = TASK_WAKING;
2738 * Fork balancing, do it here and not earlier because:
2739 * - cpus_allowed can change in the fork path
2740 * - any previously selected cpu might disappear through hotplug
2742 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2743 * without people poking at ->cpus_allowed.
2745 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2746 set_task_cpu(p, cpu);
2748 p->state = TASK_RUNNING;
2749 task_rq_unlock(rq, &flags);
2752 rq = task_rq_lock(p, &flags);
2753 activate_task(rq, p, 0);
2754 trace_sched_wakeup_new(p, 1);
2755 check_preempt_curr(rq, p, WF_FORK);
2757 if (p->sched_class->task_woken)
2758 p->sched_class->task_woken(rq, p);
2760 task_rq_unlock(rq, &flags);
2764 #ifdef CONFIG_PREEMPT_NOTIFIERS
2767 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2768 * @notifier: notifier struct to register
2770 void preempt_notifier_register(struct preempt_notifier *notifier)
2772 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2774 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2777 * preempt_notifier_unregister - no longer interested in preemption notifications
2778 * @notifier: notifier struct to unregister
2780 * This is safe to call from within a preemption notifier.
2782 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2784 hlist_del(¬ifier->link);
2786 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2788 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2790 struct preempt_notifier *notifier;
2791 struct hlist_node *node;
2793 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2794 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2798 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2799 struct task_struct *next)
2801 struct preempt_notifier *notifier;
2802 struct hlist_node *node;
2804 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2805 notifier->ops->sched_out(notifier, next);
2808 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2810 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2815 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2816 struct task_struct *next)
2820 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2823 * prepare_task_switch - prepare to switch tasks
2824 * @rq: the runqueue preparing to switch
2825 * @prev: the current task that is being switched out
2826 * @next: the task we are going to switch to.
2828 * This is called with the rq lock held and interrupts off. It must
2829 * be paired with a subsequent finish_task_switch after the context
2832 * prepare_task_switch sets up locking and calls architecture specific
2836 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2837 struct task_struct *next)
2839 fire_sched_out_preempt_notifiers(prev, next);
2840 prepare_lock_switch(rq, next);
2841 prepare_arch_switch(next);
2845 * finish_task_switch - clean up after a task-switch
2846 * @rq: runqueue associated with task-switch
2847 * @prev: the thread we just switched away from.
2849 * finish_task_switch must be called after the context switch, paired
2850 * with a prepare_task_switch call before the context switch.
2851 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2852 * and do any other architecture-specific cleanup actions.
2854 * Note that we may have delayed dropping an mm in context_switch(). If
2855 * so, we finish that here outside of the runqueue lock. (Doing it
2856 * with the lock held can cause deadlocks; see schedule() for
2859 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2860 __releases(rq->lock)
2862 struct mm_struct *mm = rq->prev_mm;
2868 * A task struct has one reference for the use as "current".
2869 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2870 * schedule one last time. The schedule call will never return, and
2871 * the scheduled task must drop that reference.
2872 * The test for TASK_DEAD must occur while the runqueue locks are
2873 * still held, otherwise prev could be scheduled on another cpu, die
2874 * there before we look at prev->state, and then the reference would
2876 * Manfred Spraul <manfred@colorfullife.com>
2878 prev_state = prev->state;
2879 finish_arch_switch(prev);
2880 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2881 local_irq_disable();
2882 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2883 perf_event_task_sched_in(current);
2884 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2886 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2887 finish_lock_switch(rq, prev);
2889 fire_sched_in_preempt_notifiers(current);
2892 if (unlikely(prev_state == TASK_DEAD)) {
2894 * Remove function-return probe instances associated with this
2895 * task and put them back on the free list.
2897 kprobe_flush_task(prev);
2898 put_task_struct(prev);
2904 /* assumes rq->lock is held */
2905 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2907 if (prev->sched_class->pre_schedule)
2908 prev->sched_class->pre_schedule(rq, prev);
2911 /* rq->lock is NOT held, but preemption is disabled */
2912 static inline void post_schedule(struct rq *rq)
2914 if (rq->post_schedule) {
2915 unsigned long flags;
2917 raw_spin_lock_irqsave(&rq->lock, flags);
2918 if (rq->curr->sched_class->post_schedule)
2919 rq->curr->sched_class->post_schedule(rq);
2920 raw_spin_unlock_irqrestore(&rq->lock, flags);
2922 rq->post_schedule = 0;
2928 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2932 static inline void post_schedule(struct rq *rq)
2939 * schedule_tail - first thing a freshly forked thread must call.
2940 * @prev: the thread we just switched away from.
2942 asmlinkage void schedule_tail(struct task_struct *prev)
2943 __releases(rq->lock)
2945 struct rq *rq = this_rq();
2947 finish_task_switch(rq, prev);
2950 * FIXME: do we need to worry about rq being invalidated by the
2955 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2956 /* In this case, finish_task_switch does not reenable preemption */
2959 if (current->set_child_tid)
2960 put_user(task_pid_vnr(current), current->set_child_tid);
2964 * context_switch - switch to the new MM and the new
2965 * thread's register state.
2968 context_switch(struct rq *rq, struct task_struct *prev,
2969 struct task_struct *next)
2971 struct mm_struct *mm, *oldmm;
2973 prepare_task_switch(rq, prev, next);
2974 trace_sched_switch(prev, next);
2976 oldmm = prev->active_mm;
2978 * For paravirt, this is coupled with an exit in switch_to to
2979 * combine the page table reload and the switch backend into
2982 arch_start_context_switch(prev);
2985 next->active_mm = oldmm;
2986 atomic_inc(&oldmm->mm_count);
2987 enter_lazy_tlb(oldmm, next);
2989 switch_mm(oldmm, mm, next);
2992 prev->active_mm = NULL;
2993 rq->prev_mm = oldmm;
2996 * Since the runqueue lock will be released by the next
2997 * task (which is an invalid locking op but in the case
2998 * of the scheduler it's an obvious special-case), so we
2999 * do an early lockdep release here:
3001 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3002 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3005 /* Here we just switch the register state and the stack. */
3006 switch_to(prev, next, prev);
3010 * this_rq must be evaluated again because prev may have moved
3011 * CPUs since it called schedule(), thus the 'rq' on its stack
3012 * frame will be invalid.
3014 finish_task_switch(this_rq(), prev);
3018 * nr_running, nr_uninterruptible and nr_context_switches:
3020 * externally visible scheduler statistics: current number of runnable
3021 * threads, current number of uninterruptible-sleeping threads, total
3022 * number of context switches performed since bootup.
3024 unsigned long nr_running(void)
3026 unsigned long i, sum = 0;
3028 for_each_online_cpu(i)
3029 sum += cpu_rq(i)->nr_running;
3034 unsigned long nr_uninterruptible(void)
3036 unsigned long i, sum = 0;
3038 for_each_possible_cpu(i)
3039 sum += cpu_rq(i)->nr_uninterruptible;
3042 * Since we read the counters lockless, it might be slightly
3043 * inaccurate. Do not allow it to go below zero though:
3045 if (unlikely((long)sum < 0))
3051 unsigned long long nr_context_switches(void)
3054 unsigned long long sum = 0;
3056 for_each_possible_cpu(i)
3057 sum += cpu_rq(i)->nr_switches;
3062 unsigned long nr_iowait(void)
3064 unsigned long i, sum = 0;
3066 for_each_possible_cpu(i)
3067 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3072 unsigned long nr_iowait_cpu(int cpu)
3074 struct rq *this = cpu_rq(cpu);
3075 return atomic_read(&this->nr_iowait);
3078 unsigned long this_cpu_load(void)
3080 struct rq *this = this_rq();
3081 return this->cpu_load[0];
3085 /* Variables and functions for calc_load */
3086 static atomic_long_t calc_load_tasks;
3087 static unsigned long calc_load_update;
3088 unsigned long avenrun[3];
3089 EXPORT_SYMBOL(avenrun);
3091 static long calc_load_fold_active(struct rq *this_rq)
3093 long nr_active, delta = 0;
3095 nr_active = this_rq->nr_running;
3096 nr_active += (long) this_rq->nr_uninterruptible;
3098 if (nr_active != this_rq->calc_load_active) {
3099 delta = nr_active - this_rq->calc_load_active;
3100 this_rq->calc_load_active = nr_active;
3108 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3110 * When making the ILB scale, we should try to pull this in as well.
3112 static atomic_long_t calc_load_tasks_idle;
3114 static void calc_load_account_idle(struct rq *this_rq)
3118 delta = calc_load_fold_active(this_rq);
3120 atomic_long_add(delta, &calc_load_tasks_idle);
3123 static long calc_load_fold_idle(void)
3128 * Its got a race, we don't care...
3130 if (atomic_long_read(&calc_load_tasks_idle))
3131 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3136 static void calc_load_account_idle(struct rq *this_rq)
3140 static inline long calc_load_fold_idle(void)
3147 * get_avenrun - get the load average array
3148 * @loads: pointer to dest load array
3149 * @offset: offset to add
3150 * @shift: shift count to shift the result left
3152 * These values are estimates at best, so no need for locking.
3154 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3156 loads[0] = (avenrun[0] + offset) << shift;
3157 loads[1] = (avenrun[1] + offset) << shift;
3158 loads[2] = (avenrun[2] + offset) << shift;
3161 static unsigned long
3162 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3165 load += active * (FIXED_1 - exp);
3166 return load >> FSHIFT;
3170 * calc_load - update the avenrun load estimates 10 ticks after the
3171 * CPUs have updated calc_load_tasks.
3173 void calc_global_load(void)
3175 unsigned long upd = calc_load_update + 10;
3178 if (time_before(jiffies, upd))
3181 active = atomic_long_read(&calc_load_tasks);
3182 active = active > 0 ? active * FIXED_1 : 0;
3184 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3185 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3186 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3188 calc_load_update += LOAD_FREQ;
3192 * Called from update_cpu_load() to periodically update this CPU's
3195 static void calc_load_account_active(struct rq *this_rq)
3199 if (time_before(jiffies, this_rq->calc_load_update))
3202 delta = calc_load_fold_active(this_rq);
3203 delta += calc_load_fold_idle();
3205 atomic_long_add(delta, &calc_load_tasks);
3207 this_rq->calc_load_update += LOAD_FREQ;
3211 * The exact cpuload at various idx values, calculated at every tick would be
3212 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3214 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3215 * on nth tick when cpu may be busy, then we have:
3216 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3217 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3219 * decay_load_missed() below does efficient calculation of
3220 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3221 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3223 * The calculation is approximated on a 128 point scale.
3224 * degrade_zero_ticks is the number of ticks after which load at any
3225 * particular idx is approximated to be zero.
3226 * degrade_factor is a precomputed table, a row for each load idx.
3227 * Each column corresponds to degradation factor for a power of two ticks,
3228 * based on 128 point scale.
3230 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3231 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3233 * With this power of 2 load factors, we can degrade the load n times
3234 * by looking at 1 bits in n and doing as many mult/shift instead of
3235 * n mult/shifts needed by the exact degradation.
3237 #define DEGRADE_SHIFT 7
3238 static const unsigned char
3239 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3240 static const unsigned char
3241 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3242 {0, 0, 0, 0, 0, 0, 0, 0},
3243 {64, 32, 8, 0, 0, 0, 0, 0},
3244 {96, 72, 40, 12, 1, 0, 0},
3245 {112, 98, 75, 43, 15, 1, 0},
3246 {120, 112, 98, 76, 45, 16, 2} };
3249 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3250 * would be when CPU is idle and so we just decay the old load without
3251 * adding any new load.
3253 static unsigned long
3254 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3258 if (!missed_updates)
3261 if (missed_updates >= degrade_zero_ticks[idx])
3265 return load >> missed_updates;
3267 while (missed_updates) {
3268 if (missed_updates % 2)
3269 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3271 missed_updates >>= 1;
3278 * Update rq->cpu_load[] statistics. This function is usually called every
3279 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3280 * every tick. We fix it up based on jiffies.
3282 static void update_cpu_load(struct rq *this_rq)
3284 unsigned long this_load = this_rq->load.weight;
3285 unsigned long curr_jiffies = jiffies;
3286 unsigned long pending_updates;
3289 this_rq->nr_load_updates++;
3291 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3292 if (curr_jiffies == this_rq->last_load_update_tick)
3295 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3296 this_rq->last_load_update_tick = curr_jiffies;
3298 /* Update our load: */
3299 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3300 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3301 unsigned long old_load, new_load;
3303 /* scale is effectively 1 << i now, and >> i divides by scale */
3305 old_load = this_rq->cpu_load[i];
3306 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3307 new_load = this_load;
3309 * Round up the averaging division if load is increasing. This
3310 * prevents us from getting stuck on 9 if the load is 10, for
3313 if (new_load > old_load)
3314 new_load += scale - 1;
3316 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3319 sched_avg_update(this_rq);
3322 static void update_cpu_load_active(struct rq *this_rq)
3324 update_cpu_load(this_rq);
3326 calc_load_account_active(this_rq);
3332 * sched_exec - execve() is a valuable balancing opportunity, because at
3333 * this point the task has the smallest effective memory and cache footprint.
3335 void sched_exec(void)
3337 struct task_struct *p = current;
3338 unsigned long flags;
3342 rq = task_rq_lock(p, &flags);
3343 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3344 if (dest_cpu == smp_processor_id())
3348 * select_task_rq() can race against ->cpus_allowed
3350 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3351 likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
3352 struct migration_arg arg = { p, dest_cpu };
3354 task_rq_unlock(rq, &flags);
3355 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3359 task_rq_unlock(rq, &flags);
3364 DEFINE_PER_CPU(struct kernel_stat, kstat);
3366 EXPORT_PER_CPU_SYMBOL(kstat);
3369 * Return any ns on the sched_clock that have not yet been accounted in
3370 * @p in case that task is currently running.
3372 * Called with task_rq_lock() held on @rq.
3374 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3378 if (task_current(rq, p)) {
3379 update_rq_clock(rq);
3380 ns = rq->clock_task - p->se.exec_start;
3388 unsigned long long task_delta_exec(struct task_struct *p)
3390 unsigned long flags;
3394 rq = task_rq_lock(p, &flags);
3395 ns = do_task_delta_exec(p, rq);
3396 task_rq_unlock(rq, &flags);
3402 * Return accounted runtime for the task.
3403 * In case the task is currently running, return the runtime plus current's
3404 * pending runtime that have not been accounted yet.
3406 unsigned long long task_sched_runtime(struct task_struct *p)
3408 unsigned long flags;
3412 rq = task_rq_lock(p, &flags);
3413 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3414 task_rq_unlock(rq, &flags);
3420 * Return sum_exec_runtime for the thread group.
3421 * In case the task is currently running, return the sum plus current's
3422 * pending runtime that have not been accounted yet.
3424 * Note that the thread group might have other running tasks as well,
3425 * so the return value not includes other pending runtime that other
3426 * running tasks might have.
3428 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3430 struct task_cputime totals;
3431 unsigned long flags;
3435 rq = task_rq_lock(p, &flags);
3436 thread_group_cputime(p, &totals);
3437 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3438 task_rq_unlock(rq, &flags);
3444 * Account user cpu time to a process.
3445 * @p: the process that the cpu time gets accounted to
3446 * @cputime: the cpu time spent in user space since the last update
3447 * @cputime_scaled: cputime scaled by cpu frequency
3449 void account_user_time(struct task_struct *p, cputime_t cputime,
3450 cputime_t cputime_scaled)
3452 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3455 /* Add user time to process. */
3456 p->utime = cputime_add(p->utime, cputime);
3457 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3458 account_group_user_time(p, cputime);
3460 /* Add user time to cpustat. */
3461 tmp = cputime_to_cputime64(cputime);
3462 if (TASK_NICE(p) > 0)
3463 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3465 cpustat->user = cputime64_add(cpustat->user, tmp);
3467 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3468 /* Account for user time used */
3469 acct_update_integrals(p);
3473 * Account guest cpu time to a process.
3474 * @p: the process that the cpu time gets accounted to
3475 * @cputime: the cpu time spent in virtual machine since the last update
3476 * @cputime_scaled: cputime scaled by cpu frequency
3478 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3479 cputime_t cputime_scaled)
3482 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3484 tmp = cputime_to_cputime64(cputime);
3486 /* Add guest time to process. */
3487 p->utime = cputime_add(p->utime, cputime);
3488 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3489 account_group_user_time(p, cputime);
3490 p->gtime = cputime_add(p->gtime, cputime);
3492 /* Add guest time to cpustat. */
3493 if (TASK_NICE(p) > 0) {
3494 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3495 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3497 cpustat->user = cputime64_add(cpustat->user, tmp);
3498 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3503 * Account system cpu time to a process.
3504 * @p: the process that the cpu time gets accounted to
3505 * @hardirq_offset: the offset to subtract from hardirq_count()
3506 * @cputime: the cpu time spent in kernel space since the last update
3507 * @cputime_scaled: cputime scaled by cpu frequency
3509 void account_system_time(struct task_struct *p, int hardirq_offset,
3510 cputime_t cputime, cputime_t cputime_scaled)
3512 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3515 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3516 account_guest_time(p, cputime, cputime_scaled);
3520 /* Add system time to process. */
3521 p->stime = cputime_add(p->stime, cputime);
3522 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3523 account_group_system_time(p, cputime);
3525 /* Add system time to cpustat. */
3526 tmp = cputime_to_cputime64(cputime);
3527 if (hardirq_count() - hardirq_offset)
3528 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3529 else if (in_serving_softirq())
3530 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3532 cpustat->system = cputime64_add(cpustat->system, tmp);
3534 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3536 /* Account for system time used */
3537 acct_update_integrals(p);
3541 * Account for involuntary wait time.
3542 * @steal: the cpu time spent in involuntary wait
3544 void account_steal_time(cputime_t cputime)
3546 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3547 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3549 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3553 * Account for idle time.
3554 * @cputime: the cpu time spent in idle wait
3556 void account_idle_time(cputime_t cputime)
3558 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3559 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3560 struct rq *rq = this_rq();
3562 if (atomic_read(&rq->nr_iowait) > 0)
3563 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3565 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3568 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3571 * Account a single tick of cpu time.
3572 * @p: the process that the cpu time gets accounted to
3573 * @user_tick: indicates if the tick is a user or a system tick
3575 void account_process_tick(struct task_struct *p, int user_tick)
3577 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3578 struct rq *rq = this_rq();
3581 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3582 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3583 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3586 account_idle_time(cputime_one_jiffy);
3590 * Account multiple ticks of steal time.
3591 * @p: the process from which the cpu time has been stolen
3592 * @ticks: number of stolen ticks
3594 void account_steal_ticks(unsigned long ticks)
3596 account_steal_time(jiffies_to_cputime(ticks));
3600 * Account multiple ticks of idle time.
3601 * @ticks: number of stolen ticks
3603 void account_idle_ticks(unsigned long ticks)
3605 account_idle_time(jiffies_to_cputime(ticks));
3611 * Use precise platform statistics if available:
3613 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3614 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3620 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3622 struct task_cputime cputime;
3624 thread_group_cputime(p, &cputime);
3626 *ut = cputime.utime;
3627 *st = cputime.stime;
3631 #ifndef nsecs_to_cputime
3632 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3635 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3637 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3640 * Use CFS's precise accounting:
3642 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3648 do_div(temp, total);
3649 utime = (cputime_t)temp;
3654 * Compare with previous values, to keep monotonicity:
3656 p->prev_utime = max(p->prev_utime, utime);
3657 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3659 *ut = p->prev_utime;
3660 *st = p->prev_stime;
3664 * Must be called with siglock held.
3666 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3668 struct signal_struct *sig = p->signal;
3669 struct task_cputime cputime;
3670 cputime_t rtime, utime, total;
3672 thread_group_cputime(p, &cputime);
3674 total = cputime_add(cputime.utime, cputime.stime);
3675 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3680 temp *= cputime.utime;
3681 do_div(temp, total);
3682 utime = (cputime_t)temp;
3686 sig->prev_utime = max(sig->prev_utime, utime);
3687 sig->prev_stime = max(sig->prev_stime,
3688 cputime_sub(rtime, sig->prev_utime));
3690 *ut = sig->prev_utime;
3691 *st = sig->prev_stime;
3696 * This function gets called by the timer code, with HZ frequency.
3697 * We call it with interrupts disabled.
3699 * It also gets called by the fork code, when changing the parent's
3702 void scheduler_tick(void)
3704 int cpu = smp_processor_id();
3705 struct rq *rq = cpu_rq(cpu);
3706 struct task_struct *curr = rq->curr;
3710 raw_spin_lock(&rq->lock);
3711 update_rq_clock(rq);
3712 update_cpu_load_active(rq);
3713 curr->sched_class->task_tick(rq, curr, 0);
3714 raw_spin_unlock(&rq->lock);
3716 perf_event_task_tick(curr);
3719 rq->idle_at_tick = idle_cpu(cpu);
3720 trigger_load_balance(rq, cpu);
3724 notrace unsigned long get_parent_ip(unsigned long addr)
3726 if (in_lock_functions(addr)) {
3727 addr = CALLER_ADDR2;
3728 if (in_lock_functions(addr))
3729 addr = CALLER_ADDR3;
3734 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3735 defined(CONFIG_PREEMPT_TRACER))
3737 void __kprobes add_preempt_count(int val)
3739 #ifdef CONFIG_DEBUG_PREEMPT
3743 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3746 preempt_count() += val;
3747 #ifdef CONFIG_DEBUG_PREEMPT
3749 * Spinlock count overflowing soon?
3751 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3754 if (preempt_count() == val)
3755 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3757 EXPORT_SYMBOL(add_preempt_count);
3759 void __kprobes sub_preempt_count(int val)
3761 #ifdef CONFIG_DEBUG_PREEMPT
3765 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3768 * Is the spinlock portion underflowing?
3770 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3771 !(preempt_count() & PREEMPT_MASK)))
3775 if (preempt_count() == val)
3776 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3777 preempt_count() -= val;
3779 EXPORT_SYMBOL(sub_preempt_count);
3784 * Print scheduling while atomic bug:
3786 static noinline void __schedule_bug(struct task_struct *prev)
3788 struct pt_regs *regs = get_irq_regs();
3790 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3791 prev->comm, prev->pid, preempt_count());
3793 debug_show_held_locks(prev);
3795 if (irqs_disabled())
3796 print_irqtrace_events(prev);
3805 * Various schedule()-time debugging checks and statistics:
3807 static inline void schedule_debug(struct task_struct *prev)
3810 * Test if we are atomic. Since do_exit() needs to call into
3811 * schedule() atomically, we ignore that path for now.
3812 * Otherwise, whine if we are scheduling when we should not be.
3814 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3815 __schedule_bug(prev);
3817 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3819 schedstat_inc(this_rq(), sched_count);
3820 #ifdef CONFIG_SCHEDSTATS
3821 if (unlikely(prev->lock_depth >= 0)) {
3822 schedstat_inc(this_rq(), bkl_count);
3823 schedstat_inc(prev, sched_info.bkl_count);
3828 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3831 update_rq_clock(rq);
3832 rq->skip_clock_update = 0;
3833 prev->sched_class->put_prev_task(rq, prev);
3837 * Pick up the highest-prio task:
3839 static inline struct task_struct *
3840 pick_next_task(struct rq *rq)
3842 const struct sched_class *class;
3843 struct task_struct *p;
3846 * Optimization: we know that if all tasks are in
3847 * the fair class we can call that function directly:
3849 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3850 p = fair_sched_class.pick_next_task(rq);
3855 for_each_class(class) {
3856 p = class->pick_next_task(rq);
3861 BUG(); /* the idle class will always have a runnable task */
3865 * schedule() is the main scheduler function.
3867 asmlinkage void __sched schedule(void)
3869 struct task_struct *prev, *next;
3870 unsigned long *switch_count;
3876 cpu = smp_processor_id();
3878 rcu_note_context_switch(cpu);
3881 release_kernel_lock(prev);
3882 need_resched_nonpreemptible:
3884 schedule_debug(prev);
3886 if (sched_feat(HRTICK))
3889 raw_spin_lock_irq(&rq->lock);
3890 clear_tsk_need_resched(prev);
3892 switch_count = &prev->nivcsw;
3893 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3894 if (unlikely(signal_pending_state(prev->state, prev))) {
3895 prev->state = TASK_RUNNING;
3898 * If a worker is going to sleep, notify and
3899 * ask workqueue whether it wants to wake up a
3900 * task to maintain concurrency. If so, wake
3903 if (prev->flags & PF_WQ_WORKER) {
3904 struct task_struct *to_wakeup;
3906 to_wakeup = wq_worker_sleeping(prev, cpu);
3908 try_to_wake_up_local(to_wakeup);
3910 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3912 switch_count = &prev->nvcsw;
3915 pre_schedule(rq, prev);
3917 if (unlikely(!rq->nr_running))
3918 idle_balance(cpu, rq);
3920 put_prev_task(rq, prev);
3921 next = pick_next_task(rq);
3923 if (likely(prev != next)) {
3924 sched_info_switch(prev, next);
3925 perf_event_task_sched_out(prev, next);
3931 context_switch(rq, prev, next); /* unlocks the rq */
3933 * The context switch have flipped the stack from under us
3934 * and restored the local variables which were saved when
3935 * this task called schedule() in the past. prev == current
3936 * is still correct, but it can be moved to another cpu/rq.
3938 cpu = smp_processor_id();
3941 raw_spin_unlock_irq(&rq->lock);
3945 if (unlikely(reacquire_kernel_lock(prev)))
3946 goto need_resched_nonpreemptible;
3948 preempt_enable_no_resched();
3952 EXPORT_SYMBOL(schedule);
3954 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3956 * Look out! "owner" is an entirely speculative pointer
3957 * access and not reliable.
3959 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3964 if (!sched_feat(OWNER_SPIN))
3967 #ifdef CONFIG_DEBUG_PAGEALLOC
3969 * Need to access the cpu field knowing that
3970 * DEBUG_PAGEALLOC could have unmapped it if
3971 * the mutex owner just released it and exited.
3973 if (probe_kernel_address(&owner->cpu, cpu))
3980 * Even if the access succeeded (likely case),
3981 * the cpu field may no longer be valid.
3983 if (cpu >= nr_cpumask_bits)
3987 * We need to validate that we can do a
3988 * get_cpu() and that we have the percpu area.
3990 if (!cpu_online(cpu))
3997 * Owner changed, break to re-assess state.
3999 if (lock->owner != owner) {
4001 * If the lock has switched to a different owner,
4002 * we likely have heavy contention. Return 0 to quit
4003 * optimistic spinning and not contend further:
4011 * Is that owner really running on that cpu?
4013 if (task_thread_info(rq->curr) != owner || need_resched())
4023 #ifdef CONFIG_PREEMPT
4025 * this is the entry point to schedule() from in-kernel preemption
4026 * off of preempt_enable. Kernel preemptions off return from interrupt
4027 * occur there and call schedule directly.
4029 asmlinkage void __sched notrace preempt_schedule(void)
4031 struct thread_info *ti = current_thread_info();
4034 * If there is a non-zero preempt_count or interrupts are disabled,
4035 * we do not want to preempt the current task. Just return..
4037 if (likely(ti->preempt_count || irqs_disabled()))
4041 add_preempt_count_notrace(PREEMPT_ACTIVE);
4043 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4046 * Check again in case we missed a preemption opportunity
4047 * between schedule and now.
4050 } while (need_resched());
4052 EXPORT_SYMBOL(preempt_schedule);
4055 * this is the entry point to schedule() from kernel preemption
4056 * off of irq context.
4057 * Note, that this is called and return with irqs disabled. This will
4058 * protect us against recursive calling from irq.
4060 asmlinkage void __sched preempt_schedule_irq(void)
4062 struct thread_info *ti = current_thread_info();
4064 /* Catch callers which need to be fixed */
4065 BUG_ON(ti->preempt_count || !irqs_disabled());
4068 add_preempt_count(PREEMPT_ACTIVE);
4071 local_irq_disable();
4072 sub_preempt_count(PREEMPT_ACTIVE);
4075 * Check again in case we missed a preemption opportunity
4076 * between schedule and now.
4079 } while (need_resched());
4082 #endif /* CONFIG_PREEMPT */
4084 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4087 return try_to_wake_up(curr->private, mode, wake_flags);
4089 EXPORT_SYMBOL(default_wake_function);
4092 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4093 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4094 * number) then we wake all the non-exclusive tasks and one exclusive task.
4096 * There are circumstances in which we can try to wake a task which has already
4097 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4098 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4100 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4101 int nr_exclusive, int wake_flags, void *key)
4103 wait_queue_t *curr, *next;
4105 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4106 unsigned flags = curr->flags;
4108 if (curr->func(curr, mode, wake_flags, key) &&
4109 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4115 * __wake_up - wake up threads blocked on a waitqueue.
4117 * @mode: which threads
4118 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4119 * @key: is directly passed to the wakeup function
4121 * It may be assumed that this function implies a write memory barrier before
4122 * changing the task state if and only if any tasks are woken up.
4124 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4125 int nr_exclusive, void *key)
4127 unsigned long flags;
4129 spin_lock_irqsave(&q->lock, flags);
4130 __wake_up_common(q, mode, nr_exclusive, 0, key);
4131 spin_unlock_irqrestore(&q->lock, flags);
4133 EXPORT_SYMBOL(__wake_up);
4136 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4138 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4140 __wake_up_common(q, mode, 1, 0, NULL);
4142 EXPORT_SYMBOL_GPL(__wake_up_locked);
4144 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4146 __wake_up_common(q, mode, 1, 0, key);
4150 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4152 * @mode: which threads
4153 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4154 * @key: opaque value to be passed to wakeup targets
4156 * The sync wakeup differs that the waker knows that it will schedule
4157 * away soon, so while the target thread will be woken up, it will not
4158 * be migrated to another CPU - ie. the two threads are 'synchronized'
4159 * with each other. This can prevent needless bouncing between CPUs.
4161 * On UP it can prevent extra preemption.
4163 * It may be assumed that this function implies a write memory barrier before
4164 * changing the task state if and only if any tasks are woken up.
4166 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4167 int nr_exclusive, void *key)
4169 unsigned long flags;
4170 int wake_flags = WF_SYNC;
4175 if (unlikely(!nr_exclusive))
4178 spin_lock_irqsave(&q->lock, flags);
4179 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4180 spin_unlock_irqrestore(&q->lock, flags);
4182 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4185 * __wake_up_sync - see __wake_up_sync_key()
4187 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4189 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4191 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4194 * complete: - signals a single thread waiting on this completion
4195 * @x: holds the state of this particular completion
4197 * This will wake up a single thread waiting on this completion. Threads will be
4198 * awakened in the same order in which they were queued.
4200 * See also complete_all(), wait_for_completion() and related routines.
4202 * It may be assumed that this function implies a write memory barrier before
4203 * changing the task state if and only if any tasks are woken up.
4205 void complete(struct completion *x)
4207 unsigned long flags;
4209 spin_lock_irqsave(&x->wait.lock, flags);
4211 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4212 spin_unlock_irqrestore(&x->wait.lock, flags);
4214 EXPORT_SYMBOL(complete);
4217 * complete_all: - signals all threads waiting on this completion
4218 * @x: holds the state of this particular completion
4220 * This will wake up all threads waiting on this particular completion event.
4222 * It may be assumed that this function implies a write memory barrier before
4223 * changing the task state if and only if any tasks are woken up.
4225 void complete_all(struct completion *x)
4227 unsigned long flags;
4229 spin_lock_irqsave(&x->wait.lock, flags);
4230 x->done += UINT_MAX/2;
4231 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4232 spin_unlock_irqrestore(&x->wait.lock, flags);
4234 EXPORT_SYMBOL(complete_all);
4236 static inline long __sched
4237 do_wait_for_common(struct completion *x, long timeout, int state)
4240 DECLARE_WAITQUEUE(wait, current);
4242 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4244 if (signal_pending_state(state, current)) {
4245 timeout = -ERESTARTSYS;
4248 __set_current_state(state);
4249 spin_unlock_irq(&x->wait.lock);
4250 timeout = schedule_timeout(timeout);
4251 spin_lock_irq(&x->wait.lock);
4252 } while (!x->done && timeout);
4253 __remove_wait_queue(&x->wait, &wait);
4258 return timeout ?: 1;
4262 wait_for_common(struct completion *x, long timeout, int state)
4266 spin_lock_irq(&x->wait.lock);
4267 timeout = do_wait_for_common(x, timeout, state);
4268 spin_unlock_irq(&x->wait.lock);
4273 * wait_for_completion: - waits for completion of a task
4274 * @x: holds the state of this particular completion
4276 * This waits to be signaled for completion of a specific task. It is NOT
4277 * interruptible and there is no timeout.
4279 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4280 * and interrupt capability. Also see complete().
4282 void __sched wait_for_completion(struct completion *x)
4284 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4286 EXPORT_SYMBOL(wait_for_completion);
4289 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4290 * @x: holds the state of this particular completion
4291 * @timeout: timeout value in jiffies
4293 * This waits for either a completion of a specific task to be signaled or for a
4294 * specified timeout to expire. The timeout is in jiffies. It is not
4297 unsigned long __sched
4298 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4300 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4302 EXPORT_SYMBOL(wait_for_completion_timeout);
4305 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4306 * @x: holds the state of this particular completion
4308 * This waits for completion of a specific task to be signaled. It is
4311 int __sched wait_for_completion_interruptible(struct completion *x)
4313 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4314 if (t == -ERESTARTSYS)
4318 EXPORT_SYMBOL(wait_for_completion_interruptible);
4321 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4322 * @x: holds the state of this particular completion
4323 * @timeout: timeout value in jiffies
4325 * This waits for either a completion of a specific task to be signaled or for a
4326 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4328 unsigned long __sched
4329 wait_for_completion_interruptible_timeout(struct completion *x,
4330 unsigned long timeout)
4332 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4334 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4337 * wait_for_completion_killable: - waits for completion of a task (killable)
4338 * @x: holds the state of this particular completion
4340 * This waits to be signaled for completion of a specific task. It can be
4341 * interrupted by a kill signal.
4343 int __sched wait_for_completion_killable(struct completion *x)
4345 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4346 if (t == -ERESTARTSYS)
4350 EXPORT_SYMBOL(wait_for_completion_killable);
4353 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4354 * @x: holds the state of this particular completion
4355 * @timeout: timeout value in jiffies
4357 * This waits for either a completion of a specific task to be
4358 * signaled or for a specified timeout to expire. It can be
4359 * interrupted by a kill signal. The timeout is in jiffies.
4361 unsigned long __sched
4362 wait_for_completion_killable_timeout(struct completion *x,
4363 unsigned long timeout)
4365 return wait_for_common(x, timeout, TASK_KILLABLE);
4367 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4370 * try_wait_for_completion - try to decrement a completion without blocking
4371 * @x: completion structure
4373 * Returns: 0 if a decrement cannot be done without blocking
4374 * 1 if a decrement succeeded.
4376 * If a completion is being used as a counting completion,
4377 * attempt to decrement the counter without blocking. This
4378 * enables us to avoid waiting if the resource the completion
4379 * is protecting is not available.
4381 bool try_wait_for_completion(struct completion *x)
4383 unsigned long flags;
4386 spin_lock_irqsave(&x->wait.lock, flags);
4391 spin_unlock_irqrestore(&x->wait.lock, flags);
4394 EXPORT_SYMBOL(try_wait_for_completion);
4397 * completion_done - Test to see if a completion has any waiters
4398 * @x: completion structure
4400 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4401 * 1 if there are no waiters.
4404 bool completion_done(struct completion *x)
4406 unsigned long flags;
4409 spin_lock_irqsave(&x->wait.lock, flags);
4412 spin_unlock_irqrestore(&x->wait.lock, flags);
4415 EXPORT_SYMBOL(completion_done);
4418 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4420 unsigned long flags;
4423 init_waitqueue_entry(&wait, current);
4425 __set_current_state(state);
4427 spin_lock_irqsave(&q->lock, flags);
4428 __add_wait_queue(q, &wait);
4429 spin_unlock(&q->lock);
4430 timeout = schedule_timeout(timeout);
4431 spin_lock_irq(&q->lock);
4432 __remove_wait_queue(q, &wait);
4433 spin_unlock_irqrestore(&q->lock, flags);
4438 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4440 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4442 EXPORT_SYMBOL(interruptible_sleep_on);
4445 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4447 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4449 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4451 void __sched sleep_on(wait_queue_head_t *q)
4453 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4455 EXPORT_SYMBOL(sleep_on);
4457 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4459 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4461 EXPORT_SYMBOL(sleep_on_timeout);
4463 #ifdef CONFIG_RT_MUTEXES
4466 * rt_mutex_setprio - set the current priority of a task
4468 * @prio: prio value (kernel-internal form)
4470 * This function changes the 'effective' priority of a task. It does
4471 * not touch ->normal_prio like __setscheduler().
4473 * Used by the rt_mutex code to implement priority inheritance logic.
4475 void rt_mutex_setprio(struct task_struct *p, int prio)
4477 unsigned long flags;
4478 int oldprio, on_rq, running;
4480 const struct sched_class *prev_class;
4482 BUG_ON(prio < 0 || prio > MAX_PRIO);
4484 rq = task_rq_lock(p, &flags);
4486 trace_sched_pi_setprio(p, prio);
4488 prev_class = p->sched_class;
4489 on_rq = p->se.on_rq;
4490 running = task_current(rq, p);
4492 dequeue_task(rq, p, 0);
4494 p->sched_class->put_prev_task(rq, p);
4497 p->sched_class = &rt_sched_class;
4499 p->sched_class = &fair_sched_class;
4504 p->sched_class->set_curr_task(rq);
4506 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4508 check_class_changed(rq, p, prev_class, oldprio, running);
4510 task_rq_unlock(rq, &flags);
4515 void set_user_nice(struct task_struct *p, long nice)
4517 int old_prio, delta, on_rq;
4518 unsigned long flags;
4521 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4524 * We have to be careful, if called from sys_setpriority(),
4525 * the task might be in the middle of scheduling on another CPU.
4527 rq = task_rq_lock(p, &flags);
4529 * The RT priorities are set via sched_setscheduler(), but we still
4530 * allow the 'normal' nice value to be set - but as expected
4531 * it wont have any effect on scheduling until the task is
4532 * SCHED_FIFO/SCHED_RR:
4534 if (task_has_rt_policy(p)) {
4535 p->static_prio = NICE_TO_PRIO(nice);
4538 on_rq = p->se.on_rq;
4540 dequeue_task(rq, p, 0);
4542 p->static_prio = NICE_TO_PRIO(nice);
4545 p->prio = effective_prio(p);
4546 delta = p->prio - old_prio;
4549 enqueue_task(rq, p, 0);
4551 * If the task increased its priority or is running and
4552 * lowered its priority, then reschedule its CPU:
4554 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4555 resched_task(rq->curr);
4558 task_rq_unlock(rq, &flags);
4560 EXPORT_SYMBOL(set_user_nice);
4563 * can_nice - check if a task can reduce its nice value
4567 int can_nice(const struct task_struct *p, const int nice)
4569 /* convert nice value [19,-20] to rlimit style value [1,40] */
4570 int nice_rlim = 20 - nice;
4572 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4573 capable(CAP_SYS_NICE));
4576 #ifdef __ARCH_WANT_SYS_NICE
4579 * sys_nice - change the priority of the current process.
4580 * @increment: priority increment
4582 * sys_setpriority is a more generic, but much slower function that
4583 * does similar things.
4585 SYSCALL_DEFINE1(nice, int, increment)
4590 * Setpriority might change our priority at the same moment.
4591 * We don't have to worry. Conceptually one call occurs first
4592 * and we have a single winner.
4594 if (increment < -40)
4599 nice = TASK_NICE(current) + increment;
4605 if (increment < 0 && !can_nice(current, nice))
4608 retval = security_task_setnice(current, nice);
4612 set_user_nice(current, nice);
4619 * task_prio - return the priority value of a given task.
4620 * @p: the task in question.
4622 * This is the priority value as seen by users in /proc.
4623 * RT tasks are offset by -200. Normal tasks are centered
4624 * around 0, value goes from -16 to +15.
4626 int task_prio(const struct task_struct *p)
4628 return p->prio - MAX_RT_PRIO;
4632 * task_nice - return the nice value of a given task.
4633 * @p: the task in question.
4635 int task_nice(const struct task_struct *p)
4637 return TASK_NICE(p);
4639 EXPORT_SYMBOL(task_nice);
4642 * idle_cpu - is a given cpu idle currently?
4643 * @cpu: the processor in question.
4645 int idle_cpu(int cpu)
4647 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4651 * idle_task - return the idle task for a given cpu.
4652 * @cpu: the processor in question.
4654 struct task_struct *idle_task(int cpu)
4656 return cpu_rq(cpu)->idle;
4660 * find_process_by_pid - find a process with a matching PID value.
4661 * @pid: the pid in question.
4663 static struct task_struct *find_process_by_pid(pid_t pid)
4665 return pid ? find_task_by_vpid(pid) : current;
4668 /* Actually do priority change: must hold rq lock. */
4670 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4672 BUG_ON(p->se.on_rq);
4675 p->rt_priority = prio;
4676 p->normal_prio = normal_prio(p);
4677 /* we are holding p->pi_lock already */
4678 p->prio = rt_mutex_getprio(p);
4679 if (rt_prio(p->prio))
4680 p->sched_class = &rt_sched_class;
4682 p->sched_class = &fair_sched_class;
4687 * check the target process has a UID that matches the current process's
4689 static bool check_same_owner(struct task_struct *p)
4691 const struct cred *cred = current_cred(), *pcred;
4695 pcred = __task_cred(p);
4696 match = (cred->euid == pcred->euid ||
4697 cred->euid == pcred->uid);
4702 static int __sched_setscheduler(struct task_struct *p, int policy,
4703 struct sched_param *param, bool user)
4705 int retval, oldprio, oldpolicy = -1, on_rq, running;
4706 unsigned long flags;
4707 const struct sched_class *prev_class;
4711 /* may grab non-irq protected spin_locks */
4712 BUG_ON(in_interrupt());
4714 /* double check policy once rq lock held */
4716 reset_on_fork = p->sched_reset_on_fork;
4717 policy = oldpolicy = p->policy;
4719 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4720 policy &= ~SCHED_RESET_ON_FORK;
4722 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4723 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4724 policy != SCHED_IDLE)
4729 * Valid priorities for SCHED_FIFO and SCHED_RR are
4730 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4731 * SCHED_BATCH and SCHED_IDLE is 0.
4733 if (param->sched_priority < 0 ||
4734 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4735 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4737 if (rt_policy(policy) != (param->sched_priority != 0))
4741 * Allow unprivileged RT tasks to decrease priority:
4743 if (user && !capable(CAP_SYS_NICE)) {
4744 if (rt_policy(policy)) {
4745 unsigned long rlim_rtprio =
4746 task_rlimit(p, RLIMIT_RTPRIO);
4748 /* can't set/change the rt policy */
4749 if (policy != p->policy && !rlim_rtprio)
4752 /* can't increase priority */
4753 if (param->sched_priority > p->rt_priority &&
4754 param->sched_priority > rlim_rtprio)
4758 * Like positive nice levels, dont allow tasks to
4759 * move out of SCHED_IDLE either:
4761 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4764 /* can't change other user's priorities */
4765 if (!check_same_owner(p))
4768 /* Normal users shall not reset the sched_reset_on_fork flag */
4769 if (p->sched_reset_on_fork && !reset_on_fork)
4774 retval = security_task_setscheduler(p, policy, param);
4780 * make sure no PI-waiters arrive (or leave) while we are
4781 * changing the priority of the task:
4783 raw_spin_lock_irqsave(&p->pi_lock, flags);
4785 * To be able to change p->policy safely, the apropriate
4786 * runqueue lock must be held.
4788 rq = __task_rq_lock(p);
4791 * Changing the policy of the stop threads its a very bad idea
4793 if (p == rq->stop) {
4794 __task_rq_unlock(rq);
4795 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4799 #ifdef CONFIG_RT_GROUP_SCHED
4802 * Do not allow realtime tasks into groups that have no runtime
4805 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4806 task_group(p)->rt_bandwidth.rt_runtime == 0) {
4807 __task_rq_unlock(rq);
4808 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4814 /* recheck policy now with rq lock held */
4815 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4816 policy = oldpolicy = -1;
4817 __task_rq_unlock(rq);
4818 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4821 on_rq = p->se.on_rq;
4822 running = task_current(rq, p);
4824 deactivate_task(rq, p, 0);
4826 p->sched_class->put_prev_task(rq, p);
4828 p->sched_reset_on_fork = reset_on_fork;
4831 prev_class = p->sched_class;
4832 __setscheduler(rq, p, policy, param->sched_priority);
4835 p->sched_class->set_curr_task(rq);
4837 activate_task(rq, p, 0);
4839 check_class_changed(rq, p, prev_class, oldprio, running);
4841 __task_rq_unlock(rq);
4842 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4844 rt_mutex_adjust_pi(p);
4850 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4851 * @p: the task in question.
4852 * @policy: new policy.
4853 * @param: structure containing the new RT priority.
4855 * NOTE that the task may be already dead.
4857 int sched_setscheduler(struct task_struct *p, int policy,
4858 struct sched_param *param)
4860 return __sched_setscheduler(p, policy, param, true);
4862 EXPORT_SYMBOL_GPL(sched_setscheduler);
4865 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4866 * @p: the task in question.
4867 * @policy: new policy.
4868 * @param: structure containing the new RT priority.
4870 * Just like sched_setscheduler, only don't bother checking if the
4871 * current context has permission. For example, this is needed in
4872 * stop_machine(): we create temporary high priority worker threads,
4873 * but our caller might not have that capability.
4875 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4876 struct sched_param *param)
4878 return __sched_setscheduler(p, policy, param, false);
4882 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4884 struct sched_param lparam;
4885 struct task_struct *p;
4888 if (!param || pid < 0)
4890 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4895 p = find_process_by_pid(pid);
4897 retval = sched_setscheduler(p, policy, &lparam);
4904 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4905 * @pid: the pid in question.
4906 * @policy: new policy.
4907 * @param: structure containing the new RT priority.
4909 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4910 struct sched_param __user *, param)
4912 /* negative values for policy are not valid */
4916 return do_sched_setscheduler(pid, policy, param);
4920 * sys_sched_setparam - set/change the RT priority of a thread
4921 * @pid: the pid in question.
4922 * @param: structure containing the new RT priority.
4924 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4926 return do_sched_setscheduler(pid, -1, param);
4930 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4931 * @pid: the pid in question.
4933 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4935 struct task_struct *p;
4943 p = find_process_by_pid(pid);
4945 retval = security_task_getscheduler(p);
4948 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4955 * sys_sched_getparam - get the RT priority of a thread
4956 * @pid: the pid in question.
4957 * @param: structure containing the RT priority.
4959 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4961 struct sched_param lp;
4962 struct task_struct *p;
4965 if (!param || pid < 0)
4969 p = find_process_by_pid(pid);
4974 retval = security_task_getscheduler(p);
4978 lp.sched_priority = p->rt_priority;
4982 * This one might sleep, we cannot do it with a spinlock held ...
4984 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4993 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4995 cpumask_var_t cpus_allowed, new_mask;
4996 struct task_struct *p;
5002 p = find_process_by_pid(pid);
5009 /* Prevent p going away */
5013 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5017 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5019 goto out_free_cpus_allowed;
5022 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5025 retval = security_task_setscheduler(p, 0, NULL);
5029 cpuset_cpus_allowed(p, cpus_allowed);
5030 cpumask_and(new_mask, in_mask, cpus_allowed);
5032 retval = set_cpus_allowed_ptr(p, new_mask);
5035 cpuset_cpus_allowed(p, cpus_allowed);
5036 if (!cpumask_subset(new_mask, cpus_allowed)) {
5038 * We must have raced with a concurrent cpuset
5039 * update. Just reset the cpus_allowed to the
5040 * cpuset's cpus_allowed
5042 cpumask_copy(new_mask, cpus_allowed);
5047 free_cpumask_var(new_mask);
5048 out_free_cpus_allowed:
5049 free_cpumask_var(cpus_allowed);
5056 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5057 struct cpumask *new_mask)
5059 if (len < cpumask_size())
5060 cpumask_clear(new_mask);
5061 else if (len > cpumask_size())
5062 len = cpumask_size();
5064 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5068 * sys_sched_setaffinity - set the cpu affinity of a process
5069 * @pid: pid of the process
5070 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5071 * @user_mask_ptr: user-space pointer to the new cpu mask
5073 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5074 unsigned long __user *, user_mask_ptr)
5076 cpumask_var_t new_mask;
5079 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5082 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5084 retval = sched_setaffinity(pid, new_mask);
5085 free_cpumask_var(new_mask);
5089 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5091 struct task_struct *p;
5092 unsigned long flags;
5100 p = find_process_by_pid(pid);
5104 retval = security_task_getscheduler(p);
5108 rq = task_rq_lock(p, &flags);
5109 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5110 task_rq_unlock(rq, &flags);
5120 * sys_sched_getaffinity - get the cpu affinity of a process
5121 * @pid: pid of the process
5122 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5123 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5125 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5126 unsigned long __user *, user_mask_ptr)
5131 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5133 if (len & (sizeof(unsigned long)-1))
5136 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5139 ret = sched_getaffinity(pid, mask);
5141 size_t retlen = min_t(size_t, len, cpumask_size());
5143 if (copy_to_user(user_mask_ptr, mask, retlen))
5148 free_cpumask_var(mask);
5154 * sys_sched_yield - yield the current processor to other threads.
5156 * This function yields the current CPU to other tasks. If there are no
5157 * other threads running on this CPU then this function will return.
5159 SYSCALL_DEFINE0(sched_yield)
5161 struct rq *rq = this_rq_lock();
5163 schedstat_inc(rq, yld_count);
5164 current->sched_class->yield_task(rq);
5167 * Since we are going to call schedule() anyway, there's
5168 * no need to preempt or enable interrupts:
5170 __release(rq->lock);
5171 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5172 do_raw_spin_unlock(&rq->lock);
5173 preempt_enable_no_resched();
5180 static inline int should_resched(void)
5182 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5185 static void __cond_resched(void)
5187 add_preempt_count(PREEMPT_ACTIVE);
5189 sub_preempt_count(PREEMPT_ACTIVE);
5192 int __sched _cond_resched(void)
5194 if (should_resched()) {
5200 EXPORT_SYMBOL(_cond_resched);
5203 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5204 * call schedule, and on return reacquire the lock.
5206 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5207 * operations here to prevent schedule() from being called twice (once via
5208 * spin_unlock(), once by hand).
5210 int __cond_resched_lock(spinlock_t *lock)
5212 int resched = should_resched();
5215 lockdep_assert_held(lock);
5217 if (spin_needbreak(lock) || resched) {
5228 EXPORT_SYMBOL(__cond_resched_lock);
5230 int __sched __cond_resched_softirq(void)
5232 BUG_ON(!in_softirq());
5234 if (should_resched()) {
5242 EXPORT_SYMBOL(__cond_resched_softirq);
5245 * yield - yield the current processor to other threads.
5247 * This is a shortcut for kernel-space yielding - it marks the
5248 * thread runnable and calls sys_sched_yield().
5250 void __sched yield(void)
5252 set_current_state(TASK_RUNNING);
5255 EXPORT_SYMBOL(yield);
5258 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5259 * that process accounting knows that this is a task in IO wait state.
5261 void __sched io_schedule(void)
5263 struct rq *rq = raw_rq();
5265 delayacct_blkio_start();
5266 atomic_inc(&rq->nr_iowait);
5267 current->in_iowait = 1;
5269 current->in_iowait = 0;
5270 atomic_dec(&rq->nr_iowait);
5271 delayacct_blkio_end();
5273 EXPORT_SYMBOL(io_schedule);
5275 long __sched io_schedule_timeout(long timeout)
5277 struct rq *rq = raw_rq();
5280 delayacct_blkio_start();
5281 atomic_inc(&rq->nr_iowait);
5282 current->in_iowait = 1;
5283 ret = schedule_timeout(timeout);
5284 current->in_iowait = 0;
5285 atomic_dec(&rq->nr_iowait);
5286 delayacct_blkio_end();
5291 * sys_sched_get_priority_max - return maximum RT priority.
5292 * @policy: scheduling class.
5294 * this syscall returns the maximum rt_priority that can be used
5295 * by a given scheduling class.
5297 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5304 ret = MAX_USER_RT_PRIO-1;
5316 * sys_sched_get_priority_min - return minimum RT priority.
5317 * @policy: scheduling class.
5319 * this syscall returns the minimum rt_priority that can be used
5320 * by a given scheduling class.
5322 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5340 * sys_sched_rr_get_interval - return the default timeslice of a process.
5341 * @pid: pid of the process.
5342 * @interval: userspace pointer to the timeslice value.
5344 * this syscall writes the default timeslice value of a given process
5345 * into the user-space timespec buffer. A value of '0' means infinity.
5347 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5348 struct timespec __user *, interval)
5350 struct task_struct *p;
5351 unsigned int time_slice;
5352 unsigned long flags;
5362 p = find_process_by_pid(pid);
5366 retval = security_task_getscheduler(p);
5370 rq = task_rq_lock(p, &flags);
5371 time_slice = p->sched_class->get_rr_interval(rq, p);
5372 task_rq_unlock(rq, &flags);
5375 jiffies_to_timespec(time_slice, &t);
5376 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5384 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5386 void sched_show_task(struct task_struct *p)
5388 unsigned long free = 0;
5391 state = p->state ? __ffs(p->state) + 1 : 0;
5392 printk(KERN_INFO "%-13.13s %c", p->comm,
5393 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5394 #if BITS_PER_LONG == 32
5395 if (state == TASK_RUNNING)
5396 printk(KERN_CONT " running ");
5398 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5400 if (state == TASK_RUNNING)
5401 printk(KERN_CONT " running task ");
5403 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5405 #ifdef CONFIG_DEBUG_STACK_USAGE
5406 free = stack_not_used(p);
5408 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5409 task_pid_nr(p), task_pid_nr(p->real_parent),
5410 (unsigned long)task_thread_info(p)->flags);
5412 show_stack(p, NULL);
5415 void show_state_filter(unsigned long state_filter)
5417 struct task_struct *g, *p;
5419 #if BITS_PER_LONG == 32
5421 " task PC stack pid father\n");
5424 " task PC stack pid father\n");
5426 read_lock(&tasklist_lock);
5427 do_each_thread(g, p) {
5429 * reset the NMI-timeout, listing all files on a slow
5430 * console might take alot of time:
5432 touch_nmi_watchdog();
5433 if (!state_filter || (p->state & state_filter))
5435 } while_each_thread(g, p);
5437 touch_all_softlockup_watchdogs();
5439 #ifdef CONFIG_SCHED_DEBUG
5440 sysrq_sched_debug_show();
5442 read_unlock(&tasklist_lock);
5444 * Only show locks if all tasks are dumped:
5447 debug_show_all_locks();
5450 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5452 idle->sched_class = &idle_sched_class;
5456 * init_idle - set up an idle thread for a given CPU
5457 * @idle: task in question
5458 * @cpu: cpu the idle task belongs to
5460 * NOTE: this function does not set the idle thread's NEED_RESCHED
5461 * flag, to make booting more robust.
5463 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5465 struct rq *rq = cpu_rq(cpu);
5466 unsigned long flags;
5468 raw_spin_lock_irqsave(&rq->lock, flags);
5471 idle->state = TASK_RUNNING;
5472 idle->se.exec_start = sched_clock();
5474 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5475 __set_task_cpu(idle, cpu);
5477 rq->curr = rq->idle = idle;
5478 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5481 raw_spin_unlock_irqrestore(&rq->lock, flags);
5483 /* Set the preempt count _outside_ the spinlocks! */
5484 #if defined(CONFIG_PREEMPT)
5485 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5487 task_thread_info(idle)->preempt_count = 0;
5490 * The idle tasks have their own, simple scheduling class:
5492 idle->sched_class = &idle_sched_class;
5493 ftrace_graph_init_task(idle);
5497 * In a system that switches off the HZ timer nohz_cpu_mask
5498 * indicates which cpus entered this state. This is used
5499 * in the rcu update to wait only for active cpus. For system
5500 * which do not switch off the HZ timer nohz_cpu_mask should
5501 * always be CPU_BITS_NONE.
5503 cpumask_var_t nohz_cpu_mask;
5506 * Increase the granularity value when there are more CPUs,
5507 * because with more CPUs the 'effective latency' as visible
5508 * to users decreases. But the relationship is not linear,
5509 * so pick a second-best guess by going with the log2 of the
5512 * This idea comes from the SD scheduler of Con Kolivas:
5514 static int get_update_sysctl_factor(void)
5516 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5517 unsigned int factor;
5519 switch (sysctl_sched_tunable_scaling) {
5520 case SCHED_TUNABLESCALING_NONE:
5523 case SCHED_TUNABLESCALING_LINEAR:
5526 case SCHED_TUNABLESCALING_LOG:
5528 factor = 1 + ilog2(cpus);
5535 static void update_sysctl(void)
5537 unsigned int factor = get_update_sysctl_factor();
5539 #define SET_SYSCTL(name) \
5540 (sysctl_##name = (factor) * normalized_sysctl_##name)
5541 SET_SYSCTL(sched_min_granularity);
5542 SET_SYSCTL(sched_latency);
5543 SET_SYSCTL(sched_wakeup_granularity);
5544 SET_SYSCTL(sched_shares_ratelimit);
5548 static inline void sched_init_granularity(void)
5555 * This is how migration works:
5557 * 1) we invoke migration_cpu_stop() on the target CPU using
5559 * 2) stopper starts to run (implicitly forcing the migrated thread
5561 * 3) it checks whether the migrated task is still in the wrong runqueue.
5562 * 4) if it's in the wrong runqueue then the migration thread removes
5563 * it and puts it into the right queue.
5564 * 5) stopper completes and stop_one_cpu() returns and the migration
5569 * Change a given task's CPU affinity. Migrate the thread to a
5570 * proper CPU and schedule it away if the CPU it's executing on
5571 * is removed from the allowed bitmask.
5573 * NOTE: the caller must have a valid reference to the task, the
5574 * task must not exit() & deallocate itself prematurely. The
5575 * call is not atomic; no spinlocks may be held.
5577 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5579 unsigned long flags;
5581 unsigned int dest_cpu;
5585 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5586 * drop the rq->lock and still rely on ->cpus_allowed.
5589 while (task_is_waking(p))
5591 rq = task_rq_lock(p, &flags);
5592 if (task_is_waking(p)) {
5593 task_rq_unlock(rq, &flags);
5597 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5602 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5603 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5608 if (p->sched_class->set_cpus_allowed)
5609 p->sched_class->set_cpus_allowed(p, new_mask);
5611 cpumask_copy(&p->cpus_allowed, new_mask);
5612 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5615 /* Can the task run on the task's current CPU? If so, we're done */
5616 if (cpumask_test_cpu(task_cpu(p), new_mask))
5619 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5620 if (migrate_task(p, dest_cpu)) {
5621 struct migration_arg arg = { p, dest_cpu };
5622 /* Need help from migration thread: drop lock and wait. */
5623 task_rq_unlock(rq, &flags);
5624 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5625 tlb_migrate_finish(p->mm);
5629 task_rq_unlock(rq, &flags);
5633 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5636 * Move (not current) task off this cpu, onto dest cpu. We're doing
5637 * this because either it can't run here any more (set_cpus_allowed()
5638 * away from this CPU, or CPU going down), or because we're
5639 * attempting to rebalance this task on exec (sched_exec).
5641 * So we race with normal scheduler movements, but that's OK, as long
5642 * as the task is no longer on this CPU.
5644 * Returns non-zero if task was successfully migrated.
5646 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5648 struct rq *rq_dest, *rq_src;
5651 if (unlikely(!cpu_active(dest_cpu)))
5654 rq_src = cpu_rq(src_cpu);
5655 rq_dest = cpu_rq(dest_cpu);
5657 double_rq_lock(rq_src, rq_dest);
5658 /* Already moved. */
5659 if (task_cpu(p) != src_cpu)
5661 /* Affinity changed (again). */
5662 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5666 * If we're not on a rq, the next wake-up will ensure we're
5670 deactivate_task(rq_src, p, 0);
5671 set_task_cpu(p, dest_cpu);
5672 activate_task(rq_dest, p, 0);
5673 check_preempt_curr(rq_dest, p, 0);
5678 double_rq_unlock(rq_src, rq_dest);
5683 * migration_cpu_stop - this will be executed by a highprio stopper thread
5684 * and performs thread migration by bumping thread off CPU then
5685 * 'pushing' onto another runqueue.
5687 static int migration_cpu_stop(void *data)
5689 struct migration_arg *arg = data;
5692 * The original target cpu might have gone down and we might
5693 * be on another cpu but it doesn't matter.
5695 local_irq_disable();
5696 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5701 #ifdef CONFIG_HOTPLUG_CPU
5703 * Figure out where task on dead CPU should go, use force if necessary.
5705 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5707 struct rq *rq = cpu_rq(dead_cpu);
5708 int needs_cpu, uninitialized_var(dest_cpu);
5709 unsigned long flags;
5711 local_irq_save(flags);
5713 raw_spin_lock(&rq->lock);
5714 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5716 dest_cpu = select_fallback_rq(dead_cpu, p);
5717 raw_spin_unlock(&rq->lock);
5719 * It can only fail if we race with set_cpus_allowed(),
5720 * in the racer should migrate the task anyway.
5723 __migrate_task(p, dead_cpu, dest_cpu);
5724 local_irq_restore(flags);
5728 * While a dead CPU has no uninterruptible tasks queued at this point,
5729 * it might still have a nonzero ->nr_uninterruptible counter, because
5730 * for performance reasons the counter is not stricly tracking tasks to
5731 * their home CPUs. So we just add the counter to another CPU's counter,
5732 * to keep the global sum constant after CPU-down:
5734 static void migrate_nr_uninterruptible(struct rq *rq_src)
5736 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5737 unsigned long flags;
5739 local_irq_save(flags);
5740 double_rq_lock(rq_src, rq_dest);
5741 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5742 rq_src->nr_uninterruptible = 0;
5743 double_rq_unlock(rq_src, rq_dest);
5744 local_irq_restore(flags);
5747 /* Run through task list and migrate tasks from the dead cpu. */
5748 static void migrate_live_tasks(int src_cpu)
5750 struct task_struct *p, *t;
5752 read_lock(&tasklist_lock);
5754 do_each_thread(t, p) {
5758 if (task_cpu(p) == src_cpu)
5759 move_task_off_dead_cpu(src_cpu, p);
5760 } while_each_thread(t, p);
5762 read_unlock(&tasklist_lock);
5766 * Schedules idle task to be the next runnable task on current CPU.
5767 * It does so by boosting its priority to highest possible.
5768 * Used by CPU offline code.
5770 void sched_idle_next(void)
5772 int this_cpu = smp_processor_id();
5773 struct rq *rq = cpu_rq(this_cpu);
5774 struct task_struct *p = rq->idle;
5775 unsigned long flags;
5777 /* cpu has to be offline */
5778 BUG_ON(cpu_online(this_cpu));
5781 * Strictly not necessary since rest of the CPUs are stopped by now
5782 * and interrupts disabled on the current cpu.
5784 raw_spin_lock_irqsave(&rq->lock, flags);
5786 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5788 activate_task(rq, p, 0);
5790 raw_spin_unlock_irqrestore(&rq->lock, flags);
5794 * Ensures that the idle task is using init_mm right before its cpu goes
5797 void idle_task_exit(void)
5799 struct mm_struct *mm = current->active_mm;
5801 BUG_ON(cpu_online(smp_processor_id()));
5804 switch_mm(mm, &init_mm, current);
5808 /* called under rq->lock with disabled interrupts */
5809 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5811 struct rq *rq = cpu_rq(dead_cpu);
5813 /* Must be exiting, otherwise would be on tasklist. */
5814 BUG_ON(!p->exit_state);
5816 /* Cannot have done final schedule yet: would have vanished. */
5817 BUG_ON(p->state == TASK_DEAD);
5822 * Drop lock around migration; if someone else moves it,
5823 * that's OK. No task can be added to this CPU, so iteration is
5826 raw_spin_unlock_irq(&rq->lock);
5827 move_task_off_dead_cpu(dead_cpu, p);
5828 raw_spin_lock_irq(&rq->lock);
5833 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5834 static void migrate_dead_tasks(unsigned int dead_cpu)
5836 struct rq *rq = cpu_rq(dead_cpu);
5837 struct task_struct *next;
5840 if (!rq->nr_running)
5842 next = pick_next_task(rq);
5845 next->sched_class->put_prev_task(rq, next);
5846 migrate_dead(dead_cpu, next);
5852 * remove the tasks which were accounted by rq from calc_load_tasks.
5854 static void calc_global_load_remove(struct rq *rq)
5856 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5857 rq->calc_load_active = 0;
5859 #endif /* CONFIG_HOTPLUG_CPU */
5861 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5863 static struct ctl_table sd_ctl_dir[] = {
5865 .procname = "sched_domain",
5871 static struct ctl_table sd_ctl_root[] = {
5873 .procname = "kernel",
5875 .child = sd_ctl_dir,
5880 static struct ctl_table *sd_alloc_ctl_entry(int n)
5882 struct ctl_table *entry =
5883 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5888 static void sd_free_ctl_entry(struct ctl_table **tablep)
5890 struct ctl_table *entry;
5893 * In the intermediate directories, both the child directory and
5894 * procname are dynamically allocated and could fail but the mode
5895 * will always be set. In the lowest directory the names are
5896 * static strings and all have proc handlers.
5898 for (entry = *tablep; entry->mode; entry++) {
5900 sd_free_ctl_entry(&entry->child);
5901 if (entry->proc_handler == NULL)
5902 kfree(entry->procname);
5910 set_table_entry(struct ctl_table *entry,
5911 const char *procname, void *data, int maxlen,
5912 mode_t mode, proc_handler *proc_handler)
5914 entry->procname = procname;
5916 entry->maxlen = maxlen;
5918 entry->proc_handler = proc_handler;
5921 static struct ctl_table *
5922 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5924 struct ctl_table *table = sd_alloc_ctl_entry(13);
5929 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5930 sizeof(long), 0644, proc_doulongvec_minmax);
5931 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5932 sizeof(long), 0644, proc_doulongvec_minmax);
5933 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5934 sizeof(int), 0644, proc_dointvec_minmax);
5935 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5936 sizeof(int), 0644, proc_dointvec_minmax);
5937 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5938 sizeof(int), 0644, proc_dointvec_minmax);
5939 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5940 sizeof(int), 0644, proc_dointvec_minmax);
5941 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5942 sizeof(int), 0644, proc_dointvec_minmax);
5943 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5944 sizeof(int), 0644, proc_dointvec_minmax);
5945 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5946 sizeof(int), 0644, proc_dointvec_minmax);
5947 set_table_entry(&table[9], "cache_nice_tries",
5948 &sd->cache_nice_tries,
5949 sizeof(int), 0644, proc_dointvec_minmax);
5950 set_table_entry(&table[10], "flags", &sd->flags,
5951 sizeof(int), 0644, proc_dointvec_minmax);
5952 set_table_entry(&table[11], "name", sd->name,
5953 CORENAME_MAX_SIZE, 0444, proc_dostring);
5954 /* &table[12] is terminator */
5959 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5961 struct ctl_table *entry, *table;
5962 struct sched_domain *sd;
5963 int domain_num = 0, i;
5966 for_each_domain(cpu, sd)
5968 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5973 for_each_domain(cpu, sd) {
5974 snprintf(buf, 32, "domain%d", i);
5975 entry->procname = kstrdup(buf, GFP_KERNEL);
5977 entry->child = sd_alloc_ctl_domain_table(sd);
5984 static struct ctl_table_header *sd_sysctl_header;
5985 static void register_sched_domain_sysctl(void)
5987 int i, cpu_num = num_possible_cpus();
5988 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5991 WARN_ON(sd_ctl_dir[0].child);
5992 sd_ctl_dir[0].child = entry;
5997 for_each_possible_cpu(i) {
5998 snprintf(buf, 32, "cpu%d", i);
5999 entry->procname = kstrdup(buf, GFP_KERNEL);
6001 entry->child = sd_alloc_ctl_cpu_table(i);
6005 WARN_ON(sd_sysctl_header);
6006 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6009 /* may be called multiple times per register */
6010 static void unregister_sched_domain_sysctl(void)
6012 if (sd_sysctl_header)
6013 unregister_sysctl_table(sd_sysctl_header);
6014 sd_sysctl_header = NULL;
6015 if (sd_ctl_dir[0].child)
6016 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6019 static void register_sched_domain_sysctl(void)
6022 static void unregister_sched_domain_sysctl(void)
6027 static void set_rq_online(struct rq *rq)
6030 const struct sched_class *class;
6032 cpumask_set_cpu(rq->cpu, rq->rd->online);
6035 for_each_class(class) {
6036 if (class->rq_online)
6037 class->rq_online(rq);
6042 static void set_rq_offline(struct rq *rq)
6045 const struct sched_class *class;
6047 for_each_class(class) {
6048 if (class->rq_offline)
6049 class->rq_offline(rq);
6052 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6058 * migration_call - callback that gets triggered when a CPU is added.
6059 * Here we can start up the necessary migration thread for the new CPU.
6061 static int __cpuinit
6062 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6064 int cpu = (long)hcpu;
6065 unsigned long flags;
6066 struct rq *rq = cpu_rq(cpu);
6070 case CPU_UP_PREPARE:
6071 case CPU_UP_PREPARE_FROZEN:
6072 rq->calc_load_update = calc_load_update;
6076 case CPU_ONLINE_FROZEN:
6077 /* Update our root-domain */
6078 raw_spin_lock_irqsave(&rq->lock, flags);
6080 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6084 raw_spin_unlock_irqrestore(&rq->lock, flags);
6087 #ifdef CONFIG_HOTPLUG_CPU
6089 case CPU_DEAD_FROZEN:
6090 migrate_live_tasks(cpu);
6091 /* Idle task back to normal (off runqueue, low prio) */
6092 raw_spin_lock_irq(&rq->lock);
6093 deactivate_task(rq, rq->idle, 0);
6094 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6095 rq->idle->sched_class = &idle_sched_class;
6096 migrate_dead_tasks(cpu);
6097 raw_spin_unlock_irq(&rq->lock);
6098 migrate_nr_uninterruptible(rq);
6099 BUG_ON(rq->nr_running != 0);
6100 calc_global_load_remove(rq);
6104 case CPU_DYING_FROZEN:
6105 /* Update our root-domain */
6106 raw_spin_lock_irqsave(&rq->lock, flags);
6108 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6111 raw_spin_unlock_irqrestore(&rq->lock, flags);
6119 * Register at high priority so that task migration (migrate_all_tasks)
6120 * happens before everything else. This has to be lower priority than
6121 * the notifier in the perf_event subsystem, though.
6123 static struct notifier_block __cpuinitdata migration_notifier = {
6124 .notifier_call = migration_call,
6125 .priority = CPU_PRI_MIGRATION,
6128 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6129 unsigned long action, void *hcpu)
6131 switch (action & ~CPU_TASKS_FROZEN) {
6133 case CPU_DOWN_FAILED:
6134 set_cpu_active((long)hcpu, true);
6141 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6142 unsigned long action, void *hcpu)
6144 switch (action & ~CPU_TASKS_FROZEN) {
6145 case CPU_DOWN_PREPARE:
6146 set_cpu_active((long)hcpu, false);
6153 static int __init migration_init(void)
6155 void *cpu = (void *)(long)smp_processor_id();
6158 /* Initialize migration for the boot CPU */
6159 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6160 BUG_ON(err == NOTIFY_BAD);
6161 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6162 register_cpu_notifier(&migration_notifier);
6164 /* Register cpu active notifiers */
6165 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6166 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6170 early_initcall(migration_init);
6175 #ifdef CONFIG_SCHED_DEBUG
6177 static __read_mostly int sched_domain_debug_enabled;
6179 static int __init sched_domain_debug_setup(char *str)
6181 sched_domain_debug_enabled = 1;
6185 early_param("sched_debug", sched_domain_debug_setup);
6187 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6188 struct cpumask *groupmask)
6190 struct sched_group *group = sd->groups;
6193 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6194 cpumask_clear(groupmask);
6196 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6198 if (!(sd->flags & SD_LOAD_BALANCE)) {
6199 printk("does not load-balance\n");
6201 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6206 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6208 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6209 printk(KERN_ERR "ERROR: domain->span does not contain "
6212 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6213 printk(KERN_ERR "ERROR: domain->groups does not contain"
6217 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6221 printk(KERN_ERR "ERROR: group is NULL\n");
6225 if (!group->cpu_power) {
6226 printk(KERN_CONT "\n");
6227 printk(KERN_ERR "ERROR: domain->cpu_power not "
6232 if (!cpumask_weight(sched_group_cpus(group))) {
6233 printk(KERN_CONT "\n");
6234 printk(KERN_ERR "ERROR: empty group\n");
6238 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6239 printk(KERN_CONT "\n");
6240 printk(KERN_ERR "ERROR: repeated CPUs\n");
6244 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6246 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6248 printk(KERN_CONT " %s", str);
6249 if (group->cpu_power != SCHED_LOAD_SCALE) {
6250 printk(KERN_CONT " (cpu_power = %d)",
6254 group = group->next;
6255 } while (group != sd->groups);
6256 printk(KERN_CONT "\n");
6258 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6259 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6262 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6263 printk(KERN_ERR "ERROR: parent span is not a superset "
6264 "of domain->span\n");
6268 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6270 cpumask_var_t groupmask;
6273 if (!sched_domain_debug_enabled)
6277 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6281 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6283 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6284 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6289 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6296 free_cpumask_var(groupmask);
6298 #else /* !CONFIG_SCHED_DEBUG */
6299 # define sched_domain_debug(sd, cpu) do { } while (0)
6300 #endif /* CONFIG_SCHED_DEBUG */
6302 static int sd_degenerate(struct sched_domain *sd)
6304 if (cpumask_weight(sched_domain_span(sd)) == 1)
6307 /* Following flags need at least 2 groups */
6308 if (sd->flags & (SD_LOAD_BALANCE |
6309 SD_BALANCE_NEWIDLE |
6313 SD_SHARE_PKG_RESOURCES)) {
6314 if (sd->groups != sd->groups->next)
6318 /* Following flags don't use groups */
6319 if (sd->flags & (SD_WAKE_AFFINE))
6326 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6328 unsigned long cflags = sd->flags, pflags = parent->flags;
6330 if (sd_degenerate(parent))
6333 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6336 /* Flags needing groups don't count if only 1 group in parent */
6337 if (parent->groups == parent->groups->next) {
6338 pflags &= ~(SD_LOAD_BALANCE |
6339 SD_BALANCE_NEWIDLE |
6343 SD_SHARE_PKG_RESOURCES);
6344 if (nr_node_ids == 1)
6345 pflags &= ~SD_SERIALIZE;
6347 if (~cflags & pflags)
6353 static void free_rootdomain(struct root_domain *rd)
6355 synchronize_sched();
6357 cpupri_cleanup(&rd->cpupri);
6359 free_cpumask_var(rd->rto_mask);
6360 free_cpumask_var(rd->online);
6361 free_cpumask_var(rd->span);
6365 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6367 struct root_domain *old_rd = NULL;
6368 unsigned long flags;
6370 raw_spin_lock_irqsave(&rq->lock, flags);
6375 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6378 cpumask_clear_cpu(rq->cpu, old_rd->span);
6381 * If we dont want to free the old_rt yet then
6382 * set old_rd to NULL to skip the freeing later
6385 if (!atomic_dec_and_test(&old_rd->refcount))
6389 atomic_inc(&rd->refcount);
6392 cpumask_set_cpu(rq->cpu, rd->span);
6393 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6396 raw_spin_unlock_irqrestore(&rq->lock, flags);
6399 free_rootdomain(old_rd);
6402 static int init_rootdomain(struct root_domain *rd)
6404 memset(rd, 0, sizeof(*rd));
6406 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6408 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6410 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6413 if (cpupri_init(&rd->cpupri) != 0)
6418 free_cpumask_var(rd->rto_mask);
6420 free_cpumask_var(rd->online);
6422 free_cpumask_var(rd->span);
6427 static void init_defrootdomain(void)
6429 init_rootdomain(&def_root_domain);
6431 atomic_set(&def_root_domain.refcount, 1);
6434 static struct root_domain *alloc_rootdomain(void)
6436 struct root_domain *rd;
6438 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6442 if (init_rootdomain(rd) != 0) {
6451 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6452 * hold the hotplug lock.
6455 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6457 struct rq *rq = cpu_rq(cpu);
6458 struct sched_domain *tmp;
6460 for (tmp = sd; tmp; tmp = tmp->parent)
6461 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6463 /* Remove the sched domains which do not contribute to scheduling. */
6464 for (tmp = sd; tmp; ) {
6465 struct sched_domain *parent = tmp->parent;
6469 if (sd_parent_degenerate(tmp, parent)) {
6470 tmp->parent = parent->parent;
6472 parent->parent->child = tmp;
6477 if (sd && sd_degenerate(sd)) {
6483 sched_domain_debug(sd, cpu);
6485 rq_attach_root(rq, rd);
6486 rcu_assign_pointer(rq->sd, sd);
6489 /* cpus with isolated domains */
6490 static cpumask_var_t cpu_isolated_map;
6492 /* Setup the mask of cpus configured for isolated domains */
6493 static int __init isolated_cpu_setup(char *str)
6495 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6496 cpulist_parse(str, cpu_isolated_map);
6500 __setup("isolcpus=", isolated_cpu_setup);
6503 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6504 * to a function which identifies what group(along with sched group) a CPU
6505 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6506 * (due to the fact that we keep track of groups covered with a struct cpumask).
6508 * init_sched_build_groups will build a circular linked list of the groups
6509 * covered by the given span, and will set each group's ->cpumask correctly,
6510 * and ->cpu_power to 0.
6513 init_sched_build_groups(const struct cpumask *span,
6514 const struct cpumask *cpu_map,
6515 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6516 struct sched_group **sg,
6517 struct cpumask *tmpmask),
6518 struct cpumask *covered, struct cpumask *tmpmask)
6520 struct sched_group *first = NULL, *last = NULL;
6523 cpumask_clear(covered);
6525 for_each_cpu(i, span) {
6526 struct sched_group *sg;
6527 int group = group_fn(i, cpu_map, &sg, tmpmask);
6530 if (cpumask_test_cpu(i, covered))
6533 cpumask_clear(sched_group_cpus(sg));
6536 for_each_cpu(j, span) {
6537 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6540 cpumask_set_cpu(j, covered);
6541 cpumask_set_cpu(j, sched_group_cpus(sg));
6552 #define SD_NODES_PER_DOMAIN 16
6557 * find_next_best_node - find the next node to include in a sched_domain
6558 * @node: node whose sched_domain we're building
6559 * @used_nodes: nodes already in the sched_domain
6561 * Find the next node to include in a given scheduling domain. Simply
6562 * finds the closest node not already in the @used_nodes map.
6564 * Should use nodemask_t.
6566 static int find_next_best_node(int node, nodemask_t *used_nodes)
6568 int i, n, val, min_val, best_node = 0;
6572 for (i = 0; i < nr_node_ids; i++) {
6573 /* Start at @node */
6574 n = (node + i) % nr_node_ids;
6576 if (!nr_cpus_node(n))
6579 /* Skip already used nodes */
6580 if (node_isset(n, *used_nodes))
6583 /* Simple min distance search */
6584 val = node_distance(node, n);
6586 if (val < min_val) {
6592 node_set(best_node, *used_nodes);
6597 * sched_domain_node_span - get a cpumask for a node's sched_domain
6598 * @node: node whose cpumask we're constructing
6599 * @span: resulting cpumask
6601 * Given a node, construct a good cpumask for its sched_domain to span. It
6602 * should be one that prevents unnecessary balancing, but also spreads tasks
6605 static void sched_domain_node_span(int node, struct cpumask *span)
6607 nodemask_t used_nodes;
6610 cpumask_clear(span);
6611 nodes_clear(used_nodes);
6613 cpumask_or(span, span, cpumask_of_node(node));
6614 node_set(node, used_nodes);
6616 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6617 int next_node = find_next_best_node(node, &used_nodes);
6619 cpumask_or(span, span, cpumask_of_node(next_node));
6622 #endif /* CONFIG_NUMA */
6624 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6627 * The cpus mask in sched_group and sched_domain hangs off the end.
6629 * ( See the the comments in include/linux/sched.h:struct sched_group
6630 * and struct sched_domain. )
6632 struct static_sched_group {
6633 struct sched_group sg;
6634 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6637 struct static_sched_domain {
6638 struct sched_domain sd;
6639 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6645 cpumask_var_t domainspan;
6646 cpumask_var_t covered;
6647 cpumask_var_t notcovered;
6649 cpumask_var_t nodemask;
6650 cpumask_var_t this_sibling_map;
6651 cpumask_var_t this_core_map;
6652 cpumask_var_t this_book_map;
6653 cpumask_var_t send_covered;
6654 cpumask_var_t tmpmask;
6655 struct sched_group **sched_group_nodes;
6656 struct root_domain *rd;
6660 sa_sched_groups = 0,
6666 sa_this_sibling_map,
6668 sa_sched_group_nodes,
6678 * SMT sched-domains:
6680 #ifdef CONFIG_SCHED_SMT
6681 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6682 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6685 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6686 struct sched_group **sg, struct cpumask *unused)
6689 *sg = &per_cpu(sched_groups, cpu).sg;
6692 #endif /* CONFIG_SCHED_SMT */
6695 * multi-core sched-domains:
6697 #ifdef CONFIG_SCHED_MC
6698 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6699 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6702 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6703 struct sched_group **sg, struct cpumask *mask)
6706 #ifdef CONFIG_SCHED_SMT
6707 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6708 group = cpumask_first(mask);
6713 *sg = &per_cpu(sched_group_core, group).sg;
6716 #endif /* CONFIG_SCHED_MC */
6719 * book sched-domains:
6721 #ifdef CONFIG_SCHED_BOOK
6722 static DEFINE_PER_CPU(struct static_sched_domain, book_domains);
6723 static DEFINE_PER_CPU(struct static_sched_group, sched_group_book);
6726 cpu_to_book_group(int cpu, const struct cpumask *cpu_map,
6727 struct sched_group **sg, struct cpumask *mask)
6730 #ifdef CONFIG_SCHED_MC
6731 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6732 group = cpumask_first(mask);
6733 #elif defined(CONFIG_SCHED_SMT)
6734 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6735 group = cpumask_first(mask);
6738 *sg = &per_cpu(sched_group_book, group).sg;
6741 #endif /* CONFIG_SCHED_BOOK */
6743 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6744 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6747 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6748 struct sched_group **sg, struct cpumask *mask)
6751 #ifdef CONFIG_SCHED_BOOK
6752 cpumask_and(mask, cpu_book_mask(cpu), cpu_map);
6753 group = cpumask_first(mask);
6754 #elif defined(CONFIG_SCHED_MC)
6755 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6756 group = cpumask_first(mask);
6757 #elif defined(CONFIG_SCHED_SMT)
6758 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6759 group = cpumask_first(mask);
6764 *sg = &per_cpu(sched_group_phys, group).sg;
6770 * The init_sched_build_groups can't handle what we want to do with node
6771 * groups, so roll our own. Now each node has its own list of groups which
6772 * gets dynamically allocated.
6774 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6775 static struct sched_group ***sched_group_nodes_bycpu;
6777 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6778 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6780 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6781 struct sched_group **sg,
6782 struct cpumask *nodemask)
6786 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6787 group = cpumask_first(nodemask);
6790 *sg = &per_cpu(sched_group_allnodes, group).sg;
6794 static void init_numa_sched_groups_power(struct sched_group *group_head)
6796 struct sched_group *sg = group_head;
6802 for_each_cpu(j, sched_group_cpus(sg)) {
6803 struct sched_domain *sd;
6805 sd = &per_cpu(phys_domains, j).sd;
6806 if (j != group_first_cpu(sd->groups)) {
6808 * Only add "power" once for each
6814 sg->cpu_power += sd->groups->cpu_power;
6817 } while (sg != group_head);
6820 static int build_numa_sched_groups(struct s_data *d,
6821 const struct cpumask *cpu_map, int num)
6823 struct sched_domain *sd;
6824 struct sched_group *sg, *prev;
6827 cpumask_clear(d->covered);
6828 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6829 if (cpumask_empty(d->nodemask)) {
6830 d->sched_group_nodes[num] = NULL;
6834 sched_domain_node_span(num, d->domainspan);
6835 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6837 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6840 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6844 d->sched_group_nodes[num] = sg;
6846 for_each_cpu(j, d->nodemask) {
6847 sd = &per_cpu(node_domains, j).sd;
6852 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6854 cpumask_or(d->covered, d->covered, d->nodemask);
6857 for (j = 0; j < nr_node_ids; j++) {
6858 n = (num + j) % nr_node_ids;
6859 cpumask_complement(d->notcovered, d->covered);
6860 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6861 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6862 if (cpumask_empty(d->tmpmask))
6864 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6865 if (cpumask_empty(d->tmpmask))
6867 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6871 "Can not alloc domain group for node %d\n", j);
6875 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6876 sg->next = prev->next;
6877 cpumask_or(d->covered, d->covered, d->tmpmask);
6884 #endif /* CONFIG_NUMA */
6887 /* Free memory allocated for various sched_group structures */
6888 static void free_sched_groups(const struct cpumask *cpu_map,
6889 struct cpumask *nodemask)
6893 for_each_cpu(cpu, cpu_map) {
6894 struct sched_group **sched_group_nodes
6895 = sched_group_nodes_bycpu[cpu];
6897 if (!sched_group_nodes)
6900 for (i = 0; i < nr_node_ids; i++) {
6901 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6903 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6904 if (cpumask_empty(nodemask))
6914 if (oldsg != sched_group_nodes[i])
6917 kfree(sched_group_nodes);
6918 sched_group_nodes_bycpu[cpu] = NULL;
6921 #else /* !CONFIG_NUMA */
6922 static void free_sched_groups(const struct cpumask *cpu_map,
6923 struct cpumask *nodemask)
6926 #endif /* CONFIG_NUMA */
6929 * Initialize sched groups cpu_power.
6931 * cpu_power indicates the capacity of sched group, which is used while
6932 * distributing the load between different sched groups in a sched domain.
6933 * Typically cpu_power for all the groups in a sched domain will be same unless
6934 * there are asymmetries in the topology. If there are asymmetries, group
6935 * having more cpu_power will pickup more load compared to the group having
6938 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6940 struct sched_domain *child;
6941 struct sched_group *group;
6945 WARN_ON(!sd || !sd->groups);
6947 if (cpu != group_first_cpu(sd->groups))
6952 sd->groups->cpu_power = 0;
6955 power = SCHED_LOAD_SCALE;
6956 weight = cpumask_weight(sched_domain_span(sd));
6958 * SMT siblings share the power of a single core.
6959 * Usually multiple threads get a better yield out of
6960 * that one core than a single thread would have,
6961 * reflect that in sd->smt_gain.
6963 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6964 power *= sd->smt_gain;
6966 power >>= SCHED_LOAD_SHIFT;
6968 sd->groups->cpu_power += power;
6973 * Add cpu_power of each child group to this groups cpu_power.
6975 group = child->groups;
6977 sd->groups->cpu_power += group->cpu_power;
6978 group = group->next;
6979 } while (group != child->groups);
6983 * Initializers for schedule domains
6984 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6987 #ifdef CONFIG_SCHED_DEBUG
6988 # define SD_INIT_NAME(sd, type) sd->name = #type
6990 # define SD_INIT_NAME(sd, type) do { } while (0)
6993 #define SD_INIT(sd, type) sd_init_##type(sd)
6995 #define SD_INIT_FUNC(type) \
6996 static noinline void sd_init_##type(struct sched_domain *sd) \
6998 memset(sd, 0, sizeof(*sd)); \
6999 *sd = SD_##type##_INIT; \
7000 sd->level = SD_LV_##type; \
7001 SD_INIT_NAME(sd, type); \
7006 SD_INIT_FUNC(ALLNODES)
7009 #ifdef CONFIG_SCHED_SMT
7010 SD_INIT_FUNC(SIBLING)
7012 #ifdef CONFIG_SCHED_MC
7015 #ifdef CONFIG_SCHED_BOOK
7019 static int default_relax_domain_level = -1;
7021 static int __init setup_relax_domain_level(char *str)
7025 val = simple_strtoul(str, NULL, 0);
7026 if (val < SD_LV_MAX)
7027 default_relax_domain_level = val;
7031 __setup("relax_domain_level=", setup_relax_domain_level);
7033 static void set_domain_attribute(struct sched_domain *sd,
7034 struct sched_domain_attr *attr)
7038 if (!attr || attr->relax_domain_level < 0) {
7039 if (default_relax_domain_level < 0)
7042 request = default_relax_domain_level;
7044 request = attr->relax_domain_level;
7045 if (request < sd->level) {
7046 /* turn off idle balance on this domain */
7047 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7049 /* turn on idle balance on this domain */
7050 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7054 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7055 const struct cpumask *cpu_map)
7058 case sa_sched_groups:
7059 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
7060 d->sched_group_nodes = NULL;
7062 free_rootdomain(d->rd); /* fall through */
7064 free_cpumask_var(d->tmpmask); /* fall through */
7065 case sa_send_covered:
7066 free_cpumask_var(d->send_covered); /* fall through */
7067 case sa_this_book_map:
7068 free_cpumask_var(d->this_book_map); /* fall through */
7069 case sa_this_core_map:
7070 free_cpumask_var(d->this_core_map); /* fall through */
7071 case sa_this_sibling_map:
7072 free_cpumask_var(d->this_sibling_map); /* fall through */
7074 free_cpumask_var(d->nodemask); /* fall through */
7075 case sa_sched_group_nodes:
7077 kfree(d->sched_group_nodes); /* fall through */
7079 free_cpumask_var(d->notcovered); /* fall through */
7081 free_cpumask_var(d->covered); /* fall through */
7083 free_cpumask_var(d->domainspan); /* fall through */
7090 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7091 const struct cpumask *cpu_map)
7094 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
7096 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
7097 return sa_domainspan;
7098 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
7100 /* Allocate the per-node list of sched groups */
7101 d->sched_group_nodes = kcalloc(nr_node_ids,
7102 sizeof(struct sched_group *), GFP_KERNEL);
7103 if (!d->sched_group_nodes) {
7104 printk(KERN_WARNING "Can not alloc sched group node list\n");
7105 return sa_notcovered;
7107 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
7109 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
7110 return sa_sched_group_nodes;
7111 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
7113 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
7114 return sa_this_sibling_map;
7115 if (!alloc_cpumask_var(&d->this_book_map, GFP_KERNEL))
7116 return sa_this_core_map;
7117 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
7118 return sa_this_book_map;
7119 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
7120 return sa_send_covered;
7121 d->rd = alloc_rootdomain();
7123 printk(KERN_WARNING "Cannot alloc root domain\n");
7126 return sa_rootdomain;
7129 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
7130 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
7132 struct sched_domain *sd = NULL;
7134 struct sched_domain *parent;
7137 if (cpumask_weight(cpu_map) >
7138 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
7139 sd = &per_cpu(allnodes_domains, i).sd;
7140 SD_INIT(sd, ALLNODES);
7141 set_domain_attribute(sd, attr);
7142 cpumask_copy(sched_domain_span(sd), cpu_map);
7143 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
7148 sd = &per_cpu(node_domains, i).sd;
7150 set_domain_attribute(sd, attr);
7151 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7152 sd->parent = parent;
7155 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7160 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7161 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7162 struct sched_domain *parent, int i)
7164 struct sched_domain *sd;
7165 sd = &per_cpu(phys_domains, i).sd;
7167 set_domain_attribute(sd, attr);
7168 cpumask_copy(sched_domain_span(sd), d->nodemask);
7169 sd->parent = parent;
7172 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7176 static struct sched_domain *__build_book_sched_domain(struct s_data *d,
7177 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7178 struct sched_domain *parent, int i)
7180 struct sched_domain *sd = parent;
7181 #ifdef CONFIG_SCHED_BOOK
7182 sd = &per_cpu(book_domains, i).sd;
7184 set_domain_attribute(sd, attr);
7185 cpumask_and(sched_domain_span(sd), cpu_map, cpu_book_mask(i));
7186 sd->parent = parent;
7188 cpu_to_book_group(i, cpu_map, &sd->groups, d->tmpmask);
7193 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7194 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7195 struct sched_domain *parent, int i)
7197 struct sched_domain *sd = parent;
7198 #ifdef CONFIG_SCHED_MC
7199 sd = &per_cpu(core_domains, i).sd;
7201 set_domain_attribute(sd, attr);
7202 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7203 sd->parent = parent;
7205 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7210 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7211 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7212 struct sched_domain *parent, int i)
7214 struct sched_domain *sd = parent;
7215 #ifdef CONFIG_SCHED_SMT
7216 sd = &per_cpu(cpu_domains, i).sd;
7217 SD_INIT(sd, SIBLING);
7218 set_domain_attribute(sd, attr);
7219 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7220 sd->parent = parent;
7222 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7227 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7228 const struct cpumask *cpu_map, int cpu)
7231 #ifdef CONFIG_SCHED_SMT
7232 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7233 cpumask_and(d->this_sibling_map, cpu_map,
7234 topology_thread_cpumask(cpu));
7235 if (cpu == cpumask_first(d->this_sibling_map))
7236 init_sched_build_groups(d->this_sibling_map, cpu_map,
7238 d->send_covered, d->tmpmask);
7241 #ifdef CONFIG_SCHED_MC
7242 case SD_LV_MC: /* set up multi-core groups */
7243 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7244 if (cpu == cpumask_first(d->this_core_map))
7245 init_sched_build_groups(d->this_core_map, cpu_map,
7247 d->send_covered, d->tmpmask);
7250 #ifdef CONFIG_SCHED_BOOK
7251 case SD_LV_BOOK: /* set up book groups */
7252 cpumask_and(d->this_book_map, cpu_map, cpu_book_mask(cpu));
7253 if (cpu == cpumask_first(d->this_book_map))
7254 init_sched_build_groups(d->this_book_map, cpu_map,
7256 d->send_covered, d->tmpmask);
7259 case SD_LV_CPU: /* set up physical groups */
7260 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7261 if (!cpumask_empty(d->nodemask))
7262 init_sched_build_groups(d->nodemask, cpu_map,
7264 d->send_covered, d->tmpmask);
7267 case SD_LV_ALLNODES:
7268 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7269 d->send_covered, d->tmpmask);
7278 * Build sched domains for a given set of cpus and attach the sched domains
7279 * to the individual cpus
7281 static int __build_sched_domains(const struct cpumask *cpu_map,
7282 struct sched_domain_attr *attr)
7284 enum s_alloc alloc_state = sa_none;
7286 struct sched_domain *sd;
7292 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7293 if (alloc_state != sa_rootdomain)
7295 alloc_state = sa_sched_groups;
7298 * Set up domains for cpus specified by the cpu_map.
7300 for_each_cpu(i, cpu_map) {
7301 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7304 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7305 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7306 sd = __build_book_sched_domain(&d, cpu_map, attr, sd, i);
7307 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7308 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7311 for_each_cpu(i, cpu_map) {
7312 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7313 build_sched_groups(&d, SD_LV_BOOK, cpu_map, i);
7314 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7317 /* Set up physical groups */
7318 for (i = 0; i < nr_node_ids; i++)
7319 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7322 /* Set up node groups */
7324 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7326 for (i = 0; i < nr_node_ids; i++)
7327 if (build_numa_sched_groups(&d, cpu_map, i))
7331 /* Calculate CPU power for physical packages and nodes */
7332 #ifdef CONFIG_SCHED_SMT
7333 for_each_cpu(i, cpu_map) {
7334 sd = &per_cpu(cpu_domains, i).sd;
7335 init_sched_groups_power(i, sd);
7338 #ifdef CONFIG_SCHED_MC
7339 for_each_cpu(i, cpu_map) {
7340 sd = &per_cpu(core_domains, i).sd;
7341 init_sched_groups_power(i, sd);
7344 #ifdef CONFIG_SCHED_BOOK
7345 for_each_cpu(i, cpu_map) {
7346 sd = &per_cpu(book_domains, i).sd;
7347 init_sched_groups_power(i, sd);
7351 for_each_cpu(i, cpu_map) {
7352 sd = &per_cpu(phys_domains, i).sd;
7353 init_sched_groups_power(i, sd);
7357 for (i = 0; i < nr_node_ids; i++)
7358 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7360 if (d.sd_allnodes) {
7361 struct sched_group *sg;
7363 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7365 init_numa_sched_groups_power(sg);
7369 /* Attach the domains */
7370 for_each_cpu(i, cpu_map) {
7371 #ifdef CONFIG_SCHED_SMT
7372 sd = &per_cpu(cpu_domains, i).sd;
7373 #elif defined(CONFIG_SCHED_MC)
7374 sd = &per_cpu(core_domains, i).sd;
7375 #elif defined(CONFIG_SCHED_BOOK)
7376 sd = &per_cpu(book_domains, i).sd;
7378 sd = &per_cpu(phys_domains, i).sd;
7380 cpu_attach_domain(sd, d.rd, i);
7383 d.sched_group_nodes = NULL; /* don't free this we still need it */
7384 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7388 __free_domain_allocs(&d, alloc_state, cpu_map);
7392 static int build_sched_domains(const struct cpumask *cpu_map)
7394 return __build_sched_domains(cpu_map, NULL);
7397 static cpumask_var_t *doms_cur; /* current sched domains */
7398 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7399 static struct sched_domain_attr *dattr_cur;
7400 /* attribues of custom domains in 'doms_cur' */
7403 * Special case: If a kmalloc of a doms_cur partition (array of
7404 * cpumask) fails, then fallback to a single sched domain,
7405 * as determined by the single cpumask fallback_doms.
7407 static cpumask_var_t fallback_doms;
7410 * arch_update_cpu_topology lets virtualized architectures update the
7411 * cpu core maps. It is supposed to return 1 if the topology changed
7412 * or 0 if it stayed the same.
7414 int __attribute__((weak)) arch_update_cpu_topology(void)
7419 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7422 cpumask_var_t *doms;
7424 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7427 for (i = 0; i < ndoms; i++) {
7428 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7429 free_sched_domains(doms, i);
7436 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7439 for (i = 0; i < ndoms; i++)
7440 free_cpumask_var(doms[i]);
7445 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7446 * For now this just excludes isolated cpus, but could be used to
7447 * exclude other special cases in the future.
7449 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7453 arch_update_cpu_topology();
7455 doms_cur = alloc_sched_domains(ndoms_cur);
7457 doms_cur = &fallback_doms;
7458 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7460 err = build_sched_domains(doms_cur[0]);
7461 register_sched_domain_sysctl();
7466 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7467 struct cpumask *tmpmask)
7469 free_sched_groups(cpu_map, tmpmask);
7473 * Detach sched domains from a group of cpus specified in cpu_map
7474 * These cpus will now be attached to the NULL domain
7476 static void detach_destroy_domains(const struct cpumask *cpu_map)
7478 /* Save because hotplug lock held. */
7479 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7482 for_each_cpu(i, cpu_map)
7483 cpu_attach_domain(NULL, &def_root_domain, i);
7484 synchronize_sched();
7485 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7488 /* handle null as "default" */
7489 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7490 struct sched_domain_attr *new, int idx_new)
7492 struct sched_domain_attr tmp;
7499 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7500 new ? (new + idx_new) : &tmp,
7501 sizeof(struct sched_domain_attr));
7505 * Partition sched domains as specified by the 'ndoms_new'
7506 * cpumasks in the array doms_new[] of cpumasks. This compares
7507 * doms_new[] to the current sched domain partitioning, doms_cur[].
7508 * It destroys each deleted domain and builds each new domain.
7510 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7511 * The masks don't intersect (don't overlap.) We should setup one
7512 * sched domain for each mask. CPUs not in any of the cpumasks will
7513 * not be load balanced. If the same cpumask appears both in the
7514 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7517 * The passed in 'doms_new' should be allocated using
7518 * alloc_sched_domains. This routine takes ownership of it and will
7519 * free_sched_domains it when done with it. If the caller failed the
7520 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7521 * and partition_sched_domains() will fallback to the single partition
7522 * 'fallback_doms', it also forces the domains to be rebuilt.
7524 * If doms_new == NULL it will be replaced with cpu_online_mask.
7525 * ndoms_new == 0 is a special case for destroying existing domains,
7526 * and it will not create the default domain.
7528 * Call with hotplug lock held
7530 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7531 struct sched_domain_attr *dattr_new)
7536 mutex_lock(&sched_domains_mutex);
7538 /* always unregister in case we don't destroy any domains */
7539 unregister_sched_domain_sysctl();
7541 /* Let architecture update cpu core mappings. */
7542 new_topology = arch_update_cpu_topology();
7544 n = doms_new ? ndoms_new : 0;
7546 /* Destroy deleted domains */
7547 for (i = 0; i < ndoms_cur; i++) {
7548 for (j = 0; j < n && !new_topology; j++) {
7549 if (cpumask_equal(doms_cur[i], doms_new[j])
7550 && dattrs_equal(dattr_cur, i, dattr_new, j))
7553 /* no match - a current sched domain not in new doms_new[] */
7554 detach_destroy_domains(doms_cur[i]);
7559 if (doms_new == NULL) {
7561 doms_new = &fallback_doms;
7562 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7563 WARN_ON_ONCE(dattr_new);
7566 /* Build new domains */
7567 for (i = 0; i < ndoms_new; i++) {
7568 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7569 if (cpumask_equal(doms_new[i], doms_cur[j])
7570 && dattrs_equal(dattr_new, i, dattr_cur, j))
7573 /* no match - add a new doms_new */
7574 __build_sched_domains(doms_new[i],
7575 dattr_new ? dattr_new + i : NULL);
7580 /* Remember the new sched domains */
7581 if (doms_cur != &fallback_doms)
7582 free_sched_domains(doms_cur, ndoms_cur);
7583 kfree(dattr_cur); /* kfree(NULL) is safe */
7584 doms_cur = doms_new;
7585 dattr_cur = dattr_new;
7586 ndoms_cur = ndoms_new;
7588 register_sched_domain_sysctl();
7590 mutex_unlock(&sched_domains_mutex);
7593 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7594 static void arch_reinit_sched_domains(void)
7598 /* Destroy domains first to force the rebuild */
7599 partition_sched_domains(0, NULL, NULL);
7601 rebuild_sched_domains();
7605 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7607 unsigned int level = 0;
7609 if (sscanf(buf, "%u", &level) != 1)
7613 * level is always be positive so don't check for
7614 * level < POWERSAVINGS_BALANCE_NONE which is 0
7615 * What happens on 0 or 1 byte write,
7616 * need to check for count as well?
7619 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7623 sched_smt_power_savings = level;
7625 sched_mc_power_savings = level;
7627 arch_reinit_sched_domains();
7632 #ifdef CONFIG_SCHED_MC
7633 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7634 struct sysdev_class_attribute *attr,
7637 return sprintf(page, "%u\n", sched_mc_power_savings);
7639 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7640 struct sysdev_class_attribute *attr,
7641 const char *buf, size_t count)
7643 return sched_power_savings_store(buf, count, 0);
7645 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7646 sched_mc_power_savings_show,
7647 sched_mc_power_savings_store);
7650 #ifdef CONFIG_SCHED_SMT
7651 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7652 struct sysdev_class_attribute *attr,
7655 return sprintf(page, "%u\n", sched_smt_power_savings);
7657 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7658 struct sysdev_class_attribute *attr,
7659 const char *buf, size_t count)
7661 return sched_power_savings_store(buf, count, 1);
7663 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7664 sched_smt_power_savings_show,
7665 sched_smt_power_savings_store);
7668 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7672 #ifdef CONFIG_SCHED_SMT
7674 err = sysfs_create_file(&cls->kset.kobj,
7675 &attr_sched_smt_power_savings.attr);
7677 #ifdef CONFIG_SCHED_MC
7678 if (!err && mc_capable())
7679 err = sysfs_create_file(&cls->kset.kobj,
7680 &attr_sched_mc_power_savings.attr);
7684 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7687 * Update cpusets according to cpu_active mask. If cpusets are
7688 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7689 * around partition_sched_domains().
7691 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7694 switch (action & ~CPU_TASKS_FROZEN) {
7696 case CPU_DOWN_FAILED:
7697 cpuset_update_active_cpus();
7704 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7707 switch (action & ~CPU_TASKS_FROZEN) {
7708 case CPU_DOWN_PREPARE:
7709 cpuset_update_active_cpus();
7716 static int update_runtime(struct notifier_block *nfb,
7717 unsigned long action, void *hcpu)
7719 int cpu = (int)(long)hcpu;
7722 case CPU_DOWN_PREPARE:
7723 case CPU_DOWN_PREPARE_FROZEN:
7724 disable_runtime(cpu_rq(cpu));
7727 case CPU_DOWN_FAILED:
7728 case CPU_DOWN_FAILED_FROZEN:
7730 case CPU_ONLINE_FROZEN:
7731 enable_runtime(cpu_rq(cpu));
7739 void __init sched_init_smp(void)
7741 cpumask_var_t non_isolated_cpus;
7743 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7744 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7746 #if defined(CONFIG_NUMA)
7747 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7749 BUG_ON(sched_group_nodes_bycpu == NULL);
7752 mutex_lock(&sched_domains_mutex);
7753 arch_init_sched_domains(cpu_active_mask);
7754 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7755 if (cpumask_empty(non_isolated_cpus))
7756 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7757 mutex_unlock(&sched_domains_mutex);
7760 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7761 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7763 /* RT runtime code needs to handle some hotplug events */
7764 hotcpu_notifier(update_runtime, 0);
7768 /* Move init over to a non-isolated CPU */
7769 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7771 sched_init_granularity();
7772 free_cpumask_var(non_isolated_cpus);
7774 init_sched_rt_class();
7777 void __init sched_init_smp(void)
7779 sched_init_granularity();
7781 #endif /* CONFIG_SMP */
7783 const_debug unsigned int sysctl_timer_migration = 1;
7785 int in_sched_functions(unsigned long addr)
7787 return in_lock_functions(addr) ||
7788 (addr >= (unsigned long)__sched_text_start
7789 && addr < (unsigned long)__sched_text_end);
7792 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7794 cfs_rq->tasks_timeline = RB_ROOT;
7795 INIT_LIST_HEAD(&cfs_rq->tasks);
7796 #ifdef CONFIG_FAIR_GROUP_SCHED
7799 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7802 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7804 struct rt_prio_array *array;
7807 array = &rt_rq->active;
7808 for (i = 0; i < MAX_RT_PRIO; i++) {
7809 INIT_LIST_HEAD(array->queue + i);
7810 __clear_bit(i, array->bitmap);
7812 /* delimiter for bitsearch: */
7813 __set_bit(MAX_RT_PRIO, array->bitmap);
7815 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7816 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7818 rt_rq->highest_prio.next = MAX_RT_PRIO;
7822 rt_rq->rt_nr_migratory = 0;
7823 rt_rq->overloaded = 0;
7824 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7828 rt_rq->rt_throttled = 0;
7829 rt_rq->rt_runtime = 0;
7830 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7832 #ifdef CONFIG_RT_GROUP_SCHED
7833 rt_rq->rt_nr_boosted = 0;
7838 #ifdef CONFIG_FAIR_GROUP_SCHED
7839 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7840 struct sched_entity *se, int cpu, int add,
7841 struct sched_entity *parent)
7843 struct rq *rq = cpu_rq(cpu);
7844 tg->cfs_rq[cpu] = cfs_rq;
7845 init_cfs_rq(cfs_rq, rq);
7848 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7851 /* se could be NULL for init_task_group */
7856 se->cfs_rq = &rq->cfs;
7858 se->cfs_rq = parent->my_q;
7861 se->load.weight = tg->shares;
7862 se->load.inv_weight = 0;
7863 se->parent = parent;
7867 #ifdef CONFIG_RT_GROUP_SCHED
7868 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7869 struct sched_rt_entity *rt_se, int cpu, int add,
7870 struct sched_rt_entity *parent)
7872 struct rq *rq = cpu_rq(cpu);
7874 tg->rt_rq[cpu] = rt_rq;
7875 init_rt_rq(rt_rq, rq);
7877 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7879 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7881 tg->rt_se[cpu] = rt_se;
7886 rt_se->rt_rq = &rq->rt;
7888 rt_se->rt_rq = parent->my_q;
7890 rt_se->my_q = rt_rq;
7891 rt_se->parent = parent;
7892 INIT_LIST_HEAD(&rt_se->run_list);
7896 void __init sched_init(void)
7899 unsigned long alloc_size = 0, ptr;
7901 #ifdef CONFIG_FAIR_GROUP_SCHED
7902 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7904 #ifdef CONFIG_RT_GROUP_SCHED
7905 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7907 #ifdef CONFIG_CPUMASK_OFFSTACK
7908 alloc_size += num_possible_cpus() * cpumask_size();
7911 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7913 #ifdef CONFIG_FAIR_GROUP_SCHED
7914 init_task_group.se = (struct sched_entity **)ptr;
7915 ptr += nr_cpu_ids * sizeof(void **);
7917 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7918 ptr += nr_cpu_ids * sizeof(void **);
7920 #endif /* CONFIG_FAIR_GROUP_SCHED */
7921 #ifdef CONFIG_RT_GROUP_SCHED
7922 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7923 ptr += nr_cpu_ids * sizeof(void **);
7925 init_task_group.rt_rq = (struct rt_rq **)ptr;
7926 ptr += nr_cpu_ids * sizeof(void **);
7928 #endif /* CONFIG_RT_GROUP_SCHED */
7929 #ifdef CONFIG_CPUMASK_OFFSTACK
7930 for_each_possible_cpu(i) {
7931 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7932 ptr += cpumask_size();
7934 #endif /* CONFIG_CPUMASK_OFFSTACK */
7938 init_defrootdomain();
7941 init_rt_bandwidth(&def_rt_bandwidth,
7942 global_rt_period(), global_rt_runtime());
7944 #ifdef CONFIG_RT_GROUP_SCHED
7945 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7946 global_rt_period(), global_rt_runtime());
7947 #endif /* CONFIG_RT_GROUP_SCHED */
7949 #ifdef CONFIG_CGROUP_SCHED
7950 list_add(&init_task_group.list, &task_groups);
7951 INIT_LIST_HEAD(&init_task_group.children);
7953 #endif /* CONFIG_CGROUP_SCHED */
7955 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7956 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7957 __alignof__(unsigned long));
7959 for_each_possible_cpu(i) {
7963 raw_spin_lock_init(&rq->lock);
7965 rq->calc_load_active = 0;
7966 rq->calc_load_update = jiffies + LOAD_FREQ;
7967 init_cfs_rq(&rq->cfs, rq);
7968 init_rt_rq(&rq->rt, rq);
7969 #ifdef CONFIG_FAIR_GROUP_SCHED
7970 init_task_group.shares = init_task_group_load;
7971 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7972 #ifdef CONFIG_CGROUP_SCHED
7974 * How much cpu bandwidth does init_task_group get?
7976 * In case of task-groups formed thr' the cgroup filesystem, it
7977 * gets 100% of the cpu resources in the system. This overall
7978 * system cpu resource is divided among the tasks of
7979 * init_task_group and its child task-groups in a fair manner,
7980 * based on each entity's (task or task-group's) weight
7981 * (se->load.weight).
7983 * In other words, if init_task_group has 10 tasks of weight
7984 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7985 * then A0's share of the cpu resource is:
7987 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7989 * We achieve this by letting init_task_group's tasks sit
7990 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7992 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7994 #endif /* CONFIG_FAIR_GROUP_SCHED */
7996 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7997 #ifdef CONFIG_RT_GROUP_SCHED
7998 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7999 #ifdef CONFIG_CGROUP_SCHED
8000 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8004 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8005 rq->cpu_load[j] = 0;
8007 rq->last_load_update_tick = jiffies;
8012 rq->cpu_power = SCHED_LOAD_SCALE;
8013 rq->post_schedule = 0;
8014 rq->active_balance = 0;
8015 rq->next_balance = jiffies;
8020 rq->avg_idle = 2*sysctl_sched_migration_cost;
8021 rq_attach_root(rq, &def_root_domain);
8023 rq->nohz_balance_kick = 0;
8024 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8028 atomic_set(&rq->nr_iowait, 0);
8031 set_load_weight(&init_task);
8033 #ifdef CONFIG_PREEMPT_NOTIFIERS
8034 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8038 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8041 #ifdef CONFIG_RT_MUTEXES
8042 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
8046 * The boot idle thread does lazy MMU switching as well:
8048 atomic_inc(&init_mm.mm_count);
8049 enter_lazy_tlb(&init_mm, current);
8052 * Make us the idle thread. Technically, schedule() should not be
8053 * called from this thread, however somewhere below it might be,
8054 * but because we are the idle thread, we just pick up running again
8055 * when this runqueue becomes "idle".
8057 init_idle(current, smp_processor_id());
8059 calc_load_update = jiffies + LOAD_FREQ;
8062 * During early bootup we pretend to be a normal task:
8064 current->sched_class = &fair_sched_class;
8066 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8067 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8070 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8071 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8072 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8073 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8074 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8076 /* May be allocated at isolcpus cmdline parse time */
8077 if (cpu_isolated_map == NULL)
8078 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8083 scheduler_running = 1;
8086 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8087 static inline int preempt_count_equals(int preempt_offset)
8089 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8091 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
8094 void __might_sleep(const char *file, int line, int preempt_offset)
8097 static unsigned long prev_jiffy; /* ratelimiting */
8099 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8100 system_state != SYSTEM_RUNNING || oops_in_progress)
8102 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8104 prev_jiffy = jiffies;
8107 "BUG: sleeping function called from invalid context at %s:%d\n",
8110 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8111 in_atomic(), irqs_disabled(),
8112 current->pid, current->comm);
8114 debug_show_held_locks(current);
8115 if (irqs_disabled())
8116 print_irqtrace_events(current);
8120 EXPORT_SYMBOL(__might_sleep);
8123 #ifdef CONFIG_MAGIC_SYSRQ
8124 static void normalize_task(struct rq *rq, struct task_struct *p)
8128 on_rq = p->se.on_rq;
8130 deactivate_task(rq, p, 0);
8131 __setscheduler(rq, p, SCHED_NORMAL, 0);
8133 activate_task(rq, p, 0);
8134 resched_task(rq->curr);
8138 void normalize_rt_tasks(void)
8140 struct task_struct *g, *p;
8141 unsigned long flags;
8144 read_lock_irqsave(&tasklist_lock, flags);
8145 do_each_thread(g, p) {
8147 * Only normalize user tasks:
8152 p->se.exec_start = 0;
8153 #ifdef CONFIG_SCHEDSTATS
8154 p->se.statistics.wait_start = 0;
8155 p->se.statistics.sleep_start = 0;
8156 p->se.statistics.block_start = 0;
8161 * Renice negative nice level userspace
8164 if (TASK_NICE(p) < 0 && p->mm)
8165 set_user_nice(p, 0);
8169 raw_spin_lock(&p->pi_lock);
8170 rq = __task_rq_lock(p);
8172 normalize_task(rq, p);
8174 __task_rq_unlock(rq);
8175 raw_spin_unlock(&p->pi_lock);
8176 } while_each_thread(g, p);
8178 read_unlock_irqrestore(&tasklist_lock, flags);
8181 #endif /* CONFIG_MAGIC_SYSRQ */
8183 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8185 * These functions are only useful for the IA64 MCA handling, or kdb.
8187 * They can only be called when the whole system has been
8188 * stopped - every CPU needs to be quiescent, and no scheduling
8189 * activity can take place. Using them for anything else would
8190 * be a serious bug, and as a result, they aren't even visible
8191 * under any other configuration.
8195 * curr_task - return the current task for a given cpu.
8196 * @cpu: the processor in question.
8198 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8200 struct task_struct *curr_task(int cpu)
8202 return cpu_curr(cpu);
8205 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8209 * set_curr_task - set the current task for a given cpu.
8210 * @cpu: the processor in question.
8211 * @p: the task pointer to set.
8213 * Description: This function must only be used when non-maskable interrupts
8214 * are serviced on a separate stack. It allows the architecture to switch the
8215 * notion of the current task on a cpu in a non-blocking manner. This function
8216 * must be called with all CPU's synchronized, and interrupts disabled, the
8217 * and caller must save the original value of the current task (see
8218 * curr_task() above) and restore that value before reenabling interrupts and
8219 * re-starting the system.
8221 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8223 void set_curr_task(int cpu, struct task_struct *p)
8230 #ifdef CONFIG_FAIR_GROUP_SCHED
8231 static void free_fair_sched_group(struct task_group *tg)
8235 for_each_possible_cpu(i) {
8237 kfree(tg->cfs_rq[i]);
8247 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8249 struct cfs_rq *cfs_rq;
8250 struct sched_entity *se;
8254 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8257 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8261 tg->shares = NICE_0_LOAD;
8263 for_each_possible_cpu(i) {
8266 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8267 GFP_KERNEL, cpu_to_node(i));
8271 se = kzalloc_node(sizeof(struct sched_entity),
8272 GFP_KERNEL, cpu_to_node(i));
8276 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8287 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8289 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8290 &cpu_rq(cpu)->leaf_cfs_rq_list);
8293 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8295 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8297 #else /* !CONFG_FAIR_GROUP_SCHED */
8298 static inline void free_fair_sched_group(struct task_group *tg)
8303 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8308 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8312 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8315 #endif /* CONFIG_FAIR_GROUP_SCHED */
8317 #ifdef CONFIG_RT_GROUP_SCHED
8318 static void free_rt_sched_group(struct task_group *tg)
8322 destroy_rt_bandwidth(&tg->rt_bandwidth);
8324 for_each_possible_cpu(i) {
8326 kfree(tg->rt_rq[i]);
8328 kfree(tg->rt_se[i]);
8336 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8338 struct rt_rq *rt_rq;
8339 struct sched_rt_entity *rt_se;
8343 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8346 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8350 init_rt_bandwidth(&tg->rt_bandwidth,
8351 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8353 for_each_possible_cpu(i) {
8356 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8357 GFP_KERNEL, cpu_to_node(i));
8361 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8362 GFP_KERNEL, cpu_to_node(i));
8366 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8377 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8379 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8380 &cpu_rq(cpu)->leaf_rt_rq_list);
8383 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8385 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8387 #else /* !CONFIG_RT_GROUP_SCHED */
8388 static inline void free_rt_sched_group(struct task_group *tg)
8393 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8398 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8402 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8405 #endif /* CONFIG_RT_GROUP_SCHED */
8407 #ifdef CONFIG_CGROUP_SCHED
8408 static void free_sched_group(struct task_group *tg)
8410 free_fair_sched_group(tg);
8411 free_rt_sched_group(tg);
8415 /* allocate runqueue etc for a new task group */
8416 struct task_group *sched_create_group(struct task_group *parent)
8418 struct task_group *tg;
8419 unsigned long flags;
8422 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8424 return ERR_PTR(-ENOMEM);
8426 if (!alloc_fair_sched_group(tg, parent))
8429 if (!alloc_rt_sched_group(tg, parent))
8432 spin_lock_irqsave(&task_group_lock, flags);
8433 for_each_possible_cpu(i) {
8434 register_fair_sched_group(tg, i);
8435 register_rt_sched_group(tg, i);
8437 list_add_rcu(&tg->list, &task_groups);
8439 WARN_ON(!parent); /* root should already exist */
8441 tg->parent = parent;
8442 INIT_LIST_HEAD(&tg->children);
8443 list_add_rcu(&tg->siblings, &parent->children);
8444 spin_unlock_irqrestore(&task_group_lock, flags);
8449 free_sched_group(tg);
8450 return ERR_PTR(-ENOMEM);
8453 /* rcu callback to free various structures associated with a task group */
8454 static void free_sched_group_rcu(struct rcu_head *rhp)
8456 /* now it should be safe to free those cfs_rqs */
8457 free_sched_group(container_of(rhp, struct task_group, rcu));
8460 /* Destroy runqueue etc associated with a task group */
8461 void sched_destroy_group(struct task_group *tg)
8463 unsigned long flags;
8466 spin_lock_irqsave(&task_group_lock, flags);
8467 for_each_possible_cpu(i) {
8468 unregister_fair_sched_group(tg, i);
8469 unregister_rt_sched_group(tg, i);
8471 list_del_rcu(&tg->list);
8472 list_del_rcu(&tg->siblings);
8473 spin_unlock_irqrestore(&task_group_lock, flags);
8475 /* wait for possible concurrent references to cfs_rqs complete */
8476 call_rcu(&tg->rcu, free_sched_group_rcu);
8479 /* change task's runqueue when it moves between groups.
8480 * The caller of this function should have put the task in its new group
8481 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8482 * reflect its new group.
8484 void sched_move_task(struct task_struct *tsk)
8487 unsigned long flags;
8490 rq = task_rq_lock(tsk, &flags);
8492 running = task_current(rq, tsk);
8493 on_rq = tsk->se.on_rq;
8496 dequeue_task(rq, tsk, 0);
8497 if (unlikely(running))
8498 tsk->sched_class->put_prev_task(rq, tsk);
8500 set_task_rq(tsk, task_cpu(tsk));
8502 #ifdef CONFIG_FAIR_GROUP_SCHED
8503 if (tsk->sched_class->moved_group)
8504 tsk->sched_class->moved_group(tsk, on_rq);
8507 if (unlikely(running))
8508 tsk->sched_class->set_curr_task(rq);
8510 enqueue_task(rq, tsk, 0);
8512 task_rq_unlock(rq, &flags);
8514 #endif /* CONFIG_CGROUP_SCHED */
8516 #ifdef CONFIG_FAIR_GROUP_SCHED
8517 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8519 struct cfs_rq *cfs_rq = se->cfs_rq;
8524 dequeue_entity(cfs_rq, se, 0);
8526 se->load.weight = shares;
8527 se->load.inv_weight = 0;
8530 enqueue_entity(cfs_rq, se, 0);
8533 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8535 struct cfs_rq *cfs_rq = se->cfs_rq;
8536 struct rq *rq = cfs_rq->rq;
8537 unsigned long flags;
8539 raw_spin_lock_irqsave(&rq->lock, flags);
8540 __set_se_shares(se, shares);
8541 raw_spin_unlock_irqrestore(&rq->lock, flags);
8544 static DEFINE_MUTEX(shares_mutex);
8546 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8549 unsigned long flags;
8552 * We can't change the weight of the root cgroup.
8557 if (shares < MIN_SHARES)
8558 shares = MIN_SHARES;
8559 else if (shares > MAX_SHARES)
8560 shares = MAX_SHARES;
8562 mutex_lock(&shares_mutex);
8563 if (tg->shares == shares)
8566 spin_lock_irqsave(&task_group_lock, flags);
8567 for_each_possible_cpu(i)
8568 unregister_fair_sched_group(tg, i);
8569 list_del_rcu(&tg->siblings);
8570 spin_unlock_irqrestore(&task_group_lock, flags);
8572 /* wait for any ongoing reference to this group to finish */
8573 synchronize_sched();
8576 * Now we are free to modify the group's share on each cpu
8577 * w/o tripping rebalance_share or load_balance_fair.
8579 tg->shares = shares;
8580 for_each_possible_cpu(i) {
8584 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8585 set_se_shares(tg->se[i], shares);
8589 * Enable load balance activity on this group, by inserting it back on
8590 * each cpu's rq->leaf_cfs_rq_list.
8592 spin_lock_irqsave(&task_group_lock, flags);
8593 for_each_possible_cpu(i)
8594 register_fair_sched_group(tg, i);
8595 list_add_rcu(&tg->siblings, &tg->parent->children);
8596 spin_unlock_irqrestore(&task_group_lock, flags);
8598 mutex_unlock(&shares_mutex);
8602 unsigned long sched_group_shares(struct task_group *tg)
8608 #ifdef CONFIG_RT_GROUP_SCHED
8610 * Ensure that the real time constraints are schedulable.
8612 static DEFINE_MUTEX(rt_constraints_mutex);
8614 static unsigned long to_ratio(u64 period, u64 runtime)
8616 if (runtime == RUNTIME_INF)
8619 return div64_u64(runtime << 20, period);
8622 /* Must be called with tasklist_lock held */
8623 static inline int tg_has_rt_tasks(struct task_group *tg)
8625 struct task_struct *g, *p;
8627 do_each_thread(g, p) {
8628 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8630 } while_each_thread(g, p);
8635 struct rt_schedulable_data {
8636 struct task_group *tg;
8641 static int tg_schedulable(struct task_group *tg, void *data)
8643 struct rt_schedulable_data *d = data;
8644 struct task_group *child;
8645 unsigned long total, sum = 0;
8646 u64 period, runtime;
8648 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8649 runtime = tg->rt_bandwidth.rt_runtime;
8652 period = d->rt_period;
8653 runtime = d->rt_runtime;
8657 * Cannot have more runtime than the period.
8659 if (runtime > period && runtime != RUNTIME_INF)
8663 * Ensure we don't starve existing RT tasks.
8665 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8668 total = to_ratio(period, runtime);
8671 * Nobody can have more than the global setting allows.
8673 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8677 * The sum of our children's runtime should not exceed our own.
8679 list_for_each_entry_rcu(child, &tg->children, siblings) {
8680 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8681 runtime = child->rt_bandwidth.rt_runtime;
8683 if (child == d->tg) {
8684 period = d->rt_period;
8685 runtime = d->rt_runtime;
8688 sum += to_ratio(period, runtime);
8697 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8699 struct rt_schedulable_data data = {
8701 .rt_period = period,
8702 .rt_runtime = runtime,
8705 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8708 static int tg_set_bandwidth(struct task_group *tg,
8709 u64 rt_period, u64 rt_runtime)
8713 mutex_lock(&rt_constraints_mutex);
8714 read_lock(&tasklist_lock);
8715 err = __rt_schedulable(tg, rt_period, rt_runtime);
8719 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8720 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8721 tg->rt_bandwidth.rt_runtime = rt_runtime;
8723 for_each_possible_cpu(i) {
8724 struct rt_rq *rt_rq = tg->rt_rq[i];
8726 raw_spin_lock(&rt_rq->rt_runtime_lock);
8727 rt_rq->rt_runtime = rt_runtime;
8728 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8730 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8732 read_unlock(&tasklist_lock);
8733 mutex_unlock(&rt_constraints_mutex);
8738 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8740 u64 rt_runtime, rt_period;
8742 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8743 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8744 if (rt_runtime_us < 0)
8745 rt_runtime = RUNTIME_INF;
8747 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8750 long sched_group_rt_runtime(struct task_group *tg)
8754 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8757 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8758 do_div(rt_runtime_us, NSEC_PER_USEC);
8759 return rt_runtime_us;
8762 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8764 u64 rt_runtime, rt_period;
8766 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8767 rt_runtime = tg->rt_bandwidth.rt_runtime;
8772 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8775 long sched_group_rt_period(struct task_group *tg)
8779 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8780 do_div(rt_period_us, NSEC_PER_USEC);
8781 return rt_period_us;
8784 static int sched_rt_global_constraints(void)
8786 u64 runtime, period;
8789 if (sysctl_sched_rt_period <= 0)
8792 runtime = global_rt_runtime();
8793 period = global_rt_period();
8796 * Sanity check on the sysctl variables.
8798 if (runtime > period && runtime != RUNTIME_INF)
8801 mutex_lock(&rt_constraints_mutex);
8802 read_lock(&tasklist_lock);
8803 ret = __rt_schedulable(NULL, 0, 0);
8804 read_unlock(&tasklist_lock);
8805 mutex_unlock(&rt_constraints_mutex);
8810 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8812 /* Don't accept realtime tasks when there is no way for them to run */
8813 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8819 #else /* !CONFIG_RT_GROUP_SCHED */
8820 static int sched_rt_global_constraints(void)
8822 unsigned long flags;
8825 if (sysctl_sched_rt_period <= 0)
8829 * There's always some RT tasks in the root group
8830 * -- migration, kstopmachine etc..
8832 if (sysctl_sched_rt_runtime == 0)
8835 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8836 for_each_possible_cpu(i) {
8837 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8839 raw_spin_lock(&rt_rq->rt_runtime_lock);
8840 rt_rq->rt_runtime = global_rt_runtime();
8841 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8843 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8847 #endif /* CONFIG_RT_GROUP_SCHED */
8849 int sched_rt_handler(struct ctl_table *table, int write,
8850 void __user *buffer, size_t *lenp,
8854 int old_period, old_runtime;
8855 static DEFINE_MUTEX(mutex);
8858 old_period = sysctl_sched_rt_period;
8859 old_runtime = sysctl_sched_rt_runtime;
8861 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8863 if (!ret && write) {
8864 ret = sched_rt_global_constraints();
8866 sysctl_sched_rt_period = old_period;
8867 sysctl_sched_rt_runtime = old_runtime;
8869 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8870 def_rt_bandwidth.rt_period =
8871 ns_to_ktime(global_rt_period());
8874 mutex_unlock(&mutex);
8879 #ifdef CONFIG_CGROUP_SCHED
8881 /* return corresponding task_group object of a cgroup */
8882 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8884 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8885 struct task_group, css);
8888 static struct cgroup_subsys_state *
8889 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8891 struct task_group *tg, *parent;
8893 if (!cgrp->parent) {
8894 /* This is early initialization for the top cgroup */
8895 return &init_task_group.css;
8898 parent = cgroup_tg(cgrp->parent);
8899 tg = sched_create_group(parent);
8901 return ERR_PTR(-ENOMEM);
8907 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8909 struct task_group *tg = cgroup_tg(cgrp);
8911 sched_destroy_group(tg);
8915 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8917 #ifdef CONFIG_RT_GROUP_SCHED
8918 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8921 /* We don't support RT-tasks being in separate groups */
8922 if (tsk->sched_class != &fair_sched_class)
8929 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8930 struct task_struct *tsk, bool threadgroup)
8932 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8936 struct task_struct *c;
8938 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8939 retval = cpu_cgroup_can_attach_task(cgrp, c);
8951 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8952 struct cgroup *old_cont, struct task_struct *tsk,
8955 sched_move_task(tsk);
8957 struct task_struct *c;
8959 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8966 #ifdef CONFIG_FAIR_GROUP_SCHED
8967 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8970 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8973 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8975 struct task_group *tg = cgroup_tg(cgrp);
8977 return (u64) tg->shares;
8979 #endif /* CONFIG_FAIR_GROUP_SCHED */
8981 #ifdef CONFIG_RT_GROUP_SCHED
8982 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8985 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8988 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8990 return sched_group_rt_runtime(cgroup_tg(cgrp));
8993 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8996 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8999 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9001 return sched_group_rt_period(cgroup_tg(cgrp));
9003 #endif /* CONFIG_RT_GROUP_SCHED */
9005 static struct cftype cpu_files[] = {
9006 #ifdef CONFIG_FAIR_GROUP_SCHED
9009 .read_u64 = cpu_shares_read_u64,
9010 .write_u64 = cpu_shares_write_u64,
9013 #ifdef CONFIG_RT_GROUP_SCHED
9015 .name = "rt_runtime_us",
9016 .read_s64 = cpu_rt_runtime_read,
9017 .write_s64 = cpu_rt_runtime_write,
9020 .name = "rt_period_us",
9021 .read_u64 = cpu_rt_period_read_uint,
9022 .write_u64 = cpu_rt_period_write_uint,
9027 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9029 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9032 struct cgroup_subsys cpu_cgroup_subsys = {
9034 .create = cpu_cgroup_create,
9035 .destroy = cpu_cgroup_destroy,
9036 .can_attach = cpu_cgroup_can_attach,
9037 .attach = cpu_cgroup_attach,
9038 .populate = cpu_cgroup_populate,
9039 .subsys_id = cpu_cgroup_subsys_id,
9043 #endif /* CONFIG_CGROUP_SCHED */
9045 #ifdef CONFIG_CGROUP_CPUACCT
9048 * CPU accounting code for task groups.
9050 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9051 * (balbir@in.ibm.com).
9054 /* track cpu usage of a group of tasks and its child groups */
9056 struct cgroup_subsys_state css;
9057 /* cpuusage holds pointer to a u64-type object on every cpu */
9058 u64 __percpu *cpuusage;
9059 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9060 struct cpuacct *parent;
9063 struct cgroup_subsys cpuacct_subsys;
9065 /* return cpu accounting group corresponding to this container */
9066 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9068 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9069 struct cpuacct, css);
9072 /* return cpu accounting group to which this task belongs */
9073 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9075 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9076 struct cpuacct, css);
9079 /* create a new cpu accounting group */
9080 static struct cgroup_subsys_state *cpuacct_create(
9081 struct cgroup_subsys *ss, struct cgroup *cgrp)
9083 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9089 ca->cpuusage = alloc_percpu(u64);
9093 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9094 if (percpu_counter_init(&ca->cpustat[i], 0))
9095 goto out_free_counters;
9098 ca->parent = cgroup_ca(cgrp->parent);
9104 percpu_counter_destroy(&ca->cpustat[i]);
9105 free_percpu(ca->cpuusage);
9109 return ERR_PTR(-ENOMEM);
9112 /* destroy an existing cpu accounting group */
9114 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9116 struct cpuacct *ca = cgroup_ca(cgrp);
9119 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9120 percpu_counter_destroy(&ca->cpustat[i]);
9121 free_percpu(ca->cpuusage);
9125 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9127 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9130 #ifndef CONFIG_64BIT
9132 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9134 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9136 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9144 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9146 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9148 #ifndef CONFIG_64BIT
9150 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9152 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9154 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9160 /* return total cpu usage (in nanoseconds) of a group */
9161 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9163 struct cpuacct *ca = cgroup_ca(cgrp);
9164 u64 totalcpuusage = 0;
9167 for_each_present_cpu(i)
9168 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9170 return totalcpuusage;
9173 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9176 struct cpuacct *ca = cgroup_ca(cgrp);
9185 for_each_present_cpu(i)
9186 cpuacct_cpuusage_write(ca, i, 0);
9192 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9195 struct cpuacct *ca = cgroup_ca(cgroup);
9199 for_each_present_cpu(i) {
9200 percpu = cpuacct_cpuusage_read(ca, i);
9201 seq_printf(m, "%llu ", (unsigned long long) percpu);
9203 seq_printf(m, "\n");
9207 static const char *cpuacct_stat_desc[] = {
9208 [CPUACCT_STAT_USER] = "user",
9209 [CPUACCT_STAT_SYSTEM] = "system",
9212 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9213 struct cgroup_map_cb *cb)
9215 struct cpuacct *ca = cgroup_ca(cgrp);
9218 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9219 s64 val = percpu_counter_read(&ca->cpustat[i]);
9220 val = cputime64_to_clock_t(val);
9221 cb->fill(cb, cpuacct_stat_desc[i], val);
9226 static struct cftype files[] = {
9229 .read_u64 = cpuusage_read,
9230 .write_u64 = cpuusage_write,
9233 .name = "usage_percpu",
9234 .read_seq_string = cpuacct_percpu_seq_read,
9238 .read_map = cpuacct_stats_show,
9242 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9244 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9248 * charge this task's execution time to its accounting group.
9250 * called with rq->lock held.
9252 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9257 if (unlikely(!cpuacct_subsys.active))
9260 cpu = task_cpu(tsk);
9266 for (; ca; ca = ca->parent) {
9267 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9268 *cpuusage += cputime;
9275 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9276 * in cputime_t units. As a result, cpuacct_update_stats calls
9277 * percpu_counter_add with values large enough to always overflow the
9278 * per cpu batch limit causing bad SMP scalability.
9280 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9281 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9282 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9285 #define CPUACCT_BATCH \
9286 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9288 #define CPUACCT_BATCH 0
9292 * Charge the system/user time to the task's accounting group.
9294 static void cpuacct_update_stats(struct task_struct *tsk,
9295 enum cpuacct_stat_index idx, cputime_t val)
9298 int batch = CPUACCT_BATCH;
9300 if (unlikely(!cpuacct_subsys.active))
9307 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9313 struct cgroup_subsys cpuacct_subsys = {
9315 .create = cpuacct_create,
9316 .destroy = cpuacct_destroy,
9317 .populate = cpuacct_populate,
9318 .subsys_id = cpuacct_subsys_id,
9320 #endif /* CONFIG_CGROUP_CPUACCT */
9324 void synchronize_sched_expedited(void)
9328 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9330 #else /* #ifndef CONFIG_SMP */
9332 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
9334 static int synchronize_sched_expedited_cpu_stop(void *data)
9337 * There must be a full memory barrier on each affected CPU
9338 * between the time that try_stop_cpus() is called and the
9339 * time that it returns.
9341 * In the current initial implementation of cpu_stop, the
9342 * above condition is already met when the control reaches
9343 * this point and the following smp_mb() is not strictly
9344 * necessary. Do smp_mb() anyway for documentation and
9345 * robustness against future implementation changes.
9347 smp_mb(); /* See above comment block. */
9352 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9353 * approach to force grace period to end quickly. This consumes
9354 * significant time on all CPUs, and is thus not recommended for
9355 * any sort of common-case code.
9357 * Note that it is illegal to call this function while holding any
9358 * lock that is acquired by a CPU-hotplug notifier. Failing to
9359 * observe this restriction will result in deadlock.
9361 void synchronize_sched_expedited(void)
9363 int snap, trycount = 0;
9365 smp_mb(); /* ensure prior mod happens before capturing snap. */
9366 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
9368 while (try_stop_cpus(cpu_online_mask,
9369 synchronize_sched_expedited_cpu_stop,
9372 if (trycount++ < 10)
9373 udelay(trycount * num_online_cpus());
9375 synchronize_sched();
9378 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
9379 smp_mb(); /* ensure test happens before caller kfree */
9384 atomic_inc(&synchronize_sched_expedited_count);
9385 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9388 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9390 #endif /* #else #ifndef CONFIG_SMP */