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 <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/stop_machine.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/debugfs.h>
71 #include <linux/ctype.h>
72 #include <linux/ftrace.h>
73 #include <linux/slab.h>
76 #include <asm/irq_regs.h>
77 #include <asm/mutex.h>
78 #ifdef CONFIG_PARAVIRT
79 #include <asm/paravirt.h>
82 #include "sched_cpupri.h"
83 #include "workqueue_sched.h"
84 #include "sched_autogroup.h"
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/sched.h>
90 * Convert user-nice values [ -20 ... 0 ... 19 ]
91 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
94 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
95 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
96 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
99 * 'User priority' is the nice value converted to something we
100 * can work with better when scaling various scheduler parameters,
101 * it's a [ 0 ... 39 ] range.
103 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
104 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
105 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
108 * Helpers for converting nanosecond timing to jiffy resolution
110 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
112 #define NICE_0_LOAD SCHED_LOAD_SCALE
113 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
116 * These are the 'tuning knobs' of the scheduler:
118 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
119 * Timeslices get refilled after they expire.
121 #define DEF_TIMESLICE (100 * HZ / 1000)
124 * single value that denotes runtime == period, ie unlimited time.
126 #define RUNTIME_INF ((u64)~0ULL)
128 static inline int rt_policy(int policy)
130 if (policy == SCHED_FIFO || policy == SCHED_RR)
135 static inline int task_has_rt_policy(struct task_struct *p)
137 return rt_policy(p->policy);
141 * This is the priority-queue data structure of the RT scheduling class:
143 struct rt_prio_array {
144 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
145 struct list_head queue[MAX_RT_PRIO];
148 struct rt_bandwidth {
149 /* nests inside the rq lock: */
150 raw_spinlock_t rt_runtime_lock;
153 struct hrtimer rt_period_timer;
156 static struct rt_bandwidth def_rt_bandwidth;
158 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
160 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
162 struct rt_bandwidth *rt_b =
163 container_of(timer, struct rt_bandwidth, rt_period_timer);
169 now = hrtimer_cb_get_time(timer);
170 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
175 idle = do_sched_rt_period_timer(rt_b, overrun);
178 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
182 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
184 rt_b->rt_period = ns_to_ktime(period);
185 rt_b->rt_runtime = runtime;
187 raw_spin_lock_init(&rt_b->rt_runtime_lock);
189 hrtimer_init(&rt_b->rt_period_timer,
190 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
191 rt_b->rt_period_timer.function = sched_rt_period_timer;
194 static inline int rt_bandwidth_enabled(void)
196 return sysctl_sched_rt_runtime >= 0;
199 static void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
202 ktime_t soft, hard, now;
205 if (hrtimer_active(period_timer))
208 now = hrtimer_cb_get_time(period_timer);
209 hrtimer_forward(period_timer, now, period);
211 soft = hrtimer_get_softexpires(period_timer);
212 hard = hrtimer_get_expires(period_timer);
213 delta = ktime_to_ns(ktime_sub(hard, soft));
214 __hrtimer_start_range_ns(period_timer, soft, delta,
215 HRTIMER_MODE_ABS_PINNED, 0);
219 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
221 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
224 if (hrtimer_active(&rt_b->rt_period_timer))
227 raw_spin_lock(&rt_b->rt_runtime_lock);
228 start_bandwidth_timer(&rt_b->rt_period_timer, rt_b->rt_period);
229 raw_spin_unlock(&rt_b->rt_runtime_lock);
232 #ifdef CONFIG_RT_GROUP_SCHED
233 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
235 hrtimer_cancel(&rt_b->rt_period_timer);
240 * sched_domains_mutex serializes calls to init_sched_domains,
241 * detach_destroy_domains and partition_sched_domains.
243 static DEFINE_MUTEX(sched_domains_mutex);
245 #ifdef CONFIG_CGROUP_SCHED
247 #include <linux/cgroup.h>
251 static LIST_HEAD(task_groups);
253 struct cfs_bandwidth {
254 #ifdef CONFIG_CFS_BANDWIDTH
258 s64 hierarchal_quota;
261 int idle, timer_active;
262 struct hrtimer period_timer, slack_timer;
263 struct list_head throttled_cfs_rq;
266 int nr_periods, nr_throttled;
271 /* task group related information */
273 struct cgroup_subsys_state css;
275 #ifdef CONFIG_FAIR_GROUP_SCHED
276 /* schedulable entities of this group on each cpu */
277 struct sched_entity **se;
278 /* runqueue "owned" by this group on each cpu */
279 struct cfs_rq **cfs_rq;
280 unsigned long shares;
282 atomic_t load_weight;
285 #ifdef CONFIG_RT_GROUP_SCHED
286 struct sched_rt_entity **rt_se;
287 struct rt_rq **rt_rq;
289 struct rt_bandwidth rt_bandwidth;
293 struct list_head list;
295 struct task_group *parent;
296 struct list_head siblings;
297 struct list_head children;
299 #ifdef CONFIG_SCHED_AUTOGROUP
300 struct autogroup *autogroup;
303 struct cfs_bandwidth cfs_bandwidth;
306 /* task_group_lock serializes the addition/removal of task groups */
307 static DEFINE_SPINLOCK(task_group_lock);
309 #ifdef CONFIG_FAIR_GROUP_SCHED
311 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
314 * A weight of 0 or 1 can cause arithmetics problems.
315 * A weight of a cfs_rq is the sum of weights of which entities
316 * are queued on this cfs_rq, so a weight of a entity should not be
317 * too large, so as the shares value of a task group.
318 * (The default weight is 1024 - so there's no practical
319 * limitation from this.)
321 #define MIN_SHARES (1UL << 1)
322 #define MAX_SHARES (1UL << 18)
324 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
327 /* Default task group.
328 * Every task in system belong to this group at bootup.
330 struct task_group root_task_group;
332 #endif /* CONFIG_CGROUP_SCHED */
334 /* CFS-related fields in a runqueue */
336 struct load_weight load;
337 unsigned long nr_running, h_nr_running;
342 u64 min_vruntime_copy;
345 struct rb_root tasks_timeline;
346 struct rb_node *rb_leftmost;
348 struct list_head tasks;
349 struct list_head *balance_iterator;
352 * 'curr' points to currently running entity on this cfs_rq.
353 * It is set to NULL otherwise (i.e when none are currently running).
355 struct sched_entity *curr, *next, *last, *skip;
357 #ifdef CONFIG_SCHED_DEBUG
358 unsigned int nr_spread_over;
361 #ifdef CONFIG_FAIR_GROUP_SCHED
362 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
365 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
366 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
367 * (like users, containers etc.)
369 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
370 * list is used during load balance.
373 struct list_head leaf_cfs_rq_list;
374 struct task_group *tg; /* group that "owns" this runqueue */
378 * the part of load.weight contributed by tasks
380 unsigned long task_weight;
383 * h_load = weight * f(tg)
385 * Where f(tg) is the recursive weight fraction assigned to
388 unsigned long h_load;
391 * Maintaining per-cpu shares distribution for group scheduling
393 * load_stamp is the last time we updated the load average
394 * load_last is the last time we updated the load average and saw load
395 * load_unacc_exec_time is currently unaccounted execution time
399 u64 load_stamp, load_last, load_unacc_exec_time;
401 unsigned long load_contribution;
403 #ifdef CONFIG_CFS_BANDWIDTH
406 s64 runtime_remaining;
408 u64 throttled_timestamp;
409 int throttled, throttle_count;
410 struct list_head throttled_list;
415 #ifdef CONFIG_FAIR_GROUP_SCHED
416 #ifdef CONFIG_CFS_BANDWIDTH
417 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
419 return &tg->cfs_bandwidth;
422 static inline u64 default_cfs_period(void);
423 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
424 static void do_sched_cfs_slack_timer(struct cfs_bandwidth *cfs_b);
426 static enum hrtimer_restart sched_cfs_slack_timer(struct hrtimer *timer)
428 struct cfs_bandwidth *cfs_b =
429 container_of(timer, struct cfs_bandwidth, slack_timer);
430 do_sched_cfs_slack_timer(cfs_b);
432 return HRTIMER_NORESTART;
435 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
437 struct cfs_bandwidth *cfs_b =
438 container_of(timer, struct cfs_bandwidth, period_timer);
444 now = hrtimer_cb_get_time(timer);
445 overrun = hrtimer_forward(timer, now, cfs_b->period);
450 idle = do_sched_cfs_period_timer(cfs_b, overrun);
453 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
456 static void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
458 raw_spin_lock_init(&cfs_b->lock);
460 cfs_b->quota = RUNTIME_INF;
461 cfs_b->period = ns_to_ktime(default_cfs_period());
463 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
464 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
465 cfs_b->period_timer.function = sched_cfs_period_timer;
466 hrtimer_init(&cfs_b->slack_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
467 cfs_b->slack_timer.function = sched_cfs_slack_timer;
470 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
472 cfs_rq->runtime_enabled = 0;
473 INIT_LIST_HEAD(&cfs_rq->throttled_list);
476 /* requires cfs_b->lock, may release to reprogram timer */
477 static void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
480 * The timer may be active because we're trying to set a new bandwidth
481 * period or because we're racing with the tear-down path
482 * (timer_active==0 becomes visible before the hrtimer call-back
483 * terminates). In either case we ensure that it's re-programmed
485 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
486 raw_spin_unlock(&cfs_b->lock);
487 /* ensure cfs_b->lock is available while we wait */
488 hrtimer_cancel(&cfs_b->period_timer);
490 raw_spin_lock(&cfs_b->lock);
491 /* if someone else restarted the timer then we're done */
492 if (cfs_b->timer_active)
496 cfs_b->timer_active = 1;
497 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
500 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
502 hrtimer_cancel(&cfs_b->period_timer);
503 hrtimer_cancel(&cfs_b->slack_timer);
506 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
507 static void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
508 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
510 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
514 #endif /* CONFIG_CFS_BANDWIDTH */
515 #endif /* CONFIG_FAIR_GROUP_SCHED */
517 /* Real-Time classes' related field in a runqueue: */
519 struct rt_prio_array active;
520 unsigned long rt_nr_running;
521 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
523 int curr; /* highest queued rt task prio */
525 int next; /* next highest */
530 unsigned long rt_nr_migratory;
531 unsigned long rt_nr_total;
533 struct plist_head pushable_tasks;
538 /* Nests inside the rq lock: */
539 raw_spinlock_t rt_runtime_lock;
541 #ifdef CONFIG_RT_GROUP_SCHED
542 unsigned long rt_nr_boosted;
545 struct list_head leaf_rt_rq_list;
546 struct task_group *tg;
553 * We add the notion of a root-domain which will be used to define per-domain
554 * variables. Each exclusive cpuset essentially defines an island domain by
555 * fully partitioning the member cpus from any other cpuset. Whenever a new
556 * exclusive cpuset is created, we also create and attach a new root-domain
565 cpumask_var_t online;
568 * The "RT overload" flag: it gets set if a CPU has more than
569 * one runnable RT task.
571 cpumask_var_t rto_mask;
572 struct cpupri cpupri;
576 * By default the system creates a single root-domain with all cpus as
577 * members (mimicking the global state we have today).
579 static struct root_domain def_root_domain;
581 #endif /* CONFIG_SMP */
584 * This is the main, per-CPU runqueue data structure.
586 * Locking rule: those places that want to lock multiple runqueues
587 * (such as the load balancing or the thread migration code), lock
588 * acquire operations must be ordered by ascending &runqueue.
595 * nr_running and cpu_load should be in the same cacheline because
596 * remote CPUs use both these fields when doing load calculation.
598 unsigned long nr_running;
599 #define CPU_LOAD_IDX_MAX 5
600 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
601 unsigned long last_load_update_tick;
604 unsigned char nohz_balance_kick;
606 int skip_clock_update;
608 /* capture load from *all* tasks on this cpu: */
609 struct load_weight load;
610 unsigned long nr_load_updates;
616 #ifdef CONFIG_FAIR_GROUP_SCHED
617 /* list of leaf cfs_rq on this cpu: */
618 struct list_head leaf_cfs_rq_list;
620 #ifdef CONFIG_RT_GROUP_SCHED
621 struct list_head leaf_rt_rq_list;
625 * This is part of a global counter where only the total sum
626 * over all CPUs matters. A task can increase this counter on
627 * one CPU and if it got migrated afterwards it may decrease
628 * it on another CPU. Always updated under the runqueue lock:
630 unsigned long nr_uninterruptible;
632 struct task_struct *curr, *idle, *stop;
633 unsigned long next_balance;
634 struct mm_struct *prev_mm;
642 struct root_domain *rd;
643 struct sched_domain *sd;
645 unsigned long cpu_power;
647 unsigned char idle_at_tick;
648 /* For active balancing */
652 struct cpu_stop_work active_balance_work;
653 /* cpu of this runqueue: */
663 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
666 #ifdef CONFIG_PARAVIRT
669 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
670 u64 prev_steal_time_rq;
673 /* calc_load related fields */
674 unsigned long calc_load_update;
675 long calc_load_active;
677 #ifdef CONFIG_SCHED_HRTICK
679 int hrtick_csd_pending;
680 struct call_single_data hrtick_csd;
682 struct hrtimer hrtick_timer;
685 #ifdef CONFIG_SCHEDSTATS
687 struct sched_info rq_sched_info;
688 unsigned long long rq_cpu_time;
689 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
691 /* sys_sched_yield() stats */
692 unsigned int yld_count;
694 /* schedule() stats */
695 unsigned int sched_switch;
696 unsigned int sched_count;
697 unsigned int sched_goidle;
699 /* try_to_wake_up() stats */
700 unsigned int ttwu_count;
701 unsigned int ttwu_local;
705 struct task_struct *wake_list;
709 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
712 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
714 static inline int cpu_of(struct rq *rq)
723 #define rcu_dereference_check_sched_domain(p) \
724 rcu_dereference_check((p), \
725 lockdep_is_held(&sched_domains_mutex))
728 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
729 * See detach_destroy_domains: synchronize_sched for details.
731 * The domain tree of any CPU may only be accessed from within
732 * preempt-disabled sections.
734 #define for_each_domain(cpu, __sd) \
735 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
737 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
738 #define this_rq() (&__get_cpu_var(runqueues))
739 #define task_rq(p) cpu_rq(task_cpu(p))
740 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
741 #define raw_rq() (&__raw_get_cpu_var(runqueues))
743 #ifdef CONFIG_CGROUP_SCHED
746 * Return the group to which this tasks belongs.
748 * We use task_subsys_state_check() and extend the RCU verification with
749 * pi->lock and rq->lock because cpu_cgroup_attach() holds those locks for each
750 * task it moves into the cgroup. Therefore by holding either of those locks,
751 * we pin the task to the current cgroup.
753 static inline struct task_group *task_group(struct task_struct *p)
755 struct task_group *tg;
756 struct cgroup_subsys_state *css;
758 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
759 lockdep_is_held(&p->pi_lock) ||
760 lockdep_is_held(&task_rq(p)->lock));
761 tg = container_of(css, struct task_group, css);
763 return autogroup_task_group(p, tg);
766 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
767 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
769 #ifdef CONFIG_FAIR_GROUP_SCHED
770 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
771 p->se.parent = task_group(p)->se[cpu];
774 #ifdef CONFIG_RT_GROUP_SCHED
775 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
776 p->rt.parent = task_group(p)->rt_se[cpu];
780 #else /* CONFIG_CGROUP_SCHED */
782 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
783 static inline struct task_group *task_group(struct task_struct *p)
788 #endif /* CONFIG_CGROUP_SCHED */
790 static void update_rq_clock_task(struct rq *rq, s64 delta);
792 static void update_rq_clock(struct rq *rq)
796 if (rq->skip_clock_update > 0)
799 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
801 update_rq_clock_task(rq, delta);
805 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
807 #ifdef CONFIG_SCHED_DEBUG
808 # define const_debug __read_mostly
810 # define const_debug static const
814 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
815 * @cpu: the processor in question.
817 * This interface allows printk to be called with the runqueue lock
818 * held and know whether or not it is OK to wake up the klogd.
820 int runqueue_is_locked(int cpu)
822 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
826 * Debugging: various feature bits
829 #define SCHED_FEAT(name, enabled) \
830 __SCHED_FEAT_##name ,
833 #include "sched_features.h"
838 #define SCHED_FEAT(name, enabled) \
839 (1UL << __SCHED_FEAT_##name) * enabled |
841 const_debug unsigned int sysctl_sched_features =
842 #include "sched_features.h"
847 #ifdef CONFIG_SCHED_DEBUG
848 #define SCHED_FEAT(name, enabled) \
851 static __read_mostly char *sched_feat_names[] = {
852 #include "sched_features.h"
858 static int sched_feat_show(struct seq_file *m, void *v)
862 for (i = 0; sched_feat_names[i]; i++) {
863 if (!(sysctl_sched_features & (1UL << i)))
865 seq_printf(m, "%s ", sched_feat_names[i]);
873 sched_feat_write(struct file *filp, const char __user *ubuf,
874 size_t cnt, loff_t *ppos)
884 if (copy_from_user(&buf, ubuf, cnt))
890 if (strncmp(cmp, "NO_", 3) == 0) {
895 for (i = 0; sched_feat_names[i]; i++) {
896 if (strcmp(cmp, sched_feat_names[i]) == 0) {
898 sysctl_sched_features &= ~(1UL << i);
900 sysctl_sched_features |= (1UL << i);
905 if (!sched_feat_names[i])
913 static int sched_feat_open(struct inode *inode, struct file *filp)
915 return single_open(filp, sched_feat_show, NULL);
918 static const struct file_operations sched_feat_fops = {
919 .open = sched_feat_open,
920 .write = sched_feat_write,
923 .release = single_release,
926 static __init int sched_init_debug(void)
928 debugfs_create_file("sched_features", 0644, NULL, NULL,
933 late_initcall(sched_init_debug);
937 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
940 * Number of tasks to iterate in a single balance run.
941 * Limited because this is done with IRQs disabled.
943 const_debug unsigned int sysctl_sched_nr_migrate = 32;
946 * period over which we average the RT time consumption, measured
951 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
954 * period over which we measure -rt task cpu usage in us.
957 unsigned int sysctl_sched_rt_period = 1000000;
959 static __read_mostly int scheduler_running;
962 * part of the period that we allow rt tasks to run in us.
965 int sysctl_sched_rt_runtime = 950000;
967 static inline u64 global_rt_period(void)
969 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
972 static inline u64 global_rt_runtime(void)
974 if (sysctl_sched_rt_runtime < 0)
977 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
980 #ifndef prepare_arch_switch
981 # define prepare_arch_switch(next) do { } while (0)
983 #ifndef finish_arch_switch
984 # define finish_arch_switch(prev) do { } while (0)
987 static inline int task_current(struct rq *rq, struct task_struct *p)
989 return rq->curr == p;
992 static inline int task_running(struct rq *rq, struct task_struct *p)
997 return task_current(rq, p);
1001 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1002 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1006 * We can optimise this out completely for !SMP, because the
1007 * SMP rebalancing from interrupt is the only thing that cares
1014 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1018 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1019 * We must ensure this doesn't happen until the switch is completely
1025 #ifdef CONFIG_DEBUG_SPINLOCK
1026 /* this is a valid case when another task releases the spinlock */
1027 rq->lock.owner = current;
1030 * If we are tracking spinlock dependencies then we have to
1031 * fix up the runqueue lock - which gets 'carried over' from
1032 * prev into current:
1034 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
1036 raw_spin_unlock_irq(&rq->lock);
1039 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
1040 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1044 * We can optimise this out completely for !SMP, because the
1045 * SMP rebalancing from interrupt is the only thing that cares
1050 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1051 raw_spin_unlock_irq(&rq->lock);
1053 raw_spin_unlock(&rq->lock);
1057 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1061 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1062 * We must ensure this doesn't happen until the switch is completely
1068 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1072 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1075 * __task_rq_lock - lock the rq @p resides on.
1077 static inline struct rq *__task_rq_lock(struct task_struct *p)
1078 __acquires(rq->lock)
1082 lockdep_assert_held(&p->pi_lock);
1086 raw_spin_lock(&rq->lock);
1087 if (likely(rq == task_rq(p)))
1089 raw_spin_unlock(&rq->lock);
1094 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
1096 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1097 __acquires(p->pi_lock)
1098 __acquires(rq->lock)
1103 raw_spin_lock_irqsave(&p->pi_lock, *flags);
1105 raw_spin_lock(&rq->lock);
1106 if (likely(rq == task_rq(p)))
1108 raw_spin_unlock(&rq->lock);
1109 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1113 static void __task_rq_unlock(struct rq *rq)
1114 __releases(rq->lock)
1116 raw_spin_unlock(&rq->lock);
1120 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
1121 __releases(rq->lock)
1122 __releases(p->pi_lock)
1124 raw_spin_unlock(&rq->lock);
1125 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1129 * this_rq_lock - lock this runqueue and disable interrupts.
1131 static struct rq *this_rq_lock(void)
1132 __acquires(rq->lock)
1136 local_irq_disable();
1138 raw_spin_lock(&rq->lock);
1143 #ifdef CONFIG_SCHED_HRTICK
1145 * Use HR-timers to deliver accurate preemption points.
1147 * Its all a bit involved since we cannot program an hrt while holding the
1148 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1151 * When we get rescheduled we reprogram the hrtick_timer outside of the
1157 * - enabled by features
1158 * - hrtimer is actually high res
1160 static inline int hrtick_enabled(struct rq *rq)
1162 if (!sched_feat(HRTICK))
1164 if (!cpu_active(cpu_of(rq)))
1166 return hrtimer_is_hres_active(&rq->hrtick_timer);
1169 static void hrtick_clear(struct rq *rq)
1171 if (hrtimer_active(&rq->hrtick_timer))
1172 hrtimer_cancel(&rq->hrtick_timer);
1176 * High-resolution timer tick.
1177 * Runs from hardirq context with interrupts disabled.
1179 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1181 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1183 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1185 raw_spin_lock(&rq->lock);
1186 update_rq_clock(rq);
1187 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1188 raw_spin_unlock(&rq->lock);
1190 return HRTIMER_NORESTART;
1195 * called from hardirq (IPI) context
1197 static void __hrtick_start(void *arg)
1199 struct rq *rq = arg;
1201 raw_spin_lock(&rq->lock);
1202 hrtimer_restart(&rq->hrtick_timer);
1203 rq->hrtick_csd_pending = 0;
1204 raw_spin_unlock(&rq->lock);
1208 * Called to set the hrtick timer state.
1210 * called with rq->lock held and irqs disabled
1212 static void hrtick_start(struct rq *rq, u64 delay)
1214 struct hrtimer *timer = &rq->hrtick_timer;
1215 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1217 hrtimer_set_expires(timer, time);
1219 if (rq == this_rq()) {
1220 hrtimer_restart(timer);
1221 } else if (!rq->hrtick_csd_pending) {
1222 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1223 rq->hrtick_csd_pending = 1;
1228 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1230 int cpu = (int)(long)hcpu;
1233 case CPU_UP_CANCELED:
1234 case CPU_UP_CANCELED_FROZEN:
1235 case CPU_DOWN_PREPARE:
1236 case CPU_DOWN_PREPARE_FROZEN:
1238 case CPU_DEAD_FROZEN:
1239 hrtick_clear(cpu_rq(cpu));
1246 static __init void init_hrtick(void)
1248 hotcpu_notifier(hotplug_hrtick, 0);
1252 * Called to set the hrtick timer state.
1254 * called with rq->lock held and irqs disabled
1256 static void hrtick_start(struct rq *rq, u64 delay)
1258 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1259 HRTIMER_MODE_REL_PINNED, 0);
1262 static inline void init_hrtick(void)
1265 #endif /* CONFIG_SMP */
1267 static void init_rq_hrtick(struct rq *rq)
1270 rq->hrtick_csd_pending = 0;
1272 rq->hrtick_csd.flags = 0;
1273 rq->hrtick_csd.func = __hrtick_start;
1274 rq->hrtick_csd.info = rq;
1277 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1278 rq->hrtick_timer.function = hrtick;
1280 #else /* CONFIG_SCHED_HRTICK */
1281 static inline void hrtick_clear(struct rq *rq)
1285 static inline void init_rq_hrtick(struct rq *rq)
1289 static inline void init_hrtick(void)
1292 #endif /* CONFIG_SCHED_HRTICK */
1295 * resched_task - mark a task 'to be rescheduled now'.
1297 * On UP this means the setting of the need_resched flag, on SMP it
1298 * might also involve a cross-CPU call to trigger the scheduler on
1303 #ifndef tsk_is_polling
1304 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1307 static void resched_task(struct task_struct *p)
1311 assert_raw_spin_locked(&task_rq(p)->lock);
1313 if (test_tsk_need_resched(p))
1316 set_tsk_need_resched(p);
1319 if (cpu == smp_processor_id())
1322 /* NEED_RESCHED must be visible before we test polling */
1324 if (!tsk_is_polling(p))
1325 smp_send_reschedule(cpu);
1328 static void resched_cpu(int cpu)
1330 struct rq *rq = cpu_rq(cpu);
1331 unsigned long flags;
1333 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1335 resched_task(cpu_curr(cpu));
1336 raw_spin_unlock_irqrestore(&rq->lock, flags);
1341 * In the semi idle case, use the nearest busy cpu for migrating timers
1342 * from an idle cpu. This is good for power-savings.
1344 * We don't do similar optimization for completely idle system, as
1345 * selecting an idle cpu will add more delays to the timers than intended
1346 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1348 int get_nohz_timer_target(void)
1350 int cpu = smp_processor_id();
1352 struct sched_domain *sd;
1355 for_each_domain(cpu, sd) {
1356 for_each_cpu(i, sched_domain_span(sd)) {
1368 * When add_timer_on() enqueues a timer into the timer wheel of an
1369 * idle CPU then this timer might expire before the next timer event
1370 * which is scheduled to wake up that CPU. In case of a completely
1371 * idle system the next event might even be infinite time into the
1372 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1373 * leaves the inner idle loop so the newly added timer is taken into
1374 * account when the CPU goes back to idle and evaluates the timer
1375 * wheel for the next timer event.
1377 void wake_up_idle_cpu(int cpu)
1379 struct rq *rq = cpu_rq(cpu);
1381 if (cpu == smp_processor_id())
1385 * This is safe, as this function is called with the timer
1386 * wheel base lock of (cpu) held. When the CPU is on the way
1387 * to idle and has not yet set rq->curr to idle then it will
1388 * be serialized on the timer wheel base lock and take the new
1389 * timer into account automatically.
1391 if (rq->curr != rq->idle)
1395 * We can set TIF_RESCHED on the idle task of the other CPU
1396 * lockless. The worst case is that the other CPU runs the
1397 * idle task through an additional NOOP schedule()
1399 set_tsk_need_resched(rq->idle);
1401 /* NEED_RESCHED must be visible before we test polling */
1403 if (!tsk_is_polling(rq->idle))
1404 smp_send_reschedule(cpu);
1407 #endif /* CONFIG_NO_HZ */
1409 static u64 sched_avg_period(void)
1411 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1414 static void sched_avg_update(struct rq *rq)
1416 s64 period = sched_avg_period();
1418 while ((s64)(rq->clock - rq->age_stamp) > period) {
1420 * Inline assembly required to prevent the compiler
1421 * optimising this loop into a divmod call.
1422 * See __iter_div_u64_rem() for another example of this.
1424 asm("" : "+rm" (rq->age_stamp));
1425 rq->age_stamp += period;
1430 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1432 rq->rt_avg += rt_delta;
1433 sched_avg_update(rq);
1436 #else /* !CONFIG_SMP */
1437 static void resched_task(struct task_struct *p)
1439 assert_raw_spin_locked(&task_rq(p)->lock);
1440 set_tsk_need_resched(p);
1443 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1447 static void sched_avg_update(struct rq *rq)
1450 #endif /* CONFIG_SMP */
1452 #if BITS_PER_LONG == 32
1453 # define WMULT_CONST (~0UL)
1455 # define WMULT_CONST (1UL << 32)
1458 #define WMULT_SHIFT 32
1461 * Shift right and round:
1463 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1466 * delta *= weight / lw
1468 static unsigned long
1469 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1470 struct load_weight *lw)
1475 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1476 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1477 * 2^SCHED_LOAD_RESOLUTION.
1479 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
1480 tmp = (u64)delta_exec * scale_load_down(weight);
1482 tmp = (u64)delta_exec;
1484 if (!lw->inv_weight) {
1485 unsigned long w = scale_load_down(lw->weight);
1487 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
1489 else if (unlikely(!w))
1490 lw->inv_weight = WMULT_CONST;
1492 lw->inv_weight = WMULT_CONST / w;
1496 * Check whether we'd overflow the 64-bit multiplication:
1498 if (unlikely(tmp > WMULT_CONST))
1499 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1502 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1504 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1507 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1513 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1519 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1526 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1527 * of tasks with abnormal "nice" values across CPUs the contribution that
1528 * each task makes to its run queue's load is weighted according to its
1529 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1530 * scaled version of the new time slice allocation that they receive on time
1534 #define WEIGHT_IDLEPRIO 3
1535 #define WMULT_IDLEPRIO 1431655765
1538 * Nice levels are multiplicative, with a gentle 10% change for every
1539 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1540 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1541 * that remained on nice 0.
1543 * The "10% effect" is relative and cumulative: from _any_ nice level,
1544 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1545 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1546 * If a task goes up by ~10% and another task goes down by ~10% then
1547 * the relative distance between them is ~25%.)
1549 static const int prio_to_weight[40] = {
1550 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1551 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1552 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1553 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1554 /* 0 */ 1024, 820, 655, 526, 423,
1555 /* 5 */ 335, 272, 215, 172, 137,
1556 /* 10 */ 110, 87, 70, 56, 45,
1557 /* 15 */ 36, 29, 23, 18, 15,
1561 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1563 * In cases where the weight does not change often, we can use the
1564 * precalculated inverse to speed up arithmetics by turning divisions
1565 * into multiplications:
1567 static const u32 prio_to_wmult[40] = {
1568 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1569 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1570 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1571 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1572 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1573 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1574 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1575 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1578 /* Time spent by the tasks of the cpu accounting group executing in ... */
1579 enum cpuacct_stat_index {
1580 CPUACCT_STAT_USER, /* ... user mode */
1581 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1583 CPUACCT_STAT_NSTATS,
1586 #ifdef CONFIG_CGROUP_CPUACCT
1587 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1588 static void cpuacct_update_stats(struct task_struct *tsk,
1589 enum cpuacct_stat_index idx, cputime_t val);
1591 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1592 static inline void cpuacct_update_stats(struct task_struct *tsk,
1593 enum cpuacct_stat_index idx, cputime_t val) {}
1596 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1598 update_load_add(&rq->load, load);
1601 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1603 update_load_sub(&rq->load, load);
1606 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1607 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1608 typedef int (*tg_visitor)(struct task_group *, void *);
1611 * Iterate task_group tree rooted at *from, calling @down when first entering a
1612 * node and @up when leaving it for the final time.
1614 * Caller must hold rcu_lock or sufficient equivalent.
1616 static int walk_tg_tree_from(struct task_group *from,
1617 tg_visitor down, tg_visitor up, void *data)
1619 struct task_group *parent, *child;
1625 ret = (*down)(parent, data);
1628 list_for_each_entry_rcu(child, &parent->children, siblings) {
1635 ret = (*up)(parent, data);
1636 if (ret || parent == from)
1640 parent = parent->parent;
1648 * Iterate the full tree, calling @down when first entering a node and @up when
1649 * leaving it for the final time.
1651 * Caller must hold rcu_lock or sufficient equivalent.
1654 static inline int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1656 return walk_tg_tree_from(&root_task_group, down, up, data);
1659 static int tg_nop(struct task_group *tg, void *data)
1666 /* Used instead of source_load when we know the type == 0 */
1667 static unsigned long weighted_cpuload(const int cpu)
1669 return cpu_rq(cpu)->load.weight;
1673 * Return a low guess at the load of a migration-source cpu weighted
1674 * according to the scheduling class and "nice" value.
1676 * We want to under-estimate the load of migration sources, to
1677 * balance conservatively.
1679 static unsigned long source_load(int cpu, int type)
1681 struct rq *rq = cpu_rq(cpu);
1682 unsigned long total = weighted_cpuload(cpu);
1684 if (type == 0 || !sched_feat(LB_BIAS))
1687 return min(rq->cpu_load[type-1], total);
1691 * Return a high guess at the load of a migration-target cpu weighted
1692 * according to the scheduling class and "nice" value.
1694 static unsigned long target_load(int cpu, int type)
1696 struct rq *rq = cpu_rq(cpu);
1697 unsigned long total = weighted_cpuload(cpu);
1699 if (type == 0 || !sched_feat(LB_BIAS))
1702 return max(rq->cpu_load[type-1], total);
1705 static unsigned long power_of(int cpu)
1707 return cpu_rq(cpu)->cpu_power;
1710 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1712 static unsigned long cpu_avg_load_per_task(int cpu)
1714 struct rq *rq = cpu_rq(cpu);
1715 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1718 return rq->load.weight / nr_running;
1723 #ifdef CONFIG_PREEMPT
1725 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1728 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1729 * way at the expense of forcing extra atomic operations in all
1730 * invocations. This assures that the double_lock is acquired using the
1731 * same underlying policy as the spinlock_t on this architecture, which
1732 * reduces latency compared to the unfair variant below. However, it
1733 * also adds more overhead and therefore may reduce throughput.
1735 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1736 __releases(this_rq->lock)
1737 __acquires(busiest->lock)
1738 __acquires(this_rq->lock)
1740 raw_spin_unlock(&this_rq->lock);
1741 double_rq_lock(this_rq, busiest);
1748 * Unfair double_lock_balance: Optimizes throughput at the expense of
1749 * latency by eliminating extra atomic operations when the locks are
1750 * already in proper order on entry. This favors lower cpu-ids and will
1751 * grant the double lock to lower cpus over higher ids under contention,
1752 * regardless of entry order into the function.
1754 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1755 __releases(this_rq->lock)
1756 __acquires(busiest->lock)
1757 __acquires(this_rq->lock)
1761 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1762 if (busiest < this_rq) {
1763 raw_spin_unlock(&this_rq->lock);
1764 raw_spin_lock(&busiest->lock);
1765 raw_spin_lock_nested(&this_rq->lock,
1766 SINGLE_DEPTH_NESTING);
1769 raw_spin_lock_nested(&busiest->lock,
1770 SINGLE_DEPTH_NESTING);
1775 #endif /* CONFIG_PREEMPT */
1778 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1780 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1782 if (unlikely(!irqs_disabled())) {
1783 /* printk() doesn't work good under rq->lock */
1784 raw_spin_unlock(&this_rq->lock);
1788 return _double_lock_balance(this_rq, busiest);
1791 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1792 __releases(busiest->lock)
1794 raw_spin_unlock(&busiest->lock);
1795 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1799 * double_rq_lock - safely lock two runqueues
1801 * Note this does not disable interrupts like task_rq_lock,
1802 * you need to do so manually before calling.
1804 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1805 __acquires(rq1->lock)
1806 __acquires(rq2->lock)
1808 BUG_ON(!irqs_disabled());
1810 raw_spin_lock(&rq1->lock);
1811 __acquire(rq2->lock); /* Fake it out ;) */
1814 raw_spin_lock(&rq1->lock);
1815 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1817 raw_spin_lock(&rq2->lock);
1818 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1824 * double_rq_unlock - safely unlock two runqueues
1826 * Note this does not restore interrupts like task_rq_unlock,
1827 * you need to do so manually after calling.
1829 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1830 __releases(rq1->lock)
1831 __releases(rq2->lock)
1833 raw_spin_unlock(&rq1->lock);
1835 raw_spin_unlock(&rq2->lock);
1837 __release(rq2->lock);
1840 #else /* CONFIG_SMP */
1843 * double_rq_lock - safely lock two runqueues
1845 * Note this does not disable interrupts like task_rq_lock,
1846 * you need to do so manually before calling.
1848 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1849 __acquires(rq1->lock)
1850 __acquires(rq2->lock)
1852 BUG_ON(!irqs_disabled());
1854 raw_spin_lock(&rq1->lock);
1855 __acquire(rq2->lock); /* Fake it out ;) */
1859 * double_rq_unlock - safely unlock two runqueues
1861 * Note this does not restore interrupts like task_rq_unlock,
1862 * you need to do so manually after calling.
1864 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1865 __releases(rq1->lock)
1866 __releases(rq2->lock)
1869 raw_spin_unlock(&rq1->lock);
1870 __release(rq2->lock);
1875 static void calc_load_account_idle(struct rq *this_rq);
1876 static void update_sysctl(void);
1877 static int get_update_sysctl_factor(void);
1878 static void update_cpu_load(struct rq *this_rq);
1880 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1882 set_task_rq(p, cpu);
1885 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1886 * successfuly executed on another CPU. We must ensure that updates of
1887 * per-task data have been completed by this moment.
1890 task_thread_info(p)->cpu = cpu;
1894 static const struct sched_class rt_sched_class;
1896 #define sched_class_highest (&stop_sched_class)
1897 #define for_each_class(class) \
1898 for (class = sched_class_highest; class; class = class->next)
1900 #include "sched_stats.h"
1902 static void inc_nr_running(struct rq *rq)
1907 static void dec_nr_running(struct rq *rq)
1912 static void set_load_weight(struct task_struct *p)
1914 int prio = p->static_prio - MAX_RT_PRIO;
1915 struct load_weight *load = &p->se.load;
1918 * SCHED_IDLE tasks get minimal weight:
1920 if (p->policy == SCHED_IDLE) {
1921 load->weight = scale_load(WEIGHT_IDLEPRIO);
1922 load->inv_weight = WMULT_IDLEPRIO;
1926 load->weight = scale_load(prio_to_weight[prio]);
1927 load->inv_weight = prio_to_wmult[prio];
1930 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1932 update_rq_clock(rq);
1933 sched_info_queued(p);
1934 p->sched_class->enqueue_task(rq, p, flags);
1937 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1939 update_rq_clock(rq);
1940 sched_info_dequeued(p);
1941 p->sched_class->dequeue_task(rq, p, flags);
1945 * activate_task - move a task to the runqueue.
1947 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1949 if (task_contributes_to_load(p))
1950 rq->nr_uninterruptible--;
1952 enqueue_task(rq, p, flags);
1956 * deactivate_task - remove a task from the runqueue.
1958 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1960 if (task_contributes_to_load(p))
1961 rq->nr_uninterruptible++;
1963 dequeue_task(rq, p, flags);
1966 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1969 * There are no locks covering percpu hardirq/softirq time.
1970 * They are only modified in account_system_vtime, on corresponding CPU
1971 * with interrupts disabled. So, writes are safe.
1972 * They are read and saved off onto struct rq in update_rq_clock().
1973 * This may result in other CPU reading this CPU's irq time and can
1974 * race with irq/account_system_vtime on this CPU. We would either get old
1975 * or new value with a side effect of accounting a slice of irq time to wrong
1976 * task when irq is in progress while we read rq->clock. That is a worthy
1977 * compromise in place of having locks on each irq in account_system_time.
1979 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1980 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1982 static DEFINE_PER_CPU(u64, irq_start_time);
1983 static int sched_clock_irqtime;
1985 void enable_sched_clock_irqtime(void)
1987 sched_clock_irqtime = 1;
1990 void disable_sched_clock_irqtime(void)
1992 sched_clock_irqtime = 0;
1995 #ifndef CONFIG_64BIT
1996 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1998 static inline void irq_time_write_begin(void)
2000 __this_cpu_inc(irq_time_seq.sequence);
2004 static inline void irq_time_write_end(void)
2007 __this_cpu_inc(irq_time_seq.sequence);
2010 static inline u64 irq_time_read(int cpu)
2016 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
2017 irq_time = per_cpu(cpu_softirq_time, cpu) +
2018 per_cpu(cpu_hardirq_time, cpu);
2019 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
2023 #else /* CONFIG_64BIT */
2024 static inline void irq_time_write_begin(void)
2028 static inline void irq_time_write_end(void)
2032 static inline u64 irq_time_read(int cpu)
2034 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
2036 #endif /* CONFIG_64BIT */
2039 * Called before incrementing preempt_count on {soft,}irq_enter
2040 * and before decrementing preempt_count on {soft,}irq_exit.
2042 void account_system_vtime(struct task_struct *curr)
2044 unsigned long flags;
2048 if (!sched_clock_irqtime)
2051 local_irq_save(flags);
2053 cpu = smp_processor_id();
2054 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
2055 __this_cpu_add(irq_start_time, delta);
2057 irq_time_write_begin();
2059 * We do not account for softirq time from ksoftirqd here.
2060 * We want to continue accounting softirq time to ksoftirqd thread
2061 * in that case, so as not to confuse scheduler with a special task
2062 * that do not consume any time, but still wants to run.
2064 if (hardirq_count())
2065 __this_cpu_add(cpu_hardirq_time, delta);
2066 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
2067 __this_cpu_add(cpu_softirq_time, delta);
2069 irq_time_write_end();
2070 local_irq_restore(flags);
2072 EXPORT_SYMBOL_GPL(account_system_vtime);
2074 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2076 #ifdef CONFIG_PARAVIRT
2077 static inline u64 steal_ticks(u64 steal)
2079 if (unlikely(steal > NSEC_PER_SEC))
2080 return div_u64(steal, TICK_NSEC);
2082 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
2086 static void update_rq_clock_task(struct rq *rq, s64 delta)
2089 * In theory, the compile should just see 0 here, and optimize out the call
2090 * to sched_rt_avg_update. But I don't trust it...
2092 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2093 s64 steal = 0, irq_delta = 0;
2095 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2096 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
2099 * Since irq_time is only updated on {soft,}irq_exit, we might run into
2100 * this case when a previous update_rq_clock() happened inside a
2101 * {soft,}irq region.
2103 * When this happens, we stop ->clock_task and only update the
2104 * prev_irq_time stamp to account for the part that fit, so that a next
2105 * update will consume the rest. This ensures ->clock_task is
2108 * It does however cause some slight miss-attribution of {soft,}irq
2109 * time, a more accurate solution would be to update the irq_time using
2110 * the current rq->clock timestamp, except that would require using
2113 if (irq_delta > delta)
2116 rq->prev_irq_time += irq_delta;
2119 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
2120 if (static_branch((¶virt_steal_rq_enabled))) {
2123 steal = paravirt_steal_clock(cpu_of(rq));
2124 steal -= rq->prev_steal_time_rq;
2126 if (unlikely(steal > delta))
2129 st = steal_ticks(steal);
2130 steal = st * TICK_NSEC;
2132 rq->prev_steal_time_rq += steal;
2138 rq->clock_task += delta;
2140 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2141 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
2142 sched_rt_avg_update(rq, irq_delta + steal);
2146 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2147 static int irqtime_account_hi_update(void)
2149 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2150 unsigned long flags;
2154 local_irq_save(flags);
2155 latest_ns = this_cpu_read(cpu_hardirq_time);
2156 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
2158 local_irq_restore(flags);
2162 static int irqtime_account_si_update(void)
2164 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2165 unsigned long flags;
2169 local_irq_save(flags);
2170 latest_ns = this_cpu_read(cpu_softirq_time);
2171 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2173 local_irq_restore(flags);
2177 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2179 #define sched_clock_irqtime (0)
2183 #include "sched_idletask.c"
2184 #include "sched_fair.c"
2185 #include "sched_rt.c"
2186 #include "sched_autogroup.c"
2187 #include "sched_stoptask.c"
2188 #ifdef CONFIG_SCHED_DEBUG
2189 # include "sched_debug.c"
2192 void sched_set_stop_task(int cpu, struct task_struct *stop)
2194 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2195 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2199 * Make it appear like a SCHED_FIFO task, its something
2200 * userspace knows about and won't get confused about.
2202 * Also, it will make PI more or less work without too
2203 * much confusion -- but then, stop work should not
2204 * rely on PI working anyway.
2206 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2208 stop->sched_class = &stop_sched_class;
2211 cpu_rq(cpu)->stop = stop;
2215 * Reset it back to a normal scheduling class so that
2216 * it can die in pieces.
2218 old_stop->sched_class = &rt_sched_class;
2223 * __normal_prio - return the priority that is based on the static prio
2225 static inline int __normal_prio(struct task_struct *p)
2227 return p->static_prio;
2231 * Calculate the expected normal priority: i.e. priority
2232 * without taking RT-inheritance into account. Might be
2233 * boosted by interactivity modifiers. Changes upon fork,
2234 * setprio syscalls, and whenever the interactivity
2235 * estimator recalculates.
2237 static inline int normal_prio(struct task_struct *p)
2241 if (task_has_rt_policy(p))
2242 prio = MAX_RT_PRIO-1 - p->rt_priority;
2244 prio = __normal_prio(p);
2249 * Calculate the current priority, i.e. the priority
2250 * taken into account by the scheduler. This value might
2251 * be boosted by RT tasks, or might be boosted by
2252 * interactivity modifiers. Will be RT if the task got
2253 * RT-boosted. If not then it returns p->normal_prio.
2255 static int effective_prio(struct task_struct *p)
2257 p->normal_prio = normal_prio(p);
2259 * If we are RT tasks or we were boosted to RT priority,
2260 * keep the priority unchanged. Otherwise, update priority
2261 * to the normal priority:
2263 if (!rt_prio(p->prio))
2264 return p->normal_prio;
2269 * task_curr - is this task currently executing on a CPU?
2270 * @p: the task in question.
2272 inline int task_curr(const struct task_struct *p)
2274 return cpu_curr(task_cpu(p)) == p;
2277 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2278 const struct sched_class *prev_class,
2281 if (prev_class != p->sched_class) {
2282 if (prev_class->switched_from)
2283 prev_class->switched_from(rq, p);
2284 p->sched_class->switched_to(rq, p);
2285 } else if (oldprio != p->prio)
2286 p->sched_class->prio_changed(rq, p, oldprio);
2289 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2291 const struct sched_class *class;
2293 if (p->sched_class == rq->curr->sched_class) {
2294 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2296 for_each_class(class) {
2297 if (class == rq->curr->sched_class)
2299 if (class == p->sched_class) {
2300 resched_task(rq->curr);
2307 * A queue event has occurred, and we're going to schedule. In
2308 * this case, we can save a useless back to back clock update.
2310 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2311 rq->skip_clock_update = 1;
2316 * Is this task likely cache-hot:
2319 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2323 if (p->sched_class != &fair_sched_class)
2326 if (unlikely(p->policy == SCHED_IDLE))
2330 * Buddy candidates are cache hot:
2332 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2333 (&p->se == cfs_rq_of(&p->se)->next ||
2334 &p->se == cfs_rq_of(&p->se)->last))
2337 if (sysctl_sched_migration_cost == -1)
2339 if (sysctl_sched_migration_cost == 0)
2342 delta = now - p->se.exec_start;
2344 return delta < (s64)sysctl_sched_migration_cost;
2347 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2349 #ifdef CONFIG_SCHED_DEBUG
2351 * We should never call set_task_cpu() on a blocked task,
2352 * ttwu() will sort out the placement.
2354 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2355 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2357 #ifdef CONFIG_LOCKDEP
2359 * The caller should hold either p->pi_lock or rq->lock, when changing
2360 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2362 * sched_move_task() holds both and thus holding either pins the cgroup,
2363 * see set_task_rq().
2365 * Furthermore, all task_rq users should acquire both locks, see
2368 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2369 lockdep_is_held(&task_rq(p)->lock)));
2373 trace_sched_migrate_task(p, new_cpu);
2375 if (task_cpu(p) != new_cpu) {
2376 p->se.nr_migrations++;
2377 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
2380 __set_task_cpu(p, new_cpu);
2383 struct migration_arg {
2384 struct task_struct *task;
2388 static int migration_cpu_stop(void *data);
2391 * wait_task_inactive - wait for a thread to unschedule.
2393 * If @match_state is nonzero, it's the @p->state value just checked and
2394 * not expected to change. If it changes, i.e. @p might have woken up,
2395 * then return zero. When we succeed in waiting for @p to be off its CPU,
2396 * we return a positive number (its total switch count). If a second call
2397 * a short while later returns the same number, the caller can be sure that
2398 * @p has remained unscheduled the whole time.
2400 * The caller must ensure that the task *will* unschedule sometime soon,
2401 * else this function might spin for a *long* time. This function can't
2402 * be called with interrupts off, or it may introduce deadlock with
2403 * smp_call_function() if an IPI is sent by the same process we are
2404 * waiting to become inactive.
2406 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2408 unsigned long flags;
2415 * We do the initial early heuristics without holding
2416 * any task-queue locks at all. We'll only try to get
2417 * the runqueue lock when things look like they will
2423 * If the task is actively running on another CPU
2424 * still, just relax and busy-wait without holding
2427 * NOTE! Since we don't hold any locks, it's not
2428 * even sure that "rq" stays as the right runqueue!
2429 * But we don't care, since "task_running()" will
2430 * return false if the runqueue has changed and p
2431 * is actually now running somewhere else!
2433 while (task_running(rq, p)) {
2434 if (match_state && unlikely(p->state != match_state))
2440 * Ok, time to look more closely! We need the rq
2441 * lock now, to be *sure*. If we're wrong, we'll
2442 * just go back and repeat.
2444 rq = task_rq_lock(p, &flags);
2445 trace_sched_wait_task(p);
2446 running = task_running(rq, p);
2449 if (!match_state || p->state == match_state)
2450 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2451 task_rq_unlock(rq, p, &flags);
2454 * If it changed from the expected state, bail out now.
2456 if (unlikely(!ncsw))
2460 * Was it really running after all now that we
2461 * checked with the proper locks actually held?
2463 * Oops. Go back and try again..
2465 if (unlikely(running)) {
2471 * It's not enough that it's not actively running,
2472 * it must be off the runqueue _entirely_, and not
2475 * So if it was still runnable (but just not actively
2476 * running right now), it's preempted, and we should
2477 * yield - it could be a while.
2479 if (unlikely(on_rq)) {
2480 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2482 set_current_state(TASK_UNINTERRUPTIBLE);
2483 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2488 * Ahh, all good. It wasn't running, and it wasn't
2489 * runnable, which means that it will never become
2490 * running in the future either. We're all done!
2499 * kick_process - kick a running thread to enter/exit the kernel
2500 * @p: the to-be-kicked thread
2502 * Cause a process which is running on another CPU to enter
2503 * kernel-mode, without any delay. (to get signals handled.)
2505 * NOTE: this function doesn't have to take the runqueue lock,
2506 * because all it wants to ensure is that the remote task enters
2507 * the kernel. If the IPI races and the task has been migrated
2508 * to another CPU then no harm is done and the purpose has been
2511 void kick_process(struct task_struct *p)
2517 if ((cpu != smp_processor_id()) && task_curr(p))
2518 smp_send_reschedule(cpu);
2521 EXPORT_SYMBOL_GPL(kick_process);
2522 #endif /* CONFIG_SMP */
2526 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2528 static int select_fallback_rq(int cpu, struct task_struct *p)
2531 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2533 /* Look for allowed, online CPU in same node. */
2534 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2535 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2538 /* Any allowed, online CPU? */
2539 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2540 if (dest_cpu < nr_cpu_ids)
2543 /* No more Mr. Nice Guy. */
2544 dest_cpu = cpuset_cpus_allowed_fallback(p);
2546 * Don't tell them about moving exiting tasks or
2547 * kernel threads (both mm NULL), since they never
2550 if (p->mm && printk_ratelimit()) {
2551 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2552 task_pid_nr(p), p->comm, cpu);
2559 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2562 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2564 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2567 * In order not to call set_task_cpu() on a blocking task we need
2568 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2571 * Since this is common to all placement strategies, this lives here.
2573 * [ this allows ->select_task() to simply return task_cpu(p) and
2574 * not worry about this generic constraint ]
2576 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2578 cpu = select_fallback_rq(task_cpu(p), p);
2583 static void update_avg(u64 *avg, u64 sample)
2585 s64 diff = sample - *avg;
2591 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2593 #ifdef CONFIG_SCHEDSTATS
2594 struct rq *rq = this_rq();
2597 int this_cpu = smp_processor_id();
2599 if (cpu == this_cpu) {
2600 schedstat_inc(rq, ttwu_local);
2601 schedstat_inc(p, se.statistics.nr_wakeups_local);
2603 struct sched_domain *sd;
2605 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2607 for_each_domain(this_cpu, sd) {
2608 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2609 schedstat_inc(sd, ttwu_wake_remote);
2616 if (wake_flags & WF_MIGRATED)
2617 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2619 #endif /* CONFIG_SMP */
2621 schedstat_inc(rq, ttwu_count);
2622 schedstat_inc(p, se.statistics.nr_wakeups);
2624 if (wake_flags & WF_SYNC)
2625 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2627 #endif /* CONFIG_SCHEDSTATS */
2630 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2632 activate_task(rq, p, en_flags);
2635 /* if a worker is waking up, notify workqueue */
2636 if (p->flags & PF_WQ_WORKER)
2637 wq_worker_waking_up(p, cpu_of(rq));
2641 * Mark the task runnable and perform wakeup-preemption.
2644 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2646 trace_sched_wakeup(p, true);
2647 check_preempt_curr(rq, p, wake_flags);
2649 p->state = TASK_RUNNING;
2651 if (p->sched_class->task_woken)
2652 p->sched_class->task_woken(rq, p);
2654 if (rq->idle_stamp) {
2655 u64 delta = rq->clock - rq->idle_stamp;
2656 u64 max = 2*sysctl_sched_migration_cost;
2661 update_avg(&rq->avg_idle, delta);
2668 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2671 if (p->sched_contributes_to_load)
2672 rq->nr_uninterruptible--;
2675 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2676 ttwu_do_wakeup(rq, p, wake_flags);
2680 * Called in case the task @p isn't fully descheduled from its runqueue,
2681 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2682 * since all we need to do is flip p->state to TASK_RUNNING, since
2683 * the task is still ->on_rq.
2685 static int ttwu_remote(struct task_struct *p, int wake_flags)
2690 rq = __task_rq_lock(p);
2692 ttwu_do_wakeup(rq, p, wake_flags);
2695 __task_rq_unlock(rq);
2701 static void sched_ttwu_do_pending(struct task_struct *list)
2703 struct rq *rq = this_rq();
2705 raw_spin_lock(&rq->lock);
2708 struct task_struct *p = list;
2709 list = list->wake_entry;
2710 ttwu_do_activate(rq, p, 0);
2713 raw_spin_unlock(&rq->lock);
2716 #ifdef CONFIG_HOTPLUG_CPU
2718 static void sched_ttwu_pending(void)
2720 struct rq *rq = this_rq();
2721 struct task_struct *list = xchg(&rq->wake_list, NULL);
2726 sched_ttwu_do_pending(list);
2729 #endif /* CONFIG_HOTPLUG_CPU */
2731 void scheduler_ipi(void)
2733 struct rq *rq = this_rq();
2734 struct task_struct *list = xchg(&rq->wake_list, NULL);
2740 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2741 * traditionally all their work was done from the interrupt return
2742 * path. Now that we actually do some work, we need to make sure
2745 * Some archs already do call them, luckily irq_enter/exit nest
2748 * Arguably we should visit all archs and update all handlers,
2749 * however a fair share of IPIs are still resched only so this would
2750 * somewhat pessimize the simple resched case.
2753 sched_ttwu_do_pending(list);
2757 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2759 struct rq *rq = cpu_rq(cpu);
2760 struct task_struct *next = rq->wake_list;
2763 struct task_struct *old = next;
2765 p->wake_entry = next;
2766 next = cmpxchg(&rq->wake_list, old, p);
2772 smp_send_reschedule(cpu);
2775 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2776 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2781 rq = __task_rq_lock(p);
2783 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2784 ttwu_do_wakeup(rq, p, wake_flags);
2787 __task_rq_unlock(rq);
2792 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2793 #endif /* CONFIG_SMP */
2795 static void ttwu_queue(struct task_struct *p, int cpu)
2797 struct rq *rq = cpu_rq(cpu);
2799 #if defined(CONFIG_SMP)
2800 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2801 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2802 ttwu_queue_remote(p, cpu);
2807 raw_spin_lock(&rq->lock);
2808 ttwu_do_activate(rq, p, 0);
2809 raw_spin_unlock(&rq->lock);
2813 * try_to_wake_up - wake up a thread
2814 * @p: the thread to be awakened
2815 * @state: the mask of task states that can be woken
2816 * @wake_flags: wake modifier flags (WF_*)
2818 * Put it on the run-queue if it's not already there. The "current"
2819 * thread is always on the run-queue (except when the actual
2820 * re-schedule is in progress), and as such you're allowed to do
2821 * the simpler "current->state = TASK_RUNNING" to mark yourself
2822 * runnable without the overhead of this.
2824 * Returns %true if @p was woken up, %false if it was already running
2825 * or @state didn't match @p's state.
2828 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2830 unsigned long flags;
2831 int cpu, success = 0;
2834 raw_spin_lock_irqsave(&p->pi_lock, flags);
2835 if (!(p->state & state))
2838 success = 1; /* we're going to change ->state */
2841 if (p->on_rq && ttwu_remote(p, wake_flags))
2846 * If the owning (remote) cpu is still in the middle of schedule() with
2847 * this task as prev, wait until its done referencing the task.
2850 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2852 * In case the architecture enables interrupts in
2853 * context_switch(), we cannot busy wait, since that
2854 * would lead to deadlocks when an interrupt hits and
2855 * tries to wake up @prev. So bail and do a complete
2858 if (ttwu_activate_remote(p, wake_flags))
2865 * Pairs with the smp_wmb() in finish_lock_switch().
2869 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2870 p->state = TASK_WAKING;
2872 if (p->sched_class->task_waking)
2873 p->sched_class->task_waking(p);
2875 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2876 if (task_cpu(p) != cpu) {
2877 wake_flags |= WF_MIGRATED;
2878 set_task_cpu(p, cpu);
2880 #endif /* CONFIG_SMP */
2884 ttwu_stat(p, cpu, wake_flags);
2886 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2892 * try_to_wake_up_local - try to wake up a local task with rq lock held
2893 * @p: the thread to be awakened
2895 * Put @p on the run-queue if it's not already there. The caller must
2896 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2899 static void try_to_wake_up_local(struct task_struct *p)
2901 struct rq *rq = task_rq(p);
2903 BUG_ON(rq != this_rq());
2904 BUG_ON(p == current);
2905 lockdep_assert_held(&rq->lock);
2907 if (!raw_spin_trylock(&p->pi_lock)) {
2908 raw_spin_unlock(&rq->lock);
2909 raw_spin_lock(&p->pi_lock);
2910 raw_spin_lock(&rq->lock);
2913 if (!(p->state & TASK_NORMAL))
2917 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2919 ttwu_do_wakeup(rq, p, 0);
2920 ttwu_stat(p, smp_processor_id(), 0);
2922 raw_spin_unlock(&p->pi_lock);
2926 * wake_up_process - Wake up a specific process
2927 * @p: The process to be woken up.
2929 * Attempt to wake up the nominated process and move it to the set of runnable
2930 * processes. Returns 1 if the process was woken up, 0 if it was already
2933 * It may be assumed that this function implies a write memory barrier before
2934 * changing the task state if and only if any tasks are woken up.
2936 int wake_up_process(struct task_struct *p)
2938 return try_to_wake_up(p, TASK_ALL, 0);
2940 EXPORT_SYMBOL(wake_up_process);
2942 int wake_up_state(struct task_struct *p, unsigned int state)
2944 return try_to_wake_up(p, state, 0);
2948 * Perform scheduler related setup for a newly forked process p.
2949 * p is forked by current.
2951 * __sched_fork() is basic setup used by init_idle() too:
2953 static void __sched_fork(struct task_struct *p)
2958 p->se.exec_start = 0;
2959 p->se.sum_exec_runtime = 0;
2960 p->se.prev_sum_exec_runtime = 0;
2961 p->se.nr_migrations = 0;
2963 INIT_LIST_HEAD(&p->se.group_node);
2965 #ifdef CONFIG_SCHEDSTATS
2966 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2969 INIT_LIST_HEAD(&p->rt.run_list);
2971 #ifdef CONFIG_PREEMPT_NOTIFIERS
2972 INIT_HLIST_HEAD(&p->preempt_notifiers);
2977 * fork()/clone()-time setup:
2979 void sched_fork(struct task_struct *p)
2981 unsigned long flags;
2982 int cpu = get_cpu();
2986 * We mark the process as running here. This guarantees that
2987 * nobody will actually run it, and a signal or other external
2988 * event cannot wake it up and insert it on the runqueue either.
2990 p->state = TASK_RUNNING;
2993 * Make sure we do not leak PI boosting priority to the child.
2995 p->prio = current->normal_prio;
2998 * Revert to default priority/policy on fork if requested.
3000 if (unlikely(p->sched_reset_on_fork)) {
3001 if (task_has_rt_policy(p)) {
3002 p->policy = SCHED_NORMAL;
3003 p->static_prio = NICE_TO_PRIO(0);
3005 } else if (PRIO_TO_NICE(p->static_prio) < 0)
3006 p->static_prio = NICE_TO_PRIO(0);
3008 p->prio = p->normal_prio = __normal_prio(p);
3012 * We don't need the reset flag anymore after the fork. It has
3013 * fulfilled its duty:
3015 p->sched_reset_on_fork = 0;
3018 if (!rt_prio(p->prio))
3019 p->sched_class = &fair_sched_class;
3021 if (p->sched_class->task_fork)
3022 p->sched_class->task_fork(p);
3025 * The child is not yet in the pid-hash so no cgroup attach races,
3026 * and the cgroup is pinned to this child due to cgroup_fork()
3027 * is ran before sched_fork().
3029 * Silence PROVE_RCU.
3031 raw_spin_lock_irqsave(&p->pi_lock, flags);
3032 set_task_cpu(p, cpu);
3033 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3035 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
3036 if (likely(sched_info_on()))
3037 memset(&p->sched_info, 0, sizeof(p->sched_info));
3039 #if defined(CONFIG_SMP)
3042 #ifdef CONFIG_PREEMPT_COUNT
3043 /* Want to start with kernel preemption disabled. */
3044 task_thread_info(p)->preempt_count = 1;
3047 plist_node_init(&p->pushable_tasks, MAX_PRIO);
3054 * wake_up_new_task - wake up a newly created task for the first time.
3056 * This function will do some initial scheduler statistics housekeeping
3057 * that must be done for every newly created context, then puts the task
3058 * on the runqueue and wakes it.
3060 void wake_up_new_task(struct task_struct *p)
3062 unsigned long flags;
3065 raw_spin_lock_irqsave(&p->pi_lock, flags);
3068 * Fork balancing, do it here and not earlier because:
3069 * - cpus_allowed can change in the fork path
3070 * - any previously selected cpu might disappear through hotplug
3072 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
3075 rq = __task_rq_lock(p);
3076 activate_task(rq, p, 0);
3078 trace_sched_wakeup_new(p, true);
3079 check_preempt_curr(rq, p, WF_FORK);
3081 if (p->sched_class->task_woken)
3082 p->sched_class->task_woken(rq, p);
3084 task_rq_unlock(rq, p, &flags);
3087 #ifdef CONFIG_PREEMPT_NOTIFIERS
3090 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3091 * @notifier: notifier struct to register
3093 void preempt_notifier_register(struct preempt_notifier *notifier)
3095 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3097 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3100 * preempt_notifier_unregister - no longer interested in preemption notifications
3101 * @notifier: notifier struct to unregister
3103 * This is safe to call from within a preemption notifier.
3105 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3107 hlist_del(¬ifier->link);
3109 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3111 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3113 struct preempt_notifier *notifier;
3114 struct hlist_node *node;
3116 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3117 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3121 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3122 struct task_struct *next)
3124 struct preempt_notifier *notifier;
3125 struct hlist_node *node;
3127 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3128 notifier->ops->sched_out(notifier, next);
3131 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3133 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3138 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3139 struct task_struct *next)
3143 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3146 * prepare_task_switch - prepare to switch tasks
3147 * @rq: the runqueue preparing to switch
3148 * @prev: the current task that is being switched out
3149 * @next: the task we are going to switch to.
3151 * This is called with the rq lock held and interrupts off. It must
3152 * be paired with a subsequent finish_task_switch after the context
3155 * prepare_task_switch sets up locking and calls architecture specific
3159 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3160 struct task_struct *next)
3162 sched_info_switch(prev, next);
3163 perf_event_task_sched_out(prev, next);
3164 fire_sched_out_preempt_notifiers(prev, next);
3165 prepare_lock_switch(rq, next);
3166 prepare_arch_switch(next);
3167 trace_sched_switch(prev, next);
3171 * finish_task_switch - clean up after a task-switch
3172 * @rq: runqueue associated with task-switch
3173 * @prev: the thread we just switched away from.
3175 * finish_task_switch must be called after the context switch, paired
3176 * with a prepare_task_switch call before the context switch.
3177 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3178 * and do any other architecture-specific cleanup actions.
3180 * Note that we may have delayed dropping an mm in context_switch(). If
3181 * so, we finish that here outside of the runqueue lock. (Doing it
3182 * with the lock held can cause deadlocks; see schedule() for
3185 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3186 __releases(rq->lock)
3188 struct mm_struct *mm = rq->prev_mm;
3194 * A task struct has one reference for the use as "current".
3195 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3196 * schedule one last time. The schedule call will never return, and
3197 * the scheduled task must drop that reference.
3198 * The test for TASK_DEAD must occur while the runqueue locks are
3199 * still held, otherwise prev could be scheduled on another cpu, die
3200 * there before we look at prev->state, and then the reference would
3202 * Manfred Spraul <manfred@colorfullife.com>
3204 prev_state = prev->state;
3205 finish_arch_switch(prev);
3206 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3207 local_irq_disable();
3208 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3209 perf_event_task_sched_in(prev, current);
3210 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3212 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3213 finish_lock_switch(rq, prev);
3215 fire_sched_in_preempt_notifiers(current);
3218 if (unlikely(prev_state == TASK_DEAD)) {
3220 * Remove function-return probe instances associated with this
3221 * task and put them back on the free list.
3223 kprobe_flush_task(prev);
3224 put_task_struct(prev);
3230 /* assumes rq->lock is held */
3231 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3233 if (prev->sched_class->pre_schedule)
3234 prev->sched_class->pre_schedule(rq, prev);
3237 /* rq->lock is NOT held, but preemption is disabled */
3238 static inline void post_schedule(struct rq *rq)
3240 if (rq->post_schedule) {
3241 unsigned long flags;
3243 raw_spin_lock_irqsave(&rq->lock, flags);
3244 if (rq->curr->sched_class->post_schedule)
3245 rq->curr->sched_class->post_schedule(rq);
3246 raw_spin_unlock_irqrestore(&rq->lock, flags);
3248 rq->post_schedule = 0;
3254 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3258 static inline void post_schedule(struct rq *rq)
3265 * schedule_tail - first thing a freshly forked thread must call.
3266 * @prev: the thread we just switched away from.
3268 asmlinkage void schedule_tail(struct task_struct *prev)
3269 __releases(rq->lock)
3271 struct rq *rq = this_rq();
3273 finish_task_switch(rq, prev);
3276 * FIXME: do we need to worry about rq being invalidated by the
3281 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3282 /* In this case, finish_task_switch does not reenable preemption */
3285 if (current->set_child_tid)
3286 put_user(task_pid_vnr(current), current->set_child_tid);
3290 * context_switch - switch to the new MM and the new
3291 * thread's register state.
3294 context_switch(struct rq *rq, struct task_struct *prev,
3295 struct task_struct *next)
3297 struct mm_struct *mm, *oldmm;
3299 prepare_task_switch(rq, prev, next);
3302 oldmm = prev->active_mm;
3304 * For paravirt, this is coupled with an exit in switch_to to
3305 * combine the page table reload and the switch backend into
3308 arch_start_context_switch(prev);
3311 next->active_mm = oldmm;
3312 atomic_inc(&oldmm->mm_count);
3313 enter_lazy_tlb(oldmm, next);
3315 switch_mm(oldmm, mm, next);
3318 prev->active_mm = NULL;
3319 rq->prev_mm = oldmm;
3322 * Since the runqueue lock will be released by the next
3323 * task (which is an invalid locking op but in the case
3324 * of the scheduler it's an obvious special-case), so we
3325 * do an early lockdep release here:
3327 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3328 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3331 /* Here we just switch the register state and the stack. */
3332 switch_to(prev, next, prev);
3336 * this_rq must be evaluated again because prev may have moved
3337 * CPUs since it called schedule(), thus the 'rq' on its stack
3338 * frame will be invalid.
3340 finish_task_switch(this_rq(), prev);
3344 * nr_running, nr_uninterruptible and nr_context_switches:
3346 * externally visible scheduler statistics: current number of runnable
3347 * threads, current number of uninterruptible-sleeping threads, total
3348 * number of context switches performed since bootup.
3350 unsigned long nr_running(void)
3352 unsigned long i, sum = 0;
3354 for_each_online_cpu(i)
3355 sum += cpu_rq(i)->nr_running;
3360 unsigned long nr_uninterruptible(void)
3362 unsigned long i, sum = 0;
3364 for_each_possible_cpu(i)
3365 sum += cpu_rq(i)->nr_uninterruptible;
3368 * Since we read the counters lockless, it might be slightly
3369 * inaccurate. Do not allow it to go below zero though:
3371 if (unlikely((long)sum < 0))
3377 unsigned long long nr_context_switches(void)
3380 unsigned long long sum = 0;
3382 for_each_possible_cpu(i)
3383 sum += cpu_rq(i)->nr_switches;
3388 unsigned long nr_iowait(void)
3390 unsigned long i, sum = 0;
3392 for_each_possible_cpu(i)
3393 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3398 unsigned long nr_iowait_cpu(int cpu)
3400 struct rq *this = cpu_rq(cpu);
3401 return atomic_read(&this->nr_iowait);
3404 unsigned long this_cpu_load(void)
3406 struct rq *this = this_rq();
3407 return this->cpu_load[0];
3411 /* Variables and functions for calc_load */
3412 static atomic_long_t calc_load_tasks;
3413 static unsigned long calc_load_update;
3414 unsigned long avenrun[3];
3415 EXPORT_SYMBOL(avenrun);
3417 static long calc_load_fold_active(struct rq *this_rq)
3419 long nr_active, delta = 0;
3421 nr_active = this_rq->nr_running;
3422 nr_active += (long) this_rq->nr_uninterruptible;
3424 if (nr_active != this_rq->calc_load_active) {
3425 delta = nr_active - this_rq->calc_load_active;
3426 this_rq->calc_load_active = nr_active;
3432 static unsigned long
3433 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3436 load += active * (FIXED_1 - exp);
3437 load += 1UL << (FSHIFT - 1);
3438 return load >> FSHIFT;
3443 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3445 * When making the ILB scale, we should try to pull this in as well.
3447 static atomic_long_t calc_load_tasks_idle;
3449 static void calc_load_account_idle(struct rq *this_rq)
3453 delta = calc_load_fold_active(this_rq);
3455 atomic_long_add(delta, &calc_load_tasks_idle);
3458 static long calc_load_fold_idle(void)
3463 * Its got a race, we don't care...
3465 if (atomic_long_read(&calc_load_tasks_idle))
3466 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3472 * fixed_power_int - compute: x^n, in O(log n) time
3474 * @x: base of the power
3475 * @frac_bits: fractional bits of @x
3476 * @n: power to raise @x to.
3478 * By exploiting the relation between the definition of the natural power
3479 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3480 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3481 * (where: n_i \elem {0, 1}, the binary vector representing n),
3482 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3483 * of course trivially computable in O(log_2 n), the length of our binary
3486 static unsigned long
3487 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3489 unsigned long result = 1UL << frac_bits;
3494 result += 1UL << (frac_bits - 1);
3495 result >>= frac_bits;
3501 x += 1UL << (frac_bits - 1);
3509 * a1 = a0 * e + a * (1 - e)
3511 * a2 = a1 * e + a * (1 - e)
3512 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3513 * = a0 * e^2 + a * (1 - e) * (1 + e)
3515 * a3 = a2 * e + a * (1 - e)
3516 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3517 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3521 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3522 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3523 * = a0 * e^n + a * (1 - e^n)
3525 * [1] application of the geometric series:
3528 * S_n := \Sum x^i = -------------
3531 static unsigned long
3532 calc_load_n(unsigned long load, unsigned long exp,
3533 unsigned long active, unsigned int n)
3536 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3540 * NO_HZ can leave us missing all per-cpu ticks calling
3541 * calc_load_account_active(), but since an idle CPU folds its delta into
3542 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3543 * in the pending idle delta if our idle period crossed a load cycle boundary.
3545 * Once we've updated the global active value, we need to apply the exponential
3546 * weights adjusted to the number of cycles missed.
3548 static void calc_global_nohz(unsigned long ticks)
3550 long delta, active, n;
3552 if (time_before(jiffies, calc_load_update))
3556 * If we crossed a calc_load_update boundary, make sure to fold
3557 * any pending idle changes, the respective CPUs might have
3558 * missed the tick driven calc_load_account_active() update
3561 delta = calc_load_fold_idle();
3563 atomic_long_add(delta, &calc_load_tasks);
3566 * If we were idle for multiple load cycles, apply them.
3568 if (ticks >= LOAD_FREQ) {
3569 n = ticks / LOAD_FREQ;
3571 active = atomic_long_read(&calc_load_tasks);
3572 active = active > 0 ? active * FIXED_1 : 0;
3574 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3575 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3576 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3578 calc_load_update += n * LOAD_FREQ;
3582 * Its possible the remainder of the above division also crosses
3583 * a LOAD_FREQ period, the regular check in calc_global_load()
3584 * which comes after this will take care of that.
3586 * Consider us being 11 ticks before a cycle completion, and us
3587 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3588 * age us 4 cycles, and the test in calc_global_load() will
3589 * pick up the final one.
3593 static void calc_load_account_idle(struct rq *this_rq)
3597 static inline long calc_load_fold_idle(void)
3602 static void calc_global_nohz(unsigned long ticks)
3608 * get_avenrun - get the load average array
3609 * @loads: pointer to dest load array
3610 * @offset: offset to add
3611 * @shift: shift count to shift the result left
3613 * These values are estimates at best, so no need for locking.
3615 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3617 loads[0] = (avenrun[0] + offset) << shift;
3618 loads[1] = (avenrun[1] + offset) << shift;
3619 loads[2] = (avenrun[2] + offset) << shift;
3623 * calc_load - update the avenrun load estimates 10 ticks after the
3624 * CPUs have updated calc_load_tasks.
3626 void calc_global_load(unsigned long ticks)
3630 calc_global_nohz(ticks);
3632 if (time_before(jiffies, calc_load_update + 10))
3635 active = atomic_long_read(&calc_load_tasks);
3636 active = active > 0 ? active * FIXED_1 : 0;
3638 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3639 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3640 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3642 calc_load_update += LOAD_FREQ;
3646 * Called from update_cpu_load() to periodically update this CPU's
3649 static void calc_load_account_active(struct rq *this_rq)
3653 if (time_before(jiffies, this_rq->calc_load_update))
3656 delta = calc_load_fold_active(this_rq);
3657 delta += calc_load_fold_idle();
3659 atomic_long_add(delta, &calc_load_tasks);
3661 this_rq->calc_load_update += LOAD_FREQ;
3665 * The exact cpuload at various idx values, calculated at every tick would be
3666 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3668 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3669 * on nth tick when cpu may be busy, then we have:
3670 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3671 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3673 * decay_load_missed() below does efficient calculation of
3674 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3675 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3677 * The calculation is approximated on a 128 point scale.
3678 * degrade_zero_ticks is the number of ticks after which load at any
3679 * particular idx is approximated to be zero.
3680 * degrade_factor is a precomputed table, a row for each load idx.
3681 * Each column corresponds to degradation factor for a power of two ticks,
3682 * based on 128 point scale.
3684 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3685 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3687 * With this power of 2 load factors, we can degrade the load n times
3688 * by looking at 1 bits in n and doing as many mult/shift instead of
3689 * n mult/shifts needed by the exact degradation.
3691 #define DEGRADE_SHIFT 7
3692 static const unsigned char
3693 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3694 static const unsigned char
3695 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3696 {0, 0, 0, 0, 0, 0, 0, 0},
3697 {64, 32, 8, 0, 0, 0, 0, 0},
3698 {96, 72, 40, 12, 1, 0, 0},
3699 {112, 98, 75, 43, 15, 1, 0},
3700 {120, 112, 98, 76, 45, 16, 2} };
3703 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3704 * would be when CPU is idle and so we just decay the old load without
3705 * adding any new load.
3707 static unsigned long
3708 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3712 if (!missed_updates)
3715 if (missed_updates >= degrade_zero_ticks[idx])
3719 return load >> missed_updates;
3721 while (missed_updates) {
3722 if (missed_updates % 2)
3723 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3725 missed_updates >>= 1;
3732 * Update rq->cpu_load[] statistics. This function is usually called every
3733 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3734 * every tick. We fix it up based on jiffies.
3736 static void update_cpu_load(struct rq *this_rq)
3738 unsigned long this_load = this_rq->load.weight;
3739 unsigned long curr_jiffies = jiffies;
3740 unsigned long pending_updates;
3743 this_rq->nr_load_updates++;
3745 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3746 if (curr_jiffies == this_rq->last_load_update_tick)
3749 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3750 this_rq->last_load_update_tick = curr_jiffies;
3752 /* Update our load: */
3753 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3754 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3755 unsigned long old_load, new_load;
3757 /* scale is effectively 1 << i now, and >> i divides by scale */
3759 old_load = this_rq->cpu_load[i];
3760 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3761 new_load = this_load;
3763 * Round up the averaging division if load is increasing. This
3764 * prevents us from getting stuck on 9 if the load is 10, for
3767 if (new_load > old_load)
3768 new_load += scale - 1;
3770 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3773 sched_avg_update(this_rq);
3776 static void update_cpu_load_active(struct rq *this_rq)
3778 update_cpu_load(this_rq);
3780 calc_load_account_active(this_rq);
3786 * sched_exec - execve() is a valuable balancing opportunity, because at
3787 * this point the task has the smallest effective memory and cache footprint.
3789 void sched_exec(void)
3791 struct task_struct *p = current;
3792 unsigned long flags;
3795 raw_spin_lock_irqsave(&p->pi_lock, flags);
3796 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3797 if (dest_cpu == smp_processor_id())
3800 if (likely(cpu_active(dest_cpu))) {
3801 struct migration_arg arg = { p, dest_cpu };
3803 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3804 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3808 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3813 DEFINE_PER_CPU(struct kernel_stat, kstat);
3815 EXPORT_PER_CPU_SYMBOL(kstat);
3818 * Return any ns on the sched_clock that have not yet been accounted in
3819 * @p in case that task is currently running.
3821 * Called with task_rq_lock() held on @rq.
3823 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3827 if (task_current(rq, p)) {
3828 update_rq_clock(rq);
3829 ns = rq->clock_task - p->se.exec_start;
3837 unsigned long long task_delta_exec(struct task_struct *p)
3839 unsigned long flags;
3843 rq = task_rq_lock(p, &flags);
3844 ns = do_task_delta_exec(p, rq);
3845 task_rq_unlock(rq, p, &flags);
3851 * Return accounted runtime for the task.
3852 * In case the task is currently running, return the runtime plus current's
3853 * pending runtime that have not been accounted yet.
3855 unsigned long long task_sched_runtime(struct task_struct *p)
3857 unsigned long flags;
3861 rq = task_rq_lock(p, &flags);
3862 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3863 task_rq_unlock(rq, p, &flags);
3869 * Account user cpu time to a process.
3870 * @p: the process that the cpu time gets accounted to
3871 * @cputime: the cpu time spent in user space since the last update
3872 * @cputime_scaled: cputime scaled by cpu frequency
3874 void account_user_time(struct task_struct *p, cputime_t cputime,
3875 cputime_t cputime_scaled)
3877 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3880 /* Add user time to process. */
3881 p->utime = cputime_add(p->utime, cputime);
3882 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3883 account_group_user_time(p, cputime);
3885 /* Add user time to cpustat. */
3886 tmp = cputime_to_cputime64(cputime);
3887 if (TASK_NICE(p) > 0)
3888 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3890 cpustat->user = cputime64_add(cpustat->user, tmp);
3892 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3893 /* Account for user time used */
3894 acct_update_integrals(p);
3898 * Account guest cpu time to a process.
3899 * @p: the process that the cpu time gets accounted to
3900 * @cputime: the cpu time spent in virtual machine since the last update
3901 * @cputime_scaled: cputime scaled by cpu frequency
3903 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3904 cputime_t cputime_scaled)
3907 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3909 tmp = cputime_to_cputime64(cputime);
3911 /* Add guest time to process. */
3912 p->utime = cputime_add(p->utime, cputime);
3913 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3914 account_group_user_time(p, cputime);
3915 p->gtime = cputime_add(p->gtime, cputime);
3917 /* Add guest time to cpustat. */
3918 if (TASK_NICE(p) > 0) {
3919 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3920 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3922 cpustat->user = cputime64_add(cpustat->user, tmp);
3923 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3928 * Account system cpu time to a process and desired cpustat field
3929 * @p: the process that the cpu time gets accounted to
3930 * @cputime: the cpu time spent in kernel space since the last update
3931 * @cputime_scaled: cputime scaled by cpu frequency
3932 * @target_cputime64: pointer to cpustat field that has to be updated
3935 void __account_system_time(struct task_struct *p, cputime_t cputime,
3936 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3938 cputime64_t tmp = cputime_to_cputime64(cputime);
3940 /* Add system time to process. */
3941 p->stime = cputime_add(p->stime, cputime);
3942 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3943 account_group_system_time(p, cputime);
3945 /* Add system time to cpustat. */
3946 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3947 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3949 /* Account for system time used */
3950 acct_update_integrals(p);
3954 * Account system cpu time to a process.
3955 * @p: the process that the cpu time gets accounted to
3956 * @hardirq_offset: the offset to subtract from hardirq_count()
3957 * @cputime: the cpu time spent in kernel space since the last update
3958 * @cputime_scaled: cputime scaled by cpu frequency
3960 void account_system_time(struct task_struct *p, int hardirq_offset,
3961 cputime_t cputime, cputime_t cputime_scaled)
3963 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3964 cputime64_t *target_cputime64;
3966 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3967 account_guest_time(p, cputime, cputime_scaled);
3971 if (hardirq_count() - hardirq_offset)
3972 target_cputime64 = &cpustat->irq;
3973 else if (in_serving_softirq())
3974 target_cputime64 = &cpustat->softirq;
3976 target_cputime64 = &cpustat->system;
3978 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3982 * Account for involuntary wait time.
3983 * @cputime: the cpu time spent in involuntary wait
3985 void account_steal_time(cputime_t cputime)
3987 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3988 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3990 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3994 * Account for idle time.
3995 * @cputime: the cpu time spent in idle wait
3997 void account_idle_time(cputime_t cputime)
3999 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4000 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4001 struct rq *rq = this_rq();
4003 if (atomic_read(&rq->nr_iowait) > 0)
4004 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4006 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4009 static __always_inline bool steal_account_process_tick(void)
4011 #ifdef CONFIG_PARAVIRT
4012 if (static_branch(¶virt_steal_enabled)) {
4015 steal = paravirt_steal_clock(smp_processor_id());
4016 steal -= this_rq()->prev_steal_time;
4018 st = steal_ticks(steal);
4019 this_rq()->prev_steal_time += st * TICK_NSEC;
4021 account_steal_time(st);
4028 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4030 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
4032 * Account a tick to a process and cpustat
4033 * @p: the process that the cpu time gets accounted to
4034 * @user_tick: is the tick from userspace
4035 * @rq: the pointer to rq
4037 * Tick demultiplexing follows the order
4038 * - pending hardirq update
4039 * - pending softirq update
4043 * - check for guest_time
4044 * - else account as system_time
4046 * Check for hardirq is done both for system and user time as there is
4047 * no timer going off while we are on hardirq and hence we may never get an
4048 * opportunity to update it solely in system time.
4049 * p->stime and friends are only updated on system time and not on irq
4050 * softirq as those do not count in task exec_runtime any more.
4052 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4055 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4056 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
4057 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4059 if (steal_account_process_tick())
4062 if (irqtime_account_hi_update()) {
4063 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4064 } else if (irqtime_account_si_update()) {
4065 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4066 } else if (this_cpu_ksoftirqd() == p) {
4068 * ksoftirqd time do not get accounted in cpu_softirq_time.
4069 * So, we have to handle it separately here.
4070 * Also, p->stime needs to be updated for ksoftirqd.
4072 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4074 } else if (user_tick) {
4075 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4076 } else if (p == rq->idle) {
4077 account_idle_time(cputime_one_jiffy);
4078 } else if (p->flags & PF_VCPU) { /* System time or guest time */
4079 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
4081 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4086 static void irqtime_account_idle_ticks(int ticks)
4089 struct rq *rq = this_rq();
4091 for (i = 0; i < ticks; i++)
4092 irqtime_account_process_tick(current, 0, rq);
4094 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
4095 static void irqtime_account_idle_ticks(int ticks) {}
4096 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4098 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
4101 * Account a single tick of cpu time.
4102 * @p: the process that the cpu time gets accounted to
4103 * @user_tick: indicates if the tick is a user or a system tick
4105 void account_process_tick(struct task_struct *p, int user_tick)
4107 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4108 struct rq *rq = this_rq();
4110 if (sched_clock_irqtime) {
4111 irqtime_account_process_tick(p, user_tick, rq);
4115 if (steal_account_process_tick())
4119 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4120 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
4121 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
4124 account_idle_time(cputime_one_jiffy);
4128 * Account multiple ticks of steal time.
4129 * @p: the process from which the cpu time has been stolen
4130 * @ticks: number of stolen ticks
4132 void account_steal_ticks(unsigned long ticks)
4134 account_steal_time(jiffies_to_cputime(ticks));
4138 * Account multiple ticks of idle time.
4139 * @ticks: number of stolen ticks
4141 void account_idle_ticks(unsigned long ticks)
4144 if (sched_clock_irqtime) {
4145 irqtime_account_idle_ticks(ticks);
4149 account_idle_time(jiffies_to_cputime(ticks));
4155 * Use precise platform statistics if available:
4157 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4158 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4164 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4166 struct task_cputime cputime;
4168 thread_group_cputime(p, &cputime);
4170 *ut = cputime.utime;
4171 *st = cputime.stime;
4175 #ifndef nsecs_to_cputime
4176 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
4179 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4181 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
4184 * Use CFS's precise accounting:
4186 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
4192 do_div(temp, total);
4193 utime = (cputime_t)temp;
4198 * Compare with previous values, to keep monotonicity:
4200 p->prev_utime = max(p->prev_utime, utime);
4201 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4203 *ut = p->prev_utime;
4204 *st = p->prev_stime;
4208 * Must be called with siglock held.
4210 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4212 struct signal_struct *sig = p->signal;
4213 struct task_cputime cputime;
4214 cputime_t rtime, utime, total;
4216 thread_group_cputime(p, &cputime);
4218 total = cputime_add(cputime.utime, cputime.stime);
4219 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4224 temp *= cputime.utime;
4225 do_div(temp, total);
4226 utime = (cputime_t)temp;
4230 sig->prev_utime = max(sig->prev_utime, utime);
4231 sig->prev_stime = max(sig->prev_stime,
4232 cputime_sub(rtime, sig->prev_utime));
4234 *ut = sig->prev_utime;
4235 *st = sig->prev_stime;
4240 * This function gets called by the timer code, with HZ frequency.
4241 * We call it with interrupts disabled.
4243 void scheduler_tick(void)
4245 int cpu = smp_processor_id();
4246 struct rq *rq = cpu_rq(cpu);
4247 struct task_struct *curr = rq->curr;
4251 raw_spin_lock(&rq->lock);
4252 update_rq_clock(rq);
4253 update_cpu_load_active(rq);
4254 curr->sched_class->task_tick(rq, curr, 0);
4255 raw_spin_unlock(&rq->lock);
4257 perf_event_task_tick();
4260 rq->idle_at_tick = idle_cpu(cpu);
4261 trigger_load_balance(rq, cpu);
4265 notrace unsigned long get_parent_ip(unsigned long addr)
4267 if (in_lock_functions(addr)) {
4268 addr = CALLER_ADDR2;
4269 if (in_lock_functions(addr))
4270 addr = CALLER_ADDR3;
4275 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4276 defined(CONFIG_PREEMPT_TRACER))
4278 void __kprobes add_preempt_count(int val)
4280 #ifdef CONFIG_DEBUG_PREEMPT
4284 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4287 preempt_count() += val;
4288 #ifdef CONFIG_DEBUG_PREEMPT
4290 * Spinlock count overflowing soon?
4292 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4295 if (preempt_count() == val)
4296 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4298 EXPORT_SYMBOL(add_preempt_count);
4300 void __kprobes sub_preempt_count(int val)
4302 #ifdef CONFIG_DEBUG_PREEMPT
4306 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4309 * Is the spinlock portion underflowing?
4311 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4312 !(preempt_count() & PREEMPT_MASK)))
4316 if (preempt_count() == val)
4317 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4318 preempt_count() -= val;
4320 EXPORT_SYMBOL(sub_preempt_count);
4325 * Print scheduling while atomic bug:
4327 static noinline void __schedule_bug(struct task_struct *prev)
4329 struct pt_regs *regs = get_irq_regs();
4331 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4332 prev->comm, prev->pid, preempt_count());
4334 debug_show_held_locks(prev);
4336 if (irqs_disabled())
4337 print_irqtrace_events(prev);
4346 * Various schedule()-time debugging checks and statistics:
4348 static inline void schedule_debug(struct task_struct *prev)
4351 * Test if we are atomic. Since do_exit() needs to call into
4352 * schedule() atomically, we ignore that path for now.
4353 * Otherwise, whine if we are scheduling when we should not be.
4355 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4356 __schedule_bug(prev);
4358 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4360 schedstat_inc(this_rq(), sched_count);
4363 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4365 if (prev->on_rq || rq->skip_clock_update < 0)
4366 update_rq_clock(rq);
4367 prev->sched_class->put_prev_task(rq, prev);
4371 * Pick up the highest-prio task:
4373 static inline struct task_struct *
4374 pick_next_task(struct rq *rq)
4376 const struct sched_class *class;
4377 struct task_struct *p;
4380 * Optimization: we know that if all tasks are in
4381 * the fair class we can call that function directly:
4383 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
4384 p = fair_sched_class.pick_next_task(rq);
4389 for_each_class(class) {
4390 p = class->pick_next_task(rq);
4395 BUG(); /* the idle class will always have a runnable task */
4399 * __schedule() is the main scheduler function.
4401 static void __sched __schedule(void)
4403 struct task_struct *prev, *next;
4404 unsigned long *switch_count;
4410 cpu = smp_processor_id();
4412 rcu_note_context_switch(cpu);
4415 schedule_debug(prev);
4417 if (sched_feat(HRTICK))
4420 raw_spin_lock_irq(&rq->lock);
4422 switch_count = &prev->nivcsw;
4423 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4424 if (unlikely(signal_pending_state(prev->state, prev))) {
4425 prev->state = TASK_RUNNING;
4427 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4431 * If a worker went to sleep, notify and ask workqueue
4432 * whether it wants to wake up a task to maintain
4435 if (prev->flags & PF_WQ_WORKER) {
4436 struct task_struct *to_wakeup;
4438 to_wakeup = wq_worker_sleeping(prev, cpu);
4440 try_to_wake_up_local(to_wakeup);
4443 switch_count = &prev->nvcsw;
4446 pre_schedule(rq, prev);
4448 if (unlikely(!rq->nr_running))
4449 idle_balance(cpu, rq);
4451 put_prev_task(rq, prev);
4452 next = pick_next_task(rq);
4453 clear_tsk_need_resched(prev);
4454 rq->skip_clock_update = 0;
4456 if (likely(prev != next)) {
4461 context_switch(rq, prev, next); /* unlocks the rq */
4463 * The context switch have flipped the stack from under us
4464 * and restored the local variables which were saved when
4465 * this task called schedule() in the past. prev == current
4466 * is still correct, but it can be moved to another cpu/rq.
4468 cpu = smp_processor_id();
4471 raw_spin_unlock_irq(&rq->lock);
4475 preempt_enable_no_resched();
4480 static inline void sched_submit_work(struct task_struct *tsk)
4485 * If we are going to sleep and we have plugged IO queued,
4486 * make sure to submit it to avoid deadlocks.
4488 if (blk_needs_flush_plug(tsk))
4489 blk_schedule_flush_plug(tsk);
4492 asmlinkage void __sched schedule(void)
4494 struct task_struct *tsk = current;
4496 sched_submit_work(tsk);
4499 EXPORT_SYMBOL(schedule);
4501 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4503 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4505 if (lock->owner != owner)
4509 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4510 * lock->owner still matches owner, if that fails, owner might
4511 * point to free()d memory, if it still matches, the rcu_read_lock()
4512 * ensures the memory stays valid.
4516 return owner->on_cpu;
4520 * Look out! "owner" is an entirely speculative pointer
4521 * access and not reliable.
4523 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4525 if (!sched_feat(OWNER_SPIN))
4529 while (owner_running(lock, owner)) {
4533 arch_mutex_cpu_relax();
4538 * We break out the loop above on need_resched() and when the
4539 * owner changed, which is a sign for heavy contention. Return
4540 * success only when lock->owner is NULL.
4542 return lock->owner == NULL;
4546 #ifdef CONFIG_PREEMPT
4548 * this is the entry point to schedule() from in-kernel preemption
4549 * off of preempt_enable. Kernel preemptions off return from interrupt
4550 * occur there and call schedule directly.
4552 asmlinkage void __sched notrace preempt_schedule(void)
4554 struct thread_info *ti = current_thread_info();
4557 * If there is a non-zero preempt_count or interrupts are disabled,
4558 * we do not want to preempt the current task. Just return..
4560 if (likely(ti->preempt_count || irqs_disabled()))
4564 add_preempt_count_notrace(PREEMPT_ACTIVE);
4566 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4569 * Check again in case we missed a preemption opportunity
4570 * between schedule and now.
4573 } while (need_resched());
4575 EXPORT_SYMBOL(preempt_schedule);
4578 * this is the entry point to schedule() from kernel preemption
4579 * off of irq context.
4580 * Note, that this is called and return with irqs disabled. This will
4581 * protect us against recursive calling from irq.
4583 asmlinkage void __sched preempt_schedule_irq(void)
4585 struct thread_info *ti = current_thread_info();
4587 /* Catch callers which need to be fixed */
4588 BUG_ON(ti->preempt_count || !irqs_disabled());
4591 add_preempt_count(PREEMPT_ACTIVE);
4594 local_irq_disable();
4595 sub_preempt_count(PREEMPT_ACTIVE);
4598 * Check again in case we missed a preemption opportunity
4599 * between schedule and now.
4602 } while (need_resched());
4605 #endif /* CONFIG_PREEMPT */
4607 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4610 return try_to_wake_up(curr->private, mode, wake_flags);
4612 EXPORT_SYMBOL(default_wake_function);
4615 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4616 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4617 * number) then we wake all the non-exclusive tasks and one exclusive task.
4619 * There are circumstances in which we can try to wake a task which has already
4620 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4621 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4623 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4624 int nr_exclusive, int wake_flags, void *key)
4626 wait_queue_t *curr, *next;
4628 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4629 unsigned flags = curr->flags;
4631 if (curr->func(curr, mode, wake_flags, key) &&
4632 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4638 * __wake_up - wake up threads blocked on a waitqueue.
4640 * @mode: which threads
4641 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4642 * @key: is directly passed to the wakeup function
4644 * It may be assumed that this function implies a write memory barrier before
4645 * changing the task state if and only if any tasks are woken up.
4647 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4648 int nr_exclusive, void *key)
4650 unsigned long flags;
4652 spin_lock_irqsave(&q->lock, flags);
4653 __wake_up_common(q, mode, nr_exclusive, 0, key);
4654 spin_unlock_irqrestore(&q->lock, flags);
4656 EXPORT_SYMBOL(__wake_up);
4659 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4661 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4663 __wake_up_common(q, mode, 1, 0, NULL);
4665 EXPORT_SYMBOL_GPL(__wake_up_locked);
4667 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4669 __wake_up_common(q, mode, 1, 0, key);
4671 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4674 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4676 * @mode: which threads
4677 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4678 * @key: opaque value to be passed to wakeup targets
4680 * The sync wakeup differs that the waker knows that it will schedule
4681 * away soon, so while the target thread will be woken up, it will not
4682 * be migrated to another CPU - ie. the two threads are 'synchronized'
4683 * with each other. This can prevent needless bouncing between CPUs.
4685 * On UP it can prevent extra preemption.
4687 * It may be assumed that this function implies a write memory barrier before
4688 * changing the task state if and only if any tasks are woken up.
4690 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4691 int nr_exclusive, void *key)
4693 unsigned long flags;
4694 int wake_flags = WF_SYNC;
4699 if (unlikely(!nr_exclusive))
4702 spin_lock_irqsave(&q->lock, flags);
4703 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4704 spin_unlock_irqrestore(&q->lock, flags);
4706 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4709 * __wake_up_sync - see __wake_up_sync_key()
4711 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4713 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4715 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4718 * complete: - signals a single thread waiting on this completion
4719 * @x: holds the state of this particular completion
4721 * This will wake up a single thread waiting on this completion. Threads will be
4722 * awakened in the same order in which they were queued.
4724 * See also complete_all(), wait_for_completion() and related routines.
4726 * It may be assumed that this function implies a write memory barrier before
4727 * changing the task state if and only if any tasks are woken up.
4729 void complete(struct completion *x)
4731 unsigned long flags;
4733 spin_lock_irqsave(&x->wait.lock, flags);
4735 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4736 spin_unlock_irqrestore(&x->wait.lock, flags);
4738 EXPORT_SYMBOL(complete);
4741 * complete_all: - signals all threads waiting on this completion
4742 * @x: holds the state of this particular completion
4744 * This will wake up all threads waiting on this particular completion event.
4746 * It may be assumed that this function implies a write memory barrier before
4747 * changing the task state if and only if any tasks are woken up.
4749 void complete_all(struct completion *x)
4751 unsigned long flags;
4753 spin_lock_irqsave(&x->wait.lock, flags);
4754 x->done += UINT_MAX/2;
4755 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4756 spin_unlock_irqrestore(&x->wait.lock, flags);
4758 EXPORT_SYMBOL(complete_all);
4760 static inline long __sched
4761 do_wait_for_common(struct completion *x, long timeout, int state)
4764 DECLARE_WAITQUEUE(wait, current);
4766 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4768 if (signal_pending_state(state, current)) {
4769 timeout = -ERESTARTSYS;
4772 __set_current_state(state);
4773 spin_unlock_irq(&x->wait.lock);
4774 timeout = schedule_timeout(timeout);
4775 spin_lock_irq(&x->wait.lock);
4776 } while (!x->done && timeout);
4777 __remove_wait_queue(&x->wait, &wait);
4782 return timeout ?: 1;
4786 wait_for_common(struct completion *x, long timeout, int state)
4790 spin_lock_irq(&x->wait.lock);
4791 timeout = do_wait_for_common(x, timeout, state);
4792 spin_unlock_irq(&x->wait.lock);
4797 * wait_for_completion: - waits for completion of a task
4798 * @x: holds the state of this particular completion
4800 * This waits to be signaled for completion of a specific task. It is NOT
4801 * interruptible and there is no timeout.
4803 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4804 * and interrupt capability. Also see complete().
4806 void __sched wait_for_completion(struct completion *x)
4808 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4810 EXPORT_SYMBOL(wait_for_completion);
4813 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4814 * @x: holds the state of this particular completion
4815 * @timeout: timeout value in jiffies
4817 * This waits for either a completion of a specific task to be signaled or for a
4818 * specified timeout to expire. The timeout is in jiffies. It is not
4821 unsigned long __sched
4822 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4824 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4826 EXPORT_SYMBOL(wait_for_completion_timeout);
4829 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4830 * @x: holds the state of this particular completion
4832 * This waits for completion of a specific task to be signaled. It is
4835 int __sched wait_for_completion_interruptible(struct completion *x)
4837 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4838 if (t == -ERESTARTSYS)
4842 EXPORT_SYMBOL(wait_for_completion_interruptible);
4845 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4846 * @x: holds the state of this particular completion
4847 * @timeout: timeout value in jiffies
4849 * This waits for either a completion of a specific task to be signaled or for a
4850 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4853 wait_for_completion_interruptible_timeout(struct completion *x,
4854 unsigned long timeout)
4856 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4858 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4861 * wait_for_completion_killable: - waits for completion of a task (killable)
4862 * @x: holds the state of this particular completion
4864 * This waits to be signaled for completion of a specific task. It can be
4865 * interrupted by a kill signal.
4867 int __sched wait_for_completion_killable(struct completion *x)
4869 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4870 if (t == -ERESTARTSYS)
4874 EXPORT_SYMBOL(wait_for_completion_killable);
4877 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4878 * @x: holds the state of this particular completion
4879 * @timeout: timeout value in jiffies
4881 * This waits for either a completion of a specific task to be
4882 * signaled or for a specified timeout to expire. It can be
4883 * interrupted by a kill signal. The timeout is in jiffies.
4886 wait_for_completion_killable_timeout(struct completion *x,
4887 unsigned long timeout)
4889 return wait_for_common(x, timeout, TASK_KILLABLE);
4891 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4894 * try_wait_for_completion - try to decrement a completion without blocking
4895 * @x: completion structure
4897 * Returns: 0 if a decrement cannot be done without blocking
4898 * 1 if a decrement succeeded.
4900 * If a completion is being used as a counting completion,
4901 * attempt to decrement the counter without blocking. This
4902 * enables us to avoid waiting if the resource the completion
4903 * is protecting is not available.
4905 bool try_wait_for_completion(struct completion *x)
4907 unsigned long flags;
4910 spin_lock_irqsave(&x->wait.lock, flags);
4915 spin_unlock_irqrestore(&x->wait.lock, flags);
4918 EXPORT_SYMBOL(try_wait_for_completion);
4921 * completion_done - Test to see if a completion has any waiters
4922 * @x: completion structure
4924 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4925 * 1 if there are no waiters.
4928 bool completion_done(struct completion *x)
4930 unsigned long flags;
4933 spin_lock_irqsave(&x->wait.lock, flags);
4936 spin_unlock_irqrestore(&x->wait.lock, flags);
4939 EXPORT_SYMBOL(completion_done);
4942 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4944 unsigned long flags;
4947 init_waitqueue_entry(&wait, current);
4949 __set_current_state(state);
4951 spin_lock_irqsave(&q->lock, flags);
4952 __add_wait_queue(q, &wait);
4953 spin_unlock(&q->lock);
4954 timeout = schedule_timeout(timeout);
4955 spin_lock_irq(&q->lock);
4956 __remove_wait_queue(q, &wait);
4957 spin_unlock_irqrestore(&q->lock, flags);
4962 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4964 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4966 EXPORT_SYMBOL(interruptible_sleep_on);
4969 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4971 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4973 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4975 void __sched sleep_on(wait_queue_head_t *q)
4977 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4979 EXPORT_SYMBOL(sleep_on);
4981 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4983 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4985 EXPORT_SYMBOL(sleep_on_timeout);
4987 #ifdef CONFIG_RT_MUTEXES
4990 * rt_mutex_setprio - set the current priority of a task
4992 * @prio: prio value (kernel-internal form)
4994 * This function changes the 'effective' priority of a task. It does
4995 * not touch ->normal_prio like __setscheduler().
4997 * Used by the rt_mutex code to implement priority inheritance logic.
4999 void rt_mutex_setprio(struct task_struct *p, int prio)
5001 int oldprio, on_rq, running;
5003 const struct sched_class *prev_class;
5005 BUG_ON(prio < 0 || prio > MAX_PRIO);
5007 rq = __task_rq_lock(p);
5009 trace_sched_pi_setprio(p, prio);
5011 prev_class = p->sched_class;
5013 running = task_current(rq, p);
5015 dequeue_task(rq, p, 0);
5017 p->sched_class->put_prev_task(rq, p);
5020 p->sched_class = &rt_sched_class;
5022 p->sched_class = &fair_sched_class;
5027 p->sched_class->set_curr_task(rq);
5029 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
5031 check_class_changed(rq, p, prev_class, oldprio);
5032 __task_rq_unlock(rq);
5037 void set_user_nice(struct task_struct *p, long nice)
5039 int old_prio, delta, on_rq;
5040 unsigned long flags;
5043 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5046 * We have to be careful, if called from sys_setpriority(),
5047 * the task might be in the middle of scheduling on another CPU.
5049 rq = task_rq_lock(p, &flags);
5051 * The RT priorities are set via sched_setscheduler(), but we still
5052 * allow the 'normal' nice value to be set - but as expected
5053 * it wont have any effect on scheduling until the task is
5054 * SCHED_FIFO/SCHED_RR:
5056 if (task_has_rt_policy(p)) {
5057 p->static_prio = NICE_TO_PRIO(nice);
5062 dequeue_task(rq, p, 0);
5064 p->static_prio = NICE_TO_PRIO(nice);
5067 p->prio = effective_prio(p);
5068 delta = p->prio - old_prio;
5071 enqueue_task(rq, p, 0);
5073 * If the task increased its priority or is running and
5074 * lowered its priority, then reschedule its CPU:
5076 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5077 resched_task(rq->curr);
5080 task_rq_unlock(rq, p, &flags);
5082 EXPORT_SYMBOL(set_user_nice);
5085 * can_nice - check if a task can reduce its nice value
5089 int can_nice(const struct task_struct *p, const int nice)
5091 /* convert nice value [19,-20] to rlimit style value [1,40] */
5092 int nice_rlim = 20 - nice;
5094 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
5095 capable(CAP_SYS_NICE));
5098 #ifdef __ARCH_WANT_SYS_NICE
5101 * sys_nice - change the priority of the current process.
5102 * @increment: priority increment
5104 * sys_setpriority is a more generic, but much slower function that
5105 * does similar things.
5107 SYSCALL_DEFINE1(nice, int, increment)
5112 * Setpriority might change our priority at the same moment.
5113 * We don't have to worry. Conceptually one call occurs first
5114 * and we have a single winner.
5116 if (increment < -40)
5121 nice = TASK_NICE(current) + increment;
5127 if (increment < 0 && !can_nice(current, nice))
5130 retval = security_task_setnice(current, nice);
5134 set_user_nice(current, nice);
5141 * task_prio - return the priority value of a given task.
5142 * @p: the task in question.
5144 * This is the priority value as seen by users in /proc.
5145 * RT tasks are offset by -200. Normal tasks are centered
5146 * around 0, value goes from -16 to +15.
5148 int task_prio(const struct task_struct *p)
5150 return p->prio - MAX_RT_PRIO;
5154 * task_nice - return the nice value of a given task.
5155 * @p: the task in question.
5157 int task_nice(const struct task_struct *p)
5159 return TASK_NICE(p);
5161 EXPORT_SYMBOL(task_nice);
5164 * idle_cpu - is a given cpu idle currently?
5165 * @cpu: the processor in question.
5167 int idle_cpu(int cpu)
5169 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5173 * idle_task - return the idle task for a given cpu.
5174 * @cpu: the processor in question.
5176 struct task_struct *idle_task(int cpu)
5178 return cpu_rq(cpu)->idle;
5182 * find_process_by_pid - find a process with a matching PID value.
5183 * @pid: the pid in question.
5185 static struct task_struct *find_process_by_pid(pid_t pid)
5187 return pid ? find_task_by_vpid(pid) : current;
5190 /* Actually do priority change: must hold rq lock. */
5192 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5195 p->rt_priority = prio;
5196 p->normal_prio = normal_prio(p);
5197 /* we are holding p->pi_lock already */
5198 p->prio = rt_mutex_getprio(p);
5199 if (rt_prio(p->prio))
5200 p->sched_class = &rt_sched_class;
5202 p->sched_class = &fair_sched_class;
5207 * check the target process has a UID that matches the current process's
5209 static bool check_same_owner(struct task_struct *p)
5211 const struct cred *cred = current_cred(), *pcred;
5215 pcred = __task_cred(p);
5216 if (cred->user->user_ns == pcred->user->user_ns)
5217 match = (cred->euid == pcred->euid ||
5218 cred->euid == pcred->uid);
5225 static int __sched_setscheduler(struct task_struct *p, int policy,
5226 const struct sched_param *param, bool user)
5228 int retval, oldprio, oldpolicy = -1, on_rq, running;
5229 unsigned long flags;
5230 const struct sched_class *prev_class;
5234 /* may grab non-irq protected spin_locks */
5235 BUG_ON(in_interrupt());
5237 /* double check policy once rq lock held */
5239 reset_on_fork = p->sched_reset_on_fork;
5240 policy = oldpolicy = p->policy;
5242 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5243 policy &= ~SCHED_RESET_ON_FORK;
5245 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5246 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5247 policy != SCHED_IDLE)
5252 * Valid priorities for SCHED_FIFO and SCHED_RR are
5253 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5254 * SCHED_BATCH and SCHED_IDLE is 0.
5256 if (param->sched_priority < 0 ||
5257 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5258 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5260 if (rt_policy(policy) != (param->sched_priority != 0))
5264 * Allow unprivileged RT tasks to decrease priority:
5266 if (user && !capable(CAP_SYS_NICE)) {
5267 if (rt_policy(policy)) {
5268 unsigned long rlim_rtprio =
5269 task_rlimit(p, RLIMIT_RTPRIO);
5271 /* can't set/change the rt policy */
5272 if (policy != p->policy && !rlim_rtprio)
5275 /* can't increase priority */
5276 if (param->sched_priority > p->rt_priority &&
5277 param->sched_priority > rlim_rtprio)
5282 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5283 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5285 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5286 if (!can_nice(p, TASK_NICE(p)))
5290 /* can't change other user's priorities */
5291 if (!check_same_owner(p))
5294 /* Normal users shall not reset the sched_reset_on_fork flag */
5295 if (p->sched_reset_on_fork && !reset_on_fork)
5300 retval = security_task_setscheduler(p);
5306 * make sure no PI-waiters arrive (or leave) while we are
5307 * changing the priority of the task:
5309 * To be able to change p->policy safely, the appropriate
5310 * runqueue lock must be held.
5312 rq = task_rq_lock(p, &flags);
5315 * Changing the policy of the stop threads its a very bad idea
5317 if (p == rq->stop) {
5318 task_rq_unlock(rq, p, &flags);
5323 * If not changing anything there's no need to proceed further:
5325 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5326 param->sched_priority == p->rt_priority))) {
5328 __task_rq_unlock(rq);
5329 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5333 #ifdef CONFIG_RT_GROUP_SCHED
5336 * Do not allow realtime tasks into groups that have no runtime
5339 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5340 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5341 !task_group_is_autogroup(task_group(p))) {
5342 task_rq_unlock(rq, p, &flags);
5348 /* recheck policy now with rq lock held */
5349 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5350 policy = oldpolicy = -1;
5351 task_rq_unlock(rq, p, &flags);
5355 running = task_current(rq, p);
5357 deactivate_task(rq, p, 0);
5359 p->sched_class->put_prev_task(rq, p);
5361 p->sched_reset_on_fork = reset_on_fork;
5364 prev_class = p->sched_class;
5365 __setscheduler(rq, p, policy, param->sched_priority);
5368 p->sched_class->set_curr_task(rq);
5370 activate_task(rq, p, 0);
5372 check_class_changed(rq, p, prev_class, oldprio);
5373 task_rq_unlock(rq, p, &flags);
5375 rt_mutex_adjust_pi(p);
5381 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5382 * @p: the task in question.
5383 * @policy: new policy.
5384 * @param: structure containing the new RT priority.
5386 * NOTE that the task may be already dead.
5388 int sched_setscheduler(struct task_struct *p, int policy,
5389 const struct sched_param *param)
5391 return __sched_setscheduler(p, policy, param, true);
5393 EXPORT_SYMBOL_GPL(sched_setscheduler);
5396 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5397 * @p: the task in question.
5398 * @policy: new policy.
5399 * @param: structure containing the new RT priority.
5401 * Just like sched_setscheduler, only don't bother checking if the
5402 * current context has permission. For example, this is needed in
5403 * stop_machine(): we create temporary high priority worker threads,
5404 * but our caller might not have that capability.
5406 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5407 const struct sched_param *param)
5409 return __sched_setscheduler(p, policy, param, false);
5413 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5415 struct sched_param lparam;
5416 struct task_struct *p;
5419 if (!param || pid < 0)
5421 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5426 p = find_process_by_pid(pid);
5428 retval = sched_setscheduler(p, policy, &lparam);
5435 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5436 * @pid: the pid in question.
5437 * @policy: new policy.
5438 * @param: structure containing the new RT priority.
5440 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5441 struct sched_param __user *, param)
5443 /* negative values for policy are not valid */
5447 return do_sched_setscheduler(pid, policy, param);
5451 * sys_sched_setparam - set/change the RT priority of a thread
5452 * @pid: the pid in question.
5453 * @param: structure containing the new RT priority.
5455 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5457 return do_sched_setscheduler(pid, -1, param);
5461 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5462 * @pid: the pid in question.
5464 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5466 struct task_struct *p;
5474 p = find_process_by_pid(pid);
5476 retval = security_task_getscheduler(p);
5479 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5486 * sys_sched_getparam - get the RT priority of a thread
5487 * @pid: the pid in question.
5488 * @param: structure containing the RT priority.
5490 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5492 struct sched_param lp;
5493 struct task_struct *p;
5496 if (!param || pid < 0)
5500 p = find_process_by_pid(pid);
5505 retval = security_task_getscheduler(p);
5509 lp.sched_priority = p->rt_priority;
5513 * This one might sleep, we cannot do it with a spinlock held ...
5515 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5524 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5526 cpumask_var_t cpus_allowed, new_mask;
5527 struct task_struct *p;
5533 p = find_process_by_pid(pid);
5540 /* Prevent p going away */
5544 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5548 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5550 goto out_free_cpus_allowed;
5553 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5556 retval = security_task_setscheduler(p);
5560 cpuset_cpus_allowed(p, cpus_allowed);
5561 cpumask_and(new_mask, in_mask, cpus_allowed);
5563 retval = set_cpus_allowed_ptr(p, new_mask);
5566 cpuset_cpus_allowed(p, cpus_allowed);
5567 if (!cpumask_subset(new_mask, cpus_allowed)) {
5569 * We must have raced with a concurrent cpuset
5570 * update. Just reset the cpus_allowed to the
5571 * cpuset's cpus_allowed
5573 cpumask_copy(new_mask, cpus_allowed);
5578 free_cpumask_var(new_mask);
5579 out_free_cpus_allowed:
5580 free_cpumask_var(cpus_allowed);
5587 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5588 struct cpumask *new_mask)
5590 if (len < cpumask_size())
5591 cpumask_clear(new_mask);
5592 else if (len > cpumask_size())
5593 len = cpumask_size();
5595 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5599 * sys_sched_setaffinity - set the cpu affinity of a process
5600 * @pid: pid of the process
5601 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5602 * @user_mask_ptr: user-space pointer to the new cpu mask
5604 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5605 unsigned long __user *, user_mask_ptr)
5607 cpumask_var_t new_mask;
5610 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5613 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5615 retval = sched_setaffinity(pid, new_mask);
5616 free_cpumask_var(new_mask);
5620 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5622 struct task_struct *p;
5623 unsigned long flags;
5630 p = find_process_by_pid(pid);
5634 retval = security_task_getscheduler(p);
5638 raw_spin_lock_irqsave(&p->pi_lock, flags);
5639 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5640 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5650 * sys_sched_getaffinity - get the cpu affinity of a process
5651 * @pid: pid of the process
5652 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5653 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5655 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5656 unsigned long __user *, user_mask_ptr)
5661 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5663 if (len & (sizeof(unsigned long)-1))
5666 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5669 ret = sched_getaffinity(pid, mask);
5671 size_t retlen = min_t(size_t, len, cpumask_size());
5673 if (copy_to_user(user_mask_ptr, mask, retlen))
5678 free_cpumask_var(mask);
5684 * sys_sched_yield - yield the current processor to other threads.
5686 * This function yields the current CPU to other tasks. If there are no
5687 * other threads running on this CPU then this function will return.
5689 SYSCALL_DEFINE0(sched_yield)
5691 struct rq *rq = this_rq_lock();
5693 schedstat_inc(rq, yld_count);
5694 current->sched_class->yield_task(rq);
5697 * Since we are going to call schedule() anyway, there's
5698 * no need to preempt or enable interrupts:
5700 __release(rq->lock);
5701 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5702 do_raw_spin_unlock(&rq->lock);
5703 preempt_enable_no_resched();
5710 static inline int should_resched(void)
5712 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5715 static void __cond_resched(void)
5717 add_preempt_count(PREEMPT_ACTIVE);
5719 sub_preempt_count(PREEMPT_ACTIVE);
5722 int __sched _cond_resched(void)
5724 if (should_resched()) {
5730 EXPORT_SYMBOL(_cond_resched);
5733 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5734 * call schedule, and on return reacquire the lock.
5736 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5737 * operations here to prevent schedule() from being called twice (once via
5738 * spin_unlock(), once by hand).
5740 int __cond_resched_lock(spinlock_t *lock)
5742 int resched = should_resched();
5745 lockdep_assert_held(lock);
5747 if (spin_needbreak(lock) || resched) {
5758 EXPORT_SYMBOL(__cond_resched_lock);
5760 int __sched __cond_resched_softirq(void)
5762 BUG_ON(!in_softirq());
5764 if (should_resched()) {
5772 EXPORT_SYMBOL(__cond_resched_softirq);
5775 * yield - yield the current processor to other threads.
5777 * This is a shortcut for kernel-space yielding - it marks the
5778 * thread runnable and calls sys_sched_yield().
5780 void __sched yield(void)
5782 set_current_state(TASK_RUNNING);
5785 EXPORT_SYMBOL(yield);
5788 * yield_to - yield the current processor to another thread in
5789 * your thread group, or accelerate that thread toward the
5790 * processor it's on.
5792 * @preempt: whether task preemption is allowed or not
5794 * It's the caller's job to ensure that the target task struct
5795 * can't go away on us before we can do any checks.
5797 * Returns true if we indeed boosted the target task.
5799 bool __sched yield_to(struct task_struct *p, bool preempt)
5801 struct task_struct *curr = current;
5802 struct rq *rq, *p_rq;
5803 unsigned long flags;
5806 local_irq_save(flags);
5811 double_rq_lock(rq, p_rq);
5812 while (task_rq(p) != p_rq) {
5813 double_rq_unlock(rq, p_rq);
5817 if (!curr->sched_class->yield_to_task)
5820 if (curr->sched_class != p->sched_class)
5823 if (task_running(p_rq, p) || p->state)
5826 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5828 schedstat_inc(rq, yld_count);
5830 * Make p's CPU reschedule; pick_next_entity takes care of
5833 if (preempt && rq != p_rq)
5834 resched_task(p_rq->curr);
5838 double_rq_unlock(rq, p_rq);
5839 local_irq_restore(flags);
5846 EXPORT_SYMBOL_GPL(yield_to);
5849 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5850 * that process accounting knows that this is a task in IO wait state.
5852 void __sched io_schedule(void)
5854 struct rq *rq = raw_rq();
5856 delayacct_blkio_start();
5857 atomic_inc(&rq->nr_iowait);
5858 blk_flush_plug(current);
5859 current->in_iowait = 1;
5861 current->in_iowait = 0;
5862 atomic_dec(&rq->nr_iowait);
5863 delayacct_blkio_end();
5865 EXPORT_SYMBOL(io_schedule);
5867 long __sched io_schedule_timeout(long timeout)
5869 struct rq *rq = raw_rq();
5872 delayacct_blkio_start();
5873 atomic_inc(&rq->nr_iowait);
5874 blk_flush_plug(current);
5875 current->in_iowait = 1;
5876 ret = schedule_timeout(timeout);
5877 current->in_iowait = 0;
5878 atomic_dec(&rq->nr_iowait);
5879 delayacct_blkio_end();
5884 * sys_sched_get_priority_max - return maximum RT priority.
5885 * @policy: scheduling class.
5887 * this syscall returns the maximum rt_priority that can be used
5888 * by a given scheduling class.
5890 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5897 ret = MAX_USER_RT_PRIO-1;
5909 * sys_sched_get_priority_min - return minimum RT priority.
5910 * @policy: scheduling class.
5912 * this syscall returns the minimum rt_priority that can be used
5913 * by a given scheduling class.
5915 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5933 * sys_sched_rr_get_interval - return the default timeslice of a process.
5934 * @pid: pid of the process.
5935 * @interval: userspace pointer to the timeslice value.
5937 * this syscall writes the default timeslice value of a given process
5938 * into the user-space timespec buffer. A value of '0' means infinity.
5940 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5941 struct timespec __user *, interval)
5943 struct task_struct *p;
5944 unsigned int time_slice;
5945 unsigned long flags;
5955 p = find_process_by_pid(pid);
5959 retval = security_task_getscheduler(p);
5963 rq = task_rq_lock(p, &flags);
5964 time_slice = p->sched_class->get_rr_interval(rq, p);
5965 task_rq_unlock(rq, p, &flags);
5968 jiffies_to_timespec(time_slice, &t);
5969 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5977 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5979 void sched_show_task(struct task_struct *p)
5981 unsigned long free = 0;
5984 state = p->state ? __ffs(p->state) + 1 : 0;
5985 printk(KERN_INFO "%-15.15s %c", p->comm,
5986 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5987 #if BITS_PER_LONG == 32
5988 if (state == TASK_RUNNING)
5989 printk(KERN_CONT " running ");
5991 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5993 if (state == TASK_RUNNING)
5994 printk(KERN_CONT " running task ");
5996 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5998 #ifdef CONFIG_DEBUG_STACK_USAGE
5999 free = stack_not_used(p);
6001 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6002 task_pid_nr(p), task_pid_nr(p->real_parent),
6003 (unsigned long)task_thread_info(p)->flags);
6005 show_stack(p, NULL);
6008 void show_state_filter(unsigned long state_filter)
6010 struct task_struct *g, *p;
6012 #if BITS_PER_LONG == 32
6014 " task PC stack pid father\n");
6017 " task PC stack pid father\n");
6019 read_lock(&tasklist_lock);
6020 do_each_thread(g, p) {
6022 * reset the NMI-timeout, listing all files on a slow
6023 * console might take a lot of time:
6025 touch_nmi_watchdog();
6026 if (!state_filter || (p->state & state_filter))
6028 } while_each_thread(g, p);
6030 touch_all_softlockup_watchdogs();
6032 #ifdef CONFIG_SCHED_DEBUG
6033 sysrq_sched_debug_show();
6035 read_unlock(&tasklist_lock);
6037 * Only show locks if all tasks are dumped:
6040 debug_show_all_locks();
6043 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6045 idle->sched_class = &idle_sched_class;
6049 * init_idle - set up an idle thread for a given CPU
6050 * @idle: task in question
6051 * @cpu: cpu the idle task belongs to
6053 * NOTE: this function does not set the idle thread's NEED_RESCHED
6054 * flag, to make booting more robust.
6056 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6058 struct rq *rq = cpu_rq(cpu);
6059 unsigned long flags;
6061 raw_spin_lock_irqsave(&rq->lock, flags);
6064 idle->state = TASK_RUNNING;
6065 idle->se.exec_start = sched_clock();
6067 do_set_cpus_allowed(idle, cpumask_of(cpu));
6069 * We're having a chicken and egg problem, even though we are
6070 * holding rq->lock, the cpu isn't yet set to this cpu so the
6071 * lockdep check in task_group() will fail.
6073 * Similar case to sched_fork(). / Alternatively we could
6074 * use task_rq_lock() here and obtain the other rq->lock.
6079 __set_task_cpu(idle, cpu);
6082 rq->curr = rq->idle = idle;
6083 #if defined(CONFIG_SMP)
6086 raw_spin_unlock_irqrestore(&rq->lock, flags);
6088 /* Set the preempt count _outside_ the spinlocks! */
6089 task_thread_info(idle)->preempt_count = 0;
6092 * The idle tasks have their own, simple scheduling class:
6094 idle->sched_class = &idle_sched_class;
6095 ftrace_graph_init_idle_task(idle, cpu);
6099 * In a system that switches off the HZ timer nohz_cpu_mask
6100 * indicates which cpus entered this state. This is used
6101 * in the rcu update to wait only for active cpus. For system
6102 * which do not switch off the HZ timer nohz_cpu_mask should
6103 * always be CPU_BITS_NONE.
6105 cpumask_var_t nohz_cpu_mask;
6108 * Increase the granularity value when there are more CPUs,
6109 * because with more CPUs the 'effective latency' as visible
6110 * to users decreases. But the relationship is not linear,
6111 * so pick a second-best guess by going with the log2 of the
6114 * This idea comes from the SD scheduler of Con Kolivas:
6116 static int get_update_sysctl_factor(void)
6118 unsigned int cpus = min_t(int, num_online_cpus(), 8);
6119 unsigned int factor;
6121 switch (sysctl_sched_tunable_scaling) {
6122 case SCHED_TUNABLESCALING_NONE:
6125 case SCHED_TUNABLESCALING_LINEAR:
6128 case SCHED_TUNABLESCALING_LOG:
6130 factor = 1 + ilog2(cpus);
6137 static void update_sysctl(void)
6139 unsigned int factor = get_update_sysctl_factor();
6141 #define SET_SYSCTL(name) \
6142 (sysctl_##name = (factor) * normalized_sysctl_##name)
6143 SET_SYSCTL(sched_min_granularity);
6144 SET_SYSCTL(sched_latency);
6145 SET_SYSCTL(sched_wakeup_granularity);
6149 static inline void sched_init_granularity(void)
6155 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
6157 if (p->sched_class && p->sched_class->set_cpus_allowed)
6158 p->sched_class->set_cpus_allowed(p, new_mask);
6160 cpumask_copy(&p->cpus_allowed, new_mask);
6161 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6166 * This is how migration works:
6168 * 1) we invoke migration_cpu_stop() on the target CPU using
6170 * 2) stopper starts to run (implicitly forcing the migrated thread
6172 * 3) it checks whether the migrated task is still in the wrong runqueue.
6173 * 4) if it's in the wrong runqueue then the migration thread removes
6174 * it and puts it into the right queue.
6175 * 5) stopper completes and stop_one_cpu() returns and the migration
6180 * Change a given task's CPU affinity. Migrate the thread to a
6181 * proper CPU and schedule it away if the CPU it's executing on
6182 * is removed from the allowed bitmask.
6184 * NOTE: the caller must have a valid reference to the task, the
6185 * task must not exit() & deallocate itself prematurely. The
6186 * call is not atomic; no spinlocks may be held.
6188 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6190 unsigned long flags;
6192 unsigned int dest_cpu;
6195 rq = task_rq_lock(p, &flags);
6197 if (cpumask_equal(&p->cpus_allowed, new_mask))
6200 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6205 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6210 do_set_cpus_allowed(p, new_mask);
6212 /* Can the task run on the task's current CPU? If so, we're done */
6213 if (cpumask_test_cpu(task_cpu(p), new_mask))
6216 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6218 struct migration_arg arg = { p, dest_cpu };
6219 /* Need help from migration thread: drop lock and wait. */
6220 task_rq_unlock(rq, p, &flags);
6221 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6222 tlb_migrate_finish(p->mm);
6226 task_rq_unlock(rq, p, &flags);
6230 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6233 * Move (not current) task off this cpu, onto dest cpu. We're doing
6234 * this because either it can't run here any more (set_cpus_allowed()
6235 * away from this CPU, or CPU going down), or because we're
6236 * attempting to rebalance this task on exec (sched_exec).
6238 * So we race with normal scheduler movements, but that's OK, as long
6239 * as the task is no longer on this CPU.
6241 * Returns non-zero if task was successfully migrated.
6243 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6245 struct rq *rq_dest, *rq_src;
6248 if (unlikely(!cpu_active(dest_cpu)))
6251 rq_src = cpu_rq(src_cpu);
6252 rq_dest = cpu_rq(dest_cpu);
6254 raw_spin_lock(&p->pi_lock);
6255 double_rq_lock(rq_src, rq_dest);
6256 /* Already moved. */
6257 if (task_cpu(p) != src_cpu)
6259 /* Affinity changed (again). */
6260 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6264 * If we're not on a rq, the next wake-up will ensure we're
6268 deactivate_task(rq_src, p, 0);
6269 set_task_cpu(p, dest_cpu);
6270 activate_task(rq_dest, p, 0);
6271 check_preempt_curr(rq_dest, p, 0);
6276 double_rq_unlock(rq_src, rq_dest);
6277 raw_spin_unlock(&p->pi_lock);
6282 * migration_cpu_stop - this will be executed by a highprio stopper thread
6283 * and performs thread migration by bumping thread off CPU then
6284 * 'pushing' onto another runqueue.
6286 static int migration_cpu_stop(void *data)
6288 struct migration_arg *arg = data;
6291 * The original target cpu might have gone down and we might
6292 * be on another cpu but it doesn't matter.
6294 local_irq_disable();
6295 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6300 #ifdef CONFIG_HOTPLUG_CPU
6303 * Ensures that the idle task is using init_mm right before its cpu goes
6306 void idle_task_exit(void)
6308 struct mm_struct *mm = current->active_mm;
6310 BUG_ON(cpu_online(smp_processor_id()));
6313 switch_mm(mm, &init_mm, current);
6318 * While a dead CPU has no uninterruptible tasks queued at this point,
6319 * it might still have a nonzero ->nr_uninterruptible counter, because
6320 * for performance reasons the counter is not stricly tracking tasks to
6321 * their home CPUs. So we just add the counter to another CPU's counter,
6322 * to keep the global sum constant after CPU-down:
6324 static void migrate_nr_uninterruptible(struct rq *rq_src)
6326 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6328 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6329 rq_src->nr_uninterruptible = 0;
6333 * remove the tasks which were accounted by rq from calc_load_tasks.
6335 static void calc_global_load_remove(struct rq *rq)
6337 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6338 rq->calc_load_active = 0;
6341 #ifdef CONFIG_CFS_BANDWIDTH
6342 static void unthrottle_offline_cfs_rqs(struct rq *rq)
6344 struct cfs_rq *cfs_rq;
6346 for_each_leaf_cfs_rq(rq, cfs_rq) {
6347 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
6349 if (!cfs_rq->runtime_enabled)
6353 * clock_task is not advancing so we just need to make sure
6354 * there's some valid quota amount
6356 cfs_rq->runtime_remaining = cfs_b->quota;
6357 if (cfs_rq_throttled(cfs_rq))
6358 unthrottle_cfs_rq(cfs_rq);
6362 static void unthrottle_offline_cfs_rqs(struct rq *rq) {}
6366 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6367 * try_to_wake_up()->select_task_rq().
6369 * Called with rq->lock held even though we'er in stop_machine() and
6370 * there's no concurrency possible, we hold the required locks anyway
6371 * because of lock validation efforts.
6373 static void migrate_tasks(unsigned int dead_cpu)
6375 struct rq *rq = cpu_rq(dead_cpu);
6376 struct task_struct *next, *stop = rq->stop;
6380 * Fudge the rq selection such that the below task selection loop
6381 * doesn't get stuck on the currently eligible stop task.
6383 * We're currently inside stop_machine() and the rq is either stuck
6384 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6385 * either way we should never end up calling schedule() until we're
6390 /* Ensure any throttled groups are reachable by pick_next_task */
6391 unthrottle_offline_cfs_rqs(rq);
6395 * There's this thread running, bail when that's the only
6398 if (rq->nr_running == 1)
6401 next = pick_next_task(rq);
6403 next->sched_class->put_prev_task(rq, next);
6405 /* Find suitable destination for @next, with force if needed. */
6406 dest_cpu = select_fallback_rq(dead_cpu, next);
6407 raw_spin_unlock(&rq->lock);
6409 __migrate_task(next, dead_cpu, dest_cpu);
6411 raw_spin_lock(&rq->lock);
6417 #endif /* CONFIG_HOTPLUG_CPU */
6419 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6421 static struct ctl_table sd_ctl_dir[] = {
6423 .procname = "sched_domain",
6429 static struct ctl_table sd_ctl_root[] = {
6431 .procname = "kernel",
6433 .child = sd_ctl_dir,
6438 static struct ctl_table *sd_alloc_ctl_entry(int n)
6440 struct ctl_table *entry =
6441 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6446 static void sd_free_ctl_entry(struct ctl_table **tablep)
6448 struct ctl_table *entry;
6451 * In the intermediate directories, both the child directory and
6452 * procname are dynamically allocated and could fail but the mode
6453 * will always be set. In the lowest directory the names are
6454 * static strings and all have proc handlers.
6456 for (entry = *tablep; entry->mode; entry++) {
6458 sd_free_ctl_entry(&entry->child);
6459 if (entry->proc_handler == NULL)
6460 kfree(entry->procname);
6468 set_table_entry(struct ctl_table *entry,
6469 const char *procname, void *data, int maxlen,
6470 mode_t mode, proc_handler *proc_handler)
6472 entry->procname = procname;
6474 entry->maxlen = maxlen;
6476 entry->proc_handler = proc_handler;
6479 static struct ctl_table *
6480 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6482 struct ctl_table *table = sd_alloc_ctl_entry(13);
6487 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6488 sizeof(long), 0644, proc_doulongvec_minmax);
6489 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6490 sizeof(long), 0644, proc_doulongvec_minmax);
6491 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6492 sizeof(int), 0644, proc_dointvec_minmax);
6493 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6494 sizeof(int), 0644, proc_dointvec_minmax);
6495 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6496 sizeof(int), 0644, proc_dointvec_minmax);
6497 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6498 sizeof(int), 0644, proc_dointvec_minmax);
6499 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6500 sizeof(int), 0644, proc_dointvec_minmax);
6501 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6502 sizeof(int), 0644, proc_dointvec_minmax);
6503 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6504 sizeof(int), 0644, proc_dointvec_minmax);
6505 set_table_entry(&table[9], "cache_nice_tries",
6506 &sd->cache_nice_tries,
6507 sizeof(int), 0644, proc_dointvec_minmax);
6508 set_table_entry(&table[10], "flags", &sd->flags,
6509 sizeof(int), 0644, proc_dointvec_minmax);
6510 set_table_entry(&table[11], "name", sd->name,
6511 CORENAME_MAX_SIZE, 0444, proc_dostring);
6512 /* &table[12] is terminator */
6517 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6519 struct ctl_table *entry, *table;
6520 struct sched_domain *sd;
6521 int domain_num = 0, i;
6524 for_each_domain(cpu, sd)
6526 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6531 for_each_domain(cpu, sd) {
6532 snprintf(buf, 32, "domain%d", i);
6533 entry->procname = kstrdup(buf, GFP_KERNEL);
6535 entry->child = sd_alloc_ctl_domain_table(sd);
6542 static struct ctl_table_header *sd_sysctl_header;
6543 static void register_sched_domain_sysctl(void)
6545 int i, cpu_num = num_possible_cpus();
6546 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6549 WARN_ON(sd_ctl_dir[0].child);
6550 sd_ctl_dir[0].child = entry;
6555 for_each_possible_cpu(i) {
6556 snprintf(buf, 32, "cpu%d", i);
6557 entry->procname = kstrdup(buf, GFP_KERNEL);
6559 entry->child = sd_alloc_ctl_cpu_table(i);
6563 WARN_ON(sd_sysctl_header);
6564 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6567 /* may be called multiple times per register */
6568 static void unregister_sched_domain_sysctl(void)
6570 if (sd_sysctl_header)
6571 unregister_sysctl_table(sd_sysctl_header);
6572 sd_sysctl_header = NULL;
6573 if (sd_ctl_dir[0].child)
6574 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6577 static void register_sched_domain_sysctl(void)
6580 static void unregister_sched_domain_sysctl(void)
6585 static void set_rq_online(struct rq *rq)
6588 const struct sched_class *class;
6590 cpumask_set_cpu(rq->cpu, rq->rd->online);
6593 for_each_class(class) {
6594 if (class->rq_online)
6595 class->rq_online(rq);
6600 static void set_rq_offline(struct rq *rq)
6603 const struct sched_class *class;
6605 for_each_class(class) {
6606 if (class->rq_offline)
6607 class->rq_offline(rq);
6610 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6616 * migration_call - callback that gets triggered when a CPU is added.
6617 * Here we can start up the necessary migration thread for the new CPU.
6619 static int __cpuinit
6620 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6622 int cpu = (long)hcpu;
6623 unsigned long flags;
6624 struct rq *rq = cpu_rq(cpu);
6626 switch (action & ~CPU_TASKS_FROZEN) {
6628 case CPU_UP_PREPARE:
6629 rq->calc_load_update = calc_load_update;
6633 /* Update our root-domain */
6634 raw_spin_lock_irqsave(&rq->lock, flags);
6636 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6640 raw_spin_unlock_irqrestore(&rq->lock, flags);
6643 #ifdef CONFIG_HOTPLUG_CPU
6645 sched_ttwu_pending();
6646 /* Update our root-domain */
6647 raw_spin_lock_irqsave(&rq->lock, flags);
6649 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6653 BUG_ON(rq->nr_running != 1); /* the migration thread */
6654 raw_spin_unlock_irqrestore(&rq->lock, flags);
6656 migrate_nr_uninterruptible(rq);
6657 calc_global_load_remove(rq);
6662 update_max_interval();
6668 * Register at high priority so that task migration (migrate_all_tasks)
6669 * happens before everything else. This has to be lower priority than
6670 * the notifier in the perf_event subsystem, though.
6672 static struct notifier_block __cpuinitdata migration_notifier = {
6673 .notifier_call = migration_call,
6674 .priority = CPU_PRI_MIGRATION,
6677 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6678 unsigned long action, void *hcpu)
6680 switch (action & ~CPU_TASKS_FROZEN) {
6682 case CPU_DOWN_FAILED:
6683 set_cpu_active((long)hcpu, true);
6690 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6691 unsigned long action, void *hcpu)
6693 switch (action & ~CPU_TASKS_FROZEN) {
6694 case CPU_DOWN_PREPARE:
6695 set_cpu_active((long)hcpu, false);
6702 static int __init migration_init(void)
6704 void *cpu = (void *)(long)smp_processor_id();
6707 /* Initialize migration for the boot CPU */
6708 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6709 BUG_ON(err == NOTIFY_BAD);
6710 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6711 register_cpu_notifier(&migration_notifier);
6713 /* Register cpu active notifiers */
6714 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6715 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6719 early_initcall(migration_init);
6724 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6726 #ifdef CONFIG_SCHED_DEBUG
6728 static __read_mostly int sched_domain_debug_enabled;
6730 static int __init sched_domain_debug_setup(char *str)
6732 sched_domain_debug_enabled = 1;
6736 early_param("sched_debug", sched_domain_debug_setup);
6738 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6739 struct cpumask *groupmask)
6741 struct sched_group *group = sd->groups;
6744 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6745 cpumask_clear(groupmask);
6747 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6749 if (!(sd->flags & SD_LOAD_BALANCE)) {
6750 printk("does not load-balance\n");
6752 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6757 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6759 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6760 printk(KERN_ERR "ERROR: domain->span does not contain "
6763 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6764 printk(KERN_ERR "ERROR: domain->groups does not contain"
6768 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6772 printk(KERN_ERR "ERROR: group is NULL\n");
6776 if (!group->sgp->power) {
6777 printk(KERN_CONT "\n");
6778 printk(KERN_ERR "ERROR: domain->cpu_power not "
6783 if (!cpumask_weight(sched_group_cpus(group))) {
6784 printk(KERN_CONT "\n");
6785 printk(KERN_ERR "ERROR: empty group\n");
6789 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6790 printk(KERN_CONT "\n");
6791 printk(KERN_ERR "ERROR: repeated CPUs\n");
6795 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6797 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6799 printk(KERN_CONT " %s", str);
6800 if (group->sgp->power != SCHED_POWER_SCALE) {
6801 printk(KERN_CONT " (cpu_power = %d)",
6805 group = group->next;
6806 } while (group != sd->groups);
6807 printk(KERN_CONT "\n");
6809 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6810 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6813 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6814 printk(KERN_ERR "ERROR: parent span is not a superset "
6815 "of domain->span\n");
6819 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6823 if (!sched_domain_debug_enabled)
6827 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6831 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6834 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6842 #else /* !CONFIG_SCHED_DEBUG */
6843 # define sched_domain_debug(sd, cpu) do { } while (0)
6844 #endif /* CONFIG_SCHED_DEBUG */
6846 static int sd_degenerate(struct sched_domain *sd)
6848 if (cpumask_weight(sched_domain_span(sd)) == 1)
6851 /* Following flags need at least 2 groups */
6852 if (sd->flags & (SD_LOAD_BALANCE |
6853 SD_BALANCE_NEWIDLE |
6857 SD_SHARE_PKG_RESOURCES)) {
6858 if (sd->groups != sd->groups->next)
6862 /* Following flags don't use groups */
6863 if (sd->flags & (SD_WAKE_AFFINE))
6870 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6872 unsigned long cflags = sd->flags, pflags = parent->flags;
6874 if (sd_degenerate(parent))
6877 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6880 /* Flags needing groups don't count if only 1 group in parent */
6881 if (parent->groups == parent->groups->next) {
6882 pflags &= ~(SD_LOAD_BALANCE |
6883 SD_BALANCE_NEWIDLE |
6887 SD_SHARE_PKG_RESOURCES);
6888 if (nr_node_ids == 1)
6889 pflags &= ~SD_SERIALIZE;
6891 if (~cflags & pflags)
6897 static void free_rootdomain(struct rcu_head *rcu)
6899 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6901 cpupri_cleanup(&rd->cpupri);
6902 free_cpumask_var(rd->rto_mask);
6903 free_cpumask_var(rd->online);
6904 free_cpumask_var(rd->span);
6908 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6910 struct root_domain *old_rd = NULL;
6911 unsigned long flags;
6913 raw_spin_lock_irqsave(&rq->lock, flags);
6918 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6921 cpumask_clear_cpu(rq->cpu, old_rd->span);
6924 * If we dont want to free the old_rt yet then
6925 * set old_rd to NULL to skip the freeing later
6928 if (!atomic_dec_and_test(&old_rd->refcount))
6932 atomic_inc(&rd->refcount);
6935 cpumask_set_cpu(rq->cpu, rd->span);
6936 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6939 raw_spin_unlock_irqrestore(&rq->lock, flags);
6942 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6945 static int init_rootdomain(struct root_domain *rd)
6947 memset(rd, 0, sizeof(*rd));
6949 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6951 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6953 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6956 if (cpupri_init(&rd->cpupri) != 0)
6961 free_cpumask_var(rd->rto_mask);
6963 free_cpumask_var(rd->online);
6965 free_cpumask_var(rd->span);
6970 static void init_defrootdomain(void)
6972 init_rootdomain(&def_root_domain);
6974 atomic_set(&def_root_domain.refcount, 1);
6977 static struct root_domain *alloc_rootdomain(void)
6979 struct root_domain *rd;
6981 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6985 if (init_rootdomain(rd) != 0) {
6993 static void free_sched_groups(struct sched_group *sg, int free_sgp)
6995 struct sched_group *tmp, *first;
7004 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
7009 } while (sg != first);
7012 static void free_sched_domain(struct rcu_head *rcu)
7014 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
7017 * If its an overlapping domain it has private groups, iterate and
7020 if (sd->flags & SD_OVERLAP) {
7021 free_sched_groups(sd->groups, 1);
7022 } else if (atomic_dec_and_test(&sd->groups->ref)) {
7023 kfree(sd->groups->sgp);
7029 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
7031 call_rcu(&sd->rcu, free_sched_domain);
7034 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
7036 for (; sd; sd = sd->parent)
7037 destroy_sched_domain(sd, cpu);
7041 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7042 * hold the hotplug lock.
7045 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7047 struct rq *rq = cpu_rq(cpu);
7048 struct sched_domain *tmp;
7050 /* Remove the sched domains which do not contribute to scheduling. */
7051 for (tmp = sd; tmp; ) {
7052 struct sched_domain *parent = tmp->parent;
7056 if (sd_parent_degenerate(tmp, parent)) {
7057 tmp->parent = parent->parent;
7059 parent->parent->child = tmp;
7060 destroy_sched_domain(parent, cpu);
7065 if (sd && sd_degenerate(sd)) {
7068 destroy_sched_domain(tmp, cpu);
7073 sched_domain_debug(sd, cpu);
7075 rq_attach_root(rq, rd);
7077 rcu_assign_pointer(rq->sd, sd);
7078 destroy_sched_domains(tmp, cpu);
7081 /* cpus with isolated domains */
7082 static cpumask_var_t cpu_isolated_map;
7084 /* Setup the mask of cpus configured for isolated domains */
7085 static int __init isolated_cpu_setup(char *str)
7087 alloc_bootmem_cpumask_var(&cpu_isolated_map);
7088 cpulist_parse(str, cpu_isolated_map);
7092 __setup("isolcpus=", isolated_cpu_setup);
7094 #define SD_NODES_PER_DOMAIN 16
7099 * find_next_best_node - find the next node to include in a sched_domain
7100 * @node: node whose sched_domain we're building
7101 * @used_nodes: nodes already in the sched_domain
7103 * Find the next node to include in a given scheduling domain. Simply
7104 * finds the closest node not already in the @used_nodes map.
7106 * Should use nodemask_t.
7108 static int find_next_best_node(int node, nodemask_t *used_nodes)
7110 int i, n, val, min_val, best_node = -1;
7114 for (i = 0; i < nr_node_ids; i++) {
7115 /* Start at @node */
7116 n = (node + i) % nr_node_ids;
7118 if (!nr_cpus_node(n))
7121 /* Skip already used nodes */
7122 if (node_isset(n, *used_nodes))
7125 /* Simple min distance search */
7126 val = node_distance(node, n);
7128 if (val < min_val) {
7134 if (best_node != -1)
7135 node_set(best_node, *used_nodes);
7140 * sched_domain_node_span - get a cpumask for a node's sched_domain
7141 * @node: node whose cpumask we're constructing
7142 * @span: resulting cpumask
7144 * Given a node, construct a good cpumask for its sched_domain to span. It
7145 * should be one that prevents unnecessary balancing, but also spreads tasks
7148 static void sched_domain_node_span(int node, struct cpumask *span)
7150 nodemask_t used_nodes;
7153 cpumask_clear(span);
7154 nodes_clear(used_nodes);
7156 cpumask_or(span, span, cpumask_of_node(node));
7157 node_set(node, used_nodes);
7159 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7160 int next_node = find_next_best_node(node, &used_nodes);
7163 cpumask_or(span, span, cpumask_of_node(next_node));
7167 static const struct cpumask *cpu_node_mask(int cpu)
7169 lockdep_assert_held(&sched_domains_mutex);
7171 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
7173 return sched_domains_tmpmask;
7176 static const struct cpumask *cpu_allnodes_mask(int cpu)
7178 return cpu_possible_mask;
7180 #endif /* CONFIG_NUMA */
7182 static const struct cpumask *cpu_cpu_mask(int cpu)
7184 return cpumask_of_node(cpu_to_node(cpu));
7187 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7190 struct sched_domain **__percpu sd;
7191 struct sched_group **__percpu sg;
7192 struct sched_group_power **__percpu sgp;
7196 struct sched_domain ** __percpu sd;
7197 struct root_domain *rd;
7207 struct sched_domain_topology_level;
7209 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
7210 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
7212 #define SDTL_OVERLAP 0x01
7214 struct sched_domain_topology_level {
7215 sched_domain_init_f init;
7216 sched_domain_mask_f mask;
7218 struct sd_data data;
7222 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
7224 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
7225 const struct cpumask *span = sched_domain_span(sd);
7226 struct cpumask *covered = sched_domains_tmpmask;
7227 struct sd_data *sdd = sd->private;
7228 struct sched_domain *child;
7231 cpumask_clear(covered);
7233 for_each_cpu(i, span) {
7234 struct cpumask *sg_span;
7236 if (cpumask_test_cpu(i, covered))
7239 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7240 GFP_KERNEL, cpu_to_node(i));
7245 sg_span = sched_group_cpus(sg);
7247 child = *per_cpu_ptr(sdd->sd, i);
7249 child = child->child;
7250 cpumask_copy(sg_span, sched_domain_span(child));
7252 cpumask_set_cpu(i, sg_span);
7254 cpumask_or(covered, covered, sg_span);
7256 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
7257 atomic_inc(&sg->sgp->ref);
7259 if (cpumask_test_cpu(cpu, sg_span))
7269 sd->groups = groups;
7274 free_sched_groups(first, 0);
7279 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
7281 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
7282 struct sched_domain *child = sd->child;
7285 cpu = cpumask_first(sched_domain_span(child));
7288 *sg = *per_cpu_ptr(sdd->sg, cpu);
7289 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
7290 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
7297 * build_sched_groups will build a circular linked list of the groups
7298 * covered by the given span, and will set each group's ->cpumask correctly,
7299 * and ->cpu_power to 0.
7301 * Assumes the sched_domain tree is fully constructed
7304 build_sched_groups(struct sched_domain *sd, int cpu)
7306 struct sched_group *first = NULL, *last = NULL;
7307 struct sd_data *sdd = sd->private;
7308 const struct cpumask *span = sched_domain_span(sd);
7309 struct cpumask *covered;
7312 get_group(cpu, sdd, &sd->groups);
7313 atomic_inc(&sd->groups->ref);
7315 if (cpu != cpumask_first(sched_domain_span(sd)))
7318 lockdep_assert_held(&sched_domains_mutex);
7319 covered = sched_domains_tmpmask;
7321 cpumask_clear(covered);
7323 for_each_cpu(i, span) {
7324 struct sched_group *sg;
7325 int group = get_group(i, sdd, &sg);
7328 if (cpumask_test_cpu(i, covered))
7331 cpumask_clear(sched_group_cpus(sg));
7334 for_each_cpu(j, span) {
7335 if (get_group(j, sdd, NULL) != group)
7338 cpumask_set_cpu(j, covered);
7339 cpumask_set_cpu(j, sched_group_cpus(sg));
7354 * Initialize sched groups cpu_power.
7356 * cpu_power indicates the capacity of sched group, which is used while
7357 * distributing the load between different sched groups in a sched domain.
7358 * Typically cpu_power for all the groups in a sched domain will be same unless
7359 * there are asymmetries in the topology. If there are asymmetries, group
7360 * having more cpu_power will pickup more load compared to the group having
7363 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7365 struct sched_group *sg = sd->groups;
7367 WARN_ON(!sd || !sg);
7370 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
7372 } while (sg != sd->groups);
7374 if (cpu != group_first_cpu(sg))
7377 update_group_power(sd, cpu);
7381 * Initializers for schedule domains
7382 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7385 #ifdef CONFIG_SCHED_DEBUG
7386 # define SD_INIT_NAME(sd, type) sd->name = #type
7388 # define SD_INIT_NAME(sd, type) do { } while (0)
7391 #define SD_INIT_FUNC(type) \
7392 static noinline struct sched_domain * \
7393 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7395 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7396 *sd = SD_##type##_INIT; \
7397 SD_INIT_NAME(sd, type); \
7398 sd->private = &tl->data; \
7404 SD_INIT_FUNC(ALLNODES)
7407 #ifdef CONFIG_SCHED_SMT
7408 SD_INIT_FUNC(SIBLING)
7410 #ifdef CONFIG_SCHED_MC
7413 #ifdef CONFIG_SCHED_BOOK
7417 static int default_relax_domain_level = -1;
7418 int sched_domain_level_max;
7420 static int __init setup_relax_domain_level(char *str)
7424 val = simple_strtoul(str, NULL, 0);
7425 if (val < sched_domain_level_max)
7426 default_relax_domain_level = val;
7430 __setup("relax_domain_level=", setup_relax_domain_level);
7432 static void set_domain_attribute(struct sched_domain *sd,
7433 struct sched_domain_attr *attr)
7437 if (!attr || attr->relax_domain_level < 0) {
7438 if (default_relax_domain_level < 0)
7441 request = default_relax_domain_level;
7443 request = attr->relax_domain_level;
7444 if (request < sd->level) {
7445 /* turn off idle balance on this domain */
7446 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7448 /* turn on idle balance on this domain */
7449 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7453 static void __sdt_free(const struct cpumask *cpu_map);
7454 static int __sdt_alloc(const struct cpumask *cpu_map);
7456 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7457 const struct cpumask *cpu_map)
7461 if (!atomic_read(&d->rd->refcount))
7462 free_rootdomain(&d->rd->rcu); /* fall through */
7464 free_percpu(d->sd); /* fall through */
7466 __sdt_free(cpu_map); /* fall through */
7472 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7473 const struct cpumask *cpu_map)
7475 memset(d, 0, sizeof(*d));
7477 if (__sdt_alloc(cpu_map))
7478 return sa_sd_storage;
7479 d->sd = alloc_percpu(struct sched_domain *);
7481 return sa_sd_storage;
7482 d->rd = alloc_rootdomain();
7485 return sa_rootdomain;
7489 * NULL the sd_data elements we've used to build the sched_domain and
7490 * sched_group structure so that the subsequent __free_domain_allocs()
7491 * will not free the data we're using.
7493 static void claim_allocations(int cpu, struct sched_domain *sd)
7495 struct sd_data *sdd = sd->private;
7497 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7498 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7500 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
7501 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7503 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
7504 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
7507 #ifdef CONFIG_SCHED_SMT
7508 static const struct cpumask *cpu_smt_mask(int cpu)
7510 return topology_thread_cpumask(cpu);
7515 * Topology list, bottom-up.
7517 static struct sched_domain_topology_level default_topology[] = {
7518 #ifdef CONFIG_SCHED_SMT
7519 { sd_init_SIBLING, cpu_smt_mask, },
7521 #ifdef CONFIG_SCHED_MC
7522 { sd_init_MC, cpu_coregroup_mask, },
7524 #ifdef CONFIG_SCHED_BOOK
7525 { sd_init_BOOK, cpu_book_mask, },
7527 { sd_init_CPU, cpu_cpu_mask, },
7529 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
7530 { sd_init_ALLNODES, cpu_allnodes_mask, },
7535 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7537 static int __sdt_alloc(const struct cpumask *cpu_map)
7539 struct sched_domain_topology_level *tl;
7542 for (tl = sched_domain_topology; tl->init; tl++) {
7543 struct sd_data *sdd = &tl->data;
7545 sdd->sd = alloc_percpu(struct sched_domain *);
7549 sdd->sg = alloc_percpu(struct sched_group *);
7553 sdd->sgp = alloc_percpu(struct sched_group_power *);
7557 for_each_cpu(j, cpu_map) {
7558 struct sched_domain *sd;
7559 struct sched_group *sg;
7560 struct sched_group_power *sgp;
7562 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7563 GFP_KERNEL, cpu_to_node(j));
7567 *per_cpu_ptr(sdd->sd, j) = sd;
7569 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7570 GFP_KERNEL, cpu_to_node(j));
7574 *per_cpu_ptr(sdd->sg, j) = sg;
7576 sgp = kzalloc_node(sizeof(struct sched_group_power),
7577 GFP_KERNEL, cpu_to_node(j));
7581 *per_cpu_ptr(sdd->sgp, j) = sgp;
7588 static void __sdt_free(const struct cpumask *cpu_map)
7590 struct sched_domain_topology_level *tl;
7593 for (tl = sched_domain_topology; tl->init; tl++) {
7594 struct sd_data *sdd = &tl->data;
7596 for_each_cpu(j, cpu_map) {
7597 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
7598 if (sd && (sd->flags & SD_OVERLAP))
7599 free_sched_groups(sd->groups, 0);
7600 kfree(*per_cpu_ptr(sdd->sd, j));
7601 kfree(*per_cpu_ptr(sdd->sg, j));
7602 kfree(*per_cpu_ptr(sdd->sgp, j));
7604 free_percpu(sdd->sd);
7605 free_percpu(sdd->sg);
7606 free_percpu(sdd->sgp);
7610 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7611 struct s_data *d, const struct cpumask *cpu_map,
7612 struct sched_domain_attr *attr, struct sched_domain *child,
7615 struct sched_domain *sd = tl->init(tl, cpu);
7619 set_domain_attribute(sd, attr);
7620 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7622 sd->level = child->level + 1;
7623 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7632 * Build sched domains for a given set of cpus and attach the sched domains
7633 * to the individual cpus
7635 static int build_sched_domains(const struct cpumask *cpu_map,
7636 struct sched_domain_attr *attr)
7638 enum s_alloc alloc_state = sa_none;
7639 struct sched_domain *sd;
7641 int i, ret = -ENOMEM;
7643 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7644 if (alloc_state != sa_rootdomain)
7647 /* Set up domains for cpus specified by the cpu_map. */
7648 for_each_cpu(i, cpu_map) {
7649 struct sched_domain_topology_level *tl;
7652 for (tl = sched_domain_topology; tl->init; tl++) {
7653 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7654 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7655 sd->flags |= SD_OVERLAP;
7656 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7663 *per_cpu_ptr(d.sd, i) = sd;
7666 /* Build the groups for the domains */
7667 for_each_cpu(i, cpu_map) {
7668 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7669 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7670 if (sd->flags & SD_OVERLAP) {
7671 if (build_overlap_sched_groups(sd, i))
7674 if (build_sched_groups(sd, i))
7680 /* Calculate CPU power for physical packages and nodes */
7681 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7682 if (!cpumask_test_cpu(i, cpu_map))
7685 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7686 claim_allocations(i, sd);
7687 init_sched_groups_power(i, sd);
7691 /* Attach the domains */
7693 for_each_cpu(i, cpu_map) {
7694 sd = *per_cpu_ptr(d.sd, i);
7695 cpu_attach_domain(sd, d.rd, i);
7701 __free_domain_allocs(&d, alloc_state, cpu_map);
7705 static cpumask_var_t *doms_cur; /* current sched domains */
7706 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7707 static struct sched_domain_attr *dattr_cur;
7708 /* attribues of custom domains in 'doms_cur' */
7711 * Special case: If a kmalloc of a doms_cur partition (array of
7712 * cpumask) fails, then fallback to a single sched domain,
7713 * as determined by the single cpumask fallback_doms.
7715 static cpumask_var_t fallback_doms;
7718 * arch_update_cpu_topology lets virtualized architectures update the
7719 * cpu core maps. It is supposed to return 1 if the topology changed
7720 * or 0 if it stayed the same.
7722 int __attribute__((weak)) arch_update_cpu_topology(void)
7727 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7730 cpumask_var_t *doms;
7732 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7735 for (i = 0; i < ndoms; i++) {
7736 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7737 free_sched_domains(doms, i);
7744 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7747 for (i = 0; i < ndoms; i++)
7748 free_cpumask_var(doms[i]);
7753 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7754 * For now this just excludes isolated cpus, but could be used to
7755 * exclude other special cases in the future.
7757 static int init_sched_domains(const struct cpumask *cpu_map)
7761 arch_update_cpu_topology();
7763 doms_cur = alloc_sched_domains(ndoms_cur);
7765 doms_cur = &fallback_doms;
7766 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7768 err = build_sched_domains(doms_cur[0], NULL);
7769 register_sched_domain_sysctl();
7775 * Detach sched domains from a group of cpus specified in cpu_map
7776 * These cpus will now be attached to the NULL domain
7778 static void detach_destroy_domains(const struct cpumask *cpu_map)
7783 for_each_cpu(i, cpu_map)
7784 cpu_attach_domain(NULL, &def_root_domain, i);
7788 /* handle null as "default" */
7789 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7790 struct sched_domain_attr *new, int idx_new)
7792 struct sched_domain_attr tmp;
7799 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7800 new ? (new + idx_new) : &tmp,
7801 sizeof(struct sched_domain_attr));
7805 * Partition sched domains as specified by the 'ndoms_new'
7806 * cpumasks in the array doms_new[] of cpumasks. This compares
7807 * doms_new[] to the current sched domain partitioning, doms_cur[].
7808 * It destroys each deleted domain and builds each new domain.
7810 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7811 * The masks don't intersect (don't overlap.) We should setup one
7812 * sched domain for each mask. CPUs not in any of the cpumasks will
7813 * not be load balanced. If the same cpumask appears both in the
7814 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7817 * The passed in 'doms_new' should be allocated using
7818 * alloc_sched_domains. This routine takes ownership of it and will
7819 * free_sched_domains it when done with it. If the caller failed the
7820 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7821 * and partition_sched_domains() will fallback to the single partition
7822 * 'fallback_doms', it also forces the domains to be rebuilt.
7824 * If doms_new == NULL it will be replaced with cpu_online_mask.
7825 * ndoms_new == 0 is a special case for destroying existing domains,
7826 * and it will not create the default domain.
7828 * Call with hotplug lock held
7830 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7831 struct sched_domain_attr *dattr_new)
7836 mutex_lock(&sched_domains_mutex);
7838 /* always unregister in case we don't destroy any domains */
7839 unregister_sched_domain_sysctl();
7841 /* Let architecture update cpu core mappings. */
7842 new_topology = arch_update_cpu_topology();
7844 n = doms_new ? ndoms_new : 0;
7846 /* Destroy deleted domains */
7847 for (i = 0; i < ndoms_cur; i++) {
7848 for (j = 0; j < n && !new_topology; j++) {
7849 if (cpumask_equal(doms_cur[i], doms_new[j])
7850 && dattrs_equal(dattr_cur, i, dattr_new, j))
7853 /* no match - a current sched domain not in new doms_new[] */
7854 detach_destroy_domains(doms_cur[i]);
7859 if (doms_new == NULL) {
7861 doms_new = &fallback_doms;
7862 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7863 WARN_ON_ONCE(dattr_new);
7866 /* Build new domains */
7867 for (i = 0; i < ndoms_new; i++) {
7868 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7869 if (cpumask_equal(doms_new[i], doms_cur[j])
7870 && dattrs_equal(dattr_new, i, dattr_cur, j))
7873 /* no match - add a new doms_new */
7874 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7879 /* Remember the new sched domains */
7880 if (doms_cur != &fallback_doms)
7881 free_sched_domains(doms_cur, ndoms_cur);
7882 kfree(dattr_cur); /* kfree(NULL) is safe */
7883 doms_cur = doms_new;
7884 dattr_cur = dattr_new;
7885 ndoms_cur = ndoms_new;
7887 register_sched_domain_sysctl();
7889 mutex_unlock(&sched_domains_mutex);
7892 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7893 static void reinit_sched_domains(void)
7897 /* Destroy domains first to force the rebuild */
7898 partition_sched_domains(0, NULL, NULL);
7900 rebuild_sched_domains();
7904 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7906 unsigned int level = 0;
7908 if (sscanf(buf, "%u", &level) != 1)
7912 * level is always be positive so don't check for
7913 * level < POWERSAVINGS_BALANCE_NONE which is 0
7914 * What happens on 0 or 1 byte write,
7915 * need to check for count as well?
7918 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7922 sched_smt_power_savings = level;
7924 sched_mc_power_savings = level;
7926 reinit_sched_domains();
7931 #ifdef CONFIG_SCHED_MC
7932 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7933 struct sysdev_class_attribute *attr,
7936 return sprintf(page, "%u\n", sched_mc_power_savings);
7938 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7939 struct sysdev_class_attribute *attr,
7940 const char *buf, size_t count)
7942 return sched_power_savings_store(buf, count, 0);
7944 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7945 sched_mc_power_savings_show,
7946 sched_mc_power_savings_store);
7949 #ifdef CONFIG_SCHED_SMT
7950 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7951 struct sysdev_class_attribute *attr,
7954 return sprintf(page, "%u\n", sched_smt_power_savings);
7956 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7957 struct sysdev_class_attribute *attr,
7958 const char *buf, size_t count)
7960 return sched_power_savings_store(buf, count, 1);
7962 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7963 sched_smt_power_savings_show,
7964 sched_smt_power_savings_store);
7967 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7971 #ifdef CONFIG_SCHED_SMT
7973 err = sysfs_create_file(&cls->kset.kobj,
7974 &attr_sched_smt_power_savings.attr);
7976 #ifdef CONFIG_SCHED_MC
7977 if (!err && mc_capable())
7978 err = sysfs_create_file(&cls->kset.kobj,
7979 &attr_sched_mc_power_savings.attr);
7983 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7986 * Update cpusets according to cpu_active mask. If cpusets are
7987 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7988 * around partition_sched_domains().
7990 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7993 switch (action & ~CPU_TASKS_FROZEN) {
7995 case CPU_DOWN_FAILED:
7996 cpuset_update_active_cpus();
8003 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
8006 switch (action & ~CPU_TASKS_FROZEN) {
8007 case CPU_DOWN_PREPARE:
8008 cpuset_update_active_cpus();
8015 static int update_runtime(struct notifier_block *nfb,
8016 unsigned long action, void *hcpu)
8018 int cpu = (int)(long)hcpu;
8021 case CPU_DOWN_PREPARE:
8022 case CPU_DOWN_PREPARE_FROZEN:
8023 disable_runtime(cpu_rq(cpu));
8026 case CPU_DOWN_FAILED:
8027 case CPU_DOWN_FAILED_FROZEN:
8029 case CPU_ONLINE_FROZEN:
8030 enable_runtime(cpu_rq(cpu));
8038 void __init sched_init_smp(void)
8040 cpumask_var_t non_isolated_cpus;
8042 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8043 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8046 mutex_lock(&sched_domains_mutex);
8047 init_sched_domains(cpu_active_mask);
8048 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8049 if (cpumask_empty(non_isolated_cpus))
8050 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8051 mutex_unlock(&sched_domains_mutex);
8054 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
8055 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
8057 /* RT runtime code needs to handle some hotplug events */
8058 hotcpu_notifier(update_runtime, 0);
8062 /* Move init over to a non-isolated CPU */
8063 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8065 sched_init_granularity();
8066 free_cpumask_var(non_isolated_cpus);
8068 init_sched_rt_class();
8071 void __init sched_init_smp(void)
8073 sched_init_granularity();
8075 #endif /* CONFIG_SMP */
8077 const_debug unsigned int sysctl_timer_migration = 1;
8079 int in_sched_functions(unsigned long addr)
8081 return in_lock_functions(addr) ||
8082 (addr >= (unsigned long)__sched_text_start
8083 && addr < (unsigned long)__sched_text_end);
8086 static void init_cfs_rq(struct cfs_rq *cfs_rq)
8088 cfs_rq->tasks_timeline = RB_ROOT;
8089 INIT_LIST_HEAD(&cfs_rq->tasks);
8090 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8091 #ifndef CONFIG_64BIT
8092 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8096 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8098 struct rt_prio_array *array;
8101 array = &rt_rq->active;
8102 for (i = 0; i < MAX_RT_PRIO; i++) {
8103 INIT_LIST_HEAD(array->queue + i);
8104 __clear_bit(i, array->bitmap);
8106 /* delimiter for bitsearch: */
8107 __set_bit(MAX_RT_PRIO, array->bitmap);
8109 #if defined CONFIG_SMP
8110 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8111 rt_rq->highest_prio.next = MAX_RT_PRIO;
8112 rt_rq->rt_nr_migratory = 0;
8113 rt_rq->overloaded = 0;
8114 plist_head_init(&rt_rq->pushable_tasks);
8118 rt_rq->rt_throttled = 0;
8119 rt_rq->rt_runtime = 0;
8120 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
8123 #ifdef CONFIG_FAIR_GROUP_SCHED
8124 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8125 struct sched_entity *se, int cpu,
8126 struct sched_entity *parent)
8128 struct rq *rq = cpu_rq(cpu);
8133 /* allow initial update_cfs_load() to truncate */
8134 cfs_rq->load_stamp = 1;
8136 init_cfs_rq_runtime(cfs_rq);
8138 tg->cfs_rq[cpu] = cfs_rq;
8141 /* se could be NULL for root_task_group */
8146 se->cfs_rq = &rq->cfs;
8148 se->cfs_rq = parent->my_q;
8151 update_load_set(&se->load, 0);
8152 se->parent = parent;
8156 #ifdef CONFIG_RT_GROUP_SCHED
8157 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8158 struct sched_rt_entity *rt_se, int cpu,
8159 struct sched_rt_entity *parent)
8161 struct rq *rq = cpu_rq(cpu);
8163 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8164 rt_rq->rt_nr_boosted = 0;
8168 tg->rt_rq[cpu] = rt_rq;
8169 tg->rt_se[cpu] = rt_se;
8175 rt_se->rt_rq = &rq->rt;
8177 rt_se->rt_rq = parent->my_q;
8179 rt_se->my_q = rt_rq;
8180 rt_se->parent = parent;
8181 INIT_LIST_HEAD(&rt_se->run_list);
8185 void __init sched_init(void)
8188 unsigned long alloc_size = 0, ptr;
8190 #ifdef CONFIG_FAIR_GROUP_SCHED
8191 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8193 #ifdef CONFIG_RT_GROUP_SCHED
8194 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8196 #ifdef CONFIG_CPUMASK_OFFSTACK
8197 alloc_size += num_possible_cpus() * cpumask_size();
8200 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8202 #ifdef CONFIG_FAIR_GROUP_SCHED
8203 root_task_group.se = (struct sched_entity **)ptr;
8204 ptr += nr_cpu_ids * sizeof(void **);
8206 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8207 ptr += nr_cpu_ids * sizeof(void **);
8209 #endif /* CONFIG_FAIR_GROUP_SCHED */
8210 #ifdef CONFIG_RT_GROUP_SCHED
8211 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8212 ptr += nr_cpu_ids * sizeof(void **);
8214 root_task_group.rt_rq = (struct rt_rq **)ptr;
8215 ptr += nr_cpu_ids * sizeof(void **);
8217 #endif /* CONFIG_RT_GROUP_SCHED */
8218 #ifdef CONFIG_CPUMASK_OFFSTACK
8219 for_each_possible_cpu(i) {
8220 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8221 ptr += cpumask_size();
8223 #endif /* CONFIG_CPUMASK_OFFSTACK */
8227 init_defrootdomain();
8230 init_rt_bandwidth(&def_rt_bandwidth,
8231 global_rt_period(), global_rt_runtime());
8233 #ifdef CONFIG_RT_GROUP_SCHED
8234 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8235 global_rt_period(), global_rt_runtime());
8236 #endif /* CONFIG_RT_GROUP_SCHED */
8238 #ifdef CONFIG_CGROUP_SCHED
8239 list_add(&root_task_group.list, &task_groups);
8240 INIT_LIST_HEAD(&root_task_group.children);
8241 autogroup_init(&init_task);
8242 #endif /* CONFIG_CGROUP_SCHED */
8244 for_each_possible_cpu(i) {
8248 raw_spin_lock_init(&rq->lock);
8250 rq->calc_load_active = 0;
8251 rq->calc_load_update = jiffies + LOAD_FREQ;
8252 init_cfs_rq(&rq->cfs);
8253 init_rt_rq(&rq->rt, rq);
8254 #ifdef CONFIG_FAIR_GROUP_SCHED
8255 root_task_group.shares = root_task_group_load;
8256 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8258 * How much cpu bandwidth does root_task_group get?
8260 * In case of task-groups formed thr' the cgroup filesystem, it
8261 * gets 100% of the cpu resources in the system. This overall
8262 * system cpu resource is divided among the tasks of
8263 * root_task_group and its child task-groups in a fair manner,
8264 * based on each entity's (task or task-group's) weight
8265 * (se->load.weight).
8267 * In other words, if root_task_group has 10 tasks of weight
8268 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8269 * then A0's share of the cpu resource is:
8271 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8273 * We achieve this by letting root_task_group's tasks sit
8274 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8276 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
8277 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8278 #endif /* CONFIG_FAIR_GROUP_SCHED */
8280 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8281 #ifdef CONFIG_RT_GROUP_SCHED
8282 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8283 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8286 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8287 rq->cpu_load[j] = 0;
8289 rq->last_load_update_tick = jiffies;
8294 rq->cpu_power = SCHED_POWER_SCALE;
8295 rq->post_schedule = 0;
8296 rq->active_balance = 0;
8297 rq->next_balance = jiffies;
8302 rq->avg_idle = 2*sysctl_sched_migration_cost;
8303 rq_attach_root(rq, &def_root_domain);
8305 rq->nohz_balance_kick = 0;
8306 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8310 atomic_set(&rq->nr_iowait, 0);
8313 set_load_weight(&init_task);
8315 #ifdef CONFIG_PREEMPT_NOTIFIERS
8316 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8320 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8323 #ifdef CONFIG_RT_MUTEXES
8324 plist_head_init(&init_task.pi_waiters);
8328 * The boot idle thread does lazy MMU switching as well:
8330 atomic_inc(&init_mm.mm_count);
8331 enter_lazy_tlb(&init_mm, current);
8334 * Make us the idle thread. Technically, schedule() should not be
8335 * called from this thread, however somewhere below it might be,
8336 * but because we are the idle thread, we just pick up running again
8337 * when this runqueue becomes "idle".
8339 init_idle(current, smp_processor_id());
8341 calc_load_update = jiffies + LOAD_FREQ;
8344 * During early bootup we pretend to be a normal task:
8346 current->sched_class = &fair_sched_class;
8348 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8349 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8351 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8353 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8354 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8355 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8356 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8357 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8359 /* May be allocated at isolcpus cmdline parse time */
8360 if (cpu_isolated_map == NULL)
8361 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8364 scheduler_running = 1;
8367 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8368 static inline int preempt_count_equals(int preempt_offset)
8370 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8372 return (nested == preempt_offset);
8375 void __might_sleep(const char *file, int line, int preempt_offset)
8377 static unsigned long prev_jiffy; /* ratelimiting */
8379 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8380 system_state != SYSTEM_RUNNING || oops_in_progress)
8382 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8384 prev_jiffy = jiffies;
8387 "BUG: sleeping function called from invalid context at %s:%d\n",
8390 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8391 in_atomic(), irqs_disabled(),
8392 current->pid, current->comm);
8394 debug_show_held_locks(current);
8395 if (irqs_disabled())
8396 print_irqtrace_events(current);
8399 EXPORT_SYMBOL(__might_sleep);
8402 #ifdef CONFIG_MAGIC_SYSRQ
8403 static void normalize_task(struct rq *rq, struct task_struct *p)
8405 const struct sched_class *prev_class = p->sched_class;
8406 int old_prio = p->prio;
8411 deactivate_task(rq, p, 0);
8412 __setscheduler(rq, p, SCHED_NORMAL, 0);
8414 activate_task(rq, p, 0);
8415 resched_task(rq->curr);
8418 check_class_changed(rq, p, prev_class, old_prio);
8421 void normalize_rt_tasks(void)
8423 struct task_struct *g, *p;
8424 unsigned long flags;
8427 read_lock_irqsave(&tasklist_lock, flags);
8428 do_each_thread(g, p) {
8430 * Only normalize user tasks:
8435 p->se.exec_start = 0;
8436 #ifdef CONFIG_SCHEDSTATS
8437 p->se.statistics.wait_start = 0;
8438 p->se.statistics.sleep_start = 0;
8439 p->se.statistics.block_start = 0;
8444 * Renice negative nice level userspace
8447 if (TASK_NICE(p) < 0 && p->mm)
8448 set_user_nice(p, 0);
8452 raw_spin_lock(&p->pi_lock);
8453 rq = __task_rq_lock(p);
8455 normalize_task(rq, p);
8457 __task_rq_unlock(rq);
8458 raw_spin_unlock(&p->pi_lock);
8459 } while_each_thread(g, p);
8461 read_unlock_irqrestore(&tasklist_lock, flags);
8464 #endif /* CONFIG_MAGIC_SYSRQ */
8466 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8468 * These functions are only useful for the IA64 MCA handling, or kdb.
8470 * They can only be called when the whole system has been
8471 * stopped - every CPU needs to be quiescent, and no scheduling
8472 * activity can take place. Using them for anything else would
8473 * be a serious bug, and as a result, they aren't even visible
8474 * under any other configuration.
8478 * curr_task - return the current task for a given cpu.
8479 * @cpu: the processor in question.
8481 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8483 struct task_struct *curr_task(int cpu)
8485 return cpu_curr(cpu);
8488 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8492 * set_curr_task - set the current task for a given cpu.
8493 * @cpu: the processor in question.
8494 * @p: the task pointer to set.
8496 * Description: This function must only be used when non-maskable interrupts
8497 * are serviced on a separate stack. It allows the architecture to switch the
8498 * notion of the current task on a cpu in a non-blocking manner. This function
8499 * must be called with all CPU's synchronized, and interrupts disabled, the
8500 * and caller must save the original value of the current task (see
8501 * curr_task() above) and restore that value before reenabling interrupts and
8502 * re-starting the system.
8504 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8506 void set_curr_task(int cpu, struct task_struct *p)
8513 #ifdef CONFIG_FAIR_GROUP_SCHED
8514 static void free_fair_sched_group(struct task_group *tg)
8518 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8520 for_each_possible_cpu(i) {
8522 kfree(tg->cfs_rq[i]);
8532 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8534 struct cfs_rq *cfs_rq;
8535 struct sched_entity *se;
8538 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8541 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8545 tg->shares = NICE_0_LOAD;
8547 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8549 for_each_possible_cpu(i) {
8550 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8551 GFP_KERNEL, cpu_to_node(i));
8555 se = kzalloc_node(sizeof(struct sched_entity),
8556 GFP_KERNEL, cpu_to_node(i));
8560 init_cfs_rq(cfs_rq);
8561 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8572 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8574 struct rq *rq = cpu_rq(cpu);
8575 unsigned long flags;
8578 * Only empty task groups can be destroyed; so we can speculatively
8579 * check on_list without danger of it being re-added.
8581 if (!tg->cfs_rq[cpu]->on_list)
8584 raw_spin_lock_irqsave(&rq->lock, flags);
8585 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8586 raw_spin_unlock_irqrestore(&rq->lock, flags);
8588 #else /* !CONFIG_FAIR_GROUP_SCHED */
8589 static inline void free_fair_sched_group(struct task_group *tg)
8594 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8599 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8602 #endif /* CONFIG_FAIR_GROUP_SCHED */
8604 #ifdef CONFIG_RT_GROUP_SCHED
8605 static void free_rt_sched_group(struct task_group *tg)
8610 destroy_rt_bandwidth(&tg->rt_bandwidth);
8612 for_each_possible_cpu(i) {
8614 kfree(tg->rt_rq[i]);
8616 kfree(tg->rt_se[i]);
8624 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8626 struct rt_rq *rt_rq;
8627 struct sched_rt_entity *rt_se;
8630 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8633 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8637 init_rt_bandwidth(&tg->rt_bandwidth,
8638 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8640 for_each_possible_cpu(i) {
8641 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8642 GFP_KERNEL, cpu_to_node(i));
8646 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8647 GFP_KERNEL, cpu_to_node(i));
8651 init_rt_rq(rt_rq, cpu_rq(i));
8652 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8653 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8663 #else /* !CONFIG_RT_GROUP_SCHED */
8664 static inline void free_rt_sched_group(struct task_group *tg)
8669 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8673 #endif /* CONFIG_RT_GROUP_SCHED */
8675 #ifdef CONFIG_CGROUP_SCHED
8676 static void free_sched_group(struct task_group *tg)
8678 free_fair_sched_group(tg);
8679 free_rt_sched_group(tg);
8684 /* allocate runqueue etc for a new task group */
8685 struct task_group *sched_create_group(struct task_group *parent)
8687 struct task_group *tg;
8688 unsigned long flags;
8690 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8692 return ERR_PTR(-ENOMEM);
8694 if (!alloc_fair_sched_group(tg, parent))
8697 if (!alloc_rt_sched_group(tg, parent))
8700 spin_lock_irqsave(&task_group_lock, flags);
8701 list_add_rcu(&tg->list, &task_groups);
8703 WARN_ON(!parent); /* root should already exist */
8705 tg->parent = parent;
8706 INIT_LIST_HEAD(&tg->children);
8707 list_add_rcu(&tg->siblings, &parent->children);
8708 spin_unlock_irqrestore(&task_group_lock, flags);
8713 free_sched_group(tg);
8714 return ERR_PTR(-ENOMEM);
8717 /* rcu callback to free various structures associated with a task group */
8718 static void free_sched_group_rcu(struct rcu_head *rhp)
8720 /* now it should be safe to free those cfs_rqs */
8721 free_sched_group(container_of(rhp, struct task_group, rcu));
8724 /* Destroy runqueue etc associated with a task group */
8725 void sched_destroy_group(struct task_group *tg)
8727 unsigned long flags;
8730 /* end participation in shares distribution */
8731 for_each_possible_cpu(i)
8732 unregister_fair_sched_group(tg, i);
8734 spin_lock_irqsave(&task_group_lock, flags);
8735 list_del_rcu(&tg->list);
8736 list_del_rcu(&tg->siblings);
8737 spin_unlock_irqrestore(&task_group_lock, flags);
8739 /* wait for possible concurrent references to cfs_rqs complete */
8740 call_rcu(&tg->rcu, free_sched_group_rcu);
8743 /* change task's runqueue when it moves between groups.
8744 * The caller of this function should have put the task in its new group
8745 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8746 * reflect its new group.
8748 void sched_move_task(struct task_struct *tsk)
8751 unsigned long flags;
8754 rq = task_rq_lock(tsk, &flags);
8756 running = task_current(rq, tsk);
8760 dequeue_task(rq, tsk, 0);
8761 if (unlikely(running))
8762 tsk->sched_class->put_prev_task(rq, tsk);
8764 #ifdef CONFIG_FAIR_GROUP_SCHED
8765 if (tsk->sched_class->task_move_group)
8766 tsk->sched_class->task_move_group(tsk, on_rq);
8769 set_task_rq(tsk, task_cpu(tsk));
8771 if (unlikely(running))
8772 tsk->sched_class->set_curr_task(rq);
8774 enqueue_task(rq, tsk, 0);
8776 task_rq_unlock(rq, tsk, &flags);
8778 #endif /* CONFIG_CGROUP_SCHED */
8780 #ifdef CONFIG_FAIR_GROUP_SCHED
8781 static DEFINE_MUTEX(shares_mutex);
8783 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8786 unsigned long flags;
8789 * We can't change the weight of the root cgroup.
8794 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8796 mutex_lock(&shares_mutex);
8797 if (tg->shares == shares)
8800 tg->shares = shares;
8801 for_each_possible_cpu(i) {
8802 struct rq *rq = cpu_rq(i);
8803 struct sched_entity *se;
8806 /* Propagate contribution to hierarchy */
8807 raw_spin_lock_irqsave(&rq->lock, flags);
8808 for_each_sched_entity(se)
8809 update_cfs_shares(group_cfs_rq(se));
8810 raw_spin_unlock_irqrestore(&rq->lock, flags);
8814 mutex_unlock(&shares_mutex);
8818 unsigned long sched_group_shares(struct task_group *tg)
8824 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
8825 static unsigned long to_ratio(u64 period, u64 runtime)
8827 if (runtime == RUNTIME_INF)
8830 return div64_u64(runtime << 20, period);
8834 #ifdef CONFIG_RT_GROUP_SCHED
8836 * Ensure that the real time constraints are schedulable.
8838 static DEFINE_MUTEX(rt_constraints_mutex);
8840 /* Must be called with tasklist_lock held */
8841 static inline int tg_has_rt_tasks(struct task_group *tg)
8843 struct task_struct *g, *p;
8845 do_each_thread(g, p) {
8846 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8848 } while_each_thread(g, p);
8853 struct rt_schedulable_data {
8854 struct task_group *tg;
8859 static int tg_rt_schedulable(struct task_group *tg, void *data)
8861 struct rt_schedulable_data *d = data;
8862 struct task_group *child;
8863 unsigned long total, sum = 0;
8864 u64 period, runtime;
8866 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8867 runtime = tg->rt_bandwidth.rt_runtime;
8870 period = d->rt_period;
8871 runtime = d->rt_runtime;
8875 * Cannot have more runtime than the period.
8877 if (runtime > period && runtime != RUNTIME_INF)
8881 * Ensure we don't starve existing RT tasks.
8883 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8886 total = to_ratio(period, runtime);
8889 * Nobody can have more than the global setting allows.
8891 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8895 * The sum of our children's runtime should not exceed our own.
8897 list_for_each_entry_rcu(child, &tg->children, siblings) {
8898 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8899 runtime = child->rt_bandwidth.rt_runtime;
8901 if (child == d->tg) {
8902 period = d->rt_period;
8903 runtime = d->rt_runtime;
8906 sum += to_ratio(period, runtime);
8915 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8919 struct rt_schedulable_data data = {
8921 .rt_period = period,
8922 .rt_runtime = runtime,
8926 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8932 static int tg_set_rt_bandwidth(struct task_group *tg,
8933 u64 rt_period, u64 rt_runtime)
8937 mutex_lock(&rt_constraints_mutex);
8938 read_lock(&tasklist_lock);
8939 err = __rt_schedulable(tg, rt_period, rt_runtime);
8943 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8944 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8945 tg->rt_bandwidth.rt_runtime = rt_runtime;
8947 for_each_possible_cpu(i) {
8948 struct rt_rq *rt_rq = tg->rt_rq[i];
8950 raw_spin_lock(&rt_rq->rt_runtime_lock);
8951 rt_rq->rt_runtime = rt_runtime;
8952 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8954 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8956 read_unlock(&tasklist_lock);
8957 mutex_unlock(&rt_constraints_mutex);
8962 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8964 u64 rt_runtime, rt_period;
8966 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8967 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8968 if (rt_runtime_us < 0)
8969 rt_runtime = RUNTIME_INF;
8971 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8974 long sched_group_rt_runtime(struct task_group *tg)
8978 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8981 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8982 do_div(rt_runtime_us, NSEC_PER_USEC);
8983 return rt_runtime_us;
8986 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8988 u64 rt_runtime, rt_period;
8990 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8991 rt_runtime = tg->rt_bandwidth.rt_runtime;
8996 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8999 long sched_group_rt_period(struct task_group *tg)
9003 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9004 do_div(rt_period_us, NSEC_PER_USEC);
9005 return rt_period_us;
9008 static int sched_rt_global_constraints(void)
9010 u64 runtime, period;
9013 if (sysctl_sched_rt_period <= 0)
9016 runtime = global_rt_runtime();
9017 period = global_rt_period();
9020 * Sanity check on the sysctl variables.
9022 if (runtime > period && runtime != RUNTIME_INF)
9025 mutex_lock(&rt_constraints_mutex);
9026 read_lock(&tasklist_lock);
9027 ret = __rt_schedulable(NULL, 0, 0);
9028 read_unlock(&tasklist_lock);
9029 mutex_unlock(&rt_constraints_mutex);
9034 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
9036 /* Don't accept realtime tasks when there is no way for them to run */
9037 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
9043 #else /* !CONFIG_RT_GROUP_SCHED */
9044 static int sched_rt_global_constraints(void)
9046 unsigned long flags;
9049 if (sysctl_sched_rt_period <= 0)
9053 * There's always some RT tasks in the root group
9054 * -- migration, kstopmachine etc..
9056 if (sysctl_sched_rt_runtime == 0)
9059 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9060 for_each_possible_cpu(i) {
9061 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9063 raw_spin_lock(&rt_rq->rt_runtime_lock);
9064 rt_rq->rt_runtime = global_rt_runtime();
9065 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9067 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9071 #endif /* CONFIG_RT_GROUP_SCHED */
9073 int sched_rt_handler(struct ctl_table *table, int write,
9074 void __user *buffer, size_t *lenp,
9078 int old_period, old_runtime;
9079 static DEFINE_MUTEX(mutex);
9082 old_period = sysctl_sched_rt_period;
9083 old_runtime = sysctl_sched_rt_runtime;
9085 ret = proc_dointvec(table, write, buffer, lenp, ppos);
9087 if (!ret && write) {
9088 ret = sched_rt_global_constraints();
9090 sysctl_sched_rt_period = old_period;
9091 sysctl_sched_rt_runtime = old_runtime;
9093 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9094 def_rt_bandwidth.rt_period =
9095 ns_to_ktime(global_rt_period());
9098 mutex_unlock(&mutex);
9103 #ifdef CONFIG_CGROUP_SCHED
9105 /* return corresponding task_group object of a cgroup */
9106 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9108 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9109 struct task_group, css);
9112 static struct cgroup_subsys_state *
9113 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9115 struct task_group *tg, *parent;
9117 if (!cgrp->parent) {
9118 /* This is early initialization for the top cgroup */
9119 return &root_task_group.css;
9122 parent = cgroup_tg(cgrp->parent);
9123 tg = sched_create_group(parent);
9125 return ERR_PTR(-ENOMEM);
9131 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9133 struct task_group *tg = cgroup_tg(cgrp);
9135 sched_destroy_group(tg);
9139 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9141 #ifdef CONFIG_RT_GROUP_SCHED
9142 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9145 /* We don't support RT-tasks being in separate groups */
9146 if (tsk->sched_class != &fair_sched_class)
9153 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9155 sched_move_task(tsk);
9159 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
9160 struct cgroup *old_cgrp, struct task_struct *task)
9163 * cgroup_exit() is called in the copy_process() failure path.
9164 * Ignore this case since the task hasn't ran yet, this avoids
9165 * trying to poke a half freed task state from generic code.
9167 if (!(task->flags & PF_EXITING))
9170 sched_move_task(task);
9173 #ifdef CONFIG_FAIR_GROUP_SCHED
9174 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9177 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
9180 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9182 struct task_group *tg = cgroup_tg(cgrp);
9184 return (u64) scale_load_down(tg->shares);
9187 #ifdef CONFIG_CFS_BANDWIDTH
9188 static DEFINE_MUTEX(cfs_constraints_mutex);
9190 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
9191 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
9193 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
9195 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
9197 int i, ret = 0, runtime_enabled;
9198 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9200 if (tg == &root_task_group)
9204 * Ensure we have at some amount of bandwidth every period. This is
9205 * to prevent reaching a state of large arrears when throttled via
9206 * entity_tick() resulting in prolonged exit starvation.
9208 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
9212 * Likewise, bound things on the otherside by preventing insane quota
9213 * periods. This also allows us to normalize in computing quota
9216 if (period > max_cfs_quota_period)
9219 mutex_lock(&cfs_constraints_mutex);
9220 ret = __cfs_schedulable(tg, period, quota);
9224 runtime_enabled = quota != RUNTIME_INF;
9225 raw_spin_lock_irq(&cfs_b->lock);
9226 cfs_b->period = ns_to_ktime(period);
9227 cfs_b->quota = quota;
9229 __refill_cfs_bandwidth_runtime(cfs_b);
9230 /* restart the period timer (if active) to handle new period expiry */
9231 if (runtime_enabled && cfs_b->timer_active) {
9232 /* force a reprogram */
9233 cfs_b->timer_active = 0;
9234 __start_cfs_bandwidth(cfs_b);
9236 raw_spin_unlock_irq(&cfs_b->lock);
9238 for_each_possible_cpu(i) {
9239 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
9240 struct rq *rq = rq_of(cfs_rq);
9242 raw_spin_lock_irq(&rq->lock);
9243 cfs_rq->runtime_enabled = runtime_enabled;
9244 cfs_rq->runtime_remaining = 0;
9246 if (cfs_rq_throttled(cfs_rq))
9247 unthrottle_cfs_rq(cfs_rq);
9248 raw_spin_unlock_irq(&rq->lock);
9251 mutex_unlock(&cfs_constraints_mutex);
9256 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
9260 period = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9261 if (cfs_quota_us < 0)
9262 quota = RUNTIME_INF;
9264 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
9266 return tg_set_cfs_bandwidth(tg, period, quota);
9269 long tg_get_cfs_quota(struct task_group *tg)
9273 if (tg_cfs_bandwidth(tg)->quota == RUNTIME_INF)
9276 quota_us = tg_cfs_bandwidth(tg)->quota;
9277 do_div(quota_us, NSEC_PER_USEC);
9282 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
9286 period = (u64)cfs_period_us * NSEC_PER_USEC;
9287 quota = tg_cfs_bandwidth(tg)->quota;
9292 return tg_set_cfs_bandwidth(tg, period, quota);
9295 long tg_get_cfs_period(struct task_group *tg)
9299 cfs_period_us = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9300 do_div(cfs_period_us, NSEC_PER_USEC);
9302 return cfs_period_us;
9305 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
9307 return tg_get_cfs_quota(cgroup_tg(cgrp));
9310 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
9313 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
9316 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
9318 return tg_get_cfs_period(cgroup_tg(cgrp));
9321 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9324 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
9327 struct cfs_schedulable_data {
9328 struct task_group *tg;
9333 * normalize group quota/period to be quota/max_period
9334 * note: units are usecs
9336 static u64 normalize_cfs_quota(struct task_group *tg,
9337 struct cfs_schedulable_data *d)
9345 period = tg_get_cfs_period(tg);
9346 quota = tg_get_cfs_quota(tg);
9349 /* note: these should typically be equivalent */
9350 if (quota == RUNTIME_INF || quota == -1)
9353 return to_ratio(period, quota);
9356 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
9358 struct cfs_schedulable_data *d = data;
9359 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9360 s64 quota = 0, parent_quota = -1;
9363 quota = RUNTIME_INF;
9365 struct cfs_bandwidth *parent_b = tg_cfs_bandwidth(tg->parent);
9367 quota = normalize_cfs_quota(tg, d);
9368 parent_quota = parent_b->hierarchal_quota;
9371 * ensure max(child_quota) <= parent_quota, inherit when no
9374 if (quota == RUNTIME_INF)
9375 quota = parent_quota;
9376 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
9379 cfs_b->hierarchal_quota = quota;
9384 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
9387 struct cfs_schedulable_data data = {
9393 if (quota != RUNTIME_INF) {
9394 do_div(data.period, NSEC_PER_USEC);
9395 do_div(data.quota, NSEC_PER_USEC);
9399 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
9405 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
9406 struct cgroup_map_cb *cb)
9408 struct task_group *tg = cgroup_tg(cgrp);
9409 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9411 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
9412 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
9413 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
9417 #endif /* CONFIG_CFS_BANDWIDTH */
9418 #endif /* CONFIG_FAIR_GROUP_SCHED */
9420 #ifdef CONFIG_RT_GROUP_SCHED
9421 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9424 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9427 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9429 return sched_group_rt_runtime(cgroup_tg(cgrp));
9432 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9435 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9438 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9440 return sched_group_rt_period(cgroup_tg(cgrp));
9442 #endif /* CONFIG_RT_GROUP_SCHED */
9444 static struct cftype cpu_files[] = {
9445 #ifdef CONFIG_FAIR_GROUP_SCHED
9448 .read_u64 = cpu_shares_read_u64,
9449 .write_u64 = cpu_shares_write_u64,
9452 #ifdef CONFIG_CFS_BANDWIDTH
9454 .name = "cfs_quota_us",
9455 .read_s64 = cpu_cfs_quota_read_s64,
9456 .write_s64 = cpu_cfs_quota_write_s64,
9459 .name = "cfs_period_us",
9460 .read_u64 = cpu_cfs_period_read_u64,
9461 .write_u64 = cpu_cfs_period_write_u64,
9465 .read_map = cpu_stats_show,
9468 #ifdef CONFIG_RT_GROUP_SCHED
9470 .name = "rt_runtime_us",
9471 .read_s64 = cpu_rt_runtime_read,
9472 .write_s64 = cpu_rt_runtime_write,
9475 .name = "rt_period_us",
9476 .read_u64 = cpu_rt_period_read_uint,
9477 .write_u64 = cpu_rt_period_write_uint,
9482 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9484 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9487 struct cgroup_subsys cpu_cgroup_subsys = {
9489 .create = cpu_cgroup_create,
9490 .destroy = cpu_cgroup_destroy,
9491 .can_attach_task = cpu_cgroup_can_attach_task,
9492 .attach_task = cpu_cgroup_attach_task,
9493 .exit = cpu_cgroup_exit,
9494 .populate = cpu_cgroup_populate,
9495 .subsys_id = cpu_cgroup_subsys_id,
9499 #endif /* CONFIG_CGROUP_SCHED */
9501 #ifdef CONFIG_CGROUP_CPUACCT
9504 * CPU accounting code for task groups.
9506 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9507 * (balbir@in.ibm.com).
9510 /* track cpu usage of a group of tasks and its child groups */
9512 struct cgroup_subsys_state css;
9513 /* cpuusage holds pointer to a u64-type object on every cpu */
9514 u64 __percpu *cpuusage;
9515 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9516 struct cpuacct *parent;
9519 struct cgroup_subsys cpuacct_subsys;
9521 /* return cpu accounting group corresponding to this container */
9522 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9524 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9525 struct cpuacct, css);
9528 /* return cpu accounting group to which this task belongs */
9529 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9531 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9532 struct cpuacct, css);
9535 /* create a new cpu accounting group */
9536 static struct cgroup_subsys_state *cpuacct_create(
9537 struct cgroup_subsys *ss, struct cgroup *cgrp)
9539 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9545 ca->cpuusage = alloc_percpu(u64);
9549 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9550 if (percpu_counter_init(&ca->cpustat[i], 0))
9551 goto out_free_counters;
9554 ca->parent = cgroup_ca(cgrp->parent);
9560 percpu_counter_destroy(&ca->cpustat[i]);
9561 free_percpu(ca->cpuusage);
9565 return ERR_PTR(-ENOMEM);
9568 /* destroy an existing cpu accounting group */
9570 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9572 struct cpuacct *ca = cgroup_ca(cgrp);
9575 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9576 percpu_counter_destroy(&ca->cpustat[i]);
9577 free_percpu(ca->cpuusage);
9581 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9583 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9586 #ifndef CONFIG_64BIT
9588 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9590 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9592 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9600 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9602 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9604 #ifndef CONFIG_64BIT
9606 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9608 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9610 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9616 /* return total cpu usage (in nanoseconds) of a group */
9617 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9619 struct cpuacct *ca = cgroup_ca(cgrp);
9620 u64 totalcpuusage = 0;
9623 for_each_present_cpu(i)
9624 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9626 return totalcpuusage;
9629 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9632 struct cpuacct *ca = cgroup_ca(cgrp);
9641 for_each_present_cpu(i)
9642 cpuacct_cpuusage_write(ca, i, 0);
9648 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9651 struct cpuacct *ca = cgroup_ca(cgroup);
9655 for_each_present_cpu(i) {
9656 percpu = cpuacct_cpuusage_read(ca, i);
9657 seq_printf(m, "%llu ", (unsigned long long) percpu);
9659 seq_printf(m, "\n");
9663 static const char *cpuacct_stat_desc[] = {
9664 [CPUACCT_STAT_USER] = "user",
9665 [CPUACCT_STAT_SYSTEM] = "system",
9668 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9669 struct cgroup_map_cb *cb)
9671 struct cpuacct *ca = cgroup_ca(cgrp);
9674 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9675 s64 val = percpu_counter_read(&ca->cpustat[i]);
9676 val = cputime64_to_clock_t(val);
9677 cb->fill(cb, cpuacct_stat_desc[i], val);
9682 static struct cftype files[] = {
9685 .read_u64 = cpuusage_read,
9686 .write_u64 = cpuusage_write,
9689 .name = "usage_percpu",
9690 .read_seq_string = cpuacct_percpu_seq_read,
9694 .read_map = cpuacct_stats_show,
9698 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9700 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9704 * charge this task's execution time to its accounting group.
9706 * called with rq->lock held.
9708 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9713 if (unlikely(!cpuacct_subsys.active))
9716 cpu = task_cpu(tsk);
9722 for (; ca; ca = ca->parent) {
9723 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9724 *cpuusage += cputime;
9731 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9732 * in cputime_t units. As a result, cpuacct_update_stats calls
9733 * percpu_counter_add with values large enough to always overflow the
9734 * per cpu batch limit causing bad SMP scalability.
9736 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9737 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9738 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9741 #define CPUACCT_BATCH \
9742 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9744 #define CPUACCT_BATCH 0
9748 * Charge the system/user time to the task's accounting group.
9750 static void cpuacct_update_stats(struct task_struct *tsk,
9751 enum cpuacct_stat_index idx, cputime_t val)
9754 int batch = CPUACCT_BATCH;
9756 if (unlikely(!cpuacct_subsys.active))
9763 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9769 struct cgroup_subsys cpuacct_subsys = {
9771 .create = cpuacct_create,
9772 .destroy = cpuacct_destroy,
9773 .populate = cpuacct_populate,
9774 .subsys_id = cpuacct_subsys_id,
9776 #endif /* CONFIG_CGROUP_CPUACCT */