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/nmi.h>
31 #include <linux/init.h>
32 #include <linux/uaccess.h>
33 #include <linux/highmem.h>
34 #include <asm/mmu_context.h>
35 #include <linux/interrupt.h>
36 #include <linux/capability.h>
37 #include <linux/completion.h>
38 #include <linux/kernel_stat.h>
39 #include <linux/debug_locks.h>
40 #include <linux/perf_event.h>
41 #include <linux/security.h>
42 #include <linux/notifier.h>
43 #include <linux/profile.h>
44 #include <linux/freezer.h>
45 #include <linux/vmalloc.h>
46 #include <linux/blkdev.h>
47 #include <linux/module.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 llist_head 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 * successfully 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_pending(void)
2703 struct rq *rq = this_rq();
2704 struct llist_node *llist = llist_del_all(&rq->wake_list);
2705 struct task_struct *p;
2707 raw_spin_lock(&rq->lock);
2710 p = llist_entry(llist, struct task_struct, wake_entry);
2711 llist = llist_next(llist);
2712 ttwu_do_activate(rq, p, 0);
2715 raw_spin_unlock(&rq->lock);
2718 void scheduler_ipi(void)
2720 if (llist_empty(&this_rq()->wake_list))
2724 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2725 * traditionally all their work was done from the interrupt return
2726 * path. Now that we actually do some work, we need to make sure
2729 * Some archs already do call them, luckily irq_enter/exit nest
2732 * Arguably we should visit all archs and update all handlers,
2733 * however a fair share of IPIs are still resched only so this would
2734 * somewhat pessimize the simple resched case.
2737 sched_ttwu_pending();
2741 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2743 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
2744 smp_send_reschedule(cpu);
2747 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2748 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2753 rq = __task_rq_lock(p);
2755 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2756 ttwu_do_wakeup(rq, p, wake_flags);
2759 __task_rq_unlock(rq);
2764 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2765 #endif /* CONFIG_SMP */
2767 static void ttwu_queue(struct task_struct *p, int cpu)
2769 struct rq *rq = cpu_rq(cpu);
2771 #if defined(CONFIG_SMP)
2772 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2773 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2774 ttwu_queue_remote(p, cpu);
2779 raw_spin_lock(&rq->lock);
2780 ttwu_do_activate(rq, p, 0);
2781 raw_spin_unlock(&rq->lock);
2785 * try_to_wake_up - wake up a thread
2786 * @p: the thread to be awakened
2787 * @state: the mask of task states that can be woken
2788 * @wake_flags: wake modifier flags (WF_*)
2790 * Put it on the run-queue if it's not already there. The "current"
2791 * thread is always on the run-queue (except when the actual
2792 * re-schedule is in progress), and as such you're allowed to do
2793 * the simpler "current->state = TASK_RUNNING" to mark yourself
2794 * runnable without the overhead of this.
2796 * Returns %true if @p was woken up, %false if it was already running
2797 * or @state didn't match @p's state.
2800 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2802 unsigned long flags;
2803 int cpu, success = 0;
2806 raw_spin_lock_irqsave(&p->pi_lock, flags);
2807 if (!(p->state & state))
2810 success = 1; /* we're going to change ->state */
2813 if (p->on_rq && ttwu_remote(p, wake_flags))
2818 * If the owning (remote) cpu is still in the middle of schedule() with
2819 * this task as prev, wait until its done referencing the task.
2822 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2824 * In case the architecture enables interrupts in
2825 * context_switch(), we cannot busy wait, since that
2826 * would lead to deadlocks when an interrupt hits and
2827 * tries to wake up @prev. So bail and do a complete
2830 if (ttwu_activate_remote(p, wake_flags))
2837 * Pairs with the smp_wmb() in finish_lock_switch().
2841 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2842 p->state = TASK_WAKING;
2844 if (p->sched_class->task_waking)
2845 p->sched_class->task_waking(p);
2847 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2848 if (task_cpu(p) != cpu) {
2849 wake_flags |= WF_MIGRATED;
2850 set_task_cpu(p, cpu);
2852 #endif /* CONFIG_SMP */
2856 ttwu_stat(p, cpu, wake_flags);
2858 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2864 * try_to_wake_up_local - try to wake up a local task with rq lock held
2865 * @p: the thread to be awakened
2867 * Put @p on the run-queue if it's not already there. The caller must
2868 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2871 static void try_to_wake_up_local(struct task_struct *p)
2873 struct rq *rq = task_rq(p);
2875 BUG_ON(rq != this_rq());
2876 BUG_ON(p == current);
2877 lockdep_assert_held(&rq->lock);
2879 if (!raw_spin_trylock(&p->pi_lock)) {
2880 raw_spin_unlock(&rq->lock);
2881 raw_spin_lock(&p->pi_lock);
2882 raw_spin_lock(&rq->lock);
2885 if (!(p->state & TASK_NORMAL))
2889 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2891 ttwu_do_wakeup(rq, p, 0);
2892 ttwu_stat(p, smp_processor_id(), 0);
2894 raw_spin_unlock(&p->pi_lock);
2898 * wake_up_process - Wake up a specific process
2899 * @p: The process to be woken up.
2901 * Attempt to wake up the nominated process and move it to the set of runnable
2902 * processes. Returns 1 if the process was woken up, 0 if it was already
2905 * It may be assumed that this function implies a write memory barrier before
2906 * changing the task state if and only if any tasks are woken up.
2908 int wake_up_process(struct task_struct *p)
2910 return try_to_wake_up(p, TASK_ALL, 0);
2912 EXPORT_SYMBOL(wake_up_process);
2914 int wake_up_state(struct task_struct *p, unsigned int state)
2916 return try_to_wake_up(p, state, 0);
2920 * Perform scheduler related setup for a newly forked process p.
2921 * p is forked by current.
2923 * __sched_fork() is basic setup used by init_idle() too:
2925 static void __sched_fork(struct task_struct *p)
2930 p->se.exec_start = 0;
2931 p->se.sum_exec_runtime = 0;
2932 p->se.prev_sum_exec_runtime = 0;
2933 p->se.nr_migrations = 0;
2935 INIT_LIST_HEAD(&p->se.group_node);
2937 #ifdef CONFIG_SCHEDSTATS
2938 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2941 INIT_LIST_HEAD(&p->rt.run_list);
2943 #ifdef CONFIG_PREEMPT_NOTIFIERS
2944 INIT_HLIST_HEAD(&p->preempt_notifiers);
2949 * fork()/clone()-time setup:
2951 void sched_fork(struct task_struct *p)
2953 unsigned long flags;
2954 int cpu = get_cpu();
2958 * We mark the process as running here. This guarantees that
2959 * nobody will actually run it, and a signal or other external
2960 * event cannot wake it up and insert it on the runqueue either.
2962 p->state = TASK_RUNNING;
2965 * Make sure we do not leak PI boosting priority to the child.
2967 p->prio = current->normal_prio;
2970 * Revert to default priority/policy on fork if requested.
2972 if (unlikely(p->sched_reset_on_fork)) {
2973 if (task_has_rt_policy(p)) {
2974 p->policy = SCHED_NORMAL;
2975 p->static_prio = NICE_TO_PRIO(0);
2977 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2978 p->static_prio = NICE_TO_PRIO(0);
2980 p->prio = p->normal_prio = __normal_prio(p);
2984 * We don't need the reset flag anymore after the fork. It has
2985 * fulfilled its duty:
2987 p->sched_reset_on_fork = 0;
2990 if (!rt_prio(p->prio))
2991 p->sched_class = &fair_sched_class;
2993 if (p->sched_class->task_fork)
2994 p->sched_class->task_fork(p);
2997 * The child is not yet in the pid-hash so no cgroup attach races,
2998 * and the cgroup is pinned to this child due to cgroup_fork()
2999 * is ran before sched_fork().
3001 * Silence PROVE_RCU.
3003 raw_spin_lock_irqsave(&p->pi_lock, flags);
3004 set_task_cpu(p, cpu);
3005 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3007 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
3008 if (likely(sched_info_on()))
3009 memset(&p->sched_info, 0, sizeof(p->sched_info));
3011 #if defined(CONFIG_SMP)
3014 #ifdef CONFIG_PREEMPT_COUNT
3015 /* Want to start with kernel preemption disabled. */
3016 task_thread_info(p)->preempt_count = 1;
3019 plist_node_init(&p->pushable_tasks, MAX_PRIO);
3026 * wake_up_new_task - wake up a newly created task for the first time.
3028 * This function will do some initial scheduler statistics housekeeping
3029 * that must be done for every newly created context, then puts the task
3030 * on the runqueue and wakes it.
3032 void wake_up_new_task(struct task_struct *p)
3034 unsigned long flags;
3037 raw_spin_lock_irqsave(&p->pi_lock, flags);
3040 * Fork balancing, do it here and not earlier because:
3041 * - cpus_allowed can change in the fork path
3042 * - any previously selected cpu might disappear through hotplug
3044 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
3047 rq = __task_rq_lock(p);
3048 activate_task(rq, p, 0);
3050 trace_sched_wakeup_new(p, true);
3051 check_preempt_curr(rq, p, WF_FORK);
3053 if (p->sched_class->task_woken)
3054 p->sched_class->task_woken(rq, p);
3056 task_rq_unlock(rq, p, &flags);
3059 #ifdef CONFIG_PREEMPT_NOTIFIERS
3062 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3063 * @notifier: notifier struct to register
3065 void preempt_notifier_register(struct preempt_notifier *notifier)
3067 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3069 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3072 * preempt_notifier_unregister - no longer interested in preemption notifications
3073 * @notifier: notifier struct to unregister
3075 * This is safe to call from within a preemption notifier.
3077 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3079 hlist_del(¬ifier->link);
3081 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3083 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3085 struct preempt_notifier *notifier;
3086 struct hlist_node *node;
3088 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3089 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3093 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3094 struct task_struct *next)
3096 struct preempt_notifier *notifier;
3097 struct hlist_node *node;
3099 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3100 notifier->ops->sched_out(notifier, next);
3103 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3105 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3110 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3111 struct task_struct *next)
3115 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3118 * prepare_task_switch - prepare to switch tasks
3119 * @rq: the runqueue preparing to switch
3120 * @prev: the current task that is being switched out
3121 * @next: the task we are going to switch to.
3123 * This is called with the rq lock held and interrupts off. It must
3124 * be paired with a subsequent finish_task_switch after the context
3127 * prepare_task_switch sets up locking and calls architecture specific
3131 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3132 struct task_struct *next)
3134 sched_info_switch(prev, next);
3135 perf_event_task_sched_out(prev, next);
3136 fire_sched_out_preempt_notifiers(prev, next);
3137 prepare_lock_switch(rq, next);
3138 prepare_arch_switch(next);
3139 trace_sched_switch(prev, next);
3143 * finish_task_switch - clean up after a task-switch
3144 * @rq: runqueue associated with task-switch
3145 * @prev: the thread we just switched away from.
3147 * finish_task_switch must be called after the context switch, paired
3148 * with a prepare_task_switch call before the context switch.
3149 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3150 * and do any other architecture-specific cleanup actions.
3152 * Note that we may have delayed dropping an mm in context_switch(). If
3153 * so, we finish that here outside of the runqueue lock. (Doing it
3154 * with the lock held can cause deadlocks; see schedule() for
3157 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3158 __releases(rq->lock)
3160 struct mm_struct *mm = rq->prev_mm;
3166 * A task struct has one reference for the use as "current".
3167 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3168 * schedule one last time. The schedule call will never return, and
3169 * the scheduled task must drop that reference.
3170 * The test for TASK_DEAD must occur while the runqueue locks are
3171 * still held, otherwise prev could be scheduled on another cpu, die
3172 * there before we look at prev->state, and then the reference would
3174 * Manfred Spraul <manfred@colorfullife.com>
3176 prev_state = prev->state;
3177 finish_arch_switch(prev);
3178 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3179 local_irq_disable();
3180 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3181 perf_event_task_sched_in(prev, current);
3182 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3184 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3185 finish_lock_switch(rq, prev);
3187 fire_sched_in_preempt_notifiers(current);
3190 if (unlikely(prev_state == TASK_DEAD)) {
3192 * Remove function-return probe instances associated with this
3193 * task and put them back on the free list.
3195 kprobe_flush_task(prev);
3196 put_task_struct(prev);
3202 /* assumes rq->lock is held */
3203 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3205 if (prev->sched_class->pre_schedule)
3206 prev->sched_class->pre_schedule(rq, prev);
3209 /* rq->lock is NOT held, but preemption is disabled */
3210 static inline void post_schedule(struct rq *rq)
3212 if (rq->post_schedule) {
3213 unsigned long flags;
3215 raw_spin_lock_irqsave(&rq->lock, flags);
3216 if (rq->curr->sched_class->post_schedule)
3217 rq->curr->sched_class->post_schedule(rq);
3218 raw_spin_unlock_irqrestore(&rq->lock, flags);
3220 rq->post_schedule = 0;
3226 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3230 static inline void post_schedule(struct rq *rq)
3237 * schedule_tail - first thing a freshly forked thread must call.
3238 * @prev: the thread we just switched away from.
3240 asmlinkage void schedule_tail(struct task_struct *prev)
3241 __releases(rq->lock)
3243 struct rq *rq = this_rq();
3245 finish_task_switch(rq, prev);
3248 * FIXME: do we need to worry about rq being invalidated by the
3253 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3254 /* In this case, finish_task_switch does not reenable preemption */
3257 if (current->set_child_tid)
3258 put_user(task_pid_vnr(current), current->set_child_tid);
3262 * context_switch - switch to the new MM and the new
3263 * thread's register state.
3266 context_switch(struct rq *rq, struct task_struct *prev,
3267 struct task_struct *next)
3269 struct mm_struct *mm, *oldmm;
3271 prepare_task_switch(rq, prev, next);
3274 oldmm = prev->active_mm;
3276 * For paravirt, this is coupled with an exit in switch_to to
3277 * combine the page table reload and the switch backend into
3280 arch_start_context_switch(prev);
3283 next->active_mm = oldmm;
3284 atomic_inc(&oldmm->mm_count);
3285 enter_lazy_tlb(oldmm, next);
3287 switch_mm(oldmm, mm, next);
3290 prev->active_mm = NULL;
3291 rq->prev_mm = oldmm;
3294 * Since the runqueue lock will be released by the next
3295 * task (which is an invalid locking op but in the case
3296 * of the scheduler it's an obvious special-case), so we
3297 * do an early lockdep release here:
3299 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3300 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3303 /* Here we just switch the register state and the stack. */
3304 switch_to(prev, next, prev);
3308 * this_rq must be evaluated again because prev may have moved
3309 * CPUs since it called schedule(), thus the 'rq' on its stack
3310 * frame will be invalid.
3312 finish_task_switch(this_rq(), prev);
3316 * nr_running, nr_uninterruptible and nr_context_switches:
3318 * externally visible scheduler statistics: current number of runnable
3319 * threads, current number of uninterruptible-sleeping threads, total
3320 * number of context switches performed since bootup.
3322 unsigned long nr_running(void)
3324 unsigned long i, sum = 0;
3326 for_each_online_cpu(i)
3327 sum += cpu_rq(i)->nr_running;
3332 unsigned long nr_uninterruptible(void)
3334 unsigned long i, sum = 0;
3336 for_each_possible_cpu(i)
3337 sum += cpu_rq(i)->nr_uninterruptible;
3340 * Since we read the counters lockless, it might be slightly
3341 * inaccurate. Do not allow it to go below zero though:
3343 if (unlikely((long)sum < 0))
3349 unsigned long long nr_context_switches(void)
3352 unsigned long long sum = 0;
3354 for_each_possible_cpu(i)
3355 sum += cpu_rq(i)->nr_switches;
3360 unsigned long nr_iowait(void)
3362 unsigned long i, sum = 0;
3364 for_each_possible_cpu(i)
3365 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3370 unsigned long nr_iowait_cpu(int cpu)
3372 struct rq *this = cpu_rq(cpu);
3373 return atomic_read(&this->nr_iowait);
3376 unsigned long this_cpu_load(void)
3378 struct rq *this = this_rq();
3379 return this->cpu_load[0];
3383 /* Variables and functions for calc_load */
3384 static atomic_long_t calc_load_tasks;
3385 static unsigned long calc_load_update;
3386 unsigned long avenrun[3];
3387 EXPORT_SYMBOL(avenrun);
3389 static long calc_load_fold_active(struct rq *this_rq)
3391 long nr_active, delta = 0;
3393 nr_active = this_rq->nr_running;
3394 nr_active += (long) this_rq->nr_uninterruptible;
3396 if (nr_active != this_rq->calc_load_active) {
3397 delta = nr_active - this_rq->calc_load_active;
3398 this_rq->calc_load_active = nr_active;
3404 static unsigned long
3405 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3408 load += active * (FIXED_1 - exp);
3409 load += 1UL << (FSHIFT - 1);
3410 return load >> FSHIFT;
3415 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3417 * When making the ILB scale, we should try to pull this in as well.
3419 static atomic_long_t calc_load_tasks_idle;
3421 static void calc_load_account_idle(struct rq *this_rq)
3425 delta = calc_load_fold_active(this_rq);
3427 atomic_long_add(delta, &calc_load_tasks_idle);
3430 static long calc_load_fold_idle(void)
3435 * Its got a race, we don't care...
3437 if (atomic_long_read(&calc_load_tasks_idle))
3438 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3444 * fixed_power_int - compute: x^n, in O(log n) time
3446 * @x: base of the power
3447 * @frac_bits: fractional bits of @x
3448 * @n: power to raise @x to.
3450 * By exploiting the relation between the definition of the natural power
3451 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3452 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3453 * (where: n_i \elem {0, 1}, the binary vector representing n),
3454 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3455 * of course trivially computable in O(log_2 n), the length of our binary
3458 static unsigned long
3459 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3461 unsigned long result = 1UL << frac_bits;
3466 result += 1UL << (frac_bits - 1);
3467 result >>= frac_bits;
3473 x += 1UL << (frac_bits - 1);
3481 * a1 = a0 * e + a * (1 - e)
3483 * a2 = a1 * e + a * (1 - e)
3484 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3485 * = a0 * e^2 + a * (1 - e) * (1 + e)
3487 * a3 = a2 * e + a * (1 - e)
3488 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3489 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3493 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3494 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3495 * = a0 * e^n + a * (1 - e^n)
3497 * [1] application of the geometric series:
3500 * S_n := \Sum x^i = -------------
3503 static unsigned long
3504 calc_load_n(unsigned long load, unsigned long exp,
3505 unsigned long active, unsigned int n)
3508 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3512 * NO_HZ can leave us missing all per-cpu ticks calling
3513 * calc_load_account_active(), but since an idle CPU folds its delta into
3514 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3515 * in the pending idle delta if our idle period crossed a load cycle boundary.
3517 * Once we've updated the global active value, we need to apply the exponential
3518 * weights adjusted to the number of cycles missed.
3520 static void calc_global_nohz(unsigned long ticks)
3522 long delta, active, n;
3524 if (time_before(jiffies, calc_load_update))
3528 * If we crossed a calc_load_update boundary, make sure to fold
3529 * any pending idle changes, the respective CPUs might have
3530 * missed the tick driven calc_load_account_active() update
3533 delta = calc_load_fold_idle();
3535 atomic_long_add(delta, &calc_load_tasks);
3538 * If we were idle for multiple load cycles, apply them.
3540 if (ticks >= LOAD_FREQ) {
3541 n = ticks / LOAD_FREQ;
3543 active = atomic_long_read(&calc_load_tasks);
3544 active = active > 0 ? active * FIXED_1 : 0;
3546 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3547 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3548 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3550 calc_load_update += n * LOAD_FREQ;
3554 * Its possible the remainder of the above division also crosses
3555 * a LOAD_FREQ period, the regular check in calc_global_load()
3556 * which comes after this will take care of that.
3558 * Consider us being 11 ticks before a cycle completion, and us
3559 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3560 * age us 4 cycles, and the test in calc_global_load() will
3561 * pick up the final one.
3565 static void calc_load_account_idle(struct rq *this_rq)
3569 static inline long calc_load_fold_idle(void)
3574 static void calc_global_nohz(unsigned long ticks)
3580 * get_avenrun - get the load average array
3581 * @loads: pointer to dest load array
3582 * @offset: offset to add
3583 * @shift: shift count to shift the result left
3585 * These values are estimates at best, so no need for locking.
3587 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3589 loads[0] = (avenrun[0] + offset) << shift;
3590 loads[1] = (avenrun[1] + offset) << shift;
3591 loads[2] = (avenrun[2] + offset) << shift;
3595 * calc_load - update the avenrun load estimates 10 ticks after the
3596 * CPUs have updated calc_load_tasks.
3598 void calc_global_load(unsigned long ticks)
3602 calc_global_nohz(ticks);
3604 if (time_before(jiffies, calc_load_update + 10))
3607 active = atomic_long_read(&calc_load_tasks);
3608 active = active > 0 ? active * FIXED_1 : 0;
3610 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3611 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3612 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3614 calc_load_update += LOAD_FREQ;
3618 * Called from update_cpu_load() to periodically update this CPU's
3621 static void calc_load_account_active(struct rq *this_rq)
3625 if (time_before(jiffies, this_rq->calc_load_update))
3628 delta = calc_load_fold_active(this_rq);
3629 delta += calc_load_fold_idle();
3631 atomic_long_add(delta, &calc_load_tasks);
3633 this_rq->calc_load_update += LOAD_FREQ;
3637 * The exact cpuload at various idx values, calculated at every tick would be
3638 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3640 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3641 * on nth tick when cpu may be busy, then we have:
3642 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3643 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3645 * decay_load_missed() below does efficient calculation of
3646 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3647 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3649 * The calculation is approximated on a 128 point scale.
3650 * degrade_zero_ticks is the number of ticks after which load at any
3651 * particular idx is approximated to be zero.
3652 * degrade_factor is a precomputed table, a row for each load idx.
3653 * Each column corresponds to degradation factor for a power of two ticks,
3654 * based on 128 point scale.
3656 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3657 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3659 * With this power of 2 load factors, we can degrade the load n times
3660 * by looking at 1 bits in n and doing as many mult/shift instead of
3661 * n mult/shifts needed by the exact degradation.
3663 #define DEGRADE_SHIFT 7
3664 static const unsigned char
3665 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3666 static const unsigned char
3667 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3668 {0, 0, 0, 0, 0, 0, 0, 0},
3669 {64, 32, 8, 0, 0, 0, 0, 0},
3670 {96, 72, 40, 12, 1, 0, 0},
3671 {112, 98, 75, 43, 15, 1, 0},
3672 {120, 112, 98, 76, 45, 16, 2} };
3675 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3676 * would be when CPU is idle and so we just decay the old load without
3677 * adding any new load.
3679 static unsigned long
3680 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3684 if (!missed_updates)
3687 if (missed_updates >= degrade_zero_ticks[idx])
3691 return load >> missed_updates;
3693 while (missed_updates) {
3694 if (missed_updates % 2)
3695 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3697 missed_updates >>= 1;
3704 * Update rq->cpu_load[] statistics. This function is usually called every
3705 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3706 * every tick. We fix it up based on jiffies.
3708 static void update_cpu_load(struct rq *this_rq)
3710 unsigned long this_load = this_rq->load.weight;
3711 unsigned long curr_jiffies = jiffies;
3712 unsigned long pending_updates;
3715 this_rq->nr_load_updates++;
3717 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3718 if (curr_jiffies == this_rq->last_load_update_tick)
3721 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3722 this_rq->last_load_update_tick = curr_jiffies;
3724 /* Update our load: */
3725 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3726 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3727 unsigned long old_load, new_load;
3729 /* scale is effectively 1 << i now, and >> i divides by scale */
3731 old_load = this_rq->cpu_load[i];
3732 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3733 new_load = this_load;
3735 * Round up the averaging division if load is increasing. This
3736 * prevents us from getting stuck on 9 if the load is 10, for
3739 if (new_load > old_load)
3740 new_load += scale - 1;
3742 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3745 sched_avg_update(this_rq);
3748 static void update_cpu_load_active(struct rq *this_rq)
3750 update_cpu_load(this_rq);
3752 calc_load_account_active(this_rq);
3758 * sched_exec - execve() is a valuable balancing opportunity, because at
3759 * this point the task has the smallest effective memory and cache footprint.
3761 void sched_exec(void)
3763 struct task_struct *p = current;
3764 unsigned long flags;
3767 raw_spin_lock_irqsave(&p->pi_lock, flags);
3768 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3769 if (dest_cpu == smp_processor_id())
3772 if (likely(cpu_active(dest_cpu))) {
3773 struct migration_arg arg = { p, dest_cpu };
3775 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3776 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3780 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3785 DEFINE_PER_CPU(struct kernel_stat, kstat);
3787 EXPORT_PER_CPU_SYMBOL(kstat);
3790 * Return any ns on the sched_clock that have not yet been accounted in
3791 * @p in case that task is currently running.
3793 * Called with task_rq_lock() held on @rq.
3795 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3799 if (task_current(rq, p)) {
3800 update_rq_clock(rq);
3801 ns = rq->clock_task - p->se.exec_start;
3809 unsigned long long task_delta_exec(struct task_struct *p)
3811 unsigned long flags;
3815 rq = task_rq_lock(p, &flags);
3816 ns = do_task_delta_exec(p, rq);
3817 task_rq_unlock(rq, p, &flags);
3823 * Return accounted runtime for the task.
3824 * In case the task is currently running, return the runtime plus current's
3825 * pending runtime that have not been accounted yet.
3827 unsigned long long task_sched_runtime(struct task_struct *p)
3829 unsigned long flags;
3833 rq = task_rq_lock(p, &flags);
3834 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3835 task_rq_unlock(rq, p, &flags);
3841 * Account user cpu time to a process.
3842 * @p: the process that the cpu time gets accounted to
3843 * @cputime: the cpu time spent in user space since the last update
3844 * @cputime_scaled: cputime scaled by cpu frequency
3846 void account_user_time(struct task_struct *p, cputime_t cputime,
3847 cputime_t cputime_scaled)
3849 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3852 /* Add user time to process. */
3853 p->utime = cputime_add(p->utime, cputime);
3854 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3855 account_group_user_time(p, cputime);
3857 /* Add user time to cpustat. */
3858 tmp = cputime_to_cputime64(cputime);
3859 if (TASK_NICE(p) > 0)
3860 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3862 cpustat->user = cputime64_add(cpustat->user, tmp);
3864 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3865 /* Account for user time used */
3866 acct_update_integrals(p);
3870 * Account guest cpu time to a process.
3871 * @p: the process that the cpu time gets accounted to
3872 * @cputime: the cpu time spent in virtual machine since the last update
3873 * @cputime_scaled: cputime scaled by cpu frequency
3875 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3876 cputime_t cputime_scaled)
3879 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3881 tmp = cputime_to_cputime64(cputime);
3883 /* Add guest time to process. */
3884 p->utime = cputime_add(p->utime, cputime);
3885 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3886 account_group_user_time(p, cputime);
3887 p->gtime = cputime_add(p->gtime, cputime);
3889 /* Add guest time to cpustat. */
3890 if (TASK_NICE(p) > 0) {
3891 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3892 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3894 cpustat->user = cputime64_add(cpustat->user, tmp);
3895 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3900 * Account system cpu time to a process and desired cpustat field
3901 * @p: the process that the cpu time gets accounted to
3902 * @cputime: the cpu time spent in kernel space since the last update
3903 * @cputime_scaled: cputime scaled by cpu frequency
3904 * @target_cputime64: pointer to cpustat field that has to be updated
3907 void __account_system_time(struct task_struct *p, cputime_t cputime,
3908 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3910 cputime64_t tmp = cputime_to_cputime64(cputime);
3912 /* Add system time to process. */
3913 p->stime = cputime_add(p->stime, cputime);
3914 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3915 account_group_system_time(p, cputime);
3917 /* Add system time to cpustat. */
3918 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3919 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3921 /* Account for system time used */
3922 acct_update_integrals(p);
3926 * Account system cpu time to a process.
3927 * @p: the process that the cpu time gets accounted to
3928 * @hardirq_offset: the offset to subtract from hardirq_count()
3929 * @cputime: the cpu time spent in kernel space since the last update
3930 * @cputime_scaled: cputime scaled by cpu frequency
3932 void account_system_time(struct task_struct *p, int hardirq_offset,
3933 cputime_t cputime, cputime_t cputime_scaled)
3935 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3936 cputime64_t *target_cputime64;
3938 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3939 account_guest_time(p, cputime, cputime_scaled);
3943 if (hardirq_count() - hardirq_offset)
3944 target_cputime64 = &cpustat->irq;
3945 else if (in_serving_softirq())
3946 target_cputime64 = &cpustat->softirq;
3948 target_cputime64 = &cpustat->system;
3950 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3954 * Account for involuntary wait time.
3955 * @cputime: the cpu time spent in involuntary wait
3957 void account_steal_time(cputime_t cputime)
3959 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3960 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3962 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3966 * Account for idle time.
3967 * @cputime: the cpu time spent in idle wait
3969 void account_idle_time(cputime_t cputime)
3971 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3972 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3973 struct rq *rq = this_rq();
3975 if (atomic_read(&rq->nr_iowait) > 0)
3976 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3978 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3981 static __always_inline bool steal_account_process_tick(void)
3983 #ifdef CONFIG_PARAVIRT
3984 if (static_branch(¶virt_steal_enabled)) {
3987 steal = paravirt_steal_clock(smp_processor_id());
3988 steal -= this_rq()->prev_steal_time;
3990 st = steal_ticks(steal);
3991 this_rq()->prev_steal_time += st * TICK_NSEC;
3993 account_steal_time(st);
4000 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4002 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
4004 * Account a tick to a process and cpustat
4005 * @p: the process that the cpu time gets accounted to
4006 * @user_tick: is the tick from userspace
4007 * @rq: the pointer to rq
4009 * Tick demultiplexing follows the order
4010 * - pending hardirq update
4011 * - pending softirq update
4015 * - check for guest_time
4016 * - else account as system_time
4018 * Check for hardirq is done both for system and user time as there is
4019 * no timer going off while we are on hardirq and hence we may never get an
4020 * opportunity to update it solely in system time.
4021 * p->stime and friends are only updated on system time and not on irq
4022 * softirq as those do not count in task exec_runtime any more.
4024 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4027 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4028 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
4029 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4031 if (steal_account_process_tick())
4034 if (irqtime_account_hi_update()) {
4035 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4036 } else if (irqtime_account_si_update()) {
4037 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4038 } else if (this_cpu_ksoftirqd() == p) {
4040 * ksoftirqd time do not get accounted in cpu_softirq_time.
4041 * So, we have to handle it separately here.
4042 * Also, p->stime needs to be updated for ksoftirqd.
4044 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4046 } else if (user_tick) {
4047 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4048 } else if (p == rq->idle) {
4049 account_idle_time(cputime_one_jiffy);
4050 } else if (p->flags & PF_VCPU) { /* System time or guest time */
4051 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
4053 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4058 static void irqtime_account_idle_ticks(int ticks)
4061 struct rq *rq = this_rq();
4063 for (i = 0; i < ticks; i++)
4064 irqtime_account_process_tick(current, 0, rq);
4066 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
4067 static void irqtime_account_idle_ticks(int ticks) {}
4068 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4070 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
4073 * Account a single tick of cpu time.
4074 * @p: the process that the cpu time gets accounted to
4075 * @user_tick: indicates if the tick is a user or a system tick
4077 void account_process_tick(struct task_struct *p, int user_tick)
4079 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4080 struct rq *rq = this_rq();
4082 if (sched_clock_irqtime) {
4083 irqtime_account_process_tick(p, user_tick, rq);
4087 if (steal_account_process_tick())
4091 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4092 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
4093 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
4096 account_idle_time(cputime_one_jiffy);
4100 * Account multiple ticks of steal time.
4101 * @p: the process from which the cpu time has been stolen
4102 * @ticks: number of stolen ticks
4104 void account_steal_ticks(unsigned long ticks)
4106 account_steal_time(jiffies_to_cputime(ticks));
4110 * Account multiple ticks of idle time.
4111 * @ticks: number of stolen ticks
4113 void account_idle_ticks(unsigned long ticks)
4116 if (sched_clock_irqtime) {
4117 irqtime_account_idle_ticks(ticks);
4121 account_idle_time(jiffies_to_cputime(ticks));
4127 * Use precise platform statistics if available:
4129 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4130 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4136 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4138 struct task_cputime cputime;
4140 thread_group_cputime(p, &cputime);
4142 *ut = cputime.utime;
4143 *st = cputime.stime;
4147 #ifndef nsecs_to_cputime
4148 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
4151 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4153 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
4156 * Use CFS's precise accounting:
4158 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
4164 do_div(temp, total);
4165 utime = (cputime_t)temp;
4170 * Compare with previous values, to keep monotonicity:
4172 p->prev_utime = max(p->prev_utime, utime);
4173 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4175 *ut = p->prev_utime;
4176 *st = p->prev_stime;
4180 * Must be called with siglock held.
4182 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4184 struct signal_struct *sig = p->signal;
4185 struct task_cputime cputime;
4186 cputime_t rtime, utime, total;
4188 thread_group_cputime(p, &cputime);
4190 total = cputime_add(cputime.utime, cputime.stime);
4191 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4196 temp *= cputime.utime;
4197 do_div(temp, total);
4198 utime = (cputime_t)temp;
4202 sig->prev_utime = max(sig->prev_utime, utime);
4203 sig->prev_stime = max(sig->prev_stime,
4204 cputime_sub(rtime, sig->prev_utime));
4206 *ut = sig->prev_utime;
4207 *st = sig->prev_stime;
4212 * This function gets called by the timer code, with HZ frequency.
4213 * We call it with interrupts disabled.
4215 void scheduler_tick(void)
4217 int cpu = smp_processor_id();
4218 struct rq *rq = cpu_rq(cpu);
4219 struct task_struct *curr = rq->curr;
4223 raw_spin_lock(&rq->lock);
4224 update_rq_clock(rq);
4225 update_cpu_load_active(rq);
4226 curr->sched_class->task_tick(rq, curr, 0);
4227 raw_spin_unlock(&rq->lock);
4229 perf_event_task_tick();
4232 rq->idle_at_tick = idle_cpu(cpu);
4233 trigger_load_balance(rq, cpu);
4237 notrace unsigned long get_parent_ip(unsigned long addr)
4239 if (in_lock_functions(addr)) {
4240 addr = CALLER_ADDR2;
4241 if (in_lock_functions(addr))
4242 addr = CALLER_ADDR3;
4247 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4248 defined(CONFIG_PREEMPT_TRACER))
4250 void __kprobes add_preempt_count(int val)
4252 #ifdef CONFIG_DEBUG_PREEMPT
4256 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4259 preempt_count() += val;
4260 #ifdef CONFIG_DEBUG_PREEMPT
4262 * Spinlock count overflowing soon?
4264 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4267 if (preempt_count() == val)
4268 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4270 EXPORT_SYMBOL(add_preempt_count);
4272 void __kprobes sub_preempt_count(int val)
4274 #ifdef CONFIG_DEBUG_PREEMPT
4278 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4281 * Is the spinlock portion underflowing?
4283 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4284 !(preempt_count() & PREEMPT_MASK)))
4288 if (preempt_count() == val)
4289 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4290 preempt_count() -= val;
4292 EXPORT_SYMBOL(sub_preempt_count);
4297 * Print scheduling while atomic bug:
4299 static noinline void __schedule_bug(struct task_struct *prev)
4301 struct pt_regs *regs = get_irq_regs();
4303 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4304 prev->comm, prev->pid, preempt_count());
4306 debug_show_held_locks(prev);
4308 if (irqs_disabled())
4309 print_irqtrace_events(prev);
4318 * Various schedule()-time debugging checks and statistics:
4320 static inline void schedule_debug(struct task_struct *prev)
4323 * Test if we are atomic. Since do_exit() needs to call into
4324 * schedule() atomically, we ignore that path for now.
4325 * Otherwise, whine if we are scheduling when we should not be.
4327 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4328 __schedule_bug(prev);
4331 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4333 schedstat_inc(this_rq(), sched_count);
4336 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4338 if (prev->on_rq || rq->skip_clock_update < 0)
4339 update_rq_clock(rq);
4340 prev->sched_class->put_prev_task(rq, prev);
4344 * Pick up the highest-prio task:
4346 static inline struct task_struct *
4347 pick_next_task(struct rq *rq)
4349 const struct sched_class *class;
4350 struct task_struct *p;
4353 * Optimization: we know that if all tasks are in
4354 * the fair class we can call that function directly:
4356 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
4357 p = fair_sched_class.pick_next_task(rq);
4362 for_each_class(class) {
4363 p = class->pick_next_task(rq);
4368 BUG(); /* the idle class will always have a runnable task */
4372 * __schedule() is the main scheduler function.
4374 static void __sched __schedule(void)
4376 struct task_struct *prev, *next;
4377 unsigned long *switch_count;
4383 cpu = smp_processor_id();
4385 rcu_note_context_switch(cpu);
4388 schedule_debug(prev);
4390 if (sched_feat(HRTICK))
4393 raw_spin_lock_irq(&rq->lock);
4395 switch_count = &prev->nivcsw;
4396 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4397 if (unlikely(signal_pending_state(prev->state, prev))) {
4398 prev->state = TASK_RUNNING;
4400 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4404 * If a worker went to sleep, notify and ask workqueue
4405 * whether it wants to wake up a task to maintain
4408 if (prev->flags & PF_WQ_WORKER) {
4409 struct task_struct *to_wakeup;
4411 to_wakeup = wq_worker_sleeping(prev, cpu);
4413 try_to_wake_up_local(to_wakeup);
4416 switch_count = &prev->nvcsw;
4419 pre_schedule(rq, prev);
4421 if (unlikely(!rq->nr_running))
4422 idle_balance(cpu, rq);
4424 put_prev_task(rq, prev);
4425 next = pick_next_task(rq);
4426 clear_tsk_need_resched(prev);
4427 rq->skip_clock_update = 0;
4429 if (likely(prev != next)) {
4434 context_switch(rq, prev, next); /* unlocks the rq */
4436 * The context switch have flipped the stack from under us
4437 * and restored the local variables which were saved when
4438 * this task called schedule() in the past. prev == current
4439 * is still correct, but it can be moved to another cpu/rq.
4441 cpu = smp_processor_id();
4444 raw_spin_unlock_irq(&rq->lock);
4448 preempt_enable_no_resched();
4453 static inline void sched_submit_work(struct task_struct *tsk)
4458 * If we are going to sleep and we have plugged IO queued,
4459 * make sure to submit it to avoid deadlocks.
4461 if (blk_needs_flush_plug(tsk))
4462 blk_schedule_flush_plug(tsk);
4465 asmlinkage void __sched schedule(void)
4467 struct task_struct *tsk = current;
4469 sched_submit_work(tsk);
4472 EXPORT_SYMBOL(schedule);
4474 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4476 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4478 if (lock->owner != owner)
4482 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4483 * lock->owner still matches owner, if that fails, owner might
4484 * point to free()d memory, if it still matches, the rcu_read_lock()
4485 * ensures the memory stays valid.
4489 return owner->on_cpu;
4493 * Look out! "owner" is an entirely speculative pointer
4494 * access and not reliable.
4496 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4498 if (!sched_feat(OWNER_SPIN))
4502 while (owner_running(lock, owner)) {
4506 arch_mutex_cpu_relax();
4511 * We break out the loop above on need_resched() and when the
4512 * owner changed, which is a sign for heavy contention. Return
4513 * success only when lock->owner is NULL.
4515 return lock->owner == NULL;
4519 #ifdef CONFIG_PREEMPT
4521 * this is the entry point to schedule() from in-kernel preemption
4522 * off of preempt_enable. Kernel preemptions off return from interrupt
4523 * occur there and call schedule directly.
4525 asmlinkage void __sched notrace preempt_schedule(void)
4527 struct thread_info *ti = current_thread_info();
4530 * If there is a non-zero preempt_count or interrupts are disabled,
4531 * we do not want to preempt the current task. Just return..
4533 if (likely(ti->preempt_count || irqs_disabled()))
4537 add_preempt_count_notrace(PREEMPT_ACTIVE);
4539 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4542 * Check again in case we missed a preemption opportunity
4543 * between schedule and now.
4546 } while (need_resched());
4548 EXPORT_SYMBOL(preempt_schedule);
4551 * this is the entry point to schedule() from kernel preemption
4552 * off of irq context.
4553 * Note, that this is called and return with irqs disabled. This will
4554 * protect us against recursive calling from irq.
4556 asmlinkage void __sched preempt_schedule_irq(void)
4558 struct thread_info *ti = current_thread_info();
4560 /* Catch callers which need to be fixed */
4561 BUG_ON(ti->preempt_count || !irqs_disabled());
4564 add_preempt_count(PREEMPT_ACTIVE);
4567 local_irq_disable();
4568 sub_preempt_count(PREEMPT_ACTIVE);
4571 * Check again in case we missed a preemption opportunity
4572 * between schedule and now.
4575 } while (need_resched());
4578 #endif /* CONFIG_PREEMPT */
4580 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4583 return try_to_wake_up(curr->private, mode, wake_flags);
4585 EXPORT_SYMBOL(default_wake_function);
4588 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4589 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4590 * number) then we wake all the non-exclusive tasks and one exclusive task.
4592 * There are circumstances in which we can try to wake a task which has already
4593 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4594 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4596 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4597 int nr_exclusive, int wake_flags, void *key)
4599 wait_queue_t *curr, *next;
4601 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4602 unsigned flags = curr->flags;
4604 if (curr->func(curr, mode, wake_flags, key) &&
4605 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4611 * __wake_up - wake up threads blocked on a waitqueue.
4613 * @mode: which threads
4614 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4615 * @key: is directly passed to the wakeup function
4617 * It may be assumed that this function implies a write memory barrier before
4618 * changing the task state if and only if any tasks are woken up.
4620 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4621 int nr_exclusive, void *key)
4623 unsigned long flags;
4625 spin_lock_irqsave(&q->lock, flags);
4626 __wake_up_common(q, mode, nr_exclusive, 0, key);
4627 spin_unlock_irqrestore(&q->lock, flags);
4629 EXPORT_SYMBOL(__wake_up);
4632 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4634 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4636 __wake_up_common(q, mode, 1, 0, NULL);
4638 EXPORT_SYMBOL_GPL(__wake_up_locked);
4640 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4642 __wake_up_common(q, mode, 1, 0, key);
4644 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4647 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4649 * @mode: which threads
4650 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4651 * @key: opaque value to be passed to wakeup targets
4653 * The sync wakeup differs that the waker knows that it will schedule
4654 * away soon, so while the target thread will be woken up, it will not
4655 * be migrated to another CPU - ie. the two threads are 'synchronized'
4656 * with each other. This can prevent needless bouncing between CPUs.
4658 * On UP it can prevent extra preemption.
4660 * It may be assumed that this function implies a write memory barrier before
4661 * changing the task state if and only if any tasks are woken up.
4663 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4664 int nr_exclusive, void *key)
4666 unsigned long flags;
4667 int wake_flags = WF_SYNC;
4672 if (unlikely(!nr_exclusive))
4675 spin_lock_irqsave(&q->lock, flags);
4676 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4677 spin_unlock_irqrestore(&q->lock, flags);
4679 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4682 * __wake_up_sync - see __wake_up_sync_key()
4684 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4686 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4688 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4691 * complete: - signals a single thread waiting on this completion
4692 * @x: holds the state of this particular completion
4694 * This will wake up a single thread waiting on this completion. Threads will be
4695 * awakened in the same order in which they were queued.
4697 * See also complete_all(), wait_for_completion() and related routines.
4699 * It may be assumed that this function implies a write memory barrier before
4700 * changing the task state if and only if any tasks are woken up.
4702 void complete(struct completion *x)
4704 unsigned long flags;
4706 spin_lock_irqsave(&x->wait.lock, flags);
4708 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4709 spin_unlock_irqrestore(&x->wait.lock, flags);
4711 EXPORT_SYMBOL(complete);
4714 * complete_all: - signals all threads waiting on this completion
4715 * @x: holds the state of this particular completion
4717 * This will wake up all threads waiting on this particular completion event.
4719 * It may be assumed that this function implies a write memory barrier before
4720 * changing the task state if and only if any tasks are woken up.
4722 void complete_all(struct completion *x)
4724 unsigned long flags;
4726 spin_lock_irqsave(&x->wait.lock, flags);
4727 x->done += UINT_MAX/2;
4728 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4729 spin_unlock_irqrestore(&x->wait.lock, flags);
4731 EXPORT_SYMBOL(complete_all);
4733 static inline long __sched
4734 do_wait_for_common(struct completion *x, long timeout, int state)
4737 DECLARE_WAITQUEUE(wait, current);
4739 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4741 if (signal_pending_state(state, current)) {
4742 timeout = -ERESTARTSYS;
4745 __set_current_state(state);
4746 spin_unlock_irq(&x->wait.lock);
4747 timeout = schedule_timeout(timeout);
4748 spin_lock_irq(&x->wait.lock);
4749 } while (!x->done && timeout);
4750 __remove_wait_queue(&x->wait, &wait);
4755 return timeout ?: 1;
4759 wait_for_common(struct completion *x, long timeout, int state)
4763 spin_lock_irq(&x->wait.lock);
4764 timeout = do_wait_for_common(x, timeout, state);
4765 spin_unlock_irq(&x->wait.lock);
4770 * wait_for_completion: - waits for completion of a task
4771 * @x: holds the state of this particular completion
4773 * This waits to be signaled for completion of a specific task. It is NOT
4774 * interruptible and there is no timeout.
4776 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4777 * and interrupt capability. Also see complete().
4779 void __sched wait_for_completion(struct completion *x)
4781 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4783 EXPORT_SYMBOL(wait_for_completion);
4786 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4787 * @x: holds the state of this particular completion
4788 * @timeout: timeout value in jiffies
4790 * This waits for either a completion of a specific task to be signaled or for a
4791 * specified timeout to expire. The timeout is in jiffies. It is not
4794 unsigned long __sched
4795 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4797 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4799 EXPORT_SYMBOL(wait_for_completion_timeout);
4802 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4803 * @x: holds the state of this particular completion
4805 * This waits for completion of a specific task to be signaled. It is
4808 int __sched wait_for_completion_interruptible(struct completion *x)
4810 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4811 if (t == -ERESTARTSYS)
4815 EXPORT_SYMBOL(wait_for_completion_interruptible);
4818 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4819 * @x: holds the state of this particular completion
4820 * @timeout: timeout value in jiffies
4822 * This waits for either a completion of a specific task to be signaled or for a
4823 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4826 wait_for_completion_interruptible_timeout(struct completion *x,
4827 unsigned long timeout)
4829 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4831 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4834 * wait_for_completion_killable: - waits for completion of a task (killable)
4835 * @x: holds the state of this particular completion
4837 * This waits to be signaled for completion of a specific task. It can be
4838 * interrupted by a kill signal.
4840 int __sched wait_for_completion_killable(struct completion *x)
4842 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4843 if (t == -ERESTARTSYS)
4847 EXPORT_SYMBOL(wait_for_completion_killable);
4850 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4851 * @x: holds the state of this particular completion
4852 * @timeout: timeout value in jiffies
4854 * This waits for either a completion of a specific task to be
4855 * signaled or for a specified timeout to expire. It can be
4856 * interrupted by a kill signal. The timeout is in jiffies.
4859 wait_for_completion_killable_timeout(struct completion *x,
4860 unsigned long timeout)
4862 return wait_for_common(x, timeout, TASK_KILLABLE);
4864 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4867 * try_wait_for_completion - try to decrement a completion without blocking
4868 * @x: completion structure
4870 * Returns: 0 if a decrement cannot be done without blocking
4871 * 1 if a decrement succeeded.
4873 * If a completion is being used as a counting completion,
4874 * attempt to decrement the counter without blocking. This
4875 * enables us to avoid waiting if the resource the completion
4876 * is protecting is not available.
4878 bool try_wait_for_completion(struct completion *x)
4880 unsigned long flags;
4883 spin_lock_irqsave(&x->wait.lock, flags);
4888 spin_unlock_irqrestore(&x->wait.lock, flags);
4891 EXPORT_SYMBOL(try_wait_for_completion);
4894 * completion_done - Test to see if a completion has any waiters
4895 * @x: completion structure
4897 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4898 * 1 if there are no waiters.
4901 bool completion_done(struct completion *x)
4903 unsigned long flags;
4906 spin_lock_irqsave(&x->wait.lock, flags);
4909 spin_unlock_irqrestore(&x->wait.lock, flags);
4912 EXPORT_SYMBOL(completion_done);
4915 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4917 unsigned long flags;
4920 init_waitqueue_entry(&wait, current);
4922 __set_current_state(state);
4924 spin_lock_irqsave(&q->lock, flags);
4925 __add_wait_queue(q, &wait);
4926 spin_unlock(&q->lock);
4927 timeout = schedule_timeout(timeout);
4928 spin_lock_irq(&q->lock);
4929 __remove_wait_queue(q, &wait);
4930 spin_unlock_irqrestore(&q->lock, flags);
4935 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4937 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4939 EXPORT_SYMBOL(interruptible_sleep_on);
4942 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4944 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4946 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4948 void __sched sleep_on(wait_queue_head_t *q)
4950 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4952 EXPORT_SYMBOL(sleep_on);
4954 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4956 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4958 EXPORT_SYMBOL(sleep_on_timeout);
4960 #ifdef CONFIG_RT_MUTEXES
4963 * rt_mutex_setprio - set the current priority of a task
4965 * @prio: prio value (kernel-internal form)
4967 * This function changes the 'effective' priority of a task. It does
4968 * not touch ->normal_prio like __setscheduler().
4970 * Used by the rt_mutex code to implement priority inheritance logic.
4972 void rt_mutex_setprio(struct task_struct *p, int prio)
4974 int oldprio, on_rq, running;
4976 const struct sched_class *prev_class;
4978 BUG_ON(prio < 0 || prio > MAX_PRIO);
4980 rq = __task_rq_lock(p);
4982 trace_sched_pi_setprio(p, prio);
4984 prev_class = p->sched_class;
4986 running = task_current(rq, p);
4988 dequeue_task(rq, p, 0);
4990 p->sched_class->put_prev_task(rq, p);
4993 p->sched_class = &rt_sched_class;
4995 p->sched_class = &fair_sched_class;
5000 p->sched_class->set_curr_task(rq);
5002 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
5004 check_class_changed(rq, p, prev_class, oldprio);
5005 __task_rq_unlock(rq);
5010 void set_user_nice(struct task_struct *p, long nice)
5012 int old_prio, delta, on_rq;
5013 unsigned long flags;
5016 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5019 * We have to be careful, if called from sys_setpriority(),
5020 * the task might be in the middle of scheduling on another CPU.
5022 rq = task_rq_lock(p, &flags);
5024 * The RT priorities are set via sched_setscheduler(), but we still
5025 * allow the 'normal' nice value to be set - but as expected
5026 * it wont have any effect on scheduling until the task is
5027 * SCHED_FIFO/SCHED_RR:
5029 if (task_has_rt_policy(p)) {
5030 p->static_prio = NICE_TO_PRIO(nice);
5035 dequeue_task(rq, p, 0);
5037 p->static_prio = NICE_TO_PRIO(nice);
5040 p->prio = effective_prio(p);
5041 delta = p->prio - old_prio;
5044 enqueue_task(rq, p, 0);
5046 * If the task increased its priority or is running and
5047 * lowered its priority, then reschedule its CPU:
5049 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5050 resched_task(rq->curr);
5053 task_rq_unlock(rq, p, &flags);
5055 EXPORT_SYMBOL(set_user_nice);
5058 * can_nice - check if a task can reduce its nice value
5062 int can_nice(const struct task_struct *p, const int nice)
5064 /* convert nice value [19,-20] to rlimit style value [1,40] */
5065 int nice_rlim = 20 - nice;
5067 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
5068 capable(CAP_SYS_NICE));
5071 #ifdef __ARCH_WANT_SYS_NICE
5074 * sys_nice - change the priority of the current process.
5075 * @increment: priority increment
5077 * sys_setpriority is a more generic, but much slower function that
5078 * does similar things.
5080 SYSCALL_DEFINE1(nice, int, increment)
5085 * Setpriority might change our priority at the same moment.
5086 * We don't have to worry. Conceptually one call occurs first
5087 * and we have a single winner.
5089 if (increment < -40)
5094 nice = TASK_NICE(current) + increment;
5100 if (increment < 0 && !can_nice(current, nice))
5103 retval = security_task_setnice(current, nice);
5107 set_user_nice(current, nice);
5114 * task_prio - return the priority value of a given task.
5115 * @p: the task in question.
5117 * This is the priority value as seen by users in /proc.
5118 * RT tasks are offset by -200. Normal tasks are centered
5119 * around 0, value goes from -16 to +15.
5121 int task_prio(const struct task_struct *p)
5123 return p->prio - MAX_RT_PRIO;
5127 * task_nice - return the nice value of a given task.
5128 * @p: the task in question.
5130 int task_nice(const struct task_struct *p)
5132 return TASK_NICE(p);
5134 EXPORT_SYMBOL(task_nice);
5137 * idle_cpu - is a given cpu idle currently?
5138 * @cpu: the processor in question.
5140 int idle_cpu(int cpu)
5142 struct rq *rq = cpu_rq(cpu);
5144 if (rq->curr != rq->idle)
5151 if (!llist_empty(&rq->wake_list))
5159 * idle_task - return the idle task for a given cpu.
5160 * @cpu: the processor in question.
5162 struct task_struct *idle_task(int cpu)
5164 return cpu_rq(cpu)->idle;
5168 * find_process_by_pid - find a process with a matching PID value.
5169 * @pid: the pid in question.
5171 static struct task_struct *find_process_by_pid(pid_t pid)
5173 return pid ? find_task_by_vpid(pid) : current;
5176 /* Actually do priority change: must hold rq lock. */
5178 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5181 p->rt_priority = prio;
5182 p->normal_prio = normal_prio(p);
5183 /* we are holding p->pi_lock already */
5184 p->prio = rt_mutex_getprio(p);
5185 if (rt_prio(p->prio))
5186 p->sched_class = &rt_sched_class;
5188 p->sched_class = &fair_sched_class;
5193 * check the target process has a UID that matches the current process's
5195 static bool check_same_owner(struct task_struct *p)
5197 const struct cred *cred = current_cred(), *pcred;
5201 pcred = __task_cred(p);
5202 if (cred->user->user_ns == pcred->user->user_ns)
5203 match = (cred->euid == pcred->euid ||
5204 cred->euid == pcred->uid);
5211 static int __sched_setscheduler(struct task_struct *p, int policy,
5212 const struct sched_param *param, bool user)
5214 int retval, oldprio, oldpolicy = -1, on_rq, running;
5215 unsigned long flags;
5216 const struct sched_class *prev_class;
5220 /* may grab non-irq protected spin_locks */
5221 BUG_ON(in_interrupt());
5223 /* double check policy once rq lock held */
5225 reset_on_fork = p->sched_reset_on_fork;
5226 policy = oldpolicy = p->policy;
5228 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5229 policy &= ~SCHED_RESET_ON_FORK;
5231 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5232 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5233 policy != SCHED_IDLE)
5238 * Valid priorities for SCHED_FIFO and SCHED_RR are
5239 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5240 * SCHED_BATCH and SCHED_IDLE is 0.
5242 if (param->sched_priority < 0 ||
5243 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5244 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5246 if (rt_policy(policy) != (param->sched_priority != 0))
5250 * Allow unprivileged RT tasks to decrease priority:
5252 if (user && !capable(CAP_SYS_NICE)) {
5253 if (rt_policy(policy)) {
5254 unsigned long rlim_rtprio =
5255 task_rlimit(p, RLIMIT_RTPRIO);
5257 /* can't set/change the rt policy */
5258 if (policy != p->policy && !rlim_rtprio)
5261 /* can't increase priority */
5262 if (param->sched_priority > p->rt_priority &&
5263 param->sched_priority > rlim_rtprio)
5268 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5269 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5271 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5272 if (!can_nice(p, TASK_NICE(p)))
5276 /* can't change other user's priorities */
5277 if (!check_same_owner(p))
5280 /* Normal users shall not reset the sched_reset_on_fork flag */
5281 if (p->sched_reset_on_fork && !reset_on_fork)
5286 retval = security_task_setscheduler(p);
5292 * make sure no PI-waiters arrive (or leave) while we are
5293 * changing the priority of the task:
5295 * To be able to change p->policy safely, the appropriate
5296 * runqueue lock must be held.
5298 rq = task_rq_lock(p, &flags);
5301 * Changing the policy of the stop threads its a very bad idea
5303 if (p == rq->stop) {
5304 task_rq_unlock(rq, p, &flags);
5309 * If not changing anything there's no need to proceed further:
5311 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5312 param->sched_priority == p->rt_priority))) {
5314 __task_rq_unlock(rq);
5315 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5319 #ifdef CONFIG_RT_GROUP_SCHED
5322 * Do not allow realtime tasks into groups that have no runtime
5325 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5326 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5327 !task_group_is_autogroup(task_group(p))) {
5328 task_rq_unlock(rq, p, &flags);
5334 /* recheck policy now with rq lock held */
5335 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5336 policy = oldpolicy = -1;
5337 task_rq_unlock(rq, p, &flags);
5341 running = task_current(rq, p);
5343 deactivate_task(rq, p, 0);
5345 p->sched_class->put_prev_task(rq, p);
5347 p->sched_reset_on_fork = reset_on_fork;
5350 prev_class = p->sched_class;
5351 __setscheduler(rq, p, policy, param->sched_priority);
5354 p->sched_class->set_curr_task(rq);
5356 activate_task(rq, p, 0);
5358 check_class_changed(rq, p, prev_class, oldprio);
5359 task_rq_unlock(rq, p, &flags);
5361 rt_mutex_adjust_pi(p);
5367 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5368 * @p: the task in question.
5369 * @policy: new policy.
5370 * @param: structure containing the new RT priority.
5372 * NOTE that the task may be already dead.
5374 int sched_setscheduler(struct task_struct *p, int policy,
5375 const struct sched_param *param)
5377 return __sched_setscheduler(p, policy, param, true);
5379 EXPORT_SYMBOL_GPL(sched_setscheduler);
5382 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5383 * @p: the task in question.
5384 * @policy: new policy.
5385 * @param: structure containing the new RT priority.
5387 * Just like sched_setscheduler, only don't bother checking if the
5388 * current context has permission. For example, this is needed in
5389 * stop_machine(): we create temporary high priority worker threads,
5390 * but our caller might not have that capability.
5392 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5393 const struct sched_param *param)
5395 return __sched_setscheduler(p, policy, param, false);
5399 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5401 struct sched_param lparam;
5402 struct task_struct *p;
5405 if (!param || pid < 0)
5407 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5412 p = find_process_by_pid(pid);
5414 retval = sched_setscheduler(p, policy, &lparam);
5421 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5422 * @pid: the pid in question.
5423 * @policy: new policy.
5424 * @param: structure containing the new RT priority.
5426 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5427 struct sched_param __user *, param)
5429 /* negative values for policy are not valid */
5433 return do_sched_setscheduler(pid, policy, param);
5437 * sys_sched_setparam - set/change the RT priority of a thread
5438 * @pid: the pid in question.
5439 * @param: structure containing the new RT priority.
5441 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5443 return do_sched_setscheduler(pid, -1, param);
5447 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5448 * @pid: the pid in question.
5450 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5452 struct task_struct *p;
5460 p = find_process_by_pid(pid);
5462 retval = security_task_getscheduler(p);
5465 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5472 * sys_sched_getparam - get the RT priority of a thread
5473 * @pid: the pid in question.
5474 * @param: structure containing the RT priority.
5476 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5478 struct sched_param lp;
5479 struct task_struct *p;
5482 if (!param || pid < 0)
5486 p = find_process_by_pid(pid);
5491 retval = security_task_getscheduler(p);
5495 lp.sched_priority = p->rt_priority;
5499 * This one might sleep, we cannot do it with a spinlock held ...
5501 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5510 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5512 cpumask_var_t cpus_allowed, new_mask;
5513 struct task_struct *p;
5519 p = find_process_by_pid(pid);
5526 /* Prevent p going away */
5530 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5534 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5536 goto out_free_cpus_allowed;
5539 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5542 retval = security_task_setscheduler(p);
5546 cpuset_cpus_allowed(p, cpus_allowed);
5547 cpumask_and(new_mask, in_mask, cpus_allowed);
5549 retval = set_cpus_allowed_ptr(p, new_mask);
5552 cpuset_cpus_allowed(p, cpus_allowed);
5553 if (!cpumask_subset(new_mask, cpus_allowed)) {
5555 * We must have raced with a concurrent cpuset
5556 * update. Just reset the cpus_allowed to the
5557 * cpuset's cpus_allowed
5559 cpumask_copy(new_mask, cpus_allowed);
5564 free_cpumask_var(new_mask);
5565 out_free_cpus_allowed:
5566 free_cpumask_var(cpus_allowed);
5573 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5574 struct cpumask *new_mask)
5576 if (len < cpumask_size())
5577 cpumask_clear(new_mask);
5578 else if (len > cpumask_size())
5579 len = cpumask_size();
5581 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5585 * sys_sched_setaffinity - set the cpu affinity of a process
5586 * @pid: pid of the process
5587 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5588 * @user_mask_ptr: user-space pointer to the new cpu mask
5590 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5591 unsigned long __user *, user_mask_ptr)
5593 cpumask_var_t new_mask;
5596 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5599 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5601 retval = sched_setaffinity(pid, new_mask);
5602 free_cpumask_var(new_mask);
5606 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5608 struct task_struct *p;
5609 unsigned long flags;
5616 p = find_process_by_pid(pid);
5620 retval = security_task_getscheduler(p);
5624 raw_spin_lock_irqsave(&p->pi_lock, flags);
5625 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5626 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5636 * sys_sched_getaffinity - get the cpu affinity of a process
5637 * @pid: pid of the process
5638 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5639 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5641 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5642 unsigned long __user *, user_mask_ptr)
5647 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5649 if (len & (sizeof(unsigned long)-1))
5652 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5655 ret = sched_getaffinity(pid, mask);
5657 size_t retlen = min_t(size_t, len, cpumask_size());
5659 if (copy_to_user(user_mask_ptr, mask, retlen))
5664 free_cpumask_var(mask);
5670 * sys_sched_yield - yield the current processor to other threads.
5672 * This function yields the current CPU to other tasks. If there are no
5673 * other threads running on this CPU then this function will return.
5675 SYSCALL_DEFINE0(sched_yield)
5677 struct rq *rq = this_rq_lock();
5679 schedstat_inc(rq, yld_count);
5680 current->sched_class->yield_task(rq);
5683 * Since we are going to call schedule() anyway, there's
5684 * no need to preempt or enable interrupts:
5686 __release(rq->lock);
5687 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5688 do_raw_spin_unlock(&rq->lock);
5689 preempt_enable_no_resched();
5696 static inline int should_resched(void)
5698 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5701 static void __cond_resched(void)
5703 add_preempt_count(PREEMPT_ACTIVE);
5705 sub_preempt_count(PREEMPT_ACTIVE);
5708 int __sched _cond_resched(void)
5710 if (should_resched()) {
5716 EXPORT_SYMBOL(_cond_resched);
5719 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5720 * call schedule, and on return reacquire the lock.
5722 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5723 * operations here to prevent schedule() from being called twice (once via
5724 * spin_unlock(), once by hand).
5726 int __cond_resched_lock(spinlock_t *lock)
5728 int resched = should_resched();
5731 lockdep_assert_held(lock);
5733 if (spin_needbreak(lock) || resched) {
5744 EXPORT_SYMBOL(__cond_resched_lock);
5746 int __sched __cond_resched_softirq(void)
5748 BUG_ON(!in_softirq());
5750 if (should_resched()) {
5758 EXPORT_SYMBOL(__cond_resched_softirq);
5761 * yield - yield the current processor to other threads.
5763 * This is a shortcut for kernel-space yielding - it marks the
5764 * thread runnable and calls sys_sched_yield().
5766 void __sched yield(void)
5768 set_current_state(TASK_RUNNING);
5771 EXPORT_SYMBOL(yield);
5774 * yield_to - yield the current processor to another thread in
5775 * your thread group, or accelerate that thread toward the
5776 * processor it's on.
5778 * @preempt: whether task preemption is allowed or not
5780 * It's the caller's job to ensure that the target task struct
5781 * can't go away on us before we can do any checks.
5783 * Returns true if we indeed boosted the target task.
5785 bool __sched yield_to(struct task_struct *p, bool preempt)
5787 struct task_struct *curr = current;
5788 struct rq *rq, *p_rq;
5789 unsigned long flags;
5792 local_irq_save(flags);
5797 double_rq_lock(rq, p_rq);
5798 while (task_rq(p) != p_rq) {
5799 double_rq_unlock(rq, p_rq);
5803 if (!curr->sched_class->yield_to_task)
5806 if (curr->sched_class != p->sched_class)
5809 if (task_running(p_rq, p) || p->state)
5812 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5814 schedstat_inc(rq, yld_count);
5816 * Make p's CPU reschedule; pick_next_entity takes care of
5819 if (preempt && rq != p_rq)
5820 resched_task(p_rq->curr);
5824 double_rq_unlock(rq, p_rq);
5825 local_irq_restore(flags);
5832 EXPORT_SYMBOL_GPL(yield_to);
5835 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5836 * that process accounting knows that this is a task in IO wait state.
5838 void __sched io_schedule(void)
5840 struct rq *rq = raw_rq();
5842 delayacct_blkio_start();
5843 atomic_inc(&rq->nr_iowait);
5844 blk_flush_plug(current);
5845 current->in_iowait = 1;
5847 current->in_iowait = 0;
5848 atomic_dec(&rq->nr_iowait);
5849 delayacct_blkio_end();
5851 EXPORT_SYMBOL(io_schedule);
5853 long __sched io_schedule_timeout(long timeout)
5855 struct rq *rq = raw_rq();
5858 delayacct_blkio_start();
5859 atomic_inc(&rq->nr_iowait);
5860 blk_flush_plug(current);
5861 current->in_iowait = 1;
5862 ret = schedule_timeout(timeout);
5863 current->in_iowait = 0;
5864 atomic_dec(&rq->nr_iowait);
5865 delayacct_blkio_end();
5870 * sys_sched_get_priority_max - return maximum RT priority.
5871 * @policy: scheduling class.
5873 * this syscall returns the maximum rt_priority that can be used
5874 * by a given scheduling class.
5876 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5883 ret = MAX_USER_RT_PRIO-1;
5895 * sys_sched_get_priority_min - return minimum RT priority.
5896 * @policy: scheduling class.
5898 * this syscall returns the minimum rt_priority that can be used
5899 * by a given scheduling class.
5901 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5919 * sys_sched_rr_get_interval - return the default timeslice of a process.
5920 * @pid: pid of the process.
5921 * @interval: userspace pointer to the timeslice value.
5923 * this syscall writes the default timeslice value of a given process
5924 * into the user-space timespec buffer. A value of '0' means infinity.
5926 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5927 struct timespec __user *, interval)
5929 struct task_struct *p;
5930 unsigned int time_slice;
5931 unsigned long flags;
5941 p = find_process_by_pid(pid);
5945 retval = security_task_getscheduler(p);
5949 rq = task_rq_lock(p, &flags);
5950 time_slice = p->sched_class->get_rr_interval(rq, p);
5951 task_rq_unlock(rq, p, &flags);
5954 jiffies_to_timespec(time_slice, &t);
5955 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5963 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5965 void sched_show_task(struct task_struct *p)
5967 unsigned long free = 0;
5970 state = p->state ? __ffs(p->state) + 1 : 0;
5971 printk(KERN_INFO "%-15.15s %c", p->comm,
5972 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5973 #if BITS_PER_LONG == 32
5974 if (state == TASK_RUNNING)
5975 printk(KERN_CONT " running ");
5977 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5979 if (state == TASK_RUNNING)
5980 printk(KERN_CONT " running task ");
5982 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5984 #ifdef CONFIG_DEBUG_STACK_USAGE
5985 free = stack_not_used(p);
5987 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5988 task_pid_nr(p), task_pid_nr(p->real_parent),
5989 (unsigned long)task_thread_info(p)->flags);
5991 show_stack(p, NULL);
5994 void show_state_filter(unsigned long state_filter)
5996 struct task_struct *g, *p;
5998 #if BITS_PER_LONG == 32
6000 " task PC stack pid father\n");
6003 " task PC stack pid father\n");
6005 read_lock(&tasklist_lock);
6006 do_each_thread(g, p) {
6008 * reset the NMI-timeout, listing all files on a slow
6009 * console might take a lot of time:
6011 touch_nmi_watchdog();
6012 if (!state_filter || (p->state & state_filter))
6014 } while_each_thread(g, p);
6016 touch_all_softlockup_watchdogs();
6018 #ifdef CONFIG_SCHED_DEBUG
6019 sysrq_sched_debug_show();
6021 read_unlock(&tasklist_lock);
6023 * Only show locks if all tasks are dumped:
6026 debug_show_all_locks();
6029 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6031 idle->sched_class = &idle_sched_class;
6035 * init_idle - set up an idle thread for a given CPU
6036 * @idle: task in question
6037 * @cpu: cpu the idle task belongs to
6039 * NOTE: this function does not set the idle thread's NEED_RESCHED
6040 * flag, to make booting more robust.
6042 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6044 struct rq *rq = cpu_rq(cpu);
6045 unsigned long flags;
6047 raw_spin_lock_irqsave(&rq->lock, flags);
6050 idle->state = TASK_RUNNING;
6051 idle->se.exec_start = sched_clock();
6053 do_set_cpus_allowed(idle, cpumask_of(cpu));
6055 * We're having a chicken and egg problem, even though we are
6056 * holding rq->lock, the cpu isn't yet set to this cpu so the
6057 * lockdep check in task_group() will fail.
6059 * Similar case to sched_fork(). / Alternatively we could
6060 * use task_rq_lock() here and obtain the other rq->lock.
6065 __set_task_cpu(idle, cpu);
6068 rq->curr = rq->idle = idle;
6069 #if defined(CONFIG_SMP)
6072 raw_spin_unlock_irqrestore(&rq->lock, flags);
6074 /* Set the preempt count _outside_ the spinlocks! */
6075 task_thread_info(idle)->preempt_count = 0;
6078 * The idle tasks have their own, simple scheduling class:
6080 idle->sched_class = &idle_sched_class;
6081 ftrace_graph_init_idle_task(idle, cpu);
6085 * Increase the granularity value when there are more CPUs,
6086 * because with more CPUs the 'effective latency' as visible
6087 * to users decreases. But the relationship is not linear,
6088 * so pick a second-best guess by going with the log2 of the
6091 * This idea comes from the SD scheduler of Con Kolivas:
6093 static int get_update_sysctl_factor(void)
6095 unsigned int cpus = min_t(int, num_online_cpus(), 8);
6096 unsigned int factor;
6098 switch (sysctl_sched_tunable_scaling) {
6099 case SCHED_TUNABLESCALING_NONE:
6102 case SCHED_TUNABLESCALING_LINEAR:
6105 case SCHED_TUNABLESCALING_LOG:
6107 factor = 1 + ilog2(cpus);
6114 static void update_sysctl(void)
6116 unsigned int factor = get_update_sysctl_factor();
6118 #define SET_SYSCTL(name) \
6119 (sysctl_##name = (factor) * normalized_sysctl_##name)
6120 SET_SYSCTL(sched_min_granularity);
6121 SET_SYSCTL(sched_latency);
6122 SET_SYSCTL(sched_wakeup_granularity);
6126 static inline void sched_init_granularity(void)
6132 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
6134 if (p->sched_class && p->sched_class->set_cpus_allowed)
6135 p->sched_class->set_cpus_allowed(p, new_mask);
6137 cpumask_copy(&p->cpus_allowed, new_mask);
6138 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6143 * This is how migration works:
6145 * 1) we invoke migration_cpu_stop() on the target CPU using
6147 * 2) stopper starts to run (implicitly forcing the migrated thread
6149 * 3) it checks whether the migrated task is still in the wrong runqueue.
6150 * 4) if it's in the wrong runqueue then the migration thread removes
6151 * it and puts it into the right queue.
6152 * 5) stopper completes and stop_one_cpu() returns and the migration
6157 * Change a given task's CPU affinity. Migrate the thread to a
6158 * proper CPU and schedule it away if the CPU it's executing on
6159 * is removed from the allowed bitmask.
6161 * NOTE: the caller must have a valid reference to the task, the
6162 * task must not exit() & deallocate itself prematurely. The
6163 * call is not atomic; no spinlocks may be held.
6165 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6167 unsigned long flags;
6169 unsigned int dest_cpu;
6172 rq = task_rq_lock(p, &flags);
6174 if (cpumask_equal(&p->cpus_allowed, new_mask))
6177 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6182 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6187 do_set_cpus_allowed(p, new_mask);
6189 /* Can the task run on the task's current CPU? If so, we're done */
6190 if (cpumask_test_cpu(task_cpu(p), new_mask))
6193 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6195 struct migration_arg arg = { p, dest_cpu };
6196 /* Need help from migration thread: drop lock and wait. */
6197 task_rq_unlock(rq, p, &flags);
6198 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6199 tlb_migrate_finish(p->mm);
6203 task_rq_unlock(rq, p, &flags);
6207 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6210 * Move (not current) task off this cpu, onto dest cpu. We're doing
6211 * this because either it can't run here any more (set_cpus_allowed()
6212 * away from this CPU, or CPU going down), or because we're
6213 * attempting to rebalance this task on exec (sched_exec).
6215 * So we race with normal scheduler movements, but that's OK, as long
6216 * as the task is no longer on this CPU.
6218 * Returns non-zero if task was successfully migrated.
6220 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6222 struct rq *rq_dest, *rq_src;
6225 if (unlikely(!cpu_active(dest_cpu)))
6228 rq_src = cpu_rq(src_cpu);
6229 rq_dest = cpu_rq(dest_cpu);
6231 raw_spin_lock(&p->pi_lock);
6232 double_rq_lock(rq_src, rq_dest);
6233 /* Already moved. */
6234 if (task_cpu(p) != src_cpu)
6236 /* Affinity changed (again). */
6237 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6241 * If we're not on a rq, the next wake-up will ensure we're
6245 deactivate_task(rq_src, p, 0);
6246 set_task_cpu(p, dest_cpu);
6247 activate_task(rq_dest, p, 0);
6248 check_preempt_curr(rq_dest, p, 0);
6253 double_rq_unlock(rq_src, rq_dest);
6254 raw_spin_unlock(&p->pi_lock);
6259 * migration_cpu_stop - this will be executed by a highprio stopper thread
6260 * and performs thread migration by bumping thread off CPU then
6261 * 'pushing' onto another runqueue.
6263 static int migration_cpu_stop(void *data)
6265 struct migration_arg *arg = data;
6268 * The original target cpu might have gone down and we might
6269 * be on another cpu but it doesn't matter.
6271 local_irq_disable();
6272 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6277 #ifdef CONFIG_HOTPLUG_CPU
6280 * Ensures that the idle task is using init_mm right before its cpu goes
6283 void idle_task_exit(void)
6285 struct mm_struct *mm = current->active_mm;
6287 BUG_ON(cpu_online(smp_processor_id()));
6290 switch_mm(mm, &init_mm, current);
6295 * While a dead CPU has no uninterruptible tasks queued at this point,
6296 * it might still have a nonzero ->nr_uninterruptible counter, because
6297 * for performance reasons the counter is not stricly tracking tasks to
6298 * their home CPUs. So we just add the counter to another CPU's counter,
6299 * to keep the global sum constant after CPU-down:
6301 static void migrate_nr_uninterruptible(struct rq *rq_src)
6303 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6305 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6306 rq_src->nr_uninterruptible = 0;
6310 * remove the tasks which were accounted by rq from calc_load_tasks.
6312 static void calc_global_load_remove(struct rq *rq)
6314 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6315 rq->calc_load_active = 0;
6318 #ifdef CONFIG_CFS_BANDWIDTH
6319 static void unthrottle_offline_cfs_rqs(struct rq *rq)
6321 struct cfs_rq *cfs_rq;
6323 for_each_leaf_cfs_rq(rq, cfs_rq) {
6324 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(cfs_rq->tg);
6326 if (!cfs_rq->runtime_enabled)
6330 * clock_task is not advancing so we just need to make sure
6331 * there's some valid quota amount
6333 cfs_rq->runtime_remaining = cfs_b->quota;
6334 if (cfs_rq_throttled(cfs_rq))
6335 unthrottle_cfs_rq(cfs_rq);
6339 static void unthrottle_offline_cfs_rqs(struct rq *rq) {}
6343 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6344 * try_to_wake_up()->select_task_rq().
6346 * Called with rq->lock held even though we'er in stop_machine() and
6347 * there's no concurrency possible, we hold the required locks anyway
6348 * because of lock validation efforts.
6350 static void migrate_tasks(unsigned int dead_cpu)
6352 struct rq *rq = cpu_rq(dead_cpu);
6353 struct task_struct *next, *stop = rq->stop;
6357 * Fudge the rq selection such that the below task selection loop
6358 * doesn't get stuck on the currently eligible stop task.
6360 * We're currently inside stop_machine() and the rq is either stuck
6361 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6362 * either way we should never end up calling schedule() until we're
6367 /* Ensure any throttled groups are reachable by pick_next_task */
6368 unthrottle_offline_cfs_rqs(rq);
6372 * There's this thread running, bail when that's the only
6375 if (rq->nr_running == 1)
6378 next = pick_next_task(rq);
6380 next->sched_class->put_prev_task(rq, next);
6382 /* Find suitable destination for @next, with force if needed. */
6383 dest_cpu = select_fallback_rq(dead_cpu, next);
6384 raw_spin_unlock(&rq->lock);
6386 __migrate_task(next, dead_cpu, dest_cpu);
6388 raw_spin_lock(&rq->lock);
6394 #endif /* CONFIG_HOTPLUG_CPU */
6396 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6398 static struct ctl_table sd_ctl_dir[] = {
6400 .procname = "sched_domain",
6406 static struct ctl_table sd_ctl_root[] = {
6408 .procname = "kernel",
6410 .child = sd_ctl_dir,
6415 static struct ctl_table *sd_alloc_ctl_entry(int n)
6417 struct ctl_table *entry =
6418 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6423 static void sd_free_ctl_entry(struct ctl_table **tablep)
6425 struct ctl_table *entry;
6428 * In the intermediate directories, both the child directory and
6429 * procname are dynamically allocated and could fail but the mode
6430 * will always be set. In the lowest directory the names are
6431 * static strings and all have proc handlers.
6433 for (entry = *tablep; entry->mode; entry++) {
6435 sd_free_ctl_entry(&entry->child);
6436 if (entry->proc_handler == NULL)
6437 kfree(entry->procname);
6445 set_table_entry(struct ctl_table *entry,
6446 const char *procname, void *data, int maxlen,
6447 mode_t mode, proc_handler *proc_handler)
6449 entry->procname = procname;
6451 entry->maxlen = maxlen;
6453 entry->proc_handler = proc_handler;
6456 static struct ctl_table *
6457 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6459 struct ctl_table *table = sd_alloc_ctl_entry(13);
6464 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6465 sizeof(long), 0644, proc_doulongvec_minmax);
6466 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6467 sizeof(long), 0644, proc_doulongvec_minmax);
6468 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6469 sizeof(int), 0644, proc_dointvec_minmax);
6470 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6471 sizeof(int), 0644, proc_dointvec_minmax);
6472 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6473 sizeof(int), 0644, proc_dointvec_minmax);
6474 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6475 sizeof(int), 0644, proc_dointvec_minmax);
6476 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6477 sizeof(int), 0644, proc_dointvec_minmax);
6478 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6479 sizeof(int), 0644, proc_dointvec_minmax);
6480 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6481 sizeof(int), 0644, proc_dointvec_minmax);
6482 set_table_entry(&table[9], "cache_nice_tries",
6483 &sd->cache_nice_tries,
6484 sizeof(int), 0644, proc_dointvec_minmax);
6485 set_table_entry(&table[10], "flags", &sd->flags,
6486 sizeof(int), 0644, proc_dointvec_minmax);
6487 set_table_entry(&table[11], "name", sd->name,
6488 CORENAME_MAX_SIZE, 0444, proc_dostring);
6489 /* &table[12] is terminator */
6494 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6496 struct ctl_table *entry, *table;
6497 struct sched_domain *sd;
6498 int domain_num = 0, i;
6501 for_each_domain(cpu, sd)
6503 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6508 for_each_domain(cpu, sd) {
6509 snprintf(buf, 32, "domain%d", i);
6510 entry->procname = kstrdup(buf, GFP_KERNEL);
6512 entry->child = sd_alloc_ctl_domain_table(sd);
6519 static struct ctl_table_header *sd_sysctl_header;
6520 static void register_sched_domain_sysctl(void)
6522 int i, cpu_num = num_possible_cpus();
6523 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6526 WARN_ON(sd_ctl_dir[0].child);
6527 sd_ctl_dir[0].child = entry;
6532 for_each_possible_cpu(i) {
6533 snprintf(buf, 32, "cpu%d", i);
6534 entry->procname = kstrdup(buf, GFP_KERNEL);
6536 entry->child = sd_alloc_ctl_cpu_table(i);
6540 WARN_ON(sd_sysctl_header);
6541 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6544 /* may be called multiple times per register */
6545 static void unregister_sched_domain_sysctl(void)
6547 if (sd_sysctl_header)
6548 unregister_sysctl_table(sd_sysctl_header);
6549 sd_sysctl_header = NULL;
6550 if (sd_ctl_dir[0].child)
6551 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6554 static void register_sched_domain_sysctl(void)
6557 static void unregister_sched_domain_sysctl(void)
6562 static void set_rq_online(struct rq *rq)
6565 const struct sched_class *class;
6567 cpumask_set_cpu(rq->cpu, rq->rd->online);
6570 for_each_class(class) {
6571 if (class->rq_online)
6572 class->rq_online(rq);
6577 static void set_rq_offline(struct rq *rq)
6580 const struct sched_class *class;
6582 for_each_class(class) {
6583 if (class->rq_offline)
6584 class->rq_offline(rq);
6587 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6593 * migration_call - callback that gets triggered when a CPU is added.
6594 * Here we can start up the necessary migration thread for the new CPU.
6596 static int __cpuinit
6597 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6599 int cpu = (long)hcpu;
6600 unsigned long flags;
6601 struct rq *rq = cpu_rq(cpu);
6603 switch (action & ~CPU_TASKS_FROZEN) {
6605 case CPU_UP_PREPARE:
6606 rq->calc_load_update = calc_load_update;
6610 /* Update our root-domain */
6611 raw_spin_lock_irqsave(&rq->lock, flags);
6613 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6617 raw_spin_unlock_irqrestore(&rq->lock, flags);
6620 #ifdef CONFIG_HOTPLUG_CPU
6622 sched_ttwu_pending();
6623 /* Update our root-domain */
6624 raw_spin_lock_irqsave(&rq->lock, flags);
6626 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6630 BUG_ON(rq->nr_running != 1); /* the migration thread */
6631 raw_spin_unlock_irqrestore(&rq->lock, flags);
6633 migrate_nr_uninterruptible(rq);
6634 calc_global_load_remove(rq);
6639 update_max_interval();
6645 * Register at high priority so that task migration (migrate_all_tasks)
6646 * happens before everything else. This has to be lower priority than
6647 * the notifier in the perf_event subsystem, though.
6649 static struct notifier_block __cpuinitdata migration_notifier = {
6650 .notifier_call = migration_call,
6651 .priority = CPU_PRI_MIGRATION,
6654 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6655 unsigned long action, void *hcpu)
6657 switch (action & ~CPU_TASKS_FROZEN) {
6659 case CPU_DOWN_FAILED:
6660 set_cpu_active((long)hcpu, true);
6667 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6668 unsigned long action, void *hcpu)
6670 switch (action & ~CPU_TASKS_FROZEN) {
6671 case CPU_DOWN_PREPARE:
6672 set_cpu_active((long)hcpu, false);
6679 static int __init migration_init(void)
6681 void *cpu = (void *)(long)smp_processor_id();
6684 /* Initialize migration for the boot CPU */
6685 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6686 BUG_ON(err == NOTIFY_BAD);
6687 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6688 register_cpu_notifier(&migration_notifier);
6690 /* Register cpu active notifiers */
6691 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6692 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6696 early_initcall(migration_init);
6701 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6703 #ifdef CONFIG_SCHED_DEBUG
6705 static __read_mostly int sched_domain_debug_enabled;
6707 static int __init sched_domain_debug_setup(char *str)
6709 sched_domain_debug_enabled = 1;
6713 early_param("sched_debug", sched_domain_debug_setup);
6715 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6716 struct cpumask *groupmask)
6718 struct sched_group *group = sd->groups;
6721 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6722 cpumask_clear(groupmask);
6724 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6726 if (!(sd->flags & SD_LOAD_BALANCE)) {
6727 printk("does not load-balance\n");
6729 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6734 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6736 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6737 printk(KERN_ERR "ERROR: domain->span does not contain "
6740 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6741 printk(KERN_ERR "ERROR: domain->groups does not contain"
6745 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6749 printk(KERN_ERR "ERROR: group is NULL\n");
6753 if (!group->sgp->power) {
6754 printk(KERN_CONT "\n");
6755 printk(KERN_ERR "ERROR: domain->cpu_power not "
6760 if (!cpumask_weight(sched_group_cpus(group))) {
6761 printk(KERN_CONT "\n");
6762 printk(KERN_ERR "ERROR: empty group\n");
6766 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6767 printk(KERN_CONT "\n");
6768 printk(KERN_ERR "ERROR: repeated CPUs\n");
6772 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6774 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6776 printk(KERN_CONT " %s", str);
6777 if (group->sgp->power != SCHED_POWER_SCALE) {
6778 printk(KERN_CONT " (cpu_power = %d)",
6782 group = group->next;
6783 } while (group != sd->groups);
6784 printk(KERN_CONT "\n");
6786 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6787 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6790 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6791 printk(KERN_ERR "ERROR: parent span is not a superset "
6792 "of domain->span\n");
6796 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6800 if (!sched_domain_debug_enabled)
6804 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6808 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6811 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6819 #else /* !CONFIG_SCHED_DEBUG */
6820 # define sched_domain_debug(sd, cpu) do { } while (0)
6821 #endif /* CONFIG_SCHED_DEBUG */
6823 static int sd_degenerate(struct sched_domain *sd)
6825 if (cpumask_weight(sched_domain_span(sd)) == 1)
6828 /* Following flags need at least 2 groups */
6829 if (sd->flags & (SD_LOAD_BALANCE |
6830 SD_BALANCE_NEWIDLE |
6834 SD_SHARE_PKG_RESOURCES)) {
6835 if (sd->groups != sd->groups->next)
6839 /* Following flags don't use groups */
6840 if (sd->flags & (SD_WAKE_AFFINE))
6847 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6849 unsigned long cflags = sd->flags, pflags = parent->flags;
6851 if (sd_degenerate(parent))
6854 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6857 /* Flags needing groups don't count if only 1 group in parent */
6858 if (parent->groups == parent->groups->next) {
6859 pflags &= ~(SD_LOAD_BALANCE |
6860 SD_BALANCE_NEWIDLE |
6864 SD_SHARE_PKG_RESOURCES);
6865 if (nr_node_ids == 1)
6866 pflags &= ~SD_SERIALIZE;
6868 if (~cflags & pflags)
6874 static void free_rootdomain(struct rcu_head *rcu)
6876 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6878 cpupri_cleanup(&rd->cpupri);
6879 free_cpumask_var(rd->rto_mask);
6880 free_cpumask_var(rd->online);
6881 free_cpumask_var(rd->span);
6885 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6887 struct root_domain *old_rd = NULL;
6888 unsigned long flags;
6890 raw_spin_lock_irqsave(&rq->lock, flags);
6895 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6898 cpumask_clear_cpu(rq->cpu, old_rd->span);
6901 * If we dont want to free the old_rt yet then
6902 * set old_rd to NULL to skip the freeing later
6905 if (!atomic_dec_and_test(&old_rd->refcount))
6909 atomic_inc(&rd->refcount);
6912 cpumask_set_cpu(rq->cpu, rd->span);
6913 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6916 raw_spin_unlock_irqrestore(&rq->lock, flags);
6919 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6922 static int init_rootdomain(struct root_domain *rd)
6924 memset(rd, 0, sizeof(*rd));
6926 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6928 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6930 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6933 if (cpupri_init(&rd->cpupri) != 0)
6938 free_cpumask_var(rd->rto_mask);
6940 free_cpumask_var(rd->online);
6942 free_cpumask_var(rd->span);
6947 static void init_defrootdomain(void)
6949 init_rootdomain(&def_root_domain);
6951 atomic_set(&def_root_domain.refcount, 1);
6954 static struct root_domain *alloc_rootdomain(void)
6956 struct root_domain *rd;
6958 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6962 if (init_rootdomain(rd) != 0) {
6970 static void free_sched_groups(struct sched_group *sg, int free_sgp)
6972 struct sched_group *tmp, *first;
6981 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
6986 } while (sg != first);
6989 static void free_sched_domain(struct rcu_head *rcu)
6991 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6994 * If its an overlapping domain it has private groups, iterate and
6997 if (sd->flags & SD_OVERLAP) {
6998 free_sched_groups(sd->groups, 1);
6999 } else if (atomic_dec_and_test(&sd->groups->ref)) {
7000 kfree(sd->groups->sgp);
7006 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
7008 call_rcu(&sd->rcu, free_sched_domain);
7011 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
7013 for (; sd; sd = sd->parent)
7014 destroy_sched_domain(sd, cpu);
7018 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7019 * hold the hotplug lock.
7022 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7024 struct rq *rq = cpu_rq(cpu);
7025 struct sched_domain *tmp;
7027 /* Remove the sched domains which do not contribute to scheduling. */
7028 for (tmp = sd; tmp; ) {
7029 struct sched_domain *parent = tmp->parent;
7033 if (sd_parent_degenerate(tmp, parent)) {
7034 tmp->parent = parent->parent;
7036 parent->parent->child = tmp;
7037 destroy_sched_domain(parent, cpu);
7042 if (sd && sd_degenerate(sd)) {
7045 destroy_sched_domain(tmp, cpu);
7050 sched_domain_debug(sd, cpu);
7052 rq_attach_root(rq, rd);
7054 rcu_assign_pointer(rq->sd, sd);
7055 destroy_sched_domains(tmp, cpu);
7058 /* cpus with isolated domains */
7059 static cpumask_var_t cpu_isolated_map;
7061 /* Setup the mask of cpus configured for isolated domains */
7062 static int __init isolated_cpu_setup(char *str)
7064 alloc_bootmem_cpumask_var(&cpu_isolated_map);
7065 cpulist_parse(str, cpu_isolated_map);
7069 __setup("isolcpus=", isolated_cpu_setup);
7074 * find_next_best_node - find the next node to include in a sched_domain
7075 * @node: node whose sched_domain we're building
7076 * @used_nodes: nodes already in the sched_domain
7078 * Find the next node to include in a given scheduling domain. Simply
7079 * finds the closest node not already in the @used_nodes map.
7081 * Should use nodemask_t.
7083 static int find_next_best_node(int node, nodemask_t *used_nodes)
7085 int i, n, val, min_val, best_node = -1;
7089 for (i = 0; i < nr_node_ids; i++) {
7090 /* Start at @node */
7091 n = (node + i) % nr_node_ids;
7093 if (!nr_cpus_node(n))
7096 /* Skip already used nodes */
7097 if (node_isset(n, *used_nodes))
7100 /* Simple min distance search */
7101 val = node_distance(node, n);
7103 if (val < min_val) {
7109 if (best_node != -1)
7110 node_set(best_node, *used_nodes);
7115 * sched_domain_node_span - get a cpumask for a node's sched_domain
7116 * @node: node whose cpumask we're constructing
7117 * @span: resulting cpumask
7119 * Given a node, construct a good cpumask for its sched_domain to span. It
7120 * should be one that prevents unnecessary balancing, but also spreads tasks
7123 static void sched_domain_node_span(int node, struct cpumask *span)
7125 nodemask_t used_nodes;
7128 cpumask_clear(span);
7129 nodes_clear(used_nodes);
7131 cpumask_or(span, span, cpumask_of_node(node));
7132 node_set(node, used_nodes);
7134 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7135 int next_node = find_next_best_node(node, &used_nodes);
7138 cpumask_or(span, span, cpumask_of_node(next_node));
7142 static const struct cpumask *cpu_node_mask(int cpu)
7144 lockdep_assert_held(&sched_domains_mutex);
7146 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
7148 return sched_domains_tmpmask;
7151 static const struct cpumask *cpu_allnodes_mask(int cpu)
7153 return cpu_possible_mask;
7155 #endif /* CONFIG_NUMA */
7157 static const struct cpumask *cpu_cpu_mask(int cpu)
7159 return cpumask_of_node(cpu_to_node(cpu));
7162 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7165 struct sched_domain **__percpu sd;
7166 struct sched_group **__percpu sg;
7167 struct sched_group_power **__percpu sgp;
7171 struct sched_domain ** __percpu sd;
7172 struct root_domain *rd;
7182 struct sched_domain_topology_level;
7184 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
7185 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
7187 #define SDTL_OVERLAP 0x01
7189 struct sched_domain_topology_level {
7190 sched_domain_init_f init;
7191 sched_domain_mask_f mask;
7193 struct sd_data data;
7197 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
7199 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
7200 const struct cpumask *span = sched_domain_span(sd);
7201 struct cpumask *covered = sched_domains_tmpmask;
7202 struct sd_data *sdd = sd->private;
7203 struct sched_domain *child;
7206 cpumask_clear(covered);
7208 for_each_cpu(i, span) {
7209 struct cpumask *sg_span;
7211 if (cpumask_test_cpu(i, covered))
7214 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7215 GFP_KERNEL, cpu_to_node(i));
7220 sg_span = sched_group_cpus(sg);
7222 child = *per_cpu_ptr(sdd->sd, i);
7224 child = child->child;
7225 cpumask_copy(sg_span, sched_domain_span(child));
7227 cpumask_set_cpu(i, sg_span);
7229 cpumask_or(covered, covered, sg_span);
7231 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
7232 atomic_inc(&sg->sgp->ref);
7234 if (cpumask_test_cpu(cpu, sg_span))
7244 sd->groups = groups;
7249 free_sched_groups(first, 0);
7254 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
7256 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
7257 struct sched_domain *child = sd->child;
7260 cpu = cpumask_first(sched_domain_span(child));
7263 *sg = *per_cpu_ptr(sdd->sg, cpu);
7264 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
7265 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
7272 * build_sched_groups will build a circular linked list of the groups
7273 * covered by the given span, and will set each group's ->cpumask correctly,
7274 * and ->cpu_power to 0.
7276 * Assumes the sched_domain tree is fully constructed
7279 build_sched_groups(struct sched_domain *sd, int cpu)
7281 struct sched_group *first = NULL, *last = NULL;
7282 struct sd_data *sdd = sd->private;
7283 const struct cpumask *span = sched_domain_span(sd);
7284 struct cpumask *covered;
7287 get_group(cpu, sdd, &sd->groups);
7288 atomic_inc(&sd->groups->ref);
7290 if (cpu != cpumask_first(sched_domain_span(sd)))
7293 lockdep_assert_held(&sched_domains_mutex);
7294 covered = sched_domains_tmpmask;
7296 cpumask_clear(covered);
7298 for_each_cpu(i, span) {
7299 struct sched_group *sg;
7300 int group = get_group(i, sdd, &sg);
7303 if (cpumask_test_cpu(i, covered))
7306 cpumask_clear(sched_group_cpus(sg));
7309 for_each_cpu(j, span) {
7310 if (get_group(j, sdd, NULL) != group)
7313 cpumask_set_cpu(j, covered);
7314 cpumask_set_cpu(j, sched_group_cpus(sg));
7329 * Initialize sched groups cpu_power.
7331 * cpu_power indicates the capacity of sched group, which is used while
7332 * distributing the load between different sched groups in a sched domain.
7333 * Typically cpu_power for all the groups in a sched domain will be same unless
7334 * there are asymmetries in the topology. If there are asymmetries, group
7335 * having more cpu_power will pickup more load compared to the group having
7338 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7340 struct sched_group *sg = sd->groups;
7342 WARN_ON(!sd || !sg);
7345 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
7347 } while (sg != sd->groups);
7349 if (cpu != group_first_cpu(sg))
7352 update_group_power(sd, cpu);
7356 * Initializers for schedule domains
7357 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7360 #ifdef CONFIG_SCHED_DEBUG
7361 # define SD_INIT_NAME(sd, type) sd->name = #type
7363 # define SD_INIT_NAME(sd, type) do { } while (0)
7366 #define SD_INIT_FUNC(type) \
7367 static noinline struct sched_domain * \
7368 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7370 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7371 *sd = SD_##type##_INIT; \
7372 SD_INIT_NAME(sd, type); \
7373 sd->private = &tl->data; \
7379 SD_INIT_FUNC(ALLNODES)
7382 #ifdef CONFIG_SCHED_SMT
7383 SD_INIT_FUNC(SIBLING)
7385 #ifdef CONFIG_SCHED_MC
7388 #ifdef CONFIG_SCHED_BOOK
7392 static int default_relax_domain_level = -1;
7393 int sched_domain_level_max;
7395 static int __init setup_relax_domain_level(char *str)
7399 val = simple_strtoul(str, NULL, 0);
7400 if (val < sched_domain_level_max)
7401 default_relax_domain_level = val;
7405 __setup("relax_domain_level=", setup_relax_domain_level);
7407 static void set_domain_attribute(struct sched_domain *sd,
7408 struct sched_domain_attr *attr)
7412 if (!attr || attr->relax_domain_level < 0) {
7413 if (default_relax_domain_level < 0)
7416 request = default_relax_domain_level;
7418 request = attr->relax_domain_level;
7419 if (request < sd->level) {
7420 /* turn off idle balance on this domain */
7421 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7423 /* turn on idle balance on this domain */
7424 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7428 static void __sdt_free(const struct cpumask *cpu_map);
7429 static int __sdt_alloc(const struct cpumask *cpu_map);
7431 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7432 const struct cpumask *cpu_map)
7436 if (!atomic_read(&d->rd->refcount))
7437 free_rootdomain(&d->rd->rcu); /* fall through */
7439 free_percpu(d->sd); /* fall through */
7441 __sdt_free(cpu_map); /* fall through */
7447 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7448 const struct cpumask *cpu_map)
7450 memset(d, 0, sizeof(*d));
7452 if (__sdt_alloc(cpu_map))
7453 return sa_sd_storage;
7454 d->sd = alloc_percpu(struct sched_domain *);
7456 return sa_sd_storage;
7457 d->rd = alloc_rootdomain();
7460 return sa_rootdomain;
7464 * NULL the sd_data elements we've used to build the sched_domain and
7465 * sched_group structure so that the subsequent __free_domain_allocs()
7466 * will not free the data we're using.
7468 static void claim_allocations(int cpu, struct sched_domain *sd)
7470 struct sd_data *sdd = sd->private;
7472 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7473 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7475 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
7476 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7478 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
7479 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
7482 #ifdef CONFIG_SCHED_SMT
7483 static const struct cpumask *cpu_smt_mask(int cpu)
7485 return topology_thread_cpumask(cpu);
7490 * Topology list, bottom-up.
7492 static struct sched_domain_topology_level default_topology[] = {
7493 #ifdef CONFIG_SCHED_SMT
7494 { sd_init_SIBLING, cpu_smt_mask, },
7496 #ifdef CONFIG_SCHED_MC
7497 { sd_init_MC, cpu_coregroup_mask, },
7499 #ifdef CONFIG_SCHED_BOOK
7500 { sd_init_BOOK, cpu_book_mask, },
7502 { sd_init_CPU, cpu_cpu_mask, },
7504 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
7505 { sd_init_ALLNODES, cpu_allnodes_mask, },
7510 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7512 static int __sdt_alloc(const struct cpumask *cpu_map)
7514 struct sched_domain_topology_level *tl;
7517 for (tl = sched_domain_topology; tl->init; tl++) {
7518 struct sd_data *sdd = &tl->data;
7520 sdd->sd = alloc_percpu(struct sched_domain *);
7524 sdd->sg = alloc_percpu(struct sched_group *);
7528 sdd->sgp = alloc_percpu(struct sched_group_power *);
7532 for_each_cpu(j, cpu_map) {
7533 struct sched_domain *sd;
7534 struct sched_group *sg;
7535 struct sched_group_power *sgp;
7537 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7538 GFP_KERNEL, cpu_to_node(j));
7542 *per_cpu_ptr(sdd->sd, j) = sd;
7544 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7545 GFP_KERNEL, cpu_to_node(j));
7549 *per_cpu_ptr(sdd->sg, j) = sg;
7551 sgp = kzalloc_node(sizeof(struct sched_group_power),
7552 GFP_KERNEL, cpu_to_node(j));
7556 *per_cpu_ptr(sdd->sgp, j) = sgp;
7563 static void __sdt_free(const struct cpumask *cpu_map)
7565 struct sched_domain_topology_level *tl;
7568 for (tl = sched_domain_topology; tl->init; tl++) {
7569 struct sd_data *sdd = &tl->data;
7571 for_each_cpu(j, cpu_map) {
7572 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
7573 if (sd && (sd->flags & SD_OVERLAP))
7574 free_sched_groups(sd->groups, 0);
7575 kfree(*per_cpu_ptr(sdd->sd, j));
7576 kfree(*per_cpu_ptr(sdd->sg, j));
7577 kfree(*per_cpu_ptr(sdd->sgp, j));
7579 free_percpu(sdd->sd);
7580 free_percpu(sdd->sg);
7581 free_percpu(sdd->sgp);
7585 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7586 struct s_data *d, const struct cpumask *cpu_map,
7587 struct sched_domain_attr *attr, struct sched_domain *child,
7590 struct sched_domain *sd = tl->init(tl, cpu);
7594 set_domain_attribute(sd, attr);
7595 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7597 sd->level = child->level + 1;
7598 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7607 * Build sched domains for a given set of cpus and attach the sched domains
7608 * to the individual cpus
7610 static int build_sched_domains(const struct cpumask *cpu_map,
7611 struct sched_domain_attr *attr)
7613 enum s_alloc alloc_state = sa_none;
7614 struct sched_domain *sd;
7616 int i, ret = -ENOMEM;
7618 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7619 if (alloc_state != sa_rootdomain)
7622 /* Set up domains for cpus specified by the cpu_map. */
7623 for_each_cpu(i, cpu_map) {
7624 struct sched_domain_topology_level *tl;
7627 for (tl = sched_domain_topology; tl->init; tl++) {
7628 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7629 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7630 sd->flags |= SD_OVERLAP;
7631 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7638 *per_cpu_ptr(d.sd, i) = sd;
7641 /* Build the groups for the domains */
7642 for_each_cpu(i, cpu_map) {
7643 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7644 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7645 if (sd->flags & SD_OVERLAP) {
7646 if (build_overlap_sched_groups(sd, i))
7649 if (build_sched_groups(sd, i))
7655 /* Calculate CPU power for physical packages and nodes */
7656 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7657 if (!cpumask_test_cpu(i, cpu_map))
7660 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7661 claim_allocations(i, sd);
7662 init_sched_groups_power(i, sd);
7666 /* Attach the domains */
7668 for_each_cpu(i, cpu_map) {
7669 sd = *per_cpu_ptr(d.sd, i);
7670 cpu_attach_domain(sd, d.rd, i);
7676 __free_domain_allocs(&d, alloc_state, cpu_map);
7680 static cpumask_var_t *doms_cur; /* current sched domains */
7681 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7682 static struct sched_domain_attr *dattr_cur;
7683 /* attribues of custom domains in 'doms_cur' */
7686 * Special case: If a kmalloc of a doms_cur partition (array of
7687 * cpumask) fails, then fallback to a single sched domain,
7688 * as determined by the single cpumask fallback_doms.
7690 static cpumask_var_t fallback_doms;
7693 * arch_update_cpu_topology lets virtualized architectures update the
7694 * cpu core maps. It is supposed to return 1 if the topology changed
7695 * or 0 if it stayed the same.
7697 int __attribute__((weak)) arch_update_cpu_topology(void)
7702 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7705 cpumask_var_t *doms;
7707 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7710 for (i = 0; i < ndoms; i++) {
7711 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7712 free_sched_domains(doms, i);
7719 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7722 for (i = 0; i < ndoms; i++)
7723 free_cpumask_var(doms[i]);
7728 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7729 * For now this just excludes isolated cpus, but could be used to
7730 * exclude other special cases in the future.
7732 static int init_sched_domains(const struct cpumask *cpu_map)
7736 arch_update_cpu_topology();
7738 doms_cur = alloc_sched_domains(ndoms_cur);
7740 doms_cur = &fallback_doms;
7741 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7743 err = build_sched_domains(doms_cur[0], NULL);
7744 register_sched_domain_sysctl();
7750 * Detach sched domains from a group of cpus specified in cpu_map
7751 * These cpus will now be attached to the NULL domain
7753 static void detach_destroy_domains(const struct cpumask *cpu_map)
7758 for_each_cpu(i, cpu_map)
7759 cpu_attach_domain(NULL, &def_root_domain, i);
7763 /* handle null as "default" */
7764 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7765 struct sched_domain_attr *new, int idx_new)
7767 struct sched_domain_attr tmp;
7774 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7775 new ? (new + idx_new) : &tmp,
7776 sizeof(struct sched_domain_attr));
7780 * Partition sched domains as specified by the 'ndoms_new'
7781 * cpumasks in the array doms_new[] of cpumasks. This compares
7782 * doms_new[] to the current sched domain partitioning, doms_cur[].
7783 * It destroys each deleted domain and builds each new domain.
7785 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7786 * The masks don't intersect (don't overlap.) We should setup one
7787 * sched domain for each mask. CPUs not in any of the cpumasks will
7788 * not be load balanced. If the same cpumask appears both in the
7789 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7792 * The passed in 'doms_new' should be allocated using
7793 * alloc_sched_domains. This routine takes ownership of it and will
7794 * free_sched_domains it when done with it. If the caller failed the
7795 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7796 * and partition_sched_domains() will fallback to the single partition
7797 * 'fallback_doms', it also forces the domains to be rebuilt.
7799 * If doms_new == NULL it will be replaced with cpu_online_mask.
7800 * ndoms_new == 0 is a special case for destroying existing domains,
7801 * and it will not create the default domain.
7803 * Call with hotplug lock held
7805 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7806 struct sched_domain_attr *dattr_new)
7811 mutex_lock(&sched_domains_mutex);
7813 /* always unregister in case we don't destroy any domains */
7814 unregister_sched_domain_sysctl();
7816 /* Let architecture update cpu core mappings. */
7817 new_topology = arch_update_cpu_topology();
7819 n = doms_new ? ndoms_new : 0;
7821 /* Destroy deleted domains */
7822 for (i = 0; i < ndoms_cur; i++) {
7823 for (j = 0; j < n && !new_topology; j++) {
7824 if (cpumask_equal(doms_cur[i], doms_new[j])
7825 && dattrs_equal(dattr_cur, i, dattr_new, j))
7828 /* no match - a current sched domain not in new doms_new[] */
7829 detach_destroy_domains(doms_cur[i]);
7834 if (doms_new == NULL) {
7836 doms_new = &fallback_doms;
7837 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7838 WARN_ON_ONCE(dattr_new);
7841 /* Build new domains */
7842 for (i = 0; i < ndoms_new; i++) {
7843 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7844 if (cpumask_equal(doms_new[i], doms_cur[j])
7845 && dattrs_equal(dattr_new, i, dattr_cur, j))
7848 /* no match - add a new doms_new */
7849 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7854 /* Remember the new sched domains */
7855 if (doms_cur != &fallback_doms)
7856 free_sched_domains(doms_cur, ndoms_cur);
7857 kfree(dattr_cur); /* kfree(NULL) is safe */
7858 doms_cur = doms_new;
7859 dattr_cur = dattr_new;
7860 ndoms_cur = ndoms_new;
7862 register_sched_domain_sysctl();
7864 mutex_unlock(&sched_domains_mutex);
7867 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7868 static void reinit_sched_domains(void)
7872 /* Destroy domains first to force the rebuild */
7873 partition_sched_domains(0, NULL, NULL);
7875 rebuild_sched_domains();
7879 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7881 unsigned int level = 0;
7883 if (sscanf(buf, "%u", &level) != 1)
7887 * level is always be positive so don't check for
7888 * level < POWERSAVINGS_BALANCE_NONE which is 0
7889 * What happens on 0 or 1 byte write,
7890 * need to check for count as well?
7893 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7897 sched_smt_power_savings = level;
7899 sched_mc_power_savings = level;
7901 reinit_sched_domains();
7906 #ifdef CONFIG_SCHED_MC
7907 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7908 struct sysdev_class_attribute *attr,
7911 return sprintf(page, "%u\n", sched_mc_power_savings);
7913 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7914 struct sysdev_class_attribute *attr,
7915 const char *buf, size_t count)
7917 return sched_power_savings_store(buf, count, 0);
7919 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7920 sched_mc_power_savings_show,
7921 sched_mc_power_savings_store);
7924 #ifdef CONFIG_SCHED_SMT
7925 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7926 struct sysdev_class_attribute *attr,
7929 return sprintf(page, "%u\n", sched_smt_power_savings);
7931 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7932 struct sysdev_class_attribute *attr,
7933 const char *buf, size_t count)
7935 return sched_power_savings_store(buf, count, 1);
7937 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7938 sched_smt_power_savings_show,
7939 sched_smt_power_savings_store);
7942 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7946 #ifdef CONFIG_SCHED_SMT
7948 err = sysfs_create_file(&cls->kset.kobj,
7949 &attr_sched_smt_power_savings.attr);
7951 #ifdef CONFIG_SCHED_MC
7952 if (!err && mc_capable())
7953 err = sysfs_create_file(&cls->kset.kobj,
7954 &attr_sched_mc_power_savings.attr);
7958 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7961 * Update cpusets according to cpu_active mask. If cpusets are
7962 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7963 * around partition_sched_domains().
7965 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7968 switch (action & ~CPU_TASKS_FROZEN) {
7970 case CPU_DOWN_FAILED:
7971 cpuset_update_active_cpus();
7978 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7981 switch (action & ~CPU_TASKS_FROZEN) {
7982 case CPU_DOWN_PREPARE:
7983 cpuset_update_active_cpus();
7990 static int update_runtime(struct notifier_block *nfb,
7991 unsigned long action, void *hcpu)
7993 int cpu = (int)(long)hcpu;
7996 case CPU_DOWN_PREPARE:
7997 case CPU_DOWN_PREPARE_FROZEN:
7998 disable_runtime(cpu_rq(cpu));
8001 case CPU_DOWN_FAILED:
8002 case CPU_DOWN_FAILED_FROZEN:
8004 case CPU_ONLINE_FROZEN:
8005 enable_runtime(cpu_rq(cpu));
8013 void __init sched_init_smp(void)
8015 cpumask_var_t non_isolated_cpus;
8017 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8018 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8021 mutex_lock(&sched_domains_mutex);
8022 init_sched_domains(cpu_active_mask);
8023 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8024 if (cpumask_empty(non_isolated_cpus))
8025 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8026 mutex_unlock(&sched_domains_mutex);
8029 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
8030 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
8032 /* RT runtime code needs to handle some hotplug events */
8033 hotcpu_notifier(update_runtime, 0);
8037 /* Move init over to a non-isolated CPU */
8038 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8040 sched_init_granularity();
8041 free_cpumask_var(non_isolated_cpus);
8043 init_sched_rt_class();
8046 void __init sched_init_smp(void)
8048 sched_init_granularity();
8050 #endif /* CONFIG_SMP */
8052 const_debug unsigned int sysctl_timer_migration = 1;
8054 int in_sched_functions(unsigned long addr)
8056 return in_lock_functions(addr) ||
8057 (addr >= (unsigned long)__sched_text_start
8058 && addr < (unsigned long)__sched_text_end);
8061 static void init_cfs_rq(struct cfs_rq *cfs_rq)
8063 cfs_rq->tasks_timeline = RB_ROOT;
8064 INIT_LIST_HEAD(&cfs_rq->tasks);
8065 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8066 #ifndef CONFIG_64BIT
8067 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8071 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8073 struct rt_prio_array *array;
8076 array = &rt_rq->active;
8077 for (i = 0; i < MAX_RT_PRIO; i++) {
8078 INIT_LIST_HEAD(array->queue + i);
8079 __clear_bit(i, array->bitmap);
8081 /* delimiter for bitsearch: */
8082 __set_bit(MAX_RT_PRIO, array->bitmap);
8084 #if defined CONFIG_SMP
8085 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8086 rt_rq->highest_prio.next = MAX_RT_PRIO;
8087 rt_rq->rt_nr_migratory = 0;
8088 rt_rq->overloaded = 0;
8089 plist_head_init(&rt_rq->pushable_tasks);
8093 rt_rq->rt_throttled = 0;
8094 rt_rq->rt_runtime = 0;
8095 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
8098 #ifdef CONFIG_FAIR_GROUP_SCHED
8099 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8100 struct sched_entity *se, int cpu,
8101 struct sched_entity *parent)
8103 struct rq *rq = cpu_rq(cpu);
8108 /* allow initial update_cfs_load() to truncate */
8109 cfs_rq->load_stamp = 1;
8111 init_cfs_rq_runtime(cfs_rq);
8113 tg->cfs_rq[cpu] = cfs_rq;
8116 /* se could be NULL for root_task_group */
8121 se->cfs_rq = &rq->cfs;
8123 se->cfs_rq = parent->my_q;
8126 update_load_set(&se->load, 0);
8127 se->parent = parent;
8131 #ifdef CONFIG_RT_GROUP_SCHED
8132 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8133 struct sched_rt_entity *rt_se, int cpu,
8134 struct sched_rt_entity *parent)
8136 struct rq *rq = cpu_rq(cpu);
8138 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8139 rt_rq->rt_nr_boosted = 0;
8143 tg->rt_rq[cpu] = rt_rq;
8144 tg->rt_se[cpu] = rt_se;
8150 rt_se->rt_rq = &rq->rt;
8152 rt_se->rt_rq = parent->my_q;
8154 rt_se->my_q = rt_rq;
8155 rt_se->parent = parent;
8156 INIT_LIST_HEAD(&rt_se->run_list);
8160 void __init sched_init(void)
8163 unsigned long alloc_size = 0, ptr;
8165 #ifdef CONFIG_FAIR_GROUP_SCHED
8166 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8168 #ifdef CONFIG_RT_GROUP_SCHED
8169 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8171 #ifdef CONFIG_CPUMASK_OFFSTACK
8172 alloc_size += num_possible_cpus() * cpumask_size();
8175 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8177 #ifdef CONFIG_FAIR_GROUP_SCHED
8178 root_task_group.se = (struct sched_entity **)ptr;
8179 ptr += nr_cpu_ids * sizeof(void **);
8181 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8182 ptr += nr_cpu_ids * sizeof(void **);
8184 #endif /* CONFIG_FAIR_GROUP_SCHED */
8185 #ifdef CONFIG_RT_GROUP_SCHED
8186 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8187 ptr += nr_cpu_ids * sizeof(void **);
8189 root_task_group.rt_rq = (struct rt_rq **)ptr;
8190 ptr += nr_cpu_ids * sizeof(void **);
8192 #endif /* CONFIG_RT_GROUP_SCHED */
8193 #ifdef CONFIG_CPUMASK_OFFSTACK
8194 for_each_possible_cpu(i) {
8195 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8196 ptr += cpumask_size();
8198 #endif /* CONFIG_CPUMASK_OFFSTACK */
8202 init_defrootdomain();
8205 init_rt_bandwidth(&def_rt_bandwidth,
8206 global_rt_period(), global_rt_runtime());
8208 #ifdef CONFIG_RT_GROUP_SCHED
8209 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8210 global_rt_period(), global_rt_runtime());
8211 #endif /* CONFIG_RT_GROUP_SCHED */
8213 #ifdef CONFIG_CGROUP_SCHED
8214 list_add(&root_task_group.list, &task_groups);
8215 INIT_LIST_HEAD(&root_task_group.children);
8216 autogroup_init(&init_task);
8217 #endif /* CONFIG_CGROUP_SCHED */
8219 for_each_possible_cpu(i) {
8223 raw_spin_lock_init(&rq->lock);
8225 rq->calc_load_active = 0;
8226 rq->calc_load_update = jiffies + LOAD_FREQ;
8227 init_cfs_rq(&rq->cfs);
8228 init_rt_rq(&rq->rt, rq);
8229 #ifdef CONFIG_FAIR_GROUP_SCHED
8230 root_task_group.shares = root_task_group_load;
8231 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8233 * How much cpu bandwidth does root_task_group get?
8235 * In case of task-groups formed thr' the cgroup filesystem, it
8236 * gets 100% of the cpu resources in the system. This overall
8237 * system cpu resource is divided among the tasks of
8238 * root_task_group and its child task-groups in a fair manner,
8239 * based on each entity's (task or task-group's) weight
8240 * (se->load.weight).
8242 * In other words, if root_task_group has 10 tasks of weight
8243 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8244 * then A0's share of the cpu resource is:
8246 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8248 * We achieve this by letting root_task_group's tasks sit
8249 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8251 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
8252 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8253 #endif /* CONFIG_FAIR_GROUP_SCHED */
8255 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8256 #ifdef CONFIG_RT_GROUP_SCHED
8257 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8258 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8261 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8262 rq->cpu_load[j] = 0;
8264 rq->last_load_update_tick = jiffies;
8269 rq->cpu_power = SCHED_POWER_SCALE;
8270 rq->post_schedule = 0;
8271 rq->active_balance = 0;
8272 rq->next_balance = jiffies;
8277 rq->avg_idle = 2*sysctl_sched_migration_cost;
8278 rq_attach_root(rq, &def_root_domain);
8280 rq->nohz_balance_kick = 0;
8281 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8285 atomic_set(&rq->nr_iowait, 0);
8288 set_load_weight(&init_task);
8290 #ifdef CONFIG_PREEMPT_NOTIFIERS
8291 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8295 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8298 #ifdef CONFIG_RT_MUTEXES
8299 plist_head_init(&init_task.pi_waiters);
8303 * The boot idle thread does lazy MMU switching as well:
8305 atomic_inc(&init_mm.mm_count);
8306 enter_lazy_tlb(&init_mm, current);
8309 * Make us the idle thread. Technically, schedule() should not be
8310 * called from this thread, however somewhere below it might be,
8311 * but because we are the idle thread, we just pick up running again
8312 * when this runqueue becomes "idle".
8314 init_idle(current, smp_processor_id());
8316 calc_load_update = jiffies + LOAD_FREQ;
8319 * During early bootup we pretend to be a normal task:
8321 current->sched_class = &fair_sched_class;
8324 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8326 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8327 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8328 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8329 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8330 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8332 /* May be allocated at isolcpus cmdline parse time */
8333 if (cpu_isolated_map == NULL)
8334 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8337 scheduler_running = 1;
8340 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8341 static inline int preempt_count_equals(int preempt_offset)
8343 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8345 return (nested == preempt_offset);
8348 void __might_sleep(const char *file, int line, int preempt_offset)
8350 static unsigned long prev_jiffy; /* ratelimiting */
8352 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
8353 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8354 system_state != SYSTEM_RUNNING || oops_in_progress)
8356 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8358 prev_jiffy = jiffies;
8361 "BUG: sleeping function called from invalid context at %s:%d\n",
8364 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8365 in_atomic(), irqs_disabled(),
8366 current->pid, current->comm);
8368 debug_show_held_locks(current);
8369 if (irqs_disabled())
8370 print_irqtrace_events(current);
8373 EXPORT_SYMBOL(__might_sleep);
8376 #ifdef CONFIG_MAGIC_SYSRQ
8377 static void normalize_task(struct rq *rq, struct task_struct *p)
8379 const struct sched_class *prev_class = p->sched_class;
8380 int old_prio = p->prio;
8385 deactivate_task(rq, p, 0);
8386 __setscheduler(rq, p, SCHED_NORMAL, 0);
8388 activate_task(rq, p, 0);
8389 resched_task(rq->curr);
8392 check_class_changed(rq, p, prev_class, old_prio);
8395 void normalize_rt_tasks(void)
8397 struct task_struct *g, *p;
8398 unsigned long flags;
8401 read_lock_irqsave(&tasklist_lock, flags);
8402 do_each_thread(g, p) {
8404 * Only normalize user tasks:
8409 p->se.exec_start = 0;
8410 #ifdef CONFIG_SCHEDSTATS
8411 p->se.statistics.wait_start = 0;
8412 p->se.statistics.sleep_start = 0;
8413 p->se.statistics.block_start = 0;
8418 * Renice negative nice level userspace
8421 if (TASK_NICE(p) < 0 && p->mm)
8422 set_user_nice(p, 0);
8426 raw_spin_lock(&p->pi_lock);
8427 rq = __task_rq_lock(p);
8429 normalize_task(rq, p);
8431 __task_rq_unlock(rq);
8432 raw_spin_unlock(&p->pi_lock);
8433 } while_each_thread(g, p);
8435 read_unlock_irqrestore(&tasklist_lock, flags);
8438 #endif /* CONFIG_MAGIC_SYSRQ */
8440 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8442 * These functions are only useful for the IA64 MCA handling, or kdb.
8444 * They can only be called when the whole system has been
8445 * stopped - every CPU needs to be quiescent, and no scheduling
8446 * activity can take place. Using them for anything else would
8447 * be a serious bug, and as a result, they aren't even visible
8448 * under any other configuration.
8452 * curr_task - return the current task for a given cpu.
8453 * @cpu: the processor in question.
8455 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8457 struct task_struct *curr_task(int cpu)
8459 return cpu_curr(cpu);
8462 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8466 * set_curr_task - set the current task for a given cpu.
8467 * @cpu: the processor in question.
8468 * @p: the task pointer to set.
8470 * Description: This function must only be used when non-maskable interrupts
8471 * are serviced on a separate stack. It allows the architecture to switch the
8472 * notion of the current task on a cpu in a non-blocking manner. This function
8473 * must be called with all CPU's synchronized, and interrupts disabled, the
8474 * and caller must save the original value of the current task (see
8475 * curr_task() above) and restore that value before reenabling interrupts and
8476 * re-starting the system.
8478 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8480 void set_curr_task(int cpu, struct task_struct *p)
8487 #ifdef CONFIG_FAIR_GROUP_SCHED
8488 static void free_fair_sched_group(struct task_group *tg)
8492 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8494 for_each_possible_cpu(i) {
8496 kfree(tg->cfs_rq[i]);
8506 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8508 struct cfs_rq *cfs_rq;
8509 struct sched_entity *se;
8512 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8515 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8519 tg->shares = NICE_0_LOAD;
8521 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8523 for_each_possible_cpu(i) {
8524 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8525 GFP_KERNEL, cpu_to_node(i));
8529 se = kzalloc_node(sizeof(struct sched_entity),
8530 GFP_KERNEL, cpu_to_node(i));
8534 init_cfs_rq(cfs_rq);
8535 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8546 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8548 struct rq *rq = cpu_rq(cpu);
8549 unsigned long flags;
8552 * Only empty task groups can be destroyed; so we can speculatively
8553 * check on_list without danger of it being re-added.
8555 if (!tg->cfs_rq[cpu]->on_list)
8558 raw_spin_lock_irqsave(&rq->lock, flags);
8559 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8560 raw_spin_unlock_irqrestore(&rq->lock, flags);
8562 #else /* !CONFIG_FAIR_GROUP_SCHED */
8563 static inline void free_fair_sched_group(struct task_group *tg)
8568 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8573 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8576 #endif /* CONFIG_FAIR_GROUP_SCHED */
8578 #ifdef CONFIG_RT_GROUP_SCHED
8579 static void free_rt_sched_group(struct task_group *tg)
8584 destroy_rt_bandwidth(&tg->rt_bandwidth);
8586 for_each_possible_cpu(i) {
8588 kfree(tg->rt_rq[i]);
8590 kfree(tg->rt_se[i]);
8598 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8600 struct rt_rq *rt_rq;
8601 struct sched_rt_entity *rt_se;
8604 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8607 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8611 init_rt_bandwidth(&tg->rt_bandwidth,
8612 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8614 for_each_possible_cpu(i) {
8615 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8616 GFP_KERNEL, cpu_to_node(i));
8620 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8621 GFP_KERNEL, cpu_to_node(i));
8625 init_rt_rq(rt_rq, cpu_rq(i));
8626 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8627 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8637 #else /* !CONFIG_RT_GROUP_SCHED */
8638 static inline void free_rt_sched_group(struct task_group *tg)
8643 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8647 #endif /* CONFIG_RT_GROUP_SCHED */
8649 #ifdef CONFIG_CGROUP_SCHED
8650 static void free_sched_group(struct task_group *tg)
8652 free_fair_sched_group(tg);
8653 free_rt_sched_group(tg);
8658 /* allocate runqueue etc for a new task group */
8659 struct task_group *sched_create_group(struct task_group *parent)
8661 struct task_group *tg;
8662 unsigned long flags;
8664 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8666 return ERR_PTR(-ENOMEM);
8668 if (!alloc_fair_sched_group(tg, parent))
8671 if (!alloc_rt_sched_group(tg, parent))
8674 spin_lock_irqsave(&task_group_lock, flags);
8675 list_add_rcu(&tg->list, &task_groups);
8677 WARN_ON(!parent); /* root should already exist */
8679 tg->parent = parent;
8680 INIT_LIST_HEAD(&tg->children);
8681 list_add_rcu(&tg->siblings, &parent->children);
8682 spin_unlock_irqrestore(&task_group_lock, flags);
8687 free_sched_group(tg);
8688 return ERR_PTR(-ENOMEM);
8691 /* rcu callback to free various structures associated with a task group */
8692 static void free_sched_group_rcu(struct rcu_head *rhp)
8694 /* now it should be safe to free those cfs_rqs */
8695 free_sched_group(container_of(rhp, struct task_group, rcu));
8698 /* Destroy runqueue etc associated with a task group */
8699 void sched_destroy_group(struct task_group *tg)
8701 unsigned long flags;
8704 /* end participation in shares distribution */
8705 for_each_possible_cpu(i)
8706 unregister_fair_sched_group(tg, i);
8708 spin_lock_irqsave(&task_group_lock, flags);
8709 list_del_rcu(&tg->list);
8710 list_del_rcu(&tg->siblings);
8711 spin_unlock_irqrestore(&task_group_lock, flags);
8713 /* wait for possible concurrent references to cfs_rqs complete */
8714 call_rcu(&tg->rcu, free_sched_group_rcu);
8717 /* change task's runqueue when it moves between groups.
8718 * The caller of this function should have put the task in its new group
8719 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8720 * reflect its new group.
8722 void sched_move_task(struct task_struct *tsk)
8725 unsigned long flags;
8728 rq = task_rq_lock(tsk, &flags);
8730 running = task_current(rq, tsk);
8734 dequeue_task(rq, tsk, 0);
8735 if (unlikely(running))
8736 tsk->sched_class->put_prev_task(rq, tsk);
8738 #ifdef CONFIG_FAIR_GROUP_SCHED
8739 if (tsk->sched_class->task_move_group)
8740 tsk->sched_class->task_move_group(tsk, on_rq);
8743 set_task_rq(tsk, task_cpu(tsk));
8745 if (unlikely(running))
8746 tsk->sched_class->set_curr_task(rq);
8748 enqueue_task(rq, tsk, 0);
8750 task_rq_unlock(rq, tsk, &flags);
8752 #endif /* CONFIG_CGROUP_SCHED */
8754 #ifdef CONFIG_FAIR_GROUP_SCHED
8755 static DEFINE_MUTEX(shares_mutex);
8757 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8760 unsigned long flags;
8763 * We can't change the weight of the root cgroup.
8768 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8770 mutex_lock(&shares_mutex);
8771 if (tg->shares == shares)
8774 tg->shares = shares;
8775 for_each_possible_cpu(i) {
8776 struct rq *rq = cpu_rq(i);
8777 struct sched_entity *se;
8780 /* Propagate contribution to hierarchy */
8781 raw_spin_lock_irqsave(&rq->lock, flags);
8782 for_each_sched_entity(se)
8783 update_cfs_shares(group_cfs_rq(se));
8784 raw_spin_unlock_irqrestore(&rq->lock, flags);
8788 mutex_unlock(&shares_mutex);
8792 unsigned long sched_group_shares(struct task_group *tg)
8798 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
8799 static unsigned long to_ratio(u64 period, u64 runtime)
8801 if (runtime == RUNTIME_INF)
8804 return div64_u64(runtime << 20, period);
8808 #ifdef CONFIG_RT_GROUP_SCHED
8810 * Ensure that the real time constraints are schedulable.
8812 static DEFINE_MUTEX(rt_constraints_mutex);
8814 /* Must be called with tasklist_lock held */
8815 static inline int tg_has_rt_tasks(struct task_group *tg)
8817 struct task_struct *g, *p;
8819 do_each_thread(g, p) {
8820 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8822 } while_each_thread(g, p);
8827 struct rt_schedulable_data {
8828 struct task_group *tg;
8833 static int tg_rt_schedulable(struct task_group *tg, void *data)
8835 struct rt_schedulable_data *d = data;
8836 struct task_group *child;
8837 unsigned long total, sum = 0;
8838 u64 period, runtime;
8840 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8841 runtime = tg->rt_bandwidth.rt_runtime;
8844 period = d->rt_period;
8845 runtime = d->rt_runtime;
8849 * Cannot have more runtime than the period.
8851 if (runtime > period && runtime != RUNTIME_INF)
8855 * Ensure we don't starve existing RT tasks.
8857 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8860 total = to_ratio(period, runtime);
8863 * Nobody can have more than the global setting allows.
8865 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8869 * The sum of our children's runtime should not exceed our own.
8871 list_for_each_entry_rcu(child, &tg->children, siblings) {
8872 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8873 runtime = child->rt_bandwidth.rt_runtime;
8875 if (child == d->tg) {
8876 period = d->rt_period;
8877 runtime = d->rt_runtime;
8880 sum += to_ratio(period, runtime);
8889 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8893 struct rt_schedulable_data data = {
8895 .rt_period = period,
8896 .rt_runtime = runtime,
8900 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8906 static int tg_set_rt_bandwidth(struct task_group *tg,
8907 u64 rt_period, u64 rt_runtime)
8911 mutex_lock(&rt_constraints_mutex);
8912 read_lock(&tasklist_lock);
8913 err = __rt_schedulable(tg, rt_period, rt_runtime);
8917 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8918 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8919 tg->rt_bandwidth.rt_runtime = rt_runtime;
8921 for_each_possible_cpu(i) {
8922 struct rt_rq *rt_rq = tg->rt_rq[i];
8924 raw_spin_lock(&rt_rq->rt_runtime_lock);
8925 rt_rq->rt_runtime = rt_runtime;
8926 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8928 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8930 read_unlock(&tasklist_lock);
8931 mutex_unlock(&rt_constraints_mutex);
8936 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8938 u64 rt_runtime, rt_period;
8940 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8941 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8942 if (rt_runtime_us < 0)
8943 rt_runtime = RUNTIME_INF;
8945 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8948 long sched_group_rt_runtime(struct task_group *tg)
8952 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8955 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8956 do_div(rt_runtime_us, NSEC_PER_USEC);
8957 return rt_runtime_us;
8960 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8962 u64 rt_runtime, rt_period;
8964 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8965 rt_runtime = tg->rt_bandwidth.rt_runtime;
8970 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8973 long sched_group_rt_period(struct task_group *tg)
8977 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8978 do_div(rt_period_us, NSEC_PER_USEC);
8979 return rt_period_us;
8982 static int sched_rt_global_constraints(void)
8984 u64 runtime, period;
8987 if (sysctl_sched_rt_period <= 0)
8990 runtime = global_rt_runtime();
8991 period = global_rt_period();
8994 * Sanity check on the sysctl variables.
8996 if (runtime > period && runtime != RUNTIME_INF)
8999 mutex_lock(&rt_constraints_mutex);
9000 read_lock(&tasklist_lock);
9001 ret = __rt_schedulable(NULL, 0, 0);
9002 read_unlock(&tasklist_lock);
9003 mutex_unlock(&rt_constraints_mutex);
9008 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
9010 /* Don't accept realtime tasks when there is no way for them to run */
9011 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
9017 #else /* !CONFIG_RT_GROUP_SCHED */
9018 static int sched_rt_global_constraints(void)
9020 unsigned long flags;
9023 if (sysctl_sched_rt_period <= 0)
9027 * There's always some RT tasks in the root group
9028 * -- migration, kstopmachine etc..
9030 if (sysctl_sched_rt_runtime == 0)
9033 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9034 for_each_possible_cpu(i) {
9035 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9037 raw_spin_lock(&rt_rq->rt_runtime_lock);
9038 rt_rq->rt_runtime = global_rt_runtime();
9039 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9041 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9045 #endif /* CONFIG_RT_GROUP_SCHED */
9047 int sched_rt_handler(struct ctl_table *table, int write,
9048 void __user *buffer, size_t *lenp,
9052 int old_period, old_runtime;
9053 static DEFINE_MUTEX(mutex);
9056 old_period = sysctl_sched_rt_period;
9057 old_runtime = sysctl_sched_rt_runtime;
9059 ret = proc_dointvec(table, write, buffer, lenp, ppos);
9061 if (!ret && write) {
9062 ret = sched_rt_global_constraints();
9064 sysctl_sched_rt_period = old_period;
9065 sysctl_sched_rt_runtime = old_runtime;
9067 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9068 def_rt_bandwidth.rt_period =
9069 ns_to_ktime(global_rt_period());
9072 mutex_unlock(&mutex);
9077 #ifdef CONFIG_CGROUP_SCHED
9079 /* return corresponding task_group object of a cgroup */
9080 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9082 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9083 struct task_group, css);
9086 static struct cgroup_subsys_state *
9087 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9089 struct task_group *tg, *parent;
9091 if (!cgrp->parent) {
9092 /* This is early initialization for the top cgroup */
9093 return &root_task_group.css;
9096 parent = cgroup_tg(cgrp->parent);
9097 tg = sched_create_group(parent);
9099 return ERR_PTR(-ENOMEM);
9105 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9107 struct task_group *tg = cgroup_tg(cgrp);
9109 sched_destroy_group(tg);
9113 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct cgroup *old_cgrp,
9114 struct task_struct *tsk)
9116 #ifdef CONFIG_RT_GROUP_SCHED
9117 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9120 /* We don't support RT-tasks being in separate groups */
9121 if (tsk->sched_class != &fair_sched_class)
9128 cpu_cgroup_attach_task(struct cgroup *cgrp, struct cgroup *old_cgrp,
9129 struct task_struct *tsk)
9131 sched_move_task(tsk);
9135 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
9136 struct cgroup *old_cgrp, struct task_struct *task)
9139 * cgroup_exit() is called in the copy_process() failure path.
9140 * Ignore this case since the task hasn't ran yet, this avoids
9141 * trying to poke a half freed task state from generic code.
9143 if (!(task->flags & PF_EXITING))
9146 sched_move_task(task);
9149 #ifdef CONFIG_FAIR_GROUP_SCHED
9150 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9153 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
9156 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9158 struct task_group *tg = cgroup_tg(cgrp);
9160 return (u64) scale_load_down(tg->shares);
9163 #ifdef CONFIG_CFS_BANDWIDTH
9164 static DEFINE_MUTEX(cfs_constraints_mutex);
9166 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
9167 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
9169 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
9171 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
9173 int i, ret = 0, runtime_enabled;
9174 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9176 if (tg == &root_task_group)
9180 * Ensure we have at some amount of bandwidth every period. This is
9181 * to prevent reaching a state of large arrears when throttled via
9182 * entity_tick() resulting in prolonged exit starvation.
9184 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
9188 * Likewise, bound things on the otherside by preventing insane quota
9189 * periods. This also allows us to normalize in computing quota
9192 if (period > max_cfs_quota_period)
9195 mutex_lock(&cfs_constraints_mutex);
9196 ret = __cfs_schedulable(tg, period, quota);
9200 runtime_enabled = quota != RUNTIME_INF;
9201 raw_spin_lock_irq(&cfs_b->lock);
9202 cfs_b->period = ns_to_ktime(period);
9203 cfs_b->quota = quota;
9205 __refill_cfs_bandwidth_runtime(cfs_b);
9206 /* restart the period timer (if active) to handle new period expiry */
9207 if (runtime_enabled && cfs_b->timer_active) {
9208 /* force a reprogram */
9209 cfs_b->timer_active = 0;
9210 __start_cfs_bandwidth(cfs_b);
9212 raw_spin_unlock_irq(&cfs_b->lock);
9214 for_each_possible_cpu(i) {
9215 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
9216 struct rq *rq = rq_of(cfs_rq);
9218 raw_spin_lock_irq(&rq->lock);
9219 cfs_rq->runtime_enabled = runtime_enabled;
9220 cfs_rq->runtime_remaining = 0;
9222 if (cfs_rq_throttled(cfs_rq))
9223 unthrottle_cfs_rq(cfs_rq);
9224 raw_spin_unlock_irq(&rq->lock);
9227 mutex_unlock(&cfs_constraints_mutex);
9232 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
9236 period = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9237 if (cfs_quota_us < 0)
9238 quota = RUNTIME_INF;
9240 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
9242 return tg_set_cfs_bandwidth(tg, period, quota);
9245 long tg_get_cfs_quota(struct task_group *tg)
9249 if (tg_cfs_bandwidth(tg)->quota == RUNTIME_INF)
9252 quota_us = tg_cfs_bandwidth(tg)->quota;
9253 do_div(quota_us, NSEC_PER_USEC);
9258 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
9262 period = (u64)cfs_period_us * NSEC_PER_USEC;
9263 quota = tg_cfs_bandwidth(tg)->quota;
9268 return tg_set_cfs_bandwidth(tg, period, quota);
9271 long tg_get_cfs_period(struct task_group *tg)
9275 cfs_period_us = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9276 do_div(cfs_period_us, NSEC_PER_USEC);
9278 return cfs_period_us;
9281 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
9283 return tg_get_cfs_quota(cgroup_tg(cgrp));
9286 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
9289 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
9292 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
9294 return tg_get_cfs_period(cgroup_tg(cgrp));
9297 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9300 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
9303 struct cfs_schedulable_data {
9304 struct task_group *tg;
9309 * normalize group quota/period to be quota/max_period
9310 * note: units are usecs
9312 static u64 normalize_cfs_quota(struct task_group *tg,
9313 struct cfs_schedulable_data *d)
9321 period = tg_get_cfs_period(tg);
9322 quota = tg_get_cfs_quota(tg);
9325 /* note: these should typically be equivalent */
9326 if (quota == RUNTIME_INF || quota == -1)
9329 return to_ratio(period, quota);
9332 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
9334 struct cfs_schedulable_data *d = data;
9335 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9336 s64 quota = 0, parent_quota = -1;
9339 quota = RUNTIME_INF;
9341 struct cfs_bandwidth *parent_b = tg_cfs_bandwidth(tg->parent);
9343 quota = normalize_cfs_quota(tg, d);
9344 parent_quota = parent_b->hierarchal_quota;
9347 * ensure max(child_quota) <= parent_quota, inherit when no
9350 if (quota == RUNTIME_INF)
9351 quota = parent_quota;
9352 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
9355 cfs_b->hierarchal_quota = quota;
9360 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
9363 struct cfs_schedulable_data data = {
9369 if (quota != RUNTIME_INF) {
9370 do_div(data.period, NSEC_PER_USEC);
9371 do_div(data.quota, NSEC_PER_USEC);
9375 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
9381 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
9382 struct cgroup_map_cb *cb)
9384 struct task_group *tg = cgroup_tg(cgrp);
9385 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9387 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
9388 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
9389 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
9393 #endif /* CONFIG_CFS_BANDWIDTH */
9394 #endif /* CONFIG_FAIR_GROUP_SCHED */
9396 #ifdef CONFIG_RT_GROUP_SCHED
9397 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9400 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9403 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9405 return sched_group_rt_runtime(cgroup_tg(cgrp));
9408 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9411 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9414 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9416 return sched_group_rt_period(cgroup_tg(cgrp));
9418 #endif /* CONFIG_RT_GROUP_SCHED */
9420 static struct cftype cpu_files[] = {
9421 #ifdef CONFIG_FAIR_GROUP_SCHED
9424 .read_u64 = cpu_shares_read_u64,
9425 .write_u64 = cpu_shares_write_u64,
9428 #ifdef CONFIG_CFS_BANDWIDTH
9430 .name = "cfs_quota_us",
9431 .read_s64 = cpu_cfs_quota_read_s64,
9432 .write_s64 = cpu_cfs_quota_write_s64,
9435 .name = "cfs_period_us",
9436 .read_u64 = cpu_cfs_period_read_u64,
9437 .write_u64 = cpu_cfs_period_write_u64,
9441 .read_map = cpu_stats_show,
9444 #ifdef CONFIG_RT_GROUP_SCHED
9446 .name = "rt_runtime_us",
9447 .read_s64 = cpu_rt_runtime_read,
9448 .write_s64 = cpu_rt_runtime_write,
9451 .name = "rt_period_us",
9452 .read_u64 = cpu_rt_period_read_uint,
9453 .write_u64 = cpu_rt_period_write_uint,
9458 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9460 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9463 struct cgroup_subsys cpu_cgroup_subsys = {
9465 .create = cpu_cgroup_create,
9466 .destroy = cpu_cgroup_destroy,
9467 .can_attach_task = cpu_cgroup_can_attach_task,
9468 .attach_task = cpu_cgroup_attach_task,
9469 .exit = cpu_cgroup_exit,
9470 .populate = cpu_cgroup_populate,
9471 .subsys_id = cpu_cgroup_subsys_id,
9475 #endif /* CONFIG_CGROUP_SCHED */
9477 #ifdef CONFIG_CGROUP_CPUACCT
9480 * CPU accounting code for task groups.
9482 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9483 * (balbir@in.ibm.com).
9486 /* track cpu usage of a group of tasks and its child groups */
9488 struct cgroup_subsys_state css;
9489 /* cpuusage holds pointer to a u64-type object on every cpu */
9490 u64 __percpu *cpuusage;
9491 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9492 struct cpuacct *parent;
9495 struct cgroup_subsys cpuacct_subsys;
9497 /* return cpu accounting group corresponding to this container */
9498 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9500 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9501 struct cpuacct, css);
9504 /* return cpu accounting group to which this task belongs */
9505 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9507 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9508 struct cpuacct, css);
9511 /* create a new cpu accounting group */
9512 static struct cgroup_subsys_state *cpuacct_create(
9513 struct cgroup_subsys *ss, struct cgroup *cgrp)
9515 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9521 ca->cpuusage = alloc_percpu(u64);
9525 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9526 if (percpu_counter_init(&ca->cpustat[i], 0))
9527 goto out_free_counters;
9530 ca->parent = cgroup_ca(cgrp->parent);
9536 percpu_counter_destroy(&ca->cpustat[i]);
9537 free_percpu(ca->cpuusage);
9541 return ERR_PTR(-ENOMEM);
9544 /* destroy an existing cpu accounting group */
9546 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9548 struct cpuacct *ca = cgroup_ca(cgrp);
9551 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9552 percpu_counter_destroy(&ca->cpustat[i]);
9553 free_percpu(ca->cpuusage);
9557 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9559 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9562 #ifndef CONFIG_64BIT
9564 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9566 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9568 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9576 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9578 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9580 #ifndef CONFIG_64BIT
9582 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9584 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9586 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9592 /* return total cpu usage (in nanoseconds) of a group */
9593 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9595 struct cpuacct *ca = cgroup_ca(cgrp);
9596 u64 totalcpuusage = 0;
9599 for_each_present_cpu(i)
9600 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9602 return totalcpuusage;
9605 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9608 struct cpuacct *ca = cgroup_ca(cgrp);
9617 for_each_present_cpu(i)
9618 cpuacct_cpuusage_write(ca, i, 0);
9624 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9627 struct cpuacct *ca = cgroup_ca(cgroup);
9631 for_each_present_cpu(i) {
9632 percpu = cpuacct_cpuusage_read(ca, i);
9633 seq_printf(m, "%llu ", (unsigned long long) percpu);
9635 seq_printf(m, "\n");
9639 static const char *cpuacct_stat_desc[] = {
9640 [CPUACCT_STAT_USER] = "user",
9641 [CPUACCT_STAT_SYSTEM] = "system",
9644 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9645 struct cgroup_map_cb *cb)
9647 struct cpuacct *ca = cgroup_ca(cgrp);
9650 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9651 s64 val = percpu_counter_read(&ca->cpustat[i]);
9652 val = cputime64_to_clock_t(val);
9653 cb->fill(cb, cpuacct_stat_desc[i], val);
9658 static struct cftype files[] = {
9661 .read_u64 = cpuusage_read,
9662 .write_u64 = cpuusage_write,
9665 .name = "usage_percpu",
9666 .read_seq_string = cpuacct_percpu_seq_read,
9670 .read_map = cpuacct_stats_show,
9674 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9676 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9680 * charge this task's execution time to its accounting group.
9682 * called with rq->lock held.
9684 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9689 if (unlikely(!cpuacct_subsys.active))
9692 cpu = task_cpu(tsk);
9698 for (; ca; ca = ca->parent) {
9699 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9700 *cpuusage += cputime;
9707 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9708 * in cputime_t units. As a result, cpuacct_update_stats calls
9709 * percpu_counter_add with values large enough to always overflow the
9710 * per cpu batch limit causing bad SMP scalability.
9712 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9713 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9714 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9717 #define CPUACCT_BATCH \
9718 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9720 #define CPUACCT_BATCH 0
9724 * Charge the system/user time to the task's accounting group.
9726 static void cpuacct_update_stats(struct task_struct *tsk,
9727 enum cpuacct_stat_index idx, cputime_t val)
9730 int batch = CPUACCT_BATCH;
9732 if (unlikely(!cpuacct_subsys.active))
9739 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9745 struct cgroup_subsys cpuacct_subsys = {
9747 .create = cpuacct_create,
9748 .destroy = cpuacct_destroy,
9749 .populate = cpuacct_populate,
9750 .subsys_id = cpuacct_subsys_id,
9752 #endif /* CONFIG_CGROUP_CPUACCT */