4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/stop_machine.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/debugfs.h>
71 #include <linux/ctype.h>
72 #include <linux/ftrace.h>
73 #include <linux/slab.h>
76 #include <asm/irq_regs.h>
77 #include <asm/mutex.h>
78 #ifdef CONFIG_PARAVIRT
79 #include <asm/paravirt.h>
82 #include "sched_cpupri.h"
83 #include "workqueue_sched.h"
84 #include "sched_autogroup.h"
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/sched.h>
90 * Convert user-nice values [ -20 ... 0 ... 19 ]
91 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
94 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
95 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
96 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
99 * 'User priority' is the nice value converted to something we
100 * can work with better when scaling various scheduler parameters,
101 * it's a [ 0 ... 39 ] range.
103 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
104 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
105 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
108 * Helpers for converting nanosecond timing to jiffy resolution
110 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
112 #define NICE_0_LOAD SCHED_LOAD_SCALE
113 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
116 * These are the 'tuning knobs' of the scheduler:
118 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
119 * Timeslices get refilled after they expire.
121 #define DEF_TIMESLICE (100 * HZ / 1000)
124 * single value that denotes runtime == period, ie unlimited time.
126 #define RUNTIME_INF ((u64)~0ULL)
128 static inline int rt_policy(int policy)
130 if (policy == SCHED_FIFO || policy == SCHED_RR)
135 static inline int task_has_rt_policy(struct task_struct *p)
137 return rt_policy(p->policy);
141 * This is the priority-queue data structure of the RT scheduling class:
143 struct rt_prio_array {
144 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
145 struct list_head queue[MAX_RT_PRIO];
148 struct rt_bandwidth {
149 /* nests inside the rq lock: */
150 raw_spinlock_t rt_runtime_lock;
153 struct hrtimer rt_period_timer;
156 static struct rt_bandwidth def_rt_bandwidth;
158 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
160 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
162 struct rt_bandwidth *rt_b =
163 container_of(timer, struct rt_bandwidth, rt_period_timer);
169 now = hrtimer_cb_get_time(timer);
170 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
175 idle = do_sched_rt_period_timer(rt_b, overrun);
178 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
182 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
184 rt_b->rt_period = ns_to_ktime(period);
185 rt_b->rt_runtime = runtime;
187 raw_spin_lock_init(&rt_b->rt_runtime_lock);
189 hrtimer_init(&rt_b->rt_period_timer,
190 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
191 rt_b->rt_period_timer.function = sched_rt_period_timer;
194 static inline int rt_bandwidth_enabled(void)
196 return sysctl_sched_rt_runtime >= 0;
199 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
203 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
206 if (hrtimer_active(&rt_b->rt_period_timer))
209 raw_spin_lock(&rt_b->rt_runtime_lock);
214 if (hrtimer_active(&rt_b->rt_period_timer))
217 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
218 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
220 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
221 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
222 delta = ktime_to_ns(ktime_sub(hard, soft));
223 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
224 HRTIMER_MODE_ABS_PINNED, 0);
226 raw_spin_unlock(&rt_b->rt_runtime_lock);
229 #ifdef CONFIG_RT_GROUP_SCHED
230 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
232 hrtimer_cancel(&rt_b->rt_period_timer);
237 * sched_domains_mutex serializes calls to init_sched_domains,
238 * detach_destroy_domains and partition_sched_domains.
240 static DEFINE_MUTEX(sched_domains_mutex);
242 #ifdef CONFIG_CGROUP_SCHED
244 #include <linux/cgroup.h>
248 static LIST_HEAD(task_groups);
250 /* task group related information */
252 struct cgroup_subsys_state css;
254 #ifdef CONFIG_FAIR_GROUP_SCHED
255 /* schedulable entities of this group on each cpu */
256 struct sched_entity **se;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq **cfs_rq;
259 unsigned long shares;
261 atomic_t load_weight;
264 #ifdef CONFIG_RT_GROUP_SCHED
265 struct sched_rt_entity **rt_se;
266 struct rt_rq **rt_rq;
268 struct rt_bandwidth rt_bandwidth;
272 struct list_head list;
274 struct task_group *parent;
275 struct list_head siblings;
276 struct list_head children;
278 #ifdef CONFIG_SCHED_AUTOGROUP
279 struct autogroup *autogroup;
283 /* task_group_lock serializes the addition/removal of task groups */
284 static DEFINE_SPINLOCK(task_group_lock);
286 #ifdef CONFIG_FAIR_GROUP_SCHED
288 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
291 * A weight of 0 or 1 can cause arithmetics problems.
292 * A weight of a cfs_rq is the sum of weights of which entities
293 * are queued on this cfs_rq, so a weight of a entity should not be
294 * too large, so as the shares value of a task group.
295 * (The default weight is 1024 - so there's no practical
296 * limitation from this.)
298 #define MIN_SHARES (1UL << 1)
299 #define MAX_SHARES (1UL << 18)
301 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
304 /* Default task group.
305 * Every task in system belong to this group at bootup.
307 struct task_group root_task_group;
309 #endif /* CONFIG_CGROUP_SCHED */
311 /* CFS-related fields in a runqueue */
313 struct load_weight load;
314 unsigned long nr_running;
319 u64 min_vruntime_copy;
322 struct rb_root tasks_timeline;
323 struct rb_node *rb_leftmost;
325 struct list_head tasks;
326 struct list_head *balance_iterator;
329 * 'curr' points to currently running entity on this cfs_rq.
330 * It is set to NULL otherwise (i.e when none are currently running).
332 struct sched_entity *curr, *next, *last, *skip;
334 #ifdef CONFIG_SCHED_DEBUG
335 unsigned int nr_spread_over;
338 #ifdef CONFIG_FAIR_GROUP_SCHED
339 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
342 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
343 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
344 * (like users, containers etc.)
346 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
347 * list is used during load balance.
350 struct list_head leaf_cfs_rq_list;
351 struct task_group *tg; /* group that "owns" this runqueue */
355 * the part of load.weight contributed by tasks
357 unsigned long task_weight;
360 * h_load = weight * f(tg)
362 * Where f(tg) is the recursive weight fraction assigned to
365 unsigned long h_load;
368 * Maintaining per-cpu shares distribution for group scheduling
370 * load_stamp is the last time we updated the load average
371 * load_last is the last time we updated the load average and saw load
372 * load_unacc_exec_time is currently unaccounted execution time
376 u64 load_stamp, load_last, load_unacc_exec_time;
378 unsigned long load_contribution;
383 /* Real-Time classes' related field in a runqueue: */
385 struct rt_prio_array active;
386 unsigned long rt_nr_running;
387 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
389 int curr; /* highest queued rt task prio */
391 int next; /* next highest */
396 unsigned long rt_nr_migratory;
397 unsigned long rt_nr_total;
399 struct plist_head pushable_tasks;
404 /* Nests inside the rq lock: */
405 raw_spinlock_t rt_runtime_lock;
407 #ifdef CONFIG_RT_GROUP_SCHED
408 unsigned long rt_nr_boosted;
411 struct list_head leaf_rt_rq_list;
412 struct task_group *tg;
419 * We add the notion of a root-domain which will be used to define per-domain
420 * variables. Each exclusive cpuset essentially defines an island domain by
421 * fully partitioning the member cpus from any other cpuset. Whenever a new
422 * exclusive cpuset is created, we also create and attach a new root-domain
431 cpumask_var_t online;
434 * The "RT overload" flag: it gets set if a CPU has more than
435 * one runnable RT task.
437 cpumask_var_t rto_mask;
438 struct cpupri cpupri;
442 * By default the system creates a single root-domain with all cpus as
443 * members (mimicking the global state we have today).
445 static struct root_domain def_root_domain;
447 #endif /* CONFIG_SMP */
450 * This is the main, per-CPU runqueue data structure.
452 * Locking rule: those places that want to lock multiple runqueues
453 * (such as the load balancing or the thread migration code), lock
454 * acquire operations must be ordered by ascending &runqueue.
461 * nr_running and cpu_load should be in the same cacheline because
462 * remote CPUs use both these fields when doing load calculation.
464 unsigned long nr_running;
465 #define CPU_LOAD_IDX_MAX 5
466 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
467 unsigned long last_load_update_tick;
470 unsigned char nohz_balance_kick;
472 int skip_clock_update;
474 /* capture load from *all* tasks on this cpu: */
475 struct load_weight load;
476 unsigned long nr_load_updates;
482 #ifdef CONFIG_FAIR_GROUP_SCHED
483 /* list of leaf cfs_rq on this cpu: */
484 struct list_head leaf_cfs_rq_list;
486 #ifdef CONFIG_RT_GROUP_SCHED
487 struct list_head leaf_rt_rq_list;
491 * This is part of a global counter where only the total sum
492 * over all CPUs matters. A task can increase this counter on
493 * one CPU and if it got migrated afterwards it may decrease
494 * it on another CPU. Always updated under the runqueue lock:
496 unsigned long nr_uninterruptible;
498 struct task_struct *curr, *idle, *stop;
499 unsigned long next_balance;
500 struct mm_struct *prev_mm;
508 struct root_domain *rd;
509 struct sched_domain *sd;
511 unsigned long cpu_power;
513 unsigned char idle_at_tick;
514 /* For active balancing */
518 struct cpu_stop_work active_balance_work;
519 /* cpu of this runqueue: */
523 unsigned long avg_load_per_task;
531 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
534 #ifdef CONFIG_PARAVIRT
537 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
538 u64 prev_steal_time_rq;
541 /* calc_load related fields */
542 unsigned long calc_load_update;
543 long calc_load_active;
545 #ifdef CONFIG_SCHED_HRTICK
547 int hrtick_csd_pending;
548 struct call_single_data hrtick_csd;
550 struct hrtimer hrtick_timer;
553 #ifdef CONFIG_SCHEDSTATS
555 struct sched_info rq_sched_info;
556 unsigned long long rq_cpu_time;
557 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
559 /* sys_sched_yield() stats */
560 unsigned int yld_count;
562 /* schedule() stats */
563 unsigned int sched_switch;
564 unsigned int sched_count;
565 unsigned int sched_goidle;
567 /* try_to_wake_up() stats */
568 unsigned int ttwu_count;
569 unsigned int ttwu_local;
573 struct task_struct *wake_list;
577 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
580 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
582 static inline int cpu_of(struct rq *rq)
591 #define rcu_dereference_check_sched_domain(p) \
592 rcu_dereference_check((p), \
593 lockdep_is_held(&sched_domains_mutex))
596 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
597 * See detach_destroy_domains: synchronize_sched for details.
599 * The domain tree of any CPU may only be accessed from within
600 * preempt-disabled sections.
602 #define for_each_domain(cpu, __sd) \
603 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
605 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
606 #define this_rq() (&__get_cpu_var(runqueues))
607 #define task_rq(p) cpu_rq(task_cpu(p))
608 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
609 #define raw_rq() (&__raw_get_cpu_var(runqueues))
611 #ifdef CONFIG_CGROUP_SCHED
614 * Return the group to which this tasks belongs.
616 * We use task_subsys_state_check() and extend the RCU verification with
617 * pi->lock and rq->lock because cpu_cgroup_attach() holds those locks for each
618 * task it moves into the cgroup. Therefore by holding either of those locks,
619 * we pin the task to the current cgroup.
621 static inline struct task_group *task_group(struct task_struct *p)
623 struct task_group *tg;
624 struct cgroup_subsys_state *css;
626 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
627 lockdep_is_held(&p->pi_lock) ||
628 lockdep_is_held(&task_rq(p)->lock));
629 tg = container_of(css, struct task_group, css);
631 return autogroup_task_group(p, tg);
634 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
635 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
637 #ifdef CONFIG_FAIR_GROUP_SCHED
638 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
639 p->se.parent = task_group(p)->se[cpu];
642 #ifdef CONFIG_RT_GROUP_SCHED
643 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
644 p->rt.parent = task_group(p)->rt_se[cpu];
648 #else /* CONFIG_CGROUP_SCHED */
650 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
651 static inline struct task_group *task_group(struct task_struct *p)
656 #endif /* CONFIG_CGROUP_SCHED */
658 static void update_rq_clock_task(struct rq *rq, s64 delta);
660 static void update_rq_clock(struct rq *rq)
664 if (rq->skip_clock_update > 0)
667 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
669 update_rq_clock_task(rq, delta);
673 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
675 #ifdef CONFIG_SCHED_DEBUG
676 # define const_debug __read_mostly
678 # define const_debug static const
682 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
683 * @cpu: the processor in question.
685 * This interface allows printk to be called with the runqueue lock
686 * held and know whether or not it is OK to wake up the klogd.
688 int runqueue_is_locked(int cpu)
690 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
694 * Debugging: various feature bits
697 #define SCHED_FEAT(name, enabled) \
698 __SCHED_FEAT_##name ,
701 #include "sched_features.h"
706 #define SCHED_FEAT(name, enabled) \
707 (1UL << __SCHED_FEAT_##name) * enabled |
709 const_debug unsigned int sysctl_sched_features =
710 #include "sched_features.h"
715 #ifdef CONFIG_SCHED_DEBUG
716 #define SCHED_FEAT(name, enabled) \
719 static __read_mostly char *sched_feat_names[] = {
720 #include "sched_features.h"
726 static int sched_feat_show(struct seq_file *m, void *v)
730 for (i = 0; sched_feat_names[i]; i++) {
731 if (!(sysctl_sched_features & (1UL << i)))
733 seq_printf(m, "%s ", sched_feat_names[i]);
741 sched_feat_write(struct file *filp, const char __user *ubuf,
742 size_t cnt, loff_t *ppos)
752 if (copy_from_user(&buf, ubuf, cnt))
758 if (strncmp(cmp, "NO_", 3) == 0) {
763 for (i = 0; sched_feat_names[i]; i++) {
764 if (strcmp(cmp, sched_feat_names[i]) == 0) {
766 sysctl_sched_features &= ~(1UL << i);
768 sysctl_sched_features |= (1UL << i);
773 if (!sched_feat_names[i])
781 static int sched_feat_open(struct inode *inode, struct file *filp)
783 return single_open(filp, sched_feat_show, NULL);
786 static const struct file_operations sched_feat_fops = {
787 .open = sched_feat_open,
788 .write = sched_feat_write,
791 .release = single_release,
794 static __init int sched_init_debug(void)
796 debugfs_create_file("sched_features", 0644, NULL, NULL,
801 late_initcall(sched_init_debug);
805 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
808 * Number of tasks to iterate in a single balance run.
809 * Limited because this is done with IRQs disabled.
811 const_debug unsigned int sysctl_sched_nr_migrate = 32;
814 * period over which we average the RT time consumption, measured
819 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
822 * period over which we measure -rt task cpu usage in us.
825 unsigned int sysctl_sched_rt_period = 1000000;
827 static __read_mostly int scheduler_running;
830 * part of the period that we allow rt tasks to run in us.
833 int sysctl_sched_rt_runtime = 950000;
835 static inline u64 global_rt_period(void)
837 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
840 static inline u64 global_rt_runtime(void)
842 if (sysctl_sched_rt_runtime < 0)
845 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
848 #ifndef prepare_arch_switch
849 # define prepare_arch_switch(next) do { } while (0)
851 #ifndef finish_arch_switch
852 # define finish_arch_switch(prev) do { } while (0)
855 static inline int task_current(struct rq *rq, struct task_struct *p)
857 return rq->curr == p;
860 static inline int task_running(struct rq *rq, struct task_struct *p)
865 return task_current(rq, p);
869 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
870 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
874 * We can optimise this out completely for !SMP, because the
875 * SMP rebalancing from interrupt is the only thing that cares
882 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
886 * After ->on_cpu is cleared, the task can be moved to a different CPU.
887 * We must ensure this doesn't happen until the switch is completely
893 #ifdef CONFIG_DEBUG_SPINLOCK
894 /* this is a valid case when another task releases the spinlock */
895 rq->lock.owner = current;
898 * If we are tracking spinlock dependencies then we have to
899 * fix up the runqueue lock - which gets 'carried over' from
902 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
904 raw_spin_unlock_irq(&rq->lock);
907 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
908 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
912 * We can optimise this out completely for !SMP, because the
913 * SMP rebalancing from interrupt is the only thing that cares
918 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
919 raw_spin_unlock_irq(&rq->lock);
921 raw_spin_unlock(&rq->lock);
925 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
929 * After ->on_cpu is cleared, the task can be moved to a different CPU.
930 * We must ensure this doesn't happen until the switch is completely
936 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
940 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
943 * __task_rq_lock - lock the rq @p resides on.
945 static inline struct rq *__task_rq_lock(struct task_struct *p)
950 lockdep_assert_held(&p->pi_lock);
954 raw_spin_lock(&rq->lock);
955 if (likely(rq == task_rq(p)))
957 raw_spin_unlock(&rq->lock);
962 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
964 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
965 __acquires(p->pi_lock)
971 raw_spin_lock_irqsave(&p->pi_lock, *flags);
973 raw_spin_lock(&rq->lock);
974 if (likely(rq == task_rq(p)))
976 raw_spin_unlock(&rq->lock);
977 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
981 static void __task_rq_unlock(struct rq *rq)
984 raw_spin_unlock(&rq->lock);
988 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
990 __releases(p->pi_lock)
992 raw_spin_unlock(&rq->lock);
993 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
997 * this_rq_lock - lock this runqueue and disable interrupts.
999 static struct rq *this_rq_lock(void)
1000 __acquires(rq->lock)
1004 local_irq_disable();
1006 raw_spin_lock(&rq->lock);
1011 #ifdef CONFIG_SCHED_HRTICK
1013 * Use HR-timers to deliver accurate preemption points.
1015 * Its all a bit involved since we cannot program an hrt while holding the
1016 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1019 * When we get rescheduled we reprogram the hrtick_timer outside of the
1025 * - enabled by features
1026 * - hrtimer is actually high res
1028 static inline int hrtick_enabled(struct rq *rq)
1030 if (!sched_feat(HRTICK))
1032 if (!cpu_active(cpu_of(rq)))
1034 return hrtimer_is_hres_active(&rq->hrtick_timer);
1037 static void hrtick_clear(struct rq *rq)
1039 if (hrtimer_active(&rq->hrtick_timer))
1040 hrtimer_cancel(&rq->hrtick_timer);
1044 * High-resolution timer tick.
1045 * Runs from hardirq context with interrupts disabled.
1047 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1049 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1051 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1053 raw_spin_lock(&rq->lock);
1054 update_rq_clock(rq);
1055 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1056 raw_spin_unlock(&rq->lock);
1058 return HRTIMER_NORESTART;
1063 * called from hardirq (IPI) context
1065 static void __hrtick_start(void *arg)
1067 struct rq *rq = arg;
1069 raw_spin_lock(&rq->lock);
1070 hrtimer_restart(&rq->hrtick_timer);
1071 rq->hrtick_csd_pending = 0;
1072 raw_spin_unlock(&rq->lock);
1076 * Called to set the hrtick timer state.
1078 * called with rq->lock held and irqs disabled
1080 static void hrtick_start(struct rq *rq, u64 delay)
1082 struct hrtimer *timer = &rq->hrtick_timer;
1083 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1085 hrtimer_set_expires(timer, time);
1087 if (rq == this_rq()) {
1088 hrtimer_restart(timer);
1089 } else if (!rq->hrtick_csd_pending) {
1090 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1091 rq->hrtick_csd_pending = 1;
1096 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1098 int cpu = (int)(long)hcpu;
1101 case CPU_UP_CANCELED:
1102 case CPU_UP_CANCELED_FROZEN:
1103 case CPU_DOWN_PREPARE:
1104 case CPU_DOWN_PREPARE_FROZEN:
1106 case CPU_DEAD_FROZEN:
1107 hrtick_clear(cpu_rq(cpu));
1114 static __init void init_hrtick(void)
1116 hotcpu_notifier(hotplug_hrtick, 0);
1120 * Called to set the hrtick timer state.
1122 * called with rq->lock held and irqs disabled
1124 static void hrtick_start(struct rq *rq, u64 delay)
1126 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1127 HRTIMER_MODE_REL_PINNED, 0);
1130 static inline void init_hrtick(void)
1133 #endif /* CONFIG_SMP */
1135 static void init_rq_hrtick(struct rq *rq)
1138 rq->hrtick_csd_pending = 0;
1140 rq->hrtick_csd.flags = 0;
1141 rq->hrtick_csd.func = __hrtick_start;
1142 rq->hrtick_csd.info = rq;
1145 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1146 rq->hrtick_timer.function = hrtick;
1148 #else /* CONFIG_SCHED_HRTICK */
1149 static inline void hrtick_clear(struct rq *rq)
1153 static inline void init_rq_hrtick(struct rq *rq)
1157 static inline void init_hrtick(void)
1160 #endif /* CONFIG_SCHED_HRTICK */
1163 * resched_task - mark a task 'to be rescheduled now'.
1165 * On UP this means the setting of the need_resched flag, on SMP it
1166 * might also involve a cross-CPU call to trigger the scheduler on
1171 #ifndef tsk_is_polling
1172 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1175 static void resched_task(struct task_struct *p)
1179 assert_raw_spin_locked(&task_rq(p)->lock);
1181 if (test_tsk_need_resched(p))
1184 set_tsk_need_resched(p);
1187 if (cpu == smp_processor_id())
1190 /* NEED_RESCHED must be visible before we test polling */
1192 if (!tsk_is_polling(p))
1193 smp_send_reschedule(cpu);
1196 static void resched_cpu(int cpu)
1198 struct rq *rq = cpu_rq(cpu);
1199 unsigned long flags;
1201 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1203 resched_task(cpu_curr(cpu));
1204 raw_spin_unlock_irqrestore(&rq->lock, flags);
1209 * In the semi idle case, use the nearest busy cpu for migrating timers
1210 * from an idle cpu. This is good for power-savings.
1212 * We don't do similar optimization for completely idle system, as
1213 * selecting an idle cpu will add more delays to the timers than intended
1214 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1216 int get_nohz_timer_target(void)
1218 int cpu = smp_processor_id();
1220 struct sched_domain *sd;
1223 for_each_domain(cpu, sd) {
1224 for_each_cpu(i, sched_domain_span(sd)) {
1236 * When add_timer_on() enqueues a timer into the timer wheel of an
1237 * idle CPU then this timer might expire before the next timer event
1238 * which is scheduled to wake up that CPU. In case of a completely
1239 * idle system the next event might even be infinite time into the
1240 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1241 * leaves the inner idle loop so the newly added timer is taken into
1242 * account when the CPU goes back to idle and evaluates the timer
1243 * wheel for the next timer event.
1245 void wake_up_idle_cpu(int cpu)
1247 struct rq *rq = cpu_rq(cpu);
1249 if (cpu == smp_processor_id())
1253 * This is safe, as this function is called with the timer
1254 * wheel base lock of (cpu) held. When the CPU is on the way
1255 * to idle and has not yet set rq->curr to idle then it will
1256 * be serialized on the timer wheel base lock and take the new
1257 * timer into account automatically.
1259 if (rq->curr != rq->idle)
1263 * We can set TIF_RESCHED on the idle task of the other CPU
1264 * lockless. The worst case is that the other CPU runs the
1265 * idle task through an additional NOOP schedule()
1267 set_tsk_need_resched(rq->idle);
1269 /* NEED_RESCHED must be visible before we test polling */
1271 if (!tsk_is_polling(rq->idle))
1272 smp_send_reschedule(cpu);
1275 #endif /* CONFIG_NO_HZ */
1277 static u64 sched_avg_period(void)
1279 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1282 static void sched_avg_update(struct rq *rq)
1284 s64 period = sched_avg_period();
1286 while ((s64)(rq->clock - rq->age_stamp) > period) {
1288 * Inline assembly required to prevent the compiler
1289 * optimising this loop into a divmod call.
1290 * See __iter_div_u64_rem() for another example of this.
1292 asm("" : "+rm" (rq->age_stamp));
1293 rq->age_stamp += period;
1298 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1300 rq->rt_avg += rt_delta;
1301 sched_avg_update(rq);
1304 #else /* !CONFIG_SMP */
1305 static void resched_task(struct task_struct *p)
1307 assert_raw_spin_locked(&task_rq(p)->lock);
1308 set_tsk_need_resched(p);
1311 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1315 static void sched_avg_update(struct rq *rq)
1318 #endif /* CONFIG_SMP */
1320 #if BITS_PER_LONG == 32
1321 # define WMULT_CONST (~0UL)
1323 # define WMULT_CONST (1UL << 32)
1326 #define WMULT_SHIFT 32
1329 * Shift right and round:
1331 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1334 * delta *= weight / lw
1336 static unsigned long
1337 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1338 struct load_weight *lw)
1343 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1344 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1345 * 2^SCHED_LOAD_RESOLUTION.
1347 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
1348 tmp = (u64)delta_exec * scale_load_down(weight);
1350 tmp = (u64)delta_exec;
1352 if (!lw->inv_weight) {
1353 unsigned long w = scale_load_down(lw->weight);
1355 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
1357 else if (unlikely(!w))
1358 lw->inv_weight = WMULT_CONST;
1360 lw->inv_weight = WMULT_CONST / w;
1364 * Check whether we'd overflow the 64-bit multiplication:
1366 if (unlikely(tmp > WMULT_CONST))
1367 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1370 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1372 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1375 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1381 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1387 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1394 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1395 * of tasks with abnormal "nice" values across CPUs the contribution that
1396 * each task makes to its run queue's load is weighted according to its
1397 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1398 * scaled version of the new time slice allocation that they receive on time
1402 #define WEIGHT_IDLEPRIO 3
1403 #define WMULT_IDLEPRIO 1431655765
1406 * Nice levels are multiplicative, with a gentle 10% change for every
1407 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1408 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1409 * that remained on nice 0.
1411 * The "10% effect" is relative and cumulative: from _any_ nice level,
1412 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1413 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1414 * If a task goes up by ~10% and another task goes down by ~10% then
1415 * the relative distance between them is ~25%.)
1417 static const int prio_to_weight[40] = {
1418 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1419 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1420 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1421 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1422 /* 0 */ 1024, 820, 655, 526, 423,
1423 /* 5 */ 335, 272, 215, 172, 137,
1424 /* 10 */ 110, 87, 70, 56, 45,
1425 /* 15 */ 36, 29, 23, 18, 15,
1429 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1431 * In cases where the weight does not change often, we can use the
1432 * precalculated inverse to speed up arithmetics by turning divisions
1433 * into multiplications:
1435 static const u32 prio_to_wmult[40] = {
1436 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1437 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1438 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1439 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1440 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1441 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1442 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1443 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1446 /* Time spent by the tasks of the cpu accounting group executing in ... */
1447 enum cpuacct_stat_index {
1448 CPUACCT_STAT_USER, /* ... user mode */
1449 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1451 CPUACCT_STAT_NSTATS,
1454 #ifdef CONFIG_CGROUP_CPUACCT
1455 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1456 static void cpuacct_update_stats(struct task_struct *tsk,
1457 enum cpuacct_stat_index idx, cputime_t val);
1459 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1460 static inline void cpuacct_update_stats(struct task_struct *tsk,
1461 enum cpuacct_stat_index idx, cputime_t val) {}
1464 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1466 update_load_add(&rq->load, load);
1469 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1471 update_load_sub(&rq->load, load);
1474 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1475 typedef int (*tg_visitor)(struct task_group *, void *);
1478 * Iterate the full tree, calling @down when first entering a node and @up when
1479 * leaving it for the final time.
1481 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1483 struct task_group *parent, *child;
1487 parent = &root_task_group;
1489 ret = (*down)(parent, data);
1492 list_for_each_entry_rcu(child, &parent->children, siblings) {
1499 ret = (*up)(parent, data);
1504 parent = parent->parent;
1513 static int tg_nop(struct task_group *tg, void *data)
1520 /* Used instead of source_load when we know the type == 0 */
1521 static unsigned long weighted_cpuload(const int cpu)
1523 return cpu_rq(cpu)->load.weight;
1527 * Return a low guess at the load of a migration-source cpu weighted
1528 * according to the scheduling class and "nice" value.
1530 * We want to under-estimate the load of migration sources, to
1531 * balance conservatively.
1533 static unsigned long source_load(int cpu, int type)
1535 struct rq *rq = cpu_rq(cpu);
1536 unsigned long total = weighted_cpuload(cpu);
1538 if (type == 0 || !sched_feat(LB_BIAS))
1541 return min(rq->cpu_load[type-1], total);
1545 * Return a high guess at the load of a migration-target cpu weighted
1546 * according to the scheduling class and "nice" value.
1548 static unsigned long target_load(int cpu, int type)
1550 struct rq *rq = cpu_rq(cpu);
1551 unsigned long total = weighted_cpuload(cpu);
1553 if (type == 0 || !sched_feat(LB_BIAS))
1556 return max(rq->cpu_load[type-1], total);
1559 static unsigned long power_of(int cpu)
1561 return cpu_rq(cpu)->cpu_power;
1564 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1566 static unsigned long cpu_avg_load_per_task(int cpu)
1568 struct rq *rq = cpu_rq(cpu);
1569 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1572 rq->avg_load_per_task = rq->load.weight / nr_running;
1574 rq->avg_load_per_task = 0;
1576 return rq->avg_load_per_task;
1579 #ifdef CONFIG_PREEMPT
1581 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1584 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1585 * way at the expense of forcing extra atomic operations in all
1586 * invocations. This assures that the double_lock is acquired using the
1587 * same underlying policy as the spinlock_t on this architecture, which
1588 * reduces latency compared to the unfair variant below. However, it
1589 * also adds more overhead and therefore may reduce throughput.
1591 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1592 __releases(this_rq->lock)
1593 __acquires(busiest->lock)
1594 __acquires(this_rq->lock)
1596 raw_spin_unlock(&this_rq->lock);
1597 double_rq_lock(this_rq, busiest);
1604 * Unfair double_lock_balance: Optimizes throughput at the expense of
1605 * latency by eliminating extra atomic operations when the locks are
1606 * already in proper order on entry. This favors lower cpu-ids and will
1607 * grant the double lock to lower cpus over higher ids under contention,
1608 * regardless of entry order into the function.
1610 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1611 __releases(this_rq->lock)
1612 __acquires(busiest->lock)
1613 __acquires(this_rq->lock)
1617 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1618 if (busiest < this_rq) {
1619 raw_spin_unlock(&this_rq->lock);
1620 raw_spin_lock(&busiest->lock);
1621 raw_spin_lock_nested(&this_rq->lock,
1622 SINGLE_DEPTH_NESTING);
1625 raw_spin_lock_nested(&busiest->lock,
1626 SINGLE_DEPTH_NESTING);
1631 #endif /* CONFIG_PREEMPT */
1634 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1636 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1638 if (unlikely(!irqs_disabled())) {
1639 /* printk() doesn't work good under rq->lock */
1640 raw_spin_unlock(&this_rq->lock);
1644 return _double_lock_balance(this_rq, busiest);
1647 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1648 __releases(busiest->lock)
1650 raw_spin_unlock(&busiest->lock);
1651 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1655 * double_rq_lock - safely lock two runqueues
1657 * Note this does not disable interrupts like task_rq_lock,
1658 * you need to do so manually before calling.
1660 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1661 __acquires(rq1->lock)
1662 __acquires(rq2->lock)
1664 BUG_ON(!irqs_disabled());
1666 raw_spin_lock(&rq1->lock);
1667 __acquire(rq2->lock); /* Fake it out ;) */
1670 raw_spin_lock(&rq1->lock);
1671 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1673 raw_spin_lock(&rq2->lock);
1674 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1680 * double_rq_unlock - safely unlock two runqueues
1682 * Note this does not restore interrupts like task_rq_unlock,
1683 * you need to do so manually after calling.
1685 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1686 __releases(rq1->lock)
1687 __releases(rq2->lock)
1689 raw_spin_unlock(&rq1->lock);
1691 raw_spin_unlock(&rq2->lock);
1693 __release(rq2->lock);
1696 #else /* CONFIG_SMP */
1699 * double_rq_lock - safely lock two runqueues
1701 * Note this does not disable interrupts like task_rq_lock,
1702 * you need to do so manually before calling.
1704 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1705 __acquires(rq1->lock)
1706 __acquires(rq2->lock)
1708 BUG_ON(!irqs_disabled());
1710 raw_spin_lock(&rq1->lock);
1711 __acquire(rq2->lock); /* Fake it out ;) */
1715 * double_rq_unlock - safely unlock two runqueues
1717 * Note this does not restore interrupts like task_rq_unlock,
1718 * you need to do so manually after calling.
1720 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1721 __releases(rq1->lock)
1722 __releases(rq2->lock)
1725 raw_spin_unlock(&rq1->lock);
1726 __release(rq2->lock);
1731 static void calc_load_account_idle(struct rq *this_rq);
1732 static void update_sysctl(void);
1733 static int get_update_sysctl_factor(void);
1734 static void update_cpu_load(struct rq *this_rq);
1736 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1738 set_task_rq(p, cpu);
1741 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1742 * successfuly executed on another CPU. We must ensure that updates of
1743 * per-task data have been completed by this moment.
1746 task_thread_info(p)->cpu = cpu;
1750 static const struct sched_class rt_sched_class;
1752 #define sched_class_highest (&stop_sched_class)
1753 #define for_each_class(class) \
1754 for (class = sched_class_highest; class; class = class->next)
1756 #include "sched_stats.h"
1758 static void inc_nr_running(struct rq *rq)
1763 static void dec_nr_running(struct rq *rq)
1768 static void set_load_weight(struct task_struct *p)
1770 int prio = p->static_prio - MAX_RT_PRIO;
1771 struct load_weight *load = &p->se.load;
1774 * SCHED_IDLE tasks get minimal weight:
1776 if (p->policy == SCHED_IDLE) {
1777 load->weight = scale_load(WEIGHT_IDLEPRIO);
1778 load->inv_weight = WMULT_IDLEPRIO;
1782 load->weight = scale_load(prio_to_weight[prio]);
1783 load->inv_weight = prio_to_wmult[prio];
1786 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1788 update_rq_clock(rq);
1789 sched_info_queued(p);
1790 p->sched_class->enqueue_task(rq, p, flags);
1793 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1795 update_rq_clock(rq);
1796 sched_info_dequeued(p);
1797 p->sched_class->dequeue_task(rq, p, flags);
1801 * activate_task - move a task to the runqueue.
1803 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1805 if (task_contributes_to_load(p))
1806 rq->nr_uninterruptible--;
1808 enqueue_task(rq, p, flags);
1813 * deactivate_task - remove a task from the runqueue.
1815 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1817 if (task_contributes_to_load(p))
1818 rq->nr_uninterruptible++;
1820 dequeue_task(rq, p, flags);
1824 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1827 * There are no locks covering percpu hardirq/softirq time.
1828 * They are only modified in account_system_vtime, on corresponding CPU
1829 * with interrupts disabled. So, writes are safe.
1830 * They are read and saved off onto struct rq in update_rq_clock().
1831 * This may result in other CPU reading this CPU's irq time and can
1832 * race with irq/account_system_vtime on this CPU. We would either get old
1833 * or new value with a side effect of accounting a slice of irq time to wrong
1834 * task when irq is in progress while we read rq->clock. That is a worthy
1835 * compromise in place of having locks on each irq in account_system_time.
1837 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1838 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1840 static DEFINE_PER_CPU(u64, irq_start_time);
1841 static int sched_clock_irqtime;
1843 void enable_sched_clock_irqtime(void)
1845 sched_clock_irqtime = 1;
1848 void disable_sched_clock_irqtime(void)
1850 sched_clock_irqtime = 0;
1853 #ifndef CONFIG_64BIT
1854 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1856 static inline void irq_time_write_begin(void)
1858 __this_cpu_inc(irq_time_seq.sequence);
1862 static inline void irq_time_write_end(void)
1865 __this_cpu_inc(irq_time_seq.sequence);
1868 static inline u64 irq_time_read(int cpu)
1874 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1875 irq_time = per_cpu(cpu_softirq_time, cpu) +
1876 per_cpu(cpu_hardirq_time, cpu);
1877 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1881 #else /* CONFIG_64BIT */
1882 static inline void irq_time_write_begin(void)
1886 static inline void irq_time_write_end(void)
1890 static inline u64 irq_time_read(int cpu)
1892 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1894 #endif /* CONFIG_64BIT */
1897 * Called before incrementing preempt_count on {soft,}irq_enter
1898 * and before decrementing preempt_count on {soft,}irq_exit.
1900 void account_system_vtime(struct task_struct *curr)
1902 unsigned long flags;
1906 if (!sched_clock_irqtime)
1909 local_irq_save(flags);
1911 cpu = smp_processor_id();
1912 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1913 __this_cpu_add(irq_start_time, delta);
1915 irq_time_write_begin();
1917 * We do not account for softirq time from ksoftirqd here.
1918 * We want to continue accounting softirq time to ksoftirqd thread
1919 * in that case, so as not to confuse scheduler with a special task
1920 * that do not consume any time, but still wants to run.
1922 if (hardirq_count())
1923 __this_cpu_add(cpu_hardirq_time, delta);
1924 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
1925 __this_cpu_add(cpu_softirq_time, delta);
1927 irq_time_write_end();
1928 local_irq_restore(flags);
1930 EXPORT_SYMBOL_GPL(account_system_vtime);
1932 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
1934 #ifdef CONFIG_PARAVIRT
1935 static inline u64 steal_ticks(u64 steal)
1937 if (unlikely(steal > NSEC_PER_SEC))
1938 return div_u64(steal, TICK_NSEC);
1940 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
1944 static void update_rq_clock_task(struct rq *rq, s64 delta)
1947 * In theory, the compile should just see 0 here, and optimize out the call
1948 * to sched_rt_avg_update. But I don't trust it...
1950 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
1951 s64 steal = 0, irq_delta = 0;
1953 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1954 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
1957 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1958 * this case when a previous update_rq_clock() happened inside a
1959 * {soft,}irq region.
1961 * When this happens, we stop ->clock_task and only update the
1962 * prev_irq_time stamp to account for the part that fit, so that a next
1963 * update will consume the rest. This ensures ->clock_task is
1966 * It does however cause some slight miss-attribution of {soft,}irq
1967 * time, a more accurate solution would be to update the irq_time using
1968 * the current rq->clock timestamp, except that would require using
1971 if (irq_delta > delta)
1974 rq->prev_irq_time += irq_delta;
1977 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
1978 if (static_branch((¶virt_steal_rq_enabled))) {
1981 steal = paravirt_steal_clock(cpu_of(rq));
1982 steal -= rq->prev_steal_time_rq;
1984 if (unlikely(steal > delta))
1987 st = steal_ticks(steal);
1988 steal = st * TICK_NSEC;
1990 rq->prev_steal_time_rq += steal;
1996 rq->clock_task += delta;
1998 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
1999 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
2000 sched_rt_avg_update(rq, irq_delta + steal);
2004 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2005 static int irqtime_account_hi_update(void)
2007 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2008 unsigned long flags;
2012 local_irq_save(flags);
2013 latest_ns = this_cpu_read(cpu_hardirq_time);
2014 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
2016 local_irq_restore(flags);
2020 static int irqtime_account_si_update(void)
2022 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2023 unsigned long flags;
2027 local_irq_save(flags);
2028 latest_ns = this_cpu_read(cpu_softirq_time);
2029 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2031 local_irq_restore(flags);
2035 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2037 #define sched_clock_irqtime (0)
2041 #include "sched_idletask.c"
2042 #include "sched_fair.c"
2043 #include "sched_rt.c"
2044 #include "sched_autogroup.c"
2045 #include "sched_stoptask.c"
2046 #ifdef CONFIG_SCHED_DEBUG
2047 # include "sched_debug.c"
2050 void sched_set_stop_task(int cpu, struct task_struct *stop)
2052 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2053 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2057 * Make it appear like a SCHED_FIFO task, its something
2058 * userspace knows about and won't get confused about.
2060 * Also, it will make PI more or less work without too
2061 * much confusion -- but then, stop work should not
2062 * rely on PI working anyway.
2064 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2066 stop->sched_class = &stop_sched_class;
2069 cpu_rq(cpu)->stop = stop;
2073 * Reset it back to a normal scheduling class so that
2074 * it can die in pieces.
2076 old_stop->sched_class = &rt_sched_class;
2081 * __normal_prio - return the priority that is based on the static prio
2083 static inline int __normal_prio(struct task_struct *p)
2085 return p->static_prio;
2089 * Calculate the expected normal priority: i.e. priority
2090 * without taking RT-inheritance into account. Might be
2091 * boosted by interactivity modifiers. Changes upon fork,
2092 * setprio syscalls, and whenever the interactivity
2093 * estimator recalculates.
2095 static inline int normal_prio(struct task_struct *p)
2099 if (task_has_rt_policy(p))
2100 prio = MAX_RT_PRIO-1 - p->rt_priority;
2102 prio = __normal_prio(p);
2107 * Calculate the current priority, i.e. the priority
2108 * taken into account by the scheduler. This value might
2109 * be boosted by RT tasks, or might be boosted by
2110 * interactivity modifiers. Will be RT if the task got
2111 * RT-boosted. If not then it returns p->normal_prio.
2113 static int effective_prio(struct task_struct *p)
2115 p->normal_prio = normal_prio(p);
2117 * If we are RT tasks or we were boosted to RT priority,
2118 * keep the priority unchanged. Otherwise, update priority
2119 * to the normal priority:
2121 if (!rt_prio(p->prio))
2122 return p->normal_prio;
2127 * task_curr - is this task currently executing on a CPU?
2128 * @p: the task in question.
2130 inline int task_curr(const struct task_struct *p)
2132 return cpu_curr(task_cpu(p)) == p;
2135 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2136 const struct sched_class *prev_class,
2139 if (prev_class != p->sched_class) {
2140 if (prev_class->switched_from)
2141 prev_class->switched_from(rq, p);
2142 p->sched_class->switched_to(rq, p);
2143 } else if (oldprio != p->prio)
2144 p->sched_class->prio_changed(rq, p, oldprio);
2147 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2149 const struct sched_class *class;
2151 if (p->sched_class == rq->curr->sched_class) {
2152 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2154 for_each_class(class) {
2155 if (class == rq->curr->sched_class)
2157 if (class == p->sched_class) {
2158 resched_task(rq->curr);
2165 * A queue event has occurred, and we're going to schedule. In
2166 * this case, we can save a useless back to back clock update.
2168 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2169 rq->skip_clock_update = 1;
2174 * Is this task likely cache-hot:
2177 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2181 if (p->sched_class != &fair_sched_class)
2184 if (unlikely(p->policy == SCHED_IDLE))
2188 * Buddy candidates are cache hot:
2190 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2191 (&p->se == cfs_rq_of(&p->se)->next ||
2192 &p->se == cfs_rq_of(&p->se)->last))
2195 if (sysctl_sched_migration_cost == -1)
2197 if (sysctl_sched_migration_cost == 0)
2200 delta = now - p->se.exec_start;
2202 return delta < (s64)sysctl_sched_migration_cost;
2205 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2207 #ifdef CONFIG_SCHED_DEBUG
2209 * We should never call set_task_cpu() on a blocked task,
2210 * ttwu() will sort out the placement.
2212 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2213 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2215 #ifdef CONFIG_LOCKDEP
2217 * The caller should hold either p->pi_lock or rq->lock, when changing
2218 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2220 * sched_move_task() holds both and thus holding either pins the cgroup,
2221 * see set_task_rq().
2223 * Furthermore, all task_rq users should acquire both locks, see
2226 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2227 lockdep_is_held(&task_rq(p)->lock)));
2231 trace_sched_migrate_task(p, new_cpu);
2233 if (task_cpu(p) != new_cpu) {
2234 p->se.nr_migrations++;
2235 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
2238 __set_task_cpu(p, new_cpu);
2241 struct migration_arg {
2242 struct task_struct *task;
2246 static int migration_cpu_stop(void *data);
2249 * wait_task_inactive - wait for a thread to unschedule.
2251 * If @match_state is nonzero, it's the @p->state value just checked and
2252 * not expected to change. If it changes, i.e. @p might have woken up,
2253 * then return zero. When we succeed in waiting for @p to be off its CPU,
2254 * we return a positive number (its total switch count). If a second call
2255 * a short while later returns the same number, the caller can be sure that
2256 * @p has remained unscheduled the whole time.
2258 * The caller must ensure that the task *will* unschedule sometime soon,
2259 * else this function might spin for a *long* time. This function can't
2260 * be called with interrupts off, or it may introduce deadlock with
2261 * smp_call_function() if an IPI is sent by the same process we are
2262 * waiting to become inactive.
2264 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2266 unsigned long flags;
2273 * We do the initial early heuristics without holding
2274 * any task-queue locks at all. We'll only try to get
2275 * the runqueue lock when things look like they will
2281 * If the task is actively running on another CPU
2282 * still, just relax and busy-wait without holding
2285 * NOTE! Since we don't hold any locks, it's not
2286 * even sure that "rq" stays as the right runqueue!
2287 * But we don't care, since "task_running()" will
2288 * return false if the runqueue has changed and p
2289 * is actually now running somewhere else!
2291 while (task_running(rq, p)) {
2292 if (match_state && unlikely(p->state != match_state))
2298 * Ok, time to look more closely! We need the rq
2299 * lock now, to be *sure*. If we're wrong, we'll
2300 * just go back and repeat.
2302 rq = task_rq_lock(p, &flags);
2303 trace_sched_wait_task(p);
2304 running = task_running(rq, p);
2307 if (!match_state || p->state == match_state)
2308 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2309 task_rq_unlock(rq, p, &flags);
2312 * If it changed from the expected state, bail out now.
2314 if (unlikely(!ncsw))
2318 * Was it really running after all now that we
2319 * checked with the proper locks actually held?
2321 * Oops. Go back and try again..
2323 if (unlikely(running)) {
2329 * It's not enough that it's not actively running,
2330 * it must be off the runqueue _entirely_, and not
2333 * So if it was still runnable (but just not actively
2334 * running right now), it's preempted, and we should
2335 * yield - it could be a while.
2337 if (unlikely(on_rq)) {
2338 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2340 set_current_state(TASK_UNINTERRUPTIBLE);
2341 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2346 * Ahh, all good. It wasn't running, and it wasn't
2347 * runnable, which means that it will never become
2348 * running in the future either. We're all done!
2357 * kick_process - kick a running thread to enter/exit the kernel
2358 * @p: the to-be-kicked thread
2360 * Cause a process which is running on another CPU to enter
2361 * kernel-mode, without any delay. (to get signals handled.)
2363 * NOTE: this function doesn't have to take the runqueue lock,
2364 * because all it wants to ensure is that the remote task enters
2365 * the kernel. If the IPI races and the task has been migrated
2366 * to another CPU then no harm is done and the purpose has been
2369 void kick_process(struct task_struct *p)
2375 if ((cpu != smp_processor_id()) && task_curr(p))
2376 smp_send_reschedule(cpu);
2379 EXPORT_SYMBOL_GPL(kick_process);
2380 #endif /* CONFIG_SMP */
2384 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2386 static int select_fallback_rq(int cpu, struct task_struct *p)
2389 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2391 /* Look for allowed, online CPU in same node. */
2392 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2393 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2396 /* Any allowed, online CPU? */
2397 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2398 if (dest_cpu < nr_cpu_ids)
2401 /* No more Mr. Nice Guy. */
2402 dest_cpu = cpuset_cpus_allowed_fallback(p);
2404 * Don't tell them about moving exiting tasks or
2405 * kernel threads (both mm NULL), since they never
2408 if (p->mm && printk_ratelimit()) {
2409 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2410 task_pid_nr(p), p->comm, cpu);
2417 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2420 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2422 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2425 * In order not to call set_task_cpu() on a blocking task we need
2426 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2429 * Since this is common to all placement strategies, this lives here.
2431 * [ this allows ->select_task() to simply return task_cpu(p) and
2432 * not worry about this generic constraint ]
2434 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2436 cpu = select_fallback_rq(task_cpu(p), p);
2441 static void update_avg(u64 *avg, u64 sample)
2443 s64 diff = sample - *avg;
2449 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2451 #ifdef CONFIG_SCHEDSTATS
2452 struct rq *rq = this_rq();
2455 int this_cpu = smp_processor_id();
2457 if (cpu == this_cpu) {
2458 schedstat_inc(rq, ttwu_local);
2459 schedstat_inc(p, se.statistics.nr_wakeups_local);
2461 struct sched_domain *sd;
2463 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2465 for_each_domain(this_cpu, sd) {
2466 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2467 schedstat_inc(sd, ttwu_wake_remote);
2474 if (wake_flags & WF_MIGRATED)
2475 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2477 #endif /* CONFIG_SMP */
2479 schedstat_inc(rq, ttwu_count);
2480 schedstat_inc(p, se.statistics.nr_wakeups);
2482 if (wake_flags & WF_SYNC)
2483 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2485 #endif /* CONFIG_SCHEDSTATS */
2488 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2490 activate_task(rq, p, en_flags);
2493 /* if a worker is waking up, notify workqueue */
2494 if (p->flags & PF_WQ_WORKER)
2495 wq_worker_waking_up(p, cpu_of(rq));
2499 * Mark the task runnable and perform wakeup-preemption.
2502 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2504 trace_sched_wakeup(p, true);
2505 check_preempt_curr(rq, p, wake_flags);
2507 p->state = TASK_RUNNING;
2509 if (p->sched_class->task_woken)
2510 p->sched_class->task_woken(rq, p);
2512 if (rq->idle_stamp) {
2513 u64 delta = rq->clock - rq->idle_stamp;
2514 u64 max = 2*sysctl_sched_migration_cost;
2519 update_avg(&rq->avg_idle, delta);
2526 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2529 if (p->sched_contributes_to_load)
2530 rq->nr_uninterruptible--;
2533 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2534 ttwu_do_wakeup(rq, p, wake_flags);
2538 * Called in case the task @p isn't fully descheduled from its runqueue,
2539 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2540 * since all we need to do is flip p->state to TASK_RUNNING, since
2541 * the task is still ->on_rq.
2543 static int ttwu_remote(struct task_struct *p, int wake_flags)
2548 rq = __task_rq_lock(p);
2550 ttwu_do_wakeup(rq, p, wake_flags);
2553 __task_rq_unlock(rq);
2559 static void sched_ttwu_do_pending(struct task_struct *list)
2561 struct rq *rq = this_rq();
2563 raw_spin_lock(&rq->lock);
2566 struct task_struct *p = list;
2567 list = list->wake_entry;
2568 ttwu_do_activate(rq, p, 0);
2571 raw_spin_unlock(&rq->lock);
2574 #ifdef CONFIG_HOTPLUG_CPU
2576 static void sched_ttwu_pending(void)
2578 struct rq *rq = this_rq();
2579 struct task_struct *list = xchg(&rq->wake_list, NULL);
2584 sched_ttwu_do_pending(list);
2587 #endif /* CONFIG_HOTPLUG_CPU */
2589 void scheduler_ipi(void)
2591 struct rq *rq = this_rq();
2592 struct task_struct *list = xchg(&rq->wake_list, NULL);
2598 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2599 * traditionally all their work was done from the interrupt return
2600 * path. Now that we actually do some work, we need to make sure
2603 * Some archs already do call them, luckily irq_enter/exit nest
2606 * Arguably we should visit all archs and update all handlers,
2607 * however a fair share of IPIs are still resched only so this would
2608 * somewhat pessimize the simple resched case.
2611 sched_ttwu_do_pending(list);
2615 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2617 struct rq *rq = cpu_rq(cpu);
2618 struct task_struct *next = rq->wake_list;
2621 struct task_struct *old = next;
2623 p->wake_entry = next;
2624 next = cmpxchg(&rq->wake_list, old, p);
2630 smp_send_reschedule(cpu);
2633 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2634 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2639 rq = __task_rq_lock(p);
2641 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2642 ttwu_do_wakeup(rq, p, wake_flags);
2645 __task_rq_unlock(rq);
2650 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2651 #endif /* CONFIG_SMP */
2653 static void ttwu_queue(struct task_struct *p, int cpu)
2655 struct rq *rq = cpu_rq(cpu);
2657 #if defined(CONFIG_SMP)
2658 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2659 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2660 ttwu_queue_remote(p, cpu);
2665 raw_spin_lock(&rq->lock);
2666 ttwu_do_activate(rq, p, 0);
2667 raw_spin_unlock(&rq->lock);
2671 * try_to_wake_up - wake up a thread
2672 * @p: the thread to be awakened
2673 * @state: the mask of task states that can be woken
2674 * @wake_flags: wake modifier flags (WF_*)
2676 * Put it on the run-queue if it's not already there. The "current"
2677 * thread is always on the run-queue (except when the actual
2678 * re-schedule is in progress), and as such you're allowed to do
2679 * the simpler "current->state = TASK_RUNNING" to mark yourself
2680 * runnable without the overhead of this.
2682 * Returns %true if @p was woken up, %false if it was already running
2683 * or @state didn't match @p's state.
2686 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2688 unsigned long flags;
2689 int cpu, success = 0;
2692 raw_spin_lock_irqsave(&p->pi_lock, flags);
2693 if (!(p->state & state))
2696 success = 1; /* we're going to change ->state */
2699 if (p->on_rq && ttwu_remote(p, wake_flags))
2704 * If the owning (remote) cpu is still in the middle of schedule() with
2705 * this task as prev, wait until its done referencing the task.
2708 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2710 * In case the architecture enables interrupts in
2711 * context_switch(), we cannot busy wait, since that
2712 * would lead to deadlocks when an interrupt hits and
2713 * tries to wake up @prev. So bail and do a complete
2716 if (ttwu_activate_remote(p, wake_flags))
2723 * Pairs with the smp_wmb() in finish_lock_switch().
2727 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2728 p->state = TASK_WAKING;
2730 if (p->sched_class->task_waking)
2731 p->sched_class->task_waking(p);
2733 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2734 if (task_cpu(p) != cpu) {
2735 wake_flags |= WF_MIGRATED;
2736 set_task_cpu(p, cpu);
2738 #endif /* CONFIG_SMP */
2742 ttwu_stat(p, cpu, wake_flags);
2744 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2750 * try_to_wake_up_local - try to wake up a local task with rq lock held
2751 * @p: the thread to be awakened
2753 * Put @p on the run-queue if it's not already there. The caller must
2754 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2757 static void try_to_wake_up_local(struct task_struct *p)
2759 struct rq *rq = task_rq(p);
2761 BUG_ON(rq != this_rq());
2762 BUG_ON(p == current);
2763 lockdep_assert_held(&rq->lock);
2765 if (!raw_spin_trylock(&p->pi_lock)) {
2766 raw_spin_unlock(&rq->lock);
2767 raw_spin_lock(&p->pi_lock);
2768 raw_spin_lock(&rq->lock);
2771 if (!(p->state & TASK_NORMAL))
2775 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2777 ttwu_do_wakeup(rq, p, 0);
2778 ttwu_stat(p, smp_processor_id(), 0);
2780 raw_spin_unlock(&p->pi_lock);
2784 * wake_up_process - Wake up a specific process
2785 * @p: The process to be woken up.
2787 * Attempt to wake up the nominated process and move it to the set of runnable
2788 * processes. Returns 1 if the process was woken up, 0 if it was already
2791 * It may be assumed that this function implies a write memory barrier before
2792 * changing the task state if and only if any tasks are woken up.
2794 int wake_up_process(struct task_struct *p)
2796 return try_to_wake_up(p, TASK_ALL, 0);
2798 EXPORT_SYMBOL(wake_up_process);
2800 int wake_up_state(struct task_struct *p, unsigned int state)
2802 return try_to_wake_up(p, state, 0);
2806 * Perform scheduler related setup for a newly forked process p.
2807 * p is forked by current.
2809 * __sched_fork() is basic setup used by init_idle() too:
2811 static void __sched_fork(struct task_struct *p)
2816 p->se.exec_start = 0;
2817 p->se.sum_exec_runtime = 0;
2818 p->se.prev_sum_exec_runtime = 0;
2819 p->se.nr_migrations = 0;
2821 INIT_LIST_HEAD(&p->se.group_node);
2823 #ifdef CONFIG_SCHEDSTATS
2824 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2827 INIT_LIST_HEAD(&p->rt.run_list);
2829 #ifdef CONFIG_PREEMPT_NOTIFIERS
2830 INIT_HLIST_HEAD(&p->preempt_notifiers);
2835 * fork()/clone()-time setup:
2837 void sched_fork(struct task_struct *p)
2839 unsigned long flags;
2840 int cpu = get_cpu();
2844 * We mark the process as running here. This guarantees that
2845 * nobody will actually run it, and a signal or other external
2846 * event cannot wake it up and insert it on the runqueue either.
2848 p->state = TASK_RUNNING;
2851 * Revert to default priority/policy on fork if requested.
2853 if (unlikely(p->sched_reset_on_fork)) {
2854 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2855 p->policy = SCHED_NORMAL;
2856 p->normal_prio = p->static_prio;
2859 if (PRIO_TO_NICE(p->static_prio) < 0) {
2860 p->static_prio = NICE_TO_PRIO(0);
2861 p->normal_prio = p->static_prio;
2866 * We don't need the reset flag anymore after the fork. It has
2867 * fulfilled its duty:
2869 p->sched_reset_on_fork = 0;
2873 * Make sure we do not leak PI boosting priority to the child.
2875 p->prio = current->normal_prio;
2877 if (!rt_prio(p->prio))
2878 p->sched_class = &fair_sched_class;
2880 if (p->sched_class->task_fork)
2881 p->sched_class->task_fork(p);
2884 * The child is not yet in the pid-hash so no cgroup attach races,
2885 * and the cgroup is pinned to this child due to cgroup_fork()
2886 * is ran before sched_fork().
2888 * Silence PROVE_RCU.
2890 raw_spin_lock_irqsave(&p->pi_lock, flags);
2891 set_task_cpu(p, cpu);
2892 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2894 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2895 if (likely(sched_info_on()))
2896 memset(&p->sched_info, 0, sizeof(p->sched_info));
2898 #if defined(CONFIG_SMP)
2901 #ifdef CONFIG_PREEMPT_COUNT
2902 /* Want to start with kernel preemption disabled. */
2903 task_thread_info(p)->preempt_count = 1;
2906 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2913 * wake_up_new_task - wake up a newly created task for the first time.
2915 * This function will do some initial scheduler statistics housekeeping
2916 * that must be done for every newly created context, then puts the task
2917 * on the runqueue and wakes it.
2919 void wake_up_new_task(struct task_struct *p)
2921 unsigned long flags;
2924 raw_spin_lock_irqsave(&p->pi_lock, flags);
2927 * Fork balancing, do it here and not earlier because:
2928 * - cpus_allowed can change in the fork path
2929 * - any previously selected cpu might disappear through hotplug
2931 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
2934 rq = __task_rq_lock(p);
2935 activate_task(rq, p, 0);
2937 trace_sched_wakeup_new(p, true);
2938 check_preempt_curr(rq, p, WF_FORK);
2940 if (p->sched_class->task_woken)
2941 p->sched_class->task_woken(rq, p);
2943 task_rq_unlock(rq, p, &flags);
2946 #ifdef CONFIG_PREEMPT_NOTIFIERS
2949 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2950 * @notifier: notifier struct to register
2952 void preempt_notifier_register(struct preempt_notifier *notifier)
2954 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2956 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2959 * preempt_notifier_unregister - no longer interested in preemption notifications
2960 * @notifier: notifier struct to unregister
2962 * This is safe to call from within a preemption notifier.
2964 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2966 hlist_del(¬ifier->link);
2968 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2970 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2972 struct preempt_notifier *notifier;
2973 struct hlist_node *node;
2975 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2976 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2980 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2981 struct task_struct *next)
2983 struct preempt_notifier *notifier;
2984 struct hlist_node *node;
2986 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2987 notifier->ops->sched_out(notifier, next);
2990 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2992 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2997 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2998 struct task_struct *next)
3002 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3005 * prepare_task_switch - prepare to switch tasks
3006 * @rq: the runqueue preparing to switch
3007 * @prev: the current task that is being switched out
3008 * @next: the task we are going to switch to.
3010 * This is called with the rq lock held and interrupts off. It must
3011 * be paired with a subsequent finish_task_switch after the context
3014 * prepare_task_switch sets up locking and calls architecture specific
3018 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3019 struct task_struct *next)
3021 sched_info_switch(prev, next);
3022 perf_event_task_sched_out(prev, next);
3023 fire_sched_out_preempt_notifiers(prev, next);
3024 prepare_lock_switch(rq, next);
3025 prepare_arch_switch(next);
3026 trace_sched_switch(prev, next);
3030 * finish_task_switch - clean up after a task-switch
3031 * @rq: runqueue associated with task-switch
3032 * @prev: the thread we just switched away from.
3034 * finish_task_switch must be called after the context switch, paired
3035 * with a prepare_task_switch call before the context switch.
3036 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3037 * and do any other architecture-specific cleanup actions.
3039 * Note that we may have delayed dropping an mm in context_switch(). If
3040 * so, we finish that here outside of the runqueue lock. (Doing it
3041 * with the lock held can cause deadlocks; see schedule() for
3044 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3045 __releases(rq->lock)
3047 struct mm_struct *mm = rq->prev_mm;
3053 * A task struct has one reference for the use as "current".
3054 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3055 * schedule one last time. The schedule call will never return, and
3056 * the scheduled task must drop that reference.
3057 * The test for TASK_DEAD must occur while the runqueue locks are
3058 * still held, otherwise prev could be scheduled on another cpu, die
3059 * there before we look at prev->state, and then the reference would
3061 * Manfred Spraul <manfred@colorfullife.com>
3063 prev_state = prev->state;
3064 finish_arch_switch(prev);
3065 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3066 local_irq_disable();
3067 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3068 perf_event_task_sched_in(prev, current);
3069 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3071 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3072 finish_lock_switch(rq, prev);
3074 fire_sched_in_preempt_notifiers(current);
3077 if (unlikely(prev_state == TASK_DEAD)) {
3079 * Remove function-return probe instances associated with this
3080 * task and put them back on the free list.
3082 kprobe_flush_task(prev);
3083 put_task_struct(prev);
3089 /* assumes rq->lock is held */
3090 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3092 if (prev->sched_class->pre_schedule)
3093 prev->sched_class->pre_schedule(rq, prev);
3096 /* rq->lock is NOT held, but preemption is disabled */
3097 static inline void post_schedule(struct rq *rq)
3099 if (rq->post_schedule) {
3100 unsigned long flags;
3102 raw_spin_lock_irqsave(&rq->lock, flags);
3103 if (rq->curr->sched_class->post_schedule)
3104 rq->curr->sched_class->post_schedule(rq);
3105 raw_spin_unlock_irqrestore(&rq->lock, flags);
3107 rq->post_schedule = 0;
3113 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3117 static inline void post_schedule(struct rq *rq)
3124 * schedule_tail - first thing a freshly forked thread must call.
3125 * @prev: the thread we just switched away from.
3127 asmlinkage void schedule_tail(struct task_struct *prev)
3128 __releases(rq->lock)
3130 struct rq *rq = this_rq();
3132 finish_task_switch(rq, prev);
3135 * FIXME: do we need to worry about rq being invalidated by the
3140 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3141 /* In this case, finish_task_switch does not reenable preemption */
3144 if (current->set_child_tid)
3145 put_user(task_pid_vnr(current), current->set_child_tid);
3149 * context_switch - switch to the new MM and the new
3150 * thread's register state.
3153 context_switch(struct rq *rq, struct task_struct *prev,
3154 struct task_struct *next)
3156 struct mm_struct *mm, *oldmm;
3158 prepare_task_switch(rq, prev, next);
3161 oldmm = prev->active_mm;
3163 * For paravirt, this is coupled with an exit in switch_to to
3164 * combine the page table reload and the switch backend into
3167 arch_start_context_switch(prev);
3170 next->active_mm = oldmm;
3171 atomic_inc(&oldmm->mm_count);
3172 enter_lazy_tlb(oldmm, next);
3174 switch_mm(oldmm, mm, next);
3177 prev->active_mm = NULL;
3178 rq->prev_mm = oldmm;
3181 * Since the runqueue lock will be released by the next
3182 * task (which is an invalid locking op but in the case
3183 * of the scheduler it's an obvious special-case), so we
3184 * do an early lockdep release here:
3186 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3187 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3190 /* Here we just switch the register state and the stack. */
3191 switch_to(prev, next, prev);
3195 * this_rq must be evaluated again because prev may have moved
3196 * CPUs since it called schedule(), thus the 'rq' on its stack
3197 * frame will be invalid.
3199 finish_task_switch(this_rq(), prev);
3203 * nr_running, nr_uninterruptible and nr_context_switches:
3205 * externally visible scheduler statistics: current number of runnable
3206 * threads, current number of uninterruptible-sleeping threads, total
3207 * number of context switches performed since bootup.
3209 unsigned long nr_running(void)
3211 unsigned long i, sum = 0;
3213 for_each_online_cpu(i)
3214 sum += cpu_rq(i)->nr_running;
3219 unsigned long nr_uninterruptible(void)
3221 unsigned long i, sum = 0;
3223 for_each_possible_cpu(i)
3224 sum += cpu_rq(i)->nr_uninterruptible;
3227 * Since we read the counters lockless, it might be slightly
3228 * inaccurate. Do not allow it to go below zero though:
3230 if (unlikely((long)sum < 0))
3236 unsigned long long nr_context_switches(void)
3239 unsigned long long sum = 0;
3241 for_each_possible_cpu(i)
3242 sum += cpu_rq(i)->nr_switches;
3247 unsigned long nr_iowait(void)
3249 unsigned long i, sum = 0;
3251 for_each_possible_cpu(i)
3252 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3257 unsigned long nr_iowait_cpu(int cpu)
3259 struct rq *this = cpu_rq(cpu);
3260 return atomic_read(&this->nr_iowait);
3263 unsigned long this_cpu_load(void)
3265 struct rq *this = this_rq();
3266 return this->cpu_load[0];
3270 /* Variables and functions for calc_load */
3271 static atomic_long_t calc_load_tasks;
3272 static unsigned long calc_load_update;
3273 unsigned long avenrun[3];
3274 EXPORT_SYMBOL(avenrun);
3276 static long calc_load_fold_active(struct rq *this_rq)
3278 long nr_active, delta = 0;
3280 nr_active = this_rq->nr_running;
3281 nr_active += (long) this_rq->nr_uninterruptible;
3283 if (nr_active != this_rq->calc_load_active) {
3284 delta = nr_active - this_rq->calc_load_active;
3285 this_rq->calc_load_active = nr_active;
3291 static unsigned long
3292 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3295 load += active * (FIXED_1 - exp);
3296 load += 1UL << (FSHIFT - 1);
3297 return load >> FSHIFT;
3302 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3304 * When making the ILB scale, we should try to pull this in as well.
3306 static atomic_long_t calc_load_tasks_idle;
3308 static void calc_load_account_idle(struct rq *this_rq)
3312 delta = calc_load_fold_active(this_rq);
3314 atomic_long_add(delta, &calc_load_tasks_idle);
3317 static long calc_load_fold_idle(void)
3322 * Its got a race, we don't care...
3324 if (atomic_long_read(&calc_load_tasks_idle))
3325 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3331 * fixed_power_int - compute: x^n, in O(log n) time
3333 * @x: base of the power
3334 * @frac_bits: fractional bits of @x
3335 * @n: power to raise @x to.
3337 * By exploiting the relation between the definition of the natural power
3338 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3339 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3340 * (where: n_i \elem {0, 1}, the binary vector representing n),
3341 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3342 * of course trivially computable in O(log_2 n), the length of our binary
3345 static unsigned long
3346 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3348 unsigned long result = 1UL << frac_bits;
3353 result += 1UL << (frac_bits - 1);
3354 result >>= frac_bits;
3360 x += 1UL << (frac_bits - 1);
3368 * a1 = a0 * e + a * (1 - e)
3370 * a2 = a1 * e + a * (1 - e)
3371 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3372 * = a0 * e^2 + a * (1 - e) * (1 + e)
3374 * a3 = a2 * e + a * (1 - e)
3375 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3376 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3380 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3381 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3382 * = a0 * e^n + a * (1 - e^n)
3384 * [1] application of the geometric series:
3387 * S_n := \Sum x^i = -------------
3390 static unsigned long
3391 calc_load_n(unsigned long load, unsigned long exp,
3392 unsigned long active, unsigned int n)
3395 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3399 * NO_HZ can leave us missing all per-cpu ticks calling
3400 * calc_load_account_active(), but since an idle CPU folds its delta into
3401 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3402 * in the pending idle delta if our idle period crossed a load cycle boundary.
3404 * Once we've updated the global active value, we need to apply the exponential
3405 * weights adjusted to the number of cycles missed.
3407 static void calc_global_nohz(unsigned long ticks)
3409 long delta, active, n;
3411 if (time_before(jiffies, calc_load_update))
3415 * If we crossed a calc_load_update boundary, make sure to fold
3416 * any pending idle changes, the respective CPUs might have
3417 * missed the tick driven calc_load_account_active() update
3420 delta = calc_load_fold_idle();
3422 atomic_long_add(delta, &calc_load_tasks);
3425 * If we were idle for multiple load cycles, apply them.
3427 if (ticks >= LOAD_FREQ) {
3428 n = ticks / LOAD_FREQ;
3430 active = atomic_long_read(&calc_load_tasks);
3431 active = active > 0 ? active * FIXED_1 : 0;
3433 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3434 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3435 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3437 calc_load_update += n * LOAD_FREQ;
3441 * Its possible the remainder of the above division also crosses
3442 * a LOAD_FREQ period, the regular check in calc_global_load()
3443 * which comes after this will take care of that.
3445 * Consider us being 11 ticks before a cycle completion, and us
3446 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3447 * age us 4 cycles, and the test in calc_global_load() will
3448 * pick up the final one.
3452 static void calc_load_account_idle(struct rq *this_rq)
3456 static inline long calc_load_fold_idle(void)
3461 static void calc_global_nohz(unsigned long ticks)
3467 * get_avenrun - get the load average array
3468 * @loads: pointer to dest load array
3469 * @offset: offset to add
3470 * @shift: shift count to shift the result left
3472 * These values are estimates at best, so no need for locking.
3474 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3476 loads[0] = (avenrun[0] + offset) << shift;
3477 loads[1] = (avenrun[1] + offset) << shift;
3478 loads[2] = (avenrun[2] + offset) << shift;
3482 * calc_load - update the avenrun load estimates 10 ticks after the
3483 * CPUs have updated calc_load_tasks.
3485 void calc_global_load(unsigned long ticks)
3489 calc_global_nohz(ticks);
3491 if (time_before(jiffies, calc_load_update + 10))
3494 active = atomic_long_read(&calc_load_tasks);
3495 active = active > 0 ? active * FIXED_1 : 0;
3497 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3498 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3499 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3501 calc_load_update += LOAD_FREQ;
3505 * Called from update_cpu_load() to periodically update this CPU's
3508 static void calc_load_account_active(struct rq *this_rq)
3512 if (time_before(jiffies, this_rq->calc_load_update))
3515 delta = calc_load_fold_active(this_rq);
3516 delta += calc_load_fold_idle();
3518 atomic_long_add(delta, &calc_load_tasks);
3520 this_rq->calc_load_update += LOAD_FREQ;
3524 * The exact cpuload at various idx values, calculated at every tick would be
3525 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3527 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3528 * on nth tick when cpu may be busy, then we have:
3529 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3530 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3532 * decay_load_missed() below does efficient calculation of
3533 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3534 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3536 * The calculation is approximated on a 128 point scale.
3537 * degrade_zero_ticks is the number of ticks after which load at any
3538 * particular idx is approximated to be zero.
3539 * degrade_factor is a precomputed table, a row for each load idx.
3540 * Each column corresponds to degradation factor for a power of two ticks,
3541 * based on 128 point scale.
3543 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3544 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3546 * With this power of 2 load factors, we can degrade the load n times
3547 * by looking at 1 bits in n and doing as many mult/shift instead of
3548 * n mult/shifts needed by the exact degradation.
3550 #define DEGRADE_SHIFT 7
3551 static const unsigned char
3552 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3553 static const unsigned char
3554 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3555 {0, 0, 0, 0, 0, 0, 0, 0},
3556 {64, 32, 8, 0, 0, 0, 0, 0},
3557 {96, 72, 40, 12, 1, 0, 0},
3558 {112, 98, 75, 43, 15, 1, 0},
3559 {120, 112, 98, 76, 45, 16, 2} };
3562 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3563 * would be when CPU is idle and so we just decay the old load without
3564 * adding any new load.
3566 static unsigned long
3567 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3571 if (!missed_updates)
3574 if (missed_updates >= degrade_zero_ticks[idx])
3578 return load >> missed_updates;
3580 while (missed_updates) {
3581 if (missed_updates % 2)
3582 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3584 missed_updates >>= 1;
3591 * Update rq->cpu_load[] statistics. This function is usually called every
3592 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3593 * every tick. We fix it up based on jiffies.
3595 static void update_cpu_load(struct rq *this_rq)
3597 unsigned long this_load = this_rq->load.weight;
3598 unsigned long curr_jiffies = jiffies;
3599 unsigned long pending_updates;
3602 this_rq->nr_load_updates++;
3604 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3605 if (curr_jiffies == this_rq->last_load_update_tick)
3608 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3609 this_rq->last_load_update_tick = curr_jiffies;
3611 /* Update our load: */
3612 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3613 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3614 unsigned long old_load, new_load;
3616 /* scale is effectively 1 << i now, and >> i divides by scale */
3618 old_load = this_rq->cpu_load[i];
3619 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3620 new_load = this_load;
3622 * Round up the averaging division if load is increasing. This
3623 * prevents us from getting stuck on 9 if the load is 10, for
3626 if (new_load > old_load)
3627 new_load += scale - 1;
3629 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3632 sched_avg_update(this_rq);
3635 static void update_cpu_load_active(struct rq *this_rq)
3637 update_cpu_load(this_rq);
3639 calc_load_account_active(this_rq);
3645 * sched_exec - execve() is a valuable balancing opportunity, because at
3646 * this point the task has the smallest effective memory and cache footprint.
3648 void sched_exec(void)
3650 struct task_struct *p = current;
3651 unsigned long flags;
3654 raw_spin_lock_irqsave(&p->pi_lock, flags);
3655 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3656 if (dest_cpu == smp_processor_id())
3659 if (likely(cpu_active(dest_cpu))) {
3660 struct migration_arg arg = { p, dest_cpu };
3662 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3663 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3667 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3672 DEFINE_PER_CPU(struct kernel_stat, kstat);
3674 EXPORT_PER_CPU_SYMBOL(kstat);
3677 * Return any ns on the sched_clock that have not yet been accounted in
3678 * @p in case that task is currently running.
3680 * Called with task_rq_lock() held on @rq.
3682 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3686 if (task_current(rq, p)) {
3687 update_rq_clock(rq);
3688 ns = rq->clock_task - p->se.exec_start;
3696 unsigned long long task_delta_exec(struct task_struct *p)
3698 unsigned long flags;
3702 rq = task_rq_lock(p, &flags);
3703 ns = do_task_delta_exec(p, rq);
3704 task_rq_unlock(rq, p, &flags);
3710 * Return accounted runtime for the task.
3711 * In case the task is currently running, return the runtime plus current's
3712 * pending runtime that have not been accounted yet.
3714 unsigned long long task_sched_runtime(struct task_struct *p)
3716 unsigned long flags;
3720 rq = task_rq_lock(p, &flags);
3721 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3722 task_rq_unlock(rq, p, &flags);
3728 * Return sum_exec_runtime for the thread group.
3729 * In case the task is currently running, return the sum plus current's
3730 * pending runtime that have not been accounted yet.
3732 * Note that the thread group might have other running tasks as well,
3733 * so the return value not includes other pending runtime that other
3734 * running tasks might have.
3736 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3738 struct task_cputime totals;
3739 unsigned long flags;
3743 rq = task_rq_lock(p, &flags);
3744 thread_group_cputime(p, &totals);
3745 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3746 task_rq_unlock(rq, p, &flags);
3752 * Account user cpu time to a process.
3753 * @p: the process that the cpu time gets accounted to
3754 * @cputime: the cpu time spent in user space since the last update
3755 * @cputime_scaled: cputime scaled by cpu frequency
3757 void account_user_time(struct task_struct *p, cputime_t cputime,
3758 cputime_t cputime_scaled)
3760 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3763 /* Add user time to process. */
3764 p->utime = cputime_add(p->utime, cputime);
3765 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3766 account_group_user_time(p, cputime);
3768 /* Add user time to cpustat. */
3769 tmp = cputime_to_cputime64(cputime);
3770 if (TASK_NICE(p) > 0)
3771 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3773 cpustat->user = cputime64_add(cpustat->user, tmp);
3775 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3776 /* Account for user time used */
3777 acct_update_integrals(p);
3781 * Account guest cpu time to a process.
3782 * @p: the process that the cpu time gets accounted to
3783 * @cputime: the cpu time spent in virtual machine since the last update
3784 * @cputime_scaled: cputime scaled by cpu frequency
3786 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3787 cputime_t cputime_scaled)
3790 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3792 tmp = cputime_to_cputime64(cputime);
3794 /* Add guest time to process. */
3795 p->utime = cputime_add(p->utime, cputime);
3796 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3797 account_group_user_time(p, cputime);
3798 p->gtime = cputime_add(p->gtime, cputime);
3800 /* Add guest time to cpustat. */
3801 if (TASK_NICE(p) > 0) {
3802 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3803 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3805 cpustat->user = cputime64_add(cpustat->user, tmp);
3806 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3811 * Account system cpu time to a process and desired cpustat field
3812 * @p: the process that the cpu time gets accounted to
3813 * @cputime: the cpu time spent in kernel space since the last update
3814 * @cputime_scaled: cputime scaled by cpu frequency
3815 * @target_cputime64: pointer to cpustat field that has to be updated
3818 void __account_system_time(struct task_struct *p, cputime_t cputime,
3819 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3821 cputime64_t tmp = cputime_to_cputime64(cputime);
3823 /* Add system time to process. */
3824 p->stime = cputime_add(p->stime, cputime);
3825 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3826 account_group_system_time(p, cputime);
3828 /* Add system time to cpustat. */
3829 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3830 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3832 /* Account for system time used */
3833 acct_update_integrals(p);
3837 * Account system cpu time to a process.
3838 * @p: the process that the cpu time gets accounted to
3839 * @hardirq_offset: the offset to subtract from hardirq_count()
3840 * @cputime: the cpu time spent in kernel space since the last update
3841 * @cputime_scaled: cputime scaled by cpu frequency
3843 void account_system_time(struct task_struct *p, int hardirq_offset,
3844 cputime_t cputime, cputime_t cputime_scaled)
3846 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3847 cputime64_t *target_cputime64;
3849 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3850 account_guest_time(p, cputime, cputime_scaled);
3854 if (hardirq_count() - hardirq_offset)
3855 target_cputime64 = &cpustat->irq;
3856 else if (in_serving_softirq())
3857 target_cputime64 = &cpustat->softirq;
3859 target_cputime64 = &cpustat->system;
3861 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3865 * Account for involuntary wait time.
3866 * @cputime: the cpu time spent in involuntary wait
3868 void account_steal_time(cputime_t cputime)
3870 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3871 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3873 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3877 * Account for idle time.
3878 * @cputime: the cpu time spent in idle wait
3880 void account_idle_time(cputime_t cputime)
3882 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3883 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3884 struct rq *rq = this_rq();
3886 if (atomic_read(&rq->nr_iowait) > 0)
3887 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3889 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3892 static __always_inline bool steal_account_process_tick(void)
3894 #ifdef CONFIG_PARAVIRT
3895 if (static_branch(¶virt_steal_enabled)) {
3898 steal = paravirt_steal_clock(smp_processor_id());
3899 steal -= this_rq()->prev_steal_time;
3901 st = steal_ticks(steal);
3902 this_rq()->prev_steal_time += st * TICK_NSEC;
3904 account_steal_time(st);
3911 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3913 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3915 * Account a tick to a process and cpustat
3916 * @p: the process that the cpu time gets accounted to
3917 * @user_tick: is the tick from userspace
3918 * @rq: the pointer to rq
3920 * Tick demultiplexing follows the order
3921 * - pending hardirq update
3922 * - pending softirq update
3926 * - check for guest_time
3927 * - else account as system_time
3929 * Check for hardirq is done both for system and user time as there is
3930 * no timer going off while we are on hardirq and hence we may never get an
3931 * opportunity to update it solely in system time.
3932 * p->stime and friends are only updated on system time and not on irq
3933 * softirq as those do not count in task exec_runtime any more.
3935 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3938 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3939 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
3940 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3942 if (steal_account_process_tick())
3945 if (irqtime_account_hi_update()) {
3946 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3947 } else if (irqtime_account_si_update()) {
3948 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3949 } else if (this_cpu_ksoftirqd() == p) {
3951 * ksoftirqd time do not get accounted in cpu_softirq_time.
3952 * So, we have to handle it separately here.
3953 * Also, p->stime needs to be updated for ksoftirqd.
3955 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3957 } else if (user_tick) {
3958 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3959 } else if (p == rq->idle) {
3960 account_idle_time(cputime_one_jiffy);
3961 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3962 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3964 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3969 static void irqtime_account_idle_ticks(int ticks)
3972 struct rq *rq = this_rq();
3974 for (i = 0; i < ticks; i++)
3975 irqtime_account_process_tick(current, 0, rq);
3977 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
3978 static void irqtime_account_idle_ticks(int ticks) {}
3979 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3981 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3984 * Account a single tick of cpu time.
3985 * @p: the process that the cpu time gets accounted to
3986 * @user_tick: indicates if the tick is a user or a system tick
3988 void account_process_tick(struct task_struct *p, int user_tick)
3990 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3991 struct rq *rq = this_rq();
3993 if (sched_clock_irqtime) {
3994 irqtime_account_process_tick(p, user_tick, rq);
3998 if (steal_account_process_tick())
4002 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4003 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
4004 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
4007 account_idle_time(cputime_one_jiffy);
4011 * Account multiple ticks of steal time.
4012 * @p: the process from which the cpu time has been stolen
4013 * @ticks: number of stolen ticks
4015 void account_steal_ticks(unsigned long ticks)
4017 account_steal_time(jiffies_to_cputime(ticks));
4021 * Account multiple ticks of idle time.
4022 * @ticks: number of stolen ticks
4024 void account_idle_ticks(unsigned long ticks)
4027 if (sched_clock_irqtime) {
4028 irqtime_account_idle_ticks(ticks);
4032 account_idle_time(jiffies_to_cputime(ticks));
4038 * Use precise platform statistics if available:
4040 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4041 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4047 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4049 struct task_cputime cputime;
4051 thread_group_cputime(p, &cputime);
4053 *ut = cputime.utime;
4054 *st = cputime.stime;
4058 #ifndef nsecs_to_cputime
4059 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
4062 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4064 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
4067 * Use CFS's precise accounting:
4069 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
4075 do_div(temp, total);
4076 utime = (cputime_t)temp;
4081 * Compare with previous values, to keep monotonicity:
4083 p->prev_utime = max(p->prev_utime, utime);
4084 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4086 *ut = p->prev_utime;
4087 *st = p->prev_stime;
4091 * Must be called with siglock held.
4093 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4095 struct signal_struct *sig = p->signal;
4096 struct task_cputime cputime;
4097 cputime_t rtime, utime, total;
4099 thread_group_cputime(p, &cputime);
4101 total = cputime_add(cputime.utime, cputime.stime);
4102 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4107 temp *= cputime.utime;
4108 do_div(temp, total);
4109 utime = (cputime_t)temp;
4113 sig->prev_utime = max(sig->prev_utime, utime);
4114 sig->prev_stime = max(sig->prev_stime,
4115 cputime_sub(rtime, sig->prev_utime));
4117 *ut = sig->prev_utime;
4118 *st = sig->prev_stime;
4123 * This function gets called by the timer code, with HZ frequency.
4124 * We call it with interrupts disabled.
4126 void scheduler_tick(void)
4128 int cpu = smp_processor_id();
4129 struct rq *rq = cpu_rq(cpu);
4130 struct task_struct *curr = rq->curr;
4134 raw_spin_lock(&rq->lock);
4135 update_rq_clock(rq);
4136 update_cpu_load_active(rq);
4137 curr->sched_class->task_tick(rq, curr, 0);
4138 raw_spin_unlock(&rq->lock);
4140 perf_event_task_tick();
4143 rq->idle_at_tick = idle_cpu(cpu);
4144 trigger_load_balance(rq, cpu);
4148 notrace unsigned long get_parent_ip(unsigned long addr)
4150 if (in_lock_functions(addr)) {
4151 addr = CALLER_ADDR2;
4152 if (in_lock_functions(addr))
4153 addr = CALLER_ADDR3;
4158 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4159 defined(CONFIG_PREEMPT_TRACER))
4161 void __kprobes add_preempt_count(int val)
4163 #ifdef CONFIG_DEBUG_PREEMPT
4167 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4170 preempt_count() += val;
4171 #ifdef CONFIG_DEBUG_PREEMPT
4173 * Spinlock count overflowing soon?
4175 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4178 if (preempt_count() == val)
4179 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4181 EXPORT_SYMBOL(add_preempt_count);
4183 void __kprobes sub_preempt_count(int val)
4185 #ifdef CONFIG_DEBUG_PREEMPT
4189 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4192 * Is the spinlock portion underflowing?
4194 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4195 !(preempt_count() & PREEMPT_MASK)))
4199 if (preempt_count() == val)
4200 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4201 preempt_count() -= val;
4203 EXPORT_SYMBOL(sub_preempt_count);
4208 * Print scheduling while atomic bug:
4210 static noinline void __schedule_bug(struct task_struct *prev)
4212 struct pt_regs *regs = get_irq_regs();
4214 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4215 prev->comm, prev->pid, preempt_count());
4217 debug_show_held_locks(prev);
4219 if (irqs_disabled())
4220 print_irqtrace_events(prev);
4229 * Various schedule()-time debugging checks and statistics:
4231 static inline void schedule_debug(struct task_struct *prev)
4234 * Test if we are atomic. Since do_exit() needs to call into
4235 * schedule() atomically, we ignore that path for now.
4236 * Otherwise, whine if we are scheduling when we should not be.
4238 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4239 __schedule_bug(prev);
4241 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4243 schedstat_inc(this_rq(), sched_count);
4246 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4248 if (prev->on_rq || rq->skip_clock_update < 0)
4249 update_rq_clock(rq);
4250 prev->sched_class->put_prev_task(rq, prev);
4254 * Pick up the highest-prio task:
4256 static inline struct task_struct *
4257 pick_next_task(struct rq *rq)
4259 const struct sched_class *class;
4260 struct task_struct *p;
4263 * Optimization: we know that if all tasks are in
4264 * the fair class we can call that function directly:
4266 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4267 p = fair_sched_class.pick_next_task(rq);
4272 for_each_class(class) {
4273 p = class->pick_next_task(rq);
4278 BUG(); /* the idle class will always have a runnable task */
4282 * __schedule() is the main scheduler function.
4284 static void __sched __schedule(void)
4286 struct task_struct *prev, *next;
4287 unsigned long *switch_count;
4293 cpu = smp_processor_id();
4295 rcu_note_context_switch(cpu);
4298 schedule_debug(prev);
4300 if (sched_feat(HRTICK))
4303 raw_spin_lock_irq(&rq->lock);
4305 switch_count = &prev->nivcsw;
4306 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4307 if (unlikely(signal_pending_state(prev->state, prev))) {
4308 prev->state = TASK_RUNNING;
4310 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4314 * If a worker went to sleep, notify and ask workqueue
4315 * whether it wants to wake up a task to maintain
4318 if (prev->flags & PF_WQ_WORKER) {
4319 struct task_struct *to_wakeup;
4321 to_wakeup = wq_worker_sleeping(prev, cpu);
4323 try_to_wake_up_local(to_wakeup);
4326 switch_count = &prev->nvcsw;
4329 pre_schedule(rq, prev);
4331 if (unlikely(!rq->nr_running))
4332 idle_balance(cpu, rq);
4334 put_prev_task(rq, prev);
4335 next = pick_next_task(rq);
4336 clear_tsk_need_resched(prev);
4337 rq->skip_clock_update = 0;
4339 if (likely(prev != next)) {
4344 context_switch(rq, prev, next); /* unlocks the rq */
4346 * The context switch have flipped the stack from under us
4347 * and restored the local variables which were saved when
4348 * this task called schedule() in the past. prev == current
4349 * is still correct, but it can be moved to another cpu/rq.
4351 cpu = smp_processor_id();
4354 raw_spin_unlock_irq(&rq->lock);
4358 preempt_enable_no_resched();
4363 static inline void sched_submit_work(struct task_struct *tsk)
4368 * If we are going to sleep and we have plugged IO queued,
4369 * make sure to submit it to avoid deadlocks.
4371 if (blk_needs_flush_plug(tsk))
4372 blk_schedule_flush_plug(tsk);
4375 asmlinkage void __sched schedule(void)
4377 struct task_struct *tsk = current;
4379 sched_submit_work(tsk);
4382 EXPORT_SYMBOL(schedule);
4384 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4386 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4388 if (lock->owner != owner)
4392 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4393 * lock->owner still matches owner, if that fails, owner might
4394 * point to free()d memory, if it still matches, the rcu_read_lock()
4395 * ensures the memory stays valid.
4399 return owner->on_cpu;
4403 * Look out! "owner" is an entirely speculative pointer
4404 * access and not reliable.
4406 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4408 if (!sched_feat(OWNER_SPIN))
4412 while (owner_running(lock, owner)) {
4416 arch_mutex_cpu_relax();
4421 * We break out the loop above on need_resched() and when the
4422 * owner changed, which is a sign for heavy contention. Return
4423 * success only when lock->owner is NULL.
4425 return lock->owner == NULL;
4429 #ifdef CONFIG_PREEMPT
4431 * this is the entry point to schedule() from in-kernel preemption
4432 * off of preempt_enable. Kernel preemptions off return from interrupt
4433 * occur there and call schedule directly.
4435 asmlinkage void __sched notrace preempt_schedule(void)
4437 struct thread_info *ti = current_thread_info();
4440 * If there is a non-zero preempt_count or interrupts are disabled,
4441 * we do not want to preempt the current task. Just return..
4443 if (likely(ti->preempt_count || irqs_disabled()))
4447 add_preempt_count_notrace(PREEMPT_ACTIVE);
4449 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4452 * Check again in case we missed a preemption opportunity
4453 * between schedule and now.
4456 } while (need_resched());
4458 EXPORT_SYMBOL(preempt_schedule);
4461 * this is the entry point to schedule() from kernel preemption
4462 * off of irq context.
4463 * Note, that this is called and return with irqs disabled. This will
4464 * protect us against recursive calling from irq.
4466 asmlinkage void __sched preempt_schedule_irq(void)
4468 struct thread_info *ti = current_thread_info();
4470 /* Catch callers which need to be fixed */
4471 BUG_ON(ti->preempt_count || !irqs_disabled());
4474 add_preempt_count(PREEMPT_ACTIVE);
4477 local_irq_disable();
4478 sub_preempt_count(PREEMPT_ACTIVE);
4481 * Check again in case we missed a preemption opportunity
4482 * between schedule and now.
4485 } while (need_resched());
4488 #endif /* CONFIG_PREEMPT */
4490 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4493 return try_to_wake_up(curr->private, mode, wake_flags);
4495 EXPORT_SYMBOL(default_wake_function);
4498 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4499 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4500 * number) then we wake all the non-exclusive tasks and one exclusive task.
4502 * There are circumstances in which we can try to wake a task which has already
4503 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4504 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4506 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4507 int nr_exclusive, int wake_flags, void *key)
4509 wait_queue_t *curr, *next;
4511 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4512 unsigned flags = curr->flags;
4514 if (curr->func(curr, mode, wake_flags, key) &&
4515 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4521 * __wake_up - wake up threads blocked on a waitqueue.
4523 * @mode: which threads
4524 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4525 * @key: is directly passed to the wakeup function
4527 * It may be assumed that this function implies a write memory barrier before
4528 * changing the task state if and only if any tasks are woken up.
4530 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4531 int nr_exclusive, void *key)
4533 unsigned long flags;
4535 spin_lock_irqsave(&q->lock, flags);
4536 __wake_up_common(q, mode, nr_exclusive, 0, key);
4537 spin_unlock_irqrestore(&q->lock, flags);
4539 EXPORT_SYMBOL(__wake_up);
4542 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4544 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4546 __wake_up_common(q, mode, 1, 0, NULL);
4548 EXPORT_SYMBOL_GPL(__wake_up_locked);
4550 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4552 __wake_up_common(q, mode, 1, 0, key);
4554 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4557 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4559 * @mode: which threads
4560 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4561 * @key: opaque value to be passed to wakeup targets
4563 * The sync wakeup differs that the waker knows that it will schedule
4564 * away soon, so while the target thread will be woken up, it will not
4565 * be migrated to another CPU - ie. the two threads are 'synchronized'
4566 * with each other. This can prevent needless bouncing between CPUs.
4568 * On UP it can prevent extra preemption.
4570 * It may be assumed that this function implies a write memory barrier before
4571 * changing the task state if and only if any tasks are woken up.
4573 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4574 int nr_exclusive, void *key)
4576 unsigned long flags;
4577 int wake_flags = WF_SYNC;
4582 if (unlikely(!nr_exclusive))
4585 spin_lock_irqsave(&q->lock, flags);
4586 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4587 spin_unlock_irqrestore(&q->lock, flags);
4589 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4592 * __wake_up_sync - see __wake_up_sync_key()
4594 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4596 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4598 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4601 * complete: - signals a single thread waiting on this completion
4602 * @x: holds the state of this particular completion
4604 * This will wake up a single thread waiting on this completion. Threads will be
4605 * awakened in the same order in which they were queued.
4607 * See also complete_all(), wait_for_completion() and related routines.
4609 * It may be assumed that this function implies a write memory barrier before
4610 * changing the task state if and only if any tasks are woken up.
4612 void complete(struct completion *x)
4614 unsigned long flags;
4616 spin_lock_irqsave(&x->wait.lock, flags);
4618 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4619 spin_unlock_irqrestore(&x->wait.lock, flags);
4621 EXPORT_SYMBOL(complete);
4624 * complete_all: - signals all threads waiting on this completion
4625 * @x: holds the state of this particular completion
4627 * This will wake up all threads waiting on this particular completion event.
4629 * It may be assumed that this function implies a write memory barrier before
4630 * changing the task state if and only if any tasks are woken up.
4632 void complete_all(struct completion *x)
4634 unsigned long flags;
4636 spin_lock_irqsave(&x->wait.lock, flags);
4637 x->done += UINT_MAX/2;
4638 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4639 spin_unlock_irqrestore(&x->wait.lock, flags);
4641 EXPORT_SYMBOL(complete_all);
4643 static inline long __sched
4644 do_wait_for_common(struct completion *x, long timeout, int state)
4647 DECLARE_WAITQUEUE(wait, current);
4649 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4651 if (signal_pending_state(state, current)) {
4652 timeout = -ERESTARTSYS;
4655 __set_current_state(state);
4656 spin_unlock_irq(&x->wait.lock);
4657 timeout = schedule_timeout(timeout);
4658 spin_lock_irq(&x->wait.lock);
4659 } while (!x->done && timeout);
4660 __remove_wait_queue(&x->wait, &wait);
4665 return timeout ?: 1;
4669 wait_for_common(struct completion *x, long timeout, int state)
4673 spin_lock_irq(&x->wait.lock);
4674 timeout = do_wait_for_common(x, timeout, state);
4675 spin_unlock_irq(&x->wait.lock);
4680 * wait_for_completion: - waits for completion of a task
4681 * @x: holds the state of this particular completion
4683 * This waits to be signaled for completion of a specific task. It is NOT
4684 * interruptible and there is no timeout.
4686 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4687 * and interrupt capability. Also see complete().
4689 void __sched wait_for_completion(struct completion *x)
4691 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4693 EXPORT_SYMBOL(wait_for_completion);
4696 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4697 * @x: holds the state of this particular completion
4698 * @timeout: timeout value in jiffies
4700 * This waits for either a completion of a specific task to be signaled or for a
4701 * specified timeout to expire. The timeout is in jiffies. It is not
4704 unsigned long __sched
4705 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4707 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4709 EXPORT_SYMBOL(wait_for_completion_timeout);
4712 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4713 * @x: holds the state of this particular completion
4715 * This waits for completion of a specific task to be signaled. It is
4718 int __sched wait_for_completion_interruptible(struct completion *x)
4720 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4721 if (t == -ERESTARTSYS)
4725 EXPORT_SYMBOL(wait_for_completion_interruptible);
4728 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4729 * @x: holds the state of this particular completion
4730 * @timeout: timeout value in jiffies
4732 * This waits for either a completion of a specific task to be signaled or for a
4733 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4736 wait_for_completion_interruptible_timeout(struct completion *x,
4737 unsigned long timeout)
4739 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4741 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4744 * wait_for_completion_killable: - waits for completion of a task (killable)
4745 * @x: holds the state of this particular completion
4747 * This waits to be signaled for completion of a specific task. It can be
4748 * interrupted by a kill signal.
4750 int __sched wait_for_completion_killable(struct completion *x)
4752 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4753 if (t == -ERESTARTSYS)
4757 EXPORT_SYMBOL(wait_for_completion_killable);
4760 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4761 * @x: holds the state of this particular completion
4762 * @timeout: timeout value in jiffies
4764 * This waits for either a completion of a specific task to be
4765 * signaled or for a specified timeout to expire. It can be
4766 * interrupted by a kill signal. The timeout is in jiffies.
4769 wait_for_completion_killable_timeout(struct completion *x,
4770 unsigned long timeout)
4772 return wait_for_common(x, timeout, TASK_KILLABLE);
4774 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4777 * try_wait_for_completion - try to decrement a completion without blocking
4778 * @x: completion structure
4780 * Returns: 0 if a decrement cannot be done without blocking
4781 * 1 if a decrement succeeded.
4783 * If a completion is being used as a counting completion,
4784 * attempt to decrement the counter without blocking. This
4785 * enables us to avoid waiting if the resource the completion
4786 * is protecting is not available.
4788 bool try_wait_for_completion(struct completion *x)
4790 unsigned long flags;
4793 spin_lock_irqsave(&x->wait.lock, flags);
4798 spin_unlock_irqrestore(&x->wait.lock, flags);
4801 EXPORT_SYMBOL(try_wait_for_completion);
4804 * completion_done - Test to see if a completion has any waiters
4805 * @x: completion structure
4807 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4808 * 1 if there are no waiters.
4811 bool completion_done(struct completion *x)
4813 unsigned long flags;
4816 spin_lock_irqsave(&x->wait.lock, flags);
4819 spin_unlock_irqrestore(&x->wait.lock, flags);
4822 EXPORT_SYMBOL(completion_done);
4825 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4827 unsigned long flags;
4830 init_waitqueue_entry(&wait, current);
4832 __set_current_state(state);
4834 spin_lock_irqsave(&q->lock, flags);
4835 __add_wait_queue(q, &wait);
4836 spin_unlock(&q->lock);
4837 timeout = schedule_timeout(timeout);
4838 spin_lock_irq(&q->lock);
4839 __remove_wait_queue(q, &wait);
4840 spin_unlock_irqrestore(&q->lock, flags);
4845 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4847 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4849 EXPORT_SYMBOL(interruptible_sleep_on);
4852 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4854 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4856 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4858 void __sched sleep_on(wait_queue_head_t *q)
4860 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4862 EXPORT_SYMBOL(sleep_on);
4864 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4866 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4868 EXPORT_SYMBOL(sleep_on_timeout);
4870 #ifdef CONFIG_RT_MUTEXES
4873 * rt_mutex_setprio - set the current priority of a task
4875 * @prio: prio value (kernel-internal form)
4877 * This function changes the 'effective' priority of a task. It does
4878 * not touch ->normal_prio like __setscheduler().
4880 * Used by the rt_mutex code to implement priority inheritance logic.
4882 void rt_mutex_setprio(struct task_struct *p, int prio)
4884 int oldprio, on_rq, running;
4886 const struct sched_class *prev_class;
4888 BUG_ON(prio < 0 || prio > MAX_PRIO);
4890 rq = __task_rq_lock(p);
4892 trace_sched_pi_setprio(p, prio);
4894 prev_class = p->sched_class;
4896 running = task_current(rq, p);
4898 dequeue_task(rq, p, 0);
4900 p->sched_class->put_prev_task(rq, p);
4903 p->sched_class = &rt_sched_class;
4905 p->sched_class = &fair_sched_class;
4910 p->sched_class->set_curr_task(rq);
4912 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4914 check_class_changed(rq, p, prev_class, oldprio);
4915 __task_rq_unlock(rq);
4920 void set_user_nice(struct task_struct *p, long nice)
4922 int old_prio, delta, on_rq;
4923 unsigned long flags;
4926 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4929 * We have to be careful, if called from sys_setpriority(),
4930 * the task might be in the middle of scheduling on another CPU.
4932 rq = task_rq_lock(p, &flags);
4934 * The RT priorities are set via sched_setscheduler(), but we still
4935 * allow the 'normal' nice value to be set - but as expected
4936 * it wont have any effect on scheduling until the task is
4937 * SCHED_FIFO/SCHED_RR:
4939 if (task_has_rt_policy(p)) {
4940 p->static_prio = NICE_TO_PRIO(nice);
4945 dequeue_task(rq, p, 0);
4947 p->static_prio = NICE_TO_PRIO(nice);
4950 p->prio = effective_prio(p);
4951 delta = p->prio - old_prio;
4954 enqueue_task(rq, p, 0);
4956 * If the task increased its priority or is running and
4957 * lowered its priority, then reschedule its CPU:
4959 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4960 resched_task(rq->curr);
4963 task_rq_unlock(rq, p, &flags);
4965 EXPORT_SYMBOL(set_user_nice);
4968 * can_nice - check if a task can reduce its nice value
4972 int can_nice(const struct task_struct *p, const int nice)
4974 /* convert nice value [19,-20] to rlimit style value [1,40] */
4975 int nice_rlim = 20 - nice;
4977 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4978 capable(CAP_SYS_NICE));
4981 #ifdef __ARCH_WANT_SYS_NICE
4984 * sys_nice - change the priority of the current process.
4985 * @increment: priority increment
4987 * sys_setpriority is a more generic, but much slower function that
4988 * does similar things.
4990 SYSCALL_DEFINE1(nice, int, increment)
4995 * Setpriority might change our priority at the same moment.
4996 * We don't have to worry. Conceptually one call occurs first
4997 * and we have a single winner.
4999 if (increment < -40)
5004 nice = TASK_NICE(current) + increment;
5010 if (increment < 0 && !can_nice(current, nice))
5013 retval = security_task_setnice(current, nice);
5017 set_user_nice(current, nice);
5024 * task_prio - return the priority value of a given task.
5025 * @p: the task in question.
5027 * This is the priority value as seen by users in /proc.
5028 * RT tasks are offset by -200. Normal tasks are centered
5029 * around 0, value goes from -16 to +15.
5031 int task_prio(const struct task_struct *p)
5033 return p->prio - MAX_RT_PRIO;
5037 * task_nice - return the nice value of a given task.
5038 * @p: the task in question.
5040 int task_nice(const struct task_struct *p)
5042 return TASK_NICE(p);
5044 EXPORT_SYMBOL(task_nice);
5047 * idle_cpu - is a given cpu idle currently?
5048 * @cpu: the processor in question.
5050 int idle_cpu(int cpu)
5052 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5056 * idle_task - return the idle task for a given cpu.
5057 * @cpu: the processor in question.
5059 struct task_struct *idle_task(int cpu)
5061 return cpu_rq(cpu)->idle;
5065 * find_process_by_pid - find a process with a matching PID value.
5066 * @pid: the pid in question.
5068 static struct task_struct *find_process_by_pid(pid_t pid)
5070 return pid ? find_task_by_vpid(pid) : current;
5073 /* Actually do priority change: must hold rq lock. */
5075 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5078 p->rt_priority = prio;
5079 p->normal_prio = normal_prio(p);
5080 /* we are holding p->pi_lock already */
5081 p->prio = rt_mutex_getprio(p);
5082 if (rt_prio(p->prio))
5083 p->sched_class = &rt_sched_class;
5085 p->sched_class = &fair_sched_class;
5090 * check the target process has a UID that matches the current process's
5092 static bool check_same_owner(struct task_struct *p)
5094 const struct cred *cred = current_cred(), *pcred;
5098 pcred = __task_cred(p);
5099 if (cred->user->user_ns == pcred->user->user_ns)
5100 match = (cred->euid == pcred->euid ||
5101 cred->euid == pcred->uid);
5108 static int __sched_setscheduler(struct task_struct *p, int policy,
5109 const struct sched_param *param, bool user)
5111 int retval, oldprio, oldpolicy = -1, on_rq, running;
5112 unsigned long flags;
5113 const struct sched_class *prev_class;
5117 /* may grab non-irq protected spin_locks */
5118 BUG_ON(in_interrupt());
5120 /* double check policy once rq lock held */
5122 reset_on_fork = p->sched_reset_on_fork;
5123 policy = oldpolicy = p->policy;
5125 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5126 policy &= ~SCHED_RESET_ON_FORK;
5128 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5129 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5130 policy != SCHED_IDLE)
5135 * Valid priorities for SCHED_FIFO and SCHED_RR are
5136 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5137 * SCHED_BATCH and SCHED_IDLE is 0.
5139 if (param->sched_priority < 0 ||
5140 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5141 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5143 if (rt_policy(policy) != (param->sched_priority != 0))
5147 * Allow unprivileged RT tasks to decrease priority:
5149 if (user && !capable(CAP_SYS_NICE)) {
5150 if (rt_policy(policy)) {
5151 unsigned long rlim_rtprio =
5152 task_rlimit(p, RLIMIT_RTPRIO);
5154 /* can't set/change the rt policy */
5155 if (policy != p->policy && !rlim_rtprio)
5158 /* can't increase priority */
5159 if (param->sched_priority > p->rt_priority &&
5160 param->sched_priority > rlim_rtprio)
5165 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5166 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5168 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5169 if (!can_nice(p, TASK_NICE(p)))
5173 /* can't change other user's priorities */
5174 if (!check_same_owner(p))
5177 /* Normal users shall not reset the sched_reset_on_fork flag */
5178 if (p->sched_reset_on_fork && !reset_on_fork)
5183 retval = security_task_setscheduler(p);
5189 * make sure no PI-waiters arrive (or leave) while we are
5190 * changing the priority of the task:
5192 * To be able to change p->policy safely, the appropriate
5193 * runqueue lock must be held.
5195 rq = task_rq_lock(p, &flags);
5198 * Changing the policy of the stop threads its a very bad idea
5200 if (p == rq->stop) {
5201 task_rq_unlock(rq, p, &flags);
5206 * If not changing anything there's no need to proceed further:
5208 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5209 param->sched_priority == p->rt_priority))) {
5211 __task_rq_unlock(rq);
5212 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5216 #ifdef CONFIG_RT_GROUP_SCHED
5219 * Do not allow realtime tasks into groups that have no runtime
5222 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5223 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5224 !task_group_is_autogroup(task_group(p))) {
5225 task_rq_unlock(rq, p, &flags);
5231 /* recheck policy now with rq lock held */
5232 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5233 policy = oldpolicy = -1;
5234 task_rq_unlock(rq, p, &flags);
5238 running = task_current(rq, p);
5240 deactivate_task(rq, p, 0);
5242 p->sched_class->put_prev_task(rq, p);
5244 p->sched_reset_on_fork = reset_on_fork;
5247 prev_class = p->sched_class;
5248 __setscheduler(rq, p, policy, param->sched_priority);
5251 p->sched_class->set_curr_task(rq);
5253 activate_task(rq, p, 0);
5255 check_class_changed(rq, p, prev_class, oldprio);
5256 task_rq_unlock(rq, p, &flags);
5258 rt_mutex_adjust_pi(p);
5264 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5265 * @p: the task in question.
5266 * @policy: new policy.
5267 * @param: structure containing the new RT priority.
5269 * NOTE that the task may be already dead.
5271 int sched_setscheduler(struct task_struct *p, int policy,
5272 const struct sched_param *param)
5274 return __sched_setscheduler(p, policy, param, true);
5276 EXPORT_SYMBOL_GPL(sched_setscheduler);
5279 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5280 * @p: the task in question.
5281 * @policy: new policy.
5282 * @param: structure containing the new RT priority.
5284 * Just like sched_setscheduler, only don't bother checking if the
5285 * current context has permission. For example, this is needed in
5286 * stop_machine(): we create temporary high priority worker threads,
5287 * but our caller might not have that capability.
5289 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5290 const struct sched_param *param)
5292 return __sched_setscheduler(p, policy, param, false);
5296 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5298 struct sched_param lparam;
5299 struct task_struct *p;
5302 if (!param || pid < 0)
5304 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5309 p = find_process_by_pid(pid);
5311 retval = sched_setscheduler(p, policy, &lparam);
5318 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5319 * @pid: the pid in question.
5320 * @policy: new policy.
5321 * @param: structure containing the new RT priority.
5323 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5324 struct sched_param __user *, param)
5326 /* negative values for policy are not valid */
5330 return do_sched_setscheduler(pid, policy, param);
5334 * sys_sched_setparam - set/change the RT priority of a thread
5335 * @pid: the pid in question.
5336 * @param: structure containing the new RT priority.
5338 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5340 return do_sched_setscheduler(pid, -1, param);
5344 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5345 * @pid: the pid in question.
5347 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5349 struct task_struct *p;
5357 p = find_process_by_pid(pid);
5359 retval = security_task_getscheduler(p);
5362 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5369 * sys_sched_getparam - get the RT priority of a thread
5370 * @pid: the pid in question.
5371 * @param: structure containing the RT priority.
5373 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5375 struct sched_param lp;
5376 struct task_struct *p;
5379 if (!param || pid < 0)
5383 p = find_process_by_pid(pid);
5388 retval = security_task_getscheduler(p);
5392 lp.sched_priority = p->rt_priority;
5396 * This one might sleep, we cannot do it with a spinlock held ...
5398 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5407 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5409 cpumask_var_t cpus_allowed, new_mask;
5410 struct task_struct *p;
5416 p = find_process_by_pid(pid);
5423 /* Prevent p going away */
5427 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5431 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5433 goto out_free_cpus_allowed;
5436 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5439 retval = security_task_setscheduler(p);
5443 cpuset_cpus_allowed(p, cpus_allowed);
5444 cpumask_and(new_mask, in_mask, cpus_allowed);
5446 retval = set_cpus_allowed_ptr(p, new_mask);
5449 cpuset_cpus_allowed(p, cpus_allowed);
5450 if (!cpumask_subset(new_mask, cpus_allowed)) {
5452 * We must have raced with a concurrent cpuset
5453 * update. Just reset the cpus_allowed to the
5454 * cpuset's cpus_allowed
5456 cpumask_copy(new_mask, cpus_allowed);
5461 free_cpumask_var(new_mask);
5462 out_free_cpus_allowed:
5463 free_cpumask_var(cpus_allowed);
5470 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5471 struct cpumask *new_mask)
5473 if (len < cpumask_size())
5474 cpumask_clear(new_mask);
5475 else if (len > cpumask_size())
5476 len = cpumask_size();
5478 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5482 * sys_sched_setaffinity - set the cpu affinity of a process
5483 * @pid: pid of the process
5484 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5485 * @user_mask_ptr: user-space pointer to the new cpu mask
5487 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5488 unsigned long __user *, user_mask_ptr)
5490 cpumask_var_t new_mask;
5493 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5496 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5498 retval = sched_setaffinity(pid, new_mask);
5499 free_cpumask_var(new_mask);
5503 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5505 struct task_struct *p;
5506 unsigned long flags;
5513 p = find_process_by_pid(pid);
5517 retval = security_task_getscheduler(p);
5521 raw_spin_lock_irqsave(&p->pi_lock, flags);
5522 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5523 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5533 * sys_sched_getaffinity - get the cpu affinity of a process
5534 * @pid: pid of the process
5535 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5536 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5538 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5539 unsigned long __user *, user_mask_ptr)
5544 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5546 if (len & (sizeof(unsigned long)-1))
5549 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5552 ret = sched_getaffinity(pid, mask);
5554 size_t retlen = min_t(size_t, len, cpumask_size());
5556 if (copy_to_user(user_mask_ptr, mask, retlen))
5561 free_cpumask_var(mask);
5567 * sys_sched_yield - yield the current processor to other threads.
5569 * This function yields the current CPU to other tasks. If there are no
5570 * other threads running on this CPU then this function will return.
5572 SYSCALL_DEFINE0(sched_yield)
5574 struct rq *rq = this_rq_lock();
5576 schedstat_inc(rq, yld_count);
5577 current->sched_class->yield_task(rq);
5580 * Since we are going to call schedule() anyway, there's
5581 * no need to preempt or enable interrupts:
5583 __release(rq->lock);
5584 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5585 do_raw_spin_unlock(&rq->lock);
5586 preempt_enable_no_resched();
5593 static inline int should_resched(void)
5595 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5598 static void __cond_resched(void)
5600 add_preempt_count(PREEMPT_ACTIVE);
5602 sub_preempt_count(PREEMPT_ACTIVE);
5605 int __sched _cond_resched(void)
5607 if (should_resched()) {
5613 EXPORT_SYMBOL(_cond_resched);
5616 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5617 * call schedule, and on return reacquire the lock.
5619 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5620 * operations here to prevent schedule() from being called twice (once via
5621 * spin_unlock(), once by hand).
5623 int __cond_resched_lock(spinlock_t *lock)
5625 int resched = should_resched();
5628 lockdep_assert_held(lock);
5630 if (spin_needbreak(lock) || resched) {
5641 EXPORT_SYMBOL(__cond_resched_lock);
5643 int __sched __cond_resched_softirq(void)
5645 BUG_ON(!in_softirq());
5647 if (should_resched()) {
5655 EXPORT_SYMBOL(__cond_resched_softirq);
5658 * yield - yield the current processor to other threads.
5660 * This is a shortcut for kernel-space yielding - it marks the
5661 * thread runnable and calls sys_sched_yield().
5663 void __sched yield(void)
5665 set_current_state(TASK_RUNNING);
5668 EXPORT_SYMBOL(yield);
5671 * yield_to - yield the current processor to another thread in
5672 * your thread group, or accelerate that thread toward the
5673 * processor it's on.
5675 * @preempt: whether task preemption is allowed or not
5677 * It's the caller's job to ensure that the target task struct
5678 * can't go away on us before we can do any checks.
5680 * Returns true if we indeed boosted the target task.
5682 bool __sched yield_to(struct task_struct *p, bool preempt)
5684 struct task_struct *curr = current;
5685 struct rq *rq, *p_rq;
5686 unsigned long flags;
5689 local_irq_save(flags);
5694 double_rq_lock(rq, p_rq);
5695 while (task_rq(p) != p_rq) {
5696 double_rq_unlock(rq, p_rq);
5700 if (!curr->sched_class->yield_to_task)
5703 if (curr->sched_class != p->sched_class)
5706 if (task_running(p_rq, p) || p->state)
5709 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5711 schedstat_inc(rq, yld_count);
5713 * Make p's CPU reschedule; pick_next_entity takes care of
5716 if (preempt && rq != p_rq)
5717 resched_task(p_rq->curr);
5721 double_rq_unlock(rq, p_rq);
5722 local_irq_restore(flags);
5729 EXPORT_SYMBOL_GPL(yield_to);
5732 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5733 * that process accounting knows that this is a task in IO wait state.
5735 void __sched io_schedule(void)
5737 struct rq *rq = raw_rq();
5739 delayacct_blkio_start();
5740 atomic_inc(&rq->nr_iowait);
5741 blk_flush_plug(current);
5742 current->in_iowait = 1;
5744 current->in_iowait = 0;
5745 atomic_dec(&rq->nr_iowait);
5746 delayacct_blkio_end();
5748 EXPORT_SYMBOL(io_schedule);
5750 long __sched io_schedule_timeout(long timeout)
5752 struct rq *rq = raw_rq();
5755 delayacct_blkio_start();
5756 atomic_inc(&rq->nr_iowait);
5757 blk_flush_plug(current);
5758 current->in_iowait = 1;
5759 ret = schedule_timeout(timeout);
5760 current->in_iowait = 0;
5761 atomic_dec(&rq->nr_iowait);
5762 delayacct_blkio_end();
5767 * sys_sched_get_priority_max - return maximum RT priority.
5768 * @policy: scheduling class.
5770 * this syscall returns the maximum rt_priority that can be used
5771 * by a given scheduling class.
5773 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5780 ret = MAX_USER_RT_PRIO-1;
5792 * sys_sched_get_priority_min - return minimum RT priority.
5793 * @policy: scheduling class.
5795 * this syscall returns the minimum rt_priority that can be used
5796 * by a given scheduling class.
5798 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5816 * sys_sched_rr_get_interval - return the default timeslice of a process.
5817 * @pid: pid of the process.
5818 * @interval: userspace pointer to the timeslice value.
5820 * this syscall writes the default timeslice value of a given process
5821 * into the user-space timespec buffer. A value of '0' means infinity.
5823 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5824 struct timespec __user *, interval)
5826 struct task_struct *p;
5827 unsigned int time_slice;
5828 unsigned long flags;
5838 p = find_process_by_pid(pid);
5842 retval = security_task_getscheduler(p);
5846 rq = task_rq_lock(p, &flags);
5847 time_slice = p->sched_class->get_rr_interval(rq, p);
5848 task_rq_unlock(rq, p, &flags);
5851 jiffies_to_timespec(time_slice, &t);
5852 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5860 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5862 void sched_show_task(struct task_struct *p)
5864 unsigned long free = 0;
5867 state = p->state ? __ffs(p->state) + 1 : 0;
5868 printk(KERN_INFO "%-15.15s %c", p->comm,
5869 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5870 #if BITS_PER_LONG == 32
5871 if (state == TASK_RUNNING)
5872 printk(KERN_CONT " running ");
5874 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5876 if (state == TASK_RUNNING)
5877 printk(KERN_CONT " running task ");
5879 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5881 #ifdef CONFIG_DEBUG_STACK_USAGE
5882 free = stack_not_used(p);
5884 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5885 task_pid_nr(p), task_pid_nr(p->real_parent),
5886 (unsigned long)task_thread_info(p)->flags);
5888 show_stack(p, NULL);
5891 void show_state_filter(unsigned long state_filter)
5893 struct task_struct *g, *p;
5895 #if BITS_PER_LONG == 32
5897 " task PC stack pid father\n");
5900 " task PC stack pid father\n");
5902 read_lock(&tasklist_lock);
5903 do_each_thread(g, p) {
5905 * reset the NMI-timeout, listing all files on a slow
5906 * console might take a lot of time:
5908 touch_nmi_watchdog();
5909 if (!state_filter || (p->state & state_filter))
5911 } while_each_thread(g, p);
5913 touch_all_softlockup_watchdogs();
5915 #ifdef CONFIG_SCHED_DEBUG
5916 sysrq_sched_debug_show();
5918 read_unlock(&tasklist_lock);
5920 * Only show locks if all tasks are dumped:
5923 debug_show_all_locks();
5926 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5928 idle->sched_class = &idle_sched_class;
5932 * init_idle - set up an idle thread for a given CPU
5933 * @idle: task in question
5934 * @cpu: cpu the idle task belongs to
5936 * NOTE: this function does not set the idle thread's NEED_RESCHED
5937 * flag, to make booting more robust.
5939 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5941 struct rq *rq = cpu_rq(cpu);
5942 unsigned long flags;
5944 raw_spin_lock_irqsave(&rq->lock, flags);
5947 idle->state = TASK_RUNNING;
5948 idle->se.exec_start = sched_clock();
5950 do_set_cpus_allowed(idle, cpumask_of(cpu));
5952 * We're having a chicken and egg problem, even though we are
5953 * holding rq->lock, the cpu isn't yet set to this cpu so the
5954 * lockdep check in task_group() will fail.
5956 * Similar case to sched_fork(). / Alternatively we could
5957 * use task_rq_lock() here and obtain the other rq->lock.
5962 __set_task_cpu(idle, cpu);
5965 rq->curr = rq->idle = idle;
5966 #if defined(CONFIG_SMP)
5969 raw_spin_unlock_irqrestore(&rq->lock, flags);
5971 /* Set the preempt count _outside_ the spinlocks! */
5972 task_thread_info(idle)->preempt_count = 0;
5975 * The idle tasks have their own, simple scheduling class:
5977 idle->sched_class = &idle_sched_class;
5978 ftrace_graph_init_idle_task(idle, cpu);
5982 * In a system that switches off the HZ timer nohz_cpu_mask
5983 * indicates which cpus entered this state. This is used
5984 * in the rcu update to wait only for active cpus. For system
5985 * which do not switch off the HZ timer nohz_cpu_mask should
5986 * always be CPU_BITS_NONE.
5988 cpumask_var_t nohz_cpu_mask;
5991 * Increase the granularity value when there are more CPUs,
5992 * because with more CPUs the 'effective latency' as visible
5993 * to users decreases. But the relationship is not linear,
5994 * so pick a second-best guess by going with the log2 of the
5997 * This idea comes from the SD scheduler of Con Kolivas:
5999 static int get_update_sysctl_factor(void)
6001 unsigned int cpus = min_t(int, num_online_cpus(), 8);
6002 unsigned int factor;
6004 switch (sysctl_sched_tunable_scaling) {
6005 case SCHED_TUNABLESCALING_NONE:
6008 case SCHED_TUNABLESCALING_LINEAR:
6011 case SCHED_TUNABLESCALING_LOG:
6013 factor = 1 + ilog2(cpus);
6020 static void update_sysctl(void)
6022 unsigned int factor = get_update_sysctl_factor();
6024 #define SET_SYSCTL(name) \
6025 (sysctl_##name = (factor) * normalized_sysctl_##name)
6026 SET_SYSCTL(sched_min_granularity);
6027 SET_SYSCTL(sched_latency);
6028 SET_SYSCTL(sched_wakeup_granularity);
6032 static inline void sched_init_granularity(void)
6038 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
6040 if (p->sched_class && p->sched_class->set_cpus_allowed)
6041 p->sched_class->set_cpus_allowed(p, new_mask);
6043 cpumask_copy(&p->cpus_allowed, new_mask);
6044 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6049 * This is how migration works:
6051 * 1) we invoke migration_cpu_stop() on the target CPU using
6053 * 2) stopper starts to run (implicitly forcing the migrated thread
6055 * 3) it checks whether the migrated task is still in the wrong runqueue.
6056 * 4) if it's in the wrong runqueue then the migration thread removes
6057 * it and puts it into the right queue.
6058 * 5) stopper completes and stop_one_cpu() returns and the migration
6063 * Change a given task's CPU affinity. Migrate the thread to a
6064 * proper CPU and schedule it away if the CPU it's executing on
6065 * is removed from the allowed bitmask.
6067 * NOTE: the caller must have a valid reference to the task, the
6068 * task must not exit() & deallocate itself prematurely. The
6069 * call is not atomic; no spinlocks may be held.
6071 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6073 unsigned long flags;
6075 unsigned int dest_cpu;
6078 rq = task_rq_lock(p, &flags);
6080 if (cpumask_equal(&p->cpus_allowed, new_mask))
6083 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6088 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6093 do_set_cpus_allowed(p, new_mask);
6095 /* Can the task run on the task's current CPU? If so, we're done */
6096 if (cpumask_test_cpu(task_cpu(p), new_mask))
6099 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6101 struct migration_arg arg = { p, dest_cpu };
6102 /* Need help from migration thread: drop lock and wait. */
6103 task_rq_unlock(rq, p, &flags);
6104 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6105 tlb_migrate_finish(p->mm);
6109 task_rq_unlock(rq, p, &flags);
6113 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6116 * Move (not current) task off this cpu, onto dest cpu. We're doing
6117 * this because either it can't run here any more (set_cpus_allowed()
6118 * away from this CPU, or CPU going down), or because we're
6119 * attempting to rebalance this task on exec (sched_exec).
6121 * So we race with normal scheduler movements, but that's OK, as long
6122 * as the task is no longer on this CPU.
6124 * Returns non-zero if task was successfully migrated.
6126 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6128 struct rq *rq_dest, *rq_src;
6131 if (unlikely(!cpu_active(dest_cpu)))
6134 rq_src = cpu_rq(src_cpu);
6135 rq_dest = cpu_rq(dest_cpu);
6137 raw_spin_lock(&p->pi_lock);
6138 double_rq_lock(rq_src, rq_dest);
6139 /* Already moved. */
6140 if (task_cpu(p) != src_cpu)
6142 /* Affinity changed (again). */
6143 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6147 * If we're not on a rq, the next wake-up will ensure we're
6151 deactivate_task(rq_src, p, 0);
6152 set_task_cpu(p, dest_cpu);
6153 activate_task(rq_dest, p, 0);
6154 check_preempt_curr(rq_dest, p, 0);
6159 double_rq_unlock(rq_src, rq_dest);
6160 raw_spin_unlock(&p->pi_lock);
6165 * migration_cpu_stop - this will be executed by a highprio stopper thread
6166 * and performs thread migration by bumping thread off CPU then
6167 * 'pushing' onto another runqueue.
6169 static int migration_cpu_stop(void *data)
6171 struct migration_arg *arg = data;
6174 * The original target cpu might have gone down and we might
6175 * be on another cpu but it doesn't matter.
6177 local_irq_disable();
6178 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6183 #ifdef CONFIG_HOTPLUG_CPU
6186 * Ensures that the idle task is using init_mm right before its cpu goes
6189 void idle_task_exit(void)
6191 struct mm_struct *mm = current->active_mm;
6193 BUG_ON(cpu_online(smp_processor_id()));
6196 switch_mm(mm, &init_mm, current);
6201 * While a dead CPU has no uninterruptible tasks queued at this point,
6202 * it might still have a nonzero ->nr_uninterruptible counter, because
6203 * for performance reasons the counter is not stricly tracking tasks to
6204 * their home CPUs. So we just add the counter to another CPU's counter,
6205 * to keep the global sum constant after CPU-down:
6207 static void migrate_nr_uninterruptible(struct rq *rq_src)
6209 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6211 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6212 rq_src->nr_uninterruptible = 0;
6216 * remove the tasks which were accounted by rq from calc_load_tasks.
6218 static void calc_global_load_remove(struct rq *rq)
6220 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6221 rq->calc_load_active = 0;
6225 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6226 * try_to_wake_up()->select_task_rq().
6228 * Called with rq->lock held even though we'er in stop_machine() and
6229 * there's no concurrency possible, we hold the required locks anyway
6230 * because of lock validation efforts.
6232 static void migrate_tasks(unsigned int dead_cpu)
6234 struct rq *rq = cpu_rq(dead_cpu);
6235 struct task_struct *next, *stop = rq->stop;
6239 * Fudge the rq selection such that the below task selection loop
6240 * doesn't get stuck on the currently eligible stop task.
6242 * We're currently inside stop_machine() and the rq is either stuck
6243 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6244 * either way we should never end up calling schedule() until we're
6251 * There's this thread running, bail when that's the only
6254 if (rq->nr_running == 1)
6257 next = pick_next_task(rq);
6259 next->sched_class->put_prev_task(rq, next);
6261 /* Find suitable destination for @next, with force if needed. */
6262 dest_cpu = select_fallback_rq(dead_cpu, next);
6263 raw_spin_unlock(&rq->lock);
6265 __migrate_task(next, dead_cpu, dest_cpu);
6267 raw_spin_lock(&rq->lock);
6273 #endif /* CONFIG_HOTPLUG_CPU */
6275 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6277 static struct ctl_table sd_ctl_dir[] = {
6279 .procname = "sched_domain",
6285 static struct ctl_table sd_ctl_root[] = {
6287 .procname = "kernel",
6289 .child = sd_ctl_dir,
6294 static struct ctl_table *sd_alloc_ctl_entry(int n)
6296 struct ctl_table *entry =
6297 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6302 static void sd_free_ctl_entry(struct ctl_table **tablep)
6304 struct ctl_table *entry;
6307 * In the intermediate directories, both the child directory and
6308 * procname are dynamically allocated and could fail but the mode
6309 * will always be set. In the lowest directory the names are
6310 * static strings and all have proc handlers.
6312 for (entry = *tablep; entry->mode; entry++) {
6314 sd_free_ctl_entry(&entry->child);
6315 if (entry->proc_handler == NULL)
6316 kfree(entry->procname);
6324 set_table_entry(struct ctl_table *entry,
6325 const char *procname, void *data, int maxlen,
6326 mode_t mode, proc_handler *proc_handler)
6328 entry->procname = procname;
6330 entry->maxlen = maxlen;
6332 entry->proc_handler = proc_handler;
6335 static struct ctl_table *
6336 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6338 struct ctl_table *table = sd_alloc_ctl_entry(13);
6343 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6344 sizeof(long), 0644, proc_doulongvec_minmax);
6345 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6346 sizeof(long), 0644, proc_doulongvec_minmax);
6347 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6348 sizeof(int), 0644, proc_dointvec_minmax);
6349 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6350 sizeof(int), 0644, proc_dointvec_minmax);
6351 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6352 sizeof(int), 0644, proc_dointvec_minmax);
6353 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6354 sizeof(int), 0644, proc_dointvec_minmax);
6355 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6356 sizeof(int), 0644, proc_dointvec_minmax);
6357 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6358 sizeof(int), 0644, proc_dointvec_minmax);
6359 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6360 sizeof(int), 0644, proc_dointvec_minmax);
6361 set_table_entry(&table[9], "cache_nice_tries",
6362 &sd->cache_nice_tries,
6363 sizeof(int), 0644, proc_dointvec_minmax);
6364 set_table_entry(&table[10], "flags", &sd->flags,
6365 sizeof(int), 0644, proc_dointvec_minmax);
6366 set_table_entry(&table[11], "name", sd->name,
6367 CORENAME_MAX_SIZE, 0444, proc_dostring);
6368 /* &table[12] is terminator */
6373 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6375 struct ctl_table *entry, *table;
6376 struct sched_domain *sd;
6377 int domain_num = 0, i;
6380 for_each_domain(cpu, sd)
6382 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6387 for_each_domain(cpu, sd) {
6388 snprintf(buf, 32, "domain%d", i);
6389 entry->procname = kstrdup(buf, GFP_KERNEL);
6391 entry->child = sd_alloc_ctl_domain_table(sd);
6398 static struct ctl_table_header *sd_sysctl_header;
6399 static void register_sched_domain_sysctl(void)
6401 int i, cpu_num = num_possible_cpus();
6402 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6405 WARN_ON(sd_ctl_dir[0].child);
6406 sd_ctl_dir[0].child = entry;
6411 for_each_possible_cpu(i) {
6412 snprintf(buf, 32, "cpu%d", i);
6413 entry->procname = kstrdup(buf, GFP_KERNEL);
6415 entry->child = sd_alloc_ctl_cpu_table(i);
6419 WARN_ON(sd_sysctl_header);
6420 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6423 /* may be called multiple times per register */
6424 static void unregister_sched_domain_sysctl(void)
6426 if (sd_sysctl_header)
6427 unregister_sysctl_table(sd_sysctl_header);
6428 sd_sysctl_header = NULL;
6429 if (sd_ctl_dir[0].child)
6430 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6433 static void register_sched_domain_sysctl(void)
6436 static void unregister_sched_domain_sysctl(void)
6441 static void set_rq_online(struct rq *rq)
6444 const struct sched_class *class;
6446 cpumask_set_cpu(rq->cpu, rq->rd->online);
6449 for_each_class(class) {
6450 if (class->rq_online)
6451 class->rq_online(rq);
6456 static void set_rq_offline(struct rq *rq)
6459 const struct sched_class *class;
6461 for_each_class(class) {
6462 if (class->rq_offline)
6463 class->rq_offline(rq);
6466 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6472 * migration_call - callback that gets triggered when a CPU is added.
6473 * Here we can start up the necessary migration thread for the new CPU.
6475 static int __cpuinit
6476 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6478 int cpu = (long)hcpu;
6479 unsigned long flags;
6480 struct rq *rq = cpu_rq(cpu);
6482 switch (action & ~CPU_TASKS_FROZEN) {
6484 case CPU_UP_PREPARE:
6485 rq->calc_load_update = calc_load_update;
6489 /* Update our root-domain */
6490 raw_spin_lock_irqsave(&rq->lock, flags);
6492 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6496 raw_spin_unlock_irqrestore(&rq->lock, flags);
6499 #ifdef CONFIG_HOTPLUG_CPU
6501 sched_ttwu_pending();
6502 /* Update our root-domain */
6503 raw_spin_lock_irqsave(&rq->lock, flags);
6505 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6509 BUG_ON(rq->nr_running != 1); /* the migration thread */
6510 raw_spin_unlock_irqrestore(&rq->lock, flags);
6512 migrate_nr_uninterruptible(rq);
6513 calc_global_load_remove(rq);
6518 update_max_interval();
6524 * Register at high priority so that task migration (migrate_all_tasks)
6525 * happens before everything else. This has to be lower priority than
6526 * the notifier in the perf_event subsystem, though.
6528 static struct notifier_block __cpuinitdata migration_notifier = {
6529 .notifier_call = migration_call,
6530 .priority = CPU_PRI_MIGRATION,
6533 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6534 unsigned long action, void *hcpu)
6536 switch (action & ~CPU_TASKS_FROZEN) {
6538 case CPU_DOWN_FAILED:
6539 set_cpu_active((long)hcpu, true);
6546 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6547 unsigned long action, void *hcpu)
6549 switch (action & ~CPU_TASKS_FROZEN) {
6550 case CPU_DOWN_PREPARE:
6551 set_cpu_active((long)hcpu, false);
6558 static int __init migration_init(void)
6560 void *cpu = (void *)(long)smp_processor_id();
6563 /* Initialize migration for the boot CPU */
6564 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6565 BUG_ON(err == NOTIFY_BAD);
6566 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6567 register_cpu_notifier(&migration_notifier);
6569 /* Register cpu active notifiers */
6570 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6571 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6575 early_initcall(migration_init);
6580 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6582 #ifdef CONFIG_SCHED_DEBUG
6584 static __read_mostly int sched_domain_debug_enabled;
6586 static int __init sched_domain_debug_setup(char *str)
6588 sched_domain_debug_enabled = 1;
6592 early_param("sched_debug", sched_domain_debug_setup);
6594 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6595 struct cpumask *groupmask)
6597 struct sched_group *group = sd->groups;
6600 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6601 cpumask_clear(groupmask);
6603 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6605 if (!(sd->flags & SD_LOAD_BALANCE)) {
6606 printk("does not load-balance\n");
6608 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6613 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6615 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6616 printk(KERN_ERR "ERROR: domain->span does not contain "
6619 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6620 printk(KERN_ERR "ERROR: domain->groups does not contain"
6624 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6628 printk(KERN_ERR "ERROR: group is NULL\n");
6632 if (!group->sgp->power) {
6633 printk(KERN_CONT "\n");
6634 printk(KERN_ERR "ERROR: domain->cpu_power not "
6639 if (!cpumask_weight(sched_group_cpus(group))) {
6640 printk(KERN_CONT "\n");
6641 printk(KERN_ERR "ERROR: empty group\n");
6645 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6646 printk(KERN_CONT "\n");
6647 printk(KERN_ERR "ERROR: repeated CPUs\n");
6651 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6653 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6655 printk(KERN_CONT " %s", str);
6656 if (group->sgp->power != SCHED_POWER_SCALE) {
6657 printk(KERN_CONT " (cpu_power = %d)",
6661 group = group->next;
6662 } while (group != sd->groups);
6663 printk(KERN_CONT "\n");
6665 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6666 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6669 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6670 printk(KERN_ERR "ERROR: parent span is not a superset "
6671 "of domain->span\n");
6675 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6679 if (!sched_domain_debug_enabled)
6683 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6687 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6690 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6698 #else /* !CONFIG_SCHED_DEBUG */
6699 # define sched_domain_debug(sd, cpu) do { } while (0)
6700 #endif /* CONFIG_SCHED_DEBUG */
6702 static int sd_degenerate(struct sched_domain *sd)
6704 if (cpumask_weight(sched_domain_span(sd)) == 1)
6707 /* Following flags need at least 2 groups */
6708 if (sd->flags & (SD_LOAD_BALANCE |
6709 SD_BALANCE_NEWIDLE |
6713 SD_SHARE_PKG_RESOURCES)) {
6714 if (sd->groups != sd->groups->next)
6718 /* Following flags don't use groups */
6719 if (sd->flags & (SD_WAKE_AFFINE))
6726 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6728 unsigned long cflags = sd->flags, pflags = parent->flags;
6730 if (sd_degenerate(parent))
6733 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6736 /* Flags needing groups don't count if only 1 group in parent */
6737 if (parent->groups == parent->groups->next) {
6738 pflags &= ~(SD_LOAD_BALANCE |
6739 SD_BALANCE_NEWIDLE |
6743 SD_SHARE_PKG_RESOURCES);
6744 if (nr_node_ids == 1)
6745 pflags &= ~SD_SERIALIZE;
6747 if (~cflags & pflags)
6753 static void free_rootdomain(struct rcu_head *rcu)
6755 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6757 cpupri_cleanup(&rd->cpupri);
6758 free_cpumask_var(rd->rto_mask);
6759 free_cpumask_var(rd->online);
6760 free_cpumask_var(rd->span);
6764 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6766 struct root_domain *old_rd = NULL;
6767 unsigned long flags;
6769 raw_spin_lock_irqsave(&rq->lock, flags);
6774 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6777 cpumask_clear_cpu(rq->cpu, old_rd->span);
6780 * If we dont want to free the old_rt yet then
6781 * set old_rd to NULL to skip the freeing later
6784 if (!atomic_dec_and_test(&old_rd->refcount))
6788 atomic_inc(&rd->refcount);
6791 cpumask_set_cpu(rq->cpu, rd->span);
6792 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6795 raw_spin_unlock_irqrestore(&rq->lock, flags);
6798 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6801 static int init_rootdomain(struct root_domain *rd)
6803 memset(rd, 0, sizeof(*rd));
6805 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6807 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6809 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6812 if (cpupri_init(&rd->cpupri) != 0)
6817 free_cpumask_var(rd->rto_mask);
6819 free_cpumask_var(rd->online);
6821 free_cpumask_var(rd->span);
6826 static void init_defrootdomain(void)
6828 init_rootdomain(&def_root_domain);
6830 atomic_set(&def_root_domain.refcount, 1);
6833 static struct root_domain *alloc_rootdomain(void)
6835 struct root_domain *rd;
6837 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6841 if (init_rootdomain(rd) != 0) {
6849 static void free_sched_groups(struct sched_group *sg, int free_sgp)
6851 struct sched_group *tmp, *first;
6860 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
6865 } while (sg != first);
6868 static void free_sched_domain(struct rcu_head *rcu)
6870 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6873 * If its an overlapping domain it has private groups, iterate and
6876 if (sd->flags & SD_OVERLAP) {
6877 free_sched_groups(sd->groups, 1);
6878 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6879 kfree(sd->groups->sgp);
6885 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6887 call_rcu(&sd->rcu, free_sched_domain);
6890 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6892 for (; sd; sd = sd->parent)
6893 destroy_sched_domain(sd, cpu);
6897 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6898 * hold the hotplug lock.
6901 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6903 struct rq *rq = cpu_rq(cpu);
6904 struct sched_domain *tmp;
6906 /* Remove the sched domains which do not contribute to scheduling. */
6907 for (tmp = sd; tmp; ) {
6908 struct sched_domain *parent = tmp->parent;
6912 if (sd_parent_degenerate(tmp, parent)) {
6913 tmp->parent = parent->parent;
6915 parent->parent->child = tmp;
6916 destroy_sched_domain(parent, cpu);
6921 if (sd && sd_degenerate(sd)) {
6924 destroy_sched_domain(tmp, cpu);
6929 sched_domain_debug(sd, cpu);
6931 rq_attach_root(rq, rd);
6933 rcu_assign_pointer(rq->sd, sd);
6934 destroy_sched_domains(tmp, cpu);
6937 /* cpus with isolated domains */
6938 static cpumask_var_t cpu_isolated_map;
6940 /* Setup the mask of cpus configured for isolated domains */
6941 static int __init isolated_cpu_setup(char *str)
6943 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6944 cpulist_parse(str, cpu_isolated_map);
6948 __setup("isolcpus=", isolated_cpu_setup);
6950 #define SD_NODES_PER_DOMAIN 16
6955 * find_next_best_node - find the next node to include in a sched_domain
6956 * @node: node whose sched_domain we're building
6957 * @used_nodes: nodes already in the sched_domain
6959 * Find the next node to include in a given scheduling domain. Simply
6960 * finds the closest node not already in the @used_nodes map.
6962 * Should use nodemask_t.
6964 static int find_next_best_node(int node, nodemask_t *used_nodes)
6966 int i, n, val, min_val, best_node = -1;
6970 for (i = 0; i < nr_node_ids; i++) {
6971 /* Start at @node */
6972 n = (node + i) % nr_node_ids;
6974 if (!nr_cpus_node(n))
6977 /* Skip already used nodes */
6978 if (node_isset(n, *used_nodes))
6981 /* Simple min distance search */
6982 val = node_distance(node, n);
6984 if (val < min_val) {
6990 if (best_node != -1)
6991 node_set(best_node, *used_nodes);
6996 * sched_domain_node_span - get a cpumask for a node's sched_domain
6997 * @node: node whose cpumask we're constructing
6998 * @span: resulting cpumask
7000 * Given a node, construct a good cpumask for its sched_domain to span. It
7001 * should be one that prevents unnecessary balancing, but also spreads tasks
7004 static void sched_domain_node_span(int node, struct cpumask *span)
7006 nodemask_t used_nodes;
7009 cpumask_clear(span);
7010 nodes_clear(used_nodes);
7012 cpumask_or(span, span, cpumask_of_node(node));
7013 node_set(node, used_nodes);
7015 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7016 int next_node = find_next_best_node(node, &used_nodes);
7019 cpumask_or(span, span, cpumask_of_node(next_node));
7023 static const struct cpumask *cpu_node_mask(int cpu)
7025 lockdep_assert_held(&sched_domains_mutex);
7027 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
7029 return sched_domains_tmpmask;
7032 static const struct cpumask *cpu_allnodes_mask(int cpu)
7034 return cpu_possible_mask;
7036 #endif /* CONFIG_NUMA */
7038 static const struct cpumask *cpu_cpu_mask(int cpu)
7040 return cpumask_of_node(cpu_to_node(cpu));
7043 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7046 struct sched_domain **__percpu sd;
7047 struct sched_group **__percpu sg;
7048 struct sched_group_power **__percpu sgp;
7052 struct sched_domain ** __percpu sd;
7053 struct root_domain *rd;
7063 struct sched_domain_topology_level;
7065 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
7066 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
7068 #define SDTL_OVERLAP 0x01
7070 struct sched_domain_topology_level {
7071 sched_domain_init_f init;
7072 sched_domain_mask_f mask;
7074 struct sd_data data;
7078 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
7080 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
7081 const struct cpumask *span = sched_domain_span(sd);
7082 struct cpumask *covered = sched_domains_tmpmask;
7083 struct sd_data *sdd = sd->private;
7084 struct sched_domain *child;
7087 cpumask_clear(covered);
7089 for_each_cpu(i, span) {
7090 struct cpumask *sg_span;
7092 if (cpumask_test_cpu(i, covered))
7095 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7096 GFP_KERNEL, cpu_to_node(i));
7101 sg_span = sched_group_cpus(sg);
7103 child = *per_cpu_ptr(sdd->sd, i);
7105 child = child->child;
7106 cpumask_copy(sg_span, sched_domain_span(child));
7108 cpumask_set_cpu(i, sg_span);
7110 cpumask_or(covered, covered, sg_span);
7112 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
7113 atomic_inc(&sg->sgp->ref);
7115 if (cpumask_test_cpu(cpu, sg_span))
7125 sd->groups = groups;
7130 free_sched_groups(first, 0);
7135 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
7137 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
7138 struct sched_domain *child = sd->child;
7141 cpu = cpumask_first(sched_domain_span(child));
7144 *sg = *per_cpu_ptr(sdd->sg, cpu);
7145 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
7146 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
7153 * build_sched_groups will build a circular linked list of the groups
7154 * covered by the given span, and will set each group's ->cpumask correctly,
7155 * and ->cpu_power to 0.
7157 * Assumes the sched_domain tree is fully constructed
7160 build_sched_groups(struct sched_domain *sd, int cpu)
7162 struct sched_group *first = NULL, *last = NULL;
7163 struct sd_data *sdd = sd->private;
7164 const struct cpumask *span = sched_domain_span(sd);
7165 struct cpumask *covered;
7168 get_group(cpu, sdd, &sd->groups);
7169 atomic_inc(&sd->groups->ref);
7171 if (cpu != cpumask_first(sched_domain_span(sd)))
7174 lockdep_assert_held(&sched_domains_mutex);
7175 covered = sched_domains_tmpmask;
7177 cpumask_clear(covered);
7179 for_each_cpu(i, span) {
7180 struct sched_group *sg;
7181 int group = get_group(i, sdd, &sg);
7184 if (cpumask_test_cpu(i, covered))
7187 cpumask_clear(sched_group_cpus(sg));
7190 for_each_cpu(j, span) {
7191 if (get_group(j, sdd, NULL) != group)
7194 cpumask_set_cpu(j, covered);
7195 cpumask_set_cpu(j, sched_group_cpus(sg));
7210 * Initialize sched groups cpu_power.
7212 * cpu_power indicates the capacity of sched group, which is used while
7213 * distributing the load between different sched groups in a sched domain.
7214 * Typically cpu_power for all the groups in a sched domain will be same unless
7215 * there are asymmetries in the topology. If there are asymmetries, group
7216 * having more cpu_power will pickup more load compared to the group having
7219 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7221 struct sched_group *sg = sd->groups;
7223 WARN_ON(!sd || !sg);
7226 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
7228 } while (sg != sd->groups);
7230 if (cpu != group_first_cpu(sg))
7233 update_group_power(sd, cpu);
7237 * Initializers for schedule domains
7238 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7241 #ifdef CONFIG_SCHED_DEBUG
7242 # define SD_INIT_NAME(sd, type) sd->name = #type
7244 # define SD_INIT_NAME(sd, type) do { } while (0)
7247 #define SD_INIT_FUNC(type) \
7248 static noinline struct sched_domain * \
7249 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7251 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7252 *sd = SD_##type##_INIT; \
7253 SD_INIT_NAME(sd, type); \
7254 sd->private = &tl->data; \
7260 SD_INIT_FUNC(ALLNODES)
7263 #ifdef CONFIG_SCHED_SMT
7264 SD_INIT_FUNC(SIBLING)
7266 #ifdef CONFIG_SCHED_MC
7269 #ifdef CONFIG_SCHED_BOOK
7273 static int default_relax_domain_level = -1;
7274 int sched_domain_level_max;
7276 static int __init setup_relax_domain_level(char *str)
7280 val = simple_strtoul(str, NULL, 0);
7281 if (val < sched_domain_level_max)
7282 default_relax_domain_level = val;
7286 __setup("relax_domain_level=", setup_relax_domain_level);
7288 static void set_domain_attribute(struct sched_domain *sd,
7289 struct sched_domain_attr *attr)
7293 if (!attr || attr->relax_domain_level < 0) {
7294 if (default_relax_domain_level < 0)
7297 request = default_relax_domain_level;
7299 request = attr->relax_domain_level;
7300 if (request < sd->level) {
7301 /* turn off idle balance on this domain */
7302 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7304 /* turn on idle balance on this domain */
7305 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7309 static void __sdt_free(const struct cpumask *cpu_map);
7310 static int __sdt_alloc(const struct cpumask *cpu_map);
7312 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7313 const struct cpumask *cpu_map)
7317 if (!atomic_read(&d->rd->refcount))
7318 free_rootdomain(&d->rd->rcu); /* fall through */
7320 free_percpu(d->sd); /* fall through */
7322 __sdt_free(cpu_map); /* fall through */
7328 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7329 const struct cpumask *cpu_map)
7331 memset(d, 0, sizeof(*d));
7333 if (__sdt_alloc(cpu_map))
7334 return sa_sd_storage;
7335 d->sd = alloc_percpu(struct sched_domain *);
7337 return sa_sd_storage;
7338 d->rd = alloc_rootdomain();
7341 return sa_rootdomain;
7345 * NULL the sd_data elements we've used to build the sched_domain and
7346 * sched_group structure so that the subsequent __free_domain_allocs()
7347 * will not free the data we're using.
7349 static void claim_allocations(int cpu, struct sched_domain *sd)
7351 struct sd_data *sdd = sd->private;
7353 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7354 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7356 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
7357 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7359 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
7360 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
7363 #ifdef CONFIG_SCHED_SMT
7364 static const struct cpumask *cpu_smt_mask(int cpu)
7366 return topology_thread_cpumask(cpu);
7371 * Topology list, bottom-up.
7373 static struct sched_domain_topology_level default_topology[] = {
7374 #ifdef CONFIG_SCHED_SMT
7375 { sd_init_SIBLING, cpu_smt_mask, },
7377 #ifdef CONFIG_SCHED_MC
7378 { sd_init_MC, cpu_coregroup_mask, },
7380 #ifdef CONFIG_SCHED_BOOK
7381 { sd_init_BOOK, cpu_book_mask, },
7383 { sd_init_CPU, cpu_cpu_mask, },
7385 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
7386 { sd_init_ALLNODES, cpu_allnodes_mask, },
7391 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7393 static int __sdt_alloc(const struct cpumask *cpu_map)
7395 struct sched_domain_topology_level *tl;
7398 for (tl = sched_domain_topology; tl->init; tl++) {
7399 struct sd_data *sdd = &tl->data;
7401 sdd->sd = alloc_percpu(struct sched_domain *);
7405 sdd->sg = alloc_percpu(struct sched_group *);
7409 sdd->sgp = alloc_percpu(struct sched_group_power *);
7413 for_each_cpu(j, cpu_map) {
7414 struct sched_domain *sd;
7415 struct sched_group *sg;
7416 struct sched_group_power *sgp;
7418 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7419 GFP_KERNEL, cpu_to_node(j));
7423 *per_cpu_ptr(sdd->sd, j) = sd;
7425 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7426 GFP_KERNEL, cpu_to_node(j));
7430 *per_cpu_ptr(sdd->sg, j) = sg;
7432 sgp = kzalloc_node(sizeof(struct sched_group_power),
7433 GFP_KERNEL, cpu_to_node(j));
7437 *per_cpu_ptr(sdd->sgp, j) = sgp;
7444 static void __sdt_free(const struct cpumask *cpu_map)
7446 struct sched_domain_topology_level *tl;
7449 for (tl = sched_domain_topology; tl->init; tl++) {
7450 struct sd_data *sdd = &tl->data;
7452 for_each_cpu(j, cpu_map) {
7453 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
7454 if (sd && (sd->flags & SD_OVERLAP))
7455 free_sched_groups(sd->groups, 0);
7456 kfree(*per_cpu_ptr(sdd->sd, j));
7457 kfree(*per_cpu_ptr(sdd->sg, j));
7458 kfree(*per_cpu_ptr(sdd->sgp, j));
7460 free_percpu(sdd->sd);
7461 free_percpu(sdd->sg);
7462 free_percpu(sdd->sgp);
7466 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7467 struct s_data *d, const struct cpumask *cpu_map,
7468 struct sched_domain_attr *attr, struct sched_domain *child,
7471 struct sched_domain *sd = tl->init(tl, cpu);
7475 set_domain_attribute(sd, attr);
7476 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7478 sd->level = child->level + 1;
7479 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7488 * Build sched domains for a given set of cpus and attach the sched domains
7489 * to the individual cpus
7491 static int build_sched_domains(const struct cpumask *cpu_map,
7492 struct sched_domain_attr *attr)
7494 enum s_alloc alloc_state = sa_none;
7495 struct sched_domain *sd;
7497 int i, ret = -ENOMEM;
7499 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7500 if (alloc_state != sa_rootdomain)
7503 /* Set up domains for cpus specified by the cpu_map. */
7504 for_each_cpu(i, cpu_map) {
7505 struct sched_domain_topology_level *tl;
7508 for (tl = sched_domain_topology; tl->init; tl++) {
7509 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7510 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7511 sd->flags |= SD_OVERLAP;
7512 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7519 *per_cpu_ptr(d.sd, i) = sd;
7522 /* Build the groups for the domains */
7523 for_each_cpu(i, cpu_map) {
7524 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7525 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7526 if (sd->flags & SD_OVERLAP) {
7527 if (build_overlap_sched_groups(sd, i))
7530 if (build_sched_groups(sd, i))
7536 /* Calculate CPU power for physical packages and nodes */
7537 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7538 if (!cpumask_test_cpu(i, cpu_map))
7541 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7542 claim_allocations(i, sd);
7543 init_sched_groups_power(i, sd);
7547 /* Attach the domains */
7549 for_each_cpu(i, cpu_map) {
7550 sd = *per_cpu_ptr(d.sd, i);
7551 cpu_attach_domain(sd, d.rd, i);
7557 __free_domain_allocs(&d, alloc_state, cpu_map);
7561 static cpumask_var_t *doms_cur; /* current sched domains */
7562 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7563 static struct sched_domain_attr *dattr_cur;
7564 /* attribues of custom domains in 'doms_cur' */
7567 * Special case: If a kmalloc of a doms_cur partition (array of
7568 * cpumask) fails, then fallback to a single sched domain,
7569 * as determined by the single cpumask fallback_doms.
7571 static cpumask_var_t fallback_doms;
7574 * arch_update_cpu_topology lets virtualized architectures update the
7575 * cpu core maps. It is supposed to return 1 if the topology changed
7576 * or 0 if it stayed the same.
7578 int __attribute__((weak)) arch_update_cpu_topology(void)
7583 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7586 cpumask_var_t *doms;
7588 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7591 for (i = 0; i < ndoms; i++) {
7592 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7593 free_sched_domains(doms, i);
7600 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7603 for (i = 0; i < ndoms; i++)
7604 free_cpumask_var(doms[i]);
7609 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7610 * For now this just excludes isolated cpus, but could be used to
7611 * exclude other special cases in the future.
7613 static int init_sched_domains(const struct cpumask *cpu_map)
7617 arch_update_cpu_topology();
7619 doms_cur = alloc_sched_domains(ndoms_cur);
7621 doms_cur = &fallback_doms;
7622 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7624 err = build_sched_domains(doms_cur[0], NULL);
7625 register_sched_domain_sysctl();
7631 * Detach sched domains from a group of cpus specified in cpu_map
7632 * These cpus will now be attached to the NULL domain
7634 static void detach_destroy_domains(const struct cpumask *cpu_map)
7639 for_each_cpu(i, cpu_map)
7640 cpu_attach_domain(NULL, &def_root_domain, i);
7644 /* handle null as "default" */
7645 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7646 struct sched_domain_attr *new, int idx_new)
7648 struct sched_domain_attr tmp;
7655 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7656 new ? (new + idx_new) : &tmp,
7657 sizeof(struct sched_domain_attr));
7661 * Partition sched domains as specified by the 'ndoms_new'
7662 * cpumasks in the array doms_new[] of cpumasks. This compares
7663 * doms_new[] to the current sched domain partitioning, doms_cur[].
7664 * It destroys each deleted domain and builds each new domain.
7666 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7667 * The masks don't intersect (don't overlap.) We should setup one
7668 * sched domain for each mask. CPUs not in any of the cpumasks will
7669 * not be load balanced. If the same cpumask appears both in the
7670 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7673 * The passed in 'doms_new' should be allocated using
7674 * alloc_sched_domains. This routine takes ownership of it and will
7675 * free_sched_domains it when done with it. If the caller failed the
7676 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7677 * and partition_sched_domains() will fallback to the single partition
7678 * 'fallback_doms', it also forces the domains to be rebuilt.
7680 * If doms_new == NULL it will be replaced with cpu_online_mask.
7681 * ndoms_new == 0 is a special case for destroying existing domains,
7682 * and it will not create the default domain.
7684 * Call with hotplug lock held
7686 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7687 struct sched_domain_attr *dattr_new)
7692 mutex_lock(&sched_domains_mutex);
7694 /* always unregister in case we don't destroy any domains */
7695 unregister_sched_domain_sysctl();
7697 /* Let architecture update cpu core mappings. */
7698 new_topology = arch_update_cpu_topology();
7700 n = doms_new ? ndoms_new : 0;
7702 /* Destroy deleted domains */
7703 for (i = 0; i < ndoms_cur; i++) {
7704 for (j = 0; j < n && !new_topology; j++) {
7705 if (cpumask_equal(doms_cur[i], doms_new[j])
7706 && dattrs_equal(dattr_cur, i, dattr_new, j))
7709 /* no match - a current sched domain not in new doms_new[] */
7710 detach_destroy_domains(doms_cur[i]);
7715 if (doms_new == NULL) {
7717 doms_new = &fallback_doms;
7718 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7719 WARN_ON_ONCE(dattr_new);
7722 /* Build new domains */
7723 for (i = 0; i < ndoms_new; i++) {
7724 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7725 if (cpumask_equal(doms_new[i], doms_cur[j])
7726 && dattrs_equal(dattr_new, i, dattr_cur, j))
7729 /* no match - add a new doms_new */
7730 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7735 /* Remember the new sched domains */
7736 if (doms_cur != &fallback_doms)
7737 free_sched_domains(doms_cur, ndoms_cur);
7738 kfree(dattr_cur); /* kfree(NULL) is safe */
7739 doms_cur = doms_new;
7740 dattr_cur = dattr_new;
7741 ndoms_cur = ndoms_new;
7743 register_sched_domain_sysctl();
7745 mutex_unlock(&sched_domains_mutex);
7748 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7749 static void reinit_sched_domains(void)
7753 /* Destroy domains first to force the rebuild */
7754 partition_sched_domains(0, NULL, NULL);
7756 rebuild_sched_domains();
7760 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7762 unsigned int level = 0;
7764 if (sscanf(buf, "%u", &level) != 1)
7768 * level is always be positive so don't check for
7769 * level < POWERSAVINGS_BALANCE_NONE which is 0
7770 * What happens on 0 or 1 byte write,
7771 * need to check for count as well?
7774 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7778 sched_smt_power_savings = level;
7780 sched_mc_power_savings = level;
7782 reinit_sched_domains();
7787 #ifdef CONFIG_SCHED_MC
7788 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7789 struct sysdev_class_attribute *attr,
7792 return sprintf(page, "%u\n", sched_mc_power_savings);
7794 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7795 struct sysdev_class_attribute *attr,
7796 const char *buf, size_t count)
7798 return sched_power_savings_store(buf, count, 0);
7800 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7801 sched_mc_power_savings_show,
7802 sched_mc_power_savings_store);
7805 #ifdef CONFIG_SCHED_SMT
7806 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7807 struct sysdev_class_attribute *attr,
7810 return sprintf(page, "%u\n", sched_smt_power_savings);
7812 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7813 struct sysdev_class_attribute *attr,
7814 const char *buf, size_t count)
7816 return sched_power_savings_store(buf, count, 1);
7818 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7819 sched_smt_power_savings_show,
7820 sched_smt_power_savings_store);
7823 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7827 #ifdef CONFIG_SCHED_SMT
7829 err = sysfs_create_file(&cls->kset.kobj,
7830 &attr_sched_smt_power_savings.attr);
7832 #ifdef CONFIG_SCHED_MC
7833 if (!err && mc_capable())
7834 err = sysfs_create_file(&cls->kset.kobj,
7835 &attr_sched_mc_power_savings.attr);
7839 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7842 * Update cpusets according to cpu_active mask. If cpusets are
7843 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7844 * around partition_sched_domains().
7846 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7849 switch (action & ~CPU_TASKS_FROZEN) {
7851 case CPU_DOWN_FAILED:
7852 cpuset_update_active_cpus();
7859 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7862 switch (action & ~CPU_TASKS_FROZEN) {
7863 case CPU_DOWN_PREPARE:
7864 cpuset_update_active_cpus();
7871 static int update_runtime(struct notifier_block *nfb,
7872 unsigned long action, void *hcpu)
7874 int cpu = (int)(long)hcpu;
7877 case CPU_DOWN_PREPARE:
7878 case CPU_DOWN_PREPARE_FROZEN:
7879 disable_runtime(cpu_rq(cpu));
7882 case CPU_DOWN_FAILED:
7883 case CPU_DOWN_FAILED_FROZEN:
7885 case CPU_ONLINE_FROZEN:
7886 enable_runtime(cpu_rq(cpu));
7894 void __init sched_init_smp(void)
7896 cpumask_var_t non_isolated_cpus;
7898 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7899 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7902 mutex_lock(&sched_domains_mutex);
7903 init_sched_domains(cpu_active_mask);
7904 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7905 if (cpumask_empty(non_isolated_cpus))
7906 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7907 mutex_unlock(&sched_domains_mutex);
7910 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7911 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7913 /* RT runtime code needs to handle some hotplug events */
7914 hotcpu_notifier(update_runtime, 0);
7918 /* Move init over to a non-isolated CPU */
7919 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7921 sched_init_granularity();
7922 free_cpumask_var(non_isolated_cpus);
7924 init_sched_rt_class();
7927 void __init sched_init_smp(void)
7929 sched_init_granularity();
7931 #endif /* CONFIG_SMP */
7933 const_debug unsigned int sysctl_timer_migration = 1;
7935 int in_sched_functions(unsigned long addr)
7937 return in_lock_functions(addr) ||
7938 (addr >= (unsigned long)__sched_text_start
7939 && addr < (unsigned long)__sched_text_end);
7942 static void init_cfs_rq(struct cfs_rq *cfs_rq)
7944 cfs_rq->tasks_timeline = RB_ROOT;
7945 INIT_LIST_HEAD(&cfs_rq->tasks);
7946 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7947 #ifndef CONFIG_64BIT
7948 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7952 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7954 struct rt_prio_array *array;
7957 array = &rt_rq->active;
7958 for (i = 0; i < MAX_RT_PRIO; i++) {
7959 INIT_LIST_HEAD(array->queue + i);
7960 __clear_bit(i, array->bitmap);
7962 /* delimiter for bitsearch: */
7963 __set_bit(MAX_RT_PRIO, array->bitmap);
7965 #if defined CONFIG_SMP
7966 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7967 rt_rq->highest_prio.next = MAX_RT_PRIO;
7968 rt_rq->rt_nr_migratory = 0;
7969 rt_rq->overloaded = 0;
7970 plist_head_init(&rt_rq->pushable_tasks);
7974 rt_rq->rt_throttled = 0;
7975 rt_rq->rt_runtime = 0;
7976 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7979 #ifdef CONFIG_FAIR_GROUP_SCHED
7980 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7981 struct sched_entity *se, int cpu,
7982 struct sched_entity *parent)
7984 struct rq *rq = cpu_rq(cpu);
7989 /* allow initial update_cfs_load() to truncate */
7990 cfs_rq->load_stamp = 1;
7993 tg->cfs_rq[cpu] = cfs_rq;
7996 /* se could be NULL for root_task_group */
8001 se->cfs_rq = &rq->cfs;
8003 se->cfs_rq = parent->my_q;
8006 update_load_set(&se->load, 0);
8007 se->parent = parent;
8011 #ifdef CONFIG_RT_GROUP_SCHED
8012 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8013 struct sched_rt_entity *rt_se, int cpu,
8014 struct sched_rt_entity *parent)
8016 struct rq *rq = cpu_rq(cpu);
8018 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8019 rt_rq->rt_nr_boosted = 0;
8023 tg->rt_rq[cpu] = rt_rq;
8024 tg->rt_se[cpu] = rt_se;
8030 rt_se->rt_rq = &rq->rt;
8032 rt_se->rt_rq = parent->my_q;
8034 rt_se->my_q = rt_rq;
8035 rt_se->parent = parent;
8036 INIT_LIST_HEAD(&rt_se->run_list);
8040 void __init sched_init(void)
8043 unsigned long alloc_size = 0, ptr;
8045 #ifdef CONFIG_FAIR_GROUP_SCHED
8046 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8048 #ifdef CONFIG_RT_GROUP_SCHED
8049 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8051 #ifdef CONFIG_CPUMASK_OFFSTACK
8052 alloc_size += num_possible_cpus() * cpumask_size();
8055 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8057 #ifdef CONFIG_FAIR_GROUP_SCHED
8058 root_task_group.se = (struct sched_entity **)ptr;
8059 ptr += nr_cpu_ids * sizeof(void **);
8061 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8062 ptr += nr_cpu_ids * sizeof(void **);
8064 #endif /* CONFIG_FAIR_GROUP_SCHED */
8065 #ifdef CONFIG_RT_GROUP_SCHED
8066 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8067 ptr += nr_cpu_ids * sizeof(void **);
8069 root_task_group.rt_rq = (struct rt_rq **)ptr;
8070 ptr += nr_cpu_ids * sizeof(void **);
8072 #endif /* CONFIG_RT_GROUP_SCHED */
8073 #ifdef CONFIG_CPUMASK_OFFSTACK
8074 for_each_possible_cpu(i) {
8075 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8076 ptr += cpumask_size();
8078 #endif /* CONFIG_CPUMASK_OFFSTACK */
8082 init_defrootdomain();
8085 init_rt_bandwidth(&def_rt_bandwidth,
8086 global_rt_period(), global_rt_runtime());
8088 #ifdef CONFIG_RT_GROUP_SCHED
8089 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8090 global_rt_period(), global_rt_runtime());
8091 #endif /* CONFIG_RT_GROUP_SCHED */
8093 #ifdef CONFIG_CGROUP_SCHED
8094 list_add(&root_task_group.list, &task_groups);
8095 INIT_LIST_HEAD(&root_task_group.children);
8096 autogroup_init(&init_task);
8097 #endif /* CONFIG_CGROUP_SCHED */
8099 for_each_possible_cpu(i) {
8103 raw_spin_lock_init(&rq->lock);
8105 rq->calc_load_active = 0;
8106 rq->calc_load_update = jiffies + LOAD_FREQ;
8107 init_cfs_rq(&rq->cfs);
8108 init_rt_rq(&rq->rt, rq);
8109 #ifdef CONFIG_FAIR_GROUP_SCHED
8110 root_task_group.shares = root_task_group_load;
8111 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8113 * How much cpu bandwidth does root_task_group get?
8115 * In case of task-groups formed thr' the cgroup filesystem, it
8116 * gets 100% of the cpu resources in the system. This overall
8117 * system cpu resource is divided among the tasks of
8118 * root_task_group and its child task-groups in a fair manner,
8119 * based on each entity's (task or task-group's) weight
8120 * (se->load.weight).
8122 * In other words, if root_task_group has 10 tasks of weight
8123 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8124 * then A0's share of the cpu resource is:
8126 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8128 * We achieve this by letting root_task_group's tasks sit
8129 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8131 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8132 #endif /* CONFIG_FAIR_GROUP_SCHED */
8134 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8135 #ifdef CONFIG_RT_GROUP_SCHED
8136 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8137 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8140 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8141 rq->cpu_load[j] = 0;
8143 rq->last_load_update_tick = jiffies;
8148 rq->cpu_power = SCHED_POWER_SCALE;
8149 rq->post_schedule = 0;
8150 rq->active_balance = 0;
8151 rq->next_balance = jiffies;
8156 rq->avg_idle = 2*sysctl_sched_migration_cost;
8157 rq_attach_root(rq, &def_root_domain);
8159 rq->nohz_balance_kick = 0;
8160 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8164 atomic_set(&rq->nr_iowait, 0);
8167 set_load_weight(&init_task);
8169 #ifdef CONFIG_PREEMPT_NOTIFIERS
8170 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8174 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8177 #ifdef CONFIG_RT_MUTEXES
8178 plist_head_init(&init_task.pi_waiters);
8182 * The boot idle thread does lazy MMU switching as well:
8184 atomic_inc(&init_mm.mm_count);
8185 enter_lazy_tlb(&init_mm, current);
8188 * Make us the idle thread. Technically, schedule() should not be
8189 * called from this thread, however somewhere below it might be,
8190 * but because we are the idle thread, we just pick up running again
8191 * when this runqueue becomes "idle".
8193 init_idle(current, smp_processor_id());
8195 calc_load_update = jiffies + LOAD_FREQ;
8198 * During early bootup we pretend to be a normal task:
8200 current->sched_class = &fair_sched_class;
8202 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8203 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8205 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8207 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8208 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8209 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8210 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8211 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8213 /* May be allocated at isolcpus cmdline parse time */
8214 if (cpu_isolated_map == NULL)
8215 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8218 scheduler_running = 1;
8221 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8222 static inline int preempt_count_equals(int preempt_offset)
8224 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8226 return (nested == preempt_offset);
8229 void __might_sleep(const char *file, int line, int preempt_offset)
8231 static unsigned long prev_jiffy; /* ratelimiting */
8233 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8234 system_state != SYSTEM_RUNNING || oops_in_progress)
8236 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8238 prev_jiffy = jiffies;
8241 "BUG: sleeping function called from invalid context at %s:%d\n",
8244 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8245 in_atomic(), irqs_disabled(),
8246 current->pid, current->comm);
8248 debug_show_held_locks(current);
8249 if (irqs_disabled())
8250 print_irqtrace_events(current);
8253 EXPORT_SYMBOL(__might_sleep);
8256 #ifdef CONFIG_MAGIC_SYSRQ
8257 static void normalize_task(struct rq *rq, struct task_struct *p)
8259 const struct sched_class *prev_class = p->sched_class;
8260 int old_prio = p->prio;
8265 deactivate_task(rq, p, 0);
8266 __setscheduler(rq, p, SCHED_NORMAL, 0);
8268 activate_task(rq, p, 0);
8269 resched_task(rq->curr);
8272 check_class_changed(rq, p, prev_class, old_prio);
8275 void normalize_rt_tasks(void)
8277 struct task_struct *g, *p;
8278 unsigned long flags;
8281 read_lock_irqsave(&tasklist_lock, flags);
8282 do_each_thread(g, p) {
8284 * Only normalize user tasks:
8289 p->se.exec_start = 0;
8290 #ifdef CONFIG_SCHEDSTATS
8291 p->se.statistics.wait_start = 0;
8292 p->se.statistics.sleep_start = 0;
8293 p->se.statistics.block_start = 0;
8298 * Renice negative nice level userspace
8301 if (TASK_NICE(p) < 0 && p->mm)
8302 set_user_nice(p, 0);
8306 raw_spin_lock(&p->pi_lock);
8307 rq = __task_rq_lock(p);
8309 normalize_task(rq, p);
8311 __task_rq_unlock(rq);
8312 raw_spin_unlock(&p->pi_lock);
8313 } while_each_thread(g, p);
8315 read_unlock_irqrestore(&tasklist_lock, flags);
8318 #endif /* CONFIG_MAGIC_SYSRQ */
8320 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8322 * These functions are only useful for the IA64 MCA handling, or kdb.
8324 * They can only be called when the whole system has been
8325 * stopped - every CPU needs to be quiescent, and no scheduling
8326 * activity can take place. Using them for anything else would
8327 * be a serious bug, and as a result, they aren't even visible
8328 * under any other configuration.
8332 * curr_task - return the current task for a given cpu.
8333 * @cpu: the processor in question.
8335 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8337 struct task_struct *curr_task(int cpu)
8339 return cpu_curr(cpu);
8342 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8346 * set_curr_task - set the current task for a given cpu.
8347 * @cpu: the processor in question.
8348 * @p: the task pointer to set.
8350 * Description: This function must only be used when non-maskable interrupts
8351 * are serviced on a separate stack. It allows the architecture to switch the
8352 * notion of the current task on a cpu in a non-blocking manner. This function
8353 * must be called with all CPU's synchronized, and interrupts disabled, the
8354 * and caller must save the original value of the current task (see
8355 * curr_task() above) and restore that value before reenabling interrupts and
8356 * re-starting the system.
8358 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8360 void set_curr_task(int cpu, struct task_struct *p)
8367 #ifdef CONFIG_FAIR_GROUP_SCHED
8368 static void free_fair_sched_group(struct task_group *tg)
8372 for_each_possible_cpu(i) {
8374 kfree(tg->cfs_rq[i]);
8384 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8386 struct cfs_rq *cfs_rq;
8387 struct sched_entity *se;
8390 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8393 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8397 tg->shares = NICE_0_LOAD;
8399 for_each_possible_cpu(i) {
8400 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8401 GFP_KERNEL, cpu_to_node(i));
8405 se = kzalloc_node(sizeof(struct sched_entity),
8406 GFP_KERNEL, cpu_to_node(i));
8410 init_cfs_rq(cfs_rq);
8411 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8422 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8424 struct rq *rq = cpu_rq(cpu);
8425 unsigned long flags;
8428 * Only empty task groups can be destroyed; so we can speculatively
8429 * check on_list without danger of it being re-added.
8431 if (!tg->cfs_rq[cpu]->on_list)
8434 raw_spin_lock_irqsave(&rq->lock, flags);
8435 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8436 raw_spin_unlock_irqrestore(&rq->lock, flags);
8438 #else /* !CONFIG_FAIR_GROUP_SCHED */
8439 static inline void free_fair_sched_group(struct task_group *tg)
8444 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8449 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8452 #endif /* CONFIG_FAIR_GROUP_SCHED */
8454 #ifdef CONFIG_RT_GROUP_SCHED
8455 static void free_rt_sched_group(struct task_group *tg)
8460 destroy_rt_bandwidth(&tg->rt_bandwidth);
8462 for_each_possible_cpu(i) {
8464 kfree(tg->rt_rq[i]);
8466 kfree(tg->rt_se[i]);
8474 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8476 struct rt_rq *rt_rq;
8477 struct sched_rt_entity *rt_se;
8480 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8483 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8487 init_rt_bandwidth(&tg->rt_bandwidth,
8488 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8490 for_each_possible_cpu(i) {
8491 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8492 GFP_KERNEL, cpu_to_node(i));
8496 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8497 GFP_KERNEL, cpu_to_node(i));
8501 init_rt_rq(rt_rq, cpu_rq(i));
8502 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8503 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8513 #else /* !CONFIG_RT_GROUP_SCHED */
8514 static inline void free_rt_sched_group(struct task_group *tg)
8519 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8523 #endif /* CONFIG_RT_GROUP_SCHED */
8525 #ifdef CONFIG_CGROUP_SCHED
8526 static void free_sched_group(struct task_group *tg)
8528 free_fair_sched_group(tg);
8529 free_rt_sched_group(tg);
8534 /* allocate runqueue etc for a new task group */
8535 struct task_group *sched_create_group(struct task_group *parent)
8537 struct task_group *tg;
8538 unsigned long flags;
8540 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8542 return ERR_PTR(-ENOMEM);
8544 if (!alloc_fair_sched_group(tg, parent))
8547 if (!alloc_rt_sched_group(tg, parent))
8550 spin_lock_irqsave(&task_group_lock, flags);
8551 list_add_rcu(&tg->list, &task_groups);
8553 WARN_ON(!parent); /* root should already exist */
8555 tg->parent = parent;
8556 INIT_LIST_HEAD(&tg->children);
8557 list_add_rcu(&tg->siblings, &parent->children);
8558 spin_unlock_irqrestore(&task_group_lock, flags);
8563 free_sched_group(tg);
8564 return ERR_PTR(-ENOMEM);
8567 /* rcu callback to free various structures associated with a task group */
8568 static void free_sched_group_rcu(struct rcu_head *rhp)
8570 /* now it should be safe to free those cfs_rqs */
8571 free_sched_group(container_of(rhp, struct task_group, rcu));
8574 /* Destroy runqueue etc associated with a task group */
8575 void sched_destroy_group(struct task_group *tg)
8577 unsigned long flags;
8580 /* end participation in shares distribution */
8581 for_each_possible_cpu(i)
8582 unregister_fair_sched_group(tg, i);
8584 spin_lock_irqsave(&task_group_lock, flags);
8585 list_del_rcu(&tg->list);
8586 list_del_rcu(&tg->siblings);
8587 spin_unlock_irqrestore(&task_group_lock, flags);
8589 /* wait for possible concurrent references to cfs_rqs complete */
8590 call_rcu(&tg->rcu, free_sched_group_rcu);
8593 /* change task's runqueue when it moves between groups.
8594 * The caller of this function should have put the task in its new group
8595 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8596 * reflect its new group.
8598 void sched_move_task(struct task_struct *tsk)
8601 unsigned long flags;
8604 rq = task_rq_lock(tsk, &flags);
8606 running = task_current(rq, tsk);
8610 dequeue_task(rq, tsk, 0);
8611 if (unlikely(running))
8612 tsk->sched_class->put_prev_task(rq, tsk);
8614 #ifdef CONFIG_FAIR_GROUP_SCHED
8615 if (tsk->sched_class->task_move_group)
8616 tsk->sched_class->task_move_group(tsk, on_rq);
8619 set_task_rq(tsk, task_cpu(tsk));
8621 if (unlikely(running))
8622 tsk->sched_class->set_curr_task(rq);
8624 enqueue_task(rq, tsk, 0);
8626 task_rq_unlock(rq, tsk, &flags);
8628 #endif /* CONFIG_CGROUP_SCHED */
8630 #ifdef CONFIG_FAIR_GROUP_SCHED
8631 static DEFINE_MUTEX(shares_mutex);
8633 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8636 unsigned long flags;
8639 * We can't change the weight of the root cgroup.
8644 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8646 mutex_lock(&shares_mutex);
8647 if (tg->shares == shares)
8650 tg->shares = shares;
8651 for_each_possible_cpu(i) {
8652 struct rq *rq = cpu_rq(i);
8653 struct sched_entity *se;
8656 /* Propagate contribution to hierarchy */
8657 raw_spin_lock_irqsave(&rq->lock, flags);
8658 for_each_sched_entity(se)
8659 update_cfs_shares(group_cfs_rq(se));
8660 raw_spin_unlock_irqrestore(&rq->lock, flags);
8664 mutex_unlock(&shares_mutex);
8668 unsigned long sched_group_shares(struct task_group *tg)
8674 #ifdef CONFIG_RT_GROUP_SCHED
8676 * Ensure that the real time constraints are schedulable.
8678 static DEFINE_MUTEX(rt_constraints_mutex);
8680 static unsigned long to_ratio(u64 period, u64 runtime)
8682 if (runtime == RUNTIME_INF)
8685 return div64_u64(runtime << 20, period);
8688 /* Must be called with tasklist_lock held */
8689 static inline int tg_has_rt_tasks(struct task_group *tg)
8691 struct task_struct *g, *p;
8693 do_each_thread(g, p) {
8694 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8696 } while_each_thread(g, p);
8701 struct rt_schedulable_data {
8702 struct task_group *tg;
8707 static int tg_schedulable(struct task_group *tg, void *data)
8709 struct rt_schedulable_data *d = data;
8710 struct task_group *child;
8711 unsigned long total, sum = 0;
8712 u64 period, runtime;
8714 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8715 runtime = tg->rt_bandwidth.rt_runtime;
8718 period = d->rt_period;
8719 runtime = d->rt_runtime;
8723 * Cannot have more runtime than the period.
8725 if (runtime > period && runtime != RUNTIME_INF)
8729 * Ensure we don't starve existing RT tasks.
8731 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8734 total = to_ratio(period, runtime);
8737 * Nobody can have more than the global setting allows.
8739 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8743 * The sum of our children's runtime should not exceed our own.
8745 list_for_each_entry_rcu(child, &tg->children, siblings) {
8746 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8747 runtime = child->rt_bandwidth.rt_runtime;
8749 if (child == d->tg) {
8750 period = d->rt_period;
8751 runtime = d->rt_runtime;
8754 sum += to_ratio(period, runtime);
8763 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8765 struct rt_schedulable_data data = {
8767 .rt_period = period,
8768 .rt_runtime = runtime,
8771 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8774 static int tg_set_bandwidth(struct task_group *tg,
8775 u64 rt_period, u64 rt_runtime)
8779 mutex_lock(&rt_constraints_mutex);
8780 read_lock(&tasklist_lock);
8781 err = __rt_schedulable(tg, rt_period, rt_runtime);
8785 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8786 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8787 tg->rt_bandwidth.rt_runtime = rt_runtime;
8789 for_each_possible_cpu(i) {
8790 struct rt_rq *rt_rq = tg->rt_rq[i];
8792 raw_spin_lock(&rt_rq->rt_runtime_lock);
8793 rt_rq->rt_runtime = rt_runtime;
8794 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8796 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8798 read_unlock(&tasklist_lock);
8799 mutex_unlock(&rt_constraints_mutex);
8804 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8806 u64 rt_runtime, rt_period;
8808 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8809 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8810 if (rt_runtime_us < 0)
8811 rt_runtime = RUNTIME_INF;
8813 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8816 long sched_group_rt_runtime(struct task_group *tg)
8820 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8823 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8824 do_div(rt_runtime_us, NSEC_PER_USEC);
8825 return rt_runtime_us;
8828 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8830 u64 rt_runtime, rt_period;
8832 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8833 rt_runtime = tg->rt_bandwidth.rt_runtime;
8838 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8841 long sched_group_rt_period(struct task_group *tg)
8845 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8846 do_div(rt_period_us, NSEC_PER_USEC);
8847 return rt_period_us;
8850 static int sched_rt_global_constraints(void)
8852 u64 runtime, period;
8855 if (sysctl_sched_rt_period <= 0)
8858 runtime = global_rt_runtime();
8859 period = global_rt_period();
8862 * Sanity check on the sysctl variables.
8864 if (runtime > period && runtime != RUNTIME_INF)
8867 mutex_lock(&rt_constraints_mutex);
8868 read_lock(&tasklist_lock);
8869 ret = __rt_schedulable(NULL, 0, 0);
8870 read_unlock(&tasklist_lock);
8871 mutex_unlock(&rt_constraints_mutex);
8876 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8878 /* Don't accept realtime tasks when there is no way for them to run */
8879 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8885 #else /* !CONFIG_RT_GROUP_SCHED */
8886 static int sched_rt_global_constraints(void)
8888 unsigned long flags;
8891 if (sysctl_sched_rt_period <= 0)
8895 * There's always some RT tasks in the root group
8896 * -- migration, kstopmachine etc..
8898 if (sysctl_sched_rt_runtime == 0)
8901 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8902 for_each_possible_cpu(i) {
8903 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8905 raw_spin_lock(&rt_rq->rt_runtime_lock);
8906 rt_rq->rt_runtime = global_rt_runtime();
8907 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8909 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8913 #endif /* CONFIG_RT_GROUP_SCHED */
8915 int sched_rt_handler(struct ctl_table *table, int write,
8916 void __user *buffer, size_t *lenp,
8920 int old_period, old_runtime;
8921 static DEFINE_MUTEX(mutex);
8924 old_period = sysctl_sched_rt_period;
8925 old_runtime = sysctl_sched_rt_runtime;
8927 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8929 if (!ret && write) {
8930 ret = sched_rt_global_constraints();
8932 sysctl_sched_rt_period = old_period;
8933 sysctl_sched_rt_runtime = old_runtime;
8935 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8936 def_rt_bandwidth.rt_period =
8937 ns_to_ktime(global_rt_period());
8940 mutex_unlock(&mutex);
8945 #ifdef CONFIG_CGROUP_SCHED
8947 /* return corresponding task_group object of a cgroup */
8948 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8950 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8951 struct task_group, css);
8954 static struct cgroup_subsys_state *
8955 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8957 struct task_group *tg, *parent;
8959 if (!cgrp->parent) {
8960 /* This is early initialization for the top cgroup */
8961 return &root_task_group.css;
8964 parent = cgroup_tg(cgrp->parent);
8965 tg = sched_create_group(parent);
8967 return ERR_PTR(-ENOMEM);
8973 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8975 struct task_group *tg = cgroup_tg(cgrp);
8977 sched_destroy_group(tg);
8981 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8983 #ifdef CONFIG_RT_GROUP_SCHED
8984 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8987 /* We don't support RT-tasks being in separate groups */
8988 if (tsk->sched_class != &fair_sched_class)
8995 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8997 sched_move_task(tsk);
9001 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
9002 struct cgroup *old_cgrp, struct task_struct *task)
9005 * cgroup_exit() is called in the copy_process() failure path.
9006 * Ignore this case since the task hasn't ran yet, this avoids
9007 * trying to poke a half freed task state from generic code.
9009 if (!(task->flags & PF_EXITING))
9012 sched_move_task(task);
9015 #ifdef CONFIG_FAIR_GROUP_SCHED
9016 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9019 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
9022 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9024 struct task_group *tg = cgroup_tg(cgrp);
9026 return (u64) scale_load_down(tg->shares);
9028 #endif /* CONFIG_FAIR_GROUP_SCHED */
9030 #ifdef CONFIG_RT_GROUP_SCHED
9031 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9034 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9037 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9039 return sched_group_rt_runtime(cgroup_tg(cgrp));
9042 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9045 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9048 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9050 return sched_group_rt_period(cgroup_tg(cgrp));
9052 #endif /* CONFIG_RT_GROUP_SCHED */
9054 static struct cftype cpu_files[] = {
9055 #ifdef CONFIG_FAIR_GROUP_SCHED
9058 .read_u64 = cpu_shares_read_u64,
9059 .write_u64 = cpu_shares_write_u64,
9062 #ifdef CONFIG_RT_GROUP_SCHED
9064 .name = "rt_runtime_us",
9065 .read_s64 = cpu_rt_runtime_read,
9066 .write_s64 = cpu_rt_runtime_write,
9069 .name = "rt_period_us",
9070 .read_u64 = cpu_rt_period_read_uint,
9071 .write_u64 = cpu_rt_period_write_uint,
9076 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9078 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9081 struct cgroup_subsys cpu_cgroup_subsys = {
9083 .create = cpu_cgroup_create,
9084 .destroy = cpu_cgroup_destroy,
9085 .can_attach_task = cpu_cgroup_can_attach_task,
9086 .attach_task = cpu_cgroup_attach_task,
9087 .exit = cpu_cgroup_exit,
9088 .populate = cpu_cgroup_populate,
9089 .subsys_id = cpu_cgroup_subsys_id,
9093 #endif /* CONFIG_CGROUP_SCHED */
9095 #ifdef CONFIG_CGROUP_CPUACCT
9098 * CPU accounting code for task groups.
9100 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9101 * (balbir@in.ibm.com).
9104 /* track cpu usage of a group of tasks and its child groups */
9106 struct cgroup_subsys_state css;
9107 /* cpuusage holds pointer to a u64-type object on every cpu */
9108 u64 __percpu *cpuusage;
9109 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9110 struct cpuacct *parent;
9113 struct cgroup_subsys cpuacct_subsys;
9115 /* return cpu accounting group corresponding to this container */
9116 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9118 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9119 struct cpuacct, css);
9122 /* return cpu accounting group to which this task belongs */
9123 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9125 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9126 struct cpuacct, css);
9129 /* create a new cpu accounting group */
9130 static struct cgroup_subsys_state *cpuacct_create(
9131 struct cgroup_subsys *ss, struct cgroup *cgrp)
9133 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9139 ca->cpuusage = alloc_percpu(u64);
9143 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9144 if (percpu_counter_init(&ca->cpustat[i], 0))
9145 goto out_free_counters;
9148 ca->parent = cgroup_ca(cgrp->parent);
9154 percpu_counter_destroy(&ca->cpustat[i]);
9155 free_percpu(ca->cpuusage);
9159 return ERR_PTR(-ENOMEM);
9162 /* destroy an existing cpu accounting group */
9164 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9166 struct cpuacct *ca = cgroup_ca(cgrp);
9169 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9170 percpu_counter_destroy(&ca->cpustat[i]);
9171 free_percpu(ca->cpuusage);
9175 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9177 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9180 #ifndef CONFIG_64BIT
9182 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9184 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9186 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9194 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9196 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9198 #ifndef CONFIG_64BIT
9200 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9202 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9204 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9210 /* return total cpu usage (in nanoseconds) of a group */
9211 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9213 struct cpuacct *ca = cgroup_ca(cgrp);
9214 u64 totalcpuusage = 0;
9217 for_each_present_cpu(i)
9218 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9220 return totalcpuusage;
9223 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9226 struct cpuacct *ca = cgroup_ca(cgrp);
9235 for_each_present_cpu(i)
9236 cpuacct_cpuusage_write(ca, i, 0);
9242 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9245 struct cpuacct *ca = cgroup_ca(cgroup);
9249 for_each_present_cpu(i) {
9250 percpu = cpuacct_cpuusage_read(ca, i);
9251 seq_printf(m, "%llu ", (unsigned long long) percpu);
9253 seq_printf(m, "\n");
9257 static const char *cpuacct_stat_desc[] = {
9258 [CPUACCT_STAT_USER] = "user",
9259 [CPUACCT_STAT_SYSTEM] = "system",
9262 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9263 struct cgroup_map_cb *cb)
9265 struct cpuacct *ca = cgroup_ca(cgrp);
9268 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9269 s64 val = percpu_counter_read(&ca->cpustat[i]);
9270 val = cputime64_to_clock_t(val);
9271 cb->fill(cb, cpuacct_stat_desc[i], val);
9276 static struct cftype files[] = {
9279 .read_u64 = cpuusage_read,
9280 .write_u64 = cpuusage_write,
9283 .name = "usage_percpu",
9284 .read_seq_string = cpuacct_percpu_seq_read,
9288 .read_map = cpuacct_stats_show,
9292 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9294 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9298 * charge this task's execution time to its accounting group.
9300 * called with rq->lock held.
9302 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9307 if (unlikely(!cpuacct_subsys.active))
9310 cpu = task_cpu(tsk);
9316 for (; ca; ca = ca->parent) {
9317 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9318 *cpuusage += cputime;
9325 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9326 * in cputime_t units. As a result, cpuacct_update_stats calls
9327 * percpu_counter_add with values large enough to always overflow the
9328 * per cpu batch limit causing bad SMP scalability.
9330 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9331 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9332 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9335 #define CPUACCT_BATCH \
9336 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9338 #define CPUACCT_BATCH 0
9342 * Charge the system/user time to the task's accounting group.
9344 static void cpuacct_update_stats(struct task_struct *tsk,
9345 enum cpuacct_stat_index idx, cputime_t val)
9348 int batch = CPUACCT_BATCH;
9350 if (unlikely(!cpuacct_subsys.active))
9357 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9363 struct cgroup_subsys cpuacct_subsys = {
9365 .create = cpuacct_create,
9366 .destroy = cpuacct_destroy,
9367 .populate = cpuacct_populate,
9368 .subsys_id = cpuacct_subsys_id,
9370 #endif /* CONFIG_CGROUP_CPUACCT */