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>
74 #include <linux/cpuacct.h>
77 #include <asm/irq_regs.h>
78 #include <asm/mutex.h>
80 #include "sched_cpupri.h"
81 #include "workqueue_sched.h"
82 #include "sched_autogroup.h"
84 #define CREATE_TRACE_POINTS
85 #include <trace/events/sched.h>
88 * Convert user-nice values [ -20 ... 0 ... 19 ]
89 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
92 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
93 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
94 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
97 * 'User priority' is the nice value converted to something we
98 * can work with better when scaling various scheduler parameters,
99 * it's a [ 0 ... 39 ] range.
101 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
102 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
103 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
106 * Helpers for converting nanosecond timing to jiffy resolution
108 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
110 #define NICE_0_LOAD SCHED_LOAD_SCALE
111 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
114 * These are the 'tuning knobs' of the scheduler:
116 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
117 * Timeslices get refilled after they expire.
119 #define DEF_TIMESLICE (100 * HZ / 1000)
122 * single value that denotes runtime == period, ie unlimited time.
124 #define RUNTIME_INF ((u64)~0ULL)
126 static inline int rt_policy(int policy)
128 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
133 static inline int task_has_rt_policy(struct task_struct *p)
135 return rt_policy(p->policy);
139 * This is the priority-queue data structure of the RT scheduling class:
141 struct rt_prio_array {
142 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
143 struct list_head queue[MAX_RT_PRIO];
146 struct rt_bandwidth {
147 /* nests inside the rq lock: */
148 raw_spinlock_t rt_runtime_lock;
151 struct hrtimer rt_period_timer;
154 static struct rt_bandwidth def_rt_bandwidth;
156 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
158 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
160 struct rt_bandwidth *rt_b =
161 container_of(timer, struct rt_bandwidth, rt_period_timer);
167 now = hrtimer_cb_get_time(timer);
168 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
173 idle = do_sched_rt_period_timer(rt_b, overrun);
176 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
180 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
182 rt_b->rt_period = ns_to_ktime(period);
183 rt_b->rt_runtime = runtime;
185 raw_spin_lock_init(&rt_b->rt_runtime_lock);
187 hrtimer_init(&rt_b->rt_period_timer,
188 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
189 rt_b->rt_period_timer.function = sched_rt_period_timer;
192 static inline int rt_bandwidth_enabled(void)
194 return sysctl_sched_rt_runtime >= 0;
197 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
201 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
204 if (hrtimer_active(&rt_b->rt_period_timer))
207 raw_spin_lock(&rt_b->rt_runtime_lock);
212 if (hrtimer_active(&rt_b->rt_period_timer))
215 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
216 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
218 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
219 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
220 delta = ktime_to_ns(ktime_sub(hard, soft));
221 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
222 HRTIMER_MODE_ABS_PINNED, 0);
224 raw_spin_unlock(&rt_b->rt_runtime_lock);
227 #ifdef CONFIG_RT_GROUP_SCHED
228 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
230 hrtimer_cancel(&rt_b->rt_period_timer);
235 * sched_domains_mutex serializes calls to init_sched_domains,
236 * detach_destroy_domains and partition_sched_domains.
238 static DEFINE_MUTEX(sched_domains_mutex);
240 #ifdef CONFIG_CGROUP_SCHED
242 #include <linux/cgroup.h>
246 static LIST_HEAD(task_groups);
248 /* task group related information */
250 struct cgroup_subsys_state css;
252 #ifdef CONFIG_FAIR_GROUP_SCHED
253 /* schedulable entities of this group on each cpu */
254 struct sched_entity **se;
255 /* runqueue "owned" by this group on each cpu */
256 struct cfs_rq **cfs_rq;
257 unsigned long shares;
259 atomic_t load_weight;
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity **rt_se;
264 struct rt_rq **rt_rq;
266 struct rt_bandwidth rt_bandwidth;
270 struct list_head list;
272 struct task_group *parent;
273 struct list_head siblings;
274 struct list_head children;
276 #ifdef CONFIG_SCHED_AUTOGROUP
277 struct autogroup *autogroup;
281 /* task_group_lock serializes the addition/removal of task groups */
282 static DEFINE_SPINLOCK(task_group_lock);
284 #ifdef CONFIG_FAIR_GROUP_SCHED
286 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
289 * A weight of 0 or 1 can cause arithmetics problems.
290 * A weight of a cfs_rq is the sum of weights of which entities
291 * are queued on this cfs_rq, so a weight of a entity should not be
292 * too large, so as the shares value of a task group.
293 * (The default weight is 1024 - so there's no practical
294 * limitation from this.)
296 #define MIN_SHARES (1UL << 1)
297 #define MAX_SHARES (1UL << 18)
299 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
302 /* Default task group.
303 * Every task in system belong to this group at bootup.
305 struct task_group root_task_group;
307 #endif /* CONFIG_CGROUP_SCHED */
309 /* CFS-related fields in a runqueue */
311 struct load_weight load;
312 unsigned long nr_running;
317 u64 min_vruntime_copy;
320 struct rb_root tasks_timeline;
321 struct rb_node *rb_leftmost;
323 struct list_head tasks;
324 struct list_head *balance_iterator;
327 * 'curr' points to currently running entity on this cfs_rq.
328 * It is set to NULL otherwise (i.e when none are currently running).
330 struct sched_entity *curr, *next, *last, *skip;
332 #ifdef CONFIG_SCHED_DEBUG
333 unsigned int nr_spread_over;
336 #ifdef CONFIG_FAIR_GROUP_SCHED
337 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
340 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
341 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
342 * (like users, containers etc.)
344 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
345 * list is used during load balance.
348 struct list_head leaf_cfs_rq_list;
349 struct task_group *tg; /* group that "owns" this runqueue */
353 * the part of load.weight contributed by tasks
355 unsigned long task_weight;
358 * h_load = weight * f(tg)
360 * Where f(tg) is the recursive weight fraction assigned to
363 unsigned long h_load;
366 * Maintaining per-cpu shares distribution for group scheduling
368 * load_stamp is the last time we updated the load average
369 * load_last is the last time we updated the load average and saw load
370 * load_unacc_exec_time is currently unaccounted execution time
374 u64 load_stamp, load_last, load_unacc_exec_time;
376 unsigned long load_contribution;
381 /* Real-Time classes' related field in a runqueue: */
383 struct rt_prio_array active;
384 unsigned long rt_nr_running;
385 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
387 int curr; /* highest queued rt task prio */
389 int next; /* next highest */
394 unsigned long rt_nr_migratory;
395 unsigned long rt_nr_total;
397 struct plist_head pushable_tasks;
402 /* Nests inside the rq lock: */
403 raw_spinlock_t rt_runtime_lock;
405 #ifdef CONFIG_RT_GROUP_SCHED
406 unsigned long rt_nr_boosted;
409 struct list_head leaf_rt_rq_list;
410 struct task_group *tg;
417 * We add the notion of a root-domain which will be used to define per-domain
418 * variables. Each exclusive cpuset essentially defines an island domain by
419 * fully partitioning the member cpus from any other cpuset. Whenever a new
420 * exclusive cpuset is created, we also create and attach a new root-domain
428 cpumask_var_t online;
431 * The "RT overload" flag: it gets set if a CPU has more than
432 * one runnable RT task.
434 cpumask_var_t rto_mask;
436 struct cpupri cpupri;
440 * By default the system creates a single root-domain with all cpus as
441 * members (mimicking the global state we have today).
443 static struct root_domain def_root_domain;
445 #endif /* CONFIG_SMP */
448 * This is the main, per-CPU runqueue data structure.
450 * Locking rule: those places that want to lock multiple runqueues
451 * (such as the load balancing or the thread migration code), lock
452 * acquire operations must be ordered by ascending &runqueue.
459 * nr_running and cpu_load should be in the same cacheline because
460 * remote CPUs use both these fields when doing load calculation.
462 unsigned long nr_running;
463 #define CPU_LOAD_IDX_MAX 5
464 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
465 unsigned long last_load_update_tick;
468 unsigned char nohz_balance_kick;
470 int skip_clock_update;
472 /* capture load from *all* tasks on this cpu: */
473 struct load_weight load;
474 unsigned long nr_load_updates;
480 #ifdef CONFIG_FAIR_GROUP_SCHED
481 /* list of leaf cfs_rq on this cpu: */
482 struct list_head leaf_cfs_rq_list;
484 #ifdef CONFIG_RT_GROUP_SCHED
485 struct list_head leaf_rt_rq_list;
489 * This is part of a global counter where only the total sum
490 * over all CPUs matters. A task can increase this counter on
491 * one CPU and if it got migrated afterwards it may decrease
492 * it on another CPU. Always updated under the runqueue lock:
494 unsigned long nr_uninterruptible;
496 struct task_struct *curr, *idle, *stop;
497 unsigned long next_balance;
498 struct mm_struct *prev_mm;
506 struct root_domain *rd;
507 struct sched_domain *sd;
509 unsigned long cpu_power;
511 unsigned char idle_at_tick;
512 /* For active balancing */
516 struct cpu_stop_work active_balance_work;
517 /* cpu of this runqueue: */
521 unsigned long avg_load_per_task;
529 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
533 /* calc_load related fields */
534 unsigned long calc_load_update;
535 long calc_load_active;
537 #ifdef CONFIG_SCHED_HRTICK
539 int hrtick_csd_pending;
540 struct call_single_data hrtick_csd;
542 struct hrtimer hrtick_timer;
545 #ifdef CONFIG_SCHEDSTATS
547 struct sched_info rq_sched_info;
548 unsigned long long rq_cpu_time;
549 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
551 /* sys_sched_yield() stats */
552 unsigned int yld_count;
554 /* schedule() stats */
555 unsigned int sched_switch;
556 unsigned int sched_count;
557 unsigned int sched_goidle;
559 /* try_to_wake_up() stats */
560 unsigned int ttwu_count;
561 unsigned int ttwu_local;
565 struct task_struct *wake_list;
569 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
572 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
574 static inline int cpu_of(struct rq *rq)
583 #define rcu_dereference_check_sched_domain(p) \
584 rcu_dereference_check((p), \
585 rcu_read_lock_held() || \
586 lockdep_is_held(&sched_domains_mutex))
589 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
590 * See detach_destroy_domains: synchronize_sched for details.
592 * The domain tree of any CPU may only be accessed from within
593 * preempt-disabled sections.
595 #define for_each_domain(cpu, __sd) \
596 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
598 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
599 #define this_rq() (&__get_cpu_var(runqueues))
600 #define task_rq(p) cpu_rq(task_cpu(p))
601 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
602 #define raw_rq() (&__raw_get_cpu_var(runqueues))
604 #ifdef CONFIG_CGROUP_SCHED
607 * Return the group to which this tasks belongs.
609 * We use task_subsys_state_check() and extend the RCU verification with
610 * pi->lock and rq->lock because cpu_cgroup_attach() holds those locks for each
611 * task it moves into the cgroup. Therefore by holding either of those locks,
612 * we pin the task to the current cgroup.
614 static inline struct task_group *task_group(struct task_struct *p)
616 struct task_group *tg;
617 struct cgroup_subsys_state *css;
619 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
620 lockdep_is_held(&p->pi_lock) ||
621 lockdep_is_held(&task_rq(p)->lock));
622 tg = container_of(css, struct task_group, css);
624 return autogroup_task_group(p, tg);
627 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
628 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
630 #ifdef CONFIG_FAIR_GROUP_SCHED
631 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
632 p->se.parent = task_group(p)->se[cpu];
635 #ifdef CONFIG_RT_GROUP_SCHED
636 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
637 p->rt.parent = task_group(p)->rt_se[cpu];
641 #else /* CONFIG_CGROUP_SCHED */
643 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
644 static inline struct task_group *task_group(struct task_struct *p)
649 #endif /* CONFIG_CGROUP_SCHED */
651 static void update_rq_clock_task(struct rq *rq, s64 delta);
653 static void update_rq_clock(struct rq *rq)
657 if (rq->skip_clock_update > 0)
660 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
662 update_rq_clock_task(rq, delta);
666 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
668 #ifdef CONFIG_SCHED_DEBUG
669 # define const_debug __read_mostly
671 # define const_debug static const
675 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
676 * @cpu: the processor in question.
678 * This interface allows printk to be called with the runqueue lock
679 * held and know whether or not it is OK to wake up the klogd.
681 int runqueue_is_locked(int cpu)
683 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
687 * Debugging: various feature bits
690 #define SCHED_FEAT(name, enabled) \
691 __SCHED_FEAT_##name ,
694 #include "sched_features.h"
699 #define SCHED_FEAT(name, enabled) \
700 (1UL << __SCHED_FEAT_##name) * enabled |
702 const_debug unsigned int sysctl_sched_features =
703 #include "sched_features.h"
708 #ifdef CONFIG_SCHED_DEBUG
709 #define SCHED_FEAT(name, enabled) \
712 static __read_mostly char *sched_feat_names[] = {
713 #include "sched_features.h"
719 static int sched_feat_show(struct seq_file *m, void *v)
723 for (i = 0; sched_feat_names[i]; i++) {
724 if (!(sysctl_sched_features & (1UL << i)))
726 seq_printf(m, "%s ", sched_feat_names[i]);
734 sched_feat_write(struct file *filp, const char __user *ubuf,
735 size_t cnt, loff_t *ppos)
745 if (copy_from_user(&buf, ubuf, cnt))
751 if (strncmp(cmp, "NO_", 3) == 0) {
756 for (i = 0; sched_feat_names[i]; i++) {
757 if (strcmp(cmp, sched_feat_names[i]) == 0) {
759 sysctl_sched_features &= ~(1UL << i);
761 sysctl_sched_features |= (1UL << i);
766 if (!sched_feat_names[i])
774 static int sched_feat_open(struct inode *inode, struct file *filp)
776 return single_open(filp, sched_feat_show, NULL);
779 static const struct file_operations sched_feat_fops = {
780 .open = sched_feat_open,
781 .write = sched_feat_write,
784 .release = single_release,
787 static __init int sched_init_debug(void)
789 debugfs_create_file("sched_features", 0644, NULL, NULL,
794 late_initcall(sched_init_debug);
798 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
801 * Number of tasks to iterate in a single balance run.
802 * Limited because this is done with IRQs disabled.
804 const_debug unsigned int sysctl_sched_nr_migrate = 32;
807 * period over which we average the RT time consumption, measured
812 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
815 * period over which we measure -rt task cpu usage in us.
818 unsigned int sysctl_sched_rt_period = 1000000;
820 static __read_mostly int scheduler_running;
823 * part of the period that we allow rt tasks to run in us.
826 int sysctl_sched_rt_runtime = 950000;
828 static inline u64 global_rt_period(void)
830 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
833 static inline u64 global_rt_runtime(void)
835 if (sysctl_sched_rt_runtime < 0)
838 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
841 #ifndef prepare_arch_switch
842 # define prepare_arch_switch(next) do { } while (0)
844 #ifndef finish_arch_switch
845 # define finish_arch_switch(prev) do { } while (0)
848 static inline int task_current(struct rq *rq, struct task_struct *p)
850 return rq->curr == p;
853 static inline int task_running(struct rq *rq, struct task_struct *p)
858 return task_current(rq, p);
862 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
863 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
867 * We can optimise this out completely for !SMP, because the
868 * SMP rebalancing from interrupt is the only thing that cares
875 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
879 * After ->on_cpu is cleared, the task can be moved to a different CPU.
880 * We must ensure this doesn't happen until the switch is completely
886 #ifdef CONFIG_DEBUG_SPINLOCK
887 /* this is a valid case when another task releases the spinlock */
888 rq->lock.owner = current;
891 * If we are tracking spinlock dependencies then we have to
892 * fix up the runqueue lock - which gets 'carried over' from
895 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
897 raw_spin_unlock_irq(&rq->lock);
900 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
901 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
905 * We can optimise this out completely for !SMP, because the
906 * SMP rebalancing from interrupt is the only thing that cares
911 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
912 raw_spin_unlock_irq(&rq->lock);
914 raw_spin_unlock(&rq->lock);
918 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
922 * After ->on_cpu is cleared, the task can be moved to a different CPU.
923 * We must ensure this doesn't happen until the switch is completely
929 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
933 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
936 * __task_rq_lock - lock the rq @p resides on.
938 static inline struct rq *__task_rq_lock(struct task_struct *p)
943 lockdep_assert_held(&p->pi_lock);
947 raw_spin_lock(&rq->lock);
948 if (likely(rq == task_rq(p)))
950 raw_spin_unlock(&rq->lock);
955 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
957 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
958 __acquires(p->pi_lock)
964 raw_spin_lock_irqsave(&p->pi_lock, *flags);
966 raw_spin_lock(&rq->lock);
967 if (likely(rq == task_rq(p)))
969 raw_spin_unlock(&rq->lock);
970 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
974 static void __task_rq_unlock(struct rq *rq)
977 raw_spin_unlock(&rq->lock);
981 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
983 __releases(p->pi_lock)
985 raw_spin_unlock(&rq->lock);
986 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
990 * this_rq_lock - lock this runqueue and disable interrupts.
992 static struct rq *this_rq_lock(void)
999 raw_spin_lock(&rq->lock);
1004 #ifdef CONFIG_SCHED_HRTICK
1006 * Use HR-timers to deliver accurate preemption points.
1008 * Its all a bit involved since we cannot program an hrt while holding the
1009 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1012 * When we get rescheduled we reprogram the hrtick_timer outside of the
1018 * - enabled by features
1019 * - hrtimer is actually high res
1021 static inline int hrtick_enabled(struct rq *rq)
1023 if (!sched_feat(HRTICK))
1025 if (!cpu_active(cpu_of(rq)))
1027 return hrtimer_is_hres_active(&rq->hrtick_timer);
1030 static void hrtick_clear(struct rq *rq)
1032 if (hrtimer_active(&rq->hrtick_timer))
1033 hrtimer_cancel(&rq->hrtick_timer);
1037 * High-resolution timer tick.
1038 * Runs from hardirq context with interrupts disabled.
1040 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1042 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1044 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1046 raw_spin_lock(&rq->lock);
1047 update_rq_clock(rq);
1048 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1049 raw_spin_unlock(&rq->lock);
1051 return HRTIMER_NORESTART;
1056 * called from hardirq (IPI) context
1058 static void __hrtick_start(void *arg)
1060 struct rq *rq = arg;
1062 raw_spin_lock(&rq->lock);
1063 hrtimer_restart(&rq->hrtick_timer);
1064 rq->hrtick_csd_pending = 0;
1065 raw_spin_unlock(&rq->lock);
1069 * Called to set the hrtick timer state.
1071 * called with rq->lock held and irqs disabled
1073 static void hrtick_start(struct rq *rq, u64 delay)
1075 struct hrtimer *timer = &rq->hrtick_timer;
1076 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1078 hrtimer_set_expires(timer, time);
1080 if (rq == this_rq()) {
1081 hrtimer_restart(timer);
1082 } else if (!rq->hrtick_csd_pending) {
1083 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1084 rq->hrtick_csd_pending = 1;
1089 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1091 int cpu = (int)(long)hcpu;
1094 case CPU_UP_CANCELED:
1095 case CPU_UP_CANCELED_FROZEN:
1096 case CPU_DOWN_PREPARE:
1097 case CPU_DOWN_PREPARE_FROZEN:
1099 case CPU_DEAD_FROZEN:
1100 hrtick_clear(cpu_rq(cpu));
1107 static __init void init_hrtick(void)
1109 hotcpu_notifier(hotplug_hrtick, 0);
1113 * Called to set the hrtick timer state.
1115 * called with rq->lock held and irqs disabled
1117 static void hrtick_start(struct rq *rq, u64 delay)
1119 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1120 HRTIMER_MODE_REL_PINNED, 0);
1123 static inline void init_hrtick(void)
1126 #endif /* CONFIG_SMP */
1128 static void init_rq_hrtick(struct rq *rq)
1131 rq->hrtick_csd_pending = 0;
1133 rq->hrtick_csd.flags = 0;
1134 rq->hrtick_csd.func = __hrtick_start;
1135 rq->hrtick_csd.info = rq;
1138 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1139 rq->hrtick_timer.function = hrtick;
1141 #else /* CONFIG_SCHED_HRTICK */
1142 static inline void hrtick_clear(struct rq *rq)
1146 static inline void init_rq_hrtick(struct rq *rq)
1150 static inline void init_hrtick(void)
1153 #endif /* CONFIG_SCHED_HRTICK */
1156 * resched_task - mark a task 'to be rescheduled now'.
1158 * On UP this means the setting of the need_resched flag, on SMP it
1159 * might also involve a cross-CPU call to trigger the scheduler on
1164 #ifndef tsk_is_polling
1165 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1168 static void resched_task(struct task_struct *p)
1172 assert_raw_spin_locked(&task_rq(p)->lock);
1174 if (test_tsk_need_resched(p))
1177 set_tsk_need_resched(p);
1180 if (cpu == smp_processor_id())
1183 /* NEED_RESCHED must be visible before we test polling */
1185 if (!tsk_is_polling(p))
1186 smp_send_reschedule(cpu);
1189 static void resched_cpu(int cpu)
1191 struct rq *rq = cpu_rq(cpu);
1192 unsigned long flags;
1194 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1196 resched_task(cpu_curr(cpu));
1197 raw_spin_unlock_irqrestore(&rq->lock, flags);
1202 * In the semi idle case, use the nearest busy cpu for migrating timers
1203 * from an idle cpu. This is good for power-savings.
1205 * We don't do similar optimization for completely idle system, as
1206 * selecting an idle cpu will add more delays to the timers than intended
1207 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1209 int get_nohz_timer_target(void)
1211 int cpu = smp_processor_id();
1213 struct sched_domain *sd;
1216 for_each_domain(cpu, sd) {
1217 for_each_cpu(i, sched_domain_span(sd)) {
1229 * When add_timer_on() enqueues a timer into the timer wheel of an
1230 * idle CPU then this timer might expire before the next timer event
1231 * which is scheduled to wake up that CPU. In case of a completely
1232 * idle system the next event might even be infinite time into the
1233 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1234 * leaves the inner idle loop so the newly added timer is taken into
1235 * account when the CPU goes back to idle and evaluates the timer
1236 * wheel for the next timer event.
1238 void wake_up_idle_cpu(int cpu)
1240 struct rq *rq = cpu_rq(cpu);
1242 if (cpu == smp_processor_id())
1246 * This is safe, as this function is called with the timer
1247 * wheel base lock of (cpu) held. When the CPU is on the way
1248 * to idle and has not yet set rq->curr to idle then it will
1249 * be serialized on the timer wheel base lock and take the new
1250 * timer into account automatically.
1252 if (rq->curr != rq->idle)
1256 * We can set TIF_RESCHED on the idle task of the other CPU
1257 * lockless. The worst case is that the other CPU runs the
1258 * idle task through an additional NOOP schedule()
1260 set_tsk_need_resched(rq->idle);
1262 /* NEED_RESCHED must be visible before we test polling */
1264 if (!tsk_is_polling(rq->idle))
1265 smp_send_reschedule(cpu);
1268 #endif /* CONFIG_NO_HZ */
1270 static u64 sched_avg_period(void)
1272 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1275 static void sched_avg_update(struct rq *rq)
1277 s64 period = sched_avg_period();
1279 while ((s64)(rq->clock - rq->age_stamp) > period) {
1281 * Inline assembly required to prevent the compiler
1282 * optimising this loop into a divmod call.
1283 * See __iter_div_u64_rem() for another example of this.
1285 asm("" : "+rm" (rq->age_stamp));
1286 rq->age_stamp += period;
1291 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1293 rq->rt_avg += rt_delta;
1294 sched_avg_update(rq);
1297 #else /* !CONFIG_SMP */
1298 static void resched_task(struct task_struct *p)
1300 assert_raw_spin_locked(&task_rq(p)->lock);
1301 set_tsk_need_resched(p);
1304 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1308 static void sched_avg_update(struct rq *rq)
1311 #endif /* CONFIG_SMP */
1313 #if BITS_PER_LONG == 32
1314 # define WMULT_CONST (~0UL)
1316 # define WMULT_CONST (1UL << 32)
1319 #define WMULT_SHIFT 32
1322 * Shift right and round:
1324 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1327 * delta *= weight / lw
1329 static unsigned long
1330 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1331 struct load_weight *lw)
1336 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1337 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1338 * 2^SCHED_LOAD_RESOLUTION.
1340 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
1341 tmp = (u64)delta_exec * scale_load_down(weight);
1343 tmp = (u64)delta_exec;
1345 if (!lw->inv_weight) {
1346 unsigned long w = scale_load_down(lw->weight);
1348 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
1350 else if (unlikely(!w))
1351 lw->inv_weight = WMULT_CONST;
1353 lw->inv_weight = WMULT_CONST / w;
1357 * Check whether we'd overflow the 64-bit multiplication:
1359 if (unlikely(tmp > WMULT_CONST))
1360 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1363 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1365 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1368 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1374 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1380 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1387 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1388 * of tasks with abnormal "nice" values across CPUs the contribution that
1389 * each task makes to its run queue's load is weighted according to its
1390 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1391 * scaled version of the new time slice allocation that they receive on time
1395 #define WEIGHT_IDLEPRIO 3
1396 #define WMULT_IDLEPRIO 1431655765
1399 * Nice levels are multiplicative, with a gentle 10% change for every
1400 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1401 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1402 * that remained on nice 0.
1404 * The "10% effect" is relative and cumulative: from _any_ nice level,
1405 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1406 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1407 * If a task goes up by ~10% and another task goes down by ~10% then
1408 * the relative distance between them is ~25%.)
1410 static const int prio_to_weight[40] = {
1411 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1412 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1413 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1414 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1415 /* 0 */ 1024, 820, 655, 526, 423,
1416 /* 5 */ 335, 272, 215, 172, 137,
1417 /* 10 */ 110, 87, 70, 56, 45,
1418 /* 15 */ 36, 29, 23, 18, 15,
1422 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1424 * In cases where the weight does not change often, we can use the
1425 * precalculated inverse to speed up arithmetics by turning divisions
1426 * into multiplications:
1428 static const u32 prio_to_wmult[40] = {
1429 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1430 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1431 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1432 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1433 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1434 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1435 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1436 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1439 /* Time spent by the tasks of the cpu accounting group executing in ... */
1440 enum cpuacct_stat_index {
1441 CPUACCT_STAT_USER, /* ... user mode */
1442 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1444 CPUACCT_STAT_NSTATS,
1447 #ifdef CONFIG_CGROUP_CPUACCT
1448 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1449 static void cpuacct_update_stats(struct task_struct *tsk,
1450 enum cpuacct_stat_index idx, cputime_t val);
1452 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1453 static inline void cpuacct_update_stats(struct task_struct *tsk,
1454 enum cpuacct_stat_index idx, cputime_t val) {}
1457 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1459 update_load_add(&rq->load, load);
1462 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1464 update_load_sub(&rq->load, load);
1467 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1468 typedef int (*tg_visitor)(struct task_group *, void *);
1471 * Iterate the full tree, calling @down when first entering a node and @up when
1472 * leaving it for the final time.
1474 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1476 struct task_group *parent, *child;
1480 parent = &root_task_group;
1482 ret = (*down)(parent, data);
1485 list_for_each_entry_rcu(child, &parent->children, siblings) {
1492 ret = (*up)(parent, data);
1497 parent = parent->parent;
1506 static int tg_nop(struct task_group *tg, void *data)
1513 /* Used instead of source_load when we know the type == 0 */
1514 static unsigned long weighted_cpuload(const int cpu)
1516 return cpu_rq(cpu)->load.weight;
1520 * Return a low guess at the load of a migration-source cpu weighted
1521 * according to the scheduling class and "nice" value.
1523 * We want to under-estimate the load of migration sources, to
1524 * balance conservatively.
1526 static unsigned long source_load(int cpu, int type)
1528 struct rq *rq = cpu_rq(cpu);
1529 unsigned long total = weighted_cpuload(cpu);
1531 if (type == 0 || !sched_feat(LB_BIAS))
1534 return min(rq->cpu_load[type-1], total);
1538 * Return a high guess at the load of a migration-target cpu weighted
1539 * according to the scheduling class and "nice" value.
1541 static unsigned long target_load(int cpu, int type)
1543 struct rq *rq = cpu_rq(cpu);
1544 unsigned long total = weighted_cpuload(cpu);
1546 if (type == 0 || !sched_feat(LB_BIAS))
1549 return max(rq->cpu_load[type-1], total);
1552 static unsigned long power_of(int cpu)
1554 return cpu_rq(cpu)->cpu_power;
1557 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1559 static unsigned long cpu_avg_load_per_task(int cpu)
1561 struct rq *rq = cpu_rq(cpu);
1562 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1565 rq->avg_load_per_task = rq->load.weight / nr_running;
1567 rq->avg_load_per_task = 0;
1569 return rq->avg_load_per_task;
1572 #ifdef CONFIG_FAIR_GROUP_SCHED
1575 * Compute the cpu's hierarchical load factor for each task group.
1576 * This needs to be done in a top-down fashion because the load of a child
1577 * group is a fraction of its parents load.
1579 static int tg_load_down(struct task_group *tg, void *data)
1582 long cpu = (long)data;
1585 load = cpu_rq(cpu)->load.weight;
1587 load = tg->parent->cfs_rq[cpu]->h_load;
1588 load *= tg->se[cpu]->load.weight;
1589 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1592 tg->cfs_rq[cpu]->h_load = load;
1597 static void update_h_load(long cpu)
1599 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1604 #ifdef CONFIG_PREEMPT
1606 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1609 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1610 * way at the expense of forcing extra atomic operations in all
1611 * invocations. This assures that the double_lock is acquired using the
1612 * same underlying policy as the spinlock_t on this architecture, which
1613 * reduces latency compared to the unfair variant below. However, it
1614 * also adds more overhead and therefore may reduce throughput.
1616 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1617 __releases(this_rq->lock)
1618 __acquires(busiest->lock)
1619 __acquires(this_rq->lock)
1621 raw_spin_unlock(&this_rq->lock);
1622 double_rq_lock(this_rq, busiest);
1629 * Unfair double_lock_balance: Optimizes throughput at the expense of
1630 * latency by eliminating extra atomic operations when the locks are
1631 * already in proper order on entry. This favors lower cpu-ids and will
1632 * grant the double lock to lower cpus over higher ids under contention,
1633 * regardless of entry order into the function.
1635 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1636 __releases(this_rq->lock)
1637 __acquires(busiest->lock)
1638 __acquires(this_rq->lock)
1642 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1643 if (busiest < this_rq) {
1644 raw_spin_unlock(&this_rq->lock);
1645 raw_spin_lock(&busiest->lock);
1646 raw_spin_lock_nested(&this_rq->lock,
1647 SINGLE_DEPTH_NESTING);
1650 raw_spin_lock_nested(&busiest->lock,
1651 SINGLE_DEPTH_NESTING);
1656 #endif /* CONFIG_PREEMPT */
1659 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1661 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1663 if (unlikely(!irqs_disabled())) {
1664 /* printk() doesn't work good under rq->lock */
1665 raw_spin_unlock(&this_rq->lock);
1669 return _double_lock_balance(this_rq, busiest);
1672 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1673 __releases(busiest->lock)
1675 raw_spin_unlock(&busiest->lock);
1676 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1680 * double_rq_lock - safely lock two runqueues
1682 * Note this does not disable interrupts like task_rq_lock,
1683 * you need to do so manually before calling.
1685 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1686 __acquires(rq1->lock)
1687 __acquires(rq2->lock)
1689 BUG_ON(!irqs_disabled());
1691 raw_spin_lock(&rq1->lock);
1692 __acquire(rq2->lock); /* Fake it out ;) */
1695 raw_spin_lock(&rq1->lock);
1696 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1698 raw_spin_lock(&rq2->lock);
1699 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1705 * double_rq_unlock - safely unlock two runqueues
1707 * Note this does not restore interrupts like task_rq_unlock,
1708 * you need to do so manually after calling.
1710 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1711 __releases(rq1->lock)
1712 __releases(rq2->lock)
1714 raw_spin_unlock(&rq1->lock);
1716 raw_spin_unlock(&rq2->lock);
1718 __release(rq2->lock);
1721 #else /* CONFIG_SMP */
1724 * double_rq_lock - safely lock two runqueues
1726 * Note this does not disable interrupts like task_rq_lock,
1727 * you need to do so manually before calling.
1729 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1730 __acquires(rq1->lock)
1731 __acquires(rq2->lock)
1733 BUG_ON(!irqs_disabled());
1735 raw_spin_lock(&rq1->lock);
1736 __acquire(rq2->lock); /* Fake it out ;) */
1740 * double_rq_unlock - safely unlock two runqueues
1742 * Note this does not restore interrupts like task_rq_unlock,
1743 * you need to do so manually after calling.
1745 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1746 __releases(rq1->lock)
1747 __releases(rq2->lock)
1750 raw_spin_unlock(&rq1->lock);
1751 __release(rq2->lock);
1756 static void calc_load_account_idle(struct rq *this_rq);
1757 static void update_sysctl(void);
1758 static int get_update_sysctl_factor(void);
1759 static void update_cpu_load(struct rq *this_rq);
1761 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1763 set_task_rq(p, cpu);
1766 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1767 * successfuly executed on another CPU. We must ensure that updates of
1768 * per-task data have been completed by this moment.
1771 task_thread_info(p)->cpu = cpu;
1775 static const struct sched_class rt_sched_class;
1777 #define sched_class_highest (&stop_sched_class)
1778 #define for_each_class(class) \
1779 for (class = sched_class_highest; class; class = class->next)
1781 #include "sched_stats.h"
1783 static void inc_nr_running(struct rq *rq)
1788 static void dec_nr_running(struct rq *rq)
1793 static void set_load_weight(struct task_struct *p)
1795 int prio = p->static_prio - MAX_RT_PRIO;
1796 struct load_weight *load = &p->se.load;
1799 * SCHED_IDLE tasks get minimal weight:
1801 if (p->policy == SCHED_IDLE) {
1802 load->weight = scale_load(WEIGHT_IDLEPRIO);
1803 load->inv_weight = WMULT_IDLEPRIO;
1807 load->weight = scale_load(prio_to_weight[prio]);
1808 load->inv_weight = prio_to_wmult[prio];
1811 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1813 update_rq_clock(rq);
1814 sched_info_queued(p);
1815 p->sched_class->enqueue_task(rq, p, flags);
1818 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1820 update_rq_clock(rq);
1821 sched_info_dequeued(p);
1822 p->sched_class->dequeue_task(rq, p, flags);
1826 * activate_task - move a task to the runqueue.
1828 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1830 if (task_contributes_to_load(p))
1831 rq->nr_uninterruptible--;
1833 enqueue_task(rq, p, flags);
1838 * deactivate_task - remove a task from the runqueue.
1840 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1842 if (task_contributes_to_load(p))
1843 rq->nr_uninterruptible++;
1845 dequeue_task(rq, p, flags);
1849 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1852 * There are no locks covering percpu hardirq/softirq time.
1853 * They are only modified in account_system_vtime, on corresponding CPU
1854 * with interrupts disabled. So, writes are safe.
1855 * They are read and saved off onto struct rq in update_rq_clock().
1856 * This may result in other CPU reading this CPU's irq time and can
1857 * race with irq/account_system_vtime on this CPU. We would either get old
1858 * or new value with a side effect of accounting a slice of irq time to wrong
1859 * task when irq is in progress while we read rq->clock. That is a worthy
1860 * compromise in place of having locks on each irq in account_system_time.
1862 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1863 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1865 static DEFINE_PER_CPU(u64, irq_start_time);
1866 static int sched_clock_irqtime;
1868 void enable_sched_clock_irqtime(void)
1870 sched_clock_irqtime = 1;
1873 void disable_sched_clock_irqtime(void)
1875 sched_clock_irqtime = 0;
1878 #ifndef CONFIG_64BIT
1879 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1881 static inline void irq_time_write_begin(void)
1883 __this_cpu_inc(irq_time_seq.sequence);
1887 static inline void irq_time_write_end(void)
1890 __this_cpu_inc(irq_time_seq.sequence);
1893 static inline u64 irq_time_read(int cpu)
1899 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1900 irq_time = per_cpu(cpu_softirq_time, cpu) +
1901 per_cpu(cpu_hardirq_time, cpu);
1902 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1906 #else /* CONFIG_64BIT */
1907 static inline void irq_time_write_begin(void)
1911 static inline void irq_time_write_end(void)
1915 static inline u64 irq_time_read(int cpu)
1917 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1919 #endif /* CONFIG_64BIT */
1922 * Called before incrementing preempt_count on {soft,}irq_enter
1923 * and before decrementing preempt_count on {soft,}irq_exit.
1925 void account_system_vtime(struct task_struct *curr)
1927 unsigned long flags;
1931 if (!sched_clock_irqtime)
1934 local_irq_save(flags);
1936 cpu = smp_processor_id();
1937 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1938 __this_cpu_add(irq_start_time, delta);
1940 irq_time_write_begin();
1942 * We do not account for softirq time from ksoftirqd here.
1943 * We want to continue accounting softirq time to ksoftirqd thread
1944 * in that case, so as not to confuse scheduler with a special task
1945 * that do not consume any time, but still wants to run.
1947 if (hardirq_count())
1948 __this_cpu_add(cpu_hardirq_time, delta);
1949 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
1950 __this_cpu_add(cpu_softirq_time, delta);
1952 irq_time_write_end();
1953 local_irq_restore(flags);
1955 EXPORT_SYMBOL_GPL(account_system_vtime);
1957 static void update_rq_clock_task(struct rq *rq, s64 delta)
1961 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
1964 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1965 * this case when a previous update_rq_clock() happened inside a
1966 * {soft,}irq region.
1968 * When this happens, we stop ->clock_task and only update the
1969 * prev_irq_time stamp to account for the part that fit, so that a next
1970 * update will consume the rest. This ensures ->clock_task is
1973 * It does however cause some slight miss-attribution of {soft,}irq
1974 * time, a more accurate solution would be to update the irq_time using
1975 * the current rq->clock timestamp, except that would require using
1978 if (irq_delta > delta)
1981 rq->prev_irq_time += irq_delta;
1983 rq->clock_task += delta;
1985 if (irq_delta && sched_feat(NONIRQ_POWER))
1986 sched_rt_avg_update(rq, irq_delta);
1989 static int irqtime_account_hi_update(void)
1991 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
1992 unsigned long flags;
1996 local_irq_save(flags);
1997 latest_ns = this_cpu_read(cpu_hardirq_time);
1998 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
2000 local_irq_restore(flags);
2004 static int irqtime_account_si_update(void)
2006 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2007 unsigned long flags;
2011 local_irq_save(flags);
2012 latest_ns = this_cpu_read(cpu_softirq_time);
2013 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2015 local_irq_restore(flags);
2019 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2021 #define sched_clock_irqtime (0)
2023 static void update_rq_clock_task(struct rq *rq, s64 delta)
2025 rq->clock_task += delta;
2028 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2030 #include "sched_idletask.c"
2031 #include "sched_fair.c"
2032 #include "sched_rt.c"
2033 #include "sched_autogroup.c"
2034 #include "sched_stoptask.c"
2035 #ifdef CONFIG_SCHED_DEBUG
2036 # include "sched_debug.c"
2039 void sched_set_stop_task(int cpu, struct task_struct *stop)
2041 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2042 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2046 * Make it appear like a SCHED_FIFO task, its something
2047 * userspace knows about and won't get confused about.
2049 * Also, it will make PI more or less work without too
2050 * much confusion -- but then, stop work should not
2051 * rely on PI working anyway.
2053 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2055 stop->sched_class = &stop_sched_class;
2058 cpu_rq(cpu)->stop = stop;
2062 * Reset it back to a normal scheduling class so that
2063 * it can die in pieces.
2065 old_stop->sched_class = &rt_sched_class;
2070 * __normal_prio - return the priority that is based on the static prio
2072 static inline int __normal_prio(struct task_struct *p)
2074 return p->static_prio;
2078 * Calculate the expected normal priority: i.e. priority
2079 * without taking RT-inheritance into account. Might be
2080 * boosted by interactivity modifiers. Changes upon fork,
2081 * setprio syscalls, and whenever the interactivity
2082 * estimator recalculates.
2084 static inline int normal_prio(struct task_struct *p)
2088 if (task_has_rt_policy(p))
2089 prio = MAX_RT_PRIO-1 - p->rt_priority;
2091 prio = __normal_prio(p);
2096 * Calculate the current priority, i.e. the priority
2097 * taken into account by the scheduler. This value might
2098 * be boosted by RT tasks, or might be boosted by
2099 * interactivity modifiers. Will be RT if the task got
2100 * RT-boosted. If not then it returns p->normal_prio.
2102 static int effective_prio(struct task_struct *p)
2104 p->normal_prio = normal_prio(p);
2106 * If we are RT tasks or we were boosted to RT priority,
2107 * keep the priority unchanged. Otherwise, update priority
2108 * to the normal priority:
2110 if (!rt_prio(p->prio))
2111 return p->normal_prio;
2116 * task_curr - is this task currently executing on a CPU?
2117 * @p: the task in question.
2119 inline int task_curr(const struct task_struct *p)
2121 return cpu_curr(task_cpu(p)) == p;
2124 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2125 const struct sched_class *prev_class,
2128 if (prev_class != p->sched_class) {
2129 if (prev_class->switched_from)
2130 prev_class->switched_from(rq, p);
2131 p->sched_class->switched_to(rq, p);
2132 } else if (oldprio != p->prio)
2133 p->sched_class->prio_changed(rq, p, oldprio);
2136 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2138 const struct sched_class *class;
2140 if (p->sched_class == rq->curr->sched_class) {
2141 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2143 for_each_class(class) {
2144 if (class == rq->curr->sched_class)
2146 if (class == p->sched_class) {
2147 resched_task(rq->curr);
2154 * A queue event has occurred, and we're going to schedule. In
2155 * this case, we can save a useless back to back clock update.
2157 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2158 rq->skip_clock_update = 1;
2163 * Is this task likely cache-hot:
2166 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2170 if (p->sched_class != &fair_sched_class)
2173 if (unlikely(p->policy == SCHED_IDLE))
2177 * Buddy candidates are cache hot:
2179 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2180 (&p->se == cfs_rq_of(&p->se)->next ||
2181 &p->se == cfs_rq_of(&p->se)->last))
2184 if (sysctl_sched_migration_cost == -1)
2186 if (sysctl_sched_migration_cost == 0)
2189 delta = now - p->se.exec_start;
2191 return delta < (s64)sysctl_sched_migration_cost;
2194 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2196 #ifdef CONFIG_SCHED_DEBUG
2198 * We should never call set_task_cpu() on a blocked task,
2199 * ttwu() will sort out the placement.
2201 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2202 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2204 #ifdef CONFIG_LOCKDEP
2206 * The caller should hold either p->pi_lock or rq->lock, when changing
2207 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2209 * sched_move_task() holds both and thus holding either pins the cgroup,
2210 * see set_task_rq().
2212 * Furthermore, all task_rq users should acquire both locks, see
2215 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2216 lockdep_is_held(&task_rq(p)->lock)));
2220 trace_sched_migrate_task(p, new_cpu);
2222 if (task_cpu(p) != new_cpu) {
2223 p->se.nr_migrations++;
2224 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2227 __set_task_cpu(p, new_cpu);
2230 struct migration_arg {
2231 struct task_struct *task;
2235 static int migration_cpu_stop(void *data);
2238 * wait_task_inactive - wait for a thread to unschedule.
2240 * If @match_state is nonzero, it's the @p->state value just checked and
2241 * not expected to change. If it changes, i.e. @p might have woken up,
2242 * then return zero. When we succeed in waiting for @p to be off its CPU,
2243 * we return a positive number (its total switch count). If a second call
2244 * a short while later returns the same number, the caller can be sure that
2245 * @p has remained unscheduled the whole time.
2247 * The caller must ensure that the task *will* unschedule sometime soon,
2248 * else this function might spin for a *long* time. This function can't
2249 * be called with interrupts off, or it may introduce deadlock with
2250 * smp_call_function() if an IPI is sent by the same process we are
2251 * waiting to become inactive.
2253 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2255 unsigned long flags;
2262 * We do the initial early heuristics without holding
2263 * any task-queue locks at all. We'll only try to get
2264 * the runqueue lock when things look like they will
2270 * If the task is actively running on another CPU
2271 * still, just relax and busy-wait without holding
2274 * NOTE! Since we don't hold any locks, it's not
2275 * even sure that "rq" stays as the right runqueue!
2276 * But we don't care, since "task_running()" will
2277 * return false if the runqueue has changed and p
2278 * is actually now running somewhere else!
2280 while (task_running(rq, p)) {
2281 if (match_state && unlikely(p->state != match_state))
2287 * Ok, time to look more closely! We need the rq
2288 * lock now, to be *sure*. If we're wrong, we'll
2289 * just go back and repeat.
2291 rq = task_rq_lock(p, &flags);
2292 trace_sched_wait_task(p);
2293 running = task_running(rq, p);
2296 if (!match_state || p->state == match_state)
2297 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2298 task_rq_unlock(rq, p, &flags);
2301 * If it changed from the expected state, bail out now.
2303 if (unlikely(!ncsw))
2307 * Was it really running after all now that we
2308 * checked with the proper locks actually held?
2310 * Oops. Go back and try again..
2312 if (unlikely(running)) {
2318 * It's not enough that it's not actively running,
2319 * it must be off the runqueue _entirely_, and not
2322 * So if it was still runnable (but just not actively
2323 * running right now), it's preempted, and we should
2324 * yield - it could be a while.
2326 if (unlikely(on_rq)) {
2327 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2329 set_current_state(TASK_UNINTERRUPTIBLE);
2330 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2335 * Ahh, all good. It wasn't running, and it wasn't
2336 * runnable, which means that it will never become
2337 * running in the future either. We're all done!
2346 * kick_process - kick a running thread to enter/exit the kernel
2347 * @p: the to-be-kicked thread
2349 * Cause a process which is running on another CPU to enter
2350 * kernel-mode, without any delay. (to get signals handled.)
2352 * NOTE: this function doesn't have to take the runqueue lock,
2353 * because all it wants to ensure is that the remote task enters
2354 * the kernel. If the IPI races and the task has been migrated
2355 * to another CPU then no harm is done and the purpose has been
2358 void kick_process(struct task_struct *p)
2364 if ((cpu != smp_processor_id()) && task_curr(p))
2365 smp_send_reschedule(cpu);
2368 EXPORT_SYMBOL_GPL(kick_process);
2369 #endif /* CONFIG_SMP */
2373 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2375 static int select_fallback_rq(int cpu, struct task_struct *p)
2378 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2380 /* Look for allowed, online CPU in same node. */
2381 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2382 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2385 /* Any allowed, online CPU? */
2386 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2387 if (dest_cpu < nr_cpu_ids)
2390 /* No more Mr. Nice Guy. */
2391 dest_cpu = cpuset_cpus_allowed_fallback(p);
2393 * Don't tell them about moving exiting tasks or
2394 * kernel threads (both mm NULL), since they never
2397 if (p->mm && printk_ratelimit()) {
2398 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2399 task_pid_nr(p), p->comm, cpu);
2406 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2409 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2411 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2414 * In order not to call set_task_cpu() on a blocking task we need
2415 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2418 * Since this is common to all placement strategies, this lives here.
2420 * [ this allows ->select_task() to simply return task_cpu(p) and
2421 * not worry about this generic constraint ]
2423 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2425 cpu = select_fallback_rq(task_cpu(p), p);
2430 static void update_avg(u64 *avg, u64 sample)
2432 s64 diff = sample - *avg;
2438 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2440 #ifdef CONFIG_SCHEDSTATS
2441 struct rq *rq = this_rq();
2444 int this_cpu = smp_processor_id();
2446 if (cpu == this_cpu) {
2447 schedstat_inc(rq, ttwu_local);
2448 schedstat_inc(p, se.statistics.nr_wakeups_local);
2450 struct sched_domain *sd;
2452 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2454 for_each_domain(this_cpu, sd) {
2455 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2456 schedstat_inc(sd, ttwu_wake_remote);
2463 if (wake_flags & WF_MIGRATED)
2464 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2466 #endif /* CONFIG_SMP */
2468 schedstat_inc(rq, ttwu_count);
2469 schedstat_inc(p, se.statistics.nr_wakeups);
2471 if (wake_flags & WF_SYNC)
2472 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2474 #endif /* CONFIG_SCHEDSTATS */
2477 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2479 activate_task(rq, p, en_flags);
2482 /* if a worker is waking up, notify workqueue */
2483 if (p->flags & PF_WQ_WORKER)
2484 wq_worker_waking_up(p, cpu_of(rq));
2488 * Mark the task runnable and perform wakeup-preemption.
2491 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2493 trace_sched_wakeup(p, true);
2494 check_preempt_curr(rq, p, wake_flags);
2496 p->state = TASK_RUNNING;
2498 if (p->sched_class->task_woken)
2499 p->sched_class->task_woken(rq, p);
2501 if (unlikely(rq->idle_stamp)) {
2502 u64 delta = rq->clock - rq->idle_stamp;
2503 u64 max = 2*sysctl_sched_migration_cost;
2508 update_avg(&rq->avg_idle, delta);
2515 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2518 if (p->sched_contributes_to_load)
2519 rq->nr_uninterruptible--;
2522 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2523 ttwu_do_wakeup(rq, p, wake_flags);
2527 * Called in case the task @p isn't fully descheduled from its runqueue,
2528 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2529 * since all we need to do is flip p->state to TASK_RUNNING, since
2530 * the task is still ->on_rq.
2532 static int ttwu_remote(struct task_struct *p, int wake_flags)
2537 rq = __task_rq_lock(p);
2539 ttwu_do_wakeup(rq, p, wake_flags);
2542 __task_rq_unlock(rq);
2548 static void sched_ttwu_do_pending(struct task_struct *list)
2550 struct rq *rq = this_rq();
2552 raw_spin_lock(&rq->lock);
2555 struct task_struct *p = list;
2556 list = list->wake_entry;
2557 ttwu_do_activate(rq, p, 0);
2560 raw_spin_unlock(&rq->lock);
2563 #ifdef CONFIG_HOTPLUG_CPU
2565 static void sched_ttwu_pending(void)
2567 struct rq *rq = this_rq();
2568 struct task_struct *list = xchg(&rq->wake_list, NULL);
2573 sched_ttwu_do_pending(list);
2576 #endif /* CONFIG_HOTPLUG_CPU */
2578 void scheduler_ipi(void)
2580 struct rq *rq = this_rq();
2581 struct task_struct *list = xchg(&rq->wake_list, NULL);
2587 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2588 * traditionally all their work was done from the interrupt return
2589 * path. Now that we actually do some work, we need to make sure
2592 * Some archs already do call them, luckily irq_enter/exit nest
2595 * Arguably we should visit all archs and update all handlers,
2596 * however a fair share of IPIs are still resched only so this would
2597 * somewhat pessimize the simple resched case.
2600 sched_ttwu_do_pending(list);
2604 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2606 struct rq *rq = cpu_rq(cpu);
2607 struct task_struct *next = rq->wake_list;
2610 struct task_struct *old = next;
2612 p->wake_entry = next;
2613 next = cmpxchg(&rq->wake_list, old, p);
2619 smp_send_reschedule(cpu);
2622 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2623 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2628 rq = __task_rq_lock(p);
2630 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2631 ttwu_do_wakeup(rq, p, wake_flags);
2634 __task_rq_unlock(rq);
2639 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2640 #endif /* CONFIG_SMP */
2642 static void ttwu_queue(struct task_struct *p, int cpu)
2644 struct rq *rq = cpu_rq(cpu);
2646 #if defined(CONFIG_SMP)
2647 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2648 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2649 ttwu_queue_remote(p, cpu);
2654 raw_spin_lock(&rq->lock);
2655 ttwu_do_activate(rq, p, 0);
2656 raw_spin_unlock(&rq->lock);
2660 * try_to_wake_up - wake up a thread
2661 * @p: the thread to be awakened
2662 * @state: the mask of task states that can be woken
2663 * @wake_flags: wake modifier flags (WF_*)
2665 * Put it on the run-queue if it's not already there. The "current"
2666 * thread is always on the run-queue (except when the actual
2667 * re-schedule is in progress), and as such you're allowed to do
2668 * the simpler "current->state = TASK_RUNNING" to mark yourself
2669 * runnable without the overhead of this.
2671 * Returns %true if @p was woken up, %false if it was already running
2672 * or @state didn't match @p's state.
2675 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2677 unsigned long flags;
2678 int cpu, success = 0;
2681 raw_spin_lock_irqsave(&p->pi_lock, flags);
2682 if (!(p->state & state))
2685 success = 1; /* we're going to change ->state */
2688 if (p->on_rq && ttwu_remote(p, wake_flags))
2693 * If the owning (remote) cpu is still in the middle of schedule() with
2694 * this task as prev, wait until its done referencing the task.
2697 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2699 * In case the architecture enables interrupts in
2700 * context_switch(), we cannot busy wait, since that
2701 * would lead to deadlocks when an interrupt hits and
2702 * tries to wake up @prev. So bail and do a complete
2705 if (ttwu_activate_remote(p, wake_flags))
2712 * Pairs with the smp_wmb() in finish_lock_switch().
2716 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2717 p->state = TASK_WAKING;
2719 if (p->sched_class->task_waking)
2720 p->sched_class->task_waking(p);
2722 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2723 if (task_cpu(p) != cpu) {
2724 wake_flags |= WF_MIGRATED;
2725 set_task_cpu(p, cpu);
2727 #endif /* CONFIG_SMP */
2731 ttwu_stat(p, cpu, wake_flags);
2733 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2739 * try_to_wake_up_local - try to wake up a local task with rq lock held
2740 * @p: the thread to be awakened
2742 * Put @p on the run-queue if it's not already there. The caller must
2743 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2746 static void try_to_wake_up_local(struct task_struct *p)
2748 struct rq *rq = task_rq(p);
2750 BUG_ON(rq != this_rq());
2751 BUG_ON(p == current);
2752 lockdep_assert_held(&rq->lock);
2754 if (!raw_spin_trylock(&p->pi_lock)) {
2755 raw_spin_unlock(&rq->lock);
2756 raw_spin_lock(&p->pi_lock);
2757 raw_spin_lock(&rq->lock);
2760 if (!(p->state & TASK_NORMAL))
2764 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2766 ttwu_do_wakeup(rq, p, 0);
2767 ttwu_stat(p, smp_processor_id(), 0);
2769 raw_spin_unlock(&p->pi_lock);
2773 * wake_up_process - Wake up a specific process
2774 * @p: The process to be woken up.
2776 * Attempt to wake up the nominated process and move it to the set of runnable
2777 * processes. Returns 1 if the process was woken up, 0 if it was already
2780 * It may be assumed that this function implies a write memory barrier before
2781 * changing the task state if and only if any tasks are woken up.
2783 int wake_up_process(struct task_struct *p)
2785 return try_to_wake_up(p, TASK_ALL, 0);
2787 EXPORT_SYMBOL(wake_up_process);
2789 int wake_up_state(struct task_struct *p, unsigned int state)
2791 return try_to_wake_up(p, state, 0);
2795 * Perform scheduler related setup for a newly forked process p.
2796 * p is forked by current.
2798 * __sched_fork() is basic setup used by init_idle() too:
2800 static void __sched_fork(struct task_struct *p)
2805 p->se.exec_start = 0;
2806 p->se.sum_exec_runtime = 0;
2807 p->se.prev_sum_exec_runtime = 0;
2808 p->se.nr_migrations = 0;
2810 INIT_LIST_HEAD(&p->se.group_node);
2812 #ifdef CONFIG_SCHEDSTATS
2813 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2816 INIT_LIST_HEAD(&p->rt.run_list);
2818 #ifdef CONFIG_PREEMPT_NOTIFIERS
2819 INIT_HLIST_HEAD(&p->preempt_notifiers);
2824 * fork()/clone()-time setup:
2826 void sched_fork(struct task_struct *p)
2828 unsigned long flags;
2829 int cpu = get_cpu();
2833 * We mark the process as running here. This guarantees that
2834 * nobody will actually run it, and a signal or other external
2835 * event cannot wake it up and insert it on the runqueue either.
2837 p->state = TASK_RUNNING;
2840 * Revert to default priority/policy on fork if requested.
2842 if (unlikely(p->sched_reset_on_fork)) {
2843 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2844 p->policy = SCHED_NORMAL;
2845 p->normal_prio = p->static_prio;
2848 if (PRIO_TO_NICE(p->static_prio) < 0) {
2849 p->static_prio = NICE_TO_PRIO(0);
2850 p->normal_prio = p->static_prio;
2855 * We don't need the reset flag anymore after the fork. It has
2856 * fulfilled its duty:
2858 p->sched_reset_on_fork = 0;
2862 * Make sure we do not leak PI boosting priority to the child.
2864 p->prio = current->normal_prio;
2866 if (!rt_prio(p->prio))
2867 p->sched_class = &fair_sched_class;
2869 if (p->sched_class->task_fork)
2870 p->sched_class->task_fork(p);
2873 * The child is not yet in the pid-hash so no cgroup attach races,
2874 * and the cgroup is pinned to this child due to cgroup_fork()
2875 * is ran before sched_fork().
2877 * Silence PROVE_RCU.
2879 raw_spin_lock_irqsave(&p->pi_lock, flags);
2880 set_task_cpu(p, cpu);
2881 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2883 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2884 if (likely(sched_info_on()))
2885 memset(&p->sched_info, 0, sizeof(p->sched_info));
2887 #if defined(CONFIG_SMP)
2890 #ifdef CONFIG_PREEMPT
2891 /* Want to start with kernel preemption disabled. */
2892 task_thread_info(p)->preempt_count = 1;
2895 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2902 * wake_up_new_task - wake up a newly created task for the first time.
2904 * This function will do some initial scheduler statistics housekeeping
2905 * that must be done for every newly created context, then puts the task
2906 * on the runqueue and wakes it.
2908 void wake_up_new_task(struct task_struct *p)
2910 unsigned long flags;
2913 raw_spin_lock_irqsave(&p->pi_lock, flags);
2916 * Fork balancing, do it here and not earlier because:
2917 * - cpus_allowed can change in the fork path
2918 * - any previously selected cpu might disappear through hotplug
2920 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
2923 rq = __task_rq_lock(p);
2924 activate_task(rq, p, 0);
2926 trace_sched_wakeup_new(p, true);
2927 check_preempt_curr(rq, p, WF_FORK);
2929 if (p->sched_class->task_woken)
2930 p->sched_class->task_woken(rq, p);
2932 task_rq_unlock(rq, p, &flags);
2935 #ifdef CONFIG_PREEMPT_NOTIFIERS
2938 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2939 * @notifier: notifier struct to register
2941 void preempt_notifier_register(struct preempt_notifier *notifier)
2943 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2945 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2948 * preempt_notifier_unregister - no longer interested in preemption notifications
2949 * @notifier: notifier struct to unregister
2951 * This is safe to call from within a preemption notifier.
2953 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2955 hlist_del(¬ifier->link);
2957 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2959 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2961 struct preempt_notifier *notifier;
2962 struct hlist_node *node;
2964 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2965 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2969 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2970 struct task_struct *next)
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_out(notifier, next);
2979 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2981 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2986 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2987 struct task_struct *next)
2991 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2994 * prepare_task_switch - prepare to switch tasks
2995 * @rq: the runqueue preparing to switch
2996 * @prev: the current task that is being switched out
2997 * @next: the task we are going to switch to.
2999 * This is called with the rq lock held and interrupts off. It must
3000 * be paired with a subsequent finish_task_switch after the context
3003 * prepare_task_switch sets up locking and calls architecture specific
3007 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3008 struct task_struct *next)
3010 sched_info_switch(prev, next);
3011 perf_event_task_sched_out(prev, next);
3012 fire_sched_out_preempt_notifiers(prev, next);
3013 prepare_lock_switch(rq, next);
3014 prepare_arch_switch(next);
3015 trace_sched_switch(prev, next);
3019 * finish_task_switch - clean up after a task-switch
3020 * @rq: runqueue associated with task-switch
3021 * @prev: the thread we just switched away from.
3023 * finish_task_switch must be called after the context switch, paired
3024 * with a prepare_task_switch call before the context switch.
3025 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3026 * and do any other architecture-specific cleanup actions.
3028 * Note that we may have delayed dropping an mm in context_switch(). If
3029 * so, we finish that here outside of the runqueue lock. (Doing it
3030 * with the lock held can cause deadlocks; see schedule() for
3033 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3034 __releases(rq->lock)
3036 struct mm_struct *mm = rq->prev_mm;
3042 * A task struct has one reference for the use as "current".
3043 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3044 * schedule one last time. The schedule call will never return, and
3045 * the scheduled task must drop that reference.
3046 * The test for TASK_DEAD must occur while the runqueue locks are
3047 * still held, otherwise prev could be scheduled on another cpu, die
3048 * there before we look at prev->state, and then the reference would
3050 * Manfred Spraul <manfred@colorfullife.com>
3052 prev_state = prev->state;
3053 finish_arch_switch(prev);
3054 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3055 local_irq_disable();
3056 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3057 perf_event_task_sched_in(current);
3058 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3060 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3061 finish_lock_switch(rq, prev);
3063 fire_sched_in_preempt_notifiers(current);
3066 if (unlikely(prev_state == TASK_DEAD)) {
3068 * Remove function-return probe instances associated with this
3069 * task and put them back on the free list.
3071 kprobe_flush_task(prev);
3072 put_task_struct(prev);
3078 /* assumes rq->lock is held */
3079 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3081 if (prev->sched_class->pre_schedule)
3082 prev->sched_class->pre_schedule(rq, prev);
3085 /* rq->lock is NOT held, but preemption is disabled */
3086 static inline void post_schedule(struct rq *rq)
3088 if (rq->post_schedule) {
3089 unsigned long flags;
3091 raw_spin_lock_irqsave(&rq->lock, flags);
3092 if (rq->curr->sched_class->post_schedule)
3093 rq->curr->sched_class->post_schedule(rq);
3094 raw_spin_unlock_irqrestore(&rq->lock, flags);
3096 rq->post_schedule = 0;
3102 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3106 static inline void post_schedule(struct rq *rq)
3113 * schedule_tail - first thing a freshly forked thread must call.
3114 * @prev: the thread we just switched away from.
3116 asmlinkage void schedule_tail(struct task_struct *prev)
3117 __releases(rq->lock)
3119 struct rq *rq = this_rq();
3121 finish_task_switch(rq, prev);
3124 * FIXME: do we need to worry about rq being invalidated by the
3129 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3130 /* In this case, finish_task_switch does not reenable preemption */
3133 if (current->set_child_tid)
3134 put_user(task_pid_vnr(current), current->set_child_tid);
3138 * context_switch - switch to the new MM and the new
3139 * thread's register state.
3142 context_switch(struct rq *rq, struct task_struct *prev,
3143 struct task_struct *next)
3145 struct mm_struct *mm, *oldmm;
3147 prepare_task_switch(rq, prev, next);
3150 oldmm = prev->active_mm;
3152 * For paravirt, this is coupled with an exit in switch_to to
3153 * combine the page table reload and the switch backend into
3156 arch_start_context_switch(prev);
3159 next->active_mm = oldmm;
3160 atomic_inc(&oldmm->mm_count);
3161 enter_lazy_tlb(oldmm, next);
3163 switch_mm(oldmm, mm, next);
3166 prev->active_mm = NULL;
3167 rq->prev_mm = oldmm;
3170 * Since the runqueue lock will be released by the next
3171 * task (which is an invalid locking op but in the case
3172 * of the scheduler it's an obvious special-case), so we
3173 * do an early lockdep release here:
3175 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3176 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3179 /* Here we just switch the register state and the stack. */
3180 switch_to(prev, next, prev);
3184 * this_rq must be evaluated again because prev may have moved
3185 * CPUs since it called schedule(), thus the 'rq' on its stack
3186 * frame will be invalid.
3188 finish_task_switch(this_rq(), prev);
3192 * nr_running, nr_uninterruptible and nr_context_switches:
3194 * externally visible scheduler statistics: current number of runnable
3195 * threads, current number of uninterruptible-sleeping threads, total
3196 * number of context switches performed since bootup.
3198 unsigned long nr_running(void)
3200 unsigned long i, sum = 0;
3202 for_each_online_cpu(i)
3203 sum += cpu_rq(i)->nr_running;
3208 unsigned long nr_uninterruptible(void)
3210 unsigned long i, sum = 0;
3212 for_each_possible_cpu(i)
3213 sum += cpu_rq(i)->nr_uninterruptible;
3216 * Since we read the counters lockless, it might be slightly
3217 * inaccurate. Do not allow it to go below zero though:
3219 if (unlikely((long)sum < 0))
3225 unsigned long long nr_context_switches(void)
3228 unsigned long long sum = 0;
3230 for_each_possible_cpu(i)
3231 sum += cpu_rq(i)->nr_switches;
3236 unsigned long nr_iowait(void)
3238 unsigned long i, sum = 0;
3240 for_each_possible_cpu(i)
3241 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3246 unsigned long nr_iowait_cpu(int cpu)
3248 struct rq *this = cpu_rq(cpu);
3249 return atomic_read(&this->nr_iowait);
3252 unsigned long this_cpu_load(void)
3254 struct rq *this = this_rq();
3255 return this->cpu_load[0];
3259 /* Variables and functions for calc_load */
3260 static atomic_long_t calc_load_tasks;
3261 static unsigned long calc_load_update;
3262 unsigned long avenrun[3];
3263 EXPORT_SYMBOL(avenrun);
3265 static long calc_load_fold_active(struct rq *this_rq)
3267 long nr_active, delta = 0;
3269 nr_active = this_rq->nr_running;
3270 nr_active += (long) this_rq->nr_uninterruptible;
3272 if (nr_active != this_rq->calc_load_active) {
3273 delta = nr_active - this_rq->calc_load_active;
3274 this_rq->calc_load_active = nr_active;
3280 static unsigned long
3281 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3284 load += active * (FIXED_1 - exp);
3285 load += 1UL << (FSHIFT - 1);
3286 return load >> FSHIFT;
3291 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3293 * When making the ILB scale, we should try to pull this in as well.
3295 static atomic_long_t calc_load_tasks_idle;
3297 static void calc_load_account_idle(struct rq *this_rq)
3301 delta = calc_load_fold_active(this_rq);
3303 atomic_long_add(delta, &calc_load_tasks_idle);
3306 static long calc_load_fold_idle(void)
3311 * Its got a race, we don't care...
3313 if (atomic_long_read(&calc_load_tasks_idle))
3314 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3320 * fixed_power_int - compute: x^n, in O(log n) time
3322 * @x: base of the power
3323 * @frac_bits: fractional bits of @x
3324 * @n: power to raise @x to.
3326 * By exploiting the relation between the definition of the natural power
3327 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3328 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3329 * (where: n_i \elem {0, 1}, the binary vector representing n),
3330 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3331 * of course trivially computable in O(log_2 n), the length of our binary
3334 static unsigned long
3335 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3337 unsigned long result = 1UL << frac_bits;
3342 result += 1UL << (frac_bits - 1);
3343 result >>= frac_bits;
3349 x += 1UL << (frac_bits - 1);
3357 * a1 = a0 * e + a * (1 - e)
3359 * a2 = a1 * e + a * (1 - e)
3360 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3361 * = a0 * e^2 + a * (1 - e) * (1 + e)
3363 * a3 = a2 * e + a * (1 - e)
3364 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3365 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3369 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3370 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3371 * = a0 * e^n + a * (1 - e^n)
3373 * [1] application of the geometric series:
3376 * S_n := \Sum x^i = -------------
3379 static unsigned long
3380 calc_load_n(unsigned long load, unsigned long exp,
3381 unsigned long active, unsigned int n)
3384 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3388 * NO_HZ can leave us missing all per-cpu ticks calling
3389 * calc_load_account_active(), but since an idle CPU folds its delta into
3390 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3391 * in the pending idle delta if our idle period crossed a load cycle boundary.
3393 * Once we've updated the global active value, we need to apply the exponential
3394 * weights adjusted to the number of cycles missed.
3396 static void calc_global_nohz(void)
3398 long delta, active, n;
3401 * If we crossed a calc_load_update boundary, make sure to fold
3402 * any pending idle changes, the respective CPUs might have
3403 * missed the tick driven calc_load_account_active() update
3406 delta = calc_load_fold_idle();
3408 atomic_long_add(delta, &calc_load_tasks);
3411 * It could be the one fold was all it took, we done!
3413 if (time_before(jiffies, calc_load_update + 10))
3417 * Catch-up, fold however many we are behind still
3419 delta = jiffies - calc_load_update - 10;
3420 n = 1 + (delta / LOAD_FREQ);
3422 active = atomic_long_read(&calc_load_tasks);
3423 active = active > 0 ? active * FIXED_1 : 0;
3425 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3426 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3427 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3429 calc_load_update += n * LOAD_FREQ;
3432 static void calc_load_account_idle(struct rq *this_rq)
3436 static inline long calc_load_fold_idle(void)
3441 static void calc_global_nohz(void)
3447 * get_avenrun - get the load average array
3448 * @loads: pointer to dest load array
3449 * @offset: offset to add
3450 * @shift: shift count to shift the result left
3452 * These values are estimates at best, so no need for locking.
3454 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3456 loads[0] = (avenrun[0] + offset) << shift;
3457 loads[1] = (avenrun[1] + offset) << shift;
3458 loads[2] = (avenrun[2] + offset) << shift;
3462 * calc_load - update the avenrun load estimates 10 ticks after the
3463 * CPUs have updated calc_load_tasks.
3465 void calc_global_load(unsigned long ticks)
3469 if (time_before(jiffies, calc_load_update + 10))
3472 active = atomic_long_read(&calc_load_tasks);
3473 active = active > 0 ? active * FIXED_1 : 0;
3475 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3476 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3477 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3479 calc_load_update += LOAD_FREQ;
3482 * Account one period with whatever state we found before
3483 * folding in the nohz state and ageing the entire idle period.
3485 * This avoids loosing a sample when we go idle between
3486 * calc_load_account_active() (10 ticks ago) and now and thus
3493 * Called from update_cpu_load() to periodically update this CPU's
3496 static void calc_load_account_active(struct rq *this_rq)
3500 if (time_before(jiffies, this_rq->calc_load_update))
3503 delta = calc_load_fold_active(this_rq);
3504 delta += calc_load_fold_idle();
3506 atomic_long_add(delta, &calc_load_tasks);
3508 this_rq->calc_load_update += LOAD_FREQ;
3512 * The exact cpuload at various idx values, calculated at every tick would be
3513 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3515 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3516 * on nth tick when cpu may be busy, then we have:
3517 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3518 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3520 * decay_load_missed() below does efficient calculation of
3521 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3522 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3524 * The calculation is approximated on a 128 point scale.
3525 * degrade_zero_ticks is the number of ticks after which load at any
3526 * particular idx is approximated to be zero.
3527 * degrade_factor is a precomputed table, a row for each load idx.
3528 * Each column corresponds to degradation factor for a power of two ticks,
3529 * based on 128 point scale.
3531 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3532 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3534 * With this power of 2 load factors, we can degrade the load n times
3535 * by looking at 1 bits in n and doing as many mult/shift instead of
3536 * n mult/shifts needed by the exact degradation.
3538 #define DEGRADE_SHIFT 7
3539 static const unsigned char
3540 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3541 static const unsigned char
3542 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3543 {0, 0, 0, 0, 0, 0, 0, 0},
3544 {64, 32, 8, 0, 0, 0, 0, 0},
3545 {96, 72, 40, 12, 1, 0, 0},
3546 {112, 98, 75, 43, 15, 1, 0},
3547 {120, 112, 98, 76, 45, 16, 2} };
3550 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3551 * would be when CPU is idle and so we just decay the old load without
3552 * adding any new load.
3554 static unsigned long
3555 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3559 if (!missed_updates)
3562 if (missed_updates >= degrade_zero_ticks[idx])
3566 return load >> missed_updates;
3568 while (missed_updates) {
3569 if (missed_updates % 2)
3570 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3572 missed_updates >>= 1;
3579 * Update rq->cpu_load[] statistics. This function is usually called every
3580 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3581 * every tick. We fix it up based on jiffies.
3583 static void update_cpu_load(struct rq *this_rq)
3585 unsigned long this_load = this_rq->load.weight;
3586 unsigned long curr_jiffies = jiffies;
3587 unsigned long pending_updates;
3590 this_rq->nr_load_updates++;
3592 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3593 if (curr_jiffies == this_rq->last_load_update_tick)
3596 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3597 this_rq->last_load_update_tick = curr_jiffies;
3599 /* Update our load: */
3600 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3601 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3602 unsigned long old_load, new_load;
3604 /* scale is effectively 1 << i now, and >> i divides by scale */
3606 old_load = this_rq->cpu_load[i];
3607 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3608 new_load = this_load;
3610 * Round up the averaging division if load is increasing. This
3611 * prevents us from getting stuck on 9 if the load is 10, for
3614 if (new_load > old_load)
3615 new_load += scale - 1;
3617 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3620 sched_avg_update(this_rq);
3623 static void update_cpu_load_active(struct rq *this_rq)
3625 update_cpu_load(this_rq);
3627 calc_load_account_active(this_rq);
3633 * sched_exec - execve() is a valuable balancing opportunity, because at
3634 * this point the task has the smallest effective memory and cache footprint.
3636 void sched_exec(void)
3638 struct task_struct *p = current;
3639 unsigned long flags;
3642 raw_spin_lock_irqsave(&p->pi_lock, flags);
3643 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3644 if (dest_cpu == smp_processor_id())
3647 if (likely(cpu_active(dest_cpu))) {
3648 struct migration_arg arg = { p, dest_cpu };
3650 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3651 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3655 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3660 DEFINE_PER_CPU(struct kernel_stat, kstat);
3662 EXPORT_PER_CPU_SYMBOL(kstat);
3665 * Return any ns on the sched_clock that have not yet been accounted in
3666 * @p in case that task is currently running.
3668 * Called with task_rq_lock() held on @rq.
3670 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3674 if (task_current(rq, p)) {
3675 update_rq_clock(rq);
3676 ns = rq->clock_task - p->se.exec_start;
3684 unsigned long long task_delta_exec(struct task_struct *p)
3686 unsigned long flags;
3690 rq = task_rq_lock(p, &flags);
3691 ns = do_task_delta_exec(p, rq);
3692 task_rq_unlock(rq, p, &flags);
3698 * Return accounted runtime for the task.
3699 * In case the task is currently running, return the runtime plus current's
3700 * pending runtime that have not been accounted yet.
3702 unsigned long long task_sched_runtime(struct task_struct *p)
3704 unsigned long flags;
3708 rq = task_rq_lock(p, &flags);
3709 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3710 task_rq_unlock(rq, p, &flags);
3716 * Account user cpu time to a process.
3717 * @p: the process that the cpu time gets accounted to
3718 * @cputime: the cpu time spent in user space since the last update
3719 * @cputime_scaled: cputime scaled by cpu frequency
3721 void account_user_time(struct task_struct *p, cputime_t cputime,
3722 cputime_t cputime_scaled)
3724 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3727 /* Add user time to process. */
3728 p->utime = cputime_add(p->utime, cputime);
3729 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3730 account_group_user_time(p, cputime);
3732 /* Add user time to cpustat. */
3733 tmp = cputime_to_cputime64(cputime);
3734 if (TASK_NICE(p) > 0)
3735 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3737 cpustat->user = cputime64_add(cpustat->user, tmp);
3739 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3740 /* Account for user time used */
3741 acct_update_integrals(p);
3745 * Account guest cpu time to a process.
3746 * @p: the process that the cpu time gets accounted to
3747 * @cputime: the cpu time spent in virtual machine since the last update
3748 * @cputime_scaled: cputime scaled by cpu frequency
3750 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3751 cputime_t cputime_scaled)
3754 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3756 tmp = cputime_to_cputime64(cputime);
3758 /* Add guest time to process. */
3759 p->utime = cputime_add(p->utime, cputime);
3760 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3761 account_group_user_time(p, cputime);
3762 p->gtime = cputime_add(p->gtime, cputime);
3764 /* Add guest time to cpustat. */
3765 if (TASK_NICE(p) > 0) {
3766 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3767 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3769 cpustat->user = cputime64_add(cpustat->user, tmp);
3770 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3775 * Account system cpu time to a process and desired cpustat field
3776 * @p: the process that the cpu time gets accounted to
3777 * @cputime: the cpu time spent in kernel space since the last update
3778 * @cputime_scaled: cputime scaled by cpu frequency
3779 * @target_cputime64: pointer to cpustat field that has to be updated
3782 void __account_system_time(struct task_struct *p, cputime_t cputime,
3783 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3785 cputime64_t tmp = cputime_to_cputime64(cputime);
3787 /* Add system time to process. */
3788 p->stime = cputime_add(p->stime, cputime);
3789 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3790 account_group_system_time(p, cputime);
3792 /* Add system time to cpustat. */
3793 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3794 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3796 /* Account for system time used */
3797 acct_update_integrals(p);
3801 * Account system cpu time to a process.
3802 * @p: the process that the cpu time gets accounted to
3803 * @hardirq_offset: the offset to subtract from hardirq_count()
3804 * @cputime: the cpu time spent in kernel space since the last update
3805 * @cputime_scaled: cputime scaled by cpu frequency
3807 void account_system_time(struct task_struct *p, int hardirq_offset,
3808 cputime_t cputime, cputime_t cputime_scaled)
3810 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3811 cputime64_t *target_cputime64;
3813 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3814 account_guest_time(p, cputime, cputime_scaled);
3818 if (hardirq_count() - hardirq_offset)
3819 target_cputime64 = &cpustat->irq;
3820 else if (in_serving_softirq())
3821 target_cputime64 = &cpustat->softirq;
3823 target_cputime64 = &cpustat->system;
3825 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3829 * Account for involuntary wait time.
3830 * @cputime: the cpu time spent in involuntary wait
3832 void account_steal_time(cputime_t cputime)
3834 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3835 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3837 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3841 * Account for idle time.
3842 * @cputime: the cpu time spent in idle wait
3844 void account_idle_time(cputime_t cputime)
3846 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3847 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3848 struct rq *rq = this_rq();
3850 if (atomic_read(&rq->nr_iowait) > 0)
3851 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3853 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3856 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3858 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3860 * Account a tick to a process and cpustat
3861 * @p: the process that the cpu time gets accounted to
3862 * @user_tick: is the tick from userspace
3863 * @rq: the pointer to rq
3865 * Tick demultiplexing follows the order
3866 * - pending hardirq update
3867 * - pending softirq update
3871 * - check for guest_time
3872 * - else account as system_time
3874 * Check for hardirq is done both for system and user time as there is
3875 * no timer going off while we are on hardirq and hence we may never get an
3876 * opportunity to update it solely in system time.
3877 * p->stime and friends are only updated on system time and not on irq
3878 * softirq as those do not count in task exec_runtime any more.
3880 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3883 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3884 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
3885 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3887 if (irqtime_account_hi_update()) {
3888 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3889 } else if (irqtime_account_si_update()) {
3890 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3891 } else if (this_cpu_ksoftirqd() == p) {
3893 * ksoftirqd time do not get accounted in cpu_softirq_time.
3894 * So, we have to handle it separately here.
3895 * Also, p->stime needs to be updated for ksoftirqd.
3897 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3899 } else if (user_tick) {
3900 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3901 } else if (p == rq->idle) {
3902 account_idle_time(cputime_one_jiffy);
3903 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3904 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3906 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3911 static void irqtime_account_idle_ticks(int ticks)
3914 struct rq *rq = this_rq();
3916 for (i = 0; i < ticks; i++)
3917 irqtime_account_process_tick(current, 0, rq);
3919 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
3920 static void irqtime_account_idle_ticks(int ticks) {}
3921 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3923 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3926 * Account a single tick of cpu time.
3927 * @p: the process that the cpu time gets accounted to
3928 * @user_tick: indicates if the tick is a user or a system tick
3930 void account_process_tick(struct task_struct *p, int user_tick)
3932 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3933 struct rq *rq = this_rq();
3935 if (sched_clock_irqtime) {
3936 irqtime_account_process_tick(p, user_tick, rq);
3941 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3942 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3943 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3946 account_idle_time(cputime_one_jiffy);
3950 * Account multiple ticks of steal time.
3951 * @p: the process from which the cpu time has been stolen
3952 * @ticks: number of stolen ticks
3954 void account_steal_ticks(unsigned long ticks)
3956 account_steal_time(jiffies_to_cputime(ticks));
3960 * Account multiple ticks of idle time.
3961 * @ticks: number of stolen ticks
3963 void account_idle_ticks(unsigned long ticks)
3966 if (sched_clock_irqtime) {
3967 irqtime_account_idle_ticks(ticks);
3971 account_idle_time(jiffies_to_cputime(ticks));
3977 * Use precise platform statistics if available:
3979 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3980 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3986 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3988 struct task_cputime cputime;
3990 thread_group_cputime(p, &cputime);
3992 *ut = cputime.utime;
3993 *st = cputime.stime;
3997 #ifndef nsecs_to_cputime
3998 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
4001 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4003 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
4006 * Use CFS's precise accounting:
4008 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
4014 do_div(temp, total);
4015 utime = (cputime_t)temp;
4020 * Compare with previous values, to keep monotonicity:
4022 p->prev_utime = max(p->prev_utime, utime);
4023 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4025 *ut = p->prev_utime;
4026 *st = p->prev_stime;
4030 * Must be called with siglock held.
4032 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4034 struct signal_struct *sig = p->signal;
4035 struct task_cputime cputime;
4036 cputime_t rtime, utime, total;
4038 thread_group_cputime(p, &cputime);
4040 total = cputime_add(cputime.utime, cputime.stime);
4041 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4046 temp *= cputime.utime;
4047 do_div(temp, total);
4048 utime = (cputime_t)temp;
4052 sig->prev_utime = max(sig->prev_utime, utime);
4053 sig->prev_stime = max(sig->prev_stime,
4054 cputime_sub(rtime, sig->prev_utime));
4056 *ut = sig->prev_utime;
4057 *st = sig->prev_stime;
4062 * This function gets called by the timer code, with HZ frequency.
4063 * We call it with interrupts disabled.
4065 void scheduler_tick(void)
4067 int cpu = smp_processor_id();
4068 struct rq *rq = cpu_rq(cpu);
4069 struct task_struct *curr = rq->curr;
4073 raw_spin_lock(&rq->lock);
4074 update_rq_clock(rq);
4075 update_cpu_load_active(rq);
4076 curr->sched_class->task_tick(rq, curr, 0);
4077 raw_spin_unlock(&rq->lock);
4079 perf_event_task_tick();
4082 rq->idle_at_tick = idle_cpu(cpu);
4083 trigger_load_balance(rq, cpu);
4087 notrace unsigned long get_parent_ip(unsigned long addr)
4089 if (in_lock_functions(addr)) {
4090 addr = CALLER_ADDR2;
4091 if (in_lock_functions(addr))
4092 addr = CALLER_ADDR3;
4097 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4098 defined(CONFIG_PREEMPT_TRACER))
4100 void __kprobes add_preempt_count(int val)
4102 #ifdef CONFIG_DEBUG_PREEMPT
4106 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4109 preempt_count() += val;
4110 #ifdef CONFIG_DEBUG_PREEMPT
4112 * Spinlock count overflowing soon?
4114 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4117 if (preempt_count() == val)
4118 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4120 EXPORT_SYMBOL(add_preempt_count);
4122 void __kprobes sub_preempt_count(int val)
4124 #ifdef CONFIG_DEBUG_PREEMPT
4128 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4131 * Is the spinlock portion underflowing?
4133 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4134 !(preempt_count() & PREEMPT_MASK)))
4138 if (preempt_count() == val)
4139 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4140 preempt_count() -= val;
4142 EXPORT_SYMBOL(sub_preempt_count);
4147 * Print scheduling while atomic bug:
4149 static noinline void __schedule_bug(struct task_struct *prev)
4151 struct pt_regs *regs = get_irq_regs();
4153 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4154 prev->comm, prev->pid, preempt_count());
4156 debug_show_held_locks(prev);
4158 if (irqs_disabled())
4159 print_irqtrace_events(prev);
4168 * Various schedule()-time debugging checks and statistics:
4170 static inline void schedule_debug(struct task_struct *prev)
4173 * Test if we are atomic. Since do_exit() needs to call into
4174 * schedule() atomically, we ignore that path for now.
4175 * Otherwise, whine if we are scheduling when we should not be.
4177 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4178 __schedule_bug(prev);
4180 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4182 schedstat_inc(this_rq(), sched_count);
4185 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4187 if (prev->on_rq || rq->skip_clock_update < 0)
4188 update_rq_clock(rq);
4189 prev->sched_class->put_prev_task(rq, prev);
4193 * Pick up the highest-prio task:
4195 static inline struct task_struct *
4196 pick_next_task(struct rq *rq)
4198 const struct sched_class *class;
4199 struct task_struct *p;
4202 * Optimization: we know that if all tasks are in
4203 * the fair class we can call that function directly:
4205 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4206 p = fair_sched_class.pick_next_task(rq);
4211 for_each_class(class) {
4212 p = class->pick_next_task(rq);
4217 BUG(); /* the idle class will always have a runnable task */
4221 * __schedule() is the main scheduler function.
4223 static void __sched __schedule(void)
4225 struct task_struct *prev, *next;
4226 unsigned long *switch_count;
4232 cpu = smp_processor_id();
4234 rcu_note_context_switch(cpu);
4237 schedule_debug(prev);
4239 if (sched_feat(HRTICK))
4242 raw_spin_lock_irq(&rq->lock);
4244 switch_count = &prev->nivcsw;
4245 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4246 if (unlikely(signal_pending_state(prev->state, prev))) {
4247 prev->state = TASK_RUNNING;
4249 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4253 * If a worker went to sleep, notify and ask workqueue
4254 * whether it wants to wake up a task to maintain
4257 if (prev->flags & PF_WQ_WORKER) {
4258 struct task_struct *to_wakeup;
4260 to_wakeup = wq_worker_sleeping(prev, cpu);
4262 try_to_wake_up_local(to_wakeup);
4265 switch_count = &prev->nvcsw;
4268 pre_schedule(rq, prev);
4270 if (unlikely(!rq->nr_running))
4271 idle_balance(cpu, rq);
4273 put_prev_task(rq, prev);
4274 next = pick_next_task(rq);
4275 clear_tsk_need_resched(prev);
4276 rq->skip_clock_update = 0;
4278 if (likely(prev != next)) {
4283 context_switch(rq, prev, next); /* unlocks the rq */
4285 * The context switch have flipped the stack from under us
4286 * and restored the local variables which were saved when
4287 * this task called schedule() in the past. prev == current
4288 * is still correct, but it can be moved to another cpu/rq.
4290 cpu = smp_processor_id();
4293 raw_spin_unlock_irq(&rq->lock);
4297 preempt_enable_no_resched();
4302 static inline void sched_submit_work(struct task_struct *tsk)
4307 * If we are going to sleep and we have plugged IO queued,
4308 * make sure to submit it to avoid deadlocks.
4310 if (blk_needs_flush_plug(tsk))
4311 blk_schedule_flush_plug(tsk);
4314 asmlinkage void __sched schedule(void)
4316 struct task_struct *tsk = current;
4318 sched_submit_work(tsk);
4321 EXPORT_SYMBOL(schedule);
4323 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4325 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4330 if (lock->owner != owner)
4334 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4335 * lock->owner still matches owner, if that fails, owner might
4336 * point to free()d memory, if it still matches, the rcu_read_lock()
4337 * ensures the memory stays valid.
4341 ret = owner->on_cpu;
4349 * Look out! "owner" is an entirely speculative pointer
4350 * access and not reliable.
4352 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4354 if (!sched_feat(OWNER_SPIN))
4357 while (owner_running(lock, owner)) {
4361 arch_mutex_cpu_relax();
4365 * If the owner changed to another task there is likely
4366 * heavy contention, stop spinning.
4375 #ifdef CONFIG_PREEMPT
4377 * this is the entry point to schedule() from in-kernel preemption
4378 * off of preempt_enable. Kernel preemptions off return from interrupt
4379 * occur there and call schedule directly.
4381 asmlinkage void __sched notrace preempt_schedule(void)
4383 struct thread_info *ti = current_thread_info();
4386 * If there is a non-zero preempt_count or interrupts are disabled,
4387 * we do not want to preempt the current task. Just return..
4389 if (likely(ti->preempt_count || irqs_disabled()))
4393 add_preempt_count_notrace(PREEMPT_ACTIVE);
4395 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4398 * Check again in case we missed a preemption opportunity
4399 * between schedule and now.
4402 } while (need_resched());
4404 EXPORT_SYMBOL(preempt_schedule);
4407 * this is the entry point to schedule() from kernel preemption
4408 * off of irq context.
4409 * Note, that this is called and return with irqs disabled. This will
4410 * protect us against recursive calling from irq.
4412 asmlinkage void __sched preempt_schedule_irq(void)
4414 struct thread_info *ti = current_thread_info();
4416 /* Catch callers which need to be fixed */
4417 BUG_ON(ti->preempt_count || !irqs_disabled());
4420 add_preempt_count(PREEMPT_ACTIVE);
4423 local_irq_disable();
4424 sub_preempt_count(PREEMPT_ACTIVE);
4427 * Check again in case we missed a preemption opportunity
4428 * between schedule and now.
4431 } while (need_resched());
4434 #endif /* CONFIG_PREEMPT */
4436 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4439 return try_to_wake_up(curr->private, mode, wake_flags);
4441 EXPORT_SYMBOL(default_wake_function);
4444 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4445 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4446 * number) then we wake all the non-exclusive tasks and one exclusive task.
4448 * There are circumstances in which we can try to wake a task which has already
4449 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4450 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4452 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4453 int nr_exclusive, int wake_flags, void *key)
4455 wait_queue_t *curr, *next;
4457 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4458 unsigned flags = curr->flags;
4460 if (curr->func(curr, mode, wake_flags, key) &&
4461 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4467 * __wake_up - wake up threads blocked on a waitqueue.
4469 * @mode: which threads
4470 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4471 * @key: is directly passed to the wakeup function
4473 * It may be assumed that this function implies a write memory barrier before
4474 * changing the task state if and only if any tasks are woken up.
4476 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4477 int nr_exclusive, void *key)
4479 unsigned long flags;
4481 spin_lock_irqsave(&q->lock, flags);
4482 __wake_up_common(q, mode, nr_exclusive, 0, key);
4483 spin_unlock_irqrestore(&q->lock, flags);
4485 EXPORT_SYMBOL(__wake_up);
4488 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4490 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4492 __wake_up_common(q, mode, 1, 0, NULL);
4494 EXPORT_SYMBOL_GPL(__wake_up_locked);
4496 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4498 __wake_up_common(q, mode, 1, 0, key);
4500 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4503 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4505 * @mode: which threads
4506 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4507 * @key: opaque value to be passed to wakeup targets
4509 * The sync wakeup differs that the waker knows that it will schedule
4510 * away soon, so while the target thread will be woken up, it will not
4511 * be migrated to another CPU - ie. the two threads are 'synchronized'
4512 * with each other. This can prevent needless bouncing between CPUs.
4514 * On UP it can prevent extra preemption.
4516 * It may be assumed that this function implies a write memory barrier before
4517 * changing the task state if and only if any tasks are woken up.
4519 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4520 int nr_exclusive, void *key)
4522 unsigned long flags;
4523 int wake_flags = WF_SYNC;
4528 if (unlikely(!nr_exclusive))
4531 spin_lock_irqsave(&q->lock, flags);
4532 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4533 spin_unlock_irqrestore(&q->lock, flags);
4535 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4538 * __wake_up_sync - see __wake_up_sync_key()
4540 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4542 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4544 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4547 * complete: - signals a single thread waiting on this completion
4548 * @x: holds the state of this particular completion
4550 * This will wake up a single thread waiting on this completion. Threads will be
4551 * awakened in the same order in which they were queued.
4553 * See also complete_all(), wait_for_completion() and related routines.
4555 * It may be assumed that this function implies a write memory barrier before
4556 * changing the task state if and only if any tasks are woken up.
4558 void complete(struct completion *x)
4560 unsigned long flags;
4562 spin_lock_irqsave(&x->wait.lock, flags);
4564 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4565 spin_unlock_irqrestore(&x->wait.lock, flags);
4567 EXPORT_SYMBOL(complete);
4570 * complete_all: - signals all threads waiting on this completion
4571 * @x: holds the state of this particular completion
4573 * This will wake up all threads waiting on this particular completion event.
4575 * It may be assumed that this function implies a write memory barrier before
4576 * changing the task state if and only if any tasks are woken up.
4578 void complete_all(struct completion *x)
4580 unsigned long flags;
4582 spin_lock_irqsave(&x->wait.lock, flags);
4583 x->done += UINT_MAX/2;
4584 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4585 spin_unlock_irqrestore(&x->wait.lock, flags);
4587 EXPORT_SYMBOL(complete_all);
4589 static inline long __sched
4590 do_wait_for_common(struct completion *x, long timeout, int state)
4593 DECLARE_WAITQUEUE(wait, current);
4595 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4597 if (signal_pending_state(state, current)) {
4598 timeout = -ERESTARTSYS;
4601 __set_current_state(state);
4602 spin_unlock_irq(&x->wait.lock);
4603 timeout = schedule_timeout(timeout);
4604 spin_lock_irq(&x->wait.lock);
4605 } while (!x->done && timeout);
4606 __remove_wait_queue(&x->wait, &wait);
4611 return timeout ?: 1;
4615 wait_for_common(struct completion *x, long timeout, int state)
4619 spin_lock_irq(&x->wait.lock);
4620 timeout = do_wait_for_common(x, timeout, state);
4621 spin_unlock_irq(&x->wait.lock);
4626 * wait_for_completion: - waits for completion of a task
4627 * @x: holds the state of this particular completion
4629 * This waits to be signaled for completion of a specific task. It is NOT
4630 * interruptible and there is no timeout.
4632 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4633 * and interrupt capability. Also see complete().
4635 void __sched wait_for_completion(struct completion *x)
4637 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4639 EXPORT_SYMBOL(wait_for_completion);
4642 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4643 * @x: holds the state of this particular completion
4644 * @timeout: timeout value in jiffies
4646 * This waits for either a completion of a specific task to be signaled or for a
4647 * specified timeout to expire. The timeout is in jiffies. It is not
4650 unsigned long __sched
4651 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4653 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4655 EXPORT_SYMBOL(wait_for_completion_timeout);
4658 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4659 * @x: holds the state of this particular completion
4661 * This waits for completion of a specific task to be signaled. It is
4664 int __sched wait_for_completion_interruptible(struct completion *x)
4666 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4667 if (t == -ERESTARTSYS)
4671 EXPORT_SYMBOL(wait_for_completion_interruptible);
4674 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4675 * @x: holds the state of this particular completion
4676 * @timeout: timeout value in jiffies
4678 * This waits for either a completion of a specific task to be signaled or for a
4679 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4682 wait_for_completion_interruptible_timeout(struct completion *x,
4683 unsigned long timeout)
4685 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4687 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4690 * wait_for_completion_killable: - waits for completion of a task (killable)
4691 * @x: holds the state of this particular completion
4693 * This waits to be signaled for completion of a specific task. It can be
4694 * interrupted by a kill signal.
4696 int __sched wait_for_completion_killable(struct completion *x)
4698 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4699 if (t == -ERESTARTSYS)
4703 EXPORT_SYMBOL(wait_for_completion_killable);
4706 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4707 * @x: holds the state of this particular completion
4708 * @timeout: timeout value in jiffies
4710 * This waits for either a completion of a specific task to be
4711 * signaled or for a specified timeout to expire. It can be
4712 * interrupted by a kill signal. The timeout is in jiffies.
4715 wait_for_completion_killable_timeout(struct completion *x,
4716 unsigned long timeout)
4718 return wait_for_common(x, timeout, TASK_KILLABLE);
4720 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4723 * try_wait_for_completion - try to decrement a completion without blocking
4724 * @x: completion structure
4726 * Returns: 0 if a decrement cannot be done without blocking
4727 * 1 if a decrement succeeded.
4729 * If a completion is being used as a counting completion,
4730 * attempt to decrement the counter without blocking. This
4731 * enables us to avoid waiting if the resource the completion
4732 * is protecting is not available.
4734 bool try_wait_for_completion(struct completion *x)
4736 unsigned long flags;
4739 spin_lock_irqsave(&x->wait.lock, flags);
4744 spin_unlock_irqrestore(&x->wait.lock, flags);
4747 EXPORT_SYMBOL(try_wait_for_completion);
4750 * completion_done - Test to see if a completion has any waiters
4751 * @x: completion structure
4753 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4754 * 1 if there are no waiters.
4757 bool completion_done(struct completion *x)
4759 unsigned long flags;
4762 spin_lock_irqsave(&x->wait.lock, flags);
4765 spin_unlock_irqrestore(&x->wait.lock, flags);
4768 EXPORT_SYMBOL(completion_done);
4771 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4773 unsigned long flags;
4776 init_waitqueue_entry(&wait, current);
4778 __set_current_state(state);
4780 spin_lock_irqsave(&q->lock, flags);
4781 __add_wait_queue(q, &wait);
4782 spin_unlock(&q->lock);
4783 timeout = schedule_timeout(timeout);
4784 spin_lock_irq(&q->lock);
4785 __remove_wait_queue(q, &wait);
4786 spin_unlock_irqrestore(&q->lock, flags);
4791 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4793 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4795 EXPORT_SYMBOL(interruptible_sleep_on);
4798 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4800 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4802 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4804 void __sched sleep_on(wait_queue_head_t *q)
4806 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4808 EXPORT_SYMBOL(sleep_on);
4810 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4812 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4814 EXPORT_SYMBOL(sleep_on_timeout);
4816 #ifdef CONFIG_RT_MUTEXES
4819 * rt_mutex_setprio - set the current priority of a task
4821 * @prio: prio value (kernel-internal form)
4823 * This function changes the 'effective' priority of a task. It does
4824 * not touch ->normal_prio like __setscheduler().
4826 * Used by the rt_mutex code to implement priority inheritance logic.
4828 void rt_mutex_setprio(struct task_struct *p, int prio)
4830 int oldprio, on_rq, running;
4832 const struct sched_class *prev_class;
4834 BUG_ON(prio < 0 || prio > MAX_PRIO);
4836 rq = __task_rq_lock(p);
4838 trace_sched_pi_setprio(p, prio);
4840 prev_class = p->sched_class;
4842 running = task_current(rq, p);
4844 dequeue_task(rq, p, 0);
4846 p->sched_class->put_prev_task(rq, p);
4849 p->sched_class = &rt_sched_class;
4851 p->sched_class = &fair_sched_class;
4856 p->sched_class->set_curr_task(rq);
4858 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4860 check_class_changed(rq, p, prev_class, oldprio);
4861 __task_rq_unlock(rq);
4866 void set_user_nice(struct task_struct *p, long nice)
4868 int old_prio, delta, on_rq;
4869 unsigned long flags;
4872 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4875 * We have to be careful, if called from sys_setpriority(),
4876 * the task might be in the middle of scheduling on another CPU.
4878 rq = task_rq_lock(p, &flags);
4880 * The RT priorities are set via sched_setscheduler(), but we still
4881 * allow the 'normal' nice value to be set - but as expected
4882 * it wont have any effect on scheduling until the task is
4883 * SCHED_FIFO/SCHED_RR:
4885 if (task_has_rt_policy(p)) {
4886 p->static_prio = NICE_TO_PRIO(nice);
4891 dequeue_task(rq, p, 0);
4893 p->static_prio = NICE_TO_PRIO(nice);
4896 p->prio = effective_prio(p);
4897 delta = p->prio - old_prio;
4900 enqueue_task(rq, p, 0);
4902 * If the task increased its priority or is running and
4903 * lowered its priority, then reschedule its CPU:
4905 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4906 resched_task(rq->curr);
4909 task_rq_unlock(rq, p, &flags);
4911 EXPORT_SYMBOL(set_user_nice);
4914 * can_nice - check if a task can reduce its nice value
4918 int can_nice(const struct task_struct *p, const int nice)
4920 /* convert nice value [19,-20] to rlimit style value [1,40] */
4921 int nice_rlim = 20 - nice;
4923 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4924 capable(CAP_SYS_NICE));
4927 #ifdef __ARCH_WANT_SYS_NICE
4930 * sys_nice - change the priority of the current process.
4931 * @increment: priority increment
4933 * sys_setpriority is a more generic, but much slower function that
4934 * does similar things.
4936 SYSCALL_DEFINE1(nice, int, increment)
4941 * Setpriority might change our priority at the same moment.
4942 * We don't have to worry. Conceptually one call occurs first
4943 * and we have a single winner.
4945 if (increment < -40)
4950 nice = TASK_NICE(current) + increment;
4956 if (increment < 0 && !can_nice(current, nice))
4959 retval = security_task_setnice(current, nice);
4963 set_user_nice(current, nice);
4970 * task_prio - return the priority value of a given task.
4971 * @p: the task in question.
4973 * This is the priority value as seen by users in /proc.
4974 * RT tasks are offset by -200. Normal tasks are centered
4975 * around 0, value goes from -16 to +15.
4977 int task_prio(const struct task_struct *p)
4979 return p->prio - MAX_RT_PRIO;
4983 * task_nice - return the nice value of a given task.
4984 * @p: the task in question.
4986 int task_nice(const struct task_struct *p)
4988 return TASK_NICE(p);
4990 EXPORT_SYMBOL(task_nice);
4993 * idle_cpu - is a given cpu idle currently?
4994 * @cpu: the processor in question.
4996 int idle_cpu(int cpu)
4998 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5002 * idle_task - return the idle task for a given cpu.
5003 * @cpu: the processor in question.
5005 struct task_struct *idle_task(int cpu)
5007 return cpu_rq(cpu)->idle;
5011 * find_process_by_pid - find a process with a matching PID value.
5012 * @pid: the pid in question.
5014 static struct task_struct *find_process_by_pid(pid_t pid)
5016 return pid ? find_task_by_vpid(pid) : current;
5019 /* Actually do priority change: must hold rq lock. */
5021 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5024 p->rt_priority = prio;
5025 p->normal_prio = normal_prio(p);
5026 /* we are holding p->pi_lock already */
5027 p->prio = rt_mutex_getprio(p);
5028 if (rt_prio(p->prio))
5029 p->sched_class = &rt_sched_class;
5031 p->sched_class = &fair_sched_class;
5036 * check the target process has a UID that matches the current process's
5038 static bool check_same_owner(struct task_struct *p)
5040 const struct cred *cred = current_cred(), *pcred;
5044 pcred = __task_cred(p);
5045 if (cred->user->user_ns == pcred->user->user_ns)
5046 match = (cred->euid == pcred->euid ||
5047 cred->euid == pcred->uid);
5054 static int __sched_setscheduler(struct task_struct *p, int policy,
5055 const struct sched_param *param, bool user)
5057 int retval, oldprio, oldpolicy = -1, on_rq, running;
5058 unsigned long flags;
5059 const struct sched_class *prev_class;
5063 /* may grab non-irq protected spin_locks */
5064 BUG_ON(in_interrupt());
5066 /* double check policy once rq lock held */
5068 reset_on_fork = p->sched_reset_on_fork;
5069 policy = oldpolicy = p->policy;
5071 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5072 policy &= ~SCHED_RESET_ON_FORK;
5074 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5075 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5076 policy != SCHED_IDLE)
5081 * Valid priorities for SCHED_FIFO and SCHED_RR are
5082 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5083 * SCHED_BATCH and SCHED_IDLE is 0.
5085 if (param->sched_priority < 0 ||
5086 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5087 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5089 if (rt_policy(policy) != (param->sched_priority != 0))
5093 * Allow unprivileged RT tasks to decrease priority:
5095 if (user && !capable(CAP_SYS_NICE)) {
5096 if (rt_policy(policy)) {
5097 unsigned long rlim_rtprio =
5098 task_rlimit(p, RLIMIT_RTPRIO);
5100 /* can't set/change the rt policy */
5101 if (policy != p->policy && !rlim_rtprio)
5104 /* can't increase priority */
5105 if (param->sched_priority > p->rt_priority &&
5106 param->sched_priority > rlim_rtprio)
5111 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5112 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5114 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5115 if (!can_nice(p, TASK_NICE(p)))
5119 /* can't change other user's priorities */
5120 if (!check_same_owner(p))
5123 /* Normal users shall not reset the sched_reset_on_fork flag */
5124 if (p->sched_reset_on_fork && !reset_on_fork)
5129 retval = security_task_setscheduler(p);
5135 * make sure no PI-waiters arrive (or leave) while we are
5136 * changing the priority of the task:
5138 * To be able to change p->policy safely, the appropriate
5139 * runqueue lock must be held.
5141 rq = task_rq_lock(p, &flags);
5144 * Changing the policy of the stop threads its a very bad idea
5146 if (p == rq->stop) {
5147 task_rq_unlock(rq, p, &flags);
5152 * If not changing anything there's no need to proceed further:
5154 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5155 param->sched_priority == p->rt_priority))) {
5157 __task_rq_unlock(rq);
5158 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5162 #ifdef CONFIG_RT_GROUP_SCHED
5165 * Do not allow realtime tasks into groups that have no runtime
5168 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5169 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5170 !task_group_is_autogroup(task_group(p))) {
5171 task_rq_unlock(rq, p, &flags);
5177 /* recheck policy now with rq lock held */
5178 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5179 policy = oldpolicy = -1;
5180 task_rq_unlock(rq, p, &flags);
5184 running = task_current(rq, p);
5186 deactivate_task(rq, p, 0);
5188 p->sched_class->put_prev_task(rq, p);
5190 p->sched_reset_on_fork = reset_on_fork;
5193 prev_class = p->sched_class;
5194 __setscheduler(rq, p, policy, param->sched_priority);
5197 p->sched_class->set_curr_task(rq);
5199 activate_task(rq, p, 0);
5201 check_class_changed(rq, p, prev_class, oldprio);
5202 task_rq_unlock(rq, p, &flags);
5204 rt_mutex_adjust_pi(p);
5210 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5211 * @p: the task in question.
5212 * @policy: new policy.
5213 * @param: structure containing the new RT priority.
5215 * NOTE that the task may be already dead.
5217 int sched_setscheduler(struct task_struct *p, int policy,
5218 const struct sched_param *param)
5220 return __sched_setscheduler(p, policy, param, true);
5222 EXPORT_SYMBOL_GPL(sched_setscheduler);
5225 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5226 * @p: the task in question.
5227 * @policy: new policy.
5228 * @param: structure containing the new RT priority.
5230 * Just like sched_setscheduler, only don't bother checking if the
5231 * current context has permission. For example, this is needed in
5232 * stop_machine(): we create temporary high priority worker threads,
5233 * but our caller might not have that capability.
5235 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5236 const struct sched_param *param)
5238 return __sched_setscheduler(p, policy, param, false);
5242 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5244 struct sched_param lparam;
5245 struct task_struct *p;
5248 if (!param || pid < 0)
5250 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5255 p = find_process_by_pid(pid);
5257 retval = sched_setscheduler(p, policy, &lparam);
5264 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5265 * @pid: the pid in question.
5266 * @policy: new policy.
5267 * @param: structure containing the new RT priority.
5269 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5270 struct sched_param __user *, param)
5272 /* negative values for policy are not valid */
5276 return do_sched_setscheduler(pid, policy, param);
5280 * sys_sched_setparam - set/change the RT priority of a thread
5281 * @pid: the pid in question.
5282 * @param: structure containing the new RT priority.
5284 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5286 return do_sched_setscheduler(pid, -1, param);
5290 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5291 * @pid: the pid in question.
5293 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5295 struct task_struct *p;
5303 p = find_process_by_pid(pid);
5305 retval = security_task_getscheduler(p);
5308 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5315 * sys_sched_getparam - get the RT priority of a thread
5316 * @pid: the pid in question.
5317 * @param: structure containing the RT priority.
5319 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5321 struct sched_param lp;
5322 struct task_struct *p;
5325 if (!param || pid < 0)
5329 p = find_process_by_pid(pid);
5334 retval = security_task_getscheduler(p);
5338 lp.sched_priority = p->rt_priority;
5342 * This one might sleep, we cannot do it with a spinlock held ...
5344 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5353 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5355 cpumask_var_t cpus_allowed, new_mask;
5356 struct task_struct *p;
5362 p = find_process_by_pid(pid);
5369 /* Prevent p going away */
5373 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5377 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5379 goto out_free_cpus_allowed;
5382 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5385 retval = security_task_setscheduler(p);
5389 cpuset_cpus_allowed(p, cpus_allowed);
5390 cpumask_and(new_mask, in_mask, cpus_allowed);
5392 retval = set_cpus_allowed_ptr(p, new_mask);
5395 cpuset_cpus_allowed(p, cpus_allowed);
5396 if (!cpumask_subset(new_mask, cpus_allowed)) {
5398 * We must have raced with a concurrent cpuset
5399 * update. Just reset the cpus_allowed to the
5400 * cpuset's cpus_allowed
5402 cpumask_copy(new_mask, cpus_allowed);
5407 free_cpumask_var(new_mask);
5408 out_free_cpus_allowed:
5409 free_cpumask_var(cpus_allowed);
5416 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5417 struct cpumask *new_mask)
5419 if (len < cpumask_size())
5420 cpumask_clear(new_mask);
5421 else if (len > cpumask_size())
5422 len = cpumask_size();
5424 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5428 * sys_sched_setaffinity - set the cpu affinity of a process
5429 * @pid: pid of the process
5430 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5431 * @user_mask_ptr: user-space pointer to the new cpu mask
5433 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5434 unsigned long __user *, user_mask_ptr)
5436 cpumask_var_t new_mask;
5439 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5442 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5444 retval = sched_setaffinity(pid, new_mask);
5445 free_cpumask_var(new_mask);
5449 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5451 struct task_struct *p;
5452 unsigned long flags;
5459 p = find_process_by_pid(pid);
5463 retval = security_task_getscheduler(p);
5467 raw_spin_lock_irqsave(&p->pi_lock, flags);
5468 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5469 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5479 * sys_sched_getaffinity - get the cpu affinity of a process
5480 * @pid: pid of the process
5481 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5482 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5484 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5485 unsigned long __user *, user_mask_ptr)
5490 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5492 if (len & (sizeof(unsigned long)-1))
5495 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5498 ret = sched_getaffinity(pid, mask);
5500 size_t retlen = min_t(size_t, len, cpumask_size());
5502 if (copy_to_user(user_mask_ptr, mask, retlen))
5507 free_cpumask_var(mask);
5513 * sys_sched_yield - yield the current processor to other threads.
5515 * This function yields the current CPU to other tasks. If there are no
5516 * other threads running on this CPU then this function will return.
5518 SYSCALL_DEFINE0(sched_yield)
5520 struct rq *rq = this_rq_lock();
5522 schedstat_inc(rq, yld_count);
5523 current->sched_class->yield_task(rq);
5526 * Since we are going to call schedule() anyway, there's
5527 * no need to preempt or enable interrupts:
5529 __release(rq->lock);
5530 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5531 do_raw_spin_unlock(&rq->lock);
5532 preempt_enable_no_resched();
5539 static inline int should_resched(void)
5541 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5544 static void __cond_resched(void)
5546 add_preempt_count(PREEMPT_ACTIVE);
5548 sub_preempt_count(PREEMPT_ACTIVE);
5551 int __sched _cond_resched(void)
5553 if (should_resched()) {
5559 EXPORT_SYMBOL(_cond_resched);
5562 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5563 * call schedule, and on return reacquire the lock.
5565 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5566 * operations here to prevent schedule() from being called twice (once via
5567 * spin_unlock(), once by hand).
5569 int __cond_resched_lock(spinlock_t *lock)
5571 int resched = should_resched();
5574 lockdep_assert_held(lock);
5576 if (spin_needbreak(lock) || resched) {
5587 EXPORT_SYMBOL(__cond_resched_lock);
5589 int __sched __cond_resched_softirq(void)
5591 BUG_ON(!in_softirq());
5593 if (should_resched()) {
5601 EXPORT_SYMBOL(__cond_resched_softirq);
5604 * yield - yield the current processor to other threads.
5606 * This is a shortcut for kernel-space yielding - it marks the
5607 * thread runnable and calls sys_sched_yield().
5609 void __sched yield(void)
5611 set_current_state(TASK_RUNNING);
5614 EXPORT_SYMBOL(yield);
5617 * yield_to - yield the current processor to another thread in
5618 * your thread group, or accelerate that thread toward the
5619 * processor it's on.
5621 * @preempt: whether task preemption is allowed or not
5623 * It's the caller's job to ensure that the target task struct
5624 * can't go away on us before we can do any checks.
5626 * Returns true if we indeed boosted the target task.
5628 bool __sched yield_to(struct task_struct *p, bool preempt)
5630 struct task_struct *curr = current;
5631 struct rq *rq, *p_rq;
5632 unsigned long flags;
5635 local_irq_save(flags);
5640 double_rq_lock(rq, p_rq);
5641 while (task_rq(p) != p_rq) {
5642 double_rq_unlock(rq, p_rq);
5646 if (!curr->sched_class->yield_to_task)
5649 if (curr->sched_class != p->sched_class)
5652 if (task_running(p_rq, p) || p->state)
5655 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5657 schedstat_inc(rq, yld_count);
5659 * Make p's CPU reschedule; pick_next_entity takes care of
5662 if (preempt && rq != p_rq)
5663 resched_task(p_rq->curr);
5667 double_rq_unlock(rq, p_rq);
5668 local_irq_restore(flags);
5675 EXPORT_SYMBOL_GPL(yield_to);
5678 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5679 * that process accounting knows that this is a task in IO wait state.
5681 void __sched io_schedule(void)
5683 struct rq *rq = raw_rq();
5685 delayacct_blkio_start();
5686 atomic_inc(&rq->nr_iowait);
5687 blk_flush_plug(current);
5688 current->in_iowait = 1;
5690 current->in_iowait = 0;
5691 atomic_dec(&rq->nr_iowait);
5692 delayacct_blkio_end();
5694 EXPORT_SYMBOL(io_schedule);
5696 long __sched io_schedule_timeout(long timeout)
5698 struct rq *rq = raw_rq();
5701 delayacct_blkio_start();
5702 atomic_inc(&rq->nr_iowait);
5703 blk_flush_plug(current);
5704 current->in_iowait = 1;
5705 ret = schedule_timeout(timeout);
5706 current->in_iowait = 0;
5707 atomic_dec(&rq->nr_iowait);
5708 delayacct_blkio_end();
5713 * sys_sched_get_priority_max - return maximum RT priority.
5714 * @policy: scheduling class.
5716 * this syscall returns the maximum rt_priority that can be used
5717 * by a given scheduling class.
5719 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5726 ret = MAX_USER_RT_PRIO-1;
5738 * sys_sched_get_priority_min - return minimum RT priority.
5739 * @policy: scheduling class.
5741 * this syscall returns the minimum rt_priority that can be used
5742 * by a given scheduling class.
5744 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5762 * sys_sched_rr_get_interval - return the default timeslice of a process.
5763 * @pid: pid of the process.
5764 * @interval: userspace pointer to the timeslice value.
5766 * this syscall writes the default timeslice value of a given process
5767 * into the user-space timespec buffer. A value of '0' means infinity.
5769 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5770 struct timespec __user *, interval)
5772 struct task_struct *p;
5773 unsigned int time_slice;
5774 unsigned long flags;
5784 p = find_process_by_pid(pid);
5788 retval = security_task_getscheduler(p);
5792 rq = task_rq_lock(p, &flags);
5793 time_slice = p->sched_class->get_rr_interval(rq, p);
5794 task_rq_unlock(rq, p, &flags);
5797 jiffies_to_timespec(time_slice, &t);
5798 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5806 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5808 void sched_show_task(struct task_struct *p)
5810 unsigned long free = 0;
5813 state = p->state ? __ffs(p->state) + 1 : 0;
5814 printk(KERN_INFO "%-15.15s %c", p->comm,
5815 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5816 #if BITS_PER_LONG == 32
5817 if (state == TASK_RUNNING)
5818 printk(KERN_CONT " running ");
5820 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5822 if (state == TASK_RUNNING)
5823 printk(KERN_CONT " running task ");
5825 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5827 #ifdef CONFIG_DEBUG_STACK_USAGE
5828 free = stack_not_used(p);
5830 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5831 task_pid_nr(p), task_pid_nr(p->real_parent),
5832 (unsigned long)task_thread_info(p)->flags);
5834 show_stack(p, NULL);
5837 void show_state_filter(unsigned long state_filter)
5839 struct task_struct *g, *p;
5841 #if BITS_PER_LONG == 32
5843 " task PC stack pid father\n");
5846 " task PC stack pid father\n");
5848 read_lock(&tasklist_lock);
5849 do_each_thread(g, p) {
5851 * reset the NMI-timeout, listing all files on a slow
5852 * console might take a lot of time:
5854 touch_nmi_watchdog();
5855 if (!state_filter || (p->state & state_filter))
5857 } while_each_thread(g, p);
5859 touch_all_softlockup_watchdogs();
5861 #ifdef CONFIG_SCHED_DEBUG
5862 sysrq_sched_debug_show();
5864 read_unlock(&tasklist_lock);
5866 * Only show locks if all tasks are dumped:
5869 debug_show_all_locks();
5872 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5874 idle->sched_class = &idle_sched_class;
5878 * init_idle - set up an idle thread for a given CPU
5879 * @idle: task in question
5880 * @cpu: cpu the idle task belongs to
5882 * NOTE: this function does not set the idle thread's NEED_RESCHED
5883 * flag, to make booting more robust.
5885 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5887 struct rq *rq = cpu_rq(cpu);
5888 unsigned long flags;
5890 raw_spin_lock_irqsave(&rq->lock, flags);
5893 idle->state = TASK_RUNNING;
5894 idle->se.exec_start = sched_clock();
5896 do_set_cpus_allowed(idle, cpumask_of(cpu));
5898 * We're having a chicken and egg problem, even though we are
5899 * holding rq->lock, the cpu isn't yet set to this cpu so the
5900 * lockdep check in task_group() will fail.
5902 * Similar case to sched_fork(). / Alternatively we could
5903 * use task_rq_lock() here and obtain the other rq->lock.
5908 __set_task_cpu(idle, cpu);
5911 rq->curr = rq->idle = idle;
5912 #if defined(CONFIG_SMP)
5915 raw_spin_unlock_irqrestore(&rq->lock, flags);
5917 /* Set the preempt count _outside_ the spinlocks! */
5918 task_thread_info(idle)->preempt_count = 0;
5921 * The idle tasks have their own, simple scheduling class:
5923 idle->sched_class = &idle_sched_class;
5924 ftrace_graph_init_idle_task(idle, cpu);
5928 * In a system that switches off the HZ timer nohz_cpu_mask
5929 * indicates which cpus entered this state. This is used
5930 * in the rcu update to wait only for active cpus. For system
5931 * which do not switch off the HZ timer nohz_cpu_mask should
5932 * always be CPU_BITS_NONE.
5934 cpumask_var_t nohz_cpu_mask;
5937 * Increase the granularity value when there are more CPUs,
5938 * because with more CPUs the 'effective latency' as visible
5939 * to users decreases. But the relationship is not linear,
5940 * so pick a second-best guess by going with the log2 of the
5943 * This idea comes from the SD scheduler of Con Kolivas:
5945 static int get_update_sysctl_factor(void)
5947 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5948 unsigned int factor;
5950 switch (sysctl_sched_tunable_scaling) {
5951 case SCHED_TUNABLESCALING_NONE:
5954 case SCHED_TUNABLESCALING_LINEAR:
5957 case SCHED_TUNABLESCALING_LOG:
5959 factor = 1 + ilog2(cpus);
5966 static void update_sysctl(void)
5968 unsigned int factor = get_update_sysctl_factor();
5970 #define SET_SYSCTL(name) \
5971 (sysctl_##name = (factor) * normalized_sysctl_##name)
5972 SET_SYSCTL(sched_min_granularity);
5973 SET_SYSCTL(sched_latency);
5974 SET_SYSCTL(sched_wakeup_granularity);
5978 static inline void sched_init_granularity(void)
5984 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
5986 if (p->sched_class && p->sched_class->set_cpus_allowed)
5987 p->sched_class->set_cpus_allowed(p, new_mask);
5989 cpumask_copy(&p->cpus_allowed, new_mask);
5990 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5995 * This is how migration works:
5997 * 1) we invoke migration_cpu_stop() on the target CPU using
5999 * 2) stopper starts to run (implicitly forcing the migrated thread
6001 * 3) it checks whether the migrated task is still in the wrong runqueue.
6002 * 4) if it's in the wrong runqueue then the migration thread removes
6003 * it and puts it into the right queue.
6004 * 5) stopper completes and stop_one_cpu() returns and the migration
6009 * Change a given task's CPU affinity. Migrate the thread to a
6010 * proper CPU and schedule it away if the CPU it's executing on
6011 * is removed from the allowed bitmask.
6013 * NOTE: the caller must have a valid reference to the task, the
6014 * task must not exit() & deallocate itself prematurely. The
6015 * call is not atomic; no spinlocks may be held.
6017 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6019 unsigned long flags;
6021 unsigned int dest_cpu;
6024 rq = task_rq_lock(p, &flags);
6026 if (cpumask_equal(&p->cpus_allowed, new_mask))
6029 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6034 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6039 do_set_cpus_allowed(p, new_mask);
6041 /* Can the task run on the task's current CPU? If so, we're done */
6042 if (cpumask_test_cpu(task_cpu(p), new_mask))
6045 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6047 struct migration_arg arg = { p, dest_cpu };
6048 /* Need help from migration thread: drop lock and wait. */
6049 task_rq_unlock(rq, p, &flags);
6050 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6051 tlb_migrate_finish(p->mm);
6055 task_rq_unlock(rq, p, &flags);
6059 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6062 * Move (not current) task off this cpu, onto dest cpu. We're doing
6063 * this because either it can't run here any more (set_cpus_allowed()
6064 * away from this CPU, or CPU going down), or because we're
6065 * attempting to rebalance this task on exec (sched_exec).
6067 * So we race with normal scheduler movements, but that's OK, as long
6068 * as the task is no longer on this CPU.
6070 * Returns non-zero if task was successfully migrated.
6072 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6074 struct rq *rq_dest, *rq_src;
6077 if (unlikely(!cpu_active(dest_cpu)))
6080 rq_src = cpu_rq(src_cpu);
6081 rq_dest = cpu_rq(dest_cpu);
6083 raw_spin_lock(&p->pi_lock);
6084 double_rq_lock(rq_src, rq_dest);
6085 /* Already moved. */
6086 if (task_cpu(p) != src_cpu)
6088 /* Affinity changed (again). */
6089 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6093 * If we're not on a rq, the next wake-up will ensure we're
6097 deactivate_task(rq_src, p, 0);
6098 set_task_cpu(p, dest_cpu);
6099 activate_task(rq_dest, p, 0);
6100 check_preempt_curr(rq_dest, p, 0);
6105 double_rq_unlock(rq_src, rq_dest);
6106 raw_spin_unlock(&p->pi_lock);
6111 * migration_cpu_stop - this will be executed by a highprio stopper thread
6112 * and performs thread migration by bumping thread off CPU then
6113 * 'pushing' onto another runqueue.
6115 static int migration_cpu_stop(void *data)
6117 struct migration_arg *arg = data;
6120 * The original target cpu might have gone down and we might
6121 * be on another cpu but it doesn't matter.
6123 local_irq_disable();
6124 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6129 #ifdef CONFIG_HOTPLUG_CPU
6132 * Ensures that the idle task is using init_mm right before its cpu goes
6135 void idle_task_exit(void)
6137 struct mm_struct *mm = current->active_mm;
6139 BUG_ON(cpu_online(smp_processor_id()));
6142 switch_mm(mm, &init_mm, current);
6147 * While a dead CPU has no uninterruptible tasks queued at this point,
6148 * it might still have a nonzero ->nr_uninterruptible counter, because
6149 * for performance reasons the counter is not stricly tracking tasks to
6150 * their home CPUs. So we just add the counter to another CPU's counter,
6151 * to keep the global sum constant after CPU-down:
6153 static void migrate_nr_uninterruptible(struct rq *rq_src)
6155 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6157 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6158 rq_src->nr_uninterruptible = 0;
6162 * remove the tasks which were accounted by rq from calc_load_tasks.
6164 static void calc_global_load_remove(struct rq *rq)
6166 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6167 rq->calc_load_active = 0;
6171 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6172 * try_to_wake_up()->select_task_rq().
6174 * Called with rq->lock held even though we'er in stop_machine() and
6175 * there's no concurrency possible, we hold the required locks anyway
6176 * because of lock validation efforts.
6178 static void migrate_tasks(unsigned int dead_cpu)
6180 struct rq *rq = cpu_rq(dead_cpu);
6181 struct task_struct *next, *stop = rq->stop;
6185 * Fudge the rq selection such that the below task selection loop
6186 * doesn't get stuck on the currently eligible stop task.
6188 * We're currently inside stop_machine() and the rq is either stuck
6189 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6190 * either way we should never end up calling schedule() until we're
6197 * There's this thread running, bail when that's the only
6200 if (rq->nr_running == 1)
6203 next = pick_next_task(rq);
6205 next->sched_class->put_prev_task(rq, next);
6207 /* Find suitable destination for @next, with force if needed. */
6208 dest_cpu = select_fallback_rq(dead_cpu, next);
6209 raw_spin_unlock(&rq->lock);
6211 __migrate_task(next, dead_cpu, dest_cpu);
6213 raw_spin_lock(&rq->lock);
6219 #endif /* CONFIG_HOTPLUG_CPU */
6221 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6223 static struct ctl_table sd_ctl_dir[] = {
6225 .procname = "sched_domain",
6231 static struct ctl_table sd_ctl_root[] = {
6233 .procname = "kernel",
6235 .child = sd_ctl_dir,
6240 static struct ctl_table *sd_alloc_ctl_entry(int n)
6242 struct ctl_table *entry =
6243 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6248 static void sd_free_ctl_entry(struct ctl_table **tablep)
6250 struct ctl_table *entry;
6253 * In the intermediate directories, both the child directory and
6254 * procname are dynamically allocated and could fail but the mode
6255 * will always be set. In the lowest directory the names are
6256 * static strings and all have proc handlers.
6258 for (entry = *tablep; entry->mode; entry++) {
6260 sd_free_ctl_entry(&entry->child);
6261 if (entry->proc_handler == NULL)
6262 kfree(entry->procname);
6270 set_table_entry(struct ctl_table *entry,
6271 const char *procname, void *data, int maxlen,
6272 mode_t mode, proc_handler *proc_handler)
6274 entry->procname = procname;
6276 entry->maxlen = maxlen;
6278 entry->proc_handler = proc_handler;
6281 static struct ctl_table *
6282 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6284 struct ctl_table *table = sd_alloc_ctl_entry(13);
6289 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6290 sizeof(long), 0644, proc_doulongvec_minmax);
6291 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6292 sizeof(long), 0644, proc_doulongvec_minmax);
6293 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6294 sizeof(int), 0644, proc_dointvec_minmax);
6295 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6296 sizeof(int), 0644, proc_dointvec_minmax);
6297 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6298 sizeof(int), 0644, proc_dointvec_minmax);
6299 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6300 sizeof(int), 0644, proc_dointvec_minmax);
6301 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6302 sizeof(int), 0644, proc_dointvec_minmax);
6303 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6304 sizeof(int), 0644, proc_dointvec_minmax);
6305 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6306 sizeof(int), 0644, proc_dointvec_minmax);
6307 set_table_entry(&table[9], "cache_nice_tries",
6308 &sd->cache_nice_tries,
6309 sizeof(int), 0644, proc_dointvec_minmax);
6310 set_table_entry(&table[10], "flags", &sd->flags,
6311 sizeof(int), 0644, proc_dointvec_minmax);
6312 set_table_entry(&table[11], "name", sd->name,
6313 CORENAME_MAX_SIZE, 0444, proc_dostring);
6314 /* &table[12] is terminator */
6319 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6321 struct ctl_table *entry, *table;
6322 struct sched_domain *sd;
6323 int domain_num = 0, i;
6326 for_each_domain(cpu, sd)
6328 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6333 for_each_domain(cpu, sd) {
6334 snprintf(buf, 32, "domain%d", i);
6335 entry->procname = kstrdup(buf, GFP_KERNEL);
6337 entry->child = sd_alloc_ctl_domain_table(sd);
6344 static struct ctl_table_header *sd_sysctl_header;
6345 static void register_sched_domain_sysctl(void)
6347 int i, cpu_num = num_possible_cpus();
6348 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6351 WARN_ON(sd_ctl_dir[0].child);
6352 sd_ctl_dir[0].child = entry;
6357 for_each_possible_cpu(i) {
6358 snprintf(buf, 32, "cpu%d", i);
6359 entry->procname = kstrdup(buf, GFP_KERNEL);
6361 entry->child = sd_alloc_ctl_cpu_table(i);
6365 WARN_ON(sd_sysctl_header);
6366 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6369 /* may be called multiple times per register */
6370 static void unregister_sched_domain_sysctl(void)
6372 if (sd_sysctl_header)
6373 unregister_sysctl_table(sd_sysctl_header);
6374 sd_sysctl_header = NULL;
6375 if (sd_ctl_dir[0].child)
6376 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6379 static void register_sched_domain_sysctl(void)
6382 static void unregister_sched_domain_sysctl(void)
6387 static void set_rq_online(struct rq *rq)
6390 const struct sched_class *class;
6392 cpumask_set_cpu(rq->cpu, rq->rd->online);
6395 for_each_class(class) {
6396 if (class->rq_online)
6397 class->rq_online(rq);
6402 static void set_rq_offline(struct rq *rq)
6405 const struct sched_class *class;
6407 for_each_class(class) {
6408 if (class->rq_offline)
6409 class->rq_offline(rq);
6412 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6418 * migration_call - callback that gets triggered when a CPU is added.
6419 * Here we can start up the necessary migration thread for the new CPU.
6421 static int __cpuinit
6422 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6424 int cpu = (long)hcpu;
6425 unsigned long flags;
6426 struct rq *rq = cpu_rq(cpu);
6428 switch (action & ~CPU_TASKS_FROZEN) {
6430 case CPU_UP_PREPARE:
6431 rq->calc_load_update = calc_load_update;
6435 /* Update our root-domain */
6436 raw_spin_lock_irqsave(&rq->lock, flags);
6438 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6442 raw_spin_unlock_irqrestore(&rq->lock, flags);
6445 #ifdef CONFIG_HOTPLUG_CPU
6447 sched_ttwu_pending();
6448 /* Update our root-domain */
6449 raw_spin_lock_irqsave(&rq->lock, flags);
6451 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6455 BUG_ON(rq->nr_running != 1); /* the migration thread */
6456 raw_spin_unlock_irqrestore(&rq->lock, flags);
6458 migrate_nr_uninterruptible(rq);
6459 calc_global_load_remove(rq);
6464 update_max_interval();
6470 * Register at high priority so that task migration (migrate_all_tasks)
6471 * happens before everything else. This has to be lower priority than
6472 * the notifier in the perf_event subsystem, though.
6474 static struct notifier_block __cpuinitdata migration_notifier = {
6475 .notifier_call = migration_call,
6476 .priority = CPU_PRI_MIGRATION,
6479 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6480 unsigned long action, void *hcpu)
6482 switch (action & ~CPU_TASKS_FROZEN) {
6484 case CPU_DOWN_FAILED:
6485 set_cpu_active((long)hcpu, true);
6492 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6493 unsigned long action, void *hcpu)
6495 switch (action & ~CPU_TASKS_FROZEN) {
6496 case CPU_DOWN_PREPARE:
6497 set_cpu_active((long)hcpu, false);
6504 static int __init migration_init(void)
6506 void *cpu = (void *)(long)smp_processor_id();
6509 /* Initialize migration for the boot CPU */
6510 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6511 BUG_ON(err == NOTIFY_BAD);
6512 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6513 register_cpu_notifier(&migration_notifier);
6515 /* Register cpu active notifiers */
6516 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6517 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6521 early_initcall(migration_init);
6526 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6528 #ifdef CONFIG_SCHED_DEBUG
6530 static __read_mostly int sched_domain_debug_enabled;
6532 static int __init sched_domain_debug_setup(char *str)
6534 sched_domain_debug_enabled = 1;
6538 early_param("sched_debug", sched_domain_debug_setup);
6540 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6541 struct cpumask *groupmask)
6543 struct sched_group *group = sd->groups;
6546 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6547 cpumask_clear(groupmask);
6549 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6551 if (!(sd->flags & SD_LOAD_BALANCE)) {
6552 printk("does not load-balance\n");
6554 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6559 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6561 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6562 printk(KERN_ERR "ERROR: domain->span does not contain "
6565 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6566 printk(KERN_ERR "ERROR: domain->groups does not contain"
6570 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6574 printk(KERN_ERR "ERROR: group is NULL\n");
6578 if (!group->sgp->power) {
6579 printk(KERN_CONT "\n");
6580 printk(KERN_ERR "ERROR: domain->cpu_power not "
6585 if (!cpumask_weight(sched_group_cpus(group))) {
6586 printk(KERN_CONT "\n");
6587 printk(KERN_ERR "ERROR: empty group\n");
6591 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6592 printk(KERN_CONT "\n");
6593 printk(KERN_ERR "ERROR: repeated CPUs\n");
6597 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6599 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6601 printk(KERN_CONT " %s", str);
6602 if (group->sgp->power != SCHED_POWER_SCALE) {
6603 printk(KERN_CONT " (cpu_power = %d)",
6607 group = group->next;
6608 } while (group != sd->groups);
6609 printk(KERN_CONT "\n");
6611 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6612 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6615 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6616 printk(KERN_ERR "ERROR: parent span is not a superset "
6617 "of domain->span\n");
6621 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6625 if (!sched_domain_debug_enabled)
6629 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6633 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6636 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6644 #else /* !CONFIG_SCHED_DEBUG */
6645 # define sched_domain_debug(sd, cpu) do { } while (0)
6646 #endif /* CONFIG_SCHED_DEBUG */
6648 static int sd_degenerate(struct sched_domain *sd)
6650 if (cpumask_weight(sched_domain_span(sd)) == 1)
6653 /* Following flags need at least 2 groups */
6654 if (sd->flags & (SD_LOAD_BALANCE |
6655 SD_BALANCE_NEWIDLE |
6659 SD_SHARE_PKG_RESOURCES)) {
6660 if (sd->groups != sd->groups->next)
6664 /* Following flags don't use groups */
6665 if (sd->flags & (SD_WAKE_AFFINE))
6672 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6674 unsigned long cflags = sd->flags, pflags = parent->flags;
6676 if (sd_degenerate(parent))
6679 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6682 /* Flags needing groups don't count if only 1 group in parent */
6683 if (parent->groups == parent->groups->next) {
6684 pflags &= ~(SD_LOAD_BALANCE |
6685 SD_BALANCE_NEWIDLE |
6689 SD_SHARE_PKG_RESOURCES);
6690 if (nr_node_ids == 1)
6691 pflags &= ~SD_SERIALIZE;
6693 if (~cflags & pflags)
6699 static void free_rootdomain(struct rcu_head *rcu)
6701 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6703 cpupri_cleanup(&rd->cpupri);
6704 free_cpumask_var(rd->rto_mask);
6705 free_cpumask_var(rd->online);
6706 free_cpumask_var(rd->span);
6710 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6712 struct root_domain *old_rd = NULL;
6713 unsigned long flags;
6715 raw_spin_lock_irqsave(&rq->lock, flags);
6720 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6723 cpumask_clear_cpu(rq->cpu, old_rd->span);
6726 * If we dont want to free the old_rt yet then
6727 * set old_rd to NULL to skip the freeing later
6730 if (!atomic_dec_and_test(&old_rd->refcount))
6734 atomic_inc(&rd->refcount);
6737 cpumask_set_cpu(rq->cpu, rd->span);
6738 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6741 raw_spin_unlock_irqrestore(&rq->lock, flags);
6744 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6747 static int init_rootdomain(struct root_domain *rd)
6749 memset(rd, 0, sizeof(*rd));
6751 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6753 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6755 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6758 if (cpupri_init(&rd->cpupri) != 0)
6763 free_cpumask_var(rd->rto_mask);
6765 free_cpumask_var(rd->online);
6767 free_cpumask_var(rd->span);
6772 static void init_defrootdomain(void)
6774 init_rootdomain(&def_root_domain);
6776 atomic_set(&def_root_domain.refcount, 1);
6779 static struct root_domain *alloc_rootdomain(void)
6781 struct root_domain *rd;
6783 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6787 if (init_rootdomain(rd) != 0) {
6795 static void free_sched_groups(struct sched_group *sg, int free_sgp)
6797 struct sched_group *tmp, *first;
6806 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
6811 } while (sg != first);
6814 static void free_sched_domain(struct rcu_head *rcu)
6816 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6819 * If its an overlapping domain it has private groups, iterate and
6822 if (sd->flags & SD_OVERLAP) {
6823 free_sched_groups(sd->groups, 1);
6824 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6825 kfree(sd->groups->sgp);
6831 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6833 call_rcu(&sd->rcu, free_sched_domain);
6836 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6838 for (; sd; sd = sd->parent)
6839 destroy_sched_domain(sd, cpu);
6843 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6844 * hold the hotplug lock.
6847 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6849 struct rq *rq = cpu_rq(cpu);
6850 struct sched_domain *tmp;
6852 /* Remove the sched domains which do not contribute to scheduling. */
6853 for (tmp = sd; tmp; ) {
6854 struct sched_domain *parent = tmp->parent;
6858 if (sd_parent_degenerate(tmp, parent)) {
6859 tmp->parent = parent->parent;
6861 parent->parent->child = tmp;
6862 destroy_sched_domain(parent, cpu);
6867 if (sd && sd_degenerate(sd)) {
6870 destroy_sched_domain(tmp, cpu);
6875 sched_domain_debug(sd, cpu);
6877 rq_attach_root(rq, rd);
6879 rcu_assign_pointer(rq->sd, sd);
6880 destroy_sched_domains(tmp, cpu);
6883 /* cpus with isolated domains */
6884 static cpumask_var_t cpu_isolated_map;
6886 /* Setup the mask of cpus configured for isolated domains */
6887 static int __init isolated_cpu_setup(char *str)
6889 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6890 cpulist_parse(str, cpu_isolated_map);
6894 __setup("isolcpus=", isolated_cpu_setup);
6896 #define SD_NODES_PER_DOMAIN 16
6901 * find_next_best_node - find the next node to include in a sched_domain
6902 * @node: node whose sched_domain we're building
6903 * @used_nodes: nodes already in the sched_domain
6905 * Find the next node to include in a given scheduling domain. Simply
6906 * finds the closest node not already in the @used_nodes map.
6908 * Should use nodemask_t.
6910 static int find_next_best_node(int node, nodemask_t *used_nodes)
6912 int i, n, val, min_val, best_node = -1;
6916 for (i = 0; i < nr_node_ids; i++) {
6917 /* Start at @node */
6918 n = (node + i) % nr_node_ids;
6920 if (!nr_cpus_node(n))
6923 /* Skip already used nodes */
6924 if (node_isset(n, *used_nodes))
6927 /* Simple min distance search */
6928 val = node_distance(node, n);
6930 if (val < min_val) {
6936 if (best_node != -1)
6937 node_set(best_node, *used_nodes);
6942 * sched_domain_node_span - get a cpumask for a node's sched_domain
6943 * @node: node whose cpumask we're constructing
6944 * @span: resulting cpumask
6946 * Given a node, construct a good cpumask for its sched_domain to span. It
6947 * should be one that prevents unnecessary balancing, but also spreads tasks
6950 static void sched_domain_node_span(int node, struct cpumask *span)
6952 nodemask_t used_nodes;
6955 cpumask_clear(span);
6956 nodes_clear(used_nodes);
6958 cpumask_or(span, span, cpumask_of_node(node));
6959 node_set(node, used_nodes);
6961 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6962 int next_node = find_next_best_node(node, &used_nodes);
6965 cpumask_or(span, span, cpumask_of_node(next_node));
6969 static const struct cpumask *cpu_node_mask(int cpu)
6971 lockdep_assert_held(&sched_domains_mutex);
6973 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
6975 return sched_domains_tmpmask;
6978 static const struct cpumask *cpu_allnodes_mask(int cpu)
6980 return cpu_possible_mask;
6982 #endif /* CONFIG_NUMA */
6984 static const struct cpumask *cpu_cpu_mask(int cpu)
6986 return cpumask_of_node(cpu_to_node(cpu));
6989 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6992 struct sched_domain **__percpu sd;
6993 struct sched_group **__percpu sg;
6994 struct sched_group_power **__percpu sgp;
6998 struct sched_domain ** __percpu sd;
6999 struct root_domain *rd;
7009 struct sched_domain_topology_level;
7011 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
7012 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
7014 #define SDTL_OVERLAP 0x01
7016 struct sched_domain_topology_level {
7017 sched_domain_init_f init;
7018 sched_domain_mask_f mask;
7020 struct sd_data data;
7024 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
7026 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
7027 const struct cpumask *span = sched_domain_span(sd);
7028 struct cpumask *covered = sched_domains_tmpmask;
7029 struct sd_data *sdd = sd->private;
7030 struct sched_domain *child;
7033 cpumask_clear(covered);
7035 for_each_cpu(i, span) {
7036 struct cpumask *sg_span;
7038 if (cpumask_test_cpu(i, covered))
7041 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7042 GFP_KERNEL, cpu_to_node(i));
7047 sg_span = sched_group_cpus(sg);
7049 child = *per_cpu_ptr(sdd->sd, i);
7051 child = child->child;
7052 cpumask_copy(sg_span, sched_domain_span(child));
7054 cpumask_set_cpu(i, sg_span);
7056 cpumask_or(covered, covered, sg_span);
7058 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
7059 atomic_inc(&sg->sgp->ref);
7061 if (cpumask_test_cpu(cpu, sg_span))
7071 sd->groups = groups;
7076 free_sched_groups(first, 0);
7081 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
7083 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
7084 struct sched_domain *child = sd->child;
7087 cpu = cpumask_first(sched_domain_span(child));
7090 *sg = *per_cpu_ptr(sdd->sg, cpu);
7091 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
7092 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
7099 * build_sched_groups will build a circular linked list of the groups
7100 * covered by the given span, and will set each group's ->cpumask correctly,
7101 * and ->cpu_power to 0.
7103 * Assumes the sched_domain tree is fully constructed
7106 build_sched_groups(struct sched_domain *sd, int cpu)
7108 struct sched_group *first = NULL, *last = NULL;
7109 struct sd_data *sdd = sd->private;
7110 const struct cpumask *span = sched_domain_span(sd);
7111 struct cpumask *covered;
7114 get_group(cpu, sdd, &sd->groups);
7115 atomic_inc(&sd->groups->ref);
7117 if (cpu != cpumask_first(sched_domain_span(sd)))
7120 lockdep_assert_held(&sched_domains_mutex);
7121 covered = sched_domains_tmpmask;
7123 cpumask_clear(covered);
7125 for_each_cpu(i, span) {
7126 struct sched_group *sg;
7127 int group = get_group(i, sdd, &sg);
7130 if (cpumask_test_cpu(i, covered))
7133 cpumask_clear(sched_group_cpus(sg));
7136 for_each_cpu(j, span) {
7137 if (get_group(j, sdd, NULL) != group)
7140 cpumask_set_cpu(j, covered);
7141 cpumask_set_cpu(j, sched_group_cpus(sg));
7156 * Initialize sched groups cpu_power.
7158 * cpu_power indicates the capacity of sched group, which is used while
7159 * distributing the load between different sched groups in a sched domain.
7160 * Typically cpu_power for all the groups in a sched domain will be same unless
7161 * there are asymmetries in the topology. If there are asymmetries, group
7162 * having more cpu_power will pickup more load compared to the group having
7165 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7167 struct sched_group *sg = sd->groups;
7169 WARN_ON(!sd || !sg);
7172 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
7174 } while (sg != sd->groups);
7176 if (cpu != group_first_cpu(sg))
7179 update_group_power(sd, cpu);
7183 * Initializers for schedule domains
7184 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7187 #ifdef CONFIG_SCHED_DEBUG
7188 # define SD_INIT_NAME(sd, type) sd->name = #type
7190 # define SD_INIT_NAME(sd, type) do { } while (0)
7193 #define SD_INIT_FUNC(type) \
7194 static noinline struct sched_domain * \
7195 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7197 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7198 *sd = SD_##type##_INIT; \
7199 SD_INIT_NAME(sd, type); \
7200 sd->private = &tl->data; \
7206 SD_INIT_FUNC(ALLNODES)
7209 #ifdef CONFIG_SCHED_SMT
7210 SD_INIT_FUNC(SIBLING)
7212 #ifdef CONFIG_SCHED_MC
7215 #ifdef CONFIG_SCHED_BOOK
7219 static int default_relax_domain_level = -1;
7220 int sched_domain_level_max;
7222 static int __init setup_relax_domain_level(char *str)
7224 if (kstrtoint(str, 0, &default_relax_domain_level))
7225 pr_warn("Unable to set relax_domain_level\n");
7229 __setup("relax_domain_level=", setup_relax_domain_level);
7231 static void set_domain_attribute(struct sched_domain *sd,
7232 struct sched_domain_attr *attr)
7236 if (!attr || attr->relax_domain_level < 0) {
7237 if (default_relax_domain_level < 0)
7240 request = default_relax_domain_level;
7242 request = attr->relax_domain_level;
7243 if (request < sd->level) {
7244 /* turn off idle balance on this domain */
7245 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7247 /* turn on idle balance on this domain */
7248 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7252 static void __sdt_free(const struct cpumask *cpu_map);
7253 static int __sdt_alloc(const struct cpumask *cpu_map);
7255 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7256 const struct cpumask *cpu_map)
7260 if (!atomic_read(&d->rd->refcount))
7261 free_rootdomain(&d->rd->rcu); /* fall through */
7263 free_percpu(d->sd); /* fall through */
7265 __sdt_free(cpu_map); /* fall through */
7271 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7272 const struct cpumask *cpu_map)
7274 memset(d, 0, sizeof(*d));
7276 if (__sdt_alloc(cpu_map))
7277 return sa_sd_storage;
7278 d->sd = alloc_percpu(struct sched_domain *);
7280 return sa_sd_storage;
7281 d->rd = alloc_rootdomain();
7284 return sa_rootdomain;
7288 * NULL the sd_data elements we've used to build the sched_domain and
7289 * sched_group structure so that the subsequent __free_domain_allocs()
7290 * will not free the data we're using.
7292 static void claim_allocations(int cpu, struct sched_domain *sd)
7294 struct sd_data *sdd = sd->private;
7296 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7297 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7299 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
7300 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7302 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
7303 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
7306 #ifdef CONFIG_SCHED_SMT
7307 static const struct cpumask *cpu_smt_mask(int cpu)
7309 return topology_thread_cpumask(cpu);
7314 * Topology list, bottom-up.
7316 static struct sched_domain_topology_level default_topology[] = {
7317 #ifdef CONFIG_SCHED_SMT
7318 { sd_init_SIBLING, cpu_smt_mask, },
7320 #ifdef CONFIG_SCHED_MC
7321 { sd_init_MC, cpu_coregroup_mask, },
7323 #ifdef CONFIG_SCHED_BOOK
7324 { sd_init_BOOK, cpu_book_mask, },
7326 { sd_init_CPU, cpu_cpu_mask, },
7328 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
7329 { sd_init_ALLNODES, cpu_allnodes_mask, },
7334 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7336 static int __sdt_alloc(const struct cpumask *cpu_map)
7338 struct sched_domain_topology_level *tl;
7341 for (tl = sched_domain_topology; tl->init; tl++) {
7342 struct sd_data *sdd = &tl->data;
7344 sdd->sd = alloc_percpu(struct sched_domain *);
7348 sdd->sg = alloc_percpu(struct sched_group *);
7352 sdd->sgp = alloc_percpu(struct sched_group_power *);
7356 for_each_cpu(j, cpu_map) {
7357 struct sched_domain *sd;
7358 struct sched_group *sg;
7359 struct sched_group_power *sgp;
7361 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7362 GFP_KERNEL, cpu_to_node(j));
7366 *per_cpu_ptr(sdd->sd, j) = sd;
7368 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7369 GFP_KERNEL, cpu_to_node(j));
7373 *per_cpu_ptr(sdd->sg, j) = sg;
7375 sgp = kzalloc_node(sizeof(struct sched_group_power),
7376 GFP_KERNEL, cpu_to_node(j));
7380 *per_cpu_ptr(sdd->sgp, j) = sgp;
7387 static void __sdt_free(const struct cpumask *cpu_map)
7389 struct sched_domain_topology_level *tl;
7392 for (tl = sched_domain_topology; tl->init; tl++) {
7393 struct sd_data *sdd = &tl->data;
7395 for_each_cpu(j, cpu_map) {
7396 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
7397 if (sd && (sd->flags & SD_OVERLAP))
7398 free_sched_groups(sd->groups, 0);
7399 kfree(*per_cpu_ptr(sdd->sd, j));
7400 kfree(*per_cpu_ptr(sdd->sg, j));
7401 kfree(*per_cpu_ptr(sdd->sgp, j));
7403 free_percpu(sdd->sd);
7404 free_percpu(sdd->sg);
7405 free_percpu(sdd->sgp);
7409 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7410 struct s_data *d, const struct cpumask *cpu_map,
7411 struct sched_domain_attr *attr, struct sched_domain *child,
7414 struct sched_domain *sd = tl->init(tl, cpu);
7418 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7420 sd->level = child->level + 1;
7421 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7425 set_domain_attribute(sd, attr);
7431 * Build sched domains for a given set of cpus and attach the sched domains
7432 * to the individual cpus
7434 static int build_sched_domains(const struct cpumask *cpu_map,
7435 struct sched_domain_attr *attr)
7437 enum s_alloc alloc_state = sa_none;
7438 struct sched_domain *sd;
7440 int i, ret = -ENOMEM;
7442 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7443 if (alloc_state != sa_rootdomain)
7446 /* Set up domains for cpus specified by the cpu_map. */
7447 for_each_cpu(i, cpu_map) {
7448 struct sched_domain_topology_level *tl;
7451 for (tl = sched_domain_topology; tl->init; tl++) {
7452 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7453 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7454 sd->flags |= SD_OVERLAP;
7455 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7462 *per_cpu_ptr(d.sd, i) = sd;
7465 /* Build the groups for the domains */
7466 for_each_cpu(i, cpu_map) {
7467 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7468 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7469 if (sd->flags & SD_OVERLAP) {
7470 if (build_overlap_sched_groups(sd, i))
7473 if (build_sched_groups(sd, i))
7479 /* Calculate CPU power for physical packages and nodes */
7480 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7481 if (!cpumask_test_cpu(i, cpu_map))
7484 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7485 claim_allocations(i, sd);
7486 init_sched_groups_power(i, sd);
7490 /* Attach the domains */
7492 for_each_cpu(i, cpu_map) {
7493 sd = *per_cpu_ptr(d.sd, i);
7494 cpu_attach_domain(sd, d.rd, i);
7500 __free_domain_allocs(&d, alloc_state, cpu_map);
7504 static cpumask_var_t *doms_cur; /* current sched domains */
7505 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7506 static struct sched_domain_attr *dattr_cur;
7507 /* attribues of custom domains in 'doms_cur' */
7510 * Special case: If a kmalloc of a doms_cur partition (array of
7511 * cpumask) fails, then fallback to a single sched domain,
7512 * as determined by the single cpumask fallback_doms.
7514 static cpumask_var_t fallback_doms;
7517 * arch_update_cpu_topology lets virtualized architectures update the
7518 * cpu core maps. It is supposed to return 1 if the topology changed
7519 * or 0 if it stayed the same.
7521 int __attribute__((weak)) arch_update_cpu_topology(void)
7526 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7529 cpumask_var_t *doms;
7531 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7534 for (i = 0; i < ndoms; i++) {
7535 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7536 free_sched_domains(doms, i);
7543 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7546 for (i = 0; i < ndoms; i++)
7547 free_cpumask_var(doms[i]);
7552 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7553 * For now this just excludes isolated cpus, but could be used to
7554 * exclude other special cases in the future.
7556 static int init_sched_domains(const struct cpumask *cpu_map)
7560 arch_update_cpu_topology();
7562 doms_cur = alloc_sched_domains(ndoms_cur);
7564 doms_cur = &fallback_doms;
7565 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7567 err = build_sched_domains(doms_cur[0], NULL);
7568 register_sched_domain_sysctl();
7574 * Detach sched domains from a group of cpus specified in cpu_map
7575 * These cpus will now be attached to the NULL domain
7577 static void detach_destroy_domains(const struct cpumask *cpu_map)
7582 for_each_cpu(i, cpu_map)
7583 cpu_attach_domain(NULL, &def_root_domain, i);
7587 /* handle null as "default" */
7588 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7589 struct sched_domain_attr *new, int idx_new)
7591 struct sched_domain_attr tmp;
7598 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7599 new ? (new + idx_new) : &tmp,
7600 sizeof(struct sched_domain_attr));
7604 * Partition sched domains as specified by the 'ndoms_new'
7605 * cpumasks in the array doms_new[] of cpumasks. This compares
7606 * doms_new[] to the current sched domain partitioning, doms_cur[].
7607 * It destroys each deleted domain and builds each new domain.
7609 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7610 * The masks don't intersect (don't overlap.) We should setup one
7611 * sched domain for each mask. CPUs not in any of the cpumasks will
7612 * not be load balanced. If the same cpumask appears both in the
7613 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7616 * The passed in 'doms_new' should be allocated using
7617 * alloc_sched_domains. This routine takes ownership of it and will
7618 * free_sched_domains it when done with it. If the caller failed the
7619 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7620 * and partition_sched_domains() will fallback to the single partition
7621 * 'fallback_doms', it also forces the domains to be rebuilt.
7623 * If doms_new == NULL it will be replaced with cpu_online_mask.
7624 * ndoms_new == 0 is a special case for destroying existing domains,
7625 * and it will not create the default domain.
7627 * Call with hotplug lock held
7629 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7630 struct sched_domain_attr *dattr_new)
7635 mutex_lock(&sched_domains_mutex);
7637 /* always unregister in case we don't destroy any domains */
7638 unregister_sched_domain_sysctl();
7640 /* Let architecture update cpu core mappings. */
7641 new_topology = arch_update_cpu_topology();
7643 n = doms_new ? ndoms_new : 0;
7645 /* Destroy deleted domains */
7646 for (i = 0; i < ndoms_cur; i++) {
7647 for (j = 0; j < n && !new_topology; j++) {
7648 if (cpumask_equal(doms_cur[i], doms_new[j])
7649 && dattrs_equal(dattr_cur, i, dattr_new, j))
7652 /* no match - a current sched domain not in new doms_new[] */
7653 detach_destroy_domains(doms_cur[i]);
7658 if (doms_new == NULL) {
7660 doms_new = &fallback_doms;
7661 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7662 WARN_ON_ONCE(dattr_new);
7665 /* Build new domains */
7666 for (i = 0; i < ndoms_new; i++) {
7667 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7668 if (cpumask_equal(doms_new[i], doms_cur[j])
7669 && dattrs_equal(dattr_new, i, dattr_cur, j))
7672 /* no match - add a new doms_new */
7673 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7678 /* Remember the new sched domains */
7679 if (doms_cur != &fallback_doms)
7680 free_sched_domains(doms_cur, ndoms_cur);
7681 kfree(dattr_cur); /* kfree(NULL) is safe */
7682 doms_cur = doms_new;
7683 dattr_cur = dattr_new;
7684 ndoms_cur = ndoms_new;
7686 register_sched_domain_sysctl();
7688 mutex_unlock(&sched_domains_mutex);
7691 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7692 static void reinit_sched_domains(void)
7696 /* Destroy domains first to force the rebuild */
7697 partition_sched_domains(0, NULL, NULL);
7699 rebuild_sched_domains();
7703 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7705 unsigned int level = 0;
7707 if (sscanf(buf, "%u", &level) != 1)
7711 * level is always be positive so don't check for
7712 * level < POWERSAVINGS_BALANCE_NONE which is 0
7713 * What happens on 0 or 1 byte write,
7714 * need to check for count as well?
7717 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7721 sched_smt_power_savings = level;
7723 sched_mc_power_savings = level;
7725 reinit_sched_domains();
7730 #ifdef CONFIG_SCHED_MC
7731 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7732 struct sysdev_class_attribute *attr,
7735 return sprintf(page, "%u\n", sched_mc_power_savings);
7737 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7738 struct sysdev_class_attribute *attr,
7739 const char *buf, size_t count)
7741 return sched_power_savings_store(buf, count, 0);
7743 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7744 sched_mc_power_savings_show,
7745 sched_mc_power_savings_store);
7748 #ifdef CONFIG_SCHED_SMT
7749 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7750 struct sysdev_class_attribute *attr,
7753 return sprintf(page, "%u\n", sched_smt_power_savings);
7755 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7756 struct sysdev_class_attribute *attr,
7757 const char *buf, size_t count)
7759 return sched_power_savings_store(buf, count, 1);
7761 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7762 sched_smt_power_savings_show,
7763 sched_smt_power_savings_store);
7766 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7770 #ifdef CONFIG_SCHED_SMT
7772 err = sysfs_create_file(&cls->kset.kobj,
7773 &attr_sched_smt_power_savings.attr);
7775 #ifdef CONFIG_SCHED_MC
7776 if (!err && mc_capable())
7777 err = sysfs_create_file(&cls->kset.kobj,
7778 &attr_sched_mc_power_savings.attr);
7782 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7785 * Update cpusets according to cpu_active mask. If cpusets are
7786 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7787 * around partition_sched_domains().
7789 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7792 switch (action & ~CPU_TASKS_FROZEN) {
7794 case CPU_DOWN_FAILED:
7795 cpuset_update_active_cpus();
7802 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7805 switch (action & ~CPU_TASKS_FROZEN) {
7806 case CPU_DOWN_PREPARE:
7807 cpuset_update_active_cpus();
7814 static int update_runtime(struct notifier_block *nfb,
7815 unsigned long action, void *hcpu)
7817 int cpu = (int)(long)hcpu;
7820 case CPU_DOWN_PREPARE:
7821 case CPU_DOWN_PREPARE_FROZEN:
7822 disable_runtime(cpu_rq(cpu));
7825 case CPU_DOWN_FAILED:
7826 case CPU_DOWN_FAILED_FROZEN:
7828 case CPU_ONLINE_FROZEN:
7829 enable_runtime(cpu_rq(cpu));
7837 void __init sched_init_smp(void)
7839 cpumask_var_t non_isolated_cpus;
7841 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7842 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7845 mutex_lock(&sched_domains_mutex);
7846 init_sched_domains(cpu_active_mask);
7847 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7848 if (cpumask_empty(non_isolated_cpus))
7849 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7850 mutex_unlock(&sched_domains_mutex);
7853 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7854 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7856 /* RT runtime code needs to handle some hotplug events */
7857 hotcpu_notifier(update_runtime, 0);
7861 /* Move init over to a non-isolated CPU */
7862 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7864 sched_init_granularity();
7865 free_cpumask_var(non_isolated_cpus);
7867 init_sched_rt_class();
7870 void __init sched_init_smp(void)
7872 sched_init_granularity();
7874 #endif /* CONFIG_SMP */
7876 const_debug unsigned int sysctl_timer_migration = 1;
7878 int in_sched_functions(unsigned long addr)
7880 return in_lock_functions(addr) ||
7881 (addr >= (unsigned long)__sched_text_start
7882 && addr < (unsigned long)__sched_text_end);
7885 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7887 cfs_rq->tasks_timeline = RB_ROOT;
7888 INIT_LIST_HEAD(&cfs_rq->tasks);
7889 #ifdef CONFIG_FAIR_GROUP_SCHED
7891 /* allow initial update_cfs_load() to truncate */
7893 cfs_rq->load_stamp = 1;
7896 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7897 #ifndef CONFIG_64BIT
7898 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7902 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7904 struct rt_prio_array *array;
7907 array = &rt_rq->active;
7908 for (i = 0; i < MAX_RT_PRIO; i++) {
7909 INIT_LIST_HEAD(array->queue + i);
7910 __clear_bit(i, array->bitmap);
7912 /* delimiter for bitsearch: */
7913 __set_bit(MAX_RT_PRIO, array->bitmap);
7915 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7916 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7918 rt_rq->highest_prio.next = MAX_RT_PRIO;
7922 rt_rq->rt_nr_migratory = 0;
7923 rt_rq->overloaded = 0;
7924 plist_head_init(&rt_rq->pushable_tasks);
7928 rt_rq->rt_throttled = 0;
7929 rt_rq->rt_runtime = 0;
7930 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7932 #ifdef CONFIG_RT_GROUP_SCHED
7933 rt_rq->rt_nr_boosted = 0;
7938 #ifdef CONFIG_FAIR_GROUP_SCHED
7939 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7940 struct sched_entity *se, int cpu,
7941 struct sched_entity *parent)
7943 struct rq *rq = cpu_rq(cpu);
7944 tg->cfs_rq[cpu] = cfs_rq;
7945 init_cfs_rq(cfs_rq, rq);
7949 /* se could be NULL for root_task_group */
7954 se->cfs_rq = &rq->cfs;
7956 se->cfs_rq = parent->my_q;
7959 update_load_set(&se->load, 0);
7960 se->parent = parent;
7964 #ifdef CONFIG_RT_GROUP_SCHED
7965 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7966 struct sched_rt_entity *rt_se, int cpu,
7967 struct sched_rt_entity *parent)
7969 struct rq *rq = cpu_rq(cpu);
7971 tg->rt_rq[cpu] = rt_rq;
7972 init_rt_rq(rt_rq, rq);
7974 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7976 tg->rt_se[cpu] = rt_se;
7981 rt_se->rt_rq = &rq->rt;
7983 rt_se->rt_rq = parent->my_q;
7985 rt_se->my_q = rt_rq;
7986 rt_se->parent = parent;
7987 INIT_LIST_HEAD(&rt_se->run_list);
7991 void __init sched_init(void)
7994 unsigned long alloc_size = 0, ptr;
7996 #ifdef CONFIG_FAIR_GROUP_SCHED
7997 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7999 #ifdef CONFIG_RT_GROUP_SCHED
8000 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8002 #ifdef CONFIG_CPUMASK_OFFSTACK
8003 alloc_size += num_possible_cpus() * cpumask_size();
8006 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8008 #ifdef CONFIG_FAIR_GROUP_SCHED
8009 root_task_group.se = (struct sched_entity **)ptr;
8010 ptr += nr_cpu_ids * sizeof(void **);
8012 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8013 ptr += nr_cpu_ids * sizeof(void **);
8015 #endif /* CONFIG_FAIR_GROUP_SCHED */
8016 #ifdef CONFIG_RT_GROUP_SCHED
8017 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8018 ptr += nr_cpu_ids * sizeof(void **);
8020 root_task_group.rt_rq = (struct rt_rq **)ptr;
8021 ptr += nr_cpu_ids * sizeof(void **);
8023 #endif /* CONFIG_RT_GROUP_SCHED */
8024 #ifdef CONFIG_CPUMASK_OFFSTACK
8025 for_each_possible_cpu(i) {
8026 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8027 ptr += cpumask_size();
8029 #endif /* CONFIG_CPUMASK_OFFSTACK */
8033 init_defrootdomain();
8036 init_rt_bandwidth(&def_rt_bandwidth,
8037 global_rt_period(), global_rt_runtime());
8039 #ifdef CONFIG_RT_GROUP_SCHED
8040 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8041 global_rt_period(), global_rt_runtime());
8042 #endif /* CONFIG_RT_GROUP_SCHED */
8044 #ifdef CONFIG_CGROUP_SCHED
8045 list_add(&root_task_group.list, &task_groups);
8046 INIT_LIST_HEAD(&root_task_group.children);
8047 autogroup_init(&init_task);
8048 #endif /* CONFIG_CGROUP_SCHED */
8050 for_each_possible_cpu(i) {
8054 raw_spin_lock_init(&rq->lock);
8056 rq->calc_load_active = 0;
8057 rq->calc_load_update = jiffies + LOAD_FREQ;
8058 init_cfs_rq(&rq->cfs, rq);
8059 init_rt_rq(&rq->rt, rq);
8060 #ifdef CONFIG_FAIR_GROUP_SCHED
8061 root_task_group.shares = root_task_group_load;
8062 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8064 * How much cpu bandwidth does root_task_group get?
8066 * In case of task-groups formed thr' the cgroup filesystem, it
8067 * gets 100% of the cpu resources in the system. This overall
8068 * system cpu resource is divided among the tasks of
8069 * root_task_group and its child task-groups in a fair manner,
8070 * based on each entity's (task or task-group's) weight
8071 * (se->load.weight).
8073 * In other words, if root_task_group has 10 tasks of weight
8074 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8075 * then A0's share of the cpu resource is:
8077 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8079 * We achieve this by letting root_task_group's tasks sit
8080 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8082 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8083 #endif /* CONFIG_FAIR_GROUP_SCHED */
8085 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8086 #ifdef CONFIG_RT_GROUP_SCHED
8087 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8088 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8091 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8092 rq->cpu_load[j] = 0;
8094 rq->last_load_update_tick = jiffies;
8099 rq->cpu_power = SCHED_POWER_SCALE;
8100 rq->post_schedule = 0;
8101 rq->active_balance = 0;
8102 rq->next_balance = jiffies;
8107 rq->avg_idle = 2*sysctl_sched_migration_cost;
8108 rq_attach_root(rq, &def_root_domain);
8110 rq->nohz_balance_kick = 0;
8111 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8115 atomic_set(&rq->nr_iowait, 0);
8118 set_load_weight(&init_task);
8120 #ifdef CONFIG_PREEMPT_NOTIFIERS
8121 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8125 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8128 #ifdef CONFIG_RT_MUTEXES
8129 plist_head_init(&init_task.pi_waiters);
8133 * The boot idle thread does lazy MMU switching as well:
8135 atomic_inc(&init_mm.mm_count);
8136 enter_lazy_tlb(&init_mm, current);
8139 * Make us the idle thread. Technically, schedule() should not be
8140 * called from this thread, however somewhere below it might be,
8141 * but because we are the idle thread, we just pick up running again
8142 * when this runqueue becomes "idle".
8144 init_idle(current, smp_processor_id());
8146 calc_load_update = jiffies + LOAD_FREQ;
8149 * During early bootup we pretend to be a normal task:
8151 current->sched_class = &fair_sched_class;
8153 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8154 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8156 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8158 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8159 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8160 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8161 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8162 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8164 /* May be allocated at isolcpus cmdline parse time */
8165 if (cpu_isolated_map == NULL)
8166 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8169 scheduler_running = 1;
8172 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8173 static inline int preempt_count_equals(int preempt_offset)
8175 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8177 return (nested == preempt_offset);
8180 static int __might_sleep_init_called;
8181 int __init __might_sleep_init(void)
8183 __might_sleep_init_called = 1;
8186 early_initcall(__might_sleep_init);
8188 void __might_sleep(const char *file, int line, int preempt_offset)
8191 static unsigned long prev_jiffy; /* ratelimiting */
8193 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8196 if (system_state != SYSTEM_RUNNING &&
8197 (!__might_sleep_init_called || system_state != SYSTEM_BOOTING))
8199 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8201 prev_jiffy = jiffies;
8204 "BUG: sleeping function called from invalid context at %s:%d\n",
8207 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8208 in_atomic(), irqs_disabled(),
8209 current->pid, current->comm);
8211 debug_show_held_locks(current);
8212 if (irqs_disabled())
8213 print_irqtrace_events(current);
8217 EXPORT_SYMBOL(__might_sleep);
8220 #ifdef CONFIG_MAGIC_SYSRQ
8221 static void normalize_task(struct rq *rq, struct task_struct *p)
8223 const struct sched_class *prev_class = p->sched_class;
8224 int old_prio = p->prio;
8229 deactivate_task(rq, p, 0);
8230 __setscheduler(rq, p, SCHED_NORMAL, 0);
8232 activate_task(rq, p, 0);
8233 resched_task(rq->curr);
8236 check_class_changed(rq, p, prev_class, old_prio);
8239 void normalize_rt_tasks(void)
8241 struct task_struct *g, *p;
8242 unsigned long flags;
8245 read_lock_irqsave(&tasklist_lock, flags);
8246 do_each_thread(g, p) {
8248 * Only normalize user tasks:
8253 p->se.exec_start = 0;
8254 #ifdef CONFIG_SCHEDSTATS
8255 p->se.statistics.wait_start = 0;
8256 p->se.statistics.sleep_start = 0;
8257 p->se.statistics.block_start = 0;
8262 * Renice negative nice level userspace
8265 if (TASK_NICE(p) < 0 && p->mm)
8266 set_user_nice(p, 0);
8270 raw_spin_lock(&p->pi_lock);
8271 rq = __task_rq_lock(p);
8273 normalize_task(rq, p);
8275 __task_rq_unlock(rq);
8276 raw_spin_unlock(&p->pi_lock);
8277 } while_each_thread(g, p);
8279 read_unlock_irqrestore(&tasklist_lock, flags);
8282 #endif /* CONFIG_MAGIC_SYSRQ */
8284 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8286 * These functions are only useful for the IA64 MCA handling, or kdb.
8288 * They can only be called when the whole system has been
8289 * stopped - every CPU needs to be quiescent, and no scheduling
8290 * activity can take place. Using them for anything else would
8291 * be a serious bug, and as a result, they aren't even visible
8292 * under any other configuration.
8296 * curr_task - return the current task for a given cpu.
8297 * @cpu: the processor in question.
8299 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8301 struct task_struct *curr_task(int cpu)
8303 return cpu_curr(cpu);
8306 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8310 * set_curr_task - set the current task for a given cpu.
8311 * @cpu: the processor in question.
8312 * @p: the task pointer to set.
8314 * Description: This function must only be used when non-maskable interrupts
8315 * are serviced on a separate stack. It allows the architecture to switch the
8316 * notion of the current task on a cpu in a non-blocking manner. This function
8317 * must be called with all CPU's synchronized, and interrupts disabled, the
8318 * and caller must save the original value of the current task (see
8319 * curr_task() above) and restore that value before reenabling interrupts and
8320 * re-starting the system.
8322 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8324 void set_curr_task(int cpu, struct task_struct *p)
8331 #ifdef CONFIG_FAIR_GROUP_SCHED
8332 static void free_fair_sched_group(struct task_group *tg)
8336 for_each_possible_cpu(i) {
8338 kfree(tg->cfs_rq[i]);
8348 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8350 struct cfs_rq *cfs_rq;
8351 struct sched_entity *se;
8354 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8357 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8361 tg->shares = NICE_0_LOAD;
8363 for_each_possible_cpu(i) {
8364 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8365 GFP_KERNEL, cpu_to_node(i));
8369 se = kzalloc_node(sizeof(struct sched_entity),
8370 GFP_KERNEL, cpu_to_node(i));
8374 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8385 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8387 struct rq *rq = cpu_rq(cpu);
8388 unsigned long flags;
8391 * Only empty task groups can be destroyed; so we can speculatively
8392 * check on_list without danger of it being re-added.
8394 if (!tg->cfs_rq[cpu]->on_list)
8397 raw_spin_lock_irqsave(&rq->lock, flags);
8398 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8399 raw_spin_unlock_irqrestore(&rq->lock, flags);
8401 #else /* !CONFG_FAIR_GROUP_SCHED */
8402 static inline void free_fair_sched_group(struct task_group *tg)
8407 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8412 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8415 #endif /* CONFIG_FAIR_GROUP_SCHED */
8417 #ifdef CONFIG_RT_GROUP_SCHED
8418 static void free_rt_sched_group(struct task_group *tg)
8422 destroy_rt_bandwidth(&tg->rt_bandwidth);
8424 for_each_possible_cpu(i) {
8426 kfree(tg->rt_rq[i]);
8428 kfree(tg->rt_se[i]);
8436 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8438 struct rt_rq *rt_rq;
8439 struct sched_rt_entity *rt_se;
8442 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8445 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8449 init_rt_bandwidth(&tg->rt_bandwidth,
8450 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8452 for_each_possible_cpu(i) {
8453 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8454 GFP_KERNEL, cpu_to_node(i));
8458 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8459 GFP_KERNEL, cpu_to_node(i));
8463 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8473 #else /* !CONFIG_RT_GROUP_SCHED */
8474 static inline void free_rt_sched_group(struct task_group *tg)
8479 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8483 #endif /* CONFIG_RT_GROUP_SCHED */
8485 #ifdef CONFIG_CGROUP_SCHED
8486 static void free_sched_group(struct task_group *tg)
8488 free_fair_sched_group(tg);
8489 free_rt_sched_group(tg);
8494 /* allocate runqueue etc for a new task group */
8495 struct task_group *sched_create_group(struct task_group *parent)
8497 struct task_group *tg;
8498 unsigned long flags;
8500 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8502 return ERR_PTR(-ENOMEM);
8504 if (!alloc_fair_sched_group(tg, parent))
8507 if (!alloc_rt_sched_group(tg, parent))
8510 spin_lock_irqsave(&task_group_lock, flags);
8511 list_add_rcu(&tg->list, &task_groups);
8513 WARN_ON(!parent); /* root should already exist */
8515 tg->parent = parent;
8516 INIT_LIST_HEAD(&tg->children);
8517 list_add_rcu(&tg->siblings, &parent->children);
8518 spin_unlock_irqrestore(&task_group_lock, flags);
8523 free_sched_group(tg);
8524 return ERR_PTR(-ENOMEM);
8527 /* rcu callback to free various structures associated with a task group */
8528 static void free_sched_group_rcu(struct rcu_head *rhp)
8530 /* now it should be safe to free those cfs_rqs */
8531 free_sched_group(container_of(rhp, struct task_group, rcu));
8534 /* Destroy runqueue etc associated with a task group */
8535 void sched_destroy_group(struct task_group *tg)
8537 unsigned long flags;
8540 /* end participation in shares distribution */
8541 for_each_possible_cpu(i)
8542 unregister_fair_sched_group(tg, i);
8544 spin_lock_irqsave(&task_group_lock, flags);
8545 list_del_rcu(&tg->list);
8546 list_del_rcu(&tg->siblings);
8547 spin_unlock_irqrestore(&task_group_lock, flags);
8549 /* wait for possible concurrent references to cfs_rqs complete */
8550 call_rcu(&tg->rcu, free_sched_group_rcu);
8553 /* change task's runqueue when it moves between groups.
8554 * The caller of this function should have put the task in its new group
8555 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8556 * reflect its new group.
8558 void sched_move_task(struct task_struct *tsk)
8561 unsigned long flags;
8564 rq = task_rq_lock(tsk, &flags);
8566 running = task_current(rq, tsk);
8570 dequeue_task(rq, tsk, 0);
8571 if (unlikely(running))
8572 tsk->sched_class->put_prev_task(rq, tsk);
8574 #ifdef CONFIG_FAIR_GROUP_SCHED
8575 if (tsk->sched_class->task_move_group)
8576 tsk->sched_class->task_move_group(tsk, on_rq);
8579 set_task_rq(tsk, task_cpu(tsk));
8581 if (unlikely(running))
8582 tsk->sched_class->set_curr_task(rq);
8584 enqueue_task(rq, tsk, 0);
8586 task_rq_unlock(rq, tsk, &flags);
8588 #endif /* CONFIG_CGROUP_SCHED */
8590 #ifdef CONFIG_FAIR_GROUP_SCHED
8591 static DEFINE_MUTEX(shares_mutex);
8593 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8596 unsigned long flags;
8599 * We can't change the weight of the root cgroup.
8604 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8606 mutex_lock(&shares_mutex);
8607 if (tg->shares == shares)
8610 tg->shares = shares;
8611 for_each_possible_cpu(i) {
8612 struct rq *rq = cpu_rq(i);
8613 struct sched_entity *se;
8616 /* Propagate contribution to hierarchy */
8617 raw_spin_lock_irqsave(&rq->lock, flags);
8618 for_each_sched_entity(se)
8619 update_cfs_shares(group_cfs_rq(se));
8620 raw_spin_unlock_irqrestore(&rq->lock, flags);
8624 mutex_unlock(&shares_mutex);
8628 unsigned long sched_group_shares(struct task_group *tg)
8634 #ifdef CONFIG_RT_GROUP_SCHED
8636 * Ensure that the real time constraints are schedulable.
8638 static DEFINE_MUTEX(rt_constraints_mutex);
8640 static unsigned long to_ratio(u64 period, u64 runtime)
8642 if (runtime == RUNTIME_INF)
8645 return div64_u64(runtime << 20, period);
8648 /* Must be called with tasklist_lock held */
8649 static inline int tg_has_rt_tasks(struct task_group *tg)
8651 struct task_struct *g, *p;
8653 do_each_thread(g, p) {
8654 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8656 } while_each_thread(g, p);
8661 struct rt_schedulable_data {
8662 struct task_group *tg;
8667 static int tg_schedulable(struct task_group *tg, void *data)
8669 struct rt_schedulable_data *d = data;
8670 struct task_group *child;
8671 unsigned long total, sum = 0;
8672 u64 period, runtime;
8674 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8675 runtime = tg->rt_bandwidth.rt_runtime;
8678 period = d->rt_period;
8679 runtime = d->rt_runtime;
8683 * Cannot have more runtime than the period.
8685 if (runtime > period && runtime != RUNTIME_INF)
8689 * Ensure we don't starve existing RT tasks.
8691 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8694 total = to_ratio(period, runtime);
8697 * Nobody can have more than the global setting allows.
8699 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8703 * The sum of our children's runtime should not exceed our own.
8705 list_for_each_entry_rcu(child, &tg->children, siblings) {
8706 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8707 runtime = child->rt_bandwidth.rt_runtime;
8709 if (child == d->tg) {
8710 period = d->rt_period;
8711 runtime = d->rt_runtime;
8714 sum += to_ratio(period, runtime);
8723 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8725 struct rt_schedulable_data data = {
8727 .rt_period = period,
8728 .rt_runtime = runtime,
8731 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8734 static int tg_set_bandwidth(struct task_group *tg,
8735 u64 rt_period, u64 rt_runtime)
8739 mutex_lock(&rt_constraints_mutex);
8740 read_lock(&tasklist_lock);
8741 err = __rt_schedulable(tg, rt_period, rt_runtime);
8745 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8746 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8747 tg->rt_bandwidth.rt_runtime = rt_runtime;
8749 for_each_possible_cpu(i) {
8750 struct rt_rq *rt_rq = tg->rt_rq[i];
8752 raw_spin_lock(&rt_rq->rt_runtime_lock);
8753 rt_rq->rt_runtime = rt_runtime;
8754 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8756 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8758 read_unlock(&tasklist_lock);
8759 mutex_unlock(&rt_constraints_mutex);
8764 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8766 u64 rt_runtime, rt_period;
8768 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8769 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8770 if (rt_runtime_us < 0)
8771 rt_runtime = RUNTIME_INF;
8773 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8776 long sched_group_rt_runtime(struct task_group *tg)
8780 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8783 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8784 do_div(rt_runtime_us, NSEC_PER_USEC);
8785 return rt_runtime_us;
8788 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8790 u64 rt_runtime, rt_period;
8792 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8793 rt_runtime = tg->rt_bandwidth.rt_runtime;
8798 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8801 long sched_group_rt_period(struct task_group *tg)
8805 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8806 do_div(rt_period_us, NSEC_PER_USEC);
8807 return rt_period_us;
8810 static int sched_rt_global_constraints(void)
8812 u64 runtime, period;
8815 if (sysctl_sched_rt_period <= 0)
8818 runtime = global_rt_runtime();
8819 period = global_rt_period();
8822 * Sanity check on the sysctl variables.
8824 if (runtime > period && runtime != RUNTIME_INF)
8827 mutex_lock(&rt_constraints_mutex);
8828 read_lock(&tasklist_lock);
8829 ret = __rt_schedulable(NULL, 0, 0);
8830 read_unlock(&tasklist_lock);
8831 mutex_unlock(&rt_constraints_mutex);
8836 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8838 /* Don't accept realtime tasks when there is no way for them to run */
8839 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8845 #else /* !CONFIG_RT_GROUP_SCHED */
8846 static int sched_rt_global_constraints(void)
8848 unsigned long flags;
8851 if (sysctl_sched_rt_period <= 0)
8855 * There's always some RT tasks in the root group
8856 * -- migration, kstopmachine etc..
8858 if (sysctl_sched_rt_runtime == 0)
8861 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8862 for_each_possible_cpu(i) {
8863 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8865 raw_spin_lock(&rt_rq->rt_runtime_lock);
8866 rt_rq->rt_runtime = global_rt_runtime();
8867 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8869 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8873 #endif /* CONFIG_RT_GROUP_SCHED */
8875 int sched_rt_handler(struct ctl_table *table, int write,
8876 void __user *buffer, size_t *lenp,
8880 int old_period, old_runtime;
8881 static DEFINE_MUTEX(mutex);
8884 old_period = sysctl_sched_rt_period;
8885 old_runtime = sysctl_sched_rt_runtime;
8887 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8889 if (!ret && write) {
8890 ret = sched_rt_global_constraints();
8892 sysctl_sched_rt_period = old_period;
8893 sysctl_sched_rt_runtime = old_runtime;
8895 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8896 def_rt_bandwidth.rt_period =
8897 ns_to_ktime(global_rt_period());
8900 mutex_unlock(&mutex);
8905 #ifdef CONFIG_CGROUP_SCHED
8907 /* return corresponding task_group object of a cgroup */
8908 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8910 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8911 struct task_group, css);
8914 static struct cgroup_subsys_state *
8915 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8917 struct task_group *tg, *parent;
8919 if (!cgrp->parent) {
8920 /* This is early initialization for the top cgroup */
8921 return &root_task_group.css;
8924 parent = cgroup_tg(cgrp->parent);
8925 tg = sched_create_group(parent);
8927 return ERR_PTR(-ENOMEM);
8933 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8935 struct task_group *tg = cgroup_tg(cgrp);
8937 sched_destroy_group(tg);
8941 cpu_cgroup_allow_attach(struct cgroup *cgrp, struct task_struct *tsk)
8943 const struct cred *cred = current_cred(), *tcred;
8945 tcred = __task_cred(tsk);
8947 if ((current != tsk) && !capable(CAP_SYS_NICE) &&
8948 cred->euid != tcred->uid && cred->euid != tcred->suid)
8955 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8957 #ifdef CONFIG_RT_GROUP_SCHED
8958 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8961 /* We don't support RT-tasks being in separate groups */
8962 if (tsk->sched_class != &fair_sched_class)
8969 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8971 sched_move_task(tsk);
8975 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
8976 struct cgroup *old_cgrp, struct task_struct *task)
8979 * cgroup_exit() is called in the copy_process() failure path.
8980 * Ignore this case since the task hasn't ran yet, this avoids
8981 * trying to poke a half freed task state from generic code.
8983 if (!(task->flags & PF_EXITING))
8986 sched_move_task(task);
8989 #ifdef CONFIG_FAIR_GROUP_SCHED
8990 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8993 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
8996 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8998 struct task_group *tg = cgroup_tg(cgrp);
9000 return (u64) scale_load_down(tg->shares);
9002 #endif /* CONFIG_FAIR_GROUP_SCHED */
9004 #ifdef CONFIG_RT_GROUP_SCHED
9005 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9008 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9011 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9013 return sched_group_rt_runtime(cgroup_tg(cgrp));
9016 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9019 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9022 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9024 return sched_group_rt_period(cgroup_tg(cgrp));
9026 #endif /* CONFIG_RT_GROUP_SCHED */
9028 static struct cftype cpu_files[] = {
9029 #ifdef CONFIG_FAIR_GROUP_SCHED
9032 .read_u64 = cpu_shares_read_u64,
9033 .write_u64 = cpu_shares_write_u64,
9036 #ifdef CONFIG_RT_GROUP_SCHED
9038 .name = "rt_runtime_us",
9039 .read_s64 = cpu_rt_runtime_read,
9040 .write_s64 = cpu_rt_runtime_write,
9043 .name = "rt_period_us",
9044 .read_u64 = cpu_rt_period_read_uint,
9045 .write_u64 = cpu_rt_period_write_uint,
9050 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9052 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9055 struct cgroup_subsys cpu_cgroup_subsys = {
9057 .create = cpu_cgroup_create,
9058 .destroy = cpu_cgroup_destroy,
9059 .allow_attach = cpu_cgroup_allow_attach,
9060 .can_attach_task = cpu_cgroup_can_attach_task,
9061 .attach_task = cpu_cgroup_attach_task,
9062 .exit = cpu_cgroup_exit,
9063 .populate = cpu_cgroup_populate,
9064 .subsys_id = cpu_cgroup_subsys_id,
9068 #endif /* CONFIG_CGROUP_SCHED */
9070 #ifdef CONFIG_CGROUP_CPUACCT
9073 * CPU accounting code for task groups.
9075 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9076 * (balbir@in.ibm.com).
9079 /* track cpu usage of a group of tasks and its child groups */
9081 struct cgroup_subsys_state css;
9082 /* cpuusage holds pointer to a u64-type object on every cpu */
9083 u64 __percpu *cpuusage;
9084 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9085 struct cpuacct *parent;
9086 struct cpuacct_charge_calls *cpufreq_fn;
9090 static struct cpuacct *cpuacct_root;
9092 /* Default calls for cpufreq accounting */
9093 static struct cpuacct_charge_calls *cpuacct_cpufreq;
9094 int cpuacct_register_cpufreq(struct cpuacct_charge_calls *fn)
9096 cpuacct_cpufreq = fn;
9099 * Root node is created before platform can register callbacks,
9102 if (cpuacct_root && fn) {
9103 cpuacct_root->cpufreq_fn = fn;
9105 fn->init(&cpuacct_root->cpuacct_data);
9110 struct cgroup_subsys cpuacct_subsys;
9112 /* return cpu accounting group corresponding to this container */
9113 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9115 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9116 struct cpuacct, css);
9119 /* return cpu accounting group to which this task belongs */
9120 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9122 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9123 struct cpuacct, css);
9126 /* create a new cpu accounting group */
9127 static struct cgroup_subsys_state *cpuacct_create(
9128 struct cgroup_subsys *ss, struct cgroup *cgrp)
9130 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9136 ca->cpuusage = alloc_percpu(u64);
9140 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9141 if (percpu_counter_init(&ca->cpustat[i], 0))
9142 goto out_free_counters;
9144 ca->cpufreq_fn = cpuacct_cpufreq;
9146 /* If available, have platform code initalize cpu frequency table */
9147 if (ca->cpufreq_fn && ca->cpufreq_fn->init)
9148 ca->cpufreq_fn->init(&ca->cpuacct_data);
9151 ca->parent = cgroup_ca(cgrp->parent);
9159 percpu_counter_destroy(&ca->cpustat[i]);
9160 free_percpu(ca->cpuusage);
9164 return ERR_PTR(-ENOMEM);
9167 /* destroy an existing cpu accounting group */
9169 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9171 struct cpuacct *ca = cgroup_ca(cgrp);
9174 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9175 percpu_counter_destroy(&ca->cpustat[i]);
9176 free_percpu(ca->cpuusage);
9180 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9182 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9185 #ifndef CONFIG_64BIT
9187 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9189 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9191 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9199 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9201 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9203 #ifndef CONFIG_64BIT
9205 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9207 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9209 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9215 /* return total cpu usage (in nanoseconds) of a group */
9216 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9218 struct cpuacct *ca = cgroup_ca(cgrp);
9219 u64 totalcpuusage = 0;
9222 for_each_present_cpu(i)
9223 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9225 return totalcpuusage;
9228 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9231 struct cpuacct *ca = cgroup_ca(cgrp);
9240 for_each_present_cpu(i)
9241 cpuacct_cpuusage_write(ca, i, 0);
9247 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9250 struct cpuacct *ca = cgroup_ca(cgroup);
9254 for_each_present_cpu(i) {
9255 percpu = cpuacct_cpuusage_read(ca, i);
9256 seq_printf(m, "%llu ", (unsigned long long) percpu);
9258 seq_printf(m, "\n");
9262 static const char *cpuacct_stat_desc[] = {
9263 [CPUACCT_STAT_USER] = "user",
9264 [CPUACCT_STAT_SYSTEM] = "system",
9267 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9268 struct cgroup_map_cb *cb)
9270 struct cpuacct *ca = cgroup_ca(cgrp);
9273 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9274 s64 val = percpu_counter_read(&ca->cpustat[i]);
9275 val = cputime64_to_clock_t(val);
9276 cb->fill(cb, cpuacct_stat_desc[i], val);
9281 static int cpuacct_cpufreq_show(struct cgroup *cgrp, struct cftype *cft,
9282 struct cgroup_map_cb *cb)
9284 struct cpuacct *ca = cgroup_ca(cgrp);
9285 if (ca->cpufreq_fn && ca->cpufreq_fn->cpufreq_show)
9286 ca->cpufreq_fn->cpufreq_show(ca->cpuacct_data, cb);
9291 /* return total cpu power usage (milliWatt second) of a group */
9292 static u64 cpuacct_powerusage_read(struct cgroup *cgrp, struct cftype *cft)
9295 struct cpuacct *ca = cgroup_ca(cgrp);
9298 if (ca->cpufreq_fn && ca->cpufreq_fn->power_usage)
9299 for_each_present_cpu(i) {
9300 totalpower += ca->cpufreq_fn->power_usage(
9307 static struct cftype files[] = {
9310 .read_u64 = cpuusage_read,
9311 .write_u64 = cpuusage_write,
9314 .name = "usage_percpu",
9315 .read_seq_string = cpuacct_percpu_seq_read,
9319 .read_map = cpuacct_stats_show,
9323 .read_map = cpuacct_cpufreq_show,
9327 .read_u64 = cpuacct_powerusage_read
9331 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9333 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9337 * charge this task's execution time to its accounting group.
9339 * called with rq->lock held.
9341 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9346 if (unlikely(!cpuacct_subsys.active))
9349 cpu = task_cpu(tsk);
9355 for (; ca; ca = ca->parent) {
9356 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9357 *cpuusage += cputime;
9359 /* Call back into platform code to account for CPU speeds */
9360 if (ca->cpufreq_fn && ca->cpufreq_fn->charge)
9361 ca->cpufreq_fn->charge(ca->cpuacct_data, cputime, cpu);
9368 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9369 * in cputime_t units. As a result, cpuacct_update_stats calls
9370 * percpu_counter_add with values large enough to always overflow the
9371 * per cpu batch limit causing bad SMP scalability.
9373 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9374 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9375 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9378 #define CPUACCT_BATCH \
9379 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9381 #define CPUACCT_BATCH 0
9385 * Charge the system/user time to the task's accounting group.
9387 static void cpuacct_update_stats(struct task_struct *tsk,
9388 enum cpuacct_stat_index idx, cputime_t val)
9391 int batch = CPUACCT_BATCH;
9393 if (unlikely(!cpuacct_subsys.active))
9400 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9406 struct cgroup_subsys cpuacct_subsys = {
9408 .create = cpuacct_create,
9409 .destroy = cpuacct_destroy,
9410 .populate = cpuacct_populate,
9411 .subsys_id = cpuacct_subsys_id,
9413 #endif /* CONFIG_CGROUP_CPUACCT */