4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/stop_machine.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/debugfs.h>
71 #include <linux/ctype.h>
72 #include <linux/ftrace.h>
73 #include <linux/slab.h>
76 #include <asm/irq_regs.h>
77 #include <asm/mutex.h>
78 #ifdef CONFIG_PARAVIRT
79 #include <asm/paravirt.h>
82 #include "sched_cpupri.h"
83 #include "workqueue_sched.h"
84 #include "sched_autogroup.h"
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/sched.h>
90 * Convert user-nice values [ -20 ... 0 ... 19 ]
91 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
94 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
95 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
96 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
99 * 'User priority' is the nice value converted to something we
100 * can work with better when scaling various scheduler parameters,
101 * it's a [ 0 ... 39 ] range.
103 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
104 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
105 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
108 * Helpers for converting nanosecond timing to jiffy resolution
110 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
112 #define NICE_0_LOAD SCHED_LOAD_SCALE
113 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
116 * These are the 'tuning knobs' of the scheduler:
118 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
119 * Timeslices get refilled after they expire.
121 #define DEF_TIMESLICE (100 * HZ / 1000)
124 * single value that denotes runtime == period, ie unlimited time.
126 #define RUNTIME_INF ((u64)~0ULL)
128 static inline int rt_policy(int policy)
130 if (policy == SCHED_FIFO || policy == SCHED_RR)
135 static inline int task_has_rt_policy(struct task_struct *p)
137 return rt_policy(p->policy);
141 * This is the priority-queue data structure of the RT scheduling class:
143 struct rt_prio_array {
144 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
145 struct list_head queue[MAX_RT_PRIO];
148 struct rt_bandwidth {
149 /* nests inside the rq lock: */
150 raw_spinlock_t rt_runtime_lock;
153 struct hrtimer rt_period_timer;
156 static struct rt_bandwidth def_rt_bandwidth;
158 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
160 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
162 struct rt_bandwidth *rt_b =
163 container_of(timer, struct rt_bandwidth, rt_period_timer);
169 now = hrtimer_cb_get_time(timer);
170 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
175 idle = do_sched_rt_period_timer(rt_b, overrun);
178 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
182 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
184 rt_b->rt_period = ns_to_ktime(period);
185 rt_b->rt_runtime = runtime;
187 raw_spin_lock_init(&rt_b->rt_runtime_lock);
189 hrtimer_init(&rt_b->rt_period_timer,
190 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
191 rt_b->rt_period_timer.function = sched_rt_period_timer;
194 static inline int rt_bandwidth_enabled(void)
196 return sysctl_sched_rt_runtime >= 0;
199 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
203 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
206 if (hrtimer_active(&rt_b->rt_period_timer))
209 raw_spin_lock(&rt_b->rt_runtime_lock);
214 if (hrtimer_active(&rt_b->rt_period_timer))
217 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
218 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
220 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
221 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
222 delta = ktime_to_ns(ktime_sub(hard, soft));
223 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
224 HRTIMER_MODE_ABS_PINNED, 0);
226 raw_spin_unlock(&rt_b->rt_runtime_lock);
229 #ifdef CONFIG_RT_GROUP_SCHED
230 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
232 hrtimer_cancel(&rt_b->rt_period_timer);
237 * sched_domains_mutex serializes calls to init_sched_domains,
238 * detach_destroy_domains and partition_sched_domains.
240 static DEFINE_MUTEX(sched_domains_mutex);
242 #ifdef CONFIG_CGROUP_SCHED
244 #include <linux/cgroup.h>
248 static LIST_HEAD(task_groups);
250 /* task group related information */
252 struct cgroup_subsys_state css;
254 #ifdef CONFIG_FAIR_GROUP_SCHED
255 /* schedulable entities of this group on each cpu */
256 struct sched_entity **se;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq **cfs_rq;
259 unsigned long shares;
261 atomic_t load_weight;
264 #ifdef CONFIG_RT_GROUP_SCHED
265 struct sched_rt_entity **rt_se;
266 struct rt_rq **rt_rq;
268 struct rt_bandwidth rt_bandwidth;
272 struct list_head list;
274 struct task_group *parent;
275 struct list_head siblings;
276 struct list_head children;
278 #ifdef CONFIG_SCHED_AUTOGROUP
279 struct autogroup *autogroup;
283 /* task_group_lock serializes the addition/removal of task groups */
284 static DEFINE_SPINLOCK(task_group_lock);
286 #ifdef CONFIG_FAIR_GROUP_SCHED
288 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
291 * A weight of 0 or 1 can cause arithmetics problems.
292 * A weight of a cfs_rq is the sum of weights of which entities
293 * are queued on this cfs_rq, so a weight of a entity should not be
294 * too large, so as the shares value of a task group.
295 * (The default weight is 1024 - so there's no practical
296 * limitation from this.)
298 #define MIN_SHARES (1UL << 1)
299 #define MAX_SHARES (1UL << 18)
301 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
304 /* Default task group.
305 * Every task in system belong to this group at bootup.
307 struct task_group root_task_group;
309 #endif /* CONFIG_CGROUP_SCHED */
311 /* CFS-related fields in a runqueue */
313 struct load_weight load;
314 unsigned long nr_running;
319 u64 min_vruntime_copy;
322 struct rb_root tasks_timeline;
323 struct rb_node *rb_leftmost;
325 struct list_head tasks;
326 struct list_head *balance_iterator;
329 * 'curr' points to currently running entity on this cfs_rq.
330 * It is set to NULL otherwise (i.e when none are currently running).
332 struct sched_entity *curr, *next, *last, *skip;
334 #ifdef CONFIG_SCHED_DEBUG
335 unsigned int nr_spread_over;
338 #ifdef CONFIG_FAIR_GROUP_SCHED
339 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
342 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
343 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
344 * (like users, containers etc.)
346 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
347 * list is used during load balance.
350 struct list_head leaf_cfs_rq_list;
351 struct task_group *tg; /* group that "owns" this runqueue */
355 * the part of load.weight contributed by tasks
357 unsigned long task_weight;
360 * h_load = weight * f(tg)
362 * Where f(tg) is the recursive weight fraction assigned to
365 unsigned long h_load;
368 * Maintaining per-cpu shares distribution for group scheduling
370 * load_stamp is the last time we updated the load average
371 * load_last is the last time we updated the load average and saw load
372 * load_unacc_exec_time is currently unaccounted execution time
376 u64 load_stamp, load_last, load_unacc_exec_time;
378 unsigned long load_contribution;
383 /* Real-Time classes' related field in a runqueue: */
385 struct rt_prio_array active;
386 unsigned long rt_nr_running;
387 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
389 int curr; /* highest queued rt task prio */
391 int next; /* next highest */
396 unsigned long rt_nr_migratory;
397 unsigned long rt_nr_total;
399 struct plist_head pushable_tasks;
404 /* Nests inside the rq lock: */
405 raw_spinlock_t rt_runtime_lock;
407 #ifdef CONFIG_RT_GROUP_SCHED
408 unsigned long rt_nr_boosted;
411 struct list_head leaf_rt_rq_list;
412 struct task_group *tg;
419 * We add the notion of a root-domain which will be used to define per-domain
420 * variables. Each exclusive cpuset essentially defines an island domain by
421 * fully partitioning the member cpus from any other cpuset. Whenever a new
422 * exclusive cpuset is created, we also create and attach a new root-domain
431 cpumask_var_t online;
434 * The "RT overload" flag: it gets set if a CPU has more than
435 * one runnable RT task.
437 cpumask_var_t rto_mask;
438 struct cpupri cpupri;
442 * By default the system creates a single root-domain with all cpus as
443 * members (mimicking the global state we have today).
445 static struct root_domain def_root_domain;
447 #endif /* CONFIG_SMP */
450 * This is the main, per-CPU runqueue data structure.
452 * Locking rule: those places that want to lock multiple runqueues
453 * (such as the load balancing or the thread migration code), lock
454 * acquire operations must be ordered by ascending &runqueue.
461 * nr_running and cpu_load should be in the same cacheline because
462 * remote CPUs use both these fields when doing load calculation.
464 unsigned long nr_running;
465 #define CPU_LOAD_IDX_MAX 5
466 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
467 unsigned long last_load_update_tick;
470 unsigned char nohz_balance_kick;
472 int skip_clock_update;
474 /* capture load from *all* tasks on this cpu: */
475 struct load_weight load;
476 unsigned long nr_load_updates;
482 #ifdef CONFIG_FAIR_GROUP_SCHED
483 /* list of leaf cfs_rq on this cpu: */
484 struct list_head leaf_cfs_rq_list;
486 #ifdef CONFIG_RT_GROUP_SCHED
487 struct list_head leaf_rt_rq_list;
491 * This is part of a global counter where only the total sum
492 * over all CPUs matters. A task can increase this counter on
493 * one CPU and if it got migrated afterwards it may decrease
494 * it on another CPU. Always updated under the runqueue lock:
496 unsigned long nr_uninterruptible;
498 struct task_struct *curr, *idle, *stop;
499 unsigned long next_balance;
500 struct mm_struct *prev_mm;
508 struct root_domain *rd;
509 struct sched_domain *sd;
511 unsigned long cpu_power;
513 unsigned char idle_at_tick;
514 /* For active balancing */
518 struct cpu_stop_work active_balance_work;
519 /* cpu of this runqueue: */
529 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
532 #ifdef CONFIG_PARAVIRT
535 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
536 u64 prev_steal_time_rq;
539 /* calc_load related fields */
540 unsigned long calc_load_update;
541 long calc_load_active;
543 #ifdef CONFIG_SCHED_HRTICK
545 int hrtick_csd_pending;
546 struct call_single_data hrtick_csd;
548 struct hrtimer hrtick_timer;
551 #ifdef CONFIG_SCHEDSTATS
553 struct sched_info rq_sched_info;
554 unsigned long long rq_cpu_time;
555 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
557 /* sys_sched_yield() stats */
558 unsigned int yld_count;
560 /* schedule() stats */
561 unsigned int sched_switch;
562 unsigned int sched_count;
563 unsigned int sched_goidle;
565 /* try_to_wake_up() stats */
566 unsigned int ttwu_count;
567 unsigned int ttwu_local;
571 struct task_struct *wake_list;
575 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
578 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
580 static inline int cpu_of(struct rq *rq)
589 #define rcu_dereference_check_sched_domain(p) \
590 rcu_dereference_check((p), \
591 lockdep_is_held(&sched_domains_mutex))
594 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
595 * See detach_destroy_domains: synchronize_sched for details.
597 * The domain tree of any CPU may only be accessed from within
598 * preempt-disabled sections.
600 #define for_each_domain(cpu, __sd) \
601 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
603 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
604 #define this_rq() (&__get_cpu_var(runqueues))
605 #define task_rq(p) cpu_rq(task_cpu(p))
606 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
607 #define raw_rq() (&__raw_get_cpu_var(runqueues))
609 #ifdef CONFIG_CGROUP_SCHED
612 * Return the group to which this tasks belongs.
614 * We use task_subsys_state_check() and extend the RCU verification with
615 * pi->lock and rq->lock because cpu_cgroup_attach() holds those locks for each
616 * task it moves into the cgroup. Therefore by holding either of those locks,
617 * we pin the task to the current cgroup.
619 static inline struct task_group *task_group(struct task_struct *p)
621 struct task_group *tg;
622 struct cgroup_subsys_state *css;
624 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
625 lockdep_is_held(&p->pi_lock) ||
626 lockdep_is_held(&task_rq(p)->lock));
627 tg = container_of(css, struct task_group, css);
629 return autogroup_task_group(p, tg);
632 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
633 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
635 #ifdef CONFIG_FAIR_GROUP_SCHED
636 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
637 p->se.parent = task_group(p)->se[cpu];
640 #ifdef CONFIG_RT_GROUP_SCHED
641 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
642 p->rt.parent = task_group(p)->rt_se[cpu];
646 #else /* CONFIG_CGROUP_SCHED */
648 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
649 static inline struct task_group *task_group(struct task_struct *p)
654 #endif /* CONFIG_CGROUP_SCHED */
656 static void update_rq_clock_task(struct rq *rq, s64 delta);
658 static void update_rq_clock(struct rq *rq)
662 if (rq->skip_clock_update > 0)
665 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
667 update_rq_clock_task(rq, delta);
671 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
673 #ifdef CONFIG_SCHED_DEBUG
674 # define const_debug __read_mostly
676 # define const_debug static const
680 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
681 * @cpu: the processor in question.
683 * This interface allows printk to be called with the runqueue lock
684 * held and know whether or not it is OK to wake up the klogd.
686 int runqueue_is_locked(int cpu)
688 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
692 * Debugging: various feature bits
695 #define SCHED_FEAT(name, enabled) \
696 __SCHED_FEAT_##name ,
699 #include "sched_features.h"
704 #define SCHED_FEAT(name, enabled) \
705 (1UL << __SCHED_FEAT_##name) * enabled |
707 const_debug unsigned int sysctl_sched_features =
708 #include "sched_features.h"
713 #ifdef CONFIG_SCHED_DEBUG
714 #define SCHED_FEAT(name, enabled) \
717 static __read_mostly char *sched_feat_names[] = {
718 #include "sched_features.h"
724 static int sched_feat_show(struct seq_file *m, void *v)
728 for (i = 0; sched_feat_names[i]; i++) {
729 if (!(sysctl_sched_features & (1UL << i)))
731 seq_printf(m, "%s ", sched_feat_names[i]);
739 sched_feat_write(struct file *filp, const char __user *ubuf,
740 size_t cnt, loff_t *ppos)
750 if (copy_from_user(&buf, ubuf, cnt))
756 if (strncmp(cmp, "NO_", 3) == 0) {
761 for (i = 0; sched_feat_names[i]; i++) {
762 if (strcmp(cmp, sched_feat_names[i]) == 0) {
764 sysctl_sched_features &= ~(1UL << i);
766 sysctl_sched_features |= (1UL << i);
771 if (!sched_feat_names[i])
779 static int sched_feat_open(struct inode *inode, struct file *filp)
781 return single_open(filp, sched_feat_show, NULL);
784 static const struct file_operations sched_feat_fops = {
785 .open = sched_feat_open,
786 .write = sched_feat_write,
789 .release = single_release,
792 static __init int sched_init_debug(void)
794 debugfs_create_file("sched_features", 0644, NULL, NULL,
799 late_initcall(sched_init_debug);
803 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
806 * Number of tasks to iterate in a single balance run.
807 * Limited because this is done with IRQs disabled.
809 const_debug unsigned int sysctl_sched_nr_migrate = 32;
812 * period over which we average the RT time consumption, measured
817 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
820 * period over which we measure -rt task cpu usage in us.
823 unsigned int sysctl_sched_rt_period = 1000000;
825 static __read_mostly int scheduler_running;
828 * part of the period that we allow rt tasks to run in us.
831 int sysctl_sched_rt_runtime = 950000;
833 static inline u64 global_rt_period(void)
835 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
838 static inline u64 global_rt_runtime(void)
840 if (sysctl_sched_rt_runtime < 0)
843 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
846 #ifndef prepare_arch_switch
847 # define prepare_arch_switch(next) do { } while (0)
849 #ifndef finish_arch_switch
850 # define finish_arch_switch(prev) do { } while (0)
853 static inline int task_current(struct rq *rq, struct task_struct *p)
855 return rq->curr == p;
858 static inline int task_running(struct rq *rq, struct task_struct *p)
863 return task_current(rq, p);
867 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
868 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
872 * We can optimise this out completely for !SMP, because the
873 * SMP rebalancing from interrupt is the only thing that cares
880 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
884 * After ->on_cpu is cleared, the task can be moved to a different CPU.
885 * We must ensure this doesn't happen until the switch is completely
891 #ifdef CONFIG_DEBUG_SPINLOCK
892 /* this is a valid case when another task releases the spinlock */
893 rq->lock.owner = current;
896 * If we are tracking spinlock dependencies then we have to
897 * fix up the runqueue lock - which gets 'carried over' from
900 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
902 raw_spin_unlock_irq(&rq->lock);
905 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
906 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
910 * We can optimise this out completely for !SMP, because the
911 * SMP rebalancing from interrupt is the only thing that cares
916 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
917 raw_spin_unlock_irq(&rq->lock);
919 raw_spin_unlock(&rq->lock);
923 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
927 * After ->on_cpu is cleared, the task can be moved to a different CPU.
928 * We must ensure this doesn't happen until the switch is completely
934 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
938 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
941 * __task_rq_lock - lock the rq @p resides on.
943 static inline struct rq *__task_rq_lock(struct task_struct *p)
948 lockdep_assert_held(&p->pi_lock);
952 raw_spin_lock(&rq->lock);
953 if (likely(rq == task_rq(p)))
955 raw_spin_unlock(&rq->lock);
960 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
962 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
963 __acquires(p->pi_lock)
969 raw_spin_lock_irqsave(&p->pi_lock, *flags);
971 raw_spin_lock(&rq->lock);
972 if (likely(rq == task_rq(p)))
974 raw_spin_unlock(&rq->lock);
975 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
979 static void __task_rq_unlock(struct rq *rq)
982 raw_spin_unlock(&rq->lock);
986 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
988 __releases(p->pi_lock)
990 raw_spin_unlock(&rq->lock);
991 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
995 * this_rq_lock - lock this runqueue and disable interrupts.
997 static struct rq *this_rq_lock(void)
1002 local_irq_disable();
1004 raw_spin_lock(&rq->lock);
1009 #ifdef CONFIG_SCHED_HRTICK
1011 * Use HR-timers to deliver accurate preemption points.
1013 * Its all a bit involved since we cannot program an hrt while holding the
1014 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1017 * When we get rescheduled we reprogram the hrtick_timer outside of the
1023 * - enabled by features
1024 * - hrtimer is actually high res
1026 static inline int hrtick_enabled(struct rq *rq)
1028 if (!sched_feat(HRTICK))
1030 if (!cpu_active(cpu_of(rq)))
1032 return hrtimer_is_hres_active(&rq->hrtick_timer);
1035 static void hrtick_clear(struct rq *rq)
1037 if (hrtimer_active(&rq->hrtick_timer))
1038 hrtimer_cancel(&rq->hrtick_timer);
1042 * High-resolution timer tick.
1043 * Runs from hardirq context with interrupts disabled.
1045 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1047 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1049 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1051 raw_spin_lock(&rq->lock);
1052 update_rq_clock(rq);
1053 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1054 raw_spin_unlock(&rq->lock);
1056 return HRTIMER_NORESTART;
1061 * called from hardirq (IPI) context
1063 static void __hrtick_start(void *arg)
1065 struct rq *rq = arg;
1067 raw_spin_lock(&rq->lock);
1068 hrtimer_restart(&rq->hrtick_timer);
1069 rq->hrtick_csd_pending = 0;
1070 raw_spin_unlock(&rq->lock);
1074 * Called to set the hrtick timer state.
1076 * called with rq->lock held and irqs disabled
1078 static void hrtick_start(struct rq *rq, u64 delay)
1080 struct hrtimer *timer = &rq->hrtick_timer;
1081 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1083 hrtimer_set_expires(timer, time);
1085 if (rq == this_rq()) {
1086 hrtimer_restart(timer);
1087 } else if (!rq->hrtick_csd_pending) {
1088 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1089 rq->hrtick_csd_pending = 1;
1094 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1096 int cpu = (int)(long)hcpu;
1099 case CPU_UP_CANCELED:
1100 case CPU_UP_CANCELED_FROZEN:
1101 case CPU_DOWN_PREPARE:
1102 case CPU_DOWN_PREPARE_FROZEN:
1104 case CPU_DEAD_FROZEN:
1105 hrtick_clear(cpu_rq(cpu));
1112 static __init void init_hrtick(void)
1114 hotcpu_notifier(hotplug_hrtick, 0);
1118 * Called to set the hrtick timer state.
1120 * called with rq->lock held and irqs disabled
1122 static void hrtick_start(struct rq *rq, u64 delay)
1124 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1125 HRTIMER_MODE_REL_PINNED, 0);
1128 static inline void init_hrtick(void)
1131 #endif /* CONFIG_SMP */
1133 static void init_rq_hrtick(struct rq *rq)
1136 rq->hrtick_csd_pending = 0;
1138 rq->hrtick_csd.flags = 0;
1139 rq->hrtick_csd.func = __hrtick_start;
1140 rq->hrtick_csd.info = rq;
1143 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1144 rq->hrtick_timer.function = hrtick;
1146 #else /* CONFIG_SCHED_HRTICK */
1147 static inline void hrtick_clear(struct rq *rq)
1151 static inline void init_rq_hrtick(struct rq *rq)
1155 static inline void init_hrtick(void)
1158 #endif /* CONFIG_SCHED_HRTICK */
1161 * resched_task - mark a task 'to be rescheduled now'.
1163 * On UP this means the setting of the need_resched flag, on SMP it
1164 * might also involve a cross-CPU call to trigger the scheduler on
1169 #ifndef tsk_is_polling
1170 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1173 static void resched_task(struct task_struct *p)
1177 assert_raw_spin_locked(&task_rq(p)->lock);
1179 if (test_tsk_need_resched(p))
1182 set_tsk_need_resched(p);
1185 if (cpu == smp_processor_id())
1188 /* NEED_RESCHED must be visible before we test polling */
1190 if (!tsk_is_polling(p))
1191 smp_send_reschedule(cpu);
1194 static void resched_cpu(int cpu)
1196 struct rq *rq = cpu_rq(cpu);
1197 unsigned long flags;
1199 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1201 resched_task(cpu_curr(cpu));
1202 raw_spin_unlock_irqrestore(&rq->lock, flags);
1207 * In the semi idle case, use the nearest busy cpu for migrating timers
1208 * from an idle cpu. This is good for power-savings.
1210 * We don't do similar optimization for completely idle system, as
1211 * selecting an idle cpu will add more delays to the timers than intended
1212 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1214 int get_nohz_timer_target(void)
1216 int cpu = smp_processor_id();
1218 struct sched_domain *sd;
1221 for_each_domain(cpu, sd) {
1222 for_each_cpu(i, sched_domain_span(sd)) {
1234 * When add_timer_on() enqueues a timer into the timer wheel of an
1235 * idle CPU then this timer might expire before the next timer event
1236 * which is scheduled to wake up that CPU. In case of a completely
1237 * idle system the next event might even be infinite time into the
1238 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1239 * leaves the inner idle loop so the newly added timer is taken into
1240 * account when the CPU goes back to idle and evaluates the timer
1241 * wheel for the next timer event.
1243 void wake_up_idle_cpu(int cpu)
1245 struct rq *rq = cpu_rq(cpu);
1247 if (cpu == smp_processor_id())
1251 * This is safe, as this function is called with the timer
1252 * wheel base lock of (cpu) held. When the CPU is on the way
1253 * to idle and has not yet set rq->curr to idle then it will
1254 * be serialized on the timer wheel base lock and take the new
1255 * timer into account automatically.
1257 if (rq->curr != rq->idle)
1261 * We can set TIF_RESCHED on the idle task of the other CPU
1262 * lockless. The worst case is that the other CPU runs the
1263 * idle task through an additional NOOP schedule()
1265 set_tsk_need_resched(rq->idle);
1267 /* NEED_RESCHED must be visible before we test polling */
1269 if (!tsk_is_polling(rq->idle))
1270 smp_send_reschedule(cpu);
1273 #endif /* CONFIG_NO_HZ */
1275 static u64 sched_avg_period(void)
1277 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1280 static void sched_avg_update(struct rq *rq)
1282 s64 period = sched_avg_period();
1284 while ((s64)(rq->clock - rq->age_stamp) > period) {
1286 * Inline assembly required to prevent the compiler
1287 * optimising this loop into a divmod call.
1288 * See __iter_div_u64_rem() for another example of this.
1290 asm("" : "+rm" (rq->age_stamp));
1291 rq->age_stamp += period;
1296 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1298 rq->rt_avg += rt_delta;
1299 sched_avg_update(rq);
1302 #else /* !CONFIG_SMP */
1303 static void resched_task(struct task_struct *p)
1305 assert_raw_spin_locked(&task_rq(p)->lock);
1306 set_tsk_need_resched(p);
1309 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1313 static void sched_avg_update(struct rq *rq)
1316 #endif /* CONFIG_SMP */
1318 #if BITS_PER_LONG == 32
1319 # define WMULT_CONST (~0UL)
1321 # define WMULT_CONST (1UL << 32)
1324 #define WMULT_SHIFT 32
1327 * Shift right and round:
1329 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1332 * delta *= weight / lw
1334 static unsigned long
1335 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1336 struct load_weight *lw)
1341 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1342 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1343 * 2^SCHED_LOAD_RESOLUTION.
1345 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
1346 tmp = (u64)delta_exec * scale_load_down(weight);
1348 tmp = (u64)delta_exec;
1350 if (!lw->inv_weight) {
1351 unsigned long w = scale_load_down(lw->weight);
1353 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
1355 else if (unlikely(!w))
1356 lw->inv_weight = WMULT_CONST;
1358 lw->inv_weight = WMULT_CONST / w;
1362 * Check whether we'd overflow the 64-bit multiplication:
1364 if (unlikely(tmp > WMULT_CONST))
1365 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1368 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1370 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1373 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1379 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1385 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1392 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1393 * of tasks with abnormal "nice" values across CPUs the contribution that
1394 * each task makes to its run queue's load is weighted according to its
1395 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1396 * scaled version of the new time slice allocation that they receive on time
1400 #define WEIGHT_IDLEPRIO 3
1401 #define WMULT_IDLEPRIO 1431655765
1404 * Nice levels are multiplicative, with a gentle 10% change for every
1405 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1406 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1407 * that remained on nice 0.
1409 * The "10% effect" is relative and cumulative: from _any_ nice level,
1410 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1411 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1412 * If a task goes up by ~10% and another task goes down by ~10% then
1413 * the relative distance between them is ~25%.)
1415 static const int prio_to_weight[40] = {
1416 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1417 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1418 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1419 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1420 /* 0 */ 1024, 820, 655, 526, 423,
1421 /* 5 */ 335, 272, 215, 172, 137,
1422 /* 10 */ 110, 87, 70, 56, 45,
1423 /* 15 */ 36, 29, 23, 18, 15,
1427 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1429 * In cases where the weight does not change often, we can use the
1430 * precalculated inverse to speed up arithmetics by turning divisions
1431 * into multiplications:
1433 static const u32 prio_to_wmult[40] = {
1434 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1435 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1436 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1437 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1438 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1439 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1440 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1441 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1444 /* Time spent by the tasks of the cpu accounting group executing in ... */
1445 enum cpuacct_stat_index {
1446 CPUACCT_STAT_USER, /* ... user mode */
1447 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1449 CPUACCT_STAT_NSTATS,
1452 #ifdef CONFIG_CGROUP_CPUACCT
1453 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1454 static void cpuacct_update_stats(struct task_struct *tsk,
1455 enum cpuacct_stat_index idx, cputime_t val);
1457 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1458 static inline void cpuacct_update_stats(struct task_struct *tsk,
1459 enum cpuacct_stat_index idx, cputime_t val) {}
1462 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1464 update_load_add(&rq->load, load);
1467 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1469 update_load_sub(&rq->load, load);
1472 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1473 typedef int (*tg_visitor)(struct task_group *, void *);
1476 * Iterate the full tree, calling @down when first entering a node and @up when
1477 * leaving it for the final time.
1479 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1481 struct task_group *parent, *child;
1485 parent = &root_task_group;
1487 ret = (*down)(parent, data);
1490 list_for_each_entry_rcu(child, &parent->children, siblings) {
1497 ret = (*up)(parent, data);
1502 parent = parent->parent;
1511 static int tg_nop(struct task_group *tg, void *data)
1518 /* Used instead of source_load when we know the type == 0 */
1519 static unsigned long weighted_cpuload(const int cpu)
1521 return cpu_rq(cpu)->load.weight;
1525 * Return a low guess at the load of a migration-source cpu weighted
1526 * according to the scheduling class and "nice" value.
1528 * We want to under-estimate the load of migration sources, to
1529 * balance conservatively.
1531 static unsigned long source_load(int cpu, int type)
1533 struct rq *rq = cpu_rq(cpu);
1534 unsigned long total = weighted_cpuload(cpu);
1536 if (type == 0 || !sched_feat(LB_BIAS))
1539 return min(rq->cpu_load[type-1], total);
1543 * Return a high guess at the load of a migration-target cpu weighted
1544 * according to the scheduling class and "nice" value.
1546 static unsigned long target_load(int cpu, int type)
1548 struct rq *rq = cpu_rq(cpu);
1549 unsigned long total = weighted_cpuload(cpu);
1551 if (type == 0 || !sched_feat(LB_BIAS))
1554 return max(rq->cpu_load[type-1], total);
1557 static unsigned long power_of(int cpu)
1559 return cpu_rq(cpu)->cpu_power;
1562 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1564 static unsigned long cpu_avg_load_per_task(int cpu)
1566 struct rq *rq = cpu_rq(cpu);
1567 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1570 return rq->load.weight / nr_running;
1575 #ifdef CONFIG_PREEMPT
1577 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1580 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1581 * way at the expense of forcing extra atomic operations in all
1582 * invocations. This assures that the double_lock is acquired using the
1583 * same underlying policy as the spinlock_t on this architecture, which
1584 * reduces latency compared to the unfair variant below. However, it
1585 * also adds more overhead and therefore may reduce throughput.
1587 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1588 __releases(this_rq->lock)
1589 __acquires(busiest->lock)
1590 __acquires(this_rq->lock)
1592 raw_spin_unlock(&this_rq->lock);
1593 double_rq_lock(this_rq, busiest);
1600 * Unfair double_lock_balance: Optimizes throughput at the expense of
1601 * latency by eliminating extra atomic operations when the locks are
1602 * already in proper order on entry. This favors lower cpu-ids and will
1603 * grant the double lock to lower cpus over higher ids under contention,
1604 * regardless of entry order into the function.
1606 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1607 __releases(this_rq->lock)
1608 __acquires(busiest->lock)
1609 __acquires(this_rq->lock)
1613 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1614 if (busiest < this_rq) {
1615 raw_spin_unlock(&this_rq->lock);
1616 raw_spin_lock(&busiest->lock);
1617 raw_spin_lock_nested(&this_rq->lock,
1618 SINGLE_DEPTH_NESTING);
1621 raw_spin_lock_nested(&busiest->lock,
1622 SINGLE_DEPTH_NESTING);
1627 #endif /* CONFIG_PREEMPT */
1630 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1632 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1634 if (unlikely(!irqs_disabled())) {
1635 /* printk() doesn't work good under rq->lock */
1636 raw_spin_unlock(&this_rq->lock);
1640 return _double_lock_balance(this_rq, busiest);
1643 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1644 __releases(busiest->lock)
1646 raw_spin_unlock(&busiest->lock);
1647 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1651 * double_rq_lock - safely lock two runqueues
1653 * Note this does not disable interrupts like task_rq_lock,
1654 * you need to do so manually before calling.
1656 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1657 __acquires(rq1->lock)
1658 __acquires(rq2->lock)
1660 BUG_ON(!irqs_disabled());
1662 raw_spin_lock(&rq1->lock);
1663 __acquire(rq2->lock); /* Fake it out ;) */
1666 raw_spin_lock(&rq1->lock);
1667 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1669 raw_spin_lock(&rq2->lock);
1670 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1676 * double_rq_unlock - safely unlock two runqueues
1678 * Note this does not restore interrupts like task_rq_unlock,
1679 * you need to do so manually after calling.
1681 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1682 __releases(rq1->lock)
1683 __releases(rq2->lock)
1685 raw_spin_unlock(&rq1->lock);
1687 raw_spin_unlock(&rq2->lock);
1689 __release(rq2->lock);
1692 #else /* CONFIG_SMP */
1695 * double_rq_lock - safely lock two runqueues
1697 * Note this does not disable interrupts like task_rq_lock,
1698 * you need to do so manually before calling.
1700 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1701 __acquires(rq1->lock)
1702 __acquires(rq2->lock)
1704 BUG_ON(!irqs_disabled());
1706 raw_spin_lock(&rq1->lock);
1707 __acquire(rq2->lock); /* Fake it out ;) */
1711 * double_rq_unlock - safely unlock two runqueues
1713 * Note this does not restore interrupts like task_rq_unlock,
1714 * you need to do so manually after calling.
1716 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1717 __releases(rq1->lock)
1718 __releases(rq2->lock)
1721 raw_spin_unlock(&rq1->lock);
1722 __release(rq2->lock);
1727 static void calc_load_account_idle(struct rq *this_rq);
1728 static void update_sysctl(void);
1729 static int get_update_sysctl_factor(void);
1730 static void update_cpu_load(struct rq *this_rq);
1732 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1734 set_task_rq(p, cpu);
1737 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1738 * successfuly executed on another CPU. We must ensure that updates of
1739 * per-task data have been completed by this moment.
1742 task_thread_info(p)->cpu = cpu;
1746 static const struct sched_class rt_sched_class;
1748 #define sched_class_highest (&stop_sched_class)
1749 #define for_each_class(class) \
1750 for (class = sched_class_highest; class; class = class->next)
1752 #include "sched_stats.h"
1754 static void inc_nr_running(struct rq *rq)
1759 static void dec_nr_running(struct rq *rq)
1764 static void set_load_weight(struct task_struct *p)
1766 int prio = p->static_prio - MAX_RT_PRIO;
1767 struct load_weight *load = &p->se.load;
1770 * SCHED_IDLE tasks get minimal weight:
1772 if (p->policy == SCHED_IDLE) {
1773 load->weight = scale_load(WEIGHT_IDLEPRIO);
1774 load->inv_weight = WMULT_IDLEPRIO;
1778 load->weight = scale_load(prio_to_weight[prio]);
1779 load->inv_weight = prio_to_wmult[prio];
1782 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1784 update_rq_clock(rq);
1785 sched_info_queued(p);
1786 p->sched_class->enqueue_task(rq, p, flags);
1789 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1791 update_rq_clock(rq);
1792 sched_info_dequeued(p);
1793 p->sched_class->dequeue_task(rq, p, flags);
1797 * activate_task - move a task to the runqueue.
1799 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1801 if (task_contributes_to_load(p))
1802 rq->nr_uninterruptible--;
1804 enqueue_task(rq, p, flags);
1809 * deactivate_task - remove a task from the runqueue.
1811 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1813 if (task_contributes_to_load(p))
1814 rq->nr_uninterruptible++;
1816 dequeue_task(rq, p, flags);
1820 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1823 * There are no locks covering percpu hardirq/softirq time.
1824 * They are only modified in account_system_vtime, on corresponding CPU
1825 * with interrupts disabled. So, writes are safe.
1826 * They are read and saved off onto struct rq in update_rq_clock().
1827 * This may result in other CPU reading this CPU's irq time and can
1828 * race with irq/account_system_vtime on this CPU. We would either get old
1829 * or new value with a side effect of accounting a slice of irq time to wrong
1830 * task when irq is in progress while we read rq->clock. That is a worthy
1831 * compromise in place of having locks on each irq in account_system_time.
1833 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1834 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1836 static DEFINE_PER_CPU(u64, irq_start_time);
1837 static int sched_clock_irqtime;
1839 void enable_sched_clock_irqtime(void)
1841 sched_clock_irqtime = 1;
1844 void disable_sched_clock_irqtime(void)
1846 sched_clock_irqtime = 0;
1849 #ifndef CONFIG_64BIT
1850 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1852 static inline void irq_time_write_begin(void)
1854 __this_cpu_inc(irq_time_seq.sequence);
1858 static inline void irq_time_write_end(void)
1861 __this_cpu_inc(irq_time_seq.sequence);
1864 static inline u64 irq_time_read(int cpu)
1870 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1871 irq_time = per_cpu(cpu_softirq_time, cpu) +
1872 per_cpu(cpu_hardirq_time, cpu);
1873 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1877 #else /* CONFIG_64BIT */
1878 static inline void irq_time_write_begin(void)
1882 static inline void irq_time_write_end(void)
1886 static inline u64 irq_time_read(int cpu)
1888 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1890 #endif /* CONFIG_64BIT */
1893 * Called before incrementing preempt_count on {soft,}irq_enter
1894 * and before decrementing preempt_count on {soft,}irq_exit.
1896 void account_system_vtime(struct task_struct *curr)
1898 unsigned long flags;
1902 if (!sched_clock_irqtime)
1905 local_irq_save(flags);
1907 cpu = smp_processor_id();
1908 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1909 __this_cpu_add(irq_start_time, delta);
1911 irq_time_write_begin();
1913 * We do not account for softirq time from ksoftirqd here.
1914 * We want to continue accounting softirq time to ksoftirqd thread
1915 * in that case, so as not to confuse scheduler with a special task
1916 * that do not consume any time, but still wants to run.
1918 if (hardirq_count())
1919 __this_cpu_add(cpu_hardirq_time, delta);
1920 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
1921 __this_cpu_add(cpu_softirq_time, delta);
1923 irq_time_write_end();
1924 local_irq_restore(flags);
1926 EXPORT_SYMBOL_GPL(account_system_vtime);
1928 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
1930 #ifdef CONFIG_PARAVIRT
1931 static inline u64 steal_ticks(u64 steal)
1933 if (unlikely(steal > NSEC_PER_SEC))
1934 return div_u64(steal, TICK_NSEC);
1936 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
1940 static void update_rq_clock_task(struct rq *rq, s64 delta)
1943 * In theory, the compile should just see 0 here, and optimize out the call
1944 * to sched_rt_avg_update. But I don't trust it...
1946 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
1947 s64 steal = 0, irq_delta = 0;
1949 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1950 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
1953 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1954 * this case when a previous update_rq_clock() happened inside a
1955 * {soft,}irq region.
1957 * When this happens, we stop ->clock_task and only update the
1958 * prev_irq_time stamp to account for the part that fit, so that a next
1959 * update will consume the rest. This ensures ->clock_task is
1962 * It does however cause some slight miss-attribution of {soft,}irq
1963 * time, a more accurate solution would be to update the irq_time using
1964 * the current rq->clock timestamp, except that would require using
1967 if (irq_delta > delta)
1970 rq->prev_irq_time += irq_delta;
1973 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
1974 if (static_branch((¶virt_steal_rq_enabled))) {
1977 steal = paravirt_steal_clock(cpu_of(rq));
1978 steal -= rq->prev_steal_time_rq;
1980 if (unlikely(steal > delta))
1983 st = steal_ticks(steal);
1984 steal = st * TICK_NSEC;
1986 rq->prev_steal_time_rq += steal;
1992 rq->clock_task += delta;
1994 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
1995 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
1996 sched_rt_avg_update(rq, irq_delta + steal);
2000 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2001 static int irqtime_account_hi_update(void)
2003 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2004 unsigned long flags;
2008 local_irq_save(flags);
2009 latest_ns = this_cpu_read(cpu_hardirq_time);
2010 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
2012 local_irq_restore(flags);
2016 static int irqtime_account_si_update(void)
2018 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2019 unsigned long flags;
2023 local_irq_save(flags);
2024 latest_ns = this_cpu_read(cpu_softirq_time);
2025 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2027 local_irq_restore(flags);
2031 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2033 #define sched_clock_irqtime (0)
2037 #include "sched_idletask.c"
2038 #include "sched_fair.c"
2039 #include "sched_rt.c"
2040 #include "sched_autogroup.c"
2041 #include "sched_stoptask.c"
2042 #ifdef CONFIG_SCHED_DEBUG
2043 # include "sched_debug.c"
2046 void sched_set_stop_task(int cpu, struct task_struct *stop)
2048 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2049 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2053 * Make it appear like a SCHED_FIFO task, its something
2054 * userspace knows about and won't get confused about.
2056 * Also, it will make PI more or less work without too
2057 * much confusion -- but then, stop work should not
2058 * rely on PI working anyway.
2060 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2062 stop->sched_class = &stop_sched_class;
2065 cpu_rq(cpu)->stop = stop;
2069 * Reset it back to a normal scheduling class so that
2070 * it can die in pieces.
2072 old_stop->sched_class = &rt_sched_class;
2077 * __normal_prio - return the priority that is based on the static prio
2079 static inline int __normal_prio(struct task_struct *p)
2081 return p->static_prio;
2085 * Calculate the expected normal priority: i.e. priority
2086 * without taking RT-inheritance into account. Might be
2087 * boosted by interactivity modifiers. Changes upon fork,
2088 * setprio syscalls, and whenever the interactivity
2089 * estimator recalculates.
2091 static inline int normal_prio(struct task_struct *p)
2095 if (task_has_rt_policy(p))
2096 prio = MAX_RT_PRIO-1 - p->rt_priority;
2098 prio = __normal_prio(p);
2103 * Calculate the current priority, i.e. the priority
2104 * taken into account by the scheduler. This value might
2105 * be boosted by RT tasks, or might be boosted by
2106 * interactivity modifiers. Will be RT if the task got
2107 * RT-boosted. If not then it returns p->normal_prio.
2109 static int effective_prio(struct task_struct *p)
2111 p->normal_prio = normal_prio(p);
2113 * If we are RT tasks or we were boosted to RT priority,
2114 * keep the priority unchanged. Otherwise, update priority
2115 * to the normal priority:
2117 if (!rt_prio(p->prio))
2118 return p->normal_prio;
2123 * task_curr - is this task currently executing on a CPU?
2124 * @p: the task in question.
2126 inline int task_curr(const struct task_struct *p)
2128 return cpu_curr(task_cpu(p)) == p;
2131 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2132 const struct sched_class *prev_class,
2135 if (prev_class != p->sched_class) {
2136 if (prev_class->switched_from)
2137 prev_class->switched_from(rq, p);
2138 p->sched_class->switched_to(rq, p);
2139 } else if (oldprio != p->prio)
2140 p->sched_class->prio_changed(rq, p, oldprio);
2143 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2145 const struct sched_class *class;
2147 if (p->sched_class == rq->curr->sched_class) {
2148 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2150 for_each_class(class) {
2151 if (class == rq->curr->sched_class)
2153 if (class == p->sched_class) {
2154 resched_task(rq->curr);
2161 * A queue event has occurred, and we're going to schedule. In
2162 * this case, we can save a useless back to back clock update.
2164 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2165 rq->skip_clock_update = 1;
2170 * Is this task likely cache-hot:
2173 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2177 if (p->sched_class != &fair_sched_class)
2180 if (unlikely(p->policy == SCHED_IDLE))
2184 * Buddy candidates are cache hot:
2186 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2187 (&p->se == cfs_rq_of(&p->se)->next ||
2188 &p->se == cfs_rq_of(&p->se)->last))
2191 if (sysctl_sched_migration_cost == -1)
2193 if (sysctl_sched_migration_cost == 0)
2196 delta = now - p->se.exec_start;
2198 return delta < (s64)sysctl_sched_migration_cost;
2201 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2203 #ifdef CONFIG_SCHED_DEBUG
2205 * We should never call set_task_cpu() on a blocked task,
2206 * ttwu() will sort out the placement.
2208 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2209 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2211 #ifdef CONFIG_LOCKDEP
2213 * The caller should hold either p->pi_lock or rq->lock, when changing
2214 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2216 * sched_move_task() holds both and thus holding either pins the cgroup,
2217 * see set_task_rq().
2219 * Furthermore, all task_rq users should acquire both locks, see
2222 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2223 lockdep_is_held(&task_rq(p)->lock)));
2227 trace_sched_migrate_task(p, new_cpu);
2229 if (task_cpu(p) != new_cpu) {
2230 p->se.nr_migrations++;
2231 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
2234 __set_task_cpu(p, new_cpu);
2237 struct migration_arg {
2238 struct task_struct *task;
2242 static int migration_cpu_stop(void *data);
2245 * wait_task_inactive - wait for a thread to unschedule.
2247 * If @match_state is nonzero, it's the @p->state value just checked and
2248 * not expected to change. If it changes, i.e. @p might have woken up,
2249 * then return zero. When we succeed in waiting for @p to be off its CPU,
2250 * we return a positive number (its total switch count). If a second call
2251 * a short while later returns the same number, the caller can be sure that
2252 * @p has remained unscheduled the whole time.
2254 * The caller must ensure that the task *will* unschedule sometime soon,
2255 * else this function might spin for a *long* time. This function can't
2256 * be called with interrupts off, or it may introduce deadlock with
2257 * smp_call_function() if an IPI is sent by the same process we are
2258 * waiting to become inactive.
2260 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2262 unsigned long flags;
2269 * We do the initial early heuristics without holding
2270 * any task-queue locks at all. We'll only try to get
2271 * the runqueue lock when things look like they will
2277 * If the task is actively running on another CPU
2278 * still, just relax and busy-wait without holding
2281 * NOTE! Since we don't hold any locks, it's not
2282 * even sure that "rq" stays as the right runqueue!
2283 * But we don't care, since "task_running()" will
2284 * return false if the runqueue has changed and p
2285 * is actually now running somewhere else!
2287 while (task_running(rq, p)) {
2288 if (match_state && unlikely(p->state != match_state))
2294 * Ok, time to look more closely! We need the rq
2295 * lock now, to be *sure*. If we're wrong, we'll
2296 * just go back and repeat.
2298 rq = task_rq_lock(p, &flags);
2299 trace_sched_wait_task(p);
2300 running = task_running(rq, p);
2303 if (!match_state || p->state == match_state)
2304 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2305 task_rq_unlock(rq, p, &flags);
2308 * If it changed from the expected state, bail out now.
2310 if (unlikely(!ncsw))
2314 * Was it really running after all now that we
2315 * checked with the proper locks actually held?
2317 * Oops. Go back and try again..
2319 if (unlikely(running)) {
2325 * It's not enough that it's not actively running,
2326 * it must be off the runqueue _entirely_, and not
2329 * So if it was still runnable (but just not actively
2330 * running right now), it's preempted, and we should
2331 * yield - it could be a while.
2333 if (unlikely(on_rq)) {
2334 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2336 set_current_state(TASK_UNINTERRUPTIBLE);
2337 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2342 * Ahh, all good. It wasn't running, and it wasn't
2343 * runnable, which means that it will never become
2344 * running in the future either. We're all done!
2353 * kick_process - kick a running thread to enter/exit the kernel
2354 * @p: the to-be-kicked thread
2356 * Cause a process which is running on another CPU to enter
2357 * kernel-mode, without any delay. (to get signals handled.)
2359 * NOTE: this function doesn't have to take the runqueue lock,
2360 * because all it wants to ensure is that the remote task enters
2361 * the kernel. If the IPI races and the task has been migrated
2362 * to another CPU then no harm is done and the purpose has been
2365 void kick_process(struct task_struct *p)
2371 if ((cpu != smp_processor_id()) && task_curr(p))
2372 smp_send_reschedule(cpu);
2375 EXPORT_SYMBOL_GPL(kick_process);
2376 #endif /* CONFIG_SMP */
2380 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2382 static int select_fallback_rq(int cpu, struct task_struct *p)
2385 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2387 /* Look for allowed, online CPU in same node. */
2388 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2389 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2392 /* Any allowed, online CPU? */
2393 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2394 if (dest_cpu < nr_cpu_ids)
2397 /* No more Mr. Nice Guy. */
2398 dest_cpu = cpuset_cpus_allowed_fallback(p);
2400 * Don't tell them about moving exiting tasks or
2401 * kernel threads (both mm NULL), since they never
2404 if (p->mm && printk_ratelimit()) {
2405 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2406 task_pid_nr(p), p->comm, cpu);
2413 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2416 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2418 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2421 * In order not to call set_task_cpu() on a blocking task we need
2422 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2425 * Since this is common to all placement strategies, this lives here.
2427 * [ this allows ->select_task() to simply return task_cpu(p) and
2428 * not worry about this generic constraint ]
2430 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2432 cpu = select_fallback_rq(task_cpu(p), p);
2437 static void update_avg(u64 *avg, u64 sample)
2439 s64 diff = sample - *avg;
2445 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2447 #ifdef CONFIG_SCHEDSTATS
2448 struct rq *rq = this_rq();
2451 int this_cpu = smp_processor_id();
2453 if (cpu == this_cpu) {
2454 schedstat_inc(rq, ttwu_local);
2455 schedstat_inc(p, se.statistics.nr_wakeups_local);
2457 struct sched_domain *sd;
2459 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2461 for_each_domain(this_cpu, sd) {
2462 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2463 schedstat_inc(sd, ttwu_wake_remote);
2470 if (wake_flags & WF_MIGRATED)
2471 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2473 #endif /* CONFIG_SMP */
2475 schedstat_inc(rq, ttwu_count);
2476 schedstat_inc(p, se.statistics.nr_wakeups);
2478 if (wake_flags & WF_SYNC)
2479 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2481 #endif /* CONFIG_SCHEDSTATS */
2484 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2486 activate_task(rq, p, en_flags);
2489 /* if a worker is waking up, notify workqueue */
2490 if (p->flags & PF_WQ_WORKER)
2491 wq_worker_waking_up(p, cpu_of(rq));
2495 * Mark the task runnable and perform wakeup-preemption.
2498 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2500 trace_sched_wakeup(p, true);
2501 check_preempt_curr(rq, p, wake_flags);
2503 p->state = TASK_RUNNING;
2505 if (p->sched_class->task_woken)
2506 p->sched_class->task_woken(rq, p);
2508 if (rq->idle_stamp) {
2509 u64 delta = rq->clock - rq->idle_stamp;
2510 u64 max = 2*sysctl_sched_migration_cost;
2515 update_avg(&rq->avg_idle, delta);
2522 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2525 if (p->sched_contributes_to_load)
2526 rq->nr_uninterruptible--;
2529 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2530 ttwu_do_wakeup(rq, p, wake_flags);
2534 * Called in case the task @p isn't fully descheduled from its runqueue,
2535 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2536 * since all we need to do is flip p->state to TASK_RUNNING, since
2537 * the task is still ->on_rq.
2539 static int ttwu_remote(struct task_struct *p, int wake_flags)
2544 rq = __task_rq_lock(p);
2546 ttwu_do_wakeup(rq, p, wake_flags);
2549 __task_rq_unlock(rq);
2555 static void sched_ttwu_do_pending(struct task_struct *list)
2557 struct rq *rq = this_rq();
2559 raw_spin_lock(&rq->lock);
2562 struct task_struct *p = list;
2563 list = list->wake_entry;
2564 ttwu_do_activate(rq, p, 0);
2567 raw_spin_unlock(&rq->lock);
2570 #ifdef CONFIG_HOTPLUG_CPU
2572 static void sched_ttwu_pending(void)
2574 struct rq *rq = this_rq();
2575 struct task_struct *list = xchg(&rq->wake_list, NULL);
2580 sched_ttwu_do_pending(list);
2583 #endif /* CONFIG_HOTPLUG_CPU */
2585 void scheduler_ipi(void)
2587 struct rq *rq = this_rq();
2588 struct task_struct *list = xchg(&rq->wake_list, NULL);
2594 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2595 * traditionally all their work was done from the interrupt return
2596 * path. Now that we actually do some work, we need to make sure
2599 * Some archs already do call them, luckily irq_enter/exit nest
2602 * Arguably we should visit all archs and update all handlers,
2603 * however a fair share of IPIs are still resched only so this would
2604 * somewhat pessimize the simple resched case.
2607 sched_ttwu_do_pending(list);
2611 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2613 struct rq *rq = cpu_rq(cpu);
2614 struct task_struct *next = rq->wake_list;
2617 struct task_struct *old = next;
2619 p->wake_entry = next;
2620 next = cmpxchg(&rq->wake_list, old, p);
2626 smp_send_reschedule(cpu);
2629 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2630 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2635 rq = __task_rq_lock(p);
2637 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2638 ttwu_do_wakeup(rq, p, wake_flags);
2641 __task_rq_unlock(rq);
2646 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2647 #endif /* CONFIG_SMP */
2649 static void ttwu_queue(struct task_struct *p, int cpu)
2651 struct rq *rq = cpu_rq(cpu);
2653 #if defined(CONFIG_SMP)
2654 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2655 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2656 ttwu_queue_remote(p, cpu);
2661 raw_spin_lock(&rq->lock);
2662 ttwu_do_activate(rq, p, 0);
2663 raw_spin_unlock(&rq->lock);
2667 * try_to_wake_up - wake up a thread
2668 * @p: the thread to be awakened
2669 * @state: the mask of task states that can be woken
2670 * @wake_flags: wake modifier flags (WF_*)
2672 * Put it on the run-queue if it's not already there. The "current"
2673 * thread is always on the run-queue (except when the actual
2674 * re-schedule is in progress), and as such you're allowed to do
2675 * the simpler "current->state = TASK_RUNNING" to mark yourself
2676 * runnable without the overhead of this.
2678 * Returns %true if @p was woken up, %false if it was already running
2679 * or @state didn't match @p's state.
2682 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2684 unsigned long flags;
2685 int cpu, success = 0;
2688 raw_spin_lock_irqsave(&p->pi_lock, flags);
2689 if (!(p->state & state))
2692 success = 1; /* we're going to change ->state */
2695 if (p->on_rq && ttwu_remote(p, wake_flags))
2700 * If the owning (remote) cpu is still in the middle of schedule() with
2701 * this task as prev, wait until its done referencing the task.
2704 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2706 * In case the architecture enables interrupts in
2707 * context_switch(), we cannot busy wait, since that
2708 * would lead to deadlocks when an interrupt hits and
2709 * tries to wake up @prev. So bail and do a complete
2712 if (ttwu_activate_remote(p, wake_flags))
2719 * Pairs with the smp_wmb() in finish_lock_switch().
2723 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2724 p->state = TASK_WAKING;
2726 if (p->sched_class->task_waking)
2727 p->sched_class->task_waking(p);
2729 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2730 if (task_cpu(p) != cpu) {
2731 wake_flags |= WF_MIGRATED;
2732 set_task_cpu(p, cpu);
2734 #endif /* CONFIG_SMP */
2738 ttwu_stat(p, cpu, wake_flags);
2740 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2746 * try_to_wake_up_local - try to wake up a local task with rq lock held
2747 * @p: the thread to be awakened
2749 * Put @p on the run-queue if it's not already there. The caller must
2750 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2753 static void try_to_wake_up_local(struct task_struct *p)
2755 struct rq *rq = task_rq(p);
2757 BUG_ON(rq != this_rq());
2758 BUG_ON(p == current);
2759 lockdep_assert_held(&rq->lock);
2761 if (!raw_spin_trylock(&p->pi_lock)) {
2762 raw_spin_unlock(&rq->lock);
2763 raw_spin_lock(&p->pi_lock);
2764 raw_spin_lock(&rq->lock);
2767 if (!(p->state & TASK_NORMAL))
2771 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2773 ttwu_do_wakeup(rq, p, 0);
2774 ttwu_stat(p, smp_processor_id(), 0);
2776 raw_spin_unlock(&p->pi_lock);
2780 * wake_up_process - Wake up a specific process
2781 * @p: The process to be woken up.
2783 * Attempt to wake up the nominated process and move it to the set of runnable
2784 * processes. Returns 1 if the process was woken up, 0 if it was already
2787 * It may be assumed that this function implies a write memory barrier before
2788 * changing the task state if and only if any tasks are woken up.
2790 int wake_up_process(struct task_struct *p)
2792 return try_to_wake_up(p, TASK_ALL, 0);
2794 EXPORT_SYMBOL(wake_up_process);
2796 int wake_up_state(struct task_struct *p, unsigned int state)
2798 return try_to_wake_up(p, state, 0);
2802 * Perform scheduler related setup for a newly forked process p.
2803 * p is forked by current.
2805 * __sched_fork() is basic setup used by init_idle() too:
2807 static void __sched_fork(struct task_struct *p)
2812 p->se.exec_start = 0;
2813 p->se.sum_exec_runtime = 0;
2814 p->se.prev_sum_exec_runtime = 0;
2815 p->se.nr_migrations = 0;
2817 INIT_LIST_HEAD(&p->se.group_node);
2819 #ifdef CONFIG_SCHEDSTATS
2820 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2823 INIT_LIST_HEAD(&p->rt.run_list);
2825 #ifdef CONFIG_PREEMPT_NOTIFIERS
2826 INIT_HLIST_HEAD(&p->preempt_notifiers);
2831 * fork()/clone()-time setup:
2833 void sched_fork(struct task_struct *p)
2835 unsigned long flags;
2836 int cpu = get_cpu();
2840 * We mark the process as running here. This guarantees that
2841 * nobody will actually run it, and a signal or other external
2842 * event cannot wake it up and insert it on the runqueue either.
2844 p->state = TASK_RUNNING;
2847 * Revert to default priority/policy on fork if requested.
2849 if (unlikely(p->sched_reset_on_fork)) {
2850 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2851 p->policy = SCHED_NORMAL;
2852 p->normal_prio = p->static_prio;
2855 if (PRIO_TO_NICE(p->static_prio) < 0) {
2856 p->static_prio = NICE_TO_PRIO(0);
2857 p->normal_prio = p->static_prio;
2862 * We don't need the reset flag anymore after the fork. It has
2863 * fulfilled its duty:
2865 p->sched_reset_on_fork = 0;
2869 * Make sure we do not leak PI boosting priority to the child.
2871 p->prio = current->normal_prio;
2873 if (!rt_prio(p->prio))
2874 p->sched_class = &fair_sched_class;
2876 if (p->sched_class->task_fork)
2877 p->sched_class->task_fork(p);
2880 * The child is not yet in the pid-hash so no cgroup attach races,
2881 * and the cgroup is pinned to this child due to cgroup_fork()
2882 * is ran before sched_fork().
2884 * Silence PROVE_RCU.
2886 raw_spin_lock_irqsave(&p->pi_lock, flags);
2887 set_task_cpu(p, cpu);
2888 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2890 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2891 if (likely(sched_info_on()))
2892 memset(&p->sched_info, 0, sizeof(p->sched_info));
2894 #if defined(CONFIG_SMP)
2897 #ifdef CONFIG_PREEMPT_COUNT
2898 /* Want to start with kernel preemption disabled. */
2899 task_thread_info(p)->preempt_count = 1;
2902 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2909 * wake_up_new_task - wake up a newly created task for the first time.
2911 * This function will do some initial scheduler statistics housekeeping
2912 * that must be done for every newly created context, then puts the task
2913 * on the runqueue and wakes it.
2915 void wake_up_new_task(struct task_struct *p)
2917 unsigned long flags;
2920 raw_spin_lock_irqsave(&p->pi_lock, flags);
2923 * Fork balancing, do it here and not earlier because:
2924 * - cpus_allowed can change in the fork path
2925 * - any previously selected cpu might disappear through hotplug
2927 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
2930 rq = __task_rq_lock(p);
2931 activate_task(rq, p, 0);
2933 trace_sched_wakeup_new(p, true);
2934 check_preempt_curr(rq, p, WF_FORK);
2936 if (p->sched_class->task_woken)
2937 p->sched_class->task_woken(rq, p);
2939 task_rq_unlock(rq, p, &flags);
2942 #ifdef CONFIG_PREEMPT_NOTIFIERS
2945 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2946 * @notifier: notifier struct to register
2948 void preempt_notifier_register(struct preempt_notifier *notifier)
2950 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2952 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2955 * preempt_notifier_unregister - no longer interested in preemption notifications
2956 * @notifier: notifier struct to unregister
2958 * This is safe to call from within a preemption notifier.
2960 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2962 hlist_del(¬ifier->link);
2964 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2966 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2968 struct preempt_notifier *notifier;
2969 struct hlist_node *node;
2971 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2972 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2976 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2977 struct task_struct *next)
2979 struct preempt_notifier *notifier;
2980 struct hlist_node *node;
2982 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2983 notifier->ops->sched_out(notifier, next);
2986 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2988 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2993 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2994 struct task_struct *next)
2998 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3001 * prepare_task_switch - prepare to switch tasks
3002 * @rq: the runqueue preparing to switch
3003 * @prev: the current task that is being switched out
3004 * @next: the task we are going to switch to.
3006 * This is called with the rq lock held and interrupts off. It must
3007 * be paired with a subsequent finish_task_switch after the context
3010 * prepare_task_switch sets up locking and calls architecture specific
3014 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3015 struct task_struct *next)
3017 sched_info_switch(prev, next);
3018 perf_event_task_sched_out(prev, next);
3019 fire_sched_out_preempt_notifiers(prev, next);
3020 prepare_lock_switch(rq, next);
3021 prepare_arch_switch(next);
3022 trace_sched_switch(prev, next);
3026 * finish_task_switch - clean up after a task-switch
3027 * @rq: runqueue associated with task-switch
3028 * @prev: the thread we just switched away from.
3030 * finish_task_switch must be called after the context switch, paired
3031 * with a prepare_task_switch call before the context switch.
3032 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3033 * and do any other architecture-specific cleanup actions.
3035 * Note that we may have delayed dropping an mm in context_switch(). If
3036 * so, we finish that here outside of the runqueue lock. (Doing it
3037 * with the lock held can cause deadlocks; see schedule() for
3040 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3041 __releases(rq->lock)
3043 struct mm_struct *mm = rq->prev_mm;
3049 * A task struct has one reference for the use as "current".
3050 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3051 * schedule one last time. The schedule call will never return, and
3052 * the scheduled task must drop that reference.
3053 * The test for TASK_DEAD must occur while the runqueue locks are
3054 * still held, otherwise prev could be scheduled on another cpu, die
3055 * there before we look at prev->state, and then the reference would
3057 * Manfred Spraul <manfred@colorfullife.com>
3059 prev_state = prev->state;
3060 finish_arch_switch(prev);
3061 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3062 local_irq_disable();
3063 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3064 perf_event_task_sched_in(current);
3065 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3067 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3068 finish_lock_switch(rq, prev);
3070 fire_sched_in_preempt_notifiers(current);
3073 if (unlikely(prev_state == TASK_DEAD)) {
3075 * Remove function-return probe instances associated with this
3076 * task and put them back on the free list.
3078 kprobe_flush_task(prev);
3079 put_task_struct(prev);
3085 /* assumes rq->lock is held */
3086 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3088 if (prev->sched_class->pre_schedule)
3089 prev->sched_class->pre_schedule(rq, prev);
3092 /* rq->lock is NOT held, but preemption is disabled */
3093 static inline void post_schedule(struct rq *rq)
3095 if (rq->post_schedule) {
3096 unsigned long flags;
3098 raw_spin_lock_irqsave(&rq->lock, flags);
3099 if (rq->curr->sched_class->post_schedule)
3100 rq->curr->sched_class->post_schedule(rq);
3101 raw_spin_unlock_irqrestore(&rq->lock, flags);
3103 rq->post_schedule = 0;
3109 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3113 static inline void post_schedule(struct rq *rq)
3120 * schedule_tail - first thing a freshly forked thread must call.
3121 * @prev: the thread we just switched away from.
3123 asmlinkage void schedule_tail(struct task_struct *prev)
3124 __releases(rq->lock)
3126 struct rq *rq = this_rq();
3128 finish_task_switch(rq, prev);
3131 * FIXME: do we need to worry about rq being invalidated by the
3136 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3137 /* In this case, finish_task_switch does not reenable preemption */
3140 if (current->set_child_tid)
3141 put_user(task_pid_vnr(current), current->set_child_tid);
3145 * context_switch - switch to the new MM and the new
3146 * thread's register state.
3149 context_switch(struct rq *rq, struct task_struct *prev,
3150 struct task_struct *next)
3152 struct mm_struct *mm, *oldmm;
3154 prepare_task_switch(rq, prev, next);
3157 oldmm = prev->active_mm;
3159 * For paravirt, this is coupled with an exit in switch_to to
3160 * combine the page table reload and the switch backend into
3163 arch_start_context_switch(prev);
3166 next->active_mm = oldmm;
3167 atomic_inc(&oldmm->mm_count);
3168 enter_lazy_tlb(oldmm, next);
3170 switch_mm(oldmm, mm, next);
3173 prev->active_mm = NULL;
3174 rq->prev_mm = oldmm;
3177 * Since the runqueue lock will be released by the next
3178 * task (which is an invalid locking op but in the case
3179 * of the scheduler it's an obvious special-case), so we
3180 * do an early lockdep release here:
3182 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3183 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3186 /* Here we just switch the register state and the stack. */
3187 switch_to(prev, next, prev);
3191 * this_rq must be evaluated again because prev may have moved
3192 * CPUs since it called schedule(), thus the 'rq' on its stack
3193 * frame will be invalid.
3195 finish_task_switch(this_rq(), prev);
3199 * nr_running, nr_uninterruptible and nr_context_switches:
3201 * externally visible scheduler statistics: current number of runnable
3202 * threads, current number of uninterruptible-sleeping threads, total
3203 * number of context switches performed since bootup.
3205 unsigned long nr_running(void)
3207 unsigned long i, sum = 0;
3209 for_each_online_cpu(i)
3210 sum += cpu_rq(i)->nr_running;
3215 unsigned long nr_uninterruptible(void)
3217 unsigned long i, sum = 0;
3219 for_each_possible_cpu(i)
3220 sum += cpu_rq(i)->nr_uninterruptible;
3223 * Since we read the counters lockless, it might be slightly
3224 * inaccurate. Do not allow it to go below zero though:
3226 if (unlikely((long)sum < 0))
3232 unsigned long long nr_context_switches(void)
3235 unsigned long long sum = 0;
3237 for_each_possible_cpu(i)
3238 sum += cpu_rq(i)->nr_switches;
3243 unsigned long nr_iowait(void)
3245 unsigned long i, sum = 0;
3247 for_each_possible_cpu(i)
3248 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3253 unsigned long nr_iowait_cpu(int cpu)
3255 struct rq *this = cpu_rq(cpu);
3256 return atomic_read(&this->nr_iowait);
3259 unsigned long this_cpu_load(void)
3261 struct rq *this = this_rq();
3262 return this->cpu_load[0];
3266 /* Variables and functions for calc_load */
3267 static atomic_long_t calc_load_tasks;
3268 static unsigned long calc_load_update;
3269 unsigned long avenrun[3];
3270 EXPORT_SYMBOL(avenrun);
3272 static long calc_load_fold_active(struct rq *this_rq)
3274 long nr_active, delta = 0;
3276 nr_active = this_rq->nr_running;
3277 nr_active += (long) this_rq->nr_uninterruptible;
3279 if (nr_active != this_rq->calc_load_active) {
3280 delta = nr_active - this_rq->calc_load_active;
3281 this_rq->calc_load_active = nr_active;
3287 static unsigned long
3288 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3291 load += active * (FIXED_1 - exp);
3292 load += 1UL << (FSHIFT - 1);
3293 return load >> FSHIFT;
3298 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3300 * When making the ILB scale, we should try to pull this in as well.
3302 static atomic_long_t calc_load_tasks_idle;
3304 static void calc_load_account_idle(struct rq *this_rq)
3308 delta = calc_load_fold_active(this_rq);
3310 atomic_long_add(delta, &calc_load_tasks_idle);
3313 static long calc_load_fold_idle(void)
3318 * Its got a race, we don't care...
3320 if (atomic_long_read(&calc_load_tasks_idle))
3321 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3327 * fixed_power_int - compute: x^n, in O(log n) time
3329 * @x: base of the power
3330 * @frac_bits: fractional bits of @x
3331 * @n: power to raise @x to.
3333 * By exploiting the relation between the definition of the natural power
3334 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3335 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3336 * (where: n_i \elem {0, 1}, the binary vector representing n),
3337 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3338 * of course trivially computable in O(log_2 n), the length of our binary
3341 static unsigned long
3342 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3344 unsigned long result = 1UL << frac_bits;
3349 result += 1UL << (frac_bits - 1);
3350 result >>= frac_bits;
3356 x += 1UL << (frac_bits - 1);
3364 * a1 = a0 * e + a * (1 - e)
3366 * a2 = a1 * e + a * (1 - e)
3367 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3368 * = a0 * e^2 + a * (1 - e) * (1 + e)
3370 * a3 = a2 * e + a * (1 - e)
3371 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3372 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3376 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3377 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3378 * = a0 * e^n + a * (1 - e^n)
3380 * [1] application of the geometric series:
3383 * S_n := \Sum x^i = -------------
3386 static unsigned long
3387 calc_load_n(unsigned long load, unsigned long exp,
3388 unsigned long active, unsigned int n)
3391 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3395 * NO_HZ can leave us missing all per-cpu ticks calling
3396 * calc_load_account_active(), but since an idle CPU folds its delta into
3397 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3398 * in the pending idle delta if our idle period crossed a load cycle boundary.
3400 * Once we've updated the global active value, we need to apply the exponential
3401 * weights adjusted to the number of cycles missed.
3403 static void calc_global_nohz(unsigned long ticks)
3405 long delta, active, n;
3407 if (time_before(jiffies, calc_load_update))
3411 * If we crossed a calc_load_update boundary, make sure to fold
3412 * any pending idle changes, the respective CPUs might have
3413 * missed the tick driven calc_load_account_active() update
3416 delta = calc_load_fold_idle();
3418 atomic_long_add(delta, &calc_load_tasks);
3421 * If we were idle for multiple load cycles, apply them.
3423 if (ticks >= LOAD_FREQ) {
3424 n = ticks / LOAD_FREQ;
3426 active = atomic_long_read(&calc_load_tasks);
3427 active = active > 0 ? active * FIXED_1 : 0;
3429 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3430 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3431 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3433 calc_load_update += n * LOAD_FREQ;
3437 * Its possible the remainder of the above division also crosses
3438 * a LOAD_FREQ period, the regular check in calc_global_load()
3439 * which comes after this will take care of that.
3441 * Consider us being 11 ticks before a cycle completion, and us
3442 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3443 * age us 4 cycles, and the test in calc_global_load() will
3444 * pick up the final one.
3448 static void calc_load_account_idle(struct rq *this_rq)
3452 static inline long calc_load_fold_idle(void)
3457 static void calc_global_nohz(unsigned long ticks)
3463 * get_avenrun - get the load average array
3464 * @loads: pointer to dest load array
3465 * @offset: offset to add
3466 * @shift: shift count to shift the result left
3468 * These values are estimates at best, so no need for locking.
3470 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3472 loads[0] = (avenrun[0] + offset) << shift;
3473 loads[1] = (avenrun[1] + offset) << shift;
3474 loads[2] = (avenrun[2] + offset) << shift;
3478 * calc_load - update the avenrun load estimates 10 ticks after the
3479 * CPUs have updated calc_load_tasks.
3481 void calc_global_load(unsigned long ticks)
3485 calc_global_nohz(ticks);
3487 if (time_before(jiffies, calc_load_update + 10))
3490 active = atomic_long_read(&calc_load_tasks);
3491 active = active > 0 ? active * FIXED_1 : 0;
3493 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3494 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3495 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3497 calc_load_update += LOAD_FREQ;
3501 * Called from update_cpu_load() to periodically update this CPU's
3504 static void calc_load_account_active(struct rq *this_rq)
3508 if (time_before(jiffies, this_rq->calc_load_update))
3511 delta = calc_load_fold_active(this_rq);
3512 delta += calc_load_fold_idle();
3514 atomic_long_add(delta, &calc_load_tasks);
3516 this_rq->calc_load_update += LOAD_FREQ;
3520 * The exact cpuload at various idx values, calculated at every tick would be
3521 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3523 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3524 * on nth tick when cpu may be busy, then we have:
3525 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3526 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3528 * decay_load_missed() below does efficient calculation of
3529 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3530 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3532 * The calculation is approximated on a 128 point scale.
3533 * degrade_zero_ticks is the number of ticks after which load at any
3534 * particular idx is approximated to be zero.
3535 * degrade_factor is a precomputed table, a row for each load idx.
3536 * Each column corresponds to degradation factor for a power of two ticks,
3537 * based on 128 point scale.
3539 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3540 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3542 * With this power of 2 load factors, we can degrade the load n times
3543 * by looking at 1 bits in n and doing as many mult/shift instead of
3544 * n mult/shifts needed by the exact degradation.
3546 #define DEGRADE_SHIFT 7
3547 static const unsigned char
3548 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3549 static const unsigned char
3550 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3551 {0, 0, 0, 0, 0, 0, 0, 0},
3552 {64, 32, 8, 0, 0, 0, 0, 0},
3553 {96, 72, 40, 12, 1, 0, 0},
3554 {112, 98, 75, 43, 15, 1, 0},
3555 {120, 112, 98, 76, 45, 16, 2} };
3558 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3559 * would be when CPU is idle and so we just decay the old load without
3560 * adding any new load.
3562 static unsigned long
3563 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3567 if (!missed_updates)
3570 if (missed_updates >= degrade_zero_ticks[idx])
3574 return load >> missed_updates;
3576 while (missed_updates) {
3577 if (missed_updates % 2)
3578 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3580 missed_updates >>= 1;
3587 * Update rq->cpu_load[] statistics. This function is usually called every
3588 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3589 * every tick. We fix it up based on jiffies.
3591 static void update_cpu_load(struct rq *this_rq)
3593 unsigned long this_load = this_rq->load.weight;
3594 unsigned long curr_jiffies = jiffies;
3595 unsigned long pending_updates;
3598 this_rq->nr_load_updates++;
3600 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3601 if (curr_jiffies == this_rq->last_load_update_tick)
3604 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3605 this_rq->last_load_update_tick = curr_jiffies;
3607 /* Update our load: */
3608 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3609 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3610 unsigned long old_load, new_load;
3612 /* scale is effectively 1 << i now, and >> i divides by scale */
3614 old_load = this_rq->cpu_load[i];
3615 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3616 new_load = this_load;
3618 * Round up the averaging division if load is increasing. This
3619 * prevents us from getting stuck on 9 if the load is 10, for
3622 if (new_load > old_load)
3623 new_load += scale - 1;
3625 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3628 sched_avg_update(this_rq);
3631 static void update_cpu_load_active(struct rq *this_rq)
3633 update_cpu_load(this_rq);
3635 calc_load_account_active(this_rq);
3641 * sched_exec - execve() is a valuable balancing opportunity, because at
3642 * this point the task has the smallest effective memory and cache footprint.
3644 void sched_exec(void)
3646 struct task_struct *p = current;
3647 unsigned long flags;
3650 raw_spin_lock_irqsave(&p->pi_lock, flags);
3651 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3652 if (dest_cpu == smp_processor_id())
3655 if (likely(cpu_active(dest_cpu))) {
3656 struct migration_arg arg = { p, dest_cpu };
3658 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3659 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3663 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3668 DEFINE_PER_CPU(struct kernel_stat, kstat);
3670 EXPORT_PER_CPU_SYMBOL(kstat);
3673 * Return any ns on the sched_clock that have not yet been accounted in
3674 * @p in case that task is currently running.
3676 * Called with task_rq_lock() held on @rq.
3678 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3682 if (task_current(rq, p)) {
3683 update_rq_clock(rq);
3684 ns = rq->clock_task - p->se.exec_start;
3692 unsigned long long task_delta_exec(struct task_struct *p)
3694 unsigned long flags;
3698 rq = task_rq_lock(p, &flags);
3699 ns = do_task_delta_exec(p, rq);
3700 task_rq_unlock(rq, p, &flags);
3706 * Return accounted runtime for the task.
3707 * In case the task is currently running, return the runtime plus current's
3708 * pending runtime that have not been accounted yet.
3710 unsigned long long task_sched_runtime(struct task_struct *p)
3712 unsigned long flags;
3716 rq = task_rq_lock(p, &flags);
3717 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3718 task_rq_unlock(rq, p, &flags);
3724 * Return sum_exec_runtime for the thread group.
3725 * In case the task is currently running, return the sum plus current's
3726 * pending runtime that have not been accounted yet.
3728 * Note that the thread group might have other running tasks as well,
3729 * so the return value not includes other pending runtime that other
3730 * running tasks might have.
3732 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3734 struct task_cputime totals;
3735 unsigned long flags;
3739 rq = task_rq_lock(p, &flags);
3740 thread_group_cputime(p, &totals);
3741 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3742 task_rq_unlock(rq, p, &flags);
3748 * Account user cpu time to a process.
3749 * @p: the process that the cpu time gets accounted to
3750 * @cputime: the cpu time spent in user space since the last update
3751 * @cputime_scaled: cputime scaled by cpu frequency
3753 void account_user_time(struct task_struct *p, cputime_t cputime,
3754 cputime_t cputime_scaled)
3756 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3759 /* Add user time to process. */
3760 p->utime = cputime_add(p->utime, cputime);
3761 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3762 account_group_user_time(p, cputime);
3764 /* Add user time to cpustat. */
3765 tmp = cputime_to_cputime64(cputime);
3766 if (TASK_NICE(p) > 0)
3767 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3769 cpustat->user = cputime64_add(cpustat->user, tmp);
3771 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3772 /* Account for user time used */
3773 acct_update_integrals(p);
3777 * Account guest cpu time to a process.
3778 * @p: the process that the cpu time gets accounted to
3779 * @cputime: the cpu time spent in virtual machine since the last update
3780 * @cputime_scaled: cputime scaled by cpu frequency
3782 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3783 cputime_t cputime_scaled)
3786 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3788 tmp = cputime_to_cputime64(cputime);
3790 /* Add guest time to process. */
3791 p->utime = cputime_add(p->utime, cputime);
3792 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3793 account_group_user_time(p, cputime);
3794 p->gtime = cputime_add(p->gtime, cputime);
3796 /* Add guest time to cpustat. */
3797 if (TASK_NICE(p) > 0) {
3798 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3799 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3801 cpustat->user = cputime64_add(cpustat->user, tmp);
3802 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3807 * Account system cpu time to a process and desired cpustat field
3808 * @p: the process that the cpu time gets accounted to
3809 * @cputime: the cpu time spent in kernel space since the last update
3810 * @cputime_scaled: cputime scaled by cpu frequency
3811 * @target_cputime64: pointer to cpustat field that has to be updated
3814 void __account_system_time(struct task_struct *p, cputime_t cputime,
3815 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3817 cputime64_t tmp = cputime_to_cputime64(cputime);
3819 /* Add system time to process. */
3820 p->stime = cputime_add(p->stime, cputime);
3821 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3822 account_group_system_time(p, cputime);
3824 /* Add system time to cpustat. */
3825 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3826 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3828 /* Account for system time used */
3829 acct_update_integrals(p);
3833 * Account system cpu time to a process.
3834 * @p: the process that the cpu time gets accounted to
3835 * @hardirq_offset: the offset to subtract from hardirq_count()
3836 * @cputime: the cpu time spent in kernel space since the last update
3837 * @cputime_scaled: cputime scaled by cpu frequency
3839 void account_system_time(struct task_struct *p, int hardirq_offset,
3840 cputime_t cputime, cputime_t cputime_scaled)
3842 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3843 cputime64_t *target_cputime64;
3845 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3846 account_guest_time(p, cputime, cputime_scaled);
3850 if (hardirq_count() - hardirq_offset)
3851 target_cputime64 = &cpustat->irq;
3852 else if (in_serving_softirq())
3853 target_cputime64 = &cpustat->softirq;
3855 target_cputime64 = &cpustat->system;
3857 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3861 * Account for involuntary wait time.
3862 * @cputime: the cpu time spent in involuntary wait
3864 void account_steal_time(cputime_t cputime)
3866 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3867 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3869 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3873 * Account for idle time.
3874 * @cputime: the cpu time spent in idle wait
3876 void account_idle_time(cputime_t cputime)
3878 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3879 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3880 struct rq *rq = this_rq();
3882 if (atomic_read(&rq->nr_iowait) > 0)
3883 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3885 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3888 static __always_inline bool steal_account_process_tick(void)
3890 #ifdef CONFIG_PARAVIRT
3891 if (static_branch(¶virt_steal_enabled)) {
3894 steal = paravirt_steal_clock(smp_processor_id());
3895 steal -= this_rq()->prev_steal_time;
3897 st = steal_ticks(steal);
3898 this_rq()->prev_steal_time += st * TICK_NSEC;
3900 account_steal_time(st);
3907 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3909 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3911 * Account a tick to a process and cpustat
3912 * @p: the process that the cpu time gets accounted to
3913 * @user_tick: is the tick from userspace
3914 * @rq: the pointer to rq
3916 * Tick demultiplexing follows the order
3917 * - pending hardirq update
3918 * - pending softirq update
3922 * - check for guest_time
3923 * - else account as system_time
3925 * Check for hardirq is done both for system and user time as there is
3926 * no timer going off while we are on hardirq and hence we may never get an
3927 * opportunity to update it solely in system time.
3928 * p->stime and friends are only updated on system time and not on irq
3929 * softirq as those do not count in task exec_runtime any more.
3931 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3934 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3935 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
3936 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3938 if (steal_account_process_tick())
3941 if (irqtime_account_hi_update()) {
3942 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3943 } else if (irqtime_account_si_update()) {
3944 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3945 } else if (this_cpu_ksoftirqd() == p) {
3947 * ksoftirqd time do not get accounted in cpu_softirq_time.
3948 * So, we have to handle it separately here.
3949 * Also, p->stime needs to be updated for ksoftirqd.
3951 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3953 } else if (user_tick) {
3954 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3955 } else if (p == rq->idle) {
3956 account_idle_time(cputime_one_jiffy);
3957 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3958 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3960 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3965 static void irqtime_account_idle_ticks(int ticks)
3968 struct rq *rq = this_rq();
3970 for (i = 0; i < ticks; i++)
3971 irqtime_account_process_tick(current, 0, rq);
3973 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
3974 static void irqtime_account_idle_ticks(int ticks) {}
3975 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3977 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3980 * Account a single tick of cpu time.
3981 * @p: the process that the cpu time gets accounted to
3982 * @user_tick: indicates if the tick is a user or a system tick
3984 void account_process_tick(struct task_struct *p, int user_tick)
3986 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3987 struct rq *rq = this_rq();
3989 if (sched_clock_irqtime) {
3990 irqtime_account_process_tick(p, user_tick, rq);
3994 if (steal_account_process_tick())
3998 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3999 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
4000 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
4003 account_idle_time(cputime_one_jiffy);
4007 * Account multiple ticks of steal time.
4008 * @p: the process from which the cpu time has been stolen
4009 * @ticks: number of stolen ticks
4011 void account_steal_ticks(unsigned long ticks)
4013 account_steal_time(jiffies_to_cputime(ticks));
4017 * Account multiple ticks of idle time.
4018 * @ticks: number of stolen ticks
4020 void account_idle_ticks(unsigned long ticks)
4023 if (sched_clock_irqtime) {
4024 irqtime_account_idle_ticks(ticks);
4028 account_idle_time(jiffies_to_cputime(ticks));
4034 * Use precise platform statistics if available:
4036 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4037 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4043 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4045 struct task_cputime cputime;
4047 thread_group_cputime(p, &cputime);
4049 *ut = cputime.utime;
4050 *st = cputime.stime;
4054 #ifndef nsecs_to_cputime
4055 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
4058 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4060 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
4063 * Use CFS's precise accounting:
4065 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
4071 do_div(temp, total);
4072 utime = (cputime_t)temp;
4077 * Compare with previous values, to keep monotonicity:
4079 p->prev_utime = max(p->prev_utime, utime);
4080 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4082 *ut = p->prev_utime;
4083 *st = p->prev_stime;
4087 * Must be called with siglock held.
4089 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4091 struct signal_struct *sig = p->signal;
4092 struct task_cputime cputime;
4093 cputime_t rtime, utime, total;
4095 thread_group_cputime(p, &cputime);
4097 total = cputime_add(cputime.utime, cputime.stime);
4098 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4103 temp *= cputime.utime;
4104 do_div(temp, total);
4105 utime = (cputime_t)temp;
4109 sig->prev_utime = max(sig->prev_utime, utime);
4110 sig->prev_stime = max(sig->prev_stime,
4111 cputime_sub(rtime, sig->prev_utime));
4113 *ut = sig->prev_utime;
4114 *st = sig->prev_stime;
4119 * This function gets called by the timer code, with HZ frequency.
4120 * We call it with interrupts disabled.
4122 void scheduler_tick(void)
4124 int cpu = smp_processor_id();
4125 struct rq *rq = cpu_rq(cpu);
4126 struct task_struct *curr = rq->curr;
4130 raw_spin_lock(&rq->lock);
4131 update_rq_clock(rq);
4132 update_cpu_load_active(rq);
4133 curr->sched_class->task_tick(rq, curr, 0);
4134 raw_spin_unlock(&rq->lock);
4136 perf_event_task_tick();
4139 rq->idle_at_tick = idle_cpu(cpu);
4140 trigger_load_balance(rq, cpu);
4144 notrace unsigned long get_parent_ip(unsigned long addr)
4146 if (in_lock_functions(addr)) {
4147 addr = CALLER_ADDR2;
4148 if (in_lock_functions(addr))
4149 addr = CALLER_ADDR3;
4154 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4155 defined(CONFIG_PREEMPT_TRACER))
4157 void __kprobes add_preempt_count(int val)
4159 #ifdef CONFIG_DEBUG_PREEMPT
4163 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4166 preempt_count() += val;
4167 #ifdef CONFIG_DEBUG_PREEMPT
4169 * Spinlock count overflowing soon?
4171 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4174 if (preempt_count() == val)
4175 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4177 EXPORT_SYMBOL(add_preempt_count);
4179 void __kprobes sub_preempt_count(int val)
4181 #ifdef CONFIG_DEBUG_PREEMPT
4185 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4188 * Is the spinlock portion underflowing?
4190 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4191 !(preempt_count() & PREEMPT_MASK)))
4195 if (preempt_count() == val)
4196 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4197 preempt_count() -= val;
4199 EXPORT_SYMBOL(sub_preempt_count);
4204 * Print scheduling while atomic bug:
4206 static noinline void __schedule_bug(struct task_struct *prev)
4208 struct pt_regs *regs = get_irq_regs();
4210 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4211 prev->comm, prev->pid, preempt_count());
4213 debug_show_held_locks(prev);
4215 if (irqs_disabled())
4216 print_irqtrace_events(prev);
4225 * Various schedule()-time debugging checks and statistics:
4227 static inline void schedule_debug(struct task_struct *prev)
4230 * Test if we are atomic. Since do_exit() needs to call into
4231 * schedule() atomically, we ignore that path for now.
4232 * Otherwise, whine if we are scheduling when we should not be.
4234 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4235 __schedule_bug(prev);
4237 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4239 schedstat_inc(this_rq(), sched_count);
4242 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4244 if (prev->on_rq || rq->skip_clock_update < 0)
4245 update_rq_clock(rq);
4246 prev->sched_class->put_prev_task(rq, prev);
4250 * Pick up the highest-prio task:
4252 static inline struct task_struct *
4253 pick_next_task(struct rq *rq)
4255 const struct sched_class *class;
4256 struct task_struct *p;
4259 * Optimization: we know that if all tasks are in
4260 * the fair class we can call that function directly:
4262 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4263 p = fair_sched_class.pick_next_task(rq);
4268 for_each_class(class) {
4269 p = class->pick_next_task(rq);
4274 BUG(); /* the idle class will always have a runnable task */
4278 * schedule() is the main scheduler function.
4280 asmlinkage void __sched schedule(void)
4282 struct task_struct *prev, *next;
4283 unsigned long *switch_count;
4289 cpu = smp_processor_id();
4291 rcu_note_context_switch(cpu);
4294 schedule_debug(prev);
4296 if (sched_feat(HRTICK))
4299 raw_spin_lock_irq(&rq->lock);
4301 switch_count = &prev->nivcsw;
4302 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4303 if (unlikely(signal_pending_state(prev->state, prev))) {
4304 prev->state = TASK_RUNNING;
4306 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4310 * If a worker went to sleep, notify and ask workqueue
4311 * whether it wants to wake up a task to maintain
4314 if (prev->flags & PF_WQ_WORKER) {
4315 struct task_struct *to_wakeup;
4317 to_wakeup = wq_worker_sleeping(prev, cpu);
4319 try_to_wake_up_local(to_wakeup);
4323 * If we are going to sleep and we have plugged IO
4324 * queued, make sure to submit it to avoid deadlocks.
4326 if (blk_needs_flush_plug(prev)) {
4327 raw_spin_unlock(&rq->lock);
4328 blk_schedule_flush_plug(prev);
4329 raw_spin_lock(&rq->lock);
4332 switch_count = &prev->nvcsw;
4335 pre_schedule(rq, prev);
4337 if (unlikely(!rq->nr_running))
4338 idle_balance(cpu, rq);
4340 put_prev_task(rq, prev);
4341 next = pick_next_task(rq);
4342 clear_tsk_need_resched(prev);
4343 rq->skip_clock_update = 0;
4345 if (likely(prev != next)) {
4350 context_switch(rq, prev, next); /* unlocks the rq */
4352 * The context switch have flipped the stack from under us
4353 * and restored the local variables which were saved when
4354 * this task called schedule() in the past. prev == current
4355 * is still correct, but it can be moved to another cpu/rq.
4357 cpu = smp_processor_id();
4360 raw_spin_unlock_irq(&rq->lock);
4364 preempt_enable_no_resched();
4368 EXPORT_SYMBOL(schedule);
4370 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4372 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4374 if (lock->owner != owner)
4378 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4379 * lock->owner still matches owner, if that fails, owner might
4380 * point to free()d memory, if it still matches, the rcu_read_lock()
4381 * ensures the memory stays valid.
4385 return owner->on_cpu;
4389 * Look out! "owner" is an entirely speculative pointer
4390 * access and not reliable.
4392 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4394 if (!sched_feat(OWNER_SPIN))
4398 while (owner_running(lock, owner)) {
4402 arch_mutex_cpu_relax();
4407 * We break out the loop above on need_resched() and when the
4408 * owner changed, which is a sign for heavy contention. Return
4409 * success only when lock->owner is NULL.
4411 return lock->owner == NULL;
4415 #ifdef CONFIG_PREEMPT
4417 * this is the entry point to schedule() from in-kernel preemption
4418 * off of preempt_enable. Kernel preemptions off return from interrupt
4419 * occur there and call schedule directly.
4421 asmlinkage void __sched notrace preempt_schedule(void)
4423 struct thread_info *ti = current_thread_info();
4426 * If there is a non-zero preempt_count or interrupts are disabled,
4427 * we do not want to preempt the current task. Just return..
4429 if (likely(ti->preempt_count || irqs_disabled()))
4433 add_preempt_count_notrace(PREEMPT_ACTIVE);
4435 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4438 * Check again in case we missed a preemption opportunity
4439 * between schedule and now.
4442 } while (need_resched());
4444 EXPORT_SYMBOL(preempt_schedule);
4447 * this is the entry point to schedule() from kernel preemption
4448 * off of irq context.
4449 * Note, that this is called and return with irqs disabled. This will
4450 * protect us against recursive calling from irq.
4452 asmlinkage void __sched preempt_schedule_irq(void)
4454 struct thread_info *ti = current_thread_info();
4456 /* Catch callers which need to be fixed */
4457 BUG_ON(ti->preempt_count || !irqs_disabled());
4460 add_preempt_count(PREEMPT_ACTIVE);
4463 local_irq_disable();
4464 sub_preempt_count(PREEMPT_ACTIVE);
4467 * Check again in case we missed a preemption opportunity
4468 * between schedule and now.
4471 } while (need_resched());
4474 #endif /* CONFIG_PREEMPT */
4476 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4479 return try_to_wake_up(curr->private, mode, wake_flags);
4481 EXPORT_SYMBOL(default_wake_function);
4484 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4485 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4486 * number) then we wake all the non-exclusive tasks and one exclusive task.
4488 * There are circumstances in which we can try to wake a task which has already
4489 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4490 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4492 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4493 int nr_exclusive, int wake_flags, void *key)
4495 wait_queue_t *curr, *next;
4497 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4498 unsigned flags = curr->flags;
4500 if (curr->func(curr, mode, wake_flags, key) &&
4501 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4507 * __wake_up - wake up threads blocked on a waitqueue.
4509 * @mode: which threads
4510 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4511 * @key: is directly passed to the wakeup function
4513 * It may be assumed that this function implies a write memory barrier before
4514 * changing the task state if and only if any tasks are woken up.
4516 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4517 int nr_exclusive, void *key)
4519 unsigned long flags;
4521 spin_lock_irqsave(&q->lock, flags);
4522 __wake_up_common(q, mode, nr_exclusive, 0, key);
4523 spin_unlock_irqrestore(&q->lock, flags);
4525 EXPORT_SYMBOL(__wake_up);
4528 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4530 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4532 __wake_up_common(q, mode, 1, 0, NULL);
4534 EXPORT_SYMBOL_GPL(__wake_up_locked);
4536 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4538 __wake_up_common(q, mode, 1, 0, key);
4540 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4543 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4545 * @mode: which threads
4546 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4547 * @key: opaque value to be passed to wakeup targets
4549 * The sync wakeup differs that the waker knows that it will schedule
4550 * away soon, so while the target thread will be woken up, it will not
4551 * be migrated to another CPU - ie. the two threads are 'synchronized'
4552 * with each other. This can prevent needless bouncing between CPUs.
4554 * On UP it can prevent extra preemption.
4556 * It may be assumed that this function implies a write memory barrier before
4557 * changing the task state if and only if any tasks are woken up.
4559 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4560 int nr_exclusive, void *key)
4562 unsigned long flags;
4563 int wake_flags = WF_SYNC;
4568 if (unlikely(!nr_exclusive))
4571 spin_lock_irqsave(&q->lock, flags);
4572 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4573 spin_unlock_irqrestore(&q->lock, flags);
4575 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4578 * __wake_up_sync - see __wake_up_sync_key()
4580 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4582 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4584 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4587 * complete: - signals a single thread waiting on this completion
4588 * @x: holds the state of this particular completion
4590 * This will wake up a single thread waiting on this completion. Threads will be
4591 * awakened in the same order in which they were queued.
4593 * See also complete_all(), wait_for_completion() and related routines.
4595 * It may be assumed that this function implies a write memory barrier before
4596 * changing the task state if and only if any tasks are woken up.
4598 void complete(struct completion *x)
4600 unsigned long flags;
4602 spin_lock_irqsave(&x->wait.lock, flags);
4604 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4605 spin_unlock_irqrestore(&x->wait.lock, flags);
4607 EXPORT_SYMBOL(complete);
4610 * complete_all: - signals all threads waiting on this completion
4611 * @x: holds the state of this particular completion
4613 * This will wake up all threads waiting on this particular completion event.
4615 * It may be assumed that this function implies a write memory barrier before
4616 * changing the task state if and only if any tasks are woken up.
4618 void complete_all(struct completion *x)
4620 unsigned long flags;
4622 spin_lock_irqsave(&x->wait.lock, flags);
4623 x->done += UINT_MAX/2;
4624 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4625 spin_unlock_irqrestore(&x->wait.lock, flags);
4627 EXPORT_SYMBOL(complete_all);
4629 static inline long __sched
4630 do_wait_for_common(struct completion *x, long timeout, int state)
4633 DECLARE_WAITQUEUE(wait, current);
4635 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4637 if (signal_pending_state(state, current)) {
4638 timeout = -ERESTARTSYS;
4641 __set_current_state(state);
4642 spin_unlock_irq(&x->wait.lock);
4643 timeout = schedule_timeout(timeout);
4644 spin_lock_irq(&x->wait.lock);
4645 } while (!x->done && timeout);
4646 __remove_wait_queue(&x->wait, &wait);
4651 return timeout ?: 1;
4655 wait_for_common(struct completion *x, long timeout, int state)
4659 spin_lock_irq(&x->wait.lock);
4660 timeout = do_wait_for_common(x, timeout, state);
4661 spin_unlock_irq(&x->wait.lock);
4666 * wait_for_completion: - waits for completion of a task
4667 * @x: holds the state of this particular completion
4669 * This waits to be signaled for completion of a specific task. It is NOT
4670 * interruptible and there is no timeout.
4672 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4673 * and interrupt capability. Also see complete().
4675 void __sched wait_for_completion(struct completion *x)
4677 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4679 EXPORT_SYMBOL(wait_for_completion);
4682 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4683 * @x: holds the state of this particular completion
4684 * @timeout: timeout value in jiffies
4686 * This waits for either a completion of a specific task to be signaled or for a
4687 * specified timeout to expire. The timeout is in jiffies. It is not
4690 unsigned long __sched
4691 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4693 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4695 EXPORT_SYMBOL(wait_for_completion_timeout);
4698 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4699 * @x: holds the state of this particular completion
4701 * This waits for completion of a specific task to be signaled. It is
4704 int __sched wait_for_completion_interruptible(struct completion *x)
4706 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4707 if (t == -ERESTARTSYS)
4711 EXPORT_SYMBOL(wait_for_completion_interruptible);
4714 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4715 * @x: holds the state of this particular completion
4716 * @timeout: timeout value in jiffies
4718 * This waits for either a completion of a specific task to be signaled or for a
4719 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4722 wait_for_completion_interruptible_timeout(struct completion *x,
4723 unsigned long timeout)
4725 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4727 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4730 * wait_for_completion_killable: - waits for completion of a task (killable)
4731 * @x: holds the state of this particular completion
4733 * This waits to be signaled for completion of a specific task. It can be
4734 * interrupted by a kill signal.
4736 int __sched wait_for_completion_killable(struct completion *x)
4738 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4739 if (t == -ERESTARTSYS)
4743 EXPORT_SYMBOL(wait_for_completion_killable);
4746 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4747 * @x: holds the state of this particular completion
4748 * @timeout: timeout value in jiffies
4750 * This waits for either a completion of a specific task to be
4751 * signaled or for a specified timeout to expire. It can be
4752 * interrupted by a kill signal. The timeout is in jiffies.
4755 wait_for_completion_killable_timeout(struct completion *x,
4756 unsigned long timeout)
4758 return wait_for_common(x, timeout, TASK_KILLABLE);
4760 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4763 * try_wait_for_completion - try to decrement a completion without blocking
4764 * @x: completion structure
4766 * Returns: 0 if a decrement cannot be done without blocking
4767 * 1 if a decrement succeeded.
4769 * If a completion is being used as a counting completion,
4770 * attempt to decrement the counter without blocking. This
4771 * enables us to avoid waiting if the resource the completion
4772 * is protecting is not available.
4774 bool try_wait_for_completion(struct completion *x)
4776 unsigned long flags;
4779 spin_lock_irqsave(&x->wait.lock, flags);
4784 spin_unlock_irqrestore(&x->wait.lock, flags);
4787 EXPORT_SYMBOL(try_wait_for_completion);
4790 * completion_done - Test to see if a completion has any waiters
4791 * @x: completion structure
4793 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4794 * 1 if there are no waiters.
4797 bool completion_done(struct completion *x)
4799 unsigned long flags;
4802 spin_lock_irqsave(&x->wait.lock, flags);
4805 spin_unlock_irqrestore(&x->wait.lock, flags);
4808 EXPORT_SYMBOL(completion_done);
4811 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4813 unsigned long flags;
4816 init_waitqueue_entry(&wait, current);
4818 __set_current_state(state);
4820 spin_lock_irqsave(&q->lock, flags);
4821 __add_wait_queue(q, &wait);
4822 spin_unlock(&q->lock);
4823 timeout = schedule_timeout(timeout);
4824 spin_lock_irq(&q->lock);
4825 __remove_wait_queue(q, &wait);
4826 spin_unlock_irqrestore(&q->lock, flags);
4831 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4833 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4835 EXPORT_SYMBOL(interruptible_sleep_on);
4838 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4840 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4842 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4844 void __sched sleep_on(wait_queue_head_t *q)
4846 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4848 EXPORT_SYMBOL(sleep_on);
4850 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4852 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4854 EXPORT_SYMBOL(sleep_on_timeout);
4856 #ifdef CONFIG_RT_MUTEXES
4859 * rt_mutex_setprio - set the current priority of a task
4861 * @prio: prio value (kernel-internal form)
4863 * This function changes the 'effective' priority of a task. It does
4864 * not touch ->normal_prio like __setscheduler().
4866 * Used by the rt_mutex code to implement priority inheritance logic.
4868 void rt_mutex_setprio(struct task_struct *p, int prio)
4870 int oldprio, on_rq, running;
4872 const struct sched_class *prev_class;
4874 BUG_ON(prio < 0 || prio > MAX_PRIO);
4876 rq = __task_rq_lock(p);
4878 trace_sched_pi_setprio(p, prio);
4880 prev_class = p->sched_class;
4882 running = task_current(rq, p);
4884 dequeue_task(rq, p, 0);
4886 p->sched_class->put_prev_task(rq, p);
4889 p->sched_class = &rt_sched_class;
4891 p->sched_class = &fair_sched_class;
4896 p->sched_class->set_curr_task(rq);
4898 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4900 check_class_changed(rq, p, prev_class, oldprio);
4901 __task_rq_unlock(rq);
4906 void set_user_nice(struct task_struct *p, long nice)
4908 int old_prio, delta, on_rq;
4909 unsigned long flags;
4912 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4915 * We have to be careful, if called from sys_setpriority(),
4916 * the task might be in the middle of scheduling on another CPU.
4918 rq = task_rq_lock(p, &flags);
4920 * The RT priorities are set via sched_setscheduler(), but we still
4921 * allow the 'normal' nice value to be set - but as expected
4922 * it wont have any effect on scheduling until the task is
4923 * SCHED_FIFO/SCHED_RR:
4925 if (task_has_rt_policy(p)) {
4926 p->static_prio = NICE_TO_PRIO(nice);
4931 dequeue_task(rq, p, 0);
4933 p->static_prio = NICE_TO_PRIO(nice);
4936 p->prio = effective_prio(p);
4937 delta = p->prio - old_prio;
4940 enqueue_task(rq, p, 0);
4942 * If the task increased its priority or is running and
4943 * lowered its priority, then reschedule its CPU:
4945 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4946 resched_task(rq->curr);
4949 task_rq_unlock(rq, p, &flags);
4951 EXPORT_SYMBOL(set_user_nice);
4954 * can_nice - check if a task can reduce its nice value
4958 int can_nice(const struct task_struct *p, const int nice)
4960 /* convert nice value [19,-20] to rlimit style value [1,40] */
4961 int nice_rlim = 20 - nice;
4963 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4964 capable(CAP_SYS_NICE));
4967 #ifdef __ARCH_WANT_SYS_NICE
4970 * sys_nice - change the priority of the current process.
4971 * @increment: priority increment
4973 * sys_setpriority is a more generic, but much slower function that
4974 * does similar things.
4976 SYSCALL_DEFINE1(nice, int, increment)
4981 * Setpriority might change our priority at the same moment.
4982 * We don't have to worry. Conceptually one call occurs first
4983 * and we have a single winner.
4985 if (increment < -40)
4990 nice = TASK_NICE(current) + increment;
4996 if (increment < 0 && !can_nice(current, nice))
4999 retval = security_task_setnice(current, nice);
5003 set_user_nice(current, nice);
5010 * task_prio - return the priority value of a given task.
5011 * @p: the task in question.
5013 * This is the priority value as seen by users in /proc.
5014 * RT tasks are offset by -200. Normal tasks are centered
5015 * around 0, value goes from -16 to +15.
5017 int task_prio(const struct task_struct *p)
5019 return p->prio - MAX_RT_PRIO;
5023 * task_nice - return the nice value of a given task.
5024 * @p: the task in question.
5026 int task_nice(const struct task_struct *p)
5028 return TASK_NICE(p);
5030 EXPORT_SYMBOL(task_nice);
5033 * idle_cpu - is a given cpu idle currently?
5034 * @cpu: the processor in question.
5036 int idle_cpu(int cpu)
5038 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5042 * idle_task - return the idle task for a given cpu.
5043 * @cpu: the processor in question.
5045 struct task_struct *idle_task(int cpu)
5047 return cpu_rq(cpu)->idle;
5051 * find_process_by_pid - find a process with a matching PID value.
5052 * @pid: the pid in question.
5054 static struct task_struct *find_process_by_pid(pid_t pid)
5056 return pid ? find_task_by_vpid(pid) : current;
5059 /* Actually do priority change: must hold rq lock. */
5061 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5064 p->rt_priority = prio;
5065 p->normal_prio = normal_prio(p);
5066 /* we are holding p->pi_lock already */
5067 p->prio = rt_mutex_getprio(p);
5068 if (rt_prio(p->prio))
5069 p->sched_class = &rt_sched_class;
5071 p->sched_class = &fair_sched_class;
5076 * check the target process has a UID that matches the current process's
5078 static bool check_same_owner(struct task_struct *p)
5080 const struct cred *cred = current_cred(), *pcred;
5084 pcred = __task_cred(p);
5085 if (cred->user->user_ns == pcred->user->user_ns)
5086 match = (cred->euid == pcred->euid ||
5087 cred->euid == pcred->uid);
5094 static int __sched_setscheduler(struct task_struct *p, int policy,
5095 const struct sched_param *param, bool user)
5097 int retval, oldprio, oldpolicy = -1, on_rq, running;
5098 unsigned long flags;
5099 const struct sched_class *prev_class;
5103 /* may grab non-irq protected spin_locks */
5104 BUG_ON(in_interrupt());
5106 /* double check policy once rq lock held */
5108 reset_on_fork = p->sched_reset_on_fork;
5109 policy = oldpolicy = p->policy;
5111 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5112 policy &= ~SCHED_RESET_ON_FORK;
5114 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5115 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5116 policy != SCHED_IDLE)
5121 * Valid priorities for SCHED_FIFO and SCHED_RR are
5122 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5123 * SCHED_BATCH and SCHED_IDLE is 0.
5125 if (param->sched_priority < 0 ||
5126 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5127 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5129 if (rt_policy(policy) != (param->sched_priority != 0))
5133 * Allow unprivileged RT tasks to decrease priority:
5135 if (user && !capable(CAP_SYS_NICE)) {
5136 if (rt_policy(policy)) {
5137 unsigned long rlim_rtprio =
5138 task_rlimit(p, RLIMIT_RTPRIO);
5140 /* can't set/change the rt policy */
5141 if (policy != p->policy && !rlim_rtprio)
5144 /* can't increase priority */
5145 if (param->sched_priority > p->rt_priority &&
5146 param->sched_priority > rlim_rtprio)
5151 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5152 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5154 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5155 if (!can_nice(p, TASK_NICE(p)))
5159 /* can't change other user's priorities */
5160 if (!check_same_owner(p))
5163 /* Normal users shall not reset the sched_reset_on_fork flag */
5164 if (p->sched_reset_on_fork && !reset_on_fork)
5169 retval = security_task_setscheduler(p);
5175 * make sure no PI-waiters arrive (or leave) while we are
5176 * changing the priority of the task:
5178 * To be able to change p->policy safely, the appropriate
5179 * runqueue lock must be held.
5181 rq = task_rq_lock(p, &flags);
5184 * Changing the policy of the stop threads its a very bad idea
5186 if (p == rq->stop) {
5187 task_rq_unlock(rq, p, &flags);
5192 * If not changing anything there's no need to proceed further:
5194 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5195 param->sched_priority == p->rt_priority))) {
5197 __task_rq_unlock(rq);
5198 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5202 #ifdef CONFIG_RT_GROUP_SCHED
5205 * Do not allow realtime tasks into groups that have no runtime
5208 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5209 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5210 !task_group_is_autogroup(task_group(p))) {
5211 task_rq_unlock(rq, p, &flags);
5217 /* recheck policy now with rq lock held */
5218 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5219 policy = oldpolicy = -1;
5220 task_rq_unlock(rq, p, &flags);
5224 running = task_current(rq, p);
5226 deactivate_task(rq, p, 0);
5228 p->sched_class->put_prev_task(rq, p);
5230 p->sched_reset_on_fork = reset_on_fork;
5233 prev_class = p->sched_class;
5234 __setscheduler(rq, p, policy, param->sched_priority);
5237 p->sched_class->set_curr_task(rq);
5239 activate_task(rq, p, 0);
5241 check_class_changed(rq, p, prev_class, oldprio);
5242 task_rq_unlock(rq, p, &flags);
5244 rt_mutex_adjust_pi(p);
5250 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5251 * @p: the task in question.
5252 * @policy: new policy.
5253 * @param: structure containing the new RT priority.
5255 * NOTE that the task may be already dead.
5257 int sched_setscheduler(struct task_struct *p, int policy,
5258 const struct sched_param *param)
5260 return __sched_setscheduler(p, policy, param, true);
5262 EXPORT_SYMBOL_GPL(sched_setscheduler);
5265 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5266 * @p: the task in question.
5267 * @policy: new policy.
5268 * @param: structure containing the new RT priority.
5270 * Just like sched_setscheduler, only don't bother checking if the
5271 * current context has permission. For example, this is needed in
5272 * stop_machine(): we create temporary high priority worker threads,
5273 * but our caller might not have that capability.
5275 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5276 const struct sched_param *param)
5278 return __sched_setscheduler(p, policy, param, false);
5282 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5284 struct sched_param lparam;
5285 struct task_struct *p;
5288 if (!param || pid < 0)
5290 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5295 p = find_process_by_pid(pid);
5297 retval = sched_setscheduler(p, policy, &lparam);
5304 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5305 * @pid: the pid in question.
5306 * @policy: new policy.
5307 * @param: structure containing the new RT priority.
5309 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5310 struct sched_param __user *, param)
5312 /* negative values for policy are not valid */
5316 return do_sched_setscheduler(pid, policy, param);
5320 * sys_sched_setparam - set/change the RT priority of a thread
5321 * @pid: the pid in question.
5322 * @param: structure containing the new RT priority.
5324 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5326 return do_sched_setscheduler(pid, -1, param);
5330 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5331 * @pid: the pid in question.
5333 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5335 struct task_struct *p;
5343 p = find_process_by_pid(pid);
5345 retval = security_task_getscheduler(p);
5348 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5355 * sys_sched_getparam - get the RT priority of a thread
5356 * @pid: the pid in question.
5357 * @param: structure containing the RT priority.
5359 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5361 struct sched_param lp;
5362 struct task_struct *p;
5365 if (!param || pid < 0)
5369 p = find_process_by_pid(pid);
5374 retval = security_task_getscheduler(p);
5378 lp.sched_priority = p->rt_priority;
5382 * This one might sleep, we cannot do it with a spinlock held ...
5384 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5393 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5395 cpumask_var_t cpus_allowed, new_mask;
5396 struct task_struct *p;
5402 p = find_process_by_pid(pid);
5409 /* Prevent p going away */
5413 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5417 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5419 goto out_free_cpus_allowed;
5422 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5425 retval = security_task_setscheduler(p);
5429 cpuset_cpus_allowed(p, cpus_allowed);
5430 cpumask_and(new_mask, in_mask, cpus_allowed);
5432 retval = set_cpus_allowed_ptr(p, new_mask);
5435 cpuset_cpus_allowed(p, cpus_allowed);
5436 if (!cpumask_subset(new_mask, cpus_allowed)) {
5438 * We must have raced with a concurrent cpuset
5439 * update. Just reset the cpus_allowed to the
5440 * cpuset's cpus_allowed
5442 cpumask_copy(new_mask, cpus_allowed);
5447 free_cpumask_var(new_mask);
5448 out_free_cpus_allowed:
5449 free_cpumask_var(cpus_allowed);
5456 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5457 struct cpumask *new_mask)
5459 if (len < cpumask_size())
5460 cpumask_clear(new_mask);
5461 else if (len > cpumask_size())
5462 len = cpumask_size();
5464 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5468 * sys_sched_setaffinity - set the cpu affinity of a process
5469 * @pid: pid of the process
5470 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5471 * @user_mask_ptr: user-space pointer to the new cpu mask
5473 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5474 unsigned long __user *, user_mask_ptr)
5476 cpumask_var_t new_mask;
5479 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5482 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5484 retval = sched_setaffinity(pid, new_mask);
5485 free_cpumask_var(new_mask);
5489 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5491 struct task_struct *p;
5492 unsigned long flags;
5499 p = find_process_by_pid(pid);
5503 retval = security_task_getscheduler(p);
5507 raw_spin_lock_irqsave(&p->pi_lock, flags);
5508 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5509 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5519 * sys_sched_getaffinity - get the cpu affinity of a process
5520 * @pid: pid of the process
5521 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5522 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5524 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5525 unsigned long __user *, user_mask_ptr)
5530 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5532 if (len & (sizeof(unsigned long)-1))
5535 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5538 ret = sched_getaffinity(pid, mask);
5540 size_t retlen = min_t(size_t, len, cpumask_size());
5542 if (copy_to_user(user_mask_ptr, mask, retlen))
5547 free_cpumask_var(mask);
5553 * sys_sched_yield - yield the current processor to other threads.
5555 * This function yields the current CPU to other tasks. If there are no
5556 * other threads running on this CPU then this function will return.
5558 SYSCALL_DEFINE0(sched_yield)
5560 struct rq *rq = this_rq_lock();
5562 schedstat_inc(rq, yld_count);
5563 current->sched_class->yield_task(rq);
5566 * Since we are going to call schedule() anyway, there's
5567 * no need to preempt or enable interrupts:
5569 __release(rq->lock);
5570 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5571 do_raw_spin_unlock(&rq->lock);
5572 preempt_enable_no_resched();
5579 static inline int should_resched(void)
5581 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5584 static void __cond_resched(void)
5586 add_preempt_count(PREEMPT_ACTIVE);
5588 sub_preempt_count(PREEMPT_ACTIVE);
5591 int __sched _cond_resched(void)
5593 if (should_resched()) {
5599 EXPORT_SYMBOL(_cond_resched);
5602 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5603 * call schedule, and on return reacquire the lock.
5605 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5606 * operations here to prevent schedule() from being called twice (once via
5607 * spin_unlock(), once by hand).
5609 int __cond_resched_lock(spinlock_t *lock)
5611 int resched = should_resched();
5614 lockdep_assert_held(lock);
5616 if (spin_needbreak(lock) || resched) {
5627 EXPORT_SYMBOL(__cond_resched_lock);
5629 int __sched __cond_resched_softirq(void)
5631 BUG_ON(!in_softirq());
5633 if (should_resched()) {
5641 EXPORT_SYMBOL(__cond_resched_softirq);
5644 * yield - yield the current processor to other threads.
5646 * This is a shortcut for kernel-space yielding - it marks the
5647 * thread runnable and calls sys_sched_yield().
5649 void __sched yield(void)
5651 set_current_state(TASK_RUNNING);
5654 EXPORT_SYMBOL(yield);
5657 * yield_to - yield the current processor to another thread in
5658 * your thread group, or accelerate that thread toward the
5659 * processor it's on.
5661 * @preempt: whether task preemption is allowed or not
5663 * It's the caller's job to ensure that the target task struct
5664 * can't go away on us before we can do any checks.
5666 * Returns true if we indeed boosted the target task.
5668 bool __sched yield_to(struct task_struct *p, bool preempt)
5670 struct task_struct *curr = current;
5671 struct rq *rq, *p_rq;
5672 unsigned long flags;
5675 local_irq_save(flags);
5680 double_rq_lock(rq, p_rq);
5681 while (task_rq(p) != p_rq) {
5682 double_rq_unlock(rq, p_rq);
5686 if (!curr->sched_class->yield_to_task)
5689 if (curr->sched_class != p->sched_class)
5692 if (task_running(p_rq, p) || p->state)
5695 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5697 schedstat_inc(rq, yld_count);
5699 * Make p's CPU reschedule; pick_next_entity takes care of
5702 if (preempt && rq != p_rq)
5703 resched_task(p_rq->curr);
5707 double_rq_unlock(rq, p_rq);
5708 local_irq_restore(flags);
5715 EXPORT_SYMBOL_GPL(yield_to);
5718 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5719 * that process accounting knows that this is a task in IO wait state.
5721 void __sched io_schedule(void)
5723 struct rq *rq = raw_rq();
5725 delayacct_blkio_start();
5726 atomic_inc(&rq->nr_iowait);
5727 blk_flush_plug(current);
5728 current->in_iowait = 1;
5730 current->in_iowait = 0;
5731 atomic_dec(&rq->nr_iowait);
5732 delayacct_blkio_end();
5734 EXPORT_SYMBOL(io_schedule);
5736 long __sched io_schedule_timeout(long timeout)
5738 struct rq *rq = raw_rq();
5741 delayacct_blkio_start();
5742 atomic_inc(&rq->nr_iowait);
5743 blk_flush_plug(current);
5744 current->in_iowait = 1;
5745 ret = schedule_timeout(timeout);
5746 current->in_iowait = 0;
5747 atomic_dec(&rq->nr_iowait);
5748 delayacct_blkio_end();
5753 * sys_sched_get_priority_max - return maximum RT priority.
5754 * @policy: scheduling class.
5756 * this syscall returns the maximum rt_priority that can be used
5757 * by a given scheduling class.
5759 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5766 ret = MAX_USER_RT_PRIO-1;
5778 * sys_sched_get_priority_min - return minimum RT priority.
5779 * @policy: scheduling class.
5781 * this syscall returns the minimum rt_priority that can be used
5782 * by a given scheduling class.
5784 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5802 * sys_sched_rr_get_interval - return the default timeslice of a process.
5803 * @pid: pid of the process.
5804 * @interval: userspace pointer to the timeslice value.
5806 * this syscall writes the default timeslice value of a given process
5807 * into the user-space timespec buffer. A value of '0' means infinity.
5809 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5810 struct timespec __user *, interval)
5812 struct task_struct *p;
5813 unsigned int time_slice;
5814 unsigned long flags;
5824 p = find_process_by_pid(pid);
5828 retval = security_task_getscheduler(p);
5832 rq = task_rq_lock(p, &flags);
5833 time_slice = p->sched_class->get_rr_interval(rq, p);
5834 task_rq_unlock(rq, p, &flags);
5837 jiffies_to_timespec(time_slice, &t);
5838 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5846 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5848 void sched_show_task(struct task_struct *p)
5850 unsigned long free = 0;
5853 state = p->state ? __ffs(p->state) + 1 : 0;
5854 printk(KERN_INFO "%-15.15s %c", p->comm,
5855 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5856 #if BITS_PER_LONG == 32
5857 if (state == TASK_RUNNING)
5858 printk(KERN_CONT " running ");
5860 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5862 if (state == TASK_RUNNING)
5863 printk(KERN_CONT " running task ");
5865 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5867 #ifdef CONFIG_DEBUG_STACK_USAGE
5868 free = stack_not_used(p);
5870 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5871 task_pid_nr(p), task_pid_nr(p->real_parent),
5872 (unsigned long)task_thread_info(p)->flags);
5874 show_stack(p, NULL);
5877 void show_state_filter(unsigned long state_filter)
5879 struct task_struct *g, *p;
5881 #if BITS_PER_LONG == 32
5883 " task PC stack pid father\n");
5886 " task PC stack pid father\n");
5888 read_lock(&tasklist_lock);
5889 do_each_thread(g, p) {
5891 * reset the NMI-timeout, listing all files on a slow
5892 * console might take a lot of time:
5894 touch_nmi_watchdog();
5895 if (!state_filter || (p->state & state_filter))
5897 } while_each_thread(g, p);
5899 touch_all_softlockup_watchdogs();
5901 #ifdef CONFIG_SCHED_DEBUG
5902 sysrq_sched_debug_show();
5904 read_unlock(&tasklist_lock);
5906 * Only show locks if all tasks are dumped:
5909 debug_show_all_locks();
5912 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5914 idle->sched_class = &idle_sched_class;
5918 * init_idle - set up an idle thread for a given CPU
5919 * @idle: task in question
5920 * @cpu: cpu the idle task belongs to
5922 * NOTE: this function does not set the idle thread's NEED_RESCHED
5923 * flag, to make booting more robust.
5925 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5927 struct rq *rq = cpu_rq(cpu);
5928 unsigned long flags;
5930 raw_spin_lock_irqsave(&rq->lock, flags);
5933 idle->state = TASK_RUNNING;
5934 idle->se.exec_start = sched_clock();
5936 do_set_cpus_allowed(idle, cpumask_of(cpu));
5938 * We're having a chicken and egg problem, even though we are
5939 * holding rq->lock, the cpu isn't yet set to this cpu so the
5940 * lockdep check in task_group() will fail.
5942 * Similar case to sched_fork(). / Alternatively we could
5943 * use task_rq_lock() here and obtain the other rq->lock.
5948 __set_task_cpu(idle, cpu);
5951 rq->curr = rq->idle = idle;
5952 #if defined(CONFIG_SMP)
5955 raw_spin_unlock_irqrestore(&rq->lock, flags);
5957 /* Set the preempt count _outside_ the spinlocks! */
5958 task_thread_info(idle)->preempt_count = 0;
5961 * The idle tasks have their own, simple scheduling class:
5963 idle->sched_class = &idle_sched_class;
5964 ftrace_graph_init_idle_task(idle, cpu);
5968 * In a system that switches off the HZ timer nohz_cpu_mask
5969 * indicates which cpus entered this state. This is used
5970 * in the rcu update to wait only for active cpus. For system
5971 * which do not switch off the HZ timer nohz_cpu_mask should
5972 * always be CPU_BITS_NONE.
5974 cpumask_var_t nohz_cpu_mask;
5977 * Increase the granularity value when there are more CPUs,
5978 * because with more CPUs the 'effective latency' as visible
5979 * to users decreases. But the relationship is not linear,
5980 * so pick a second-best guess by going with the log2 of the
5983 * This idea comes from the SD scheduler of Con Kolivas:
5985 static int get_update_sysctl_factor(void)
5987 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5988 unsigned int factor;
5990 switch (sysctl_sched_tunable_scaling) {
5991 case SCHED_TUNABLESCALING_NONE:
5994 case SCHED_TUNABLESCALING_LINEAR:
5997 case SCHED_TUNABLESCALING_LOG:
5999 factor = 1 + ilog2(cpus);
6006 static void update_sysctl(void)
6008 unsigned int factor = get_update_sysctl_factor();
6010 #define SET_SYSCTL(name) \
6011 (sysctl_##name = (factor) * normalized_sysctl_##name)
6012 SET_SYSCTL(sched_min_granularity);
6013 SET_SYSCTL(sched_latency);
6014 SET_SYSCTL(sched_wakeup_granularity);
6018 static inline void sched_init_granularity(void)
6024 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
6026 if (p->sched_class && p->sched_class->set_cpus_allowed)
6027 p->sched_class->set_cpus_allowed(p, new_mask);
6029 cpumask_copy(&p->cpus_allowed, new_mask);
6030 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6035 * This is how migration works:
6037 * 1) we invoke migration_cpu_stop() on the target CPU using
6039 * 2) stopper starts to run (implicitly forcing the migrated thread
6041 * 3) it checks whether the migrated task is still in the wrong runqueue.
6042 * 4) if it's in the wrong runqueue then the migration thread removes
6043 * it and puts it into the right queue.
6044 * 5) stopper completes and stop_one_cpu() returns and the migration
6049 * Change a given task's CPU affinity. Migrate the thread to a
6050 * proper CPU and schedule it away if the CPU it's executing on
6051 * is removed from the allowed bitmask.
6053 * NOTE: the caller must have a valid reference to the task, the
6054 * task must not exit() & deallocate itself prematurely. The
6055 * call is not atomic; no spinlocks may be held.
6057 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6059 unsigned long flags;
6061 unsigned int dest_cpu;
6064 rq = task_rq_lock(p, &flags);
6066 if (cpumask_equal(&p->cpus_allowed, new_mask))
6069 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6074 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6079 do_set_cpus_allowed(p, new_mask);
6081 /* Can the task run on the task's current CPU? If so, we're done */
6082 if (cpumask_test_cpu(task_cpu(p), new_mask))
6085 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6087 struct migration_arg arg = { p, dest_cpu };
6088 /* Need help from migration thread: drop lock and wait. */
6089 task_rq_unlock(rq, p, &flags);
6090 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6091 tlb_migrate_finish(p->mm);
6095 task_rq_unlock(rq, p, &flags);
6099 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6102 * Move (not current) task off this cpu, onto dest cpu. We're doing
6103 * this because either it can't run here any more (set_cpus_allowed()
6104 * away from this CPU, or CPU going down), or because we're
6105 * attempting to rebalance this task on exec (sched_exec).
6107 * So we race with normal scheduler movements, but that's OK, as long
6108 * as the task is no longer on this CPU.
6110 * Returns non-zero if task was successfully migrated.
6112 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6114 struct rq *rq_dest, *rq_src;
6117 if (unlikely(!cpu_active(dest_cpu)))
6120 rq_src = cpu_rq(src_cpu);
6121 rq_dest = cpu_rq(dest_cpu);
6123 raw_spin_lock(&p->pi_lock);
6124 double_rq_lock(rq_src, rq_dest);
6125 /* Already moved. */
6126 if (task_cpu(p) != src_cpu)
6128 /* Affinity changed (again). */
6129 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6133 * If we're not on a rq, the next wake-up will ensure we're
6137 deactivate_task(rq_src, p, 0);
6138 set_task_cpu(p, dest_cpu);
6139 activate_task(rq_dest, p, 0);
6140 check_preempt_curr(rq_dest, p, 0);
6145 double_rq_unlock(rq_src, rq_dest);
6146 raw_spin_unlock(&p->pi_lock);
6151 * migration_cpu_stop - this will be executed by a highprio stopper thread
6152 * and performs thread migration by bumping thread off CPU then
6153 * 'pushing' onto another runqueue.
6155 static int migration_cpu_stop(void *data)
6157 struct migration_arg *arg = data;
6160 * The original target cpu might have gone down and we might
6161 * be on another cpu but it doesn't matter.
6163 local_irq_disable();
6164 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6169 #ifdef CONFIG_HOTPLUG_CPU
6172 * Ensures that the idle task is using init_mm right before its cpu goes
6175 void idle_task_exit(void)
6177 struct mm_struct *mm = current->active_mm;
6179 BUG_ON(cpu_online(smp_processor_id()));
6182 switch_mm(mm, &init_mm, current);
6187 * While a dead CPU has no uninterruptible tasks queued at this point,
6188 * it might still have a nonzero ->nr_uninterruptible counter, because
6189 * for performance reasons the counter is not stricly tracking tasks to
6190 * their home CPUs. So we just add the counter to another CPU's counter,
6191 * to keep the global sum constant after CPU-down:
6193 static void migrate_nr_uninterruptible(struct rq *rq_src)
6195 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6197 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6198 rq_src->nr_uninterruptible = 0;
6202 * remove the tasks which were accounted by rq from calc_load_tasks.
6204 static void calc_global_load_remove(struct rq *rq)
6206 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6207 rq->calc_load_active = 0;
6211 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6212 * try_to_wake_up()->select_task_rq().
6214 * Called with rq->lock held even though we'er in stop_machine() and
6215 * there's no concurrency possible, we hold the required locks anyway
6216 * because of lock validation efforts.
6218 static void migrate_tasks(unsigned int dead_cpu)
6220 struct rq *rq = cpu_rq(dead_cpu);
6221 struct task_struct *next, *stop = rq->stop;
6225 * Fudge the rq selection such that the below task selection loop
6226 * doesn't get stuck on the currently eligible stop task.
6228 * We're currently inside stop_machine() and the rq is either stuck
6229 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6230 * either way we should never end up calling schedule() until we're
6237 * There's this thread running, bail when that's the only
6240 if (rq->nr_running == 1)
6243 next = pick_next_task(rq);
6245 next->sched_class->put_prev_task(rq, next);
6247 /* Find suitable destination for @next, with force if needed. */
6248 dest_cpu = select_fallback_rq(dead_cpu, next);
6249 raw_spin_unlock(&rq->lock);
6251 __migrate_task(next, dead_cpu, dest_cpu);
6253 raw_spin_lock(&rq->lock);
6259 #endif /* CONFIG_HOTPLUG_CPU */
6261 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6263 static struct ctl_table sd_ctl_dir[] = {
6265 .procname = "sched_domain",
6271 static struct ctl_table sd_ctl_root[] = {
6273 .procname = "kernel",
6275 .child = sd_ctl_dir,
6280 static struct ctl_table *sd_alloc_ctl_entry(int n)
6282 struct ctl_table *entry =
6283 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6288 static void sd_free_ctl_entry(struct ctl_table **tablep)
6290 struct ctl_table *entry;
6293 * In the intermediate directories, both the child directory and
6294 * procname are dynamically allocated and could fail but the mode
6295 * will always be set. In the lowest directory the names are
6296 * static strings and all have proc handlers.
6298 for (entry = *tablep; entry->mode; entry++) {
6300 sd_free_ctl_entry(&entry->child);
6301 if (entry->proc_handler == NULL)
6302 kfree(entry->procname);
6310 set_table_entry(struct ctl_table *entry,
6311 const char *procname, void *data, int maxlen,
6312 mode_t mode, proc_handler *proc_handler)
6314 entry->procname = procname;
6316 entry->maxlen = maxlen;
6318 entry->proc_handler = proc_handler;
6321 static struct ctl_table *
6322 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6324 struct ctl_table *table = sd_alloc_ctl_entry(13);
6329 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6330 sizeof(long), 0644, proc_doulongvec_minmax);
6331 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6332 sizeof(long), 0644, proc_doulongvec_minmax);
6333 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6334 sizeof(int), 0644, proc_dointvec_minmax);
6335 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6336 sizeof(int), 0644, proc_dointvec_minmax);
6337 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6338 sizeof(int), 0644, proc_dointvec_minmax);
6339 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6340 sizeof(int), 0644, proc_dointvec_minmax);
6341 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6342 sizeof(int), 0644, proc_dointvec_minmax);
6343 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6344 sizeof(int), 0644, proc_dointvec_minmax);
6345 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6346 sizeof(int), 0644, proc_dointvec_minmax);
6347 set_table_entry(&table[9], "cache_nice_tries",
6348 &sd->cache_nice_tries,
6349 sizeof(int), 0644, proc_dointvec_minmax);
6350 set_table_entry(&table[10], "flags", &sd->flags,
6351 sizeof(int), 0644, proc_dointvec_minmax);
6352 set_table_entry(&table[11], "name", sd->name,
6353 CORENAME_MAX_SIZE, 0444, proc_dostring);
6354 /* &table[12] is terminator */
6359 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6361 struct ctl_table *entry, *table;
6362 struct sched_domain *sd;
6363 int domain_num = 0, i;
6366 for_each_domain(cpu, sd)
6368 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6373 for_each_domain(cpu, sd) {
6374 snprintf(buf, 32, "domain%d", i);
6375 entry->procname = kstrdup(buf, GFP_KERNEL);
6377 entry->child = sd_alloc_ctl_domain_table(sd);
6384 static struct ctl_table_header *sd_sysctl_header;
6385 static void register_sched_domain_sysctl(void)
6387 int i, cpu_num = num_possible_cpus();
6388 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6391 WARN_ON(sd_ctl_dir[0].child);
6392 sd_ctl_dir[0].child = entry;
6397 for_each_possible_cpu(i) {
6398 snprintf(buf, 32, "cpu%d", i);
6399 entry->procname = kstrdup(buf, GFP_KERNEL);
6401 entry->child = sd_alloc_ctl_cpu_table(i);
6405 WARN_ON(sd_sysctl_header);
6406 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6409 /* may be called multiple times per register */
6410 static void unregister_sched_domain_sysctl(void)
6412 if (sd_sysctl_header)
6413 unregister_sysctl_table(sd_sysctl_header);
6414 sd_sysctl_header = NULL;
6415 if (sd_ctl_dir[0].child)
6416 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6419 static void register_sched_domain_sysctl(void)
6422 static void unregister_sched_domain_sysctl(void)
6427 static void set_rq_online(struct rq *rq)
6430 const struct sched_class *class;
6432 cpumask_set_cpu(rq->cpu, rq->rd->online);
6435 for_each_class(class) {
6436 if (class->rq_online)
6437 class->rq_online(rq);
6442 static void set_rq_offline(struct rq *rq)
6445 const struct sched_class *class;
6447 for_each_class(class) {
6448 if (class->rq_offline)
6449 class->rq_offline(rq);
6452 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6458 * migration_call - callback that gets triggered when a CPU is added.
6459 * Here we can start up the necessary migration thread for the new CPU.
6461 static int __cpuinit
6462 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6464 int cpu = (long)hcpu;
6465 unsigned long flags;
6466 struct rq *rq = cpu_rq(cpu);
6468 switch (action & ~CPU_TASKS_FROZEN) {
6470 case CPU_UP_PREPARE:
6471 rq->calc_load_update = calc_load_update;
6475 /* Update our root-domain */
6476 raw_spin_lock_irqsave(&rq->lock, flags);
6478 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6482 raw_spin_unlock_irqrestore(&rq->lock, flags);
6485 #ifdef CONFIG_HOTPLUG_CPU
6487 sched_ttwu_pending();
6488 /* Update our root-domain */
6489 raw_spin_lock_irqsave(&rq->lock, flags);
6491 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6495 BUG_ON(rq->nr_running != 1); /* the migration thread */
6496 raw_spin_unlock_irqrestore(&rq->lock, flags);
6498 migrate_nr_uninterruptible(rq);
6499 calc_global_load_remove(rq);
6504 update_max_interval();
6510 * Register at high priority so that task migration (migrate_all_tasks)
6511 * happens before everything else. This has to be lower priority than
6512 * the notifier in the perf_event subsystem, though.
6514 static struct notifier_block __cpuinitdata migration_notifier = {
6515 .notifier_call = migration_call,
6516 .priority = CPU_PRI_MIGRATION,
6519 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6520 unsigned long action, void *hcpu)
6522 switch (action & ~CPU_TASKS_FROZEN) {
6524 case CPU_DOWN_FAILED:
6525 set_cpu_active((long)hcpu, true);
6532 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6533 unsigned long action, void *hcpu)
6535 switch (action & ~CPU_TASKS_FROZEN) {
6536 case CPU_DOWN_PREPARE:
6537 set_cpu_active((long)hcpu, false);
6544 static int __init migration_init(void)
6546 void *cpu = (void *)(long)smp_processor_id();
6549 /* Initialize migration for the boot CPU */
6550 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6551 BUG_ON(err == NOTIFY_BAD);
6552 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6553 register_cpu_notifier(&migration_notifier);
6555 /* Register cpu active notifiers */
6556 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6557 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6561 early_initcall(migration_init);
6566 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6568 #ifdef CONFIG_SCHED_DEBUG
6570 static __read_mostly int sched_domain_debug_enabled;
6572 static int __init sched_domain_debug_setup(char *str)
6574 sched_domain_debug_enabled = 1;
6578 early_param("sched_debug", sched_domain_debug_setup);
6580 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6581 struct cpumask *groupmask)
6583 struct sched_group *group = sd->groups;
6586 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6587 cpumask_clear(groupmask);
6589 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6591 if (!(sd->flags & SD_LOAD_BALANCE)) {
6592 printk("does not load-balance\n");
6594 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6599 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6601 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6602 printk(KERN_ERR "ERROR: domain->span does not contain "
6605 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6606 printk(KERN_ERR "ERROR: domain->groups does not contain"
6610 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6614 printk(KERN_ERR "ERROR: group is NULL\n");
6618 if (!group->sgp->power) {
6619 printk(KERN_CONT "\n");
6620 printk(KERN_ERR "ERROR: domain->cpu_power not "
6625 if (!cpumask_weight(sched_group_cpus(group))) {
6626 printk(KERN_CONT "\n");
6627 printk(KERN_ERR "ERROR: empty group\n");
6631 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6632 printk(KERN_CONT "\n");
6633 printk(KERN_ERR "ERROR: repeated CPUs\n");
6637 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6639 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6641 printk(KERN_CONT " %s", str);
6642 if (group->sgp->power != SCHED_POWER_SCALE) {
6643 printk(KERN_CONT " (cpu_power = %d)",
6647 group = group->next;
6648 } while (group != sd->groups);
6649 printk(KERN_CONT "\n");
6651 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6652 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6655 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6656 printk(KERN_ERR "ERROR: parent span is not a superset "
6657 "of domain->span\n");
6661 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6665 if (!sched_domain_debug_enabled)
6669 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6673 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6676 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6684 #else /* !CONFIG_SCHED_DEBUG */
6685 # define sched_domain_debug(sd, cpu) do { } while (0)
6686 #endif /* CONFIG_SCHED_DEBUG */
6688 static int sd_degenerate(struct sched_domain *sd)
6690 if (cpumask_weight(sched_domain_span(sd)) == 1)
6693 /* Following flags need at least 2 groups */
6694 if (sd->flags & (SD_LOAD_BALANCE |
6695 SD_BALANCE_NEWIDLE |
6699 SD_SHARE_PKG_RESOURCES)) {
6700 if (sd->groups != sd->groups->next)
6704 /* Following flags don't use groups */
6705 if (sd->flags & (SD_WAKE_AFFINE))
6712 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6714 unsigned long cflags = sd->flags, pflags = parent->flags;
6716 if (sd_degenerate(parent))
6719 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6722 /* Flags needing groups don't count if only 1 group in parent */
6723 if (parent->groups == parent->groups->next) {
6724 pflags &= ~(SD_LOAD_BALANCE |
6725 SD_BALANCE_NEWIDLE |
6729 SD_SHARE_PKG_RESOURCES);
6730 if (nr_node_ids == 1)
6731 pflags &= ~SD_SERIALIZE;
6733 if (~cflags & pflags)
6739 static void free_rootdomain(struct rcu_head *rcu)
6741 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6743 cpupri_cleanup(&rd->cpupri);
6744 free_cpumask_var(rd->rto_mask);
6745 free_cpumask_var(rd->online);
6746 free_cpumask_var(rd->span);
6750 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6752 struct root_domain *old_rd = NULL;
6753 unsigned long flags;
6755 raw_spin_lock_irqsave(&rq->lock, flags);
6760 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6763 cpumask_clear_cpu(rq->cpu, old_rd->span);
6766 * If we dont want to free the old_rt yet then
6767 * set old_rd to NULL to skip the freeing later
6770 if (!atomic_dec_and_test(&old_rd->refcount))
6774 atomic_inc(&rd->refcount);
6777 cpumask_set_cpu(rq->cpu, rd->span);
6778 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6781 raw_spin_unlock_irqrestore(&rq->lock, flags);
6784 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6787 static int init_rootdomain(struct root_domain *rd)
6789 memset(rd, 0, sizeof(*rd));
6791 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6793 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6795 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6798 if (cpupri_init(&rd->cpupri) != 0)
6803 free_cpumask_var(rd->rto_mask);
6805 free_cpumask_var(rd->online);
6807 free_cpumask_var(rd->span);
6812 static void init_defrootdomain(void)
6814 init_rootdomain(&def_root_domain);
6816 atomic_set(&def_root_domain.refcount, 1);
6819 static struct root_domain *alloc_rootdomain(void)
6821 struct root_domain *rd;
6823 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6827 if (init_rootdomain(rd) != 0) {
6835 static void free_sched_groups(struct sched_group *sg, int free_sgp)
6837 struct sched_group *tmp, *first;
6846 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
6851 } while (sg != first);
6854 static void free_sched_domain(struct rcu_head *rcu)
6856 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6859 * If its an overlapping domain it has private groups, iterate and
6862 if (sd->flags & SD_OVERLAP) {
6863 free_sched_groups(sd->groups, 1);
6864 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6865 kfree(sd->groups->sgp);
6871 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6873 call_rcu(&sd->rcu, free_sched_domain);
6876 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6878 for (; sd; sd = sd->parent)
6879 destroy_sched_domain(sd, cpu);
6883 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6884 * hold the hotplug lock.
6887 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6889 struct rq *rq = cpu_rq(cpu);
6890 struct sched_domain *tmp;
6892 /* Remove the sched domains which do not contribute to scheduling. */
6893 for (tmp = sd; tmp; ) {
6894 struct sched_domain *parent = tmp->parent;
6898 if (sd_parent_degenerate(tmp, parent)) {
6899 tmp->parent = parent->parent;
6901 parent->parent->child = tmp;
6902 destroy_sched_domain(parent, cpu);
6907 if (sd && sd_degenerate(sd)) {
6910 destroy_sched_domain(tmp, cpu);
6915 sched_domain_debug(sd, cpu);
6917 rq_attach_root(rq, rd);
6919 rcu_assign_pointer(rq->sd, sd);
6920 destroy_sched_domains(tmp, cpu);
6923 /* cpus with isolated domains */
6924 static cpumask_var_t cpu_isolated_map;
6926 /* Setup the mask of cpus configured for isolated domains */
6927 static int __init isolated_cpu_setup(char *str)
6929 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6930 cpulist_parse(str, cpu_isolated_map);
6934 __setup("isolcpus=", isolated_cpu_setup);
6936 #define SD_NODES_PER_DOMAIN 16
6941 * find_next_best_node - find the next node to include in a sched_domain
6942 * @node: node whose sched_domain we're building
6943 * @used_nodes: nodes already in the sched_domain
6945 * Find the next node to include in a given scheduling domain. Simply
6946 * finds the closest node not already in the @used_nodes map.
6948 * Should use nodemask_t.
6950 static int find_next_best_node(int node, nodemask_t *used_nodes)
6952 int i, n, val, min_val, best_node = -1;
6956 for (i = 0; i < nr_node_ids; i++) {
6957 /* Start at @node */
6958 n = (node + i) % nr_node_ids;
6960 if (!nr_cpus_node(n))
6963 /* Skip already used nodes */
6964 if (node_isset(n, *used_nodes))
6967 /* Simple min distance search */
6968 val = node_distance(node, n);
6970 if (val < min_val) {
6976 if (best_node != -1)
6977 node_set(best_node, *used_nodes);
6982 * sched_domain_node_span - get a cpumask for a node's sched_domain
6983 * @node: node whose cpumask we're constructing
6984 * @span: resulting cpumask
6986 * Given a node, construct a good cpumask for its sched_domain to span. It
6987 * should be one that prevents unnecessary balancing, but also spreads tasks
6990 static void sched_domain_node_span(int node, struct cpumask *span)
6992 nodemask_t used_nodes;
6995 cpumask_clear(span);
6996 nodes_clear(used_nodes);
6998 cpumask_or(span, span, cpumask_of_node(node));
6999 node_set(node, used_nodes);
7001 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7002 int next_node = find_next_best_node(node, &used_nodes);
7005 cpumask_or(span, span, cpumask_of_node(next_node));
7009 static const struct cpumask *cpu_node_mask(int cpu)
7011 lockdep_assert_held(&sched_domains_mutex);
7013 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
7015 return sched_domains_tmpmask;
7018 static const struct cpumask *cpu_allnodes_mask(int cpu)
7020 return cpu_possible_mask;
7022 #endif /* CONFIG_NUMA */
7024 static const struct cpumask *cpu_cpu_mask(int cpu)
7026 return cpumask_of_node(cpu_to_node(cpu));
7029 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7032 struct sched_domain **__percpu sd;
7033 struct sched_group **__percpu sg;
7034 struct sched_group_power **__percpu sgp;
7038 struct sched_domain ** __percpu sd;
7039 struct root_domain *rd;
7049 struct sched_domain_topology_level;
7051 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
7052 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
7054 #define SDTL_OVERLAP 0x01
7056 struct sched_domain_topology_level {
7057 sched_domain_init_f init;
7058 sched_domain_mask_f mask;
7060 struct sd_data data;
7064 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
7066 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
7067 const struct cpumask *span = sched_domain_span(sd);
7068 struct cpumask *covered = sched_domains_tmpmask;
7069 struct sd_data *sdd = sd->private;
7070 struct sched_domain *child;
7073 cpumask_clear(covered);
7075 for_each_cpu(i, span) {
7076 struct cpumask *sg_span;
7078 if (cpumask_test_cpu(i, covered))
7081 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7082 GFP_KERNEL, cpu_to_node(i));
7087 sg_span = sched_group_cpus(sg);
7089 child = *per_cpu_ptr(sdd->sd, i);
7091 child = child->child;
7092 cpumask_copy(sg_span, sched_domain_span(child));
7094 cpumask_set_cpu(i, sg_span);
7096 cpumask_or(covered, covered, sg_span);
7098 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
7099 atomic_inc(&sg->sgp->ref);
7101 if (cpumask_test_cpu(cpu, sg_span))
7111 sd->groups = groups;
7116 free_sched_groups(first, 0);
7121 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
7123 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
7124 struct sched_domain *child = sd->child;
7127 cpu = cpumask_first(sched_domain_span(child));
7130 *sg = *per_cpu_ptr(sdd->sg, cpu);
7131 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
7132 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
7139 * build_sched_groups will build a circular linked list of the groups
7140 * covered by the given span, and will set each group's ->cpumask correctly,
7141 * and ->cpu_power to 0.
7143 * Assumes the sched_domain tree is fully constructed
7146 build_sched_groups(struct sched_domain *sd, int cpu)
7148 struct sched_group *first = NULL, *last = NULL;
7149 struct sd_data *sdd = sd->private;
7150 const struct cpumask *span = sched_domain_span(sd);
7151 struct cpumask *covered;
7154 get_group(cpu, sdd, &sd->groups);
7155 atomic_inc(&sd->groups->ref);
7157 if (cpu != cpumask_first(sched_domain_span(sd)))
7160 lockdep_assert_held(&sched_domains_mutex);
7161 covered = sched_domains_tmpmask;
7163 cpumask_clear(covered);
7165 for_each_cpu(i, span) {
7166 struct sched_group *sg;
7167 int group = get_group(i, sdd, &sg);
7170 if (cpumask_test_cpu(i, covered))
7173 cpumask_clear(sched_group_cpus(sg));
7176 for_each_cpu(j, span) {
7177 if (get_group(j, sdd, NULL) != group)
7180 cpumask_set_cpu(j, covered);
7181 cpumask_set_cpu(j, sched_group_cpus(sg));
7196 * Initialize sched groups cpu_power.
7198 * cpu_power indicates the capacity of sched group, which is used while
7199 * distributing the load between different sched groups in a sched domain.
7200 * Typically cpu_power for all the groups in a sched domain will be same unless
7201 * there are asymmetries in the topology. If there are asymmetries, group
7202 * having more cpu_power will pickup more load compared to the group having
7205 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7207 struct sched_group *sg = sd->groups;
7209 WARN_ON(!sd || !sg);
7212 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
7214 } while (sg != sd->groups);
7216 if (cpu != group_first_cpu(sg))
7219 update_group_power(sd, cpu);
7223 * Initializers for schedule domains
7224 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7227 #ifdef CONFIG_SCHED_DEBUG
7228 # define SD_INIT_NAME(sd, type) sd->name = #type
7230 # define SD_INIT_NAME(sd, type) do { } while (0)
7233 #define SD_INIT_FUNC(type) \
7234 static noinline struct sched_domain * \
7235 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7237 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7238 *sd = SD_##type##_INIT; \
7239 SD_INIT_NAME(sd, type); \
7240 sd->private = &tl->data; \
7246 SD_INIT_FUNC(ALLNODES)
7249 #ifdef CONFIG_SCHED_SMT
7250 SD_INIT_FUNC(SIBLING)
7252 #ifdef CONFIG_SCHED_MC
7255 #ifdef CONFIG_SCHED_BOOK
7259 static int default_relax_domain_level = -1;
7260 int sched_domain_level_max;
7262 static int __init setup_relax_domain_level(char *str)
7266 val = simple_strtoul(str, NULL, 0);
7267 if (val < sched_domain_level_max)
7268 default_relax_domain_level = val;
7272 __setup("relax_domain_level=", setup_relax_domain_level);
7274 static void set_domain_attribute(struct sched_domain *sd,
7275 struct sched_domain_attr *attr)
7279 if (!attr || attr->relax_domain_level < 0) {
7280 if (default_relax_domain_level < 0)
7283 request = default_relax_domain_level;
7285 request = attr->relax_domain_level;
7286 if (request < sd->level) {
7287 /* turn off idle balance on this domain */
7288 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7290 /* turn on idle balance on this domain */
7291 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7295 static void __sdt_free(const struct cpumask *cpu_map);
7296 static int __sdt_alloc(const struct cpumask *cpu_map);
7298 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7299 const struct cpumask *cpu_map)
7303 if (!atomic_read(&d->rd->refcount))
7304 free_rootdomain(&d->rd->rcu); /* fall through */
7306 free_percpu(d->sd); /* fall through */
7308 __sdt_free(cpu_map); /* fall through */
7314 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7315 const struct cpumask *cpu_map)
7317 memset(d, 0, sizeof(*d));
7319 if (__sdt_alloc(cpu_map))
7320 return sa_sd_storage;
7321 d->sd = alloc_percpu(struct sched_domain *);
7323 return sa_sd_storage;
7324 d->rd = alloc_rootdomain();
7327 return sa_rootdomain;
7331 * NULL the sd_data elements we've used to build the sched_domain and
7332 * sched_group structure so that the subsequent __free_domain_allocs()
7333 * will not free the data we're using.
7335 static void claim_allocations(int cpu, struct sched_domain *sd)
7337 struct sd_data *sdd = sd->private;
7339 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7340 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7342 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
7343 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7345 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
7346 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
7349 #ifdef CONFIG_SCHED_SMT
7350 static const struct cpumask *cpu_smt_mask(int cpu)
7352 return topology_thread_cpumask(cpu);
7357 * Topology list, bottom-up.
7359 static struct sched_domain_topology_level default_topology[] = {
7360 #ifdef CONFIG_SCHED_SMT
7361 { sd_init_SIBLING, cpu_smt_mask, },
7363 #ifdef CONFIG_SCHED_MC
7364 { sd_init_MC, cpu_coregroup_mask, },
7366 #ifdef CONFIG_SCHED_BOOK
7367 { sd_init_BOOK, cpu_book_mask, },
7369 { sd_init_CPU, cpu_cpu_mask, },
7371 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
7372 { sd_init_ALLNODES, cpu_allnodes_mask, },
7377 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7379 static int __sdt_alloc(const struct cpumask *cpu_map)
7381 struct sched_domain_topology_level *tl;
7384 for (tl = sched_domain_topology; tl->init; tl++) {
7385 struct sd_data *sdd = &tl->data;
7387 sdd->sd = alloc_percpu(struct sched_domain *);
7391 sdd->sg = alloc_percpu(struct sched_group *);
7395 sdd->sgp = alloc_percpu(struct sched_group_power *);
7399 for_each_cpu(j, cpu_map) {
7400 struct sched_domain *sd;
7401 struct sched_group *sg;
7402 struct sched_group_power *sgp;
7404 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7405 GFP_KERNEL, cpu_to_node(j));
7409 *per_cpu_ptr(sdd->sd, j) = sd;
7411 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7412 GFP_KERNEL, cpu_to_node(j));
7416 *per_cpu_ptr(sdd->sg, j) = sg;
7418 sgp = kzalloc_node(sizeof(struct sched_group_power),
7419 GFP_KERNEL, cpu_to_node(j));
7423 *per_cpu_ptr(sdd->sgp, j) = sgp;
7430 static void __sdt_free(const struct cpumask *cpu_map)
7432 struct sched_domain_topology_level *tl;
7435 for (tl = sched_domain_topology; tl->init; tl++) {
7436 struct sd_data *sdd = &tl->data;
7438 for_each_cpu(j, cpu_map) {
7439 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
7440 if (sd && (sd->flags & SD_OVERLAP))
7441 free_sched_groups(sd->groups, 0);
7442 kfree(*per_cpu_ptr(sdd->sg, j));
7443 kfree(*per_cpu_ptr(sdd->sgp, j));
7445 free_percpu(sdd->sd);
7446 free_percpu(sdd->sg);
7447 free_percpu(sdd->sgp);
7451 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7452 struct s_data *d, const struct cpumask *cpu_map,
7453 struct sched_domain_attr *attr, struct sched_domain *child,
7456 struct sched_domain *sd = tl->init(tl, cpu);
7460 set_domain_attribute(sd, attr);
7461 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7463 sd->level = child->level + 1;
7464 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7473 * Build sched domains for a given set of cpus and attach the sched domains
7474 * to the individual cpus
7476 static int build_sched_domains(const struct cpumask *cpu_map,
7477 struct sched_domain_attr *attr)
7479 enum s_alloc alloc_state = sa_none;
7480 struct sched_domain *sd;
7482 int i, ret = -ENOMEM;
7484 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7485 if (alloc_state != sa_rootdomain)
7488 /* Set up domains for cpus specified by the cpu_map. */
7489 for_each_cpu(i, cpu_map) {
7490 struct sched_domain_topology_level *tl;
7493 for (tl = sched_domain_topology; tl->init; tl++) {
7494 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7495 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7496 sd->flags |= SD_OVERLAP;
7497 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7504 *per_cpu_ptr(d.sd, i) = sd;
7507 /* Build the groups for the domains */
7508 for_each_cpu(i, cpu_map) {
7509 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7510 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7511 if (sd->flags & SD_OVERLAP) {
7512 if (build_overlap_sched_groups(sd, i))
7515 if (build_sched_groups(sd, i))
7521 /* Calculate CPU power for physical packages and nodes */
7522 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7523 if (!cpumask_test_cpu(i, cpu_map))
7526 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7527 claim_allocations(i, sd);
7528 init_sched_groups_power(i, sd);
7532 /* Attach the domains */
7534 for_each_cpu(i, cpu_map) {
7535 sd = *per_cpu_ptr(d.sd, i);
7536 cpu_attach_domain(sd, d.rd, i);
7542 __free_domain_allocs(&d, alloc_state, cpu_map);
7546 static cpumask_var_t *doms_cur; /* current sched domains */
7547 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7548 static struct sched_domain_attr *dattr_cur;
7549 /* attribues of custom domains in 'doms_cur' */
7552 * Special case: If a kmalloc of a doms_cur partition (array of
7553 * cpumask) fails, then fallback to a single sched domain,
7554 * as determined by the single cpumask fallback_doms.
7556 static cpumask_var_t fallback_doms;
7559 * arch_update_cpu_topology lets virtualized architectures update the
7560 * cpu core maps. It is supposed to return 1 if the topology changed
7561 * or 0 if it stayed the same.
7563 int __attribute__((weak)) arch_update_cpu_topology(void)
7568 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7571 cpumask_var_t *doms;
7573 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7576 for (i = 0; i < ndoms; i++) {
7577 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7578 free_sched_domains(doms, i);
7585 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7588 for (i = 0; i < ndoms; i++)
7589 free_cpumask_var(doms[i]);
7594 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7595 * For now this just excludes isolated cpus, but could be used to
7596 * exclude other special cases in the future.
7598 static int init_sched_domains(const struct cpumask *cpu_map)
7602 arch_update_cpu_topology();
7604 doms_cur = alloc_sched_domains(ndoms_cur);
7606 doms_cur = &fallback_doms;
7607 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7609 err = build_sched_domains(doms_cur[0], NULL);
7610 register_sched_domain_sysctl();
7616 * Detach sched domains from a group of cpus specified in cpu_map
7617 * These cpus will now be attached to the NULL domain
7619 static void detach_destroy_domains(const struct cpumask *cpu_map)
7624 for_each_cpu(i, cpu_map)
7625 cpu_attach_domain(NULL, &def_root_domain, i);
7629 /* handle null as "default" */
7630 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7631 struct sched_domain_attr *new, int idx_new)
7633 struct sched_domain_attr tmp;
7640 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7641 new ? (new + idx_new) : &tmp,
7642 sizeof(struct sched_domain_attr));
7646 * Partition sched domains as specified by the 'ndoms_new'
7647 * cpumasks in the array doms_new[] of cpumasks. This compares
7648 * doms_new[] to the current sched domain partitioning, doms_cur[].
7649 * It destroys each deleted domain and builds each new domain.
7651 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7652 * The masks don't intersect (don't overlap.) We should setup one
7653 * sched domain for each mask. CPUs not in any of the cpumasks will
7654 * not be load balanced. If the same cpumask appears both in the
7655 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7658 * The passed in 'doms_new' should be allocated using
7659 * alloc_sched_domains. This routine takes ownership of it and will
7660 * free_sched_domains it when done with it. If the caller failed the
7661 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7662 * and partition_sched_domains() will fallback to the single partition
7663 * 'fallback_doms', it also forces the domains to be rebuilt.
7665 * If doms_new == NULL it will be replaced with cpu_online_mask.
7666 * ndoms_new == 0 is a special case for destroying existing domains,
7667 * and it will not create the default domain.
7669 * Call with hotplug lock held
7671 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7672 struct sched_domain_attr *dattr_new)
7677 mutex_lock(&sched_domains_mutex);
7679 /* always unregister in case we don't destroy any domains */
7680 unregister_sched_domain_sysctl();
7682 /* Let architecture update cpu core mappings. */
7683 new_topology = arch_update_cpu_topology();
7685 n = doms_new ? ndoms_new : 0;
7687 /* Destroy deleted domains */
7688 for (i = 0; i < ndoms_cur; i++) {
7689 for (j = 0; j < n && !new_topology; j++) {
7690 if (cpumask_equal(doms_cur[i], doms_new[j])
7691 && dattrs_equal(dattr_cur, i, dattr_new, j))
7694 /* no match - a current sched domain not in new doms_new[] */
7695 detach_destroy_domains(doms_cur[i]);
7700 if (doms_new == NULL) {
7702 doms_new = &fallback_doms;
7703 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7704 WARN_ON_ONCE(dattr_new);
7707 /* Build new domains */
7708 for (i = 0; i < ndoms_new; i++) {
7709 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7710 if (cpumask_equal(doms_new[i], doms_cur[j])
7711 && dattrs_equal(dattr_new, i, dattr_cur, j))
7714 /* no match - add a new doms_new */
7715 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7720 /* Remember the new sched domains */
7721 if (doms_cur != &fallback_doms)
7722 free_sched_domains(doms_cur, ndoms_cur);
7723 kfree(dattr_cur); /* kfree(NULL) is safe */
7724 doms_cur = doms_new;
7725 dattr_cur = dattr_new;
7726 ndoms_cur = ndoms_new;
7728 register_sched_domain_sysctl();
7730 mutex_unlock(&sched_domains_mutex);
7733 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7734 static void reinit_sched_domains(void)
7738 /* Destroy domains first to force the rebuild */
7739 partition_sched_domains(0, NULL, NULL);
7741 rebuild_sched_domains();
7745 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7747 unsigned int level = 0;
7749 if (sscanf(buf, "%u", &level) != 1)
7753 * level is always be positive so don't check for
7754 * level < POWERSAVINGS_BALANCE_NONE which is 0
7755 * What happens on 0 or 1 byte write,
7756 * need to check for count as well?
7759 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7763 sched_smt_power_savings = level;
7765 sched_mc_power_savings = level;
7767 reinit_sched_domains();
7772 #ifdef CONFIG_SCHED_MC
7773 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7774 struct sysdev_class_attribute *attr,
7777 return sprintf(page, "%u\n", sched_mc_power_savings);
7779 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7780 struct sysdev_class_attribute *attr,
7781 const char *buf, size_t count)
7783 return sched_power_savings_store(buf, count, 0);
7785 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7786 sched_mc_power_savings_show,
7787 sched_mc_power_savings_store);
7790 #ifdef CONFIG_SCHED_SMT
7791 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7792 struct sysdev_class_attribute *attr,
7795 return sprintf(page, "%u\n", sched_smt_power_savings);
7797 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7798 struct sysdev_class_attribute *attr,
7799 const char *buf, size_t count)
7801 return sched_power_savings_store(buf, count, 1);
7803 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7804 sched_smt_power_savings_show,
7805 sched_smt_power_savings_store);
7808 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7812 #ifdef CONFIG_SCHED_SMT
7814 err = sysfs_create_file(&cls->kset.kobj,
7815 &attr_sched_smt_power_savings.attr);
7817 #ifdef CONFIG_SCHED_MC
7818 if (!err && mc_capable())
7819 err = sysfs_create_file(&cls->kset.kobj,
7820 &attr_sched_mc_power_savings.attr);
7824 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7827 * Update cpusets according to cpu_active mask. If cpusets are
7828 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7829 * around partition_sched_domains().
7831 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7834 switch (action & ~CPU_TASKS_FROZEN) {
7836 case CPU_DOWN_FAILED:
7837 cpuset_update_active_cpus();
7844 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7847 switch (action & ~CPU_TASKS_FROZEN) {
7848 case CPU_DOWN_PREPARE:
7849 cpuset_update_active_cpus();
7856 static int update_runtime(struct notifier_block *nfb,
7857 unsigned long action, void *hcpu)
7859 int cpu = (int)(long)hcpu;
7862 case CPU_DOWN_PREPARE:
7863 case CPU_DOWN_PREPARE_FROZEN:
7864 disable_runtime(cpu_rq(cpu));
7867 case CPU_DOWN_FAILED:
7868 case CPU_DOWN_FAILED_FROZEN:
7870 case CPU_ONLINE_FROZEN:
7871 enable_runtime(cpu_rq(cpu));
7879 void __init sched_init_smp(void)
7881 cpumask_var_t non_isolated_cpus;
7883 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7884 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7887 mutex_lock(&sched_domains_mutex);
7888 init_sched_domains(cpu_active_mask);
7889 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7890 if (cpumask_empty(non_isolated_cpus))
7891 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7892 mutex_unlock(&sched_domains_mutex);
7895 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7896 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7898 /* RT runtime code needs to handle some hotplug events */
7899 hotcpu_notifier(update_runtime, 0);
7903 /* Move init over to a non-isolated CPU */
7904 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7906 sched_init_granularity();
7907 free_cpumask_var(non_isolated_cpus);
7909 init_sched_rt_class();
7912 void __init sched_init_smp(void)
7914 sched_init_granularity();
7916 #endif /* CONFIG_SMP */
7918 const_debug unsigned int sysctl_timer_migration = 1;
7920 int in_sched_functions(unsigned long addr)
7922 return in_lock_functions(addr) ||
7923 (addr >= (unsigned long)__sched_text_start
7924 && addr < (unsigned long)__sched_text_end);
7927 static void init_cfs_rq(struct cfs_rq *cfs_rq)
7929 cfs_rq->tasks_timeline = RB_ROOT;
7930 INIT_LIST_HEAD(&cfs_rq->tasks);
7931 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7932 #ifndef CONFIG_64BIT
7933 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7937 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7939 struct rt_prio_array *array;
7942 array = &rt_rq->active;
7943 for (i = 0; i < MAX_RT_PRIO; i++) {
7944 INIT_LIST_HEAD(array->queue + i);
7945 __clear_bit(i, array->bitmap);
7947 /* delimiter for bitsearch: */
7948 __set_bit(MAX_RT_PRIO, array->bitmap);
7950 #if defined CONFIG_SMP
7951 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7952 rt_rq->highest_prio.next = MAX_RT_PRIO;
7953 rt_rq->rt_nr_migratory = 0;
7954 rt_rq->overloaded = 0;
7955 plist_head_init(&rt_rq->pushable_tasks);
7959 rt_rq->rt_throttled = 0;
7960 rt_rq->rt_runtime = 0;
7961 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7964 #ifdef CONFIG_FAIR_GROUP_SCHED
7965 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7966 struct sched_entity *se, int cpu,
7967 struct sched_entity *parent)
7969 struct rq *rq = cpu_rq(cpu);
7974 /* allow initial update_cfs_load() to truncate */
7975 cfs_rq->load_stamp = 1;
7978 tg->cfs_rq[cpu] = cfs_rq;
7981 /* se could be NULL for root_task_group */
7986 se->cfs_rq = &rq->cfs;
7988 se->cfs_rq = parent->my_q;
7991 update_load_set(&se->load, 0);
7992 se->parent = parent;
7996 #ifdef CONFIG_RT_GROUP_SCHED
7997 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7998 struct sched_rt_entity *rt_se, int cpu,
7999 struct sched_rt_entity *parent)
8001 struct rq *rq = cpu_rq(cpu);
8003 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8004 rt_rq->rt_nr_boosted = 0;
8008 tg->rt_rq[cpu] = rt_rq;
8009 tg->rt_se[cpu] = rt_se;
8015 rt_se->rt_rq = &rq->rt;
8017 rt_se->rt_rq = parent->my_q;
8019 rt_se->my_q = rt_rq;
8020 rt_se->parent = parent;
8021 INIT_LIST_HEAD(&rt_se->run_list);
8025 void __init sched_init(void)
8028 unsigned long alloc_size = 0, ptr;
8030 #ifdef CONFIG_FAIR_GROUP_SCHED
8031 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8033 #ifdef CONFIG_RT_GROUP_SCHED
8034 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8036 #ifdef CONFIG_CPUMASK_OFFSTACK
8037 alloc_size += num_possible_cpus() * cpumask_size();
8040 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8042 #ifdef CONFIG_FAIR_GROUP_SCHED
8043 root_task_group.se = (struct sched_entity **)ptr;
8044 ptr += nr_cpu_ids * sizeof(void **);
8046 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8047 ptr += nr_cpu_ids * sizeof(void **);
8049 #endif /* CONFIG_FAIR_GROUP_SCHED */
8050 #ifdef CONFIG_RT_GROUP_SCHED
8051 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8052 ptr += nr_cpu_ids * sizeof(void **);
8054 root_task_group.rt_rq = (struct rt_rq **)ptr;
8055 ptr += nr_cpu_ids * sizeof(void **);
8057 #endif /* CONFIG_RT_GROUP_SCHED */
8058 #ifdef CONFIG_CPUMASK_OFFSTACK
8059 for_each_possible_cpu(i) {
8060 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8061 ptr += cpumask_size();
8063 #endif /* CONFIG_CPUMASK_OFFSTACK */
8067 init_defrootdomain();
8070 init_rt_bandwidth(&def_rt_bandwidth,
8071 global_rt_period(), global_rt_runtime());
8073 #ifdef CONFIG_RT_GROUP_SCHED
8074 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8075 global_rt_period(), global_rt_runtime());
8076 #endif /* CONFIG_RT_GROUP_SCHED */
8078 #ifdef CONFIG_CGROUP_SCHED
8079 list_add(&root_task_group.list, &task_groups);
8080 INIT_LIST_HEAD(&root_task_group.children);
8081 autogroup_init(&init_task);
8082 #endif /* CONFIG_CGROUP_SCHED */
8084 for_each_possible_cpu(i) {
8088 raw_spin_lock_init(&rq->lock);
8090 rq->calc_load_active = 0;
8091 rq->calc_load_update = jiffies + LOAD_FREQ;
8092 init_cfs_rq(&rq->cfs);
8093 init_rt_rq(&rq->rt, rq);
8094 #ifdef CONFIG_FAIR_GROUP_SCHED
8095 root_task_group.shares = root_task_group_load;
8096 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8098 * How much cpu bandwidth does root_task_group get?
8100 * In case of task-groups formed thr' the cgroup filesystem, it
8101 * gets 100% of the cpu resources in the system. This overall
8102 * system cpu resource is divided among the tasks of
8103 * root_task_group and its child task-groups in a fair manner,
8104 * based on each entity's (task or task-group's) weight
8105 * (se->load.weight).
8107 * In other words, if root_task_group has 10 tasks of weight
8108 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8109 * then A0's share of the cpu resource is:
8111 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8113 * We achieve this by letting root_task_group's tasks sit
8114 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8116 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8117 #endif /* CONFIG_FAIR_GROUP_SCHED */
8119 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8120 #ifdef CONFIG_RT_GROUP_SCHED
8121 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8122 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8125 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8126 rq->cpu_load[j] = 0;
8128 rq->last_load_update_tick = jiffies;
8133 rq->cpu_power = SCHED_POWER_SCALE;
8134 rq->post_schedule = 0;
8135 rq->active_balance = 0;
8136 rq->next_balance = jiffies;
8141 rq->avg_idle = 2*sysctl_sched_migration_cost;
8142 rq_attach_root(rq, &def_root_domain);
8144 rq->nohz_balance_kick = 0;
8145 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8149 atomic_set(&rq->nr_iowait, 0);
8152 set_load_weight(&init_task);
8154 #ifdef CONFIG_PREEMPT_NOTIFIERS
8155 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8159 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8162 #ifdef CONFIG_RT_MUTEXES
8163 plist_head_init(&init_task.pi_waiters);
8167 * The boot idle thread does lazy MMU switching as well:
8169 atomic_inc(&init_mm.mm_count);
8170 enter_lazy_tlb(&init_mm, current);
8173 * Make us the idle thread. Technically, schedule() should not be
8174 * called from this thread, however somewhere below it might be,
8175 * but because we are the idle thread, we just pick up running again
8176 * when this runqueue becomes "idle".
8178 init_idle(current, smp_processor_id());
8180 calc_load_update = jiffies + LOAD_FREQ;
8183 * During early bootup we pretend to be a normal task:
8185 current->sched_class = &fair_sched_class;
8187 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8188 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8190 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8192 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8193 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8194 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8195 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8196 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8198 /* May be allocated at isolcpus cmdline parse time */
8199 if (cpu_isolated_map == NULL)
8200 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8203 scheduler_running = 1;
8206 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8207 static inline int preempt_count_equals(int preempt_offset)
8209 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8211 return (nested == preempt_offset);
8214 void __might_sleep(const char *file, int line, int preempt_offset)
8216 static unsigned long prev_jiffy; /* ratelimiting */
8218 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8219 system_state != SYSTEM_RUNNING || oops_in_progress)
8221 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8223 prev_jiffy = jiffies;
8226 "BUG: sleeping function called from invalid context at %s:%d\n",
8229 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8230 in_atomic(), irqs_disabled(),
8231 current->pid, current->comm);
8233 debug_show_held_locks(current);
8234 if (irqs_disabled())
8235 print_irqtrace_events(current);
8238 EXPORT_SYMBOL(__might_sleep);
8241 #ifdef CONFIG_MAGIC_SYSRQ
8242 static void normalize_task(struct rq *rq, struct task_struct *p)
8244 const struct sched_class *prev_class = p->sched_class;
8245 int old_prio = p->prio;
8250 deactivate_task(rq, p, 0);
8251 __setscheduler(rq, p, SCHED_NORMAL, 0);
8253 activate_task(rq, p, 0);
8254 resched_task(rq->curr);
8257 check_class_changed(rq, p, prev_class, old_prio);
8260 void normalize_rt_tasks(void)
8262 struct task_struct *g, *p;
8263 unsigned long flags;
8266 read_lock_irqsave(&tasklist_lock, flags);
8267 do_each_thread(g, p) {
8269 * Only normalize user tasks:
8274 p->se.exec_start = 0;
8275 #ifdef CONFIG_SCHEDSTATS
8276 p->se.statistics.wait_start = 0;
8277 p->se.statistics.sleep_start = 0;
8278 p->se.statistics.block_start = 0;
8283 * Renice negative nice level userspace
8286 if (TASK_NICE(p) < 0 && p->mm)
8287 set_user_nice(p, 0);
8291 raw_spin_lock(&p->pi_lock);
8292 rq = __task_rq_lock(p);
8294 normalize_task(rq, p);
8296 __task_rq_unlock(rq);
8297 raw_spin_unlock(&p->pi_lock);
8298 } while_each_thread(g, p);
8300 read_unlock_irqrestore(&tasklist_lock, flags);
8303 #endif /* CONFIG_MAGIC_SYSRQ */
8305 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8307 * These functions are only useful for the IA64 MCA handling, or kdb.
8309 * They can only be called when the whole system has been
8310 * stopped - every CPU needs to be quiescent, and no scheduling
8311 * activity can take place. Using them for anything else would
8312 * be a serious bug, and as a result, they aren't even visible
8313 * under any other configuration.
8317 * curr_task - return the current task for a given cpu.
8318 * @cpu: the processor in question.
8320 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8322 struct task_struct *curr_task(int cpu)
8324 return cpu_curr(cpu);
8327 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8331 * set_curr_task - set the current task for a given cpu.
8332 * @cpu: the processor in question.
8333 * @p: the task pointer to set.
8335 * Description: This function must only be used when non-maskable interrupts
8336 * are serviced on a separate stack. It allows the architecture to switch the
8337 * notion of the current task on a cpu in a non-blocking manner. This function
8338 * must be called with all CPU's synchronized, and interrupts disabled, the
8339 * and caller must save the original value of the current task (see
8340 * curr_task() above) and restore that value before reenabling interrupts and
8341 * re-starting the system.
8343 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8345 void set_curr_task(int cpu, struct task_struct *p)
8352 #ifdef CONFIG_FAIR_GROUP_SCHED
8353 static void free_fair_sched_group(struct task_group *tg)
8357 for_each_possible_cpu(i) {
8359 kfree(tg->cfs_rq[i]);
8369 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8371 struct cfs_rq *cfs_rq;
8372 struct sched_entity *se;
8375 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8378 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8382 tg->shares = NICE_0_LOAD;
8384 for_each_possible_cpu(i) {
8385 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8386 GFP_KERNEL, cpu_to_node(i));
8390 se = kzalloc_node(sizeof(struct sched_entity),
8391 GFP_KERNEL, cpu_to_node(i));
8395 init_cfs_rq(cfs_rq);
8396 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8407 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8409 struct rq *rq = cpu_rq(cpu);
8410 unsigned long flags;
8413 * Only empty task groups can be destroyed; so we can speculatively
8414 * check on_list without danger of it being re-added.
8416 if (!tg->cfs_rq[cpu]->on_list)
8419 raw_spin_lock_irqsave(&rq->lock, flags);
8420 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8421 raw_spin_unlock_irqrestore(&rq->lock, flags);
8423 #else /* !CONFIG_FAIR_GROUP_SCHED */
8424 static inline void free_fair_sched_group(struct task_group *tg)
8429 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8434 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8437 #endif /* CONFIG_FAIR_GROUP_SCHED */
8439 #ifdef CONFIG_RT_GROUP_SCHED
8440 static void free_rt_sched_group(struct task_group *tg)
8445 destroy_rt_bandwidth(&tg->rt_bandwidth);
8447 for_each_possible_cpu(i) {
8449 kfree(tg->rt_rq[i]);
8451 kfree(tg->rt_se[i]);
8459 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8461 struct rt_rq *rt_rq;
8462 struct sched_rt_entity *rt_se;
8465 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8468 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8472 init_rt_bandwidth(&tg->rt_bandwidth,
8473 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8475 for_each_possible_cpu(i) {
8476 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8477 GFP_KERNEL, cpu_to_node(i));
8481 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8482 GFP_KERNEL, cpu_to_node(i));
8486 init_rt_rq(rt_rq, cpu_rq(i));
8487 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8488 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8498 #else /* !CONFIG_RT_GROUP_SCHED */
8499 static inline void free_rt_sched_group(struct task_group *tg)
8504 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8508 #endif /* CONFIG_RT_GROUP_SCHED */
8510 #ifdef CONFIG_CGROUP_SCHED
8511 static void free_sched_group(struct task_group *tg)
8513 free_fair_sched_group(tg);
8514 free_rt_sched_group(tg);
8519 /* allocate runqueue etc for a new task group */
8520 struct task_group *sched_create_group(struct task_group *parent)
8522 struct task_group *tg;
8523 unsigned long flags;
8525 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8527 return ERR_PTR(-ENOMEM);
8529 if (!alloc_fair_sched_group(tg, parent))
8532 if (!alloc_rt_sched_group(tg, parent))
8535 spin_lock_irqsave(&task_group_lock, flags);
8536 list_add_rcu(&tg->list, &task_groups);
8538 WARN_ON(!parent); /* root should already exist */
8540 tg->parent = parent;
8541 INIT_LIST_HEAD(&tg->children);
8542 list_add_rcu(&tg->siblings, &parent->children);
8543 spin_unlock_irqrestore(&task_group_lock, flags);
8548 free_sched_group(tg);
8549 return ERR_PTR(-ENOMEM);
8552 /* rcu callback to free various structures associated with a task group */
8553 static void free_sched_group_rcu(struct rcu_head *rhp)
8555 /* now it should be safe to free those cfs_rqs */
8556 free_sched_group(container_of(rhp, struct task_group, rcu));
8559 /* Destroy runqueue etc associated with a task group */
8560 void sched_destroy_group(struct task_group *tg)
8562 unsigned long flags;
8565 /* end participation in shares distribution */
8566 for_each_possible_cpu(i)
8567 unregister_fair_sched_group(tg, i);
8569 spin_lock_irqsave(&task_group_lock, flags);
8570 list_del_rcu(&tg->list);
8571 list_del_rcu(&tg->siblings);
8572 spin_unlock_irqrestore(&task_group_lock, flags);
8574 /* wait for possible concurrent references to cfs_rqs complete */
8575 call_rcu(&tg->rcu, free_sched_group_rcu);
8578 /* change task's runqueue when it moves between groups.
8579 * The caller of this function should have put the task in its new group
8580 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8581 * reflect its new group.
8583 void sched_move_task(struct task_struct *tsk)
8586 unsigned long flags;
8589 rq = task_rq_lock(tsk, &flags);
8591 running = task_current(rq, tsk);
8595 dequeue_task(rq, tsk, 0);
8596 if (unlikely(running))
8597 tsk->sched_class->put_prev_task(rq, tsk);
8599 #ifdef CONFIG_FAIR_GROUP_SCHED
8600 if (tsk->sched_class->task_move_group)
8601 tsk->sched_class->task_move_group(tsk, on_rq);
8604 set_task_rq(tsk, task_cpu(tsk));
8606 if (unlikely(running))
8607 tsk->sched_class->set_curr_task(rq);
8609 enqueue_task(rq, tsk, 0);
8611 task_rq_unlock(rq, tsk, &flags);
8613 #endif /* CONFIG_CGROUP_SCHED */
8615 #ifdef CONFIG_FAIR_GROUP_SCHED
8616 static DEFINE_MUTEX(shares_mutex);
8618 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8621 unsigned long flags;
8624 * We can't change the weight of the root cgroup.
8629 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8631 mutex_lock(&shares_mutex);
8632 if (tg->shares == shares)
8635 tg->shares = shares;
8636 for_each_possible_cpu(i) {
8637 struct rq *rq = cpu_rq(i);
8638 struct sched_entity *se;
8641 /* Propagate contribution to hierarchy */
8642 raw_spin_lock_irqsave(&rq->lock, flags);
8643 for_each_sched_entity(se)
8644 update_cfs_shares(group_cfs_rq(se));
8645 raw_spin_unlock_irqrestore(&rq->lock, flags);
8649 mutex_unlock(&shares_mutex);
8653 unsigned long sched_group_shares(struct task_group *tg)
8659 #ifdef CONFIG_RT_GROUP_SCHED
8661 * Ensure that the real time constraints are schedulable.
8663 static DEFINE_MUTEX(rt_constraints_mutex);
8665 static unsigned long to_ratio(u64 period, u64 runtime)
8667 if (runtime == RUNTIME_INF)
8670 return div64_u64(runtime << 20, period);
8673 /* Must be called with tasklist_lock held */
8674 static inline int tg_has_rt_tasks(struct task_group *tg)
8676 struct task_struct *g, *p;
8678 do_each_thread(g, p) {
8679 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8681 } while_each_thread(g, p);
8686 struct rt_schedulable_data {
8687 struct task_group *tg;
8692 static int tg_schedulable(struct task_group *tg, void *data)
8694 struct rt_schedulable_data *d = data;
8695 struct task_group *child;
8696 unsigned long total, sum = 0;
8697 u64 period, runtime;
8699 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8700 runtime = tg->rt_bandwidth.rt_runtime;
8703 period = d->rt_period;
8704 runtime = d->rt_runtime;
8708 * Cannot have more runtime than the period.
8710 if (runtime > period && runtime != RUNTIME_INF)
8714 * Ensure we don't starve existing RT tasks.
8716 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8719 total = to_ratio(period, runtime);
8722 * Nobody can have more than the global setting allows.
8724 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8728 * The sum of our children's runtime should not exceed our own.
8730 list_for_each_entry_rcu(child, &tg->children, siblings) {
8731 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8732 runtime = child->rt_bandwidth.rt_runtime;
8734 if (child == d->tg) {
8735 period = d->rt_period;
8736 runtime = d->rt_runtime;
8739 sum += to_ratio(period, runtime);
8748 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8750 struct rt_schedulable_data data = {
8752 .rt_period = period,
8753 .rt_runtime = runtime,
8756 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8759 static int tg_set_bandwidth(struct task_group *tg,
8760 u64 rt_period, u64 rt_runtime)
8764 mutex_lock(&rt_constraints_mutex);
8765 read_lock(&tasklist_lock);
8766 err = __rt_schedulable(tg, rt_period, rt_runtime);
8770 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8771 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8772 tg->rt_bandwidth.rt_runtime = rt_runtime;
8774 for_each_possible_cpu(i) {
8775 struct rt_rq *rt_rq = tg->rt_rq[i];
8777 raw_spin_lock(&rt_rq->rt_runtime_lock);
8778 rt_rq->rt_runtime = rt_runtime;
8779 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8781 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8783 read_unlock(&tasklist_lock);
8784 mutex_unlock(&rt_constraints_mutex);
8789 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8791 u64 rt_runtime, rt_period;
8793 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8794 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8795 if (rt_runtime_us < 0)
8796 rt_runtime = RUNTIME_INF;
8798 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8801 long sched_group_rt_runtime(struct task_group *tg)
8805 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8808 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8809 do_div(rt_runtime_us, NSEC_PER_USEC);
8810 return rt_runtime_us;
8813 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8815 u64 rt_runtime, rt_period;
8817 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8818 rt_runtime = tg->rt_bandwidth.rt_runtime;
8823 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8826 long sched_group_rt_period(struct task_group *tg)
8830 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8831 do_div(rt_period_us, NSEC_PER_USEC);
8832 return rt_period_us;
8835 static int sched_rt_global_constraints(void)
8837 u64 runtime, period;
8840 if (sysctl_sched_rt_period <= 0)
8843 runtime = global_rt_runtime();
8844 period = global_rt_period();
8847 * Sanity check on the sysctl variables.
8849 if (runtime > period && runtime != RUNTIME_INF)
8852 mutex_lock(&rt_constraints_mutex);
8853 read_lock(&tasklist_lock);
8854 ret = __rt_schedulable(NULL, 0, 0);
8855 read_unlock(&tasklist_lock);
8856 mutex_unlock(&rt_constraints_mutex);
8861 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8863 /* Don't accept realtime tasks when there is no way for them to run */
8864 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8870 #else /* !CONFIG_RT_GROUP_SCHED */
8871 static int sched_rt_global_constraints(void)
8873 unsigned long flags;
8876 if (sysctl_sched_rt_period <= 0)
8880 * There's always some RT tasks in the root group
8881 * -- migration, kstopmachine etc..
8883 if (sysctl_sched_rt_runtime == 0)
8886 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8887 for_each_possible_cpu(i) {
8888 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8890 raw_spin_lock(&rt_rq->rt_runtime_lock);
8891 rt_rq->rt_runtime = global_rt_runtime();
8892 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8894 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8898 #endif /* CONFIG_RT_GROUP_SCHED */
8900 int sched_rt_handler(struct ctl_table *table, int write,
8901 void __user *buffer, size_t *lenp,
8905 int old_period, old_runtime;
8906 static DEFINE_MUTEX(mutex);
8909 old_period = sysctl_sched_rt_period;
8910 old_runtime = sysctl_sched_rt_runtime;
8912 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8914 if (!ret && write) {
8915 ret = sched_rt_global_constraints();
8917 sysctl_sched_rt_period = old_period;
8918 sysctl_sched_rt_runtime = old_runtime;
8920 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8921 def_rt_bandwidth.rt_period =
8922 ns_to_ktime(global_rt_period());
8925 mutex_unlock(&mutex);
8930 #ifdef CONFIG_CGROUP_SCHED
8932 /* return corresponding task_group object of a cgroup */
8933 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8935 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8936 struct task_group, css);
8939 static struct cgroup_subsys_state *
8940 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8942 struct task_group *tg, *parent;
8944 if (!cgrp->parent) {
8945 /* This is early initialization for the top cgroup */
8946 return &root_task_group.css;
8949 parent = cgroup_tg(cgrp->parent);
8950 tg = sched_create_group(parent);
8952 return ERR_PTR(-ENOMEM);
8958 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8960 struct task_group *tg = cgroup_tg(cgrp);
8962 sched_destroy_group(tg);
8966 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8968 #ifdef CONFIG_RT_GROUP_SCHED
8969 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8972 /* We don't support RT-tasks being in separate groups */
8973 if (tsk->sched_class != &fair_sched_class)
8980 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8982 sched_move_task(tsk);
8986 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
8987 struct cgroup *old_cgrp, struct task_struct *task)
8990 * cgroup_exit() is called in the copy_process() failure path.
8991 * Ignore this case since the task hasn't ran yet, this avoids
8992 * trying to poke a half freed task state from generic code.
8994 if (!(task->flags & PF_EXITING))
8997 sched_move_task(task);
9000 #ifdef CONFIG_FAIR_GROUP_SCHED
9001 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9004 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
9007 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9009 struct task_group *tg = cgroup_tg(cgrp);
9011 return (u64) scale_load_down(tg->shares);
9013 #endif /* CONFIG_FAIR_GROUP_SCHED */
9015 #ifdef CONFIG_RT_GROUP_SCHED
9016 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9019 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9022 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9024 return sched_group_rt_runtime(cgroup_tg(cgrp));
9027 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9030 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9033 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9035 return sched_group_rt_period(cgroup_tg(cgrp));
9037 #endif /* CONFIG_RT_GROUP_SCHED */
9039 static struct cftype cpu_files[] = {
9040 #ifdef CONFIG_FAIR_GROUP_SCHED
9043 .read_u64 = cpu_shares_read_u64,
9044 .write_u64 = cpu_shares_write_u64,
9047 #ifdef CONFIG_RT_GROUP_SCHED
9049 .name = "rt_runtime_us",
9050 .read_s64 = cpu_rt_runtime_read,
9051 .write_s64 = cpu_rt_runtime_write,
9054 .name = "rt_period_us",
9055 .read_u64 = cpu_rt_period_read_uint,
9056 .write_u64 = cpu_rt_period_write_uint,
9061 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9063 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9066 struct cgroup_subsys cpu_cgroup_subsys = {
9068 .create = cpu_cgroup_create,
9069 .destroy = cpu_cgroup_destroy,
9070 .can_attach_task = cpu_cgroup_can_attach_task,
9071 .attach_task = cpu_cgroup_attach_task,
9072 .exit = cpu_cgroup_exit,
9073 .populate = cpu_cgroup_populate,
9074 .subsys_id = cpu_cgroup_subsys_id,
9078 #endif /* CONFIG_CGROUP_SCHED */
9080 #ifdef CONFIG_CGROUP_CPUACCT
9083 * CPU accounting code for task groups.
9085 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9086 * (balbir@in.ibm.com).
9089 /* track cpu usage of a group of tasks and its child groups */
9091 struct cgroup_subsys_state css;
9092 /* cpuusage holds pointer to a u64-type object on every cpu */
9093 u64 __percpu *cpuusage;
9094 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9095 struct cpuacct *parent;
9098 struct cgroup_subsys cpuacct_subsys;
9100 /* return cpu accounting group corresponding to this container */
9101 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9103 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9104 struct cpuacct, css);
9107 /* return cpu accounting group to which this task belongs */
9108 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9110 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9111 struct cpuacct, css);
9114 /* create a new cpu accounting group */
9115 static struct cgroup_subsys_state *cpuacct_create(
9116 struct cgroup_subsys *ss, struct cgroup *cgrp)
9118 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9124 ca->cpuusage = alloc_percpu(u64);
9128 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9129 if (percpu_counter_init(&ca->cpustat[i], 0))
9130 goto out_free_counters;
9133 ca->parent = cgroup_ca(cgrp->parent);
9139 percpu_counter_destroy(&ca->cpustat[i]);
9140 free_percpu(ca->cpuusage);
9144 return ERR_PTR(-ENOMEM);
9147 /* destroy an existing cpu accounting group */
9149 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9151 struct cpuacct *ca = cgroup_ca(cgrp);
9154 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9155 percpu_counter_destroy(&ca->cpustat[i]);
9156 free_percpu(ca->cpuusage);
9160 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9162 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9165 #ifndef CONFIG_64BIT
9167 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9169 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9171 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9179 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9181 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9183 #ifndef CONFIG_64BIT
9185 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9187 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9189 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9195 /* return total cpu usage (in nanoseconds) of a group */
9196 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9198 struct cpuacct *ca = cgroup_ca(cgrp);
9199 u64 totalcpuusage = 0;
9202 for_each_present_cpu(i)
9203 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9205 return totalcpuusage;
9208 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9211 struct cpuacct *ca = cgroup_ca(cgrp);
9220 for_each_present_cpu(i)
9221 cpuacct_cpuusage_write(ca, i, 0);
9227 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9230 struct cpuacct *ca = cgroup_ca(cgroup);
9234 for_each_present_cpu(i) {
9235 percpu = cpuacct_cpuusage_read(ca, i);
9236 seq_printf(m, "%llu ", (unsigned long long) percpu);
9238 seq_printf(m, "\n");
9242 static const char *cpuacct_stat_desc[] = {
9243 [CPUACCT_STAT_USER] = "user",
9244 [CPUACCT_STAT_SYSTEM] = "system",
9247 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9248 struct cgroup_map_cb *cb)
9250 struct cpuacct *ca = cgroup_ca(cgrp);
9253 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9254 s64 val = percpu_counter_read(&ca->cpustat[i]);
9255 val = cputime64_to_clock_t(val);
9256 cb->fill(cb, cpuacct_stat_desc[i], val);
9261 static struct cftype files[] = {
9264 .read_u64 = cpuusage_read,
9265 .write_u64 = cpuusage_write,
9268 .name = "usage_percpu",
9269 .read_seq_string = cpuacct_percpu_seq_read,
9273 .read_map = cpuacct_stats_show,
9277 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9279 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9283 * charge this task's execution time to its accounting group.
9285 * called with rq->lock held.
9287 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9292 if (unlikely(!cpuacct_subsys.active))
9295 cpu = task_cpu(tsk);
9301 for (; ca; ca = ca->parent) {
9302 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9303 *cpuusage += cputime;
9310 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9311 * in cputime_t units. As a result, cpuacct_update_stats calls
9312 * percpu_counter_add with values large enough to always overflow the
9313 * per cpu batch limit causing bad SMP scalability.
9315 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9316 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9317 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9320 #define CPUACCT_BATCH \
9321 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9323 #define CPUACCT_BATCH 0
9327 * Charge the system/user time to the task's accounting group.
9329 static void cpuacct_update_stats(struct task_struct *tsk,
9330 enum cpuacct_stat_index idx, cputime_t val)
9333 int batch = CPUACCT_BATCH;
9335 if (unlikely(!cpuacct_subsys.active))
9342 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9348 struct cgroup_subsys cpuacct_subsys = {
9350 .create = cpuacct_create,
9351 .destroy = cpuacct_destroy,
9352 .populate = cpuacct_populate,
9353 .subsys_id = cpuacct_subsys_id,
9355 #endif /* CONFIG_CGROUP_CPUACCT */