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 (unlikely(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.)
299 #define MAX_SHARES (1UL << (18 + SCHED_LOAD_RESOLUTION))
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
430 cpumask_var_t online;
433 * The "RT overload" flag: it gets set if a CPU has more than
434 * one runnable RT task.
436 cpumask_var_t rto_mask;
438 struct cpupri cpupri;
442 * By default the system creates a single root-domain with all cpus as
443 * members (mimicking the global state we have today).
445 static struct root_domain def_root_domain;
447 #endif /* CONFIG_SMP */
450 * This is the main, per-CPU runqueue data structure.
452 * Locking rule: those places that want to lock multiple runqueues
453 * (such as the load balancing or the thread migration code), lock
454 * acquire operations must be ordered by ascending &runqueue.
461 * nr_running and cpu_load should be in the same cacheline because
462 * remote CPUs use both these fields when doing load calculation.
464 unsigned long nr_running;
465 #define CPU_LOAD_IDX_MAX 5
466 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
467 unsigned long last_load_update_tick;
470 unsigned char nohz_balance_kick;
472 int skip_clock_update;
474 /* capture load from *all* tasks on this cpu: */
475 struct load_weight load;
476 unsigned long nr_load_updates;
482 #ifdef CONFIG_FAIR_GROUP_SCHED
483 /* list of leaf cfs_rq on this cpu: */
484 struct list_head leaf_cfs_rq_list;
486 #ifdef CONFIG_RT_GROUP_SCHED
487 struct list_head leaf_rt_rq_list;
491 * This is part of a global counter where only the total sum
492 * over all CPUs matters. A task can increase this counter on
493 * one CPU and if it got migrated afterwards it may decrease
494 * it on another CPU. Always updated under the runqueue lock:
496 unsigned long nr_uninterruptible;
498 struct task_struct *curr, *idle, *stop;
499 unsigned long next_balance;
500 struct mm_struct *prev_mm;
508 struct root_domain *rd;
509 struct sched_domain *sd;
511 unsigned long cpu_power;
513 unsigned char idle_at_tick;
514 /* For active balancing */
518 struct cpu_stop_work active_balance_work;
519 /* cpu of this runqueue: */
523 unsigned long avg_load_per_task;
531 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
534 #ifdef CONFIG_PARAVIRT
538 /* calc_load related fields */
539 unsigned long calc_load_update;
540 long calc_load_active;
542 #ifdef CONFIG_SCHED_HRTICK
544 int hrtick_csd_pending;
545 struct call_single_data hrtick_csd;
547 struct hrtimer hrtick_timer;
550 #ifdef CONFIG_SCHEDSTATS
552 struct sched_info rq_sched_info;
553 unsigned long long rq_cpu_time;
554 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
556 /* sys_sched_yield() stats */
557 unsigned int yld_count;
559 /* schedule() stats */
560 unsigned int sched_switch;
561 unsigned int sched_count;
562 unsigned int sched_goidle;
564 /* try_to_wake_up() stats */
565 unsigned int ttwu_count;
566 unsigned int ttwu_local;
570 struct task_struct *wake_list;
574 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
577 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
579 static inline int cpu_of(struct rq *rq)
588 #define rcu_dereference_check_sched_domain(p) \
589 rcu_dereference_check((p), \
590 rcu_read_lock_held() || \
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 rq->avg_load_per_task = rq->load.weight / nr_running;
1572 rq->avg_load_per_task = 0;
1574 return rq->avg_load_per_task;
1577 #ifdef CONFIG_FAIR_GROUP_SCHED
1580 * Compute the cpu's hierarchical load factor for each task group.
1581 * This needs to be done in a top-down fashion because the load of a child
1582 * group is a fraction of its parents load.
1584 static int tg_load_down(struct task_group *tg, void *data)
1587 long cpu = (long)data;
1590 load = cpu_rq(cpu)->load.weight;
1592 load = tg->parent->cfs_rq[cpu]->h_load;
1593 load *= tg->se[cpu]->load.weight;
1594 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1597 tg->cfs_rq[cpu]->h_load = load;
1602 static void update_h_load(long cpu)
1604 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1609 #ifdef CONFIG_PREEMPT
1611 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1614 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1615 * way at the expense of forcing extra atomic operations in all
1616 * invocations. This assures that the double_lock is acquired using the
1617 * same underlying policy as the spinlock_t on this architecture, which
1618 * reduces latency compared to the unfair variant below. However, it
1619 * also adds more overhead and therefore may reduce throughput.
1621 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1622 __releases(this_rq->lock)
1623 __acquires(busiest->lock)
1624 __acquires(this_rq->lock)
1626 raw_spin_unlock(&this_rq->lock);
1627 double_rq_lock(this_rq, busiest);
1634 * Unfair double_lock_balance: Optimizes throughput at the expense of
1635 * latency by eliminating extra atomic operations when the locks are
1636 * already in proper order on entry. This favors lower cpu-ids and will
1637 * grant the double lock to lower cpus over higher ids under contention,
1638 * regardless of entry order into the function.
1640 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1641 __releases(this_rq->lock)
1642 __acquires(busiest->lock)
1643 __acquires(this_rq->lock)
1647 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1648 if (busiest < this_rq) {
1649 raw_spin_unlock(&this_rq->lock);
1650 raw_spin_lock(&busiest->lock);
1651 raw_spin_lock_nested(&this_rq->lock,
1652 SINGLE_DEPTH_NESTING);
1655 raw_spin_lock_nested(&busiest->lock,
1656 SINGLE_DEPTH_NESTING);
1661 #endif /* CONFIG_PREEMPT */
1664 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1666 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1668 if (unlikely(!irqs_disabled())) {
1669 /* printk() doesn't work good under rq->lock */
1670 raw_spin_unlock(&this_rq->lock);
1674 return _double_lock_balance(this_rq, busiest);
1677 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1678 __releases(busiest->lock)
1680 raw_spin_unlock(&busiest->lock);
1681 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1685 * double_rq_lock - safely lock two runqueues
1687 * Note this does not disable interrupts like task_rq_lock,
1688 * you need to do so manually before calling.
1690 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1691 __acquires(rq1->lock)
1692 __acquires(rq2->lock)
1694 BUG_ON(!irqs_disabled());
1696 raw_spin_lock(&rq1->lock);
1697 __acquire(rq2->lock); /* Fake it out ;) */
1700 raw_spin_lock(&rq1->lock);
1701 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1703 raw_spin_lock(&rq2->lock);
1704 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1710 * double_rq_unlock - safely unlock two runqueues
1712 * Note this does not restore interrupts like task_rq_unlock,
1713 * you need to do so manually after calling.
1715 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1716 __releases(rq1->lock)
1717 __releases(rq2->lock)
1719 raw_spin_unlock(&rq1->lock);
1721 raw_spin_unlock(&rq2->lock);
1723 __release(rq2->lock);
1726 #else /* CONFIG_SMP */
1729 * double_rq_lock - safely lock two runqueues
1731 * Note this does not disable interrupts like task_rq_lock,
1732 * you need to do so manually before calling.
1734 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1735 __acquires(rq1->lock)
1736 __acquires(rq2->lock)
1738 BUG_ON(!irqs_disabled());
1740 raw_spin_lock(&rq1->lock);
1741 __acquire(rq2->lock); /* Fake it out ;) */
1745 * double_rq_unlock - safely unlock two runqueues
1747 * Note this does not restore interrupts like task_rq_unlock,
1748 * you need to do so manually after calling.
1750 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1751 __releases(rq1->lock)
1752 __releases(rq2->lock)
1755 raw_spin_unlock(&rq1->lock);
1756 __release(rq2->lock);
1761 static void calc_load_account_idle(struct rq *this_rq);
1762 static void update_sysctl(void);
1763 static int get_update_sysctl_factor(void);
1764 static void update_cpu_load(struct rq *this_rq);
1766 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1768 set_task_rq(p, cpu);
1771 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1772 * successfuly executed on another CPU. We must ensure that updates of
1773 * per-task data have been completed by this moment.
1776 task_thread_info(p)->cpu = cpu;
1780 static const struct sched_class rt_sched_class;
1782 #define sched_class_highest (&stop_sched_class)
1783 #define for_each_class(class) \
1784 for (class = sched_class_highest; class; class = class->next)
1786 #include "sched_stats.h"
1788 static void inc_nr_running(struct rq *rq)
1793 static void dec_nr_running(struct rq *rq)
1798 static void set_load_weight(struct task_struct *p)
1800 int prio = p->static_prio - MAX_RT_PRIO;
1801 struct load_weight *load = &p->se.load;
1804 * SCHED_IDLE tasks get minimal weight:
1806 if (p->policy == SCHED_IDLE) {
1807 load->weight = scale_load(WEIGHT_IDLEPRIO);
1808 load->inv_weight = WMULT_IDLEPRIO;
1812 load->weight = scale_load(prio_to_weight[prio]);
1813 load->inv_weight = prio_to_wmult[prio];
1816 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1818 update_rq_clock(rq);
1819 sched_info_queued(p);
1820 p->sched_class->enqueue_task(rq, p, flags);
1823 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1825 update_rq_clock(rq);
1826 sched_info_dequeued(p);
1827 p->sched_class->dequeue_task(rq, p, flags);
1831 * activate_task - move a task to the runqueue.
1833 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1835 if (task_contributes_to_load(p))
1836 rq->nr_uninterruptible--;
1838 enqueue_task(rq, p, flags);
1843 * deactivate_task - remove a task from the runqueue.
1845 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1847 if (task_contributes_to_load(p))
1848 rq->nr_uninterruptible++;
1850 dequeue_task(rq, p, flags);
1854 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1857 * There are no locks covering percpu hardirq/softirq time.
1858 * They are only modified in account_system_vtime, on corresponding CPU
1859 * with interrupts disabled. So, writes are safe.
1860 * They are read and saved off onto struct rq in update_rq_clock().
1861 * This may result in other CPU reading this CPU's irq time and can
1862 * race with irq/account_system_vtime on this CPU. We would either get old
1863 * or new value with a side effect of accounting a slice of irq time to wrong
1864 * task when irq is in progress while we read rq->clock. That is a worthy
1865 * compromise in place of having locks on each irq in account_system_time.
1867 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1868 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1870 static DEFINE_PER_CPU(u64, irq_start_time);
1871 static int sched_clock_irqtime;
1873 void enable_sched_clock_irqtime(void)
1875 sched_clock_irqtime = 1;
1878 void disable_sched_clock_irqtime(void)
1880 sched_clock_irqtime = 0;
1883 #ifndef CONFIG_64BIT
1884 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1886 static inline void irq_time_write_begin(void)
1888 __this_cpu_inc(irq_time_seq.sequence);
1892 static inline void irq_time_write_end(void)
1895 __this_cpu_inc(irq_time_seq.sequence);
1898 static inline u64 irq_time_read(int cpu)
1904 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1905 irq_time = per_cpu(cpu_softirq_time, cpu) +
1906 per_cpu(cpu_hardirq_time, cpu);
1907 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1911 #else /* CONFIG_64BIT */
1912 static inline void irq_time_write_begin(void)
1916 static inline void irq_time_write_end(void)
1920 static inline u64 irq_time_read(int cpu)
1922 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1924 #endif /* CONFIG_64BIT */
1927 * Called before incrementing preempt_count on {soft,}irq_enter
1928 * and before decrementing preempt_count on {soft,}irq_exit.
1930 void account_system_vtime(struct task_struct *curr)
1932 unsigned long flags;
1936 if (!sched_clock_irqtime)
1939 local_irq_save(flags);
1941 cpu = smp_processor_id();
1942 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1943 __this_cpu_add(irq_start_time, delta);
1945 irq_time_write_begin();
1947 * We do not account for softirq time from ksoftirqd here.
1948 * We want to continue accounting softirq time to ksoftirqd thread
1949 * in that case, so as not to confuse scheduler with a special task
1950 * that do not consume any time, but still wants to run.
1952 if (hardirq_count())
1953 __this_cpu_add(cpu_hardirq_time, delta);
1954 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
1955 __this_cpu_add(cpu_softirq_time, delta);
1957 irq_time_write_end();
1958 local_irq_restore(flags);
1960 EXPORT_SYMBOL_GPL(account_system_vtime);
1962 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
1964 #ifdef CONFIG_PARAVIRT
1965 static inline u64 steal_ticks(u64 steal)
1967 if (unlikely(steal > NSEC_PER_SEC))
1968 return div_u64(steal, TICK_NSEC);
1970 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
1974 static void update_rq_clock_task(struct rq *rq, s64 delta)
1978 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
1981 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1982 * this case when a previous update_rq_clock() happened inside a
1983 * {soft,}irq region.
1985 * When this happens, we stop ->clock_task and only update the
1986 * prev_irq_time stamp to account for the part that fit, so that a next
1987 * update will consume the rest. This ensures ->clock_task is
1990 * It does however cause some slight miss-attribution of {soft,}irq
1991 * time, a more accurate solution would be to update the irq_time using
1992 * the current rq->clock timestamp, except that would require using
1995 if (irq_delta > delta)
1998 rq->prev_irq_time += irq_delta;
2000 rq->clock_task += delta;
2002 if (irq_delta && sched_feat(NONIRQ_POWER))
2003 sched_rt_avg_update(rq, irq_delta);
2006 static int irqtime_account_hi_update(void)
2008 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2009 unsigned long flags;
2013 local_irq_save(flags);
2014 latest_ns = this_cpu_read(cpu_hardirq_time);
2015 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
2017 local_irq_restore(flags);
2021 static int irqtime_account_si_update(void)
2023 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2024 unsigned long flags;
2028 local_irq_save(flags);
2029 latest_ns = this_cpu_read(cpu_softirq_time);
2030 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2032 local_irq_restore(flags);
2036 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2038 #define sched_clock_irqtime (0)
2040 static void update_rq_clock_task(struct rq *rq, s64 delta)
2042 rq->clock_task += delta;
2045 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2047 #include "sched_idletask.c"
2048 #include "sched_fair.c"
2049 #include "sched_rt.c"
2050 #include "sched_autogroup.c"
2051 #include "sched_stoptask.c"
2052 #ifdef CONFIG_SCHED_DEBUG
2053 # include "sched_debug.c"
2056 void sched_set_stop_task(int cpu, struct task_struct *stop)
2058 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2059 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2063 * Make it appear like a SCHED_FIFO task, its something
2064 * userspace knows about and won't get confused about.
2066 * Also, it will make PI more or less work without too
2067 * much confusion -- but then, stop work should not
2068 * rely on PI working anyway.
2070 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2072 stop->sched_class = &stop_sched_class;
2075 cpu_rq(cpu)->stop = stop;
2079 * Reset it back to a normal scheduling class so that
2080 * it can die in pieces.
2082 old_stop->sched_class = &rt_sched_class;
2087 * __normal_prio - return the priority that is based on the static prio
2089 static inline int __normal_prio(struct task_struct *p)
2091 return p->static_prio;
2095 * Calculate the expected normal priority: i.e. priority
2096 * without taking RT-inheritance into account. Might be
2097 * boosted by interactivity modifiers. Changes upon fork,
2098 * setprio syscalls, and whenever the interactivity
2099 * estimator recalculates.
2101 static inline int normal_prio(struct task_struct *p)
2105 if (task_has_rt_policy(p))
2106 prio = MAX_RT_PRIO-1 - p->rt_priority;
2108 prio = __normal_prio(p);
2113 * Calculate the current priority, i.e. the priority
2114 * taken into account by the scheduler. This value might
2115 * be boosted by RT tasks, or might be boosted by
2116 * interactivity modifiers. Will be RT if the task got
2117 * RT-boosted. If not then it returns p->normal_prio.
2119 static int effective_prio(struct task_struct *p)
2121 p->normal_prio = normal_prio(p);
2123 * If we are RT tasks or we were boosted to RT priority,
2124 * keep the priority unchanged. Otherwise, update priority
2125 * to the normal priority:
2127 if (!rt_prio(p->prio))
2128 return p->normal_prio;
2133 * task_curr - is this task currently executing on a CPU?
2134 * @p: the task in question.
2136 inline int task_curr(const struct task_struct *p)
2138 return cpu_curr(task_cpu(p)) == p;
2141 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2142 const struct sched_class *prev_class,
2145 if (prev_class != p->sched_class) {
2146 if (prev_class->switched_from)
2147 prev_class->switched_from(rq, p);
2148 p->sched_class->switched_to(rq, p);
2149 } else if (oldprio != p->prio)
2150 p->sched_class->prio_changed(rq, p, oldprio);
2153 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2155 const struct sched_class *class;
2157 if (p->sched_class == rq->curr->sched_class) {
2158 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2160 for_each_class(class) {
2161 if (class == rq->curr->sched_class)
2163 if (class == p->sched_class) {
2164 resched_task(rq->curr);
2171 * A queue event has occurred, and we're going to schedule. In
2172 * this case, we can save a useless back to back clock update.
2174 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2175 rq->skip_clock_update = 1;
2180 * Is this task likely cache-hot:
2183 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2187 if (p->sched_class != &fair_sched_class)
2190 if (unlikely(p->policy == SCHED_IDLE))
2194 * Buddy candidates are cache hot:
2196 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2197 (&p->se == cfs_rq_of(&p->se)->next ||
2198 &p->se == cfs_rq_of(&p->se)->last))
2201 if (sysctl_sched_migration_cost == -1)
2203 if (sysctl_sched_migration_cost == 0)
2206 delta = now - p->se.exec_start;
2208 return delta < (s64)sysctl_sched_migration_cost;
2211 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2213 #ifdef CONFIG_SCHED_DEBUG
2215 * We should never call set_task_cpu() on a blocked task,
2216 * ttwu() will sort out the placement.
2218 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2219 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2221 #ifdef CONFIG_LOCKDEP
2223 * The caller should hold either p->pi_lock or rq->lock, when changing
2224 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2226 * sched_move_task() holds both and thus holding either pins the cgroup,
2227 * see set_task_rq().
2229 * Furthermore, all task_rq users should acquire both locks, see
2232 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2233 lockdep_is_held(&task_rq(p)->lock)));
2237 trace_sched_migrate_task(p, new_cpu);
2239 if (task_cpu(p) != new_cpu) {
2240 p->se.nr_migrations++;
2241 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2244 __set_task_cpu(p, new_cpu);
2247 struct migration_arg {
2248 struct task_struct *task;
2252 static int migration_cpu_stop(void *data);
2255 * wait_task_inactive - wait for a thread to unschedule.
2257 * If @match_state is nonzero, it's the @p->state value just checked and
2258 * not expected to change. If it changes, i.e. @p might have woken up,
2259 * then return zero. When we succeed in waiting for @p to be off its CPU,
2260 * we return a positive number (its total switch count). If a second call
2261 * a short while later returns the same number, the caller can be sure that
2262 * @p has remained unscheduled the whole time.
2264 * The caller must ensure that the task *will* unschedule sometime soon,
2265 * else this function might spin for a *long* time. This function can't
2266 * be called with interrupts off, or it may introduce deadlock with
2267 * smp_call_function() if an IPI is sent by the same process we are
2268 * waiting to become inactive.
2270 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2272 unsigned long flags;
2279 * We do the initial early heuristics without holding
2280 * any task-queue locks at all. We'll only try to get
2281 * the runqueue lock when things look like they will
2287 * If the task is actively running on another CPU
2288 * still, just relax and busy-wait without holding
2291 * NOTE! Since we don't hold any locks, it's not
2292 * even sure that "rq" stays as the right runqueue!
2293 * But we don't care, since "task_running()" will
2294 * return false if the runqueue has changed and p
2295 * is actually now running somewhere else!
2297 while (task_running(rq, p)) {
2298 if (match_state && unlikely(p->state != match_state))
2304 * Ok, time to look more closely! We need the rq
2305 * lock now, to be *sure*. If we're wrong, we'll
2306 * just go back and repeat.
2308 rq = task_rq_lock(p, &flags);
2309 trace_sched_wait_task(p);
2310 running = task_running(rq, p);
2313 if (!match_state || p->state == match_state)
2314 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2315 task_rq_unlock(rq, p, &flags);
2318 * If it changed from the expected state, bail out now.
2320 if (unlikely(!ncsw))
2324 * Was it really running after all now that we
2325 * checked with the proper locks actually held?
2327 * Oops. Go back and try again..
2329 if (unlikely(running)) {
2335 * It's not enough that it's not actively running,
2336 * it must be off the runqueue _entirely_, and not
2339 * So if it was still runnable (but just not actively
2340 * running right now), it's preempted, and we should
2341 * yield - it could be a while.
2343 if (unlikely(on_rq)) {
2344 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2346 set_current_state(TASK_UNINTERRUPTIBLE);
2347 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2352 * Ahh, all good. It wasn't running, and it wasn't
2353 * runnable, which means that it will never become
2354 * running in the future either. We're all done!
2363 * kick_process - kick a running thread to enter/exit the kernel
2364 * @p: the to-be-kicked thread
2366 * Cause a process which is running on another CPU to enter
2367 * kernel-mode, without any delay. (to get signals handled.)
2369 * NOTE: this function doesn't have to take the runqueue lock,
2370 * because all it wants to ensure is that the remote task enters
2371 * the kernel. If the IPI races and the task has been migrated
2372 * to another CPU then no harm is done and the purpose has been
2375 void kick_process(struct task_struct *p)
2381 if ((cpu != smp_processor_id()) && task_curr(p))
2382 smp_send_reschedule(cpu);
2385 EXPORT_SYMBOL_GPL(kick_process);
2386 #endif /* CONFIG_SMP */
2390 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2392 static int select_fallback_rq(int cpu, struct task_struct *p)
2395 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2397 /* Look for allowed, online CPU in same node. */
2398 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2399 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2402 /* Any allowed, online CPU? */
2403 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2404 if (dest_cpu < nr_cpu_ids)
2407 /* No more Mr. Nice Guy. */
2408 dest_cpu = cpuset_cpus_allowed_fallback(p);
2410 * Don't tell them about moving exiting tasks or
2411 * kernel threads (both mm NULL), since they never
2414 if (p->mm && printk_ratelimit()) {
2415 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2416 task_pid_nr(p), p->comm, cpu);
2423 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2426 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2428 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2431 * In order not to call set_task_cpu() on a blocking task we need
2432 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2435 * Since this is common to all placement strategies, this lives here.
2437 * [ this allows ->select_task() to simply return task_cpu(p) and
2438 * not worry about this generic constraint ]
2440 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2442 cpu = select_fallback_rq(task_cpu(p), p);
2447 static void update_avg(u64 *avg, u64 sample)
2449 s64 diff = sample - *avg;
2455 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2457 #ifdef CONFIG_SCHEDSTATS
2458 struct rq *rq = this_rq();
2461 int this_cpu = smp_processor_id();
2463 if (cpu == this_cpu) {
2464 schedstat_inc(rq, ttwu_local);
2465 schedstat_inc(p, se.statistics.nr_wakeups_local);
2467 struct sched_domain *sd;
2469 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2471 for_each_domain(this_cpu, sd) {
2472 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2473 schedstat_inc(sd, ttwu_wake_remote);
2480 if (wake_flags & WF_MIGRATED)
2481 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2483 #endif /* CONFIG_SMP */
2485 schedstat_inc(rq, ttwu_count);
2486 schedstat_inc(p, se.statistics.nr_wakeups);
2488 if (wake_flags & WF_SYNC)
2489 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2491 #endif /* CONFIG_SCHEDSTATS */
2494 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2496 activate_task(rq, p, en_flags);
2499 /* if a worker is waking up, notify workqueue */
2500 if (p->flags & PF_WQ_WORKER)
2501 wq_worker_waking_up(p, cpu_of(rq));
2505 * Mark the task runnable and perform wakeup-preemption.
2508 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2510 trace_sched_wakeup(p, true);
2511 check_preempt_curr(rq, p, wake_flags);
2513 p->state = TASK_RUNNING;
2515 if (p->sched_class->task_woken)
2516 p->sched_class->task_woken(rq, p);
2518 if (unlikely(rq->idle_stamp)) {
2519 u64 delta = rq->clock - rq->idle_stamp;
2520 u64 max = 2*sysctl_sched_migration_cost;
2525 update_avg(&rq->avg_idle, delta);
2532 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2535 if (p->sched_contributes_to_load)
2536 rq->nr_uninterruptible--;
2539 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2540 ttwu_do_wakeup(rq, p, wake_flags);
2544 * Called in case the task @p isn't fully descheduled from its runqueue,
2545 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2546 * since all we need to do is flip p->state to TASK_RUNNING, since
2547 * the task is still ->on_rq.
2549 static int ttwu_remote(struct task_struct *p, int wake_flags)
2554 rq = __task_rq_lock(p);
2556 ttwu_do_wakeup(rq, p, wake_flags);
2559 __task_rq_unlock(rq);
2565 static void sched_ttwu_pending(void)
2567 struct rq *rq = this_rq();
2568 struct task_struct *list = xchg(&rq->wake_list, NULL);
2573 raw_spin_lock(&rq->lock);
2576 struct task_struct *p = list;
2577 list = list->wake_entry;
2578 ttwu_do_activate(rq, p, 0);
2581 raw_spin_unlock(&rq->lock);
2584 void scheduler_ipi(void)
2586 sched_ttwu_pending();
2589 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2591 struct rq *rq = cpu_rq(cpu);
2592 struct task_struct *next = rq->wake_list;
2595 struct task_struct *old = next;
2597 p->wake_entry = next;
2598 next = cmpxchg(&rq->wake_list, old, p);
2604 smp_send_reschedule(cpu);
2607 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2608 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2613 rq = __task_rq_lock(p);
2615 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2616 ttwu_do_wakeup(rq, p, wake_flags);
2619 __task_rq_unlock(rq);
2624 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2625 #endif /* CONFIG_SMP */
2627 static void ttwu_queue(struct task_struct *p, int cpu)
2629 struct rq *rq = cpu_rq(cpu);
2631 #if defined(CONFIG_SMP)
2632 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2633 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2634 ttwu_queue_remote(p, cpu);
2639 raw_spin_lock(&rq->lock);
2640 ttwu_do_activate(rq, p, 0);
2641 raw_spin_unlock(&rq->lock);
2645 * try_to_wake_up - wake up a thread
2646 * @p: the thread to be awakened
2647 * @state: the mask of task states that can be woken
2648 * @wake_flags: wake modifier flags (WF_*)
2650 * Put it on the run-queue if it's not already there. The "current"
2651 * thread is always on the run-queue (except when the actual
2652 * re-schedule is in progress), and as such you're allowed to do
2653 * the simpler "current->state = TASK_RUNNING" to mark yourself
2654 * runnable without the overhead of this.
2656 * Returns %true if @p was woken up, %false if it was already running
2657 * or @state didn't match @p's state.
2660 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2662 unsigned long flags;
2663 int cpu, success = 0;
2666 raw_spin_lock_irqsave(&p->pi_lock, flags);
2667 if (!(p->state & state))
2670 success = 1; /* we're going to change ->state */
2673 if (p->on_rq && ttwu_remote(p, wake_flags))
2678 * If the owning (remote) cpu is still in the middle of schedule() with
2679 * this task as prev, wait until its done referencing the task.
2682 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2684 * In case the architecture enables interrupts in
2685 * context_switch(), we cannot busy wait, since that
2686 * would lead to deadlocks when an interrupt hits and
2687 * tries to wake up @prev. So bail and do a complete
2690 if (ttwu_activate_remote(p, wake_flags))
2697 * Pairs with the smp_wmb() in finish_lock_switch().
2701 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2702 p->state = TASK_WAKING;
2704 if (p->sched_class->task_waking)
2705 p->sched_class->task_waking(p);
2707 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2708 if (task_cpu(p) != cpu) {
2709 wake_flags |= WF_MIGRATED;
2710 set_task_cpu(p, cpu);
2712 #endif /* CONFIG_SMP */
2716 ttwu_stat(p, cpu, wake_flags);
2718 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2724 * try_to_wake_up_local - try to wake up a local task with rq lock held
2725 * @p: the thread to be awakened
2727 * Put @p on the run-queue if it's not already there. The caller must
2728 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2731 static void try_to_wake_up_local(struct task_struct *p)
2733 struct rq *rq = task_rq(p);
2735 BUG_ON(rq != this_rq());
2736 BUG_ON(p == current);
2737 lockdep_assert_held(&rq->lock);
2739 if (!raw_spin_trylock(&p->pi_lock)) {
2740 raw_spin_unlock(&rq->lock);
2741 raw_spin_lock(&p->pi_lock);
2742 raw_spin_lock(&rq->lock);
2745 if (!(p->state & TASK_NORMAL))
2749 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2751 ttwu_do_wakeup(rq, p, 0);
2752 ttwu_stat(p, smp_processor_id(), 0);
2754 raw_spin_unlock(&p->pi_lock);
2758 * wake_up_process - Wake up a specific process
2759 * @p: The process to be woken up.
2761 * Attempt to wake up the nominated process and move it to the set of runnable
2762 * processes. Returns 1 if the process was woken up, 0 if it was already
2765 * It may be assumed that this function implies a write memory barrier before
2766 * changing the task state if and only if any tasks are woken up.
2768 int wake_up_process(struct task_struct *p)
2770 return try_to_wake_up(p, TASK_ALL, 0);
2772 EXPORT_SYMBOL(wake_up_process);
2774 int wake_up_state(struct task_struct *p, unsigned int state)
2776 return try_to_wake_up(p, state, 0);
2780 * Perform scheduler related setup for a newly forked process p.
2781 * p is forked by current.
2783 * __sched_fork() is basic setup used by init_idle() too:
2785 static void __sched_fork(struct task_struct *p)
2790 p->se.exec_start = 0;
2791 p->se.sum_exec_runtime = 0;
2792 p->se.prev_sum_exec_runtime = 0;
2793 p->se.nr_migrations = 0;
2795 INIT_LIST_HEAD(&p->se.group_node);
2797 #ifdef CONFIG_SCHEDSTATS
2798 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2801 INIT_LIST_HEAD(&p->rt.run_list);
2803 #ifdef CONFIG_PREEMPT_NOTIFIERS
2804 INIT_HLIST_HEAD(&p->preempt_notifiers);
2809 * fork()/clone()-time setup:
2811 void sched_fork(struct task_struct *p)
2813 unsigned long flags;
2814 int cpu = get_cpu();
2818 * We mark the process as running here. This guarantees that
2819 * nobody will actually run it, and a signal or other external
2820 * event cannot wake it up and insert it on the runqueue either.
2822 p->state = TASK_RUNNING;
2825 * Revert to default priority/policy on fork if requested.
2827 if (unlikely(p->sched_reset_on_fork)) {
2828 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2829 p->policy = SCHED_NORMAL;
2830 p->normal_prio = p->static_prio;
2833 if (PRIO_TO_NICE(p->static_prio) < 0) {
2834 p->static_prio = NICE_TO_PRIO(0);
2835 p->normal_prio = p->static_prio;
2840 * We don't need the reset flag anymore after the fork. It has
2841 * fulfilled its duty:
2843 p->sched_reset_on_fork = 0;
2847 * Make sure we do not leak PI boosting priority to the child.
2849 p->prio = current->normal_prio;
2851 if (!rt_prio(p->prio))
2852 p->sched_class = &fair_sched_class;
2854 if (p->sched_class->task_fork)
2855 p->sched_class->task_fork(p);
2858 * The child is not yet in the pid-hash so no cgroup attach races,
2859 * and the cgroup is pinned to this child due to cgroup_fork()
2860 * is ran before sched_fork().
2862 * Silence PROVE_RCU.
2864 raw_spin_lock_irqsave(&p->pi_lock, flags);
2865 set_task_cpu(p, cpu);
2866 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2868 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2869 if (likely(sched_info_on()))
2870 memset(&p->sched_info, 0, sizeof(p->sched_info));
2872 #if defined(CONFIG_SMP)
2875 #ifdef CONFIG_PREEMPT
2876 /* Want to start with kernel preemption disabled. */
2877 task_thread_info(p)->preempt_count = 1;
2880 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2887 * wake_up_new_task - wake up a newly created task for the first time.
2889 * This function will do some initial scheduler statistics housekeeping
2890 * that must be done for every newly created context, then puts the task
2891 * on the runqueue and wakes it.
2893 void wake_up_new_task(struct task_struct *p)
2895 unsigned long flags;
2898 raw_spin_lock_irqsave(&p->pi_lock, flags);
2901 * Fork balancing, do it here and not earlier because:
2902 * - cpus_allowed can change in the fork path
2903 * - any previously selected cpu might disappear through hotplug
2905 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
2908 rq = __task_rq_lock(p);
2909 activate_task(rq, p, 0);
2911 trace_sched_wakeup_new(p, true);
2912 check_preempt_curr(rq, p, WF_FORK);
2914 if (p->sched_class->task_woken)
2915 p->sched_class->task_woken(rq, p);
2917 task_rq_unlock(rq, p, &flags);
2920 #ifdef CONFIG_PREEMPT_NOTIFIERS
2923 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2924 * @notifier: notifier struct to register
2926 void preempt_notifier_register(struct preempt_notifier *notifier)
2928 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2930 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2933 * preempt_notifier_unregister - no longer interested in preemption notifications
2934 * @notifier: notifier struct to unregister
2936 * This is safe to call from within a preemption notifier.
2938 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2940 hlist_del(¬ifier->link);
2942 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2944 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2946 struct preempt_notifier *notifier;
2947 struct hlist_node *node;
2949 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2950 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2954 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2955 struct task_struct *next)
2957 struct preempt_notifier *notifier;
2958 struct hlist_node *node;
2960 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2961 notifier->ops->sched_out(notifier, next);
2964 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2966 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2971 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2972 struct task_struct *next)
2976 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2979 * prepare_task_switch - prepare to switch tasks
2980 * @rq: the runqueue preparing to switch
2981 * @prev: the current task that is being switched out
2982 * @next: the task we are going to switch to.
2984 * This is called with the rq lock held and interrupts off. It must
2985 * be paired with a subsequent finish_task_switch after the context
2988 * prepare_task_switch sets up locking and calls architecture specific
2992 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2993 struct task_struct *next)
2995 sched_info_switch(prev, next);
2996 perf_event_task_sched_out(prev, next);
2997 fire_sched_out_preempt_notifiers(prev, next);
2998 prepare_lock_switch(rq, next);
2999 prepare_arch_switch(next);
3000 trace_sched_switch(prev, next);
3004 * finish_task_switch - clean up after a task-switch
3005 * @rq: runqueue associated with task-switch
3006 * @prev: the thread we just switched away from.
3008 * finish_task_switch must be called after the context switch, paired
3009 * with a prepare_task_switch call before the context switch.
3010 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3011 * and do any other architecture-specific cleanup actions.
3013 * Note that we may have delayed dropping an mm in context_switch(). If
3014 * so, we finish that here outside of the runqueue lock. (Doing it
3015 * with the lock held can cause deadlocks; see schedule() for
3018 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3019 __releases(rq->lock)
3021 struct mm_struct *mm = rq->prev_mm;
3027 * A task struct has one reference for the use as "current".
3028 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3029 * schedule one last time. The schedule call will never return, and
3030 * the scheduled task must drop that reference.
3031 * The test for TASK_DEAD must occur while the runqueue locks are
3032 * still held, otherwise prev could be scheduled on another cpu, die
3033 * there before we look at prev->state, and then the reference would
3035 * Manfred Spraul <manfred@colorfullife.com>
3037 prev_state = prev->state;
3038 finish_arch_switch(prev);
3039 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3040 local_irq_disable();
3041 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3042 perf_event_task_sched_in(current);
3043 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3045 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3046 finish_lock_switch(rq, prev);
3048 fire_sched_in_preempt_notifiers(current);
3051 if (unlikely(prev_state == TASK_DEAD)) {
3053 * Remove function-return probe instances associated with this
3054 * task and put them back on the free list.
3056 kprobe_flush_task(prev);
3057 put_task_struct(prev);
3063 /* assumes rq->lock is held */
3064 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3066 if (prev->sched_class->pre_schedule)
3067 prev->sched_class->pre_schedule(rq, prev);
3070 /* rq->lock is NOT held, but preemption is disabled */
3071 static inline void post_schedule(struct rq *rq)
3073 if (rq->post_schedule) {
3074 unsigned long flags;
3076 raw_spin_lock_irqsave(&rq->lock, flags);
3077 if (rq->curr->sched_class->post_schedule)
3078 rq->curr->sched_class->post_schedule(rq);
3079 raw_spin_unlock_irqrestore(&rq->lock, flags);
3081 rq->post_schedule = 0;
3087 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3091 static inline void post_schedule(struct rq *rq)
3098 * schedule_tail - first thing a freshly forked thread must call.
3099 * @prev: the thread we just switched away from.
3101 asmlinkage void schedule_tail(struct task_struct *prev)
3102 __releases(rq->lock)
3104 struct rq *rq = this_rq();
3106 finish_task_switch(rq, prev);
3109 * FIXME: do we need to worry about rq being invalidated by the
3114 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3115 /* In this case, finish_task_switch does not reenable preemption */
3118 if (current->set_child_tid)
3119 put_user(task_pid_vnr(current), current->set_child_tid);
3123 * context_switch - switch to the new MM and the new
3124 * thread's register state.
3127 context_switch(struct rq *rq, struct task_struct *prev,
3128 struct task_struct *next)
3130 struct mm_struct *mm, *oldmm;
3132 prepare_task_switch(rq, prev, next);
3135 oldmm = prev->active_mm;
3137 * For paravirt, this is coupled with an exit in switch_to to
3138 * combine the page table reload and the switch backend into
3141 arch_start_context_switch(prev);
3144 next->active_mm = oldmm;
3145 atomic_inc(&oldmm->mm_count);
3146 enter_lazy_tlb(oldmm, next);
3148 switch_mm(oldmm, mm, next);
3151 prev->active_mm = NULL;
3152 rq->prev_mm = oldmm;
3155 * Since the runqueue lock will be released by the next
3156 * task (which is an invalid locking op but in the case
3157 * of the scheduler it's an obvious special-case), so we
3158 * do an early lockdep release here:
3160 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3161 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3164 /* Here we just switch the register state and the stack. */
3165 switch_to(prev, next, prev);
3169 * this_rq must be evaluated again because prev may have moved
3170 * CPUs since it called schedule(), thus the 'rq' on its stack
3171 * frame will be invalid.
3173 finish_task_switch(this_rq(), prev);
3177 * nr_running, nr_uninterruptible and nr_context_switches:
3179 * externally visible scheduler statistics: current number of runnable
3180 * threads, current number of uninterruptible-sleeping threads, total
3181 * number of context switches performed since bootup.
3183 unsigned long nr_running(void)
3185 unsigned long i, sum = 0;
3187 for_each_online_cpu(i)
3188 sum += cpu_rq(i)->nr_running;
3193 unsigned long nr_uninterruptible(void)
3195 unsigned long i, sum = 0;
3197 for_each_possible_cpu(i)
3198 sum += cpu_rq(i)->nr_uninterruptible;
3201 * Since we read the counters lockless, it might be slightly
3202 * inaccurate. Do not allow it to go below zero though:
3204 if (unlikely((long)sum < 0))
3210 unsigned long long nr_context_switches(void)
3213 unsigned long long sum = 0;
3215 for_each_possible_cpu(i)
3216 sum += cpu_rq(i)->nr_switches;
3221 unsigned long nr_iowait(void)
3223 unsigned long i, sum = 0;
3225 for_each_possible_cpu(i)
3226 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3231 unsigned long nr_iowait_cpu(int cpu)
3233 struct rq *this = cpu_rq(cpu);
3234 return atomic_read(&this->nr_iowait);
3237 unsigned long this_cpu_load(void)
3239 struct rq *this = this_rq();
3240 return this->cpu_load[0];
3244 /* Variables and functions for calc_load */
3245 static atomic_long_t calc_load_tasks;
3246 static unsigned long calc_load_update;
3247 unsigned long avenrun[3];
3248 EXPORT_SYMBOL(avenrun);
3250 static long calc_load_fold_active(struct rq *this_rq)
3252 long nr_active, delta = 0;
3254 nr_active = this_rq->nr_running;
3255 nr_active += (long) this_rq->nr_uninterruptible;
3257 if (nr_active != this_rq->calc_load_active) {
3258 delta = nr_active - this_rq->calc_load_active;
3259 this_rq->calc_load_active = nr_active;
3265 static unsigned long
3266 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3269 load += active * (FIXED_1 - exp);
3270 load += 1UL << (FSHIFT - 1);
3271 return load >> FSHIFT;
3276 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3278 * When making the ILB scale, we should try to pull this in as well.
3280 static atomic_long_t calc_load_tasks_idle;
3282 static void calc_load_account_idle(struct rq *this_rq)
3286 delta = calc_load_fold_active(this_rq);
3288 atomic_long_add(delta, &calc_load_tasks_idle);
3291 static long calc_load_fold_idle(void)
3296 * Its got a race, we don't care...
3298 if (atomic_long_read(&calc_load_tasks_idle))
3299 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3305 * fixed_power_int - compute: x^n, in O(log n) time
3307 * @x: base of the power
3308 * @frac_bits: fractional bits of @x
3309 * @n: power to raise @x to.
3311 * By exploiting the relation between the definition of the natural power
3312 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3313 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3314 * (where: n_i \elem {0, 1}, the binary vector representing n),
3315 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3316 * of course trivially computable in O(log_2 n), the length of our binary
3319 static unsigned long
3320 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3322 unsigned long result = 1UL << frac_bits;
3327 result += 1UL << (frac_bits - 1);
3328 result >>= frac_bits;
3334 x += 1UL << (frac_bits - 1);
3342 * a1 = a0 * e + a * (1 - e)
3344 * a2 = a1 * e + a * (1 - e)
3345 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3346 * = a0 * e^2 + a * (1 - e) * (1 + e)
3348 * a3 = a2 * e + a * (1 - e)
3349 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3350 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3354 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3355 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3356 * = a0 * e^n + a * (1 - e^n)
3358 * [1] application of the geometric series:
3361 * S_n := \Sum x^i = -------------
3364 static unsigned long
3365 calc_load_n(unsigned long load, unsigned long exp,
3366 unsigned long active, unsigned int n)
3369 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3373 * NO_HZ can leave us missing all per-cpu ticks calling
3374 * calc_load_account_active(), but since an idle CPU folds its delta into
3375 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3376 * in the pending idle delta if our idle period crossed a load cycle boundary.
3378 * Once we've updated the global active value, we need to apply the exponential
3379 * weights adjusted to the number of cycles missed.
3381 static void calc_global_nohz(unsigned long ticks)
3383 long delta, active, n;
3385 if (time_before(jiffies, calc_load_update))
3389 * If we crossed a calc_load_update boundary, make sure to fold
3390 * any pending idle changes, the respective CPUs might have
3391 * missed the tick driven calc_load_account_active() update
3394 delta = calc_load_fold_idle();
3396 atomic_long_add(delta, &calc_load_tasks);
3399 * If we were idle for multiple load cycles, apply them.
3401 if (ticks >= LOAD_FREQ) {
3402 n = ticks / LOAD_FREQ;
3404 active = atomic_long_read(&calc_load_tasks);
3405 active = active > 0 ? active * FIXED_1 : 0;
3407 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3408 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3409 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3411 calc_load_update += n * LOAD_FREQ;
3415 * Its possible the remainder of the above division also crosses
3416 * a LOAD_FREQ period, the regular check in calc_global_load()
3417 * which comes after this will take care of that.
3419 * Consider us being 11 ticks before a cycle completion, and us
3420 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3421 * age us 4 cycles, and the test in calc_global_load() will
3422 * pick up the final one.
3426 static void calc_load_account_idle(struct rq *this_rq)
3430 static inline long calc_load_fold_idle(void)
3435 static void calc_global_nohz(unsigned long ticks)
3441 * get_avenrun - get the load average array
3442 * @loads: pointer to dest load array
3443 * @offset: offset to add
3444 * @shift: shift count to shift the result left
3446 * These values are estimates at best, so no need for locking.
3448 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3450 loads[0] = (avenrun[0] + offset) << shift;
3451 loads[1] = (avenrun[1] + offset) << shift;
3452 loads[2] = (avenrun[2] + offset) << shift;
3456 * calc_load - update the avenrun load estimates 10 ticks after the
3457 * CPUs have updated calc_load_tasks.
3459 void calc_global_load(unsigned long ticks)
3463 calc_global_nohz(ticks);
3465 if (time_before(jiffies, calc_load_update + 10))
3468 active = atomic_long_read(&calc_load_tasks);
3469 active = active > 0 ? active * FIXED_1 : 0;
3471 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3472 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3473 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3475 calc_load_update += LOAD_FREQ;
3479 * Called from update_cpu_load() to periodically update this CPU's
3482 static void calc_load_account_active(struct rq *this_rq)
3486 if (time_before(jiffies, this_rq->calc_load_update))
3489 delta = calc_load_fold_active(this_rq);
3490 delta += calc_load_fold_idle();
3492 atomic_long_add(delta, &calc_load_tasks);
3494 this_rq->calc_load_update += LOAD_FREQ;
3498 * The exact cpuload at various idx values, calculated at every tick would be
3499 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3501 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3502 * on nth tick when cpu may be busy, then we have:
3503 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3504 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3506 * decay_load_missed() below does efficient calculation of
3507 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3508 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3510 * The calculation is approximated on a 128 point scale.
3511 * degrade_zero_ticks is the number of ticks after which load at any
3512 * particular idx is approximated to be zero.
3513 * degrade_factor is a precomputed table, a row for each load idx.
3514 * Each column corresponds to degradation factor for a power of two ticks,
3515 * based on 128 point scale.
3517 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3518 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3520 * With this power of 2 load factors, we can degrade the load n times
3521 * by looking at 1 bits in n and doing as many mult/shift instead of
3522 * n mult/shifts needed by the exact degradation.
3524 #define DEGRADE_SHIFT 7
3525 static const unsigned char
3526 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3527 static const unsigned char
3528 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3529 {0, 0, 0, 0, 0, 0, 0, 0},
3530 {64, 32, 8, 0, 0, 0, 0, 0},
3531 {96, 72, 40, 12, 1, 0, 0},
3532 {112, 98, 75, 43, 15, 1, 0},
3533 {120, 112, 98, 76, 45, 16, 2} };
3536 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3537 * would be when CPU is idle and so we just decay the old load without
3538 * adding any new load.
3540 static unsigned long
3541 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3545 if (!missed_updates)
3548 if (missed_updates >= degrade_zero_ticks[idx])
3552 return load >> missed_updates;
3554 while (missed_updates) {
3555 if (missed_updates % 2)
3556 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3558 missed_updates >>= 1;
3565 * Update rq->cpu_load[] statistics. This function is usually called every
3566 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3567 * every tick. We fix it up based on jiffies.
3569 static void update_cpu_load(struct rq *this_rq)
3571 unsigned long this_load = this_rq->load.weight;
3572 unsigned long curr_jiffies = jiffies;
3573 unsigned long pending_updates;
3576 this_rq->nr_load_updates++;
3578 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3579 if (curr_jiffies == this_rq->last_load_update_tick)
3582 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3583 this_rq->last_load_update_tick = curr_jiffies;
3585 /* Update our load: */
3586 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3587 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3588 unsigned long old_load, new_load;
3590 /* scale is effectively 1 << i now, and >> i divides by scale */
3592 old_load = this_rq->cpu_load[i];
3593 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3594 new_load = this_load;
3596 * Round up the averaging division if load is increasing. This
3597 * prevents us from getting stuck on 9 if the load is 10, for
3600 if (new_load > old_load)
3601 new_load += scale - 1;
3603 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3606 sched_avg_update(this_rq);
3609 static void update_cpu_load_active(struct rq *this_rq)
3611 update_cpu_load(this_rq);
3613 calc_load_account_active(this_rq);
3619 * sched_exec - execve() is a valuable balancing opportunity, because at
3620 * this point the task has the smallest effective memory and cache footprint.
3622 void sched_exec(void)
3624 struct task_struct *p = current;
3625 unsigned long flags;
3628 raw_spin_lock_irqsave(&p->pi_lock, flags);
3629 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3630 if (dest_cpu == smp_processor_id())
3633 if (likely(cpu_active(dest_cpu))) {
3634 struct migration_arg arg = { p, dest_cpu };
3636 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3637 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3641 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3646 DEFINE_PER_CPU(struct kernel_stat, kstat);
3648 EXPORT_PER_CPU_SYMBOL(kstat);
3651 * Return any ns on the sched_clock that have not yet been accounted in
3652 * @p in case that task is currently running.
3654 * Called with task_rq_lock() held on @rq.
3656 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3660 if (task_current(rq, p)) {
3661 update_rq_clock(rq);
3662 ns = rq->clock_task - p->se.exec_start;
3670 unsigned long long task_delta_exec(struct task_struct *p)
3672 unsigned long flags;
3676 rq = task_rq_lock(p, &flags);
3677 ns = do_task_delta_exec(p, rq);
3678 task_rq_unlock(rq, p, &flags);
3684 * Return accounted runtime for the task.
3685 * In case the task is currently running, return the runtime plus current's
3686 * pending runtime that have not been accounted yet.
3688 unsigned long long task_sched_runtime(struct task_struct *p)
3690 unsigned long flags;
3694 rq = task_rq_lock(p, &flags);
3695 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3696 task_rq_unlock(rq, p, &flags);
3702 * Return sum_exec_runtime for the thread group.
3703 * In case the task is currently running, return the sum plus current's
3704 * pending runtime that have not been accounted yet.
3706 * Note that the thread group might have other running tasks as well,
3707 * so the return value not includes other pending runtime that other
3708 * running tasks might have.
3710 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3712 struct task_cputime totals;
3713 unsigned long flags;
3717 rq = task_rq_lock(p, &flags);
3718 thread_group_cputime(p, &totals);
3719 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3720 task_rq_unlock(rq, p, &flags);
3726 * Account user cpu time to a process.
3727 * @p: the process that the cpu time gets accounted to
3728 * @cputime: the cpu time spent in user space since the last update
3729 * @cputime_scaled: cputime scaled by cpu frequency
3731 void account_user_time(struct task_struct *p, cputime_t cputime,
3732 cputime_t cputime_scaled)
3734 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3737 /* Add user time to process. */
3738 p->utime = cputime_add(p->utime, cputime);
3739 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3740 account_group_user_time(p, cputime);
3742 /* Add user time to cpustat. */
3743 tmp = cputime_to_cputime64(cputime);
3744 if (TASK_NICE(p) > 0)
3745 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3747 cpustat->user = cputime64_add(cpustat->user, tmp);
3749 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3750 /* Account for user time used */
3751 acct_update_integrals(p);
3755 * Account guest cpu time to a process.
3756 * @p: the process that the cpu time gets accounted to
3757 * @cputime: the cpu time spent in virtual machine since the last update
3758 * @cputime_scaled: cputime scaled by cpu frequency
3760 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3761 cputime_t cputime_scaled)
3764 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3766 tmp = cputime_to_cputime64(cputime);
3768 /* Add guest time to process. */
3769 p->utime = cputime_add(p->utime, cputime);
3770 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3771 account_group_user_time(p, cputime);
3772 p->gtime = cputime_add(p->gtime, cputime);
3774 /* Add guest time to cpustat. */
3775 if (TASK_NICE(p) > 0) {
3776 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3777 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3779 cpustat->user = cputime64_add(cpustat->user, tmp);
3780 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3785 * Account system cpu time to a process and desired cpustat field
3786 * @p: the process that the cpu time gets accounted to
3787 * @cputime: the cpu time spent in kernel space since the last update
3788 * @cputime_scaled: cputime scaled by cpu frequency
3789 * @target_cputime64: pointer to cpustat field that has to be updated
3792 void __account_system_time(struct task_struct *p, cputime_t cputime,
3793 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3795 cputime64_t tmp = cputime_to_cputime64(cputime);
3797 /* Add system time to process. */
3798 p->stime = cputime_add(p->stime, cputime);
3799 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3800 account_group_system_time(p, cputime);
3802 /* Add system time to cpustat. */
3803 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3804 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3806 /* Account for system time used */
3807 acct_update_integrals(p);
3811 * Account system cpu time to a process.
3812 * @p: the process that the cpu time gets accounted to
3813 * @hardirq_offset: the offset to subtract from hardirq_count()
3814 * @cputime: the cpu time spent in kernel space since the last update
3815 * @cputime_scaled: cputime scaled by cpu frequency
3817 void account_system_time(struct task_struct *p, int hardirq_offset,
3818 cputime_t cputime, cputime_t cputime_scaled)
3820 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3821 cputime64_t *target_cputime64;
3823 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3824 account_guest_time(p, cputime, cputime_scaled);
3828 if (hardirq_count() - hardirq_offset)
3829 target_cputime64 = &cpustat->irq;
3830 else if (in_serving_softirq())
3831 target_cputime64 = &cpustat->softirq;
3833 target_cputime64 = &cpustat->system;
3835 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3839 * Account for involuntary wait time.
3840 * @cputime: the cpu time spent in involuntary wait
3842 void account_steal_time(cputime_t cputime)
3844 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3845 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3847 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3851 * Account for idle time.
3852 * @cputime: the cpu time spent in idle wait
3854 void account_idle_time(cputime_t cputime)
3856 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3857 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3858 struct rq *rq = this_rq();
3860 if (atomic_read(&rq->nr_iowait) > 0)
3861 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3863 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3866 static __always_inline bool steal_account_process_tick(void)
3868 #ifdef CONFIG_PARAVIRT
3869 if (static_branch(¶virt_steal_enabled)) {
3872 steal = paravirt_steal_clock(smp_processor_id());
3873 steal -= this_rq()->prev_steal_time;
3875 st = steal_ticks(steal);
3876 this_rq()->prev_steal_time += st * TICK_NSEC;
3878 account_steal_time(st);
3885 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3887 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3889 * Account a tick to a process and cpustat
3890 * @p: the process that the cpu time gets accounted to
3891 * @user_tick: is the tick from userspace
3892 * @rq: the pointer to rq
3894 * Tick demultiplexing follows the order
3895 * - pending hardirq update
3896 * - pending softirq update
3900 * - check for guest_time
3901 * - else account as system_time
3903 * Check for hardirq is done both for system and user time as there is
3904 * no timer going off while we are on hardirq and hence we may never get an
3905 * opportunity to update it solely in system time.
3906 * p->stime and friends are only updated on system time and not on irq
3907 * softirq as those do not count in task exec_runtime any more.
3909 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3912 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3913 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
3914 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3916 if (steal_account_process_tick())
3919 if (irqtime_account_hi_update()) {
3920 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3921 } else if (irqtime_account_si_update()) {
3922 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3923 } else if (this_cpu_ksoftirqd() == p) {
3925 * ksoftirqd time do not get accounted in cpu_softirq_time.
3926 * So, we have to handle it separately here.
3927 * Also, p->stime needs to be updated for ksoftirqd.
3929 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3931 } else if (user_tick) {
3932 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3933 } else if (p == rq->idle) {
3934 account_idle_time(cputime_one_jiffy);
3935 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3936 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3938 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3943 static void irqtime_account_idle_ticks(int ticks)
3946 struct rq *rq = this_rq();
3948 for (i = 0; i < ticks; i++)
3949 irqtime_account_process_tick(current, 0, rq);
3951 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
3952 static void irqtime_account_idle_ticks(int ticks) {}
3953 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3955 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3958 * Account a single tick of cpu time.
3959 * @p: the process that the cpu time gets accounted to
3960 * @user_tick: indicates if the tick is a user or a system tick
3962 void account_process_tick(struct task_struct *p, int user_tick)
3964 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3965 struct rq *rq = this_rq();
3967 if (sched_clock_irqtime) {
3968 irqtime_account_process_tick(p, user_tick, rq);
3972 if (steal_account_process_tick())
3976 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3977 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3978 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3981 account_idle_time(cputime_one_jiffy);
3985 * Account multiple ticks of steal time.
3986 * @p: the process from which the cpu time has been stolen
3987 * @ticks: number of stolen ticks
3989 void account_steal_ticks(unsigned long ticks)
3991 account_steal_time(jiffies_to_cputime(ticks));
3995 * Account multiple ticks of idle time.
3996 * @ticks: number of stolen ticks
3998 void account_idle_ticks(unsigned long ticks)
4001 if (sched_clock_irqtime) {
4002 irqtime_account_idle_ticks(ticks);
4006 account_idle_time(jiffies_to_cputime(ticks));
4012 * Use precise platform statistics if available:
4014 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4015 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4021 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4023 struct task_cputime cputime;
4025 thread_group_cputime(p, &cputime);
4027 *ut = cputime.utime;
4028 *st = cputime.stime;
4032 #ifndef nsecs_to_cputime
4033 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
4036 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4038 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
4041 * Use CFS's precise accounting:
4043 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
4049 do_div(temp, total);
4050 utime = (cputime_t)temp;
4055 * Compare with previous values, to keep monotonicity:
4057 p->prev_utime = max(p->prev_utime, utime);
4058 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4060 *ut = p->prev_utime;
4061 *st = p->prev_stime;
4065 * Must be called with siglock held.
4067 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4069 struct signal_struct *sig = p->signal;
4070 struct task_cputime cputime;
4071 cputime_t rtime, utime, total;
4073 thread_group_cputime(p, &cputime);
4075 total = cputime_add(cputime.utime, cputime.stime);
4076 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4081 temp *= cputime.utime;
4082 do_div(temp, total);
4083 utime = (cputime_t)temp;
4087 sig->prev_utime = max(sig->prev_utime, utime);
4088 sig->prev_stime = max(sig->prev_stime,
4089 cputime_sub(rtime, sig->prev_utime));
4091 *ut = sig->prev_utime;
4092 *st = sig->prev_stime;
4097 * This function gets called by the timer code, with HZ frequency.
4098 * We call it with interrupts disabled.
4100 void scheduler_tick(void)
4102 int cpu = smp_processor_id();
4103 struct rq *rq = cpu_rq(cpu);
4104 struct task_struct *curr = rq->curr;
4108 raw_spin_lock(&rq->lock);
4109 update_rq_clock(rq);
4110 update_cpu_load_active(rq);
4111 curr->sched_class->task_tick(rq, curr, 0);
4112 raw_spin_unlock(&rq->lock);
4114 perf_event_task_tick();
4117 rq->idle_at_tick = idle_cpu(cpu);
4118 trigger_load_balance(rq, cpu);
4122 notrace unsigned long get_parent_ip(unsigned long addr)
4124 if (in_lock_functions(addr)) {
4125 addr = CALLER_ADDR2;
4126 if (in_lock_functions(addr))
4127 addr = CALLER_ADDR3;
4132 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4133 defined(CONFIG_PREEMPT_TRACER))
4135 void __kprobes add_preempt_count(int val)
4137 #ifdef CONFIG_DEBUG_PREEMPT
4141 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4144 preempt_count() += val;
4145 #ifdef CONFIG_DEBUG_PREEMPT
4147 * Spinlock count overflowing soon?
4149 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4152 if (preempt_count() == val)
4153 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4155 EXPORT_SYMBOL(add_preempt_count);
4157 void __kprobes sub_preempt_count(int val)
4159 #ifdef CONFIG_DEBUG_PREEMPT
4163 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4166 * Is the spinlock portion underflowing?
4168 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4169 !(preempt_count() & PREEMPT_MASK)))
4173 if (preempt_count() == val)
4174 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4175 preempt_count() -= val;
4177 EXPORT_SYMBOL(sub_preempt_count);
4182 * Print scheduling while atomic bug:
4184 static noinline void __schedule_bug(struct task_struct *prev)
4186 struct pt_regs *regs = get_irq_regs();
4188 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4189 prev->comm, prev->pid, preempt_count());
4191 debug_show_held_locks(prev);
4193 if (irqs_disabled())
4194 print_irqtrace_events(prev);
4203 * Various schedule()-time debugging checks and statistics:
4205 static inline void schedule_debug(struct task_struct *prev)
4208 * Test if we are atomic. Since do_exit() needs to call into
4209 * schedule() atomically, we ignore that path for now.
4210 * Otherwise, whine if we are scheduling when we should not be.
4212 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4213 __schedule_bug(prev);
4215 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4217 schedstat_inc(this_rq(), sched_count);
4220 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4222 if (prev->on_rq || rq->skip_clock_update < 0)
4223 update_rq_clock(rq);
4224 prev->sched_class->put_prev_task(rq, prev);
4228 * Pick up the highest-prio task:
4230 static inline struct task_struct *
4231 pick_next_task(struct rq *rq)
4233 const struct sched_class *class;
4234 struct task_struct *p;
4237 * Optimization: we know that if all tasks are in
4238 * the fair class we can call that function directly:
4240 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4241 p = fair_sched_class.pick_next_task(rq);
4246 for_each_class(class) {
4247 p = class->pick_next_task(rq);
4252 BUG(); /* the idle class will always have a runnable task */
4256 * schedule() is the main scheduler function.
4258 asmlinkage void __sched schedule(void)
4260 struct task_struct *prev, *next;
4261 unsigned long *switch_count;
4267 cpu = smp_processor_id();
4269 rcu_note_context_switch(cpu);
4272 schedule_debug(prev);
4274 if (sched_feat(HRTICK))
4277 raw_spin_lock_irq(&rq->lock);
4279 switch_count = &prev->nivcsw;
4280 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4281 if (unlikely(signal_pending_state(prev->state, prev))) {
4282 prev->state = TASK_RUNNING;
4284 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4288 * If a worker went to sleep, notify and ask workqueue
4289 * whether it wants to wake up a task to maintain
4292 if (prev->flags & PF_WQ_WORKER) {
4293 struct task_struct *to_wakeup;
4295 to_wakeup = wq_worker_sleeping(prev, cpu);
4297 try_to_wake_up_local(to_wakeup);
4301 * If we are going to sleep and we have plugged IO
4302 * queued, make sure to submit it to avoid deadlocks.
4304 if (blk_needs_flush_plug(prev)) {
4305 raw_spin_unlock(&rq->lock);
4306 blk_schedule_flush_plug(prev);
4307 raw_spin_lock(&rq->lock);
4310 switch_count = &prev->nvcsw;
4313 pre_schedule(rq, prev);
4315 if (unlikely(!rq->nr_running))
4316 idle_balance(cpu, rq);
4318 put_prev_task(rq, prev);
4319 next = pick_next_task(rq);
4320 clear_tsk_need_resched(prev);
4321 rq->skip_clock_update = 0;
4323 if (likely(prev != next)) {
4328 context_switch(rq, prev, next); /* unlocks the rq */
4330 * The context switch have flipped the stack from under us
4331 * and restored the local variables which were saved when
4332 * this task called schedule() in the past. prev == current
4333 * is still correct, but it can be moved to another cpu/rq.
4335 cpu = smp_processor_id();
4338 raw_spin_unlock_irq(&rq->lock);
4342 preempt_enable_no_resched();
4346 EXPORT_SYMBOL(schedule);
4348 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4350 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4355 if (lock->owner != owner)
4359 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4360 * lock->owner still matches owner, if that fails, owner might
4361 * point to free()d memory, if it still matches, the rcu_read_lock()
4362 * ensures the memory stays valid.
4366 ret = owner->on_cpu;
4374 * Look out! "owner" is an entirely speculative pointer
4375 * access and not reliable.
4377 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4379 if (!sched_feat(OWNER_SPIN))
4382 while (owner_running(lock, owner)) {
4386 arch_mutex_cpu_relax();
4390 * If the owner changed to another task there is likely
4391 * heavy contention, stop spinning.
4400 #ifdef CONFIG_PREEMPT
4402 * this is the entry point to schedule() from in-kernel preemption
4403 * off of preempt_enable. Kernel preemptions off return from interrupt
4404 * occur there and call schedule directly.
4406 asmlinkage void __sched notrace preempt_schedule(void)
4408 struct thread_info *ti = current_thread_info();
4411 * If there is a non-zero preempt_count or interrupts are disabled,
4412 * we do not want to preempt the current task. Just return..
4414 if (likely(ti->preempt_count || irqs_disabled()))
4418 add_preempt_count_notrace(PREEMPT_ACTIVE);
4420 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4423 * Check again in case we missed a preemption opportunity
4424 * between schedule and now.
4427 } while (need_resched());
4429 EXPORT_SYMBOL(preempt_schedule);
4432 * this is the entry point to schedule() from kernel preemption
4433 * off of irq context.
4434 * Note, that this is called and return with irqs disabled. This will
4435 * protect us against recursive calling from irq.
4437 asmlinkage void __sched preempt_schedule_irq(void)
4439 struct thread_info *ti = current_thread_info();
4441 /* Catch callers which need to be fixed */
4442 BUG_ON(ti->preempt_count || !irqs_disabled());
4445 add_preempt_count(PREEMPT_ACTIVE);
4448 local_irq_disable();
4449 sub_preempt_count(PREEMPT_ACTIVE);
4452 * Check again in case we missed a preemption opportunity
4453 * between schedule and now.
4456 } while (need_resched());
4459 #endif /* CONFIG_PREEMPT */
4461 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4464 return try_to_wake_up(curr->private, mode, wake_flags);
4466 EXPORT_SYMBOL(default_wake_function);
4469 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4470 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4471 * number) then we wake all the non-exclusive tasks and one exclusive task.
4473 * There are circumstances in which we can try to wake a task which has already
4474 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4475 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4477 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4478 int nr_exclusive, int wake_flags, void *key)
4480 wait_queue_t *curr, *next;
4482 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4483 unsigned flags = curr->flags;
4485 if (curr->func(curr, mode, wake_flags, key) &&
4486 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4492 * __wake_up - wake up threads blocked on a waitqueue.
4494 * @mode: which threads
4495 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4496 * @key: is directly passed to the wakeup function
4498 * It may be assumed that this function implies a write memory barrier before
4499 * changing the task state if and only if any tasks are woken up.
4501 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4502 int nr_exclusive, void *key)
4504 unsigned long flags;
4506 spin_lock_irqsave(&q->lock, flags);
4507 __wake_up_common(q, mode, nr_exclusive, 0, key);
4508 spin_unlock_irqrestore(&q->lock, flags);
4510 EXPORT_SYMBOL(__wake_up);
4513 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4515 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4517 __wake_up_common(q, mode, 1, 0, NULL);
4519 EXPORT_SYMBOL_GPL(__wake_up_locked);
4521 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4523 __wake_up_common(q, mode, 1, 0, key);
4525 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4528 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4530 * @mode: which threads
4531 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4532 * @key: opaque value to be passed to wakeup targets
4534 * The sync wakeup differs that the waker knows that it will schedule
4535 * away soon, so while the target thread will be woken up, it will not
4536 * be migrated to another CPU - ie. the two threads are 'synchronized'
4537 * with each other. This can prevent needless bouncing between CPUs.
4539 * On UP it can prevent extra preemption.
4541 * It may be assumed that this function implies a write memory barrier before
4542 * changing the task state if and only if any tasks are woken up.
4544 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4545 int nr_exclusive, void *key)
4547 unsigned long flags;
4548 int wake_flags = WF_SYNC;
4553 if (unlikely(!nr_exclusive))
4556 spin_lock_irqsave(&q->lock, flags);
4557 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4558 spin_unlock_irqrestore(&q->lock, flags);
4560 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4563 * __wake_up_sync - see __wake_up_sync_key()
4565 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4567 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4569 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4572 * complete: - signals a single thread waiting on this completion
4573 * @x: holds the state of this particular completion
4575 * This will wake up a single thread waiting on this completion. Threads will be
4576 * awakened in the same order in which they were queued.
4578 * See also complete_all(), wait_for_completion() and related routines.
4580 * It may be assumed that this function implies a write memory barrier before
4581 * changing the task state if and only if any tasks are woken up.
4583 void complete(struct completion *x)
4585 unsigned long flags;
4587 spin_lock_irqsave(&x->wait.lock, flags);
4589 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4590 spin_unlock_irqrestore(&x->wait.lock, flags);
4592 EXPORT_SYMBOL(complete);
4595 * complete_all: - signals all threads waiting on this completion
4596 * @x: holds the state of this particular completion
4598 * This will wake up all threads waiting on this particular completion event.
4600 * It may be assumed that this function implies a write memory barrier before
4601 * changing the task state if and only if any tasks are woken up.
4603 void complete_all(struct completion *x)
4605 unsigned long flags;
4607 spin_lock_irqsave(&x->wait.lock, flags);
4608 x->done += UINT_MAX/2;
4609 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4610 spin_unlock_irqrestore(&x->wait.lock, flags);
4612 EXPORT_SYMBOL(complete_all);
4614 static inline long __sched
4615 do_wait_for_common(struct completion *x, long timeout, int state)
4618 DECLARE_WAITQUEUE(wait, current);
4620 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4622 if (signal_pending_state(state, current)) {
4623 timeout = -ERESTARTSYS;
4626 __set_current_state(state);
4627 spin_unlock_irq(&x->wait.lock);
4628 timeout = schedule_timeout(timeout);
4629 spin_lock_irq(&x->wait.lock);
4630 } while (!x->done && timeout);
4631 __remove_wait_queue(&x->wait, &wait);
4636 return timeout ?: 1;
4640 wait_for_common(struct completion *x, long timeout, int state)
4644 spin_lock_irq(&x->wait.lock);
4645 timeout = do_wait_for_common(x, timeout, state);
4646 spin_unlock_irq(&x->wait.lock);
4651 * wait_for_completion: - waits for completion of a task
4652 * @x: holds the state of this particular completion
4654 * This waits to be signaled for completion of a specific task. It is NOT
4655 * interruptible and there is no timeout.
4657 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4658 * and interrupt capability. Also see complete().
4660 void __sched wait_for_completion(struct completion *x)
4662 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4664 EXPORT_SYMBOL(wait_for_completion);
4667 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4668 * @x: holds the state of this particular completion
4669 * @timeout: timeout value in jiffies
4671 * This waits for either a completion of a specific task to be signaled or for a
4672 * specified timeout to expire. The timeout is in jiffies. It is not
4675 unsigned long __sched
4676 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4678 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4680 EXPORT_SYMBOL(wait_for_completion_timeout);
4683 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4684 * @x: holds the state of this particular completion
4686 * This waits for completion of a specific task to be signaled. It is
4689 int __sched wait_for_completion_interruptible(struct completion *x)
4691 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4692 if (t == -ERESTARTSYS)
4696 EXPORT_SYMBOL(wait_for_completion_interruptible);
4699 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4700 * @x: holds the state of this particular completion
4701 * @timeout: timeout value in jiffies
4703 * This waits for either a completion of a specific task to be signaled or for a
4704 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4707 wait_for_completion_interruptible_timeout(struct completion *x,
4708 unsigned long timeout)
4710 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4712 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4715 * wait_for_completion_killable: - waits for completion of a task (killable)
4716 * @x: holds the state of this particular completion
4718 * This waits to be signaled for completion of a specific task. It can be
4719 * interrupted by a kill signal.
4721 int __sched wait_for_completion_killable(struct completion *x)
4723 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4724 if (t == -ERESTARTSYS)
4728 EXPORT_SYMBOL(wait_for_completion_killable);
4731 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4732 * @x: holds the state of this particular completion
4733 * @timeout: timeout value in jiffies
4735 * This waits for either a completion of a specific task to be
4736 * signaled or for a specified timeout to expire. It can be
4737 * interrupted by a kill signal. The timeout is in jiffies.
4740 wait_for_completion_killable_timeout(struct completion *x,
4741 unsigned long timeout)
4743 return wait_for_common(x, timeout, TASK_KILLABLE);
4745 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4748 * try_wait_for_completion - try to decrement a completion without blocking
4749 * @x: completion structure
4751 * Returns: 0 if a decrement cannot be done without blocking
4752 * 1 if a decrement succeeded.
4754 * If a completion is being used as a counting completion,
4755 * attempt to decrement the counter without blocking. This
4756 * enables us to avoid waiting if the resource the completion
4757 * is protecting is not available.
4759 bool try_wait_for_completion(struct completion *x)
4761 unsigned long flags;
4764 spin_lock_irqsave(&x->wait.lock, flags);
4769 spin_unlock_irqrestore(&x->wait.lock, flags);
4772 EXPORT_SYMBOL(try_wait_for_completion);
4775 * completion_done - Test to see if a completion has any waiters
4776 * @x: completion structure
4778 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4779 * 1 if there are no waiters.
4782 bool completion_done(struct completion *x)
4784 unsigned long flags;
4787 spin_lock_irqsave(&x->wait.lock, flags);
4790 spin_unlock_irqrestore(&x->wait.lock, flags);
4793 EXPORT_SYMBOL(completion_done);
4796 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4798 unsigned long flags;
4801 init_waitqueue_entry(&wait, current);
4803 __set_current_state(state);
4805 spin_lock_irqsave(&q->lock, flags);
4806 __add_wait_queue(q, &wait);
4807 spin_unlock(&q->lock);
4808 timeout = schedule_timeout(timeout);
4809 spin_lock_irq(&q->lock);
4810 __remove_wait_queue(q, &wait);
4811 spin_unlock_irqrestore(&q->lock, flags);
4816 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4818 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4820 EXPORT_SYMBOL(interruptible_sleep_on);
4823 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4825 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4827 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4829 void __sched sleep_on(wait_queue_head_t *q)
4831 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4833 EXPORT_SYMBOL(sleep_on);
4835 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4837 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4839 EXPORT_SYMBOL(sleep_on_timeout);
4841 #ifdef CONFIG_RT_MUTEXES
4844 * rt_mutex_setprio - set the current priority of a task
4846 * @prio: prio value (kernel-internal form)
4848 * This function changes the 'effective' priority of a task. It does
4849 * not touch ->normal_prio like __setscheduler().
4851 * Used by the rt_mutex code to implement priority inheritance logic.
4853 void rt_mutex_setprio(struct task_struct *p, int prio)
4855 int oldprio, on_rq, running;
4857 const struct sched_class *prev_class;
4859 BUG_ON(prio < 0 || prio > MAX_PRIO);
4861 rq = __task_rq_lock(p);
4863 trace_sched_pi_setprio(p, prio);
4865 prev_class = p->sched_class;
4867 running = task_current(rq, p);
4869 dequeue_task(rq, p, 0);
4871 p->sched_class->put_prev_task(rq, p);
4874 p->sched_class = &rt_sched_class;
4876 p->sched_class = &fair_sched_class;
4881 p->sched_class->set_curr_task(rq);
4883 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4885 check_class_changed(rq, p, prev_class, oldprio);
4886 __task_rq_unlock(rq);
4891 void set_user_nice(struct task_struct *p, long nice)
4893 int old_prio, delta, on_rq;
4894 unsigned long flags;
4897 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4900 * We have to be careful, if called from sys_setpriority(),
4901 * the task might be in the middle of scheduling on another CPU.
4903 rq = task_rq_lock(p, &flags);
4905 * The RT priorities are set via sched_setscheduler(), but we still
4906 * allow the 'normal' nice value to be set - but as expected
4907 * it wont have any effect on scheduling until the task is
4908 * SCHED_FIFO/SCHED_RR:
4910 if (task_has_rt_policy(p)) {
4911 p->static_prio = NICE_TO_PRIO(nice);
4916 dequeue_task(rq, p, 0);
4918 p->static_prio = NICE_TO_PRIO(nice);
4921 p->prio = effective_prio(p);
4922 delta = p->prio - old_prio;
4925 enqueue_task(rq, p, 0);
4927 * If the task increased its priority or is running and
4928 * lowered its priority, then reschedule its CPU:
4930 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4931 resched_task(rq->curr);
4934 task_rq_unlock(rq, p, &flags);
4936 EXPORT_SYMBOL(set_user_nice);
4939 * can_nice - check if a task can reduce its nice value
4943 int can_nice(const struct task_struct *p, const int nice)
4945 /* convert nice value [19,-20] to rlimit style value [1,40] */
4946 int nice_rlim = 20 - nice;
4948 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4949 capable(CAP_SYS_NICE));
4952 #ifdef __ARCH_WANT_SYS_NICE
4955 * sys_nice - change the priority of the current process.
4956 * @increment: priority increment
4958 * sys_setpriority is a more generic, but much slower function that
4959 * does similar things.
4961 SYSCALL_DEFINE1(nice, int, increment)
4966 * Setpriority might change our priority at the same moment.
4967 * We don't have to worry. Conceptually one call occurs first
4968 * and we have a single winner.
4970 if (increment < -40)
4975 nice = TASK_NICE(current) + increment;
4981 if (increment < 0 && !can_nice(current, nice))
4984 retval = security_task_setnice(current, nice);
4988 set_user_nice(current, nice);
4995 * task_prio - return the priority value of a given task.
4996 * @p: the task in question.
4998 * This is the priority value as seen by users in /proc.
4999 * RT tasks are offset by -200. Normal tasks are centered
5000 * around 0, value goes from -16 to +15.
5002 int task_prio(const struct task_struct *p)
5004 return p->prio - MAX_RT_PRIO;
5008 * task_nice - return the nice value of a given task.
5009 * @p: the task in question.
5011 int task_nice(const struct task_struct *p)
5013 return TASK_NICE(p);
5015 EXPORT_SYMBOL(task_nice);
5018 * idle_cpu - is a given cpu idle currently?
5019 * @cpu: the processor in question.
5021 int idle_cpu(int cpu)
5023 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5027 * idle_task - return the idle task for a given cpu.
5028 * @cpu: the processor in question.
5030 struct task_struct *idle_task(int cpu)
5032 return cpu_rq(cpu)->idle;
5036 * find_process_by_pid - find a process with a matching PID value.
5037 * @pid: the pid in question.
5039 static struct task_struct *find_process_by_pid(pid_t pid)
5041 return pid ? find_task_by_vpid(pid) : current;
5044 /* Actually do priority change: must hold rq lock. */
5046 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5049 p->rt_priority = prio;
5050 p->normal_prio = normal_prio(p);
5051 /* we are holding p->pi_lock already */
5052 p->prio = rt_mutex_getprio(p);
5053 if (rt_prio(p->prio))
5054 p->sched_class = &rt_sched_class;
5056 p->sched_class = &fair_sched_class;
5061 * check the target process has a UID that matches the current process's
5063 static bool check_same_owner(struct task_struct *p)
5065 const struct cred *cred = current_cred(), *pcred;
5069 pcred = __task_cred(p);
5070 if (cred->user->user_ns == pcred->user->user_ns)
5071 match = (cred->euid == pcred->euid ||
5072 cred->euid == pcred->uid);
5079 static int __sched_setscheduler(struct task_struct *p, int policy,
5080 const struct sched_param *param, bool user)
5082 int retval, oldprio, oldpolicy = -1, on_rq, running;
5083 unsigned long flags;
5084 const struct sched_class *prev_class;
5088 /* may grab non-irq protected spin_locks */
5089 BUG_ON(in_interrupt());
5091 /* double check policy once rq lock held */
5093 reset_on_fork = p->sched_reset_on_fork;
5094 policy = oldpolicy = p->policy;
5096 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5097 policy &= ~SCHED_RESET_ON_FORK;
5099 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5100 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5101 policy != SCHED_IDLE)
5106 * Valid priorities for SCHED_FIFO and SCHED_RR are
5107 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5108 * SCHED_BATCH and SCHED_IDLE is 0.
5110 if (param->sched_priority < 0 ||
5111 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5112 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5114 if (rt_policy(policy) != (param->sched_priority != 0))
5118 * Allow unprivileged RT tasks to decrease priority:
5120 if (user && !capable(CAP_SYS_NICE)) {
5121 if (rt_policy(policy)) {
5122 unsigned long rlim_rtprio =
5123 task_rlimit(p, RLIMIT_RTPRIO);
5125 /* can't set/change the rt policy */
5126 if (policy != p->policy && !rlim_rtprio)
5129 /* can't increase priority */
5130 if (param->sched_priority > p->rt_priority &&
5131 param->sched_priority > rlim_rtprio)
5136 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5137 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5139 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5140 if (!can_nice(p, TASK_NICE(p)))
5144 /* can't change other user's priorities */
5145 if (!check_same_owner(p))
5148 /* Normal users shall not reset the sched_reset_on_fork flag */
5149 if (p->sched_reset_on_fork && !reset_on_fork)
5154 retval = security_task_setscheduler(p);
5160 * make sure no PI-waiters arrive (or leave) while we are
5161 * changing the priority of the task:
5163 * To be able to change p->policy safely, the appropriate
5164 * runqueue lock must be held.
5166 rq = task_rq_lock(p, &flags);
5169 * Changing the policy of the stop threads its a very bad idea
5171 if (p == rq->stop) {
5172 task_rq_unlock(rq, p, &flags);
5177 * If not changing anything there's no need to proceed further:
5179 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5180 param->sched_priority == p->rt_priority))) {
5182 __task_rq_unlock(rq);
5183 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5187 #ifdef CONFIG_RT_GROUP_SCHED
5190 * Do not allow realtime tasks into groups that have no runtime
5193 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5194 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5195 !task_group_is_autogroup(task_group(p))) {
5196 task_rq_unlock(rq, p, &flags);
5202 /* recheck policy now with rq lock held */
5203 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5204 policy = oldpolicy = -1;
5205 task_rq_unlock(rq, p, &flags);
5209 running = task_current(rq, p);
5211 deactivate_task(rq, p, 0);
5213 p->sched_class->put_prev_task(rq, p);
5215 p->sched_reset_on_fork = reset_on_fork;
5218 prev_class = p->sched_class;
5219 __setscheduler(rq, p, policy, param->sched_priority);
5222 p->sched_class->set_curr_task(rq);
5224 activate_task(rq, p, 0);
5226 check_class_changed(rq, p, prev_class, oldprio);
5227 task_rq_unlock(rq, p, &flags);
5229 rt_mutex_adjust_pi(p);
5235 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5236 * @p: the task in question.
5237 * @policy: new policy.
5238 * @param: structure containing the new RT priority.
5240 * NOTE that the task may be already dead.
5242 int sched_setscheduler(struct task_struct *p, int policy,
5243 const struct sched_param *param)
5245 return __sched_setscheduler(p, policy, param, true);
5247 EXPORT_SYMBOL_GPL(sched_setscheduler);
5250 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5251 * @p: the task in question.
5252 * @policy: new policy.
5253 * @param: structure containing the new RT priority.
5255 * Just like sched_setscheduler, only don't bother checking if the
5256 * current context has permission. For example, this is needed in
5257 * stop_machine(): we create temporary high priority worker threads,
5258 * but our caller might not have that capability.
5260 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5261 const struct sched_param *param)
5263 return __sched_setscheduler(p, policy, param, false);
5267 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5269 struct sched_param lparam;
5270 struct task_struct *p;
5273 if (!param || pid < 0)
5275 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5280 p = find_process_by_pid(pid);
5282 retval = sched_setscheduler(p, policy, &lparam);
5289 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5290 * @pid: the pid in question.
5291 * @policy: new policy.
5292 * @param: structure containing the new RT priority.
5294 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5295 struct sched_param __user *, param)
5297 /* negative values for policy are not valid */
5301 return do_sched_setscheduler(pid, policy, param);
5305 * sys_sched_setparam - set/change the RT priority of a thread
5306 * @pid: the pid in question.
5307 * @param: structure containing the new RT priority.
5309 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5311 return do_sched_setscheduler(pid, -1, param);
5315 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5316 * @pid: the pid in question.
5318 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5320 struct task_struct *p;
5328 p = find_process_by_pid(pid);
5330 retval = security_task_getscheduler(p);
5333 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5340 * sys_sched_getparam - get the RT priority of a thread
5341 * @pid: the pid in question.
5342 * @param: structure containing the RT priority.
5344 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5346 struct sched_param lp;
5347 struct task_struct *p;
5350 if (!param || pid < 0)
5354 p = find_process_by_pid(pid);
5359 retval = security_task_getscheduler(p);
5363 lp.sched_priority = p->rt_priority;
5367 * This one might sleep, we cannot do it with a spinlock held ...
5369 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5378 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5380 cpumask_var_t cpus_allowed, new_mask;
5381 struct task_struct *p;
5387 p = find_process_by_pid(pid);
5394 /* Prevent p going away */
5398 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5402 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5404 goto out_free_cpus_allowed;
5407 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5410 retval = security_task_setscheduler(p);
5414 cpuset_cpus_allowed(p, cpus_allowed);
5415 cpumask_and(new_mask, in_mask, cpus_allowed);
5417 retval = set_cpus_allowed_ptr(p, new_mask);
5420 cpuset_cpus_allowed(p, cpus_allowed);
5421 if (!cpumask_subset(new_mask, cpus_allowed)) {
5423 * We must have raced with a concurrent cpuset
5424 * update. Just reset the cpus_allowed to the
5425 * cpuset's cpus_allowed
5427 cpumask_copy(new_mask, cpus_allowed);
5432 free_cpumask_var(new_mask);
5433 out_free_cpus_allowed:
5434 free_cpumask_var(cpus_allowed);
5441 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5442 struct cpumask *new_mask)
5444 if (len < cpumask_size())
5445 cpumask_clear(new_mask);
5446 else if (len > cpumask_size())
5447 len = cpumask_size();
5449 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5453 * sys_sched_setaffinity - set the cpu affinity of a process
5454 * @pid: pid of the process
5455 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5456 * @user_mask_ptr: user-space pointer to the new cpu mask
5458 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5459 unsigned long __user *, user_mask_ptr)
5461 cpumask_var_t new_mask;
5464 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5467 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5469 retval = sched_setaffinity(pid, new_mask);
5470 free_cpumask_var(new_mask);
5474 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5476 struct task_struct *p;
5477 unsigned long flags;
5484 p = find_process_by_pid(pid);
5488 retval = security_task_getscheduler(p);
5492 raw_spin_lock_irqsave(&p->pi_lock, flags);
5493 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5494 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5504 * sys_sched_getaffinity - get the cpu affinity of a process
5505 * @pid: pid of the process
5506 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5507 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5509 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5510 unsigned long __user *, user_mask_ptr)
5515 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5517 if (len & (sizeof(unsigned long)-1))
5520 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5523 ret = sched_getaffinity(pid, mask);
5525 size_t retlen = min_t(size_t, len, cpumask_size());
5527 if (copy_to_user(user_mask_ptr, mask, retlen))
5532 free_cpumask_var(mask);
5538 * sys_sched_yield - yield the current processor to other threads.
5540 * This function yields the current CPU to other tasks. If there are no
5541 * other threads running on this CPU then this function will return.
5543 SYSCALL_DEFINE0(sched_yield)
5545 struct rq *rq = this_rq_lock();
5547 schedstat_inc(rq, yld_count);
5548 current->sched_class->yield_task(rq);
5551 * Since we are going to call schedule() anyway, there's
5552 * no need to preempt or enable interrupts:
5554 __release(rq->lock);
5555 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5556 do_raw_spin_unlock(&rq->lock);
5557 preempt_enable_no_resched();
5564 static inline int should_resched(void)
5566 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5569 static void __cond_resched(void)
5571 add_preempt_count(PREEMPT_ACTIVE);
5573 sub_preempt_count(PREEMPT_ACTIVE);
5576 int __sched _cond_resched(void)
5578 if (should_resched()) {
5584 EXPORT_SYMBOL(_cond_resched);
5587 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5588 * call schedule, and on return reacquire the lock.
5590 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5591 * operations here to prevent schedule() from being called twice (once via
5592 * spin_unlock(), once by hand).
5594 int __cond_resched_lock(spinlock_t *lock)
5596 int resched = should_resched();
5599 lockdep_assert_held(lock);
5601 if (spin_needbreak(lock) || resched) {
5612 EXPORT_SYMBOL(__cond_resched_lock);
5614 int __sched __cond_resched_softirq(void)
5616 BUG_ON(!in_softirq());
5618 if (should_resched()) {
5626 EXPORT_SYMBOL(__cond_resched_softirq);
5629 * yield - yield the current processor to other threads.
5631 * This is a shortcut for kernel-space yielding - it marks the
5632 * thread runnable and calls sys_sched_yield().
5634 void __sched yield(void)
5636 set_current_state(TASK_RUNNING);
5639 EXPORT_SYMBOL(yield);
5642 * yield_to - yield the current processor to another thread in
5643 * your thread group, or accelerate that thread toward the
5644 * processor it's on.
5646 * @preempt: whether task preemption is allowed or not
5648 * It's the caller's job to ensure that the target task struct
5649 * can't go away on us before we can do any checks.
5651 * Returns true if we indeed boosted the target task.
5653 bool __sched yield_to(struct task_struct *p, bool preempt)
5655 struct task_struct *curr = current;
5656 struct rq *rq, *p_rq;
5657 unsigned long flags;
5660 local_irq_save(flags);
5665 double_rq_lock(rq, p_rq);
5666 while (task_rq(p) != p_rq) {
5667 double_rq_unlock(rq, p_rq);
5671 if (!curr->sched_class->yield_to_task)
5674 if (curr->sched_class != p->sched_class)
5677 if (task_running(p_rq, p) || p->state)
5680 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5682 schedstat_inc(rq, yld_count);
5684 * Make p's CPU reschedule; pick_next_entity takes care of
5687 if (preempt && rq != p_rq)
5688 resched_task(p_rq->curr);
5692 double_rq_unlock(rq, p_rq);
5693 local_irq_restore(flags);
5700 EXPORT_SYMBOL_GPL(yield_to);
5703 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5704 * that process accounting knows that this is a task in IO wait state.
5706 void __sched io_schedule(void)
5708 struct rq *rq = raw_rq();
5710 delayacct_blkio_start();
5711 atomic_inc(&rq->nr_iowait);
5712 blk_flush_plug(current);
5713 current->in_iowait = 1;
5715 current->in_iowait = 0;
5716 atomic_dec(&rq->nr_iowait);
5717 delayacct_blkio_end();
5719 EXPORT_SYMBOL(io_schedule);
5721 long __sched io_schedule_timeout(long timeout)
5723 struct rq *rq = raw_rq();
5726 delayacct_blkio_start();
5727 atomic_inc(&rq->nr_iowait);
5728 blk_flush_plug(current);
5729 current->in_iowait = 1;
5730 ret = schedule_timeout(timeout);
5731 current->in_iowait = 0;
5732 atomic_dec(&rq->nr_iowait);
5733 delayacct_blkio_end();
5738 * sys_sched_get_priority_max - return maximum RT priority.
5739 * @policy: scheduling class.
5741 * this syscall returns the maximum rt_priority that can be used
5742 * by a given scheduling class.
5744 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5751 ret = MAX_USER_RT_PRIO-1;
5763 * sys_sched_get_priority_min - return minimum RT priority.
5764 * @policy: scheduling class.
5766 * this syscall returns the minimum rt_priority that can be used
5767 * by a given scheduling class.
5769 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5787 * sys_sched_rr_get_interval - return the default timeslice of a process.
5788 * @pid: pid of the process.
5789 * @interval: userspace pointer to the timeslice value.
5791 * this syscall writes the default timeslice value of a given process
5792 * into the user-space timespec buffer. A value of '0' means infinity.
5794 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5795 struct timespec __user *, interval)
5797 struct task_struct *p;
5798 unsigned int time_slice;
5799 unsigned long flags;
5809 p = find_process_by_pid(pid);
5813 retval = security_task_getscheduler(p);
5817 rq = task_rq_lock(p, &flags);
5818 time_slice = p->sched_class->get_rr_interval(rq, p);
5819 task_rq_unlock(rq, p, &flags);
5822 jiffies_to_timespec(time_slice, &t);
5823 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5831 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5833 void sched_show_task(struct task_struct *p)
5835 unsigned long free = 0;
5838 state = p->state ? __ffs(p->state) + 1 : 0;
5839 printk(KERN_INFO "%-15.15s %c", p->comm,
5840 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5841 #if BITS_PER_LONG == 32
5842 if (state == TASK_RUNNING)
5843 printk(KERN_CONT " running ");
5845 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5847 if (state == TASK_RUNNING)
5848 printk(KERN_CONT " running task ");
5850 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5852 #ifdef CONFIG_DEBUG_STACK_USAGE
5853 free = stack_not_used(p);
5855 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5856 task_pid_nr(p), task_pid_nr(p->real_parent),
5857 (unsigned long)task_thread_info(p)->flags);
5859 show_stack(p, NULL);
5862 void show_state_filter(unsigned long state_filter)
5864 struct task_struct *g, *p;
5866 #if BITS_PER_LONG == 32
5868 " task PC stack pid father\n");
5871 " task PC stack pid father\n");
5873 read_lock(&tasklist_lock);
5874 do_each_thread(g, p) {
5876 * reset the NMI-timeout, listing all files on a slow
5877 * console might take a lot of time:
5879 touch_nmi_watchdog();
5880 if (!state_filter || (p->state & state_filter))
5882 } while_each_thread(g, p);
5884 touch_all_softlockup_watchdogs();
5886 #ifdef CONFIG_SCHED_DEBUG
5887 sysrq_sched_debug_show();
5889 read_unlock(&tasklist_lock);
5891 * Only show locks if all tasks are dumped:
5894 debug_show_all_locks();
5897 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5899 idle->sched_class = &idle_sched_class;
5903 * init_idle - set up an idle thread for a given CPU
5904 * @idle: task in question
5905 * @cpu: cpu the idle task belongs to
5907 * NOTE: this function does not set the idle thread's NEED_RESCHED
5908 * flag, to make booting more robust.
5910 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5912 struct rq *rq = cpu_rq(cpu);
5913 unsigned long flags;
5915 raw_spin_lock_irqsave(&rq->lock, flags);
5918 idle->state = TASK_RUNNING;
5919 idle->se.exec_start = sched_clock();
5921 do_set_cpus_allowed(idle, cpumask_of(cpu));
5923 * We're having a chicken and egg problem, even though we are
5924 * holding rq->lock, the cpu isn't yet set to this cpu so the
5925 * lockdep check in task_group() will fail.
5927 * Similar case to sched_fork(). / Alternatively we could
5928 * use task_rq_lock() here and obtain the other rq->lock.
5933 __set_task_cpu(idle, cpu);
5936 rq->curr = rq->idle = idle;
5937 #if defined(CONFIG_SMP)
5940 raw_spin_unlock_irqrestore(&rq->lock, flags);
5942 /* Set the preempt count _outside_ the spinlocks! */
5943 task_thread_info(idle)->preempt_count = 0;
5946 * The idle tasks have their own, simple scheduling class:
5948 idle->sched_class = &idle_sched_class;
5949 ftrace_graph_init_idle_task(idle, cpu);
5953 * In a system that switches off the HZ timer nohz_cpu_mask
5954 * indicates which cpus entered this state. This is used
5955 * in the rcu update to wait only for active cpus. For system
5956 * which do not switch off the HZ timer nohz_cpu_mask should
5957 * always be CPU_BITS_NONE.
5959 cpumask_var_t nohz_cpu_mask;
5962 * Increase the granularity value when there are more CPUs,
5963 * because with more CPUs the 'effective latency' as visible
5964 * to users decreases. But the relationship is not linear,
5965 * so pick a second-best guess by going with the log2 of the
5968 * This idea comes from the SD scheduler of Con Kolivas:
5970 static int get_update_sysctl_factor(void)
5972 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5973 unsigned int factor;
5975 switch (sysctl_sched_tunable_scaling) {
5976 case SCHED_TUNABLESCALING_NONE:
5979 case SCHED_TUNABLESCALING_LINEAR:
5982 case SCHED_TUNABLESCALING_LOG:
5984 factor = 1 + ilog2(cpus);
5991 static void update_sysctl(void)
5993 unsigned int factor = get_update_sysctl_factor();
5995 #define SET_SYSCTL(name) \
5996 (sysctl_##name = (factor) * normalized_sysctl_##name)
5997 SET_SYSCTL(sched_min_granularity);
5998 SET_SYSCTL(sched_latency);
5999 SET_SYSCTL(sched_wakeup_granularity);
6003 static inline void sched_init_granularity(void)
6009 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
6011 if (p->sched_class && p->sched_class->set_cpus_allowed)
6012 p->sched_class->set_cpus_allowed(p, new_mask);
6014 cpumask_copy(&p->cpus_allowed, new_mask);
6015 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6020 * This is how migration works:
6022 * 1) we invoke migration_cpu_stop() on the target CPU using
6024 * 2) stopper starts to run (implicitly forcing the migrated thread
6026 * 3) it checks whether the migrated task is still in the wrong runqueue.
6027 * 4) if it's in the wrong runqueue then the migration thread removes
6028 * it and puts it into the right queue.
6029 * 5) stopper completes and stop_one_cpu() returns and the migration
6034 * Change a given task's CPU affinity. Migrate the thread to a
6035 * proper CPU and schedule it away if the CPU it's executing on
6036 * is removed from the allowed bitmask.
6038 * NOTE: the caller must have a valid reference to the task, the
6039 * task must not exit() & deallocate itself prematurely. The
6040 * call is not atomic; no spinlocks may be held.
6042 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6044 unsigned long flags;
6046 unsigned int dest_cpu;
6049 rq = task_rq_lock(p, &flags);
6051 if (cpumask_equal(&p->cpus_allowed, new_mask))
6054 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6059 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6064 do_set_cpus_allowed(p, new_mask);
6066 /* Can the task run on the task's current CPU? If so, we're done */
6067 if (cpumask_test_cpu(task_cpu(p), new_mask))
6070 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6072 struct migration_arg arg = { p, dest_cpu };
6073 /* Need help from migration thread: drop lock and wait. */
6074 task_rq_unlock(rq, p, &flags);
6075 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6076 tlb_migrate_finish(p->mm);
6080 task_rq_unlock(rq, p, &flags);
6084 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6087 * Move (not current) task off this cpu, onto dest cpu. We're doing
6088 * this because either it can't run here any more (set_cpus_allowed()
6089 * away from this CPU, or CPU going down), or because we're
6090 * attempting to rebalance this task on exec (sched_exec).
6092 * So we race with normal scheduler movements, but that's OK, as long
6093 * as the task is no longer on this CPU.
6095 * Returns non-zero if task was successfully migrated.
6097 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6099 struct rq *rq_dest, *rq_src;
6102 if (unlikely(!cpu_active(dest_cpu)))
6105 rq_src = cpu_rq(src_cpu);
6106 rq_dest = cpu_rq(dest_cpu);
6108 raw_spin_lock(&p->pi_lock);
6109 double_rq_lock(rq_src, rq_dest);
6110 /* Already moved. */
6111 if (task_cpu(p) != src_cpu)
6113 /* Affinity changed (again). */
6114 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6118 * If we're not on a rq, the next wake-up will ensure we're
6122 deactivate_task(rq_src, p, 0);
6123 set_task_cpu(p, dest_cpu);
6124 activate_task(rq_dest, p, 0);
6125 check_preempt_curr(rq_dest, p, 0);
6130 double_rq_unlock(rq_src, rq_dest);
6131 raw_spin_unlock(&p->pi_lock);
6136 * migration_cpu_stop - this will be executed by a highprio stopper thread
6137 * and performs thread migration by bumping thread off CPU then
6138 * 'pushing' onto another runqueue.
6140 static int migration_cpu_stop(void *data)
6142 struct migration_arg *arg = data;
6145 * The original target cpu might have gone down and we might
6146 * be on another cpu but it doesn't matter.
6148 local_irq_disable();
6149 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6154 #ifdef CONFIG_HOTPLUG_CPU
6157 * Ensures that the idle task is using init_mm right before its cpu goes
6160 void idle_task_exit(void)
6162 struct mm_struct *mm = current->active_mm;
6164 BUG_ON(cpu_online(smp_processor_id()));
6167 switch_mm(mm, &init_mm, current);
6172 * While a dead CPU has no uninterruptible tasks queued at this point,
6173 * it might still have a nonzero ->nr_uninterruptible counter, because
6174 * for performance reasons the counter is not stricly tracking tasks to
6175 * their home CPUs. So we just add the counter to another CPU's counter,
6176 * to keep the global sum constant after CPU-down:
6178 static void migrate_nr_uninterruptible(struct rq *rq_src)
6180 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6182 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6183 rq_src->nr_uninterruptible = 0;
6187 * remove the tasks which were accounted by rq from calc_load_tasks.
6189 static void calc_global_load_remove(struct rq *rq)
6191 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6192 rq->calc_load_active = 0;
6196 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6197 * try_to_wake_up()->select_task_rq().
6199 * Called with rq->lock held even though we'er in stop_machine() and
6200 * there's no concurrency possible, we hold the required locks anyway
6201 * because of lock validation efforts.
6203 static void migrate_tasks(unsigned int dead_cpu)
6205 struct rq *rq = cpu_rq(dead_cpu);
6206 struct task_struct *next, *stop = rq->stop;
6210 * Fudge the rq selection such that the below task selection loop
6211 * doesn't get stuck on the currently eligible stop task.
6213 * We're currently inside stop_machine() and the rq is either stuck
6214 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6215 * either way we should never end up calling schedule() until we're
6222 * There's this thread running, bail when that's the only
6225 if (rq->nr_running == 1)
6228 next = pick_next_task(rq);
6230 next->sched_class->put_prev_task(rq, next);
6232 /* Find suitable destination for @next, with force if needed. */
6233 dest_cpu = select_fallback_rq(dead_cpu, next);
6234 raw_spin_unlock(&rq->lock);
6236 __migrate_task(next, dead_cpu, dest_cpu);
6238 raw_spin_lock(&rq->lock);
6244 #endif /* CONFIG_HOTPLUG_CPU */
6246 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6248 static struct ctl_table sd_ctl_dir[] = {
6250 .procname = "sched_domain",
6256 static struct ctl_table sd_ctl_root[] = {
6258 .procname = "kernel",
6260 .child = sd_ctl_dir,
6265 static struct ctl_table *sd_alloc_ctl_entry(int n)
6267 struct ctl_table *entry =
6268 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6273 static void sd_free_ctl_entry(struct ctl_table **tablep)
6275 struct ctl_table *entry;
6278 * In the intermediate directories, both the child directory and
6279 * procname are dynamically allocated and could fail but the mode
6280 * will always be set. In the lowest directory the names are
6281 * static strings and all have proc handlers.
6283 for (entry = *tablep; entry->mode; entry++) {
6285 sd_free_ctl_entry(&entry->child);
6286 if (entry->proc_handler == NULL)
6287 kfree(entry->procname);
6295 set_table_entry(struct ctl_table *entry,
6296 const char *procname, void *data, int maxlen,
6297 mode_t mode, proc_handler *proc_handler)
6299 entry->procname = procname;
6301 entry->maxlen = maxlen;
6303 entry->proc_handler = proc_handler;
6306 static struct ctl_table *
6307 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6309 struct ctl_table *table = sd_alloc_ctl_entry(13);
6314 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6315 sizeof(long), 0644, proc_doulongvec_minmax);
6316 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6317 sizeof(long), 0644, proc_doulongvec_minmax);
6318 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6319 sizeof(int), 0644, proc_dointvec_minmax);
6320 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6321 sizeof(int), 0644, proc_dointvec_minmax);
6322 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6323 sizeof(int), 0644, proc_dointvec_minmax);
6324 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6325 sizeof(int), 0644, proc_dointvec_minmax);
6326 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6327 sizeof(int), 0644, proc_dointvec_minmax);
6328 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6329 sizeof(int), 0644, proc_dointvec_minmax);
6330 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6331 sizeof(int), 0644, proc_dointvec_minmax);
6332 set_table_entry(&table[9], "cache_nice_tries",
6333 &sd->cache_nice_tries,
6334 sizeof(int), 0644, proc_dointvec_minmax);
6335 set_table_entry(&table[10], "flags", &sd->flags,
6336 sizeof(int), 0644, proc_dointvec_minmax);
6337 set_table_entry(&table[11], "name", sd->name,
6338 CORENAME_MAX_SIZE, 0444, proc_dostring);
6339 /* &table[12] is terminator */
6344 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6346 struct ctl_table *entry, *table;
6347 struct sched_domain *sd;
6348 int domain_num = 0, i;
6351 for_each_domain(cpu, sd)
6353 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6358 for_each_domain(cpu, sd) {
6359 snprintf(buf, 32, "domain%d", i);
6360 entry->procname = kstrdup(buf, GFP_KERNEL);
6362 entry->child = sd_alloc_ctl_domain_table(sd);
6369 static struct ctl_table_header *sd_sysctl_header;
6370 static void register_sched_domain_sysctl(void)
6372 int i, cpu_num = num_possible_cpus();
6373 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6376 WARN_ON(sd_ctl_dir[0].child);
6377 sd_ctl_dir[0].child = entry;
6382 for_each_possible_cpu(i) {
6383 snprintf(buf, 32, "cpu%d", i);
6384 entry->procname = kstrdup(buf, GFP_KERNEL);
6386 entry->child = sd_alloc_ctl_cpu_table(i);
6390 WARN_ON(sd_sysctl_header);
6391 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6394 /* may be called multiple times per register */
6395 static void unregister_sched_domain_sysctl(void)
6397 if (sd_sysctl_header)
6398 unregister_sysctl_table(sd_sysctl_header);
6399 sd_sysctl_header = NULL;
6400 if (sd_ctl_dir[0].child)
6401 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6404 static void register_sched_domain_sysctl(void)
6407 static void unregister_sched_domain_sysctl(void)
6412 static void set_rq_online(struct rq *rq)
6415 const struct sched_class *class;
6417 cpumask_set_cpu(rq->cpu, rq->rd->online);
6420 for_each_class(class) {
6421 if (class->rq_online)
6422 class->rq_online(rq);
6427 static void set_rq_offline(struct rq *rq)
6430 const struct sched_class *class;
6432 for_each_class(class) {
6433 if (class->rq_offline)
6434 class->rq_offline(rq);
6437 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6443 * migration_call - callback that gets triggered when a CPU is added.
6444 * Here we can start up the necessary migration thread for the new CPU.
6446 static int __cpuinit
6447 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6449 int cpu = (long)hcpu;
6450 unsigned long flags;
6451 struct rq *rq = cpu_rq(cpu);
6453 switch (action & ~CPU_TASKS_FROZEN) {
6455 case CPU_UP_PREPARE:
6456 rq->calc_load_update = calc_load_update;
6460 /* Update our root-domain */
6461 raw_spin_lock_irqsave(&rq->lock, flags);
6463 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6467 raw_spin_unlock_irqrestore(&rq->lock, flags);
6470 #ifdef CONFIG_HOTPLUG_CPU
6472 sched_ttwu_pending();
6473 /* Update our root-domain */
6474 raw_spin_lock_irqsave(&rq->lock, flags);
6476 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6480 BUG_ON(rq->nr_running != 1); /* the migration thread */
6481 raw_spin_unlock_irqrestore(&rq->lock, flags);
6483 migrate_nr_uninterruptible(rq);
6484 calc_global_load_remove(rq);
6489 update_max_interval();
6495 * Register at high priority so that task migration (migrate_all_tasks)
6496 * happens before everything else. This has to be lower priority than
6497 * the notifier in the perf_event subsystem, though.
6499 static struct notifier_block __cpuinitdata migration_notifier = {
6500 .notifier_call = migration_call,
6501 .priority = CPU_PRI_MIGRATION,
6504 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6505 unsigned long action, void *hcpu)
6507 switch (action & ~CPU_TASKS_FROZEN) {
6509 case CPU_DOWN_FAILED:
6510 set_cpu_active((long)hcpu, true);
6517 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6518 unsigned long action, void *hcpu)
6520 switch (action & ~CPU_TASKS_FROZEN) {
6521 case CPU_DOWN_PREPARE:
6522 set_cpu_active((long)hcpu, false);
6529 static int __init migration_init(void)
6531 void *cpu = (void *)(long)smp_processor_id();
6534 /* Initialize migration for the boot CPU */
6535 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6536 BUG_ON(err == NOTIFY_BAD);
6537 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6538 register_cpu_notifier(&migration_notifier);
6540 /* Register cpu active notifiers */
6541 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6542 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6546 early_initcall(migration_init);
6551 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6553 #ifdef CONFIG_SCHED_DEBUG
6555 static __read_mostly int sched_domain_debug_enabled;
6557 static int __init sched_domain_debug_setup(char *str)
6559 sched_domain_debug_enabled = 1;
6563 early_param("sched_debug", sched_domain_debug_setup);
6565 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6566 struct cpumask *groupmask)
6568 struct sched_group *group = sd->groups;
6571 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6572 cpumask_clear(groupmask);
6574 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6576 if (!(sd->flags & SD_LOAD_BALANCE)) {
6577 printk("does not load-balance\n");
6579 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6584 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6586 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6587 printk(KERN_ERR "ERROR: domain->span does not contain "
6590 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6591 printk(KERN_ERR "ERROR: domain->groups does not contain"
6595 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6599 printk(KERN_ERR "ERROR: group is NULL\n");
6603 if (!group->cpu_power) {
6604 printk(KERN_CONT "\n");
6605 printk(KERN_ERR "ERROR: domain->cpu_power not "
6610 if (!cpumask_weight(sched_group_cpus(group))) {
6611 printk(KERN_CONT "\n");
6612 printk(KERN_ERR "ERROR: empty group\n");
6616 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6617 printk(KERN_CONT "\n");
6618 printk(KERN_ERR "ERROR: repeated CPUs\n");
6622 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6624 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6626 printk(KERN_CONT " %s", str);
6627 if (group->cpu_power != SCHED_POWER_SCALE) {
6628 printk(KERN_CONT " (cpu_power = %d)",
6632 group = group->next;
6633 } while (group != sd->groups);
6634 printk(KERN_CONT "\n");
6636 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6637 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6640 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6641 printk(KERN_ERR "ERROR: parent span is not a superset "
6642 "of domain->span\n");
6646 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6650 if (!sched_domain_debug_enabled)
6654 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6658 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6661 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6669 #else /* !CONFIG_SCHED_DEBUG */
6670 # define sched_domain_debug(sd, cpu) do { } while (0)
6671 #endif /* CONFIG_SCHED_DEBUG */
6673 static int sd_degenerate(struct sched_domain *sd)
6675 if (cpumask_weight(sched_domain_span(sd)) == 1)
6678 /* Following flags need at least 2 groups */
6679 if (sd->flags & (SD_LOAD_BALANCE |
6680 SD_BALANCE_NEWIDLE |
6684 SD_SHARE_PKG_RESOURCES)) {
6685 if (sd->groups != sd->groups->next)
6689 /* Following flags don't use groups */
6690 if (sd->flags & (SD_WAKE_AFFINE))
6697 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6699 unsigned long cflags = sd->flags, pflags = parent->flags;
6701 if (sd_degenerate(parent))
6704 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6707 /* Flags needing groups don't count if only 1 group in parent */
6708 if (parent->groups == parent->groups->next) {
6709 pflags &= ~(SD_LOAD_BALANCE |
6710 SD_BALANCE_NEWIDLE |
6714 SD_SHARE_PKG_RESOURCES);
6715 if (nr_node_ids == 1)
6716 pflags &= ~SD_SERIALIZE;
6718 if (~cflags & pflags)
6724 static void free_rootdomain(struct rcu_head *rcu)
6726 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6728 cpupri_cleanup(&rd->cpupri);
6729 free_cpumask_var(rd->rto_mask);
6730 free_cpumask_var(rd->online);
6731 free_cpumask_var(rd->span);
6735 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6737 struct root_domain *old_rd = NULL;
6738 unsigned long flags;
6740 raw_spin_lock_irqsave(&rq->lock, flags);
6745 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6748 cpumask_clear_cpu(rq->cpu, old_rd->span);
6751 * If we dont want to free the old_rt yet then
6752 * set old_rd to NULL to skip the freeing later
6755 if (!atomic_dec_and_test(&old_rd->refcount))
6759 atomic_inc(&rd->refcount);
6762 cpumask_set_cpu(rq->cpu, rd->span);
6763 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6766 raw_spin_unlock_irqrestore(&rq->lock, flags);
6769 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6772 static int init_rootdomain(struct root_domain *rd)
6774 memset(rd, 0, sizeof(*rd));
6776 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6778 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6780 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6783 if (cpupri_init(&rd->cpupri) != 0)
6788 free_cpumask_var(rd->rto_mask);
6790 free_cpumask_var(rd->online);
6792 free_cpumask_var(rd->span);
6797 static void init_defrootdomain(void)
6799 init_rootdomain(&def_root_domain);
6801 atomic_set(&def_root_domain.refcount, 1);
6804 static struct root_domain *alloc_rootdomain(void)
6806 struct root_domain *rd;
6808 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6812 if (init_rootdomain(rd) != 0) {
6820 static void free_sched_domain(struct rcu_head *rcu)
6822 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6823 if (atomic_dec_and_test(&sd->groups->ref))
6828 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6830 call_rcu(&sd->rcu, free_sched_domain);
6833 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6835 for (; sd; sd = sd->parent)
6836 destroy_sched_domain(sd, cpu);
6840 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6841 * hold the hotplug lock.
6844 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6846 struct rq *rq = cpu_rq(cpu);
6847 struct sched_domain *tmp;
6849 /* Remove the sched domains which do not contribute to scheduling. */
6850 for (tmp = sd; tmp; ) {
6851 struct sched_domain *parent = tmp->parent;
6855 if (sd_parent_degenerate(tmp, parent)) {
6856 tmp->parent = parent->parent;
6858 parent->parent->child = tmp;
6859 destroy_sched_domain(parent, cpu);
6864 if (sd && sd_degenerate(sd)) {
6867 destroy_sched_domain(tmp, cpu);
6872 sched_domain_debug(sd, cpu);
6874 rq_attach_root(rq, rd);
6876 rcu_assign_pointer(rq->sd, sd);
6877 destroy_sched_domains(tmp, cpu);
6880 /* cpus with isolated domains */
6881 static cpumask_var_t cpu_isolated_map;
6883 /* Setup the mask of cpus configured for isolated domains */
6884 static int __init isolated_cpu_setup(char *str)
6886 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6887 cpulist_parse(str, cpu_isolated_map);
6891 __setup("isolcpus=", isolated_cpu_setup);
6893 #define SD_NODES_PER_DOMAIN 16
6898 * find_next_best_node - find the next node to include in a sched_domain
6899 * @node: node whose sched_domain we're building
6900 * @used_nodes: nodes already in the sched_domain
6902 * Find the next node to include in a given scheduling domain. Simply
6903 * finds the closest node not already in the @used_nodes map.
6905 * Should use nodemask_t.
6907 static int find_next_best_node(int node, nodemask_t *used_nodes)
6909 int i, n, val, min_val, best_node = -1;
6913 for (i = 0; i < nr_node_ids; i++) {
6914 /* Start at @node */
6915 n = (node + i) % nr_node_ids;
6917 if (!nr_cpus_node(n))
6920 /* Skip already used nodes */
6921 if (node_isset(n, *used_nodes))
6924 /* Simple min distance search */
6925 val = node_distance(node, n);
6927 if (val < min_val) {
6933 if (best_node != -1)
6934 node_set(best_node, *used_nodes);
6939 * sched_domain_node_span - get a cpumask for a node's sched_domain
6940 * @node: node whose cpumask we're constructing
6941 * @span: resulting cpumask
6943 * Given a node, construct a good cpumask for its sched_domain to span. It
6944 * should be one that prevents unnecessary balancing, but also spreads tasks
6947 static void sched_domain_node_span(int node, struct cpumask *span)
6949 nodemask_t used_nodes;
6952 cpumask_clear(span);
6953 nodes_clear(used_nodes);
6955 cpumask_or(span, span, cpumask_of_node(node));
6956 node_set(node, used_nodes);
6958 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6959 int next_node = find_next_best_node(node, &used_nodes);
6962 cpumask_or(span, span, cpumask_of_node(next_node));
6966 static const struct cpumask *cpu_node_mask(int cpu)
6968 lockdep_assert_held(&sched_domains_mutex);
6970 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
6972 return sched_domains_tmpmask;
6975 static const struct cpumask *cpu_allnodes_mask(int cpu)
6977 return cpu_possible_mask;
6979 #endif /* CONFIG_NUMA */
6981 static const struct cpumask *cpu_cpu_mask(int cpu)
6983 return cpumask_of_node(cpu_to_node(cpu));
6986 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6989 struct sched_domain **__percpu sd;
6990 struct sched_group **__percpu sg;
6994 struct sched_domain ** __percpu sd;
6995 struct root_domain *rd;
7005 struct sched_domain_topology_level;
7007 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
7008 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
7010 struct sched_domain_topology_level {
7011 sched_domain_init_f init;
7012 sched_domain_mask_f mask;
7013 struct sd_data data;
7017 * Assumes the sched_domain tree is fully constructed
7019 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
7021 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
7022 struct sched_domain *child = sd->child;
7025 cpu = cpumask_first(sched_domain_span(child));
7028 *sg = *per_cpu_ptr(sdd->sg, cpu);
7034 * build_sched_groups takes the cpumask we wish to span, and a pointer
7035 * to a function which identifies what group(along with sched group) a CPU
7036 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7037 * (due to the fact that we keep track of groups covered with a struct cpumask).
7039 * build_sched_groups will build a circular linked list of the groups
7040 * covered by the given span, and will set each group's ->cpumask correctly,
7041 * and ->cpu_power to 0.
7044 build_sched_groups(struct sched_domain *sd)
7046 struct sched_group *first = NULL, *last = NULL;
7047 struct sd_data *sdd = sd->private;
7048 const struct cpumask *span = sched_domain_span(sd);
7049 struct cpumask *covered;
7052 lockdep_assert_held(&sched_domains_mutex);
7053 covered = sched_domains_tmpmask;
7055 cpumask_clear(covered);
7057 for_each_cpu(i, span) {
7058 struct sched_group *sg;
7059 int group = get_group(i, sdd, &sg);
7062 if (cpumask_test_cpu(i, covered))
7065 cpumask_clear(sched_group_cpus(sg));
7068 for_each_cpu(j, span) {
7069 if (get_group(j, sdd, NULL) != group)
7072 cpumask_set_cpu(j, covered);
7073 cpumask_set_cpu(j, sched_group_cpus(sg));
7086 * Initialize sched groups cpu_power.
7088 * cpu_power indicates the capacity of sched group, which is used while
7089 * distributing the load between different sched groups in a sched domain.
7090 * Typically cpu_power for all the groups in a sched domain will be same unless
7091 * there are asymmetries in the topology. If there are asymmetries, group
7092 * having more cpu_power will pickup more load compared to the group having
7095 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7097 WARN_ON(!sd || !sd->groups);
7099 if (cpu != group_first_cpu(sd->groups))
7102 sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups));
7104 update_group_power(sd, cpu);
7108 * Initializers for schedule domains
7109 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7112 #ifdef CONFIG_SCHED_DEBUG
7113 # define SD_INIT_NAME(sd, type) sd->name = #type
7115 # define SD_INIT_NAME(sd, type) do { } while (0)
7118 #define SD_INIT_FUNC(type) \
7119 static noinline struct sched_domain * \
7120 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7122 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7123 *sd = SD_##type##_INIT; \
7124 SD_INIT_NAME(sd, type); \
7125 sd->private = &tl->data; \
7131 SD_INIT_FUNC(ALLNODES)
7134 #ifdef CONFIG_SCHED_SMT
7135 SD_INIT_FUNC(SIBLING)
7137 #ifdef CONFIG_SCHED_MC
7140 #ifdef CONFIG_SCHED_BOOK
7144 static int default_relax_domain_level = -1;
7145 int sched_domain_level_max;
7147 static int __init setup_relax_domain_level(char *str)
7151 val = simple_strtoul(str, NULL, 0);
7152 if (val < sched_domain_level_max)
7153 default_relax_domain_level = val;
7157 __setup("relax_domain_level=", setup_relax_domain_level);
7159 static void set_domain_attribute(struct sched_domain *sd,
7160 struct sched_domain_attr *attr)
7164 if (!attr || attr->relax_domain_level < 0) {
7165 if (default_relax_domain_level < 0)
7168 request = default_relax_domain_level;
7170 request = attr->relax_domain_level;
7171 if (request < sd->level) {
7172 /* turn off idle balance on this domain */
7173 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7175 /* turn on idle balance on this domain */
7176 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7180 static void __sdt_free(const struct cpumask *cpu_map);
7181 static int __sdt_alloc(const struct cpumask *cpu_map);
7183 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7184 const struct cpumask *cpu_map)
7188 if (!atomic_read(&d->rd->refcount))
7189 free_rootdomain(&d->rd->rcu); /* fall through */
7191 free_percpu(d->sd); /* fall through */
7193 __sdt_free(cpu_map); /* fall through */
7199 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7200 const struct cpumask *cpu_map)
7202 memset(d, 0, sizeof(*d));
7204 if (__sdt_alloc(cpu_map))
7205 return sa_sd_storage;
7206 d->sd = alloc_percpu(struct sched_domain *);
7208 return sa_sd_storage;
7209 d->rd = alloc_rootdomain();
7212 return sa_rootdomain;
7216 * NULL the sd_data elements we've used to build the sched_domain and
7217 * sched_group structure so that the subsequent __free_domain_allocs()
7218 * will not free the data we're using.
7220 static void claim_allocations(int cpu, struct sched_domain *sd)
7222 struct sd_data *sdd = sd->private;
7223 struct sched_group *sg = sd->groups;
7225 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7226 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7228 if (cpu == cpumask_first(sched_group_cpus(sg))) {
7229 WARN_ON_ONCE(*per_cpu_ptr(sdd->sg, cpu) != sg);
7230 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7234 #ifdef CONFIG_SCHED_SMT
7235 static const struct cpumask *cpu_smt_mask(int cpu)
7237 return topology_thread_cpumask(cpu);
7242 * Topology list, bottom-up.
7244 static struct sched_domain_topology_level default_topology[] = {
7245 #ifdef CONFIG_SCHED_SMT
7246 { sd_init_SIBLING, cpu_smt_mask, },
7248 #ifdef CONFIG_SCHED_MC
7249 { sd_init_MC, cpu_coregroup_mask, },
7251 #ifdef CONFIG_SCHED_BOOK
7252 { sd_init_BOOK, cpu_book_mask, },
7254 { sd_init_CPU, cpu_cpu_mask, },
7256 { sd_init_NODE, cpu_node_mask, },
7257 { sd_init_ALLNODES, cpu_allnodes_mask, },
7262 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7264 static int __sdt_alloc(const struct cpumask *cpu_map)
7266 struct sched_domain_topology_level *tl;
7269 for (tl = sched_domain_topology; tl->init; tl++) {
7270 struct sd_data *sdd = &tl->data;
7272 sdd->sd = alloc_percpu(struct sched_domain *);
7276 sdd->sg = alloc_percpu(struct sched_group *);
7280 for_each_cpu(j, cpu_map) {
7281 struct sched_domain *sd;
7282 struct sched_group *sg;
7284 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7285 GFP_KERNEL, cpu_to_node(j));
7289 *per_cpu_ptr(sdd->sd, j) = sd;
7291 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7292 GFP_KERNEL, cpu_to_node(j));
7296 *per_cpu_ptr(sdd->sg, j) = sg;
7303 static void __sdt_free(const struct cpumask *cpu_map)
7305 struct sched_domain_topology_level *tl;
7308 for (tl = sched_domain_topology; tl->init; tl++) {
7309 struct sd_data *sdd = &tl->data;
7311 for_each_cpu(j, cpu_map) {
7312 kfree(*per_cpu_ptr(sdd->sd, j));
7313 kfree(*per_cpu_ptr(sdd->sg, j));
7315 free_percpu(sdd->sd);
7316 free_percpu(sdd->sg);
7320 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7321 struct s_data *d, const struct cpumask *cpu_map,
7322 struct sched_domain_attr *attr, struct sched_domain *child,
7325 struct sched_domain *sd = tl->init(tl, cpu);
7329 set_domain_attribute(sd, attr);
7330 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7332 sd->level = child->level + 1;
7333 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7342 * Build sched domains for a given set of cpus and attach the sched domains
7343 * to the individual cpus
7345 static int build_sched_domains(const struct cpumask *cpu_map,
7346 struct sched_domain_attr *attr)
7348 enum s_alloc alloc_state = sa_none;
7349 struct sched_domain *sd;
7351 int i, ret = -ENOMEM;
7353 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7354 if (alloc_state != sa_rootdomain)
7357 /* Set up domains for cpus specified by the cpu_map. */
7358 for_each_cpu(i, cpu_map) {
7359 struct sched_domain_topology_level *tl;
7362 for (tl = sched_domain_topology; tl->init; tl++)
7363 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7368 *per_cpu_ptr(d.sd, i) = sd;
7371 /* Build the groups for the domains */
7372 for_each_cpu(i, cpu_map) {
7373 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7374 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7375 get_group(i, sd->private, &sd->groups);
7376 atomic_inc(&sd->groups->ref);
7378 if (i != cpumask_first(sched_domain_span(sd)))
7381 build_sched_groups(sd);
7385 /* Calculate CPU power for physical packages and nodes */
7386 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7387 if (!cpumask_test_cpu(i, cpu_map))
7390 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7391 claim_allocations(i, sd);
7392 init_sched_groups_power(i, sd);
7396 /* Attach the domains */
7398 for_each_cpu(i, cpu_map) {
7399 sd = *per_cpu_ptr(d.sd, i);
7400 cpu_attach_domain(sd, d.rd, i);
7406 __free_domain_allocs(&d, alloc_state, cpu_map);
7410 static cpumask_var_t *doms_cur; /* current sched domains */
7411 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7412 static struct sched_domain_attr *dattr_cur;
7413 /* attribues of custom domains in 'doms_cur' */
7416 * Special case: If a kmalloc of a doms_cur partition (array of
7417 * cpumask) fails, then fallback to a single sched domain,
7418 * as determined by the single cpumask fallback_doms.
7420 static cpumask_var_t fallback_doms;
7423 * arch_update_cpu_topology lets virtualized architectures update the
7424 * cpu core maps. It is supposed to return 1 if the topology changed
7425 * or 0 if it stayed the same.
7427 int __attribute__((weak)) arch_update_cpu_topology(void)
7432 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7435 cpumask_var_t *doms;
7437 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7440 for (i = 0; i < ndoms; i++) {
7441 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7442 free_sched_domains(doms, i);
7449 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7452 for (i = 0; i < ndoms; i++)
7453 free_cpumask_var(doms[i]);
7458 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7459 * For now this just excludes isolated cpus, but could be used to
7460 * exclude other special cases in the future.
7462 static int init_sched_domains(const struct cpumask *cpu_map)
7466 arch_update_cpu_topology();
7468 doms_cur = alloc_sched_domains(ndoms_cur);
7470 doms_cur = &fallback_doms;
7471 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7473 err = build_sched_domains(doms_cur[0], NULL);
7474 register_sched_domain_sysctl();
7480 * Detach sched domains from a group of cpus specified in cpu_map
7481 * These cpus will now be attached to the NULL domain
7483 static void detach_destroy_domains(const struct cpumask *cpu_map)
7488 for_each_cpu(i, cpu_map)
7489 cpu_attach_domain(NULL, &def_root_domain, i);
7493 /* handle null as "default" */
7494 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7495 struct sched_domain_attr *new, int idx_new)
7497 struct sched_domain_attr tmp;
7504 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7505 new ? (new + idx_new) : &tmp,
7506 sizeof(struct sched_domain_attr));
7510 * Partition sched domains as specified by the 'ndoms_new'
7511 * cpumasks in the array doms_new[] of cpumasks. This compares
7512 * doms_new[] to the current sched domain partitioning, doms_cur[].
7513 * It destroys each deleted domain and builds each new domain.
7515 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7516 * The masks don't intersect (don't overlap.) We should setup one
7517 * sched domain for each mask. CPUs not in any of the cpumasks will
7518 * not be load balanced. If the same cpumask appears both in the
7519 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7522 * The passed in 'doms_new' should be allocated using
7523 * alloc_sched_domains. This routine takes ownership of it and will
7524 * free_sched_domains it when done with it. If the caller failed the
7525 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7526 * and partition_sched_domains() will fallback to the single partition
7527 * 'fallback_doms', it also forces the domains to be rebuilt.
7529 * If doms_new == NULL it will be replaced with cpu_online_mask.
7530 * ndoms_new == 0 is a special case for destroying existing domains,
7531 * and it will not create the default domain.
7533 * Call with hotplug lock held
7535 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7536 struct sched_domain_attr *dattr_new)
7541 mutex_lock(&sched_domains_mutex);
7543 /* always unregister in case we don't destroy any domains */
7544 unregister_sched_domain_sysctl();
7546 /* Let architecture update cpu core mappings. */
7547 new_topology = arch_update_cpu_topology();
7549 n = doms_new ? ndoms_new : 0;
7551 /* Destroy deleted domains */
7552 for (i = 0; i < ndoms_cur; i++) {
7553 for (j = 0; j < n && !new_topology; j++) {
7554 if (cpumask_equal(doms_cur[i], doms_new[j])
7555 && dattrs_equal(dattr_cur, i, dattr_new, j))
7558 /* no match - a current sched domain not in new doms_new[] */
7559 detach_destroy_domains(doms_cur[i]);
7564 if (doms_new == NULL) {
7566 doms_new = &fallback_doms;
7567 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7568 WARN_ON_ONCE(dattr_new);
7571 /* Build new domains */
7572 for (i = 0; i < ndoms_new; i++) {
7573 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7574 if (cpumask_equal(doms_new[i], doms_cur[j])
7575 && dattrs_equal(dattr_new, i, dattr_cur, j))
7578 /* no match - add a new doms_new */
7579 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7584 /* Remember the new sched domains */
7585 if (doms_cur != &fallback_doms)
7586 free_sched_domains(doms_cur, ndoms_cur);
7587 kfree(dattr_cur); /* kfree(NULL) is safe */
7588 doms_cur = doms_new;
7589 dattr_cur = dattr_new;
7590 ndoms_cur = ndoms_new;
7592 register_sched_domain_sysctl();
7594 mutex_unlock(&sched_domains_mutex);
7597 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7598 static void reinit_sched_domains(void)
7602 /* Destroy domains first to force the rebuild */
7603 partition_sched_domains(0, NULL, NULL);
7605 rebuild_sched_domains();
7609 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7611 unsigned int level = 0;
7613 if (sscanf(buf, "%u", &level) != 1)
7617 * level is always be positive so don't check for
7618 * level < POWERSAVINGS_BALANCE_NONE which is 0
7619 * What happens on 0 or 1 byte write,
7620 * need to check for count as well?
7623 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7627 sched_smt_power_savings = level;
7629 sched_mc_power_savings = level;
7631 reinit_sched_domains();
7636 #ifdef CONFIG_SCHED_MC
7637 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7638 struct sysdev_class_attribute *attr,
7641 return sprintf(page, "%u\n", sched_mc_power_savings);
7643 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7644 struct sysdev_class_attribute *attr,
7645 const char *buf, size_t count)
7647 return sched_power_savings_store(buf, count, 0);
7649 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7650 sched_mc_power_savings_show,
7651 sched_mc_power_savings_store);
7654 #ifdef CONFIG_SCHED_SMT
7655 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7656 struct sysdev_class_attribute *attr,
7659 return sprintf(page, "%u\n", sched_smt_power_savings);
7661 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7662 struct sysdev_class_attribute *attr,
7663 const char *buf, size_t count)
7665 return sched_power_savings_store(buf, count, 1);
7667 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7668 sched_smt_power_savings_show,
7669 sched_smt_power_savings_store);
7672 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7676 #ifdef CONFIG_SCHED_SMT
7678 err = sysfs_create_file(&cls->kset.kobj,
7679 &attr_sched_smt_power_savings.attr);
7681 #ifdef CONFIG_SCHED_MC
7682 if (!err && mc_capable())
7683 err = sysfs_create_file(&cls->kset.kobj,
7684 &attr_sched_mc_power_savings.attr);
7688 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7691 * Update cpusets according to cpu_active mask. If cpusets are
7692 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7693 * around partition_sched_domains().
7695 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7698 switch (action & ~CPU_TASKS_FROZEN) {
7700 case CPU_DOWN_FAILED:
7701 cpuset_update_active_cpus();
7708 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7711 switch (action & ~CPU_TASKS_FROZEN) {
7712 case CPU_DOWN_PREPARE:
7713 cpuset_update_active_cpus();
7720 static int update_runtime(struct notifier_block *nfb,
7721 unsigned long action, void *hcpu)
7723 int cpu = (int)(long)hcpu;
7726 case CPU_DOWN_PREPARE:
7727 case CPU_DOWN_PREPARE_FROZEN:
7728 disable_runtime(cpu_rq(cpu));
7731 case CPU_DOWN_FAILED:
7732 case CPU_DOWN_FAILED_FROZEN:
7734 case CPU_ONLINE_FROZEN:
7735 enable_runtime(cpu_rq(cpu));
7743 void __init sched_init_smp(void)
7745 cpumask_var_t non_isolated_cpus;
7747 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7748 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7751 mutex_lock(&sched_domains_mutex);
7752 init_sched_domains(cpu_active_mask);
7753 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7754 if (cpumask_empty(non_isolated_cpus))
7755 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7756 mutex_unlock(&sched_domains_mutex);
7759 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7760 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7762 /* RT runtime code needs to handle some hotplug events */
7763 hotcpu_notifier(update_runtime, 0);
7767 /* Move init over to a non-isolated CPU */
7768 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7770 sched_init_granularity();
7771 free_cpumask_var(non_isolated_cpus);
7773 init_sched_rt_class();
7776 void __init sched_init_smp(void)
7778 sched_init_granularity();
7780 #endif /* CONFIG_SMP */
7782 const_debug unsigned int sysctl_timer_migration = 1;
7784 int in_sched_functions(unsigned long addr)
7786 return in_lock_functions(addr) ||
7787 (addr >= (unsigned long)__sched_text_start
7788 && addr < (unsigned long)__sched_text_end);
7791 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7793 cfs_rq->tasks_timeline = RB_ROOT;
7794 INIT_LIST_HEAD(&cfs_rq->tasks);
7795 #ifdef CONFIG_FAIR_GROUP_SCHED
7797 /* allow initial update_cfs_load() to truncate */
7799 cfs_rq->load_stamp = 1;
7802 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7805 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7807 struct rt_prio_array *array;
7810 array = &rt_rq->active;
7811 for (i = 0; i < MAX_RT_PRIO; i++) {
7812 INIT_LIST_HEAD(array->queue + i);
7813 __clear_bit(i, array->bitmap);
7815 /* delimiter for bitsearch: */
7816 __set_bit(MAX_RT_PRIO, array->bitmap);
7818 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7819 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7821 rt_rq->highest_prio.next = MAX_RT_PRIO;
7825 rt_rq->rt_nr_migratory = 0;
7826 rt_rq->overloaded = 0;
7827 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7831 rt_rq->rt_throttled = 0;
7832 rt_rq->rt_runtime = 0;
7833 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7835 #ifdef CONFIG_RT_GROUP_SCHED
7836 rt_rq->rt_nr_boosted = 0;
7841 #ifdef CONFIG_FAIR_GROUP_SCHED
7842 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7843 struct sched_entity *se, int cpu,
7844 struct sched_entity *parent)
7846 struct rq *rq = cpu_rq(cpu);
7847 tg->cfs_rq[cpu] = cfs_rq;
7848 init_cfs_rq(cfs_rq, rq);
7852 /* se could be NULL for root_task_group */
7857 se->cfs_rq = &rq->cfs;
7859 se->cfs_rq = parent->my_q;
7862 update_load_set(&se->load, 0);
7863 se->parent = parent;
7867 #ifdef CONFIG_RT_GROUP_SCHED
7868 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7869 struct sched_rt_entity *rt_se, int cpu,
7870 struct sched_rt_entity *parent)
7872 struct rq *rq = cpu_rq(cpu);
7874 tg->rt_rq[cpu] = rt_rq;
7875 init_rt_rq(rt_rq, rq);
7877 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7879 tg->rt_se[cpu] = rt_se;
7884 rt_se->rt_rq = &rq->rt;
7886 rt_se->rt_rq = parent->my_q;
7888 rt_se->my_q = rt_rq;
7889 rt_se->parent = parent;
7890 INIT_LIST_HEAD(&rt_se->run_list);
7894 void __init sched_init(void)
7897 unsigned long alloc_size = 0, ptr;
7899 #ifdef CONFIG_FAIR_GROUP_SCHED
7900 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7902 #ifdef CONFIG_RT_GROUP_SCHED
7903 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7905 #ifdef CONFIG_CPUMASK_OFFSTACK
7906 alloc_size += num_possible_cpus() * cpumask_size();
7909 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7911 #ifdef CONFIG_FAIR_GROUP_SCHED
7912 root_task_group.se = (struct sched_entity **)ptr;
7913 ptr += nr_cpu_ids * sizeof(void **);
7915 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7916 ptr += nr_cpu_ids * sizeof(void **);
7918 #endif /* CONFIG_FAIR_GROUP_SCHED */
7919 #ifdef CONFIG_RT_GROUP_SCHED
7920 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7921 ptr += nr_cpu_ids * sizeof(void **);
7923 root_task_group.rt_rq = (struct rt_rq **)ptr;
7924 ptr += nr_cpu_ids * sizeof(void **);
7926 #endif /* CONFIG_RT_GROUP_SCHED */
7927 #ifdef CONFIG_CPUMASK_OFFSTACK
7928 for_each_possible_cpu(i) {
7929 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7930 ptr += cpumask_size();
7932 #endif /* CONFIG_CPUMASK_OFFSTACK */
7936 init_defrootdomain();
7939 init_rt_bandwidth(&def_rt_bandwidth,
7940 global_rt_period(), global_rt_runtime());
7942 #ifdef CONFIG_RT_GROUP_SCHED
7943 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7944 global_rt_period(), global_rt_runtime());
7945 #endif /* CONFIG_RT_GROUP_SCHED */
7947 #ifdef CONFIG_CGROUP_SCHED
7948 list_add(&root_task_group.list, &task_groups);
7949 INIT_LIST_HEAD(&root_task_group.children);
7950 autogroup_init(&init_task);
7951 #endif /* CONFIG_CGROUP_SCHED */
7953 for_each_possible_cpu(i) {
7957 raw_spin_lock_init(&rq->lock);
7959 rq->calc_load_active = 0;
7960 rq->calc_load_update = jiffies + LOAD_FREQ;
7961 init_cfs_rq(&rq->cfs, rq);
7962 init_rt_rq(&rq->rt, rq);
7963 #ifdef CONFIG_FAIR_GROUP_SCHED
7964 root_task_group.shares = root_task_group_load;
7965 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7967 * How much cpu bandwidth does root_task_group get?
7969 * In case of task-groups formed thr' the cgroup filesystem, it
7970 * gets 100% of the cpu resources in the system. This overall
7971 * system cpu resource is divided among the tasks of
7972 * root_task_group and its child task-groups in a fair manner,
7973 * based on each entity's (task or task-group's) weight
7974 * (se->load.weight).
7976 * In other words, if root_task_group has 10 tasks of weight
7977 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7978 * then A0's share of the cpu resource is:
7980 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7982 * We achieve this by letting root_task_group's tasks sit
7983 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7985 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7986 #endif /* CONFIG_FAIR_GROUP_SCHED */
7988 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7989 #ifdef CONFIG_RT_GROUP_SCHED
7990 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7991 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7994 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7995 rq->cpu_load[j] = 0;
7997 rq->last_load_update_tick = jiffies;
8002 rq->cpu_power = SCHED_POWER_SCALE;
8003 rq->post_schedule = 0;
8004 rq->active_balance = 0;
8005 rq->next_balance = jiffies;
8010 rq->avg_idle = 2*sysctl_sched_migration_cost;
8011 rq_attach_root(rq, &def_root_domain);
8013 rq->nohz_balance_kick = 0;
8014 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8018 atomic_set(&rq->nr_iowait, 0);
8021 set_load_weight(&init_task);
8023 #ifdef CONFIG_PREEMPT_NOTIFIERS
8024 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8028 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8031 #ifdef CONFIG_RT_MUTEXES
8032 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
8036 * The boot idle thread does lazy MMU switching as well:
8038 atomic_inc(&init_mm.mm_count);
8039 enter_lazy_tlb(&init_mm, current);
8042 * Make us the idle thread. Technically, schedule() should not be
8043 * called from this thread, however somewhere below it might be,
8044 * but because we are the idle thread, we just pick up running again
8045 * when this runqueue becomes "idle".
8047 init_idle(current, smp_processor_id());
8049 calc_load_update = jiffies + LOAD_FREQ;
8052 * During early bootup we pretend to be a normal task:
8054 current->sched_class = &fair_sched_class;
8056 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8057 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8059 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8061 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8062 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8063 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8064 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8065 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8067 /* May be allocated at isolcpus cmdline parse time */
8068 if (cpu_isolated_map == NULL)
8069 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8072 scheduler_running = 1;
8075 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8076 static inline int preempt_count_equals(int preempt_offset)
8078 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8080 return (nested == preempt_offset);
8083 void __might_sleep(const char *file, int line, int preempt_offset)
8086 static unsigned long prev_jiffy; /* ratelimiting */
8088 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8089 system_state != SYSTEM_RUNNING || oops_in_progress)
8091 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8093 prev_jiffy = jiffies;
8096 "BUG: sleeping function called from invalid context at %s:%d\n",
8099 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8100 in_atomic(), irqs_disabled(),
8101 current->pid, current->comm);
8103 debug_show_held_locks(current);
8104 if (irqs_disabled())
8105 print_irqtrace_events(current);
8109 EXPORT_SYMBOL(__might_sleep);
8112 #ifdef CONFIG_MAGIC_SYSRQ
8113 static void normalize_task(struct rq *rq, struct task_struct *p)
8115 const struct sched_class *prev_class = p->sched_class;
8116 int old_prio = p->prio;
8121 deactivate_task(rq, p, 0);
8122 __setscheduler(rq, p, SCHED_NORMAL, 0);
8124 activate_task(rq, p, 0);
8125 resched_task(rq->curr);
8128 check_class_changed(rq, p, prev_class, old_prio);
8131 void normalize_rt_tasks(void)
8133 struct task_struct *g, *p;
8134 unsigned long flags;
8137 read_lock_irqsave(&tasklist_lock, flags);
8138 do_each_thread(g, p) {
8140 * Only normalize user tasks:
8145 p->se.exec_start = 0;
8146 #ifdef CONFIG_SCHEDSTATS
8147 p->se.statistics.wait_start = 0;
8148 p->se.statistics.sleep_start = 0;
8149 p->se.statistics.block_start = 0;
8154 * Renice negative nice level userspace
8157 if (TASK_NICE(p) < 0 && p->mm)
8158 set_user_nice(p, 0);
8162 raw_spin_lock(&p->pi_lock);
8163 rq = __task_rq_lock(p);
8165 normalize_task(rq, p);
8167 __task_rq_unlock(rq);
8168 raw_spin_unlock(&p->pi_lock);
8169 } while_each_thread(g, p);
8171 read_unlock_irqrestore(&tasklist_lock, flags);
8174 #endif /* CONFIG_MAGIC_SYSRQ */
8176 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8178 * These functions are only useful for the IA64 MCA handling, or kdb.
8180 * They can only be called when the whole system has been
8181 * stopped - every CPU needs to be quiescent, and no scheduling
8182 * activity can take place. Using them for anything else would
8183 * be a serious bug, and as a result, they aren't even visible
8184 * under any other configuration.
8188 * curr_task - return the current task for a given cpu.
8189 * @cpu: the processor in question.
8191 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8193 struct task_struct *curr_task(int cpu)
8195 return cpu_curr(cpu);
8198 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8202 * set_curr_task - set the current task for a given cpu.
8203 * @cpu: the processor in question.
8204 * @p: the task pointer to set.
8206 * Description: This function must only be used when non-maskable interrupts
8207 * are serviced on a separate stack. It allows the architecture to switch the
8208 * notion of the current task on a cpu in a non-blocking manner. This function
8209 * must be called with all CPU's synchronized, and interrupts disabled, the
8210 * and caller must save the original value of the current task (see
8211 * curr_task() above) and restore that value before reenabling interrupts and
8212 * re-starting the system.
8214 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8216 void set_curr_task(int cpu, struct task_struct *p)
8223 #ifdef CONFIG_FAIR_GROUP_SCHED
8224 static void free_fair_sched_group(struct task_group *tg)
8228 for_each_possible_cpu(i) {
8230 kfree(tg->cfs_rq[i]);
8240 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8242 struct cfs_rq *cfs_rq;
8243 struct sched_entity *se;
8246 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8249 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8253 tg->shares = NICE_0_LOAD;
8255 for_each_possible_cpu(i) {
8256 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8257 GFP_KERNEL, cpu_to_node(i));
8261 se = kzalloc_node(sizeof(struct sched_entity),
8262 GFP_KERNEL, cpu_to_node(i));
8266 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8277 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8279 struct rq *rq = cpu_rq(cpu);
8280 unsigned long flags;
8283 * Only empty task groups can be destroyed; so we can speculatively
8284 * check on_list without danger of it being re-added.
8286 if (!tg->cfs_rq[cpu]->on_list)
8289 raw_spin_lock_irqsave(&rq->lock, flags);
8290 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8291 raw_spin_unlock_irqrestore(&rq->lock, flags);
8293 #else /* !CONFG_FAIR_GROUP_SCHED */
8294 static inline void free_fair_sched_group(struct task_group *tg)
8299 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8304 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8307 #endif /* CONFIG_FAIR_GROUP_SCHED */
8309 #ifdef CONFIG_RT_GROUP_SCHED
8310 static void free_rt_sched_group(struct task_group *tg)
8314 destroy_rt_bandwidth(&tg->rt_bandwidth);
8316 for_each_possible_cpu(i) {
8318 kfree(tg->rt_rq[i]);
8320 kfree(tg->rt_se[i]);
8328 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8330 struct rt_rq *rt_rq;
8331 struct sched_rt_entity *rt_se;
8334 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8337 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8341 init_rt_bandwidth(&tg->rt_bandwidth,
8342 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8344 for_each_possible_cpu(i) {
8345 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8346 GFP_KERNEL, cpu_to_node(i));
8350 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8351 GFP_KERNEL, cpu_to_node(i));
8355 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8365 #else /* !CONFIG_RT_GROUP_SCHED */
8366 static inline void free_rt_sched_group(struct task_group *tg)
8371 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8375 #endif /* CONFIG_RT_GROUP_SCHED */
8377 #ifdef CONFIG_CGROUP_SCHED
8378 static void free_sched_group(struct task_group *tg)
8380 free_fair_sched_group(tg);
8381 free_rt_sched_group(tg);
8386 /* allocate runqueue etc for a new task group */
8387 struct task_group *sched_create_group(struct task_group *parent)
8389 struct task_group *tg;
8390 unsigned long flags;
8392 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8394 return ERR_PTR(-ENOMEM);
8396 if (!alloc_fair_sched_group(tg, parent))
8399 if (!alloc_rt_sched_group(tg, parent))
8402 spin_lock_irqsave(&task_group_lock, flags);
8403 list_add_rcu(&tg->list, &task_groups);
8405 WARN_ON(!parent); /* root should already exist */
8407 tg->parent = parent;
8408 INIT_LIST_HEAD(&tg->children);
8409 list_add_rcu(&tg->siblings, &parent->children);
8410 spin_unlock_irqrestore(&task_group_lock, flags);
8415 free_sched_group(tg);
8416 return ERR_PTR(-ENOMEM);
8419 /* rcu callback to free various structures associated with a task group */
8420 static void free_sched_group_rcu(struct rcu_head *rhp)
8422 /* now it should be safe to free those cfs_rqs */
8423 free_sched_group(container_of(rhp, struct task_group, rcu));
8426 /* Destroy runqueue etc associated with a task group */
8427 void sched_destroy_group(struct task_group *tg)
8429 unsigned long flags;
8432 /* end participation in shares distribution */
8433 for_each_possible_cpu(i)
8434 unregister_fair_sched_group(tg, i);
8436 spin_lock_irqsave(&task_group_lock, flags);
8437 list_del_rcu(&tg->list);
8438 list_del_rcu(&tg->siblings);
8439 spin_unlock_irqrestore(&task_group_lock, flags);
8441 /* wait for possible concurrent references to cfs_rqs complete */
8442 call_rcu(&tg->rcu, free_sched_group_rcu);
8445 /* change task's runqueue when it moves between groups.
8446 * The caller of this function should have put the task in its new group
8447 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8448 * reflect its new group.
8450 void sched_move_task(struct task_struct *tsk)
8453 unsigned long flags;
8456 rq = task_rq_lock(tsk, &flags);
8458 running = task_current(rq, tsk);
8462 dequeue_task(rq, tsk, 0);
8463 if (unlikely(running))
8464 tsk->sched_class->put_prev_task(rq, tsk);
8466 #ifdef CONFIG_FAIR_GROUP_SCHED
8467 if (tsk->sched_class->task_move_group)
8468 tsk->sched_class->task_move_group(tsk, on_rq);
8471 set_task_rq(tsk, task_cpu(tsk));
8473 if (unlikely(running))
8474 tsk->sched_class->set_curr_task(rq);
8476 enqueue_task(rq, tsk, 0);
8478 task_rq_unlock(rq, tsk, &flags);
8480 #endif /* CONFIG_CGROUP_SCHED */
8482 #ifdef CONFIG_FAIR_GROUP_SCHED
8483 static DEFINE_MUTEX(shares_mutex);
8485 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8488 unsigned long flags;
8491 * We can't change the weight of the root cgroup.
8496 if (shares < MIN_SHARES)
8497 shares = MIN_SHARES;
8498 else if (shares > MAX_SHARES)
8499 shares = MAX_SHARES;
8501 mutex_lock(&shares_mutex);
8502 if (tg->shares == shares)
8505 tg->shares = shares;
8506 for_each_possible_cpu(i) {
8507 struct rq *rq = cpu_rq(i);
8508 struct sched_entity *se;
8511 /* Propagate contribution to hierarchy */
8512 raw_spin_lock_irqsave(&rq->lock, flags);
8513 for_each_sched_entity(se)
8514 update_cfs_shares(group_cfs_rq(se));
8515 raw_spin_unlock_irqrestore(&rq->lock, flags);
8519 mutex_unlock(&shares_mutex);
8523 unsigned long sched_group_shares(struct task_group *tg)
8529 #ifdef CONFIG_RT_GROUP_SCHED
8531 * Ensure that the real time constraints are schedulable.
8533 static DEFINE_MUTEX(rt_constraints_mutex);
8535 static unsigned long to_ratio(u64 period, u64 runtime)
8537 if (runtime == RUNTIME_INF)
8540 return div64_u64(runtime << 20, period);
8543 /* Must be called with tasklist_lock held */
8544 static inline int tg_has_rt_tasks(struct task_group *tg)
8546 struct task_struct *g, *p;
8548 do_each_thread(g, p) {
8549 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8551 } while_each_thread(g, p);
8556 struct rt_schedulable_data {
8557 struct task_group *tg;
8562 static int tg_schedulable(struct task_group *tg, void *data)
8564 struct rt_schedulable_data *d = data;
8565 struct task_group *child;
8566 unsigned long total, sum = 0;
8567 u64 period, runtime;
8569 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8570 runtime = tg->rt_bandwidth.rt_runtime;
8573 period = d->rt_period;
8574 runtime = d->rt_runtime;
8578 * Cannot have more runtime than the period.
8580 if (runtime > period && runtime != RUNTIME_INF)
8584 * Ensure we don't starve existing RT tasks.
8586 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8589 total = to_ratio(period, runtime);
8592 * Nobody can have more than the global setting allows.
8594 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8598 * The sum of our children's runtime should not exceed our own.
8600 list_for_each_entry_rcu(child, &tg->children, siblings) {
8601 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8602 runtime = child->rt_bandwidth.rt_runtime;
8604 if (child == d->tg) {
8605 period = d->rt_period;
8606 runtime = d->rt_runtime;
8609 sum += to_ratio(period, runtime);
8618 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8620 struct rt_schedulable_data data = {
8622 .rt_period = period,
8623 .rt_runtime = runtime,
8626 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8629 static int tg_set_bandwidth(struct task_group *tg,
8630 u64 rt_period, u64 rt_runtime)
8634 mutex_lock(&rt_constraints_mutex);
8635 read_lock(&tasklist_lock);
8636 err = __rt_schedulable(tg, rt_period, rt_runtime);
8640 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8641 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8642 tg->rt_bandwidth.rt_runtime = rt_runtime;
8644 for_each_possible_cpu(i) {
8645 struct rt_rq *rt_rq = tg->rt_rq[i];
8647 raw_spin_lock(&rt_rq->rt_runtime_lock);
8648 rt_rq->rt_runtime = rt_runtime;
8649 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8651 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8653 read_unlock(&tasklist_lock);
8654 mutex_unlock(&rt_constraints_mutex);
8659 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8661 u64 rt_runtime, rt_period;
8663 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8664 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8665 if (rt_runtime_us < 0)
8666 rt_runtime = RUNTIME_INF;
8668 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8671 long sched_group_rt_runtime(struct task_group *tg)
8675 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8678 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8679 do_div(rt_runtime_us, NSEC_PER_USEC);
8680 return rt_runtime_us;
8683 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8685 u64 rt_runtime, rt_period;
8687 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8688 rt_runtime = tg->rt_bandwidth.rt_runtime;
8693 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8696 long sched_group_rt_period(struct task_group *tg)
8700 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8701 do_div(rt_period_us, NSEC_PER_USEC);
8702 return rt_period_us;
8705 static int sched_rt_global_constraints(void)
8707 u64 runtime, period;
8710 if (sysctl_sched_rt_period <= 0)
8713 runtime = global_rt_runtime();
8714 period = global_rt_period();
8717 * Sanity check on the sysctl variables.
8719 if (runtime > period && runtime != RUNTIME_INF)
8722 mutex_lock(&rt_constraints_mutex);
8723 read_lock(&tasklist_lock);
8724 ret = __rt_schedulable(NULL, 0, 0);
8725 read_unlock(&tasklist_lock);
8726 mutex_unlock(&rt_constraints_mutex);
8731 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8733 /* Don't accept realtime tasks when there is no way for them to run */
8734 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8740 #else /* !CONFIG_RT_GROUP_SCHED */
8741 static int sched_rt_global_constraints(void)
8743 unsigned long flags;
8746 if (sysctl_sched_rt_period <= 0)
8750 * There's always some RT tasks in the root group
8751 * -- migration, kstopmachine etc..
8753 if (sysctl_sched_rt_runtime == 0)
8756 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8757 for_each_possible_cpu(i) {
8758 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8760 raw_spin_lock(&rt_rq->rt_runtime_lock);
8761 rt_rq->rt_runtime = global_rt_runtime();
8762 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8764 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8768 #endif /* CONFIG_RT_GROUP_SCHED */
8770 int sched_rt_handler(struct ctl_table *table, int write,
8771 void __user *buffer, size_t *lenp,
8775 int old_period, old_runtime;
8776 static DEFINE_MUTEX(mutex);
8779 old_period = sysctl_sched_rt_period;
8780 old_runtime = sysctl_sched_rt_runtime;
8782 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8784 if (!ret && write) {
8785 ret = sched_rt_global_constraints();
8787 sysctl_sched_rt_period = old_period;
8788 sysctl_sched_rt_runtime = old_runtime;
8790 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8791 def_rt_bandwidth.rt_period =
8792 ns_to_ktime(global_rt_period());
8795 mutex_unlock(&mutex);
8800 #ifdef CONFIG_CGROUP_SCHED
8802 /* return corresponding task_group object of a cgroup */
8803 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8805 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8806 struct task_group, css);
8809 static struct cgroup_subsys_state *
8810 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8812 struct task_group *tg, *parent;
8814 if (!cgrp->parent) {
8815 /* This is early initialization for the top cgroup */
8816 return &root_task_group.css;
8819 parent = cgroup_tg(cgrp->parent);
8820 tg = sched_create_group(parent);
8822 return ERR_PTR(-ENOMEM);
8828 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8830 struct task_group *tg = cgroup_tg(cgrp);
8832 sched_destroy_group(tg);
8836 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8838 #ifdef CONFIG_RT_GROUP_SCHED
8839 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8842 /* We don't support RT-tasks being in separate groups */
8843 if (tsk->sched_class != &fair_sched_class)
8850 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8852 sched_move_task(tsk);
8856 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
8857 struct cgroup *old_cgrp, struct task_struct *task)
8860 * cgroup_exit() is called in the copy_process() failure path.
8861 * Ignore this case since the task hasn't ran yet, this avoids
8862 * trying to poke a half freed task state from generic code.
8864 if (!(task->flags & PF_EXITING))
8867 sched_move_task(task);
8870 #ifdef CONFIG_FAIR_GROUP_SCHED
8871 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8874 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
8877 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8879 struct task_group *tg = cgroup_tg(cgrp);
8881 return (u64) scale_load_down(tg->shares);
8883 #endif /* CONFIG_FAIR_GROUP_SCHED */
8885 #ifdef CONFIG_RT_GROUP_SCHED
8886 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8889 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8892 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8894 return sched_group_rt_runtime(cgroup_tg(cgrp));
8897 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8900 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8903 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8905 return sched_group_rt_period(cgroup_tg(cgrp));
8907 #endif /* CONFIG_RT_GROUP_SCHED */
8909 static struct cftype cpu_files[] = {
8910 #ifdef CONFIG_FAIR_GROUP_SCHED
8913 .read_u64 = cpu_shares_read_u64,
8914 .write_u64 = cpu_shares_write_u64,
8917 #ifdef CONFIG_RT_GROUP_SCHED
8919 .name = "rt_runtime_us",
8920 .read_s64 = cpu_rt_runtime_read,
8921 .write_s64 = cpu_rt_runtime_write,
8924 .name = "rt_period_us",
8925 .read_u64 = cpu_rt_period_read_uint,
8926 .write_u64 = cpu_rt_period_write_uint,
8931 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8933 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8936 struct cgroup_subsys cpu_cgroup_subsys = {
8938 .create = cpu_cgroup_create,
8939 .destroy = cpu_cgroup_destroy,
8940 .can_attach_task = cpu_cgroup_can_attach_task,
8941 .attach_task = cpu_cgroup_attach_task,
8942 .exit = cpu_cgroup_exit,
8943 .populate = cpu_cgroup_populate,
8944 .subsys_id = cpu_cgroup_subsys_id,
8948 #endif /* CONFIG_CGROUP_SCHED */
8950 #ifdef CONFIG_CGROUP_CPUACCT
8953 * CPU accounting code for task groups.
8955 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8956 * (balbir@in.ibm.com).
8959 /* track cpu usage of a group of tasks and its child groups */
8961 struct cgroup_subsys_state css;
8962 /* cpuusage holds pointer to a u64-type object on every cpu */
8963 u64 __percpu *cpuusage;
8964 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8965 struct cpuacct *parent;
8968 struct cgroup_subsys cpuacct_subsys;
8970 /* return cpu accounting group corresponding to this container */
8971 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8973 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8974 struct cpuacct, css);
8977 /* return cpu accounting group to which this task belongs */
8978 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8980 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8981 struct cpuacct, css);
8984 /* create a new cpu accounting group */
8985 static struct cgroup_subsys_state *cpuacct_create(
8986 struct cgroup_subsys *ss, struct cgroup *cgrp)
8988 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8994 ca->cpuusage = alloc_percpu(u64);
8998 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8999 if (percpu_counter_init(&ca->cpustat[i], 0))
9000 goto out_free_counters;
9003 ca->parent = cgroup_ca(cgrp->parent);
9009 percpu_counter_destroy(&ca->cpustat[i]);
9010 free_percpu(ca->cpuusage);
9014 return ERR_PTR(-ENOMEM);
9017 /* destroy an existing cpu accounting group */
9019 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9021 struct cpuacct *ca = cgroup_ca(cgrp);
9024 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9025 percpu_counter_destroy(&ca->cpustat[i]);
9026 free_percpu(ca->cpuusage);
9030 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9032 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9035 #ifndef CONFIG_64BIT
9037 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9039 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9041 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9049 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9051 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9053 #ifndef CONFIG_64BIT
9055 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9057 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9059 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9065 /* return total cpu usage (in nanoseconds) of a group */
9066 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9068 struct cpuacct *ca = cgroup_ca(cgrp);
9069 u64 totalcpuusage = 0;
9072 for_each_present_cpu(i)
9073 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9075 return totalcpuusage;
9078 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9081 struct cpuacct *ca = cgroup_ca(cgrp);
9090 for_each_present_cpu(i)
9091 cpuacct_cpuusage_write(ca, i, 0);
9097 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9100 struct cpuacct *ca = cgroup_ca(cgroup);
9104 for_each_present_cpu(i) {
9105 percpu = cpuacct_cpuusage_read(ca, i);
9106 seq_printf(m, "%llu ", (unsigned long long) percpu);
9108 seq_printf(m, "\n");
9112 static const char *cpuacct_stat_desc[] = {
9113 [CPUACCT_STAT_USER] = "user",
9114 [CPUACCT_STAT_SYSTEM] = "system",
9117 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9118 struct cgroup_map_cb *cb)
9120 struct cpuacct *ca = cgroup_ca(cgrp);
9123 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9124 s64 val = percpu_counter_read(&ca->cpustat[i]);
9125 val = cputime64_to_clock_t(val);
9126 cb->fill(cb, cpuacct_stat_desc[i], val);
9131 static struct cftype files[] = {
9134 .read_u64 = cpuusage_read,
9135 .write_u64 = cpuusage_write,
9138 .name = "usage_percpu",
9139 .read_seq_string = cpuacct_percpu_seq_read,
9143 .read_map = cpuacct_stats_show,
9147 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9149 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9153 * charge this task's execution time to its accounting group.
9155 * called with rq->lock held.
9157 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9162 if (unlikely(!cpuacct_subsys.active))
9165 cpu = task_cpu(tsk);
9171 for (; ca; ca = ca->parent) {
9172 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9173 *cpuusage += cputime;
9180 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9181 * in cputime_t units. As a result, cpuacct_update_stats calls
9182 * percpu_counter_add with values large enough to always overflow the
9183 * per cpu batch limit causing bad SMP scalability.
9185 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9186 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9187 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9190 #define CPUACCT_BATCH \
9191 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9193 #define CPUACCT_BATCH 0
9197 * Charge the system/user time to the task's accounting group.
9199 static void cpuacct_update_stats(struct task_struct *tsk,
9200 enum cpuacct_stat_index idx, cputime_t val)
9203 int batch = CPUACCT_BATCH;
9205 if (unlikely(!cpuacct_subsys.active))
9212 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9218 struct cgroup_subsys cpuacct_subsys = {
9220 .create = cpuacct_create,
9221 .destroy = cpuacct_destroy,
9222 .populate = cpuacct_populate,
9223 .subsys_id = cpuacct_subsys_id,
9225 #endif /* CONFIG_CGROUP_CPUACCT */