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 <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/stop_machine.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
74 #include <linux/slab.h>
77 #include <asm/irq_regs.h>
78 #include <asm/mutex.h>
80 #include "sched_cpupri.h"
81 #include "workqueue_sched.h"
82 #include "sched_autogroup.h"
84 #define CREATE_TRACE_POINTS
85 #include <trace/events/sched.h>
88 * Convert user-nice values [ -20 ... 0 ... 19 ]
89 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
92 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
93 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
94 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
97 * 'User priority' is the nice value converted to something we
98 * can work with better when scaling various scheduler parameters,
99 * it's a [ 0 ... 39 ] range.
101 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
102 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
103 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
106 * Helpers for converting nanosecond timing to jiffy resolution
108 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
110 #define NICE_0_LOAD SCHED_LOAD_SCALE
111 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
114 * These are the 'tuning knobs' of the scheduler:
116 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
117 * Timeslices get refilled after they expire.
119 #define DEF_TIMESLICE (100 * HZ / 1000)
122 * single value that denotes runtime == period, ie unlimited time.
124 #define RUNTIME_INF ((u64)~0ULL)
126 static inline int rt_policy(int policy)
128 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
133 static inline int task_has_rt_policy(struct task_struct *p)
135 return rt_policy(p->policy);
139 * This is the priority-queue data structure of the RT scheduling class:
141 struct rt_prio_array {
142 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
143 struct list_head queue[MAX_RT_PRIO];
146 struct rt_bandwidth {
147 /* nests inside the rq lock: */
148 raw_spinlock_t rt_runtime_lock;
151 struct hrtimer rt_period_timer;
154 static struct rt_bandwidth def_rt_bandwidth;
156 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
158 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
160 struct rt_bandwidth *rt_b =
161 container_of(timer, struct rt_bandwidth, rt_period_timer);
167 now = hrtimer_cb_get_time(timer);
168 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
173 idle = do_sched_rt_period_timer(rt_b, overrun);
176 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
180 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
182 rt_b->rt_period = ns_to_ktime(period);
183 rt_b->rt_runtime = runtime;
185 raw_spin_lock_init(&rt_b->rt_runtime_lock);
187 hrtimer_init(&rt_b->rt_period_timer,
188 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
189 rt_b->rt_period_timer.function = sched_rt_period_timer;
192 static inline int rt_bandwidth_enabled(void)
194 return sysctl_sched_rt_runtime >= 0;
197 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
201 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
204 if (hrtimer_active(&rt_b->rt_period_timer))
207 raw_spin_lock(&rt_b->rt_runtime_lock);
212 if (hrtimer_active(&rt_b->rt_period_timer))
215 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
216 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
218 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
219 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
220 delta = ktime_to_ns(ktime_sub(hard, soft));
221 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
222 HRTIMER_MODE_ABS_PINNED, 0);
224 raw_spin_unlock(&rt_b->rt_runtime_lock);
227 #ifdef CONFIG_RT_GROUP_SCHED
228 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
230 hrtimer_cancel(&rt_b->rt_period_timer);
235 * sched_domains_mutex serializes calls to arch_init_sched_domains,
236 * detach_destroy_domains and partition_sched_domains.
238 static DEFINE_MUTEX(sched_domains_mutex);
240 #ifdef CONFIG_CGROUP_SCHED
242 #include <linux/cgroup.h>
246 static LIST_HEAD(task_groups);
248 /* task group related information */
250 struct cgroup_subsys_state css;
252 #ifdef CONFIG_FAIR_GROUP_SCHED
253 /* schedulable entities of this group on each cpu */
254 struct sched_entity **se;
255 /* runqueue "owned" by this group on each cpu */
256 struct cfs_rq **cfs_rq;
257 unsigned long shares;
259 atomic_t load_weight;
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity **rt_se;
264 struct rt_rq **rt_rq;
266 struct rt_bandwidth rt_bandwidth;
270 struct list_head list;
272 struct task_group *parent;
273 struct list_head siblings;
274 struct list_head children;
276 #ifdef CONFIG_SCHED_AUTOGROUP
277 struct autogroup *autogroup;
281 /* task_group_lock serializes the addition/removal of task groups */
282 static DEFINE_SPINLOCK(task_group_lock);
284 #ifdef CONFIG_FAIR_GROUP_SCHED
286 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
289 * A weight of 0 or 1 can cause arithmetics problems.
290 * A weight of a cfs_rq is the sum of weights of which entities
291 * are queued on this cfs_rq, so a weight of a entity should not be
292 * too large, so as the shares value of a task group.
293 * (The default weight is 1024 - so there's no practical
294 * limitation from this.)
297 #define MAX_SHARES (1UL << 18)
299 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
302 /* Default task group.
303 * Every task in system belong to this group at bootup.
305 struct task_group root_task_group;
307 #endif /* CONFIG_CGROUP_SCHED */
309 /* CFS-related fields in a runqueue */
311 struct load_weight load;
312 unsigned long nr_running;
317 struct rb_root tasks_timeline;
318 struct rb_node *rb_leftmost;
320 struct list_head tasks;
321 struct list_head *balance_iterator;
324 * 'curr' points to currently running entity on this cfs_rq.
325 * It is set to NULL otherwise (i.e when none are currently running).
327 struct sched_entity *curr, *next, *last;
329 unsigned int nr_spread_over;
331 #ifdef CONFIG_FAIR_GROUP_SCHED
332 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
335 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
336 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
337 * (like users, containers etc.)
339 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
340 * list is used during load balance.
343 struct list_head leaf_cfs_rq_list;
344 struct task_group *tg; /* group that "owns" this runqueue */
348 * the part of load.weight contributed by tasks
350 unsigned long task_weight;
353 * h_load = weight * f(tg)
355 * Where f(tg) is the recursive weight fraction assigned to
358 unsigned long h_load;
361 * Maintaining per-cpu shares distribution for group scheduling
363 * load_stamp is the last time we updated the load average
364 * load_last is the last time we updated the load average and saw load
365 * load_unacc_exec_time is currently unaccounted execution time
369 u64 load_stamp, load_last, load_unacc_exec_time;
371 unsigned long load_contribution;
376 /* Real-Time classes' related field in a runqueue: */
378 struct rt_prio_array active;
379 unsigned long rt_nr_running;
380 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
382 int curr; /* highest queued rt task prio */
384 int next; /* next highest */
389 unsigned long rt_nr_migratory;
390 unsigned long rt_nr_total;
392 struct plist_head pushable_tasks;
397 /* Nests inside the rq lock: */
398 raw_spinlock_t rt_runtime_lock;
400 #ifdef CONFIG_RT_GROUP_SCHED
401 unsigned long rt_nr_boosted;
404 struct list_head leaf_rt_rq_list;
405 struct task_group *tg;
412 * We add the notion of a root-domain which will be used to define per-domain
413 * variables. Each exclusive cpuset essentially defines an island domain by
414 * fully partitioning the member cpus from any other cpuset. Whenever a new
415 * exclusive cpuset is created, we also create and attach a new root-domain
422 cpumask_var_t online;
425 * The "RT overload" flag: it gets set if a CPU has more than
426 * one runnable RT task.
428 cpumask_var_t rto_mask;
430 struct cpupri cpupri;
434 * By default the system creates a single root-domain with all cpus as
435 * members (mimicking the global state we have today).
437 static struct root_domain def_root_domain;
439 #endif /* CONFIG_SMP */
442 * This is the main, per-CPU runqueue data structure.
444 * Locking rule: those places that want to lock multiple runqueues
445 * (such as the load balancing or the thread migration code), lock
446 * acquire operations must be ordered by ascending &runqueue.
453 * nr_running and cpu_load should be in the same cacheline because
454 * remote CPUs use both these fields when doing load calculation.
456 unsigned long nr_running;
457 #define CPU_LOAD_IDX_MAX 5
458 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
459 unsigned long last_load_update_tick;
462 unsigned char nohz_balance_kick;
464 unsigned int skip_clock_update;
466 /* capture load from *all* tasks on this cpu: */
467 struct load_weight load;
468 unsigned long nr_load_updates;
474 #ifdef CONFIG_FAIR_GROUP_SCHED
475 /* list of leaf cfs_rq on this cpu: */
476 struct list_head leaf_cfs_rq_list;
478 #ifdef CONFIG_RT_GROUP_SCHED
479 struct list_head leaf_rt_rq_list;
483 * This is part of a global counter where only the total sum
484 * over all CPUs matters. A task can increase this counter on
485 * one CPU and if it got migrated afterwards it may decrease
486 * it on another CPU. Always updated under the runqueue lock:
488 unsigned long nr_uninterruptible;
490 struct task_struct *curr, *idle, *stop;
491 unsigned long next_balance;
492 struct mm_struct *prev_mm;
500 struct root_domain *rd;
501 struct sched_domain *sd;
503 unsigned long cpu_power;
505 unsigned char idle_at_tick;
506 /* For active balancing */
510 struct cpu_stop_work active_balance_work;
511 /* cpu of this runqueue: */
515 unsigned long avg_load_per_task;
523 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
527 /* calc_load related fields */
528 unsigned long calc_load_update;
529 long calc_load_active;
531 #ifdef CONFIG_SCHED_HRTICK
533 int hrtick_csd_pending;
534 struct call_single_data hrtick_csd;
536 struct hrtimer hrtick_timer;
539 #ifdef CONFIG_SCHEDSTATS
541 struct sched_info rq_sched_info;
542 unsigned long long rq_cpu_time;
543 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
545 /* sys_sched_yield() stats */
546 unsigned int yld_count;
548 /* schedule() stats */
549 unsigned int sched_switch;
550 unsigned int sched_count;
551 unsigned int sched_goidle;
553 /* try_to_wake_up() stats */
554 unsigned int ttwu_count;
555 unsigned int ttwu_local;
559 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
562 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
564 static inline int cpu_of(struct rq *rq)
573 #define rcu_dereference_check_sched_domain(p) \
574 rcu_dereference_check((p), \
575 rcu_read_lock_sched_held() || \
576 lockdep_is_held(&sched_domains_mutex))
579 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
580 * See detach_destroy_domains: synchronize_sched for details.
582 * The domain tree of any CPU may only be accessed from within
583 * preempt-disabled sections.
585 #define for_each_domain(cpu, __sd) \
586 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
588 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
589 #define this_rq() (&__get_cpu_var(runqueues))
590 #define task_rq(p) cpu_rq(task_cpu(p))
591 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
592 #define raw_rq() (&__raw_get_cpu_var(runqueues))
594 #ifdef CONFIG_CGROUP_SCHED
597 * Return the group to which this tasks belongs.
599 * We use task_subsys_state_check() and extend the RCU verification
600 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
601 * holds that lock for each task it moves into the cgroup. Therefore
602 * by holding that lock, we pin the task to the current cgroup.
604 static inline struct task_group *task_group(struct task_struct *p)
606 struct task_group *tg;
607 struct cgroup_subsys_state *css;
609 if (p->flags & PF_EXITING)
610 return &root_task_group;
612 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
613 lockdep_is_held(&task_rq(p)->lock));
614 tg = container_of(css, struct task_group, css);
616 return autogroup_task_group(p, tg);
619 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
620 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
622 #ifdef CONFIG_FAIR_GROUP_SCHED
623 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
624 p->se.parent = task_group(p)->se[cpu];
627 #ifdef CONFIG_RT_GROUP_SCHED
628 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
629 p->rt.parent = task_group(p)->rt_se[cpu];
633 #else /* CONFIG_CGROUP_SCHED */
635 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
636 static inline struct task_group *task_group(struct task_struct *p)
641 #endif /* CONFIG_CGROUP_SCHED */
643 static void update_rq_clock_task(struct rq *rq, s64 delta);
645 static void update_rq_clock(struct rq *rq)
649 if (rq->skip_clock_update)
652 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
654 update_rq_clock_task(rq, delta);
658 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
660 #ifdef CONFIG_SCHED_DEBUG
661 # define const_debug __read_mostly
663 # define const_debug static const
668 * @cpu: the processor in question.
670 * Returns true if the current cpu runqueue is locked.
671 * This interface allows printk to be called with the runqueue lock
672 * held and know whether or not it is OK to wake up the klogd.
674 int runqueue_is_locked(int cpu)
676 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
680 * Debugging: various feature bits
683 #define SCHED_FEAT(name, enabled) \
684 __SCHED_FEAT_##name ,
687 #include "sched_features.h"
692 #define SCHED_FEAT(name, enabled) \
693 (1UL << __SCHED_FEAT_##name) * enabled |
695 const_debug unsigned int sysctl_sched_features =
696 #include "sched_features.h"
701 #ifdef CONFIG_SCHED_DEBUG
702 #define SCHED_FEAT(name, enabled) \
705 static __read_mostly char *sched_feat_names[] = {
706 #include "sched_features.h"
712 static int sched_feat_show(struct seq_file *m, void *v)
716 for (i = 0; sched_feat_names[i]; i++) {
717 if (!(sysctl_sched_features & (1UL << i)))
719 seq_printf(m, "%s ", sched_feat_names[i]);
727 sched_feat_write(struct file *filp, const char __user *ubuf,
728 size_t cnt, loff_t *ppos)
738 if (copy_from_user(&buf, ubuf, cnt))
744 if (strncmp(cmp, "NO_", 3) == 0) {
749 for (i = 0; sched_feat_names[i]; i++) {
750 if (strcmp(cmp, sched_feat_names[i]) == 0) {
752 sysctl_sched_features &= ~(1UL << i);
754 sysctl_sched_features |= (1UL << i);
759 if (!sched_feat_names[i])
767 static int sched_feat_open(struct inode *inode, struct file *filp)
769 return single_open(filp, sched_feat_show, NULL);
772 static const struct file_operations sched_feat_fops = {
773 .open = sched_feat_open,
774 .write = sched_feat_write,
777 .release = single_release,
780 static __init int sched_init_debug(void)
782 debugfs_create_file("sched_features", 0644, NULL, NULL,
787 late_initcall(sched_init_debug);
791 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
794 * Number of tasks to iterate in a single balance run.
795 * Limited because this is done with IRQs disabled.
797 const_debug unsigned int sysctl_sched_nr_migrate = 32;
800 * period over which we average the RT time consumption, measured
805 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
808 * period over which we measure -rt task cpu usage in us.
811 unsigned int sysctl_sched_rt_period = 1000000;
813 static __read_mostly int scheduler_running;
816 * part of the period that we allow rt tasks to run in us.
819 int sysctl_sched_rt_runtime = 950000;
821 static inline u64 global_rt_period(void)
823 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
826 static inline u64 global_rt_runtime(void)
828 if (sysctl_sched_rt_runtime < 0)
831 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
834 #ifndef prepare_arch_switch
835 # define prepare_arch_switch(next) do { } while (0)
837 #ifndef finish_arch_switch
838 # define finish_arch_switch(prev) do { } while (0)
841 static inline int task_current(struct rq *rq, struct task_struct *p)
843 return rq->curr == p;
846 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
847 static inline int task_running(struct rq *rq, struct task_struct *p)
849 return task_current(rq, p);
852 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
856 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
858 #ifdef CONFIG_DEBUG_SPINLOCK
859 /* this is a valid case when another task releases the spinlock */
860 rq->lock.owner = current;
863 * If we are tracking spinlock dependencies then we have to
864 * fix up the runqueue lock - which gets 'carried over' from
867 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
869 raw_spin_unlock_irq(&rq->lock);
872 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
873 static inline int task_running(struct rq *rq, struct task_struct *p)
878 return task_current(rq, p);
882 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
886 * We can optimise this out completely for !SMP, because the
887 * SMP rebalancing from interrupt is the only thing that cares
892 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
893 raw_spin_unlock_irq(&rq->lock);
895 raw_spin_unlock(&rq->lock);
899 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
903 * After ->oncpu is cleared, the task can be moved to a different CPU.
904 * We must ensure this doesn't happen until the switch is completely
910 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
914 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
917 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
920 static inline int task_is_waking(struct task_struct *p)
922 return unlikely(p->state == TASK_WAKING);
926 * __task_rq_lock - lock the runqueue a given task resides on.
927 * Must be called interrupts disabled.
929 static inline struct rq *__task_rq_lock(struct task_struct *p)
936 raw_spin_lock(&rq->lock);
937 if (likely(rq == task_rq(p)))
939 raw_spin_unlock(&rq->lock);
944 * task_rq_lock - lock the runqueue a given task resides on and disable
945 * interrupts. Note the ordering: we can safely lookup the task_rq without
946 * explicitly disabling preemption.
948 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
954 local_irq_save(*flags);
956 raw_spin_lock(&rq->lock);
957 if (likely(rq == task_rq(p)))
959 raw_spin_unlock_irqrestore(&rq->lock, *flags);
963 static void __task_rq_unlock(struct rq *rq)
966 raw_spin_unlock(&rq->lock);
969 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
972 raw_spin_unlock_irqrestore(&rq->lock, *flags);
976 * this_rq_lock - lock this runqueue and disable interrupts.
978 static struct rq *this_rq_lock(void)
985 raw_spin_lock(&rq->lock);
990 #ifdef CONFIG_SCHED_HRTICK
992 * Use HR-timers to deliver accurate preemption points.
994 * Its all a bit involved since we cannot program an hrt while holding the
995 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
998 * When we get rescheduled we reprogram the hrtick_timer outside of the
1004 * - enabled by features
1005 * - hrtimer is actually high res
1007 static inline int hrtick_enabled(struct rq *rq)
1009 if (!sched_feat(HRTICK))
1011 if (!cpu_active(cpu_of(rq)))
1013 return hrtimer_is_hres_active(&rq->hrtick_timer);
1016 static void hrtick_clear(struct rq *rq)
1018 if (hrtimer_active(&rq->hrtick_timer))
1019 hrtimer_cancel(&rq->hrtick_timer);
1023 * High-resolution timer tick.
1024 * Runs from hardirq context with interrupts disabled.
1026 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1028 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1030 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1032 raw_spin_lock(&rq->lock);
1033 update_rq_clock(rq);
1034 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1035 raw_spin_unlock(&rq->lock);
1037 return HRTIMER_NORESTART;
1042 * called from hardirq (IPI) context
1044 static void __hrtick_start(void *arg)
1046 struct rq *rq = arg;
1048 raw_spin_lock(&rq->lock);
1049 hrtimer_restart(&rq->hrtick_timer);
1050 rq->hrtick_csd_pending = 0;
1051 raw_spin_unlock(&rq->lock);
1055 * Called to set the hrtick timer state.
1057 * called with rq->lock held and irqs disabled
1059 static void hrtick_start(struct rq *rq, u64 delay)
1061 struct hrtimer *timer = &rq->hrtick_timer;
1062 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1064 hrtimer_set_expires(timer, time);
1066 if (rq == this_rq()) {
1067 hrtimer_restart(timer);
1068 } else if (!rq->hrtick_csd_pending) {
1069 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1070 rq->hrtick_csd_pending = 1;
1075 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1077 int cpu = (int)(long)hcpu;
1080 case CPU_UP_CANCELED:
1081 case CPU_UP_CANCELED_FROZEN:
1082 case CPU_DOWN_PREPARE:
1083 case CPU_DOWN_PREPARE_FROZEN:
1085 case CPU_DEAD_FROZEN:
1086 hrtick_clear(cpu_rq(cpu));
1093 static __init void init_hrtick(void)
1095 hotcpu_notifier(hotplug_hrtick, 0);
1099 * Called to set the hrtick timer state.
1101 * called with rq->lock held and irqs disabled
1103 static void hrtick_start(struct rq *rq, u64 delay)
1105 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1106 HRTIMER_MODE_REL_PINNED, 0);
1109 static inline void init_hrtick(void)
1112 #endif /* CONFIG_SMP */
1114 static void init_rq_hrtick(struct rq *rq)
1117 rq->hrtick_csd_pending = 0;
1119 rq->hrtick_csd.flags = 0;
1120 rq->hrtick_csd.func = __hrtick_start;
1121 rq->hrtick_csd.info = rq;
1124 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1125 rq->hrtick_timer.function = hrtick;
1127 #else /* CONFIG_SCHED_HRTICK */
1128 static inline void hrtick_clear(struct rq *rq)
1132 static inline void init_rq_hrtick(struct rq *rq)
1136 static inline void init_hrtick(void)
1139 #endif /* CONFIG_SCHED_HRTICK */
1142 * resched_task - mark a task 'to be rescheduled now'.
1144 * On UP this means the setting of the need_resched flag, on SMP it
1145 * might also involve a cross-CPU call to trigger the scheduler on
1150 #ifndef tsk_is_polling
1151 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1154 static void resched_task(struct task_struct *p)
1158 assert_raw_spin_locked(&task_rq(p)->lock);
1160 if (test_tsk_need_resched(p))
1163 set_tsk_need_resched(p);
1166 if (cpu == smp_processor_id())
1169 /* NEED_RESCHED must be visible before we test polling */
1171 if (!tsk_is_polling(p))
1172 smp_send_reschedule(cpu);
1175 static void resched_cpu(int cpu)
1177 struct rq *rq = cpu_rq(cpu);
1178 unsigned long flags;
1180 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1182 resched_task(cpu_curr(cpu));
1183 raw_spin_unlock_irqrestore(&rq->lock, flags);
1188 * In the semi idle case, use the nearest busy cpu for migrating timers
1189 * from an idle cpu. This is good for power-savings.
1191 * We don't do similar optimization for completely idle system, as
1192 * selecting an idle cpu will add more delays to the timers than intended
1193 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1195 int get_nohz_timer_target(void)
1197 int cpu = smp_processor_id();
1199 struct sched_domain *sd;
1201 for_each_domain(cpu, sd) {
1202 for_each_cpu(i, sched_domain_span(sd))
1209 * When add_timer_on() enqueues a timer into the timer wheel of an
1210 * idle CPU then this timer might expire before the next timer event
1211 * which is scheduled to wake up that CPU. In case of a completely
1212 * idle system the next event might even be infinite time into the
1213 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1214 * leaves the inner idle loop so the newly added timer is taken into
1215 * account when the CPU goes back to idle and evaluates the timer
1216 * wheel for the next timer event.
1218 void wake_up_idle_cpu(int cpu)
1220 struct rq *rq = cpu_rq(cpu);
1222 if (cpu == smp_processor_id())
1226 * This is safe, as this function is called with the timer
1227 * wheel base lock of (cpu) held. When the CPU is on the way
1228 * to idle and has not yet set rq->curr to idle then it will
1229 * be serialized on the timer wheel base lock and take the new
1230 * timer into account automatically.
1232 if (rq->curr != rq->idle)
1236 * We can set TIF_RESCHED on the idle task of the other CPU
1237 * lockless. The worst case is that the other CPU runs the
1238 * idle task through an additional NOOP schedule()
1240 set_tsk_need_resched(rq->idle);
1242 /* NEED_RESCHED must be visible before we test polling */
1244 if (!tsk_is_polling(rq->idle))
1245 smp_send_reschedule(cpu);
1248 #endif /* CONFIG_NO_HZ */
1250 static u64 sched_avg_period(void)
1252 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1255 static void sched_avg_update(struct rq *rq)
1257 s64 period = sched_avg_period();
1259 while ((s64)(rq->clock - rq->age_stamp) > period) {
1261 * Inline assembly required to prevent the compiler
1262 * optimising this loop into a divmod call.
1263 * See __iter_div_u64_rem() for another example of this.
1265 asm("" : "+rm" (rq->age_stamp));
1266 rq->age_stamp += period;
1271 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1273 rq->rt_avg += rt_delta;
1274 sched_avg_update(rq);
1277 #else /* !CONFIG_SMP */
1278 static void resched_task(struct task_struct *p)
1280 assert_raw_spin_locked(&task_rq(p)->lock);
1281 set_tsk_need_resched(p);
1284 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1288 static void sched_avg_update(struct rq *rq)
1291 #endif /* CONFIG_SMP */
1293 #if BITS_PER_LONG == 32
1294 # define WMULT_CONST (~0UL)
1296 # define WMULT_CONST (1UL << 32)
1299 #define WMULT_SHIFT 32
1302 * Shift right and round:
1304 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1307 * delta *= weight / lw
1309 static unsigned long
1310 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1311 struct load_weight *lw)
1315 if (!lw->inv_weight) {
1316 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1319 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1323 tmp = (u64)delta_exec * weight;
1325 * Check whether we'd overflow the 64-bit multiplication:
1327 if (unlikely(tmp > WMULT_CONST))
1328 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1331 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1333 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1336 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1342 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1348 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1355 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1356 * of tasks with abnormal "nice" values across CPUs the contribution that
1357 * each task makes to its run queue's load is weighted according to its
1358 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1359 * scaled version of the new time slice allocation that they receive on time
1363 #define WEIGHT_IDLEPRIO 3
1364 #define WMULT_IDLEPRIO 1431655765
1367 * Nice levels are multiplicative, with a gentle 10% change for every
1368 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1369 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1370 * that remained on nice 0.
1372 * The "10% effect" is relative and cumulative: from _any_ nice level,
1373 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1374 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1375 * If a task goes up by ~10% and another task goes down by ~10% then
1376 * the relative distance between them is ~25%.)
1378 static const int prio_to_weight[40] = {
1379 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1380 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1381 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1382 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1383 /* 0 */ 1024, 820, 655, 526, 423,
1384 /* 5 */ 335, 272, 215, 172, 137,
1385 /* 10 */ 110, 87, 70, 56, 45,
1386 /* 15 */ 36, 29, 23, 18, 15,
1390 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1392 * In cases where the weight does not change often, we can use the
1393 * precalculated inverse to speed up arithmetics by turning divisions
1394 * into multiplications:
1396 static const u32 prio_to_wmult[40] = {
1397 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1398 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1399 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1400 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1401 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1402 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1403 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1404 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1407 /* Time spent by the tasks of the cpu accounting group executing in ... */
1408 enum cpuacct_stat_index {
1409 CPUACCT_STAT_USER, /* ... user mode */
1410 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1412 CPUACCT_STAT_NSTATS,
1415 #ifdef CONFIG_CGROUP_CPUACCT
1416 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1417 static void cpuacct_update_stats(struct task_struct *tsk,
1418 enum cpuacct_stat_index idx, cputime_t val);
1420 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1421 static inline void cpuacct_update_stats(struct task_struct *tsk,
1422 enum cpuacct_stat_index idx, cputime_t val) {}
1425 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1427 update_load_add(&rq->load, load);
1430 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1432 update_load_sub(&rq->load, load);
1435 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1436 typedef int (*tg_visitor)(struct task_group *, void *);
1439 * Iterate the full tree, calling @down when first entering a node and @up when
1440 * leaving it for the final time.
1442 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1444 struct task_group *parent, *child;
1448 parent = &root_task_group;
1450 ret = (*down)(parent, data);
1453 list_for_each_entry_rcu(child, &parent->children, siblings) {
1460 ret = (*up)(parent, data);
1465 parent = parent->parent;
1474 static int tg_nop(struct task_group *tg, void *data)
1481 /* Used instead of source_load when we know the type == 0 */
1482 static unsigned long weighted_cpuload(const int cpu)
1484 return cpu_rq(cpu)->load.weight;
1488 * Return a low guess at the load of a migration-source cpu weighted
1489 * according to the scheduling class and "nice" value.
1491 * We want to under-estimate the load of migration sources, to
1492 * balance conservatively.
1494 static unsigned long source_load(int cpu, int type)
1496 struct rq *rq = cpu_rq(cpu);
1497 unsigned long total = weighted_cpuload(cpu);
1499 if (type == 0 || !sched_feat(LB_BIAS))
1502 return min(rq->cpu_load[type-1], total);
1506 * Return a high guess at the load of a migration-target cpu weighted
1507 * according to the scheduling class and "nice" value.
1509 static unsigned long target_load(int cpu, int type)
1511 struct rq *rq = cpu_rq(cpu);
1512 unsigned long total = weighted_cpuload(cpu);
1514 if (type == 0 || !sched_feat(LB_BIAS))
1517 return max(rq->cpu_load[type-1], total);
1520 static unsigned long power_of(int cpu)
1522 return cpu_rq(cpu)->cpu_power;
1525 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1527 static unsigned long cpu_avg_load_per_task(int cpu)
1529 struct rq *rq = cpu_rq(cpu);
1530 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1533 rq->avg_load_per_task = rq->load.weight / nr_running;
1535 rq->avg_load_per_task = 0;
1537 return rq->avg_load_per_task;
1540 #ifdef CONFIG_FAIR_GROUP_SCHED
1543 * Compute the cpu's hierarchical load factor for each task group.
1544 * This needs to be done in a top-down fashion because the load of a child
1545 * group is a fraction of its parents load.
1547 static int tg_load_down(struct task_group *tg, void *data)
1550 long cpu = (long)data;
1553 load = cpu_rq(cpu)->load.weight;
1555 load = tg->parent->cfs_rq[cpu]->h_load;
1556 load *= tg->se[cpu]->load.weight;
1557 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1560 tg->cfs_rq[cpu]->h_load = load;
1565 static void update_h_load(long cpu)
1567 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1572 #ifdef CONFIG_PREEMPT
1574 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1577 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1578 * way at the expense of forcing extra atomic operations in all
1579 * invocations. This assures that the double_lock is acquired using the
1580 * same underlying policy as the spinlock_t on this architecture, which
1581 * reduces latency compared to the unfair variant below. However, it
1582 * also adds more overhead and therefore may reduce throughput.
1584 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1585 __releases(this_rq->lock)
1586 __acquires(busiest->lock)
1587 __acquires(this_rq->lock)
1589 raw_spin_unlock(&this_rq->lock);
1590 double_rq_lock(this_rq, busiest);
1597 * Unfair double_lock_balance: Optimizes throughput at the expense of
1598 * latency by eliminating extra atomic operations when the locks are
1599 * already in proper order on entry. This favors lower cpu-ids and will
1600 * grant the double lock to lower cpus over higher ids under contention,
1601 * regardless of entry order into the function.
1603 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1604 __releases(this_rq->lock)
1605 __acquires(busiest->lock)
1606 __acquires(this_rq->lock)
1610 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1611 if (busiest < this_rq) {
1612 raw_spin_unlock(&this_rq->lock);
1613 raw_spin_lock(&busiest->lock);
1614 raw_spin_lock_nested(&this_rq->lock,
1615 SINGLE_DEPTH_NESTING);
1618 raw_spin_lock_nested(&busiest->lock,
1619 SINGLE_DEPTH_NESTING);
1624 #endif /* CONFIG_PREEMPT */
1627 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1629 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1631 if (unlikely(!irqs_disabled())) {
1632 /* printk() doesn't work good under rq->lock */
1633 raw_spin_unlock(&this_rq->lock);
1637 return _double_lock_balance(this_rq, busiest);
1640 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1641 __releases(busiest->lock)
1643 raw_spin_unlock(&busiest->lock);
1644 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1648 * double_rq_lock - safely lock two runqueues
1650 * Note this does not disable interrupts like task_rq_lock,
1651 * you need to do so manually before calling.
1653 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1654 __acquires(rq1->lock)
1655 __acquires(rq2->lock)
1657 BUG_ON(!irqs_disabled());
1659 raw_spin_lock(&rq1->lock);
1660 __acquire(rq2->lock); /* Fake it out ;) */
1663 raw_spin_lock(&rq1->lock);
1664 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1666 raw_spin_lock(&rq2->lock);
1667 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1673 * double_rq_unlock - safely unlock two runqueues
1675 * Note this does not restore interrupts like task_rq_unlock,
1676 * you need to do so manually after calling.
1678 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1679 __releases(rq1->lock)
1680 __releases(rq2->lock)
1682 raw_spin_unlock(&rq1->lock);
1684 raw_spin_unlock(&rq2->lock);
1686 __release(rq2->lock);
1691 static void calc_load_account_idle(struct rq *this_rq);
1692 static void update_sysctl(void);
1693 static int get_update_sysctl_factor(void);
1694 static void update_cpu_load(struct rq *this_rq);
1696 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1698 set_task_rq(p, cpu);
1701 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1702 * successfuly executed on another CPU. We must ensure that updates of
1703 * per-task data have been completed by this moment.
1706 task_thread_info(p)->cpu = cpu;
1710 static const struct sched_class rt_sched_class;
1712 #define sched_class_highest (&stop_sched_class)
1713 #define for_each_class(class) \
1714 for (class = sched_class_highest; class; class = class->next)
1716 #include "sched_stats.h"
1718 static void inc_nr_running(struct rq *rq)
1723 static void dec_nr_running(struct rq *rq)
1728 static void set_load_weight(struct task_struct *p)
1731 * SCHED_IDLE tasks get minimal weight:
1733 if (p->policy == SCHED_IDLE) {
1734 p->se.load.weight = WEIGHT_IDLEPRIO;
1735 p->se.load.inv_weight = WMULT_IDLEPRIO;
1739 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1740 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1743 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1745 update_rq_clock(rq);
1746 sched_info_queued(p);
1747 p->sched_class->enqueue_task(rq, p, flags);
1751 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1753 update_rq_clock(rq);
1754 sched_info_dequeued(p);
1755 p->sched_class->dequeue_task(rq, p, flags);
1760 * activate_task - move a task to the runqueue.
1762 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1764 if (task_contributes_to_load(p))
1765 rq->nr_uninterruptible--;
1767 enqueue_task(rq, p, flags);
1772 * deactivate_task - remove a task from the runqueue.
1774 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1776 if (task_contributes_to_load(p))
1777 rq->nr_uninterruptible++;
1779 dequeue_task(rq, p, flags);
1783 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1786 * There are no locks covering percpu hardirq/softirq time.
1787 * They are only modified in account_system_vtime, on corresponding CPU
1788 * with interrupts disabled. So, writes are safe.
1789 * They are read and saved off onto struct rq in update_rq_clock().
1790 * This may result in other CPU reading this CPU's irq time and can
1791 * race with irq/account_system_vtime on this CPU. We would either get old
1792 * or new value with a side effect of accounting a slice of irq time to wrong
1793 * task when irq is in progress while we read rq->clock. That is a worthy
1794 * compromise in place of having locks on each irq in account_system_time.
1796 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1797 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1799 static DEFINE_PER_CPU(u64, irq_start_time);
1800 static int sched_clock_irqtime;
1802 void enable_sched_clock_irqtime(void)
1804 sched_clock_irqtime = 1;
1807 void disable_sched_clock_irqtime(void)
1809 sched_clock_irqtime = 0;
1812 #ifndef CONFIG_64BIT
1813 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1815 static inline void irq_time_write_begin(void)
1817 __this_cpu_inc(irq_time_seq.sequence);
1821 static inline void irq_time_write_end(void)
1824 __this_cpu_inc(irq_time_seq.sequence);
1827 static inline u64 irq_time_read(int cpu)
1833 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1834 irq_time = per_cpu(cpu_softirq_time, cpu) +
1835 per_cpu(cpu_hardirq_time, cpu);
1836 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1840 #else /* CONFIG_64BIT */
1841 static inline void irq_time_write_begin(void)
1845 static inline void irq_time_write_end(void)
1849 static inline u64 irq_time_read(int cpu)
1851 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1853 #endif /* CONFIG_64BIT */
1856 * Called before incrementing preempt_count on {soft,}irq_enter
1857 * and before decrementing preempt_count on {soft,}irq_exit.
1859 void account_system_vtime(struct task_struct *curr)
1861 unsigned long flags;
1865 if (!sched_clock_irqtime)
1868 local_irq_save(flags);
1870 cpu = smp_processor_id();
1871 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1872 __this_cpu_add(irq_start_time, delta);
1874 irq_time_write_begin();
1876 * We do not account for softirq time from ksoftirqd here.
1877 * We want to continue accounting softirq time to ksoftirqd thread
1878 * in that case, so as not to confuse scheduler with a special task
1879 * that do not consume any time, but still wants to run.
1881 if (hardirq_count())
1882 __this_cpu_add(cpu_hardirq_time, delta);
1883 else if (in_serving_softirq() && !(curr->flags & PF_KSOFTIRQD))
1884 __this_cpu_add(cpu_softirq_time, delta);
1886 irq_time_write_end();
1887 local_irq_restore(flags);
1889 EXPORT_SYMBOL_GPL(account_system_vtime);
1891 static void update_rq_clock_task(struct rq *rq, s64 delta)
1895 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
1898 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1899 * this case when a previous update_rq_clock() happened inside a
1900 * {soft,}irq region.
1902 * When this happens, we stop ->clock_task and only update the
1903 * prev_irq_time stamp to account for the part that fit, so that a next
1904 * update will consume the rest. This ensures ->clock_task is
1907 * It does however cause some slight miss-attribution of {soft,}irq
1908 * time, a more accurate solution would be to update the irq_time using
1909 * the current rq->clock timestamp, except that would require using
1912 if (irq_delta > delta)
1915 rq->prev_irq_time += irq_delta;
1917 rq->clock_task += delta;
1919 if (irq_delta && sched_feat(NONIRQ_POWER))
1920 sched_rt_avg_update(rq, irq_delta);
1923 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
1925 static void update_rq_clock_task(struct rq *rq, s64 delta)
1927 rq->clock_task += delta;
1930 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
1932 #include "sched_idletask.c"
1933 #include "sched_fair.c"
1934 #include "sched_rt.c"
1935 #include "sched_autogroup.c"
1936 #include "sched_stoptask.c"
1937 #ifdef CONFIG_SCHED_DEBUG
1938 # include "sched_debug.c"
1941 void sched_set_stop_task(int cpu, struct task_struct *stop)
1943 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
1944 struct task_struct *old_stop = cpu_rq(cpu)->stop;
1948 * Make it appear like a SCHED_FIFO task, its something
1949 * userspace knows about and won't get confused about.
1951 * Also, it will make PI more or less work without too
1952 * much confusion -- but then, stop work should not
1953 * rely on PI working anyway.
1955 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
1957 stop->sched_class = &stop_sched_class;
1960 cpu_rq(cpu)->stop = stop;
1964 * Reset it back to a normal scheduling class so that
1965 * it can die in pieces.
1967 old_stop->sched_class = &rt_sched_class;
1972 * __normal_prio - return the priority that is based on the static prio
1974 static inline int __normal_prio(struct task_struct *p)
1976 return p->static_prio;
1980 * Calculate the expected normal priority: i.e. priority
1981 * without taking RT-inheritance into account. Might be
1982 * boosted by interactivity modifiers. Changes upon fork,
1983 * setprio syscalls, and whenever the interactivity
1984 * estimator recalculates.
1986 static inline int normal_prio(struct task_struct *p)
1990 if (task_has_rt_policy(p))
1991 prio = MAX_RT_PRIO-1 - p->rt_priority;
1993 prio = __normal_prio(p);
1998 * Calculate the current priority, i.e. the priority
1999 * taken into account by the scheduler. This value might
2000 * be boosted by RT tasks, or might be boosted by
2001 * interactivity modifiers. Will be RT if the task got
2002 * RT-boosted. If not then it returns p->normal_prio.
2004 static int effective_prio(struct task_struct *p)
2006 p->normal_prio = normal_prio(p);
2008 * If we are RT tasks or we were boosted to RT priority,
2009 * keep the priority unchanged. Otherwise, update priority
2010 * to the normal priority:
2012 if (!rt_prio(p->prio))
2013 return p->normal_prio;
2018 * task_curr - is this task currently executing on a CPU?
2019 * @p: the task in question.
2021 inline int task_curr(const struct task_struct *p)
2023 return cpu_curr(task_cpu(p)) == p;
2026 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2027 const struct sched_class *prev_class,
2028 int oldprio, int running)
2030 if (prev_class != p->sched_class) {
2031 if (prev_class->switched_from)
2032 prev_class->switched_from(rq, p, running);
2033 p->sched_class->switched_to(rq, p, running);
2035 p->sched_class->prio_changed(rq, p, oldprio, running);
2038 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2040 const struct sched_class *class;
2042 if (p->sched_class == rq->curr->sched_class) {
2043 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2045 for_each_class(class) {
2046 if (class == rq->curr->sched_class)
2048 if (class == p->sched_class) {
2049 resched_task(rq->curr);
2056 * A queue event has occurred, and we're going to schedule. In
2057 * this case, we can save a useless back to back clock update.
2059 if (rq->curr->se.on_rq && test_tsk_need_resched(rq->curr))
2060 rq->skip_clock_update = 1;
2065 * Is this task likely cache-hot:
2068 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2072 if (p->sched_class != &fair_sched_class)
2075 if (unlikely(p->policy == SCHED_IDLE))
2079 * Buddy candidates are cache hot:
2081 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2082 (&p->se == cfs_rq_of(&p->se)->next ||
2083 &p->se == cfs_rq_of(&p->se)->last))
2086 if (sysctl_sched_migration_cost == -1)
2088 if (sysctl_sched_migration_cost == 0)
2091 delta = now - p->se.exec_start;
2093 return delta < (s64)sysctl_sched_migration_cost;
2096 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2098 #ifdef CONFIG_SCHED_DEBUG
2100 * We should never call set_task_cpu() on a blocked task,
2101 * ttwu() will sort out the placement.
2103 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2104 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2107 trace_sched_migrate_task(p, new_cpu);
2109 if (task_cpu(p) != new_cpu) {
2110 p->se.nr_migrations++;
2111 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2114 __set_task_cpu(p, new_cpu);
2117 struct migration_arg {
2118 struct task_struct *task;
2122 static int migration_cpu_stop(void *data);
2125 * The task's runqueue lock must be held.
2126 * Returns true if you have to wait for migration thread.
2128 static bool migrate_task(struct task_struct *p, struct rq *rq)
2131 * If the task is not on a runqueue (and not running), then
2132 * the next wake-up will properly place the task.
2134 return p->se.on_rq || task_running(rq, p);
2138 * wait_task_inactive - wait for a thread to unschedule.
2140 * If @match_state is nonzero, it's the @p->state value just checked and
2141 * not expected to change. If it changes, i.e. @p might have woken up,
2142 * then return zero. When we succeed in waiting for @p to be off its CPU,
2143 * we return a positive number (its total switch count). If a second call
2144 * a short while later returns the same number, the caller can be sure that
2145 * @p has remained unscheduled the whole time.
2147 * The caller must ensure that the task *will* unschedule sometime soon,
2148 * else this function might spin for a *long* time. This function can't
2149 * be called with interrupts off, or it may introduce deadlock with
2150 * smp_call_function() if an IPI is sent by the same process we are
2151 * waiting to become inactive.
2153 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2155 unsigned long flags;
2162 * We do the initial early heuristics without holding
2163 * any task-queue locks at all. We'll only try to get
2164 * the runqueue lock when things look like they will
2170 * If the task is actively running on another CPU
2171 * still, just relax and busy-wait without holding
2174 * NOTE! Since we don't hold any locks, it's not
2175 * even sure that "rq" stays as the right runqueue!
2176 * But we don't care, since "task_running()" will
2177 * return false if the runqueue has changed and p
2178 * is actually now running somewhere else!
2180 while (task_running(rq, p)) {
2181 if (match_state && unlikely(p->state != match_state))
2187 * Ok, time to look more closely! We need the rq
2188 * lock now, to be *sure*. If we're wrong, we'll
2189 * just go back and repeat.
2191 rq = task_rq_lock(p, &flags);
2192 trace_sched_wait_task(p);
2193 running = task_running(rq, p);
2194 on_rq = p->se.on_rq;
2196 if (!match_state || p->state == match_state)
2197 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2198 task_rq_unlock(rq, &flags);
2201 * If it changed from the expected state, bail out now.
2203 if (unlikely(!ncsw))
2207 * Was it really running after all now that we
2208 * checked with the proper locks actually held?
2210 * Oops. Go back and try again..
2212 if (unlikely(running)) {
2218 * It's not enough that it's not actively running,
2219 * it must be off the runqueue _entirely_, and not
2222 * So if it was still runnable (but just not actively
2223 * running right now), it's preempted, and we should
2224 * yield - it could be a while.
2226 if (unlikely(on_rq)) {
2227 schedule_timeout_uninterruptible(1);
2232 * Ahh, all good. It wasn't running, and it wasn't
2233 * runnable, which means that it will never become
2234 * running in the future either. We're all done!
2243 * kick_process - kick a running thread to enter/exit the kernel
2244 * @p: the to-be-kicked thread
2246 * Cause a process which is running on another CPU to enter
2247 * kernel-mode, without any delay. (to get signals handled.)
2249 * NOTE: this function doesnt have to take the runqueue lock,
2250 * because all it wants to ensure is that the remote task enters
2251 * the kernel. If the IPI races and the task has been migrated
2252 * to another CPU then no harm is done and the purpose has been
2255 void kick_process(struct task_struct *p)
2261 if ((cpu != smp_processor_id()) && task_curr(p))
2262 smp_send_reschedule(cpu);
2265 EXPORT_SYMBOL_GPL(kick_process);
2266 #endif /* CONFIG_SMP */
2270 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2272 static int select_fallback_rq(int cpu, struct task_struct *p)
2275 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2277 /* Look for allowed, online CPU in same node. */
2278 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2279 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2282 /* Any allowed, online CPU? */
2283 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2284 if (dest_cpu < nr_cpu_ids)
2287 /* No more Mr. Nice Guy. */
2288 dest_cpu = cpuset_cpus_allowed_fallback(p);
2290 * Don't tell them about moving exiting tasks or
2291 * kernel threads (both mm NULL), since they never
2294 if (p->mm && printk_ratelimit()) {
2295 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2296 task_pid_nr(p), p->comm, cpu);
2303 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2306 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2308 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2311 * In order not to call set_task_cpu() on a blocking task we need
2312 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2315 * Since this is common to all placement strategies, this lives here.
2317 * [ this allows ->select_task() to simply return task_cpu(p) and
2318 * not worry about this generic constraint ]
2320 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2322 cpu = select_fallback_rq(task_cpu(p), p);
2327 static void update_avg(u64 *avg, u64 sample)
2329 s64 diff = sample - *avg;
2334 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2335 bool is_sync, bool is_migrate, bool is_local,
2336 unsigned long en_flags)
2338 schedstat_inc(p, se.statistics.nr_wakeups);
2340 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2342 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2344 schedstat_inc(p, se.statistics.nr_wakeups_local);
2346 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2348 activate_task(rq, p, en_flags);
2351 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2352 int wake_flags, bool success)
2354 trace_sched_wakeup(p, success);
2355 check_preempt_curr(rq, p, wake_flags);
2357 p->state = TASK_RUNNING;
2359 if (p->sched_class->task_woken)
2360 p->sched_class->task_woken(rq, p);
2362 if (unlikely(rq->idle_stamp)) {
2363 u64 delta = rq->clock - rq->idle_stamp;
2364 u64 max = 2*sysctl_sched_migration_cost;
2369 update_avg(&rq->avg_idle, delta);
2373 /* if a worker is waking up, notify workqueue */
2374 if ((p->flags & PF_WQ_WORKER) && success)
2375 wq_worker_waking_up(p, cpu_of(rq));
2379 * try_to_wake_up - wake up a thread
2380 * @p: the thread to be awakened
2381 * @state: the mask of task states that can be woken
2382 * @wake_flags: wake modifier flags (WF_*)
2384 * Put it on the run-queue if it's not already there. The "current"
2385 * thread is always on the run-queue (except when the actual
2386 * re-schedule is in progress), and as such you're allowed to do
2387 * the simpler "current->state = TASK_RUNNING" to mark yourself
2388 * runnable without the overhead of this.
2390 * Returns %true if @p was woken up, %false if it was already running
2391 * or @state didn't match @p's state.
2393 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2396 int cpu, orig_cpu, this_cpu, success = 0;
2397 unsigned long flags;
2398 unsigned long en_flags = ENQUEUE_WAKEUP;
2401 this_cpu = get_cpu();
2404 rq = task_rq_lock(p, &flags);
2405 if (!(p->state & state))
2415 if (unlikely(task_running(rq, p)))
2419 * In order to handle concurrent wakeups and release the rq->lock
2420 * we put the task in TASK_WAKING state.
2422 * First fix up the nr_uninterruptible count:
2424 if (task_contributes_to_load(p)) {
2425 if (likely(cpu_online(orig_cpu)))
2426 rq->nr_uninterruptible--;
2428 this_rq()->nr_uninterruptible--;
2430 p->state = TASK_WAKING;
2432 if (p->sched_class->task_waking) {
2433 p->sched_class->task_waking(rq, p);
2434 en_flags |= ENQUEUE_WAKING;
2437 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2438 if (cpu != orig_cpu)
2439 set_task_cpu(p, cpu);
2440 __task_rq_unlock(rq);
2443 raw_spin_lock(&rq->lock);
2446 * We migrated the task without holding either rq->lock, however
2447 * since the task is not on the task list itself, nobody else
2448 * will try and migrate the task, hence the rq should match the
2449 * cpu we just moved it to.
2451 WARN_ON(task_cpu(p) != cpu);
2452 WARN_ON(p->state != TASK_WAKING);
2454 #ifdef CONFIG_SCHEDSTATS
2455 schedstat_inc(rq, ttwu_count);
2456 if (cpu == this_cpu)
2457 schedstat_inc(rq, ttwu_local);
2459 struct sched_domain *sd;
2460 for_each_domain(this_cpu, sd) {
2461 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2462 schedstat_inc(sd, ttwu_wake_remote);
2467 #endif /* CONFIG_SCHEDSTATS */
2470 #endif /* CONFIG_SMP */
2471 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2472 cpu == this_cpu, en_flags);
2475 ttwu_post_activation(p, rq, wake_flags, success);
2477 task_rq_unlock(rq, &flags);
2484 * try_to_wake_up_local - try to wake up a local task with rq lock held
2485 * @p: the thread to be awakened
2487 * Put @p on the run-queue if it's not already there. The caller must
2488 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2489 * the current task. this_rq() stays locked over invocation.
2491 static void try_to_wake_up_local(struct task_struct *p)
2493 struct rq *rq = task_rq(p);
2494 bool success = false;
2496 BUG_ON(rq != this_rq());
2497 BUG_ON(p == current);
2498 lockdep_assert_held(&rq->lock);
2500 if (!(p->state & TASK_NORMAL))
2504 if (likely(!task_running(rq, p))) {
2505 schedstat_inc(rq, ttwu_count);
2506 schedstat_inc(rq, ttwu_local);
2508 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2511 ttwu_post_activation(p, rq, 0, success);
2515 * wake_up_process - Wake up a specific process
2516 * @p: The process to be woken up.
2518 * Attempt to wake up the nominated process and move it to the set of runnable
2519 * processes. Returns 1 if the process was woken up, 0 if it was already
2522 * It may be assumed that this function implies a write memory barrier before
2523 * changing the task state if and only if any tasks are woken up.
2525 int wake_up_process(struct task_struct *p)
2527 return try_to_wake_up(p, TASK_ALL, 0);
2529 EXPORT_SYMBOL(wake_up_process);
2531 int wake_up_state(struct task_struct *p, unsigned int state)
2533 return try_to_wake_up(p, state, 0);
2537 * Perform scheduler related setup for a newly forked process p.
2538 * p is forked by current.
2540 * __sched_fork() is basic setup used by init_idle() too:
2542 static void __sched_fork(struct task_struct *p)
2544 p->se.exec_start = 0;
2545 p->se.sum_exec_runtime = 0;
2546 p->se.prev_sum_exec_runtime = 0;
2547 p->se.nr_migrations = 0;
2549 #ifdef CONFIG_SCHEDSTATS
2550 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2553 INIT_LIST_HEAD(&p->rt.run_list);
2555 INIT_LIST_HEAD(&p->se.group_node);
2557 #ifdef CONFIG_PREEMPT_NOTIFIERS
2558 INIT_HLIST_HEAD(&p->preempt_notifiers);
2563 * fork()/clone()-time setup:
2565 void sched_fork(struct task_struct *p, int clone_flags)
2567 int cpu = get_cpu();
2571 * We mark the process as running here. This guarantees that
2572 * nobody will actually run it, and a signal or other external
2573 * event cannot wake it up and insert it on the runqueue either.
2575 p->state = TASK_RUNNING;
2578 * Revert to default priority/policy on fork if requested.
2580 if (unlikely(p->sched_reset_on_fork)) {
2581 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2582 p->policy = SCHED_NORMAL;
2583 p->normal_prio = p->static_prio;
2586 if (PRIO_TO_NICE(p->static_prio) < 0) {
2587 p->static_prio = NICE_TO_PRIO(0);
2588 p->normal_prio = p->static_prio;
2593 * We don't need the reset flag anymore after the fork. It has
2594 * fulfilled its duty:
2596 p->sched_reset_on_fork = 0;
2600 * Make sure we do not leak PI boosting priority to the child.
2602 p->prio = current->normal_prio;
2604 if (!rt_prio(p->prio))
2605 p->sched_class = &fair_sched_class;
2607 if (p->sched_class->task_fork)
2608 p->sched_class->task_fork(p);
2611 * The child is not yet in the pid-hash so no cgroup attach races,
2612 * and the cgroup is pinned to this child due to cgroup_fork()
2613 * is ran before sched_fork().
2615 * Silence PROVE_RCU.
2618 set_task_cpu(p, cpu);
2621 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2622 if (likely(sched_info_on()))
2623 memset(&p->sched_info, 0, sizeof(p->sched_info));
2625 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2628 #ifdef CONFIG_PREEMPT
2629 /* Want to start with kernel preemption disabled. */
2630 task_thread_info(p)->preempt_count = 1;
2633 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2640 * wake_up_new_task - wake up a newly created task for the first time.
2642 * This function will do some initial scheduler statistics housekeeping
2643 * that must be done for every newly created context, then puts the task
2644 * on the runqueue and wakes it.
2646 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2648 unsigned long flags;
2650 int cpu __maybe_unused = get_cpu();
2653 rq = task_rq_lock(p, &flags);
2654 p->state = TASK_WAKING;
2657 * Fork balancing, do it here and not earlier because:
2658 * - cpus_allowed can change in the fork path
2659 * - any previously selected cpu might disappear through hotplug
2661 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2662 * without people poking at ->cpus_allowed.
2664 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2665 set_task_cpu(p, cpu);
2667 p->state = TASK_RUNNING;
2668 task_rq_unlock(rq, &flags);
2671 rq = task_rq_lock(p, &flags);
2672 activate_task(rq, p, 0);
2673 trace_sched_wakeup_new(p, 1);
2674 check_preempt_curr(rq, p, WF_FORK);
2676 if (p->sched_class->task_woken)
2677 p->sched_class->task_woken(rq, p);
2679 task_rq_unlock(rq, &flags);
2683 #ifdef CONFIG_PREEMPT_NOTIFIERS
2686 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2687 * @notifier: notifier struct to register
2689 void preempt_notifier_register(struct preempt_notifier *notifier)
2691 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2693 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2696 * preempt_notifier_unregister - no longer interested in preemption notifications
2697 * @notifier: notifier struct to unregister
2699 * This is safe to call from within a preemption notifier.
2701 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2703 hlist_del(¬ifier->link);
2705 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2707 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2709 struct preempt_notifier *notifier;
2710 struct hlist_node *node;
2712 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2713 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2717 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2718 struct task_struct *next)
2720 struct preempt_notifier *notifier;
2721 struct hlist_node *node;
2723 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2724 notifier->ops->sched_out(notifier, next);
2727 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2729 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2734 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2735 struct task_struct *next)
2739 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2742 * prepare_task_switch - prepare to switch tasks
2743 * @rq: the runqueue preparing to switch
2744 * @prev: the current task that is being switched out
2745 * @next: the task we are going to switch to.
2747 * This is called with the rq lock held and interrupts off. It must
2748 * be paired with a subsequent finish_task_switch after the context
2751 * prepare_task_switch sets up locking and calls architecture specific
2755 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2756 struct task_struct *next)
2758 sched_info_switch(prev, next);
2759 perf_event_task_sched_out(prev, next);
2760 fire_sched_out_preempt_notifiers(prev, next);
2761 prepare_lock_switch(rq, next);
2762 prepare_arch_switch(next);
2763 trace_sched_switch(prev, next);
2767 * finish_task_switch - clean up after a task-switch
2768 * @rq: runqueue associated with task-switch
2769 * @prev: the thread we just switched away from.
2771 * finish_task_switch must be called after the context switch, paired
2772 * with a prepare_task_switch call before the context switch.
2773 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2774 * and do any other architecture-specific cleanup actions.
2776 * Note that we may have delayed dropping an mm in context_switch(). If
2777 * so, we finish that here outside of the runqueue lock. (Doing it
2778 * with the lock held can cause deadlocks; see schedule() for
2781 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2782 __releases(rq->lock)
2784 struct mm_struct *mm = rq->prev_mm;
2790 * A task struct has one reference for the use as "current".
2791 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2792 * schedule one last time. The schedule call will never return, and
2793 * the scheduled task must drop that reference.
2794 * The test for TASK_DEAD must occur while the runqueue locks are
2795 * still held, otherwise prev could be scheduled on another cpu, die
2796 * there before we look at prev->state, and then the reference would
2798 * Manfred Spraul <manfred@colorfullife.com>
2800 prev_state = prev->state;
2801 finish_arch_switch(prev);
2802 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2803 local_irq_disable();
2804 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2805 perf_event_task_sched_in(current);
2806 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2808 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2809 finish_lock_switch(rq, prev);
2811 fire_sched_in_preempt_notifiers(current);
2814 if (unlikely(prev_state == TASK_DEAD)) {
2816 * Remove function-return probe instances associated with this
2817 * task and put them back on the free list.
2819 kprobe_flush_task(prev);
2820 put_task_struct(prev);
2826 /* assumes rq->lock is held */
2827 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2829 if (prev->sched_class->pre_schedule)
2830 prev->sched_class->pre_schedule(rq, prev);
2833 /* rq->lock is NOT held, but preemption is disabled */
2834 static inline void post_schedule(struct rq *rq)
2836 if (rq->post_schedule) {
2837 unsigned long flags;
2839 raw_spin_lock_irqsave(&rq->lock, flags);
2840 if (rq->curr->sched_class->post_schedule)
2841 rq->curr->sched_class->post_schedule(rq);
2842 raw_spin_unlock_irqrestore(&rq->lock, flags);
2844 rq->post_schedule = 0;
2850 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2854 static inline void post_schedule(struct rq *rq)
2861 * schedule_tail - first thing a freshly forked thread must call.
2862 * @prev: the thread we just switched away from.
2864 asmlinkage void schedule_tail(struct task_struct *prev)
2865 __releases(rq->lock)
2867 struct rq *rq = this_rq();
2869 finish_task_switch(rq, prev);
2872 * FIXME: do we need to worry about rq being invalidated by the
2877 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2878 /* In this case, finish_task_switch does not reenable preemption */
2881 if (current->set_child_tid)
2882 put_user(task_pid_vnr(current), current->set_child_tid);
2886 * context_switch - switch to the new MM and the new
2887 * thread's register state.
2890 context_switch(struct rq *rq, struct task_struct *prev,
2891 struct task_struct *next)
2893 struct mm_struct *mm, *oldmm;
2895 prepare_task_switch(rq, prev, next);
2898 oldmm = prev->active_mm;
2900 * For paravirt, this is coupled with an exit in switch_to to
2901 * combine the page table reload and the switch backend into
2904 arch_start_context_switch(prev);
2907 next->active_mm = oldmm;
2908 atomic_inc(&oldmm->mm_count);
2909 enter_lazy_tlb(oldmm, next);
2911 switch_mm(oldmm, mm, next);
2914 prev->active_mm = NULL;
2915 rq->prev_mm = oldmm;
2918 * Since the runqueue lock will be released by the next
2919 * task (which is an invalid locking op but in the case
2920 * of the scheduler it's an obvious special-case), so we
2921 * do an early lockdep release here:
2923 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2924 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2927 /* Here we just switch the register state and the stack. */
2928 switch_to(prev, next, prev);
2932 * this_rq must be evaluated again because prev may have moved
2933 * CPUs since it called schedule(), thus the 'rq' on its stack
2934 * frame will be invalid.
2936 finish_task_switch(this_rq(), prev);
2940 * nr_running, nr_uninterruptible and nr_context_switches:
2942 * externally visible scheduler statistics: current number of runnable
2943 * threads, current number of uninterruptible-sleeping threads, total
2944 * number of context switches performed since bootup.
2946 unsigned long nr_running(void)
2948 unsigned long i, sum = 0;
2950 for_each_online_cpu(i)
2951 sum += cpu_rq(i)->nr_running;
2956 unsigned long nr_uninterruptible(void)
2958 unsigned long i, sum = 0;
2960 for_each_possible_cpu(i)
2961 sum += cpu_rq(i)->nr_uninterruptible;
2964 * Since we read the counters lockless, it might be slightly
2965 * inaccurate. Do not allow it to go below zero though:
2967 if (unlikely((long)sum < 0))
2973 unsigned long long nr_context_switches(void)
2976 unsigned long long sum = 0;
2978 for_each_possible_cpu(i)
2979 sum += cpu_rq(i)->nr_switches;
2984 unsigned long nr_iowait(void)
2986 unsigned long i, sum = 0;
2988 for_each_possible_cpu(i)
2989 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2994 unsigned long nr_iowait_cpu(int cpu)
2996 struct rq *this = cpu_rq(cpu);
2997 return atomic_read(&this->nr_iowait);
3000 unsigned long this_cpu_load(void)
3002 struct rq *this = this_rq();
3003 return this->cpu_load[0];
3007 /* Variables and functions for calc_load */
3008 static atomic_long_t calc_load_tasks;
3009 static unsigned long calc_load_update;
3010 unsigned long avenrun[3];
3011 EXPORT_SYMBOL(avenrun);
3013 static long calc_load_fold_active(struct rq *this_rq)
3015 long nr_active, delta = 0;
3017 nr_active = this_rq->nr_running;
3018 nr_active += (long) this_rq->nr_uninterruptible;
3020 if (nr_active != this_rq->calc_load_active) {
3021 delta = nr_active - this_rq->calc_load_active;
3022 this_rq->calc_load_active = nr_active;
3028 static unsigned long
3029 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3032 load += active * (FIXED_1 - exp);
3033 load += 1UL << (FSHIFT - 1);
3034 return load >> FSHIFT;
3039 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3041 * When making the ILB scale, we should try to pull this in as well.
3043 static atomic_long_t calc_load_tasks_idle;
3045 static void calc_load_account_idle(struct rq *this_rq)
3049 delta = calc_load_fold_active(this_rq);
3051 atomic_long_add(delta, &calc_load_tasks_idle);
3054 static long calc_load_fold_idle(void)
3059 * Its got a race, we don't care...
3061 if (atomic_long_read(&calc_load_tasks_idle))
3062 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3068 * fixed_power_int - compute: x^n, in O(log n) time
3070 * @x: base of the power
3071 * @frac_bits: fractional bits of @x
3072 * @n: power to raise @x to.
3074 * By exploiting the relation between the definition of the natural power
3075 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3076 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3077 * (where: n_i \elem {0, 1}, the binary vector representing n),
3078 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3079 * of course trivially computable in O(log_2 n), the length of our binary
3082 static unsigned long
3083 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3085 unsigned long result = 1UL << frac_bits;
3090 result += 1UL << (frac_bits - 1);
3091 result >>= frac_bits;
3097 x += 1UL << (frac_bits - 1);
3105 * a1 = a0 * e + a * (1 - e)
3107 * a2 = a1 * e + a * (1 - e)
3108 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3109 * = a0 * e^2 + a * (1 - e) * (1 + e)
3111 * a3 = a2 * e + a * (1 - e)
3112 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3113 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3117 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3118 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3119 * = a0 * e^n + a * (1 - e^n)
3121 * [1] application of the geometric series:
3124 * S_n := \Sum x^i = -------------
3127 static unsigned long
3128 calc_load_n(unsigned long load, unsigned long exp,
3129 unsigned long active, unsigned int n)
3132 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3136 * NO_HZ can leave us missing all per-cpu ticks calling
3137 * calc_load_account_active(), but since an idle CPU folds its delta into
3138 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3139 * in the pending idle delta if our idle period crossed a load cycle boundary.
3141 * Once we've updated the global active value, we need to apply the exponential
3142 * weights adjusted to the number of cycles missed.
3144 static void calc_global_nohz(unsigned long ticks)
3146 long delta, active, n;
3148 if (time_before(jiffies, calc_load_update))
3152 * If we crossed a calc_load_update boundary, make sure to fold
3153 * any pending idle changes, the respective CPUs might have
3154 * missed the tick driven calc_load_account_active() update
3157 delta = calc_load_fold_idle();
3159 atomic_long_add(delta, &calc_load_tasks);
3162 * If we were idle for multiple load cycles, apply them.
3164 if (ticks >= LOAD_FREQ) {
3165 n = ticks / LOAD_FREQ;
3167 active = atomic_long_read(&calc_load_tasks);
3168 active = active > 0 ? active * FIXED_1 : 0;
3170 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3171 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3172 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3174 calc_load_update += n * LOAD_FREQ;
3178 * Its possible the remainder of the above division also crosses
3179 * a LOAD_FREQ period, the regular check in calc_global_load()
3180 * which comes after this will take care of that.
3182 * Consider us being 11 ticks before a cycle completion, and us
3183 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3184 * age us 4 cycles, and the test in calc_global_load() will
3185 * pick up the final one.
3189 static void calc_load_account_idle(struct rq *this_rq)
3193 static inline long calc_load_fold_idle(void)
3198 static void calc_global_nohz(unsigned long ticks)
3204 * get_avenrun - get the load average array
3205 * @loads: pointer to dest load array
3206 * @offset: offset to add
3207 * @shift: shift count to shift the result left
3209 * These values are estimates at best, so no need for locking.
3211 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3213 loads[0] = (avenrun[0] + offset) << shift;
3214 loads[1] = (avenrun[1] + offset) << shift;
3215 loads[2] = (avenrun[2] + offset) << shift;
3219 * calc_load - update the avenrun load estimates 10 ticks after the
3220 * CPUs have updated calc_load_tasks.
3222 void calc_global_load(unsigned long ticks)
3226 calc_global_nohz(ticks);
3228 if (time_before(jiffies, calc_load_update + 10))
3231 active = atomic_long_read(&calc_load_tasks);
3232 active = active > 0 ? active * FIXED_1 : 0;
3234 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3235 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3236 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3238 calc_load_update += LOAD_FREQ;
3242 * Called from update_cpu_load() to periodically update this CPU's
3245 static void calc_load_account_active(struct rq *this_rq)
3249 if (time_before(jiffies, this_rq->calc_load_update))
3252 delta = calc_load_fold_active(this_rq);
3253 delta += calc_load_fold_idle();
3255 atomic_long_add(delta, &calc_load_tasks);
3257 this_rq->calc_load_update += LOAD_FREQ;
3261 * The exact cpuload at various idx values, calculated at every tick would be
3262 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3264 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3265 * on nth tick when cpu may be busy, then we have:
3266 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3267 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3269 * decay_load_missed() below does efficient calculation of
3270 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3271 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3273 * The calculation is approximated on a 128 point scale.
3274 * degrade_zero_ticks is the number of ticks after which load at any
3275 * particular idx is approximated to be zero.
3276 * degrade_factor is a precomputed table, a row for each load idx.
3277 * Each column corresponds to degradation factor for a power of two ticks,
3278 * based on 128 point scale.
3280 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3281 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3283 * With this power of 2 load factors, we can degrade the load n times
3284 * by looking at 1 bits in n and doing as many mult/shift instead of
3285 * n mult/shifts needed by the exact degradation.
3287 #define DEGRADE_SHIFT 7
3288 static const unsigned char
3289 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3290 static const unsigned char
3291 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3292 {0, 0, 0, 0, 0, 0, 0, 0},
3293 {64, 32, 8, 0, 0, 0, 0, 0},
3294 {96, 72, 40, 12, 1, 0, 0},
3295 {112, 98, 75, 43, 15, 1, 0},
3296 {120, 112, 98, 76, 45, 16, 2} };
3299 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3300 * would be when CPU is idle and so we just decay the old load without
3301 * adding any new load.
3303 static unsigned long
3304 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3308 if (!missed_updates)
3311 if (missed_updates >= degrade_zero_ticks[idx])
3315 return load >> missed_updates;
3317 while (missed_updates) {
3318 if (missed_updates % 2)
3319 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3321 missed_updates >>= 1;
3328 * Update rq->cpu_load[] statistics. This function is usually called every
3329 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3330 * every tick. We fix it up based on jiffies.
3332 static void update_cpu_load(struct rq *this_rq)
3334 unsigned long this_load = this_rq->load.weight;
3335 unsigned long curr_jiffies = jiffies;
3336 unsigned long pending_updates;
3339 this_rq->nr_load_updates++;
3341 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3342 if (curr_jiffies == this_rq->last_load_update_tick)
3345 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3346 this_rq->last_load_update_tick = curr_jiffies;
3348 /* Update our load: */
3349 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3350 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3351 unsigned long old_load, new_load;
3353 /* scale is effectively 1 << i now, and >> i divides by scale */
3355 old_load = this_rq->cpu_load[i];
3356 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3357 new_load = this_load;
3359 * Round up the averaging division if load is increasing. This
3360 * prevents us from getting stuck on 9 if the load is 10, for
3363 if (new_load > old_load)
3364 new_load += scale - 1;
3366 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3369 sched_avg_update(this_rq);
3372 static void update_cpu_load_active(struct rq *this_rq)
3374 update_cpu_load(this_rq);
3376 calc_load_account_active(this_rq);
3382 * sched_exec - execve() is a valuable balancing opportunity, because at
3383 * this point the task has the smallest effective memory and cache footprint.
3385 void sched_exec(void)
3387 struct task_struct *p = current;
3388 unsigned long flags;
3392 rq = task_rq_lock(p, &flags);
3393 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3394 if (dest_cpu == smp_processor_id())
3398 * select_task_rq() can race against ->cpus_allowed
3400 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3401 likely(cpu_active(dest_cpu)) && migrate_task(p, rq)) {
3402 struct migration_arg arg = { p, dest_cpu };
3404 task_rq_unlock(rq, &flags);
3405 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3409 task_rq_unlock(rq, &flags);
3414 DEFINE_PER_CPU(struct kernel_stat, kstat);
3416 EXPORT_PER_CPU_SYMBOL(kstat);
3419 * Return any ns on the sched_clock that have not yet been accounted in
3420 * @p in case that task is currently running.
3422 * Called with task_rq_lock() held on @rq.
3424 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3428 if (task_current(rq, p)) {
3429 update_rq_clock(rq);
3430 ns = rq->clock_task - p->se.exec_start;
3438 unsigned long long task_delta_exec(struct task_struct *p)
3440 unsigned long flags;
3444 rq = task_rq_lock(p, &flags);
3445 ns = do_task_delta_exec(p, rq);
3446 task_rq_unlock(rq, &flags);
3452 * Return accounted runtime for the task.
3453 * In case the task is currently running, return the runtime plus current's
3454 * pending runtime that have not been accounted yet.
3456 unsigned long long task_sched_runtime(struct task_struct *p)
3458 unsigned long flags;
3462 rq = task_rq_lock(p, &flags);
3463 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3464 task_rq_unlock(rq, &flags);
3470 * Return sum_exec_runtime for the thread group.
3471 * In case the task is currently running, return the sum plus current's
3472 * pending runtime that have not been accounted yet.
3474 * Note that the thread group might have other running tasks as well,
3475 * so the return value not includes other pending runtime that other
3476 * running tasks might have.
3478 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3480 struct task_cputime totals;
3481 unsigned long flags;
3485 rq = task_rq_lock(p, &flags);
3486 thread_group_cputime(p, &totals);
3487 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3488 task_rq_unlock(rq, &flags);
3494 * Account user cpu time to a process.
3495 * @p: the process that the cpu time gets accounted to
3496 * @cputime: the cpu time spent in user space since the last update
3497 * @cputime_scaled: cputime scaled by cpu frequency
3499 void account_user_time(struct task_struct *p, cputime_t cputime,
3500 cputime_t cputime_scaled)
3502 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3505 /* Add user time to process. */
3506 p->utime = cputime_add(p->utime, cputime);
3507 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3508 account_group_user_time(p, cputime);
3510 /* Add user time to cpustat. */
3511 tmp = cputime_to_cputime64(cputime);
3512 if (TASK_NICE(p) > 0)
3513 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3515 cpustat->user = cputime64_add(cpustat->user, tmp);
3517 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3518 /* Account for user time used */
3519 acct_update_integrals(p);
3523 * Account guest cpu time to a process.
3524 * @p: the process that the cpu time gets accounted to
3525 * @cputime: the cpu time spent in virtual machine since the last update
3526 * @cputime_scaled: cputime scaled by cpu frequency
3528 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3529 cputime_t cputime_scaled)
3532 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3534 tmp = cputime_to_cputime64(cputime);
3536 /* Add guest time to process. */
3537 p->utime = cputime_add(p->utime, cputime);
3538 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3539 account_group_user_time(p, cputime);
3540 p->gtime = cputime_add(p->gtime, cputime);
3542 /* Add guest time to cpustat. */
3543 if (TASK_NICE(p) > 0) {
3544 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3545 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3547 cpustat->user = cputime64_add(cpustat->user, tmp);
3548 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3553 * Account system cpu time to a process.
3554 * @p: the process that the cpu time gets accounted to
3555 * @hardirq_offset: the offset to subtract from hardirq_count()
3556 * @cputime: the cpu time spent in kernel space since the last update
3557 * @cputime_scaled: cputime scaled by cpu frequency
3559 void account_system_time(struct task_struct *p, int hardirq_offset,
3560 cputime_t cputime, cputime_t cputime_scaled)
3562 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3565 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3566 account_guest_time(p, cputime, cputime_scaled);
3570 /* Add system time to process. */
3571 p->stime = cputime_add(p->stime, cputime);
3572 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3573 account_group_system_time(p, cputime);
3575 /* Add system time to cpustat. */
3576 tmp = cputime_to_cputime64(cputime);
3577 if (hardirq_count() - hardirq_offset)
3578 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3579 else if (in_serving_softirq())
3580 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3582 cpustat->system = cputime64_add(cpustat->system, tmp);
3584 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3586 /* Account for system time used */
3587 acct_update_integrals(p);
3591 * Account for involuntary wait time.
3592 * @steal: the cpu time spent in involuntary wait
3594 void account_steal_time(cputime_t cputime)
3596 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3597 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3599 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3603 * Account for idle time.
3604 * @cputime: the cpu time spent in idle wait
3606 void account_idle_time(cputime_t cputime)
3608 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3609 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3610 struct rq *rq = this_rq();
3612 if (atomic_read(&rq->nr_iowait) > 0)
3613 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3615 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3618 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3621 * Account a single tick of cpu time.
3622 * @p: the process that the cpu time gets accounted to
3623 * @user_tick: indicates if the tick is a user or a system tick
3625 void account_process_tick(struct task_struct *p, int user_tick)
3627 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3628 struct rq *rq = this_rq();
3631 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3632 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3633 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3636 account_idle_time(cputime_one_jiffy);
3640 * Account multiple ticks of steal time.
3641 * @p: the process from which the cpu time has been stolen
3642 * @ticks: number of stolen ticks
3644 void account_steal_ticks(unsigned long ticks)
3646 account_steal_time(jiffies_to_cputime(ticks));
3650 * Account multiple ticks of idle time.
3651 * @ticks: number of stolen ticks
3653 void account_idle_ticks(unsigned long ticks)
3655 account_idle_time(jiffies_to_cputime(ticks));
3661 * Use precise platform statistics if available:
3663 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3664 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3670 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3672 struct task_cputime cputime;
3674 thread_group_cputime(p, &cputime);
3676 *ut = cputime.utime;
3677 *st = cputime.stime;
3681 #ifndef nsecs_to_cputime
3682 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3685 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3687 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3690 * Use CFS's precise accounting:
3692 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3698 do_div(temp, total);
3699 utime = (cputime_t)temp;
3704 * Compare with previous values, to keep monotonicity:
3706 p->prev_utime = max(p->prev_utime, utime);
3707 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3709 *ut = p->prev_utime;
3710 *st = p->prev_stime;
3714 * Must be called with siglock held.
3716 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3718 struct signal_struct *sig = p->signal;
3719 struct task_cputime cputime;
3720 cputime_t rtime, utime, total;
3722 thread_group_cputime(p, &cputime);
3724 total = cputime_add(cputime.utime, cputime.stime);
3725 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3730 temp *= cputime.utime;
3731 do_div(temp, total);
3732 utime = (cputime_t)temp;
3736 sig->prev_utime = max(sig->prev_utime, utime);
3737 sig->prev_stime = max(sig->prev_stime,
3738 cputime_sub(rtime, sig->prev_utime));
3740 *ut = sig->prev_utime;
3741 *st = sig->prev_stime;
3746 * This function gets called by the timer code, with HZ frequency.
3747 * We call it with interrupts disabled.
3749 * It also gets called by the fork code, when changing the parent's
3752 void scheduler_tick(void)
3754 int cpu = smp_processor_id();
3755 struct rq *rq = cpu_rq(cpu);
3756 struct task_struct *curr = rq->curr;
3760 raw_spin_lock(&rq->lock);
3761 update_rq_clock(rq);
3762 update_cpu_load_active(rq);
3763 curr->sched_class->task_tick(rq, curr, 0);
3764 raw_spin_unlock(&rq->lock);
3766 perf_event_task_tick();
3769 rq->idle_at_tick = idle_cpu(cpu);
3770 trigger_load_balance(rq, cpu);
3774 notrace unsigned long get_parent_ip(unsigned long addr)
3776 if (in_lock_functions(addr)) {
3777 addr = CALLER_ADDR2;
3778 if (in_lock_functions(addr))
3779 addr = CALLER_ADDR3;
3784 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3785 defined(CONFIG_PREEMPT_TRACER))
3787 void __kprobes add_preempt_count(int val)
3789 #ifdef CONFIG_DEBUG_PREEMPT
3793 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3796 preempt_count() += val;
3797 #ifdef CONFIG_DEBUG_PREEMPT
3799 * Spinlock count overflowing soon?
3801 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3804 if (preempt_count() == val)
3805 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3807 EXPORT_SYMBOL(add_preempt_count);
3809 void __kprobes sub_preempt_count(int val)
3811 #ifdef CONFIG_DEBUG_PREEMPT
3815 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3818 * Is the spinlock portion underflowing?
3820 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3821 !(preempt_count() & PREEMPT_MASK)))
3825 if (preempt_count() == val)
3826 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3827 preempt_count() -= val;
3829 EXPORT_SYMBOL(sub_preempt_count);
3834 * Print scheduling while atomic bug:
3836 static noinline void __schedule_bug(struct task_struct *prev)
3838 struct pt_regs *regs = get_irq_regs();
3840 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3841 prev->comm, prev->pid, preempt_count());
3843 debug_show_held_locks(prev);
3845 if (irqs_disabled())
3846 print_irqtrace_events(prev);
3855 * Various schedule()-time debugging checks and statistics:
3857 static inline void schedule_debug(struct task_struct *prev)
3860 * Test if we are atomic. Since do_exit() needs to call into
3861 * schedule() atomically, we ignore that path for now.
3862 * Otherwise, whine if we are scheduling when we should not be.
3864 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3865 __schedule_bug(prev);
3867 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3869 schedstat_inc(this_rq(), sched_count);
3870 #ifdef CONFIG_SCHEDSTATS
3871 if (unlikely(prev->lock_depth >= 0)) {
3872 schedstat_inc(this_rq(), rq_sched_info.bkl_count);
3873 schedstat_inc(prev, sched_info.bkl_count);
3878 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3881 update_rq_clock(rq);
3882 prev->sched_class->put_prev_task(rq, prev);
3886 * Pick up the highest-prio task:
3888 static inline struct task_struct *
3889 pick_next_task(struct rq *rq)
3891 const struct sched_class *class;
3892 struct task_struct *p;
3895 * Optimization: we know that if all tasks are in
3896 * the fair class we can call that function directly:
3898 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3899 p = fair_sched_class.pick_next_task(rq);
3904 for_each_class(class) {
3905 p = class->pick_next_task(rq);
3910 BUG(); /* the idle class will always have a runnable task */
3914 * schedule() is the main scheduler function.
3916 asmlinkage void __sched schedule(void)
3918 struct task_struct *prev, *next;
3919 unsigned long *switch_count;
3925 cpu = smp_processor_id();
3927 rcu_note_context_switch(cpu);
3930 release_kernel_lock(prev);
3931 need_resched_nonpreemptible:
3933 schedule_debug(prev);
3935 if (sched_feat(HRTICK))
3938 raw_spin_lock_irq(&rq->lock);
3940 switch_count = &prev->nivcsw;
3941 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3942 if (unlikely(signal_pending_state(prev->state, prev))) {
3943 prev->state = TASK_RUNNING;
3946 * If a worker is going to sleep, notify and
3947 * ask workqueue whether it wants to wake up a
3948 * task to maintain concurrency. If so, wake
3951 if (prev->flags & PF_WQ_WORKER) {
3952 struct task_struct *to_wakeup;
3954 to_wakeup = wq_worker_sleeping(prev, cpu);
3956 try_to_wake_up_local(to_wakeup);
3958 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3960 switch_count = &prev->nvcsw;
3963 pre_schedule(rq, prev);
3965 if (unlikely(!rq->nr_running))
3966 idle_balance(cpu, rq);
3968 put_prev_task(rq, prev);
3969 next = pick_next_task(rq);
3970 clear_tsk_need_resched(prev);
3971 rq->skip_clock_update = 0;
3973 if (likely(prev != next)) {
3978 context_switch(rq, prev, next); /* unlocks the rq */
3980 * The context switch have flipped the stack from under us
3981 * and restored the local variables which were saved when
3982 * this task called schedule() in the past. prev == current
3983 * is still correct, but it can be moved to another cpu/rq.
3985 cpu = smp_processor_id();
3988 raw_spin_unlock_irq(&rq->lock);
3992 if (unlikely(reacquire_kernel_lock(prev)))
3993 goto need_resched_nonpreemptible;
3995 preempt_enable_no_resched();
3999 EXPORT_SYMBOL(schedule);
4001 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4003 * Look out! "owner" is an entirely speculative pointer
4004 * access and not reliable.
4006 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
4011 if (!sched_feat(OWNER_SPIN))
4014 #ifdef CONFIG_DEBUG_PAGEALLOC
4016 * Need to access the cpu field knowing that
4017 * DEBUG_PAGEALLOC could have unmapped it if
4018 * the mutex owner just released it and exited.
4020 if (probe_kernel_address(&owner->cpu, cpu))
4027 * Even if the access succeeded (likely case),
4028 * the cpu field may no longer be valid.
4030 if (cpu >= nr_cpumask_bits)
4034 * We need to validate that we can do a
4035 * get_cpu() and that we have the percpu area.
4037 if (!cpu_online(cpu))
4044 * Owner changed, break to re-assess state.
4046 if (lock->owner != owner) {
4048 * If the lock has switched to a different owner,
4049 * we likely have heavy contention. Return 0 to quit
4050 * optimistic spinning and not contend further:
4058 * Is that owner really running on that cpu?
4060 if (task_thread_info(rq->curr) != owner || need_resched())
4063 arch_mutex_cpu_relax();
4070 #ifdef CONFIG_PREEMPT
4072 * this is the entry point to schedule() from in-kernel preemption
4073 * off of preempt_enable. Kernel preemptions off return from interrupt
4074 * occur there and call schedule directly.
4076 asmlinkage void __sched notrace preempt_schedule(void)
4078 struct thread_info *ti = current_thread_info();
4081 * If there is a non-zero preempt_count or interrupts are disabled,
4082 * we do not want to preempt the current task. Just return..
4084 if (likely(ti->preempt_count || irqs_disabled()))
4088 add_preempt_count_notrace(PREEMPT_ACTIVE);
4090 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4093 * Check again in case we missed a preemption opportunity
4094 * between schedule and now.
4097 } while (need_resched());
4099 EXPORT_SYMBOL(preempt_schedule);
4102 * this is the entry point to schedule() from kernel preemption
4103 * off of irq context.
4104 * Note, that this is called and return with irqs disabled. This will
4105 * protect us against recursive calling from irq.
4107 asmlinkage void __sched preempt_schedule_irq(void)
4109 struct thread_info *ti = current_thread_info();
4111 /* Catch callers which need to be fixed */
4112 BUG_ON(ti->preempt_count || !irqs_disabled());
4115 add_preempt_count(PREEMPT_ACTIVE);
4118 local_irq_disable();
4119 sub_preempt_count(PREEMPT_ACTIVE);
4122 * Check again in case we missed a preemption opportunity
4123 * between schedule and now.
4126 } while (need_resched());
4129 #endif /* CONFIG_PREEMPT */
4131 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4134 return try_to_wake_up(curr->private, mode, wake_flags);
4136 EXPORT_SYMBOL(default_wake_function);
4139 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4140 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4141 * number) then we wake all the non-exclusive tasks and one exclusive task.
4143 * There are circumstances in which we can try to wake a task which has already
4144 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4145 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4147 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4148 int nr_exclusive, int wake_flags, void *key)
4150 wait_queue_t *curr, *next;
4152 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4153 unsigned flags = curr->flags;
4155 if (curr->func(curr, mode, wake_flags, key) &&
4156 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4162 * __wake_up - wake up threads blocked on a waitqueue.
4164 * @mode: which threads
4165 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4166 * @key: is directly passed to the wakeup function
4168 * It may be assumed that this function implies a write memory barrier before
4169 * changing the task state if and only if any tasks are woken up.
4171 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4172 int nr_exclusive, void *key)
4174 unsigned long flags;
4176 spin_lock_irqsave(&q->lock, flags);
4177 __wake_up_common(q, mode, nr_exclusive, 0, key);
4178 spin_unlock_irqrestore(&q->lock, flags);
4180 EXPORT_SYMBOL(__wake_up);
4183 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4185 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4187 __wake_up_common(q, mode, 1, 0, NULL);
4189 EXPORT_SYMBOL_GPL(__wake_up_locked);
4191 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4193 __wake_up_common(q, mode, 1, 0, key);
4197 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4199 * @mode: which threads
4200 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4201 * @key: opaque value to be passed to wakeup targets
4203 * The sync wakeup differs that the waker knows that it will schedule
4204 * away soon, so while the target thread will be woken up, it will not
4205 * be migrated to another CPU - ie. the two threads are 'synchronized'
4206 * with each other. This can prevent needless bouncing between CPUs.
4208 * On UP it can prevent extra preemption.
4210 * It may be assumed that this function implies a write memory barrier before
4211 * changing the task state if and only if any tasks are woken up.
4213 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4214 int nr_exclusive, void *key)
4216 unsigned long flags;
4217 int wake_flags = WF_SYNC;
4222 if (unlikely(!nr_exclusive))
4225 spin_lock_irqsave(&q->lock, flags);
4226 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4227 spin_unlock_irqrestore(&q->lock, flags);
4229 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4232 * __wake_up_sync - see __wake_up_sync_key()
4234 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4236 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4238 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4241 * complete: - signals a single thread waiting on this completion
4242 * @x: holds the state of this particular completion
4244 * This will wake up a single thread waiting on this completion. Threads will be
4245 * awakened in the same order in which they were queued.
4247 * See also complete_all(), wait_for_completion() and related routines.
4249 * It may be assumed that this function implies a write memory barrier before
4250 * changing the task state if and only if any tasks are woken up.
4252 void complete(struct completion *x)
4254 unsigned long flags;
4256 spin_lock_irqsave(&x->wait.lock, flags);
4258 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4259 spin_unlock_irqrestore(&x->wait.lock, flags);
4261 EXPORT_SYMBOL(complete);
4264 * complete_all: - signals all threads waiting on this completion
4265 * @x: holds the state of this particular completion
4267 * This will wake up all threads waiting on this particular completion event.
4269 * It may be assumed that this function implies a write memory barrier before
4270 * changing the task state if and only if any tasks are woken up.
4272 void complete_all(struct completion *x)
4274 unsigned long flags;
4276 spin_lock_irqsave(&x->wait.lock, flags);
4277 x->done += UINT_MAX/2;
4278 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4279 spin_unlock_irqrestore(&x->wait.lock, flags);
4281 EXPORT_SYMBOL(complete_all);
4283 static inline long __sched
4284 do_wait_for_common(struct completion *x, long timeout, int state)
4287 DECLARE_WAITQUEUE(wait, current);
4289 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4291 if (signal_pending_state(state, current)) {
4292 timeout = -ERESTARTSYS;
4295 __set_current_state(state);
4296 spin_unlock_irq(&x->wait.lock);
4297 timeout = schedule_timeout(timeout);
4298 spin_lock_irq(&x->wait.lock);
4299 } while (!x->done && timeout);
4300 __remove_wait_queue(&x->wait, &wait);
4305 return timeout ?: 1;
4309 wait_for_common(struct completion *x, long timeout, int state)
4313 spin_lock_irq(&x->wait.lock);
4314 timeout = do_wait_for_common(x, timeout, state);
4315 spin_unlock_irq(&x->wait.lock);
4320 * wait_for_completion: - waits for completion of a task
4321 * @x: holds the state of this particular completion
4323 * This waits to be signaled for completion of a specific task. It is NOT
4324 * interruptible and there is no timeout.
4326 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4327 * and interrupt capability. Also see complete().
4329 void __sched wait_for_completion(struct completion *x)
4331 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4333 EXPORT_SYMBOL(wait_for_completion);
4336 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4337 * @x: holds the state of this particular completion
4338 * @timeout: timeout value in jiffies
4340 * This waits for either a completion of a specific task to be signaled or for a
4341 * specified timeout to expire. The timeout is in jiffies. It is not
4344 unsigned long __sched
4345 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4347 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4349 EXPORT_SYMBOL(wait_for_completion_timeout);
4352 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4353 * @x: holds the state of this particular completion
4355 * This waits for completion of a specific task to be signaled. It is
4358 int __sched wait_for_completion_interruptible(struct completion *x)
4360 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4361 if (t == -ERESTARTSYS)
4365 EXPORT_SYMBOL(wait_for_completion_interruptible);
4368 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4369 * @x: holds the state of this particular completion
4370 * @timeout: timeout value in jiffies
4372 * This waits for either a completion of a specific task to be signaled or for a
4373 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4376 wait_for_completion_interruptible_timeout(struct completion *x,
4377 unsigned long timeout)
4379 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4381 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4384 * wait_for_completion_killable: - waits for completion of a task (killable)
4385 * @x: holds the state of this particular completion
4387 * This waits to be signaled for completion of a specific task. It can be
4388 * interrupted by a kill signal.
4390 int __sched wait_for_completion_killable(struct completion *x)
4392 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4393 if (t == -ERESTARTSYS)
4397 EXPORT_SYMBOL(wait_for_completion_killable);
4400 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4401 * @x: holds the state of this particular completion
4402 * @timeout: timeout value in jiffies
4404 * This waits for either a completion of a specific task to be
4405 * signaled or for a specified timeout to expire. It can be
4406 * interrupted by a kill signal. The timeout is in jiffies.
4409 wait_for_completion_killable_timeout(struct completion *x,
4410 unsigned long timeout)
4412 return wait_for_common(x, timeout, TASK_KILLABLE);
4414 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4417 * try_wait_for_completion - try to decrement a completion without blocking
4418 * @x: completion structure
4420 * Returns: 0 if a decrement cannot be done without blocking
4421 * 1 if a decrement succeeded.
4423 * If a completion is being used as a counting completion,
4424 * attempt to decrement the counter without blocking. This
4425 * enables us to avoid waiting if the resource the completion
4426 * is protecting is not available.
4428 bool try_wait_for_completion(struct completion *x)
4430 unsigned long flags;
4433 spin_lock_irqsave(&x->wait.lock, flags);
4438 spin_unlock_irqrestore(&x->wait.lock, flags);
4441 EXPORT_SYMBOL(try_wait_for_completion);
4444 * completion_done - Test to see if a completion has any waiters
4445 * @x: completion structure
4447 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4448 * 1 if there are no waiters.
4451 bool completion_done(struct completion *x)
4453 unsigned long flags;
4456 spin_lock_irqsave(&x->wait.lock, flags);
4459 spin_unlock_irqrestore(&x->wait.lock, flags);
4462 EXPORT_SYMBOL(completion_done);
4465 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4467 unsigned long flags;
4470 init_waitqueue_entry(&wait, current);
4472 __set_current_state(state);
4474 spin_lock_irqsave(&q->lock, flags);
4475 __add_wait_queue(q, &wait);
4476 spin_unlock(&q->lock);
4477 timeout = schedule_timeout(timeout);
4478 spin_lock_irq(&q->lock);
4479 __remove_wait_queue(q, &wait);
4480 spin_unlock_irqrestore(&q->lock, flags);
4485 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4487 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4489 EXPORT_SYMBOL(interruptible_sleep_on);
4492 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4494 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4496 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4498 void __sched sleep_on(wait_queue_head_t *q)
4500 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4502 EXPORT_SYMBOL(sleep_on);
4504 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4506 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4508 EXPORT_SYMBOL(sleep_on_timeout);
4510 #ifdef CONFIG_RT_MUTEXES
4513 * rt_mutex_setprio - set the current priority of a task
4515 * @prio: prio value (kernel-internal form)
4517 * This function changes the 'effective' priority of a task. It does
4518 * not touch ->normal_prio like __setscheduler().
4520 * Used by the rt_mutex code to implement priority inheritance logic.
4522 void rt_mutex_setprio(struct task_struct *p, int prio)
4524 unsigned long flags;
4525 int oldprio, on_rq, running;
4527 const struct sched_class *prev_class;
4529 BUG_ON(prio < 0 || prio > MAX_PRIO);
4531 rq = task_rq_lock(p, &flags);
4533 trace_sched_pi_setprio(p, prio);
4535 prev_class = p->sched_class;
4536 on_rq = p->se.on_rq;
4537 running = task_current(rq, p);
4539 dequeue_task(rq, p, 0);
4541 p->sched_class->put_prev_task(rq, p);
4544 p->sched_class = &rt_sched_class;
4546 p->sched_class = &fair_sched_class;
4551 p->sched_class->set_curr_task(rq);
4553 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4555 check_class_changed(rq, p, prev_class, oldprio, running);
4557 task_rq_unlock(rq, &flags);
4562 void set_user_nice(struct task_struct *p, long nice)
4564 int old_prio, delta, on_rq;
4565 unsigned long flags;
4568 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4571 * We have to be careful, if called from sys_setpriority(),
4572 * the task might be in the middle of scheduling on another CPU.
4574 rq = task_rq_lock(p, &flags);
4576 * The RT priorities are set via sched_setscheduler(), but we still
4577 * allow the 'normal' nice value to be set - but as expected
4578 * it wont have any effect on scheduling until the task is
4579 * SCHED_FIFO/SCHED_RR:
4581 if (task_has_rt_policy(p)) {
4582 p->static_prio = NICE_TO_PRIO(nice);
4585 on_rq = p->se.on_rq;
4587 dequeue_task(rq, p, 0);
4589 p->static_prio = NICE_TO_PRIO(nice);
4592 p->prio = effective_prio(p);
4593 delta = p->prio - old_prio;
4596 enqueue_task(rq, p, 0);
4598 * If the task increased its priority or is running and
4599 * lowered its priority, then reschedule its CPU:
4601 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4602 resched_task(rq->curr);
4605 task_rq_unlock(rq, &flags);
4607 EXPORT_SYMBOL(set_user_nice);
4610 * can_nice - check if a task can reduce its nice value
4614 int can_nice(const struct task_struct *p, const int nice)
4616 /* convert nice value [19,-20] to rlimit style value [1,40] */
4617 int nice_rlim = 20 - nice;
4619 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4620 capable(CAP_SYS_NICE));
4623 #ifdef __ARCH_WANT_SYS_NICE
4626 * sys_nice - change the priority of the current process.
4627 * @increment: priority increment
4629 * sys_setpriority is a more generic, but much slower function that
4630 * does similar things.
4632 SYSCALL_DEFINE1(nice, int, increment)
4637 * Setpriority might change our priority at the same moment.
4638 * We don't have to worry. Conceptually one call occurs first
4639 * and we have a single winner.
4641 if (increment < -40)
4646 nice = TASK_NICE(current) + increment;
4652 if (increment < 0 && !can_nice(current, nice))
4655 retval = security_task_setnice(current, nice);
4659 set_user_nice(current, nice);
4666 * task_prio - return the priority value of a given task.
4667 * @p: the task in question.
4669 * This is the priority value as seen by users in /proc.
4670 * RT tasks are offset by -200. Normal tasks are centered
4671 * around 0, value goes from -16 to +15.
4673 int task_prio(const struct task_struct *p)
4675 return p->prio - MAX_RT_PRIO;
4679 * task_nice - return the nice value of a given task.
4680 * @p: the task in question.
4682 int task_nice(const struct task_struct *p)
4684 return TASK_NICE(p);
4686 EXPORT_SYMBOL(task_nice);
4689 * idle_cpu - is a given cpu idle currently?
4690 * @cpu: the processor in question.
4692 int idle_cpu(int cpu)
4694 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4698 * idle_task - return the idle task for a given cpu.
4699 * @cpu: the processor in question.
4701 struct task_struct *idle_task(int cpu)
4703 return cpu_rq(cpu)->idle;
4707 * find_process_by_pid - find a process with a matching PID value.
4708 * @pid: the pid in question.
4710 static struct task_struct *find_process_by_pid(pid_t pid)
4712 return pid ? find_task_by_vpid(pid) : current;
4715 /* Actually do priority change: must hold rq lock. */
4717 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4719 BUG_ON(p->se.on_rq);
4722 p->rt_priority = prio;
4723 p->normal_prio = normal_prio(p);
4724 /* we are holding p->pi_lock already */
4725 p->prio = rt_mutex_getprio(p);
4726 if (rt_prio(p->prio))
4727 p->sched_class = &rt_sched_class;
4729 p->sched_class = &fair_sched_class;
4734 * check the target process has a UID that matches the current process's
4736 static bool check_same_owner(struct task_struct *p)
4738 const struct cred *cred = current_cred(), *pcred;
4742 pcred = __task_cred(p);
4743 match = (cred->euid == pcred->euid ||
4744 cred->euid == pcred->uid);
4749 static int __sched_setscheduler(struct task_struct *p, int policy,
4750 const struct sched_param *param, bool user)
4752 int retval, oldprio, oldpolicy = -1, on_rq, running;
4753 unsigned long flags;
4754 const struct sched_class *prev_class;
4758 /* may grab non-irq protected spin_locks */
4759 BUG_ON(in_interrupt());
4761 /* double check policy once rq lock held */
4763 reset_on_fork = p->sched_reset_on_fork;
4764 policy = oldpolicy = p->policy;
4766 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4767 policy &= ~SCHED_RESET_ON_FORK;
4769 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4770 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4771 policy != SCHED_IDLE)
4776 * Valid priorities for SCHED_FIFO and SCHED_RR are
4777 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4778 * SCHED_BATCH and SCHED_IDLE is 0.
4780 if (param->sched_priority < 0 ||
4781 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4782 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4784 if (rt_policy(policy) != (param->sched_priority != 0))
4788 * Allow unprivileged RT tasks to decrease priority:
4790 if (user && !capable(CAP_SYS_NICE)) {
4791 if (rt_policy(policy)) {
4792 unsigned long rlim_rtprio =
4793 task_rlimit(p, RLIMIT_RTPRIO);
4795 /* can't set/change the rt policy */
4796 if (policy != p->policy && !rlim_rtprio)
4799 /* can't increase priority */
4800 if (param->sched_priority > p->rt_priority &&
4801 param->sched_priority > rlim_rtprio)
4805 * Like positive nice levels, dont allow tasks to
4806 * move out of SCHED_IDLE either:
4808 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4811 /* can't change other user's priorities */
4812 if (!check_same_owner(p))
4815 /* Normal users shall not reset the sched_reset_on_fork flag */
4816 if (p->sched_reset_on_fork && !reset_on_fork)
4821 retval = security_task_setscheduler(p);
4827 * make sure no PI-waiters arrive (or leave) while we are
4828 * changing the priority of the task:
4830 raw_spin_lock_irqsave(&p->pi_lock, flags);
4832 * To be able to change p->policy safely, the apropriate
4833 * runqueue lock must be held.
4835 rq = __task_rq_lock(p);
4838 * Changing the policy of the stop threads its a very bad idea
4840 if (p == rq->stop) {
4841 __task_rq_unlock(rq);
4842 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4846 #ifdef CONFIG_RT_GROUP_SCHED
4849 * Do not allow realtime tasks into groups that have no runtime
4852 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4853 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4854 !task_group_is_autogroup(task_group(p))) {
4855 __task_rq_unlock(rq);
4856 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4862 /* recheck policy now with rq lock held */
4863 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4864 policy = oldpolicy = -1;
4865 __task_rq_unlock(rq);
4866 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4869 on_rq = p->se.on_rq;
4870 running = task_current(rq, p);
4872 deactivate_task(rq, p, 0);
4874 p->sched_class->put_prev_task(rq, p);
4876 p->sched_reset_on_fork = reset_on_fork;
4879 prev_class = p->sched_class;
4880 __setscheduler(rq, p, policy, param->sched_priority);
4883 p->sched_class->set_curr_task(rq);
4885 activate_task(rq, p, 0);
4887 check_class_changed(rq, p, prev_class, oldprio, running);
4889 __task_rq_unlock(rq);
4890 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4892 rt_mutex_adjust_pi(p);
4898 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4899 * @p: the task in question.
4900 * @policy: new policy.
4901 * @param: structure containing the new RT priority.
4903 * NOTE that the task may be already dead.
4905 int sched_setscheduler(struct task_struct *p, int policy,
4906 const struct sched_param *param)
4908 return __sched_setscheduler(p, policy, param, true);
4910 EXPORT_SYMBOL_GPL(sched_setscheduler);
4913 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4914 * @p: the task in question.
4915 * @policy: new policy.
4916 * @param: structure containing the new RT priority.
4918 * Just like sched_setscheduler, only don't bother checking if the
4919 * current context has permission. For example, this is needed in
4920 * stop_machine(): we create temporary high priority worker threads,
4921 * but our caller might not have that capability.
4923 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4924 const struct sched_param *param)
4926 return __sched_setscheduler(p, policy, param, false);
4930 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4932 struct sched_param lparam;
4933 struct task_struct *p;
4936 if (!param || pid < 0)
4938 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4943 p = find_process_by_pid(pid);
4945 retval = sched_setscheduler(p, policy, &lparam);
4952 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4953 * @pid: the pid in question.
4954 * @policy: new policy.
4955 * @param: structure containing the new RT priority.
4957 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4958 struct sched_param __user *, param)
4960 /* negative values for policy are not valid */
4964 return do_sched_setscheduler(pid, policy, param);
4968 * sys_sched_setparam - set/change the RT priority of a thread
4969 * @pid: the pid in question.
4970 * @param: structure containing the new RT priority.
4972 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4974 return do_sched_setscheduler(pid, -1, param);
4978 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4979 * @pid: the pid in question.
4981 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4983 struct task_struct *p;
4991 p = find_process_by_pid(pid);
4993 retval = security_task_getscheduler(p);
4996 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5003 * sys_sched_getparam - get the RT priority of a thread
5004 * @pid: the pid in question.
5005 * @param: structure containing the RT priority.
5007 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5009 struct sched_param lp;
5010 struct task_struct *p;
5013 if (!param || pid < 0)
5017 p = find_process_by_pid(pid);
5022 retval = security_task_getscheduler(p);
5026 lp.sched_priority = p->rt_priority;
5030 * This one might sleep, we cannot do it with a spinlock held ...
5032 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5041 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5043 cpumask_var_t cpus_allowed, new_mask;
5044 struct task_struct *p;
5050 p = find_process_by_pid(pid);
5057 /* Prevent p going away */
5061 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5065 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5067 goto out_free_cpus_allowed;
5070 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5073 retval = security_task_setscheduler(p);
5077 cpuset_cpus_allowed(p, cpus_allowed);
5078 cpumask_and(new_mask, in_mask, cpus_allowed);
5080 retval = set_cpus_allowed_ptr(p, new_mask);
5083 cpuset_cpus_allowed(p, cpus_allowed);
5084 if (!cpumask_subset(new_mask, cpus_allowed)) {
5086 * We must have raced with a concurrent cpuset
5087 * update. Just reset the cpus_allowed to the
5088 * cpuset's cpus_allowed
5090 cpumask_copy(new_mask, cpus_allowed);
5095 free_cpumask_var(new_mask);
5096 out_free_cpus_allowed:
5097 free_cpumask_var(cpus_allowed);
5104 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5105 struct cpumask *new_mask)
5107 if (len < cpumask_size())
5108 cpumask_clear(new_mask);
5109 else if (len > cpumask_size())
5110 len = cpumask_size();
5112 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5116 * sys_sched_setaffinity - set the cpu affinity of a process
5117 * @pid: pid of the process
5118 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5119 * @user_mask_ptr: user-space pointer to the new cpu mask
5121 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5122 unsigned long __user *, user_mask_ptr)
5124 cpumask_var_t new_mask;
5127 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5130 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5132 retval = sched_setaffinity(pid, new_mask);
5133 free_cpumask_var(new_mask);
5137 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5139 struct task_struct *p;
5140 unsigned long flags;
5148 p = find_process_by_pid(pid);
5152 retval = security_task_getscheduler(p);
5156 rq = task_rq_lock(p, &flags);
5157 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5158 task_rq_unlock(rq, &flags);
5168 * sys_sched_getaffinity - get the cpu affinity of a process
5169 * @pid: pid of the process
5170 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5171 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5173 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5174 unsigned long __user *, user_mask_ptr)
5179 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5181 if (len & (sizeof(unsigned long)-1))
5184 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5187 ret = sched_getaffinity(pid, mask);
5189 size_t retlen = min_t(size_t, len, cpumask_size());
5191 if (copy_to_user(user_mask_ptr, mask, retlen))
5196 free_cpumask_var(mask);
5202 * sys_sched_yield - yield the current processor to other threads.
5204 * This function yields the current CPU to other tasks. If there are no
5205 * other threads running on this CPU then this function will return.
5207 SYSCALL_DEFINE0(sched_yield)
5209 struct rq *rq = this_rq_lock();
5211 schedstat_inc(rq, yld_count);
5212 current->sched_class->yield_task(rq);
5215 * Since we are going to call schedule() anyway, there's
5216 * no need to preempt or enable interrupts:
5218 __release(rq->lock);
5219 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5220 do_raw_spin_unlock(&rq->lock);
5221 preempt_enable_no_resched();
5228 static inline int should_resched(void)
5230 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5233 static void __cond_resched(void)
5235 add_preempt_count(PREEMPT_ACTIVE);
5237 sub_preempt_count(PREEMPT_ACTIVE);
5240 int __sched _cond_resched(void)
5242 if (should_resched()) {
5248 EXPORT_SYMBOL(_cond_resched);
5251 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5252 * call schedule, and on return reacquire the lock.
5254 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5255 * operations here to prevent schedule() from being called twice (once via
5256 * spin_unlock(), once by hand).
5258 int __cond_resched_lock(spinlock_t *lock)
5260 int resched = should_resched();
5263 lockdep_assert_held(lock);
5265 if (spin_needbreak(lock) || resched) {
5276 EXPORT_SYMBOL(__cond_resched_lock);
5278 int __sched __cond_resched_softirq(void)
5280 BUG_ON(!in_softirq());
5282 if (should_resched()) {
5290 EXPORT_SYMBOL(__cond_resched_softirq);
5293 * yield - yield the current processor to other threads.
5295 * This is a shortcut for kernel-space yielding - it marks the
5296 * thread runnable and calls sys_sched_yield().
5298 void __sched yield(void)
5300 set_current_state(TASK_RUNNING);
5303 EXPORT_SYMBOL(yield);
5306 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5307 * that process accounting knows that this is a task in IO wait state.
5309 void __sched io_schedule(void)
5311 struct rq *rq = raw_rq();
5313 delayacct_blkio_start();
5314 atomic_inc(&rq->nr_iowait);
5315 current->in_iowait = 1;
5317 current->in_iowait = 0;
5318 atomic_dec(&rq->nr_iowait);
5319 delayacct_blkio_end();
5321 EXPORT_SYMBOL(io_schedule);
5323 long __sched io_schedule_timeout(long timeout)
5325 struct rq *rq = raw_rq();
5328 delayacct_blkio_start();
5329 atomic_inc(&rq->nr_iowait);
5330 current->in_iowait = 1;
5331 ret = schedule_timeout(timeout);
5332 current->in_iowait = 0;
5333 atomic_dec(&rq->nr_iowait);
5334 delayacct_blkio_end();
5339 * sys_sched_get_priority_max - return maximum RT priority.
5340 * @policy: scheduling class.
5342 * this syscall returns the maximum rt_priority that can be used
5343 * by a given scheduling class.
5345 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5352 ret = MAX_USER_RT_PRIO-1;
5364 * sys_sched_get_priority_min - return minimum RT priority.
5365 * @policy: scheduling class.
5367 * this syscall returns the minimum rt_priority that can be used
5368 * by a given scheduling class.
5370 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5388 * sys_sched_rr_get_interval - return the default timeslice of a process.
5389 * @pid: pid of the process.
5390 * @interval: userspace pointer to the timeslice value.
5392 * this syscall writes the default timeslice value of a given process
5393 * into the user-space timespec buffer. A value of '0' means infinity.
5395 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5396 struct timespec __user *, interval)
5398 struct task_struct *p;
5399 unsigned int time_slice;
5400 unsigned long flags;
5410 p = find_process_by_pid(pid);
5414 retval = security_task_getscheduler(p);
5418 rq = task_rq_lock(p, &flags);
5419 time_slice = p->sched_class->get_rr_interval(rq, p);
5420 task_rq_unlock(rq, &flags);
5423 jiffies_to_timespec(time_slice, &t);
5424 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5432 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5434 void sched_show_task(struct task_struct *p)
5436 unsigned long free = 0;
5439 state = p->state ? __ffs(p->state) + 1 : 0;
5440 printk(KERN_INFO "%-15.15s %c", p->comm,
5441 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5442 #if BITS_PER_LONG == 32
5443 if (state == TASK_RUNNING)
5444 printk(KERN_CONT " running ");
5446 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5448 if (state == TASK_RUNNING)
5449 printk(KERN_CONT " running task ");
5451 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5453 #ifdef CONFIG_DEBUG_STACK_USAGE
5454 free = stack_not_used(p);
5456 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5457 task_pid_nr(p), task_pid_nr(p->real_parent),
5458 (unsigned long)task_thread_info(p)->flags);
5460 show_stack(p, NULL);
5463 void show_state_filter(unsigned long state_filter)
5465 struct task_struct *g, *p;
5467 #if BITS_PER_LONG == 32
5469 " task PC stack pid father\n");
5472 " task PC stack pid father\n");
5474 read_lock(&tasklist_lock);
5475 do_each_thread(g, p) {
5477 * reset the NMI-timeout, listing all files on a slow
5478 * console might take alot of time:
5480 touch_nmi_watchdog();
5481 if (!state_filter || (p->state & state_filter))
5483 } while_each_thread(g, p);
5485 touch_all_softlockup_watchdogs();
5487 #ifdef CONFIG_SCHED_DEBUG
5488 sysrq_sched_debug_show();
5490 read_unlock(&tasklist_lock);
5492 * Only show locks if all tasks are dumped:
5495 debug_show_all_locks();
5498 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5500 idle->sched_class = &idle_sched_class;
5504 * init_idle - set up an idle thread for a given CPU
5505 * @idle: task in question
5506 * @cpu: cpu the idle task belongs to
5508 * NOTE: this function does not set the idle thread's NEED_RESCHED
5509 * flag, to make booting more robust.
5511 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5513 struct rq *rq = cpu_rq(cpu);
5514 unsigned long flags;
5516 raw_spin_lock_irqsave(&rq->lock, flags);
5519 idle->state = TASK_RUNNING;
5520 idle->se.exec_start = sched_clock();
5522 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5524 * We're having a chicken and egg problem, even though we are
5525 * holding rq->lock, the cpu isn't yet set to this cpu so the
5526 * lockdep check in task_group() will fail.
5528 * Similar case to sched_fork(). / Alternatively we could
5529 * use task_rq_lock() here and obtain the other rq->lock.
5534 __set_task_cpu(idle, cpu);
5537 rq->curr = rq->idle = idle;
5538 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5541 raw_spin_unlock_irqrestore(&rq->lock, flags);
5543 /* Set the preempt count _outside_ the spinlocks! */
5544 #if defined(CONFIG_PREEMPT)
5545 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5547 task_thread_info(idle)->preempt_count = 0;
5550 * The idle tasks have their own, simple scheduling class:
5552 idle->sched_class = &idle_sched_class;
5553 ftrace_graph_init_task(idle);
5557 * In a system that switches off the HZ timer nohz_cpu_mask
5558 * indicates which cpus entered this state. This is used
5559 * in the rcu update to wait only for active cpus. For system
5560 * which do not switch off the HZ timer nohz_cpu_mask should
5561 * always be CPU_BITS_NONE.
5563 cpumask_var_t nohz_cpu_mask;
5566 * Increase the granularity value when there are more CPUs,
5567 * because with more CPUs the 'effective latency' as visible
5568 * to users decreases. But the relationship is not linear,
5569 * so pick a second-best guess by going with the log2 of the
5572 * This idea comes from the SD scheduler of Con Kolivas:
5574 static int get_update_sysctl_factor(void)
5576 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5577 unsigned int factor;
5579 switch (sysctl_sched_tunable_scaling) {
5580 case SCHED_TUNABLESCALING_NONE:
5583 case SCHED_TUNABLESCALING_LINEAR:
5586 case SCHED_TUNABLESCALING_LOG:
5588 factor = 1 + ilog2(cpus);
5595 static void update_sysctl(void)
5597 unsigned int factor = get_update_sysctl_factor();
5599 #define SET_SYSCTL(name) \
5600 (sysctl_##name = (factor) * normalized_sysctl_##name)
5601 SET_SYSCTL(sched_min_granularity);
5602 SET_SYSCTL(sched_latency);
5603 SET_SYSCTL(sched_wakeup_granularity);
5607 static inline void sched_init_granularity(void)
5614 * This is how migration works:
5616 * 1) we invoke migration_cpu_stop() on the target CPU using
5618 * 2) stopper starts to run (implicitly forcing the migrated thread
5620 * 3) it checks whether the migrated task is still in the wrong runqueue.
5621 * 4) if it's in the wrong runqueue then the migration thread removes
5622 * it and puts it into the right queue.
5623 * 5) stopper completes and stop_one_cpu() returns and the migration
5628 * Change a given task's CPU affinity. Migrate the thread to a
5629 * proper CPU and schedule it away if the CPU it's executing on
5630 * is removed from the allowed bitmask.
5632 * NOTE: the caller must have a valid reference to the task, the
5633 * task must not exit() & deallocate itself prematurely. The
5634 * call is not atomic; no spinlocks may be held.
5636 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5638 unsigned long flags;
5640 unsigned int dest_cpu;
5644 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5645 * drop the rq->lock and still rely on ->cpus_allowed.
5648 while (task_is_waking(p))
5650 rq = task_rq_lock(p, &flags);
5651 if (task_is_waking(p)) {
5652 task_rq_unlock(rq, &flags);
5656 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5661 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5662 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5667 if (p->sched_class->set_cpus_allowed)
5668 p->sched_class->set_cpus_allowed(p, new_mask);
5670 cpumask_copy(&p->cpus_allowed, new_mask);
5671 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5674 /* Can the task run on the task's current CPU? If so, we're done */
5675 if (cpumask_test_cpu(task_cpu(p), new_mask))
5678 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5679 if (migrate_task(p, rq)) {
5680 struct migration_arg arg = { p, dest_cpu };
5681 /* Need help from migration thread: drop lock and wait. */
5682 task_rq_unlock(rq, &flags);
5683 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5684 tlb_migrate_finish(p->mm);
5688 task_rq_unlock(rq, &flags);
5692 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5695 * Move (not current) task off this cpu, onto dest cpu. We're doing
5696 * this because either it can't run here any more (set_cpus_allowed()
5697 * away from this CPU, or CPU going down), or because we're
5698 * attempting to rebalance this task on exec (sched_exec).
5700 * So we race with normal scheduler movements, but that's OK, as long
5701 * as the task is no longer on this CPU.
5703 * Returns non-zero if task was successfully migrated.
5705 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5707 struct rq *rq_dest, *rq_src;
5710 if (unlikely(!cpu_active(dest_cpu)))
5713 rq_src = cpu_rq(src_cpu);
5714 rq_dest = cpu_rq(dest_cpu);
5716 double_rq_lock(rq_src, rq_dest);
5717 /* Already moved. */
5718 if (task_cpu(p) != src_cpu)
5720 /* Affinity changed (again). */
5721 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5725 * If we're not on a rq, the next wake-up will ensure we're
5729 deactivate_task(rq_src, p, 0);
5730 set_task_cpu(p, dest_cpu);
5731 activate_task(rq_dest, p, 0);
5732 check_preempt_curr(rq_dest, p, 0);
5737 double_rq_unlock(rq_src, rq_dest);
5742 * migration_cpu_stop - this will be executed by a highprio stopper thread
5743 * and performs thread migration by bumping thread off CPU then
5744 * 'pushing' onto another runqueue.
5746 static int migration_cpu_stop(void *data)
5748 struct migration_arg *arg = data;
5751 * The original target cpu might have gone down and we might
5752 * be on another cpu but it doesn't matter.
5754 local_irq_disable();
5755 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5760 #ifdef CONFIG_HOTPLUG_CPU
5763 * Ensures that the idle task is using init_mm right before its cpu goes
5766 void idle_task_exit(void)
5768 struct mm_struct *mm = current->active_mm;
5770 BUG_ON(cpu_online(smp_processor_id()));
5773 switch_mm(mm, &init_mm, current);
5778 * While a dead CPU has no uninterruptible tasks queued at this point,
5779 * it might still have a nonzero ->nr_uninterruptible counter, because
5780 * for performance reasons the counter is not stricly tracking tasks to
5781 * their home CPUs. So we just add the counter to another CPU's counter,
5782 * to keep the global sum constant after CPU-down:
5784 static void migrate_nr_uninterruptible(struct rq *rq_src)
5786 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5788 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5789 rq_src->nr_uninterruptible = 0;
5793 * remove the tasks which were accounted by rq from calc_load_tasks.
5795 static void calc_global_load_remove(struct rq *rq)
5797 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5798 rq->calc_load_active = 0;
5802 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5803 * try_to_wake_up()->select_task_rq().
5805 * Called with rq->lock held even though we'er in stop_machine() and
5806 * there's no concurrency possible, we hold the required locks anyway
5807 * because of lock validation efforts.
5809 static void migrate_tasks(unsigned int dead_cpu)
5811 struct rq *rq = cpu_rq(dead_cpu);
5812 struct task_struct *next, *stop = rq->stop;
5816 * Fudge the rq selection such that the below task selection loop
5817 * doesn't get stuck on the currently eligible stop task.
5819 * We're currently inside stop_machine() and the rq is either stuck
5820 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5821 * either way we should never end up calling schedule() until we're
5828 * There's this thread running, bail when that's the only
5831 if (rq->nr_running == 1)
5834 next = pick_next_task(rq);
5836 next->sched_class->put_prev_task(rq, next);
5838 /* Find suitable destination for @next, with force if needed. */
5839 dest_cpu = select_fallback_rq(dead_cpu, next);
5840 raw_spin_unlock(&rq->lock);
5842 __migrate_task(next, dead_cpu, dest_cpu);
5844 raw_spin_lock(&rq->lock);
5850 #endif /* CONFIG_HOTPLUG_CPU */
5852 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5854 static struct ctl_table sd_ctl_dir[] = {
5856 .procname = "sched_domain",
5862 static struct ctl_table sd_ctl_root[] = {
5864 .procname = "kernel",
5866 .child = sd_ctl_dir,
5871 static struct ctl_table *sd_alloc_ctl_entry(int n)
5873 struct ctl_table *entry =
5874 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5879 static void sd_free_ctl_entry(struct ctl_table **tablep)
5881 struct ctl_table *entry;
5884 * In the intermediate directories, both the child directory and
5885 * procname are dynamically allocated and could fail but the mode
5886 * will always be set. In the lowest directory the names are
5887 * static strings and all have proc handlers.
5889 for (entry = *tablep; entry->mode; entry++) {
5891 sd_free_ctl_entry(&entry->child);
5892 if (entry->proc_handler == NULL)
5893 kfree(entry->procname);
5901 set_table_entry(struct ctl_table *entry,
5902 const char *procname, void *data, int maxlen,
5903 mode_t mode, proc_handler *proc_handler)
5905 entry->procname = procname;
5907 entry->maxlen = maxlen;
5909 entry->proc_handler = proc_handler;
5912 static struct ctl_table *
5913 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5915 struct ctl_table *table = sd_alloc_ctl_entry(13);
5920 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5921 sizeof(long), 0644, proc_doulongvec_minmax);
5922 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5923 sizeof(long), 0644, proc_doulongvec_minmax);
5924 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5925 sizeof(int), 0644, proc_dointvec_minmax);
5926 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5927 sizeof(int), 0644, proc_dointvec_minmax);
5928 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5929 sizeof(int), 0644, proc_dointvec_minmax);
5930 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5931 sizeof(int), 0644, proc_dointvec_minmax);
5932 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5933 sizeof(int), 0644, proc_dointvec_minmax);
5934 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5935 sizeof(int), 0644, proc_dointvec_minmax);
5936 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5937 sizeof(int), 0644, proc_dointvec_minmax);
5938 set_table_entry(&table[9], "cache_nice_tries",
5939 &sd->cache_nice_tries,
5940 sizeof(int), 0644, proc_dointvec_minmax);
5941 set_table_entry(&table[10], "flags", &sd->flags,
5942 sizeof(int), 0644, proc_dointvec_minmax);
5943 set_table_entry(&table[11], "name", sd->name,
5944 CORENAME_MAX_SIZE, 0444, proc_dostring);
5945 /* &table[12] is terminator */
5950 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5952 struct ctl_table *entry, *table;
5953 struct sched_domain *sd;
5954 int domain_num = 0, i;
5957 for_each_domain(cpu, sd)
5959 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5964 for_each_domain(cpu, sd) {
5965 snprintf(buf, 32, "domain%d", i);
5966 entry->procname = kstrdup(buf, GFP_KERNEL);
5968 entry->child = sd_alloc_ctl_domain_table(sd);
5975 static struct ctl_table_header *sd_sysctl_header;
5976 static void register_sched_domain_sysctl(void)
5978 int i, cpu_num = num_possible_cpus();
5979 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5982 WARN_ON(sd_ctl_dir[0].child);
5983 sd_ctl_dir[0].child = entry;
5988 for_each_possible_cpu(i) {
5989 snprintf(buf, 32, "cpu%d", i);
5990 entry->procname = kstrdup(buf, GFP_KERNEL);
5992 entry->child = sd_alloc_ctl_cpu_table(i);
5996 WARN_ON(sd_sysctl_header);
5997 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6000 /* may be called multiple times per register */
6001 static void unregister_sched_domain_sysctl(void)
6003 if (sd_sysctl_header)
6004 unregister_sysctl_table(sd_sysctl_header);
6005 sd_sysctl_header = NULL;
6006 if (sd_ctl_dir[0].child)
6007 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6010 static void register_sched_domain_sysctl(void)
6013 static void unregister_sched_domain_sysctl(void)
6018 static void set_rq_online(struct rq *rq)
6021 const struct sched_class *class;
6023 cpumask_set_cpu(rq->cpu, rq->rd->online);
6026 for_each_class(class) {
6027 if (class->rq_online)
6028 class->rq_online(rq);
6033 static void set_rq_offline(struct rq *rq)
6036 const struct sched_class *class;
6038 for_each_class(class) {
6039 if (class->rq_offline)
6040 class->rq_offline(rq);
6043 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6049 * migration_call - callback that gets triggered when a CPU is added.
6050 * Here we can start up the necessary migration thread for the new CPU.
6052 static int __cpuinit
6053 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6055 int cpu = (long)hcpu;
6056 unsigned long flags;
6057 struct rq *rq = cpu_rq(cpu);
6059 switch (action & ~CPU_TASKS_FROZEN) {
6061 case CPU_UP_PREPARE:
6062 rq->calc_load_update = calc_load_update;
6066 /* Update our root-domain */
6067 raw_spin_lock_irqsave(&rq->lock, flags);
6069 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6073 raw_spin_unlock_irqrestore(&rq->lock, flags);
6076 #ifdef CONFIG_HOTPLUG_CPU
6078 /* Update our root-domain */
6079 raw_spin_lock_irqsave(&rq->lock, flags);
6081 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6085 BUG_ON(rq->nr_running != 1); /* the migration thread */
6086 raw_spin_unlock_irqrestore(&rq->lock, flags);
6088 migrate_nr_uninterruptible(rq);
6089 calc_global_load_remove(rq);
6097 * Register at high priority so that task migration (migrate_all_tasks)
6098 * happens before everything else. This has to be lower priority than
6099 * the notifier in the perf_event subsystem, though.
6101 static struct notifier_block __cpuinitdata migration_notifier = {
6102 .notifier_call = migration_call,
6103 .priority = CPU_PRI_MIGRATION,
6106 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6107 unsigned long action, void *hcpu)
6109 switch (action & ~CPU_TASKS_FROZEN) {
6111 case CPU_DOWN_FAILED:
6112 set_cpu_active((long)hcpu, true);
6119 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6120 unsigned long action, void *hcpu)
6122 switch (action & ~CPU_TASKS_FROZEN) {
6123 case CPU_DOWN_PREPARE:
6124 set_cpu_active((long)hcpu, false);
6131 static int __init migration_init(void)
6133 void *cpu = (void *)(long)smp_processor_id();
6136 /* Initialize migration for the boot CPU */
6137 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6138 BUG_ON(err == NOTIFY_BAD);
6139 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6140 register_cpu_notifier(&migration_notifier);
6142 /* Register cpu active notifiers */
6143 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6144 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6148 early_initcall(migration_init);
6153 #ifdef CONFIG_SCHED_DEBUG
6155 static __read_mostly int sched_domain_debug_enabled;
6157 static int __init sched_domain_debug_setup(char *str)
6159 sched_domain_debug_enabled = 1;
6163 early_param("sched_debug", sched_domain_debug_setup);
6165 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6166 struct cpumask *groupmask)
6168 struct sched_group *group = sd->groups;
6171 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6172 cpumask_clear(groupmask);
6174 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6176 if (!(sd->flags & SD_LOAD_BALANCE)) {
6177 printk("does not load-balance\n");
6179 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6184 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6186 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6187 printk(KERN_ERR "ERROR: domain->span does not contain "
6190 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6191 printk(KERN_ERR "ERROR: domain->groups does not contain"
6195 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6199 printk(KERN_ERR "ERROR: group is NULL\n");
6203 if (!group->cpu_power) {
6204 printk(KERN_CONT "\n");
6205 printk(KERN_ERR "ERROR: domain->cpu_power not "
6210 if (!cpumask_weight(sched_group_cpus(group))) {
6211 printk(KERN_CONT "\n");
6212 printk(KERN_ERR "ERROR: empty group\n");
6216 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6217 printk(KERN_CONT "\n");
6218 printk(KERN_ERR "ERROR: repeated CPUs\n");
6222 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6224 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6226 printk(KERN_CONT " %s", str);
6227 if (group->cpu_power != SCHED_LOAD_SCALE) {
6228 printk(KERN_CONT " (cpu_power = %d)",
6232 group = group->next;
6233 } while (group != sd->groups);
6234 printk(KERN_CONT "\n");
6236 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6237 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6240 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6241 printk(KERN_ERR "ERROR: parent span is not a superset "
6242 "of domain->span\n");
6246 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6248 cpumask_var_t groupmask;
6251 if (!sched_domain_debug_enabled)
6255 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6259 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6261 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6262 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6267 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6274 free_cpumask_var(groupmask);
6276 #else /* !CONFIG_SCHED_DEBUG */
6277 # define sched_domain_debug(sd, cpu) do { } while (0)
6278 #endif /* CONFIG_SCHED_DEBUG */
6280 static int sd_degenerate(struct sched_domain *sd)
6282 if (cpumask_weight(sched_domain_span(sd)) == 1)
6285 /* Following flags need at least 2 groups */
6286 if (sd->flags & (SD_LOAD_BALANCE |
6287 SD_BALANCE_NEWIDLE |
6291 SD_SHARE_PKG_RESOURCES)) {
6292 if (sd->groups != sd->groups->next)
6296 /* Following flags don't use groups */
6297 if (sd->flags & (SD_WAKE_AFFINE))
6304 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6306 unsigned long cflags = sd->flags, pflags = parent->flags;
6308 if (sd_degenerate(parent))
6311 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6314 /* Flags needing groups don't count if only 1 group in parent */
6315 if (parent->groups == parent->groups->next) {
6316 pflags &= ~(SD_LOAD_BALANCE |
6317 SD_BALANCE_NEWIDLE |
6321 SD_SHARE_PKG_RESOURCES);
6322 if (nr_node_ids == 1)
6323 pflags &= ~SD_SERIALIZE;
6325 if (~cflags & pflags)
6331 static void free_rootdomain(struct root_domain *rd)
6333 synchronize_sched();
6335 cpupri_cleanup(&rd->cpupri);
6337 free_cpumask_var(rd->rto_mask);
6338 free_cpumask_var(rd->online);
6339 free_cpumask_var(rd->span);
6343 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6345 struct root_domain *old_rd = NULL;
6346 unsigned long flags;
6348 raw_spin_lock_irqsave(&rq->lock, flags);
6353 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6356 cpumask_clear_cpu(rq->cpu, old_rd->span);
6359 * If we dont want to free the old_rt yet then
6360 * set old_rd to NULL to skip the freeing later
6363 if (!atomic_dec_and_test(&old_rd->refcount))
6367 atomic_inc(&rd->refcount);
6370 cpumask_set_cpu(rq->cpu, rd->span);
6371 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6374 raw_spin_unlock_irqrestore(&rq->lock, flags);
6377 free_rootdomain(old_rd);
6380 static int init_rootdomain(struct root_domain *rd)
6382 memset(rd, 0, sizeof(*rd));
6384 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6386 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6388 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6391 if (cpupri_init(&rd->cpupri) != 0)
6396 free_cpumask_var(rd->rto_mask);
6398 free_cpumask_var(rd->online);
6400 free_cpumask_var(rd->span);
6405 static void init_defrootdomain(void)
6407 init_rootdomain(&def_root_domain);
6409 atomic_set(&def_root_domain.refcount, 1);
6412 static struct root_domain *alloc_rootdomain(void)
6414 struct root_domain *rd;
6416 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6420 if (init_rootdomain(rd) != 0) {
6429 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6430 * hold the hotplug lock.
6433 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6435 struct rq *rq = cpu_rq(cpu);
6436 struct sched_domain *tmp;
6438 for (tmp = sd; tmp; tmp = tmp->parent)
6439 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6441 /* Remove the sched domains which do not contribute to scheduling. */
6442 for (tmp = sd; tmp; ) {
6443 struct sched_domain *parent = tmp->parent;
6447 if (sd_parent_degenerate(tmp, parent)) {
6448 tmp->parent = parent->parent;
6450 parent->parent->child = tmp;
6455 if (sd && sd_degenerate(sd)) {
6461 sched_domain_debug(sd, cpu);
6463 rq_attach_root(rq, rd);
6464 rcu_assign_pointer(rq->sd, sd);
6467 /* cpus with isolated domains */
6468 static cpumask_var_t cpu_isolated_map;
6470 /* Setup the mask of cpus configured for isolated domains */
6471 static int __init isolated_cpu_setup(char *str)
6473 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6474 cpulist_parse(str, cpu_isolated_map);
6478 __setup("isolcpus=", isolated_cpu_setup);
6481 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6482 * to a function which identifies what group(along with sched group) a CPU
6483 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6484 * (due to the fact that we keep track of groups covered with a struct cpumask).
6486 * init_sched_build_groups will build a circular linked list of the groups
6487 * covered by the given span, and will set each group's ->cpumask correctly,
6488 * and ->cpu_power to 0.
6491 init_sched_build_groups(const struct cpumask *span,
6492 const struct cpumask *cpu_map,
6493 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6494 struct sched_group **sg,
6495 struct cpumask *tmpmask),
6496 struct cpumask *covered, struct cpumask *tmpmask)
6498 struct sched_group *first = NULL, *last = NULL;
6501 cpumask_clear(covered);
6503 for_each_cpu(i, span) {
6504 struct sched_group *sg;
6505 int group = group_fn(i, cpu_map, &sg, tmpmask);
6508 if (cpumask_test_cpu(i, covered))
6511 cpumask_clear(sched_group_cpus(sg));
6514 for_each_cpu(j, span) {
6515 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6518 cpumask_set_cpu(j, covered);
6519 cpumask_set_cpu(j, sched_group_cpus(sg));
6530 #define SD_NODES_PER_DOMAIN 16
6535 * find_next_best_node - find the next node to include in a sched_domain
6536 * @node: node whose sched_domain we're building
6537 * @used_nodes: nodes already in the sched_domain
6539 * Find the next node to include in a given scheduling domain. Simply
6540 * finds the closest node not already in the @used_nodes map.
6542 * Should use nodemask_t.
6544 static int find_next_best_node(int node, nodemask_t *used_nodes)
6546 int i, n, val, min_val, best_node = 0;
6550 for (i = 0; i < nr_node_ids; i++) {
6551 /* Start at @node */
6552 n = (node + i) % nr_node_ids;
6554 if (!nr_cpus_node(n))
6557 /* Skip already used nodes */
6558 if (node_isset(n, *used_nodes))
6561 /* Simple min distance search */
6562 val = node_distance(node, n);
6564 if (val < min_val) {
6570 node_set(best_node, *used_nodes);
6575 * sched_domain_node_span - get a cpumask for a node's sched_domain
6576 * @node: node whose cpumask we're constructing
6577 * @span: resulting cpumask
6579 * Given a node, construct a good cpumask for its sched_domain to span. It
6580 * should be one that prevents unnecessary balancing, but also spreads tasks
6583 static void sched_domain_node_span(int node, struct cpumask *span)
6585 nodemask_t used_nodes;
6588 cpumask_clear(span);
6589 nodes_clear(used_nodes);
6591 cpumask_or(span, span, cpumask_of_node(node));
6592 node_set(node, used_nodes);
6594 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6595 int next_node = find_next_best_node(node, &used_nodes);
6597 cpumask_or(span, span, cpumask_of_node(next_node));
6600 #endif /* CONFIG_NUMA */
6602 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6605 * The cpus mask in sched_group and sched_domain hangs off the end.
6607 * ( See the the comments in include/linux/sched.h:struct sched_group
6608 * and struct sched_domain. )
6610 struct static_sched_group {
6611 struct sched_group sg;
6612 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6615 struct static_sched_domain {
6616 struct sched_domain sd;
6617 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6623 cpumask_var_t domainspan;
6624 cpumask_var_t covered;
6625 cpumask_var_t notcovered;
6627 cpumask_var_t nodemask;
6628 cpumask_var_t this_sibling_map;
6629 cpumask_var_t this_core_map;
6630 cpumask_var_t this_book_map;
6631 cpumask_var_t send_covered;
6632 cpumask_var_t tmpmask;
6633 struct sched_group **sched_group_nodes;
6634 struct root_domain *rd;
6638 sa_sched_groups = 0,
6644 sa_this_sibling_map,
6646 sa_sched_group_nodes,
6656 * SMT sched-domains:
6658 #ifdef CONFIG_SCHED_SMT
6659 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6660 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6663 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6664 struct sched_group **sg, struct cpumask *unused)
6667 *sg = &per_cpu(sched_groups, cpu).sg;
6670 #endif /* CONFIG_SCHED_SMT */
6673 * multi-core sched-domains:
6675 #ifdef CONFIG_SCHED_MC
6676 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6677 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6680 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6681 struct sched_group **sg, struct cpumask *mask)
6684 #ifdef CONFIG_SCHED_SMT
6685 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6686 group = cpumask_first(mask);
6691 *sg = &per_cpu(sched_group_core, group).sg;
6694 #endif /* CONFIG_SCHED_MC */
6697 * book sched-domains:
6699 #ifdef CONFIG_SCHED_BOOK
6700 static DEFINE_PER_CPU(struct static_sched_domain, book_domains);
6701 static DEFINE_PER_CPU(struct static_sched_group, sched_group_book);
6704 cpu_to_book_group(int cpu, const struct cpumask *cpu_map,
6705 struct sched_group **sg, struct cpumask *mask)
6708 #ifdef CONFIG_SCHED_MC
6709 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6710 group = cpumask_first(mask);
6711 #elif defined(CONFIG_SCHED_SMT)
6712 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6713 group = cpumask_first(mask);
6716 *sg = &per_cpu(sched_group_book, group).sg;
6719 #endif /* CONFIG_SCHED_BOOK */
6721 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6722 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6725 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6726 struct sched_group **sg, struct cpumask *mask)
6729 #ifdef CONFIG_SCHED_BOOK
6730 cpumask_and(mask, cpu_book_mask(cpu), cpu_map);
6731 group = cpumask_first(mask);
6732 #elif defined(CONFIG_SCHED_MC)
6733 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6734 group = cpumask_first(mask);
6735 #elif defined(CONFIG_SCHED_SMT)
6736 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6737 group = cpumask_first(mask);
6742 *sg = &per_cpu(sched_group_phys, group).sg;
6748 * The init_sched_build_groups can't handle what we want to do with node
6749 * groups, so roll our own. Now each node has its own list of groups which
6750 * gets dynamically allocated.
6752 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6753 static struct sched_group ***sched_group_nodes_bycpu;
6755 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6756 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6758 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6759 struct sched_group **sg,
6760 struct cpumask *nodemask)
6764 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6765 group = cpumask_first(nodemask);
6768 *sg = &per_cpu(sched_group_allnodes, group).sg;
6772 static void init_numa_sched_groups_power(struct sched_group *group_head)
6774 struct sched_group *sg = group_head;
6780 for_each_cpu(j, sched_group_cpus(sg)) {
6781 struct sched_domain *sd;
6783 sd = &per_cpu(phys_domains, j).sd;
6784 if (j != group_first_cpu(sd->groups)) {
6786 * Only add "power" once for each
6792 sg->cpu_power += sd->groups->cpu_power;
6795 } while (sg != group_head);
6798 static int build_numa_sched_groups(struct s_data *d,
6799 const struct cpumask *cpu_map, int num)
6801 struct sched_domain *sd;
6802 struct sched_group *sg, *prev;
6805 cpumask_clear(d->covered);
6806 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6807 if (cpumask_empty(d->nodemask)) {
6808 d->sched_group_nodes[num] = NULL;
6812 sched_domain_node_span(num, d->domainspan);
6813 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6815 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6818 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6822 d->sched_group_nodes[num] = sg;
6824 for_each_cpu(j, d->nodemask) {
6825 sd = &per_cpu(node_domains, j).sd;
6830 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6832 cpumask_or(d->covered, d->covered, d->nodemask);
6835 for (j = 0; j < nr_node_ids; j++) {
6836 n = (num + j) % nr_node_ids;
6837 cpumask_complement(d->notcovered, d->covered);
6838 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6839 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6840 if (cpumask_empty(d->tmpmask))
6842 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6843 if (cpumask_empty(d->tmpmask))
6845 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6849 "Can not alloc domain group for node %d\n", j);
6853 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6854 sg->next = prev->next;
6855 cpumask_or(d->covered, d->covered, d->tmpmask);
6862 #endif /* CONFIG_NUMA */
6865 /* Free memory allocated for various sched_group structures */
6866 static void free_sched_groups(const struct cpumask *cpu_map,
6867 struct cpumask *nodemask)
6871 for_each_cpu(cpu, cpu_map) {
6872 struct sched_group **sched_group_nodes
6873 = sched_group_nodes_bycpu[cpu];
6875 if (!sched_group_nodes)
6878 for (i = 0; i < nr_node_ids; i++) {
6879 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6881 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6882 if (cpumask_empty(nodemask))
6892 if (oldsg != sched_group_nodes[i])
6895 kfree(sched_group_nodes);
6896 sched_group_nodes_bycpu[cpu] = NULL;
6899 #else /* !CONFIG_NUMA */
6900 static void free_sched_groups(const struct cpumask *cpu_map,
6901 struct cpumask *nodemask)
6904 #endif /* CONFIG_NUMA */
6907 * Initialize sched groups cpu_power.
6909 * cpu_power indicates the capacity of sched group, which is used while
6910 * distributing the load between different sched groups in a sched domain.
6911 * Typically cpu_power for all the groups in a sched domain will be same unless
6912 * there are asymmetries in the topology. If there are asymmetries, group
6913 * having more cpu_power will pickup more load compared to the group having
6916 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6918 struct sched_domain *child;
6919 struct sched_group *group;
6923 WARN_ON(!sd || !sd->groups);
6925 if (cpu != group_first_cpu(sd->groups))
6928 sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups));
6932 sd->groups->cpu_power = 0;
6935 power = SCHED_LOAD_SCALE;
6936 weight = cpumask_weight(sched_domain_span(sd));
6938 * SMT siblings share the power of a single core.
6939 * Usually multiple threads get a better yield out of
6940 * that one core than a single thread would have,
6941 * reflect that in sd->smt_gain.
6943 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6944 power *= sd->smt_gain;
6946 power >>= SCHED_LOAD_SHIFT;
6948 sd->groups->cpu_power += power;
6953 * Add cpu_power of each child group to this groups cpu_power.
6955 group = child->groups;
6957 sd->groups->cpu_power += group->cpu_power;
6958 group = group->next;
6959 } while (group != child->groups);
6963 * Initializers for schedule domains
6964 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6967 #ifdef CONFIG_SCHED_DEBUG
6968 # define SD_INIT_NAME(sd, type) sd->name = #type
6970 # define SD_INIT_NAME(sd, type) do { } while (0)
6973 #define SD_INIT(sd, type) sd_init_##type(sd)
6975 #define SD_INIT_FUNC(type) \
6976 static noinline void sd_init_##type(struct sched_domain *sd) \
6978 memset(sd, 0, sizeof(*sd)); \
6979 *sd = SD_##type##_INIT; \
6980 sd->level = SD_LV_##type; \
6981 SD_INIT_NAME(sd, type); \
6986 SD_INIT_FUNC(ALLNODES)
6989 #ifdef CONFIG_SCHED_SMT
6990 SD_INIT_FUNC(SIBLING)
6992 #ifdef CONFIG_SCHED_MC
6995 #ifdef CONFIG_SCHED_BOOK
6999 static int default_relax_domain_level = -1;
7001 static int __init setup_relax_domain_level(char *str)
7005 val = simple_strtoul(str, NULL, 0);
7006 if (val < SD_LV_MAX)
7007 default_relax_domain_level = val;
7011 __setup("relax_domain_level=", setup_relax_domain_level);
7013 static void set_domain_attribute(struct sched_domain *sd,
7014 struct sched_domain_attr *attr)
7018 if (!attr || attr->relax_domain_level < 0) {
7019 if (default_relax_domain_level < 0)
7022 request = default_relax_domain_level;
7024 request = attr->relax_domain_level;
7025 if (request < sd->level) {
7026 /* turn off idle balance on this domain */
7027 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7029 /* turn on idle balance on this domain */
7030 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7034 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7035 const struct cpumask *cpu_map)
7038 case sa_sched_groups:
7039 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
7040 d->sched_group_nodes = NULL;
7042 free_rootdomain(d->rd); /* fall through */
7044 free_cpumask_var(d->tmpmask); /* fall through */
7045 case sa_send_covered:
7046 free_cpumask_var(d->send_covered); /* fall through */
7047 case sa_this_book_map:
7048 free_cpumask_var(d->this_book_map); /* fall through */
7049 case sa_this_core_map:
7050 free_cpumask_var(d->this_core_map); /* fall through */
7051 case sa_this_sibling_map:
7052 free_cpumask_var(d->this_sibling_map); /* fall through */
7054 free_cpumask_var(d->nodemask); /* fall through */
7055 case sa_sched_group_nodes:
7057 kfree(d->sched_group_nodes); /* fall through */
7059 free_cpumask_var(d->notcovered); /* fall through */
7061 free_cpumask_var(d->covered); /* fall through */
7063 free_cpumask_var(d->domainspan); /* fall through */
7070 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7071 const struct cpumask *cpu_map)
7074 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
7076 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
7077 return sa_domainspan;
7078 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
7080 /* Allocate the per-node list of sched groups */
7081 d->sched_group_nodes = kcalloc(nr_node_ids,
7082 sizeof(struct sched_group *), GFP_KERNEL);
7083 if (!d->sched_group_nodes) {
7084 printk(KERN_WARNING "Can not alloc sched group node list\n");
7085 return sa_notcovered;
7087 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
7089 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
7090 return sa_sched_group_nodes;
7091 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
7093 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
7094 return sa_this_sibling_map;
7095 if (!alloc_cpumask_var(&d->this_book_map, GFP_KERNEL))
7096 return sa_this_core_map;
7097 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
7098 return sa_this_book_map;
7099 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
7100 return sa_send_covered;
7101 d->rd = alloc_rootdomain();
7103 printk(KERN_WARNING "Cannot alloc root domain\n");
7106 return sa_rootdomain;
7109 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
7110 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
7112 struct sched_domain *sd = NULL;
7114 struct sched_domain *parent;
7117 if (cpumask_weight(cpu_map) >
7118 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
7119 sd = &per_cpu(allnodes_domains, i).sd;
7120 SD_INIT(sd, ALLNODES);
7121 set_domain_attribute(sd, attr);
7122 cpumask_copy(sched_domain_span(sd), cpu_map);
7123 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
7128 sd = &per_cpu(node_domains, i).sd;
7130 set_domain_attribute(sd, attr);
7131 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7132 sd->parent = parent;
7135 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7140 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7141 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7142 struct sched_domain *parent, int i)
7144 struct sched_domain *sd;
7145 sd = &per_cpu(phys_domains, i).sd;
7147 set_domain_attribute(sd, attr);
7148 cpumask_copy(sched_domain_span(sd), d->nodemask);
7149 sd->parent = parent;
7152 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7156 static struct sched_domain *__build_book_sched_domain(struct s_data *d,
7157 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7158 struct sched_domain *parent, int i)
7160 struct sched_domain *sd = parent;
7161 #ifdef CONFIG_SCHED_BOOK
7162 sd = &per_cpu(book_domains, i).sd;
7164 set_domain_attribute(sd, attr);
7165 cpumask_and(sched_domain_span(sd), cpu_map, cpu_book_mask(i));
7166 sd->parent = parent;
7168 cpu_to_book_group(i, cpu_map, &sd->groups, d->tmpmask);
7173 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7174 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7175 struct sched_domain *parent, int i)
7177 struct sched_domain *sd = parent;
7178 #ifdef CONFIG_SCHED_MC
7179 sd = &per_cpu(core_domains, i).sd;
7181 set_domain_attribute(sd, attr);
7182 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7183 sd->parent = parent;
7185 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7190 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7191 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7192 struct sched_domain *parent, int i)
7194 struct sched_domain *sd = parent;
7195 #ifdef CONFIG_SCHED_SMT
7196 sd = &per_cpu(cpu_domains, i).sd;
7197 SD_INIT(sd, SIBLING);
7198 set_domain_attribute(sd, attr);
7199 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7200 sd->parent = parent;
7202 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7207 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7208 const struct cpumask *cpu_map, int cpu)
7211 #ifdef CONFIG_SCHED_SMT
7212 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7213 cpumask_and(d->this_sibling_map, cpu_map,
7214 topology_thread_cpumask(cpu));
7215 if (cpu == cpumask_first(d->this_sibling_map))
7216 init_sched_build_groups(d->this_sibling_map, cpu_map,
7218 d->send_covered, d->tmpmask);
7221 #ifdef CONFIG_SCHED_MC
7222 case SD_LV_MC: /* set up multi-core groups */
7223 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7224 if (cpu == cpumask_first(d->this_core_map))
7225 init_sched_build_groups(d->this_core_map, cpu_map,
7227 d->send_covered, d->tmpmask);
7230 #ifdef CONFIG_SCHED_BOOK
7231 case SD_LV_BOOK: /* set up book groups */
7232 cpumask_and(d->this_book_map, cpu_map, cpu_book_mask(cpu));
7233 if (cpu == cpumask_first(d->this_book_map))
7234 init_sched_build_groups(d->this_book_map, cpu_map,
7236 d->send_covered, d->tmpmask);
7239 case SD_LV_CPU: /* set up physical groups */
7240 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7241 if (!cpumask_empty(d->nodemask))
7242 init_sched_build_groups(d->nodemask, cpu_map,
7244 d->send_covered, d->tmpmask);
7247 case SD_LV_ALLNODES:
7248 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7249 d->send_covered, d->tmpmask);
7258 * Build sched domains for a given set of cpus and attach the sched domains
7259 * to the individual cpus
7261 static int __build_sched_domains(const struct cpumask *cpu_map,
7262 struct sched_domain_attr *attr)
7264 enum s_alloc alloc_state = sa_none;
7266 struct sched_domain *sd;
7272 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7273 if (alloc_state != sa_rootdomain)
7275 alloc_state = sa_sched_groups;
7278 * Set up domains for cpus specified by the cpu_map.
7280 for_each_cpu(i, cpu_map) {
7281 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7284 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7285 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7286 sd = __build_book_sched_domain(&d, cpu_map, attr, sd, i);
7287 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7288 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7291 for_each_cpu(i, cpu_map) {
7292 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7293 build_sched_groups(&d, SD_LV_BOOK, cpu_map, i);
7294 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7297 /* Set up physical groups */
7298 for (i = 0; i < nr_node_ids; i++)
7299 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7302 /* Set up node groups */
7304 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7306 for (i = 0; i < nr_node_ids; i++)
7307 if (build_numa_sched_groups(&d, cpu_map, i))
7311 /* Calculate CPU power for physical packages and nodes */
7312 #ifdef CONFIG_SCHED_SMT
7313 for_each_cpu(i, cpu_map) {
7314 sd = &per_cpu(cpu_domains, i).sd;
7315 init_sched_groups_power(i, sd);
7318 #ifdef CONFIG_SCHED_MC
7319 for_each_cpu(i, cpu_map) {
7320 sd = &per_cpu(core_domains, i).sd;
7321 init_sched_groups_power(i, sd);
7324 #ifdef CONFIG_SCHED_BOOK
7325 for_each_cpu(i, cpu_map) {
7326 sd = &per_cpu(book_domains, i).sd;
7327 init_sched_groups_power(i, sd);
7331 for_each_cpu(i, cpu_map) {
7332 sd = &per_cpu(phys_domains, i).sd;
7333 init_sched_groups_power(i, sd);
7337 for (i = 0; i < nr_node_ids; i++)
7338 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7340 if (d.sd_allnodes) {
7341 struct sched_group *sg;
7343 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7345 init_numa_sched_groups_power(sg);
7349 /* Attach the domains */
7350 for_each_cpu(i, cpu_map) {
7351 #ifdef CONFIG_SCHED_SMT
7352 sd = &per_cpu(cpu_domains, i).sd;
7353 #elif defined(CONFIG_SCHED_MC)
7354 sd = &per_cpu(core_domains, i).sd;
7355 #elif defined(CONFIG_SCHED_BOOK)
7356 sd = &per_cpu(book_domains, i).sd;
7358 sd = &per_cpu(phys_domains, i).sd;
7360 cpu_attach_domain(sd, d.rd, i);
7363 d.sched_group_nodes = NULL; /* don't free this we still need it */
7364 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7368 __free_domain_allocs(&d, alloc_state, cpu_map);
7372 static int build_sched_domains(const struct cpumask *cpu_map)
7374 return __build_sched_domains(cpu_map, NULL);
7377 static cpumask_var_t *doms_cur; /* current sched domains */
7378 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7379 static struct sched_domain_attr *dattr_cur;
7380 /* attribues of custom domains in 'doms_cur' */
7383 * Special case: If a kmalloc of a doms_cur partition (array of
7384 * cpumask) fails, then fallback to a single sched domain,
7385 * as determined by the single cpumask fallback_doms.
7387 static cpumask_var_t fallback_doms;
7390 * arch_update_cpu_topology lets virtualized architectures update the
7391 * cpu core maps. It is supposed to return 1 if the topology changed
7392 * or 0 if it stayed the same.
7394 int __attribute__((weak)) arch_update_cpu_topology(void)
7399 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7402 cpumask_var_t *doms;
7404 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7407 for (i = 0; i < ndoms; i++) {
7408 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7409 free_sched_domains(doms, i);
7416 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7419 for (i = 0; i < ndoms; i++)
7420 free_cpumask_var(doms[i]);
7425 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7426 * For now this just excludes isolated cpus, but could be used to
7427 * exclude other special cases in the future.
7429 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7433 arch_update_cpu_topology();
7435 doms_cur = alloc_sched_domains(ndoms_cur);
7437 doms_cur = &fallback_doms;
7438 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7440 err = build_sched_domains(doms_cur[0]);
7441 register_sched_domain_sysctl();
7446 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7447 struct cpumask *tmpmask)
7449 free_sched_groups(cpu_map, tmpmask);
7453 * Detach sched domains from a group of cpus specified in cpu_map
7454 * These cpus will now be attached to the NULL domain
7456 static void detach_destroy_domains(const struct cpumask *cpu_map)
7458 /* Save because hotplug lock held. */
7459 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7462 for_each_cpu(i, cpu_map)
7463 cpu_attach_domain(NULL, &def_root_domain, i);
7464 synchronize_sched();
7465 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7468 /* handle null as "default" */
7469 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7470 struct sched_domain_attr *new, int idx_new)
7472 struct sched_domain_attr tmp;
7479 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7480 new ? (new + idx_new) : &tmp,
7481 sizeof(struct sched_domain_attr));
7485 * Partition sched domains as specified by the 'ndoms_new'
7486 * cpumasks in the array doms_new[] of cpumasks. This compares
7487 * doms_new[] to the current sched domain partitioning, doms_cur[].
7488 * It destroys each deleted domain and builds each new domain.
7490 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7491 * The masks don't intersect (don't overlap.) We should setup one
7492 * sched domain for each mask. CPUs not in any of the cpumasks will
7493 * not be load balanced. If the same cpumask appears both in the
7494 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7497 * The passed in 'doms_new' should be allocated using
7498 * alloc_sched_domains. This routine takes ownership of it and will
7499 * free_sched_domains it when done with it. If the caller failed the
7500 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7501 * and partition_sched_domains() will fallback to the single partition
7502 * 'fallback_doms', it also forces the domains to be rebuilt.
7504 * If doms_new == NULL it will be replaced with cpu_online_mask.
7505 * ndoms_new == 0 is a special case for destroying existing domains,
7506 * and it will not create the default domain.
7508 * Call with hotplug lock held
7510 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7511 struct sched_domain_attr *dattr_new)
7516 mutex_lock(&sched_domains_mutex);
7518 /* always unregister in case we don't destroy any domains */
7519 unregister_sched_domain_sysctl();
7521 /* Let architecture update cpu core mappings. */
7522 new_topology = arch_update_cpu_topology();
7524 n = doms_new ? ndoms_new : 0;
7526 /* Destroy deleted domains */
7527 for (i = 0; i < ndoms_cur; i++) {
7528 for (j = 0; j < n && !new_topology; j++) {
7529 if (cpumask_equal(doms_cur[i], doms_new[j])
7530 && dattrs_equal(dattr_cur, i, dattr_new, j))
7533 /* no match - a current sched domain not in new doms_new[] */
7534 detach_destroy_domains(doms_cur[i]);
7539 if (doms_new == NULL) {
7541 doms_new = &fallback_doms;
7542 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7543 WARN_ON_ONCE(dattr_new);
7546 /* Build new domains */
7547 for (i = 0; i < ndoms_new; i++) {
7548 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7549 if (cpumask_equal(doms_new[i], doms_cur[j])
7550 && dattrs_equal(dattr_new, i, dattr_cur, j))
7553 /* no match - add a new doms_new */
7554 __build_sched_domains(doms_new[i],
7555 dattr_new ? dattr_new + i : NULL);
7560 /* Remember the new sched domains */
7561 if (doms_cur != &fallback_doms)
7562 free_sched_domains(doms_cur, ndoms_cur);
7563 kfree(dattr_cur); /* kfree(NULL) is safe */
7564 doms_cur = doms_new;
7565 dattr_cur = dattr_new;
7566 ndoms_cur = ndoms_new;
7568 register_sched_domain_sysctl();
7570 mutex_unlock(&sched_domains_mutex);
7573 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7574 static void arch_reinit_sched_domains(void)
7578 /* Destroy domains first to force the rebuild */
7579 partition_sched_domains(0, NULL, NULL);
7581 rebuild_sched_domains();
7585 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7587 unsigned int level = 0;
7589 if (sscanf(buf, "%u", &level) != 1)
7593 * level is always be positive so don't check for
7594 * level < POWERSAVINGS_BALANCE_NONE which is 0
7595 * What happens on 0 or 1 byte write,
7596 * need to check for count as well?
7599 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7603 sched_smt_power_savings = level;
7605 sched_mc_power_savings = level;
7607 arch_reinit_sched_domains();
7612 #ifdef CONFIG_SCHED_MC
7613 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7614 struct sysdev_class_attribute *attr,
7617 return sprintf(page, "%u\n", sched_mc_power_savings);
7619 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7620 struct sysdev_class_attribute *attr,
7621 const char *buf, size_t count)
7623 return sched_power_savings_store(buf, count, 0);
7625 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7626 sched_mc_power_savings_show,
7627 sched_mc_power_savings_store);
7630 #ifdef CONFIG_SCHED_SMT
7631 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7632 struct sysdev_class_attribute *attr,
7635 return sprintf(page, "%u\n", sched_smt_power_savings);
7637 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7638 struct sysdev_class_attribute *attr,
7639 const char *buf, size_t count)
7641 return sched_power_savings_store(buf, count, 1);
7643 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7644 sched_smt_power_savings_show,
7645 sched_smt_power_savings_store);
7648 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7652 #ifdef CONFIG_SCHED_SMT
7654 err = sysfs_create_file(&cls->kset.kobj,
7655 &attr_sched_smt_power_savings.attr);
7657 #ifdef CONFIG_SCHED_MC
7658 if (!err && mc_capable())
7659 err = sysfs_create_file(&cls->kset.kobj,
7660 &attr_sched_mc_power_savings.attr);
7664 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7667 * Update cpusets according to cpu_active mask. If cpusets are
7668 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7669 * around partition_sched_domains().
7671 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7674 switch (action & ~CPU_TASKS_FROZEN) {
7676 case CPU_DOWN_FAILED:
7677 cpuset_update_active_cpus();
7684 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7687 switch (action & ~CPU_TASKS_FROZEN) {
7688 case CPU_DOWN_PREPARE:
7689 cpuset_update_active_cpus();
7696 static int update_runtime(struct notifier_block *nfb,
7697 unsigned long action, void *hcpu)
7699 int cpu = (int)(long)hcpu;
7702 case CPU_DOWN_PREPARE:
7703 case CPU_DOWN_PREPARE_FROZEN:
7704 disable_runtime(cpu_rq(cpu));
7707 case CPU_DOWN_FAILED:
7708 case CPU_DOWN_FAILED_FROZEN:
7710 case CPU_ONLINE_FROZEN:
7711 enable_runtime(cpu_rq(cpu));
7719 void __init sched_init_smp(void)
7721 cpumask_var_t non_isolated_cpus;
7723 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7724 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7726 #if defined(CONFIG_NUMA)
7727 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7729 BUG_ON(sched_group_nodes_bycpu == NULL);
7732 mutex_lock(&sched_domains_mutex);
7733 arch_init_sched_domains(cpu_active_mask);
7734 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7735 if (cpumask_empty(non_isolated_cpus))
7736 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7737 mutex_unlock(&sched_domains_mutex);
7740 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7741 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7743 /* RT runtime code needs to handle some hotplug events */
7744 hotcpu_notifier(update_runtime, 0);
7748 /* Move init over to a non-isolated CPU */
7749 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7751 sched_init_granularity();
7752 free_cpumask_var(non_isolated_cpus);
7754 init_sched_rt_class();
7757 void __init sched_init_smp(void)
7759 sched_init_granularity();
7761 #endif /* CONFIG_SMP */
7763 const_debug unsigned int sysctl_timer_migration = 1;
7765 int in_sched_functions(unsigned long addr)
7767 return in_lock_functions(addr) ||
7768 (addr >= (unsigned long)__sched_text_start
7769 && addr < (unsigned long)__sched_text_end);
7772 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7774 cfs_rq->tasks_timeline = RB_ROOT;
7775 INIT_LIST_HEAD(&cfs_rq->tasks);
7776 #ifdef CONFIG_FAIR_GROUP_SCHED
7779 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7782 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7784 struct rt_prio_array *array;
7787 array = &rt_rq->active;
7788 for (i = 0; i < MAX_RT_PRIO; i++) {
7789 INIT_LIST_HEAD(array->queue + i);
7790 __clear_bit(i, array->bitmap);
7792 /* delimiter for bitsearch: */
7793 __set_bit(MAX_RT_PRIO, array->bitmap);
7795 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7796 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7798 rt_rq->highest_prio.next = MAX_RT_PRIO;
7802 rt_rq->rt_nr_migratory = 0;
7803 rt_rq->overloaded = 0;
7804 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7808 rt_rq->rt_throttled = 0;
7809 rt_rq->rt_runtime = 0;
7810 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7812 #ifdef CONFIG_RT_GROUP_SCHED
7813 rt_rq->rt_nr_boosted = 0;
7818 #ifdef CONFIG_FAIR_GROUP_SCHED
7819 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7820 struct sched_entity *se, int cpu,
7821 struct sched_entity *parent)
7823 struct rq *rq = cpu_rq(cpu);
7824 tg->cfs_rq[cpu] = cfs_rq;
7825 init_cfs_rq(cfs_rq, rq);
7829 /* se could be NULL for root_task_group */
7834 se->cfs_rq = &rq->cfs;
7836 se->cfs_rq = parent->my_q;
7839 update_load_set(&se->load, 0);
7840 se->parent = parent;
7844 #ifdef CONFIG_RT_GROUP_SCHED
7845 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7846 struct sched_rt_entity *rt_se, int cpu,
7847 struct sched_rt_entity *parent)
7849 struct rq *rq = cpu_rq(cpu);
7851 tg->rt_rq[cpu] = rt_rq;
7852 init_rt_rq(rt_rq, rq);
7854 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7856 tg->rt_se[cpu] = rt_se;
7861 rt_se->rt_rq = &rq->rt;
7863 rt_se->rt_rq = parent->my_q;
7865 rt_se->my_q = rt_rq;
7866 rt_se->parent = parent;
7867 INIT_LIST_HEAD(&rt_se->run_list);
7871 void __init sched_init(void)
7874 unsigned long alloc_size = 0, ptr;
7876 #ifdef CONFIG_FAIR_GROUP_SCHED
7877 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7879 #ifdef CONFIG_RT_GROUP_SCHED
7880 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7882 #ifdef CONFIG_CPUMASK_OFFSTACK
7883 alloc_size += num_possible_cpus() * cpumask_size();
7886 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7888 #ifdef CONFIG_FAIR_GROUP_SCHED
7889 root_task_group.se = (struct sched_entity **)ptr;
7890 ptr += nr_cpu_ids * sizeof(void **);
7892 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7893 ptr += nr_cpu_ids * sizeof(void **);
7895 #endif /* CONFIG_FAIR_GROUP_SCHED */
7896 #ifdef CONFIG_RT_GROUP_SCHED
7897 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7898 ptr += nr_cpu_ids * sizeof(void **);
7900 root_task_group.rt_rq = (struct rt_rq **)ptr;
7901 ptr += nr_cpu_ids * sizeof(void **);
7903 #endif /* CONFIG_RT_GROUP_SCHED */
7904 #ifdef CONFIG_CPUMASK_OFFSTACK
7905 for_each_possible_cpu(i) {
7906 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7907 ptr += cpumask_size();
7909 #endif /* CONFIG_CPUMASK_OFFSTACK */
7913 init_defrootdomain();
7916 init_rt_bandwidth(&def_rt_bandwidth,
7917 global_rt_period(), global_rt_runtime());
7919 #ifdef CONFIG_RT_GROUP_SCHED
7920 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7921 global_rt_period(), global_rt_runtime());
7922 #endif /* CONFIG_RT_GROUP_SCHED */
7924 #ifdef CONFIG_CGROUP_SCHED
7925 list_add(&root_task_group.list, &task_groups);
7926 INIT_LIST_HEAD(&root_task_group.children);
7927 autogroup_init(&init_task);
7928 #endif /* CONFIG_CGROUP_SCHED */
7930 for_each_possible_cpu(i) {
7934 raw_spin_lock_init(&rq->lock);
7936 rq->calc_load_active = 0;
7937 rq->calc_load_update = jiffies + LOAD_FREQ;
7938 init_cfs_rq(&rq->cfs, rq);
7939 init_rt_rq(&rq->rt, rq);
7940 #ifdef CONFIG_FAIR_GROUP_SCHED
7941 root_task_group.shares = root_task_group_load;
7942 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7944 * How much cpu bandwidth does root_task_group get?
7946 * In case of task-groups formed thr' the cgroup filesystem, it
7947 * gets 100% of the cpu resources in the system. This overall
7948 * system cpu resource is divided among the tasks of
7949 * root_task_group and its child task-groups in a fair manner,
7950 * based on each entity's (task or task-group's) weight
7951 * (se->load.weight).
7953 * In other words, if root_task_group has 10 tasks of weight
7954 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7955 * then A0's share of the cpu resource is:
7957 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7959 * We achieve this by letting root_task_group's tasks sit
7960 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7962 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7963 #endif /* CONFIG_FAIR_GROUP_SCHED */
7965 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7966 #ifdef CONFIG_RT_GROUP_SCHED
7967 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7968 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7971 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7972 rq->cpu_load[j] = 0;
7974 rq->last_load_update_tick = jiffies;
7979 rq->cpu_power = SCHED_LOAD_SCALE;
7980 rq->post_schedule = 0;
7981 rq->active_balance = 0;
7982 rq->next_balance = jiffies;
7987 rq->avg_idle = 2*sysctl_sched_migration_cost;
7988 rq_attach_root(rq, &def_root_domain);
7990 rq->nohz_balance_kick = 0;
7991 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
7995 atomic_set(&rq->nr_iowait, 0);
7998 set_load_weight(&init_task);
8000 #ifdef CONFIG_PREEMPT_NOTIFIERS
8001 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8005 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8008 #ifdef CONFIG_RT_MUTEXES
8009 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
8013 * The boot idle thread does lazy MMU switching as well:
8015 atomic_inc(&init_mm.mm_count);
8016 enter_lazy_tlb(&init_mm, current);
8019 * Make us the idle thread. Technically, schedule() should not be
8020 * called from this thread, however somewhere below it might be,
8021 * but because we are the idle thread, we just pick up running again
8022 * when this runqueue becomes "idle".
8024 init_idle(current, smp_processor_id());
8026 calc_load_update = jiffies + LOAD_FREQ;
8029 * During early bootup we pretend to be a normal task:
8031 current->sched_class = &fair_sched_class;
8033 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8034 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8037 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8038 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8039 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8040 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8041 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8043 /* May be allocated at isolcpus cmdline parse time */
8044 if (cpu_isolated_map == NULL)
8045 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8048 scheduler_running = 1;
8051 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8052 static inline int preempt_count_equals(int preempt_offset)
8054 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8056 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
8059 void __might_sleep(const char *file, int line, int preempt_offset)
8062 static unsigned long prev_jiffy; /* ratelimiting */
8064 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8065 system_state != SYSTEM_RUNNING || oops_in_progress)
8067 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8069 prev_jiffy = jiffies;
8072 "BUG: sleeping function called from invalid context at %s:%d\n",
8075 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8076 in_atomic(), irqs_disabled(),
8077 current->pid, current->comm);
8079 debug_show_held_locks(current);
8080 if (irqs_disabled())
8081 print_irqtrace_events(current);
8085 EXPORT_SYMBOL(__might_sleep);
8088 #ifdef CONFIG_MAGIC_SYSRQ
8089 static void normalize_task(struct rq *rq, struct task_struct *p)
8093 on_rq = p->se.on_rq;
8095 deactivate_task(rq, p, 0);
8096 __setscheduler(rq, p, SCHED_NORMAL, 0);
8098 activate_task(rq, p, 0);
8099 resched_task(rq->curr);
8103 void normalize_rt_tasks(void)
8105 struct task_struct *g, *p;
8106 unsigned long flags;
8109 read_lock_irqsave(&tasklist_lock, flags);
8110 do_each_thread(g, p) {
8112 * Only normalize user tasks:
8117 p->se.exec_start = 0;
8118 #ifdef CONFIG_SCHEDSTATS
8119 p->se.statistics.wait_start = 0;
8120 p->se.statistics.sleep_start = 0;
8121 p->se.statistics.block_start = 0;
8126 * Renice negative nice level userspace
8129 if (TASK_NICE(p) < 0 && p->mm)
8130 set_user_nice(p, 0);
8134 raw_spin_lock(&p->pi_lock);
8135 rq = __task_rq_lock(p);
8137 normalize_task(rq, p);
8139 __task_rq_unlock(rq);
8140 raw_spin_unlock(&p->pi_lock);
8141 } while_each_thread(g, p);
8143 read_unlock_irqrestore(&tasklist_lock, flags);
8146 #endif /* CONFIG_MAGIC_SYSRQ */
8148 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8150 * These functions are only useful for the IA64 MCA handling, or kdb.
8152 * They can only be called when the whole system has been
8153 * stopped - every CPU needs to be quiescent, and no scheduling
8154 * activity can take place. Using them for anything else would
8155 * be a serious bug, and as a result, they aren't even visible
8156 * under any other configuration.
8160 * curr_task - return the current task for a given cpu.
8161 * @cpu: the processor in question.
8163 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8165 struct task_struct *curr_task(int cpu)
8167 return cpu_curr(cpu);
8170 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8174 * set_curr_task - set the current task for a given cpu.
8175 * @cpu: the processor in question.
8176 * @p: the task pointer to set.
8178 * Description: This function must only be used when non-maskable interrupts
8179 * are serviced on a separate stack. It allows the architecture to switch the
8180 * notion of the current task on a cpu in a non-blocking manner. This function
8181 * must be called with all CPU's synchronized, and interrupts disabled, the
8182 * and caller must save the original value of the current task (see
8183 * curr_task() above) and restore that value before reenabling interrupts and
8184 * re-starting the system.
8186 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8188 void set_curr_task(int cpu, struct task_struct *p)
8195 #ifdef CONFIG_FAIR_GROUP_SCHED
8196 static void free_fair_sched_group(struct task_group *tg)
8200 for_each_possible_cpu(i) {
8202 kfree(tg->cfs_rq[i]);
8212 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8214 struct cfs_rq *cfs_rq;
8215 struct sched_entity *se;
8219 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8222 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8226 tg->shares = NICE_0_LOAD;
8228 for_each_possible_cpu(i) {
8231 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8232 GFP_KERNEL, cpu_to_node(i));
8236 se = kzalloc_node(sizeof(struct sched_entity),
8237 GFP_KERNEL, cpu_to_node(i));
8241 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8252 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8254 struct rq *rq = cpu_rq(cpu);
8255 unsigned long flags;
8258 * Only empty task groups can be destroyed; so we can speculatively
8259 * check on_list without danger of it being re-added.
8261 if (!tg->cfs_rq[cpu]->on_list)
8264 raw_spin_lock_irqsave(&rq->lock, flags);
8265 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8266 raw_spin_unlock_irqrestore(&rq->lock, flags);
8268 #else /* !CONFG_FAIR_GROUP_SCHED */
8269 static inline void free_fair_sched_group(struct task_group *tg)
8274 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8279 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8282 #endif /* CONFIG_FAIR_GROUP_SCHED */
8284 #ifdef CONFIG_RT_GROUP_SCHED
8285 static void free_rt_sched_group(struct task_group *tg)
8289 destroy_rt_bandwidth(&tg->rt_bandwidth);
8291 for_each_possible_cpu(i) {
8293 kfree(tg->rt_rq[i]);
8295 kfree(tg->rt_se[i]);
8303 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8305 struct rt_rq *rt_rq;
8306 struct sched_rt_entity *rt_se;
8310 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8313 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8317 init_rt_bandwidth(&tg->rt_bandwidth,
8318 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8320 for_each_possible_cpu(i) {
8323 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8324 GFP_KERNEL, cpu_to_node(i));
8328 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8329 GFP_KERNEL, cpu_to_node(i));
8333 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8343 #else /* !CONFIG_RT_GROUP_SCHED */
8344 static inline void free_rt_sched_group(struct task_group *tg)
8349 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8353 #endif /* CONFIG_RT_GROUP_SCHED */
8355 #ifdef CONFIG_CGROUP_SCHED
8356 static void free_sched_group(struct task_group *tg)
8358 free_fair_sched_group(tg);
8359 free_rt_sched_group(tg);
8364 /* allocate runqueue etc for a new task group */
8365 struct task_group *sched_create_group(struct task_group *parent)
8367 struct task_group *tg;
8368 unsigned long flags;
8370 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8372 return ERR_PTR(-ENOMEM);
8374 if (!alloc_fair_sched_group(tg, parent))
8377 if (!alloc_rt_sched_group(tg, parent))
8380 spin_lock_irqsave(&task_group_lock, flags);
8381 list_add_rcu(&tg->list, &task_groups);
8383 WARN_ON(!parent); /* root should already exist */
8385 tg->parent = parent;
8386 INIT_LIST_HEAD(&tg->children);
8387 list_add_rcu(&tg->siblings, &parent->children);
8388 spin_unlock_irqrestore(&task_group_lock, flags);
8393 free_sched_group(tg);
8394 return ERR_PTR(-ENOMEM);
8397 /* rcu callback to free various structures associated with a task group */
8398 static void free_sched_group_rcu(struct rcu_head *rhp)
8400 /* now it should be safe to free those cfs_rqs */
8401 free_sched_group(container_of(rhp, struct task_group, rcu));
8404 /* Destroy runqueue etc associated with a task group */
8405 void sched_destroy_group(struct task_group *tg)
8407 unsigned long flags;
8410 /* end participation in shares distribution */
8411 for_each_possible_cpu(i)
8412 unregister_fair_sched_group(tg, i);
8414 spin_lock_irqsave(&task_group_lock, flags);
8415 list_del_rcu(&tg->list);
8416 list_del_rcu(&tg->siblings);
8417 spin_unlock_irqrestore(&task_group_lock, flags);
8419 /* wait for possible concurrent references to cfs_rqs complete */
8420 call_rcu(&tg->rcu, free_sched_group_rcu);
8423 /* change task's runqueue when it moves between groups.
8424 * The caller of this function should have put the task in its new group
8425 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8426 * reflect its new group.
8428 void sched_move_task(struct task_struct *tsk)
8431 unsigned long flags;
8434 rq = task_rq_lock(tsk, &flags);
8436 running = task_current(rq, tsk);
8437 on_rq = tsk->se.on_rq;
8440 dequeue_task(rq, tsk, 0);
8441 if (unlikely(running))
8442 tsk->sched_class->put_prev_task(rq, tsk);
8444 #ifdef CONFIG_FAIR_GROUP_SCHED
8445 if (tsk->sched_class->task_move_group)
8446 tsk->sched_class->task_move_group(tsk, on_rq);
8449 set_task_rq(tsk, task_cpu(tsk));
8451 if (unlikely(running))
8452 tsk->sched_class->set_curr_task(rq);
8454 enqueue_task(rq, tsk, 0);
8456 task_rq_unlock(rq, &flags);
8458 #endif /* CONFIG_CGROUP_SCHED */
8460 #ifdef CONFIG_FAIR_GROUP_SCHED
8461 static DEFINE_MUTEX(shares_mutex);
8463 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8466 unsigned long flags;
8469 * We can't change the weight of the root cgroup.
8474 if (shares < MIN_SHARES)
8475 shares = MIN_SHARES;
8476 else if (shares > MAX_SHARES)
8477 shares = MAX_SHARES;
8479 mutex_lock(&shares_mutex);
8480 if (tg->shares == shares)
8483 tg->shares = shares;
8484 for_each_possible_cpu(i) {
8485 struct rq *rq = cpu_rq(i);
8486 struct sched_entity *se;
8489 /* Propagate contribution to hierarchy */
8490 raw_spin_lock_irqsave(&rq->lock, flags);
8491 for_each_sched_entity(se)
8492 update_cfs_shares(group_cfs_rq(se), 0);
8493 raw_spin_unlock_irqrestore(&rq->lock, flags);
8497 mutex_unlock(&shares_mutex);
8501 unsigned long sched_group_shares(struct task_group *tg)
8507 #ifdef CONFIG_RT_GROUP_SCHED
8509 * Ensure that the real time constraints are schedulable.
8511 static DEFINE_MUTEX(rt_constraints_mutex);
8513 static unsigned long to_ratio(u64 period, u64 runtime)
8515 if (runtime == RUNTIME_INF)
8518 return div64_u64(runtime << 20, period);
8521 /* Must be called with tasklist_lock held */
8522 static inline int tg_has_rt_tasks(struct task_group *tg)
8524 struct task_struct *g, *p;
8526 do_each_thread(g, p) {
8527 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8529 } while_each_thread(g, p);
8534 struct rt_schedulable_data {
8535 struct task_group *tg;
8540 static int tg_schedulable(struct task_group *tg, void *data)
8542 struct rt_schedulable_data *d = data;
8543 struct task_group *child;
8544 unsigned long total, sum = 0;
8545 u64 period, runtime;
8547 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8548 runtime = tg->rt_bandwidth.rt_runtime;
8551 period = d->rt_period;
8552 runtime = d->rt_runtime;
8556 * Cannot have more runtime than the period.
8558 if (runtime > period && runtime != RUNTIME_INF)
8562 * Ensure we don't starve existing RT tasks.
8564 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8567 total = to_ratio(period, runtime);
8570 * Nobody can have more than the global setting allows.
8572 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8576 * The sum of our children's runtime should not exceed our own.
8578 list_for_each_entry_rcu(child, &tg->children, siblings) {
8579 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8580 runtime = child->rt_bandwidth.rt_runtime;
8582 if (child == d->tg) {
8583 period = d->rt_period;
8584 runtime = d->rt_runtime;
8587 sum += to_ratio(period, runtime);
8596 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8598 struct rt_schedulable_data data = {
8600 .rt_period = period,
8601 .rt_runtime = runtime,
8604 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8607 static int tg_set_bandwidth(struct task_group *tg,
8608 u64 rt_period, u64 rt_runtime)
8612 mutex_lock(&rt_constraints_mutex);
8613 read_lock(&tasklist_lock);
8614 err = __rt_schedulable(tg, rt_period, rt_runtime);
8618 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8619 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8620 tg->rt_bandwidth.rt_runtime = rt_runtime;
8622 for_each_possible_cpu(i) {
8623 struct rt_rq *rt_rq = tg->rt_rq[i];
8625 raw_spin_lock(&rt_rq->rt_runtime_lock);
8626 rt_rq->rt_runtime = rt_runtime;
8627 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8629 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8631 read_unlock(&tasklist_lock);
8632 mutex_unlock(&rt_constraints_mutex);
8637 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8639 u64 rt_runtime, rt_period;
8641 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8642 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8643 if (rt_runtime_us < 0)
8644 rt_runtime = RUNTIME_INF;
8646 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8649 long sched_group_rt_runtime(struct task_group *tg)
8653 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8656 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8657 do_div(rt_runtime_us, NSEC_PER_USEC);
8658 return rt_runtime_us;
8661 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8663 u64 rt_runtime, rt_period;
8665 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8666 rt_runtime = tg->rt_bandwidth.rt_runtime;
8671 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8674 long sched_group_rt_period(struct task_group *tg)
8678 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8679 do_div(rt_period_us, NSEC_PER_USEC);
8680 return rt_period_us;
8683 static int sched_rt_global_constraints(void)
8685 u64 runtime, period;
8688 if (sysctl_sched_rt_period <= 0)
8691 runtime = global_rt_runtime();
8692 period = global_rt_period();
8695 * Sanity check on the sysctl variables.
8697 if (runtime > period && runtime != RUNTIME_INF)
8700 mutex_lock(&rt_constraints_mutex);
8701 read_lock(&tasklist_lock);
8702 ret = __rt_schedulable(NULL, 0, 0);
8703 read_unlock(&tasklist_lock);
8704 mutex_unlock(&rt_constraints_mutex);
8709 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8711 /* Don't accept realtime tasks when there is no way for them to run */
8712 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8718 #else /* !CONFIG_RT_GROUP_SCHED */
8719 static int sched_rt_global_constraints(void)
8721 unsigned long flags;
8724 if (sysctl_sched_rt_period <= 0)
8728 * There's always some RT tasks in the root group
8729 * -- migration, kstopmachine etc..
8731 if (sysctl_sched_rt_runtime == 0)
8734 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8735 for_each_possible_cpu(i) {
8736 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8738 raw_spin_lock(&rt_rq->rt_runtime_lock);
8739 rt_rq->rt_runtime = global_rt_runtime();
8740 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8742 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8746 #endif /* CONFIG_RT_GROUP_SCHED */
8748 int sched_rt_handler(struct ctl_table *table, int write,
8749 void __user *buffer, size_t *lenp,
8753 int old_period, old_runtime;
8754 static DEFINE_MUTEX(mutex);
8757 old_period = sysctl_sched_rt_period;
8758 old_runtime = sysctl_sched_rt_runtime;
8760 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8762 if (!ret && write) {
8763 ret = sched_rt_global_constraints();
8765 sysctl_sched_rt_period = old_period;
8766 sysctl_sched_rt_runtime = old_runtime;
8768 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8769 def_rt_bandwidth.rt_period =
8770 ns_to_ktime(global_rt_period());
8773 mutex_unlock(&mutex);
8778 #ifdef CONFIG_CGROUP_SCHED
8780 /* return corresponding task_group object of a cgroup */
8781 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8783 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8784 struct task_group, css);
8787 static struct cgroup_subsys_state *
8788 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8790 struct task_group *tg, *parent;
8792 if (!cgrp->parent) {
8793 /* This is early initialization for the top cgroup */
8794 return &root_task_group.css;
8797 parent = cgroup_tg(cgrp->parent);
8798 tg = sched_create_group(parent);
8800 return ERR_PTR(-ENOMEM);
8806 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8808 struct task_group *tg = cgroup_tg(cgrp);
8810 sched_destroy_group(tg);
8814 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8816 #ifdef CONFIG_RT_GROUP_SCHED
8817 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8820 /* We don't support RT-tasks being in separate groups */
8821 if (tsk->sched_class != &fair_sched_class)
8828 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8829 struct task_struct *tsk, bool threadgroup)
8831 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8835 struct task_struct *c;
8837 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8838 retval = cpu_cgroup_can_attach_task(cgrp, c);
8850 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8851 struct cgroup *old_cont, struct task_struct *tsk,
8854 sched_move_task(tsk);
8856 struct task_struct *c;
8858 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8866 cpu_cgroup_exit(struct cgroup_subsys *ss, struct task_struct *task)
8869 * cgroup_exit() is called in the copy_process() failure path.
8870 * Ignore this case since the task hasn't ran yet, this avoids
8871 * trying to poke a half freed task state from generic code.
8873 if (!(task->flags & PF_EXITING))
8876 sched_move_task(task);
8879 #ifdef CONFIG_FAIR_GROUP_SCHED
8880 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8883 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8886 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8888 struct task_group *tg = cgroup_tg(cgrp);
8890 return (u64) tg->shares;
8892 #endif /* CONFIG_FAIR_GROUP_SCHED */
8894 #ifdef CONFIG_RT_GROUP_SCHED
8895 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8898 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8901 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8903 return sched_group_rt_runtime(cgroup_tg(cgrp));
8906 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8909 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8912 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8914 return sched_group_rt_period(cgroup_tg(cgrp));
8916 #endif /* CONFIG_RT_GROUP_SCHED */
8918 static struct cftype cpu_files[] = {
8919 #ifdef CONFIG_FAIR_GROUP_SCHED
8922 .read_u64 = cpu_shares_read_u64,
8923 .write_u64 = cpu_shares_write_u64,
8926 #ifdef CONFIG_RT_GROUP_SCHED
8928 .name = "rt_runtime_us",
8929 .read_s64 = cpu_rt_runtime_read,
8930 .write_s64 = cpu_rt_runtime_write,
8933 .name = "rt_period_us",
8934 .read_u64 = cpu_rt_period_read_uint,
8935 .write_u64 = cpu_rt_period_write_uint,
8940 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8942 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8945 struct cgroup_subsys cpu_cgroup_subsys = {
8947 .create = cpu_cgroup_create,
8948 .destroy = cpu_cgroup_destroy,
8949 .can_attach = cpu_cgroup_can_attach,
8950 .attach = cpu_cgroup_attach,
8951 .exit = cpu_cgroup_exit,
8952 .populate = cpu_cgroup_populate,
8953 .subsys_id = cpu_cgroup_subsys_id,
8957 #endif /* CONFIG_CGROUP_SCHED */
8959 #ifdef CONFIG_CGROUP_CPUACCT
8962 * CPU accounting code for task groups.
8964 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8965 * (balbir@in.ibm.com).
8968 /* track cpu usage of a group of tasks and its child groups */
8970 struct cgroup_subsys_state css;
8971 /* cpuusage holds pointer to a u64-type object on every cpu */
8972 u64 __percpu *cpuusage;
8973 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8974 struct cpuacct *parent;
8977 struct cgroup_subsys cpuacct_subsys;
8979 /* return cpu accounting group corresponding to this container */
8980 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8982 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8983 struct cpuacct, css);
8986 /* return cpu accounting group to which this task belongs */
8987 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8989 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8990 struct cpuacct, css);
8993 /* create a new cpu accounting group */
8994 static struct cgroup_subsys_state *cpuacct_create(
8995 struct cgroup_subsys *ss, struct cgroup *cgrp)
8997 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9003 ca->cpuusage = alloc_percpu(u64);
9007 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9008 if (percpu_counter_init(&ca->cpustat[i], 0))
9009 goto out_free_counters;
9012 ca->parent = cgroup_ca(cgrp->parent);
9018 percpu_counter_destroy(&ca->cpustat[i]);
9019 free_percpu(ca->cpuusage);
9023 return ERR_PTR(-ENOMEM);
9026 /* destroy an existing cpu accounting group */
9028 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9030 struct cpuacct *ca = cgroup_ca(cgrp);
9033 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9034 percpu_counter_destroy(&ca->cpustat[i]);
9035 free_percpu(ca->cpuusage);
9039 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9041 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9044 #ifndef CONFIG_64BIT
9046 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9048 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9050 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9058 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9060 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9062 #ifndef CONFIG_64BIT
9064 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9066 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9068 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9074 /* return total cpu usage (in nanoseconds) of a group */
9075 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9077 struct cpuacct *ca = cgroup_ca(cgrp);
9078 u64 totalcpuusage = 0;
9081 for_each_present_cpu(i)
9082 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9084 return totalcpuusage;
9087 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9090 struct cpuacct *ca = cgroup_ca(cgrp);
9099 for_each_present_cpu(i)
9100 cpuacct_cpuusage_write(ca, i, 0);
9106 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9109 struct cpuacct *ca = cgroup_ca(cgroup);
9113 for_each_present_cpu(i) {
9114 percpu = cpuacct_cpuusage_read(ca, i);
9115 seq_printf(m, "%llu ", (unsigned long long) percpu);
9117 seq_printf(m, "\n");
9121 static const char *cpuacct_stat_desc[] = {
9122 [CPUACCT_STAT_USER] = "user",
9123 [CPUACCT_STAT_SYSTEM] = "system",
9126 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9127 struct cgroup_map_cb *cb)
9129 struct cpuacct *ca = cgroup_ca(cgrp);
9132 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9133 s64 val = percpu_counter_read(&ca->cpustat[i]);
9134 val = cputime64_to_clock_t(val);
9135 cb->fill(cb, cpuacct_stat_desc[i], val);
9140 static struct cftype files[] = {
9143 .read_u64 = cpuusage_read,
9144 .write_u64 = cpuusage_write,
9147 .name = "usage_percpu",
9148 .read_seq_string = cpuacct_percpu_seq_read,
9152 .read_map = cpuacct_stats_show,
9156 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9158 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9162 * charge this task's execution time to its accounting group.
9164 * called with rq->lock held.
9166 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9171 if (unlikely(!cpuacct_subsys.active))
9174 cpu = task_cpu(tsk);
9180 for (; ca; ca = ca->parent) {
9181 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9182 *cpuusage += cputime;
9189 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9190 * in cputime_t units. As a result, cpuacct_update_stats calls
9191 * percpu_counter_add with values large enough to always overflow the
9192 * per cpu batch limit causing bad SMP scalability.
9194 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9195 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9196 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9199 #define CPUACCT_BATCH \
9200 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9202 #define CPUACCT_BATCH 0
9206 * Charge the system/user time to the task's accounting group.
9208 static void cpuacct_update_stats(struct task_struct *tsk,
9209 enum cpuacct_stat_index idx, cputime_t val)
9212 int batch = CPUACCT_BATCH;
9214 if (unlikely(!cpuacct_subsys.active))
9221 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9227 struct cgroup_subsys cpuacct_subsys = {
9229 .create = cpuacct_create,
9230 .destroy = cpuacct_destroy,
9231 .populate = cpuacct_populate,
9232 .subsys_id = cpuacct_subsys_id,
9234 #endif /* CONFIG_CGROUP_CPUACCT */