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>
79 #include "sched_cpupri.h"
81 #define CREATE_TRACE_POINTS
82 #include <trace/events/sched.h>
85 * Convert user-nice values [ -20 ... 0 ... 19 ]
86 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
89 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
90 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
91 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
94 * 'User priority' is the nice value converted to something we
95 * can work with better when scaling various scheduler parameters,
96 * it's a [ 0 ... 39 ] range.
98 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
99 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
100 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
103 * Helpers for converting nanosecond timing to jiffy resolution
105 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
107 #define NICE_0_LOAD SCHED_LOAD_SCALE
108 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
111 * These are the 'tuning knobs' of the scheduler:
113 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
114 * Timeslices get refilled after they expire.
116 #define DEF_TIMESLICE (100 * HZ / 1000)
119 * single value that denotes runtime == period, ie unlimited time.
121 #define RUNTIME_INF ((u64)~0ULL)
123 static inline int rt_policy(int policy)
125 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
130 static inline int task_has_rt_policy(struct task_struct *p)
132 return rt_policy(p->policy);
136 * This is the priority-queue data structure of the RT scheduling class:
138 struct rt_prio_array {
139 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
140 struct list_head queue[MAX_RT_PRIO];
143 struct rt_bandwidth {
144 /* nests inside the rq lock: */
145 raw_spinlock_t rt_runtime_lock;
148 struct hrtimer rt_period_timer;
151 static struct rt_bandwidth def_rt_bandwidth;
153 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
155 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
157 struct rt_bandwidth *rt_b =
158 container_of(timer, struct rt_bandwidth, rt_period_timer);
164 now = hrtimer_cb_get_time(timer);
165 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
170 idle = do_sched_rt_period_timer(rt_b, overrun);
173 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
177 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
179 rt_b->rt_period = ns_to_ktime(period);
180 rt_b->rt_runtime = runtime;
182 raw_spin_lock_init(&rt_b->rt_runtime_lock);
184 hrtimer_init(&rt_b->rt_period_timer,
185 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
186 rt_b->rt_period_timer.function = sched_rt_period_timer;
189 static inline int rt_bandwidth_enabled(void)
191 return sysctl_sched_rt_runtime >= 0;
194 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
198 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
201 if (hrtimer_active(&rt_b->rt_period_timer))
204 raw_spin_lock(&rt_b->rt_runtime_lock);
209 if (hrtimer_active(&rt_b->rt_period_timer))
212 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
213 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
215 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
216 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
217 delta = ktime_to_ns(ktime_sub(hard, soft));
218 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
219 HRTIMER_MODE_ABS_PINNED, 0);
221 raw_spin_unlock(&rt_b->rt_runtime_lock);
224 #ifdef CONFIG_RT_GROUP_SCHED
225 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
227 hrtimer_cancel(&rt_b->rt_period_timer);
232 * sched_domains_mutex serializes calls to arch_init_sched_domains,
233 * detach_destroy_domains and partition_sched_domains.
235 static DEFINE_MUTEX(sched_domains_mutex);
237 #ifdef CONFIG_CGROUP_SCHED
239 #include <linux/cgroup.h>
243 static LIST_HEAD(task_groups);
245 /* task group related information */
247 struct cgroup_subsys_state css;
249 #ifdef CONFIG_FAIR_GROUP_SCHED
250 /* schedulable entities of this group on each cpu */
251 struct sched_entity **se;
252 /* runqueue "owned" by this group on each cpu */
253 struct cfs_rq **cfs_rq;
254 unsigned long shares;
257 #ifdef CONFIG_RT_GROUP_SCHED
258 struct sched_rt_entity **rt_se;
259 struct rt_rq **rt_rq;
261 struct rt_bandwidth rt_bandwidth;
265 struct list_head list;
267 struct task_group *parent;
268 struct list_head siblings;
269 struct list_head children;
272 #define root_task_group init_task_group
274 /* task_group_lock serializes add/remove of task groups and also changes to
275 * a task group's cpu shares.
277 static DEFINE_SPINLOCK(task_group_lock);
279 #ifdef CONFIG_FAIR_GROUP_SCHED
282 static int root_task_group_empty(void)
284 return list_empty(&root_task_group.children);
288 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
291 * A weight of 0 or 1 can cause arithmetics problems.
292 * A weight of a cfs_rq is the sum of weights of which entities
293 * are queued on this cfs_rq, so a weight of a entity should not be
294 * too large, so as the shares value of a task group.
295 * (The default weight is 1024 - so there's no practical
296 * limitation from this.)
299 #define MAX_SHARES (1UL << 18)
301 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
304 /* Default task group.
305 * Every task in system belong to this group at bootup.
307 struct task_group init_task_group;
309 #endif /* CONFIG_CGROUP_SCHED */
311 /* CFS-related fields in a runqueue */
313 struct load_weight load;
314 unsigned long nr_running;
319 struct rb_root tasks_timeline;
320 struct rb_node *rb_leftmost;
322 struct list_head tasks;
323 struct list_head *balance_iterator;
326 * 'curr' points to currently running entity on this cfs_rq.
327 * It is set to NULL otherwise (i.e when none are currently running).
329 struct sched_entity *curr, *next, *last;
331 unsigned int nr_spread_over;
333 #ifdef CONFIG_FAIR_GROUP_SCHED
334 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
337 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
338 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
339 * (like users, containers etc.)
341 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
342 * list is used during load balance.
344 struct list_head leaf_cfs_rq_list;
345 struct task_group *tg; /* group that "owns" this runqueue */
349 * the part of load.weight contributed by tasks
351 unsigned long task_weight;
354 * h_load = weight * f(tg)
356 * Where f(tg) is the recursive weight fraction assigned to
359 unsigned long h_load;
362 * this cpu's part of tg->shares
364 unsigned long shares;
367 * load.weight at the time we set shares
369 unsigned long rq_weight;
374 /* Real-Time classes' related field in a runqueue: */
376 struct rt_prio_array active;
377 unsigned long rt_nr_running;
378 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
380 int curr; /* highest queued rt task prio */
382 int next; /* next highest */
387 unsigned long rt_nr_migratory;
388 unsigned long rt_nr_total;
390 struct plist_head pushable_tasks;
395 /* Nests inside the rq lock: */
396 raw_spinlock_t rt_runtime_lock;
398 #ifdef CONFIG_RT_GROUP_SCHED
399 unsigned long rt_nr_boosted;
402 struct list_head leaf_rt_rq_list;
403 struct task_group *tg;
410 * We add the notion of a root-domain which will be used to define per-domain
411 * variables. Each exclusive cpuset essentially defines an island domain by
412 * fully partitioning the member cpus from any other cpuset. Whenever a new
413 * exclusive cpuset is created, we also create and attach a new root-domain
420 cpumask_var_t online;
423 * The "RT overload" flag: it gets set if a CPU has more than
424 * one runnable RT task.
426 cpumask_var_t rto_mask;
429 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;
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];
461 unsigned char in_nohz_recently;
463 unsigned int skip_clock_update;
465 /* capture load from *all* tasks on this cpu: */
466 struct load_weight load;
467 unsigned long nr_load_updates;
473 #ifdef CONFIG_FAIR_GROUP_SCHED
474 /* list of leaf cfs_rq on this cpu: */
475 struct list_head leaf_cfs_rq_list;
477 #ifdef CONFIG_RT_GROUP_SCHED
478 struct list_head leaf_rt_rq_list;
482 * This is part of a global counter where only the total sum
483 * over all CPUs matters. A task can increase this counter on
484 * one CPU and if it got migrated afterwards it may decrease
485 * it on another CPU. Always updated under the runqueue lock:
487 unsigned long nr_uninterruptible;
489 struct task_struct *curr, *idle;
490 unsigned long next_balance;
491 struct mm_struct *prev_mm;
499 struct root_domain *rd;
500 struct sched_domain *sd;
502 unsigned long cpu_power;
504 unsigned char idle_at_tick;
505 /* For active balancing */
509 struct cpu_stop_work active_balance_work;
510 /* cpu of this runqueue: */
514 unsigned long avg_load_per_task;
522 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
526 /* calc_load related fields */
527 unsigned long calc_load_update;
528 long calc_load_active;
530 #ifdef CONFIG_SCHED_HRTICK
532 int hrtick_csd_pending;
533 struct call_single_data hrtick_csd;
535 struct hrtimer hrtick_timer;
538 #ifdef CONFIG_SCHEDSTATS
540 struct sched_info rq_sched_info;
541 unsigned long long rq_cpu_time;
542 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
544 /* sys_sched_yield() stats */
545 unsigned int yld_count;
547 /* schedule() stats */
548 unsigned int sched_switch;
549 unsigned int sched_count;
550 unsigned int sched_goidle;
552 /* try_to_wake_up() stats */
553 unsigned int ttwu_count;
554 unsigned int ttwu_local;
557 unsigned int bkl_count;
561 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
564 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
566 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
569 * A queue event has occurred, and we're going to schedule. In
570 * this case, we can save a useless back to back clock update.
572 if (rq->curr->se.on_rq && test_tsk_need_resched(rq->curr))
573 rq->skip_clock_update = 1;
576 static inline int cpu_of(struct rq *rq)
585 #define rcu_dereference_check_sched_domain(p) \
586 rcu_dereference_check((p), \
587 rcu_read_lock_sched_held() || \
588 lockdep_is_held(&sched_domains_mutex))
591 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
592 * See detach_destroy_domains: synchronize_sched for details.
594 * The domain tree of any CPU may only be accessed from within
595 * preempt-disabled sections.
597 #define for_each_domain(cpu, __sd) \
598 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
600 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
601 #define this_rq() (&__get_cpu_var(runqueues))
602 #define task_rq(p) cpu_rq(task_cpu(p))
603 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
604 #define raw_rq() (&__raw_get_cpu_var(runqueues))
606 #ifdef CONFIG_CGROUP_SCHED
609 * Return the group to which this tasks belongs.
611 * We use task_subsys_state_check() and extend the RCU verification
612 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
613 * holds that lock for each task it moves into the cgroup. Therefore
614 * by holding that lock, we pin the task to the current cgroup.
616 static inline struct task_group *task_group(struct task_struct *p)
618 struct cgroup_subsys_state *css;
620 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
621 lockdep_is_held(&task_rq(p)->lock));
622 return container_of(css, struct task_group, css);
625 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
626 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
628 #ifdef CONFIG_FAIR_GROUP_SCHED
629 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
630 p->se.parent = task_group(p)->se[cpu];
633 #ifdef CONFIG_RT_GROUP_SCHED
634 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
635 p->rt.parent = task_group(p)->rt_se[cpu];
639 #else /* CONFIG_CGROUP_SCHED */
641 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
642 static inline struct task_group *task_group(struct task_struct *p)
647 #endif /* CONFIG_CGROUP_SCHED */
649 static u64 irq_time_cpu(int cpu);
650 static void sched_irq_time_avg_update(struct rq *rq, u64 irq_time);
652 inline void update_rq_clock(struct rq *rq)
654 int cpu = cpu_of(rq);
657 if (!rq->skip_clock_update)
658 rq->clock = sched_clock_cpu(cpu_of(rq));
659 irq_time = irq_time_cpu(cpu);
660 if (rq->clock - irq_time > rq->clock_task)
661 rq->clock_task = rq->clock - irq_time;
663 sched_irq_time_avg_update(rq, irq_time);
667 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
669 #ifdef CONFIG_SCHED_DEBUG
670 # define const_debug __read_mostly
672 # define const_debug static const
677 * @cpu: the processor in question.
679 * Returns true if the current cpu runqueue is locked.
680 * This interface allows printk to be called with the runqueue lock
681 * held and know whether or not it is OK to wake up the klogd.
683 int runqueue_is_locked(int cpu)
685 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
689 * Debugging: various feature bits
692 #define SCHED_FEAT(name, enabled) \
693 __SCHED_FEAT_##name ,
696 #include "sched_features.h"
701 #define SCHED_FEAT(name, enabled) \
702 (1UL << __SCHED_FEAT_##name) * enabled |
704 const_debug unsigned int sysctl_sched_features =
705 #include "sched_features.h"
710 #ifdef CONFIG_SCHED_DEBUG
711 #define SCHED_FEAT(name, enabled) \
714 static __read_mostly char *sched_feat_names[] = {
715 #include "sched_features.h"
721 static int sched_feat_show(struct seq_file *m, void *v)
725 for (i = 0; sched_feat_names[i]; i++) {
726 if (!(sysctl_sched_features & (1UL << i)))
728 seq_printf(m, "%s ", sched_feat_names[i]);
736 sched_feat_write(struct file *filp, const char __user *ubuf,
737 size_t cnt, loff_t *ppos)
747 if (copy_from_user(&buf, ubuf, cnt))
753 if (strncmp(buf, "NO_", 3) == 0) {
758 for (i = 0; sched_feat_names[i]; i++) {
759 if (strcmp(cmp, sched_feat_names[i]) == 0) {
761 sysctl_sched_features &= ~(1UL << i);
763 sysctl_sched_features |= (1UL << i);
768 if (!sched_feat_names[i])
776 static int sched_feat_open(struct inode *inode, struct file *filp)
778 return single_open(filp, sched_feat_show, NULL);
781 static const struct file_operations sched_feat_fops = {
782 .open = sched_feat_open,
783 .write = sched_feat_write,
786 .release = single_release,
789 static __init int sched_init_debug(void)
791 debugfs_create_file("sched_features", 0644, NULL, NULL,
796 late_initcall(sched_init_debug);
800 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
803 * Number of tasks to iterate in a single balance run.
804 * Limited because this is done with IRQs disabled.
806 const_debug unsigned int sysctl_sched_nr_migrate = 32;
809 * ratelimit for updating the group shares.
812 unsigned int sysctl_sched_shares_ratelimit = 250000;
813 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
816 * Inject some fuzzyness into changing the per-cpu group shares
817 * this avoids remote rq-locks at the expense of fairness.
820 unsigned int sysctl_sched_shares_thresh = 4;
823 * period over which we average the RT time consumption, measured
828 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
831 * period over which we measure -rt task cpu usage in us.
834 unsigned int sysctl_sched_rt_period = 1000000;
836 static __read_mostly int scheduler_running;
839 * part of the period that we allow rt tasks to run in us.
842 int sysctl_sched_rt_runtime = 950000;
844 static inline u64 global_rt_period(void)
846 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
849 static inline u64 global_rt_runtime(void)
851 if (sysctl_sched_rt_runtime < 0)
854 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
857 #ifndef prepare_arch_switch
858 # define prepare_arch_switch(next) do { } while (0)
860 #ifndef finish_arch_switch
861 # define finish_arch_switch(prev) do { } while (0)
864 static inline int task_current(struct rq *rq, struct task_struct *p)
866 return rq->curr == p;
869 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
870 static inline int task_running(struct rq *rq, struct task_struct *p)
872 return task_current(rq, p);
875 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
879 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
881 #ifdef CONFIG_DEBUG_SPINLOCK
882 /* this is a valid case when another task releases the spinlock */
883 rq->lock.owner = current;
886 * If we are tracking spinlock dependencies then we have to
887 * fix up the runqueue lock - which gets 'carried over' from
890 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
892 raw_spin_unlock_irq(&rq->lock);
895 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
896 static inline int task_running(struct rq *rq, struct task_struct *p)
901 return task_current(rq, p);
905 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
909 * We can optimise this out completely for !SMP, because the
910 * SMP rebalancing from interrupt is the only thing that cares
915 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
916 raw_spin_unlock_irq(&rq->lock);
918 raw_spin_unlock(&rq->lock);
922 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
926 * After ->oncpu is cleared, the task can be moved to a different CPU.
927 * We must ensure this doesn't happen until the switch is completely
933 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
937 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
940 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
943 static inline int task_is_waking(struct task_struct *p)
945 return unlikely(p->state == TASK_WAKING);
949 * __task_rq_lock - lock the runqueue a given task resides on.
950 * Must be called interrupts disabled.
952 static inline struct rq *__task_rq_lock(struct task_struct *p)
959 raw_spin_lock(&rq->lock);
960 if (likely(rq == task_rq(p)))
962 raw_spin_unlock(&rq->lock);
967 * task_rq_lock - lock the runqueue a given task resides on and disable
968 * interrupts. Note the ordering: we can safely lookup the task_rq without
969 * explicitly disabling preemption.
971 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
977 local_irq_save(*flags);
979 raw_spin_lock(&rq->lock);
980 if (likely(rq == task_rq(p)))
982 raw_spin_unlock_irqrestore(&rq->lock, *flags);
986 static void __task_rq_unlock(struct rq *rq)
989 raw_spin_unlock(&rq->lock);
992 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
995 raw_spin_unlock_irqrestore(&rq->lock, *flags);
999 * this_rq_lock - lock this runqueue and disable interrupts.
1001 static struct rq *this_rq_lock(void)
1002 __acquires(rq->lock)
1006 local_irq_disable();
1008 raw_spin_lock(&rq->lock);
1013 #ifdef CONFIG_SCHED_HRTICK
1015 * Use HR-timers to deliver accurate preemption points.
1017 * Its all a bit involved since we cannot program an hrt while holding the
1018 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1021 * When we get rescheduled we reprogram the hrtick_timer outside of the
1027 * - enabled by features
1028 * - hrtimer is actually high res
1030 static inline int hrtick_enabled(struct rq *rq)
1032 if (!sched_feat(HRTICK))
1034 if (!cpu_active(cpu_of(rq)))
1036 return hrtimer_is_hres_active(&rq->hrtick_timer);
1039 static void hrtick_clear(struct rq *rq)
1041 if (hrtimer_active(&rq->hrtick_timer))
1042 hrtimer_cancel(&rq->hrtick_timer);
1046 * High-resolution timer tick.
1047 * Runs from hardirq context with interrupts disabled.
1049 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1051 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1053 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1055 raw_spin_lock(&rq->lock);
1056 update_rq_clock(rq);
1057 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1058 raw_spin_unlock(&rq->lock);
1060 return HRTIMER_NORESTART;
1065 * called from hardirq (IPI) context
1067 static void __hrtick_start(void *arg)
1069 struct rq *rq = arg;
1071 raw_spin_lock(&rq->lock);
1072 hrtimer_restart(&rq->hrtick_timer);
1073 rq->hrtick_csd_pending = 0;
1074 raw_spin_unlock(&rq->lock);
1078 * Called to set the hrtick timer state.
1080 * called with rq->lock held and irqs disabled
1082 static void hrtick_start(struct rq *rq, u64 delay)
1084 struct hrtimer *timer = &rq->hrtick_timer;
1085 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1087 hrtimer_set_expires(timer, time);
1089 if (rq == this_rq()) {
1090 hrtimer_restart(timer);
1091 } else if (!rq->hrtick_csd_pending) {
1092 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1093 rq->hrtick_csd_pending = 1;
1098 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1100 int cpu = (int)(long)hcpu;
1103 case CPU_UP_CANCELED:
1104 case CPU_UP_CANCELED_FROZEN:
1105 case CPU_DOWN_PREPARE:
1106 case CPU_DOWN_PREPARE_FROZEN:
1108 case CPU_DEAD_FROZEN:
1109 hrtick_clear(cpu_rq(cpu));
1116 static __init void init_hrtick(void)
1118 hotcpu_notifier(hotplug_hrtick, 0);
1122 * Called to set the hrtick timer state.
1124 * called with rq->lock held and irqs disabled
1126 static void hrtick_start(struct rq *rq, u64 delay)
1128 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1129 HRTIMER_MODE_REL_PINNED, 0);
1132 static inline void init_hrtick(void)
1135 #endif /* CONFIG_SMP */
1137 static void init_rq_hrtick(struct rq *rq)
1140 rq->hrtick_csd_pending = 0;
1142 rq->hrtick_csd.flags = 0;
1143 rq->hrtick_csd.func = __hrtick_start;
1144 rq->hrtick_csd.info = rq;
1147 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1148 rq->hrtick_timer.function = hrtick;
1150 #else /* CONFIG_SCHED_HRTICK */
1151 static inline void hrtick_clear(struct rq *rq)
1155 static inline void init_rq_hrtick(struct rq *rq)
1159 static inline void init_hrtick(void)
1162 #endif /* CONFIG_SCHED_HRTICK */
1165 * resched_task - mark a task 'to be rescheduled now'.
1167 * On UP this means the setting of the need_resched flag, on SMP it
1168 * might also involve a cross-CPU call to trigger the scheduler on
1173 #ifndef tsk_is_polling
1174 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1177 static void resched_task(struct task_struct *p)
1181 assert_raw_spin_locked(&task_rq(p)->lock);
1183 if (test_tsk_need_resched(p))
1186 set_tsk_need_resched(p);
1189 if (cpu == smp_processor_id())
1192 /* NEED_RESCHED must be visible before we test polling */
1194 if (!tsk_is_polling(p))
1195 smp_send_reschedule(cpu);
1198 static void resched_cpu(int cpu)
1200 struct rq *rq = cpu_rq(cpu);
1201 unsigned long flags;
1203 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1205 resched_task(cpu_curr(cpu));
1206 raw_spin_unlock_irqrestore(&rq->lock, flags);
1211 * When add_timer_on() enqueues a timer into the timer wheel of an
1212 * idle CPU then this timer might expire before the next timer event
1213 * which is scheduled to wake up that CPU. In case of a completely
1214 * idle system the next event might even be infinite time into the
1215 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1216 * leaves the inner idle loop so the newly added timer is taken into
1217 * account when the CPU goes back to idle and evaluates the timer
1218 * wheel for the next timer event.
1220 void wake_up_idle_cpu(int cpu)
1222 struct rq *rq = cpu_rq(cpu);
1224 if (cpu == smp_processor_id())
1228 * This is safe, as this function is called with the timer
1229 * wheel base lock of (cpu) held. When the CPU is on the way
1230 * to idle and has not yet set rq->curr to idle then it will
1231 * be serialized on the timer wheel base lock and take the new
1232 * timer into account automatically.
1234 if (rq->curr != rq->idle)
1238 * We can set TIF_RESCHED on the idle task of the other CPU
1239 * lockless. The worst case is that the other CPU runs the
1240 * idle task through an additional NOOP schedule()
1242 set_tsk_need_resched(rq->idle);
1244 /* NEED_RESCHED must be visible before we test polling */
1246 if (!tsk_is_polling(rq->idle))
1247 smp_send_reschedule(cpu);
1250 #endif /* CONFIG_NO_HZ */
1252 static u64 sched_avg_period(void)
1254 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1257 static void sched_avg_update(struct rq *rq)
1259 s64 period = sched_avg_period();
1261 while ((s64)(rq->clock - rq->age_stamp) > period) {
1263 * Inline assembly required to prevent the compiler
1264 * optimising this loop into a divmod call.
1265 * See __iter_div_u64_rem() for another example of this.
1267 asm("" : "+rm" (rq->age_stamp));
1268 rq->age_stamp += period;
1273 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1275 rq->rt_avg += rt_delta;
1276 sched_avg_update(rq);
1279 #else /* !CONFIG_SMP */
1280 static void resched_task(struct task_struct *p)
1282 assert_raw_spin_locked(&task_rq(p)->lock);
1283 set_tsk_need_resched(p);
1286 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1290 static void sched_avg_update(struct rq *rq)
1293 #endif /* CONFIG_SMP */
1295 #if BITS_PER_LONG == 32
1296 # define WMULT_CONST (~0UL)
1298 # define WMULT_CONST (1UL << 32)
1301 #define WMULT_SHIFT 32
1304 * Shift right and round:
1306 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1309 * delta *= weight / lw
1311 static unsigned long
1312 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1313 struct load_weight *lw)
1317 if (!lw->inv_weight) {
1318 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1321 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1325 tmp = (u64)delta_exec * weight;
1327 * Check whether we'd overflow the 64-bit multiplication:
1329 if (unlikely(tmp > WMULT_CONST))
1330 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1333 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1335 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1338 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1344 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1351 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1352 * of tasks with abnormal "nice" values across CPUs the contribution that
1353 * each task makes to its run queue's load is weighted according to its
1354 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1355 * scaled version of the new time slice allocation that they receive on time
1359 #define WEIGHT_IDLEPRIO 3
1360 #define WMULT_IDLEPRIO 1431655765
1363 * Nice levels are multiplicative, with a gentle 10% change for every
1364 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1365 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1366 * that remained on nice 0.
1368 * The "10% effect" is relative and cumulative: from _any_ nice level,
1369 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1370 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1371 * If a task goes up by ~10% and another task goes down by ~10% then
1372 * the relative distance between them is ~25%.)
1374 static const int prio_to_weight[40] = {
1375 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1376 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1377 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1378 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1379 /* 0 */ 1024, 820, 655, 526, 423,
1380 /* 5 */ 335, 272, 215, 172, 137,
1381 /* 10 */ 110, 87, 70, 56, 45,
1382 /* 15 */ 36, 29, 23, 18, 15,
1386 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1388 * In cases where the weight does not change often, we can use the
1389 * precalculated inverse to speed up arithmetics by turning divisions
1390 * into multiplications:
1392 static const u32 prio_to_wmult[40] = {
1393 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1394 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1395 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1396 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1397 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1398 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1399 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1400 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1403 /* Time spent by the tasks of the cpu accounting group executing in ... */
1404 enum cpuacct_stat_index {
1405 CPUACCT_STAT_USER, /* ... user mode */
1406 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1408 CPUACCT_STAT_NSTATS,
1411 #ifdef CONFIG_CGROUP_CPUACCT
1412 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1413 static void cpuacct_update_stats(struct task_struct *tsk,
1414 enum cpuacct_stat_index idx, cputime_t val);
1416 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1417 static inline void cpuacct_update_stats(struct task_struct *tsk,
1418 enum cpuacct_stat_index idx, cputime_t val) {}
1421 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1423 update_load_add(&rq->load, load);
1426 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1428 update_load_sub(&rq->load, load);
1431 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1432 typedef int (*tg_visitor)(struct task_group *, void *);
1435 * Iterate the full tree, calling @down when first entering a node and @up when
1436 * leaving it for the final time.
1438 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1440 struct task_group *parent, *child;
1444 parent = &root_task_group;
1446 ret = (*down)(parent, data);
1449 list_for_each_entry_rcu(child, &parent->children, siblings) {
1456 ret = (*up)(parent, data);
1461 parent = parent->parent;
1470 static int tg_nop(struct task_group *tg, void *data)
1477 /* Used instead of source_load when we know the type == 0 */
1478 static unsigned long weighted_cpuload(const int cpu)
1480 return cpu_rq(cpu)->load.weight;
1484 * Return a low guess at the load of a migration-source cpu weighted
1485 * according to the scheduling class and "nice" value.
1487 * We want to under-estimate the load of migration sources, to
1488 * balance conservatively.
1490 static unsigned long source_load(int cpu, int type)
1492 struct rq *rq = cpu_rq(cpu);
1493 unsigned long total = weighted_cpuload(cpu);
1495 if (type == 0 || !sched_feat(LB_BIAS))
1498 return min(rq->cpu_load[type-1], total);
1502 * Return a high guess at the load of a migration-target cpu weighted
1503 * according to the scheduling class and "nice" value.
1505 static unsigned long target_load(int cpu, int type)
1507 struct rq *rq = cpu_rq(cpu);
1508 unsigned long total = weighted_cpuload(cpu);
1510 if (type == 0 || !sched_feat(LB_BIAS))
1513 return max(rq->cpu_load[type-1], total);
1516 static unsigned long power_of(int cpu)
1518 return cpu_rq(cpu)->cpu_power;
1521 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1523 static unsigned long cpu_avg_load_per_task(int cpu)
1525 struct rq *rq = cpu_rq(cpu);
1526 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1529 rq->avg_load_per_task = rq->load.weight / nr_running;
1531 rq->avg_load_per_task = 0;
1533 return rq->avg_load_per_task;
1536 #ifdef CONFIG_FAIR_GROUP_SCHED
1538 static __read_mostly unsigned long __percpu *update_shares_data;
1540 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1543 * Calculate and set the cpu's group shares.
1545 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1546 unsigned long sd_shares,
1547 unsigned long sd_rq_weight,
1548 unsigned long *usd_rq_weight)
1550 unsigned long shares, rq_weight;
1553 rq_weight = usd_rq_weight[cpu];
1556 rq_weight = NICE_0_LOAD;
1560 * \Sum_j shares_j * rq_weight_i
1561 * shares_i = -----------------------------
1562 * \Sum_j rq_weight_j
1564 shares = (sd_shares * rq_weight) / sd_rq_weight;
1565 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1567 if (abs(shares - tg->se[cpu]->load.weight) >
1568 sysctl_sched_shares_thresh) {
1569 struct rq *rq = cpu_rq(cpu);
1570 unsigned long flags;
1572 raw_spin_lock_irqsave(&rq->lock, flags);
1573 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1574 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1575 __set_se_shares(tg->se[cpu], shares);
1576 raw_spin_unlock_irqrestore(&rq->lock, flags);
1581 * Re-compute the task group their per cpu shares over the given domain.
1582 * This needs to be done in a bottom-up fashion because the rq weight of a
1583 * parent group depends on the shares of its child groups.
1585 static int tg_shares_up(struct task_group *tg, void *data)
1587 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1588 unsigned long *usd_rq_weight;
1589 struct sched_domain *sd = data;
1590 unsigned long flags;
1596 local_irq_save(flags);
1597 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1599 for_each_cpu(i, sched_domain_span(sd)) {
1600 weight = tg->cfs_rq[i]->load.weight;
1601 usd_rq_weight[i] = weight;
1603 rq_weight += weight;
1605 * If there are currently no tasks on the cpu pretend there
1606 * is one of average load so that when a new task gets to
1607 * run here it will not get delayed by group starvation.
1610 weight = NICE_0_LOAD;
1612 sum_weight += weight;
1613 shares += tg->cfs_rq[i]->shares;
1617 rq_weight = sum_weight;
1619 if ((!shares && rq_weight) || shares > tg->shares)
1620 shares = tg->shares;
1622 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1623 shares = tg->shares;
1625 for_each_cpu(i, sched_domain_span(sd))
1626 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1628 local_irq_restore(flags);
1634 * Compute the cpu's hierarchical load factor for each task group.
1635 * This needs to be done in a top-down fashion because the load of a child
1636 * group is a fraction of its parents load.
1638 static int tg_load_down(struct task_group *tg, void *data)
1641 long cpu = (long)data;
1644 load = cpu_rq(cpu)->load.weight;
1646 load = tg->parent->cfs_rq[cpu]->h_load;
1647 load *= tg->cfs_rq[cpu]->shares;
1648 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1651 tg->cfs_rq[cpu]->h_load = load;
1656 static void update_shares(struct sched_domain *sd)
1661 if (root_task_group_empty())
1664 now = cpu_clock(raw_smp_processor_id());
1665 elapsed = now - sd->last_update;
1667 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1668 sd->last_update = now;
1669 walk_tg_tree(tg_nop, tg_shares_up, sd);
1673 static void update_h_load(long cpu)
1675 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1680 static inline void update_shares(struct sched_domain *sd)
1686 #ifdef CONFIG_PREEMPT
1688 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1691 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1692 * way at the expense of forcing extra atomic operations in all
1693 * invocations. This assures that the double_lock is acquired using the
1694 * same underlying policy as the spinlock_t on this architecture, which
1695 * reduces latency compared to the unfair variant below. However, it
1696 * also adds more overhead and therefore may reduce throughput.
1698 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1699 __releases(this_rq->lock)
1700 __acquires(busiest->lock)
1701 __acquires(this_rq->lock)
1703 raw_spin_unlock(&this_rq->lock);
1704 double_rq_lock(this_rq, busiest);
1711 * Unfair double_lock_balance: Optimizes throughput at the expense of
1712 * latency by eliminating extra atomic operations when the locks are
1713 * already in proper order on entry. This favors lower cpu-ids and will
1714 * grant the double lock to lower cpus over higher ids under contention,
1715 * regardless of entry order into the function.
1717 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1718 __releases(this_rq->lock)
1719 __acquires(busiest->lock)
1720 __acquires(this_rq->lock)
1724 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1725 if (busiest < this_rq) {
1726 raw_spin_unlock(&this_rq->lock);
1727 raw_spin_lock(&busiest->lock);
1728 raw_spin_lock_nested(&this_rq->lock,
1729 SINGLE_DEPTH_NESTING);
1732 raw_spin_lock_nested(&busiest->lock,
1733 SINGLE_DEPTH_NESTING);
1738 #endif /* CONFIG_PREEMPT */
1741 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1743 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1745 if (unlikely(!irqs_disabled())) {
1746 /* printk() doesn't work good under rq->lock */
1747 raw_spin_unlock(&this_rq->lock);
1751 return _double_lock_balance(this_rq, busiest);
1754 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1755 __releases(busiest->lock)
1757 raw_spin_unlock(&busiest->lock);
1758 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1762 * double_rq_lock - safely lock two runqueues
1764 * Note this does not disable interrupts like task_rq_lock,
1765 * you need to do so manually before calling.
1767 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1768 __acquires(rq1->lock)
1769 __acquires(rq2->lock)
1771 BUG_ON(!irqs_disabled());
1773 raw_spin_lock(&rq1->lock);
1774 __acquire(rq2->lock); /* Fake it out ;) */
1777 raw_spin_lock(&rq1->lock);
1778 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1780 raw_spin_lock(&rq2->lock);
1781 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1787 * double_rq_unlock - safely unlock two runqueues
1789 * Note this does not restore interrupts like task_rq_unlock,
1790 * you need to do so manually after calling.
1792 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1793 __releases(rq1->lock)
1794 __releases(rq2->lock)
1796 raw_spin_unlock(&rq1->lock);
1798 raw_spin_unlock(&rq2->lock);
1800 __release(rq2->lock);
1805 #ifdef CONFIG_FAIR_GROUP_SCHED
1806 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1809 cfs_rq->shares = shares;
1814 static void calc_load_account_idle(struct rq *this_rq);
1815 static void update_sysctl(void);
1816 static int get_update_sysctl_factor(void);
1818 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1820 set_task_rq(p, cpu);
1823 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1824 * successfuly executed on another CPU. We must ensure that updates of
1825 * per-task data have been completed by this moment.
1828 task_thread_info(p)->cpu = cpu;
1832 static const struct sched_class rt_sched_class;
1834 #define sched_class_highest (&rt_sched_class)
1835 #define for_each_class(class) \
1836 for (class = sched_class_highest; class; class = class->next)
1838 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1841 * There are no locks covering percpu hardirq/softirq time.
1842 * They are only modified in account_system_vtime, on corresponding CPU
1843 * with interrupts disabled. So, writes are safe.
1844 * They are read and saved off onto struct rq in update_rq_clock().
1845 * This may result in other CPU reading this CPU's irq time and can
1846 * race with irq/account_system_vtime on this CPU. We would either get old
1847 * or new value (or semi updated value on 32 bit) with a side effect of
1848 * accounting a slice of irq time to wrong task when irq is in progress
1849 * while we read rq->clock. That is a worthy compromise in place of having
1850 * locks on each irq in account_system_time.
1852 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1853 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1855 static DEFINE_PER_CPU(u64, irq_start_time);
1856 static int sched_clock_irqtime;
1858 void enable_sched_clock_irqtime(void)
1860 sched_clock_irqtime = 1;
1863 void disable_sched_clock_irqtime(void)
1865 sched_clock_irqtime = 0;
1868 static u64 irq_time_cpu(int cpu)
1870 if (!sched_clock_irqtime)
1873 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1876 void account_system_vtime(struct task_struct *curr)
1878 unsigned long flags;
1882 if (!sched_clock_irqtime)
1885 local_irq_save(flags);
1887 cpu = smp_processor_id();
1888 now = sched_clock_cpu(cpu);
1889 delta = now - per_cpu(irq_start_time, cpu);
1890 per_cpu(irq_start_time, cpu) = now;
1892 * We do not account for softirq time from ksoftirqd here.
1893 * We want to continue accounting softirq time to ksoftirqd thread
1894 * in that case, so as not to confuse scheduler with a special task
1895 * that do not consume any time, but still wants to run.
1897 if (hardirq_count())
1898 per_cpu(cpu_hardirq_time, cpu) += delta;
1899 else if (in_serving_softirq() && !(curr->flags & PF_KSOFTIRQD))
1900 per_cpu(cpu_softirq_time, cpu) += delta;
1902 local_irq_restore(flags);
1905 static void sched_irq_time_avg_update(struct rq *rq, u64 curr_irq_time)
1907 if (sched_clock_irqtime && sched_feat(NONIRQ_POWER)) {
1908 u64 delta_irq = curr_irq_time - rq->prev_irq_time;
1909 rq->prev_irq_time = curr_irq_time;
1910 sched_rt_avg_update(rq, delta_irq);
1916 static u64 irq_time_cpu(int cpu)
1921 static void sched_irq_time_avg_update(struct rq *rq, u64 curr_irq_time) { }
1925 #include "sched_stats.h"
1927 static void inc_nr_running(struct rq *rq)
1932 static void dec_nr_running(struct rq *rq)
1937 static void set_load_weight(struct task_struct *p)
1940 * SCHED_IDLE tasks get minimal weight:
1942 if (p->policy == SCHED_IDLE) {
1943 p->se.load.weight = WEIGHT_IDLEPRIO;
1944 p->se.load.inv_weight = WMULT_IDLEPRIO;
1948 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1949 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1952 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1954 update_rq_clock(rq);
1955 sched_info_queued(p);
1956 p->sched_class->enqueue_task(rq, p, flags);
1960 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1962 update_rq_clock(rq);
1963 sched_info_dequeued(p);
1964 p->sched_class->dequeue_task(rq, p, flags);
1969 * activate_task - move a task to the runqueue.
1971 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1973 if (task_contributes_to_load(p))
1974 rq->nr_uninterruptible--;
1976 enqueue_task(rq, p, flags);
1981 * deactivate_task - remove a task from the runqueue.
1983 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1985 if (task_contributes_to_load(p))
1986 rq->nr_uninterruptible++;
1988 dequeue_task(rq, p, flags);
1992 #include "sched_idletask.c"
1993 #include "sched_fair.c"
1994 #include "sched_rt.c"
1995 #ifdef CONFIG_SCHED_DEBUG
1996 # include "sched_debug.c"
2000 * __normal_prio - return the priority that is based on the static prio
2002 static inline int __normal_prio(struct task_struct *p)
2004 return p->static_prio;
2008 * Calculate the expected normal priority: i.e. priority
2009 * without taking RT-inheritance into account. Might be
2010 * boosted by interactivity modifiers. Changes upon fork,
2011 * setprio syscalls, and whenever the interactivity
2012 * estimator recalculates.
2014 static inline int normal_prio(struct task_struct *p)
2018 if (task_has_rt_policy(p))
2019 prio = MAX_RT_PRIO-1 - p->rt_priority;
2021 prio = __normal_prio(p);
2026 * Calculate the current priority, i.e. the priority
2027 * taken into account by the scheduler. This value might
2028 * be boosted by RT tasks, or might be boosted by
2029 * interactivity modifiers. Will be RT if the task got
2030 * RT-boosted. If not then it returns p->normal_prio.
2032 static int effective_prio(struct task_struct *p)
2034 p->normal_prio = normal_prio(p);
2036 * If we are RT tasks or we were boosted to RT priority,
2037 * keep the priority unchanged. Otherwise, update priority
2038 * to the normal priority:
2040 if (!rt_prio(p->prio))
2041 return p->normal_prio;
2046 * task_curr - is this task currently executing on a CPU?
2047 * @p: the task in question.
2049 inline int task_curr(const struct task_struct *p)
2051 return cpu_curr(task_cpu(p)) == p;
2054 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2055 const struct sched_class *prev_class,
2056 int oldprio, int running)
2058 if (prev_class != p->sched_class) {
2059 if (prev_class->switched_from)
2060 prev_class->switched_from(rq, p, running);
2061 p->sched_class->switched_to(rq, p, running);
2063 p->sched_class->prio_changed(rq, p, oldprio, running);
2068 * Is this task likely cache-hot:
2071 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2075 if (p->sched_class != &fair_sched_class)
2078 if (unlikely(p->policy == SCHED_IDLE))
2082 * Buddy candidates are cache hot:
2084 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2085 (&p->se == cfs_rq_of(&p->se)->next ||
2086 &p->se == cfs_rq_of(&p->se)->last))
2089 if (sysctl_sched_migration_cost == -1)
2091 if (sysctl_sched_migration_cost == 0)
2094 delta = now - p->se.exec_start;
2096 return delta < (s64)sysctl_sched_migration_cost;
2099 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2101 #ifdef CONFIG_SCHED_DEBUG
2103 * We should never call set_task_cpu() on a blocked task,
2104 * ttwu() will sort out the placement.
2106 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2107 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2110 trace_sched_migrate_task(p, new_cpu);
2112 if (task_cpu(p) != new_cpu) {
2113 p->se.nr_migrations++;
2114 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2117 __set_task_cpu(p, new_cpu);
2120 struct migration_arg {
2121 struct task_struct *task;
2125 static int migration_cpu_stop(void *data);
2128 * The task's runqueue lock must be held.
2129 * Returns true if you have to wait for migration thread.
2131 static bool migrate_task(struct task_struct *p, int dest_cpu)
2133 struct rq *rq = task_rq(p);
2136 * If the task is not on a runqueue (and not running), then
2137 * the next wake-up will properly place the task.
2139 return p->se.on_rq || task_running(rq, p);
2143 * wait_task_inactive - wait for a thread to unschedule.
2145 * If @match_state is nonzero, it's the @p->state value just checked and
2146 * not expected to change. If it changes, i.e. @p might have woken up,
2147 * then return zero. When we succeed in waiting for @p to be off its CPU,
2148 * we return a positive number (its total switch count). If a second call
2149 * a short while later returns the same number, the caller can be sure that
2150 * @p has remained unscheduled the whole time.
2152 * The caller must ensure that the task *will* unschedule sometime soon,
2153 * else this function might spin for a *long* time. This function can't
2154 * be called with interrupts off, or it may introduce deadlock with
2155 * smp_call_function() if an IPI is sent by the same process we are
2156 * waiting to become inactive.
2158 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2160 unsigned long flags;
2167 * We do the initial early heuristics without holding
2168 * any task-queue locks at all. We'll only try to get
2169 * the runqueue lock when things look like they will
2175 * If the task is actively running on another CPU
2176 * still, just relax and busy-wait without holding
2179 * NOTE! Since we don't hold any locks, it's not
2180 * even sure that "rq" stays as the right runqueue!
2181 * But we don't care, since "task_running()" will
2182 * return false if the runqueue has changed and p
2183 * is actually now running somewhere else!
2185 while (task_running(rq, p)) {
2186 if (match_state && unlikely(p->state != match_state))
2192 * Ok, time to look more closely! We need the rq
2193 * lock now, to be *sure*. If we're wrong, we'll
2194 * just go back and repeat.
2196 rq = task_rq_lock(p, &flags);
2197 trace_sched_wait_task(p);
2198 running = task_running(rq, p);
2199 on_rq = p->se.on_rq;
2201 if (!match_state || p->state == match_state)
2202 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2203 task_rq_unlock(rq, &flags);
2206 * If it changed from the expected state, bail out now.
2208 if (unlikely(!ncsw))
2212 * Was it really running after all now that we
2213 * checked with the proper locks actually held?
2215 * Oops. Go back and try again..
2217 if (unlikely(running)) {
2223 * It's not enough that it's not actively running,
2224 * it must be off the runqueue _entirely_, and not
2227 * So if it was still runnable (but just not actively
2228 * running right now), it's preempted, and we should
2229 * yield - it could be a while.
2231 if (unlikely(on_rq)) {
2232 schedule_timeout_uninterruptible(1);
2237 * Ahh, all good. It wasn't running, and it wasn't
2238 * runnable, which means that it will never become
2239 * running in the future either. We're all done!
2248 * kick_process - kick a running thread to enter/exit the kernel
2249 * @p: the to-be-kicked thread
2251 * Cause a process which is running on another CPU to enter
2252 * kernel-mode, without any delay. (to get signals handled.)
2254 * NOTE: this function doesnt have to take the runqueue lock,
2255 * because all it wants to ensure is that the remote task enters
2256 * the kernel. If the IPI races and the task has been migrated
2257 * to another CPU then no harm is done and the purpose has been
2260 void kick_process(struct task_struct *p)
2266 if ((cpu != smp_processor_id()) && task_curr(p))
2267 smp_send_reschedule(cpu);
2270 EXPORT_SYMBOL_GPL(kick_process);
2271 #endif /* CONFIG_SMP */
2274 * task_oncpu_function_call - call a function on the cpu on which a task runs
2275 * @p: the task to evaluate
2276 * @func: the function to be called
2277 * @info: the function call argument
2279 * Calls the function @func when the task is currently running. This might
2280 * be on the current CPU, which just calls the function directly
2282 void task_oncpu_function_call(struct task_struct *p,
2283 void (*func) (void *info), void *info)
2290 smp_call_function_single(cpu, func, info, 1);
2296 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2298 static int select_fallback_rq(int cpu, struct task_struct *p)
2301 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2303 /* Look for allowed, online CPU in same node. */
2304 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2305 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2308 /* Any allowed, online CPU? */
2309 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2310 if (dest_cpu < nr_cpu_ids)
2313 /* No more Mr. Nice Guy. */
2314 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2315 dest_cpu = cpuset_cpus_allowed_fallback(p);
2317 * Don't tell them about moving exiting tasks or
2318 * kernel threads (both mm NULL), since they never
2321 if (p->mm && printk_ratelimit()) {
2322 printk(KERN_INFO "process %d (%s) no "
2323 "longer affine to cpu%d\n",
2324 task_pid_nr(p), p->comm, cpu);
2332 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2335 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2337 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2340 * In order not to call set_task_cpu() on a blocking task we need
2341 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2344 * Since this is common to all placement strategies, this lives here.
2346 * [ this allows ->select_task() to simply return task_cpu(p) and
2347 * not worry about this generic constraint ]
2349 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2351 cpu = select_fallback_rq(task_cpu(p), p);
2356 static void update_avg(u64 *avg, u64 sample)
2358 s64 diff = sample - *avg;
2364 * try_to_wake_up - wake up a thread
2365 * @p: the to-be-woken-up thread
2366 * @state: the mask of task states that can be woken
2367 * @sync: do a synchronous wakeup?
2369 * Put it on the run-queue if it's not already there. The "current"
2370 * thread is always on the run-queue (except when the actual
2371 * re-schedule is in progress), and as such you're allowed to do
2372 * the simpler "current->state = TASK_RUNNING" to mark yourself
2373 * runnable without the overhead of this.
2375 * returns failure only if the task is already active.
2377 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2380 int cpu, orig_cpu, this_cpu, success = 0;
2381 unsigned long flags;
2382 unsigned long en_flags = ENQUEUE_WAKEUP;
2385 this_cpu = get_cpu();
2388 rq = task_rq_lock(p, &flags);
2389 if (!(p->state & state))
2399 if (unlikely(task_running(rq, p)))
2403 * In order to handle concurrent wakeups and release the rq->lock
2404 * we put the task in TASK_WAKING state.
2406 * First fix up the nr_uninterruptible count:
2408 if (task_contributes_to_load(p)) {
2409 if (likely(cpu_online(orig_cpu)))
2410 rq->nr_uninterruptible--;
2412 this_rq()->nr_uninterruptible--;
2414 p->state = TASK_WAKING;
2416 if (p->sched_class->task_waking) {
2417 p->sched_class->task_waking(rq, p);
2418 en_flags |= ENQUEUE_WAKING;
2421 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2422 if (cpu != orig_cpu)
2423 set_task_cpu(p, cpu);
2424 __task_rq_unlock(rq);
2427 raw_spin_lock(&rq->lock);
2430 * We migrated the task without holding either rq->lock, however
2431 * since the task is not on the task list itself, nobody else
2432 * will try and migrate the task, hence the rq should match the
2433 * cpu we just moved it to.
2435 WARN_ON(task_cpu(p) != cpu);
2436 WARN_ON(p->state != TASK_WAKING);
2438 #ifdef CONFIG_SCHEDSTATS
2439 schedstat_inc(rq, ttwu_count);
2440 if (cpu == this_cpu)
2441 schedstat_inc(rq, ttwu_local);
2443 struct sched_domain *sd;
2444 for_each_domain(this_cpu, sd) {
2445 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2446 schedstat_inc(sd, ttwu_wake_remote);
2451 #endif /* CONFIG_SCHEDSTATS */
2454 #endif /* CONFIG_SMP */
2455 schedstat_inc(p, se.statistics.nr_wakeups);
2456 if (wake_flags & WF_SYNC)
2457 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2458 if (orig_cpu != cpu)
2459 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2460 if (cpu == this_cpu)
2461 schedstat_inc(p, se.statistics.nr_wakeups_local);
2463 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2464 activate_task(rq, p, en_flags);
2468 trace_sched_wakeup(p, success);
2469 check_preempt_curr(rq, p, wake_flags);
2471 p->state = TASK_RUNNING;
2473 if (p->sched_class->task_woken)
2474 p->sched_class->task_woken(rq, p);
2476 if (unlikely(rq->idle_stamp)) {
2477 u64 delta = rq->clock - rq->idle_stamp;
2478 u64 max = 2*sysctl_sched_migration_cost;
2483 update_avg(&rq->avg_idle, delta);
2488 task_rq_unlock(rq, &flags);
2495 * wake_up_process - Wake up a specific process
2496 * @p: The process to be woken up.
2498 * Attempt to wake up the nominated process and move it to the set of runnable
2499 * processes. Returns 1 if the process was woken up, 0 if it was already
2502 * It may be assumed that this function implies a write memory barrier before
2503 * changing the task state if and only if any tasks are woken up.
2505 int wake_up_process(struct task_struct *p)
2507 return try_to_wake_up(p, TASK_ALL, 0);
2509 EXPORT_SYMBOL(wake_up_process);
2511 int wake_up_state(struct task_struct *p, unsigned int state)
2513 return try_to_wake_up(p, state, 0);
2517 * Perform scheduler related setup for a newly forked process p.
2518 * p is forked by current.
2520 * __sched_fork() is basic setup used by init_idle() too:
2522 static void __sched_fork(struct task_struct *p)
2524 p->se.exec_start = 0;
2525 p->se.sum_exec_runtime = 0;
2526 p->se.prev_sum_exec_runtime = 0;
2527 p->se.nr_migrations = 0;
2529 #ifdef CONFIG_SCHEDSTATS
2530 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2533 INIT_LIST_HEAD(&p->rt.run_list);
2535 INIT_LIST_HEAD(&p->se.group_node);
2537 #ifdef CONFIG_PREEMPT_NOTIFIERS
2538 INIT_HLIST_HEAD(&p->preempt_notifiers);
2543 * fork()/clone()-time setup:
2545 void sched_fork(struct task_struct *p, int clone_flags)
2547 int cpu = get_cpu();
2551 * We mark the process as running here. This guarantees that
2552 * nobody will actually run it, and a signal or other external
2553 * event cannot wake it up and insert it on the runqueue either.
2555 p->state = TASK_RUNNING;
2558 * Revert to default priority/policy on fork if requested.
2560 if (unlikely(p->sched_reset_on_fork)) {
2561 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2562 p->policy = SCHED_NORMAL;
2563 p->normal_prio = p->static_prio;
2566 if (PRIO_TO_NICE(p->static_prio) < 0) {
2567 p->static_prio = NICE_TO_PRIO(0);
2568 p->normal_prio = p->static_prio;
2573 * We don't need the reset flag anymore after the fork. It has
2574 * fulfilled its duty:
2576 p->sched_reset_on_fork = 0;
2580 * Make sure we do not leak PI boosting priority to the child.
2582 p->prio = current->normal_prio;
2584 if (!rt_prio(p->prio))
2585 p->sched_class = &fair_sched_class;
2587 if (p->sched_class->task_fork)
2588 p->sched_class->task_fork(p);
2591 * The child is not yet in the pid-hash so no cgroup attach races,
2592 * and the cgroup is pinned to this child due to cgroup_fork()
2593 * is ran before sched_fork().
2595 * Silence PROVE_RCU.
2598 set_task_cpu(p, cpu);
2601 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2602 if (likely(sched_info_on()))
2603 memset(&p->sched_info, 0, sizeof(p->sched_info));
2605 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2608 #ifdef CONFIG_PREEMPT
2609 /* Want to start with kernel preemption disabled. */
2610 task_thread_info(p)->preempt_count = 1;
2612 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2618 * wake_up_new_task - wake up a newly created task for the first time.
2620 * This function will do some initial scheduler statistics housekeeping
2621 * that must be done for every newly created context, then puts the task
2622 * on the runqueue and wakes it.
2624 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2626 unsigned long flags;
2628 int cpu __maybe_unused = get_cpu();
2631 rq = task_rq_lock(p, &flags);
2632 p->state = TASK_WAKING;
2635 * Fork balancing, do it here and not earlier because:
2636 * - cpus_allowed can change in the fork path
2637 * - any previously selected cpu might disappear through hotplug
2639 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2640 * without people poking at ->cpus_allowed.
2642 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2643 set_task_cpu(p, cpu);
2645 p->state = TASK_RUNNING;
2646 task_rq_unlock(rq, &flags);
2649 rq = task_rq_lock(p, &flags);
2650 activate_task(rq, p, 0);
2651 trace_sched_wakeup_new(p, 1);
2652 check_preempt_curr(rq, p, WF_FORK);
2654 if (p->sched_class->task_woken)
2655 p->sched_class->task_woken(rq, p);
2657 task_rq_unlock(rq, &flags);
2661 #ifdef CONFIG_PREEMPT_NOTIFIERS
2664 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2665 * @notifier: notifier struct to register
2667 void preempt_notifier_register(struct preempt_notifier *notifier)
2669 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2671 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2674 * preempt_notifier_unregister - no longer interested in preemption notifications
2675 * @notifier: notifier struct to unregister
2677 * This is safe to call from within a preemption notifier.
2679 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2681 hlist_del(¬ifier->link);
2683 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2685 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2687 struct preempt_notifier *notifier;
2688 struct hlist_node *node;
2690 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2691 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2695 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2696 struct task_struct *next)
2698 struct preempt_notifier *notifier;
2699 struct hlist_node *node;
2701 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2702 notifier->ops->sched_out(notifier, next);
2705 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2707 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2712 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2713 struct task_struct *next)
2717 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2720 * prepare_task_switch - prepare to switch tasks
2721 * @rq: the runqueue preparing to switch
2722 * @prev: the current task that is being switched out
2723 * @next: the task we are going to switch to.
2725 * This is called with the rq lock held and interrupts off. It must
2726 * be paired with a subsequent finish_task_switch after the context
2729 * prepare_task_switch sets up locking and calls architecture specific
2733 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2734 struct task_struct *next)
2736 fire_sched_out_preempt_notifiers(prev, next);
2737 prepare_lock_switch(rq, next);
2738 prepare_arch_switch(next);
2742 * finish_task_switch - clean up after a task-switch
2743 * @rq: runqueue associated with task-switch
2744 * @prev: the thread we just switched away from.
2746 * finish_task_switch must be called after the context switch, paired
2747 * with a prepare_task_switch call before the context switch.
2748 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2749 * and do any other architecture-specific cleanup actions.
2751 * Note that we may have delayed dropping an mm in context_switch(). If
2752 * so, we finish that here outside of the runqueue lock. (Doing it
2753 * with the lock held can cause deadlocks; see schedule() for
2756 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2757 __releases(rq->lock)
2759 struct mm_struct *mm = rq->prev_mm;
2765 * A task struct has one reference for the use as "current".
2766 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2767 * schedule one last time. The schedule call will never return, and
2768 * the scheduled task must drop that reference.
2769 * The test for TASK_DEAD must occur while the runqueue locks are
2770 * still held, otherwise prev could be scheduled on another cpu, die
2771 * there before we look at prev->state, and then the reference would
2773 * Manfred Spraul <manfred@colorfullife.com>
2775 prev_state = prev->state;
2776 finish_arch_switch(prev);
2777 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2778 local_irq_disable();
2779 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2780 perf_event_task_sched_in(current);
2781 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2783 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2784 finish_lock_switch(rq, prev);
2786 fire_sched_in_preempt_notifiers(current);
2789 if (unlikely(prev_state == TASK_DEAD)) {
2791 * Remove function-return probe instances associated with this
2792 * task and put them back on the free list.
2794 kprobe_flush_task(prev);
2795 put_task_struct(prev);
2801 /* assumes rq->lock is held */
2802 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2804 if (prev->sched_class->pre_schedule)
2805 prev->sched_class->pre_schedule(rq, prev);
2808 /* rq->lock is NOT held, but preemption is disabled */
2809 static inline void post_schedule(struct rq *rq)
2811 if (rq->post_schedule) {
2812 unsigned long flags;
2814 raw_spin_lock_irqsave(&rq->lock, flags);
2815 if (rq->curr->sched_class->post_schedule)
2816 rq->curr->sched_class->post_schedule(rq);
2817 raw_spin_unlock_irqrestore(&rq->lock, flags);
2819 rq->post_schedule = 0;
2825 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2829 static inline void post_schedule(struct rq *rq)
2836 * schedule_tail - first thing a freshly forked thread must call.
2837 * @prev: the thread we just switched away from.
2839 asmlinkage void schedule_tail(struct task_struct *prev)
2840 __releases(rq->lock)
2842 struct rq *rq = this_rq();
2844 finish_task_switch(rq, prev);
2847 * FIXME: do we need to worry about rq being invalidated by the
2852 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2853 /* In this case, finish_task_switch does not reenable preemption */
2856 if (current->set_child_tid)
2857 put_user(task_pid_vnr(current), current->set_child_tid);
2861 * context_switch - switch to the new MM and the new
2862 * thread's register state.
2865 context_switch(struct rq *rq, struct task_struct *prev,
2866 struct task_struct *next)
2868 struct mm_struct *mm, *oldmm;
2870 prepare_task_switch(rq, prev, next);
2871 trace_sched_switch(prev, next);
2873 oldmm = prev->active_mm;
2875 * For paravirt, this is coupled with an exit in switch_to to
2876 * combine the page table reload and the switch backend into
2879 arch_start_context_switch(prev);
2882 next->active_mm = oldmm;
2883 atomic_inc(&oldmm->mm_count);
2884 enter_lazy_tlb(oldmm, next);
2886 switch_mm(oldmm, mm, next);
2888 if (likely(!prev->mm)) {
2889 prev->active_mm = NULL;
2890 rq->prev_mm = oldmm;
2893 * Since the runqueue lock will be released by the next
2894 * task (which is an invalid locking op but in the case
2895 * of the scheduler it's an obvious special-case), so we
2896 * do an early lockdep release here:
2898 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2899 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2902 /* Here we just switch the register state and the stack. */
2903 switch_to(prev, next, prev);
2907 * this_rq must be evaluated again because prev may have moved
2908 * CPUs since it called schedule(), thus the 'rq' on its stack
2909 * frame will be invalid.
2911 finish_task_switch(this_rq(), prev);
2915 * nr_running, nr_uninterruptible and nr_context_switches:
2917 * externally visible scheduler statistics: current number of runnable
2918 * threads, current number of uninterruptible-sleeping threads, total
2919 * number of context switches performed since bootup.
2921 unsigned long nr_running(void)
2923 unsigned long i, sum = 0;
2925 for_each_online_cpu(i)
2926 sum += cpu_rq(i)->nr_running;
2931 unsigned long nr_uninterruptible(void)
2933 unsigned long i, sum = 0;
2935 for_each_possible_cpu(i)
2936 sum += cpu_rq(i)->nr_uninterruptible;
2939 * Since we read the counters lockless, it might be slightly
2940 * inaccurate. Do not allow it to go below zero though:
2942 if (unlikely((long)sum < 0))
2948 unsigned long long nr_context_switches(void)
2951 unsigned long long sum = 0;
2953 for_each_possible_cpu(i)
2954 sum += cpu_rq(i)->nr_switches;
2959 unsigned long nr_iowait(void)
2961 unsigned long i, sum = 0;
2963 for_each_possible_cpu(i)
2964 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2969 unsigned long nr_iowait_cpu(int cpu)
2971 struct rq *this = cpu_rq(cpu);
2972 return atomic_read(&this->nr_iowait);
2975 unsigned long this_cpu_load(void)
2977 struct rq *this = this_rq();
2978 return this->cpu_load[0];
2982 /* Variables and functions for calc_load */
2983 static atomic_long_t calc_load_tasks;
2984 static unsigned long calc_load_update;
2985 unsigned long avenrun[3];
2986 EXPORT_SYMBOL(avenrun);
2988 static long calc_load_fold_active(struct rq *this_rq)
2990 long nr_active, delta = 0;
2992 nr_active = this_rq->nr_running;
2993 nr_active += (long) this_rq->nr_uninterruptible;
2995 if (nr_active != this_rq->calc_load_active) {
2996 delta = nr_active - this_rq->calc_load_active;
2997 this_rq->calc_load_active = nr_active;
3003 static unsigned long
3004 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3007 load += active * (FIXED_1 - exp);
3008 load += 1UL << (FSHIFT - 1);
3009 return load >> FSHIFT;
3014 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3016 * When making the ILB scale, we should try to pull this in as well.
3018 static atomic_long_t calc_load_tasks_idle;
3020 static void calc_load_account_idle(struct rq *this_rq)
3024 delta = calc_load_fold_active(this_rq);
3026 atomic_long_add(delta, &calc_load_tasks_idle);
3029 static long calc_load_fold_idle(void)
3034 * Its got a race, we don't care...
3036 if (atomic_long_read(&calc_load_tasks_idle))
3037 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3043 * fixed_power_int - compute: x^n, in O(log n) time
3045 * @x: base of the power
3046 * @frac_bits: fractional bits of @x
3047 * @n: power to raise @x to.
3049 * By exploiting the relation between the definition of the natural power
3050 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3051 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3052 * (where: n_i \elem {0, 1}, the binary vector representing n),
3053 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3054 * of course trivially computable in O(log_2 n), the length of our binary
3057 static unsigned long
3058 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3060 unsigned long result = 1UL << frac_bits;
3065 result += 1UL << (frac_bits - 1);
3066 result >>= frac_bits;
3072 x += 1UL << (frac_bits - 1);
3080 * a1 = a0 * e + a * (1 - e)
3082 * a2 = a1 * e + a * (1 - e)
3083 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3084 * = a0 * e^2 + a * (1 - e) * (1 + e)
3086 * a3 = a2 * e + a * (1 - e)
3087 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3088 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3092 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3093 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3094 * = a0 * e^n + a * (1 - e^n)
3096 * [1] application of the geometric series:
3099 * S_n := \Sum x^i = -------------
3102 static unsigned long
3103 calc_load_n(unsigned long load, unsigned long exp,
3104 unsigned long active, unsigned int n)
3107 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3111 * NO_HZ can leave us missing all per-cpu ticks calling
3112 * calc_load_account_active(), but since an idle CPU folds its delta into
3113 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3114 * in the pending idle delta if our idle period crossed a load cycle boundary.
3116 * Once we've updated the global active value, we need to apply the exponential
3117 * weights adjusted to the number of cycles missed.
3119 static void calc_global_nohz(unsigned long ticks)
3121 long delta, active, n;
3123 if (time_before(jiffies, calc_load_update))
3127 * If we crossed a calc_load_update boundary, make sure to fold
3128 * any pending idle changes, the respective CPUs might have
3129 * missed the tick driven calc_load_account_active() update
3132 delta = calc_load_fold_idle();
3134 atomic_long_add(delta, &calc_load_tasks);
3137 * If we were idle for multiple load cycles, apply them.
3139 if (ticks >= LOAD_FREQ) {
3140 n = ticks / LOAD_FREQ;
3142 active = atomic_long_read(&calc_load_tasks);
3143 active = active > 0 ? active * FIXED_1 : 0;
3145 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3146 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3147 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3149 calc_load_update += n * LOAD_FREQ;
3153 * Its possible the remainder of the above division also crosses
3154 * a LOAD_FREQ period, the regular check in calc_global_load()
3155 * which comes after this will take care of that.
3157 * Consider us being 11 ticks before a cycle completion, and us
3158 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3159 * age us 4 cycles, and the test in calc_global_load() will
3160 * pick up the final one.
3164 static void calc_load_account_idle(struct rq *this_rq)
3168 static inline long calc_load_fold_idle(void)
3173 static void calc_global_nohz(unsigned long ticks)
3179 * get_avenrun - get the load average array
3180 * @loads: pointer to dest load array
3181 * @offset: offset to add
3182 * @shift: shift count to shift the result left
3184 * These values are estimates at best, so no need for locking.
3186 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3188 loads[0] = (avenrun[0] + offset) << shift;
3189 loads[1] = (avenrun[1] + offset) << shift;
3190 loads[2] = (avenrun[2] + offset) << shift;
3194 * calc_load - update the avenrun load estimates 10 ticks after the
3195 * CPUs have updated calc_load_tasks.
3197 void calc_global_load(unsigned long ticks)
3201 calc_global_nohz(ticks);
3203 if (time_before(jiffies, calc_load_update + 10))
3206 active = atomic_long_read(&calc_load_tasks);
3207 active = active > 0 ? active * FIXED_1 : 0;
3209 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3210 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3211 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3213 calc_load_update += LOAD_FREQ;
3217 * Called from update_cpu_load() to periodically update this CPU's
3220 static void calc_load_account_active(struct rq *this_rq)
3224 if (time_before(jiffies, this_rq->calc_load_update))
3227 delta = calc_load_fold_active(this_rq);
3228 delta += calc_load_fold_idle();
3230 atomic_long_add(delta, &calc_load_tasks);
3232 this_rq->calc_load_update += LOAD_FREQ;
3236 * Update rq->cpu_load[] statistics. This function is usually called every
3237 * scheduler tick (TICK_NSEC).
3239 static void update_cpu_load(struct rq *this_rq)
3241 unsigned long this_load = this_rq->load.weight;
3244 this_rq->nr_load_updates++;
3246 /* Update our load: */
3247 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3248 unsigned long old_load, new_load;
3250 /* scale is effectively 1 << i now, and >> i divides by scale */
3252 old_load = this_rq->cpu_load[i];
3253 new_load = this_load;
3255 * Round up the averaging division if load is increasing. This
3256 * prevents us from getting stuck on 9 if the load is 10, for
3259 if (new_load > old_load)
3260 new_load += scale-1;
3261 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3264 calc_load_account_active(this_rq);
3266 sched_avg_update(this_rq);
3272 * sched_exec - execve() is a valuable balancing opportunity, because at
3273 * this point the task has the smallest effective memory and cache footprint.
3275 void sched_exec(void)
3277 struct task_struct *p = current;
3278 unsigned long flags;
3282 rq = task_rq_lock(p, &flags);
3283 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3284 if (dest_cpu == smp_processor_id())
3288 * select_task_rq() can race against ->cpus_allowed
3290 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3291 likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
3292 struct migration_arg arg = { p, dest_cpu };
3294 task_rq_unlock(rq, &flags);
3295 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3299 task_rq_unlock(rq, &flags);
3304 DEFINE_PER_CPU(struct kernel_stat, kstat);
3306 EXPORT_PER_CPU_SYMBOL(kstat);
3309 * Return any ns on the sched_clock that have not yet been accounted in
3310 * @p in case that task is currently running.
3312 * Called with task_rq_lock() held on @rq.
3314 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3318 if (task_current(rq, p)) {
3319 update_rq_clock(rq);
3320 ns = rq->clock_task - p->se.exec_start;
3328 unsigned long long task_delta_exec(struct task_struct *p)
3330 unsigned long flags;
3334 rq = task_rq_lock(p, &flags);
3335 ns = do_task_delta_exec(p, rq);
3336 task_rq_unlock(rq, &flags);
3342 * Return accounted runtime for the task.
3343 * In case the task is currently running, return the runtime plus current's
3344 * pending runtime that have not been accounted yet.
3346 unsigned long long task_sched_runtime(struct task_struct *p)
3348 unsigned long flags;
3352 rq = task_rq_lock(p, &flags);
3353 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3354 task_rq_unlock(rq, &flags);
3360 * Return sum_exec_runtime for the thread group.
3361 * In case the task is currently running, return the sum plus current's
3362 * pending runtime that have not been accounted yet.
3364 * Note that the thread group might have other running tasks as well,
3365 * so the return value not includes other pending runtime that other
3366 * running tasks might have.
3368 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3370 struct task_cputime totals;
3371 unsigned long flags;
3375 rq = task_rq_lock(p, &flags);
3376 thread_group_cputime(p, &totals);
3377 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3378 task_rq_unlock(rq, &flags);
3384 * Account user cpu time to a process.
3385 * @p: the process that the cpu time gets accounted to
3386 * @cputime: the cpu time spent in user space since the last update
3387 * @cputime_scaled: cputime scaled by cpu frequency
3389 void account_user_time(struct task_struct *p, cputime_t cputime,
3390 cputime_t cputime_scaled)
3392 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3395 /* Add user time to process. */
3396 p->utime = cputime_add(p->utime, cputime);
3397 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3398 account_group_user_time(p, cputime);
3400 /* Add user time to cpustat. */
3401 tmp = cputime_to_cputime64(cputime);
3402 if (TASK_NICE(p) > 0)
3403 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3405 cpustat->user = cputime64_add(cpustat->user, tmp);
3407 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3408 /* Account for user time used */
3409 acct_update_integrals(p);
3413 * Account guest cpu time to a process.
3414 * @p: the process that the cpu time gets accounted to
3415 * @cputime: the cpu time spent in virtual machine since the last update
3416 * @cputime_scaled: cputime scaled by cpu frequency
3418 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3419 cputime_t cputime_scaled)
3422 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3424 tmp = cputime_to_cputime64(cputime);
3426 /* Add guest time to process. */
3427 p->utime = cputime_add(p->utime, cputime);
3428 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3429 account_group_user_time(p, cputime);
3430 p->gtime = cputime_add(p->gtime, cputime);
3432 /* Add guest time to cpustat. */
3433 if (TASK_NICE(p) > 0) {
3434 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3435 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3437 cpustat->user = cputime64_add(cpustat->user, tmp);
3438 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3443 * Account system cpu time to a process.
3444 * @p: the process that the cpu time gets accounted to
3445 * @hardirq_offset: the offset to subtract from hardirq_count()
3446 * @cputime: the cpu time spent in kernel space since the last update
3447 * @cputime_scaled: cputime scaled by cpu frequency
3449 void account_system_time(struct task_struct *p, int hardirq_offset,
3450 cputime_t cputime, cputime_t cputime_scaled)
3452 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3455 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3456 account_guest_time(p, cputime, cputime_scaled);
3460 /* Add system time to process. */
3461 p->stime = cputime_add(p->stime, cputime);
3462 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3463 account_group_system_time(p, cputime);
3465 /* Add system time to cpustat. */
3466 tmp = cputime_to_cputime64(cputime);
3467 if (hardirq_count() - hardirq_offset)
3468 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3469 else if (in_serving_softirq())
3470 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3472 cpustat->system = cputime64_add(cpustat->system, tmp);
3474 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3476 /* Account for system time used */
3477 acct_update_integrals(p);
3481 * Account for involuntary wait time.
3482 * @steal: the cpu time spent in involuntary wait
3484 void account_steal_time(cputime_t cputime)
3486 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3487 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3489 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3493 * Account for idle time.
3494 * @cputime: the cpu time spent in idle wait
3496 void account_idle_time(cputime_t cputime)
3498 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3499 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3500 struct rq *rq = this_rq();
3502 if (atomic_read(&rq->nr_iowait) > 0)
3503 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3505 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3508 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3511 * Account a single tick of cpu time.
3512 * @p: the process that the cpu time gets accounted to
3513 * @user_tick: indicates if the tick is a user or a system tick
3515 void account_process_tick(struct task_struct *p, int user_tick)
3517 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3518 struct rq *rq = this_rq();
3521 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3522 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3523 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3526 account_idle_time(cputime_one_jiffy);
3530 * Account multiple ticks of steal time.
3531 * @p: the process from which the cpu time has been stolen
3532 * @ticks: number of stolen ticks
3534 void account_steal_ticks(unsigned long ticks)
3536 account_steal_time(jiffies_to_cputime(ticks));
3540 * Account multiple ticks of idle time.
3541 * @ticks: number of stolen ticks
3543 void account_idle_ticks(unsigned long ticks)
3545 account_idle_time(jiffies_to_cputime(ticks));
3551 * Use precise platform statistics if available:
3553 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3554 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3560 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3562 struct task_cputime cputime;
3564 thread_group_cputime(p, &cputime);
3566 *ut = cputime.utime;
3567 *st = cputime.stime;
3571 #ifndef nsecs_to_cputime
3572 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3575 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3577 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3580 * Use CFS's precise accounting:
3582 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3588 do_div(temp, total);
3589 utime = (cputime_t)temp;
3594 * Compare with previous values, to keep monotonicity:
3596 p->prev_utime = max(p->prev_utime, utime);
3597 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3599 *ut = p->prev_utime;
3600 *st = p->prev_stime;
3604 * Must be called with siglock held.
3606 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3608 struct signal_struct *sig = p->signal;
3609 struct task_cputime cputime;
3610 cputime_t rtime, utime, total;
3612 thread_group_cputime(p, &cputime);
3614 total = cputime_add(cputime.utime, cputime.stime);
3615 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3620 temp *= cputime.utime;
3621 do_div(temp, total);
3622 utime = (cputime_t)temp;
3626 sig->prev_utime = max(sig->prev_utime, utime);
3627 sig->prev_stime = max(sig->prev_stime,
3628 cputime_sub(rtime, sig->prev_utime));
3630 *ut = sig->prev_utime;
3631 *st = sig->prev_stime;
3636 * This function gets called by the timer code, with HZ frequency.
3637 * We call it with interrupts disabled.
3639 * It also gets called by the fork code, when changing the parent's
3642 void scheduler_tick(void)
3644 int cpu = smp_processor_id();
3645 struct rq *rq = cpu_rq(cpu);
3646 struct task_struct *curr = rq->curr;
3650 raw_spin_lock(&rq->lock);
3651 update_rq_clock(rq);
3652 update_cpu_load(rq);
3653 curr->sched_class->task_tick(rq, curr, 0);
3654 raw_spin_unlock(&rq->lock);
3656 perf_event_task_tick(curr);
3659 rq->idle_at_tick = idle_cpu(cpu);
3660 trigger_load_balance(rq, cpu);
3664 notrace unsigned long get_parent_ip(unsigned long addr)
3666 if (in_lock_functions(addr)) {
3667 addr = CALLER_ADDR2;
3668 if (in_lock_functions(addr))
3669 addr = CALLER_ADDR3;
3674 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3675 defined(CONFIG_PREEMPT_TRACER))
3677 void __kprobes add_preempt_count(int val)
3679 #ifdef CONFIG_DEBUG_PREEMPT
3683 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3686 preempt_count() += val;
3687 #ifdef CONFIG_DEBUG_PREEMPT
3689 * Spinlock count overflowing soon?
3691 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3694 if (preempt_count() == val)
3695 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3697 EXPORT_SYMBOL(add_preempt_count);
3699 void __kprobes sub_preempt_count(int val)
3701 #ifdef CONFIG_DEBUG_PREEMPT
3705 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3708 * Is the spinlock portion underflowing?
3710 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3711 !(preempt_count() & PREEMPT_MASK)))
3715 if (preempt_count() == val)
3716 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3717 preempt_count() -= val;
3719 EXPORT_SYMBOL(sub_preempt_count);
3724 * Print scheduling while atomic bug:
3726 static noinline void __schedule_bug(struct task_struct *prev)
3728 struct pt_regs *regs = get_irq_regs();
3730 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3731 prev->comm, prev->pid, preempt_count());
3733 debug_show_held_locks(prev);
3735 if (irqs_disabled())
3736 print_irqtrace_events(prev);
3745 * Various schedule()-time debugging checks and statistics:
3747 static inline void schedule_debug(struct task_struct *prev)
3750 * Test if we are atomic. Since do_exit() needs to call into
3751 * schedule() atomically, we ignore that path for now.
3752 * Otherwise, whine if we are scheduling when we should not be.
3754 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3755 __schedule_bug(prev);
3757 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3759 schedstat_inc(this_rq(), sched_count);
3760 #ifdef CONFIG_SCHEDSTATS
3761 if (unlikely(prev->lock_depth >= 0)) {
3762 schedstat_inc(this_rq(), bkl_count);
3763 schedstat_inc(prev, sched_info.bkl_count);
3768 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3771 update_rq_clock(rq);
3772 prev->sched_class->put_prev_task(rq, prev);
3776 * Pick up the highest-prio task:
3778 static inline struct task_struct *
3779 pick_next_task(struct rq *rq)
3781 const struct sched_class *class;
3782 struct task_struct *p;
3785 * Optimization: we know that if all tasks are in
3786 * the fair class we can call that function directly:
3788 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3789 p = fair_sched_class.pick_next_task(rq);
3794 class = sched_class_highest;
3796 p = class->pick_next_task(rq);
3800 * Will never be NULL as the idle class always
3801 * returns a non-NULL p:
3803 class = class->next;
3808 * schedule() is the main scheduler function.
3810 asmlinkage void __sched schedule(void)
3812 struct task_struct *prev, *next;
3813 unsigned long *switch_count;
3819 cpu = smp_processor_id();
3821 rcu_note_context_switch(cpu);
3823 switch_count = &prev->nivcsw;
3825 release_kernel_lock(prev);
3826 need_resched_nonpreemptible:
3828 schedule_debug(prev);
3830 if (sched_feat(HRTICK))
3833 raw_spin_lock_irq(&rq->lock);
3835 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3836 if (unlikely(signal_pending_state(prev->state, prev)))
3837 prev->state = TASK_RUNNING;
3839 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3840 switch_count = &prev->nvcsw;
3843 pre_schedule(rq, prev);
3845 if (unlikely(!rq->nr_running))
3846 idle_balance(cpu, rq);
3848 put_prev_task(rq, prev);
3849 next = pick_next_task(rq);
3850 clear_tsk_need_resched(prev);
3851 rq->skip_clock_update = 0;
3853 if (likely(prev != next)) {
3854 sched_info_switch(prev, next);
3855 perf_event_task_sched_out(prev, next);
3861 context_switch(rq, prev, next); /* unlocks the rq */
3863 * the context switch might have flipped the stack from under
3864 * us, hence refresh the local variables.
3866 cpu = smp_processor_id();
3869 raw_spin_unlock_irq(&rq->lock);
3873 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3875 switch_count = &prev->nivcsw;
3876 goto need_resched_nonpreemptible;
3879 preempt_enable_no_resched();
3883 EXPORT_SYMBOL(schedule);
3885 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3887 * Look out! "owner" is an entirely speculative pointer
3888 * access and not reliable.
3890 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3895 if (!sched_feat(OWNER_SPIN))
3898 #ifdef CONFIG_DEBUG_PAGEALLOC
3900 * Need to access the cpu field knowing that
3901 * DEBUG_PAGEALLOC could have unmapped it if
3902 * the mutex owner just released it and exited.
3904 if (probe_kernel_address(&owner->cpu, cpu))
3911 * Even if the access succeeded (likely case),
3912 * the cpu field may no longer be valid.
3914 if (cpu >= nr_cpumask_bits)
3918 * We need to validate that we can do a
3919 * get_cpu() and that we have the percpu area.
3921 if (!cpu_online(cpu))
3928 * Owner changed, break to re-assess state.
3930 if (lock->owner != owner) {
3932 * If the lock has switched to a different owner,
3933 * we likely have heavy contention. Return 0 to quit
3934 * optimistic spinning and not contend further:
3942 * Is that owner really running on that cpu?
3944 if (task_thread_info(rq->curr) != owner || need_resched())
3954 #ifdef CONFIG_PREEMPT
3956 * this is the entry point to schedule() from in-kernel preemption
3957 * off of preempt_enable. Kernel preemptions off return from interrupt
3958 * occur there and call schedule directly.
3960 asmlinkage void __sched preempt_schedule(void)
3962 struct thread_info *ti = current_thread_info();
3965 * If there is a non-zero preempt_count or interrupts are disabled,
3966 * we do not want to preempt the current task. Just return..
3968 if (likely(ti->preempt_count || irqs_disabled()))
3972 add_preempt_count(PREEMPT_ACTIVE);
3974 sub_preempt_count(PREEMPT_ACTIVE);
3977 * Check again in case we missed a preemption opportunity
3978 * between schedule and now.
3981 } while (need_resched());
3983 EXPORT_SYMBOL(preempt_schedule);
3986 * this is the entry point to schedule() from kernel preemption
3987 * off of irq context.
3988 * Note, that this is called and return with irqs disabled. This will
3989 * protect us against recursive calling from irq.
3991 asmlinkage void __sched preempt_schedule_irq(void)
3993 struct thread_info *ti = current_thread_info();
3995 /* Catch callers which need to be fixed */
3996 BUG_ON(ti->preempt_count || !irqs_disabled());
3999 add_preempt_count(PREEMPT_ACTIVE);
4002 local_irq_disable();
4003 sub_preempt_count(PREEMPT_ACTIVE);
4006 * Check again in case we missed a preemption opportunity
4007 * between schedule and now.
4010 } while (need_resched());
4013 #endif /* CONFIG_PREEMPT */
4015 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4018 return try_to_wake_up(curr->private, mode, wake_flags);
4020 EXPORT_SYMBOL(default_wake_function);
4023 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4024 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4025 * number) then we wake all the non-exclusive tasks and one exclusive task.
4027 * There are circumstances in which we can try to wake a task which has already
4028 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4029 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4031 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4032 int nr_exclusive, int wake_flags, void *key)
4034 wait_queue_t *curr, *next;
4036 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4037 unsigned flags = curr->flags;
4039 if (curr->func(curr, mode, wake_flags, key) &&
4040 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4046 * __wake_up - wake up threads blocked on a waitqueue.
4048 * @mode: which threads
4049 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4050 * @key: is directly passed to the wakeup function
4052 * It may be assumed that this function implies a write memory barrier before
4053 * changing the task state if and only if any tasks are woken up.
4055 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4056 int nr_exclusive, void *key)
4058 unsigned long flags;
4060 spin_lock_irqsave(&q->lock, flags);
4061 __wake_up_common(q, mode, nr_exclusive, 0, key);
4062 spin_unlock_irqrestore(&q->lock, flags);
4064 EXPORT_SYMBOL(__wake_up);
4067 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4069 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4071 __wake_up_common(q, mode, 1, 0, NULL);
4073 EXPORT_SYMBOL_GPL(__wake_up_locked);
4075 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4077 __wake_up_common(q, mode, 1, 0, key);
4081 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4083 * @mode: which threads
4084 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4085 * @key: opaque value to be passed to wakeup targets
4087 * The sync wakeup differs that the waker knows that it will schedule
4088 * away soon, so while the target thread will be woken up, it will not
4089 * be migrated to another CPU - ie. the two threads are 'synchronized'
4090 * with each other. This can prevent needless bouncing between CPUs.
4092 * On UP it can prevent extra preemption.
4094 * It may be assumed that this function implies a write memory barrier before
4095 * changing the task state if and only if any tasks are woken up.
4097 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4098 int nr_exclusive, void *key)
4100 unsigned long flags;
4101 int wake_flags = WF_SYNC;
4106 if (unlikely(!nr_exclusive))
4109 spin_lock_irqsave(&q->lock, flags);
4110 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4111 spin_unlock_irqrestore(&q->lock, flags);
4113 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4116 * __wake_up_sync - see __wake_up_sync_key()
4118 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4120 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4122 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4125 * complete: - signals a single thread waiting on this completion
4126 * @x: holds the state of this particular completion
4128 * This will wake up a single thread waiting on this completion. Threads will be
4129 * awakened in the same order in which they were queued.
4131 * See also complete_all(), wait_for_completion() and related routines.
4133 * It may be assumed that this function implies a write memory barrier before
4134 * changing the task state if and only if any tasks are woken up.
4136 void complete(struct completion *x)
4138 unsigned long flags;
4140 spin_lock_irqsave(&x->wait.lock, flags);
4142 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4143 spin_unlock_irqrestore(&x->wait.lock, flags);
4145 EXPORT_SYMBOL(complete);
4148 * complete_all: - signals all threads waiting on this completion
4149 * @x: holds the state of this particular completion
4151 * This will wake up all threads waiting on this particular completion event.
4153 * It may be assumed that this function implies a write memory barrier before
4154 * changing the task state if and only if any tasks are woken up.
4156 void complete_all(struct completion *x)
4158 unsigned long flags;
4160 spin_lock_irqsave(&x->wait.lock, flags);
4161 x->done += UINT_MAX/2;
4162 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4163 spin_unlock_irqrestore(&x->wait.lock, flags);
4165 EXPORT_SYMBOL(complete_all);
4167 static inline long __sched
4168 do_wait_for_common(struct completion *x, long timeout, int state)
4171 DECLARE_WAITQUEUE(wait, current);
4173 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4175 if (signal_pending_state(state, current)) {
4176 timeout = -ERESTARTSYS;
4179 __set_current_state(state);
4180 spin_unlock_irq(&x->wait.lock);
4181 timeout = schedule_timeout(timeout);
4182 spin_lock_irq(&x->wait.lock);
4183 } while (!x->done && timeout);
4184 __remove_wait_queue(&x->wait, &wait);
4189 return timeout ?: 1;
4193 wait_for_common(struct completion *x, long timeout, int state)
4197 spin_lock_irq(&x->wait.lock);
4198 timeout = do_wait_for_common(x, timeout, state);
4199 spin_unlock_irq(&x->wait.lock);
4204 * wait_for_completion: - waits for completion of a task
4205 * @x: holds the state of this particular completion
4207 * This waits to be signaled for completion of a specific task. It is NOT
4208 * interruptible and there is no timeout.
4210 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4211 * and interrupt capability. Also see complete().
4213 void __sched wait_for_completion(struct completion *x)
4215 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4217 EXPORT_SYMBOL(wait_for_completion);
4220 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4221 * @x: holds the state of this particular completion
4222 * @timeout: timeout value in jiffies
4224 * This waits for either a completion of a specific task to be signaled or for a
4225 * specified timeout to expire. The timeout is in jiffies. It is not
4228 unsigned long __sched
4229 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4231 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4233 EXPORT_SYMBOL(wait_for_completion_timeout);
4236 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4237 * @x: holds the state of this particular completion
4239 * This waits for completion of a specific task to be signaled. It is
4242 int __sched wait_for_completion_interruptible(struct completion *x)
4244 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4245 if (t == -ERESTARTSYS)
4249 EXPORT_SYMBOL(wait_for_completion_interruptible);
4252 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4253 * @x: holds the state of this particular completion
4254 * @timeout: timeout value in jiffies
4256 * This waits for either a completion of a specific task to be signaled or for a
4257 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4259 unsigned long __sched
4260 wait_for_completion_interruptible_timeout(struct completion *x,
4261 unsigned long timeout)
4263 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4265 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4268 * wait_for_completion_killable: - waits for completion of a task (killable)
4269 * @x: holds the state of this particular completion
4271 * This waits to be signaled for completion of a specific task. It can be
4272 * interrupted by a kill signal.
4274 int __sched wait_for_completion_killable(struct completion *x)
4276 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4277 if (t == -ERESTARTSYS)
4281 EXPORT_SYMBOL(wait_for_completion_killable);
4284 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4285 * @x: holds the state of this particular completion
4286 * @timeout: timeout value in jiffies
4288 * This waits for either a completion of a specific task to be
4289 * signaled or for a specified timeout to expire. It can be
4290 * interrupted by a kill signal. The timeout is in jiffies.
4292 unsigned long __sched
4293 wait_for_completion_killable_timeout(struct completion *x,
4294 unsigned long timeout)
4296 return wait_for_common(x, timeout, TASK_KILLABLE);
4298 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4301 * try_wait_for_completion - try to decrement a completion without blocking
4302 * @x: completion structure
4304 * Returns: 0 if a decrement cannot be done without blocking
4305 * 1 if a decrement succeeded.
4307 * If a completion is being used as a counting completion,
4308 * attempt to decrement the counter without blocking. This
4309 * enables us to avoid waiting if the resource the completion
4310 * is protecting is not available.
4312 bool try_wait_for_completion(struct completion *x)
4314 unsigned long flags;
4317 spin_lock_irqsave(&x->wait.lock, flags);
4322 spin_unlock_irqrestore(&x->wait.lock, flags);
4325 EXPORT_SYMBOL(try_wait_for_completion);
4328 * completion_done - Test to see if a completion has any waiters
4329 * @x: completion structure
4331 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4332 * 1 if there are no waiters.
4335 bool completion_done(struct completion *x)
4337 unsigned long flags;
4340 spin_lock_irqsave(&x->wait.lock, flags);
4343 spin_unlock_irqrestore(&x->wait.lock, flags);
4346 EXPORT_SYMBOL(completion_done);
4349 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4351 unsigned long flags;
4354 init_waitqueue_entry(&wait, current);
4356 __set_current_state(state);
4358 spin_lock_irqsave(&q->lock, flags);
4359 __add_wait_queue(q, &wait);
4360 spin_unlock(&q->lock);
4361 timeout = schedule_timeout(timeout);
4362 spin_lock_irq(&q->lock);
4363 __remove_wait_queue(q, &wait);
4364 spin_unlock_irqrestore(&q->lock, flags);
4369 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4371 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4373 EXPORT_SYMBOL(interruptible_sleep_on);
4376 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4378 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4380 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4382 void __sched sleep_on(wait_queue_head_t *q)
4384 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4386 EXPORT_SYMBOL(sleep_on);
4388 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4390 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4392 EXPORT_SYMBOL(sleep_on_timeout);
4394 #ifdef CONFIG_RT_MUTEXES
4397 * rt_mutex_setprio - set the current priority of a task
4399 * @prio: prio value (kernel-internal form)
4401 * This function changes the 'effective' priority of a task. It does
4402 * not touch ->normal_prio like __setscheduler().
4404 * Used by the rt_mutex code to implement priority inheritance logic.
4406 void rt_mutex_setprio(struct task_struct *p, int prio)
4408 unsigned long flags;
4409 int oldprio, on_rq, running;
4411 const struct sched_class *prev_class;
4413 BUG_ON(prio < 0 || prio > MAX_PRIO);
4415 rq = task_rq_lock(p, &flags);
4418 prev_class = p->sched_class;
4419 on_rq = p->se.on_rq;
4420 running = task_current(rq, p);
4422 dequeue_task(rq, p, 0);
4424 p->sched_class->put_prev_task(rq, p);
4427 p->sched_class = &rt_sched_class;
4429 p->sched_class = &fair_sched_class;
4434 p->sched_class->set_curr_task(rq);
4436 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4438 check_class_changed(rq, p, prev_class, oldprio, running);
4440 task_rq_unlock(rq, &flags);
4445 void set_user_nice(struct task_struct *p, long nice)
4447 int old_prio, delta, on_rq;
4448 unsigned long flags;
4451 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4454 * We have to be careful, if called from sys_setpriority(),
4455 * the task might be in the middle of scheduling on another CPU.
4457 rq = task_rq_lock(p, &flags);
4459 * The RT priorities are set via sched_setscheduler(), but we still
4460 * allow the 'normal' nice value to be set - but as expected
4461 * it wont have any effect on scheduling until the task is
4462 * SCHED_FIFO/SCHED_RR:
4464 if (task_has_rt_policy(p)) {
4465 p->static_prio = NICE_TO_PRIO(nice);
4468 on_rq = p->se.on_rq;
4470 dequeue_task(rq, p, 0);
4472 p->static_prio = NICE_TO_PRIO(nice);
4475 p->prio = effective_prio(p);
4476 delta = p->prio - old_prio;
4479 enqueue_task(rq, p, 0);
4481 * If the task increased its priority or is running and
4482 * lowered its priority, then reschedule its CPU:
4484 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4485 resched_task(rq->curr);
4488 task_rq_unlock(rq, &flags);
4490 EXPORT_SYMBOL(set_user_nice);
4493 * can_nice - check if a task can reduce its nice value
4497 int can_nice(const struct task_struct *p, const int nice)
4499 /* convert nice value [19,-20] to rlimit style value [1,40] */
4500 int nice_rlim = 20 - nice;
4502 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4503 capable(CAP_SYS_NICE));
4506 #ifdef __ARCH_WANT_SYS_NICE
4509 * sys_nice - change the priority of the current process.
4510 * @increment: priority increment
4512 * sys_setpriority is a more generic, but much slower function that
4513 * does similar things.
4515 SYSCALL_DEFINE1(nice, int, increment)
4520 * Setpriority might change our priority at the same moment.
4521 * We don't have to worry. Conceptually one call occurs first
4522 * and we have a single winner.
4524 if (increment < -40)
4529 nice = TASK_NICE(current) + increment;
4535 if (increment < 0 && !can_nice(current, nice))
4538 retval = security_task_setnice(current, nice);
4542 set_user_nice(current, nice);
4549 * task_prio - return the priority value of a given task.
4550 * @p: the task in question.
4552 * This is the priority value as seen by users in /proc.
4553 * RT tasks are offset by -200. Normal tasks are centered
4554 * around 0, value goes from -16 to +15.
4556 int task_prio(const struct task_struct *p)
4558 return p->prio - MAX_RT_PRIO;
4562 * task_nice - return the nice value of a given task.
4563 * @p: the task in question.
4565 int task_nice(const struct task_struct *p)
4567 return TASK_NICE(p);
4569 EXPORT_SYMBOL(task_nice);
4572 * idle_cpu - is a given cpu idle currently?
4573 * @cpu: the processor in question.
4575 int idle_cpu(int cpu)
4577 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4581 * idle_task - return the idle task for a given cpu.
4582 * @cpu: the processor in question.
4584 struct task_struct *idle_task(int cpu)
4586 return cpu_rq(cpu)->idle;
4590 * find_process_by_pid - find a process with a matching PID value.
4591 * @pid: the pid in question.
4593 static struct task_struct *find_process_by_pid(pid_t pid)
4595 return pid ? find_task_by_vpid(pid) : current;
4598 /* Actually do priority change: must hold rq lock. */
4600 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4602 BUG_ON(p->se.on_rq);
4605 p->rt_priority = prio;
4606 p->normal_prio = normal_prio(p);
4607 /* we are holding p->pi_lock already */
4608 p->prio = rt_mutex_getprio(p);
4609 if (rt_prio(p->prio))
4610 p->sched_class = &rt_sched_class;
4612 p->sched_class = &fair_sched_class;
4617 * check the target process has a UID that matches the current process's
4619 static bool check_same_owner(struct task_struct *p)
4621 const struct cred *cred = current_cred(), *pcred;
4625 pcred = __task_cred(p);
4626 match = (cred->euid == pcred->euid ||
4627 cred->euid == pcred->uid);
4632 static int __sched_setscheduler(struct task_struct *p, int policy,
4633 struct sched_param *param, bool user)
4635 int retval, oldprio, oldpolicy = -1, on_rq, running;
4636 unsigned long flags;
4637 const struct sched_class *prev_class;
4641 /* may grab non-irq protected spin_locks */
4642 BUG_ON(in_interrupt());
4644 /* double check policy once rq lock held */
4646 reset_on_fork = p->sched_reset_on_fork;
4647 policy = oldpolicy = p->policy;
4649 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4650 policy &= ~SCHED_RESET_ON_FORK;
4652 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4653 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4654 policy != SCHED_IDLE)
4659 * Valid priorities for SCHED_FIFO and SCHED_RR are
4660 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4661 * SCHED_BATCH and SCHED_IDLE is 0.
4663 if (param->sched_priority < 0 ||
4664 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4665 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4667 if (rt_policy(policy) != (param->sched_priority != 0))
4671 * Allow unprivileged RT tasks to decrease priority:
4673 if (user && !capable(CAP_SYS_NICE)) {
4674 if (rt_policy(policy)) {
4675 unsigned long rlim_rtprio;
4677 if (!lock_task_sighand(p, &flags))
4679 rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
4680 unlock_task_sighand(p, &flags);
4682 /* can't set/change the rt policy */
4683 if (policy != p->policy && !rlim_rtprio)
4686 /* can't increase priority */
4687 if (param->sched_priority > p->rt_priority &&
4688 param->sched_priority > rlim_rtprio)
4692 * Like positive nice levels, dont allow tasks to
4693 * move out of SCHED_IDLE either:
4695 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4698 /* can't change other user's priorities */
4699 if (!check_same_owner(p))
4702 /* Normal users shall not reset the sched_reset_on_fork flag */
4703 if (p->sched_reset_on_fork && !reset_on_fork)
4708 retval = security_task_setscheduler(p, policy, param);
4714 * make sure no PI-waiters arrive (or leave) while we are
4715 * changing the priority of the task:
4717 raw_spin_lock_irqsave(&p->pi_lock, flags);
4719 * To be able to change p->policy safely, the apropriate
4720 * runqueue lock must be held.
4722 rq = __task_rq_lock(p);
4724 #ifdef CONFIG_RT_GROUP_SCHED
4727 * Do not allow realtime tasks into groups that have no runtime
4730 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4731 task_group(p)->rt_bandwidth.rt_runtime == 0) {
4732 __task_rq_unlock(rq);
4733 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4739 /* recheck policy now with rq lock held */
4740 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4741 policy = oldpolicy = -1;
4742 __task_rq_unlock(rq);
4743 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4746 on_rq = p->se.on_rq;
4747 running = task_current(rq, p);
4749 deactivate_task(rq, p, 0);
4751 p->sched_class->put_prev_task(rq, p);
4753 p->sched_reset_on_fork = reset_on_fork;
4756 prev_class = p->sched_class;
4757 __setscheduler(rq, p, policy, param->sched_priority);
4760 p->sched_class->set_curr_task(rq);
4762 activate_task(rq, p, 0);
4764 check_class_changed(rq, p, prev_class, oldprio, running);
4766 __task_rq_unlock(rq);
4767 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4769 rt_mutex_adjust_pi(p);
4775 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4776 * @p: the task in question.
4777 * @policy: new policy.
4778 * @param: structure containing the new RT priority.
4780 * NOTE that the task may be already dead.
4782 int sched_setscheduler(struct task_struct *p, int policy,
4783 struct sched_param *param)
4785 return __sched_setscheduler(p, policy, param, true);
4787 EXPORT_SYMBOL_GPL(sched_setscheduler);
4790 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4791 * @p: the task in question.
4792 * @policy: new policy.
4793 * @param: structure containing the new RT priority.
4795 * Just like sched_setscheduler, only don't bother checking if the
4796 * current context has permission. For example, this is needed in
4797 * stop_machine(): we create temporary high priority worker threads,
4798 * but our caller might not have that capability.
4800 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4801 struct sched_param *param)
4803 return __sched_setscheduler(p, policy, param, false);
4807 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4809 struct sched_param lparam;
4810 struct task_struct *p;
4813 if (!param || pid < 0)
4815 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4820 p = find_process_by_pid(pid);
4822 retval = sched_setscheduler(p, policy, &lparam);
4829 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4830 * @pid: the pid in question.
4831 * @policy: new policy.
4832 * @param: structure containing the new RT priority.
4834 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4835 struct sched_param __user *, param)
4837 /* negative values for policy are not valid */
4841 return do_sched_setscheduler(pid, policy, param);
4845 * sys_sched_setparam - set/change the RT priority of a thread
4846 * @pid: the pid in question.
4847 * @param: structure containing the new RT priority.
4849 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4851 return do_sched_setscheduler(pid, -1, param);
4855 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4856 * @pid: the pid in question.
4858 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4860 struct task_struct *p;
4868 p = find_process_by_pid(pid);
4870 retval = security_task_getscheduler(p);
4873 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4880 * sys_sched_getparam - get the RT priority of a thread
4881 * @pid: the pid in question.
4882 * @param: structure containing the RT priority.
4884 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4886 struct sched_param lp;
4887 struct task_struct *p;
4890 if (!param || pid < 0)
4894 p = find_process_by_pid(pid);
4899 retval = security_task_getscheduler(p);
4903 lp.sched_priority = p->rt_priority;
4907 * This one might sleep, we cannot do it with a spinlock held ...
4909 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4918 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4920 cpumask_var_t cpus_allowed, new_mask;
4921 struct task_struct *p;
4927 p = find_process_by_pid(pid);
4934 /* Prevent p going away */
4938 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4942 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4944 goto out_free_cpus_allowed;
4947 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4950 retval = security_task_setscheduler(p, 0, NULL);
4954 cpuset_cpus_allowed(p, cpus_allowed);
4955 cpumask_and(new_mask, in_mask, cpus_allowed);
4957 retval = set_cpus_allowed_ptr(p, new_mask);
4960 cpuset_cpus_allowed(p, cpus_allowed);
4961 if (!cpumask_subset(new_mask, cpus_allowed)) {
4963 * We must have raced with a concurrent cpuset
4964 * update. Just reset the cpus_allowed to the
4965 * cpuset's cpus_allowed
4967 cpumask_copy(new_mask, cpus_allowed);
4972 free_cpumask_var(new_mask);
4973 out_free_cpus_allowed:
4974 free_cpumask_var(cpus_allowed);
4981 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4982 struct cpumask *new_mask)
4984 if (len < cpumask_size())
4985 cpumask_clear(new_mask);
4986 else if (len > cpumask_size())
4987 len = cpumask_size();
4989 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4993 * sys_sched_setaffinity - set the cpu affinity of a process
4994 * @pid: pid of the process
4995 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4996 * @user_mask_ptr: user-space pointer to the new cpu mask
4998 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4999 unsigned long __user *, user_mask_ptr)
5001 cpumask_var_t new_mask;
5004 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5007 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5009 retval = sched_setaffinity(pid, new_mask);
5010 free_cpumask_var(new_mask);
5014 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5016 struct task_struct *p;
5017 unsigned long flags;
5025 p = find_process_by_pid(pid);
5029 retval = security_task_getscheduler(p);
5033 rq = task_rq_lock(p, &flags);
5034 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5035 task_rq_unlock(rq, &flags);
5045 * sys_sched_getaffinity - get the cpu affinity of a process
5046 * @pid: pid of the process
5047 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5048 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5050 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5051 unsigned long __user *, user_mask_ptr)
5056 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5058 if (len & (sizeof(unsigned long)-1))
5061 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5064 ret = sched_getaffinity(pid, mask);
5066 size_t retlen = min_t(size_t, len, cpumask_size());
5068 if (copy_to_user(user_mask_ptr, mask, retlen))
5073 free_cpumask_var(mask);
5079 * sys_sched_yield - yield the current processor to other threads.
5081 * This function yields the current CPU to other tasks. If there are no
5082 * other threads running on this CPU then this function will return.
5084 SYSCALL_DEFINE0(sched_yield)
5086 struct rq *rq = this_rq_lock();
5088 schedstat_inc(rq, yld_count);
5089 current->sched_class->yield_task(rq);
5092 * Since we are going to call schedule() anyway, there's
5093 * no need to preempt or enable interrupts:
5095 __release(rq->lock);
5096 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5097 do_raw_spin_unlock(&rq->lock);
5098 preempt_enable_no_resched();
5105 static inline int should_resched(void)
5107 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5110 static void __cond_resched(void)
5112 add_preempt_count(PREEMPT_ACTIVE);
5114 sub_preempt_count(PREEMPT_ACTIVE);
5117 int __sched _cond_resched(void)
5119 if (should_resched()) {
5125 EXPORT_SYMBOL(_cond_resched);
5128 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5129 * call schedule, and on return reacquire the lock.
5131 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5132 * operations here to prevent schedule() from being called twice (once via
5133 * spin_unlock(), once by hand).
5135 int __cond_resched_lock(spinlock_t *lock)
5137 int resched = should_resched();
5140 lockdep_assert_held(lock);
5142 if (spin_needbreak(lock) || resched) {
5153 EXPORT_SYMBOL(__cond_resched_lock);
5155 int __sched __cond_resched_softirq(void)
5157 BUG_ON(!in_softirq());
5159 if (should_resched()) {
5167 EXPORT_SYMBOL(__cond_resched_softirq);
5170 * yield - yield the current processor to other threads.
5172 * This is a shortcut for kernel-space yielding - it marks the
5173 * thread runnable and calls sys_sched_yield().
5175 void __sched yield(void)
5177 set_current_state(TASK_RUNNING);
5180 EXPORT_SYMBOL(yield);
5183 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5184 * that process accounting knows that this is a task in IO wait state.
5186 void __sched io_schedule(void)
5188 struct rq *rq = raw_rq();
5190 delayacct_blkio_start();
5191 atomic_inc(&rq->nr_iowait);
5192 current->in_iowait = 1;
5194 current->in_iowait = 0;
5195 atomic_dec(&rq->nr_iowait);
5196 delayacct_blkio_end();
5198 EXPORT_SYMBOL(io_schedule);
5200 long __sched io_schedule_timeout(long timeout)
5202 struct rq *rq = raw_rq();
5205 delayacct_blkio_start();
5206 atomic_inc(&rq->nr_iowait);
5207 current->in_iowait = 1;
5208 ret = schedule_timeout(timeout);
5209 current->in_iowait = 0;
5210 atomic_dec(&rq->nr_iowait);
5211 delayacct_blkio_end();
5216 * sys_sched_get_priority_max - return maximum RT priority.
5217 * @policy: scheduling class.
5219 * this syscall returns the maximum rt_priority that can be used
5220 * by a given scheduling class.
5222 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5229 ret = MAX_USER_RT_PRIO-1;
5241 * sys_sched_get_priority_min - return minimum RT priority.
5242 * @policy: scheduling class.
5244 * this syscall returns the minimum rt_priority that can be used
5245 * by a given scheduling class.
5247 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5265 * sys_sched_rr_get_interval - return the default timeslice of a process.
5266 * @pid: pid of the process.
5267 * @interval: userspace pointer to the timeslice value.
5269 * this syscall writes the default timeslice value of a given process
5270 * into the user-space timespec buffer. A value of '0' means infinity.
5272 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5273 struct timespec __user *, interval)
5275 struct task_struct *p;
5276 unsigned int time_slice;
5277 unsigned long flags;
5287 p = find_process_by_pid(pid);
5291 retval = security_task_getscheduler(p);
5295 rq = task_rq_lock(p, &flags);
5296 time_slice = p->sched_class->get_rr_interval(rq, p);
5297 task_rq_unlock(rq, &flags);
5300 jiffies_to_timespec(time_slice, &t);
5301 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5309 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5311 void sched_show_task(struct task_struct *p)
5313 unsigned long free = 0;
5316 state = p->state ? __ffs(p->state) + 1 : 0;
5317 printk(KERN_INFO "%-13.13s %c", p->comm,
5318 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5319 #if BITS_PER_LONG == 32
5320 if (state == TASK_RUNNING)
5321 printk(KERN_CONT " running ");
5323 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5325 if (state == TASK_RUNNING)
5326 printk(KERN_CONT " running task ");
5328 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5330 #ifdef CONFIG_DEBUG_STACK_USAGE
5331 free = stack_not_used(p);
5333 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5334 task_pid_nr(p), task_pid_nr(p->real_parent),
5335 (unsigned long)task_thread_info(p)->flags);
5337 show_stack(p, NULL);
5340 void show_state_filter(unsigned long state_filter)
5342 struct task_struct *g, *p;
5344 #if BITS_PER_LONG == 32
5346 " task PC stack pid father\n");
5349 " task PC stack pid father\n");
5351 read_lock(&tasklist_lock);
5352 do_each_thread(g, p) {
5354 * reset the NMI-timeout, listing all files on a slow
5355 * console might take alot of time:
5357 touch_nmi_watchdog();
5358 if (!state_filter || (p->state & state_filter))
5360 } while_each_thread(g, p);
5362 touch_all_softlockup_watchdogs();
5364 #ifdef CONFIG_SCHED_DEBUG
5365 sysrq_sched_debug_show();
5367 read_unlock(&tasklist_lock);
5369 * Only show locks if all tasks are dumped:
5372 debug_show_all_locks();
5375 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5377 idle->sched_class = &idle_sched_class;
5381 * init_idle - set up an idle thread for a given CPU
5382 * @idle: task in question
5383 * @cpu: cpu the idle task belongs to
5385 * NOTE: this function does not set the idle thread's NEED_RESCHED
5386 * flag, to make booting more robust.
5388 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5390 struct rq *rq = cpu_rq(cpu);
5391 unsigned long flags;
5393 raw_spin_lock_irqsave(&rq->lock, flags);
5396 idle->state = TASK_RUNNING;
5397 idle->se.exec_start = sched_clock();
5399 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5401 * We're having a chicken and egg problem, even though we are
5402 * holding rq->lock, the cpu isn't yet set to this cpu so the
5403 * lockdep check in task_group() will fail.
5405 * Similar case to sched_fork(). / Alternatively we could
5406 * use task_rq_lock() here and obtain the other rq->lock.
5411 __set_task_cpu(idle, cpu);
5414 rq->curr = rq->idle = idle;
5415 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5418 raw_spin_unlock_irqrestore(&rq->lock, flags);
5420 /* Set the preempt count _outside_ the spinlocks! */
5421 #if defined(CONFIG_PREEMPT)
5422 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5424 task_thread_info(idle)->preempt_count = 0;
5427 * The idle tasks have their own, simple scheduling class:
5429 idle->sched_class = &idle_sched_class;
5430 ftrace_graph_init_task(idle);
5434 * In a system that switches off the HZ timer nohz_cpu_mask
5435 * indicates which cpus entered this state. This is used
5436 * in the rcu update to wait only for active cpus. For system
5437 * which do not switch off the HZ timer nohz_cpu_mask should
5438 * always be CPU_BITS_NONE.
5440 cpumask_var_t nohz_cpu_mask;
5443 * Increase the granularity value when there are more CPUs,
5444 * because with more CPUs the 'effective latency' as visible
5445 * to users decreases. But the relationship is not linear,
5446 * so pick a second-best guess by going with the log2 of the
5449 * This idea comes from the SD scheduler of Con Kolivas:
5451 static int get_update_sysctl_factor(void)
5453 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5454 unsigned int factor;
5456 switch (sysctl_sched_tunable_scaling) {
5457 case SCHED_TUNABLESCALING_NONE:
5460 case SCHED_TUNABLESCALING_LINEAR:
5463 case SCHED_TUNABLESCALING_LOG:
5465 factor = 1 + ilog2(cpus);
5472 static void update_sysctl(void)
5474 unsigned int factor = get_update_sysctl_factor();
5476 #define SET_SYSCTL(name) \
5477 (sysctl_##name = (factor) * normalized_sysctl_##name)
5478 SET_SYSCTL(sched_min_granularity);
5479 SET_SYSCTL(sched_latency);
5480 SET_SYSCTL(sched_wakeup_granularity);
5481 SET_SYSCTL(sched_shares_ratelimit);
5485 static inline void sched_init_granularity(void)
5492 * This is how migration works:
5494 * 1) we invoke migration_cpu_stop() on the target CPU using
5496 * 2) stopper starts to run (implicitly forcing the migrated thread
5498 * 3) it checks whether the migrated task is still in the wrong runqueue.
5499 * 4) if it's in the wrong runqueue then the migration thread removes
5500 * it and puts it into the right queue.
5501 * 5) stopper completes and stop_one_cpu() returns and the migration
5506 * Change a given task's CPU affinity. Migrate the thread to a
5507 * proper CPU and schedule it away if the CPU it's executing on
5508 * is removed from the allowed bitmask.
5510 * NOTE: the caller must have a valid reference to the task, the
5511 * task must not exit() & deallocate itself prematurely. The
5512 * call is not atomic; no spinlocks may be held.
5514 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5516 unsigned long flags;
5518 unsigned int dest_cpu;
5522 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5523 * drop the rq->lock and still rely on ->cpus_allowed.
5526 while (task_is_waking(p))
5528 rq = task_rq_lock(p, &flags);
5529 if (task_is_waking(p)) {
5530 task_rq_unlock(rq, &flags);
5534 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5539 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5540 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5545 if (p->sched_class->set_cpus_allowed)
5546 p->sched_class->set_cpus_allowed(p, new_mask);
5548 cpumask_copy(&p->cpus_allowed, new_mask);
5549 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5552 /* Can the task run on the task's current CPU? If so, we're done */
5553 if (cpumask_test_cpu(task_cpu(p), new_mask))
5556 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5557 if (migrate_task(p, dest_cpu)) {
5558 struct migration_arg arg = { p, dest_cpu };
5559 /* Need help from migration thread: drop lock and wait. */
5560 task_rq_unlock(rq, &flags);
5561 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5562 tlb_migrate_finish(p->mm);
5566 task_rq_unlock(rq, &flags);
5570 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5573 * Move (not current) task off this cpu, onto dest cpu. We're doing
5574 * this because either it can't run here any more (set_cpus_allowed()
5575 * away from this CPU, or CPU going down), or because we're
5576 * attempting to rebalance this task on exec (sched_exec).
5578 * So we race with normal scheduler movements, but that's OK, as long
5579 * as the task is no longer on this CPU.
5581 * Returns non-zero if task was successfully migrated.
5583 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5585 struct rq *rq_dest, *rq_src;
5588 if (unlikely(!cpu_active(dest_cpu)))
5591 rq_src = cpu_rq(src_cpu);
5592 rq_dest = cpu_rq(dest_cpu);
5594 double_rq_lock(rq_src, rq_dest);
5595 /* Already moved. */
5596 if (task_cpu(p) != src_cpu)
5598 /* Affinity changed (again). */
5599 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5603 * If we're not on a rq, the next wake-up will ensure we're
5607 deactivate_task(rq_src, p, 0);
5608 set_task_cpu(p, dest_cpu);
5609 activate_task(rq_dest, p, 0);
5610 check_preempt_curr(rq_dest, p, 0);
5615 double_rq_unlock(rq_src, rq_dest);
5620 * migration_cpu_stop - this will be executed by a highprio stopper thread
5621 * and performs thread migration by bumping thread off CPU then
5622 * 'pushing' onto another runqueue.
5624 static int migration_cpu_stop(void *data)
5626 struct migration_arg *arg = data;
5629 * The original target cpu might have gone down and we might
5630 * be on another cpu but it doesn't matter.
5632 local_irq_disable();
5633 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5638 #ifdef CONFIG_HOTPLUG_CPU
5640 * Figure out where task on dead CPU should go, use force if necessary.
5642 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5644 struct rq *rq = cpu_rq(dead_cpu);
5645 int needs_cpu, uninitialized_var(dest_cpu);
5646 unsigned long flags;
5648 local_irq_save(flags);
5650 raw_spin_lock(&rq->lock);
5651 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5653 dest_cpu = select_fallback_rq(dead_cpu, p);
5654 raw_spin_unlock(&rq->lock);
5656 * It can only fail if we race with set_cpus_allowed(),
5657 * in the racer should migrate the task anyway.
5660 __migrate_task(p, dead_cpu, dest_cpu);
5661 local_irq_restore(flags);
5665 * While a dead CPU has no uninterruptible tasks queued at this point,
5666 * it might still have a nonzero ->nr_uninterruptible counter, because
5667 * for performance reasons the counter is not stricly tracking tasks to
5668 * their home CPUs. So we just add the counter to another CPU's counter,
5669 * to keep the global sum constant after CPU-down:
5671 static void migrate_nr_uninterruptible(struct rq *rq_src)
5673 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5674 unsigned long flags;
5676 local_irq_save(flags);
5677 double_rq_lock(rq_src, rq_dest);
5678 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5679 rq_src->nr_uninterruptible = 0;
5680 double_rq_unlock(rq_src, rq_dest);
5681 local_irq_restore(flags);
5684 /* Run through task list and migrate tasks from the dead cpu. */
5685 static void migrate_live_tasks(int src_cpu)
5687 struct task_struct *p, *t;
5689 read_lock(&tasklist_lock);
5691 do_each_thread(t, p) {
5695 if (task_cpu(p) == src_cpu)
5696 move_task_off_dead_cpu(src_cpu, p);
5697 } while_each_thread(t, p);
5699 read_unlock(&tasklist_lock);
5703 * Schedules idle task to be the next runnable task on current CPU.
5704 * It does so by boosting its priority to highest possible.
5705 * Used by CPU offline code.
5707 void sched_idle_next(void)
5709 int this_cpu = smp_processor_id();
5710 struct rq *rq = cpu_rq(this_cpu);
5711 struct task_struct *p = rq->idle;
5712 unsigned long flags;
5714 /* cpu has to be offline */
5715 BUG_ON(cpu_online(this_cpu));
5718 * Strictly not necessary since rest of the CPUs are stopped by now
5719 * and interrupts disabled on the current cpu.
5721 raw_spin_lock_irqsave(&rq->lock, flags);
5723 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5725 activate_task(rq, p, 0);
5727 raw_spin_unlock_irqrestore(&rq->lock, flags);
5731 * Ensures that the idle task is using init_mm right before its cpu goes
5734 void idle_task_exit(void)
5736 struct mm_struct *mm = current->active_mm;
5738 BUG_ON(cpu_online(smp_processor_id()));
5741 switch_mm(mm, &init_mm, current);
5745 /* called under rq->lock with disabled interrupts */
5746 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5748 struct rq *rq = cpu_rq(dead_cpu);
5750 /* Must be exiting, otherwise would be on tasklist. */
5751 BUG_ON(!p->exit_state);
5753 /* Cannot have done final schedule yet: would have vanished. */
5754 BUG_ON(p->state == TASK_DEAD);
5759 * Drop lock around migration; if someone else moves it,
5760 * that's OK. No task can be added to this CPU, so iteration is
5763 raw_spin_unlock_irq(&rq->lock);
5764 move_task_off_dead_cpu(dead_cpu, p);
5765 raw_spin_lock_irq(&rq->lock);
5770 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5771 static void migrate_dead_tasks(unsigned int dead_cpu)
5773 struct rq *rq = cpu_rq(dead_cpu);
5774 struct task_struct *next;
5777 if (!rq->nr_running)
5779 next = pick_next_task(rq);
5782 next->sched_class->put_prev_task(rq, next);
5783 migrate_dead(dead_cpu, next);
5789 * remove the tasks which were accounted by rq from calc_load_tasks.
5791 static void calc_global_load_remove(struct rq *rq)
5793 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5794 rq->calc_load_active = 0;
5796 #endif /* CONFIG_HOTPLUG_CPU */
5798 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5800 static struct ctl_table sd_ctl_dir[] = {
5802 .procname = "sched_domain",
5808 static struct ctl_table sd_ctl_root[] = {
5810 .procname = "kernel",
5812 .child = sd_ctl_dir,
5817 static struct ctl_table *sd_alloc_ctl_entry(int n)
5819 struct ctl_table *entry =
5820 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5825 static void sd_free_ctl_entry(struct ctl_table **tablep)
5827 struct ctl_table *entry;
5830 * In the intermediate directories, both the child directory and
5831 * procname are dynamically allocated and could fail but the mode
5832 * will always be set. In the lowest directory the names are
5833 * static strings and all have proc handlers.
5835 for (entry = *tablep; entry->mode; entry++) {
5837 sd_free_ctl_entry(&entry->child);
5838 if (entry->proc_handler == NULL)
5839 kfree(entry->procname);
5847 set_table_entry(struct ctl_table *entry,
5848 const char *procname, void *data, int maxlen,
5849 mode_t mode, proc_handler *proc_handler)
5851 entry->procname = procname;
5853 entry->maxlen = maxlen;
5855 entry->proc_handler = proc_handler;
5858 static struct ctl_table *
5859 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5861 struct ctl_table *table = sd_alloc_ctl_entry(13);
5866 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5867 sizeof(long), 0644, proc_doulongvec_minmax);
5868 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5869 sizeof(long), 0644, proc_doulongvec_minmax);
5870 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5871 sizeof(int), 0644, proc_dointvec_minmax);
5872 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5873 sizeof(int), 0644, proc_dointvec_minmax);
5874 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5875 sizeof(int), 0644, proc_dointvec_minmax);
5876 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5877 sizeof(int), 0644, proc_dointvec_minmax);
5878 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5879 sizeof(int), 0644, proc_dointvec_minmax);
5880 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5881 sizeof(int), 0644, proc_dointvec_minmax);
5882 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5883 sizeof(int), 0644, proc_dointvec_minmax);
5884 set_table_entry(&table[9], "cache_nice_tries",
5885 &sd->cache_nice_tries,
5886 sizeof(int), 0644, proc_dointvec_minmax);
5887 set_table_entry(&table[10], "flags", &sd->flags,
5888 sizeof(int), 0644, proc_dointvec_minmax);
5889 set_table_entry(&table[11], "name", sd->name,
5890 CORENAME_MAX_SIZE, 0444, proc_dostring);
5891 /* &table[12] is terminator */
5896 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5898 struct ctl_table *entry, *table;
5899 struct sched_domain *sd;
5900 int domain_num = 0, i;
5903 for_each_domain(cpu, sd)
5905 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5910 for_each_domain(cpu, sd) {
5911 snprintf(buf, 32, "domain%d", i);
5912 entry->procname = kstrdup(buf, GFP_KERNEL);
5914 entry->child = sd_alloc_ctl_domain_table(sd);
5921 static struct ctl_table_header *sd_sysctl_header;
5922 static void register_sched_domain_sysctl(void)
5924 int i, cpu_num = num_possible_cpus();
5925 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5928 WARN_ON(sd_ctl_dir[0].child);
5929 sd_ctl_dir[0].child = entry;
5934 for_each_possible_cpu(i) {
5935 snprintf(buf, 32, "cpu%d", i);
5936 entry->procname = kstrdup(buf, GFP_KERNEL);
5938 entry->child = sd_alloc_ctl_cpu_table(i);
5942 WARN_ON(sd_sysctl_header);
5943 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5946 /* may be called multiple times per register */
5947 static void unregister_sched_domain_sysctl(void)
5949 if (sd_sysctl_header)
5950 unregister_sysctl_table(sd_sysctl_header);
5951 sd_sysctl_header = NULL;
5952 if (sd_ctl_dir[0].child)
5953 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5956 static void register_sched_domain_sysctl(void)
5959 static void unregister_sched_domain_sysctl(void)
5964 static void set_rq_online(struct rq *rq)
5967 const struct sched_class *class;
5969 cpumask_set_cpu(rq->cpu, rq->rd->online);
5972 for_each_class(class) {
5973 if (class->rq_online)
5974 class->rq_online(rq);
5979 static void set_rq_offline(struct rq *rq)
5982 const struct sched_class *class;
5984 for_each_class(class) {
5985 if (class->rq_offline)
5986 class->rq_offline(rq);
5989 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5995 * migration_call - callback that gets triggered when a CPU is added.
5996 * Here we can start up the necessary migration thread for the new CPU.
5998 static int __cpuinit
5999 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6001 int cpu = (long)hcpu;
6002 unsigned long flags;
6003 struct rq *rq = cpu_rq(cpu);
6007 case CPU_UP_PREPARE:
6008 case CPU_UP_PREPARE_FROZEN:
6009 rq->calc_load_update = calc_load_update;
6013 case CPU_ONLINE_FROZEN:
6014 /* Update our root-domain */
6015 raw_spin_lock_irqsave(&rq->lock, flags);
6017 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6021 raw_spin_unlock_irqrestore(&rq->lock, flags);
6024 #ifdef CONFIG_HOTPLUG_CPU
6026 case CPU_DEAD_FROZEN:
6027 migrate_live_tasks(cpu);
6028 /* Idle task back to normal (off runqueue, low prio) */
6029 raw_spin_lock_irq(&rq->lock);
6030 deactivate_task(rq, rq->idle, 0);
6031 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6032 rq->idle->sched_class = &idle_sched_class;
6033 migrate_dead_tasks(cpu);
6034 raw_spin_unlock_irq(&rq->lock);
6035 migrate_nr_uninterruptible(rq);
6036 BUG_ON(rq->nr_running != 0);
6037 calc_global_load_remove(rq);
6041 case CPU_DYING_FROZEN:
6042 /* Update our root-domain */
6043 raw_spin_lock_irqsave(&rq->lock, flags);
6045 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6048 raw_spin_unlock_irqrestore(&rq->lock, flags);
6056 * Register at high priority so that task migration (migrate_all_tasks)
6057 * happens before everything else. This has to be lower priority than
6058 * the notifier in the perf_event subsystem, though.
6060 static struct notifier_block __cpuinitdata migration_notifier = {
6061 .notifier_call = migration_call,
6065 static int __init migration_init(void)
6067 void *cpu = (void *)(long)smp_processor_id();
6070 /* Start one for the boot CPU: */
6071 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6072 BUG_ON(err == NOTIFY_BAD);
6073 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6074 register_cpu_notifier(&migration_notifier);
6078 early_initcall(migration_init);
6083 #ifdef CONFIG_SCHED_DEBUG
6085 static __read_mostly int sched_domain_debug_enabled;
6087 static int __init sched_domain_debug_setup(char *str)
6089 sched_domain_debug_enabled = 1;
6093 early_param("sched_debug", sched_domain_debug_setup);
6095 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6096 struct cpumask *groupmask)
6098 struct sched_group *group = sd->groups;
6101 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6102 cpumask_clear(groupmask);
6104 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6106 if (!(sd->flags & SD_LOAD_BALANCE)) {
6107 printk("does not load-balance\n");
6109 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6114 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6116 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6117 printk(KERN_ERR "ERROR: domain->span does not contain "
6120 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6121 printk(KERN_ERR "ERROR: domain->groups does not contain"
6125 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6129 printk(KERN_ERR "ERROR: group is NULL\n");
6133 if (!group->cpu_power) {
6134 printk(KERN_CONT "\n");
6135 printk(KERN_ERR "ERROR: domain->cpu_power not "
6140 if (!cpumask_weight(sched_group_cpus(group))) {
6141 printk(KERN_CONT "\n");
6142 printk(KERN_ERR "ERROR: empty group\n");
6146 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6147 printk(KERN_CONT "\n");
6148 printk(KERN_ERR "ERROR: repeated CPUs\n");
6152 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6154 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6156 printk(KERN_CONT " %s", str);
6157 if (group->cpu_power != SCHED_LOAD_SCALE) {
6158 printk(KERN_CONT " (cpu_power = %d)",
6162 group = group->next;
6163 } while (group != sd->groups);
6164 printk(KERN_CONT "\n");
6166 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6167 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6170 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6171 printk(KERN_ERR "ERROR: parent span is not a superset "
6172 "of domain->span\n");
6176 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6178 cpumask_var_t groupmask;
6181 if (!sched_domain_debug_enabled)
6185 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6189 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6191 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6192 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6197 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6204 free_cpumask_var(groupmask);
6206 #else /* !CONFIG_SCHED_DEBUG */
6207 # define sched_domain_debug(sd, cpu) do { } while (0)
6208 #endif /* CONFIG_SCHED_DEBUG */
6210 static int sd_degenerate(struct sched_domain *sd)
6212 if (cpumask_weight(sched_domain_span(sd)) == 1)
6215 /* Following flags need at least 2 groups */
6216 if (sd->flags & (SD_LOAD_BALANCE |
6217 SD_BALANCE_NEWIDLE |
6221 SD_SHARE_PKG_RESOURCES)) {
6222 if (sd->groups != sd->groups->next)
6226 /* Following flags don't use groups */
6227 if (sd->flags & (SD_WAKE_AFFINE))
6234 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6236 unsigned long cflags = sd->flags, pflags = parent->flags;
6238 if (sd_degenerate(parent))
6241 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6244 /* Flags needing groups don't count if only 1 group in parent */
6245 if (parent->groups == parent->groups->next) {
6246 pflags &= ~(SD_LOAD_BALANCE |
6247 SD_BALANCE_NEWIDLE |
6251 SD_SHARE_PKG_RESOURCES);
6252 if (nr_node_ids == 1)
6253 pflags &= ~SD_SERIALIZE;
6255 if (~cflags & pflags)
6261 static void free_rootdomain(struct root_domain *rd)
6263 synchronize_sched();
6265 cpupri_cleanup(&rd->cpupri);
6267 free_cpumask_var(rd->rto_mask);
6268 free_cpumask_var(rd->online);
6269 free_cpumask_var(rd->span);
6273 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6275 struct root_domain *old_rd = NULL;
6276 unsigned long flags;
6278 raw_spin_lock_irqsave(&rq->lock, flags);
6283 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6286 cpumask_clear_cpu(rq->cpu, old_rd->span);
6289 * If we dont want to free the old_rt yet then
6290 * set old_rd to NULL to skip the freeing later
6293 if (!atomic_dec_and_test(&old_rd->refcount))
6297 atomic_inc(&rd->refcount);
6300 cpumask_set_cpu(rq->cpu, rd->span);
6301 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6304 raw_spin_unlock_irqrestore(&rq->lock, flags);
6307 free_rootdomain(old_rd);
6310 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6312 gfp_t gfp = GFP_KERNEL;
6314 memset(rd, 0, sizeof(*rd));
6319 if (!alloc_cpumask_var(&rd->span, gfp))
6321 if (!alloc_cpumask_var(&rd->online, gfp))
6323 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6326 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6331 free_cpumask_var(rd->rto_mask);
6333 free_cpumask_var(rd->online);
6335 free_cpumask_var(rd->span);
6340 static void init_defrootdomain(void)
6342 init_rootdomain(&def_root_domain, true);
6344 atomic_set(&def_root_domain.refcount, 1);
6347 static struct root_domain *alloc_rootdomain(void)
6349 struct root_domain *rd;
6351 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6355 if (init_rootdomain(rd, false) != 0) {
6364 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6365 * hold the hotplug lock.
6368 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6370 struct rq *rq = cpu_rq(cpu);
6371 struct sched_domain *tmp;
6373 for (tmp = sd; tmp; tmp = tmp->parent)
6374 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6376 /* Remove the sched domains which do not contribute to scheduling. */
6377 for (tmp = sd; tmp; ) {
6378 struct sched_domain *parent = tmp->parent;
6382 if (sd_parent_degenerate(tmp, parent)) {
6383 tmp->parent = parent->parent;
6385 parent->parent->child = tmp;
6390 if (sd && sd_degenerate(sd)) {
6396 sched_domain_debug(sd, cpu);
6398 rq_attach_root(rq, rd);
6399 rcu_assign_pointer(rq->sd, sd);
6402 /* cpus with isolated domains */
6403 static cpumask_var_t cpu_isolated_map;
6405 /* Setup the mask of cpus configured for isolated domains */
6406 static int __init isolated_cpu_setup(char *str)
6408 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6409 cpulist_parse(str, cpu_isolated_map);
6413 __setup("isolcpus=", isolated_cpu_setup);
6416 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6417 * to a function which identifies what group(along with sched group) a CPU
6418 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6419 * (due to the fact that we keep track of groups covered with a struct cpumask).
6421 * init_sched_build_groups will build a circular linked list of the groups
6422 * covered by the given span, and will set each group's ->cpumask correctly,
6423 * and ->cpu_power to 0.
6426 init_sched_build_groups(const struct cpumask *span,
6427 const struct cpumask *cpu_map,
6428 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6429 struct sched_group **sg,
6430 struct cpumask *tmpmask),
6431 struct cpumask *covered, struct cpumask *tmpmask)
6433 struct sched_group *first = NULL, *last = NULL;
6436 cpumask_clear(covered);
6438 for_each_cpu(i, span) {
6439 struct sched_group *sg;
6440 int group = group_fn(i, cpu_map, &sg, tmpmask);
6443 if (cpumask_test_cpu(i, covered))
6446 cpumask_clear(sched_group_cpus(sg));
6449 for_each_cpu(j, span) {
6450 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6453 cpumask_set_cpu(j, covered);
6454 cpumask_set_cpu(j, sched_group_cpus(sg));
6465 #define SD_NODES_PER_DOMAIN 16
6470 * find_next_best_node - find the next node to include in a sched_domain
6471 * @node: node whose sched_domain we're building
6472 * @used_nodes: nodes already in the sched_domain
6474 * Find the next node to include in a given scheduling domain. Simply
6475 * finds the closest node not already in the @used_nodes map.
6477 * Should use nodemask_t.
6479 static int find_next_best_node(int node, nodemask_t *used_nodes)
6481 int i, n, val, min_val, best_node = 0;
6485 for (i = 0; i < nr_node_ids; i++) {
6486 /* Start at @node */
6487 n = (node + i) % nr_node_ids;
6489 if (!nr_cpus_node(n))
6492 /* Skip already used nodes */
6493 if (node_isset(n, *used_nodes))
6496 /* Simple min distance search */
6497 val = node_distance(node, n);
6499 if (val < min_val) {
6505 node_set(best_node, *used_nodes);
6510 * sched_domain_node_span - get a cpumask for a node's sched_domain
6511 * @node: node whose cpumask we're constructing
6512 * @span: resulting cpumask
6514 * Given a node, construct a good cpumask for its sched_domain to span. It
6515 * should be one that prevents unnecessary balancing, but also spreads tasks
6518 static void sched_domain_node_span(int node, struct cpumask *span)
6520 nodemask_t used_nodes;
6523 cpumask_clear(span);
6524 nodes_clear(used_nodes);
6526 cpumask_or(span, span, cpumask_of_node(node));
6527 node_set(node, used_nodes);
6529 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6530 int next_node = find_next_best_node(node, &used_nodes);
6532 cpumask_or(span, span, cpumask_of_node(next_node));
6535 #endif /* CONFIG_NUMA */
6537 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6540 * The cpus mask in sched_group and sched_domain hangs off the end.
6542 * ( See the the comments in include/linux/sched.h:struct sched_group
6543 * and struct sched_domain. )
6545 struct static_sched_group {
6546 struct sched_group sg;
6547 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6550 struct static_sched_domain {
6551 struct sched_domain sd;
6552 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6558 cpumask_var_t domainspan;
6559 cpumask_var_t covered;
6560 cpumask_var_t notcovered;
6562 cpumask_var_t nodemask;
6563 cpumask_var_t this_sibling_map;
6564 cpumask_var_t this_core_map;
6565 cpumask_var_t send_covered;
6566 cpumask_var_t tmpmask;
6567 struct sched_group **sched_group_nodes;
6568 struct root_domain *rd;
6572 sa_sched_groups = 0,
6577 sa_this_sibling_map,
6579 sa_sched_group_nodes,
6589 * SMT sched-domains:
6591 #ifdef CONFIG_SCHED_SMT
6592 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6593 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6596 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6597 struct sched_group **sg, struct cpumask *unused)
6600 *sg = &per_cpu(sched_groups, cpu).sg;
6603 #endif /* CONFIG_SCHED_SMT */
6606 * multi-core sched-domains:
6608 #ifdef CONFIG_SCHED_MC
6609 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6610 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6611 #endif /* CONFIG_SCHED_MC */
6613 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6615 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6616 struct sched_group **sg, struct cpumask *mask)
6620 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6621 group = cpumask_first(mask);
6623 *sg = &per_cpu(sched_group_core, group).sg;
6626 #elif defined(CONFIG_SCHED_MC)
6628 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6629 struct sched_group **sg, struct cpumask *unused)
6632 *sg = &per_cpu(sched_group_core, cpu).sg;
6637 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6638 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6641 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6642 struct sched_group **sg, struct cpumask *mask)
6645 #ifdef CONFIG_SCHED_MC
6646 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6647 group = cpumask_first(mask);
6648 #elif defined(CONFIG_SCHED_SMT)
6649 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6650 group = cpumask_first(mask);
6655 *sg = &per_cpu(sched_group_phys, group).sg;
6661 * The init_sched_build_groups can't handle what we want to do with node
6662 * groups, so roll our own. Now each node has its own list of groups which
6663 * gets dynamically allocated.
6665 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6666 static struct sched_group ***sched_group_nodes_bycpu;
6668 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6669 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6671 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6672 struct sched_group **sg,
6673 struct cpumask *nodemask)
6677 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6678 group = cpumask_first(nodemask);
6681 *sg = &per_cpu(sched_group_allnodes, group).sg;
6685 static void init_numa_sched_groups_power(struct sched_group *group_head)
6687 struct sched_group *sg = group_head;
6693 for_each_cpu(j, sched_group_cpus(sg)) {
6694 struct sched_domain *sd;
6696 sd = &per_cpu(phys_domains, j).sd;
6697 if (j != group_first_cpu(sd->groups)) {
6699 * Only add "power" once for each
6705 sg->cpu_power += sd->groups->cpu_power;
6708 } while (sg != group_head);
6711 static int build_numa_sched_groups(struct s_data *d,
6712 const struct cpumask *cpu_map, int num)
6714 struct sched_domain *sd;
6715 struct sched_group *sg, *prev;
6718 cpumask_clear(d->covered);
6719 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6720 if (cpumask_empty(d->nodemask)) {
6721 d->sched_group_nodes[num] = NULL;
6725 sched_domain_node_span(num, d->domainspan);
6726 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6728 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6731 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6735 d->sched_group_nodes[num] = sg;
6737 for_each_cpu(j, d->nodemask) {
6738 sd = &per_cpu(node_domains, j).sd;
6743 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6745 cpumask_or(d->covered, d->covered, d->nodemask);
6748 for (j = 0; j < nr_node_ids; j++) {
6749 n = (num + j) % nr_node_ids;
6750 cpumask_complement(d->notcovered, d->covered);
6751 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6752 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6753 if (cpumask_empty(d->tmpmask))
6755 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6756 if (cpumask_empty(d->tmpmask))
6758 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6762 "Can not alloc domain group for node %d\n", j);
6766 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6767 sg->next = prev->next;
6768 cpumask_or(d->covered, d->covered, d->tmpmask);
6775 #endif /* CONFIG_NUMA */
6778 /* Free memory allocated for various sched_group structures */
6779 static void free_sched_groups(const struct cpumask *cpu_map,
6780 struct cpumask *nodemask)
6784 for_each_cpu(cpu, cpu_map) {
6785 struct sched_group **sched_group_nodes
6786 = sched_group_nodes_bycpu[cpu];
6788 if (!sched_group_nodes)
6791 for (i = 0; i < nr_node_ids; i++) {
6792 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6794 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6795 if (cpumask_empty(nodemask))
6805 if (oldsg != sched_group_nodes[i])
6808 kfree(sched_group_nodes);
6809 sched_group_nodes_bycpu[cpu] = NULL;
6812 #else /* !CONFIG_NUMA */
6813 static void free_sched_groups(const struct cpumask *cpu_map,
6814 struct cpumask *nodemask)
6817 #endif /* CONFIG_NUMA */
6820 * Initialize sched groups cpu_power.
6822 * cpu_power indicates the capacity of sched group, which is used while
6823 * distributing the load between different sched groups in a sched domain.
6824 * Typically cpu_power for all the groups in a sched domain will be same unless
6825 * there are asymmetries in the topology. If there are asymmetries, group
6826 * having more cpu_power will pickup more load compared to the group having
6829 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6831 struct sched_domain *child;
6832 struct sched_group *group;
6836 WARN_ON(!sd || !sd->groups);
6838 if (cpu != group_first_cpu(sd->groups))
6843 sd->groups->cpu_power = 0;
6846 power = SCHED_LOAD_SCALE;
6847 weight = cpumask_weight(sched_domain_span(sd));
6849 * SMT siblings share the power of a single core.
6850 * Usually multiple threads get a better yield out of
6851 * that one core than a single thread would have,
6852 * reflect that in sd->smt_gain.
6854 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6855 power *= sd->smt_gain;
6857 power >>= SCHED_LOAD_SHIFT;
6859 sd->groups->cpu_power += power;
6864 * Add cpu_power of each child group to this groups cpu_power.
6866 group = child->groups;
6868 sd->groups->cpu_power += group->cpu_power;
6869 group = group->next;
6870 } while (group != child->groups);
6874 * Initializers for schedule domains
6875 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6878 #ifdef CONFIG_SCHED_DEBUG
6879 # define SD_INIT_NAME(sd, type) sd->name = #type
6881 # define SD_INIT_NAME(sd, type) do { } while (0)
6884 #define SD_INIT(sd, type) sd_init_##type(sd)
6886 #define SD_INIT_FUNC(type) \
6887 static noinline void sd_init_##type(struct sched_domain *sd) \
6889 memset(sd, 0, sizeof(*sd)); \
6890 *sd = SD_##type##_INIT; \
6891 sd->level = SD_LV_##type; \
6892 SD_INIT_NAME(sd, type); \
6897 SD_INIT_FUNC(ALLNODES)
6900 #ifdef CONFIG_SCHED_SMT
6901 SD_INIT_FUNC(SIBLING)
6903 #ifdef CONFIG_SCHED_MC
6907 static int default_relax_domain_level = -1;
6909 static int __init setup_relax_domain_level(char *str)
6913 val = simple_strtoul(str, NULL, 0);
6914 if (val < SD_LV_MAX)
6915 default_relax_domain_level = val;
6919 __setup("relax_domain_level=", setup_relax_domain_level);
6921 static void set_domain_attribute(struct sched_domain *sd,
6922 struct sched_domain_attr *attr)
6926 if (!attr || attr->relax_domain_level < 0) {
6927 if (default_relax_domain_level < 0)
6930 request = default_relax_domain_level;
6932 request = attr->relax_domain_level;
6933 if (request < sd->level) {
6934 /* turn off idle balance on this domain */
6935 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6937 /* turn on idle balance on this domain */
6938 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6942 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6943 const struct cpumask *cpu_map)
6946 case sa_sched_groups:
6947 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6948 d->sched_group_nodes = NULL;
6950 free_rootdomain(d->rd); /* fall through */
6952 free_cpumask_var(d->tmpmask); /* fall through */
6953 case sa_send_covered:
6954 free_cpumask_var(d->send_covered); /* fall through */
6955 case sa_this_core_map:
6956 free_cpumask_var(d->this_core_map); /* fall through */
6957 case sa_this_sibling_map:
6958 free_cpumask_var(d->this_sibling_map); /* fall through */
6960 free_cpumask_var(d->nodemask); /* fall through */
6961 case sa_sched_group_nodes:
6963 kfree(d->sched_group_nodes); /* fall through */
6965 free_cpumask_var(d->notcovered); /* fall through */
6967 free_cpumask_var(d->covered); /* fall through */
6969 free_cpumask_var(d->domainspan); /* fall through */
6976 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6977 const struct cpumask *cpu_map)
6980 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6982 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6983 return sa_domainspan;
6984 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6986 /* Allocate the per-node list of sched groups */
6987 d->sched_group_nodes = kcalloc(nr_node_ids,
6988 sizeof(struct sched_group *), GFP_KERNEL);
6989 if (!d->sched_group_nodes) {
6990 printk(KERN_WARNING "Can not alloc sched group node list\n");
6991 return sa_notcovered;
6993 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6995 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6996 return sa_sched_group_nodes;
6997 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6999 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
7000 return sa_this_sibling_map;
7001 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
7002 return sa_this_core_map;
7003 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
7004 return sa_send_covered;
7005 d->rd = alloc_rootdomain();
7007 printk(KERN_WARNING "Cannot alloc root domain\n");
7010 return sa_rootdomain;
7013 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
7014 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
7016 struct sched_domain *sd = NULL;
7018 struct sched_domain *parent;
7021 if (cpumask_weight(cpu_map) >
7022 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
7023 sd = &per_cpu(allnodes_domains, i).sd;
7024 SD_INIT(sd, ALLNODES);
7025 set_domain_attribute(sd, attr);
7026 cpumask_copy(sched_domain_span(sd), cpu_map);
7027 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
7032 sd = &per_cpu(node_domains, i).sd;
7034 set_domain_attribute(sd, attr);
7035 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7036 sd->parent = parent;
7039 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7044 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7045 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7046 struct sched_domain *parent, int i)
7048 struct sched_domain *sd;
7049 sd = &per_cpu(phys_domains, i).sd;
7051 set_domain_attribute(sd, attr);
7052 cpumask_copy(sched_domain_span(sd), d->nodemask);
7053 sd->parent = parent;
7056 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7060 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7061 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7062 struct sched_domain *parent, int i)
7064 struct sched_domain *sd = parent;
7065 #ifdef CONFIG_SCHED_MC
7066 sd = &per_cpu(core_domains, i).sd;
7068 set_domain_attribute(sd, attr);
7069 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7070 sd->parent = parent;
7072 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7077 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7078 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7079 struct sched_domain *parent, int i)
7081 struct sched_domain *sd = parent;
7082 #ifdef CONFIG_SCHED_SMT
7083 sd = &per_cpu(cpu_domains, i).sd;
7084 SD_INIT(sd, SIBLING);
7085 set_domain_attribute(sd, attr);
7086 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7087 sd->parent = parent;
7089 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7094 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7095 const struct cpumask *cpu_map, int cpu)
7098 #ifdef CONFIG_SCHED_SMT
7099 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7100 cpumask_and(d->this_sibling_map, cpu_map,
7101 topology_thread_cpumask(cpu));
7102 if (cpu == cpumask_first(d->this_sibling_map))
7103 init_sched_build_groups(d->this_sibling_map, cpu_map,
7105 d->send_covered, d->tmpmask);
7108 #ifdef CONFIG_SCHED_MC
7109 case SD_LV_MC: /* set up multi-core groups */
7110 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7111 if (cpu == cpumask_first(d->this_core_map))
7112 init_sched_build_groups(d->this_core_map, cpu_map,
7114 d->send_covered, d->tmpmask);
7117 case SD_LV_CPU: /* set up physical groups */
7118 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7119 if (!cpumask_empty(d->nodemask))
7120 init_sched_build_groups(d->nodemask, cpu_map,
7122 d->send_covered, d->tmpmask);
7125 case SD_LV_ALLNODES:
7126 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7127 d->send_covered, d->tmpmask);
7136 * Build sched domains for a given set of cpus and attach the sched domains
7137 * to the individual cpus
7139 static int __build_sched_domains(const struct cpumask *cpu_map,
7140 struct sched_domain_attr *attr)
7142 enum s_alloc alloc_state = sa_none;
7144 struct sched_domain *sd;
7150 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7151 if (alloc_state != sa_rootdomain)
7153 alloc_state = sa_sched_groups;
7156 * Set up domains for cpus specified by the cpu_map.
7158 for_each_cpu(i, cpu_map) {
7159 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7162 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7163 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7164 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7165 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7168 for_each_cpu(i, cpu_map) {
7169 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7170 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7173 /* Set up physical groups */
7174 for (i = 0; i < nr_node_ids; i++)
7175 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7178 /* Set up node groups */
7180 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7182 for (i = 0; i < nr_node_ids; i++)
7183 if (build_numa_sched_groups(&d, cpu_map, i))
7187 /* Calculate CPU power for physical packages and nodes */
7188 #ifdef CONFIG_SCHED_SMT
7189 for_each_cpu(i, cpu_map) {
7190 sd = &per_cpu(cpu_domains, i).sd;
7191 init_sched_groups_power(i, sd);
7194 #ifdef CONFIG_SCHED_MC
7195 for_each_cpu(i, cpu_map) {
7196 sd = &per_cpu(core_domains, i).sd;
7197 init_sched_groups_power(i, sd);
7201 for_each_cpu(i, cpu_map) {
7202 sd = &per_cpu(phys_domains, i).sd;
7203 init_sched_groups_power(i, sd);
7207 for (i = 0; i < nr_node_ids; i++)
7208 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7210 if (d.sd_allnodes) {
7211 struct sched_group *sg;
7213 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7215 init_numa_sched_groups_power(sg);
7219 /* Attach the domains */
7220 for_each_cpu(i, cpu_map) {
7221 #ifdef CONFIG_SCHED_SMT
7222 sd = &per_cpu(cpu_domains, i).sd;
7223 #elif defined(CONFIG_SCHED_MC)
7224 sd = &per_cpu(core_domains, i).sd;
7226 sd = &per_cpu(phys_domains, i).sd;
7228 cpu_attach_domain(sd, d.rd, i);
7231 d.sched_group_nodes = NULL; /* don't free this we still need it */
7232 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7236 __free_domain_allocs(&d, alloc_state, cpu_map);
7240 static int build_sched_domains(const struct cpumask *cpu_map)
7242 return __build_sched_domains(cpu_map, NULL);
7245 static cpumask_var_t *doms_cur; /* current sched domains */
7246 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7247 static struct sched_domain_attr *dattr_cur;
7248 /* attribues of custom domains in 'doms_cur' */
7251 * Special case: If a kmalloc of a doms_cur partition (array of
7252 * cpumask) fails, then fallback to a single sched domain,
7253 * as determined by the single cpumask fallback_doms.
7255 static cpumask_var_t fallback_doms;
7258 * arch_update_cpu_topology lets virtualized architectures update the
7259 * cpu core maps. It is supposed to return 1 if the topology changed
7260 * or 0 if it stayed the same.
7262 int __attribute__((weak)) arch_update_cpu_topology(void)
7267 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7270 cpumask_var_t *doms;
7272 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7275 for (i = 0; i < ndoms; i++) {
7276 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7277 free_sched_domains(doms, i);
7284 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7287 for (i = 0; i < ndoms; i++)
7288 free_cpumask_var(doms[i]);
7293 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7294 * For now this just excludes isolated cpus, but could be used to
7295 * exclude other special cases in the future.
7297 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7301 arch_update_cpu_topology();
7303 doms_cur = alloc_sched_domains(ndoms_cur);
7305 doms_cur = &fallback_doms;
7306 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7308 err = build_sched_domains(doms_cur[0]);
7309 register_sched_domain_sysctl();
7314 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7315 struct cpumask *tmpmask)
7317 free_sched_groups(cpu_map, tmpmask);
7321 * Detach sched domains from a group of cpus specified in cpu_map
7322 * These cpus will now be attached to the NULL domain
7324 static void detach_destroy_domains(const struct cpumask *cpu_map)
7326 /* Save because hotplug lock held. */
7327 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7330 for_each_cpu(i, cpu_map)
7331 cpu_attach_domain(NULL, &def_root_domain, i);
7332 synchronize_sched();
7333 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7336 /* handle null as "default" */
7337 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7338 struct sched_domain_attr *new, int idx_new)
7340 struct sched_domain_attr tmp;
7347 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7348 new ? (new + idx_new) : &tmp,
7349 sizeof(struct sched_domain_attr));
7353 * Partition sched domains as specified by the 'ndoms_new'
7354 * cpumasks in the array doms_new[] of cpumasks. This compares
7355 * doms_new[] to the current sched domain partitioning, doms_cur[].
7356 * It destroys each deleted domain and builds each new domain.
7358 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7359 * The masks don't intersect (don't overlap.) We should setup one
7360 * sched domain for each mask. CPUs not in any of the cpumasks will
7361 * not be load balanced. If the same cpumask appears both in the
7362 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7365 * The passed in 'doms_new' should be allocated using
7366 * alloc_sched_domains. This routine takes ownership of it and will
7367 * free_sched_domains it when done with it. If the caller failed the
7368 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7369 * and partition_sched_domains() will fallback to the single partition
7370 * 'fallback_doms', it also forces the domains to be rebuilt.
7372 * If doms_new == NULL it will be replaced with cpu_online_mask.
7373 * ndoms_new == 0 is a special case for destroying existing domains,
7374 * and it will not create the default domain.
7376 * Call with hotplug lock held
7378 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7379 struct sched_domain_attr *dattr_new)
7384 mutex_lock(&sched_domains_mutex);
7386 /* always unregister in case we don't destroy any domains */
7387 unregister_sched_domain_sysctl();
7389 /* Let architecture update cpu core mappings. */
7390 new_topology = arch_update_cpu_topology();
7392 n = doms_new ? ndoms_new : 0;
7394 /* Destroy deleted domains */
7395 for (i = 0; i < ndoms_cur; i++) {
7396 for (j = 0; j < n && !new_topology; j++) {
7397 if (cpumask_equal(doms_cur[i], doms_new[j])
7398 && dattrs_equal(dattr_cur, i, dattr_new, j))
7401 /* no match - a current sched domain not in new doms_new[] */
7402 detach_destroy_domains(doms_cur[i]);
7407 if (doms_new == NULL) {
7409 doms_new = &fallback_doms;
7410 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7411 WARN_ON_ONCE(dattr_new);
7414 /* Build new domains */
7415 for (i = 0; i < ndoms_new; i++) {
7416 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7417 if (cpumask_equal(doms_new[i], doms_cur[j])
7418 && dattrs_equal(dattr_new, i, dattr_cur, j))
7421 /* no match - add a new doms_new */
7422 __build_sched_domains(doms_new[i],
7423 dattr_new ? dattr_new + i : NULL);
7428 /* Remember the new sched domains */
7429 if (doms_cur != &fallback_doms)
7430 free_sched_domains(doms_cur, ndoms_cur);
7431 kfree(dattr_cur); /* kfree(NULL) is safe */
7432 doms_cur = doms_new;
7433 dattr_cur = dattr_new;
7434 ndoms_cur = ndoms_new;
7436 register_sched_domain_sysctl();
7438 mutex_unlock(&sched_domains_mutex);
7441 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7442 static void arch_reinit_sched_domains(void)
7446 /* Destroy domains first to force the rebuild */
7447 partition_sched_domains(0, NULL, NULL);
7449 rebuild_sched_domains();
7453 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7455 unsigned int level = 0;
7457 if (sscanf(buf, "%u", &level) != 1)
7461 * level is always be positive so don't check for
7462 * level < POWERSAVINGS_BALANCE_NONE which is 0
7463 * What happens on 0 or 1 byte write,
7464 * need to check for count as well?
7467 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7471 sched_smt_power_savings = level;
7473 sched_mc_power_savings = level;
7475 arch_reinit_sched_domains();
7480 #ifdef CONFIG_SCHED_MC
7481 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7482 struct sysdev_class_attribute *attr,
7485 return sprintf(page, "%u\n", sched_mc_power_savings);
7487 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7488 struct sysdev_class_attribute *attr,
7489 const char *buf, size_t count)
7491 return sched_power_savings_store(buf, count, 0);
7493 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7494 sched_mc_power_savings_show,
7495 sched_mc_power_savings_store);
7498 #ifdef CONFIG_SCHED_SMT
7499 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7500 struct sysdev_class_attribute *attr,
7503 return sprintf(page, "%u\n", sched_smt_power_savings);
7505 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7506 struct sysdev_class_attribute *attr,
7507 const char *buf, size_t count)
7509 return sched_power_savings_store(buf, count, 1);
7511 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7512 sched_smt_power_savings_show,
7513 sched_smt_power_savings_store);
7516 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7520 #ifdef CONFIG_SCHED_SMT
7522 err = sysfs_create_file(&cls->kset.kobj,
7523 &attr_sched_smt_power_savings.attr);
7525 #ifdef CONFIG_SCHED_MC
7526 if (!err && mc_capable())
7527 err = sysfs_create_file(&cls->kset.kobj,
7528 &attr_sched_mc_power_savings.attr);
7532 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7534 #ifndef CONFIG_CPUSETS
7536 * Add online and remove offline CPUs from the scheduler domains.
7537 * When cpusets are enabled they take over this function.
7539 static int update_sched_domains(struct notifier_block *nfb,
7540 unsigned long action, void *hcpu)
7544 case CPU_ONLINE_FROZEN:
7545 case CPU_DOWN_PREPARE:
7546 case CPU_DOWN_PREPARE_FROZEN:
7547 case CPU_DOWN_FAILED:
7548 case CPU_DOWN_FAILED_FROZEN:
7549 partition_sched_domains(1, NULL, NULL);
7558 static int update_runtime(struct notifier_block *nfb,
7559 unsigned long action, void *hcpu)
7561 int cpu = (int)(long)hcpu;
7564 case CPU_DOWN_PREPARE:
7565 case CPU_DOWN_PREPARE_FROZEN:
7566 disable_runtime(cpu_rq(cpu));
7569 case CPU_DOWN_FAILED:
7570 case CPU_DOWN_FAILED_FROZEN:
7572 case CPU_ONLINE_FROZEN:
7573 enable_runtime(cpu_rq(cpu));
7581 void __init sched_init_smp(void)
7583 cpumask_var_t non_isolated_cpus;
7585 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7586 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7588 #if defined(CONFIG_NUMA)
7589 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7591 BUG_ON(sched_group_nodes_bycpu == NULL);
7594 mutex_lock(&sched_domains_mutex);
7595 arch_init_sched_domains(cpu_active_mask);
7596 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7597 if (cpumask_empty(non_isolated_cpus))
7598 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7599 mutex_unlock(&sched_domains_mutex);
7602 #ifndef CONFIG_CPUSETS
7603 /* XXX: Theoretical race here - CPU may be hotplugged now */
7604 hotcpu_notifier(update_sched_domains, 0);
7607 /* RT runtime code needs to handle some hotplug events */
7608 hotcpu_notifier(update_runtime, 0);
7612 /* Move init over to a non-isolated CPU */
7613 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7615 sched_init_granularity();
7616 free_cpumask_var(non_isolated_cpus);
7618 init_sched_rt_class();
7621 void __init sched_init_smp(void)
7623 sched_init_granularity();
7625 #endif /* CONFIG_SMP */
7627 const_debug unsigned int sysctl_timer_migration = 1;
7629 int in_sched_functions(unsigned long addr)
7631 return in_lock_functions(addr) ||
7632 (addr >= (unsigned long)__sched_text_start
7633 && addr < (unsigned long)__sched_text_end);
7636 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7638 cfs_rq->tasks_timeline = RB_ROOT;
7639 INIT_LIST_HEAD(&cfs_rq->tasks);
7640 #ifdef CONFIG_FAIR_GROUP_SCHED
7643 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7646 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7648 struct rt_prio_array *array;
7651 array = &rt_rq->active;
7652 for (i = 0; i < MAX_RT_PRIO; i++) {
7653 INIT_LIST_HEAD(array->queue + i);
7654 __clear_bit(i, array->bitmap);
7656 /* delimiter for bitsearch: */
7657 __set_bit(MAX_RT_PRIO, array->bitmap);
7659 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7660 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7662 rt_rq->highest_prio.next = MAX_RT_PRIO;
7666 rt_rq->rt_nr_migratory = 0;
7667 rt_rq->overloaded = 0;
7668 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7672 rt_rq->rt_throttled = 0;
7673 rt_rq->rt_runtime = 0;
7674 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7676 #ifdef CONFIG_RT_GROUP_SCHED
7677 rt_rq->rt_nr_boosted = 0;
7682 #ifdef CONFIG_FAIR_GROUP_SCHED
7683 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7684 struct sched_entity *se, int cpu, int add,
7685 struct sched_entity *parent)
7687 struct rq *rq = cpu_rq(cpu);
7688 tg->cfs_rq[cpu] = cfs_rq;
7689 init_cfs_rq(cfs_rq, rq);
7692 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7695 /* se could be NULL for init_task_group */
7700 se->cfs_rq = &rq->cfs;
7702 se->cfs_rq = parent->my_q;
7705 se->load.weight = tg->shares;
7706 se->load.inv_weight = 0;
7707 se->parent = parent;
7711 #ifdef CONFIG_RT_GROUP_SCHED
7712 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7713 struct sched_rt_entity *rt_se, int cpu, int add,
7714 struct sched_rt_entity *parent)
7716 struct rq *rq = cpu_rq(cpu);
7718 tg->rt_rq[cpu] = rt_rq;
7719 init_rt_rq(rt_rq, rq);
7721 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7723 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7725 tg->rt_se[cpu] = rt_se;
7730 rt_se->rt_rq = &rq->rt;
7732 rt_se->rt_rq = parent->my_q;
7734 rt_se->my_q = rt_rq;
7735 rt_se->parent = parent;
7736 INIT_LIST_HEAD(&rt_se->run_list);
7740 void __init sched_init(void)
7743 unsigned long alloc_size = 0, ptr;
7745 #ifdef CONFIG_FAIR_GROUP_SCHED
7746 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7748 #ifdef CONFIG_RT_GROUP_SCHED
7749 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7751 #ifdef CONFIG_CPUMASK_OFFSTACK
7752 alloc_size += num_possible_cpus() * cpumask_size();
7755 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7757 #ifdef CONFIG_FAIR_GROUP_SCHED
7758 init_task_group.se = (struct sched_entity **)ptr;
7759 ptr += nr_cpu_ids * sizeof(void **);
7761 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7762 ptr += nr_cpu_ids * sizeof(void **);
7764 #endif /* CONFIG_FAIR_GROUP_SCHED */
7765 #ifdef CONFIG_RT_GROUP_SCHED
7766 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7767 ptr += nr_cpu_ids * sizeof(void **);
7769 init_task_group.rt_rq = (struct rt_rq **)ptr;
7770 ptr += nr_cpu_ids * sizeof(void **);
7772 #endif /* CONFIG_RT_GROUP_SCHED */
7773 #ifdef CONFIG_CPUMASK_OFFSTACK
7774 for_each_possible_cpu(i) {
7775 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7776 ptr += cpumask_size();
7778 #endif /* CONFIG_CPUMASK_OFFSTACK */
7782 init_defrootdomain();
7785 init_rt_bandwidth(&def_rt_bandwidth,
7786 global_rt_period(), global_rt_runtime());
7788 #ifdef CONFIG_RT_GROUP_SCHED
7789 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7790 global_rt_period(), global_rt_runtime());
7791 #endif /* CONFIG_RT_GROUP_SCHED */
7793 #ifdef CONFIG_CGROUP_SCHED
7794 list_add(&init_task_group.list, &task_groups);
7795 INIT_LIST_HEAD(&init_task_group.children);
7797 #endif /* CONFIG_CGROUP_SCHED */
7799 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7800 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7801 __alignof__(unsigned long));
7803 for_each_possible_cpu(i) {
7807 raw_spin_lock_init(&rq->lock);
7809 rq->calc_load_active = 0;
7810 rq->calc_load_update = jiffies + LOAD_FREQ;
7811 init_cfs_rq(&rq->cfs, rq);
7812 init_rt_rq(&rq->rt, rq);
7813 #ifdef CONFIG_FAIR_GROUP_SCHED
7814 init_task_group.shares = init_task_group_load;
7815 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7816 #ifdef CONFIG_CGROUP_SCHED
7818 * How much cpu bandwidth does init_task_group get?
7820 * In case of task-groups formed thr' the cgroup filesystem, it
7821 * gets 100% of the cpu resources in the system. This overall
7822 * system cpu resource is divided among the tasks of
7823 * init_task_group and its child task-groups in a fair manner,
7824 * based on each entity's (task or task-group's) weight
7825 * (se->load.weight).
7827 * In other words, if init_task_group has 10 tasks of weight
7828 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7829 * then A0's share of the cpu resource is:
7831 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7833 * We achieve this by letting init_task_group's tasks sit
7834 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7836 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7838 #endif /* CONFIG_FAIR_GROUP_SCHED */
7840 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7841 #ifdef CONFIG_RT_GROUP_SCHED
7842 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7843 #ifdef CONFIG_CGROUP_SCHED
7844 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7848 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7849 rq->cpu_load[j] = 0;
7853 rq->cpu_power = SCHED_LOAD_SCALE;
7854 rq->post_schedule = 0;
7855 rq->active_balance = 0;
7856 rq->next_balance = jiffies;
7861 rq->avg_idle = 2*sysctl_sched_migration_cost;
7862 rq_attach_root(rq, &def_root_domain);
7865 atomic_set(&rq->nr_iowait, 0);
7868 set_load_weight(&init_task);
7870 #ifdef CONFIG_PREEMPT_NOTIFIERS
7871 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7875 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7878 #ifdef CONFIG_RT_MUTEXES
7879 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7883 * The boot idle thread does lazy MMU switching as well:
7885 atomic_inc(&init_mm.mm_count);
7886 enter_lazy_tlb(&init_mm, current);
7889 * Make us the idle thread. Technically, schedule() should not be
7890 * called from this thread, however somewhere below it might be,
7891 * but because we are the idle thread, we just pick up running again
7892 * when this runqueue becomes "idle".
7894 init_idle(current, smp_processor_id());
7896 calc_load_update = jiffies + LOAD_FREQ;
7899 * During early bootup we pretend to be a normal task:
7901 current->sched_class = &fair_sched_class;
7903 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7904 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7907 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
7908 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
7910 /* May be allocated at isolcpus cmdline parse time */
7911 if (cpu_isolated_map == NULL)
7912 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7917 scheduler_running = 1;
7920 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7921 static inline int preempt_count_equals(int preempt_offset)
7923 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7925 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7928 void __might_sleep(const char *file, int line, int preempt_offset)
7931 static unsigned long prev_jiffy; /* ratelimiting */
7933 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7934 system_state != SYSTEM_RUNNING || oops_in_progress)
7936 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7938 prev_jiffy = jiffies;
7941 "BUG: sleeping function called from invalid context at %s:%d\n",
7944 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7945 in_atomic(), irqs_disabled(),
7946 current->pid, current->comm);
7948 debug_show_held_locks(current);
7949 if (irqs_disabled())
7950 print_irqtrace_events(current);
7954 EXPORT_SYMBOL(__might_sleep);
7957 #ifdef CONFIG_MAGIC_SYSRQ
7958 static void normalize_task(struct rq *rq, struct task_struct *p)
7962 on_rq = p->se.on_rq;
7964 deactivate_task(rq, p, 0);
7965 __setscheduler(rq, p, SCHED_NORMAL, 0);
7967 activate_task(rq, p, 0);
7968 resched_task(rq->curr);
7972 void normalize_rt_tasks(void)
7974 struct task_struct *g, *p;
7975 unsigned long flags;
7978 read_lock_irqsave(&tasklist_lock, flags);
7979 do_each_thread(g, p) {
7981 * Only normalize user tasks:
7986 p->se.exec_start = 0;
7987 #ifdef CONFIG_SCHEDSTATS
7988 p->se.statistics.wait_start = 0;
7989 p->se.statistics.sleep_start = 0;
7990 p->se.statistics.block_start = 0;
7995 * Renice negative nice level userspace
7998 if (TASK_NICE(p) < 0 && p->mm)
7999 set_user_nice(p, 0);
8003 raw_spin_lock(&p->pi_lock);
8004 rq = __task_rq_lock(p);
8006 normalize_task(rq, p);
8008 __task_rq_unlock(rq);
8009 raw_spin_unlock(&p->pi_lock);
8010 } while_each_thread(g, p);
8012 read_unlock_irqrestore(&tasklist_lock, flags);
8015 #endif /* CONFIG_MAGIC_SYSRQ */
8017 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8019 * These functions are only useful for the IA64 MCA handling, or kdb.
8021 * They can only be called when the whole system has been
8022 * stopped - every CPU needs to be quiescent, and no scheduling
8023 * activity can take place. Using them for anything else would
8024 * be a serious bug, and as a result, they aren't even visible
8025 * under any other configuration.
8029 * curr_task - return the current task for a given cpu.
8030 * @cpu: the processor in question.
8032 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8034 struct task_struct *curr_task(int cpu)
8036 return cpu_curr(cpu);
8039 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8043 * set_curr_task - set the current task for a given cpu.
8044 * @cpu: the processor in question.
8045 * @p: the task pointer to set.
8047 * Description: This function must only be used when non-maskable interrupts
8048 * are serviced on a separate stack. It allows the architecture to switch the
8049 * notion of the current task on a cpu in a non-blocking manner. This function
8050 * must be called with all CPU's synchronized, and interrupts disabled, the
8051 * and caller must save the original value of the current task (see
8052 * curr_task() above) and restore that value before reenabling interrupts and
8053 * re-starting the system.
8055 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8057 void set_curr_task(int cpu, struct task_struct *p)
8064 #ifdef CONFIG_FAIR_GROUP_SCHED
8065 static void free_fair_sched_group(struct task_group *tg)
8069 for_each_possible_cpu(i) {
8071 kfree(tg->cfs_rq[i]);
8081 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8083 struct cfs_rq *cfs_rq;
8084 struct sched_entity *se;
8088 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8091 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8095 tg->shares = NICE_0_LOAD;
8097 for_each_possible_cpu(i) {
8100 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8101 GFP_KERNEL, cpu_to_node(i));
8105 se = kzalloc_node(sizeof(struct sched_entity),
8106 GFP_KERNEL, cpu_to_node(i));
8110 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8121 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8123 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8124 &cpu_rq(cpu)->leaf_cfs_rq_list);
8127 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8129 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8131 #else /* !CONFG_FAIR_GROUP_SCHED */
8132 static inline void free_fair_sched_group(struct task_group *tg)
8137 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8142 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8146 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8149 #endif /* CONFIG_FAIR_GROUP_SCHED */
8151 #ifdef CONFIG_RT_GROUP_SCHED
8152 static void free_rt_sched_group(struct task_group *tg)
8156 destroy_rt_bandwidth(&tg->rt_bandwidth);
8158 for_each_possible_cpu(i) {
8160 kfree(tg->rt_rq[i]);
8162 kfree(tg->rt_se[i]);
8170 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8172 struct rt_rq *rt_rq;
8173 struct sched_rt_entity *rt_se;
8177 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8180 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8184 init_rt_bandwidth(&tg->rt_bandwidth,
8185 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8187 for_each_possible_cpu(i) {
8190 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8191 GFP_KERNEL, cpu_to_node(i));
8195 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8196 GFP_KERNEL, cpu_to_node(i));
8200 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8211 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8213 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8214 &cpu_rq(cpu)->leaf_rt_rq_list);
8217 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8219 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8221 #else /* !CONFIG_RT_GROUP_SCHED */
8222 static inline void free_rt_sched_group(struct task_group *tg)
8227 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8232 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8236 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8239 #endif /* CONFIG_RT_GROUP_SCHED */
8241 #ifdef CONFIG_CGROUP_SCHED
8242 static void free_sched_group(struct task_group *tg)
8244 free_fair_sched_group(tg);
8245 free_rt_sched_group(tg);
8249 /* allocate runqueue etc for a new task group */
8250 struct task_group *sched_create_group(struct task_group *parent)
8252 struct task_group *tg;
8253 unsigned long flags;
8256 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8258 return ERR_PTR(-ENOMEM);
8260 if (!alloc_fair_sched_group(tg, parent))
8263 if (!alloc_rt_sched_group(tg, parent))
8266 spin_lock_irqsave(&task_group_lock, flags);
8267 for_each_possible_cpu(i) {
8268 register_fair_sched_group(tg, i);
8269 register_rt_sched_group(tg, i);
8271 list_add_rcu(&tg->list, &task_groups);
8273 WARN_ON(!parent); /* root should already exist */
8275 tg->parent = parent;
8276 INIT_LIST_HEAD(&tg->children);
8277 list_add_rcu(&tg->siblings, &parent->children);
8278 spin_unlock_irqrestore(&task_group_lock, flags);
8283 free_sched_group(tg);
8284 return ERR_PTR(-ENOMEM);
8287 /* rcu callback to free various structures associated with a task group */
8288 static void free_sched_group_rcu(struct rcu_head *rhp)
8290 /* now it should be safe to free those cfs_rqs */
8291 free_sched_group(container_of(rhp, struct task_group, rcu));
8294 /* Destroy runqueue etc associated with a task group */
8295 void sched_destroy_group(struct task_group *tg)
8297 unsigned long flags;
8300 spin_lock_irqsave(&task_group_lock, flags);
8301 for_each_possible_cpu(i) {
8302 unregister_fair_sched_group(tg, i);
8303 unregister_rt_sched_group(tg, i);
8305 list_del_rcu(&tg->list);
8306 list_del_rcu(&tg->siblings);
8307 spin_unlock_irqrestore(&task_group_lock, flags);
8309 /* wait for possible concurrent references to cfs_rqs complete */
8310 call_rcu(&tg->rcu, free_sched_group_rcu);
8313 /* change task's runqueue when it moves between groups.
8314 * The caller of this function should have put the task in its new group
8315 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8316 * reflect its new group.
8318 void sched_move_task(struct task_struct *tsk)
8321 unsigned long flags;
8324 rq = task_rq_lock(tsk, &flags);
8326 running = task_current(rq, tsk);
8327 on_rq = tsk->se.on_rq;
8330 dequeue_task(rq, tsk, 0);
8331 if (unlikely(running))
8332 tsk->sched_class->put_prev_task(rq, tsk);
8334 set_task_rq(tsk, task_cpu(tsk));
8336 #ifdef CONFIG_FAIR_GROUP_SCHED
8337 if (tsk->sched_class->moved_group)
8338 tsk->sched_class->moved_group(tsk, on_rq);
8341 if (unlikely(running))
8342 tsk->sched_class->set_curr_task(rq);
8344 enqueue_task(rq, tsk, 0);
8346 task_rq_unlock(rq, &flags);
8348 #endif /* CONFIG_CGROUP_SCHED */
8350 #ifdef CONFIG_FAIR_GROUP_SCHED
8351 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8353 struct cfs_rq *cfs_rq = se->cfs_rq;
8358 dequeue_entity(cfs_rq, se, 0);
8360 se->load.weight = shares;
8361 se->load.inv_weight = 0;
8364 enqueue_entity(cfs_rq, se, 0);
8367 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8369 struct cfs_rq *cfs_rq = se->cfs_rq;
8370 struct rq *rq = cfs_rq->rq;
8371 unsigned long flags;
8373 raw_spin_lock_irqsave(&rq->lock, flags);
8374 __set_se_shares(se, shares);
8375 raw_spin_unlock_irqrestore(&rq->lock, flags);
8378 static DEFINE_MUTEX(shares_mutex);
8380 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8383 unsigned long flags;
8386 * We can't change the weight of the root cgroup.
8391 if (shares < MIN_SHARES)
8392 shares = MIN_SHARES;
8393 else if (shares > MAX_SHARES)
8394 shares = MAX_SHARES;
8396 mutex_lock(&shares_mutex);
8397 if (tg->shares == shares)
8400 spin_lock_irqsave(&task_group_lock, flags);
8401 for_each_possible_cpu(i)
8402 unregister_fair_sched_group(tg, i);
8403 list_del_rcu(&tg->siblings);
8404 spin_unlock_irqrestore(&task_group_lock, flags);
8406 /* wait for any ongoing reference to this group to finish */
8407 synchronize_sched();
8410 * Now we are free to modify the group's share on each cpu
8411 * w/o tripping rebalance_share or load_balance_fair.
8413 tg->shares = shares;
8414 for_each_possible_cpu(i) {
8418 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8419 set_se_shares(tg->se[i], shares);
8423 * Enable load balance activity on this group, by inserting it back on
8424 * each cpu's rq->leaf_cfs_rq_list.
8426 spin_lock_irqsave(&task_group_lock, flags);
8427 for_each_possible_cpu(i)
8428 register_fair_sched_group(tg, i);
8429 list_add_rcu(&tg->siblings, &tg->parent->children);
8430 spin_unlock_irqrestore(&task_group_lock, flags);
8432 mutex_unlock(&shares_mutex);
8436 unsigned long sched_group_shares(struct task_group *tg)
8442 #ifdef CONFIG_RT_GROUP_SCHED
8444 * Ensure that the real time constraints are schedulable.
8446 static DEFINE_MUTEX(rt_constraints_mutex);
8448 static unsigned long to_ratio(u64 period, u64 runtime)
8450 if (runtime == RUNTIME_INF)
8453 return div64_u64(runtime << 20, period);
8456 /* Must be called with tasklist_lock held */
8457 static inline int tg_has_rt_tasks(struct task_group *tg)
8459 struct task_struct *g, *p;
8461 do_each_thread(g, p) {
8462 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8464 } while_each_thread(g, p);
8469 struct rt_schedulable_data {
8470 struct task_group *tg;
8475 static int tg_schedulable(struct task_group *tg, void *data)
8477 struct rt_schedulable_data *d = data;
8478 struct task_group *child;
8479 unsigned long total, sum = 0;
8480 u64 period, runtime;
8482 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8483 runtime = tg->rt_bandwidth.rt_runtime;
8486 period = d->rt_period;
8487 runtime = d->rt_runtime;
8491 * Cannot have more runtime than the period.
8493 if (runtime > period && runtime != RUNTIME_INF)
8497 * Ensure we don't starve existing RT tasks.
8499 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8502 total = to_ratio(period, runtime);
8505 * Nobody can have more than the global setting allows.
8507 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8511 * The sum of our children's runtime should not exceed our own.
8513 list_for_each_entry_rcu(child, &tg->children, siblings) {
8514 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8515 runtime = child->rt_bandwidth.rt_runtime;
8517 if (child == d->tg) {
8518 period = d->rt_period;
8519 runtime = d->rt_runtime;
8522 sum += to_ratio(period, runtime);
8531 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8533 struct rt_schedulable_data data = {
8535 .rt_period = period,
8536 .rt_runtime = runtime,
8539 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8542 static int tg_set_bandwidth(struct task_group *tg,
8543 u64 rt_period, u64 rt_runtime)
8547 mutex_lock(&rt_constraints_mutex);
8548 read_lock(&tasklist_lock);
8549 err = __rt_schedulable(tg, rt_period, rt_runtime);
8553 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8554 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8555 tg->rt_bandwidth.rt_runtime = rt_runtime;
8557 for_each_possible_cpu(i) {
8558 struct rt_rq *rt_rq = tg->rt_rq[i];
8560 raw_spin_lock(&rt_rq->rt_runtime_lock);
8561 rt_rq->rt_runtime = rt_runtime;
8562 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8564 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8566 read_unlock(&tasklist_lock);
8567 mutex_unlock(&rt_constraints_mutex);
8572 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8574 u64 rt_runtime, rt_period;
8576 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8577 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8578 if (rt_runtime_us < 0)
8579 rt_runtime = RUNTIME_INF;
8581 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8584 long sched_group_rt_runtime(struct task_group *tg)
8588 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8591 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8592 do_div(rt_runtime_us, NSEC_PER_USEC);
8593 return rt_runtime_us;
8596 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8598 u64 rt_runtime, rt_period;
8600 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8601 rt_runtime = tg->rt_bandwidth.rt_runtime;
8606 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8609 long sched_group_rt_period(struct task_group *tg)
8613 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8614 do_div(rt_period_us, NSEC_PER_USEC);
8615 return rt_period_us;
8618 static int sched_rt_global_constraints(void)
8620 u64 runtime, period;
8623 if (sysctl_sched_rt_period <= 0)
8626 runtime = global_rt_runtime();
8627 period = global_rt_period();
8630 * Sanity check on the sysctl variables.
8632 if (runtime > period && runtime != RUNTIME_INF)
8635 mutex_lock(&rt_constraints_mutex);
8636 read_lock(&tasklist_lock);
8637 ret = __rt_schedulable(NULL, 0, 0);
8638 read_unlock(&tasklist_lock);
8639 mutex_unlock(&rt_constraints_mutex);
8644 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8646 /* Don't accept realtime tasks when there is no way for them to run */
8647 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8653 #else /* !CONFIG_RT_GROUP_SCHED */
8654 static int sched_rt_global_constraints(void)
8656 unsigned long flags;
8659 if (sysctl_sched_rt_period <= 0)
8663 * There's always some RT tasks in the root group
8664 * -- migration, kstopmachine etc..
8666 if (sysctl_sched_rt_runtime == 0)
8669 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8670 for_each_possible_cpu(i) {
8671 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8673 raw_spin_lock(&rt_rq->rt_runtime_lock);
8674 rt_rq->rt_runtime = global_rt_runtime();
8675 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8677 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8681 #endif /* CONFIG_RT_GROUP_SCHED */
8683 int sched_rt_handler(struct ctl_table *table, int write,
8684 void __user *buffer, size_t *lenp,
8688 int old_period, old_runtime;
8689 static DEFINE_MUTEX(mutex);
8692 old_period = sysctl_sched_rt_period;
8693 old_runtime = sysctl_sched_rt_runtime;
8695 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8697 if (!ret && write) {
8698 ret = sched_rt_global_constraints();
8700 sysctl_sched_rt_period = old_period;
8701 sysctl_sched_rt_runtime = old_runtime;
8703 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8704 def_rt_bandwidth.rt_period =
8705 ns_to_ktime(global_rt_period());
8708 mutex_unlock(&mutex);
8713 #ifdef CONFIG_CGROUP_SCHED
8715 /* return corresponding task_group object of a cgroup */
8716 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8718 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8719 struct task_group, css);
8722 static struct cgroup_subsys_state *
8723 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8725 struct task_group *tg, *parent;
8727 if (!cgrp->parent) {
8728 /* This is early initialization for the top cgroup */
8729 return &init_task_group.css;
8732 parent = cgroup_tg(cgrp->parent);
8733 tg = sched_create_group(parent);
8735 return ERR_PTR(-ENOMEM);
8741 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8743 struct task_group *tg = cgroup_tg(cgrp);
8745 sched_destroy_group(tg);
8749 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8751 #ifdef CONFIG_RT_GROUP_SCHED
8752 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8755 /* We don't support RT-tasks being in separate groups */
8756 if (tsk->sched_class != &fair_sched_class)
8763 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8764 struct task_struct *tsk, bool threadgroup)
8766 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8770 struct task_struct *c;
8772 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8773 retval = cpu_cgroup_can_attach_task(cgrp, c);
8785 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8786 struct cgroup *old_cont, struct task_struct *tsk,
8789 sched_move_task(tsk);
8791 struct task_struct *c;
8793 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8800 #ifdef CONFIG_FAIR_GROUP_SCHED
8801 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8804 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8807 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8809 struct task_group *tg = cgroup_tg(cgrp);
8811 return (u64) tg->shares;
8813 #endif /* CONFIG_FAIR_GROUP_SCHED */
8815 #ifdef CONFIG_RT_GROUP_SCHED
8816 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8819 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8822 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8824 return sched_group_rt_runtime(cgroup_tg(cgrp));
8827 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8830 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8833 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8835 return sched_group_rt_period(cgroup_tg(cgrp));
8837 #endif /* CONFIG_RT_GROUP_SCHED */
8839 static struct cftype cpu_files[] = {
8840 #ifdef CONFIG_FAIR_GROUP_SCHED
8843 .read_u64 = cpu_shares_read_u64,
8844 .write_u64 = cpu_shares_write_u64,
8847 #ifdef CONFIG_RT_GROUP_SCHED
8849 .name = "rt_runtime_us",
8850 .read_s64 = cpu_rt_runtime_read,
8851 .write_s64 = cpu_rt_runtime_write,
8854 .name = "rt_period_us",
8855 .read_u64 = cpu_rt_period_read_uint,
8856 .write_u64 = cpu_rt_period_write_uint,
8861 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8863 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8866 struct cgroup_subsys cpu_cgroup_subsys = {
8868 .create = cpu_cgroup_create,
8869 .destroy = cpu_cgroup_destroy,
8870 .can_attach = cpu_cgroup_can_attach,
8871 .attach = cpu_cgroup_attach,
8872 .populate = cpu_cgroup_populate,
8873 .subsys_id = cpu_cgroup_subsys_id,
8877 #endif /* CONFIG_CGROUP_SCHED */
8879 #ifdef CONFIG_CGROUP_CPUACCT
8882 * CPU accounting code for task groups.
8884 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8885 * (balbir@in.ibm.com).
8888 /* track cpu usage of a group of tasks and its child groups */
8890 struct cgroup_subsys_state css;
8891 /* cpuusage holds pointer to a u64-type object on every cpu */
8892 u64 __percpu *cpuusage;
8893 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8894 struct cpuacct *parent;
8897 struct cgroup_subsys cpuacct_subsys;
8899 /* return cpu accounting group corresponding to this container */
8900 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8902 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8903 struct cpuacct, css);
8906 /* return cpu accounting group to which this task belongs */
8907 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8909 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8910 struct cpuacct, css);
8913 /* create a new cpu accounting group */
8914 static struct cgroup_subsys_state *cpuacct_create(
8915 struct cgroup_subsys *ss, struct cgroup *cgrp)
8917 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8923 ca->cpuusage = alloc_percpu(u64);
8927 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8928 if (percpu_counter_init(&ca->cpustat[i], 0))
8929 goto out_free_counters;
8932 ca->parent = cgroup_ca(cgrp->parent);
8938 percpu_counter_destroy(&ca->cpustat[i]);
8939 free_percpu(ca->cpuusage);
8943 return ERR_PTR(-ENOMEM);
8946 /* destroy an existing cpu accounting group */
8948 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8950 struct cpuacct *ca = cgroup_ca(cgrp);
8953 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8954 percpu_counter_destroy(&ca->cpustat[i]);
8955 free_percpu(ca->cpuusage);
8959 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8961 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8964 #ifndef CONFIG_64BIT
8966 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8968 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8970 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8978 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8980 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8982 #ifndef CONFIG_64BIT
8984 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8986 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8988 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8994 /* return total cpu usage (in nanoseconds) of a group */
8995 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8997 struct cpuacct *ca = cgroup_ca(cgrp);
8998 u64 totalcpuusage = 0;
9001 for_each_present_cpu(i)
9002 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9004 return totalcpuusage;
9007 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9010 struct cpuacct *ca = cgroup_ca(cgrp);
9019 for_each_present_cpu(i)
9020 cpuacct_cpuusage_write(ca, i, 0);
9026 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9029 struct cpuacct *ca = cgroup_ca(cgroup);
9033 for_each_present_cpu(i) {
9034 percpu = cpuacct_cpuusage_read(ca, i);
9035 seq_printf(m, "%llu ", (unsigned long long) percpu);
9037 seq_printf(m, "\n");
9041 static const char *cpuacct_stat_desc[] = {
9042 [CPUACCT_STAT_USER] = "user",
9043 [CPUACCT_STAT_SYSTEM] = "system",
9046 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9047 struct cgroup_map_cb *cb)
9049 struct cpuacct *ca = cgroup_ca(cgrp);
9052 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9053 s64 val = percpu_counter_read(&ca->cpustat[i]);
9054 val = cputime64_to_clock_t(val);
9055 cb->fill(cb, cpuacct_stat_desc[i], val);
9060 static struct cftype files[] = {
9063 .read_u64 = cpuusage_read,
9064 .write_u64 = cpuusage_write,
9067 .name = "usage_percpu",
9068 .read_seq_string = cpuacct_percpu_seq_read,
9072 .read_map = cpuacct_stats_show,
9076 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9078 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9082 * charge this task's execution time to its accounting group.
9084 * called with rq->lock held.
9086 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9091 if (unlikely(!cpuacct_subsys.active))
9094 cpu = task_cpu(tsk);
9100 for (; ca; ca = ca->parent) {
9101 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9102 *cpuusage += cputime;
9109 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9110 * in cputime_t units. As a result, cpuacct_update_stats calls
9111 * percpu_counter_add with values large enough to always overflow the
9112 * per cpu batch limit causing bad SMP scalability.
9114 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9115 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9116 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9119 #define CPUACCT_BATCH \
9120 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9122 #define CPUACCT_BATCH 0
9126 * Charge the system/user time to the task's accounting group.
9128 static void cpuacct_update_stats(struct task_struct *tsk,
9129 enum cpuacct_stat_index idx, cputime_t val)
9132 int batch = CPUACCT_BATCH;
9134 if (unlikely(!cpuacct_subsys.active))
9141 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9147 struct cgroup_subsys cpuacct_subsys = {
9149 .create = cpuacct_create,
9150 .destroy = cpuacct_destroy,
9151 .populate = cpuacct_populate,
9152 .subsys_id = cpuacct_subsys_id,
9154 #endif /* CONFIG_CGROUP_CPUACCT */
9158 void synchronize_sched_expedited(void)
9162 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9164 #else /* #ifndef CONFIG_SMP */
9166 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
9168 static int synchronize_sched_expedited_cpu_stop(void *data)
9171 * There must be a full memory barrier on each affected CPU
9172 * between the time that try_stop_cpus() is called and the
9173 * time that it returns.
9175 * In the current initial implementation of cpu_stop, the
9176 * above condition is already met when the control reaches
9177 * this point and the following smp_mb() is not strictly
9178 * necessary. Do smp_mb() anyway for documentation and
9179 * robustness against future implementation changes.
9181 smp_mb(); /* See above comment block. */
9186 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9187 * approach to force grace period to end quickly. This consumes
9188 * significant time on all CPUs, and is thus not recommended for
9189 * any sort of common-case code.
9191 * Note that it is illegal to call this function while holding any
9192 * lock that is acquired by a CPU-hotplug notifier. Failing to
9193 * observe this restriction will result in deadlock.
9195 void synchronize_sched_expedited(void)
9197 int snap, trycount = 0;
9199 smp_mb(); /* ensure prior mod happens before capturing snap. */
9200 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
9202 while (try_stop_cpus(cpu_online_mask,
9203 synchronize_sched_expedited_cpu_stop,
9206 if (trycount++ < 10)
9207 udelay(trycount * num_online_cpus());
9209 synchronize_sched();
9212 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
9213 smp_mb(); /* ensure test happens before caller kfree */
9218 atomic_inc(&synchronize_sched_expedited_count);
9219 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9222 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9224 #endif /* #else #ifndef CONFIG_SMP */