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"
80 #include "workqueue_sched.h"
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/sched.h>
86 * Convert user-nice values [ -20 ... 0 ... 19 ]
87 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
90 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
91 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
92 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
95 * 'User priority' is the nice value converted to something we
96 * can work with better when scaling various scheduler parameters,
97 * it's a [ 0 ... 39 ] range.
99 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
100 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
101 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
104 * Helpers for converting nanosecond timing to jiffy resolution
106 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
108 #define NICE_0_LOAD SCHED_LOAD_SCALE
109 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
112 * These are the 'tuning knobs' of the scheduler:
114 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
115 * Timeslices get refilled after they expire.
117 #define DEF_TIMESLICE (100 * HZ / 1000)
120 * single value that denotes runtime == period, ie unlimited time.
122 #define RUNTIME_INF ((u64)~0ULL)
124 static inline int rt_policy(int policy)
126 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
131 static inline int task_has_rt_policy(struct task_struct *p)
133 return rt_policy(p->policy);
137 * This is the priority-queue data structure of the RT scheduling class:
139 struct rt_prio_array {
140 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
141 struct list_head queue[MAX_RT_PRIO];
144 struct rt_bandwidth {
145 /* nests inside the rq lock: */
146 raw_spinlock_t rt_runtime_lock;
149 struct hrtimer rt_period_timer;
152 static struct rt_bandwidth def_rt_bandwidth;
154 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
156 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
158 struct rt_bandwidth *rt_b =
159 container_of(timer, struct rt_bandwidth, rt_period_timer);
165 now = hrtimer_cb_get_time(timer);
166 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
171 idle = do_sched_rt_period_timer(rt_b, overrun);
174 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
178 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
180 rt_b->rt_period = ns_to_ktime(period);
181 rt_b->rt_runtime = runtime;
183 raw_spin_lock_init(&rt_b->rt_runtime_lock);
185 hrtimer_init(&rt_b->rt_period_timer,
186 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
187 rt_b->rt_period_timer.function = sched_rt_period_timer;
190 static inline int rt_bandwidth_enabled(void)
192 return sysctl_sched_rt_runtime >= 0;
195 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
199 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
202 if (hrtimer_active(&rt_b->rt_period_timer))
205 raw_spin_lock(&rt_b->rt_runtime_lock);
210 if (hrtimer_active(&rt_b->rt_period_timer))
213 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
214 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
216 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
217 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
218 delta = ktime_to_ns(ktime_sub(hard, soft));
219 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
220 HRTIMER_MODE_ABS_PINNED, 0);
222 raw_spin_unlock(&rt_b->rt_runtime_lock);
225 #ifdef CONFIG_RT_GROUP_SCHED
226 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
228 hrtimer_cancel(&rt_b->rt_period_timer);
233 * sched_domains_mutex serializes calls to arch_init_sched_domains,
234 * detach_destroy_domains and partition_sched_domains.
236 static DEFINE_MUTEX(sched_domains_mutex);
238 #ifdef CONFIG_CGROUP_SCHED
240 #include <linux/cgroup.h>
244 static LIST_HEAD(task_groups);
246 /* task group related information */
248 struct cgroup_subsys_state css;
250 #ifdef CONFIG_FAIR_GROUP_SCHED
251 /* schedulable entities of this group on each cpu */
252 struct sched_entity **se;
253 /* runqueue "owned" by this group on each cpu */
254 struct cfs_rq **cfs_rq;
255 unsigned long shares;
257 atomic_t load_weight;
260 #ifdef CONFIG_RT_GROUP_SCHED
261 struct sched_rt_entity **rt_se;
262 struct rt_rq **rt_rq;
264 struct rt_bandwidth rt_bandwidth;
268 struct list_head list;
270 struct task_group *parent;
271 struct list_head siblings;
272 struct list_head children;
275 #define root_task_group init_task_group
277 /* task_group_lock serializes the addition/removal of task groups */
278 static DEFINE_SPINLOCK(task_group_lock);
280 #ifdef CONFIG_FAIR_GROUP_SCHED
283 static int root_task_group_empty(void)
285 return list_empty(&root_task_group.children);
289 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
292 * A weight of 0 or 1 can cause arithmetics problems.
293 * A weight of a cfs_rq is the sum of weights of which entities
294 * are queued on this cfs_rq, so a weight of a entity should not be
295 * too large, so as the shares value of a task group.
296 * (The default weight is 1024 - so there's no practical
297 * limitation from this.)
300 #define MAX_SHARES (1UL << 18)
302 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
305 /* Default task group.
306 * Every task in system belong to this group at bootup.
308 struct task_group init_task_group;
310 #endif /* CONFIG_CGROUP_SCHED */
312 /* CFS-related fields in a runqueue */
314 struct load_weight load;
315 unsigned long nr_running;
320 struct rb_root tasks_timeline;
321 struct rb_node *rb_leftmost;
323 struct list_head tasks;
324 struct list_head *balance_iterator;
327 * 'curr' points to currently running entity on this cfs_rq.
328 * It is set to NULL otherwise (i.e when none are currently running).
330 struct sched_entity *curr, *next, *last;
332 unsigned int nr_spread_over;
334 #ifdef CONFIG_FAIR_GROUP_SCHED
335 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
338 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
339 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
340 * (like users, containers etc.)
342 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
343 * list is used during load balance.
346 struct list_head leaf_cfs_rq_list;
347 struct task_group *tg; /* group that "owns" this runqueue */
351 * the part of load.weight contributed by tasks
353 unsigned long task_weight;
356 * h_load = weight * f(tg)
358 * Where f(tg) is the recursive weight fraction assigned to
361 unsigned long h_load;
367 unsigned long load_contribution;
372 /* Real-Time classes' related field in a runqueue: */
374 struct rt_prio_array active;
375 unsigned long rt_nr_running;
376 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
378 int curr; /* highest queued rt task prio */
380 int next; /* next highest */
385 unsigned long rt_nr_migratory;
386 unsigned long rt_nr_total;
388 struct plist_head pushable_tasks;
393 /* Nests inside the rq lock: */
394 raw_spinlock_t rt_runtime_lock;
396 #ifdef CONFIG_RT_GROUP_SCHED
397 unsigned long rt_nr_boosted;
400 struct list_head leaf_rt_rq_list;
401 struct task_group *tg;
408 * We add the notion of a root-domain which will be used to define per-domain
409 * variables. Each exclusive cpuset essentially defines an island domain by
410 * fully partitioning the member cpus from any other cpuset. Whenever a new
411 * exclusive cpuset is created, we also create and attach a new root-domain
418 cpumask_var_t online;
421 * The "RT overload" flag: it gets set if a CPU has more than
422 * one runnable RT task.
424 cpumask_var_t rto_mask;
426 struct cpupri cpupri;
430 * By default the system creates a single root-domain with all cpus as
431 * members (mimicking the global state we have today).
433 static struct root_domain def_root_domain;
435 #endif /* CONFIG_SMP */
438 * This is the main, per-CPU runqueue data structure.
440 * Locking rule: those places that want to lock multiple runqueues
441 * (such as the load balancing or the thread migration code), lock
442 * acquire operations must be ordered by ascending &runqueue.
449 * nr_running and cpu_load should be in the same cacheline because
450 * remote CPUs use both these fields when doing load calculation.
452 unsigned long nr_running;
453 #define CPU_LOAD_IDX_MAX 5
454 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
455 unsigned long last_load_update_tick;
458 unsigned char nohz_balance_kick;
460 unsigned int skip_clock_update;
462 /* capture load from *all* tasks on this cpu: */
463 struct load_weight load;
464 unsigned long nr_load_updates;
470 #ifdef CONFIG_FAIR_GROUP_SCHED
471 /* list of leaf cfs_rq on this cpu: */
472 struct list_head leaf_cfs_rq_list;
474 #ifdef CONFIG_RT_GROUP_SCHED
475 struct list_head leaf_rt_rq_list;
479 * This is part of a global counter where only the total sum
480 * over all CPUs matters. A task can increase this counter on
481 * one CPU and if it got migrated afterwards it may decrease
482 * it on another CPU. Always updated under the runqueue lock:
484 unsigned long nr_uninterruptible;
486 struct task_struct *curr, *idle, *stop;
487 unsigned long next_balance;
488 struct mm_struct *prev_mm;
496 struct root_domain *rd;
497 struct sched_domain *sd;
499 unsigned long cpu_power;
501 unsigned char idle_at_tick;
502 /* For active balancing */
506 struct cpu_stop_work active_balance_work;
507 /* cpu of this runqueue: */
511 unsigned long avg_load_per_task;
519 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
523 /* calc_load related fields */
524 unsigned long calc_load_update;
525 long calc_load_active;
527 #ifdef CONFIG_SCHED_HRTICK
529 int hrtick_csd_pending;
530 struct call_single_data hrtick_csd;
532 struct hrtimer hrtick_timer;
535 #ifdef CONFIG_SCHEDSTATS
537 struct sched_info rq_sched_info;
538 unsigned long long rq_cpu_time;
539 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
541 /* sys_sched_yield() stats */
542 unsigned int yld_count;
544 /* schedule() stats */
545 unsigned int sched_switch;
546 unsigned int sched_count;
547 unsigned int sched_goidle;
549 /* try_to_wake_up() stats */
550 unsigned int ttwu_count;
551 unsigned int ttwu_local;
554 unsigned int bkl_count;
558 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
561 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
563 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
566 * A queue event has occurred, and we're going to schedule. In
567 * this case, we can save a useless back to back clock update.
569 if (test_tsk_need_resched(p))
570 rq->skip_clock_update = 1;
573 static inline int cpu_of(struct rq *rq)
582 #define rcu_dereference_check_sched_domain(p) \
583 rcu_dereference_check((p), \
584 rcu_read_lock_sched_held() || \
585 lockdep_is_held(&sched_domains_mutex))
588 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
589 * See detach_destroy_domains: synchronize_sched for details.
591 * The domain tree of any CPU may only be accessed from within
592 * preempt-disabled sections.
594 #define for_each_domain(cpu, __sd) \
595 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
597 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
598 #define this_rq() (&__get_cpu_var(runqueues))
599 #define task_rq(p) cpu_rq(task_cpu(p))
600 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
601 #define raw_rq() (&__raw_get_cpu_var(runqueues))
603 #ifdef CONFIG_CGROUP_SCHED
606 * Return the group to which this tasks belongs.
608 * We use task_subsys_state_check() and extend the RCU verification
609 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
610 * holds that lock for each task it moves into the cgroup. Therefore
611 * by holding that lock, we pin the task to the current cgroup.
613 static inline struct task_group *task_group(struct task_struct *p)
615 struct cgroup_subsys_state *css;
617 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
618 lockdep_is_held(&task_rq(p)->lock));
619 return container_of(css, struct task_group, css);
622 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
623 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
625 #ifdef CONFIG_FAIR_GROUP_SCHED
626 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
627 p->se.parent = task_group(p)->se[cpu];
630 #ifdef CONFIG_RT_GROUP_SCHED
631 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
632 p->rt.parent = task_group(p)->rt_se[cpu];
636 #else /* CONFIG_CGROUP_SCHED */
638 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
639 static inline struct task_group *task_group(struct task_struct *p)
644 #endif /* CONFIG_CGROUP_SCHED */
646 static u64 irq_time_cpu(int cpu);
647 static void sched_irq_time_avg_update(struct rq *rq, u64 irq_time);
649 inline void update_rq_clock(struct rq *rq)
651 if (!rq->skip_clock_update) {
652 int cpu = cpu_of(rq);
655 rq->clock = sched_clock_cpu(cpu);
656 irq_time = irq_time_cpu(cpu);
657 if (rq->clock - irq_time > rq->clock_task)
658 rq->clock_task = rq->clock - irq_time;
660 sched_irq_time_avg_update(rq, irq_time);
665 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
667 #ifdef CONFIG_SCHED_DEBUG
668 # define const_debug __read_mostly
670 # define const_debug static const
675 * @cpu: the processor in question.
677 * Returns true if the current cpu runqueue is locked.
678 * This interface allows printk to be called with the runqueue lock
679 * held and know whether or not it is OK to wake up the klogd.
681 int runqueue_is_locked(int cpu)
683 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
687 * Debugging: various feature bits
690 #define SCHED_FEAT(name, enabled) \
691 __SCHED_FEAT_##name ,
694 #include "sched_features.h"
699 #define SCHED_FEAT(name, enabled) \
700 (1UL << __SCHED_FEAT_##name) * enabled |
702 const_debug unsigned int sysctl_sched_features =
703 #include "sched_features.h"
708 #ifdef CONFIG_SCHED_DEBUG
709 #define SCHED_FEAT(name, enabled) \
712 static __read_mostly char *sched_feat_names[] = {
713 #include "sched_features.h"
719 static int sched_feat_show(struct seq_file *m, void *v)
723 for (i = 0; sched_feat_names[i]; i++) {
724 if (!(sysctl_sched_features & (1UL << i)))
726 seq_printf(m, "%s ", sched_feat_names[i]);
734 sched_feat_write(struct file *filp, const char __user *ubuf,
735 size_t cnt, loff_t *ppos)
745 if (copy_from_user(&buf, ubuf, cnt))
751 if (strncmp(buf, "NO_", 3) == 0) {
756 for (i = 0; sched_feat_names[i]; i++) {
757 if (strcmp(cmp, sched_feat_names[i]) == 0) {
759 sysctl_sched_features &= ~(1UL << i);
761 sysctl_sched_features |= (1UL << i);
766 if (!sched_feat_names[i])
774 static int sched_feat_open(struct inode *inode, struct file *filp)
776 return single_open(filp, sched_feat_show, NULL);
779 static const struct file_operations sched_feat_fops = {
780 .open = sched_feat_open,
781 .write = sched_feat_write,
784 .release = single_release,
787 static __init int sched_init_debug(void)
789 debugfs_create_file("sched_features", 0644, NULL, NULL,
794 late_initcall(sched_init_debug);
798 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
801 * Number of tasks to iterate in a single balance run.
802 * Limited because this is done with IRQs disabled.
804 const_debug unsigned int sysctl_sched_nr_migrate = 32;
807 * period over which we average the RT time consumption, measured
812 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
815 * period over which we measure -rt task cpu usage in us.
818 unsigned int sysctl_sched_rt_period = 1000000;
820 static __read_mostly int scheduler_running;
823 * part of the period that we allow rt tasks to run in us.
826 int sysctl_sched_rt_runtime = 950000;
828 static inline u64 global_rt_period(void)
830 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
833 static inline u64 global_rt_runtime(void)
835 if (sysctl_sched_rt_runtime < 0)
838 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
841 #ifndef prepare_arch_switch
842 # define prepare_arch_switch(next) do { } while (0)
844 #ifndef finish_arch_switch
845 # define finish_arch_switch(prev) do { } while (0)
848 static inline int task_current(struct rq *rq, struct task_struct *p)
850 return rq->curr == p;
853 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
854 static inline int task_running(struct rq *rq, struct task_struct *p)
856 return task_current(rq, p);
859 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
863 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
865 #ifdef CONFIG_DEBUG_SPINLOCK
866 /* this is a valid case when another task releases the spinlock */
867 rq->lock.owner = current;
870 * If we are tracking spinlock dependencies then we have to
871 * fix up the runqueue lock - which gets 'carried over' from
874 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
876 raw_spin_unlock_irq(&rq->lock);
879 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
880 static inline int task_running(struct rq *rq, struct task_struct *p)
885 return task_current(rq, p);
889 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
893 * We can optimise this out completely for !SMP, because the
894 * SMP rebalancing from interrupt is the only thing that cares
899 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
900 raw_spin_unlock_irq(&rq->lock);
902 raw_spin_unlock(&rq->lock);
906 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
910 * After ->oncpu is cleared, the task can be moved to a different CPU.
911 * We must ensure this doesn't happen until the switch is completely
917 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
921 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
924 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
927 static inline int task_is_waking(struct task_struct *p)
929 return unlikely(p->state == TASK_WAKING);
933 * __task_rq_lock - lock the runqueue a given task resides on.
934 * Must be called interrupts disabled.
936 static inline struct rq *__task_rq_lock(struct task_struct *p)
943 raw_spin_lock(&rq->lock);
944 if (likely(rq == task_rq(p)))
946 raw_spin_unlock(&rq->lock);
951 * task_rq_lock - lock the runqueue a given task resides on and disable
952 * interrupts. Note the ordering: we can safely lookup the task_rq without
953 * explicitly disabling preemption.
955 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
961 local_irq_save(*flags);
963 raw_spin_lock(&rq->lock);
964 if (likely(rq == task_rq(p)))
966 raw_spin_unlock_irqrestore(&rq->lock, *flags);
970 static void __task_rq_unlock(struct rq *rq)
973 raw_spin_unlock(&rq->lock);
976 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
979 raw_spin_unlock_irqrestore(&rq->lock, *flags);
983 * this_rq_lock - lock this runqueue and disable interrupts.
985 static struct rq *this_rq_lock(void)
992 raw_spin_lock(&rq->lock);
997 #ifdef CONFIG_SCHED_HRTICK
999 * Use HR-timers to deliver accurate preemption points.
1001 * Its all a bit involved since we cannot program an hrt while holding the
1002 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1005 * When we get rescheduled we reprogram the hrtick_timer outside of the
1011 * - enabled by features
1012 * - hrtimer is actually high res
1014 static inline int hrtick_enabled(struct rq *rq)
1016 if (!sched_feat(HRTICK))
1018 if (!cpu_active(cpu_of(rq)))
1020 return hrtimer_is_hres_active(&rq->hrtick_timer);
1023 static void hrtick_clear(struct rq *rq)
1025 if (hrtimer_active(&rq->hrtick_timer))
1026 hrtimer_cancel(&rq->hrtick_timer);
1030 * High-resolution timer tick.
1031 * Runs from hardirq context with interrupts disabled.
1033 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1035 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1037 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1039 raw_spin_lock(&rq->lock);
1040 update_rq_clock(rq);
1041 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1042 raw_spin_unlock(&rq->lock);
1044 return HRTIMER_NORESTART;
1049 * called from hardirq (IPI) context
1051 static void __hrtick_start(void *arg)
1053 struct rq *rq = arg;
1055 raw_spin_lock(&rq->lock);
1056 hrtimer_restart(&rq->hrtick_timer);
1057 rq->hrtick_csd_pending = 0;
1058 raw_spin_unlock(&rq->lock);
1062 * Called to set the hrtick timer state.
1064 * called with rq->lock held and irqs disabled
1066 static void hrtick_start(struct rq *rq, u64 delay)
1068 struct hrtimer *timer = &rq->hrtick_timer;
1069 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1071 hrtimer_set_expires(timer, time);
1073 if (rq == this_rq()) {
1074 hrtimer_restart(timer);
1075 } else if (!rq->hrtick_csd_pending) {
1076 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1077 rq->hrtick_csd_pending = 1;
1082 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1084 int cpu = (int)(long)hcpu;
1087 case CPU_UP_CANCELED:
1088 case CPU_UP_CANCELED_FROZEN:
1089 case CPU_DOWN_PREPARE:
1090 case CPU_DOWN_PREPARE_FROZEN:
1092 case CPU_DEAD_FROZEN:
1093 hrtick_clear(cpu_rq(cpu));
1100 static __init void init_hrtick(void)
1102 hotcpu_notifier(hotplug_hrtick, 0);
1106 * Called to set the hrtick timer state.
1108 * called with rq->lock held and irqs disabled
1110 static void hrtick_start(struct rq *rq, u64 delay)
1112 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1113 HRTIMER_MODE_REL_PINNED, 0);
1116 static inline void init_hrtick(void)
1119 #endif /* CONFIG_SMP */
1121 static void init_rq_hrtick(struct rq *rq)
1124 rq->hrtick_csd_pending = 0;
1126 rq->hrtick_csd.flags = 0;
1127 rq->hrtick_csd.func = __hrtick_start;
1128 rq->hrtick_csd.info = rq;
1131 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1132 rq->hrtick_timer.function = hrtick;
1134 #else /* CONFIG_SCHED_HRTICK */
1135 static inline void hrtick_clear(struct rq *rq)
1139 static inline void init_rq_hrtick(struct rq *rq)
1143 static inline void init_hrtick(void)
1146 #endif /* CONFIG_SCHED_HRTICK */
1149 * resched_task - mark a task 'to be rescheduled now'.
1151 * On UP this means the setting of the need_resched flag, on SMP it
1152 * might also involve a cross-CPU call to trigger the scheduler on
1157 #ifndef tsk_is_polling
1158 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1161 static void resched_task(struct task_struct *p)
1165 assert_raw_spin_locked(&task_rq(p)->lock);
1167 if (test_tsk_need_resched(p))
1170 set_tsk_need_resched(p);
1173 if (cpu == smp_processor_id())
1176 /* NEED_RESCHED must be visible before we test polling */
1178 if (!tsk_is_polling(p))
1179 smp_send_reschedule(cpu);
1182 static void resched_cpu(int cpu)
1184 struct rq *rq = cpu_rq(cpu);
1185 unsigned long flags;
1187 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1189 resched_task(cpu_curr(cpu));
1190 raw_spin_unlock_irqrestore(&rq->lock, flags);
1195 * In the semi idle case, use the nearest busy cpu for migrating timers
1196 * from an idle cpu. This is good for power-savings.
1198 * We don't do similar optimization for completely idle system, as
1199 * selecting an idle cpu will add more delays to the timers than intended
1200 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1202 int get_nohz_timer_target(void)
1204 int cpu = smp_processor_id();
1206 struct sched_domain *sd;
1208 for_each_domain(cpu, sd) {
1209 for_each_cpu(i, sched_domain_span(sd))
1216 * When add_timer_on() enqueues a timer into the timer wheel of an
1217 * idle CPU then this timer might expire before the next timer event
1218 * which is scheduled to wake up that CPU. In case of a completely
1219 * idle system the next event might even be infinite time into the
1220 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1221 * leaves the inner idle loop so the newly added timer is taken into
1222 * account when the CPU goes back to idle and evaluates the timer
1223 * wheel for the next timer event.
1225 void wake_up_idle_cpu(int cpu)
1227 struct rq *rq = cpu_rq(cpu);
1229 if (cpu == smp_processor_id())
1233 * This is safe, as this function is called with the timer
1234 * wheel base lock of (cpu) held. When the CPU is on the way
1235 * to idle and has not yet set rq->curr to idle then it will
1236 * be serialized on the timer wheel base lock and take the new
1237 * timer into account automatically.
1239 if (rq->curr != rq->idle)
1243 * We can set TIF_RESCHED on the idle task of the other CPU
1244 * lockless. The worst case is that the other CPU runs the
1245 * idle task through an additional NOOP schedule()
1247 set_tsk_need_resched(rq->idle);
1249 /* NEED_RESCHED must be visible before we test polling */
1251 if (!tsk_is_polling(rq->idle))
1252 smp_send_reschedule(cpu);
1255 #endif /* CONFIG_NO_HZ */
1257 static u64 sched_avg_period(void)
1259 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1262 static void sched_avg_update(struct rq *rq)
1264 s64 period = sched_avg_period();
1266 while ((s64)(rq->clock - rq->age_stamp) > period) {
1268 * Inline assembly required to prevent the compiler
1269 * optimising this loop into a divmod call.
1270 * See __iter_div_u64_rem() for another example of this.
1272 asm("" : "+rm" (rq->age_stamp));
1273 rq->age_stamp += period;
1278 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1280 rq->rt_avg += rt_delta;
1281 sched_avg_update(rq);
1284 #else /* !CONFIG_SMP */
1285 static void resched_task(struct task_struct *p)
1287 assert_raw_spin_locked(&task_rq(p)->lock);
1288 set_tsk_need_resched(p);
1291 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1295 static void sched_avg_update(struct rq *rq)
1298 #endif /* CONFIG_SMP */
1300 #if BITS_PER_LONG == 32
1301 # define WMULT_CONST (~0UL)
1303 # define WMULT_CONST (1UL << 32)
1306 #define WMULT_SHIFT 32
1309 * Shift right and round:
1311 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1314 * delta *= weight / lw
1316 static unsigned long
1317 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1318 struct load_weight *lw)
1322 if (!lw->inv_weight) {
1323 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1326 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1330 tmp = (u64)delta_exec * weight;
1332 * Check whether we'd overflow the 64-bit multiplication:
1334 if (unlikely(tmp > WMULT_CONST))
1335 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1338 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1340 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1343 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1349 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1355 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1362 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1363 * of tasks with abnormal "nice" values across CPUs the contribution that
1364 * each task makes to its run queue's load is weighted according to its
1365 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1366 * scaled version of the new time slice allocation that they receive on time
1370 #define WEIGHT_IDLEPRIO 3
1371 #define WMULT_IDLEPRIO 1431655765
1374 * Nice levels are multiplicative, with a gentle 10% change for every
1375 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1376 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1377 * that remained on nice 0.
1379 * The "10% effect" is relative and cumulative: from _any_ nice level,
1380 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1381 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1382 * If a task goes up by ~10% and another task goes down by ~10% then
1383 * the relative distance between them is ~25%.)
1385 static const int prio_to_weight[40] = {
1386 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1387 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1388 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1389 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1390 /* 0 */ 1024, 820, 655, 526, 423,
1391 /* 5 */ 335, 272, 215, 172, 137,
1392 /* 10 */ 110, 87, 70, 56, 45,
1393 /* 15 */ 36, 29, 23, 18, 15,
1397 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1399 * In cases where the weight does not change often, we can use the
1400 * precalculated inverse to speed up arithmetics by turning divisions
1401 * into multiplications:
1403 static const u32 prio_to_wmult[40] = {
1404 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1405 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1406 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1407 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1408 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1409 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1410 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1411 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1414 /* Time spent by the tasks of the cpu accounting group executing in ... */
1415 enum cpuacct_stat_index {
1416 CPUACCT_STAT_USER, /* ... user mode */
1417 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1419 CPUACCT_STAT_NSTATS,
1422 #ifdef CONFIG_CGROUP_CPUACCT
1423 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1424 static void cpuacct_update_stats(struct task_struct *tsk,
1425 enum cpuacct_stat_index idx, cputime_t val);
1427 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1428 static inline void cpuacct_update_stats(struct task_struct *tsk,
1429 enum cpuacct_stat_index idx, cputime_t val) {}
1432 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1434 update_load_add(&rq->load, load);
1437 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1439 update_load_sub(&rq->load, load);
1442 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1443 typedef int (*tg_visitor)(struct task_group *, void *);
1446 * Iterate the full tree, calling @down when first entering a node and @up when
1447 * leaving it for the final time.
1449 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1451 struct task_group *parent, *child;
1455 parent = &root_task_group;
1457 ret = (*down)(parent, data);
1460 list_for_each_entry_rcu(child, &parent->children, siblings) {
1467 ret = (*up)(parent, data);
1472 parent = parent->parent;
1481 static int tg_nop(struct task_group *tg, void *data)
1488 /* Used instead of source_load when we know the type == 0 */
1489 static unsigned long weighted_cpuload(const int cpu)
1491 return cpu_rq(cpu)->load.weight;
1495 * Return a low guess at the load of a migration-source cpu weighted
1496 * according to the scheduling class and "nice" value.
1498 * We want to under-estimate the load of migration sources, to
1499 * balance conservatively.
1501 static unsigned long source_load(int cpu, int type)
1503 struct rq *rq = cpu_rq(cpu);
1504 unsigned long total = weighted_cpuload(cpu);
1506 if (type == 0 || !sched_feat(LB_BIAS))
1509 return min(rq->cpu_load[type-1], total);
1513 * Return a high guess at the load of a migration-target cpu weighted
1514 * according to the scheduling class and "nice" value.
1516 static unsigned long target_load(int cpu, int type)
1518 struct rq *rq = cpu_rq(cpu);
1519 unsigned long total = weighted_cpuload(cpu);
1521 if (type == 0 || !sched_feat(LB_BIAS))
1524 return max(rq->cpu_load[type-1], total);
1527 static unsigned long power_of(int cpu)
1529 return cpu_rq(cpu)->cpu_power;
1532 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1534 static unsigned long cpu_avg_load_per_task(int cpu)
1536 struct rq *rq = cpu_rq(cpu);
1537 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1540 rq->avg_load_per_task = rq->load.weight / nr_running;
1542 rq->avg_load_per_task = 0;
1544 return rq->avg_load_per_task;
1547 #ifdef CONFIG_FAIR_GROUP_SCHED
1549 static void update_cfs_load(struct cfs_rq *cfs_rq, int lb);
1550 static void update_cfs_shares(struct cfs_rq *cfs_rq);
1553 * update tg->load_weight by folding this cpu's load_avg
1555 static int tg_shares_up(struct task_group *tg, void *data)
1558 struct cfs_rq *cfs_rq;
1559 unsigned long flags;
1560 int cpu = (long)data;
1567 cfs_rq = tg->cfs_rq[cpu];
1569 raw_spin_lock_irqsave(&rq->lock, flags);
1571 update_rq_clock(rq);
1572 update_cfs_load(cfs_rq, 1);
1574 load_avg = div64_u64(cfs_rq->load_avg, cfs_rq->load_period+1);
1575 load_avg -= cfs_rq->load_contribution;
1577 atomic_add(load_avg, &tg->load_weight);
1578 cfs_rq->load_contribution += load_avg;
1581 * We need to update shares after updating tg->load_weight in
1582 * order to adjust the weight of groups with long running tasks.
1584 update_cfs_shares(cfs_rq);
1586 raw_spin_unlock_irqrestore(&rq->lock, flags);
1592 * Compute the cpu's hierarchical load factor for each task group.
1593 * This needs to be done in a top-down fashion because the load of a child
1594 * group is a fraction of its parents load.
1596 static int tg_load_down(struct task_group *tg, void *data)
1599 long cpu = (long)data;
1602 load = cpu_rq(cpu)->load.weight;
1604 load = tg->parent->cfs_rq[cpu]->h_load;
1605 load *= tg->se[cpu]->load.weight;
1606 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1609 tg->cfs_rq[cpu]->h_load = load;
1614 static void update_shares(long cpu)
1616 if (root_task_group_empty())
1620 * XXX: replace with an on-demand list
1623 walk_tg_tree(tg_nop, tg_shares_up, (void *)cpu);
1626 static void update_h_load(long cpu)
1628 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1633 static inline void update_shares(int cpu)
1639 #ifdef CONFIG_PREEMPT
1641 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1644 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1645 * way at the expense of forcing extra atomic operations in all
1646 * invocations. This assures that the double_lock is acquired using the
1647 * same underlying policy as the spinlock_t on this architecture, which
1648 * reduces latency compared to the unfair variant below. However, it
1649 * also adds more overhead and therefore may reduce throughput.
1651 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1652 __releases(this_rq->lock)
1653 __acquires(busiest->lock)
1654 __acquires(this_rq->lock)
1656 raw_spin_unlock(&this_rq->lock);
1657 double_rq_lock(this_rq, busiest);
1664 * Unfair double_lock_balance: Optimizes throughput at the expense of
1665 * latency by eliminating extra atomic operations when the locks are
1666 * already in proper order on entry. This favors lower cpu-ids and will
1667 * grant the double lock to lower cpus over higher ids under contention,
1668 * regardless of entry order into the function.
1670 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1671 __releases(this_rq->lock)
1672 __acquires(busiest->lock)
1673 __acquires(this_rq->lock)
1677 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1678 if (busiest < this_rq) {
1679 raw_spin_unlock(&this_rq->lock);
1680 raw_spin_lock(&busiest->lock);
1681 raw_spin_lock_nested(&this_rq->lock,
1682 SINGLE_DEPTH_NESTING);
1685 raw_spin_lock_nested(&busiest->lock,
1686 SINGLE_DEPTH_NESTING);
1691 #endif /* CONFIG_PREEMPT */
1694 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1696 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1698 if (unlikely(!irqs_disabled())) {
1699 /* printk() doesn't work good under rq->lock */
1700 raw_spin_unlock(&this_rq->lock);
1704 return _double_lock_balance(this_rq, busiest);
1707 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1708 __releases(busiest->lock)
1710 raw_spin_unlock(&busiest->lock);
1711 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1715 * double_rq_lock - safely lock two runqueues
1717 * Note this does not disable interrupts like task_rq_lock,
1718 * you need to do so manually before calling.
1720 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1721 __acquires(rq1->lock)
1722 __acquires(rq2->lock)
1724 BUG_ON(!irqs_disabled());
1726 raw_spin_lock(&rq1->lock);
1727 __acquire(rq2->lock); /* Fake it out ;) */
1730 raw_spin_lock(&rq1->lock);
1731 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1733 raw_spin_lock(&rq2->lock);
1734 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1740 * double_rq_unlock - safely unlock two runqueues
1742 * Note this does not restore interrupts like task_rq_unlock,
1743 * you need to do so manually after calling.
1745 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1746 __releases(rq1->lock)
1747 __releases(rq2->lock)
1749 raw_spin_unlock(&rq1->lock);
1751 raw_spin_unlock(&rq2->lock);
1753 __release(rq2->lock);
1758 static void calc_load_account_idle(struct rq *this_rq);
1759 static void update_sysctl(void);
1760 static int get_update_sysctl_factor(void);
1761 static void update_cpu_load(struct rq *this_rq);
1763 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1765 set_task_rq(p, cpu);
1768 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1769 * successfuly executed on another CPU. We must ensure that updates of
1770 * per-task data have been completed by this moment.
1773 task_thread_info(p)->cpu = cpu;
1777 static const struct sched_class rt_sched_class;
1779 #define sched_class_highest (&stop_sched_class)
1780 #define for_each_class(class) \
1781 for (class = sched_class_highest; class; class = class->next)
1783 #include "sched_stats.h"
1785 static void inc_nr_running(struct rq *rq)
1790 static void dec_nr_running(struct rq *rq)
1795 static void set_load_weight(struct task_struct *p)
1798 * SCHED_IDLE tasks get minimal weight:
1800 if (p->policy == SCHED_IDLE) {
1801 p->se.load.weight = WEIGHT_IDLEPRIO;
1802 p->se.load.inv_weight = WMULT_IDLEPRIO;
1806 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1807 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1810 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1812 update_rq_clock(rq);
1813 sched_info_queued(p);
1814 p->sched_class->enqueue_task(rq, p, flags);
1818 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1820 update_rq_clock(rq);
1821 sched_info_dequeued(p);
1822 p->sched_class->dequeue_task(rq, p, flags);
1827 * activate_task - move a task to the runqueue.
1829 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1831 if (task_contributes_to_load(p))
1832 rq->nr_uninterruptible--;
1834 enqueue_task(rq, p, flags);
1839 * deactivate_task - remove a task from the runqueue.
1841 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1843 if (task_contributes_to_load(p))
1844 rq->nr_uninterruptible++;
1846 dequeue_task(rq, p, flags);
1850 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1853 * There are no locks covering percpu hardirq/softirq time.
1854 * They are only modified in account_system_vtime, on corresponding CPU
1855 * with interrupts disabled. So, writes are safe.
1856 * They are read and saved off onto struct rq in update_rq_clock().
1857 * This may result in other CPU reading this CPU's irq time and can
1858 * race with irq/account_system_vtime on this CPU. We would either get old
1859 * or new value (or semi updated value on 32 bit) with a side effect of
1860 * accounting a slice of irq time to wrong task when irq is in progress
1861 * while we read rq->clock. That is a worthy compromise in place of having
1862 * locks on each irq in account_system_time.
1864 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1865 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1867 static DEFINE_PER_CPU(u64, irq_start_time);
1868 static int sched_clock_irqtime;
1870 void enable_sched_clock_irqtime(void)
1872 sched_clock_irqtime = 1;
1875 void disable_sched_clock_irqtime(void)
1877 sched_clock_irqtime = 0;
1880 static u64 irq_time_cpu(int cpu)
1882 if (!sched_clock_irqtime)
1885 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1888 void account_system_vtime(struct task_struct *curr)
1890 unsigned long flags;
1894 if (!sched_clock_irqtime)
1897 local_irq_save(flags);
1899 cpu = smp_processor_id();
1900 now = sched_clock_cpu(cpu);
1901 delta = now - per_cpu(irq_start_time, cpu);
1902 per_cpu(irq_start_time, cpu) = now;
1904 * We do not account for softirq time from ksoftirqd here.
1905 * We want to continue accounting softirq time to ksoftirqd thread
1906 * in that case, so as not to confuse scheduler with a special task
1907 * that do not consume any time, but still wants to run.
1909 if (hardirq_count())
1910 per_cpu(cpu_hardirq_time, cpu) += delta;
1911 else if (in_serving_softirq() && !(curr->flags & PF_KSOFTIRQD))
1912 per_cpu(cpu_softirq_time, cpu) += delta;
1914 local_irq_restore(flags);
1916 EXPORT_SYMBOL_GPL(account_system_vtime);
1918 static void sched_irq_time_avg_update(struct rq *rq, u64 curr_irq_time)
1920 if (sched_clock_irqtime && sched_feat(NONIRQ_POWER)) {
1921 u64 delta_irq = curr_irq_time - rq->prev_irq_time;
1922 rq->prev_irq_time = curr_irq_time;
1923 sched_rt_avg_update(rq, delta_irq);
1929 static u64 irq_time_cpu(int cpu)
1934 static void sched_irq_time_avg_update(struct rq *rq, u64 curr_irq_time) { }
1938 #include "sched_idletask.c"
1939 #include "sched_fair.c"
1940 #include "sched_rt.c"
1941 #include "sched_stoptask.c"
1942 #ifdef CONFIG_SCHED_DEBUG
1943 # include "sched_debug.c"
1946 void sched_set_stop_task(int cpu, struct task_struct *stop)
1948 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
1949 struct task_struct *old_stop = cpu_rq(cpu)->stop;
1953 * Make it appear like a SCHED_FIFO task, its something
1954 * userspace knows about and won't get confused about.
1956 * Also, it will make PI more or less work without too
1957 * much confusion -- but then, stop work should not
1958 * rely on PI working anyway.
1960 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
1962 stop->sched_class = &stop_sched_class;
1965 cpu_rq(cpu)->stop = stop;
1969 * Reset it back to a normal scheduling class so that
1970 * it can die in pieces.
1972 old_stop->sched_class = &rt_sched_class;
1977 * __normal_prio - return the priority that is based on the static prio
1979 static inline int __normal_prio(struct task_struct *p)
1981 return p->static_prio;
1985 * Calculate the expected normal priority: i.e. priority
1986 * without taking RT-inheritance into account. Might be
1987 * boosted by interactivity modifiers. Changes upon fork,
1988 * setprio syscalls, and whenever the interactivity
1989 * estimator recalculates.
1991 static inline int normal_prio(struct task_struct *p)
1995 if (task_has_rt_policy(p))
1996 prio = MAX_RT_PRIO-1 - p->rt_priority;
1998 prio = __normal_prio(p);
2003 * Calculate the current priority, i.e. the priority
2004 * taken into account by the scheduler. This value might
2005 * be boosted by RT tasks, or might be boosted by
2006 * interactivity modifiers. Will be RT if the task got
2007 * RT-boosted. If not then it returns p->normal_prio.
2009 static int effective_prio(struct task_struct *p)
2011 p->normal_prio = normal_prio(p);
2013 * If we are RT tasks or we were boosted to RT priority,
2014 * keep the priority unchanged. Otherwise, update priority
2015 * to the normal priority:
2017 if (!rt_prio(p->prio))
2018 return p->normal_prio;
2023 * task_curr - is this task currently executing on a CPU?
2024 * @p: the task in question.
2026 inline int task_curr(const struct task_struct *p)
2028 return cpu_curr(task_cpu(p)) == p;
2031 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2032 const struct sched_class *prev_class,
2033 int oldprio, int running)
2035 if (prev_class != p->sched_class) {
2036 if (prev_class->switched_from)
2037 prev_class->switched_from(rq, p, running);
2038 p->sched_class->switched_to(rq, p, running);
2040 p->sched_class->prio_changed(rq, p, oldprio, running);
2045 * Is this task likely cache-hot:
2048 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2052 if (p->sched_class != &fair_sched_class)
2055 if (unlikely(p->policy == SCHED_IDLE))
2059 * Buddy candidates are cache hot:
2061 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2062 (&p->se == cfs_rq_of(&p->se)->next ||
2063 &p->se == cfs_rq_of(&p->se)->last))
2066 if (sysctl_sched_migration_cost == -1)
2068 if (sysctl_sched_migration_cost == 0)
2071 delta = now - p->se.exec_start;
2073 return delta < (s64)sysctl_sched_migration_cost;
2076 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2078 #ifdef CONFIG_SCHED_DEBUG
2080 * We should never call set_task_cpu() on a blocked task,
2081 * ttwu() will sort out the placement.
2083 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2084 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2087 trace_sched_migrate_task(p, new_cpu);
2089 if (task_cpu(p) != new_cpu) {
2090 p->se.nr_migrations++;
2091 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2094 __set_task_cpu(p, new_cpu);
2097 struct migration_arg {
2098 struct task_struct *task;
2102 static int migration_cpu_stop(void *data);
2105 * The task's runqueue lock must be held.
2106 * Returns true if you have to wait for migration thread.
2108 static bool migrate_task(struct task_struct *p, int dest_cpu)
2110 struct rq *rq = task_rq(p);
2113 * If the task is not on a runqueue (and not running), then
2114 * the next wake-up will properly place the task.
2116 return p->se.on_rq || task_running(rq, p);
2120 * wait_task_inactive - wait for a thread to unschedule.
2122 * If @match_state is nonzero, it's the @p->state value just checked and
2123 * not expected to change. If it changes, i.e. @p might have woken up,
2124 * then return zero. When we succeed in waiting for @p to be off its CPU,
2125 * we return a positive number (its total switch count). If a second call
2126 * a short while later returns the same number, the caller can be sure that
2127 * @p has remained unscheduled the whole time.
2129 * The caller must ensure that the task *will* unschedule sometime soon,
2130 * else this function might spin for a *long* time. This function can't
2131 * be called with interrupts off, or it may introduce deadlock with
2132 * smp_call_function() if an IPI is sent by the same process we are
2133 * waiting to become inactive.
2135 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2137 unsigned long flags;
2144 * We do the initial early heuristics without holding
2145 * any task-queue locks at all. We'll only try to get
2146 * the runqueue lock when things look like they will
2152 * If the task is actively running on another CPU
2153 * still, just relax and busy-wait without holding
2156 * NOTE! Since we don't hold any locks, it's not
2157 * even sure that "rq" stays as the right runqueue!
2158 * But we don't care, since "task_running()" will
2159 * return false if the runqueue has changed and p
2160 * is actually now running somewhere else!
2162 while (task_running(rq, p)) {
2163 if (match_state && unlikely(p->state != match_state))
2169 * Ok, time to look more closely! We need the rq
2170 * lock now, to be *sure*. If we're wrong, we'll
2171 * just go back and repeat.
2173 rq = task_rq_lock(p, &flags);
2174 trace_sched_wait_task(p);
2175 running = task_running(rq, p);
2176 on_rq = p->se.on_rq;
2178 if (!match_state || p->state == match_state)
2179 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2180 task_rq_unlock(rq, &flags);
2183 * If it changed from the expected state, bail out now.
2185 if (unlikely(!ncsw))
2189 * Was it really running after all now that we
2190 * checked with the proper locks actually held?
2192 * Oops. Go back and try again..
2194 if (unlikely(running)) {
2200 * It's not enough that it's not actively running,
2201 * it must be off the runqueue _entirely_, and not
2204 * So if it was still runnable (but just not actively
2205 * running right now), it's preempted, and we should
2206 * yield - it could be a while.
2208 if (unlikely(on_rq)) {
2209 schedule_timeout_uninterruptible(1);
2214 * Ahh, all good. It wasn't running, and it wasn't
2215 * runnable, which means that it will never become
2216 * running in the future either. We're all done!
2225 * kick_process - kick a running thread to enter/exit the kernel
2226 * @p: the to-be-kicked thread
2228 * Cause a process which is running on another CPU to enter
2229 * kernel-mode, without any delay. (to get signals handled.)
2231 * NOTE: this function doesnt have to take the runqueue lock,
2232 * because all it wants to ensure is that the remote task enters
2233 * the kernel. If the IPI races and the task has been migrated
2234 * to another CPU then no harm is done and the purpose has been
2237 void kick_process(struct task_struct *p)
2243 if ((cpu != smp_processor_id()) && task_curr(p))
2244 smp_send_reschedule(cpu);
2247 EXPORT_SYMBOL_GPL(kick_process);
2248 #endif /* CONFIG_SMP */
2251 * task_oncpu_function_call - call a function on the cpu on which a task runs
2252 * @p: the task to evaluate
2253 * @func: the function to be called
2254 * @info: the function call argument
2256 * Calls the function @func when the task is currently running. This might
2257 * be on the current CPU, which just calls the function directly
2259 void task_oncpu_function_call(struct task_struct *p,
2260 void (*func) (void *info), void *info)
2267 smp_call_function_single(cpu, func, info, 1);
2273 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2275 static int select_fallback_rq(int cpu, struct task_struct *p)
2278 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2280 /* Look for allowed, online CPU in same node. */
2281 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2282 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2285 /* Any allowed, online CPU? */
2286 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2287 if (dest_cpu < nr_cpu_ids)
2290 /* No more Mr. Nice Guy. */
2291 dest_cpu = cpuset_cpus_allowed_fallback(p);
2293 * Don't tell them about moving exiting tasks or
2294 * kernel threads (both mm NULL), since they never
2297 if (p->mm && printk_ratelimit()) {
2298 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2299 task_pid_nr(p), p->comm, cpu);
2306 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2309 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2311 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2314 * In order not to call set_task_cpu() on a blocking task we need
2315 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2318 * Since this is common to all placement strategies, this lives here.
2320 * [ this allows ->select_task() to simply return task_cpu(p) and
2321 * not worry about this generic constraint ]
2323 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2325 cpu = select_fallback_rq(task_cpu(p), p);
2330 static void update_avg(u64 *avg, u64 sample)
2332 s64 diff = sample - *avg;
2337 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2338 bool is_sync, bool is_migrate, bool is_local,
2339 unsigned long en_flags)
2341 schedstat_inc(p, se.statistics.nr_wakeups);
2343 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2345 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2347 schedstat_inc(p, se.statistics.nr_wakeups_local);
2349 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2351 activate_task(rq, p, en_flags);
2354 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2355 int wake_flags, bool success)
2357 trace_sched_wakeup(p, success);
2358 check_preempt_curr(rq, p, wake_flags);
2360 p->state = TASK_RUNNING;
2362 if (p->sched_class->task_woken)
2363 p->sched_class->task_woken(rq, p);
2365 if (unlikely(rq->idle_stamp)) {
2366 u64 delta = rq->clock - rq->idle_stamp;
2367 u64 max = 2*sysctl_sched_migration_cost;
2372 update_avg(&rq->avg_idle, delta);
2376 /* if a worker is waking up, notify workqueue */
2377 if ((p->flags & PF_WQ_WORKER) && success)
2378 wq_worker_waking_up(p, cpu_of(rq));
2382 * try_to_wake_up - wake up a thread
2383 * @p: the thread to be awakened
2384 * @state: the mask of task states that can be woken
2385 * @wake_flags: wake modifier flags (WF_*)
2387 * Put it on the run-queue if it's not already there. The "current"
2388 * thread is always on the run-queue (except when the actual
2389 * re-schedule is in progress), and as such you're allowed to do
2390 * the simpler "current->state = TASK_RUNNING" to mark yourself
2391 * runnable without the overhead of this.
2393 * Returns %true if @p was woken up, %false if it was already running
2394 * or @state didn't match @p's state.
2396 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2399 int cpu, orig_cpu, this_cpu, success = 0;
2400 unsigned long flags;
2401 unsigned long en_flags = ENQUEUE_WAKEUP;
2404 this_cpu = get_cpu();
2407 rq = task_rq_lock(p, &flags);
2408 if (!(p->state & state))
2418 if (unlikely(task_running(rq, p)))
2422 * In order to handle concurrent wakeups and release the rq->lock
2423 * we put the task in TASK_WAKING state.
2425 * First fix up the nr_uninterruptible count:
2427 if (task_contributes_to_load(p)) {
2428 if (likely(cpu_online(orig_cpu)))
2429 rq->nr_uninterruptible--;
2431 this_rq()->nr_uninterruptible--;
2433 p->state = TASK_WAKING;
2435 if (p->sched_class->task_waking) {
2436 p->sched_class->task_waking(rq, p);
2437 en_flags |= ENQUEUE_WAKING;
2440 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2441 if (cpu != orig_cpu)
2442 set_task_cpu(p, cpu);
2443 __task_rq_unlock(rq);
2446 raw_spin_lock(&rq->lock);
2449 * We migrated the task without holding either rq->lock, however
2450 * since the task is not on the task list itself, nobody else
2451 * will try and migrate the task, hence the rq should match the
2452 * cpu we just moved it to.
2454 WARN_ON(task_cpu(p) != cpu);
2455 WARN_ON(p->state != TASK_WAKING);
2457 #ifdef CONFIG_SCHEDSTATS
2458 schedstat_inc(rq, ttwu_count);
2459 if (cpu == this_cpu)
2460 schedstat_inc(rq, ttwu_local);
2462 struct sched_domain *sd;
2463 for_each_domain(this_cpu, sd) {
2464 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2465 schedstat_inc(sd, ttwu_wake_remote);
2470 #endif /* CONFIG_SCHEDSTATS */
2473 #endif /* CONFIG_SMP */
2474 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2475 cpu == this_cpu, en_flags);
2478 ttwu_post_activation(p, rq, wake_flags, success);
2480 task_rq_unlock(rq, &flags);
2487 * try_to_wake_up_local - try to wake up a local task with rq lock held
2488 * @p: the thread to be awakened
2490 * Put @p on the run-queue if it's not alredy there. The caller must
2491 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2492 * the current task. this_rq() stays locked over invocation.
2494 static void try_to_wake_up_local(struct task_struct *p)
2496 struct rq *rq = task_rq(p);
2497 bool success = false;
2499 BUG_ON(rq != this_rq());
2500 BUG_ON(p == current);
2501 lockdep_assert_held(&rq->lock);
2503 if (!(p->state & TASK_NORMAL))
2507 if (likely(!task_running(rq, p))) {
2508 schedstat_inc(rq, ttwu_count);
2509 schedstat_inc(rq, ttwu_local);
2511 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2514 ttwu_post_activation(p, rq, 0, success);
2518 * wake_up_process - Wake up a specific process
2519 * @p: The process to be woken up.
2521 * Attempt to wake up the nominated process and move it to the set of runnable
2522 * processes. Returns 1 if the process was woken up, 0 if it was already
2525 * It may be assumed that this function implies a write memory barrier before
2526 * changing the task state if and only if any tasks are woken up.
2528 int wake_up_process(struct task_struct *p)
2530 return try_to_wake_up(p, TASK_ALL, 0);
2532 EXPORT_SYMBOL(wake_up_process);
2534 int wake_up_state(struct task_struct *p, unsigned int state)
2536 return try_to_wake_up(p, state, 0);
2540 * Perform scheduler related setup for a newly forked process p.
2541 * p is forked by current.
2543 * __sched_fork() is basic setup used by init_idle() too:
2545 static void __sched_fork(struct task_struct *p)
2547 p->se.exec_start = 0;
2548 p->se.sum_exec_runtime = 0;
2549 p->se.prev_sum_exec_runtime = 0;
2550 p->se.nr_migrations = 0;
2552 #ifdef CONFIG_SCHEDSTATS
2553 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2556 INIT_LIST_HEAD(&p->rt.run_list);
2558 INIT_LIST_HEAD(&p->se.group_node);
2560 #ifdef CONFIG_PREEMPT_NOTIFIERS
2561 INIT_HLIST_HEAD(&p->preempt_notifiers);
2566 * fork()/clone()-time setup:
2568 void sched_fork(struct task_struct *p, int clone_flags)
2570 int cpu = get_cpu();
2574 * We mark the process as running here. This guarantees that
2575 * nobody will actually run it, and a signal or other external
2576 * event cannot wake it up and insert it on the runqueue either.
2578 p->state = TASK_RUNNING;
2581 * Revert to default priority/policy on fork if requested.
2583 if (unlikely(p->sched_reset_on_fork)) {
2584 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2585 p->policy = SCHED_NORMAL;
2586 p->normal_prio = p->static_prio;
2589 if (PRIO_TO_NICE(p->static_prio) < 0) {
2590 p->static_prio = NICE_TO_PRIO(0);
2591 p->normal_prio = p->static_prio;
2596 * We don't need the reset flag anymore after the fork. It has
2597 * fulfilled its duty:
2599 p->sched_reset_on_fork = 0;
2603 * Make sure we do not leak PI boosting priority to the child.
2605 p->prio = current->normal_prio;
2607 if (!rt_prio(p->prio))
2608 p->sched_class = &fair_sched_class;
2610 if (p->sched_class->task_fork)
2611 p->sched_class->task_fork(p);
2614 * The child is not yet in the pid-hash so no cgroup attach races,
2615 * and the cgroup is pinned to this child due to cgroup_fork()
2616 * is ran before sched_fork().
2618 * Silence PROVE_RCU.
2621 set_task_cpu(p, cpu);
2624 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2625 if (likely(sched_info_on()))
2626 memset(&p->sched_info, 0, sizeof(p->sched_info));
2628 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2631 #ifdef CONFIG_PREEMPT
2632 /* Want to start with kernel preemption disabled. */
2633 task_thread_info(p)->preempt_count = 1;
2635 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2641 * wake_up_new_task - wake up a newly created task for the first time.
2643 * This function will do some initial scheduler statistics housekeeping
2644 * that must be done for every newly created context, then puts the task
2645 * on the runqueue and wakes it.
2647 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2649 unsigned long flags;
2651 int cpu __maybe_unused = get_cpu();
2654 rq = task_rq_lock(p, &flags);
2655 p->state = TASK_WAKING;
2658 * Fork balancing, do it here and not earlier because:
2659 * - cpus_allowed can change in the fork path
2660 * - any previously selected cpu might disappear through hotplug
2662 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2663 * without people poking at ->cpus_allowed.
2665 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2666 set_task_cpu(p, cpu);
2668 p->state = TASK_RUNNING;
2669 task_rq_unlock(rq, &flags);
2672 rq = task_rq_lock(p, &flags);
2673 activate_task(rq, p, 0);
2674 trace_sched_wakeup_new(p, 1);
2675 check_preempt_curr(rq, p, WF_FORK);
2677 if (p->sched_class->task_woken)
2678 p->sched_class->task_woken(rq, p);
2680 task_rq_unlock(rq, &flags);
2684 #ifdef CONFIG_PREEMPT_NOTIFIERS
2687 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2688 * @notifier: notifier struct to register
2690 void preempt_notifier_register(struct preempt_notifier *notifier)
2692 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2694 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2697 * preempt_notifier_unregister - no longer interested in preemption notifications
2698 * @notifier: notifier struct to unregister
2700 * This is safe to call from within a preemption notifier.
2702 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2704 hlist_del(¬ifier->link);
2706 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2708 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2710 struct preempt_notifier *notifier;
2711 struct hlist_node *node;
2713 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2714 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2718 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2719 struct task_struct *next)
2721 struct preempt_notifier *notifier;
2722 struct hlist_node *node;
2724 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2725 notifier->ops->sched_out(notifier, next);
2728 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2730 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2735 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2736 struct task_struct *next)
2740 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2743 * prepare_task_switch - prepare to switch tasks
2744 * @rq: the runqueue preparing to switch
2745 * @prev: the current task that is being switched out
2746 * @next: the task we are going to switch to.
2748 * This is called with the rq lock held and interrupts off. It must
2749 * be paired with a subsequent finish_task_switch after the context
2752 * prepare_task_switch sets up locking and calls architecture specific
2756 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2757 struct task_struct *next)
2759 fire_sched_out_preempt_notifiers(prev, next);
2760 prepare_lock_switch(rq, next);
2761 prepare_arch_switch(next);
2765 * finish_task_switch - clean up after a task-switch
2766 * @rq: runqueue associated with task-switch
2767 * @prev: the thread we just switched away from.
2769 * finish_task_switch must be called after the context switch, paired
2770 * with a prepare_task_switch call before the context switch.
2771 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2772 * and do any other architecture-specific cleanup actions.
2774 * Note that we may have delayed dropping an mm in context_switch(). If
2775 * so, we finish that here outside of the runqueue lock. (Doing it
2776 * with the lock held can cause deadlocks; see schedule() for
2779 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2780 __releases(rq->lock)
2782 struct mm_struct *mm = rq->prev_mm;
2788 * A task struct has one reference for the use as "current".
2789 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2790 * schedule one last time. The schedule call will never return, and
2791 * the scheduled task must drop that reference.
2792 * The test for TASK_DEAD must occur while the runqueue locks are
2793 * still held, otherwise prev could be scheduled on another cpu, die
2794 * there before we look at prev->state, and then the reference would
2796 * Manfred Spraul <manfred@colorfullife.com>
2798 prev_state = prev->state;
2799 finish_arch_switch(prev);
2800 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2801 local_irq_disable();
2802 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2803 perf_event_task_sched_in(current);
2804 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2806 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2807 finish_lock_switch(rq, prev);
2809 fire_sched_in_preempt_notifiers(current);
2812 if (unlikely(prev_state == TASK_DEAD)) {
2814 * Remove function-return probe instances associated with this
2815 * task and put them back on the free list.
2817 kprobe_flush_task(prev);
2818 put_task_struct(prev);
2824 /* assumes rq->lock is held */
2825 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2827 if (prev->sched_class->pre_schedule)
2828 prev->sched_class->pre_schedule(rq, prev);
2831 /* rq->lock is NOT held, but preemption is disabled */
2832 static inline void post_schedule(struct rq *rq)
2834 if (rq->post_schedule) {
2835 unsigned long flags;
2837 raw_spin_lock_irqsave(&rq->lock, flags);
2838 if (rq->curr->sched_class->post_schedule)
2839 rq->curr->sched_class->post_schedule(rq);
2840 raw_spin_unlock_irqrestore(&rq->lock, flags);
2842 rq->post_schedule = 0;
2848 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2852 static inline void post_schedule(struct rq *rq)
2859 * schedule_tail - first thing a freshly forked thread must call.
2860 * @prev: the thread we just switched away from.
2862 asmlinkage void schedule_tail(struct task_struct *prev)
2863 __releases(rq->lock)
2865 struct rq *rq = this_rq();
2867 finish_task_switch(rq, prev);
2870 * FIXME: do we need to worry about rq being invalidated by the
2875 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2876 /* In this case, finish_task_switch does not reenable preemption */
2879 if (current->set_child_tid)
2880 put_user(task_pid_vnr(current), current->set_child_tid);
2884 * context_switch - switch to the new MM and the new
2885 * thread's register state.
2888 context_switch(struct rq *rq, struct task_struct *prev,
2889 struct task_struct *next)
2891 struct mm_struct *mm, *oldmm;
2893 prepare_task_switch(rq, prev, next);
2894 trace_sched_switch(prev, next);
2896 oldmm = prev->active_mm;
2898 * For paravirt, this is coupled with an exit in switch_to to
2899 * combine the page table reload and the switch backend into
2902 arch_start_context_switch(prev);
2905 next->active_mm = oldmm;
2906 atomic_inc(&oldmm->mm_count);
2907 enter_lazy_tlb(oldmm, next);
2909 switch_mm(oldmm, mm, next);
2912 prev->active_mm = NULL;
2913 rq->prev_mm = oldmm;
2916 * Since the runqueue lock will be released by the next
2917 * task (which is an invalid locking op but in the case
2918 * of the scheduler it's an obvious special-case), so we
2919 * do an early lockdep release here:
2921 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2922 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2925 /* Here we just switch the register state and the stack. */
2926 switch_to(prev, next, prev);
2930 * this_rq must be evaluated again because prev may have moved
2931 * CPUs since it called schedule(), thus the 'rq' on its stack
2932 * frame will be invalid.
2934 finish_task_switch(this_rq(), prev);
2938 * nr_running, nr_uninterruptible and nr_context_switches:
2940 * externally visible scheduler statistics: current number of runnable
2941 * threads, current number of uninterruptible-sleeping threads, total
2942 * number of context switches performed since bootup.
2944 unsigned long nr_running(void)
2946 unsigned long i, sum = 0;
2948 for_each_online_cpu(i)
2949 sum += cpu_rq(i)->nr_running;
2954 unsigned long nr_uninterruptible(void)
2956 unsigned long i, sum = 0;
2958 for_each_possible_cpu(i)
2959 sum += cpu_rq(i)->nr_uninterruptible;
2962 * Since we read the counters lockless, it might be slightly
2963 * inaccurate. Do not allow it to go below zero though:
2965 if (unlikely((long)sum < 0))
2971 unsigned long long nr_context_switches(void)
2974 unsigned long long sum = 0;
2976 for_each_possible_cpu(i)
2977 sum += cpu_rq(i)->nr_switches;
2982 unsigned long nr_iowait(void)
2984 unsigned long i, sum = 0;
2986 for_each_possible_cpu(i)
2987 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2992 unsigned long nr_iowait_cpu(int cpu)
2994 struct rq *this = cpu_rq(cpu);
2995 return atomic_read(&this->nr_iowait);
2998 unsigned long this_cpu_load(void)
3000 struct rq *this = this_rq();
3001 return this->cpu_load[0];
3005 /* Variables and functions for calc_load */
3006 static atomic_long_t calc_load_tasks;
3007 static unsigned long calc_load_update;
3008 unsigned long avenrun[3];
3009 EXPORT_SYMBOL(avenrun);
3011 static long calc_load_fold_active(struct rq *this_rq)
3013 long nr_active, delta = 0;
3015 nr_active = this_rq->nr_running;
3016 nr_active += (long) this_rq->nr_uninterruptible;
3018 if (nr_active != this_rq->calc_load_active) {
3019 delta = nr_active - this_rq->calc_load_active;
3020 this_rq->calc_load_active = nr_active;
3028 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3030 * When making the ILB scale, we should try to pull this in as well.
3032 static atomic_long_t calc_load_tasks_idle;
3034 static void calc_load_account_idle(struct rq *this_rq)
3038 delta = calc_load_fold_active(this_rq);
3040 atomic_long_add(delta, &calc_load_tasks_idle);
3043 static long calc_load_fold_idle(void)
3048 * Its got a race, we don't care...
3050 if (atomic_long_read(&calc_load_tasks_idle))
3051 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3056 static void calc_load_account_idle(struct rq *this_rq)
3060 static inline long calc_load_fold_idle(void)
3067 * get_avenrun - get the load average array
3068 * @loads: pointer to dest load array
3069 * @offset: offset to add
3070 * @shift: shift count to shift the result left
3072 * These values are estimates at best, so no need for locking.
3074 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3076 loads[0] = (avenrun[0] + offset) << shift;
3077 loads[1] = (avenrun[1] + offset) << shift;
3078 loads[2] = (avenrun[2] + offset) << shift;
3081 static unsigned long
3082 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3085 load += active * (FIXED_1 - exp);
3086 return load >> FSHIFT;
3090 * calc_load - update the avenrun load estimates 10 ticks after the
3091 * CPUs have updated calc_load_tasks.
3093 void calc_global_load(void)
3095 unsigned long upd = calc_load_update + 10;
3098 if (time_before(jiffies, upd))
3101 active = atomic_long_read(&calc_load_tasks);
3102 active = active > 0 ? active * FIXED_1 : 0;
3104 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3105 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3106 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3108 calc_load_update += LOAD_FREQ;
3112 * Called from update_cpu_load() to periodically update this CPU's
3115 static void calc_load_account_active(struct rq *this_rq)
3119 if (time_before(jiffies, this_rq->calc_load_update))
3122 delta = calc_load_fold_active(this_rq);
3123 delta += calc_load_fold_idle();
3125 atomic_long_add(delta, &calc_load_tasks);
3127 this_rq->calc_load_update += LOAD_FREQ;
3131 * The exact cpuload at various idx values, calculated at every tick would be
3132 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3134 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3135 * on nth tick when cpu may be busy, then we have:
3136 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3137 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3139 * decay_load_missed() below does efficient calculation of
3140 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3141 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3143 * The calculation is approximated on a 128 point scale.
3144 * degrade_zero_ticks is the number of ticks after which load at any
3145 * particular idx is approximated to be zero.
3146 * degrade_factor is a precomputed table, a row for each load idx.
3147 * Each column corresponds to degradation factor for a power of two ticks,
3148 * based on 128 point scale.
3150 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3151 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3153 * With this power of 2 load factors, we can degrade the load n times
3154 * by looking at 1 bits in n and doing as many mult/shift instead of
3155 * n mult/shifts needed by the exact degradation.
3157 #define DEGRADE_SHIFT 7
3158 static const unsigned char
3159 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3160 static const unsigned char
3161 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3162 {0, 0, 0, 0, 0, 0, 0, 0},
3163 {64, 32, 8, 0, 0, 0, 0, 0},
3164 {96, 72, 40, 12, 1, 0, 0},
3165 {112, 98, 75, 43, 15, 1, 0},
3166 {120, 112, 98, 76, 45, 16, 2} };
3169 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3170 * would be when CPU is idle and so we just decay the old load without
3171 * adding any new load.
3173 static unsigned long
3174 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3178 if (!missed_updates)
3181 if (missed_updates >= degrade_zero_ticks[idx])
3185 return load >> missed_updates;
3187 while (missed_updates) {
3188 if (missed_updates % 2)
3189 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3191 missed_updates >>= 1;
3198 * Update rq->cpu_load[] statistics. This function is usually called every
3199 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3200 * every tick. We fix it up based on jiffies.
3202 static void update_cpu_load(struct rq *this_rq)
3204 unsigned long this_load = this_rq->load.weight;
3205 unsigned long curr_jiffies = jiffies;
3206 unsigned long pending_updates;
3209 this_rq->nr_load_updates++;
3211 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3212 if (curr_jiffies == this_rq->last_load_update_tick)
3215 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3216 this_rq->last_load_update_tick = curr_jiffies;
3218 /* Update our load: */
3219 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3220 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3221 unsigned long old_load, new_load;
3223 /* scale is effectively 1 << i now, and >> i divides by scale */
3225 old_load = this_rq->cpu_load[i];
3226 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3227 new_load = this_load;
3229 * Round up the averaging division if load is increasing. This
3230 * prevents us from getting stuck on 9 if the load is 10, for
3233 if (new_load > old_load)
3234 new_load += scale - 1;
3236 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3239 sched_avg_update(this_rq);
3242 static void update_cpu_load_active(struct rq *this_rq)
3244 update_cpu_load(this_rq);
3246 calc_load_account_active(this_rq);
3252 * sched_exec - execve() is a valuable balancing opportunity, because at
3253 * this point the task has the smallest effective memory and cache footprint.
3255 void sched_exec(void)
3257 struct task_struct *p = current;
3258 unsigned long flags;
3262 rq = task_rq_lock(p, &flags);
3263 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3264 if (dest_cpu == smp_processor_id())
3268 * select_task_rq() can race against ->cpus_allowed
3270 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3271 likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
3272 struct migration_arg arg = { p, dest_cpu };
3274 task_rq_unlock(rq, &flags);
3275 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3279 task_rq_unlock(rq, &flags);
3284 DEFINE_PER_CPU(struct kernel_stat, kstat);
3286 EXPORT_PER_CPU_SYMBOL(kstat);
3289 * Return any ns on the sched_clock that have not yet been accounted in
3290 * @p in case that task is currently running.
3292 * Called with task_rq_lock() held on @rq.
3294 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3298 if (task_current(rq, p)) {
3299 update_rq_clock(rq);
3300 ns = rq->clock_task - p->se.exec_start;
3308 unsigned long long task_delta_exec(struct task_struct *p)
3310 unsigned long flags;
3314 rq = task_rq_lock(p, &flags);
3315 ns = do_task_delta_exec(p, rq);
3316 task_rq_unlock(rq, &flags);
3322 * Return accounted runtime for the task.
3323 * In case the task is currently running, return the runtime plus current's
3324 * pending runtime that have not been accounted yet.
3326 unsigned long long task_sched_runtime(struct task_struct *p)
3328 unsigned long flags;
3332 rq = task_rq_lock(p, &flags);
3333 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3334 task_rq_unlock(rq, &flags);
3340 * Return sum_exec_runtime for the thread group.
3341 * In case the task is currently running, return the sum plus current's
3342 * pending runtime that have not been accounted yet.
3344 * Note that the thread group might have other running tasks as well,
3345 * so the return value not includes other pending runtime that other
3346 * running tasks might have.
3348 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3350 struct task_cputime totals;
3351 unsigned long flags;
3355 rq = task_rq_lock(p, &flags);
3356 thread_group_cputime(p, &totals);
3357 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3358 task_rq_unlock(rq, &flags);
3364 * Account user cpu time to a process.
3365 * @p: the process that the cpu time gets accounted to
3366 * @cputime: the cpu time spent in user space since the last update
3367 * @cputime_scaled: cputime scaled by cpu frequency
3369 void account_user_time(struct task_struct *p, cputime_t cputime,
3370 cputime_t cputime_scaled)
3372 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3375 /* Add user time to process. */
3376 p->utime = cputime_add(p->utime, cputime);
3377 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3378 account_group_user_time(p, cputime);
3380 /* Add user time to cpustat. */
3381 tmp = cputime_to_cputime64(cputime);
3382 if (TASK_NICE(p) > 0)
3383 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3385 cpustat->user = cputime64_add(cpustat->user, tmp);
3387 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3388 /* Account for user time used */
3389 acct_update_integrals(p);
3393 * Account guest cpu time to a process.
3394 * @p: the process that the cpu time gets accounted to
3395 * @cputime: the cpu time spent in virtual machine since the last update
3396 * @cputime_scaled: cputime scaled by cpu frequency
3398 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3399 cputime_t cputime_scaled)
3402 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3404 tmp = cputime_to_cputime64(cputime);
3406 /* Add guest time to process. */
3407 p->utime = cputime_add(p->utime, cputime);
3408 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3409 account_group_user_time(p, cputime);
3410 p->gtime = cputime_add(p->gtime, cputime);
3412 /* Add guest time to cpustat. */
3413 if (TASK_NICE(p) > 0) {
3414 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3415 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3417 cpustat->user = cputime64_add(cpustat->user, tmp);
3418 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3423 * Account system cpu time to a process.
3424 * @p: the process that the cpu time gets accounted to
3425 * @hardirq_offset: the offset to subtract from hardirq_count()
3426 * @cputime: the cpu time spent in kernel space since the last update
3427 * @cputime_scaled: cputime scaled by cpu frequency
3429 void account_system_time(struct task_struct *p, int hardirq_offset,
3430 cputime_t cputime, cputime_t cputime_scaled)
3432 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3435 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3436 account_guest_time(p, cputime, cputime_scaled);
3440 /* Add system time to process. */
3441 p->stime = cputime_add(p->stime, cputime);
3442 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3443 account_group_system_time(p, cputime);
3445 /* Add system time to cpustat. */
3446 tmp = cputime_to_cputime64(cputime);
3447 if (hardirq_count() - hardirq_offset)
3448 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3449 else if (in_serving_softirq())
3450 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3452 cpustat->system = cputime64_add(cpustat->system, tmp);
3454 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3456 /* Account for system time used */
3457 acct_update_integrals(p);
3461 * Account for involuntary wait time.
3462 * @steal: the cpu time spent in involuntary wait
3464 void account_steal_time(cputime_t cputime)
3466 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3467 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3469 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3473 * Account for idle time.
3474 * @cputime: the cpu time spent in idle wait
3476 void account_idle_time(cputime_t cputime)
3478 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3479 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3480 struct rq *rq = this_rq();
3482 if (atomic_read(&rq->nr_iowait) > 0)
3483 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3485 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3488 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3491 * Account a single tick of cpu time.
3492 * @p: the process that the cpu time gets accounted to
3493 * @user_tick: indicates if the tick is a user or a system tick
3495 void account_process_tick(struct task_struct *p, int user_tick)
3497 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3498 struct rq *rq = this_rq();
3501 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3502 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3503 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3506 account_idle_time(cputime_one_jiffy);
3510 * Account multiple ticks of steal time.
3511 * @p: the process from which the cpu time has been stolen
3512 * @ticks: number of stolen ticks
3514 void account_steal_ticks(unsigned long ticks)
3516 account_steal_time(jiffies_to_cputime(ticks));
3520 * Account multiple ticks of idle time.
3521 * @ticks: number of stolen ticks
3523 void account_idle_ticks(unsigned long ticks)
3525 account_idle_time(jiffies_to_cputime(ticks));
3531 * Use precise platform statistics if available:
3533 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3534 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3540 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3542 struct task_cputime cputime;
3544 thread_group_cputime(p, &cputime);
3546 *ut = cputime.utime;
3547 *st = cputime.stime;
3551 #ifndef nsecs_to_cputime
3552 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3555 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3557 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3560 * Use CFS's precise accounting:
3562 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3568 do_div(temp, total);
3569 utime = (cputime_t)temp;
3574 * Compare with previous values, to keep monotonicity:
3576 p->prev_utime = max(p->prev_utime, utime);
3577 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3579 *ut = p->prev_utime;
3580 *st = p->prev_stime;
3584 * Must be called with siglock held.
3586 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3588 struct signal_struct *sig = p->signal;
3589 struct task_cputime cputime;
3590 cputime_t rtime, utime, total;
3592 thread_group_cputime(p, &cputime);
3594 total = cputime_add(cputime.utime, cputime.stime);
3595 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3600 temp *= cputime.utime;
3601 do_div(temp, total);
3602 utime = (cputime_t)temp;
3606 sig->prev_utime = max(sig->prev_utime, utime);
3607 sig->prev_stime = max(sig->prev_stime,
3608 cputime_sub(rtime, sig->prev_utime));
3610 *ut = sig->prev_utime;
3611 *st = sig->prev_stime;
3616 * This function gets called by the timer code, with HZ frequency.
3617 * We call it with interrupts disabled.
3619 * It also gets called by the fork code, when changing the parent's
3622 void scheduler_tick(void)
3624 int cpu = smp_processor_id();
3625 struct rq *rq = cpu_rq(cpu);
3626 struct task_struct *curr = rq->curr;
3630 raw_spin_lock(&rq->lock);
3631 update_rq_clock(rq);
3632 update_cpu_load_active(rq);
3633 curr->sched_class->task_tick(rq, curr, 0);
3634 raw_spin_unlock(&rq->lock);
3636 perf_event_task_tick();
3639 rq->idle_at_tick = idle_cpu(cpu);
3640 trigger_load_balance(rq, cpu);
3644 notrace unsigned long get_parent_ip(unsigned long addr)
3646 if (in_lock_functions(addr)) {
3647 addr = CALLER_ADDR2;
3648 if (in_lock_functions(addr))
3649 addr = CALLER_ADDR3;
3654 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3655 defined(CONFIG_PREEMPT_TRACER))
3657 void __kprobes add_preempt_count(int val)
3659 #ifdef CONFIG_DEBUG_PREEMPT
3663 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3666 preempt_count() += val;
3667 #ifdef CONFIG_DEBUG_PREEMPT
3669 * Spinlock count overflowing soon?
3671 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3674 if (preempt_count() == val)
3675 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3677 EXPORT_SYMBOL(add_preempt_count);
3679 void __kprobes sub_preempt_count(int val)
3681 #ifdef CONFIG_DEBUG_PREEMPT
3685 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3688 * Is the spinlock portion underflowing?
3690 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3691 !(preempt_count() & PREEMPT_MASK)))
3695 if (preempt_count() == val)
3696 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3697 preempt_count() -= val;
3699 EXPORT_SYMBOL(sub_preempt_count);
3704 * Print scheduling while atomic bug:
3706 static noinline void __schedule_bug(struct task_struct *prev)
3708 struct pt_regs *regs = get_irq_regs();
3710 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3711 prev->comm, prev->pid, preempt_count());
3713 debug_show_held_locks(prev);
3715 if (irqs_disabled())
3716 print_irqtrace_events(prev);
3725 * Various schedule()-time debugging checks and statistics:
3727 static inline void schedule_debug(struct task_struct *prev)
3730 * Test if we are atomic. Since do_exit() needs to call into
3731 * schedule() atomically, we ignore that path for now.
3732 * Otherwise, whine if we are scheduling when we should not be.
3734 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3735 __schedule_bug(prev);
3737 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3739 schedstat_inc(this_rq(), sched_count);
3740 #ifdef CONFIG_SCHEDSTATS
3741 if (unlikely(prev->lock_depth >= 0)) {
3742 schedstat_inc(this_rq(), bkl_count);
3743 schedstat_inc(prev, sched_info.bkl_count);
3748 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3751 update_rq_clock(rq);
3752 rq->skip_clock_update = 0;
3753 prev->sched_class->put_prev_task(rq, prev);
3757 * Pick up the highest-prio task:
3759 static inline struct task_struct *
3760 pick_next_task(struct rq *rq)
3762 const struct sched_class *class;
3763 struct task_struct *p;
3766 * Optimization: we know that if all tasks are in
3767 * the fair class we can call that function directly:
3769 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3770 p = fair_sched_class.pick_next_task(rq);
3775 for_each_class(class) {
3776 p = class->pick_next_task(rq);
3781 BUG(); /* the idle class will always have a runnable task */
3785 * schedule() is the main scheduler function.
3787 asmlinkage void __sched schedule(void)
3789 struct task_struct *prev, *next;
3790 unsigned long *switch_count;
3796 cpu = smp_processor_id();
3798 rcu_note_context_switch(cpu);
3801 release_kernel_lock(prev);
3802 need_resched_nonpreemptible:
3804 schedule_debug(prev);
3806 if (sched_feat(HRTICK))
3809 raw_spin_lock_irq(&rq->lock);
3810 clear_tsk_need_resched(prev);
3812 switch_count = &prev->nivcsw;
3813 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3814 if (unlikely(signal_pending_state(prev->state, prev))) {
3815 prev->state = TASK_RUNNING;
3818 * If a worker is going to sleep, notify and
3819 * ask workqueue whether it wants to wake up a
3820 * task to maintain concurrency. If so, wake
3823 if (prev->flags & PF_WQ_WORKER) {
3824 struct task_struct *to_wakeup;
3826 to_wakeup = wq_worker_sleeping(prev, cpu);
3828 try_to_wake_up_local(to_wakeup);
3830 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3832 switch_count = &prev->nvcsw;
3835 pre_schedule(rq, prev);
3837 if (unlikely(!rq->nr_running))
3838 idle_balance(cpu, rq);
3840 put_prev_task(rq, prev);
3841 next = pick_next_task(rq);
3843 if (likely(prev != next)) {
3844 sched_info_switch(prev, next);
3845 perf_event_task_sched_out(prev, next);
3851 context_switch(rq, prev, next); /* unlocks the rq */
3853 * The context switch have flipped the stack from under us
3854 * and restored the local variables which were saved when
3855 * this task called schedule() in the past. prev == current
3856 * is still correct, but it can be moved to another cpu/rq.
3858 cpu = smp_processor_id();
3861 raw_spin_unlock_irq(&rq->lock);
3865 if (unlikely(reacquire_kernel_lock(prev)))
3866 goto need_resched_nonpreemptible;
3868 preempt_enable_no_resched();
3872 EXPORT_SYMBOL(schedule);
3874 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3876 * Look out! "owner" is an entirely speculative pointer
3877 * access and not reliable.
3879 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3884 if (!sched_feat(OWNER_SPIN))
3887 #ifdef CONFIG_DEBUG_PAGEALLOC
3889 * Need to access the cpu field knowing that
3890 * DEBUG_PAGEALLOC could have unmapped it if
3891 * the mutex owner just released it and exited.
3893 if (probe_kernel_address(&owner->cpu, cpu))
3900 * Even if the access succeeded (likely case),
3901 * the cpu field may no longer be valid.
3903 if (cpu >= nr_cpumask_bits)
3907 * We need to validate that we can do a
3908 * get_cpu() and that we have the percpu area.
3910 if (!cpu_online(cpu))
3917 * Owner changed, break to re-assess state.
3919 if (lock->owner != owner) {
3921 * If the lock has switched to a different owner,
3922 * we likely have heavy contention. Return 0 to quit
3923 * optimistic spinning and not contend further:
3931 * Is that owner really running on that cpu?
3933 if (task_thread_info(rq->curr) != owner || need_resched())
3943 #ifdef CONFIG_PREEMPT
3945 * this is the entry point to schedule() from in-kernel preemption
3946 * off of preempt_enable. Kernel preemptions off return from interrupt
3947 * occur there and call schedule directly.
3949 asmlinkage void __sched notrace preempt_schedule(void)
3951 struct thread_info *ti = current_thread_info();
3954 * If there is a non-zero preempt_count or interrupts are disabled,
3955 * we do not want to preempt the current task. Just return..
3957 if (likely(ti->preempt_count || irqs_disabled()))
3961 add_preempt_count_notrace(PREEMPT_ACTIVE);
3963 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3966 * Check again in case we missed a preemption opportunity
3967 * between schedule and now.
3970 } while (need_resched());
3972 EXPORT_SYMBOL(preempt_schedule);
3975 * this is the entry point to schedule() from kernel preemption
3976 * off of irq context.
3977 * Note, that this is called and return with irqs disabled. This will
3978 * protect us against recursive calling from irq.
3980 asmlinkage void __sched preempt_schedule_irq(void)
3982 struct thread_info *ti = current_thread_info();
3984 /* Catch callers which need to be fixed */
3985 BUG_ON(ti->preempt_count || !irqs_disabled());
3988 add_preempt_count(PREEMPT_ACTIVE);
3991 local_irq_disable();
3992 sub_preempt_count(PREEMPT_ACTIVE);
3995 * Check again in case we missed a preemption opportunity
3996 * between schedule and now.
3999 } while (need_resched());
4002 #endif /* CONFIG_PREEMPT */
4004 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4007 return try_to_wake_up(curr->private, mode, wake_flags);
4009 EXPORT_SYMBOL(default_wake_function);
4012 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4013 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4014 * number) then we wake all the non-exclusive tasks and one exclusive task.
4016 * There are circumstances in which we can try to wake a task which has already
4017 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4018 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4020 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4021 int nr_exclusive, int wake_flags, void *key)
4023 wait_queue_t *curr, *next;
4025 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4026 unsigned flags = curr->flags;
4028 if (curr->func(curr, mode, wake_flags, key) &&
4029 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4035 * __wake_up - wake up threads blocked on a waitqueue.
4037 * @mode: which threads
4038 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4039 * @key: is directly passed to the wakeup function
4041 * It may be assumed that this function implies a write memory barrier before
4042 * changing the task state if and only if any tasks are woken up.
4044 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4045 int nr_exclusive, void *key)
4047 unsigned long flags;
4049 spin_lock_irqsave(&q->lock, flags);
4050 __wake_up_common(q, mode, nr_exclusive, 0, key);
4051 spin_unlock_irqrestore(&q->lock, flags);
4053 EXPORT_SYMBOL(__wake_up);
4056 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4058 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4060 __wake_up_common(q, mode, 1, 0, NULL);
4062 EXPORT_SYMBOL_GPL(__wake_up_locked);
4064 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4066 __wake_up_common(q, mode, 1, 0, key);
4070 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4072 * @mode: which threads
4073 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4074 * @key: opaque value to be passed to wakeup targets
4076 * The sync wakeup differs that the waker knows that it will schedule
4077 * away soon, so while the target thread will be woken up, it will not
4078 * be migrated to another CPU - ie. the two threads are 'synchronized'
4079 * with each other. This can prevent needless bouncing between CPUs.
4081 * On UP it can prevent extra preemption.
4083 * It may be assumed that this function implies a write memory barrier before
4084 * changing the task state if and only if any tasks are woken up.
4086 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4087 int nr_exclusive, void *key)
4089 unsigned long flags;
4090 int wake_flags = WF_SYNC;
4095 if (unlikely(!nr_exclusive))
4098 spin_lock_irqsave(&q->lock, flags);
4099 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4100 spin_unlock_irqrestore(&q->lock, flags);
4102 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4105 * __wake_up_sync - see __wake_up_sync_key()
4107 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4109 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4111 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4114 * complete: - signals a single thread waiting on this completion
4115 * @x: holds the state of this particular completion
4117 * This will wake up a single thread waiting on this completion. Threads will be
4118 * awakened in the same order in which they were queued.
4120 * See also complete_all(), wait_for_completion() and related routines.
4122 * It may be assumed that this function implies a write memory barrier before
4123 * changing the task state if and only if any tasks are woken up.
4125 void complete(struct completion *x)
4127 unsigned long flags;
4129 spin_lock_irqsave(&x->wait.lock, flags);
4131 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4132 spin_unlock_irqrestore(&x->wait.lock, flags);
4134 EXPORT_SYMBOL(complete);
4137 * complete_all: - signals all threads waiting on this completion
4138 * @x: holds the state of this particular completion
4140 * This will wake up all threads waiting on this particular completion event.
4142 * It may be assumed that this function implies a write memory barrier before
4143 * changing the task state if and only if any tasks are woken up.
4145 void complete_all(struct completion *x)
4147 unsigned long flags;
4149 spin_lock_irqsave(&x->wait.lock, flags);
4150 x->done += UINT_MAX/2;
4151 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4152 spin_unlock_irqrestore(&x->wait.lock, flags);
4154 EXPORT_SYMBOL(complete_all);
4156 static inline long __sched
4157 do_wait_for_common(struct completion *x, long timeout, int state)
4160 DECLARE_WAITQUEUE(wait, current);
4162 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4164 if (signal_pending_state(state, current)) {
4165 timeout = -ERESTARTSYS;
4168 __set_current_state(state);
4169 spin_unlock_irq(&x->wait.lock);
4170 timeout = schedule_timeout(timeout);
4171 spin_lock_irq(&x->wait.lock);
4172 } while (!x->done && timeout);
4173 __remove_wait_queue(&x->wait, &wait);
4178 return timeout ?: 1;
4182 wait_for_common(struct completion *x, long timeout, int state)
4186 spin_lock_irq(&x->wait.lock);
4187 timeout = do_wait_for_common(x, timeout, state);
4188 spin_unlock_irq(&x->wait.lock);
4193 * wait_for_completion: - waits for completion of a task
4194 * @x: holds the state of this particular completion
4196 * This waits to be signaled for completion of a specific task. It is NOT
4197 * interruptible and there is no timeout.
4199 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4200 * and interrupt capability. Also see complete().
4202 void __sched wait_for_completion(struct completion *x)
4204 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4206 EXPORT_SYMBOL(wait_for_completion);
4209 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4210 * @x: holds the state of this particular completion
4211 * @timeout: timeout value in jiffies
4213 * This waits for either a completion of a specific task to be signaled or for a
4214 * specified timeout to expire. The timeout is in jiffies. It is not
4217 unsigned long __sched
4218 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4220 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4222 EXPORT_SYMBOL(wait_for_completion_timeout);
4225 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4226 * @x: holds the state of this particular completion
4228 * This waits for completion of a specific task to be signaled. It is
4231 int __sched wait_for_completion_interruptible(struct completion *x)
4233 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4234 if (t == -ERESTARTSYS)
4238 EXPORT_SYMBOL(wait_for_completion_interruptible);
4241 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4242 * @x: holds the state of this particular completion
4243 * @timeout: timeout value in jiffies
4245 * This waits for either a completion of a specific task to be signaled or for a
4246 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4248 unsigned long __sched
4249 wait_for_completion_interruptible_timeout(struct completion *x,
4250 unsigned long timeout)
4252 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4254 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4257 * wait_for_completion_killable: - waits for completion of a task (killable)
4258 * @x: holds the state of this particular completion
4260 * This waits to be signaled for completion of a specific task. It can be
4261 * interrupted by a kill signal.
4263 int __sched wait_for_completion_killable(struct completion *x)
4265 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4266 if (t == -ERESTARTSYS)
4270 EXPORT_SYMBOL(wait_for_completion_killable);
4273 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4274 * @x: holds the state of this particular completion
4275 * @timeout: timeout value in jiffies
4277 * This waits for either a completion of a specific task to be
4278 * signaled or for a specified timeout to expire. It can be
4279 * interrupted by a kill signal. The timeout is in jiffies.
4281 unsigned long __sched
4282 wait_for_completion_killable_timeout(struct completion *x,
4283 unsigned long timeout)
4285 return wait_for_common(x, timeout, TASK_KILLABLE);
4287 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4290 * try_wait_for_completion - try to decrement a completion without blocking
4291 * @x: completion structure
4293 * Returns: 0 if a decrement cannot be done without blocking
4294 * 1 if a decrement succeeded.
4296 * If a completion is being used as a counting completion,
4297 * attempt to decrement the counter without blocking. This
4298 * enables us to avoid waiting if the resource the completion
4299 * is protecting is not available.
4301 bool try_wait_for_completion(struct completion *x)
4303 unsigned long flags;
4306 spin_lock_irqsave(&x->wait.lock, flags);
4311 spin_unlock_irqrestore(&x->wait.lock, flags);
4314 EXPORT_SYMBOL(try_wait_for_completion);
4317 * completion_done - Test to see if a completion has any waiters
4318 * @x: completion structure
4320 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4321 * 1 if there are no waiters.
4324 bool completion_done(struct completion *x)
4326 unsigned long flags;
4329 spin_lock_irqsave(&x->wait.lock, flags);
4332 spin_unlock_irqrestore(&x->wait.lock, flags);
4335 EXPORT_SYMBOL(completion_done);
4338 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4340 unsigned long flags;
4343 init_waitqueue_entry(&wait, current);
4345 __set_current_state(state);
4347 spin_lock_irqsave(&q->lock, flags);
4348 __add_wait_queue(q, &wait);
4349 spin_unlock(&q->lock);
4350 timeout = schedule_timeout(timeout);
4351 spin_lock_irq(&q->lock);
4352 __remove_wait_queue(q, &wait);
4353 spin_unlock_irqrestore(&q->lock, flags);
4358 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4360 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4362 EXPORT_SYMBOL(interruptible_sleep_on);
4365 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4367 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4369 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4371 void __sched sleep_on(wait_queue_head_t *q)
4373 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4375 EXPORT_SYMBOL(sleep_on);
4377 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4379 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4381 EXPORT_SYMBOL(sleep_on_timeout);
4383 #ifdef CONFIG_RT_MUTEXES
4386 * rt_mutex_setprio - set the current priority of a task
4388 * @prio: prio value (kernel-internal form)
4390 * This function changes the 'effective' priority of a task. It does
4391 * not touch ->normal_prio like __setscheduler().
4393 * Used by the rt_mutex code to implement priority inheritance logic.
4395 void rt_mutex_setprio(struct task_struct *p, int prio)
4397 unsigned long flags;
4398 int oldprio, on_rq, running;
4400 const struct sched_class *prev_class;
4402 BUG_ON(prio < 0 || prio > MAX_PRIO);
4404 rq = task_rq_lock(p, &flags);
4406 trace_sched_pi_setprio(p, prio);
4408 prev_class = p->sched_class;
4409 on_rq = p->se.on_rq;
4410 running = task_current(rq, p);
4412 dequeue_task(rq, p, 0);
4414 p->sched_class->put_prev_task(rq, p);
4417 p->sched_class = &rt_sched_class;
4419 p->sched_class = &fair_sched_class;
4424 p->sched_class->set_curr_task(rq);
4426 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4428 check_class_changed(rq, p, prev_class, oldprio, running);
4430 task_rq_unlock(rq, &flags);
4435 void set_user_nice(struct task_struct *p, long nice)
4437 int old_prio, delta, on_rq;
4438 unsigned long flags;
4441 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4444 * We have to be careful, if called from sys_setpriority(),
4445 * the task might be in the middle of scheduling on another CPU.
4447 rq = task_rq_lock(p, &flags);
4449 * The RT priorities are set via sched_setscheduler(), but we still
4450 * allow the 'normal' nice value to be set - but as expected
4451 * it wont have any effect on scheduling until the task is
4452 * SCHED_FIFO/SCHED_RR:
4454 if (task_has_rt_policy(p)) {
4455 p->static_prio = NICE_TO_PRIO(nice);
4458 on_rq = p->se.on_rq;
4460 dequeue_task(rq, p, 0);
4462 p->static_prio = NICE_TO_PRIO(nice);
4465 p->prio = effective_prio(p);
4466 delta = p->prio - old_prio;
4469 enqueue_task(rq, p, 0);
4471 * If the task increased its priority or is running and
4472 * lowered its priority, then reschedule its CPU:
4474 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4475 resched_task(rq->curr);
4478 task_rq_unlock(rq, &flags);
4480 EXPORT_SYMBOL(set_user_nice);
4483 * can_nice - check if a task can reduce its nice value
4487 int can_nice(const struct task_struct *p, const int nice)
4489 /* convert nice value [19,-20] to rlimit style value [1,40] */
4490 int nice_rlim = 20 - nice;
4492 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4493 capable(CAP_SYS_NICE));
4496 #ifdef __ARCH_WANT_SYS_NICE
4499 * sys_nice - change the priority of the current process.
4500 * @increment: priority increment
4502 * sys_setpriority is a more generic, but much slower function that
4503 * does similar things.
4505 SYSCALL_DEFINE1(nice, int, increment)
4510 * Setpriority might change our priority at the same moment.
4511 * We don't have to worry. Conceptually one call occurs first
4512 * and we have a single winner.
4514 if (increment < -40)
4519 nice = TASK_NICE(current) + increment;
4525 if (increment < 0 && !can_nice(current, nice))
4528 retval = security_task_setnice(current, nice);
4532 set_user_nice(current, nice);
4539 * task_prio - return the priority value of a given task.
4540 * @p: the task in question.
4542 * This is the priority value as seen by users in /proc.
4543 * RT tasks are offset by -200. Normal tasks are centered
4544 * around 0, value goes from -16 to +15.
4546 int task_prio(const struct task_struct *p)
4548 return p->prio - MAX_RT_PRIO;
4552 * task_nice - return the nice value of a given task.
4553 * @p: the task in question.
4555 int task_nice(const struct task_struct *p)
4557 return TASK_NICE(p);
4559 EXPORT_SYMBOL(task_nice);
4562 * idle_cpu - is a given cpu idle currently?
4563 * @cpu: the processor in question.
4565 int idle_cpu(int cpu)
4567 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4571 * idle_task - return the idle task for a given cpu.
4572 * @cpu: the processor in question.
4574 struct task_struct *idle_task(int cpu)
4576 return cpu_rq(cpu)->idle;
4580 * find_process_by_pid - find a process with a matching PID value.
4581 * @pid: the pid in question.
4583 static struct task_struct *find_process_by_pid(pid_t pid)
4585 return pid ? find_task_by_vpid(pid) : current;
4588 /* Actually do priority change: must hold rq lock. */
4590 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4592 BUG_ON(p->se.on_rq);
4595 p->rt_priority = prio;
4596 p->normal_prio = normal_prio(p);
4597 /* we are holding p->pi_lock already */
4598 p->prio = rt_mutex_getprio(p);
4599 if (rt_prio(p->prio))
4600 p->sched_class = &rt_sched_class;
4602 p->sched_class = &fair_sched_class;
4607 * check the target process has a UID that matches the current process's
4609 static bool check_same_owner(struct task_struct *p)
4611 const struct cred *cred = current_cred(), *pcred;
4615 pcred = __task_cred(p);
4616 match = (cred->euid == pcred->euid ||
4617 cred->euid == pcred->uid);
4622 static int __sched_setscheduler(struct task_struct *p, int policy,
4623 const struct sched_param *param, bool user)
4625 int retval, oldprio, oldpolicy = -1, on_rq, running;
4626 unsigned long flags;
4627 const struct sched_class *prev_class;
4631 /* may grab non-irq protected spin_locks */
4632 BUG_ON(in_interrupt());
4634 /* double check policy once rq lock held */
4636 reset_on_fork = p->sched_reset_on_fork;
4637 policy = oldpolicy = p->policy;
4639 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4640 policy &= ~SCHED_RESET_ON_FORK;
4642 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4643 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4644 policy != SCHED_IDLE)
4649 * Valid priorities for SCHED_FIFO and SCHED_RR are
4650 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4651 * SCHED_BATCH and SCHED_IDLE is 0.
4653 if (param->sched_priority < 0 ||
4654 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4655 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4657 if (rt_policy(policy) != (param->sched_priority != 0))
4661 * Allow unprivileged RT tasks to decrease priority:
4663 if (user && !capable(CAP_SYS_NICE)) {
4664 if (rt_policy(policy)) {
4665 unsigned long rlim_rtprio =
4666 task_rlimit(p, RLIMIT_RTPRIO);
4668 /* can't set/change the rt policy */
4669 if (policy != p->policy && !rlim_rtprio)
4672 /* can't increase priority */
4673 if (param->sched_priority > p->rt_priority &&
4674 param->sched_priority > rlim_rtprio)
4678 * Like positive nice levels, dont allow tasks to
4679 * move out of SCHED_IDLE either:
4681 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4684 /* can't change other user's priorities */
4685 if (!check_same_owner(p))
4688 /* Normal users shall not reset the sched_reset_on_fork flag */
4689 if (p->sched_reset_on_fork && !reset_on_fork)
4694 retval = security_task_setscheduler(p);
4700 * make sure no PI-waiters arrive (or leave) while we are
4701 * changing the priority of the task:
4703 raw_spin_lock_irqsave(&p->pi_lock, flags);
4705 * To be able to change p->policy safely, the apropriate
4706 * runqueue lock must be held.
4708 rq = __task_rq_lock(p);
4711 * Changing the policy of the stop threads its a very bad idea
4713 if (p == rq->stop) {
4714 __task_rq_unlock(rq);
4715 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4719 #ifdef CONFIG_RT_GROUP_SCHED
4722 * Do not allow realtime tasks into groups that have no runtime
4725 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4726 task_group(p)->rt_bandwidth.rt_runtime == 0) {
4727 __task_rq_unlock(rq);
4728 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4734 /* recheck policy now with rq lock held */
4735 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4736 policy = oldpolicy = -1;
4737 __task_rq_unlock(rq);
4738 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4741 on_rq = p->se.on_rq;
4742 running = task_current(rq, p);
4744 deactivate_task(rq, p, 0);
4746 p->sched_class->put_prev_task(rq, p);
4748 p->sched_reset_on_fork = reset_on_fork;
4751 prev_class = p->sched_class;
4752 __setscheduler(rq, p, policy, param->sched_priority);
4755 p->sched_class->set_curr_task(rq);
4757 activate_task(rq, p, 0);
4759 check_class_changed(rq, p, prev_class, oldprio, running);
4761 __task_rq_unlock(rq);
4762 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4764 rt_mutex_adjust_pi(p);
4770 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4771 * @p: the task in question.
4772 * @policy: new policy.
4773 * @param: structure containing the new RT priority.
4775 * NOTE that the task may be already dead.
4777 int sched_setscheduler(struct task_struct *p, int policy,
4778 const struct sched_param *param)
4780 return __sched_setscheduler(p, policy, param, true);
4782 EXPORT_SYMBOL_GPL(sched_setscheduler);
4785 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4786 * @p: the task in question.
4787 * @policy: new policy.
4788 * @param: structure containing the new RT priority.
4790 * Just like sched_setscheduler, only don't bother checking if the
4791 * current context has permission. For example, this is needed in
4792 * stop_machine(): we create temporary high priority worker threads,
4793 * but our caller might not have that capability.
4795 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4796 const struct sched_param *param)
4798 return __sched_setscheduler(p, policy, param, false);
4802 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4804 struct sched_param lparam;
4805 struct task_struct *p;
4808 if (!param || pid < 0)
4810 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4815 p = find_process_by_pid(pid);
4817 retval = sched_setscheduler(p, policy, &lparam);
4824 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4825 * @pid: the pid in question.
4826 * @policy: new policy.
4827 * @param: structure containing the new RT priority.
4829 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4830 struct sched_param __user *, param)
4832 /* negative values for policy are not valid */
4836 return do_sched_setscheduler(pid, policy, param);
4840 * sys_sched_setparam - set/change the RT priority of a thread
4841 * @pid: the pid in question.
4842 * @param: structure containing the new RT priority.
4844 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4846 return do_sched_setscheduler(pid, -1, param);
4850 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4851 * @pid: the pid in question.
4853 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4855 struct task_struct *p;
4863 p = find_process_by_pid(pid);
4865 retval = security_task_getscheduler(p);
4868 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4875 * sys_sched_getparam - get the RT priority of a thread
4876 * @pid: the pid in question.
4877 * @param: structure containing the RT priority.
4879 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4881 struct sched_param lp;
4882 struct task_struct *p;
4885 if (!param || pid < 0)
4889 p = find_process_by_pid(pid);
4894 retval = security_task_getscheduler(p);
4898 lp.sched_priority = p->rt_priority;
4902 * This one might sleep, we cannot do it with a spinlock held ...
4904 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4913 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4915 cpumask_var_t cpus_allowed, new_mask;
4916 struct task_struct *p;
4922 p = find_process_by_pid(pid);
4929 /* Prevent p going away */
4933 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4937 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4939 goto out_free_cpus_allowed;
4942 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4945 retval = security_task_setscheduler(p);
4949 cpuset_cpus_allowed(p, cpus_allowed);
4950 cpumask_and(new_mask, in_mask, cpus_allowed);
4952 retval = set_cpus_allowed_ptr(p, new_mask);
4955 cpuset_cpus_allowed(p, cpus_allowed);
4956 if (!cpumask_subset(new_mask, cpus_allowed)) {
4958 * We must have raced with a concurrent cpuset
4959 * update. Just reset the cpus_allowed to the
4960 * cpuset's cpus_allowed
4962 cpumask_copy(new_mask, cpus_allowed);
4967 free_cpumask_var(new_mask);
4968 out_free_cpus_allowed:
4969 free_cpumask_var(cpus_allowed);
4976 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4977 struct cpumask *new_mask)
4979 if (len < cpumask_size())
4980 cpumask_clear(new_mask);
4981 else if (len > cpumask_size())
4982 len = cpumask_size();
4984 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4988 * sys_sched_setaffinity - set the cpu affinity of a process
4989 * @pid: pid of the process
4990 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4991 * @user_mask_ptr: user-space pointer to the new cpu mask
4993 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4994 unsigned long __user *, user_mask_ptr)
4996 cpumask_var_t new_mask;
4999 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5002 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5004 retval = sched_setaffinity(pid, new_mask);
5005 free_cpumask_var(new_mask);
5009 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5011 struct task_struct *p;
5012 unsigned long flags;
5020 p = find_process_by_pid(pid);
5024 retval = security_task_getscheduler(p);
5028 rq = task_rq_lock(p, &flags);
5029 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5030 task_rq_unlock(rq, &flags);
5040 * sys_sched_getaffinity - get the cpu affinity of a process
5041 * @pid: pid of the process
5042 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5043 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5045 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5046 unsigned long __user *, user_mask_ptr)
5051 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5053 if (len & (sizeof(unsigned long)-1))
5056 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5059 ret = sched_getaffinity(pid, mask);
5061 size_t retlen = min_t(size_t, len, cpumask_size());
5063 if (copy_to_user(user_mask_ptr, mask, retlen))
5068 free_cpumask_var(mask);
5074 * sys_sched_yield - yield the current processor to other threads.
5076 * This function yields the current CPU to other tasks. If there are no
5077 * other threads running on this CPU then this function will return.
5079 SYSCALL_DEFINE0(sched_yield)
5081 struct rq *rq = this_rq_lock();
5083 schedstat_inc(rq, yld_count);
5084 current->sched_class->yield_task(rq);
5087 * Since we are going to call schedule() anyway, there's
5088 * no need to preempt or enable interrupts:
5090 __release(rq->lock);
5091 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5092 do_raw_spin_unlock(&rq->lock);
5093 preempt_enable_no_resched();
5100 static inline int should_resched(void)
5102 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5105 static void __cond_resched(void)
5107 add_preempt_count(PREEMPT_ACTIVE);
5109 sub_preempt_count(PREEMPT_ACTIVE);
5112 int __sched _cond_resched(void)
5114 if (should_resched()) {
5120 EXPORT_SYMBOL(_cond_resched);
5123 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5124 * call schedule, and on return reacquire the lock.
5126 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5127 * operations here to prevent schedule() from being called twice (once via
5128 * spin_unlock(), once by hand).
5130 int __cond_resched_lock(spinlock_t *lock)
5132 int resched = should_resched();
5135 lockdep_assert_held(lock);
5137 if (spin_needbreak(lock) || resched) {
5148 EXPORT_SYMBOL(__cond_resched_lock);
5150 int __sched __cond_resched_softirq(void)
5152 BUG_ON(!in_softirq());
5154 if (should_resched()) {
5162 EXPORT_SYMBOL(__cond_resched_softirq);
5165 * yield - yield the current processor to other threads.
5167 * This is a shortcut for kernel-space yielding - it marks the
5168 * thread runnable and calls sys_sched_yield().
5170 void __sched yield(void)
5172 set_current_state(TASK_RUNNING);
5175 EXPORT_SYMBOL(yield);
5178 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5179 * that process accounting knows that this is a task in IO wait state.
5181 void __sched io_schedule(void)
5183 struct rq *rq = raw_rq();
5185 delayacct_blkio_start();
5186 atomic_inc(&rq->nr_iowait);
5187 current->in_iowait = 1;
5189 current->in_iowait = 0;
5190 atomic_dec(&rq->nr_iowait);
5191 delayacct_blkio_end();
5193 EXPORT_SYMBOL(io_schedule);
5195 long __sched io_schedule_timeout(long timeout)
5197 struct rq *rq = raw_rq();
5200 delayacct_blkio_start();
5201 atomic_inc(&rq->nr_iowait);
5202 current->in_iowait = 1;
5203 ret = schedule_timeout(timeout);
5204 current->in_iowait = 0;
5205 atomic_dec(&rq->nr_iowait);
5206 delayacct_blkio_end();
5211 * sys_sched_get_priority_max - return maximum RT priority.
5212 * @policy: scheduling class.
5214 * this syscall returns the maximum rt_priority that can be used
5215 * by a given scheduling class.
5217 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5224 ret = MAX_USER_RT_PRIO-1;
5236 * sys_sched_get_priority_min - return minimum RT priority.
5237 * @policy: scheduling class.
5239 * this syscall returns the minimum rt_priority that can be used
5240 * by a given scheduling class.
5242 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5260 * sys_sched_rr_get_interval - return the default timeslice of a process.
5261 * @pid: pid of the process.
5262 * @interval: userspace pointer to the timeslice value.
5264 * this syscall writes the default timeslice value of a given process
5265 * into the user-space timespec buffer. A value of '0' means infinity.
5267 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5268 struct timespec __user *, interval)
5270 struct task_struct *p;
5271 unsigned int time_slice;
5272 unsigned long flags;
5282 p = find_process_by_pid(pid);
5286 retval = security_task_getscheduler(p);
5290 rq = task_rq_lock(p, &flags);
5291 time_slice = p->sched_class->get_rr_interval(rq, p);
5292 task_rq_unlock(rq, &flags);
5295 jiffies_to_timespec(time_slice, &t);
5296 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5304 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5306 void sched_show_task(struct task_struct *p)
5308 unsigned long free = 0;
5311 state = p->state ? __ffs(p->state) + 1 : 0;
5312 printk(KERN_INFO "%-13.13s %c", p->comm,
5313 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5314 #if BITS_PER_LONG == 32
5315 if (state == TASK_RUNNING)
5316 printk(KERN_CONT " running ");
5318 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5320 if (state == TASK_RUNNING)
5321 printk(KERN_CONT " running task ");
5323 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5325 #ifdef CONFIG_DEBUG_STACK_USAGE
5326 free = stack_not_used(p);
5328 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5329 task_pid_nr(p), task_pid_nr(p->real_parent),
5330 (unsigned long)task_thread_info(p)->flags);
5332 show_stack(p, NULL);
5335 void show_state_filter(unsigned long state_filter)
5337 struct task_struct *g, *p;
5339 #if BITS_PER_LONG == 32
5341 " task PC stack pid father\n");
5344 " task PC stack pid father\n");
5346 read_lock(&tasklist_lock);
5347 do_each_thread(g, p) {
5349 * reset the NMI-timeout, listing all files on a slow
5350 * console might take alot of time:
5352 touch_nmi_watchdog();
5353 if (!state_filter || (p->state & state_filter))
5355 } while_each_thread(g, p);
5357 touch_all_softlockup_watchdogs();
5359 #ifdef CONFIG_SCHED_DEBUG
5360 sysrq_sched_debug_show();
5362 read_unlock(&tasklist_lock);
5364 * Only show locks if all tasks are dumped:
5367 debug_show_all_locks();
5370 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5372 idle->sched_class = &idle_sched_class;
5376 * init_idle - set up an idle thread for a given CPU
5377 * @idle: task in question
5378 * @cpu: cpu the idle task belongs to
5380 * NOTE: this function does not set the idle thread's NEED_RESCHED
5381 * flag, to make booting more robust.
5383 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5385 struct rq *rq = cpu_rq(cpu);
5386 unsigned long flags;
5388 raw_spin_lock_irqsave(&rq->lock, flags);
5391 idle->state = TASK_RUNNING;
5392 idle->se.exec_start = sched_clock();
5394 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5396 * We're having a chicken and egg problem, even though we are
5397 * holding rq->lock, the cpu isn't yet set to this cpu so the
5398 * lockdep check in task_group() will fail.
5400 * Similar case to sched_fork(). / Alternatively we could
5401 * use task_rq_lock() here and obtain the other rq->lock.
5406 __set_task_cpu(idle, cpu);
5409 rq->curr = rq->idle = idle;
5410 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5413 raw_spin_unlock_irqrestore(&rq->lock, flags);
5415 /* Set the preempt count _outside_ the spinlocks! */
5416 #if defined(CONFIG_PREEMPT)
5417 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5419 task_thread_info(idle)->preempt_count = 0;
5422 * The idle tasks have their own, simple scheduling class:
5424 idle->sched_class = &idle_sched_class;
5425 ftrace_graph_init_task(idle);
5429 * In a system that switches off the HZ timer nohz_cpu_mask
5430 * indicates which cpus entered this state. This is used
5431 * in the rcu update to wait only for active cpus. For system
5432 * which do not switch off the HZ timer nohz_cpu_mask should
5433 * always be CPU_BITS_NONE.
5435 cpumask_var_t nohz_cpu_mask;
5438 * Increase the granularity value when there are more CPUs,
5439 * because with more CPUs the 'effective latency' as visible
5440 * to users decreases. But the relationship is not linear,
5441 * so pick a second-best guess by going with the log2 of the
5444 * This idea comes from the SD scheduler of Con Kolivas:
5446 static int get_update_sysctl_factor(void)
5448 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5449 unsigned int factor;
5451 switch (sysctl_sched_tunable_scaling) {
5452 case SCHED_TUNABLESCALING_NONE:
5455 case SCHED_TUNABLESCALING_LINEAR:
5458 case SCHED_TUNABLESCALING_LOG:
5460 factor = 1 + ilog2(cpus);
5467 static void update_sysctl(void)
5469 unsigned int factor = get_update_sysctl_factor();
5471 #define SET_SYSCTL(name) \
5472 (sysctl_##name = (factor) * normalized_sysctl_##name)
5473 SET_SYSCTL(sched_min_granularity);
5474 SET_SYSCTL(sched_latency);
5475 SET_SYSCTL(sched_wakeup_granularity);
5479 static inline void sched_init_granularity(void)
5486 * This is how migration works:
5488 * 1) we invoke migration_cpu_stop() on the target CPU using
5490 * 2) stopper starts to run (implicitly forcing the migrated thread
5492 * 3) it checks whether the migrated task is still in the wrong runqueue.
5493 * 4) if it's in the wrong runqueue then the migration thread removes
5494 * it and puts it into the right queue.
5495 * 5) stopper completes and stop_one_cpu() returns and the migration
5500 * Change a given task's CPU affinity. Migrate the thread to a
5501 * proper CPU and schedule it away if the CPU it's executing on
5502 * is removed from the allowed bitmask.
5504 * NOTE: the caller must have a valid reference to the task, the
5505 * task must not exit() & deallocate itself prematurely. The
5506 * call is not atomic; no spinlocks may be held.
5508 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5510 unsigned long flags;
5512 unsigned int dest_cpu;
5516 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5517 * drop the rq->lock and still rely on ->cpus_allowed.
5520 while (task_is_waking(p))
5522 rq = task_rq_lock(p, &flags);
5523 if (task_is_waking(p)) {
5524 task_rq_unlock(rq, &flags);
5528 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5533 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5534 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5539 if (p->sched_class->set_cpus_allowed)
5540 p->sched_class->set_cpus_allowed(p, new_mask);
5542 cpumask_copy(&p->cpus_allowed, new_mask);
5543 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5546 /* Can the task run on the task's current CPU? If so, we're done */
5547 if (cpumask_test_cpu(task_cpu(p), new_mask))
5550 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5551 if (migrate_task(p, dest_cpu)) {
5552 struct migration_arg arg = { p, dest_cpu };
5553 /* Need help from migration thread: drop lock and wait. */
5554 task_rq_unlock(rq, &flags);
5555 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5556 tlb_migrate_finish(p->mm);
5560 task_rq_unlock(rq, &flags);
5564 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5567 * Move (not current) task off this cpu, onto dest cpu. We're doing
5568 * this because either it can't run here any more (set_cpus_allowed()
5569 * away from this CPU, or CPU going down), or because we're
5570 * attempting to rebalance this task on exec (sched_exec).
5572 * So we race with normal scheduler movements, but that's OK, as long
5573 * as the task is no longer on this CPU.
5575 * Returns non-zero if task was successfully migrated.
5577 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5579 struct rq *rq_dest, *rq_src;
5582 if (unlikely(!cpu_active(dest_cpu)))
5585 rq_src = cpu_rq(src_cpu);
5586 rq_dest = cpu_rq(dest_cpu);
5588 double_rq_lock(rq_src, rq_dest);
5589 /* Already moved. */
5590 if (task_cpu(p) != src_cpu)
5592 /* Affinity changed (again). */
5593 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5597 * If we're not on a rq, the next wake-up will ensure we're
5601 deactivate_task(rq_src, p, 0);
5602 set_task_cpu(p, dest_cpu);
5603 activate_task(rq_dest, p, 0);
5604 check_preempt_curr(rq_dest, p, 0);
5609 double_rq_unlock(rq_src, rq_dest);
5614 * migration_cpu_stop - this will be executed by a highprio stopper thread
5615 * and performs thread migration by bumping thread off CPU then
5616 * 'pushing' onto another runqueue.
5618 static int migration_cpu_stop(void *data)
5620 struct migration_arg *arg = data;
5623 * The original target cpu might have gone down and we might
5624 * be on another cpu but it doesn't matter.
5626 local_irq_disable();
5627 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5632 #ifdef CONFIG_HOTPLUG_CPU
5635 * Ensures that the idle task is using init_mm right before its cpu goes
5638 void idle_task_exit(void)
5640 struct mm_struct *mm = current->active_mm;
5642 BUG_ON(cpu_online(smp_processor_id()));
5645 switch_mm(mm, &init_mm, current);
5650 * While a dead CPU has no uninterruptible tasks queued at this point,
5651 * it might still have a nonzero ->nr_uninterruptible counter, because
5652 * for performance reasons the counter is not stricly tracking tasks to
5653 * their home CPUs. So we just add the counter to another CPU's counter,
5654 * to keep the global sum constant after CPU-down:
5656 static void migrate_nr_uninterruptible(struct rq *rq_src)
5658 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5660 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5661 rq_src->nr_uninterruptible = 0;
5665 * remove the tasks which were accounted by rq from calc_load_tasks.
5667 static void calc_global_load_remove(struct rq *rq)
5669 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5670 rq->calc_load_active = 0;
5674 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5675 * try_to_wake_up()->select_task_rq().
5677 * Called with rq->lock held even though we'er in stop_machine() and
5678 * there's no concurrency possible, we hold the required locks anyway
5679 * because of lock validation efforts.
5681 static void migrate_tasks(unsigned int dead_cpu)
5683 struct rq *rq = cpu_rq(dead_cpu);
5684 struct task_struct *next, *stop = rq->stop;
5688 * Fudge the rq selection such that the below task selection loop
5689 * doesn't get stuck on the currently eligible stop task.
5691 * We're currently inside stop_machine() and the rq is either stuck
5692 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5693 * either way we should never end up calling schedule() until we're
5700 * There's this thread running, bail when that's the only
5703 if (rq->nr_running == 1)
5706 next = pick_next_task(rq);
5708 next->sched_class->put_prev_task(rq, next);
5710 /* Find suitable destination for @next, with force if needed. */
5711 dest_cpu = select_fallback_rq(dead_cpu, next);
5712 raw_spin_unlock(&rq->lock);
5714 __migrate_task(next, dead_cpu, dest_cpu);
5716 raw_spin_lock(&rq->lock);
5722 #endif /* CONFIG_HOTPLUG_CPU */
5724 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5726 static struct ctl_table sd_ctl_dir[] = {
5728 .procname = "sched_domain",
5734 static struct ctl_table sd_ctl_root[] = {
5736 .procname = "kernel",
5738 .child = sd_ctl_dir,
5743 static struct ctl_table *sd_alloc_ctl_entry(int n)
5745 struct ctl_table *entry =
5746 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5751 static void sd_free_ctl_entry(struct ctl_table **tablep)
5753 struct ctl_table *entry;
5756 * In the intermediate directories, both the child directory and
5757 * procname are dynamically allocated and could fail but the mode
5758 * will always be set. In the lowest directory the names are
5759 * static strings and all have proc handlers.
5761 for (entry = *tablep; entry->mode; entry++) {
5763 sd_free_ctl_entry(&entry->child);
5764 if (entry->proc_handler == NULL)
5765 kfree(entry->procname);
5773 set_table_entry(struct ctl_table *entry,
5774 const char *procname, void *data, int maxlen,
5775 mode_t mode, proc_handler *proc_handler)
5777 entry->procname = procname;
5779 entry->maxlen = maxlen;
5781 entry->proc_handler = proc_handler;
5784 static struct ctl_table *
5785 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5787 struct ctl_table *table = sd_alloc_ctl_entry(13);
5792 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5793 sizeof(long), 0644, proc_doulongvec_minmax);
5794 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5795 sizeof(long), 0644, proc_doulongvec_minmax);
5796 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5797 sizeof(int), 0644, proc_dointvec_minmax);
5798 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5799 sizeof(int), 0644, proc_dointvec_minmax);
5800 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5801 sizeof(int), 0644, proc_dointvec_minmax);
5802 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5803 sizeof(int), 0644, proc_dointvec_minmax);
5804 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5805 sizeof(int), 0644, proc_dointvec_minmax);
5806 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5807 sizeof(int), 0644, proc_dointvec_minmax);
5808 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5809 sizeof(int), 0644, proc_dointvec_minmax);
5810 set_table_entry(&table[9], "cache_nice_tries",
5811 &sd->cache_nice_tries,
5812 sizeof(int), 0644, proc_dointvec_minmax);
5813 set_table_entry(&table[10], "flags", &sd->flags,
5814 sizeof(int), 0644, proc_dointvec_minmax);
5815 set_table_entry(&table[11], "name", sd->name,
5816 CORENAME_MAX_SIZE, 0444, proc_dostring);
5817 /* &table[12] is terminator */
5822 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5824 struct ctl_table *entry, *table;
5825 struct sched_domain *sd;
5826 int domain_num = 0, i;
5829 for_each_domain(cpu, sd)
5831 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5836 for_each_domain(cpu, sd) {
5837 snprintf(buf, 32, "domain%d", i);
5838 entry->procname = kstrdup(buf, GFP_KERNEL);
5840 entry->child = sd_alloc_ctl_domain_table(sd);
5847 static struct ctl_table_header *sd_sysctl_header;
5848 static void register_sched_domain_sysctl(void)
5850 int i, cpu_num = num_possible_cpus();
5851 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5854 WARN_ON(sd_ctl_dir[0].child);
5855 sd_ctl_dir[0].child = entry;
5860 for_each_possible_cpu(i) {
5861 snprintf(buf, 32, "cpu%d", i);
5862 entry->procname = kstrdup(buf, GFP_KERNEL);
5864 entry->child = sd_alloc_ctl_cpu_table(i);
5868 WARN_ON(sd_sysctl_header);
5869 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5872 /* may be called multiple times per register */
5873 static void unregister_sched_domain_sysctl(void)
5875 if (sd_sysctl_header)
5876 unregister_sysctl_table(sd_sysctl_header);
5877 sd_sysctl_header = NULL;
5878 if (sd_ctl_dir[0].child)
5879 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5882 static void register_sched_domain_sysctl(void)
5885 static void unregister_sched_domain_sysctl(void)
5890 static void set_rq_online(struct rq *rq)
5893 const struct sched_class *class;
5895 cpumask_set_cpu(rq->cpu, rq->rd->online);
5898 for_each_class(class) {
5899 if (class->rq_online)
5900 class->rq_online(rq);
5905 static void set_rq_offline(struct rq *rq)
5908 const struct sched_class *class;
5910 for_each_class(class) {
5911 if (class->rq_offline)
5912 class->rq_offline(rq);
5915 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5921 * migration_call - callback that gets triggered when a CPU is added.
5922 * Here we can start up the necessary migration thread for the new CPU.
5924 static int __cpuinit
5925 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5927 int cpu = (long)hcpu;
5928 unsigned long flags;
5929 struct rq *rq = cpu_rq(cpu);
5931 switch (action & ~CPU_TASKS_FROZEN) {
5933 case CPU_UP_PREPARE:
5934 rq->calc_load_update = calc_load_update;
5938 /* Update our root-domain */
5939 raw_spin_lock_irqsave(&rq->lock, flags);
5941 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5945 raw_spin_unlock_irqrestore(&rq->lock, flags);
5948 #ifdef CONFIG_HOTPLUG_CPU
5950 /* Update our root-domain */
5951 raw_spin_lock_irqsave(&rq->lock, flags);
5953 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5957 BUG_ON(rq->nr_running != 1); /* the migration thread */
5958 raw_spin_unlock_irqrestore(&rq->lock, flags);
5960 migrate_nr_uninterruptible(rq);
5961 calc_global_load_remove(rq);
5969 * Register at high priority so that task migration (migrate_all_tasks)
5970 * happens before everything else. This has to be lower priority than
5971 * the notifier in the perf_event subsystem, though.
5973 static struct notifier_block __cpuinitdata migration_notifier = {
5974 .notifier_call = migration_call,
5975 .priority = CPU_PRI_MIGRATION,
5978 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5979 unsigned long action, void *hcpu)
5981 switch (action & ~CPU_TASKS_FROZEN) {
5983 case CPU_DOWN_FAILED:
5984 set_cpu_active((long)hcpu, true);
5991 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5992 unsigned long action, void *hcpu)
5994 switch (action & ~CPU_TASKS_FROZEN) {
5995 case CPU_DOWN_PREPARE:
5996 set_cpu_active((long)hcpu, false);
6003 static int __init migration_init(void)
6005 void *cpu = (void *)(long)smp_processor_id();
6008 /* Initialize migration for the boot CPU */
6009 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6010 BUG_ON(err == NOTIFY_BAD);
6011 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6012 register_cpu_notifier(&migration_notifier);
6014 /* Register cpu active notifiers */
6015 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6016 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6020 early_initcall(migration_init);
6025 #ifdef CONFIG_SCHED_DEBUG
6027 static __read_mostly int sched_domain_debug_enabled;
6029 static int __init sched_domain_debug_setup(char *str)
6031 sched_domain_debug_enabled = 1;
6035 early_param("sched_debug", sched_domain_debug_setup);
6037 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6038 struct cpumask *groupmask)
6040 struct sched_group *group = sd->groups;
6043 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6044 cpumask_clear(groupmask);
6046 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6048 if (!(sd->flags & SD_LOAD_BALANCE)) {
6049 printk("does not load-balance\n");
6051 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6056 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6058 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6059 printk(KERN_ERR "ERROR: domain->span does not contain "
6062 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6063 printk(KERN_ERR "ERROR: domain->groups does not contain"
6067 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6071 printk(KERN_ERR "ERROR: group is NULL\n");
6075 if (!group->cpu_power) {
6076 printk(KERN_CONT "\n");
6077 printk(KERN_ERR "ERROR: domain->cpu_power not "
6082 if (!cpumask_weight(sched_group_cpus(group))) {
6083 printk(KERN_CONT "\n");
6084 printk(KERN_ERR "ERROR: empty group\n");
6088 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6089 printk(KERN_CONT "\n");
6090 printk(KERN_ERR "ERROR: repeated CPUs\n");
6094 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6096 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6098 printk(KERN_CONT " %s", str);
6099 if (group->cpu_power != SCHED_LOAD_SCALE) {
6100 printk(KERN_CONT " (cpu_power = %d)",
6104 group = group->next;
6105 } while (group != sd->groups);
6106 printk(KERN_CONT "\n");
6108 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6109 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6112 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6113 printk(KERN_ERR "ERROR: parent span is not a superset "
6114 "of domain->span\n");
6118 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6120 cpumask_var_t groupmask;
6123 if (!sched_domain_debug_enabled)
6127 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6131 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6133 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6134 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6139 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6146 free_cpumask_var(groupmask);
6148 #else /* !CONFIG_SCHED_DEBUG */
6149 # define sched_domain_debug(sd, cpu) do { } while (0)
6150 #endif /* CONFIG_SCHED_DEBUG */
6152 static int sd_degenerate(struct sched_domain *sd)
6154 if (cpumask_weight(sched_domain_span(sd)) == 1)
6157 /* Following flags need at least 2 groups */
6158 if (sd->flags & (SD_LOAD_BALANCE |
6159 SD_BALANCE_NEWIDLE |
6163 SD_SHARE_PKG_RESOURCES)) {
6164 if (sd->groups != sd->groups->next)
6168 /* Following flags don't use groups */
6169 if (sd->flags & (SD_WAKE_AFFINE))
6176 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6178 unsigned long cflags = sd->flags, pflags = parent->flags;
6180 if (sd_degenerate(parent))
6183 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6186 /* Flags needing groups don't count if only 1 group in parent */
6187 if (parent->groups == parent->groups->next) {
6188 pflags &= ~(SD_LOAD_BALANCE |
6189 SD_BALANCE_NEWIDLE |
6193 SD_SHARE_PKG_RESOURCES);
6194 if (nr_node_ids == 1)
6195 pflags &= ~SD_SERIALIZE;
6197 if (~cflags & pflags)
6203 static void free_rootdomain(struct root_domain *rd)
6205 synchronize_sched();
6207 cpupri_cleanup(&rd->cpupri);
6209 free_cpumask_var(rd->rto_mask);
6210 free_cpumask_var(rd->online);
6211 free_cpumask_var(rd->span);
6215 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6217 struct root_domain *old_rd = NULL;
6218 unsigned long flags;
6220 raw_spin_lock_irqsave(&rq->lock, flags);
6225 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6228 cpumask_clear_cpu(rq->cpu, old_rd->span);
6231 * If we dont want to free the old_rt yet then
6232 * set old_rd to NULL to skip the freeing later
6235 if (!atomic_dec_and_test(&old_rd->refcount))
6239 atomic_inc(&rd->refcount);
6242 cpumask_set_cpu(rq->cpu, rd->span);
6243 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6246 raw_spin_unlock_irqrestore(&rq->lock, flags);
6249 free_rootdomain(old_rd);
6252 static int init_rootdomain(struct root_domain *rd)
6254 memset(rd, 0, sizeof(*rd));
6256 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6258 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6260 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6263 if (cpupri_init(&rd->cpupri) != 0)
6268 free_cpumask_var(rd->rto_mask);
6270 free_cpumask_var(rd->online);
6272 free_cpumask_var(rd->span);
6277 static void init_defrootdomain(void)
6279 init_rootdomain(&def_root_domain);
6281 atomic_set(&def_root_domain.refcount, 1);
6284 static struct root_domain *alloc_rootdomain(void)
6286 struct root_domain *rd;
6288 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6292 if (init_rootdomain(rd) != 0) {
6301 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6302 * hold the hotplug lock.
6305 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6307 struct rq *rq = cpu_rq(cpu);
6308 struct sched_domain *tmp;
6310 for (tmp = sd; tmp; tmp = tmp->parent)
6311 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6313 /* Remove the sched domains which do not contribute to scheduling. */
6314 for (tmp = sd; tmp; ) {
6315 struct sched_domain *parent = tmp->parent;
6319 if (sd_parent_degenerate(tmp, parent)) {
6320 tmp->parent = parent->parent;
6322 parent->parent->child = tmp;
6327 if (sd && sd_degenerate(sd)) {
6333 sched_domain_debug(sd, cpu);
6335 rq_attach_root(rq, rd);
6336 rcu_assign_pointer(rq->sd, sd);
6339 /* cpus with isolated domains */
6340 static cpumask_var_t cpu_isolated_map;
6342 /* Setup the mask of cpus configured for isolated domains */
6343 static int __init isolated_cpu_setup(char *str)
6345 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6346 cpulist_parse(str, cpu_isolated_map);
6350 __setup("isolcpus=", isolated_cpu_setup);
6353 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6354 * to a function which identifies what group(along with sched group) a CPU
6355 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6356 * (due to the fact that we keep track of groups covered with a struct cpumask).
6358 * init_sched_build_groups will build a circular linked list of the groups
6359 * covered by the given span, and will set each group's ->cpumask correctly,
6360 * and ->cpu_power to 0.
6363 init_sched_build_groups(const struct cpumask *span,
6364 const struct cpumask *cpu_map,
6365 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6366 struct sched_group **sg,
6367 struct cpumask *tmpmask),
6368 struct cpumask *covered, struct cpumask *tmpmask)
6370 struct sched_group *first = NULL, *last = NULL;
6373 cpumask_clear(covered);
6375 for_each_cpu(i, span) {
6376 struct sched_group *sg;
6377 int group = group_fn(i, cpu_map, &sg, tmpmask);
6380 if (cpumask_test_cpu(i, covered))
6383 cpumask_clear(sched_group_cpus(sg));
6386 for_each_cpu(j, span) {
6387 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6390 cpumask_set_cpu(j, covered);
6391 cpumask_set_cpu(j, sched_group_cpus(sg));
6402 #define SD_NODES_PER_DOMAIN 16
6407 * find_next_best_node - find the next node to include in a sched_domain
6408 * @node: node whose sched_domain we're building
6409 * @used_nodes: nodes already in the sched_domain
6411 * Find the next node to include in a given scheduling domain. Simply
6412 * finds the closest node not already in the @used_nodes map.
6414 * Should use nodemask_t.
6416 static int find_next_best_node(int node, nodemask_t *used_nodes)
6418 int i, n, val, min_val, best_node = 0;
6422 for (i = 0; i < nr_node_ids; i++) {
6423 /* Start at @node */
6424 n = (node + i) % nr_node_ids;
6426 if (!nr_cpus_node(n))
6429 /* Skip already used nodes */
6430 if (node_isset(n, *used_nodes))
6433 /* Simple min distance search */
6434 val = node_distance(node, n);
6436 if (val < min_val) {
6442 node_set(best_node, *used_nodes);
6447 * sched_domain_node_span - get a cpumask for a node's sched_domain
6448 * @node: node whose cpumask we're constructing
6449 * @span: resulting cpumask
6451 * Given a node, construct a good cpumask for its sched_domain to span. It
6452 * should be one that prevents unnecessary balancing, but also spreads tasks
6455 static void sched_domain_node_span(int node, struct cpumask *span)
6457 nodemask_t used_nodes;
6460 cpumask_clear(span);
6461 nodes_clear(used_nodes);
6463 cpumask_or(span, span, cpumask_of_node(node));
6464 node_set(node, used_nodes);
6466 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6467 int next_node = find_next_best_node(node, &used_nodes);
6469 cpumask_or(span, span, cpumask_of_node(next_node));
6472 #endif /* CONFIG_NUMA */
6474 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6477 * The cpus mask in sched_group and sched_domain hangs off the end.
6479 * ( See the the comments in include/linux/sched.h:struct sched_group
6480 * and struct sched_domain. )
6482 struct static_sched_group {
6483 struct sched_group sg;
6484 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6487 struct static_sched_domain {
6488 struct sched_domain sd;
6489 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6495 cpumask_var_t domainspan;
6496 cpumask_var_t covered;
6497 cpumask_var_t notcovered;
6499 cpumask_var_t nodemask;
6500 cpumask_var_t this_sibling_map;
6501 cpumask_var_t this_core_map;
6502 cpumask_var_t this_book_map;
6503 cpumask_var_t send_covered;
6504 cpumask_var_t tmpmask;
6505 struct sched_group **sched_group_nodes;
6506 struct root_domain *rd;
6510 sa_sched_groups = 0,
6516 sa_this_sibling_map,
6518 sa_sched_group_nodes,
6528 * SMT sched-domains:
6530 #ifdef CONFIG_SCHED_SMT
6531 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6532 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6535 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6536 struct sched_group **sg, struct cpumask *unused)
6539 *sg = &per_cpu(sched_groups, cpu).sg;
6542 #endif /* CONFIG_SCHED_SMT */
6545 * multi-core sched-domains:
6547 #ifdef CONFIG_SCHED_MC
6548 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6549 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6552 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6553 struct sched_group **sg, struct cpumask *mask)
6556 #ifdef CONFIG_SCHED_SMT
6557 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6558 group = cpumask_first(mask);
6563 *sg = &per_cpu(sched_group_core, group).sg;
6566 #endif /* CONFIG_SCHED_MC */
6569 * book sched-domains:
6571 #ifdef CONFIG_SCHED_BOOK
6572 static DEFINE_PER_CPU(struct static_sched_domain, book_domains);
6573 static DEFINE_PER_CPU(struct static_sched_group, sched_group_book);
6576 cpu_to_book_group(int cpu, const struct cpumask *cpu_map,
6577 struct sched_group **sg, struct cpumask *mask)
6580 #ifdef CONFIG_SCHED_MC
6581 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6582 group = cpumask_first(mask);
6583 #elif defined(CONFIG_SCHED_SMT)
6584 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6585 group = cpumask_first(mask);
6588 *sg = &per_cpu(sched_group_book, group).sg;
6591 #endif /* CONFIG_SCHED_BOOK */
6593 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6594 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6597 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6598 struct sched_group **sg, struct cpumask *mask)
6601 #ifdef CONFIG_SCHED_BOOK
6602 cpumask_and(mask, cpu_book_mask(cpu), cpu_map);
6603 group = cpumask_first(mask);
6604 #elif defined(CONFIG_SCHED_MC)
6605 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6606 group = cpumask_first(mask);
6607 #elif defined(CONFIG_SCHED_SMT)
6608 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6609 group = cpumask_first(mask);
6614 *sg = &per_cpu(sched_group_phys, group).sg;
6620 * The init_sched_build_groups can't handle what we want to do with node
6621 * groups, so roll our own. Now each node has its own list of groups which
6622 * gets dynamically allocated.
6624 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6625 static struct sched_group ***sched_group_nodes_bycpu;
6627 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6628 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6630 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6631 struct sched_group **sg,
6632 struct cpumask *nodemask)
6636 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6637 group = cpumask_first(nodemask);
6640 *sg = &per_cpu(sched_group_allnodes, group).sg;
6644 static void init_numa_sched_groups_power(struct sched_group *group_head)
6646 struct sched_group *sg = group_head;
6652 for_each_cpu(j, sched_group_cpus(sg)) {
6653 struct sched_domain *sd;
6655 sd = &per_cpu(phys_domains, j).sd;
6656 if (j != group_first_cpu(sd->groups)) {
6658 * Only add "power" once for each
6664 sg->cpu_power += sd->groups->cpu_power;
6667 } while (sg != group_head);
6670 static int build_numa_sched_groups(struct s_data *d,
6671 const struct cpumask *cpu_map, int num)
6673 struct sched_domain *sd;
6674 struct sched_group *sg, *prev;
6677 cpumask_clear(d->covered);
6678 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6679 if (cpumask_empty(d->nodemask)) {
6680 d->sched_group_nodes[num] = NULL;
6684 sched_domain_node_span(num, d->domainspan);
6685 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6687 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6690 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6694 d->sched_group_nodes[num] = sg;
6696 for_each_cpu(j, d->nodemask) {
6697 sd = &per_cpu(node_domains, j).sd;
6702 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6704 cpumask_or(d->covered, d->covered, d->nodemask);
6707 for (j = 0; j < nr_node_ids; j++) {
6708 n = (num + j) % nr_node_ids;
6709 cpumask_complement(d->notcovered, d->covered);
6710 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6711 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6712 if (cpumask_empty(d->tmpmask))
6714 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6715 if (cpumask_empty(d->tmpmask))
6717 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6721 "Can not alloc domain group for node %d\n", j);
6725 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6726 sg->next = prev->next;
6727 cpumask_or(d->covered, d->covered, d->tmpmask);
6734 #endif /* CONFIG_NUMA */
6737 /* Free memory allocated for various sched_group structures */
6738 static void free_sched_groups(const struct cpumask *cpu_map,
6739 struct cpumask *nodemask)
6743 for_each_cpu(cpu, cpu_map) {
6744 struct sched_group **sched_group_nodes
6745 = sched_group_nodes_bycpu[cpu];
6747 if (!sched_group_nodes)
6750 for (i = 0; i < nr_node_ids; i++) {
6751 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6753 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6754 if (cpumask_empty(nodemask))
6764 if (oldsg != sched_group_nodes[i])
6767 kfree(sched_group_nodes);
6768 sched_group_nodes_bycpu[cpu] = NULL;
6771 #else /* !CONFIG_NUMA */
6772 static void free_sched_groups(const struct cpumask *cpu_map,
6773 struct cpumask *nodemask)
6776 #endif /* CONFIG_NUMA */
6779 * Initialize sched groups cpu_power.
6781 * cpu_power indicates the capacity of sched group, which is used while
6782 * distributing the load between different sched groups in a sched domain.
6783 * Typically cpu_power for all the groups in a sched domain will be same unless
6784 * there are asymmetries in the topology. If there are asymmetries, group
6785 * having more cpu_power will pickup more load compared to the group having
6788 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6790 struct sched_domain *child;
6791 struct sched_group *group;
6795 WARN_ON(!sd || !sd->groups);
6797 if (cpu != group_first_cpu(sd->groups))
6802 sd->groups->cpu_power = 0;
6805 power = SCHED_LOAD_SCALE;
6806 weight = cpumask_weight(sched_domain_span(sd));
6808 * SMT siblings share the power of a single core.
6809 * Usually multiple threads get a better yield out of
6810 * that one core than a single thread would have,
6811 * reflect that in sd->smt_gain.
6813 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6814 power *= sd->smt_gain;
6816 power >>= SCHED_LOAD_SHIFT;
6818 sd->groups->cpu_power += power;
6823 * Add cpu_power of each child group to this groups cpu_power.
6825 group = child->groups;
6827 sd->groups->cpu_power += group->cpu_power;
6828 group = group->next;
6829 } while (group != child->groups);
6833 * Initializers for schedule domains
6834 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6837 #ifdef CONFIG_SCHED_DEBUG
6838 # define SD_INIT_NAME(sd, type) sd->name = #type
6840 # define SD_INIT_NAME(sd, type) do { } while (0)
6843 #define SD_INIT(sd, type) sd_init_##type(sd)
6845 #define SD_INIT_FUNC(type) \
6846 static noinline void sd_init_##type(struct sched_domain *sd) \
6848 memset(sd, 0, sizeof(*sd)); \
6849 *sd = SD_##type##_INIT; \
6850 sd->level = SD_LV_##type; \
6851 SD_INIT_NAME(sd, type); \
6856 SD_INIT_FUNC(ALLNODES)
6859 #ifdef CONFIG_SCHED_SMT
6860 SD_INIT_FUNC(SIBLING)
6862 #ifdef CONFIG_SCHED_MC
6865 #ifdef CONFIG_SCHED_BOOK
6869 static int default_relax_domain_level = -1;
6871 static int __init setup_relax_domain_level(char *str)
6875 val = simple_strtoul(str, NULL, 0);
6876 if (val < SD_LV_MAX)
6877 default_relax_domain_level = val;
6881 __setup("relax_domain_level=", setup_relax_domain_level);
6883 static void set_domain_attribute(struct sched_domain *sd,
6884 struct sched_domain_attr *attr)
6888 if (!attr || attr->relax_domain_level < 0) {
6889 if (default_relax_domain_level < 0)
6892 request = default_relax_domain_level;
6894 request = attr->relax_domain_level;
6895 if (request < sd->level) {
6896 /* turn off idle balance on this domain */
6897 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6899 /* turn on idle balance on this domain */
6900 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6904 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6905 const struct cpumask *cpu_map)
6908 case sa_sched_groups:
6909 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6910 d->sched_group_nodes = NULL;
6912 free_rootdomain(d->rd); /* fall through */
6914 free_cpumask_var(d->tmpmask); /* fall through */
6915 case sa_send_covered:
6916 free_cpumask_var(d->send_covered); /* fall through */
6917 case sa_this_book_map:
6918 free_cpumask_var(d->this_book_map); /* fall through */
6919 case sa_this_core_map:
6920 free_cpumask_var(d->this_core_map); /* fall through */
6921 case sa_this_sibling_map:
6922 free_cpumask_var(d->this_sibling_map); /* fall through */
6924 free_cpumask_var(d->nodemask); /* fall through */
6925 case sa_sched_group_nodes:
6927 kfree(d->sched_group_nodes); /* fall through */
6929 free_cpumask_var(d->notcovered); /* fall through */
6931 free_cpumask_var(d->covered); /* fall through */
6933 free_cpumask_var(d->domainspan); /* fall through */
6940 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6941 const struct cpumask *cpu_map)
6944 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6946 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6947 return sa_domainspan;
6948 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6950 /* Allocate the per-node list of sched groups */
6951 d->sched_group_nodes = kcalloc(nr_node_ids,
6952 sizeof(struct sched_group *), GFP_KERNEL);
6953 if (!d->sched_group_nodes) {
6954 printk(KERN_WARNING "Can not alloc sched group node list\n");
6955 return sa_notcovered;
6957 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6959 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6960 return sa_sched_group_nodes;
6961 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6963 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6964 return sa_this_sibling_map;
6965 if (!alloc_cpumask_var(&d->this_book_map, GFP_KERNEL))
6966 return sa_this_core_map;
6967 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6968 return sa_this_book_map;
6969 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6970 return sa_send_covered;
6971 d->rd = alloc_rootdomain();
6973 printk(KERN_WARNING "Cannot alloc root domain\n");
6976 return sa_rootdomain;
6979 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6980 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6982 struct sched_domain *sd = NULL;
6984 struct sched_domain *parent;
6987 if (cpumask_weight(cpu_map) >
6988 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6989 sd = &per_cpu(allnodes_domains, i).sd;
6990 SD_INIT(sd, ALLNODES);
6991 set_domain_attribute(sd, attr);
6992 cpumask_copy(sched_domain_span(sd), cpu_map);
6993 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6998 sd = &per_cpu(node_domains, i).sd;
7000 set_domain_attribute(sd, attr);
7001 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7002 sd->parent = parent;
7005 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7010 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7011 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7012 struct sched_domain *parent, int i)
7014 struct sched_domain *sd;
7015 sd = &per_cpu(phys_domains, i).sd;
7017 set_domain_attribute(sd, attr);
7018 cpumask_copy(sched_domain_span(sd), d->nodemask);
7019 sd->parent = parent;
7022 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7026 static struct sched_domain *__build_book_sched_domain(struct s_data *d,
7027 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7028 struct sched_domain *parent, int i)
7030 struct sched_domain *sd = parent;
7031 #ifdef CONFIG_SCHED_BOOK
7032 sd = &per_cpu(book_domains, i).sd;
7034 set_domain_attribute(sd, attr);
7035 cpumask_and(sched_domain_span(sd), cpu_map, cpu_book_mask(i));
7036 sd->parent = parent;
7038 cpu_to_book_group(i, cpu_map, &sd->groups, d->tmpmask);
7043 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7044 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7045 struct sched_domain *parent, int i)
7047 struct sched_domain *sd = parent;
7048 #ifdef CONFIG_SCHED_MC
7049 sd = &per_cpu(core_domains, i).sd;
7051 set_domain_attribute(sd, attr);
7052 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7053 sd->parent = parent;
7055 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7060 static struct sched_domain *__build_smt_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_SMT
7066 sd = &per_cpu(cpu_domains, i).sd;
7067 SD_INIT(sd, SIBLING);
7068 set_domain_attribute(sd, attr);
7069 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7070 sd->parent = parent;
7072 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7077 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7078 const struct cpumask *cpu_map, int cpu)
7081 #ifdef CONFIG_SCHED_SMT
7082 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7083 cpumask_and(d->this_sibling_map, cpu_map,
7084 topology_thread_cpumask(cpu));
7085 if (cpu == cpumask_first(d->this_sibling_map))
7086 init_sched_build_groups(d->this_sibling_map, cpu_map,
7088 d->send_covered, d->tmpmask);
7091 #ifdef CONFIG_SCHED_MC
7092 case SD_LV_MC: /* set up multi-core groups */
7093 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7094 if (cpu == cpumask_first(d->this_core_map))
7095 init_sched_build_groups(d->this_core_map, cpu_map,
7097 d->send_covered, d->tmpmask);
7100 #ifdef CONFIG_SCHED_BOOK
7101 case SD_LV_BOOK: /* set up book groups */
7102 cpumask_and(d->this_book_map, cpu_map, cpu_book_mask(cpu));
7103 if (cpu == cpumask_first(d->this_book_map))
7104 init_sched_build_groups(d->this_book_map, cpu_map,
7106 d->send_covered, d->tmpmask);
7109 case SD_LV_CPU: /* set up physical groups */
7110 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7111 if (!cpumask_empty(d->nodemask))
7112 init_sched_build_groups(d->nodemask, cpu_map,
7114 d->send_covered, d->tmpmask);
7117 case SD_LV_ALLNODES:
7118 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7119 d->send_covered, d->tmpmask);
7128 * Build sched domains for a given set of cpus and attach the sched domains
7129 * to the individual cpus
7131 static int __build_sched_domains(const struct cpumask *cpu_map,
7132 struct sched_domain_attr *attr)
7134 enum s_alloc alloc_state = sa_none;
7136 struct sched_domain *sd;
7142 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7143 if (alloc_state != sa_rootdomain)
7145 alloc_state = sa_sched_groups;
7148 * Set up domains for cpus specified by the cpu_map.
7150 for_each_cpu(i, cpu_map) {
7151 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7154 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7155 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7156 sd = __build_book_sched_domain(&d, cpu_map, attr, sd, i);
7157 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7158 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7161 for_each_cpu(i, cpu_map) {
7162 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7163 build_sched_groups(&d, SD_LV_BOOK, cpu_map, i);
7164 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7167 /* Set up physical groups */
7168 for (i = 0; i < nr_node_ids; i++)
7169 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7172 /* Set up node groups */
7174 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7176 for (i = 0; i < nr_node_ids; i++)
7177 if (build_numa_sched_groups(&d, cpu_map, i))
7181 /* Calculate CPU power for physical packages and nodes */
7182 #ifdef CONFIG_SCHED_SMT
7183 for_each_cpu(i, cpu_map) {
7184 sd = &per_cpu(cpu_domains, i).sd;
7185 init_sched_groups_power(i, sd);
7188 #ifdef CONFIG_SCHED_MC
7189 for_each_cpu(i, cpu_map) {
7190 sd = &per_cpu(core_domains, i).sd;
7191 init_sched_groups_power(i, sd);
7194 #ifdef CONFIG_SCHED_BOOK
7195 for_each_cpu(i, cpu_map) {
7196 sd = &per_cpu(book_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;
7225 #elif defined(CONFIG_SCHED_BOOK)
7226 sd = &per_cpu(book_domains, i).sd;
7228 sd = &per_cpu(phys_domains, i).sd;
7230 cpu_attach_domain(sd, d.rd, i);
7233 d.sched_group_nodes = NULL; /* don't free this we still need it */
7234 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7238 __free_domain_allocs(&d, alloc_state, cpu_map);
7242 static int build_sched_domains(const struct cpumask *cpu_map)
7244 return __build_sched_domains(cpu_map, NULL);
7247 static cpumask_var_t *doms_cur; /* current sched domains */
7248 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7249 static struct sched_domain_attr *dattr_cur;
7250 /* attribues of custom domains in 'doms_cur' */
7253 * Special case: If a kmalloc of a doms_cur partition (array of
7254 * cpumask) fails, then fallback to a single sched domain,
7255 * as determined by the single cpumask fallback_doms.
7257 static cpumask_var_t fallback_doms;
7260 * arch_update_cpu_topology lets virtualized architectures update the
7261 * cpu core maps. It is supposed to return 1 if the topology changed
7262 * or 0 if it stayed the same.
7264 int __attribute__((weak)) arch_update_cpu_topology(void)
7269 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7272 cpumask_var_t *doms;
7274 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7277 for (i = 0; i < ndoms; i++) {
7278 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7279 free_sched_domains(doms, i);
7286 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7289 for (i = 0; i < ndoms; i++)
7290 free_cpumask_var(doms[i]);
7295 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7296 * For now this just excludes isolated cpus, but could be used to
7297 * exclude other special cases in the future.
7299 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7303 arch_update_cpu_topology();
7305 doms_cur = alloc_sched_domains(ndoms_cur);
7307 doms_cur = &fallback_doms;
7308 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7310 err = build_sched_domains(doms_cur[0]);
7311 register_sched_domain_sysctl();
7316 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7317 struct cpumask *tmpmask)
7319 free_sched_groups(cpu_map, tmpmask);
7323 * Detach sched domains from a group of cpus specified in cpu_map
7324 * These cpus will now be attached to the NULL domain
7326 static void detach_destroy_domains(const struct cpumask *cpu_map)
7328 /* Save because hotplug lock held. */
7329 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7332 for_each_cpu(i, cpu_map)
7333 cpu_attach_domain(NULL, &def_root_domain, i);
7334 synchronize_sched();
7335 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7338 /* handle null as "default" */
7339 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7340 struct sched_domain_attr *new, int idx_new)
7342 struct sched_domain_attr tmp;
7349 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7350 new ? (new + idx_new) : &tmp,
7351 sizeof(struct sched_domain_attr));
7355 * Partition sched domains as specified by the 'ndoms_new'
7356 * cpumasks in the array doms_new[] of cpumasks. This compares
7357 * doms_new[] to the current sched domain partitioning, doms_cur[].
7358 * It destroys each deleted domain and builds each new domain.
7360 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7361 * The masks don't intersect (don't overlap.) We should setup one
7362 * sched domain for each mask. CPUs not in any of the cpumasks will
7363 * not be load balanced. If the same cpumask appears both in the
7364 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7367 * The passed in 'doms_new' should be allocated using
7368 * alloc_sched_domains. This routine takes ownership of it and will
7369 * free_sched_domains it when done with it. If the caller failed the
7370 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7371 * and partition_sched_domains() will fallback to the single partition
7372 * 'fallback_doms', it also forces the domains to be rebuilt.
7374 * If doms_new == NULL it will be replaced with cpu_online_mask.
7375 * ndoms_new == 0 is a special case for destroying existing domains,
7376 * and it will not create the default domain.
7378 * Call with hotplug lock held
7380 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7381 struct sched_domain_attr *dattr_new)
7386 mutex_lock(&sched_domains_mutex);
7388 /* always unregister in case we don't destroy any domains */
7389 unregister_sched_domain_sysctl();
7391 /* Let architecture update cpu core mappings. */
7392 new_topology = arch_update_cpu_topology();
7394 n = doms_new ? ndoms_new : 0;
7396 /* Destroy deleted domains */
7397 for (i = 0; i < ndoms_cur; i++) {
7398 for (j = 0; j < n && !new_topology; j++) {
7399 if (cpumask_equal(doms_cur[i], doms_new[j])
7400 && dattrs_equal(dattr_cur, i, dattr_new, j))
7403 /* no match - a current sched domain not in new doms_new[] */
7404 detach_destroy_domains(doms_cur[i]);
7409 if (doms_new == NULL) {
7411 doms_new = &fallback_doms;
7412 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7413 WARN_ON_ONCE(dattr_new);
7416 /* Build new domains */
7417 for (i = 0; i < ndoms_new; i++) {
7418 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7419 if (cpumask_equal(doms_new[i], doms_cur[j])
7420 && dattrs_equal(dattr_new, i, dattr_cur, j))
7423 /* no match - add a new doms_new */
7424 __build_sched_domains(doms_new[i],
7425 dattr_new ? dattr_new + i : NULL);
7430 /* Remember the new sched domains */
7431 if (doms_cur != &fallback_doms)
7432 free_sched_domains(doms_cur, ndoms_cur);
7433 kfree(dattr_cur); /* kfree(NULL) is safe */
7434 doms_cur = doms_new;
7435 dattr_cur = dattr_new;
7436 ndoms_cur = ndoms_new;
7438 register_sched_domain_sysctl();
7440 mutex_unlock(&sched_domains_mutex);
7443 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7444 static void arch_reinit_sched_domains(void)
7448 /* Destroy domains first to force the rebuild */
7449 partition_sched_domains(0, NULL, NULL);
7451 rebuild_sched_domains();
7455 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7457 unsigned int level = 0;
7459 if (sscanf(buf, "%u", &level) != 1)
7463 * level is always be positive so don't check for
7464 * level < POWERSAVINGS_BALANCE_NONE which is 0
7465 * What happens on 0 or 1 byte write,
7466 * need to check for count as well?
7469 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7473 sched_smt_power_savings = level;
7475 sched_mc_power_savings = level;
7477 arch_reinit_sched_domains();
7482 #ifdef CONFIG_SCHED_MC
7483 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7484 struct sysdev_class_attribute *attr,
7487 return sprintf(page, "%u\n", sched_mc_power_savings);
7489 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7490 struct sysdev_class_attribute *attr,
7491 const char *buf, size_t count)
7493 return sched_power_savings_store(buf, count, 0);
7495 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7496 sched_mc_power_savings_show,
7497 sched_mc_power_savings_store);
7500 #ifdef CONFIG_SCHED_SMT
7501 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7502 struct sysdev_class_attribute *attr,
7505 return sprintf(page, "%u\n", sched_smt_power_savings);
7507 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7508 struct sysdev_class_attribute *attr,
7509 const char *buf, size_t count)
7511 return sched_power_savings_store(buf, count, 1);
7513 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7514 sched_smt_power_savings_show,
7515 sched_smt_power_savings_store);
7518 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7522 #ifdef CONFIG_SCHED_SMT
7524 err = sysfs_create_file(&cls->kset.kobj,
7525 &attr_sched_smt_power_savings.attr);
7527 #ifdef CONFIG_SCHED_MC
7528 if (!err && mc_capable())
7529 err = sysfs_create_file(&cls->kset.kobj,
7530 &attr_sched_mc_power_savings.attr);
7534 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7537 * Update cpusets according to cpu_active mask. If cpusets are
7538 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7539 * around partition_sched_domains().
7541 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7544 switch (action & ~CPU_TASKS_FROZEN) {
7546 case CPU_DOWN_FAILED:
7547 cpuset_update_active_cpus();
7554 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7557 switch (action & ~CPU_TASKS_FROZEN) {
7558 case CPU_DOWN_PREPARE:
7559 cpuset_update_active_cpus();
7566 static int update_runtime(struct notifier_block *nfb,
7567 unsigned long action, void *hcpu)
7569 int cpu = (int)(long)hcpu;
7572 case CPU_DOWN_PREPARE:
7573 case CPU_DOWN_PREPARE_FROZEN:
7574 disable_runtime(cpu_rq(cpu));
7577 case CPU_DOWN_FAILED:
7578 case CPU_DOWN_FAILED_FROZEN:
7580 case CPU_ONLINE_FROZEN:
7581 enable_runtime(cpu_rq(cpu));
7589 void __init sched_init_smp(void)
7591 cpumask_var_t non_isolated_cpus;
7593 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7594 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7596 #if defined(CONFIG_NUMA)
7597 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7599 BUG_ON(sched_group_nodes_bycpu == NULL);
7602 mutex_lock(&sched_domains_mutex);
7603 arch_init_sched_domains(cpu_active_mask);
7604 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7605 if (cpumask_empty(non_isolated_cpus))
7606 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7607 mutex_unlock(&sched_domains_mutex);
7610 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7611 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7613 /* RT runtime code needs to handle some hotplug events */
7614 hotcpu_notifier(update_runtime, 0);
7618 /* Move init over to a non-isolated CPU */
7619 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7621 sched_init_granularity();
7622 free_cpumask_var(non_isolated_cpus);
7624 init_sched_rt_class();
7627 void __init sched_init_smp(void)
7629 sched_init_granularity();
7631 #endif /* CONFIG_SMP */
7633 const_debug unsigned int sysctl_timer_migration = 1;
7635 int in_sched_functions(unsigned long addr)
7637 return in_lock_functions(addr) ||
7638 (addr >= (unsigned long)__sched_text_start
7639 && addr < (unsigned long)__sched_text_end);
7642 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7644 cfs_rq->tasks_timeline = RB_ROOT;
7645 INIT_LIST_HEAD(&cfs_rq->tasks);
7646 #ifdef CONFIG_FAIR_GROUP_SCHED
7649 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7652 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7654 struct rt_prio_array *array;
7657 array = &rt_rq->active;
7658 for (i = 0; i < MAX_RT_PRIO; i++) {
7659 INIT_LIST_HEAD(array->queue + i);
7660 __clear_bit(i, array->bitmap);
7662 /* delimiter for bitsearch: */
7663 __set_bit(MAX_RT_PRIO, array->bitmap);
7665 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7666 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7668 rt_rq->highest_prio.next = MAX_RT_PRIO;
7672 rt_rq->rt_nr_migratory = 0;
7673 rt_rq->overloaded = 0;
7674 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7678 rt_rq->rt_throttled = 0;
7679 rt_rq->rt_runtime = 0;
7680 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7682 #ifdef CONFIG_RT_GROUP_SCHED
7683 rt_rq->rt_nr_boosted = 0;
7688 #ifdef CONFIG_FAIR_GROUP_SCHED
7689 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7690 struct sched_entity *se, int cpu,
7691 struct sched_entity *parent)
7693 struct rq *rq = cpu_rq(cpu);
7694 tg->cfs_rq[cpu] = cfs_rq;
7695 init_cfs_rq(cfs_rq, rq);
7699 /* se could be NULL for init_task_group */
7704 se->cfs_rq = &rq->cfs;
7706 se->cfs_rq = parent->my_q;
7709 update_load_set(&se->load, tg->shares);
7710 se->parent = parent;
7714 #ifdef CONFIG_RT_GROUP_SCHED
7715 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7716 struct sched_rt_entity *rt_se, int cpu,
7717 struct sched_rt_entity *parent)
7719 struct rq *rq = cpu_rq(cpu);
7721 tg->rt_rq[cpu] = rt_rq;
7722 init_rt_rq(rt_rq, rq);
7724 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7726 tg->rt_se[cpu] = rt_se;
7731 rt_se->rt_rq = &rq->rt;
7733 rt_se->rt_rq = parent->my_q;
7735 rt_se->my_q = rt_rq;
7736 rt_se->parent = parent;
7737 INIT_LIST_HEAD(&rt_se->run_list);
7741 void __init sched_init(void)
7744 unsigned long alloc_size = 0, ptr;
7746 #ifdef CONFIG_FAIR_GROUP_SCHED
7747 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7749 #ifdef CONFIG_RT_GROUP_SCHED
7750 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7752 #ifdef CONFIG_CPUMASK_OFFSTACK
7753 alloc_size += num_possible_cpus() * cpumask_size();
7756 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7758 #ifdef CONFIG_FAIR_GROUP_SCHED
7759 init_task_group.se = (struct sched_entity **)ptr;
7760 ptr += nr_cpu_ids * sizeof(void **);
7762 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7763 ptr += nr_cpu_ids * sizeof(void **);
7765 #endif /* CONFIG_FAIR_GROUP_SCHED */
7766 #ifdef CONFIG_RT_GROUP_SCHED
7767 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7768 ptr += nr_cpu_ids * sizeof(void **);
7770 init_task_group.rt_rq = (struct rt_rq **)ptr;
7771 ptr += nr_cpu_ids * sizeof(void **);
7773 #endif /* CONFIG_RT_GROUP_SCHED */
7774 #ifdef CONFIG_CPUMASK_OFFSTACK
7775 for_each_possible_cpu(i) {
7776 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7777 ptr += cpumask_size();
7779 #endif /* CONFIG_CPUMASK_OFFSTACK */
7783 init_defrootdomain();
7786 init_rt_bandwidth(&def_rt_bandwidth,
7787 global_rt_period(), global_rt_runtime());
7789 #ifdef CONFIG_RT_GROUP_SCHED
7790 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7791 global_rt_period(), global_rt_runtime());
7792 #endif /* CONFIG_RT_GROUP_SCHED */
7794 #ifdef CONFIG_CGROUP_SCHED
7795 list_add(&init_task_group.list, &task_groups);
7796 INIT_LIST_HEAD(&init_task_group.children);
7798 #endif /* CONFIG_CGROUP_SCHED */
7800 for_each_possible_cpu(i) {
7804 raw_spin_lock_init(&rq->lock);
7806 rq->calc_load_active = 0;
7807 rq->calc_load_update = jiffies + LOAD_FREQ;
7808 init_cfs_rq(&rq->cfs, rq);
7809 init_rt_rq(&rq->rt, rq);
7810 #ifdef CONFIG_FAIR_GROUP_SCHED
7811 init_task_group.shares = init_task_group_load;
7812 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7813 #ifdef CONFIG_CGROUP_SCHED
7815 * How much cpu bandwidth does init_task_group get?
7817 * In case of task-groups formed thr' the cgroup filesystem, it
7818 * gets 100% of the cpu resources in the system. This overall
7819 * system cpu resource is divided among the tasks of
7820 * init_task_group and its child task-groups in a fair manner,
7821 * based on each entity's (task or task-group's) weight
7822 * (se->load.weight).
7824 * In other words, if init_task_group has 10 tasks of weight
7825 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7826 * then A0's share of the cpu resource is:
7828 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7830 * We achieve this by letting init_task_group's tasks sit
7831 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7833 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, NULL);
7835 #endif /* CONFIG_FAIR_GROUP_SCHED */
7837 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7838 #ifdef CONFIG_RT_GROUP_SCHED
7839 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7840 #ifdef CONFIG_CGROUP_SCHED
7841 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, NULL);
7845 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7846 rq->cpu_load[j] = 0;
7848 rq->last_load_update_tick = jiffies;
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);
7864 rq->nohz_balance_kick = 0;
7865 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
7869 atomic_set(&rq->nr_iowait, 0);
7872 set_load_weight(&init_task);
7874 #ifdef CONFIG_PREEMPT_NOTIFIERS
7875 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7879 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7882 #ifdef CONFIG_RT_MUTEXES
7883 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7887 * The boot idle thread does lazy MMU switching as well:
7889 atomic_inc(&init_mm.mm_count);
7890 enter_lazy_tlb(&init_mm, current);
7893 * Make us the idle thread. Technically, schedule() should not be
7894 * called from this thread, however somewhere below it might be,
7895 * but because we are the idle thread, we just pick up running again
7896 * when this runqueue becomes "idle".
7898 init_idle(current, smp_processor_id());
7900 calc_load_update = jiffies + LOAD_FREQ;
7903 * During early bootup we pretend to be a normal task:
7905 current->sched_class = &fair_sched_class;
7907 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7908 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7911 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7912 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
7913 atomic_set(&nohz.load_balancer, nr_cpu_ids);
7914 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
7915 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
7917 /* May be allocated at isolcpus cmdline parse time */
7918 if (cpu_isolated_map == NULL)
7919 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7924 scheduler_running = 1;
7927 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7928 static inline int preempt_count_equals(int preempt_offset)
7930 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7932 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7935 void __might_sleep(const char *file, int line, int preempt_offset)
7938 static unsigned long prev_jiffy; /* ratelimiting */
7940 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7941 system_state != SYSTEM_RUNNING || oops_in_progress)
7943 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7945 prev_jiffy = jiffies;
7948 "BUG: sleeping function called from invalid context at %s:%d\n",
7951 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7952 in_atomic(), irqs_disabled(),
7953 current->pid, current->comm);
7955 debug_show_held_locks(current);
7956 if (irqs_disabled())
7957 print_irqtrace_events(current);
7961 EXPORT_SYMBOL(__might_sleep);
7964 #ifdef CONFIG_MAGIC_SYSRQ
7965 static void normalize_task(struct rq *rq, struct task_struct *p)
7969 on_rq = p->se.on_rq;
7971 deactivate_task(rq, p, 0);
7972 __setscheduler(rq, p, SCHED_NORMAL, 0);
7974 activate_task(rq, p, 0);
7975 resched_task(rq->curr);
7979 void normalize_rt_tasks(void)
7981 struct task_struct *g, *p;
7982 unsigned long flags;
7985 read_lock_irqsave(&tasklist_lock, flags);
7986 do_each_thread(g, p) {
7988 * Only normalize user tasks:
7993 p->se.exec_start = 0;
7994 #ifdef CONFIG_SCHEDSTATS
7995 p->se.statistics.wait_start = 0;
7996 p->se.statistics.sleep_start = 0;
7997 p->se.statistics.block_start = 0;
8002 * Renice negative nice level userspace
8005 if (TASK_NICE(p) < 0 && p->mm)
8006 set_user_nice(p, 0);
8010 raw_spin_lock(&p->pi_lock);
8011 rq = __task_rq_lock(p);
8013 normalize_task(rq, p);
8015 __task_rq_unlock(rq);
8016 raw_spin_unlock(&p->pi_lock);
8017 } while_each_thread(g, p);
8019 read_unlock_irqrestore(&tasklist_lock, flags);
8022 #endif /* CONFIG_MAGIC_SYSRQ */
8024 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8026 * These functions are only useful for the IA64 MCA handling, or kdb.
8028 * They can only be called when the whole system has been
8029 * stopped - every CPU needs to be quiescent, and no scheduling
8030 * activity can take place. Using them for anything else would
8031 * be a serious bug, and as a result, they aren't even visible
8032 * under any other configuration.
8036 * curr_task - return the current task for a given cpu.
8037 * @cpu: the processor in question.
8039 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8041 struct task_struct *curr_task(int cpu)
8043 return cpu_curr(cpu);
8046 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8050 * set_curr_task - set the current task for a given cpu.
8051 * @cpu: the processor in question.
8052 * @p: the task pointer to set.
8054 * Description: This function must only be used when non-maskable interrupts
8055 * are serviced on a separate stack. It allows the architecture to switch the
8056 * notion of the current task on a cpu in a non-blocking manner. This function
8057 * must be called with all CPU's synchronized, and interrupts disabled, the
8058 * and caller must save the original value of the current task (see
8059 * curr_task() above) and restore that value before reenabling interrupts and
8060 * re-starting the system.
8062 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8064 void set_curr_task(int cpu, struct task_struct *p)
8071 #ifdef CONFIG_FAIR_GROUP_SCHED
8072 static void free_fair_sched_group(struct task_group *tg)
8076 for_each_possible_cpu(i) {
8078 kfree(tg->cfs_rq[i]);
8088 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8090 struct cfs_rq *cfs_rq;
8091 struct sched_entity *se;
8095 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8098 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8102 tg->shares = NICE_0_LOAD;
8104 for_each_possible_cpu(i) {
8107 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8108 GFP_KERNEL, cpu_to_node(i));
8112 se = kzalloc_node(sizeof(struct sched_entity),
8113 GFP_KERNEL, cpu_to_node(i));
8117 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8128 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8130 struct rq *rq = cpu_rq(cpu);
8131 unsigned long flags;
8135 * Only empty task groups can be destroyed; so we can speculatively
8136 * check on_list without danger of it being re-added.
8138 if (!tg->cfs_rq[cpu]->on_list)
8141 raw_spin_lock_irqsave(&rq->lock, flags);
8142 list_del_leaf_cfs_rq(tg->cfs_rq[i]);
8143 raw_spin_unlock_irqrestore(&rq->lock, flags);
8145 #else /* !CONFG_FAIR_GROUP_SCHED */
8146 static inline void free_fair_sched_group(struct task_group *tg)
8151 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8156 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8159 #endif /* CONFIG_FAIR_GROUP_SCHED */
8161 #ifdef CONFIG_RT_GROUP_SCHED
8162 static void free_rt_sched_group(struct task_group *tg)
8166 destroy_rt_bandwidth(&tg->rt_bandwidth);
8168 for_each_possible_cpu(i) {
8170 kfree(tg->rt_rq[i]);
8172 kfree(tg->rt_se[i]);
8180 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8182 struct rt_rq *rt_rq;
8183 struct sched_rt_entity *rt_se;
8187 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8190 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8194 init_rt_bandwidth(&tg->rt_bandwidth,
8195 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8197 for_each_possible_cpu(i) {
8200 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8201 GFP_KERNEL, cpu_to_node(i));
8205 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8206 GFP_KERNEL, cpu_to_node(i));
8210 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8220 #else /* !CONFIG_RT_GROUP_SCHED */
8221 static inline void free_rt_sched_group(struct task_group *tg)
8226 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8230 #endif /* CONFIG_RT_GROUP_SCHED */
8232 #ifdef CONFIG_CGROUP_SCHED
8233 static void free_sched_group(struct task_group *tg)
8235 free_fair_sched_group(tg);
8236 free_rt_sched_group(tg);
8240 /* allocate runqueue etc for a new task group */
8241 struct task_group *sched_create_group(struct task_group *parent)
8243 struct task_group *tg;
8244 unsigned long flags;
8246 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8248 return ERR_PTR(-ENOMEM);
8250 if (!alloc_fair_sched_group(tg, parent))
8253 if (!alloc_rt_sched_group(tg, parent))
8256 spin_lock_irqsave(&task_group_lock, flags);
8257 list_add_rcu(&tg->list, &task_groups);
8259 WARN_ON(!parent); /* root should already exist */
8261 tg->parent = parent;
8262 INIT_LIST_HEAD(&tg->children);
8263 list_add_rcu(&tg->siblings, &parent->children);
8264 spin_unlock_irqrestore(&task_group_lock, flags);
8269 free_sched_group(tg);
8270 return ERR_PTR(-ENOMEM);
8273 /* rcu callback to free various structures associated with a task group */
8274 static void free_sched_group_rcu(struct rcu_head *rhp)
8276 /* now it should be safe to free those cfs_rqs */
8277 free_sched_group(container_of(rhp, struct task_group, rcu));
8280 /* Destroy runqueue etc associated with a task group */
8281 void sched_destroy_group(struct task_group *tg)
8283 unsigned long flags;
8286 /* end participation in shares distribution */
8287 for_each_possible_cpu(i)
8288 unregister_fair_sched_group(tg, i);
8290 spin_lock_irqsave(&task_group_lock, flags);
8291 list_del_rcu(&tg->list);
8292 list_del_rcu(&tg->siblings);
8293 spin_unlock_irqrestore(&task_group_lock, flags);
8295 /* wait for possible concurrent references to cfs_rqs complete */
8296 call_rcu(&tg->rcu, free_sched_group_rcu);
8299 /* change task's runqueue when it moves between groups.
8300 * The caller of this function should have put the task in its new group
8301 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8302 * reflect its new group.
8304 void sched_move_task(struct task_struct *tsk)
8307 unsigned long flags;
8310 rq = task_rq_lock(tsk, &flags);
8312 running = task_current(rq, tsk);
8313 on_rq = tsk->se.on_rq;
8316 dequeue_task(rq, tsk, 0);
8317 if (unlikely(running))
8318 tsk->sched_class->put_prev_task(rq, tsk);
8320 #ifdef CONFIG_FAIR_GROUP_SCHED
8321 if (tsk->sched_class->task_move_group)
8322 tsk->sched_class->task_move_group(tsk, on_rq);
8325 set_task_rq(tsk, task_cpu(tsk));
8327 if (unlikely(running))
8328 tsk->sched_class->set_curr_task(rq);
8330 enqueue_task(rq, tsk, 0);
8332 task_rq_unlock(rq, &flags);
8334 #endif /* CONFIG_CGROUP_SCHED */
8336 #ifdef CONFIG_FAIR_GROUP_SCHED
8337 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8339 struct cfs_rq *cfs_rq = se->cfs_rq;
8344 dequeue_entity(cfs_rq, se, 0);
8346 update_load_set(&se->load, shares);
8349 enqueue_entity(cfs_rq, se, 0);
8352 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8354 struct cfs_rq *cfs_rq = se->cfs_rq;
8355 struct rq *rq = cfs_rq->rq;
8356 unsigned long flags;
8358 raw_spin_lock_irqsave(&rq->lock, flags);
8359 __set_se_shares(se, shares);
8360 raw_spin_unlock_irqrestore(&rq->lock, flags);
8363 static DEFINE_MUTEX(shares_mutex);
8365 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8370 * We can't change the weight of the root cgroup.
8375 if (shares < MIN_SHARES)
8376 shares = MIN_SHARES;
8377 else if (shares > MAX_SHARES)
8378 shares = MAX_SHARES;
8380 mutex_lock(&shares_mutex);
8381 if (tg->shares == shares)
8384 tg->shares = shares;
8385 for_each_possible_cpu(i) {
8389 set_se_shares(tg->se[i], shares);
8393 mutex_unlock(&shares_mutex);
8397 unsigned long sched_group_shares(struct task_group *tg)
8403 #ifdef CONFIG_RT_GROUP_SCHED
8405 * Ensure that the real time constraints are schedulable.
8407 static DEFINE_MUTEX(rt_constraints_mutex);
8409 static unsigned long to_ratio(u64 period, u64 runtime)
8411 if (runtime == RUNTIME_INF)
8414 return div64_u64(runtime << 20, period);
8417 /* Must be called with tasklist_lock held */
8418 static inline int tg_has_rt_tasks(struct task_group *tg)
8420 struct task_struct *g, *p;
8422 do_each_thread(g, p) {
8423 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8425 } while_each_thread(g, p);
8430 struct rt_schedulable_data {
8431 struct task_group *tg;
8436 static int tg_schedulable(struct task_group *tg, void *data)
8438 struct rt_schedulable_data *d = data;
8439 struct task_group *child;
8440 unsigned long total, sum = 0;
8441 u64 period, runtime;
8443 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8444 runtime = tg->rt_bandwidth.rt_runtime;
8447 period = d->rt_period;
8448 runtime = d->rt_runtime;
8452 * Cannot have more runtime than the period.
8454 if (runtime > period && runtime != RUNTIME_INF)
8458 * Ensure we don't starve existing RT tasks.
8460 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8463 total = to_ratio(period, runtime);
8466 * Nobody can have more than the global setting allows.
8468 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8472 * The sum of our children's runtime should not exceed our own.
8474 list_for_each_entry_rcu(child, &tg->children, siblings) {
8475 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8476 runtime = child->rt_bandwidth.rt_runtime;
8478 if (child == d->tg) {
8479 period = d->rt_period;
8480 runtime = d->rt_runtime;
8483 sum += to_ratio(period, runtime);
8492 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8494 struct rt_schedulable_data data = {
8496 .rt_period = period,
8497 .rt_runtime = runtime,
8500 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8503 static int tg_set_bandwidth(struct task_group *tg,
8504 u64 rt_period, u64 rt_runtime)
8508 mutex_lock(&rt_constraints_mutex);
8509 read_lock(&tasklist_lock);
8510 err = __rt_schedulable(tg, rt_period, rt_runtime);
8514 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8515 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8516 tg->rt_bandwidth.rt_runtime = rt_runtime;
8518 for_each_possible_cpu(i) {
8519 struct rt_rq *rt_rq = tg->rt_rq[i];
8521 raw_spin_lock(&rt_rq->rt_runtime_lock);
8522 rt_rq->rt_runtime = rt_runtime;
8523 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8525 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8527 read_unlock(&tasklist_lock);
8528 mutex_unlock(&rt_constraints_mutex);
8533 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8535 u64 rt_runtime, rt_period;
8537 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8538 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8539 if (rt_runtime_us < 0)
8540 rt_runtime = RUNTIME_INF;
8542 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8545 long sched_group_rt_runtime(struct task_group *tg)
8549 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8552 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8553 do_div(rt_runtime_us, NSEC_PER_USEC);
8554 return rt_runtime_us;
8557 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8559 u64 rt_runtime, rt_period;
8561 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8562 rt_runtime = tg->rt_bandwidth.rt_runtime;
8567 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8570 long sched_group_rt_period(struct task_group *tg)
8574 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8575 do_div(rt_period_us, NSEC_PER_USEC);
8576 return rt_period_us;
8579 static int sched_rt_global_constraints(void)
8581 u64 runtime, period;
8584 if (sysctl_sched_rt_period <= 0)
8587 runtime = global_rt_runtime();
8588 period = global_rt_period();
8591 * Sanity check on the sysctl variables.
8593 if (runtime > period && runtime != RUNTIME_INF)
8596 mutex_lock(&rt_constraints_mutex);
8597 read_lock(&tasklist_lock);
8598 ret = __rt_schedulable(NULL, 0, 0);
8599 read_unlock(&tasklist_lock);
8600 mutex_unlock(&rt_constraints_mutex);
8605 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8607 /* Don't accept realtime tasks when there is no way for them to run */
8608 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8614 #else /* !CONFIG_RT_GROUP_SCHED */
8615 static int sched_rt_global_constraints(void)
8617 unsigned long flags;
8620 if (sysctl_sched_rt_period <= 0)
8624 * There's always some RT tasks in the root group
8625 * -- migration, kstopmachine etc..
8627 if (sysctl_sched_rt_runtime == 0)
8630 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8631 for_each_possible_cpu(i) {
8632 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8634 raw_spin_lock(&rt_rq->rt_runtime_lock);
8635 rt_rq->rt_runtime = global_rt_runtime();
8636 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8638 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8642 #endif /* CONFIG_RT_GROUP_SCHED */
8644 int sched_rt_handler(struct ctl_table *table, int write,
8645 void __user *buffer, size_t *lenp,
8649 int old_period, old_runtime;
8650 static DEFINE_MUTEX(mutex);
8653 old_period = sysctl_sched_rt_period;
8654 old_runtime = sysctl_sched_rt_runtime;
8656 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8658 if (!ret && write) {
8659 ret = sched_rt_global_constraints();
8661 sysctl_sched_rt_period = old_period;
8662 sysctl_sched_rt_runtime = old_runtime;
8664 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8665 def_rt_bandwidth.rt_period =
8666 ns_to_ktime(global_rt_period());
8669 mutex_unlock(&mutex);
8674 #ifdef CONFIG_CGROUP_SCHED
8676 /* return corresponding task_group object of a cgroup */
8677 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8679 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8680 struct task_group, css);
8683 static struct cgroup_subsys_state *
8684 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8686 struct task_group *tg, *parent;
8688 if (!cgrp->parent) {
8689 /* This is early initialization for the top cgroup */
8690 return &init_task_group.css;
8693 parent = cgroup_tg(cgrp->parent);
8694 tg = sched_create_group(parent);
8696 return ERR_PTR(-ENOMEM);
8702 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8704 struct task_group *tg = cgroup_tg(cgrp);
8706 sched_destroy_group(tg);
8710 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8712 #ifdef CONFIG_RT_GROUP_SCHED
8713 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8716 /* We don't support RT-tasks being in separate groups */
8717 if (tsk->sched_class != &fair_sched_class)
8724 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8725 struct task_struct *tsk, bool threadgroup)
8727 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8731 struct task_struct *c;
8733 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8734 retval = cpu_cgroup_can_attach_task(cgrp, c);
8746 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8747 struct cgroup *old_cont, struct task_struct *tsk,
8750 sched_move_task(tsk);
8752 struct task_struct *c;
8754 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8761 #ifdef CONFIG_FAIR_GROUP_SCHED
8762 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8765 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8768 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8770 struct task_group *tg = cgroup_tg(cgrp);
8772 return (u64) tg->shares;
8774 #endif /* CONFIG_FAIR_GROUP_SCHED */
8776 #ifdef CONFIG_RT_GROUP_SCHED
8777 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8780 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8783 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8785 return sched_group_rt_runtime(cgroup_tg(cgrp));
8788 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8791 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8794 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8796 return sched_group_rt_period(cgroup_tg(cgrp));
8798 #endif /* CONFIG_RT_GROUP_SCHED */
8800 static struct cftype cpu_files[] = {
8801 #ifdef CONFIG_FAIR_GROUP_SCHED
8804 .read_u64 = cpu_shares_read_u64,
8805 .write_u64 = cpu_shares_write_u64,
8808 #ifdef CONFIG_RT_GROUP_SCHED
8810 .name = "rt_runtime_us",
8811 .read_s64 = cpu_rt_runtime_read,
8812 .write_s64 = cpu_rt_runtime_write,
8815 .name = "rt_period_us",
8816 .read_u64 = cpu_rt_period_read_uint,
8817 .write_u64 = cpu_rt_period_write_uint,
8822 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8824 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8827 struct cgroup_subsys cpu_cgroup_subsys = {
8829 .create = cpu_cgroup_create,
8830 .destroy = cpu_cgroup_destroy,
8831 .can_attach = cpu_cgroup_can_attach,
8832 .attach = cpu_cgroup_attach,
8833 .populate = cpu_cgroup_populate,
8834 .subsys_id = cpu_cgroup_subsys_id,
8838 #endif /* CONFIG_CGROUP_SCHED */
8840 #ifdef CONFIG_CGROUP_CPUACCT
8843 * CPU accounting code for task groups.
8845 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8846 * (balbir@in.ibm.com).
8849 /* track cpu usage of a group of tasks and its child groups */
8851 struct cgroup_subsys_state css;
8852 /* cpuusage holds pointer to a u64-type object on every cpu */
8853 u64 __percpu *cpuusage;
8854 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8855 struct cpuacct *parent;
8858 struct cgroup_subsys cpuacct_subsys;
8860 /* return cpu accounting group corresponding to this container */
8861 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8863 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8864 struct cpuacct, css);
8867 /* return cpu accounting group to which this task belongs */
8868 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8870 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8871 struct cpuacct, css);
8874 /* create a new cpu accounting group */
8875 static struct cgroup_subsys_state *cpuacct_create(
8876 struct cgroup_subsys *ss, struct cgroup *cgrp)
8878 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8884 ca->cpuusage = alloc_percpu(u64);
8888 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8889 if (percpu_counter_init(&ca->cpustat[i], 0))
8890 goto out_free_counters;
8893 ca->parent = cgroup_ca(cgrp->parent);
8899 percpu_counter_destroy(&ca->cpustat[i]);
8900 free_percpu(ca->cpuusage);
8904 return ERR_PTR(-ENOMEM);
8907 /* destroy an existing cpu accounting group */
8909 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8911 struct cpuacct *ca = cgroup_ca(cgrp);
8914 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8915 percpu_counter_destroy(&ca->cpustat[i]);
8916 free_percpu(ca->cpuusage);
8920 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8922 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8925 #ifndef CONFIG_64BIT
8927 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8929 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8931 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8939 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8941 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8943 #ifndef CONFIG_64BIT
8945 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8947 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8949 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8955 /* return total cpu usage (in nanoseconds) of a group */
8956 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8958 struct cpuacct *ca = cgroup_ca(cgrp);
8959 u64 totalcpuusage = 0;
8962 for_each_present_cpu(i)
8963 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8965 return totalcpuusage;
8968 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8971 struct cpuacct *ca = cgroup_ca(cgrp);
8980 for_each_present_cpu(i)
8981 cpuacct_cpuusage_write(ca, i, 0);
8987 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8990 struct cpuacct *ca = cgroup_ca(cgroup);
8994 for_each_present_cpu(i) {
8995 percpu = cpuacct_cpuusage_read(ca, i);
8996 seq_printf(m, "%llu ", (unsigned long long) percpu);
8998 seq_printf(m, "\n");
9002 static const char *cpuacct_stat_desc[] = {
9003 [CPUACCT_STAT_USER] = "user",
9004 [CPUACCT_STAT_SYSTEM] = "system",
9007 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9008 struct cgroup_map_cb *cb)
9010 struct cpuacct *ca = cgroup_ca(cgrp);
9013 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9014 s64 val = percpu_counter_read(&ca->cpustat[i]);
9015 val = cputime64_to_clock_t(val);
9016 cb->fill(cb, cpuacct_stat_desc[i], val);
9021 static struct cftype files[] = {
9024 .read_u64 = cpuusage_read,
9025 .write_u64 = cpuusage_write,
9028 .name = "usage_percpu",
9029 .read_seq_string = cpuacct_percpu_seq_read,
9033 .read_map = cpuacct_stats_show,
9037 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9039 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9043 * charge this task's execution time to its accounting group.
9045 * called with rq->lock held.
9047 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9052 if (unlikely(!cpuacct_subsys.active))
9055 cpu = task_cpu(tsk);
9061 for (; ca; ca = ca->parent) {
9062 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9063 *cpuusage += cputime;
9070 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9071 * in cputime_t units. As a result, cpuacct_update_stats calls
9072 * percpu_counter_add with values large enough to always overflow the
9073 * per cpu batch limit causing bad SMP scalability.
9075 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9076 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9077 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9080 #define CPUACCT_BATCH \
9081 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9083 #define CPUACCT_BATCH 0
9087 * Charge the system/user time to the task's accounting group.
9089 static void cpuacct_update_stats(struct task_struct *tsk,
9090 enum cpuacct_stat_index idx, cputime_t val)
9093 int batch = CPUACCT_BATCH;
9095 if (unlikely(!cpuacct_subsys.active))
9102 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9108 struct cgroup_subsys cpuacct_subsys = {
9110 .create = cpuacct_create,
9111 .destroy = cpuacct_destroy,
9112 .populate = cpuacct_populate,
9113 .subsys_id = cpuacct_subsys_id,
9115 #endif /* CONFIG_CGROUP_CPUACCT */
9119 void synchronize_sched_expedited(void)
9123 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9125 #else /* #ifndef CONFIG_SMP */
9127 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
9129 static int synchronize_sched_expedited_cpu_stop(void *data)
9132 * There must be a full memory barrier on each affected CPU
9133 * between the time that try_stop_cpus() is called and the
9134 * time that it returns.
9136 * In the current initial implementation of cpu_stop, the
9137 * above condition is already met when the control reaches
9138 * this point and the following smp_mb() is not strictly
9139 * necessary. Do smp_mb() anyway for documentation and
9140 * robustness against future implementation changes.
9142 smp_mb(); /* See above comment block. */
9147 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9148 * approach to force grace period to end quickly. This consumes
9149 * significant time on all CPUs, and is thus not recommended for
9150 * any sort of common-case code.
9152 * Note that it is illegal to call this function while holding any
9153 * lock that is acquired by a CPU-hotplug notifier. Failing to
9154 * observe this restriction will result in deadlock.
9156 void synchronize_sched_expedited(void)
9158 int snap, trycount = 0;
9160 smp_mb(); /* ensure prior mod happens before capturing snap. */
9161 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
9163 while (try_stop_cpus(cpu_online_mask,
9164 synchronize_sched_expedited_cpu_stop,
9167 if (trycount++ < 10)
9168 udelay(trycount * num_online_cpus());
9170 synchronize_sched();
9173 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
9174 smp_mb(); /* ensure test happens before caller kfree */
9179 atomic_inc(&synchronize_sched_expedited_count);
9180 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9183 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9185 #endif /* #else #ifndef CONFIG_SMP */