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;
258 #ifdef CONFIG_RT_GROUP_SCHED
259 struct sched_rt_entity **rt_se;
260 struct rt_rq **rt_rq;
262 struct rt_bandwidth rt_bandwidth;
266 struct list_head list;
268 struct task_group *parent;
269 struct list_head siblings;
270 struct list_head children;
273 #define root_task_group init_task_group
275 /* task_group_lock serializes add/remove of task groups and also changes to
276 * a task group's cpu shares.
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.
345 struct list_head leaf_cfs_rq_list;
346 struct task_group *tg; /* group that "owns" this runqueue */
350 * the part of load.weight contributed by tasks
352 unsigned long task_weight;
355 * h_load = weight * f(tg)
357 * Where f(tg) is the recursive weight fraction assigned to
360 unsigned long h_load;
363 * this cpu's part of tg->shares
365 unsigned long shares;
368 * load.weight at the time we set shares
370 unsigned long rq_weight;
375 /* Real-Time classes' related field in a runqueue: */
377 struct rt_prio_array active;
378 unsigned long rt_nr_running;
379 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
381 int curr; /* highest queued rt task prio */
383 int next; /* next highest */
388 unsigned long rt_nr_migratory;
389 unsigned long rt_nr_total;
391 struct plist_head pushable_tasks;
396 /* Nests inside the rq lock: */
397 raw_spinlock_t rt_runtime_lock;
399 #ifdef CONFIG_RT_GROUP_SCHED
400 unsigned long rt_nr_boosted;
403 struct list_head leaf_rt_rq_list;
404 struct task_group *tg;
411 * We add the notion of a root-domain which will be used to define per-domain
412 * variables. Each exclusive cpuset essentially defines an island domain by
413 * fully partitioning the member cpus from any other cpuset. Whenever a new
414 * exclusive cpuset is created, we also create and attach a new root-domain
421 cpumask_var_t online;
424 * The "RT overload" flag: it gets set if a CPU has more than
425 * one runnable RT task.
427 cpumask_var_t rto_mask;
429 struct cpupri cpupri;
433 * By default the system creates a single root-domain with all cpus as
434 * members (mimicking the global state we have today).
436 static struct root_domain def_root_domain;
438 #endif /* CONFIG_SMP */
441 * This is the main, per-CPU runqueue data structure.
443 * Locking rule: those places that want to lock multiple runqueues
444 * (such as the load balancing or the thread migration code), lock
445 * acquire operations must be ordered by ascending &runqueue.
452 * nr_running and cpu_load should be in the same cacheline because
453 * remote CPUs use both these fields when doing load calculation.
455 unsigned long nr_running;
456 #define CPU_LOAD_IDX_MAX 5
457 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
458 unsigned long last_load_update_tick;
461 unsigned char nohz_balance_kick;
463 unsigned int skip_clock_update;
465 /* capture load from *all* tasks on this cpu: */
466 struct load_weight load;
467 unsigned long nr_load_updates;
473 #ifdef CONFIG_FAIR_GROUP_SCHED
474 /* list of leaf cfs_rq on this cpu: */
475 struct list_head leaf_cfs_rq_list;
477 #ifdef CONFIG_RT_GROUP_SCHED
478 struct list_head leaf_rt_rq_list;
482 * This is part of a global counter where only the total sum
483 * over all CPUs matters. A task can increase this counter on
484 * one CPU and if it got migrated afterwards it may decrease
485 * it on another CPU. Always updated under the runqueue lock:
487 unsigned long nr_uninterruptible;
489 struct task_struct *curr, *idle, *stop;
490 unsigned long next_balance;
491 struct mm_struct *prev_mm;
499 struct root_domain *rd;
500 struct sched_domain *sd;
502 unsigned long cpu_power;
504 unsigned char idle_at_tick;
505 /* For active balancing */
509 struct cpu_stop_work active_balance_work;
510 /* cpu of this runqueue: */
514 unsigned long avg_load_per_task;
522 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
526 /* calc_load related fields */
527 unsigned long calc_load_update;
528 long calc_load_active;
530 #ifdef CONFIG_SCHED_HRTICK
532 int hrtick_csd_pending;
533 struct call_single_data hrtick_csd;
535 struct hrtimer hrtick_timer;
538 #ifdef CONFIG_SCHEDSTATS
540 struct sched_info rq_sched_info;
541 unsigned long long rq_cpu_time;
542 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
544 /* sys_sched_yield() stats */
545 unsigned int yld_count;
547 /* schedule() stats */
548 unsigned int sched_switch;
549 unsigned int sched_count;
550 unsigned int sched_goidle;
552 /* try_to_wake_up() stats */
553 unsigned int ttwu_count;
554 unsigned int ttwu_local;
557 unsigned int bkl_count;
561 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
564 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
566 static inline int cpu_of(struct rq *rq)
575 #define rcu_dereference_check_sched_domain(p) \
576 rcu_dereference_check((p), \
577 rcu_read_lock_sched_held() || \
578 lockdep_is_held(&sched_domains_mutex))
581 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
582 * See detach_destroy_domains: synchronize_sched for details.
584 * The domain tree of any CPU may only be accessed from within
585 * preempt-disabled sections.
587 #define for_each_domain(cpu, __sd) \
588 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
590 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
591 #define this_rq() (&__get_cpu_var(runqueues))
592 #define task_rq(p) cpu_rq(task_cpu(p))
593 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
594 #define raw_rq() (&__raw_get_cpu_var(runqueues))
596 #ifdef CONFIG_CGROUP_SCHED
599 * Return the group to which this tasks belongs.
601 * We use task_subsys_state_check() and extend the RCU verification
602 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
603 * holds that lock for each task it moves into the cgroup. Therefore
604 * by holding that lock, we pin the task to the current cgroup.
606 static inline struct task_group *task_group(struct task_struct *p)
608 struct cgroup_subsys_state *css;
610 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
611 lockdep_is_held(&task_rq(p)->lock));
612 return container_of(css, struct task_group, css);
615 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
616 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
618 #ifdef CONFIG_FAIR_GROUP_SCHED
619 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
620 p->se.parent = task_group(p)->se[cpu];
623 #ifdef CONFIG_RT_GROUP_SCHED
624 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
625 p->rt.parent = task_group(p)->rt_se[cpu];
629 #else /* CONFIG_CGROUP_SCHED */
631 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
632 static inline struct task_group *task_group(struct task_struct *p)
637 #endif /* CONFIG_CGROUP_SCHED */
639 static void update_rq_clock_task(struct rq *rq, s64 delta);
641 static void update_rq_clock(struct rq *rq)
645 if (rq->skip_clock_update)
648 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
650 update_rq_clock_task(rq, delta);
654 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
656 #ifdef CONFIG_SCHED_DEBUG
657 # define const_debug __read_mostly
659 # define const_debug static const
664 * @cpu: the processor in question.
666 * Returns true if the current cpu runqueue is locked.
667 * This interface allows printk to be called with the runqueue lock
668 * held and know whether or not it is OK to wake up the klogd.
670 int runqueue_is_locked(int cpu)
672 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
676 * Debugging: various feature bits
679 #define SCHED_FEAT(name, enabled) \
680 __SCHED_FEAT_##name ,
683 #include "sched_features.h"
688 #define SCHED_FEAT(name, enabled) \
689 (1UL << __SCHED_FEAT_##name) * enabled |
691 const_debug unsigned int sysctl_sched_features =
692 #include "sched_features.h"
697 #ifdef CONFIG_SCHED_DEBUG
698 #define SCHED_FEAT(name, enabled) \
701 static __read_mostly char *sched_feat_names[] = {
702 #include "sched_features.h"
708 static int sched_feat_show(struct seq_file *m, void *v)
712 for (i = 0; sched_feat_names[i]; i++) {
713 if (!(sysctl_sched_features & (1UL << i)))
715 seq_printf(m, "%s ", sched_feat_names[i]);
723 sched_feat_write(struct file *filp, const char __user *ubuf,
724 size_t cnt, loff_t *ppos)
734 if (copy_from_user(&buf, ubuf, cnt))
740 if (strncmp(buf, "NO_", 3) == 0) {
745 for (i = 0; sched_feat_names[i]; i++) {
746 if (strcmp(cmp, sched_feat_names[i]) == 0) {
748 sysctl_sched_features &= ~(1UL << i);
750 sysctl_sched_features |= (1UL << i);
755 if (!sched_feat_names[i])
763 static int sched_feat_open(struct inode *inode, struct file *filp)
765 return single_open(filp, sched_feat_show, NULL);
768 static const struct file_operations sched_feat_fops = {
769 .open = sched_feat_open,
770 .write = sched_feat_write,
773 .release = single_release,
776 static __init int sched_init_debug(void)
778 debugfs_create_file("sched_features", 0644, NULL, NULL,
783 late_initcall(sched_init_debug);
787 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
790 * Number of tasks to iterate in a single balance run.
791 * Limited because this is done with IRQs disabled.
793 const_debug unsigned int sysctl_sched_nr_migrate = 32;
796 * ratelimit for updating the group shares.
799 unsigned int sysctl_sched_shares_ratelimit = 250000;
800 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
803 * Inject some fuzzyness into changing the per-cpu group shares
804 * this avoids remote rq-locks at the expense of fairness.
807 unsigned int sysctl_sched_shares_thresh = 4;
810 * period over which we average the RT time consumption, measured
815 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
818 * period over which we measure -rt task cpu usage in us.
821 unsigned int sysctl_sched_rt_period = 1000000;
823 static __read_mostly int scheduler_running;
826 * part of the period that we allow rt tasks to run in us.
829 int sysctl_sched_rt_runtime = 950000;
831 static inline u64 global_rt_period(void)
833 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
836 static inline u64 global_rt_runtime(void)
838 if (sysctl_sched_rt_runtime < 0)
841 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
844 #ifndef prepare_arch_switch
845 # define prepare_arch_switch(next) do { } while (0)
847 #ifndef finish_arch_switch
848 # define finish_arch_switch(prev) do { } while (0)
851 static inline int task_current(struct rq *rq, struct task_struct *p)
853 return rq->curr == p;
856 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
857 static inline int task_running(struct rq *rq, struct task_struct *p)
859 return task_current(rq, p);
862 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
866 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
868 #ifdef CONFIG_DEBUG_SPINLOCK
869 /* this is a valid case when another task releases the spinlock */
870 rq->lock.owner = current;
873 * If we are tracking spinlock dependencies then we have to
874 * fix up the runqueue lock - which gets 'carried over' from
877 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
879 raw_spin_unlock_irq(&rq->lock);
882 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
883 static inline int task_running(struct rq *rq, struct task_struct *p)
888 return task_current(rq, p);
892 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
896 * We can optimise this out completely for !SMP, because the
897 * SMP rebalancing from interrupt is the only thing that cares
902 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
903 raw_spin_unlock_irq(&rq->lock);
905 raw_spin_unlock(&rq->lock);
909 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
913 * After ->oncpu is cleared, the task can be moved to a different CPU.
914 * We must ensure this doesn't happen until the switch is completely
920 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
924 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
927 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
930 static inline int task_is_waking(struct task_struct *p)
932 return unlikely(p->state == TASK_WAKING);
936 * __task_rq_lock - lock the runqueue a given task resides on.
937 * Must be called interrupts disabled.
939 static inline struct rq *__task_rq_lock(struct task_struct *p)
946 raw_spin_lock(&rq->lock);
947 if (likely(rq == task_rq(p)))
949 raw_spin_unlock(&rq->lock);
954 * task_rq_lock - lock the runqueue a given task resides on and disable
955 * interrupts. Note the ordering: we can safely lookup the task_rq without
956 * explicitly disabling preemption.
958 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
964 local_irq_save(*flags);
966 raw_spin_lock(&rq->lock);
967 if (likely(rq == task_rq(p)))
969 raw_spin_unlock_irqrestore(&rq->lock, *flags);
973 static void __task_rq_unlock(struct rq *rq)
976 raw_spin_unlock(&rq->lock);
979 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
982 raw_spin_unlock_irqrestore(&rq->lock, *flags);
986 * this_rq_lock - lock this runqueue and disable interrupts.
988 static struct rq *this_rq_lock(void)
995 raw_spin_lock(&rq->lock);
1000 #ifdef CONFIG_SCHED_HRTICK
1002 * Use HR-timers to deliver accurate preemption points.
1004 * Its all a bit involved since we cannot program an hrt while holding the
1005 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1008 * When we get rescheduled we reprogram the hrtick_timer outside of the
1014 * - enabled by features
1015 * - hrtimer is actually high res
1017 static inline int hrtick_enabled(struct rq *rq)
1019 if (!sched_feat(HRTICK))
1021 if (!cpu_active(cpu_of(rq)))
1023 return hrtimer_is_hres_active(&rq->hrtick_timer);
1026 static void hrtick_clear(struct rq *rq)
1028 if (hrtimer_active(&rq->hrtick_timer))
1029 hrtimer_cancel(&rq->hrtick_timer);
1033 * High-resolution timer tick.
1034 * Runs from hardirq context with interrupts disabled.
1036 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1038 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1040 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1042 raw_spin_lock(&rq->lock);
1043 update_rq_clock(rq);
1044 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1045 raw_spin_unlock(&rq->lock);
1047 return HRTIMER_NORESTART;
1052 * called from hardirq (IPI) context
1054 static void __hrtick_start(void *arg)
1056 struct rq *rq = arg;
1058 raw_spin_lock(&rq->lock);
1059 hrtimer_restart(&rq->hrtick_timer);
1060 rq->hrtick_csd_pending = 0;
1061 raw_spin_unlock(&rq->lock);
1065 * Called to set the hrtick timer state.
1067 * called with rq->lock held and irqs disabled
1069 static void hrtick_start(struct rq *rq, u64 delay)
1071 struct hrtimer *timer = &rq->hrtick_timer;
1072 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1074 hrtimer_set_expires(timer, time);
1076 if (rq == this_rq()) {
1077 hrtimer_restart(timer);
1078 } else if (!rq->hrtick_csd_pending) {
1079 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1080 rq->hrtick_csd_pending = 1;
1085 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1087 int cpu = (int)(long)hcpu;
1090 case CPU_UP_CANCELED:
1091 case CPU_UP_CANCELED_FROZEN:
1092 case CPU_DOWN_PREPARE:
1093 case CPU_DOWN_PREPARE_FROZEN:
1095 case CPU_DEAD_FROZEN:
1096 hrtick_clear(cpu_rq(cpu));
1103 static __init void init_hrtick(void)
1105 hotcpu_notifier(hotplug_hrtick, 0);
1109 * Called to set the hrtick timer state.
1111 * called with rq->lock held and irqs disabled
1113 static void hrtick_start(struct rq *rq, u64 delay)
1115 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1116 HRTIMER_MODE_REL_PINNED, 0);
1119 static inline void init_hrtick(void)
1122 #endif /* CONFIG_SMP */
1124 static void init_rq_hrtick(struct rq *rq)
1127 rq->hrtick_csd_pending = 0;
1129 rq->hrtick_csd.flags = 0;
1130 rq->hrtick_csd.func = __hrtick_start;
1131 rq->hrtick_csd.info = rq;
1134 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1135 rq->hrtick_timer.function = hrtick;
1137 #else /* CONFIG_SCHED_HRTICK */
1138 static inline void hrtick_clear(struct rq *rq)
1142 static inline void init_rq_hrtick(struct rq *rq)
1146 static inline void init_hrtick(void)
1149 #endif /* CONFIG_SCHED_HRTICK */
1152 * resched_task - mark a task 'to be rescheduled now'.
1154 * On UP this means the setting of the need_resched flag, on SMP it
1155 * might also involve a cross-CPU call to trigger the scheduler on
1160 #ifndef tsk_is_polling
1161 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1164 static void resched_task(struct task_struct *p)
1168 assert_raw_spin_locked(&task_rq(p)->lock);
1170 if (test_tsk_need_resched(p))
1173 set_tsk_need_resched(p);
1176 if (cpu == smp_processor_id())
1179 /* NEED_RESCHED must be visible before we test polling */
1181 if (!tsk_is_polling(p))
1182 smp_send_reschedule(cpu);
1185 static void resched_cpu(int cpu)
1187 struct rq *rq = cpu_rq(cpu);
1188 unsigned long flags;
1190 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1192 resched_task(cpu_curr(cpu));
1193 raw_spin_unlock_irqrestore(&rq->lock, flags);
1198 * In the semi idle case, use the nearest busy cpu for migrating timers
1199 * from an idle cpu. This is good for power-savings.
1201 * We don't do similar optimization for completely idle system, as
1202 * selecting an idle cpu will add more delays to the timers than intended
1203 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1205 int get_nohz_timer_target(void)
1207 int cpu = smp_processor_id();
1209 struct sched_domain *sd;
1211 for_each_domain(cpu, sd) {
1212 for_each_cpu(i, sched_domain_span(sd))
1219 * When add_timer_on() enqueues a timer into the timer wheel of an
1220 * idle CPU then this timer might expire before the next timer event
1221 * which is scheduled to wake up that CPU. In case of a completely
1222 * idle system the next event might even be infinite time into the
1223 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1224 * leaves the inner idle loop so the newly added timer is taken into
1225 * account when the CPU goes back to idle and evaluates the timer
1226 * wheel for the next timer event.
1228 void wake_up_idle_cpu(int cpu)
1230 struct rq *rq = cpu_rq(cpu);
1232 if (cpu == smp_processor_id())
1236 * This is safe, as this function is called with the timer
1237 * wheel base lock of (cpu) held. When the CPU is on the way
1238 * to idle and has not yet set rq->curr to idle then it will
1239 * be serialized on the timer wheel base lock and take the new
1240 * timer into account automatically.
1242 if (rq->curr != rq->idle)
1246 * We can set TIF_RESCHED on the idle task of the other CPU
1247 * lockless. The worst case is that the other CPU runs the
1248 * idle task through an additional NOOP schedule()
1250 set_tsk_need_resched(rq->idle);
1252 /* NEED_RESCHED must be visible before we test polling */
1254 if (!tsk_is_polling(rq->idle))
1255 smp_send_reschedule(cpu);
1258 #endif /* CONFIG_NO_HZ */
1260 static u64 sched_avg_period(void)
1262 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1265 static void sched_avg_update(struct rq *rq)
1267 s64 period = sched_avg_period();
1269 while ((s64)(rq->clock - rq->age_stamp) > period) {
1271 * Inline assembly required to prevent the compiler
1272 * optimising this loop into a divmod call.
1273 * See __iter_div_u64_rem() for another example of this.
1275 asm("" : "+rm" (rq->age_stamp));
1276 rq->age_stamp += period;
1281 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1283 rq->rt_avg += rt_delta;
1284 sched_avg_update(rq);
1287 #else /* !CONFIG_SMP */
1288 static void resched_task(struct task_struct *p)
1290 assert_raw_spin_locked(&task_rq(p)->lock);
1291 set_tsk_need_resched(p);
1294 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1298 static void sched_avg_update(struct rq *rq)
1301 #endif /* CONFIG_SMP */
1303 #if BITS_PER_LONG == 32
1304 # define WMULT_CONST (~0UL)
1306 # define WMULT_CONST (1UL << 32)
1309 #define WMULT_SHIFT 32
1312 * Shift right and round:
1314 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1317 * delta *= weight / lw
1319 static unsigned long
1320 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1321 struct load_weight *lw)
1325 if (!lw->inv_weight) {
1326 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1329 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1333 tmp = (u64)delta_exec * weight;
1335 * Check whether we'd overflow the 64-bit multiplication:
1337 if (unlikely(tmp > WMULT_CONST))
1338 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1341 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1343 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1346 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1352 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1359 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1360 * of tasks with abnormal "nice" values across CPUs the contribution that
1361 * each task makes to its run queue's load is weighted according to its
1362 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1363 * scaled version of the new time slice allocation that they receive on time
1367 #define WEIGHT_IDLEPRIO 3
1368 #define WMULT_IDLEPRIO 1431655765
1371 * Nice levels are multiplicative, with a gentle 10% change for every
1372 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1373 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1374 * that remained on nice 0.
1376 * The "10% effect" is relative and cumulative: from _any_ nice level,
1377 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1378 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1379 * If a task goes up by ~10% and another task goes down by ~10% then
1380 * the relative distance between them is ~25%.)
1382 static const int prio_to_weight[40] = {
1383 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1384 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1385 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1386 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1387 /* 0 */ 1024, 820, 655, 526, 423,
1388 /* 5 */ 335, 272, 215, 172, 137,
1389 /* 10 */ 110, 87, 70, 56, 45,
1390 /* 15 */ 36, 29, 23, 18, 15,
1394 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1396 * In cases where the weight does not change often, we can use the
1397 * precalculated inverse to speed up arithmetics by turning divisions
1398 * into multiplications:
1400 static const u32 prio_to_wmult[40] = {
1401 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1402 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1403 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1404 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1405 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1406 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1407 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1408 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1411 /* Time spent by the tasks of the cpu accounting group executing in ... */
1412 enum cpuacct_stat_index {
1413 CPUACCT_STAT_USER, /* ... user mode */
1414 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1416 CPUACCT_STAT_NSTATS,
1419 #ifdef CONFIG_CGROUP_CPUACCT
1420 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1421 static void cpuacct_update_stats(struct task_struct *tsk,
1422 enum cpuacct_stat_index idx, cputime_t val);
1424 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1425 static inline void cpuacct_update_stats(struct task_struct *tsk,
1426 enum cpuacct_stat_index idx, cputime_t val) {}
1429 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1431 update_load_add(&rq->load, load);
1434 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1436 update_load_sub(&rq->load, load);
1439 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1440 typedef int (*tg_visitor)(struct task_group *, void *);
1443 * Iterate the full tree, calling @down when first entering a node and @up when
1444 * leaving it for the final time.
1446 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1448 struct task_group *parent, *child;
1452 parent = &root_task_group;
1454 ret = (*down)(parent, data);
1457 list_for_each_entry_rcu(child, &parent->children, siblings) {
1464 ret = (*up)(parent, data);
1469 parent = parent->parent;
1478 static int tg_nop(struct task_group *tg, void *data)
1485 /* Used instead of source_load when we know the type == 0 */
1486 static unsigned long weighted_cpuload(const int cpu)
1488 return cpu_rq(cpu)->load.weight;
1492 * Return a low guess at the load of a migration-source cpu weighted
1493 * according to the scheduling class and "nice" value.
1495 * We want to under-estimate the load of migration sources, to
1496 * balance conservatively.
1498 static unsigned long source_load(int cpu, int type)
1500 struct rq *rq = cpu_rq(cpu);
1501 unsigned long total = weighted_cpuload(cpu);
1503 if (type == 0 || !sched_feat(LB_BIAS))
1506 return min(rq->cpu_load[type-1], total);
1510 * Return a high guess at the load of a migration-target cpu weighted
1511 * according to the scheduling class and "nice" value.
1513 static unsigned long target_load(int cpu, int type)
1515 struct rq *rq = cpu_rq(cpu);
1516 unsigned long total = weighted_cpuload(cpu);
1518 if (type == 0 || !sched_feat(LB_BIAS))
1521 return max(rq->cpu_load[type-1], total);
1524 static unsigned long power_of(int cpu)
1526 return cpu_rq(cpu)->cpu_power;
1529 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1531 static unsigned long cpu_avg_load_per_task(int cpu)
1533 struct rq *rq = cpu_rq(cpu);
1534 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1537 rq->avg_load_per_task = rq->load.weight / nr_running;
1539 rq->avg_load_per_task = 0;
1541 return rq->avg_load_per_task;
1544 #ifdef CONFIG_FAIR_GROUP_SCHED
1546 static __read_mostly unsigned long __percpu *update_shares_data;
1548 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1551 * Calculate and set the cpu's group shares.
1553 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1554 unsigned long sd_shares,
1555 unsigned long sd_rq_weight,
1556 unsigned long *usd_rq_weight)
1558 unsigned long shares, rq_weight;
1561 rq_weight = usd_rq_weight[cpu];
1564 rq_weight = NICE_0_LOAD;
1568 * \Sum_j shares_j * rq_weight_i
1569 * shares_i = -----------------------------
1570 * \Sum_j rq_weight_j
1572 shares = (sd_shares * rq_weight) / sd_rq_weight;
1573 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1575 if (abs(shares - tg->se[cpu]->load.weight) >
1576 sysctl_sched_shares_thresh) {
1577 struct rq *rq = cpu_rq(cpu);
1578 unsigned long flags;
1580 raw_spin_lock_irqsave(&rq->lock, flags);
1581 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1582 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1583 __set_se_shares(tg->se[cpu], shares);
1584 raw_spin_unlock_irqrestore(&rq->lock, flags);
1589 * Re-compute the task group their per cpu shares over the given domain.
1590 * This needs to be done in a bottom-up fashion because the rq weight of a
1591 * parent group depends on the shares of its child groups.
1593 static int tg_shares_up(struct task_group *tg, void *data)
1595 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1596 unsigned long *usd_rq_weight;
1597 struct sched_domain *sd = data;
1598 unsigned long flags;
1604 local_irq_save(flags);
1605 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1607 for_each_cpu(i, sched_domain_span(sd)) {
1608 weight = tg->cfs_rq[i]->load.weight;
1609 usd_rq_weight[i] = weight;
1611 rq_weight += weight;
1613 * If there are currently no tasks on the cpu pretend there
1614 * is one of average load so that when a new task gets to
1615 * run here it will not get delayed by group starvation.
1618 weight = NICE_0_LOAD;
1620 sum_weight += weight;
1621 shares += tg->cfs_rq[i]->shares;
1625 rq_weight = sum_weight;
1627 if ((!shares && rq_weight) || shares > tg->shares)
1628 shares = tg->shares;
1630 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1631 shares = tg->shares;
1633 for_each_cpu(i, sched_domain_span(sd))
1634 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1636 local_irq_restore(flags);
1642 * Compute the cpu's hierarchical load factor for each task group.
1643 * This needs to be done in a top-down fashion because the load of a child
1644 * group is a fraction of its parents load.
1646 static int tg_load_down(struct task_group *tg, void *data)
1649 long cpu = (long)data;
1652 load = cpu_rq(cpu)->load.weight;
1654 load = tg->parent->cfs_rq[cpu]->h_load;
1655 load *= tg->cfs_rq[cpu]->shares;
1656 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1659 tg->cfs_rq[cpu]->h_load = load;
1664 static void update_shares(struct sched_domain *sd)
1669 if (root_task_group_empty())
1672 now = local_clock();
1673 elapsed = now - sd->last_update;
1675 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1676 sd->last_update = now;
1677 walk_tg_tree(tg_nop, tg_shares_up, sd);
1681 static void update_h_load(long cpu)
1683 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1688 static inline void update_shares(struct sched_domain *sd)
1694 #ifdef CONFIG_PREEMPT
1696 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1699 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1700 * way at the expense of forcing extra atomic operations in all
1701 * invocations. This assures that the double_lock is acquired using the
1702 * same underlying policy as the spinlock_t on this architecture, which
1703 * reduces latency compared to the unfair variant below. However, it
1704 * also adds more overhead and therefore may reduce throughput.
1706 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1707 __releases(this_rq->lock)
1708 __acquires(busiest->lock)
1709 __acquires(this_rq->lock)
1711 raw_spin_unlock(&this_rq->lock);
1712 double_rq_lock(this_rq, busiest);
1719 * Unfair double_lock_balance: Optimizes throughput at the expense of
1720 * latency by eliminating extra atomic operations when the locks are
1721 * already in proper order on entry. This favors lower cpu-ids and will
1722 * grant the double lock to lower cpus over higher ids under contention,
1723 * regardless of entry order into the function.
1725 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1726 __releases(this_rq->lock)
1727 __acquires(busiest->lock)
1728 __acquires(this_rq->lock)
1732 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1733 if (busiest < this_rq) {
1734 raw_spin_unlock(&this_rq->lock);
1735 raw_spin_lock(&busiest->lock);
1736 raw_spin_lock_nested(&this_rq->lock,
1737 SINGLE_DEPTH_NESTING);
1740 raw_spin_lock_nested(&busiest->lock,
1741 SINGLE_DEPTH_NESTING);
1746 #endif /* CONFIG_PREEMPT */
1749 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1751 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1753 if (unlikely(!irqs_disabled())) {
1754 /* printk() doesn't work good under rq->lock */
1755 raw_spin_unlock(&this_rq->lock);
1759 return _double_lock_balance(this_rq, busiest);
1762 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1763 __releases(busiest->lock)
1765 raw_spin_unlock(&busiest->lock);
1766 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1770 * double_rq_lock - safely lock two runqueues
1772 * Note this does not disable interrupts like task_rq_lock,
1773 * you need to do so manually before calling.
1775 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1776 __acquires(rq1->lock)
1777 __acquires(rq2->lock)
1779 BUG_ON(!irqs_disabled());
1781 raw_spin_lock(&rq1->lock);
1782 __acquire(rq2->lock); /* Fake it out ;) */
1785 raw_spin_lock(&rq1->lock);
1786 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1788 raw_spin_lock(&rq2->lock);
1789 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1795 * double_rq_unlock - safely unlock two runqueues
1797 * Note this does not restore interrupts like task_rq_unlock,
1798 * you need to do so manually after calling.
1800 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1801 __releases(rq1->lock)
1802 __releases(rq2->lock)
1804 raw_spin_unlock(&rq1->lock);
1806 raw_spin_unlock(&rq2->lock);
1808 __release(rq2->lock);
1813 #ifdef CONFIG_FAIR_GROUP_SCHED
1814 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1817 cfs_rq->shares = shares;
1822 static void calc_load_account_idle(struct rq *this_rq);
1823 static void update_sysctl(void);
1824 static int get_update_sysctl_factor(void);
1825 static void update_cpu_load(struct rq *this_rq);
1827 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1829 set_task_rq(p, cpu);
1832 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1833 * successfuly executed on another CPU. We must ensure that updates of
1834 * per-task data have been completed by this moment.
1837 task_thread_info(p)->cpu = cpu;
1841 static const struct sched_class rt_sched_class;
1843 #define sched_class_highest (&stop_sched_class)
1844 #define for_each_class(class) \
1845 for (class = sched_class_highest; class; class = class->next)
1847 #include "sched_stats.h"
1849 static void inc_nr_running(struct rq *rq)
1854 static void dec_nr_running(struct rq *rq)
1859 static void set_load_weight(struct task_struct *p)
1862 * SCHED_IDLE tasks get minimal weight:
1864 if (p->policy == SCHED_IDLE) {
1865 p->se.load.weight = WEIGHT_IDLEPRIO;
1866 p->se.load.inv_weight = WMULT_IDLEPRIO;
1870 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1871 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1874 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1876 update_rq_clock(rq);
1877 sched_info_queued(p);
1878 p->sched_class->enqueue_task(rq, p, flags);
1882 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1884 update_rq_clock(rq);
1885 sched_info_dequeued(p);
1886 p->sched_class->dequeue_task(rq, p, flags);
1891 * activate_task - move a task to the runqueue.
1893 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1895 if (task_contributes_to_load(p))
1896 rq->nr_uninterruptible--;
1898 enqueue_task(rq, p, flags);
1903 * deactivate_task - remove a task from the runqueue.
1905 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1907 if (task_contributes_to_load(p))
1908 rq->nr_uninterruptible++;
1910 dequeue_task(rq, p, flags);
1914 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1917 * There are no locks covering percpu hardirq/softirq time.
1918 * They are only modified in account_system_vtime, on corresponding CPU
1919 * with interrupts disabled. So, writes are safe.
1920 * They are read and saved off onto struct rq in update_rq_clock().
1921 * This may result in other CPU reading this CPU's irq time and can
1922 * race with irq/account_system_vtime on this CPU. We would either get old
1923 * or new value with a side effect of accounting a slice of irq time to wrong
1924 * task when irq is in progress while we read rq->clock. That is a worthy
1925 * compromise in place of having locks on each irq in account_system_time.
1927 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1928 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1930 static DEFINE_PER_CPU(u64, irq_start_time);
1931 static int sched_clock_irqtime;
1933 void enable_sched_clock_irqtime(void)
1935 sched_clock_irqtime = 1;
1938 void disable_sched_clock_irqtime(void)
1940 sched_clock_irqtime = 0;
1943 #ifndef CONFIG_64BIT
1944 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1946 static inline void irq_time_write_begin(void)
1948 __this_cpu_inc(irq_time_seq.sequence);
1952 static inline void irq_time_write_end(void)
1955 __this_cpu_inc(irq_time_seq.sequence);
1958 static inline u64 irq_time_read(int cpu)
1964 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1965 irq_time = per_cpu(cpu_softirq_time, cpu) +
1966 per_cpu(cpu_hardirq_time, cpu);
1967 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1971 #else /* CONFIG_64BIT */
1972 static inline void irq_time_write_begin(void)
1976 static inline void irq_time_write_end(void)
1980 static inline u64 irq_time_read(int cpu)
1982 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1984 #endif /* CONFIG_64BIT */
1987 * Called before incrementing preempt_count on {soft,}irq_enter
1988 * and before decrementing preempt_count on {soft,}irq_exit.
1990 void account_system_vtime(struct task_struct *curr)
1992 unsigned long flags;
1996 if (!sched_clock_irqtime)
1999 local_irq_save(flags);
2001 cpu = smp_processor_id();
2002 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
2003 __this_cpu_add(irq_start_time, delta);
2005 irq_time_write_begin();
2007 * We do not account for softirq time from ksoftirqd here.
2008 * We want to continue accounting softirq time to ksoftirqd thread
2009 * in that case, so as not to confuse scheduler with a special task
2010 * that do not consume any time, but still wants to run.
2012 if (hardirq_count())
2013 __this_cpu_add(cpu_hardirq_time, delta);
2014 else if (in_serving_softirq() && !(curr->flags & PF_KSOFTIRQD))
2015 __this_cpu_add(cpu_softirq_time, delta);
2017 irq_time_write_end();
2018 local_irq_restore(flags);
2020 EXPORT_SYMBOL_GPL(account_system_vtime);
2022 static void update_rq_clock_task(struct rq *rq, s64 delta)
2026 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
2029 * Since irq_time is only updated on {soft,}irq_exit, we might run into
2030 * this case when a previous update_rq_clock() happened inside a
2031 * {soft,}irq region.
2033 * When this happens, we stop ->clock_task and only update the
2034 * prev_irq_time stamp to account for the part that fit, so that a next
2035 * update will consume the rest. This ensures ->clock_task is
2038 * It does however cause some slight miss-attribution of {soft,}irq
2039 * time, a more accurate solution would be to update the irq_time using
2040 * the current rq->clock timestamp, except that would require using
2043 if (irq_delta > delta)
2046 rq->prev_irq_time += irq_delta;
2048 rq->clock_task += delta;
2050 if (irq_delta && sched_feat(NONIRQ_POWER))
2051 sched_rt_avg_update(rq, irq_delta);
2054 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2056 static void update_rq_clock_task(struct rq *rq, s64 delta)
2058 rq->clock_task += delta;
2061 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2063 #include "sched_idletask.c"
2064 #include "sched_fair.c"
2065 #include "sched_rt.c"
2066 #include "sched_stoptask.c"
2067 #ifdef CONFIG_SCHED_DEBUG
2068 # include "sched_debug.c"
2071 void sched_set_stop_task(int cpu, struct task_struct *stop)
2073 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2074 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2078 * Make it appear like a SCHED_FIFO task, its something
2079 * userspace knows about and won't get confused about.
2081 * Also, it will make PI more or less work without too
2082 * much confusion -- but then, stop work should not
2083 * rely on PI working anyway.
2085 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2087 stop->sched_class = &stop_sched_class;
2090 cpu_rq(cpu)->stop = stop;
2094 * Reset it back to a normal scheduling class so that
2095 * it can die in pieces.
2097 old_stop->sched_class = &rt_sched_class;
2102 * __normal_prio - return the priority that is based on the static prio
2104 static inline int __normal_prio(struct task_struct *p)
2106 return p->static_prio;
2110 * Calculate the expected normal priority: i.e. priority
2111 * without taking RT-inheritance into account. Might be
2112 * boosted by interactivity modifiers. Changes upon fork,
2113 * setprio syscalls, and whenever the interactivity
2114 * estimator recalculates.
2116 static inline int normal_prio(struct task_struct *p)
2120 if (task_has_rt_policy(p))
2121 prio = MAX_RT_PRIO-1 - p->rt_priority;
2123 prio = __normal_prio(p);
2128 * Calculate the current priority, i.e. the priority
2129 * taken into account by the scheduler. This value might
2130 * be boosted by RT tasks, or might be boosted by
2131 * interactivity modifiers. Will be RT if the task got
2132 * RT-boosted. If not then it returns p->normal_prio.
2134 static int effective_prio(struct task_struct *p)
2136 p->normal_prio = normal_prio(p);
2138 * If we are RT tasks or we were boosted to RT priority,
2139 * keep the priority unchanged. Otherwise, update priority
2140 * to the normal priority:
2142 if (!rt_prio(p->prio))
2143 return p->normal_prio;
2148 * task_curr - is this task currently executing on a CPU?
2149 * @p: the task in question.
2151 inline int task_curr(const struct task_struct *p)
2153 return cpu_curr(task_cpu(p)) == p;
2156 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2157 const struct sched_class *prev_class,
2158 int oldprio, int running)
2160 if (prev_class != p->sched_class) {
2161 if (prev_class->switched_from)
2162 prev_class->switched_from(rq, p, running);
2163 p->sched_class->switched_to(rq, p, running);
2165 p->sched_class->prio_changed(rq, p, oldprio, running);
2168 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2170 const struct sched_class *class;
2172 if (p->sched_class == rq->curr->sched_class) {
2173 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2175 for_each_class(class) {
2176 if (class == rq->curr->sched_class)
2178 if (class == p->sched_class) {
2179 resched_task(rq->curr);
2186 * A queue event has occurred, and we're going to schedule. In
2187 * this case, we can save a useless back to back clock update.
2189 if (rq->curr->se.on_rq && test_tsk_need_resched(rq->curr))
2190 rq->skip_clock_update = 1;
2195 * Is this task likely cache-hot:
2198 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2202 if (p->sched_class != &fair_sched_class)
2205 if (unlikely(p->policy == SCHED_IDLE))
2209 * Buddy candidates are cache hot:
2211 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2212 (&p->se == cfs_rq_of(&p->se)->next ||
2213 &p->se == cfs_rq_of(&p->se)->last))
2216 if (sysctl_sched_migration_cost == -1)
2218 if (sysctl_sched_migration_cost == 0)
2221 delta = now - p->se.exec_start;
2223 return delta < (s64)sysctl_sched_migration_cost;
2226 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2228 #ifdef CONFIG_SCHED_DEBUG
2230 * We should never call set_task_cpu() on a blocked task,
2231 * ttwu() will sort out the placement.
2233 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2234 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2237 trace_sched_migrate_task(p, new_cpu);
2239 if (task_cpu(p) != new_cpu) {
2240 p->se.nr_migrations++;
2241 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2244 __set_task_cpu(p, new_cpu);
2247 struct migration_arg {
2248 struct task_struct *task;
2252 static int migration_cpu_stop(void *data);
2255 * The task's runqueue lock must be held.
2256 * Returns true if you have to wait for migration thread.
2258 static bool migrate_task(struct task_struct *p, int dest_cpu)
2260 struct rq *rq = task_rq(p);
2263 * If the task is not on a runqueue (and not running), then
2264 * the next wake-up will properly place the task.
2266 return p->se.on_rq || task_running(rq, p);
2270 * wait_task_inactive - wait for a thread to unschedule.
2272 * If @match_state is nonzero, it's the @p->state value just checked and
2273 * not expected to change. If it changes, i.e. @p might have woken up,
2274 * then return zero. When we succeed in waiting for @p to be off its CPU,
2275 * we return a positive number (its total switch count). If a second call
2276 * a short while later returns the same number, the caller can be sure that
2277 * @p has remained unscheduled the whole time.
2279 * The caller must ensure that the task *will* unschedule sometime soon,
2280 * else this function might spin for a *long* time. This function can't
2281 * be called with interrupts off, or it may introduce deadlock with
2282 * smp_call_function() if an IPI is sent by the same process we are
2283 * waiting to become inactive.
2285 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2287 unsigned long flags;
2294 * We do the initial early heuristics without holding
2295 * any task-queue locks at all. We'll only try to get
2296 * the runqueue lock when things look like they will
2302 * If the task is actively running on another CPU
2303 * still, just relax and busy-wait without holding
2306 * NOTE! Since we don't hold any locks, it's not
2307 * even sure that "rq" stays as the right runqueue!
2308 * But we don't care, since "task_running()" will
2309 * return false if the runqueue has changed and p
2310 * is actually now running somewhere else!
2312 while (task_running(rq, p)) {
2313 if (match_state && unlikely(p->state != match_state))
2319 * Ok, time to look more closely! We need the rq
2320 * lock now, to be *sure*. If we're wrong, we'll
2321 * just go back and repeat.
2323 rq = task_rq_lock(p, &flags);
2324 trace_sched_wait_task(p);
2325 running = task_running(rq, p);
2326 on_rq = p->se.on_rq;
2328 if (!match_state || p->state == match_state)
2329 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2330 task_rq_unlock(rq, &flags);
2333 * If it changed from the expected state, bail out now.
2335 if (unlikely(!ncsw))
2339 * Was it really running after all now that we
2340 * checked with the proper locks actually held?
2342 * Oops. Go back and try again..
2344 if (unlikely(running)) {
2350 * It's not enough that it's not actively running,
2351 * it must be off the runqueue _entirely_, and not
2354 * So if it was still runnable (but just not actively
2355 * running right now), it's preempted, and we should
2356 * yield - it could be a while.
2358 if (unlikely(on_rq)) {
2359 schedule_timeout_uninterruptible(1);
2364 * Ahh, all good. It wasn't running, and it wasn't
2365 * runnable, which means that it will never become
2366 * running in the future either. We're all done!
2375 * kick_process - kick a running thread to enter/exit the kernel
2376 * @p: the to-be-kicked thread
2378 * Cause a process which is running on another CPU to enter
2379 * kernel-mode, without any delay. (to get signals handled.)
2381 * NOTE: this function doesnt have to take the runqueue lock,
2382 * because all it wants to ensure is that the remote task enters
2383 * the kernel. If the IPI races and the task has been migrated
2384 * to another CPU then no harm is done and the purpose has been
2387 void kick_process(struct task_struct *p)
2393 if ((cpu != smp_processor_id()) && task_curr(p))
2394 smp_send_reschedule(cpu);
2397 EXPORT_SYMBOL_GPL(kick_process);
2398 #endif /* CONFIG_SMP */
2401 * task_oncpu_function_call - call a function on the cpu on which a task runs
2402 * @p: the task to evaluate
2403 * @func: the function to be called
2404 * @info: the function call argument
2406 * Calls the function @func when the task is currently running. This might
2407 * be on the current CPU, which just calls the function directly
2409 void task_oncpu_function_call(struct task_struct *p,
2410 void (*func) (void *info), void *info)
2417 smp_call_function_single(cpu, func, info, 1);
2423 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2425 static int select_fallback_rq(int cpu, struct task_struct *p)
2428 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2430 /* Look for allowed, online CPU in same node. */
2431 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2432 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2435 /* Any allowed, online CPU? */
2436 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2437 if (dest_cpu < nr_cpu_ids)
2440 /* No more Mr. Nice Guy. */
2441 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2442 dest_cpu = cpuset_cpus_allowed_fallback(p);
2444 * Don't tell them about moving exiting tasks or
2445 * kernel threads (both mm NULL), since they never
2448 if (p->mm && printk_ratelimit()) {
2449 printk(KERN_INFO "process %d (%s) no "
2450 "longer affine to cpu%d\n",
2451 task_pid_nr(p), p->comm, cpu);
2459 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2462 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2464 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2467 * In order not to call set_task_cpu() on a blocking task we need
2468 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2471 * Since this is common to all placement strategies, this lives here.
2473 * [ this allows ->select_task() to simply return task_cpu(p) and
2474 * not worry about this generic constraint ]
2476 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2478 cpu = select_fallback_rq(task_cpu(p), p);
2483 static void update_avg(u64 *avg, u64 sample)
2485 s64 diff = sample - *avg;
2490 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2491 bool is_sync, bool is_migrate, bool is_local,
2492 unsigned long en_flags)
2494 schedstat_inc(p, se.statistics.nr_wakeups);
2496 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2498 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2500 schedstat_inc(p, se.statistics.nr_wakeups_local);
2502 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2504 activate_task(rq, p, en_flags);
2507 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2508 int wake_flags, bool success)
2510 trace_sched_wakeup(p, success);
2511 check_preempt_curr(rq, p, wake_flags);
2513 p->state = TASK_RUNNING;
2515 if (p->sched_class->task_woken)
2516 p->sched_class->task_woken(rq, p);
2518 if (unlikely(rq->idle_stamp)) {
2519 u64 delta = rq->clock - rq->idle_stamp;
2520 u64 max = 2*sysctl_sched_migration_cost;
2525 update_avg(&rq->avg_idle, delta);
2529 /* if a worker is waking up, notify workqueue */
2530 if ((p->flags & PF_WQ_WORKER) && success)
2531 wq_worker_waking_up(p, cpu_of(rq));
2535 * try_to_wake_up - wake up a thread
2536 * @p: the thread to be awakened
2537 * @state: the mask of task states that can be woken
2538 * @wake_flags: wake modifier flags (WF_*)
2540 * Put it on the run-queue if it's not already there. The "current"
2541 * thread is always on the run-queue (except when the actual
2542 * re-schedule is in progress), and as such you're allowed to do
2543 * the simpler "current->state = TASK_RUNNING" to mark yourself
2544 * runnable without the overhead of this.
2546 * Returns %true if @p was woken up, %false if it was already running
2547 * or @state didn't match @p's state.
2549 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2552 int cpu, orig_cpu, this_cpu, success = 0;
2553 unsigned long flags;
2554 unsigned long en_flags = ENQUEUE_WAKEUP;
2557 this_cpu = get_cpu();
2560 rq = task_rq_lock(p, &flags);
2561 if (!(p->state & state))
2571 if (unlikely(task_running(rq, p)))
2575 * In order to handle concurrent wakeups and release the rq->lock
2576 * we put the task in TASK_WAKING state.
2578 * First fix up the nr_uninterruptible count:
2580 if (task_contributes_to_load(p)) {
2581 if (likely(cpu_online(orig_cpu)))
2582 rq->nr_uninterruptible--;
2584 this_rq()->nr_uninterruptible--;
2586 p->state = TASK_WAKING;
2588 if (p->sched_class->task_waking) {
2589 p->sched_class->task_waking(rq, p);
2590 en_flags |= ENQUEUE_WAKING;
2593 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2594 if (cpu != orig_cpu)
2595 set_task_cpu(p, cpu);
2596 __task_rq_unlock(rq);
2599 raw_spin_lock(&rq->lock);
2602 * We migrated the task without holding either rq->lock, however
2603 * since the task is not on the task list itself, nobody else
2604 * will try and migrate the task, hence the rq should match the
2605 * cpu we just moved it to.
2607 WARN_ON(task_cpu(p) != cpu);
2608 WARN_ON(p->state != TASK_WAKING);
2610 #ifdef CONFIG_SCHEDSTATS
2611 schedstat_inc(rq, ttwu_count);
2612 if (cpu == this_cpu)
2613 schedstat_inc(rq, ttwu_local);
2615 struct sched_domain *sd;
2616 for_each_domain(this_cpu, sd) {
2617 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2618 schedstat_inc(sd, ttwu_wake_remote);
2623 #endif /* CONFIG_SCHEDSTATS */
2626 #endif /* CONFIG_SMP */
2627 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2628 cpu == this_cpu, en_flags);
2631 ttwu_post_activation(p, rq, wake_flags, success);
2633 task_rq_unlock(rq, &flags);
2640 * try_to_wake_up_local - try to wake up a local task with rq lock held
2641 * @p: the thread to be awakened
2643 * Put @p on the run-queue if it's not alredy there. The caller must
2644 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2645 * the current task. this_rq() stays locked over invocation.
2647 static void try_to_wake_up_local(struct task_struct *p)
2649 struct rq *rq = task_rq(p);
2650 bool success = false;
2652 BUG_ON(rq != this_rq());
2653 BUG_ON(p == current);
2654 lockdep_assert_held(&rq->lock);
2656 if (!(p->state & TASK_NORMAL))
2660 if (likely(!task_running(rq, p))) {
2661 schedstat_inc(rq, ttwu_count);
2662 schedstat_inc(rq, ttwu_local);
2664 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2667 ttwu_post_activation(p, rq, 0, success);
2671 * wake_up_process - Wake up a specific process
2672 * @p: The process to be woken up.
2674 * Attempt to wake up the nominated process and move it to the set of runnable
2675 * processes. Returns 1 if the process was woken up, 0 if it was already
2678 * It may be assumed that this function implies a write memory barrier before
2679 * changing the task state if and only if any tasks are woken up.
2681 int wake_up_process(struct task_struct *p)
2683 return try_to_wake_up(p, TASK_ALL, 0);
2685 EXPORT_SYMBOL(wake_up_process);
2687 int wake_up_state(struct task_struct *p, unsigned int state)
2689 return try_to_wake_up(p, state, 0);
2693 * Perform scheduler related setup for a newly forked process p.
2694 * p is forked by current.
2696 * __sched_fork() is basic setup used by init_idle() too:
2698 static void __sched_fork(struct task_struct *p)
2700 p->se.exec_start = 0;
2701 p->se.sum_exec_runtime = 0;
2702 p->se.prev_sum_exec_runtime = 0;
2703 p->se.nr_migrations = 0;
2705 #ifdef CONFIG_SCHEDSTATS
2706 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2709 INIT_LIST_HEAD(&p->rt.run_list);
2711 INIT_LIST_HEAD(&p->se.group_node);
2713 #ifdef CONFIG_PREEMPT_NOTIFIERS
2714 INIT_HLIST_HEAD(&p->preempt_notifiers);
2719 * fork()/clone()-time setup:
2721 void sched_fork(struct task_struct *p, int clone_flags)
2723 int cpu = get_cpu();
2727 * We mark the process as running here. This guarantees that
2728 * nobody will actually run it, and a signal or other external
2729 * event cannot wake it up and insert it on the runqueue either.
2731 p->state = TASK_RUNNING;
2734 * Revert to default priority/policy on fork if requested.
2736 if (unlikely(p->sched_reset_on_fork)) {
2737 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2738 p->policy = SCHED_NORMAL;
2739 p->normal_prio = p->static_prio;
2742 if (PRIO_TO_NICE(p->static_prio) < 0) {
2743 p->static_prio = NICE_TO_PRIO(0);
2744 p->normal_prio = p->static_prio;
2749 * We don't need the reset flag anymore after the fork. It has
2750 * fulfilled its duty:
2752 p->sched_reset_on_fork = 0;
2756 * Make sure we do not leak PI boosting priority to the child.
2758 p->prio = current->normal_prio;
2760 if (!rt_prio(p->prio))
2761 p->sched_class = &fair_sched_class;
2763 if (p->sched_class->task_fork)
2764 p->sched_class->task_fork(p);
2767 * The child is not yet in the pid-hash so no cgroup attach races,
2768 * and the cgroup is pinned to this child due to cgroup_fork()
2769 * is ran before sched_fork().
2771 * Silence PROVE_RCU.
2774 set_task_cpu(p, cpu);
2777 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2778 if (likely(sched_info_on()))
2779 memset(&p->sched_info, 0, sizeof(p->sched_info));
2781 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2784 #ifdef CONFIG_PREEMPT
2785 /* Want to start with kernel preemption disabled. */
2786 task_thread_info(p)->preempt_count = 1;
2788 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2794 * wake_up_new_task - wake up a newly created task for the first time.
2796 * This function will do some initial scheduler statistics housekeeping
2797 * that must be done for every newly created context, then puts the task
2798 * on the runqueue and wakes it.
2800 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2802 unsigned long flags;
2804 int cpu __maybe_unused = get_cpu();
2807 rq = task_rq_lock(p, &flags);
2808 p->state = TASK_WAKING;
2811 * Fork balancing, do it here and not earlier because:
2812 * - cpus_allowed can change in the fork path
2813 * - any previously selected cpu might disappear through hotplug
2815 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2816 * without people poking at ->cpus_allowed.
2818 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2819 set_task_cpu(p, cpu);
2821 p->state = TASK_RUNNING;
2822 task_rq_unlock(rq, &flags);
2825 rq = task_rq_lock(p, &flags);
2826 activate_task(rq, p, 0);
2827 trace_sched_wakeup_new(p, 1);
2828 check_preempt_curr(rq, p, WF_FORK);
2830 if (p->sched_class->task_woken)
2831 p->sched_class->task_woken(rq, p);
2833 task_rq_unlock(rq, &flags);
2837 #ifdef CONFIG_PREEMPT_NOTIFIERS
2840 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2841 * @notifier: notifier struct to register
2843 void preempt_notifier_register(struct preempt_notifier *notifier)
2845 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2847 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2850 * preempt_notifier_unregister - no longer interested in preemption notifications
2851 * @notifier: notifier struct to unregister
2853 * This is safe to call from within a preemption notifier.
2855 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2857 hlist_del(¬ifier->link);
2859 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2861 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2863 struct preempt_notifier *notifier;
2864 struct hlist_node *node;
2866 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2867 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2871 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2872 struct task_struct *next)
2874 struct preempt_notifier *notifier;
2875 struct hlist_node *node;
2877 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2878 notifier->ops->sched_out(notifier, next);
2881 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2883 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2888 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2889 struct task_struct *next)
2893 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2896 * prepare_task_switch - prepare to switch tasks
2897 * @rq: the runqueue preparing to switch
2898 * @prev: the current task that is being switched out
2899 * @next: the task we are going to switch to.
2901 * This is called with the rq lock held and interrupts off. It must
2902 * be paired with a subsequent finish_task_switch after the context
2905 * prepare_task_switch sets up locking and calls architecture specific
2909 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2910 struct task_struct *next)
2912 fire_sched_out_preempt_notifiers(prev, next);
2913 prepare_lock_switch(rq, next);
2914 prepare_arch_switch(next);
2918 * finish_task_switch - clean up after a task-switch
2919 * @rq: runqueue associated with task-switch
2920 * @prev: the thread we just switched away from.
2922 * finish_task_switch must be called after the context switch, paired
2923 * with a prepare_task_switch call before the context switch.
2924 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2925 * and do any other architecture-specific cleanup actions.
2927 * Note that we may have delayed dropping an mm in context_switch(). If
2928 * so, we finish that here outside of the runqueue lock. (Doing it
2929 * with the lock held can cause deadlocks; see schedule() for
2932 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2933 __releases(rq->lock)
2935 struct mm_struct *mm = rq->prev_mm;
2941 * A task struct has one reference for the use as "current".
2942 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2943 * schedule one last time. The schedule call will never return, and
2944 * the scheduled task must drop that reference.
2945 * The test for TASK_DEAD must occur while the runqueue locks are
2946 * still held, otherwise prev could be scheduled on another cpu, die
2947 * there before we look at prev->state, and then the reference would
2949 * Manfred Spraul <manfred@colorfullife.com>
2951 prev_state = prev->state;
2952 finish_arch_switch(prev);
2953 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2954 local_irq_disable();
2955 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2956 perf_event_task_sched_in(current);
2957 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2959 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2960 finish_lock_switch(rq, prev);
2962 fire_sched_in_preempt_notifiers(current);
2965 if (unlikely(prev_state == TASK_DEAD)) {
2967 * Remove function-return probe instances associated with this
2968 * task and put them back on the free list.
2970 kprobe_flush_task(prev);
2971 put_task_struct(prev);
2977 /* assumes rq->lock is held */
2978 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2980 if (prev->sched_class->pre_schedule)
2981 prev->sched_class->pre_schedule(rq, prev);
2984 /* rq->lock is NOT held, but preemption is disabled */
2985 static inline void post_schedule(struct rq *rq)
2987 if (rq->post_schedule) {
2988 unsigned long flags;
2990 raw_spin_lock_irqsave(&rq->lock, flags);
2991 if (rq->curr->sched_class->post_schedule)
2992 rq->curr->sched_class->post_schedule(rq);
2993 raw_spin_unlock_irqrestore(&rq->lock, flags);
2995 rq->post_schedule = 0;
3001 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3005 static inline void post_schedule(struct rq *rq)
3012 * schedule_tail - first thing a freshly forked thread must call.
3013 * @prev: the thread we just switched away from.
3015 asmlinkage void schedule_tail(struct task_struct *prev)
3016 __releases(rq->lock)
3018 struct rq *rq = this_rq();
3020 finish_task_switch(rq, prev);
3023 * FIXME: do we need to worry about rq being invalidated by the
3028 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3029 /* In this case, finish_task_switch does not reenable preemption */
3032 if (current->set_child_tid)
3033 put_user(task_pid_vnr(current), current->set_child_tid);
3037 * context_switch - switch to the new MM and the new
3038 * thread's register state.
3041 context_switch(struct rq *rq, struct task_struct *prev,
3042 struct task_struct *next)
3044 struct mm_struct *mm, *oldmm;
3046 prepare_task_switch(rq, prev, next);
3047 trace_sched_switch(prev, next);
3049 oldmm = prev->active_mm;
3051 * For paravirt, this is coupled with an exit in switch_to to
3052 * combine the page table reload and the switch backend into
3055 arch_start_context_switch(prev);
3058 next->active_mm = oldmm;
3059 atomic_inc(&oldmm->mm_count);
3060 enter_lazy_tlb(oldmm, next);
3062 switch_mm(oldmm, mm, next);
3065 prev->active_mm = NULL;
3066 rq->prev_mm = oldmm;
3069 * Since the runqueue lock will be released by the next
3070 * task (which is an invalid locking op but in the case
3071 * of the scheduler it's an obvious special-case), so we
3072 * do an early lockdep release here:
3074 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3075 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3078 /* Here we just switch the register state and the stack. */
3079 switch_to(prev, next, prev);
3083 * this_rq must be evaluated again because prev may have moved
3084 * CPUs since it called schedule(), thus the 'rq' on its stack
3085 * frame will be invalid.
3087 finish_task_switch(this_rq(), prev);
3091 * nr_running, nr_uninterruptible and nr_context_switches:
3093 * externally visible scheduler statistics: current number of runnable
3094 * threads, current number of uninterruptible-sleeping threads, total
3095 * number of context switches performed since bootup.
3097 unsigned long nr_running(void)
3099 unsigned long i, sum = 0;
3101 for_each_online_cpu(i)
3102 sum += cpu_rq(i)->nr_running;
3107 unsigned long nr_uninterruptible(void)
3109 unsigned long i, sum = 0;
3111 for_each_possible_cpu(i)
3112 sum += cpu_rq(i)->nr_uninterruptible;
3115 * Since we read the counters lockless, it might be slightly
3116 * inaccurate. Do not allow it to go below zero though:
3118 if (unlikely((long)sum < 0))
3124 unsigned long long nr_context_switches(void)
3127 unsigned long long sum = 0;
3129 for_each_possible_cpu(i)
3130 sum += cpu_rq(i)->nr_switches;
3135 unsigned long nr_iowait(void)
3137 unsigned long i, sum = 0;
3139 for_each_possible_cpu(i)
3140 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3145 unsigned long nr_iowait_cpu(int cpu)
3147 struct rq *this = cpu_rq(cpu);
3148 return atomic_read(&this->nr_iowait);
3151 unsigned long this_cpu_load(void)
3153 struct rq *this = this_rq();
3154 return this->cpu_load[0];
3158 /* Variables and functions for calc_load */
3159 static atomic_long_t calc_load_tasks;
3160 static unsigned long calc_load_update;
3161 unsigned long avenrun[3];
3162 EXPORT_SYMBOL(avenrun);
3164 static long calc_load_fold_active(struct rq *this_rq)
3166 long nr_active, delta = 0;
3168 nr_active = this_rq->nr_running;
3169 nr_active += (long) this_rq->nr_uninterruptible;
3171 if (nr_active != this_rq->calc_load_active) {
3172 delta = nr_active - this_rq->calc_load_active;
3173 this_rq->calc_load_active = nr_active;
3179 static unsigned long
3180 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3183 load += active * (FIXED_1 - exp);
3184 load += 1UL << (FSHIFT - 1);
3185 return load >> FSHIFT;
3190 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3192 * When making the ILB scale, we should try to pull this in as well.
3194 static atomic_long_t calc_load_tasks_idle;
3196 static void calc_load_account_idle(struct rq *this_rq)
3200 delta = calc_load_fold_active(this_rq);
3202 atomic_long_add(delta, &calc_load_tasks_idle);
3205 static long calc_load_fold_idle(void)
3210 * Its got a race, we don't care...
3212 if (atomic_long_read(&calc_load_tasks_idle))
3213 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3219 * fixed_power_int - compute: x^n, in O(log n) time
3221 * @x: base of the power
3222 * @frac_bits: fractional bits of @x
3223 * @n: power to raise @x to.
3225 * By exploiting the relation between the definition of the natural power
3226 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3227 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3228 * (where: n_i \elem {0, 1}, the binary vector representing n),
3229 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3230 * of course trivially computable in O(log_2 n), the length of our binary
3233 static unsigned long
3234 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3236 unsigned long result = 1UL << frac_bits;
3241 result += 1UL << (frac_bits - 1);
3242 result >>= frac_bits;
3248 x += 1UL << (frac_bits - 1);
3256 * a1 = a0 * e + a * (1 - e)
3258 * a2 = a1 * e + a * (1 - e)
3259 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3260 * = a0 * e^2 + a * (1 - e) * (1 + e)
3262 * a3 = a2 * e + a * (1 - e)
3263 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3264 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3268 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3269 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3270 * = a0 * e^n + a * (1 - e^n)
3272 * [1] application of the geometric series:
3275 * S_n := \Sum x^i = -------------
3278 static unsigned long
3279 calc_load_n(unsigned long load, unsigned long exp,
3280 unsigned long active, unsigned int n)
3283 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3287 * NO_HZ can leave us missing all per-cpu ticks calling
3288 * calc_load_account_active(), but since an idle CPU folds its delta into
3289 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3290 * in the pending idle delta if our idle period crossed a load cycle boundary.
3292 * Once we've updated the global active value, we need to apply the exponential
3293 * weights adjusted to the number of cycles missed.
3295 static void calc_global_nohz(unsigned long ticks)
3297 long delta, active, n;
3299 if (time_before(jiffies, calc_load_update))
3303 * If we crossed a calc_load_update boundary, make sure to fold
3304 * any pending idle changes, the respective CPUs might have
3305 * missed the tick driven calc_load_account_active() update
3308 delta = calc_load_fold_idle();
3310 atomic_long_add(delta, &calc_load_tasks);
3313 * If we were idle for multiple load cycles, apply them.
3315 if (ticks >= LOAD_FREQ) {
3316 n = ticks / LOAD_FREQ;
3318 active = atomic_long_read(&calc_load_tasks);
3319 active = active > 0 ? active * FIXED_1 : 0;
3321 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3322 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3323 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3325 calc_load_update += n * LOAD_FREQ;
3329 * Its possible the remainder of the above division also crosses
3330 * a LOAD_FREQ period, the regular check in calc_global_load()
3331 * which comes after this will take care of that.
3333 * Consider us being 11 ticks before a cycle completion, and us
3334 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3335 * age us 4 cycles, and the test in calc_global_load() will
3336 * pick up the final one.
3340 static void calc_load_account_idle(struct rq *this_rq)
3344 static inline long calc_load_fold_idle(void)
3349 static void calc_global_nohz(unsigned long ticks)
3355 * get_avenrun - get the load average array
3356 * @loads: pointer to dest load array
3357 * @offset: offset to add
3358 * @shift: shift count to shift the result left
3360 * These values are estimates at best, so no need for locking.
3362 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3364 loads[0] = (avenrun[0] + offset) << shift;
3365 loads[1] = (avenrun[1] + offset) << shift;
3366 loads[2] = (avenrun[2] + offset) << shift;
3370 * calc_load - update the avenrun load estimates 10 ticks after the
3371 * CPUs have updated calc_load_tasks.
3373 void calc_global_load(unsigned long ticks)
3377 calc_global_nohz(ticks);
3379 if (time_before(jiffies, calc_load_update + 10))
3382 active = atomic_long_read(&calc_load_tasks);
3383 active = active > 0 ? active * FIXED_1 : 0;
3385 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3386 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3387 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3389 calc_load_update += LOAD_FREQ;
3393 * Called from update_cpu_load() to periodically update this CPU's
3396 static void calc_load_account_active(struct rq *this_rq)
3400 if (time_before(jiffies, this_rq->calc_load_update))
3403 delta = calc_load_fold_active(this_rq);
3404 delta += calc_load_fold_idle();
3406 atomic_long_add(delta, &calc_load_tasks);
3408 this_rq->calc_load_update += LOAD_FREQ;
3412 * The exact cpuload at various idx values, calculated at every tick would be
3413 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3415 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3416 * on nth tick when cpu may be busy, then we have:
3417 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3418 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3420 * decay_load_missed() below does efficient calculation of
3421 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3422 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3424 * The calculation is approximated on a 128 point scale.
3425 * degrade_zero_ticks is the number of ticks after which load at any
3426 * particular idx is approximated to be zero.
3427 * degrade_factor is a precomputed table, a row for each load idx.
3428 * Each column corresponds to degradation factor for a power of two ticks,
3429 * based on 128 point scale.
3431 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3432 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3434 * With this power of 2 load factors, we can degrade the load n times
3435 * by looking at 1 bits in n and doing as many mult/shift instead of
3436 * n mult/shifts needed by the exact degradation.
3438 #define DEGRADE_SHIFT 7
3439 static const unsigned char
3440 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3441 static const unsigned char
3442 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3443 {0, 0, 0, 0, 0, 0, 0, 0},
3444 {64, 32, 8, 0, 0, 0, 0, 0},
3445 {96, 72, 40, 12, 1, 0, 0},
3446 {112, 98, 75, 43, 15, 1, 0},
3447 {120, 112, 98, 76, 45, 16, 2} };
3450 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3451 * would be when CPU is idle and so we just decay the old load without
3452 * adding any new load.
3454 static unsigned long
3455 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3459 if (!missed_updates)
3462 if (missed_updates >= degrade_zero_ticks[idx])
3466 return load >> missed_updates;
3468 while (missed_updates) {
3469 if (missed_updates % 2)
3470 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3472 missed_updates >>= 1;
3479 * Update rq->cpu_load[] statistics. This function is usually called every
3480 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3481 * every tick. We fix it up based on jiffies.
3483 static void update_cpu_load(struct rq *this_rq)
3485 unsigned long this_load = this_rq->load.weight;
3486 unsigned long curr_jiffies = jiffies;
3487 unsigned long pending_updates;
3490 this_rq->nr_load_updates++;
3492 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3493 if (curr_jiffies == this_rq->last_load_update_tick)
3496 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3497 this_rq->last_load_update_tick = curr_jiffies;
3499 /* Update our load: */
3500 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3501 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3502 unsigned long old_load, new_load;
3504 /* scale is effectively 1 << i now, and >> i divides by scale */
3506 old_load = this_rq->cpu_load[i];
3507 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3508 new_load = this_load;
3510 * Round up the averaging division if load is increasing. This
3511 * prevents us from getting stuck on 9 if the load is 10, for
3514 if (new_load > old_load)
3515 new_load += scale - 1;
3517 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3520 sched_avg_update(this_rq);
3523 static void update_cpu_load_active(struct rq *this_rq)
3525 update_cpu_load(this_rq);
3527 calc_load_account_active(this_rq);
3533 * sched_exec - execve() is a valuable balancing opportunity, because at
3534 * this point the task has the smallest effective memory and cache footprint.
3536 void sched_exec(void)
3538 struct task_struct *p = current;
3539 unsigned long flags;
3543 rq = task_rq_lock(p, &flags);
3544 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3545 if (dest_cpu == smp_processor_id())
3549 * select_task_rq() can race against ->cpus_allowed
3551 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3552 likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
3553 struct migration_arg arg = { p, dest_cpu };
3555 task_rq_unlock(rq, &flags);
3556 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3560 task_rq_unlock(rq, &flags);
3565 DEFINE_PER_CPU(struct kernel_stat, kstat);
3567 EXPORT_PER_CPU_SYMBOL(kstat);
3570 * Return any ns on the sched_clock that have not yet been accounted in
3571 * @p in case that task is currently running.
3573 * Called with task_rq_lock() held on @rq.
3575 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3579 if (task_current(rq, p)) {
3580 update_rq_clock(rq);
3581 ns = rq->clock_task - p->se.exec_start;
3589 unsigned long long task_delta_exec(struct task_struct *p)
3591 unsigned long flags;
3595 rq = task_rq_lock(p, &flags);
3596 ns = do_task_delta_exec(p, rq);
3597 task_rq_unlock(rq, &flags);
3603 * Return accounted runtime for the task.
3604 * In case the task is currently running, return the runtime plus current's
3605 * pending runtime that have not been accounted yet.
3607 unsigned long long task_sched_runtime(struct task_struct *p)
3609 unsigned long flags;
3613 rq = task_rq_lock(p, &flags);
3614 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3615 task_rq_unlock(rq, &flags);
3621 * Return sum_exec_runtime for the thread group.
3622 * In case the task is currently running, return the sum plus current's
3623 * pending runtime that have not been accounted yet.
3625 * Note that the thread group might have other running tasks as well,
3626 * so the return value not includes other pending runtime that other
3627 * running tasks might have.
3629 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3631 struct task_cputime totals;
3632 unsigned long flags;
3636 rq = task_rq_lock(p, &flags);
3637 thread_group_cputime(p, &totals);
3638 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3639 task_rq_unlock(rq, &flags);
3645 * Account user cpu time to a process.
3646 * @p: the process that the cpu time gets accounted to
3647 * @cputime: the cpu time spent in user space since the last update
3648 * @cputime_scaled: cputime scaled by cpu frequency
3650 void account_user_time(struct task_struct *p, cputime_t cputime,
3651 cputime_t cputime_scaled)
3653 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3656 /* Add user time to process. */
3657 p->utime = cputime_add(p->utime, cputime);
3658 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3659 account_group_user_time(p, cputime);
3661 /* Add user time to cpustat. */
3662 tmp = cputime_to_cputime64(cputime);
3663 if (TASK_NICE(p) > 0)
3664 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3666 cpustat->user = cputime64_add(cpustat->user, tmp);
3668 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3669 /* Account for user time used */
3670 acct_update_integrals(p);
3674 * Account guest cpu time to a process.
3675 * @p: the process that the cpu time gets accounted to
3676 * @cputime: the cpu time spent in virtual machine since the last update
3677 * @cputime_scaled: cputime scaled by cpu frequency
3679 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3680 cputime_t cputime_scaled)
3683 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3685 tmp = cputime_to_cputime64(cputime);
3687 /* Add guest time to process. */
3688 p->utime = cputime_add(p->utime, cputime);
3689 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3690 account_group_user_time(p, cputime);
3691 p->gtime = cputime_add(p->gtime, cputime);
3693 /* Add guest time to cpustat. */
3694 if (TASK_NICE(p) > 0) {
3695 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3696 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3698 cpustat->user = cputime64_add(cpustat->user, tmp);
3699 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3704 * Account system cpu time to a process.
3705 * @p: the process that the cpu time gets accounted to
3706 * @hardirq_offset: the offset to subtract from hardirq_count()
3707 * @cputime: the cpu time spent in kernel space since the last update
3708 * @cputime_scaled: cputime scaled by cpu frequency
3710 void account_system_time(struct task_struct *p, int hardirq_offset,
3711 cputime_t cputime, cputime_t cputime_scaled)
3713 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3716 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3717 account_guest_time(p, cputime, cputime_scaled);
3721 /* Add system time to process. */
3722 p->stime = cputime_add(p->stime, cputime);
3723 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3724 account_group_system_time(p, cputime);
3726 /* Add system time to cpustat. */
3727 tmp = cputime_to_cputime64(cputime);
3728 if (hardirq_count() - hardirq_offset)
3729 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3730 else if (in_serving_softirq())
3731 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3733 cpustat->system = cputime64_add(cpustat->system, tmp);
3735 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3737 /* Account for system time used */
3738 acct_update_integrals(p);
3742 * Account for involuntary wait time.
3743 * @steal: the cpu time spent in involuntary wait
3745 void account_steal_time(cputime_t cputime)
3747 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3748 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3750 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3754 * Account for idle time.
3755 * @cputime: the cpu time spent in idle wait
3757 void account_idle_time(cputime_t cputime)
3759 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3760 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3761 struct rq *rq = this_rq();
3763 if (atomic_read(&rq->nr_iowait) > 0)
3764 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3766 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3769 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3772 * Account a single tick of cpu time.
3773 * @p: the process that the cpu time gets accounted to
3774 * @user_tick: indicates if the tick is a user or a system tick
3776 void account_process_tick(struct task_struct *p, int user_tick)
3778 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3779 struct rq *rq = this_rq();
3782 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3783 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3784 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3787 account_idle_time(cputime_one_jiffy);
3791 * Account multiple ticks of steal time.
3792 * @p: the process from which the cpu time has been stolen
3793 * @ticks: number of stolen ticks
3795 void account_steal_ticks(unsigned long ticks)
3797 account_steal_time(jiffies_to_cputime(ticks));
3801 * Account multiple ticks of idle time.
3802 * @ticks: number of stolen ticks
3804 void account_idle_ticks(unsigned long ticks)
3806 account_idle_time(jiffies_to_cputime(ticks));
3812 * Use precise platform statistics if available:
3814 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3815 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3821 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3823 struct task_cputime cputime;
3825 thread_group_cputime(p, &cputime);
3827 *ut = cputime.utime;
3828 *st = cputime.stime;
3832 #ifndef nsecs_to_cputime
3833 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3836 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3838 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3841 * Use CFS's precise accounting:
3843 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3849 do_div(temp, total);
3850 utime = (cputime_t)temp;
3855 * Compare with previous values, to keep monotonicity:
3857 p->prev_utime = max(p->prev_utime, utime);
3858 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3860 *ut = p->prev_utime;
3861 *st = p->prev_stime;
3865 * Must be called with siglock held.
3867 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3869 struct signal_struct *sig = p->signal;
3870 struct task_cputime cputime;
3871 cputime_t rtime, utime, total;
3873 thread_group_cputime(p, &cputime);
3875 total = cputime_add(cputime.utime, cputime.stime);
3876 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3881 temp *= cputime.utime;
3882 do_div(temp, total);
3883 utime = (cputime_t)temp;
3887 sig->prev_utime = max(sig->prev_utime, utime);
3888 sig->prev_stime = max(sig->prev_stime,
3889 cputime_sub(rtime, sig->prev_utime));
3891 *ut = sig->prev_utime;
3892 *st = sig->prev_stime;
3897 * This function gets called by the timer code, with HZ frequency.
3898 * We call it with interrupts disabled.
3900 * It also gets called by the fork code, when changing the parent's
3903 void scheduler_tick(void)
3905 int cpu = smp_processor_id();
3906 struct rq *rq = cpu_rq(cpu);
3907 struct task_struct *curr = rq->curr;
3911 raw_spin_lock(&rq->lock);
3912 update_rq_clock(rq);
3913 update_cpu_load_active(rq);
3914 curr->sched_class->task_tick(rq, curr, 0);
3915 raw_spin_unlock(&rq->lock);
3917 perf_event_task_tick();
3920 rq->idle_at_tick = idle_cpu(cpu);
3921 trigger_load_balance(rq, cpu);
3925 notrace unsigned long get_parent_ip(unsigned long addr)
3927 if (in_lock_functions(addr)) {
3928 addr = CALLER_ADDR2;
3929 if (in_lock_functions(addr))
3930 addr = CALLER_ADDR3;
3935 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3936 defined(CONFIG_PREEMPT_TRACER))
3938 void __kprobes add_preempt_count(int val)
3940 #ifdef CONFIG_DEBUG_PREEMPT
3944 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3947 preempt_count() += val;
3948 #ifdef CONFIG_DEBUG_PREEMPT
3950 * Spinlock count overflowing soon?
3952 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3955 if (preempt_count() == val)
3956 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3958 EXPORT_SYMBOL(add_preempt_count);
3960 void __kprobes sub_preempt_count(int val)
3962 #ifdef CONFIG_DEBUG_PREEMPT
3966 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3969 * Is the spinlock portion underflowing?
3971 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3972 !(preempt_count() & PREEMPT_MASK)))
3976 if (preempt_count() == val)
3977 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3978 preempt_count() -= val;
3980 EXPORT_SYMBOL(sub_preempt_count);
3985 * Print scheduling while atomic bug:
3987 static noinline void __schedule_bug(struct task_struct *prev)
3989 struct pt_regs *regs = get_irq_regs();
3991 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3992 prev->comm, prev->pid, preempt_count());
3994 debug_show_held_locks(prev);
3996 if (irqs_disabled())
3997 print_irqtrace_events(prev);
4006 * Various schedule()-time debugging checks and statistics:
4008 static inline void schedule_debug(struct task_struct *prev)
4011 * Test if we are atomic. Since do_exit() needs to call into
4012 * schedule() atomically, we ignore that path for now.
4013 * Otherwise, whine if we are scheduling when we should not be.
4015 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4016 __schedule_bug(prev);
4018 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4020 schedstat_inc(this_rq(), sched_count);
4021 #ifdef CONFIG_SCHEDSTATS
4022 if (unlikely(prev->lock_depth >= 0)) {
4023 schedstat_inc(this_rq(), bkl_count);
4024 schedstat_inc(prev, sched_info.bkl_count);
4029 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4032 update_rq_clock(rq);
4033 prev->sched_class->put_prev_task(rq, prev);
4037 * Pick up the highest-prio task:
4039 static inline struct task_struct *
4040 pick_next_task(struct rq *rq)
4042 const struct sched_class *class;
4043 struct task_struct *p;
4046 * Optimization: we know that if all tasks are in
4047 * the fair class we can call that function directly:
4049 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4050 p = fair_sched_class.pick_next_task(rq);
4055 for_each_class(class) {
4056 p = class->pick_next_task(rq);
4061 BUG(); /* the idle class will always have a runnable task */
4065 * schedule() is the main scheduler function.
4067 asmlinkage void __sched schedule(void)
4069 struct task_struct *prev, *next;
4070 unsigned long *switch_count;
4076 cpu = smp_processor_id();
4078 rcu_note_context_switch(cpu);
4081 release_kernel_lock(prev);
4082 need_resched_nonpreemptible:
4084 schedule_debug(prev);
4086 if (sched_feat(HRTICK))
4089 raw_spin_lock_irq(&rq->lock);
4091 switch_count = &prev->nivcsw;
4092 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4093 if (unlikely(signal_pending_state(prev->state, prev))) {
4094 prev->state = TASK_RUNNING;
4097 * If a worker is going to sleep, notify and
4098 * ask workqueue whether it wants to wake up a
4099 * task to maintain concurrency. If so, wake
4102 if (prev->flags & PF_WQ_WORKER) {
4103 struct task_struct *to_wakeup;
4105 to_wakeup = wq_worker_sleeping(prev, cpu);
4107 try_to_wake_up_local(to_wakeup);
4109 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4111 switch_count = &prev->nvcsw;
4114 pre_schedule(rq, prev);
4116 if (unlikely(!rq->nr_running))
4117 idle_balance(cpu, rq);
4119 put_prev_task(rq, prev);
4120 next = pick_next_task(rq);
4121 clear_tsk_need_resched(prev);
4122 rq->skip_clock_update = 0;
4124 if (likely(prev != next)) {
4125 sched_info_switch(prev, next);
4126 perf_event_task_sched_out(prev, next);
4131 WARN_ON_ONCE(test_tsk_need_resched(next));
4133 context_switch(rq, prev, next); /* unlocks the rq */
4135 * The context switch have flipped the stack from under us
4136 * and restored the local variables which were saved when
4137 * this task called schedule() in the past. prev == current
4138 * is still correct, but it can be moved to another cpu/rq.
4140 cpu = smp_processor_id();
4143 raw_spin_unlock_irq(&rq->lock);
4147 if (unlikely(reacquire_kernel_lock(prev)))
4148 goto need_resched_nonpreemptible;
4150 preempt_enable_no_resched();
4154 EXPORT_SYMBOL(schedule);
4156 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4158 * Look out! "owner" is an entirely speculative pointer
4159 * access and not reliable.
4161 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
4166 if (!sched_feat(OWNER_SPIN))
4169 #ifdef CONFIG_DEBUG_PAGEALLOC
4171 * Need to access the cpu field knowing that
4172 * DEBUG_PAGEALLOC could have unmapped it if
4173 * the mutex owner just released it and exited.
4175 if (probe_kernel_address(&owner->cpu, cpu))
4182 * Even if the access succeeded (likely case),
4183 * the cpu field may no longer be valid.
4185 if (cpu >= nr_cpumask_bits)
4189 * We need to validate that we can do a
4190 * get_cpu() and that we have the percpu area.
4192 if (!cpu_online(cpu))
4199 * Owner changed, break to re-assess state.
4201 if (lock->owner != owner) {
4203 * If the lock has switched to a different owner,
4204 * we likely have heavy contention. Return 0 to quit
4205 * optimistic spinning and not contend further:
4213 * Is that owner really running on that cpu?
4215 if (task_thread_info(rq->curr) != owner || need_resched())
4225 #ifdef CONFIG_PREEMPT
4227 * this is the entry point to schedule() from in-kernel preemption
4228 * off of preempt_enable. Kernel preemptions off return from interrupt
4229 * occur there and call schedule directly.
4231 asmlinkage void __sched notrace preempt_schedule(void)
4233 struct thread_info *ti = current_thread_info();
4236 * If there is a non-zero preempt_count or interrupts are disabled,
4237 * we do not want to preempt the current task. Just return..
4239 if (likely(ti->preempt_count || irqs_disabled()))
4243 add_preempt_count_notrace(PREEMPT_ACTIVE);
4245 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4248 * Check again in case we missed a preemption opportunity
4249 * between schedule and now.
4252 } while (need_resched());
4254 EXPORT_SYMBOL(preempt_schedule);
4257 * this is the entry point to schedule() from kernel preemption
4258 * off of irq context.
4259 * Note, that this is called and return with irqs disabled. This will
4260 * protect us against recursive calling from irq.
4262 asmlinkage void __sched preempt_schedule_irq(void)
4264 struct thread_info *ti = current_thread_info();
4266 /* Catch callers which need to be fixed */
4267 BUG_ON(ti->preempt_count || !irqs_disabled());
4270 add_preempt_count(PREEMPT_ACTIVE);
4273 local_irq_disable();
4274 sub_preempt_count(PREEMPT_ACTIVE);
4277 * Check again in case we missed a preemption opportunity
4278 * between schedule and now.
4281 } while (need_resched());
4284 #endif /* CONFIG_PREEMPT */
4286 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4289 return try_to_wake_up(curr->private, mode, wake_flags);
4291 EXPORT_SYMBOL(default_wake_function);
4294 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4295 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4296 * number) then we wake all the non-exclusive tasks and one exclusive task.
4298 * There are circumstances in which we can try to wake a task which has already
4299 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4300 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4302 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4303 int nr_exclusive, int wake_flags, void *key)
4305 wait_queue_t *curr, *next;
4307 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4308 unsigned flags = curr->flags;
4310 if (curr->func(curr, mode, wake_flags, key) &&
4311 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4317 * __wake_up - wake up threads blocked on a waitqueue.
4319 * @mode: which threads
4320 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4321 * @key: is directly passed to the wakeup function
4323 * It may be assumed that this function implies a write memory barrier before
4324 * changing the task state if and only if any tasks are woken up.
4326 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4327 int nr_exclusive, void *key)
4329 unsigned long flags;
4331 spin_lock_irqsave(&q->lock, flags);
4332 __wake_up_common(q, mode, nr_exclusive, 0, key);
4333 spin_unlock_irqrestore(&q->lock, flags);
4335 EXPORT_SYMBOL(__wake_up);
4338 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4340 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4342 __wake_up_common(q, mode, 1, 0, NULL);
4344 EXPORT_SYMBOL_GPL(__wake_up_locked);
4346 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4348 __wake_up_common(q, mode, 1, 0, key);
4352 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4354 * @mode: which threads
4355 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4356 * @key: opaque value to be passed to wakeup targets
4358 * The sync wakeup differs that the waker knows that it will schedule
4359 * away soon, so while the target thread will be woken up, it will not
4360 * be migrated to another CPU - ie. the two threads are 'synchronized'
4361 * with each other. This can prevent needless bouncing between CPUs.
4363 * On UP it can prevent extra preemption.
4365 * It may be assumed that this function implies a write memory barrier before
4366 * changing the task state if and only if any tasks are woken up.
4368 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4369 int nr_exclusive, void *key)
4371 unsigned long flags;
4372 int wake_flags = WF_SYNC;
4377 if (unlikely(!nr_exclusive))
4380 spin_lock_irqsave(&q->lock, flags);
4381 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4382 spin_unlock_irqrestore(&q->lock, flags);
4384 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4387 * __wake_up_sync - see __wake_up_sync_key()
4389 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4391 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4393 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4396 * complete: - signals a single thread waiting on this completion
4397 * @x: holds the state of this particular completion
4399 * This will wake up a single thread waiting on this completion. Threads will be
4400 * awakened in the same order in which they were queued.
4402 * See also complete_all(), wait_for_completion() and related routines.
4404 * It may be assumed that this function implies a write memory barrier before
4405 * changing the task state if and only if any tasks are woken up.
4407 void complete(struct completion *x)
4409 unsigned long flags;
4411 spin_lock_irqsave(&x->wait.lock, flags);
4413 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4414 spin_unlock_irqrestore(&x->wait.lock, flags);
4416 EXPORT_SYMBOL(complete);
4419 * complete_all: - signals all threads waiting on this completion
4420 * @x: holds the state of this particular completion
4422 * This will wake up all threads waiting on this particular completion event.
4424 * It may be assumed that this function implies a write memory barrier before
4425 * changing the task state if and only if any tasks are woken up.
4427 void complete_all(struct completion *x)
4429 unsigned long flags;
4431 spin_lock_irqsave(&x->wait.lock, flags);
4432 x->done += UINT_MAX/2;
4433 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4434 spin_unlock_irqrestore(&x->wait.lock, flags);
4436 EXPORT_SYMBOL(complete_all);
4438 static inline long __sched
4439 do_wait_for_common(struct completion *x, long timeout, int state)
4442 DECLARE_WAITQUEUE(wait, current);
4444 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4446 if (signal_pending_state(state, current)) {
4447 timeout = -ERESTARTSYS;
4450 __set_current_state(state);
4451 spin_unlock_irq(&x->wait.lock);
4452 timeout = schedule_timeout(timeout);
4453 spin_lock_irq(&x->wait.lock);
4454 } while (!x->done && timeout);
4455 __remove_wait_queue(&x->wait, &wait);
4460 return timeout ?: 1;
4464 wait_for_common(struct completion *x, long timeout, int state)
4468 spin_lock_irq(&x->wait.lock);
4469 timeout = do_wait_for_common(x, timeout, state);
4470 spin_unlock_irq(&x->wait.lock);
4475 * wait_for_completion: - waits for completion of a task
4476 * @x: holds the state of this particular completion
4478 * This waits to be signaled for completion of a specific task. It is NOT
4479 * interruptible and there is no timeout.
4481 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4482 * and interrupt capability. Also see complete().
4484 void __sched wait_for_completion(struct completion *x)
4486 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4488 EXPORT_SYMBOL(wait_for_completion);
4491 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4492 * @x: holds the state of this particular completion
4493 * @timeout: timeout value in jiffies
4495 * This waits for either a completion of a specific task to be signaled or for a
4496 * specified timeout to expire. The timeout is in jiffies. It is not
4499 unsigned long __sched
4500 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4502 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4504 EXPORT_SYMBOL(wait_for_completion_timeout);
4507 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4508 * @x: holds the state of this particular completion
4510 * This waits for completion of a specific task to be signaled. It is
4513 int __sched wait_for_completion_interruptible(struct completion *x)
4515 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4516 if (t == -ERESTARTSYS)
4520 EXPORT_SYMBOL(wait_for_completion_interruptible);
4523 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4524 * @x: holds the state of this particular completion
4525 * @timeout: timeout value in jiffies
4527 * This waits for either a completion of a specific task to be signaled or for a
4528 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4530 unsigned long __sched
4531 wait_for_completion_interruptible_timeout(struct completion *x,
4532 unsigned long timeout)
4534 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4536 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4539 * wait_for_completion_killable: - waits for completion of a task (killable)
4540 * @x: holds the state of this particular completion
4542 * This waits to be signaled for completion of a specific task. It can be
4543 * interrupted by a kill signal.
4545 int __sched wait_for_completion_killable(struct completion *x)
4547 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4548 if (t == -ERESTARTSYS)
4552 EXPORT_SYMBOL(wait_for_completion_killable);
4555 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4556 * @x: holds the state of this particular completion
4557 * @timeout: timeout value in jiffies
4559 * This waits for either a completion of a specific task to be
4560 * signaled or for a specified timeout to expire. It can be
4561 * interrupted by a kill signal. The timeout is in jiffies.
4563 unsigned long __sched
4564 wait_for_completion_killable_timeout(struct completion *x,
4565 unsigned long timeout)
4567 return wait_for_common(x, timeout, TASK_KILLABLE);
4569 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4572 * try_wait_for_completion - try to decrement a completion without blocking
4573 * @x: completion structure
4575 * Returns: 0 if a decrement cannot be done without blocking
4576 * 1 if a decrement succeeded.
4578 * If a completion is being used as a counting completion,
4579 * attempt to decrement the counter without blocking. This
4580 * enables us to avoid waiting if the resource the completion
4581 * is protecting is not available.
4583 bool try_wait_for_completion(struct completion *x)
4585 unsigned long flags;
4588 spin_lock_irqsave(&x->wait.lock, flags);
4593 spin_unlock_irqrestore(&x->wait.lock, flags);
4596 EXPORT_SYMBOL(try_wait_for_completion);
4599 * completion_done - Test to see if a completion has any waiters
4600 * @x: completion structure
4602 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4603 * 1 if there are no waiters.
4606 bool completion_done(struct completion *x)
4608 unsigned long flags;
4611 spin_lock_irqsave(&x->wait.lock, flags);
4614 spin_unlock_irqrestore(&x->wait.lock, flags);
4617 EXPORT_SYMBOL(completion_done);
4620 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4622 unsigned long flags;
4625 init_waitqueue_entry(&wait, current);
4627 __set_current_state(state);
4629 spin_lock_irqsave(&q->lock, flags);
4630 __add_wait_queue(q, &wait);
4631 spin_unlock(&q->lock);
4632 timeout = schedule_timeout(timeout);
4633 spin_lock_irq(&q->lock);
4634 __remove_wait_queue(q, &wait);
4635 spin_unlock_irqrestore(&q->lock, flags);
4640 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4642 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4644 EXPORT_SYMBOL(interruptible_sleep_on);
4647 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4649 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4651 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4653 void __sched sleep_on(wait_queue_head_t *q)
4655 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4657 EXPORT_SYMBOL(sleep_on);
4659 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4661 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4663 EXPORT_SYMBOL(sleep_on_timeout);
4665 #ifdef CONFIG_RT_MUTEXES
4668 * rt_mutex_setprio - set the current priority of a task
4670 * @prio: prio value (kernel-internal form)
4672 * This function changes the 'effective' priority of a task. It does
4673 * not touch ->normal_prio like __setscheduler().
4675 * Used by the rt_mutex code to implement priority inheritance logic.
4677 void rt_mutex_setprio(struct task_struct *p, int prio)
4679 unsigned long flags;
4680 int oldprio, on_rq, running;
4682 const struct sched_class *prev_class;
4684 BUG_ON(prio < 0 || prio > MAX_PRIO);
4686 rq = task_rq_lock(p, &flags);
4688 trace_sched_pi_setprio(p, prio);
4690 prev_class = p->sched_class;
4691 on_rq = p->se.on_rq;
4692 running = task_current(rq, p);
4694 dequeue_task(rq, p, 0);
4696 p->sched_class->put_prev_task(rq, p);
4699 p->sched_class = &rt_sched_class;
4701 p->sched_class = &fair_sched_class;
4706 p->sched_class->set_curr_task(rq);
4708 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4710 check_class_changed(rq, p, prev_class, oldprio, running);
4712 task_rq_unlock(rq, &flags);
4717 void set_user_nice(struct task_struct *p, long nice)
4719 int old_prio, delta, on_rq;
4720 unsigned long flags;
4723 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4726 * We have to be careful, if called from sys_setpriority(),
4727 * the task might be in the middle of scheduling on another CPU.
4729 rq = task_rq_lock(p, &flags);
4731 * The RT priorities are set via sched_setscheduler(), but we still
4732 * allow the 'normal' nice value to be set - but as expected
4733 * it wont have any effect on scheduling until the task is
4734 * SCHED_FIFO/SCHED_RR:
4736 if (task_has_rt_policy(p)) {
4737 p->static_prio = NICE_TO_PRIO(nice);
4740 on_rq = p->se.on_rq;
4742 dequeue_task(rq, p, 0);
4744 p->static_prio = NICE_TO_PRIO(nice);
4747 p->prio = effective_prio(p);
4748 delta = p->prio - old_prio;
4751 enqueue_task(rq, p, 0);
4753 * If the task increased its priority or is running and
4754 * lowered its priority, then reschedule its CPU:
4756 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4757 resched_task(rq->curr);
4760 task_rq_unlock(rq, &flags);
4762 EXPORT_SYMBOL(set_user_nice);
4765 * can_nice - check if a task can reduce its nice value
4769 int can_nice(const struct task_struct *p, const int nice)
4771 /* convert nice value [19,-20] to rlimit style value [1,40] */
4772 int nice_rlim = 20 - nice;
4774 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4775 capable(CAP_SYS_NICE));
4778 #ifdef __ARCH_WANT_SYS_NICE
4781 * sys_nice - change the priority of the current process.
4782 * @increment: priority increment
4784 * sys_setpriority is a more generic, but much slower function that
4785 * does similar things.
4787 SYSCALL_DEFINE1(nice, int, increment)
4792 * Setpriority might change our priority at the same moment.
4793 * We don't have to worry. Conceptually one call occurs first
4794 * and we have a single winner.
4796 if (increment < -40)
4801 nice = TASK_NICE(current) + increment;
4807 if (increment < 0 && !can_nice(current, nice))
4810 retval = security_task_setnice(current, nice);
4814 set_user_nice(current, nice);
4821 * task_prio - return the priority value of a given task.
4822 * @p: the task in question.
4824 * This is the priority value as seen by users in /proc.
4825 * RT tasks are offset by -200. Normal tasks are centered
4826 * around 0, value goes from -16 to +15.
4828 int task_prio(const struct task_struct *p)
4830 return p->prio - MAX_RT_PRIO;
4834 * task_nice - return the nice value of a given task.
4835 * @p: the task in question.
4837 int task_nice(const struct task_struct *p)
4839 return TASK_NICE(p);
4841 EXPORT_SYMBOL(task_nice);
4844 * idle_cpu - is a given cpu idle currently?
4845 * @cpu: the processor in question.
4847 int idle_cpu(int cpu)
4849 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4853 * idle_task - return the idle task for a given cpu.
4854 * @cpu: the processor in question.
4856 struct task_struct *idle_task(int cpu)
4858 return cpu_rq(cpu)->idle;
4862 * find_process_by_pid - find a process with a matching PID value.
4863 * @pid: the pid in question.
4865 static struct task_struct *find_process_by_pid(pid_t pid)
4867 return pid ? find_task_by_vpid(pid) : current;
4870 /* Actually do priority change: must hold rq lock. */
4872 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4874 BUG_ON(p->se.on_rq);
4877 p->rt_priority = prio;
4878 p->normal_prio = normal_prio(p);
4879 /* we are holding p->pi_lock already */
4880 p->prio = rt_mutex_getprio(p);
4881 if (rt_prio(p->prio))
4882 p->sched_class = &rt_sched_class;
4884 p->sched_class = &fair_sched_class;
4889 * check the target process has a UID that matches the current process's
4891 static bool check_same_owner(struct task_struct *p)
4893 const struct cred *cred = current_cred(), *pcred;
4897 pcred = __task_cred(p);
4898 match = (cred->euid == pcred->euid ||
4899 cred->euid == pcred->uid);
4904 static int __sched_setscheduler(struct task_struct *p, int policy,
4905 struct sched_param *param, bool user)
4907 int retval, oldprio, oldpolicy = -1, on_rq, running;
4908 unsigned long flags;
4909 const struct sched_class *prev_class;
4913 /* may grab non-irq protected spin_locks */
4914 BUG_ON(in_interrupt());
4916 /* double check policy once rq lock held */
4918 reset_on_fork = p->sched_reset_on_fork;
4919 policy = oldpolicy = p->policy;
4921 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4922 policy &= ~SCHED_RESET_ON_FORK;
4924 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4925 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4926 policy != SCHED_IDLE)
4931 * Valid priorities for SCHED_FIFO and SCHED_RR are
4932 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4933 * SCHED_BATCH and SCHED_IDLE is 0.
4935 if (param->sched_priority < 0 ||
4936 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4937 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4939 if (rt_policy(policy) != (param->sched_priority != 0))
4943 * Allow unprivileged RT tasks to decrease priority:
4945 if (user && !capable(CAP_SYS_NICE)) {
4946 if (rt_policy(policy)) {
4947 unsigned long rlim_rtprio =
4948 task_rlimit(p, RLIMIT_RTPRIO);
4950 /* can't set/change the rt policy */
4951 if (policy != p->policy && !rlim_rtprio)
4954 /* can't increase priority */
4955 if (param->sched_priority > p->rt_priority &&
4956 param->sched_priority > rlim_rtprio)
4960 * Like positive nice levels, dont allow tasks to
4961 * move out of SCHED_IDLE either:
4963 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4966 /* can't change other user's priorities */
4967 if (!check_same_owner(p))
4970 /* Normal users shall not reset the sched_reset_on_fork flag */
4971 if (p->sched_reset_on_fork && !reset_on_fork)
4976 retval = security_task_setscheduler(p);
4982 * make sure no PI-waiters arrive (or leave) while we are
4983 * changing the priority of the task:
4985 raw_spin_lock_irqsave(&p->pi_lock, flags);
4987 * To be able to change p->policy safely, the apropriate
4988 * runqueue lock must be held.
4990 rq = __task_rq_lock(p);
4993 * Changing the policy of the stop threads its a very bad idea
4995 if (p == rq->stop) {
4996 __task_rq_unlock(rq);
4997 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5001 #ifdef CONFIG_RT_GROUP_SCHED
5004 * Do not allow realtime tasks into groups that have no runtime
5007 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5008 task_group(p)->rt_bandwidth.rt_runtime == 0) {
5009 __task_rq_unlock(rq);
5010 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5016 /* recheck policy now with rq lock held */
5017 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5018 policy = oldpolicy = -1;
5019 __task_rq_unlock(rq);
5020 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5023 on_rq = p->se.on_rq;
5024 running = task_current(rq, p);
5026 deactivate_task(rq, p, 0);
5028 p->sched_class->put_prev_task(rq, p);
5030 p->sched_reset_on_fork = reset_on_fork;
5033 prev_class = p->sched_class;
5034 __setscheduler(rq, p, policy, param->sched_priority);
5037 p->sched_class->set_curr_task(rq);
5039 activate_task(rq, p, 0);
5041 check_class_changed(rq, p, prev_class, oldprio, running);
5043 __task_rq_unlock(rq);
5044 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5046 rt_mutex_adjust_pi(p);
5052 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5053 * @p: the task in question.
5054 * @policy: new policy.
5055 * @param: structure containing the new RT priority.
5057 * NOTE that the task may be already dead.
5059 int sched_setscheduler(struct task_struct *p, int policy,
5060 struct sched_param *param)
5062 return __sched_setscheduler(p, policy, param, true);
5064 EXPORT_SYMBOL_GPL(sched_setscheduler);
5067 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5068 * @p: the task in question.
5069 * @policy: new policy.
5070 * @param: structure containing the new RT priority.
5072 * Just like sched_setscheduler, only don't bother checking if the
5073 * current context has permission. For example, this is needed in
5074 * stop_machine(): we create temporary high priority worker threads,
5075 * but our caller might not have that capability.
5077 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5078 struct sched_param *param)
5080 return __sched_setscheduler(p, policy, param, false);
5084 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5086 struct sched_param lparam;
5087 struct task_struct *p;
5090 if (!param || pid < 0)
5092 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5097 p = find_process_by_pid(pid);
5099 retval = sched_setscheduler(p, policy, &lparam);
5106 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5107 * @pid: the pid in question.
5108 * @policy: new policy.
5109 * @param: structure containing the new RT priority.
5111 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5112 struct sched_param __user *, param)
5114 /* negative values for policy are not valid */
5118 return do_sched_setscheduler(pid, policy, param);
5122 * sys_sched_setparam - set/change the RT priority of a thread
5123 * @pid: the pid in question.
5124 * @param: structure containing the new RT priority.
5126 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5128 return do_sched_setscheduler(pid, -1, param);
5132 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5133 * @pid: the pid in question.
5135 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5137 struct task_struct *p;
5145 p = find_process_by_pid(pid);
5147 retval = security_task_getscheduler(p);
5150 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5157 * sys_sched_getparam - get the RT priority of a thread
5158 * @pid: the pid in question.
5159 * @param: structure containing the RT priority.
5161 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5163 struct sched_param lp;
5164 struct task_struct *p;
5167 if (!param || pid < 0)
5171 p = find_process_by_pid(pid);
5176 retval = security_task_getscheduler(p);
5180 lp.sched_priority = p->rt_priority;
5184 * This one might sleep, we cannot do it with a spinlock held ...
5186 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5195 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5197 cpumask_var_t cpus_allowed, new_mask;
5198 struct task_struct *p;
5204 p = find_process_by_pid(pid);
5211 /* Prevent p going away */
5215 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5219 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5221 goto out_free_cpus_allowed;
5224 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5227 retval = security_task_setscheduler(p);
5231 cpuset_cpus_allowed(p, cpus_allowed);
5232 cpumask_and(new_mask, in_mask, cpus_allowed);
5234 retval = set_cpus_allowed_ptr(p, new_mask);
5237 cpuset_cpus_allowed(p, cpus_allowed);
5238 if (!cpumask_subset(new_mask, cpus_allowed)) {
5240 * We must have raced with a concurrent cpuset
5241 * update. Just reset the cpus_allowed to the
5242 * cpuset's cpus_allowed
5244 cpumask_copy(new_mask, cpus_allowed);
5249 free_cpumask_var(new_mask);
5250 out_free_cpus_allowed:
5251 free_cpumask_var(cpus_allowed);
5258 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5259 struct cpumask *new_mask)
5261 if (len < cpumask_size())
5262 cpumask_clear(new_mask);
5263 else if (len > cpumask_size())
5264 len = cpumask_size();
5266 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5270 * sys_sched_setaffinity - set the cpu affinity of a process
5271 * @pid: pid of the process
5272 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5273 * @user_mask_ptr: user-space pointer to the new cpu mask
5275 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5276 unsigned long __user *, user_mask_ptr)
5278 cpumask_var_t new_mask;
5281 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5284 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5286 retval = sched_setaffinity(pid, new_mask);
5287 free_cpumask_var(new_mask);
5291 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5293 struct task_struct *p;
5294 unsigned long flags;
5302 p = find_process_by_pid(pid);
5306 retval = security_task_getscheduler(p);
5310 rq = task_rq_lock(p, &flags);
5311 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5312 task_rq_unlock(rq, &flags);
5322 * sys_sched_getaffinity - get the cpu affinity of a process
5323 * @pid: pid of the process
5324 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5325 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5327 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5328 unsigned long __user *, user_mask_ptr)
5333 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5335 if (len & (sizeof(unsigned long)-1))
5338 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5341 ret = sched_getaffinity(pid, mask);
5343 size_t retlen = min_t(size_t, len, cpumask_size());
5345 if (copy_to_user(user_mask_ptr, mask, retlen))
5350 free_cpumask_var(mask);
5356 * sys_sched_yield - yield the current processor to other threads.
5358 * This function yields the current CPU to other tasks. If there are no
5359 * other threads running on this CPU then this function will return.
5361 SYSCALL_DEFINE0(sched_yield)
5363 struct rq *rq = this_rq_lock();
5365 schedstat_inc(rq, yld_count);
5366 current->sched_class->yield_task(rq);
5369 * Since we are going to call schedule() anyway, there's
5370 * no need to preempt or enable interrupts:
5372 __release(rq->lock);
5373 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5374 do_raw_spin_unlock(&rq->lock);
5375 preempt_enable_no_resched();
5382 static inline int should_resched(void)
5384 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5387 static void __cond_resched(void)
5389 add_preempt_count(PREEMPT_ACTIVE);
5391 sub_preempt_count(PREEMPT_ACTIVE);
5394 int __sched _cond_resched(void)
5396 if (should_resched()) {
5402 EXPORT_SYMBOL(_cond_resched);
5405 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5406 * call schedule, and on return reacquire the lock.
5408 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5409 * operations here to prevent schedule() from being called twice (once via
5410 * spin_unlock(), once by hand).
5412 int __cond_resched_lock(spinlock_t *lock)
5414 int resched = should_resched();
5417 lockdep_assert_held(lock);
5419 if (spin_needbreak(lock) || resched) {
5430 EXPORT_SYMBOL(__cond_resched_lock);
5432 int __sched __cond_resched_softirq(void)
5434 BUG_ON(!in_softirq());
5436 if (should_resched()) {
5444 EXPORT_SYMBOL(__cond_resched_softirq);
5447 * yield - yield the current processor to other threads.
5449 * This is a shortcut for kernel-space yielding - it marks the
5450 * thread runnable and calls sys_sched_yield().
5452 void __sched yield(void)
5454 set_current_state(TASK_RUNNING);
5457 EXPORT_SYMBOL(yield);
5460 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5461 * that process accounting knows that this is a task in IO wait state.
5463 void __sched io_schedule(void)
5465 struct rq *rq = raw_rq();
5467 delayacct_blkio_start();
5468 atomic_inc(&rq->nr_iowait);
5469 current->in_iowait = 1;
5471 current->in_iowait = 0;
5472 atomic_dec(&rq->nr_iowait);
5473 delayacct_blkio_end();
5475 EXPORT_SYMBOL(io_schedule);
5477 long __sched io_schedule_timeout(long timeout)
5479 struct rq *rq = raw_rq();
5482 delayacct_blkio_start();
5483 atomic_inc(&rq->nr_iowait);
5484 current->in_iowait = 1;
5485 ret = schedule_timeout(timeout);
5486 current->in_iowait = 0;
5487 atomic_dec(&rq->nr_iowait);
5488 delayacct_blkio_end();
5493 * sys_sched_get_priority_max - return maximum RT priority.
5494 * @policy: scheduling class.
5496 * this syscall returns the maximum rt_priority that can be used
5497 * by a given scheduling class.
5499 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5506 ret = MAX_USER_RT_PRIO-1;
5518 * sys_sched_get_priority_min - return minimum RT priority.
5519 * @policy: scheduling class.
5521 * this syscall returns the minimum rt_priority that can be used
5522 * by a given scheduling class.
5524 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5542 * sys_sched_rr_get_interval - return the default timeslice of a process.
5543 * @pid: pid of the process.
5544 * @interval: userspace pointer to the timeslice value.
5546 * this syscall writes the default timeslice value of a given process
5547 * into the user-space timespec buffer. A value of '0' means infinity.
5549 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5550 struct timespec __user *, interval)
5552 struct task_struct *p;
5553 unsigned int time_slice;
5554 unsigned long flags;
5564 p = find_process_by_pid(pid);
5568 retval = security_task_getscheduler(p);
5572 rq = task_rq_lock(p, &flags);
5573 time_slice = p->sched_class->get_rr_interval(rq, p);
5574 task_rq_unlock(rq, &flags);
5577 jiffies_to_timespec(time_slice, &t);
5578 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5586 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5588 void sched_show_task(struct task_struct *p)
5590 unsigned long free = 0;
5593 state = p->state ? __ffs(p->state) + 1 : 0;
5594 printk(KERN_INFO "%-13.13s %c", p->comm,
5595 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5596 #if BITS_PER_LONG == 32
5597 if (state == TASK_RUNNING)
5598 printk(KERN_CONT " running ");
5600 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5602 if (state == TASK_RUNNING)
5603 printk(KERN_CONT " running task ");
5605 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5607 #ifdef CONFIG_DEBUG_STACK_USAGE
5608 free = stack_not_used(p);
5610 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5611 task_pid_nr(p), task_pid_nr(p->real_parent),
5612 (unsigned long)task_thread_info(p)->flags);
5614 show_stack(p, NULL);
5617 void show_state_filter(unsigned long state_filter)
5619 struct task_struct *g, *p;
5621 #if BITS_PER_LONG == 32
5623 " task PC stack pid father\n");
5626 " task PC stack pid father\n");
5628 read_lock(&tasklist_lock);
5629 do_each_thread(g, p) {
5631 * reset the NMI-timeout, listing all files on a slow
5632 * console might take alot of time:
5634 touch_nmi_watchdog();
5635 if (!state_filter || (p->state & state_filter))
5637 } while_each_thread(g, p);
5639 touch_all_softlockup_watchdogs();
5641 #ifdef CONFIG_SCHED_DEBUG
5642 sysrq_sched_debug_show();
5644 read_unlock(&tasklist_lock);
5646 * Only show locks if all tasks are dumped:
5649 debug_show_all_locks();
5652 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5654 idle->sched_class = &idle_sched_class;
5658 * init_idle - set up an idle thread for a given CPU
5659 * @idle: task in question
5660 * @cpu: cpu the idle task belongs to
5662 * NOTE: this function does not set the idle thread's NEED_RESCHED
5663 * flag, to make booting more robust.
5665 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5667 struct rq *rq = cpu_rq(cpu);
5668 unsigned long flags;
5670 raw_spin_lock_irqsave(&rq->lock, flags);
5673 idle->state = TASK_RUNNING;
5674 idle->se.exec_start = sched_clock();
5676 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5678 * We're having a chicken and egg problem, even though we are
5679 * holding rq->lock, the cpu isn't yet set to this cpu so the
5680 * lockdep check in task_group() will fail.
5682 * Similar case to sched_fork(). / Alternatively we could
5683 * use task_rq_lock() here and obtain the other rq->lock.
5688 __set_task_cpu(idle, cpu);
5691 rq->curr = rq->idle = idle;
5692 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5695 raw_spin_unlock_irqrestore(&rq->lock, flags);
5697 /* Set the preempt count _outside_ the spinlocks! */
5698 #if defined(CONFIG_PREEMPT)
5699 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5701 task_thread_info(idle)->preempt_count = 0;
5704 * The idle tasks have their own, simple scheduling class:
5706 idle->sched_class = &idle_sched_class;
5707 ftrace_graph_init_task(idle);
5711 * In a system that switches off the HZ timer nohz_cpu_mask
5712 * indicates which cpus entered this state. This is used
5713 * in the rcu update to wait only for active cpus. For system
5714 * which do not switch off the HZ timer nohz_cpu_mask should
5715 * always be CPU_BITS_NONE.
5717 cpumask_var_t nohz_cpu_mask;
5720 * Increase the granularity value when there are more CPUs,
5721 * because with more CPUs the 'effective latency' as visible
5722 * to users decreases. But the relationship is not linear,
5723 * so pick a second-best guess by going with the log2 of the
5726 * This idea comes from the SD scheduler of Con Kolivas:
5728 static int get_update_sysctl_factor(void)
5730 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5731 unsigned int factor;
5733 switch (sysctl_sched_tunable_scaling) {
5734 case SCHED_TUNABLESCALING_NONE:
5737 case SCHED_TUNABLESCALING_LINEAR:
5740 case SCHED_TUNABLESCALING_LOG:
5742 factor = 1 + ilog2(cpus);
5749 static void update_sysctl(void)
5751 unsigned int factor = get_update_sysctl_factor();
5753 #define SET_SYSCTL(name) \
5754 (sysctl_##name = (factor) * normalized_sysctl_##name)
5755 SET_SYSCTL(sched_min_granularity);
5756 SET_SYSCTL(sched_latency);
5757 SET_SYSCTL(sched_wakeup_granularity);
5758 SET_SYSCTL(sched_shares_ratelimit);
5762 static inline void sched_init_granularity(void)
5769 * This is how migration works:
5771 * 1) we invoke migration_cpu_stop() on the target CPU using
5773 * 2) stopper starts to run (implicitly forcing the migrated thread
5775 * 3) it checks whether the migrated task is still in the wrong runqueue.
5776 * 4) if it's in the wrong runqueue then the migration thread removes
5777 * it and puts it into the right queue.
5778 * 5) stopper completes and stop_one_cpu() returns and the migration
5783 * Change a given task's CPU affinity. Migrate the thread to a
5784 * proper CPU and schedule it away if the CPU it's executing on
5785 * is removed from the allowed bitmask.
5787 * NOTE: the caller must have a valid reference to the task, the
5788 * task must not exit() & deallocate itself prematurely. The
5789 * call is not atomic; no spinlocks may be held.
5791 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5793 unsigned long flags;
5795 unsigned int dest_cpu;
5799 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5800 * drop the rq->lock and still rely on ->cpus_allowed.
5803 while (task_is_waking(p))
5805 rq = task_rq_lock(p, &flags);
5806 if (task_is_waking(p)) {
5807 task_rq_unlock(rq, &flags);
5811 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5816 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5817 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5822 if (p->sched_class->set_cpus_allowed)
5823 p->sched_class->set_cpus_allowed(p, new_mask);
5825 cpumask_copy(&p->cpus_allowed, new_mask);
5826 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5829 /* Can the task run on the task's current CPU? If so, we're done */
5830 if (cpumask_test_cpu(task_cpu(p), new_mask))
5833 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5834 if (migrate_task(p, dest_cpu)) {
5835 struct migration_arg arg = { p, dest_cpu };
5836 /* Need help from migration thread: drop lock and wait. */
5837 task_rq_unlock(rq, &flags);
5838 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5839 tlb_migrate_finish(p->mm);
5843 task_rq_unlock(rq, &flags);
5847 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5850 * Move (not current) task off this cpu, onto dest cpu. We're doing
5851 * this because either it can't run here any more (set_cpus_allowed()
5852 * away from this CPU, or CPU going down), or because we're
5853 * attempting to rebalance this task on exec (sched_exec).
5855 * So we race with normal scheduler movements, but that's OK, as long
5856 * as the task is no longer on this CPU.
5858 * Returns non-zero if task was successfully migrated.
5860 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5862 struct rq *rq_dest, *rq_src;
5865 if (unlikely(!cpu_active(dest_cpu)))
5868 rq_src = cpu_rq(src_cpu);
5869 rq_dest = cpu_rq(dest_cpu);
5871 double_rq_lock(rq_src, rq_dest);
5872 /* Already moved. */
5873 if (task_cpu(p) != src_cpu)
5875 /* Affinity changed (again). */
5876 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5880 * If we're not on a rq, the next wake-up will ensure we're
5884 deactivate_task(rq_src, p, 0);
5885 set_task_cpu(p, dest_cpu);
5886 activate_task(rq_dest, p, 0);
5887 check_preempt_curr(rq_dest, p, 0);
5892 double_rq_unlock(rq_src, rq_dest);
5897 * migration_cpu_stop - this will be executed by a highprio stopper thread
5898 * and performs thread migration by bumping thread off CPU then
5899 * 'pushing' onto another runqueue.
5901 static int migration_cpu_stop(void *data)
5903 struct migration_arg *arg = data;
5906 * The original target cpu might have gone down and we might
5907 * be on another cpu but it doesn't matter.
5909 local_irq_disable();
5910 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5915 #ifdef CONFIG_HOTPLUG_CPU
5917 * Figure out where task on dead CPU should go, use force if necessary.
5919 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5921 struct rq *rq = cpu_rq(dead_cpu);
5922 int needs_cpu, uninitialized_var(dest_cpu);
5923 unsigned long flags;
5925 local_irq_save(flags);
5927 raw_spin_lock(&rq->lock);
5928 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5930 dest_cpu = select_fallback_rq(dead_cpu, p);
5931 raw_spin_unlock(&rq->lock);
5933 * It can only fail if we race with set_cpus_allowed(),
5934 * in the racer should migrate the task anyway.
5937 __migrate_task(p, dead_cpu, dest_cpu);
5938 local_irq_restore(flags);
5942 * While a dead CPU has no uninterruptible tasks queued at this point,
5943 * it might still have a nonzero ->nr_uninterruptible counter, because
5944 * for performance reasons the counter is not stricly tracking tasks to
5945 * their home CPUs. So we just add the counter to another CPU's counter,
5946 * to keep the global sum constant after CPU-down:
5948 static void migrate_nr_uninterruptible(struct rq *rq_src)
5950 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5951 unsigned long flags;
5953 local_irq_save(flags);
5954 double_rq_lock(rq_src, rq_dest);
5955 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5956 rq_src->nr_uninterruptible = 0;
5957 double_rq_unlock(rq_src, rq_dest);
5958 local_irq_restore(flags);
5961 /* Run through task list and migrate tasks from the dead cpu. */
5962 static void migrate_live_tasks(int src_cpu)
5964 struct task_struct *p, *t;
5966 read_lock(&tasklist_lock);
5968 do_each_thread(t, p) {
5972 if (task_cpu(p) == src_cpu)
5973 move_task_off_dead_cpu(src_cpu, p);
5974 } while_each_thread(t, p);
5976 read_unlock(&tasklist_lock);
5980 * Schedules idle task to be the next runnable task on current CPU.
5981 * It does so by boosting its priority to highest possible.
5982 * Used by CPU offline code.
5984 void sched_idle_next(void)
5986 int this_cpu = smp_processor_id();
5987 struct rq *rq = cpu_rq(this_cpu);
5988 struct task_struct *p = rq->idle;
5989 unsigned long flags;
5991 /* cpu has to be offline */
5992 BUG_ON(cpu_online(this_cpu));
5995 * Strictly not necessary since rest of the CPUs are stopped by now
5996 * and interrupts disabled on the current cpu.
5998 raw_spin_lock_irqsave(&rq->lock, flags);
6000 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6002 activate_task(rq, p, 0);
6004 raw_spin_unlock_irqrestore(&rq->lock, flags);
6008 * Ensures that the idle task is using init_mm right before its cpu goes
6011 void idle_task_exit(void)
6013 struct mm_struct *mm = current->active_mm;
6015 BUG_ON(cpu_online(smp_processor_id()));
6018 switch_mm(mm, &init_mm, current);
6022 /* called under rq->lock with disabled interrupts */
6023 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6025 struct rq *rq = cpu_rq(dead_cpu);
6027 /* Must be exiting, otherwise would be on tasklist. */
6028 BUG_ON(!p->exit_state);
6030 /* Cannot have done final schedule yet: would have vanished. */
6031 BUG_ON(p->state == TASK_DEAD);
6036 * Drop lock around migration; if someone else moves it,
6037 * that's OK. No task can be added to this CPU, so iteration is
6040 raw_spin_unlock_irq(&rq->lock);
6041 move_task_off_dead_cpu(dead_cpu, p);
6042 raw_spin_lock_irq(&rq->lock);
6047 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6048 static void migrate_dead_tasks(unsigned int dead_cpu)
6050 struct rq *rq = cpu_rq(dead_cpu);
6051 struct task_struct *next;
6054 if (!rq->nr_running)
6056 next = pick_next_task(rq);
6059 next->sched_class->put_prev_task(rq, next);
6060 migrate_dead(dead_cpu, next);
6066 * remove the tasks which were accounted by rq from calc_load_tasks.
6068 static void calc_global_load_remove(struct rq *rq)
6070 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6071 rq->calc_load_active = 0;
6073 #endif /* CONFIG_HOTPLUG_CPU */
6075 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6077 static struct ctl_table sd_ctl_dir[] = {
6079 .procname = "sched_domain",
6085 static struct ctl_table sd_ctl_root[] = {
6087 .procname = "kernel",
6089 .child = sd_ctl_dir,
6094 static struct ctl_table *sd_alloc_ctl_entry(int n)
6096 struct ctl_table *entry =
6097 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6102 static void sd_free_ctl_entry(struct ctl_table **tablep)
6104 struct ctl_table *entry;
6107 * In the intermediate directories, both the child directory and
6108 * procname are dynamically allocated and could fail but the mode
6109 * will always be set. In the lowest directory the names are
6110 * static strings and all have proc handlers.
6112 for (entry = *tablep; entry->mode; entry++) {
6114 sd_free_ctl_entry(&entry->child);
6115 if (entry->proc_handler == NULL)
6116 kfree(entry->procname);
6124 set_table_entry(struct ctl_table *entry,
6125 const char *procname, void *data, int maxlen,
6126 mode_t mode, proc_handler *proc_handler)
6128 entry->procname = procname;
6130 entry->maxlen = maxlen;
6132 entry->proc_handler = proc_handler;
6135 static struct ctl_table *
6136 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6138 struct ctl_table *table = sd_alloc_ctl_entry(13);
6143 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6144 sizeof(long), 0644, proc_doulongvec_minmax);
6145 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6146 sizeof(long), 0644, proc_doulongvec_minmax);
6147 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6148 sizeof(int), 0644, proc_dointvec_minmax);
6149 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6150 sizeof(int), 0644, proc_dointvec_minmax);
6151 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6152 sizeof(int), 0644, proc_dointvec_minmax);
6153 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6154 sizeof(int), 0644, proc_dointvec_minmax);
6155 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6156 sizeof(int), 0644, proc_dointvec_minmax);
6157 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6158 sizeof(int), 0644, proc_dointvec_minmax);
6159 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6160 sizeof(int), 0644, proc_dointvec_minmax);
6161 set_table_entry(&table[9], "cache_nice_tries",
6162 &sd->cache_nice_tries,
6163 sizeof(int), 0644, proc_dointvec_minmax);
6164 set_table_entry(&table[10], "flags", &sd->flags,
6165 sizeof(int), 0644, proc_dointvec_minmax);
6166 set_table_entry(&table[11], "name", sd->name,
6167 CORENAME_MAX_SIZE, 0444, proc_dostring);
6168 /* &table[12] is terminator */
6173 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6175 struct ctl_table *entry, *table;
6176 struct sched_domain *sd;
6177 int domain_num = 0, i;
6180 for_each_domain(cpu, sd)
6182 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6187 for_each_domain(cpu, sd) {
6188 snprintf(buf, 32, "domain%d", i);
6189 entry->procname = kstrdup(buf, GFP_KERNEL);
6191 entry->child = sd_alloc_ctl_domain_table(sd);
6198 static struct ctl_table_header *sd_sysctl_header;
6199 static void register_sched_domain_sysctl(void)
6201 int i, cpu_num = num_possible_cpus();
6202 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6205 WARN_ON(sd_ctl_dir[0].child);
6206 sd_ctl_dir[0].child = entry;
6211 for_each_possible_cpu(i) {
6212 snprintf(buf, 32, "cpu%d", i);
6213 entry->procname = kstrdup(buf, GFP_KERNEL);
6215 entry->child = sd_alloc_ctl_cpu_table(i);
6219 WARN_ON(sd_sysctl_header);
6220 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6223 /* may be called multiple times per register */
6224 static void unregister_sched_domain_sysctl(void)
6226 if (sd_sysctl_header)
6227 unregister_sysctl_table(sd_sysctl_header);
6228 sd_sysctl_header = NULL;
6229 if (sd_ctl_dir[0].child)
6230 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6233 static void register_sched_domain_sysctl(void)
6236 static void unregister_sched_domain_sysctl(void)
6241 static void set_rq_online(struct rq *rq)
6244 const struct sched_class *class;
6246 cpumask_set_cpu(rq->cpu, rq->rd->online);
6249 for_each_class(class) {
6250 if (class->rq_online)
6251 class->rq_online(rq);
6256 static void set_rq_offline(struct rq *rq)
6259 const struct sched_class *class;
6261 for_each_class(class) {
6262 if (class->rq_offline)
6263 class->rq_offline(rq);
6266 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6272 * migration_call - callback that gets triggered when a CPU is added.
6273 * Here we can start up the necessary migration thread for the new CPU.
6275 static int __cpuinit
6276 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6278 int cpu = (long)hcpu;
6279 unsigned long flags;
6280 struct rq *rq = cpu_rq(cpu);
6284 case CPU_UP_PREPARE:
6285 case CPU_UP_PREPARE_FROZEN:
6286 rq->calc_load_update = calc_load_update;
6290 case CPU_ONLINE_FROZEN:
6291 /* Update our root-domain */
6292 raw_spin_lock_irqsave(&rq->lock, flags);
6294 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6298 raw_spin_unlock_irqrestore(&rq->lock, flags);
6301 #ifdef CONFIG_HOTPLUG_CPU
6303 case CPU_DEAD_FROZEN:
6304 migrate_live_tasks(cpu);
6305 /* Idle task back to normal (off runqueue, low prio) */
6306 raw_spin_lock_irq(&rq->lock);
6307 deactivate_task(rq, rq->idle, 0);
6308 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6309 rq->idle->sched_class = &idle_sched_class;
6310 migrate_dead_tasks(cpu);
6311 raw_spin_unlock_irq(&rq->lock);
6312 migrate_nr_uninterruptible(rq);
6313 BUG_ON(rq->nr_running != 0);
6314 calc_global_load_remove(rq);
6318 case CPU_DYING_FROZEN:
6319 /* Update our root-domain */
6320 raw_spin_lock_irqsave(&rq->lock, flags);
6322 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6325 raw_spin_unlock_irqrestore(&rq->lock, flags);
6333 * Register at high priority so that task migration (migrate_all_tasks)
6334 * happens before everything else. This has to be lower priority than
6335 * the notifier in the perf_event subsystem, though.
6337 static struct notifier_block __cpuinitdata migration_notifier = {
6338 .notifier_call = migration_call,
6339 .priority = CPU_PRI_MIGRATION,
6342 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6343 unsigned long action, void *hcpu)
6345 switch (action & ~CPU_TASKS_FROZEN) {
6347 case CPU_DOWN_FAILED:
6348 set_cpu_active((long)hcpu, true);
6355 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6356 unsigned long action, void *hcpu)
6358 switch (action & ~CPU_TASKS_FROZEN) {
6359 case CPU_DOWN_PREPARE:
6360 set_cpu_active((long)hcpu, false);
6367 static int __init migration_init(void)
6369 void *cpu = (void *)(long)smp_processor_id();
6372 /* Initialize migration for the boot CPU */
6373 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6374 BUG_ON(err == NOTIFY_BAD);
6375 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6376 register_cpu_notifier(&migration_notifier);
6378 /* Register cpu active notifiers */
6379 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6380 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6384 early_initcall(migration_init);
6389 #ifdef CONFIG_SCHED_DEBUG
6391 static __read_mostly int sched_domain_debug_enabled;
6393 static int __init sched_domain_debug_setup(char *str)
6395 sched_domain_debug_enabled = 1;
6399 early_param("sched_debug", sched_domain_debug_setup);
6401 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6402 struct cpumask *groupmask)
6404 struct sched_group *group = sd->groups;
6407 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6408 cpumask_clear(groupmask);
6410 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6412 if (!(sd->flags & SD_LOAD_BALANCE)) {
6413 printk("does not load-balance\n");
6415 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6420 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6422 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6423 printk(KERN_ERR "ERROR: domain->span does not contain "
6426 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6427 printk(KERN_ERR "ERROR: domain->groups does not contain"
6431 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6435 printk(KERN_ERR "ERROR: group is NULL\n");
6439 if (!group->cpu_power) {
6440 printk(KERN_CONT "\n");
6441 printk(KERN_ERR "ERROR: domain->cpu_power not "
6446 if (!cpumask_weight(sched_group_cpus(group))) {
6447 printk(KERN_CONT "\n");
6448 printk(KERN_ERR "ERROR: empty group\n");
6452 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6453 printk(KERN_CONT "\n");
6454 printk(KERN_ERR "ERROR: repeated CPUs\n");
6458 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6460 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6462 printk(KERN_CONT " %s", str);
6463 if (group->cpu_power != SCHED_LOAD_SCALE) {
6464 printk(KERN_CONT " (cpu_power = %d)",
6468 group = group->next;
6469 } while (group != sd->groups);
6470 printk(KERN_CONT "\n");
6472 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6473 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6476 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6477 printk(KERN_ERR "ERROR: parent span is not a superset "
6478 "of domain->span\n");
6482 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6484 cpumask_var_t groupmask;
6487 if (!sched_domain_debug_enabled)
6491 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6495 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6497 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6498 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6503 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6510 free_cpumask_var(groupmask);
6512 #else /* !CONFIG_SCHED_DEBUG */
6513 # define sched_domain_debug(sd, cpu) do { } while (0)
6514 #endif /* CONFIG_SCHED_DEBUG */
6516 static int sd_degenerate(struct sched_domain *sd)
6518 if (cpumask_weight(sched_domain_span(sd)) == 1)
6521 /* Following flags need at least 2 groups */
6522 if (sd->flags & (SD_LOAD_BALANCE |
6523 SD_BALANCE_NEWIDLE |
6527 SD_SHARE_PKG_RESOURCES)) {
6528 if (sd->groups != sd->groups->next)
6532 /* Following flags don't use groups */
6533 if (sd->flags & (SD_WAKE_AFFINE))
6540 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6542 unsigned long cflags = sd->flags, pflags = parent->flags;
6544 if (sd_degenerate(parent))
6547 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6550 /* Flags needing groups don't count if only 1 group in parent */
6551 if (parent->groups == parent->groups->next) {
6552 pflags &= ~(SD_LOAD_BALANCE |
6553 SD_BALANCE_NEWIDLE |
6557 SD_SHARE_PKG_RESOURCES);
6558 if (nr_node_ids == 1)
6559 pflags &= ~SD_SERIALIZE;
6561 if (~cflags & pflags)
6567 static void free_rootdomain(struct root_domain *rd)
6569 synchronize_sched();
6571 cpupri_cleanup(&rd->cpupri);
6573 free_cpumask_var(rd->rto_mask);
6574 free_cpumask_var(rd->online);
6575 free_cpumask_var(rd->span);
6579 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6581 struct root_domain *old_rd = NULL;
6582 unsigned long flags;
6584 raw_spin_lock_irqsave(&rq->lock, flags);
6589 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6592 cpumask_clear_cpu(rq->cpu, old_rd->span);
6595 * If we dont want to free the old_rt yet then
6596 * set old_rd to NULL to skip the freeing later
6599 if (!atomic_dec_and_test(&old_rd->refcount))
6603 atomic_inc(&rd->refcount);
6606 cpumask_set_cpu(rq->cpu, rd->span);
6607 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6610 raw_spin_unlock_irqrestore(&rq->lock, flags);
6613 free_rootdomain(old_rd);
6616 static int init_rootdomain(struct root_domain *rd)
6618 memset(rd, 0, sizeof(*rd));
6620 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6622 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6624 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6627 if (cpupri_init(&rd->cpupri) != 0)
6632 free_cpumask_var(rd->rto_mask);
6634 free_cpumask_var(rd->online);
6636 free_cpumask_var(rd->span);
6641 static void init_defrootdomain(void)
6643 init_rootdomain(&def_root_domain);
6645 atomic_set(&def_root_domain.refcount, 1);
6648 static struct root_domain *alloc_rootdomain(void)
6650 struct root_domain *rd;
6652 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6656 if (init_rootdomain(rd) != 0) {
6665 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6666 * hold the hotplug lock.
6669 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6671 struct rq *rq = cpu_rq(cpu);
6672 struct sched_domain *tmp;
6674 for (tmp = sd; tmp; tmp = tmp->parent)
6675 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6677 /* Remove the sched domains which do not contribute to scheduling. */
6678 for (tmp = sd; tmp; ) {
6679 struct sched_domain *parent = tmp->parent;
6683 if (sd_parent_degenerate(tmp, parent)) {
6684 tmp->parent = parent->parent;
6686 parent->parent->child = tmp;
6691 if (sd && sd_degenerate(sd)) {
6697 sched_domain_debug(sd, cpu);
6699 rq_attach_root(rq, rd);
6700 rcu_assign_pointer(rq->sd, sd);
6703 /* cpus with isolated domains */
6704 static cpumask_var_t cpu_isolated_map;
6706 /* Setup the mask of cpus configured for isolated domains */
6707 static int __init isolated_cpu_setup(char *str)
6709 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6710 cpulist_parse(str, cpu_isolated_map);
6714 __setup("isolcpus=", isolated_cpu_setup);
6717 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6718 * to a function which identifies what group(along with sched group) a CPU
6719 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6720 * (due to the fact that we keep track of groups covered with a struct cpumask).
6722 * init_sched_build_groups will build a circular linked list of the groups
6723 * covered by the given span, and will set each group's ->cpumask correctly,
6724 * and ->cpu_power to 0.
6727 init_sched_build_groups(const struct cpumask *span,
6728 const struct cpumask *cpu_map,
6729 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6730 struct sched_group **sg,
6731 struct cpumask *tmpmask),
6732 struct cpumask *covered, struct cpumask *tmpmask)
6734 struct sched_group *first = NULL, *last = NULL;
6737 cpumask_clear(covered);
6739 for_each_cpu(i, span) {
6740 struct sched_group *sg;
6741 int group = group_fn(i, cpu_map, &sg, tmpmask);
6744 if (cpumask_test_cpu(i, covered))
6747 cpumask_clear(sched_group_cpus(sg));
6750 for_each_cpu(j, span) {
6751 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6754 cpumask_set_cpu(j, covered);
6755 cpumask_set_cpu(j, sched_group_cpus(sg));
6766 #define SD_NODES_PER_DOMAIN 16
6771 * find_next_best_node - find the next node to include in a sched_domain
6772 * @node: node whose sched_domain we're building
6773 * @used_nodes: nodes already in the sched_domain
6775 * Find the next node to include in a given scheduling domain. Simply
6776 * finds the closest node not already in the @used_nodes map.
6778 * Should use nodemask_t.
6780 static int find_next_best_node(int node, nodemask_t *used_nodes)
6782 int i, n, val, min_val, best_node = 0;
6786 for (i = 0; i < nr_node_ids; i++) {
6787 /* Start at @node */
6788 n = (node + i) % nr_node_ids;
6790 if (!nr_cpus_node(n))
6793 /* Skip already used nodes */
6794 if (node_isset(n, *used_nodes))
6797 /* Simple min distance search */
6798 val = node_distance(node, n);
6800 if (val < min_val) {
6806 node_set(best_node, *used_nodes);
6811 * sched_domain_node_span - get a cpumask for a node's sched_domain
6812 * @node: node whose cpumask we're constructing
6813 * @span: resulting cpumask
6815 * Given a node, construct a good cpumask for its sched_domain to span. It
6816 * should be one that prevents unnecessary balancing, but also spreads tasks
6819 static void sched_domain_node_span(int node, struct cpumask *span)
6821 nodemask_t used_nodes;
6824 cpumask_clear(span);
6825 nodes_clear(used_nodes);
6827 cpumask_or(span, span, cpumask_of_node(node));
6828 node_set(node, used_nodes);
6830 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6831 int next_node = find_next_best_node(node, &used_nodes);
6833 cpumask_or(span, span, cpumask_of_node(next_node));
6836 #endif /* CONFIG_NUMA */
6838 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6841 * The cpus mask in sched_group and sched_domain hangs off the end.
6843 * ( See the the comments in include/linux/sched.h:struct sched_group
6844 * and struct sched_domain. )
6846 struct static_sched_group {
6847 struct sched_group sg;
6848 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6851 struct static_sched_domain {
6852 struct sched_domain sd;
6853 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6859 cpumask_var_t domainspan;
6860 cpumask_var_t covered;
6861 cpumask_var_t notcovered;
6863 cpumask_var_t nodemask;
6864 cpumask_var_t this_sibling_map;
6865 cpumask_var_t this_core_map;
6866 cpumask_var_t this_book_map;
6867 cpumask_var_t send_covered;
6868 cpumask_var_t tmpmask;
6869 struct sched_group **sched_group_nodes;
6870 struct root_domain *rd;
6874 sa_sched_groups = 0,
6880 sa_this_sibling_map,
6882 sa_sched_group_nodes,
6892 * SMT sched-domains:
6894 #ifdef CONFIG_SCHED_SMT
6895 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6896 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6899 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6900 struct sched_group **sg, struct cpumask *unused)
6903 *sg = &per_cpu(sched_groups, cpu).sg;
6906 #endif /* CONFIG_SCHED_SMT */
6909 * multi-core sched-domains:
6911 #ifdef CONFIG_SCHED_MC
6912 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6913 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6916 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6917 struct sched_group **sg, struct cpumask *mask)
6920 #ifdef CONFIG_SCHED_SMT
6921 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6922 group = cpumask_first(mask);
6927 *sg = &per_cpu(sched_group_core, group).sg;
6930 #endif /* CONFIG_SCHED_MC */
6933 * book sched-domains:
6935 #ifdef CONFIG_SCHED_BOOK
6936 static DEFINE_PER_CPU(struct static_sched_domain, book_domains);
6937 static DEFINE_PER_CPU(struct static_sched_group, sched_group_book);
6940 cpu_to_book_group(int cpu, const struct cpumask *cpu_map,
6941 struct sched_group **sg, struct cpumask *mask)
6944 #ifdef CONFIG_SCHED_MC
6945 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6946 group = cpumask_first(mask);
6947 #elif defined(CONFIG_SCHED_SMT)
6948 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6949 group = cpumask_first(mask);
6952 *sg = &per_cpu(sched_group_book, group).sg;
6955 #endif /* CONFIG_SCHED_BOOK */
6957 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6958 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6961 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6962 struct sched_group **sg, struct cpumask *mask)
6965 #ifdef CONFIG_SCHED_BOOK
6966 cpumask_and(mask, cpu_book_mask(cpu), cpu_map);
6967 group = cpumask_first(mask);
6968 #elif defined(CONFIG_SCHED_MC)
6969 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6970 group = cpumask_first(mask);
6971 #elif defined(CONFIG_SCHED_SMT)
6972 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6973 group = cpumask_first(mask);
6978 *sg = &per_cpu(sched_group_phys, group).sg;
6984 * The init_sched_build_groups can't handle what we want to do with node
6985 * groups, so roll our own. Now each node has its own list of groups which
6986 * gets dynamically allocated.
6988 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6989 static struct sched_group ***sched_group_nodes_bycpu;
6991 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6992 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6994 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6995 struct sched_group **sg,
6996 struct cpumask *nodemask)
7000 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
7001 group = cpumask_first(nodemask);
7004 *sg = &per_cpu(sched_group_allnodes, group).sg;
7008 static void init_numa_sched_groups_power(struct sched_group *group_head)
7010 struct sched_group *sg = group_head;
7016 for_each_cpu(j, sched_group_cpus(sg)) {
7017 struct sched_domain *sd;
7019 sd = &per_cpu(phys_domains, j).sd;
7020 if (j != group_first_cpu(sd->groups)) {
7022 * Only add "power" once for each
7028 sg->cpu_power += sd->groups->cpu_power;
7031 } while (sg != group_head);
7034 static int build_numa_sched_groups(struct s_data *d,
7035 const struct cpumask *cpu_map, int num)
7037 struct sched_domain *sd;
7038 struct sched_group *sg, *prev;
7041 cpumask_clear(d->covered);
7042 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
7043 if (cpumask_empty(d->nodemask)) {
7044 d->sched_group_nodes[num] = NULL;
7048 sched_domain_node_span(num, d->domainspan);
7049 cpumask_and(d->domainspan, d->domainspan, cpu_map);
7051 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7054 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
7058 d->sched_group_nodes[num] = sg;
7060 for_each_cpu(j, d->nodemask) {
7061 sd = &per_cpu(node_domains, j).sd;
7066 cpumask_copy(sched_group_cpus(sg), d->nodemask);
7068 cpumask_or(d->covered, d->covered, d->nodemask);
7071 for (j = 0; j < nr_node_ids; j++) {
7072 n = (num + j) % nr_node_ids;
7073 cpumask_complement(d->notcovered, d->covered);
7074 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
7075 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
7076 if (cpumask_empty(d->tmpmask))
7078 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
7079 if (cpumask_empty(d->tmpmask))
7081 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7085 "Can not alloc domain group for node %d\n", j);
7089 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
7090 sg->next = prev->next;
7091 cpumask_or(d->covered, d->covered, d->tmpmask);
7098 #endif /* CONFIG_NUMA */
7101 /* Free memory allocated for various sched_group structures */
7102 static void free_sched_groups(const struct cpumask *cpu_map,
7103 struct cpumask *nodemask)
7107 for_each_cpu(cpu, cpu_map) {
7108 struct sched_group **sched_group_nodes
7109 = sched_group_nodes_bycpu[cpu];
7111 if (!sched_group_nodes)
7114 for (i = 0; i < nr_node_ids; i++) {
7115 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7117 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7118 if (cpumask_empty(nodemask))
7128 if (oldsg != sched_group_nodes[i])
7131 kfree(sched_group_nodes);
7132 sched_group_nodes_bycpu[cpu] = NULL;
7135 #else /* !CONFIG_NUMA */
7136 static void free_sched_groups(const struct cpumask *cpu_map,
7137 struct cpumask *nodemask)
7140 #endif /* CONFIG_NUMA */
7143 * Initialize sched groups cpu_power.
7145 * cpu_power indicates the capacity of sched group, which is used while
7146 * distributing the load between different sched groups in a sched domain.
7147 * Typically cpu_power for all the groups in a sched domain will be same unless
7148 * there are asymmetries in the topology. If there are asymmetries, group
7149 * having more cpu_power will pickup more load compared to the group having
7152 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7154 struct sched_domain *child;
7155 struct sched_group *group;
7159 WARN_ON(!sd || !sd->groups);
7161 if (cpu != group_first_cpu(sd->groups))
7164 sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups));
7168 sd->groups->cpu_power = 0;
7171 power = SCHED_LOAD_SCALE;
7172 weight = cpumask_weight(sched_domain_span(sd));
7174 * SMT siblings share the power of a single core.
7175 * Usually multiple threads get a better yield out of
7176 * that one core than a single thread would have,
7177 * reflect that in sd->smt_gain.
7179 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
7180 power *= sd->smt_gain;
7182 power >>= SCHED_LOAD_SHIFT;
7184 sd->groups->cpu_power += power;
7189 * Add cpu_power of each child group to this groups cpu_power.
7191 group = child->groups;
7193 sd->groups->cpu_power += group->cpu_power;
7194 group = group->next;
7195 } while (group != child->groups);
7199 * Initializers for schedule domains
7200 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7203 #ifdef CONFIG_SCHED_DEBUG
7204 # define SD_INIT_NAME(sd, type) sd->name = #type
7206 # define SD_INIT_NAME(sd, type) do { } while (0)
7209 #define SD_INIT(sd, type) sd_init_##type(sd)
7211 #define SD_INIT_FUNC(type) \
7212 static noinline void sd_init_##type(struct sched_domain *sd) \
7214 memset(sd, 0, sizeof(*sd)); \
7215 *sd = SD_##type##_INIT; \
7216 sd->level = SD_LV_##type; \
7217 SD_INIT_NAME(sd, type); \
7222 SD_INIT_FUNC(ALLNODES)
7225 #ifdef CONFIG_SCHED_SMT
7226 SD_INIT_FUNC(SIBLING)
7228 #ifdef CONFIG_SCHED_MC
7231 #ifdef CONFIG_SCHED_BOOK
7235 static int default_relax_domain_level = -1;
7237 static int __init setup_relax_domain_level(char *str)
7241 val = simple_strtoul(str, NULL, 0);
7242 if (val < SD_LV_MAX)
7243 default_relax_domain_level = val;
7247 __setup("relax_domain_level=", setup_relax_domain_level);
7249 static void set_domain_attribute(struct sched_domain *sd,
7250 struct sched_domain_attr *attr)
7254 if (!attr || attr->relax_domain_level < 0) {
7255 if (default_relax_domain_level < 0)
7258 request = default_relax_domain_level;
7260 request = attr->relax_domain_level;
7261 if (request < sd->level) {
7262 /* turn off idle balance on this domain */
7263 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7265 /* turn on idle balance on this domain */
7266 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7270 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7271 const struct cpumask *cpu_map)
7274 case sa_sched_groups:
7275 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
7276 d->sched_group_nodes = NULL;
7278 free_rootdomain(d->rd); /* fall through */
7280 free_cpumask_var(d->tmpmask); /* fall through */
7281 case sa_send_covered:
7282 free_cpumask_var(d->send_covered); /* fall through */
7283 case sa_this_book_map:
7284 free_cpumask_var(d->this_book_map); /* fall through */
7285 case sa_this_core_map:
7286 free_cpumask_var(d->this_core_map); /* fall through */
7287 case sa_this_sibling_map:
7288 free_cpumask_var(d->this_sibling_map); /* fall through */
7290 free_cpumask_var(d->nodemask); /* fall through */
7291 case sa_sched_group_nodes:
7293 kfree(d->sched_group_nodes); /* fall through */
7295 free_cpumask_var(d->notcovered); /* fall through */
7297 free_cpumask_var(d->covered); /* fall through */
7299 free_cpumask_var(d->domainspan); /* fall through */
7306 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7307 const struct cpumask *cpu_map)
7310 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
7312 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
7313 return sa_domainspan;
7314 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
7316 /* Allocate the per-node list of sched groups */
7317 d->sched_group_nodes = kcalloc(nr_node_ids,
7318 sizeof(struct sched_group *), GFP_KERNEL);
7319 if (!d->sched_group_nodes) {
7320 printk(KERN_WARNING "Can not alloc sched group node list\n");
7321 return sa_notcovered;
7323 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
7325 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
7326 return sa_sched_group_nodes;
7327 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
7329 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
7330 return sa_this_sibling_map;
7331 if (!alloc_cpumask_var(&d->this_book_map, GFP_KERNEL))
7332 return sa_this_core_map;
7333 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
7334 return sa_this_book_map;
7335 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
7336 return sa_send_covered;
7337 d->rd = alloc_rootdomain();
7339 printk(KERN_WARNING "Cannot alloc root domain\n");
7342 return sa_rootdomain;
7345 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
7346 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
7348 struct sched_domain *sd = NULL;
7350 struct sched_domain *parent;
7353 if (cpumask_weight(cpu_map) >
7354 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
7355 sd = &per_cpu(allnodes_domains, i).sd;
7356 SD_INIT(sd, ALLNODES);
7357 set_domain_attribute(sd, attr);
7358 cpumask_copy(sched_domain_span(sd), cpu_map);
7359 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
7364 sd = &per_cpu(node_domains, i).sd;
7366 set_domain_attribute(sd, attr);
7367 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7368 sd->parent = parent;
7371 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7376 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7377 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7378 struct sched_domain *parent, int i)
7380 struct sched_domain *sd;
7381 sd = &per_cpu(phys_domains, i).sd;
7383 set_domain_attribute(sd, attr);
7384 cpumask_copy(sched_domain_span(sd), d->nodemask);
7385 sd->parent = parent;
7388 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7392 static struct sched_domain *__build_book_sched_domain(struct s_data *d,
7393 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7394 struct sched_domain *parent, int i)
7396 struct sched_domain *sd = parent;
7397 #ifdef CONFIG_SCHED_BOOK
7398 sd = &per_cpu(book_domains, i).sd;
7400 set_domain_attribute(sd, attr);
7401 cpumask_and(sched_domain_span(sd), cpu_map, cpu_book_mask(i));
7402 sd->parent = parent;
7404 cpu_to_book_group(i, cpu_map, &sd->groups, d->tmpmask);
7409 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7410 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7411 struct sched_domain *parent, int i)
7413 struct sched_domain *sd = parent;
7414 #ifdef CONFIG_SCHED_MC
7415 sd = &per_cpu(core_domains, i).sd;
7417 set_domain_attribute(sd, attr);
7418 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7419 sd->parent = parent;
7421 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7426 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7427 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7428 struct sched_domain *parent, int i)
7430 struct sched_domain *sd = parent;
7431 #ifdef CONFIG_SCHED_SMT
7432 sd = &per_cpu(cpu_domains, i).sd;
7433 SD_INIT(sd, SIBLING);
7434 set_domain_attribute(sd, attr);
7435 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7436 sd->parent = parent;
7438 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7443 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7444 const struct cpumask *cpu_map, int cpu)
7447 #ifdef CONFIG_SCHED_SMT
7448 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7449 cpumask_and(d->this_sibling_map, cpu_map,
7450 topology_thread_cpumask(cpu));
7451 if (cpu == cpumask_first(d->this_sibling_map))
7452 init_sched_build_groups(d->this_sibling_map, cpu_map,
7454 d->send_covered, d->tmpmask);
7457 #ifdef CONFIG_SCHED_MC
7458 case SD_LV_MC: /* set up multi-core groups */
7459 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7460 if (cpu == cpumask_first(d->this_core_map))
7461 init_sched_build_groups(d->this_core_map, cpu_map,
7463 d->send_covered, d->tmpmask);
7466 #ifdef CONFIG_SCHED_BOOK
7467 case SD_LV_BOOK: /* set up book groups */
7468 cpumask_and(d->this_book_map, cpu_map, cpu_book_mask(cpu));
7469 if (cpu == cpumask_first(d->this_book_map))
7470 init_sched_build_groups(d->this_book_map, cpu_map,
7472 d->send_covered, d->tmpmask);
7475 case SD_LV_CPU: /* set up physical groups */
7476 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7477 if (!cpumask_empty(d->nodemask))
7478 init_sched_build_groups(d->nodemask, cpu_map,
7480 d->send_covered, d->tmpmask);
7483 case SD_LV_ALLNODES:
7484 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7485 d->send_covered, d->tmpmask);
7494 * Build sched domains for a given set of cpus and attach the sched domains
7495 * to the individual cpus
7497 static int __build_sched_domains(const struct cpumask *cpu_map,
7498 struct sched_domain_attr *attr)
7500 enum s_alloc alloc_state = sa_none;
7502 struct sched_domain *sd;
7508 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7509 if (alloc_state != sa_rootdomain)
7511 alloc_state = sa_sched_groups;
7514 * Set up domains for cpus specified by the cpu_map.
7516 for_each_cpu(i, cpu_map) {
7517 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7520 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7521 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7522 sd = __build_book_sched_domain(&d, cpu_map, attr, sd, i);
7523 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7524 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7527 for_each_cpu(i, cpu_map) {
7528 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7529 build_sched_groups(&d, SD_LV_BOOK, cpu_map, i);
7530 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7533 /* Set up physical groups */
7534 for (i = 0; i < nr_node_ids; i++)
7535 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7538 /* Set up node groups */
7540 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7542 for (i = 0; i < nr_node_ids; i++)
7543 if (build_numa_sched_groups(&d, cpu_map, i))
7547 /* Calculate CPU power for physical packages and nodes */
7548 #ifdef CONFIG_SCHED_SMT
7549 for_each_cpu(i, cpu_map) {
7550 sd = &per_cpu(cpu_domains, i).sd;
7551 init_sched_groups_power(i, sd);
7554 #ifdef CONFIG_SCHED_MC
7555 for_each_cpu(i, cpu_map) {
7556 sd = &per_cpu(core_domains, i).sd;
7557 init_sched_groups_power(i, sd);
7560 #ifdef CONFIG_SCHED_BOOK
7561 for_each_cpu(i, cpu_map) {
7562 sd = &per_cpu(book_domains, i).sd;
7563 init_sched_groups_power(i, sd);
7567 for_each_cpu(i, cpu_map) {
7568 sd = &per_cpu(phys_domains, i).sd;
7569 init_sched_groups_power(i, sd);
7573 for (i = 0; i < nr_node_ids; i++)
7574 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7576 if (d.sd_allnodes) {
7577 struct sched_group *sg;
7579 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7581 init_numa_sched_groups_power(sg);
7585 /* Attach the domains */
7586 for_each_cpu(i, cpu_map) {
7587 #ifdef CONFIG_SCHED_SMT
7588 sd = &per_cpu(cpu_domains, i).sd;
7589 #elif defined(CONFIG_SCHED_MC)
7590 sd = &per_cpu(core_domains, i).sd;
7591 #elif defined(CONFIG_SCHED_BOOK)
7592 sd = &per_cpu(book_domains, i).sd;
7594 sd = &per_cpu(phys_domains, i).sd;
7596 cpu_attach_domain(sd, d.rd, i);
7599 d.sched_group_nodes = NULL; /* don't free this we still need it */
7600 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7604 __free_domain_allocs(&d, alloc_state, cpu_map);
7608 static int build_sched_domains(const struct cpumask *cpu_map)
7610 return __build_sched_domains(cpu_map, NULL);
7613 static cpumask_var_t *doms_cur; /* current sched domains */
7614 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7615 static struct sched_domain_attr *dattr_cur;
7616 /* attribues of custom domains in 'doms_cur' */
7619 * Special case: If a kmalloc of a doms_cur partition (array of
7620 * cpumask) fails, then fallback to a single sched domain,
7621 * as determined by the single cpumask fallback_doms.
7623 static cpumask_var_t fallback_doms;
7626 * arch_update_cpu_topology lets virtualized architectures update the
7627 * cpu core maps. It is supposed to return 1 if the topology changed
7628 * or 0 if it stayed the same.
7630 int __attribute__((weak)) arch_update_cpu_topology(void)
7635 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7638 cpumask_var_t *doms;
7640 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7643 for (i = 0; i < ndoms; i++) {
7644 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7645 free_sched_domains(doms, i);
7652 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7655 for (i = 0; i < ndoms; i++)
7656 free_cpumask_var(doms[i]);
7661 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7662 * For now this just excludes isolated cpus, but could be used to
7663 * exclude other special cases in the future.
7665 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7669 arch_update_cpu_topology();
7671 doms_cur = alloc_sched_domains(ndoms_cur);
7673 doms_cur = &fallback_doms;
7674 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7676 err = build_sched_domains(doms_cur[0]);
7677 register_sched_domain_sysctl();
7682 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7683 struct cpumask *tmpmask)
7685 free_sched_groups(cpu_map, tmpmask);
7689 * Detach sched domains from a group of cpus specified in cpu_map
7690 * These cpus will now be attached to the NULL domain
7692 static void detach_destroy_domains(const struct cpumask *cpu_map)
7694 /* Save because hotplug lock held. */
7695 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7698 for_each_cpu(i, cpu_map)
7699 cpu_attach_domain(NULL, &def_root_domain, i);
7700 synchronize_sched();
7701 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7704 /* handle null as "default" */
7705 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7706 struct sched_domain_attr *new, int idx_new)
7708 struct sched_domain_attr tmp;
7715 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7716 new ? (new + idx_new) : &tmp,
7717 sizeof(struct sched_domain_attr));
7721 * Partition sched domains as specified by the 'ndoms_new'
7722 * cpumasks in the array doms_new[] of cpumasks. This compares
7723 * doms_new[] to the current sched domain partitioning, doms_cur[].
7724 * It destroys each deleted domain and builds each new domain.
7726 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7727 * The masks don't intersect (don't overlap.) We should setup one
7728 * sched domain for each mask. CPUs not in any of the cpumasks will
7729 * not be load balanced. If the same cpumask appears both in the
7730 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7733 * The passed in 'doms_new' should be allocated using
7734 * alloc_sched_domains. This routine takes ownership of it and will
7735 * free_sched_domains it when done with it. If the caller failed the
7736 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7737 * and partition_sched_domains() will fallback to the single partition
7738 * 'fallback_doms', it also forces the domains to be rebuilt.
7740 * If doms_new == NULL it will be replaced with cpu_online_mask.
7741 * ndoms_new == 0 is a special case for destroying existing domains,
7742 * and it will not create the default domain.
7744 * Call with hotplug lock held
7746 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7747 struct sched_domain_attr *dattr_new)
7752 mutex_lock(&sched_domains_mutex);
7754 /* always unregister in case we don't destroy any domains */
7755 unregister_sched_domain_sysctl();
7757 /* Let architecture update cpu core mappings. */
7758 new_topology = arch_update_cpu_topology();
7760 n = doms_new ? ndoms_new : 0;
7762 /* Destroy deleted domains */
7763 for (i = 0; i < ndoms_cur; i++) {
7764 for (j = 0; j < n && !new_topology; j++) {
7765 if (cpumask_equal(doms_cur[i], doms_new[j])
7766 && dattrs_equal(dattr_cur, i, dattr_new, j))
7769 /* no match - a current sched domain not in new doms_new[] */
7770 detach_destroy_domains(doms_cur[i]);
7775 if (doms_new == NULL) {
7777 doms_new = &fallback_doms;
7778 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7779 WARN_ON_ONCE(dattr_new);
7782 /* Build new domains */
7783 for (i = 0; i < ndoms_new; i++) {
7784 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7785 if (cpumask_equal(doms_new[i], doms_cur[j])
7786 && dattrs_equal(dattr_new, i, dattr_cur, j))
7789 /* no match - add a new doms_new */
7790 __build_sched_domains(doms_new[i],
7791 dattr_new ? dattr_new + i : NULL);
7796 /* Remember the new sched domains */
7797 if (doms_cur != &fallback_doms)
7798 free_sched_domains(doms_cur, ndoms_cur);
7799 kfree(dattr_cur); /* kfree(NULL) is safe */
7800 doms_cur = doms_new;
7801 dattr_cur = dattr_new;
7802 ndoms_cur = ndoms_new;
7804 register_sched_domain_sysctl();
7806 mutex_unlock(&sched_domains_mutex);
7809 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7810 static void arch_reinit_sched_domains(void)
7814 /* Destroy domains first to force the rebuild */
7815 partition_sched_domains(0, NULL, NULL);
7817 rebuild_sched_domains();
7821 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7823 unsigned int level = 0;
7825 if (sscanf(buf, "%u", &level) != 1)
7829 * level is always be positive so don't check for
7830 * level < POWERSAVINGS_BALANCE_NONE which is 0
7831 * What happens on 0 or 1 byte write,
7832 * need to check for count as well?
7835 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7839 sched_smt_power_savings = level;
7841 sched_mc_power_savings = level;
7843 arch_reinit_sched_domains();
7848 #ifdef CONFIG_SCHED_MC
7849 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7850 struct sysdev_class_attribute *attr,
7853 return sprintf(page, "%u\n", sched_mc_power_savings);
7855 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7856 struct sysdev_class_attribute *attr,
7857 const char *buf, size_t count)
7859 return sched_power_savings_store(buf, count, 0);
7861 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7862 sched_mc_power_savings_show,
7863 sched_mc_power_savings_store);
7866 #ifdef CONFIG_SCHED_SMT
7867 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7868 struct sysdev_class_attribute *attr,
7871 return sprintf(page, "%u\n", sched_smt_power_savings);
7873 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7874 struct sysdev_class_attribute *attr,
7875 const char *buf, size_t count)
7877 return sched_power_savings_store(buf, count, 1);
7879 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7880 sched_smt_power_savings_show,
7881 sched_smt_power_savings_store);
7884 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7888 #ifdef CONFIG_SCHED_SMT
7890 err = sysfs_create_file(&cls->kset.kobj,
7891 &attr_sched_smt_power_savings.attr);
7893 #ifdef CONFIG_SCHED_MC
7894 if (!err && mc_capable())
7895 err = sysfs_create_file(&cls->kset.kobj,
7896 &attr_sched_mc_power_savings.attr);
7900 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7903 * Update cpusets according to cpu_active mask. If cpusets are
7904 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7905 * around partition_sched_domains().
7907 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7910 switch (action & ~CPU_TASKS_FROZEN) {
7912 case CPU_DOWN_FAILED:
7913 cpuset_update_active_cpus();
7920 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7923 switch (action & ~CPU_TASKS_FROZEN) {
7924 case CPU_DOWN_PREPARE:
7925 cpuset_update_active_cpus();
7932 static int update_runtime(struct notifier_block *nfb,
7933 unsigned long action, void *hcpu)
7935 int cpu = (int)(long)hcpu;
7938 case CPU_DOWN_PREPARE:
7939 case CPU_DOWN_PREPARE_FROZEN:
7940 disable_runtime(cpu_rq(cpu));
7943 case CPU_DOWN_FAILED:
7944 case CPU_DOWN_FAILED_FROZEN:
7946 case CPU_ONLINE_FROZEN:
7947 enable_runtime(cpu_rq(cpu));
7955 void __init sched_init_smp(void)
7957 cpumask_var_t non_isolated_cpus;
7959 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7960 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7962 #if defined(CONFIG_NUMA)
7963 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7965 BUG_ON(sched_group_nodes_bycpu == NULL);
7968 mutex_lock(&sched_domains_mutex);
7969 arch_init_sched_domains(cpu_active_mask);
7970 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7971 if (cpumask_empty(non_isolated_cpus))
7972 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7973 mutex_unlock(&sched_domains_mutex);
7976 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7977 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7979 /* RT runtime code needs to handle some hotplug events */
7980 hotcpu_notifier(update_runtime, 0);
7984 /* Move init over to a non-isolated CPU */
7985 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7987 sched_init_granularity();
7988 free_cpumask_var(non_isolated_cpus);
7990 init_sched_rt_class();
7993 void __init sched_init_smp(void)
7995 sched_init_granularity();
7997 #endif /* CONFIG_SMP */
7999 const_debug unsigned int sysctl_timer_migration = 1;
8001 int in_sched_functions(unsigned long addr)
8003 return in_lock_functions(addr) ||
8004 (addr >= (unsigned long)__sched_text_start
8005 && addr < (unsigned long)__sched_text_end);
8008 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8010 cfs_rq->tasks_timeline = RB_ROOT;
8011 INIT_LIST_HEAD(&cfs_rq->tasks);
8012 #ifdef CONFIG_FAIR_GROUP_SCHED
8015 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8018 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8020 struct rt_prio_array *array;
8023 array = &rt_rq->active;
8024 for (i = 0; i < MAX_RT_PRIO; i++) {
8025 INIT_LIST_HEAD(array->queue + i);
8026 __clear_bit(i, array->bitmap);
8028 /* delimiter for bitsearch: */
8029 __set_bit(MAX_RT_PRIO, array->bitmap);
8031 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8032 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8034 rt_rq->highest_prio.next = MAX_RT_PRIO;
8038 rt_rq->rt_nr_migratory = 0;
8039 rt_rq->overloaded = 0;
8040 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
8044 rt_rq->rt_throttled = 0;
8045 rt_rq->rt_runtime = 0;
8046 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
8048 #ifdef CONFIG_RT_GROUP_SCHED
8049 rt_rq->rt_nr_boosted = 0;
8054 #ifdef CONFIG_FAIR_GROUP_SCHED
8055 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8056 struct sched_entity *se, int cpu, int add,
8057 struct sched_entity *parent)
8059 struct rq *rq = cpu_rq(cpu);
8060 tg->cfs_rq[cpu] = cfs_rq;
8061 init_cfs_rq(cfs_rq, rq);
8064 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8067 /* se could be NULL for init_task_group */
8072 se->cfs_rq = &rq->cfs;
8074 se->cfs_rq = parent->my_q;
8077 se->load.weight = tg->shares;
8078 se->load.inv_weight = 0;
8079 se->parent = parent;
8083 #ifdef CONFIG_RT_GROUP_SCHED
8084 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8085 struct sched_rt_entity *rt_se, int cpu, int add,
8086 struct sched_rt_entity *parent)
8088 struct rq *rq = cpu_rq(cpu);
8090 tg->rt_rq[cpu] = rt_rq;
8091 init_rt_rq(rt_rq, rq);
8093 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8095 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8097 tg->rt_se[cpu] = rt_se;
8102 rt_se->rt_rq = &rq->rt;
8104 rt_se->rt_rq = parent->my_q;
8106 rt_se->my_q = rt_rq;
8107 rt_se->parent = parent;
8108 INIT_LIST_HEAD(&rt_se->run_list);
8112 void __init sched_init(void)
8115 unsigned long alloc_size = 0, ptr;
8117 #ifdef CONFIG_FAIR_GROUP_SCHED
8118 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8120 #ifdef CONFIG_RT_GROUP_SCHED
8121 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8123 #ifdef CONFIG_CPUMASK_OFFSTACK
8124 alloc_size += num_possible_cpus() * cpumask_size();
8127 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8129 #ifdef CONFIG_FAIR_GROUP_SCHED
8130 init_task_group.se = (struct sched_entity **)ptr;
8131 ptr += nr_cpu_ids * sizeof(void **);
8133 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8134 ptr += nr_cpu_ids * sizeof(void **);
8136 #endif /* CONFIG_FAIR_GROUP_SCHED */
8137 #ifdef CONFIG_RT_GROUP_SCHED
8138 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8139 ptr += nr_cpu_ids * sizeof(void **);
8141 init_task_group.rt_rq = (struct rt_rq **)ptr;
8142 ptr += nr_cpu_ids * sizeof(void **);
8144 #endif /* CONFIG_RT_GROUP_SCHED */
8145 #ifdef CONFIG_CPUMASK_OFFSTACK
8146 for_each_possible_cpu(i) {
8147 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8148 ptr += cpumask_size();
8150 #endif /* CONFIG_CPUMASK_OFFSTACK */
8154 init_defrootdomain();
8157 init_rt_bandwidth(&def_rt_bandwidth,
8158 global_rt_period(), global_rt_runtime());
8160 #ifdef CONFIG_RT_GROUP_SCHED
8161 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8162 global_rt_period(), global_rt_runtime());
8163 #endif /* CONFIG_RT_GROUP_SCHED */
8165 #ifdef CONFIG_CGROUP_SCHED
8166 list_add(&init_task_group.list, &task_groups);
8167 INIT_LIST_HEAD(&init_task_group.children);
8169 #endif /* CONFIG_CGROUP_SCHED */
8171 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
8172 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
8173 __alignof__(unsigned long));
8175 for_each_possible_cpu(i) {
8179 raw_spin_lock_init(&rq->lock);
8181 rq->calc_load_active = 0;
8182 rq->calc_load_update = jiffies + LOAD_FREQ;
8183 init_cfs_rq(&rq->cfs, rq);
8184 init_rt_rq(&rq->rt, rq);
8185 #ifdef CONFIG_FAIR_GROUP_SCHED
8186 init_task_group.shares = init_task_group_load;
8187 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8188 #ifdef CONFIG_CGROUP_SCHED
8190 * How much cpu bandwidth does init_task_group get?
8192 * In case of task-groups formed thr' the cgroup filesystem, it
8193 * gets 100% of the cpu resources in the system. This overall
8194 * system cpu resource is divided among the tasks of
8195 * init_task_group and its child task-groups in a fair manner,
8196 * based on each entity's (task or task-group's) weight
8197 * (se->load.weight).
8199 * In other words, if init_task_group has 10 tasks of weight
8200 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8201 * then A0's share of the cpu resource is:
8203 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8205 * We achieve this by letting init_task_group's tasks sit
8206 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8208 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8210 #endif /* CONFIG_FAIR_GROUP_SCHED */
8212 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8213 #ifdef CONFIG_RT_GROUP_SCHED
8214 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8215 #ifdef CONFIG_CGROUP_SCHED
8216 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8220 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8221 rq->cpu_load[j] = 0;
8223 rq->last_load_update_tick = jiffies;
8228 rq->cpu_power = SCHED_LOAD_SCALE;
8229 rq->post_schedule = 0;
8230 rq->active_balance = 0;
8231 rq->next_balance = jiffies;
8236 rq->avg_idle = 2*sysctl_sched_migration_cost;
8237 rq_attach_root(rq, &def_root_domain);
8239 rq->nohz_balance_kick = 0;
8240 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8244 atomic_set(&rq->nr_iowait, 0);
8247 set_load_weight(&init_task);
8249 #ifdef CONFIG_PREEMPT_NOTIFIERS
8250 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8254 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8257 #ifdef CONFIG_RT_MUTEXES
8258 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
8262 * The boot idle thread does lazy MMU switching as well:
8264 atomic_inc(&init_mm.mm_count);
8265 enter_lazy_tlb(&init_mm, current);
8268 * Make us the idle thread. Technically, schedule() should not be
8269 * called from this thread, however somewhere below it might be,
8270 * but because we are the idle thread, we just pick up running again
8271 * when this runqueue becomes "idle".
8273 init_idle(current, smp_processor_id());
8275 calc_load_update = jiffies + LOAD_FREQ;
8278 * During early bootup we pretend to be a normal task:
8280 current->sched_class = &fair_sched_class;
8282 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8283 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8286 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8287 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8288 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8289 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8290 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8292 /* May be allocated at isolcpus cmdline parse time */
8293 if (cpu_isolated_map == NULL)
8294 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8299 scheduler_running = 1;
8302 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8303 static inline int preempt_count_equals(int preempt_offset)
8305 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8307 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
8310 void __might_sleep(const char *file, int line, int preempt_offset)
8313 static unsigned long prev_jiffy; /* ratelimiting */
8315 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8316 system_state != SYSTEM_RUNNING || oops_in_progress)
8318 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8320 prev_jiffy = jiffies;
8323 "BUG: sleeping function called from invalid context at %s:%d\n",
8326 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8327 in_atomic(), irqs_disabled(),
8328 current->pid, current->comm);
8330 debug_show_held_locks(current);
8331 if (irqs_disabled())
8332 print_irqtrace_events(current);
8336 EXPORT_SYMBOL(__might_sleep);
8339 #ifdef CONFIG_MAGIC_SYSRQ
8340 static void normalize_task(struct rq *rq, struct task_struct *p)
8344 on_rq = p->se.on_rq;
8346 deactivate_task(rq, p, 0);
8347 __setscheduler(rq, p, SCHED_NORMAL, 0);
8349 activate_task(rq, p, 0);
8350 resched_task(rq->curr);
8354 void normalize_rt_tasks(void)
8356 struct task_struct *g, *p;
8357 unsigned long flags;
8360 read_lock_irqsave(&tasklist_lock, flags);
8361 do_each_thread(g, p) {
8363 * Only normalize user tasks:
8368 p->se.exec_start = 0;
8369 #ifdef CONFIG_SCHEDSTATS
8370 p->se.statistics.wait_start = 0;
8371 p->se.statistics.sleep_start = 0;
8372 p->se.statistics.block_start = 0;
8377 * Renice negative nice level userspace
8380 if (TASK_NICE(p) < 0 && p->mm)
8381 set_user_nice(p, 0);
8385 raw_spin_lock(&p->pi_lock);
8386 rq = __task_rq_lock(p);
8388 normalize_task(rq, p);
8390 __task_rq_unlock(rq);
8391 raw_spin_unlock(&p->pi_lock);
8392 } while_each_thread(g, p);
8394 read_unlock_irqrestore(&tasklist_lock, flags);
8397 #endif /* CONFIG_MAGIC_SYSRQ */
8399 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8401 * These functions are only useful for the IA64 MCA handling, or kdb.
8403 * They can only be called when the whole system has been
8404 * stopped - every CPU needs to be quiescent, and no scheduling
8405 * activity can take place. Using them for anything else would
8406 * be a serious bug, and as a result, they aren't even visible
8407 * under any other configuration.
8411 * curr_task - return the current task for a given cpu.
8412 * @cpu: the processor in question.
8414 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8416 struct task_struct *curr_task(int cpu)
8418 return cpu_curr(cpu);
8421 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8425 * set_curr_task - set the current task for a given cpu.
8426 * @cpu: the processor in question.
8427 * @p: the task pointer to set.
8429 * Description: This function must only be used when non-maskable interrupts
8430 * are serviced on a separate stack. It allows the architecture to switch the
8431 * notion of the current task on a cpu in a non-blocking manner. This function
8432 * must be called with all CPU's synchronized, and interrupts disabled, the
8433 * and caller must save the original value of the current task (see
8434 * curr_task() above) and restore that value before reenabling interrupts and
8435 * re-starting the system.
8437 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8439 void set_curr_task(int cpu, struct task_struct *p)
8446 #ifdef CONFIG_FAIR_GROUP_SCHED
8447 static void free_fair_sched_group(struct task_group *tg)
8451 for_each_possible_cpu(i) {
8453 kfree(tg->cfs_rq[i]);
8463 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8465 struct cfs_rq *cfs_rq;
8466 struct sched_entity *se;
8470 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8473 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8477 tg->shares = NICE_0_LOAD;
8479 for_each_possible_cpu(i) {
8482 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8483 GFP_KERNEL, cpu_to_node(i));
8487 se = kzalloc_node(sizeof(struct sched_entity),
8488 GFP_KERNEL, cpu_to_node(i));
8492 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8503 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8505 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8506 &cpu_rq(cpu)->leaf_cfs_rq_list);
8509 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8511 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8513 #else /* !CONFG_FAIR_GROUP_SCHED */
8514 static inline void free_fair_sched_group(struct task_group *tg)
8519 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8524 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8528 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8531 #endif /* CONFIG_FAIR_GROUP_SCHED */
8533 #ifdef CONFIG_RT_GROUP_SCHED
8534 static void free_rt_sched_group(struct task_group *tg)
8538 destroy_rt_bandwidth(&tg->rt_bandwidth);
8540 for_each_possible_cpu(i) {
8542 kfree(tg->rt_rq[i]);
8544 kfree(tg->rt_se[i]);
8552 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8554 struct rt_rq *rt_rq;
8555 struct sched_rt_entity *rt_se;
8559 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8562 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8566 init_rt_bandwidth(&tg->rt_bandwidth,
8567 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8569 for_each_possible_cpu(i) {
8572 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8573 GFP_KERNEL, cpu_to_node(i));
8577 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8578 GFP_KERNEL, cpu_to_node(i));
8582 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8593 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8595 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8596 &cpu_rq(cpu)->leaf_rt_rq_list);
8599 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8601 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8603 #else /* !CONFIG_RT_GROUP_SCHED */
8604 static inline void free_rt_sched_group(struct task_group *tg)
8609 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8614 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8618 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8621 #endif /* CONFIG_RT_GROUP_SCHED */
8623 #ifdef CONFIG_CGROUP_SCHED
8624 static void free_sched_group(struct task_group *tg)
8626 free_fair_sched_group(tg);
8627 free_rt_sched_group(tg);
8631 /* allocate runqueue etc for a new task group */
8632 struct task_group *sched_create_group(struct task_group *parent)
8634 struct task_group *tg;
8635 unsigned long flags;
8638 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8640 return ERR_PTR(-ENOMEM);
8642 if (!alloc_fair_sched_group(tg, parent))
8645 if (!alloc_rt_sched_group(tg, parent))
8648 spin_lock_irqsave(&task_group_lock, flags);
8649 for_each_possible_cpu(i) {
8650 register_fair_sched_group(tg, i);
8651 register_rt_sched_group(tg, i);
8653 list_add_rcu(&tg->list, &task_groups);
8655 WARN_ON(!parent); /* root should already exist */
8657 tg->parent = parent;
8658 INIT_LIST_HEAD(&tg->children);
8659 list_add_rcu(&tg->siblings, &parent->children);
8660 spin_unlock_irqrestore(&task_group_lock, flags);
8665 free_sched_group(tg);
8666 return ERR_PTR(-ENOMEM);
8669 /* rcu callback to free various structures associated with a task group */
8670 static void free_sched_group_rcu(struct rcu_head *rhp)
8672 /* now it should be safe to free those cfs_rqs */
8673 free_sched_group(container_of(rhp, struct task_group, rcu));
8676 /* Destroy runqueue etc associated with a task group */
8677 void sched_destroy_group(struct task_group *tg)
8679 unsigned long flags;
8682 spin_lock_irqsave(&task_group_lock, flags);
8683 for_each_possible_cpu(i) {
8684 unregister_fair_sched_group(tg, i);
8685 unregister_rt_sched_group(tg, i);
8687 list_del_rcu(&tg->list);
8688 list_del_rcu(&tg->siblings);
8689 spin_unlock_irqrestore(&task_group_lock, flags);
8691 /* wait for possible concurrent references to cfs_rqs complete */
8692 call_rcu(&tg->rcu, free_sched_group_rcu);
8695 /* change task's runqueue when it moves between groups.
8696 * The caller of this function should have put the task in its new group
8697 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8698 * reflect its new group.
8700 void sched_move_task(struct task_struct *tsk)
8703 unsigned long flags;
8706 rq = task_rq_lock(tsk, &flags);
8708 running = task_current(rq, tsk);
8709 on_rq = tsk->se.on_rq;
8712 dequeue_task(rq, tsk, 0);
8713 if (unlikely(running))
8714 tsk->sched_class->put_prev_task(rq, tsk);
8716 #ifdef CONFIG_FAIR_GROUP_SCHED
8717 if (tsk->sched_class->task_move_group)
8718 tsk->sched_class->task_move_group(tsk, on_rq);
8721 set_task_rq(tsk, task_cpu(tsk));
8723 if (unlikely(running))
8724 tsk->sched_class->set_curr_task(rq);
8726 enqueue_task(rq, tsk, 0);
8728 task_rq_unlock(rq, &flags);
8730 #endif /* CONFIG_CGROUP_SCHED */
8732 #ifdef CONFIG_FAIR_GROUP_SCHED
8733 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8735 struct cfs_rq *cfs_rq = se->cfs_rq;
8740 dequeue_entity(cfs_rq, se, 0);
8742 se->load.weight = shares;
8743 se->load.inv_weight = 0;
8746 enqueue_entity(cfs_rq, se, 0);
8749 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8751 struct cfs_rq *cfs_rq = se->cfs_rq;
8752 struct rq *rq = cfs_rq->rq;
8753 unsigned long flags;
8755 raw_spin_lock_irqsave(&rq->lock, flags);
8756 __set_se_shares(se, shares);
8757 raw_spin_unlock_irqrestore(&rq->lock, flags);
8760 static DEFINE_MUTEX(shares_mutex);
8762 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8765 unsigned long flags;
8768 * We can't change the weight of the root cgroup.
8773 if (shares < MIN_SHARES)
8774 shares = MIN_SHARES;
8775 else if (shares > MAX_SHARES)
8776 shares = MAX_SHARES;
8778 mutex_lock(&shares_mutex);
8779 if (tg->shares == shares)
8782 spin_lock_irqsave(&task_group_lock, flags);
8783 for_each_possible_cpu(i)
8784 unregister_fair_sched_group(tg, i);
8785 list_del_rcu(&tg->siblings);
8786 spin_unlock_irqrestore(&task_group_lock, flags);
8788 /* wait for any ongoing reference to this group to finish */
8789 synchronize_sched();
8792 * Now we are free to modify the group's share on each cpu
8793 * w/o tripping rebalance_share or load_balance_fair.
8795 tg->shares = shares;
8796 for_each_possible_cpu(i) {
8800 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8801 set_se_shares(tg->se[i], shares);
8805 * Enable load balance activity on this group, by inserting it back on
8806 * each cpu's rq->leaf_cfs_rq_list.
8808 spin_lock_irqsave(&task_group_lock, flags);
8809 for_each_possible_cpu(i)
8810 register_fair_sched_group(tg, i);
8811 list_add_rcu(&tg->siblings, &tg->parent->children);
8812 spin_unlock_irqrestore(&task_group_lock, flags);
8814 mutex_unlock(&shares_mutex);
8818 unsigned long sched_group_shares(struct task_group *tg)
8824 #ifdef CONFIG_RT_GROUP_SCHED
8826 * Ensure that the real time constraints are schedulable.
8828 static DEFINE_MUTEX(rt_constraints_mutex);
8830 static unsigned long to_ratio(u64 period, u64 runtime)
8832 if (runtime == RUNTIME_INF)
8835 return div64_u64(runtime << 20, period);
8838 /* Must be called with tasklist_lock held */
8839 static inline int tg_has_rt_tasks(struct task_group *tg)
8841 struct task_struct *g, *p;
8843 do_each_thread(g, p) {
8844 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8846 } while_each_thread(g, p);
8851 struct rt_schedulable_data {
8852 struct task_group *tg;
8857 static int tg_schedulable(struct task_group *tg, void *data)
8859 struct rt_schedulable_data *d = data;
8860 struct task_group *child;
8861 unsigned long total, sum = 0;
8862 u64 period, runtime;
8864 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8865 runtime = tg->rt_bandwidth.rt_runtime;
8868 period = d->rt_period;
8869 runtime = d->rt_runtime;
8873 * Cannot have more runtime than the period.
8875 if (runtime > period && runtime != RUNTIME_INF)
8879 * Ensure we don't starve existing RT tasks.
8881 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8884 total = to_ratio(period, runtime);
8887 * Nobody can have more than the global setting allows.
8889 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8893 * The sum of our children's runtime should not exceed our own.
8895 list_for_each_entry_rcu(child, &tg->children, siblings) {
8896 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8897 runtime = child->rt_bandwidth.rt_runtime;
8899 if (child == d->tg) {
8900 period = d->rt_period;
8901 runtime = d->rt_runtime;
8904 sum += to_ratio(period, runtime);
8913 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8915 struct rt_schedulable_data data = {
8917 .rt_period = period,
8918 .rt_runtime = runtime,
8921 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8924 static int tg_set_bandwidth(struct task_group *tg,
8925 u64 rt_period, u64 rt_runtime)
8929 mutex_lock(&rt_constraints_mutex);
8930 read_lock(&tasklist_lock);
8931 err = __rt_schedulable(tg, rt_period, rt_runtime);
8935 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8936 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8937 tg->rt_bandwidth.rt_runtime = rt_runtime;
8939 for_each_possible_cpu(i) {
8940 struct rt_rq *rt_rq = tg->rt_rq[i];
8942 raw_spin_lock(&rt_rq->rt_runtime_lock);
8943 rt_rq->rt_runtime = rt_runtime;
8944 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8946 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8948 read_unlock(&tasklist_lock);
8949 mutex_unlock(&rt_constraints_mutex);
8954 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8956 u64 rt_runtime, rt_period;
8958 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8959 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8960 if (rt_runtime_us < 0)
8961 rt_runtime = RUNTIME_INF;
8963 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8966 long sched_group_rt_runtime(struct task_group *tg)
8970 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8973 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8974 do_div(rt_runtime_us, NSEC_PER_USEC);
8975 return rt_runtime_us;
8978 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8980 u64 rt_runtime, rt_period;
8982 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8983 rt_runtime = tg->rt_bandwidth.rt_runtime;
8988 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8991 long sched_group_rt_period(struct task_group *tg)
8995 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8996 do_div(rt_period_us, NSEC_PER_USEC);
8997 return rt_period_us;
9000 static int sched_rt_global_constraints(void)
9002 u64 runtime, period;
9005 if (sysctl_sched_rt_period <= 0)
9008 runtime = global_rt_runtime();
9009 period = global_rt_period();
9012 * Sanity check on the sysctl variables.
9014 if (runtime > period && runtime != RUNTIME_INF)
9017 mutex_lock(&rt_constraints_mutex);
9018 read_lock(&tasklist_lock);
9019 ret = __rt_schedulable(NULL, 0, 0);
9020 read_unlock(&tasklist_lock);
9021 mutex_unlock(&rt_constraints_mutex);
9026 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
9028 /* Don't accept realtime tasks when there is no way for them to run */
9029 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
9035 #else /* !CONFIG_RT_GROUP_SCHED */
9036 static int sched_rt_global_constraints(void)
9038 unsigned long flags;
9041 if (sysctl_sched_rt_period <= 0)
9045 * There's always some RT tasks in the root group
9046 * -- migration, kstopmachine etc..
9048 if (sysctl_sched_rt_runtime == 0)
9051 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9052 for_each_possible_cpu(i) {
9053 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9055 raw_spin_lock(&rt_rq->rt_runtime_lock);
9056 rt_rq->rt_runtime = global_rt_runtime();
9057 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9059 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9063 #endif /* CONFIG_RT_GROUP_SCHED */
9065 int sched_rt_handler(struct ctl_table *table, int write,
9066 void __user *buffer, size_t *lenp,
9070 int old_period, old_runtime;
9071 static DEFINE_MUTEX(mutex);
9074 old_period = sysctl_sched_rt_period;
9075 old_runtime = sysctl_sched_rt_runtime;
9077 ret = proc_dointvec(table, write, buffer, lenp, ppos);
9079 if (!ret && write) {
9080 ret = sched_rt_global_constraints();
9082 sysctl_sched_rt_period = old_period;
9083 sysctl_sched_rt_runtime = old_runtime;
9085 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9086 def_rt_bandwidth.rt_period =
9087 ns_to_ktime(global_rt_period());
9090 mutex_unlock(&mutex);
9095 #ifdef CONFIG_CGROUP_SCHED
9097 /* return corresponding task_group object of a cgroup */
9098 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9100 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9101 struct task_group, css);
9104 static struct cgroup_subsys_state *
9105 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9107 struct task_group *tg, *parent;
9109 if (!cgrp->parent) {
9110 /* This is early initialization for the top cgroup */
9111 return &init_task_group.css;
9114 parent = cgroup_tg(cgrp->parent);
9115 tg = sched_create_group(parent);
9117 return ERR_PTR(-ENOMEM);
9123 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9125 struct task_group *tg = cgroup_tg(cgrp);
9127 sched_destroy_group(tg);
9131 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9133 #ifdef CONFIG_RT_GROUP_SCHED
9134 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9137 /* We don't support RT-tasks being in separate groups */
9138 if (tsk->sched_class != &fair_sched_class)
9145 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9146 struct task_struct *tsk, bool threadgroup)
9148 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
9152 struct task_struct *c;
9154 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
9155 retval = cpu_cgroup_can_attach_task(cgrp, c);
9167 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9168 struct cgroup *old_cont, struct task_struct *tsk,
9171 sched_move_task(tsk);
9173 struct task_struct *c;
9175 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
9182 #ifdef CONFIG_FAIR_GROUP_SCHED
9183 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9186 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9189 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9191 struct task_group *tg = cgroup_tg(cgrp);
9193 return (u64) tg->shares;
9195 #endif /* CONFIG_FAIR_GROUP_SCHED */
9197 #ifdef CONFIG_RT_GROUP_SCHED
9198 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9201 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9204 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9206 return sched_group_rt_runtime(cgroup_tg(cgrp));
9209 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9212 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9215 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9217 return sched_group_rt_period(cgroup_tg(cgrp));
9219 #endif /* CONFIG_RT_GROUP_SCHED */
9221 static struct cftype cpu_files[] = {
9222 #ifdef CONFIG_FAIR_GROUP_SCHED
9225 .read_u64 = cpu_shares_read_u64,
9226 .write_u64 = cpu_shares_write_u64,
9229 #ifdef CONFIG_RT_GROUP_SCHED
9231 .name = "rt_runtime_us",
9232 .read_s64 = cpu_rt_runtime_read,
9233 .write_s64 = cpu_rt_runtime_write,
9236 .name = "rt_period_us",
9237 .read_u64 = cpu_rt_period_read_uint,
9238 .write_u64 = cpu_rt_period_write_uint,
9243 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9245 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9248 struct cgroup_subsys cpu_cgroup_subsys = {
9250 .create = cpu_cgroup_create,
9251 .destroy = cpu_cgroup_destroy,
9252 .can_attach = cpu_cgroup_can_attach,
9253 .attach = cpu_cgroup_attach,
9254 .populate = cpu_cgroup_populate,
9255 .subsys_id = cpu_cgroup_subsys_id,
9259 #endif /* CONFIG_CGROUP_SCHED */
9261 #ifdef CONFIG_CGROUP_CPUACCT
9264 * CPU accounting code for task groups.
9266 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9267 * (balbir@in.ibm.com).
9270 /* track cpu usage of a group of tasks and its child groups */
9272 struct cgroup_subsys_state css;
9273 /* cpuusage holds pointer to a u64-type object on every cpu */
9274 u64 __percpu *cpuusage;
9275 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9276 struct cpuacct *parent;
9279 struct cgroup_subsys cpuacct_subsys;
9281 /* return cpu accounting group corresponding to this container */
9282 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9284 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9285 struct cpuacct, css);
9288 /* return cpu accounting group to which this task belongs */
9289 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9291 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9292 struct cpuacct, css);
9295 /* create a new cpu accounting group */
9296 static struct cgroup_subsys_state *cpuacct_create(
9297 struct cgroup_subsys *ss, struct cgroup *cgrp)
9299 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9305 ca->cpuusage = alloc_percpu(u64);
9309 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9310 if (percpu_counter_init(&ca->cpustat[i], 0))
9311 goto out_free_counters;
9314 ca->parent = cgroup_ca(cgrp->parent);
9320 percpu_counter_destroy(&ca->cpustat[i]);
9321 free_percpu(ca->cpuusage);
9325 return ERR_PTR(-ENOMEM);
9328 /* destroy an existing cpu accounting group */
9330 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9332 struct cpuacct *ca = cgroup_ca(cgrp);
9335 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9336 percpu_counter_destroy(&ca->cpustat[i]);
9337 free_percpu(ca->cpuusage);
9341 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9343 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9346 #ifndef CONFIG_64BIT
9348 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9350 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9352 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9360 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9362 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9364 #ifndef CONFIG_64BIT
9366 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9368 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9370 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9376 /* return total cpu usage (in nanoseconds) of a group */
9377 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9379 struct cpuacct *ca = cgroup_ca(cgrp);
9380 u64 totalcpuusage = 0;
9383 for_each_present_cpu(i)
9384 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9386 return totalcpuusage;
9389 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9392 struct cpuacct *ca = cgroup_ca(cgrp);
9401 for_each_present_cpu(i)
9402 cpuacct_cpuusage_write(ca, i, 0);
9408 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9411 struct cpuacct *ca = cgroup_ca(cgroup);
9415 for_each_present_cpu(i) {
9416 percpu = cpuacct_cpuusage_read(ca, i);
9417 seq_printf(m, "%llu ", (unsigned long long) percpu);
9419 seq_printf(m, "\n");
9423 static const char *cpuacct_stat_desc[] = {
9424 [CPUACCT_STAT_USER] = "user",
9425 [CPUACCT_STAT_SYSTEM] = "system",
9428 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9429 struct cgroup_map_cb *cb)
9431 struct cpuacct *ca = cgroup_ca(cgrp);
9434 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9435 s64 val = percpu_counter_read(&ca->cpustat[i]);
9436 val = cputime64_to_clock_t(val);
9437 cb->fill(cb, cpuacct_stat_desc[i], val);
9442 static struct cftype files[] = {
9445 .read_u64 = cpuusage_read,
9446 .write_u64 = cpuusage_write,
9449 .name = "usage_percpu",
9450 .read_seq_string = cpuacct_percpu_seq_read,
9454 .read_map = cpuacct_stats_show,
9458 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9460 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9464 * charge this task's execution time to its accounting group.
9466 * called with rq->lock held.
9468 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9473 if (unlikely(!cpuacct_subsys.active))
9476 cpu = task_cpu(tsk);
9482 for (; ca; ca = ca->parent) {
9483 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9484 *cpuusage += cputime;
9491 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9492 * in cputime_t units. As a result, cpuacct_update_stats calls
9493 * percpu_counter_add with values large enough to always overflow the
9494 * per cpu batch limit causing bad SMP scalability.
9496 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9497 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9498 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9501 #define CPUACCT_BATCH \
9502 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9504 #define CPUACCT_BATCH 0
9508 * Charge the system/user time to the task's accounting group.
9510 static void cpuacct_update_stats(struct task_struct *tsk,
9511 enum cpuacct_stat_index idx, cputime_t val)
9514 int batch = CPUACCT_BATCH;
9516 if (unlikely(!cpuacct_subsys.active))
9523 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9529 struct cgroup_subsys cpuacct_subsys = {
9531 .create = cpuacct_create,
9532 .destroy = cpuacct_destroy,
9533 .populate = cpuacct_populate,
9534 .subsys_id = cpuacct_subsys_id,
9536 #endif /* CONFIG_CGROUP_CPUACCT */
9540 void synchronize_sched_expedited(void)
9544 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9546 #else /* #ifndef CONFIG_SMP */
9548 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
9550 static int synchronize_sched_expedited_cpu_stop(void *data)
9553 * There must be a full memory barrier on each affected CPU
9554 * between the time that try_stop_cpus() is called and the
9555 * time that it returns.
9557 * In the current initial implementation of cpu_stop, the
9558 * above condition is already met when the control reaches
9559 * this point and the following smp_mb() is not strictly
9560 * necessary. Do smp_mb() anyway for documentation and
9561 * robustness against future implementation changes.
9563 smp_mb(); /* See above comment block. */
9568 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9569 * approach to force grace period to end quickly. This consumes
9570 * significant time on all CPUs, and is thus not recommended for
9571 * any sort of common-case code.
9573 * Note that it is illegal to call this function while holding any
9574 * lock that is acquired by a CPU-hotplug notifier. Failing to
9575 * observe this restriction will result in deadlock.
9577 void synchronize_sched_expedited(void)
9579 int snap, trycount = 0;
9581 smp_mb(); /* ensure prior mod happens before capturing snap. */
9582 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
9584 while (try_stop_cpus(cpu_online_mask,
9585 synchronize_sched_expedited_cpu_stop,
9588 if (trycount++ < 10)
9589 udelay(trycount * num_online_cpus());
9591 synchronize_sched();
9594 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
9595 smp_mb(); /* ensure test happens before caller kfree */
9600 atomic_inc(&synchronize_sched_expedited_count);
9601 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9604 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9606 #endif /* #else #ifndef CONFIG_SMP */