4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/stop_machine.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/debugfs.h>
71 #include <linux/ctype.h>
72 #include <linux/ftrace.h>
73 #include <linux/slab.h>
74 #include <linux/cpuacct.h>
77 #include <asm/irq_regs.h>
78 #include <asm/mutex.h>
80 #include "sched_cpupri.h"
81 #include "workqueue_sched.h"
82 #include "sched_autogroup.h"
84 #define CREATE_TRACE_POINTS
85 #include <trace/events/sched.h>
88 * Convert user-nice values [ -20 ... 0 ... 19 ]
89 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
92 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
93 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
94 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
97 * 'User priority' is the nice value converted to something we
98 * can work with better when scaling various scheduler parameters,
99 * it's a [ 0 ... 39 ] range.
101 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
102 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
103 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
106 * Helpers for converting nanosecond timing to jiffy resolution
108 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
110 #define NICE_0_LOAD SCHED_LOAD_SCALE
111 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
114 * These are the 'tuning knobs' of the scheduler:
116 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
117 * Timeslices get refilled after they expire.
119 #define DEF_TIMESLICE (100 * HZ / 1000)
122 * single value that denotes runtime == period, ie unlimited time.
124 #define RUNTIME_INF ((u64)~0ULL)
126 static inline int rt_policy(int policy)
128 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
133 static inline int task_has_rt_policy(struct task_struct *p)
135 return rt_policy(p->policy);
139 * This is the priority-queue data structure of the RT scheduling class:
141 struct rt_prio_array {
142 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
143 struct list_head queue[MAX_RT_PRIO];
146 struct rt_bandwidth {
147 /* nests inside the rq lock: */
148 raw_spinlock_t rt_runtime_lock;
151 struct hrtimer rt_period_timer;
154 static struct rt_bandwidth def_rt_bandwidth;
156 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
158 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
160 struct rt_bandwidth *rt_b =
161 container_of(timer, struct rt_bandwidth, rt_period_timer);
167 now = hrtimer_cb_get_time(timer);
168 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
173 idle = do_sched_rt_period_timer(rt_b, overrun);
176 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
180 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
182 rt_b->rt_period = ns_to_ktime(period);
183 rt_b->rt_runtime = runtime;
185 raw_spin_lock_init(&rt_b->rt_runtime_lock);
187 hrtimer_init(&rt_b->rt_period_timer,
188 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
189 rt_b->rt_period_timer.function = sched_rt_period_timer;
192 static inline int rt_bandwidth_enabled(void)
194 return sysctl_sched_rt_runtime >= 0;
197 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
201 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
204 if (hrtimer_active(&rt_b->rt_period_timer))
207 raw_spin_lock(&rt_b->rt_runtime_lock);
212 if (hrtimer_active(&rt_b->rt_period_timer))
215 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
216 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
218 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
219 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
220 delta = ktime_to_ns(ktime_sub(hard, soft));
221 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
222 HRTIMER_MODE_ABS_PINNED, 0);
224 raw_spin_unlock(&rt_b->rt_runtime_lock);
227 #ifdef CONFIG_RT_GROUP_SCHED
228 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
230 hrtimer_cancel(&rt_b->rt_period_timer);
235 * sched_domains_mutex serializes calls to init_sched_domains,
236 * detach_destroy_domains and partition_sched_domains.
238 static DEFINE_MUTEX(sched_domains_mutex);
240 #ifdef CONFIG_CGROUP_SCHED
242 #include <linux/cgroup.h>
246 static LIST_HEAD(task_groups);
248 /* task group related information */
250 struct cgroup_subsys_state css;
252 #ifdef CONFIG_FAIR_GROUP_SCHED
253 /* schedulable entities of this group on each cpu */
254 struct sched_entity **se;
255 /* runqueue "owned" by this group on each cpu */
256 struct cfs_rq **cfs_rq;
257 unsigned long shares;
259 atomic_t load_weight;
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity **rt_se;
264 struct rt_rq **rt_rq;
266 struct rt_bandwidth rt_bandwidth;
270 struct list_head list;
272 struct task_group *parent;
273 struct list_head siblings;
274 struct list_head children;
276 #ifdef CONFIG_SCHED_AUTOGROUP
277 struct autogroup *autogroup;
281 /* task_group_lock serializes the addition/removal of task groups */
282 static DEFINE_SPINLOCK(task_group_lock);
284 #ifdef CONFIG_FAIR_GROUP_SCHED
286 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
289 * A weight of 0 or 1 can cause arithmetics problems.
290 * A weight of a cfs_rq is the sum of weights of which entities
291 * are queued on this cfs_rq, so a weight of a entity should not be
292 * too large, so as the shares value of a task group.
293 * (The default weight is 1024 - so there's no practical
294 * limitation from this.)
296 #define MIN_SHARES (1UL << 1)
297 #define MAX_SHARES (1UL << 18)
299 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
302 /* Default task group.
303 * Every task in system belong to this group at bootup.
305 struct task_group root_task_group;
307 #endif /* CONFIG_CGROUP_SCHED */
309 /* CFS-related fields in a runqueue */
311 struct load_weight load;
312 unsigned long nr_running;
317 u64 min_vruntime_copy;
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, *skip;
332 #ifdef CONFIG_SCHED_DEBUG
333 unsigned int nr_spread_over;
336 #ifdef CONFIG_FAIR_GROUP_SCHED
337 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
340 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
341 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
342 * (like users, containers etc.)
344 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
345 * list is used during load balance.
348 struct list_head leaf_cfs_rq_list;
349 struct task_group *tg; /* group that "owns" this runqueue */
353 * the part of load.weight contributed by tasks
355 unsigned long task_weight;
358 * h_load = weight * f(tg)
360 * Where f(tg) is the recursive weight fraction assigned to
363 unsigned long h_load;
366 * Maintaining per-cpu shares distribution for group scheduling
368 * load_stamp is the last time we updated the load average
369 * load_last is the last time we updated the load average and saw load
370 * load_unacc_exec_time is currently unaccounted execution time
374 u64 load_stamp, load_last, load_unacc_exec_time;
376 unsigned long load_contribution;
381 /* Real-Time classes' related field in a runqueue: */
383 struct rt_prio_array active;
384 unsigned long rt_nr_running;
385 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
387 int curr; /* highest queued rt task prio */
389 int next; /* next highest */
394 unsigned long rt_nr_migratory;
395 unsigned long rt_nr_total;
397 struct plist_head pushable_tasks;
402 /* Nests inside the rq lock: */
403 raw_spinlock_t rt_runtime_lock;
405 #ifdef CONFIG_RT_GROUP_SCHED
406 unsigned long rt_nr_boosted;
409 struct list_head leaf_rt_rq_list;
410 struct task_group *tg;
417 * We add the notion of a root-domain which will be used to define per-domain
418 * variables. Each exclusive cpuset essentially defines an island domain by
419 * fully partitioning the member cpus from any other cpuset. Whenever a new
420 * exclusive cpuset is created, we also create and attach a new root-domain
428 cpumask_var_t online;
431 * The "RT overload" flag: it gets set if a CPU has more than
432 * one runnable RT task.
434 cpumask_var_t rto_mask;
436 struct cpupri cpupri;
440 * By default the system creates a single root-domain with all cpus as
441 * members (mimicking the global state we have today).
443 static struct root_domain def_root_domain;
445 #endif /* CONFIG_SMP */
448 * This is the main, per-CPU runqueue data structure.
450 * Locking rule: those places that want to lock multiple runqueues
451 * (such as the load balancing or the thread migration code), lock
452 * acquire operations must be ordered by ascending &runqueue.
459 * nr_running and cpu_load should be in the same cacheline because
460 * remote CPUs use both these fields when doing load calculation.
462 unsigned long nr_running;
463 #define CPU_LOAD_IDX_MAX 5
464 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
465 unsigned long last_load_update_tick;
468 unsigned char nohz_balance_kick;
470 int skip_clock_update;
472 /* capture load from *all* tasks on this cpu: */
473 struct load_weight load;
474 unsigned long nr_load_updates;
480 #ifdef CONFIG_FAIR_GROUP_SCHED
481 /* list of leaf cfs_rq on this cpu: */
482 struct list_head leaf_cfs_rq_list;
484 #ifdef CONFIG_RT_GROUP_SCHED
485 struct list_head leaf_rt_rq_list;
489 * This is part of a global counter where only the total sum
490 * over all CPUs matters. A task can increase this counter on
491 * one CPU and if it got migrated afterwards it may decrease
492 * it on another CPU. Always updated under the runqueue lock:
494 unsigned long nr_uninterruptible;
496 struct task_struct *curr, *idle, *stop;
497 unsigned long next_balance;
498 struct mm_struct *prev_mm;
506 struct root_domain *rd;
507 struct sched_domain *sd;
509 unsigned long cpu_power;
511 unsigned char idle_at_tick;
512 /* For active balancing */
516 struct cpu_stop_work active_balance_work;
517 /* cpu of this runqueue: */
521 unsigned long avg_load_per_task;
529 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
533 /* calc_load related fields */
534 unsigned long calc_load_update;
535 long calc_load_active;
537 #ifdef CONFIG_SCHED_HRTICK
539 int hrtick_csd_pending;
540 struct call_single_data hrtick_csd;
542 struct hrtimer hrtick_timer;
545 #ifdef CONFIG_SCHEDSTATS
547 struct sched_info rq_sched_info;
548 unsigned long long rq_cpu_time;
549 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
551 /* sys_sched_yield() stats */
552 unsigned int yld_count;
554 /* schedule() stats */
555 unsigned int sched_switch;
556 unsigned int sched_count;
557 unsigned int sched_goidle;
559 /* try_to_wake_up() stats */
560 unsigned int ttwu_count;
561 unsigned int ttwu_local;
565 struct task_struct *wake_list;
569 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
572 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
574 static inline int cpu_of(struct rq *rq)
583 #define rcu_dereference_check_sched_domain(p) \
584 rcu_dereference_check((p), \
585 rcu_read_lock_held() || \
586 lockdep_is_held(&sched_domains_mutex))
589 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
590 * See detach_destroy_domains: synchronize_sched for details.
592 * The domain tree of any CPU may only be accessed from within
593 * preempt-disabled sections.
595 #define for_each_domain(cpu, __sd) \
596 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
598 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
599 #define this_rq() (&__get_cpu_var(runqueues))
600 #define task_rq(p) cpu_rq(task_cpu(p))
601 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
602 #define raw_rq() (&__raw_get_cpu_var(runqueues))
604 #ifdef CONFIG_CGROUP_SCHED
607 * Return the group to which this tasks belongs.
609 * We cannot use task_subsys_state() and friends because the cgroup
610 * subsystem changes that value before the cgroup_subsys::attach() method
611 * is called, therefore we cannot pin it and might observe the wrong value.
613 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
614 * core changes this before calling sched_move_task().
616 * Instead we use a 'copy' which is updated from sched_move_task() while
617 * holding both task_struct::pi_lock and rq::lock.
619 static inline struct task_group *task_group(struct task_struct *p)
621 return p->sched_task_group;
624 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
625 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
627 #ifdef CONFIG_FAIR_GROUP_SCHED
628 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
629 p->se.parent = task_group(p)->se[cpu];
632 #ifdef CONFIG_RT_GROUP_SCHED
633 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
634 p->rt.parent = task_group(p)->rt_se[cpu];
638 #else /* CONFIG_CGROUP_SCHED */
640 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
641 static inline struct task_group *task_group(struct task_struct *p)
646 #endif /* CONFIG_CGROUP_SCHED */
648 static void update_rq_clock_task(struct rq *rq, s64 delta);
650 static void update_rq_clock(struct rq *rq)
654 if (rq->skip_clock_update > 0)
657 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
659 update_rq_clock_task(rq, delta);
663 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
665 #ifdef CONFIG_SCHED_DEBUG
666 # define const_debug __read_mostly
668 # define const_debug static const
672 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
673 * @cpu: the processor in question.
675 * This interface allows printk to be called with the runqueue lock
676 * held and know whether or not it is OK to wake up the klogd.
678 int runqueue_is_locked(int cpu)
680 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
684 * Debugging: various feature bits
687 #define SCHED_FEAT(name, enabled) \
688 __SCHED_FEAT_##name ,
691 #include "sched_features.h"
696 #define SCHED_FEAT(name, enabled) \
697 (1UL << __SCHED_FEAT_##name) * enabled |
699 const_debug unsigned int sysctl_sched_features =
700 #include "sched_features.h"
705 #ifdef CONFIG_SCHED_DEBUG
706 #define SCHED_FEAT(name, enabled) \
709 static __read_mostly char *sched_feat_names[] = {
710 #include "sched_features.h"
716 static int sched_feat_show(struct seq_file *m, void *v)
720 for (i = 0; sched_feat_names[i]; i++) {
721 if (!(sysctl_sched_features & (1UL << i)))
723 seq_printf(m, "%s ", sched_feat_names[i]);
731 sched_feat_write(struct file *filp, const char __user *ubuf,
732 size_t cnt, loff_t *ppos)
742 if (copy_from_user(&buf, ubuf, cnt))
748 if (strncmp(cmp, "NO_", 3) == 0) {
753 for (i = 0; sched_feat_names[i]; i++) {
754 if (strcmp(cmp, sched_feat_names[i]) == 0) {
756 sysctl_sched_features &= ~(1UL << i);
758 sysctl_sched_features |= (1UL << i);
763 if (!sched_feat_names[i])
771 static int sched_feat_open(struct inode *inode, struct file *filp)
773 return single_open(filp, sched_feat_show, NULL);
776 static const struct file_operations sched_feat_fops = {
777 .open = sched_feat_open,
778 .write = sched_feat_write,
781 .release = single_release,
784 static __init int sched_init_debug(void)
786 debugfs_create_file("sched_features", 0644, NULL, NULL,
791 late_initcall(sched_init_debug);
795 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
798 * Number of tasks to iterate in a single balance run.
799 * Limited because this is done with IRQs disabled.
801 const_debug unsigned int sysctl_sched_nr_migrate = 32;
804 * period over which we average the RT time consumption, measured
809 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
812 * period over which we measure -rt task cpu usage in us.
815 unsigned int sysctl_sched_rt_period = 1000000;
817 static __read_mostly int scheduler_running;
820 * part of the period that we allow rt tasks to run in us.
823 int sysctl_sched_rt_runtime = 950000;
825 static inline u64 global_rt_period(void)
827 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
830 static inline u64 global_rt_runtime(void)
832 if (sysctl_sched_rt_runtime < 0)
835 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
838 #ifndef prepare_arch_switch
839 # define prepare_arch_switch(next) do { } while (0)
841 #ifndef finish_arch_switch
842 # define finish_arch_switch(prev) do { } while (0)
845 static inline int task_current(struct rq *rq, struct task_struct *p)
847 return rq->curr == p;
850 static inline int task_running(struct rq *rq, struct task_struct *p)
855 return task_current(rq, p);
859 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
860 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
864 * We can optimise this out completely for !SMP, because the
865 * SMP rebalancing from interrupt is the only thing that cares
872 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
876 * After ->on_cpu is cleared, the task can be moved to a different CPU.
877 * We must ensure this doesn't happen until the switch is completely
883 #ifdef CONFIG_DEBUG_SPINLOCK
884 /* this is a valid case when another task releases the spinlock */
885 rq->lock.owner = current;
888 * If we are tracking spinlock dependencies then we have to
889 * fix up the runqueue lock - which gets 'carried over' from
892 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
894 raw_spin_unlock_irq(&rq->lock);
897 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
898 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
902 * We can optimise this out completely for !SMP, because the
903 * SMP rebalancing from interrupt is the only thing that cares
908 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
909 raw_spin_unlock_irq(&rq->lock);
911 raw_spin_unlock(&rq->lock);
915 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
919 * After ->on_cpu is cleared, the task can be moved to a different CPU.
920 * We must ensure this doesn't happen until the switch is completely
926 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
930 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
933 * __task_rq_lock - lock the rq @p resides on.
935 static inline struct rq *__task_rq_lock(struct task_struct *p)
940 lockdep_assert_held(&p->pi_lock);
944 raw_spin_lock(&rq->lock);
945 if (likely(rq == task_rq(p)))
947 raw_spin_unlock(&rq->lock);
952 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
954 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
955 __acquires(p->pi_lock)
961 raw_spin_lock_irqsave(&p->pi_lock, *flags);
963 raw_spin_lock(&rq->lock);
964 if (likely(rq == task_rq(p)))
966 raw_spin_unlock(&rq->lock);
967 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
971 static void __task_rq_unlock(struct rq *rq)
974 raw_spin_unlock(&rq->lock);
978 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
980 __releases(p->pi_lock)
982 raw_spin_unlock(&rq->lock);
983 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
987 * this_rq_lock - lock this runqueue and disable interrupts.
989 static struct rq *this_rq_lock(void)
996 raw_spin_lock(&rq->lock);
1001 #ifdef CONFIG_SCHED_HRTICK
1003 * Use HR-timers to deliver accurate preemption points.
1005 * Its all a bit involved since we cannot program an hrt while holding the
1006 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1009 * When we get rescheduled we reprogram the hrtick_timer outside of the
1015 * - enabled by features
1016 * - hrtimer is actually high res
1018 static inline int hrtick_enabled(struct rq *rq)
1020 if (!sched_feat(HRTICK))
1022 if (!cpu_active(cpu_of(rq)))
1024 return hrtimer_is_hres_active(&rq->hrtick_timer);
1027 static void hrtick_clear(struct rq *rq)
1029 if (hrtimer_active(&rq->hrtick_timer))
1030 hrtimer_cancel(&rq->hrtick_timer);
1034 * High-resolution timer tick.
1035 * Runs from hardirq context with interrupts disabled.
1037 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1039 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1041 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1043 raw_spin_lock(&rq->lock);
1044 update_rq_clock(rq);
1045 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1046 raw_spin_unlock(&rq->lock);
1048 return HRTIMER_NORESTART;
1053 * called from hardirq (IPI) context
1055 static void __hrtick_start(void *arg)
1057 struct rq *rq = arg;
1059 raw_spin_lock(&rq->lock);
1060 hrtimer_restart(&rq->hrtick_timer);
1061 rq->hrtick_csd_pending = 0;
1062 raw_spin_unlock(&rq->lock);
1066 * Called to set the hrtick timer state.
1068 * called with rq->lock held and irqs disabled
1070 static void hrtick_start(struct rq *rq, u64 delay)
1072 struct hrtimer *timer = &rq->hrtick_timer;
1073 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1075 hrtimer_set_expires(timer, time);
1077 if (rq == this_rq()) {
1078 hrtimer_restart(timer);
1079 } else if (!rq->hrtick_csd_pending) {
1080 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1081 rq->hrtick_csd_pending = 1;
1086 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1088 int cpu = (int)(long)hcpu;
1091 case CPU_UP_CANCELED:
1092 case CPU_UP_CANCELED_FROZEN:
1093 case CPU_DOWN_PREPARE:
1094 case CPU_DOWN_PREPARE_FROZEN:
1096 case CPU_DEAD_FROZEN:
1097 hrtick_clear(cpu_rq(cpu));
1104 static __init void init_hrtick(void)
1106 hotcpu_notifier(hotplug_hrtick, 0);
1110 * Called to set the hrtick timer state.
1112 * called with rq->lock held and irqs disabled
1114 static void hrtick_start(struct rq *rq, u64 delay)
1116 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1117 HRTIMER_MODE_REL_PINNED, 0);
1120 static inline void init_hrtick(void)
1123 #endif /* CONFIG_SMP */
1125 static void init_rq_hrtick(struct rq *rq)
1128 rq->hrtick_csd_pending = 0;
1130 rq->hrtick_csd.flags = 0;
1131 rq->hrtick_csd.func = __hrtick_start;
1132 rq->hrtick_csd.info = rq;
1135 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1136 rq->hrtick_timer.function = hrtick;
1138 #else /* CONFIG_SCHED_HRTICK */
1139 static inline void hrtick_clear(struct rq *rq)
1143 static inline void init_rq_hrtick(struct rq *rq)
1147 static inline void init_hrtick(void)
1150 #endif /* CONFIG_SCHED_HRTICK */
1153 * resched_task - mark a task 'to be rescheduled now'.
1155 * On UP this means the setting of the need_resched flag, on SMP it
1156 * might also involve a cross-CPU call to trigger the scheduler on
1161 #ifndef tsk_is_polling
1162 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1165 static void resched_task(struct task_struct *p)
1169 assert_raw_spin_locked(&task_rq(p)->lock);
1171 if (test_tsk_need_resched(p))
1174 set_tsk_need_resched(p);
1177 if (cpu == smp_processor_id())
1180 /* NEED_RESCHED must be visible before we test polling */
1182 if (!tsk_is_polling(p))
1183 smp_send_reschedule(cpu);
1186 static void resched_cpu(int cpu)
1188 struct rq *rq = cpu_rq(cpu);
1189 unsigned long flags;
1191 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1193 resched_task(cpu_curr(cpu));
1194 raw_spin_unlock_irqrestore(&rq->lock, flags);
1199 * In the semi idle case, use the nearest busy cpu for migrating timers
1200 * from an idle cpu. This is good for power-savings.
1202 * We don't do similar optimization for completely idle system, as
1203 * selecting an idle cpu will add more delays to the timers than intended
1204 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1206 int get_nohz_timer_target(void)
1208 int cpu = smp_processor_id();
1210 struct sched_domain *sd;
1213 for_each_domain(cpu, sd) {
1214 for_each_cpu(i, sched_domain_span(sd)) {
1226 * When add_timer_on() enqueues a timer into the timer wheel of an
1227 * idle CPU then this timer might expire before the next timer event
1228 * which is scheduled to wake up that CPU. In case of a completely
1229 * idle system the next event might even be infinite time into the
1230 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1231 * leaves the inner idle loop so the newly added timer is taken into
1232 * account when the CPU goes back to idle and evaluates the timer
1233 * wheel for the next timer event.
1235 void wake_up_idle_cpu(int cpu)
1237 struct rq *rq = cpu_rq(cpu);
1239 if (cpu == smp_processor_id())
1243 * This is safe, as this function is called with the timer
1244 * wheel base lock of (cpu) held. When the CPU is on the way
1245 * to idle and has not yet set rq->curr to idle then it will
1246 * be serialized on the timer wheel base lock and take the new
1247 * timer into account automatically.
1249 if (rq->curr != rq->idle)
1253 * We can set TIF_RESCHED on the idle task of the other CPU
1254 * lockless. The worst case is that the other CPU runs the
1255 * idle task through an additional NOOP schedule()
1257 set_tsk_need_resched(rq->idle);
1259 /* NEED_RESCHED must be visible before we test polling */
1261 if (!tsk_is_polling(rq->idle))
1262 smp_send_reschedule(cpu);
1265 #endif /* CONFIG_NO_HZ */
1267 static u64 sched_avg_period(void)
1269 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1272 static void sched_avg_update(struct rq *rq)
1274 s64 period = sched_avg_period();
1276 while ((s64)(rq->clock - rq->age_stamp) > period) {
1278 * Inline assembly required to prevent the compiler
1279 * optimising this loop into a divmod call.
1280 * See __iter_div_u64_rem() for another example of this.
1282 asm("" : "+rm" (rq->age_stamp));
1283 rq->age_stamp += period;
1288 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1290 rq->rt_avg += rt_delta;
1291 sched_avg_update(rq);
1294 #else /* !CONFIG_SMP */
1295 static void resched_task(struct task_struct *p)
1297 assert_raw_spin_locked(&task_rq(p)->lock);
1298 set_tsk_need_resched(p);
1301 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1305 static void sched_avg_update(struct rq *rq)
1308 #endif /* CONFIG_SMP */
1310 #if BITS_PER_LONG == 32
1311 # define WMULT_CONST (~0UL)
1313 # define WMULT_CONST (1UL << 32)
1316 #define WMULT_SHIFT 32
1319 * Shift right and round:
1321 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1324 * delta *= weight / lw
1326 static unsigned long
1327 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1328 struct load_weight *lw)
1333 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1334 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1335 * 2^SCHED_LOAD_RESOLUTION.
1337 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
1338 tmp = (u64)delta_exec * scale_load_down(weight);
1340 tmp = (u64)delta_exec;
1342 if (!lw->inv_weight) {
1343 unsigned long w = scale_load_down(lw->weight);
1345 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
1347 else if (unlikely(!w))
1348 lw->inv_weight = WMULT_CONST;
1350 lw->inv_weight = WMULT_CONST / w;
1354 * Check whether we'd overflow the 64-bit multiplication:
1356 if (unlikely(tmp > WMULT_CONST))
1357 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1360 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1362 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1365 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1371 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1377 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1384 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1385 * of tasks with abnormal "nice" values across CPUs the contribution that
1386 * each task makes to its run queue's load is weighted according to its
1387 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1388 * scaled version of the new time slice allocation that they receive on time
1392 #define WEIGHT_IDLEPRIO 3
1393 #define WMULT_IDLEPRIO 1431655765
1396 * Nice levels are multiplicative, with a gentle 10% change for every
1397 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1398 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1399 * that remained on nice 0.
1401 * The "10% effect" is relative and cumulative: from _any_ nice level,
1402 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1403 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1404 * If a task goes up by ~10% and another task goes down by ~10% then
1405 * the relative distance between them is ~25%.)
1407 static const int prio_to_weight[40] = {
1408 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1409 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1410 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1411 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1412 /* 0 */ 1024, 820, 655, 526, 423,
1413 /* 5 */ 335, 272, 215, 172, 137,
1414 /* 10 */ 110, 87, 70, 56, 45,
1415 /* 15 */ 36, 29, 23, 18, 15,
1419 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1421 * In cases where the weight does not change often, we can use the
1422 * precalculated inverse to speed up arithmetics by turning divisions
1423 * into multiplications:
1425 static const u32 prio_to_wmult[40] = {
1426 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1427 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1428 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1429 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1430 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1431 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1432 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1433 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1436 /* Time spent by the tasks of the cpu accounting group executing in ... */
1437 enum cpuacct_stat_index {
1438 CPUACCT_STAT_USER, /* ... user mode */
1439 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1441 CPUACCT_STAT_NSTATS,
1444 #ifdef CONFIG_CGROUP_CPUACCT
1445 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1446 static void cpuacct_update_stats(struct task_struct *tsk,
1447 enum cpuacct_stat_index idx, cputime_t val);
1449 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1450 static inline void cpuacct_update_stats(struct task_struct *tsk,
1451 enum cpuacct_stat_index idx, cputime_t val) {}
1454 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1456 update_load_add(&rq->load, load);
1459 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1461 update_load_sub(&rq->load, load);
1464 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1465 typedef int (*tg_visitor)(struct task_group *, void *);
1468 * Iterate the full tree, calling @down when first entering a node and @up when
1469 * leaving it for the final time.
1471 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1473 struct task_group *parent, *child;
1477 parent = &root_task_group;
1479 ret = (*down)(parent, data);
1482 list_for_each_entry_rcu(child, &parent->children, siblings) {
1489 ret = (*up)(parent, data);
1494 parent = parent->parent;
1503 static int tg_nop(struct task_group *tg, void *data)
1510 /* Used instead of source_load when we know the type == 0 */
1511 static unsigned long weighted_cpuload(const int cpu)
1513 return cpu_rq(cpu)->load.weight;
1517 * Return a low guess at the load of a migration-source cpu weighted
1518 * according to the scheduling class and "nice" value.
1520 * We want to under-estimate the load of migration sources, to
1521 * balance conservatively.
1523 static unsigned long source_load(int cpu, int type)
1525 struct rq *rq = cpu_rq(cpu);
1526 unsigned long total = weighted_cpuload(cpu);
1528 if (type == 0 || !sched_feat(LB_BIAS))
1531 return min(rq->cpu_load[type-1], total);
1535 * Return a high guess at the load of a migration-target cpu weighted
1536 * according to the scheduling class and "nice" value.
1538 static unsigned long target_load(int cpu, int type)
1540 struct rq *rq = cpu_rq(cpu);
1541 unsigned long total = weighted_cpuload(cpu);
1543 if (type == 0 || !sched_feat(LB_BIAS))
1546 return max(rq->cpu_load[type-1], total);
1549 static unsigned long power_of(int cpu)
1551 return cpu_rq(cpu)->cpu_power;
1554 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1556 static unsigned long cpu_avg_load_per_task(int cpu)
1558 struct rq *rq = cpu_rq(cpu);
1559 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1562 rq->avg_load_per_task = rq->load.weight / nr_running;
1564 rq->avg_load_per_task = 0;
1566 return rq->avg_load_per_task;
1569 #ifdef CONFIG_FAIR_GROUP_SCHED
1572 * Compute the cpu's hierarchical load factor for each task group.
1573 * This needs to be done in a top-down fashion because the load of a child
1574 * group is a fraction of its parents load.
1576 static int tg_load_down(struct task_group *tg, void *data)
1579 long cpu = (long)data;
1582 load = cpu_rq(cpu)->load.weight;
1584 load = tg->parent->cfs_rq[cpu]->h_load;
1585 load *= tg->se[cpu]->load.weight;
1586 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1589 tg->cfs_rq[cpu]->h_load = load;
1594 static void update_h_load(long cpu)
1596 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1601 #ifdef CONFIG_PREEMPT
1603 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1606 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1607 * way at the expense of forcing extra atomic operations in all
1608 * invocations. This assures that the double_lock is acquired using the
1609 * same underlying policy as the spinlock_t on this architecture, which
1610 * reduces latency compared to the unfair variant below. However, it
1611 * also adds more overhead and therefore may reduce throughput.
1613 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1614 __releases(this_rq->lock)
1615 __acquires(busiest->lock)
1616 __acquires(this_rq->lock)
1618 raw_spin_unlock(&this_rq->lock);
1619 double_rq_lock(this_rq, busiest);
1626 * Unfair double_lock_balance: Optimizes throughput at the expense of
1627 * latency by eliminating extra atomic operations when the locks are
1628 * already in proper order on entry. This favors lower cpu-ids and will
1629 * grant the double lock to lower cpus over higher ids under contention,
1630 * regardless of entry order into the function.
1632 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1633 __releases(this_rq->lock)
1634 __acquires(busiest->lock)
1635 __acquires(this_rq->lock)
1639 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1640 if (busiest < this_rq) {
1641 raw_spin_unlock(&this_rq->lock);
1642 raw_spin_lock(&busiest->lock);
1643 raw_spin_lock_nested(&this_rq->lock,
1644 SINGLE_DEPTH_NESTING);
1647 raw_spin_lock_nested(&busiest->lock,
1648 SINGLE_DEPTH_NESTING);
1653 #endif /* CONFIG_PREEMPT */
1656 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1658 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1660 if (unlikely(!irqs_disabled())) {
1661 /* printk() doesn't work good under rq->lock */
1662 raw_spin_unlock(&this_rq->lock);
1666 return _double_lock_balance(this_rq, busiest);
1669 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1670 __releases(busiest->lock)
1672 raw_spin_unlock(&busiest->lock);
1673 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1677 * double_rq_lock - safely lock two runqueues
1679 * Note this does not disable interrupts like task_rq_lock,
1680 * you need to do so manually before calling.
1682 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1683 __acquires(rq1->lock)
1684 __acquires(rq2->lock)
1686 BUG_ON(!irqs_disabled());
1688 raw_spin_lock(&rq1->lock);
1689 __acquire(rq2->lock); /* Fake it out ;) */
1692 raw_spin_lock(&rq1->lock);
1693 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1695 raw_spin_lock(&rq2->lock);
1696 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1702 * double_rq_unlock - safely unlock two runqueues
1704 * Note this does not restore interrupts like task_rq_unlock,
1705 * you need to do so manually after calling.
1707 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1708 __releases(rq1->lock)
1709 __releases(rq2->lock)
1711 raw_spin_unlock(&rq1->lock);
1713 raw_spin_unlock(&rq2->lock);
1715 __release(rq2->lock);
1718 #else /* CONFIG_SMP */
1721 * double_rq_lock - safely lock two runqueues
1723 * Note this does not disable interrupts like task_rq_lock,
1724 * you need to do so manually before calling.
1726 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1727 __acquires(rq1->lock)
1728 __acquires(rq2->lock)
1730 BUG_ON(!irqs_disabled());
1732 raw_spin_lock(&rq1->lock);
1733 __acquire(rq2->lock); /* Fake it out ;) */
1737 * double_rq_unlock - safely unlock two runqueues
1739 * Note this does not restore interrupts like task_rq_unlock,
1740 * you need to do so manually after calling.
1742 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1743 __releases(rq1->lock)
1744 __releases(rq2->lock)
1747 raw_spin_unlock(&rq1->lock);
1748 __release(rq2->lock);
1753 static void calc_load_account_idle(struct rq *this_rq);
1754 static void update_sysctl(void);
1755 static int get_update_sysctl_factor(void);
1756 static void update_cpu_load(struct rq *this_rq);
1758 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1760 set_task_rq(p, cpu);
1763 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1764 * successfuly executed on another CPU. We must ensure that updates of
1765 * per-task data have been completed by this moment.
1768 task_thread_info(p)->cpu = cpu;
1772 static const struct sched_class rt_sched_class;
1774 #define sched_class_highest (&stop_sched_class)
1775 #define for_each_class(class) \
1776 for (class = sched_class_highest; class; class = class->next)
1778 #include "sched_stats.h"
1780 static void inc_nr_running(struct rq *rq)
1785 static void dec_nr_running(struct rq *rq)
1790 static void set_load_weight(struct task_struct *p)
1792 int prio = p->static_prio - MAX_RT_PRIO;
1793 struct load_weight *load = &p->se.load;
1796 * SCHED_IDLE tasks get minimal weight:
1798 if (p->policy == SCHED_IDLE) {
1799 load->weight = scale_load(WEIGHT_IDLEPRIO);
1800 load->inv_weight = WMULT_IDLEPRIO;
1804 load->weight = scale_load(prio_to_weight[prio]);
1805 load->inv_weight = prio_to_wmult[prio];
1808 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1810 update_rq_clock(rq);
1811 sched_info_queued(p);
1812 p->sched_class->enqueue_task(rq, p, flags);
1815 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1817 update_rq_clock(rq);
1818 sched_info_dequeued(p);
1819 p->sched_class->dequeue_task(rq, p, flags);
1823 * activate_task - move a task to the runqueue.
1825 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1827 if (task_contributes_to_load(p))
1828 rq->nr_uninterruptible--;
1830 enqueue_task(rq, p, flags);
1835 * deactivate_task - remove a task from the runqueue.
1837 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1839 if (task_contributes_to_load(p))
1840 rq->nr_uninterruptible++;
1842 dequeue_task(rq, p, flags);
1846 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1849 * There are no locks covering percpu hardirq/softirq time.
1850 * They are only modified in account_system_vtime, on corresponding CPU
1851 * with interrupts disabled. So, writes are safe.
1852 * They are read and saved off onto struct rq in update_rq_clock().
1853 * This may result in other CPU reading this CPU's irq time and can
1854 * race with irq/account_system_vtime on this CPU. We would either get old
1855 * or new value with a side effect of accounting a slice of irq time to wrong
1856 * task when irq is in progress while we read rq->clock. That is a worthy
1857 * compromise in place of having locks on each irq in account_system_time.
1859 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1860 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1862 static DEFINE_PER_CPU(u64, irq_start_time);
1863 static int sched_clock_irqtime;
1865 void enable_sched_clock_irqtime(void)
1867 sched_clock_irqtime = 1;
1870 void disable_sched_clock_irqtime(void)
1872 sched_clock_irqtime = 0;
1875 #ifndef CONFIG_64BIT
1876 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1878 static inline void irq_time_write_begin(void)
1880 __this_cpu_inc(irq_time_seq.sequence);
1884 static inline void irq_time_write_end(void)
1887 __this_cpu_inc(irq_time_seq.sequence);
1890 static inline u64 irq_time_read(int cpu)
1896 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1897 irq_time = per_cpu(cpu_softirq_time, cpu) +
1898 per_cpu(cpu_hardirq_time, cpu);
1899 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1903 #else /* CONFIG_64BIT */
1904 static inline void irq_time_write_begin(void)
1908 static inline void irq_time_write_end(void)
1912 static inline u64 irq_time_read(int cpu)
1914 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1916 #endif /* CONFIG_64BIT */
1919 * Called before incrementing preempt_count on {soft,}irq_enter
1920 * and before decrementing preempt_count on {soft,}irq_exit.
1922 void account_system_vtime(struct task_struct *curr)
1924 unsigned long flags;
1928 if (!sched_clock_irqtime)
1931 local_irq_save(flags);
1933 cpu = smp_processor_id();
1934 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1935 __this_cpu_add(irq_start_time, delta);
1937 irq_time_write_begin();
1939 * We do not account for softirq time from ksoftirqd here.
1940 * We want to continue accounting softirq time to ksoftirqd thread
1941 * in that case, so as not to confuse scheduler with a special task
1942 * that do not consume any time, but still wants to run.
1944 if (hardirq_count())
1945 __this_cpu_add(cpu_hardirq_time, delta);
1946 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
1947 __this_cpu_add(cpu_softirq_time, delta);
1949 irq_time_write_end();
1950 local_irq_restore(flags);
1952 EXPORT_SYMBOL_GPL(account_system_vtime);
1954 static void update_rq_clock_task(struct rq *rq, s64 delta)
1958 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
1961 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1962 * this case when a previous update_rq_clock() happened inside a
1963 * {soft,}irq region.
1965 * When this happens, we stop ->clock_task and only update the
1966 * prev_irq_time stamp to account for the part that fit, so that a next
1967 * update will consume the rest. This ensures ->clock_task is
1970 * It does however cause some slight miss-attribution of {soft,}irq
1971 * time, a more accurate solution would be to update the irq_time using
1972 * the current rq->clock timestamp, except that would require using
1975 if (irq_delta > delta)
1978 rq->prev_irq_time += irq_delta;
1980 rq->clock_task += delta;
1982 if (irq_delta && sched_feat(NONIRQ_POWER))
1983 sched_rt_avg_update(rq, irq_delta);
1986 static int irqtime_account_hi_update(void)
1988 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
1989 unsigned long flags;
1993 local_irq_save(flags);
1994 latest_ns = this_cpu_read(cpu_hardirq_time);
1995 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
1997 local_irq_restore(flags);
2001 static int irqtime_account_si_update(void)
2003 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2004 unsigned long flags;
2008 local_irq_save(flags);
2009 latest_ns = this_cpu_read(cpu_softirq_time);
2010 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2012 local_irq_restore(flags);
2016 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2018 #define sched_clock_irqtime (0)
2020 static void update_rq_clock_task(struct rq *rq, s64 delta)
2022 rq->clock_task += delta;
2025 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2027 #include "sched_idletask.c"
2028 #include "sched_fair.c"
2029 #include "sched_rt.c"
2030 #include "sched_autogroup.c"
2031 #include "sched_stoptask.c"
2032 #ifdef CONFIG_SCHED_DEBUG
2033 # include "sched_debug.c"
2036 void sched_set_stop_task(int cpu, struct task_struct *stop)
2038 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2039 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2043 * Make it appear like a SCHED_FIFO task, its something
2044 * userspace knows about and won't get confused about.
2046 * Also, it will make PI more or less work without too
2047 * much confusion -- but then, stop work should not
2048 * rely on PI working anyway.
2050 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2052 stop->sched_class = &stop_sched_class;
2055 cpu_rq(cpu)->stop = stop;
2059 * Reset it back to a normal scheduling class so that
2060 * it can die in pieces.
2062 old_stop->sched_class = &rt_sched_class;
2067 * __normal_prio - return the priority that is based on the static prio
2069 static inline int __normal_prio(struct task_struct *p)
2071 return p->static_prio;
2075 * Calculate the expected normal priority: i.e. priority
2076 * without taking RT-inheritance into account. Might be
2077 * boosted by interactivity modifiers. Changes upon fork,
2078 * setprio syscalls, and whenever the interactivity
2079 * estimator recalculates.
2081 static inline int normal_prio(struct task_struct *p)
2085 if (task_has_rt_policy(p))
2086 prio = MAX_RT_PRIO-1 - p->rt_priority;
2088 prio = __normal_prio(p);
2093 * Calculate the current priority, i.e. the priority
2094 * taken into account by the scheduler. This value might
2095 * be boosted by RT tasks, or might be boosted by
2096 * interactivity modifiers. Will be RT if the task got
2097 * RT-boosted. If not then it returns p->normal_prio.
2099 static int effective_prio(struct task_struct *p)
2101 p->normal_prio = normal_prio(p);
2103 * If we are RT tasks or we were boosted to RT priority,
2104 * keep the priority unchanged. Otherwise, update priority
2105 * to the normal priority:
2107 if (!rt_prio(p->prio))
2108 return p->normal_prio;
2113 * task_curr - is this task currently executing on a CPU?
2114 * @p: the task in question.
2116 inline int task_curr(const struct task_struct *p)
2118 return cpu_curr(task_cpu(p)) == p;
2121 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2122 const struct sched_class *prev_class,
2125 if (prev_class != p->sched_class) {
2126 if (prev_class->switched_from)
2127 prev_class->switched_from(rq, p);
2128 p->sched_class->switched_to(rq, p);
2129 } else if (oldprio != p->prio)
2130 p->sched_class->prio_changed(rq, p, oldprio);
2133 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2135 const struct sched_class *class;
2137 if (p->sched_class == rq->curr->sched_class) {
2138 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2140 for_each_class(class) {
2141 if (class == rq->curr->sched_class)
2143 if (class == p->sched_class) {
2144 resched_task(rq->curr);
2151 * A queue event has occurred, and we're going to schedule. In
2152 * this case, we can save a useless back to back clock update.
2154 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2155 rq->skip_clock_update = 1;
2160 * Is this task likely cache-hot:
2163 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2167 if (p->sched_class != &fair_sched_class)
2170 if (unlikely(p->policy == SCHED_IDLE))
2174 * Buddy candidates are cache hot:
2176 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2177 (&p->se == cfs_rq_of(&p->se)->next ||
2178 &p->se == cfs_rq_of(&p->se)->last))
2181 if (sysctl_sched_migration_cost == -1)
2183 if (sysctl_sched_migration_cost == 0)
2186 delta = now - p->se.exec_start;
2188 return delta < (s64)sysctl_sched_migration_cost;
2191 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2193 #ifdef CONFIG_SCHED_DEBUG
2195 * We should never call set_task_cpu() on a blocked task,
2196 * ttwu() will sort out the placement.
2198 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2199 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2201 #ifdef CONFIG_LOCKDEP
2203 * The caller should hold either p->pi_lock or rq->lock, when changing
2204 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2206 * sched_move_task() holds both and thus holding either pins the cgroup,
2209 * Furthermore, all task_rq users should acquire both locks, see
2212 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2213 lockdep_is_held(&task_rq(p)->lock)));
2217 trace_sched_migrate_task(p, new_cpu);
2219 if (task_cpu(p) != new_cpu) {
2220 p->se.nr_migrations++;
2221 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2224 __set_task_cpu(p, new_cpu);
2227 struct migration_arg {
2228 struct task_struct *task;
2232 static int migration_cpu_stop(void *data);
2235 * wait_task_inactive - wait for a thread to unschedule.
2237 * If @match_state is nonzero, it's the @p->state value just checked and
2238 * not expected to change. If it changes, i.e. @p might have woken up,
2239 * then return zero. When we succeed in waiting for @p to be off its CPU,
2240 * we return a positive number (its total switch count). If a second call
2241 * a short while later returns the same number, the caller can be sure that
2242 * @p has remained unscheduled the whole time.
2244 * The caller must ensure that the task *will* unschedule sometime soon,
2245 * else this function might spin for a *long* time. This function can't
2246 * be called with interrupts off, or it may introduce deadlock with
2247 * smp_call_function() if an IPI is sent by the same process we are
2248 * waiting to become inactive.
2250 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2252 unsigned long flags;
2259 * We do the initial early heuristics without holding
2260 * any task-queue locks at all. We'll only try to get
2261 * the runqueue lock when things look like they will
2267 * If the task is actively running on another CPU
2268 * still, just relax and busy-wait without holding
2271 * NOTE! Since we don't hold any locks, it's not
2272 * even sure that "rq" stays as the right runqueue!
2273 * But we don't care, since "task_running()" will
2274 * return false if the runqueue has changed and p
2275 * is actually now running somewhere else!
2277 while (task_running(rq, p)) {
2278 if (match_state && unlikely(p->state != match_state))
2284 * Ok, time to look more closely! We need the rq
2285 * lock now, to be *sure*. If we're wrong, we'll
2286 * just go back and repeat.
2288 rq = task_rq_lock(p, &flags);
2289 trace_sched_wait_task(p);
2290 running = task_running(rq, p);
2293 if (!match_state || p->state == match_state)
2294 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2295 task_rq_unlock(rq, p, &flags);
2298 * If it changed from the expected state, bail out now.
2300 if (unlikely(!ncsw))
2304 * Was it really running after all now that we
2305 * checked with the proper locks actually held?
2307 * Oops. Go back and try again..
2309 if (unlikely(running)) {
2315 * It's not enough that it's not actively running,
2316 * it must be off the runqueue _entirely_, and not
2319 * So if it was still runnable (but just not actively
2320 * running right now), it's preempted, and we should
2321 * yield - it could be a while.
2323 if (unlikely(on_rq)) {
2324 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2326 set_current_state(TASK_UNINTERRUPTIBLE);
2327 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2332 * Ahh, all good. It wasn't running, and it wasn't
2333 * runnable, which means that it will never become
2334 * running in the future either. We're all done!
2343 * kick_process - kick a running thread to enter/exit the kernel
2344 * @p: the to-be-kicked thread
2346 * Cause a process which is running on another CPU to enter
2347 * kernel-mode, without any delay. (to get signals handled.)
2349 * NOTE: this function doesn't have to take the runqueue lock,
2350 * because all it wants to ensure is that the remote task enters
2351 * the kernel. If the IPI races and the task has been migrated
2352 * to another CPU then no harm is done and the purpose has been
2355 void kick_process(struct task_struct *p)
2361 if ((cpu != smp_processor_id()) && task_curr(p))
2362 smp_send_reschedule(cpu);
2365 EXPORT_SYMBOL_GPL(kick_process);
2366 #endif /* CONFIG_SMP */
2370 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2372 static int select_fallback_rq(int cpu, struct task_struct *p)
2375 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2377 /* Look for allowed, online CPU in same node. */
2378 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2379 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2382 /* Any allowed, online CPU? */
2383 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2384 if (dest_cpu < nr_cpu_ids)
2387 /* No more Mr. Nice Guy. */
2388 dest_cpu = cpuset_cpus_allowed_fallback(p);
2390 * Don't tell them about moving exiting tasks or
2391 * kernel threads (both mm NULL), since they never
2394 if (p->mm && printk_ratelimit()) {
2395 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2396 task_pid_nr(p), p->comm, cpu);
2403 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2406 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2408 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2411 * In order not to call set_task_cpu() on a blocking task we need
2412 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2415 * Since this is common to all placement strategies, this lives here.
2417 * [ this allows ->select_task() to simply return task_cpu(p) and
2418 * not worry about this generic constraint ]
2420 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2422 cpu = select_fallback_rq(task_cpu(p), p);
2427 static void update_avg(u64 *avg, u64 sample)
2429 s64 diff = sample - *avg;
2435 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2437 #ifdef CONFIG_SCHEDSTATS
2438 struct rq *rq = this_rq();
2441 int this_cpu = smp_processor_id();
2443 if (cpu == this_cpu) {
2444 schedstat_inc(rq, ttwu_local);
2445 schedstat_inc(p, se.statistics.nr_wakeups_local);
2447 struct sched_domain *sd;
2449 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2451 for_each_domain(this_cpu, sd) {
2452 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2453 schedstat_inc(sd, ttwu_wake_remote);
2460 if (wake_flags & WF_MIGRATED)
2461 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2463 #endif /* CONFIG_SMP */
2465 schedstat_inc(rq, ttwu_count);
2466 schedstat_inc(p, se.statistics.nr_wakeups);
2468 if (wake_flags & WF_SYNC)
2469 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2471 #endif /* CONFIG_SCHEDSTATS */
2474 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2476 activate_task(rq, p, en_flags);
2479 /* if a worker is waking up, notify workqueue */
2480 if (p->flags & PF_WQ_WORKER)
2481 wq_worker_waking_up(p, cpu_of(rq));
2485 * Mark the task runnable and perform wakeup-preemption.
2488 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2490 trace_sched_wakeup(p, true);
2491 check_preempt_curr(rq, p, wake_flags);
2493 p->state = TASK_RUNNING;
2495 if (p->sched_class->task_woken)
2496 p->sched_class->task_woken(rq, p);
2498 if (unlikely(rq->idle_stamp)) {
2499 u64 delta = rq->clock - rq->idle_stamp;
2500 u64 max = 2*sysctl_sched_migration_cost;
2505 update_avg(&rq->avg_idle, delta);
2512 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2515 if (p->sched_contributes_to_load)
2516 rq->nr_uninterruptible--;
2519 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2520 ttwu_do_wakeup(rq, p, wake_flags);
2524 * Called in case the task @p isn't fully descheduled from its runqueue,
2525 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2526 * since all we need to do is flip p->state to TASK_RUNNING, since
2527 * the task is still ->on_rq.
2529 static int ttwu_remote(struct task_struct *p, int wake_flags)
2534 rq = __task_rq_lock(p);
2536 ttwu_do_wakeup(rq, p, wake_flags);
2539 __task_rq_unlock(rq);
2545 static void sched_ttwu_do_pending(struct task_struct *list)
2547 struct rq *rq = this_rq();
2549 raw_spin_lock(&rq->lock);
2552 struct task_struct *p = list;
2553 list = list->wake_entry;
2554 ttwu_do_activate(rq, p, 0);
2557 raw_spin_unlock(&rq->lock);
2560 #ifdef CONFIG_HOTPLUG_CPU
2562 static void sched_ttwu_pending(void)
2564 struct rq *rq = this_rq();
2565 struct task_struct *list = xchg(&rq->wake_list, NULL);
2570 sched_ttwu_do_pending(list);
2573 #endif /* CONFIG_HOTPLUG_CPU */
2575 void scheduler_ipi(void)
2577 struct rq *rq = this_rq();
2578 struct task_struct *list = xchg(&rq->wake_list, NULL);
2584 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2585 * traditionally all their work was done from the interrupt return
2586 * path. Now that we actually do some work, we need to make sure
2589 * Some archs already do call them, luckily irq_enter/exit nest
2592 * Arguably we should visit all archs and update all handlers,
2593 * however a fair share of IPIs are still resched only so this would
2594 * somewhat pessimize the simple resched case.
2597 sched_ttwu_do_pending(list);
2601 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2603 struct rq *rq = cpu_rq(cpu);
2604 struct task_struct *next = rq->wake_list;
2607 struct task_struct *old = next;
2609 p->wake_entry = next;
2610 next = cmpxchg(&rq->wake_list, old, p);
2616 smp_send_reschedule(cpu);
2619 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2620 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2625 rq = __task_rq_lock(p);
2627 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2628 ttwu_do_wakeup(rq, p, wake_flags);
2631 __task_rq_unlock(rq);
2636 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2637 #endif /* CONFIG_SMP */
2639 static void ttwu_queue(struct task_struct *p, int cpu)
2641 struct rq *rq = cpu_rq(cpu);
2643 #if defined(CONFIG_SMP)
2644 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2645 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2646 ttwu_queue_remote(p, cpu);
2651 raw_spin_lock(&rq->lock);
2652 ttwu_do_activate(rq, p, 0);
2653 raw_spin_unlock(&rq->lock);
2657 * try_to_wake_up - wake up a thread
2658 * @p: the thread to be awakened
2659 * @state: the mask of task states that can be woken
2660 * @wake_flags: wake modifier flags (WF_*)
2662 * Put it on the run-queue if it's not already there. The "current"
2663 * thread is always on the run-queue (except when the actual
2664 * re-schedule is in progress), and as such you're allowed to do
2665 * the simpler "current->state = TASK_RUNNING" to mark yourself
2666 * runnable without the overhead of this.
2668 * Returns %true if @p was woken up, %false if it was already running
2669 * or @state didn't match @p's state.
2672 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2674 unsigned long flags;
2675 int cpu, success = 0;
2678 raw_spin_lock_irqsave(&p->pi_lock, flags);
2679 if (!(p->state & state))
2682 success = 1; /* we're going to change ->state */
2685 if (p->on_rq && ttwu_remote(p, wake_flags))
2690 * If the owning (remote) cpu is still in the middle of schedule() with
2691 * this task as prev, wait until its done referencing the task.
2694 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2696 * In case the architecture enables interrupts in
2697 * context_switch(), we cannot busy wait, since that
2698 * would lead to deadlocks when an interrupt hits and
2699 * tries to wake up @prev. So bail and do a complete
2702 if (ttwu_activate_remote(p, wake_flags))
2709 * Pairs with the smp_wmb() in finish_lock_switch().
2713 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2714 p->state = TASK_WAKING;
2716 if (p->sched_class->task_waking)
2717 p->sched_class->task_waking(p);
2719 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2720 if (task_cpu(p) != cpu) {
2721 wake_flags |= WF_MIGRATED;
2722 set_task_cpu(p, cpu);
2724 #endif /* CONFIG_SMP */
2728 ttwu_stat(p, cpu, wake_flags);
2730 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2736 * try_to_wake_up_local - try to wake up a local task with rq lock held
2737 * @p: the thread to be awakened
2739 * Put @p on the run-queue if it's not already there. The caller must
2740 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2743 static void try_to_wake_up_local(struct task_struct *p)
2745 struct rq *rq = task_rq(p);
2747 BUG_ON(rq != this_rq());
2748 BUG_ON(p == current);
2749 lockdep_assert_held(&rq->lock);
2751 if (!raw_spin_trylock(&p->pi_lock)) {
2752 raw_spin_unlock(&rq->lock);
2753 raw_spin_lock(&p->pi_lock);
2754 raw_spin_lock(&rq->lock);
2757 if (!(p->state & TASK_NORMAL))
2761 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2763 ttwu_do_wakeup(rq, p, 0);
2764 ttwu_stat(p, smp_processor_id(), 0);
2766 raw_spin_unlock(&p->pi_lock);
2770 * wake_up_process - Wake up a specific process
2771 * @p: The process to be woken up.
2773 * Attempt to wake up the nominated process and move it to the set of runnable
2774 * processes. Returns 1 if the process was woken up, 0 if it was already
2777 * It may be assumed that this function implies a write memory barrier before
2778 * changing the task state if and only if any tasks are woken up.
2780 int wake_up_process(struct task_struct *p)
2782 return try_to_wake_up(p, TASK_ALL, 0);
2784 EXPORT_SYMBOL(wake_up_process);
2786 int wake_up_state(struct task_struct *p, unsigned int state)
2788 return try_to_wake_up(p, state, 0);
2792 * Perform scheduler related setup for a newly forked process p.
2793 * p is forked by current.
2795 * __sched_fork() is basic setup used by init_idle() too:
2797 static void __sched_fork(struct task_struct *p)
2802 p->se.exec_start = 0;
2803 p->se.sum_exec_runtime = 0;
2804 p->se.prev_sum_exec_runtime = 0;
2805 p->se.nr_migrations = 0;
2807 INIT_LIST_HEAD(&p->se.group_node);
2809 #ifdef CONFIG_SCHEDSTATS
2810 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2813 INIT_LIST_HEAD(&p->rt.run_list);
2815 #ifdef CONFIG_PREEMPT_NOTIFIERS
2816 INIT_HLIST_HEAD(&p->preempt_notifiers);
2821 * fork()/clone()-time setup:
2823 void sched_fork(struct task_struct *p)
2825 unsigned long flags;
2826 int cpu = get_cpu();
2830 * We mark the process as running here. This guarantees that
2831 * nobody will actually run it, and a signal or other external
2832 * event cannot wake it up and insert it on the runqueue either.
2834 p->state = TASK_RUNNING;
2837 * Revert to default priority/policy on fork if requested.
2839 if (unlikely(p->sched_reset_on_fork)) {
2840 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2841 p->policy = SCHED_NORMAL;
2842 p->normal_prio = p->static_prio;
2845 if (PRIO_TO_NICE(p->static_prio) < 0) {
2846 p->static_prio = NICE_TO_PRIO(0);
2847 p->normal_prio = p->static_prio;
2852 * We don't need the reset flag anymore after the fork. It has
2853 * fulfilled its duty:
2855 p->sched_reset_on_fork = 0;
2859 * Make sure we do not leak PI boosting priority to the child.
2861 p->prio = current->normal_prio;
2863 if (!rt_prio(p->prio))
2864 p->sched_class = &fair_sched_class;
2866 if (p->sched_class->task_fork)
2867 p->sched_class->task_fork(p);
2870 * The child is not yet in the pid-hash so no cgroup attach races,
2871 * and the cgroup is pinned to this child due to cgroup_fork()
2872 * is ran before sched_fork().
2874 * Silence PROVE_RCU.
2876 raw_spin_lock_irqsave(&p->pi_lock, flags);
2877 set_task_cpu(p, cpu);
2878 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2880 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2881 if (likely(sched_info_on()))
2882 memset(&p->sched_info, 0, sizeof(p->sched_info));
2884 #if defined(CONFIG_SMP)
2887 #ifdef CONFIG_PREEMPT
2888 /* Want to start with kernel preemption disabled. */
2889 task_thread_info(p)->preempt_count = 1;
2892 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2899 * wake_up_new_task - wake up a newly created task for the first time.
2901 * This function will do some initial scheduler statistics housekeeping
2902 * that must be done for every newly created context, then puts the task
2903 * on the runqueue and wakes it.
2905 void wake_up_new_task(struct task_struct *p)
2907 unsigned long flags;
2910 raw_spin_lock_irqsave(&p->pi_lock, flags);
2913 * Fork balancing, do it here and not earlier because:
2914 * - cpus_allowed can change in the fork path
2915 * - any previously selected cpu might disappear through hotplug
2917 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
2920 rq = __task_rq_lock(p);
2921 activate_task(rq, p, 0);
2923 trace_sched_wakeup_new(p, true);
2924 check_preempt_curr(rq, p, WF_FORK);
2926 if (p->sched_class->task_woken)
2927 p->sched_class->task_woken(rq, p);
2929 task_rq_unlock(rq, p, &flags);
2932 #ifdef CONFIG_PREEMPT_NOTIFIERS
2935 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2936 * @notifier: notifier struct to register
2938 void preempt_notifier_register(struct preempt_notifier *notifier)
2940 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2942 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2945 * preempt_notifier_unregister - no longer interested in preemption notifications
2946 * @notifier: notifier struct to unregister
2948 * This is safe to call from within a preemption notifier.
2950 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2952 hlist_del(¬ifier->link);
2954 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2956 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2958 struct preempt_notifier *notifier;
2959 struct hlist_node *node;
2961 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2962 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2966 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2967 struct task_struct *next)
2969 struct preempt_notifier *notifier;
2970 struct hlist_node *node;
2972 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2973 notifier->ops->sched_out(notifier, next);
2976 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2978 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2983 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2984 struct task_struct *next)
2988 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2991 * prepare_task_switch - prepare to switch tasks
2992 * @rq: the runqueue preparing to switch
2993 * @prev: the current task that is being switched out
2994 * @next: the task we are going to switch to.
2996 * This is called with the rq lock held and interrupts off. It must
2997 * be paired with a subsequent finish_task_switch after the context
3000 * prepare_task_switch sets up locking and calls architecture specific
3004 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3005 struct task_struct *next)
3007 sched_info_switch(prev, next);
3008 perf_event_task_sched_out(prev, next);
3009 fire_sched_out_preempt_notifiers(prev, next);
3010 prepare_lock_switch(rq, next);
3011 prepare_arch_switch(next);
3012 trace_sched_switch(prev, next);
3016 * finish_task_switch - clean up after a task-switch
3017 * @rq: runqueue associated with task-switch
3018 * @prev: the thread we just switched away from.
3020 * finish_task_switch must be called after the context switch, paired
3021 * with a prepare_task_switch call before the context switch.
3022 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3023 * and do any other architecture-specific cleanup actions.
3025 * Note that we may have delayed dropping an mm in context_switch(). If
3026 * so, we finish that here outside of the runqueue lock. (Doing it
3027 * with the lock held can cause deadlocks; see schedule() for
3030 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3031 __releases(rq->lock)
3033 struct mm_struct *mm = rq->prev_mm;
3039 * A task struct has one reference for the use as "current".
3040 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3041 * schedule one last time. The schedule call will never return, and
3042 * the scheduled task must drop that reference.
3043 * The test for TASK_DEAD must occur while the runqueue locks are
3044 * still held, otherwise prev could be scheduled on another cpu, die
3045 * there before we look at prev->state, and then the reference would
3047 * Manfred Spraul <manfred@colorfullife.com>
3049 prev_state = prev->state;
3050 finish_arch_switch(prev);
3051 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3052 local_irq_disable();
3053 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3054 perf_event_task_sched_in(current);
3055 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3057 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3058 finish_lock_switch(rq, prev);
3060 fire_sched_in_preempt_notifiers(current);
3063 if (unlikely(prev_state == TASK_DEAD)) {
3065 * Remove function-return probe instances associated with this
3066 * task and put them back on the free list.
3068 kprobe_flush_task(prev);
3069 put_task_struct(prev);
3075 /* assumes rq->lock is held */
3076 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3078 if (prev->sched_class->pre_schedule)
3079 prev->sched_class->pre_schedule(rq, prev);
3082 /* rq->lock is NOT held, but preemption is disabled */
3083 static inline void post_schedule(struct rq *rq)
3085 if (rq->post_schedule) {
3086 unsigned long flags;
3088 raw_spin_lock_irqsave(&rq->lock, flags);
3089 if (rq->curr->sched_class->post_schedule)
3090 rq->curr->sched_class->post_schedule(rq);
3091 raw_spin_unlock_irqrestore(&rq->lock, flags);
3093 rq->post_schedule = 0;
3099 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3103 static inline void post_schedule(struct rq *rq)
3110 * schedule_tail - first thing a freshly forked thread must call.
3111 * @prev: the thread we just switched away from.
3113 asmlinkage void schedule_tail(struct task_struct *prev)
3114 __releases(rq->lock)
3116 struct rq *rq = this_rq();
3118 finish_task_switch(rq, prev);
3121 * FIXME: do we need to worry about rq being invalidated by the
3126 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3127 /* In this case, finish_task_switch does not reenable preemption */
3130 if (current->set_child_tid)
3131 put_user(task_pid_vnr(current), current->set_child_tid);
3135 * context_switch - switch to the new MM and the new
3136 * thread's register state.
3139 context_switch(struct rq *rq, struct task_struct *prev,
3140 struct task_struct *next)
3142 struct mm_struct *mm, *oldmm;
3144 prepare_task_switch(rq, prev, next);
3147 oldmm = prev->active_mm;
3149 * For paravirt, this is coupled with an exit in switch_to to
3150 * combine the page table reload and the switch backend into
3153 arch_start_context_switch(prev);
3156 next->active_mm = oldmm;
3157 atomic_inc(&oldmm->mm_count);
3158 enter_lazy_tlb(oldmm, next);
3160 switch_mm(oldmm, mm, next);
3163 prev->active_mm = NULL;
3164 rq->prev_mm = oldmm;
3167 * Since the runqueue lock will be released by the next
3168 * task (which is an invalid locking op but in the case
3169 * of the scheduler it's an obvious special-case), so we
3170 * do an early lockdep release here:
3172 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3173 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3176 /* Here we just switch the register state and the stack. */
3177 switch_to(prev, next, prev);
3181 * this_rq must be evaluated again because prev may have moved
3182 * CPUs since it called schedule(), thus the 'rq' on its stack
3183 * frame will be invalid.
3185 finish_task_switch(this_rq(), prev);
3189 * nr_running, nr_uninterruptible and nr_context_switches:
3191 * externally visible scheduler statistics: current number of runnable
3192 * threads, current number of uninterruptible-sleeping threads, total
3193 * number of context switches performed since bootup.
3195 unsigned long nr_running(void)
3197 unsigned long i, sum = 0;
3199 for_each_online_cpu(i)
3200 sum += cpu_rq(i)->nr_running;
3205 unsigned long nr_uninterruptible(void)
3207 unsigned long i, sum = 0;
3209 for_each_possible_cpu(i)
3210 sum += cpu_rq(i)->nr_uninterruptible;
3213 * Since we read the counters lockless, it might be slightly
3214 * inaccurate. Do not allow it to go below zero though:
3216 if (unlikely((long)sum < 0))
3222 unsigned long long nr_context_switches(void)
3225 unsigned long long sum = 0;
3227 for_each_possible_cpu(i)
3228 sum += cpu_rq(i)->nr_switches;
3233 unsigned long nr_iowait(void)
3235 unsigned long i, sum = 0;
3237 for_each_possible_cpu(i)
3238 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3243 unsigned long nr_iowait_cpu(int cpu)
3245 struct rq *this = cpu_rq(cpu);
3246 return atomic_read(&this->nr_iowait);
3249 unsigned long this_cpu_load(void)
3251 struct rq *this = this_rq();
3252 return this->cpu_load[0];
3256 /* Variables and functions for calc_load */
3257 static atomic_long_t calc_load_tasks;
3258 static unsigned long calc_load_update;
3259 unsigned long avenrun[3];
3260 EXPORT_SYMBOL(avenrun);
3262 static long calc_load_fold_active(struct rq *this_rq)
3264 long nr_active, delta = 0;
3266 nr_active = this_rq->nr_running;
3267 nr_active += (long) this_rq->nr_uninterruptible;
3269 if (nr_active != this_rq->calc_load_active) {
3270 delta = nr_active - this_rq->calc_load_active;
3271 this_rq->calc_load_active = nr_active;
3277 static unsigned long
3278 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3281 load += active * (FIXED_1 - exp);
3282 load += 1UL << (FSHIFT - 1);
3283 return load >> FSHIFT;
3288 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3290 * When making the ILB scale, we should try to pull this in as well.
3292 static atomic_long_t calc_load_tasks_idle;
3294 static void calc_load_account_idle(struct rq *this_rq)
3298 delta = calc_load_fold_active(this_rq);
3300 atomic_long_add(delta, &calc_load_tasks_idle);
3303 static long calc_load_fold_idle(void)
3308 * Its got a race, we don't care...
3310 if (atomic_long_read(&calc_load_tasks_idle))
3311 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3317 * fixed_power_int - compute: x^n, in O(log n) time
3319 * @x: base of the power
3320 * @frac_bits: fractional bits of @x
3321 * @n: power to raise @x to.
3323 * By exploiting the relation between the definition of the natural power
3324 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3325 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3326 * (where: n_i \elem {0, 1}, the binary vector representing n),
3327 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3328 * of course trivially computable in O(log_2 n), the length of our binary
3331 static unsigned long
3332 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3334 unsigned long result = 1UL << frac_bits;
3339 result += 1UL << (frac_bits - 1);
3340 result >>= frac_bits;
3346 x += 1UL << (frac_bits - 1);
3354 * a1 = a0 * e + a * (1 - e)
3356 * a2 = a1 * e + a * (1 - e)
3357 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3358 * = a0 * e^2 + a * (1 - e) * (1 + e)
3360 * a3 = a2 * e + a * (1 - e)
3361 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3362 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3366 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3367 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3368 * = a0 * e^n + a * (1 - e^n)
3370 * [1] application of the geometric series:
3373 * S_n := \Sum x^i = -------------
3376 static unsigned long
3377 calc_load_n(unsigned long load, unsigned long exp,
3378 unsigned long active, unsigned int n)
3381 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3385 * NO_HZ can leave us missing all per-cpu ticks calling
3386 * calc_load_account_active(), but since an idle CPU folds its delta into
3387 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3388 * in the pending idle delta if our idle period crossed a load cycle boundary.
3390 * Once we've updated the global active value, we need to apply the exponential
3391 * weights adjusted to the number of cycles missed.
3393 static void calc_global_nohz(void)
3395 long delta, active, n;
3398 * If we crossed a calc_load_update boundary, make sure to fold
3399 * any pending idle changes, the respective CPUs might have
3400 * missed the tick driven calc_load_account_active() update
3403 delta = calc_load_fold_idle();
3405 atomic_long_add(delta, &calc_load_tasks);
3408 * It could be the one fold was all it took, we done!
3410 if (time_before(jiffies, calc_load_update + 10))
3414 * Catch-up, fold however many we are behind still
3416 delta = jiffies - calc_load_update - 10;
3417 n = 1 + (delta / LOAD_FREQ);
3419 active = atomic_long_read(&calc_load_tasks);
3420 active = active > 0 ? active * FIXED_1 : 0;
3422 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3423 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3424 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3426 calc_load_update += n * LOAD_FREQ;
3429 static void calc_load_account_idle(struct rq *this_rq)
3433 static inline long calc_load_fold_idle(void)
3438 static void calc_global_nohz(void)
3444 * get_avenrun - get the load average array
3445 * @loads: pointer to dest load array
3446 * @offset: offset to add
3447 * @shift: shift count to shift the result left
3449 * These values are estimates at best, so no need for locking.
3451 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3453 loads[0] = (avenrun[0] + offset) << shift;
3454 loads[1] = (avenrun[1] + offset) << shift;
3455 loads[2] = (avenrun[2] + offset) << shift;
3459 * calc_load - update the avenrun load estimates 10 ticks after the
3460 * CPUs have updated calc_load_tasks.
3462 void calc_global_load(unsigned long ticks)
3466 if (time_before(jiffies, calc_load_update + 10))
3469 active = atomic_long_read(&calc_load_tasks);
3470 active = active > 0 ? active * FIXED_1 : 0;
3472 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3473 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3474 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3476 calc_load_update += LOAD_FREQ;
3479 * Account one period with whatever state we found before
3480 * folding in the nohz state and ageing the entire idle period.
3482 * This avoids loosing a sample when we go idle between
3483 * calc_load_account_active() (10 ticks ago) and now and thus
3490 * Called from update_cpu_load() to periodically update this CPU's
3493 static void calc_load_account_active(struct rq *this_rq)
3497 if (time_before(jiffies, this_rq->calc_load_update))
3500 delta = calc_load_fold_active(this_rq);
3501 delta += calc_load_fold_idle();
3503 atomic_long_add(delta, &calc_load_tasks);
3505 this_rq->calc_load_update += LOAD_FREQ;
3509 * The exact cpuload at various idx values, calculated at every tick would be
3510 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3512 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3513 * on nth tick when cpu may be busy, then we have:
3514 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3515 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3517 * decay_load_missed() below does efficient calculation of
3518 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3519 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3521 * The calculation is approximated on a 128 point scale.
3522 * degrade_zero_ticks is the number of ticks after which load at any
3523 * particular idx is approximated to be zero.
3524 * degrade_factor is a precomputed table, a row for each load idx.
3525 * Each column corresponds to degradation factor for a power of two ticks,
3526 * based on 128 point scale.
3528 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3529 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3531 * With this power of 2 load factors, we can degrade the load n times
3532 * by looking at 1 bits in n and doing as many mult/shift instead of
3533 * n mult/shifts needed by the exact degradation.
3535 #define DEGRADE_SHIFT 7
3536 static const unsigned char
3537 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3538 static const unsigned char
3539 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3540 {0, 0, 0, 0, 0, 0, 0, 0},
3541 {64, 32, 8, 0, 0, 0, 0, 0},
3542 {96, 72, 40, 12, 1, 0, 0},
3543 {112, 98, 75, 43, 15, 1, 0},
3544 {120, 112, 98, 76, 45, 16, 2} };
3547 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3548 * would be when CPU is idle and so we just decay the old load without
3549 * adding any new load.
3551 static unsigned long
3552 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3556 if (!missed_updates)
3559 if (missed_updates >= degrade_zero_ticks[idx])
3563 return load >> missed_updates;
3565 while (missed_updates) {
3566 if (missed_updates % 2)
3567 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3569 missed_updates >>= 1;
3576 * Update rq->cpu_load[] statistics. This function is usually called every
3577 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3578 * every tick. We fix it up based on jiffies.
3580 static void update_cpu_load(struct rq *this_rq)
3582 unsigned long this_load = this_rq->load.weight;
3583 unsigned long curr_jiffies = jiffies;
3584 unsigned long pending_updates;
3587 this_rq->nr_load_updates++;
3589 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3590 if (curr_jiffies == this_rq->last_load_update_tick)
3593 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3594 this_rq->last_load_update_tick = curr_jiffies;
3596 /* Update our load: */
3597 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3598 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3599 unsigned long old_load, new_load;
3601 /* scale is effectively 1 << i now, and >> i divides by scale */
3603 old_load = this_rq->cpu_load[i];
3604 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3605 new_load = this_load;
3607 * Round up the averaging division if load is increasing. This
3608 * prevents us from getting stuck on 9 if the load is 10, for
3611 if (new_load > old_load)
3612 new_load += scale - 1;
3614 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3617 sched_avg_update(this_rq);
3620 static void update_cpu_load_active(struct rq *this_rq)
3622 update_cpu_load(this_rq);
3624 calc_load_account_active(this_rq);
3630 * sched_exec - execve() is a valuable balancing opportunity, because at
3631 * this point the task has the smallest effective memory and cache footprint.
3633 void sched_exec(void)
3635 struct task_struct *p = current;
3636 unsigned long flags;
3639 raw_spin_lock_irqsave(&p->pi_lock, flags);
3640 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3641 if (dest_cpu == smp_processor_id())
3644 if (likely(cpu_active(dest_cpu))) {
3645 struct migration_arg arg = { p, dest_cpu };
3647 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3648 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3652 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3657 DEFINE_PER_CPU(struct kernel_stat, kstat);
3659 EXPORT_PER_CPU_SYMBOL(kstat);
3662 * Return any ns on the sched_clock that have not yet been accounted in
3663 * @p in case that task is currently running.
3665 * Called with task_rq_lock() held on @rq.
3667 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3671 if (task_current(rq, p)) {
3672 update_rq_clock(rq);
3673 ns = rq->clock_task - p->se.exec_start;
3681 unsigned long long task_delta_exec(struct task_struct *p)
3683 unsigned long flags;
3687 rq = task_rq_lock(p, &flags);
3688 ns = do_task_delta_exec(p, rq);
3689 task_rq_unlock(rq, p, &flags);
3695 * Return accounted runtime for the task.
3696 * In case the task is currently running, return the runtime plus current's
3697 * pending runtime that have not been accounted yet.
3699 unsigned long long task_sched_runtime(struct task_struct *p)
3701 unsigned long flags;
3705 rq = task_rq_lock(p, &flags);
3706 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3707 task_rq_unlock(rq, p, &flags);
3713 * Account user cpu time to a process.
3714 * @p: the process that the cpu time gets accounted to
3715 * @cputime: the cpu time spent in user space since the last update
3716 * @cputime_scaled: cputime scaled by cpu frequency
3718 void account_user_time(struct task_struct *p, cputime_t cputime,
3719 cputime_t cputime_scaled)
3721 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3724 /* Add user time to process. */
3725 p->utime = cputime_add(p->utime, cputime);
3726 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3727 account_group_user_time(p, cputime);
3729 /* Add user time to cpustat. */
3730 tmp = cputime_to_cputime64(cputime);
3731 if (TASK_NICE(p) > 0)
3732 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3734 cpustat->user = cputime64_add(cpustat->user, tmp);
3736 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3737 /* Account for user time used */
3738 acct_update_integrals(p);
3742 * Account guest cpu time to a process.
3743 * @p: the process that the cpu time gets accounted to
3744 * @cputime: the cpu time spent in virtual machine since the last update
3745 * @cputime_scaled: cputime scaled by cpu frequency
3747 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3748 cputime_t cputime_scaled)
3751 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3753 tmp = cputime_to_cputime64(cputime);
3755 /* Add guest time to process. */
3756 p->utime = cputime_add(p->utime, cputime);
3757 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3758 account_group_user_time(p, cputime);
3759 p->gtime = cputime_add(p->gtime, cputime);
3761 /* Add guest time to cpustat. */
3762 if (TASK_NICE(p) > 0) {
3763 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3764 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3766 cpustat->user = cputime64_add(cpustat->user, tmp);
3767 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3772 * Account system cpu time to a process and desired cpustat field
3773 * @p: the process that the cpu time gets accounted to
3774 * @cputime: the cpu time spent in kernel space since the last update
3775 * @cputime_scaled: cputime scaled by cpu frequency
3776 * @target_cputime64: pointer to cpustat field that has to be updated
3779 void __account_system_time(struct task_struct *p, cputime_t cputime,
3780 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3782 cputime64_t tmp = cputime_to_cputime64(cputime);
3784 /* Add system time to process. */
3785 p->stime = cputime_add(p->stime, cputime);
3786 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3787 account_group_system_time(p, cputime);
3789 /* Add system time to cpustat. */
3790 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3791 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3793 /* Account for system time used */
3794 acct_update_integrals(p);
3798 * Account system cpu time to a process.
3799 * @p: the process that the cpu time gets accounted to
3800 * @hardirq_offset: the offset to subtract from hardirq_count()
3801 * @cputime: the cpu time spent in kernel space since the last update
3802 * @cputime_scaled: cputime scaled by cpu frequency
3804 void account_system_time(struct task_struct *p, int hardirq_offset,
3805 cputime_t cputime, cputime_t cputime_scaled)
3807 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3808 cputime64_t *target_cputime64;
3810 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3811 account_guest_time(p, cputime, cputime_scaled);
3815 if (hardirq_count() - hardirq_offset)
3816 target_cputime64 = &cpustat->irq;
3817 else if (in_serving_softirq())
3818 target_cputime64 = &cpustat->softirq;
3820 target_cputime64 = &cpustat->system;
3822 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3826 * Account for involuntary wait time.
3827 * @cputime: the cpu time spent in involuntary wait
3829 void account_steal_time(cputime_t cputime)
3831 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3832 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3834 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3838 * Account for idle time.
3839 * @cputime: the cpu time spent in idle wait
3841 void account_idle_time(cputime_t cputime)
3843 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3844 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3845 struct rq *rq = this_rq();
3847 if (atomic_read(&rq->nr_iowait) > 0)
3848 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3850 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3853 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3855 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3857 * Account a tick to a process and cpustat
3858 * @p: the process that the cpu time gets accounted to
3859 * @user_tick: is the tick from userspace
3860 * @rq: the pointer to rq
3862 * Tick demultiplexing follows the order
3863 * - pending hardirq update
3864 * - pending softirq update
3868 * - check for guest_time
3869 * - else account as system_time
3871 * Check for hardirq is done both for system and user time as there is
3872 * no timer going off while we are on hardirq and hence we may never get an
3873 * opportunity to update it solely in system time.
3874 * p->stime and friends are only updated on system time and not on irq
3875 * softirq as those do not count in task exec_runtime any more.
3877 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3880 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3881 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
3882 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3884 if (irqtime_account_hi_update()) {
3885 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3886 } else if (irqtime_account_si_update()) {
3887 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3888 } else if (this_cpu_ksoftirqd() == p) {
3890 * ksoftirqd time do not get accounted in cpu_softirq_time.
3891 * So, we have to handle it separately here.
3892 * Also, p->stime needs to be updated for ksoftirqd.
3894 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3896 } else if (user_tick) {
3897 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3898 } else if (p == rq->idle) {
3899 account_idle_time(cputime_one_jiffy);
3900 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3901 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3903 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3908 static void irqtime_account_idle_ticks(int ticks)
3911 struct rq *rq = this_rq();
3913 for (i = 0; i < ticks; i++)
3914 irqtime_account_process_tick(current, 0, rq);
3916 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
3917 static void irqtime_account_idle_ticks(int ticks) {}
3918 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3920 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3923 * Account a single tick of cpu time.
3924 * @p: the process that the cpu time gets accounted to
3925 * @user_tick: indicates if the tick is a user or a system tick
3927 void account_process_tick(struct task_struct *p, int user_tick)
3929 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3930 struct rq *rq = this_rq();
3932 if (sched_clock_irqtime) {
3933 irqtime_account_process_tick(p, user_tick, rq);
3938 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3939 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3940 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3943 account_idle_time(cputime_one_jiffy);
3947 * Account multiple ticks of steal time.
3948 * @p: the process from which the cpu time has been stolen
3949 * @ticks: number of stolen ticks
3951 void account_steal_ticks(unsigned long ticks)
3953 account_steal_time(jiffies_to_cputime(ticks));
3957 * Account multiple ticks of idle time.
3958 * @ticks: number of stolen ticks
3960 void account_idle_ticks(unsigned long ticks)
3963 if (sched_clock_irqtime) {
3964 irqtime_account_idle_ticks(ticks);
3968 account_idle_time(jiffies_to_cputime(ticks));
3974 * Use precise platform statistics if available:
3976 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3977 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3983 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3985 struct task_cputime cputime;
3987 thread_group_cputime(p, &cputime);
3989 *ut = cputime.utime;
3990 *st = cputime.stime;
3994 #ifndef nsecs_to_cputime
3995 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3998 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4000 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
4003 * Use CFS's precise accounting:
4005 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
4011 do_div(temp, total);
4012 utime = (cputime_t)temp;
4017 * Compare with previous values, to keep monotonicity:
4019 p->prev_utime = max(p->prev_utime, utime);
4020 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4022 *ut = p->prev_utime;
4023 *st = p->prev_stime;
4027 * Must be called with siglock held.
4029 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4031 struct signal_struct *sig = p->signal;
4032 struct task_cputime cputime;
4033 cputime_t rtime, utime, total;
4035 thread_group_cputime(p, &cputime);
4037 total = cputime_add(cputime.utime, cputime.stime);
4038 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4043 temp *= cputime.utime;
4044 do_div(temp, total);
4045 utime = (cputime_t)temp;
4049 sig->prev_utime = max(sig->prev_utime, utime);
4050 sig->prev_stime = max(sig->prev_stime,
4051 cputime_sub(rtime, sig->prev_utime));
4053 *ut = sig->prev_utime;
4054 *st = sig->prev_stime;
4059 * This function gets called by the timer code, with HZ frequency.
4060 * We call it with interrupts disabled.
4062 void scheduler_tick(void)
4064 int cpu = smp_processor_id();
4065 struct rq *rq = cpu_rq(cpu);
4066 struct task_struct *curr = rq->curr;
4070 raw_spin_lock(&rq->lock);
4071 update_rq_clock(rq);
4072 update_cpu_load_active(rq);
4073 curr->sched_class->task_tick(rq, curr, 0);
4074 raw_spin_unlock(&rq->lock);
4076 perf_event_task_tick();
4079 rq->idle_at_tick = idle_cpu(cpu);
4080 trigger_load_balance(rq, cpu);
4084 notrace unsigned long get_parent_ip(unsigned long addr)
4086 if (in_lock_functions(addr)) {
4087 addr = CALLER_ADDR2;
4088 if (in_lock_functions(addr))
4089 addr = CALLER_ADDR3;
4094 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4095 defined(CONFIG_PREEMPT_TRACER))
4097 void __kprobes add_preempt_count(int val)
4099 #ifdef CONFIG_DEBUG_PREEMPT
4103 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4106 preempt_count() += val;
4107 #ifdef CONFIG_DEBUG_PREEMPT
4109 * Spinlock count overflowing soon?
4111 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4114 if (preempt_count() == val)
4115 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4117 EXPORT_SYMBOL(add_preempt_count);
4119 void __kprobes sub_preempt_count(int val)
4121 #ifdef CONFIG_DEBUG_PREEMPT
4125 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4128 * Is the spinlock portion underflowing?
4130 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4131 !(preempt_count() & PREEMPT_MASK)))
4135 if (preempt_count() == val)
4136 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4137 preempt_count() -= val;
4139 EXPORT_SYMBOL(sub_preempt_count);
4144 * Print scheduling while atomic bug:
4146 static noinline void __schedule_bug(struct task_struct *prev)
4148 struct pt_regs *regs = get_irq_regs();
4150 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4151 prev->comm, prev->pid, preempt_count());
4153 debug_show_held_locks(prev);
4155 if (irqs_disabled())
4156 print_irqtrace_events(prev);
4165 * Various schedule()-time debugging checks and statistics:
4167 static inline void schedule_debug(struct task_struct *prev)
4170 * Test if we are atomic. Since do_exit() needs to call into
4171 * schedule() atomically, we ignore that path for now.
4172 * Otherwise, whine if we are scheduling when we should not be.
4174 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4175 __schedule_bug(prev);
4177 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4179 schedstat_inc(this_rq(), sched_count);
4182 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4184 if (prev->on_rq || rq->skip_clock_update < 0)
4185 update_rq_clock(rq);
4186 prev->sched_class->put_prev_task(rq, prev);
4190 * Pick up the highest-prio task:
4192 static inline struct task_struct *
4193 pick_next_task(struct rq *rq)
4195 const struct sched_class *class;
4196 struct task_struct *p;
4199 * Optimization: we know that if all tasks are in
4200 * the fair class we can call that function directly:
4202 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4203 p = fair_sched_class.pick_next_task(rq);
4208 for_each_class(class) {
4209 p = class->pick_next_task(rq);
4214 BUG(); /* the idle class will always have a runnable task */
4218 * __schedule() is the main scheduler function.
4220 static void __sched __schedule(void)
4222 struct task_struct *prev, *next;
4223 unsigned long *switch_count;
4229 cpu = smp_processor_id();
4231 rcu_note_context_switch(cpu);
4234 schedule_debug(prev);
4236 if (sched_feat(HRTICK))
4239 raw_spin_lock_irq(&rq->lock);
4241 switch_count = &prev->nivcsw;
4242 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4243 if (unlikely(signal_pending_state(prev->state, prev))) {
4244 prev->state = TASK_RUNNING;
4246 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4250 * If a worker went to sleep, notify and ask workqueue
4251 * whether it wants to wake up a task to maintain
4254 if (prev->flags & PF_WQ_WORKER) {
4255 struct task_struct *to_wakeup;
4257 to_wakeup = wq_worker_sleeping(prev, cpu);
4259 try_to_wake_up_local(to_wakeup);
4262 switch_count = &prev->nvcsw;
4265 pre_schedule(rq, prev);
4267 if (unlikely(!rq->nr_running))
4268 idle_balance(cpu, rq);
4270 put_prev_task(rq, prev);
4271 next = pick_next_task(rq);
4272 clear_tsk_need_resched(prev);
4273 rq->skip_clock_update = 0;
4275 if (likely(prev != next)) {
4280 context_switch(rq, prev, next); /* unlocks the rq */
4282 * The context switch have flipped the stack from under us
4283 * and restored the local variables which were saved when
4284 * this task called schedule() in the past. prev == current
4285 * is still correct, but it can be moved to another cpu/rq.
4287 cpu = smp_processor_id();
4290 raw_spin_unlock_irq(&rq->lock);
4294 preempt_enable_no_resched();
4299 static inline void sched_submit_work(struct task_struct *tsk)
4304 * If we are going to sleep and we have plugged IO queued,
4305 * make sure to submit it to avoid deadlocks.
4307 if (blk_needs_flush_plug(tsk))
4308 blk_schedule_flush_plug(tsk);
4311 asmlinkage void __sched schedule(void)
4313 struct task_struct *tsk = current;
4315 sched_submit_work(tsk);
4318 EXPORT_SYMBOL(schedule);
4320 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4322 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4327 if (lock->owner != owner)
4331 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4332 * lock->owner still matches owner, if that fails, owner might
4333 * point to free()d memory, if it still matches, the rcu_read_lock()
4334 * ensures the memory stays valid.
4338 ret = owner->on_cpu;
4346 * Look out! "owner" is an entirely speculative pointer
4347 * access and not reliable.
4349 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4351 if (!sched_feat(OWNER_SPIN))
4354 while (owner_running(lock, owner)) {
4358 arch_mutex_cpu_relax();
4362 * If the owner changed to another task there is likely
4363 * heavy contention, stop spinning.
4372 #ifdef CONFIG_PREEMPT
4374 * this is the entry point to schedule() from in-kernel preemption
4375 * off of preempt_enable. Kernel preemptions off return from interrupt
4376 * occur there and call schedule directly.
4378 asmlinkage void __sched notrace preempt_schedule(void)
4380 struct thread_info *ti = current_thread_info();
4383 * If there is a non-zero preempt_count or interrupts are disabled,
4384 * we do not want to preempt the current task. Just return..
4386 if (likely(ti->preempt_count || irqs_disabled()))
4390 add_preempt_count_notrace(PREEMPT_ACTIVE);
4392 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4395 * Check again in case we missed a preemption opportunity
4396 * between schedule and now.
4399 } while (need_resched());
4401 EXPORT_SYMBOL(preempt_schedule);
4404 * this is the entry point to schedule() from kernel preemption
4405 * off of irq context.
4406 * Note, that this is called and return with irqs disabled. This will
4407 * protect us against recursive calling from irq.
4409 asmlinkage void __sched preempt_schedule_irq(void)
4411 struct thread_info *ti = current_thread_info();
4413 /* Catch callers which need to be fixed */
4414 BUG_ON(ti->preempt_count || !irqs_disabled());
4417 add_preempt_count(PREEMPT_ACTIVE);
4420 local_irq_disable();
4421 sub_preempt_count(PREEMPT_ACTIVE);
4424 * Check again in case we missed a preemption opportunity
4425 * between schedule and now.
4428 } while (need_resched());
4431 #endif /* CONFIG_PREEMPT */
4433 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4436 return try_to_wake_up(curr->private, mode, wake_flags);
4438 EXPORT_SYMBOL(default_wake_function);
4441 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4442 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4443 * number) then we wake all the non-exclusive tasks and one exclusive task.
4445 * There are circumstances in which we can try to wake a task which has already
4446 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4447 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4449 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4450 int nr_exclusive, int wake_flags, void *key)
4452 wait_queue_t *curr, *next;
4454 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4455 unsigned flags = curr->flags;
4457 if (curr->func(curr, mode, wake_flags, key) &&
4458 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4464 * __wake_up - wake up threads blocked on a waitqueue.
4466 * @mode: which threads
4467 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4468 * @key: is directly passed to the wakeup function
4470 * It may be assumed that this function implies a write memory barrier before
4471 * changing the task state if and only if any tasks are woken up.
4473 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4474 int nr_exclusive, void *key)
4476 unsigned long flags;
4478 spin_lock_irqsave(&q->lock, flags);
4479 __wake_up_common(q, mode, nr_exclusive, 0, key);
4480 spin_unlock_irqrestore(&q->lock, flags);
4482 EXPORT_SYMBOL(__wake_up);
4485 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4487 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4489 __wake_up_common(q, mode, 1, 0, NULL);
4491 EXPORT_SYMBOL_GPL(__wake_up_locked);
4493 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4495 __wake_up_common(q, mode, 1, 0, key);
4497 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4500 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4502 * @mode: which threads
4503 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4504 * @key: opaque value to be passed to wakeup targets
4506 * The sync wakeup differs that the waker knows that it will schedule
4507 * away soon, so while the target thread will be woken up, it will not
4508 * be migrated to another CPU - ie. the two threads are 'synchronized'
4509 * with each other. This can prevent needless bouncing between CPUs.
4511 * On UP it can prevent extra preemption.
4513 * It may be assumed that this function implies a write memory barrier before
4514 * changing the task state if and only if any tasks are woken up.
4516 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4517 int nr_exclusive, void *key)
4519 unsigned long flags;
4520 int wake_flags = WF_SYNC;
4525 if (unlikely(!nr_exclusive))
4528 spin_lock_irqsave(&q->lock, flags);
4529 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4530 spin_unlock_irqrestore(&q->lock, flags);
4532 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4535 * __wake_up_sync - see __wake_up_sync_key()
4537 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4539 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4541 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4544 * complete: - signals a single thread waiting on this completion
4545 * @x: holds the state of this particular completion
4547 * This will wake up a single thread waiting on this completion. Threads will be
4548 * awakened in the same order in which they were queued.
4550 * See also complete_all(), wait_for_completion() and related routines.
4552 * It may be assumed that this function implies a write memory barrier before
4553 * changing the task state if and only if any tasks are woken up.
4555 void complete(struct completion *x)
4557 unsigned long flags;
4559 spin_lock_irqsave(&x->wait.lock, flags);
4561 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4562 spin_unlock_irqrestore(&x->wait.lock, flags);
4564 EXPORT_SYMBOL(complete);
4567 * complete_all: - signals all threads waiting on this completion
4568 * @x: holds the state of this particular completion
4570 * This will wake up all threads waiting on this particular completion event.
4572 * It may be assumed that this function implies a write memory barrier before
4573 * changing the task state if and only if any tasks are woken up.
4575 void complete_all(struct completion *x)
4577 unsigned long flags;
4579 spin_lock_irqsave(&x->wait.lock, flags);
4580 x->done += UINT_MAX/2;
4581 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4582 spin_unlock_irqrestore(&x->wait.lock, flags);
4584 EXPORT_SYMBOL(complete_all);
4586 static inline long __sched
4587 do_wait_for_common(struct completion *x, long timeout, int state)
4590 DECLARE_WAITQUEUE(wait, current);
4592 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4594 if (signal_pending_state(state, current)) {
4595 timeout = -ERESTARTSYS;
4598 __set_current_state(state);
4599 spin_unlock_irq(&x->wait.lock);
4600 timeout = schedule_timeout(timeout);
4601 spin_lock_irq(&x->wait.lock);
4602 } while (!x->done && timeout);
4603 __remove_wait_queue(&x->wait, &wait);
4608 return timeout ?: 1;
4612 wait_for_common(struct completion *x, long timeout, int state)
4616 spin_lock_irq(&x->wait.lock);
4617 timeout = do_wait_for_common(x, timeout, state);
4618 spin_unlock_irq(&x->wait.lock);
4623 * wait_for_completion: - waits for completion of a task
4624 * @x: holds the state of this particular completion
4626 * This waits to be signaled for completion of a specific task. It is NOT
4627 * interruptible and there is no timeout.
4629 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4630 * and interrupt capability. Also see complete().
4632 void __sched wait_for_completion(struct completion *x)
4634 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4636 EXPORT_SYMBOL(wait_for_completion);
4639 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4640 * @x: holds the state of this particular completion
4641 * @timeout: timeout value in jiffies
4643 * This waits for either a completion of a specific task to be signaled or for a
4644 * specified timeout to expire. The timeout is in jiffies. It is not
4647 unsigned long __sched
4648 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4650 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4652 EXPORT_SYMBOL(wait_for_completion_timeout);
4655 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4656 * @x: holds the state of this particular completion
4658 * This waits for completion of a specific task to be signaled. It is
4661 int __sched wait_for_completion_interruptible(struct completion *x)
4663 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4664 if (t == -ERESTARTSYS)
4668 EXPORT_SYMBOL(wait_for_completion_interruptible);
4671 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4672 * @x: holds the state of this particular completion
4673 * @timeout: timeout value in jiffies
4675 * This waits for either a completion of a specific task to be signaled or for a
4676 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4679 wait_for_completion_interruptible_timeout(struct completion *x,
4680 unsigned long timeout)
4682 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4684 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4687 * wait_for_completion_killable: - waits for completion of a task (killable)
4688 * @x: holds the state of this particular completion
4690 * This waits to be signaled for completion of a specific task. It can be
4691 * interrupted by a kill signal.
4693 int __sched wait_for_completion_killable(struct completion *x)
4695 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4696 if (t == -ERESTARTSYS)
4700 EXPORT_SYMBOL(wait_for_completion_killable);
4703 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4704 * @x: holds the state of this particular completion
4705 * @timeout: timeout value in jiffies
4707 * This waits for either a completion of a specific task to be
4708 * signaled or for a specified timeout to expire. It can be
4709 * interrupted by a kill signal. The timeout is in jiffies.
4712 wait_for_completion_killable_timeout(struct completion *x,
4713 unsigned long timeout)
4715 return wait_for_common(x, timeout, TASK_KILLABLE);
4717 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4720 * try_wait_for_completion - try to decrement a completion without blocking
4721 * @x: completion structure
4723 * Returns: 0 if a decrement cannot be done without blocking
4724 * 1 if a decrement succeeded.
4726 * If a completion is being used as a counting completion,
4727 * attempt to decrement the counter without blocking. This
4728 * enables us to avoid waiting if the resource the completion
4729 * is protecting is not available.
4731 bool try_wait_for_completion(struct completion *x)
4733 unsigned long flags;
4736 spin_lock_irqsave(&x->wait.lock, flags);
4741 spin_unlock_irqrestore(&x->wait.lock, flags);
4744 EXPORT_SYMBOL(try_wait_for_completion);
4747 * completion_done - Test to see if a completion has any waiters
4748 * @x: completion structure
4750 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4751 * 1 if there are no waiters.
4754 bool completion_done(struct completion *x)
4756 unsigned long flags;
4759 spin_lock_irqsave(&x->wait.lock, flags);
4762 spin_unlock_irqrestore(&x->wait.lock, flags);
4765 EXPORT_SYMBOL(completion_done);
4768 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4770 unsigned long flags;
4773 init_waitqueue_entry(&wait, current);
4775 __set_current_state(state);
4777 spin_lock_irqsave(&q->lock, flags);
4778 __add_wait_queue(q, &wait);
4779 spin_unlock(&q->lock);
4780 timeout = schedule_timeout(timeout);
4781 spin_lock_irq(&q->lock);
4782 __remove_wait_queue(q, &wait);
4783 spin_unlock_irqrestore(&q->lock, flags);
4788 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4790 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4792 EXPORT_SYMBOL(interruptible_sleep_on);
4795 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4797 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4799 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4801 void __sched sleep_on(wait_queue_head_t *q)
4803 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4805 EXPORT_SYMBOL(sleep_on);
4807 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4809 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4811 EXPORT_SYMBOL(sleep_on_timeout);
4813 #ifdef CONFIG_RT_MUTEXES
4816 * rt_mutex_setprio - set the current priority of a task
4818 * @prio: prio value (kernel-internal form)
4820 * This function changes the 'effective' priority of a task. It does
4821 * not touch ->normal_prio like __setscheduler().
4823 * Used by the rt_mutex code to implement priority inheritance logic.
4825 void rt_mutex_setprio(struct task_struct *p, int prio)
4827 int oldprio, on_rq, running;
4829 const struct sched_class *prev_class;
4831 BUG_ON(prio < 0 || prio > MAX_PRIO);
4833 rq = __task_rq_lock(p);
4835 trace_sched_pi_setprio(p, prio);
4837 prev_class = p->sched_class;
4839 running = task_current(rq, p);
4841 dequeue_task(rq, p, 0);
4843 p->sched_class->put_prev_task(rq, p);
4846 p->sched_class = &rt_sched_class;
4848 p->sched_class = &fair_sched_class;
4853 p->sched_class->set_curr_task(rq);
4855 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4857 check_class_changed(rq, p, prev_class, oldprio);
4858 __task_rq_unlock(rq);
4863 void set_user_nice(struct task_struct *p, long nice)
4865 int old_prio, delta, on_rq;
4866 unsigned long flags;
4869 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4872 * We have to be careful, if called from sys_setpriority(),
4873 * the task might be in the middle of scheduling on another CPU.
4875 rq = task_rq_lock(p, &flags);
4877 * The RT priorities are set via sched_setscheduler(), but we still
4878 * allow the 'normal' nice value to be set - but as expected
4879 * it wont have any effect on scheduling until the task is
4880 * SCHED_FIFO/SCHED_RR:
4882 if (task_has_rt_policy(p)) {
4883 p->static_prio = NICE_TO_PRIO(nice);
4888 dequeue_task(rq, p, 0);
4890 p->static_prio = NICE_TO_PRIO(nice);
4893 p->prio = effective_prio(p);
4894 delta = p->prio - old_prio;
4897 enqueue_task(rq, p, 0);
4899 * If the task increased its priority or is running and
4900 * lowered its priority, then reschedule its CPU:
4902 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4903 resched_task(rq->curr);
4906 task_rq_unlock(rq, p, &flags);
4908 EXPORT_SYMBOL(set_user_nice);
4911 * can_nice - check if a task can reduce its nice value
4915 int can_nice(const struct task_struct *p, const int nice)
4917 /* convert nice value [19,-20] to rlimit style value [1,40] */
4918 int nice_rlim = 20 - nice;
4920 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4921 capable(CAP_SYS_NICE));
4924 #ifdef __ARCH_WANT_SYS_NICE
4927 * sys_nice - change the priority of the current process.
4928 * @increment: priority increment
4930 * sys_setpriority is a more generic, but much slower function that
4931 * does similar things.
4933 SYSCALL_DEFINE1(nice, int, increment)
4938 * Setpriority might change our priority at the same moment.
4939 * We don't have to worry. Conceptually one call occurs first
4940 * and we have a single winner.
4942 if (increment < -40)
4947 nice = TASK_NICE(current) + increment;
4953 if (increment < 0 && !can_nice(current, nice))
4956 retval = security_task_setnice(current, nice);
4960 set_user_nice(current, nice);
4967 * task_prio - return the priority value of a given task.
4968 * @p: the task in question.
4970 * This is the priority value as seen by users in /proc.
4971 * RT tasks are offset by -200. Normal tasks are centered
4972 * around 0, value goes from -16 to +15.
4974 int task_prio(const struct task_struct *p)
4976 return p->prio - MAX_RT_PRIO;
4980 * task_nice - return the nice value of a given task.
4981 * @p: the task in question.
4983 int task_nice(const struct task_struct *p)
4985 return TASK_NICE(p);
4987 EXPORT_SYMBOL(task_nice);
4990 * idle_cpu - is a given cpu idle currently?
4991 * @cpu: the processor in question.
4993 int idle_cpu(int cpu)
4995 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4999 * idle_task - return the idle task for a given cpu.
5000 * @cpu: the processor in question.
5002 struct task_struct *idle_task(int cpu)
5004 return cpu_rq(cpu)->idle;
5008 * find_process_by_pid - find a process with a matching PID value.
5009 * @pid: the pid in question.
5011 static struct task_struct *find_process_by_pid(pid_t pid)
5013 return pid ? find_task_by_vpid(pid) : current;
5016 /* Actually do priority change: must hold rq lock. */
5018 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5021 p->rt_priority = prio;
5022 p->normal_prio = normal_prio(p);
5023 /* we are holding p->pi_lock already */
5024 p->prio = rt_mutex_getprio(p);
5025 if (rt_prio(p->prio))
5026 p->sched_class = &rt_sched_class;
5028 p->sched_class = &fair_sched_class;
5033 * check the target process has a UID that matches the current process's
5035 static bool check_same_owner(struct task_struct *p)
5037 const struct cred *cred = current_cred(), *pcred;
5041 pcred = __task_cred(p);
5042 if (cred->user->user_ns == pcred->user->user_ns)
5043 match = (cred->euid == pcred->euid ||
5044 cred->euid == pcred->uid);
5051 static int __sched_setscheduler(struct task_struct *p, int policy,
5052 const struct sched_param *param, bool user)
5054 int retval, oldprio, oldpolicy = -1, on_rq, running;
5055 unsigned long flags;
5056 const struct sched_class *prev_class;
5060 /* may grab non-irq protected spin_locks */
5061 BUG_ON(in_interrupt());
5063 /* double check policy once rq lock held */
5065 reset_on_fork = p->sched_reset_on_fork;
5066 policy = oldpolicy = p->policy;
5068 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5069 policy &= ~SCHED_RESET_ON_FORK;
5071 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5072 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5073 policy != SCHED_IDLE)
5078 * Valid priorities for SCHED_FIFO and SCHED_RR are
5079 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5080 * SCHED_BATCH and SCHED_IDLE is 0.
5082 if (param->sched_priority < 0 ||
5083 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5084 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5086 if (rt_policy(policy) != (param->sched_priority != 0))
5090 * Allow unprivileged RT tasks to decrease priority:
5092 if (user && !capable(CAP_SYS_NICE)) {
5093 if (rt_policy(policy)) {
5094 unsigned long rlim_rtprio =
5095 task_rlimit(p, RLIMIT_RTPRIO);
5097 /* can't set/change the rt policy */
5098 if (policy != p->policy && !rlim_rtprio)
5101 /* can't increase priority */
5102 if (param->sched_priority > p->rt_priority &&
5103 param->sched_priority > rlim_rtprio)
5108 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5109 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5111 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5112 if (!can_nice(p, TASK_NICE(p)))
5116 /* can't change other user's priorities */
5117 if (!check_same_owner(p))
5120 /* Normal users shall not reset the sched_reset_on_fork flag */
5121 if (p->sched_reset_on_fork && !reset_on_fork)
5126 retval = security_task_setscheduler(p);
5132 * make sure no PI-waiters arrive (or leave) while we are
5133 * changing the priority of the task:
5135 * To be able to change p->policy safely, the appropriate
5136 * runqueue lock must be held.
5138 rq = task_rq_lock(p, &flags);
5141 * Changing the policy of the stop threads its a very bad idea
5143 if (p == rq->stop) {
5144 task_rq_unlock(rq, p, &flags);
5149 * If not changing anything there's no need to proceed further:
5151 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5152 param->sched_priority == p->rt_priority))) {
5154 __task_rq_unlock(rq);
5155 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5159 #ifdef CONFIG_RT_GROUP_SCHED
5162 * Do not allow realtime tasks into groups that have no runtime
5165 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5166 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5167 !task_group_is_autogroup(task_group(p))) {
5168 task_rq_unlock(rq, p, &flags);
5174 /* recheck policy now with rq lock held */
5175 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5176 policy = oldpolicy = -1;
5177 task_rq_unlock(rq, p, &flags);
5181 running = task_current(rq, p);
5183 deactivate_task(rq, p, 0);
5185 p->sched_class->put_prev_task(rq, p);
5187 p->sched_reset_on_fork = reset_on_fork;
5190 prev_class = p->sched_class;
5191 __setscheduler(rq, p, policy, param->sched_priority);
5194 p->sched_class->set_curr_task(rq);
5196 activate_task(rq, p, 0);
5198 check_class_changed(rq, p, prev_class, oldprio);
5199 task_rq_unlock(rq, p, &flags);
5201 rt_mutex_adjust_pi(p);
5207 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5208 * @p: the task in question.
5209 * @policy: new policy.
5210 * @param: structure containing the new RT priority.
5212 * NOTE that the task may be already dead.
5214 int sched_setscheduler(struct task_struct *p, int policy,
5215 const struct sched_param *param)
5217 return __sched_setscheduler(p, policy, param, true);
5219 EXPORT_SYMBOL_GPL(sched_setscheduler);
5222 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5223 * @p: the task in question.
5224 * @policy: new policy.
5225 * @param: structure containing the new RT priority.
5227 * Just like sched_setscheduler, only don't bother checking if the
5228 * current context has permission. For example, this is needed in
5229 * stop_machine(): we create temporary high priority worker threads,
5230 * but our caller might not have that capability.
5232 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5233 const struct sched_param *param)
5235 return __sched_setscheduler(p, policy, param, false);
5239 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5241 struct sched_param lparam;
5242 struct task_struct *p;
5245 if (!param || pid < 0)
5247 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5252 p = find_process_by_pid(pid);
5254 retval = sched_setscheduler(p, policy, &lparam);
5261 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5262 * @pid: the pid in question.
5263 * @policy: new policy.
5264 * @param: structure containing the new RT priority.
5266 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5267 struct sched_param __user *, param)
5269 /* negative values for policy are not valid */
5273 return do_sched_setscheduler(pid, policy, param);
5277 * sys_sched_setparam - set/change the RT priority of a thread
5278 * @pid: the pid in question.
5279 * @param: structure containing the new RT priority.
5281 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5283 return do_sched_setscheduler(pid, -1, param);
5287 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5288 * @pid: the pid in question.
5290 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5292 struct task_struct *p;
5300 p = find_process_by_pid(pid);
5302 retval = security_task_getscheduler(p);
5305 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5312 * sys_sched_getparam - get the RT priority of a thread
5313 * @pid: the pid in question.
5314 * @param: structure containing the RT priority.
5316 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5318 struct sched_param lp;
5319 struct task_struct *p;
5322 if (!param || pid < 0)
5326 p = find_process_by_pid(pid);
5331 retval = security_task_getscheduler(p);
5335 lp.sched_priority = p->rt_priority;
5339 * This one might sleep, we cannot do it with a spinlock held ...
5341 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5350 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5352 cpumask_var_t cpus_allowed, new_mask;
5353 struct task_struct *p;
5359 p = find_process_by_pid(pid);
5366 /* Prevent p going away */
5370 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5374 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5376 goto out_free_cpus_allowed;
5379 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5382 retval = security_task_setscheduler(p);
5386 cpuset_cpus_allowed(p, cpus_allowed);
5387 cpumask_and(new_mask, in_mask, cpus_allowed);
5389 retval = set_cpus_allowed_ptr(p, new_mask);
5392 cpuset_cpus_allowed(p, cpus_allowed);
5393 if (!cpumask_subset(new_mask, cpus_allowed)) {
5395 * We must have raced with a concurrent cpuset
5396 * update. Just reset the cpus_allowed to the
5397 * cpuset's cpus_allowed
5399 cpumask_copy(new_mask, cpus_allowed);
5404 free_cpumask_var(new_mask);
5405 out_free_cpus_allowed:
5406 free_cpumask_var(cpus_allowed);
5413 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5414 struct cpumask *new_mask)
5416 if (len < cpumask_size())
5417 cpumask_clear(new_mask);
5418 else if (len > cpumask_size())
5419 len = cpumask_size();
5421 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5425 * sys_sched_setaffinity - set the cpu affinity of a process
5426 * @pid: pid of the process
5427 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5428 * @user_mask_ptr: user-space pointer to the new cpu mask
5430 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5431 unsigned long __user *, user_mask_ptr)
5433 cpumask_var_t new_mask;
5436 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5439 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5441 retval = sched_setaffinity(pid, new_mask);
5442 free_cpumask_var(new_mask);
5446 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5448 struct task_struct *p;
5449 unsigned long flags;
5456 p = find_process_by_pid(pid);
5460 retval = security_task_getscheduler(p);
5464 raw_spin_lock_irqsave(&p->pi_lock, flags);
5465 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5466 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5476 * sys_sched_getaffinity - get the cpu affinity of a process
5477 * @pid: pid of the process
5478 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5479 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5481 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5482 unsigned long __user *, user_mask_ptr)
5487 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5489 if (len & (sizeof(unsigned long)-1))
5492 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5495 ret = sched_getaffinity(pid, mask);
5497 size_t retlen = min_t(size_t, len, cpumask_size());
5499 if (copy_to_user(user_mask_ptr, mask, retlen))
5504 free_cpumask_var(mask);
5510 * sys_sched_yield - yield the current processor to other threads.
5512 * This function yields the current CPU to other tasks. If there are no
5513 * other threads running on this CPU then this function will return.
5515 SYSCALL_DEFINE0(sched_yield)
5517 struct rq *rq = this_rq_lock();
5519 schedstat_inc(rq, yld_count);
5520 current->sched_class->yield_task(rq);
5523 * Since we are going to call schedule() anyway, there's
5524 * no need to preempt or enable interrupts:
5526 __release(rq->lock);
5527 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5528 do_raw_spin_unlock(&rq->lock);
5529 preempt_enable_no_resched();
5536 static inline int should_resched(void)
5538 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5541 static void __cond_resched(void)
5543 add_preempt_count(PREEMPT_ACTIVE);
5545 sub_preempt_count(PREEMPT_ACTIVE);
5548 int __sched _cond_resched(void)
5550 if (should_resched()) {
5556 EXPORT_SYMBOL(_cond_resched);
5559 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5560 * call schedule, and on return reacquire the lock.
5562 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5563 * operations here to prevent schedule() from being called twice (once via
5564 * spin_unlock(), once by hand).
5566 int __cond_resched_lock(spinlock_t *lock)
5568 int resched = should_resched();
5571 lockdep_assert_held(lock);
5573 if (spin_needbreak(lock) || resched) {
5584 EXPORT_SYMBOL(__cond_resched_lock);
5586 int __sched __cond_resched_softirq(void)
5588 BUG_ON(!in_softirq());
5590 if (should_resched()) {
5598 EXPORT_SYMBOL(__cond_resched_softirq);
5601 * yield - yield the current processor to other threads.
5603 * This is a shortcut for kernel-space yielding - it marks the
5604 * thread runnable and calls sys_sched_yield().
5606 void __sched yield(void)
5608 set_current_state(TASK_RUNNING);
5611 EXPORT_SYMBOL(yield);
5614 * yield_to - yield the current processor to another thread in
5615 * your thread group, or accelerate that thread toward the
5616 * processor it's on.
5618 * @preempt: whether task preemption is allowed or not
5620 * It's the caller's job to ensure that the target task struct
5621 * can't go away on us before we can do any checks.
5623 * Returns true if we indeed boosted the target task.
5625 bool __sched yield_to(struct task_struct *p, bool preempt)
5627 struct task_struct *curr = current;
5628 struct rq *rq, *p_rq;
5629 unsigned long flags;
5632 local_irq_save(flags);
5637 double_rq_lock(rq, p_rq);
5638 while (task_rq(p) != p_rq) {
5639 double_rq_unlock(rq, p_rq);
5643 if (!curr->sched_class->yield_to_task)
5646 if (curr->sched_class != p->sched_class)
5649 if (task_running(p_rq, p) || p->state)
5652 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5654 schedstat_inc(rq, yld_count);
5656 * Make p's CPU reschedule; pick_next_entity takes care of
5659 if (preempt && rq != p_rq)
5660 resched_task(p_rq->curr);
5664 double_rq_unlock(rq, p_rq);
5665 local_irq_restore(flags);
5672 EXPORT_SYMBOL_GPL(yield_to);
5675 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5676 * that process accounting knows that this is a task in IO wait state.
5678 void __sched io_schedule(void)
5680 struct rq *rq = raw_rq();
5682 delayacct_blkio_start();
5683 atomic_inc(&rq->nr_iowait);
5684 blk_flush_plug(current);
5685 current->in_iowait = 1;
5687 current->in_iowait = 0;
5688 atomic_dec(&rq->nr_iowait);
5689 delayacct_blkio_end();
5691 EXPORT_SYMBOL(io_schedule);
5693 long __sched io_schedule_timeout(long timeout)
5695 struct rq *rq = raw_rq();
5698 delayacct_blkio_start();
5699 atomic_inc(&rq->nr_iowait);
5700 blk_flush_plug(current);
5701 current->in_iowait = 1;
5702 ret = schedule_timeout(timeout);
5703 current->in_iowait = 0;
5704 atomic_dec(&rq->nr_iowait);
5705 delayacct_blkio_end();
5710 * sys_sched_get_priority_max - return maximum RT priority.
5711 * @policy: scheduling class.
5713 * this syscall returns the maximum rt_priority that can be used
5714 * by a given scheduling class.
5716 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5723 ret = MAX_USER_RT_PRIO-1;
5735 * sys_sched_get_priority_min - return minimum RT priority.
5736 * @policy: scheduling class.
5738 * this syscall returns the minimum rt_priority that can be used
5739 * by a given scheduling class.
5741 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5759 * sys_sched_rr_get_interval - return the default timeslice of a process.
5760 * @pid: pid of the process.
5761 * @interval: userspace pointer to the timeslice value.
5763 * this syscall writes the default timeslice value of a given process
5764 * into the user-space timespec buffer. A value of '0' means infinity.
5766 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5767 struct timespec __user *, interval)
5769 struct task_struct *p;
5770 unsigned int time_slice;
5771 unsigned long flags;
5781 p = find_process_by_pid(pid);
5785 retval = security_task_getscheduler(p);
5789 rq = task_rq_lock(p, &flags);
5790 time_slice = p->sched_class->get_rr_interval(rq, p);
5791 task_rq_unlock(rq, p, &flags);
5794 jiffies_to_timespec(time_slice, &t);
5795 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5803 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5805 void sched_show_task(struct task_struct *p)
5807 unsigned long free = 0;
5810 state = p->state ? __ffs(p->state) + 1 : 0;
5811 printk(KERN_INFO "%-15.15s %c", p->comm,
5812 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5813 #if BITS_PER_LONG == 32
5814 if (state == TASK_RUNNING)
5815 printk(KERN_CONT " running ");
5817 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5819 if (state == TASK_RUNNING)
5820 printk(KERN_CONT " running task ");
5822 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5824 #ifdef CONFIG_DEBUG_STACK_USAGE
5825 free = stack_not_used(p);
5827 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5828 task_pid_nr(p), task_pid_nr(p->real_parent),
5829 (unsigned long)task_thread_info(p)->flags);
5831 show_stack(p, NULL);
5834 void show_state_filter(unsigned long state_filter)
5836 struct task_struct *g, *p;
5838 #if BITS_PER_LONG == 32
5840 " task PC stack pid father\n");
5843 " task PC stack pid father\n");
5845 read_lock(&tasklist_lock);
5846 do_each_thread(g, p) {
5848 * reset the NMI-timeout, listing all files on a slow
5849 * console might take a lot of time:
5851 touch_nmi_watchdog();
5852 if (!state_filter || (p->state & state_filter))
5854 } while_each_thread(g, p);
5856 touch_all_softlockup_watchdogs();
5858 #ifdef CONFIG_SCHED_DEBUG
5859 sysrq_sched_debug_show();
5861 read_unlock(&tasklist_lock);
5863 * Only show locks if all tasks are dumped:
5866 debug_show_all_locks();
5869 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5871 idle->sched_class = &idle_sched_class;
5875 * init_idle - set up an idle thread for a given CPU
5876 * @idle: task in question
5877 * @cpu: cpu the idle task belongs to
5879 * NOTE: this function does not set the idle thread's NEED_RESCHED
5880 * flag, to make booting more robust.
5882 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5884 struct rq *rq = cpu_rq(cpu);
5885 unsigned long flags;
5887 raw_spin_lock_irqsave(&rq->lock, flags);
5890 idle->state = TASK_RUNNING;
5891 idle->se.exec_start = sched_clock();
5893 do_set_cpus_allowed(idle, cpumask_of(cpu));
5895 * We're having a chicken and egg problem, even though we are
5896 * holding rq->lock, the cpu isn't yet set to this cpu so the
5897 * lockdep check in task_group() will fail.
5899 * Similar case to sched_fork(). / Alternatively we could
5900 * use task_rq_lock() here and obtain the other rq->lock.
5905 __set_task_cpu(idle, cpu);
5908 rq->curr = rq->idle = idle;
5909 #if defined(CONFIG_SMP)
5912 raw_spin_unlock_irqrestore(&rq->lock, flags);
5914 /* Set the preempt count _outside_ the spinlocks! */
5915 task_thread_info(idle)->preempt_count = 0;
5918 * The idle tasks have their own, simple scheduling class:
5920 idle->sched_class = &idle_sched_class;
5921 ftrace_graph_init_idle_task(idle, cpu);
5925 * In a system that switches off the HZ timer nohz_cpu_mask
5926 * indicates which cpus entered this state. This is used
5927 * in the rcu update to wait only for active cpus. For system
5928 * which do not switch off the HZ timer nohz_cpu_mask should
5929 * always be CPU_BITS_NONE.
5931 cpumask_var_t nohz_cpu_mask;
5934 * Increase the granularity value when there are more CPUs,
5935 * because with more CPUs the 'effective latency' as visible
5936 * to users decreases. But the relationship is not linear,
5937 * so pick a second-best guess by going with the log2 of the
5940 * This idea comes from the SD scheduler of Con Kolivas:
5942 static int get_update_sysctl_factor(void)
5944 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5945 unsigned int factor;
5947 switch (sysctl_sched_tunable_scaling) {
5948 case SCHED_TUNABLESCALING_NONE:
5951 case SCHED_TUNABLESCALING_LINEAR:
5954 case SCHED_TUNABLESCALING_LOG:
5956 factor = 1 + ilog2(cpus);
5963 static void update_sysctl(void)
5965 unsigned int factor = get_update_sysctl_factor();
5967 #define SET_SYSCTL(name) \
5968 (sysctl_##name = (factor) * normalized_sysctl_##name)
5969 SET_SYSCTL(sched_min_granularity);
5970 SET_SYSCTL(sched_latency);
5971 SET_SYSCTL(sched_wakeup_granularity);
5975 static inline void sched_init_granularity(void)
5981 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
5983 if (p->sched_class && p->sched_class->set_cpus_allowed)
5984 p->sched_class->set_cpus_allowed(p, new_mask);
5986 cpumask_copy(&p->cpus_allowed, new_mask);
5987 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5992 * This is how migration works:
5994 * 1) we invoke migration_cpu_stop() on the target CPU using
5996 * 2) stopper starts to run (implicitly forcing the migrated thread
5998 * 3) it checks whether the migrated task is still in the wrong runqueue.
5999 * 4) if it's in the wrong runqueue then the migration thread removes
6000 * it and puts it into the right queue.
6001 * 5) stopper completes and stop_one_cpu() returns and the migration
6006 * Change a given task's CPU affinity. Migrate the thread to a
6007 * proper CPU and schedule it away if the CPU it's executing on
6008 * is removed from the allowed bitmask.
6010 * NOTE: the caller must have a valid reference to the task, the
6011 * task must not exit() & deallocate itself prematurely. The
6012 * call is not atomic; no spinlocks may be held.
6014 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6016 unsigned long flags;
6018 unsigned int dest_cpu;
6021 rq = task_rq_lock(p, &flags);
6023 if (cpumask_equal(&p->cpus_allowed, new_mask))
6026 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6031 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6036 do_set_cpus_allowed(p, new_mask);
6038 /* Can the task run on the task's current CPU? If so, we're done */
6039 if (cpumask_test_cpu(task_cpu(p), new_mask))
6042 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6044 struct migration_arg arg = { p, dest_cpu };
6045 /* Need help from migration thread: drop lock and wait. */
6046 task_rq_unlock(rq, p, &flags);
6047 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6048 tlb_migrate_finish(p->mm);
6052 task_rq_unlock(rq, p, &flags);
6056 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6059 * Move (not current) task off this cpu, onto dest cpu. We're doing
6060 * this because either it can't run here any more (set_cpus_allowed()
6061 * away from this CPU, or CPU going down), or because we're
6062 * attempting to rebalance this task on exec (sched_exec).
6064 * So we race with normal scheduler movements, but that's OK, as long
6065 * as the task is no longer on this CPU.
6067 * Returns non-zero if task was successfully migrated.
6069 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6071 struct rq *rq_dest, *rq_src;
6074 if (unlikely(!cpu_active(dest_cpu)))
6077 rq_src = cpu_rq(src_cpu);
6078 rq_dest = cpu_rq(dest_cpu);
6080 raw_spin_lock(&p->pi_lock);
6081 double_rq_lock(rq_src, rq_dest);
6082 /* Already moved. */
6083 if (task_cpu(p) != src_cpu)
6085 /* Affinity changed (again). */
6086 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6090 * If we're not on a rq, the next wake-up will ensure we're
6094 deactivate_task(rq_src, p, 0);
6095 set_task_cpu(p, dest_cpu);
6096 activate_task(rq_dest, p, 0);
6097 check_preempt_curr(rq_dest, p, 0);
6102 double_rq_unlock(rq_src, rq_dest);
6103 raw_spin_unlock(&p->pi_lock);
6108 * migration_cpu_stop - this will be executed by a highprio stopper thread
6109 * and performs thread migration by bumping thread off CPU then
6110 * 'pushing' onto another runqueue.
6112 static int migration_cpu_stop(void *data)
6114 struct migration_arg *arg = data;
6117 * The original target cpu might have gone down and we might
6118 * be on another cpu but it doesn't matter.
6120 local_irq_disable();
6121 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6126 #ifdef CONFIG_HOTPLUG_CPU
6129 * Ensures that the idle task is using init_mm right before its cpu goes
6132 void idle_task_exit(void)
6134 struct mm_struct *mm = current->active_mm;
6136 BUG_ON(cpu_online(smp_processor_id()));
6139 switch_mm(mm, &init_mm, current);
6144 * While a dead CPU has no uninterruptible tasks queued at this point,
6145 * it might still have a nonzero ->nr_uninterruptible counter, because
6146 * for performance reasons the counter is not stricly tracking tasks to
6147 * their home CPUs. So we just add the counter to another CPU's counter,
6148 * to keep the global sum constant after CPU-down:
6150 static void migrate_nr_uninterruptible(struct rq *rq_src)
6152 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6154 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6155 rq_src->nr_uninterruptible = 0;
6159 * remove the tasks which were accounted by rq from calc_load_tasks.
6161 static void calc_global_load_remove(struct rq *rq)
6163 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6164 rq->calc_load_active = 0;
6168 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6169 * try_to_wake_up()->select_task_rq().
6171 * Called with rq->lock held even though we'er in stop_machine() and
6172 * there's no concurrency possible, we hold the required locks anyway
6173 * because of lock validation efforts.
6175 static void migrate_tasks(unsigned int dead_cpu)
6177 struct rq *rq = cpu_rq(dead_cpu);
6178 struct task_struct *next, *stop = rq->stop;
6182 * Fudge the rq selection such that the below task selection loop
6183 * doesn't get stuck on the currently eligible stop task.
6185 * We're currently inside stop_machine() and the rq is either stuck
6186 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6187 * either way we should never end up calling schedule() until we're
6194 * There's this thread running, bail when that's the only
6197 if (rq->nr_running == 1)
6200 next = pick_next_task(rq);
6202 next->sched_class->put_prev_task(rq, next);
6204 /* Find suitable destination for @next, with force if needed. */
6205 dest_cpu = select_fallback_rq(dead_cpu, next);
6206 raw_spin_unlock(&rq->lock);
6208 __migrate_task(next, dead_cpu, dest_cpu);
6210 raw_spin_lock(&rq->lock);
6216 #endif /* CONFIG_HOTPLUG_CPU */
6218 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6220 static struct ctl_table sd_ctl_dir[] = {
6222 .procname = "sched_domain",
6228 static struct ctl_table sd_ctl_root[] = {
6230 .procname = "kernel",
6232 .child = sd_ctl_dir,
6237 static struct ctl_table *sd_alloc_ctl_entry(int n)
6239 struct ctl_table *entry =
6240 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6245 static void sd_free_ctl_entry(struct ctl_table **tablep)
6247 struct ctl_table *entry;
6250 * In the intermediate directories, both the child directory and
6251 * procname are dynamically allocated and could fail but the mode
6252 * will always be set. In the lowest directory the names are
6253 * static strings and all have proc handlers.
6255 for (entry = *tablep; entry->mode; entry++) {
6257 sd_free_ctl_entry(&entry->child);
6258 if (entry->proc_handler == NULL)
6259 kfree(entry->procname);
6267 set_table_entry(struct ctl_table *entry,
6268 const char *procname, void *data, int maxlen,
6269 mode_t mode, proc_handler *proc_handler)
6271 entry->procname = procname;
6273 entry->maxlen = maxlen;
6275 entry->proc_handler = proc_handler;
6278 static struct ctl_table *
6279 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6281 struct ctl_table *table = sd_alloc_ctl_entry(13);
6286 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6287 sizeof(long), 0644, proc_doulongvec_minmax);
6288 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6289 sizeof(long), 0644, proc_doulongvec_minmax);
6290 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6291 sizeof(int), 0644, proc_dointvec_minmax);
6292 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6293 sizeof(int), 0644, proc_dointvec_minmax);
6294 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6295 sizeof(int), 0644, proc_dointvec_minmax);
6296 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6297 sizeof(int), 0644, proc_dointvec_minmax);
6298 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6299 sizeof(int), 0644, proc_dointvec_minmax);
6300 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6301 sizeof(int), 0644, proc_dointvec_minmax);
6302 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6303 sizeof(int), 0644, proc_dointvec_minmax);
6304 set_table_entry(&table[9], "cache_nice_tries",
6305 &sd->cache_nice_tries,
6306 sizeof(int), 0644, proc_dointvec_minmax);
6307 set_table_entry(&table[10], "flags", &sd->flags,
6308 sizeof(int), 0644, proc_dointvec_minmax);
6309 set_table_entry(&table[11], "name", sd->name,
6310 CORENAME_MAX_SIZE, 0444, proc_dostring);
6311 /* &table[12] is terminator */
6316 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6318 struct ctl_table *entry, *table;
6319 struct sched_domain *sd;
6320 int domain_num = 0, i;
6323 for_each_domain(cpu, sd)
6325 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6330 for_each_domain(cpu, sd) {
6331 snprintf(buf, 32, "domain%d", i);
6332 entry->procname = kstrdup(buf, GFP_KERNEL);
6334 entry->child = sd_alloc_ctl_domain_table(sd);
6341 static struct ctl_table_header *sd_sysctl_header;
6342 static void register_sched_domain_sysctl(void)
6344 int i, cpu_num = num_possible_cpus();
6345 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6348 WARN_ON(sd_ctl_dir[0].child);
6349 sd_ctl_dir[0].child = entry;
6354 for_each_possible_cpu(i) {
6355 snprintf(buf, 32, "cpu%d", i);
6356 entry->procname = kstrdup(buf, GFP_KERNEL);
6358 entry->child = sd_alloc_ctl_cpu_table(i);
6362 WARN_ON(sd_sysctl_header);
6363 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6366 /* may be called multiple times per register */
6367 static void unregister_sched_domain_sysctl(void)
6369 if (sd_sysctl_header)
6370 unregister_sysctl_table(sd_sysctl_header);
6371 sd_sysctl_header = NULL;
6372 if (sd_ctl_dir[0].child)
6373 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6376 static void register_sched_domain_sysctl(void)
6379 static void unregister_sched_domain_sysctl(void)
6384 static void set_rq_online(struct rq *rq)
6387 const struct sched_class *class;
6389 cpumask_set_cpu(rq->cpu, rq->rd->online);
6392 for_each_class(class) {
6393 if (class->rq_online)
6394 class->rq_online(rq);
6399 static void set_rq_offline(struct rq *rq)
6402 const struct sched_class *class;
6404 for_each_class(class) {
6405 if (class->rq_offline)
6406 class->rq_offline(rq);
6409 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6415 * migration_call - callback that gets triggered when a CPU is added.
6416 * Here we can start up the necessary migration thread for the new CPU.
6418 static int __cpuinit
6419 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6421 int cpu = (long)hcpu;
6422 unsigned long flags;
6423 struct rq *rq = cpu_rq(cpu);
6425 switch (action & ~CPU_TASKS_FROZEN) {
6427 case CPU_UP_PREPARE:
6428 rq->calc_load_update = calc_load_update;
6432 /* Update our root-domain */
6433 raw_spin_lock_irqsave(&rq->lock, flags);
6435 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6439 raw_spin_unlock_irqrestore(&rq->lock, flags);
6442 #ifdef CONFIG_HOTPLUG_CPU
6444 sched_ttwu_pending();
6445 /* Update our root-domain */
6446 raw_spin_lock_irqsave(&rq->lock, flags);
6448 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6452 BUG_ON(rq->nr_running != 1); /* the migration thread */
6453 raw_spin_unlock_irqrestore(&rq->lock, flags);
6455 migrate_nr_uninterruptible(rq);
6456 calc_global_load_remove(rq);
6461 update_max_interval();
6467 * Register at high priority so that task migration (migrate_all_tasks)
6468 * happens before everything else. This has to be lower priority than
6469 * the notifier in the perf_event subsystem, though.
6471 static struct notifier_block __cpuinitdata migration_notifier = {
6472 .notifier_call = migration_call,
6473 .priority = CPU_PRI_MIGRATION,
6476 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6477 unsigned long action, void *hcpu)
6479 switch (action & ~CPU_TASKS_FROZEN) {
6481 case CPU_DOWN_FAILED:
6482 set_cpu_active((long)hcpu, true);
6489 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6490 unsigned long action, void *hcpu)
6492 switch (action & ~CPU_TASKS_FROZEN) {
6493 case CPU_DOWN_PREPARE:
6494 set_cpu_active((long)hcpu, false);
6501 static int __init migration_init(void)
6503 void *cpu = (void *)(long)smp_processor_id();
6506 /* Initialize migration for the boot CPU */
6507 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6508 BUG_ON(err == NOTIFY_BAD);
6509 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6510 register_cpu_notifier(&migration_notifier);
6512 /* Register cpu active notifiers */
6513 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6514 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6518 early_initcall(migration_init);
6523 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6525 #ifdef CONFIG_SCHED_DEBUG
6527 static __read_mostly int sched_domain_debug_enabled;
6529 static int __init sched_domain_debug_setup(char *str)
6531 sched_domain_debug_enabled = 1;
6535 early_param("sched_debug", sched_domain_debug_setup);
6537 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6538 struct cpumask *groupmask)
6540 struct sched_group *group = sd->groups;
6543 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6544 cpumask_clear(groupmask);
6546 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6548 if (!(sd->flags & SD_LOAD_BALANCE)) {
6549 printk("does not load-balance\n");
6551 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6556 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6558 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6559 printk(KERN_ERR "ERROR: domain->span does not contain "
6562 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6563 printk(KERN_ERR "ERROR: domain->groups does not contain"
6567 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6571 printk(KERN_ERR "ERROR: group is NULL\n");
6575 if (!group->sgp->power) {
6576 printk(KERN_CONT "\n");
6577 printk(KERN_ERR "ERROR: domain->cpu_power not "
6582 if (!cpumask_weight(sched_group_cpus(group))) {
6583 printk(KERN_CONT "\n");
6584 printk(KERN_ERR "ERROR: empty group\n");
6588 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6589 printk(KERN_CONT "\n");
6590 printk(KERN_ERR "ERROR: repeated CPUs\n");
6594 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6596 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6598 printk(KERN_CONT " %s", str);
6599 if (group->sgp->power != SCHED_POWER_SCALE) {
6600 printk(KERN_CONT " (cpu_power = %d)",
6604 group = group->next;
6605 } while (group != sd->groups);
6606 printk(KERN_CONT "\n");
6608 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6609 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6612 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6613 printk(KERN_ERR "ERROR: parent span is not a superset "
6614 "of domain->span\n");
6618 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6622 if (!sched_domain_debug_enabled)
6626 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6630 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6633 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6641 #else /* !CONFIG_SCHED_DEBUG */
6642 # define sched_domain_debug(sd, cpu) do { } while (0)
6643 #endif /* CONFIG_SCHED_DEBUG */
6645 static int sd_degenerate(struct sched_domain *sd)
6647 if (cpumask_weight(sched_domain_span(sd)) == 1)
6650 /* Following flags need at least 2 groups */
6651 if (sd->flags & (SD_LOAD_BALANCE |
6652 SD_BALANCE_NEWIDLE |
6656 SD_SHARE_PKG_RESOURCES)) {
6657 if (sd->groups != sd->groups->next)
6661 /* Following flags don't use groups */
6662 if (sd->flags & (SD_WAKE_AFFINE))
6669 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6671 unsigned long cflags = sd->flags, pflags = parent->flags;
6673 if (sd_degenerate(parent))
6676 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6679 /* Flags needing groups don't count if only 1 group in parent */
6680 if (parent->groups == parent->groups->next) {
6681 pflags &= ~(SD_LOAD_BALANCE |
6682 SD_BALANCE_NEWIDLE |
6686 SD_SHARE_PKG_RESOURCES);
6687 if (nr_node_ids == 1)
6688 pflags &= ~SD_SERIALIZE;
6690 if (~cflags & pflags)
6696 static void free_rootdomain(struct rcu_head *rcu)
6698 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6700 cpupri_cleanup(&rd->cpupri);
6701 free_cpumask_var(rd->rto_mask);
6702 free_cpumask_var(rd->online);
6703 free_cpumask_var(rd->span);
6707 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6709 struct root_domain *old_rd = NULL;
6710 unsigned long flags;
6712 raw_spin_lock_irqsave(&rq->lock, flags);
6717 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6720 cpumask_clear_cpu(rq->cpu, old_rd->span);
6723 * If we dont want to free the old_rt yet then
6724 * set old_rd to NULL to skip the freeing later
6727 if (!atomic_dec_and_test(&old_rd->refcount))
6731 atomic_inc(&rd->refcount);
6734 cpumask_set_cpu(rq->cpu, rd->span);
6735 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6738 raw_spin_unlock_irqrestore(&rq->lock, flags);
6741 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6744 static int init_rootdomain(struct root_domain *rd)
6746 memset(rd, 0, sizeof(*rd));
6748 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6750 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6752 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6755 if (cpupri_init(&rd->cpupri) != 0)
6760 free_cpumask_var(rd->rto_mask);
6762 free_cpumask_var(rd->online);
6764 free_cpumask_var(rd->span);
6769 static void init_defrootdomain(void)
6771 init_rootdomain(&def_root_domain);
6773 atomic_set(&def_root_domain.refcount, 1);
6776 static struct root_domain *alloc_rootdomain(void)
6778 struct root_domain *rd;
6780 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6784 if (init_rootdomain(rd) != 0) {
6792 static void free_sched_groups(struct sched_group *sg, int free_sgp)
6794 struct sched_group *tmp, *first;
6803 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
6808 } while (sg != first);
6811 static void free_sched_domain(struct rcu_head *rcu)
6813 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6816 * If its an overlapping domain it has private groups, iterate and
6819 if (sd->flags & SD_OVERLAP) {
6820 free_sched_groups(sd->groups, 1);
6821 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6822 kfree(sd->groups->sgp);
6828 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6830 call_rcu(&sd->rcu, free_sched_domain);
6833 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6835 for (; sd; sd = sd->parent)
6836 destroy_sched_domain(sd, cpu);
6840 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6841 * hold the hotplug lock.
6844 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6846 struct rq *rq = cpu_rq(cpu);
6847 struct sched_domain *tmp;
6849 /* Remove the sched domains which do not contribute to scheduling. */
6850 for (tmp = sd; tmp; ) {
6851 struct sched_domain *parent = tmp->parent;
6855 if (sd_parent_degenerate(tmp, parent)) {
6856 tmp->parent = parent->parent;
6858 parent->parent->child = tmp;
6859 destroy_sched_domain(parent, cpu);
6864 if (sd && sd_degenerate(sd)) {
6867 destroy_sched_domain(tmp, cpu);
6872 sched_domain_debug(sd, cpu);
6874 rq_attach_root(rq, rd);
6876 rcu_assign_pointer(rq->sd, sd);
6877 destroy_sched_domains(tmp, cpu);
6880 /* cpus with isolated domains */
6881 static cpumask_var_t cpu_isolated_map;
6883 /* Setup the mask of cpus configured for isolated domains */
6884 static int __init isolated_cpu_setup(char *str)
6886 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6887 cpulist_parse(str, cpu_isolated_map);
6891 __setup("isolcpus=", isolated_cpu_setup);
6893 #define SD_NODES_PER_DOMAIN 16
6898 * find_next_best_node - find the next node to include in a sched_domain
6899 * @node: node whose sched_domain we're building
6900 * @used_nodes: nodes already in the sched_domain
6902 * Find the next node to include in a given scheduling domain. Simply
6903 * finds the closest node not already in the @used_nodes map.
6905 * Should use nodemask_t.
6907 static int find_next_best_node(int node, nodemask_t *used_nodes)
6909 int i, n, val, min_val, best_node = -1;
6913 for (i = 0; i < nr_node_ids; i++) {
6914 /* Start at @node */
6915 n = (node + i) % nr_node_ids;
6917 if (!nr_cpus_node(n))
6920 /* Skip already used nodes */
6921 if (node_isset(n, *used_nodes))
6924 /* Simple min distance search */
6925 val = node_distance(node, n);
6927 if (val < min_val) {
6933 if (best_node != -1)
6934 node_set(best_node, *used_nodes);
6939 * sched_domain_node_span - get a cpumask for a node's sched_domain
6940 * @node: node whose cpumask we're constructing
6941 * @span: resulting cpumask
6943 * Given a node, construct a good cpumask for its sched_domain to span. It
6944 * should be one that prevents unnecessary balancing, but also spreads tasks
6947 static void sched_domain_node_span(int node, struct cpumask *span)
6949 nodemask_t used_nodes;
6952 cpumask_clear(span);
6953 nodes_clear(used_nodes);
6955 cpumask_or(span, span, cpumask_of_node(node));
6956 node_set(node, used_nodes);
6958 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6959 int next_node = find_next_best_node(node, &used_nodes);
6962 cpumask_or(span, span, cpumask_of_node(next_node));
6966 static const struct cpumask *cpu_node_mask(int cpu)
6968 lockdep_assert_held(&sched_domains_mutex);
6970 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
6972 return sched_domains_tmpmask;
6975 static const struct cpumask *cpu_allnodes_mask(int cpu)
6977 return cpu_possible_mask;
6979 #endif /* CONFIG_NUMA */
6981 static const struct cpumask *cpu_cpu_mask(int cpu)
6983 return cpumask_of_node(cpu_to_node(cpu));
6986 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6989 struct sched_domain **__percpu sd;
6990 struct sched_group **__percpu sg;
6991 struct sched_group_power **__percpu sgp;
6995 struct sched_domain ** __percpu sd;
6996 struct root_domain *rd;
7006 struct sched_domain_topology_level;
7008 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
7009 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
7011 #define SDTL_OVERLAP 0x01
7013 struct sched_domain_topology_level {
7014 sched_domain_init_f init;
7015 sched_domain_mask_f mask;
7017 struct sd_data data;
7021 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
7023 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
7024 const struct cpumask *span = sched_domain_span(sd);
7025 struct cpumask *covered = sched_domains_tmpmask;
7026 struct sd_data *sdd = sd->private;
7027 struct sched_domain *child;
7030 cpumask_clear(covered);
7032 for_each_cpu(i, span) {
7033 struct cpumask *sg_span;
7035 if (cpumask_test_cpu(i, covered))
7038 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7039 GFP_KERNEL, cpu_to_node(i));
7044 sg_span = sched_group_cpus(sg);
7046 child = *per_cpu_ptr(sdd->sd, i);
7048 child = child->child;
7049 cpumask_copy(sg_span, sched_domain_span(child));
7051 cpumask_set_cpu(i, sg_span);
7053 cpumask_or(covered, covered, sg_span);
7055 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
7056 atomic_inc(&sg->sgp->ref);
7058 if (cpumask_test_cpu(cpu, sg_span))
7068 sd->groups = groups;
7073 free_sched_groups(first, 0);
7078 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
7080 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
7081 struct sched_domain *child = sd->child;
7084 cpu = cpumask_first(sched_domain_span(child));
7087 *sg = *per_cpu_ptr(sdd->sg, cpu);
7088 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
7089 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
7096 * build_sched_groups will build a circular linked list of the groups
7097 * covered by the given span, and will set each group's ->cpumask correctly,
7098 * and ->cpu_power to 0.
7100 * Assumes the sched_domain tree is fully constructed
7103 build_sched_groups(struct sched_domain *sd, int cpu)
7105 struct sched_group *first = NULL, *last = NULL;
7106 struct sd_data *sdd = sd->private;
7107 const struct cpumask *span = sched_domain_span(sd);
7108 struct cpumask *covered;
7111 get_group(cpu, sdd, &sd->groups);
7112 atomic_inc(&sd->groups->ref);
7114 if (cpu != cpumask_first(sched_domain_span(sd)))
7117 lockdep_assert_held(&sched_domains_mutex);
7118 covered = sched_domains_tmpmask;
7120 cpumask_clear(covered);
7122 for_each_cpu(i, span) {
7123 struct sched_group *sg;
7124 int group = get_group(i, sdd, &sg);
7127 if (cpumask_test_cpu(i, covered))
7130 cpumask_clear(sched_group_cpus(sg));
7133 for_each_cpu(j, span) {
7134 if (get_group(j, sdd, NULL) != group)
7137 cpumask_set_cpu(j, covered);
7138 cpumask_set_cpu(j, sched_group_cpus(sg));
7153 * Initialize sched groups cpu_power.
7155 * cpu_power indicates the capacity of sched group, which is used while
7156 * distributing the load between different sched groups in a sched domain.
7157 * Typically cpu_power for all the groups in a sched domain will be same unless
7158 * there are asymmetries in the topology. If there are asymmetries, group
7159 * having more cpu_power will pickup more load compared to the group having
7162 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7164 struct sched_group *sg = sd->groups;
7166 WARN_ON(!sd || !sg);
7169 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
7171 } while (sg != sd->groups);
7173 if (cpu != group_first_cpu(sg))
7176 update_group_power(sd, cpu);
7180 * Initializers for schedule domains
7181 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7184 #ifdef CONFIG_SCHED_DEBUG
7185 # define SD_INIT_NAME(sd, type) sd->name = #type
7187 # define SD_INIT_NAME(sd, type) do { } while (0)
7190 #define SD_INIT_FUNC(type) \
7191 static noinline struct sched_domain * \
7192 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7194 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7195 *sd = SD_##type##_INIT; \
7196 SD_INIT_NAME(sd, type); \
7197 sd->private = &tl->data; \
7203 SD_INIT_FUNC(ALLNODES)
7206 #ifdef CONFIG_SCHED_SMT
7207 SD_INIT_FUNC(SIBLING)
7209 #ifdef CONFIG_SCHED_MC
7212 #ifdef CONFIG_SCHED_BOOK
7216 static int default_relax_domain_level = -1;
7217 int sched_domain_level_max;
7219 static int __init setup_relax_domain_level(char *str)
7221 if (kstrtoint(str, 0, &default_relax_domain_level))
7222 pr_warn("Unable to set relax_domain_level\n");
7226 __setup("relax_domain_level=", setup_relax_domain_level);
7228 static void set_domain_attribute(struct sched_domain *sd,
7229 struct sched_domain_attr *attr)
7233 if (!attr || attr->relax_domain_level < 0) {
7234 if (default_relax_domain_level < 0)
7237 request = default_relax_domain_level;
7239 request = attr->relax_domain_level;
7240 if (request < sd->level) {
7241 /* turn off idle balance on this domain */
7242 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7244 /* turn on idle balance on this domain */
7245 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7249 static void __sdt_free(const struct cpumask *cpu_map);
7250 static int __sdt_alloc(const struct cpumask *cpu_map);
7252 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7253 const struct cpumask *cpu_map)
7257 if (!atomic_read(&d->rd->refcount))
7258 free_rootdomain(&d->rd->rcu); /* fall through */
7260 free_percpu(d->sd); /* fall through */
7262 __sdt_free(cpu_map); /* fall through */
7268 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7269 const struct cpumask *cpu_map)
7271 memset(d, 0, sizeof(*d));
7273 if (__sdt_alloc(cpu_map))
7274 return sa_sd_storage;
7275 d->sd = alloc_percpu(struct sched_domain *);
7277 return sa_sd_storage;
7278 d->rd = alloc_rootdomain();
7281 return sa_rootdomain;
7285 * NULL the sd_data elements we've used to build the sched_domain and
7286 * sched_group structure so that the subsequent __free_domain_allocs()
7287 * will not free the data we're using.
7289 static void claim_allocations(int cpu, struct sched_domain *sd)
7291 struct sd_data *sdd = sd->private;
7293 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7294 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7296 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
7297 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7299 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
7300 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
7303 #ifdef CONFIG_SCHED_SMT
7304 static const struct cpumask *cpu_smt_mask(int cpu)
7306 return topology_thread_cpumask(cpu);
7311 * Topology list, bottom-up.
7313 static struct sched_domain_topology_level default_topology[] = {
7314 #ifdef CONFIG_SCHED_SMT
7315 { sd_init_SIBLING, cpu_smt_mask, },
7317 #ifdef CONFIG_SCHED_MC
7318 { sd_init_MC, cpu_coregroup_mask, },
7320 #ifdef CONFIG_SCHED_BOOK
7321 { sd_init_BOOK, cpu_book_mask, },
7323 { sd_init_CPU, cpu_cpu_mask, },
7325 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
7326 { sd_init_ALLNODES, cpu_allnodes_mask, },
7331 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7333 static int __sdt_alloc(const struct cpumask *cpu_map)
7335 struct sched_domain_topology_level *tl;
7338 for (tl = sched_domain_topology; tl->init; tl++) {
7339 struct sd_data *sdd = &tl->data;
7341 sdd->sd = alloc_percpu(struct sched_domain *);
7345 sdd->sg = alloc_percpu(struct sched_group *);
7349 sdd->sgp = alloc_percpu(struct sched_group_power *);
7353 for_each_cpu(j, cpu_map) {
7354 struct sched_domain *sd;
7355 struct sched_group *sg;
7356 struct sched_group_power *sgp;
7358 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7359 GFP_KERNEL, cpu_to_node(j));
7363 *per_cpu_ptr(sdd->sd, j) = sd;
7365 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7366 GFP_KERNEL, cpu_to_node(j));
7370 *per_cpu_ptr(sdd->sg, j) = sg;
7372 sgp = kzalloc_node(sizeof(struct sched_group_power),
7373 GFP_KERNEL, cpu_to_node(j));
7377 *per_cpu_ptr(sdd->sgp, j) = sgp;
7384 static void __sdt_free(const struct cpumask *cpu_map)
7386 struct sched_domain_topology_level *tl;
7389 for (tl = sched_domain_topology; tl->init; tl++) {
7390 struct sd_data *sdd = &tl->data;
7392 for_each_cpu(j, cpu_map) {
7393 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
7394 if (sd && (sd->flags & SD_OVERLAP))
7395 free_sched_groups(sd->groups, 0);
7396 kfree(*per_cpu_ptr(sdd->sd, j));
7397 kfree(*per_cpu_ptr(sdd->sg, j));
7398 kfree(*per_cpu_ptr(sdd->sgp, j));
7400 free_percpu(sdd->sd);
7401 free_percpu(sdd->sg);
7402 free_percpu(sdd->sgp);
7406 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7407 struct s_data *d, const struct cpumask *cpu_map,
7408 struct sched_domain_attr *attr, struct sched_domain *child,
7411 struct sched_domain *sd = tl->init(tl, cpu);
7415 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7417 sd->level = child->level + 1;
7418 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7422 set_domain_attribute(sd, attr);
7428 * Build sched domains for a given set of cpus and attach the sched domains
7429 * to the individual cpus
7431 static int build_sched_domains(const struct cpumask *cpu_map,
7432 struct sched_domain_attr *attr)
7434 enum s_alloc alloc_state = sa_none;
7435 struct sched_domain *sd;
7437 int i, ret = -ENOMEM;
7439 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7440 if (alloc_state != sa_rootdomain)
7443 /* Set up domains for cpus specified by the cpu_map. */
7444 for_each_cpu(i, cpu_map) {
7445 struct sched_domain_topology_level *tl;
7448 for (tl = sched_domain_topology; tl->init; tl++) {
7449 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7450 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7451 sd->flags |= SD_OVERLAP;
7452 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7459 *per_cpu_ptr(d.sd, i) = sd;
7462 /* Build the groups for the domains */
7463 for_each_cpu(i, cpu_map) {
7464 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7465 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7466 if (sd->flags & SD_OVERLAP) {
7467 if (build_overlap_sched_groups(sd, i))
7470 if (build_sched_groups(sd, i))
7476 /* Calculate CPU power for physical packages and nodes */
7477 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7478 if (!cpumask_test_cpu(i, cpu_map))
7481 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7482 claim_allocations(i, sd);
7483 init_sched_groups_power(i, sd);
7487 /* Attach the domains */
7489 for_each_cpu(i, cpu_map) {
7490 sd = *per_cpu_ptr(d.sd, i);
7491 cpu_attach_domain(sd, d.rd, i);
7497 __free_domain_allocs(&d, alloc_state, cpu_map);
7501 static cpumask_var_t *doms_cur; /* current sched domains */
7502 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7503 static struct sched_domain_attr *dattr_cur;
7504 /* attribues of custom domains in 'doms_cur' */
7507 * Special case: If a kmalloc of a doms_cur partition (array of
7508 * cpumask) fails, then fallback to a single sched domain,
7509 * as determined by the single cpumask fallback_doms.
7511 static cpumask_var_t fallback_doms;
7514 * arch_update_cpu_topology lets virtualized architectures update the
7515 * cpu core maps. It is supposed to return 1 if the topology changed
7516 * or 0 if it stayed the same.
7518 int __attribute__((weak)) arch_update_cpu_topology(void)
7523 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7526 cpumask_var_t *doms;
7528 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7531 for (i = 0; i < ndoms; i++) {
7532 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7533 free_sched_domains(doms, i);
7540 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7543 for (i = 0; i < ndoms; i++)
7544 free_cpumask_var(doms[i]);
7549 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7550 * For now this just excludes isolated cpus, but could be used to
7551 * exclude other special cases in the future.
7553 static int init_sched_domains(const struct cpumask *cpu_map)
7557 arch_update_cpu_topology();
7559 doms_cur = alloc_sched_domains(ndoms_cur);
7561 doms_cur = &fallback_doms;
7562 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7564 err = build_sched_domains(doms_cur[0], NULL);
7565 register_sched_domain_sysctl();
7571 * Detach sched domains from a group of cpus specified in cpu_map
7572 * These cpus will now be attached to the NULL domain
7574 static void detach_destroy_domains(const struct cpumask *cpu_map)
7579 for_each_cpu(i, cpu_map)
7580 cpu_attach_domain(NULL, &def_root_domain, i);
7584 /* handle null as "default" */
7585 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7586 struct sched_domain_attr *new, int idx_new)
7588 struct sched_domain_attr tmp;
7595 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7596 new ? (new + idx_new) : &tmp,
7597 sizeof(struct sched_domain_attr));
7601 * Partition sched domains as specified by the 'ndoms_new'
7602 * cpumasks in the array doms_new[] of cpumasks. This compares
7603 * doms_new[] to the current sched domain partitioning, doms_cur[].
7604 * It destroys each deleted domain and builds each new domain.
7606 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7607 * The masks don't intersect (don't overlap.) We should setup one
7608 * sched domain for each mask. CPUs not in any of the cpumasks will
7609 * not be load balanced. If the same cpumask appears both in the
7610 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7613 * The passed in 'doms_new' should be allocated using
7614 * alloc_sched_domains. This routine takes ownership of it and will
7615 * free_sched_domains it when done with it. If the caller failed the
7616 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7617 * and partition_sched_domains() will fallback to the single partition
7618 * 'fallback_doms', it also forces the domains to be rebuilt.
7620 * If doms_new == NULL it will be replaced with cpu_online_mask.
7621 * ndoms_new == 0 is a special case for destroying existing domains,
7622 * and it will not create the default domain.
7624 * Call with hotplug lock held
7626 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7627 struct sched_domain_attr *dattr_new)
7632 mutex_lock(&sched_domains_mutex);
7634 /* always unregister in case we don't destroy any domains */
7635 unregister_sched_domain_sysctl();
7637 /* Let architecture update cpu core mappings. */
7638 new_topology = arch_update_cpu_topology();
7640 n = doms_new ? ndoms_new : 0;
7642 /* Destroy deleted domains */
7643 for (i = 0; i < ndoms_cur; i++) {
7644 for (j = 0; j < n && !new_topology; j++) {
7645 if (cpumask_equal(doms_cur[i], doms_new[j])
7646 && dattrs_equal(dattr_cur, i, dattr_new, j))
7649 /* no match - a current sched domain not in new doms_new[] */
7650 detach_destroy_domains(doms_cur[i]);
7655 if (doms_new == NULL) {
7657 doms_new = &fallback_doms;
7658 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7659 WARN_ON_ONCE(dattr_new);
7662 /* Build new domains */
7663 for (i = 0; i < ndoms_new; i++) {
7664 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7665 if (cpumask_equal(doms_new[i], doms_cur[j])
7666 && dattrs_equal(dattr_new, i, dattr_cur, j))
7669 /* no match - add a new doms_new */
7670 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7675 /* Remember the new sched domains */
7676 if (doms_cur != &fallback_doms)
7677 free_sched_domains(doms_cur, ndoms_cur);
7678 kfree(dattr_cur); /* kfree(NULL) is safe */
7679 doms_cur = doms_new;
7680 dattr_cur = dattr_new;
7681 ndoms_cur = ndoms_new;
7683 register_sched_domain_sysctl();
7685 mutex_unlock(&sched_domains_mutex);
7688 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7689 static void reinit_sched_domains(void)
7693 /* Destroy domains first to force the rebuild */
7694 partition_sched_domains(0, NULL, NULL);
7696 rebuild_sched_domains();
7700 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7702 unsigned int level = 0;
7704 if (sscanf(buf, "%u", &level) != 1)
7708 * level is always be positive so don't check for
7709 * level < POWERSAVINGS_BALANCE_NONE which is 0
7710 * What happens on 0 or 1 byte write,
7711 * need to check for count as well?
7714 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7718 sched_smt_power_savings = level;
7720 sched_mc_power_savings = level;
7722 reinit_sched_domains();
7727 #ifdef CONFIG_SCHED_MC
7728 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7729 struct sysdev_class_attribute *attr,
7732 return sprintf(page, "%u\n", sched_mc_power_savings);
7734 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7735 struct sysdev_class_attribute *attr,
7736 const char *buf, size_t count)
7738 return sched_power_savings_store(buf, count, 0);
7740 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7741 sched_mc_power_savings_show,
7742 sched_mc_power_savings_store);
7745 #ifdef CONFIG_SCHED_SMT
7746 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7747 struct sysdev_class_attribute *attr,
7750 return sprintf(page, "%u\n", sched_smt_power_savings);
7752 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7753 struct sysdev_class_attribute *attr,
7754 const char *buf, size_t count)
7756 return sched_power_savings_store(buf, count, 1);
7758 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7759 sched_smt_power_savings_show,
7760 sched_smt_power_savings_store);
7763 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7767 #ifdef CONFIG_SCHED_SMT
7769 err = sysfs_create_file(&cls->kset.kobj,
7770 &attr_sched_smt_power_savings.attr);
7772 #ifdef CONFIG_SCHED_MC
7773 if (!err && mc_capable())
7774 err = sysfs_create_file(&cls->kset.kobj,
7775 &attr_sched_mc_power_savings.attr);
7779 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7781 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7784 * Update cpusets according to cpu_active mask. If cpusets are
7785 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7786 * around partition_sched_domains().
7788 * If we come here as part of a suspend/resume, don't touch cpusets because we
7789 * want to restore it back to its original state upon resume anyway.
7791 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7795 case CPU_ONLINE_FROZEN:
7796 case CPU_DOWN_FAILED_FROZEN:
7799 * num_cpus_frozen tracks how many CPUs are involved in suspend
7800 * resume sequence. As long as this is not the last online
7801 * operation in the resume sequence, just build a single sched
7802 * domain, ignoring cpusets.
7805 if (likely(num_cpus_frozen)) {
7806 partition_sched_domains(1, NULL, NULL);
7811 * This is the last CPU online operation. So fall through and
7812 * restore the original sched domains by considering the
7813 * cpuset configurations.
7817 case CPU_DOWN_FAILED:
7818 cpuset_update_active_cpus();
7826 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7830 case CPU_DOWN_PREPARE:
7831 cpuset_update_active_cpus();
7833 case CPU_DOWN_PREPARE_FROZEN:
7835 partition_sched_domains(1, NULL, NULL);
7843 static int update_runtime(struct notifier_block *nfb,
7844 unsigned long action, void *hcpu)
7846 int cpu = (int)(long)hcpu;
7849 case CPU_DOWN_PREPARE:
7850 case CPU_DOWN_PREPARE_FROZEN:
7851 disable_runtime(cpu_rq(cpu));
7854 case CPU_DOWN_FAILED:
7855 case CPU_DOWN_FAILED_FROZEN:
7857 case CPU_ONLINE_FROZEN:
7858 enable_runtime(cpu_rq(cpu));
7866 void __init sched_init_smp(void)
7868 cpumask_var_t non_isolated_cpus;
7870 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7871 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7874 mutex_lock(&sched_domains_mutex);
7875 init_sched_domains(cpu_active_mask);
7876 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7877 if (cpumask_empty(non_isolated_cpus))
7878 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7879 mutex_unlock(&sched_domains_mutex);
7882 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7883 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7885 /* RT runtime code needs to handle some hotplug events */
7886 hotcpu_notifier(update_runtime, 0);
7890 /* Move init over to a non-isolated CPU */
7891 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7893 sched_init_granularity();
7894 free_cpumask_var(non_isolated_cpus);
7896 init_sched_rt_class();
7899 void __init sched_init_smp(void)
7901 sched_init_granularity();
7903 #endif /* CONFIG_SMP */
7905 const_debug unsigned int sysctl_timer_migration = 1;
7907 int in_sched_functions(unsigned long addr)
7909 return in_lock_functions(addr) ||
7910 (addr >= (unsigned long)__sched_text_start
7911 && addr < (unsigned long)__sched_text_end);
7914 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7916 cfs_rq->tasks_timeline = RB_ROOT;
7917 INIT_LIST_HEAD(&cfs_rq->tasks);
7918 #ifdef CONFIG_FAIR_GROUP_SCHED
7920 /* allow initial update_cfs_load() to truncate */
7922 cfs_rq->load_stamp = 1;
7925 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7926 #ifndef CONFIG_64BIT
7927 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7931 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7933 struct rt_prio_array *array;
7936 array = &rt_rq->active;
7937 for (i = 0; i < MAX_RT_PRIO; i++) {
7938 INIT_LIST_HEAD(array->queue + i);
7939 __clear_bit(i, array->bitmap);
7941 /* delimiter for bitsearch: */
7942 __set_bit(MAX_RT_PRIO, array->bitmap);
7944 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7945 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7947 rt_rq->highest_prio.next = MAX_RT_PRIO;
7951 rt_rq->rt_nr_migratory = 0;
7952 rt_rq->overloaded = 0;
7953 plist_head_init(&rt_rq->pushable_tasks);
7957 rt_rq->rt_throttled = 0;
7958 rt_rq->rt_runtime = 0;
7959 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7961 #ifdef CONFIG_RT_GROUP_SCHED
7962 rt_rq->rt_nr_boosted = 0;
7967 #ifdef CONFIG_FAIR_GROUP_SCHED
7968 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7969 struct sched_entity *se, int cpu,
7970 struct sched_entity *parent)
7972 struct rq *rq = cpu_rq(cpu);
7973 tg->cfs_rq[cpu] = cfs_rq;
7974 init_cfs_rq(cfs_rq, rq);
7978 /* se could be NULL for root_task_group */
7983 se->cfs_rq = &rq->cfs;
7985 se->cfs_rq = parent->my_q;
7988 update_load_set(&se->load, 0);
7989 se->parent = parent;
7993 #ifdef CONFIG_RT_GROUP_SCHED
7994 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7995 struct sched_rt_entity *rt_se, int cpu,
7996 struct sched_rt_entity *parent)
7998 struct rq *rq = cpu_rq(cpu);
8000 tg->rt_rq[cpu] = rt_rq;
8001 init_rt_rq(rt_rq, rq);
8003 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8005 tg->rt_se[cpu] = rt_se;
8010 rt_se->rt_rq = &rq->rt;
8012 rt_se->rt_rq = parent->my_q;
8014 rt_se->my_q = rt_rq;
8015 rt_se->parent = parent;
8016 INIT_LIST_HEAD(&rt_se->run_list);
8020 void __init sched_init(void)
8023 unsigned long alloc_size = 0, ptr;
8025 #ifdef CONFIG_FAIR_GROUP_SCHED
8026 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8028 #ifdef CONFIG_RT_GROUP_SCHED
8029 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8031 #ifdef CONFIG_CPUMASK_OFFSTACK
8032 alloc_size += num_possible_cpus() * cpumask_size();
8035 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8037 #ifdef CONFIG_FAIR_GROUP_SCHED
8038 root_task_group.se = (struct sched_entity **)ptr;
8039 ptr += nr_cpu_ids * sizeof(void **);
8041 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8042 ptr += nr_cpu_ids * sizeof(void **);
8044 #endif /* CONFIG_FAIR_GROUP_SCHED */
8045 #ifdef CONFIG_RT_GROUP_SCHED
8046 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8047 ptr += nr_cpu_ids * sizeof(void **);
8049 root_task_group.rt_rq = (struct rt_rq **)ptr;
8050 ptr += nr_cpu_ids * sizeof(void **);
8052 #endif /* CONFIG_RT_GROUP_SCHED */
8053 #ifdef CONFIG_CPUMASK_OFFSTACK
8054 for_each_possible_cpu(i) {
8055 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8056 ptr += cpumask_size();
8058 #endif /* CONFIG_CPUMASK_OFFSTACK */
8062 init_defrootdomain();
8065 init_rt_bandwidth(&def_rt_bandwidth,
8066 global_rt_period(), global_rt_runtime());
8068 #ifdef CONFIG_RT_GROUP_SCHED
8069 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8070 global_rt_period(), global_rt_runtime());
8071 #endif /* CONFIG_RT_GROUP_SCHED */
8073 #ifdef CONFIG_CGROUP_SCHED
8074 list_add(&root_task_group.list, &task_groups);
8075 INIT_LIST_HEAD(&root_task_group.children);
8076 autogroup_init(&init_task);
8077 #endif /* CONFIG_CGROUP_SCHED */
8079 for_each_possible_cpu(i) {
8083 raw_spin_lock_init(&rq->lock);
8085 rq->calc_load_active = 0;
8086 rq->calc_load_update = jiffies + LOAD_FREQ;
8087 init_cfs_rq(&rq->cfs, rq);
8088 init_rt_rq(&rq->rt, rq);
8089 #ifdef CONFIG_FAIR_GROUP_SCHED
8090 root_task_group.shares = root_task_group_load;
8091 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8093 * How much cpu bandwidth does root_task_group get?
8095 * In case of task-groups formed thr' the cgroup filesystem, it
8096 * gets 100% of the cpu resources in the system. This overall
8097 * system cpu resource is divided among the tasks of
8098 * root_task_group and its child task-groups in a fair manner,
8099 * based on each entity's (task or task-group's) weight
8100 * (se->load.weight).
8102 * In other words, if root_task_group has 10 tasks of weight
8103 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8104 * then A0's share of the cpu resource is:
8106 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8108 * We achieve this by letting root_task_group's tasks sit
8109 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8111 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8112 #endif /* CONFIG_FAIR_GROUP_SCHED */
8114 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8115 #ifdef CONFIG_RT_GROUP_SCHED
8116 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8117 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8120 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8121 rq->cpu_load[j] = 0;
8123 rq->last_load_update_tick = jiffies;
8128 rq->cpu_power = SCHED_POWER_SCALE;
8129 rq->post_schedule = 0;
8130 rq->active_balance = 0;
8131 rq->next_balance = jiffies;
8136 rq->avg_idle = 2*sysctl_sched_migration_cost;
8137 rq_attach_root(rq, &def_root_domain);
8139 rq->nohz_balance_kick = 0;
8140 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8144 atomic_set(&rq->nr_iowait, 0);
8147 set_load_weight(&init_task);
8149 #ifdef CONFIG_PREEMPT_NOTIFIERS
8150 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8154 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8157 #ifdef CONFIG_RT_MUTEXES
8158 plist_head_init(&init_task.pi_waiters);
8162 * The boot idle thread does lazy MMU switching as well:
8164 atomic_inc(&init_mm.mm_count);
8165 enter_lazy_tlb(&init_mm, current);
8168 * Make us the idle thread. Technically, schedule() should not be
8169 * called from this thread, however somewhere below it might be,
8170 * but because we are the idle thread, we just pick up running again
8171 * when this runqueue becomes "idle".
8173 init_idle(current, smp_processor_id());
8175 calc_load_update = jiffies + LOAD_FREQ;
8178 * During early bootup we pretend to be a normal task:
8180 current->sched_class = &fair_sched_class;
8182 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8183 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8185 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8187 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8188 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8189 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8190 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8191 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8193 /* May be allocated at isolcpus cmdline parse time */
8194 if (cpu_isolated_map == NULL)
8195 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8198 scheduler_running = 1;
8201 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8202 static inline int preempt_count_equals(int preempt_offset)
8204 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8206 return (nested == preempt_offset);
8209 static int __might_sleep_init_called;
8210 int __init __might_sleep_init(void)
8212 __might_sleep_init_called = 1;
8215 early_initcall(__might_sleep_init);
8217 void __might_sleep(const char *file, int line, int preempt_offset)
8220 static unsigned long prev_jiffy; /* ratelimiting */
8222 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8225 if (system_state != SYSTEM_RUNNING &&
8226 (!__might_sleep_init_called || system_state != SYSTEM_BOOTING))
8228 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8230 prev_jiffy = jiffies;
8233 "BUG: sleeping function called from invalid context at %s:%d\n",
8236 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8237 in_atomic(), irqs_disabled(),
8238 current->pid, current->comm);
8240 debug_show_held_locks(current);
8241 if (irqs_disabled())
8242 print_irqtrace_events(current);
8246 EXPORT_SYMBOL(__might_sleep);
8249 #ifdef CONFIG_MAGIC_SYSRQ
8250 static void normalize_task(struct rq *rq, struct task_struct *p)
8252 const struct sched_class *prev_class = p->sched_class;
8253 int old_prio = p->prio;
8258 deactivate_task(rq, p, 0);
8259 __setscheduler(rq, p, SCHED_NORMAL, 0);
8261 activate_task(rq, p, 0);
8262 resched_task(rq->curr);
8265 check_class_changed(rq, p, prev_class, old_prio);
8268 void normalize_rt_tasks(void)
8270 struct task_struct *g, *p;
8271 unsigned long flags;
8274 read_lock_irqsave(&tasklist_lock, flags);
8275 do_each_thread(g, p) {
8277 * Only normalize user tasks:
8282 p->se.exec_start = 0;
8283 #ifdef CONFIG_SCHEDSTATS
8284 p->se.statistics.wait_start = 0;
8285 p->se.statistics.sleep_start = 0;
8286 p->se.statistics.block_start = 0;
8291 * Renice negative nice level userspace
8294 if (TASK_NICE(p) < 0 && p->mm)
8295 set_user_nice(p, 0);
8299 raw_spin_lock(&p->pi_lock);
8300 rq = __task_rq_lock(p);
8302 normalize_task(rq, p);
8304 __task_rq_unlock(rq);
8305 raw_spin_unlock(&p->pi_lock);
8306 } while_each_thread(g, p);
8308 read_unlock_irqrestore(&tasklist_lock, flags);
8311 #endif /* CONFIG_MAGIC_SYSRQ */
8313 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8315 * These functions are only useful for the IA64 MCA handling, or kdb.
8317 * They can only be called when the whole system has been
8318 * stopped - every CPU needs to be quiescent, and no scheduling
8319 * activity can take place. Using them for anything else would
8320 * be a serious bug, and as a result, they aren't even visible
8321 * under any other configuration.
8325 * curr_task - return the current task for a given cpu.
8326 * @cpu: the processor in question.
8328 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8330 struct task_struct *curr_task(int cpu)
8332 return cpu_curr(cpu);
8335 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8339 * set_curr_task - set the current task for a given cpu.
8340 * @cpu: the processor in question.
8341 * @p: the task pointer to set.
8343 * Description: This function must only be used when non-maskable interrupts
8344 * are serviced on a separate stack. It allows the architecture to switch the
8345 * notion of the current task on a cpu in a non-blocking manner. This function
8346 * must be called with all CPU's synchronized, and interrupts disabled, the
8347 * and caller must save the original value of the current task (see
8348 * curr_task() above) and restore that value before reenabling interrupts and
8349 * re-starting the system.
8351 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8353 void set_curr_task(int cpu, struct task_struct *p)
8360 #ifdef CONFIG_FAIR_GROUP_SCHED
8361 static void free_fair_sched_group(struct task_group *tg)
8365 for_each_possible_cpu(i) {
8367 kfree(tg->cfs_rq[i]);
8377 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8379 struct cfs_rq *cfs_rq;
8380 struct sched_entity *se;
8383 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8386 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8390 tg->shares = NICE_0_LOAD;
8392 for_each_possible_cpu(i) {
8393 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8394 GFP_KERNEL, cpu_to_node(i));
8398 se = kzalloc_node(sizeof(struct sched_entity),
8399 GFP_KERNEL, cpu_to_node(i));
8403 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8414 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8416 struct rq *rq = cpu_rq(cpu);
8417 unsigned long flags;
8420 * Only empty task groups can be destroyed; so we can speculatively
8421 * check on_list without danger of it being re-added.
8423 if (!tg->cfs_rq[cpu]->on_list)
8426 raw_spin_lock_irqsave(&rq->lock, flags);
8427 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8428 raw_spin_unlock_irqrestore(&rq->lock, flags);
8430 #else /* !CONFG_FAIR_GROUP_SCHED */
8431 static inline void free_fair_sched_group(struct task_group *tg)
8436 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8441 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8444 #endif /* CONFIG_FAIR_GROUP_SCHED */
8446 #ifdef CONFIG_RT_GROUP_SCHED
8447 static void free_rt_sched_group(struct task_group *tg)
8451 destroy_rt_bandwidth(&tg->rt_bandwidth);
8453 for_each_possible_cpu(i) {
8455 kfree(tg->rt_rq[i]);
8457 kfree(tg->rt_se[i]);
8465 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8467 struct rt_rq *rt_rq;
8468 struct sched_rt_entity *rt_se;
8471 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8474 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8478 init_rt_bandwidth(&tg->rt_bandwidth,
8479 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8481 for_each_possible_cpu(i) {
8482 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8483 GFP_KERNEL, cpu_to_node(i));
8487 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8488 GFP_KERNEL, cpu_to_node(i));
8492 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8502 #else /* !CONFIG_RT_GROUP_SCHED */
8503 static inline void free_rt_sched_group(struct task_group *tg)
8508 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8512 #endif /* CONFIG_RT_GROUP_SCHED */
8514 #ifdef CONFIG_CGROUP_SCHED
8515 static void free_sched_group(struct task_group *tg)
8517 free_fair_sched_group(tg);
8518 free_rt_sched_group(tg);
8523 /* allocate runqueue etc for a new task group */
8524 struct task_group *sched_create_group(struct task_group *parent)
8526 struct task_group *tg;
8527 unsigned long flags;
8529 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8531 return ERR_PTR(-ENOMEM);
8533 if (!alloc_fair_sched_group(tg, parent))
8536 if (!alloc_rt_sched_group(tg, parent))
8539 spin_lock_irqsave(&task_group_lock, flags);
8540 list_add_rcu(&tg->list, &task_groups);
8542 WARN_ON(!parent); /* root should already exist */
8544 tg->parent = parent;
8545 INIT_LIST_HEAD(&tg->children);
8546 list_add_rcu(&tg->siblings, &parent->children);
8547 spin_unlock_irqrestore(&task_group_lock, flags);
8552 free_sched_group(tg);
8553 return ERR_PTR(-ENOMEM);
8556 /* rcu callback to free various structures associated with a task group */
8557 static void free_sched_group_rcu(struct rcu_head *rhp)
8559 /* now it should be safe to free those cfs_rqs */
8560 free_sched_group(container_of(rhp, struct task_group, rcu));
8563 /* Destroy runqueue etc associated with a task group */
8564 void sched_destroy_group(struct task_group *tg)
8566 unsigned long flags;
8569 /* end participation in shares distribution */
8570 for_each_possible_cpu(i)
8571 unregister_fair_sched_group(tg, i);
8573 spin_lock_irqsave(&task_group_lock, flags);
8574 list_del_rcu(&tg->list);
8575 list_del_rcu(&tg->siblings);
8576 spin_unlock_irqrestore(&task_group_lock, flags);
8578 /* wait for possible concurrent references to cfs_rqs complete */
8579 call_rcu(&tg->rcu, free_sched_group_rcu);
8582 /* change task's runqueue when it moves between groups.
8583 * The caller of this function should have put the task in its new group
8584 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8585 * reflect its new group.
8587 void sched_move_task(struct task_struct *tsk)
8589 struct task_group *tg;
8591 unsigned long flags;
8594 rq = task_rq_lock(tsk, &flags);
8596 running = task_current(rq, tsk);
8600 dequeue_task(rq, tsk, 0);
8601 if (unlikely(running))
8602 tsk->sched_class->put_prev_task(rq, tsk);
8604 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
8605 lockdep_is_held(&tsk->sighand->siglock)),
8606 struct task_group, css);
8607 tg = autogroup_task_group(tsk, tg);
8608 tsk->sched_task_group = tg;
8610 #ifdef CONFIG_FAIR_GROUP_SCHED
8611 if (tsk->sched_class->task_move_group)
8612 tsk->sched_class->task_move_group(tsk, on_rq);
8615 set_task_rq(tsk, task_cpu(tsk));
8617 if (unlikely(running))
8618 tsk->sched_class->set_curr_task(rq);
8620 enqueue_task(rq, tsk, 0);
8622 task_rq_unlock(rq, tsk, &flags);
8624 #endif /* CONFIG_CGROUP_SCHED */
8626 #ifdef CONFIG_FAIR_GROUP_SCHED
8627 static DEFINE_MUTEX(shares_mutex);
8629 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8632 unsigned long flags;
8635 * We can't change the weight of the root cgroup.
8640 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8642 mutex_lock(&shares_mutex);
8643 if (tg->shares == shares)
8646 tg->shares = shares;
8647 for_each_possible_cpu(i) {
8648 struct rq *rq = cpu_rq(i);
8649 struct sched_entity *se;
8652 /* Propagate contribution to hierarchy */
8653 raw_spin_lock_irqsave(&rq->lock, flags);
8654 for_each_sched_entity(se)
8655 update_cfs_shares(group_cfs_rq(se));
8656 raw_spin_unlock_irqrestore(&rq->lock, flags);
8660 mutex_unlock(&shares_mutex);
8664 unsigned long sched_group_shares(struct task_group *tg)
8670 #ifdef CONFIG_RT_GROUP_SCHED
8672 * Ensure that the real time constraints are schedulable.
8674 static DEFINE_MUTEX(rt_constraints_mutex);
8676 static unsigned long to_ratio(u64 period, u64 runtime)
8678 if (runtime == RUNTIME_INF)
8681 return div64_u64(runtime << 20, period);
8684 /* Must be called with tasklist_lock held */
8685 static inline int tg_has_rt_tasks(struct task_group *tg)
8687 struct task_struct *g, *p;
8689 do_each_thread(g, p) {
8690 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8692 } while_each_thread(g, p);
8697 struct rt_schedulable_data {
8698 struct task_group *tg;
8703 static int tg_schedulable(struct task_group *tg, void *data)
8705 struct rt_schedulable_data *d = data;
8706 struct task_group *child;
8707 unsigned long total, sum = 0;
8708 u64 period, runtime;
8710 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8711 runtime = tg->rt_bandwidth.rt_runtime;
8714 period = d->rt_period;
8715 runtime = d->rt_runtime;
8719 * Cannot have more runtime than the period.
8721 if (runtime > period && runtime != RUNTIME_INF)
8725 * Ensure we don't starve existing RT tasks.
8727 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8730 total = to_ratio(period, runtime);
8733 * Nobody can have more than the global setting allows.
8735 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8739 * The sum of our children's runtime should not exceed our own.
8741 list_for_each_entry_rcu(child, &tg->children, siblings) {
8742 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8743 runtime = child->rt_bandwidth.rt_runtime;
8745 if (child == d->tg) {
8746 period = d->rt_period;
8747 runtime = d->rt_runtime;
8750 sum += to_ratio(period, runtime);
8759 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8761 struct rt_schedulable_data data = {
8763 .rt_period = period,
8764 .rt_runtime = runtime,
8767 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8770 static int tg_set_bandwidth(struct task_group *tg,
8771 u64 rt_period, u64 rt_runtime)
8775 mutex_lock(&rt_constraints_mutex);
8776 read_lock(&tasklist_lock);
8777 err = __rt_schedulable(tg, rt_period, rt_runtime);
8781 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8782 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8783 tg->rt_bandwidth.rt_runtime = rt_runtime;
8785 for_each_possible_cpu(i) {
8786 struct rt_rq *rt_rq = tg->rt_rq[i];
8788 raw_spin_lock(&rt_rq->rt_runtime_lock);
8789 rt_rq->rt_runtime = rt_runtime;
8790 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8792 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8794 read_unlock(&tasklist_lock);
8795 mutex_unlock(&rt_constraints_mutex);
8800 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8802 u64 rt_runtime, rt_period;
8804 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8805 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8806 if (rt_runtime_us < 0)
8807 rt_runtime = RUNTIME_INF;
8809 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8812 long sched_group_rt_runtime(struct task_group *tg)
8816 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8819 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8820 do_div(rt_runtime_us, NSEC_PER_USEC);
8821 return rt_runtime_us;
8824 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8826 u64 rt_runtime, rt_period;
8828 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8829 rt_runtime = tg->rt_bandwidth.rt_runtime;
8834 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8837 long sched_group_rt_period(struct task_group *tg)
8841 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8842 do_div(rt_period_us, NSEC_PER_USEC);
8843 return rt_period_us;
8846 static int sched_rt_global_constraints(void)
8848 u64 runtime, period;
8851 if (sysctl_sched_rt_period <= 0)
8854 runtime = global_rt_runtime();
8855 period = global_rt_period();
8858 * Sanity check on the sysctl variables.
8860 if (runtime > period && runtime != RUNTIME_INF)
8863 mutex_lock(&rt_constraints_mutex);
8864 read_lock(&tasklist_lock);
8865 ret = __rt_schedulable(NULL, 0, 0);
8866 read_unlock(&tasklist_lock);
8867 mutex_unlock(&rt_constraints_mutex);
8872 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8874 /* Don't accept realtime tasks when there is no way for them to run */
8875 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8881 #else /* !CONFIG_RT_GROUP_SCHED */
8882 static int sched_rt_global_constraints(void)
8884 unsigned long flags;
8887 if (sysctl_sched_rt_period <= 0)
8891 * There's always some RT tasks in the root group
8892 * -- migration, kstopmachine etc..
8894 if (sysctl_sched_rt_runtime == 0)
8897 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8898 for_each_possible_cpu(i) {
8899 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8901 raw_spin_lock(&rt_rq->rt_runtime_lock);
8902 rt_rq->rt_runtime = global_rt_runtime();
8903 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8905 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8909 #endif /* CONFIG_RT_GROUP_SCHED */
8911 int sched_rt_handler(struct ctl_table *table, int write,
8912 void __user *buffer, size_t *lenp,
8916 int old_period, old_runtime;
8917 static DEFINE_MUTEX(mutex);
8920 old_period = sysctl_sched_rt_period;
8921 old_runtime = sysctl_sched_rt_runtime;
8923 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8925 if (!ret && write) {
8926 ret = sched_rt_global_constraints();
8928 sysctl_sched_rt_period = old_period;
8929 sysctl_sched_rt_runtime = old_runtime;
8931 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8932 def_rt_bandwidth.rt_period =
8933 ns_to_ktime(global_rt_period());
8936 mutex_unlock(&mutex);
8941 #ifdef CONFIG_CGROUP_SCHED
8943 /* return corresponding task_group object of a cgroup */
8944 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8946 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8947 struct task_group, css);
8950 static struct cgroup_subsys_state *
8951 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8953 struct task_group *tg, *parent;
8955 if (!cgrp->parent) {
8956 /* This is early initialization for the top cgroup */
8957 return &root_task_group.css;
8960 parent = cgroup_tg(cgrp->parent);
8961 tg = sched_create_group(parent);
8963 return ERR_PTR(-ENOMEM);
8969 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8971 struct task_group *tg = cgroup_tg(cgrp);
8973 sched_destroy_group(tg);
8977 cpu_cgroup_allow_attach(struct cgroup *cgrp, struct task_struct *tsk)
8979 const struct cred *cred = current_cred(), *tcred;
8981 tcred = __task_cred(tsk);
8983 if ((current != tsk) && !capable(CAP_SYS_NICE) &&
8984 cred->euid != tcred->uid && cred->euid != tcred->suid)
8991 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8993 #ifdef CONFIG_RT_GROUP_SCHED
8994 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8997 /* We don't support RT-tasks being in separate groups */
8998 if (tsk->sched_class != &fair_sched_class)
9005 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9007 sched_move_task(tsk);
9011 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
9012 struct cgroup *old_cgrp, struct task_struct *task)
9015 * cgroup_exit() is called in the copy_process() failure path.
9016 * Ignore this case since the task hasn't ran yet, this avoids
9017 * trying to poke a half freed task state from generic code.
9019 if (!(task->flags & PF_EXITING))
9022 sched_move_task(task);
9025 #ifdef CONFIG_FAIR_GROUP_SCHED
9026 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9029 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
9032 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9034 struct task_group *tg = cgroup_tg(cgrp);
9036 return (u64) scale_load_down(tg->shares);
9038 #endif /* CONFIG_FAIR_GROUP_SCHED */
9040 #ifdef CONFIG_RT_GROUP_SCHED
9041 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9044 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9047 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9049 return sched_group_rt_runtime(cgroup_tg(cgrp));
9052 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9055 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9058 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9060 return sched_group_rt_period(cgroup_tg(cgrp));
9062 #endif /* CONFIG_RT_GROUP_SCHED */
9064 static struct cftype cpu_files[] = {
9065 #ifdef CONFIG_FAIR_GROUP_SCHED
9068 .read_u64 = cpu_shares_read_u64,
9069 .write_u64 = cpu_shares_write_u64,
9072 #ifdef CONFIG_RT_GROUP_SCHED
9074 .name = "rt_runtime_us",
9075 .read_s64 = cpu_rt_runtime_read,
9076 .write_s64 = cpu_rt_runtime_write,
9079 .name = "rt_period_us",
9080 .read_u64 = cpu_rt_period_read_uint,
9081 .write_u64 = cpu_rt_period_write_uint,
9086 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9088 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9091 struct cgroup_subsys cpu_cgroup_subsys = {
9093 .create = cpu_cgroup_create,
9094 .destroy = cpu_cgroup_destroy,
9095 .allow_attach = cpu_cgroup_allow_attach,
9096 .can_attach_task = cpu_cgroup_can_attach_task,
9097 .attach_task = cpu_cgroup_attach_task,
9098 .exit = cpu_cgroup_exit,
9099 .populate = cpu_cgroup_populate,
9100 .subsys_id = cpu_cgroup_subsys_id,
9104 #endif /* CONFIG_CGROUP_SCHED */
9106 #ifdef CONFIG_CGROUP_CPUACCT
9109 * CPU accounting code for task groups.
9111 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9112 * (balbir@in.ibm.com).
9115 /* track cpu usage of a group of tasks and its child groups */
9117 struct cgroup_subsys_state css;
9118 /* cpuusage holds pointer to a u64-type object on every cpu */
9119 u64 __percpu *cpuusage;
9120 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9121 struct cpuacct *parent;
9122 struct cpuacct_charge_calls *cpufreq_fn;
9126 static struct cpuacct *cpuacct_root;
9128 /* Default calls for cpufreq accounting */
9129 static struct cpuacct_charge_calls *cpuacct_cpufreq;
9130 int cpuacct_register_cpufreq(struct cpuacct_charge_calls *fn)
9132 cpuacct_cpufreq = fn;
9135 * Root node is created before platform can register callbacks,
9138 if (cpuacct_root && fn) {
9139 cpuacct_root->cpufreq_fn = fn;
9141 fn->init(&cpuacct_root->cpuacct_data);
9146 struct cgroup_subsys cpuacct_subsys;
9148 /* return cpu accounting group corresponding to this container */
9149 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9151 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9152 struct cpuacct, css);
9155 /* return cpu accounting group to which this task belongs */
9156 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9158 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9159 struct cpuacct, css);
9162 /* create a new cpu accounting group */
9163 static struct cgroup_subsys_state *cpuacct_create(
9164 struct cgroup_subsys *ss, struct cgroup *cgrp)
9166 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9172 ca->cpuusage = alloc_percpu(u64);
9176 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9177 if (percpu_counter_init(&ca->cpustat[i], 0))
9178 goto out_free_counters;
9180 ca->cpufreq_fn = cpuacct_cpufreq;
9182 /* If available, have platform code initalize cpu frequency table */
9183 if (ca->cpufreq_fn && ca->cpufreq_fn->init)
9184 ca->cpufreq_fn->init(&ca->cpuacct_data);
9187 ca->parent = cgroup_ca(cgrp->parent);
9195 percpu_counter_destroy(&ca->cpustat[i]);
9196 free_percpu(ca->cpuusage);
9200 return ERR_PTR(-ENOMEM);
9203 /* destroy an existing cpu accounting group */
9205 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9207 struct cpuacct *ca = cgroup_ca(cgrp);
9210 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9211 percpu_counter_destroy(&ca->cpustat[i]);
9212 free_percpu(ca->cpuusage);
9216 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9218 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9221 #ifndef CONFIG_64BIT
9223 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9225 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9227 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9235 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9237 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9239 #ifndef CONFIG_64BIT
9241 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9243 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9245 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9251 /* return total cpu usage (in nanoseconds) of a group */
9252 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9254 struct cpuacct *ca = cgroup_ca(cgrp);
9255 u64 totalcpuusage = 0;
9258 for_each_present_cpu(i)
9259 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9261 return totalcpuusage;
9264 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9267 struct cpuacct *ca = cgroup_ca(cgrp);
9276 for_each_present_cpu(i)
9277 cpuacct_cpuusage_write(ca, i, 0);
9283 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9286 struct cpuacct *ca = cgroup_ca(cgroup);
9290 for_each_present_cpu(i) {
9291 percpu = cpuacct_cpuusage_read(ca, i);
9292 seq_printf(m, "%llu ", (unsigned long long) percpu);
9294 seq_printf(m, "\n");
9298 static const char *cpuacct_stat_desc[] = {
9299 [CPUACCT_STAT_USER] = "user",
9300 [CPUACCT_STAT_SYSTEM] = "system",
9303 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9304 struct cgroup_map_cb *cb)
9306 struct cpuacct *ca = cgroup_ca(cgrp);
9309 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9310 s64 val = percpu_counter_read(&ca->cpustat[i]);
9311 val = cputime64_to_clock_t(val);
9312 cb->fill(cb, cpuacct_stat_desc[i], val);
9317 static int cpuacct_cpufreq_show(struct cgroup *cgrp, struct cftype *cft,
9318 struct cgroup_map_cb *cb)
9320 struct cpuacct *ca = cgroup_ca(cgrp);
9321 if (ca->cpufreq_fn && ca->cpufreq_fn->cpufreq_show)
9322 ca->cpufreq_fn->cpufreq_show(ca->cpuacct_data, cb);
9327 /* return total cpu power usage (milliWatt second) of a group */
9328 static u64 cpuacct_powerusage_read(struct cgroup *cgrp, struct cftype *cft)
9331 struct cpuacct *ca = cgroup_ca(cgrp);
9334 if (ca->cpufreq_fn && ca->cpufreq_fn->power_usage)
9335 for_each_present_cpu(i) {
9336 totalpower += ca->cpufreq_fn->power_usage(
9343 static struct cftype files[] = {
9346 .read_u64 = cpuusage_read,
9347 .write_u64 = cpuusage_write,
9350 .name = "usage_percpu",
9351 .read_seq_string = cpuacct_percpu_seq_read,
9355 .read_map = cpuacct_stats_show,
9359 .read_map = cpuacct_cpufreq_show,
9363 .read_u64 = cpuacct_powerusage_read
9367 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9369 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9373 * charge this task's execution time to its accounting group.
9375 * called with rq->lock held.
9377 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9382 if (unlikely(!cpuacct_subsys.active))
9385 cpu = task_cpu(tsk);
9391 for (; ca; ca = ca->parent) {
9392 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9393 *cpuusage += cputime;
9395 /* Call back into platform code to account for CPU speeds */
9396 if (ca->cpufreq_fn && ca->cpufreq_fn->charge)
9397 ca->cpufreq_fn->charge(ca->cpuacct_data, cputime, cpu);
9404 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9405 * in cputime_t units. As a result, cpuacct_update_stats calls
9406 * percpu_counter_add with values large enough to always overflow the
9407 * per cpu batch limit causing bad SMP scalability.
9409 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9410 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9411 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9414 #define CPUACCT_BATCH \
9415 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9417 #define CPUACCT_BATCH 0
9421 * Charge the system/user time to the task's accounting group.
9423 static void cpuacct_update_stats(struct task_struct *tsk,
9424 enum cpuacct_stat_index idx, cputime_t val)
9427 int batch = CPUACCT_BATCH;
9429 if (unlikely(!cpuacct_subsys.active))
9436 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9442 struct cgroup_subsys cpuacct_subsys = {
9444 .create = cpuacct_create,
9445 .destroy = cpuacct_destroy,
9446 .populate = cpuacct_populate,
9447 .subsys_id = cpuacct_subsys_id,
9449 #endif /* CONFIG_CGROUP_CPUACCT */