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>
76 #include <asm/irq_regs.h>
77 #include <asm/mutex.h>
79 #include "sched_cpupri.h"
80 #include "workqueue_sched.h"
81 #include "sched_autogroup.h"
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/sched.h>
87 * Convert user-nice values [ -20 ... 0 ... 19 ]
88 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
91 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
92 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
93 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
96 * 'User priority' is the nice value converted to something we
97 * can work with better when scaling various scheduler parameters,
98 * it's a [ 0 ... 39 ] range.
100 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
101 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
102 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
105 * Helpers for converting nanosecond timing to jiffy resolution
107 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
109 #define NICE_0_LOAD SCHED_LOAD_SCALE
110 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
113 * These are the 'tuning knobs' of the scheduler:
115 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
116 * Timeslices get refilled after they expire.
118 #define DEF_TIMESLICE (100 * HZ / 1000)
121 * single value that denotes runtime == period, ie unlimited time.
123 #define RUNTIME_INF ((u64)~0ULL)
125 static inline int rt_policy(int policy)
127 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
132 static inline int task_has_rt_policy(struct task_struct *p)
134 return rt_policy(p->policy);
138 * This is the priority-queue data structure of the RT scheduling class:
140 struct rt_prio_array {
141 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
142 struct list_head queue[MAX_RT_PRIO];
145 struct rt_bandwidth {
146 /* nests inside the rq lock: */
147 raw_spinlock_t rt_runtime_lock;
150 struct hrtimer rt_period_timer;
153 static struct rt_bandwidth def_rt_bandwidth;
155 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
157 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
159 struct rt_bandwidth *rt_b =
160 container_of(timer, struct rt_bandwidth, rt_period_timer);
166 now = hrtimer_cb_get_time(timer);
167 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
172 idle = do_sched_rt_period_timer(rt_b, overrun);
175 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
179 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
181 rt_b->rt_period = ns_to_ktime(period);
182 rt_b->rt_runtime = runtime;
184 raw_spin_lock_init(&rt_b->rt_runtime_lock);
186 hrtimer_init(&rt_b->rt_period_timer,
187 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
188 rt_b->rt_period_timer.function = sched_rt_period_timer;
191 static inline int rt_bandwidth_enabled(void)
193 return sysctl_sched_rt_runtime >= 0;
196 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
200 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
203 if (hrtimer_active(&rt_b->rt_period_timer))
206 raw_spin_lock(&rt_b->rt_runtime_lock);
211 if (hrtimer_active(&rt_b->rt_period_timer))
214 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
215 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
217 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
218 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
219 delta = ktime_to_ns(ktime_sub(hard, soft));
220 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
221 HRTIMER_MODE_ABS_PINNED, 0);
223 raw_spin_unlock(&rt_b->rt_runtime_lock);
226 #ifdef CONFIG_RT_GROUP_SCHED
227 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
229 hrtimer_cancel(&rt_b->rt_period_timer);
234 * sched_domains_mutex serializes calls to init_sched_domains,
235 * detach_destroy_domains and partition_sched_domains.
237 static DEFINE_MUTEX(sched_domains_mutex);
239 #ifdef CONFIG_CGROUP_SCHED
241 #include <linux/cgroup.h>
245 static LIST_HEAD(task_groups);
247 /* task group related information */
249 struct cgroup_subsys_state css;
251 #ifdef CONFIG_FAIR_GROUP_SCHED
252 /* schedulable entities of this group on each cpu */
253 struct sched_entity **se;
254 /* runqueue "owned" by this group on each cpu */
255 struct cfs_rq **cfs_rq;
256 unsigned long shares;
258 atomic_t load_weight;
261 #ifdef CONFIG_RT_GROUP_SCHED
262 struct sched_rt_entity **rt_se;
263 struct rt_rq **rt_rq;
265 struct rt_bandwidth rt_bandwidth;
269 struct list_head list;
271 struct task_group *parent;
272 struct list_head siblings;
273 struct list_head children;
275 #ifdef CONFIG_SCHED_AUTOGROUP
276 struct autogroup *autogroup;
280 /* task_group_lock serializes the addition/removal of task groups */
281 static DEFINE_SPINLOCK(task_group_lock);
283 #ifdef CONFIG_FAIR_GROUP_SCHED
285 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
288 * A weight of 0 or 1 can cause arithmetics problems.
289 * A weight of a cfs_rq is the sum of weights of which entities
290 * are queued on this cfs_rq, so a weight of a entity should not be
291 * too large, so as the shares value of a task group.
292 * (The default weight is 1024 - so there's no practical
293 * limitation from this.)
295 #define MIN_SHARES (1UL << 1)
296 #define MAX_SHARES (1UL << 18)
298 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
301 /* Default task group.
302 * Every task in system belong to this group at bootup.
304 struct task_group root_task_group;
306 #endif /* CONFIG_CGROUP_SCHED */
308 /* CFS-related fields in a runqueue */
310 struct load_weight load;
311 unsigned long nr_running;
316 u64 min_vruntime_copy;
319 struct rb_root tasks_timeline;
320 struct rb_node *rb_leftmost;
322 struct list_head tasks;
323 struct list_head *balance_iterator;
326 * 'curr' points to currently running entity on this cfs_rq.
327 * It is set to NULL otherwise (i.e when none are currently running).
329 struct sched_entity *curr, *next, *last, *skip;
331 #ifdef CONFIG_SCHED_DEBUG
332 unsigned int nr_spread_over;
335 #ifdef CONFIG_FAIR_GROUP_SCHED
336 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
339 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
340 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
341 * (like users, containers etc.)
343 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
344 * list is used during load balance.
347 struct list_head leaf_cfs_rq_list;
348 struct task_group *tg; /* group that "owns" this runqueue */
352 * the part of load.weight contributed by tasks
354 unsigned long task_weight;
357 * h_load = weight * f(tg)
359 * Where f(tg) is the recursive weight fraction assigned to
362 unsigned long h_load;
365 * Maintaining per-cpu shares distribution for group scheduling
367 * load_stamp is the last time we updated the load average
368 * load_last is the last time we updated the load average and saw load
369 * load_unacc_exec_time is currently unaccounted execution time
373 u64 load_stamp, load_last, load_unacc_exec_time;
375 unsigned long load_contribution;
380 /* Real-Time classes' related field in a runqueue: */
382 struct rt_prio_array active;
383 unsigned long rt_nr_running;
384 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
386 int curr; /* highest queued rt task prio */
388 int next; /* next highest */
393 unsigned long rt_nr_migratory;
394 unsigned long rt_nr_total;
396 struct plist_head pushable_tasks;
401 /* Nests inside the rq lock: */
402 raw_spinlock_t rt_runtime_lock;
404 #ifdef CONFIG_RT_GROUP_SCHED
405 unsigned long rt_nr_boosted;
408 struct list_head leaf_rt_rq_list;
409 struct task_group *tg;
416 * We add the notion of a root-domain which will be used to define per-domain
417 * variables. Each exclusive cpuset essentially defines an island domain by
418 * fully partitioning the member cpus from any other cpuset. Whenever a new
419 * exclusive cpuset is created, we also create and attach a new root-domain
427 cpumask_var_t online;
430 * The "RT overload" flag: it gets set if a CPU has more than
431 * one runnable RT task.
433 cpumask_var_t rto_mask;
435 struct cpupri cpupri;
439 * By default the system creates a single root-domain with all cpus as
440 * members (mimicking the global state we have today).
442 static struct root_domain def_root_domain;
444 #endif /* CONFIG_SMP */
447 * This is the main, per-CPU runqueue data structure.
449 * Locking rule: those places that want to lock multiple runqueues
450 * (such as the load balancing or the thread migration code), lock
451 * acquire operations must be ordered by ascending &runqueue.
458 * nr_running and cpu_load should be in the same cacheline because
459 * remote CPUs use both these fields when doing load calculation.
461 unsigned long nr_running;
462 #define CPU_LOAD_IDX_MAX 5
463 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
464 unsigned long last_load_update_tick;
467 unsigned char nohz_balance_kick;
469 int skip_clock_update;
471 /* capture load from *all* tasks on this cpu: */
472 struct load_weight load;
473 unsigned long nr_load_updates;
479 #ifdef CONFIG_FAIR_GROUP_SCHED
480 /* list of leaf cfs_rq on this cpu: */
481 struct list_head leaf_cfs_rq_list;
483 #ifdef CONFIG_RT_GROUP_SCHED
484 struct list_head leaf_rt_rq_list;
488 * This is part of a global counter where only the total sum
489 * over all CPUs matters. A task can increase this counter on
490 * one CPU and if it got migrated afterwards it may decrease
491 * it on another CPU. Always updated under the runqueue lock:
493 unsigned long nr_uninterruptible;
495 struct task_struct *curr, *idle, *stop;
496 unsigned long next_balance;
497 struct mm_struct *prev_mm;
505 struct root_domain *rd;
506 struct sched_domain *sd;
508 unsigned long cpu_power;
510 unsigned char idle_at_tick;
511 /* For active balancing */
515 struct cpu_stop_work active_balance_work;
516 /* cpu of this runqueue: */
520 unsigned long avg_load_per_task;
528 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
532 /* calc_load related fields */
533 unsigned long calc_load_update;
534 long calc_load_active;
536 #ifdef CONFIG_SCHED_HRTICK
538 int hrtick_csd_pending;
539 struct call_single_data hrtick_csd;
541 struct hrtimer hrtick_timer;
544 #ifdef CONFIG_SCHEDSTATS
546 struct sched_info rq_sched_info;
547 unsigned long long rq_cpu_time;
548 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
550 /* sys_sched_yield() stats */
551 unsigned int yld_count;
553 /* schedule() stats */
554 unsigned int sched_switch;
555 unsigned int sched_count;
556 unsigned int sched_goidle;
558 /* try_to_wake_up() stats */
559 unsigned int ttwu_count;
560 unsigned int ttwu_local;
564 struct task_struct *wake_list;
568 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
571 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
573 static inline int cpu_of(struct rq *rq)
582 #define rcu_dereference_check_sched_domain(p) \
583 rcu_dereference_check((p), \
584 rcu_read_lock_held() || \
585 lockdep_is_held(&sched_domains_mutex))
588 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
589 * See detach_destroy_domains: synchronize_sched for details.
591 * The domain tree of any CPU may only be accessed from within
592 * preempt-disabled sections.
594 #define for_each_domain(cpu, __sd) \
595 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
597 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
598 #define this_rq() (&__get_cpu_var(runqueues))
599 #define task_rq(p) cpu_rq(task_cpu(p))
600 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
601 #define raw_rq() (&__raw_get_cpu_var(runqueues))
603 #ifdef CONFIG_CGROUP_SCHED
606 * Return the group to which this tasks belongs.
608 * We cannot use task_subsys_state() and friends because the cgroup
609 * subsystem changes that value before the cgroup_subsys::attach() method
610 * is called, therefore we cannot pin it and might observe the wrong value.
612 * The same is true for autogroup's p->signal->autogroup->tg, the autogroup
613 * core changes this before calling sched_move_task().
615 * Instead we use a 'copy' which is updated from sched_move_task() while
616 * holding both task_struct::pi_lock and rq::lock.
618 static inline struct task_group *task_group(struct task_struct *p)
620 return p->sched_task_group;
623 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
624 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
626 #ifdef CONFIG_FAIR_GROUP_SCHED
627 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
628 p->se.parent = task_group(p)->se[cpu];
631 #ifdef CONFIG_RT_GROUP_SCHED
632 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
633 p->rt.parent = task_group(p)->rt_se[cpu];
637 #else /* CONFIG_CGROUP_SCHED */
639 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
640 static inline struct task_group *task_group(struct task_struct *p)
645 #endif /* CONFIG_CGROUP_SCHED */
647 static void update_rq_clock_task(struct rq *rq, s64 delta);
649 static void update_rq_clock(struct rq *rq)
653 if (rq->skip_clock_update > 0)
656 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
658 update_rq_clock_task(rq, delta);
662 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
664 #ifdef CONFIG_SCHED_DEBUG
665 # define const_debug __read_mostly
667 # define const_debug static const
671 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
672 * @cpu: the processor in question.
674 * This interface allows printk to be called with the runqueue lock
675 * held and know whether or not it is OK to wake up the klogd.
677 int runqueue_is_locked(int cpu)
679 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
683 * Debugging: various feature bits
686 #define SCHED_FEAT(name, enabled) \
687 __SCHED_FEAT_##name ,
690 #include "sched_features.h"
695 #define SCHED_FEAT(name, enabled) \
696 (1UL << __SCHED_FEAT_##name) * enabled |
698 const_debug unsigned int sysctl_sched_features =
699 #include "sched_features.h"
704 #ifdef CONFIG_SCHED_DEBUG
705 #define SCHED_FEAT(name, enabled) \
708 static __read_mostly char *sched_feat_names[] = {
709 #include "sched_features.h"
715 static int sched_feat_show(struct seq_file *m, void *v)
719 for (i = 0; sched_feat_names[i]; i++) {
720 if (!(sysctl_sched_features & (1UL << i)))
722 seq_printf(m, "%s ", sched_feat_names[i]);
730 sched_feat_write(struct file *filp, const char __user *ubuf,
731 size_t cnt, loff_t *ppos)
741 if (copy_from_user(&buf, ubuf, cnt))
747 if (strncmp(cmp, "NO_", 3) == 0) {
752 for (i = 0; sched_feat_names[i]; i++) {
753 if (strcmp(cmp, sched_feat_names[i]) == 0) {
755 sysctl_sched_features &= ~(1UL << i);
757 sysctl_sched_features |= (1UL << i);
762 if (!sched_feat_names[i])
770 static int sched_feat_open(struct inode *inode, struct file *filp)
772 return single_open(filp, sched_feat_show, NULL);
775 static const struct file_operations sched_feat_fops = {
776 .open = sched_feat_open,
777 .write = sched_feat_write,
780 .release = single_release,
783 static __init int sched_init_debug(void)
785 debugfs_create_file("sched_features", 0644, NULL, NULL,
790 late_initcall(sched_init_debug);
794 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
797 * Number of tasks to iterate in a single balance run.
798 * Limited because this is done with IRQs disabled.
800 const_debug unsigned int sysctl_sched_nr_migrate = 32;
803 * period over which we average the RT time consumption, measured
808 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
811 * period over which we measure -rt task cpu usage in us.
814 unsigned int sysctl_sched_rt_period = 1000000;
816 static __read_mostly int scheduler_running;
819 * part of the period that we allow rt tasks to run in us.
822 int sysctl_sched_rt_runtime = 950000;
824 static inline u64 global_rt_period(void)
826 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
829 static inline u64 global_rt_runtime(void)
831 if (sysctl_sched_rt_runtime < 0)
834 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
837 #ifndef prepare_arch_switch
838 # define prepare_arch_switch(next) do { } while (0)
840 #ifndef finish_arch_switch
841 # define finish_arch_switch(prev) do { } while (0)
844 static inline int task_current(struct rq *rq, struct task_struct *p)
846 return rq->curr == p;
849 static inline int task_running(struct rq *rq, struct task_struct *p)
854 return task_current(rq, p);
858 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
859 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
863 * We can optimise this out completely for !SMP, because the
864 * SMP rebalancing from interrupt is the only thing that cares
871 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
875 * After ->on_cpu is cleared, the task can be moved to a different CPU.
876 * We must ensure this doesn't happen until the switch is completely
882 #ifdef CONFIG_DEBUG_SPINLOCK
883 /* this is a valid case when another task releases the spinlock */
884 rq->lock.owner = current;
887 * If we are tracking spinlock dependencies then we have to
888 * fix up the runqueue lock - which gets 'carried over' from
891 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
893 raw_spin_unlock_irq(&rq->lock);
896 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
897 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
901 * We can optimise this out completely for !SMP, because the
902 * SMP rebalancing from interrupt is the only thing that cares
907 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
908 raw_spin_unlock_irq(&rq->lock);
910 raw_spin_unlock(&rq->lock);
914 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
918 * After ->on_cpu is cleared, the task can be moved to a different CPU.
919 * We must ensure this doesn't happen until the switch is completely
925 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
929 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
932 * __task_rq_lock - lock the rq @p resides on.
934 static inline struct rq *__task_rq_lock(struct task_struct *p)
939 lockdep_assert_held(&p->pi_lock);
943 raw_spin_lock(&rq->lock);
944 if (likely(rq == task_rq(p)))
946 raw_spin_unlock(&rq->lock);
951 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
953 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
954 __acquires(p->pi_lock)
960 raw_spin_lock_irqsave(&p->pi_lock, *flags);
962 raw_spin_lock(&rq->lock);
963 if (likely(rq == task_rq(p)))
965 raw_spin_unlock(&rq->lock);
966 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
970 static void __task_rq_unlock(struct rq *rq)
973 raw_spin_unlock(&rq->lock);
977 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
979 __releases(p->pi_lock)
981 raw_spin_unlock(&rq->lock);
982 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
986 * this_rq_lock - lock this runqueue and disable interrupts.
988 static struct rq *this_rq_lock(void)
995 raw_spin_lock(&rq->lock);
1000 #ifdef CONFIG_SCHED_HRTICK
1002 * Use HR-timers to deliver accurate preemption points.
1004 * Its all a bit involved since we cannot program an hrt while holding the
1005 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1008 * When we get rescheduled we reprogram the hrtick_timer outside of the
1014 * - enabled by features
1015 * - hrtimer is actually high res
1017 static inline int hrtick_enabled(struct rq *rq)
1019 if (!sched_feat(HRTICK))
1021 if (!cpu_active(cpu_of(rq)))
1023 return hrtimer_is_hres_active(&rq->hrtick_timer);
1026 static void hrtick_clear(struct rq *rq)
1028 if (hrtimer_active(&rq->hrtick_timer))
1029 hrtimer_cancel(&rq->hrtick_timer);
1033 * High-resolution timer tick.
1034 * Runs from hardirq context with interrupts disabled.
1036 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1038 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1040 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1042 raw_spin_lock(&rq->lock);
1043 update_rq_clock(rq);
1044 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1045 raw_spin_unlock(&rq->lock);
1047 return HRTIMER_NORESTART;
1052 * called from hardirq (IPI) context
1054 static void __hrtick_start(void *arg)
1056 struct rq *rq = arg;
1058 raw_spin_lock(&rq->lock);
1059 hrtimer_restart(&rq->hrtick_timer);
1060 rq->hrtick_csd_pending = 0;
1061 raw_spin_unlock(&rq->lock);
1065 * Called to set the hrtick timer state.
1067 * called with rq->lock held and irqs disabled
1069 static void hrtick_start(struct rq *rq, u64 delay)
1071 struct hrtimer *timer = &rq->hrtick_timer;
1072 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1074 hrtimer_set_expires(timer, time);
1076 if (rq == this_rq()) {
1077 hrtimer_restart(timer);
1078 } else if (!rq->hrtick_csd_pending) {
1079 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1080 rq->hrtick_csd_pending = 1;
1085 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1087 int cpu = (int)(long)hcpu;
1090 case CPU_UP_CANCELED:
1091 case CPU_UP_CANCELED_FROZEN:
1092 case CPU_DOWN_PREPARE:
1093 case CPU_DOWN_PREPARE_FROZEN:
1095 case CPU_DEAD_FROZEN:
1096 hrtick_clear(cpu_rq(cpu));
1103 static __init void init_hrtick(void)
1105 hotcpu_notifier(hotplug_hrtick, 0);
1109 * Called to set the hrtick timer state.
1111 * called with rq->lock held and irqs disabled
1113 static void hrtick_start(struct rq *rq, u64 delay)
1115 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1116 HRTIMER_MODE_REL_PINNED, 0);
1119 static inline void init_hrtick(void)
1122 #endif /* CONFIG_SMP */
1124 static void init_rq_hrtick(struct rq *rq)
1127 rq->hrtick_csd_pending = 0;
1129 rq->hrtick_csd.flags = 0;
1130 rq->hrtick_csd.func = __hrtick_start;
1131 rq->hrtick_csd.info = rq;
1134 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1135 rq->hrtick_timer.function = hrtick;
1137 #else /* CONFIG_SCHED_HRTICK */
1138 static inline void hrtick_clear(struct rq *rq)
1142 static inline void init_rq_hrtick(struct rq *rq)
1146 static inline void init_hrtick(void)
1149 #endif /* CONFIG_SCHED_HRTICK */
1152 * resched_task - mark a task 'to be rescheduled now'.
1154 * On UP this means the setting of the need_resched flag, on SMP it
1155 * might also involve a cross-CPU call to trigger the scheduler on
1160 #ifndef tsk_is_polling
1161 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1164 static void resched_task(struct task_struct *p)
1168 assert_raw_spin_locked(&task_rq(p)->lock);
1170 if (test_tsk_need_resched(p))
1173 set_tsk_need_resched(p);
1176 if (cpu == smp_processor_id())
1179 /* NEED_RESCHED must be visible before we test polling */
1181 if (!tsk_is_polling(p))
1182 smp_send_reschedule(cpu);
1185 static void resched_cpu(int cpu)
1187 struct rq *rq = cpu_rq(cpu);
1188 unsigned long flags;
1190 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1192 resched_task(cpu_curr(cpu));
1193 raw_spin_unlock_irqrestore(&rq->lock, flags);
1198 * In the semi idle case, use the nearest busy cpu for migrating timers
1199 * from an idle cpu. This is good for power-savings.
1201 * We don't do similar optimization for completely idle system, as
1202 * selecting an idle cpu will add more delays to the timers than intended
1203 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1205 int get_nohz_timer_target(void)
1207 int cpu = smp_processor_id();
1209 struct sched_domain *sd;
1212 for_each_domain(cpu, sd) {
1213 for_each_cpu(i, sched_domain_span(sd)) {
1225 * When add_timer_on() enqueues a timer into the timer wheel of an
1226 * idle CPU then this timer might expire before the next timer event
1227 * which is scheduled to wake up that CPU. In case of a completely
1228 * idle system the next event might even be infinite time into the
1229 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1230 * leaves the inner idle loop so the newly added timer is taken into
1231 * account when the CPU goes back to idle and evaluates the timer
1232 * wheel for the next timer event.
1234 void wake_up_idle_cpu(int cpu)
1236 struct rq *rq = cpu_rq(cpu);
1238 if (cpu == smp_processor_id())
1242 * This is safe, as this function is called with the timer
1243 * wheel base lock of (cpu) held. When the CPU is on the way
1244 * to idle and has not yet set rq->curr to idle then it will
1245 * be serialized on the timer wheel base lock and take the new
1246 * timer into account automatically.
1248 if (rq->curr != rq->idle)
1252 * We can set TIF_RESCHED on the idle task of the other CPU
1253 * lockless. The worst case is that the other CPU runs the
1254 * idle task through an additional NOOP schedule()
1256 set_tsk_need_resched(rq->idle);
1258 /* NEED_RESCHED must be visible before we test polling */
1260 if (!tsk_is_polling(rq->idle))
1261 smp_send_reschedule(cpu);
1264 #endif /* CONFIG_NO_HZ */
1266 static u64 sched_avg_period(void)
1268 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1271 static void sched_avg_update(struct rq *rq)
1273 s64 period = sched_avg_period();
1275 while ((s64)(rq->clock - rq->age_stamp) > period) {
1277 * Inline assembly required to prevent the compiler
1278 * optimising this loop into a divmod call.
1279 * See __iter_div_u64_rem() for another example of this.
1281 asm("" : "+rm" (rq->age_stamp));
1282 rq->age_stamp += period;
1287 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1289 rq->rt_avg += rt_delta;
1290 sched_avg_update(rq);
1293 #else /* !CONFIG_SMP */
1294 static void resched_task(struct task_struct *p)
1296 assert_raw_spin_locked(&task_rq(p)->lock);
1297 set_tsk_need_resched(p);
1300 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1304 static void sched_avg_update(struct rq *rq)
1307 #endif /* CONFIG_SMP */
1309 #if BITS_PER_LONG == 32
1310 # define WMULT_CONST (~0UL)
1312 # define WMULT_CONST (1UL << 32)
1315 #define WMULT_SHIFT 32
1318 * Shift right and round:
1320 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1323 * delta *= weight / lw
1325 static unsigned long
1326 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1327 struct load_weight *lw)
1332 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1333 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1334 * 2^SCHED_LOAD_RESOLUTION.
1336 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
1337 tmp = (u64)delta_exec * scale_load_down(weight);
1339 tmp = (u64)delta_exec;
1341 if (!lw->inv_weight) {
1342 unsigned long w = scale_load_down(lw->weight);
1344 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
1346 else if (unlikely(!w))
1347 lw->inv_weight = WMULT_CONST;
1349 lw->inv_weight = WMULT_CONST / w;
1353 * Check whether we'd overflow the 64-bit multiplication:
1355 if (unlikely(tmp > WMULT_CONST))
1356 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1359 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1361 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1364 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1370 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1376 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1383 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1384 * of tasks with abnormal "nice" values across CPUs the contribution that
1385 * each task makes to its run queue's load is weighted according to its
1386 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1387 * scaled version of the new time slice allocation that they receive on time
1391 #define WEIGHT_IDLEPRIO 3
1392 #define WMULT_IDLEPRIO 1431655765
1395 * Nice levels are multiplicative, with a gentle 10% change for every
1396 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1397 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1398 * that remained on nice 0.
1400 * The "10% effect" is relative and cumulative: from _any_ nice level,
1401 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1402 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1403 * If a task goes up by ~10% and another task goes down by ~10% then
1404 * the relative distance between them is ~25%.)
1406 static const int prio_to_weight[40] = {
1407 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1408 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1409 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1410 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1411 /* 0 */ 1024, 820, 655, 526, 423,
1412 /* 5 */ 335, 272, 215, 172, 137,
1413 /* 10 */ 110, 87, 70, 56, 45,
1414 /* 15 */ 36, 29, 23, 18, 15,
1418 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1420 * In cases where the weight does not change often, we can use the
1421 * precalculated inverse to speed up arithmetics by turning divisions
1422 * into multiplications:
1424 static const u32 prio_to_wmult[40] = {
1425 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1426 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1427 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1428 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1429 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1430 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1431 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1432 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1435 /* Time spent by the tasks of the cpu accounting group executing in ... */
1436 enum cpuacct_stat_index {
1437 CPUACCT_STAT_USER, /* ... user mode */
1438 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1440 CPUACCT_STAT_NSTATS,
1443 #ifdef CONFIG_CGROUP_CPUACCT
1444 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1445 static void cpuacct_update_stats(struct task_struct *tsk,
1446 enum cpuacct_stat_index idx, cputime_t val);
1448 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1449 static inline void cpuacct_update_stats(struct task_struct *tsk,
1450 enum cpuacct_stat_index idx, cputime_t val) {}
1453 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1455 update_load_add(&rq->load, load);
1458 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1460 update_load_sub(&rq->load, load);
1463 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1464 typedef int (*tg_visitor)(struct task_group *, void *);
1467 * Iterate the full tree, calling @down when first entering a node and @up when
1468 * leaving it for the final time.
1470 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1472 struct task_group *parent, *child;
1476 parent = &root_task_group;
1478 ret = (*down)(parent, data);
1481 list_for_each_entry_rcu(child, &parent->children, siblings) {
1488 ret = (*up)(parent, data);
1493 parent = parent->parent;
1502 static int tg_nop(struct task_group *tg, void *data)
1509 /* Used instead of source_load when we know the type == 0 */
1510 static unsigned long weighted_cpuload(const int cpu)
1512 return cpu_rq(cpu)->load.weight;
1516 * Return a low guess at the load of a migration-source cpu weighted
1517 * according to the scheduling class and "nice" value.
1519 * We want to under-estimate the load of migration sources, to
1520 * balance conservatively.
1522 static unsigned long source_load(int cpu, int type)
1524 struct rq *rq = cpu_rq(cpu);
1525 unsigned long total = weighted_cpuload(cpu);
1527 if (type == 0 || !sched_feat(LB_BIAS))
1530 return min(rq->cpu_load[type-1], total);
1534 * Return a high guess at the load of a migration-target cpu weighted
1535 * according to the scheduling class and "nice" value.
1537 static unsigned long target_load(int cpu, int type)
1539 struct rq *rq = cpu_rq(cpu);
1540 unsigned long total = weighted_cpuload(cpu);
1542 if (type == 0 || !sched_feat(LB_BIAS))
1545 return max(rq->cpu_load[type-1], total);
1548 static unsigned long power_of(int cpu)
1550 return cpu_rq(cpu)->cpu_power;
1553 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1555 static unsigned long cpu_avg_load_per_task(int cpu)
1557 struct rq *rq = cpu_rq(cpu);
1558 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1561 rq->avg_load_per_task = rq->load.weight / nr_running;
1563 rq->avg_load_per_task = 0;
1565 return rq->avg_load_per_task;
1568 #ifdef CONFIG_FAIR_GROUP_SCHED
1571 * Compute the cpu's hierarchical load factor for each task group.
1572 * This needs to be done in a top-down fashion because the load of a child
1573 * group is a fraction of its parents load.
1575 static int tg_load_down(struct task_group *tg, void *data)
1578 long cpu = (long)data;
1581 load = cpu_rq(cpu)->load.weight;
1583 load = tg->parent->cfs_rq[cpu]->h_load;
1584 load *= tg->se[cpu]->load.weight;
1585 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1588 tg->cfs_rq[cpu]->h_load = load;
1593 static void update_h_load(long cpu)
1595 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1600 #ifdef CONFIG_PREEMPT
1602 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1605 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1606 * way at the expense of forcing extra atomic operations in all
1607 * invocations. This assures that the double_lock is acquired using the
1608 * same underlying policy as the spinlock_t on this architecture, which
1609 * reduces latency compared to the unfair variant below. However, it
1610 * also adds more overhead and therefore may reduce throughput.
1612 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1613 __releases(this_rq->lock)
1614 __acquires(busiest->lock)
1615 __acquires(this_rq->lock)
1617 raw_spin_unlock(&this_rq->lock);
1618 double_rq_lock(this_rq, busiest);
1625 * Unfair double_lock_balance: Optimizes throughput at the expense of
1626 * latency by eliminating extra atomic operations when the locks are
1627 * already in proper order on entry. This favors lower cpu-ids and will
1628 * grant the double lock to lower cpus over higher ids under contention,
1629 * regardless of entry order into the function.
1631 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1632 __releases(this_rq->lock)
1633 __acquires(busiest->lock)
1634 __acquires(this_rq->lock)
1638 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1639 if (busiest < this_rq) {
1640 raw_spin_unlock(&this_rq->lock);
1641 raw_spin_lock(&busiest->lock);
1642 raw_spin_lock_nested(&this_rq->lock,
1643 SINGLE_DEPTH_NESTING);
1646 raw_spin_lock_nested(&busiest->lock,
1647 SINGLE_DEPTH_NESTING);
1652 #endif /* CONFIG_PREEMPT */
1655 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1657 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1659 if (unlikely(!irqs_disabled())) {
1660 /* printk() doesn't work good under rq->lock */
1661 raw_spin_unlock(&this_rq->lock);
1665 return _double_lock_balance(this_rq, busiest);
1668 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1669 __releases(busiest->lock)
1671 raw_spin_unlock(&busiest->lock);
1672 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1676 * double_rq_lock - safely lock two runqueues
1678 * Note this does not disable interrupts like task_rq_lock,
1679 * you need to do so manually before calling.
1681 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1682 __acquires(rq1->lock)
1683 __acquires(rq2->lock)
1685 BUG_ON(!irqs_disabled());
1687 raw_spin_lock(&rq1->lock);
1688 __acquire(rq2->lock); /* Fake it out ;) */
1691 raw_spin_lock(&rq1->lock);
1692 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1694 raw_spin_lock(&rq2->lock);
1695 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1701 * double_rq_unlock - safely unlock two runqueues
1703 * Note this does not restore interrupts like task_rq_unlock,
1704 * you need to do so manually after calling.
1706 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1707 __releases(rq1->lock)
1708 __releases(rq2->lock)
1710 raw_spin_unlock(&rq1->lock);
1712 raw_spin_unlock(&rq2->lock);
1714 __release(rq2->lock);
1717 #else /* CONFIG_SMP */
1720 * double_rq_lock - safely lock two runqueues
1722 * Note this does not disable interrupts like task_rq_lock,
1723 * you need to do so manually before calling.
1725 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1726 __acquires(rq1->lock)
1727 __acquires(rq2->lock)
1729 BUG_ON(!irqs_disabled());
1731 raw_spin_lock(&rq1->lock);
1732 __acquire(rq2->lock); /* Fake it out ;) */
1736 * double_rq_unlock - safely unlock two runqueues
1738 * Note this does not restore interrupts like task_rq_unlock,
1739 * you need to do so manually after calling.
1741 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1742 __releases(rq1->lock)
1743 __releases(rq2->lock)
1746 raw_spin_unlock(&rq1->lock);
1747 __release(rq2->lock);
1752 static void calc_load_account_idle(struct rq *this_rq);
1753 static void update_sysctl(void);
1754 static int get_update_sysctl_factor(void);
1755 static void update_cpu_load(struct rq *this_rq);
1757 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1759 set_task_rq(p, cpu);
1762 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1763 * successfuly executed on another CPU. We must ensure that updates of
1764 * per-task data have been completed by this moment.
1767 task_thread_info(p)->cpu = cpu;
1771 static const struct sched_class rt_sched_class;
1773 #define sched_class_highest (&stop_sched_class)
1774 #define for_each_class(class) \
1775 for (class = sched_class_highest; class; class = class->next)
1777 #include "sched_stats.h"
1779 static void inc_nr_running(struct rq *rq)
1784 static void dec_nr_running(struct rq *rq)
1789 static void set_load_weight(struct task_struct *p)
1791 int prio = p->static_prio - MAX_RT_PRIO;
1792 struct load_weight *load = &p->se.load;
1795 * SCHED_IDLE tasks get minimal weight:
1797 if (p->policy == SCHED_IDLE) {
1798 load->weight = scale_load(WEIGHT_IDLEPRIO);
1799 load->inv_weight = WMULT_IDLEPRIO;
1803 load->weight = scale_load(prio_to_weight[prio]);
1804 load->inv_weight = prio_to_wmult[prio];
1807 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1809 update_rq_clock(rq);
1810 sched_info_queued(p);
1811 p->sched_class->enqueue_task(rq, p, flags);
1814 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1816 update_rq_clock(rq);
1817 sched_info_dequeued(p);
1818 p->sched_class->dequeue_task(rq, p, flags);
1822 * activate_task - move a task to the runqueue.
1824 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1826 if (task_contributes_to_load(p))
1827 rq->nr_uninterruptible--;
1829 enqueue_task(rq, p, flags);
1834 * deactivate_task - remove a task from the runqueue.
1836 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1838 if (task_contributes_to_load(p))
1839 rq->nr_uninterruptible++;
1841 dequeue_task(rq, p, flags);
1845 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1848 * There are no locks covering percpu hardirq/softirq time.
1849 * They are only modified in account_system_vtime, on corresponding CPU
1850 * with interrupts disabled. So, writes are safe.
1851 * They are read and saved off onto struct rq in update_rq_clock().
1852 * This may result in other CPU reading this CPU's irq time and can
1853 * race with irq/account_system_vtime on this CPU. We would either get old
1854 * or new value with a side effect of accounting a slice of irq time to wrong
1855 * task when irq is in progress while we read rq->clock. That is a worthy
1856 * compromise in place of having locks on each irq in account_system_time.
1858 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1859 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1861 static DEFINE_PER_CPU(u64, irq_start_time);
1862 static int sched_clock_irqtime;
1864 void enable_sched_clock_irqtime(void)
1866 sched_clock_irqtime = 1;
1869 void disable_sched_clock_irqtime(void)
1871 sched_clock_irqtime = 0;
1874 #ifndef CONFIG_64BIT
1875 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1877 static inline void irq_time_write_begin(void)
1879 __this_cpu_inc(irq_time_seq.sequence);
1883 static inline void irq_time_write_end(void)
1886 __this_cpu_inc(irq_time_seq.sequence);
1889 static inline u64 irq_time_read(int cpu)
1895 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1896 irq_time = per_cpu(cpu_softirq_time, cpu) +
1897 per_cpu(cpu_hardirq_time, cpu);
1898 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1902 #else /* CONFIG_64BIT */
1903 static inline void irq_time_write_begin(void)
1907 static inline void irq_time_write_end(void)
1911 static inline u64 irq_time_read(int cpu)
1913 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1915 #endif /* CONFIG_64BIT */
1918 * Called before incrementing preempt_count on {soft,}irq_enter
1919 * and before decrementing preempt_count on {soft,}irq_exit.
1921 void account_system_vtime(struct task_struct *curr)
1923 unsigned long flags;
1927 if (!sched_clock_irqtime)
1930 local_irq_save(flags);
1932 cpu = smp_processor_id();
1933 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1934 __this_cpu_add(irq_start_time, delta);
1936 irq_time_write_begin();
1938 * We do not account for softirq time from ksoftirqd here.
1939 * We want to continue accounting softirq time to ksoftirqd thread
1940 * in that case, so as not to confuse scheduler with a special task
1941 * that do not consume any time, but still wants to run.
1943 if (hardirq_count())
1944 __this_cpu_add(cpu_hardirq_time, delta);
1945 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
1946 __this_cpu_add(cpu_softirq_time, delta);
1948 irq_time_write_end();
1949 local_irq_restore(flags);
1951 EXPORT_SYMBOL_GPL(account_system_vtime);
1953 static void update_rq_clock_task(struct rq *rq, s64 delta)
1957 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
1960 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1961 * this case when a previous update_rq_clock() happened inside a
1962 * {soft,}irq region.
1964 * When this happens, we stop ->clock_task and only update the
1965 * prev_irq_time stamp to account for the part that fit, so that a next
1966 * update will consume the rest. This ensures ->clock_task is
1969 * It does however cause some slight miss-attribution of {soft,}irq
1970 * time, a more accurate solution would be to update the irq_time using
1971 * the current rq->clock timestamp, except that would require using
1974 if (irq_delta > delta)
1977 rq->prev_irq_time += irq_delta;
1979 rq->clock_task += delta;
1981 if (irq_delta && sched_feat(NONIRQ_POWER))
1982 sched_rt_avg_update(rq, irq_delta);
1985 static int irqtime_account_hi_update(void)
1987 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
1988 unsigned long flags;
1992 local_irq_save(flags);
1993 latest_ns = this_cpu_read(cpu_hardirq_time);
1994 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
1996 local_irq_restore(flags);
2000 static int irqtime_account_si_update(void)
2002 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2003 unsigned long flags;
2007 local_irq_save(flags);
2008 latest_ns = this_cpu_read(cpu_softirq_time);
2009 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2011 local_irq_restore(flags);
2015 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2017 #define sched_clock_irqtime (0)
2019 static void update_rq_clock_task(struct rq *rq, s64 delta)
2021 rq->clock_task += delta;
2024 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2026 #include "sched_idletask.c"
2027 #include "sched_fair.c"
2028 #include "sched_rt.c"
2029 #include "sched_autogroup.c"
2030 #include "sched_stoptask.c"
2031 #ifdef CONFIG_SCHED_DEBUG
2032 # include "sched_debug.c"
2035 void sched_set_stop_task(int cpu, struct task_struct *stop)
2037 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2038 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2042 * Make it appear like a SCHED_FIFO task, its something
2043 * userspace knows about and won't get confused about.
2045 * Also, it will make PI more or less work without too
2046 * much confusion -- but then, stop work should not
2047 * rely on PI working anyway.
2049 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2051 stop->sched_class = &stop_sched_class;
2054 cpu_rq(cpu)->stop = stop;
2058 * Reset it back to a normal scheduling class so that
2059 * it can die in pieces.
2061 old_stop->sched_class = &rt_sched_class;
2066 * __normal_prio - return the priority that is based on the static prio
2068 static inline int __normal_prio(struct task_struct *p)
2070 return p->static_prio;
2074 * Calculate the expected normal priority: i.e. priority
2075 * without taking RT-inheritance into account. Might be
2076 * boosted by interactivity modifiers. Changes upon fork,
2077 * setprio syscalls, and whenever the interactivity
2078 * estimator recalculates.
2080 static inline int normal_prio(struct task_struct *p)
2084 if (task_has_rt_policy(p))
2085 prio = MAX_RT_PRIO-1 - p->rt_priority;
2087 prio = __normal_prio(p);
2092 * Calculate the current priority, i.e. the priority
2093 * taken into account by the scheduler. This value might
2094 * be boosted by RT tasks, or might be boosted by
2095 * interactivity modifiers. Will be RT if the task got
2096 * RT-boosted. If not then it returns p->normal_prio.
2098 static int effective_prio(struct task_struct *p)
2100 p->normal_prio = normal_prio(p);
2102 * If we are RT tasks or we were boosted to RT priority,
2103 * keep the priority unchanged. Otherwise, update priority
2104 * to the normal priority:
2106 if (!rt_prio(p->prio))
2107 return p->normal_prio;
2112 * task_curr - is this task currently executing on a CPU?
2113 * @p: the task in question.
2115 inline int task_curr(const struct task_struct *p)
2117 return cpu_curr(task_cpu(p)) == p;
2120 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2121 const struct sched_class *prev_class,
2124 if (prev_class != p->sched_class) {
2125 if (prev_class->switched_from)
2126 prev_class->switched_from(rq, p);
2127 p->sched_class->switched_to(rq, p);
2128 } else if (oldprio != p->prio)
2129 p->sched_class->prio_changed(rq, p, oldprio);
2132 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2134 const struct sched_class *class;
2136 if (p->sched_class == rq->curr->sched_class) {
2137 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2139 for_each_class(class) {
2140 if (class == rq->curr->sched_class)
2142 if (class == p->sched_class) {
2143 resched_task(rq->curr);
2150 * A queue event has occurred, and we're going to schedule. In
2151 * this case, we can save a useless back to back clock update.
2153 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2154 rq->skip_clock_update = 1;
2159 * Is this task likely cache-hot:
2162 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2166 if (p->sched_class != &fair_sched_class)
2169 if (unlikely(p->policy == SCHED_IDLE))
2173 * Buddy candidates are cache hot:
2175 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2176 (&p->se == cfs_rq_of(&p->se)->next ||
2177 &p->se == cfs_rq_of(&p->se)->last))
2180 if (sysctl_sched_migration_cost == -1)
2182 if (sysctl_sched_migration_cost == 0)
2185 delta = now - p->se.exec_start;
2187 return delta < (s64)sysctl_sched_migration_cost;
2190 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2192 #ifdef CONFIG_SCHED_DEBUG
2194 * We should never call set_task_cpu() on a blocked task,
2195 * ttwu() will sort out the placement.
2197 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2198 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2200 #ifdef CONFIG_LOCKDEP
2202 * The caller should hold either p->pi_lock or rq->lock, when changing
2203 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2205 * sched_move_task() holds both and thus holding either pins the cgroup,
2208 * Furthermore, all task_rq users should acquire both locks, see
2211 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2212 lockdep_is_held(&task_rq(p)->lock)));
2216 trace_sched_migrate_task(p, new_cpu);
2218 if (task_cpu(p) != new_cpu) {
2219 p->se.nr_migrations++;
2220 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2223 __set_task_cpu(p, new_cpu);
2226 struct migration_arg {
2227 struct task_struct *task;
2231 static int migration_cpu_stop(void *data);
2234 * wait_task_inactive - wait for a thread to unschedule.
2236 * If @match_state is nonzero, it's the @p->state value just checked and
2237 * not expected to change. If it changes, i.e. @p might have woken up,
2238 * then return zero. When we succeed in waiting for @p to be off its CPU,
2239 * we return a positive number (its total switch count). If a second call
2240 * a short while later returns the same number, the caller can be sure that
2241 * @p has remained unscheduled the whole time.
2243 * The caller must ensure that the task *will* unschedule sometime soon,
2244 * else this function might spin for a *long* time. This function can't
2245 * be called with interrupts off, or it may introduce deadlock with
2246 * smp_call_function() if an IPI is sent by the same process we are
2247 * waiting to become inactive.
2249 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2251 unsigned long flags;
2258 * We do the initial early heuristics without holding
2259 * any task-queue locks at all. We'll only try to get
2260 * the runqueue lock when things look like they will
2266 * If the task is actively running on another CPU
2267 * still, just relax and busy-wait without holding
2270 * NOTE! Since we don't hold any locks, it's not
2271 * even sure that "rq" stays as the right runqueue!
2272 * But we don't care, since "task_running()" will
2273 * return false if the runqueue has changed and p
2274 * is actually now running somewhere else!
2276 while (task_running(rq, p)) {
2277 if (match_state && unlikely(p->state != match_state))
2283 * Ok, time to look more closely! We need the rq
2284 * lock now, to be *sure*. If we're wrong, we'll
2285 * just go back and repeat.
2287 rq = task_rq_lock(p, &flags);
2288 trace_sched_wait_task(p);
2289 running = task_running(rq, p);
2292 if (!match_state || p->state == match_state)
2293 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2294 task_rq_unlock(rq, p, &flags);
2297 * If it changed from the expected state, bail out now.
2299 if (unlikely(!ncsw))
2303 * Was it really running after all now that we
2304 * checked with the proper locks actually held?
2306 * Oops. Go back and try again..
2308 if (unlikely(running)) {
2314 * It's not enough that it's not actively running,
2315 * it must be off the runqueue _entirely_, and not
2318 * So if it was still runnable (but just not actively
2319 * running right now), it's preempted, and we should
2320 * yield - it could be a while.
2322 if (unlikely(on_rq)) {
2323 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2325 set_current_state(TASK_UNINTERRUPTIBLE);
2326 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2331 * Ahh, all good. It wasn't running, and it wasn't
2332 * runnable, which means that it will never become
2333 * running in the future either. We're all done!
2342 * kick_process - kick a running thread to enter/exit the kernel
2343 * @p: the to-be-kicked thread
2345 * Cause a process which is running on another CPU to enter
2346 * kernel-mode, without any delay. (to get signals handled.)
2348 * NOTE: this function doesn't have to take the runqueue lock,
2349 * because all it wants to ensure is that the remote task enters
2350 * the kernel. If the IPI races and the task has been migrated
2351 * to another CPU then no harm is done and the purpose has been
2354 void kick_process(struct task_struct *p)
2360 if ((cpu != smp_processor_id()) && task_curr(p))
2361 smp_send_reschedule(cpu);
2364 EXPORT_SYMBOL_GPL(kick_process);
2365 #endif /* CONFIG_SMP */
2369 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2371 static int select_fallback_rq(int cpu, struct task_struct *p)
2374 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2376 /* Look for allowed, online CPU in same node. */
2377 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2378 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2381 /* Any allowed, online CPU? */
2382 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2383 if (dest_cpu < nr_cpu_ids)
2386 /* No more Mr. Nice Guy. */
2387 dest_cpu = cpuset_cpus_allowed_fallback(p);
2389 * Don't tell them about moving exiting tasks or
2390 * kernel threads (both mm NULL), since they never
2393 if (p->mm && printk_ratelimit()) {
2394 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2395 task_pid_nr(p), p->comm, cpu);
2402 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2405 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2407 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2410 * In order not to call set_task_cpu() on a blocking task we need
2411 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2414 * Since this is common to all placement strategies, this lives here.
2416 * [ this allows ->select_task() to simply return task_cpu(p) and
2417 * not worry about this generic constraint ]
2419 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2421 cpu = select_fallback_rq(task_cpu(p), p);
2426 static void update_avg(u64 *avg, u64 sample)
2428 s64 diff = sample - *avg;
2434 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2436 #ifdef CONFIG_SCHEDSTATS
2437 struct rq *rq = this_rq();
2440 int this_cpu = smp_processor_id();
2442 if (cpu == this_cpu) {
2443 schedstat_inc(rq, ttwu_local);
2444 schedstat_inc(p, se.statistics.nr_wakeups_local);
2446 struct sched_domain *sd;
2448 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2450 for_each_domain(this_cpu, sd) {
2451 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2452 schedstat_inc(sd, ttwu_wake_remote);
2459 if (wake_flags & WF_MIGRATED)
2460 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2462 #endif /* CONFIG_SMP */
2464 schedstat_inc(rq, ttwu_count);
2465 schedstat_inc(p, se.statistics.nr_wakeups);
2467 if (wake_flags & WF_SYNC)
2468 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2470 #endif /* CONFIG_SCHEDSTATS */
2473 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2475 activate_task(rq, p, en_flags);
2478 /* if a worker is waking up, notify workqueue */
2479 if (p->flags & PF_WQ_WORKER)
2480 wq_worker_waking_up(p, cpu_of(rq));
2484 * Mark the task runnable and perform wakeup-preemption.
2487 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2489 trace_sched_wakeup(p, true);
2490 check_preempt_curr(rq, p, wake_flags);
2492 p->state = TASK_RUNNING;
2494 if (p->sched_class->task_woken)
2495 p->sched_class->task_woken(rq, p);
2497 if (unlikely(rq->idle_stamp)) {
2498 u64 delta = rq->clock - rq->idle_stamp;
2499 u64 max = 2*sysctl_sched_migration_cost;
2504 update_avg(&rq->avg_idle, delta);
2511 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2514 if (p->sched_contributes_to_load)
2515 rq->nr_uninterruptible--;
2518 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2519 ttwu_do_wakeup(rq, p, wake_flags);
2523 * Called in case the task @p isn't fully descheduled from its runqueue,
2524 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2525 * since all we need to do is flip p->state to TASK_RUNNING, since
2526 * the task is still ->on_rq.
2528 static int ttwu_remote(struct task_struct *p, int wake_flags)
2533 rq = __task_rq_lock(p);
2535 ttwu_do_wakeup(rq, p, wake_flags);
2538 __task_rq_unlock(rq);
2544 static void sched_ttwu_do_pending(struct task_struct *list)
2546 struct rq *rq = this_rq();
2548 raw_spin_lock(&rq->lock);
2551 struct task_struct *p = list;
2552 list = list->wake_entry;
2553 ttwu_do_activate(rq, p, 0);
2556 raw_spin_unlock(&rq->lock);
2559 #ifdef CONFIG_HOTPLUG_CPU
2561 static void sched_ttwu_pending(void)
2563 struct rq *rq = this_rq();
2564 struct task_struct *list = xchg(&rq->wake_list, NULL);
2569 sched_ttwu_do_pending(list);
2572 #endif /* CONFIG_HOTPLUG_CPU */
2574 void scheduler_ipi(void)
2576 struct rq *rq = this_rq();
2577 struct task_struct *list = xchg(&rq->wake_list, NULL);
2583 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2584 * traditionally all their work was done from the interrupt return
2585 * path. Now that we actually do some work, we need to make sure
2588 * Some archs already do call them, luckily irq_enter/exit nest
2591 * Arguably we should visit all archs and update all handlers,
2592 * however a fair share of IPIs are still resched only so this would
2593 * somewhat pessimize the simple resched case.
2596 sched_ttwu_do_pending(list);
2600 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2602 struct rq *rq = cpu_rq(cpu);
2603 struct task_struct *next = rq->wake_list;
2606 struct task_struct *old = next;
2608 p->wake_entry = next;
2609 next = cmpxchg(&rq->wake_list, old, p);
2615 smp_send_reschedule(cpu);
2618 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2619 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2624 rq = __task_rq_lock(p);
2626 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2627 ttwu_do_wakeup(rq, p, wake_flags);
2630 __task_rq_unlock(rq);
2635 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2636 #endif /* CONFIG_SMP */
2638 static void ttwu_queue(struct task_struct *p, int cpu)
2640 struct rq *rq = cpu_rq(cpu);
2642 #if defined(CONFIG_SMP)
2643 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2644 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2645 ttwu_queue_remote(p, cpu);
2650 raw_spin_lock(&rq->lock);
2651 ttwu_do_activate(rq, p, 0);
2652 raw_spin_unlock(&rq->lock);
2656 * try_to_wake_up - wake up a thread
2657 * @p: the thread to be awakened
2658 * @state: the mask of task states that can be woken
2659 * @wake_flags: wake modifier flags (WF_*)
2661 * Put it on the run-queue if it's not already there. The "current"
2662 * thread is always on the run-queue (except when the actual
2663 * re-schedule is in progress), and as such you're allowed to do
2664 * the simpler "current->state = TASK_RUNNING" to mark yourself
2665 * runnable without the overhead of this.
2667 * Returns %true if @p was woken up, %false if it was already running
2668 * or @state didn't match @p's state.
2671 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2673 unsigned long flags;
2674 int cpu, success = 0;
2677 raw_spin_lock_irqsave(&p->pi_lock, flags);
2678 if (!(p->state & state))
2681 success = 1; /* we're going to change ->state */
2684 if (p->on_rq && ttwu_remote(p, wake_flags))
2689 * If the owning (remote) cpu is still in the middle of schedule() with
2690 * this task as prev, wait until its done referencing the task.
2693 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2695 * In case the architecture enables interrupts in
2696 * context_switch(), we cannot busy wait, since that
2697 * would lead to deadlocks when an interrupt hits and
2698 * tries to wake up @prev. So bail and do a complete
2701 if (ttwu_activate_remote(p, wake_flags))
2708 * Pairs with the smp_wmb() in finish_lock_switch().
2712 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2713 p->state = TASK_WAKING;
2715 if (p->sched_class->task_waking)
2716 p->sched_class->task_waking(p);
2718 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2719 if (task_cpu(p) != cpu) {
2720 wake_flags |= WF_MIGRATED;
2721 set_task_cpu(p, cpu);
2723 #endif /* CONFIG_SMP */
2727 ttwu_stat(p, cpu, wake_flags);
2729 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2735 * try_to_wake_up_local - try to wake up a local task with rq lock held
2736 * @p: the thread to be awakened
2738 * Put @p on the run-queue if it's not already there. The caller must
2739 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2742 static void try_to_wake_up_local(struct task_struct *p)
2744 struct rq *rq = task_rq(p);
2746 BUG_ON(rq != this_rq());
2747 BUG_ON(p == current);
2748 lockdep_assert_held(&rq->lock);
2750 if (!raw_spin_trylock(&p->pi_lock)) {
2751 raw_spin_unlock(&rq->lock);
2752 raw_spin_lock(&p->pi_lock);
2753 raw_spin_lock(&rq->lock);
2756 if (!(p->state & TASK_NORMAL))
2760 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2762 ttwu_do_wakeup(rq, p, 0);
2763 ttwu_stat(p, smp_processor_id(), 0);
2765 raw_spin_unlock(&p->pi_lock);
2769 * wake_up_process - Wake up a specific process
2770 * @p: The process to be woken up.
2772 * Attempt to wake up the nominated process and move it to the set of runnable
2773 * processes. Returns 1 if the process was woken up, 0 if it was already
2776 * It may be assumed that this function implies a write memory barrier before
2777 * changing the task state if and only if any tasks are woken up.
2779 int wake_up_process(struct task_struct *p)
2781 return try_to_wake_up(p, TASK_ALL, 0);
2783 EXPORT_SYMBOL(wake_up_process);
2785 int wake_up_state(struct task_struct *p, unsigned int state)
2787 return try_to_wake_up(p, state, 0);
2791 * Perform scheduler related setup for a newly forked process p.
2792 * p is forked by current.
2794 * __sched_fork() is basic setup used by init_idle() too:
2796 static void __sched_fork(struct task_struct *p)
2801 p->se.exec_start = 0;
2802 p->se.sum_exec_runtime = 0;
2803 p->se.prev_sum_exec_runtime = 0;
2804 p->se.nr_migrations = 0;
2806 INIT_LIST_HEAD(&p->se.group_node);
2808 #ifdef CONFIG_SCHEDSTATS
2809 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2812 INIT_LIST_HEAD(&p->rt.run_list);
2814 #ifdef CONFIG_PREEMPT_NOTIFIERS
2815 INIT_HLIST_HEAD(&p->preempt_notifiers);
2820 * fork()/clone()-time setup:
2822 void sched_fork(struct task_struct *p)
2824 unsigned long flags;
2825 int cpu = get_cpu();
2829 * We mark the process as running here. This guarantees that
2830 * nobody will actually run it, and a signal or other external
2831 * event cannot wake it up and insert it on the runqueue either.
2833 p->state = TASK_RUNNING;
2836 * Revert to default priority/policy on fork if requested.
2838 if (unlikely(p->sched_reset_on_fork)) {
2839 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2840 p->policy = SCHED_NORMAL;
2841 p->normal_prio = p->static_prio;
2844 if (PRIO_TO_NICE(p->static_prio) < 0) {
2845 p->static_prio = NICE_TO_PRIO(0);
2846 p->normal_prio = p->static_prio;
2851 * We don't need the reset flag anymore after the fork. It has
2852 * fulfilled its duty:
2854 p->sched_reset_on_fork = 0;
2858 * Make sure we do not leak PI boosting priority to the child.
2860 p->prio = current->normal_prio;
2862 if (!rt_prio(p->prio))
2863 p->sched_class = &fair_sched_class;
2865 if (p->sched_class->task_fork)
2866 p->sched_class->task_fork(p);
2869 * The child is not yet in the pid-hash so no cgroup attach races,
2870 * and the cgroup is pinned to this child due to cgroup_fork()
2871 * is ran before sched_fork().
2873 * Silence PROVE_RCU.
2875 raw_spin_lock_irqsave(&p->pi_lock, flags);
2876 set_task_cpu(p, cpu);
2877 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2879 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2880 if (likely(sched_info_on()))
2881 memset(&p->sched_info, 0, sizeof(p->sched_info));
2883 #if defined(CONFIG_SMP)
2886 #ifdef CONFIG_PREEMPT
2887 /* Want to start with kernel preemption disabled. */
2888 task_thread_info(p)->preempt_count = 1;
2891 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2898 * wake_up_new_task - wake up a newly created task for the first time.
2900 * This function will do some initial scheduler statistics housekeeping
2901 * that must be done for every newly created context, then puts the task
2902 * on the runqueue and wakes it.
2904 void wake_up_new_task(struct task_struct *p)
2906 unsigned long flags;
2909 raw_spin_lock_irqsave(&p->pi_lock, flags);
2912 * Fork balancing, do it here and not earlier because:
2913 * - cpus_allowed can change in the fork path
2914 * - any previously selected cpu might disappear through hotplug
2916 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
2919 rq = __task_rq_lock(p);
2920 activate_task(rq, p, 0);
2922 trace_sched_wakeup_new(p, true);
2923 check_preempt_curr(rq, p, WF_FORK);
2925 if (p->sched_class->task_woken)
2926 p->sched_class->task_woken(rq, p);
2928 task_rq_unlock(rq, p, &flags);
2931 #ifdef CONFIG_PREEMPT_NOTIFIERS
2934 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2935 * @notifier: notifier struct to register
2937 void preempt_notifier_register(struct preempt_notifier *notifier)
2939 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2941 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2944 * preempt_notifier_unregister - no longer interested in preemption notifications
2945 * @notifier: notifier struct to unregister
2947 * This is safe to call from within a preemption notifier.
2949 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2951 hlist_del(¬ifier->link);
2953 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2955 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2957 struct preempt_notifier *notifier;
2958 struct hlist_node *node;
2960 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2961 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2965 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2966 struct task_struct *next)
2968 struct preempt_notifier *notifier;
2969 struct hlist_node *node;
2971 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2972 notifier->ops->sched_out(notifier, next);
2975 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2977 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2982 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2983 struct task_struct *next)
2987 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2990 * prepare_task_switch - prepare to switch tasks
2991 * @rq: the runqueue preparing to switch
2992 * @prev: the current task that is being switched out
2993 * @next: the task we are going to switch to.
2995 * This is called with the rq lock held and interrupts off. It must
2996 * be paired with a subsequent finish_task_switch after the context
2999 * prepare_task_switch sets up locking and calls architecture specific
3003 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3004 struct task_struct *next)
3006 sched_info_switch(prev, next);
3007 perf_event_task_sched_out(prev, next);
3008 fire_sched_out_preempt_notifiers(prev, next);
3009 prepare_lock_switch(rq, next);
3010 prepare_arch_switch(next);
3011 trace_sched_switch(prev, next);
3015 * finish_task_switch - clean up after a task-switch
3016 * @rq: runqueue associated with task-switch
3017 * @prev: the thread we just switched away from.
3019 * finish_task_switch must be called after the context switch, paired
3020 * with a prepare_task_switch call before the context switch.
3021 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3022 * and do any other architecture-specific cleanup actions.
3024 * Note that we may have delayed dropping an mm in context_switch(). If
3025 * so, we finish that here outside of the runqueue lock. (Doing it
3026 * with the lock held can cause deadlocks; see schedule() for
3029 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3030 __releases(rq->lock)
3032 struct mm_struct *mm = rq->prev_mm;
3038 * A task struct has one reference for the use as "current".
3039 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3040 * schedule one last time. The schedule call will never return, and
3041 * the scheduled task must drop that reference.
3042 * The test for TASK_DEAD must occur while the runqueue locks are
3043 * still held, otherwise prev could be scheduled on another cpu, die
3044 * there before we look at prev->state, and then the reference would
3046 * Manfred Spraul <manfred@colorfullife.com>
3048 prev_state = prev->state;
3049 finish_arch_switch(prev);
3050 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3051 local_irq_disable();
3052 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3053 perf_event_task_sched_in(current);
3054 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3056 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3057 finish_lock_switch(rq, prev);
3059 fire_sched_in_preempt_notifiers(current);
3062 if (unlikely(prev_state == TASK_DEAD)) {
3064 * Remove function-return probe instances associated with this
3065 * task and put them back on the free list.
3067 kprobe_flush_task(prev);
3068 put_task_struct(prev);
3074 /* assumes rq->lock is held */
3075 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3077 if (prev->sched_class->pre_schedule)
3078 prev->sched_class->pre_schedule(rq, prev);
3081 /* rq->lock is NOT held, but preemption is disabled */
3082 static inline void post_schedule(struct rq *rq)
3084 if (rq->post_schedule) {
3085 unsigned long flags;
3087 raw_spin_lock_irqsave(&rq->lock, flags);
3088 if (rq->curr->sched_class->post_schedule)
3089 rq->curr->sched_class->post_schedule(rq);
3090 raw_spin_unlock_irqrestore(&rq->lock, flags);
3092 rq->post_schedule = 0;
3098 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3102 static inline void post_schedule(struct rq *rq)
3109 * schedule_tail - first thing a freshly forked thread must call.
3110 * @prev: the thread we just switched away from.
3112 asmlinkage void schedule_tail(struct task_struct *prev)
3113 __releases(rq->lock)
3115 struct rq *rq = this_rq();
3117 finish_task_switch(rq, prev);
3120 * FIXME: do we need to worry about rq being invalidated by the
3125 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3126 /* In this case, finish_task_switch does not reenable preemption */
3129 if (current->set_child_tid)
3130 put_user(task_pid_vnr(current), current->set_child_tid);
3134 * context_switch - switch to the new MM and the new
3135 * thread's register state.
3138 context_switch(struct rq *rq, struct task_struct *prev,
3139 struct task_struct *next)
3141 struct mm_struct *mm, *oldmm;
3143 prepare_task_switch(rq, prev, next);
3146 oldmm = prev->active_mm;
3148 * For paravirt, this is coupled with an exit in switch_to to
3149 * combine the page table reload and the switch backend into
3152 arch_start_context_switch(prev);
3155 next->active_mm = oldmm;
3156 atomic_inc(&oldmm->mm_count);
3157 enter_lazy_tlb(oldmm, next);
3159 switch_mm(oldmm, mm, next);
3162 prev->active_mm = NULL;
3163 rq->prev_mm = oldmm;
3166 * Since the runqueue lock will be released by the next
3167 * task (which is an invalid locking op but in the case
3168 * of the scheduler it's an obvious special-case), so we
3169 * do an early lockdep release here:
3171 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3172 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3175 /* Here we just switch the register state and the stack. */
3176 switch_to(prev, next, prev);
3180 * this_rq must be evaluated again because prev may have moved
3181 * CPUs since it called schedule(), thus the 'rq' on its stack
3182 * frame will be invalid.
3184 finish_task_switch(this_rq(), prev);
3188 * nr_running, nr_uninterruptible and nr_context_switches:
3190 * externally visible scheduler statistics: current number of runnable
3191 * threads, current number of uninterruptible-sleeping threads, total
3192 * number of context switches performed since bootup.
3194 unsigned long nr_running(void)
3196 unsigned long i, sum = 0;
3198 for_each_online_cpu(i)
3199 sum += cpu_rq(i)->nr_running;
3204 unsigned long nr_uninterruptible(void)
3206 unsigned long i, sum = 0;
3208 for_each_possible_cpu(i)
3209 sum += cpu_rq(i)->nr_uninterruptible;
3212 * Since we read the counters lockless, it might be slightly
3213 * inaccurate. Do not allow it to go below zero though:
3215 if (unlikely((long)sum < 0))
3221 unsigned long long nr_context_switches(void)
3224 unsigned long long sum = 0;
3226 for_each_possible_cpu(i)
3227 sum += cpu_rq(i)->nr_switches;
3232 unsigned long nr_iowait(void)
3234 unsigned long i, sum = 0;
3236 for_each_possible_cpu(i)
3237 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3242 unsigned long nr_iowait_cpu(int cpu)
3244 struct rq *this = cpu_rq(cpu);
3245 return atomic_read(&this->nr_iowait);
3248 unsigned long this_cpu_load(void)
3250 struct rq *this = this_rq();
3251 return this->cpu_load[0];
3255 /* Variables and functions for calc_load */
3256 static atomic_long_t calc_load_tasks;
3257 static unsigned long calc_load_update;
3258 unsigned long avenrun[3];
3259 EXPORT_SYMBOL(avenrun);
3261 static long calc_load_fold_active(struct rq *this_rq)
3263 long nr_active, delta = 0;
3265 nr_active = this_rq->nr_running;
3266 nr_active += (long) this_rq->nr_uninterruptible;
3268 if (nr_active != this_rq->calc_load_active) {
3269 delta = nr_active - this_rq->calc_load_active;
3270 this_rq->calc_load_active = nr_active;
3276 static unsigned long
3277 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3280 load += active * (FIXED_1 - exp);
3281 load += 1UL << (FSHIFT - 1);
3282 return load >> FSHIFT;
3287 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3289 * When making the ILB scale, we should try to pull this in as well.
3291 static atomic_long_t calc_load_tasks_idle;
3293 static void calc_load_account_idle(struct rq *this_rq)
3297 delta = calc_load_fold_active(this_rq);
3299 atomic_long_add(delta, &calc_load_tasks_idle);
3302 static long calc_load_fold_idle(void)
3307 * Its got a race, we don't care...
3309 if (atomic_long_read(&calc_load_tasks_idle))
3310 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3316 * fixed_power_int - compute: x^n, in O(log n) time
3318 * @x: base of the power
3319 * @frac_bits: fractional bits of @x
3320 * @n: power to raise @x to.
3322 * By exploiting the relation between the definition of the natural power
3323 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3324 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3325 * (where: n_i \elem {0, 1}, the binary vector representing n),
3326 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3327 * of course trivially computable in O(log_2 n), the length of our binary
3330 static unsigned long
3331 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3333 unsigned long result = 1UL << frac_bits;
3338 result += 1UL << (frac_bits - 1);
3339 result >>= frac_bits;
3345 x += 1UL << (frac_bits - 1);
3353 * a1 = a0 * e + a * (1 - e)
3355 * a2 = a1 * e + a * (1 - e)
3356 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3357 * = a0 * e^2 + a * (1 - e) * (1 + e)
3359 * a3 = a2 * e + a * (1 - e)
3360 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3361 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3365 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3366 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3367 * = a0 * e^n + a * (1 - e^n)
3369 * [1] application of the geometric series:
3372 * S_n := \Sum x^i = -------------
3375 static unsigned long
3376 calc_load_n(unsigned long load, unsigned long exp,
3377 unsigned long active, unsigned int n)
3380 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3384 * NO_HZ can leave us missing all per-cpu ticks calling
3385 * calc_load_account_active(), but since an idle CPU folds its delta into
3386 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3387 * in the pending idle delta if our idle period crossed a load cycle boundary.
3389 * Once we've updated the global active value, we need to apply the exponential
3390 * weights adjusted to the number of cycles missed.
3392 static void calc_global_nohz(void)
3394 long delta, active, n;
3397 * If we crossed a calc_load_update boundary, make sure to fold
3398 * any pending idle changes, the respective CPUs might have
3399 * missed the tick driven calc_load_account_active() update
3402 delta = calc_load_fold_idle();
3404 atomic_long_add(delta, &calc_load_tasks);
3407 * It could be the one fold was all it took, we done!
3409 if (time_before(jiffies, calc_load_update + 10))
3413 * Catch-up, fold however many we are behind still
3415 delta = jiffies - calc_load_update - 10;
3416 n = 1 + (delta / LOAD_FREQ);
3418 active = atomic_long_read(&calc_load_tasks);
3419 active = active > 0 ? active * FIXED_1 : 0;
3421 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3422 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3423 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3425 calc_load_update += n * LOAD_FREQ;
3428 static void calc_load_account_idle(struct rq *this_rq)
3432 static inline long calc_load_fold_idle(void)
3437 static void calc_global_nohz(void)
3443 * get_avenrun - get the load average array
3444 * @loads: pointer to dest load array
3445 * @offset: offset to add
3446 * @shift: shift count to shift the result left
3448 * These values are estimates at best, so no need for locking.
3450 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3452 loads[0] = (avenrun[0] + offset) << shift;
3453 loads[1] = (avenrun[1] + offset) << shift;
3454 loads[2] = (avenrun[2] + offset) << shift;
3458 * calc_load - update the avenrun load estimates 10 ticks after the
3459 * CPUs have updated calc_load_tasks.
3461 void calc_global_load(unsigned long ticks)
3465 if (time_before(jiffies, calc_load_update + 10))
3468 active = atomic_long_read(&calc_load_tasks);
3469 active = active > 0 ? active * FIXED_1 : 0;
3471 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3472 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3473 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3475 calc_load_update += LOAD_FREQ;
3478 * Account one period with whatever state we found before
3479 * folding in the nohz state and ageing the entire idle period.
3481 * This avoids loosing a sample when we go idle between
3482 * calc_load_account_active() (10 ticks ago) and now and thus
3489 * Called from update_cpu_load() to periodically update this CPU's
3492 static void calc_load_account_active(struct rq *this_rq)
3496 if (time_before(jiffies, this_rq->calc_load_update))
3499 delta = calc_load_fold_active(this_rq);
3500 delta += calc_load_fold_idle();
3502 atomic_long_add(delta, &calc_load_tasks);
3504 this_rq->calc_load_update += LOAD_FREQ;
3508 * The exact cpuload at various idx values, calculated at every tick would be
3509 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3511 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3512 * on nth tick when cpu may be busy, then we have:
3513 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3514 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3516 * decay_load_missed() below does efficient calculation of
3517 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3518 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3520 * The calculation is approximated on a 128 point scale.
3521 * degrade_zero_ticks is the number of ticks after which load at any
3522 * particular idx is approximated to be zero.
3523 * degrade_factor is a precomputed table, a row for each load idx.
3524 * Each column corresponds to degradation factor for a power of two ticks,
3525 * based on 128 point scale.
3527 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3528 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3530 * With this power of 2 load factors, we can degrade the load n times
3531 * by looking at 1 bits in n and doing as many mult/shift instead of
3532 * n mult/shifts needed by the exact degradation.
3534 #define DEGRADE_SHIFT 7
3535 static const unsigned char
3536 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3537 static const unsigned char
3538 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3539 {0, 0, 0, 0, 0, 0, 0, 0},
3540 {64, 32, 8, 0, 0, 0, 0, 0},
3541 {96, 72, 40, 12, 1, 0, 0},
3542 {112, 98, 75, 43, 15, 1, 0},
3543 {120, 112, 98, 76, 45, 16, 2} };
3546 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3547 * would be when CPU is idle and so we just decay the old load without
3548 * adding any new load.
3550 static unsigned long
3551 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3555 if (!missed_updates)
3558 if (missed_updates >= degrade_zero_ticks[idx])
3562 return load >> missed_updates;
3564 while (missed_updates) {
3565 if (missed_updates % 2)
3566 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3568 missed_updates >>= 1;
3575 * Update rq->cpu_load[] statistics. This function is usually called every
3576 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3577 * every tick. We fix it up based on jiffies.
3579 static void update_cpu_load(struct rq *this_rq)
3581 unsigned long this_load = this_rq->load.weight;
3582 unsigned long curr_jiffies = jiffies;
3583 unsigned long pending_updates;
3586 this_rq->nr_load_updates++;
3588 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3589 if (curr_jiffies == this_rq->last_load_update_tick)
3592 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3593 this_rq->last_load_update_tick = curr_jiffies;
3595 /* Update our load: */
3596 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3597 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3598 unsigned long old_load, new_load;
3600 /* scale is effectively 1 << i now, and >> i divides by scale */
3602 old_load = this_rq->cpu_load[i];
3603 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3604 new_load = this_load;
3606 * Round up the averaging division if load is increasing. This
3607 * prevents us from getting stuck on 9 if the load is 10, for
3610 if (new_load > old_load)
3611 new_load += scale - 1;
3613 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3616 sched_avg_update(this_rq);
3619 static void update_cpu_load_active(struct rq *this_rq)
3621 update_cpu_load(this_rq);
3623 calc_load_account_active(this_rq);
3629 * sched_exec - execve() is a valuable balancing opportunity, because at
3630 * this point the task has the smallest effective memory and cache footprint.
3632 void sched_exec(void)
3634 struct task_struct *p = current;
3635 unsigned long flags;
3638 raw_spin_lock_irqsave(&p->pi_lock, flags);
3639 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3640 if (dest_cpu == smp_processor_id())
3643 if (likely(cpu_active(dest_cpu))) {
3644 struct migration_arg arg = { p, dest_cpu };
3646 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3647 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3651 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3656 DEFINE_PER_CPU(struct kernel_stat, kstat);
3658 EXPORT_PER_CPU_SYMBOL(kstat);
3661 * Return any ns on the sched_clock that have not yet been accounted in
3662 * @p in case that task is currently running.
3664 * Called with task_rq_lock() held on @rq.
3666 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3670 if (task_current(rq, p)) {
3671 update_rq_clock(rq);
3672 ns = rq->clock_task - p->se.exec_start;
3680 unsigned long long task_delta_exec(struct task_struct *p)
3682 unsigned long flags;
3686 rq = task_rq_lock(p, &flags);
3687 ns = do_task_delta_exec(p, rq);
3688 task_rq_unlock(rq, p, &flags);
3694 * Return accounted runtime for the task.
3695 * In case the task is currently running, return the runtime plus current's
3696 * pending runtime that have not been accounted yet.
3698 unsigned long long task_sched_runtime(struct task_struct *p)
3700 unsigned long flags;
3704 rq = task_rq_lock(p, &flags);
3705 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3706 task_rq_unlock(rq, p, &flags);
3712 * Account user cpu time to a process.
3713 * @p: the process that the cpu time gets accounted to
3714 * @cputime: the cpu time spent in user space since the last update
3715 * @cputime_scaled: cputime scaled by cpu frequency
3717 void account_user_time(struct task_struct *p, cputime_t cputime,
3718 cputime_t cputime_scaled)
3720 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3723 /* Add user time to process. */
3724 p->utime = cputime_add(p->utime, cputime);
3725 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3726 account_group_user_time(p, cputime);
3728 /* Add user time to cpustat. */
3729 tmp = cputime_to_cputime64(cputime);
3730 if (TASK_NICE(p) > 0)
3731 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3733 cpustat->user = cputime64_add(cpustat->user, tmp);
3735 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3736 /* Account for user time used */
3737 acct_update_integrals(p);
3741 * Account guest cpu time to a process.
3742 * @p: the process that the cpu time gets accounted to
3743 * @cputime: the cpu time spent in virtual machine since the last update
3744 * @cputime_scaled: cputime scaled by cpu frequency
3746 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3747 cputime_t cputime_scaled)
3750 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3752 tmp = cputime_to_cputime64(cputime);
3754 /* Add guest time to process. */
3755 p->utime = cputime_add(p->utime, cputime);
3756 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3757 account_group_user_time(p, cputime);
3758 p->gtime = cputime_add(p->gtime, cputime);
3760 /* Add guest time to cpustat. */
3761 if (TASK_NICE(p) > 0) {
3762 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3763 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3765 cpustat->user = cputime64_add(cpustat->user, tmp);
3766 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3771 * Account system cpu time to a process and desired cpustat field
3772 * @p: the process that the cpu time gets accounted to
3773 * @cputime: the cpu time spent in kernel space since the last update
3774 * @cputime_scaled: cputime scaled by cpu frequency
3775 * @target_cputime64: pointer to cpustat field that has to be updated
3778 void __account_system_time(struct task_struct *p, cputime_t cputime,
3779 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3781 cputime64_t tmp = cputime_to_cputime64(cputime);
3783 /* Add system time to process. */
3784 p->stime = cputime_add(p->stime, cputime);
3785 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3786 account_group_system_time(p, cputime);
3788 /* Add system time to cpustat. */
3789 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3790 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3792 /* Account for system time used */
3793 acct_update_integrals(p);
3797 * Account system cpu time to a process.
3798 * @p: the process that the cpu time gets accounted to
3799 * @hardirq_offset: the offset to subtract from hardirq_count()
3800 * @cputime: the cpu time spent in kernel space since the last update
3801 * @cputime_scaled: cputime scaled by cpu frequency
3803 void account_system_time(struct task_struct *p, int hardirq_offset,
3804 cputime_t cputime, cputime_t cputime_scaled)
3806 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3807 cputime64_t *target_cputime64;
3809 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3810 account_guest_time(p, cputime, cputime_scaled);
3814 if (hardirq_count() - hardirq_offset)
3815 target_cputime64 = &cpustat->irq;
3816 else if (in_serving_softirq())
3817 target_cputime64 = &cpustat->softirq;
3819 target_cputime64 = &cpustat->system;
3821 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3825 * Account for involuntary wait time.
3826 * @cputime: the cpu time spent in involuntary wait
3828 void account_steal_time(cputime_t cputime)
3830 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3831 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3833 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3837 * Account for idle time.
3838 * @cputime: the cpu time spent in idle wait
3840 void account_idle_time(cputime_t cputime)
3842 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3843 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3844 struct rq *rq = this_rq();
3846 if (atomic_read(&rq->nr_iowait) > 0)
3847 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3849 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3852 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3854 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3856 * Account a tick to a process and cpustat
3857 * @p: the process that the cpu time gets accounted to
3858 * @user_tick: is the tick from userspace
3859 * @rq: the pointer to rq
3861 * Tick demultiplexing follows the order
3862 * - pending hardirq update
3863 * - pending softirq update
3867 * - check for guest_time
3868 * - else account as system_time
3870 * Check for hardirq is done both for system and user time as there is
3871 * no timer going off while we are on hardirq and hence we may never get an
3872 * opportunity to update it solely in system time.
3873 * p->stime and friends are only updated on system time and not on irq
3874 * softirq as those do not count in task exec_runtime any more.
3876 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3879 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3880 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
3881 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3883 if (irqtime_account_hi_update()) {
3884 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3885 } else if (irqtime_account_si_update()) {
3886 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3887 } else if (this_cpu_ksoftirqd() == p) {
3889 * ksoftirqd time do not get accounted in cpu_softirq_time.
3890 * So, we have to handle it separately here.
3891 * Also, p->stime needs to be updated for ksoftirqd.
3893 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3895 } else if (user_tick) {
3896 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3897 } else if (p == rq->idle) {
3898 account_idle_time(cputime_one_jiffy);
3899 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3900 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3902 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3907 static void irqtime_account_idle_ticks(int ticks)
3910 struct rq *rq = this_rq();
3912 for (i = 0; i < ticks; i++)
3913 irqtime_account_process_tick(current, 0, rq);
3915 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
3916 static void irqtime_account_idle_ticks(int ticks) {}
3917 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3919 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3922 * Account a single tick of cpu time.
3923 * @p: the process that the cpu time gets accounted to
3924 * @user_tick: indicates if the tick is a user or a system tick
3926 void account_process_tick(struct task_struct *p, int user_tick)
3928 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3929 struct rq *rq = this_rq();
3931 if (sched_clock_irqtime) {
3932 irqtime_account_process_tick(p, user_tick, rq);
3937 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3938 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3939 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3942 account_idle_time(cputime_one_jiffy);
3946 * Account multiple ticks of steal time.
3947 * @p: the process from which the cpu time has been stolen
3948 * @ticks: number of stolen ticks
3950 void account_steal_ticks(unsigned long ticks)
3952 account_steal_time(jiffies_to_cputime(ticks));
3956 * Account multiple ticks of idle time.
3957 * @ticks: number of stolen ticks
3959 void account_idle_ticks(unsigned long ticks)
3962 if (sched_clock_irqtime) {
3963 irqtime_account_idle_ticks(ticks);
3967 account_idle_time(jiffies_to_cputime(ticks));
3973 * Use precise platform statistics if available:
3975 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3976 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3982 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3984 struct task_cputime cputime;
3986 thread_group_cputime(p, &cputime);
3988 *ut = cputime.utime;
3989 *st = cputime.stime;
3993 #ifndef nsecs_to_cputime
3994 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3997 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3999 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
4002 * Use CFS's precise accounting:
4004 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
4010 do_div(temp, total);
4011 utime = (cputime_t)temp;
4016 * Compare with previous values, to keep monotonicity:
4018 p->prev_utime = max(p->prev_utime, utime);
4019 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4021 *ut = p->prev_utime;
4022 *st = p->prev_stime;
4026 * Must be called with siglock held.
4028 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4030 struct signal_struct *sig = p->signal;
4031 struct task_cputime cputime;
4032 cputime_t rtime, utime, total;
4034 thread_group_cputime(p, &cputime);
4036 total = cputime_add(cputime.utime, cputime.stime);
4037 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4042 temp *= cputime.utime;
4043 do_div(temp, total);
4044 utime = (cputime_t)temp;
4048 sig->prev_utime = max(sig->prev_utime, utime);
4049 sig->prev_stime = max(sig->prev_stime,
4050 cputime_sub(rtime, sig->prev_utime));
4052 *ut = sig->prev_utime;
4053 *st = sig->prev_stime;
4058 * This function gets called by the timer code, with HZ frequency.
4059 * We call it with interrupts disabled.
4061 void scheduler_tick(void)
4063 int cpu = smp_processor_id();
4064 struct rq *rq = cpu_rq(cpu);
4065 struct task_struct *curr = rq->curr;
4069 raw_spin_lock(&rq->lock);
4070 update_rq_clock(rq);
4071 update_cpu_load_active(rq);
4072 curr->sched_class->task_tick(rq, curr, 0);
4073 raw_spin_unlock(&rq->lock);
4075 perf_event_task_tick();
4078 rq->idle_at_tick = idle_cpu(cpu);
4079 trigger_load_balance(rq, cpu);
4083 notrace unsigned long get_parent_ip(unsigned long addr)
4085 if (in_lock_functions(addr)) {
4086 addr = CALLER_ADDR2;
4087 if (in_lock_functions(addr))
4088 addr = CALLER_ADDR3;
4093 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4094 defined(CONFIG_PREEMPT_TRACER))
4096 void __kprobes add_preempt_count(int val)
4098 #ifdef CONFIG_DEBUG_PREEMPT
4102 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4105 preempt_count() += val;
4106 #ifdef CONFIG_DEBUG_PREEMPT
4108 * Spinlock count overflowing soon?
4110 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4113 if (preempt_count() == val)
4114 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4116 EXPORT_SYMBOL(add_preempt_count);
4118 void __kprobes sub_preempt_count(int val)
4120 #ifdef CONFIG_DEBUG_PREEMPT
4124 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4127 * Is the spinlock portion underflowing?
4129 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4130 !(preempt_count() & PREEMPT_MASK)))
4134 if (preempt_count() == val)
4135 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4136 preempt_count() -= val;
4138 EXPORT_SYMBOL(sub_preempt_count);
4143 * Print scheduling while atomic bug:
4145 static noinline void __schedule_bug(struct task_struct *prev)
4147 struct pt_regs *regs = get_irq_regs();
4149 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4150 prev->comm, prev->pid, preempt_count());
4152 debug_show_held_locks(prev);
4154 if (irqs_disabled())
4155 print_irqtrace_events(prev);
4164 * Various schedule()-time debugging checks and statistics:
4166 static inline void schedule_debug(struct task_struct *prev)
4169 * Test if we are atomic. Since do_exit() needs to call into
4170 * schedule() atomically, we ignore that path for now.
4171 * Otherwise, whine if we are scheduling when we should not be.
4173 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4174 __schedule_bug(prev);
4176 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4178 schedstat_inc(this_rq(), sched_count);
4181 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4183 if (prev->on_rq || rq->skip_clock_update < 0)
4184 update_rq_clock(rq);
4185 prev->sched_class->put_prev_task(rq, prev);
4189 * Pick up the highest-prio task:
4191 static inline struct task_struct *
4192 pick_next_task(struct rq *rq)
4194 const struct sched_class *class;
4195 struct task_struct *p;
4198 * Optimization: we know that if all tasks are in
4199 * the fair class we can call that function directly:
4201 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4202 p = fair_sched_class.pick_next_task(rq);
4207 for_each_class(class) {
4208 p = class->pick_next_task(rq);
4213 BUG(); /* the idle class will always have a runnable task */
4217 * __schedule() is the main scheduler function.
4219 static void __sched __schedule(void)
4221 struct task_struct *prev, *next;
4222 unsigned long *switch_count;
4228 cpu = smp_processor_id();
4230 rcu_note_context_switch(cpu);
4233 schedule_debug(prev);
4235 if (sched_feat(HRTICK))
4238 raw_spin_lock_irq(&rq->lock);
4240 switch_count = &prev->nivcsw;
4241 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4242 if (unlikely(signal_pending_state(prev->state, prev))) {
4243 prev->state = TASK_RUNNING;
4245 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4249 * If a worker went to sleep, notify and ask workqueue
4250 * whether it wants to wake up a task to maintain
4253 if (prev->flags & PF_WQ_WORKER) {
4254 struct task_struct *to_wakeup;
4256 to_wakeup = wq_worker_sleeping(prev, cpu);
4258 try_to_wake_up_local(to_wakeup);
4261 switch_count = &prev->nvcsw;
4264 pre_schedule(rq, prev);
4266 if (unlikely(!rq->nr_running))
4267 idle_balance(cpu, rq);
4269 put_prev_task(rq, prev);
4270 next = pick_next_task(rq);
4271 clear_tsk_need_resched(prev);
4272 rq->skip_clock_update = 0;
4274 if (likely(prev != next)) {
4279 context_switch(rq, prev, next); /* unlocks the rq */
4281 * The context switch have flipped the stack from under us
4282 * and restored the local variables which were saved when
4283 * this task called schedule() in the past. prev == current
4284 * is still correct, but it can be moved to another cpu/rq.
4286 cpu = smp_processor_id();
4289 raw_spin_unlock_irq(&rq->lock);
4293 preempt_enable_no_resched();
4298 static inline void sched_submit_work(struct task_struct *tsk)
4303 * If we are going to sleep and we have plugged IO queued,
4304 * make sure to submit it to avoid deadlocks.
4306 if (blk_needs_flush_plug(tsk))
4307 blk_schedule_flush_plug(tsk);
4310 asmlinkage void __sched schedule(void)
4312 struct task_struct *tsk = current;
4314 sched_submit_work(tsk);
4317 EXPORT_SYMBOL(schedule);
4319 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4321 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4326 if (lock->owner != owner)
4330 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4331 * lock->owner still matches owner, if that fails, owner might
4332 * point to free()d memory, if it still matches, the rcu_read_lock()
4333 * ensures the memory stays valid.
4337 ret = owner->on_cpu;
4345 * Look out! "owner" is an entirely speculative pointer
4346 * access and not reliable.
4348 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4350 if (!sched_feat(OWNER_SPIN))
4353 while (owner_running(lock, owner)) {
4357 arch_mutex_cpu_relax();
4361 * If the owner changed to another task there is likely
4362 * heavy contention, stop spinning.
4371 #ifdef CONFIG_PREEMPT
4373 * this is the entry point to schedule() from in-kernel preemption
4374 * off of preempt_enable. Kernel preemptions off return from interrupt
4375 * occur there and call schedule directly.
4377 asmlinkage void __sched notrace preempt_schedule(void)
4379 struct thread_info *ti = current_thread_info();
4382 * If there is a non-zero preempt_count or interrupts are disabled,
4383 * we do not want to preempt the current task. Just return..
4385 if (likely(ti->preempt_count || irqs_disabled()))
4389 add_preempt_count_notrace(PREEMPT_ACTIVE);
4391 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4394 * Check again in case we missed a preemption opportunity
4395 * between schedule and now.
4398 } while (need_resched());
4400 EXPORT_SYMBOL(preempt_schedule);
4403 * this is the entry point to schedule() from kernel preemption
4404 * off of irq context.
4405 * Note, that this is called and return with irqs disabled. This will
4406 * protect us against recursive calling from irq.
4408 asmlinkage void __sched preempt_schedule_irq(void)
4410 struct thread_info *ti = current_thread_info();
4412 /* Catch callers which need to be fixed */
4413 BUG_ON(ti->preempt_count || !irqs_disabled());
4416 add_preempt_count(PREEMPT_ACTIVE);
4419 local_irq_disable();
4420 sub_preempt_count(PREEMPT_ACTIVE);
4423 * Check again in case we missed a preemption opportunity
4424 * between schedule and now.
4427 } while (need_resched());
4430 #endif /* CONFIG_PREEMPT */
4432 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4435 return try_to_wake_up(curr->private, mode, wake_flags);
4437 EXPORT_SYMBOL(default_wake_function);
4440 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4441 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4442 * number) then we wake all the non-exclusive tasks and one exclusive task.
4444 * There are circumstances in which we can try to wake a task which has already
4445 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4446 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4448 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4449 int nr_exclusive, int wake_flags, void *key)
4451 wait_queue_t *curr, *next;
4453 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4454 unsigned flags = curr->flags;
4456 if (curr->func(curr, mode, wake_flags, key) &&
4457 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4463 * __wake_up - wake up threads blocked on a waitqueue.
4465 * @mode: which threads
4466 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4467 * @key: is directly passed to the wakeup function
4469 * It may be assumed that this function implies a write memory barrier before
4470 * changing the task state if and only if any tasks are woken up.
4472 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4473 int nr_exclusive, void *key)
4475 unsigned long flags;
4477 spin_lock_irqsave(&q->lock, flags);
4478 __wake_up_common(q, mode, nr_exclusive, 0, key);
4479 spin_unlock_irqrestore(&q->lock, flags);
4481 EXPORT_SYMBOL(__wake_up);
4484 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4486 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4488 __wake_up_common(q, mode, 1, 0, NULL);
4490 EXPORT_SYMBOL_GPL(__wake_up_locked);
4492 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4494 __wake_up_common(q, mode, 1, 0, key);
4496 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4499 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4501 * @mode: which threads
4502 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4503 * @key: opaque value to be passed to wakeup targets
4505 * The sync wakeup differs that the waker knows that it will schedule
4506 * away soon, so while the target thread will be woken up, it will not
4507 * be migrated to another CPU - ie. the two threads are 'synchronized'
4508 * with each other. This can prevent needless bouncing between CPUs.
4510 * On UP it can prevent extra preemption.
4512 * It may be assumed that this function implies a write memory barrier before
4513 * changing the task state if and only if any tasks are woken up.
4515 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4516 int nr_exclusive, void *key)
4518 unsigned long flags;
4519 int wake_flags = WF_SYNC;
4524 if (unlikely(!nr_exclusive))
4527 spin_lock_irqsave(&q->lock, flags);
4528 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4529 spin_unlock_irqrestore(&q->lock, flags);
4531 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4534 * __wake_up_sync - see __wake_up_sync_key()
4536 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4538 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4540 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4543 * complete: - signals a single thread waiting on this completion
4544 * @x: holds the state of this particular completion
4546 * This will wake up a single thread waiting on this completion. Threads will be
4547 * awakened in the same order in which they were queued.
4549 * See also complete_all(), wait_for_completion() and related routines.
4551 * It may be assumed that this function implies a write memory barrier before
4552 * changing the task state if and only if any tasks are woken up.
4554 void complete(struct completion *x)
4556 unsigned long flags;
4558 spin_lock_irqsave(&x->wait.lock, flags);
4560 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4561 spin_unlock_irqrestore(&x->wait.lock, flags);
4563 EXPORT_SYMBOL(complete);
4566 * complete_all: - signals all threads waiting on this completion
4567 * @x: holds the state of this particular completion
4569 * This will wake up all threads waiting on this particular completion event.
4571 * It may be assumed that this function implies a write memory barrier before
4572 * changing the task state if and only if any tasks are woken up.
4574 void complete_all(struct completion *x)
4576 unsigned long flags;
4578 spin_lock_irqsave(&x->wait.lock, flags);
4579 x->done += UINT_MAX/2;
4580 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4581 spin_unlock_irqrestore(&x->wait.lock, flags);
4583 EXPORT_SYMBOL(complete_all);
4585 static inline long __sched
4586 do_wait_for_common(struct completion *x, long timeout, int state)
4589 DECLARE_WAITQUEUE(wait, current);
4591 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4593 if (signal_pending_state(state, current)) {
4594 timeout = -ERESTARTSYS;
4597 __set_current_state(state);
4598 spin_unlock_irq(&x->wait.lock);
4599 timeout = schedule_timeout(timeout);
4600 spin_lock_irq(&x->wait.lock);
4601 } while (!x->done && timeout);
4602 __remove_wait_queue(&x->wait, &wait);
4607 return timeout ?: 1;
4611 wait_for_common(struct completion *x, long timeout, int state)
4615 spin_lock_irq(&x->wait.lock);
4616 timeout = do_wait_for_common(x, timeout, state);
4617 spin_unlock_irq(&x->wait.lock);
4622 * wait_for_completion: - waits for completion of a task
4623 * @x: holds the state of this particular completion
4625 * This waits to be signaled for completion of a specific task. It is NOT
4626 * interruptible and there is no timeout.
4628 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4629 * and interrupt capability. Also see complete().
4631 void __sched wait_for_completion(struct completion *x)
4633 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4635 EXPORT_SYMBOL(wait_for_completion);
4638 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4639 * @x: holds the state of this particular completion
4640 * @timeout: timeout value in jiffies
4642 * This waits for either a completion of a specific task to be signaled or for a
4643 * specified timeout to expire. The timeout is in jiffies. It is not
4646 unsigned long __sched
4647 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4649 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4651 EXPORT_SYMBOL(wait_for_completion_timeout);
4654 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4655 * @x: holds the state of this particular completion
4657 * This waits for completion of a specific task to be signaled. It is
4660 int __sched wait_for_completion_interruptible(struct completion *x)
4662 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4663 if (t == -ERESTARTSYS)
4667 EXPORT_SYMBOL(wait_for_completion_interruptible);
4670 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4671 * @x: holds the state of this particular completion
4672 * @timeout: timeout value in jiffies
4674 * This waits for either a completion of a specific task to be signaled or for a
4675 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4678 wait_for_completion_interruptible_timeout(struct completion *x,
4679 unsigned long timeout)
4681 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4683 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4686 * wait_for_completion_killable: - waits for completion of a task (killable)
4687 * @x: holds the state of this particular completion
4689 * This waits to be signaled for completion of a specific task. It can be
4690 * interrupted by a kill signal.
4692 int __sched wait_for_completion_killable(struct completion *x)
4694 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4695 if (t == -ERESTARTSYS)
4699 EXPORT_SYMBOL(wait_for_completion_killable);
4702 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4703 * @x: holds the state of this particular completion
4704 * @timeout: timeout value in jiffies
4706 * This waits for either a completion of a specific task to be
4707 * signaled or for a specified timeout to expire. It can be
4708 * interrupted by a kill signal. The timeout is in jiffies.
4711 wait_for_completion_killable_timeout(struct completion *x,
4712 unsigned long timeout)
4714 return wait_for_common(x, timeout, TASK_KILLABLE);
4716 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4719 * try_wait_for_completion - try to decrement a completion without blocking
4720 * @x: completion structure
4722 * Returns: 0 if a decrement cannot be done without blocking
4723 * 1 if a decrement succeeded.
4725 * If a completion is being used as a counting completion,
4726 * attempt to decrement the counter without blocking. This
4727 * enables us to avoid waiting if the resource the completion
4728 * is protecting is not available.
4730 bool try_wait_for_completion(struct completion *x)
4732 unsigned long flags;
4735 spin_lock_irqsave(&x->wait.lock, flags);
4740 spin_unlock_irqrestore(&x->wait.lock, flags);
4743 EXPORT_SYMBOL(try_wait_for_completion);
4746 * completion_done - Test to see if a completion has any waiters
4747 * @x: completion structure
4749 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4750 * 1 if there are no waiters.
4753 bool completion_done(struct completion *x)
4755 unsigned long flags;
4758 spin_lock_irqsave(&x->wait.lock, flags);
4761 spin_unlock_irqrestore(&x->wait.lock, flags);
4764 EXPORT_SYMBOL(completion_done);
4767 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4769 unsigned long flags;
4772 init_waitqueue_entry(&wait, current);
4774 __set_current_state(state);
4776 spin_lock_irqsave(&q->lock, flags);
4777 __add_wait_queue(q, &wait);
4778 spin_unlock(&q->lock);
4779 timeout = schedule_timeout(timeout);
4780 spin_lock_irq(&q->lock);
4781 __remove_wait_queue(q, &wait);
4782 spin_unlock_irqrestore(&q->lock, flags);
4787 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4789 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4791 EXPORT_SYMBOL(interruptible_sleep_on);
4794 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4796 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4798 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4800 void __sched sleep_on(wait_queue_head_t *q)
4802 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4804 EXPORT_SYMBOL(sleep_on);
4806 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4808 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4810 EXPORT_SYMBOL(sleep_on_timeout);
4812 #ifdef CONFIG_RT_MUTEXES
4815 * rt_mutex_setprio - set the current priority of a task
4817 * @prio: prio value (kernel-internal form)
4819 * This function changes the 'effective' priority of a task. It does
4820 * not touch ->normal_prio like __setscheduler().
4822 * Used by the rt_mutex code to implement priority inheritance logic.
4824 void rt_mutex_setprio(struct task_struct *p, int prio)
4826 int oldprio, on_rq, running;
4828 const struct sched_class *prev_class;
4830 BUG_ON(prio < 0 || prio > MAX_PRIO);
4832 rq = __task_rq_lock(p);
4834 trace_sched_pi_setprio(p, prio);
4836 prev_class = p->sched_class;
4838 running = task_current(rq, p);
4840 dequeue_task(rq, p, 0);
4842 p->sched_class->put_prev_task(rq, p);
4845 p->sched_class = &rt_sched_class;
4847 p->sched_class = &fair_sched_class;
4852 p->sched_class->set_curr_task(rq);
4854 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4856 check_class_changed(rq, p, prev_class, oldprio);
4857 __task_rq_unlock(rq);
4862 void set_user_nice(struct task_struct *p, long nice)
4864 int old_prio, delta, on_rq;
4865 unsigned long flags;
4868 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4871 * We have to be careful, if called from sys_setpriority(),
4872 * the task might be in the middle of scheduling on another CPU.
4874 rq = task_rq_lock(p, &flags);
4876 * The RT priorities are set via sched_setscheduler(), but we still
4877 * allow the 'normal' nice value to be set - but as expected
4878 * it wont have any effect on scheduling until the task is
4879 * SCHED_FIFO/SCHED_RR:
4881 if (task_has_rt_policy(p)) {
4882 p->static_prio = NICE_TO_PRIO(nice);
4887 dequeue_task(rq, p, 0);
4889 p->static_prio = NICE_TO_PRIO(nice);
4892 p->prio = effective_prio(p);
4893 delta = p->prio - old_prio;
4896 enqueue_task(rq, p, 0);
4898 * If the task increased its priority or is running and
4899 * lowered its priority, then reschedule its CPU:
4901 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4902 resched_task(rq->curr);
4905 task_rq_unlock(rq, p, &flags);
4907 EXPORT_SYMBOL(set_user_nice);
4910 * can_nice - check if a task can reduce its nice value
4914 int can_nice(const struct task_struct *p, const int nice)
4916 /* convert nice value [19,-20] to rlimit style value [1,40] */
4917 int nice_rlim = 20 - nice;
4919 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4920 capable(CAP_SYS_NICE));
4923 #ifdef __ARCH_WANT_SYS_NICE
4926 * sys_nice - change the priority of the current process.
4927 * @increment: priority increment
4929 * sys_setpriority is a more generic, but much slower function that
4930 * does similar things.
4932 SYSCALL_DEFINE1(nice, int, increment)
4937 * Setpriority might change our priority at the same moment.
4938 * We don't have to worry. Conceptually one call occurs first
4939 * and we have a single winner.
4941 if (increment < -40)
4946 nice = TASK_NICE(current) + increment;
4952 if (increment < 0 && !can_nice(current, nice))
4955 retval = security_task_setnice(current, nice);
4959 set_user_nice(current, nice);
4966 * task_prio - return the priority value of a given task.
4967 * @p: the task in question.
4969 * This is the priority value as seen by users in /proc.
4970 * RT tasks are offset by -200. Normal tasks are centered
4971 * around 0, value goes from -16 to +15.
4973 int task_prio(const struct task_struct *p)
4975 return p->prio - MAX_RT_PRIO;
4979 * task_nice - return the nice value of a given task.
4980 * @p: the task in question.
4982 int task_nice(const struct task_struct *p)
4984 return TASK_NICE(p);
4986 EXPORT_SYMBOL(task_nice);
4989 * idle_cpu - is a given cpu idle currently?
4990 * @cpu: the processor in question.
4992 int idle_cpu(int cpu)
4994 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4998 * idle_task - return the idle task for a given cpu.
4999 * @cpu: the processor in question.
5001 struct task_struct *idle_task(int cpu)
5003 return cpu_rq(cpu)->idle;
5007 * find_process_by_pid - find a process with a matching PID value.
5008 * @pid: the pid in question.
5010 static struct task_struct *find_process_by_pid(pid_t pid)
5012 return pid ? find_task_by_vpid(pid) : current;
5015 /* Actually do priority change: must hold rq lock. */
5017 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5020 p->rt_priority = prio;
5021 p->normal_prio = normal_prio(p);
5022 /* we are holding p->pi_lock already */
5023 p->prio = rt_mutex_getprio(p);
5024 if (rt_prio(p->prio))
5025 p->sched_class = &rt_sched_class;
5027 p->sched_class = &fair_sched_class;
5032 * check the target process has a UID that matches the current process's
5034 static bool check_same_owner(struct task_struct *p)
5036 const struct cred *cred = current_cred(), *pcred;
5040 pcred = __task_cred(p);
5041 if (cred->user->user_ns == pcred->user->user_ns)
5042 match = (cred->euid == pcred->euid ||
5043 cred->euid == pcred->uid);
5050 static int __sched_setscheduler(struct task_struct *p, int policy,
5051 const struct sched_param *param, bool user)
5053 int retval, oldprio, oldpolicy = -1, on_rq, running;
5054 unsigned long flags;
5055 const struct sched_class *prev_class;
5059 /* may grab non-irq protected spin_locks */
5060 BUG_ON(in_interrupt());
5062 /* double check policy once rq lock held */
5064 reset_on_fork = p->sched_reset_on_fork;
5065 policy = oldpolicy = p->policy;
5067 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5068 policy &= ~SCHED_RESET_ON_FORK;
5070 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5071 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5072 policy != SCHED_IDLE)
5077 * Valid priorities for SCHED_FIFO and SCHED_RR are
5078 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5079 * SCHED_BATCH and SCHED_IDLE is 0.
5081 if (param->sched_priority < 0 ||
5082 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5083 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5085 if (rt_policy(policy) != (param->sched_priority != 0))
5089 * Allow unprivileged RT tasks to decrease priority:
5091 if (user && !capable(CAP_SYS_NICE)) {
5092 if (rt_policy(policy)) {
5093 unsigned long rlim_rtprio =
5094 task_rlimit(p, RLIMIT_RTPRIO);
5096 /* can't set/change the rt policy */
5097 if (policy != p->policy && !rlim_rtprio)
5100 /* can't increase priority */
5101 if (param->sched_priority > p->rt_priority &&
5102 param->sched_priority > rlim_rtprio)
5107 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5108 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5110 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5111 if (!can_nice(p, TASK_NICE(p)))
5115 /* can't change other user's priorities */
5116 if (!check_same_owner(p))
5119 /* Normal users shall not reset the sched_reset_on_fork flag */
5120 if (p->sched_reset_on_fork && !reset_on_fork)
5125 retval = security_task_setscheduler(p);
5131 * make sure no PI-waiters arrive (or leave) while we are
5132 * changing the priority of the task:
5134 * To be able to change p->policy safely, the appropriate
5135 * runqueue lock must be held.
5137 rq = task_rq_lock(p, &flags);
5140 * Changing the policy of the stop threads its a very bad idea
5142 if (p == rq->stop) {
5143 task_rq_unlock(rq, p, &flags);
5148 * If not changing anything there's no need to proceed further:
5150 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5151 param->sched_priority == p->rt_priority))) {
5153 __task_rq_unlock(rq);
5154 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5158 #ifdef CONFIG_RT_GROUP_SCHED
5161 * Do not allow realtime tasks into groups that have no runtime
5164 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5165 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5166 !task_group_is_autogroup(task_group(p))) {
5167 task_rq_unlock(rq, p, &flags);
5173 /* recheck policy now with rq lock held */
5174 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5175 policy = oldpolicy = -1;
5176 task_rq_unlock(rq, p, &flags);
5180 running = task_current(rq, p);
5182 deactivate_task(rq, p, 0);
5184 p->sched_class->put_prev_task(rq, p);
5186 p->sched_reset_on_fork = reset_on_fork;
5189 prev_class = p->sched_class;
5190 __setscheduler(rq, p, policy, param->sched_priority);
5193 p->sched_class->set_curr_task(rq);
5195 activate_task(rq, p, 0);
5197 check_class_changed(rq, p, prev_class, oldprio);
5198 task_rq_unlock(rq, p, &flags);
5200 rt_mutex_adjust_pi(p);
5206 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5207 * @p: the task in question.
5208 * @policy: new policy.
5209 * @param: structure containing the new RT priority.
5211 * NOTE that the task may be already dead.
5213 int sched_setscheduler(struct task_struct *p, int policy,
5214 const struct sched_param *param)
5216 return __sched_setscheduler(p, policy, param, true);
5218 EXPORT_SYMBOL_GPL(sched_setscheduler);
5221 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5222 * @p: the task in question.
5223 * @policy: new policy.
5224 * @param: structure containing the new RT priority.
5226 * Just like sched_setscheduler, only don't bother checking if the
5227 * current context has permission. For example, this is needed in
5228 * stop_machine(): we create temporary high priority worker threads,
5229 * but our caller might not have that capability.
5231 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5232 const struct sched_param *param)
5234 return __sched_setscheduler(p, policy, param, false);
5238 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5240 struct sched_param lparam;
5241 struct task_struct *p;
5244 if (!param || pid < 0)
5246 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5251 p = find_process_by_pid(pid);
5253 retval = sched_setscheduler(p, policy, &lparam);
5260 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5261 * @pid: the pid in question.
5262 * @policy: new policy.
5263 * @param: structure containing the new RT priority.
5265 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5266 struct sched_param __user *, param)
5268 /* negative values for policy are not valid */
5272 return do_sched_setscheduler(pid, policy, param);
5276 * sys_sched_setparam - set/change the RT priority of a thread
5277 * @pid: the pid in question.
5278 * @param: structure containing the new RT priority.
5280 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5282 return do_sched_setscheduler(pid, -1, param);
5286 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5287 * @pid: the pid in question.
5289 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5291 struct task_struct *p;
5299 p = find_process_by_pid(pid);
5301 retval = security_task_getscheduler(p);
5304 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5311 * sys_sched_getparam - get the RT priority of a thread
5312 * @pid: the pid in question.
5313 * @param: structure containing the RT priority.
5315 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5317 struct sched_param lp;
5318 struct task_struct *p;
5321 if (!param || pid < 0)
5325 p = find_process_by_pid(pid);
5330 retval = security_task_getscheduler(p);
5334 lp.sched_priority = p->rt_priority;
5338 * This one might sleep, we cannot do it with a spinlock held ...
5340 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5349 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5351 cpumask_var_t cpus_allowed, new_mask;
5352 struct task_struct *p;
5358 p = find_process_by_pid(pid);
5365 /* Prevent p going away */
5369 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5373 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5375 goto out_free_cpus_allowed;
5378 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5381 retval = security_task_setscheduler(p);
5385 cpuset_cpus_allowed(p, cpus_allowed);
5386 cpumask_and(new_mask, in_mask, cpus_allowed);
5388 retval = set_cpus_allowed_ptr(p, new_mask);
5391 cpuset_cpus_allowed(p, cpus_allowed);
5392 if (!cpumask_subset(new_mask, cpus_allowed)) {
5394 * We must have raced with a concurrent cpuset
5395 * update. Just reset the cpus_allowed to the
5396 * cpuset's cpus_allowed
5398 cpumask_copy(new_mask, cpus_allowed);
5403 free_cpumask_var(new_mask);
5404 out_free_cpus_allowed:
5405 free_cpumask_var(cpus_allowed);
5412 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5413 struct cpumask *new_mask)
5415 if (len < cpumask_size())
5416 cpumask_clear(new_mask);
5417 else if (len > cpumask_size())
5418 len = cpumask_size();
5420 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5424 * sys_sched_setaffinity - set the cpu affinity of a process
5425 * @pid: pid of the process
5426 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5427 * @user_mask_ptr: user-space pointer to the new cpu mask
5429 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5430 unsigned long __user *, user_mask_ptr)
5432 cpumask_var_t new_mask;
5435 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5438 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5440 retval = sched_setaffinity(pid, new_mask);
5441 free_cpumask_var(new_mask);
5445 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5447 struct task_struct *p;
5448 unsigned long flags;
5455 p = find_process_by_pid(pid);
5459 retval = security_task_getscheduler(p);
5463 raw_spin_lock_irqsave(&p->pi_lock, flags);
5464 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5465 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5475 * sys_sched_getaffinity - get the cpu affinity of a process
5476 * @pid: pid of the process
5477 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5478 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5480 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5481 unsigned long __user *, user_mask_ptr)
5486 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5488 if (len & (sizeof(unsigned long)-1))
5491 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5494 ret = sched_getaffinity(pid, mask);
5496 size_t retlen = min_t(size_t, len, cpumask_size());
5498 if (copy_to_user(user_mask_ptr, mask, retlen))
5503 free_cpumask_var(mask);
5509 * sys_sched_yield - yield the current processor to other threads.
5511 * This function yields the current CPU to other tasks. If there are no
5512 * other threads running on this CPU then this function will return.
5514 SYSCALL_DEFINE0(sched_yield)
5516 struct rq *rq = this_rq_lock();
5518 schedstat_inc(rq, yld_count);
5519 current->sched_class->yield_task(rq);
5522 * Since we are going to call schedule() anyway, there's
5523 * no need to preempt or enable interrupts:
5525 __release(rq->lock);
5526 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5527 do_raw_spin_unlock(&rq->lock);
5528 preempt_enable_no_resched();
5535 static inline int should_resched(void)
5537 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5540 static void __cond_resched(void)
5542 add_preempt_count(PREEMPT_ACTIVE);
5544 sub_preempt_count(PREEMPT_ACTIVE);
5547 int __sched _cond_resched(void)
5549 if (should_resched()) {
5555 EXPORT_SYMBOL(_cond_resched);
5558 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5559 * call schedule, and on return reacquire the lock.
5561 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5562 * operations here to prevent schedule() from being called twice (once via
5563 * spin_unlock(), once by hand).
5565 int __cond_resched_lock(spinlock_t *lock)
5567 int resched = should_resched();
5570 lockdep_assert_held(lock);
5572 if (spin_needbreak(lock) || resched) {
5583 EXPORT_SYMBOL(__cond_resched_lock);
5585 int __sched __cond_resched_softirq(void)
5587 BUG_ON(!in_softirq());
5589 if (should_resched()) {
5597 EXPORT_SYMBOL(__cond_resched_softirq);
5600 * yield - yield the current processor to other threads.
5602 * This is a shortcut for kernel-space yielding - it marks the
5603 * thread runnable and calls sys_sched_yield().
5605 void __sched yield(void)
5607 set_current_state(TASK_RUNNING);
5610 EXPORT_SYMBOL(yield);
5613 * yield_to - yield the current processor to another thread in
5614 * your thread group, or accelerate that thread toward the
5615 * processor it's on.
5617 * @preempt: whether task preemption is allowed or not
5619 * It's the caller's job to ensure that the target task struct
5620 * can't go away on us before we can do any checks.
5622 * Returns true if we indeed boosted the target task.
5624 bool __sched yield_to(struct task_struct *p, bool preempt)
5626 struct task_struct *curr = current;
5627 struct rq *rq, *p_rq;
5628 unsigned long flags;
5631 local_irq_save(flags);
5636 double_rq_lock(rq, p_rq);
5637 while (task_rq(p) != p_rq) {
5638 double_rq_unlock(rq, p_rq);
5642 if (!curr->sched_class->yield_to_task)
5645 if (curr->sched_class != p->sched_class)
5648 if (task_running(p_rq, p) || p->state)
5651 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5653 schedstat_inc(rq, yld_count);
5655 * Make p's CPU reschedule; pick_next_entity takes care of
5658 if (preempt && rq != p_rq)
5659 resched_task(p_rq->curr);
5663 double_rq_unlock(rq, p_rq);
5664 local_irq_restore(flags);
5671 EXPORT_SYMBOL_GPL(yield_to);
5674 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5675 * that process accounting knows that this is a task in IO wait state.
5677 void __sched io_schedule(void)
5679 struct rq *rq = raw_rq();
5681 delayacct_blkio_start();
5682 atomic_inc(&rq->nr_iowait);
5683 blk_flush_plug(current);
5684 current->in_iowait = 1;
5686 current->in_iowait = 0;
5687 atomic_dec(&rq->nr_iowait);
5688 delayacct_blkio_end();
5690 EXPORT_SYMBOL(io_schedule);
5692 long __sched io_schedule_timeout(long timeout)
5694 struct rq *rq = raw_rq();
5697 delayacct_blkio_start();
5698 atomic_inc(&rq->nr_iowait);
5699 blk_flush_plug(current);
5700 current->in_iowait = 1;
5701 ret = schedule_timeout(timeout);
5702 current->in_iowait = 0;
5703 atomic_dec(&rq->nr_iowait);
5704 delayacct_blkio_end();
5709 * sys_sched_get_priority_max - return maximum RT priority.
5710 * @policy: scheduling class.
5712 * this syscall returns the maximum rt_priority that can be used
5713 * by a given scheduling class.
5715 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5722 ret = MAX_USER_RT_PRIO-1;
5734 * sys_sched_get_priority_min - return minimum RT priority.
5735 * @policy: scheduling class.
5737 * this syscall returns the minimum rt_priority that can be used
5738 * by a given scheduling class.
5740 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5758 * sys_sched_rr_get_interval - return the default timeslice of a process.
5759 * @pid: pid of the process.
5760 * @interval: userspace pointer to the timeslice value.
5762 * this syscall writes the default timeslice value of a given process
5763 * into the user-space timespec buffer. A value of '0' means infinity.
5765 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5766 struct timespec __user *, interval)
5768 struct task_struct *p;
5769 unsigned int time_slice;
5770 unsigned long flags;
5780 p = find_process_by_pid(pid);
5784 retval = security_task_getscheduler(p);
5788 rq = task_rq_lock(p, &flags);
5789 time_slice = p->sched_class->get_rr_interval(rq, p);
5790 task_rq_unlock(rq, p, &flags);
5793 jiffies_to_timespec(time_slice, &t);
5794 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5802 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5804 void sched_show_task(struct task_struct *p)
5806 unsigned long free = 0;
5809 state = p->state ? __ffs(p->state) + 1 : 0;
5810 printk(KERN_INFO "%-15.15s %c", p->comm,
5811 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5812 #if BITS_PER_LONG == 32
5813 if (state == TASK_RUNNING)
5814 printk(KERN_CONT " running ");
5816 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5818 if (state == TASK_RUNNING)
5819 printk(KERN_CONT " running task ");
5821 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5823 #ifdef CONFIG_DEBUG_STACK_USAGE
5824 free = stack_not_used(p);
5826 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5827 task_pid_nr(p), task_pid_nr(p->real_parent),
5828 (unsigned long)task_thread_info(p)->flags);
5830 show_stack(p, NULL);
5833 void show_state_filter(unsigned long state_filter)
5835 struct task_struct *g, *p;
5837 #if BITS_PER_LONG == 32
5839 " task PC stack pid father\n");
5842 " task PC stack pid father\n");
5844 read_lock(&tasklist_lock);
5845 do_each_thread(g, p) {
5847 * reset the NMI-timeout, listing all files on a slow
5848 * console might take a lot of time:
5850 touch_nmi_watchdog();
5851 if (!state_filter || (p->state & state_filter))
5853 } while_each_thread(g, p);
5855 touch_all_softlockup_watchdogs();
5857 #ifdef CONFIG_SCHED_DEBUG
5858 sysrq_sched_debug_show();
5860 read_unlock(&tasklist_lock);
5862 * Only show locks if all tasks are dumped:
5865 debug_show_all_locks();
5868 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5870 idle->sched_class = &idle_sched_class;
5874 * init_idle - set up an idle thread for a given CPU
5875 * @idle: task in question
5876 * @cpu: cpu the idle task belongs to
5878 * NOTE: this function does not set the idle thread's NEED_RESCHED
5879 * flag, to make booting more robust.
5881 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5883 struct rq *rq = cpu_rq(cpu);
5884 unsigned long flags;
5886 raw_spin_lock_irqsave(&rq->lock, flags);
5889 idle->state = TASK_RUNNING;
5890 idle->se.exec_start = sched_clock();
5892 do_set_cpus_allowed(idle, cpumask_of(cpu));
5894 * We're having a chicken and egg problem, even though we are
5895 * holding rq->lock, the cpu isn't yet set to this cpu so the
5896 * lockdep check in task_group() will fail.
5898 * Similar case to sched_fork(). / Alternatively we could
5899 * use task_rq_lock() here and obtain the other rq->lock.
5904 __set_task_cpu(idle, cpu);
5907 rq->curr = rq->idle = idle;
5908 #if defined(CONFIG_SMP)
5911 raw_spin_unlock_irqrestore(&rq->lock, flags);
5913 /* Set the preempt count _outside_ the spinlocks! */
5914 task_thread_info(idle)->preempt_count = 0;
5917 * The idle tasks have their own, simple scheduling class:
5919 idle->sched_class = &idle_sched_class;
5920 ftrace_graph_init_idle_task(idle, cpu);
5924 * In a system that switches off the HZ timer nohz_cpu_mask
5925 * indicates which cpus entered this state. This is used
5926 * in the rcu update to wait only for active cpus. For system
5927 * which do not switch off the HZ timer nohz_cpu_mask should
5928 * always be CPU_BITS_NONE.
5930 cpumask_var_t nohz_cpu_mask;
5933 * Increase the granularity value when there are more CPUs,
5934 * because with more CPUs the 'effective latency' as visible
5935 * to users decreases. But the relationship is not linear,
5936 * so pick a second-best guess by going with the log2 of the
5939 * This idea comes from the SD scheduler of Con Kolivas:
5941 static int get_update_sysctl_factor(void)
5943 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5944 unsigned int factor;
5946 switch (sysctl_sched_tunable_scaling) {
5947 case SCHED_TUNABLESCALING_NONE:
5950 case SCHED_TUNABLESCALING_LINEAR:
5953 case SCHED_TUNABLESCALING_LOG:
5955 factor = 1 + ilog2(cpus);
5962 static void update_sysctl(void)
5964 unsigned int factor = get_update_sysctl_factor();
5966 #define SET_SYSCTL(name) \
5967 (sysctl_##name = (factor) * normalized_sysctl_##name)
5968 SET_SYSCTL(sched_min_granularity);
5969 SET_SYSCTL(sched_latency);
5970 SET_SYSCTL(sched_wakeup_granularity);
5974 static inline void sched_init_granularity(void)
5980 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
5982 if (p->sched_class && p->sched_class->set_cpus_allowed)
5983 p->sched_class->set_cpus_allowed(p, new_mask);
5985 cpumask_copy(&p->cpus_allowed, new_mask);
5986 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5991 * This is how migration works:
5993 * 1) we invoke migration_cpu_stop() on the target CPU using
5995 * 2) stopper starts to run (implicitly forcing the migrated thread
5997 * 3) it checks whether the migrated task is still in the wrong runqueue.
5998 * 4) if it's in the wrong runqueue then the migration thread removes
5999 * it and puts it into the right queue.
6000 * 5) stopper completes and stop_one_cpu() returns and the migration
6005 * Change a given task's CPU affinity. Migrate the thread to a
6006 * proper CPU and schedule it away if the CPU it's executing on
6007 * is removed from the allowed bitmask.
6009 * NOTE: the caller must have a valid reference to the task, the
6010 * task must not exit() & deallocate itself prematurely. The
6011 * call is not atomic; no spinlocks may be held.
6013 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6015 unsigned long flags;
6017 unsigned int dest_cpu;
6020 rq = task_rq_lock(p, &flags);
6022 if (cpumask_equal(&p->cpus_allowed, new_mask))
6025 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6030 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6035 do_set_cpus_allowed(p, new_mask);
6037 /* Can the task run on the task's current CPU? If so, we're done */
6038 if (cpumask_test_cpu(task_cpu(p), new_mask))
6041 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6043 struct migration_arg arg = { p, dest_cpu };
6044 /* Need help from migration thread: drop lock and wait. */
6045 task_rq_unlock(rq, p, &flags);
6046 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6047 tlb_migrate_finish(p->mm);
6051 task_rq_unlock(rq, p, &flags);
6055 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6058 * Move (not current) task off this cpu, onto dest cpu. We're doing
6059 * this because either it can't run here any more (set_cpus_allowed()
6060 * away from this CPU, or CPU going down), or because we're
6061 * attempting to rebalance this task on exec (sched_exec).
6063 * So we race with normal scheduler movements, but that's OK, as long
6064 * as the task is no longer on this CPU.
6066 * Returns non-zero if task was successfully migrated.
6068 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6070 struct rq *rq_dest, *rq_src;
6073 if (unlikely(!cpu_active(dest_cpu)))
6076 rq_src = cpu_rq(src_cpu);
6077 rq_dest = cpu_rq(dest_cpu);
6079 raw_spin_lock(&p->pi_lock);
6080 double_rq_lock(rq_src, rq_dest);
6081 /* Already moved. */
6082 if (task_cpu(p) != src_cpu)
6084 /* Affinity changed (again). */
6085 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6089 * If we're not on a rq, the next wake-up will ensure we're
6093 deactivate_task(rq_src, p, 0);
6094 set_task_cpu(p, dest_cpu);
6095 activate_task(rq_dest, p, 0);
6096 check_preempt_curr(rq_dest, p, 0);
6101 double_rq_unlock(rq_src, rq_dest);
6102 raw_spin_unlock(&p->pi_lock);
6107 * migration_cpu_stop - this will be executed by a highprio stopper thread
6108 * and performs thread migration by bumping thread off CPU then
6109 * 'pushing' onto another runqueue.
6111 static int migration_cpu_stop(void *data)
6113 struct migration_arg *arg = data;
6116 * The original target cpu might have gone down and we might
6117 * be on another cpu but it doesn't matter.
6119 local_irq_disable();
6120 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6125 #ifdef CONFIG_HOTPLUG_CPU
6128 * Ensures that the idle task is using init_mm right before its cpu goes
6131 void idle_task_exit(void)
6133 struct mm_struct *mm = current->active_mm;
6135 BUG_ON(cpu_online(smp_processor_id()));
6138 switch_mm(mm, &init_mm, current);
6143 * While a dead CPU has no uninterruptible tasks queued at this point,
6144 * it might still have a nonzero ->nr_uninterruptible counter, because
6145 * for performance reasons the counter is not stricly tracking tasks to
6146 * their home CPUs. So we just add the counter to another CPU's counter,
6147 * to keep the global sum constant after CPU-down:
6149 static void migrate_nr_uninterruptible(struct rq *rq_src)
6151 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6153 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6154 rq_src->nr_uninterruptible = 0;
6158 * remove the tasks which were accounted by rq from calc_load_tasks.
6160 static void calc_global_load_remove(struct rq *rq)
6162 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6163 rq->calc_load_active = 0;
6167 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6168 * try_to_wake_up()->select_task_rq().
6170 * Called with rq->lock held even though we'er in stop_machine() and
6171 * there's no concurrency possible, we hold the required locks anyway
6172 * because of lock validation efforts.
6174 static void migrate_tasks(unsigned int dead_cpu)
6176 struct rq *rq = cpu_rq(dead_cpu);
6177 struct task_struct *next, *stop = rq->stop;
6181 * Fudge the rq selection such that the below task selection loop
6182 * doesn't get stuck on the currently eligible stop task.
6184 * We're currently inside stop_machine() and the rq is either stuck
6185 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6186 * either way we should never end up calling schedule() until we're
6193 * There's this thread running, bail when that's the only
6196 if (rq->nr_running == 1)
6199 next = pick_next_task(rq);
6201 next->sched_class->put_prev_task(rq, next);
6203 /* Find suitable destination for @next, with force if needed. */
6204 dest_cpu = select_fallback_rq(dead_cpu, next);
6205 raw_spin_unlock(&rq->lock);
6207 __migrate_task(next, dead_cpu, dest_cpu);
6209 raw_spin_lock(&rq->lock);
6215 #endif /* CONFIG_HOTPLUG_CPU */
6217 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6219 static struct ctl_table sd_ctl_dir[] = {
6221 .procname = "sched_domain",
6227 static struct ctl_table sd_ctl_root[] = {
6229 .procname = "kernel",
6231 .child = sd_ctl_dir,
6236 static struct ctl_table *sd_alloc_ctl_entry(int n)
6238 struct ctl_table *entry =
6239 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6244 static void sd_free_ctl_entry(struct ctl_table **tablep)
6246 struct ctl_table *entry;
6249 * In the intermediate directories, both the child directory and
6250 * procname are dynamically allocated and could fail but the mode
6251 * will always be set. In the lowest directory the names are
6252 * static strings and all have proc handlers.
6254 for (entry = *tablep; entry->mode; entry++) {
6256 sd_free_ctl_entry(&entry->child);
6257 if (entry->proc_handler == NULL)
6258 kfree(entry->procname);
6266 set_table_entry(struct ctl_table *entry,
6267 const char *procname, void *data, int maxlen,
6268 mode_t mode, proc_handler *proc_handler)
6270 entry->procname = procname;
6272 entry->maxlen = maxlen;
6274 entry->proc_handler = proc_handler;
6277 static struct ctl_table *
6278 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6280 struct ctl_table *table = sd_alloc_ctl_entry(13);
6285 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6286 sizeof(long), 0644, proc_doulongvec_minmax);
6287 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6288 sizeof(long), 0644, proc_doulongvec_minmax);
6289 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6290 sizeof(int), 0644, proc_dointvec_minmax);
6291 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6292 sizeof(int), 0644, proc_dointvec_minmax);
6293 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6294 sizeof(int), 0644, proc_dointvec_minmax);
6295 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6296 sizeof(int), 0644, proc_dointvec_minmax);
6297 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6298 sizeof(int), 0644, proc_dointvec_minmax);
6299 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6300 sizeof(int), 0644, proc_dointvec_minmax);
6301 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6302 sizeof(int), 0644, proc_dointvec_minmax);
6303 set_table_entry(&table[9], "cache_nice_tries",
6304 &sd->cache_nice_tries,
6305 sizeof(int), 0644, proc_dointvec_minmax);
6306 set_table_entry(&table[10], "flags", &sd->flags,
6307 sizeof(int), 0644, proc_dointvec_minmax);
6308 set_table_entry(&table[11], "name", sd->name,
6309 CORENAME_MAX_SIZE, 0444, proc_dostring);
6310 /* &table[12] is terminator */
6315 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6317 struct ctl_table *entry, *table;
6318 struct sched_domain *sd;
6319 int domain_num = 0, i;
6322 for_each_domain(cpu, sd)
6324 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6329 for_each_domain(cpu, sd) {
6330 snprintf(buf, 32, "domain%d", i);
6331 entry->procname = kstrdup(buf, GFP_KERNEL);
6333 entry->child = sd_alloc_ctl_domain_table(sd);
6340 static struct ctl_table_header *sd_sysctl_header;
6341 static void register_sched_domain_sysctl(void)
6343 int i, cpu_num = num_possible_cpus();
6344 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6347 WARN_ON(sd_ctl_dir[0].child);
6348 sd_ctl_dir[0].child = entry;
6353 for_each_possible_cpu(i) {
6354 snprintf(buf, 32, "cpu%d", i);
6355 entry->procname = kstrdup(buf, GFP_KERNEL);
6357 entry->child = sd_alloc_ctl_cpu_table(i);
6361 WARN_ON(sd_sysctl_header);
6362 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6365 /* may be called multiple times per register */
6366 static void unregister_sched_domain_sysctl(void)
6368 if (sd_sysctl_header)
6369 unregister_sysctl_table(sd_sysctl_header);
6370 sd_sysctl_header = NULL;
6371 if (sd_ctl_dir[0].child)
6372 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6375 static void register_sched_domain_sysctl(void)
6378 static void unregister_sched_domain_sysctl(void)
6383 static void set_rq_online(struct rq *rq)
6386 const struct sched_class *class;
6388 cpumask_set_cpu(rq->cpu, rq->rd->online);
6391 for_each_class(class) {
6392 if (class->rq_online)
6393 class->rq_online(rq);
6398 static void set_rq_offline(struct rq *rq)
6401 const struct sched_class *class;
6403 for_each_class(class) {
6404 if (class->rq_offline)
6405 class->rq_offline(rq);
6408 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6414 * migration_call - callback that gets triggered when a CPU is added.
6415 * Here we can start up the necessary migration thread for the new CPU.
6417 static int __cpuinit
6418 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6420 int cpu = (long)hcpu;
6421 unsigned long flags;
6422 struct rq *rq = cpu_rq(cpu);
6424 switch (action & ~CPU_TASKS_FROZEN) {
6426 case CPU_UP_PREPARE:
6427 rq->calc_load_update = calc_load_update;
6431 /* Update our root-domain */
6432 raw_spin_lock_irqsave(&rq->lock, flags);
6434 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6438 raw_spin_unlock_irqrestore(&rq->lock, flags);
6441 #ifdef CONFIG_HOTPLUG_CPU
6443 sched_ttwu_pending();
6444 /* Update our root-domain */
6445 raw_spin_lock_irqsave(&rq->lock, flags);
6447 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6451 BUG_ON(rq->nr_running != 1); /* the migration thread */
6452 raw_spin_unlock_irqrestore(&rq->lock, flags);
6454 migrate_nr_uninterruptible(rq);
6455 calc_global_load_remove(rq);
6460 update_max_interval();
6466 * Register at high priority so that task migration (migrate_all_tasks)
6467 * happens before everything else. This has to be lower priority than
6468 * the notifier in the perf_event subsystem, though.
6470 static struct notifier_block __cpuinitdata migration_notifier = {
6471 .notifier_call = migration_call,
6472 .priority = CPU_PRI_MIGRATION,
6475 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6476 unsigned long action, void *hcpu)
6478 switch (action & ~CPU_TASKS_FROZEN) {
6480 case CPU_DOWN_FAILED:
6481 set_cpu_active((long)hcpu, true);
6488 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6489 unsigned long action, void *hcpu)
6491 switch (action & ~CPU_TASKS_FROZEN) {
6492 case CPU_DOWN_PREPARE:
6493 set_cpu_active((long)hcpu, false);
6500 static int __init migration_init(void)
6502 void *cpu = (void *)(long)smp_processor_id();
6505 /* Initialize migration for the boot CPU */
6506 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6507 BUG_ON(err == NOTIFY_BAD);
6508 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6509 register_cpu_notifier(&migration_notifier);
6511 /* Register cpu active notifiers */
6512 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6513 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6517 early_initcall(migration_init);
6522 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6524 #ifdef CONFIG_SCHED_DEBUG
6526 static __read_mostly int sched_domain_debug_enabled;
6528 static int __init sched_domain_debug_setup(char *str)
6530 sched_domain_debug_enabled = 1;
6534 early_param("sched_debug", sched_domain_debug_setup);
6536 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6537 struct cpumask *groupmask)
6539 struct sched_group *group = sd->groups;
6542 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6543 cpumask_clear(groupmask);
6545 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6547 if (!(sd->flags & SD_LOAD_BALANCE)) {
6548 printk("does not load-balance\n");
6550 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6555 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6557 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6558 printk(KERN_ERR "ERROR: domain->span does not contain "
6561 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6562 printk(KERN_ERR "ERROR: domain->groups does not contain"
6566 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6570 printk(KERN_ERR "ERROR: group is NULL\n");
6574 if (!group->sgp->power) {
6575 printk(KERN_CONT "\n");
6576 printk(KERN_ERR "ERROR: domain->cpu_power not "
6581 if (!cpumask_weight(sched_group_cpus(group))) {
6582 printk(KERN_CONT "\n");
6583 printk(KERN_ERR "ERROR: empty group\n");
6587 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6588 printk(KERN_CONT "\n");
6589 printk(KERN_ERR "ERROR: repeated CPUs\n");
6593 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6595 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6597 printk(KERN_CONT " %s", str);
6598 if (group->sgp->power != SCHED_POWER_SCALE) {
6599 printk(KERN_CONT " (cpu_power = %d)",
6603 group = group->next;
6604 } while (group != sd->groups);
6605 printk(KERN_CONT "\n");
6607 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6608 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6611 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6612 printk(KERN_ERR "ERROR: parent span is not a superset "
6613 "of domain->span\n");
6617 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6621 if (!sched_domain_debug_enabled)
6625 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6629 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6632 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6640 #else /* !CONFIG_SCHED_DEBUG */
6641 # define sched_domain_debug(sd, cpu) do { } while (0)
6642 #endif /* CONFIG_SCHED_DEBUG */
6644 static int sd_degenerate(struct sched_domain *sd)
6646 if (cpumask_weight(sched_domain_span(sd)) == 1)
6649 /* Following flags need at least 2 groups */
6650 if (sd->flags & (SD_LOAD_BALANCE |
6651 SD_BALANCE_NEWIDLE |
6655 SD_SHARE_PKG_RESOURCES)) {
6656 if (sd->groups != sd->groups->next)
6660 /* Following flags don't use groups */
6661 if (sd->flags & (SD_WAKE_AFFINE))
6668 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6670 unsigned long cflags = sd->flags, pflags = parent->flags;
6672 if (sd_degenerate(parent))
6675 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6678 /* Flags needing groups don't count if only 1 group in parent */
6679 if (parent->groups == parent->groups->next) {
6680 pflags &= ~(SD_LOAD_BALANCE |
6681 SD_BALANCE_NEWIDLE |
6685 SD_SHARE_PKG_RESOURCES);
6686 if (nr_node_ids == 1)
6687 pflags &= ~SD_SERIALIZE;
6689 if (~cflags & pflags)
6695 static void free_rootdomain(struct rcu_head *rcu)
6697 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6699 cpupri_cleanup(&rd->cpupri);
6700 free_cpumask_var(rd->rto_mask);
6701 free_cpumask_var(rd->online);
6702 free_cpumask_var(rd->span);
6706 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6708 struct root_domain *old_rd = NULL;
6709 unsigned long flags;
6711 raw_spin_lock_irqsave(&rq->lock, flags);
6716 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6719 cpumask_clear_cpu(rq->cpu, old_rd->span);
6722 * If we dont want to free the old_rt yet then
6723 * set old_rd to NULL to skip the freeing later
6726 if (!atomic_dec_and_test(&old_rd->refcount))
6730 atomic_inc(&rd->refcount);
6733 cpumask_set_cpu(rq->cpu, rd->span);
6734 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6737 raw_spin_unlock_irqrestore(&rq->lock, flags);
6740 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6743 static int init_rootdomain(struct root_domain *rd)
6745 memset(rd, 0, sizeof(*rd));
6747 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6749 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6751 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6754 if (cpupri_init(&rd->cpupri) != 0)
6759 free_cpumask_var(rd->rto_mask);
6761 free_cpumask_var(rd->online);
6763 free_cpumask_var(rd->span);
6768 static void init_defrootdomain(void)
6770 init_rootdomain(&def_root_domain);
6772 atomic_set(&def_root_domain.refcount, 1);
6775 static struct root_domain *alloc_rootdomain(void)
6777 struct root_domain *rd;
6779 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6783 if (init_rootdomain(rd) != 0) {
6791 static void free_sched_groups(struct sched_group *sg, int free_sgp)
6793 struct sched_group *tmp, *first;
6802 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
6807 } while (sg != first);
6810 static void free_sched_domain(struct rcu_head *rcu)
6812 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6815 * If its an overlapping domain it has private groups, iterate and
6818 if (sd->flags & SD_OVERLAP) {
6819 free_sched_groups(sd->groups, 1);
6820 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6821 kfree(sd->groups->sgp);
6827 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6829 call_rcu(&sd->rcu, free_sched_domain);
6832 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6834 for (; sd; sd = sd->parent)
6835 destroy_sched_domain(sd, cpu);
6839 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6840 * hold the hotplug lock.
6843 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6845 struct rq *rq = cpu_rq(cpu);
6846 struct sched_domain *tmp;
6848 /* Remove the sched domains which do not contribute to scheduling. */
6849 for (tmp = sd; tmp; ) {
6850 struct sched_domain *parent = tmp->parent;
6854 if (sd_parent_degenerate(tmp, parent)) {
6855 tmp->parent = parent->parent;
6857 parent->parent->child = tmp;
6858 destroy_sched_domain(parent, cpu);
6863 if (sd && sd_degenerate(sd)) {
6866 destroy_sched_domain(tmp, cpu);
6871 sched_domain_debug(sd, cpu);
6873 rq_attach_root(rq, rd);
6875 rcu_assign_pointer(rq->sd, sd);
6876 destroy_sched_domains(tmp, cpu);
6879 /* cpus with isolated domains */
6880 static cpumask_var_t cpu_isolated_map;
6882 /* Setup the mask of cpus configured for isolated domains */
6883 static int __init isolated_cpu_setup(char *str)
6885 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6886 cpulist_parse(str, cpu_isolated_map);
6890 __setup("isolcpus=", isolated_cpu_setup);
6892 #define SD_NODES_PER_DOMAIN 16
6897 * find_next_best_node - find the next node to include in a sched_domain
6898 * @node: node whose sched_domain we're building
6899 * @used_nodes: nodes already in the sched_domain
6901 * Find the next node to include in a given scheduling domain. Simply
6902 * finds the closest node not already in the @used_nodes map.
6904 * Should use nodemask_t.
6906 static int find_next_best_node(int node, nodemask_t *used_nodes)
6908 int i, n, val, min_val, best_node = -1;
6912 for (i = 0; i < nr_node_ids; i++) {
6913 /* Start at @node */
6914 n = (node + i) % nr_node_ids;
6916 if (!nr_cpus_node(n))
6919 /* Skip already used nodes */
6920 if (node_isset(n, *used_nodes))
6923 /* Simple min distance search */
6924 val = node_distance(node, n);
6926 if (val < min_val) {
6932 if (best_node != -1)
6933 node_set(best_node, *used_nodes);
6938 * sched_domain_node_span - get a cpumask for a node's sched_domain
6939 * @node: node whose cpumask we're constructing
6940 * @span: resulting cpumask
6942 * Given a node, construct a good cpumask for its sched_domain to span. It
6943 * should be one that prevents unnecessary balancing, but also spreads tasks
6946 static void sched_domain_node_span(int node, struct cpumask *span)
6948 nodemask_t used_nodes;
6951 cpumask_clear(span);
6952 nodes_clear(used_nodes);
6954 cpumask_or(span, span, cpumask_of_node(node));
6955 node_set(node, used_nodes);
6957 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6958 int next_node = find_next_best_node(node, &used_nodes);
6961 cpumask_or(span, span, cpumask_of_node(next_node));
6965 static const struct cpumask *cpu_node_mask(int cpu)
6967 lockdep_assert_held(&sched_domains_mutex);
6969 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
6971 return sched_domains_tmpmask;
6974 static const struct cpumask *cpu_allnodes_mask(int cpu)
6976 return cpu_possible_mask;
6978 #endif /* CONFIG_NUMA */
6980 static const struct cpumask *cpu_cpu_mask(int cpu)
6982 return cpumask_of_node(cpu_to_node(cpu));
6985 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6988 struct sched_domain **__percpu sd;
6989 struct sched_group **__percpu sg;
6990 struct sched_group_power **__percpu sgp;
6994 struct sched_domain ** __percpu sd;
6995 struct root_domain *rd;
7005 struct sched_domain_topology_level;
7007 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
7008 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
7010 #define SDTL_OVERLAP 0x01
7012 struct sched_domain_topology_level {
7013 sched_domain_init_f init;
7014 sched_domain_mask_f mask;
7016 struct sd_data data;
7020 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
7022 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
7023 const struct cpumask *span = sched_domain_span(sd);
7024 struct cpumask *covered = sched_domains_tmpmask;
7025 struct sd_data *sdd = sd->private;
7026 struct sched_domain *child;
7029 cpumask_clear(covered);
7031 for_each_cpu(i, span) {
7032 struct cpumask *sg_span;
7034 if (cpumask_test_cpu(i, covered))
7037 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7038 GFP_KERNEL, cpu_to_node(i));
7043 sg_span = sched_group_cpus(sg);
7045 child = *per_cpu_ptr(sdd->sd, i);
7047 child = child->child;
7048 cpumask_copy(sg_span, sched_domain_span(child));
7050 cpumask_set_cpu(i, sg_span);
7052 cpumask_or(covered, covered, sg_span);
7054 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
7055 atomic_inc(&sg->sgp->ref);
7057 if (cpumask_test_cpu(cpu, sg_span))
7067 sd->groups = groups;
7072 free_sched_groups(first, 0);
7077 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
7079 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
7080 struct sched_domain *child = sd->child;
7083 cpu = cpumask_first(sched_domain_span(child));
7086 *sg = *per_cpu_ptr(sdd->sg, cpu);
7087 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
7088 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
7095 * build_sched_groups will build a circular linked list of the groups
7096 * covered by the given span, and will set each group's ->cpumask correctly,
7097 * and ->cpu_power to 0.
7099 * Assumes the sched_domain tree is fully constructed
7102 build_sched_groups(struct sched_domain *sd, int cpu)
7104 struct sched_group *first = NULL, *last = NULL;
7105 struct sd_data *sdd = sd->private;
7106 const struct cpumask *span = sched_domain_span(sd);
7107 struct cpumask *covered;
7110 get_group(cpu, sdd, &sd->groups);
7111 atomic_inc(&sd->groups->ref);
7113 if (cpu != cpumask_first(sched_domain_span(sd)))
7116 lockdep_assert_held(&sched_domains_mutex);
7117 covered = sched_domains_tmpmask;
7119 cpumask_clear(covered);
7121 for_each_cpu(i, span) {
7122 struct sched_group *sg;
7123 int group = get_group(i, sdd, &sg);
7126 if (cpumask_test_cpu(i, covered))
7129 cpumask_clear(sched_group_cpus(sg));
7132 for_each_cpu(j, span) {
7133 if (get_group(j, sdd, NULL) != group)
7136 cpumask_set_cpu(j, covered);
7137 cpumask_set_cpu(j, sched_group_cpus(sg));
7152 * Initialize sched groups cpu_power.
7154 * cpu_power indicates the capacity of sched group, which is used while
7155 * distributing the load between different sched groups in a sched domain.
7156 * Typically cpu_power for all the groups in a sched domain will be same unless
7157 * there are asymmetries in the topology. If there are asymmetries, group
7158 * having more cpu_power will pickup more load compared to the group having
7161 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7163 struct sched_group *sg = sd->groups;
7165 WARN_ON(!sd || !sg);
7168 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
7170 } while (sg != sd->groups);
7172 if (cpu != group_first_cpu(sg))
7175 update_group_power(sd, cpu);
7179 * Initializers for schedule domains
7180 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7183 #ifdef CONFIG_SCHED_DEBUG
7184 # define SD_INIT_NAME(sd, type) sd->name = #type
7186 # define SD_INIT_NAME(sd, type) do { } while (0)
7189 #define SD_INIT_FUNC(type) \
7190 static noinline struct sched_domain * \
7191 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7193 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7194 *sd = SD_##type##_INIT; \
7195 SD_INIT_NAME(sd, type); \
7196 sd->private = &tl->data; \
7202 SD_INIT_FUNC(ALLNODES)
7205 #ifdef CONFIG_SCHED_SMT
7206 SD_INIT_FUNC(SIBLING)
7208 #ifdef CONFIG_SCHED_MC
7211 #ifdef CONFIG_SCHED_BOOK
7215 static int default_relax_domain_level = -1;
7216 int sched_domain_level_max;
7218 static int __init setup_relax_domain_level(char *str)
7220 if (kstrtoint(str, 0, &default_relax_domain_level))
7221 pr_warn("Unable to set relax_domain_level\n");
7225 __setup("relax_domain_level=", setup_relax_domain_level);
7227 static void set_domain_attribute(struct sched_domain *sd,
7228 struct sched_domain_attr *attr)
7232 if (!attr || attr->relax_domain_level < 0) {
7233 if (default_relax_domain_level < 0)
7236 request = default_relax_domain_level;
7238 request = attr->relax_domain_level;
7239 if (request < sd->level) {
7240 /* turn off idle balance on this domain */
7241 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7243 /* turn on idle balance on this domain */
7244 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7248 static void __sdt_free(const struct cpumask *cpu_map);
7249 static int __sdt_alloc(const struct cpumask *cpu_map);
7251 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7252 const struct cpumask *cpu_map)
7256 if (!atomic_read(&d->rd->refcount))
7257 free_rootdomain(&d->rd->rcu); /* fall through */
7259 free_percpu(d->sd); /* fall through */
7261 __sdt_free(cpu_map); /* fall through */
7267 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7268 const struct cpumask *cpu_map)
7270 memset(d, 0, sizeof(*d));
7272 if (__sdt_alloc(cpu_map))
7273 return sa_sd_storage;
7274 d->sd = alloc_percpu(struct sched_domain *);
7276 return sa_sd_storage;
7277 d->rd = alloc_rootdomain();
7280 return sa_rootdomain;
7284 * NULL the sd_data elements we've used to build the sched_domain and
7285 * sched_group structure so that the subsequent __free_domain_allocs()
7286 * will not free the data we're using.
7288 static void claim_allocations(int cpu, struct sched_domain *sd)
7290 struct sd_data *sdd = sd->private;
7292 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7293 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7295 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
7296 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7298 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
7299 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
7302 #ifdef CONFIG_SCHED_SMT
7303 static const struct cpumask *cpu_smt_mask(int cpu)
7305 return topology_thread_cpumask(cpu);
7310 * Topology list, bottom-up.
7312 static struct sched_domain_topology_level default_topology[] = {
7313 #ifdef CONFIG_SCHED_SMT
7314 { sd_init_SIBLING, cpu_smt_mask, },
7316 #ifdef CONFIG_SCHED_MC
7317 { sd_init_MC, cpu_coregroup_mask, },
7319 #ifdef CONFIG_SCHED_BOOK
7320 { sd_init_BOOK, cpu_book_mask, },
7322 { sd_init_CPU, cpu_cpu_mask, },
7324 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
7325 { sd_init_ALLNODES, cpu_allnodes_mask, },
7330 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7332 static int __sdt_alloc(const struct cpumask *cpu_map)
7334 struct sched_domain_topology_level *tl;
7337 for (tl = sched_domain_topology; tl->init; tl++) {
7338 struct sd_data *sdd = &tl->data;
7340 sdd->sd = alloc_percpu(struct sched_domain *);
7344 sdd->sg = alloc_percpu(struct sched_group *);
7348 sdd->sgp = alloc_percpu(struct sched_group_power *);
7352 for_each_cpu(j, cpu_map) {
7353 struct sched_domain *sd;
7354 struct sched_group *sg;
7355 struct sched_group_power *sgp;
7357 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7358 GFP_KERNEL, cpu_to_node(j));
7362 *per_cpu_ptr(sdd->sd, j) = sd;
7364 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7365 GFP_KERNEL, cpu_to_node(j));
7369 *per_cpu_ptr(sdd->sg, j) = sg;
7371 sgp = kzalloc_node(sizeof(struct sched_group_power),
7372 GFP_KERNEL, cpu_to_node(j));
7376 *per_cpu_ptr(sdd->sgp, j) = sgp;
7383 static void __sdt_free(const struct cpumask *cpu_map)
7385 struct sched_domain_topology_level *tl;
7388 for (tl = sched_domain_topology; tl->init; tl++) {
7389 struct sd_data *sdd = &tl->data;
7391 for_each_cpu(j, cpu_map) {
7392 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
7393 if (sd && (sd->flags & SD_OVERLAP))
7394 free_sched_groups(sd->groups, 0);
7395 kfree(*per_cpu_ptr(sdd->sd, j));
7396 kfree(*per_cpu_ptr(sdd->sg, j));
7397 kfree(*per_cpu_ptr(sdd->sgp, j));
7399 free_percpu(sdd->sd);
7400 free_percpu(sdd->sg);
7401 free_percpu(sdd->sgp);
7405 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7406 struct s_data *d, const struct cpumask *cpu_map,
7407 struct sched_domain_attr *attr, struct sched_domain *child,
7410 struct sched_domain *sd = tl->init(tl, cpu);
7414 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7416 sd->level = child->level + 1;
7417 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7421 set_domain_attribute(sd, attr);
7427 * Build sched domains for a given set of cpus and attach the sched domains
7428 * to the individual cpus
7430 static int build_sched_domains(const struct cpumask *cpu_map,
7431 struct sched_domain_attr *attr)
7433 enum s_alloc alloc_state = sa_none;
7434 struct sched_domain *sd;
7436 int i, ret = -ENOMEM;
7438 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7439 if (alloc_state != sa_rootdomain)
7442 /* Set up domains for cpus specified by the cpu_map. */
7443 for_each_cpu(i, cpu_map) {
7444 struct sched_domain_topology_level *tl;
7447 for (tl = sched_domain_topology; tl->init; tl++) {
7448 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7449 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7450 sd->flags |= SD_OVERLAP;
7451 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7458 *per_cpu_ptr(d.sd, i) = sd;
7461 /* Build the groups for the domains */
7462 for_each_cpu(i, cpu_map) {
7463 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7464 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7465 if (sd->flags & SD_OVERLAP) {
7466 if (build_overlap_sched_groups(sd, i))
7469 if (build_sched_groups(sd, i))
7475 /* Calculate CPU power for physical packages and nodes */
7476 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7477 if (!cpumask_test_cpu(i, cpu_map))
7480 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7481 claim_allocations(i, sd);
7482 init_sched_groups_power(i, sd);
7486 /* Attach the domains */
7488 for_each_cpu(i, cpu_map) {
7489 sd = *per_cpu_ptr(d.sd, i);
7490 cpu_attach_domain(sd, d.rd, i);
7496 __free_domain_allocs(&d, alloc_state, cpu_map);
7500 static cpumask_var_t *doms_cur; /* current sched domains */
7501 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7502 static struct sched_domain_attr *dattr_cur;
7503 /* attribues of custom domains in 'doms_cur' */
7506 * Special case: If a kmalloc of a doms_cur partition (array of
7507 * cpumask) fails, then fallback to a single sched domain,
7508 * as determined by the single cpumask fallback_doms.
7510 static cpumask_var_t fallback_doms;
7513 * arch_update_cpu_topology lets virtualized architectures update the
7514 * cpu core maps. It is supposed to return 1 if the topology changed
7515 * or 0 if it stayed the same.
7517 int __attribute__((weak)) arch_update_cpu_topology(void)
7522 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7525 cpumask_var_t *doms;
7527 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7530 for (i = 0; i < ndoms; i++) {
7531 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7532 free_sched_domains(doms, i);
7539 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7542 for (i = 0; i < ndoms; i++)
7543 free_cpumask_var(doms[i]);
7548 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7549 * For now this just excludes isolated cpus, but could be used to
7550 * exclude other special cases in the future.
7552 static int init_sched_domains(const struct cpumask *cpu_map)
7556 arch_update_cpu_topology();
7558 doms_cur = alloc_sched_domains(ndoms_cur);
7560 doms_cur = &fallback_doms;
7561 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7563 err = build_sched_domains(doms_cur[0], NULL);
7564 register_sched_domain_sysctl();
7570 * Detach sched domains from a group of cpus specified in cpu_map
7571 * These cpus will now be attached to the NULL domain
7573 static void detach_destroy_domains(const struct cpumask *cpu_map)
7578 for_each_cpu(i, cpu_map)
7579 cpu_attach_domain(NULL, &def_root_domain, i);
7583 /* handle null as "default" */
7584 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7585 struct sched_domain_attr *new, int idx_new)
7587 struct sched_domain_attr tmp;
7594 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7595 new ? (new + idx_new) : &tmp,
7596 sizeof(struct sched_domain_attr));
7600 * Partition sched domains as specified by the 'ndoms_new'
7601 * cpumasks in the array doms_new[] of cpumasks. This compares
7602 * doms_new[] to the current sched domain partitioning, doms_cur[].
7603 * It destroys each deleted domain and builds each new domain.
7605 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7606 * The masks don't intersect (don't overlap.) We should setup one
7607 * sched domain for each mask. CPUs not in any of the cpumasks will
7608 * not be load balanced. If the same cpumask appears both in the
7609 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7612 * The passed in 'doms_new' should be allocated using
7613 * alloc_sched_domains. This routine takes ownership of it and will
7614 * free_sched_domains it when done with it. If the caller failed the
7615 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7616 * and partition_sched_domains() will fallback to the single partition
7617 * 'fallback_doms', it also forces the domains to be rebuilt.
7619 * If doms_new == NULL it will be replaced with cpu_online_mask.
7620 * ndoms_new == 0 is a special case for destroying existing domains,
7621 * and it will not create the default domain.
7623 * Call with hotplug lock held
7625 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7626 struct sched_domain_attr *dattr_new)
7631 mutex_lock(&sched_domains_mutex);
7633 /* always unregister in case we don't destroy any domains */
7634 unregister_sched_domain_sysctl();
7636 /* Let architecture update cpu core mappings. */
7637 new_topology = arch_update_cpu_topology();
7639 n = doms_new ? ndoms_new : 0;
7641 /* Destroy deleted domains */
7642 for (i = 0; i < ndoms_cur; i++) {
7643 for (j = 0; j < n && !new_topology; j++) {
7644 if (cpumask_equal(doms_cur[i], doms_new[j])
7645 && dattrs_equal(dattr_cur, i, dattr_new, j))
7648 /* no match - a current sched domain not in new doms_new[] */
7649 detach_destroy_domains(doms_cur[i]);
7654 if (doms_new == NULL) {
7656 doms_new = &fallback_doms;
7657 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7658 WARN_ON_ONCE(dattr_new);
7661 /* Build new domains */
7662 for (i = 0; i < ndoms_new; i++) {
7663 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7664 if (cpumask_equal(doms_new[i], doms_cur[j])
7665 && dattrs_equal(dattr_new, i, dattr_cur, j))
7668 /* no match - add a new doms_new */
7669 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7674 /* Remember the new sched domains */
7675 if (doms_cur != &fallback_doms)
7676 free_sched_domains(doms_cur, ndoms_cur);
7677 kfree(dattr_cur); /* kfree(NULL) is safe */
7678 doms_cur = doms_new;
7679 dattr_cur = dattr_new;
7680 ndoms_cur = ndoms_new;
7682 register_sched_domain_sysctl();
7684 mutex_unlock(&sched_domains_mutex);
7687 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7688 static void reinit_sched_domains(void)
7692 /* Destroy domains first to force the rebuild */
7693 partition_sched_domains(0, NULL, NULL);
7695 rebuild_sched_domains();
7699 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7701 unsigned int level = 0;
7703 if (sscanf(buf, "%u", &level) != 1)
7707 * level is always be positive so don't check for
7708 * level < POWERSAVINGS_BALANCE_NONE which is 0
7709 * What happens on 0 or 1 byte write,
7710 * need to check for count as well?
7713 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7717 sched_smt_power_savings = level;
7719 sched_mc_power_savings = level;
7721 reinit_sched_domains();
7726 #ifdef CONFIG_SCHED_MC
7727 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7728 struct sysdev_class_attribute *attr,
7731 return sprintf(page, "%u\n", sched_mc_power_savings);
7733 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7734 struct sysdev_class_attribute *attr,
7735 const char *buf, size_t count)
7737 return sched_power_savings_store(buf, count, 0);
7739 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7740 sched_mc_power_savings_show,
7741 sched_mc_power_savings_store);
7744 #ifdef CONFIG_SCHED_SMT
7745 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7746 struct sysdev_class_attribute *attr,
7749 return sprintf(page, "%u\n", sched_smt_power_savings);
7751 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7752 struct sysdev_class_attribute *attr,
7753 const char *buf, size_t count)
7755 return sched_power_savings_store(buf, count, 1);
7757 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7758 sched_smt_power_savings_show,
7759 sched_smt_power_savings_store);
7762 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7766 #ifdef CONFIG_SCHED_SMT
7768 err = sysfs_create_file(&cls->kset.kobj,
7769 &attr_sched_smt_power_savings.attr);
7771 #ifdef CONFIG_SCHED_MC
7772 if (!err && mc_capable())
7773 err = sysfs_create_file(&cls->kset.kobj,
7774 &attr_sched_mc_power_savings.attr);
7778 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7781 * Update cpusets according to cpu_active mask. If cpusets are
7782 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7783 * around partition_sched_domains().
7785 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7788 switch (action & ~CPU_TASKS_FROZEN) {
7790 case CPU_DOWN_FAILED:
7791 cpuset_update_active_cpus();
7798 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7801 switch (action & ~CPU_TASKS_FROZEN) {
7802 case CPU_DOWN_PREPARE:
7803 cpuset_update_active_cpus();
7810 static int update_runtime(struct notifier_block *nfb,
7811 unsigned long action, void *hcpu)
7813 int cpu = (int)(long)hcpu;
7816 case CPU_DOWN_PREPARE:
7817 case CPU_DOWN_PREPARE_FROZEN:
7818 disable_runtime(cpu_rq(cpu));
7821 case CPU_DOWN_FAILED:
7822 case CPU_DOWN_FAILED_FROZEN:
7824 case CPU_ONLINE_FROZEN:
7825 enable_runtime(cpu_rq(cpu));
7833 void __init sched_init_smp(void)
7835 cpumask_var_t non_isolated_cpus;
7837 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7838 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7841 mutex_lock(&sched_domains_mutex);
7842 init_sched_domains(cpu_active_mask);
7843 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7844 if (cpumask_empty(non_isolated_cpus))
7845 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7846 mutex_unlock(&sched_domains_mutex);
7849 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7850 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7852 /* RT runtime code needs to handle some hotplug events */
7853 hotcpu_notifier(update_runtime, 0);
7857 /* Move init over to a non-isolated CPU */
7858 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7860 sched_init_granularity();
7861 free_cpumask_var(non_isolated_cpus);
7863 init_sched_rt_class();
7866 void __init sched_init_smp(void)
7868 sched_init_granularity();
7870 #endif /* CONFIG_SMP */
7872 const_debug unsigned int sysctl_timer_migration = 1;
7874 int in_sched_functions(unsigned long addr)
7876 return in_lock_functions(addr) ||
7877 (addr >= (unsigned long)__sched_text_start
7878 && addr < (unsigned long)__sched_text_end);
7881 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7883 cfs_rq->tasks_timeline = RB_ROOT;
7884 INIT_LIST_HEAD(&cfs_rq->tasks);
7885 #ifdef CONFIG_FAIR_GROUP_SCHED
7887 /* allow initial update_cfs_load() to truncate */
7889 cfs_rq->load_stamp = 1;
7892 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7893 #ifndef CONFIG_64BIT
7894 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
7898 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7900 struct rt_prio_array *array;
7903 array = &rt_rq->active;
7904 for (i = 0; i < MAX_RT_PRIO; i++) {
7905 INIT_LIST_HEAD(array->queue + i);
7906 __clear_bit(i, array->bitmap);
7908 /* delimiter for bitsearch: */
7909 __set_bit(MAX_RT_PRIO, array->bitmap);
7911 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7912 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7914 rt_rq->highest_prio.next = MAX_RT_PRIO;
7918 rt_rq->rt_nr_migratory = 0;
7919 rt_rq->overloaded = 0;
7920 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7924 rt_rq->rt_throttled = 0;
7925 rt_rq->rt_runtime = 0;
7926 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7928 #ifdef CONFIG_RT_GROUP_SCHED
7929 rt_rq->rt_nr_boosted = 0;
7934 #ifdef CONFIG_FAIR_GROUP_SCHED
7935 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7936 struct sched_entity *se, int cpu,
7937 struct sched_entity *parent)
7939 struct rq *rq = cpu_rq(cpu);
7940 tg->cfs_rq[cpu] = cfs_rq;
7941 init_cfs_rq(cfs_rq, rq);
7945 /* se could be NULL for root_task_group */
7950 se->cfs_rq = &rq->cfs;
7952 se->cfs_rq = parent->my_q;
7955 update_load_set(&se->load, 0);
7956 se->parent = parent;
7960 #ifdef CONFIG_RT_GROUP_SCHED
7961 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7962 struct sched_rt_entity *rt_se, int cpu,
7963 struct sched_rt_entity *parent)
7965 struct rq *rq = cpu_rq(cpu);
7967 tg->rt_rq[cpu] = rt_rq;
7968 init_rt_rq(rt_rq, rq);
7970 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7972 tg->rt_se[cpu] = rt_se;
7977 rt_se->rt_rq = &rq->rt;
7979 rt_se->rt_rq = parent->my_q;
7981 rt_se->my_q = rt_rq;
7982 rt_se->parent = parent;
7983 INIT_LIST_HEAD(&rt_se->run_list);
7987 void __init sched_init(void)
7990 unsigned long alloc_size = 0, ptr;
7992 #ifdef CONFIG_FAIR_GROUP_SCHED
7993 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7995 #ifdef CONFIG_RT_GROUP_SCHED
7996 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7998 #ifdef CONFIG_CPUMASK_OFFSTACK
7999 alloc_size += num_possible_cpus() * cpumask_size();
8002 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8004 #ifdef CONFIG_FAIR_GROUP_SCHED
8005 root_task_group.se = (struct sched_entity **)ptr;
8006 ptr += nr_cpu_ids * sizeof(void **);
8008 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8009 ptr += nr_cpu_ids * sizeof(void **);
8011 #endif /* CONFIG_FAIR_GROUP_SCHED */
8012 #ifdef CONFIG_RT_GROUP_SCHED
8013 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8014 ptr += nr_cpu_ids * sizeof(void **);
8016 root_task_group.rt_rq = (struct rt_rq **)ptr;
8017 ptr += nr_cpu_ids * sizeof(void **);
8019 #endif /* CONFIG_RT_GROUP_SCHED */
8020 #ifdef CONFIG_CPUMASK_OFFSTACK
8021 for_each_possible_cpu(i) {
8022 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8023 ptr += cpumask_size();
8025 #endif /* CONFIG_CPUMASK_OFFSTACK */
8029 init_defrootdomain();
8032 init_rt_bandwidth(&def_rt_bandwidth,
8033 global_rt_period(), global_rt_runtime());
8035 #ifdef CONFIG_RT_GROUP_SCHED
8036 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8037 global_rt_period(), global_rt_runtime());
8038 #endif /* CONFIG_RT_GROUP_SCHED */
8040 #ifdef CONFIG_CGROUP_SCHED
8041 list_add(&root_task_group.list, &task_groups);
8042 INIT_LIST_HEAD(&root_task_group.children);
8043 autogroup_init(&init_task);
8044 #endif /* CONFIG_CGROUP_SCHED */
8046 for_each_possible_cpu(i) {
8050 raw_spin_lock_init(&rq->lock);
8052 rq->calc_load_active = 0;
8053 rq->calc_load_update = jiffies + LOAD_FREQ;
8054 init_cfs_rq(&rq->cfs, rq);
8055 init_rt_rq(&rq->rt, rq);
8056 #ifdef CONFIG_FAIR_GROUP_SCHED
8057 root_task_group.shares = root_task_group_load;
8058 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8060 * How much cpu bandwidth does root_task_group get?
8062 * In case of task-groups formed thr' the cgroup filesystem, it
8063 * gets 100% of the cpu resources in the system. This overall
8064 * system cpu resource is divided among the tasks of
8065 * root_task_group and its child task-groups in a fair manner,
8066 * based on each entity's (task or task-group's) weight
8067 * (se->load.weight).
8069 * In other words, if root_task_group has 10 tasks of weight
8070 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8071 * then A0's share of the cpu resource is:
8073 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8075 * We achieve this by letting root_task_group's tasks sit
8076 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8078 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8079 #endif /* CONFIG_FAIR_GROUP_SCHED */
8081 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8082 #ifdef CONFIG_RT_GROUP_SCHED
8083 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8084 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8087 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8088 rq->cpu_load[j] = 0;
8090 rq->last_load_update_tick = jiffies;
8095 rq->cpu_power = SCHED_POWER_SCALE;
8096 rq->post_schedule = 0;
8097 rq->active_balance = 0;
8098 rq->next_balance = jiffies;
8103 rq->avg_idle = 2*sysctl_sched_migration_cost;
8104 rq_attach_root(rq, &def_root_domain);
8106 rq->nohz_balance_kick = 0;
8107 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8111 atomic_set(&rq->nr_iowait, 0);
8114 set_load_weight(&init_task);
8116 #ifdef CONFIG_PREEMPT_NOTIFIERS
8117 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8121 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8124 #ifdef CONFIG_RT_MUTEXES
8125 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
8129 * The boot idle thread does lazy MMU switching as well:
8131 atomic_inc(&init_mm.mm_count);
8132 enter_lazy_tlb(&init_mm, current);
8135 * Make us the idle thread. Technically, schedule() should not be
8136 * called from this thread, however somewhere below it might be,
8137 * but because we are the idle thread, we just pick up running again
8138 * when this runqueue becomes "idle".
8140 init_idle(current, smp_processor_id());
8142 calc_load_update = jiffies + LOAD_FREQ;
8145 * During early bootup we pretend to be a normal task:
8147 current->sched_class = &fair_sched_class;
8149 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8150 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8152 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8154 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8155 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8156 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8157 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8158 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8160 /* May be allocated at isolcpus cmdline parse time */
8161 if (cpu_isolated_map == NULL)
8162 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8165 scheduler_running = 1;
8168 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8169 static inline int preempt_count_equals(int preempt_offset)
8171 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8173 return (nested == preempt_offset);
8176 void __might_sleep(const char *file, int line, int preempt_offset)
8179 static unsigned long prev_jiffy; /* ratelimiting */
8181 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8182 system_state != SYSTEM_RUNNING || oops_in_progress)
8184 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8186 prev_jiffy = jiffies;
8189 "BUG: sleeping function called from invalid context at %s:%d\n",
8192 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8193 in_atomic(), irqs_disabled(),
8194 current->pid, current->comm);
8196 debug_show_held_locks(current);
8197 if (irqs_disabled())
8198 print_irqtrace_events(current);
8202 EXPORT_SYMBOL(__might_sleep);
8205 #ifdef CONFIG_MAGIC_SYSRQ
8206 static void normalize_task(struct rq *rq, struct task_struct *p)
8208 const struct sched_class *prev_class = p->sched_class;
8209 int old_prio = p->prio;
8214 deactivate_task(rq, p, 0);
8215 __setscheduler(rq, p, SCHED_NORMAL, 0);
8217 activate_task(rq, p, 0);
8218 resched_task(rq->curr);
8221 check_class_changed(rq, p, prev_class, old_prio);
8224 void normalize_rt_tasks(void)
8226 struct task_struct *g, *p;
8227 unsigned long flags;
8230 read_lock_irqsave(&tasklist_lock, flags);
8231 do_each_thread(g, p) {
8233 * Only normalize user tasks:
8238 p->se.exec_start = 0;
8239 #ifdef CONFIG_SCHEDSTATS
8240 p->se.statistics.wait_start = 0;
8241 p->se.statistics.sleep_start = 0;
8242 p->se.statistics.block_start = 0;
8247 * Renice negative nice level userspace
8250 if (TASK_NICE(p) < 0 && p->mm)
8251 set_user_nice(p, 0);
8255 raw_spin_lock(&p->pi_lock);
8256 rq = __task_rq_lock(p);
8258 normalize_task(rq, p);
8260 __task_rq_unlock(rq);
8261 raw_spin_unlock(&p->pi_lock);
8262 } while_each_thread(g, p);
8264 read_unlock_irqrestore(&tasklist_lock, flags);
8267 #endif /* CONFIG_MAGIC_SYSRQ */
8269 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8271 * These functions are only useful for the IA64 MCA handling, or kdb.
8273 * They can only be called when the whole system has been
8274 * stopped - every CPU needs to be quiescent, and no scheduling
8275 * activity can take place. Using them for anything else would
8276 * be a serious bug, and as a result, they aren't even visible
8277 * under any other configuration.
8281 * curr_task - return the current task for a given cpu.
8282 * @cpu: the processor in question.
8284 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8286 struct task_struct *curr_task(int cpu)
8288 return cpu_curr(cpu);
8291 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8295 * set_curr_task - set the current task for a given cpu.
8296 * @cpu: the processor in question.
8297 * @p: the task pointer to set.
8299 * Description: This function must only be used when non-maskable interrupts
8300 * are serviced on a separate stack. It allows the architecture to switch the
8301 * notion of the current task on a cpu in a non-blocking manner. This function
8302 * must be called with all CPU's synchronized, and interrupts disabled, the
8303 * and caller must save the original value of the current task (see
8304 * curr_task() above) and restore that value before reenabling interrupts and
8305 * re-starting the system.
8307 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8309 void set_curr_task(int cpu, struct task_struct *p)
8316 #ifdef CONFIG_FAIR_GROUP_SCHED
8317 static void free_fair_sched_group(struct task_group *tg)
8321 for_each_possible_cpu(i) {
8323 kfree(tg->cfs_rq[i]);
8333 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8335 struct cfs_rq *cfs_rq;
8336 struct sched_entity *se;
8339 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8342 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8346 tg->shares = NICE_0_LOAD;
8348 for_each_possible_cpu(i) {
8349 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8350 GFP_KERNEL, cpu_to_node(i));
8354 se = kzalloc_node(sizeof(struct sched_entity),
8355 GFP_KERNEL, cpu_to_node(i));
8359 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8370 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8372 struct rq *rq = cpu_rq(cpu);
8373 unsigned long flags;
8376 * Only empty task groups can be destroyed; so we can speculatively
8377 * check on_list without danger of it being re-added.
8379 if (!tg->cfs_rq[cpu]->on_list)
8382 raw_spin_lock_irqsave(&rq->lock, flags);
8383 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8384 raw_spin_unlock_irqrestore(&rq->lock, flags);
8386 #else /* !CONFG_FAIR_GROUP_SCHED */
8387 static inline void free_fair_sched_group(struct task_group *tg)
8392 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8397 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8400 #endif /* CONFIG_FAIR_GROUP_SCHED */
8402 #ifdef CONFIG_RT_GROUP_SCHED
8403 static void free_rt_sched_group(struct task_group *tg)
8407 destroy_rt_bandwidth(&tg->rt_bandwidth);
8409 for_each_possible_cpu(i) {
8411 kfree(tg->rt_rq[i]);
8413 kfree(tg->rt_se[i]);
8421 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8423 struct rt_rq *rt_rq;
8424 struct sched_rt_entity *rt_se;
8427 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8430 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8434 init_rt_bandwidth(&tg->rt_bandwidth,
8435 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8437 for_each_possible_cpu(i) {
8438 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8439 GFP_KERNEL, cpu_to_node(i));
8443 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8444 GFP_KERNEL, cpu_to_node(i));
8448 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8458 #else /* !CONFIG_RT_GROUP_SCHED */
8459 static inline void free_rt_sched_group(struct task_group *tg)
8464 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8468 #endif /* CONFIG_RT_GROUP_SCHED */
8470 #ifdef CONFIG_CGROUP_SCHED
8471 static void free_sched_group(struct task_group *tg)
8473 free_fair_sched_group(tg);
8474 free_rt_sched_group(tg);
8479 /* allocate runqueue etc for a new task group */
8480 struct task_group *sched_create_group(struct task_group *parent)
8482 struct task_group *tg;
8483 unsigned long flags;
8485 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8487 return ERR_PTR(-ENOMEM);
8489 if (!alloc_fair_sched_group(tg, parent))
8492 if (!alloc_rt_sched_group(tg, parent))
8495 spin_lock_irqsave(&task_group_lock, flags);
8496 list_add_rcu(&tg->list, &task_groups);
8498 WARN_ON(!parent); /* root should already exist */
8500 tg->parent = parent;
8501 INIT_LIST_HEAD(&tg->children);
8502 list_add_rcu(&tg->siblings, &parent->children);
8503 spin_unlock_irqrestore(&task_group_lock, flags);
8508 free_sched_group(tg);
8509 return ERR_PTR(-ENOMEM);
8512 /* rcu callback to free various structures associated with a task group */
8513 static void free_sched_group_rcu(struct rcu_head *rhp)
8515 /* now it should be safe to free those cfs_rqs */
8516 free_sched_group(container_of(rhp, struct task_group, rcu));
8519 /* Destroy runqueue etc associated with a task group */
8520 void sched_destroy_group(struct task_group *tg)
8522 unsigned long flags;
8525 /* end participation in shares distribution */
8526 for_each_possible_cpu(i)
8527 unregister_fair_sched_group(tg, i);
8529 spin_lock_irqsave(&task_group_lock, flags);
8530 list_del_rcu(&tg->list);
8531 list_del_rcu(&tg->siblings);
8532 spin_unlock_irqrestore(&task_group_lock, flags);
8534 /* wait for possible concurrent references to cfs_rqs complete */
8535 call_rcu(&tg->rcu, free_sched_group_rcu);
8538 /* change task's runqueue when it moves between groups.
8539 * The caller of this function should have put the task in its new group
8540 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8541 * reflect its new group.
8543 void sched_move_task(struct task_struct *tsk)
8545 struct task_group *tg;
8547 unsigned long flags;
8550 rq = task_rq_lock(tsk, &flags);
8552 running = task_current(rq, tsk);
8556 dequeue_task(rq, tsk, 0);
8557 if (unlikely(running))
8558 tsk->sched_class->put_prev_task(rq, tsk);
8560 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
8561 lockdep_is_held(&tsk->sighand->siglock)),
8562 struct task_group, css);
8563 tg = autogroup_task_group(tsk, tg);
8564 tsk->sched_task_group = tg;
8566 #ifdef CONFIG_FAIR_GROUP_SCHED
8567 if (tsk->sched_class->task_move_group)
8568 tsk->sched_class->task_move_group(tsk, on_rq);
8571 set_task_rq(tsk, task_cpu(tsk));
8573 if (unlikely(running))
8574 tsk->sched_class->set_curr_task(rq);
8576 enqueue_task(rq, tsk, 0);
8578 task_rq_unlock(rq, tsk, &flags);
8580 #endif /* CONFIG_CGROUP_SCHED */
8582 #ifdef CONFIG_FAIR_GROUP_SCHED
8583 static DEFINE_MUTEX(shares_mutex);
8585 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8588 unsigned long flags;
8591 * We can't change the weight of the root cgroup.
8596 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8598 mutex_lock(&shares_mutex);
8599 if (tg->shares == shares)
8602 tg->shares = shares;
8603 for_each_possible_cpu(i) {
8604 struct rq *rq = cpu_rq(i);
8605 struct sched_entity *se;
8608 /* Propagate contribution to hierarchy */
8609 raw_spin_lock_irqsave(&rq->lock, flags);
8610 for_each_sched_entity(se)
8611 update_cfs_shares(group_cfs_rq(se));
8612 raw_spin_unlock_irqrestore(&rq->lock, flags);
8616 mutex_unlock(&shares_mutex);
8620 unsigned long sched_group_shares(struct task_group *tg)
8626 #ifdef CONFIG_RT_GROUP_SCHED
8628 * Ensure that the real time constraints are schedulable.
8630 static DEFINE_MUTEX(rt_constraints_mutex);
8632 static unsigned long to_ratio(u64 period, u64 runtime)
8634 if (runtime == RUNTIME_INF)
8637 return div64_u64(runtime << 20, period);
8640 /* Must be called with tasklist_lock held */
8641 static inline int tg_has_rt_tasks(struct task_group *tg)
8643 struct task_struct *g, *p;
8645 do_each_thread(g, p) {
8646 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8648 } while_each_thread(g, p);
8653 struct rt_schedulable_data {
8654 struct task_group *tg;
8659 static int tg_schedulable(struct task_group *tg, void *data)
8661 struct rt_schedulable_data *d = data;
8662 struct task_group *child;
8663 unsigned long total, sum = 0;
8664 u64 period, runtime;
8666 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8667 runtime = tg->rt_bandwidth.rt_runtime;
8670 period = d->rt_period;
8671 runtime = d->rt_runtime;
8675 * Cannot have more runtime than the period.
8677 if (runtime > period && runtime != RUNTIME_INF)
8681 * Ensure we don't starve existing RT tasks.
8683 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8686 total = to_ratio(period, runtime);
8689 * Nobody can have more than the global setting allows.
8691 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8695 * The sum of our children's runtime should not exceed our own.
8697 list_for_each_entry_rcu(child, &tg->children, siblings) {
8698 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8699 runtime = child->rt_bandwidth.rt_runtime;
8701 if (child == d->tg) {
8702 period = d->rt_period;
8703 runtime = d->rt_runtime;
8706 sum += to_ratio(period, runtime);
8715 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8717 struct rt_schedulable_data data = {
8719 .rt_period = period,
8720 .rt_runtime = runtime,
8723 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8726 static int tg_set_bandwidth(struct task_group *tg,
8727 u64 rt_period, u64 rt_runtime)
8731 mutex_lock(&rt_constraints_mutex);
8732 read_lock(&tasklist_lock);
8733 err = __rt_schedulable(tg, rt_period, rt_runtime);
8737 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8738 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8739 tg->rt_bandwidth.rt_runtime = rt_runtime;
8741 for_each_possible_cpu(i) {
8742 struct rt_rq *rt_rq = tg->rt_rq[i];
8744 raw_spin_lock(&rt_rq->rt_runtime_lock);
8745 rt_rq->rt_runtime = rt_runtime;
8746 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8748 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8750 read_unlock(&tasklist_lock);
8751 mutex_unlock(&rt_constraints_mutex);
8756 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8758 u64 rt_runtime, rt_period;
8760 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8761 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8762 if (rt_runtime_us < 0)
8763 rt_runtime = RUNTIME_INF;
8765 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8768 long sched_group_rt_runtime(struct task_group *tg)
8772 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8775 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8776 do_div(rt_runtime_us, NSEC_PER_USEC);
8777 return rt_runtime_us;
8780 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8782 u64 rt_runtime, rt_period;
8784 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8785 rt_runtime = tg->rt_bandwidth.rt_runtime;
8790 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8793 long sched_group_rt_period(struct task_group *tg)
8797 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8798 do_div(rt_period_us, NSEC_PER_USEC);
8799 return rt_period_us;
8802 static int sched_rt_global_constraints(void)
8804 u64 runtime, period;
8807 if (sysctl_sched_rt_period <= 0)
8810 runtime = global_rt_runtime();
8811 period = global_rt_period();
8814 * Sanity check on the sysctl variables.
8816 if (runtime > period && runtime != RUNTIME_INF)
8819 mutex_lock(&rt_constraints_mutex);
8820 read_lock(&tasklist_lock);
8821 ret = __rt_schedulable(NULL, 0, 0);
8822 read_unlock(&tasklist_lock);
8823 mutex_unlock(&rt_constraints_mutex);
8828 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8830 /* Don't accept realtime tasks when there is no way for them to run */
8831 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8837 #else /* !CONFIG_RT_GROUP_SCHED */
8838 static int sched_rt_global_constraints(void)
8840 unsigned long flags;
8843 if (sysctl_sched_rt_period <= 0)
8847 * There's always some RT tasks in the root group
8848 * -- migration, kstopmachine etc..
8850 if (sysctl_sched_rt_runtime == 0)
8853 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8854 for_each_possible_cpu(i) {
8855 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8857 raw_spin_lock(&rt_rq->rt_runtime_lock);
8858 rt_rq->rt_runtime = global_rt_runtime();
8859 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8861 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8865 #endif /* CONFIG_RT_GROUP_SCHED */
8867 int sched_rt_handler(struct ctl_table *table, int write,
8868 void __user *buffer, size_t *lenp,
8872 int old_period, old_runtime;
8873 static DEFINE_MUTEX(mutex);
8876 old_period = sysctl_sched_rt_period;
8877 old_runtime = sysctl_sched_rt_runtime;
8879 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8881 if (!ret && write) {
8882 ret = sched_rt_global_constraints();
8884 sysctl_sched_rt_period = old_period;
8885 sysctl_sched_rt_runtime = old_runtime;
8887 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8888 def_rt_bandwidth.rt_period =
8889 ns_to_ktime(global_rt_period());
8892 mutex_unlock(&mutex);
8897 #ifdef CONFIG_CGROUP_SCHED
8899 /* return corresponding task_group object of a cgroup */
8900 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8902 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8903 struct task_group, css);
8906 static struct cgroup_subsys_state *
8907 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8909 struct task_group *tg, *parent;
8911 if (!cgrp->parent) {
8912 /* This is early initialization for the top cgroup */
8913 return &root_task_group.css;
8916 parent = cgroup_tg(cgrp->parent);
8917 tg = sched_create_group(parent);
8919 return ERR_PTR(-ENOMEM);
8925 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8927 struct task_group *tg = cgroup_tg(cgrp);
8929 sched_destroy_group(tg);
8933 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8935 #ifdef CONFIG_RT_GROUP_SCHED
8936 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8939 /* We don't support RT-tasks being in separate groups */
8940 if (tsk->sched_class != &fair_sched_class)
8947 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8949 sched_move_task(tsk);
8953 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
8954 struct cgroup *old_cgrp, struct task_struct *task)
8957 * cgroup_exit() is called in the copy_process() failure path.
8958 * Ignore this case since the task hasn't ran yet, this avoids
8959 * trying to poke a half freed task state from generic code.
8961 if (!(task->flags & PF_EXITING))
8964 sched_move_task(task);
8967 #ifdef CONFIG_FAIR_GROUP_SCHED
8968 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8971 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
8974 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8976 struct task_group *tg = cgroup_tg(cgrp);
8978 return (u64) scale_load_down(tg->shares);
8980 #endif /* CONFIG_FAIR_GROUP_SCHED */
8982 #ifdef CONFIG_RT_GROUP_SCHED
8983 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8986 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8989 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8991 return sched_group_rt_runtime(cgroup_tg(cgrp));
8994 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8997 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9000 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9002 return sched_group_rt_period(cgroup_tg(cgrp));
9004 #endif /* CONFIG_RT_GROUP_SCHED */
9006 static struct cftype cpu_files[] = {
9007 #ifdef CONFIG_FAIR_GROUP_SCHED
9010 .read_u64 = cpu_shares_read_u64,
9011 .write_u64 = cpu_shares_write_u64,
9014 #ifdef CONFIG_RT_GROUP_SCHED
9016 .name = "rt_runtime_us",
9017 .read_s64 = cpu_rt_runtime_read,
9018 .write_s64 = cpu_rt_runtime_write,
9021 .name = "rt_period_us",
9022 .read_u64 = cpu_rt_period_read_uint,
9023 .write_u64 = cpu_rt_period_write_uint,
9028 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9030 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9033 struct cgroup_subsys cpu_cgroup_subsys = {
9035 .create = cpu_cgroup_create,
9036 .destroy = cpu_cgroup_destroy,
9037 .can_attach_task = cpu_cgroup_can_attach_task,
9038 .attach_task = cpu_cgroup_attach_task,
9039 .exit = cpu_cgroup_exit,
9040 .populate = cpu_cgroup_populate,
9041 .subsys_id = cpu_cgroup_subsys_id,
9045 #endif /* CONFIG_CGROUP_SCHED */
9047 #ifdef CONFIG_CGROUP_CPUACCT
9050 * CPU accounting code for task groups.
9052 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9053 * (balbir@in.ibm.com).
9056 /* track cpu usage of a group of tasks and its child groups */
9058 struct cgroup_subsys_state css;
9059 /* cpuusage holds pointer to a u64-type object on every cpu */
9060 u64 __percpu *cpuusage;
9061 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9062 struct cpuacct *parent;
9065 struct cgroup_subsys cpuacct_subsys;
9067 /* return cpu accounting group corresponding to this container */
9068 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9070 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9071 struct cpuacct, css);
9074 /* return cpu accounting group to which this task belongs */
9075 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9077 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9078 struct cpuacct, css);
9081 /* create a new cpu accounting group */
9082 static struct cgroup_subsys_state *cpuacct_create(
9083 struct cgroup_subsys *ss, struct cgroup *cgrp)
9085 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9091 ca->cpuusage = alloc_percpu(u64);
9095 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9096 if (percpu_counter_init(&ca->cpustat[i], 0))
9097 goto out_free_counters;
9100 ca->parent = cgroup_ca(cgrp->parent);
9106 percpu_counter_destroy(&ca->cpustat[i]);
9107 free_percpu(ca->cpuusage);
9111 return ERR_PTR(-ENOMEM);
9114 /* destroy an existing cpu accounting group */
9116 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9118 struct cpuacct *ca = cgroup_ca(cgrp);
9121 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9122 percpu_counter_destroy(&ca->cpustat[i]);
9123 free_percpu(ca->cpuusage);
9127 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9129 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9132 #ifndef CONFIG_64BIT
9134 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9136 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9138 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9146 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9148 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9150 #ifndef CONFIG_64BIT
9152 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9154 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9156 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9162 /* return total cpu usage (in nanoseconds) of a group */
9163 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9165 struct cpuacct *ca = cgroup_ca(cgrp);
9166 u64 totalcpuusage = 0;
9169 for_each_present_cpu(i)
9170 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9172 return totalcpuusage;
9175 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9178 struct cpuacct *ca = cgroup_ca(cgrp);
9187 for_each_present_cpu(i)
9188 cpuacct_cpuusage_write(ca, i, 0);
9194 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9197 struct cpuacct *ca = cgroup_ca(cgroup);
9201 for_each_present_cpu(i) {
9202 percpu = cpuacct_cpuusage_read(ca, i);
9203 seq_printf(m, "%llu ", (unsigned long long) percpu);
9205 seq_printf(m, "\n");
9209 static const char *cpuacct_stat_desc[] = {
9210 [CPUACCT_STAT_USER] = "user",
9211 [CPUACCT_STAT_SYSTEM] = "system",
9214 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9215 struct cgroup_map_cb *cb)
9217 struct cpuacct *ca = cgroup_ca(cgrp);
9220 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9221 s64 val = percpu_counter_read(&ca->cpustat[i]);
9222 val = cputime64_to_clock_t(val);
9223 cb->fill(cb, cpuacct_stat_desc[i], val);
9228 static struct cftype files[] = {
9231 .read_u64 = cpuusage_read,
9232 .write_u64 = cpuusage_write,
9235 .name = "usage_percpu",
9236 .read_seq_string = cpuacct_percpu_seq_read,
9240 .read_map = cpuacct_stats_show,
9244 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9246 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9250 * charge this task's execution time to its accounting group.
9252 * called with rq->lock held.
9254 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9259 if (unlikely(!cpuacct_subsys.active))
9262 cpu = task_cpu(tsk);
9268 for (; ca; ca = ca->parent) {
9269 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9270 *cpuusage += cputime;
9277 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9278 * in cputime_t units. As a result, cpuacct_update_stats calls
9279 * percpu_counter_add with values large enough to always overflow the
9280 * per cpu batch limit causing bad SMP scalability.
9282 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9283 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9284 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9287 #define CPUACCT_BATCH \
9288 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9290 #define CPUACCT_BATCH 0
9294 * Charge the system/user time to the task's accounting group.
9296 static void cpuacct_update_stats(struct task_struct *tsk,
9297 enum cpuacct_stat_index idx, cputime_t val)
9300 int batch = CPUACCT_BATCH;
9302 if (unlikely(!cpuacct_subsys.active))
9309 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9315 struct cgroup_subsys cpuacct_subsys = {
9317 .create = cpuacct_create,
9318 .destroy = cpuacct_destroy,
9319 .populate = cpuacct_populate,
9320 .subsys_id = cpuacct_subsys_id,
9322 #endif /* CONFIG_CGROUP_CPUACCT */