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 arch_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.)
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 struct rb_root tasks_timeline;
317 struct rb_node *rb_leftmost;
319 struct list_head tasks;
320 struct list_head *balance_iterator;
323 * 'curr' points to currently running entity on this cfs_rq.
324 * It is set to NULL otherwise (i.e when none are currently running).
326 struct sched_entity *curr, *next, *last, *skip;
328 unsigned int nr_spread_over;
330 #ifdef CONFIG_FAIR_GROUP_SCHED
331 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
334 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
335 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
336 * (like users, containers etc.)
338 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
339 * list is used during load balance.
342 struct list_head leaf_cfs_rq_list;
343 struct task_group *tg; /* group that "owns" this runqueue */
347 * the part of load.weight contributed by tasks
349 unsigned long task_weight;
352 * h_load = weight * f(tg)
354 * Where f(tg) is the recursive weight fraction assigned to
357 unsigned long h_load;
360 * Maintaining per-cpu shares distribution for group scheduling
362 * load_stamp is the last time we updated the load average
363 * load_last is the last time we updated the load average and saw load
364 * load_unacc_exec_time is currently unaccounted execution time
368 u64 load_stamp, load_last, load_unacc_exec_time;
370 unsigned long load_contribution;
375 /* Real-Time classes' related field in a runqueue: */
377 struct rt_prio_array active;
378 unsigned long rt_nr_running;
379 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
381 int curr; /* highest queued rt task prio */
383 int next; /* next highest */
388 unsigned long rt_nr_migratory;
389 unsigned long rt_nr_total;
391 struct plist_head pushable_tasks;
396 /* Nests inside the rq lock: */
397 raw_spinlock_t rt_runtime_lock;
399 #ifdef CONFIG_RT_GROUP_SCHED
400 unsigned long rt_nr_boosted;
403 struct list_head leaf_rt_rq_list;
404 struct task_group *tg;
411 * We add the notion of a root-domain which will be used to define per-domain
412 * variables. Each exclusive cpuset essentially defines an island domain by
413 * fully partitioning the member cpus from any other cpuset. Whenever a new
414 * exclusive cpuset is created, we also create and attach a new root-domain
421 cpumask_var_t online;
424 * The "RT overload" flag: it gets set if a CPU has more than
425 * one runnable RT task.
427 cpumask_var_t rto_mask;
429 struct cpupri cpupri;
433 * By default the system creates a single root-domain with all cpus as
434 * members (mimicking the global state we have today).
436 static struct root_domain def_root_domain;
438 #endif /* CONFIG_SMP */
441 * This is the main, per-CPU runqueue data structure.
443 * Locking rule: those places that want to lock multiple runqueues
444 * (such as the load balancing or the thread migration code), lock
445 * acquire operations must be ordered by ascending &runqueue.
452 * nr_running and cpu_load should be in the same cacheline because
453 * remote CPUs use both these fields when doing load calculation.
455 unsigned long nr_running;
456 #define CPU_LOAD_IDX_MAX 5
457 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
458 unsigned long last_load_update_tick;
461 unsigned char nohz_balance_kick;
463 unsigned int skip_clock_update;
465 /* capture load from *all* tasks on this cpu: */
466 struct load_weight load;
467 unsigned long nr_load_updates;
473 #ifdef CONFIG_FAIR_GROUP_SCHED
474 /* list of leaf cfs_rq on this cpu: */
475 struct list_head leaf_cfs_rq_list;
477 #ifdef CONFIG_RT_GROUP_SCHED
478 struct list_head leaf_rt_rq_list;
482 * This is part of a global counter where only the total sum
483 * over all CPUs matters. A task can increase this counter on
484 * one CPU and if it got migrated afterwards it may decrease
485 * it on another CPU. Always updated under the runqueue lock:
487 unsigned long nr_uninterruptible;
489 struct task_struct *curr, *idle, *stop;
490 unsigned long next_balance;
491 struct mm_struct *prev_mm;
499 struct root_domain *rd;
500 struct sched_domain *sd;
502 unsigned long cpu_power;
504 unsigned char idle_at_tick;
505 /* For active balancing */
509 struct cpu_stop_work active_balance_work;
510 /* cpu of this runqueue: */
514 unsigned long avg_load_per_task;
522 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
526 /* calc_load related fields */
527 unsigned long calc_load_update;
528 long calc_load_active;
530 #ifdef CONFIG_SCHED_HRTICK
532 int hrtick_csd_pending;
533 struct call_single_data hrtick_csd;
535 struct hrtimer hrtick_timer;
538 #ifdef CONFIG_SCHEDSTATS
540 struct sched_info rq_sched_info;
541 unsigned long long rq_cpu_time;
542 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
544 /* sys_sched_yield() stats */
545 unsigned int yld_count;
547 /* schedule() stats */
548 unsigned int sched_switch;
549 unsigned int sched_count;
550 unsigned int sched_goidle;
552 /* try_to_wake_up() stats */
553 unsigned int ttwu_count;
554 unsigned int ttwu_local;
558 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
561 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
563 static inline int cpu_of(struct rq *rq)
572 #define rcu_dereference_check_sched_domain(p) \
573 rcu_dereference_check((p), \
574 rcu_read_lock_sched_held() || \
575 lockdep_is_held(&sched_domains_mutex))
578 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
579 * See detach_destroy_domains: synchronize_sched for details.
581 * The domain tree of any CPU may only be accessed from within
582 * preempt-disabled sections.
584 #define for_each_domain(cpu, __sd) \
585 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
587 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
588 #define this_rq() (&__get_cpu_var(runqueues))
589 #define task_rq(p) cpu_rq(task_cpu(p))
590 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
591 #define raw_rq() (&__raw_get_cpu_var(runqueues))
593 #ifdef CONFIG_CGROUP_SCHED
596 * Return the group to which this tasks belongs.
598 * We use task_subsys_state_check() and extend the RCU verification
599 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
600 * holds that lock for each task it moves into the cgroup. Therefore
601 * by holding that lock, we pin the task to the current cgroup.
603 static inline struct task_group *task_group(struct task_struct *p)
605 struct task_group *tg;
606 struct cgroup_subsys_state *css;
608 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
609 lockdep_is_held(&task_rq(p)->lock));
610 tg = container_of(css, struct task_group, css);
612 return autogroup_task_group(p, tg);
615 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
616 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
618 #ifdef CONFIG_FAIR_GROUP_SCHED
619 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
620 p->se.parent = task_group(p)->se[cpu];
623 #ifdef CONFIG_RT_GROUP_SCHED
624 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
625 p->rt.parent = task_group(p)->rt_se[cpu];
629 #else /* CONFIG_CGROUP_SCHED */
631 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
632 static inline struct task_group *task_group(struct task_struct *p)
637 #endif /* CONFIG_CGROUP_SCHED */
639 static void update_rq_clock_task(struct rq *rq, s64 delta);
641 static void update_rq_clock(struct rq *rq)
645 if (rq->skip_clock_update)
648 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
650 update_rq_clock_task(rq, delta);
654 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
656 #ifdef CONFIG_SCHED_DEBUG
657 # define const_debug __read_mostly
659 # define const_debug static const
663 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
664 * @cpu: the processor in question.
666 * This interface allows printk to be called with the runqueue lock
667 * held and know whether or not it is OK to wake up the klogd.
669 int runqueue_is_locked(int cpu)
671 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
675 * Debugging: various feature bits
678 #define SCHED_FEAT(name, enabled) \
679 __SCHED_FEAT_##name ,
682 #include "sched_features.h"
687 #define SCHED_FEAT(name, enabled) \
688 (1UL << __SCHED_FEAT_##name) * enabled |
690 const_debug unsigned int sysctl_sched_features =
691 #include "sched_features.h"
696 #ifdef CONFIG_SCHED_DEBUG
697 #define SCHED_FEAT(name, enabled) \
700 static __read_mostly char *sched_feat_names[] = {
701 #include "sched_features.h"
707 static int sched_feat_show(struct seq_file *m, void *v)
711 for (i = 0; sched_feat_names[i]; i++) {
712 if (!(sysctl_sched_features & (1UL << i)))
714 seq_printf(m, "%s ", sched_feat_names[i]);
722 sched_feat_write(struct file *filp, const char __user *ubuf,
723 size_t cnt, loff_t *ppos)
733 if (copy_from_user(&buf, ubuf, cnt))
739 if (strncmp(cmp, "NO_", 3) == 0) {
744 for (i = 0; sched_feat_names[i]; i++) {
745 if (strcmp(cmp, sched_feat_names[i]) == 0) {
747 sysctl_sched_features &= ~(1UL << i);
749 sysctl_sched_features |= (1UL << i);
754 if (!sched_feat_names[i])
762 static int sched_feat_open(struct inode *inode, struct file *filp)
764 return single_open(filp, sched_feat_show, NULL);
767 static const struct file_operations sched_feat_fops = {
768 .open = sched_feat_open,
769 .write = sched_feat_write,
772 .release = single_release,
775 static __init int sched_init_debug(void)
777 debugfs_create_file("sched_features", 0644, NULL, NULL,
782 late_initcall(sched_init_debug);
786 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
789 * Number of tasks to iterate in a single balance run.
790 * Limited because this is done with IRQs disabled.
792 const_debug unsigned int sysctl_sched_nr_migrate = 32;
795 * period over which we average the RT time consumption, measured
800 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
803 * period over which we measure -rt task cpu usage in us.
806 unsigned int sysctl_sched_rt_period = 1000000;
808 static __read_mostly int scheduler_running;
811 * part of the period that we allow rt tasks to run in us.
814 int sysctl_sched_rt_runtime = 950000;
816 static inline u64 global_rt_period(void)
818 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
821 static inline u64 global_rt_runtime(void)
823 if (sysctl_sched_rt_runtime < 0)
826 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
829 #ifndef prepare_arch_switch
830 # define prepare_arch_switch(next) do { } while (0)
832 #ifndef finish_arch_switch
833 # define finish_arch_switch(prev) do { } while (0)
836 static inline int task_current(struct rq *rq, struct task_struct *p)
838 return rq->curr == p;
841 static inline int task_running(struct rq *rq, struct task_struct *p)
846 return task_current(rq, p);
850 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
851 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
855 * We can optimise this out completely for !SMP, because the
856 * SMP rebalancing from interrupt is the only thing that cares
863 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
867 * After ->on_cpu is cleared, the task can be moved to a different CPU.
868 * We must ensure this doesn't happen until the switch is completely
874 #ifdef CONFIG_DEBUG_SPINLOCK
875 /* this is a valid case when another task releases the spinlock */
876 rq->lock.owner = current;
879 * If we are tracking spinlock dependencies then we have to
880 * fix up the runqueue lock - which gets 'carried over' from
883 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
885 raw_spin_unlock_irq(&rq->lock);
888 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
889 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
893 * We can optimise this out completely for !SMP, because the
894 * SMP rebalancing from interrupt is the only thing that cares
899 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
900 raw_spin_unlock_irq(&rq->lock);
902 raw_spin_unlock(&rq->lock);
906 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
910 * After ->on_cpu is cleared, the task can be moved to a different CPU.
911 * We must ensure this doesn't happen until the switch is completely
917 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
921 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
924 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
927 static inline int task_is_waking(struct task_struct *p)
929 return unlikely(p->state == TASK_WAKING);
933 * __task_rq_lock - lock the runqueue a given task resides on.
934 * Must be called interrupts disabled.
936 static inline struct rq *__task_rq_lock(struct task_struct *p)
943 raw_spin_lock(&rq->lock);
944 if (likely(rq == task_rq(p)))
946 raw_spin_unlock(&rq->lock);
951 * task_rq_lock - lock the runqueue a given task resides on and disable
952 * interrupts. Note the ordering: we can safely lookup the task_rq without
953 * explicitly disabling preemption.
955 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
961 local_irq_save(*flags);
963 raw_spin_lock(&rq->lock);
964 if (likely(rq == task_rq(p)))
966 raw_spin_unlock_irqrestore(&rq->lock, *flags);
970 static void __task_rq_unlock(struct rq *rq)
973 raw_spin_unlock(&rq->lock);
976 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
979 raw_spin_unlock_irqrestore(&rq->lock, *flags);
983 * this_rq_lock - lock this runqueue and disable interrupts.
985 static struct rq *this_rq_lock(void)
992 raw_spin_lock(&rq->lock);
997 #ifdef CONFIG_SCHED_HRTICK
999 * Use HR-timers to deliver accurate preemption points.
1001 * Its all a bit involved since we cannot program an hrt while holding the
1002 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1005 * When we get rescheduled we reprogram the hrtick_timer outside of the
1011 * - enabled by features
1012 * - hrtimer is actually high res
1014 static inline int hrtick_enabled(struct rq *rq)
1016 if (!sched_feat(HRTICK))
1018 if (!cpu_active(cpu_of(rq)))
1020 return hrtimer_is_hres_active(&rq->hrtick_timer);
1023 static void hrtick_clear(struct rq *rq)
1025 if (hrtimer_active(&rq->hrtick_timer))
1026 hrtimer_cancel(&rq->hrtick_timer);
1030 * High-resolution timer tick.
1031 * Runs from hardirq context with interrupts disabled.
1033 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1035 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1037 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1039 raw_spin_lock(&rq->lock);
1040 update_rq_clock(rq);
1041 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1042 raw_spin_unlock(&rq->lock);
1044 return HRTIMER_NORESTART;
1049 * called from hardirq (IPI) context
1051 static void __hrtick_start(void *arg)
1053 struct rq *rq = arg;
1055 raw_spin_lock(&rq->lock);
1056 hrtimer_restart(&rq->hrtick_timer);
1057 rq->hrtick_csd_pending = 0;
1058 raw_spin_unlock(&rq->lock);
1062 * Called to set the hrtick timer state.
1064 * called with rq->lock held and irqs disabled
1066 static void hrtick_start(struct rq *rq, u64 delay)
1068 struct hrtimer *timer = &rq->hrtick_timer;
1069 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1071 hrtimer_set_expires(timer, time);
1073 if (rq == this_rq()) {
1074 hrtimer_restart(timer);
1075 } else if (!rq->hrtick_csd_pending) {
1076 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1077 rq->hrtick_csd_pending = 1;
1082 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1084 int cpu = (int)(long)hcpu;
1087 case CPU_UP_CANCELED:
1088 case CPU_UP_CANCELED_FROZEN:
1089 case CPU_DOWN_PREPARE:
1090 case CPU_DOWN_PREPARE_FROZEN:
1092 case CPU_DEAD_FROZEN:
1093 hrtick_clear(cpu_rq(cpu));
1100 static __init void init_hrtick(void)
1102 hotcpu_notifier(hotplug_hrtick, 0);
1106 * Called to set the hrtick timer state.
1108 * called with rq->lock held and irqs disabled
1110 static void hrtick_start(struct rq *rq, u64 delay)
1112 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1113 HRTIMER_MODE_REL_PINNED, 0);
1116 static inline void init_hrtick(void)
1119 #endif /* CONFIG_SMP */
1121 static void init_rq_hrtick(struct rq *rq)
1124 rq->hrtick_csd_pending = 0;
1126 rq->hrtick_csd.flags = 0;
1127 rq->hrtick_csd.func = __hrtick_start;
1128 rq->hrtick_csd.info = rq;
1131 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1132 rq->hrtick_timer.function = hrtick;
1134 #else /* CONFIG_SCHED_HRTICK */
1135 static inline void hrtick_clear(struct rq *rq)
1139 static inline void init_rq_hrtick(struct rq *rq)
1143 static inline void init_hrtick(void)
1146 #endif /* CONFIG_SCHED_HRTICK */
1149 * resched_task - mark a task 'to be rescheduled now'.
1151 * On UP this means the setting of the need_resched flag, on SMP it
1152 * might also involve a cross-CPU call to trigger the scheduler on
1157 #ifndef tsk_is_polling
1158 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1161 static void resched_task(struct task_struct *p)
1165 assert_raw_spin_locked(&task_rq(p)->lock);
1167 if (test_tsk_need_resched(p))
1170 set_tsk_need_resched(p);
1173 if (cpu == smp_processor_id())
1176 /* NEED_RESCHED must be visible before we test polling */
1178 if (!tsk_is_polling(p))
1179 smp_send_reschedule(cpu);
1182 static void resched_cpu(int cpu)
1184 struct rq *rq = cpu_rq(cpu);
1185 unsigned long flags;
1187 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1189 resched_task(cpu_curr(cpu));
1190 raw_spin_unlock_irqrestore(&rq->lock, flags);
1195 * In the semi idle case, use the nearest busy cpu for migrating timers
1196 * from an idle cpu. This is good for power-savings.
1198 * We don't do similar optimization for completely idle system, as
1199 * selecting an idle cpu will add more delays to the timers than intended
1200 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1202 int get_nohz_timer_target(void)
1204 int cpu = smp_processor_id();
1206 struct sched_domain *sd;
1208 for_each_domain(cpu, sd) {
1209 for_each_cpu(i, sched_domain_span(sd))
1216 * When add_timer_on() enqueues a timer into the timer wheel of an
1217 * idle CPU then this timer might expire before the next timer event
1218 * which is scheduled to wake up that CPU. In case of a completely
1219 * idle system the next event might even be infinite time into the
1220 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1221 * leaves the inner idle loop so the newly added timer is taken into
1222 * account when the CPU goes back to idle and evaluates the timer
1223 * wheel for the next timer event.
1225 void wake_up_idle_cpu(int cpu)
1227 struct rq *rq = cpu_rq(cpu);
1229 if (cpu == smp_processor_id())
1233 * This is safe, as this function is called with the timer
1234 * wheel base lock of (cpu) held. When the CPU is on the way
1235 * to idle and has not yet set rq->curr to idle then it will
1236 * be serialized on the timer wheel base lock and take the new
1237 * timer into account automatically.
1239 if (rq->curr != rq->idle)
1243 * We can set TIF_RESCHED on the idle task of the other CPU
1244 * lockless. The worst case is that the other CPU runs the
1245 * idle task through an additional NOOP schedule()
1247 set_tsk_need_resched(rq->idle);
1249 /* NEED_RESCHED must be visible before we test polling */
1251 if (!tsk_is_polling(rq->idle))
1252 smp_send_reschedule(cpu);
1255 #endif /* CONFIG_NO_HZ */
1257 static u64 sched_avg_period(void)
1259 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1262 static void sched_avg_update(struct rq *rq)
1264 s64 period = sched_avg_period();
1266 while ((s64)(rq->clock - rq->age_stamp) > period) {
1268 * Inline assembly required to prevent the compiler
1269 * optimising this loop into a divmod call.
1270 * See __iter_div_u64_rem() for another example of this.
1272 asm("" : "+rm" (rq->age_stamp));
1273 rq->age_stamp += period;
1278 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1280 rq->rt_avg += rt_delta;
1281 sched_avg_update(rq);
1284 #else /* !CONFIG_SMP */
1285 static void resched_task(struct task_struct *p)
1287 assert_raw_spin_locked(&task_rq(p)->lock);
1288 set_tsk_need_resched(p);
1291 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1295 static void sched_avg_update(struct rq *rq)
1298 #endif /* CONFIG_SMP */
1300 #if BITS_PER_LONG == 32
1301 # define WMULT_CONST (~0UL)
1303 # define WMULT_CONST (1UL << 32)
1306 #define WMULT_SHIFT 32
1309 * Shift right and round:
1311 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1314 * delta *= weight / lw
1316 static unsigned long
1317 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1318 struct load_weight *lw)
1322 if (!lw->inv_weight) {
1323 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1326 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1330 tmp = (u64)delta_exec * weight;
1332 * Check whether we'd overflow the 64-bit multiplication:
1334 if (unlikely(tmp > WMULT_CONST))
1335 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1338 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1340 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1343 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1349 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1355 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1362 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1363 * of tasks with abnormal "nice" values across CPUs the contribution that
1364 * each task makes to its run queue's load is weighted according to its
1365 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1366 * scaled version of the new time slice allocation that they receive on time
1370 #define WEIGHT_IDLEPRIO 3
1371 #define WMULT_IDLEPRIO 1431655765
1374 * Nice levels are multiplicative, with a gentle 10% change for every
1375 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1376 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1377 * that remained on nice 0.
1379 * The "10% effect" is relative and cumulative: from _any_ nice level,
1380 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1381 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1382 * If a task goes up by ~10% and another task goes down by ~10% then
1383 * the relative distance between them is ~25%.)
1385 static const int prio_to_weight[40] = {
1386 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1387 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1388 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1389 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1390 /* 0 */ 1024, 820, 655, 526, 423,
1391 /* 5 */ 335, 272, 215, 172, 137,
1392 /* 10 */ 110, 87, 70, 56, 45,
1393 /* 15 */ 36, 29, 23, 18, 15,
1397 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1399 * In cases where the weight does not change often, we can use the
1400 * precalculated inverse to speed up arithmetics by turning divisions
1401 * into multiplications:
1403 static const u32 prio_to_wmult[40] = {
1404 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1405 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1406 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1407 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1408 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1409 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1410 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1411 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1414 /* Time spent by the tasks of the cpu accounting group executing in ... */
1415 enum cpuacct_stat_index {
1416 CPUACCT_STAT_USER, /* ... user mode */
1417 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1419 CPUACCT_STAT_NSTATS,
1422 #ifdef CONFIG_CGROUP_CPUACCT
1423 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1424 static void cpuacct_update_stats(struct task_struct *tsk,
1425 enum cpuacct_stat_index idx, cputime_t val);
1427 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1428 static inline void cpuacct_update_stats(struct task_struct *tsk,
1429 enum cpuacct_stat_index idx, cputime_t val) {}
1432 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1434 update_load_add(&rq->load, load);
1437 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1439 update_load_sub(&rq->load, load);
1442 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1443 typedef int (*tg_visitor)(struct task_group *, void *);
1446 * Iterate the full tree, calling @down when first entering a node and @up when
1447 * leaving it for the final time.
1449 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1451 struct task_group *parent, *child;
1455 parent = &root_task_group;
1457 ret = (*down)(parent, data);
1460 list_for_each_entry_rcu(child, &parent->children, siblings) {
1467 ret = (*up)(parent, data);
1472 parent = parent->parent;
1481 static int tg_nop(struct task_group *tg, void *data)
1488 /* Used instead of source_load when we know the type == 0 */
1489 static unsigned long weighted_cpuload(const int cpu)
1491 return cpu_rq(cpu)->load.weight;
1495 * Return a low guess at the load of a migration-source cpu weighted
1496 * according to the scheduling class and "nice" value.
1498 * We want to under-estimate the load of migration sources, to
1499 * balance conservatively.
1501 static unsigned long source_load(int cpu, int type)
1503 struct rq *rq = cpu_rq(cpu);
1504 unsigned long total = weighted_cpuload(cpu);
1506 if (type == 0 || !sched_feat(LB_BIAS))
1509 return min(rq->cpu_load[type-1], total);
1513 * Return a high guess at the load of a migration-target cpu weighted
1514 * according to the scheduling class and "nice" value.
1516 static unsigned long target_load(int cpu, int type)
1518 struct rq *rq = cpu_rq(cpu);
1519 unsigned long total = weighted_cpuload(cpu);
1521 if (type == 0 || !sched_feat(LB_BIAS))
1524 return max(rq->cpu_load[type-1], total);
1527 static unsigned long power_of(int cpu)
1529 return cpu_rq(cpu)->cpu_power;
1532 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1534 static unsigned long cpu_avg_load_per_task(int cpu)
1536 struct rq *rq = cpu_rq(cpu);
1537 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1540 rq->avg_load_per_task = rq->load.weight / nr_running;
1542 rq->avg_load_per_task = 0;
1544 return rq->avg_load_per_task;
1547 #ifdef CONFIG_FAIR_GROUP_SCHED
1550 * Compute the cpu's hierarchical load factor for each task group.
1551 * This needs to be done in a top-down fashion because the load of a child
1552 * group is a fraction of its parents load.
1554 static int tg_load_down(struct task_group *tg, void *data)
1557 long cpu = (long)data;
1560 load = cpu_rq(cpu)->load.weight;
1562 load = tg->parent->cfs_rq[cpu]->h_load;
1563 load *= tg->se[cpu]->load.weight;
1564 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1567 tg->cfs_rq[cpu]->h_load = load;
1572 static void update_h_load(long cpu)
1574 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1579 #ifdef CONFIG_PREEMPT
1581 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1584 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1585 * way at the expense of forcing extra atomic operations in all
1586 * invocations. This assures that the double_lock is acquired using the
1587 * same underlying policy as the spinlock_t on this architecture, which
1588 * reduces latency compared to the unfair variant below. However, it
1589 * also adds more overhead and therefore may reduce throughput.
1591 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1592 __releases(this_rq->lock)
1593 __acquires(busiest->lock)
1594 __acquires(this_rq->lock)
1596 raw_spin_unlock(&this_rq->lock);
1597 double_rq_lock(this_rq, busiest);
1604 * Unfair double_lock_balance: Optimizes throughput at the expense of
1605 * latency by eliminating extra atomic operations when the locks are
1606 * already in proper order on entry. This favors lower cpu-ids and will
1607 * grant the double lock to lower cpus over higher ids under contention,
1608 * regardless of entry order into the function.
1610 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1611 __releases(this_rq->lock)
1612 __acquires(busiest->lock)
1613 __acquires(this_rq->lock)
1617 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1618 if (busiest < this_rq) {
1619 raw_spin_unlock(&this_rq->lock);
1620 raw_spin_lock(&busiest->lock);
1621 raw_spin_lock_nested(&this_rq->lock,
1622 SINGLE_DEPTH_NESTING);
1625 raw_spin_lock_nested(&busiest->lock,
1626 SINGLE_DEPTH_NESTING);
1631 #endif /* CONFIG_PREEMPT */
1634 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1636 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1638 if (unlikely(!irqs_disabled())) {
1639 /* printk() doesn't work good under rq->lock */
1640 raw_spin_unlock(&this_rq->lock);
1644 return _double_lock_balance(this_rq, busiest);
1647 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1648 __releases(busiest->lock)
1650 raw_spin_unlock(&busiest->lock);
1651 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1655 * double_rq_lock - safely lock two runqueues
1657 * Note this does not disable interrupts like task_rq_lock,
1658 * you need to do so manually before calling.
1660 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1661 __acquires(rq1->lock)
1662 __acquires(rq2->lock)
1664 BUG_ON(!irqs_disabled());
1666 raw_spin_lock(&rq1->lock);
1667 __acquire(rq2->lock); /* Fake it out ;) */
1670 raw_spin_lock(&rq1->lock);
1671 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1673 raw_spin_lock(&rq2->lock);
1674 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1680 * double_rq_unlock - safely unlock two runqueues
1682 * Note this does not restore interrupts like task_rq_unlock,
1683 * you need to do so manually after calling.
1685 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1686 __releases(rq1->lock)
1687 __releases(rq2->lock)
1689 raw_spin_unlock(&rq1->lock);
1691 raw_spin_unlock(&rq2->lock);
1693 __release(rq2->lock);
1696 #else /* CONFIG_SMP */
1699 * double_rq_lock - safely lock two runqueues
1701 * Note this does not disable interrupts like task_rq_lock,
1702 * you need to do so manually before calling.
1704 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1705 __acquires(rq1->lock)
1706 __acquires(rq2->lock)
1708 BUG_ON(!irqs_disabled());
1710 raw_spin_lock(&rq1->lock);
1711 __acquire(rq2->lock); /* Fake it out ;) */
1715 * double_rq_unlock - safely unlock two runqueues
1717 * Note this does not restore interrupts like task_rq_unlock,
1718 * you need to do so manually after calling.
1720 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1721 __releases(rq1->lock)
1722 __releases(rq2->lock)
1725 raw_spin_unlock(&rq1->lock);
1726 __release(rq2->lock);
1731 static void calc_load_account_idle(struct rq *this_rq);
1732 static void update_sysctl(void);
1733 static int get_update_sysctl_factor(void);
1734 static void update_cpu_load(struct rq *this_rq);
1736 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1738 set_task_rq(p, cpu);
1741 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1742 * successfuly executed on another CPU. We must ensure that updates of
1743 * per-task data have been completed by this moment.
1746 task_thread_info(p)->cpu = cpu;
1750 static const struct sched_class rt_sched_class;
1752 #define sched_class_highest (&stop_sched_class)
1753 #define for_each_class(class) \
1754 for (class = sched_class_highest; class; class = class->next)
1756 #include "sched_stats.h"
1758 static void inc_nr_running(struct rq *rq)
1763 static void dec_nr_running(struct rq *rq)
1768 static void set_load_weight(struct task_struct *p)
1771 * SCHED_IDLE tasks get minimal weight:
1773 if (p->policy == SCHED_IDLE) {
1774 p->se.load.weight = WEIGHT_IDLEPRIO;
1775 p->se.load.inv_weight = WMULT_IDLEPRIO;
1779 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1780 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1783 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1785 update_rq_clock(rq);
1786 sched_info_queued(p);
1787 p->sched_class->enqueue_task(rq, p, flags);
1791 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1793 update_rq_clock(rq);
1794 sched_info_dequeued(p);
1795 p->sched_class->dequeue_task(rq, p, flags);
1800 * activate_task - move a task to the runqueue.
1802 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1804 if (task_contributes_to_load(p))
1805 rq->nr_uninterruptible--;
1807 enqueue_task(rq, p, flags);
1812 * deactivate_task - remove a task from the runqueue.
1814 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1816 if (task_contributes_to_load(p))
1817 rq->nr_uninterruptible++;
1819 dequeue_task(rq, p, flags);
1823 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1826 * There are no locks covering percpu hardirq/softirq time.
1827 * They are only modified in account_system_vtime, on corresponding CPU
1828 * with interrupts disabled. So, writes are safe.
1829 * They are read and saved off onto struct rq in update_rq_clock().
1830 * This may result in other CPU reading this CPU's irq time and can
1831 * race with irq/account_system_vtime on this CPU. We would either get old
1832 * or new value with a side effect of accounting a slice of irq time to wrong
1833 * task when irq is in progress while we read rq->clock. That is a worthy
1834 * compromise in place of having locks on each irq in account_system_time.
1836 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1837 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1839 static DEFINE_PER_CPU(u64, irq_start_time);
1840 static int sched_clock_irqtime;
1842 void enable_sched_clock_irqtime(void)
1844 sched_clock_irqtime = 1;
1847 void disable_sched_clock_irqtime(void)
1849 sched_clock_irqtime = 0;
1852 #ifndef CONFIG_64BIT
1853 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1855 static inline void irq_time_write_begin(void)
1857 __this_cpu_inc(irq_time_seq.sequence);
1861 static inline void irq_time_write_end(void)
1864 __this_cpu_inc(irq_time_seq.sequence);
1867 static inline u64 irq_time_read(int cpu)
1873 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1874 irq_time = per_cpu(cpu_softirq_time, cpu) +
1875 per_cpu(cpu_hardirq_time, cpu);
1876 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1880 #else /* CONFIG_64BIT */
1881 static inline void irq_time_write_begin(void)
1885 static inline void irq_time_write_end(void)
1889 static inline u64 irq_time_read(int cpu)
1891 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1893 #endif /* CONFIG_64BIT */
1896 * Called before incrementing preempt_count on {soft,}irq_enter
1897 * and before decrementing preempt_count on {soft,}irq_exit.
1899 void account_system_vtime(struct task_struct *curr)
1901 unsigned long flags;
1905 if (!sched_clock_irqtime)
1908 local_irq_save(flags);
1910 cpu = smp_processor_id();
1911 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1912 __this_cpu_add(irq_start_time, delta);
1914 irq_time_write_begin();
1916 * We do not account for softirq time from ksoftirqd here.
1917 * We want to continue accounting softirq time to ksoftirqd thread
1918 * in that case, so as not to confuse scheduler with a special task
1919 * that do not consume any time, but still wants to run.
1921 if (hardirq_count())
1922 __this_cpu_add(cpu_hardirq_time, delta);
1923 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
1924 __this_cpu_add(cpu_softirq_time, delta);
1926 irq_time_write_end();
1927 local_irq_restore(flags);
1929 EXPORT_SYMBOL_GPL(account_system_vtime);
1931 static void update_rq_clock_task(struct rq *rq, s64 delta)
1935 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
1938 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1939 * this case when a previous update_rq_clock() happened inside a
1940 * {soft,}irq region.
1942 * When this happens, we stop ->clock_task and only update the
1943 * prev_irq_time stamp to account for the part that fit, so that a next
1944 * update will consume the rest. This ensures ->clock_task is
1947 * It does however cause some slight miss-attribution of {soft,}irq
1948 * time, a more accurate solution would be to update the irq_time using
1949 * the current rq->clock timestamp, except that would require using
1952 if (irq_delta > delta)
1955 rq->prev_irq_time += irq_delta;
1957 rq->clock_task += delta;
1959 if (irq_delta && sched_feat(NONIRQ_POWER))
1960 sched_rt_avg_update(rq, irq_delta);
1963 static int irqtime_account_hi_update(void)
1965 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
1966 unsigned long flags;
1970 local_irq_save(flags);
1971 latest_ns = this_cpu_read(cpu_hardirq_time);
1972 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
1974 local_irq_restore(flags);
1978 static int irqtime_account_si_update(void)
1980 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
1981 unsigned long flags;
1985 local_irq_save(flags);
1986 latest_ns = this_cpu_read(cpu_softirq_time);
1987 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
1989 local_irq_restore(flags);
1993 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
1995 #define sched_clock_irqtime (0)
1997 static void update_rq_clock_task(struct rq *rq, s64 delta)
1999 rq->clock_task += delta;
2002 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2004 #include "sched_idletask.c"
2005 #include "sched_fair.c"
2006 #include "sched_rt.c"
2007 #include "sched_autogroup.c"
2008 #include "sched_stoptask.c"
2009 #ifdef CONFIG_SCHED_DEBUG
2010 # include "sched_debug.c"
2013 void sched_set_stop_task(int cpu, struct task_struct *stop)
2015 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2016 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2020 * Make it appear like a SCHED_FIFO task, its something
2021 * userspace knows about and won't get confused about.
2023 * Also, it will make PI more or less work without too
2024 * much confusion -- but then, stop work should not
2025 * rely on PI working anyway.
2027 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2029 stop->sched_class = &stop_sched_class;
2032 cpu_rq(cpu)->stop = stop;
2036 * Reset it back to a normal scheduling class so that
2037 * it can die in pieces.
2039 old_stop->sched_class = &rt_sched_class;
2044 * __normal_prio - return the priority that is based on the static prio
2046 static inline int __normal_prio(struct task_struct *p)
2048 return p->static_prio;
2052 * Calculate the expected normal priority: i.e. priority
2053 * without taking RT-inheritance into account. Might be
2054 * boosted by interactivity modifiers. Changes upon fork,
2055 * setprio syscalls, and whenever the interactivity
2056 * estimator recalculates.
2058 static inline int normal_prio(struct task_struct *p)
2062 if (task_has_rt_policy(p))
2063 prio = MAX_RT_PRIO-1 - p->rt_priority;
2065 prio = __normal_prio(p);
2070 * Calculate the current priority, i.e. the priority
2071 * taken into account by the scheduler. This value might
2072 * be boosted by RT tasks, or might be boosted by
2073 * interactivity modifiers. Will be RT if the task got
2074 * RT-boosted. If not then it returns p->normal_prio.
2076 static int effective_prio(struct task_struct *p)
2078 p->normal_prio = normal_prio(p);
2080 * If we are RT tasks or we were boosted to RT priority,
2081 * keep the priority unchanged. Otherwise, update priority
2082 * to the normal priority:
2084 if (!rt_prio(p->prio))
2085 return p->normal_prio;
2090 * task_curr - is this task currently executing on a CPU?
2091 * @p: the task in question.
2093 inline int task_curr(const struct task_struct *p)
2095 return cpu_curr(task_cpu(p)) == p;
2098 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2099 const struct sched_class *prev_class,
2102 if (prev_class != p->sched_class) {
2103 if (prev_class->switched_from)
2104 prev_class->switched_from(rq, p);
2105 p->sched_class->switched_to(rq, p);
2106 } else if (oldprio != p->prio)
2107 p->sched_class->prio_changed(rq, p, oldprio);
2110 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2112 const struct sched_class *class;
2114 if (p->sched_class == rq->curr->sched_class) {
2115 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2117 for_each_class(class) {
2118 if (class == rq->curr->sched_class)
2120 if (class == p->sched_class) {
2121 resched_task(rq->curr);
2128 * A queue event has occurred, and we're going to schedule. In
2129 * this case, we can save a useless back to back clock update.
2131 if (rq->curr->se.on_rq && test_tsk_need_resched(rq->curr))
2132 rq->skip_clock_update = 1;
2137 * Is this task likely cache-hot:
2140 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2144 if (p->sched_class != &fair_sched_class)
2147 if (unlikely(p->policy == SCHED_IDLE))
2151 * Buddy candidates are cache hot:
2153 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2154 (&p->se == cfs_rq_of(&p->se)->next ||
2155 &p->se == cfs_rq_of(&p->se)->last))
2158 if (sysctl_sched_migration_cost == -1)
2160 if (sysctl_sched_migration_cost == 0)
2163 delta = now - p->se.exec_start;
2165 return delta < (s64)sysctl_sched_migration_cost;
2168 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2170 #ifdef CONFIG_SCHED_DEBUG
2172 * We should never call set_task_cpu() on a blocked task,
2173 * ttwu() will sort out the placement.
2175 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2176 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2179 trace_sched_migrate_task(p, new_cpu);
2181 if (task_cpu(p) != new_cpu) {
2182 p->se.nr_migrations++;
2183 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2186 __set_task_cpu(p, new_cpu);
2189 struct migration_arg {
2190 struct task_struct *task;
2194 static int migration_cpu_stop(void *data);
2197 * The task's runqueue lock must be held.
2198 * Returns true if you have to wait for migration thread.
2200 static bool migrate_task(struct task_struct *p, struct rq *rq)
2203 * If the task is not on a runqueue (and not running), then
2204 * the next wake-up will properly place the task.
2206 return p->se.on_rq || task_running(rq, p);
2210 * wait_task_inactive - wait for a thread to unschedule.
2212 * If @match_state is nonzero, it's the @p->state value just checked and
2213 * not expected to change. If it changes, i.e. @p might have woken up,
2214 * then return zero. When we succeed in waiting for @p to be off its CPU,
2215 * we return a positive number (its total switch count). If a second call
2216 * a short while later returns the same number, the caller can be sure that
2217 * @p has remained unscheduled the whole time.
2219 * The caller must ensure that the task *will* unschedule sometime soon,
2220 * else this function might spin for a *long* time. This function can't
2221 * be called with interrupts off, or it may introduce deadlock with
2222 * smp_call_function() if an IPI is sent by the same process we are
2223 * waiting to become inactive.
2225 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2227 unsigned long flags;
2234 * We do the initial early heuristics without holding
2235 * any task-queue locks at all. We'll only try to get
2236 * the runqueue lock when things look like they will
2242 * If the task is actively running on another CPU
2243 * still, just relax and busy-wait without holding
2246 * NOTE! Since we don't hold any locks, it's not
2247 * even sure that "rq" stays as the right runqueue!
2248 * But we don't care, since "task_running()" will
2249 * return false if the runqueue has changed and p
2250 * is actually now running somewhere else!
2252 while (task_running(rq, p)) {
2253 if (match_state && unlikely(p->state != match_state))
2259 * Ok, time to look more closely! We need the rq
2260 * lock now, to be *sure*. If we're wrong, we'll
2261 * just go back and repeat.
2263 rq = task_rq_lock(p, &flags);
2264 trace_sched_wait_task(p);
2265 running = task_running(rq, p);
2266 on_rq = p->se.on_rq;
2268 if (!match_state || p->state == match_state)
2269 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2270 task_rq_unlock(rq, &flags);
2273 * If it changed from the expected state, bail out now.
2275 if (unlikely(!ncsw))
2279 * Was it really running after all now that we
2280 * checked with the proper locks actually held?
2282 * Oops. Go back and try again..
2284 if (unlikely(running)) {
2290 * It's not enough that it's not actively running,
2291 * it must be off the runqueue _entirely_, and not
2294 * So if it was still runnable (but just not actively
2295 * running right now), it's preempted, and we should
2296 * yield - it could be a while.
2298 if (unlikely(on_rq)) {
2299 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2301 set_current_state(TASK_UNINTERRUPTIBLE);
2302 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2307 * Ahh, all good. It wasn't running, and it wasn't
2308 * runnable, which means that it will never become
2309 * running in the future either. We're all done!
2318 * kick_process - kick a running thread to enter/exit the kernel
2319 * @p: the to-be-kicked thread
2321 * Cause a process which is running on another CPU to enter
2322 * kernel-mode, without any delay. (to get signals handled.)
2324 * NOTE: this function doesn't have to take the runqueue lock,
2325 * because all it wants to ensure is that the remote task enters
2326 * the kernel. If the IPI races and the task has been migrated
2327 * to another CPU then no harm is done and the purpose has been
2330 void kick_process(struct task_struct *p)
2336 if ((cpu != smp_processor_id()) && task_curr(p))
2337 smp_send_reschedule(cpu);
2340 EXPORT_SYMBOL_GPL(kick_process);
2341 #endif /* CONFIG_SMP */
2345 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2347 static int select_fallback_rq(int cpu, struct task_struct *p)
2350 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2352 /* Look for allowed, online CPU in same node. */
2353 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2354 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2357 /* Any allowed, online CPU? */
2358 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2359 if (dest_cpu < nr_cpu_ids)
2362 /* No more Mr. Nice Guy. */
2363 dest_cpu = cpuset_cpus_allowed_fallback(p);
2365 * Don't tell them about moving exiting tasks or
2366 * kernel threads (both mm NULL), since they never
2369 if (p->mm && printk_ratelimit()) {
2370 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2371 task_pid_nr(p), p->comm, cpu);
2378 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2381 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2383 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2386 * In order not to call set_task_cpu() on a blocking task we need
2387 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2390 * Since this is common to all placement strategies, this lives here.
2392 * [ this allows ->select_task() to simply return task_cpu(p) and
2393 * not worry about this generic constraint ]
2395 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2397 cpu = select_fallback_rq(task_cpu(p), p);
2402 static void update_avg(u64 *avg, u64 sample)
2404 s64 diff = sample - *avg;
2409 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2410 bool is_sync, bool is_migrate, bool is_local,
2411 unsigned long en_flags)
2413 schedstat_inc(p, se.statistics.nr_wakeups);
2415 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2417 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2419 schedstat_inc(p, se.statistics.nr_wakeups_local);
2421 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2423 activate_task(rq, p, en_flags);
2426 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2427 int wake_flags, bool success)
2429 trace_sched_wakeup(p, success);
2430 check_preempt_curr(rq, p, wake_flags);
2432 p->state = TASK_RUNNING;
2434 if (p->sched_class->task_woken)
2435 p->sched_class->task_woken(rq, p);
2437 if (unlikely(rq->idle_stamp)) {
2438 u64 delta = rq->clock - rq->idle_stamp;
2439 u64 max = 2*sysctl_sched_migration_cost;
2444 update_avg(&rq->avg_idle, delta);
2448 /* if a worker is waking up, notify workqueue */
2449 if ((p->flags & PF_WQ_WORKER) && success)
2450 wq_worker_waking_up(p, cpu_of(rq));
2454 * try_to_wake_up - wake up a thread
2455 * @p: the thread to be awakened
2456 * @state: the mask of task states that can be woken
2457 * @wake_flags: wake modifier flags (WF_*)
2459 * Put it on the run-queue if it's not already there. The "current"
2460 * thread is always on the run-queue (except when the actual
2461 * re-schedule is in progress), and as such you're allowed to do
2462 * the simpler "current->state = TASK_RUNNING" to mark yourself
2463 * runnable without the overhead of this.
2465 * Returns %true if @p was woken up, %false if it was already running
2466 * or @state didn't match @p's state.
2468 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2471 int cpu, orig_cpu, this_cpu, success = 0;
2472 unsigned long flags;
2473 unsigned long en_flags = ENQUEUE_WAKEUP;
2476 this_cpu = get_cpu();
2479 rq = task_rq_lock(p, &flags);
2480 if (!(p->state & state))
2490 if (unlikely(task_running(rq, p)))
2494 * In order to handle concurrent wakeups and release the rq->lock
2495 * we put the task in TASK_WAKING state.
2497 * First fix up the nr_uninterruptible count:
2499 if (task_contributes_to_load(p)) {
2500 if (likely(cpu_online(orig_cpu)))
2501 rq->nr_uninterruptible--;
2503 this_rq()->nr_uninterruptible--;
2505 p->state = TASK_WAKING;
2507 if (p->sched_class->task_waking) {
2508 p->sched_class->task_waking(rq, p);
2509 en_flags |= ENQUEUE_WAKING;
2512 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2513 if (cpu != orig_cpu)
2514 set_task_cpu(p, cpu);
2515 __task_rq_unlock(rq);
2518 raw_spin_lock(&rq->lock);
2521 * We migrated the task without holding either rq->lock, however
2522 * since the task is not on the task list itself, nobody else
2523 * will try and migrate the task, hence the rq should match the
2524 * cpu we just moved it to.
2526 WARN_ON(task_cpu(p) != cpu);
2527 WARN_ON(p->state != TASK_WAKING);
2529 #ifdef CONFIG_SCHEDSTATS
2530 schedstat_inc(rq, ttwu_count);
2531 if (cpu == this_cpu)
2532 schedstat_inc(rq, ttwu_local);
2534 struct sched_domain *sd;
2535 for_each_domain(this_cpu, sd) {
2536 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2537 schedstat_inc(sd, ttwu_wake_remote);
2542 #endif /* CONFIG_SCHEDSTATS */
2545 #endif /* CONFIG_SMP */
2546 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2547 cpu == this_cpu, en_flags);
2550 ttwu_post_activation(p, rq, wake_flags, success);
2552 task_rq_unlock(rq, &flags);
2559 * try_to_wake_up_local - try to wake up a local task with rq lock held
2560 * @p: the thread to be awakened
2562 * Put @p on the run-queue if it's not already there. The caller must
2563 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2564 * the current task. this_rq() stays locked over invocation.
2566 static void try_to_wake_up_local(struct task_struct *p)
2568 struct rq *rq = task_rq(p);
2569 bool success = false;
2571 BUG_ON(rq != this_rq());
2572 BUG_ON(p == current);
2573 lockdep_assert_held(&rq->lock);
2575 if (!(p->state & TASK_NORMAL))
2579 if (likely(!task_running(rq, p))) {
2580 schedstat_inc(rq, ttwu_count);
2581 schedstat_inc(rq, ttwu_local);
2583 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2586 ttwu_post_activation(p, rq, 0, success);
2590 * wake_up_process - Wake up a specific process
2591 * @p: The process to be woken up.
2593 * Attempt to wake up the nominated process and move it to the set of runnable
2594 * processes. Returns 1 if the process was woken up, 0 if it was already
2597 * It may be assumed that this function implies a write memory barrier before
2598 * changing the task state if and only if any tasks are woken up.
2600 int wake_up_process(struct task_struct *p)
2602 return try_to_wake_up(p, TASK_ALL, 0);
2604 EXPORT_SYMBOL(wake_up_process);
2606 int wake_up_state(struct task_struct *p, unsigned int state)
2608 return try_to_wake_up(p, state, 0);
2612 * Perform scheduler related setup for a newly forked process p.
2613 * p is forked by current.
2615 * __sched_fork() is basic setup used by init_idle() too:
2617 static void __sched_fork(struct task_struct *p)
2619 p->se.exec_start = 0;
2620 p->se.sum_exec_runtime = 0;
2621 p->se.prev_sum_exec_runtime = 0;
2622 p->se.nr_migrations = 0;
2625 #ifdef CONFIG_SCHEDSTATS
2626 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2629 INIT_LIST_HEAD(&p->rt.run_list);
2631 INIT_LIST_HEAD(&p->se.group_node);
2633 #ifdef CONFIG_PREEMPT_NOTIFIERS
2634 INIT_HLIST_HEAD(&p->preempt_notifiers);
2639 * fork()/clone()-time setup:
2641 void sched_fork(struct task_struct *p, int clone_flags)
2643 int cpu = get_cpu();
2647 * We mark the process as running here. This guarantees that
2648 * nobody will actually run it, and a signal or other external
2649 * event cannot wake it up and insert it on the runqueue either.
2651 p->state = TASK_RUNNING;
2654 * Revert to default priority/policy on fork if requested.
2656 if (unlikely(p->sched_reset_on_fork)) {
2657 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2658 p->policy = SCHED_NORMAL;
2659 p->normal_prio = p->static_prio;
2662 if (PRIO_TO_NICE(p->static_prio) < 0) {
2663 p->static_prio = NICE_TO_PRIO(0);
2664 p->normal_prio = p->static_prio;
2669 * We don't need the reset flag anymore after the fork. It has
2670 * fulfilled its duty:
2672 p->sched_reset_on_fork = 0;
2676 * Make sure we do not leak PI boosting priority to the child.
2678 p->prio = current->normal_prio;
2680 if (!rt_prio(p->prio))
2681 p->sched_class = &fair_sched_class;
2683 if (p->sched_class->task_fork)
2684 p->sched_class->task_fork(p);
2687 * The child is not yet in the pid-hash so no cgroup attach races,
2688 * and the cgroup is pinned to this child due to cgroup_fork()
2689 * is ran before sched_fork().
2691 * Silence PROVE_RCU.
2694 set_task_cpu(p, cpu);
2697 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2698 if (likely(sched_info_on()))
2699 memset(&p->sched_info, 0, sizeof(p->sched_info));
2701 #if defined(CONFIG_SMP)
2704 #ifdef CONFIG_PREEMPT
2705 /* Want to start with kernel preemption disabled. */
2706 task_thread_info(p)->preempt_count = 1;
2709 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2716 * wake_up_new_task - wake up a newly created task for the first time.
2718 * This function will do some initial scheduler statistics housekeeping
2719 * that must be done for every newly created context, then puts the task
2720 * on the runqueue and wakes it.
2722 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2724 unsigned long flags;
2726 int cpu __maybe_unused = get_cpu();
2729 rq = task_rq_lock(p, &flags);
2730 p->state = TASK_WAKING;
2733 * Fork balancing, do it here and not earlier because:
2734 * - cpus_allowed can change in the fork path
2735 * - any previously selected cpu might disappear through hotplug
2737 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2738 * without people poking at ->cpus_allowed.
2740 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2741 set_task_cpu(p, cpu);
2743 p->state = TASK_RUNNING;
2744 task_rq_unlock(rq, &flags);
2747 rq = task_rq_lock(p, &flags);
2748 activate_task(rq, p, 0);
2749 trace_sched_wakeup_new(p, 1);
2750 check_preempt_curr(rq, p, WF_FORK);
2752 if (p->sched_class->task_woken)
2753 p->sched_class->task_woken(rq, p);
2755 task_rq_unlock(rq, &flags);
2759 #ifdef CONFIG_PREEMPT_NOTIFIERS
2762 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2763 * @notifier: notifier struct to register
2765 void preempt_notifier_register(struct preempt_notifier *notifier)
2767 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2769 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2772 * preempt_notifier_unregister - no longer interested in preemption notifications
2773 * @notifier: notifier struct to unregister
2775 * This is safe to call from within a preemption notifier.
2777 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2779 hlist_del(¬ifier->link);
2781 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2783 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2785 struct preempt_notifier *notifier;
2786 struct hlist_node *node;
2788 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2789 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2793 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2794 struct task_struct *next)
2796 struct preempt_notifier *notifier;
2797 struct hlist_node *node;
2799 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2800 notifier->ops->sched_out(notifier, next);
2803 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2805 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2810 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2811 struct task_struct *next)
2815 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2818 * prepare_task_switch - prepare to switch tasks
2819 * @rq: the runqueue preparing to switch
2820 * @prev: the current task that is being switched out
2821 * @next: the task we are going to switch to.
2823 * This is called with the rq lock held and interrupts off. It must
2824 * be paired with a subsequent finish_task_switch after the context
2827 * prepare_task_switch sets up locking and calls architecture specific
2831 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2832 struct task_struct *next)
2834 sched_info_switch(prev, next);
2835 perf_event_task_sched_out(prev, next);
2836 fire_sched_out_preempt_notifiers(prev, next);
2837 prepare_lock_switch(rq, next);
2838 prepare_arch_switch(next);
2839 trace_sched_switch(prev, next);
2843 * finish_task_switch - clean up after a task-switch
2844 * @rq: runqueue associated with task-switch
2845 * @prev: the thread we just switched away from.
2847 * finish_task_switch must be called after the context switch, paired
2848 * with a prepare_task_switch call before the context switch.
2849 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2850 * and do any other architecture-specific cleanup actions.
2852 * Note that we may have delayed dropping an mm in context_switch(). If
2853 * so, we finish that here outside of the runqueue lock. (Doing it
2854 * with the lock held can cause deadlocks; see schedule() for
2857 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2858 __releases(rq->lock)
2860 struct mm_struct *mm = rq->prev_mm;
2866 * A task struct has one reference for the use as "current".
2867 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2868 * schedule one last time. The schedule call will never return, and
2869 * the scheduled task must drop that reference.
2870 * The test for TASK_DEAD must occur while the runqueue locks are
2871 * still held, otherwise prev could be scheduled on another cpu, die
2872 * there before we look at prev->state, and then the reference would
2874 * Manfred Spraul <manfred@colorfullife.com>
2876 prev_state = prev->state;
2877 finish_arch_switch(prev);
2878 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2879 local_irq_disable();
2880 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2881 perf_event_task_sched_in(current);
2882 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2884 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2885 finish_lock_switch(rq, prev);
2887 fire_sched_in_preempt_notifiers(current);
2890 if (unlikely(prev_state == TASK_DEAD)) {
2892 * Remove function-return probe instances associated with this
2893 * task and put them back on the free list.
2895 kprobe_flush_task(prev);
2896 put_task_struct(prev);
2902 /* assumes rq->lock is held */
2903 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2905 if (prev->sched_class->pre_schedule)
2906 prev->sched_class->pre_schedule(rq, prev);
2909 /* rq->lock is NOT held, but preemption is disabled */
2910 static inline void post_schedule(struct rq *rq)
2912 if (rq->post_schedule) {
2913 unsigned long flags;
2915 raw_spin_lock_irqsave(&rq->lock, flags);
2916 if (rq->curr->sched_class->post_schedule)
2917 rq->curr->sched_class->post_schedule(rq);
2918 raw_spin_unlock_irqrestore(&rq->lock, flags);
2920 rq->post_schedule = 0;
2926 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2930 static inline void post_schedule(struct rq *rq)
2937 * schedule_tail - first thing a freshly forked thread must call.
2938 * @prev: the thread we just switched away from.
2940 asmlinkage void schedule_tail(struct task_struct *prev)
2941 __releases(rq->lock)
2943 struct rq *rq = this_rq();
2945 finish_task_switch(rq, prev);
2948 * FIXME: do we need to worry about rq being invalidated by the
2953 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2954 /* In this case, finish_task_switch does not reenable preemption */
2957 if (current->set_child_tid)
2958 put_user(task_pid_vnr(current), current->set_child_tid);
2962 * context_switch - switch to the new MM and the new
2963 * thread's register state.
2966 context_switch(struct rq *rq, struct task_struct *prev,
2967 struct task_struct *next)
2969 struct mm_struct *mm, *oldmm;
2971 prepare_task_switch(rq, prev, next);
2974 oldmm = prev->active_mm;
2976 * For paravirt, this is coupled with an exit in switch_to to
2977 * combine the page table reload and the switch backend into
2980 arch_start_context_switch(prev);
2983 next->active_mm = oldmm;
2984 atomic_inc(&oldmm->mm_count);
2985 enter_lazy_tlb(oldmm, next);
2987 switch_mm(oldmm, mm, next);
2990 prev->active_mm = NULL;
2991 rq->prev_mm = oldmm;
2994 * Since the runqueue lock will be released by the next
2995 * task (which is an invalid locking op but in the case
2996 * of the scheduler it's an obvious special-case), so we
2997 * do an early lockdep release here:
2999 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3000 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3003 /* Here we just switch the register state and the stack. */
3004 switch_to(prev, next, prev);
3008 * this_rq must be evaluated again because prev may have moved
3009 * CPUs since it called schedule(), thus the 'rq' on its stack
3010 * frame will be invalid.
3012 finish_task_switch(this_rq(), prev);
3016 * nr_running, nr_uninterruptible and nr_context_switches:
3018 * externally visible scheduler statistics: current number of runnable
3019 * threads, current number of uninterruptible-sleeping threads, total
3020 * number of context switches performed since bootup.
3022 unsigned long nr_running(void)
3024 unsigned long i, sum = 0;
3026 for_each_online_cpu(i)
3027 sum += cpu_rq(i)->nr_running;
3032 unsigned long nr_uninterruptible(void)
3034 unsigned long i, sum = 0;
3036 for_each_possible_cpu(i)
3037 sum += cpu_rq(i)->nr_uninterruptible;
3040 * Since we read the counters lockless, it might be slightly
3041 * inaccurate. Do not allow it to go below zero though:
3043 if (unlikely((long)sum < 0))
3049 unsigned long long nr_context_switches(void)
3052 unsigned long long sum = 0;
3054 for_each_possible_cpu(i)
3055 sum += cpu_rq(i)->nr_switches;
3060 unsigned long nr_iowait(void)
3062 unsigned long i, sum = 0;
3064 for_each_possible_cpu(i)
3065 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3070 unsigned long nr_iowait_cpu(int cpu)
3072 struct rq *this = cpu_rq(cpu);
3073 return atomic_read(&this->nr_iowait);
3076 unsigned long this_cpu_load(void)
3078 struct rq *this = this_rq();
3079 return this->cpu_load[0];
3083 /* Variables and functions for calc_load */
3084 static atomic_long_t calc_load_tasks;
3085 static unsigned long calc_load_update;
3086 unsigned long avenrun[3];
3087 EXPORT_SYMBOL(avenrun);
3089 static long calc_load_fold_active(struct rq *this_rq)
3091 long nr_active, delta = 0;
3093 nr_active = this_rq->nr_running;
3094 nr_active += (long) this_rq->nr_uninterruptible;
3096 if (nr_active != this_rq->calc_load_active) {
3097 delta = nr_active - this_rq->calc_load_active;
3098 this_rq->calc_load_active = nr_active;
3104 static unsigned long
3105 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3108 load += active * (FIXED_1 - exp);
3109 load += 1UL << (FSHIFT - 1);
3110 return load >> FSHIFT;
3115 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3117 * When making the ILB scale, we should try to pull this in as well.
3119 static atomic_long_t calc_load_tasks_idle;
3121 static void calc_load_account_idle(struct rq *this_rq)
3125 delta = calc_load_fold_active(this_rq);
3127 atomic_long_add(delta, &calc_load_tasks_idle);
3130 static long calc_load_fold_idle(void)
3135 * Its got a race, we don't care...
3137 if (atomic_long_read(&calc_load_tasks_idle))
3138 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3144 * fixed_power_int - compute: x^n, in O(log n) time
3146 * @x: base of the power
3147 * @frac_bits: fractional bits of @x
3148 * @n: power to raise @x to.
3150 * By exploiting the relation between the definition of the natural power
3151 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3152 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3153 * (where: n_i \elem {0, 1}, the binary vector representing n),
3154 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3155 * of course trivially computable in O(log_2 n), the length of our binary
3158 static unsigned long
3159 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3161 unsigned long result = 1UL << frac_bits;
3166 result += 1UL << (frac_bits - 1);
3167 result >>= frac_bits;
3173 x += 1UL << (frac_bits - 1);
3181 * a1 = a0 * e + a * (1 - e)
3183 * a2 = a1 * e + a * (1 - e)
3184 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3185 * = a0 * e^2 + a * (1 - e) * (1 + e)
3187 * a3 = a2 * e + a * (1 - e)
3188 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3189 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3193 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3194 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3195 * = a0 * e^n + a * (1 - e^n)
3197 * [1] application of the geometric series:
3200 * S_n := \Sum x^i = -------------
3203 static unsigned long
3204 calc_load_n(unsigned long load, unsigned long exp,
3205 unsigned long active, unsigned int n)
3208 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3212 * NO_HZ can leave us missing all per-cpu ticks calling
3213 * calc_load_account_active(), but since an idle CPU folds its delta into
3214 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3215 * in the pending idle delta if our idle period crossed a load cycle boundary.
3217 * Once we've updated the global active value, we need to apply the exponential
3218 * weights adjusted to the number of cycles missed.
3220 static void calc_global_nohz(unsigned long ticks)
3222 long delta, active, n;
3224 if (time_before(jiffies, calc_load_update))
3228 * If we crossed a calc_load_update boundary, make sure to fold
3229 * any pending idle changes, the respective CPUs might have
3230 * missed the tick driven calc_load_account_active() update
3233 delta = calc_load_fold_idle();
3235 atomic_long_add(delta, &calc_load_tasks);
3238 * If we were idle for multiple load cycles, apply them.
3240 if (ticks >= LOAD_FREQ) {
3241 n = ticks / LOAD_FREQ;
3243 active = atomic_long_read(&calc_load_tasks);
3244 active = active > 0 ? active * FIXED_1 : 0;
3246 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3247 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3248 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3250 calc_load_update += n * LOAD_FREQ;
3254 * Its possible the remainder of the above division also crosses
3255 * a LOAD_FREQ period, the regular check in calc_global_load()
3256 * which comes after this will take care of that.
3258 * Consider us being 11 ticks before a cycle completion, and us
3259 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3260 * age us 4 cycles, and the test in calc_global_load() will
3261 * pick up the final one.
3265 static void calc_load_account_idle(struct rq *this_rq)
3269 static inline long calc_load_fold_idle(void)
3274 static void calc_global_nohz(unsigned long ticks)
3280 * get_avenrun - get the load average array
3281 * @loads: pointer to dest load array
3282 * @offset: offset to add
3283 * @shift: shift count to shift the result left
3285 * These values are estimates at best, so no need for locking.
3287 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3289 loads[0] = (avenrun[0] + offset) << shift;
3290 loads[1] = (avenrun[1] + offset) << shift;
3291 loads[2] = (avenrun[2] + offset) << shift;
3295 * calc_load - update the avenrun load estimates 10 ticks after the
3296 * CPUs have updated calc_load_tasks.
3298 void calc_global_load(unsigned long ticks)
3302 calc_global_nohz(ticks);
3304 if (time_before(jiffies, calc_load_update + 10))
3307 active = atomic_long_read(&calc_load_tasks);
3308 active = active > 0 ? active * FIXED_1 : 0;
3310 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3311 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3312 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3314 calc_load_update += LOAD_FREQ;
3318 * Called from update_cpu_load() to periodically update this CPU's
3321 static void calc_load_account_active(struct rq *this_rq)
3325 if (time_before(jiffies, this_rq->calc_load_update))
3328 delta = calc_load_fold_active(this_rq);
3329 delta += calc_load_fold_idle();
3331 atomic_long_add(delta, &calc_load_tasks);
3333 this_rq->calc_load_update += LOAD_FREQ;
3337 * The exact cpuload at various idx values, calculated at every tick would be
3338 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3340 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3341 * on nth tick when cpu may be busy, then we have:
3342 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3343 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3345 * decay_load_missed() below does efficient calculation of
3346 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3347 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3349 * The calculation is approximated on a 128 point scale.
3350 * degrade_zero_ticks is the number of ticks after which load at any
3351 * particular idx is approximated to be zero.
3352 * degrade_factor is a precomputed table, a row for each load idx.
3353 * Each column corresponds to degradation factor for a power of two ticks,
3354 * based on 128 point scale.
3356 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3357 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3359 * With this power of 2 load factors, we can degrade the load n times
3360 * by looking at 1 bits in n and doing as many mult/shift instead of
3361 * n mult/shifts needed by the exact degradation.
3363 #define DEGRADE_SHIFT 7
3364 static const unsigned char
3365 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3366 static const unsigned char
3367 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3368 {0, 0, 0, 0, 0, 0, 0, 0},
3369 {64, 32, 8, 0, 0, 0, 0, 0},
3370 {96, 72, 40, 12, 1, 0, 0},
3371 {112, 98, 75, 43, 15, 1, 0},
3372 {120, 112, 98, 76, 45, 16, 2} };
3375 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3376 * would be when CPU is idle and so we just decay the old load without
3377 * adding any new load.
3379 static unsigned long
3380 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3384 if (!missed_updates)
3387 if (missed_updates >= degrade_zero_ticks[idx])
3391 return load >> missed_updates;
3393 while (missed_updates) {
3394 if (missed_updates % 2)
3395 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3397 missed_updates >>= 1;
3404 * Update rq->cpu_load[] statistics. This function is usually called every
3405 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3406 * every tick. We fix it up based on jiffies.
3408 static void update_cpu_load(struct rq *this_rq)
3410 unsigned long this_load = this_rq->load.weight;
3411 unsigned long curr_jiffies = jiffies;
3412 unsigned long pending_updates;
3415 this_rq->nr_load_updates++;
3417 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3418 if (curr_jiffies == this_rq->last_load_update_tick)
3421 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3422 this_rq->last_load_update_tick = curr_jiffies;
3424 /* Update our load: */
3425 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3426 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3427 unsigned long old_load, new_load;
3429 /* scale is effectively 1 << i now, and >> i divides by scale */
3431 old_load = this_rq->cpu_load[i];
3432 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3433 new_load = this_load;
3435 * Round up the averaging division if load is increasing. This
3436 * prevents us from getting stuck on 9 if the load is 10, for
3439 if (new_load > old_load)
3440 new_load += scale - 1;
3442 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3445 sched_avg_update(this_rq);
3448 static void update_cpu_load_active(struct rq *this_rq)
3450 update_cpu_load(this_rq);
3452 calc_load_account_active(this_rq);
3458 * sched_exec - execve() is a valuable balancing opportunity, because at
3459 * this point the task has the smallest effective memory and cache footprint.
3461 void sched_exec(void)
3463 struct task_struct *p = current;
3464 unsigned long flags;
3468 rq = task_rq_lock(p, &flags);
3469 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3470 if (dest_cpu == smp_processor_id())
3474 * select_task_rq() can race against ->cpus_allowed
3476 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3477 likely(cpu_active(dest_cpu)) && migrate_task(p, rq)) {
3478 struct migration_arg arg = { p, dest_cpu };
3480 task_rq_unlock(rq, &flags);
3481 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3485 task_rq_unlock(rq, &flags);
3490 DEFINE_PER_CPU(struct kernel_stat, kstat);
3492 EXPORT_PER_CPU_SYMBOL(kstat);
3495 * Return any ns on the sched_clock that have not yet been accounted in
3496 * @p in case that task is currently running.
3498 * Called with task_rq_lock() held on @rq.
3500 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3504 if (task_current(rq, p)) {
3505 update_rq_clock(rq);
3506 ns = rq->clock_task - p->se.exec_start;
3514 unsigned long long task_delta_exec(struct task_struct *p)
3516 unsigned long flags;
3520 rq = task_rq_lock(p, &flags);
3521 ns = do_task_delta_exec(p, rq);
3522 task_rq_unlock(rq, &flags);
3528 * Return accounted runtime for the task.
3529 * In case the task is currently running, return the runtime plus current's
3530 * pending runtime that have not been accounted yet.
3532 unsigned long long task_sched_runtime(struct task_struct *p)
3534 unsigned long flags;
3538 rq = task_rq_lock(p, &flags);
3539 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3540 task_rq_unlock(rq, &flags);
3546 * Return sum_exec_runtime for the thread group.
3547 * In case the task is currently running, return the sum plus current's
3548 * pending runtime that have not been accounted yet.
3550 * Note that the thread group might have other running tasks as well,
3551 * so the return value not includes other pending runtime that other
3552 * running tasks might have.
3554 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3556 struct task_cputime totals;
3557 unsigned long flags;
3561 rq = task_rq_lock(p, &flags);
3562 thread_group_cputime(p, &totals);
3563 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3564 task_rq_unlock(rq, &flags);
3570 * Account user cpu time to a process.
3571 * @p: the process that the cpu time gets accounted to
3572 * @cputime: the cpu time spent in user space since the last update
3573 * @cputime_scaled: cputime scaled by cpu frequency
3575 void account_user_time(struct task_struct *p, cputime_t cputime,
3576 cputime_t cputime_scaled)
3578 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3581 /* Add user time to process. */
3582 p->utime = cputime_add(p->utime, cputime);
3583 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3584 account_group_user_time(p, cputime);
3586 /* Add user time to cpustat. */
3587 tmp = cputime_to_cputime64(cputime);
3588 if (TASK_NICE(p) > 0)
3589 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3591 cpustat->user = cputime64_add(cpustat->user, tmp);
3593 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3594 /* Account for user time used */
3595 acct_update_integrals(p);
3599 * Account guest cpu time to a process.
3600 * @p: the process that the cpu time gets accounted to
3601 * @cputime: the cpu time spent in virtual machine since the last update
3602 * @cputime_scaled: cputime scaled by cpu frequency
3604 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3605 cputime_t cputime_scaled)
3608 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3610 tmp = cputime_to_cputime64(cputime);
3612 /* Add guest time to process. */
3613 p->utime = cputime_add(p->utime, cputime);
3614 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3615 account_group_user_time(p, cputime);
3616 p->gtime = cputime_add(p->gtime, cputime);
3618 /* Add guest time to cpustat. */
3619 if (TASK_NICE(p) > 0) {
3620 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3621 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3623 cpustat->user = cputime64_add(cpustat->user, tmp);
3624 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3629 * Account system cpu time to a process and desired cpustat field
3630 * @p: the process that the cpu time gets accounted to
3631 * @cputime: the cpu time spent in kernel space since the last update
3632 * @cputime_scaled: cputime scaled by cpu frequency
3633 * @target_cputime64: pointer to cpustat field that has to be updated
3636 void __account_system_time(struct task_struct *p, cputime_t cputime,
3637 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3639 cputime64_t tmp = cputime_to_cputime64(cputime);
3641 /* Add system time to process. */
3642 p->stime = cputime_add(p->stime, cputime);
3643 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3644 account_group_system_time(p, cputime);
3646 /* Add system time to cpustat. */
3647 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3648 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3650 /* Account for system time used */
3651 acct_update_integrals(p);
3655 * Account system cpu time to a process.
3656 * @p: the process that the cpu time gets accounted to
3657 * @hardirq_offset: the offset to subtract from hardirq_count()
3658 * @cputime: the cpu time spent in kernel space since the last update
3659 * @cputime_scaled: cputime scaled by cpu frequency
3661 void account_system_time(struct task_struct *p, int hardirq_offset,
3662 cputime_t cputime, cputime_t cputime_scaled)
3664 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3665 cputime64_t *target_cputime64;
3667 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3668 account_guest_time(p, cputime, cputime_scaled);
3672 if (hardirq_count() - hardirq_offset)
3673 target_cputime64 = &cpustat->irq;
3674 else if (in_serving_softirq())
3675 target_cputime64 = &cpustat->softirq;
3677 target_cputime64 = &cpustat->system;
3679 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3683 * Account for involuntary wait time.
3684 * @cputime: the cpu time spent in involuntary wait
3686 void account_steal_time(cputime_t cputime)
3688 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3689 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3691 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3695 * Account for idle time.
3696 * @cputime: the cpu time spent in idle wait
3698 void account_idle_time(cputime_t cputime)
3700 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3701 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3702 struct rq *rq = this_rq();
3704 if (atomic_read(&rq->nr_iowait) > 0)
3705 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3707 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3710 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3712 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3714 * Account a tick to a process and cpustat
3715 * @p: the process that the cpu time gets accounted to
3716 * @user_tick: is the tick from userspace
3717 * @rq: the pointer to rq
3719 * Tick demultiplexing follows the order
3720 * - pending hardirq update
3721 * - pending softirq update
3725 * - check for guest_time
3726 * - else account as system_time
3728 * Check for hardirq is done both for system and user time as there is
3729 * no timer going off while we are on hardirq and hence we may never get an
3730 * opportunity to update it solely in system time.
3731 * p->stime and friends are only updated on system time and not on irq
3732 * softirq as those do not count in task exec_runtime any more.
3734 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3737 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3738 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
3739 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3741 if (irqtime_account_hi_update()) {
3742 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3743 } else if (irqtime_account_si_update()) {
3744 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3745 } else if (this_cpu_ksoftirqd() == p) {
3747 * ksoftirqd time do not get accounted in cpu_softirq_time.
3748 * So, we have to handle it separately here.
3749 * Also, p->stime needs to be updated for ksoftirqd.
3751 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3753 } else if (user_tick) {
3754 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3755 } else if (p == rq->idle) {
3756 account_idle_time(cputime_one_jiffy);
3757 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3758 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3760 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3765 static void irqtime_account_idle_ticks(int ticks)
3768 struct rq *rq = this_rq();
3770 for (i = 0; i < ticks; i++)
3771 irqtime_account_process_tick(current, 0, rq);
3773 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
3774 static void irqtime_account_idle_ticks(int ticks) {}
3775 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3777 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3780 * Account a single tick of cpu time.
3781 * @p: the process that the cpu time gets accounted to
3782 * @user_tick: indicates if the tick is a user or a system tick
3784 void account_process_tick(struct task_struct *p, int user_tick)
3786 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3787 struct rq *rq = this_rq();
3789 if (sched_clock_irqtime) {
3790 irqtime_account_process_tick(p, user_tick, rq);
3795 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3796 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3797 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3800 account_idle_time(cputime_one_jiffy);
3804 * Account multiple ticks of steal time.
3805 * @p: the process from which the cpu time has been stolen
3806 * @ticks: number of stolen ticks
3808 void account_steal_ticks(unsigned long ticks)
3810 account_steal_time(jiffies_to_cputime(ticks));
3814 * Account multiple ticks of idle time.
3815 * @ticks: number of stolen ticks
3817 void account_idle_ticks(unsigned long ticks)
3820 if (sched_clock_irqtime) {
3821 irqtime_account_idle_ticks(ticks);
3825 account_idle_time(jiffies_to_cputime(ticks));
3831 * Use precise platform statistics if available:
3833 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3834 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3840 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3842 struct task_cputime cputime;
3844 thread_group_cputime(p, &cputime);
3846 *ut = cputime.utime;
3847 *st = cputime.stime;
3851 #ifndef nsecs_to_cputime
3852 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3855 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3857 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3860 * Use CFS's precise accounting:
3862 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3868 do_div(temp, total);
3869 utime = (cputime_t)temp;
3874 * Compare with previous values, to keep monotonicity:
3876 p->prev_utime = max(p->prev_utime, utime);
3877 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3879 *ut = p->prev_utime;
3880 *st = p->prev_stime;
3884 * Must be called with siglock held.
3886 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3888 struct signal_struct *sig = p->signal;
3889 struct task_cputime cputime;
3890 cputime_t rtime, utime, total;
3892 thread_group_cputime(p, &cputime);
3894 total = cputime_add(cputime.utime, cputime.stime);
3895 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3900 temp *= cputime.utime;
3901 do_div(temp, total);
3902 utime = (cputime_t)temp;
3906 sig->prev_utime = max(sig->prev_utime, utime);
3907 sig->prev_stime = max(sig->prev_stime,
3908 cputime_sub(rtime, sig->prev_utime));
3910 *ut = sig->prev_utime;
3911 *st = sig->prev_stime;
3916 * This function gets called by the timer code, with HZ frequency.
3917 * We call it with interrupts disabled.
3919 * It also gets called by the fork code, when changing the parent's
3922 void scheduler_tick(void)
3924 int cpu = smp_processor_id();
3925 struct rq *rq = cpu_rq(cpu);
3926 struct task_struct *curr = rq->curr;
3930 raw_spin_lock(&rq->lock);
3931 update_rq_clock(rq);
3932 update_cpu_load_active(rq);
3933 curr->sched_class->task_tick(rq, curr, 0);
3934 raw_spin_unlock(&rq->lock);
3936 perf_event_task_tick();
3939 rq->idle_at_tick = idle_cpu(cpu);
3940 trigger_load_balance(rq, cpu);
3944 notrace unsigned long get_parent_ip(unsigned long addr)
3946 if (in_lock_functions(addr)) {
3947 addr = CALLER_ADDR2;
3948 if (in_lock_functions(addr))
3949 addr = CALLER_ADDR3;
3954 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3955 defined(CONFIG_PREEMPT_TRACER))
3957 void __kprobes add_preempt_count(int val)
3959 #ifdef CONFIG_DEBUG_PREEMPT
3963 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3966 preempt_count() += val;
3967 #ifdef CONFIG_DEBUG_PREEMPT
3969 * Spinlock count overflowing soon?
3971 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3974 if (preempt_count() == val)
3975 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3977 EXPORT_SYMBOL(add_preempt_count);
3979 void __kprobes sub_preempt_count(int val)
3981 #ifdef CONFIG_DEBUG_PREEMPT
3985 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3988 * Is the spinlock portion underflowing?
3990 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3991 !(preempt_count() & PREEMPT_MASK)))
3995 if (preempt_count() == val)
3996 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3997 preempt_count() -= val;
3999 EXPORT_SYMBOL(sub_preempt_count);
4004 * Print scheduling while atomic bug:
4006 static noinline void __schedule_bug(struct task_struct *prev)
4008 struct pt_regs *regs = get_irq_regs();
4010 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4011 prev->comm, prev->pid, preempt_count());
4013 debug_show_held_locks(prev);
4015 if (irqs_disabled())
4016 print_irqtrace_events(prev);
4025 * Various schedule()-time debugging checks and statistics:
4027 static inline void schedule_debug(struct task_struct *prev)
4030 * Test if we are atomic. Since do_exit() needs to call into
4031 * schedule() atomically, we ignore that path for now.
4032 * Otherwise, whine if we are scheduling when we should not be.
4034 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4035 __schedule_bug(prev);
4037 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4039 schedstat_inc(this_rq(), sched_count);
4040 #ifdef CONFIG_SCHEDSTATS
4041 if (unlikely(prev->lock_depth >= 0)) {
4042 schedstat_inc(this_rq(), rq_sched_info.bkl_count);
4043 schedstat_inc(prev, sched_info.bkl_count);
4048 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4051 update_rq_clock(rq);
4052 prev->sched_class->put_prev_task(rq, prev);
4056 * Pick up the highest-prio task:
4058 static inline struct task_struct *
4059 pick_next_task(struct rq *rq)
4061 const struct sched_class *class;
4062 struct task_struct *p;
4065 * Optimization: we know that if all tasks are in
4066 * the fair class we can call that function directly:
4068 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4069 p = fair_sched_class.pick_next_task(rq);
4074 for_each_class(class) {
4075 p = class->pick_next_task(rq);
4080 BUG(); /* the idle class will always have a runnable task */
4084 * schedule() is the main scheduler function.
4086 asmlinkage void __sched schedule(void)
4088 struct task_struct *prev, *next;
4089 unsigned long *switch_count;
4095 cpu = smp_processor_id();
4097 rcu_note_context_switch(cpu);
4100 schedule_debug(prev);
4102 if (sched_feat(HRTICK))
4105 raw_spin_lock_irq(&rq->lock);
4107 switch_count = &prev->nivcsw;
4108 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4109 if (unlikely(signal_pending_state(prev->state, prev))) {
4110 prev->state = TASK_RUNNING;
4113 * If a worker is going to sleep, notify and
4114 * ask workqueue whether it wants to wake up a
4115 * task to maintain concurrency. If so, wake
4118 if (prev->flags & PF_WQ_WORKER) {
4119 struct task_struct *to_wakeup;
4121 to_wakeup = wq_worker_sleeping(prev, cpu);
4123 try_to_wake_up_local(to_wakeup);
4125 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4128 * If we are going to sleep and we have plugged IO queued, make
4129 * sure to submit it to avoid deadlocks.
4131 if (blk_needs_flush_plug(prev)) {
4132 raw_spin_unlock(&rq->lock);
4133 blk_flush_plug(prev);
4134 raw_spin_lock(&rq->lock);
4137 switch_count = &prev->nvcsw;
4140 pre_schedule(rq, prev);
4142 if (unlikely(!rq->nr_running))
4143 idle_balance(cpu, rq);
4145 put_prev_task(rq, prev);
4146 next = pick_next_task(rq);
4147 clear_tsk_need_resched(prev);
4148 rq->skip_clock_update = 0;
4150 if (likely(prev != next)) {
4155 context_switch(rq, prev, next); /* unlocks the rq */
4157 * The context switch have flipped the stack from under us
4158 * and restored the local variables which were saved when
4159 * this task called schedule() in the past. prev == current
4160 * is still correct, but it can be moved to another cpu/rq.
4162 cpu = smp_processor_id();
4165 raw_spin_unlock_irq(&rq->lock);
4169 preempt_enable_no_resched();
4173 EXPORT_SYMBOL(schedule);
4175 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4177 * Look out! "owner" is an entirely speculative pointer
4178 * access and not reliable.
4180 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
4185 if (!sched_feat(OWNER_SPIN))
4188 #ifdef CONFIG_DEBUG_PAGEALLOC
4190 * Need to access the cpu field knowing that
4191 * DEBUG_PAGEALLOC could have unmapped it if
4192 * the mutex owner just released it and exited.
4194 if (probe_kernel_address(&owner->cpu, cpu))
4201 * Even if the access succeeded (likely case),
4202 * the cpu field may no longer be valid.
4204 if (cpu >= nr_cpumask_bits)
4208 * We need to validate that we can do a
4209 * get_cpu() and that we have the percpu area.
4211 if (!cpu_online(cpu))
4218 * Owner changed, break to re-assess state.
4220 if (lock->owner != owner) {
4222 * If the lock has switched to a different owner,
4223 * we likely have heavy contention. Return 0 to quit
4224 * optimistic spinning and not contend further:
4232 * Is that owner really running on that cpu?
4234 if (task_thread_info(rq->curr) != owner || need_resched())
4237 arch_mutex_cpu_relax();
4244 #ifdef CONFIG_PREEMPT
4246 * this is the entry point to schedule() from in-kernel preemption
4247 * off of preempt_enable. Kernel preemptions off return from interrupt
4248 * occur there and call schedule directly.
4250 asmlinkage void __sched notrace preempt_schedule(void)
4252 struct thread_info *ti = current_thread_info();
4255 * If there is a non-zero preempt_count or interrupts are disabled,
4256 * we do not want to preempt the current task. Just return..
4258 if (likely(ti->preempt_count || irqs_disabled()))
4262 add_preempt_count_notrace(PREEMPT_ACTIVE);
4264 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4267 * Check again in case we missed a preemption opportunity
4268 * between schedule and now.
4271 } while (need_resched());
4273 EXPORT_SYMBOL(preempt_schedule);
4276 * this is the entry point to schedule() from kernel preemption
4277 * off of irq context.
4278 * Note, that this is called and return with irqs disabled. This will
4279 * protect us against recursive calling from irq.
4281 asmlinkage void __sched preempt_schedule_irq(void)
4283 struct thread_info *ti = current_thread_info();
4285 /* Catch callers which need to be fixed */
4286 BUG_ON(ti->preempt_count || !irqs_disabled());
4289 add_preempt_count(PREEMPT_ACTIVE);
4292 local_irq_disable();
4293 sub_preempt_count(PREEMPT_ACTIVE);
4296 * Check again in case we missed a preemption opportunity
4297 * between schedule and now.
4300 } while (need_resched());
4303 #endif /* CONFIG_PREEMPT */
4305 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4308 return try_to_wake_up(curr->private, mode, wake_flags);
4310 EXPORT_SYMBOL(default_wake_function);
4313 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4314 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4315 * number) then we wake all the non-exclusive tasks and one exclusive task.
4317 * There are circumstances in which we can try to wake a task which has already
4318 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4319 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4321 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4322 int nr_exclusive, int wake_flags, void *key)
4324 wait_queue_t *curr, *next;
4326 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4327 unsigned flags = curr->flags;
4329 if (curr->func(curr, mode, wake_flags, key) &&
4330 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4336 * __wake_up - wake up threads blocked on a waitqueue.
4338 * @mode: which threads
4339 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4340 * @key: is directly passed to the wakeup function
4342 * It may be assumed that this function implies a write memory barrier before
4343 * changing the task state if and only if any tasks are woken up.
4345 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4346 int nr_exclusive, void *key)
4348 unsigned long flags;
4350 spin_lock_irqsave(&q->lock, flags);
4351 __wake_up_common(q, mode, nr_exclusive, 0, key);
4352 spin_unlock_irqrestore(&q->lock, flags);
4354 EXPORT_SYMBOL(__wake_up);
4357 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4359 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4361 __wake_up_common(q, mode, 1, 0, NULL);
4363 EXPORT_SYMBOL_GPL(__wake_up_locked);
4365 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4367 __wake_up_common(q, mode, 1, 0, key);
4369 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4372 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4374 * @mode: which threads
4375 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4376 * @key: opaque value to be passed to wakeup targets
4378 * The sync wakeup differs that the waker knows that it will schedule
4379 * away soon, so while the target thread will be woken up, it will not
4380 * be migrated to another CPU - ie. the two threads are 'synchronized'
4381 * with each other. This can prevent needless bouncing between CPUs.
4383 * On UP it can prevent extra preemption.
4385 * It may be assumed that this function implies a write memory barrier before
4386 * changing the task state if and only if any tasks are woken up.
4388 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4389 int nr_exclusive, void *key)
4391 unsigned long flags;
4392 int wake_flags = WF_SYNC;
4397 if (unlikely(!nr_exclusive))
4400 spin_lock_irqsave(&q->lock, flags);
4401 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4402 spin_unlock_irqrestore(&q->lock, flags);
4404 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4407 * __wake_up_sync - see __wake_up_sync_key()
4409 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4411 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4413 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4416 * complete: - signals a single thread waiting on this completion
4417 * @x: holds the state of this particular completion
4419 * This will wake up a single thread waiting on this completion. Threads will be
4420 * awakened in the same order in which they were queued.
4422 * See also complete_all(), wait_for_completion() and related routines.
4424 * It may be assumed that this function implies a write memory barrier before
4425 * changing the task state if and only if any tasks are woken up.
4427 void complete(struct completion *x)
4429 unsigned long flags;
4431 spin_lock_irqsave(&x->wait.lock, flags);
4433 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4434 spin_unlock_irqrestore(&x->wait.lock, flags);
4436 EXPORT_SYMBOL(complete);
4439 * complete_all: - signals all threads waiting on this completion
4440 * @x: holds the state of this particular completion
4442 * This will wake up all threads waiting on this particular completion event.
4444 * It may be assumed that this function implies a write memory barrier before
4445 * changing the task state if and only if any tasks are woken up.
4447 void complete_all(struct completion *x)
4449 unsigned long flags;
4451 spin_lock_irqsave(&x->wait.lock, flags);
4452 x->done += UINT_MAX/2;
4453 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4454 spin_unlock_irqrestore(&x->wait.lock, flags);
4456 EXPORT_SYMBOL(complete_all);
4458 static inline long __sched
4459 do_wait_for_common(struct completion *x, long timeout, int state)
4462 DECLARE_WAITQUEUE(wait, current);
4464 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4466 if (signal_pending_state(state, current)) {
4467 timeout = -ERESTARTSYS;
4470 __set_current_state(state);
4471 spin_unlock_irq(&x->wait.lock);
4472 timeout = schedule_timeout(timeout);
4473 spin_lock_irq(&x->wait.lock);
4474 } while (!x->done && timeout);
4475 __remove_wait_queue(&x->wait, &wait);
4480 return timeout ?: 1;
4484 wait_for_common(struct completion *x, long timeout, int state)
4488 spin_lock_irq(&x->wait.lock);
4489 timeout = do_wait_for_common(x, timeout, state);
4490 spin_unlock_irq(&x->wait.lock);
4495 * wait_for_completion: - waits for completion of a task
4496 * @x: holds the state of this particular completion
4498 * This waits to be signaled for completion of a specific task. It is NOT
4499 * interruptible and there is no timeout.
4501 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4502 * and interrupt capability. Also see complete().
4504 void __sched wait_for_completion(struct completion *x)
4506 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4508 EXPORT_SYMBOL(wait_for_completion);
4511 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4512 * @x: holds the state of this particular completion
4513 * @timeout: timeout value in jiffies
4515 * This waits for either a completion of a specific task to be signaled or for a
4516 * specified timeout to expire. The timeout is in jiffies. It is not
4519 unsigned long __sched
4520 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4522 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4524 EXPORT_SYMBOL(wait_for_completion_timeout);
4527 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4528 * @x: holds the state of this particular completion
4530 * This waits for completion of a specific task to be signaled. It is
4533 int __sched wait_for_completion_interruptible(struct completion *x)
4535 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4536 if (t == -ERESTARTSYS)
4540 EXPORT_SYMBOL(wait_for_completion_interruptible);
4543 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4544 * @x: holds the state of this particular completion
4545 * @timeout: timeout value in jiffies
4547 * This waits for either a completion of a specific task to be signaled or for a
4548 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4551 wait_for_completion_interruptible_timeout(struct completion *x,
4552 unsigned long timeout)
4554 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4556 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4559 * wait_for_completion_killable: - waits for completion of a task (killable)
4560 * @x: holds the state of this particular completion
4562 * This waits to be signaled for completion of a specific task. It can be
4563 * interrupted by a kill signal.
4565 int __sched wait_for_completion_killable(struct completion *x)
4567 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4568 if (t == -ERESTARTSYS)
4572 EXPORT_SYMBOL(wait_for_completion_killable);
4575 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4576 * @x: holds the state of this particular completion
4577 * @timeout: timeout value in jiffies
4579 * This waits for either a completion of a specific task to be
4580 * signaled or for a specified timeout to expire. It can be
4581 * interrupted by a kill signal. The timeout is in jiffies.
4584 wait_for_completion_killable_timeout(struct completion *x,
4585 unsigned long timeout)
4587 return wait_for_common(x, timeout, TASK_KILLABLE);
4589 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4592 * try_wait_for_completion - try to decrement a completion without blocking
4593 * @x: completion structure
4595 * Returns: 0 if a decrement cannot be done without blocking
4596 * 1 if a decrement succeeded.
4598 * If a completion is being used as a counting completion,
4599 * attempt to decrement the counter without blocking. This
4600 * enables us to avoid waiting if the resource the completion
4601 * is protecting is not available.
4603 bool try_wait_for_completion(struct completion *x)
4605 unsigned long flags;
4608 spin_lock_irqsave(&x->wait.lock, flags);
4613 spin_unlock_irqrestore(&x->wait.lock, flags);
4616 EXPORT_SYMBOL(try_wait_for_completion);
4619 * completion_done - Test to see if a completion has any waiters
4620 * @x: completion structure
4622 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4623 * 1 if there are no waiters.
4626 bool completion_done(struct completion *x)
4628 unsigned long flags;
4631 spin_lock_irqsave(&x->wait.lock, flags);
4634 spin_unlock_irqrestore(&x->wait.lock, flags);
4637 EXPORT_SYMBOL(completion_done);
4640 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4642 unsigned long flags;
4645 init_waitqueue_entry(&wait, current);
4647 __set_current_state(state);
4649 spin_lock_irqsave(&q->lock, flags);
4650 __add_wait_queue(q, &wait);
4651 spin_unlock(&q->lock);
4652 timeout = schedule_timeout(timeout);
4653 spin_lock_irq(&q->lock);
4654 __remove_wait_queue(q, &wait);
4655 spin_unlock_irqrestore(&q->lock, flags);
4660 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4662 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4664 EXPORT_SYMBOL(interruptible_sleep_on);
4667 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4669 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4671 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4673 void __sched sleep_on(wait_queue_head_t *q)
4675 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4677 EXPORT_SYMBOL(sleep_on);
4679 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4681 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4683 EXPORT_SYMBOL(sleep_on_timeout);
4685 #ifdef CONFIG_RT_MUTEXES
4688 * rt_mutex_setprio - set the current priority of a task
4690 * @prio: prio value (kernel-internal form)
4692 * This function changes the 'effective' priority of a task. It does
4693 * not touch ->normal_prio like __setscheduler().
4695 * Used by the rt_mutex code to implement priority inheritance logic.
4697 void rt_mutex_setprio(struct task_struct *p, int prio)
4699 unsigned long flags;
4700 int oldprio, on_rq, running;
4702 const struct sched_class *prev_class;
4704 BUG_ON(prio < 0 || prio > MAX_PRIO);
4706 rq = task_rq_lock(p, &flags);
4708 trace_sched_pi_setprio(p, prio);
4710 prev_class = p->sched_class;
4711 on_rq = p->se.on_rq;
4712 running = task_current(rq, p);
4714 dequeue_task(rq, p, 0);
4716 p->sched_class->put_prev_task(rq, p);
4719 p->sched_class = &rt_sched_class;
4721 p->sched_class = &fair_sched_class;
4726 p->sched_class->set_curr_task(rq);
4728 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4730 check_class_changed(rq, p, prev_class, oldprio);
4731 task_rq_unlock(rq, &flags);
4736 void set_user_nice(struct task_struct *p, long nice)
4738 int old_prio, delta, on_rq;
4739 unsigned long flags;
4742 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4745 * We have to be careful, if called from sys_setpriority(),
4746 * the task might be in the middle of scheduling on another CPU.
4748 rq = task_rq_lock(p, &flags);
4750 * The RT priorities are set via sched_setscheduler(), but we still
4751 * allow the 'normal' nice value to be set - but as expected
4752 * it wont have any effect on scheduling until the task is
4753 * SCHED_FIFO/SCHED_RR:
4755 if (task_has_rt_policy(p)) {
4756 p->static_prio = NICE_TO_PRIO(nice);
4759 on_rq = p->se.on_rq;
4761 dequeue_task(rq, p, 0);
4763 p->static_prio = NICE_TO_PRIO(nice);
4766 p->prio = effective_prio(p);
4767 delta = p->prio - old_prio;
4770 enqueue_task(rq, p, 0);
4772 * If the task increased its priority or is running and
4773 * lowered its priority, then reschedule its CPU:
4775 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4776 resched_task(rq->curr);
4779 task_rq_unlock(rq, &flags);
4781 EXPORT_SYMBOL(set_user_nice);
4784 * can_nice - check if a task can reduce its nice value
4788 int can_nice(const struct task_struct *p, const int nice)
4790 /* convert nice value [19,-20] to rlimit style value [1,40] */
4791 int nice_rlim = 20 - nice;
4793 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4794 capable(CAP_SYS_NICE));
4797 #ifdef __ARCH_WANT_SYS_NICE
4800 * sys_nice - change the priority of the current process.
4801 * @increment: priority increment
4803 * sys_setpriority is a more generic, but much slower function that
4804 * does similar things.
4806 SYSCALL_DEFINE1(nice, int, increment)
4811 * Setpriority might change our priority at the same moment.
4812 * We don't have to worry. Conceptually one call occurs first
4813 * and we have a single winner.
4815 if (increment < -40)
4820 nice = TASK_NICE(current) + increment;
4826 if (increment < 0 && !can_nice(current, nice))
4829 retval = security_task_setnice(current, nice);
4833 set_user_nice(current, nice);
4840 * task_prio - return the priority value of a given task.
4841 * @p: the task in question.
4843 * This is the priority value as seen by users in /proc.
4844 * RT tasks are offset by -200. Normal tasks are centered
4845 * around 0, value goes from -16 to +15.
4847 int task_prio(const struct task_struct *p)
4849 return p->prio - MAX_RT_PRIO;
4853 * task_nice - return the nice value of a given task.
4854 * @p: the task in question.
4856 int task_nice(const struct task_struct *p)
4858 return TASK_NICE(p);
4860 EXPORT_SYMBOL(task_nice);
4863 * idle_cpu - is a given cpu idle currently?
4864 * @cpu: the processor in question.
4866 int idle_cpu(int cpu)
4868 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4872 * idle_task - return the idle task for a given cpu.
4873 * @cpu: the processor in question.
4875 struct task_struct *idle_task(int cpu)
4877 return cpu_rq(cpu)->idle;
4881 * find_process_by_pid - find a process with a matching PID value.
4882 * @pid: the pid in question.
4884 static struct task_struct *find_process_by_pid(pid_t pid)
4886 return pid ? find_task_by_vpid(pid) : current;
4889 /* Actually do priority change: must hold rq lock. */
4891 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4893 BUG_ON(p->se.on_rq);
4896 p->rt_priority = prio;
4897 p->normal_prio = normal_prio(p);
4898 /* we are holding p->pi_lock already */
4899 p->prio = rt_mutex_getprio(p);
4900 if (rt_prio(p->prio))
4901 p->sched_class = &rt_sched_class;
4903 p->sched_class = &fair_sched_class;
4908 * check the target process has a UID that matches the current process's
4910 static bool check_same_owner(struct task_struct *p)
4912 const struct cred *cred = current_cred(), *pcred;
4916 pcred = __task_cred(p);
4917 if (cred->user->user_ns == pcred->user->user_ns)
4918 match = (cred->euid == pcred->euid ||
4919 cred->euid == pcred->uid);
4926 static int __sched_setscheduler(struct task_struct *p, int policy,
4927 const struct sched_param *param, bool user)
4929 int retval, oldprio, oldpolicy = -1, on_rq, running;
4930 unsigned long flags;
4931 const struct sched_class *prev_class;
4935 /* may grab non-irq protected spin_locks */
4936 BUG_ON(in_interrupt());
4938 /* double check policy once rq lock held */
4940 reset_on_fork = p->sched_reset_on_fork;
4941 policy = oldpolicy = p->policy;
4943 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4944 policy &= ~SCHED_RESET_ON_FORK;
4946 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4947 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4948 policy != SCHED_IDLE)
4953 * Valid priorities for SCHED_FIFO and SCHED_RR are
4954 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4955 * SCHED_BATCH and SCHED_IDLE is 0.
4957 if (param->sched_priority < 0 ||
4958 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4959 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4961 if (rt_policy(policy) != (param->sched_priority != 0))
4965 * Allow unprivileged RT tasks to decrease priority:
4967 if (user && !capable(CAP_SYS_NICE)) {
4968 if (rt_policy(policy)) {
4969 unsigned long rlim_rtprio =
4970 task_rlimit(p, RLIMIT_RTPRIO);
4972 /* can't set/change the rt policy */
4973 if (policy != p->policy && !rlim_rtprio)
4976 /* can't increase priority */
4977 if (param->sched_priority > p->rt_priority &&
4978 param->sched_priority > rlim_rtprio)
4983 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4984 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4986 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4987 if (!can_nice(p, TASK_NICE(p)))
4991 /* can't change other user's priorities */
4992 if (!check_same_owner(p))
4995 /* Normal users shall not reset the sched_reset_on_fork flag */
4996 if (p->sched_reset_on_fork && !reset_on_fork)
5001 retval = security_task_setscheduler(p);
5007 * make sure no PI-waiters arrive (or leave) while we are
5008 * changing the priority of the task:
5010 raw_spin_lock_irqsave(&p->pi_lock, flags);
5012 * To be able to change p->policy safely, the appropriate
5013 * runqueue lock must be held.
5015 rq = __task_rq_lock(p);
5018 * Changing the policy of the stop threads its a very bad idea
5020 if (p == rq->stop) {
5021 __task_rq_unlock(rq);
5022 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5027 * If not changing anything there's no need to proceed further:
5029 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5030 param->sched_priority == p->rt_priority))) {
5032 __task_rq_unlock(rq);
5033 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5037 #ifdef CONFIG_RT_GROUP_SCHED
5040 * Do not allow realtime tasks into groups that have no runtime
5043 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5044 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5045 !task_group_is_autogroup(task_group(p))) {
5046 __task_rq_unlock(rq);
5047 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5053 /* recheck policy now with rq lock held */
5054 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5055 policy = oldpolicy = -1;
5056 __task_rq_unlock(rq);
5057 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5060 on_rq = p->se.on_rq;
5061 running = task_current(rq, p);
5063 deactivate_task(rq, p, 0);
5065 p->sched_class->put_prev_task(rq, p);
5067 p->sched_reset_on_fork = reset_on_fork;
5070 prev_class = p->sched_class;
5071 __setscheduler(rq, p, policy, param->sched_priority);
5074 p->sched_class->set_curr_task(rq);
5076 activate_task(rq, p, 0);
5078 check_class_changed(rq, p, prev_class, oldprio);
5079 __task_rq_unlock(rq);
5080 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5082 rt_mutex_adjust_pi(p);
5088 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5089 * @p: the task in question.
5090 * @policy: new policy.
5091 * @param: structure containing the new RT priority.
5093 * NOTE that the task may be already dead.
5095 int sched_setscheduler(struct task_struct *p, int policy,
5096 const struct sched_param *param)
5098 return __sched_setscheduler(p, policy, param, true);
5100 EXPORT_SYMBOL_GPL(sched_setscheduler);
5103 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5104 * @p: the task in question.
5105 * @policy: new policy.
5106 * @param: structure containing the new RT priority.
5108 * Just like sched_setscheduler, only don't bother checking if the
5109 * current context has permission. For example, this is needed in
5110 * stop_machine(): we create temporary high priority worker threads,
5111 * but our caller might not have that capability.
5113 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5114 const struct sched_param *param)
5116 return __sched_setscheduler(p, policy, param, false);
5120 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5122 struct sched_param lparam;
5123 struct task_struct *p;
5126 if (!param || pid < 0)
5128 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5133 p = find_process_by_pid(pid);
5135 retval = sched_setscheduler(p, policy, &lparam);
5142 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5143 * @pid: the pid in question.
5144 * @policy: new policy.
5145 * @param: structure containing the new RT priority.
5147 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5148 struct sched_param __user *, param)
5150 /* negative values for policy are not valid */
5154 return do_sched_setscheduler(pid, policy, param);
5158 * sys_sched_setparam - set/change the RT priority of a thread
5159 * @pid: the pid in question.
5160 * @param: structure containing the new RT priority.
5162 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5164 return do_sched_setscheduler(pid, -1, param);
5168 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5169 * @pid: the pid in question.
5171 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5173 struct task_struct *p;
5181 p = find_process_by_pid(pid);
5183 retval = security_task_getscheduler(p);
5186 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5193 * sys_sched_getparam - get the RT priority of a thread
5194 * @pid: the pid in question.
5195 * @param: structure containing the RT priority.
5197 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5199 struct sched_param lp;
5200 struct task_struct *p;
5203 if (!param || pid < 0)
5207 p = find_process_by_pid(pid);
5212 retval = security_task_getscheduler(p);
5216 lp.sched_priority = p->rt_priority;
5220 * This one might sleep, we cannot do it with a spinlock held ...
5222 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5231 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5233 cpumask_var_t cpus_allowed, new_mask;
5234 struct task_struct *p;
5240 p = find_process_by_pid(pid);
5247 /* Prevent p going away */
5251 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5255 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5257 goto out_free_cpus_allowed;
5260 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5263 retval = security_task_setscheduler(p);
5267 cpuset_cpus_allowed(p, cpus_allowed);
5268 cpumask_and(new_mask, in_mask, cpus_allowed);
5270 retval = set_cpus_allowed_ptr(p, new_mask);
5273 cpuset_cpus_allowed(p, cpus_allowed);
5274 if (!cpumask_subset(new_mask, cpus_allowed)) {
5276 * We must have raced with a concurrent cpuset
5277 * update. Just reset the cpus_allowed to the
5278 * cpuset's cpus_allowed
5280 cpumask_copy(new_mask, cpus_allowed);
5285 free_cpumask_var(new_mask);
5286 out_free_cpus_allowed:
5287 free_cpumask_var(cpus_allowed);
5294 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5295 struct cpumask *new_mask)
5297 if (len < cpumask_size())
5298 cpumask_clear(new_mask);
5299 else if (len > cpumask_size())
5300 len = cpumask_size();
5302 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5306 * sys_sched_setaffinity - set the cpu affinity of a process
5307 * @pid: pid of the process
5308 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5309 * @user_mask_ptr: user-space pointer to the new cpu mask
5311 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5312 unsigned long __user *, user_mask_ptr)
5314 cpumask_var_t new_mask;
5317 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5320 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5322 retval = sched_setaffinity(pid, new_mask);
5323 free_cpumask_var(new_mask);
5327 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5329 struct task_struct *p;
5330 unsigned long flags;
5338 p = find_process_by_pid(pid);
5342 retval = security_task_getscheduler(p);
5346 rq = task_rq_lock(p, &flags);
5347 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5348 task_rq_unlock(rq, &flags);
5358 * sys_sched_getaffinity - get the cpu affinity of a process
5359 * @pid: pid of the process
5360 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5361 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5363 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5364 unsigned long __user *, user_mask_ptr)
5369 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5371 if (len & (sizeof(unsigned long)-1))
5374 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5377 ret = sched_getaffinity(pid, mask);
5379 size_t retlen = min_t(size_t, len, cpumask_size());
5381 if (copy_to_user(user_mask_ptr, mask, retlen))
5386 free_cpumask_var(mask);
5392 * sys_sched_yield - yield the current processor to other threads.
5394 * This function yields the current CPU to other tasks. If there are no
5395 * other threads running on this CPU then this function will return.
5397 SYSCALL_DEFINE0(sched_yield)
5399 struct rq *rq = this_rq_lock();
5401 schedstat_inc(rq, yld_count);
5402 current->sched_class->yield_task(rq);
5405 * Since we are going to call schedule() anyway, there's
5406 * no need to preempt or enable interrupts:
5408 __release(rq->lock);
5409 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5410 do_raw_spin_unlock(&rq->lock);
5411 preempt_enable_no_resched();
5418 static inline int should_resched(void)
5420 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5423 static void __cond_resched(void)
5425 add_preempt_count(PREEMPT_ACTIVE);
5427 sub_preempt_count(PREEMPT_ACTIVE);
5430 int __sched _cond_resched(void)
5432 if (should_resched()) {
5438 EXPORT_SYMBOL(_cond_resched);
5441 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5442 * call schedule, and on return reacquire the lock.
5444 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5445 * operations here to prevent schedule() from being called twice (once via
5446 * spin_unlock(), once by hand).
5448 int __cond_resched_lock(spinlock_t *lock)
5450 int resched = should_resched();
5453 lockdep_assert_held(lock);
5455 if (spin_needbreak(lock) || resched) {
5466 EXPORT_SYMBOL(__cond_resched_lock);
5468 int __sched __cond_resched_softirq(void)
5470 BUG_ON(!in_softirq());
5472 if (should_resched()) {
5480 EXPORT_SYMBOL(__cond_resched_softirq);
5483 * yield - yield the current processor to other threads.
5485 * This is a shortcut for kernel-space yielding - it marks the
5486 * thread runnable and calls sys_sched_yield().
5488 void __sched yield(void)
5490 set_current_state(TASK_RUNNING);
5493 EXPORT_SYMBOL(yield);
5496 * yield_to - yield the current processor to another thread in
5497 * your thread group, or accelerate that thread toward the
5498 * processor it's on.
5500 * @preempt: whether task preemption is allowed or not
5502 * It's the caller's job to ensure that the target task struct
5503 * can't go away on us before we can do any checks.
5505 * Returns true if we indeed boosted the target task.
5507 bool __sched yield_to(struct task_struct *p, bool preempt)
5509 struct task_struct *curr = current;
5510 struct rq *rq, *p_rq;
5511 unsigned long flags;
5514 local_irq_save(flags);
5519 double_rq_lock(rq, p_rq);
5520 while (task_rq(p) != p_rq) {
5521 double_rq_unlock(rq, p_rq);
5525 if (!curr->sched_class->yield_to_task)
5528 if (curr->sched_class != p->sched_class)
5531 if (task_running(p_rq, p) || p->state)
5534 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5536 schedstat_inc(rq, yld_count);
5538 * Make p's CPU reschedule; pick_next_entity takes care of
5541 if (preempt && rq != p_rq)
5542 resched_task(p_rq->curr);
5546 double_rq_unlock(rq, p_rq);
5547 local_irq_restore(flags);
5554 EXPORT_SYMBOL_GPL(yield_to);
5557 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5558 * that process accounting knows that this is a task in IO wait state.
5560 void __sched io_schedule(void)
5562 struct rq *rq = raw_rq();
5564 delayacct_blkio_start();
5565 atomic_inc(&rq->nr_iowait);
5566 blk_flush_plug(current);
5567 current->in_iowait = 1;
5569 current->in_iowait = 0;
5570 atomic_dec(&rq->nr_iowait);
5571 delayacct_blkio_end();
5573 EXPORT_SYMBOL(io_schedule);
5575 long __sched io_schedule_timeout(long timeout)
5577 struct rq *rq = raw_rq();
5580 delayacct_blkio_start();
5581 atomic_inc(&rq->nr_iowait);
5582 blk_flush_plug(current);
5583 current->in_iowait = 1;
5584 ret = schedule_timeout(timeout);
5585 current->in_iowait = 0;
5586 atomic_dec(&rq->nr_iowait);
5587 delayacct_blkio_end();
5592 * sys_sched_get_priority_max - return maximum RT priority.
5593 * @policy: scheduling class.
5595 * this syscall returns the maximum rt_priority that can be used
5596 * by a given scheduling class.
5598 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5605 ret = MAX_USER_RT_PRIO-1;
5617 * sys_sched_get_priority_min - return minimum RT priority.
5618 * @policy: scheduling class.
5620 * this syscall returns the minimum rt_priority that can be used
5621 * by a given scheduling class.
5623 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5641 * sys_sched_rr_get_interval - return the default timeslice of a process.
5642 * @pid: pid of the process.
5643 * @interval: userspace pointer to the timeslice value.
5645 * this syscall writes the default timeslice value of a given process
5646 * into the user-space timespec buffer. A value of '0' means infinity.
5648 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5649 struct timespec __user *, interval)
5651 struct task_struct *p;
5652 unsigned int time_slice;
5653 unsigned long flags;
5663 p = find_process_by_pid(pid);
5667 retval = security_task_getscheduler(p);
5671 rq = task_rq_lock(p, &flags);
5672 time_slice = p->sched_class->get_rr_interval(rq, p);
5673 task_rq_unlock(rq, &flags);
5676 jiffies_to_timespec(time_slice, &t);
5677 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5685 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5687 void sched_show_task(struct task_struct *p)
5689 unsigned long free = 0;
5692 state = p->state ? __ffs(p->state) + 1 : 0;
5693 printk(KERN_INFO "%-15.15s %c", p->comm,
5694 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5695 #if BITS_PER_LONG == 32
5696 if (state == TASK_RUNNING)
5697 printk(KERN_CONT " running ");
5699 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5701 if (state == TASK_RUNNING)
5702 printk(KERN_CONT " running task ");
5704 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5706 #ifdef CONFIG_DEBUG_STACK_USAGE
5707 free = stack_not_used(p);
5709 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5710 task_pid_nr(p), task_pid_nr(p->real_parent),
5711 (unsigned long)task_thread_info(p)->flags);
5713 show_stack(p, NULL);
5716 void show_state_filter(unsigned long state_filter)
5718 struct task_struct *g, *p;
5720 #if BITS_PER_LONG == 32
5722 " task PC stack pid father\n");
5725 " task PC stack pid father\n");
5727 read_lock(&tasklist_lock);
5728 do_each_thread(g, p) {
5730 * reset the NMI-timeout, listing all files on a slow
5731 * console might take a lot of time:
5733 touch_nmi_watchdog();
5734 if (!state_filter || (p->state & state_filter))
5736 } while_each_thread(g, p);
5738 touch_all_softlockup_watchdogs();
5740 #ifdef CONFIG_SCHED_DEBUG
5741 sysrq_sched_debug_show();
5743 read_unlock(&tasklist_lock);
5745 * Only show locks if all tasks are dumped:
5748 debug_show_all_locks();
5751 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5753 idle->sched_class = &idle_sched_class;
5757 * init_idle - set up an idle thread for a given CPU
5758 * @idle: task in question
5759 * @cpu: cpu the idle task belongs to
5761 * NOTE: this function does not set the idle thread's NEED_RESCHED
5762 * flag, to make booting more robust.
5764 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5766 struct rq *rq = cpu_rq(cpu);
5767 unsigned long flags;
5769 raw_spin_lock_irqsave(&rq->lock, flags);
5772 idle->state = TASK_RUNNING;
5773 idle->se.exec_start = sched_clock();
5775 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5777 * We're having a chicken and egg problem, even though we are
5778 * holding rq->lock, the cpu isn't yet set to this cpu so the
5779 * lockdep check in task_group() will fail.
5781 * Similar case to sched_fork(). / Alternatively we could
5782 * use task_rq_lock() here and obtain the other rq->lock.
5787 __set_task_cpu(idle, cpu);
5790 rq->curr = rq->idle = idle;
5791 #if defined(CONFIG_SMP)
5794 raw_spin_unlock_irqrestore(&rq->lock, flags);
5796 /* Set the preempt count _outside_ the spinlocks! */
5797 #if defined(CONFIG_PREEMPT)
5798 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5800 task_thread_info(idle)->preempt_count = 0;
5803 * The idle tasks have their own, simple scheduling class:
5805 idle->sched_class = &idle_sched_class;
5806 ftrace_graph_init_idle_task(idle, cpu);
5810 * In a system that switches off the HZ timer nohz_cpu_mask
5811 * indicates which cpus entered this state. This is used
5812 * in the rcu update to wait only for active cpus. For system
5813 * which do not switch off the HZ timer nohz_cpu_mask should
5814 * always be CPU_BITS_NONE.
5816 cpumask_var_t nohz_cpu_mask;
5819 * Increase the granularity value when there are more CPUs,
5820 * because with more CPUs the 'effective latency' as visible
5821 * to users decreases. But the relationship is not linear,
5822 * so pick a second-best guess by going with the log2 of the
5825 * This idea comes from the SD scheduler of Con Kolivas:
5827 static int get_update_sysctl_factor(void)
5829 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5830 unsigned int factor;
5832 switch (sysctl_sched_tunable_scaling) {
5833 case SCHED_TUNABLESCALING_NONE:
5836 case SCHED_TUNABLESCALING_LINEAR:
5839 case SCHED_TUNABLESCALING_LOG:
5841 factor = 1 + ilog2(cpus);
5848 static void update_sysctl(void)
5850 unsigned int factor = get_update_sysctl_factor();
5852 #define SET_SYSCTL(name) \
5853 (sysctl_##name = (factor) * normalized_sysctl_##name)
5854 SET_SYSCTL(sched_min_granularity);
5855 SET_SYSCTL(sched_latency);
5856 SET_SYSCTL(sched_wakeup_granularity);
5860 static inline void sched_init_granularity(void)
5867 * This is how migration works:
5869 * 1) we invoke migration_cpu_stop() on the target CPU using
5871 * 2) stopper starts to run (implicitly forcing the migrated thread
5873 * 3) it checks whether the migrated task is still in the wrong runqueue.
5874 * 4) if it's in the wrong runqueue then the migration thread removes
5875 * it and puts it into the right queue.
5876 * 5) stopper completes and stop_one_cpu() returns and the migration
5881 * Change a given task's CPU affinity. Migrate the thread to a
5882 * proper CPU and schedule it away if the CPU it's executing on
5883 * is removed from the allowed bitmask.
5885 * NOTE: the caller must have a valid reference to the task, the
5886 * task must not exit() & deallocate itself prematurely. The
5887 * call is not atomic; no spinlocks may be held.
5889 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5891 unsigned long flags;
5893 unsigned int dest_cpu;
5897 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5898 * drop the rq->lock and still rely on ->cpus_allowed.
5901 while (task_is_waking(p))
5903 rq = task_rq_lock(p, &flags);
5904 if (task_is_waking(p)) {
5905 task_rq_unlock(rq, &flags);
5909 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5914 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5915 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5920 if (p->sched_class->set_cpus_allowed)
5921 p->sched_class->set_cpus_allowed(p, new_mask);
5923 cpumask_copy(&p->cpus_allowed, new_mask);
5924 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5927 /* Can the task run on the task's current CPU? If so, we're done */
5928 if (cpumask_test_cpu(task_cpu(p), new_mask))
5931 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5932 if (migrate_task(p, rq)) {
5933 struct migration_arg arg = { p, dest_cpu };
5934 /* Need help from migration thread: drop lock and wait. */
5935 task_rq_unlock(rq, &flags);
5936 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5937 tlb_migrate_finish(p->mm);
5941 task_rq_unlock(rq, &flags);
5945 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5948 * Move (not current) task off this cpu, onto dest cpu. We're doing
5949 * this because either it can't run here any more (set_cpus_allowed()
5950 * away from this CPU, or CPU going down), or because we're
5951 * attempting to rebalance this task on exec (sched_exec).
5953 * So we race with normal scheduler movements, but that's OK, as long
5954 * as the task is no longer on this CPU.
5956 * Returns non-zero if task was successfully migrated.
5958 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5960 struct rq *rq_dest, *rq_src;
5963 if (unlikely(!cpu_active(dest_cpu)))
5966 rq_src = cpu_rq(src_cpu);
5967 rq_dest = cpu_rq(dest_cpu);
5969 double_rq_lock(rq_src, rq_dest);
5970 /* Already moved. */
5971 if (task_cpu(p) != src_cpu)
5973 /* Affinity changed (again). */
5974 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5978 * If we're not on a rq, the next wake-up will ensure we're
5982 deactivate_task(rq_src, p, 0);
5983 set_task_cpu(p, dest_cpu);
5984 activate_task(rq_dest, p, 0);
5985 check_preempt_curr(rq_dest, p, 0);
5990 double_rq_unlock(rq_src, rq_dest);
5995 * migration_cpu_stop - this will be executed by a highprio stopper thread
5996 * and performs thread migration by bumping thread off CPU then
5997 * 'pushing' onto another runqueue.
5999 static int migration_cpu_stop(void *data)
6001 struct migration_arg *arg = data;
6004 * The original target cpu might have gone down and we might
6005 * be on another cpu but it doesn't matter.
6007 local_irq_disable();
6008 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6013 #ifdef CONFIG_HOTPLUG_CPU
6016 * Ensures that the idle task is using init_mm right before its cpu goes
6019 void idle_task_exit(void)
6021 struct mm_struct *mm = current->active_mm;
6023 BUG_ON(cpu_online(smp_processor_id()));
6026 switch_mm(mm, &init_mm, current);
6031 * While a dead CPU has no uninterruptible tasks queued at this point,
6032 * it might still have a nonzero ->nr_uninterruptible counter, because
6033 * for performance reasons the counter is not stricly tracking tasks to
6034 * their home CPUs. So we just add the counter to another CPU's counter,
6035 * to keep the global sum constant after CPU-down:
6037 static void migrate_nr_uninterruptible(struct rq *rq_src)
6039 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6041 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6042 rq_src->nr_uninterruptible = 0;
6046 * remove the tasks which were accounted by rq from calc_load_tasks.
6048 static void calc_global_load_remove(struct rq *rq)
6050 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6051 rq->calc_load_active = 0;
6055 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6056 * try_to_wake_up()->select_task_rq().
6058 * Called with rq->lock held even though we'er in stop_machine() and
6059 * there's no concurrency possible, we hold the required locks anyway
6060 * because of lock validation efforts.
6062 static void migrate_tasks(unsigned int dead_cpu)
6064 struct rq *rq = cpu_rq(dead_cpu);
6065 struct task_struct *next, *stop = rq->stop;
6069 * Fudge the rq selection such that the below task selection loop
6070 * doesn't get stuck on the currently eligible stop task.
6072 * We're currently inside stop_machine() and the rq is either stuck
6073 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6074 * either way we should never end up calling schedule() until we're
6081 * There's this thread running, bail when that's the only
6084 if (rq->nr_running == 1)
6087 next = pick_next_task(rq);
6089 next->sched_class->put_prev_task(rq, next);
6091 /* Find suitable destination for @next, with force if needed. */
6092 dest_cpu = select_fallback_rq(dead_cpu, next);
6093 raw_spin_unlock(&rq->lock);
6095 __migrate_task(next, dead_cpu, dest_cpu);
6097 raw_spin_lock(&rq->lock);
6103 #endif /* CONFIG_HOTPLUG_CPU */
6105 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6107 static struct ctl_table sd_ctl_dir[] = {
6109 .procname = "sched_domain",
6115 static struct ctl_table sd_ctl_root[] = {
6117 .procname = "kernel",
6119 .child = sd_ctl_dir,
6124 static struct ctl_table *sd_alloc_ctl_entry(int n)
6126 struct ctl_table *entry =
6127 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6132 static void sd_free_ctl_entry(struct ctl_table **tablep)
6134 struct ctl_table *entry;
6137 * In the intermediate directories, both the child directory and
6138 * procname are dynamically allocated and could fail but the mode
6139 * will always be set. In the lowest directory the names are
6140 * static strings and all have proc handlers.
6142 for (entry = *tablep; entry->mode; entry++) {
6144 sd_free_ctl_entry(&entry->child);
6145 if (entry->proc_handler == NULL)
6146 kfree(entry->procname);
6154 set_table_entry(struct ctl_table *entry,
6155 const char *procname, void *data, int maxlen,
6156 mode_t mode, proc_handler *proc_handler)
6158 entry->procname = procname;
6160 entry->maxlen = maxlen;
6162 entry->proc_handler = proc_handler;
6165 static struct ctl_table *
6166 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6168 struct ctl_table *table = sd_alloc_ctl_entry(13);
6173 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6174 sizeof(long), 0644, proc_doulongvec_minmax);
6175 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6176 sizeof(long), 0644, proc_doulongvec_minmax);
6177 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6178 sizeof(int), 0644, proc_dointvec_minmax);
6179 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6180 sizeof(int), 0644, proc_dointvec_minmax);
6181 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6182 sizeof(int), 0644, proc_dointvec_minmax);
6183 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6184 sizeof(int), 0644, proc_dointvec_minmax);
6185 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6186 sizeof(int), 0644, proc_dointvec_minmax);
6187 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6188 sizeof(int), 0644, proc_dointvec_minmax);
6189 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6190 sizeof(int), 0644, proc_dointvec_minmax);
6191 set_table_entry(&table[9], "cache_nice_tries",
6192 &sd->cache_nice_tries,
6193 sizeof(int), 0644, proc_dointvec_minmax);
6194 set_table_entry(&table[10], "flags", &sd->flags,
6195 sizeof(int), 0644, proc_dointvec_minmax);
6196 set_table_entry(&table[11], "name", sd->name,
6197 CORENAME_MAX_SIZE, 0444, proc_dostring);
6198 /* &table[12] is terminator */
6203 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6205 struct ctl_table *entry, *table;
6206 struct sched_domain *sd;
6207 int domain_num = 0, i;
6210 for_each_domain(cpu, sd)
6212 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6217 for_each_domain(cpu, sd) {
6218 snprintf(buf, 32, "domain%d", i);
6219 entry->procname = kstrdup(buf, GFP_KERNEL);
6221 entry->child = sd_alloc_ctl_domain_table(sd);
6228 static struct ctl_table_header *sd_sysctl_header;
6229 static void register_sched_domain_sysctl(void)
6231 int i, cpu_num = num_possible_cpus();
6232 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6235 WARN_ON(sd_ctl_dir[0].child);
6236 sd_ctl_dir[0].child = entry;
6241 for_each_possible_cpu(i) {
6242 snprintf(buf, 32, "cpu%d", i);
6243 entry->procname = kstrdup(buf, GFP_KERNEL);
6245 entry->child = sd_alloc_ctl_cpu_table(i);
6249 WARN_ON(sd_sysctl_header);
6250 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6253 /* may be called multiple times per register */
6254 static void unregister_sched_domain_sysctl(void)
6256 if (sd_sysctl_header)
6257 unregister_sysctl_table(sd_sysctl_header);
6258 sd_sysctl_header = NULL;
6259 if (sd_ctl_dir[0].child)
6260 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6263 static void register_sched_domain_sysctl(void)
6266 static void unregister_sched_domain_sysctl(void)
6271 static void set_rq_online(struct rq *rq)
6274 const struct sched_class *class;
6276 cpumask_set_cpu(rq->cpu, rq->rd->online);
6279 for_each_class(class) {
6280 if (class->rq_online)
6281 class->rq_online(rq);
6286 static void set_rq_offline(struct rq *rq)
6289 const struct sched_class *class;
6291 for_each_class(class) {
6292 if (class->rq_offline)
6293 class->rq_offline(rq);
6296 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6302 * migration_call - callback that gets triggered when a CPU is added.
6303 * Here we can start up the necessary migration thread for the new CPU.
6305 static int __cpuinit
6306 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6308 int cpu = (long)hcpu;
6309 unsigned long flags;
6310 struct rq *rq = cpu_rq(cpu);
6312 switch (action & ~CPU_TASKS_FROZEN) {
6314 case CPU_UP_PREPARE:
6315 rq->calc_load_update = calc_load_update;
6319 /* Update our root-domain */
6320 raw_spin_lock_irqsave(&rq->lock, flags);
6322 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6326 raw_spin_unlock_irqrestore(&rq->lock, flags);
6329 #ifdef CONFIG_HOTPLUG_CPU
6331 /* Update our root-domain */
6332 raw_spin_lock_irqsave(&rq->lock, flags);
6334 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6338 BUG_ON(rq->nr_running != 1); /* the migration thread */
6339 raw_spin_unlock_irqrestore(&rq->lock, flags);
6341 migrate_nr_uninterruptible(rq);
6342 calc_global_load_remove(rq);
6347 update_max_interval();
6353 * Register at high priority so that task migration (migrate_all_tasks)
6354 * happens before everything else. This has to be lower priority than
6355 * the notifier in the perf_event subsystem, though.
6357 static struct notifier_block __cpuinitdata migration_notifier = {
6358 .notifier_call = migration_call,
6359 .priority = CPU_PRI_MIGRATION,
6362 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6363 unsigned long action, void *hcpu)
6365 switch (action & ~CPU_TASKS_FROZEN) {
6367 case CPU_DOWN_FAILED:
6368 set_cpu_active((long)hcpu, true);
6375 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6376 unsigned long action, void *hcpu)
6378 switch (action & ~CPU_TASKS_FROZEN) {
6379 case CPU_DOWN_PREPARE:
6380 set_cpu_active((long)hcpu, false);
6387 static int __init migration_init(void)
6389 void *cpu = (void *)(long)smp_processor_id();
6392 /* Initialize migration for the boot CPU */
6393 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6394 BUG_ON(err == NOTIFY_BAD);
6395 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6396 register_cpu_notifier(&migration_notifier);
6398 /* Register cpu active notifiers */
6399 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6400 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6404 early_initcall(migration_init);
6409 #ifdef CONFIG_SCHED_DEBUG
6411 static __read_mostly int sched_domain_debug_enabled;
6413 static int __init sched_domain_debug_setup(char *str)
6415 sched_domain_debug_enabled = 1;
6419 early_param("sched_debug", sched_domain_debug_setup);
6421 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6422 struct cpumask *groupmask)
6424 struct sched_group *group = sd->groups;
6427 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6428 cpumask_clear(groupmask);
6430 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6432 if (!(sd->flags & SD_LOAD_BALANCE)) {
6433 printk("does not load-balance\n");
6435 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6440 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6442 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6443 printk(KERN_ERR "ERROR: domain->span does not contain "
6446 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6447 printk(KERN_ERR "ERROR: domain->groups does not contain"
6451 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6455 printk(KERN_ERR "ERROR: group is NULL\n");
6459 if (!group->cpu_power) {
6460 printk(KERN_CONT "\n");
6461 printk(KERN_ERR "ERROR: domain->cpu_power not "
6466 if (!cpumask_weight(sched_group_cpus(group))) {
6467 printk(KERN_CONT "\n");
6468 printk(KERN_ERR "ERROR: empty group\n");
6472 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6473 printk(KERN_CONT "\n");
6474 printk(KERN_ERR "ERROR: repeated CPUs\n");
6478 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6480 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6482 printk(KERN_CONT " %s", str);
6483 if (group->cpu_power != SCHED_LOAD_SCALE) {
6484 printk(KERN_CONT " (cpu_power = %d)",
6488 group = group->next;
6489 } while (group != sd->groups);
6490 printk(KERN_CONT "\n");
6492 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6493 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6496 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6497 printk(KERN_ERR "ERROR: parent span is not a superset "
6498 "of domain->span\n");
6502 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6504 cpumask_var_t groupmask;
6507 if (!sched_domain_debug_enabled)
6511 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6515 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6517 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6518 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6523 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6530 free_cpumask_var(groupmask);
6532 #else /* !CONFIG_SCHED_DEBUG */
6533 # define sched_domain_debug(sd, cpu) do { } while (0)
6534 #endif /* CONFIG_SCHED_DEBUG */
6536 static int sd_degenerate(struct sched_domain *sd)
6538 if (cpumask_weight(sched_domain_span(sd)) == 1)
6541 /* Following flags need at least 2 groups */
6542 if (sd->flags & (SD_LOAD_BALANCE |
6543 SD_BALANCE_NEWIDLE |
6547 SD_SHARE_PKG_RESOURCES)) {
6548 if (sd->groups != sd->groups->next)
6552 /* Following flags don't use groups */
6553 if (sd->flags & (SD_WAKE_AFFINE))
6560 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6562 unsigned long cflags = sd->flags, pflags = parent->flags;
6564 if (sd_degenerate(parent))
6567 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6570 /* Flags needing groups don't count if only 1 group in parent */
6571 if (parent->groups == parent->groups->next) {
6572 pflags &= ~(SD_LOAD_BALANCE |
6573 SD_BALANCE_NEWIDLE |
6577 SD_SHARE_PKG_RESOURCES);
6578 if (nr_node_ids == 1)
6579 pflags &= ~SD_SERIALIZE;
6581 if (~cflags & pflags)
6587 static void free_rootdomain(struct root_domain *rd)
6589 synchronize_sched();
6591 cpupri_cleanup(&rd->cpupri);
6593 free_cpumask_var(rd->rto_mask);
6594 free_cpumask_var(rd->online);
6595 free_cpumask_var(rd->span);
6599 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6601 struct root_domain *old_rd = NULL;
6602 unsigned long flags;
6604 raw_spin_lock_irqsave(&rq->lock, flags);
6609 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6612 cpumask_clear_cpu(rq->cpu, old_rd->span);
6615 * If we dont want to free the old_rt yet then
6616 * set old_rd to NULL to skip the freeing later
6619 if (!atomic_dec_and_test(&old_rd->refcount))
6623 atomic_inc(&rd->refcount);
6626 cpumask_set_cpu(rq->cpu, rd->span);
6627 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6630 raw_spin_unlock_irqrestore(&rq->lock, flags);
6633 free_rootdomain(old_rd);
6636 static int init_rootdomain(struct root_domain *rd)
6638 memset(rd, 0, sizeof(*rd));
6640 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6642 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6644 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6647 if (cpupri_init(&rd->cpupri) != 0)
6652 free_cpumask_var(rd->rto_mask);
6654 free_cpumask_var(rd->online);
6656 free_cpumask_var(rd->span);
6661 static void init_defrootdomain(void)
6663 init_rootdomain(&def_root_domain);
6665 atomic_set(&def_root_domain.refcount, 1);
6668 static struct root_domain *alloc_rootdomain(void)
6670 struct root_domain *rd;
6672 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6676 if (init_rootdomain(rd) != 0) {
6685 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6686 * hold the hotplug lock.
6689 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6691 struct rq *rq = cpu_rq(cpu);
6692 struct sched_domain *tmp;
6694 for (tmp = sd; tmp; tmp = tmp->parent)
6695 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6697 /* Remove the sched domains which do not contribute to scheduling. */
6698 for (tmp = sd; tmp; ) {
6699 struct sched_domain *parent = tmp->parent;
6703 if (sd_parent_degenerate(tmp, parent)) {
6704 tmp->parent = parent->parent;
6706 parent->parent->child = tmp;
6711 if (sd && sd_degenerate(sd)) {
6717 sched_domain_debug(sd, cpu);
6719 rq_attach_root(rq, rd);
6720 rcu_assign_pointer(rq->sd, sd);
6723 /* cpus with isolated domains */
6724 static cpumask_var_t cpu_isolated_map;
6726 /* Setup the mask of cpus configured for isolated domains */
6727 static int __init isolated_cpu_setup(char *str)
6729 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6730 cpulist_parse(str, cpu_isolated_map);
6734 __setup("isolcpus=", isolated_cpu_setup);
6737 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6738 * to a function which identifies what group(along with sched group) a CPU
6739 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6740 * (due to the fact that we keep track of groups covered with a struct cpumask).
6742 * init_sched_build_groups will build a circular linked list of the groups
6743 * covered by the given span, and will set each group's ->cpumask correctly,
6744 * and ->cpu_power to 0.
6747 init_sched_build_groups(const struct cpumask *span,
6748 const struct cpumask *cpu_map,
6749 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6750 struct sched_group **sg,
6751 struct cpumask *tmpmask),
6752 struct cpumask *covered, struct cpumask *tmpmask)
6754 struct sched_group *first = NULL, *last = NULL;
6757 cpumask_clear(covered);
6759 for_each_cpu(i, span) {
6760 struct sched_group *sg;
6761 int group = group_fn(i, cpu_map, &sg, tmpmask);
6764 if (cpumask_test_cpu(i, covered))
6767 cpumask_clear(sched_group_cpus(sg));
6770 for_each_cpu(j, span) {
6771 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6774 cpumask_set_cpu(j, covered);
6775 cpumask_set_cpu(j, sched_group_cpus(sg));
6786 #define SD_NODES_PER_DOMAIN 16
6791 * find_next_best_node - find the next node to include in a sched_domain
6792 * @node: node whose sched_domain we're building
6793 * @used_nodes: nodes already in the sched_domain
6795 * Find the next node to include in a given scheduling domain. Simply
6796 * finds the closest node not already in the @used_nodes map.
6798 * Should use nodemask_t.
6800 static int find_next_best_node(int node, nodemask_t *used_nodes)
6802 int i, n, val, min_val, best_node = 0;
6806 for (i = 0; i < nr_node_ids; i++) {
6807 /* Start at @node */
6808 n = (node + i) % nr_node_ids;
6810 if (!nr_cpus_node(n))
6813 /* Skip already used nodes */
6814 if (node_isset(n, *used_nodes))
6817 /* Simple min distance search */
6818 val = node_distance(node, n);
6820 if (val < min_val) {
6826 node_set(best_node, *used_nodes);
6831 * sched_domain_node_span - get a cpumask for a node's sched_domain
6832 * @node: node whose cpumask we're constructing
6833 * @span: resulting cpumask
6835 * Given a node, construct a good cpumask for its sched_domain to span. It
6836 * should be one that prevents unnecessary balancing, but also spreads tasks
6839 static void sched_domain_node_span(int node, struct cpumask *span)
6841 nodemask_t used_nodes;
6844 cpumask_clear(span);
6845 nodes_clear(used_nodes);
6847 cpumask_or(span, span, cpumask_of_node(node));
6848 node_set(node, used_nodes);
6850 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6851 int next_node = find_next_best_node(node, &used_nodes);
6853 cpumask_or(span, span, cpumask_of_node(next_node));
6856 #endif /* CONFIG_NUMA */
6858 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6861 * The cpus mask in sched_group and sched_domain hangs off the end.
6863 * ( See the the comments in include/linux/sched.h:struct sched_group
6864 * and struct sched_domain. )
6866 struct static_sched_group {
6867 struct sched_group sg;
6868 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6871 struct static_sched_domain {
6872 struct sched_domain sd;
6873 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6879 cpumask_var_t domainspan;
6880 cpumask_var_t covered;
6881 cpumask_var_t notcovered;
6883 cpumask_var_t nodemask;
6884 cpumask_var_t this_sibling_map;
6885 cpumask_var_t this_core_map;
6886 cpumask_var_t this_book_map;
6887 cpumask_var_t send_covered;
6888 cpumask_var_t tmpmask;
6889 struct sched_group **sched_group_nodes;
6890 struct root_domain *rd;
6894 sa_sched_groups = 0,
6900 sa_this_sibling_map,
6902 sa_sched_group_nodes,
6912 * SMT sched-domains:
6914 #ifdef CONFIG_SCHED_SMT
6915 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6916 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6919 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6920 struct sched_group **sg, struct cpumask *unused)
6923 *sg = &per_cpu(sched_groups, cpu).sg;
6926 #endif /* CONFIG_SCHED_SMT */
6929 * multi-core sched-domains:
6931 #ifdef CONFIG_SCHED_MC
6932 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6933 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6936 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6937 struct sched_group **sg, struct cpumask *mask)
6940 #ifdef CONFIG_SCHED_SMT
6941 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6942 group = cpumask_first(mask);
6947 *sg = &per_cpu(sched_group_core, group).sg;
6950 #endif /* CONFIG_SCHED_MC */
6953 * book sched-domains:
6955 #ifdef CONFIG_SCHED_BOOK
6956 static DEFINE_PER_CPU(struct static_sched_domain, book_domains);
6957 static DEFINE_PER_CPU(struct static_sched_group, sched_group_book);
6960 cpu_to_book_group(int cpu, const struct cpumask *cpu_map,
6961 struct sched_group **sg, struct cpumask *mask)
6964 #ifdef CONFIG_SCHED_MC
6965 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6966 group = cpumask_first(mask);
6967 #elif defined(CONFIG_SCHED_SMT)
6968 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6969 group = cpumask_first(mask);
6972 *sg = &per_cpu(sched_group_book, group).sg;
6975 #endif /* CONFIG_SCHED_BOOK */
6977 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6978 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6981 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6982 struct sched_group **sg, struct cpumask *mask)
6985 #ifdef CONFIG_SCHED_BOOK
6986 cpumask_and(mask, cpu_book_mask(cpu), cpu_map);
6987 group = cpumask_first(mask);
6988 #elif defined(CONFIG_SCHED_MC)
6989 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6990 group = cpumask_first(mask);
6991 #elif defined(CONFIG_SCHED_SMT)
6992 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6993 group = cpumask_first(mask);
6998 *sg = &per_cpu(sched_group_phys, group).sg;
7004 * The init_sched_build_groups can't handle what we want to do with node
7005 * groups, so roll our own. Now each node has its own list of groups which
7006 * gets dynamically allocated.
7008 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
7009 static struct sched_group ***sched_group_nodes_bycpu;
7011 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
7012 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
7014 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
7015 struct sched_group **sg,
7016 struct cpumask *nodemask)
7020 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
7021 group = cpumask_first(nodemask);
7024 *sg = &per_cpu(sched_group_allnodes, group).sg;
7028 static void init_numa_sched_groups_power(struct sched_group *group_head)
7030 struct sched_group *sg = group_head;
7036 for_each_cpu(j, sched_group_cpus(sg)) {
7037 struct sched_domain *sd;
7039 sd = &per_cpu(phys_domains, j).sd;
7040 if (j != group_first_cpu(sd->groups)) {
7042 * Only add "power" once for each
7048 sg->cpu_power += sd->groups->cpu_power;
7051 } while (sg != group_head);
7054 static int build_numa_sched_groups(struct s_data *d,
7055 const struct cpumask *cpu_map, int num)
7057 struct sched_domain *sd;
7058 struct sched_group *sg, *prev;
7061 cpumask_clear(d->covered);
7062 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
7063 if (cpumask_empty(d->nodemask)) {
7064 d->sched_group_nodes[num] = NULL;
7068 sched_domain_node_span(num, d->domainspan);
7069 cpumask_and(d->domainspan, d->domainspan, cpu_map);
7071 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7074 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
7078 d->sched_group_nodes[num] = sg;
7080 for_each_cpu(j, d->nodemask) {
7081 sd = &per_cpu(node_domains, j).sd;
7086 cpumask_copy(sched_group_cpus(sg), d->nodemask);
7088 cpumask_or(d->covered, d->covered, d->nodemask);
7091 for (j = 0; j < nr_node_ids; j++) {
7092 n = (num + j) % nr_node_ids;
7093 cpumask_complement(d->notcovered, d->covered);
7094 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
7095 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
7096 if (cpumask_empty(d->tmpmask))
7098 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
7099 if (cpumask_empty(d->tmpmask))
7101 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7105 "Can not alloc domain group for node %d\n", j);
7109 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
7110 sg->next = prev->next;
7111 cpumask_or(d->covered, d->covered, d->tmpmask);
7118 #endif /* CONFIG_NUMA */
7121 /* Free memory allocated for various sched_group structures */
7122 static void free_sched_groups(const struct cpumask *cpu_map,
7123 struct cpumask *nodemask)
7127 for_each_cpu(cpu, cpu_map) {
7128 struct sched_group **sched_group_nodes
7129 = sched_group_nodes_bycpu[cpu];
7131 if (!sched_group_nodes)
7134 for (i = 0; i < nr_node_ids; i++) {
7135 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7137 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7138 if (cpumask_empty(nodemask))
7148 if (oldsg != sched_group_nodes[i])
7151 kfree(sched_group_nodes);
7152 sched_group_nodes_bycpu[cpu] = NULL;
7155 #else /* !CONFIG_NUMA */
7156 static void free_sched_groups(const struct cpumask *cpu_map,
7157 struct cpumask *nodemask)
7160 #endif /* CONFIG_NUMA */
7163 * Initialize sched groups cpu_power.
7165 * cpu_power indicates the capacity of sched group, which is used while
7166 * distributing the load between different sched groups in a sched domain.
7167 * Typically cpu_power for all the groups in a sched domain will be same unless
7168 * there are asymmetries in the topology. If there are asymmetries, group
7169 * having more cpu_power will pickup more load compared to the group having
7172 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7174 struct sched_domain *child;
7175 struct sched_group *group;
7179 WARN_ON(!sd || !sd->groups);
7181 if (cpu != group_first_cpu(sd->groups))
7184 sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups));
7188 sd->groups->cpu_power = 0;
7191 power = SCHED_LOAD_SCALE;
7192 weight = cpumask_weight(sched_domain_span(sd));
7194 * SMT siblings share the power of a single core.
7195 * Usually multiple threads get a better yield out of
7196 * that one core than a single thread would have,
7197 * reflect that in sd->smt_gain.
7199 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
7200 power *= sd->smt_gain;
7202 power >>= SCHED_LOAD_SHIFT;
7204 sd->groups->cpu_power += power;
7209 * Add cpu_power of each child group to this groups cpu_power.
7211 group = child->groups;
7213 sd->groups->cpu_power += group->cpu_power;
7214 group = group->next;
7215 } while (group != child->groups);
7219 * Initializers for schedule domains
7220 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7223 #ifdef CONFIG_SCHED_DEBUG
7224 # define SD_INIT_NAME(sd, type) sd->name = #type
7226 # define SD_INIT_NAME(sd, type) do { } while (0)
7229 #define SD_INIT(sd, type) sd_init_##type(sd)
7231 #define SD_INIT_FUNC(type) \
7232 static noinline void sd_init_##type(struct sched_domain *sd) \
7234 memset(sd, 0, sizeof(*sd)); \
7235 *sd = SD_##type##_INIT; \
7236 sd->level = SD_LV_##type; \
7237 SD_INIT_NAME(sd, type); \
7242 SD_INIT_FUNC(ALLNODES)
7245 #ifdef CONFIG_SCHED_SMT
7246 SD_INIT_FUNC(SIBLING)
7248 #ifdef CONFIG_SCHED_MC
7251 #ifdef CONFIG_SCHED_BOOK
7255 static int default_relax_domain_level = -1;
7257 static int __init setup_relax_domain_level(char *str)
7261 val = simple_strtoul(str, NULL, 0);
7262 if (val < SD_LV_MAX)
7263 default_relax_domain_level = val;
7267 __setup("relax_domain_level=", setup_relax_domain_level);
7269 static void set_domain_attribute(struct sched_domain *sd,
7270 struct sched_domain_attr *attr)
7274 if (!attr || attr->relax_domain_level < 0) {
7275 if (default_relax_domain_level < 0)
7278 request = default_relax_domain_level;
7280 request = attr->relax_domain_level;
7281 if (request < sd->level) {
7282 /* turn off idle balance on this domain */
7283 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7285 /* turn on idle balance on this domain */
7286 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7290 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7291 const struct cpumask *cpu_map)
7294 case sa_sched_groups:
7295 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
7296 d->sched_group_nodes = NULL;
7298 free_rootdomain(d->rd); /* fall through */
7300 free_cpumask_var(d->tmpmask); /* fall through */
7301 case sa_send_covered:
7302 free_cpumask_var(d->send_covered); /* fall through */
7303 case sa_this_book_map:
7304 free_cpumask_var(d->this_book_map); /* fall through */
7305 case sa_this_core_map:
7306 free_cpumask_var(d->this_core_map); /* fall through */
7307 case sa_this_sibling_map:
7308 free_cpumask_var(d->this_sibling_map); /* fall through */
7310 free_cpumask_var(d->nodemask); /* fall through */
7311 case sa_sched_group_nodes:
7313 kfree(d->sched_group_nodes); /* fall through */
7315 free_cpumask_var(d->notcovered); /* fall through */
7317 free_cpumask_var(d->covered); /* fall through */
7319 free_cpumask_var(d->domainspan); /* fall through */
7326 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7327 const struct cpumask *cpu_map)
7330 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
7332 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
7333 return sa_domainspan;
7334 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
7336 /* Allocate the per-node list of sched groups */
7337 d->sched_group_nodes = kcalloc(nr_node_ids,
7338 sizeof(struct sched_group *), GFP_KERNEL);
7339 if (!d->sched_group_nodes) {
7340 printk(KERN_WARNING "Can not alloc sched group node list\n");
7341 return sa_notcovered;
7343 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
7345 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
7346 return sa_sched_group_nodes;
7347 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
7349 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
7350 return sa_this_sibling_map;
7351 if (!alloc_cpumask_var(&d->this_book_map, GFP_KERNEL))
7352 return sa_this_core_map;
7353 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
7354 return sa_this_book_map;
7355 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
7356 return sa_send_covered;
7357 d->rd = alloc_rootdomain();
7359 printk(KERN_WARNING "Cannot alloc root domain\n");
7362 return sa_rootdomain;
7365 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
7366 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
7368 struct sched_domain *sd = NULL;
7370 struct sched_domain *parent;
7373 if (cpumask_weight(cpu_map) >
7374 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
7375 sd = &per_cpu(allnodes_domains, i).sd;
7376 SD_INIT(sd, ALLNODES);
7377 set_domain_attribute(sd, attr);
7378 cpumask_copy(sched_domain_span(sd), cpu_map);
7379 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
7384 sd = &per_cpu(node_domains, i).sd;
7386 set_domain_attribute(sd, attr);
7387 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7388 sd->parent = parent;
7391 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7396 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7397 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7398 struct sched_domain *parent, int i)
7400 struct sched_domain *sd;
7401 sd = &per_cpu(phys_domains, i).sd;
7403 set_domain_attribute(sd, attr);
7404 cpumask_copy(sched_domain_span(sd), d->nodemask);
7405 sd->parent = parent;
7408 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7412 static struct sched_domain *__build_book_sched_domain(struct s_data *d,
7413 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7414 struct sched_domain *parent, int i)
7416 struct sched_domain *sd = parent;
7417 #ifdef CONFIG_SCHED_BOOK
7418 sd = &per_cpu(book_domains, i).sd;
7420 set_domain_attribute(sd, attr);
7421 cpumask_and(sched_domain_span(sd), cpu_map, cpu_book_mask(i));
7422 sd->parent = parent;
7424 cpu_to_book_group(i, cpu_map, &sd->groups, d->tmpmask);
7429 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7430 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7431 struct sched_domain *parent, int i)
7433 struct sched_domain *sd = parent;
7434 #ifdef CONFIG_SCHED_MC
7435 sd = &per_cpu(core_domains, i).sd;
7437 set_domain_attribute(sd, attr);
7438 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7439 sd->parent = parent;
7441 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7446 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7447 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7448 struct sched_domain *parent, int i)
7450 struct sched_domain *sd = parent;
7451 #ifdef CONFIG_SCHED_SMT
7452 sd = &per_cpu(cpu_domains, i).sd;
7453 SD_INIT(sd, SIBLING);
7454 set_domain_attribute(sd, attr);
7455 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7456 sd->parent = parent;
7458 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7463 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7464 const struct cpumask *cpu_map, int cpu)
7467 #ifdef CONFIG_SCHED_SMT
7468 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7469 cpumask_and(d->this_sibling_map, cpu_map,
7470 topology_thread_cpumask(cpu));
7471 if (cpu == cpumask_first(d->this_sibling_map))
7472 init_sched_build_groups(d->this_sibling_map, cpu_map,
7474 d->send_covered, d->tmpmask);
7477 #ifdef CONFIG_SCHED_MC
7478 case SD_LV_MC: /* set up multi-core groups */
7479 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7480 if (cpu == cpumask_first(d->this_core_map))
7481 init_sched_build_groups(d->this_core_map, cpu_map,
7483 d->send_covered, d->tmpmask);
7486 #ifdef CONFIG_SCHED_BOOK
7487 case SD_LV_BOOK: /* set up book groups */
7488 cpumask_and(d->this_book_map, cpu_map, cpu_book_mask(cpu));
7489 if (cpu == cpumask_first(d->this_book_map))
7490 init_sched_build_groups(d->this_book_map, cpu_map,
7492 d->send_covered, d->tmpmask);
7495 case SD_LV_CPU: /* set up physical groups */
7496 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7497 if (!cpumask_empty(d->nodemask))
7498 init_sched_build_groups(d->nodemask, cpu_map,
7500 d->send_covered, d->tmpmask);
7503 case SD_LV_ALLNODES:
7504 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7505 d->send_covered, d->tmpmask);
7514 * Build sched domains for a given set of cpus and attach the sched domains
7515 * to the individual cpus
7517 static int __build_sched_domains(const struct cpumask *cpu_map,
7518 struct sched_domain_attr *attr)
7520 enum s_alloc alloc_state = sa_none;
7522 struct sched_domain *sd;
7528 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7529 if (alloc_state != sa_rootdomain)
7531 alloc_state = sa_sched_groups;
7534 * Set up domains for cpus specified by the cpu_map.
7536 for_each_cpu(i, cpu_map) {
7537 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7540 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7541 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7542 sd = __build_book_sched_domain(&d, cpu_map, attr, sd, i);
7543 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7544 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7547 for_each_cpu(i, cpu_map) {
7548 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7549 build_sched_groups(&d, SD_LV_BOOK, cpu_map, i);
7550 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7553 /* Set up physical groups */
7554 for (i = 0; i < nr_node_ids; i++)
7555 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7558 /* Set up node groups */
7560 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7562 for (i = 0; i < nr_node_ids; i++)
7563 if (build_numa_sched_groups(&d, cpu_map, i))
7567 /* Calculate CPU power for physical packages and nodes */
7568 #ifdef CONFIG_SCHED_SMT
7569 for_each_cpu(i, cpu_map) {
7570 sd = &per_cpu(cpu_domains, i).sd;
7571 init_sched_groups_power(i, sd);
7574 #ifdef CONFIG_SCHED_MC
7575 for_each_cpu(i, cpu_map) {
7576 sd = &per_cpu(core_domains, i).sd;
7577 init_sched_groups_power(i, sd);
7580 #ifdef CONFIG_SCHED_BOOK
7581 for_each_cpu(i, cpu_map) {
7582 sd = &per_cpu(book_domains, i).sd;
7583 init_sched_groups_power(i, sd);
7587 for_each_cpu(i, cpu_map) {
7588 sd = &per_cpu(phys_domains, i).sd;
7589 init_sched_groups_power(i, sd);
7593 for (i = 0; i < nr_node_ids; i++)
7594 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7596 if (d.sd_allnodes) {
7597 struct sched_group *sg;
7599 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7601 init_numa_sched_groups_power(sg);
7605 /* Attach the domains */
7606 for_each_cpu(i, cpu_map) {
7607 #ifdef CONFIG_SCHED_SMT
7608 sd = &per_cpu(cpu_domains, i).sd;
7609 #elif defined(CONFIG_SCHED_MC)
7610 sd = &per_cpu(core_domains, i).sd;
7611 #elif defined(CONFIG_SCHED_BOOK)
7612 sd = &per_cpu(book_domains, i).sd;
7614 sd = &per_cpu(phys_domains, i).sd;
7616 cpu_attach_domain(sd, d.rd, i);
7619 d.sched_group_nodes = NULL; /* don't free this we still need it */
7620 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7624 __free_domain_allocs(&d, alloc_state, cpu_map);
7628 static int build_sched_domains(const struct cpumask *cpu_map)
7630 return __build_sched_domains(cpu_map, NULL);
7633 static cpumask_var_t *doms_cur; /* current sched domains */
7634 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7635 static struct sched_domain_attr *dattr_cur;
7636 /* attribues of custom domains in 'doms_cur' */
7639 * Special case: If a kmalloc of a doms_cur partition (array of
7640 * cpumask) fails, then fallback to a single sched domain,
7641 * as determined by the single cpumask fallback_doms.
7643 static cpumask_var_t fallback_doms;
7646 * arch_update_cpu_topology lets virtualized architectures update the
7647 * cpu core maps. It is supposed to return 1 if the topology changed
7648 * or 0 if it stayed the same.
7650 int __attribute__((weak)) arch_update_cpu_topology(void)
7655 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7658 cpumask_var_t *doms;
7660 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7663 for (i = 0; i < ndoms; i++) {
7664 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7665 free_sched_domains(doms, i);
7672 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7675 for (i = 0; i < ndoms; i++)
7676 free_cpumask_var(doms[i]);
7681 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7682 * For now this just excludes isolated cpus, but could be used to
7683 * exclude other special cases in the future.
7685 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7689 arch_update_cpu_topology();
7691 doms_cur = alloc_sched_domains(ndoms_cur);
7693 doms_cur = &fallback_doms;
7694 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7696 err = build_sched_domains(doms_cur[0]);
7697 register_sched_domain_sysctl();
7702 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7703 struct cpumask *tmpmask)
7705 free_sched_groups(cpu_map, tmpmask);
7709 * Detach sched domains from a group of cpus specified in cpu_map
7710 * These cpus will now be attached to the NULL domain
7712 static void detach_destroy_domains(const struct cpumask *cpu_map)
7714 /* Save because hotplug lock held. */
7715 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7718 for_each_cpu(i, cpu_map)
7719 cpu_attach_domain(NULL, &def_root_domain, i);
7720 synchronize_sched();
7721 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7724 /* handle null as "default" */
7725 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7726 struct sched_domain_attr *new, int idx_new)
7728 struct sched_domain_attr tmp;
7735 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7736 new ? (new + idx_new) : &tmp,
7737 sizeof(struct sched_domain_attr));
7741 * Partition sched domains as specified by the 'ndoms_new'
7742 * cpumasks in the array doms_new[] of cpumasks. This compares
7743 * doms_new[] to the current sched domain partitioning, doms_cur[].
7744 * It destroys each deleted domain and builds each new domain.
7746 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7747 * The masks don't intersect (don't overlap.) We should setup one
7748 * sched domain for each mask. CPUs not in any of the cpumasks will
7749 * not be load balanced. If the same cpumask appears both in the
7750 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7753 * The passed in 'doms_new' should be allocated using
7754 * alloc_sched_domains. This routine takes ownership of it and will
7755 * free_sched_domains it when done with it. If the caller failed the
7756 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7757 * and partition_sched_domains() will fallback to the single partition
7758 * 'fallback_doms', it also forces the domains to be rebuilt.
7760 * If doms_new == NULL it will be replaced with cpu_online_mask.
7761 * ndoms_new == 0 is a special case for destroying existing domains,
7762 * and it will not create the default domain.
7764 * Call with hotplug lock held
7766 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7767 struct sched_domain_attr *dattr_new)
7772 mutex_lock(&sched_domains_mutex);
7774 /* always unregister in case we don't destroy any domains */
7775 unregister_sched_domain_sysctl();
7777 /* Let architecture update cpu core mappings. */
7778 new_topology = arch_update_cpu_topology();
7780 n = doms_new ? ndoms_new : 0;
7782 /* Destroy deleted domains */
7783 for (i = 0; i < ndoms_cur; i++) {
7784 for (j = 0; j < n && !new_topology; j++) {
7785 if (cpumask_equal(doms_cur[i], doms_new[j])
7786 && dattrs_equal(dattr_cur, i, dattr_new, j))
7789 /* no match - a current sched domain not in new doms_new[] */
7790 detach_destroy_domains(doms_cur[i]);
7795 if (doms_new == NULL) {
7797 doms_new = &fallback_doms;
7798 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7799 WARN_ON_ONCE(dattr_new);
7802 /* Build new domains */
7803 for (i = 0; i < ndoms_new; i++) {
7804 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7805 if (cpumask_equal(doms_new[i], doms_cur[j])
7806 && dattrs_equal(dattr_new, i, dattr_cur, j))
7809 /* no match - add a new doms_new */
7810 __build_sched_domains(doms_new[i],
7811 dattr_new ? dattr_new + i : NULL);
7816 /* Remember the new sched domains */
7817 if (doms_cur != &fallback_doms)
7818 free_sched_domains(doms_cur, ndoms_cur);
7819 kfree(dattr_cur); /* kfree(NULL) is safe */
7820 doms_cur = doms_new;
7821 dattr_cur = dattr_new;
7822 ndoms_cur = ndoms_new;
7824 register_sched_domain_sysctl();
7826 mutex_unlock(&sched_domains_mutex);
7829 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7830 static void arch_reinit_sched_domains(void)
7834 /* Destroy domains first to force the rebuild */
7835 partition_sched_domains(0, NULL, NULL);
7837 rebuild_sched_domains();
7841 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7843 unsigned int level = 0;
7845 if (sscanf(buf, "%u", &level) != 1)
7849 * level is always be positive so don't check for
7850 * level < POWERSAVINGS_BALANCE_NONE which is 0
7851 * What happens on 0 or 1 byte write,
7852 * need to check for count as well?
7855 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7859 sched_smt_power_savings = level;
7861 sched_mc_power_savings = level;
7863 arch_reinit_sched_domains();
7868 #ifdef CONFIG_SCHED_MC
7869 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7870 struct sysdev_class_attribute *attr,
7873 return sprintf(page, "%u\n", sched_mc_power_savings);
7875 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7876 struct sysdev_class_attribute *attr,
7877 const char *buf, size_t count)
7879 return sched_power_savings_store(buf, count, 0);
7881 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7882 sched_mc_power_savings_show,
7883 sched_mc_power_savings_store);
7886 #ifdef CONFIG_SCHED_SMT
7887 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7888 struct sysdev_class_attribute *attr,
7891 return sprintf(page, "%u\n", sched_smt_power_savings);
7893 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7894 struct sysdev_class_attribute *attr,
7895 const char *buf, size_t count)
7897 return sched_power_savings_store(buf, count, 1);
7899 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7900 sched_smt_power_savings_show,
7901 sched_smt_power_savings_store);
7904 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7908 #ifdef CONFIG_SCHED_SMT
7910 err = sysfs_create_file(&cls->kset.kobj,
7911 &attr_sched_smt_power_savings.attr);
7913 #ifdef CONFIG_SCHED_MC
7914 if (!err && mc_capable())
7915 err = sysfs_create_file(&cls->kset.kobj,
7916 &attr_sched_mc_power_savings.attr);
7920 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7923 * Update cpusets according to cpu_active mask. If cpusets are
7924 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7925 * around partition_sched_domains().
7927 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7930 switch (action & ~CPU_TASKS_FROZEN) {
7932 case CPU_DOWN_FAILED:
7933 cpuset_update_active_cpus();
7940 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7943 switch (action & ~CPU_TASKS_FROZEN) {
7944 case CPU_DOWN_PREPARE:
7945 cpuset_update_active_cpus();
7952 static int update_runtime(struct notifier_block *nfb,
7953 unsigned long action, void *hcpu)
7955 int cpu = (int)(long)hcpu;
7958 case CPU_DOWN_PREPARE:
7959 case CPU_DOWN_PREPARE_FROZEN:
7960 disable_runtime(cpu_rq(cpu));
7963 case CPU_DOWN_FAILED:
7964 case CPU_DOWN_FAILED_FROZEN:
7966 case CPU_ONLINE_FROZEN:
7967 enable_runtime(cpu_rq(cpu));
7975 void __init sched_init_smp(void)
7977 cpumask_var_t non_isolated_cpus;
7979 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7980 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7982 #if defined(CONFIG_NUMA)
7983 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7985 BUG_ON(sched_group_nodes_bycpu == NULL);
7988 mutex_lock(&sched_domains_mutex);
7989 arch_init_sched_domains(cpu_active_mask);
7990 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7991 if (cpumask_empty(non_isolated_cpus))
7992 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7993 mutex_unlock(&sched_domains_mutex);
7996 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7997 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7999 /* RT runtime code needs to handle some hotplug events */
8000 hotcpu_notifier(update_runtime, 0);
8004 /* Move init over to a non-isolated CPU */
8005 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8007 sched_init_granularity();
8008 free_cpumask_var(non_isolated_cpus);
8010 init_sched_rt_class();
8013 void __init sched_init_smp(void)
8015 sched_init_granularity();
8017 #endif /* CONFIG_SMP */
8019 const_debug unsigned int sysctl_timer_migration = 1;
8021 int in_sched_functions(unsigned long addr)
8023 return in_lock_functions(addr) ||
8024 (addr >= (unsigned long)__sched_text_start
8025 && addr < (unsigned long)__sched_text_end);
8028 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8030 cfs_rq->tasks_timeline = RB_ROOT;
8031 INIT_LIST_HEAD(&cfs_rq->tasks);
8032 #ifdef CONFIG_FAIR_GROUP_SCHED
8034 /* allow initial update_cfs_load() to truncate */
8036 cfs_rq->load_stamp = 1;
8039 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8042 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8044 struct rt_prio_array *array;
8047 array = &rt_rq->active;
8048 for (i = 0; i < MAX_RT_PRIO; i++) {
8049 INIT_LIST_HEAD(array->queue + i);
8050 __clear_bit(i, array->bitmap);
8052 /* delimiter for bitsearch: */
8053 __set_bit(MAX_RT_PRIO, array->bitmap);
8055 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8056 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8058 rt_rq->highest_prio.next = MAX_RT_PRIO;
8062 rt_rq->rt_nr_migratory = 0;
8063 rt_rq->overloaded = 0;
8064 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
8068 rt_rq->rt_throttled = 0;
8069 rt_rq->rt_runtime = 0;
8070 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
8072 #ifdef CONFIG_RT_GROUP_SCHED
8073 rt_rq->rt_nr_boosted = 0;
8078 #ifdef CONFIG_FAIR_GROUP_SCHED
8079 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8080 struct sched_entity *se, int cpu,
8081 struct sched_entity *parent)
8083 struct rq *rq = cpu_rq(cpu);
8084 tg->cfs_rq[cpu] = cfs_rq;
8085 init_cfs_rq(cfs_rq, rq);
8089 /* se could be NULL for root_task_group */
8094 se->cfs_rq = &rq->cfs;
8096 se->cfs_rq = parent->my_q;
8099 update_load_set(&se->load, 0);
8100 se->parent = parent;
8104 #ifdef CONFIG_RT_GROUP_SCHED
8105 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8106 struct sched_rt_entity *rt_se, int cpu,
8107 struct sched_rt_entity *parent)
8109 struct rq *rq = cpu_rq(cpu);
8111 tg->rt_rq[cpu] = rt_rq;
8112 init_rt_rq(rt_rq, rq);
8114 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8116 tg->rt_se[cpu] = rt_se;
8121 rt_se->rt_rq = &rq->rt;
8123 rt_se->rt_rq = parent->my_q;
8125 rt_se->my_q = rt_rq;
8126 rt_se->parent = parent;
8127 INIT_LIST_HEAD(&rt_se->run_list);
8131 void __init sched_init(void)
8134 unsigned long alloc_size = 0, ptr;
8136 #ifdef CONFIG_FAIR_GROUP_SCHED
8137 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8139 #ifdef CONFIG_RT_GROUP_SCHED
8140 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8142 #ifdef CONFIG_CPUMASK_OFFSTACK
8143 alloc_size += num_possible_cpus() * cpumask_size();
8146 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8148 #ifdef CONFIG_FAIR_GROUP_SCHED
8149 root_task_group.se = (struct sched_entity **)ptr;
8150 ptr += nr_cpu_ids * sizeof(void **);
8152 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8153 ptr += nr_cpu_ids * sizeof(void **);
8155 #endif /* CONFIG_FAIR_GROUP_SCHED */
8156 #ifdef CONFIG_RT_GROUP_SCHED
8157 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8158 ptr += nr_cpu_ids * sizeof(void **);
8160 root_task_group.rt_rq = (struct rt_rq **)ptr;
8161 ptr += nr_cpu_ids * sizeof(void **);
8163 #endif /* CONFIG_RT_GROUP_SCHED */
8164 #ifdef CONFIG_CPUMASK_OFFSTACK
8165 for_each_possible_cpu(i) {
8166 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8167 ptr += cpumask_size();
8169 #endif /* CONFIG_CPUMASK_OFFSTACK */
8173 init_defrootdomain();
8176 init_rt_bandwidth(&def_rt_bandwidth,
8177 global_rt_period(), global_rt_runtime());
8179 #ifdef CONFIG_RT_GROUP_SCHED
8180 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8181 global_rt_period(), global_rt_runtime());
8182 #endif /* CONFIG_RT_GROUP_SCHED */
8184 #ifdef CONFIG_CGROUP_SCHED
8185 list_add(&root_task_group.list, &task_groups);
8186 INIT_LIST_HEAD(&root_task_group.children);
8187 autogroup_init(&init_task);
8188 #endif /* CONFIG_CGROUP_SCHED */
8190 for_each_possible_cpu(i) {
8194 raw_spin_lock_init(&rq->lock);
8196 rq->calc_load_active = 0;
8197 rq->calc_load_update = jiffies + LOAD_FREQ;
8198 init_cfs_rq(&rq->cfs, rq);
8199 init_rt_rq(&rq->rt, rq);
8200 #ifdef CONFIG_FAIR_GROUP_SCHED
8201 root_task_group.shares = root_task_group_load;
8202 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8204 * How much cpu bandwidth does root_task_group get?
8206 * In case of task-groups formed thr' the cgroup filesystem, it
8207 * gets 100% of the cpu resources in the system. This overall
8208 * system cpu resource is divided among the tasks of
8209 * root_task_group and its child task-groups in a fair manner,
8210 * based on each entity's (task or task-group's) weight
8211 * (se->load.weight).
8213 * In other words, if root_task_group has 10 tasks of weight
8214 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8215 * then A0's share of the cpu resource is:
8217 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8219 * We achieve this by letting root_task_group's tasks sit
8220 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8222 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8223 #endif /* CONFIG_FAIR_GROUP_SCHED */
8225 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8226 #ifdef CONFIG_RT_GROUP_SCHED
8227 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8228 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8231 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8232 rq->cpu_load[j] = 0;
8234 rq->last_load_update_tick = jiffies;
8239 rq->cpu_power = SCHED_LOAD_SCALE;
8240 rq->post_schedule = 0;
8241 rq->active_balance = 0;
8242 rq->next_balance = jiffies;
8247 rq->avg_idle = 2*sysctl_sched_migration_cost;
8248 rq_attach_root(rq, &def_root_domain);
8250 rq->nohz_balance_kick = 0;
8251 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8255 atomic_set(&rq->nr_iowait, 0);
8258 set_load_weight(&init_task);
8260 #ifdef CONFIG_PREEMPT_NOTIFIERS
8261 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8265 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8268 #ifdef CONFIG_RT_MUTEXES
8269 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
8273 * The boot idle thread does lazy MMU switching as well:
8275 atomic_inc(&init_mm.mm_count);
8276 enter_lazy_tlb(&init_mm, current);
8279 * Make us the idle thread. Technically, schedule() should not be
8280 * called from this thread, however somewhere below it might be,
8281 * but because we are the idle thread, we just pick up running again
8282 * when this runqueue becomes "idle".
8284 init_idle(current, smp_processor_id());
8286 calc_load_update = jiffies + LOAD_FREQ;
8289 * During early bootup we pretend to be a normal task:
8291 current->sched_class = &fair_sched_class;
8293 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8294 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8297 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8298 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8299 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8300 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8301 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8303 /* May be allocated at isolcpus cmdline parse time */
8304 if (cpu_isolated_map == NULL)
8305 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8308 scheduler_running = 1;
8311 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8312 static inline int preempt_count_equals(int preempt_offset)
8314 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8316 return (nested == preempt_offset);
8319 void __might_sleep(const char *file, int line, int preempt_offset)
8322 static unsigned long prev_jiffy; /* ratelimiting */
8324 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8325 system_state != SYSTEM_RUNNING || oops_in_progress)
8327 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8329 prev_jiffy = jiffies;
8332 "BUG: sleeping function called from invalid context at %s:%d\n",
8335 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8336 in_atomic(), irqs_disabled(),
8337 current->pid, current->comm);
8339 debug_show_held_locks(current);
8340 if (irqs_disabled())
8341 print_irqtrace_events(current);
8345 EXPORT_SYMBOL(__might_sleep);
8348 #ifdef CONFIG_MAGIC_SYSRQ
8349 static void normalize_task(struct rq *rq, struct task_struct *p)
8351 const struct sched_class *prev_class = p->sched_class;
8352 int old_prio = p->prio;
8355 on_rq = p->se.on_rq;
8357 deactivate_task(rq, p, 0);
8358 __setscheduler(rq, p, SCHED_NORMAL, 0);
8360 activate_task(rq, p, 0);
8361 resched_task(rq->curr);
8364 check_class_changed(rq, p, prev_class, old_prio);
8367 void normalize_rt_tasks(void)
8369 struct task_struct *g, *p;
8370 unsigned long flags;
8373 read_lock_irqsave(&tasklist_lock, flags);
8374 do_each_thread(g, p) {
8376 * Only normalize user tasks:
8381 p->se.exec_start = 0;
8382 #ifdef CONFIG_SCHEDSTATS
8383 p->se.statistics.wait_start = 0;
8384 p->se.statistics.sleep_start = 0;
8385 p->se.statistics.block_start = 0;
8390 * Renice negative nice level userspace
8393 if (TASK_NICE(p) < 0 && p->mm)
8394 set_user_nice(p, 0);
8398 raw_spin_lock(&p->pi_lock);
8399 rq = __task_rq_lock(p);
8401 normalize_task(rq, p);
8403 __task_rq_unlock(rq);
8404 raw_spin_unlock(&p->pi_lock);
8405 } while_each_thread(g, p);
8407 read_unlock_irqrestore(&tasklist_lock, flags);
8410 #endif /* CONFIG_MAGIC_SYSRQ */
8412 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8414 * These functions are only useful for the IA64 MCA handling, or kdb.
8416 * They can only be called when the whole system has been
8417 * stopped - every CPU needs to be quiescent, and no scheduling
8418 * activity can take place. Using them for anything else would
8419 * be a serious bug, and as a result, they aren't even visible
8420 * under any other configuration.
8424 * curr_task - return the current task for a given cpu.
8425 * @cpu: the processor in question.
8427 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8429 struct task_struct *curr_task(int cpu)
8431 return cpu_curr(cpu);
8434 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8438 * set_curr_task - set the current task for a given cpu.
8439 * @cpu: the processor in question.
8440 * @p: the task pointer to set.
8442 * Description: This function must only be used when non-maskable interrupts
8443 * are serviced on a separate stack. It allows the architecture to switch the
8444 * notion of the current task on a cpu in a non-blocking manner. This function
8445 * must be called with all CPU's synchronized, and interrupts disabled, the
8446 * and caller must save the original value of the current task (see
8447 * curr_task() above) and restore that value before reenabling interrupts and
8448 * re-starting the system.
8450 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8452 void set_curr_task(int cpu, struct task_struct *p)
8459 #ifdef CONFIG_FAIR_GROUP_SCHED
8460 static void free_fair_sched_group(struct task_group *tg)
8464 for_each_possible_cpu(i) {
8466 kfree(tg->cfs_rq[i]);
8476 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8478 struct cfs_rq *cfs_rq;
8479 struct sched_entity *se;
8482 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8485 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8489 tg->shares = NICE_0_LOAD;
8491 for_each_possible_cpu(i) {
8492 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8493 GFP_KERNEL, cpu_to_node(i));
8497 se = kzalloc_node(sizeof(struct sched_entity),
8498 GFP_KERNEL, cpu_to_node(i));
8502 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8513 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8515 struct rq *rq = cpu_rq(cpu);
8516 unsigned long flags;
8519 * Only empty task groups can be destroyed; so we can speculatively
8520 * check on_list without danger of it being re-added.
8522 if (!tg->cfs_rq[cpu]->on_list)
8525 raw_spin_lock_irqsave(&rq->lock, flags);
8526 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8527 raw_spin_unlock_irqrestore(&rq->lock, flags);
8529 #else /* !CONFG_FAIR_GROUP_SCHED */
8530 static inline void free_fair_sched_group(struct task_group *tg)
8535 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8540 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8543 #endif /* CONFIG_FAIR_GROUP_SCHED */
8545 #ifdef CONFIG_RT_GROUP_SCHED
8546 static void free_rt_sched_group(struct task_group *tg)
8550 destroy_rt_bandwidth(&tg->rt_bandwidth);
8552 for_each_possible_cpu(i) {
8554 kfree(tg->rt_rq[i]);
8556 kfree(tg->rt_se[i]);
8564 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8566 struct rt_rq *rt_rq;
8567 struct sched_rt_entity *rt_se;
8571 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8574 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8578 init_rt_bandwidth(&tg->rt_bandwidth,
8579 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8581 for_each_possible_cpu(i) {
8584 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8585 GFP_KERNEL, cpu_to_node(i));
8589 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8590 GFP_KERNEL, cpu_to_node(i));
8594 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8604 #else /* !CONFIG_RT_GROUP_SCHED */
8605 static inline void free_rt_sched_group(struct task_group *tg)
8610 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8614 #endif /* CONFIG_RT_GROUP_SCHED */
8616 #ifdef CONFIG_CGROUP_SCHED
8617 static void free_sched_group(struct task_group *tg)
8619 free_fair_sched_group(tg);
8620 free_rt_sched_group(tg);
8625 /* allocate runqueue etc for a new task group */
8626 struct task_group *sched_create_group(struct task_group *parent)
8628 struct task_group *tg;
8629 unsigned long flags;
8631 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8633 return ERR_PTR(-ENOMEM);
8635 if (!alloc_fair_sched_group(tg, parent))
8638 if (!alloc_rt_sched_group(tg, parent))
8641 spin_lock_irqsave(&task_group_lock, flags);
8642 list_add_rcu(&tg->list, &task_groups);
8644 WARN_ON(!parent); /* root should already exist */
8646 tg->parent = parent;
8647 INIT_LIST_HEAD(&tg->children);
8648 list_add_rcu(&tg->siblings, &parent->children);
8649 spin_unlock_irqrestore(&task_group_lock, flags);
8654 free_sched_group(tg);
8655 return ERR_PTR(-ENOMEM);
8658 /* rcu callback to free various structures associated with a task group */
8659 static void free_sched_group_rcu(struct rcu_head *rhp)
8661 /* now it should be safe to free those cfs_rqs */
8662 free_sched_group(container_of(rhp, struct task_group, rcu));
8665 /* Destroy runqueue etc associated with a task group */
8666 void sched_destroy_group(struct task_group *tg)
8668 unsigned long flags;
8671 /* end participation in shares distribution */
8672 for_each_possible_cpu(i)
8673 unregister_fair_sched_group(tg, i);
8675 spin_lock_irqsave(&task_group_lock, flags);
8676 list_del_rcu(&tg->list);
8677 list_del_rcu(&tg->siblings);
8678 spin_unlock_irqrestore(&task_group_lock, flags);
8680 /* wait for possible concurrent references to cfs_rqs complete */
8681 call_rcu(&tg->rcu, free_sched_group_rcu);
8684 /* change task's runqueue when it moves between groups.
8685 * The caller of this function should have put the task in its new group
8686 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8687 * reflect its new group.
8689 void sched_move_task(struct task_struct *tsk)
8692 unsigned long flags;
8695 rq = task_rq_lock(tsk, &flags);
8697 running = task_current(rq, tsk);
8698 on_rq = tsk->se.on_rq;
8701 dequeue_task(rq, tsk, 0);
8702 if (unlikely(running))
8703 tsk->sched_class->put_prev_task(rq, tsk);
8705 #ifdef CONFIG_FAIR_GROUP_SCHED
8706 if (tsk->sched_class->task_move_group)
8707 tsk->sched_class->task_move_group(tsk, on_rq);
8710 set_task_rq(tsk, task_cpu(tsk));
8712 if (unlikely(running))
8713 tsk->sched_class->set_curr_task(rq);
8715 enqueue_task(rq, tsk, 0);
8717 task_rq_unlock(rq, &flags);
8719 #endif /* CONFIG_CGROUP_SCHED */
8721 #ifdef CONFIG_FAIR_GROUP_SCHED
8722 static DEFINE_MUTEX(shares_mutex);
8724 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8727 unsigned long flags;
8730 * We can't change the weight of the root cgroup.
8735 if (shares < MIN_SHARES)
8736 shares = MIN_SHARES;
8737 else if (shares > MAX_SHARES)
8738 shares = MAX_SHARES;
8740 mutex_lock(&shares_mutex);
8741 if (tg->shares == shares)
8744 tg->shares = shares;
8745 for_each_possible_cpu(i) {
8746 struct rq *rq = cpu_rq(i);
8747 struct sched_entity *se;
8750 /* Propagate contribution to hierarchy */
8751 raw_spin_lock_irqsave(&rq->lock, flags);
8752 for_each_sched_entity(se)
8753 update_cfs_shares(group_cfs_rq(se));
8754 raw_spin_unlock_irqrestore(&rq->lock, flags);
8758 mutex_unlock(&shares_mutex);
8762 unsigned long sched_group_shares(struct task_group *tg)
8768 #ifdef CONFIG_RT_GROUP_SCHED
8770 * Ensure that the real time constraints are schedulable.
8772 static DEFINE_MUTEX(rt_constraints_mutex);
8774 static unsigned long to_ratio(u64 period, u64 runtime)
8776 if (runtime == RUNTIME_INF)
8779 return div64_u64(runtime << 20, period);
8782 /* Must be called with tasklist_lock held */
8783 static inline int tg_has_rt_tasks(struct task_group *tg)
8785 struct task_struct *g, *p;
8787 do_each_thread(g, p) {
8788 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8790 } while_each_thread(g, p);
8795 struct rt_schedulable_data {
8796 struct task_group *tg;
8801 static int tg_schedulable(struct task_group *tg, void *data)
8803 struct rt_schedulable_data *d = data;
8804 struct task_group *child;
8805 unsigned long total, sum = 0;
8806 u64 period, runtime;
8808 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8809 runtime = tg->rt_bandwidth.rt_runtime;
8812 period = d->rt_period;
8813 runtime = d->rt_runtime;
8817 * Cannot have more runtime than the period.
8819 if (runtime > period && runtime != RUNTIME_INF)
8823 * Ensure we don't starve existing RT tasks.
8825 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8828 total = to_ratio(period, runtime);
8831 * Nobody can have more than the global setting allows.
8833 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8837 * The sum of our children's runtime should not exceed our own.
8839 list_for_each_entry_rcu(child, &tg->children, siblings) {
8840 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8841 runtime = child->rt_bandwidth.rt_runtime;
8843 if (child == d->tg) {
8844 period = d->rt_period;
8845 runtime = d->rt_runtime;
8848 sum += to_ratio(period, runtime);
8857 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8859 struct rt_schedulable_data data = {
8861 .rt_period = period,
8862 .rt_runtime = runtime,
8865 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8868 static int tg_set_bandwidth(struct task_group *tg,
8869 u64 rt_period, u64 rt_runtime)
8873 mutex_lock(&rt_constraints_mutex);
8874 read_lock(&tasklist_lock);
8875 err = __rt_schedulable(tg, rt_period, rt_runtime);
8879 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8880 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8881 tg->rt_bandwidth.rt_runtime = rt_runtime;
8883 for_each_possible_cpu(i) {
8884 struct rt_rq *rt_rq = tg->rt_rq[i];
8886 raw_spin_lock(&rt_rq->rt_runtime_lock);
8887 rt_rq->rt_runtime = rt_runtime;
8888 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8890 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8892 read_unlock(&tasklist_lock);
8893 mutex_unlock(&rt_constraints_mutex);
8898 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8900 u64 rt_runtime, rt_period;
8902 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8903 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8904 if (rt_runtime_us < 0)
8905 rt_runtime = RUNTIME_INF;
8907 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8910 long sched_group_rt_runtime(struct task_group *tg)
8914 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8917 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8918 do_div(rt_runtime_us, NSEC_PER_USEC);
8919 return rt_runtime_us;
8922 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8924 u64 rt_runtime, rt_period;
8926 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8927 rt_runtime = tg->rt_bandwidth.rt_runtime;
8932 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8935 long sched_group_rt_period(struct task_group *tg)
8939 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8940 do_div(rt_period_us, NSEC_PER_USEC);
8941 return rt_period_us;
8944 static int sched_rt_global_constraints(void)
8946 u64 runtime, period;
8949 if (sysctl_sched_rt_period <= 0)
8952 runtime = global_rt_runtime();
8953 period = global_rt_period();
8956 * Sanity check on the sysctl variables.
8958 if (runtime > period && runtime != RUNTIME_INF)
8961 mutex_lock(&rt_constraints_mutex);
8962 read_lock(&tasklist_lock);
8963 ret = __rt_schedulable(NULL, 0, 0);
8964 read_unlock(&tasklist_lock);
8965 mutex_unlock(&rt_constraints_mutex);
8970 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8972 /* Don't accept realtime tasks when there is no way for them to run */
8973 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8979 #else /* !CONFIG_RT_GROUP_SCHED */
8980 static int sched_rt_global_constraints(void)
8982 unsigned long flags;
8985 if (sysctl_sched_rt_period <= 0)
8989 * There's always some RT tasks in the root group
8990 * -- migration, kstopmachine etc..
8992 if (sysctl_sched_rt_runtime == 0)
8995 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8996 for_each_possible_cpu(i) {
8997 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8999 raw_spin_lock(&rt_rq->rt_runtime_lock);
9000 rt_rq->rt_runtime = global_rt_runtime();
9001 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9003 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9007 #endif /* CONFIG_RT_GROUP_SCHED */
9009 int sched_rt_handler(struct ctl_table *table, int write,
9010 void __user *buffer, size_t *lenp,
9014 int old_period, old_runtime;
9015 static DEFINE_MUTEX(mutex);
9018 old_period = sysctl_sched_rt_period;
9019 old_runtime = sysctl_sched_rt_runtime;
9021 ret = proc_dointvec(table, write, buffer, lenp, ppos);
9023 if (!ret && write) {
9024 ret = sched_rt_global_constraints();
9026 sysctl_sched_rt_period = old_period;
9027 sysctl_sched_rt_runtime = old_runtime;
9029 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9030 def_rt_bandwidth.rt_period =
9031 ns_to_ktime(global_rt_period());
9034 mutex_unlock(&mutex);
9039 #ifdef CONFIG_CGROUP_SCHED
9041 /* return corresponding task_group object of a cgroup */
9042 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9044 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9045 struct task_group, css);
9048 static struct cgroup_subsys_state *
9049 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9051 struct task_group *tg, *parent;
9053 if (!cgrp->parent) {
9054 /* This is early initialization for the top cgroup */
9055 return &root_task_group.css;
9058 parent = cgroup_tg(cgrp->parent);
9059 tg = sched_create_group(parent);
9061 return ERR_PTR(-ENOMEM);
9067 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9069 struct task_group *tg = cgroup_tg(cgrp);
9071 sched_destroy_group(tg);
9075 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9077 #ifdef CONFIG_RT_GROUP_SCHED
9078 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9081 /* We don't support RT-tasks being in separate groups */
9082 if (tsk->sched_class != &fair_sched_class)
9089 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9090 struct task_struct *tsk, bool threadgroup)
9092 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
9096 struct task_struct *c;
9098 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
9099 retval = cpu_cgroup_can_attach_task(cgrp, c);
9111 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9112 struct cgroup *old_cont, struct task_struct *tsk,
9115 sched_move_task(tsk);
9117 struct task_struct *c;
9119 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
9127 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
9128 struct cgroup *old_cgrp, struct task_struct *task)
9131 * cgroup_exit() is called in the copy_process() failure path.
9132 * Ignore this case since the task hasn't ran yet, this avoids
9133 * trying to poke a half freed task state from generic code.
9135 if (!(task->flags & PF_EXITING))
9138 sched_move_task(task);
9141 #ifdef CONFIG_FAIR_GROUP_SCHED
9142 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9145 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9148 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9150 struct task_group *tg = cgroup_tg(cgrp);
9152 return (u64) tg->shares;
9154 #endif /* CONFIG_FAIR_GROUP_SCHED */
9156 #ifdef CONFIG_RT_GROUP_SCHED
9157 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9160 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9163 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9165 return sched_group_rt_runtime(cgroup_tg(cgrp));
9168 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9171 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9174 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9176 return sched_group_rt_period(cgroup_tg(cgrp));
9178 #endif /* CONFIG_RT_GROUP_SCHED */
9180 static struct cftype cpu_files[] = {
9181 #ifdef CONFIG_FAIR_GROUP_SCHED
9184 .read_u64 = cpu_shares_read_u64,
9185 .write_u64 = cpu_shares_write_u64,
9188 #ifdef CONFIG_RT_GROUP_SCHED
9190 .name = "rt_runtime_us",
9191 .read_s64 = cpu_rt_runtime_read,
9192 .write_s64 = cpu_rt_runtime_write,
9195 .name = "rt_period_us",
9196 .read_u64 = cpu_rt_period_read_uint,
9197 .write_u64 = cpu_rt_period_write_uint,
9202 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9204 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9207 struct cgroup_subsys cpu_cgroup_subsys = {
9209 .create = cpu_cgroup_create,
9210 .destroy = cpu_cgroup_destroy,
9211 .can_attach = cpu_cgroup_can_attach,
9212 .attach = cpu_cgroup_attach,
9213 .exit = cpu_cgroup_exit,
9214 .populate = cpu_cgroup_populate,
9215 .subsys_id = cpu_cgroup_subsys_id,
9219 #endif /* CONFIG_CGROUP_SCHED */
9221 #ifdef CONFIG_CGROUP_CPUACCT
9224 * CPU accounting code for task groups.
9226 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9227 * (balbir@in.ibm.com).
9230 /* track cpu usage of a group of tasks and its child groups */
9232 struct cgroup_subsys_state css;
9233 /* cpuusage holds pointer to a u64-type object on every cpu */
9234 u64 __percpu *cpuusage;
9235 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9236 struct cpuacct *parent;
9239 struct cgroup_subsys cpuacct_subsys;
9241 /* return cpu accounting group corresponding to this container */
9242 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9244 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9245 struct cpuacct, css);
9248 /* return cpu accounting group to which this task belongs */
9249 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9251 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9252 struct cpuacct, css);
9255 /* create a new cpu accounting group */
9256 static struct cgroup_subsys_state *cpuacct_create(
9257 struct cgroup_subsys *ss, struct cgroup *cgrp)
9259 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9265 ca->cpuusage = alloc_percpu(u64);
9269 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9270 if (percpu_counter_init(&ca->cpustat[i], 0))
9271 goto out_free_counters;
9274 ca->parent = cgroup_ca(cgrp->parent);
9280 percpu_counter_destroy(&ca->cpustat[i]);
9281 free_percpu(ca->cpuusage);
9285 return ERR_PTR(-ENOMEM);
9288 /* destroy an existing cpu accounting group */
9290 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9292 struct cpuacct *ca = cgroup_ca(cgrp);
9295 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9296 percpu_counter_destroy(&ca->cpustat[i]);
9297 free_percpu(ca->cpuusage);
9301 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9303 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9306 #ifndef CONFIG_64BIT
9308 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9310 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9312 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9320 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9322 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9324 #ifndef CONFIG_64BIT
9326 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9328 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9330 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9336 /* return total cpu usage (in nanoseconds) of a group */
9337 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9339 struct cpuacct *ca = cgroup_ca(cgrp);
9340 u64 totalcpuusage = 0;
9343 for_each_present_cpu(i)
9344 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9346 return totalcpuusage;
9349 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9352 struct cpuacct *ca = cgroup_ca(cgrp);
9361 for_each_present_cpu(i)
9362 cpuacct_cpuusage_write(ca, i, 0);
9368 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9371 struct cpuacct *ca = cgroup_ca(cgroup);
9375 for_each_present_cpu(i) {
9376 percpu = cpuacct_cpuusage_read(ca, i);
9377 seq_printf(m, "%llu ", (unsigned long long) percpu);
9379 seq_printf(m, "\n");
9383 static const char *cpuacct_stat_desc[] = {
9384 [CPUACCT_STAT_USER] = "user",
9385 [CPUACCT_STAT_SYSTEM] = "system",
9388 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9389 struct cgroup_map_cb *cb)
9391 struct cpuacct *ca = cgroup_ca(cgrp);
9394 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9395 s64 val = percpu_counter_read(&ca->cpustat[i]);
9396 val = cputime64_to_clock_t(val);
9397 cb->fill(cb, cpuacct_stat_desc[i], val);
9402 static struct cftype files[] = {
9405 .read_u64 = cpuusage_read,
9406 .write_u64 = cpuusage_write,
9409 .name = "usage_percpu",
9410 .read_seq_string = cpuacct_percpu_seq_read,
9414 .read_map = cpuacct_stats_show,
9418 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9420 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9424 * charge this task's execution time to its accounting group.
9426 * called with rq->lock held.
9428 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9433 if (unlikely(!cpuacct_subsys.active))
9436 cpu = task_cpu(tsk);
9442 for (; ca; ca = ca->parent) {
9443 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9444 *cpuusage += cputime;
9451 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9452 * in cputime_t units. As a result, cpuacct_update_stats calls
9453 * percpu_counter_add with values large enough to always overflow the
9454 * per cpu batch limit causing bad SMP scalability.
9456 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9457 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9458 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9461 #define CPUACCT_BATCH \
9462 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9464 #define CPUACCT_BATCH 0
9468 * Charge the system/user time to the task's accounting group.
9470 static void cpuacct_update_stats(struct task_struct *tsk,
9471 enum cpuacct_stat_index idx, cputime_t val)
9474 int batch = CPUACCT_BATCH;
9476 if (unlikely(!cpuacct_subsys.active))
9483 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9489 struct cgroup_subsys cpuacct_subsys = {
9491 .create = cpuacct_create,
9492 .destroy = cpuacct_destroy,
9493 .populate = cpuacct_populate,
9494 .subsys_id = cpuacct_subsys_id,
9496 #endif /* CONFIG_CGROUP_CPUACCT */