4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/stop_machine.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
74 #include <linux/slab.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
80 #include "workqueue_sched.h"
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/sched.h>
86 * Convert user-nice values [ -20 ... 0 ... 19 ]
87 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
90 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
91 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
92 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
95 * 'User priority' is the nice value converted to something we
96 * can work with better when scaling various scheduler parameters,
97 * it's a [ 0 ... 39 ] range.
99 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
100 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
101 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
104 * Helpers for converting nanosecond timing to jiffy resolution
106 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
108 #define NICE_0_LOAD SCHED_LOAD_SCALE
109 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
112 * These are the 'tuning knobs' of the scheduler:
114 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
115 * Timeslices get refilled after they expire.
117 #define DEF_TIMESLICE (100 * HZ / 1000)
120 * single value that denotes runtime == period, ie unlimited time.
122 #define RUNTIME_INF ((u64)~0ULL)
124 static inline int rt_policy(int policy)
126 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
131 static inline int task_has_rt_policy(struct task_struct *p)
133 return rt_policy(p->policy);
137 * This is the priority-queue data structure of the RT scheduling class:
139 struct rt_prio_array {
140 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
141 struct list_head queue[MAX_RT_PRIO];
144 struct rt_bandwidth {
145 /* nests inside the rq lock: */
146 raw_spinlock_t rt_runtime_lock;
149 struct hrtimer rt_period_timer;
152 static struct rt_bandwidth def_rt_bandwidth;
154 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
156 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
158 struct rt_bandwidth *rt_b =
159 container_of(timer, struct rt_bandwidth, rt_period_timer);
165 now = hrtimer_cb_get_time(timer);
166 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
171 idle = do_sched_rt_period_timer(rt_b, overrun);
174 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
178 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
180 rt_b->rt_period = ns_to_ktime(period);
181 rt_b->rt_runtime = runtime;
183 raw_spin_lock_init(&rt_b->rt_runtime_lock);
185 hrtimer_init(&rt_b->rt_period_timer,
186 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
187 rt_b->rt_period_timer.function = sched_rt_period_timer;
190 static inline int rt_bandwidth_enabled(void)
192 return sysctl_sched_rt_runtime >= 0;
195 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
199 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
202 if (hrtimer_active(&rt_b->rt_period_timer))
205 raw_spin_lock(&rt_b->rt_runtime_lock);
210 if (hrtimer_active(&rt_b->rt_period_timer))
213 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
214 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
216 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
217 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
218 delta = ktime_to_ns(ktime_sub(hard, soft));
219 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
220 HRTIMER_MODE_ABS_PINNED, 0);
222 raw_spin_unlock(&rt_b->rt_runtime_lock);
225 #ifdef CONFIG_RT_GROUP_SCHED
226 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
228 hrtimer_cancel(&rt_b->rt_period_timer);
233 * sched_domains_mutex serializes calls to arch_init_sched_domains,
234 * detach_destroy_domains and partition_sched_domains.
236 static DEFINE_MUTEX(sched_domains_mutex);
238 #ifdef CONFIG_CGROUP_SCHED
240 #include <linux/cgroup.h>
244 static LIST_HEAD(task_groups);
246 /* task group related information */
248 struct cgroup_subsys_state css;
250 #ifdef CONFIG_FAIR_GROUP_SCHED
251 /* schedulable entities of this group on each cpu */
252 struct sched_entity **se;
253 /* runqueue "owned" by this group on each cpu */
254 struct cfs_rq **cfs_rq;
255 unsigned long shares;
258 #ifdef CONFIG_RT_GROUP_SCHED
259 struct sched_rt_entity **rt_se;
260 struct rt_rq **rt_rq;
262 struct rt_bandwidth rt_bandwidth;
266 struct list_head list;
268 struct task_group *parent;
269 struct list_head siblings;
270 struct list_head children;
273 #define root_task_group init_task_group
275 /* task_group_lock serializes add/remove of task groups and also changes to
276 * a task group's cpu shares.
278 static DEFINE_SPINLOCK(task_group_lock);
280 #ifdef CONFIG_FAIR_GROUP_SCHED
283 static int root_task_group_empty(void)
285 return list_empty(&root_task_group.children);
289 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
292 * A weight of 0 or 1 can cause arithmetics problems.
293 * A weight of a cfs_rq is the sum of weights of which entities
294 * are queued on this cfs_rq, so a weight of a entity should not be
295 * too large, so as the shares value of a task group.
296 * (The default weight is 1024 - so there's no practical
297 * limitation from this.)
300 #define MAX_SHARES (1UL << 18)
302 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
305 /* Default task group.
306 * Every task in system belong to this group at bootup.
308 struct task_group init_task_group;
310 #endif /* CONFIG_CGROUP_SCHED */
312 /* CFS-related fields in a runqueue */
314 struct load_weight load;
315 unsigned long nr_running;
320 struct rb_root tasks_timeline;
321 struct rb_node *rb_leftmost;
323 struct list_head tasks;
324 struct list_head *balance_iterator;
327 * 'curr' points to currently running entity on this cfs_rq.
328 * It is set to NULL otherwise (i.e when none are currently running).
330 struct sched_entity *curr, *next, *last;
332 unsigned int nr_spread_over;
334 #ifdef CONFIG_FAIR_GROUP_SCHED
335 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
338 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
339 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
340 * (like users, containers etc.)
342 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
343 * list is used during load balance.
345 struct list_head leaf_cfs_rq_list;
346 struct task_group *tg; /* group that "owns" this runqueue */
350 * the part of load.weight contributed by tasks
352 unsigned long task_weight;
355 * h_load = weight * f(tg)
357 * Where f(tg) is the recursive weight fraction assigned to
360 unsigned long h_load;
363 * this cpu's part of tg->shares
365 unsigned long shares;
368 * load.weight at the time we set shares
370 unsigned long rq_weight;
375 /* Real-Time classes' related field in a runqueue: */
377 struct rt_prio_array active;
378 unsigned long rt_nr_running;
379 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
381 int curr; /* highest queued rt task prio */
383 int next; /* next highest */
388 unsigned long rt_nr_migratory;
389 unsigned long rt_nr_total;
391 struct plist_head pushable_tasks;
396 /* Nests inside the rq lock: */
397 raw_spinlock_t rt_runtime_lock;
399 #ifdef CONFIG_RT_GROUP_SCHED
400 unsigned long rt_nr_boosted;
403 struct list_head leaf_rt_rq_list;
404 struct task_group *tg;
411 * We add the notion of a root-domain which will be used to define per-domain
412 * variables. Each exclusive cpuset essentially defines an island domain by
413 * fully partitioning the member cpus from any other cpuset. Whenever a new
414 * exclusive cpuset is created, we also create and attach a new root-domain
421 cpumask_var_t online;
424 * The "RT overload" flag: it gets set if a CPU has more than
425 * one runnable RT task.
427 cpumask_var_t rto_mask;
429 struct cpupri cpupri;
433 * By default the system creates a single root-domain with all cpus as
434 * members (mimicking the global state we have today).
436 static struct root_domain def_root_domain;
438 #endif /* CONFIG_SMP */
441 * This is the main, per-CPU runqueue data structure.
443 * Locking rule: those places that want to lock multiple runqueues
444 * (such as the load balancing or the thread migration code), lock
445 * acquire operations must be ordered by ascending &runqueue.
452 * nr_running and cpu_load should be in the same cacheline because
453 * remote CPUs use both these fields when doing load calculation.
455 unsigned long nr_running;
456 #define CPU_LOAD_IDX_MAX 5
457 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
458 unsigned long last_load_update_tick;
461 unsigned char nohz_balance_kick;
463 unsigned int skip_clock_update;
465 /* capture load from *all* tasks on this cpu: */
466 struct load_weight load;
467 unsigned long nr_load_updates;
473 #ifdef CONFIG_FAIR_GROUP_SCHED
474 /* list of leaf cfs_rq on this cpu: */
475 struct list_head leaf_cfs_rq_list;
477 #ifdef CONFIG_RT_GROUP_SCHED
478 struct list_head leaf_rt_rq_list;
482 * This is part of a global counter where only the total sum
483 * over all CPUs matters. A task can increase this counter on
484 * one CPU and if it got migrated afterwards it may decrease
485 * it on another CPU. Always updated under the runqueue lock:
487 unsigned long nr_uninterruptible;
489 struct task_struct *curr, *idle, *stop;
490 unsigned long next_balance;
491 struct mm_struct *prev_mm;
499 struct root_domain *rd;
500 struct sched_domain *sd;
502 unsigned long cpu_power;
504 unsigned char idle_at_tick;
505 /* For active balancing */
509 struct cpu_stop_work active_balance_work;
510 /* cpu of this runqueue: */
514 unsigned long avg_load_per_task;
522 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
526 /* calc_load related fields */
527 unsigned long calc_load_update;
528 long calc_load_active;
530 #ifdef CONFIG_SCHED_HRTICK
532 int hrtick_csd_pending;
533 struct call_single_data hrtick_csd;
535 struct hrtimer hrtick_timer;
538 #ifdef CONFIG_SCHEDSTATS
540 struct sched_info rq_sched_info;
541 unsigned long long rq_cpu_time;
542 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
544 /* sys_sched_yield() stats */
545 unsigned int yld_count;
547 /* schedule() stats */
548 unsigned int sched_switch;
549 unsigned int sched_count;
550 unsigned int sched_goidle;
552 /* try_to_wake_up() stats */
553 unsigned int ttwu_count;
554 unsigned int ttwu_local;
557 unsigned int bkl_count;
561 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
564 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
566 static inline int cpu_of(struct rq *rq)
575 #define rcu_dereference_check_sched_domain(p) \
576 rcu_dereference_check((p), \
577 rcu_read_lock_sched_held() || \
578 lockdep_is_held(&sched_domains_mutex))
581 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
582 * See detach_destroy_domains: synchronize_sched for details.
584 * The domain tree of any CPU may only be accessed from within
585 * preempt-disabled sections.
587 #define for_each_domain(cpu, __sd) \
588 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
590 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
591 #define this_rq() (&__get_cpu_var(runqueues))
592 #define task_rq(p) cpu_rq(task_cpu(p))
593 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
594 #define raw_rq() (&__raw_get_cpu_var(runqueues))
596 #ifdef CONFIG_CGROUP_SCHED
599 * Return the group to which this tasks belongs.
601 * We use task_subsys_state_check() and extend the RCU verification
602 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
603 * holds that lock for each task it moves into the cgroup. Therefore
604 * by holding that lock, we pin the task to the current cgroup.
606 static inline struct task_group *task_group(struct task_struct *p)
608 struct cgroup_subsys_state *css;
610 if (p->flags & PF_EXITING)
611 return &root_task_group;
613 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
614 lockdep_is_held(&task_rq(p)->lock));
615 return container_of(css, struct task_group, css);
618 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
619 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
621 #ifdef CONFIG_FAIR_GROUP_SCHED
622 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
623 p->se.parent = task_group(p)->se[cpu];
626 #ifdef CONFIG_RT_GROUP_SCHED
627 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
628 p->rt.parent = task_group(p)->rt_se[cpu];
632 #else /* CONFIG_CGROUP_SCHED */
634 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
635 static inline struct task_group *task_group(struct task_struct *p)
640 #endif /* CONFIG_CGROUP_SCHED */
642 static void update_rq_clock_task(struct rq *rq, s64 delta);
644 static void update_rq_clock(struct rq *rq)
648 if (rq->skip_clock_update)
651 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
653 update_rq_clock_task(rq, delta);
657 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
659 #ifdef CONFIG_SCHED_DEBUG
660 # define const_debug __read_mostly
662 # define const_debug static const
667 * @cpu: the processor in question.
669 * Returns true if the current cpu runqueue is locked.
670 * This interface allows printk to be called with the runqueue lock
671 * held and know whether or not it is OK to wake up the klogd.
673 int runqueue_is_locked(int cpu)
675 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
679 * Debugging: various feature bits
682 #define SCHED_FEAT(name, enabled) \
683 __SCHED_FEAT_##name ,
686 #include "sched_features.h"
691 #define SCHED_FEAT(name, enabled) \
692 (1UL << __SCHED_FEAT_##name) * enabled |
694 const_debug unsigned int sysctl_sched_features =
695 #include "sched_features.h"
700 #ifdef CONFIG_SCHED_DEBUG
701 #define SCHED_FEAT(name, enabled) \
704 static __read_mostly char *sched_feat_names[] = {
705 #include "sched_features.h"
711 static int sched_feat_show(struct seq_file *m, void *v)
715 for (i = 0; sched_feat_names[i]; i++) {
716 if (!(sysctl_sched_features & (1UL << i)))
718 seq_printf(m, "%s ", sched_feat_names[i]);
726 sched_feat_write(struct file *filp, const char __user *ubuf,
727 size_t cnt, loff_t *ppos)
737 if (copy_from_user(&buf, ubuf, cnt))
743 if (strncmp(buf, "NO_", 3) == 0) {
748 for (i = 0; sched_feat_names[i]; i++) {
749 if (strcmp(cmp, sched_feat_names[i]) == 0) {
751 sysctl_sched_features &= ~(1UL << i);
753 sysctl_sched_features |= (1UL << i);
758 if (!sched_feat_names[i])
766 static int sched_feat_open(struct inode *inode, struct file *filp)
768 return single_open(filp, sched_feat_show, NULL);
771 static const struct file_operations sched_feat_fops = {
772 .open = sched_feat_open,
773 .write = sched_feat_write,
776 .release = single_release,
779 static __init int sched_init_debug(void)
781 debugfs_create_file("sched_features", 0644, NULL, NULL,
786 late_initcall(sched_init_debug);
790 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
793 * Number of tasks to iterate in a single balance run.
794 * Limited because this is done with IRQs disabled.
796 const_debug unsigned int sysctl_sched_nr_migrate = 32;
799 * ratelimit for updating the group shares.
802 unsigned int sysctl_sched_shares_ratelimit = 250000;
803 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
806 * Inject some fuzzyness into changing the per-cpu group shares
807 * this avoids remote rq-locks at the expense of fairness.
810 unsigned int sysctl_sched_shares_thresh = 4;
813 * period over which we average the RT time consumption, measured
818 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
821 * period over which we measure -rt task cpu usage in us.
824 unsigned int sysctl_sched_rt_period = 1000000;
826 static __read_mostly int scheduler_running;
829 * part of the period that we allow rt tasks to run in us.
832 int sysctl_sched_rt_runtime = 950000;
834 static inline u64 global_rt_period(void)
836 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
839 static inline u64 global_rt_runtime(void)
841 if (sysctl_sched_rt_runtime < 0)
844 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
847 #ifndef prepare_arch_switch
848 # define prepare_arch_switch(next) do { } while (0)
850 #ifndef finish_arch_switch
851 # define finish_arch_switch(prev) do { } while (0)
854 static inline int task_current(struct rq *rq, struct task_struct *p)
856 return rq->curr == p;
859 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
860 static inline int task_running(struct rq *rq, struct task_struct *p)
862 return task_current(rq, p);
865 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
869 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
871 #ifdef CONFIG_DEBUG_SPINLOCK
872 /* this is a valid case when another task releases the spinlock */
873 rq->lock.owner = current;
876 * If we are tracking spinlock dependencies then we have to
877 * fix up the runqueue lock - which gets 'carried over' from
880 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
882 raw_spin_unlock_irq(&rq->lock);
885 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
886 static inline int task_running(struct rq *rq, struct task_struct *p)
891 return task_current(rq, p);
895 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
899 * We can optimise this out completely for !SMP, because the
900 * SMP rebalancing from interrupt is the only thing that cares
905 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
906 raw_spin_unlock_irq(&rq->lock);
908 raw_spin_unlock(&rq->lock);
912 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
916 * After ->oncpu is cleared, the task can be moved to a different CPU.
917 * We must ensure this doesn't happen until the switch is completely
923 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
927 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
930 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
933 static inline int task_is_waking(struct task_struct *p)
935 return unlikely(p->state == TASK_WAKING);
939 * __task_rq_lock - lock the runqueue a given task resides on.
940 * Must be called interrupts disabled.
942 static inline struct rq *__task_rq_lock(struct task_struct *p)
949 raw_spin_lock(&rq->lock);
950 if (likely(rq == task_rq(p)))
952 raw_spin_unlock(&rq->lock);
957 * task_rq_lock - lock the runqueue a given task resides on and disable
958 * interrupts. Note the ordering: we can safely lookup the task_rq without
959 * explicitly disabling preemption.
961 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
967 local_irq_save(*flags);
969 raw_spin_lock(&rq->lock);
970 if (likely(rq == task_rq(p)))
972 raw_spin_unlock_irqrestore(&rq->lock, *flags);
976 static void __task_rq_unlock(struct rq *rq)
979 raw_spin_unlock(&rq->lock);
982 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
985 raw_spin_unlock_irqrestore(&rq->lock, *flags);
989 * this_rq_lock - lock this runqueue and disable interrupts.
991 static struct rq *this_rq_lock(void)
998 raw_spin_lock(&rq->lock);
1003 #ifdef CONFIG_SCHED_HRTICK
1005 * Use HR-timers to deliver accurate preemption points.
1007 * Its all a bit involved since we cannot program an hrt while holding the
1008 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1011 * When we get rescheduled we reprogram the hrtick_timer outside of the
1017 * - enabled by features
1018 * - hrtimer is actually high res
1020 static inline int hrtick_enabled(struct rq *rq)
1022 if (!sched_feat(HRTICK))
1024 if (!cpu_active(cpu_of(rq)))
1026 return hrtimer_is_hres_active(&rq->hrtick_timer);
1029 static void hrtick_clear(struct rq *rq)
1031 if (hrtimer_active(&rq->hrtick_timer))
1032 hrtimer_cancel(&rq->hrtick_timer);
1036 * High-resolution timer tick.
1037 * Runs from hardirq context with interrupts disabled.
1039 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1041 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1043 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1045 raw_spin_lock(&rq->lock);
1046 update_rq_clock(rq);
1047 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1048 raw_spin_unlock(&rq->lock);
1050 return HRTIMER_NORESTART;
1055 * called from hardirq (IPI) context
1057 static void __hrtick_start(void *arg)
1059 struct rq *rq = arg;
1061 raw_spin_lock(&rq->lock);
1062 hrtimer_restart(&rq->hrtick_timer);
1063 rq->hrtick_csd_pending = 0;
1064 raw_spin_unlock(&rq->lock);
1068 * Called to set the hrtick timer state.
1070 * called with rq->lock held and irqs disabled
1072 static void hrtick_start(struct rq *rq, u64 delay)
1074 struct hrtimer *timer = &rq->hrtick_timer;
1075 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1077 hrtimer_set_expires(timer, time);
1079 if (rq == this_rq()) {
1080 hrtimer_restart(timer);
1081 } else if (!rq->hrtick_csd_pending) {
1082 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1083 rq->hrtick_csd_pending = 1;
1088 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1090 int cpu = (int)(long)hcpu;
1093 case CPU_UP_CANCELED:
1094 case CPU_UP_CANCELED_FROZEN:
1095 case CPU_DOWN_PREPARE:
1096 case CPU_DOWN_PREPARE_FROZEN:
1098 case CPU_DEAD_FROZEN:
1099 hrtick_clear(cpu_rq(cpu));
1106 static __init void init_hrtick(void)
1108 hotcpu_notifier(hotplug_hrtick, 0);
1112 * Called to set the hrtick timer state.
1114 * called with rq->lock held and irqs disabled
1116 static void hrtick_start(struct rq *rq, u64 delay)
1118 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1119 HRTIMER_MODE_REL_PINNED, 0);
1122 static inline void init_hrtick(void)
1125 #endif /* CONFIG_SMP */
1127 static void init_rq_hrtick(struct rq *rq)
1130 rq->hrtick_csd_pending = 0;
1132 rq->hrtick_csd.flags = 0;
1133 rq->hrtick_csd.func = __hrtick_start;
1134 rq->hrtick_csd.info = rq;
1137 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1138 rq->hrtick_timer.function = hrtick;
1140 #else /* CONFIG_SCHED_HRTICK */
1141 static inline void hrtick_clear(struct rq *rq)
1145 static inline void init_rq_hrtick(struct rq *rq)
1149 static inline void init_hrtick(void)
1152 #endif /* CONFIG_SCHED_HRTICK */
1155 * resched_task - mark a task 'to be rescheduled now'.
1157 * On UP this means the setting of the need_resched flag, on SMP it
1158 * might also involve a cross-CPU call to trigger the scheduler on
1163 #ifndef tsk_is_polling
1164 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1167 static void resched_task(struct task_struct *p)
1171 assert_raw_spin_locked(&task_rq(p)->lock);
1173 if (test_tsk_need_resched(p))
1176 set_tsk_need_resched(p);
1179 if (cpu == smp_processor_id())
1182 /* NEED_RESCHED must be visible before we test polling */
1184 if (!tsk_is_polling(p))
1185 smp_send_reschedule(cpu);
1188 static void resched_cpu(int cpu)
1190 struct rq *rq = cpu_rq(cpu);
1191 unsigned long flags;
1193 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1195 resched_task(cpu_curr(cpu));
1196 raw_spin_unlock_irqrestore(&rq->lock, flags);
1201 * In the semi idle case, use the nearest busy cpu for migrating timers
1202 * from an idle cpu. This is good for power-savings.
1204 * We don't do similar optimization for completely idle system, as
1205 * selecting an idle cpu will add more delays to the timers than intended
1206 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1208 int get_nohz_timer_target(void)
1210 int cpu = smp_processor_id();
1212 struct sched_domain *sd;
1214 for_each_domain(cpu, sd) {
1215 for_each_cpu(i, sched_domain_span(sd))
1222 * When add_timer_on() enqueues a timer into the timer wheel of an
1223 * idle CPU then this timer might expire before the next timer event
1224 * which is scheduled to wake up that CPU. In case of a completely
1225 * idle system the next event might even be infinite time into the
1226 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1227 * leaves the inner idle loop so the newly added timer is taken into
1228 * account when the CPU goes back to idle and evaluates the timer
1229 * wheel for the next timer event.
1231 void wake_up_idle_cpu(int cpu)
1233 struct rq *rq = cpu_rq(cpu);
1235 if (cpu == smp_processor_id())
1239 * This is safe, as this function is called with the timer
1240 * wheel base lock of (cpu) held. When the CPU is on the way
1241 * to idle and has not yet set rq->curr to idle then it will
1242 * be serialized on the timer wheel base lock and take the new
1243 * timer into account automatically.
1245 if (rq->curr != rq->idle)
1249 * We can set TIF_RESCHED on the idle task of the other CPU
1250 * lockless. The worst case is that the other CPU runs the
1251 * idle task through an additional NOOP schedule()
1253 set_tsk_need_resched(rq->idle);
1255 /* NEED_RESCHED must be visible before we test polling */
1257 if (!tsk_is_polling(rq->idle))
1258 smp_send_reschedule(cpu);
1261 #endif /* CONFIG_NO_HZ */
1263 static u64 sched_avg_period(void)
1265 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1268 static void sched_avg_update(struct rq *rq)
1270 s64 period = sched_avg_period();
1272 while ((s64)(rq->clock - rq->age_stamp) > period) {
1274 * Inline assembly required to prevent the compiler
1275 * optimising this loop into a divmod call.
1276 * See __iter_div_u64_rem() for another example of this.
1278 asm("" : "+rm" (rq->age_stamp));
1279 rq->age_stamp += period;
1284 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1286 rq->rt_avg += rt_delta;
1287 sched_avg_update(rq);
1290 #else /* !CONFIG_SMP */
1291 static void resched_task(struct task_struct *p)
1293 assert_raw_spin_locked(&task_rq(p)->lock);
1294 set_tsk_need_resched(p);
1297 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1301 static void sched_avg_update(struct rq *rq)
1304 #endif /* CONFIG_SMP */
1306 #if BITS_PER_LONG == 32
1307 # define WMULT_CONST (~0UL)
1309 # define WMULT_CONST (1UL << 32)
1312 #define WMULT_SHIFT 32
1315 * Shift right and round:
1317 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1320 * delta *= weight / lw
1322 static unsigned long
1323 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1324 struct load_weight *lw)
1328 if (!lw->inv_weight) {
1329 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1332 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1336 tmp = (u64)delta_exec * weight;
1338 * Check whether we'd overflow the 64-bit multiplication:
1340 if (unlikely(tmp > WMULT_CONST))
1341 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1344 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1346 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1349 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1355 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1362 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1363 * of tasks with abnormal "nice" values across CPUs the contribution that
1364 * each task makes to its run queue's load is weighted according to its
1365 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1366 * scaled version of the new time slice allocation that they receive on time
1370 #define WEIGHT_IDLEPRIO 3
1371 #define WMULT_IDLEPRIO 1431655765
1374 * Nice levels are multiplicative, with a gentle 10% change for every
1375 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1376 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1377 * that remained on nice 0.
1379 * The "10% effect" is relative and cumulative: from _any_ nice level,
1380 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1381 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1382 * If a task goes up by ~10% and another task goes down by ~10% then
1383 * the relative distance between them is ~25%.)
1385 static const int prio_to_weight[40] = {
1386 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1387 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1388 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1389 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1390 /* 0 */ 1024, 820, 655, 526, 423,
1391 /* 5 */ 335, 272, 215, 172, 137,
1392 /* 10 */ 110, 87, 70, 56, 45,
1393 /* 15 */ 36, 29, 23, 18, 15,
1397 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1399 * In cases where the weight does not change often, we can use the
1400 * precalculated inverse to speed up arithmetics by turning divisions
1401 * into multiplications:
1403 static const u32 prio_to_wmult[40] = {
1404 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1405 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1406 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1407 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1408 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1409 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1410 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1411 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1414 /* Time spent by the tasks of the cpu accounting group executing in ... */
1415 enum cpuacct_stat_index {
1416 CPUACCT_STAT_USER, /* ... user mode */
1417 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1419 CPUACCT_STAT_NSTATS,
1422 #ifdef CONFIG_CGROUP_CPUACCT
1423 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1424 static void cpuacct_update_stats(struct task_struct *tsk,
1425 enum cpuacct_stat_index idx, cputime_t val);
1427 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1428 static inline void cpuacct_update_stats(struct task_struct *tsk,
1429 enum cpuacct_stat_index idx, cputime_t val) {}
1432 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1434 update_load_add(&rq->load, load);
1437 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1439 update_load_sub(&rq->load, load);
1442 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1443 typedef int (*tg_visitor)(struct task_group *, void *);
1446 * Iterate the full tree, calling @down when first entering a node and @up when
1447 * leaving it for the final time.
1449 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1451 struct task_group *parent, *child;
1455 parent = &root_task_group;
1457 ret = (*down)(parent, data);
1460 list_for_each_entry_rcu(child, &parent->children, siblings) {
1467 ret = (*up)(parent, data);
1472 parent = parent->parent;
1481 static int tg_nop(struct task_group *tg, void *data)
1488 /* Used instead of source_load when we know the type == 0 */
1489 static unsigned long weighted_cpuload(const int cpu)
1491 return cpu_rq(cpu)->load.weight;
1495 * Return a low guess at the load of a migration-source cpu weighted
1496 * according to the scheduling class and "nice" value.
1498 * We want to under-estimate the load of migration sources, to
1499 * balance conservatively.
1501 static unsigned long source_load(int cpu, int type)
1503 struct rq *rq = cpu_rq(cpu);
1504 unsigned long total = weighted_cpuload(cpu);
1506 if (type == 0 || !sched_feat(LB_BIAS))
1509 return min(rq->cpu_load[type-1], total);
1513 * Return a high guess at the load of a migration-target cpu weighted
1514 * according to the scheduling class and "nice" value.
1516 static unsigned long target_load(int cpu, int type)
1518 struct rq *rq = cpu_rq(cpu);
1519 unsigned long total = weighted_cpuload(cpu);
1521 if (type == 0 || !sched_feat(LB_BIAS))
1524 return max(rq->cpu_load[type-1], total);
1527 static unsigned long power_of(int cpu)
1529 return cpu_rq(cpu)->cpu_power;
1532 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1534 static unsigned long cpu_avg_load_per_task(int cpu)
1536 struct rq *rq = cpu_rq(cpu);
1537 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1540 rq->avg_load_per_task = rq->load.weight / nr_running;
1542 rq->avg_load_per_task = 0;
1544 return rq->avg_load_per_task;
1547 #ifdef CONFIG_FAIR_GROUP_SCHED
1549 static __read_mostly unsigned long __percpu *update_shares_data;
1551 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1554 * Calculate and set the cpu's group shares.
1556 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1557 unsigned long sd_shares,
1558 unsigned long sd_rq_weight,
1559 unsigned long *usd_rq_weight)
1561 unsigned long shares, rq_weight;
1564 rq_weight = usd_rq_weight[cpu];
1567 rq_weight = NICE_0_LOAD;
1571 * \Sum_j shares_j * rq_weight_i
1572 * shares_i = -----------------------------
1573 * \Sum_j rq_weight_j
1575 shares = (sd_shares * rq_weight) / sd_rq_weight;
1576 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1578 if (abs(shares - tg->se[cpu]->load.weight) >
1579 sysctl_sched_shares_thresh) {
1580 struct rq *rq = cpu_rq(cpu);
1581 unsigned long flags;
1583 raw_spin_lock_irqsave(&rq->lock, flags);
1584 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1585 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1586 __set_se_shares(tg->se[cpu], shares);
1587 raw_spin_unlock_irqrestore(&rq->lock, flags);
1592 * Re-compute the task group their per cpu shares over the given domain.
1593 * This needs to be done in a bottom-up fashion because the rq weight of a
1594 * parent group depends on the shares of its child groups.
1596 static int tg_shares_up(struct task_group *tg, void *data)
1598 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1599 unsigned long *usd_rq_weight;
1600 struct sched_domain *sd = data;
1601 unsigned long flags;
1607 local_irq_save(flags);
1608 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1610 for_each_cpu(i, sched_domain_span(sd)) {
1611 weight = tg->cfs_rq[i]->load.weight;
1612 usd_rq_weight[i] = weight;
1614 rq_weight += weight;
1616 * If there are currently no tasks on the cpu pretend there
1617 * is one of average load so that when a new task gets to
1618 * run here it will not get delayed by group starvation.
1621 weight = NICE_0_LOAD;
1623 sum_weight += weight;
1624 shares += tg->cfs_rq[i]->shares;
1628 rq_weight = sum_weight;
1630 if ((!shares && rq_weight) || shares > tg->shares)
1631 shares = tg->shares;
1633 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1634 shares = tg->shares;
1636 for_each_cpu(i, sched_domain_span(sd))
1637 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1639 local_irq_restore(flags);
1645 * Compute the cpu's hierarchical load factor for each task group.
1646 * This needs to be done in a top-down fashion because the load of a child
1647 * group is a fraction of its parents load.
1649 static int tg_load_down(struct task_group *tg, void *data)
1652 long cpu = (long)data;
1655 load = cpu_rq(cpu)->load.weight;
1657 load = tg->parent->cfs_rq[cpu]->h_load;
1658 load *= tg->cfs_rq[cpu]->shares;
1659 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1662 tg->cfs_rq[cpu]->h_load = load;
1667 static void update_shares(struct sched_domain *sd)
1672 if (root_task_group_empty())
1675 now = local_clock();
1676 elapsed = now - sd->last_update;
1678 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1679 sd->last_update = now;
1680 walk_tg_tree(tg_nop, tg_shares_up, sd);
1684 static void update_h_load(long cpu)
1686 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1691 static inline void update_shares(struct sched_domain *sd)
1697 #ifdef CONFIG_PREEMPT
1699 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1702 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1703 * way at the expense of forcing extra atomic operations in all
1704 * invocations. This assures that the double_lock is acquired using the
1705 * same underlying policy as the spinlock_t on this architecture, which
1706 * reduces latency compared to the unfair variant below. However, it
1707 * also adds more overhead and therefore may reduce throughput.
1709 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1710 __releases(this_rq->lock)
1711 __acquires(busiest->lock)
1712 __acquires(this_rq->lock)
1714 raw_spin_unlock(&this_rq->lock);
1715 double_rq_lock(this_rq, busiest);
1722 * Unfair double_lock_balance: Optimizes throughput at the expense of
1723 * latency by eliminating extra atomic operations when the locks are
1724 * already in proper order on entry. This favors lower cpu-ids and will
1725 * grant the double lock to lower cpus over higher ids under contention,
1726 * regardless of entry order into the function.
1728 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1729 __releases(this_rq->lock)
1730 __acquires(busiest->lock)
1731 __acquires(this_rq->lock)
1735 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1736 if (busiest < this_rq) {
1737 raw_spin_unlock(&this_rq->lock);
1738 raw_spin_lock(&busiest->lock);
1739 raw_spin_lock_nested(&this_rq->lock,
1740 SINGLE_DEPTH_NESTING);
1743 raw_spin_lock_nested(&busiest->lock,
1744 SINGLE_DEPTH_NESTING);
1749 #endif /* CONFIG_PREEMPT */
1752 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1754 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1756 if (unlikely(!irqs_disabled())) {
1757 /* printk() doesn't work good under rq->lock */
1758 raw_spin_unlock(&this_rq->lock);
1762 return _double_lock_balance(this_rq, busiest);
1765 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1766 __releases(busiest->lock)
1768 raw_spin_unlock(&busiest->lock);
1769 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1773 * double_rq_lock - safely lock two runqueues
1775 * Note this does not disable interrupts like task_rq_lock,
1776 * you need to do so manually before calling.
1778 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1779 __acquires(rq1->lock)
1780 __acquires(rq2->lock)
1782 BUG_ON(!irqs_disabled());
1784 raw_spin_lock(&rq1->lock);
1785 __acquire(rq2->lock); /* Fake it out ;) */
1788 raw_spin_lock(&rq1->lock);
1789 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1791 raw_spin_lock(&rq2->lock);
1792 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1798 * double_rq_unlock - safely unlock two runqueues
1800 * Note this does not restore interrupts like task_rq_unlock,
1801 * you need to do so manually after calling.
1803 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1804 __releases(rq1->lock)
1805 __releases(rq2->lock)
1807 raw_spin_unlock(&rq1->lock);
1809 raw_spin_unlock(&rq2->lock);
1811 __release(rq2->lock);
1816 #ifdef CONFIG_FAIR_GROUP_SCHED
1817 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1820 cfs_rq->shares = shares;
1825 static void calc_load_account_idle(struct rq *this_rq);
1826 static void update_sysctl(void);
1827 static int get_update_sysctl_factor(void);
1828 static void update_cpu_load(struct rq *this_rq);
1830 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1832 set_task_rq(p, cpu);
1835 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1836 * successfuly executed on another CPU. We must ensure that updates of
1837 * per-task data have been completed by this moment.
1840 task_thread_info(p)->cpu = cpu;
1844 static const struct sched_class rt_sched_class;
1846 #define sched_class_highest (&stop_sched_class)
1847 #define for_each_class(class) \
1848 for (class = sched_class_highest; class; class = class->next)
1850 #include "sched_stats.h"
1852 static void inc_nr_running(struct rq *rq)
1857 static void dec_nr_running(struct rq *rq)
1862 static void set_load_weight(struct task_struct *p)
1865 * SCHED_IDLE tasks get minimal weight:
1867 if (p->policy == SCHED_IDLE) {
1868 p->se.load.weight = WEIGHT_IDLEPRIO;
1869 p->se.load.inv_weight = WMULT_IDLEPRIO;
1873 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1874 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1877 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1879 update_rq_clock(rq);
1880 sched_info_queued(p);
1881 p->sched_class->enqueue_task(rq, p, flags);
1885 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1887 update_rq_clock(rq);
1888 sched_info_dequeued(p);
1889 p->sched_class->dequeue_task(rq, p, flags);
1894 * activate_task - move a task to the runqueue.
1896 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1898 if (task_contributes_to_load(p))
1899 rq->nr_uninterruptible--;
1901 enqueue_task(rq, p, flags);
1906 * deactivate_task - remove a task from the runqueue.
1908 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1910 if (task_contributes_to_load(p))
1911 rq->nr_uninterruptible++;
1913 dequeue_task(rq, p, flags);
1917 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1920 * There are no locks covering percpu hardirq/softirq time.
1921 * They are only modified in account_system_vtime, on corresponding CPU
1922 * with interrupts disabled. So, writes are safe.
1923 * They are read and saved off onto struct rq in update_rq_clock().
1924 * This may result in other CPU reading this CPU's irq time and can
1925 * race with irq/account_system_vtime on this CPU. We would either get old
1926 * or new value with a side effect of accounting a slice of irq time to wrong
1927 * task when irq is in progress while we read rq->clock. That is a worthy
1928 * compromise in place of having locks on each irq in account_system_time.
1930 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1931 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1933 static DEFINE_PER_CPU(u64, irq_start_time);
1934 static int sched_clock_irqtime;
1936 void enable_sched_clock_irqtime(void)
1938 sched_clock_irqtime = 1;
1941 void disable_sched_clock_irqtime(void)
1943 sched_clock_irqtime = 0;
1946 #ifndef CONFIG_64BIT
1947 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1949 static inline void irq_time_write_begin(void)
1951 __this_cpu_inc(irq_time_seq.sequence);
1955 static inline void irq_time_write_end(void)
1958 __this_cpu_inc(irq_time_seq.sequence);
1961 static inline u64 irq_time_read(int cpu)
1967 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1968 irq_time = per_cpu(cpu_softirq_time, cpu) +
1969 per_cpu(cpu_hardirq_time, cpu);
1970 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1974 #else /* CONFIG_64BIT */
1975 static inline void irq_time_write_begin(void)
1979 static inline void irq_time_write_end(void)
1983 static inline u64 irq_time_read(int cpu)
1985 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1987 #endif /* CONFIG_64BIT */
1990 * Called before incrementing preempt_count on {soft,}irq_enter
1991 * and before decrementing preempt_count on {soft,}irq_exit.
1993 void account_system_vtime(struct task_struct *curr)
1995 unsigned long flags;
1999 if (!sched_clock_irqtime)
2002 local_irq_save(flags);
2004 cpu = smp_processor_id();
2005 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
2006 __this_cpu_add(irq_start_time, delta);
2008 irq_time_write_begin();
2010 * We do not account for softirq time from ksoftirqd here.
2011 * We want to continue accounting softirq time to ksoftirqd thread
2012 * in that case, so as not to confuse scheduler with a special task
2013 * that do not consume any time, but still wants to run.
2015 if (hardirq_count())
2016 __this_cpu_add(cpu_hardirq_time, delta);
2017 else if (in_serving_softirq() && !(curr->flags & PF_KSOFTIRQD))
2018 __this_cpu_add(cpu_softirq_time, delta);
2020 irq_time_write_end();
2021 local_irq_restore(flags);
2023 EXPORT_SYMBOL_GPL(account_system_vtime);
2025 static void update_rq_clock_task(struct rq *rq, s64 delta)
2029 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
2032 * Since irq_time is only updated on {soft,}irq_exit, we might run into
2033 * this case when a previous update_rq_clock() happened inside a
2034 * {soft,}irq region.
2036 * When this happens, we stop ->clock_task and only update the
2037 * prev_irq_time stamp to account for the part that fit, so that a next
2038 * update will consume the rest. This ensures ->clock_task is
2041 * It does however cause some slight miss-attribution of {soft,}irq
2042 * time, a more accurate solution would be to update the irq_time using
2043 * the current rq->clock timestamp, except that would require using
2046 if (irq_delta > delta)
2049 rq->prev_irq_time += irq_delta;
2051 rq->clock_task += delta;
2053 if (irq_delta && sched_feat(NONIRQ_POWER))
2054 sched_rt_avg_update(rq, irq_delta);
2057 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2059 static void update_rq_clock_task(struct rq *rq, s64 delta)
2061 rq->clock_task += delta;
2064 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2066 #include "sched_idletask.c"
2067 #include "sched_fair.c"
2068 #include "sched_rt.c"
2069 #include "sched_stoptask.c"
2070 #ifdef CONFIG_SCHED_DEBUG
2071 # include "sched_debug.c"
2074 void sched_set_stop_task(int cpu, struct task_struct *stop)
2076 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2077 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2081 * Make it appear like a SCHED_FIFO task, its something
2082 * userspace knows about and won't get confused about.
2084 * Also, it will make PI more or less work without too
2085 * much confusion -- but then, stop work should not
2086 * rely on PI working anyway.
2088 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2090 stop->sched_class = &stop_sched_class;
2093 cpu_rq(cpu)->stop = stop;
2097 * Reset it back to a normal scheduling class so that
2098 * it can die in pieces.
2100 old_stop->sched_class = &rt_sched_class;
2105 * __normal_prio - return the priority that is based on the static prio
2107 static inline int __normal_prio(struct task_struct *p)
2109 return p->static_prio;
2113 * Calculate the expected normal priority: i.e. priority
2114 * without taking RT-inheritance into account. Might be
2115 * boosted by interactivity modifiers. Changes upon fork,
2116 * setprio syscalls, and whenever the interactivity
2117 * estimator recalculates.
2119 static inline int normal_prio(struct task_struct *p)
2123 if (task_has_rt_policy(p))
2124 prio = MAX_RT_PRIO-1 - p->rt_priority;
2126 prio = __normal_prio(p);
2131 * Calculate the current priority, i.e. the priority
2132 * taken into account by the scheduler. This value might
2133 * be boosted by RT tasks, or might be boosted by
2134 * interactivity modifiers. Will be RT if the task got
2135 * RT-boosted. If not then it returns p->normal_prio.
2137 static int effective_prio(struct task_struct *p)
2139 p->normal_prio = normal_prio(p);
2141 * If we are RT tasks or we were boosted to RT priority,
2142 * keep the priority unchanged. Otherwise, update priority
2143 * to the normal priority:
2145 if (!rt_prio(p->prio))
2146 return p->normal_prio;
2151 * task_curr - is this task currently executing on a CPU?
2152 * @p: the task in question.
2154 inline int task_curr(const struct task_struct *p)
2156 return cpu_curr(task_cpu(p)) == p;
2159 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2160 const struct sched_class *prev_class,
2161 int oldprio, int running)
2163 if (prev_class != p->sched_class) {
2164 if (prev_class->switched_from)
2165 prev_class->switched_from(rq, p, running);
2166 p->sched_class->switched_to(rq, p, running);
2168 p->sched_class->prio_changed(rq, p, oldprio, running);
2171 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2173 const struct sched_class *class;
2175 if (p->sched_class == rq->curr->sched_class) {
2176 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2178 for_each_class(class) {
2179 if (class == rq->curr->sched_class)
2181 if (class == p->sched_class) {
2182 resched_task(rq->curr);
2189 * A queue event has occurred, and we're going to schedule. In
2190 * this case, we can save a useless back to back clock update.
2192 if (rq->curr->se.on_rq && test_tsk_need_resched(rq->curr))
2193 rq->skip_clock_update = 1;
2198 * Is this task likely cache-hot:
2201 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2205 if (p->sched_class != &fair_sched_class)
2208 if (unlikely(p->policy == SCHED_IDLE))
2212 * Buddy candidates are cache hot:
2214 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2215 (&p->se == cfs_rq_of(&p->se)->next ||
2216 &p->se == cfs_rq_of(&p->se)->last))
2219 if (sysctl_sched_migration_cost == -1)
2221 if (sysctl_sched_migration_cost == 0)
2224 delta = now - p->se.exec_start;
2226 return delta < (s64)sysctl_sched_migration_cost;
2229 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2231 #ifdef CONFIG_SCHED_DEBUG
2233 * We should never call set_task_cpu() on a blocked task,
2234 * ttwu() will sort out the placement.
2236 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2237 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2240 trace_sched_migrate_task(p, new_cpu);
2242 if (task_cpu(p) != new_cpu) {
2243 p->se.nr_migrations++;
2244 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2247 __set_task_cpu(p, new_cpu);
2250 struct migration_arg {
2251 struct task_struct *task;
2255 static int migration_cpu_stop(void *data);
2258 * The task's runqueue lock must be held.
2259 * Returns true if you have to wait for migration thread.
2261 static bool migrate_task(struct task_struct *p, int dest_cpu)
2263 struct rq *rq = task_rq(p);
2266 * If the task is not on a runqueue (and not running), then
2267 * the next wake-up will properly place the task.
2269 return p->se.on_rq || task_running(rq, p);
2273 * wait_task_inactive - wait for a thread to unschedule.
2275 * If @match_state is nonzero, it's the @p->state value just checked and
2276 * not expected to change. If it changes, i.e. @p might have woken up,
2277 * then return zero. When we succeed in waiting for @p to be off its CPU,
2278 * we return a positive number (its total switch count). If a second call
2279 * a short while later returns the same number, the caller can be sure that
2280 * @p has remained unscheduled the whole time.
2282 * The caller must ensure that the task *will* unschedule sometime soon,
2283 * else this function might spin for a *long* time. This function can't
2284 * be called with interrupts off, or it may introduce deadlock with
2285 * smp_call_function() if an IPI is sent by the same process we are
2286 * waiting to become inactive.
2288 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2290 unsigned long flags;
2297 * We do the initial early heuristics without holding
2298 * any task-queue locks at all. We'll only try to get
2299 * the runqueue lock when things look like they will
2305 * If the task is actively running on another CPU
2306 * still, just relax and busy-wait without holding
2309 * NOTE! Since we don't hold any locks, it's not
2310 * even sure that "rq" stays as the right runqueue!
2311 * But we don't care, since "task_running()" will
2312 * return false if the runqueue has changed and p
2313 * is actually now running somewhere else!
2315 while (task_running(rq, p)) {
2316 if (match_state && unlikely(p->state != match_state))
2322 * Ok, time to look more closely! We need the rq
2323 * lock now, to be *sure*. If we're wrong, we'll
2324 * just go back and repeat.
2326 rq = task_rq_lock(p, &flags);
2327 trace_sched_wait_task(p);
2328 running = task_running(rq, p);
2329 on_rq = p->se.on_rq;
2331 if (!match_state || p->state == match_state)
2332 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2333 task_rq_unlock(rq, &flags);
2336 * If it changed from the expected state, bail out now.
2338 if (unlikely(!ncsw))
2342 * Was it really running after all now that we
2343 * checked with the proper locks actually held?
2345 * Oops. Go back and try again..
2347 if (unlikely(running)) {
2353 * It's not enough that it's not actively running,
2354 * it must be off the runqueue _entirely_, and not
2357 * So if it was still runnable (but just not actively
2358 * running right now), it's preempted, and we should
2359 * yield - it could be a while.
2361 if (unlikely(on_rq)) {
2362 schedule_timeout_uninterruptible(1);
2367 * Ahh, all good. It wasn't running, and it wasn't
2368 * runnable, which means that it will never become
2369 * running in the future either. We're all done!
2378 * kick_process - kick a running thread to enter/exit the kernel
2379 * @p: the to-be-kicked thread
2381 * Cause a process which is running on another CPU to enter
2382 * kernel-mode, without any delay. (to get signals handled.)
2384 * NOTE: this function doesnt have to take the runqueue lock,
2385 * because all it wants to ensure is that the remote task enters
2386 * the kernel. If the IPI races and the task has been migrated
2387 * to another CPU then no harm is done and the purpose has been
2390 void kick_process(struct task_struct *p)
2396 if ((cpu != smp_processor_id()) && task_curr(p))
2397 smp_send_reschedule(cpu);
2400 EXPORT_SYMBOL_GPL(kick_process);
2401 #endif /* CONFIG_SMP */
2404 * task_oncpu_function_call - call a function on the cpu on which a task runs
2405 * @p: the task to evaluate
2406 * @func: the function to be called
2407 * @info: the function call argument
2409 * Calls the function @func when the task is currently running. This might
2410 * be on the current CPU, which just calls the function directly
2412 void task_oncpu_function_call(struct task_struct *p,
2413 void (*func) (void *info), void *info)
2420 smp_call_function_single(cpu, func, info, 1);
2426 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2428 static int select_fallback_rq(int cpu, struct task_struct *p)
2431 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2433 /* Look for allowed, online CPU in same node. */
2434 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2435 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2438 /* Any allowed, online CPU? */
2439 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2440 if (dest_cpu < nr_cpu_ids)
2443 /* No more Mr. Nice Guy. */
2444 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2445 dest_cpu = cpuset_cpus_allowed_fallback(p);
2447 * Don't tell them about moving exiting tasks or
2448 * kernel threads (both mm NULL), since they never
2451 if (p->mm && printk_ratelimit()) {
2452 printk(KERN_INFO "process %d (%s) no "
2453 "longer affine to cpu%d\n",
2454 task_pid_nr(p), p->comm, cpu);
2462 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2465 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2467 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2470 * In order not to call set_task_cpu() on a blocking task we need
2471 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2474 * Since this is common to all placement strategies, this lives here.
2476 * [ this allows ->select_task() to simply return task_cpu(p) and
2477 * not worry about this generic constraint ]
2479 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2481 cpu = select_fallback_rq(task_cpu(p), p);
2486 static void update_avg(u64 *avg, u64 sample)
2488 s64 diff = sample - *avg;
2493 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2494 bool is_sync, bool is_migrate, bool is_local,
2495 unsigned long en_flags)
2497 schedstat_inc(p, se.statistics.nr_wakeups);
2499 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2501 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2503 schedstat_inc(p, se.statistics.nr_wakeups_local);
2505 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2507 activate_task(rq, p, en_flags);
2510 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2511 int wake_flags, bool success)
2513 trace_sched_wakeup(p, success);
2514 check_preempt_curr(rq, p, wake_flags);
2516 p->state = TASK_RUNNING;
2518 if (p->sched_class->task_woken)
2519 p->sched_class->task_woken(rq, p);
2521 if (unlikely(rq->idle_stamp)) {
2522 u64 delta = rq->clock - rq->idle_stamp;
2523 u64 max = 2*sysctl_sched_migration_cost;
2528 update_avg(&rq->avg_idle, delta);
2532 /* if a worker is waking up, notify workqueue */
2533 if ((p->flags & PF_WQ_WORKER) && success)
2534 wq_worker_waking_up(p, cpu_of(rq));
2538 * try_to_wake_up - wake up a thread
2539 * @p: the thread to be awakened
2540 * @state: the mask of task states that can be woken
2541 * @wake_flags: wake modifier flags (WF_*)
2543 * Put it on the run-queue if it's not already there. The "current"
2544 * thread is always on the run-queue (except when the actual
2545 * re-schedule is in progress), and as such you're allowed to do
2546 * the simpler "current->state = TASK_RUNNING" to mark yourself
2547 * runnable without the overhead of this.
2549 * Returns %true if @p was woken up, %false if it was already running
2550 * or @state didn't match @p's state.
2552 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2555 int cpu, orig_cpu, this_cpu, success = 0;
2556 unsigned long flags;
2557 unsigned long en_flags = ENQUEUE_WAKEUP;
2560 this_cpu = get_cpu();
2563 rq = task_rq_lock(p, &flags);
2564 if (!(p->state & state))
2574 if (unlikely(task_running(rq, p)))
2578 * In order to handle concurrent wakeups and release the rq->lock
2579 * we put the task in TASK_WAKING state.
2581 * First fix up the nr_uninterruptible count:
2583 if (task_contributes_to_load(p)) {
2584 if (likely(cpu_online(orig_cpu)))
2585 rq->nr_uninterruptible--;
2587 this_rq()->nr_uninterruptible--;
2589 p->state = TASK_WAKING;
2591 if (p->sched_class->task_waking) {
2592 p->sched_class->task_waking(rq, p);
2593 en_flags |= ENQUEUE_WAKING;
2596 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2597 if (cpu != orig_cpu)
2598 set_task_cpu(p, cpu);
2599 __task_rq_unlock(rq);
2602 raw_spin_lock(&rq->lock);
2605 * We migrated the task without holding either rq->lock, however
2606 * since the task is not on the task list itself, nobody else
2607 * will try and migrate the task, hence the rq should match the
2608 * cpu we just moved it to.
2610 WARN_ON(task_cpu(p) != cpu);
2611 WARN_ON(p->state != TASK_WAKING);
2613 #ifdef CONFIG_SCHEDSTATS
2614 schedstat_inc(rq, ttwu_count);
2615 if (cpu == this_cpu)
2616 schedstat_inc(rq, ttwu_local);
2618 struct sched_domain *sd;
2619 for_each_domain(this_cpu, sd) {
2620 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2621 schedstat_inc(sd, ttwu_wake_remote);
2626 #endif /* CONFIG_SCHEDSTATS */
2629 #endif /* CONFIG_SMP */
2630 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2631 cpu == this_cpu, en_flags);
2634 ttwu_post_activation(p, rq, wake_flags, success);
2636 task_rq_unlock(rq, &flags);
2643 * try_to_wake_up_local - try to wake up a local task with rq lock held
2644 * @p: the thread to be awakened
2646 * Put @p on the run-queue if it's not alredy there. The caller must
2647 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2648 * the current task. this_rq() stays locked over invocation.
2650 static void try_to_wake_up_local(struct task_struct *p)
2652 struct rq *rq = task_rq(p);
2653 bool success = false;
2655 BUG_ON(rq != this_rq());
2656 BUG_ON(p == current);
2657 lockdep_assert_held(&rq->lock);
2659 if (!(p->state & TASK_NORMAL))
2663 if (likely(!task_running(rq, p))) {
2664 schedstat_inc(rq, ttwu_count);
2665 schedstat_inc(rq, ttwu_local);
2667 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2670 ttwu_post_activation(p, rq, 0, success);
2674 * wake_up_process - Wake up a specific process
2675 * @p: The process to be woken up.
2677 * Attempt to wake up the nominated process and move it to the set of runnable
2678 * processes. Returns 1 if the process was woken up, 0 if it was already
2681 * It may be assumed that this function implies a write memory barrier before
2682 * changing the task state if and only if any tasks are woken up.
2684 int wake_up_process(struct task_struct *p)
2686 return try_to_wake_up(p, TASK_ALL, 0);
2688 EXPORT_SYMBOL(wake_up_process);
2690 int wake_up_state(struct task_struct *p, unsigned int state)
2692 return try_to_wake_up(p, state, 0);
2696 * Perform scheduler related setup for a newly forked process p.
2697 * p is forked by current.
2699 * __sched_fork() is basic setup used by init_idle() too:
2701 static void __sched_fork(struct task_struct *p)
2703 p->se.exec_start = 0;
2704 p->se.sum_exec_runtime = 0;
2705 p->se.prev_sum_exec_runtime = 0;
2706 p->se.nr_migrations = 0;
2708 #ifdef CONFIG_SCHEDSTATS
2709 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2712 INIT_LIST_HEAD(&p->rt.run_list);
2714 INIT_LIST_HEAD(&p->se.group_node);
2716 #ifdef CONFIG_PREEMPT_NOTIFIERS
2717 INIT_HLIST_HEAD(&p->preempt_notifiers);
2722 * fork()/clone()-time setup:
2724 void sched_fork(struct task_struct *p, int clone_flags)
2726 int cpu = get_cpu();
2730 * We mark the process as running here. This guarantees that
2731 * nobody will actually run it, and a signal or other external
2732 * event cannot wake it up and insert it on the runqueue either.
2734 p->state = TASK_RUNNING;
2737 * Revert to default priority/policy on fork if requested.
2739 if (unlikely(p->sched_reset_on_fork)) {
2740 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2741 p->policy = SCHED_NORMAL;
2742 p->normal_prio = p->static_prio;
2745 if (PRIO_TO_NICE(p->static_prio) < 0) {
2746 p->static_prio = NICE_TO_PRIO(0);
2747 p->normal_prio = p->static_prio;
2752 * We don't need the reset flag anymore after the fork. It has
2753 * fulfilled its duty:
2755 p->sched_reset_on_fork = 0;
2759 * Make sure we do not leak PI boosting priority to the child.
2761 p->prio = current->normal_prio;
2763 if (!rt_prio(p->prio))
2764 p->sched_class = &fair_sched_class;
2766 if (p->sched_class->task_fork)
2767 p->sched_class->task_fork(p);
2770 * The child is not yet in the pid-hash so no cgroup attach races,
2771 * and the cgroup is pinned to this child due to cgroup_fork()
2772 * is ran before sched_fork().
2774 * Silence PROVE_RCU.
2777 set_task_cpu(p, cpu);
2780 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2781 if (likely(sched_info_on()))
2782 memset(&p->sched_info, 0, sizeof(p->sched_info));
2784 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2787 #ifdef CONFIG_PREEMPT
2788 /* Want to start with kernel preemption disabled. */
2789 task_thread_info(p)->preempt_count = 1;
2791 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2797 * wake_up_new_task - wake up a newly created task for the first time.
2799 * This function will do some initial scheduler statistics housekeeping
2800 * that must be done for every newly created context, then puts the task
2801 * on the runqueue and wakes it.
2803 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2805 unsigned long flags;
2807 int cpu __maybe_unused = get_cpu();
2810 rq = task_rq_lock(p, &flags);
2811 p->state = TASK_WAKING;
2814 * Fork balancing, do it here and not earlier because:
2815 * - cpus_allowed can change in the fork path
2816 * - any previously selected cpu might disappear through hotplug
2818 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2819 * without people poking at ->cpus_allowed.
2821 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2822 set_task_cpu(p, cpu);
2824 p->state = TASK_RUNNING;
2825 task_rq_unlock(rq, &flags);
2828 rq = task_rq_lock(p, &flags);
2829 activate_task(rq, p, 0);
2830 trace_sched_wakeup_new(p, 1);
2831 check_preempt_curr(rq, p, WF_FORK);
2833 if (p->sched_class->task_woken)
2834 p->sched_class->task_woken(rq, p);
2836 task_rq_unlock(rq, &flags);
2840 #ifdef CONFIG_PREEMPT_NOTIFIERS
2843 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2844 * @notifier: notifier struct to register
2846 void preempt_notifier_register(struct preempt_notifier *notifier)
2848 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2850 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2853 * preempt_notifier_unregister - no longer interested in preemption notifications
2854 * @notifier: notifier struct to unregister
2856 * This is safe to call from within a preemption notifier.
2858 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2860 hlist_del(¬ifier->link);
2862 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2864 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2866 struct preempt_notifier *notifier;
2867 struct hlist_node *node;
2869 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2870 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2874 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2875 struct task_struct *next)
2877 struct preempt_notifier *notifier;
2878 struct hlist_node *node;
2880 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2881 notifier->ops->sched_out(notifier, next);
2884 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2886 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2891 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2892 struct task_struct *next)
2896 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2899 * prepare_task_switch - prepare to switch tasks
2900 * @rq: the runqueue preparing to switch
2901 * @prev: the current task that is being switched out
2902 * @next: the task we are going to switch to.
2904 * This is called with the rq lock held and interrupts off. It must
2905 * be paired with a subsequent finish_task_switch after the context
2908 * prepare_task_switch sets up locking and calls architecture specific
2912 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2913 struct task_struct *next)
2915 fire_sched_out_preempt_notifiers(prev, next);
2916 prepare_lock_switch(rq, next);
2917 prepare_arch_switch(next);
2921 * finish_task_switch - clean up after a task-switch
2922 * @rq: runqueue associated with task-switch
2923 * @prev: the thread we just switched away from.
2925 * finish_task_switch must be called after the context switch, paired
2926 * with a prepare_task_switch call before the context switch.
2927 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2928 * and do any other architecture-specific cleanup actions.
2930 * Note that we may have delayed dropping an mm in context_switch(). If
2931 * so, we finish that here outside of the runqueue lock. (Doing it
2932 * with the lock held can cause deadlocks; see schedule() for
2935 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2936 __releases(rq->lock)
2938 struct mm_struct *mm = rq->prev_mm;
2944 * A task struct has one reference for the use as "current".
2945 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2946 * schedule one last time. The schedule call will never return, and
2947 * the scheduled task must drop that reference.
2948 * The test for TASK_DEAD must occur while the runqueue locks are
2949 * still held, otherwise prev could be scheduled on another cpu, die
2950 * there before we look at prev->state, and then the reference would
2952 * Manfred Spraul <manfred@colorfullife.com>
2954 prev_state = prev->state;
2955 finish_arch_switch(prev);
2956 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2957 local_irq_disable();
2958 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2959 perf_event_task_sched_in(current);
2960 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2962 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2963 finish_lock_switch(rq, prev);
2965 fire_sched_in_preempt_notifiers(current);
2968 if (unlikely(prev_state == TASK_DEAD)) {
2970 * Remove function-return probe instances associated with this
2971 * task and put them back on the free list.
2973 kprobe_flush_task(prev);
2974 put_task_struct(prev);
2980 /* assumes rq->lock is held */
2981 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2983 if (prev->sched_class->pre_schedule)
2984 prev->sched_class->pre_schedule(rq, prev);
2987 /* rq->lock is NOT held, but preemption is disabled */
2988 static inline void post_schedule(struct rq *rq)
2990 if (rq->post_schedule) {
2991 unsigned long flags;
2993 raw_spin_lock_irqsave(&rq->lock, flags);
2994 if (rq->curr->sched_class->post_schedule)
2995 rq->curr->sched_class->post_schedule(rq);
2996 raw_spin_unlock_irqrestore(&rq->lock, flags);
2998 rq->post_schedule = 0;
3004 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3008 static inline void post_schedule(struct rq *rq)
3015 * schedule_tail - first thing a freshly forked thread must call.
3016 * @prev: the thread we just switched away from.
3018 asmlinkage void schedule_tail(struct task_struct *prev)
3019 __releases(rq->lock)
3021 struct rq *rq = this_rq();
3023 finish_task_switch(rq, prev);
3026 * FIXME: do we need to worry about rq being invalidated by the
3031 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3032 /* In this case, finish_task_switch does not reenable preemption */
3035 if (current->set_child_tid)
3036 put_user(task_pid_vnr(current), current->set_child_tid);
3040 * context_switch - switch to the new MM and the new
3041 * thread's register state.
3044 context_switch(struct rq *rq, struct task_struct *prev,
3045 struct task_struct *next)
3047 struct mm_struct *mm, *oldmm;
3049 prepare_task_switch(rq, prev, next);
3050 trace_sched_switch(prev, next);
3052 oldmm = prev->active_mm;
3054 * For paravirt, this is coupled with an exit in switch_to to
3055 * combine the page table reload and the switch backend into
3058 arch_start_context_switch(prev);
3061 next->active_mm = oldmm;
3062 atomic_inc(&oldmm->mm_count);
3063 enter_lazy_tlb(oldmm, next);
3065 switch_mm(oldmm, mm, next);
3068 prev->active_mm = NULL;
3069 rq->prev_mm = oldmm;
3072 * Since the runqueue lock will be released by the next
3073 * task (which is an invalid locking op but in the case
3074 * of the scheduler it's an obvious special-case), so we
3075 * do an early lockdep release here:
3077 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3078 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3081 /* Here we just switch the register state and the stack. */
3082 switch_to(prev, next, prev);
3086 * this_rq must be evaluated again because prev may have moved
3087 * CPUs since it called schedule(), thus the 'rq' on its stack
3088 * frame will be invalid.
3090 finish_task_switch(this_rq(), prev);
3094 * nr_running, nr_uninterruptible and nr_context_switches:
3096 * externally visible scheduler statistics: current number of runnable
3097 * threads, current number of uninterruptible-sleeping threads, total
3098 * number of context switches performed since bootup.
3100 unsigned long nr_running(void)
3102 unsigned long i, sum = 0;
3104 for_each_online_cpu(i)
3105 sum += cpu_rq(i)->nr_running;
3110 unsigned long nr_uninterruptible(void)
3112 unsigned long i, sum = 0;
3114 for_each_possible_cpu(i)
3115 sum += cpu_rq(i)->nr_uninterruptible;
3118 * Since we read the counters lockless, it might be slightly
3119 * inaccurate. Do not allow it to go below zero though:
3121 if (unlikely((long)sum < 0))
3127 unsigned long long nr_context_switches(void)
3130 unsigned long long sum = 0;
3132 for_each_possible_cpu(i)
3133 sum += cpu_rq(i)->nr_switches;
3138 unsigned long nr_iowait(void)
3140 unsigned long i, sum = 0;
3142 for_each_possible_cpu(i)
3143 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3148 unsigned long nr_iowait_cpu(int cpu)
3150 struct rq *this = cpu_rq(cpu);
3151 return atomic_read(&this->nr_iowait);
3154 unsigned long this_cpu_load(void)
3156 struct rq *this = this_rq();
3157 return this->cpu_load[0];
3161 /* Variables and functions for calc_load */
3162 static atomic_long_t calc_load_tasks;
3163 static unsigned long calc_load_update;
3164 unsigned long avenrun[3];
3165 EXPORT_SYMBOL(avenrun);
3167 static long calc_load_fold_active(struct rq *this_rq)
3169 long nr_active, delta = 0;
3171 nr_active = this_rq->nr_running;
3172 nr_active += (long) this_rq->nr_uninterruptible;
3174 if (nr_active != this_rq->calc_load_active) {
3175 delta = nr_active - this_rq->calc_load_active;
3176 this_rq->calc_load_active = nr_active;
3182 static unsigned long
3183 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3186 load += active * (FIXED_1 - exp);
3187 load += 1UL << (FSHIFT - 1);
3188 return load >> FSHIFT;
3193 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3195 * When making the ILB scale, we should try to pull this in as well.
3197 static atomic_long_t calc_load_tasks_idle;
3199 static void calc_load_account_idle(struct rq *this_rq)
3203 delta = calc_load_fold_active(this_rq);
3205 atomic_long_add(delta, &calc_load_tasks_idle);
3208 static long calc_load_fold_idle(void)
3213 * Its got a race, we don't care...
3215 if (atomic_long_read(&calc_load_tasks_idle))
3216 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3222 * fixed_power_int - compute: x^n, in O(log n) time
3224 * @x: base of the power
3225 * @frac_bits: fractional bits of @x
3226 * @n: power to raise @x to.
3228 * By exploiting the relation between the definition of the natural power
3229 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3230 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3231 * (where: n_i \elem {0, 1}, the binary vector representing n),
3232 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3233 * of course trivially computable in O(log_2 n), the length of our binary
3236 static unsigned long
3237 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3239 unsigned long result = 1UL << frac_bits;
3244 result += 1UL << (frac_bits - 1);
3245 result >>= frac_bits;
3251 x += 1UL << (frac_bits - 1);
3259 * a1 = a0 * e + a * (1 - e)
3261 * a2 = a1 * e + a * (1 - e)
3262 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3263 * = a0 * e^2 + a * (1 - e) * (1 + e)
3265 * a3 = a2 * e + a * (1 - e)
3266 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3267 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3271 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3272 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3273 * = a0 * e^n + a * (1 - e^n)
3275 * [1] application of the geometric series:
3278 * S_n := \Sum x^i = -------------
3281 static unsigned long
3282 calc_load_n(unsigned long load, unsigned long exp,
3283 unsigned long active, unsigned int n)
3286 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3290 * NO_HZ can leave us missing all per-cpu ticks calling
3291 * calc_load_account_active(), but since an idle CPU folds its delta into
3292 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3293 * in the pending idle delta if our idle period crossed a load cycle boundary.
3295 * Once we've updated the global active value, we need to apply the exponential
3296 * weights adjusted to the number of cycles missed.
3298 static void calc_global_nohz(unsigned long ticks)
3300 long delta, active, n;
3302 if (time_before(jiffies, calc_load_update))
3306 * If we crossed a calc_load_update boundary, make sure to fold
3307 * any pending idle changes, the respective CPUs might have
3308 * missed the tick driven calc_load_account_active() update
3311 delta = calc_load_fold_idle();
3313 atomic_long_add(delta, &calc_load_tasks);
3316 * If we were idle for multiple load cycles, apply them.
3318 if (ticks >= LOAD_FREQ) {
3319 n = ticks / LOAD_FREQ;
3321 active = atomic_long_read(&calc_load_tasks);
3322 active = active > 0 ? active * FIXED_1 : 0;
3324 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3325 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3326 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3328 calc_load_update += n * LOAD_FREQ;
3332 * Its possible the remainder of the above division also crosses
3333 * a LOAD_FREQ period, the regular check in calc_global_load()
3334 * which comes after this will take care of that.
3336 * Consider us being 11 ticks before a cycle completion, and us
3337 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3338 * age us 4 cycles, and the test in calc_global_load() will
3339 * pick up the final one.
3343 static void calc_load_account_idle(struct rq *this_rq)
3347 static inline long calc_load_fold_idle(void)
3352 static void calc_global_nohz(unsigned long ticks)
3358 * get_avenrun - get the load average array
3359 * @loads: pointer to dest load array
3360 * @offset: offset to add
3361 * @shift: shift count to shift the result left
3363 * These values are estimates at best, so no need for locking.
3365 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3367 loads[0] = (avenrun[0] + offset) << shift;
3368 loads[1] = (avenrun[1] + offset) << shift;
3369 loads[2] = (avenrun[2] + offset) << shift;
3373 * calc_load - update the avenrun load estimates 10 ticks after the
3374 * CPUs have updated calc_load_tasks.
3376 void calc_global_load(unsigned long ticks)
3380 calc_global_nohz(ticks);
3382 if (time_before(jiffies, calc_load_update + 10))
3385 active = atomic_long_read(&calc_load_tasks);
3386 active = active > 0 ? active * FIXED_1 : 0;
3388 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3389 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3390 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3392 calc_load_update += LOAD_FREQ;
3396 * Called from update_cpu_load() to periodically update this CPU's
3399 static void calc_load_account_active(struct rq *this_rq)
3403 if (time_before(jiffies, this_rq->calc_load_update))
3406 delta = calc_load_fold_active(this_rq);
3407 delta += calc_load_fold_idle();
3409 atomic_long_add(delta, &calc_load_tasks);
3411 this_rq->calc_load_update += LOAD_FREQ;
3415 * The exact cpuload at various idx values, calculated at every tick would be
3416 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3418 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3419 * on nth tick when cpu may be busy, then we have:
3420 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3421 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3423 * decay_load_missed() below does efficient calculation of
3424 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3425 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3427 * The calculation is approximated on a 128 point scale.
3428 * degrade_zero_ticks is the number of ticks after which load at any
3429 * particular idx is approximated to be zero.
3430 * degrade_factor is a precomputed table, a row for each load idx.
3431 * Each column corresponds to degradation factor for a power of two ticks,
3432 * based on 128 point scale.
3434 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3435 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3437 * With this power of 2 load factors, we can degrade the load n times
3438 * by looking at 1 bits in n and doing as many mult/shift instead of
3439 * n mult/shifts needed by the exact degradation.
3441 #define DEGRADE_SHIFT 7
3442 static const unsigned char
3443 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3444 static const unsigned char
3445 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3446 {0, 0, 0, 0, 0, 0, 0, 0},
3447 {64, 32, 8, 0, 0, 0, 0, 0},
3448 {96, 72, 40, 12, 1, 0, 0},
3449 {112, 98, 75, 43, 15, 1, 0},
3450 {120, 112, 98, 76, 45, 16, 2} };
3453 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3454 * would be when CPU is idle and so we just decay the old load without
3455 * adding any new load.
3457 static unsigned long
3458 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3462 if (!missed_updates)
3465 if (missed_updates >= degrade_zero_ticks[idx])
3469 return load >> missed_updates;
3471 while (missed_updates) {
3472 if (missed_updates % 2)
3473 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3475 missed_updates >>= 1;
3482 * Update rq->cpu_load[] statistics. This function is usually called every
3483 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3484 * every tick. We fix it up based on jiffies.
3486 static void update_cpu_load(struct rq *this_rq)
3488 unsigned long this_load = this_rq->load.weight;
3489 unsigned long curr_jiffies = jiffies;
3490 unsigned long pending_updates;
3493 this_rq->nr_load_updates++;
3495 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3496 if (curr_jiffies == this_rq->last_load_update_tick)
3499 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3500 this_rq->last_load_update_tick = curr_jiffies;
3502 /* Update our load: */
3503 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3504 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3505 unsigned long old_load, new_load;
3507 /* scale is effectively 1 << i now, and >> i divides by scale */
3509 old_load = this_rq->cpu_load[i];
3510 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3511 new_load = this_load;
3513 * Round up the averaging division if load is increasing. This
3514 * prevents us from getting stuck on 9 if the load is 10, for
3517 if (new_load > old_load)
3518 new_load += scale - 1;
3520 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3523 sched_avg_update(this_rq);
3526 static void update_cpu_load_active(struct rq *this_rq)
3528 update_cpu_load(this_rq);
3530 calc_load_account_active(this_rq);
3536 * sched_exec - execve() is a valuable balancing opportunity, because at
3537 * this point the task has the smallest effective memory and cache footprint.
3539 void sched_exec(void)
3541 struct task_struct *p = current;
3542 unsigned long flags;
3546 rq = task_rq_lock(p, &flags);
3547 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3548 if (dest_cpu == smp_processor_id())
3552 * select_task_rq() can race against ->cpus_allowed
3554 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3555 likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
3556 struct migration_arg arg = { p, dest_cpu };
3558 task_rq_unlock(rq, &flags);
3559 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3563 task_rq_unlock(rq, &flags);
3568 DEFINE_PER_CPU(struct kernel_stat, kstat);
3570 EXPORT_PER_CPU_SYMBOL(kstat);
3573 * Return any ns on the sched_clock that have not yet been accounted in
3574 * @p in case that task is currently running.
3576 * Called with task_rq_lock() held on @rq.
3578 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3582 if (task_current(rq, p)) {
3583 update_rq_clock(rq);
3584 ns = rq->clock_task - p->se.exec_start;
3592 unsigned long long task_delta_exec(struct task_struct *p)
3594 unsigned long flags;
3598 rq = task_rq_lock(p, &flags);
3599 ns = do_task_delta_exec(p, rq);
3600 task_rq_unlock(rq, &flags);
3606 * Return accounted runtime for the task.
3607 * In case the task is currently running, return the runtime plus current's
3608 * pending runtime that have not been accounted yet.
3610 unsigned long long task_sched_runtime(struct task_struct *p)
3612 unsigned long flags;
3616 rq = task_rq_lock(p, &flags);
3617 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3618 task_rq_unlock(rq, &flags);
3624 * Return sum_exec_runtime for the thread group.
3625 * In case the task is currently running, return the sum plus current's
3626 * pending runtime that have not been accounted yet.
3628 * Note that the thread group might have other running tasks as well,
3629 * so the return value not includes other pending runtime that other
3630 * running tasks might have.
3632 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3634 struct task_cputime totals;
3635 unsigned long flags;
3639 rq = task_rq_lock(p, &flags);
3640 thread_group_cputime(p, &totals);
3641 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3642 task_rq_unlock(rq, &flags);
3648 * Account user cpu time to a process.
3649 * @p: the process that the cpu time gets accounted to
3650 * @cputime: the cpu time spent in user space since the last update
3651 * @cputime_scaled: cputime scaled by cpu frequency
3653 void account_user_time(struct task_struct *p, cputime_t cputime,
3654 cputime_t cputime_scaled)
3656 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3659 /* Add user time to process. */
3660 p->utime = cputime_add(p->utime, cputime);
3661 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3662 account_group_user_time(p, cputime);
3664 /* Add user time to cpustat. */
3665 tmp = cputime_to_cputime64(cputime);
3666 if (TASK_NICE(p) > 0)
3667 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3669 cpustat->user = cputime64_add(cpustat->user, tmp);
3671 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3672 /* Account for user time used */
3673 acct_update_integrals(p);
3677 * Account guest cpu time to a process.
3678 * @p: the process that the cpu time gets accounted to
3679 * @cputime: the cpu time spent in virtual machine since the last update
3680 * @cputime_scaled: cputime scaled by cpu frequency
3682 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3683 cputime_t cputime_scaled)
3686 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3688 tmp = cputime_to_cputime64(cputime);
3690 /* Add guest time to process. */
3691 p->utime = cputime_add(p->utime, cputime);
3692 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3693 account_group_user_time(p, cputime);
3694 p->gtime = cputime_add(p->gtime, cputime);
3696 /* Add guest time to cpustat. */
3697 if (TASK_NICE(p) > 0) {
3698 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3699 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3701 cpustat->user = cputime64_add(cpustat->user, tmp);
3702 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3707 * Account system cpu time to a process.
3708 * @p: the process that the cpu time gets accounted to
3709 * @hardirq_offset: the offset to subtract from hardirq_count()
3710 * @cputime: the cpu time spent in kernel space since the last update
3711 * @cputime_scaled: cputime scaled by cpu frequency
3713 void account_system_time(struct task_struct *p, int hardirq_offset,
3714 cputime_t cputime, cputime_t cputime_scaled)
3716 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3719 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3720 account_guest_time(p, cputime, cputime_scaled);
3724 /* Add system time to process. */
3725 p->stime = cputime_add(p->stime, cputime);
3726 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3727 account_group_system_time(p, cputime);
3729 /* Add system time to cpustat. */
3730 tmp = cputime_to_cputime64(cputime);
3731 if (hardirq_count() - hardirq_offset)
3732 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3733 else if (in_serving_softirq())
3734 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3736 cpustat->system = cputime64_add(cpustat->system, tmp);
3738 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3740 /* Account for system time used */
3741 acct_update_integrals(p);
3745 * Account for involuntary wait time.
3746 * @steal: the cpu time spent in involuntary wait
3748 void account_steal_time(cputime_t cputime)
3750 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3751 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3753 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3757 * Account for idle time.
3758 * @cputime: the cpu time spent in idle wait
3760 void account_idle_time(cputime_t cputime)
3762 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3763 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3764 struct rq *rq = this_rq();
3766 if (atomic_read(&rq->nr_iowait) > 0)
3767 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3769 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3772 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3775 * Account a single tick of cpu time.
3776 * @p: the process that the cpu time gets accounted to
3777 * @user_tick: indicates if the tick is a user or a system tick
3779 void account_process_tick(struct task_struct *p, int user_tick)
3781 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3782 struct rq *rq = this_rq();
3785 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3786 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3787 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3790 account_idle_time(cputime_one_jiffy);
3794 * Account multiple ticks of steal time.
3795 * @p: the process from which the cpu time has been stolen
3796 * @ticks: number of stolen ticks
3798 void account_steal_ticks(unsigned long ticks)
3800 account_steal_time(jiffies_to_cputime(ticks));
3804 * Account multiple ticks of idle time.
3805 * @ticks: number of stolen ticks
3807 void account_idle_ticks(unsigned long ticks)
3809 account_idle_time(jiffies_to_cputime(ticks));
3815 * Use precise platform statistics if available:
3817 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3818 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3824 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3826 struct task_cputime cputime;
3828 thread_group_cputime(p, &cputime);
3830 *ut = cputime.utime;
3831 *st = cputime.stime;
3835 #ifndef nsecs_to_cputime
3836 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3839 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3841 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3844 * Use CFS's precise accounting:
3846 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3852 do_div(temp, total);
3853 utime = (cputime_t)temp;
3858 * Compare with previous values, to keep monotonicity:
3860 p->prev_utime = max(p->prev_utime, utime);
3861 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3863 *ut = p->prev_utime;
3864 *st = p->prev_stime;
3868 * Must be called with siglock held.
3870 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3872 struct signal_struct *sig = p->signal;
3873 struct task_cputime cputime;
3874 cputime_t rtime, utime, total;
3876 thread_group_cputime(p, &cputime);
3878 total = cputime_add(cputime.utime, cputime.stime);
3879 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3884 temp *= cputime.utime;
3885 do_div(temp, total);
3886 utime = (cputime_t)temp;
3890 sig->prev_utime = max(sig->prev_utime, utime);
3891 sig->prev_stime = max(sig->prev_stime,
3892 cputime_sub(rtime, sig->prev_utime));
3894 *ut = sig->prev_utime;
3895 *st = sig->prev_stime;
3900 * This function gets called by the timer code, with HZ frequency.
3901 * We call it with interrupts disabled.
3903 * It also gets called by the fork code, when changing the parent's
3906 void scheduler_tick(void)
3908 int cpu = smp_processor_id();
3909 struct rq *rq = cpu_rq(cpu);
3910 struct task_struct *curr = rq->curr;
3914 raw_spin_lock(&rq->lock);
3915 update_rq_clock(rq);
3916 update_cpu_load_active(rq);
3917 curr->sched_class->task_tick(rq, curr, 0);
3918 raw_spin_unlock(&rq->lock);
3920 perf_event_task_tick();
3923 rq->idle_at_tick = idle_cpu(cpu);
3924 trigger_load_balance(rq, cpu);
3928 notrace unsigned long get_parent_ip(unsigned long addr)
3930 if (in_lock_functions(addr)) {
3931 addr = CALLER_ADDR2;
3932 if (in_lock_functions(addr))
3933 addr = CALLER_ADDR3;
3938 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3939 defined(CONFIG_PREEMPT_TRACER))
3941 void __kprobes add_preempt_count(int val)
3943 #ifdef CONFIG_DEBUG_PREEMPT
3947 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3950 preempt_count() += val;
3951 #ifdef CONFIG_DEBUG_PREEMPT
3953 * Spinlock count overflowing soon?
3955 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3958 if (preempt_count() == val)
3959 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3961 EXPORT_SYMBOL(add_preempt_count);
3963 void __kprobes sub_preempt_count(int val)
3965 #ifdef CONFIG_DEBUG_PREEMPT
3969 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3972 * Is the spinlock portion underflowing?
3974 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3975 !(preempt_count() & PREEMPT_MASK)))
3979 if (preempt_count() == val)
3980 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3981 preempt_count() -= val;
3983 EXPORT_SYMBOL(sub_preempt_count);
3988 * Print scheduling while atomic bug:
3990 static noinline void __schedule_bug(struct task_struct *prev)
3992 struct pt_regs *regs = get_irq_regs();
3994 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3995 prev->comm, prev->pid, preempt_count());
3997 debug_show_held_locks(prev);
3999 if (irqs_disabled())
4000 print_irqtrace_events(prev);
4009 * Various schedule()-time debugging checks and statistics:
4011 static inline void schedule_debug(struct task_struct *prev)
4014 * Test if we are atomic. Since do_exit() needs to call into
4015 * schedule() atomically, we ignore that path for now.
4016 * Otherwise, whine if we are scheduling when we should not be.
4018 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4019 __schedule_bug(prev);
4021 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4023 schedstat_inc(this_rq(), sched_count);
4024 #ifdef CONFIG_SCHEDSTATS
4025 if (unlikely(prev->lock_depth >= 0)) {
4026 schedstat_inc(this_rq(), bkl_count);
4027 schedstat_inc(prev, sched_info.bkl_count);
4032 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4035 update_rq_clock(rq);
4036 prev->sched_class->put_prev_task(rq, prev);
4040 * Pick up the highest-prio task:
4042 static inline struct task_struct *
4043 pick_next_task(struct rq *rq)
4045 const struct sched_class *class;
4046 struct task_struct *p;
4049 * Optimization: we know that if all tasks are in
4050 * the fair class we can call that function directly:
4052 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4053 p = fair_sched_class.pick_next_task(rq);
4058 for_each_class(class) {
4059 p = class->pick_next_task(rq);
4064 BUG(); /* the idle class will always have a runnable task */
4068 * schedule() is the main scheduler function.
4070 asmlinkage void __sched schedule(void)
4072 struct task_struct *prev, *next;
4073 unsigned long *switch_count;
4079 cpu = smp_processor_id();
4081 rcu_note_context_switch(cpu);
4084 release_kernel_lock(prev);
4085 need_resched_nonpreemptible:
4087 schedule_debug(prev);
4089 if (sched_feat(HRTICK))
4092 raw_spin_lock_irq(&rq->lock);
4094 switch_count = &prev->nivcsw;
4095 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4096 if (unlikely(signal_pending_state(prev->state, prev))) {
4097 prev->state = TASK_RUNNING;
4100 * If a worker is going to sleep, notify and
4101 * ask workqueue whether it wants to wake up a
4102 * task to maintain concurrency. If so, wake
4105 if (prev->flags & PF_WQ_WORKER) {
4106 struct task_struct *to_wakeup;
4108 to_wakeup = wq_worker_sleeping(prev, cpu);
4110 try_to_wake_up_local(to_wakeup);
4112 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4114 switch_count = &prev->nvcsw;
4117 pre_schedule(rq, prev);
4119 if (unlikely(!rq->nr_running))
4120 idle_balance(cpu, rq);
4122 put_prev_task(rq, prev);
4123 next = pick_next_task(rq);
4124 clear_tsk_need_resched(prev);
4125 rq->skip_clock_update = 0;
4127 if (likely(prev != next)) {
4128 sched_info_switch(prev, next);
4129 perf_event_task_sched_out(prev, next);
4135 context_switch(rq, prev, next); /* unlocks the rq */
4137 * The context switch have flipped the stack from under us
4138 * and restored the local variables which were saved when
4139 * this task called schedule() in the past. prev == current
4140 * is still correct, but it can be moved to another cpu/rq.
4142 cpu = smp_processor_id();
4145 raw_spin_unlock_irq(&rq->lock);
4149 if (unlikely(reacquire_kernel_lock(prev)))
4150 goto need_resched_nonpreemptible;
4152 preempt_enable_no_resched();
4156 EXPORT_SYMBOL(schedule);
4158 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4160 * Look out! "owner" is an entirely speculative pointer
4161 * access and not reliable.
4163 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
4168 if (!sched_feat(OWNER_SPIN))
4171 #ifdef CONFIG_DEBUG_PAGEALLOC
4173 * Need to access the cpu field knowing that
4174 * DEBUG_PAGEALLOC could have unmapped it if
4175 * the mutex owner just released it and exited.
4177 if (probe_kernel_address(&owner->cpu, cpu))
4184 * Even if the access succeeded (likely case),
4185 * the cpu field may no longer be valid.
4187 if (cpu >= nr_cpumask_bits)
4191 * We need to validate that we can do a
4192 * get_cpu() and that we have the percpu area.
4194 if (!cpu_online(cpu))
4201 * Owner changed, break to re-assess state.
4203 if (lock->owner != owner) {
4205 * If the lock has switched to a different owner,
4206 * we likely have heavy contention. Return 0 to quit
4207 * optimistic spinning and not contend further:
4215 * Is that owner really running on that cpu?
4217 if (task_thread_info(rq->curr) != owner || need_resched())
4227 #ifdef CONFIG_PREEMPT
4229 * this is the entry point to schedule() from in-kernel preemption
4230 * off of preempt_enable. Kernel preemptions off return from interrupt
4231 * occur there and call schedule directly.
4233 asmlinkage void __sched notrace preempt_schedule(void)
4235 struct thread_info *ti = current_thread_info();
4238 * If there is a non-zero preempt_count or interrupts are disabled,
4239 * we do not want to preempt the current task. Just return..
4241 if (likely(ti->preempt_count || irqs_disabled()))
4245 add_preempt_count_notrace(PREEMPT_ACTIVE);
4247 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4250 * Check again in case we missed a preemption opportunity
4251 * between schedule and now.
4254 } while (need_resched());
4256 EXPORT_SYMBOL(preempt_schedule);
4259 * this is the entry point to schedule() from kernel preemption
4260 * off of irq context.
4261 * Note, that this is called and return with irqs disabled. This will
4262 * protect us against recursive calling from irq.
4264 asmlinkage void __sched preempt_schedule_irq(void)
4266 struct thread_info *ti = current_thread_info();
4268 /* Catch callers which need to be fixed */
4269 BUG_ON(ti->preempt_count || !irqs_disabled());
4272 add_preempt_count(PREEMPT_ACTIVE);
4275 local_irq_disable();
4276 sub_preempt_count(PREEMPT_ACTIVE);
4279 * Check again in case we missed a preemption opportunity
4280 * between schedule and now.
4283 } while (need_resched());
4286 #endif /* CONFIG_PREEMPT */
4288 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4291 return try_to_wake_up(curr->private, mode, wake_flags);
4293 EXPORT_SYMBOL(default_wake_function);
4296 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4297 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4298 * number) then we wake all the non-exclusive tasks and one exclusive task.
4300 * There are circumstances in which we can try to wake a task which has already
4301 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4302 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4304 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4305 int nr_exclusive, int wake_flags, void *key)
4307 wait_queue_t *curr, *next;
4309 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4310 unsigned flags = curr->flags;
4312 if (curr->func(curr, mode, wake_flags, key) &&
4313 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4319 * __wake_up - wake up threads blocked on a waitqueue.
4321 * @mode: which threads
4322 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4323 * @key: is directly passed to the wakeup function
4325 * It may be assumed that this function implies a write memory barrier before
4326 * changing the task state if and only if any tasks are woken up.
4328 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4329 int nr_exclusive, void *key)
4331 unsigned long flags;
4333 spin_lock_irqsave(&q->lock, flags);
4334 __wake_up_common(q, mode, nr_exclusive, 0, key);
4335 spin_unlock_irqrestore(&q->lock, flags);
4337 EXPORT_SYMBOL(__wake_up);
4340 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4342 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4344 __wake_up_common(q, mode, 1, 0, NULL);
4346 EXPORT_SYMBOL_GPL(__wake_up_locked);
4348 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4350 __wake_up_common(q, mode, 1, 0, key);
4354 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4356 * @mode: which threads
4357 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4358 * @key: opaque value to be passed to wakeup targets
4360 * The sync wakeup differs that the waker knows that it will schedule
4361 * away soon, so while the target thread will be woken up, it will not
4362 * be migrated to another CPU - ie. the two threads are 'synchronized'
4363 * with each other. This can prevent needless bouncing between CPUs.
4365 * On UP it can prevent extra preemption.
4367 * It may be assumed that this function implies a write memory barrier before
4368 * changing the task state if and only if any tasks are woken up.
4370 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4371 int nr_exclusive, void *key)
4373 unsigned long flags;
4374 int wake_flags = WF_SYNC;
4379 if (unlikely(!nr_exclusive))
4382 spin_lock_irqsave(&q->lock, flags);
4383 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4384 spin_unlock_irqrestore(&q->lock, flags);
4386 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4389 * __wake_up_sync - see __wake_up_sync_key()
4391 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4393 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4395 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4398 * complete: - signals a single thread waiting on this completion
4399 * @x: holds the state of this particular completion
4401 * This will wake up a single thread waiting on this completion. Threads will be
4402 * awakened in the same order in which they were queued.
4404 * See also complete_all(), wait_for_completion() and related routines.
4406 * It may be assumed that this function implies a write memory barrier before
4407 * changing the task state if and only if any tasks are woken up.
4409 void complete(struct completion *x)
4411 unsigned long flags;
4413 spin_lock_irqsave(&x->wait.lock, flags);
4415 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4416 spin_unlock_irqrestore(&x->wait.lock, flags);
4418 EXPORT_SYMBOL(complete);
4421 * complete_all: - signals all threads waiting on this completion
4422 * @x: holds the state of this particular completion
4424 * This will wake up all threads waiting on this particular completion event.
4426 * It may be assumed that this function implies a write memory barrier before
4427 * changing the task state if and only if any tasks are woken up.
4429 void complete_all(struct completion *x)
4431 unsigned long flags;
4433 spin_lock_irqsave(&x->wait.lock, flags);
4434 x->done += UINT_MAX/2;
4435 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4436 spin_unlock_irqrestore(&x->wait.lock, flags);
4438 EXPORT_SYMBOL(complete_all);
4440 static inline long __sched
4441 do_wait_for_common(struct completion *x, long timeout, int state)
4444 DECLARE_WAITQUEUE(wait, current);
4446 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4448 if (signal_pending_state(state, current)) {
4449 timeout = -ERESTARTSYS;
4452 __set_current_state(state);
4453 spin_unlock_irq(&x->wait.lock);
4454 timeout = schedule_timeout(timeout);
4455 spin_lock_irq(&x->wait.lock);
4456 } while (!x->done && timeout);
4457 __remove_wait_queue(&x->wait, &wait);
4462 return timeout ?: 1;
4466 wait_for_common(struct completion *x, long timeout, int state)
4470 spin_lock_irq(&x->wait.lock);
4471 timeout = do_wait_for_common(x, timeout, state);
4472 spin_unlock_irq(&x->wait.lock);
4477 * wait_for_completion: - waits for completion of a task
4478 * @x: holds the state of this particular completion
4480 * This waits to be signaled for completion of a specific task. It is NOT
4481 * interruptible and there is no timeout.
4483 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4484 * and interrupt capability. Also see complete().
4486 void __sched wait_for_completion(struct completion *x)
4488 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4490 EXPORT_SYMBOL(wait_for_completion);
4493 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4494 * @x: holds the state of this particular completion
4495 * @timeout: timeout value in jiffies
4497 * This waits for either a completion of a specific task to be signaled or for a
4498 * specified timeout to expire. The timeout is in jiffies. It is not
4501 unsigned long __sched
4502 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4504 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4506 EXPORT_SYMBOL(wait_for_completion_timeout);
4509 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4510 * @x: holds the state of this particular completion
4512 * This waits for completion of a specific task to be signaled. It is
4515 int __sched wait_for_completion_interruptible(struct completion *x)
4517 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4518 if (t == -ERESTARTSYS)
4522 EXPORT_SYMBOL(wait_for_completion_interruptible);
4525 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4526 * @x: holds the state of this particular completion
4527 * @timeout: timeout value in jiffies
4529 * This waits for either a completion of a specific task to be signaled or for a
4530 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4533 wait_for_completion_interruptible_timeout(struct completion *x,
4534 unsigned long timeout)
4536 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4538 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4541 * wait_for_completion_killable: - waits for completion of a task (killable)
4542 * @x: holds the state of this particular completion
4544 * This waits to be signaled for completion of a specific task. It can be
4545 * interrupted by a kill signal.
4547 int __sched wait_for_completion_killable(struct completion *x)
4549 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4550 if (t == -ERESTARTSYS)
4554 EXPORT_SYMBOL(wait_for_completion_killable);
4557 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4558 * @x: holds the state of this particular completion
4559 * @timeout: timeout value in jiffies
4561 * This waits for either a completion of a specific task to be
4562 * signaled or for a specified timeout to expire. It can be
4563 * interrupted by a kill signal. The timeout is in jiffies.
4566 wait_for_completion_killable_timeout(struct completion *x,
4567 unsigned long timeout)
4569 return wait_for_common(x, timeout, TASK_KILLABLE);
4571 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4574 * try_wait_for_completion - try to decrement a completion without blocking
4575 * @x: completion structure
4577 * Returns: 0 if a decrement cannot be done without blocking
4578 * 1 if a decrement succeeded.
4580 * If a completion is being used as a counting completion,
4581 * attempt to decrement the counter without blocking. This
4582 * enables us to avoid waiting if the resource the completion
4583 * is protecting is not available.
4585 bool try_wait_for_completion(struct completion *x)
4587 unsigned long flags;
4590 spin_lock_irqsave(&x->wait.lock, flags);
4595 spin_unlock_irqrestore(&x->wait.lock, flags);
4598 EXPORT_SYMBOL(try_wait_for_completion);
4601 * completion_done - Test to see if a completion has any waiters
4602 * @x: completion structure
4604 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4605 * 1 if there are no waiters.
4608 bool completion_done(struct completion *x)
4610 unsigned long flags;
4613 spin_lock_irqsave(&x->wait.lock, flags);
4616 spin_unlock_irqrestore(&x->wait.lock, flags);
4619 EXPORT_SYMBOL(completion_done);
4622 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4624 unsigned long flags;
4627 init_waitqueue_entry(&wait, current);
4629 __set_current_state(state);
4631 spin_lock_irqsave(&q->lock, flags);
4632 __add_wait_queue(q, &wait);
4633 spin_unlock(&q->lock);
4634 timeout = schedule_timeout(timeout);
4635 spin_lock_irq(&q->lock);
4636 __remove_wait_queue(q, &wait);
4637 spin_unlock_irqrestore(&q->lock, flags);
4642 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4644 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4646 EXPORT_SYMBOL(interruptible_sleep_on);
4649 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4651 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4653 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4655 void __sched sleep_on(wait_queue_head_t *q)
4657 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4659 EXPORT_SYMBOL(sleep_on);
4661 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4663 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4665 EXPORT_SYMBOL(sleep_on_timeout);
4667 #ifdef CONFIG_RT_MUTEXES
4670 * rt_mutex_setprio - set the current priority of a task
4672 * @prio: prio value (kernel-internal form)
4674 * This function changes the 'effective' priority of a task. It does
4675 * not touch ->normal_prio like __setscheduler().
4677 * Used by the rt_mutex code to implement priority inheritance logic.
4679 void rt_mutex_setprio(struct task_struct *p, int prio)
4681 unsigned long flags;
4682 int oldprio, on_rq, running;
4684 const struct sched_class *prev_class;
4686 BUG_ON(prio < 0 || prio > MAX_PRIO);
4688 rq = task_rq_lock(p, &flags);
4690 trace_sched_pi_setprio(p, prio);
4692 prev_class = p->sched_class;
4693 on_rq = p->se.on_rq;
4694 running = task_current(rq, p);
4696 dequeue_task(rq, p, 0);
4698 p->sched_class->put_prev_task(rq, p);
4701 p->sched_class = &rt_sched_class;
4703 p->sched_class = &fair_sched_class;
4708 p->sched_class->set_curr_task(rq);
4710 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4712 check_class_changed(rq, p, prev_class, oldprio, running);
4714 task_rq_unlock(rq, &flags);
4719 void set_user_nice(struct task_struct *p, long nice)
4721 int old_prio, delta, on_rq;
4722 unsigned long flags;
4725 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4728 * We have to be careful, if called from sys_setpriority(),
4729 * the task might be in the middle of scheduling on another CPU.
4731 rq = task_rq_lock(p, &flags);
4733 * The RT priorities are set via sched_setscheduler(), but we still
4734 * allow the 'normal' nice value to be set - but as expected
4735 * it wont have any effect on scheduling until the task is
4736 * SCHED_FIFO/SCHED_RR:
4738 if (task_has_rt_policy(p)) {
4739 p->static_prio = NICE_TO_PRIO(nice);
4742 on_rq = p->se.on_rq;
4744 dequeue_task(rq, p, 0);
4746 p->static_prio = NICE_TO_PRIO(nice);
4749 p->prio = effective_prio(p);
4750 delta = p->prio - old_prio;
4753 enqueue_task(rq, p, 0);
4755 * If the task increased its priority or is running and
4756 * lowered its priority, then reschedule its CPU:
4758 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4759 resched_task(rq->curr);
4762 task_rq_unlock(rq, &flags);
4764 EXPORT_SYMBOL(set_user_nice);
4767 * can_nice - check if a task can reduce its nice value
4771 int can_nice(const struct task_struct *p, const int nice)
4773 /* convert nice value [19,-20] to rlimit style value [1,40] */
4774 int nice_rlim = 20 - nice;
4776 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4777 capable(CAP_SYS_NICE));
4780 #ifdef __ARCH_WANT_SYS_NICE
4783 * sys_nice - change the priority of the current process.
4784 * @increment: priority increment
4786 * sys_setpriority is a more generic, but much slower function that
4787 * does similar things.
4789 SYSCALL_DEFINE1(nice, int, increment)
4794 * Setpriority might change our priority at the same moment.
4795 * We don't have to worry. Conceptually one call occurs first
4796 * and we have a single winner.
4798 if (increment < -40)
4803 nice = TASK_NICE(current) + increment;
4809 if (increment < 0 && !can_nice(current, nice))
4812 retval = security_task_setnice(current, nice);
4816 set_user_nice(current, nice);
4823 * task_prio - return the priority value of a given task.
4824 * @p: the task in question.
4826 * This is the priority value as seen by users in /proc.
4827 * RT tasks are offset by -200. Normal tasks are centered
4828 * around 0, value goes from -16 to +15.
4830 int task_prio(const struct task_struct *p)
4832 return p->prio - MAX_RT_PRIO;
4836 * task_nice - return the nice value of a given task.
4837 * @p: the task in question.
4839 int task_nice(const struct task_struct *p)
4841 return TASK_NICE(p);
4843 EXPORT_SYMBOL(task_nice);
4846 * idle_cpu - is a given cpu idle currently?
4847 * @cpu: the processor in question.
4849 int idle_cpu(int cpu)
4851 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4855 * idle_task - return the idle task for a given cpu.
4856 * @cpu: the processor in question.
4858 struct task_struct *idle_task(int cpu)
4860 return cpu_rq(cpu)->idle;
4864 * find_process_by_pid - find a process with a matching PID value.
4865 * @pid: the pid in question.
4867 static struct task_struct *find_process_by_pid(pid_t pid)
4869 return pid ? find_task_by_vpid(pid) : current;
4872 /* Actually do priority change: must hold rq lock. */
4874 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4876 BUG_ON(p->se.on_rq);
4879 p->rt_priority = prio;
4880 p->normal_prio = normal_prio(p);
4881 /* we are holding p->pi_lock already */
4882 p->prio = rt_mutex_getprio(p);
4883 if (rt_prio(p->prio))
4884 p->sched_class = &rt_sched_class;
4886 p->sched_class = &fair_sched_class;
4891 * check the target process has a UID that matches the current process's
4893 static bool check_same_owner(struct task_struct *p)
4895 const struct cred *cred = current_cred(), *pcred;
4899 pcred = __task_cred(p);
4900 match = (cred->euid == pcred->euid ||
4901 cred->euid == pcred->uid);
4906 static int __sched_setscheduler(struct task_struct *p, int policy,
4907 struct sched_param *param, bool user)
4909 int retval, oldprio, oldpolicy = -1, on_rq, running;
4910 unsigned long flags;
4911 const struct sched_class *prev_class;
4915 /* may grab non-irq protected spin_locks */
4916 BUG_ON(in_interrupt());
4918 /* double check policy once rq lock held */
4920 reset_on_fork = p->sched_reset_on_fork;
4921 policy = oldpolicy = p->policy;
4923 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4924 policy &= ~SCHED_RESET_ON_FORK;
4926 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4927 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4928 policy != SCHED_IDLE)
4933 * Valid priorities for SCHED_FIFO and SCHED_RR are
4934 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4935 * SCHED_BATCH and SCHED_IDLE is 0.
4937 if (param->sched_priority < 0 ||
4938 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4939 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4941 if (rt_policy(policy) != (param->sched_priority != 0))
4945 * Allow unprivileged RT tasks to decrease priority:
4947 if (user && !capable(CAP_SYS_NICE)) {
4948 if (rt_policy(policy)) {
4949 unsigned long rlim_rtprio =
4950 task_rlimit(p, RLIMIT_RTPRIO);
4952 /* can't set/change the rt policy */
4953 if (policy != p->policy && !rlim_rtprio)
4956 /* can't increase priority */
4957 if (param->sched_priority > p->rt_priority &&
4958 param->sched_priority > rlim_rtprio)
4962 * Like positive nice levels, dont allow tasks to
4963 * move out of SCHED_IDLE either:
4965 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4968 /* can't change other user's priorities */
4969 if (!check_same_owner(p))
4972 /* Normal users shall not reset the sched_reset_on_fork flag */
4973 if (p->sched_reset_on_fork && !reset_on_fork)
4978 retval = security_task_setscheduler(p);
4984 * make sure no PI-waiters arrive (or leave) while we are
4985 * changing the priority of the task:
4987 raw_spin_lock_irqsave(&p->pi_lock, flags);
4989 * To be able to change p->policy safely, the apropriate
4990 * runqueue lock must be held.
4992 rq = __task_rq_lock(p);
4995 * Changing the policy of the stop threads its a very bad idea
4997 if (p == rq->stop) {
4998 __task_rq_unlock(rq);
4999 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5003 #ifdef CONFIG_RT_GROUP_SCHED
5006 * Do not allow realtime tasks into groups that have no runtime
5009 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5010 task_group(p)->rt_bandwidth.rt_runtime == 0) {
5011 __task_rq_unlock(rq);
5012 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5018 /* recheck policy now with rq lock held */
5019 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5020 policy = oldpolicy = -1;
5021 __task_rq_unlock(rq);
5022 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5025 on_rq = p->se.on_rq;
5026 running = task_current(rq, p);
5028 deactivate_task(rq, p, 0);
5030 p->sched_class->put_prev_task(rq, p);
5032 p->sched_reset_on_fork = reset_on_fork;
5035 prev_class = p->sched_class;
5036 __setscheduler(rq, p, policy, param->sched_priority);
5039 p->sched_class->set_curr_task(rq);
5041 activate_task(rq, p, 0);
5043 check_class_changed(rq, p, prev_class, oldprio, running);
5045 __task_rq_unlock(rq);
5046 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5048 rt_mutex_adjust_pi(p);
5054 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5055 * @p: the task in question.
5056 * @policy: new policy.
5057 * @param: structure containing the new RT priority.
5059 * NOTE that the task may be already dead.
5061 int sched_setscheduler(struct task_struct *p, int policy,
5062 struct sched_param *param)
5064 return __sched_setscheduler(p, policy, param, true);
5066 EXPORT_SYMBOL_GPL(sched_setscheduler);
5069 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5070 * @p: the task in question.
5071 * @policy: new policy.
5072 * @param: structure containing the new RT priority.
5074 * Just like sched_setscheduler, only don't bother checking if the
5075 * current context has permission. For example, this is needed in
5076 * stop_machine(): we create temporary high priority worker threads,
5077 * but our caller might not have that capability.
5079 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5080 struct sched_param *param)
5082 return __sched_setscheduler(p, policy, param, false);
5086 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5088 struct sched_param lparam;
5089 struct task_struct *p;
5092 if (!param || pid < 0)
5094 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5099 p = find_process_by_pid(pid);
5101 retval = sched_setscheduler(p, policy, &lparam);
5108 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5109 * @pid: the pid in question.
5110 * @policy: new policy.
5111 * @param: structure containing the new RT priority.
5113 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5114 struct sched_param __user *, param)
5116 /* negative values for policy are not valid */
5120 return do_sched_setscheduler(pid, policy, param);
5124 * sys_sched_setparam - set/change the RT priority of a thread
5125 * @pid: the pid in question.
5126 * @param: structure containing the new RT priority.
5128 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5130 return do_sched_setscheduler(pid, -1, param);
5134 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5135 * @pid: the pid in question.
5137 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5139 struct task_struct *p;
5147 p = find_process_by_pid(pid);
5149 retval = security_task_getscheduler(p);
5152 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5159 * sys_sched_getparam - get the RT priority of a thread
5160 * @pid: the pid in question.
5161 * @param: structure containing the RT priority.
5163 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5165 struct sched_param lp;
5166 struct task_struct *p;
5169 if (!param || pid < 0)
5173 p = find_process_by_pid(pid);
5178 retval = security_task_getscheduler(p);
5182 lp.sched_priority = p->rt_priority;
5186 * This one might sleep, we cannot do it with a spinlock held ...
5188 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5197 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5199 cpumask_var_t cpus_allowed, new_mask;
5200 struct task_struct *p;
5206 p = find_process_by_pid(pid);
5213 /* Prevent p going away */
5217 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5221 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5223 goto out_free_cpus_allowed;
5226 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5229 retval = security_task_setscheduler(p);
5233 cpuset_cpus_allowed(p, cpus_allowed);
5234 cpumask_and(new_mask, in_mask, cpus_allowed);
5236 retval = set_cpus_allowed_ptr(p, new_mask);
5239 cpuset_cpus_allowed(p, cpus_allowed);
5240 if (!cpumask_subset(new_mask, cpus_allowed)) {
5242 * We must have raced with a concurrent cpuset
5243 * update. Just reset the cpus_allowed to the
5244 * cpuset's cpus_allowed
5246 cpumask_copy(new_mask, cpus_allowed);
5251 free_cpumask_var(new_mask);
5252 out_free_cpus_allowed:
5253 free_cpumask_var(cpus_allowed);
5260 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5261 struct cpumask *new_mask)
5263 if (len < cpumask_size())
5264 cpumask_clear(new_mask);
5265 else if (len > cpumask_size())
5266 len = cpumask_size();
5268 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5272 * sys_sched_setaffinity - set the cpu affinity of a process
5273 * @pid: pid of the process
5274 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5275 * @user_mask_ptr: user-space pointer to the new cpu mask
5277 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5278 unsigned long __user *, user_mask_ptr)
5280 cpumask_var_t new_mask;
5283 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5286 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5288 retval = sched_setaffinity(pid, new_mask);
5289 free_cpumask_var(new_mask);
5293 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5295 struct task_struct *p;
5296 unsigned long flags;
5304 p = find_process_by_pid(pid);
5308 retval = security_task_getscheduler(p);
5312 rq = task_rq_lock(p, &flags);
5313 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5314 task_rq_unlock(rq, &flags);
5324 * sys_sched_getaffinity - get the cpu affinity of a process
5325 * @pid: pid of the process
5326 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5327 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5329 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5330 unsigned long __user *, user_mask_ptr)
5335 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5337 if (len & (sizeof(unsigned long)-1))
5340 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5343 ret = sched_getaffinity(pid, mask);
5345 size_t retlen = min_t(size_t, len, cpumask_size());
5347 if (copy_to_user(user_mask_ptr, mask, retlen))
5352 free_cpumask_var(mask);
5358 * sys_sched_yield - yield the current processor to other threads.
5360 * This function yields the current CPU to other tasks. If there are no
5361 * other threads running on this CPU then this function will return.
5363 SYSCALL_DEFINE0(sched_yield)
5365 struct rq *rq = this_rq_lock();
5367 schedstat_inc(rq, yld_count);
5368 current->sched_class->yield_task(rq);
5371 * Since we are going to call schedule() anyway, there's
5372 * no need to preempt or enable interrupts:
5374 __release(rq->lock);
5375 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5376 do_raw_spin_unlock(&rq->lock);
5377 preempt_enable_no_resched();
5384 static inline int should_resched(void)
5386 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5389 static void __cond_resched(void)
5391 add_preempt_count(PREEMPT_ACTIVE);
5393 sub_preempt_count(PREEMPT_ACTIVE);
5396 int __sched _cond_resched(void)
5398 if (should_resched()) {
5404 EXPORT_SYMBOL(_cond_resched);
5407 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5408 * call schedule, and on return reacquire the lock.
5410 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5411 * operations here to prevent schedule() from being called twice (once via
5412 * spin_unlock(), once by hand).
5414 int __cond_resched_lock(spinlock_t *lock)
5416 int resched = should_resched();
5419 lockdep_assert_held(lock);
5421 if (spin_needbreak(lock) || resched) {
5432 EXPORT_SYMBOL(__cond_resched_lock);
5434 int __sched __cond_resched_softirq(void)
5436 BUG_ON(!in_softirq());
5438 if (should_resched()) {
5446 EXPORT_SYMBOL(__cond_resched_softirq);
5449 * yield - yield the current processor to other threads.
5451 * This is a shortcut for kernel-space yielding - it marks the
5452 * thread runnable and calls sys_sched_yield().
5454 void __sched yield(void)
5456 set_current_state(TASK_RUNNING);
5459 EXPORT_SYMBOL(yield);
5462 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5463 * that process accounting knows that this is a task in IO wait state.
5465 void __sched io_schedule(void)
5467 struct rq *rq = raw_rq();
5469 delayacct_blkio_start();
5470 atomic_inc(&rq->nr_iowait);
5471 current->in_iowait = 1;
5473 current->in_iowait = 0;
5474 atomic_dec(&rq->nr_iowait);
5475 delayacct_blkio_end();
5477 EXPORT_SYMBOL(io_schedule);
5479 long __sched io_schedule_timeout(long timeout)
5481 struct rq *rq = raw_rq();
5484 delayacct_blkio_start();
5485 atomic_inc(&rq->nr_iowait);
5486 current->in_iowait = 1;
5487 ret = schedule_timeout(timeout);
5488 current->in_iowait = 0;
5489 atomic_dec(&rq->nr_iowait);
5490 delayacct_blkio_end();
5495 * sys_sched_get_priority_max - return maximum RT priority.
5496 * @policy: scheduling class.
5498 * this syscall returns the maximum rt_priority that can be used
5499 * by a given scheduling class.
5501 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5508 ret = MAX_USER_RT_PRIO-1;
5520 * sys_sched_get_priority_min - return minimum RT priority.
5521 * @policy: scheduling class.
5523 * this syscall returns the minimum rt_priority that can be used
5524 * by a given scheduling class.
5526 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5544 * sys_sched_rr_get_interval - return the default timeslice of a process.
5545 * @pid: pid of the process.
5546 * @interval: userspace pointer to the timeslice value.
5548 * this syscall writes the default timeslice value of a given process
5549 * into the user-space timespec buffer. A value of '0' means infinity.
5551 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5552 struct timespec __user *, interval)
5554 struct task_struct *p;
5555 unsigned int time_slice;
5556 unsigned long flags;
5566 p = find_process_by_pid(pid);
5570 retval = security_task_getscheduler(p);
5574 rq = task_rq_lock(p, &flags);
5575 time_slice = p->sched_class->get_rr_interval(rq, p);
5576 task_rq_unlock(rq, &flags);
5579 jiffies_to_timespec(time_slice, &t);
5580 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5588 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5590 void sched_show_task(struct task_struct *p)
5592 unsigned long free = 0;
5595 state = p->state ? __ffs(p->state) + 1 : 0;
5596 printk(KERN_INFO "%-13.13s %c", p->comm,
5597 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5598 #if BITS_PER_LONG == 32
5599 if (state == TASK_RUNNING)
5600 printk(KERN_CONT " running ");
5602 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5604 if (state == TASK_RUNNING)
5605 printk(KERN_CONT " running task ");
5607 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5609 #ifdef CONFIG_DEBUG_STACK_USAGE
5610 free = stack_not_used(p);
5612 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5613 task_pid_nr(p), task_pid_nr(p->real_parent),
5614 (unsigned long)task_thread_info(p)->flags);
5616 show_stack(p, NULL);
5619 void show_state_filter(unsigned long state_filter)
5621 struct task_struct *g, *p;
5623 #if BITS_PER_LONG == 32
5625 " task PC stack pid father\n");
5628 " task PC stack pid father\n");
5630 read_lock(&tasklist_lock);
5631 do_each_thread(g, p) {
5633 * reset the NMI-timeout, listing all files on a slow
5634 * console might take alot of time:
5636 touch_nmi_watchdog();
5637 if (!state_filter || (p->state & state_filter))
5639 } while_each_thread(g, p);
5641 touch_all_softlockup_watchdogs();
5643 #ifdef CONFIG_SCHED_DEBUG
5644 sysrq_sched_debug_show();
5646 read_unlock(&tasklist_lock);
5648 * Only show locks if all tasks are dumped:
5651 debug_show_all_locks();
5654 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5656 idle->sched_class = &idle_sched_class;
5660 * init_idle - set up an idle thread for a given CPU
5661 * @idle: task in question
5662 * @cpu: cpu the idle task belongs to
5664 * NOTE: this function does not set the idle thread's NEED_RESCHED
5665 * flag, to make booting more robust.
5667 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5669 struct rq *rq = cpu_rq(cpu);
5670 unsigned long flags;
5672 raw_spin_lock_irqsave(&rq->lock, flags);
5675 idle->state = TASK_RUNNING;
5676 idle->se.exec_start = sched_clock();
5678 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5680 * We're having a chicken and egg problem, even though we are
5681 * holding rq->lock, the cpu isn't yet set to this cpu so the
5682 * lockdep check in task_group() will fail.
5684 * Similar case to sched_fork(). / Alternatively we could
5685 * use task_rq_lock() here and obtain the other rq->lock.
5690 __set_task_cpu(idle, cpu);
5693 rq->curr = rq->idle = idle;
5694 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5697 raw_spin_unlock_irqrestore(&rq->lock, flags);
5699 /* Set the preempt count _outside_ the spinlocks! */
5700 #if defined(CONFIG_PREEMPT)
5701 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5703 task_thread_info(idle)->preempt_count = 0;
5706 * The idle tasks have their own, simple scheduling class:
5708 idle->sched_class = &idle_sched_class;
5709 ftrace_graph_init_idle_task(idle, cpu);
5713 * In a system that switches off the HZ timer nohz_cpu_mask
5714 * indicates which cpus entered this state. This is used
5715 * in the rcu update to wait only for active cpus. For system
5716 * which do not switch off the HZ timer nohz_cpu_mask should
5717 * always be CPU_BITS_NONE.
5719 cpumask_var_t nohz_cpu_mask;
5722 * Increase the granularity value when there are more CPUs,
5723 * because with more CPUs the 'effective latency' as visible
5724 * to users decreases. But the relationship is not linear,
5725 * so pick a second-best guess by going with the log2 of the
5728 * This idea comes from the SD scheduler of Con Kolivas:
5730 static int get_update_sysctl_factor(void)
5732 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5733 unsigned int factor;
5735 switch (sysctl_sched_tunable_scaling) {
5736 case SCHED_TUNABLESCALING_NONE:
5739 case SCHED_TUNABLESCALING_LINEAR:
5742 case SCHED_TUNABLESCALING_LOG:
5744 factor = 1 + ilog2(cpus);
5751 static void update_sysctl(void)
5753 unsigned int factor = get_update_sysctl_factor();
5755 #define SET_SYSCTL(name) \
5756 (sysctl_##name = (factor) * normalized_sysctl_##name)
5757 SET_SYSCTL(sched_min_granularity);
5758 SET_SYSCTL(sched_latency);
5759 SET_SYSCTL(sched_wakeup_granularity);
5760 SET_SYSCTL(sched_shares_ratelimit);
5764 static inline void sched_init_granularity(void)
5771 * This is how migration works:
5773 * 1) we invoke migration_cpu_stop() on the target CPU using
5775 * 2) stopper starts to run (implicitly forcing the migrated thread
5777 * 3) it checks whether the migrated task is still in the wrong runqueue.
5778 * 4) if it's in the wrong runqueue then the migration thread removes
5779 * it and puts it into the right queue.
5780 * 5) stopper completes and stop_one_cpu() returns and the migration
5785 * Change a given task's CPU affinity. Migrate the thread to a
5786 * proper CPU and schedule it away if the CPU it's executing on
5787 * is removed from the allowed bitmask.
5789 * NOTE: the caller must have a valid reference to the task, the
5790 * task must not exit() & deallocate itself prematurely. The
5791 * call is not atomic; no spinlocks may be held.
5793 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5795 unsigned long flags;
5797 unsigned int dest_cpu;
5801 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5802 * drop the rq->lock and still rely on ->cpus_allowed.
5805 while (task_is_waking(p))
5807 rq = task_rq_lock(p, &flags);
5808 if (task_is_waking(p)) {
5809 task_rq_unlock(rq, &flags);
5813 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5818 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5819 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5824 if (p->sched_class->set_cpus_allowed)
5825 p->sched_class->set_cpus_allowed(p, new_mask);
5827 cpumask_copy(&p->cpus_allowed, new_mask);
5828 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5831 /* Can the task run on the task's current CPU? If so, we're done */
5832 if (cpumask_test_cpu(task_cpu(p), new_mask))
5835 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5836 if (migrate_task(p, dest_cpu)) {
5837 struct migration_arg arg = { p, dest_cpu };
5838 /* Need help from migration thread: drop lock and wait. */
5839 task_rq_unlock(rq, &flags);
5840 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5841 tlb_migrate_finish(p->mm);
5845 task_rq_unlock(rq, &flags);
5849 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5852 * Move (not current) task off this cpu, onto dest cpu. We're doing
5853 * this because either it can't run here any more (set_cpus_allowed()
5854 * away from this CPU, or CPU going down), or because we're
5855 * attempting to rebalance this task on exec (sched_exec).
5857 * So we race with normal scheduler movements, but that's OK, as long
5858 * as the task is no longer on this CPU.
5860 * Returns non-zero if task was successfully migrated.
5862 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5864 struct rq *rq_dest, *rq_src;
5867 if (unlikely(!cpu_active(dest_cpu)))
5870 rq_src = cpu_rq(src_cpu);
5871 rq_dest = cpu_rq(dest_cpu);
5873 double_rq_lock(rq_src, rq_dest);
5874 /* Already moved. */
5875 if (task_cpu(p) != src_cpu)
5877 /* Affinity changed (again). */
5878 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5882 * If we're not on a rq, the next wake-up will ensure we're
5886 deactivate_task(rq_src, p, 0);
5887 set_task_cpu(p, dest_cpu);
5888 activate_task(rq_dest, p, 0);
5889 check_preempt_curr(rq_dest, p, 0);
5894 double_rq_unlock(rq_src, rq_dest);
5899 * migration_cpu_stop - this will be executed by a highprio stopper thread
5900 * and performs thread migration by bumping thread off CPU then
5901 * 'pushing' onto another runqueue.
5903 static int migration_cpu_stop(void *data)
5905 struct migration_arg *arg = data;
5908 * The original target cpu might have gone down and we might
5909 * be on another cpu but it doesn't matter.
5911 local_irq_disable();
5912 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5917 #ifdef CONFIG_HOTPLUG_CPU
5919 * Figure out where task on dead CPU should go, use force if necessary.
5921 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5923 struct rq *rq = cpu_rq(dead_cpu);
5924 int needs_cpu, uninitialized_var(dest_cpu);
5925 unsigned long flags;
5927 local_irq_save(flags);
5929 raw_spin_lock(&rq->lock);
5930 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5932 dest_cpu = select_fallback_rq(dead_cpu, p);
5933 raw_spin_unlock(&rq->lock);
5935 * It can only fail if we race with set_cpus_allowed(),
5936 * in the racer should migrate the task anyway.
5939 __migrate_task(p, dead_cpu, dest_cpu);
5940 local_irq_restore(flags);
5944 * While a dead CPU has no uninterruptible tasks queued at this point,
5945 * it might still have a nonzero ->nr_uninterruptible counter, because
5946 * for performance reasons the counter is not stricly tracking tasks to
5947 * their home CPUs. So we just add the counter to another CPU's counter,
5948 * to keep the global sum constant after CPU-down:
5950 static void migrate_nr_uninterruptible(struct rq *rq_src)
5952 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5953 unsigned long flags;
5955 local_irq_save(flags);
5956 double_rq_lock(rq_src, rq_dest);
5957 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5958 rq_src->nr_uninterruptible = 0;
5959 double_rq_unlock(rq_src, rq_dest);
5960 local_irq_restore(flags);
5963 /* Run through task list and migrate tasks from the dead cpu. */
5964 static void migrate_live_tasks(int src_cpu)
5966 struct task_struct *p, *t;
5968 read_lock(&tasklist_lock);
5970 do_each_thread(t, p) {
5974 if (task_cpu(p) == src_cpu)
5975 move_task_off_dead_cpu(src_cpu, p);
5976 } while_each_thread(t, p);
5978 read_unlock(&tasklist_lock);
5982 * Schedules idle task to be the next runnable task on current CPU.
5983 * It does so by boosting its priority to highest possible.
5984 * Used by CPU offline code.
5986 void sched_idle_next(void)
5988 int this_cpu = smp_processor_id();
5989 struct rq *rq = cpu_rq(this_cpu);
5990 struct task_struct *p = rq->idle;
5991 unsigned long flags;
5993 /* cpu has to be offline */
5994 BUG_ON(cpu_online(this_cpu));
5997 * Strictly not necessary since rest of the CPUs are stopped by now
5998 * and interrupts disabled on the current cpu.
6000 raw_spin_lock_irqsave(&rq->lock, flags);
6002 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6004 activate_task(rq, p, 0);
6006 raw_spin_unlock_irqrestore(&rq->lock, flags);
6010 * Ensures that the idle task is using init_mm right before its cpu goes
6013 void idle_task_exit(void)
6015 struct mm_struct *mm = current->active_mm;
6017 BUG_ON(cpu_online(smp_processor_id()));
6020 switch_mm(mm, &init_mm, current);
6024 /* called under rq->lock with disabled interrupts */
6025 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6027 struct rq *rq = cpu_rq(dead_cpu);
6029 /* Must be exiting, otherwise would be on tasklist. */
6030 BUG_ON(!p->exit_state);
6032 /* Cannot have done final schedule yet: would have vanished. */
6033 BUG_ON(p->state == TASK_DEAD);
6038 * Drop lock around migration; if someone else moves it,
6039 * that's OK. No task can be added to this CPU, so iteration is
6042 raw_spin_unlock_irq(&rq->lock);
6043 move_task_off_dead_cpu(dead_cpu, p);
6044 raw_spin_lock_irq(&rq->lock);
6049 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6050 static void migrate_dead_tasks(unsigned int dead_cpu)
6052 struct rq *rq = cpu_rq(dead_cpu);
6053 struct task_struct *next;
6056 if (!rq->nr_running)
6058 next = pick_next_task(rq);
6061 next->sched_class->put_prev_task(rq, next);
6062 migrate_dead(dead_cpu, next);
6068 * remove the tasks which were accounted by rq from calc_load_tasks.
6070 static void calc_global_load_remove(struct rq *rq)
6072 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6073 rq->calc_load_active = 0;
6075 #endif /* CONFIG_HOTPLUG_CPU */
6077 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6079 static struct ctl_table sd_ctl_dir[] = {
6081 .procname = "sched_domain",
6087 static struct ctl_table sd_ctl_root[] = {
6089 .procname = "kernel",
6091 .child = sd_ctl_dir,
6096 static struct ctl_table *sd_alloc_ctl_entry(int n)
6098 struct ctl_table *entry =
6099 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6104 static void sd_free_ctl_entry(struct ctl_table **tablep)
6106 struct ctl_table *entry;
6109 * In the intermediate directories, both the child directory and
6110 * procname are dynamically allocated and could fail but the mode
6111 * will always be set. In the lowest directory the names are
6112 * static strings and all have proc handlers.
6114 for (entry = *tablep; entry->mode; entry++) {
6116 sd_free_ctl_entry(&entry->child);
6117 if (entry->proc_handler == NULL)
6118 kfree(entry->procname);
6126 set_table_entry(struct ctl_table *entry,
6127 const char *procname, void *data, int maxlen,
6128 mode_t mode, proc_handler *proc_handler)
6130 entry->procname = procname;
6132 entry->maxlen = maxlen;
6134 entry->proc_handler = proc_handler;
6137 static struct ctl_table *
6138 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6140 struct ctl_table *table = sd_alloc_ctl_entry(13);
6145 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6146 sizeof(long), 0644, proc_doulongvec_minmax);
6147 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6148 sizeof(long), 0644, proc_doulongvec_minmax);
6149 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6150 sizeof(int), 0644, proc_dointvec_minmax);
6151 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6152 sizeof(int), 0644, proc_dointvec_minmax);
6153 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6154 sizeof(int), 0644, proc_dointvec_minmax);
6155 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6156 sizeof(int), 0644, proc_dointvec_minmax);
6157 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6158 sizeof(int), 0644, proc_dointvec_minmax);
6159 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6160 sizeof(int), 0644, proc_dointvec_minmax);
6161 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6162 sizeof(int), 0644, proc_dointvec_minmax);
6163 set_table_entry(&table[9], "cache_nice_tries",
6164 &sd->cache_nice_tries,
6165 sizeof(int), 0644, proc_dointvec_minmax);
6166 set_table_entry(&table[10], "flags", &sd->flags,
6167 sizeof(int), 0644, proc_dointvec_minmax);
6168 set_table_entry(&table[11], "name", sd->name,
6169 CORENAME_MAX_SIZE, 0444, proc_dostring);
6170 /* &table[12] is terminator */
6175 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6177 struct ctl_table *entry, *table;
6178 struct sched_domain *sd;
6179 int domain_num = 0, i;
6182 for_each_domain(cpu, sd)
6184 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6189 for_each_domain(cpu, sd) {
6190 snprintf(buf, 32, "domain%d", i);
6191 entry->procname = kstrdup(buf, GFP_KERNEL);
6193 entry->child = sd_alloc_ctl_domain_table(sd);
6200 static struct ctl_table_header *sd_sysctl_header;
6201 static void register_sched_domain_sysctl(void)
6203 int i, cpu_num = num_possible_cpus();
6204 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6207 WARN_ON(sd_ctl_dir[0].child);
6208 sd_ctl_dir[0].child = entry;
6213 for_each_possible_cpu(i) {
6214 snprintf(buf, 32, "cpu%d", i);
6215 entry->procname = kstrdup(buf, GFP_KERNEL);
6217 entry->child = sd_alloc_ctl_cpu_table(i);
6221 WARN_ON(sd_sysctl_header);
6222 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6225 /* may be called multiple times per register */
6226 static void unregister_sched_domain_sysctl(void)
6228 if (sd_sysctl_header)
6229 unregister_sysctl_table(sd_sysctl_header);
6230 sd_sysctl_header = NULL;
6231 if (sd_ctl_dir[0].child)
6232 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6235 static void register_sched_domain_sysctl(void)
6238 static void unregister_sched_domain_sysctl(void)
6243 static void set_rq_online(struct rq *rq)
6246 const struct sched_class *class;
6248 cpumask_set_cpu(rq->cpu, rq->rd->online);
6251 for_each_class(class) {
6252 if (class->rq_online)
6253 class->rq_online(rq);
6258 static void set_rq_offline(struct rq *rq)
6261 const struct sched_class *class;
6263 for_each_class(class) {
6264 if (class->rq_offline)
6265 class->rq_offline(rq);
6268 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6274 * migration_call - callback that gets triggered when a CPU is added.
6275 * Here we can start up the necessary migration thread for the new CPU.
6277 static int __cpuinit
6278 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6280 int cpu = (long)hcpu;
6281 unsigned long flags;
6282 struct rq *rq = cpu_rq(cpu);
6286 case CPU_UP_PREPARE:
6287 case CPU_UP_PREPARE_FROZEN:
6288 rq->calc_load_update = calc_load_update;
6292 case CPU_ONLINE_FROZEN:
6293 /* Update our root-domain */
6294 raw_spin_lock_irqsave(&rq->lock, flags);
6296 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6300 raw_spin_unlock_irqrestore(&rq->lock, flags);
6303 #ifdef CONFIG_HOTPLUG_CPU
6305 case CPU_DEAD_FROZEN:
6306 migrate_live_tasks(cpu);
6307 /* Idle task back to normal (off runqueue, low prio) */
6308 raw_spin_lock_irq(&rq->lock);
6309 deactivate_task(rq, rq->idle, 0);
6310 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6311 rq->idle->sched_class = &idle_sched_class;
6312 migrate_dead_tasks(cpu);
6313 raw_spin_unlock_irq(&rq->lock);
6314 migrate_nr_uninterruptible(rq);
6315 BUG_ON(rq->nr_running != 0);
6316 calc_global_load_remove(rq);
6320 case CPU_DYING_FROZEN:
6321 /* Update our root-domain */
6322 raw_spin_lock_irqsave(&rq->lock, flags);
6324 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6327 raw_spin_unlock_irqrestore(&rq->lock, flags);
6335 * Register at high priority so that task migration (migrate_all_tasks)
6336 * happens before everything else. This has to be lower priority than
6337 * the notifier in the perf_event subsystem, though.
6339 static struct notifier_block __cpuinitdata migration_notifier = {
6340 .notifier_call = migration_call,
6341 .priority = CPU_PRI_MIGRATION,
6344 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6345 unsigned long action, void *hcpu)
6347 switch (action & ~CPU_TASKS_FROZEN) {
6349 case CPU_DOWN_FAILED:
6350 set_cpu_active((long)hcpu, true);
6357 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6358 unsigned long action, void *hcpu)
6360 switch (action & ~CPU_TASKS_FROZEN) {
6361 case CPU_DOWN_PREPARE:
6362 set_cpu_active((long)hcpu, false);
6369 static int __init migration_init(void)
6371 void *cpu = (void *)(long)smp_processor_id();
6374 /* Initialize migration for the boot CPU */
6375 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6376 BUG_ON(err == NOTIFY_BAD);
6377 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6378 register_cpu_notifier(&migration_notifier);
6380 /* Register cpu active notifiers */
6381 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6382 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6386 early_initcall(migration_init);
6391 #ifdef CONFIG_SCHED_DEBUG
6393 static __read_mostly int sched_domain_debug_enabled;
6395 static int __init sched_domain_debug_setup(char *str)
6397 sched_domain_debug_enabled = 1;
6401 early_param("sched_debug", sched_domain_debug_setup);
6403 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6404 struct cpumask *groupmask)
6406 struct sched_group *group = sd->groups;
6409 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6410 cpumask_clear(groupmask);
6412 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6414 if (!(sd->flags & SD_LOAD_BALANCE)) {
6415 printk("does not load-balance\n");
6417 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6422 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6424 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6425 printk(KERN_ERR "ERROR: domain->span does not contain "
6428 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6429 printk(KERN_ERR "ERROR: domain->groups does not contain"
6433 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6437 printk(KERN_ERR "ERROR: group is NULL\n");
6441 if (!group->cpu_power) {
6442 printk(KERN_CONT "\n");
6443 printk(KERN_ERR "ERROR: domain->cpu_power not "
6448 if (!cpumask_weight(sched_group_cpus(group))) {
6449 printk(KERN_CONT "\n");
6450 printk(KERN_ERR "ERROR: empty group\n");
6454 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6455 printk(KERN_CONT "\n");
6456 printk(KERN_ERR "ERROR: repeated CPUs\n");
6460 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6462 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6464 printk(KERN_CONT " %s", str);
6465 if (group->cpu_power != SCHED_LOAD_SCALE) {
6466 printk(KERN_CONT " (cpu_power = %d)",
6470 group = group->next;
6471 } while (group != sd->groups);
6472 printk(KERN_CONT "\n");
6474 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6475 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6478 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6479 printk(KERN_ERR "ERROR: parent span is not a superset "
6480 "of domain->span\n");
6484 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6486 cpumask_var_t groupmask;
6489 if (!sched_domain_debug_enabled)
6493 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6497 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6499 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6500 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6505 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6512 free_cpumask_var(groupmask);
6514 #else /* !CONFIG_SCHED_DEBUG */
6515 # define sched_domain_debug(sd, cpu) do { } while (0)
6516 #endif /* CONFIG_SCHED_DEBUG */
6518 static int sd_degenerate(struct sched_domain *sd)
6520 if (cpumask_weight(sched_domain_span(sd)) == 1)
6523 /* Following flags need at least 2 groups */
6524 if (sd->flags & (SD_LOAD_BALANCE |
6525 SD_BALANCE_NEWIDLE |
6529 SD_SHARE_PKG_RESOURCES)) {
6530 if (sd->groups != sd->groups->next)
6534 /* Following flags don't use groups */
6535 if (sd->flags & (SD_WAKE_AFFINE))
6542 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6544 unsigned long cflags = sd->flags, pflags = parent->flags;
6546 if (sd_degenerate(parent))
6549 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6552 /* Flags needing groups don't count if only 1 group in parent */
6553 if (parent->groups == parent->groups->next) {
6554 pflags &= ~(SD_LOAD_BALANCE |
6555 SD_BALANCE_NEWIDLE |
6559 SD_SHARE_PKG_RESOURCES);
6560 if (nr_node_ids == 1)
6561 pflags &= ~SD_SERIALIZE;
6563 if (~cflags & pflags)
6569 static void free_rootdomain(struct root_domain *rd)
6571 synchronize_sched();
6573 cpupri_cleanup(&rd->cpupri);
6575 free_cpumask_var(rd->rto_mask);
6576 free_cpumask_var(rd->online);
6577 free_cpumask_var(rd->span);
6581 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6583 struct root_domain *old_rd = NULL;
6584 unsigned long flags;
6586 raw_spin_lock_irqsave(&rq->lock, flags);
6591 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6594 cpumask_clear_cpu(rq->cpu, old_rd->span);
6597 * If we dont want to free the old_rt yet then
6598 * set old_rd to NULL to skip the freeing later
6601 if (!atomic_dec_and_test(&old_rd->refcount))
6605 atomic_inc(&rd->refcount);
6608 cpumask_set_cpu(rq->cpu, rd->span);
6609 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6612 raw_spin_unlock_irqrestore(&rq->lock, flags);
6615 free_rootdomain(old_rd);
6618 static int init_rootdomain(struct root_domain *rd)
6620 memset(rd, 0, sizeof(*rd));
6622 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6624 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6626 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6629 if (cpupri_init(&rd->cpupri) != 0)
6634 free_cpumask_var(rd->rto_mask);
6636 free_cpumask_var(rd->online);
6638 free_cpumask_var(rd->span);
6643 static void init_defrootdomain(void)
6645 init_rootdomain(&def_root_domain);
6647 atomic_set(&def_root_domain.refcount, 1);
6650 static struct root_domain *alloc_rootdomain(void)
6652 struct root_domain *rd;
6654 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6658 if (init_rootdomain(rd) != 0) {
6667 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6668 * hold the hotplug lock.
6671 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6673 struct rq *rq = cpu_rq(cpu);
6674 struct sched_domain *tmp;
6676 for (tmp = sd; tmp; tmp = tmp->parent)
6677 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6679 /* Remove the sched domains which do not contribute to scheduling. */
6680 for (tmp = sd; tmp; ) {
6681 struct sched_domain *parent = tmp->parent;
6685 if (sd_parent_degenerate(tmp, parent)) {
6686 tmp->parent = parent->parent;
6688 parent->parent->child = tmp;
6693 if (sd && sd_degenerate(sd)) {
6699 sched_domain_debug(sd, cpu);
6701 rq_attach_root(rq, rd);
6702 rcu_assign_pointer(rq->sd, sd);
6705 /* cpus with isolated domains */
6706 static cpumask_var_t cpu_isolated_map;
6708 /* Setup the mask of cpus configured for isolated domains */
6709 static int __init isolated_cpu_setup(char *str)
6711 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6712 cpulist_parse(str, cpu_isolated_map);
6716 __setup("isolcpus=", isolated_cpu_setup);
6719 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6720 * to a function which identifies what group(along with sched group) a CPU
6721 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6722 * (due to the fact that we keep track of groups covered with a struct cpumask).
6724 * init_sched_build_groups will build a circular linked list of the groups
6725 * covered by the given span, and will set each group's ->cpumask correctly,
6726 * and ->cpu_power to 0.
6729 init_sched_build_groups(const struct cpumask *span,
6730 const struct cpumask *cpu_map,
6731 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6732 struct sched_group **sg,
6733 struct cpumask *tmpmask),
6734 struct cpumask *covered, struct cpumask *tmpmask)
6736 struct sched_group *first = NULL, *last = NULL;
6739 cpumask_clear(covered);
6741 for_each_cpu(i, span) {
6742 struct sched_group *sg;
6743 int group = group_fn(i, cpu_map, &sg, tmpmask);
6746 if (cpumask_test_cpu(i, covered))
6749 cpumask_clear(sched_group_cpus(sg));
6752 for_each_cpu(j, span) {
6753 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6756 cpumask_set_cpu(j, covered);
6757 cpumask_set_cpu(j, sched_group_cpus(sg));
6768 #define SD_NODES_PER_DOMAIN 16
6773 * find_next_best_node - find the next node to include in a sched_domain
6774 * @node: node whose sched_domain we're building
6775 * @used_nodes: nodes already in the sched_domain
6777 * Find the next node to include in a given scheduling domain. Simply
6778 * finds the closest node not already in the @used_nodes map.
6780 * Should use nodemask_t.
6782 static int find_next_best_node(int node, nodemask_t *used_nodes)
6784 int i, n, val, min_val, best_node = 0;
6788 for (i = 0; i < nr_node_ids; i++) {
6789 /* Start at @node */
6790 n = (node + i) % nr_node_ids;
6792 if (!nr_cpus_node(n))
6795 /* Skip already used nodes */
6796 if (node_isset(n, *used_nodes))
6799 /* Simple min distance search */
6800 val = node_distance(node, n);
6802 if (val < min_val) {
6808 node_set(best_node, *used_nodes);
6813 * sched_domain_node_span - get a cpumask for a node's sched_domain
6814 * @node: node whose cpumask we're constructing
6815 * @span: resulting cpumask
6817 * Given a node, construct a good cpumask for its sched_domain to span. It
6818 * should be one that prevents unnecessary balancing, but also spreads tasks
6821 static void sched_domain_node_span(int node, struct cpumask *span)
6823 nodemask_t used_nodes;
6826 cpumask_clear(span);
6827 nodes_clear(used_nodes);
6829 cpumask_or(span, span, cpumask_of_node(node));
6830 node_set(node, used_nodes);
6832 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6833 int next_node = find_next_best_node(node, &used_nodes);
6835 cpumask_or(span, span, cpumask_of_node(next_node));
6838 #endif /* CONFIG_NUMA */
6840 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6843 * The cpus mask in sched_group and sched_domain hangs off the end.
6845 * ( See the the comments in include/linux/sched.h:struct sched_group
6846 * and struct sched_domain. )
6848 struct static_sched_group {
6849 struct sched_group sg;
6850 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6853 struct static_sched_domain {
6854 struct sched_domain sd;
6855 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6861 cpumask_var_t domainspan;
6862 cpumask_var_t covered;
6863 cpumask_var_t notcovered;
6865 cpumask_var_t nodemask;
6866 cpumask_var_t this_sibling_map;
6867 cpumask_var_t this_core_map;
6868 cpumask_var_t this_book_map;
6869 cpumask_var_t send_covered;
6870 cpumask_var_t tmpmask;
6871 struct sched_group **sched_group_nodes;
6872 struct root_domain *rd;
6876 sa_sched_groups = 0,
6882 sa_this_sibling_map,
6884 sa_sched_group_nodes,
6894 * SMT sched-domains:
6896 #ifdef CONFIG_SCHED_SMT
6897 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6898 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6901 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6902 struct sched_group **sg, struct cpumask *unused)
6905 *sg = &per_cpu(sched_groups, cpu).sg;
6908 #endif /* CONFIG_SCHED_SMT */
6911 * multi-core sched-domains:
6913 #ifdef CONFIG_SCHED_MC
6914 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6915 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6918 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6919 struct sched_group **sg, struct cpumask *mask)
6922 #ifdef CONFIG_SCHED_SMT
6923 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6924 group = cpumask_first(mask);
6929 *sg = &per_cpu(sched_group_core, group).sg;
6932 #endif /* CONFIG_SCHED_MC */
6935 * book sched-domains:
6937 #ifdef CONFIG_SCHED_BOOK
6938 static DEFINE_PER_CPU(struct static_sched_domain, book_domains);
6939 static DEFINE_PER_CPU(struct static_sched_group, sched_group_book);
6942 cpu_to_book_group(int cpu, const struct cpumask *cpu_map,
6943 struct sched_group **sg, struct cpumask *mask)
6946 #ifdef CONFIG_SCHED_MC
6947 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6948 group = cpumask_first(mask);
6949 #elif defined(CONFIG_SCHED_SMT)
6950 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6951 group = cpumask_first(mask);
6954 *sg = &per_cpu(sched_group_book, group).sg;
6957 #endif /* CONFIG_SCHED_BOOK */
6959 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6960 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6963 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6964 struct sched_group **sg, struct cpumask *mask)
6967 #ifdef CONFIG_SCHED_BOOK
6968 cpumask_and(mask, cpu_book_mask(cpu), cpu_map);
6969 group = cpumask_first(mask);
6970 #elif defined(CONFIG_SCHED_MC)
6971 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6972 group = cpumask_first(mask);
6973 #elif defined(CONFIG_SCHED_SMT)
6974 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6975 group = cpumask_first(mask);
6980 *sg = &per_cpu(sched_group_phys, group).sg;
6986 * The init_sched_build_groups can't handle what we want to do with node
6987 * groups, so roll our own. Now each node has its own list of groups which
6988 * gets dynamically allocated.
6990 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6991 static struct sched_group ***sched_group_nodes_bycpu;
6993 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6994 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6996 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6997 struct sched_group **sg,
6998 struct cpumask *nodemask)
7002 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
7003 group = cpumask_first(nodemask);
7006 *sg = &per_cpu(sched_group_allnodes, group).sg;
7010 static void init_numa_sched_groups_power(struct sched_group *group_head)
7012 struct sched_group *sg = group_head;
7018 for_each_cpu(j, sched_group_cpus(sg)) {
7019 struct sched_domain *sd;
7021 sd = &per_cpu(phys_domains, j).sd;
7022 if (j != group_first_cpu(sd->groups)) {
7024 * Only add "power" once for each
7030 sg->cpu_power += sd->groups->cpu_power;
7033 } while (sg != group_head);
7036 static int build_numa_sched_groups(struct s_data *d,
7037 const struct cpumask *cpu_map, int num)
7039 struct sched_domain *sd;
7040 struct sched_group *sg, *prev;
7043 cpumask_clear(d->covered);
7044 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
7045 if (cpumask_empty(d->nodemask)) {
7046 d->sched_group_nodes[num] = NULL;
7050 sched_domain_node_span(num, d->domainspan);
7051 cpumask_and(d->domainspan, d->domainspan, cpu_map);
7053 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7056 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
7060 d->sched_group_nodes[num] = sg;
7062 for_each_cpu(j, d->nodemask) {
7063 sd = &per_cpu(node_domains, j).sd;
7068 cpumask_copy(sched_group_cpus(sg), d->nodemask);
7070 cpumask_or(d->covered, d->covered, d->nodemask);
7073 for (j = 0; j < nr_node_ids; j++) {
7074 n = (num + j) % nr_node_ids;
7075 cpumask_complement(d->notcovered, d->covered);
7076 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
7077 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
7078 if (cpumask_empty(d->tmpmask))
7080 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
7081 if (cpumask_empty(d->tmpmask))
7083 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7087 "Can not alloc domain group for node %d\n", j);
7091 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
7092 sg->next = prev->next;
7093 cpumask_or(d->covered, d->covered, d->tmpmask);
7100 #endif /* CONFIG_NUMA */
7103 /* Free memory allocated for various sched_group structures */
7104 static void free_sched_groups(const struct cpumask *cpu_map,
7105 struct cpumask *nodemask)
7109 for_each_cpu(cpu, cpu_map) {
7110 struct sched_group **sched_group_nodes
7111 = sched_group_nodes_bycpu[cpu];
7113 if (!sched_group_nodes)
7116 for (i = 0; i < nr_node_ids; i++) {
7117 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7119 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7120 if (cpumask_empty(nodemask))
7130 if (oldsg != sched_group_nodes[i])
7133 kfree(sched_group_nodes);
7134 sched_group_nodes_bycpu[cpu] = NULL;
7137 #else /* !CONFIG_NUMA */
7138 static void free_sched_groups(const struct cpumask *cpu_map,
7139 struct cpumask *nodemask)
7142 #endif /* CONFIG_NUMA */
7145 * Initialize sched groups cpu_power.
7147 * cpu_power indicates the capacity of sched group, which is used while
7148 * distributing the load between different sched groups in a sched domain.
7149 * Typically cpu_power for all the groups in a sched domain will be same unless
7150 * there are asymmetries in the topology. If there are asymmetries, group
7151 * having more cpu_power will pickup more load compared to the group having
7154 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7156 struct sched_domain *child;
7157 struct sched_group *group;
7161 WARN_ON(!sd || !sd->groups);
7163 if (cpu != group_first_cpu(sd->groups))
7166 sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups));
7170 sd->groups->cpu_power = 0;
7173 power = SCHED_LOAD_SCALE;
7174 weight = cpumask_weight(sched_domain_span(sd));
7176 * SMT siblings share the power of a single core.
7177 * Usually multiple threads get a better yield out of
7178 * that one core than a single thread would have,
7179 * reflect that in sd->smt_gain.
7181 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
7182 power *= sd->smt_gain;
7184 power >>= SCHED_LOAD_SHIFT;
7186 sd->groups->cpu_power += power;
7191 * Add cpu_power of each child group to this groups cpu_power.
7193 group = child->groups;
7195 sd->groups->cpu_power += group->cpu_power;
7196 group = group->next;
7197 } while (group != child->groups);
7201 * Initializers for schedule domains
7202 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7205 #ifdef CONFIG_SCHED_DEBUG
7206 # define SD_INIT_NAME(sd, type) sd->name = #type
7208 # define SD_INIT_NAME(sd, type) do { } while (0)
7211 #define SD_INIT(sd, type) sd_init_##type(sd)
7213 #define SD_INIT_FUNC(type) \
7214 static noinline void sd_init_##type(struct sched_domain *sd) \
7216 memset(sd, 0, sizeof(*sd)); \
7217 *sd = SD_##type##_INIT; \
7218 sd->level = SD_LV_##type; \
7219 SD_INIT_NAME(sd, type); \
7224 SD_INIT_FUNC(ALLNODES)
7227 #ifdef CONFIG_SCHED_SMT
7228 SD_INIT_FUNC(SIBLING)
7230 #ifdef CONFIG_SCHED_MC
7233 #ifdef CONFIG_SCHED_BOOK
7237 static int default_relax_domain_level = -1;
7239 static int __init setup_relax_domain_level(char *str)
7243 val = simple_strtoul(str, NULL, 0);
7244 if (val < SD_LV_MAX)
7245 default_relax_domain_level = val;
7249 __setup("relax_domain_level=", setup_relax_domain_level);
7251 static void set_domain_attribute(struct sched_domain *sd,
7252 struct sched_domain_attr *attr)
7256 if (!attr || attr->relax_domain_level < 0) {
7257 if (default_relax_domain_level < 0)
7260 request = default_relax_domain_level;
7262 request = attr->relax_domain_level;
7263 if (request < sd->level) {
7264 /* turn off idle balance on this domain */
7265 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7267 /* turn on idle balance on this domain */
7268 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7272 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7273 const struct cpumask *cpu_map)
7276 case sa_sched_groups:
7277 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
7278 d->sched_group_nodes = NULL;
7280 free_rootdomain(d->rd); /* fall through */
7282 free_cpumask_var(d->tmpmask); /* fall through */
7283 case sa_send_covered:
7284 free_cpumask_var(d->send_covered); /* fall through */
7285 case sa_this_book_map:
7286 free_cpumask_var(d->this_book_map); /* fall through */
7287 case sa_this_core_map:
7288 free_cpumask_var(d->this_core_map); /* fall through */
7289 case sa_this_sibling_map:
7290 free_cpumask_var(d->this_sibling_map); /* fall through */
7292 free_cpumask_var(d->nodemask); /* fall through */
7293 case sa_sched_group_nodes:
7295 kfree(d->sched_group_nodes); /* fall through */
7297 free_cpumask_var(d->notcovered); /* fall through */
7299 free_cpumask_var(d->covered); /* fall through */
7301 free_cpumask_var(d->domainspan); /* fall through */
7308 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7309 const struct cpumask *cpu_map)
7312 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
7314 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
7315 return sa_domainspan;
7316 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
7318 /* Allocate the per-node list of sched groups */
7319 d->sched_group_nodes = kcalloc(nr_node_ids,
7320 sizeof(struct sched_group *), GFP_KERNEL);
7321 if (!d->sched_group_nodes) {
7322 printk(KERN_WARNING "Can not alloc sched group node list\n");
7323 return sa_notcovered;
7325 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
7327 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
7328 return sa_sched_group_nodes;
7329 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
7331 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
7332 return sa_this_sibling_map;
7333 if (!alloc_cpumask_var(&d->this_book_map, GFP_KERNEL))
7334 return sa_this_core_map;
7335 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
7336 return sa_this_book_map;
7337 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
7338 return sa_send_covered;
7339 d->rd = alloc_rootdomain();
7341 printk(KERN_WARNING "Cannot alloc root domain\n");
7344 return sa_rootdomain;
7347 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
7348 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
7350 struct sched_domain *sd = NULL;
7352 struct sched_domain *parent;
7355 if (cpumask_weight(cpu_map) >
7356 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
7357 sd = &per_cpu(allnodes_domains, i).sd;
7358 SD_INIT(sd, ALLNODES);
7359 set_domain_attribute(sd, attr);
7360 cpumask_copy(sched_domain_span(sd), cpu_map);
7361 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
7366 sd = &per_cpu(node_domains, i).sd;
7368 set_domain_attribute(sd, attr);
7369 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7370 sd->parent = parent;
7373 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7378 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7379 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7380 struct sched_domain *parent, int i)
7382 struct sched_domain *sd;
7383 sd = &per_cpu(phys_domains, i).sd;
7385 set_domain_attribute(sd, attr);
7386 cpumask_copy(sched_domain_span(sd), d->nodemask);
7387 sd->parent = parent;
7390 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7394 static struct sched_domain *__build_book_sched_domain(struct s_data *d,
7395 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7396 struct sched_domain *parent, int i)
7398 struct sched_domain *sd = parent;
7399 #ifdef CONFIG_SCHED_BOOK
7400 sd = &per_cpu(book_domains, i).sd;
7402 set_domain_attribute(sd, attr);
7403 cpumask_and(sched_domain_span(sd), cpu_map, cpu_book_mask(i));
7404 sd->parent = parent;
7406 cpu_to_book_group(i, cpu_map, &sd->groups, d->tmpmask);
7411 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7412 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7413 struct sched_domain *parent, int i)
7415 struct sched_domain *sd = parent;
7416 #ifdef CONFIG_SCHED_MC
7417 sd = &per_cpu(core_domains, i).sd;
7419 set_domain_attribute(sd, attr);
7420 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7421 sd->parent = parent;
7423 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7428 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7429 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7430 struct sched_domain *parent, int i)
7432 struct sched_domain *sd = parent;
7433 #ifdef CONFIG_SCHED_SMT
7434 sd = &per_cpu(cpu_domains, i).sd;
7435 SD_INIT(sd, SIBLING);
7436 set_domain_attribute(sd, attr);
7437 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7438 sd->parent = parent;
7440 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7445 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7446 const struct cpumask *cpu_map, int cpu)
7449 #ifdef CONFIG_SCHED_SMT
7450 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7451 cpumask_and(d->this_sibling_map, cpu_map,
7452 topology_thread_cpumask(cpu));
7453 if (cpu == cpumask_first(d->this_sibling_map))
7454 init_sched_build_groups(d->this_sibling_map, cpu_map,
7456 d->send_covered, d->tmpmask);
7459 #ifdef CONFIG_SCHED_MC
7460 case SD_LV_MC: /* set up multi-core groups */
7461 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7462 if (cpu == cpumask_first(d->this_core_map))
7463 init_sched_build_groups(d->this_core_map, cpu_map,
7465 d->send_covered, d->tmpmask);
7468 #ifdef CONFIG_SCHED_BOOK
7469 case SD_LV_BOOK: /* set up book groups */
7470 cpumask_and(d->this_book_map, cpu_map, cpu_book_mask(cpu));
7471 if (cpu == cpumask_first(d->this_book_map))
7472 init_sched_build_groups(d->this_book_map, cpu_map,
7474 d->send_covered, d->tmpmask);
7477 case SD_LV_CPU: /* set up physical groups */
7478 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7479 if (!cpumask_empty(d->nodemask))
7480 init_sched_build_groups(d->nodemask, cpu_map,
7482 d->send_covered, d->tmpmask);
7485 case SD_LV_ALLNODES:
7486 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7487 d->send_covered, d->tmpmask);
7496 * Build sched domains for a given set of cpus and attach the sched domains
7497 * to the individual cpus
7499 static int __build_sched_domains(const struct cpumask *cpu_map,
7500 struct sched_domain_attr *attr)
7502 enum s_alloc alloc_state = sa_none;
7504 struct sched_domain *sd;
7510 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7511 if (alloc_state != sa_rootdomain)
7513 alloc_state = sa_sched_groups;
7516 * Set up domains for cpus specified by the cpu_map.
7518 for_each_cpu(i, cpu_map) {
7519 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7522 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7523 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7524 sd = __build_book_sched_domain(&d, cpu_map, attr, sd, i);
7525 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7526 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7529 for_each_cpu(i, cpu_map) {
7530 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7531 build_sched_groups(&d, SD_LV_BOOK, cpu_map, i);
7532 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7535 /* Set up physical groups */
7536 for (i = 0; i < nr_node_ids; i++)
7537 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7540 /* Set up node groups */
7542 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7544 for (i = 0; i < nr_node_ids; i++)
7545 if (build_numa_sched_groups(&d, cpu_map, i))
7549 /* Calculate CPU power for physical packages and nodes */
7550 #ifdef CONFIG_SCHED_SMT
7551 for_each_cpu(i, cpu_map) {
7552 sd = &per_cpu(cpu_domains, i).sd;
7553 init_sched_groups_power(i, sd);
7556 #ifdef CONFIG_SCHED_MC
7557 for_each_cpu(i, cpu_map) {
7558 sd = &per_cpu(core_domains, i).sd;
7559 init_sched_groups_power(i, sd);
7562 #ifdef CONFIG_SCHED_BOOK
7563 for_each_cpu(i, cpu_map) {
7564 sd = &per_cpu(book_domains, i).sd;
7565 init_sched_groups_power(i, sd);
7569 for_each_cpu(i, cpu_map) {
7570 sd = &per_cpu(phys_domains, i).sd;
7571 init_sched_groups_power(i, sd);
7575 for (i = 0; i < nr_node_ids; i++)
7576 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7578 if (d.sd_allnodes) {
7579 struct sched_group *sg;
7581 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7583 init_numa_sched_groups_power(sg);
7587 /* Attach the domains */
7588 for_each_cpu(i, cpu_map) {
7589 #ifdef CONFIG_SCHED_SMT
7590 sd = &per_cpu(cpu_domains, i).sd;
7591 #elif defined(CONFIG_SCHED_MC)
7592 sd = &per_cpu(core_domains, i).sd;
7593 #elif defined(CONFIG_SCHED_BOOK)
7594 sd = &per_cpu(book_domains, i).sd;
7596 sd = &per_cpu(phys_domains, i).sd;
7598 cpu_attach_domain(sd, d.rd, i);
7601 d.sched_group_nodes = NULL; /* don't free this we still need it */
7602 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7606 __free_domain_allocs(&d, alloc_state, cpu_map);
7610 static int build_sched_domains(const struct cpumask *cpu_map)
7612 return __build_sched_domains(cpu_map, NULL);
7615 static cpumask_var_t *doms_cur; /* current sched domains */
7616 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7617 static struct sched_domain_attr *dattr_cur;
7618 /* attribues of custom domains in 'doms_cur' */
7621 * Special case: If a kmalloc of a doms_cur partition (array of
7622 * cpumask) fails, then fallback to a single sched domain,
7623 * as determined by the single cpumask fallback_doms.
7625 static cpumask_var_t fallback_doms;
7628 * arch_update_cpu_topology lets virtualized architectures update the
7629 * cpu core maps. It is supposed to return 1 if the topology changed
7630 * or 0 if it stayed the same.
7632 int __attribute__((weak)) arch_update_cpu_topology(void)
7637 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7640 cpumask_var_t *doms;
7642 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7645 for (i = 0; i < ndoms; i++) {
7646 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7647 free_sched_domains(doms, i);
7654 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7657 for (i = 0; i < ndoms; i++)
7658 free_cpumask_var(doms[i]);
7663 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7664 * For now this just excludes isolated cpus, but could be used to
7665 * exclude other special cases in the future.
7667 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7671 arch_update_cpu_topology();
7673 doms_cur = alloc_sched_domains(ndoms_cur);
7675 doms_cur = &fallback_doms;
7676 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7678 err = build_sched_domains(doms_cur[0]);
7679 register_sched_domain_sysctl();
7684 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7685 struct cpumask *tmpmask)
7687 free_sched_groups(cpu_map, tmpmask);
7691 * Detach sched domains from a group of cpus specified in cpu_map
7692 * These cpus will now be attached to the NULL domain
7694 static void detach_destroy_domains(const struct cpumask *cpu_map)
7696 /* Save because hotplug lock held. */
7697 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7700 for_each_cpu(i, cpu_map)
7701 cpu_attach_domain(NULL, &def_root_domain, i);
7702 synchronize_sched();
7703 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7706 /* handle null as "default" */
7707 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7708 struct sched_domain_attr *new, int idx_new)
7710 struct sched_domain_attr tmp;
7717 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7718 new ? (new + idx_new) : &tmp,
7719 sizeof(struct sched_domain_attr));
7723 * Partition sched domains as specified by the 'ndoms_new'
7724 * cpumasks in the array doms_new[] of cpumasks. This compares
7725 * doms_new[] to the current sched domain partitioning, doms_cur[].
7726 * It destroys each deleted domain and builds each new domain.
7728 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7729 * The masks don't intersect (don't overlap.) We should setup one
7730 * sched domain for each mask. CPUs not in any of the cpumasks will
7731 * not be load balanced. If the same cpumask appears both in the
7732 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7735 * The passed in 'doms_new' should be allocated using
7736 * alloc_sched_domains. This routine takes ownership of it and will
7737 * free_sched_domains it when done with it. If the caller failed the
7738 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7739 * and partition_sched_domains() will fallback to the single partition
7740 * 'fallback_doms', it also forces the domains to be rebuilt.
7742 * If doms_new == NULL it will be replaced with cpu_online_mask.
7743 * ndoms_new == 0 is a special case for destroying existing domains,
7744 * and it will not create the default domain.
7746 * Call with hotplug lock held
7748 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7749 struct sched_domain_attr *dattr_new)
7754 mutex_lock(&sched_domains_mutex);
7756 /* always unregister in case we don't destroy any domains */
7757 unregister_sched_domain_sysctl();
7759 /* Let architecture update cpu core mappings. */
7760 new_topology = arch_update_cpu_topology();
7762 n = doms_new ? ndoms_new : 0;
7764 /* Destroy deleted domains */
7765 for (i = 0; i < ndoms_cur; i++) {
7766 for (j = 0; j < n && !new_topology; j++) {
7767 if (cpumask_equal(doms_cur[i], doms_new[j])
7768 && dattrs_equal(dattr_cur, i, dattr_new, j))
7771 /* no match - a current sched domain not in new doms_new[] */
7772 detach_destroy_domains(doms_cur[i]);
7777 if (doms_new == NULL) {
7779 doms_new = &fallback_doms;
7780 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7781 WARN_ON_ONCE(dattr_new);
7784 /* Build new domains */
7785 for (i = 0; i < ndoms_new; i++) {
7786 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7787 if (cpumask_equal(doms_new[i], doms_cur[j])
7788 && dattrs_equal(dattr_new, i, dattr_cur, j))
7791 /* no match - add a new doms_new */
7792 __build_sched_domains(doms_new[i],
7793 dattr_new ? dattr_new + i : NULL);
7798 /* Remember the new sched domains */
7799 if (doms_cur != &fallback_doms)
7800 free_sched_domains(doms_cur, ndoms_cur);
7801 kfree(dattr_cur); /* kfree(NULL) is safe */
7802 doms_cur = doms_new;
7803 dattr_cur = dattr_new;
7804 ndoms_cur = ndoms_new;
7806 register_sched_domain_sysctl();
7808 mutex_unlock(&sched_domains_mutex);
7811 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7812 static void arch_reinit_sched_domains(void)
7816 /* Destroy domains first to force the rebuild */
7817 partition_sched_domains(0, NULL, NULL);
7819 rebuild_sched_domains();
7823 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7825 unsigned int level = 0;
7827 if (sscanf(buf, "%u", &level) != 1)
7831 * level is always be positive so don't check for
7832 * level < POWERSAVINGS_BALANCE_NONE which is 0
7833 * What happens on 0 or 1 byte write,
7834 * need to check for count as well?
7837 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7841 sched_smt_power_savings = level;
7843 sched_mc_power_savings = level;
7845 arch_reinit_sched_domains();
7850 #ifdef CONFIG_SCHED_MC
7851 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7852 struct sysdev_class_attribute *attr,
7855 return sprintf(page, "%u\n", sched_mc_power_savings);
7857 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7858 struct sysdev_class_attribute *attr,
7859 const char *buf, size_t count)
7861 return sched_power_savings_store(buf, count, 0);
7863 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7864 sched_mc_power_savings_show,
7865 sched_mc_power_savings_store);
7868 #ifdef CONFIG_SCHED_SMT
7869 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7870 struct sysdev_class_attribute *attr,
7873 return sprintf(page, "%u\n", sched_smt_power_savings);
7875 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7876 struct sysdev_class_attribute *attr,
7877 const char *buf, size_t count)
7879 return sched_power_savings_store(buf, count, 1);
7881 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7882 sched_smt_power_savings_show,
7883 sched_smt_power_savings_store);
7886 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7890 #ifdef CONFIG_SCHED_SMT
7892 err = sysfs_create_file(&cls->kset.kobj,
7893 &attr_sched_smt_power_savings.attr);
7895 #ifdef CONFIG_SCHED_MC
7896 if (!err && mc_capable())
7897 err = sysfs_create_file(&cls->kset.kobj,
7898 &attr_sched_mc_power_savings.attr);
7902 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7905 * Update cpusets according to cpu_active mask. If cpusets are
7906 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7907 * around partition_sched_domains().
7909 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7912 switch (action & ~CPU_TASKS_FROZEN) {
7914 case CPU_DOWN_FAILED:
7915 cpuset_update_active_cpus();
7922 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7925 switch (action & ~CPU_TASKS_FROZEN) {
7926 case CPU_DOWN_PREPARE:
7927 cpuset_update_active_cpus();
7934 static int update_runtime(struct notifier_block *nfb,
7935 unsigned long action, void *hcpu)
7937 int cpu = (int)(long)hcpu;
7940 case CPU_DOWN_PREPARE:
7941 case CPU_DOWN_PREPARE_FROZEN:
7942 disable_runtime(cpu_rq(cpu));
7945 case CPU_DOWN_FAILED:
7946 case CPU_DOWN_FAILED_FROZEN:
7948 case CPU_ONLINE_FROZEN:
7949 enable_runtime(cpu_rq(cpu));
7957 void __init sched_init_smp(void)
7959 cpumask_var_t non_isolated_cpus;
7961 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7962 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7964 #if defined(CONFIG_NUMA)
7965 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7967 BUG_ON(sched_group_nodes_bycpu == NULL);
7970 mutex_lock(&sched_domains_mutex);
7971 arch_init_sched_domains(cpu_active_mask);
7972 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7973 if (cpumask_empty(non_isolated_cpus))
7974 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7975 mutex_unlock(&sched_domains_mutex);
7978 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7979 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7981 /* RT runtime code needs to handle some hotplug events */
7982 hotcpu_notifier(update_runtime, 0);
7986 /* Move init over to a non-isolated CPU */
7987 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7989 sched_init_granularity();
7990 free_cpumask_var(non_isolated_cpus);
7992 init_sched_rt_class();
7995 void __init sched_init_smp(void)
7997 sched_init_granularity();
7999 #endif /* CONFIG_SMP */
8001 const_debug unsigned int sysctl_timer_migration = 1;
8003 int in_sched_functions(unsigned long addr)
8005 return in_lock_functions(addr) ||
8006 (addr >= (unsigned long)__sched_text_start
8007 && addr < (unsigned long)__sched_text_end);
8010 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8012 cfs_rq->tasks_timeline = RB_ROOT;
8013 INIT_LIST_HEAD(&cfs_rq->tasks);
8014 #ifdef CONFIG_FAIR_GROUP_SCHED
8017 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8020 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8022 struct rt_prio_array *array;
8025 array = &rt_rq->active;
8026 for (i = 0; i < MAX_RT_PRIO; i++) {
8027 INIT_LIST_HEAD(array->queue + i);
8028 __clear_bit(i, array->bitmap);
8030 /* delimiter for bitsearch: */
8031 __set_bit(MAX_RT_PRIO, array->bitmap);
8033 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8034 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8036 rt_rq->highest_prio.next = MAX_RT_PRIO;
8040 rt_rq->rt_nr_migratory = 0;
8041 rt_rq->overloaded = 0;
8042 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
8046 rt_rq->rt_throttled = 0;
8047 rt_rq->rt_runtime = 0;
8048 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
8050 #ifdef CONFIG_RT_GROUP_SCHED
8051 rt_rq->rt_nr_boosted = 0;
8056 #ifdef CONFIG_FAIR_GROUP_SCHED
8057 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8058 struct sched_entity *se, int cpu, int add,
8059 struct sched_entity *parent)
8061 struct rq *rq = cpu_rq(cpu);
8062 tg->cfs_rq[cpu] = cfs_rq;
8063 init_cfs_rq(cfs_rq, rq);
8066 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8069 /* se could be NULL for init_task_group */
8074 se->cfs_rq = &rq->cfs;
8076 se->cfs_rq = parent->my_q;
8079 se->load.weight = tg->shares;
8080 se->load.inv_weight = 0;
8081 se->parent = parent;
8085 #ifdef CONFIG_RT_GROUP_SCHED
8086 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8087 struct sched_rt_entity *rt_se, int cpu, int add,
8088 struct sched_rt_entity *parent)
8090 struct rq *rq = cpu_rq(cpu);
8092 tg->rt_rq[cpu] = rt_rq;
8093 init_rt_rq(rt_rq, rq);
8095 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8097 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8099 tg->rt_se[cpu] = rt_se;
8104 rt_se->rt_rq = &rq->rt;
8106 rt_se->rt_rq = parent->my_q;
8108 rt_se->my_q = rt_rq;
8109 rt_se->parent = parent;
8110 INIT_LIST_HEAD(&rt_se->run_list);
8114 void __init sched_init(void)
8117 unsigned long alloc_size = 0, ptr;
8119 #ifdef CONFIG_FAIR_GROUP_SCHED
8120 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8122 #ifdef CONFIG_RT_GROUP_SCHED
8123 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8125 #ifdef CONFIG_CPUMASK_OFFSTACK
8126 alloc_size += num_possible_cpus() * cpumask_size();
8129 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8131 #ifdef CONFIG_FAIR_GROUP_SCHED
8132 init_task_group.se = (struct sched_entity **)ptr;
8133 ptr += nr_cpu_ids * sizeof(void **);
8135 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8136 ptr += nr_cpu_ids * sizeof(void **);
8138 #endif /* CONFIG_FAIR_GROUP_SCHED */
8139 #ifdef CONFIG_RT_GROUP_SCHED
8140 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8141 ptr += nr_cpu_ids * sizeof(void **);
8143 init_task_group.rt_rq = (struct rt_rq **)ptr;
8144 ptr += nr_cpu_ids * sizeof(void **);
8146 #endif /* CONFIG_RT_GROUP_SCHED */
8147 #ifdef CONFIG_CPUMASK_OFFSTACK
8148 for_each_possible_cpu(i) {
8149 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8150 ptr += cpumask_size();
8152 #endif /* CONFIG_CPUMASK_OFFSTACK */
8156 init_defrootdomain();
8159 init_rt_bandwidth(&def_rt_bandwidth,
8160 global_rt_period(), global_rt_runtime());
8162 #ifdef CONFIG_RT_GROUP_SCHED
8163 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8164 global_rt_period(), global_rt_runtime());
8165 #endif /* CONFIG_RT_GROUP_SCHED */
8167 #ifdef CONFIG_CGROUP_SCHED
8168 list_add(&init_task_group.list, &task_groups);
8169 INIT_LIST_HEAD(&init_task_group.children);
8171 #endif /* CONFIG_CGROUP_SCHED */
8173 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
8174 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
8175 __alignof__(unsigned long));
8177 for_each_possible_cpu(i) {
8181 raw_spin_lock_init(&rq->lock);
8183 rq->calc_load_active = 0;
8184 rq->calc_load_update = jiffies + LOAD_FREQ;
8185 init_cfs_rq(&rq->cfs, rq);
8186 init_rt_rq(&rq->rt, rq);
8187 #ifdef CONFIG_FAIR_GROUP_SCHED
8188 init_task_group.shares = init_task_group_load;
8189 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8190 #ifdef CONFIG_CGROUP_SCHED
8192 * How much cpu bandwidth does init_task_group get?
8194 * In case of task-groups formed thr' the cgroup filesystem, it
8195 * gets 100% of the cpu resources in the system. This overall
8196 * system cpu resource is divided among the tasks of
8197 * init_task_group and its child task-groups in a fair manner,
8198 * based on each entity's (task or task-group's) weight
8199 * (se->load.weight).
8201 * In other words, if init_task_group has 10 tasks of weight
8202 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8203 * then A0's share of the cpu resource is:
8205 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8207 * We achieve this by letting init_task_group's tasks sit
8208 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8210 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8212 #endif /* CONFIG_FAIR_GROUP_SCHED */
8214 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8215 #ifdef CONFIG_RT_GROUP_SCHED
8216 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8217 #ifdef CONFIG_CGROUP_SCHED
8218 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8222 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8223 rq->cpu_load[j] = 0;
8225 rq->last_load_update_tick = jiffies;
8230 rq->cpu_power = SCHED_LOAD_SCALE;
8231 rq->post_schedule = 0;
8232 rq->active_balance = 0;
8233 rq->next_balance = jiffies;
8238 rq->avg_idle = 2*sysctl_sched_migration_cost;
8239 rq_attach_root(rq, &def_root_domain);
8241 rq->nohz_balance_kick = 0;
8242 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8246 atomic_set(&rq->nr_iowait, 0);
8249 set_load_weight(&init_task);
8251 #ifdef CONFIG_PREEMPT_NOTIFIERS
8252 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8256 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8259 #ifdef CONFIG_RT_MUTEXES
8260 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
8264 * The boot idle thread does lazy MMU switching as well:
8266 atomic_inc(&init_mm.mm_count);
8267 enter_lazy_tlb(&init_mm, current);
8270 * Make us the idle thread. Technically, schedule() should not be
8271 * called from this thread, however somewhere below it might be,
8272 * but because we are the idle thread, we just pick up running again
8273 * when this runqueue becomes "idle".
8275 init_idle(current, smp_processor_id());
8277 calc_load_update = jiffies + LOAD_FREQ;
8280 * During early bootup we pretend to be a normal task:
8282 current->sched_class = &fair_sched_class;
8284 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8285 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8288 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8289 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8290 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8291 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8292 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8294 /* May be allocated at isolcpus cmdline parse time */
8295 if (cpu_isolated_map == NULL)
8296 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8301 scheduler_running = 1;
8304 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8305 static inline int preempt_count_equals(int preempt_offset)
8307 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8309 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
8312 void __might_sleep(const char *file, int line, int preempt_offset)
8315 static unsigned long prev_jiffy; /* ratelimiting */
8317 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8318 system_state != SYSTEM_RUNNING || oops_in_progress)
8320 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8322 prev_jiffy = jiffies;
8325 "BUG: sleeping function called from invalid context at %s:%d\n",
8328 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8329 in_atomic(), irqs_disabled(),
8330 current->pid, current->comm);
8332 debug_show_held_locks(current);
8333 if (irqs_disabled())
8334 print_irqtrace_events(current);
8338 EXPORT_SYMBOL(__might_sleep);
8341 #ifdef CONFIG_MAGIC_SYSRQ
8342 static void normalize_task(struct rq *rq, struct task_struct *p)
8346 on_rq = p->se.on_rq;
8348 deactivate_task(rq, p, 0);
8349 __setscheduler(rq, p, SCHED_NORMAL, 0);
8351 activate_task(rq, p, 0);
8352 resched_task(rq->curr);
8356 void normalize_rt_tasks(void)
8358 struct task_struct *g, *p;
8359 unsigned long flags;
8362 read_lock_irqsave(&tasklist_lock, flags);
8363 do_each_thread(g, p) {
8365 * Only normalize user tasks:
8370 p->se.exec_start = 0;
8371 #ifdef CONFIG_SCHEDSTATS
8372 p->se.statistics.wait_start = 0;
8373 p->se.statistics.sleep_start = 0;
8374 p->se.statistics.block_start = 0;
8379 * Renice negative nice level userspace
8382 if (TASK_NICE(p) < 0 && p->mm)
8383 set_user_nice(p, 0);
8387 raw_spin_lock(&p->pi_lock);
8388 rq = __task_rq_lock(p);
8390 normalize_task(rq, p);
8392 __task_rq_unlock(rq);
8393 raw_spin_unlock(&p->pi_lock);
8394 } while_each_thread(g, p);
8396 read_unlock_irqrestore(&tasklist_lock, flags);
8399 #endif /* CONFIG_MAGIC_SYSRQ */
8401 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8403 * These functions are only useful for the IA64 MCA handling, or kdb.
8405 * They can only be called when the whole system has been
8406 * stopped - every CPU needs to be quiescent, and no scheduling
8407 * activity can take place. Using them for anything else would
8408 * be a serious bug, and as a result, they aren't even visible
8409 * under any other configuration.
8413 * curr_task - return the current task for a given cpu.
8414 * @cpu: the processor in question.
8416 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8418 struct task_struct *curr_task(int cpu)
8420 return cpu_curr(cpu);
8423 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8427 * set_curr_task - set the current task for a given cpu.
8428 * @cpu: the processor in question.
8429 * @p: the task pointer to set.
8431 * Description: This function must only be used when non-maskable interrupts
8432 * are serviced on a separate stack. It allows the architecture to switch the
8433 * notion of the current task on a cpu in a non-blocking manner. This function
8434 * must be called with all CPU's synchronized, and interrupts disabled, the
8435 * and caller must save the original value of the current task (see
8436 * curr_task() above) and restore that value before reenabling interrupts and
8437 * re-starting the system.
8439 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8441 void set_curr_task(int cpu, struct task_struct *p)
8448 #ifdef CONFIG_FAIR_GROUP_SCHED
8449 static void free_fair_sched_group(struct task_group *tg)
8453 for_each_possible_cpu(i) {
8455 kfree(tg->cfs_rq[i]);
8465 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8467 struct cfs_rq *cfs_rq;
8468 struct sched_entity *se;
8472 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8475 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8479 tg->shares = NICE_0_LOAD;
8481 for_each_possible_cpu(i) {
8484 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8485 GFP_KERNEL, cpu_to_node(i));
8489 se = kzalloc_node(sizeof(struct sched_entity),
8490 GFP_KERNEL, cpu_to_node(i));
8494 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8505 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8507 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8508 &cpu_rq(cpu)->leaf_cfs_rq_list);
8511 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8513 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8515 #else /* !CONFG_FAIR_GROUP_SCHED */
8516 static inline void free_fair_sched_group(struct task_group *tg)
8521 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8526 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8530 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8533 #endif /* CONFIG_FAIR_GROUP_SCHED */
8535 #ifdef CONFIG_RT_GROUP_SCHED
8536 static void free_rt_sched_group(struct task_group *tg)
8540 destroy_rt_bandwidth(&tg->rt_bandwidth);
8542 for_each_possible_cpu(i) {
8544 kfree(tg->rt_rq[i]);
8546 kfree(tg->rt_se[i]);
8554 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8556 struct rt_rq *rt_rq;
8557 struct sched_rt_entity *rt_se;
8561 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8564 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8568 init_rt_bandwidth(&tg->rt_bandwidth,
8569 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8571 for_each_possible_cpu(i) {
8574 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8575 GFP_KERNEL, cpu_to_node(i));
8579 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8580 GFP_KERNEL, cpu_to_node(i));
8584 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8595 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8597 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8598 &cpu_rq(cpu)->leaf_rt_rq_list);
8601 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8603 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8605 #else /* !CONFIG_RT_GROUP_SCHED */
8606 static inline void free_rt_sched_group(struct task_group *tg)
8611 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8616 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8620 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8623 #endif /* CONFIG_RT_GROUP_SCHED */
8625 #ifdef CONFIG_CGROUP_SCHED
8626 static void free_sched_group(struct task_group *tg)
8628 free_fair_sched_group(tg);
8629 free_rt_sched_group(tg);
8633 /* allocate runqueue etc for a new task group */
8634 struct task_group *sched_create_group(struct task_group *parent)
8636 struct task_group *tg;
8637 unsigned long flags;
8640 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8642 return ERR_PTR(-ENOMEM);
8644 if (!alloc_fair_sched_group(tg, parent))
8647 if (!alloc_rt_sched_group(tg, parent))
8650 spin_lock_irqsave(&task_group_lock, flags);
8651 for_each_possible_cpu(i) {
8652 register_fair_sched_group(tg, i);
8653 register_rt_sched_group(tg, i);
8655 list_add_rcu(&tg->list, &task_groups);
8657 WARN_ON(!parent); /* root should already exist */
8659 tg->parent = parent;
8660 INIT_LIST_HEAD(&tg->children);
8661 list_add_rcu(&tg->siblings, &parent->children);
8662 spin_unlock_irqrestore(&task_group_lock, flags);
8667 free_sched_group(tg);
8668 return ERR_PTR(-ENOMEM);
8671 /* rcu callback to free various structures associated with a task group */
8672 static void free_sched_group_rcu(struct rcu_head *rhp)
8674 /* now it should be safe to free those cfs_rqs */
8675 free_sched_group(container_of(rhp, struct task_group, rcu));
8678 /* Destroy runqueue etc associated with a task group */
8679 void sched_destroy_group(struct task_group *tg)
8681 unsigned long flags;
8684 spin_lock_irqsave(&task_group_lock, flags);
8685 for_each_possible_cpu(i) {
8686 unregister_fair_sched_group(tg, i);
8687 unregister_rt_sched_group(tg, i);
8689 list_del_rcu(&tg->list);
8690 list_del_rcu(&tg->siblings);
8691 spin_unlock_irqrestore(&task_group_lock, flags);
8693 /* wait for possible concurrent references to cfs_rqs complete */
8694 call_rcu(&tg->rcu, free_sched_group_rcu);
8697 /* change task's runqueue when it moves between groups.
8698 * The caller of this function should have put the task in its new group
8699 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8700 * reflect its new group.
8702 void sched_move_task(struct task_struct *tsk)
8705 unsigned long flags;
8708 rq = task_rq_lock(tsk, &flags);
8710 running = task_current(rq, tsk);
8711 on_rq = tsk->se.on_rq;
8714 dequeue_task(rq, tsk, 0);
8715 if (unlikely(running))
8716 tsk->sched_class->put_prev_task(rq, tsk);
8718 #ifdef CONFIG_FAIR_GROUP_SCHED
8719 if (tsk->sched_class->task_move_group)
8720 tsk->sched_class->task_move_group(tsk, on_rq);
8723 set_task_rq(tsk, task_cpu(tsk));
8725 if (unlikely(running))
8726 tsk->sched_class->set_curr_task(rq);
8728 enqueue_task(rq, tsk, 0);
8730 task_rq_unlock(rq, &flags);
8732 #endif /* CONFIG_CGROUP_SCHED */
8734 #ifdef CONFIG_FAIR_GROUP_SCHED
8735 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8737 struct cfs_rq *cfs_rq = se->cfs_rq;
8742 dequeue_entity(cfs_rq, se, 0);
8744 se->load.weight = shares;
8745 se->load.inv_weight = 0;
8748 enqueue_entity(cfs_rq, se, 0);
8751 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8753 struct cfs_rq *cfs_rq = se->cfs_rq;
8754 struct rq *rq = cfs_rq->rq;
8755 unsigned long flags;
8757 raw_spin_lock_irqsave(&rq->lock, flags);
8758 __set_se_shares(se, shares);
8759 raw_spin_unlock_irqrestore(&rq->lock, flags);
8762 static DEFINE_MUTEX(shares_mutex);
8764 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8767 unsigned long flags;
8770 * We can't change the weight of the root cgroup.
8775 if (shares < MIN_SHARES)
8776 shares = MIN_SHARES;
8777 else if (shares > MAX_SHARES)
8778 shares = MAX_SHARES;
8780 mutex_lock(&shares_mutex);
8781 if (tg->shares == shares)
8784 spin_lock_irqsave(&task_group_lock, flags);
8785 for_each_possible_cpu(i)
8786 unregister_fair_sched_group(tg, i);
8787 list_del_rcu(&tg->siblings);
8788 spin_unlock_irqrestore(&task_group_lock, flags);
8790 /* wait for any ongoing reference to this group to finish */
8791 synchronize_sched();
8794 * Now we are free to modify the group's share on each cpu
8795 * w/o tripping rebalance_share or load_balance_fair.
8797 tg->shares = shares;
8798 for_each_possible_cpu(i) {
8802 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8803 set_se_shares(tg->se[i], shares);
8807 * Enable load balance activity on this group, by inserting it back on
8808 * each cpu's rq->leaf_cfs_rq_list.
8810 spin_lock_irqsave(&task_group_lock, flags);
8811 for_each_possible_cpu(i)
8812 register_fair_sched_group(tg, i);
8813 list_add_rcu(&tg->siblings, &tg->parent->children);
8814 spin_unlock_irqrestore(&task_group_lock, flags);
8816 mutex_unlock(&shares_mutex);
8820 unsigned long sched_group_shares(struct task_group *tg)
8826 #ifdef CONFIG_RT_GROUP_SCHED
8828 * Ensure that the real time constraints are schedulable.
8830 static DEFINE_MUTEX(rt_constraints_mutex);
8832 static unsigned long to_ratio(u64 period, u64 runtime)
8834 if (runtime == RUNTIME_INF)
8837 return div64_u64(runtime << 20, period);
8840 /* Must be called with tasklist_lock held */
8841 static inline int tg_has_rt_tasks(struct task_group *tg)
8843 struct task_struct *g, *p;
8845 do_each_thread(g, p) {
8846 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8848 } while_each_thread(g, p);
8853 struct rt_schedulable_data {
8854 struct task_group *tg;
8859 static int tg_schedulable(struct task_group *tg, void *data)
8861 struct rt_schedulable_data *d = data;
8862 struct task_group *child;
8863 unsigned long total, sum = 0;
8864 u64 period, runtime;
8866 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8867 runtime = tg->rt_bandwidth.rt_runtime;
8870 period = d->rt_period;
8871 runtime = d->rt_runtime;
8875 * Cannot have more runtime than the period.
8877 if (runtime > period && runtime != RUNTIME_INF)
8881 * Ensure we don't starve existing RT tasks.
8883 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8886 total = to_ratio(period, runtime);
8889 * Nobody can have more than the global setting allows.
8891 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8895 * The sum of our children's runtime should not exceed our own.
8897 list_for_each_entry_rcu(child, &tg->children, siblings) {
8898 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8899 runtime = child->rt_bandwidth.rt_runtime;
8901 if (child == d->tg) {
8902 period = d->rt_period;
8903 runtime = d->rt_runtime;
8906 sum += to_ratio(period, runtime);
8915 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8917 struct rt_schedulable_data data = {
8919 .rt_period = period,
8920 .rt_runtime = runtime,
8923 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8926 static int tg_set_bandwidth(struct task_group *tg,
8927 u64 rt_period, u64 rt_runtime)
8931 mutex_lock(&rt_constraints_mutex);
8932 read_lock(&tasklist_lock);
8933 err = __rt_schedulable(tg, rt_period, rt_runtime);
8937 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8938 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8939 tg->rt_bandwidth.rt_runtime = rt_runtime;
8941 for_each_possible_cpu(i) {
8942 struct rt_rq *rt_rq = tg->rt_rq[i];
8944 raw_spin_lock(&rt_rq->rt_runtime_lock);
8945 rt_rq->rt_runtime = rt_runtime;
8946 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8948 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8950 read_unlock(&tasklist_lock);
8951 mutex_unlock(&rt_constraints_mutex);
8956 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8958 u64 rt_runtime, rt_period;
8960 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8961 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8962 if (rt_runtime_us < 0)
8963 rt_runtime = RUNTIME_INF;
8965 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8968 long sched_group_rt_runtime(struct task_group *tg)
8972 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8975 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8976 do_div(rt_runtime_us, NSEC_PER_USEC);
8977 return rt_runtime_us;
8980 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8982 u64 rt_runtime, rt_period;
8984 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8985 rt_runtime = tg->rt_bandwidth.rt_runtime;
8990 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8993 long sched_group_rt_period(struct task_group *tg)
8997 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8998 do_div(rt_period_us, NSEC_PER_USEC);
8999 return rt_period_us;
9002 static int sched_rt_global_constraints(void)
9004 u64 runtime, period;
9007 if (sysctl_sched_rt_period <= 0)
9010 runtime = global_rt_runtime();
9011 period = global_rt_period();
9014 * Sanity check on the sysctl variables.
9016 if (runtime > period && runtime != RUNTIME_INF)
9019 mutex_lock(&rt_constraints_mutex);
9020 read_lock(&tasklist_lock);
9021 ret = __rt_schedulable(NULL, 0, 0);
9022 read_unlock(&tasklist_lock);
9023 mutex_unlock(&rt_constraints_mutex);
9028 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
9030 /* Don't accept realtime tasks when there is no way for them to run */
9031 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
9037 #else /* !CONFIG_RT_GROUP_SCHED */
9038 static int sched_rt_global_constraints(void)
9040 unsigned long flags;
9043 if (sysctl_sched_rt_period <= 0)
9047 * There's always some RT tasks in the root group
9048 * -- migration, kstopmachine etc..
9050 if (sysctl_sched_rt_runtime == 0)
9053 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9054 for_each_possible_cpu(i) {
9055 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9057 raw_spin_lock(&rt_rq->rt_runtime_lock);
9058 rt_rq->rt_runtime = global_rt_runtime();
9059 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9061 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9065 #endif /* CONFIG_RT_GROUP_SCHED */
9067 int sched_rt_handler(struct ctl_table *table, int write,
9068 void __user *buffer, size_t *lenp,
9072 int old_period, old_runtime;
9073 static DEFINE_MUTEX(mutex);
9076 old_period = sysctl_sched_rt_period;
9077 old_runtime = sysctl_sched_rt_runtime;
9079 ret = proc_dointvec(table, write, buffer, lenp, ppos);
9081 if (!ret && write) {
9082 ret = sched_rt_global_constraints();
9084 sysctl_sched_rt_period = old_period;
9085 sysctl_sched_rt_runtime = old_runtime;
9087 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9088 def_rt_bandwidth.rt_period =
9089 ns_to_ktime(global_rt_period());
9092 mutex_unlock(&mutex);
9097 #ifdef CONFIG_CGROUP_SCHED
9099 /* return corresponding task_group object of a cgroup */
9100 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9102 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9103 struct task_group, css);
9106 static struct cgroup_subsys_state *
9107 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9109 struct task_group *tg, *parent;
9111 if (!cgrp->parent) {
9112 /* This is early initialization for the top cgroup */
9113 return &init_task_group.css;
9116 parent = cgroup_tg(cgrp->parent);
9117 tg = sched_create_group(parent);
9119 return ERR_PTR(-ENOMEM);
9125 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9127 struct task_group *tg = cgroup_tg(cgrp);
9129 sched_destroy_group(tg);
9133 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9135 #ifdef CONFIG_RT_GROUP_SCHED
9136 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9139 /* We don't support RT-tasks being in separate groups */
9140 if (tsk->sched_class != &fair_sched_class)
9147 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9148 struct task_struct *tsk, bool threadgroup)
9150 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
9154 struct task_struct *c;
9156 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
9157 retval = cpu_cgroup_can_attach_task(cgrp, c);
9169 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9170 struct cgroup *old_cont, struct task_struct *tsk,
9173 sched_move_task(tsk);
9175 struct task_struct *c;
9177 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
9185 cpu_cgroup_exit(struct cgroup_subsys *ss, struct task_struct *task)
9188 * cgroup_exit() is called in the copy_process() failure path.
9189 * Ignore this case since the task hasn't ran yet, this avoids
9190 * trying to poke a half freed task state from generic code.
9192 if (!(task->flags & PF_EXITING))
9195 sched_move_task(task);
9198 #ifdef CONFIG_FAIR_GROUP_SCHED
9199 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9202 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9205 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9207 struct task_group *tg = cgroup_tg(cgrp);
9209 return (u64) tg->shares;
9211 #endif /* CONFIG_FAIR_GROUP_SCHED */
9213 #ifdef CONFIG_RT_GROUP_SCHED
9214 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9217 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9220 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9222 return sched_group_rt_runtime(cgroup_tg(cgrp));
9225 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9228 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9231 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9233 return sched_group_rt_period(cgroup_tg(cgrp));
9235 #endif /* CONFIG_RT_GROUP_SCHED */
9237 static struct cftype cpu_files[] = {
9238 #ifdef CONFIG_FAIR_GROUP_SCHED
9241 .read_u64 = cpu_shares_read_u64,
9242 .write_u64 = cpu_shares_write_u64,
9245 #ifdef CONFIG_RT_GROUP_SCHED
9247 .name = "rt_runtime_us",
9248 .read_s64 = cpu_rt_runtime_read,
9249 .write_s64 = cpu_rt_runtime_write,
9252 .name = "rt_period_us",
9253 .read_u64 = cpu_rt_period_read_uint,
9254 .write_u64 = cpu_rt_period_write_uint,
9259 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9261 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9264 struct cgroup_subsys cpu_cgroup_subsys = {
9266 .create = cpu_cgroup_create,
9267 .destroy = cpu_cgroup_destroy,
9268 .can_attach = cpu_cgroup_can_attach,
9269 .attach = cpu_cgroup_attach,
9270 .exit = cpu_cgroup_exit,
9271 .populate = cpu_cgroup_populate,
9272 .subsys_id = cpu_cgroup_subsys_id,
9276 #endif /* CONFIG_CGROUP_SCHED */
9278 #ifdef CONFIG_CGROUP_CPUACCT
9281 * CPU accounting code for task groups.
9283 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9284 * (balbir@in.ibm.com).
9287 /* track cpu usage of a group of tasks and its child groups */
9289 struct cgroup_subsys_state css;
9290 /* cpuusage holds pointer to a u64-type object on every cpu */
9291 u64 __percpu *cpuusage;
9292 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9293 struct cpuacct *parent;
9296 struct cgroup_subsys cpuacct_subsys;
9298 /* return cpu accounting group corresponding to this container */
9299 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9301 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9302 struct cpuacct, css);
9305 /* return cpu accounting group to which this task belongs */
9306 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9308 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9309 struct cpuacct, css);
9312 /* create a new cpu accounting group */
9313 static struct cgroup_subsys_state *cpuacct_create(
9314 struct cgroup_subsys *ss, struct cgroup *cgrp)
9316 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9322 ca->cpuusage = alloc_percpu(u64);
9326 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9327 if (percpu_counter_init(&ca->cpustat[i], 0))
9328 goto out_free_counters;
9331 ca->parent = cgroup_ca(cgrp->parent);
9337 percpu_counter_destroy(&ca->cpustat[i]);
9338 free_percpu(ca->cpuusage);
9342 return ERR_PTR(-ENOMEM);
9345 /* destroy an existing cpu accounting group */
9347 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9349 struct cpuacct *ca = cgroup_ca(cgrp);
9352 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9353 percpu_counter_destroy(&ca->cpustat[i]);
9354 free_percpu(ca->cpuusage);
9358 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9360 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9363 #ifndef CONFIG_64BIT
9365 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9367 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9369 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9377 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9379 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9381 #ifndef CONFIG_64BIT
9383 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9385 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9387 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9393 /* return total cpu usage (in nanoseconds) of a group */
9394 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9396 struct cpuacct *ca = cgroup_ca(cgrp);
9397 u64 totalcpuusage = 0;
9400 for_each_present_cpu(i)
9401 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9403 return totalcpuusage;
9406 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9409 struct cpuacct *ca = cgroup_ca(cgrp);
9418 for_each_present_cpu(i)
9419 cpuacct_cpuusage_write(ca, i, 0);
9425 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9428 struct cpuacct *ca = cgroup_ca(cgroup);
9432 for_each_present_cpu(i) {
9433 percpu = cpuacct_cpuusage_read(ca, i);
9434 seq_printf(m, "%llu ", (unsigned long long) percpu);
9436 seq_printf(m, "\n");
9440 static const char *cpuacct_stat_desc[] = {
9441 [CPUACCT_STAT_USER] = "user",
9442 [CPUACCT_STAT_SYSTEM] = "system",
9445 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9446 struct cgroup_map_cb *cb)
9448 struct cpuacct *ca = cgroup_ca(cgrp);
9451 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9452 s64 val = percpu_counter_read(&ca->cpustat[i]);
9453 val = cputime64_to_clock_t(val);
9454 cb->fill(cb, cpuacct_stat_desc[i], val);
9459 static struct cftype files[] = {
9462 .read_u64 = cpuusage_read,
9463 .write_u64 = cpuusage_write,
9466 .name = "usage_percpu",
9467 .read_seq_string = cpuacct_percpu_seq_read,
9471 .read_map = cpuacct_stats_show,
9475 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9477 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9481 * charge this task's execution time to its accounting group.
9483 * called with rq->lock held.
9485 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9490 if (unlikely(!cpuacct_subsys.active))
9493 cpu = task_cpu(tsk);
9499 for (; ca; ca = ca->parent) {
9500 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9501 *cpuusage += cputime;
9508 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9509 * in cputime_t units. As a result, cpuacct_update_stats calls
9510 * percpu_counter_add with values large enough to always overflow the
9511 * per cpu batch limit causing bad SMP scalability.
9513 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9514 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9515 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9518 #define CPUACCT_BATCH \
9519 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9521 #define CPUACCT_BATCH 0
9525 * Charge the system/user time to the task's accounting group.
9527 static void cpuacct_update_stats(struct task_struct *tsk,
9528 enum cpuacct_stat_index idx, cputime_t val)
9531 int batch = CPUACCT_BATCH;
9533 if (unlikely(!cpuacct_subsys.active))
9540 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9546 struct cgroup_subsys cpuacct_subsys = {
9548 .create = cpuacct_create,
9549 .destroy = cpuacct_destroy,
9550 .populate = cpuacct_populate,
9551 .subsys_id = cpuacct_subsys_id,
9553 #endif /* CONFIG_CGROUP_CPUACCT */
9557 void synchronize_sched_expedited(void)
9561 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9563 #else /* #ifndef CONFIG_SMP */
9565 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
9567 static int synchronize_sched_expedited_cpu_stop(void *data)
9570 * There must be a full memory barrier on each affected CPU
9571 * between the time that try_stop_cpus() is called and the
9572 * time that it returns.
9574 * In the current initial implementation of cpu_stop, the
9575 * above condition is already met when the control reaches
9576 * this point and the following smp_mb() is not strictly
9577 * necessary. Do smp_mb() anyway for documentation and
9578 * robustness against future implementation changes.
9580 smp_mb(); /* See above comment block. */
9585 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9586 * approach to force grace period to end quickly. This consumes
9587 * significant time on all CPUs, and is thus not recommended for
9588 * any sort of common-case code.
9590 * Note that it is illegal to call this function while holding any
9591 * lock that is acquired by a CPU-hotplug notifier. Failing to
9592 * observe this restriction will result in deadlock.
9594 void synchronize_sched_expedited(void)
9596 int snap, trycount = 0;
9598 smp_mb(); /* ensure prior mod happens before capturing snap. */
9599 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
9601 while (try_stop_cpus(cpu_online_mask,
9602 synchronize_sched_expedited_cpu_stop,
9605 if (trycount++ < 10)
9606 udelay(trycount * num_online_cpus());
9608 synchronize_sched();
9611 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
9612 smp_mb(); /* ensure test happens before caller kfree */
9617 atomic_inc(&synchronize_sched_expedited_count);
9618 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9621 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9623 #endif /* #else #ifndef CONFIG_SMP */