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;
430 struct cpupri cpupri;
435 * By default the system creates a single root-domain with all cpus as
436 * members (mimicking the global state we have today).
438 static struct root_domain def_root_domain;
443 * This is the main, per-CPU runqueue data structure.
445 * Locking rule: those places that want to lock multiple runqueues
446 * (such as the load balancing or the thread migration code), lock
447 * acquire operations must be ordered by ascending &runqueue.
454 * nr_running and cpu_load should be in the same cacheline because
455 * remote CPUs use both these fields when doing load calculation.
457 unsigned long nr_running;
458 #define CPU_LOAD_IDX_MAX 5
459 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
460 unsigned long last_load_update_tick;
463 unsigned char nohz_balance_kick;
465 unsigned int skip_clock_update;
467 /* capture load from *all* tasks on this cpu: */
468 struct load_weight load;
469 unsigned long nr_load_updates;
475 #ifdef CONFIG_FAIR_GROUP_SCHED
476 /* list of leaf cfs_rq on this cpu: */
477 struct list_head leaf_cfs_rq_list;
479 #ifdef CONFIG_RT_GROUP_SCHED
480 struct list_head leaf_rt_rq_list;
484 * This is part of a global counter where only the total sum
485 * over all CPUs matters. A task can increase this counter on
486 * one CPU and if it got migrated afterwards it may decrease
487 * it on another CPU. Always updated under the runqueue lock:
489 unsigned long nr_uninterruptible;
491 struct task_struct *curr, *idle;
492 unsigned long next_balance;
493 struct mm_struct *prev_mm;
500 struct root_domain *rd;
501 struct sched_domain *sd;
503 unsigned long cpu_power;
505 unsigned char idle_at_tick;
506 /* For active balancing */
510 struct cpu_stop_work active_balance_work;
511 /* cpu of this runqueue: */
515 unsigned long avg_load_per_task;
523 /* calc_load related fields */
524 unsigned long calc_load_update;
525 long calc_load_active;
527 #ifdef CONFIG_SCHED_HRTICK
529 int hrtick_csd_pending;
530 struct call_single_data hrtick_csd;
532 struct hrtimer hrtick_timer;
535 #ifdef CONFIG_SCHEDSTATS
537 struct sched_info rq_sched_info;
538 unsigned long long rq_cpu_time;
539 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
541 /* sys_sched_yield() stats */
542 unsigned int yld_count;
544 /* schedule() stats */
545 unsigned int sched_switch;
546 unsigned int sched_count;
547 unsigned int sched_goidle;
549 /* try_to_wake_up() stats */
550 unsigned int ttwu_count;
551 unsigned int ttwu_local;
554 unsigned int bkl_count;
558 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
561 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
563 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
566 * A queue event has occurred, and we're going to schedule. In
567 * this case, we can save a useless back to back clock update.
569 if (test_tsk_need_resched(p))
570 rq->skip_clock_update = 1;
573 static inline int cpu_of(struct rq *rq)
582 #define rcu_dereference_check_sched_domain(p) \
583 rcu_dereference_check((p), \
584 rcu_read_lock_sched_held() || \
585 lockdep_is_held(&sched_domains_mutex))
588 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
589 * See detach_destroy_domains: synchronize_sched for details.
591 * The domain tree of any CPU may only be accessed from within
592 * preempt-disabled sections.
594 #define for_each_domain(cpu, __sd) \
595 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
597 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
598 #define this_rq() (&__get_cpu_var(runqueues))
599 #define task_rq(p) cpu_rq(task_cpu(p))
600 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
601 #define raw_rq() (&__raw_get_cpu_var(runqueues))
603 #ifdef CONFIG_CGROUP_SCHED
606 * Return the group to which this tasks belongs.
608 * We use task_subsys_state_check() and extend the RCU verification
609 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
610 * holds that lock for each task it moves into the cgroup. Therefore
611 * by holding that lock, we pin the task to the current cgroup.
613 static inline struct task_group *task_group(struct task_struct *p)
615 struct cgroup_subsys_state *css;
617 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
618 lockdep_is_held(&task_rq(p)->lock));
619 return container_of(css, struct task_group, css);
622 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
623 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
625 #ifdef CONFIG_FAIR_GROUP_SCHED
626 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
627 p->se.parent = task_group(p)->se[cpu];
630 #ifdef CONFIG_RT_GROUP_SCHED
631 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
632 p->rt.parent = task_group(p)->rt_se[cpu];
636 #else /* CONFIG_CGROUP_SCHED */
638 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
639 static inline struct task_group *task_group(struct task_struct *p)
644 #endif /* CONFIG_CGROUP_SCHED */
646 inline void update_rq_clock(struct rq *rq)
648 if (!rq->skip_clock_update)
649 rq->clock = sched_clock_cpu(cpu_of(rq));
653 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
655 #ifdef CONFIG_SCHED_DEBUG
656 # define const_debug __read_mostly
658 # define const_debug static const
663 * @cpu: the processor in question.
665 * Returns true if the current cpu runqueue is locked.
666 * This interface allows printk to be called with the runqueue lock
667 * held and know whether or not it is OK to wake up the klogd.
669 int runqueue_is_locked(int cpu)
671 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
675 * Debugging: various feature bits
678 #define SCHED_FEAT(name, enabled) \
679 __SCHED_FEAT_##name ,
682 #include "sched_features.h"
687 #define SCHED_FEAT(name, enabled) \
688 (1UL << __SCHED_FEAT_##name) * enabled |
690 const_debug unsigned int sysctl_sched_features =
691 #include "sched_features.h"
696 #ifdef CONFIG_SCHED_DEBUG
697 #define SCHED_FEAT(name, enabled) \
700 static __read_mostly char *sched_feat_names[] = {
701 #include "sched_features.h"
707 static int sched_feat_show(struct seq_file *m, void *v)
711 for (i = 0; sched_feat_names[i]; i++) {
712 if (!(sysctl_sched_features & (1UL << i)))
714 seq_printf(m, "%s ", sched_feat_names[i]);
722 sched_feat_write(struct file *filp, const char __user *ubuf,
723 size_t cnt, loff_t *ppos)
733 if (copy_from_user(&buf, ubuf, cnt))
738 if (strncmp(buf, "NO_", 3) == 0) {
743 for (i = 0; sched_feat_names[i]; i++) {
744 int len = strlen(sched_feat_names[i]);
746 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
748 sysctl_sched_features &= ~(1UL << i);
750 sysctl_sched_features |= (1UL << i);
755 if (!sched_feat_names[i])
763 static int sched_feat_open(struct inode *inode, struct file *filp)
765 return single_open(filp, sched_feat_show, NULL);
768 static const struct file_operations sched_feat_fops = {
769 .open = sched_feat_open,
770 .write = sched_feat_write,
773 .release = single_release,
776 static __init int sched_init_debug(void)
778 debugfs_create_file("sched_features", 0644, NULL, NULL,
783 late_initcall(sched_init_debug);
787 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
790 * Number of tasks to iterate in a single balance run.
791 * Limited because this is done with IRQs disabled.
793 const_debug unsigned int sysctl_sched_nr_migrate = 32;
796 * ratelimit for updating the group shares.
799 unsigned int sysctl_sched_shares_ratelimit = 250000;
800 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
803 * Inject some fuzzyness into changing the per-cpu group shares
804 * this avoids remote rq-locks at the expense of fairness.
807 unsigned int sysctl_sched_shares_thresh = 4;
810 * period over which we average the RT time consumption, measured
815 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
818 * period over which we measure -rt task cpu usage in us.
821 unsigned int sysctl_sched_rt_period = 1000000;
823 static __read_mostly int scheduler_running;
826 * part of the period that we allow rt tasks to run in us.
829 int sysctl_sched_rt_runtime = 950000;
831 static inline u64 global_rt_period(void)
833 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
836 static inline u64 global_rt_runtime(void)
838 if (sysctl_sched_rt_runtime < 0)
841 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
844 #ifndef prepare_arch_switch
845 # define prepare_arch_switch(next) do { } while (0)
847 #ifndef finish_arch_switch
848 # define finish_arch_switch(prev) do { } while (0)
851 static inline int task_current(struct rq *rq, struct task_struct *p)
853 return rq->curr == p;
856 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
857 static inline int task_running(struct rq *rq, struct task_struct *p)
859 return task_current(rq, p);
862 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
866 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
868 #ifdef CONFIG_DEBUG_SPINLOCK
869 /* this is a valid case when another task releases the spinlock */
870 rq->lock.owner = current;
873 * If we are tracking spinlock dependencies then we have to
874 * fix up the runqueue lock - which gets 'carried over' from
877 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
879 raw_spin_unlock_irq(&rq->lock);
882 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
883 static inline int task_running(struct rq *rq, struct task_struct *p)
888 return task_current(rq, p);
892 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
896 * We can optimise this out completely for !SMP, because the
897 * SMP rebalancing from interrupt is the only thing that cares
902 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
903 raw_spin_unlock_irq(&rq->lock);
905 raw_spin_unlock(&rq->lock);
909 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
913 * After ->oncpu is cleared, the task can be moved to a different CPU.
914 * We must ensure this doesn't happen until the switch is completely
920 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
924 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
927 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
930 static inline int task_is_waking(struct task_struct *p)
932 return unlikely(p->state == TASK_WAKING);
936 * __task_rq_lock - lock the runqueue a given task resides on.
937 * Must be called interrupts disabled.
939 static inline struct rq *__task_rq_lock(struct task_struct *p)
946 raw_spin_lock(&rq->lock);
947 if (likely(rq == task_rq(p)))
949 raw_spin_unlock(&rq->lock);
954 * task_rq_lock - lock the runqueue a given task resides on and disable
955 * interrupts. Note the ordering: we can safely lookup the task_rq without
956 * explicitly disabling preemption.
958 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
964 local_irq_save(*flags);
966 raw_spin_lock(&rq->lock);
967 if (likely(rq == task_rq(p)))
969 raw_spin_unlock_irqrestore(&rq->lock, *flags);
973 static void __task_rq_unlock(struct rq *rq)
976 raw_spin_unlock(&rq->lock);
979 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
982 raw_spin_unlock_irqrestore(&rq->lock, *flags);
986 * this_rq_lock - lock this runqueue and disable interrupts.
988 static struct rq *this_rq_lock(void)
995 raw_spin_lock(&rq->lock);
1000 #ifdef CONFIG_SCHED_HRTICK
1002 * Use HR-timers to deliver accurate preemption points.
1004 * Its all a bit involved since we cannot program an hrt while holding the
1005 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1008 * When we get rescheduled we reprogram the hrtick_timer outside of the
1014 * - enabled by features
1015 * - hrtimer is actually high res
1017 static inline int hrtick_enabled(struct rq *rq)
1019 if (!sched_feat(HRTICK))
1021 if (!cpu_active(cpu_of(rq)))
1023 return hrtimer_is_hres_active(&rq->hrtick_timer);
1026 static void hrtick_clear(struct rq *rq)
1028 if (hrtimer_active(&rq->hrtick_timer))
1029 hrtimer_cancel(&rq->hrtick_timer);
1033 * High-resolution timer tick.
1034 * Runs from hardirq context with interrupts disabled.
1036 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1038 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1040 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1042 raw_spin_lock(&rq->lock);
1043 update_rq_clock(rq);
1044 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1045 raw_spin_unlock(&rq->lock);
1047 return HRTIMER_NORESTART;
1052 * called from hardirq (IPI) context
1054 static void __hrtick_start(void *arg)
1056 struct rq *rq = arg;
1058 raw_spin_lock(&rq->lock);
1059 hrtimer_restart(&rq->hrtick_timer);
1060 rq->hrtick_csd_pending = 0;
1061 raw_spin_unlock(&rq->lock);
1065 * Called to set the hrtick timer state.
1067 * called with rq->lock held and irqs disabled
1069 static void hrtick_start(struct rq *rq, u64 delay)
1071 struct hrtimer *timer = &rq->hrtick_timer;
1072 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1074 hrtimer_set_expires(timer, time);
1076 if (rq == this_rq()) {
1077 hrtimer_restart(timer);
1078 } else if (!rq->hrtick_csd_pending) {
1079 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1080 rq->hrtick_csd_pending = 1;
1085 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1087 int cpu = (int)(long)hcpu;
1090 case CPU_UP_CANCELED:
1091 case CPU_UP_CANCELED_FROZEN:
1092 case CPU_DOWN_PREPARE:
1093 case CPU_DOWN_PREPARE_FROZEN:
1095 case CPU_DEAD_FROZEN:
1096 hrtick_clear(cpu_rq(cpu));
1103 static __init void init_hrtick(void)
1105 hotcpu_notifier(hotplug_hrtick, 0);
1109 * Called to set the hrtick timer state.
1111 * called with rq->lock held and irqs disabled
1113 static void hrtick_start(struct rq *rq, u64 delay)
1115 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1116 HRTIMER_MODE_REL_PINNED, 0);
1119 static inline void init_hrtick(void)
1122 #endif /* CONFIG_SMP */
1124 static void init_rq_hrtick(struct rq *rq)
1127 rq->hrtick_csd_pending = 0;
1129 rq->hrtick_csd.flags = 0;
1130 rq->hrtick_csd.func = __hrtick_start;
1131 rq->hrtick_csd.info = rq;
1134 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1135 rq->hrtick_timer.function = hrtick;
1137 #else /* CONFIG_SCHED_HRTICK */
1138 static inline void hrtick_clear(struct rq *rq)
1142 static inline void init_rq_hrtick(struct rq *rq)
1146 static inline void init_hrtick(void)
1149 #endif /* CONFIG_SCHED_HRTICK */
1152 * resched_task - mark a task 'to be rescheduled now'.
1154 * On UP this means the setting of the need_resched flag, on SMP it
1155 * might also involve a cross-CPU call to trigger the scheduler on
1160 #ifndef tsk_is_polling
1161 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1164 static void resched_task(struct task_struct *p)
1168 assert_raw_spin_locked(&task_rq(p)->lock);
1170 if (test_tsk_need_resched(p))
1173 set_tsk_need_resched(p);
1176 if (cpu == smp_processor_id())
1179 /* NEED_RESCHED must be visible before we test polling */
1181 if (!tsk_is_polling(p))
1182 smp_send_reschedule(cpu);
1185 static void resched_cpu(int cpu)
1187 struct rq *rq = cpu_rq(cpu);
1188 unsigned long flags;
1190 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1192 resched_task(cpu_curr(cpu));
1193 raw_spin_unlock_irqrestore(&rq->lock, flags);
1198 * In the semi idle case, use the nearest busy cpu for migrating timers
1199 * from an idle cpu. This is good for power-savings.
1201 * We don't do similar optimization for completely idle system, as
1202 * selecting an idle cpu will add more delays to the timers than intended
1203 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1205 int get_nohz_timer_target(void)
1207 int cpu = smp_processor_id();
1209 struct sched_domain *sd;
1211 for_each_domain(cpu, sd) {
1212 for_each_cpu(i, sched_domain_span(sd))
1219 * When add_timer_on() enqueues a timer into the timer wheel of an
1220 * idle CPU then this timer might expire before the next timer event
1221 * which is scheduled to wake up that CPU. In case of a completely
1222 * idle system the next event might even be infinite time into the
1223 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1224 * leaves the inner idle loop so the newly added timer is taken into
1225 * account when the CPU goes back to idle and evaluates the timer
1226 * wheel for the next timer event.
1228 void wake_up_idle_cpu(int cpu)
1230 struct rq *rq = cpu_rq(cpu);
1232 if (cpu == smp_processor_id())
1236 * This is safe, as this function is called with the timer
1237 * wheel base lock of (cpu) held. When the CPU is on the way
1238 * to idle and has not yet set rq->curr to idle then it will
1239 * be serialized on the timer wheel base lock and take the new
1240 * timer into account automatically.
1242 if (rq->curr != rq->idle)
1246 * We can set TIF_RESCHED on the idle task of the other CPU
1247 * lockless. The worst case is that the other CPU runs the
1248 * idle task through an additional NOOP schedule()
1250 set_tsk_need_resched(rq->idle);
1252 /* NEED_RESCHED must be visible before we test polling */
1254 if (!tsk_is_polling(rq->idle))
1255 smp_send_reschedule(cpu);
1258 int nohz_ratelimit(int cpu)
1260 struct rq *rq = cpu_rq(cpu);
1261 u64 diff = rq->clock - rq->nohz_stamp;
1263 rq->nohz_stamp = rq->clock;
1265 return diff < (NSEC_PER_SEC / HZ) >> 1;
1268 #endif /* CONFIG_NO_HZ */
1270 static u64 sched_avg_period(void)
1272 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1275 static void sched_avg_update(struct rq *rq)
1277 s64 period = sched_avg_period();
1279 while ((s64)(rq->clock - rq->age_stamp) > period) {
1281 * Inline assembly required to prevent the compiler
1282 * optimising this loop into a divmod call.
1283 * See __iter_div_u64_rem() for another example of this.
1285 asm("" : "+rm" (rq->age_stamp));
1286 rq->age_stamp += period;
1291 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1293 rq->rt_avg += rt_delta;
1294 sched_avg_update(rq);
1297 #else /* !CONFIG_SMP */
1298 static void resched_task(struct task_struct *p)
1300 assert_raw_spin_locked(&task_rq(p)->lock);
1301 set_tsk_need_resched(p);
1304 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1307 #endif /* CONFIG_SMP */
1309 #if BITS_PER_LONG == 32
1310 # define WMULT_CONST (~0UL)
1312 # define WMULT_CONST (1UL << 32)
1315 #define WMULT_SHIFT 32
1318 * Shift right and round:
1320 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1323 * delta *= weight / lw
1325 static unsigned long
1326 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1327 struct load_weight *lw)
1331 if (!lw->inv_weight) {
1332 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1335 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1339 tmp = (u64)delta_exec * weight;
1341 * Check whether we'd overflow the 64-bit multiplication:
1343 if (unlikely(tmp > WMULT_CONST))
1344 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1347 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1349 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1352 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1358 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1365 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1366 * of tasks with abnormal "nice" values across CPUs the contribution that
1367 * each task makes to its run queue's load is weighted according to its
1368 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1369 * scaled version of the new time slice allocation that they receive on time
1373 #define WEIGHT_IDLEPRIO 3
1374 #define WMULT_IDLEPRIO 1431655765
1377 * Nice levels are multiplicative, with a gentle 10% change for every
1378 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1379 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1380 * that remained on nice 0.
1382 * The "10% effect" is relative and cumulative: from _any_ nice level,
1383 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1384 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1385 * If a task goes up by ~10% and another task goes down by ~10% then
1386 * the relative distance between them is ~25%.)
1388 static const int prio_to_weight[40] = {
1389 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1390 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1391 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1392 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1393 /* 0 */ 1024, 820, 655, 526, 423,
1394 /* 5 */ 335, 272, 215, 172, 137,
1395 /* 10 */ 110, 87, 70, 56, 45,
1396 /* 15 */ 36, 29, 23, 18, 15,
1400 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1402 * In cases where the weight does not change often, we can use the
1403 * precalculated inverse to speed up arithmetics by turning divisions
1404 * into multiplications:
1406 static const u32 prio_to_wmult[40] = {
1407 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1408 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1409 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1410 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1411 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1412 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1413 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1414 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1417 /* Time spent by the tasks of the cpu accounting group executing in ... */
1418 enum cpuacct_stat_index {
1419 CPUACCT_STAT_USER, /* ... user mode */
1420 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1422 CPUACCT_STAT_NSTATS,
1425 #ifdef CONFIG_CGROUP_CPUACCT
1426 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1427 static void cpuacct_update_stats(struct task_struct *tsk,
1428 enum cpuacct_stat_index idx, cputime_t val);
1430 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1431 static inline void cpuacct_update_stats(struct task_struct *tsk,
1432 enum cpuacct_stat_index idx, cputime_t val) {}
1435 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1437 update_load_add(&rq->load, load);
1440 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1442 update_load_sub(&rq->load, load);
1445 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1446 typedef int (*tg_visitor)(struct task_group *, void *);
1449 * Iterate the full tree, calling @down when first entering a node and @up when
1450 * leaving it for the final time.
1452 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1454 struct task_group *parent, *child;
1458 parent = &root_task_group;
1460 ret = (*down)(parent, data);
1463 list_for_each_entry_rcu(child, &parent->children, siblings) {
1470 ret = (*up)(parent, data);
1475 parent = parent->parent;
1484 static int tg_nop(struct task_group *tg, void *data)
1491 /* Used instead of source_load when we know the type == 0 */
1492 static unsigned long weighted_cpuload(const int cpu)
1494 return cpu_rq(cpu)->load.weight;
1498 * Return a low guess at the load of a migration-source cpu weighted
1499 * according to the scheduling class and "nice" value.
1501 * We want to under-estimate the load of migration sources, to
1502 * balance conservatively.
1504 static unsigned long source_load(int cpu, int type)
1506 struct rq *rq = cpu_rq(cpu);
1507 unsigned long total = weighted_cpuload(cpu);
1509 if (type == 0 || !sched_feat(LB_BIAS))
1512 return min(rq->cpu_load[type-1], total);
1516 * Return a high guess at the load of a migration-target cpu weighted
1517 * according to the scheduling class and "nice" value.
1519 static unsigned long target_load(int cpu, int type)
1521 struct rq *rq = cpu_rq(cpu);
1522 unsigned long total = weighted_cpuload(cpu);
1524 if (type == 0 || !sched_feat(LB_BIAS))
1527 return max(rq->cpu_load[type-1], total);
1530 static unsigned long power_of(int cpu)
1532 return cpu_rq(cpu)->cpu_power;
1535 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1537 static unsigned long cpu_avg_load_per_task(int cpu)
1539 struct rq *rq = cpu_rq(cpu);
1540 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1543 rq->avg_load_per_task = rq->load.weight / nr_running;
1545 rq->avg_load_per_task = 0;
1547 return rq->avg_load_per_task;
1550 #ifdef CONFIG_FAIR_GROUP_SCHED
1552 static __read_mostly unsigned long __percpu *update_shares_data;
1554 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1557 * Calculate and set the cpu's group shares.
1559 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1560 unsigned long sd_shares,
1561 unsigned long sd_rq_weight,
1562 unsigned long *usd_rq_weight)
1564 unsigned long shares, rq_weight;
1567 rq_weight = usd_rq_weight[cpu];
1570 rq_weight = NICE_0_LOAD;
1574 * \Sum_j shares_j * rq_weight_i
1575 * shares_i = -----------------------------
1576 * \Sum_j rq_weight_j
1578 shares = (sd_shares * rq_weight) / sd_rq_weight;
1579 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1581 if (abs(shares - tg->se[cpu]->load.weight) >
1582 sysctl_sched_shares_thresh) {
1583 struct rq *rq = cpu_rq(cpu);
1584 unsigned long flags;
1586 raw_spin_lock_irqsave(&rq->lock, flags);
1587 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1588 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1589 __set_se_shares(tg->se[cpu], shares);
1590 raw_spin_unlock_irqrestore(&rq->lock, flags);
1595 * Re-compute the task group their per cpu shares over the given domain.
1596 * This needs to be done in a bottom-up fashion because the rq weight of a
1597 * parent group depends on the shares of its child groups.
1599 static int tg_shares_up(struct task_group *tg, void *data)
1601 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1602 unsigned long *usd_rq_weight;
1603 struct sched_domain *sd = data;
1604 unsigned long flags;
1610 local_irq_save(flags);
1611 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1613 for_each_cpu(i, sched_domain_span(sd)) {
1614 weight = tg->cfs_rq[i]->load.weight;
1615 usd_rq_weight[i] = weight;
1617 rq_weight += weight;
1619 * If there are currently no tasks on the cpu pretend there
1620 * is one of average load so that when a new task gets to
1621 * run here it will not get delayed by group starvation.
1624 weight = NICE_0_LOAD;
1626 sum_weight += weight;
1627 shares += tg->cfs_rq[i]->shares;
1631 rq_weight = sum_weight;
1633 if ((!shares && rq_weight) || shares > tg->shares)
1634 shares = tg->shares;
1636 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1637 shares = tg->shares;
1639 for_each_cpu(i, sched_domain_span(sd))
1640 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1642 local_irq_restore(flags);
1648 * Compute the cpu's hierarchical load factor for each task group.
1649 * This needs to be done in a top-down fashion because the load of a child
1650 * group is a fraction of its parents load.
1652 static int tg_load_down(struct task_group *tg, void *data)
1655 long cpu = (long)data;
1658 load = cpu_rq(cpu)->load.weight;
1660 load = tg->parent->cfs_rq[cpu]->h_load;
1661 load *= tg->cfs_rq[cpu]->shares;
1662 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1665 tg->cfs_rq[cpu]->h_load = load;
1670 static void update_shares(struct sched_domain *sd)
1675 if (root_task_group_empty())
1678 now = local_clock();
1679 elapsed = now - sd->last_update;
1681 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1682 sd->last_update = now;
1683 walk_tg_tree(tg_nop, tg_shares_up, sd);
1687 static void update_h_load(long cpu)
1689 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1694 static inline void update_shares(struct sched_domain *sd)
1700 #ifdef CONFIG_PREEMPT
1702 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1705 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1706 * way at the expense of forcing extra atomic operations in all
1707 * invocations. This assures that the double_lock is acquired using the
1708 * same underlying policy as the spinlock_t on this architecture, which
1709 * reduces latency compared to the unfair variant below. However, it
1710 * also adds more overhead and therefore may reduce throughput.
1712 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1713 __releases(this_rq->lock)
1714 __acquires(busiest->lock)
1715 __acquires(this_rq->lock)
1717 raw_spin_unlock(&this_rq->lock);
1718 double_rq_lock(this_rq, busiest);
1725 * Unfair double_lock_balance: Optimizes throughput at the expense of
1726 * latency by eliminating extra atomic operations when the locks are
1727 * already in proper order on entry. This favors lower cpu-ids and will
1728 * grant the double lock to lower cpus over higher ids under contention,
1729 * regardless of entry order into the function.
1731 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1732 __releases(this_rq->lock)
1733 __acquires(busiest->lock)
1734 __acquires(this_rq->lock)
1738 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1739 if (busiest < this_rq) {
1740 raw_spin_unlock(&this_rq->lock);
1741 raw_spin_lock(&busiest->lock);
1742 raw_spin_lock_nested(&this_rq->lock,
1743 SINGLE_DEPTH_NESTING);
1746 raw_spin_lock_nested(&busiest->lock,
1747 SINGLE_DEPTH_NESTING);
1752 #endif /* CONFIG_PREEMPT */
1755 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1757 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1759 if (unlikely(!irqs_disabled())) {
1760 /* printk() doesn't work good under rq->lock */
1761 raw_spin_unlock(&this_rq->lock);
1765 return _double_lock_balance(this_rq, busiest);
1768 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1769 __releases(busiest->lock)
1771 raw_spin_unlock(&busiest->lock);
1772 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1776 * double_rq_lock - safely lock two runqueues
1778 * Note this does not disable interrupts like task_rq_lock,
1779 * you need to do so manually before calling.
1781 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1782 __acquires(rq1->lock)
1783 __acquires(rq2->lock)
1785 BUG_ON(!irqs_disabled());
1787 raw_spin_lock(&rq1->lock);
1788 __acquire(rq2->lock); /* Fake it out ;) */
1791 raw_spin_lock(&rq1->lock);
1792 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1794 raw_spin_lock(&rq2->lock);
1795 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1801 * double_rq_unlock - safely unlock two runqueues
1803 * Note this does not restore interrupts like task_rq_unlock,
1804 * you need to do so manually after calling.
1806 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1807 __releases(rq1->lock)
1808 __releases(rq2->lock)
1810 raw_spin_unlock(&rq1->lock);
1812 raw_spin_unlock(&rq2->lock);
1814 __release(rq2->lock);
1819 #ifdef CONFIG_FAIR_GROUP_SCHED
1820 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1823 cfs_rq->shares = shares;
1828 static void calc_load_account_idle(struct rq *this_rq);
1829 static void update_sysctl(void);
1830 static int get_update_sysctl_factor(void);
1831 static void update_cpu_load(struct rq *this_rq);
1833 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1835 set_task_rq(p, cpu);
1838 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1839 * successfuly executed on another CPU. We must ensure that updates of
1840 * per-task data have been completed by this moment.
1843 task_thread_info(p)->cpu = cpu;
1847 static const struct sched_class rt_sched_class;
1849 #define sched_class_highest (&rt_sched_class)
1850 #define for_each_class(class) \
1851 for (class = sched_class_highest; class; class = class->next)
1853 #include "sched_stats.h"
1855 static void inc_nr_running(struct rq *rq)
1860 static void dec_nr_running(struct rq *rq)
1865 static void set_load_weight(struct task_struct *p)
1867 if (task_has_rt_policy(p)) {
1868 p->se.load.weight = 0;
1869 p->se.load.inv_weight = WMULT_CONST;
1874 * SCHED_IDLE tasks get minimal weight:
1876 if (p->policy == SCHED_IDLE) {
1877 p->se.load.weight = WEIGHT_IDLEPRIO;
1878 p->se.load.inv_weight = WMULT_IDLEPRIO;
1882 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1883 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1886 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1888 update_rq_clock(rq);
1889 sched_info_queued(p);
1890 p->sched_class->enqueue_task(rq, p, flags);
1894 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1896 update_rq_clock(rq);
1897 sched_info_dequeued(p);
1898 p->sched_class->dequeue_task(rq, p, flags);
1903 * activate_task - move a task to the runqueue.
1905 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1907 if (task_contributes_to_load(p))
1908 rq->nr_uninterruptible--;
1910 enqueue_task(rq, p, flags);
1915 * deactivate_task - remove a task from the runqueue.
1917 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1919 if (task_contributes_to_load(p))
1920 rq->nr_uninterruptible++;
1922 dequeue_task(rq, p, flags);
1926 #include "sched_idletask.c"
1927 #include "sched_fair.c"
1928 #include "sched_rt.c"
1929 #ifdef CONFIG_SCHED_DEBUG
1930 # include "sched_debug.c"
1934 * __normal_prio - return the priority that is based on the static prio
1936 static inline int __normal_prio(struct task_struct *p)
1938 return p->static_prio;
1942 * Calculate the expected normal priority: i.e. priority
1943 * without taking RT-inheritance into account. Might be
1944 * boosted by interactivity modifiers. Changes upon fork,
1945 * setprio syscalls, and whenever the interactivity
1946 * estimator recalculates.
1948 static inline int normal_prio(struct task_struct *p)
1952 if (task_has_rt_policy(p))
1953 prio = MAX_RT_PRIO-1 - p->rt_priority;
1955 prio = __normal_prio(p);
1960 * Calculate the current priority, i.e. the priority
1961 * taken into account by the scheduler. This value might
1962 * be boosted by RT tasks, or might be boosted by
1963 * interactivity modifiers. Will be RT if the task got
1964 * RT-boosted. If not then it returns p->normal_prio.
1966 static int effective_prio(struct task_struct *p)
1968 p->normal_prio = normal_prio(p);
1970 * If we are RT tasks or we were boosted to RT priority,
1971 * keep the priority unchanged. Otherwise, update priority
1972 * to the normal priority:
1974 if (!rt_prio(p->prio))
1975 return p->normal_prio;
1980 * task_curr - is this task currently executing on a CPU?
1981 * @p: the task in question.
1983 inline int task_curr(const struct task_struct *p)
1985 return cpu_curr(task_cpu(p)) == p;
1988 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1989 const struct sched_class *prev_class,
1990 int oldprio, int running)
1992 if (prev_class != p->sched_class) {
1993 if (prev_class->switched_from)
1994 prev_class->switched_from(rq, p, running);
1995 p->sched_class->switched_to(rq, p, running);
1997 p->sched_class->prio_changed(rq, p, oldprio, running);
2002 * Is this task likely cache-hot:
2005 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2009 if (p->sched_class != &fair_sched_class)
2013 * Buddy candidates are cache hot:
2015 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2016 (&p->se == cfs_rq_of(&p->se)->next ||
2017 &p->se == cfs_rq_of(&p->se)->last))
2020 if (sysctl_sched_migration_cost == -1)
2022 if (sysctl_sched_migration_cost == 0)
2025 delta = now - p->se.exec_start;
2027 return delta < (s64)sysctl_sched_migration_cost;
2030 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2032 #ifdef CONFIG_SCHED_DEBUG
2034 * We should never call set_task_cpu() on a blocked task,
2035 * ttwu() will sort out the placement.
2037 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2038 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2041 trace_sched_migrate_task(p, new_cpu);
2043 if (task_cpu(p) != new_cpu) {
2044 p->se.nr_migrations++;
2045 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2048 __set_task_cpu(p, new_cpu);
2051 struct migration_arg {
2052 struct task_struct *task;
2056 static int migration_cpu_stop(void *data);
2059 * The task's runqueue lock must be held.
2060 * Returns true if you have to wait for migration thread.
2062 static bool migrate_task(struct task_struct *p, int dest_cpu)
2064 struct rq *rq = task_rq(p);
2067 * If the task is not on a runqueue (and not running), then
2068 * the next wake-up will properly place the task.
2070 return p->se.on_rq || task_running(rq, p);
2074 * wait_task_inactive - wait for a thread to unschedule.
2076 * If @match_state is nonzero, it's the @p->state value just checked and
2077 * not expected to change. If it changes, i.e. @p might have woken up,
2078 * then return zero. When we succeed in waiting for @p to be off its CPU,
2079 * we return a positive number (its total switch count). If a second call
2080 * a short while later returns the same number, the caller can be sure that
2081 * @p has remained unscheduled the whole time.
2083 * The caller must ensure that the task *will* unschedule sometime soon,
2084 * else this function might spin for a *long* time. This function can't
2085 * be called with interrupts off, or it may introduce deadlock with
2086 * smp_call_function() if an IPI is sent by the same process we are
2087 * waiting to become inactive.
2089 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2091 unsigned long flags;
2098 * We do the initial early heuristics without holding
2099 * any task-queue locks at all. We'll only try to get
2100 * the runqueue lock when things look like they will
2106 * If the task is actively running on another CPU
2107 * still, just relax and busy-wait without holding
2110 * NOTE! Since we don't hold any locks, it's not
2111 * even sure that "rq" stays as the right runqueue!
2112 * But we don't care, since "task_running()" will
2113 * return false if the runqueue has changed and p
2114 * is actually now running somewhere else!
2116 while (task_running(rq, p)) {
2117 if (match_state && unlikely(p->state != match_state))
2123 * Ok, time to look more closely! We need the rq
2124 * lock now, to be *sure*. If we're wrong, we'll
2125 * just go back and repeat.
2127 rq = task_rq_lock(p, &flags);
2128 trace_sched_wait_task(p);
2129 running = task_running(rq, p);
2130 on_rq = p->se.on_rq;
2132 if (!match_state || p->state == match_state)
2133 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2134 task_rq_unlock(rq, &flags);
2137 * If it changed from the expected state, bail out now.
2139 if (unlikely(!ncsw))
2143 * Was it really running after all now that we
2144 * checked with the proper locks actually held?
2146 * Oops. Go back and try again..
2148 if (unlikely(running)) {
2154 * It's not enough that it's not actively running,
2155 * it must be off the runqueue _entirely_, and not
2158 * So if it was still runnable (but just not actively
2159 * running right now), it's preempted, and we should
2160 * yield - it could be a while.
2162 if (unlikely(on_rq)) {
2163 schedule_timeout_uninterruptible(1);
2168 * Ahh, all good. It wasn't running, and it wasn't
2169 * runnable, which means that it will never become
2170 * running in the future either. We're all done!
2179 * kick_process - kick a running thread to enter/exit the kernel
2180 * @p: the to-be-kicked thread
2182 * Cause a process which is running on another CPU to enter
2183 * kernel-mode, without any delay. (to get signals handled.)
2185 * NOTE: this function doesnt have to take the runqueue lock,
2186 * because all it wants to ensure is that the remote task enters
2187 * the kernel. If the IPI races and the task has been migrated
2188 * to another CPU then no harm is done and the purpose has been
2191 void kick_process(struct task_struct *p)
2197 if ((cpu != smp_processor_id()) && task_curr(p))
2198 smp_send_reschedule(cpu);
2201 EXPORT_SYMBOL_GPL(kick_process);
2202 #endif /* CONFIG_SMP */
2205 * task_oncpu_function_call - call a function on the cpu on which a task runs
2206 * @p: the task to evaluate
2207 * @func: the function to be called
2208 * @info: the function call argument
2210 * Calls the function @func when the task is currently running. This might
2211 * be on the current CPU, which just calls the function directly
2213 void task_oncpu_function_call(struct task_struct *p,
2214 void (*func) (void *info), void *info)
2221 smp_call_function_single(cpu, func, info, 1);
2227 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2229 static int select_fallback_rq(int cpu, struct task_struct *p)
2232 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2234 /* Look for allowed, online CPU in same node. */
2235 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2236 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2239 /* Any allowed, online CPU? */
2240 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2241 if (dest_cpu < nr_cpu_ids)
2244 /* No more Mr. Nice Guy. */
2245 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2246 dest_cpu = cpuset_cpus_allowed_fallback(p);
2248 * Don't tell them about moving exiting tasks or
2249 * kernel threads (both mm NULL), since they never
2252 if (p->mm && printk_ratelimit()) {
2253 printk(KERN_INFO "process %d (%s) no "
2254 "longer affine to cpu%d\n",
2255 task_pid_nr(p), p->comm, cpu);
2263 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2266 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2268 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2271 * In order not to call set_task_cpu() on a blocking task we need
2272 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2275 * Since this is common to all placement strategies, this lives here.
2277 * [ this allows ->select_task() to simply return task_cpu(p) and
2278 * not worry about this generic constraint ]
2280 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2282 cpu = select_fallback_rq(task_cpu(p), p);
2287 static void update_avg(u64 *avg, u64 sample)
2289 s64 diff = sample - *avg;
2294 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2295 bool is_sync, bool is_migrate, bool is_local,
2296 unsigned long en_flags)
2298 schedstat_inc(p, se.statistics.nr_wakeups);
2300 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2302 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2304 schedstat_inc(p, se.statistics.nr_wakeups_local);
2306 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2308 activate_task(rq, p, en_flags);
2311 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2312 int wake_flags, bool success)
2314 trace_sched_wakeup(p, success);
2315 check_preempt_curr(rq, p, wake_flags);
2317 p->state = TASK_RUNNING;
2319 if (p->sched_class->task_woken)
2320 p->sched_class->task_woken(rq, p);
2322 if (unlikely(rq->idle_stamp)) {
2323 u64 delta = rq->clock - rq->idle_stamp;
2324 u64 max = 2*sysctl_sched_migration_cost;
2329 update_avg(&rq->avg_idle, delta);
2333 /* if a worker is waking up, notify workqueue */
2334 if ((p->flags & PF_WQ_WORKER) && success)
2335 wq_worker_waking_up(p, cpu_of(rq));
2339 * try_to_wake_up - wake up a thread
2340 * @p: the thread to be awakened
2341 * @state: the mask of task states that can be woken
2342 * @wake_flags: wake modifier flags (WF_*)
2344 * Put it on the run-queue if it's not already there. The "current"
2345 * thread is always on the run-queue (except when the actual
2346 * re-schedule is in progress), and as such you're allowed to do
2347 * the simpler "current->state = TASK_RUNNING" to mark yourself
2348 * runnable without the overhead of this.
2350 * Returns %true if @p was woken up, %false if it was already running
2351 * or @state didn't match @p's state.
2353 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2356 int cpu, orig_cpu, this_cpu, success = 0;
2357 unsigned long flags;
2358 unsigned long en_flags = ENQUEUE_WAKEUP;
2361 this_cpu = get_cpu();
2364 rq = task_rq_lock(p, &flags);
2365 if (!(p->state & state))
2375 if (unlikely(task_running(rq, p)))
2379 * In order to handle concurrent wakeups and release the rq->lock
2380 * we put the task in TASK_WAKING state.
2382 * First fix up the nr_uninterruptible count:
2384 if (task_contributes_to_load(p)) {
2385 if (likely(cpu_online(orig_cpu)))
2386 rq->nr_uninterruptible--;
2388 this_rq()->nr_uninterruptible--;
2390 p->state = TASK_WAKING;
2392 if (p->sched_class->task_waking) {
2393 p->sched_class->task_waking(rq, p);
2394 en_flags |= ENQUEUE_WAKING;
2397 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2398 if (cpu != orig_cpu)
2399 set_task_cpu(p, cpu);
2400 __task_rq_unlock(rq);
2403 raw_spin_lock(&rq->lock);
2406 * We migrated the task without holding either rq->lock, however
2407 * since the task is not on the task list itself, nobody else
2408 * will try and migrate the task, hence the rq should match the
2409 * cpu we just moved it to.
2411 WARN_ON(task_cpu(p) != cpu);
2412 WARN_ON(p->state != TASK_WAKING);
2414 #ifdef CONFIG_SCHEDSTATS
2415 schedstat_inc(rq, ttwu_count);
2416 if (cpu == this_cpu)
2417 schedstat_inc(rq, ttwu_local);
2419 struct sched_domain *sd;
2420 for_each_domain(this_cpu, sd) {
2421 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2422 schedstat_inc(sd, ttwu_wake_remote);
2427 #endif /* CONFIG_SCHEDSTATS */
2430 #endif /* CONFIG_SMP */
2431 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2432 cpu == this_cpu, en_flags);
2435 ttwu_post_activation(p, rq, wake_flags, success);
2437 task_rq_unlock(rq, &flags);
2444 * try_to_wake_up_local - try to wake up a local task with rq lock held
2445 * @p: the thread to be awakened
2447 * Put @p on the run-queue if it's not alredy there. The caller must
2448 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2449 * the current task. this_rq() stays locked over invocation.
2451 static void try_to_wake_up_local(struct task_struct *p)
2453 struct rq *rq = task_rq(p);
2454 bool success = false;
2456 BUG_ON(rq != this_rq());
2457 BUG_ON(p == current);
2458 lockdep_assert_held(&rq->lock);
2460 if (!(p->state & TASK_NORMAL))
2464 if (likely(!task_running(rq, p))) {
2465 schedstat_inc(rq, ttwu_count);
2466 schedstat_inc(rq, ttwu_local);
2468 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2471 ttwu_post_activation(p, rq, 0, success);
2475 * wake_up_process - Wake up a specific process
2476 * @p: The process to be woken up.
2478 * Attempt to wake up the nominated process and move it to the set of runnable
2479 * processes. Returns 1 if the process was woken up, 0 if it was already
2482 * It may be assumed that this function implies a write memory barrier before
2483 * changing the task state if and only if any tasks are woken up.
2485 int wake_up_process(struct task_struct *p)
2487 return try_to_wake_up(p, TASK_ALL, 0);
2489 EXPORT_SYMBOL(wake_up_process);
2491 int wake_up_state(struct task_struct *p, unsigned int state)
2493 return try_to_wake_up(p, state, 0);
2497 * Perform scheduler related setup for a newly forked process p.
2498 * p is forked by current.
2500 * __sched_fork() is basic setup used by init_idle() too:
2502 static void __sched_fork(struct task_struct *p)
2504 p->se.exec_start = 0;
2505 p->se.sum_exec_runtime = 0;
2506 p->se.prev_sum_exec_runtime = 0;
2507 p->se.nr_migrations = 0;
2509 #ifdef CONFIG_SCHEDSTATS
2510 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2513 INIT_LIST_HEAD(&p->rt.run_list);
2515 INIT_LIST_HEAD(&p->se.group_node);
2517 #ifdef CONFIG_PREEMPT_NOTIFIERS
2518 INIT_HLIST_HEAD(&p->preempt_notifiers);
2523 * fork()/clone()-time setup:
2525 void sched_fork(struct task_struct *p, int clone_flags)
2527 int cpu = get_cpu();
2531 * We mark the process as running here. This guarantees that
2532 * nobody will actually run it, and a signal or other external
2533 * event cannot wake it up and insert it on the runqueue either.
2535 p->state = TASK_RUNNING;
2538 * Revert to default priority/policy on fork if requested.
2540 if (unlikely(p->sched_reset_on_fork)) {
2541 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2542 p->policy = SCHED_NORMAL;
2543 p->normal_prio = p->static_prio;
2546 if (PRIO_TO_NICE(p->static_prio) < 0) {
2547 p->static_prio = NICE_TO_PRIO(0);
2548 p->normal_prio = p->static_prio;
2553 * We don't need the reset flag anymore after the fork. It has
2554 * fulfilled its duty:
2556 p->sched_reset_on_fork = 0;
2560 * Make sure we do not leak PI boosting priority to the child.
2562 p->prio = current->normal_prio;
2564 if (!rt_prio(p->prio))
2565 p->sched_class = &fair_sched_class;
2567 if (p->sched_class->task_fork)
2568 p->sched_class->task_fork(p);
2571 * The child is not yet in the pid-hash so no cgroup attach races,
2572 * and the cgroup is pinned to this child due to cgroup_fork()
2573 * is ran before sched_fork().
2575 * Silence PROVE_RCU.
2578 set_task_cpu(p, cpu);
2581 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2582 if (likely(sched_info_on()))
2583 memset(&p->sched_info, 0, sizeof(p->sched_info));
2585 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2588 #ifdef CONFIG_PREEMPT
2589 /* Want to start with kernel preemption disabled. */
2590 task_thread_info(p)->preempt_count = 1;
2592 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2598 * wake_up_new_task - wake up a newly created task for the first time.
2600 * This function will do some initial scheduler statistics housekeeping
2601 * that must be done for every newly created context, then puts the task
2602 * on the runqueue and wakes it.
2604 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2606 unsigned long flags;
2608 int cpu __maybe_unused = get_cpu();
2611 rq = task_rq_lock(p, &flags);
2612 p->state = TASK_WAKING;
2615 * Fork balancing, do it here and not earlier because:
2616 * - cpus_allowed can change in the fork path
2617 * - any previously selected cpu might disappear through hotplug
2619 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2620 * without people poking at ->cpus_allowed.
2622 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2623 set_task_cpu(p, cpu);
2625 p->state = TASK_RUNNING;
2626 task_rq_unlock(rq, &flags);
2629 rq = task_rq_lock(p, &flags);
2630 activate_task(rq, p, 0);
2631 trace_sched_wakeup_new(p, 1);
2632 check_preempt_curr(rq, p, WF_FORK);
2634 if (p->sched_class->task_woken)
2635 p->sched_class->task_woken(rq, p);
2637 task_rq_unlock(rq, &flags);
2641 #ifdef CONFIG_PREEMPT_NOTIFIERS
2644 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2645 * @notifier: notifier struct to register
2647 void preempt_notifier_register(struct preempt_notifier *notifier)
2649 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2651 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2654 * preempt_notifier_unregister - no longer interested in preemption notifications
2655 * @notifier: notifier struct to unregister
2657 * This is safe to call from within a preemption notifier.
2659 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2661 hlist_del(¬ifier->link);
2663 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2665 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2667 struct preempt_notifier *notifier;
2668 struct hlist_node *node;
2670 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2671 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2675 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2676 struct task_struct *next)
2678 struct preempt_notifier *notifier;
2679 struct hlist_node *node;
2681 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2682 notifier->ops->sched_out(notifier, next);
2685 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2687 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2692 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2693 struct task_struct *next)
2697 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2700 * prepare_task_switch - prepare to switch tasks
2701 * @rq: the runqueue preparing to switch
2702 * @prev: the current task that is being switched out
2703 * @next: the task we are going to switch to.
2705 * This is called with the rq lock held and interrupts off. It must
2706 * be paired with a subsequent finish_task_switch after the context
2709 * prepare_task_switch sets up locking and calls architecture specific
2713 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2714 struct task_struct *next)
2716 fire_sched_out_preempt_notifiers(prev, next);
2717 prepare_lock_switch(rq, next);
2718 prepare_arch_switch(next);
2722 * finish_task_switch - clean up after a task-switch
2723 * @rq: runqueue associated with task-switch
2724 * @prev: the thread we just switched away from.
2726 * finish_task_switch must be called after the context switch, paired
2727 * with a prepare_task_switch call before the context switch.
2728 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2729 * and do any other architecture-specific cleanup actions.
2731 * Note that we may have delayed dropping an mm in context_switch(). If
2732 * so, we finish that here outside of the runqueue lock. (Doing it
2733 * with the lock held can cause deadlocks; see schedule() for
2736 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2737 __releases(rq->lock)
2739 struct mm_struct *mm = rq->prev_mm;
2745 * A task struct has one reference for the use as "current".
2746 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2747 * schedule one last time. The schedule call will never return, and
2748 * the scheduled task must drop that reference.
2749 * The test for TASK_DEAD must occur while the runqueue locks are
2750 * still held, otherwise prev could be scheduled on another cpu, die
2751 * there before we look at prev->state, and then the reference would
2753 * Manfred Spraul <manfred@colorfullife.com>
2755 prev_state = prev->state;
2756 finish_arch_switch(prev);
2757 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2758 local_irq_disable();
2759 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2760 perf_event_task_sched_in(current);
2761 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2763 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2764 finish_lock_switch(rq, prev);
2766 fire_sched_in_preempt_notifiers(current);
2769 if (unlikely(prev_state == TASK_DEAD)) {
2771 * Remove function-return probe instances associated with this
2772 * task and put them back on the free list.
2774 kprobe_flush_task(prev);
2775 put_task_struct(prev);
2781 /* assumes rq->lock is held */
2782 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2784 if (prev->sched_class->pre_schedule)
2785 prev->sched_class->pre_schedule(rq, prev);
2788 /* rq->lock is NOT held, but preemption is disabled */
2789 static inline void post_schedule(struct rq *rq)
2791 if (rq->post_schedule) {
2792 unsigned long flags;
2794 raw_spin_lock_irqsave(&rq->lock, flags);
2795 if (rq->curr->sched_class->post_schedule)
2796 rq->curr->sched_class->post_schedule(rq);
2797 raw_spin_unlock_irqrestore(&rq->lock, flags);
2799 rq->post_schedule = 0;
2805 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2809 static inline void post_schedule(struct rq *rq)
2816 * schedule_tail - first thing a freshly forked thread must call.
2817 * @prev: the thread we just switched away from.
2819 asmlinkage void schedule_tail(struct task_struct *prev)
2820 __releases(rq->lock)
2822 struct rq *rq = this_rq();
2824 finish_task_switch(rq, prev);
2827 * FIXME: do we need to worry about rq being invalidated by the
2832 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2833 /* In this case, finish_task_switch does not reenable preemption */
2836 if (current->set_child_tid)
2837 put_user(task_pid_vnr(current), current->set_child_tid);
2841 * context_switch - switch to the new MM and the new
2842 * thread's register state.
2845 context_switch(struct rq *rq, struct task_struct *prev,
2846 struct task_struct *next)
2848 struct mm_struct *mm, *oldmm;
2850 prepare_task_switch(rq, prev, next);
2851 trace_sched_switch(prev, next);
2853 oldmm = prev->active_mm;
2855 * For paravirt, this is coupled with an exit in switch_to to
2856 * combine the page table reload and the switch backend into
2859 arch_start_context_switch(prev);
2862 next->active_mm = oldmm;
2863 atomic_inc(&oldmm->mm_count);
2864 enter_lazy_tlb(oldmm, next);
2866 switch_mm(oldmm, mm, next);
2868 if (likely(!prev->mm)) {
2869 prev->active_mm = NULL;
2870 rq->prev_mm = oldmm;
2873 * Since the runqueue lock will be released by the next
2874 * task (which is an invalid locking op but in the case
2875 * of the scheduler it's an obvious special-case), so we
2876 * do an early lockdep release here:
2878 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2879 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2882 /* Here we just switch the register state and the stack. */
2883 switch_to(prev, next, prev);
2887 * this_rq must be evaluated again because prev may have moved
2888 * CPUs since it called schedule(), thus the 'rq' on its stack
2889 * frame will be invalid.
2891 finish_task_switch(this_rq(), prev);
2895 * nr_running, nr_uninterruptible and nr_context_switches:
2897 * externally visible scheduler statistics: current number of runnable
2898 * threads, current number of uninterruptible-sleeping threads, total
2899 * number of context switches performed since bootup.
2901 unsigned long nr_running(void)
2903 unsigned long i, sum = 0;
2905 for_each_online_cpu(i)
2906 sum += cpu_rq(i)->nr_running;
2911 unsigned long nr_uninterruptible(void)
2913 unsigned long i, sum = 0;
2915 for_each_possible_cpu(i)
2916 sum += cpu_rq(i)->nr_uninterruptible;
2919 * Since we read the counters lockless, it might be slightly
2920 * inaccurate. Do not allow it to go below zero though:
2922 if (unlikely((long)sum < 0))
2928 unsigned long long nr_context_switches(void)
2931 unsigned long long sum = 0;
2933 for_each_possible_cpu(i)
2934 sum += cpu_rq(i)->nr_switches;
2939 unsigned long nr_iowait(void)
2941 unsigned long i, sum = 0;
2943 for_each_possible_cpu(i)
2944 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2949 unsigned long nr_iowait_cpu(int cpu)
2951 struct rq *this = cpu_rq(cpu);
2952 return atomic_read(&this->nr_iowait);
2955 unsigned long this_cpu_load(void)
2957 struct rq *this = this_rq();
2958 return this->cpu_load[0];
2962 /* Variables and functions for calc_load */
2963 static atomic_long_t calc_load_tasks;
2964 static unsigned long calc_load_update;
2965 unsigned long avenrun[3];
2966 EXPORT_SYMBOL(avenrun);
2968 static long calc_load_fold_active(struct rq *this_rq)
2970 long nr_active, delta = 0;
2972 nr_active = this_rq->nr_running;
2973 nr_active += (long) this_rq->nr_uninterruptible;
2975 if (nr_active != this_rq->calc_load_active) {
2976 delta = nr_active - this_rq->calc_load_active;
2977 this_rq->calc_load_active = nr_active;
2985 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2987 * When making the ILB scale, we should try to pull this in as well.
2989 static atomic_long_t calc_load_tasks_idle;
2991 static void calc_load_account_idle(struct rq *this_rq)
2995 delta = calc_load_fold_active(this_rq);
2997 atomic_long_add(delta, &calc_load_tasks_idle);
3000 static long calc_load_fold_idle(void)
3005 * Its got a race, we don't care...
3007 if (atomic_long_read(&calc_load_tasks_idle))
3008 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3013 static void calc_load_account_idle(struct rq *this_rq)
3017 static inline long calc_load_fold_idle(void)
3024 * get_avenrun - get the load average array
3025 * @loads: pointer to dest load array
3026 * @offset: offset to add
3027 * @shift: shift count to shift the result left
3029 * These values are estimates at best, so no need for locking.
3031 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3033 loads[0] = (avenrun[0] + offset) << shift;
3034 loads[1] = (avenrun[1] + offset) << shift;
3035 loads[2] = (avenrun[2] + offset) << shift;
3038 static unsigned long
3039 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3042 load += active * (FIXED_1 - exp);
3043 return load >> FSHIFT;
3047 * calc_load - update the avenrun load estimates 10 ticks after the
3048 * CPUs have updated calc_load_tasks.
3050 void calc_global_load(void)
3052 unsigned long upd = calc_load_update + 10;
3055 if (time_before(jiffies, upd))
3058 active = atomic_long_read(&calc_load_tasks);
3059 active = active > 0 ? active * FIXED_1 : 0;
3061 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3062 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3063 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3065 calc_load_update += LOAD_FREQ;
3069 * Called from update_cpu_load() to periodically update this CPU's
3072 static void calc_load_account_active(struct rq *this_rq)
3076 if (time_before(jiffies, this_rq->calc_load_update))
3079 delta = calc_load_fold_active(this_rq);
3080 delta += calc_load_fold_idle();
3082 atomic_long_add(delta, &calc_load_tasks);
3084 this_rq->calc_load_update += LOAD_FREQ;
3088 * The exact cpuload at various idx values, calculated at every tick would be
3089 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3091 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3092 * on nth tick when cpu may be busy, then we have:
3093 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3094 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3096 * decay_load_missed() below does efficient calculation of
3097 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3098 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3100 * The calculation is approximated on a 128 point scale.
3101 * degrade_zero_ticks is the number of ticks after which load at any
3102 * particular idx is approximated to be zero.
3103 * degrade_factor is a precomputed table, a row for each load idx.
3104 * Each column corresponds to degradation factor for a power of two ticks,
3105 * based on 128 point scale.
3107 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3108 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3110 * With this power of 2 load factors, we can degrade the load n times
3111 * by looking at 1 bits in n and doing as many mult/shift instead of
3112 * n mult/shifts needed by the exact degradation.
3114 #define DEGRADE_SHIFT 7
3115 static const unsigned char
3116 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3117 static const unsigned char
3118 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3119 {0, 0, 0, 0, 0, 0, 0, 0},
3120 {64, 32, 8, 0, 0, 0, 0, 0},
3121 {96, 72, 40, 12, 1, 0, 0},
3122 {112, 98, 75, 43, 15, 1, 0},
3123 {120, 112, 98, 76, 45, 16, 2} };
3126 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3127 * would be when CPU is idle and so we just decay the old load without
3128 * adding any new load.
3130 static unsigned long
3131 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3135 if (!missed_updates)
3138 if (missed_updates >= degrade_zero_ticks[idx])
3142 return load >> missed_updates;
3144 while (missed_updates) {
3145 if (missed_updates % 2)
3146 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3148 missed_updates >>= 1;
3155 * Update rq->cpu_load[] statistics. This function is usually called every
3156 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3157 * every tick. We fix it up based on jiffies.
3159 static void update_cpu_load(struct rq *this_rq)
3161 unsigned long this_load = this_rq->load.weight;
3162 unsigned long curr_jiffies = jiffies;
3163 unsigned long pending_updates;
3166 this_rq->nr_load_updates++;
3168 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3169 if (curr_jiffies == this_rq->last_load_update_tick)
3172 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3173 this_rq->last_load_update_tick = curr_jiffies;
3175 /* Update our load: */
3176 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3177 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3178 unsigned long old_load, new_load;
3180 /* scale is effectively 1 << i now, and >> i divides by scale */
3182 old_load = this_rq->cpu_load[i];
3183 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3184 new_load = this_load;
3186 * Round up the averaging division if load is increasing. This
3187 * prevents us from getting stuck on 9 if the load is 10, for
3190 if (new_load > old_load)
3191 new_load += scale - 1;
3193 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3197 static void update_cpu_load_active(struct rq *this_rq)
3199 update_cpu_load(this_rq);
3201 calc_load_account_active(this_rq);
3207 * sched_exec - execve() is a valuable balancing opportunity, because at
3208 * this point the task has the smallest effective memory and cache footprint.
3210 void sched_exec(void)
3212 struct task_struct *p = current;
3213 unsigned long flags;
3217 rq = task_rq_lock(p, &flags);
3218 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3219 if (dest_cpu == smp_processor_id())
3223 * select_task_rq() can race against ->cpus_allowed
3225 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3226 likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
3227 struct migration_arg arg = { p, dest_cpu };
3229 task_rq_unlock(rq, &flags);
3230 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3234 task_rq_unlock(rq, &flags);
3239 DEFINE_PER_CPU(struct kernel_stat, kstat);
3241 EXPORT_PER_CPU_SYMBOL(kstat);
3244 * Return any ns on the sched_clock that have not yet been accounted in
3245 * @p in case that task is currently running.
3247 * Called with task_rq_lock() held on @rq.
3249 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3253 if (task_current(rq, p)) {
3254 update_rq_clock(rq);
3255 ns = rq->clock - p->se.exec_start;
3263 unsigned long long task_delta_exec(struct task_struct *p)
3265 unsigned long flags;
3269 rq = task_rq_lock(p, &flags);
3270 ns = do_task_delta_exec(p, rq);
3271 task_rq_unlock(rq, &flags);
3277 * Return accounted runtime for the task.
3278 * In case the task is currently running, return the runtime plus current's
3279 * pending runtime that have not been accounted yet.
3281 unsigned long long task_sched_runtime(struct task_struct *p)
3283 unsigned long flags;
3287 rq = task_rq_lock(p, &flags);
3288 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3289 task_rq_unlock(rq, &flags);
3295 * Return sum_exec_runtime for the thread group.
3296 * In case the task is currently running, return the sum plus current's
3297 * pending runtime that have not been accounted yet.
3299 * Note that the thread group might have other running tasks as well,
3300 * so the return value not includes other pending runtime that other
3301 * running tasks might have.
3303 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3305 struct task_cputime totals;
3306 unsigned long flags;
3310 rq = task_rq_lock(p, &flags);
3311 thread_group_cputime(p, &totals);
3312 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3313 task_rq_unlock(rq, &flags);
3319 * Account user cpu time to a process.
3320 * @p: the process that the cpu time gets accounted to
3321 * @cputime: the cpu time spent in user space since the last update
3322 * @cputime_scaled: cputime scaled by cpu frequency
3324 void account_user_time(struct task_struct *p, cputime_t cputime,
3325 cputime_t cputime_scaled)
3327 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3330 /* Add user time to process. */
3331 p->utime = cputime_add(p->utime, cputime);
3332 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3333 account_group_user_time(p, cputime);
3335 /* Add user time to cpustat. */
3336 tmp = cputime_to_cputime64(cputime);
3337 if (TASK_NICE(p) > 0)
3338 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3340 cpustat->user = cputime64_add(cpustat->user, tmp);
3342 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3343 /* Account for user time used */
3344 acct_update_integrals(p);
3348 * Account guest cpu time to a process.
3349 * @p: the process that the cpu time gets accounted to
3350 * @cputime: the cpu time spent in virtual machine since the last update
3351 * @cputime_scaled: cputime scaled by cpu frequency
3353 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3354 cputime_t cputime_scaled)
3357 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3359 tmp = cputime_to_cputime64(cputime);
3361 /* Add guest time to process. */
3362 p->utime = cputime_add(p->utime, cputime);
3363 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3364 account_group_user_time(p, cputime);
3365 p->gtime = cputime_add(p->gtime, cputime);
3367 /* Add guest time to cpustat. */
3368 if (TASK_NICE(p) > 0) {
3369 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3370 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3372 cpustat->user = cputime64_add(cpustat->user, tmp);
3373 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3378 * Account system cpu time to a process.
3379 * @p: the process that the cpu time gets accounted to
3380 * @hardirq_offset: the offset to subtract from hardirq_count()
3381 * @cputime: the cpu time spent in kernel space since the last update
3382 * @cputime_scaled: cputime scaled by cpu frequency
3384 void account_system_time(struct task_struct *p, int hardirq_offset,
3385 cputime_t cputime, cputime_t cputime_scaled)
3387 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3390 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3391 account_guest_time(p, cputime, cputime_scaled);
3395 /* Add system time to process. */
3396 p->stime = cputime_add(p->stime, cputime);
3397 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3398 account_group_system_time(p, cputime);
3400 /* Add system time to cpustat. */
3401 tmp = cputime_to_cputime64(cputime);
3402 if (hardirq_count() - hardirq_offset)
3403 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3404 else if (softirq_count())
3405 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3407 cpustat->system = cputime64_add(cpustat->system, tmp);
3409 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3411 /* Account for system time used */
3412 acct_update_integrals(p);
3416 * Account for involuntary wait time.
3417 * @steal: the cpu time spent in involuntary wait
3419 void account_steal_time(cputime_t cputime)
3421 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3422 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3424 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3428 * Account for idle time.
3429 * @cputime: the cpu time spent in idle wait
3431 void account_idle_time(cputime_t cputime)
3433 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3434 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3435 struct rq *rq = this_rq();
3437 if (atomic_read(&rq->nr_iowait) > 0)
3438 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3440 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3443 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3446 * Account a single tick of cpu time.
3447 * @p: the process that the cpu time gets accounted to
3448 * @user_tick: indicates if the tick is a user or a system tick
3450 void account_process_tick(struct task_struct *p, int user_tick)
3452 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3453 struct rq *rq = this_rq();
3456 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3457 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3458 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3461 account_idle_time(cputime_one_jiffy);
3465 * Account multiple ticks of steal time.
3466 * @p: the process from which the cpu time has been stolen
3467 * @ticks: number of stolen ticks
3469 void account_steal_ticks(unsigned long ticks)
3471 account_steal_time(jiffies_to_cputime(ticks));
3475 * Account multiple ticks of idle time.
3476 * @ticks: number of stolen ticks
3478 void account_idle_ticks(unsigned long ticks)
3480 account_idle_time(jiffies_to_cputime(ticks));
3486 * Use precise platform statistics if available:
3488 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3489 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3495 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3497 struct task_cputime cputime;
3499 thread_group_cputime(p, &cputime);
3501 *ut = cputime.utime;
3502 *st = cputime.stime;
3506 #ifndef nsecs_to_cputime
3507 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3510 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3512 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3515 * Use CFS's precise accounting:
3517 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3522 temp = (u64)(rtime * utime);
3523 do_div(temp, total);
3524 utime = (cputime_t)temp;
3529 * Compare with previous values, to keep monotonicity:
3531 p->prev_utime = max(p->prev_utime, utime);
3532 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3534 *ut = p->prev_utime;
3535 *st = p->prev_stime;
3539 * Must be called with siglock held.
3541 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3543 struct signal_struct *sig = p->signal;
3544 struct task_cputime cputime;
3545 cputime_t rtime, utime, total;
3547 thread_group_cputime(p, &cputime);
3549 total = cputime_add(cputime.utime, cputime.stime);
3550 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3555 temp = (u64)(rtime * cputime.utime);
3556 do_div(temp, total);
3557 utime = (cputime_t)temp;
3561 sig->prev_utime = max(sig->prev_utime, utime);
3562 sig->prev_stime = max(sig->prev_stime,
3563 cputime_sub(rtime, sig->prev_utime));
3565 *ut = sig->prev_utime;
3566 *st = sig->prev_stime;
3571 * This function gets called by the timer code, with HZ frequency.
3572 * We call it with interrupts disabled.
3574 * It also gets called by the fork code, when changing the parent's
3577 void scheduler_tick(void)
3579 int cpu = smp_processor_id();
3580 struct rq *rq = cpu_rq(cpu);
3581 struct task_struct *curr = rq->curr;
3585 raw_spin_lock(&rq->lock);
3586 update_rq_clock(rq);
3587 update_cpu_load_active(rq);
3588 curr->sched_class->task_tick(rq, curr, 0);
3589 raw_spin_unlock(&rq->lock);
3591 perf_event_task_tick(curr);
3594 rq->idle_at_tick = idle_cpu(cpu);
3595 trigger_load_balance(rq, cpu);
3599 notrace unsigned long get_parent_ip(unsigned long addr)
3601 if (in_lock_functions(addr)) {
3602 addr = CALLER_ADDR2;
3603 if (in_lock_functions(addr))
3604 addr = CALLER_ADDR3;
3609 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3610 defined(CONFIG_PREEMPT_TRACER))
3612 void __kprobes add_preempt_count(int val)
3614 #ifdef CONFIG_DEBUG_PREEMPT
3618 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3621 preempt_count() += val;
3622 #ifdef CONFIG_DEBUG_PREEMPT
3624 * Spinlock count overflowing soon?
3626 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3629 if (preempt_count() == val)
3630 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3632 EXPORT_SYMBOL(add_preempt_count);
3634 void __kprobes sub_preempt_count(int val)
3636 #ifdef CONFIG_DEBUG_PREEMPT
3640 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3643 * Is the spinlock portion underflowing?
3645 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3646 !(preempt_count() & PREEMPT_MASK)))
3650 if (preempt_count() == val)
3651 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3652 preempt_count() -= val;
3654 EXPORT_SYMBOL(sub_preempt_count);
3659 * Print scheduling while atomic bug:
3661 static noinline void __schedule_bug(struct task_struct *prev)
3663 struct pt_regs *regs = get_irq_regs();
3665 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3666 prev->comm, prev->pid, preempt_count());
3668 debug_show_held_locks(prev);
3670 if (irqs_disabled())
3671 print_irqtrace_events(prev);
3680 * Various schedule()-time debugging checks and statistics:
3682 static inline void schedule_debug(struct task_struct *prev)
3685 * Test if we are atomic. Since do_exit() needs to call into
3686 * schedule() atomically, we ignore that path for now.
3687 * Otherwise, whine if we are scheduling when we should not be.
3689 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3690 __schedule_bug(prev);
3692 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3694 schedstat_inc(this_rq(), sched_count);
3695 #ifdef CONFIG_SCHEDSTATS
3696 if (unlikely(prev->lock_depth >= 0)) {
3697 schedstat_inc(this_rq(), bkl_count);
3698 schedstat_inc(prev, sched_info.bkl_count);
3703 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3706 update_rq_clock(rq);
3707 rq->skip_clock_update = 0;
3708 prev->sched_class->put_prev_task(rq, prev);
3712 * Pick up the highest-prio task:
3714 static inline struct task_struct *
3715 pick_next_task(struct rq *rq)
3717 const struct sched_class *class;
3718 struct task_struct *p;
3721 * Optimization: we know that if all tasks are in
3722 * the fair class we can call that function directly:
3724 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3725 p = fair_sched_class.pick_next_task(rq);
3730 class = sched_class_highest;
3732 p = class->pick_next_task(rq);
3736 * Will never be NULL as the idle class always
3737 * returns a non-NULL p:
3739 class = class->next;
3744 * schedule() is the main scheduler function.
3746 asmlinkage void __sched schedule(void)
3748 struct task_struct *prev, *next;
3749 unsigned long *switch_count;
3755 cpu = smp_processor_id();
3757 rcu_note_context_switch(cpu);
3760 release_kernel_lock(prev);
3761 need_resched_nonpreemptible:
3763 schedule_debug(prev);
3765 if (sched_feat(HRTICK))
3768 raw_spin_lock_irq(&rq->lock);
3769 clear_tsk_need_resched(prev);
3771 switch_count = &prev->nivcsw;
3772 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3773 if (unlikely(signal_pending_state(prev->state, prev))) {
3774 prev->state = TASK_RUNNING;
3777 * If a worker is going to sleep, notify and
3778 * ask workqueue whether it wants to wake up a
3779 * task to maintain concurrency. If so, wake
3782 if (prev->flags & PF_WQ_WORKER) {
3783 struct task_struct *to_wakeup;
3785 to_wakeup = wq_worker_sleeping(prev, cpu);
3787 try_to_wake_up_local(to_wakeup);
3789 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3791 switch_count = &prev->nvcsw;
3794 pre_schedule(rq, prev);
3796 if (unlikely(!rq->nr_running))
3797 idle_balance(cpu, rq);
3799 put_prev_task(rq, prev);
3800 next = pick_next_task(rq);
3802 if (likely(prev != next)) {
3803 sched_info_switch(prev, next);
3804 perf_event_task_sched_out(prev, next);
3810 context_switch(rq, prev, next); /* unlocks the rq */
3812 * The context switch have flipped the stack from under us
3813 * and restored the local variables which were saved when
3814 * this task called schedule() in the past. prev == current
3815 * is still correct, but it can be moved to another cpu/rq.
3817 cpu = smp_processor_id();
3820 raw_spin_unlock_irq(&rq->lock);
3824 if (unlikely(reacquire_kernel_lock(prev)))
3825 goto need_resched_nonpreemptible;
3827 preempt_enable_no_resched();
3831 EXPORT_SYMBOL(schedule);
3833 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3835 * Look out! "owner" is an entirely speculative pointer
3836 * access and not reliable.
3838 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3843 if (!sched_feat(OWNER_SPIN))
3846 #ifdef CONFIG_DEBUG_PAGEALLOC
3848 * Need to access the cpu field knowing that
3849 * DEBUG_PAGEALLOC could have unmapped it if
3850 * the mutex owner just released it and exited.
3852 if (probe_kernel_address(&owner->cpu, cpu))
3859 * Even if the access succeeded (likely case),
3860 * the cpu field may no longer be valid.
3862 if (cpu >= nr_cpumask_bits)
3866 * We need to validate that we can do a
3867 * get_cpu() and that we have the percpu area.
3869 if (!cpu_online(cpu))
3876 * Owner changed, break to re-assess state.
3878 if (lock->owner != owner)
3882 * Is that owner really running on that cpu?
3884 if (task_thread_info(rq->curr) != owner || need_resched())
3894 #ifdef CONFIG_PREEMPT
3896 * this is the entry point to schedule() from in-kernel preemption
3897 * off of preempt_enable. Kernel preemptions off return from interrupt
3898 * occur there and call schedule directly.
3900 asmlinkage void __sched preempt_schedule(void)
3902 struct thread_info *ti = current_thread_info();
3905 * If there is a non-zero preempt_count or interrupts are disabled,
3906 * we do not want to preempt the current task. Just return..
3908 if (likely(ti->preempt_count || irqs_disabled()))
3912 add_preempt_count(PREEMPT_ACTIVE);
3914 sub_preempt_count(PREEMPT_ACTIVE);
3917 * Check again in case we missed a preemption opportunity
3918 * between schedule and now.
3921 } while (need_resched());
3923 EXPORT_SYMBOL(preempt_schedule);
3926 * this is the entry point to schedule() from kernel preemption
3927 * off of irq context.
3928 * Note, that this is called and return with irqs disabled. This will
3929 * protect us against recursive calling from irq.
3931 asmlinkage void __sched preempt_schedule_irq(void)
3933 struct thread_info *ti = current_thread_info();
3935 /* Catch callers which need to be fixed */
3936 BUG_ON(ti->preempt_count || !irqs_disabled());
3939 add_preempt_count(PREEMPT_ACTIVE);
3942 local_irq_disable();
3943 sub_preempt_count(PREEMPT_ACTIVE);
3946 * Check again in case we missed a preemption opportunity
3947 * between schedule and now.
3950 } while (need_resched());
3953 #endif /* CONFIG_PREEMPT */
3955 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3958 return try_to_wake_up(curr->private, mode, wake_flags);
3960 EXPORT_SYMBOL(default_wake_function);
3963 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3964 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3965 * number) then we wake all the non-exclusive tasks and one exclusive task.
3967 * There are circumstances in which we can try to wake a task which has already
3968 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3969 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3971 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3972 int nr_exclusive, int wake_flags, void *key)
3974 wait_queue_t *curr, *next;
3976 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3977 unsigned flags = curr->flags;
3979 if (curr->func(curr, mode, wake_flags, key) &&
3980 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3986 * __wake_up - wake up threads blocked on a waitqueue.
3988 * @mode: which threads
3989 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3990 * @key: is directly passed to the wakeup function
3992 * It may be assumed that this function implies a write memory barrier before
3993 * changing the task state if and only if any tasks are woken up.
3995 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3996 int nr_exclusive, void *key)
3998 unsigned long flags;
4000 spin_lock_irqsave(&q->lock, flags);
4001 __wake_up_common(q, mode, nr_exclusive, 0, key);
4002 spin_unlock_irqrestore(&q->lock, flags);
4004 EXPORT_SYMBOL(__wake_up);
4007 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4009 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4011 __wake_up_common(q, mode, 1, 0, NULL);
4013 EXPORT_SYMBOL_GPL(__wake_up_locked);
4015 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4017 __wake_up_common(q, mode, 1, 0, key);
4021 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4023 * @mode: which threads
4024 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4025 * @key: opaque value to be passed to wakeup targets
4027 * The sync wakeup differs that the waker knows that it will schedule
4028 * away soon, so while the target thread will be woken up, it will not
4029 * be migrated to another CPU - ie. the two threads are 'synchronized'
4030 * with each other. This can prevent needless bouncing between CPUs.
4032 * On UP it can prevent extra preemption.
4034 * It may be assumed that this function implies a write memory barrier before
4035 * changing the task state if and only if any tasks are woken up.
4037 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4038 int nr_exclusive, void *key)
4040 unsigned long flags;
4041 int wake_flags = WF_SYNC;
4046 if (unlikely(!nr_exclusive))
4049 spin_lock_irqsave(&q->lock, flags);
4050 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4051 spin_unlock_irqrestore(&q->lock, flags);
4053 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4056 * __wake_up_sync - see __wake_up_sync_key()
4058 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4060 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4062 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4065 * complete: - signals a single thread waiting on this completion
4066 * @x: holds the state of this particular completion
4068 * This will wake up a single thread waiting on this completion. Threads will be
4069 * awakened in the same order in which they were queued.
4071 * See also complete_all(), wait_for_completion() and related routines.
4073 * It may be assumed that this function implies a write memory barrier before
4074 * changing the task state if and only if any tasks are woken up.
4076 void complete(struct completion *x)
4078 unsigned long flags;
4080 spin_lock_irqsave(&x->wait.lock, flags);
4082 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4083 spin_unlock_irqrestore(&x->wait.lock, flags);
4085 EXPORT_SYMBOL(complete);
4088 * complete_all: - signals all threads waiting on this completion
4089 * @x: holds the state of this particular completion
4091 * This will wake up all threads waiting on this particular completion event.
4093 * It may be assumed that this function implies a write memory barrier before
4094 * changing the task state if and only if any tasks are woken up.
4096 void complete_all(struct completion *x)
4098 unsigned long flags;
4100 spin_lock_irqsave(&x->wait.lock, flags);
4101 x->done += UINT_MAX/2;
4102 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4103 spin_unlock_irqrestore(&x->wait.lock, flags);
4105 EXPORT_SYMBOL(complete_all);
4107 static inline long __sched
4108 do_wait_for_common(struct completion *x, long timeout, int state)
4111 DECLARE_WAITQUEUE(wait, current);
4113 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4115 if (signal_pending_state(state, current)) {
4116 timeout = -ERESTARTSYS;
4119 __set_current_state(state);
4120 spin_unlock_irq(&x->wait.lock);
4121 timeout = schedule_timeout(timeout);
4122 spin_lock_irq(&x->wait.lock);
4123 } while (!x->done && timeout);
4124 __remove_wait_queue(&x->wait, &wait);
4129 return timeout ?: 1;
4133 wait_for_common(struct completion *x, long timeout, int state)
4137 spin_lock_irq(&x->wait.lock);
4138 timeout = do_wait_for_common(x, timeout, state);
4139 spin_unlock_irq(&x->wait.lock);
4144 * wait_for_completion: - waits for completion of a task
4145 * @x: holds the state of this particular completion
4147 * This waits to be signaled for completion of a specific task. It is NOT
4148 * interruptible and there is no timeout.
4150 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4151 * and interrupt capability. Also see complete().
4153 void __sched wait_for_completion(struct completion *x)
4155 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4157 EXPORT_SYMBOL(wait_for_completion);
4160 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4161 * @x: holds the state of this particular completion
4162 * @timeout: timeout value in jiffies
4164 * This waits for either a completion of a specific task to be signaled or for a
4165 * specified timeout to expire. The timeout is in jiffies. It is not
4168 unsigned long __sched
4169 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4171 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4173 EXPORT_SYMBOL(wait_for_completion_timeout);
4176 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4177 * @x: holds the state of this particular completion
4179 * This waits for completion of a specific task to be signaled. It is
4182 int __sched wait_for_completion_interruptible(struct completion *x)
4184 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4185 if (t == -ERESTARTSYS)
4189 EXPORT_SYMBOL(wait_for_completion_interruptible);
4192 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4193 * @x: holds the state of this particular completion
4194 * @timeout: timeout value in jiffies
4196 * This waits for either a completion of a specific task to be signaled or for a
4197 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4199 unsigned long __sched
4200 wait_for_completion_interruptible_timeout(struct completion *x,
4201 unsigned long timeout)
4203 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4205 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4208 * wait_for_completion_killable: - waits for completion of a task (killable)
4209 * @x: holds the state of this particular completion
4211 * This waits to be signaled for completion of a specific task. It can be
4212 * interrupted by a kill signal.
4214 int __sched wait_for_completion_killable(struct completion *x)
4216 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4217 if (t == -ERESTARTSYS)
4221 EXPORT_SYMBOL(wait_for_completion_killable);
4224 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4225 * @x: holds the state of this particular completion
4226 * @timeout: timeout value in jiffies
4228 * This waits for either a completion of a specific task to be
4229 * signaled or for a specified timeout to expire. It can be
4230 * interrupted by a kill signal. The timeout is in jiffies.
4232 unsigned long __sched
4233 wait_for_completion_killable_timeout(struct completion *x,
4234 unsigned long timeout)
4236 return wait_for_common(x, timeout, TASK_KILLABLE);
4238 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4241 * try_wait_for_completion - try to decrement a completion without blocking
4242 * @x: completion structure
4244 * Returns: 0 if a decrement cannot be done without blocking
4245 * 1 if a decrement succeeded.
4247 * If a completion is being used as a counting completion,
4248 * attempt to decrement the counter without blocking. This
4249 * enables us to avoid waiting if the resource the completion
4250 * is protecting is not available.
4252 bool try_wait_for_completion(struct completion *x)
4254 unsigned long flags;
4257 spin_lock_irqsave(&x->wait.lock, flags);
4262 spin_unlock_irqrestore(&x->wait.lock, flags);
4265 EXPORT_SYMBOL(try_wait_for_completion);
4268 * completion_done - Test to see if a completion has any waiters
4269 * @x: completion structure
4271 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4272 * 1 if there are no waiters.
4275 bool completion_done(struct completion *x)
4277 unsigned long flags;
4280 spin_lock_irqsave(&x->wait.lock, flags);
4283 spin_unlock_irqrestore(&x->wait.lock, flags);
4286 EXPORT_SYMBOL(completion_done);
4289 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4291 unsigned long flags;
4294 init_waitqueue_entry(&wait, current);
4296 __set_current_state(state);
4298 spin_lock_irqsave(&q->lock, flags);
4299 __add_wait_queue(q, &wait);
4300 spin_unlock(&q->lock);
4301 timeout = schedule_timeout(timeout);
4302 spin_lock_irq(&q->lock);
4303 __remove_wait_queue(q, &wait);
4304 spin_unlock_irqrestore(&q->lock, flags);
4309 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4311 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4313 EXPORT_SYMBOL(interruptible_sleep_on);
4316 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4318 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4320 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4322 void __sched sleep_on(wait_queue_head_t *q)
4324 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4326 EXPORT_SYMBOL(sleep_on);
4328 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4330 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4332 EXPORT_SYMBOL(sleep_on_timeout);
4334 #ifdef CONFIG_RT_MUTEXES
4337 * rt_mutex_setprio - set the current priority of a task
4339 * @prio: prio value (kernel-internal form)
4341 * This function changes the 'effective' priority of a task. It does
4342 * not touch ->normal_prio like __setscheduler().
4344 * Used by the rt_mutex code to implement priority inheritance logic.
4346 void rt_mutex_setprio(struct task_struct *p, int prio)
4348 unsigned long flags;
4349 int oldprio, on_rq, running;
4351 const struct sched_class *prev_class;
4353 BUG_ON(prio < 0 || prio > MAX_PRIO);
4355 rq = task_rq_lock(p, &flags);
4358 prev_class = p->sched_class;
4359 on_rq = p->se.on_rq;
4360 running = task_current(rq, p);
4362 dequeue_task(rq, p, 0);
4364 p->sched_class->put_prev_task(rq, p);
4367 p->sched_class = &rt_sched_class;
4369 p->sched_class = &fair_sched_class;
4374 p->sched_class->set_curr_task(rq);
4376 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4378 check_class_changed(rq, p, prev_class, oldprio, running);
4380 task_rq_unlock(rq, &flags);
4385 void set_user_nice(struct task_struct *p, long nice)
4387 int old_prio, delta, on_rq;
4388 unsigned long flags;
4391 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4394 * We have to be careful, if called from sys_setpriority(),
4395 * the task might be in the middle of scheduling on another CPU.
4397 rq = task_rq_lock(p, &flags);
4399 * The RT priorities are set via sched_setscheduler(), but we still
4400 * allow the 'normal' nice value to be set - but as expected
4401 * it wont have any effect on scheduling until the task is
4402 * SCHED_FIFO/SCHED_RR:
4404 if (task_has_rt_policy(p)) {
4405 p->static_prio = NICE_TO_PRIO(nice);
4408 on_rq = p->se.on_rq;
4410 dequeue_task(rq, p, 0);
4412 p->static_prio = NICE_TO_PRIO(nice);
4415 p->prio = effective_prio(p);
4416 delta = p->prio - old_prio;
4419 enqueue_task(rq, p, 0);
4421 * If the task increased its priority or is running and
4422 * lowered its priority, then reschedule its CPU:
4424 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4425 resched_task(rq->curr);
4428 task_rq_unlock(rq, &flags);
4430 EXPORT_SYMBOL(set_user_nice);
4433 * can_nice - check if a task can reduce its nice value
4437 int can_nice(const struct task_struct *p, const int nice)
4439 /* convert nice value [19,-20] to rlimit style value [1,40] */
4440 int nice_rlim = 20 - nice;
4442 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4443 capable(CAP_SYS_NICE));
4446 #ifdef __ARCH_WANT_SYS_NICE
4449 * sys_nice - change the priority of the current process.
4450 * @increment: priority increment
4452 * sys_setpriority is a more generic, but much slower function that
4453 * does similar things.
4455 SYSCALL_DEFINE1(nice, int, increment)
4460 * Setpriority might change our priority at the same moment.
4461 * We don't have to worry. Conceptually one call occurs first
4462 * and we have a single winner.
4464 if (increment < -40)
4469 nice = TASK_NICE(current) + increment;
4475 if (increment < 0 && !can_nice(current, nice))
4478 retval = security_task_setnice(current, nice);
4482 set_user_nice(current, nice);
4489 * task_prio - return the priority value of a given task.
4490 * @p: the task in question.
4492 * This is the priority value as seen by users in /proc.
4493 * RT tasks are offset by -200. Normal tasks are centered
4494 * around 0, value goes from -16 to +15.
4496 int task_prio(const struct task_struct *p)
4498 return p->prio - MAX_RT_PRIO;
4502 * task_nice - return the nice value of a given task.
4503 * @p: the task in question.
4505 int task_nice(const struct task_struct *p)
4507 return TASK_NICE(p);
4509 EXPORT_SYMBOL(task_nice);
4512 * idle_cpu - is a given cpu idle currently?
4513 * @cpu: the processor in question.
4515 int idle_cpu(int cpu)
4517 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4521 * idle_task - return the idle task for a given cpu.
4522 * @cpu: the processor in question.
4524 struct task_struct *idle_task(int cpu)
4526 return cpu_rq(cpu)->idle;
4530 * find_process_by_pid - find a process with a matching PID value.
4531 * @pid: the pid in question.
4533 static struct task_struct *find_process_by_pid(pid_t pid)
4535 return pid ? find_task_by_vpid(pid) : current;
4538 /* Actually do priority change: must hold rq lock. */
4540 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4542 BUG_ON(p->se.on_rq);
4545 p->rt_priority = prio;
4546 p->normal_prio = normal_prio(p);
4547 /* we are holding p->pi_lock already */
4548 p->prio = rt_mutex_getprio(p);
4549 if (rt_prio(p->prio))
4550 p->sched_class = &rt_sched_class;
4552 p->sched_class = &fair_sched_class;
4557 * check the target process has a UID that matches the current process's
4559 static bool check_same_owner(struct task_struct *p)
4561 const struct cred *cred = current_cred(), *pcred;
4565 pcred = __task_cred(p);
4566 match = (cred->euid == pcred->euid ||
4567 cred->euid == pcred->uid);
4572 static int __sched_setscheduler(struct task_struct *p, int policy,
4573 struct sched_param *param, bool user)
4575 int retval, oldprio, oldpolicy = -1, on_rq, running;
4576 unsigned long flags;
4577 const struct sched_class *prev_class;
4581 /* may grab non-irq protected spin_locks */
4582 BUG_ON(in_interrupt());
4584 /* double check policy once rq lock held */
4586 reset_on_fork = p->sched_reset_on_fork;
4587 policy = oldpolicy = p->policy;
4589 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4590 policy &= ~SCHED_RESET_ON_FORK;
4592 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4593 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4594 policy != SCHED_IDLE)
4599 * Valid priorities for SCHED_FIFO and SCHED_RR are
4600 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4601 * SCHED_BATCH and SCHED_IDLE is 0.
4603 if (param->sched_priority < 0 ||
4604 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4605 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4607 if (rt_policy(policy) != (param->sched_priority != 0))
4611 * Allow unprivileged RT tasks to decrease priority:
4613 if (user && !capable(CAP_SYS_NICE)) {
4614 if (rt_policy(policy)) {
4615 unsigned long rlim_rtprio =
4616 task_rlimit(p, RLIMIT_RTPRIO);
4618 /* can't set/change the rt policy */
4619 if (policy != p->policy && !rlim_rtprio)
4622 /* can't increase priority */
4623 if (param->sched_priority > p->rt_priority &&
4624 param->sched_priority > rlim_rtprio)
4628 * Like positive nice levels, dont allow tasks to
4629 * move out of SCHED_IDLE either:
4631 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4634 /* can't change other user's priorities */
4635 if (!check_same_owner(p))
4638 /* Normal users shall not reset the sched_reset_on_fork flag */
4639 if (p->sched_reset_on_fork && !reset_on_fork)
4644 retval = security_task_setscheduler(p, policy, param);
4650 * make sure no PI-waiters arrive (or leave) while we are
4651 * changing the priority of the task:
4653 raw_spin_lock_irqsave(&p->pi_lock, flags);
4655 * To be able to change p->policy safely, the apropriate
4656 * runqueue lock must be held.
4658 rq = __task_rq_lock(p);
4660 #ifdef CONFIG_RT_GROUP_SCHED
4663 * Do not allow realtime tasks into groups that have no runtime
4666 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4667 task_group(p)->rt_bandwidth.rt_runtime == 0) {
4668 __task_rq_unlock(rq);
4669 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4675 /* recheck policy now with rq lock held */
4676 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4677 policy = oldpolicy = -1;
4678 __task_rq_unlock(rq);
4679 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4682 on_rq = p->se.on_rq;
4683 running = task_current(rq, p);
4685 deactivate_task(rq, p, 0);
4687 p->sched_class->put_prev_task(rq, p);
4689 p->sched_reset_on_fork = reset_on_fork;
4692 prev_class = p->sched_class;
4693 __setscheduler(rq, p, policy, param->sched_priority);
4696 p->sched_class->set_curr_task(rq);
4698 activate_task(rq, p, 0);
4700 check_class_changed(rq, p, prev_class, oldprio, running);
4702 __task_rq_unlock(rq);
4703 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4705 rt_mutex_adjust_pi(p);
4711 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4712 * @p: the task in question.
4713 * @policy: new policy.
4714 * @param: structure containing the new RT priority.
4716 * NOTE that the task may be already dead.
4718 int sched_setscheduler(struct task_struct *p, int policy,
4719 struct sched_param *param)
4721 return __sched_setscheduler(p, policy, param, true);
4723 EXPORT_SYMBOL_GPL(sched_setscheduler);
4726 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4727 * @p: the task in question.
4728 * @policy: new policy.
4729 * @param: structure containing the new RT priority.
4731 * Just like sched_setscheduler, only don't bother checking if the
4732 * current context has permission. For example, this is needed in
4733 * stop_machine(): we create temporary high priority worker threads,
4734 * but our caller might not have that capability.
4736 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4737 struct sched_param *param)
4739 return __sched_setscheduler(p, policy, param, false);
4743 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4745 struct sched_param lparam;
4746 struct task_struct *p;
4749 if (!param || pid < 0)
4751 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4756 p = find_process_by_pid(pid);
4758 retval = sched_setscheduler(p, policy, &lparam);
4765 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4766 * @pid: the pid in question.
4767 * @policy: new policy.
4768 * @param: structure containing the new RT priority.
4770 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4771 struct sched_param __user *, param)
4773 /* negative values for policy are not valid */
4777 return do_sched_setscheduler(pid, policy, param);
4781 * sys_sched_setparam - set/change the RT priority of a thread
4782 * @pid: the pid in question.
4783 * @param: structure containing the new RT priority.
4785 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4787 return do_sched_setscheduler(pid, -1, param);
4791 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4792 * @pid: the pid in question.
4794 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4796 struct task_struct *p;
4804 p = find_process_by_pid(pid);
4806 retval = security_task_getscheduler(p);
4809 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4816 * sys_sched_getparam - get the RT priority of a thread
4817 * @pid: the pid in question.
4818 * @param: structure containing the RT priority.
4820 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4822 struct sched_param lp;
4823 struct task_struct *p;
4826 if (!param || pid < 0)
4830 p = find_process_by_pid(pid);
4835 retval = security_task_getscheduler(p);
4839 lp.sched_priority = p->rt_priority;
4843 * This one might sleep, we cannot do it with a spinlock held ...
4845 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4854 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4856 cpumask_var_t cpus_allowed, new_mask;
4857 struct task_struct *p;
4863 p = find_process_by_pid(pid);
4870 /* Prevent p going away */
4874 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4878 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4880 goto out_free_cpus_allowed;
4883 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4886 retval = security_task_setscheduler(p, 0, NULL);
4890 cpuset_cpus_allowed(p, cpus_allowed);
4891 cpumask_and(new_mask, in_mask, cpus_allowed);
4893 retval = set_cpus_allowed_ptr(p, new_mask);
4896 cpuset_cpus_allowed(p, cpus_allowed);
4897 if (!cpumask_subset(new_mask, cpus_allowed)) {
4899 * We must have raced with a concurrent cpuset
4900 * update. Just reset the cpus_allowed to the
4901 * cpuset's cpus_allowed
4903 cpumask_copy(new_mask, cpus_allowed);
4908 free_cpumask_var(new_mask);
4909 out_free_cpus_allowed:
4910 free_cpumask_var(cpus_allowed);
4917 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4918 struct cpumask *new_mask)
4920 if (len < cpumask_size())
4921 cpumask_clear(new_mask);
4922 else if (len > cpumask_size())
4923 len = cpumask_size();
4925 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4929 * sys_sched_setaffinity - set the cpu affinity of a process
4930 * @pid: pid of the process
4931 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4932 * @user_mask_ptr: user-space pointer to the new cpu mask
4934 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4935 unsigned long __user *, user_mask_ptr)
4937 cpumask_var_t new_mask;
4940 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4943 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4945 retval = sched_setaffinity(pid, new_mask);
4946 free_cpumask_var(new_mask);
4950 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4952 struct task_struct *p;
4953 unsigned long flags;
4961 p = find_process_by_pid(pid);
4965 retval = security_task_getscheduler(p);
4969 rq = task_rq_lock(p, &flags);
4970 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4971 task_rq_unlock(rq, &flags);
4981 * sys_sched_getaffinity - get the cpu affinity of a process
4982 * @pid: pid of the process
4983 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4984 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4986 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4987 unsigned long __user *, user_mask_ptr)
4992 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4994 if (len & (sizeof(unsigned long)-1))
4997 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5000 ret = sched_getaffinity(pid, mask);
5002 size_t retlen = min_t(size_t, len, cpumask_size());
5004 if (copy_to_user(user_mask_ptr, mask, retlen))
5009 free_cpumask_var(mask);
5015 * sys_sched_yield - yield the current processor to other threads.
5017 * This function yields the current CPU to other tasks. If there are no
5018 * other threads running on this CPU then this function will return.
5020 SYSCALL_DEFINE0(sched_yield)
5022 struct rq *rq = this_rq_lock();
5024 schedstat_inc(rq, yld_count);
5025 current->sched_class->yield_task(rq);
5028 * Since we are going to call schedule() anyway, there's
5029 * no need to preempt or enable interrupts:
5031 __release(rq->lock);
5032 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5033 do_raw_spin_unlock(&rq->lock);
5034 preempt_enable_no_resched();
5041 static inline int should_resched(void)
5043 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5046 static void __cond_resched(void)
5048 add_preempt_count(PREEMPT_ACTIVE);
5050 sub_preempt_count(PREEMPT_ACTIVE);
5053 int __sched _cond_resched(void)
5055 if (should_resched()) {
5061 EXPORT_SYMBOL(_cond_resched);
5064 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5065 * call schedule, and on return reacquire the lock.
5067 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5068 * operations here to prevent schedule() from being called twice (once via
5069 * spin_unlock(), once by hand).
5071 int __cond_resched_lock(spinlock_t *lock)
5073 int resched = should_resched();
5076 lockdep_assert_held(lock);
5078 if (spin_needbreak(lock) || resched) {
5089 EXPORT_SYMBOL(__cond_resched_lock);
5091 int __sched __cond_resched_softirq(void)
5093 BUG_ON(!in_softirq());
5095 if (should_resched()) {
5103 EXPORT_SYMBOL(__cond_resched_softirq);
5106 * yield - yield the current processor to other threads.
5108 * This is a shortcut for kernel-space yielding - it marks the
5109 * thread runnable and calls sys_sched_yield().
5111 void __sched yield(void)
5113 set_current_state(TASK_RUNNING);
5116 EXPORT_SYMBOL(yield);
5119 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5120 * that process accounting knows that this is a task in IO wait state.
5122 void __sched io_schedule(void)
5124 struct rq *rq = raw_rq();
5126 delayacct_blkio_start();
5127 atomic_inc(&rq->nr_iowait);
5128 current->in_iowait = 1;
5130 current->in_iowait = 0;
5131 atomic_dec(&rq->nr_iowait);
5132 delayacct_blkio_end();
5134 EXPORT_SYMBOL(io_schedule);
5136 long __sched io_schedule_timeout(long timeout)
5138 struct rq *rq = raw_rq();
5141 delayacct_blkio_start();
5142 atomic_inc(&rq->nr_iowait);
5143 current->in_iowait = 1;
5144 ret = schedule_timeout(timeout);
5145 current->in_iowait = 0;
5146 atomic_dec(&rq->nr_iowait);
5147 delayacct_blkio_end();
5152 * sys_sched_get_priority_max - return maximum RT priority.
5153 * @policy: scheduling class.
5155 * this syscall returns the maximum rt_priority that can be used
5156 * by a given scheduling class.
5158 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5165 ret = MAX_USER_RT_PRIO-1;
5177 * sys_sched_get_priority_min - return minimum RT priority.
5178 * @policy: scheduling class.
5180 * this syscall returns the minimum rt_priority that can be used
5181 * by a given scheduling class.
5183 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5201 * sys_sched_rr_get_interval - return the default timeslice of a process.
5202 * @pid: pid of the process.
5203 * @interval: userspace pointer to the timeslice value.
5205 * this syscall writes the default timeslice value of a given process
5206 * into the user-space timespec buffer. A value of '0' means infinity.
5208 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5209 struct timespec __user *, interval)
5211 struct task_struct *p;
5212 unsigned int time_slice;
5213 unsigned long flags;
5223 p = find_process_by_pid(pid);
5227 retval = security_task_getscheduler(p);
5231 rq = task_rq_lock(p, &flags);
5232 time_slice = p->sched_class->get_rr_interval(rq, p);
5233 task_rq_unlock(rq, &flags);
5236 jiffies_to_timespec(time_slice, &t);
5237 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5245 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5247 void sched_show_task(struct task_struct *p)
5249 unsigned long free = 0;
5252 state = p->state ? __ffs(p->state) + 1 : 0;
5253 printk(KERN_INFO "%-13.13s %c", p->comm,
5254 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5255 #if BITS_PER_LONG == 32
5256 if (state == TASK_RUNNING)
5257 printk(KERN_CONT " running ");
5259 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5261 if (state == TASK_RUNNING)
5262 printk(KERN_CONT " running task ");
5264 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5266 #ifdef CONFIG_DEBUG_STACK_USAGE
5267 free = stack_not_used(p);
5269 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5270 task_pid_nr(p), task_pid_nr(p->real_parent),
5271 (unsigned long)task_thread_info(p)->flags);
5273 show_stack(p, NULL);
5276 void show_state_filter(unsigned long state_filter)
5278 struct task_struct *g, *p;
5280 #if BITS_PER_LONG == 32
5282 " task PC stack pid father\n");
5285 " task PC stack pid father\n");
5287 read_lock(&tasklist_lock);
5288 do_each_thread(g, p) {
5290 * reset the NMI-timeout, listing all files on a slow
5291 * console might take alot of time:
5293 touch_nmi_watchdog();
5294 if (!state_filter || (p->state & state_filter))
5296 } while_each_thread(g, p);
5298 touch_all_softlockup_watchdogs();
5300 #ifdef CONFIG_SCHED_DEBUG
5301 sysrq_sched_debug_show();
5303 read_unlock(&tasklist_lock);
5305 * Only show locks if all tasks are dumped:
5308 debug_show_all_locks();
5311 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5313 idle->sched_class = &idle_sched_class;
5317 * init_idle - set up an idle thread for a given CPU
5318 * @idle: task in question
5319 * @cpu: cpu the idle task belongs to
5321 * NOTE: this function does not set the idle thread's NEED_RESCHED
5322 * flag, to make booting more robust.
5324 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5326 struct rq *rq = cpu_rq(cpu);
5327 unsigned long flags;
5329 raw_spin_lock_irqsave(&rq->lock, flags);
5332 idle->state = TASK_RUNNING;
5333 idle->se.exec_start = sched_clock();
5335 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5336 __set_task_cpu(idle, cpu);
5338 rq->curr = rq->idle = idle;
5339 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5342 raw_spin_unlock_irqrestore(&rq->lock, flags);
5344 /* Set the preempt count _outside_ the spinlocks! */
5345 #if defined(CONFIG_PREEMPT)
5346 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5348 task_thread_info(idle)->preempt_count = 0;
5351 * The idle tasks have their own, simple scheduling class:
5353 idle->sched_class = &idle_sched_class;
5354 ftrace_graph_init_task(idle);
5358 * In a system that switches off the HZ timer nohz_cpu_mask
5359 * indicates which cpus entered this state. This is used
5360 * in the rcu update to wait only for active cpus. For system
5361 * which do not switch off the HZ timer nohz_cpu_mask should
5362 * always be CPU_BITS_NONE.
5364 cpumask_var_t nohz_cpu_mask;
5367 * Increase the granularity value when there are more CPUs,
5368 * because with more CPUs the 'effective latency' as visible
5369 * to users decreases. But the relationship is not linear,
5370 * so pick a second-best guess by going with the log2 of the
5373 * This idea comes from the SD scheduler of Con Kolivas:
5375 static int get_update_sysctl_factor(void)
5377 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5378 unsigned int factor;
5380 switch (sysctl_sched_tunable_scaling) {
5381 case SCHED_TUNABLESCALING_NONE:
5384 case SCHED_TUNABLESCALING_LINEAR:
5387 case SCHED_TUNABLESCALING_LOG:
5389 factor = 1 + ilog2(cpus);
5396 static void update_sysctl(void)
5398 unsigned int factor = get_update_sysctl_factor();
5400 #define SET_SYSCTL(name) \
5401 (sysctl_##name = (factor) * normalized_sysctl_##name)
5402 SET_SYSCTL(sched_min_granularity);
5403 SET_SYSCTL(sched_latency);
5404 SET_SYSCTL(sched_wakeup_granularity);
5405 SET_SYSCTL(sched_shares_ratelimit);
5409 static inline void sched_init_granularity(void)
5416 * This is how migration works:
5418 * 1) we invoke migration_cpu_stop() on the target CPU using
5420 * 2) stopper starts to run (implicitly forcing the migrated thread
5422 * 3) it checks whether the migrated task is still in the wrong runqueue.
5423 * 4) if it's in the wrong runqueue then the migration thread removes
5424 * it and puts it into the right queue.
5425 * 5) stopper completes and stop_one_cpu() returns and the migration
5430 * Change a given task's CPU affinity. Migrate the thread to a
5431 * proper CPU and schedule it away if the CPU it's executing on
5432 * is removed from the allowed bitmask.
5434 * NOTE: the caller must have a valid reference to the task, the
5435 * task must not exit() & deallocate itself prematurely. The
5436 * call is not atomic; no spinlocks may be held.
5438 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5440 unsigned long flags;
5442 unsigned int dest_cpu;
5446 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5447 * drop the rq->lock and still rely on ->cpus_allowed.
5450 while (task_is_waking(p))
5452 rq = task_rq_lock(p, &flags);
5453 if (task_is_waking(p)) {
5454 task_rq_unlock(rq, &flags);
5458 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5463 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5464 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5469 if (p->sched_class->set_cpus_allowed)
5470 p->sched_class->set_cpus_allowed(p, new_mask);
5472 cpumask_copy(&p->cpus_allowed, new_mask);
5473 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5476 /* Can the task run on the task's current CPU? If so, we're done */
5477 if (cpumask_test_cpu(task_cpu(p), new_mask))
5480 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5481 if (migrate_task(p, dest_cpu)) {
5482 struct migration_arg arg = { p, dest_cpu };
5483 /* Need help from migration thread: drop lock and wait. */
5484 task_rq_unlock(rq, &flags);
5485 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5486 tlb_migrate_finish(p->mm);
5490 task_rq_unlock(rq, &flags);
5494 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5497 * Move (not current) task off this cpu, onto dest cpu. We're doing
5498 * this because either it can't run here any more (set_cpus_allowed()
5499 * away from this CPU, or CPU going down), or because we're
5500 * attempting to rebalance this task on exec (sched_exec).
5502 * So we race with normal scheduler movements, but that's OK, as long
5503 * as the task is no longer on this CPU.
5505 * Returns non-zero if task was successfully migrated.
5507 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5509 struct rq *rq_dest, *rq_src;
5512 if (unlikely(!cpu_active(dest_cpu)))
5515 rq_src = cpu_rq(src_cpu);
5516 rq_dest = cpu_rq(dest_cpu);
5518 double_rq_lock(rq_src, rq_dest);
5519 /* Already moved. */
5520 if (task_cpu(p) != src_cpu)
5522 /* Affinity changed (again). */
5523 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5527 * If we're not on a rq, the next wake-up will ensure we're
5531 deactivate_task(rq_src, p, 0);
5532 set_task_cpu(p, dest_cpu);
5533 activate_task(rq_dest, p, 0);
5534 check_preempt_curr(rq_dest, p, 0);
5539 double_rq_unlock(rq_src, rq_dest);
5544 * migration_cpu_stop - this will be executed by a highprio stopper thread
5545 * and performs thread migration by bumping thread off CPU then
5546 * 'pushing' onto another runqueue.
5548 static int migration_cpu_stop(void *data)
5550 struct migration_arg *arg = data;
5553 * The original target cpu might have gone down and we might
5554 * be on another cpu but it doesn't matter.
5556 local_irq_disable();
5557 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5562 #ifdef CONFIG_HOTPLUG_CPU
5564 * Figure out where task on dead CPU should go, use force if necessary.
5566 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5568 struct rq *rq = cpu_rq(dead_cpu);
5569 int needs_cpu, uninitialized_var(dest_cpu);
5570 unsigned long flags;
5572 local_irq_save(flags);
5574 raw_spin_lock(&rq->lock);
5575 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5577 dest_cpu = select_fallback_rq(dead_cpu, p);
5578 raw_spin_unlock(&rq->lock);
5580 * It can only fail if we race with set_cpus_allowed(),
5581 * in the racer should migrate the task anyway.
5584 __migrate_task(p, dead_cpu, dest_cpu);
5585 local_irq_restore(flags);
5589 * While a dead CPU has no uninterruptible tasks queued at this point,
5590 * it might still have a nonzero ->nr_uninterruptible counter, because
5591 * for performance reasons the counter is not stricly tracking tasks to
5592 * their home CPUs. So we just add the counter to another CPU's counter,
5593 * to keep the global sum constant after CPU-down:
5595 static void migrate_nr_uninterruptible(struct rq *rq_src)
5597 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5598 unsigned long flags;
5600 local_irq_save(flags);
5601 double_rq_lock(rq_src, rq_dest);
5602 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5603 rq_src->nr_uninterruptible = 0;
5604 double_rq_unlock(rq_src, rq_dest);
5605 local_irq_restore(flags);
5608 /* Run through task list and migrate tasks from the dead cpu. */
5609 static void migrate_live_tasks(int src_cpu)
5611 struct task_struct *p, *t;
5613 read_lock(&tasklist_lock);
5615 do_each_thread(t, p) {
5619 if (task_cpu(p) == src_cpu)
5620 move_task_off_dead_cpu(src_cpu, p);
5621 } while_each_thread(t, p);
5623 read_unlock(&tasklist_lock);
5627 * Schedules idle task to be the next runnable task on current CPU.
5628 * It does so by boosting its priority to highest possible.
5629 * Used by CPU offline code.
5631 void sched_idle_next(void)
5633 int this_cpu = smp_processor_id();
5634 struct rq *rq = cpu_rq(this_cpu);
5635 struct task_struct *p = rq->idle;
5636 unsigned long flags;
5638 /* cpu has to be offline */
5639 BUG_ON(cpu_online(this_cpu));
5642 * Strictly not necessary since rest of the CPUs are stopped by now
5643 * and interrupts disabled on the current cpu.
5645 raw_spin_lock_irqsave(&rq->lock, flags);
5647 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5649 activate_task(rq, p, 0);
5651 raw_spin_unlock_irqrestore(&rq->lock, flags);
5655 * Ensures that the idle task is using init_mm right before its cpu goes
5658 void idle_task_exit(void)
5660 struct mm_struct *mm = current->active_mm;
5662 BUG_ON(cpu_online(smp_processor_id()));
5665 switch_mm(mm, &init_mm, current);
5669 /* called under rq->lock with disabled interrupts */
5670 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5672 struct rq *rq = cpu_rq(dead_cpu);
5674 /* Must be exiting, otherwise would be on tasklist. */
5675 BUG_ON(!p->exit_state);
5677 /* Cannot have done final schedule yet: would have vanished. */
5678 BUG_ON(p->state == TASK_DEAD);
5683 * Drop lock around migration; if someone else moves it,
5684 * that's OK. No task can be added to this CPU, so iteration is
5687 raw_spin_unlock_irq(&rq->lock);
5688 move_task_off_dead_cpu(dead_cpu, p);
5689 raw_spin_lock_irq(&rq->lock);
5694 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5695 static void migrate_dead_tasks(unsigned int dead_cpu)
5697 struct rq *rq = cpu_rq(dead_cpu);
5698 struct task_struct *next;
5701 if (!rq->nr_running)
5703 next = pick_next_task(rq);
5706 next->sched_class->put_prev_task(rq, next);
5707 migrate_dead(dead_cpu, next);
5713 * remove the tasks which were accounted by rq from calc_load_tasks.
5715 static void calc_global_load_remove(struct rq *rq)
5717 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5718 rq->calc_load_active = 0;
5720 #endif /* CONFIG_HOTPLUG_CPU */
5722 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5724 static struct ctl_table sd_ctl_dir[] = {
5726 .procname = "sched_domain",
5732 static struct ctl_table sd_ctl_root[] = {
5734 .procname = "kernel",
5736 .child = sd_ctl_dir,
5741 static struct ctl_table *sd_alloc_ctl_entry(int n)
5743 struct ctl_table *entry =
5744 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5749 static void sd_free_ctl_entry(struct ctl_table **tablep)
5751 struct ctl_table *entry;
5754 * In the intermediate directories, both the child directory and
5755 * procname are dynamically allocated and could fail but the mode
5756 * will always be set. In the lowest directory the names are
5757 * static strings and all have proc handlers.
5759 for (entry = *tablep; entry->mode; entry++) {
5761 sd_free_ctl_entry(&entry->child);
5762 if (entry->proc_handler == NULL)
5763 kfree(entry->procname);
5771 set_table_entry(struct ctl_table *entry,
5772 const char *procname, void *data, int maxlen,
5773 mode_t mode, proc_handler *proc_handler)
5775 entry->procname = procname;
5777 entry->maxlen = maxlen;
5779 entry->proc_handler = proc_handler;
5782 static struct ctl_table *
5783 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5785 struct ctl_table *table = sd_alloc_ctl_entry(13);
5790 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5791 sizeof(long), 0644, proc_doulongvec_minmax);
5792 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5793 sizeof(long), 0644, proc_doulongvec_minmax);
5794 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5795 sizeof(int), 0644, proc_dointvec_minmax);
5796 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5797 sizeof(int), 0644, proc_dointvec_minmax);
5798 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5799 sizeof(int), 0644, proc_dointvec_minmax);
5800 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5801 sizeof(int), 0644, proc_dointvec_minmax);
5802 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5803 sizeof(int), 0644, proc_dointvec_minmax);
5804 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5805 sizeof(int), 0644, proc_dointvec_minmax);
5806 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5807 sizeof(int), 0644, proc_dointvec_minmax);
5808 set_table_entry(&table[9], "cache_nice_tries",
5809 &sd->cache_nice_tries,
5810 sizeof(int), 0644, proc_dointvec_minmax);
5811 set_table_entry(&table[10], "flags", &sd->flags,
5812 sizeof(int), 0644, proc_dointvec_minmax);
5813 set_table_entry(&table[11], "name", sd->name,
5814 CORENAME_MAX_SIZE, 0444, proc_dostring);
5815 /* &table[12] is terminator */
5820 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5822 struct ctl_table *entry, *table;
5823 struct sched_domain *sd;
5824 int domain_num = 0, i;
5827 for_each_domain(cpu, sd)
5829 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5834 for_each_domain(cpu, sd) {
5835 snprintf(buf, 32, "domain%d", i);
5836 entry->procname = kstrdup(buf, GFP_KERNEL);
5838 entry->child = sd_alloc_ctl_domain_table(sd);
5845 static struct ctl_table_header *sd_sysctl_header;
5846 static void register_sched_domain_sysctl(void)
5848 int i, cpu_num = num_possible_cpus();
5849 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5852 WARN_ON(sd_ctl_dir[0].child);
5853 sd_ctl_dir[0].child = entry;
5858 for_each_possible_cpu(i) {
5859 snprintf(buf, 32, "cpu%d", i);
5860 entry->procname = kstrdup(buf, GFP_KERNEL);
5862 entry->child = sd_alloc_ctl_cpu_table(i);
5866 WARN_ON(sd_sysctl_header);
5867 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5870 /* may be called multiple times per register */
5871 static void unregister_sched_domain_sysctl(void)
5873 if (sd_sysctl_header)
5874 unregister_sysctl_table(sd_sysctl_header);
5875 sd_sysctl_header = NULL;
5876 if (sd_ctl_dir[0].child)
5877 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5880 static void register_sched_domain_sysctl(void)
5883 static void unregister_sched_domain_sysctl(void)
5888 static void set_rq_online(struct rq *rq)
5891 const struct sched_class *class;
5893 cpumask_set_cpu(rq->cpu, rq->rd->online);
5896 for_each_class(class) {
5897 if (class->rq_online)
5898 class->rq_online(rq);
5903 static void set_rq_offline(struct rq *rq)
5906 const struct sched_class *class;
5908 for_each_class(class) {
5909 if (class->rq_offline)
5910 class->rq_offline(rq);
5913 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5919 * migration_call - callback that gets triggered when a CPU is added.
5920 * Here we can start up the necessary migration thread for the new CPU.
5922 static int __cpuinit
5923 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5925 int cpu = (long)hcpu;
5926 unsigned long flags;
5927 struct rq *rq = cpu_rq(cpu);
5931 case CPU_UP_PREPARE:
5932 case CPU_UP_PREPARE_FROZEN:
5933 rq->calc_load_update = calc_load_update;
5937 case CPU_ONLINE_FROZEN:
5938 /* Update our root-domain */
5939 raw_spin_lock_irqsave(&rq->lock, flags);
5941 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5945 raw_spin_unlock_irqrestore(&rq->lock, flags);
5948 #ifdef CONFIG_HOTPLUG_CPU
5950 case CPU_DEAD_FROZEN:
5951 migrate_live_tasks(cpu);
5952 /* Idle task back to normal (off runqueue, low prio) */
5953 raw_spin_lock_irq(&rq->lock);
5954 deactivate_task(rq, rq->idle, 0);
5955 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5956 rq->idle->sched_class = &idle_sched_class;
5957 migrate_dead_tasks(cpu);
5958 raw_spin_unlock_irq(&rq->lock);
5959 migrate_nr_uninterruptible(rq);
5960 BUG_ON(rq->nr_running != 0);
5961 calc_global_load_remove(rq);
5965 case CPU_DYING_FROZEN:
5966 /* Update our root-domain */
5967 raw_spin_lock_irqsave(&rq->lock, flags);
5969 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5972 raw_spin_unlock_irqrestore(&rq->lock, flags);
5980 * Register at high priority so that task migration (migrate_all_tasks)
5981 * happens before everything else. This has to be lower priority than
5982 * the notifier in the perf_event subsystem, though.
5984 static struct notifier_block __cpuinitdata migration_notifier = {
5985 .notifier_call = migration_call,
5986 .priority = CPU_PRI_MIGRATION,
5989 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5990 unsigned long action, void *hcpu)
5992 switch (action & ~CPU_TASKS_FROZEN) {
5994 case CPU_DOWN_FAILED:
5995 set_cpu_active((long)hcpu, true);
6002 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6003 unsigned long action, void *hcpu)
6005 switch (action & ~CPU_TASKS_FROZEN) {
6006 case CPU_DOWN_PREPARE:
6007 set_cpu_active((long)hcpu, false);
6014 static int __init migration_init(void)
6016 void *cpu = (void *)(long)smp_processor_id();
6019 /* Initialize migration for the boot CPU */
6020 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6021 BUG_ON(err == NOTIFY_BAD);
6022 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6023 register_cpu_notifier(&migration_notifier);
6025 /* Register cpu active notifiers */
6026 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6027 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6031 early_initcall(migration_init);
6036 #ifdef CONFIG_SCHED_DEBUG
6038 static __read_mostly int sched_domain_debug_enabled;
6040 static int __init sched_domain_debug_setup(char *str)
6042 sched_domain_debug_enabled = 1;
6046 early_param("sched_debug", sched_domain_debug_setup);
6048 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6049 struct cpumask *groupmask)
6051 struct sched_group *group = sd->groups;
6054 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6055 cpumask_clear(groupmask);
6057 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6059 if (!(sd->flags & SD_LOAD_BALANCE)) {
6060 printk("does not load-balance\n");
6062 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6067 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6069 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6070 printk(KERN_ERR "ERROR: domain->span does not contain "
6073 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6074 printk(KERN_ERR "ERROR: domain->groups does not contain"
6078 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6082 printk(KERN_ERR "ERROR: group is NULL\n");
6086 if (!group->cpu_power) {
6087 printk(KERN_CONT "\n");
6088 printk(KERN_ERR "ERROR: domain->cpu_power not "
6093 if (!cpumask_weight(sched_group_cpus(group))) {
6094 printk(KERN_CONT "\n");
6095 printk(KERN_ERR "ERROR: empty group\n");
6099 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6100 printk(KERN_CONT "\n");
6101 printk(KERN_ERR "ERROR: repeated CPUs\n");
6105 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6107 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6109 printk(KERN_CONT " %s", str);
6110 if (group->cpu_power != SCHED_LOAD_SCALE) {
6111 printk(KERN_CONT " (cpu_power = %d)",
6115 group = group->next;
6116 } while (group != sd->groups);
6117 printk(KERN_CONT "\n");
6119 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6120 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6123 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6124 printk(KERN_ERR "ERROR: parent span is not a superset "
6125 "of domain->span\n");
6129 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6131 cpumask_var_t groupmask;
6134 if (!sched_domain_debug_enabled)
6138 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6142 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6144 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6145 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6150 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6157 free_cpumask_var(groupmask);
6159 #else /* !CONFIG_SCHED_DEBUG */
6160 # define sched_domain_debug(sd, cpu) do { } while (0)
6161 #endif /* CONFIG_SCHED_DEBUG */
6163 static int sd_degenerate(struct sched_domain *sd)
6165 if (cpumask_weight(sched_domain_span(sd)) == 1)
6168 /* Following flags need at least 2 groups */
6169 if (sd->flags & (SD_LOAD_BALANCE |
6170 SD_BALANCE_NEWIDLE |
6174 SD_SHARE_PKG_RESOURCES)) {
6175 if (sd->groups != sd->groups->next)
6179 /* Following flags don't use groups */
6180 if (sd->flags & (SD_WAKE_AFFINE))
6187 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6189 unsigned long cflags = sd->flags, pflags = parent->flags;
6191 if (sd_degenerate(parent))
6194 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6197 /* Flags needing groups don't count if only 1 group in parent */
6198 if (parent->groups == parent->groups->next) {
6199 pflags &= ~(SD_LOAD_BALANCE |
6200 SD_BALANCE_NEWIDLE |
6204 SD_SHARE_PKG_RESOURCES);
6205 if (nr_node_ids == 1)
6206 pflags &= ~SD_SERIALIZE;
6208 if (~cflags & pflags)
6214 static void free_rootdomain(struct root_domain *rd)
6216 synchronize_sched();
6218 cpupri_cleanup(&rd->cpupri);
6220 free_cpumask_var(rd->rto_mask);
6221 free_cpumask_var(rd->online);
6222 free_cpumask_var(rd->span);
6226 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6228 struct root_domain *old_rd = NULL;
6229 unsigned long flags;
6231 raw_spin_lock_irqsave(&rq->lock, flags);
6236 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6239 cpumask_clear_cpu(rq->cpu, old_rd->span);
6242 * If we dont want to free the old_rt yet then
6243 * set old_rd to NULL to skip the freeing later
6246 if (!atomic_dec_and_test(&old_rd->refcount))
6250 atomic_inc(&rd->refcount);
6253 cpumask_set_cpu(rq->cpu, rd->span);
6254 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6257 raw_spin_unlock_irqrestore(&rq->lock, flags);
6260 free_rootdomain(old_rd);
6263 static int init_rootdomain(struct root_domain *rd)
6265 memset(rd, 0, sizeof(*rd));
6267 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6269 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6271 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6274 if (cpupri_init(&rd->cpupri) != 0)
6279 free_cpumask_var(rd->rto_mask);
6281 free_cpumask_var(rd->online);
6283 free_cpumask_var(rd->span);
6288 static void init_defrootdomain(void)
6290 init_rootdomain(&def_root_domain);
6292 atomic_set(&def_root_domain.refcount, 1);
6295 static struct root_domain *alloc_rootdomain(void)
6297 struct root_domain *rd;
6299 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6303 if (init_rootdomain(rd) != 0) {
6312 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6313 * hold the hotplug lock.
6316 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6318 struct rq *rq = cpu_rq(cpu);
6319 struct sched_domain *tmp;
6321 for (tmp = sd; tmp; tmp = tmp->parent)
6322 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6324 /* Remove the sched domains which do not contribute to scheduling. */
6325 for (tmp = sd; tmp; ) {
6326 struct sched_domain *parent = tmp->parent;
6330 if (sd_parent_degenerate(tmp, parent)) {
6331 tmp->parent = parent->parent;
6333 parent->parent->child = tmp;
6338 if (sd && sd_degenerate(sd)) {
6344 sched_domain_debug(sd, cpu);
6346 rq_attach_root(rq, rd);
6347 rcu_assign_pointer(rq->sd, sd);
6350 /* cpus with isolated domains */
6351 static cpumask_var_t cpu_isolated_map;
6353 /* Setup the mask of cpus configured for isolated domains */
6354 static int __init isolated_cpu_setup(char *str)
6356 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6357 cpulist_parse(str, cpu_isolated_map);
6361 __setup("isolcpus=", isolated_cpu_setup);
6364 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6365 * to a function which identifies what group(along with sched group) a CPU
6366 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6367 * (due to the fact that we keep track of groups covered with a struct cpumask).
6369 * init_sched_build_groups will build a circular linked list of the groups
6370 * covered by the given span, and will set each group's ->cpumask correctly,
6371 * and ->cpu_power to 0.
6374 init_sched_build_groups(const struct cpumask *span,
6375 const struct cpumask *cpu_map,
6376 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6377 struct sched_group **sg,
6378 struct cpumask *tmpmask),
6379 struct cpumask *covered, struct cpumask *tmpmask)
6381 struct sched_group *first = NULL, *last = NULL;
6384 cpumask_clear(covered);
6386 for_each_cpu(i, span) {
6387 struct sched_group *sg;
6388 int group = group_fn(i, cpu_map, &sg, tmpmask);
6391 if (cpumask_test_cpu(i, covered))
6394 cpumask_clear(sched_group_cpus(sg));
6397 for_each_cpu(j, span) {
6398 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6401 cpumask_set_cpu(j, covered);
6402 cpumask_set_cpu(j, sched_group_cpus(sg));
6413 #define SD_NODES_PER_DOMAIN 16
6418 * find_next_best_node - find the next node to include in a sched_domain
6419 * @node: node whose sched_domain we're building
6420 * @used_nodes: nodes already in the sched_domain
6422 * Find the next node to include in a given scheduling domain. Simply
6423 * finds the closest node not already in the @used_nodes map.
6425 * Should use nodemask_t.
6427 static int find_next_best_node(int node, nodemask_t *used_nodes)
6429 int i, n, val, min_val, best_node = 0;
6433 for (i = 0; i < nr_node_ids; i++) {
6434 /* Start at @node */
6435 n = (node + i) % nr_node_ids;
6437 if (!nr_cpus_node(n))
6440 /* Skip already used nodes */
6441 if (node_isset(n, *used_nodes))
6444 /* Simple min distance search */
6445 val = node_distance(node, n);
6447 if (val < min_val) {
6453 node_set(best_node, *used_nodes);
6458 * sched_domain_node_span - get a cpumask for a node's sched_domain
6459 * @node: node whose cpumask we're constructing
6460 * @span: resulting cpumask
6462 * Given a node, construct a good cpumask for its sched_domain to span. It
6463 * should be one that prevents unnecessary balancing, but also spreads tasks
6466 static void sched_domain_node_span(int node, struct cpumask *span)
6468 nodemask_t used_nodes;
6471 cpumask_clear(span);
6472 nodes_clear(used_nodes);
6474 cpumask_or(span, span, cpumask_of_node(node));
6475 node_set(node, used_nodes);
6477 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6478 int next_node = find_next_best_node(node, &used_nodes);
6480 cpumask_or(span, span, cpumask_of_node(next_node));
6483 #endif /* CONFIG_NUMA */
6485 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6488 * The cpus mask in sched_group and sched_domain hangs off the end.
6490 * ( See the the comments in include/linux/sched.h:struct sched_group
6491 * and struct sched_domain. )
6493 struct static_sched_group {
6494 struct sched_group sg;
6495 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6498 struct static_sched_domain {
6499 struct sched_domain sd;
6500 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6506 cpumask_var_t domainspan;
6507 cpumask_var_t covered;
6508 cpumask_var_t notcovered;
6510 cpumask_var_t nodemask;
6511 cpumask_var_t this_sibling_map;
6512 cpumask_var_t this_core_map;
6513 cpumask_var_t send_covered;
6514 cpumask_var_t tmpmask;
6515 struct sched_group **sched_group_nodes;
6516 struct root_domain *rd;
6520 sa_sched_groups = 0,
6525 sa_this_sibling_map,
6527 sa_sched_group_nodes,
6537 * SMT sched-domains:
6539 #ifdef CONFIG_SCHED_SMT
6540 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6541 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6544 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6545 struct sched_group **sg, struct cpumask *unused)
6548 *sg = &per_cpu(sched_groups, cpu).sg;
6551 #endif /* CONFIG_SCHED_SMT */
6554 * multi-core sched-domains:
6556 #ifdef CONFIG_SCHED_MC
6557 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6558 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6559 #endif /* CONFIG_SCHED_MC */
6561 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6563 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6564 struct sched_group **sg, struct cpumask *mask)
6568 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6569 group = cpumask_first(mask);
6571 *sg = &per_cpu(sched_group_core, group).sg;
6574 #elif defined(CONFIG_SCHED_MC)
6576 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6577 struct sched_group **sg, struct cpumask *unused)
6580 *sg = &per_cpu(sched_group_core, cpu).sg;
6585 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6586 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6589 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6590 struct sched_group **sg, struct cpumask *mask)
6593 #ifdef CONFIG_SCHED_MC
6594 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6595 group = cpumask_first(mask);
6596 #elif defined(CONFIG_SCHED_SMT)
6597 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6598 group = cpumask_first(mask);
6603 *sg = &per_cpu(sched_group_phys, group).sg;
6609 * The init_sched_build_groups can't handle what we want to do with node
6610 * groups, so roll our own. Now each node has its own list of groups which
6611 * gets dynamically allocated.
6613 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6614 static struct sched_group ***sched_group_nodes_bycpu;
6616 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6617 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6619 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6620 struct sched_group **sg,
6621 struct cpumask *nodemask)
6625 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6626 group = cpumask_first(nodemask);
6629 *sg = &per_cpu(sched_group_allnodes, group).sg;
6633 static void init_numa_sched_groups_power(struct sched_group *group_head)
6635 struct sched_group *sg = group_head;
6641 for_each_cpu(j, sched_group_cpus(sg)) {
6642 struct sched_domain *sd;
6644 sd = &per_cpu(phys_domains, j).sd;
6645 if (j != group_first_cpu(sd->groups)) {
6647 * Only add "power" once for each
6653 sg->cpu_power += sd->groups->cpu_power;
6656 } while (sg != group_head);
6659 static int build_numa_sched_groups(struct s_data *d,
6660 const struct cpumask *cpu_map, int num)
6662 struct sched_domain *sd;
6663 struct sched_group *sg, *prev;
6666 cpumask_clear(d->covered);
6667 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6668 if (cpumask_empty(d->nodemask)) {
6669 d->sched_group_nodes[num] = NULL;
6673 sched_domain_node_span(num, d->domainspan);
6674 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6676 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6679 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6683 d->sched_group_nodes[num] = sg;
6685 for_each_cpu(j, d->nodemask) {
6686 sd = &per_cpu(node_domains, j).sd;
6691 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6693 cpumask_or(d->covered, d->covered, d->nodemask);
6696 for (j = 0; j < nr_node_ids; j++) {
6697 n = (num + j) % nr_node_ids;
6698 cpumask_complement(d->notcovered, d->covered);
6699 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6700 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6701 if (cpumask_empty(d->tmpmask))
6703 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6704 if (cpumask_empty(d->tmpmask))
6706 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6710 "Can not alloc domain group for node %d\n", j);
6714 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6715 sg->next = prev->next;
6716 cpumask_or(d->covered, d->covered, d->tmpmask);
6723 #endif /* CONFIG_NUMA */
6726 /* Free memory allocated for various sched_group structures */
6727 static void free_sched_groups(const struct cpumask *cpu_map,
6728 struct cpumask *nodemask)
6732 for_each_cpu(cpu, cpu_map) {
6733 struct sched_group **sched_group_nodes
6734 = sched_group_nodes_bycpu[cpu];
6736 if (!sched_group_nodes)
6739 for (i = 0; i < nr_node_ids; i++) {
6740 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6742 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6743 if (cpumask_empty(nodemask))
6753 if (oldsg != sched_group_nodes[i])
6756 kfree(sched_group_nodes);
6757 sched_group_nodes_bycpu[cpu] = NULL;
6760 #else /* !CONFIG_NUMA */
6761 static void free_sched_groups(const struct cpumask *cpu_map,
6762 struct cpumask *nodemask)
6765 #endif /* CONFIG_NUMA */
6768 * Initialize sched groups cpu_power.
6770 * cpu_power indicates the capacity of sched group, which is used while
6771 * distributing the load between different sched groups in a sched domain.
6772 * Typically cpu_power for all the groups in a sched domain will be same unless
6773 * there are asymmetries in the topology. If there are asymmetries, group
6774 * having more cpu_power will pickup more load compared to the group having
6777 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6779 struct sched_domain *child;
6780 struct sched_group *group;
6784 WARN_ON(!sd || !sd->groups);
6786 if (cpu != group_first_cpu(sd->groups))
6791 sd->groups->cpu_power = 0;
6794 power = SCHED_LOAD_SCALE;
6795 weight = cpumask_weight(sched_domain_span(sd));
6797 * SMT siblings share the power of a single core.
6798 * Usually multiple threads get a better yield out of
6799 * that one core than a single thread would have,
6800 * reflect that in sd->smt_gain.
6802 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6803 power *= sd->smt_gain;
6805 power >>= SCHED_LOAD_SHIFT;
6807 sd->groups->cpu_power += power;
6812 * Add cpu_power of each child group to this groups cpu_power.
6814 group = child->groups;
6816 sd->groups->cpu_power += group->cpu_power;
6817 group = group->next;
6818 } while (group != child->groups);
6822 * Initializers for schedule domains
6823 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6826 #ifdef CONFIG_SCHED_DEBUG
6827 # define SD_INIT_NAME(sd, type) sd->name = #type
6829 # define SD_INIT_NAME(sd, type) do { } while (0)
6832 #define SD_INIT(sd, type) sd_init_##type(sd)
6834 #define SD_INIT_FUNC(type) \
6835 static noinline void sd_init_##type(struct sched_domain *sd) \
6837 memset(sd, 0, sizeof(*sd)); \
6838 *sd = SD_##type##_INIT; \
6839 sd->level = SD_LV_##type; \
6840 SD_INIT_NAME(sd, type); \
6845 SD_INIT_FUNC(ALLNODES)
6848 #ifdef CONFIG_SCHED_SMT
6849 SD_INIT_FUNC(SIBLING)
6851 #ifdef CONFIG_SCHED_MC
6855 static int default_relax_domain_level = -1;
6857 static int __init setup_relax_domain_level(char *str)
6861 val = simple_strtoul(str, NULL, 0);
6862 if (val < SD_LV_MAX)
6863 default_relax_domain_level = val;
6867 __setup("relax_domain_level=", setup_relax_domain_level);
6869 static void set_domain_attribute(struct sched_domain *sd,
6870 struct sched_domain_attr *attr)
6874 if (!attr || attr->relax_domain_level < 0) {
6875 if (default_relax_domain_level < 0)
6878 request = default_relax_domain_level;
6880 request = attr->relax_domain_level;
6881 if (request < sd->level) {
6882 /* turn off idle balance on this domain */
6883 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6885 /* turn on idle balance on this domain */
6886 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6890 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6891 const struct cpumask *cpu_map)
6894 case sa_sched_groups:
6895 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6896 d->sched_group_nodes = NULL;
6898 free_rootdomain(d->rd); /* fall through */
6900 free_cpumask_var(d->tmpmask); /* fall through */
6901 case sa_send_covered:
6902 free_cpumask_var(d->send_covered); /* fall through */
6903 case sa_this_core_map:
6904 free_cpumask_var(d->this_core_map); /* fall through */
6905 case sa_this_sibling_map:
6906 free_cpumask_var(d->this_sibling_map); /* fall through */
6908 free_cpumask_var(d->nodemask); /* fall through */
6909 case sa_sched_group_nodes:
6911 kfree(d->sched_group_nodes); /* fall through */
6913 free_cpumask_var(d->notcovered); /* fall through */
6915 free_cpumask_var(d->covered); /* fall through */
6917 free_cpumask_var(d->domainspan); /* fall through */
6924 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6925 const struct cpumask *cpu_map)
6928 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6930 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6931 return sa_domainspan;
6932 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6934 /* Allocate the per-node list of sched groups */
6935 d->sched_group_nodes = kcalloc(nr_node_ids,
6936 sizeof(struct sched_group *), GFP_KERNEL);
6937 if (!d->sched_group_nodes) {
6938 printk(KERN_WARNING "Can not alloc sched group node list\n");
6939 return sa_notcovered;
6941 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6943 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6944 return sa_sched_group_nodes;
6945 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6947 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6948 return sa_this_sibling_map;
6949 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6950 return sa_this_core_map;
6951 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6952 return sa_send_covered;
6953 d->rd = alloc_rootdomain();
6955 printk(KERN_WARNING "Cannot alloc root domain\n");
6958 return sa_rootdomain;
6961 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6962 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6964 struct sched_domain *sd = NULL;
6966 struct sched_domain *parent;
6969 if (cpumask_weight(cpu_map) >
6970 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6971 sd = &per_cpu(allnodes_domains, i).sd;
6972 SD_INIT(sd, ALLNODES);
6973 set_domain_attribute(sd, attr);
6974 cpumask_copy(sched_domain_span(sd), cpu_map);
6975 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6980 sd = &per_cpu(node_domains, i).sd;
6982 set_domain_attribute(sd, attr);
6983 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6984 sd->parent = parent;
6987 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6992 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6993 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6994 struct sched_domain *parent, int i)
6996 struct sched_domain *sd;
6997 sd = &per_cpu(phys_domains, i).sd;
6999 set_domain_attribute(sd, attr);
7000 cpumask_copy(sched_domain_span(sd), d->nodemask);
7001 sd->parent = parent;
7004 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7008 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7009 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7010 struct sched_domain *parent, int i)
7012 struct sched_domain *sd = parent;
7013 #ifdef CONFIG_SCHED_MC
7014 sd = &per_cpu(core_domains, i).sd;
7016 set_domain_attribute(sd, attr);
7017 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7018 sd->parent = parent;
7020 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7025 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7026 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7027 struct sched_domain *parent, int i)
7029 struct sched_domain *sd = parent;
7030 #ifdef CONFIG_SCHED_SMT
7031 sd = &per_cpu(cpu_domains, i).sd;
7032 SD_INIT(sd, SIBLING);
7033 set_domain_attribute(sd, attr);
7034 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7035 sd->parent = parent;
7037 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7042 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7043 const struct cpumask *cpu_map, int cpu)
7046 #ifdef CONFIG_SCHED_SMT
7047 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7048 cpumask_and(d->this_sibling_map, cpu_map,
7049 topology_thread_cpumask(cpu));
7050 if (cpu == cpumask_first(d->this_sibling_map))
7051 init_sched_build_groups(d->this_sibling_map, cpu_map,
7053 d->send_covered, d->tmpmask);
7056 #ifdef CONFIG_SCHED_MC
7057 case SD_LV_MC: /* set up multi-core groups */
7058 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7059 if (cpu == cpumask_first(d->this_core_map))
7060 init_sched_build_groups(d->this_core_map, cpu_map,
7062 d->send_covered, d->tmpmask);
7065 case SD_LV_CPU: /* set up physical groups */
7066 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7067 if (!cpumask_empty(d->nodemask))
7068 init_sched_build_groups(d->nodemask, cpu_map,
7070 d->send_covered, d->tmpmask);
7073 case SD_LV_ALLNODES:
7074 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7075 d->send_covered, d->tmpmask);
7084 * Build sched domains for a given set of cpus and attach the sched domains
7085 * to the individual cpus
7087 static int __build_sched_domains(const struct cpumask *cpu_map,
7088 struct sched_domain_attr *attr)
7090 enum s_alloc alloc_state = sa_none;
7092 struct sched_domain *sd;
7098 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7099 if (alloc_state != sa_rootdomain)
7101 alloc_state = sa_sched_groups;
7104 * Set up domains for cpus specified by the cpu_map.
7106 for_each_cpu(i, cpu_map) {
7107 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7110 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7111 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7112 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7113 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7116 for_each_cpu(i, cpu_map) {
7117 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7118 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7121 /* Set up physical groups */
7122 for (i = 0; i < nr_node_ids; i++)
7123 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7126 /* Set up node groups */
7128 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7130 for (i = 0; i < nr_node_ids; i++)
7131 if (build_numa_sched_groups(&d, cpu_map, i))
7135 /* Calculate CPU power for physical packages and nodes */
7136 #ifdef CONFIG_SCHED_SMT
7137 for_each_cpu(i, cpu_map) {
7138 sd = &per_cpu(cpu_domains, i).sd;
7139 init_sched_groups_power(i, sd);
7142 #ifdef CONFIG_SCHED_MC
7143 for_each_cpu(i, cpu_map) {
7144 sd = &per_cpu(core_domains, i).sd;
7145 init_sched_groups_power(i, sd);
7149 for_each_cpu(i, cpu_map) {
7150 sd = &per_cpu(phys_domains, i).sd;
7151 init_sched_groups_power(i, sd);
7155 for (i = 0; i < nr_node_ids; i++)
7156 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7158 if (d.sd_allnodes) {
7159 struct sched_group *sg;
7161 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7163 init_numa_sched_groups_power(sg);
7167 /* Attach the domains */
7168 for_each_cpu(i, cpu_map) {
7169 #ifdef CONFIG_SCHED_SMT
7170 sd = &per_cpu(cpu_domains, i).sd;
7171 #elif defined(CONFIG_SCHED_MC)
7172 sd = &per_cpu(core_domains, i).sd;
7174 sd = &per_cpu(phys_domains, i).sd;
7176 cpu_attach_domain(sd, d.rd, i);
7179 d.sched_group_nodes = NULL; /* don't free this we still need it */
7180 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7184 __free_domain_allocs(&d, alloc_state, cpu_map);
7188 static int build_sched_domains(const struct cpumask *cpu_map)
7190 return __build_sched_domains(cpu_map, NULL);
7193 static cpumask_var_t *doms_cur; /* current sched domains */
7194 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7195 static struct sched_domain_attr *dattr_cur;
7196 /* attribues of custom domains in 'doms_cur' */
7199 * Special case: If a kmalloc of a doms_cur partition (array of
7200 * cpumask) fails, then fallback to a single sched domain,
7201 * as determined by the single cpumask fallback_doms.
7203 static cpumask_var_t fallback_doms;
7206 * arch_update_cpu_topology lets virtualized architectures update the
7207 * cpu core maps. It is supposed to return 1 if the topology changed
7208 * or 0 if it stayed the same.
7210 int __attribute__((weak)) arch_update_cpu_topology(void)
7215 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7218 cpumask_var_t *doms;
7220 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7223 for (i = 0; i < ndoms; i++) {
7224 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7225 free_sched_domains(doms, i);
7232 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7235 for (i = 0; i < ndoms; i++)
7236 free_cpumask_var(doms[i]);
7241 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7242 * For now this just excludes isolated cpus, but could be used to
7243 * exclude other special cases in the future.
7245 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7249 arch_update_cpu_topology();
7251 doms_cur = alloc_sched_domains(ndoms_cur);
7253 doms_cur = &fallback_doms;
7254 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7256 err = build_sched_domains(doms_cur[0]);
7257 register_sched_domain_sysctl();
7262 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7263 struct cpumask *tmpmask)
7265 free_sched_groups(cpu_map, tmpmask);
7269 * Detach sched domains from a group of cpus specified in cpu_map
7270 * These cpus will now be attached to the NULL domain
7272 static void detach_destroy_domains(const struct cpumask *cpu_map)
7274 /* Save because hotplug lock held. */
7275 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7278 for_each_cpu(i, cpu_map)
7279 cpu_attach_domain(NULL, &def_root_domain, i);
7280 synchronize_sched();
7281 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7284 /* handle null as "default" */
7285 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7286 struct sched_domain_attr *new, int idx_new)
7288 struct sched_domain_attr tmp;
7295 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7296 new ? (new + idx_new) : &tmp,
7297 sizeof(struct sched_domain_attr));
7301 * Partition sched domains as specified by the 'ndoms_new'
7302 * cpumasks in the array doms_new[] of cpumasks. This compares
7303 * doms_new[] to the current sched domain partitioning, doms_cur[].
7304 * It destroys each deleted domain and builds each new domain.
7306 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7307 * The masks don't intersect (don't overlap.) We should setup one
7308 * sched domain for each mask. CPUs not in any of the cpumasks will
7309 * not be load balanced. If the same cpumask appears both in the
7310 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7313 * The passed in 'doms_new' should be allocated using
7314 * alloc_sched_domains. This routine takes ownership of it and will
7315 * free_sched_domains it when done with it. If the caller failed the
7316 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7317 * and partition_sched_domains() will fallback to the single partition
7318 * 'fallback_doms', it also forces the domains to be rebuilt.
7320 * If doms_new == NULL it will be replaced with cpu_online_mask.
7321 * ndoms_new == 0 is a special case for destroying existing domains,
7322 * and it will not create the default domain.
7324 * Call with hotplug lock held
7326 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7327 struct sched_domain_attr *dattr_new)
7332 mutex_lock(&sched_domains_mutex);
7334 /* always unregister in case we don't destroy any domains */
7335 unregister_sched_domain_sysctl();
7337 /* Let architecture update cpu core mappings. */
7338 new_topology = arch_update_cpu_topology();
7340 n = doms_new ? ndoms_new : 0;
7342 /* Destroy deleted domains */
7343 for (i = 0; i < ndoms_cur; i++) {
7344 for (j = 0; j < n && !new_topology; j++) {
7345 if (cpumask_equal(doms_cur[i], doms_new[j])
7346 && dattrs_equal(dattr_cur, i, dattr_new, j))
7349 /* no match - a current sched domain not in new doms_new[] */
7350 detach_destroy_domains(doms_cur[i]);
7355 if (doms_new == NULL) {
7357 doms_new = &fallback_doms;
7358 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7359 WARN_ON_ONCE(dattr_new);
7362 /* Build new domains */
7363 for (i = 0; i < ndoms_new; i++) {
7364 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7365 if (cpumask_equal(doms_new[i], doms_cur[j])
7366 && dattrs_equal(dattr_new, i, dattr_cur, j))
7369 /* no match - add a new doms_new */
7370 __build_sched_domains(doms_new[i],
7371 dattr_new ? dattr_new + i : NULL);
7376 /* Remember the new sched domains */
7377 if (doms_cur != &fallback_doms)
7378 free_sched_domains(doms_cur, ndoms_cur);
7379 kfree(dattr_cur); /* kfree(NULL) is safe */
7380 doms_cur = doms_new;
7381 dattr_cur = dattr_new;
7382 ndoms_cur = ndoms_new;
7384 register_sched_domain_sysctl();
7386 mutex_unlock(&sched_domains_mutex);
7389 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7390 static void arch_reinit_sched_domains(void)
7394 /* Destroy domains first to force the rebuild */
7395 partition_sched_domains(0, NULL, NULL);
7397 rebuild_sched_domains();
7401 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7403 unsigned int level = 0;
7405 if (sscanf(buf, "%u", &level) != 1)
7409 * level is always be positive so don't check for
7410 * level < POWERSAVINGS_BALANCE_NONE which is 0
7411 * What happens on 0 or 1 byte write,
7412 * need to check for count as well?
7415 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7419 sched_smt_power_savings = level;
7421 sched_mc_power_savings = level;
7423 arch_reinit_sched_domains();
7428 #ifdef CONFIG_SCHED_MC
7429 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7430 struct sysdev_class_attribute *attr,
7433 return sprintf(page, "%u\n", sched_mc_power_savings);
7435 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7436 struct sysdev_class_attribute *attr,
7437 const char *buf, size_t count)
7439 return sched_power_savings_store(buf, count, 0);
7441 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7442 sched_mc_power_savings_show,
7443 sched_mc_power_savings_store);
7446 #ifdef CONFIG_SCHED_SMT
7447 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7448 struct sysdev_class_attribute *attr,
7451 return sprintf(page, "%u\n", sched_smt_power_savings);
7453 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7454 struct sysdev_class_attribute *attr,
7455 const char *buf, size_t count)
7457 return sched_power_savings_store(buf, count, 1);
7459 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7460 sched_smt_power_savings_show,
7461 sched_smt_power_savings_store);
7464 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7468 #ifdef CONFIG_SCHED_SMT
7470 err = sysfs_create_file(&cls->kset.kobj,
7471 &attr_sched_smt_power_savings.attr);
7473 #ifdef CONFIG_SCHED_MC
7474 if (!err && mc_capable())
7475 err = sysfs_create_file(&cls->kset.kobj,
7476 &attr_sched_mc_power_savings.attr);
7480 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7483 * Update cpusets according to cpu_active mask. If cpusets are
7484 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7485 * around partition_sched_domains().
7487 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7490 switch (action & ~CPU_TASKS_FROZEN) {
7492 case CPU_DOWN_FAILED:
7493 cpuset_update_active_cpus();
7500 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7503 switch (action & ~CPU_TASKS_FROZEN) {
7504 case CPU_DOWN_PREPARE:
7505 cpuset_update_active_cpus();
7512 static int update_runtime(struct notifier_block *nfb,
7513 unsigned long action, void *hcpu)
7515 int cpu = (int)(long)hcpu;
7518 case CPU_DOWN_PREPARE:
7519 case CPU_DOWN_PREPARE_FROZEN:
7520 disable_runtime(cpu_rq(cpu));
7523 case CPU_DOWN_FAILED:
7524 case CPU_DOWN_FAILED_FROZEN:
7526 case CPU_ONLINE_FROZEN:
7527 enable_runtime(cpu_rq(cpu));
7535 void __init sched_init_smp(void)
7537 cpumask_var_t non_isolated_cpus;
7539 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7540 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7542 #if defined(CONFIG_NUMA)
7543 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7545 BUG_ON(sched_group_nodes_bycpu == NULL);
7548 mutex_lock(&sched_domains_mutex);
7549 arch_init_sched_domains(cpu_active_mask);
7550 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7551 if (cpumask_empty(non_isolated_cpus))
7552 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7553 mutex_unlock(&sched_domains_mutex);
7556 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7557 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7559 /* RT runtime code needs to handle some hotplug events */
7560 hotcpu_notifier(update_runtime, 0);
7564 /* Move init over to a non-isolated CPU */
7565 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7567 sched_init_granularity();
7568 free_cpumask_var(non_isolated_cpus);
7570 init_sched_rt_class();
7573 void __init sched_init_smp(void)
7575 sched_init_granularity();
7577 #endif /* CONFIG_SMP */
7579 const_debug unsigned int sysctl_timer_migration = 1;
7581 int in_sched_functions(unsigned long addr)
7583 return in_lock_functions(addr) ||
7584 (addr >= (unsigned long)__sched_text_start
7585 && addr < (unsigned long)__sched_text_end);
7588 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7590 cfs_rq->tasks_timeline = RB_ROOT;
7591 INIT_LIST_HEAD(&cfs_rq->tasks);
7592 #ifdef CONFIG_FAIR_GROUP_SCHED
7595 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7598 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7600 struct rt_prio_array *array;
7603 array = &rt_rq->active;
7604 for (i = 0; i < MAX_RT_PRIO; i++) {
7605 INIT_LIST_HEAD(array->queue + i);
7606 __clear_bit(i, array->bitmap);
7608 /* delimiter for bitsearch: */
7609 __set_bit(MAX_RT_PRIO, array->bitmap);
7611 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7612 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7614 rt_rq->highest_prio.next = MAX_RT_PRIO;
7618 rt_rq->rt_nr_migratory = 0;
7619 rt_rq->overloaded = 0;
7620 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7624 rt_rq->rt_throttled = 0;
7625 rt_rq->rt_runtime = 0;
7626 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7628 #ifdef CONFIG_RT_GROUP_SCHED
7629 rt_rq->rt_nr_boosted = 0;
7634 #ifdef CONFIG_FAIR_GROUP_SCHED
7635 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7636 struct sched_entity *se, int cpu, int add,
7637 struct sched_entity *parent)
7639 struct rq *rq = cpu_rq(cpu);
7640 tg->cfs_rq[cpu] = cfs_rq;
7641 init_cfs_rq(cfs_rq, rq);
7644 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7647 /* se could be NULL for init_task_group */
7652 se->cfs_rq = &rq->cfs;
7654 se->cfs_rq = parent->my_q;
7657 se->load.weight = tg->shares;
7658 se->load.inv_weight = 0;
7659 se->parent = parent;
7663 #ifdef CONFIG_RT_GROUP_SCHED
7664 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7665 struct sched_rt_entity *rt_se, int cpu, int add,
7666 struct sched_rt_entity *parent)
7668 struct rq *rq = cpu_rq(cpu);
7670 tg->rt_rq[cpu] = rt_rq;
7671 init_rt_rq(rt_rq, rq);
7673 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7675 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7677 tg->rt_se[cpu] = rt_se;
7682 rt_se->rt_rq = &rq->rt;
7684 rt_se->rt_rq = parent->my_q;
7686 rt_se->my_q = rt_rq;
7687 rt_se->parent = parent;
7688 INIT_LIST_HEAD(&rt_se->run_list);
7692 void __init sched_init(void)
7695 unsigned long alloc_size = 0, ptr;
7697 #ifdef CONFIG_FAIR_GROUP_SCHED
7698 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7700 #ifdef CONFIG_RT_GROUP_SCHED
7701 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7703 #ifdef CONFIG_CPUMASK_OFFSTACK
7704 alloc_size += num_possible_cpus() * cpumask_size();
7707 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7709 #ifdef CONFIG_FAIR_GROUP_SCHED
7710 init_task_group.se = (struct sched_entity **)ptr;
7711 ptr += nr_cpu_ids * sizeof(void **);
7713 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7714 ptr += nr_cpu_ids * sizeof(void **);
7716 #endif /* CONFIG_FAIR_GROUP_SCHED */
7717 #ifdef CONFIG_RT_GROUP_SCHED
7718 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7719 ptr += nr_cpu_ids * sizeof(void **);
7721 init_task_group.rt_rq = (struct rt_rq **)ptr;
7722 ptr += nr_cpu_ids * sizeof(void **);
7724 #endif /* CONFIG_RT_GROUP_SCHED */
7725 #ifdef CONFIG_CPUMASK_OFFSTACK
7726 for_each_possible_cpu(i) {
7727 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7728 ptr += cpumask_size();
7730 #endif /* CONFIG_CPUMASK_OFFSTACK */
7734 init_defrootdomain();
7737 init_rt_bandwidth(&def_rt_bandwidth,
7738 global_rt_period(), global_rt_runtime());
7740 #ifdef CONFIG_RT_GROUP_SCHED
7741 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7742 global_rt_period(), global_rt_runtime());
7743 #endif /* CONFIG_RT_GROUP_SCHED */
7745 #ifdef CONFIG_CGROUP_SCHED
7746 list_add(&init_task_group.list, &task_groups);
7747 INIT_LIST_HEAD(&init_task_group.children);
7749 #endif /* CONFIG_CGROUP_SCHED */
7751 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7752 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7753 __alignof__(unsigned long));
7755 for_each_possible_cpu(i) {
7759 raw_spin_lock_init(&rq->lock);
7761 rq->calc_load_active = 0;
7762 rq->calc_load_update = jiffies + LOAD_FREQ;
7763 init_cfs_rq(&rq->cfs, rq);
7764 init_rt_rq(&rq->rt, rq);
7765 #ifdef CONFIG_FAIR_GROUP_SCHED
7766 init_task_group.shares = init_task_group_load;
7767 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7768 #ifdef CONFIG_CGROUP_SCHED
7770 * How much cpu bandwidth does init_task_group get?
7772 * In case of task-groups formed thr' the cgroup filesystem, it
7773 * gets 100% of the cpu resources in the system. This overall
7774 * system cpu resource is divided among the tasks of
7775 * init_task_group and its child task-groups in a fair manner,
7776 * based on each entity's (task or task-group's) weight
7777 * (se->load.weight).
7779 * In other words, if init_task_group has 10 tasks of weight
7780 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7781 * then A0's share of the cpu resource is:
7783 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7785 * We achieve this by letting init_task_group's tasks sit
7786 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7788 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7790 #endif /* CONFIG_FAIR_GROUP_SCHED */
7792 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7793 #ifdef CONFIG_RT_GROUP_SCHED
7794 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7795 #ifdef CONFIG_CGROUP_SCHED
7796 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7800 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7801 rq->cpu_load[j] = 0;
7803 rq->last_load_update_tick = jiffies;
7808 rq->cpu_power = SCHED_LOAD_SCALE;
7809 rq->post_schedule = 0;
7810 rq->active_balance = 0;
7811 rq->next_balance = jiffies;
7816 rq->avg_idle = 2*sysctl_sched_migration_cost;
7817 rq_attach_root(rq, &def_root_domain);
7819 rq->nohz_balance_kick = 0;
7820 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
7824 atomic_set(&rq->nr_iowait, 0);
7827 set_load_weight(&init_task);
7829 #ifdef CONFIG_PREEMPT_NOTIFIERS
7830 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7834 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7837 #ifdef CONFIG_RT_MUTEXES
7838 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7842 * The boot idle thread does lazy MMU switching as well:
7844 atomic_inc(&init_mm.mm_count);
7845 enter_lazy_tlb(&init_mm, current);
7848 * Make us the idle thread. Technically, schedule() should not be
7849 * called from this thread, however somewhere below it might be,
7850 * but because we are the idle thread, we just pick up running again
7851 * when this runqueue becomes "idle".
7853 init_idle(current, smp_processor_id());
7855 calc_load_update = jiffies + LOAD_FREQ;
7858 * During early bootup we pretend to be a normal task:
7860 current->sched_class = &fair_sched_class;
7862 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7863 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7866 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7867 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
7868 atomic_set(&nohz.load_balancer, nr_cpu_ids);
7869 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
7870 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
7872 /* May be allocated at isolcpus cmdline parse time */
7873 if (cpu_isolated_map == NULL)
7874 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7879 scheduler_running = 1;
7882 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7883 static inline int preempt_count_equals(int preempt_offset)
7885 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7887 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7890 void __might_sleep(const char *file, int line, int preempt_offset)
7893 static unsigned long prev_jiffy; /* ratelimiting */
7895 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7896 system_state != SYSTEM_RUNNING || oops_in_progress)
7898 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7900 prev_jiffy = jiffies;
7903 "BUG: sleeping function called from invalid context at %s:%d\n",
7906 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7907 in_atomic(), irqs_disabled(),
7908 current->pid, current->comm);
7910 debug_show_held_locks(current);
7911 if (irqs_disabled())
7912 print_irqtrace_events(current);
7916 EXPORT_SYMBOL(__might_sleep);
7919 #ifdef CONFIG_MAGIC_SYSRQ
7920 static void normalize_task(struct rq *rq, struct task_struct *p)
7924 on_rq = p->se.on_rq;
7926 deactivate_task(rq, p, 0);
7927 __setscheduler(rq, p, SCHED_NORMAL, 0);
7929 activate_task(rq, p, 0);
7930 resched_task(rq->curr);
7934 void normalize_rt_tasks(void)
7936 struct task_struct *g, *p;
7937 unsigned long flags;
7940 read_lock_irqsave(&tasklist_lock, flags);
7941 do_each_thread(g, p) {
7943 * Only normalize user tasks:
7948 p->se.exec_start = 0;
7949 #ifdef CONFIG_SCHEDSTATS
7950 p->se.statistics.wait_start = 0;
7951 p->se.statistics.sleep_start = 0;
7952 p->se.statistics.block_start = 0;
7957 * Renice negative nice level userspace
7960 if (TASK_NICE(p) < 0 && p->mm)
7961 set_user_nice(p, 0);
7965 raw_spin_lock(&p->pi_lock);
7966 rq = __task_rq_lock(p);
7968 normalize_task(rq, p);
7970 __task_rq_unlock(rq);
7971 raw_spin_unlock(&p->pi_lock);
7972 } while_each_thread(g, p);
7974 read_unlock_irqrestore(&tasklist_lock, flags);
7977 #endif /* CONFIG_MAGIC_SYSRQ */
7979 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7981 * These functions are only useful for the IA64 MCA handling, or kdb.
7983 * They can only be called when the whole system has been
7984 * stopped - every CPU needs to be quiescent, and no scheduling
7985 * activity can take place. Using them for anything else would
7986 * be a serious bug, and as a result, they aren't even visible
7987 * under any other configuration.
7991 * curr_task - return the current task for a given cpu.
7992 * @cpu: the processor in question.
7994 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7996 struct task_struct *curr_task(int cpu)
7998 return cpu_curr(cpu);
8001 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8005 * set_curr_task - set the current task for a given cpu.
8006 * @cpu: the processor in question.
8007 * @p: the task pointer to set.
8009 * Description: This function must only be used when non-maskable interrupts
8010 * are serviced on a separate stack. It allows the architecture to switch the
8011 * notion of the current task on a cpu in a non-blocking manner. This function
8012 * must be called with all CPU's synchronized, and interrupts disabled, the
8013 * and caller must save the original value of the current task (see
8014 * curr_task() above) and restore that value before reenabling interrupts and
8015 * re-starting the system.
8017 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8019 void set_curr_task(int cpu, struct task_struct *p)
8026 #ifdef CONFIG_FAIR_GROUP_SCHED
8027 static void free_fair_sched_group(struct task_group *tg)
8031 for_each_possible_cpu(i) {
8033 kfree(tg->cfs_rq[i]);
8043 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8045 struct cfs_rq *cfs_rq;
8046 struct sched_entity *se;
8050 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8053 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8057 tg->shares = NICE_0_LOAD;
8059 for_each_possible_cpu(i) {
8062 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8063 GFP_KERNEL, cpu_to_node(i));
8067 se = kzalloc_node(sizeof(struct sched_entity),
8068 GFP_KERNEL, cpu_to_node(i));
8072 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8083 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8085 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8086 &cpu_rq(cpu)->leaf_cfs_rq_list);
8089 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8091 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8093 #else /* !CONFG_FAIR_GROUP_SCHED */
8094 static inline void free_fair_sched_group(struct task_group *tg)
8099 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8104 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8108 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8111 #endif /* CONFIG_FAIR_GROUP_SCHED */
8113 #ifdef CONFIG_RT_GROUP_SCHED
8114 static void free_rt_sched_group(struct task_group *tg)
8118 destroy_rt_bandwidth(&tg->rt_bandwidth);
8120 for_each_possible_cpu(i) {
8122 kfree(tg->rt_rq[i]);
8124 kfree(tg->rt_se[i]);
8132 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8134 struct rt_rq *rt_rq;
8135 struct sched_rt_entity *rt_se;
8139 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8142 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8146 init_rt_bandwidth(&tg->rt_bandwidth,
8147 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8149 for_each_possible_cpu(i) {
8152 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8153 GFP_KERNEL, cpu_to_node(i));
8157 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8158 GFP_KERNEL, cpu_to_node(i));
8162 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8173 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8175 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8176 &cpu_rq(cpu)->leaf_rt_rq_list);
8179 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8181 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8183 #else /* !CONFIG_RT_GROUP_SCHED */
8184 static inline void free_rt_sched_group(struct task_group *tg)
8189 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8194 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8198 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8201 #endif /* CONFIG_RT_GROUP_SCHED */
8203 #ifdef CONFIG_CGROUP_SCHED
8204 static void free_sched_group(struct task_group *tg)
8206 free_fair_sched_group(tg);
8207 free_rt_sched_group(tg);
8211 /* allocate runqueue etc for a new task group */
8212 struct task_group *sched_create_group(struct task_group *parent)
8214 struct task_group *tg;
8215 unsigned long flags;
8218 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8220 return ERR_PTR(-ENOMEM);
8222 if (!alloc_fair_sched_group(tg, parent))
8225 if (!alloc_rt_sched_group(tg, parent))
8228 spin_lock_irqsave(&task_group_lock, flags);
8229 for_each_possible_cpu(i) {
8230 register_fair_sched_group(tg, i);
8231 register_rt_sched_group(tg, i);
8233 list_add_rcu(&tg->list, &task_groups);
8235 WARN_ON(!parent); /* root should already exist */
8237 tg->parent = parent;
8238 INIT_LIST_HEAD(&tg->children);
8239 list_add_rcu(&tg->siblings, &parent->children);
8240 spin_unlock_irqrestore(&task_group_lock, flags);
8245 free_sched_group(tg);
8246 return ERR_PTR(-ENOMEM);
8249 /* rcu callback to free various structures associated with a task group */
8250 static void free_sched_group_rcu(struct rcu_head *rhp)
8252 /* now it should be safe to free those cfs_rqs */
8253 free_sched_group(container_of(rhp, struct task_group, rcu));
8256 /* Destroy runqueue etc associated with a task group */
8257 void sched_destroy_group(struct task_group *tg)
8259 unsigned long flags;
8262 spin_lock_irqsave(&task_group_lock, flags);
8263 for_each_possible_cpu(i) {
8264 unregister_fair_sched_group(tg, i);
8265 unregister_rt_sched_group(tg, i);
8267 list_del_rcu(&tg->list);
8268 list_del_rcu(&tg->siblings);
8269 spin_unlock_irqrestore(&task_group_lock, flags);
8271 /* wait for possible concurrent references to cfs_rqs complete */
8272 call_rcu(&tg->rcu, free_sched_group_rcu);
8275 /* change task's runqueue when it moves between groups.
8276 * The caller of this function should have put the task in its new group
8277 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8278 * reflect its new group.
8280 void sched_move_task(struct task_struct *tsk)
8283 unsigned long flags;
8286 rq = task_rq_lock(tsk, &flags);
8288 running = task_current(rq, tsk);
8289 on_rq = tsk->se.on_rq;
8292 dequeue_task(rq, tsk, 0);
8293 if (unlikely(running))
8294 tsk->sched_class->put_prev_task(rq, tsk);
8296 set_task_rq(tsk, task_cpu(tsk));
8298 #ifdef CONFIG_FAIR_GROUP_SCHED
8299 if (tsk->sched_class->moved_group)
8300 tsk->sched_class->moved_group(tsk, on_rq);
8303 if (unlikely(running))
8304 tsk->sched_class->set_curr_task(rq);
8306 enqueue_task(rq, tsk, 0);
8308 task_rq_unlock(rq, &flags);
8310 #endif /* CONFIG_CGROUP_SCHED */
8312 #ifdef CONFIG_FAIR_GROUP_SCHED
8313 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8315 struct cfs_rq *cfs_rq = se->cfs_rq;
8320 dequeue_entity(cfs_rq, se, 0);
8322 se->load.weight = shares;
8323 se->load.inv_weight = 0;
8326 enqueue_entity(cfs_rq, se, 0);
8329 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8331 struct cfs_rq *cfs_rq = se->cfs_rq;
8332 struct rq *rq = cfs_rq->rq;
8333 unsigned long flags;
8335 raw_spin_lock_irqsave(&rq->lock, flags);
8336 __set_se_shares(se, shares);
8337 raw_spin_unlock_irqrestore(&rq->lock, flags);
8340 static DEFINE_MUTEX(shares_mutex);
8342 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8345 unsigned long flags;
8348 * We can't change the weight of the root cgroup.
8353 if (shares < MIN_SHARES)
8354 shares = MIN_SHARES;
8355 else if (shares > MAX_SHARES)
8356 shares = MAX_SHARES;
8358 mutex_lock(&shares_mutex);
8359 if (tg->shares == shares)
8362 spin_lock_irqsave(&task_group_lock, flags);
8363 for_each_possible_cpu(i)
8364 unregister_fair_sched_group(tg, i);
8365 list_del_rcu(&tg->siblings);
8366 spin_unlock_irqrestore(&task_group_lock, flags);
8368 /* wait for any ongoing reference to this group to finish */
8369 synchronize_sched();
8372 * Now we are free to modify the group's share on each cpu
8373 * w/o tripping rebalance_share or load_balance_fair.
8375 tg->shares = shares;
8376 for_each_possible_cpu(i) {
8380 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8381 set_se_shares(tg->se[i], shares);
8385 * Enable load balance activity on this group, by inserting it back on
8386 * each cpu's rq->leaf_cfs_rq_list.
8388 spin_lock_irqsave(&task_group_lock, flags);
8389 for_each_possible_cpu(i)
8390 register_fair_sched_group(tg, i);
8391 list_add_rcu(&tg->siblings, &tg->parent->children);
8392 spin_unlock_irqrestore(&task_group_lock, flags);
8394 mutex_unlock(&shares_mutex);
8398 unsigned long sched_group_shares(struct task_group *tg)
8404 #ifdef CONFIG_RT_GROUP_SCHED
8406 * Ensure that the real time constraints are schedulable.
8408 static DEFINE_MUTEX(rt_constraints_mutex);
8410 static unsigned long to_ratio(u64 period, u64 runtime)
8412 if (runtime == RUNTIME_INF)
8415 return div64_u64(runtime << 20, period);
8418 /* Must be called with tasklist_lock held */
8419 static inline int tg_has_rt_tasks(struct task_group *tg)
8421 struct task_struct *g, *p;
8423 do_each_thread(g, p) {
8424 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8426 } while_each_thread(g, p);
8431 struct rt_schedulable_data {
8432 struct task_group *tg;
8437 static int tg_schedulable(struct task_group *tg, void *data)
8439 struct rt_schedulable_data *d = data;
8440 struct task_group *child;
8441 unsigned long total, sum = 0;
8442 u64 period, runtime;
8444 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8445 runtime = tg->rt_bandwidth.rt_runtime;
8448 period = d->rt_period;
8449 runtime = d->rt_runtime;
8453 * Cannot have more runtime than the period.
8455 if (runtime > period && runtime != RUNTIME_INF)
8459 * Ensure we don't starve existing RT tasks.
8461 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8464 total = to_ratio(period, runtime);
8467 * Nobody can have more than the global setting allows.
8469 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8473 * The sum of our children's runtime should not exceed our own.
8475 list_for_each_entry_rcu(child, &tg->children, siblings) {
8476 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8477 runtime = child->rt_bandwidth.rt_runtime;
8479 if (child == d->tg) {
8480 period = d->rt_period;
8481 runtime = d->rt_runtime;
8484 sum += to_ratio(period, runtime);
8493 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8495 struct rt_schedulable_data data = {
8497 .rt_period = period,
8498 .rt_runtime = runtime,
8501 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8504 static int tg_set_bandwidth(struct task_group *tg,
8505 u64 rt_period, u64 rt_runtime)
8509 mutex_lock(&rt_constraints_mutex);
8510 read_lock(&tasklist_lock);
8511 err = __rt_schedulable(tg, rt_period, rt_runtime);
8515 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8516 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8517 tg->rt_bandwidth.rt_runtime = rt_runtime;
8519 for_each_possible_cpu(i) {
8520 struct rt_rq *rt_rq = tg->rt_rq[i];
8522 raw_spin_lock(&rt_rq->rt_runtime_lock);
8523 rt_rq->rt_runtime = rt_runtime;
8524 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8526 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8528 read_unlock(&tasklist_lock);
8529 mutex_unlock(&rt_constraints_mutex);
8534 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8536 u64 rt_runtime, rt_period;
8538 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8539 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8540 if (rt_runtime_us < 0)
8541 rt_runtime = RUNTIME_INF;
8543 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8546 long sched_group_rt_runtime(struct task_group *tg)
8550 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8553 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8554 do_div(rt_runtime_us, NSEC_PER_USEC);
8555 return rt_runtime_us;
8558 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8560 u64 rt_runtime, rt_period;
8562 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8563 rt_runtime = tg->rt_bandwidth.rt_runtime;
8568 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8571 long sched_group_rt_period(struct task_group *tg)
8575 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8576 do_div(rt_period_us, NSEC_PER_USEC);
8577 return rt_period_us;
8580 static int sched_rt_global_constraints(void)
8582 u64 runtime, period;
8585 if (sysctl_sched_rt_period <= 0)
8588 runtime = global_rt_runtime();
8589 period = global_rt_period();
8592 * Sanity check on the sysctl variables.
8594 if (runtime > period && runtime != RUNTIME_INF)
8597 mutex_lock(&rt_constraints_mutex);
8598 read_lock(&tasklist_lock);
8599 ret = __rt_schedulable(NULL, 0, 0);
8600 read_unlock(&tasklist_lock);
8601 mutex_unlock(&rt_constraints_mutex);
8606 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8608 /* Don't accept realtime tasks when there is no way for them to run */
8609 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8615 #else /* !CONFIG_RT_GROUP_SCHED */
8616 static int sched_rt_global_constraints(void)
8618 unsigned long flags;
8621 if (sysctl_sched_rt_period <= 0)
8625 * There's always some RT tasks in the root group
8626 * -- migration, kstopmachine etc..
8628 if (sysctl_sched_rt_runtime == 0)
8631 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8632 for_each_possible_cpu(i) {
8633 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8635 raw_spin_lock(&rt_rq->rt_runtime_lock);
8636 rt_rq->rt_runtime = global_rt_runtime();
8637 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8639 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8643 #endif /* CONFIG_RT_GROUP_SCHED */
8645 int sched_rt_handler(struct ctl_table *table, int write,
8646 void __user *buffer, size_t *lenp,
8650 int old_period, old_runtime;
8651 static DEFINE_MUTEX(mutex);
8654 old_period = sysctl_sched_rt_period;
8655 old_runtime = sysctl_sched_rt_runtime;
8657 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8659 if (!ret && write) {
8660 ret = sched_rt_global_constraints();
8662 sysctl_sched_rt_period = old_period;
8663 sysctl_sched_rt_runtime = old_runtime;
8665 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8666 def_rt_bandwidth.rt_period =
8667 ns_to_ktime(global_rt_period());
8670 mutex_unlock(&mutex);
8675 #ifdef CONFIG_CGROUP_SCHED
8677 /* return corresponding task_group object of a cgroup */
8678 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8680 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8681 struct task_group, css);
8684 static struct cgroup_subsys_state *
8685 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8687 struct task_group *tg, *parent;
8689 if (!cgrp->parent) {
8690 /* This is early initialization for the top cgroup */
8691 return &init_task_group.css;
8694 parent = cgroup_tg(cgrp->parent);
8695 tg = sched_create_group(parent);
8697 return ERR_PTR(-ENOMEM);
8703 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8705 struct task_group *tg = cgroup_tg(cgrp);
8707 sched_destroy_group(tg);
8711 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8713 #ifdef CONFIG_RT_GROUP_SCHED
8714 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8717 /* We don't support RT-tasks being in separate groups */
8718 if (tsk->sched_class != &fair_sched_class)
8725 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8726 struct task_struct *tsk, bool threadgroup)
8728 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8732 struct task_struct *c;
8734 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8735 retval = cpu_cgroup_can_attach_task(cgrp, c);
8747 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8748 struct cgroup *old_cont, struct task_struct *tsk,
8751 sched_move_task(tsk);
8753 struct task_struct *c;
8755 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8762 #ifdef CONFIG_FAIR_GROUP_SCHED
8763 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8766 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8769 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8771 struct task_group *tg = cgroup_tg(cgrp);
8773 return (u64) tg->shares;
8775 #endif /* CONFIG_FAIR_GROUP_SCHED */
8777 #ifdef CONFIG_RT_GROUP_SCHED
8778 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8781 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8784 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8786 return sched_group_rt_runtime(cgroup_tg(cgrp));
8789 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8792 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8795 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8797 return sched_group_rt_period(cgroup_tg(cgrp));
8799 #endif /* CONFIG_RT_GROUP_SCHED */
8801 static struct cftype cpu_files[] = {
8802 #ifdef CONFIG_FAIR_GROUP_SCHED
8805 .read_u64 = cpu_shares_read_u64,
8806 .write_u64 = cpu_shares_write_u64,
8809 #ifdef CONFIG_RT_GROUP_SCHED
8811 .name = "rt_runtime_us",
8812 .read_s64 = cpu_rt_runtime_read,
8813 .write_s64 = cpu_rt_runtime_write,
8816 .name = "rt_period_us",
8817 .read_u64 = cpu_rt_period_read_uint,
8818 .write_u64 = cpu_rt_period_write_uint,
8823 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8825 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8828 struct cgroup_subsys cpu_cgroup_subsys = {
8830 .create = cpu_cgroup_create,
8831 .destroy = cpu_cgroup_destroy,
8832 .can_attach = cpu_cgroup_can_attach,
8833 .attach = cpu_cgroup_attach,
8834 .populate = cpu_cgroup_populate,
8835 .subsys_id = cpu_cgroup_subsys_id,
8839 #endif /* CONFIG_CGROUP_SCHED */
8841 #ifdef CONFIG_CGROUP_CPUACCT
8844 * CPU accounting code for task groups.
8846 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8847 * (balbir@in.ibm.com).
8850 /* track cpu usage of a group of tasks and its child groups */
8852 struct cgroup_subsys_state css;
8853 /* cpuusage holds pointer to a u64-type object on every cpu */
8854 u64 __percpu *cpuusage;
8855 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8856 struct cpuacct *parent;
8859 struct cgroup_subsys cpuacct_subsys;
8861 /* return cpu accounting group corresponding to this container */
8862 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8864 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8865 struct cpuacct, css);
8868 /* return cpu accounting group to which this task belongs */
8869 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8871 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8872 struct cpuacct, css);
8875 /* create a new cpu accounting group */
8876 static struct cgroup_subsys_state *cpuacct_create(
8877 struct cgroup_subsys *ss, struct cgroup *cgrp)
8879 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8885 ca->cpuusage = alloc_percpu(u64);
8889 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8890 if (percpu_counter_init(&ca->cpustat[i], 0))
8891 goto out_free_counters;
8894 ca->parent = cgroup_ca(cgrp->parent);
8900 percpu_counter_destroy(&ca->cpustat[i]);
8901 free_percpu(ca->cpuusage);
8905 return ERR_PTR(-ENOMEM);
8908 /* destroy an existing cpu accounting group */
8910 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8912 struct cpuacct *ca = cgroup_ca(cgrp);
8915 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8916 percpu_counter_destroy(&ca->cpustat[i]);
8917 free_percpu(ca->cpuusage);
8921 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8923 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8926 #ifndef CONFIG_64BIT
8928 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8930 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8932 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8940 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8942 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8944 #ifndef CONFIG_64BIT
8946 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8948 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8950 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8956 /* return total cpu usage (in nanoseconds) of a group */
8957 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8959 struct cpuacct *ca = cgroup_ca(cgrp);
8960 u64 totalcpuusage = 0;
8963 for_each_present_cpu(i)
8964 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8966 return totalcpuusage;
8969 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8972 struct cpuacct *ca = cgroup_ca(cgrp);
8981 for_each_present_cpu(i)
8982 cpuacct_cpuusage_write(ca, i, 0);
8988 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8991 struct cpuacct *ca = cgroup_ca(cgroup);
8995 for_each_present_cpu(i) {
8996 percpu = cpuacct_cpuusage_read(ca, i);
8997 seq_printf(m, "%llu ", (unsigned long long) percpu);
8999 seq_printf(m, "\n");
9003 static const char *cpuacct_stat_desc[] = {
9004 [CPUACCT_STAT_USER] = "user",
9005 [CPUACCT_STAT_SYSTEM] = "system",
9008 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9009 struct cgroup_map_cb *cb)
9011 struct cpuacct *ca = cgroup_ca(cgrp);
9014 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9015 s64 val = percpu_counter_read(&ca->cpustat[i]);
9016 val = cputime64_to_clock_t(val);
9017 cb->fill(cb, cpuacct_stat_desc[i], val);
9022 static struct cftype files[] = {
9025 .read_u64 = cpuusage_read,
9026 .write_u64 = cpuusage_write,
9029 .name = "usage_percpu",
9030 .read_seq_string = cpuacct_percpu_seq_read,
9034 .read_map = cpuacct_stats_show,
9038 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9040 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9044 * charge this task's execution time to its accounting group.
9046 * called with rq->lock held.
9048 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9053 if (unlikely(!cpuacct_subsys.active))
9056 cpu = task_cpu(tsk);
9062 for (; ca; ca = ca->parent) {
9063 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9064 *cpuusage += cputime;
9071 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9072 * in cputime_t units. As a result, cpuacct_update_stats calls
9073 * percpu_counter_add with values large enough to always overflow the
9074 * per cpu batch limit causing bad SMP scalability.
9076 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9077 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9078 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9081 #define CPUACCT_BATCH \
9082 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9084 #define CPUACCT_BATCH 0
9088 * Charge the system/user time to the task's accounting group.
9090 static void cpuacct_update_stats(struct task_struct *tsk,
9091 enum cpuacct_stat_index idx, cputime_t val)
9094 int batch = CPUACCT_BATCH;
9096 if (unlikely(!cpuacct_subsys.active))
9103 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9109 struct cgroup_subsys cpuacct_subsys = {
9111 .create = cpuacct_create,
9112 .destroy = cpuacct_destroy,
9113 .populate = cpuacct_populate,
9114 .subsys_id = cpuacct_subsys_id,
9116 #endif /* CONFIG_CGROUP_CPUACCT */
9120 void synchronize_sched_expedited(void)
9124 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9126 #else /* #ifndef CONFIG_SMP */
9128 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
9130 static int synchronize_sched_expedited_cpu_stop(void *data)
9133 * There must be a full memory barrier on each affected CPU
9134 * between the time that try_stop_cpus() is called and the
9135 * time that it returns.
9137 * In the current initial implementation of cpu_stop, the
9138 * above condition is already met when the control reaches
9139 * this point and the following smp_mb() is not strictly
9140 * necessary. Do smp_mb() anyway for documentation and
9141 * robustness against future implementation changes.
9143 smp_mb(); /* See above comment block. */
9148 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9149 * approach to force grace period to end quickly. This consumes
9150 * significant time on all CPUs, and is thus not recommended for
9151 * any sort of common-case code.
9153 * Note that it is illegal to call this function while holding any
9154 * lock that is acquired by a CPU-hotplug notifier. Failing to
9155 * observe this restriction will result in deadlock.
9157 void synchronize_sched_expedited(void)
9159 int snap, trycount = 0;
9161 smp_mb(); /* ensure prior mod happens before capturing snap. */
9162 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
9164 while (try_stop_cpus(cpu_online_mask,
9165 synchronize_sched_expedited_cpu_stop,
9168 if (trycount++ < 10)
9169 udelay(trycount * num_online_cpus());
9171 synchronize_sched();
9174 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
9175 smp_mb(); /* ensure test happens before caller kfree */
9180 atomic_inc(&synchronize_sched_expedited_count);
9181 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9184 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9186 #endif /* #else #ifndef CONFIG_SMP */